id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2309.14491-en-1
|
2309.14491-en-1.jpg
|
trained vision-language model to realize open-vocabulary
3D perception in the wild.
We propose a novel paradigm of Lnsupervised 3D Perception with 2D Vision-Language distillation (UP-VL). Specifically, by incorporating a pre-trained vision-language model, UP-VL can generate auto labels with substantially higher quality for objects in arbitrary motion states, compared to the latest work by Najibi et $al.[36].$
With our auto labels, we propose to co-train a 3D object detector with a knowledge distillation task, which can achieve two goals simultaneously, $i.e.$ improving detection quality and transferring semantic features from 2D image pixels to 3D LiDAR points. The perception model therefore is capable of detecting all traffic participants and thanks to the distilled open-vocabulary features, we can flexibly query the detector's output embedding with text prompts, for preserving specific types of objects at inference time (see Figure 1 for some examples).
We summarize the contributions of UP-VL as follows:
$\bullet$ UP-VL achieves state-of-the-art performance on un-
supervised 3D perception (detection and tracking) of
moving objects for autonomous driving.
·UP-VL introduces semantic-aware unsupervised de-
tection for objects in any motion state, a first in the field of autonomous driving. This breakthrough eliminates the information bottleneck that has plagued previous work[36], where class-agnostic auto labels were used, covering only moving objects with a speed above a predetermined threshold.
· UP-VL enables 3D open-vocabulary detection of novel
objects in the wild, with queries specified by users at inference time, therefore removing the need to recollect data or re-train models.
## $2. \textbf{Related works}$
Vision-language training. Contrastive vision language training on billions of image-text training pairs resulted in impressive improvements in the tasks of open-set and zeroshot image classification and language related applications [39, 20, 58]. More recently, open-set object localization in 2D images has been shown to benefit from such abundant image-text data as well. Specifically,[21,12,25,60,61,33] used image-text training to improve the open-set capability of 2D object detectors and [24,11] explored the use of large-scale scene-level vision-language data for the task of open-set 2D semantic segmentation. Recent research [31, 22, 51, 13, 18] has begun to explore the application of 2D vision-language pre-training in 3D perception tasks. However, these studies focused on static indoor scenario where the scene is small-scale and the RGB-D data is captured in high-resolution. Here we design a multi-modal
pipeline that leverages vision-language pre-training for unsupervised open-set 3D perception in complex, sparse, and occlusion-rich environments for autonomous driving.
Unsupervised 3D object detection. Unsupervised 3D object detection from LiDAR data is largely underexplored [7,54,50,28,36]. Dewan $etal.[7]$ proposed a model-free method to detect and track the visible part of objects, by using the motion cues from LiDAR sequences. However, this approach is incapable of generating amodal bounding boxes which is essential for autonomous driving. Cen et al.[3] relied on a supervised detector to produce proposals of unknown categories. However, this approach requires full supervision to train the base detector and has limited generalization capability to only semantically similar categories. Wong et $al.[54]$ identified unknown instances via supervised segmentation and clustering, which by design cannot generate amodal boxes from partial observations. Most recently, Najibi $etal.[36]$ developed an unsupervised auto meta labeling pipeline to generate pseudo labels for moving objects, which can be used to train real-time 3D detection models. This approach fails to provide semantics to detection boxes and ignores static objects, which limits its practical utility. Compared to all previous efforts, we realize open-vocabulary unsupervised 3D detection for both static and moving objects, by leveraging vision-language pre-training, and benchmark our system on the realistic and challenging scenario of autonomous driving. While utilizing 2D vision-language models that may have been pre-trained with human annotations, we avoid the need for any additional 3D labels within our paradigm, thereby creating a pragmatically unsupervised setting.
LiDAR 3D object detection. Most previous works focused on developing performant model architectures in the fully supervised setting, without considering the generalization capability to long-tail cases and unknown object types that are prevalent in the dynamic real world. These methods can be categorized into point based [43, 38, 56, 44, 35, 26], voxelization based [8,52,46,37,55,45,63,23,53,57,59, 5, 30], perspective projection based [32, 2, 10], and feature fusion [49,6,62,14,42]. Recent research also explore transferring knowledge from image for 3D point cloud understanding [40, 29, 19, 4]. Our method is compatible with any 3D detector, extending it to handle the open-set settings.
## $3. \textbf{Method}$
We present UP-VL, a new approach for unsupervised open-vocabulary 3D detection and tracking of traffic par-
UP-VL advances the previous state-of-the-
ticipants.
art[36] which was limited to class-agnostic detection of moving-only objects in two main directions: 1) It enables class-aware open-set 3D detection by incorporating openvocabulary text queries at inference time, and 2) It is able
2
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.106689453125
],
[
0.080078125,
0.10797119140625,
0.253173828125,
0.10797119140625,
0.253173828125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.098876953125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.098876953125,
0.1363525390625
],
[
0.50244140625,
0.12371826171875,
0.85107421875,
0.12371826171875,
0.85107421875,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50244140625,
0.14453125,
0.8896484375,
0.14453125,
0.8896484375,
0.1578369140625,
0.50244140625,
0.1578369140625
],
[
0.07843017578125,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.50244140625,
0.1610107421875,
0.890625,
0.1610107421875,
0.890625,
0.1749267578125,
0.50244140625,
0.1749267578125
],
[
0.5009765625,
0.1749267578125,
0.89208984375,
0.175537109375,
0.89208984375,
0.1900634765625,
0.5009765625,
0.189453125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1900634765625,
0.89208984375,
0.1900634765625,
0.89208984375,
0.203857421875,
0.50244140625,
0.203857421875
],
[
0.07843017578125,
0.1988525390625,
0.371826171875,
0.1976318359375,
0.371826171875,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.50244140625,
0.205810546875,
0.890625,
0.205810546875,
0.890625,
0.2197265625,
0.50244140625,
0.2197265625
],
[
0.09722900390625,
0.212158203125,
0.468017578125,
0.2108154296875,
0.468017578125,
0.2259521484375,
0.09722900390625,
0.2266845703125
],
[
0.50244140625,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.50244140625,
0.236083984375,
0.89208984375,
0.236083984375,
0.89208984375,
0.25,
0.50244140625,
0.25
],
[
0.080078125,
0.24365234375,
0.46728515625,
0.24365234375,
0.46728515625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.251953125,
0.890625,
0.251953125,
0.890625,
0.26513671875,
0.50244140625,
0.26513671875
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.267578125,
0.89208984375,
0.267578125,
0.89208984375,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.50341796875,
0.2822265625,
0.89208984375,
0.2822265625,
0.89208984375,
0.29541015625,
0.50341796875,
0.29541015625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.29736328125,
0.89208984375,
0.29736328125,
0.89208984375,
0.311279296875,
0.50244140625,
0.311279296875
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.3125,
0.89208984375,
0.3125,
0.89208984375,
0.32568359375,
0.50244140625,
0.32568359375
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.333251953125
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.50341796875,
0.328857421875,
0.8896484375,
0.328857421875,
0.8896484375,
0.339111328125,
0.50341796875,
0.339111328125
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.50244140625,
0.35595703125
],
[
0.08087158203125,
0.35107421875,
0.30224609375,
0.35107421875,
0.30224609375,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50244140625,
0.356689453125,
0.890625,
0.356689453125,
0.890625,
0.37060546875,
0.50244140625,
0.37060546875
],
[
0.1004638671875,
0.363037109375,
0.45751953125,
0.363037109375,
0.45751953125,
0.376953125,
0.1004638671875,
0.376953125
],
[
0.50244140625,
0.37255859375,
0.89208984375,
0.37255859375,
0.89208984375,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.09722900390625,
0.389404296875,
0.468017578125,
0.39013671875,
0.468017578125,
0.40478515625,
0.09722900390625,
0.404052734375
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.1094970703125,
0.406005859375,
0.468017578125,
0.406005859375,
0.468017578125,
0.42041015625,
0.1094970703125,
0.42041015625
],
[
0.50341796875,
0.404052734375,
0.890625,
0.404052734375,
0.890625,
0.4140625,
0.50341796875,
0.4140625
],
[
0.1094970703125,
0.42041015625,
0.37744140625,
0.421630859375,
0.37744140625,
0.436279296875,
0.1094970703125,
0.43505859375
],
[
0.50341796875,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.50341796875,
0.43310546875,
0.890625,
0.43310546875,
0.890625,
0.4462890625,
0.50341796875,
0.4462890625
],
[
0.09722900390625,
0.443115234375,
0.46728515625,
0.4443359375,
0.46728515625,
0.458984375,
0.09722900390625,
0.458251953125
],
[
0.5009765625,
0.448974609375,
0.890625,
0.4462890625,
0.890625,
0.460205078125,
0.5009765625,
0.462158203125
],
[
0.111083984375,
0.4609375,
0.4697265625,
0.4609375,
0.4697265625,
0.474853515625,
0.111083984375,
0.474853515625
],
[
0.5009765625,
0.4609375,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.5009765625,
0.47607421875
],
[
0.111083984375,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.111083984375,
0.490478515625
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.111083984375,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.111083984375,
0.5048828125
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.11273193359375,
0.5068359375,
0.468017578125,
0.5068359375,
0.468017578125,
0.52099609375,
0.11273193359375,
0.52099609375
],
[
0.50244140625,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50244140625,
0.521484375
],
[
0.111083984375,
0.521484375,
0.4697265625,
0.521484375,
0.4697265625,
0.5361328125,
0.111083984375,
0.5361328125
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.1094970703125,
0.5361328125,
0.28759765625,
0.5361328125,
0.28759765625,
0.5498046875,
0.1094970703125,
0.5498046875
],
[
0.5009765625,
0.537109375,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.5009765625,
0.5517578125
],
[
0.5009765625,
0.55322265625,
0.85546875,
0.55322265625,
0.85546875,
0.56640625,
0.5009765625,
0.56640625
],
[
0.09722900390625,
0.5595703125,
0.468017578125,
0.56005859375,
0.468017578125,
0.57470703125,
0.09722900390625,
0.57373046875
],
[
0.111083984375,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.111083984375,
0.59033203125
],
[
0.50244140625,
0.57568359375,
0.890625,
0.57568359375,
0.890625,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.1094970703125,
0.59033203125,
0.468017578125,
0.59130859375,
0.468017578125,
0.60595703125,
0.1094970703125,
0.60498046875
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.111083984375,
0.60693359375,
0.314453125,
0.60693359375,
0.314453125,
0.62060546875,
0.111083984375,
0.62060546875
],
[
0.50341796875,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.50244140625,
0.62255859375,
0.89208984375,
0.62255859375,
0.89208984375,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.07843017578125,
0.6337890625,
0.223876953125,
0.63623046875,
0.223876953125,
0.65380859375,
0.07843017578125,
0.6513671875
],
[
0.50244140625,
0.63720703125,
0.89208984375,
0.63720703125,
0.89208984375,
0.65087890625,
0.50244140625,
0.65087890625
],
[
0.50244140625,
0.65283203125,
0.890625,
0.65283203125,
0.890625,
0.66650390625,
0.50244140625,
0.66650390625
],
[
0.08087158203125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.50341796875,
0.66943359375,
0.8896484375,
0.66943359375,
0.8896484375,
0.6796875,
0.50341796875,
0.6796875
],
[
0.080078125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.681640625,
0.89208984375,
0.68115234375,
0.89208984375,
0.69580078125,
0.50244140625,
0.6962890625
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5009765625,
0.69580078125,
0.89208984375,
0.69775390625,
0.89208984375,
0.7119140625,
0.5009765625,
0.71044921875
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.7119140625,
0.89208984375,
0.7119140625,
0.89208984375,
0.72607421875,
0.50244140625,
0.72607421875
],
[
0.08087158203125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.08087158203125,
0.734375
],
[
0.50244140625,
0.72802734375,
0.89208984375,
0.72802734375,
0.89208984375,
0.7412109375,
0.50244140625,
0.7412109375
],
[
0.07843017578125,
0.734375,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.07843017578125,
0.74755859375
],
[
0.50244140625,
0.74365234375,
0.890625,
0.74365234375,
0.890625,
0.7568359375,
0.50244140625,
0.7568359375
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.77294921875,
0.5908203125,
0.77294921875,
0.5908203125,
0.787109375,
0.50244140625,
0.787109375
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.521484375,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.468017578125,
0.88525390625,
0.468017578125,
0.8994140625,
0.07843017578125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"trained vision-language model to realize open-vocabulary",
"pipeline that leverages vision-language pre-training for un-",
"3D perception in the wild.",
"supervised open-set 3D perception in complex, sparse, and",
"We propose a novel paradigm of Unsupervised 3D",
"occlusion-rich environments for autonomous driving.",
"Perception with 2D Vision-Language distillation (UP-VL)",
"Unsupervised 3D object detection.Unsupervised 3D",
"Specifically, by incorporating a pre-trained vision-language",
"model, UP-VL can generate auto labels with substantially",
"object detection from LiDAR data is largely under-",
"explored [7, 54, 50, 28, 36]. Dewan et al. [7] proposed a",
"higher quality for objects in arbitrary motion states, com-",
"model-free method to detect and track the visible part of",
"pared to the latest work by Najibi et al. [36].",
"objects, by using the motion cues from LiDAR sequences.",
"With our auto labels, we propose to co-train a 3D ob-",
"However, this approach is incapable of generating amodal",
"ject detector with aknowledge distillation task, which can",
"bounding boxes which is essential for autonomous driv-",
"achieve two goals simultaneously, i.e. improving detection",
"ing. Cen et al. [3] relied on a supervised detector to pro-",
"quality and transferring semantic features from 2D image",
"pixels to 3D LiDAR points. The perception model there-",
"duce proposals of unknown categories. However, this ap-",
"proach requires full supervision to train the base detector",
"fore is capable of detecting all traffic participants and thanks",
"and has limited generalization capability to only semanti-",
"to the distilled open-vocabulary features, we can flexibly",
"cally similar categories. Wong et al. [54] identified un-",
"query the detector's output embedding with text prompts.",
"for preserving specific types of objects at inference time",
"known instances via supervised segmentation and cluster-",
"ing, which by design cannot generate amodal boxes from",
"(seeFigurelforsomeexamples).",
"partial observations. Most recently, Najibi et al. [36] devel-",
"We summarize the contributions of UP-VL as follows:",
"oped an unsupervised auto meta labeling pipeline to gener-",
": UP-VL achieves state-of-the-art performance on un-",
"ate pseudo labels for moving objects, which can be used to",
"supervised 3D perception (detection and tracking) of",
"trainreal-time3D detectionmodels.Thisapproachfailsto",
"moving objects for autonomous driving.",
"provide semantics to detection boxes and ignores static ob-",
"jects, which limits its practical utility. Compared to all pre-",
". UP-VL introduces semantic-aware unsupervised de-",
"vious efforts, we realize open-vocabulary unsupervised 3D",
"tection for objects in any motion state, a first in the",
"detection for both static and moving objects, by leveraging",
"field of autonomous driving. This breakthrough elimi-",
"vision-language pre-training, and benchmark our system on",
"nates the information bottleneck that has plagued pre-",
"the realistic and challenging scenario of autonomous driv-",
"vious work [36], where class-agnostic auto labels were",
"ing. While utilizing 2D vision-language models that may",
"used, covering only moving objects with a speed above",
"have been pre-trained with human annotations, we avoid",
"a predetermined threshold.",
"the need for any additional 3D labels within our paradigm,",
"thereby creating a pragmatically unsupervised setting.",
".UP-VL enables 3D open-vocabulary detection of novel",
"objects in the wild, with queries specified by users",
"LiDAR 3D object detection. Most previous works fo-",
"at inference time, therefore removing the need to re-",
"cused on developing performant model architectures in the",
"collect data or re-train models.",
"fully supervised setting, without considering the generaliza-",
"tion capability to long-tail cases and unknown object types",
"2. Related works",
"that are prevalent in the dynamic real world. These methods",
"can be categorized into point based [43, 38, 56, 44, 35, 26]",
"Vision-language training. Contrastive vision language",
"voxelizationbased[8,52,46,37,55,45,63,23,53.57,59.",
"training on billions of image-text training pairs resulted in",
"5, 30], perspective projection based [32, 2, 10], and feature",
"impressive improvements in the tasks of open-set and zero-",
"fusion [49, 6, 62, 14, 42]. Recent research also explore",
"shot image classification and language related applications",
"transferring knowledge from image for 3D point cloud un-",
"[39, 20, 58]. More recently, open-set object localization in",
"derstanding [40, 29, 19, 4]. Our method is compatible with",
"2D images has been shown to benefit from such abundant",
"any 3D detector, extending it to handle the open-set settings.",
"image-text data as well. Specifically, [21, 12, 25, 60, 61, 33]",
"used image-text training to improve the open-set capa-",
"3.Method",
"bility of 2D object detectors and [24, 11] explored the",
"use of large-scale scene-level vision-language data for the",
"We present UP-VL, a new approach for unsupervised",
"task of open-set 2D semantic segmentation. Recent re-",
"open-vocabulary 3D detection and tracking of traffic par-",
"search [31, 22, 51, 13, 18] has begun to explore the appli-",
"ticipants.UP-VL advances the previous state-of-the-",
"cation of 2D vision-language pre-training in 3D perception",
"art [36] which was limited to class-agnostic detection of",
"tasks. However, these studies focused on static indoor sce-",
"moving-onlyobjects in two main directions:1)It enables",
"nario where the scene is small-scale and the RGB-D data is",
"class-aware open-set 3D detection by incorporating open-",
"captured in high-resolution. Here we design a multi-modal",
"vocabulary text queries at inference time, and 2) It is able",
"2"
] |
{
"arxiv_id": "2309.14491",
"language": "en",
"page_idx": 1
}
|
||
2403.16092-en-1
|
2403.16092-en-1.jpg
|
effective options for offloading real-world testing.
Nevertheless, employing neural rendering techniques for AV testing gives rise to a fundamental concern: How reliably can conclusions drawn from simulated data be transferred to real data? To address this concern, it is essential to assess whether the system, trained on real data, interprets real and simulated data similarly, as highlighted in Fig. 1. This divergence, termed the real2sim gap [47], has received far less attention than its counterpart sim2real gap, which pertains to transferring knowledge acquired in simulation to the real world [11]. Traditionally, the real2sim problem has been addressed by improving the realism of rendered images. However, it is unknown how well common novel view synthesis (NVS) performance measures, such as PSNR, LPIPS [ 49] , SSIM [ 40] , or FID [ 10] , correlate with a small real2sim gap, making it difficult to state what quality a given NeRF must reach to be useful for AV testing. Yet another aspect is that we are typically interested in the rendering quality when deviating from the original trajectory, $i.e.$, in a setting where such metrics cannot be calculated due to the lack of ground truth data to compare with.
In this paper, we propose a novel perspective on reducing the gap between real and simulated data for different perception modules of an autonomous system. Rather than improving upon the rendering quality, we aim to make the perception models more robust to NeRF artifacts without degrading performance on real data. We believe this direction to be complementary to increasing NeRF performance, and a potential key for making scalable, virtual AV testing a reality. As a first step in this direction, we show that even simple data augmentation techniques can have a large effect on model robustness against NeRF artifacts.
Further, we perform the first extensive real2sim gap study on a large-scale AD dataset and assess the performance of three object detectors alongside an online mapping model on both real data and data from a state-of-the-art (SOTA) neural rendering method. Our investigation encompasses the impact of diverse data augmentation techniques during training, as well as the fidelity of NeRF renderings during inference. We find that integrating such data during model fine-tuning notably enhances their robustness to simulated data and, in some cases, even elevates performance on real data. Lastly, we investigate the correlation between the real2sim-gap and image reconstruction metrics to provide insights into what matters for applying NeRFs as simulators for AD data. We find LPIPS and FID to be strong indicators of the real2sim-gap, and that our proposed augmentations reduce the sensitivity to poor view synthesis.
## $2. \textbf{ Related work}$
Novel view synthesis for autonomous driving: NeRFs have emerged as a promising approach for simulating AD data. In contrast to game engine-based methods, NeRFs re-
move the need for manual asset creation and are optimized to create sensor-realistic renderings by design. However, a key challenge for NeRFs is handling the scale and dynamics of automotive scenes. Neural Scene Graphs [32], Panoptic Neural Fields [18] and Panoptic NeRF [7] separate the background from moving actors by modeling each component with a separate, rigid, multi-layer perceptron (MLP). Still, these methods struggle with scaling to large scenes due to the limited expressiveness of the MLP. S-NeRF[46] addresses this by building upon Mip-NeRF 360[2] to better handle unbounded scenes. However, its long training time makes it impractical to simulate many scenes. MARS [44] and UniSim [47] utilize the hash-grid representation from iNGP [29] and achieve efficient models, although with limitations on sensor configuration. NeuRAD [37] introduces efficient ways of modeling the important aspects of AD data, and achieves state-of-the-art performance across five AD datasets [1,3,8,41,45]. Some works in this domain [37, 43, 47] make efforts to tailor their evaluation to closed-loop AD simulation. Nonetheless, their testing sets are small in the context of AD perception, and their applicability to downstream tasks at a larger scale is unexplored. Studies with large simulated test sets have thus far exclusively been done using game engine-based simulators, see for instance [17], which compared to NeRFs require laboursome manual asset creation.
Perception for autonomous driving: Perception in 3D is a critical component of many autonomous driving solutions. Due to cameras' low cost and high availability, camera-only methods have been the subject of extensive research in recent years. For 3D object detection, FCOS3D [38] is a onestage monocular object detector, building upon the 2D object detector FCOS [36], but regressing targets in 3D rather than 2D. PETR [25] further supports the multi-view setting and instead builds upon the decoder architecture from DETR [4], adding encoding points in the camera frustum into the image features. BEVFormer [20] also adopts a query-based architecture for multi-view input but encodes features into a bird's eye view (BEV) representation instead.
In addition to detecting objects in 3D, many autonomous vehicles also estimate the road elements of their surroundings, also known as online mapping [19,21,22,26,48]. MapTRv2[22] is a current SOTA method that uses a vectorbased representation for detected road elements such as lane dividers and road boundaries. Similarly to BEVFormer, MapTRv2 encodes image features and lifts them into a BEV representation. The objects' class and geometry are estimated through a DETR-like[4] transformer decoder
Domain adaptation and multi-task learning: Domain adaptation (DA) is a field aiming to cope with issues arising when the training and evaluation data come from different distributions [6,24]. Although different DA definitions exist, unsupervised DA is the most commonly studied version
2
|
[
[
0.080078125,
0.0921630859375,
0.406005859375,
0.0921630859375,
0.406005859375,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.09722900390625,
0.106689453125,
0.4697265625,
0.1060791015625,
0.4697265625,
0.12060546875,
0.09722900390625,
0.121826171875
],
[
0.5009765625,
0.10797119140625,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.5009765625,
0.121826171875
],
[
0.08087158203125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.08087158203125,
0.1395263671875,
0.468017578125,
0.1395263671875,
0.468017578125,
0.149658203125,
0.08087158203125,
0.149658203125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.50244140625,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.08087158203125,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1956787109375,
0.08087158203125,
0.1956787109375
],
[
0.50341796875,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.080078125,
0.1962890625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.212158203125,
0.080078125,
0.2108154296875
],
[
0.50244140625,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50341796875,
0.2139892578125,
0.890625,
0.2139892578125,
0.890625,
0.2247314453125,
0.50341796875,
0.2247314453125
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.5009765625,
0.2279052734375,
0.890625,
0.2266845703125,
0.890625,
0.2412109375,
0.5009765625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.07843017578125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.07843017578125,
0.271484375
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.080078125,
0.28857421875,
0.4697265625,
0.28857421875,
0.4697265625,
0.3017578125,
0.080078125,
0.3017578125
],
[
0.50244140625,
0.2890625,
0.8935546875,
0.2890625,
0.8935546875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.3037109375,
0.471435546875,
0.3037109375,
0.471435546875,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.3037109375,
0.890625,
0.3037109375,
0.890625,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.07843017578125,
0.317626953125,
0.46728515625,
0.318115234375,
0.46728515625,
0.333251953125,
0.07843017578125,
0.33203125
],
[
0.5048828125,
0.32080078125,
0.890625,
0.32080078125,
0.890625,
0.330810546875,
0.5048828125,
0.330810546875
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.348388671875,
0.89208984375,
0.348388671875,
0.89208984375,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.5009765625,
0.363037109375,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.5009765625,
0.376953125
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50244140625,
0.37890625,
0.89208984375,
0.37890625,
0.89208984375,
0.392578125,
0.50244140625,
0.392578125
],
[
0.080078125,
0.39453125,
0.396240234375,
0.39453125,
0.396240234375,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.393310546875,
0.8896484375,
0.393310546875,
0.8896484375,
0.4072265625,
0.50244140625,
0.4072265625
],
[
0.09722900390625,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.09722900390625,
0.423583984375
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50341796875,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.436279296875,
0.50341796875,
0.436279296875
],
[
0.080078125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.5009765625,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.5009765625,
0.453857421875
],
[
0.07843017578125,
0.455078125,
0.4697265625,
0.45263671875,
0.4697265625,
0.46728515625,
0.07843017578125,
0.4697265625
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.07843017578125,
0.47021484375,
0.4697265625,
0.4697265625,
0.4697265625,
0.48291015625,
0.07843017578125,
0.484130859375
],
[
0.50244140625,
0.47021484375,
0.689453125,
0.47021484375,
0.689453125,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.5009765625,
0.486083984375,
0.89208984375,
0.4873046875,
0.89208984375,
0.501953125,
0.5009765625,
0.50048828125
],
[
0.080078125,
0.50048828125,
0.46728515625,
0.50048828125,
0.46728515625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50341796875,
0.5048828125,
0.8896484375,
0.5048828125,
0.8896484375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.07843017578125,
0.5146484375,
0.4697265625,
0.51513671875,
0.4697265625,
0.52978515625,
0.07843017578125,
0.52880859375
],
[
0.50244140625,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.53515625,
0.8896484375,
0.53515625,
0.8896484375,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.080078125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50244140625,
0.548828125,
0.890625,
0.548828125,
0.890625,
0.5625,
0.50244140625,
0.5625
],
[
0.080078125,
0.56005859375,
0.369384765625,
0.56005859375,
0.369384765625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.5009765625,
0.56298828125,
0.89208984375,
0.5625,
0.89208984375,
0.5771484375,
0.5009765625,
0.57763671875
],
[
0.096435546875,
0.57373046875,
0.468017578125,
0.57568359375,
0.468017578125,
0.59033203125,
0.096435546875,
0.58837890625
],
[
0.50244140625,
0.5791015625,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.080078125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50244140625,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.080078125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.080078125,
0.62060546875,
0.4697265625,
0.62060546875,
0.4697265625,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.50341796875,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.08087158203125,
0.63623046875,
0.468017578125,
0.63720703125,
0.468017578125,
0.6484375,
0.08087158203125,
0.64794921875
],
[
0.50244140625,
0.6396484375,
0.8935546875,
0.6396484375,
0.8935546875,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.080078125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66845703125,
0.50244140625,
0.66845703125
],
[
0.080078125,
0.66552734375,
0.468017578125,
0.66552734375,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.50341796875,
0.67041015625,
0.8896484375,
0.67041015625,
0.8896484375,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.07843017578125,
0.6796875,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.07843017578125,
0.693359375
],
[
0.521484375,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.521484375,
0.70068359375
],
[
0.080078125,
0.6962890625,
0.468017578125,
0.6962890625,
0.468017578125,
0.71044921875,
0.080078125,
0.71044921875
],
[
0.5048828125,
0.7041015625,
0.8896484375,
0.7041015625,
0.8896484375,
0.7138671875,
0.5048828125,
0.7138671875
],
[
0.08087158203125,
0.71337890625,
0.46728515625,
0.71337890625,
0.46728515625,
0.7236328125,
0.08087158203125,
0.7236328125
],
[
0.5009765625,
0.71630859375,
0.8896484375,
0.71630859375,
0.8896484375,
0.73095703125,
0.5009765625,
0.73095703125
],
[
0.080078125,
0.72607421875,
0.468017578125,
0.72607421875,
0.468017578125,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.50341796875,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.744140625,
0.50341796875,
0.744140625
],
[
0.080078125,
0.7412109375,
0.4697265625,
0.7412109375,
0.4697265625,
0.755859375,
0.080078125,
0.755859375
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.76025390625,
0.50341796875,
0.76025390625
],
[
0.080078125,
0.7568359375,
0.468017578125,
0.7568359375,
0.468017578125,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.50341796875,
0.7646484375,
0.8896484375,
0.7646484375,
0.8896484375,
0.7744140625,
0.50341796875,
0.7744140625
],
[
0.07843017578125,
0.77099609375,
0.4697265625,
0.771484375,
0.4697265625,
0.7861328125,
0.07843017578125,
0.78515625
],
[
0.5009765625,
0.77880859375,
0.890625,
0.7763671875,
0.890625,
0.79052734375,
0.5009765625,
0.79248046875
],
[
0.07843017578125,
0.78515625,
0.468017578125,
0.787109375,
0.468017578125,
0.8017578125,
0.07843017578125,
0.7998046875
],
[
0.50341796875,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50341796875,
0.806640625
],
[
0.07843017578125,
0.80322265625,
0.449462890625,
0.80322265625,
0.449462890625,
0.81640625,
0.07843017578125,
0.81640625
],
[
0.50244140625,
0.8095703125,
0.8505859375,
0.8095703125,
0.8505859375,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.080078125,
0.83056640625,
0.214111328125,
0.83056640625,
0.214111328125,
0.84521484375,
0.080078125,
0.84521484375
],
[
0.50341796875,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.5048828125,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.5048828125,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.46728515625,
0.88525390625,
0.46728515625,
0.8994140625,
0.07843017578125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.4755859375,
0.923828125,
0.49267578125,
0.92041015625,
0.496826171875,
0.93603515625,
0.48046875,
0.939453125
]
] |
[
"effective options for offloading real-world testing.",
"move the need for manual asset creation and are optimized",
"Nevertheless, employing neural rendering techniques for",
"to create sensor-realistic renderings by design. However, a",
"AV testing gives rise to a fundamental concern: How reli-",
"key challenge for NeRFs is handling the scale and dynamics",
"ably can conclusions drawn from simulated data be trans-",
"of automotive scenes. Neural Scene Graphs [32], Panop-",
"ferred to real data? To address this concern, it is essential",
"tic Neural Fields [18] and Panoptic NeRF [7] separate the",
"to assess whether the system, trained on real data, interprets",
"background from moving actors by modeling each compo-",
"real and simulated data similarly, as highlighted in Fig. 1.",
"nent with a separate, rigid, multi-layer perceptron (MLP).",
"This divergence, termed the real2sim gap [47], has received",
"Still, these methods struggle with scaling to large scenes",
"far less attention than its counterpart sim2real gap, which",
"duetothelimited expressivenessof theMLP.S-NeRF[46]",
"pertains to transferring knowledge acquired in simulation",
"addresses this by building upon Mip-NeRF 360 [2] to bet-",
"to the real world [11]. Traditionally, the real2sim prob-",
"ter handle unbounded scenes. However, its long training",
"lem has been addressed by improving the realism of ren-",
"time makes it impractical to simulate many scenes. MARS",
"dered images. However, it is unknown how well common",
"[44] and UniSim [47] utilize the hash-grid representation",
"novel view synthesis (NvS)performance measures, such as",
"from iNGP [29] and achieve efficient models, although with",
"PSNR, LPIPS [49], SSIM[40], or FID[10], correlate with a",
"limitations on sensor configuration. NeuRAD [37] intro-",
"small real2sim gap,making it difficult to state what quality",
"ducesefficientways ofmodelingtheimportantaspects of",
"a given NeRF must reach to be useful for AV testing. Yet",
"AD data, and achieves state-of-the-art performance across",
"another aspect is that we are typically interested in the ren-",
"five AD datasets [1, 3, 8, 41, 45]. Some works in this do-",
"dering quality when deviating from the original trajectory,",
"main[37,43,47] makeefforts totailor theirevaluationto",
"i.e.,in a setting wheresuchmetrics cannot be calculated due",
"closed-loop AD simulation. Nonetheless, their testing sets",
"to the lack of ground truth data to compare with",
"are small in the context of AD perception, and their appli-",
"In this paper, we propose a novel perspective on reduc-",
"cability to downstream tasks at a larger scale is unexplored.",
"ing the gap between real and simulated data for different",
"Studies with large simulated test sets have thus far exclu-",
"perception modules of an autonomous system. Rather than",
"sively been done using game engine-based simulators, see",
"improving upon the rendering quality, we aim to make the",
"for instance [17], which compared to NeRFs require labour-",
"perception models more robust to NeRF artifacts without",
"some manual asset creation.",
"degrading performance on real data. We believe this direc-",
"Perception for autonomous driving: Perception in 3D is a",
"tion to be complementary to increasing NeRF performance",
"critical component of many autonomous driving solutions.",
"and a potential key for making scalable, virtual AV testing",
"Due to cameras' low cost and high availability, camera-only",
"a reality. As a first step in this direction, we show that even",
"methodshavebeenthesubjectofextensiveresearchinre-",
"simple data augmentation techniques can have a large effect",
"cent years. For 3D object detection, FCOS3D [38] is a one-",
"on model robustness against NeRF artifacts.",
"stage monocular object detector, building upon the 2D ob-",
"Further, we perform the first extensive real2sim gap",
"ject detector FCOS [36], but regressing targets in 3D rather",
"study on a large-scale AD dataset and assess the perfor-",
"than 2D. PETR [25] further supports the multi-view set-",
"mance of three object detectors alongside an online map-",
"ting and instead builds upon the decoder architecture from",
"ping model on both real data and data from a state-of-the-art",
"DETR [4], adding encoding points in the camera frustum",
"(SOTA)neuralrenderingmethod.Ourinvestigationencom-",
"into the image features. BEVFormer [20] also adopts a",
"passes the impact of diverse data augmentation techniques",
"query-based architecture for multi-view input but encodes",
"during training, as well as the fidelity of NeRF renderings",
"featuresintoabird'seyeview(BEV)representationinstead.",
"during inference. We find that integrating such data during",
"In addition to detecting objects in 3D, many autonomous",
"model fine-tuning notably enhances their robustness to sim-",
"vehiclesalsoestimatetheroadelementsoftheirsurround-",
"ulateddataandin somecases,even elevatesperformance",
"ings, also known as online mapping [19, 21, 22, 26, 48]",
"on real data. Lastly, we investigate the correlation between",
"MapTRv2[22lisacurrentSOTAmethodthatusesavector-",
"the real2sim-gap and image reconstruction metrics to pro-",
"basedrepresentationfordetectedroadelementssuchaslane",
"vide insights into what matters for applying NeRFs as sim-",
"dividers and road boundaries.Similarly to BEVFormer",
"ulators for AD data. We find LPIPS and FID to be strong",
"MapTRv2encodesimagefeaturesandlifts themintoaBEV",
"indicators of the real2sim-gap, and that our proposed aug-",
"representation.The objects' class and geometry are esti-",
"mentations reduce the sensitivity to poor view synthesis.",
"mated through a DETR-like [4] transformer decoder.",
"2.Related work",
"Domain adaptation and multi-task learning: Domain",
"adaptation (DA) is a field aiming to cope with issues arising",
"Novel view synthesis for autonomous driving: NeRFs",
"when thetraining and evaluation data come from different",
"have emerged as a promising approach for simulating AD",
"distributions [6, 24]. Although different DA definitions ex-",
"data. In contrast to game engine-based methods, NeRFs re-",
"ist, unsupervised DA is the most commonly studied version",
"2"
] |
{
"arxiv_id": "2403.16092",
"language": "en",
"page_idx": 1
}
|
||
2403.16092-en-3
|
2403.16092-en-3.jpg
|
<FigureHere>
an image-to-image method. Given a real image, the model translates the image to the NeRF domain, effectively introducing typical NeRF artifacts. This enables us to vastly increase the amount of NeRF-like images during fine-tuning at a limited computational cost. We train the image-toimage model[39] using rendered images $\mathcal{D}_\mathrm{fine-tune}^\mathrm{nerf}$ and their corresponding real images. See Fig. 3 for visual examples of the different augmentation strategies.
### $4. \textbf{Results}$
In this section, we first describe our experimental setting. We then show how the real2sim gap differs for different perception models, tasks, and augmentation strategies. Next, we study how the models with the smallest gap perceive renderings with large changes in viewpoint where no corresponding real images have been collected. Finally, we show the correlation between common metrics for novel view synthesis and the real2sim gap.
Figure 3. Examples of our different data augmentation strategies to make perception models more robust.
current SOTA method for NVS on AD data, as our NeRFmethod. NeuRAD relies on a hashgrid representation [29] and upsampling techniques for efficient learning and inference of high-resolution data. To handle dynamic scenes, it models all actors to be rigid entities, which limits expressiveness for deformable actors such as pedestrians. Nonetheless, NeuRAD is currently the most performant NVS technique across multiple AD datasets. For the perception validation sequences, we train NeuRAD on all images except those held out for evaluating the perception models. For reference, statistics for the NeuRAD trainings are shown in Tab. 8, where we observe similar performance on images used for data augmentation as the ones used for evaluating the real2sim gap.
Dataset and tasks: Our evaluations are conducted on nuScenes, a widely recognized research ground for perception models. Compared to other large AD datasets such as Waymo Open Dataset [35], Argoverse2 [41] or Zenseact Open Dataset[1], nuScenes uses lower resolution cameras and lidars, posing an interesting challenge for neural rendering methods. We evaluate all perception models on a subset of the official validation split, namely scenes collected at daytime and without heavy rain. This is because lens flares and water spray cannot currently be handled by any neural rendering method for AD data. This filtering results in 111 scenes used for evaluation, see Appendix B.1 for details.
Further, we evaluate two tasks: 3D object detection and online mapping. These tasks consider complementary aspects of the real2sim gap, as the former focuses on foreground objects, while the latter targets the static background. Both tasks and corresponding models are described in more detail below.
3D object detection: We evaluate the gap across three camera-only 3D object detection (3DOD) models, chosen to represent different aspects of prevalent model architectures. Namely, we apply FCOS3D [38], a fully-convolutional monocular detector, PETR [25], a multi-view and 3D adaptation of DETR [4], and BEVFormer [20] a multi-view detector centered around the bird's-eye-view representa-
# 4.1. Experimental setup
We assess the real2sim gap by comparing the performance of multiple perception models when applied to real data versus their NeRF-rendered counterparts. We validate our approach to improving the robustness of perception models by implementing various augmentation strategies during fine-tuning of the models, as detailed in Sec. 3. Striving for a general approach adaptable to diverse models and tasks, we employ the same augmentation techniques across all models and tasks. Given the critical need to not sacrifice real-world performance, we also assess our augmentation methods on real data. Specifically, we test the three augmentation techniques described in Secs. 3.1 to 3.3; traditional image augmentations, rendered data, and imageto-image translation. See Appendix A.2 respectively Appendix A.4 for details regarding the datasets used during $\hat{\text{fine-tuning, and Appendix A.3 for details on training the}}$ image-to-image model.
Neural rendering: We opt to employ NeuRAD [37], the
4
|
[
[
0.318603515625,
0.175537109375,
0.449462890625,
0.175537109375,
0.449462890625,
0.189453125,
0.318603515625,
0.189453125
],
[
0.75732421875,
0.1749267578125,
0.82861328125,
0.1767578125,
0.82861328125,
0.1912841796875,
0.75732421875,
0.189453125
],
[
0.56396484375,
0.177978515625,
0.61376953125,
0.177978515625,
0.61376953125,
0.189453125,
0.56396484375,
0.189453125
],
[
0.146240234375,
0.177978515625,
0.2099609375,
0.177978515625,
0.2099609375,
0.189453125,
0.146240234375,
0.189453125
],
[
0.176513671875,
0.287353515625,
0.7939453125,
0.287353515625,
0.7939453125,
0.300537109375,
0.176513671875,
0.300537109375
],
[
0.080078125,
0.306884765625,
0.4697265625,
0.306884765625,
0.4697265625,
0.32080078125,
0.080078125,
0.32080078125
],
[
0.50244140625,
0.306884765625,
0.890625,
0.306884765625,
0.890625,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.08087158203125,
0.3212890625,
0.468017578125,
0.3212890625,
0.468017578125,
0.3359375,
0.08087158203125,
0.3359375
],
[
0.50244140625,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50244140625,
0.335205078125
],
[
0.08087158203125,
0.337158203125,
0.4697265625,
0.337158203125,
0.4697265625,
0.35107421875,
0.08087158203125,
0.35107421875
],
[
0.50244140625,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.35107421875,
0.50244140625,
0.35107421875
],
[
0.07843017578125,
0.35107421875,
0.468017578125,
0.3515625,
0.468017578125,
0.3662109375,
0.07843017578125,
0.365478515625
],
[
0.5009765625,
0.35107421875,
0.890625,
0.3515625,
0.890625,
0.3662109375,
0.5009765625,
0.365478515625
],
[
0.080078125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.080078125,
0.380615234375
],
[
0.50244140625,
0.367431640625,
0.8896484375,
0.367431640625,
0.8896484375,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.07843017578125,
0.3818359375,
0.4697265625,
0.380615234375,
0.4697265625,
0.395751953125,
0.07843017578125,
0.396484375
],
[
0.50341796875,
0.38330078125,
0.8896484375,
0.38330078125,
0.8896484375,
0.396484375,
0.50341796875,
0.396484375
],
[
0.080078125,
0.397705078125,
0.4697265625,
0.397705078125,
0.4697265625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.396484375,
0.89208984375,
0.410888671875,
0.50244140625,
0.410400390625
],
[
0.080078125,
0.41357421875,
0.343017578125,
0.41357421875,
0.343017578125,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50341796875,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50341796875,
0.425537109375
],
[
0.50341796875,
0.427978515625,
0.890625,
0.427978515625,
0.890625,
0.44140625,
0.50341796875,
0.44140625
],
[
0.07843017578125,
0.442626953125,
0.1641845703125,
0.4443359375,
0.1641845703125,
0.460205078125,
0.07843017578125,
0.458251953125
],
[
0.50341796875,
0.443115234375,
0.89208984375,
0.443115234375,
0.89208984375,
0.45703125,
0.50341796875,
0.45703125
],
[
0.50244140625,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.07843017578125,
0.4697265625,
0.468017578125,
0.47021484375,
0.468017578125,
0.48486328125,
0.07843017578125,
0.484130859375
],
[
0.50341796875,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.48486328125,
0.50341796875,
0.48486328125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.486083984375,
0.468017578125,
0.50048828125,
0.080078125,
0.5
],
[
0.50244140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.5009765625,
0.501953125,
0.689453125,
0.50390625,
0.689453125,
0.51806640625,
0.5009765625,
0.5166015625
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50341796875,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.50244140625,
0.5341796875,
0.8896484375,
0.53515625,
0.8896484375,
0.5498046875,
0.50244140625,
0.548828125
],
[
0.080078125,
0.546875,
0.468017578125,
0.546875,
0.468017578125,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.080078125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.5048828125,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.57763671875,
0.5048828125,
0.57763671875
],
[
0.07843017578125,
0.57568359375,
0.3203125,
0.5771484375,
0.3203125,
0.59130859375,
0.07843017578125,
0.59033203125
],
[
0.50341796875,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50341796875,
0.5947265625
],
[
0.50341796875,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.50341796875,
0.609375
],
[
0.080078125,
0.60302734375,
0.261474609375,
0.60498046875,
0.261474609375,
0.619140625,
0.080078125,
0.6171875
],
[
0.5048828125,
0.6123046875,
0.890625,
0.6123046875,
0.890625,
0.62255859375,
0.5048828125,
0.62255859375
],
[
0.07843017578125,
0.62548828125,
0.4697265625,
0.626953125,
0.4697265625,
0.64453125,
0.07843017578125,
0.64306640625
],
[
0.50244140625,
0.62548828125,
0.89208984375,
0.62548828125,
0.89208984375,
0.6396484375,
0.50244140625,
0.6396484375
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.5048828125,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65283203125,
0.5048828125,
0.65283203125
],
[
0.080078125,
0.6591796875,
0.4697265625,
0.6591796875,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.07843017578125,
0.6748046875,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.07843017578125,
0.68798828125
],
[
0.50244140625,
0.671875,
0.890625,
0.671875,
0.890625,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.5009765625,
0.68603515625,
0.87353515625,
0.68505859375,
0.87353515625,
0.69970703125,
0.5009765625,
0.70068359375
],
[
0.080078125,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.521484375,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.521484375,
0.71630859375
],
[
0.07843017578125,
0.7197265625,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.07843017578125,
0.73291015625
],
[
0.50244140625,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.50244140625,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.080078125,
0.74853515625,
0.46728515625,
0.74853515625,
0.46728515625,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50341796875,
0.75048828125,
0.88818359375,
0.75048828125,
0.88818359375,
0.76123046875,
0.50341796875,
0.76123046875
],
[
0.08087158203125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.7763671875,
0.08087158203125,
0.7763671875
],
[
0.50341796875,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.080078125,
0.77978515625,
0.471435546875,
0.77978515625,
0.471435546875,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.77880859375,
0.646484375,
0.77880859375,
0.646484375,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50244140625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.080078125,
0.82373046875
],
[
0.50341796875,
0.8125,
0.890625,
0.8125,
0.890625,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.080078125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.080078125,
0.84033203125,
0.468017578125,
0.84033203125,
0.468017578125,
0.85400390625,
0.080078125,
0.85400390625
],
[
0.50244140625,
0.84033203125,
0.89208984375,
0.84033203125,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85498046875,
0.468017578125,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.87060546875,
0.2344970703125,
0.87060546875,
0.2344970703125,
0.8837890625,
0.080078125,
0.8837890625
],
[
0.50244140625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.48193359375,
0.9267578125,
0.491943359375,
0.9267578125,
0.491943359375,
0.9345703125,
0.48193359375,
0.9345703125
]
] |
[
"(b) Image Augmentation",
"(d) Img2Img",
"(c) NeRF",
"(a) Original",
"Figure 3. Examples of our different data augmentation strategies to make perception models more robust.",
"an image-to-image method. Given a real image, the model",
"current SOTA method for NVS on AD data, as our NeRF-",
"translates the image to the NeRF domain, effectively intro-",
"method. NeuRAD relies on a hashgrid representation [29]",
"ducing typical NeRF artifacts. This enables us to vastly in-",
"and upsampling techniques for efficient learning and infer-",
"crease the amount of NeRF-like images during fine-tuning",
"ence of high-resolution data. To handle dynamic scenes,",
"at a limited computational cost. We train the image-to-",
"it models all actors to be rigid entities, which limits ex-",
"image model [39] using rendered images Dnerfune and their",
"pressiveness for deformable actors such as pedestrians.",
"corresponding real images. See Fig. 3 for visual examples",
"Nonetheless, NeuRAD is currently the most performant",
"of the different augmentation strategies.",
"NVS technique across multiple AD datasets. For the per-",
"ception validation sequences, we train NeuRAD on all im-",
"4.Results",
"ages except those held out for evaluating the perception",
"models. For reference, statistics for the NeuRAD trainings",
"In this section, we first describe our experimental setting.",
"areshowninTab.8.whereweobservesimilarperformance",
"We then show how thereal2sim gap differs for different per-",
"on images used for data augmentation as the ones used for",
"ception models, tasks, and augmentation strategies. Next,",
"evaluating the real2sim gap.",
"we study how the models with the smallest gap perceive",
"Dataset and tasks: Our evaluations are conducted on",
"renderings with large changes in viewpoint where no cor-",
"nuScenes, a widely recognized research ground for percep-",
"responding real images have been collected.Finally, we",
"tion models. Compared to other large AD datasets such",
"show the correlation between common metrics for novel",
"asWaymoOpenDataset[35l,Argoverse2[41lorZenseact",
"view synthesis and the real2sim gap.",
"Open Dataset [1], nuScenes uses lower resolution cameras",
"and lidars, posing an interesting challenge for neural render-",
"4.1.Experimental setup",
"ingmethods.Weevaluateallperceptionmodelsonasubset",
"We assess the real2sim gap by comparing the performance",
"of the official validation split, namely scenes collected at",
"of multiple perception models when applied to real data",
"daytimeand withoutheavyrain.Thisis becauselensflares",
"versus their NeRF-rendered counterparts. We validate our",
"and water spray cannot currently be handled by any neural",
"approachto improving therobustness of perception mod-",
"rendering method for AD data. This filtering results in 111",
"els by implementing various augmentation strategies dur-",
"scenes used for evaluation, see Appendix B.1 for details.",
"ing fine-tuning of the models, as detailed in Sec. 3. Striv-",
"Further, we evaluate two tasks:3D object detection",
"ing for a general approach adaptable to diverse models and",
"and online mapping:These tasks consider complemen-",
"tasks,weemploythe same augmentationtechniques across",
"tary aspects of the real2sim gap, as the former focuses on",
"all models and tasks. Given the critical need to not sacri-",
"foreground objects,whilethe lattertargetsthe static back",
"ficereal-world performance,we also assess ouraugmenta-",
"ground. Both tasks and corresponding models are described",
"tion methods on real data. Specifically, we test the three",
"inmore detail below.",
"augmentation techniques described in Secs. 3.1 to 3.3; tra-",
"3D object detection: We evaluate the gap across three",
"ditional image augmentations, rendered data, and image-",
"camera-only3Dobjectdetection(3DOD)models.chosento",
"to-image translation. See Appendix A.2 respectively Ap-",
"represent different aspects of prevalent model architectures.",
"pendix A.4 for details regarding the datasets used during",
"Namely,we apply FCOS3D [38], a fully-convolutional",
"fine-tuning, and Appendix A.3 for details on training the",
"monocular detector, PETR [25], a multi-view and 3D adap-",
"image-to-image model.",
"tation of DETR [4], and BEVFormer [20] a multi-view",
"Neural rendering: We opt to employ NeuRAD [37], the",
"detector centered around the bird's-eye-view representa-",
"4"
] |
{
"arxiv_id": "2403.16092",
"language": "en",
"page_idx": 3
}
|
||
2108.03151-en-2
|
2108.03151-en-2.jpg
|
Full-Duplex Strategy for Video Object Segmentation
the validity and rationality of our current design
under various conditions.
We provide more details to the conference version, such as backbone details (see Sec. 3.6.1), evaluation metrics (see Sec. 4.2), prediction selection ( see Sec. 4. 5. 1) , and post- processing techniques (see Sec. 4.5.2).
$\bullet$ We further present more performances under different thresholds used in the fnal results $(i.e.$, PR curve in Fig. 6). Additional test results $(i.e.$, DAVSOD$_19$-Normal25, DAVSOD$_19$-Difficult20) on a recent challenging dataset still show that our framework is superior to existing SOTA models (see Tab. 3).
### 2 Related Works
Depending on whether or not the frst frame of ground truth is given, the VOS task can be divided into two scenarios, $i. e. , \textit{semi- supervised}$ [109] and $unsupervised/zero$-shot [71]. Some typical semisupervised VOS models can be referred to [6,11, 12, 20, 25, 38, 47, 53, 64, 81, 85, 88, 102, 121, 126, 128, 133, 134]. This paper studies the unsupervised setting [140, 141], leaving the semi-supervised setting as our future work.
## 2.1 Unsupervised VOS
Although there are many works addressing the VOS task in a semi-supervised manner, $i.e.$, by supposing an object mask annotation is given in the first frame, other researchers have attempted to address the more challenging unsupervised VOS (U-VOS) problem. Early U-VOS models resort to low-level handcrafted features for heuristic segmentation inference, such as proposals [58,59,69,83], saliency priors [27,106,108], optical flow [100], or superpixels [32,33,123]. These traditional models have limited generalizability and thus low accuracy in highly dynamic and complex scenarios due to their lack of semantic information and high-level content understanding. Recently, RNNbased models [4,93,97,113,126,138] have become popular due to their better capability of capturing longterm dependencies and their use of deep learning. In this case, U-VOS is formulated as a recurrent modelling exploited with long-term temporal context.
How to combine motion cues with appearance features is a long-standing problem in this feld. To this end, Tokmakov $et\textit{al. [ 96] proposed to use the motion}$ patterns required from the video simply. However, their
3
method cannot accurately segment objects between two similar consecutive frames since it relies heavily on the guidance of optical flow. To resolve this, several works [21,92,97] have integrated the spatial and temporal features from the parallel network, which can be viewed as plain feature fusion from the independent spatial and temporal branch with an implicit modelling strategy. Li $et$ al.[62] proposed a multi-stage processing method to tackle U-VOS, which frst utilizes a fxed appearance-based network to generate objectness and then feeds this into the motionbased bilateral estimator to segment the objects.
## 2.2 Attention-based VOS
The attention-based VOS task is closely related to U-VOS since it extracts attention-aware object(s) from a video clip. Traditional methods [40,112,125, 142] first compute the single-frame saliency based on various hand-crafted static and motion features and then conduct spatial-temporal optimization to preserve coherency across consecutive frames. Recent works [55, 73, 110] aim to learn a highly semantic representation and usually perform spatial-temporal detection end-toend. Many schemes have been proposed to employ deep networks that consider temporal information, such as ConvLSTM [30,60,93], take optical-flows/adjacentframes as input $[61,110],3\dot{\mathrm{D}}$ convolutional [55, 73], or directly exploit temporally concatenated deep features [56]. Besides, long-term influences are often taken into account and combined with deep learning. Li $etal.[63]$ proposed a key-frame strategy to locate representative high-quality video frames of salient objects [7, 139] and diffused their saliency to illdetected non-key frames. Chen $etal.[15]$ improved
long sparse point trajectories [10,31,75,90,111],object saliency detection by leveraging long-term spatial-
temporal information, where high-quality“beyond-thescope frames" are aligned with the current frames. Both types of information are fed to deep neural networks for classification. Besides considering how to better leverage temporal information, other researchers have attempted to address different problems in video salient object detection (V-SOD), such as reducing the data labelling requirements [127], developing semisupervised approaches [94], or investigating relative saliency [116]. Fan$etal.[30]$ recently introduced a
issue over time, where spatial features are jointly V-SOD model equipped with a saliency shift-aware
ConvLSTM, together with an attention-consistent VSOD dataset with high-quality annotations. Zhao $et$ $al.$ [137] build a large-scale with scribble annotation for weakly supervised video salient object detection. They propose an appearance-motion fusion module to
$\textcircled{1}\underline{\mathrm{TSINGHUA}}$ Springer
|
[
[
0.09716796875,
0.055694580078125,
0.448974609375,
0.055694580078125,
0.448974609375,
0.0693359375,
0.09716796875,
0.0693359375
],
[
0.88671875,
0.05694580078125,
0.90283203125,
0.05694580078125,
0.90283203125,
0.07177734375,
0.88671875,
0.07177734375
],
[
0.1273193359375,
0.09527587890625,
0.48974609375,
0.09527587890625,
0.48974609375,
0.10888671875,
0.1273193359375,
0.10888671875
],
[
0.50927734375,
0.09527587890625,
0.90283203125,
0.09527587890625,
0.90283203125,
0.11016845703125,
0.50927734375,
0.11016845703125
],
[
0.1273193359375,
0.11199951171875,
0.311767578125,
0.11199951171875,
0.311767578125,
0.1256103515625,
0.1273193359375,
0.1256103515625
],
[
0.50927734375,
0.11199951171875,
0.90283203125,
0.11199951171875,
0.90283203125,
0.1256103515625,
0.50927734375,
0.1256103515625
],
[
0.11102294921875,
0.1268310546875,
0.48974609375,
0.1256103515625,
0.48974609375,
0.1405029296875,
0.11102294921875,
0.14111328125
],
[
0.50634765625,
0.1280517578125,
0.90185546875,
0.1268310546875,
0.90185546875,
0.14111328125,
0.50634765625,
0.142333984375
],
[
0.1273193359375,
0.1435546875,
0.485595703125,
0.1435546875,
0.485595703125,
0.1572265625,
0.1273193359375,
0.1572265625
],
[
0.50634765625,
0.1441650390625,
0.90283203125,
0.1441650390625,
0.90283203125,
0.1578369140625,
0.50634765625,
0.1578369140625
],
[
0.125732421875,
0.1578369140625,
0.4873046875,
0.1578369140625,
0.4873046875,
0.172607421875,
0.125732421875,
0.172607421875
],
[
0.5078125,
0.1602783203125,
0.90185546875,
0.1602783203125,
0.90185546875,
0.173828125,
0.5078125,
0.173828125
],
[
0.1248779296875,
0.174560546875,
0.48974609375,
0.17578125,
0.48974609375,
0.1905517578125,
0.1248779296875,
0.1893310546875
],
[
0.50927734375,
0.177001953125,
0.90283203125,
0.177001953125,
0.90283203125,
0.1905517578125,
0.50927734375,
0.1905517578125
],
[
0.1248779296875,
0.191162109375,
0.320068359375,
0.1905517578125,
0.320068359375,
0.204833984375,
0.1248779296875,
0.2060546875
],
[
0.51025390625,
0.1925048828125,
0.90283203125,
0.1925048828125,
0.90283203125,
0.2060546875,
0.51025390625,
0.2060546875
],
[
0.11346435546875,
0.2066650390625,
0.48974609375,
0.2066650390625,
0.48974609375,
0.2203369140625,
0.11346435546875,
0.2203369140625
],
[
0.50927734375,
0.2091064453125,
0.90283203125,
0.2091064453125,
0.90283203125,
0.2227783203125,
0.50927734375,
0.2227783203125
],
[
0.1273193359375,
0.223388671875,
0.489013671875,
0.223388671875,
0.489013671875,
0.237060546875,
0.1273193359375,
0.237060546875
],
[
0.50927734375,
0.224609375,
0.90283203125,
0.224609375,
0.90283203125,
0.23828125,
0.50927734375,
0.23828125
],
[
0.125732421875,
0.23828125,
0.489013671875,
0.239501953125,
0.489013671875,
0.253662109375,
0.125732421875,
0.253173828125
],
[
0.51025390625,
0.2413330078125,
0.90283203125,
0.2413330078125,
0.90283203125,
0.2548828125,
0.51025390625,
0.2548828125
],
[
0.125732421875,
0.253662109375,
0.489013671875,
0.2548828125,
0.489013671875,
0.269287109375,
0.125732421875,
0.2685546875
],
[
0.51025390625,
0.2568359375,
0.90185546875,
0.2568359375,
0.90185546875,
0.2705078125,
0.51025390625,
0.2705078125
],
[
0.1273193359375,
0.271728515625,
0.48974609375,
0.271728515625,
0.48974609375,
0.28515625,
0.1273193359375,
0.28515625
],
[
0.50927734375,
0.27294921875,
0.861328125,
0.27294921875,
0.861328125,
0.287109375,
0.50927734375,
0.287109375
],
[
0.125732421875,
0.285888671875,
0.48974609375,
0.287109375,
0.48974609375,
0.302001953125,
0.125732421875,
0.30078125
],
[
0.51171875,
0.293212890625,
0.74267578125,
0.293212890625,
0.74267578125,
0.30810546875,
0.51171875,
0.30810546875
],
[
0.1273193359375,
0.302001953125,
0.22119140625,
0.303955078125,
0.22119140625,
0.318603515625,
0.1273193359375,
0.316162109375
],
[
0.5263671875,
0.314453125,
0.904296875,
0.314453125,
0.904296875,
0.32861328125,
0.5263671875,
0.32861328125
],
[
0.09716796875,
0.3310546875,
0.12164306640625,
0.3310546875,
0.12164306640625,
0.34228515625,
0.09716796875,
0.34228515625
],
[
0.12078857421875,
0.32861328125,
0.274169921875,
0.32861328125,
0.274169921875,
0.343505859375,
0.12078857421875,
0.343505859375
],
[
0.50927734375,
0.3310546875,
0.90283203125,
0.331787109375,
0.90283203125,
0.346435546875,
0.50927734375,
0.34521484375
],
[
0.51025390625,
0.348388671875,
0.90185546875,
0.348388671875,
0.90185546875,
0.362060546875,
0.51025390625,
0.362060546875
],
[
0.11102294921875,
0.355224609375,
0.48974609375,
0.355224609375,
0.48974609375,
0.3681640625,
0.11102294921875,
0.3681640625
],
[
0.50927734375,
0.364501953125,
0.90283203125,
0.364501953125,
0.90283203125,
0.378173828125,
0.50927734375,
0.378173828125
],
[
0.09307861328125,
0.37060546875,
0.48974609375,
0.369384765625,
0.48974609375,
0.38427734375,
0.09307861328125,
0.385009765625
],
[
0.09466552734375,
0.387451171875,
0.4873046875,
0.387451171875,
0.4873046875,
0.401611328125,
0.09466552734375,
0.401611328125
],
[
0.51025390625,
0.3818359375,
0.90283203125,
0.3818359375,
0.90283203125,
0.392333984375,
0.51025390625,
0.392333984375
],
[
0.50927734375,
0.396728515625,
0.904296875,
0.396728515625,
0.904296875,
0.41015625,
0.50927734375,
0.41015625
],
[
0.09466552734375,
0.404052734375,
0.4873046875,
0.404052734375,
0.4873046875,
0.417724609375,
0.09466552734375,
0.417724609375
],
[
0.50927734375,
0.413330078125,
0.90283203125,
0.413330078125,
0.90283203125,
0.427001953125,
0.50927734375,
0.427001953125
],
[
0.09466552734375,
0.41943359375,
0.485595703125,
0.41943359375,
0.485595703125,
0.43310546875,
0.09466552734375,
0.43310546875
],
[
0.51171875,
0.428955078125,
0.90283203125,
0.428955078125,
0.90283203125,
0.4423828125,
0.51171875,
0.4423828125
],
[
0.0963134765625,
0.43505859375,
0.485595703125,
0.43505859375,
0.485595703125,
0.44873046875,
0.0963134765625,
0.44873046875
],
[
0.51025390625,
0.445556640625,
0.90283203125,
0.445556640625,
0.90283203125,
0.459228515625,
0.51025390625,
0.459228515625
],
[
0.09466552734375,
0.45166015625,
0.489013671875,
0.45166015625,
0.489013671875,
0.46533203125,
0.09466552734375,
0.46533203125
],
[
0.5078125,
0.46044921875,
0.90283203125,
0.4609375,
0.90283203125,
0.475830078125,
0.5078125,
0.474609375
],
[
0.09307861328125,
0.468505859375,
0.489013671875,
0.468505859375,
0.489013671875,
0.48193359375,
0.09307861328125,
0.48193359375
],
[
0.51171875,
0.4765625,
0.904296875,
0.4765625,
0.904296875,
0.489990234375,
0.51171875,
0.489990234375
],
[
0.09466552734375,
0.48388671875,
0.2335205078125,
0.48388671875,
0.2335205078125,
0.49755859375,
0.09466552734375,
0.49755859375
],
[
0.51025390625,
0.4931640625,
0.90185546875,
0.4931640625,
0.90185546875,
0.5068359375,
0.51025390625,
0.5068359375
],
[
0.0963134765625,
0.50634765625,
0.1290283203125,
0.50634765625,
0.1290283203125,
0.5185546875,
0.0963134765625,
0.5185546875
],
[
0.125732421875,
0.5048828125,
0.306884765625,
0.50390625,
0.306884765625,
0.5185546875,
0.125732421875,
0.51904296875
],
[
0.51171875,
0.50927734375,
0.90185546875,
0.50927734375,
0.90185546875,
0.52294921875,
0.51171875,
0.52294921875
],
[
0.1126708984375,
0.52783203125,
0.489013671875,
0.52783203125,
0.489013671875,
0.54150390625,
0.1126708984375,
0.54150390625
],
[
0.51171875,
0.52587890625,
0.90283203125,
0.52587890625,
0.90283203125,
0.5390625,
0.51171875,
0.5390625
],
[
0.09466552734375,
0.54443359375,
0.489013671875,
0.54443359375,
0.489013671875,
0.55810546875,
0.09466552734375,
0.55810546875
],
[
0.50927734375,
0.54150390625,
0.90283203125,
0.54150390625,
0.90283203125,
0.55517578125,
0.50927734375,
0.55517578125
],
[
0.09466552734375,
0.56005859375,
0.489013671875,
0.56005859375,
0.489013671875,
0.57373046875,
0.09466552734375,
0.57373046875
],
[
0.50927734375,
0.55712890625,
0.90185546875,
0.55810546875,
0.90185546875,
0.57177734375,
0.50927734375,
0.5703125
],
[
0.09466552734375,
0.57568359375,
0.48974609375,
0.57568359375,
0.48974609375,
0.5888671875,
0.09466552734375,
0.5888671875
],
[
0.50927734375,
0.572265625,
0.90283203125,
0.57373046875,
0.90283203125,
0.58837890625,
0.50927734375,
0.58740234375
],
[
0.09466552734375,
0.59228515625,
0.485595703125,
0.59228515625,
0.485595703125,
0.60595703125,
0.09466552734375,
0.60595703125
],
[
0.50927734375,
0.5888671875,
0.904296875,
0.5888671875,
0.904296875,
0.6025390625,
0.50927734375,
0.6025390625
],
[
0.09466552734375,
0.6083984375,
0.489013671875,
0.6083984375,
0.489013671875,
0.62109375,
0.09466552734375,
0.62109375
],
[
0.5078125,
0.60498046875,
0.90185546875,
0.60498046875,
0.90185546875,
0.6181640625,
0.5078125,
0.6181640625
],
[
0.09466552734375,
0.625,
0.48974609375,
0.625,
0.48974609375,
0.63818359375,
0.09466552734375,
0.63818359375
],
[
0.50927734375,
0.62109375,
0.904296875,
0.62109375,
0.904296875,
0.634765625,
0.50927734375,
0.634765625
],
[
0.09307861328125,
0.640625,
0.489013671875,
0.63916015625,
0.489013671875,
0.654296875,
0.09307861328125,
0.65478515625
],
[
0.50634765625,
0.63720703125,
0.90185546875,
0.63720703125,
0.90185546875,
0.65087890625,
0.50634765625,
0.65087890625
],
[
0.09466552734375,
0.6572265625,
0.485595703125,
0.6572265625,
0.485595703125,
0.6708984375,
0.09466552734375,
0.6708984375
],
[
0.50927734375,
0.654296875,
0.904296875,
0.654296875,
0.904296875,
0.66748046875,
0.50927734375,
0.66748046875
],
[
0.09466552734375,
0.6728515625,
0.489013671875,
0.6728515625,
0.489013671875,
0.68603515625,
0.09466552734375,
0.68603515625
],
[
0.50927734375,
0.6708984375,
0.90185546875,
0.6708984375,
0.90185546875,
0.68359375,
0.50927734375,
0.68359375
],
[
0.09466552734375,
0.68798828125,
0.48974609375,
0.68798828125,
0.48974609375,
0.70166015625,
0.09466552734375,
0.70166015625
],
[
0.50927734375,
0.68603515625,
0.90283203125,
0.68603515625,
0.90283203125,
0.69970703125,
0.50927734375,
0.69970703125
],
[
0.09307861328125,
0.70458984375,
0.48974609375,
0.70458984375,
0.48974609375,
0.71826171875,
0.09307861328125,
0.71826171875
],
[
0.50927734375,
0.703125,
0.90283203125,
0.703125,
0.90283203125,
0.716796875,
0.50927734375,
0.716796875
],
[
0.09466552734375,
0.72021484375,
0.489013671875,
0.72021484375,
0.489013671875,
0.73388671875,
0.09466552734375,
0.73388671875
],
[
0.50927734375,
0.71826171875,
0.904296875,
0.71826171875,
0.904296875,
0.73193359375,
0.50927734375,
0.73193359375
],
[
0.09466552734375,
0.73681640625,
0.489013671875,
0.73681640625,
0.489013671875,
0.75048828125,
0.09466552734375,
0.75048828125
],
[
0.51171875,
0.736328125,
0.90185546875,
0.736328125,
0.90185546875,
0.74609375,
0.51171875,
0.74609375
],
[
0.09466552734375,
0.7529296875,
0.48974609375,
0.7529296875,
0.48974609375,
0.76611328125,
0.09466552734375,
0.76611328125
],
[
0.5078125,
0.75048828125,
0.90283203125,
0.75048828125,
0.90283203125,
0.76416015625,
0.5078125,
0.76416015625
],
[
0.09466552734375,
0.76904296875,
0.48486328125,
0.76904296875,
0.48486328125,
0.78271484375,
0.09466552734375,
0.78271484375
],
[
0.51171875,
0.76611328125,
0.90283203125,
0.76611328125,
0.90283203125,
0.77978515625,
0.51171875,
0.77978515625
],
[
0.0963134765625,
0.7861328125,
0.4873046875,
0.7861328125,
0.4873046875,
0.79638671875,
0.0963134765625,
0.79638671875
],
[
0.50927734375,
0.78271484375,
0.90283203125,
0.78271484375,
0.90283203125,
0.79638671875,
0.50927734375,
0.79638671875
],
[
0.09466552734375,
0.80078125,
0.489013671875,
0.80078125,
0.489013671875,
0.814453125,
0.09466552734375,
0.814453125
],
[
0.50927734375,
0.798828125,
0.90283203125,
0.798828125,
0.90283203125,
0.8125,
0.50927734375,
0.8125
],
[
0.09466552734375,
0.8173828125,
0.489013671875,
0.8173828125,
0.489013671875,
0.8310546875,
0.09466552734375,
0.8310546875
],
[
0.50927734375,
0.81298828125,
0.904296875,
0.814453125,
0.904296875,
0.82861328125,
0.50927734375,
0.828125
],
[
0.09466552734375,
0.8330078125,
0.4072265625,
0.8330078125,
0.4072265625,
0.8466796875,
0.09466552734375,
0.8466796875
],
[
0.51025390625,
0.8310546875,
0.90283203125,
0.8310546875,
0.90283203125,
0.8447265625,
0.51025390625,
0.8447265625
],
[
0.109375,
0.84765625,
0.48974609375,
0.8486328125,
0.48974609375,
0.86328125,
0.109375,
0.86181640625
],
[
0.51025390625,
0.84521484375,
0.90625,
0.8466796875,
0.90625,
0.861328125,
0.51025390625,
0.8603515625
],
[
0.09307861328125,
0.8642578125,
0.491455078125,
0.86328125,
0.491455078125,
0.8779296875,
0.09307861328125,
0.87890625
],
[
0.51025390625,
0.86328125,
0.904296875,
0.86328125,
0.904296875,
0.8779296875,
0.51025390625,
0.8779296875
],
[
0.09466552734375,
0.88134765625,
0.48974609375,
0.88134765625,
0.48974609375,
0.89404296875,
0.09466552734375,
0.89404296875
],
[
0.50927734375,
0.8798828125,
0.90283203125,
0.8798828125,
0.90283203125,
0.8935546875,
0.50927734375,
0.8935546875
],
[
0.09466552734375,
0.89794921875,
0.48974609375,
0.89794921875,
0.48974609375,
0.9111328125,
0.09466552734375,
0.9111328125
],
[
0.50927734375,
0.89404296875,
0.904296875,
0.8955078125,
0.904296875,
0.91015625,
0.50927734375,
0.9091796875
],
[
0.6875,
0.93017578125,
0.90869140625,
0.93017578125,
0.90869140625,
0.95166015625,
0.6875,
0.95166015625
]
] |
[
"Full-Duplex Strategy for Video Object Segmentation",
"3",
"the validity and rationality of our current design",
"method cannot accurately segment objects between",
"under various conditions.",
"two similar consecutive frames since it relies heavily",
".Weprovide moredetails to the conference",
"on the guidance of optical flow.To resolve this,",
"version, such as backbone details (see Sec. 3.6.1),",
"several works [21, 92, 97] have integrated the spatial",
"evaluation metrics (see Sec. 4.2), prediction",
"and temporal features from the parallel network,",
"selection (see Sec. 4.5.1), and post-processing",
"which can be viewed as plain feature fusion from",
"techniques (see Sec. 4.5.2).",
"the independent spatial and temporal branch with an",
". We further present more performances under",
"implicit modelling strategy.Li et al.[62] proposed",
"different thresholds used in the final results (i.e.,",
"a multi-stage processing method to tackle U-VOs,",
"PR curve in Fig. 6). Additional test results (i.e.,",
"which first utilizes a fixed appearance-based network to",
"DAVSOD1g-Normal25,DAVSOD1g-Difficult20)on",
"generate objectness and then feeds this into the motion-",
"a recent challenging dataset still show that our",
"based bilateral estimator to segment the objects.",
"framework is superior to existing SOTA models",
"2.2Attention-based VOs",
"(see Tab. 3).",
"The attention-based VOS task is closely related",
"2.",
"RelatedWorks",
"to U-vOs since it extracts attention-aware object(s)",
"from a video clip. Traditional methods [40, 112, 125,",
"Depending on whether or not the first frame",
"142] first compute the single-frame saliency based on",
"of ground truth is given,thevos task canbe",
"divided into two scenarios, i.e., semi-supervised [109]",
"various hand-crafted static and motion features and",
"then conduct spatial-temporal optimization to preserve",
"and unsupervised/zero-shot [7]. Some typical semi-",
"coherency across consecutive frames. Recent works [55,",
"supervised VOS models can be referred to [6, 11.",
"73, 110l aim to learn a highly semantic representation",
"12, 20,25, 38, 47,53, 64, 81,85,88, 102,121,126.",
"and usually perform spatial-temporal detection end-to-",
"128, 133, 134]. This paper studies the unsupervised",
"end.Many schemes have been proposed to employ deep",
"setting [140, 141], leaving the semi-supervised setting",
"networks that consider temporal information, such as",
"as our future work.",
"ConvLSTM [30, 60, 93], take optical-flows/adjacent-",
"2.1",
"Unsupervisedvos",
"frames as input [61, 110], 3D convolutional [55,",
"Although there are many works addressing the VOS",
"73l, or directly exploit temporally concatenated deep",
"task in a semi-supervised manner, i.e., by supposing",
"features [56]l. Besides, long-term influences are often",
"an object mask annotation is given in the first",
"taken into account and combined with deep learning.",
"frame, other researchers have attempted to address the",
"Li et al. [63] proposed a key-frame strategy to locate",
"more challenging unsupervised VOS (U-VOS) problem.",
"representative high-quality video frames of salient",
"Early U-VOS models resort to low-level handcrafted",
"objects [7, 139] and diffused their saliency to ill-",
"features for heuristic segmentation inference, such as",
"detected non-key frames. Chen et al. [15] improved",
"long sparse point trajectories[10, 31,75,90,111],object",
"saliency detection by leveraging long-term spatial-",
"proposals [58, 59, 69, 83], saliency priors [27, 106, 108]",
"temporal information, where high-quality\"beyond-the-",
"optical fow [100], or superpixels [32, 33, 123]. These",
"scope frames\" are aligned with the current frames.",
"traditional models have limited generalizability and",
"Both types of information are fed to deep neural",
"thus low accuracy in highly dynamic and complex",
"networks for classification. Besides considering how to",
"scenarios due to their lack of semantic information",
"better leverage temporal information, other researchers",
"and high-level content understanding. Recently, RNN-",
"have attemptedto address different problems in video",
"based models [4, 93, 97,113, 126, 138] have become",
"salient object detection (V-SOD), such as reducing",
"popular due to their better capability of capturing long",
"the data labelling requirements [127l, developing semi-",
"term dependencies and their use of deep learning.In",
"supervised approaches [94], or investigating relative",
"this case, U-VOS is formulated as a recurrent modelling",
"saliency [116]. Fan et al. [30] recently introduced a",
"issue over time, where spatial features are jointly",
"V-SOD model equipped with a saliency shift-aware",
"exploited with long-term temporal context.",
"ConvLSTM, together with an attention-consistent V-",
"Howtocombine motion cueswithappearance",
"SOD dataset with high-quality annotations. Zhao et",
"features is a long-standing problem in this field. To this",
"al. [137] build a large-scale with scribble annotation",
"end, Tokmakov et al.J96l proposed to use themotion",
"for weakly supervised video salient object detection.",
"patterns required from the video simply. However, their",
"They propose an appearance-motion fusion module to",
"TSINGHREA Springer"
] |
{
"arxiv_id": "2108.03151",
"language": "en",
"page_idx": 2
}
|
||
2108.03151-en-3
|
2108.03151-en-3.jpg
|
G.-P. Ji, D.-P. Fan, K. Fu, and Z. Wu, $et\textit{al. }$
<FigureHere>
Fig. 3 The architecture of our FSNet for video object segmentation. The Relational Cross-Attention Module (RCAM) abstracts Purification Modules (BPM) are stacked to further re-calibrate inconsistencies between the motion and appearance features. Finally we utilize the decoder to generate our prediction.
3.2 Relational Cross-Attention Module
aggregate the spatial-temporal features attentively.
### 3 Methodology
## 3.1 Overview
As discussed in Sec. 1, a single-modality $(i.e.$,motion or appearance) guided stimulation may cause the model to make incorrect decisions. To alleviate this, we design a cross-attention module (RCAM) via the channel-wise attention mechanism, which focuses on distilling out effectively squeezed cues from two modalities and then modulating each other. As shown in Fig. 4 (c), the
and motion features $\{\mathcal{Y}_k\}_{k=1}^K$, which are obtained from the two different branches of the standard ResNet- 50[36]. Specifically, for each $k$-level, we frst perform global average pooling (GAP) to generate channelwise vectors $\mathcal{V}_k^X$ and $\mathcal{V}_k^Y$ from each $\mathcal{X}_k$ and $\mathcal{Y}_k.$ Next, two 1\times1 conv, $i.e.,\phi(x;\mathbf{W}_{\phi})$ and $\theta(x;\mathbf{W}_{\theta})$, with learnable parameters $\mathbf{W}_\phi$ and $\mathbf{W}_\theta$, generate two discriminated global descriptors. The sigmoid function $\sigma[x]~=~e^x/(e^x~+~1),~x~\in~\mathbb{R}~$is then applied to convert the final descriptors into the interval $[0,1],i.e.$, into the valid attention vector for channel weighting. Then, we perform outer product $\otimes$between $\mathcal{X}_k$ and $\sigma\left[\theta(\mathcal{V}_k^Y;\mathbf{W}_\theta)\right]$to generate a candidate feature $\mathcal Q_k^X$,and vice versa, as follows:
(1)
$$\mathcal{Q}_{k}^{X}=\mathcal{X}_{k}\otimes\sigma\left[\theta(\mathcal{V}_{k}^{Y};\mathbf{W}_{\theta})\right],\\\mathcal{Q}_{k}^{Y}=\mathcal{Y}_{k}\otimes\sigma\left[\phi(\mathcal{V}_{k}^{X};\mathbf{W}_{\phi})\right].$$
Suppose that a video clip contains $T$ consecutive frames $\{ \mathbf{A} ^{t}\} _{t= 1}^{T}.$ We frst utilize optical flow field generator $\mathcal{H},i.e.$, FlowNet 2.0 [42], to generate $T-1$ optical flow maps $\{\mathbf{M}^{t}\}_{t=1}^{T-1}$, which are all computed ensure the inputs match, we discard the last frame in the pipeline. Thus, the proposed pipeline takes both the appearance image $\{\mathbf{A}^{t}\}_{t=1}^{T-1}$ and its paired ${\mathrm{motion~map}}\{\mathbf{M}^{t}\}_{t=1}^{T-1}{\mathrm{~as~the~input.~First,~M}^{t}\&\mathbf{A}^{t}}$ pairs at frame $t^2$ are fed to two independent ResNet- 50 [36] branches $(i.e.$,motion and appearance blocks in Fig. 3). The appearance features$\{\mathcal{X}_k\}_k=1^K$and motion features $\{\mathcal{Y}_k\}_{k=1}^K$ extracted from $K$ layers are then sent to the Relational Cross-Attention Modules (RCAMs),which allows the network to embed spatialtemporal cross-modal features. Next, we employ the Bidirectional Purification Modules (BPMs) with $N$ cascaded units. BPMs focus on distilling representative carriers from fused features $\{\mathbf{F}_k^n\}_{n=1}^N$ and motion-based features $\{ \mathbf{G} _k^n\} _{n= 1}^N.$ Finally, the predictions $(i.e.,\mathbf{S}_M^t$ and $\mathbf{S}_A^t)$ at frame $t$ are generated from two decoder blocks
(2)
$^{2}$Here, we omit the superscript“$t^\text{"for the convenient expression.}$
Then, we combine $Q_k^X,Q_k^Y$, and lower-level fused feature $\mathcal{Z}_k-1$ for in-depth feature extraction. With an
$\textcircled{1}\underline{\mathrm{TSINGHUA}}$Springer
|
[
[
0.09716796875,
0.061248779296875,
0.10693359375,
0.061248779296875,
0.10693359375,
0.06866455078125,
0.09716796875,
0.06866455078125
],
[
0.60400390625,
0.05694580078125,
0.89892578125,
0.05694580078125,
0.89892578125,
0.070556640625,
0.60400390625,
0.070556640625
],
[
0.61083984375,
0.09466552734375,
0.6962890625,
0.09466552734375,
0.6962890625,
0.1058349609375,
0.61083984375,
0.1058349609375
],
[
0.806640625,
0.09466552734375,
0.89794921875,
0.09466552734375,
0.89794921875,
0.1058349609375,
0.806640625,
0.1058349609375
],
[
0.40576171875,
0.09527587890625,
0.4970703125,
0.09527587890625,
0.4970703125,
0.1058349609375,
0.40576171875,
0.1058349609375
],
[
0.2073974609375,
0.09527587890625,
0.29052734375,
0.09527587890625,
0.29052734375,
0.1058349609375,
0.2073974609375,
0.1058349609375
],
[
0.2457275390625,
0.1070556640625,
0.272705078125,
0.1070556640625,
0.272705078125,
0.1181640625,
0.2457275390625,
0.1181640625
],
[
0.44482421875,
0.1070556640625,
0.47265625,
0.1070556640625,
0.47265625,
0.1181640625,
0.44482421875,
0.1181640625
],
[
0.64306640625,
0.1070556640625,
0.6708984375,
0.1070556640625,
0.6708984375,
0.1181640625,
0.64306640625,
0.1181640625
],
[
0.841796875,
0.1070556640625,
0.8701171875,
0.1070556640625,
0.8701171875,
0.11944580078125,
0.841796875,
0.11944580078125
],
[
0.10040283203125,
0.133056640625,
0.147705078125,
0.128662109375,
0.1494140625,
0.1392822265625,
0.10284423828125,
0.1441650390625
],
[
0.30126953125,
0.133056640625,
0.34375,
0.1312255859375,
0.34375,
0.14111328125,
0.30126953125,
0.1435546875
],
[
0.72412109375,
0.128662109375,
0.7919921875,
0.1318359375,
0.7919921875,
0.142333984375,
0.72412109375,
0.1392822265625
],
[
0.52001953125,
0.1348876953125,
0.5419921875,
0.1348876953125,
0.5419921875,
0.1392822265625,
0.52001953125,
0.1392822265625
],
[
0.1461181640625,
0.160888671875,
0.170654296875,
0.160888671875,
0.170654296875,
0.172607421875,
0.1461181640625,
0.172607421875
],
[
0.24658203125,
0.162109375,
0.272705078125,
0.162109375,
0.272705078125,
0.1707763671875,
0.24658203125,
0.1707763671875
],
[
0.64892578125,
0.162109375,
0.68017578125,
0.162109375,
0.68017578125,
0.1695556640625,
0.64892578125,
0.1695556640625
],
[
0.7509765625,
0.163330078125,
0.77734375,
0.163330078125,
0.77734375,
0.16650390625,
0.7509765625,
0.16650390625
],
[
0.8408203125,
0.1590576171875,
0.89794921875,
0.1558837890625,
0.89892578125,
0.1707763671875,
0.841796875,
0.173828125
],
[
0.282470703125,
0.1776123046875,
0.3486328125,
0.1832275390625,
0.34619140625,
0.2017822265625,
0.279296875,
0.19677734375
],
[
0.32568359375,
0.1776123046875,
0.384521484375,
0.1776123046875,
0.384521484375,
0.1973876953125,
0.32568359375,
0.1973876953125
],
[
0.37060546875,
0.172607421875,
0.5810546875,
0.172607421875,
0.5810546875,
0.204833984375,
0.37060546875,
0.204833984375
],
[
0.63671875,
0.1788330078125,
0.6630859375,
0.1788330078125,
0.6630859375,
0.1905517578125,
0.63671875,
0.1905517578125
],
[
0.72509765625,
0.1776123046875,
0.76318359375,
0.1776123046875,
0.76318359375,
0.1893310546875,
0.72509765625,
0.1893310546875
],
[
0.6572265625,
0.1893310546875,
0.69775390625,
0.173828125,
0.70849609375,
0.1905517578125,
0.66796875,
0.2060546875
],
[
0.2481689453125,
0.1806640625,
0.27099609375,
0.1806640625,
0.27099609375,
0.1881103515625,
0.2481689453125,
0.1881103515625
],
[
0.26611328125,
0.1925048828125,
0.282470703125,
0.1925048828125,
0.282470703125,
0.195556640625,
0.26611328125,
0.195556640625
],
[
0.70849609375,
0.1893310546875,
0.72802734375,
0.1893310546875,
0.72802734375,
0.19677734375,
0.70849609375,
0.19677734375
],
[
0.252197265625,
0.2078857421875,
0.262939453125,
0.21044921875,
0.261962890625,
0.21533203125,
0.249755859375,
0.2122802734375
],
[
0.48974609375,
0.2060546875,
0.5263671875,
0.2078857421875,
0.5263671875,
0.21533203125,
0.48974609375,
0.2135009765625
],
[
0.75341796875,
0.2078857421875,
0.806640625,
0.200439453125,
0.810546875,
0.21533203125,
0.75732421875,
0.2227783203125
],
[
0.7412109375,
0.2135009765625,
0.765625,
0.2135009765625,
0.765625,
0.2171630859375,
0.7412109375,
0.2171630859375
],
[
0.36083984375,
0.21044921875,
0.386962890625,
0.21044921875,
0.386962890625,
0.2171630859375,
0.36083984375,
0.2171630859375
],
[
0.287353515625,
0.223388671875,
0.311767578125,
0.223388671875,
0.311767578125,
0.2320556640625,
0.287353515625,
0.2320556640625
],
[
0.48486328125,
0.223388671875,
0.51025390625,
0.223388671875,
0.51025390625,
0.2320556640625,
0.48486328125,
0.2320556640625
],
[
0.6875,
0.223388671875,
0.7099609375,
0.223388671875,
0.7099609375,
0.2320556640625,
0.6875,
0.2320556640625
],
[
0.28662109375,
0.237060546875,
0.315185546875,
0.237060546875,
0.315185546875,
0.251953125,
0.28662109375,
0.251953125
],
[
0.48486328125,
0.237060546875,
0.51025390625,
0.237060546875,
0.51025390625,
0.25,
0.48486328125,
0.25
],
[
0.685546875,
0.237060546875,
0.71435546875,
0.237060546875,
0.71435546875,
0.251953125,
0.685546875,
0.251953125
],
[
0.5732421875,
0.246337890625,
0.60986328125,
0.246337890625,
0.60986328125,
0.2568359375,
0.5732421875,
0.2568359375
],
[
0.287353515625,
0.253662109375,
0.3134765625,
0.253662109375,
0.3134765625,
0.26611328125,
0.287353515625,
0.26611328125
],
[
0.485595703125,
0.253173828125,
0.51025390625,
0.253173828125,
0.51025390625,
0.265380859375,
0.485595703125,
0.265380859375
],
[
0.6875,
0.253662109375,
0.7119140625,
0.253662109375,
0.7119140625,
0.267333984375,
0.6875,
0.267333984375
],
[
0.287353515625,
0.269287109375,
0.3134765625,
0.269287109375,
0.3134765625,
0.281005859375,
0.287353515625,
0.281005859375
],
[
0.485595703125,
0.269287109375,
0.51025390625,
0.269287109375,
0.51025390625,
0.281005859375,
0.485595703125,
0.281005859375
],
[
0.685546875,
0.269287109375,
0.712890625,
0.269287109375,
0.712890625,
0.2822265625,
0.685546875,
0.2822265625
],
[
0.22607421875,
0.298828125,
0.272705078125,
0.30078125,
0.27099609375,
0.315673828125,
0.225341796875,
0.313232421875
],
[
0.32568359375,
0.298828125,
0.373046875,
0.30078125,
0.372314453125,
0.315673828125,
0.323974609375,
0.313232421875
],
[
0.52392578125,
0.299560546875,
0.5732421875,
0.299560546875,
0.5732421875,
0.315673828125,
0.52392578125,
0.315673828125
],
[
0.7265625,
0.299560546875,
0.77001953125,
0.299560546875,
0.77001953125,
0.315673828125,
0.7265625,
0.315673828125
],
[
0.62451171875,
0.30078125,
0.6708984375,
0.30078125,
0.6708984375,
0.315673828125,
0.62451171875,
0.315673828125
],
[
0.424560546875,
0.30078125,
0.47265625,
0.30078125,
0.47265625,
0.315673828125,
0.424560546875,
0.315673828125
],
[
0.54443359375,
0.32177734375,
0.55078125,
0.32177734375,
0.55078125,
0.325439453125,
0.54443359375,
0.325439453125
],
[
0.8662109375,
0.35888671875,
0.88232421875,
0.35888671875,
0.88232421875,
0.36767578125,
0.8662109375,
0.36767578125
],
[
0.0963134765625,
0.385009765625,
0.147705078125,
0.385009765625,
0.147705078125,
0.396728515625,
0.0963134765625,
0.396728515625
],
[
0.1412353515625,
0.38427734375,
0.90283203125,
0.38427734375,
0.90283203125,
0.39794921875,
0.1412353515625,
0.39794921875
],
[
0.0963134765625,
0.39794921875,
0.90185546875,
0.39794921875,
0.90185546875,
0.40771484375,
0.0963134765625,
0.40771484375
],
[
0.09466552734375,
0.40771484375,
0.90283203125,
0.40771484375,
0.90283203125,
0.42138671875,
0.09466552734375,
0.42138671875
],
[
0.0963134765625,
0.42138671875,
0.3974609375,
0.42138671875,
0.3974609375,
0.431884765625,
0.0963134765625,
0.431884765625
],
[
0.09466552734375,
0.46044921875,
0.461181640625,
0.4580078125,
0.461181640625,
0.47265625,
0.09466552734375,
0.474609375
],
[
0.51025390625,
0.459228515625,
0.86181640625,
0.459228515625,
0.86181640625,
0.47265625,
0.51025390625,
0.47265625
],
[
0.1248779296875,
0.4833984375,
0.25634765625,
0.486328125,
0.25634765625,
0.501953125,
0.1248779296875,
0.498779296875
],
[
0.52587890625,
0.480224609375,
0.90283203125,
0.480712890625,
0.90283203125,
0.49560546875,
0.52587890625,
0.494384765625
],
[
0.09466552734375,
0.486328125,
0.1273193359375,
0.486328125,
0.1273193359375,
0.49755859375,
0.09466552734375,
0.49755859375
],
[
0.1248779296875,
0.489501953125,
0.1297607421875,
0.489501953125,
0.1297607421875,
0.49560546875,
0.1248779296875,
0.49560546875
],
[
0.5078125,
0.49755859375,
0.90283203125,
0.496337890625,
0.90283203125,
0.51123046875,
0.5078125,
0.51220703125
],
[
0.0963134765625,
0.51220703125,
0.1436767578125,
0.51220703125,
0.1436767578125,
0.5234375,
0.0963134765625,
0.5234375
],
[
0.1370849609375,
0.51123046875,
0.2236328125,
0.51123046875,
0.2236328125,
0.52587890625,
0.1370849609375,
0.52587890625
],
[
0.5078125,
0.51416015625,
0.90185546875,
0.51416015625,
0.90185546875,
0.52783203125,
0.5078125,
0.52783203125
],
[
0.109375,
0.53271484375,
0.48974609375,
0.5322265625,
0.48974609375,
0.54638671875,
0.109375,
0.5478515625
],
[
0.50927734375,
0.52978515625,
0.90283203125,
0.52978515625,
0.90283203125,
0.54345703125,
0.50927734375,
0.54345703125
],
[
0.09222412109375,
0.54638671875,
0.48974609375,
0.5478515625,
0.48974609375,
0.5654296875,
0.09222412109375,
0.564453125
],
[
0.50927734375,
0.5458984375,
0.90283203125,
0.5458984375,
0.90283203125,
0.55859375,
0.50927734375,
0.55859375
],
[
0.09466552734375,
0.5654296875,
0.489013671875,
0.5654296875,
0.489013671875,
0.57958984375,
0.09466552734375,
0.57958984375
],
[
0.50927734375,
0.56103515625,
0.90283203125,
0.56103515625,
0.90283203125,
0.57568359375,
0.50927734375,
0.57568359375
],
[
0.09307861328125,
0.5810546875,
0.491455078125,
0.5810546875,
0.491455078125,
0.5986328125,
0.09307861328125,
0.5986328125
],
[
0.50927734375,
0.57666015625,
0.90283203125,
0.57666015625,
0.90283203125,
0.591796875,
0.50927734375,
0.591796875
],
[
0.50634765625,
0.59033203125,
0.90283203125,
0.59228515625,
0.90283203125,
0.6103515625,
0.50634765625,
0.6083984375
],
[
0.09466552734375,
0.59765625,
0.48974609375,
0.59765625,
0.48974609375,
0.611328125,
0.09466552734375,
0.611328125
],
[
0.09222412109375,
0.61328125,
0.48974609375,
0.61181640625,
0.48974609375,
0.626953125,
0.09222412109375,
0.6279296875
],
[
0.50634765625,
0.6083984375,
0.904296875,
0.60693359375,
0.904296875,
0.625,
0.50634765625,
0.62548828125
],
[
0.225341796875,
0.6298828125,
0.48974609375,
0.6298828125,
0.48974609375,
0.64404296875,
0.225341796875,
0.64404296875
],
[
0.50927734375,
0.62548828125,
0.90283203125,
0.62548828125,
0.90283203125,
0.63916015625,
0.50927734375,
0.63916015625
],
[
0.2196044921875,
0.634765625,
0.227783203125,
0.634765625,
0.227783203125,
0.63818359375,
0.2196044921875,
0.63818359375
],
[
0.50927734375,
0.64111328125,
0.90283203125,
0.64111328125,
0.90283203125,
0.65478515625,
0.50927734375,
0.65478515625
],
[
0.09307861328125,
0.63134765625,
0.22119140625,
0.63134765625,
0.22119140625,
0.64404296875,
0.09307861328125,
0.64404296875
],
[
0.09466552734375,
0.646484375,
0.48974609375,
0.646484375,
0.48974609375,
0.66015625,
0.09466552734375,
0.66015625
],
[
0.09307861328125,
0.66015625,
0.252197265625,
0.66015625,
0.252197265625,
0.677734375,
0.09307861328125,
0.677734375
],
[
0.250732421875,
0.662109375,
0.4873046875,
0.662109375,
0.4873046875,
0.67578125,
0.250732421875,
0.67578125
],
[
0.51025390625,
0.6572265625,
0.90283203125,
0.6572265625,
0.90283203125,
0.67138671875,
0.51025390625,
0.67138671875
],
[
0.225341796875,
0.6640625,
0.2342529296875,
0.6640625,
0.2342529296875,
0.66845703125,
0.225341796875,
0.66845703125
],
[
0.50634765625,
0.6708984375,
0.904296875,
0.67138671875,
0.904296875,
0.689453125,
0.50634765625,
0.68798828125
],
[
0.09466552734375,
0.6767578125,
0.483154296875,
0.6767578125,
0.483154296875,
0.69140625,
0.09466552734375,
0.69140625
],
[
0.09466552734375,
0.6943359375,
0.48974609375,
0.6943359375,
0.48974609375,
0.708984375,
0.09466552734375,
0.708984375
],
[
0.50927734375,
0.6904296875,
0.90185546875,
0.6904296875,
0.90185546875,
0.70361328125,
0.50927734375,
0.70361328125
],
[
0.09063720703125,
0.70654296875,
0.48974609375,
0.708984375,
0.48974609375,
0.7265625,
0.09063720703125,
0.724609375
],
[
0.50927734375,
0.70458984375,
0.90283203125,
0.7060546875,
0.90283203125,
0.72021484375,
0.50927734375,
0.7197265625
],
[
0.09466552734375,
0.7265625,
0.489013671875,
0.7265625,
0.489013671875,
0.7412109375,
0.09466552734375,
0.7412109375
],
[
0.51171875,
0.72265625,
0.90283203125,
0.72265625,
0.90283203125,
0.736328125,
0.51171875,
0.736328125
],
[
0.09466552734375,
0.74267578125,
0.48974609375,
0.74267578125,
0.48974609375,
0.75634765625,
0.09466552734375,
0.75634765625
],
[
0.50927734375,
0.73681640625,
0.90283203125,
0.73681640625,
0.90283203125,
0.751953125,
0.50927734375,
0.751953125
],
[
0.5078125,
0.75390625,
0.90185546875,
0.75390625,
0.90185546875,
0.76708984375,
0.5078125,
0.76708984375
],
[
0.09716796875,
0.759765625,
0.48486328125,
0.759765625,
0.48486328125,
0.7705078125,
0.09716796875,
0.7705078125
],
[
0.09466552734375,
0.77490234375,
0.48974609375,
0.77490234375,
0.48974609375,
0.78857421875,
0.09466552734375,
0.78857421875
],
[
0.5078125,
0.76904296875,
0.90185546875,
0.7705078125,
0.90185546875,
0.78515625,
0.5078125,
0.7841796875
],
[
0.0963134765625,
0.7900390625,
0.4873046875,
0.7900390625,
0.4873046875,
0.8037109375,
0.0963134765625,
0.8037109375
],
[
0.51025390625,
0.7861328125,
0.90283203125,
0.7861328125,
0.90283203125,
0.80078125,
0.51025390625,
0.80078125
],
[
0.09466552734375,
0.80712890625,
0.48974609375,
0.80712890625,
0.48974609375,
0.82177734375,
0.09466552734375,
0.82177734375
],
[
0.51025390625,
0.802734375,
0.90283203125,
0.802734375,
0.90283203125,
0.8173828125,
0.51025390625,
0.8173828125
],
[
0.09222412109375,
0.8203125,
0.48974609375,
0.82177734375,
0.48974609375,
0.8388671875,
0.09222412109375,
0.83837890625
],
[
0.5078125,
0.81884765625,
0.6669921875,
0.8173828125,
0.6669921875,
0.83154296875,
0.5078125,
0.8330078125
],
[
0.09063720703125,
0.833984375,
0.48974609375,
0.833984375,
0.48974609375,
0.8583984375,
0.09063720703125,
0.8583984375
],
[
0.6064453125,
0.83544921875,
0.80908203125,
0.83544921875,
0.80908203125,
0.8525390625,
0.6064453125,
0.8525390625
],
[
0.87841796875,
0.83544921875,
0.904296875,
0.83544921875,
0.904296875,
0.85400390625,
0.87841796875,
0.85400390625
],
[
0.09466552734375,
0.85498046875,
0.485595703125,
0.85498046875,
0.485595703125,
0.86962890625,
0.09466552734375,
0.86962890625
],
[
0.6064453125,
0.85498046875,
0.81201171875,
0.85498046875,
0.81201171875,
0.87255859375,
0.6064453125,
0.87255859375
],
[
0.8798828125,
0.85693359375,
0.90283203125,
0.85693359375,
0.90283203125,
0.87109375,
0.8798828125,
0.87109375
],
[
0.09466552734375,
0.87255859375,
0.14208984375,
0.87255859375,
0.14208984375,
0.88427734375,
0.09466552734375,
0.88427734375
],
[
0.5263671875,
0.8798828125,
0.90283203125,
0.8798828125,
0.90283203125,
0.89404296875,
0.5263671875,
0.89404296875
],
[
0.1175537109375,
0.88623046875,
0.485595703125,
0.88720703125,
0.485595703125,
0.90087890625,
0.1175537109375,
0.89990234375
],
[
0.51025390625,
0.896484375,
0.904296875,
0.896484375,
0.904296875,
0.9111328125,
0.51025390625,
0.9111328125
],
[
0.10528564453125,
0.93017578125,
0.327392578125,
0.93017578125,
0.327392578125,
0.95166015625,
0.10528564453125,
0.95166015625
],
[
0.1575927734375,
0.9443359375,
0.2220458984375,
0.9443359375,
0.2220458984375,
0.94873046875,
0.1575927734375,
0.94873046875
]
] |
[
"4",
"G.-P. Ji, D.-P. Fan, K. Fu, and Z. Wu, et al.",
"FlowNet2.0k",
"FlowNet2.0k",
"FlowNet2.0k",
"FlowNet2.0",
"H",
"VH",
"VH",
"WH",
"",
"",
"",
"",
"+A",
"IM",
"MI",
"",
"M°+3",
"",
"",
"",
"V",
"Fxit",
"",
"",
"",
"",
"",
"Mlocato",
"(Fn1+2",
"",
"EA",
"BPM",
"BPM",
"BPM",
"BPM",
"BPM",
"BPM",
"N=4",
"BPM",
"BPM",
"BPM",
"BPM",
"BPM",
"BPM",
"Decoder",
"Decoder",
"Decoder",
"Decoder",
"Decoder",
"Deeoder",
"",
"Sf",
"Fig.3",
"The architecture of our FSNet for video object segmentation.The Relational Cross-Attention Module (RCAM) abstracts",
"more discriminative representations between the motion and appearance cues using the full-duplex strategy.Then four Bidirectional",
"Purification Modules (BPM) are stacked to further re-calibrate inconsistencies between the motion and appearance features. Finally,",
"we utilize the decoder to generate our prediction.",
"aggregate the spatial-temporal features attentively.",
"3.2 Relational Cross-Attention Module",
"Methodology",
"As discussed in Sec.1, a single-modality (i.e., motion",
"3",
"",
" or appearance) guided stimulation may cause the model",
"3.1",
"Overview",
"to make incorrect decisions. To alleviate this, we design",
"Suppose that a video clip contains T consecutive",
"a cross-attention module (RCAM) via the channel-wise",
"frames {A\"}T-1.We first utilize optical flow field",
"attention mechanism, which focuses on distilling out",
"generator H, i.e., FlowNet 2.0 [42], to generate T -1",
"effectively squeezed cues from two modalities and then",
"optical fow maps [Mt}t-1, which are all computed",
"modulating each other. As shown in Fig. 4 (c), the",
"two inputs of RCAM are appearance features [X]K-1",
"by two adjacent frames (Mt = H[At, At+1). To",
"ensure the inputs match, we discard the last frame",
"and motion features []K-1, which are obtained from",
"Thus, the proposed pipeline takes",
"the two different branches of the standard ResNet-",
"",
"50 [36]. Specifically, for each k-level, we first perform",
"in the pipeline.",
"both the appearance image (At}t-} and its paired",
"motion map {M')T-]",
"I as the input. First, Mt & At",
"global average pooling (GAP) to generate channel-",
"1:",
"wise vectors vx and vy from each Xk and Jk.",
"pairs at frame t? are fed to two independent ResNet.",
"50 [36] branches (i.e., motion and appearance blocks",
"Next, two lxl conv, i.e., @(a; W) and (a; We),",
"in Fig. 3). The appearance features [X]k-1 and",
"with learnable parameters W and Wo, generate two",
"motion features [V]k-i extracted from K layers are",
"discriminated global descriptors: The sigmoid function",
"then sent to the Relational Cross-Attention Modules",
"o[rl = e/(er + l), e R is then applied to",
"convert the final descriptors into the interval [o, 1], i.e..",
"(RCAMs), which allows thenetwork to embed spatial-",
"temporal cross-modal features. Next, we employ the",
"into the valid attention vector for channel weighting.",
"Bidirectional Purification Modules (BPMs) with N",
"Then, we perform outer product between Xk and",
"cascaded units: BPMs focus on distilling representative",
"o [o(Vy; We)] to generate a candidate feature Qx, and",
"carriers from fused features [F)n-1 and motion-based",
"vice versa, as follows:",
"features (G)N=Finally, the predictions (ie., SM",
"QX = X @o [0(VY ; We)] ,",
"(1)",
"and st) at frame t are generated from two decoder",
"Q = Vha [d(Vx; Wo)] .",
"(2)",
"blocks",
"Then, we combine Qx, Q, and lower-level fused",
"2Here, we omit the superscript \"t\" for the convenient expression",
"feature Zk-i for in-depth feature extraction. With an",
"TSINGHRASpringer",
"IERSITVPRES"
] |
{
"arxiv_id": "2108.03151",
"language": "en",
"page_idx": 3
}
|
||
2404.12065-en-1
|
2404.12065-en-1.jpg
|
existing research involves checking cross-modal consistency and unified embedding representation of multimodal features to interpret images (Yao et al., 2023a).
LLMs like BERT (Devlin et al., 2018) and BART (Lewis et al., 2020) have advanced detection, labeling, and generation of veracity explanations to combat fake news (Das et al., 2023). Multimodal LLMs that can interpret and understand images have also been developed. Although LLMs have drawbacks, research is being conducted to remove them and enable LLMs to generate factually correct explanations. For instance, adoption of reasoning techniques like Chain of Thought have been explored in the automated fact-checking domain to combat hallucinations and improve the veracity of responses (Pan et al., 2023; Chern et al., 2023).
More recently, Retrieval-Augmented Generation (RAG) has also been adopted in fact-checking using LLMs. RAG allows LLMs to access up-to-date information without needing to be constantly retrained. Recent studies have explored using RAG with LLMs for automated fact-checking of claims. Common approaches (Asai et al., 2024; Zeng and Gao, 2024) involve converting the input claim into relevant phrase queries which are used to query a search engine for information. The information is then used to assess the veracity of the claim.
Nevertheless, a merger between reasoning techniques such as that of Chain of Thought or more advanced approaches such as Tree of Thought (Yao et al., 2023c) and RAG still needs to be explored in the domain of automated fact-checking. To this end, we present two multimodal fact-checking approaches based on utilizing the RAG response to motivate the next follow-up step to determine the veracity of political claims. We term these approaches as RAG-augmented Reasoning (RAGAR). Our RAGAR approaches employ multimodal LLMs, adept at comprehending textual and visual elements, thereby incorporating information from both modalities when verifying the claim. The two RAGAR approaches we introduce are (i) Chain of RAG (CoRAG) and (ii) Tree of RAG (ToRAG). A high-level overview of the two approaches can be seen in Figure 1. To our knowledge, this is the first work incorporating multimodal LLMs and utilizing multimodal RAG coupled with reasoning in the fact-checking domain to handle multimodal claims.
### 2 $\textbf{Related Work}$
# 2.1 Retrieval-Augmented Generation (RAG) for Fact-Checking
To combat hallucination in text generation by LLMs, current fact-checking pipelines often implement a RAG approach wherein the LLM is fed data from external sources to enhance its response. This enables the LLM to condition its generation based on the external data and is no longer limited by its knowledge cutoff.
Peng et al. (2023) present the LLM-Augmenter system, which combines external knowledge integration and automated feedback mechanisms with a set of plug-and-play modules to enhance LLM performance. Further, Chern et al.(2023) develop a factuality detection framework to assess LLM-generated text spanning multiple tasks and domains. For the KBQA task, the authors make use of Google Search API to extract relevant knowledge and parse the result. Similarly, Pan et al. (2023) propose a program-guided reasoning approach leveraging the in-context learning ability of large language models. Chain of Thought reasoning is employed to guide the model in answering complex tasks Zhang and Gao (2023) propose a Hierarchical Step-by-Step (HiSS) prompting method. This approach involves breaking down a claim into sub-claims thereby creating a hierarchy and then sequentially verifying each one through multiple question-answering steps using evidence retrieved from the web. Concurrent research is being conducted to improve the capabilities of context-based question generation using RAG evidence. Zeng and Gao (2024) argue that the approach for generating justifications for fact-checking claims needs to be able to capture the complexity of generating justifications based on retrieved evidence. To address this, they propose a novel methodology focusing on retrieving evidence to support or refute claims, rather than relying solely on summarization.
2.2 Multimodal Fact-Checking using LLMs
Multimodality is an underexplored feld in factchecking (Alam et al., 2022). Prior approaches in evaluating multimodal claims, particularly text and images, include cross-modal consistency checking or incorporating a common embedding for multimodal features (Yao et al., 2023b). With the advent of multimodal instruction-finetuned LLMs such as GPT4-V and LLaVA, newer approaches towards assessing multimodal features and associating them
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.09027099609375,
0.52490234375,
0.0843505859375,
0.5322265625,
0.09320068359375,
0.51708984375,
0.09979248046875
],
[
0.52880859375,
0.08489990234375,
0.6650390625,
0.08489990234375,
0.6650390625,
0.09857177734375,
0.52880859375,
0.09857177734375
],
[
0.1175537109375,
0.10394287109375,
0.484375,
0.10394287109375,
0.484375,
0.114013671875,
0.1175537109375,
0.114013671875
],
[
0.51123046875,
0.10986328125,
0.87158203125,
0.10986328125,
0.87158203125,
0.12469482421875,
0.51123046875,
0.12469482421875
],
[
0.1158447265625,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.5498046875,
0.1258544921875,
0.7001953125,
0.1270751953125,
0.7001953125,
0.141357421875,
0.5498046875,
0.14013671875
],
[
0.1175537109375,
0.1353759765625,
0.21826171875,
0.1353759765625,
0.21826171875,
0.146728515625,
0.1175537109375,
0.146728515625
],
[
0.5087890625,
0.1478271484375,
0.87890625,
0.1478271484375,
0.87890625,
0.160888671875,
0.5087890625,
0.160888671875
],
[
0.133544921875,
0.157958984375,
0.48779296875,
0.1591796875,
0.48779296875,
0.17333984375,
0.133544921875,
0.1728515625
],
[
0.509765625,
0.163330078125,
0.8818359375,
0.1644287109375,
0.8818359375,
0.1787109375,
0.509765625,
0.1776123046875
],
[
0.1175537109375,
0.17578125,
0.489501953125,
0.17578125,
0.489501953125,
0.190673828125,
0.1175537109375,
0.190673828125
],
[
0.509765625,
0.179931640625,
0.880859375,
0.179931640625,
0.880859375,
0.193603515625,
0.509765625,
0.193603515625
],
[
0.1158447265625,
0.1923828125,
0.489501953125,
0.1923828125,
0.489501953125,
0.2060546875,
0.1158447265625,
0.2060546875
],
[
0.5087890625,
0.1953125,
0.88671875,
0.196533203125,
0.88671875,
0.21142578125,
0.5087890625,
0.210205078125
],
[
0.1175537109375,
0.2078857421875,
0.489501953125,
0.2078857421875,
0.489501953125,
0.221435546875,
0.1175537109375,
0.221435546875
],
[
0.5087890625,
0.21142578125,
0.8818359375,
0.2120361328125,
0.8818359375,
0.226806640625,
0.5087890625,
0.2257080078125
],
[
0.1175537109375,
0.2244873046875,
0.489501953125,
0.2244873046875,
0.489501953125,
0.2381591796875,
0.1175537109375,
0.2381591796875
],
[
0.51318359375,
0.2286376953125,
0.8818359375,
0.2286376953125,
0.8818359375,
0.2423095703125,
0.51318359375,
0.2423095703125
],
[
0.1175537109375,
0.240478515625,
0.48779296875,
0.240478515625,
0.48779296875,
0.254150390625,
0.1175537109375,
0.254150390625
],
[
0.51318359375,
0.24462890625,
0.69189453125,
0.24462890625,
0.69189453125,
0.25830078125,
0.51318359375,
0.25830078125
],
[
0.1158447265625,
0.257080078125,
0.48779296875,
0.257080078125,
0.48779296875,
0.270751953125,
0.1158447265625,
0.270751953125
],
[
0.52880859375,
0.260009765625,
0.8818359375,
0.260009765625,
0.8818359375,
0.27490234375,
0.52880859375,
0.27490234375
],
[
0.11419677734375,
0.271484375,
0.489501953125,
0.2724609375,
0.489501953125,
0.287353515625,
0.11419677734375,
0.2861328125
],
[
0.509765625,
0.276611328125,
0.8818359375,
0.275634765625,
0.8818359375,
0.290283203125,
0.509765625,
0.29150390625
],
[
0.1158447265625,
0.2880859375,
0.486083984375,
0.2880859375,
0.486083984375,
0.302734375,
0.1158447265625,
0.302734375
],
[
0.51123046875,
0.29345703125,
0.8818359375,
0.292236328125,
0.8818359375,
0.30712890625,
0.51123046875,
0.30810546875
],
[
0.1175537109375,
0.305908203125,
0.490234375,
0.305908203125,
0.490234375,
0.319580078125,
0.1175537109375,
0.319580078125
],
[
0.509765625,
0.308837890625,
0.8818359375,
0.30810546875,
0.8818359375,
0.322509765625,
0.509765625,
0.32373046875
],
[
0.1175537109375,
0.32177734375,
0.48779296875,
0.32177734375,
0.48779296875,
0.3349609375,
0.1175537109375,
0.3349609375
],
[
0.51318359375,
0.325927734375,
0.8818359375,
0.325927734375,
0.8818359375,
0.339111328125,
0.51318359375,
0.339111328125
],
[
0.1158447265625,
0.33740234375,
0.48779296875,
0.33740234375,
0.48779296875,
0.350830078125,
0.1158447265625,
0.350830078125
],
[
0.51123046875,
0.341552734375,
0.8818359375,
0.341552734375,
0.8818359375,
0.355224609375,
0.51123046875,
0.355224609375
],
[
0.11419677734375,
0.352783203125,
0.46337890625,
0.3515625,
0.46337890625,
0.366455078125,
0.11419677734375,
0.36767578125
],
[
0.51123046875,
0.355712890625,
0.8818359375,
0.35693359375,
0.8818359375,
0.371826171875,
0.51123046875,
0.37060546875
],
[
0.1351318359375,
0.37890625,
0.48779296875,
0.37890625,
0.48779296875,
0.393798828125,
0.1351318359375,
0.393798828125
],
[
0.51123046875,
0.37353515625,
0.8818359375,
0.37353515625,
0.8818359375,
0.38720703125,
0.51123046875,
0.38720703125
],
[
0.509765625,
0.388427734375,
0.8818359375,
0.388427734375,
0.8818359375,
0.402099609375,
0.509765625,
0.402099609375
],
[
0.1158447265625,
0.394287109375,
0.489501953125,
0.394287109375,
0.489501953125,
0.407958984375,
0.1158447265625,
0.407958984375
],
[
0.1158447265625,
0.410400390625,
0.48779296875,
0.410888671875,
0.48779296875,
0.42578125,
0.1158447265625,
0.424560546875
],
[
0.51318359375,
0.40625,
0.88330078125,
0.40625,
0.88330078125,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.509765625,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435791015625,
0.509765625,
0.435791015625
],
[
0.11419677734375,
0.427490234375,
0.486083984375,
0.427490234375,
0.486083984375,
0.441162109375,
0.11419677734375,
0.441162109375
],
[
0.11419677734375,
0.443603515625,
0.486083984375,
0.443603515625,
0.486083984375,
0.457275390625,
0.11419677734375,
0.457275390625
],
[
0.51123046875,
0.438232421875,
0.88330078125,
0.438232421875,
0.88330078125,
0.453125,
0.51123046875,
0.453125
],
[
0.1175537109375,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.1175537109375,
0.473876953125
],
[
0.509765625,
0.454833984375,
0.884765625,
0.454833984375,
0.884765625,
0.468505859375,
0.509765625,
0.468505859375
],
[
0.1175537109375,
0.4755859375,
0.48779296875,
0.474365234375,
0.48779296875,
0.4892578125,
0.1175537109375,
0.490478515625
],
[
0.5087890625,
0.47021484375,
0.8818359375,
0.47021484375,
0.8818359375,
0.48388671875,
0.5087890625,
0.48388671875
],
[
0.509765625,
0.485107421875,
0.8818359375,
0.486328125,
0.8818359375,
0.50048828125,
0.509765625,
0.49951171875
],
[
0.1158447265625,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.505859375,
0.1158447265625,
0.505859375
],
[
0.51318359375,
0.5029296875,
0.88330078125,
0.5029296875,
0.88330078125,
0.5166015625,
0.51318359375,
0.5166015625
],
[
0.118408203125,
0.51025390625,
0.486083984375,
0.51025390625,
0.486083984375,
0.52099609375,
0.118408203125,
0.52099609375
],
[
0.51123046875,
0.5185546875,
0.880859375,
0.5185546875,
0.880859375,
0.5322265625,
0.51123046875,
0.5322265625
],
[
0.1175537109375,
0.52490234375,
0.489501953125,
0.52490234375,
0.489501953125,
0.5380859375,
0.1175537109375,
0.5380859375
],
[
0.1175537109375,
0.54052734375,
0.444091796875,
0.54052734375,
0.444091796875,
0.5546875,
0.1175537109375,
0.5546875
],
[
0.51318359375,
0.53515625,
0.8818359375,
0.53515625,
0.8818359375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.51123046875,
0.55029296875,
0.880859375,
0.55029296875,
0.880859375,
0.56396484375,
0.51123046875,
0.56396484375
],
[
0.133544921875,
0.5654296875,
0.490234375,
0.56640625,
0.490234375,
0.58056640625,
0.133544921875,
0.580078125
],
[
0.509765625,
0.56787109375,
0.8818359375,
0.56787109375,
0.8818359375,
0.58056640625,
0.509765625,
0.58056640625
],
[
0.1175537109375,
0.5830078125,
0.489501953125,
0.5830078125,
0.489501953125,
0.5966796875,
0.1175537109375,
0.5966796875
],
[
0.51123046875,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.1158447265625,
0.5986328125,
0.489501953125,
0.5986328125,
0.489501953125,
0.6123046875,
0.1158447265625,
0.6123046875
],
[
0.51123046875,
0.599609375,
0.8818359375,
0.599609375,
0.8818359375,
0.61328125,
0.51123046875,
0.61328125
],
[
0.1158447265625,
0.61328125,
0.486083984375,
0.61328125,
0.486083984375,
0.626953125,
0.1158447265625,
0.626953125
],
[
0.509765625,
0.6142578125,
0.880859375,
0.6142578125,
0.880859375,
0.62744140625,
0.509765625,
0.62744140625
],
[
0.1158447265625,
0.63134765625,
0.48779296875,
0.63134765625,
0.48779296875,
0.64453125,
0.1158447265625,
0.64453125
],
[
0.509765625,
0.6298828125,
0.880859375,
0.63134765625,
0.880859375,
0.6455078125,
0.509765625,
0.64453125
],
[
0.1158447265625,
0.646484375,
0.48779296875,
0.646484375,
0.48779296875,
0.6611328125,
0.1158447265625,
0.6611328125
],
[
0.509765625,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.6611328125,
0.509765625,
0.6611328125
],
[
0.1158447265625,
0.6630859375,
0.490234375,
0.6630859375,
0.490234375,
0.677734375,
0.1158447265625,
0.677734375
],
[
0.5087890625,
0.662109375,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.5087890625,
0.6767578125
],
[
0.1158447265625,
0.6796875,
0.490234375,
0.6796875,
0.490234375,
0.693359375,
0.1158447265625,
0.693359375
],
[
0.51123046875,
0.68115234375,
0.880859375,
0.68115234375,
0.880859375,
0.69140625,
0.51123046875,
0.69140625
],
[
0.1158447265625,
0.6953125,
0.490234375,
0.6953125,
0.490234375,
0.708984375,
0.1158447265625,
0.708984375
],
[
0.509765625,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.509765625,
0.71044921875
],
[
0.1158447265625,
0.71142578125,
0.490234375,
0.71142578125,
0.490234375,
0.7255859375,
0.1158447265625,
0.7255859375
],
[
0.5087890625,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.5087890625,
0.724609375
],
[
0.1158447265625,
0.7255859375,
0.490234375,
0.72705078125,
0.490234375,
0.7412109375,
0.1158447265625,
0.74072265625
],
[
0.509765625,
0.72802734375,
0.8388671875,
0.72802734375,
0.8388671875,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.11419677734375,
0.7421875,
0.484375,
0.7421875,
0.484375,
0.755859375,
0.11419677734375,
0.755859375
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51318359375,
0.7548828125,
0.869140625,
0.7548828125,
0.869140625,
0.7685546875,
0.51318359375,
0.7685546875
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.11419677734375,
0.79052734375,
0.486083984375,
0.79052734375,
0.486083984375,
0.80419921875,
0.11419677734375,
0.80419921875
],
[
0.51123046875,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.1158447265625,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.509765625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.5087890625,
0.84033203125,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.5087890625,
0.85498046875
],
[
0.1158447265625,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.50537109375,
0.85546875,
0.88330078125,
0.85498046875,
0.88330078125,
0.869140625,
0.50537109375,
0.87060546875
],
[
0.11419677734375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.11419677734375,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.11419677734375,
0.90185546875
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90625,
0.172119140625,
0.90625,
0.172119140625,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"existing research involves checking cross-modal",
"2",
"Related Work",
"consistency andunifiedembeddingrepresentation",
"2.1 Retrieval-Augmented Generation (RAG)",
"of multimodal features to interpret images (Yao",
"for Fact-Checking",
"et al.,2023a).",
"To combat hallucination in text generation by",
"LLMs like BERT (Devlin et al., 2018) and BART",
"LLMs, current fact-checking pipelines often im-",
"(Lewis et al., 2020) have advanced detection, la-",
"plement a RAG approach wherein the LLM is fed",
"beling, and generation of veracity explanations to",
" data from external sources to enhance its response.",
"combat fake news (Das et al., 2023). Multimodal",
"This enables the LLM to condition its generation",
"LLMs that can interpret and understand images",
"based on the external data and is no longer limited",
"have also been developed. Although LLMs have",
"by its knowledge cutoff.",
"drawbacks, research is being conducted to remove",
"Peng et al. (2023) present the LLM-Augmenter",
" them and enable LLMs to generate factually correct",
"system, which combines external knowledge inte-",
"explanations. For instance, adoption of reasoning",
"gration and automated feedback mechanisms with",
"techniques like Chain of Thought have been ex-",
"a set of plug-and-play modules to enhance LLM",
"plored in the automated fact-checking domain to",
"performance. Further, Chern et al. (2023) de-",
"combat hallucinations and improve the veracity of",
"velop a factuality detection framework to assess",
"responses (Pan et al., 2023; Chern et al., 2023)",
"LLM-generated text spanning multiple tasks and",
"More recently, Retrieval-Augmented Generation",
"domains. For the KBQA task, the authors make use",
"of Google Search API to extract relevant knowl-",
"(RAG) has also been adopted in fact-checking us-",
"ing LLMs. RAG allows LLMs to access up-to-date",
"edge and parse the result.Similarly, Pan et al.",
"(2023) propose a program-guided reasoning ap-",
"information without needing to be constantly re-",
"trained. Recent studies have explored using RAG",
"proach leveraging the in-context learning ability of",
"with LLMs for automated fact-checking of claims.",
"large language models. Chain of Thought reason-",
"Common approaches (Asai et al., 2024; Zeng and",
"ing is employed to guide the model in answering",
"complex tasks Zhang and Gao (2023) propose a Hi-",
"Gao, 2024) involve converting the input claim into",
"erarchical Step-by-Step (HiSS) prompting method.",
"relevantphrasequerieswhichareusedtoquerya",
"This approach involves breaking down a claim into",
"search engine for information. The information is",
"then used to assess the veracity of the claim.",
"sub-claims thereby creating a hierarchy and then",
"sequentially verifying each one through multiple",
"Nevertheless, a merger between reasoning tech-",
"question-answering steps using evidence retrieved",
"niques such as that of Chain of Thought or more",
"from the web. Concurrent researchis being con-",
"advanced approaches such as Tree of Thought (Yao",
"ducted to improve the capabilities of context-based",
"et al., 2023c) and RAG still needs to be explored",
"question generation using RAG evidence. Zeng and",
"in the domain of automated fact-checking. To",
"Gao (2024) argue that the approach for generating",
"this end, we present two multimodal fact-checking",
"justifications for fact-checking claims needs to be",
"approaches based on utilizing the RAG response",
"able to capture the complexity of generating justi-",
"to motivate the next follow-up step to determine",
"fications based onretrieved evidence.Toaddress",
"the veracity of political claims. We term these",
"this, they propose a novel methodology focusing",
"approaches as RAG-augmented Reasoning (RA-",
"on retrieving evidence to support or refute claims,",
"GAR). Our RAGAR approaches employ multi-",
"rather than relying solely on summarization.",
" modal LLMs, adept at comprehending textual and",
"visual elements, thereby incorporating information",
"2.2 Multimodal Fact-Checking using LLMs",
"from both modalities when verifying the claim. The",
"Multimodality is an underexplored field in fact-",
" two RAGAR approaches we introduce are (i) Chain",
"checking (Alam et al., 2022). Prior approaches in",
"of RAG (CoRAG) and (ii) Tree of RAG (ToRAG).",
"evaluating multimodal claims, particularly text and",
"A high-level overview of the two approaches can",
"images, include cross-modal consistency checking",
"be seen in Figure l. To our knowledge, this is",
"or incorporating a common embedding for multi-",
"the first work incorporating multimodal LLMs and",
"modal features (Yao et al., 2023b).With the advent",
" utilizing multimodal RAG coupled with reasoning",
" of multimodal instruction-finetuned LLMs such as",
"in the fact-checking domain to handle multimodal",
"GPT4-V and LLaVA, newer approaches towards",
"claims.",
"assessing multimodal features and associating them"
] |
{
"arxiv_id": "2404.12065",
"language": "en",
"page_idx": 1
}
|
||
2404.12065-en-2
|
2404.12065-en-2.jpg
|
<FigureHere>
Figure 2: A detailed overview of the Multimodal Fact-checking pipeline
with textual input can be developed for generating
veracity labels and explanations for claims.
Guo et al. (2023) introduce Img2Prompt, which leverages LLM-agnostic models to generate textual prompts from images. These prompts effectively guide LLMs in generating meaningful responses to VQA queries, showcasing the method's efficacy in handling zero-shot learning scenarios. Yao et al. (2023a) develop an end-to-end multimodal factchecking and explanation generation pipeline The authors construct a multimodal dataset using PolitiFact and Snopes, including the source text and images of the original claims that these articles address. The multimodal evidence retrieval encodes and reranks each sentence in the document corpus in relation to the claim. A CLIP (Radford et al., 2021) encoding is used for images, and then similarity is computed. An attention model is used for multimodal claim verification, while BART (Lewis et al., 2020) is used for explanation generation.
## 3 Dataset
We use the MOCHEG dataset (Yao et al., 2023b) for our experiments. MOCHEG provides a set of 21,184 multimodal claims. The claims are sourced from two fact-checking websites, PolitiFact$^1$ and Snopes$^{2}.$ Following Hanselowski et al.(2019), the MOCHEG dataset simplifies the labels from the two websites into: supported, refuted and NEI (Not $Enough\textit{Information}) .$
MOCHEG provides a test set with 2,007 multimodal claims. We further prune this set to only include claims that were fact-checked by PolitiFact.
$^{1}$https://www.politifact.com/
$^{2}$https://www.snopes.com/
This reduces our test samples to 794. We pick 300 test samples from this set of 794 and ensure that these test samples are selected at random from supported or refuted labels only, discarding the NEI labels. Therefore, our test dataset contains 150 supported and 150 refuted multimodal claims.
The supported and refuted labels signify that sufficient web-based information exists for verifying claims with high confdence. Conversely, the NEI (Not Enough Information) label corresponds to PolitiFact's $Half\textit{True designation, which indicates}$ a claim is partially accurate but omits significant details or contexts. This study does not explore any enhancements in LLM reasoning to understand the context or the nature of claims more deeply. Furthermore, the interpretation of NEI by LLMs, which may occur due to inadequate data for factchecking or unsuccessful evidence retrieval, differs significantly from PolitiFact's "Half True."
## 4 Multimodal Fact-Checking Pipeline
Our fact-checking pipeline comprises 4 parts: (i) Multimodal Claim Generation, (ii) Multimodal Evidence Retrieval, (iii) LLM-based and RAGaugmented Reasoning for Fact-checking, and finally the (iv) Veracity Prediction and Explanation. The pipeline is visualized in Figure 2.
## 4.1 Multimodal Claim Generation
Given an input claim as text, an associated image and the date of the claim, the task of the claim generation module is to generate a suitable response detailing both the image and textual claim information. We make use of GPT-4V (OpenAI, 2023) as our multimodal LLM for this task.
|
[
[
0.5439453125,
0.1092529296875,
0.59619140625,
0.1092529296875,
0.59619140625,
0.11639404296875,
0.5439453125,
0.11639404296875
],
[
0.5146484375,
0.119384765625,
0.59423828125,
0.119384765625,
0.59423828125,
0.1383056640625,
0.5146484375,
0.1383056640625
],
[
0.274658203125,
0.157958984375,
0.3525390625,
0.157958984375,
0.3525390625,
0.168701171875,
0.274658203125,
0.168701171875
],
[
0.52490234375,
0.157958984375,
0.55322265625,
0.157958984375,
0.55322265625,
0.162109375,
0.52490234375,
0.162109375
],
[
0.6630859375,
0.1602783203125,
0.71728515625,
0.1602783203125,
0.71728515625,
0.17578125,
0.6630859375,
0.17578125
],
[
0.6767578125,
0.1644287109375,
0.69189453125,
0.1644287109375,
0.69189453125,
0.168701171875,
0.6767578125,
0.168701171875
],
[
0.57666015625,
0.1650390625,
0.59033203125,
0.1650390625,
0.59033203125,
0.169189453125,
0.57666015625,
0.169189453125
],
[
0.291259765625,
0.168701171875,
0.345947265625,
0.168701171875,
0.345947265625,
0.1728515625,
0.291259765625,
0.1728515625
],
[
0.5205078125,
0.163330078125,
0.5634765625,
0.160888671875,
0.5634765625,
0.169189453125,
0.5205078125,
0.171630859375
],
[
0.2939453125,
0.1728515625,
0.341796875,
0.1728515625,
0.341796875,
0.182861328125,
0.2939453125,
0.182861328125
],
[
0.5693359375,
0.174560546875,
0.58447265625,
0.174560546875,
0.58447265625,
0.1787109375,
0.5693359375,
0.1787109375
],
[
0.6689453125,
0.17333984375,
0.7109375,
0.17333984375,
0.7109375,
0.18115234375,
0.6689453125,
0.18115234375
],
[
0.51123046875,
0.23095703125,
0.5634765625,
0.23095703125,
0.5634765625,
0.2423095703125,
0.51123046875,
0.2423095703125
],
[
0.255126953125,
0.311279296875,
0.7421875,
0.311279296875,
0.7421875,
0.32470703125,
0.255126953125,
0.32470703125
],
[
0.1175537109375,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.51123046875,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.365234375,
0.51123046875,
0.365234375
],
[
0.1175537109375,
0.36767578125,
0.4375,
0.36767578125,
0.4375,
0.38134765625,
0.1175537109375,
0.38134765625
],
[
0.51318359375,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.1368408203125,
0.383056640625,
0.48779296875,
0.383056640625,
0.48779296875,
0.39794921875,
0.1368408203125,
0.39794921875
],
[
0.509765625,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.509765625,
0.39794921875
],
[
0.1175537109375,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.1175537109375,
0.413330078125
],
[
0.509765625,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.509765625,
0.413330078125
],
[
0.1175537109375,
0.416259765625,
0.48779296875,
0.416259765625,
0.48779296875,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.5087890625,
0.4150390625,
0.8818359375,
0.41455078125,
0.8818359375,
0.427490234375,
0.5087890625,
0.4287109375
],
[
0.1158447265625,
0.431640625,
0.486083984375,
0.431640625,
0.486083984375,
0.4453125,
0.1158447265625,
0.4453125
],
[
0.5087890625,
0.431640625,
0.8505859375,
0.431640625,
0.8505859375,
0.4453125,
0.5087890625,
0.4453125
],
[
0.1175537109375,
0.448974609375,
0.486083984375,
0.448974609375,
0.486083984375,
0.4619140625,
0.1175537109375,
0.4619140625
],
[
0.52880859375,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.52880859375,
0.4619140625
],
[
0.1175537109375,
0.46435546875,
0.489501953125,
0.46435546875,
0.489501953125,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.1158447265625,
0.478515625,
0.491943359375,
0.479736328125,
0.491943359375,
0.49462890625,
0.1158447265625,
0.493408203125
],
[
0.51123046875,
0.48095703125,
0.8818359375,
0.48095703125,
0.8818359375,
0.49462890625,
0.51123046875,
0.49462890625
],
[
0.1175537109375,
0.496337890625,
0.489501953125,
0.496337890625,
0.489501953125,
0.51123046875,
0.1175537109375,
0.51123046875
],
[
0.51123046875,
0.496337890625,
0.8818359375,
0.496337890625,
0.8818359375,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.1175537109375,
0.5126953125,
0.490234375,
0.5126953125,
0.490234375,
0.525390625,
0.1175537109375,
0.525390625
],
[
0.5087890625,
0.5126953125,
0.880859375,
0.5126953125,
0.880859375,
0.525390625,
0.5087890625,
0.525390625
],
[
0.1175537109375,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.5419921875,
0.1175537109375,
0.5419921875
],
[
0.509765625,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.509765625,
0.5419921875
],
[
0.1175537109375,
0.54443359375,
0.490234375,
0.54443359375,
0.490234375,
0.55810546875,
0.1175537109375,
0.55810546875
],
[
0.509765625,
0.54345703125,
0.8818359375,
0.54443359375,
0.8818359375,
0.55859375,
0.509765625,
0.55810546875
],
[
0.1175537109375,
0.56005859375,
0.489501953125,
0.56005859375,
0.489501953125,
0.57373046875,
0.1175537109375,
0.57373046875
],
[
0.509765625,
0.56005859375,
0.880859375,
0.56005859375,
0.880859375,
0.57373046875,
0.509765625,
0.57373046875
],
[
0.1158447265625,
0.576171875,
0.486083984375,
0.576171875,
0.486083984375,
0.5888671875,
0.1158447265625,
0.5888671875
],
[
0.51123046875,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.1175537109375,
0.5927734375,
0.489501953125,
0.5927734375,
0.489501953125,
0.60546875,
0.1175537109375,
0.60546875
],
[
0.51318359375,
0.59326171875,
0.88330078125,
0.59326171875,
0.88330078125,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.1175537109375,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62353515625,
0.1175537109375,
0.62353515625
],
[
0.509765625,
0.60791015625,
0.8818359375,
0.609375,
0.8818359375,
0.62353515625,
0.509765625,
0.62255859375
],
[
0.118408203125,
0.6259765625,
0.486083984375,
0.6259765625,
0.486083984375,
0.6357421875,
0.118408203125,
0.6357421875
],
[
0.51123046875,
0.6259765625,
0.8818359375,
0.6259765625,
0.8818359375,
0.6396484375,
0.51123046875,
0.6396484375
],
[
0.1158447265625,
0.64111328125,
0.489501953125,
0.64111328125,
0.489501953125,
0.65478515625,
0.1158447265625,
0.65478515625
],
[
0.509765625,
0.64111328125,
0.8271484375,
0.64013671875,
0.8271484375,
0.65478515625,
0.509765625,
0.65625
],
[
0.11419677734375,
0.65625,
0.46337890625,
0.65673828125,
0.46337890625,
0.67138671875,
0.11419677734375,
0.67041015625
],
[
0.51318359375,
0.66943359375,
0.8505859375,
0.66943359375,
0.8505859375,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.1158447265625,
0.68505859375,
0.214111328125,
0.68505859375,
0.214111328125,
0.69677734375,
0.1158447265625,
0.69677734375
],
[
0.51318359375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.51318359375,
0.71044921875
],
[
0.1175537109375,
0.7080078125,
0.48779296875,
0.7080078125,
0.48779296875,
0.72265625,
0.1175537109375,
0.72265625
],
[
0.51318359375,
0.71142578125,
0.880859375,
0.71142578125,
0.880859375,
0.724609375,
0.51318359375,
0.724609375
],
[
0.1158447265625,
0.7236328125,
0.486083984375,
0.7236328125,
0.486083984375,
0.73681640625,
0.1158447265625,
0.73681640625
],
[
0.51123046875,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.74072265625,
0.51123046875,
0.74072265625
],
[
0.1175537109375,
0.7412109375,
0.48779296875,
0.7412109375,
0.48779296875,
0.7548828125,
0.1175537109375,
0.7548828125
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1175537109375,
0.75732421875,
0.48779296875,
0.75732421875,
0.48779296875,
0.77099609375,
0.1175537109375,
0.77099609375
],
[
0.51123046875,
0.76025390625,
0.88330078125,
0.76025390625,
0.88330078125,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.1175537109375,
0.77392578125,
0.48779296875,
0.77392578125,
0.48779296875,
0.78759765625,
0.1175537109375,
0.78759765625
],
[
0.509765625,
0.775390625,
0.79248046875,
0.775390625,
0.79248046875,
0.7890625,
0.509765625,
0.7890625
],
[
0.1158447265625,
0.7880859375,
0.486083984375,
0.7880859375,
0.486083984375,
0.8017578125,
0.1158447265625,
0.8017578125
],
[
0.1158447265625,
0.8056640625,
0.489501953125,
0.8056640625,
0.489501953125,
0.8193359375,
0.1158447265625,
0.8193359375
],
[
0.51318359375,
0.802734375,
0.79443359375,
0.802734375,
0.79443359375,
0.81640625,
0.51318359375,
0.81640625
],
[
0.11419677734375,
0.822265625,
0.27783203125,
0.822265625,
0.27783203125,
0.8359375,
0.11419677734375,
0.8359375
],
[
0.51123046875,
0.822265625,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51123046875,
0.83740234375
],
[
0.133544921875,
0.83837890625,
0.490234375,
0.8359375,
0.490234375,
0.849609375,
0.133544921875,
0.8515625
],
[
0.5087890625,
0.8388671875,
0.88330078125,
0.84033203125,
0.88330078125,
0.85498046875,
0.5087890625,
0.85400390625
],
[
0.1158447265625,
0.8525390625,
0.48779296875,
0.85400390625,
0.48779296875,
0.8681640625,
0.1158447265625,
0.86767578125
],
[
0.51318359375,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.8681640625,
0.51318359375,
0.8681640625
],
[
0.1175537109375,
0.87060546875,
0.489501953125,
0.87060546875,
0.489501953125,
0.88427734375,
0.1175537109375,
0.88427734375
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.139404296875,
0.89111328125,
0.35107421875,
0.89111328125,
0.35107421875,
0.90478515625,
0.139404296875,
0.90478515625
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.139404296875,
0.90478515625,
0.32080078125,
0.90478515625,
0.32080078125,
0.91845703125,
0.139404296875,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.767578125,
0.90478515625,
0.767578125,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"ImageCaptlot",
"(uvdecostara,",
"MultimodalClal",
"M-h",
"Pvericttin",
"SIC",
"see",
"ceneraton",
"RAGAR-",
"(GPT-4V)",
"S",
"(GPT-4V)",
"Suported",
"Figure 2: A detailed overview of the Multimodal Fact-checking pipeline",
"with textual input can be developed for generating",
"This reduces our test samples to 794. We pick 300",
"veracity labels and explanations for claims.",
"test samples from this set of 794 and ensure that",
"Guo et al. (2023) introduce Img2Prompt, which",
"these test samples are selected at random from sup",
"leverages LLM-agnostic models to generate textual",
"ported or refuted labels only, discarding the NEI",
"prompts from images. These prompts effectively",
"labels. Therefore, our test dataset contains 150",
"guide LLMs in generating meaningful responses",
" supported and 150 refuted multimodal claims",
"to VQA queries, showcasing the method's efficacy",
"The supported and refuted labels signify that",
"in handling zero-shot learning scenarios. Yao et al.",
"sufficient web-based information exists for verify-",
"(2023a) develop an end-to-end multimodal fact-",
"ing claims with high confidence. Conversely, the",
"checking and explanation generation pipeline The",
"NEI (Not Enough Information) label corresponds to",
"authors construct a multimodal dataset using Poli-",
"PolitiFact's Half True designation, which indicates",
"tiFact and Snopes, including the source text and",
"a claim is partially accurate but omits significant",
"images of the original claims that these articles ad-",
"details or contexts. This study does not explore",
"dress. The multimodal evidence retrieval encodes",
"any enhancements in LLM reasoning to understand",
"and reranks each sentence in the document corpus",
"the context or the nature of claims more deeply",
"in relation to the claim. A CLIP (Radford et al.,",
"Furthermore, the interpretation of NEI by LLMs,",
"2021) encoding is used for images, and then simi-",
"which may occur due to inadequate data for fact-",
"larityis computed. Anattentionmodelis used for",
"checking or unsuccessful evidence retrieval, differs",
"multimodal claim verification, while BART (Lewis",
"significantly from PolitiFact's \"Half True.\"",
"et al., 2020) is used for explanation generation.",
"4Multimodal Fact-Checking Pipeline",
"3Dataset",
"Our fact-checking pipeline comprises 4 parts: (i)",
"We use the MOCHEG dataset (Yao et al., 2023b)",
"Multimodal Claim Generation, (ii) Multimodal",
"for our experiments. MOCHEG provides a set of",
"Evidence Retrieval, (iii) LLM-based and RAG-",
"21,184 multimodal claims. The claims are sourced",
"augmented Reasoning for Fact-checking, and fi-",
"from two fact-checking websites, PolitiFact' and",
"nally the (iv) Veracity Prediction and Explanation",
"Snopes?. Following Hanselowski et al. (2019), the",
"The pipeline is visualized in Figure 2.",
"MOCHEG dataset simplifies the labels from the",
"two websites into: supported, refuted and NEI (Not",
"4.1 Multimodal Claim Generation",
"Enough Information).",
"Given an input claim as text, an associated image",
"MOCHEG provides atest set with 2.007mul-",
" and the date of the claim, the task of the claim gen-",
"timodal claims. We further prune this set to only",
"eration moduleis togenerate a suitableresponse",
"include claims that were fact-checked by PolitiFact.",
"detailing both the image and textual claim informa-",
"'https://www.politifact.com/",
"tion. We make use of GPT-4V (OpenAI, 2023) as",
"2https://www.snopes.com/",
"our multimodal LLM for this task."
] |
{
"arxiv_id": "2404.12065",
"language": "en",
"page_idx": 2
}
|
||
2404.12065-en-3
|
2404.12065-en-3.jpg
|
<FigureHere>
Figure 3: Chain of RAG and Tree of RAG pipeline
The generated response is divided into two sections; claim which simply contains the original text claim and image context which contains the relevant details extracted from the image by GPT-4V and used to contextualize the textual claim in light of the image information.
The image context is able to shed light on aspects such as the speaker that the claim is quoting, extract numerical information from the figures and highlight relevant textual data mentioned in the image. In addition to this, the contextualization further provides details into whether the image is of any importance to the text claim or not.
### 4.2 Multimodal Evidence Retrieval
The questions generated by the LLM-based or RAG-augmented Reasoning techniques to factcheck the claim serve as input for the multimodal evidence retrieval. The task of evidence retrieval is to retrieve relevant text snippets from the internet and use the details associated with the image to help answer the question.
For text-based evidence retrieval, we make use of the DuckDuckGo Search tool provided by $LangChain^3. We retrieve the top 10 results from$ the API and use it to answer the question. We also restrict the time frame of search by only collecting articles published in the two years window before the claim was fact-checked by PolitiFact. To mimic a real-time fact-checking scenario, we make sure
$^{3}$https://www.langchain.com/
to remove those search results that originate from “www.politifact.com”,“www.snopes.com",and “www. factcheck.org”since it is likely that they already contain answers to the claim and would thus impact the fairness of the experiment. We also remove the following social media websites:“www.tiktok.com”,“www.facebook.com”, “www. twitter. com”, “www. youtube. com" to further improve the quality of retrieved results.
Most of the images in our dataset contain faces of politicians, pictures from debates/events, government buildings etc. In such a situation, the image itself may not provide much additional information over the text claim. However, it is useful to determine the metadata behind the image such as the date when the claim was made or the venue where it was made. For this purpose, we make use of SerpAPI$^{4}$ to conduct reverse image search over the associated images. We extract the captions for the images from the first 10 websites and use it as additional information.
4.3 LLM-Based and RAG-Augmented
Reasoning for Fact-Checking
### 4.3.1 RAG-Augmented Reasoning: Chain of $\mathbf{RAG}\left(\mathbf{CoRAG}\right)$
Chain of RAG (CoRAG) builds upon general RAG approaches by using sequential follow-up questions augmented from the RAG response to retrieve further evidence. An overview of the approach is
$^{4}$https://serpapi.com/
|
[
[
0.45751953125,
0.0843505859375,
0.54150390625,
0.0843505859375,
0.54150390625,
0.09442138671875,
0.45751953125,
0.09442138671875
],
[
0.32080078125,
0.097412109375,
0.39794921875,
0.097412109375,
0.39794921875,
0.1051025390625,
0.32080078125,
0.1051025390625
],
[
0.611328125,
0.10272216796875,
0.62451171875,
0.10272216796875,
0.62451171875,
0.105712890625,
0.611328125,
0.105712890625
],
[
0.443359375,
0.11102294921875,
0.46337890625,
0.11102294921875,
0.46337890625,
0.114013671875,
0.443359375,
0.114013671875
],
[
0.447509765625,
0.21142578125,
0.53662109375,
0.208984375,
0.53662109375,
0.223876953125,
0.447509765625,
0.2257080078125
],
[
0.420654296875,
0.236328125,
0.459228515625,
0.236328125,
0.459228515625,
0.243408203125,
0.420654296875,
0.243408203125
],
[
0.42822265625,
0.273681640625,
0.44580078125,
0.273681640625,
0.44580078125,
0.276611328125,
0.42822265625,
0.276611328125
],
[
0.568359375,
0.302734375,
0.58251953125,
0.302734375,
0.58251953125,
0.305908203125,
0.568359375,
0.305908203125
],
[
0.328369140625,
0.371826171875,
0.6708984375,
0.371826171875,
0.6708984375,
0.385498046875,
0.328369140625,
0.385498046875
],
[
0.1368408203125,
0.413330078125,
0.491943359375,
0.413330078125,
0.491943359375,
0.427490234375,
0.1368408203125,
0.427490234375
],
[
0.509765625,
0.413330078125,
0.8818359375,
0.413330078125,
0.8818359375,
0.427001953125,
0.509765625,
0.427001953125
],
[
0.1158447265625,
0.429931640625,
0.490234375,
0.429931640625,
0.490234375,
0.443603515625,
0.1158447265625,
0.443603515625
],
[
0.50537109375,
0.429931640625,
0.880859375,
0.429931640625,
0.880859375,
0.443603515625,
0.50537109375,
0.443603515625
],
[
0.1175537109375,
0.4453125,
0.49365234375,
0.4453125,
0.49365234375,
0.460205078125,
0.1175537109375,
0.460205078125
],
[
0.50732421875,
0.4453125,
0.880859375,
0.4453125,
0.880859375,
0.458984375,
0.50732421875,
0.458984375
],
[
0.11419677734375,
0.46142578125,
0.48779296875,
0.460205078125,
0.48779296875,
0.474365234375,
0.11419677734375,
0.4755859375
],
[
0.5087890625,
0.4619140625,
0.880859375,
0.4619140625,
0.880859375,
0.4755859375,
0.5087890625,
0.4755859375
],
[
0.1158447265625,
0.47802734375,
0.489501953125,
0.47802734375,
0.489501953125,
0.4912109375,
0.1158447265625,
0.4912109375
],
[
0.509765625,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.509765625,
0.4912109375
],
[
0.1158447265625,
0.49462890625,
0.306396484375,
0.49462890625,
0.306396484375,
0.50830078125,
0.1158447265625,
0.50830078125
],
[
0.509765625,
0.49462890625,
0.88330078125,
0.49462890625,
0.88330078125,
0.50830078125,
0.509765625,
0.50830078125
],
[
0.1368408203125,
0.51025390625,
0.490234375,
0.51025390625,
0.490234375,
0.52490234375,
0.1368408203125,
0.52490234375
],
[
0.509765625,
0.51025390625,
0.88330078125,
0.51025390625,
0.88330078125,
0.52392578125,
0.509765625,
0.52392578125
],
[
0.1158447265625,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.1158447265625,
0.54052734375
],
[
0.5087890625,
0.52685546875,
0.88330078125,
0.52685546875,
0.88330078125,
0.54052734375,
0.5087890625,
0.54052734375
],
[
0.1175537109375,
0.5419921875,
0.48779296875,
0.5419921875,
0.48779296875,
0.5556640625,
0.1175537109375,
0.5556640625
],
[
0.509765625,
0.5419921875,
0.83447265625,
0.5419921875,
0.83447265625,
0.5556640625,
0.509765625,
0.5556640625
],
[
0.1175537109375,
0.56005859375,
0.486083984375,
0.56005859375,
0.486083984375,
0.57080078125,
0.1175537109375,
0.57080078125
],
[
0.52880859375,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.52880859375,
0.572265625
],
[
0.1175537109375,
0.57470703125,
0.489501953125,
0.57470703125,
0.489501953125,
0.58837890625,
0.1175537109375,
0.58837890625
],
[
0.509765625,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58740234375,
0.509765625,
0.58740234375
],
[
0.1175537109375,
0.59130859375,
0.489501953125,
0.59130859375,
0.489501953125,
0.60498046875,
0.1175537109375,
0.60498046875
],
[
0.5087890625,
0.5888671875,
0.8818359375,
0.59033203125,
0.8818359375,
0.60498046875,
0.5087890625,
0.60400390625
],
[
0.1175537109375,
0.60791015625,
0.429931640625,
0.60791015625,
0.429931640625,
0.62158203125,
0.1175537109375,
0.62158203125
],
[
0.509765625,
0.60693359375,
0.8818359375,
0.60693359375,
0.8818359375,
0.62060546875,
0.509765625,
0.62060546875
],
[
0.5087890625,
0.62158203125,
0.880859375,
0.62158203125,
0.880859375,
0.63525390625,
0.5087890625,
0.63525390625
],
[
0.1175537109375,
0.63427734375,
0.405517578125,
0.63427734375,
0.405517578125,
0.64794921875,
0.1175537109375,
0.64794921875
],
[
0.51123046875,
0.6396484375,
0.8818359375,
0.6396484375,
0.8818359375,
0.65283203125,
0.51123046875,
0.65283203125
],
[
0.1158447265625,
0.65478515625,
0.489501953125,
0.65478515625,
0.489501953125,
0.66943359375,
0.1158447265625,
0.66943359375
],
[
0.5087890625,
0.65380859375,
0.8818359375,
0.65478515625,
0.8818359375,
0.66943359375,
0.5087890625,
0.66845703125
],
[
0.1175537109375,
0.67138671875,
0.491943359375,
0.67138671875,
0.491943359375,
0.68505859375,
0.1175537109375,
0.68505859375
],
[
0.509765625,
0.67138671875,
0.8818359375,
0.67138671875,
0.8818359375,
0.68505859375,
0.509765625,
0.68505859375
],
[
0.1175537109375,
0.6884765625,
0.489501953125,
0.6884765625,
0.489501953125,
0.70166015625,
0.1175537109375,
0.70166015625
],
[
0.509765625,
0.68603515625,
0.88330078125,
0.68603515625,
0.88330078125,
0.70068359375,
0.509765625,
0.70068359375
],
[
0.1158447265625,
0.703125,
0.489501953125,
0.703125,
0.489501953125,
0.71630859375,
0.1158447265625,
0.71630859375
],
[
0.51123046875,
0.70361328125,
0.88330078125,
0.70361328125,
0.88330078125,
0.71728515625,
0.51123046875,
0.71728515625
],
[
0.1158447265625,
0.72021484375,
0.489501953125,
0.72021484375,
0.489501953125,
0.73388671875,
0.1158447265625,
0.73388671875
],
[
0.51318359375,
0.72021484375,
0.880859375,
0.72021484375,
0.880859375,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.1175537109375,
0.736328125,
0.48779296875,
0.736328125,
0.48779296875,
0.75,
0.1175537109375,
0.75
],
[
0.51318359375,
0.73681640625,
0.6826171875,
0.73681640625,
0.6826171875,
0.74755859375,
0.51318359375,
0.74755859375
],
[
0.1175537109375,
0.75439453125,
0.3046875,
0.75439453125,
0.3046875,
0.76416015625,
0.1175537109375,
0.76416015625
],
[
0.51123046875,
0.76025390625,
0.82373046875,
0.76025390625,
0.82373046875,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.1368408203125,
0.76953125,
0.486083984375,
0.76953125,
0.486083984375,
0.77978515625,
0.1368408203125,
0.77978515625
],
[
0.55322265625,
0.77783203125,
0.78515625,
0.77783203125,
0.78515625,
0.7919921875,
0.55322265625,
0.7919921875
],
[
0.118408203125,
0.7861328125,
0.484375,
0.7861328125,
0.484375,
0.79638671875,
0.118408203125,
0.79638671875
],
[
0.1175537109375,
0.7998046875,
0.48779296875,
0.7998046875,
0.48779296875,
0.81396484375,
0.1175537109375,
0.81396484375
],
[
0.51123046875,
0.798828125,
0.87158203125,
0.798828125,
0.87158203125,
0.8125,
0.51123046875,
0.8125
],
[
0.1158447265625,
0.81640625,
0.48779296875,
0.81640625,
0.48779296875,
0.830078125,
0.1158447265625,
0.830078125
],
[
0.5634765625,
0.8134765625,
0.6884765625,
0.81396484375,
0.6884765625,
0.8291015625,
0.5634765625,
0.82763671875
],
[
0.1158447265625,
0.8330078125,
0.48779296875,
0.8330078125,
0.48779296875,
0.8466796875,
0.1158447265625,
0.8466796875
],
[
0.51318359375,
0.83447265625,
0.8818359375,
0.83447265625,
0.8818359375,
0.84716796875,
0.51318359375,
0.84716796875
],
[
0.1175537109375,
0.8486328125,
0.489501953125,
0.8486328125,
0.489501953125,
0.8623046875,
0.1175537109375,
0.8623046875
],
[
0.5087890625,
0.85107421875,
0.8818359375,
0.85107421875,
0.8818359375,
0.8642578125,
0.5087890625,
0.8642578125
],
[
0.1158447265625,
0.8642578125,
0.489501953125,
0.8642578125,
0.489501953125,
0.87744140625,
0.1158447265625,
0.87744140625
],
[
0.509765625,
0.8662109375,
0.88330078125,
0.8662109375,
0.88330078125,
0.8798828125,
0.509765625,
0.8798828125
],
[
0.1158447265625,
0.88134765625,
0.489501953125,
0.88134765625,
0.489501953125,
0.89453125,
0.1158447265625,
0.89453125
],
[
0.509765625,
0.8818359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.896484375,
0.509765625,
0.896484375
],
[
0.139404296875,
0.90478515625,
0.341796875,
0.90478515625,
0.341796875,
0.91845703125,
0.139404296875,
0.91845703125
],
[
0.5341796875,
0.90478515625,
0.69189453125,
0.90478515625,
0.69189453125,
0.91845703125,
0.5341796875,
0.91845703125
]
] |
[
"ChainofRAG",
"Ctalmt maae Coptext",
"O",
"",
"Tree of RAG",
"Question",
"t",
".03",
"Figure 3: Chain of RAG and Tree of RAG pipeline",
"The generated response is divided into two sec-",
"to remove those search results that originate from",
"tions; claim which simply contains the original text",
"\"www.politifact.com\",\"www.snopes.com\",and",
"claim and image context which contains the rele-",
"\"www. factcheck. org\" since it is likely that they",
"vant details extracted from the image by GPT-4V",
" already contain answers to the claim and would",
"and used to contextualize the textual claim in light",
"thus impact the fairness of the experiment. We",
"of the image information.",
"also remove the following social media web-",
"The image context is able to shed light on as-",
"sites:“www.tiktok.com\",\"www.facebook.com\",",
"pects such as the speaker that the claim is quoting,",
"\"\"www.twitter.com\",“\"www.youtube.com\"to fur-",
"extract numerical information from the figures and",
"ther improve the quality of retrieved results.",
"highlight relevant textual data mentioned in the",
"Most of the images in our dataset contain faces",
"image. In addition to this, the contextualization",
"of politicians, pictures from debates/events,govern",
"further provides details into whether the image is",
"ment buildings etc. In such a situation, the image",
"of any importance to the text claim or not.",
" itself may not provide much additional informa",
"tion over the text claim. However, it is useful to",
"4.2Multimodal Evidence Retrieval",
"determine the metadata behind the image such as",
"The questions generated by the LLM-based or",
"the date when the claim was made or the venue",
"RAG-augmented Reasoning techniques to fact-",
"where it was made. For this purpose, we make use",
"check the claim serve as input for the multimodal",
"of SerpAPI4 to conduct reverse image search over",
"evidence retrieval.The task of evidence retrieval is",
"the associated images. We extract the captions for",
"to retrieve relevant text snippets from the internet",
"the images fromthe first 10 websites and useitas",
"and use the details associated with the image to",
"additional information.",
"helpanswerthequestion.",
"4.3 LLM-Based and RAG-Augmented",
"Fortext-basedevidence retrieval, we make use",
"Reasoning for Fact-Checking",
"of the DuckDuckGo Searchtool provided by",
"LangChain3. We retrieve the top 10 results from",
"4.3.1 RAG-Augmented Reasoning: Chain of",
"the API and use it to answer the question. We also",
"RAG (CoRAG)",
"restrict the time frame of search by only collecting",
"Chain of RAG (CoRAG) builds upon general RAG",
"articles published in the two years window before",
" approaches by using sequential follow-up questions",
"the claim was fact-checked by PolitiFact. To mimic",
" augmented from the RAG response to retrieve fur-",
"a real-time fact-checking scenario, we make sure",
"ther evidence. An overview of the approach is",
"\"https://www.langchain.com/",
"thttps://serpapi.com/"
] |
{
"arxiv_id": "2404.12065",
"language": "en",
"page_idx": 3
}
|
||
2404.12065-en-4
|
2404.12065-en-4.jpg
|
<table>
<tbody>
<tr>
<td>( Tha1t) RAG$\left ( \mathrm{CoRAG}\right )$ gorifhm 21</td>
<td> </td>
</tr>
<tr>
<td>$counter\leftarrow0$</td>
<td> </td>
</tr>
<tr>
<td>$followUpNeeded\leftarrow$True</td>
<td> </td>
</tr>
<tr>
<td>$A\leftarrow$ImageQA$(\bar{Q},\bar{I},IC)$</td>
<td>DUsing l captions</td>
</tr>
<tr>
<td>endif</td>
<td>stanuand cviuc</td>
</tr>
<tr>
<td>return OAPairs</td>
<td>$\triangleright$Returns the list of O-A nalr</td>
</tr>
</tbody>
</table>
provided in Algorithm 1. The input to the CoRAG module is the claim and image context from the multimodal claim generation. The LLM first generates a question that is intended to answer an aspect of the claim. We follow a zero-shot approach wherein the LLM is not provided with any example question/answer pairs to influence or guide the reasoning process. The generated question is passed to the multimodal evidence retriever to reteieve evidence to inform the RAG answer.
Once the answer is generated, the CoRAG process undergoes an early termination check, which we term as follow-up check. The follow-up check takes as input the LLM generated claim as well as the generated question-answer pair(s) and checks whether enough information to answer the claim has been gathered. If the response from the followup check is “Yes”, it asks a follow-up question that is already informed of all the previously answered question-answer pairs.
A follow-up check occurs after each questionanswer generation step. If the follow-up check prompt finds sufficient evidence in the questions and answers generated by the CoRAG process, it terminates and passes the evidence to the veracity prediction and explanation generation module of the pipeline. We also set a constraint of a maximum of six questions, in which case the CoRAG process terminates even if it does not have enough evidence for the fact-check.
step. An overview of the approach is provided in Algorithm 2. The input to the ToRAG module is the claim and image context from the multimodal claim generation module. Upon receiving this input, the ToRAG approach branches into three, with each branch asking a unique question to fact-check the claim.
Once the three starting questions have been generated, the ToRAG approach makes use of the evidence retriever to retrieve and generate answers for these questions. The three question-answer pairs are then passed into an elimination prompt from which only one question-answer pair is chosen as the candidate evidence. This selection is based on the criteria of relevance, detail, additional information, and answer confidence as determined by GPT-4 (see Appendix A.3).
The candidate evidence is the best questionanswer pair that was returned from the question answer elimination and now acts as the basis for the follow-up question to be asked. Yet again, three follow-up questions are generated simultaneously based on the information from the candidate evidence. The evidence retriever fetches answers to these questions, and the LLM generates the answer based on the evidence retrieved. A new candidate's evidence is chosen by the elimination prompt and is added to the existing list of candidate evidences. The candidate evidence list, therefore, stores only the best of the three question-answer pairs from each step. Upon gathering sufficient information to fact-check the claim as determined by the followup check prompt or reaching a maximum of six candidate evidence question-answer pairs, the ToRAG process terminates, and the list of candidate evidence is passed onto the veracity prediction module to generate the veracity label and the explanation.
## 4.3.2 RAG-Augmented Reasoning: Tree of $\mathbf{RAG}\left(\mathbf{ToRAG}\right)$
Tree of RAG (ToRAG) builds upon our CoRAG approach by creating question branches at each step of the reasoning. Based on the retrieved evidence, the best question-answer branch is selected at each
|
[
[
0.1175537109375,
0.08489990234375,
0.4013671875,
0.08489990234375,
0.4013671875,
0.09857177734375,
0.1175537109375,
0.09857177734375
],
[
0.145263671875,
0.105712890625,
0.411376953125,
0.105712890625,
0.411376953125,
0.11639404296875,
0.145263671875,
0.11639404296875
],
[
0.1275634765625,
0.114013671875,
0.151123046875,
0.114013671875,
0.151123046875,
0.1270751953125,
0.1275634765625,
0.1270751953125
],
[
0.1435546875,
0.114013671875,
0.322509765625,
0.11517333984375,
0.322509765625,
0.1270751953125,
0.1435546875,
0.1258544921875
],
[
0.1275634765625,
0.122314453125,
0.151123046875,
0.122314453125,
0.151123046875,
0.1353759765625,
0.1275634765625,
0.1353759765625
],
[
0.145263671875,
0.122314453125,
0.22412109375,
0.1258544921875,
0.22412109375,
0.13720703125,
0.145263671875,
0.1341552734375
],
[
0.69775390625,
0.12469482421875,
0.8818359375,
0.12469482421875,
0.8818359375,
0.1353759765625,
0.69775390625,
0.1353759765625
],
[
0.1292724609375,
0.133056640625,
0.145263671875,
0.133056640625,
0.145263671875,
0.142578125,
0.1292724609375,
0.142578125
],
[
0.1485595703125,
0.135986328125,
0.2032470703125,
0.135986328125,
0.2032470703125,
0.1436767578125,
0.1485595703125,
0.1436767578125
],
[
0.130126953125,
0.1436767578125,
0.1368408203125,
0.1436767578125,
0.1368408203125,
0.149658203125,
0.130126953125,
0.149658203125
],
[
0.1502685546875,
0.1436767578125,
0.2939453125,
0.1436767578125,
0.2939453125,
0.15380859375,
0.1502685546875,
0.15380859375
],
[
0.1275634765625,
0.15380859375,
0.14697265625,
0.15380859375,
0.14697265625,
0.162109375,
0.1275634765625,
0.162109375
],
[
0.1485595703125,
0.152587890625,
0.393798828125,
0.152587890625,
0.393798828125,
0.163330078125,
0.1485595703125,
0.163330078125
],
[
0.130126953125,
0.163330078125,
0.1368408203125,
0.163330078125,
0.1368408203125,
0.168701171875,
0.130126953125,
0.168701171875
],
[
0.16455078125,
0.1644287109375,
0.202392578125,
0.162109375,
0.202392578125,
0.169189453125,
0.16455078125,
0.171630859375
],
[
0.2091064453125,
0.160888671875,
0.32666015625,
0.160888671875,
0.32666015625,
0.171630859375,
0.2091064453125,
0.171630859375
],
[
0.1275634765625,
0.17041015625,
0.14111328125,
0.17041015625,
0.14111328125,
0.182861328125,
0.1275634765625,
0.182861328125
],
[
0.185546875,
0.1728515625,
0.322509765625,
0.1728515625,
0.322509765625,
0.182861328125,
0.185546875,
0.182861328125
],
[
0.69775390625,
0.169189453125,
0.880859375,
0.17041015625,
0.880859375,
0.18408203125,
0.69775390625,
0.182861328125
],
[
0.130126953125,
0.179931640625,
0.14111328125,
0.182861328125,
0.1368408203125,
0.190673828125,
0.1259765625,
0.18701171875
],
[
0.166259765625,
0.1817626953125,
0.185546875,
0.1817626953125,
0.185546875,
0.190673828125,
0.166259765625,
0.190673828125
],
[
0.12005615234375,
0.189453125,
0.1427001953125,
0.189453125,
0.1427001953125,
0.201904296875,
0.12005615234375,
0.201904296875
],
[
0.19140625,
0.190673828125,
0.274658203125,
0.190673828125,
0.274658203125,
0.201904296875,
0.19140625,
0.201904296875
],
[
0.7421875,
0.191162109375,
0.8818359375,
0.191162109375,
0.8818359375,
0.201904296875,
0.7421875,
0.201904296875
],
[
0.1234130859375,
0.1995849609375,
0.14111328125,
0.1995849609375,
0.14111328125,
0.2078857421875,
0.1234130859375,
0.2078857421875
],
[
0.1234130859375,
0.2078857421875,
0.14111328125,
0.2078857421875,
0.14111328125,
0.21728515625,
0.1234130859375,
0.21728515625
],
[
0.1695556640625,
0.203125,
0.19482421875,
0.203125,
0.19482421875,
0.2060546875,
0.1695556640625,
0.2060546875
],
[
0.166259765625,
0.208984375,
0.297119140625,
0.208984375,
0.297119140625,
0.2197265625,
0.166259765625,
0.2197265625
],
[
0.7080078125,
0.208984375,
0.8818359375,
0.208984375,
0.8818359375,
0.2197265625,
0.7080078125,
0.2197265625
],
[
0.1234130859375,
0.218505859375,
0.1368408203125,
0.218505859375,
0.1368408203125,
0.226806640625,
0.1234130859375,
0.226806640625
],
[
0.166259765625,
0.2197265625,
0.403076171875,
0.2197265625,
0.403076171875,
0.2298583984375,
0.166259765625,
0.2298583984375
],
[
0.12005615234375,
0.22802734375,
0.14111328125,
0.226806640625,
0.14111328125,
0.236328125,
0.12176513671875,
0.2369384765625
],
[
0.166259765625,
0.2298583984375,
0.301513671875,
0.2298583984375,
0.301513671875,
0.2369384765625,
0.166259765625,
0.2369384765625
],
[
0.1234130859375,
0.2381591796875,
0.139404296875,
0.2381591796875,
0.139404296875,
0.2452392578125,
0.1234130859375,
0.2452392578125
],
[
0.183837890625,
0.236328125,
0.370361328125,
0.236328125,
0.370361328125,
0.2493896484375,
0.183837890625,
0.2493896484375
],
[
0.1234130859375,
0.2476806640625,
0.139404296875,
0.2476806640625,
0.139404296875,
0.257080078125,
0.1234130859375,
0.257080078125
],
[
0.166259765625,
0.248779296875,
0.1956787109375,
0.248779296875,
0.1956787109375,
0.257080078125,
0.166259765625,
0.257080078125
],
[
0.12176513671875,
0.257080078125,
0.14111328125,
0.257080078125,
0.14111328125,
0.2666015625,
0.12176513671875,
0.2666015625
],
[
0.166259765625,
0.257080078125,
0.3046875,
0.257080078125,
0.3046875,
0.267333984375,
0.166259765625,
0.267333984375
],
[
0.12176513671875,
0.26318359375,
0.19482421875,
0.2666015625,
0.193115234375,
0.27783203125,
0.12005615234375,
0.27490234375
],
[
0.12005615234375,
0.27490234375,
0.2283935546875,
0.27490234375,
0.2283935546875,
0.2861328125,
0.12005615234375,
0.2861328125
],
[
0.73486328125,
0.27490234375,
0.8818359375,
0.27490234375,
0.8818359375,
0.2880859375,
0.73486328125,
0.2880859375
],
[
0.1158447265625,
0.31640625,
0.486083984375,
0.3154296875,
0.486083984375,
0.330078125,
0.1158447265625,
0.330810546875
],
[
0.51123046875,
0.317626953125,
0.8818359375,
0.317626953125,
0.8818359375,
0.330810546875,
0.51123046875,
0.330810546875
],
[
0.1158447265625,
0.334228515625,
0.489501953125,
0.334228515625,
0.489501953125,
0.347412109375,
0.1158447265625,
0.347412109375
],
[
0.51318359375,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.345703125,
0.51318359375,
0.345703125
],
[
0.1158447265625,
0.349853515625,
0.491943359375,
0.349853515625,
0.491943359375,
0.363525390625,
0.1158447265625,
0.363525390625
],
[
0.5087890625,
0.3486328125,
0.880859375,
0.3486328125,
0.880859375,
0.3623046875,
0.5087890625,
0.3623046875
],
[
0.1175537109375,
0.366455078125,
0.48779296875,
0.366455078125,
0.48779296875,
0.377197265625,
0.1175537109375,
0.377197265625
],
[
0.509765625,
0.366455078125,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.509765625,
0.380126953125
],
[
0.1175537109375,
0.3818359375,
0.48779296875,
0.3818359375,
0.48779296875,
0.3955078125,
0.1175537109375,
0.3955078125
],
[
0.509765625,
0.3818359375,
0.8818359375,
0.3818359375,
0.8818359375,
0.3955078125,
0.509765625,
0.3955078125
],
[
0.1158447265625,
0.39794921875,
0.489501953125,
0.39794921875,
0.489501953125,
0.410888671875,
0.1158447265625,
0.410888671875
],
[
0.5087890625,
0.39794921875,
0.880859375,
0.39794921875,
0.880859375,
0.410888671875,
0.5087890625,
0.410888671875
],
[
0.1175537109375,
0.41455078125,
0.490234375,
0.41455078125,
0.490234375,
0.427490234375,
0.1175537109375,
0.427490234375
],
[
0.51318359375,
0.4150390625,
0.56103515625,
0.4150390625,
0.56103515625,
0.427001953125,
0.51318359375,
0.427001953125
],
[
0.11419677734375,
0.43115234375,
0.486083984375,
0.43115234375,
0.486083984375,
0.44482421875,
0.11419677734375,
0.44482421875
],
[
0.52880859375,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.44775390625,
0.52880859375,
0.44775390625
],
[
0.1158447265625,
0.446533203125,
0.48779296875,
0.446533203125,
0.48779296875,
0.460205078125,
0.1158447265625,
0.460205078125
],
[
0.51123046875,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.51123046875,
0.46435546875
],
[
0.1158447265625,
0.4619140625,
0.387939453125,
0.4619140625,
0.387939453125,
0.4755859375,
0.1158447265625,
0.4755859375
],
[
0.51123046875,
0.468505859375,
0.87890625,
0.468505859375,
0.87890625,
0.478515625,
0.51123046875,
0.478515625
],
[
0.1351318359375,
0.479736328125,
0.489501953125,
0.479736328125,
0.489501953125,
0.493408203125,
0.1351318359375,
0.493408203125
],
[
0.51318359375,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.51318359375,
0.496337890625
],
[
0.1158447265625,
0.496337890625,
0.48779296875,
0.496337890625,
0.48779296875,
0.51025390625,
0.1158447265625,
0.51025390625
],
[
0.5087890625,
0.498779296875,
0.880859375,
0.498779296875,
0.880859375,
0.5126953125,
0.5087890625,
0.5126953125
],
[
0.1158447265625,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.1158447265625,
0.525390625
],
[
0.509765625,
0.515625,
0.880859375,
0.515625,
0.880859375,
0.529296875,
0.509765625,
0.529296875
],
[
0.1158447265625,
0.52783203125,
0.489501953125,
0.52685546875,
0.489501953125,
0.54150390625,
0.1158447265625,
0.5419921875
],
[
0.509765625,
0.53076171875,
0.8818359375,
0.53076171875,
0.8818359375,
0.54443359375,
0.509765625,
0.54443359375
],
[
0.1175537109375,
0.54443359375,
0.48779296875,
0.54443359375,
0.48779296875,
0.55810546875,
0.1175537109375,
0.55810546875
],
[
0.509765625,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.509765625,
0.56005859375
],
[
0.1158447265625,
0.55859375,
0.486083984375,
0.55859375,
0.486083984375,
0.572265625,
0.1158447265625,
0.572265625
],
[
0.509765625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.509765625,
0.57666015625
],
[
0.1175537109375,
0.57666015625,
0.490234375,
0.57666015625,
0.490234375,
0.59033203125,
0.1175537109375,
0.59033203125
],
[
0.51318359375,
0.5791015625,
0.71533203125,
0.5791015625,
0.71533203125,
0.5927734375,
0.51318359375,
0.5927734375
],
[
0.1158447265625,
0.5927734375,
0.48779296875,
0.5927734375,
0.48779296875,
0.60546875,
0.1158447265625,
0.60546875
],
[
0.53076171875,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.53076171875,
0.6123046875
],
[
0.1158447265625,
0.609375,
0.486083984375,
0.609375,
0.486083984375,
0.62255859375,
0.1158447265625,
0.62255859375
],
[
0.51123046875,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.51123046875,
0.62890625
],
[
0.1158447265625,
0.6259765625,
0.285400390625,
0.6259765625,
0.285400390625,
0.6396484375,
0.1158447265625,
0.6396484375
],
[
0.509765625,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.509765625,
0.64453125
],
[
0.1351318359375,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.1351318359375,
0.65478515625
],
[
0.509765625,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.6611328125,
0.509765625,
0.6611328125
],
[
0.11419677734375,
0.65771484375,
0.486083984375,
0.65771484375,
0.486083984375,
0.67138671875,
0.11419677734375,
0.67138671875
],
[
0.509765625,
0.662109375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.509765625,
0.6767578125
],
[
0.11419677734375,
0.67578125,
0.489501953125,
0.6748046875,
0.489501953125,
0.6884765625,
0.11419677734375,
0.689453125
],
[
0.51318359375,
0.68115234375,
0.8818359375,
0.68115234375,
0.8818359375,
0.69140625,
0.51318359375,
0.69140625
],
[
0.1158447265625,
0.6904296875,
0.48779296875,
0.6904296875,
0.48779296875,
0.705078125,
0.1158447265625,
0.705078125
],
[
0.51318359375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1158447265625,
0.7060546875,
0.486083984375,
0.7060546875,
0.486083984375,
0.7197265625,
0.1158447265625,
0.7197265625
],
[
0.51123046875,
0.71142578125,
0.88330078125,
0.71044921875,
0.88330078125,
0.724609375,
0.51123046875,
0.7255859375
],
[
0.1175537109375,
0.7236328125,
0.48779296875,
0.7236328125,
0.48779296875,
0.73681640625,
0.1175537109375,
0.73681640625
],
[
0.509765625,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.1175537109375,
0.7392578125,
0.48779296875,
0.7392578125,
0.48779296875,
0.7529296875,
0.1175537109375,
0.7529296875
],
[
0.51318359375,
0.7451171875,
0.880859375,
0.7451171875,
0.880859375,
0.755859375,
0.51318359375,
0.755859375
],
[
0.1158447265625,
0.7548828125,
0.489501953125,
0.7548828125,
0.489501953125,
0.7685546875,
0.1158447265625,
0.7685546875
],
[
0.509765625,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.509765625,
0.77392578125
],
[
0.1158447265625,
0.771484375,
0.48779296875,
0.771484375,
0.48779296875,
0.78515625,
0.1158447265625,
0.78515625
],
[
0.509765625,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.509765625,
0.7880859375
],
[
0.1158447265625,
0.78759765625,
0.2493896484375,
0.78759765625,
0.2493896484375,
0.79736328125,
0.1158447265625,
0.79736328125
],
[
0.51318359375,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.81787109375,
0.462646484375,
0.81787109375,
0.462646484375,
0.83056640625,
0.1158447265625,
0.83056640625
],
[
0.5146484375,
0.826171875,
0.8818359375,
0.826171875,
0.8818359375,
0.8359375,
0.5146484375,
0.8359375
],
[
0.17041015625,
0.83447265625,
0.291259765625,
0.83447265625,
0.291259765625,
0.84716796875,
0.17041015625,
0.84716796875
],
[
0.51123046875,
0.84130859375,
0.88330078125,
0.84130859375,
0.88330078125,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.48779296875,
0.85546875,
0.48779296875,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.11419677734375,
0.87255859375,
0.48779296875,
0.87158203125,
0.48779296875,
0.8857421875,
0.11419677734375,
0.88720703125
],
[
0.51123046875,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1158447265625,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.8759765625,
0.90478515625,
0.8759765625,
0.91845703125,
0.5087890625,
0.91845703125
]
] |
[
"Algorithm 1Chain of RAG (CoRAG)",
"Input: Claim C,Image Context I, Image Captions IC",
"2:",
"QGenerateFirstQuestion(C, I)",
"3:",
"QAPairs←",
" Initialize an empty list for Q-A pairs",
"4:",
"counter",
"5",
"followUpNeeded←True",
"6:",
"whilecounter < 6 and followUpNeeded do",
"7",
"ifOues",
"onAboutlmage(Q)then",
"8:",
"A ImageQA(Q,I,IC)",
"Ung imaquestin,andaptio",
"9",
"else",
"10:",
"1 +←WebQA(Q)",
" Standard evidence retrieval",
"11:",
"12:",
"",
"QAPairs.append((Q,A))",
"> Store the Q-A pair of this iteration",
"13",
"followUpNeededFollowupCheck(Q,A)",
"14:",
"iffollowUpNeededther",
"15",
"Q+FollowupQuestion(QAPairs)",
"16",
"end if",
"17:",
"counter + counter +1",
"18: end whil",
"19:return QAPairs",
"> Returns the list of Q-A pairs",
"provided in Algorithm 1. The input to the CoRAG",
"step. An overview of the approach is provided in",
"module is the claim and image context from the",
"Algorithm2.The inputtotheToRAGmoduleis the",
"multimodal claim generation. The LLM first gen-",
" claim and image context from the multimodal claim",
"erates aquestionthatis intended toansweran as-",
"generation module. Upon receiving this input, the",
"pect of the claim. We follow a zero-shot approach",
"ToRAG approach branches into three, with each",
"wherein the LLM is not provided with any example",
" branch asking a unique question to fact-check the",
"question/answer pairs to influence or guide the rea-",
"claim.",
" soning process. The generated question is passed",
"Once the three starting questions have been gen-",
"to the multimodal evidence retriever to reteieve",
"erated, the ToRAG approach makes use of the evi-",
"evidence to inform the RAG answer.",
"denceretrievertoretrieveandgenerateanswersfor",
"Once the answer is generated, the CoRAG pro-",
"these questions. The three question-answer pairs",
"cess undergoes an early termination check, which",
" are then passed into an elimination prompt from",
"we term as follow-up check. The follow-up check",
"which only one question-answer pair is chosen as",
"takes as input the LLM generated claim as well as",
"the candidate evidence. This selection is based on",
"the generated question-answer pair(s) and checks",
"the criteria of relevance, detail, additional infor-",
"whether enough information to answer the claim",
"mation, and answer confidence as determined by",
"has been gathered. If the response from the follow-",
"GPT-4 (see Appendix A.3).",
"up check is \"Yes\", it asks a follow-up question that",
"The candidate evidence is the best question-",
"is already informed of all the previously answered",
"answer pair that was returned from the question",
"question-answer pairs.",
"answer elimination and now acts as the basis for",
"A follow-up check occurs after each question-",
"the follow-up question to be asked. Yet again, three",
"answer generation step. If the follow-up check",
"follow-up questions are generated simultaneously",
"prompt finds sufficient evidence in the questions",
"based on the informationfrom the candidate evi-",
"and answers generated by the CoRAG process, it",
"dence. The evidence retriever fetches answers to",
"terminates and passes the evidence to the veracity",
"these questions, and the LLM generates the answer",
"prediction and explanation generation module of",
"based on the evidence retrieved. A new candidate's",
"the pipeline. We also set a constraint of a maximum",
"evidence is chosen by theeliminationprompt and",
"of six questions, in which case the CoRAG process",
"is added to the existing list of candidate evidences",
"terminates even if it does not have enough evidence",
"The candidate evidence list, therefore, stores only",
"forthe fact-check",
"the best of the three question-answer pairs from",
"each step. Upon gathering sufficient information to",
"4.3.2RAG-Augmented Reasoning: Tree of",
"fact-checktheclaimasdeterminedbythefollow-",
"RAG (ToRAG)",
"up check prompt or reaching a maximum of six can-",
"Tree of RAG (ToRAG) builds upon our CoRAG",
"didate evidence question-answer pairs, the ToRAG",
"approach by creating question branches at each step",
"process terminates, and the list of candidate evi-",
"of the reasoning. Based on the retrieved evidence,",
"dence is passed onto the veracity prediction module",
"the best question-answer branch is selected at each",
" to generate the veracity label and the explanation."
] |
{
"arxiv_id": "2404.12065",
"language": "en",
"page_idx": 4
}
|
||
1909.05608-en-0
|
1909.05608-en-0.jpg
|
## $\mathbf{ABSApp:~A~Portable~Weakly-Supervised}$ $\textbf{Aspect- Based Sentiment Extraction System}$
$\textbf{Oren Pereg}^1,\textbf{Daniel Korat}^1,\textbf{Moshe Wasserblat}^1,\textbf{Jonathan Mamou}^1,\textbf{Ido Dagan}^2$
$^{1}$Intel AI Lab, Petah Tikva, Israel
$^{2}$Department of Computer Science, Bar-Ilan University, Ramat Gan, Israel
$^{1}$firstname.lastname@intel.com
$^{2}$daganêcs.biu.ac.il
### $\mathbf{Abstract}$
We present ABSApp, a portable system for weakly-supervised aspect-based sentiment extraction$^{\mathrm{l} }.$ The system is interpretable and user friendly and does not require labeled training data, hence can be rapidly and costeffectively used across different domains in applied setups. The system flow includes three stages: First, it generates domain-specific aspect and opinion lexicons based on an unlabeled dataset; second, it enables the user to view and edit those lexicons (weak supervision); and finally, it enables the user to select an unlabeled target dataset from the same domain, classify it, and generate an aspect-based sentiment report. ABSApp has been successfully used in a number of real-life use cases, among them movie review analysis and convention impact analysis.
## 1 Introduction
Aspect Based Sentiment Analysis (ABSA) is the task of extracting, from a given corpus, opinion targets (aspect terms) and the sentiment expressed towards them. For example, in the sentence “The $dessertwasincredible"$,the aspect term is dessert and the sentiment towards it is positive. This finegrained trait of ABSA makes it an effective application for measuring and monitoring the ratio between positive and negative opinions expressed towards specific aspects of a product or service.
Most work around ABSA focused on supervised sequence tagging based systems. Liu et al. (2015) showed promising results when the training and the inference data are from the same domain. However, this approach is typically not robust across different domains since aspect terms from two different domains are usually semantically different hence separated in the embedding
A demo video of ABSApp is available at
https://drive.google.com/open?id= 1BLkOxk jIOqyRhNy4UQEFQpDF\_KR\_NMAd.
vector space. For example, frequent aspect terms in the restaurant domain, like $food,menu,starters$ and salad, have little or no semantic relatedness to frequent aspect terms in the laptop domain, like screen size, keyboard and battery life. In addition to aspect terms, many opinion terms are also domain-specific. For example, opinion terms like tasty (positive), yummy (positive), flavorful (positive) and tasteless (negative) are specific to the restaurant domain whereas opinion terms like $lightweight(positive),durable(positive),compat-$ ible (positive) and heavy (negative) are specific to the laptop domain. This makes domain-agnostic ABSA a challenging task, with little work addressing it.
A recent line of work is based on transferlearning methods, in which labeled data from a source domain is used for training a model to classify data in a target domain. Ding et al. (2017) and Wang and Jialin Pan (2018) proposed using supervised RNNs for cross-domain aspect term extraction and for aspect and opinion term co-extraction. This approach showed encouraging results, however it requires a considerable amount of labeled data from the source domain which is often not practical in applied settings due to cost or legal considerations (relevant data is usually not available for commercial use).
Another approach towards domain robustness is based on unsupervised methods. Hu and Liu (2004) used association rules and Qiu et al. (2011) used syntactic rules for aspect and opinion term co-extraction. Industrial setups usually lack labeled data for training and this is where unsupervised methods excel. However, these methods can be noisy (see the ABSApp-unsup baseline in Table 2). In this paper we show that weak supervision, namely a short manual process of editing lexicons that were generated by an unsupervised method, produces results that are comparable to
|
[
[
0.29296875,
0.0831298828125,
0.70947265625,
0.0831298828125,
0.70947265625,
0.097412109375,
0.29296875,
0.097412109375
],
[
0.2822265625,
0.10272216796875,
0.72119140625,
0.10272216796875,
0.72119140625,
0.1175537109375,
0.2822265625,
0.1175537109375
],
[
0.15869140625,
0.14013671875,
0.84814453125,
0.1395263671875,
0.84814453125,
0.15673828125,
0.15869140625,
0.157958984375
],
[
0.36865234375,
0.1602783203125,
0.6396484375,
0.1602783203125,
0.6396484375,
0.17333984375,
0.36865234375,
0.17333984375
],
[
0.2032470703125,
0.174560546875,
0.8037109375,
0.17578125,
0.8037109375,
0.190673828125,
0.2032470703125,
0.189453125
],
[
0.330078125,
0.191162109375,
0.67919921875,
0.193603515625,
0.67919921875,
0.2078857421875,
0.330078125,
0.2060546875
],
[
0.38623046875,
0.2060546875,
0.61865234375,
0.2078857421875,
0.61865234375,
0.2257080078125,
0.38623046875,
0.223876953125
],
[
0.264404296875,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.264404296875,
0.279052734375
],
[
0.51123046875,
0.26416015625,
0.8818359375,
0.26416015625,
0.8818359375,
0.279052734375,
0.51123046875,
0.279052734375
],
[
0.1485595703125,
0.287353515625,
0.459228515625,
0.287353515625,
0.459228515625,
0.3017578125,
0.1485595703125,
0.3017578125
],
[
0.51318359375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.51318359375,
0.295654296875
],
[
0.14697265625,
0.3017578125,
0.45751953125,
0.3017578125,
0.45751953125,
0.3154296875,
0.14697265625,
0.3154296875
],
[
0.5146484375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.5146484375,
0.312255859375
],
[
0.019317626953125,
0.3046875,
0.06634521484375,
0.3046875,
0.06634521484375,
0.435302734375,
0.019317626953125,
0.435302734375
],
[
0.214111328125,
0.31640625,
0.459228515625,
0.31640625,
0.459228515625,
0.330078125,
0.214111328125,
0.330078125
],
[
0.51318359375,
0.3154296875,
0.8818359375,
0.3154296875,
0.8818359375,
0.328857421875,
0.51318359375,
0.328857421875
],
[
0.14697265625,
0.330810546875,
0.4609375,
0.330810546875,
0.4609375,
0.344482421875,
0.14697265625,
0.344482421875
],
[
0.1485595703125,
0.317626953125,
0.2166748046875,
0.317626953125,
0.2166748046875,
0.328857421875,
0.1485595703125,
0.328857421875
],
[
0.5146484375,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.5146484375,
0.344482421875
],
[
0.1485595703125,
0.3466796875,
0.45751953125,
0.3466796875,
0.45751953125,
0.35693359375,
0.1485595703125,
0.35693359375
],
[
0.51318359375,
0.3466796875,
0.884765625,
0.3466796875,
0.884765625,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.14697265625,
0.359375,
0.4609375,
0.359375,
0.4609375,
0.372314453125,
0.14697265625,
0.372314453125
],
[
0.5146484375,
0.363525390625,
0.88330078125,
0.363525390625,
0.88330078125,
0.377197265625,
0.5146484375,
0.377197265625
],
[
0.145263671875,
0.372314453125,
0.4609375,
0.371826171875,
0.4609375,
0.385986328125,
0.145263671875,
0.38720703125
],
[
0.5146484375,
0.37890625,
0.88330078125,
0.37890625,
0.88330078125,
0.393798828125,
0.5146484375,
0.393798828125
],
[
0.1485595703125,
0.38720703125,
0.459228515625,
0.38720703125,
0.459228515625,
0.40087890625,
0.1485595703125,
0.40087890625
],
[
0.51318359375,
0.394287109375,
0.8818359375,
0.394287109375,
0.8818359375,
0.407958984375,
0.51318359375,
0.407958984375
],
[
0.14697265625,
0.402099609375,
0.45751953125,
0.402099609375,
0.45751953125,
0.4150390625,
0.14697265625,
0.4150390625
],
[
0.145263671875,
0.41455078125,
0.4609375,
0.4150390625,
0.4609375,
0.4287109375,
0.145263671875,
0.427490234375
],
[
0.5146484375,
0.410888671875,
0.88330078125,
0.410888671875,
0.88330078125,
0.424560546875,
0.5146484375,
0.424560546875
],
[
0.1485595703125,
0.429931640625,
0.459228515625,
0.429931640625,
0.459228515625,
0.443603515625,
0.1485595703125,
0.443603515625
],
[
0.5146484375,
0.427490234375,
0.8818359375,
0.427490234375,
0.8818359375,
0.441162109375,
0.5146484375,
0.441162109375
],
[
0.0218353271484375,
0.435791015625,
0.06463623046875,
0.43701171875,
0.0621337890625,
0.5185546875,
0.019317626953125,
0.51708984375
],
[
0.1485595703125,
0.4453125,
0.459228515625,
0.4453125,
0.459228515625,
0.4560546875,
0.1485595703125,
0.4560546875
],
[
0.5146484375,
0.443603515625,
0.88330078125,
0.443603515625,
0.88330078125,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.1485595703125,
0.460205078125,
0.45751953125,
0.460205078125,
0.45751953125,
0.47021484375,
0.1485595703125,
0.47021484375
],
[
0.51318359375,
0.458984375,
0.8818359375,
0.458984375,
0.8818359375,
0.47265625,
0.51318359375,
0.47265625
],
[
0.14697265625,
0.47265625,
0.4609375,
0.47265625,
0.4609375,
0.486328125,
0.14697265625,
0.486328125
],
[
0.5146484375,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.4892578125,
0.5146484375,
0.4892578125
],
[
0.1485595703125,
0.48681640625,
0.459228515625,
0.48681640625,
0.459228515625,
0.50048828125,
0.1485595703125,
0.50048828125
],
[
0.51123046875,
0.4912109375,
0.55908203125,
0.4912109375,
0.55908203125,
0.505859375,
0.51123046875,
0.505859375
],
[
0.145263671875,
0.49951171875,
0.459228515625,
0.50048828125,
0.459228515625,
0.515625,
0.145263671875,
0.51416015625
],
[
0.5283203125,
0.50830078125,
0.8818359375,
0.50732421875,
0.8818359375,
0.521484375,
0.5283203125,
0.5224609375
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.14697265625,
0.51708984375,
0.45751953125,
0.51708984375,
0.45751953125,
0.52783203125,
0.14697265625,
0.52783203125
],
[
0.51318359375,
0.52392578125,
0.88330078125,
0.52392578125,
0.88330078125,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.1485595703125,
0.53076171875,
0.312255859375,
0.53076171875,
0.312255859375,
0.54150390625,
0.1485595703125,
0.54150390625
],
[
0.51318359375,
0.54052734375,
0.88330078125,
0.54052734375,
0.88330078125,
0.55419921875,
0.51318359375,
0.55419921875
],
[
0.118408203125,
0.55029296875,
0.26025390625,
0.55029296875,
0.26025390625,
0.5654296875,
0.118408203125,
0.5654296875
],
[
0.515625,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.568359375,
0.515625,
0.568359375
],
[
0.12005615234375,
0.57666015625,
0.489501953125,
0.57666015625,
0.489501953125,
0.59033203125,
0.12005615234375,
0.59033203125
],
[
0.51318359375,
0.57177734375,
0.88330078125,
0.572265625,
0.88330078125,
0.58740234375,
0.51318359375,
0.5859375
],
[
0.51123046875,
0.58740234375,
0.88330078125,
0.58837890625,
0.88330078125,
0.6025390625,
0.51123046875,
0.6015625
],
[
0.12005615234375,
0.59326171875,
0.490234375,
0.59326171875,
0.490234375,
0.60693359375,
0.12005615234375,
0.60693359375
],
[
0.12005615234375,
0.609375,
0.490234375,
0.609375,
0.490234375,
0.62255859375,
0.12005615234375,
0.62255859375
],
[
0.51318359375,
0.60498046875,
0.88330078125,
0.60498046875,
0.88330078125,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.118408203125,
0.62451171875,
0.490234375,
0.62451171875,
0.490234375,
0.63818359375,
0.118408203125,
0.63818359375
],
[
0.51318359375,
0.619140625,
0.88330078125,
0.62060546875,
0.88330078125,
0.63525390625,
0.51318359375,
0.63427734375
],
[
0.1175537109375,
0.6396484375,
0.490234375,
0.64111328125,
0.490234375,
0.65478515625,
0.1175537109375,
0.65283203125
],
[
0.51318359375,
0.6357421875,
0.884765625,
0.63525390625,
0.884765625,
0.6494140625,
0.51318359375,
0.65087890625
],
[
0.12005615234375,
0.65673828125,
0.489501953125,
0.65673828125,
0.489501953125,
0.67041015625,
0.12005615234375,
0.67041015625
],
[
0.51318359375,
0.65283203125,
0.88330078125,
0.65283203125,
0.88330078125,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.118408203125,
0.6728515625,
0.489501953125,
0.67138671875,
0.489501953125,
0.68603515625,
0.118408203125,
0.68701171875
],
[
0.51318359375,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.51123046875,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.51123046875,
0.69775390625
],
[
0.12005615234375,
0.6904296875,
0.490234375,
0.6904296875,
0.490234375,
0.70361328125,
0.12005615234375,
0.70361328125
],
[
0.12005615234375,
0.705078125,
0.489501953125,
0.705078125,
0.489501953125,
0.71875,
0.12005615234375,
0.71875
],
[
0.51318359375,
0.70166015625,
0.7021484375,
0.70166015625,
0.7021484375,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.1175537109375,
0.72021484375,
0.47021484375,
0.72021484375,
0.47021484375,
0.73388671875,
0.1175537109375,
0.73388671875
],
[
0.5322265625,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.5322265625,
0.73095703125
],
[
0.1351318359375,
0.736328125,
0.489501953125,
0.73681640625,
0.489501953125,
0.751953125,
0.1351318359375,
0.75048828125
],
[
0.51123046875,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.51123046875,
0.74755859375
],
[
0.118408203125,
0.75439453125,
0.489501953125,
0.75439453125,
0.489501953125,
0.76708984375,
0.118408203125,
0.76708984375
],
[
0.51318359375,
0.75,
0.88330078125,
0.75,
0.88330078125,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.12005615234375,
0.76953125,
0.489501953125,
0.76953125,
0.489501953125,
0.783203125,
0.12005615234375,
0.783203125
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.12176513671875,
0.78759765625,
0.48779296875,
0.78759765625,
0.48779296875,
0.79736328125,
0.12176513671875,
0.79736328125
],
[
0.51318359375,
0.7822265625,
0.8818359375,
0.7822265625,
0.8818359375,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.12005615234375,
0.8017578125,
0.489501953125,
0.8017578125,
0.489501953125,
0.8154296875,
0.12005615234375,
0.8154296875
],
[
0.51123046875,
0.79638671875,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51123046875,
0.81005859375
],
[
0.12005615234375,
0.818359375,
0.489501953125,
0.818359375,
0.489501953125,
0.83203125,
0.12005615234375,
0.83203125
],
[
0.5146484375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.12005615234375,
0.83447265625,
0.489501953125,
0.83447265625,
0.489501953125,
0.84716796875,
0.12005615234375,
0.84716796875
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.85107421875,
0.490234375,
0.85107421875,
0.490234375,
0.8642578125,
0.12005615234375,
0.8642578125
],
[
0.51318359375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.51318359375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.51318359375,
0.876953125
],
[
0.145263671875,
0.87158203125,
0.466796875,
0.87158203125,
0.466796875,
0.88525390625,
0.145263671875,
0.88525390625
],
[
0.466796875,
0.87451171875,
0.48779296875,
0.87451171875,
0.48779296875,
0.8828125,
0.466796875,
0.8828125
],
[
0.51318359375,
0.87744140625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.88427734375,
0.4189453125,
0.88427734375,
0.4189453125,
0.89794921875,
0.118408203125,
0.89794921875
],
[
0.12005615234375,
0.89794921875,
0.420654296875,
0.89794921875,
0.420654296875,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.5146484375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"ABSApp: A Portable Weaklv-Supervised",
"Aspect-Based Sentiment Extraction System",
"Oren Pereg', Daniel Korat', Moshe Wasserblat', Jonathan Mamou', Ido Dagan?",
"1Intel AI Lab, Petah Tikva, Israel",
"2Department of Computer Science, Bar-Ilan University, Ramat Gan, Israel",
"lfirstname.lastname@intel.com",
"2dagan@cs.biu.ac.il",
"Abstract",
"vector space. For example, frequent aspect terms",
"We present ABSApp, a portable system for",
"in the restaurant domain, like food, menu, starters",
"weakly-supervised aspect-based sentiment ex-",
"and salad, have little or no semantic relatedness",
"610z des z1",
"The system is interpretable and",
"to frequent aspect terms in the laptop domain,",
"user friendly and does not require labeled",
"traction'.",
"like screen size, keyboard and battery life. In ad-",
"training data, hence can be rapidly and cost-",
"dition to aspect terms, many opinion terms are",
"effectively used across different domains in",
"also domain-specific. For example, opinion terms",
"applied setups. The system flow includes three",
"like tasty (positive), yummy (positive), flavorful",
"stages: First, it generates domain-specific as-",
"(positive)and tasteless (negative)are specific to",
"pect and opinion lexicons based on an unla-",
"beled dataset; second, it enables the user to",
"the restaurant domain whereas opinion terms like",
"view and edit those lexicons (weak supervi-",
"lightweight (positive), durable (positive), compat-",
"[T's]",
"sion); and finally, it enables the user to select",
"ible (positive) and heavy (negative) are specific to",
"an unlabeled target datasetfrom the same do-",
"the laptop domain. This makes domain-agnostic",
"main, classify it, and generate an aspect-based",
"ABSA a challenging task, with little work address-",
"sentiment report. ABSApp has been success-",
"ing it.",
"fully used in a number of real-life use cases,",
"A recent line of work is based on transfer-",
"1A80900601A0",
"among them movie review analysis and con-",
"learning methods, in which labeled data from a",
"vention impact analysis.",
"source domain is used for training a model to clas-",
"1 Introduction",
"sify datain atarget domain.Ding et al.(2017)and",
"Aspect Based Sentiment Analysis (ABSA) is the",
"Wang and Jialin Pan (2018) proposed using super-",
"vised RNNs for cross-domain aspect term extrac-",
"task of extracting, from a given corpus, opinion",
"targets (aspect terms) and the sentiment expressed",
"tion and for aspect and opinion term co-extraction.",
"towards them. For example, in the sentence “The",
"This approach showed encouraging results, how-",
"dessert was incredible,the aspect termis dessert",
"ever it requires a considerable amount of labeled",
"and the sentiment towards it is positive. This fine-",
"data from the source domain which is often not",
"grained trait of ABSA makes it an effective ap-",
"practical in applied settings due to cost or legal",
"considerations (relevant data is usually not avail-",
"plication for measuring and monitoring the ratio",
"between positive and negative opinions expressed",
"able for commercial use).",
"towards specific aspects of a product or service.",
"Another approach towards domain robustness",
"Most work around ABSA focused on super-",
"is based on unsupervised methods. Hu and Liu",
"vised sequence tagging based systems. Liu et al.",
"(2004) used association rules and Qiu et al. (2011)",
"(2015) showed promising results when the train-",
"used syntactic rules for aspect and opinion term",
"ing and the inference data are from the same do-",
"co-extraction. Industrial setups usually lack la-",
"main. However, this approach is typically not ro-",
"beled data for training and this is where unsuper-",
"bust across different domains since aspect terms",
"vised methods excel. However, these methods can",
"from two different domains are usually semanti-",
"be noisy (see the ABSApp-unsup baseline in Ta-",
"cally different hence separated in the embedding",
"ble 2). In this paper we show that weak super-",
"vision, namely a short manual process of editing",
"'A demo videoof ABSApp is available",
"at",
"lexicons that were generated by an unsupervised",
"https://drive.google.com/open?id=",
"1BLkOxkjIOgyRhNy4UQEFQpDF_KR_NMAd",
"method, produces results that are comparable to"
] |
{
"arxiv_id": "1909.05608",
"language": "en",
"page_idx": 0
}
|
||
1909.05608-en-1
|
1909.05608-en-1.jpg
|
<FigureHere>
Figure 1: ABSApp workflow.
transfer-learning based supervised methods.
The contribution of this paper is twofold. First, it presents ABSApp, a practical weaklysupervised system that does not require labeled data for training, hence can be rapidly and costeffectively used across different domains for producing fine-grained sentiment reports. Second, it introduces a workflow that enables users to weakly-supervise the system, thus enhancing its precision. This workflow enables users to select an unlabeled input dataset from a completely new domain, produce domain-specific aspect and opinion lexicons, and edit the lexicons. The user can then select an unlabeled target dataset from the same domain, classify it, and obtain a detailed report regarding the positive and negative sentiments expressed towards each aspect in the corpus and browse through the results.
Our system has been successfully deployed in several real-life use cases. One of them is the analysis of social media opinions towards specific aspects of movies, preformed in collaboration with a major entertainment content provider. Another use case involves measuring the impact of social events, like conventions and conferences, based on opinions published in social media posts.
The system is distributed as open source software under the Apache license as part of NLP Architect by Intel AI Lab.$^2$
Step 3: Lexicon Editing. Figure 2 shows the aspect and opinion lexicon management (editing) user interface. The user can choose to edit an aspect lexicon or an opinion lexicon that was generated in step 2. As shown in Figure 2, in which the Aspect Lexicon has been selected, the Term column displays the aspect terms while the Aliasl-3 columns display aspect terms that have the same semantic meaning.
Upon selecting a specific aspect, the Examples view on the right-hand side of Figure 2 displays text snippets from the input dataset that include this term (highlighted in blue). The Examples view enables the user to verify that the selected term indeed functions as an aspect term in various contexts in the domain. Based on this, the user can delete (by unchecking the term's checkbox), add or modify the lexicon items. The recommended best practice is to keep relevant aspect terms and delete non-relevant aspect terms. For example, keep terms like 'service' and 'decor' and delete terms like 'time' and'city' from an aspect lexicon related to restaurant reviews.
In addition, the user can group together synonym aspects like 'drinks' and 'beverages' (see Figure 2). Finally, the user can save the edited lexicon. The opinion lexicon editor (not shown) functions similarly to the aspect lexicon editor except that it includes a Polarity column and a Score column (see Section 3.1) instead of the Alias columns. Both the polarity and the score can be edited by the user.
Steps 4 &5: Selecting a Target Dataset and
## 2 $\textbf{System Workflow}$
This section describes the workflow of ABSApp,
as depicted in Figure 1.
Performing Sentiment Classification. A target dataset and its classification are performed by
clicking the Classify button in Figure 2. Once the dataset is selected the system starts the sentiment classification process (see Section 3.2).
Steps 1 & 2: Selecting a Dataset and Extracting Lexicons. The first step of the flow is to select an input dataset for lexicon extraction, performed by clicking the Extract lexicons button shown in Figure 2. Once a dataset is selected, the system performs the lexicon extraction process. This step extracts aspect terms and produces an aspect lexicon. In addition, this step extracts candidate opinion terms, flters them and estimates their polarity, producing an opinion lexicon (see Section 3.1).
Step 6: Results Visualization. Figure 3 shows the output of the sentiment classification process. The upper part of the fgure displays the count of positive and negative sentiment mentions detected in the target dataset towards each aspect, as green and red bars. Hovering over a green(red) bar displays the count of the positive(negative) senti-
${^{2}\mathrm{http://n}}$lp\_architect.nervanasys.com/
absa.html
|
[
[
0.1376953125,
0.08197021484375,
0.235107421875,
0.08197021484375,
0.235107421875,
0.09320068359375,
0.1376953125,
0.09320068359375
],
[
0.28955078125,
0.08197021484375,
0.339111328125,
0.08197021484375,
0.339111328125,
0.09320068359375,
0.28955078125,
0.09320068359375
],
[
0.39794921875,
0.08197021484375,
0.48193359375,
0.08197021484375,
0.48193359375,
0.09320068359375,
0.39794921875,
0.09320068359375
],
[
0.66748046875,
0.08197021484375,
0.71875,
0.08197021484375,
0.71875,
0.09320068359375,
0.66748046875,
0.09320068359375
],
[
0.78857421875,
0.08197021484375,
0.8486328125,
0.08197021484375,
0.8486328125,
0.09320068359375,
0.78857421875,
0.09320068359375
],
[
0.5283203125,
0.09320068359375,
0.60546875,
0.0955810546875,
0.60546875,
0.10986328125,
0.5283203125,
0.10809326171875
],
[
0.52490234375,
0.0831298828125,
0.609375,
0.0831298828125,
0.609375,
0.09442138671875,
0.52490234375,
0.09442138671875
],
[
0.17626953125,
0.0955810546875,
0.198974609375,
0.0955810546875,
0.198974609375,
0.10986328125,
0.17626953125,
0.10986328125
],
[
0.273681640625,
0.0955810546875,
0.3525390625,
0.0955810546875,
0.3525390625,
0.10687255859375,
0.273681640625,
0.10687255859375
],
[
0.393798828125,
0.0968017578125,
0.483642578125,
0.0968017578125,
0.483642578125,
0.10809326171875,
0.393798828125,
0.10809326171875
],
[
0.64404296875,
0.0968017578125,
0.73876953125,
0.0968017578125,
0.73876953125,
0.10809326171875,
0.64404296875,
0.10809326171875
],
[
0.78515625,
0.0968017578125,
0.85205078125,
0.0968017578125,
0.85205078125,
0.10809326171875,
0.78515625,
0.10809326171875
],
[
0.4013671875,
0.127685546875,
0.60107421875,
0.127685546875,
0.60107421875,
0.142578125,
0.4013671875,
0.142578125
],
[
0.118408203125,
0.168701171875,
0.441650390625,
0.168701171875,
0.441650390625,
0.1817626953125,
0.118408203125,
0.1817626953125
],
[
0.51318359375,
0.168701171875,
0.88330078125,
0.168701171875,
0.88330078125,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.1368408203125,
0.185302734375,
0.48779296875,
0.185302734375,
0.48779296875,
0.198974609375,
0.1368408203125,
0.198974609375
],
[
0.51318359375,
0.185302734375,
0.8818359375,
0.185302734375,
0.8818359375,
0.198974609375,
0.51318359375,
0.198974609375
],
[
0.12005615234375,
0.20068359375,
0.489501953125,
0.20068359375,
0.489501953125,
0.215576171875,
0.12005615234375,
0.215576171875
],
[
0.51123046875,
0.20068359375,
0.87890625,
0.20068359375,
0.87890625,
0.21435546875,
0.51123046875,
0.21435546875
],
[
0.12005615234375,
0.2161865234375,
0.489501953125,
0.2161865234375,
0.489501953125,
0.2298583984375,
0.12005615234375,
0.2298583984375
],
[
0.5146484375,
0.218505859375,
0.87890625,
0.218505859375,
0.87890625,
0.2286376953125,
0.5146484375,
0.2286376953125
],
[
0.12005615234375,
0.2327880859375,
0.489501953125,
0.2327880859375,
0.489501953125,
0.2464599609375,
0.12005615234375,
0.2464599609375
],
[
0.51318359375,
0.2327880859375,
0.88330078125,
0.2327880859375,
0.88330078125,
0.2464599609375,
0.51318359375,
0.2464599609375
],
[
0.118408203125,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.51123046875,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.262451171875,
0.51123046875,
0.262451171875
],
[
0.118408203125,
0.265380859375,
0.489501953125,
0.265380859375,
0.489501953125,
0.279052734375,
0.118408203125,
0.279052734375
],
[
0.51123046875,
0.26416015625,
0.88330078125,
0.26318359375,
0.88330078125,
0.27783203125,
0.51123046875,
0.279052734375
],
[
0.118408203125,
0.283203125,
0.489501953125,
0.283203125,
0.489501953125,
0.29345703125,
0.118408203125,
0.29345703125
],
[
0.51123046875,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.51123046875,
0.295654296875
],
[
0.118408203125,
0.297607421875,
0.489501953125,
0.297607421875,
0.489501953125,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51123046875,
0.297607421875,
0.654296875,
0.297607421875,
0.654296875,
0.311279296875,
0.51123046875,
0.311279296875
],
[
0.118408203125,
0.3134765625,
0.490234375,
0.3134765625,
0.490234375,
0.327880859375,
0.118408203125,
0.327880859375
],
[
0.53076171875,
0.312255859375,
0.88330078125,
0.312255859375,
0.88330078125,
0.32666015625,
0.53076171875,
0.32666015625
],
[
0.1175537109375,
0.328857421875,
0.48779296875,
0.328857421875,
0.48779296875,
0.342529296875,
0.1175537109375,
0.342529296875
],
[
0.51318359375,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.1175537109375,
0.344482421875,
0.490234375,
0.345703125,
0.490234375,
0.35986328125,
0.1175537109375,
0.359375
],
[
0.51123046875,
0.345703125,
0.884765625,
0.344482421875,
0.884765625,
0.359375,
0.51123046875,
0.35986328125
],
[
0.12176513671875,
0.363525390625,
0.490234375,
0.363525390625,
0.490234375,
0.37353515625,
0.12176513671875,
0.37353515625
],
[
0.51318359375,
0.3623046875,
0.884765625,
0.3623046875,
0.884765625,
0.3759765625,
0.51318359375,
0.3759765625
],
[
0.12005615234375,
0.377685546875,
0.490234375,
0.377685546875,
0.490234375,
0.391357421875,
0.12005615234375,
0.391357421875
],
[
0.51318359375,
0.377685546875,
0.88330078125,
0.377685546875,
0.88330078125,
0.391357421875,
0.51318359375,
0.391357421875
],
[
0.12176513671875,
0.3955078125,
0.48779296875,
0.3955078125,
0.48779296875,
0.40625,
0.12176513671875,
0.40625
],
[
0.51318359375,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.40673828125,
0.51318359375,
0.40673828125
],
[
0.12005615234375,
0.410400390625,
0.489501953125,
0.410400390625,
0.489501953125,
0.42333984375,
0.12005615234375,
0.42333984375
],
[
0.51123046875,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.12005615234375,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.440673828125,
0.12005615234375,
0.440673828125
],
[
0.51318359375,
0.42578125,
0.8818359375,
0.42578125,
0.8818359375,
0.439453125,
0.51318359375,
0.439453125
],
[
0.1175537109375,
0.441162109375,
0.316650390625,
0.441162109375,
0.316650390625,
0.454833984375,
0.1175537109375,
0.454833984375
],
[
0.51318359375,
0.443603515625,
0.880859375,
0.443603515625,
0.880859375,
0.45361328125,
0.51318359375,
0.45361328125
],
[
0.1368408203125,
0.457763671875,
0.490234375,
0.457763671875,
0.490234375,
0.471435546875,
0.1368408203125,
0.471435546875
],
[
0.51318359375,
0.457763671875,
0.884765625,
0.457763671875,
0.884765625,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.118408203125,
0.474365234375,
0.48779296875,
0.474365234375,
0.48779296875,
0.488037109375,
0.118408203125,
0.488037109375
],
[
0.51318359375,
0.474365234375,
0.88330078125,
0.474365234375,
0.88330078125,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.1175537109375,
0.490478515625,
0.489501953125,
0.4892578125,
0.489501953125,
0.50439453125,
0.1175537109375,
0.5048828125
],
[
0.51318359375,
0.490478515625,
0.88330078125,
0.490478515625,
0.88330078125,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.12005615234375,
0.50732421875,
0.490234375,
0.50732421875,
0.490234375,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.51123046875,
0.5048828125,
0.884765625,
0.505859375,
0.884765625,
0.52099609375,
0.51123046875,
0.51953125
],
[
0.1175537109375,
0.5224609375,
0.490234375,
0.521484375,
0.490234375,
0.5361328125,
0.1175537109375,
0.53759765625
],
[
0.51318359375,
0.5224609375,
0.7900390625,
0.5224609375,
0.7900390625,
0.5361328125,
0.51318359375,
0.5361328125
],
[
0.118408203125,
0.5390625,
0.489501953125,
0.5390625,
0.489501953125,
0.552734375,
0.118408203125,
0.552734375
],
[
0.5283203125,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.5283203125,
0.5517578125
],
[
0.118408203125,
0.5546875,
0.490234375,
0.5546875,
0.490234375,
0.568359375,
0.118408203125,
0.568359375
],
[
0.51123046875,
0.55419921875,
0.88330078125,
0.55419921875,
0.88330078125,
0.56787109375,
0.51123046875,
0.56787109375
],
[
0.118408203125,
0.57177734375,
0.4248046875,
0.57177734375,
0.4248046875,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.51318359375,
0.57080078125,
0.884765625,
0.57080078125,
0.884765625,
0.58447265625,
0.51318359375,
0.58447265625
],
[
0.1368408203125,
0.58740234375,
0.490234375,
0.58740234375,
0.490234375,
0.6015625,
0.1368408203125,
0.6015625
],
[
0.5146484375,
0.58837890625,
0.8818359375,
0.58837890625,
0.8818359375,
0.5986328125,
0.5146484375,
0.5986328125
],
[
0.1175537109375,
0.6025390625,
0.490234375,
0.6015625,
0.490234375,
0.6162109375,
0.1175537109375,
0.61767578125
],
[
0.515625,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.615234375,
0.515625,
0.615234375
],
[
0.1175537109375,
0.6181640625,
0.301513671875,
0.61767578125,
0.301513671875,
0.6318359375,
0.1175537109375,
0.6328125
],
[
0.5146484375,
0.619140625,
0.884765625,
0.619140625,
0.884765625,
0.6328125,
0.5146484375,
0.6328125
],
[
0.5146484375,
0.63720703125,
0.8818359375,
0.63720703125,
0.8818359375,
0.64794921875,
0.5146484375,
0.64794921875
],
[
0.118408203125,
0.6455078125,
0.300537109375,
0.6455078125,
0.300537109375,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.65185546875,
0.88330078125,
0.65185546875,
0.88330078125,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.118408203125,
0.67138671875,
0.489501953125,
0.6728515625,
0.489501953125,
0.68701171875,
0.118408203125,
0.68603515625
],
[
0.51123046875,
0.66650390625,
0.6513671875,
0.66748046875,
0.6513671875,
0.681640625,
0.51123046875,
0.68115234375
],
[
0.118408203125,
0.689453125,
0.2939453125,
0.689453125,
0.2939453125,
0.70361328125,
0.118408203125,
0.70361328125
],
[
0.51318359375,
0.6923828125,
0.88330078125,
0.6923828125,
0.88330078125,
0.7060546875,
0.51318359375,
0.7060546875
],
[
0.118408203125,
0.71435546875,
0.489501953125,
0.71435546875,
0.489501953125,
0.728515625,
0.118408203125,
0.728515625
],
[
0.5146484375,
0.708984375,
0.88330078125,
0.708984375,
0.88330078125,
0.72265625,
0.5146484375,
0.72265625
],
[
0.51123046875,
0.7236328125,
0.8818359375,
0.724609375,
0.8818359375,
0.7392578125,
0.51123046875,
0.73828125
],
[
0.118408203125,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51318359375,
0.7412109375,
0.88330078125,
0.7412109375,
0.88330078125,
0.7548828125,
0.51318359375,
0.7548828125
],
[
0.118408203125,
0.74755859375,
0.490234375,
0.74755859375,
0.490234375,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.1175537109375,
0.7626953125,
0.489501953125,
0.7626953125,
0.489501953125,
0.775390625,
0.1175537109375,
0.775390625
],
[
0.5146484375,
0.7587890625,
0.88330078125,
0.7587890625,
0.88330078125,
0.76953125,
0.5146484375,
0.76953125
],
[
0.118408203125,
0.779296875,
0.490234375,
0.779296875,
0.490234375,
0.79345703125,
0.118408203125,
0.79345703125
],
[
0.51318359375,
0.77392578125,
0.8037109375,
0.77392578125,
0.8037109375,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.118408203125,
0.7958984375,
0.490234375,
0.7958984375,
0.490234375,
0.80859375,
0.118408203125,
0.80859375
],
[
0.51318359375,
0.79736328125,
0.88330078125,
0.79736328125,
0.88330078125,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.12005615234375,
0.8125,
0.489501953125,
0.8125,
0.489501953125,
0.826171875,
0.12005615234375,
0.826171875
],
[
0.51318359375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.118408203125,
0.82763671875,
0.489501953125,
0.82763671875,
0.489501953125,
0.84130859375,
0.118408203125,
0.84130859375
],
[
0.51318359375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.845703125,
0.51318359375,
0.845703125
],
[
0.1175537109375,
0.8427734375,
0.489501953125,
0.84326171875,
0.489501953125,
0.85791015625,
0.1175537109375,
0.85693359375
],
[
0.51123046875,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.118408203125,
0.85986328125,
0.465087890625,
0.85986328125,
0.465087890625,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1427001953125,
0.8798828125,
0.480224609375,
0.8828125,
0.480224609375,
0.90087890625,
0.1427001953125,
0.89794921875
],
[
0.5146484375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.12005615234375,
0.89794921875,
0.202392578125,
0.89794921875,
0.202392578125,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"Select dataset",
"Extract",
"Editlexicons",
"Classify",
"Visualize",
"dataset (4)",
"Selecttarget",
"(1)",
"Lexicons (2)",
"(optional)(3)",
"sentiment (5)",
"results (6)",
"Figurel:ABSApp workflow.",
"transfer-learning based supervised methods.",
"Step 3: Lexicon Editing. Figure 2 shows the",
"The contribution of this paper is twofold.",
"aspect and opinion lexicon management (editing)",
"First, it presents ABSApp, a practical weakly-",
"user interface. The user can choose to edit an as",
"supervised system that does not require labeled",
"pectlexiconor an opinionlexiconthat was gener",
"data for training, hence can be rapidly and cost-",
"ated in step 2. As shown in Figure 2, in which the",
"effectively used across different domains for pro-",
"Aspect Lexicon has been selected, the Term col-",
"ducing fine-grained sentiment reports. Second,",
"umn displays the aspect terms while the Alias1-3",
"it introduces a workflow that enables users to",
"columns display aspect terms that have the same",
"weakly-supervise the system, thus enhancing its",
" semantic meaning.",
"precision. This workflow enables users to select",
"Upon selecting a specific aspect, the Examples",
"an unlabeled input dataset from a completely new",
"view on the right-hand side of Figure 2 displays",
"domain, produce domain-specific aspect and opin-",
"text snippets from the input dataset that include",
"ionlexicons.and editthelexicons.Theusercan",
"this term (highlighted in blue). The Examples",
"then select an unlabeled target dataset from the",
"view enables the user to verify that the selected",
"samedomain,classify it,and obtaina detailedre-",
"term indeed functions as an aspect term in var-",
"port regarding the positive and negative sentiments",
"ious contexts in the domain. Based on this, the",
"expressed towards each aspect in the corpus and",
"user can delete (by unchecking the term's check-",
"browse through the results.",
"box),add ormodifythelexiconitems..Therec-",
"Our system has been successfully deployed in",
"ommended best practice is to keep relevant aspect",
"several real-life use cases. One of them is the anal-",
"terms and delete non-relevant aspect terms. For",
"ysis of social media opinions towards specific as-",
"example, keep terms like'service'and'decor'and",
"pects of movies, preformed in collaboration with",
"delete terms like 'time' and 'city' from an aspect",
"a major entertainment content provider. Another",
"lexicon related to restaurant reviews.",
"use case involves measuring the impact of social",
"In addition, the user can group together syn-",
"events, like conventions and conferences, based on",
"onym aspects like 'drinks' and \"beverages(see",
"opinions published in social media posts.",
"Figure 2). Finally, the user can save the edited",
"The system is distributed as open source soft-",
"lexicon.The opinion lexicon editor (not shown)",
"ware under the Apache license as part of NLP Ar-",
"functions similarly to the aspect lexicon editor",
"chitect by Intel AI Lab.?",
"except that it includes a Polarity column and a",
"Scorecolumn(seeSection3.l)insteadoftheAlias",
"2 System Workflow",
"columns. Both the polarity and the score can be",
"This section describes the workflow of ABSApp,",
"edited by the user.",
"as depicted in Figure 1.",
"Steps 4 & 5: Selecting a Target Dataset and",
"Steps 1 & 2: Selecting a Dataset and Extracting",
"Performing Sentiment Classification.A target",
"dataset and its classification are performed by",
"Lexicons.The first step of the flow is to select",
"clicking the Classify button in Figure 2. Once the",
"an input dataset for lexicon extraction, performed",
"by clicking the Extract lexicons button shown in",
"dataset is selected the system starts the sentiment",
"Figure 2. Once a dataset is selected, the system",
"classification process (see Section 3.2).",
"performs the lexicon extraction process. This step",
"Step 6: Results Visualization. Figure 3 shows",
"extracts aspect terms and produces an aspect lexi-",
"the output of the sentiment classification process.",
"con. In addition, this step extracts candidate opin-",
"The upper part of the figure displays the count",
"ion terms, filters them and estimates their polarity,",
"of positive and negative sentiment mentions de-",
"producing an opinion lexicon (see Section 3.1)",
"tected in the target dataset towards each aspect, as",
"2http://nlp_architect.nervanasys.com/",
"green and red bars. Hovering over a green(red) bar",
"absa.html",
"displays the count of the positive(negative) senti-"
] |
{
"arxiv_id": "1909.05608",
"language": "en",
"page_idx": 1
}
|
||
1909.05608-en-3
|
1909.05608-en-3.jpg
|
known aspect term food.
<table>
<tbody>
<tr>
<th>$\mathbf{RuleID}$</th>
<th>Rule</th>
<th>$\mathbf{Example}$</th>
</tr>
<tr>
<td>R1</td>
<td>0 amod $A(NN)$</td>
<td>nice decor amod</td>
</tr>
<tr>
<td>$\mathbb{R}2$</td>
<td>$\underline{nsubj}.$ 4 O( ADJ) </td>
<td>the food d was super tasty nsubj (food tasty)</td>
</tr>
</tbody>
</table>
Table 1: Examples of the opinion and aspect terms extraction rules. $O$ represents an opinion term and $A$ represents an aspect term.
Opinion Lexicon Scoring and Filtering. This step aims to filter the noisy candidate opinion terms extracted by the bootstrap process. It is based on using an MLP classifer for generating a score that represents the probability that a candidate is indeed an opinion term. Candidate opinion terms are qualified as opinion terms if their classification score exceeds a threshold $^{4}.$
The MLP classifer input features consists of the candidate term word embedding$^{5}$ and the mean, standard-deviation, maximum and minimum word-embedding cosine similarities between the candidate term and a pre-determined set of generic opinion terms. The MLP consists of a single hidden layer and is trained once for a binary classification task using manually labeled data that consists of a set of opinion terms (positive class) and a set of non-opinion terms (negative class) from a specific domain$^6.$ Once the model is generated it is then used for grading candidate opinion terms extracted in other domains. It is reasonable to use such model across domains, since the classification features represent semantic similarity levels that are robust across domains.
Opinion Polarity Estimation The goal of this step is to set the binary sentiment polarity (positive or negative) of the opinion terms. Following Pablos et al. (2016), an opinion term polarity is assigned based on estimating whether it is semantically closer to a set of generic positive opinion terms or to a set of generic negative opinion terms. To produce those sets we used a subset of 47 positive terms and a subset of 47 negative terms de-
$^{4}$The threshold's value was empirically determined based
on precision-recall tradeoff.
$^{5}$We used Stanford Glove embeddings https://nlp.
stanford.edu/projects/glove/
$^{6}$This training data can be downloaded from https:// github.com/NervanaSystems/nlp-architect/ blob/master/nlp\_architect/models/absa/ train/lexicons/RerankTrainingData.csv
rived from the opinion lexicon generated by Hu and Liu (2004). The semantic similarity between an opinion term and a positive or negative set is estimated by averaging the cosine similarity of the embedding of the opinion term and the embedding of each one of the terms in the positive or negative set.
In this module we used pre-trained embeddings$^{5}$ which produce overall good results but raise a drawback; some opinion terms may convey different sentiment polarities in different domains (e.g. 'delicate movie’(positive) vs. 'delicate cellphone’(negative)), while a pre-trained embeddings setup is only capable of setting a single polarity per opinion term. A suggested solution is to adapt the embeddings to the target domain or to use context embeddings. We intend to address this challenge in future work.
## 3.2 Sentiment Classification
Sentiment classification uses the opinion and aspect lexicons for detecting aspect-opinion pairs (sentiment mentions) within the input target dataset, and determining their sentiment polarity. Aspect-opinion pairs are extracted based on detecting a direct or second-order dependency relation of any type, between them. The aspectopinion pair polarity is assigned according to the polarity of the opinion term. This step also uses a pre-determined negation lexicon containing negation terms. Upon detecting a direct dependency relation between a negation term and the opinion term, the aspect-opinion pair polarity is reversed.
## 4 Evaluation
Our evaluation objective is to show that the different algorithmic steps, namely, lexicon extraction and sentiment classification, produce usable results. An additional objective is to show that weak supervision of an aspect lexicon that was generated in an unsupervised manner produces comparable results to the recent transfer-learning based methods (Ding et al., 2017; Wang and Jialin Pan, 2018). For this purpose, we leveraged the data of SemEval 2014 task4 (Pontiki et al., 2014), which tests the two main ABSA sub-tasks: aspect term extraction and aspect term polarity detection.
Datasets. The performance of the algorithm was evaluated using data from two different domains: restaurant reviews and laptop reviews. Those two domains are disjoint and therefore demonstrate
|
[
[
0.118408203125,
0.07659912109375,
0.300537109375,
0.07659912109375,
0.300537109375,
0.0914306640625,
0.118408203125,
0.0914306640625
],
[
0.51318359375,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.51318359375,
0.0914306640625,
0.884765625,
0.0926513671875,
0.884765625,
0.10687255859375,
0.51318359375,
0.105712890625
],
[
0.2359619140625,
0.1092529296875,
0.273681640625,
0.1092529296875,
0.273681640625,
0.12469482421875,
0.2359619140625,
0.12469482421875
],
[
0.51318359375,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.51318359375,
0.122314453125
],
[
0.370361328125,
0.10986328125,
0.43408203125,
0.10986328125,
0.43408203125,
0.12469482421875,
0.370361328125,
0.12469482421875
],
[
0.1275634765625,
0.10986328125,
0.183837890625,
0.10986328125,
0.183837890625,
0.12469482421875,
0.1275634765625,
0.12469482421875
],
[
0.197265625,
0.1270751953125,
0.30908203125,
0.12890625,
0.30908203125,
0.1478271484375,
0.197265625,
0.1455078125
],
[
0.365234375,
0.1312255859375,
0.4375,
0.1318359375,
0.4375,
0.146728515625,
0.365234375,
0.1455078125
],
[
0.51318359375,
0.1258544921875,
0.884765625,
0.1258544921875,
0.884765625,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.1292724609375,
0.1318359375,
0.15283203125,
0.1318359375,
0.15283203125,
0.146728515625,
0.1292724609375,
0.146728515625
],
[
0.3720703125,
0.1395263671875,
0.465087890625,
0.1484375,
0.4609375,
0.166259765625,
0.366943359375,
0.15673828125
],
[
0.51318359375,
0.141357421875,
0.88330078125,
0.141357421875,
0.88330078125,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.341796875,
0.149658203125,
0.3779296875,
0.149658203125,
0.3779296875,
0.160888671875,
0.341796875,
0.160888671875
],
[
0.2476806640625,
0.162109375,
0.316650390625,
0.1644287109375,
0.316650390625,
0.1787109375,
0.2476806640625,
0.177001953125
],
[
0.51318359375,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.130126953125,
0.1644287109375,
0.154541015625,
0.1644287109375,
0.154541015625,
0.1776123046875,
0.130126953125,
0.1776123046875
],
[
0.2166748046875,
0.163330078125,
0.2509765625,
0.163330078125,
0.2509765625,
0.17041015625,
0.2166748046875,
0.17041015625
],
[
0.32421875,
0.1644287109375,
0.47607421875,
0.1644287109375,
0.47607421875,
0.1776123046875,
0.32421875,
0.1776123046875
],
[
0.5146484375,
0.174560546875,
0.5439453125,
0.174560546875,
0.5439453125,
0.1859130859375,
0.5146484375,
0.1859130859375
],
[
0.202392578125,
0.1650390625,
0.21826171875,
0.168701171875,
0.2149658203125,
0.17578125,
0.198974609375,
0.1728515625
],
[
0.339111328125,
0.177001953125,
0.46337890625,
0.177001953125,
0.46337890625,
0.1953125,
0.339111328125,
0.1953125
],
[
0.53076171875,
0.188232421875,
0.88330078125,
0.188232421875,
0.88330078125,
0.201904296875,
0.53076171875,
0.201904296875
],
[
0.1175537109375,
0.207275390625,
0.489501953125,
0.2078857421875,
0.489501953125,
0.22265625,
0.1175537109375,
0.221435546875
],
[
0.51318359375,
0.2037353515625,
0.88671875,
0.2037353515625,
0.88671875,
0.218505859375,
0.51318359375,
0.218505859375
],
[
0.118408203125,
0.223876953125,
0.489501953125,
0.223876953125,
0.489501953125,
0.2369384765625,
0.118408203125,
0.2369384765625
],
[
0.51318359375,
0.2203369140625,
0.884765625,
0.2203369140625,
0.884765625,
0.2340087890625,
0.51318359375,
0.2340087890625
],
[
0.1175537109375,
0.2381591796875,
0.274658203125,
0.2381591796875,
0.274658203125,
0.251708984375,
0.1175537109375,
0.251708984375
],
[
0.5146484375,
0.2369384765625,
0.88330078125,
0.2369384765625,
0.88330078125,
0.25048828125,
0.5146484375,
0.25048828125
],
[
0.5146484375,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.2666015625,
0.5146484375,
0.2666015625
],
[
0.12005615234375,
0.268310546875,
0.443359375,
0.268310546875,
0.443359375,
0.281982421875,
0.12005615234375,
0.281982421875
],
[
0.435791015625,
0.26953125,
0.48779296875,
0.26953125,
0.48779296875,
0.28076171875,
0.435791015625,
0.28076171875
],
[
0.51318359375,
0.268310546875,
0.88330078125,
0.268310546875,
0.88330078125,
0.281982421875,
0.51318359375,
0.281982421875
],
[
0.118408203125,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.118408203125,
0.297607421875
],
[
0.51318359375,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.51318359375,
0.298583984375
],
[
0.12005615234375,
0.3017578125,
0.490234375,
0.3017578125,
0.490234375,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.51318359375,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.314208984375,
0.51318359375,
0.314208984375
],
[
0.118408203125,
0.31640625,
0.489501953125,
0.31640625,
0.489501953125,
0.330810546875,
0.118408203125,
0.330810546875
],
[
0.51318359375,
0.317626953125,
0.880859375,
0.317626953125,
0.880859375,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.12005615234375,
0.3349609375,
0.48779296875,
0.3349609375,
0.48779296875,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.51318359375,
0.333251953125,
0.88671875,
0.333251953125,
0.88671875,
0.3466796875,
0.51318359375,
0.3466796875
],
[
0.12005615234375,
0.349853515625,
0.490234375,
0.349853515625,
0.490234375,
0.363525390625,
0.12005615234375,
0.363525390625
],
[
0.51318359375,
0.349853515625,
0.791015625,
0.349853515625,
0.791015625,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.118408203125,
0.364013671875,
0.486083984375,
0.364013671875,
0.486083984375,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.51318359375,
0.374755859375,
0.748046875,
0.3759765625,
0.748046875,
0.39013671875,
0.51318359375,
0.3896484375
],
[
0.118408203125,
0.38134765625,
0.385498046875,
0.38134765625,
0.385498046875,
0.394287109375,
0.118408203125,
0.394287109375
],
[
0.1368408203125,
0.3955078125,
0.490234375,
0.3955078125,
0.490234375,
0.4091796875,
0.1368408203125,
0.4091796875
],
[
0.51123046875,
0.3955078125,
0.88330078125,
0.396728515625,
0.88330078125,
0.410888671875,
0.51123046875,
0.410400390625
],
[
0.12005615234375,
0.413330078125,
0.489501953125,
0.413330078125,
0.489501953125,
0.42333984375,
0.12005615234375,
0.42333984375
],
[
0.51123046875,
0.413330078125,
0.88330078125,
0.413330078125,
0.88330078125,
0.427001953125,
0.51123046875,
0.427001953125
],
[
0.1158447265625,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.441162109375,
0.1158447265625,
0.441162109375
],
[
0.51123046875,
0.427490234375,
0.88671875,
0.4287109375,
0.88671875,
0.443603515625,
0.51123046875,
0.4423828125
],
[
0.118408203125,
0.44482421875,
0.490234375,
0.44482421875,
0.490234375,
0.457763671875,
0.118408203125,
0.457763671875
],
[
0.51318359375,
0.4453125,
0.88330078125,
0.4453125,
0.88330078125,
0.460205078125,
0.51318359375,
0.460205078125
],
[
0.12005615234375,
0.46142578125,
0.490234375,
0.46142578125,
0.490234375,
0.474365234375,
0.12005615234375,
0.474365234375
],
[
0.51318359375,
0.4619140625,
0.88330078125,
0.4619140625,
0.88330078125,
0.4755859375,
0.51318359375,
0.4755859375
],
[
0.12005615234375,
0.476806640625,
0.489501953125,
0.476806640625,
0.489501953125,
0.490478515625,
0.12005615234375,
0.490478515625
],
[
0.51318359375,
0.47802734375,
0.88330078125,
0.47802734375,
0.88330078125,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.118408203125,
0.4921875,
0.486083984375,
0.4921875,
0.486083984375,
0.505859375,
0.118408203125,
0.505859375
],
[
0.51318359375,
0.493408203125,
0.88330078125,
0.493408203125,
0.88330078125,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.12005615234375,
0.51025390625,
0.489501953125,
0.51025390625,
0.489501953125,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.51123046875,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.51123046875,
0.52392578125
],
[
0.118408203125,
0.52490234375,
0.489501953125,
0.52490234375,
0.489501953125,
0.5380859375,
0.118408203125,
0.5380859375
],
[
0.51318359375,
0.525390625,
0.884765625,
0.525390625,
0.884765625,
0.5390625,
0.51318359375,
0.5390625
],
[
0.12005615234375,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5546875,
0.12005615234375,
0.5546875
],
[
0.5146484375,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.5146484375,
0.5556640625
],
[
0.118408203125,
0.55712890625,
0.489501953125,
0.55712890625,
0.489501953125,
0.57080078125,
0.118408203125,
0.57080078125
],
[
0.51318359375,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.12005615234375,
0.572265625,
0.490234375,
0.572265625,
0.490234375,
0.5859375,
0.12005615234375,
0.5859375
],
[
0.51318359375,
0.57470703125,
0.88330078125,
0.57470703125,
0.88330078125,
0.58837890625,
0.51318359375,
0.58837890625
],
[
0.118408203125,
0.5888671875,
0.489501953125,
0.5888671875,
0.489501953125,
0.6025390625,
0.118408203125,
0.6025390625
],
[
0.51318359375,
0.59033203125,
0.875,
0.59033203125,
0.875,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.12005615234375,
0.60693359375,
0.48779296875,
0.60693359375,
0.48779296875,
0.61767578125,
0.12005615234375,
0.61767578125
],
[
0.118408203125,
0.62158203125,
0.49365234375,
0.62158203125,
0.49365234375,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.51318359375,
0.62255859375,
0.52490234375,
0.619140625,
0.52880859375,
0.62744140625,
0.51708984375,
0.63134765625
],
[
0.53564453125,
0.62255859375,
0.54150390625,
0.62255859375,
0.54150390625,
0.62744140625,
0.53564453125,
0.62744140625
],
[
0.5380859375,
0.619140625,
0.6396484375,
0.619140625,
0.6396484375,
0.63134765625,
0.5380859375,
0.63134765625
],
[
0.118408203125,
0.63818359375,
0.376220703125,
0.63818359375,
0.376220703125,
0.65185546875,
0.118408203125,
0.65185546875
],
[
0.51123046875,
0.642578125,
0.880859375,
0.642578125,
0.880859375,
0.65625,
0.51123046875,
0.65625
],
[
0.1175537109375,
0.65771484375,
0.490234375,
0.6591796875,
0.490234375,
0.67333984375,
0.1175537109375,
0.6728515625
],
[
0.51318359375,
0.66015625,
0.884765625,
0.66015625,
0.884765625,
0.67333984375,
0.51318359375,
0.67333984375
],
[
0.118408203125,
0.6767578125,
0.489501953125,
0.6767578125,
0.489501953125,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.51318359375,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.1175537109375,
0.6923828125,
0.489501953125,
0.6923828125,
0.489501953125,
0.7060546875,
0.1175537109375,
0.7060546875
],
[
0.51318359375,
0.6923828125,
0.88330078125,
0.6923828125,
0.88330078125,
0.7060546875,
0.51318359375,
0.7060546875
],
[
0.12005615234375,
0.7080078125,
0.490234375,
0.7080078125,
0.490234375,
0.7216796875,
0.12005615234375,
0.7216796875
],
[
0.51318359375,
0.708984375,
0.88330078125,
0.708984375,
0.88330078125,
0.72265625,
0.51318359375,
0.72265625
],
[
0.118408203125,
0.724609375,
0.489501953125,
0.724609375,
0.489501953125,
0.73828125,
0.118408203125,
0.73828125
],
[
0.51123046875,
0.7236328125,
0.8818359375,
0.7236328125,
0.8818359375,
0.73681640625,
0.51123046875,
0.73681640625
],
[
0.118408203125,
0.74072265625,
0.490234375,
0.74072265625,
0.490234375,
0.75439453125,
0.118408203125,
0.75439453125
],
[
0.51318359375,
0.7412109375,
0.884765625,
0.7412109375,
0.884765625,
0.7548828125,
0.51318359375,
0.7548828125
],
[
0.118408203125,
0.75732421875,
0.489501953125,
0.75732421875,
0.489501953125,
0.77099609375,
0.118408203125,
0.77099609375
],
[
0.51318359375,
0.755859375,
0.8818359375,
0.755859375,
0.8818359375,
0.76953125,
0.51318359375,
0.76953125
],
[
0.118408203125,
0.7724609375,
0.490234375,
0.7724609375,
0.490234375,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.51123046875,
0.771484375,
0.884765625,
0.7724609375,
0.884765625,
0.78759765625,
0.51123046875,
0.7861328125
],
[
0.118408203125,
0.7880859375,
0.489501953125,
0.7880859375,
0.489501953125,
0.802734375,
0.118408203125,
0.802734375
],
[
0.515625,
0.79052734375,
0.88330078125,
0.79052734375,
0.88330078125,
0.80029296875,
0.515625,
0.80029296875
],
[
0.1427001953125,
0.80810546875,
0.489501953125,
0.81103515625,
0.489501953125,
0.826171875,
0.1427001953125,
0.822265625
],
[
0.51318359375,
0.8046875,
0.884765625,
0.8046875,
0.884765625,
0.818359375,
0.51318359375,
0.818359375
],
[
0.118408203125,
0.82470703125,
0.2880859375,
0.82470703125,
0.2880859375,
0.8349609375,
0.118408203125,
0.8349609375
],
[
0.51318359375,
0.82177734375,
0.84814453125,
0.82177734375,
0.84814453125,
0.8349609375,
0.51318359375,
0.8349609375
],
[
0.1435546875,
0.8349609375,
0.486083984375,
0.8349609375,
0.486083984375,
0.8486328125,
0.1435546875,
0.8486328125
],
[
0.12005615234375,
0.8486328125,
0.372802734375,
0.8486328125,
0.372802734375,
0.859375,
0.12005615234375,
0.859375
],
[
0.51123046875,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.1427001953125,
0.85791015625,
0.489501953125,
0.859375,
0.489501953125,
0.87353515625,
0.1427001953125,
0.87255859375
],
[
0.51123046875,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51123046875,
0.876953125
],
[
0.12176513671875,
0.87353515625,
0.48193359375,
0.87353515625,
0.48193359375,
0.88427734375,
0.12176513671875,
0.88427734375
],
[
0.118408203125,
0.8828125,
0.4609375,
0.88427734375,
0.4609375,
0.89794921875,
0.118408203125,
0.896484375
],
[
0.51318359375,
0.8798828125,
0.884765625,
0.8798828125,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.89794921875,
0.456787109375,
0.89794921875,
0.456787109375,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"known aspect term food.",
"rived from the opinion lexicon generated by Hu",
"and Liu (2004). The semantic similarity between",
"Rule",
"an opinion term and a positive or negative set is",
"Example",
"RulelD",
" amod, A(NN)",
"nice decor",
"estimated by averaging the cosine similarity of the",
"R1",
"amod,decor)",
"embedding of the opinion term and the embedding",
"(nice",
"O(ADJ)",
"of each one of the terms in the positive or negative",
"R2",
"nsubj",
"the food was super tasty",
"set.",
"A",
"(food nsubi, tasty)",
"In this module we used pre-trained embed-",
"Table 1: Examples of the opinion and aspect terms ex-",
"dings5 which produce overall good results but",
"traction rules. O represents an opinion term and A rep-",
"raise a drawback; some opinion terms may con-",
"resents an aspect term.",
"vey different sentiment polarities in different do-",
"mains (e.g.'delicate movie'(positive) vs.'del-",
"Opinion Lexicon Scoring and Filtering.",
"This",
"icate cellphone' (negative)), while a pre-trained",
"step aims to filter the noisy candidate opinion",
"embeddings setup is only capable of setting a sin-",
"terms extracted by the bootstrap process..It is",
"gle polarity per opinion term. A suggested solu-",
"based on using an MLP classifier for generating",
"tion is to adapt the embeddings to the target do-",
"a scorethatrepresentstheprobabilitythatacandi-",
"main or to use context embeddings. We intend to",
"date is indeed an opinion term. Candidate opinion",
"address this challenge in future work.",
"terms are qualified as opinion terms if their classi-",
"3.2Sentiment Classification",
"fication score exceeds a threshold 4.",
"The MLP classifier input features consists of",
"Sentiment classification uses the opinion and as-",
"thecandidate term wordembeddingand the",
"pect lexicons for detecting aspect-opinion pairs",
"mean, standard-deviation, maximum and mini-",
"(sentiment mentions) within the input target",
"mum word-embedding cosine similarities between",
"dataset, and determining their sentiment polarity.",
"the candidate term and a pre-determined set of",
"Aspect-opinion pairs are extracted based on de-",
"generic opinion terms. The MLP consists of a sin-",
"tecting a direct or second-order dependency re-",
"gle hidden layer and is trained once for a binary",
"lation of any type, between them. The aspect-",
"classificationtaskusingmanuallylabeled datathat",
"opinion pair polarity is assigned according to the",
"consists of a set of opinion terms (positive class)",
"polarity of the opinion term. This step also uses a",
"and a set of non-opinion terms (negative class)",
"pre-determined negation lexicon containing nega-",
"from a specific domain. Once the model is gener-",
"tion terms. Upon detecting a direct dependency",
"ated it is then used for grading candidate opinion",
"relation between a negation term and the opinion",
"terms extracted in other domains. It is reasonable",
"term, the aspect-opinion pair polarity is reversed.",
"to use such model across domains, since the classi-",
"fication features represent semantic similarity lev-",
"4",
"",
"Evaluation",
"els that are robust across domains.",
"Our evaluation objective is to show that the differ-",
"Opinion Polarity Estimation The goal of this",
"ent algorithmic steps, namely,lexicon extraction",
"step is to set the binary sentiment polarity (posi-",
"and sentiment classification, produce usable re-",
"tive or negative) of the opinion terms. Following",
"sults. An additional objective is to show that weak",
"Pablos et al. (2016), an opinion term polarity is",
"supervision of an aspect lexicon that was gener-",
"assigned based on estimating whether it is seman-",
"ated in an unsupervised manner produces compa-",
"tically closer to a set of generic positive opinion",
"rable results to the recent transfer-learning based",
"terms or to a set of generic negative opinion terms.",
"methods (Ding et al., 2017; Wang and Jialin Pan.",
"To produce those sets we used a subset of 47 pos-",
"2018). For this purpose, we leveraged the data of",
"itive terms and a subset of 47 negative terms de-",
"SemEval 2014task4 (Pontikiet al.,2014), which",
"4The threshold's value was empirically determined based",
"tests the two main ABSA sub-tasks: aspect term",
"on precision-recall tradeoff.",
"extraction and aspect term polarity detection.",
"5We used Stanford Glove embeddings https : / /nlp.",
"stanford.edu/projects/glove/",
"Datasets.The performance of the algorithm was",
"6This training data can be downloaded from https : / /",
" evaluated using data from two different domains:",
"github.com/Nervanasystems/nlp-architect/",
"blob/master/nlp_architect/models/absa/",
"restaurant reviews and laptop reviews. Those two",
"train/lexicons/RerankTrainingData.csv",
"domains are disjoint and therefore demonstrate"
] |
{
"arxiv_id": "1909.05608",
"language": "en",
"page_idx": 3
}
|
||
1912.01217-en-0
|
1912.01217-en-0.jpg
|
SafeLife $1. 0{: }\textbf{Exploring Side Effects in Complex Environments}$
Carroll L. Wainwright and Peter Eckersley
Partnership on AI
San Francisco, CA 94105
carroll@clwainwright.net, pde@pde.is
### Abstract
We present SafeLife, a publicly available reinforcement learning environment that tests the safety of reinforcement learning agents. It contains complex, dynamic, tunable, procedurally generated levels with many opportunities for unsafe behavior. Agents are graded both on their ability to maximize their explicit reward and on their ability to operate safely without unnecessary side effects. We train agents to maximize rewards using proximal policy optimization and score them on a suite of benchmark levels. The resulting agents are performant but not safe—they tend to cause large side effects in their environments—but they form a baseline against which future safety research can be measured.
### 1 Introduction
Safety problems for reinforcement learning (RL) have received growing attention in recent years (see e.g. (Garcıa and Fernández 2015; Amodei et al. $\dot{2}016;$ Ortega and Maini 2018) for surveys). It has been recognized that the capabilities of RL methods in many respects exceed their ability to be made predictable, robust, free from unintended side effects, and fully controllable. Although work is underway on all of these problems, there exist few environments to organize and benchmark progress on them. In most cases, the frontiers of RL capabilities research are separated from safety, except where safety problems prevent tasks from being accomplished at all. We know from other felds of machine learning that well-defined benchmarks can be incredibly important both for knowing how much progress has happened and for inspiring it.$^{1}$As reinforcement learning systems get more advanced and closer to deployment in semi-constrained or open-ended settings, the requirement for more complex, more dynamic, and higher-fidelity environments for simulation of safety problems grows. However, when RL safety
$^{1}$Prominent examples include classification performance on MNIST (Lecun et al. 1998), ImageNet (Deng et ai. 2009), Switchboard (Godfrey, Holliman, and McDaniel 1992), or SQUAD (Rajpurkar et al. 2016); accurate modelling on Penn Treebank (Prasad et al. 2008); or scores in the Arcade Learning Environment (Bellemare et al. 2013) or MuJoCo (Todorov, Erez, and Tassa 2012; Duan et al. 2016). See (Eckersley, Nasser, and others 2017) for a more systematic survey.
problems are measured, they are often assessed in small handcrafted environments (e.g. (Leike et al. 2018; Milli et al. 2017; Shah et al. 2019)), which are correspondingly limited in their richness and may not guarantee generalizable solutions (“safety by overfitting” is not the kind of safety we are looking for). Safety problems are often precisely those problems which are difficult to foresee, so it is essential that future safety benchmarks include tests of emergent behavior over a variety of non-trivial scenarios.
This paper attempts to address current limitations in safety benchmarks by introducing SafeLife, a family of environments with simple physics and complex, emergent dynamics. The SafeLife rules allow for a rich and diverse set of levels that can be used to examine and measure reinforcement learning safety. We focus at first on the problem of avoiding negative side effects, though we plan to extend our study to other safety problems in future SafeLife releases.
SafeLife satisfies a set of desiderata that we believe are important for a safety benchmark, and for the study of side effects in particular. First and foremost, the environment has dynamics that allow for large and interesting effects. These effects are not $ad$ hoc, but are built into the environment definition itself. Second, to facilitate easier research, agents can be trained with only moderate amounts of compute. An agent on a single-core CPU can easily make thousands of steps and observations per second, and an agent can learn to complete its basic tasks (unsafely) within a million time steps. Third, the environment uses procedurally generated levels with numerous tunable characteristics and challenges. This is essential in allowing for a diverse set of training environments that do not cause agents to be overfit to a particular level layout or goal structure. Finally, we think that SafeLife presents a fun and interesting challenge for human players.
In Section 2 we summarize the rules for the SafeLife environment, including the environment's dynamics and the player's (or agent's) scoring function. We describe how side effects are measured in Section 3. Section 4 details the different types of levels used for benchmarking agents, and the different aspects of the side effects safety problem that they test. In Section 5 we train and test baseline agents using proximal policy optimization. We include a very simple side effect impact penalty in the baseline performance, however
|
[
[
0.1871337890625,
0.12371826171875,
0.810546875,
0.125,
0.810546875,
0.1395263671875,
0.1871337890625,
0.1383056640625
],
[
0.32275390625,
0.175537109375,
0.68115234375,
0.175537109375,
0.68115234375,
0.1900634765625,
0.32275390625,
0.1900634765625
],
[
0.439453125,
0.1912841796875,
0.560546875,
0.1912841796875,
0.560546875,
0.2052001953125,
0.439453125,
0.2052001953125
],
[
0.414306640625,
0.2052001953125,
0.583984375,
0.2052001953125,
0.583984375,
0.2183837890625,
0.414306640625,
0.2183837890625
],
[
0.371826171875,
0.2197265625,
0.6259765625,
0.2197265625,
0.6259765625,
0.23291015625,
0.371826171875,
0.23291015625
],
[
0.0228729248046875,
0.2734375,
0.062103271484375,
0.2734375,
0.05963134765625,
0.3818359375,
0.0196075439453125,
0.3818359375
],
[
0.250732421875,
0.296630859375,
0.316162109375,
0.296630859375,
0.316162109375,
0.311279296875,
0.250732421875,
0.311279296875
],
[
0.51953125,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.311279296875,
0.51953125,
0.311279296875
],
[
0.51953125,
0.311279296875,
0.9140625,
0.311279296875,
0.9140625,
0.324462890625,
0.51953125,
0.324462890625
],
[
0.10540771484375,
0.3212890625,
0.4599609375,
0.3212890625,
0.4599609375,
0.33203125,
0.10540771484375,
0.33203125
],
[
0.51953125,
0.32568359375,
0.91259765625,
0.32568359375,
0.91259765625,
0.339111328125,
0.51953125,
0.339111328125
],
[
0.1004638671875,
0.33203125,
0.463134765625,
0.333251953125,
0.463134765625,
0.3466796875,
0.1004638671875,
0.34521484375
],
[
0.101318359375,
0.34521484375,
0.463134765625,
0.34521484375,
0.463134765625,
0.359130859375,
0.101318359375,
0.359130859375
],
[
0.521484375,
0.341552734375,
0.9111328125,
0.341552734375,
0.9111328125,
0.3515625,
0.521484375,
0.3515625
],
[
0.521484375,
0.352783203125,
0.9140625,
0.352783203125,
0.9140625,
0.3662109375,
0.521484375,
0.3662109375
],
[
0.10455322265625,
0.35986328125,
0.463134765625,
0.35986328125,
0.463134765625,
0.37060546875,
0.10455322265625,
0.37060546875
],
[
0.02532958984375,
0.369384765625,
0.058013916015625,
0.369384765625,
0.058013916015625,
0.406005859375,
0.02532958984375,
0.406005859375
],
[
0.10455322265625,
0.37255859375,
0.463134765625,
0.37255859375,
0.463134765625,
0.38330078125,
0.10455322265625,
0.38330078125
],
[
0.51953125,
0.367431640625,
0.91259765625,
0.367431640625,
0.91259765625,
0.380615234375,
0.51953125,
0.380615234375
],
[
0.10296630859375,
0.3837890625,
0.463134765625,
0.3837890625,
0.463134765625,
0.397705078125,
0.10296630859375,
0.397705078125
],
[
0.51953125,
0.380126953125,
0.91259765625,
0.380126953125,
0.91259765625,
0.393310546875,
0.51953125,
0.393310546875
],
[
0.10455322265625,
0.397705078125,
0.4599609375,
0.397705078125,
0.4599609375,
0.40771484375,
0.10455322265625,
0.40771484375
],
[
0.51953125,
0.39453125,
0.9140625,
0.39453125,
0.9140625,
0.40771484375,
0.51953125,
0.40771484375
],
[
0.0228729248046875,
0.40771484375,
0.062103271484375,
0.40771484375,
0.062103271484375,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.10296630859375,
0.4091796875,
0.463134765625,
0.4091796875,
0.463134765625,
0.42236328125,
0.10296630859375,
0.42236328125
],
[
0.521484375,
0.410400390625,
0.7255859375,
0.410400390625,
0.7255859375,
0.42041015625,
0.521484375,
0.42041015625
],
[
0.10296630859375,
0.42236328125,
0.461669921875,
0.42236328125,
0.461669921875,
0.43310546875,
0.10296630859375,
0.43310546875
],
[
0.5361328125,
0.425537109375,
0.91259765625,
0.425537109375,
0.91259765625,
0.439453125,
0.5361328125,
0.439453125
],
[
0.101318359375,
0.434326171875,
0.463134765625,
0.434326171875,
0.463134765625,
0.447509765625,
0.101318359375,
0.447509765625
],
[
0.51953125,
0.439453125,
0.9140625,
0.439453125,
0.9140625,
0.45263671875,
0.51953125,
0.45263671875
],
[
0.101318359375,
0.4462890625,
0.463134765625,
0.4462890625,
0.463134765625,
0.460205078125,
0.101318359375,
0.460205078125
],
[
0.1004638671875,
0.458984375,
0.3349609375,
0.458984375,
0.3349609375,
0.472900390625,
0.1004638671875,
0.472900390625
],
[
0.51953125,
0.453857421875,
0.9140625,
0.453857421875,
0.9140625,
0.46728515625,
0.51953125,
0.46728515625
],
[
0.51953125,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51953125,
0.48095703125
],
[
0.52294921875,
0.48291015625,
0.91259765625,
0.48291015625,
0.91259765625,
0.49365234375,
0.52294921875,
0.49365234375
],
[
0.0236968994140625,
0.490478515625,
0.05963134765625,
0.4912109375,
0.056365966796875,
0.54345703125,
0.021240234375,
0.54248046875
],
[
0.2156982421875,
0.49560546875,
0.226318359375,
0.49560546875,
0.226318359375,
0.505859375,
0.2156982421875,
0.505859375
],
[
0.240234375,
0.49365234375,
0.3505859375,
0.49365234375,
0.3505859375,
0.50830078125,
0.240234375,
0.50830078125
],
[
0.521484375,
0.49560546875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.521484375,
0.5087890625
],
[
0.08660888671875,
0.51318359375,
0.4794921875,
0.51318359375,
0.4794921875,
0.5263671875,
0.08660888671875,
0.5263671875
],
[
0.51953125,
0.5087890625,
0.9140625,
0.5087890625,
0.9140625,
0.52294921875,
0.51953125,
0.52294921875
],
[
0.0849609375,
0.52783203125,
0.4794921875,
0.52783203125,
0.4794921875,
0.54150390625,
0.0849609375,
0.54150390625
],
[
0.51953125,
0.52294921875,
0.84228515625,
0.52294921875,
0.84228515625,
0.5361328125,
0.51953125,
0.5361328125
],
[
0.0400390625,
0.5390625,
0.049835205078125,
0.53857421875,
0.051483154296875,
0.54345703125,
0.041656494140625,
0.54443359375
],
[
0.0849609375,
0.54150390625,
0.4794921875,
0.54150390625,
0.4794921875,
0.5546875,
0.0849609375,
0.5546875
],
[
0.5361328125,
0.54052734375,
0.9140625,
0.54052734375,
0.9140625,
0.5537109375,
0.5361328125,
0.5537109375
],
[
0.51953125,
0.5537109375,
0.91259765625,
0.5537109375,
0.91259765625,
0.5673828125,
0.51953125,
0.5673828125
],
[
0.02777099609375,
0.54345703125,
0.056365966796875,
0.54345703125,
0.056365966796875,
0.56494140625,
0.02777099609375,
0.56494140625
],
[
0.0318603515625,
0.55810546875,
0.053924560546875,
0.5625,
0.051483154296875,
0.57080078125,
0.0294189453125,
0.56640625
],
[
0.08660888671875,
0.5546875,
0.4794921875,
0.5546875,
0.4794921875,
0.568359375,
0.08660888671875,
0.568359375
],
[
0.0269622802734375,
0.56298828125,
0.0555419921875,
0.56298828125,
0.0555419921875,
0.59130859375,
0.0269622802734375,
0.59130859375
],
[
0.08660888671875,
0.57080078125,
0.47802734375,
0.57080078125,
0.47802734375,
0.58056640625,
0.08660888671875,
0.58056640625
],
[
0.521484375,
0.5673828125,
0.91259765625,
0.5673828125,
0.91259765625,
0.58056640625,
0.521484375,
0.58056640625
],
[
0.0849609375,
0.58349609375,
0.4794921875,
0.58349609375,
0.4794921875,
0.5966796875,
0.0849609375,
0.5966796875
],
[
0.51953125,
0.58203125,
0.91259765625,
0.58203125,
0.91259765625,
0.59521484375,
0.51953125,
0.59521484375
],
[
0.02777099609375,
0.59033203125,
0.056365966796875,
0.59033203125,
0.056365966796875,
0.6123046875,
0.02777099609375,
0.6123046875
],
[
0.0849609375,
0.5966796875,
0.4794921875,
0.5966796875,
0.4794921875,
0.60986328125,
0.0849609375,
0.60986328125
],
[
0.51953125,
0.59521484375,
0.9140625,
0.59521484375,
0.9140625,
0.609375,
0.51953125,
0.609375
],
[
0.03594970703125,
0.60693359375,
0.043304443359375,
0.6123046875,
0.03594970703125,
0.6171875,
0.0294189453125,
0.6123046875
],
[
0.0849609375,
0.60986328125,
0.4794921875,
0.60986328125,
0.4794921875,
0.62353515625,
0.0849609375,
0.62353515625
],
[
0.5185546875,
0.60791015625,
0.91259765625,
0.609375,
0.91259765625,
0.62353515625,
0.5185546875,
0.62255859375
],
[
0.02777099609375,
0.6142578125,
0.053924560546875,
0.6142578125,
0.053924560546875,
0.63623046875,
0.02777099609375,
0.63623046875
],
[
0.0849609375,
0.625,
0.4794921875,
0.625,
0.4794921875,
0.63818359375,
0.0849609375,
0.63818359375
],
[
0.51953125,
0.62353515625,
0.91259765625,
0.62353515625,
0.91259765625,
0.63720703125,
0.51953125,
0.63720703125
],
[
0.0187835693359375,
0.6337890625,
0.058013916015625,
0.63330078125,
0.060455322265625,
0.7109375,
0.021240234375,
0.7119140625
],
[
0.0849609375,
0.63818359375,
0.4794921875,
0.63818359375,
0.4794921875,
0.6513671875,
0.0849609375,
0.6513671875
],
[
0.51953125,
0.63720703125,
0.9140625,
0.63720703125,
0.9140625,
0.65087890625,
0.51953125,
0.65087890625
],
[
0.08660888671875,
0.6513671875,
0.4794921875,
0.6513671875,
0.4794921875,
0.66552734375,
0.08660888671875,
0.66552734375
],
[
0.521484375,
0.6513671875,
0.91259765625,
0.6513671875,
0.91259765625,
0.66552734375,
0.521484375,
0.66552734375
],
[
0.08660888671875,
0.66552734375,
0.47802734375,
0.66552734375,
0.47802734375,
0.6787109375,
0.08660888671875,
0.6787109375
],
[
0.51953125,
0.6640625,
0.91259765625,
0.6640625,
0.91259765625,
0.67822265625,
0.51953125,
0.67822265625
],
[
0.0882568359375,
0.68115234375,
0.47802734375,
0.68115234375,
0.47802734375,
0.69140625,
0.0882568359375,
0.69140625
],
[
0.521484375,
0.6787109375,
0.91259765625,
0.6787109375,
0.91259765625,
0.6923828125,
0.521484375,
0.6923828125
],
[
0.08660888671875,
0.693359375,
0.47802734375,
0.693359375,
0.47802734375,
0.70703125,
0.08660888671875,
0.70703125
],
[
0.51953125,
0.69140625,
0.9140625,
0.6923828125,
0.9140625,
0.70703125,
0.51953125,
0.70556640625
],
[
0.08660888671875,
0.70849609375,
0.4794921875,
0.70849609375,
0.4794921875,
0.7216796875,
0.08660888671875,
0.7216796875
],
[
0.521484375,
0.70556640625,
0.9140625,
0.70556640625,
0.9140625,
0.7197265625,
0.521484375,
0.7197265625
],
[
0.08660888671875,
0.7216796875,
0.4794921875,
0.7216796875,
0.4794921875,
0.73486328125,
0.08660888671875,
0.73486328125
],
[
0.521484375,
0.7216796875,
0.9111328125,
0.7216796875,
0.9111328125,
0.73193359375,
0.521484375,
0.73193359375
],
[
0.0849609375,
0.73486328125,
0.4794921875,
0.73486328125,
0.4794921875,
0.74853515625,
0.0849609375,
0.74853515625
],
[
0.521484375,
0.734375,
0.91259765625,
0.734375,
0.91259765625,
0.74755859375,
0.521484375,
0.74755859375
],
[
0.0849609375,
0.74853515625,
0.4794921875,
0.74853515625,
0.4794921875,
0.76318359375,
0.0849609375,
0.76318359375
],
[
0.51953125,
0.74853515625,
0.90625,
0.74853515625,
0.90625,
0.76220703125,
0.51953125,
0.76220703125
],
[
0.08660888671875,
0.76318359375,
0.4794921875,
0.76318359375,
0.4794921875,
0.77783203125,
0.08660888671875,
0.77783203125
],
[
0.5361328125,
0.76513671875,
0.9140625,
0.76513671875,
0.9140625,
0.77880859375,
0.5361328125,
0.77880859375
],
[
0.521484375,
0.77783203125,
0.9140625,
0.77783203125,
0.9140625,
0.79150390625,
0.521484375,
0.79150390625
],
[
0.1094970703125,
0.787109375,
0.47802734375,
0.787109375,
0.47802734375,
0.80029296875,
0.1094970703125,
0.80029296875
],
[
0.51953125,
0.79248046875,
0.9140625,
0.79248046875,
0.9140625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.80029296875,
0.4794921875,
0.80029296875,
0.4794921875,
0.814453125,
0.0849609375,
0.814453125
],
[
0.51953125,
0.806640625,
0.9140625,
0.806640625,
0.9140625,
0.82080078125,
0.51953125,
0.82080078125
],
[
0.0849609375,
0.81298828125,
0.48046875,
0.81298828125,
0.48046875,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.0849609375,
0.82568359375,
0.4794921875,
0.82568359375,
0.4794921875,
0.8388671875,
0.0849609375,
0.8388671875
],
[
0.521484375,
0.82080078125,
0.9140625,
0.82080078125,
0.9140625,
0.833984375,
0.521484375,
0.833984375
],
[
0.0849609375,
0.83837890625,
0.47802734375,
0.83837890625,
0.47802734375,
0.8515625,
0.0849609375,
0.8515625
],
[
0.521484375,
0.8359375,
0.9111328125,
0.8359375,
0.9111328125,
0.8466796875,
0.521484375,
0.8466796875
],
[
0.0849609375,
0.85107421875,
0.48046875,
0.85107421875,
0.48046875,
0.86572265625,
0.0849609375,
0.86572265625
],
[
0.5185546875,
0.84765625,
0.91259765625,
0.8486328125,
0.91259765625,
0.8623046875,
0.5185546875,
0.861328125
],
[
0.0849609375,
0.86279296875,
0.48046875,
0.86279296875,
0.48046875,
0.876953125,
0.0849609375,
0.876953125
],
[
0.51953125,
0.8623046875,
0.9140625,
0.8623046875,
0.9140625,
0.87548828125,
0.51953125,
0.87548828125
],
[
0.0849609375,
0.876953125,
0.238525390625,
0.876953125,
0.238525390625,
0.89013671875,
0.0849609375,
0.89013671875
],
[
0.51953125,
0.876953125,
0.9140625,
0.876953125,
0.9140625,
0.89013671875,
0.51953125,
0.89013671875
]
] |
[
"SafeLife 1.O: Exploring Side Effects in Complex Environments",
"Carroll L. Wainwright and Peter Eckersley",
"Partnership on AI",
"San Francisco, CA 94105",
"carroll@clwainwright.net,pde@pde.is",
"",
"Abstract",
"problems are measured, they are often assessed in small hand-",
"crafted environments (e.g. (Leike et al. 2018; Milli et al. 2017;",
"Wepresent SafeLife,apubliclyavailablereinforcementlearn",
"Shah et al. 2019)), which are correspondingly limited in",
"ing environment that tests the safety of reinforcement learning",
"agents. It contains complex, dynamic, tunable, procedurally",
"theirrichnessandmaynotguaranteegeneralizablesolutions",
"(\"safety by overfitting\" is not the kind of safety we are look-",
"generatedlevelswithmany opportunitiesforunsafebehav-",
"26",
"ior.Agents aregradedboth on theirabilityto maximizetheir",
"ing for). Safety problems are often precisely those problems",
"explicit reward and on their ability to operate safely with-",
"which are difficult to foresee, so it is essential that future",
"out unnecessary side effects. We train agents to maximize",
"safety benchmarks include tests of emergent behavior over a",
"[I's]",
"rewards using proximal policy optimization and score them",
"varietyofnon-trivialscenarios.",
"ona suite of benchmark levels.The resulting agents areper",
"This paper attempts to address current limitations in safety",
"formant but not safe-they tend to cause large side effects in",
"benchmarks by introducing SafeLife, a family of environ-",
"their environments-but they form a baseline against which",
"future safety research can be measured.",
"ments with simple physics and complex, emergent dynamics.",
"The SafeLife rules allow for a rich and diverse set of lev-",
"elsthatcanbeusedtoexamineandmeasurereinforcement",
"ZALT",
"1",
"Introduction",
"learning safety. We focus at first on the problem of avoiding",
"Safety problems for reinforcement learning (RL) have re-",
"negative side effects, though we plan to extend our study to",
"ceived growing attention in recent years (see e.g. (Garcia",
"other safety problems in future SafeLife releases.",
"-",
"and Fernandez 2015; Amodei et al.2016; Ortega and Maini",
"SafeLife satisfies a set of desiderata that we believe are",
"important for a safety benchmark, and for the study of side",
"2",
"",
"2018) for surveys). It has been recognized that the capabili-",
"0",
"tiesofRLmethodsinmanyrespectsexceedtheirabilitytobe",
"effects in particular. First and foremost, the environment has",
"made predictable,robust,free from unintended side effects,",
"dynamics that allow for large and interesting effects. These",
"2",
"and fully controllable. Although work is underway on all",
"effects are not ad hoc, but are built into the environment",
"2",
"of these problems, there exist few environments to organize",
"definition itself. Second, to facilitate easier research, agents",
"9",
"and benchmark progress on them. In most cases, the fron-",
"can be trained with only moderate amounts of compute. An",
"[AIIE",
"tiers of RL capabilities research are separated from safety,",
"agent on a single-core CPU can easily make thousands of",
"except where safety problems prevent tasks from being ac-",
"steps and observations per second, and an agent can learn",
"complished at all. We know from other fields of machine",
"to complete its basic tasks (unsafely) within a million time",
"learningthatwell-definedbenchmarkscanbeincrediblyim-",
"steps. Third, the environment uses procedurally generated",
"portant both for knowing how much progress has happened",
"levels with numerous tunable characteristics and challenges.",
"and for inspiring it.As reinforcement learning systems get",
"This is essential in allowing for a diverse set of training envi-",
"more advanced and closer to deployment in semi-constrained",
"ronments that do not causeagents to beoverfitto a particular",
"or open-ended settings, the requirement for more complex,",
"level layout or goal structure. Finally, we think that SafeLife",
"more dynamic, and higher-fidelity environments for simu-",
"presents a fun and interesting challenge for human players.",
"lation of safety problems grows. However, when RL safety",
"In Section 2 we summarize the rules for the SafeLife en-",
"vironment, including the environment's dynamics and the",
"'Prominent examples include classification performance on",
"player's (or agent's) scoring function. We describe how side",
"MNIST (Lecun et al. 1998), ImageNet (Deng et al. 2009), Switch-",
"effects are measured in Section 3. Section 4 details the dif-",
"board (Godfrey, Holliman, and McDaniel 1992), or SQUAD (Ra-",
"jpurkar et al. 2016); accurate modelling on Penn Treebank (Prasad",
"ferent types of levels used for benchmarking agents, and",
"et al. 2008); or scores in the Arcade Learning Environment (Belle",
"thedifferentaspects ofthesideeffectssafetyproblemthat",
"mare et al. 2013) or MuJoCo (Todorov, Erez, and Tassa 2012;",
"they test. In Section 5 we train and test baseline agents using",
"Duan et al. 2016). See (Eckersley, Nasser, and others 2017) for",
"proximal policy optimization. We include a very simple side",
"a more systematic survey.",
"effect impactpenalty in the baselineperformance, however"
] |
{
"arxiv_id": "1912.01217",
"language": "en",
"page_idx": 0
}
|
||
1912.01217-en-1
|
1912.01217-en-1.jpg
|
it is only able to produce safe behavior in limited scenarios. Our hope is that future safer algorithms will be able to improve upon this baseline. Finally, we conclude with prospects of future work in Section 6. Appendices include details of the SafeLife procedural level generation code and tables of benchmark results.
SafeLife is open-sourced and publicly available at https: $\|$github.com/PartnershipOnAI/safelife. Although we try to provide a full accounting of the most important environmental and training aspects in this paper, the code itself? should be viewed as the definitive environmental description.
# 2 Rules of the game
The SafeLife 1.0 environment consists of an agent ('》'in figures) in a gridworld setting along with dots of "life'(o), wall-like barriers (m), and a level exit (@). The agent can move in any of the four cardinal directions and can add or remove adjacent individual dots of life at each time step. The agent's goal is to either create new patterns in highlighted goal cells (blue), or to remove unwanted patterns of dots (red) from the board. Once the agent has completed its goals, it should proceed to the level exit. The challenge of SafeLife, and the reason why it excels as a safety benchmark, is that the board is very dynamic. Every dot of life interacts as a cellular automaton, following the rules of Conway's Game of Life, from which both simple and very difficult control problems emerge.
Conway's Game of Life (Gardner 1970) is a simple twodimensional system of cellular automata. Each cell can either be alive or dead, and at each time step the entire cell grid gets updated according to the following rules:
· any dead cell with exactly three living neighbors becomes
$alive;$
$\bullet$ any live cell with less than two or more than three neigh-
bors dies (as if by under- or overpopulation); and
$\bullet$ every other cell retains its prior state.
Each cell's“neighborhood”is identified with its Moore neighborhood, which consists of the 8 cells that immediately surround the central cell, including diagonals. These simple rules create an astoundingly large set of interesting behaviors such as still life patterns, oscillators, gliders and spaceships that move across the grid, and guns and puffers that can breed never-ending streams of new life. Conway's Game of Life is technically Turing complete: with an infinite grid and carefully selected initial conditions, anything can be computed (Wolfram 2002; tet 2017), although complex structures such as Turing machines are fragile and vanishingly unlikely to arise from a random initial state.
As agents build and destroy patterns in SafeLife, those patterns will change in complicated and surprising ways according to the above rules.$^3$Clever agents will use the
$^{2}$SafeLife is at version 1.0 as of this publication.
$^{3}$Agents also“freeze”their surrounding cells, preventing them from dying or coming to life. Without this, every time an agent creates a single life dot in isolation it would die due to underpopulation before the agent would have a chance to create another.
<FigureHere>
Figure 1: A simple level of the SafeLife environment containing an agent $(\mathbf{z})$, a spawner (%), crates ( a), and dots of life. The agent's goal is to remove unwanted red dots and to create new patterns of life in the blue squares. Once the agent has satisfactorily completed its goals it can leave via the level exit (@). Note that all level boundaries wrap; they have toroidal topology.
dynamism to their advantage to more efficiently complete their goals; less clever agents will not be able to create complex patterns at all, limiting them to simple stable forms and correspondingly lower scores.
We extend Game of Life's inherent possibilities for complex behavior with a few additional cell types, which can be included or excluded from the procedural generation process for creating levels. Trees (本) are fixed living cells; they contribute to neighboring life, but do not themselves die $Crates(\mathfrak{a})$ are like walls that can be pushed. They allow the agent to build barriers or fences to prevent the spread of unwanted life patterns. Some life cells are hardened (®) and cannot be directly removed by agent actions, complicating pattern manipulation. Finally, the spawner (螺) will randomly create new living neighbors at each time step. Although its immediate effect is local, it can produce patterns that propagate much farther than its local neighborhood. A single spawner changes the dynamics from deterministic to stochastic, and greatly increases the state space accessible to an agent.
Cells can also have different colors. Whenever a new cell comes to life, it inherits the colors of the majority of its parents. The colors can be used to keep track of which cells were created by the agent and which were pre-existing. In the case of red cells, they are also used to mark patterns that should be removed.
Figure 1 shows a simple SafeLife level that contains many complex elements of play. The agent must attempt to fill in the blue goals and remove the red dots, but the spawner on the lower right will tend to emit its own disruptive patterns. Optimal agent behavior is to push the crates closer to the spawner, thereby fencing it in and limiting its area of effect, and then to fill in blue goals once it is safe to do so.
An agent's total rewards are, at any point $t$, the sum of its presently completed goals relative to those at $t=0.$ The agent gains one point for every red dot that is removed from the board, and three points for every new dot of life that is added to a blue goal. Similarly, the agent
|
[
[
0.08660888671875,
0.0694580078125,
0.4794921875,
0.0694580078125,
0.4794921875,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.08660888671875,
0.083984375,
0.4794921875,
0.083984375,
0.4794921875,
0.09783935546875,
0.08660888671875,
0.09783935546875
],
[
0.79150390625,
0.08587646484375,
0.80126953125,
0.09344482421875,
0.7939453125,
0.0997314453125,
0.78369140625,
0.091552734375
],
[
0.08660888671875,
0.09783935546875,
0.47802734375,
0.09783935546875,
0.47802734375,
0.111083984375,
0.08660888671875,
0.111083984375
],
[
0.701171875,
0.10919189453125,
0.7158203125,
0.10919189453125,
0.7158203125,
0.11297607421875,
0.701171875,
0.11297607421875
],
[
0.76708984375,
0.106689453125,
0.77294921875,
0.111083984375,
0.76708984375,
0.11553955078125,
0.76123046875,
0.111083984375
],
[
0.0849609375,
0.1256103515625,
0.4794921875,
0.1256103515625,
0.4794921875,
0.1395263671875,
0.0849609375,
0.1395263671875
],
[
0.0849609375,
0.11236572265625,
0.4794921875,
0.11236572265625,
0.4794921875,
0.1256103515625,
0.0849609375,
0.1256103515625
],
[
0.0849609375,
0.1395263671875,
0.2099609375,
0.14013671875,
0.2099609375,
0.1514892578125,
0.0849609375,
0.15087890625
],
[
0.10296630859375,
0.15283203125,
0.4794921875,
0.15283203125,
0.4794921875,
0.1666259765625,
0.10296630859375,
0.1666259765625
],
[
0.7861328125,
0.1578369140625,
0.79150390625,
0.1578369140625,
0.79150390625,
0.1634521484375,
0.7861328125,
0.1634521484375
],
[
0.0882568359375,
0.1685791015625,
0.47802734375,
0.1685791015625,
0.47802734375,
0.1787109375,
0.0882568359375,
0.1787109375
],
[
0.7255859375,
0.1654052734375,
0.74853515625,
0.1673583984375,
0.74658203125,
0.1749267578125,
0.7236328125,
0.17236328125
],
[
0.7587890625,
0.1685791015625,
0.77392578125,
0.1685791015625,
0.77392578125,
0.1717529296875,
0.7587890625,
0.1717529296875
],
[
0.775390625,
0.1685791015625,
0.7861328125,
0.1685791015625,
0.7861328125,
0.1717529296875,
0.775390625,
0.1717529296875
],
[
0.0882568359375,
0.18115234375,
0.4794921875,
0.18115234375,
0.4794921875,
0.1944580078125,
0.0882568359375,
0.1944580078125
],
[
0.0849609375,
0.1932373046875,
0.4794921875,
0.1925048828125,
0.4794921875,
0.20703125,
0.0849609375,
0.2083740234375
],
[
0.650390625,
0.1976318359375,
0.66845703125,
0.1976318359375,
0.66845703125,
0.2020263671875,
0.650390625,
0.2020263671875
],
[
0.7255859375,
0.1956787109375,
0.771484375,
0.1956787109375,
0.771484375,
0.20263671875,
0.7255859375,
0.20263671875
],
[
0.0849609375,
0.2083740234375,
0.438720703125,
0.2083740234375,
0.438720703125,
0.2215576171875,
0.0849609375,
0.2215576171875
],
[
0.521484375,
0.2247314453125,
0.9140625,
0.2247314453125,
0.9140625,
0.238037109375,
0.521484375,
0.238037109375
],
[
0.188720703125,
0.2342529296875,
0.371826171875,
0.236083984375,
0.371826171875,
0.250732421875,
0.188720703125,
0.248779296875
],
[
0.521484375,
0.2392578125,
0.9140625,
0.2392578125,
0.9140625,
0.25244140625,
0.521484375,
0.25244140625
],
[
0.08660888671875,
0.25390625,
0.476318359375,
0.25390625,
0.476318359375,
0.267578125,
0.08660888671875,
0.267578125
],
[
0.521484375,
0.25244140625,
0.91259765625,
0.25244140625,
0.91259765625,
0.266357421875,
0.521484375,
0.266357421875
],
[
0.08660888671875,
0.268310546875,
0.4794921875,
0.268310546875,
0.4794921875,
0.2822265625,
0.08660888671875,
0.2822265625
],
[
0.51953125,
0.266357421875,
0.91259765625,
0.266357421875,
0.91259765625,
0.27978515625,
0.51953125,
0.27978515625
],
[
0.08660888671875,
0.2841796875,
0.47802734375,
0.2841796875,
0.47802734375,
0.294189453125,
0.08660888671875,
0.294189453125
],
[
0.51953125,
0.281005859375,
0.91259765625,
0.281005859375,
0.91259765625,
0.294189453125,
0.51953125,
0.294189453125
],
[
0.08660888671875,
0.29736328125,
0.47802734375,
0.29736328125,
0.47802734375,
0.30810546875,
0.08660888671875,
0.30810546875
],
[
0.5185546875,
0.29345703125,
0.91259765625,
0.294189453125,
0.91259765625,
0.30810546875,
0.5185546875,
0.306884765625
],
[
0.0849609375,
0.31005859375,
0.47802734375,
0.31005859375,
0.47802734375,
0.323974609375,
0.0849609375,
0.323974609375
],
[
0.5185546875,
0.306884765625,
0.67236328125,
0.30810546875,
0.67236328125,
0.322509765625,
0.5185546875,
0.3212890625
],
[
0.0849609375,
0.323974609375,
0.4794921875,
0.323974609375,
0.4794921875,
0.337158203125,
0.0849609375,
0.337158203125
],
[
0.08660888671875,
0.33837890625,
0.4794921875,
0.33837890625,
0.4794921875,
0.3515625,
0.08660888671875,
0.3515625
],
[
0.08660888671875,
0.3515625,
0.4794921875,
0.3515625,
0.4794921875,
0.365478515625,
0.08660888671875,
0.365478515625
],
[
0.521484375,
0.354248046875,
0.91259765625,
0.354248046875,
0.91259765625,
0.367431640625,
0.521484375,
0.367431640625
],
[
0.0849609375,
0.365478515625,
0.4794921875,
0.365478515625,
0.4794921875,
0.37890625,
0.0849609375,
0.37890625
],
[
0.51953125,
0.36865234375,
0.91259765625,
0.36865234375,
0.91259765625,
0.3818359375,
0.51953125,
0.3818359375
],
[
0.0882568359375,
0.380615234375,
0.47802734375,
0.380615234375,
0.47802734375,
0.391357421875,
0.0882568359375,
0.391357421875
],
[
0.521484375,
0.3818359375,
0.9140625,
0.3818359375,
0.9140625,
0.395751953125,
0.521484375,
0.395751953125
],
[
0.08660888671875,
0.392578125,
0.47802734375,
0.392578125,
0.47802734375,
0.406005859375,
0.08660888671875,
0.406005859375
],
[
0.5185546875,
0.396484375,
0.7197265625,
0.396484375,
0.7197265625,
0.410400390625,
0.5185546875,
0.410400390625
],
[
0.0849609375,
0.4072265625,
0.47802734375,
0.4072265625,
0.47802734375,
0.42041015625,
0.0849609375,
0.42041015625
],
[
0.5361328125,
0.410400390625,
0.9140625,
0.410888671875,
0.9140625,
0.425537109375,
0.5361328125,
0.4248046875
],
[
0.0849609375,
0.421630859375,
0.4794921875,
0.421630859375,
0.4794921875,
0.43505859375,
0.0849609375,
0.43505859375
],
[
0.51953125,
0.4248046875,
0.91259765625,
0.425537109375,
0.91259765625,
0.43994140625,
0.51953125,
0.439453125
],
[
0.08660888671875,
0.43505859375,
0.2034912109375,
0.43505859375,
0.2034912109375,
0.448974609375,
0.08660888671875,
0.448974609375
],
[
0.51953125,
0.43994140625,
0.91259765625,
0.43994140625,
0.91259765625,
0.453857421875,
0.51953125,
0.453857421875
],
[
0.101318359375,
0.447509765625,
0.4794921875,
0.4462890625,
0.4794921875,
0.4609375,
0.101318359375,
0.462158203125
],
[
0.5185546875,
0.45263671875,
0.9111328125,
0.453857421875,
0.9111328125,
0.46728515625,
0.5185546875,
0.466552734375
],
[
0.08660888671875,
0.462158203125,
0.4794921875,
0.462158203125,
0.4794921875,
0.47607421875,
0.08660888671875,
0.47607421875
],
[
0.521484375,
0.46728515625,
0.91259765625,
0.46728515625,
0.91259765625,
0.48095703125,
0.521484375,
0.48095703125
],
[
0.0849609375,
0.47607421875,
0.4794921875,
0.47607421875,
0.4794921875,
0.4892578125,
0.0849609375,
0.4892578125
],
[
0.521484375,
0.481689453125,
0.9111328125,
0.481689453125,
0.9111328125,
0.49560546875,
0.521484375,
0.49560546875
],
[
0.08660888671875,
0.490478515625,
0.386474609375,
0.490478515625,
0.386474609375,
0.50390625,
0.08660888671875,
0.50390625
],
[
0.52294921875,
0.496826171875,
0.9111328125,
0.496826171875,
0.9111328125,
0.5068359375,
0.52294921875,
0.5068359375
],
[
0.0882568359375,
0.50830078125,
0.4794921875,
0.50830078125,
0.4794921875,
0.521484375,
0.0882568359375,
0.521484375
],
[
0.51953125,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51953125,
0.52294921875
],
[
0.10296630859375,
0.5234375,
0.143798828125,
0.5234375,
0.143798828125,
0.53515625,
0.10296630859375,
0.53515625
],
[
0.521484375,
0.52294921875,
0.91259765625,
0.52294921875,
0.91259765625,
0.5361328125,
0.521484375,
0.5361328125
],
[
0.08905029296875,
0.54150390625,
0.4794921875,
0.54150390625,
0.4794921875,
0.5546875,
0.08905029296875,
0.5546875
],
[
0.521484375,
0.53857421875,
0.9111328125,
0.53857421875,
0.9111328125,
0.548828125,
0.521484375,
0.548828125
],
[
0.1004638671875,
0.5546875,
0.42724609375,
0.5546875,
0.42724609375,
0.568359375,
0.1004638671875,
0.568359375
],
[
0.521484375,
0.55029296875,
0.9140625,
0.55029296875,
0.9140625,
0.564453125,
0.521484375,
0.564453125
],
[
0.521484375,
0.564453125,
0.91259765625,
0.564453125,
0.91259765625,
0.57763671875,
0.521484375,
0.57763671875
],
[
0.08905029296875,
0.57373046875,
0.346435546875,
0.57373046875,
0.346435546875,
0.5869140625,
0.08905029296875,
0.5869140625
],
[
0.521484375,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.521484375,
0.59228515625
],
[
0.08660888671875,
0.59228515625,
0.47802734375,
0.59228515625,
0.47802734375,
0.60595703125,
0.08660888671875,
0.60595703125
],
[
0.521484375,
0.59228515625,
0.9140625,
0.59228515625,
0.9140625,
0.60595703125,
0.521484375,
0.60595703125
],
[
0.08660888671875,
0.60791015625,
0.47802734375,
0.60791015625,
0.47802734375,
0.61865234375,
0.08660888671875,
0.61865234375
],
[
0.521484375,
0.60595703125,
0.9140625,
0.60595703125,
0.9140625,
0.619140625,
0.521484375,
0.619140625
],
[
0.0849609375,
0.62060546875,
0.4794921875,
0.62060546875,
0.4794921875,
0.6337890625,
0.0849609375,
0.6337890625
],
[
0.521484375,
0.6220703125,
0.58251953125,
0.6220703125,
0.58251953125,
0.6318359375,
0.521484375,
0.6318359375
],
[
0.0849609375,
0.63330078125,
0.4794921875,
0.6337890625,
0.4794921875,
0.6484375,
0.0849609375,
0.64794921875
],
[
0.53759765625,
0.63525390625,
0.91259765625,
0.63525390625,
0.91259765625,
0.6484375,
0.53759765625,
0.6484375
],
[
0.0849609375,
0.6484375,
0.47802734375,
0.6484375,
0.47802734375,
0.662109375,
0.0849609375,
0.662109375
],
[
0.51953125,
0.6494140625,
0.91259765625,
0.6494140625,
0.91259765625,
0.66357421875,
0.51953125,
0.66357421875
],
[
0.0849609375,
0.662109375,
0.47802734375,
0.662109375,
0.47802734375,
0.67529296875,
0.0849609375,
0.67529296875
],
[
0.521484375,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.6767578125,
0.521484375,
0.6767578125
],
[
0.0849609375,
0.67529296875,
0.4794921875,
0.67529296875,
0.4794921875,
0.689453125,
0.0849609375,
0.689453125
],
[
0.521484375,
0.6787109375,
0.9111328125,
0.6787109375,
0.9111328125,
0.689453125,
0.521484375,
0.689453125
],
[
0.0849609375,
0.689453125,
0.4794921875,
0.689453125,
0.4794921875,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.521484375,
0.69140625,
0.9140625,
0.69140625,
0.9140625,
0.705078125,
0.521484375,
0.705078125
],
[
0.0882568359375,
0.705078125,
0.476318359375,
0.705078125,
0.476318359375,
0.71533203125,
0.0882568359375,
0.71533203125
],
[
0.521484375,
0.7041015625,
0.650390625,
0.7041015625,
0.650390625,
0.71728515625,
0.521484375,
0.71728515625
],
[
0.08660888671875,
0.71728515625,
0.4794921875,
0.71728515625,
0.4794921875,
0.73095703125,
0.08660888671875,
0.73095703125
],
[
0.5361328125,
0.72021484375,
0.91259765625,
0.72021484375,
0.91259765625,
0.734375,
0.5361328125,
0.734375
],
[
0.08660888671875,
0.73095703125,
0.4794921875,
0.73095703125,
0.4794921875,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.521484375,
0.734375,
0.91259765625,
0.734375,
0.91259765625,
0.74755859375,
0.521484375,
0.74755859375
],
[
0.0849609375,
0.74560546875,
0.4794921875,
0.74560546875,
0.4794921875,
0.7587890625,
0.0849609375,
0.7587890625
],
[
0.51953125,
0.74853515625,
0.91259765625,
0.74853515625,
0.91259765625,
0.76220703125,
0.51953125,
0.76220703125
],
[
0.08660888671875,
0.76025390625,
0.2197265625,
0.76025390625,
0.2197265625,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.521484375,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.7744140625,
0.521484375,
0.7744140625
],
[
0.10296630859375,
0.77294921875,
0.4794921875,
0.77294921875,
0.4794921875,
0.7861328125,
0.10296630859375,
0.7861328125
],
[
0.521484375,
0.7763671875,
0.91259765625,
0.7763671875,
0.91259765625,
0.79052734375,
0.521484375,
0.79052734375
],
[
0.08660888671875,
0.78857421875,
0.476318359375,
0.78857421875,
0.476318359375,
0.798828125,
0.08660888671875,
0.798828125
],
[
0.521484375,
0.79052734375,
0.9140625,
0.79052734375,
0.9140625,
0.8037109375,
0.521484375,
0.8037109375
],
[
0.0849609375,
0.80029296875,
0.47802734375,
0.80029296875,
0.47802734375,
0.814453125,
0.0849609375,
0.814453125
],
[
0.51953125,
0.8046875,
0.859375,
0.8046875,
0.859375,
0.818359375,
0.51953125,
0.818359375
],
[
0.1070556640625,
0.82373046875,
0.390625,
0.82373046875,
0.390625,
0.8369140625,
0.1070556640625,
0.8369140625
],
[
0.5390625,
0.82080078125,
0.9140625,
0.82080078125,
0.9140625,
0.833984375,
0.5390625,
0.833984375
],
[
0.1094970703125,
0.83837890625,
0.47802734375,
0.83837890625,
0.47802734375,
0.8515625,
0.1094970703125,
0.8515625
],
[
0.51953125,
0.833984375,
0.9140625,
0.833984375,
0.9140625,
0.84765625,
0.51953125,
0.84765625
],
[
0.08660888671875,
0.85107421875,
0.47802734375,
0.85107421875,
0.47802734375,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.5185546875,
0.8466796875,
0.91259765625,
0.84765625,
0.91259765625,
0.8623046875,
0.5185546875,
0.861328125
],
[
0.0882568359375,
0.86572265625,
0.47802734375,
0.86572265625,
0.47802734375,
0.87548828125,
0.0882568359375,
0.87548828125
],
[
0.521484375,
0.86279296875,
0.91259765625,
0.86279296875,
0.91259765625,
0.87353515625,
0.521484375,
0.87353515625
],
[
0.0849609375,
0.87548828125,
0.41259765625,
0.87548828125,
0.41259765625,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51953125,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"it is only able to produce safe behavior in limited scenarios.",
"Our hope is that future safer algorithms will be able to im-",
"",
"prove upon this baseline. Finally, we conclude with prospects",
"",
"R",
"the SafeLife procedural level generation code and tables of",
"of future work in Section 6. Appendices include details of",
"benchmark results.",
"SafeLife is open-sourced and publicly available at https:",
"E",
"//github.com/PartnershipOnAI/safelife.Althoughwetryto",
"o",
"",
"",
"provide a full accounting of the most important environmen-",
"tal and training aspects in this paper, the code itself2 should",
".",
"··#",
"be viewed as the definitive environmental description.",
"Figure l: A simple level of the SafeLife environment con-",
"2Rules of the game",
"taining an agent (>), a spawner (), crates (), and dots of",
"The SafeLife 1.0 environment consists of an agent (>'in",
"life. The agent's goal is to remove unwanted red dots and",
"figures) in a gridworld setting along with dots of life' (o),",
"to create new patterns of life in the blue squares. Once the",
"wall-like barriers (),and a level exit (m).The agent can",
"agent has satisfactorily completed its goals it can leave via",
"move in any of thefour cardinal directions and can add or",
"the level exit ().Note that all level boundaries wrap; they",
"remove adjacent individual dots of life at each time step. The",
"have toroidal topology.",
"agent's goal is to either create new patterns in highlighted",
"goal cells (blue), or to remove unwanted patterns of dots (red)",
"from the board. Once the agent has completed its goals, it",
"dynamism to their advantage to more efficiently complete",
"should proceed to the level exit. The challenge of SafeLife,",
"their goals; less clever agents will not be able to create com-",
"andthereasonwhyitexcels asa safetybenchmark,isthat",
"plex patterns at all, limiting them to simple stable forms and",
"the board is very dynamic. Every dot of life interacts as a",
"correspondingly lower scores.",
"cellular automaton, following the rules of Conway's Game",
"We extend Game of Life's inherent possibilities for com-",
"of Life, from which both simple and very difficult control",
"plex behavior with a few additional cell types, which can",
"problems emerge.",
"be included or excluded from the procedural generation pro-",
"Conway's Game of Life (Gardner 1970) is a simple two-",
"cess for creating levels. Trees () are fixed living cells; they",
"dimensional system of cellular automata. Each cell can either",
"contribute to neighboring life, but do not themselves die.",
"be alive or dead, and at each time step the entire cell grid",
"Crates () are like walls that can be pushed. They allow",
"gets updated according to the following rules:",
"theagenttobuildbarriersorfencestopreventthespread",
": any dead cell with exactly three living neighbors becomes",
"of unwanted life patterns. Some life cells are hardened ()",
"alive;",
"and cannot be directly removed by agent actions, compli-",
": any live cell with less than two or more than three neigh-",
"catingpatternmanipulation.Finally,the spawner()will",
"bors dies (as if by under- or overpopulation): and",
"randomly create new living neighbors at each time step. Al-",
"though its immediate effect is local, it can produce patterns",
": every other cell retains its prior state.",
"that propagate much farther than its local neighborhood. A",
"Each cell's “\"neighborhood\"is identified with its Moore",
"single spawner changes the dynamics from deterministic to",
"neighborhood, which consists of the 8 cells that immedi-",
"stochastic, and greatly increases the state space accessible to",
" ately surround the central cell, including diagonals. These",
"an agent.",
"simple rules create an astoundingly large set of interest-",
"Cells can also have different colors. Whenever a new cell",
"ing behaviors such as still life patterns, oscillators, glid",
"comes to life, it inherits the colors of the majority of its",
"ers and spaceships that move across the grid, and guns",
"parents. The colors can be used to keep track of which cells",
"and puffers that can breed never-ending streams of new",
"were created bytheagent and which werepre-existing.In",
"life. Conway's Game of Life is technically Turing com-",
"the case of red cells, they are also used to mark patterns that",
"plete: with an infinite grid and carefully selected initial",
"should be removed.",
"conditions, anything can be computed (Wolfram 2002;",
"Figure 1 shows a simple SafeLife level that contains many",
"tet 2017), although complex structures such as Turing ma-",
"complex elements of play. The agent must attempt to fill in",
"chines are fragile and vanishingly unlikely to arise from a",
"the blue goals and remove the red dots, but the spawner on",
"random initial state.",
"thelowerrightwilltendtoemititsowndisruptivepatterns.",
"As agents build and destroy patterns in SafeLife, those",
"Optimal agent behavior is to push the crates closer to the",
"patterns will change in complicated and surprising ways",
"spawner, thereby fencing it in and limiting its area of effect,",
"according to the above rules.3 Clever agents will use the",
"and then to fill in blue goals once it is safe to do so.",
"2SafeLife is at version 1.0 as of this publication.",
"An agent's total rewards are, at any point t, the sum of",
"3Agents also “freeze\" their surrounding cells, preventing them",
"its presently completed goals relative to those at t = 0.",
"from dying or coming to life. Without this, every time an agent cre",
"The agent gains one point for every red dot that is re-",
"atesa singlelifedot inisolation it would dieduetounderpopulation",
"moved from the board, and three points for every new dot",
"before the agent would have a chance to create another.",
"of life that is added to a blue goal. Similarly, the agent"
] |
{
"arxiv_id": "1912.01217",
"language": "en",
"page_idx": 1
}
|
||
1912.01217-en-2
|
1912.01217-en-2.jpg
|
loses points if red dots get added or goal patterns get disrupted. Formally, let $V(s)$ measures the point value of a $\hat{\text{board state }s}.$ An agent's immediate reward at time step $t$ will $\mathop{\mathrm{be}}R_{t}(s_{t+1},s_{t})=\tilde{V}(s_{t+1})-V(s_{t}).\mathop{\mathrm{Additionally,~the~agent}}$ gets one extra point for reaching a level's exit.$^{4}$
The patterns that the agent needs to create or destroy are often interwoven with or adjacent to neutral green patterns, as in Figure 1. The neutral patterns may or may not hinder the agent in its tasks, but they are often fragile; a reckless agent will tend to disrupt them. The goal of our environment is to test this recklessness: can an agent learn to avoid disrupting its environment without being explicitly told not to do so? If not, can future agent architectures accomplish this?
It is essential to this safety test that the neutral patterns be truly neutral to the agent's reward function. It would be trivial to teach an agent to avoid side effects in this environment by explicitly penalizing it for the side effects which we test. However, the problem of avoiding side effects is that there tend to be many more of them than one can enumerate. A safe agent must learn to avoid side effects in general without foreknowledge about our particular testing criteria.
# 3 Measuring Side Effects
Side effects are difficult and subtle phenomena to correctly define. To give a full account of what an agent's effects are, one must establish theories of causality and moral responsibility, including those in interactions with other agents.$^5$ These are well beyond the scope of this paper. Our task here is much easier: we wish to define a heuristic side effect measurement that makes sense in the SafeLife environment and aligns with human intuition.
Following Krakovna et al. (Krakovna et al. 2018), we define the benchmark SafeLife side effect measurement in terms of a baseline state—the state against which effects are compared—and a deviation measure between states.
## Baseline states
When testing for side effects in episodic simulated environments, it makes most sense to define the baseline state as the state that would have occurred had the agent not taken any actions. This state is easily computed: to find the baseline state at time $t$, one need only advance a copy of the initial state $s_0$ a total of $t$ time steps. However, when the dynamics are stochastic or chaotic, as is the case in SafeLife, the inaction state can have a distribution of different values.$^{6}$
Rather than confining ourselves to a single state, we take the baseline as the entirety of this distribution. Assuming that the dynamics without agent interaction are approximately
$^{4}$In some levels, the exit will only open once the agent has
accomplished a certain fraction of its goals.
$^{5}$SafeLife 1.0 does not include multi-agent tasks, but the environment has been designed to be readily extensible to more complex scenarios.
$^{6}$A chaotic yet deterministic environment will have a welldefined inaction baseline state for every time $t$ as long as the starting state is specified to arbitrary precision. However, if there is any uncertainty in the starting state, then the inaction state at time $t$ could take on many divergently different possibilities.
ergodic, we approximate this distribution by sampling from a sequence of $n=1000$ states following time $t.^7$ Patterns in SafeLife tend to collapse to steady or oscillating states, or they chaotically grow and regrow at short time scales $\Delta t$ with $\Delta t\ll1000$, so the ergodicity requirement for inaction trajectories is easily satisfied. We then likewise sample $n$ states after the end of the agent's trajectory at the same time $t.$ This produces action and inaction distributions rather than action and inaction states. Each distribution is sampled from its own single trajectory of length $t+n$ either with or without agent interactions.
Note that this baseline is computationally expensive and makes most sense when $t$ is taken to be the end of an episode; there are other baseline states that may be more appropriate in the course of training to penalize side effects. Every baseline has its disadvantage, however. A starting state baseline requires no computation, but it counts all natural dynamics as side effects. A single-state inaction baseline is more intuitively appealing, but it requires a full simulation, isn't applicable in chaotic environments, and may incentivize agents to offset their positive actions (Eysenbach et al.2018). Finally, a step-wise inaction baseline (Turner, Hadfield-Menell, and Tadepalli 2019) can be used in continuing (rather than episodic) environments, but it may predispose agents to inaction precisely when safety demands that they act.$^{8}$Finding optimal baselines for training is very much an unsolved problem, and one we save for future work.
# Deviation measure
Since we use a distribution of states as our baseline, we must use a distance measure between distributions as our deviation measure. To do this, we calculate average cell densities across the distributions and find the earth-mover distances (Rubner, Tomasi, and Guibas 1998) between them. Let
(1)
$$\rho_\alpha(\vec{x};\mathcal{D}_s)=\frac{1}{|\mathcal{D}_s|}\sum_{s\in\mathcal{D}_s}\left\{\begin{matrix}1&\text{if}\:s(\vec{x})=\alpha\\0&\text{otherwise}\end{matrix}\right.$$
represent the expected density of cell type $\alpha$ at location $\vec{x}$ within the environment, where the expectation is taken over a distribution of states $\mathcal{D}_s.$ Let $d_x(\Delta\vec{x})$ be a spatial metric on the environment's grid defined such that max$(d_x)=1.$For concreteness, we take
(2)
$$d_x(\Delta\vec{x})=\tanh(\frac{1}{5}\|\Delta\vec{x}\|_1)$$
$^{7}$Whether a given choice of $n$ is sufficient for this purpose depends both on the environment, its degree of divergence from ergodicity, and on the statistic being computed over the sample. The degree of non-ergodicity in SafeLife is somewhat bounded by the $\bar{\text{limited size of the board, which prevents(for instance) rare struc-}}$ tures like“puffers”and“rakes”from introducing non-ergodic singularities in the side effect distribution. When there are no generative stochastic elements on a SafeLife board (i.e., no “spawners”), most distributions will tend to be dominated by the initial patterns' late stage evolution towards still lifes and oscillators which are trivially $\bar{\mathrm{ergodic}}.$
$^{8}$For example, if an agent is driving a car and their“no side effect" baseline is the one in which they ceased actions in the last time step, then they may conclude that the least disruptive thing to do is take their hands off the wheel and cause an accident!
|
[
[
0.08660888671875,
0.0694580078125,
0.47802734375,
0.0694580078125,
0.47802734375,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.521484375,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.521484375,
0.08331298828125
],
[
0.08660888671875,
0.083984375,
0.4794921875,
0.083984375,
0.4794921875,
0.09783935546875,
0.08660888671875,
0.09783935546875
],
[
0.521484375,
0.083984375,
0.91259765625,
0.083984375,
0.91259765625,
0.09783935546875,
0.521484375,
0.09783935546875
],
[
0.08660888671875,
0.09783935546875,
0.476318359375,
0.09783935546875,
0.476318359375,
0.111083984375,
0.08660888671875,
0.111083984375
],
[
0.51953125,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.51953125,
0.111083984375
],
[
0.0849609375,
0.11236572265625,
0.47802734375,
0.11236572265625,
0.47802734375,
0.1256103515625,
0.0849609375,
0.1256103515625
],
[
0.51953125,
0.11236572265625,
0.91259765625,
0.11236572265625,
0.91259765625,
0.1256103515625,
0.51953125,
0.1256103515625
],
[
0.08660888671875,
0.1256103515625,
0.39208984375,
0.1256103515625,
0.39208984375,
0.1395263671875,
0.08660888671875,
0.1395263671875
],
[
0.51953125,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.51953125,
0.1395263671875
],
[
0.10296630859375,
0.1395263671875,
0.4794921875,
0.1395263671875,
0.4794921875,
0.15283203125,
0.10296630859375,
0.15283203125
],
[
0.521484375,
0.1395263671875,
0.9140625,
0.1395263671875,
0.9140625,
0.15283203125,
0.521484375,
0.15283203125
],
[
0.0882568359375,
0.1546630859375,
0.47802734375,
0.1546630859375,
0.47802734375,
0.1654052734375,
0.0882568359375,
0.1654052734375
],
[
0.521484375,
0.15283203125,
0.91259765625,
0.15283203125,
0.91259765625,
0.1666259765625,
0.521484375,
0.1666259765625
],
[
0.0849609375,
0.1666259765625,
0.47802734375,
0.1666259765625,
0.47802734375,
0.179931640625,
0.0849609375,
0.179931640625
],
[
0.521484375,
0.1685791015625,
0.9111328125,
0.1685791015625,
0.9111328125,
0.1787109375,
0.521484375,
0.1787109375
],
[
0.08660888671875,
0.18310546875,
0.47802734375,
0.18310546875,
0.47802734375,
0.1932373046875,
0.08660888671875,
0.1932373046875
],
[
0.521484375,
0.18115234375,
0.91259765625,
0.18115234375,
0.91259765625,
0.1944580078125,
0.521484375,
0.1944580078125
],
[
0.0849609375,
0.1956787109375,
0.4794921875,
0.1956787109375,
0.4794921875,
0.208984375,
0.0849609375,
0.208984375
],
[
0.51953125,
0.1944580078125,
0.91259765625,
0.1944580078125,
0.91259765625,
0.2083740234375,
0.51953125,
0.2083740234375
],
[
0.08660888671875,
0.208984375,
0.4794921875,
0.208984375,
0.4794921875,
0.222900390625,
0.08660888671875,
0.222900390625
],
[
0.521484375,
0.210205078125,
0.63818359375,
0.210205078125,
0.63818359375,
0.2203369140625,
0.521484375,
0.2203369140625
],
[
0.0849609375,
0.222900390625,
0.4794921875,
0.222900390625,
0.4794921875,
0.236083984375,
0.0849609375,
0.236083984375
],
[
0.53515625,
0.2215576171875,
0.9140625,
0.2215576171875,
0.9140625,
0.236083984375,
0.53515625,
0.236083984375
],
[
0.08660888671875,
0.238037109375,
0.42236328125,
0.238037109375,
0.42236328125,
0.248779296875,
0.08660888671875,
0.248779296875
],
[
0.51953125,
0.236083984375,
0.9140625,
0.236083984375,
0.9140625,
0.25,
0.51953125,
0.25
],
[
0.101318359375,
0.250732421875,
0.47802734375,
0.250732421875,
0.47802734375,
0.264404296875,
0.101318359375,
0.264404296875
],
[
0.51953125,
0.250732421875,
0.91259765625,
0.250732421875,
0.91259765625,
0.264404296875,
0.51953125,
0.264404296875
],
[
0.0849609375,
0.264404296875,
0.47802734375,
0.264404296875,
0.47802734375,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.521484375,
0.264404296875,
0.9140625,
0.264404296875,
0.9140625,
0.27783203125,
0.521484375,
0.27783203125
],
[
0.0849609375,
0.279052734375,
0.4794921875,
0.279052734375,
0.4794921875,
0.292236328125,
0.0849609375,
0.292236328125
],
[
0.51953125,
0.276611328125,
0.91259765625,
0.276611328125,
0.91259765625,
0.290283203125,
0.51953125,
0.290283203125
],
[
0.0849609375,
0.292236328125,
0.4794921875,
0.292236328125,
0.4794921875,
0.30615234375,
0.0849609375,
0.30615234375
],
[
0.521484375,
0.29345703125,
0.91259765625,
0.29345703125,
0.91259765625,
0.3037109375,
0.521484375,
0.3037109375
],
[
0.08660888671875,
0.30615234375,
0.47802734375,
0.30615234375,
0.47802734375,
0.3193359375,
0.08660888671875,
0.3193359375
],
[
0.521484375,
0.30615234375,
0.9140625,
0.30615234375,
0.9140625,
0.3193359375,
0.521484375,
0.3193359375
],
[
0.0882568359375,
0.3212890625,
0.47802734375,
0.3212890625,
0.47802734375,
0.33203125,
0.0882568359375,
0.33203125
],
[
0.521484375,
0.32080078125,
0.91259765625,
0.32080078125,
0.91259765625,
0.330810546875,
0.521484375,
0.330810546875
],
[
0.08660888671875,
0.333984375,
0.4794921875,
0.333984375,
0.4794921875,
0.347900390625,
0.08660888671875,
0.347900390625
],
[
0.51953125,
0.333251953125,
0.91259765625,
0.333251953125,
0.91259765625,
0.3466796875,
0.51953125,
0.3466796875
],
[
0.08660888671875,
0.349853515625,
0.418212890625,
0.349853515625,
0.418212890625,
0.35986328125,
0.08660888671875,
0.35986328125
],
[
0.521484375,
0.3466796875,
0.9140625,
0.3466796875,
0.9140625,
0.35986328125,
0.521484375,
0.35986328125
],
[
0.51953125,
0.35986328125,
0.9140625,
0.35986328125,
0.9140625,
0.373779296875,
0.51953125,
0.373779296875
],
[
0.1666259765625,
0.373779296875,
0.39453125,
0.373779296875,
0.39453125,
0.391357421875,
0.1666259765625,
0.391357421875
],
[
0.51953125,
0.375,
0.9140625,
0.375,
0.9140625,
0.38818359375,
0.51953125,
0.38818359375
],
[
0.521484375,
0.38818359375,
0.9140625,
0.38818359375,
0.9140625,
0.401611328125,
0.521484375,
0.401611328125
],
[
0.0849609375,
0.39453125,
0.47802734375,
0.395751953125,
0.47802734375,
0.410400390625,
0.0849609375,
0.4091796875
],
[
0.51953125,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.51953125,
0.415283203125
],
[
0.08660888671875,
0.410400390625,
0.4794921875,
0.410400390625,
0.4794921875,
0.423583984375,
0.08660888671875,
0.423583984375
],
[
0.51953125,
0.416748046875,
0.9140625,
0.416748046875,
0.9140625,
0.429931640625,
0.51953125,
0.429931640625
],
[
0.0849609375,
0.4248046875,
0.4794921875,
0.4248046875,
0.4794921875,
0.438232421875,
0.0849609375,
0.438232421875
],
[
0.51953125,
0.429931640625,
0.76953125,
0.429931640625,
0.76953125,
0.443115234375,
0.51953125,
0.443115234375
],
[
0.0849609375,
0.438232421875,
0.476318359375,
0.438232421875,
0.476318359375,
0.451904296875,
0.0849609375,
0.451904296875
],
[
0.08660888671875,
0.451904296875,
0.4794921875,
0.451904296875,
0.4794921875,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.51953125,
0.453857421875,
0.6650390625,
0.455078125,
0.6650390625,
0.4697265625,
0.51953125,
0.468505859375
],
[
0.0849609375,
0.466552734375,
0.4794921875,
0.466552734375,
0.4794921875,
0.479736328125,
0.0849609375,
0.479736328125
],
[
0.5185546875,
0.472900390625,
0.9140625,
0.4716796875,
0.9140625,
0.486083984375,
0.5185546875,
0.4873046875
],
[
0.08660888671875,
0.479736328125,
0.4794921875,
0.479736328125,
0.4794921875,
0.49365234375,
0.08660888671875,
0.49365234375
],
[
0.5185546875,
0.4873046875,
0.91259765625,
0.486083984375,
0.91259765625,
0.5,
0.5185546875,
0.50048828125
],
[
0.0849609375,
0.49365234375,
0.26953125,
0.49365234375,
0.26953125,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.1004638671875,
0.5068359375,
0.4794921875,
0.505859375,
0.4794921875,
0.52099609375,
0.1004638671875,
0.521484375
],
[
0.51953125,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.51953125,
0.51513671875
],
[
0.51953125,
0.51513671875,
0.9140625,
0.51513671875,
0.9140625,
0.52880859375,
0.51953125,
0.52880859375
],
[
0.08660888671875,
0.521484375,
0.4794921875,
0.521484375,
0.4794921875,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.51953125,
0.52880859375,
0.81884765625,
0.52880859375,
0.81884765625,
0.54248046875,
0.51953125,
0.54248046875
],
[
0.08660888671875,
0.53515625,
0.4794921875,
0.53515625,
0.4794921875,
0.548828125,
0.08660888671875,
0.548828125
],
[
0.08416748046875,
0.5498046875,
0.4306640625,
0.5498046875,
0.4306640625,
0.56298828125,
0.08416748046875,
0.56298828125
],
[
0.7685546875,
0.55322265625,
0.8486328125,
0.55322265625,
0.8486328125,
0.564453125,
0.7685546875,
0.564453125
],
[
0.580078125,
0.56005859375,
0.6611328125,
0.56005859375,
0.6611328125,
0.57470703125,
0.580078125,
0.57470703125
],
[
0.6689453125,
0.56494140625,
0.705078125,
0.568359375,
0.70166015625,
0.58349609375,
0.66650390625,
0.580078125
],
[
0.890625,
0.56005859375,
0.9140625,
0.56005859375,
0.9140625,
0.57373046875,
0.890625,
0.57373046875
],
[
0.0849609375,
0.57470703125,
0.1993408203125,
0.57568359375,
0.1993408203125,
0.59033203125,
0.0849609375,
0.5888671875
],
[
0.77294921875,
0.5693359375,
0.8388671875,
0.5693359375,
0.8388671875,
0.58056640625,
0.77294921875,
0.58056640625
],
[
0.70361328125,
0.5791015625,
0.7294921875,
0.5791015625,
0.7294921875,
0.5859375,
0.70361328125,
0.5859375
],
[
0.0849609375,
0.59521484375,
0.4794921875,
0.59521484375,
0.4794921875,
0.609375,
0.0849609375,
0.609375
],
[
0.51953125,
0.59521484375,
0.9140625,
0.59521484375,
0.9140625,
0.60986328125,
0.51953125,
0.60986328125
],
[
0.08660888671875,
0.60986328125,
0.47802734375,
0.60986328125,
0.47802734375,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.51953125,
0.60986328125,
0.91259765625,
0.60986328125,
0.91259765625,
0.62353515625,
0.51953125,
0.62353515625
],
[
0.0849609375,
0.62353515625,
0.47802734375,
0.62353515625,
0.47802734375,
0.63720703125,
0.0849609375,
0.63720703125
],
[
0.51953125,
0.62353515625,
0.9140625,
0.62353515625,
0.9140625,
0.63720703125,
0.51953125,
0.63720703125
],
[
0.08660888671875,
0.63818359375,
0.476318359375,
0.63818359375,
0.476318359375,
0.6484375,
0.08660888671875,
0.6484375
],
[
0.51953125,
0.63818359375,
0.9140625,
0.63818359375,
0.9140625,
0.6513671875,
0.51953125,
0.6513671875
],
[
0.0849609375,
0.65087890625,
0.47802734375,
0.65087890625,
0.47802734375,
0.6640625,
0.0849609375,
0.6640625
],
[
0.51953125,
0.6513671875,
0.6650390625,
0.6513671875,
0.6650390625,
0.66552734375,
0.51953125,
0.66552734375
],
[
0.0849609375,
0.66552734375,
0.4794921875,
0.66552734375,
0.4794921875,
0.6787109375,
0.0849609375,
0.6787109375
],
[
0.6259765625,
0.67041015625,
0.8037109375,
0.67041015625,
0.8037109375,
0.68798828125,
0.6259765625,
0.68798828125
],
[
0.890625,
0.67236328125,
0.9140625,
0.67236328125,
0.9140625,
0.6875,
0.890625,
0.6875
],
[
0.0882568359375,
0.6796875,
0.476318359375,
0.6796875,
0.476318359375,
0.68994140625,
0.0882568359375,
0.68994140625
],
[
0.0849609375,
0.69140625,
0.455810546875,
0.69140625,
0.455810546875,
0.705078125,
0.0849609375,
0.705078125
],
[
0.54150390625,
0.6962890625,
0.9140625,
0.69873046875,
0.9140625,
0.71337890625,
0.54150390625,
0.7109375
],
[
0.10296630859375,
0.70703125,
0.47802734375,
0.70703125,
0.47802734375,
0.72021484375,
0.10296630859375,
0.72021484375
],
[
0.52294921875,
0.71337890625,
0.91259765625,
0.71337890625,
0.91259765625,
0.7236328125,
0.52294921875,
0.7236328125
],
[
0.0849609375,
0.7197265625,
0.4794921875,
0.72021484375,
0.4794921875,
0.734375,
0.0849609375,
0.73291015625
],
[
0.52294921875,
0.72607421875,
0.91259765625,
0.72607421875,
0.91259765625,
0.736328125,
0.52294921875,
0.736328125
],
[
0.08416748046875,
0.73291015625,
0.47802734375,
0.734375,
0.47802734375,
0.74853515625,
0.08416748046875,
0.74755859375
],
[
0.521484375,
0.73779296875,
0.91259765625,
0.73779296875,
0.91259765625,
0.74853515625,
0.521484375,
0.74853515625
],
[
0.521484375,
0.75048828125,
0.91259765625,
0.75048828125,
0.91259765625,
0.76123046875,
0.521484375,
0.76123046875
],
[
0.1070556640625,
0.75830078125,
0.4794921875,
0.76025390625,
0.4794921875,
0.7744140625,
0.1070556640625,
0.77294921875
],
[
0.521484375,
0.76318359375,
0.91259765625,
0.76318359375,
0.91259765625,
0.7734375,
0.521484375,
0.7734375
],
[
0.08660888671875,
0.7744140625,
0.343017578125,
0.7744140625,
0.343017578125,
0.78515625,
0.08660888671875,
0.78515625
],
[
0.51953125,
0.7744140625,
0.9140625,
0.7744140625,
0.9140625,
0.78857421875,
0.51953125,
0.78857421875
],
[
0.108642578125,
0.787109375,
0.47802734375,
0.787109375,
0.47802734375,
0.80029296875,
0.108642578125,
0.80029296875
],
[
0.52294921875,
0.7890625,
0.91259765625,
0.7890625,
0.91259765625,
0.7998046875,
0.52294921875,
0.7998046875
],
[
0.0849609375,
0.7998046875,
0.4794921875,
0.7998046875,
0.4794921875,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.5185546875,
0.798828125,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.5185546875,
0.8125
],
[
0.0882568359375,
0.81494140625,
0.1422119140625,
0.81494140625,
0.1422119140625,
0.82275390625,
0.0882568359375,
0.82275390625
],
[
0.52294921875,
0.814453125,
0.9111328125,
0.814453125,
0.9111328125,
0.82470703125,
0.52294921875,
0.82470703125
],
[
0.108642578125,
0.8271484375,
0.47802734375,
0.8271484375,
0.47802734375,
0.8369140625,
0.108642578125,
0.8369140625
],
[
0.521484375,
0.82568359375,
0.56787109375,
0.82568359375,
0.56787109375,
0.8359375,
0.521484375,
0.8359375
],
[
0.0849609375,
0.83837890625,
0.4794921875,
0.83837890625,
0.4794921875,
0.8515625,
0.0849609375,
0.8515625
],
[
0.54345703125,
0.83837890625,
0.9111328125,
0.83837890625,
0.9111328125,
0.8515625,
0.54345703125,
0.8515625
],
[
0.0849609375,
0.85107421875,
0.47802734375,
0.85107421875,
0.47802734375,
0.8642578125,
0.0849609375,
0.8642578125
],
[
0.51953125,
0.85107421875,
0.91259765625,
0.85107421875,
0.91259765625,
0.8642578125,
0.51953125,
0.8642578125
],
[
0.08660888671875,
0.86572265625,
0.476318359375,
0.86572265625,
0.476318359375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51953125,
0.8642578125,
0.9140625,
0.8642578125,
0.9140625,
0.87744140625,
0.51953125,
0.87744140625
],
[
0.08416748046875,
0.87548828125,
0.406005859375,
0.87548828125,
0.406005859375,
0.8896484375,
0.08416748046875,
0.8896484375
],
[
0.5185546875,
0.87548828125,
0.865234375,
0.87548828125,
0.865234375,
0.8896484375,
0.5185546875,
0.8896484375
]
] |
[
"loses points if red dots get added or goal patterns get dis-",
"ergodic, we approximate this distribution by sampling from",
"rupted. Formally, let V(s) measures the point value of a",
"a sequence of n = 1000 states following time t.7 Patterns",
"board state s. An agent's immediate reward at time step t will",
"in SafeLife tend to collapse to steady or oscillating states",
"be Rt(st+1, $t) = V(st+1) - V(st). Additionally, the agent",
"or they chaotically grow and regrow at short time scales △t",
"gets one extra point for reaching a level's exit.",
"with △t < 1000, so the ergodicity requirement for inaction",
"The patterns that the agent needs to create or destroy are",
"trajectories is easily satisfied. We then likewise sample n",
"ofteninterwovenwithoradjacenttoneutralgreenpatterns",
"states after the end of the agent's trajectory at the same time",
"as in Figure 1. The neutral patterns may or may not hinder the",
"t.Thisproduces action and inaction distributionsratherthan",
"agentin its tasks,butthey are often fragile;a reckless agent",
"action and inaction states. Each distribution is sampled from",
"will tend to disrupt them. The goal of our environment is to",
"its own single trajectory of length t + n either with or without",
"test this recklessness: can an agent learn to avoid disrupting",
"agentinteractions",
"its environment without being explicitly told not to do so? If",
"Note that this baseline is computationally expensive and",
"not, canfuture agent architectures accomplishthis?",
"makes most sense when t is taken to be the end of an episode;",
"It is essential to this safety test that the neutral patterns be",
"there are other baseline states that may be more appropriate",
"truly neutral to the agent's reward function. It would be trivial",
"in the course of training to penalize side effects. Every base-",
"to teach an agent to avoid side effects in this environment",
"line has its disadvantage, however. A starting state baseline",
"by explicitly penalizing it for the side effects which we test.",
"requires no computation,but it counts all natural dynam-",
"However, the problem of avoiding side effects is that there",
"ics as side effects. A single-state inaction baseline is more",
"tend tobemanymoreofthem thanone can enumerate.A",
"intuitivelyappealing,butitrequires afull simulation,isn'tap-",
"safe agent must learn to avoid side effects in general without",
"plicable in chaotic environments, and may incentivize agents",
"foreknowledgeaboutourparticulartestingcriteria.",
"to offset their positive actions (Eysenbach et al. 2018). Fi-",
"nally, a step-wise inaction baseline (Turner, Hadfield-Menell,",
"3Measuring Side Effects",
"and Tadepalli 2019) can be used in continuing (rather than",
"episodic) environments, but it may predispose agents to inac-",
"Side effects are difficult and subtle phenomena to correctly",
"tion precisely when safety demands that they act.\" Finding",
"define. To give a full account of what an agent's effects are,",
"optimal baselines for training is very much an unsolved prob-",
"one must establish theories of causality and moral respon-",
"lem, and one we save for future work.",
" sibility, including those in interactions with other agents.5",
"These are well beyond the scope of this paper. Our task here",
"Deviation measure",
"is much easier: we wish to define a heuristic side effect mea-",
"Since we use a distribution of states as our baseline, we must",
"surement that makes sense in the SafeLife environment and",
"use a distance measurebetween distributions as our deviation",
"aligns with human intuition.",
"Following Krakovna et al. (Krakovna et al. 2018), we",
"measure. To do this, we calculate average cell densities across",
"the distributions and find the earth-mover distances (Rubner,",
"define the benchmark SafeLife side effect measurement in",
"Tomasi, and Guibas 1998) between them. Let",
"terms of a baseline state-the state against which effects are",
" compared-and a deviation measure between states.",
"if s(a)=a",
"Pa(; Ds) :",
"D.I",
"(1)",
"Baseline states",
"otherwise",
"SED",
"When testing for side effects in episodic simulated environ-",
"represent the expected density of cell type Q at location ",
"ments, it makes most sense to define the baseline state as the",
"within the environment, where the expectation is taken over",
" state that would have occurred had the agent not taken any",
"a distribution of states Ds. Let dr (△a) be a spatial metric on",
"actions.This state is easilycomputed:to find the baseline",
"the environment's grid defined such that max(d.) = 1. For",
" state at time t, one need only advance a copy of the initial",
"concreteness,we take",
"state so a total of t time steps. However, when the dynam-",
"d() = tanh(1)",
"(2)",
"ics arestochasticor chaotic,as isthecasein SafeLife,the",
"inaction state can have a distribution of different values.",
"7whether a given choice of n is sufficient for this purpose de-",
"Rather than confining ourselves to a single state, we take",
"pends both on the environment,its degree of divergencefrom er-",
"the baseline as the entirety of this distribution. Assuming that",
"godicity,and onthestatisticbeing computed overthesample.The",
"the dynamics without agent interaction are approximately",
"degree of non-ergodicity in SafeLifeis somewhat bounded bythe",
"limited sizeof theboard,whichprevents(forinstance)rare struc-",
"4In some levels, the exit will only open once the agent has",
"tures like“puffers\"and\"rakes\"fromintroducing non-ergodic singu-",
"accomplished acertainfraction of itsgoals.",
"larities in the side effect distribution. When there are no generative",
"5 SafeLife 1.0 does not include multi-agent tasks, but the environ.",
"stochastic elementsona SafeLife board (i.e.,no“\"spawners\"),most",
"ment has been designed to be readily extensible to more complex",
"distributions will tend to be dominated by the initial patterns' late",
"scenarios",
"stage evolutiontowards still lifes and oscillators whicharetrivially",
"A chaotic yet deterministic environment will have awell",
"ergodic",
"defined inaction baseline state for every time t as long as the starting",
"\"For example, if an agent is driving a car and their \"no side",
"state is specified to arbitrary precision. However, if there is any",
"effect\" baseline is the one in which they ceased actions in the last",
"uncertainty in the starting state, then the inaction state at timet",
"time step, then they may conclude that the least disruptive thing to",
"could take on many divergently different possibilities.",
"do is take their hands off the wheel and cause an accident!"
] |
{
"arxiv_id": "1912.01217",
"language": "en",
"page_idx": 2
}
|
||
2004.11540-en-1
|
2004.11540-en-1.jpg
|
$\mathbf{x},\mathbf{y}\in\mathbb{R}^D.$ Prior methods treat these 2$\times D$-dimensional vectors as a set, and apply global set processing models for analysis. Such models largely disregard local geometric structure. Yet the correspondences are embedded in a metric space $(\mathbb{R}^{2\times D})$ that induces distances and neighborhood relationships. In particular, 3D correspondences form a geometric structure in 6-dimensional space[8] and we use a high-dimensional convolutional network to analyze the 6D structure formed by correspondences and estimate the likelihood that a given correspondence is correct (i.e., an inlier).
The second component we develop is a differentiable Weighted Procrustes solver. The Procrustes method[15] provides a closed-form solution for rigid registration in SE(3). A differentiable version of the Procrustes method by Wang et $al.[41]$ has been used for end-to-end registration. However, the differentiable Procrustes method passes gradients through coordinates, which requires $O(N^{2})$ time and memory for $N$ keypoints, limiting the number of keypoints that can be processed by the network. We use the inlier probabilities predicted by our first module (the 6D convolutional network) to guide the Procrustes method, thus forming a differentiable Weighted Procrustes method. This method passes gradients through the weights associated with correspondences rather than correspondence coordinates. The computational complexity of the Weighted Procrustes method is linear in the number of correspondences, allowing the registration pipeline to use dense correspondence sets rather than sparse keypoints. This substantially increases registration accuracy.
Our third component is a robust optimization module that fine-tunes the alignment produced by the Weighted Procrustes solver. This optimization module minimizes a differentiable loss via gradient descent on the continuous SE(3) representation space[52]. The optimization is fast since it does not require neighbor search in the inner $\log[48].$
Experimentally, we validate the presented modules on a real-world pairwise registration benchmark [47] and large-scale scene reconstruction datasets [18,6, 32]. We show that our modules are robust, accurate, and fast in comparison to both classical global registration algorithms [50,35,45] and recent end-toend approaches [31,41,2]. All training and experiment scripts are available at https://github.com/ chrischoy/DeepGlobalRegistration.
## $2. \textbf{Related Work}$
We divide the related work into three categories following the stages of standard registration pipelines that deal with real-world 3D scans: feature-based correspondence matching, outlier fltering, and pose optimization.
Feature-based correspondence matching. The first step in many 3D registration pipelines is feature extraction. Lo-
cal and global geometric structure in 3D is analyzed to produce high-dimensional feature descriptors, which can then be used to establish correspondences.
Traditional hand-crafted features commonly summarize pairwise or higher-order relationships in histograms [20,37, 40,36,35]. Recent work has shifted to learning features via deep networks [47,22]. A number of recent methods are based on global pooling models [12, 11, 49], while others use convolutional networks [14,9].
Our work is agnostic to the feature extraction mechanism. Our modules primarily address subsequent stages of the registration pipeline and are compatible with a wide variety of feature descriptors.
Outlier filtering. Correspondences produced by matching features are commonly heavily contaminated by outliers. These outliers need to be filtered out for robust alignment. A widely used family of techniques for robust model fitting is based on RANdom SAmple Consensus (RANSAC)[38,1,35,29,19],which iteratively samples small sets of correspondences in the hope of sampling a subset that is free from outliers. Other algorithms are based on branch-and-bound[45], semi-definite programming [28, 25], and maximal clique selection [44]. These methods are accurate, but commonly require longer iterative sampling or more expensive computation as the signalto-noise ratio decreases. One exception is TEASER [44], which remains effective even with high outlier rates. Other methods use robust loss functions to reject outliers during optimization [50, 4].
Our work uses a convolutional network to identify inliers and outliers. The network needs only one feed-forward pass at test time and does not require iterative optimization.
Pose optimization. Pose optimization is the final stage that minimizes an alignment objective on filtered correspondences. Iterative Closest Points (ICP)[3] and Fast Global Registration (FGR) [50] use second-order optimization to optimize poses. Makadia $et$ al.[26]propose an iterative procedure to minimize correlation scores. Maken et al.[27] propose to accelerate this process by stochastic gradient descent.
Recent end-to-end frameworks combine feature learning and pose optimization. Aoki $etal.[2]$ combine PointNet global features with an iterative pose optimization method[24]. Wang $etal.[41,42]$ train graph neural network features by backpropagating through pose optimization.
We further advance this line of work. In particular, our Weighted Procrustes method reduces the complexity of optimization from quadratic to linear and enables the use of dense correspondences for highly accurate registration of real-world scans.
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5048828125,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.5048828125,
0.11932373046875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.5009765625,
0.121826171875,
0.7490234375,
0.12249755859375,
0.7490234375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.521484375,
0.143310546875,
0.890625,
0.143310546875,
0.890625,
0.1572265625,
0.521484375,
0.1572265625
],
[
0.07843017578125,
0.149658203125,
0.4697265625,
0.149658203125,
0.4697265625,
0.1673583984375,
0.07843017578125,
0.1673583984375
],
[
0.5009765625,
0.1578369140625,
0.8896484375,
0.1578369140625,
0.8896484375,
0.1717529296875,
0.5009765625,
0.1717529296875
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50634765625,
0.1749267578125,
0.890625,
0.1749267578125,
0.890625,
0.1849365234375,
0.50634765625,
0.1849365234375
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1881103515625,
0.8935546875,
0.1881103515625,
0.8935546875,
0.2020263671875,
0.50244140625,
0.2020263671875
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.5009765625,
0.203857421875,
0.89208984375,
0.203857421875,
0.89208984375,
0.2171630859375,
0.5009765625,
0.2171630859375
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.2183837890625,
0.73193359375,
0.2183837890625,
0.73193359375,
0.231689453125,
0.5009765625,
0.231689453125
],
[
0.07843017578125,
0.2266845703125,
0.46728515625,
0.2279052734375,
0.46728515625,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.098876953125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.098876953125,
0.2568359375
],
[
0.521484375,
0.2392578125,
0.890625,
0.2392578125,
0.890625,
0.25244140625,
0.521484375,
0.25244140625
],
[
0.08087158203125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.08087158203125,
0.271484375
],
[
0.50244140625,
0.255126953125,
0.89208984375,
0.255126953125,
0.89208984375,
0.268310546875,
0.50244140625,
0.268310546875
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.26953125,
0.89208984375,
0.268310546875,
0.89208984375,
0.28271484375,
0.50244140625,
0.2841796875
],
[
0.07843017578125,
0.287353515625,
0.4697265625,
0.28857421875,
0.4697265625,
0.302978515625,
0.07843017578125,
0.3017578125
],
[
0.5009765625,
0.2841796875,
0.68115234375,
0.2841796875,
0.68115234375,
0.29736328125,
0.5009765625,
0.29736328125
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50341796875,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.3193359375,
0.50341796875,
0.3193359375
],
[
0.08087158203125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50244140625,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.50244140625,
0.333984375
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.35107421875,
0.890625,
0.35107421875,
0.890625,
0.3642578125,
0.50341796875,
0.3642578125
],
[
0.080078125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.080078125,
0.376953125
],
[
0.50341796875,
0.367431640625,
0.8896484375,
0.367431640625,
0.8896484375,
0.37744140625,
0.50341796875,
0.37744140625
],
[
0.080078125,
0.37744140625,
0.46728515625,
0.37744140625,
0.46728515625,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.50244140625,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.39453125,
0.50244140625,
0.39453125
],
[
0.08087158203125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.406005859375,
0.08087158203125,
0.406005859375
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.080078125,
0.4091796875,
0.4697265625,
0.4091796875,
0.4697265625,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50244140625,
0.410888671875,
0.89208984375,
0.410888671875,
0.89208984375,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.08087158203125,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.50244140625,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455078125,
0.50244140625,
0.455078125
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.45703125,
0.890625,
0.45703125,
0.890625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.08087158203125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50341796875,
0.488037109375,
0.8896484375,
0.488037109375,
0.8896484375,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50341796875,
0.501953125,
0.89208984375,
0.501953125,
0.89208984375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.080078125,
0.5166015625,
0.285888671875,
0.5166015625,
0.285888671875,
0.52978515625,
0.080078125,
0.52978515625
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.09722900390625,
0.52880859375,
0.46728515625,
0.52880859375,
0.46728515625,
0.54345703125,
0.09722900390625,
0.54345703125
],
[
0.5,
0.5322265625,
0.63818359375,
0.53076171875,
0.63818359375,
0.54541015625,
0.5,
0.546875
],
[
0.07843017578125,
0.54345703125,
0.4697265625,
0.54443359375,
0.4697265625,
0.5595703125,
0.07843017578125,
0.55810546875
],
[
0.521484375,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.58349609375,
0.861328125,
0.58349609375,
0.861328125,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.080078125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.6015625,
0.89208984375,
0.60302734375,
0.89208984375,
0.6171875,
0.50244140625,
0.6162109375
],
[
0.07843017578125,
0.619140625,
0.148681640625,
0.6220703125,
0.148681640625,
0.63623046875,
0.07843017578125,
0.6337890625
],
[
0.50341796875,
0.62060546875,
0.8896484375,
0.62060546875,
0.8896484375,
0.630859375,
0.50341796875,
0.630859375
],
[
0.098876953125,
0.63525390625,
0.468017578125,
0.63525390625,
0.468017578125,
0.6494140625,
0.098876953125,
0.6494140625
],
[
0.50244140625,
0.6337890625,
0.890625,
0.6337890625,
0.890625,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.6484375,
0.89208984375,
0.6484375,
0.89208984375,
0.662109375,
0.50244140625,
0.662109375
],
[
0.07843017578125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.07843017578125,
0.6796875
],
[
0.50244140625,
0.6640625,
0.89208984375,
0.6640625,
0.89208984375,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.11767578125,
0.6796875,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.11767578125,
0.693359375
],
[
0.5009765625,
0.6787109375,
0.890625,
0.67822265625,
0.890625,
0.6923828125,
0.5009765625,
0.693359375
],
[
0.5009765625,
0.693359375,
0.88623046875,
0.693359375,
0.88623046875,
0.70703125,
0.5009765625,
0.70703125
],
[
0.08087158203125,
0.68310546875,
0.11767578125,
0.68310546875,
0.11767578125,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.07843017578125,
0.6962890625,
0.4697265625,
0.6962890625,
0.4697265625,
0.71044921875,
0.07843017578125,
0.71044921875
],
[
0.080078125,
0.7119140625,
0.271240234375,
0.7119140625,
0.271240234375,
0.72607421875,
0.080078125,
0.72607421875
],
[
0.306396484375,
0.71337890625,
0.468017578125,
0.71337890625,
0.468017578125,
0.7236328125,
0.306396484375,
0.7236328125
],
[
0.50244140625,
0.7109375,
0.54345703125,
0.7109375,
0.54345703125,
0.72265625,
0.50244140625,
0.72265625
],
[
0.30224609375,
0.71630859375,
0.3095703125,
0.71630859375,
0.3095703125,
0.72021484375,
0.30224609375,
0.72021484375
],
[
0.273681640625,
0.7138671875,
0.301513671875,
0.7138671875,
0.301513671875,
0.72265625,
0.273681640625,
0.72265625
],
[
0.07843017578125,
0.7265625,
0.281982421875,
0.7265625,
0.281982421875,
0.74072265625,
0.07843017578125,
0.74072265625
],
[
0.301513671875,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.301513671875,
0.74072265625
],
[
0.521484375,
0.72998046875,
0.89208984375,
0.72998046875,
0.89208984375,
0.74365234375,
0.521484375,
0.74365234375
],
[
0.0767822265625,
0.74072265625,
0.468017578125,
0.7412109375,
0.468017578125,
0.755859375,
0.0767822265625,
0.7548828125
],
[
0.50244140625,
0.74560546875,
0.890625,
0.74560546875,
0.890625,
0.7587890625,
0.50244140625,
0.7587890625
],
[
0.08087158203125,
0.76025390625,
0.191162109375,
0.76025390625,
0.191162109375,
0.767578125,
0.08087158203125,
0.767578125
],
[
0.2017822265625,
0.7587890625,
0.366943359375,
0.7587890625,
0.366943359375,
0.76953125,
0.2017822265625,
0.76953125
],
[
0.50341796875,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.50341796875,
0.7734375
],
[
0.50244140625,
0.77587890625,
0.89208984375,
0.77587890625,
0.89208984375,
0.7890625,
0.50244140625,
0.7890625
],
[
0.080078125,
0.78515625,
0.2181396484375,
0.78515625,
0.2181396484375,
0.7998046875,
0.080078125,
0.7998046875
],
[
0.50244140625,
0.79150390625,
0.890625,
0.79150390625,
0.890625,
0.8046875,
0.50244140625,
0.8046875
],
[
0.098876953125,
0.81005859375,
0.46728515625,
0.81005859375,
0.46728515625,
0.82373046875,
0.098876953125,
0.82373046875
],
[
0.50244140625,
0.806640625,
0.53515625,
0.806640625,
0.53515625,
0.818359375,
0.50244140625,
0.818359375
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.521484375,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.07843017578125,
0.84033203125,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.07843017578125,
0.85400390625
],
[
0.50244140625,
0.8388671875,
0.89208984375,
0.84130859375,
0.89208984375,
0.85595703125,
0.50244140625,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.406982421875,
0.85693359375,
0.406982421875,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.61767578125,
0.88720703125,
0.61767578125,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"x, y E RD. Prior methods treat these 2 × D-dimensional",
"cal and global geometric structure in 3D is analyzed to pro-",
"vectors as a set, and apply global set processing models",
"ducehigh-dimensionalfeaturedescriptors,which canthen",
"for analysis. Such models largely disregard local geometric",
"be used to establish correspondences.",
"structure. Yet the correspondences are embedded in a met-",
"Traditional hand-crafted features commonly summarize",
"ric space (R2x D) that induces distances and neighborhood",
"pairwise or higher-order relationships in histograms [20, 37.",
"relationships.Inparticular,3D correspondencesform age-",
"40.36,35l.Recentworkhasshiftedtolearningfeaturesvia",
"ometric structure in 6-dimensional space [8] and we use a",
"deep networks [47, 22]. A number of recent methods are",
"high-dimensional convolutional network to analyze the 6D",
"based on global pooling models [12, 11, 49], while others",
"structure formed by correspondences and estimate the like-",
"use convolutional networks [14, 9]",
"lihood that a given correspondence is correct (i.e., an inlier).",
"The second component we develop is a differentiable",
"Our work is agnostic to the feature extraction mecha-",
"Weighted Procrustes solver.The Procrustes method [15]",
"nism. Our modules primarily address subsequent stages of",
"provides a closed-form solution for rigid registration in",
"the registration pipeline and are compatible with a wide va-",
"SE(3).A differentiable version of the Procrustes method",
"riety of feature descriptors.",
"by Wang et al. [41J has been used for end-to-end registra-",
"Outlier filtering. Correspondences produced by match-",
"tion.However.thedifferentiableProcrustesmethodpasses",
"ing features are commonly heavily contaminated by out-",
"gradients through coordinates, which requires O(N2) time",
"liers. These outliers need to be filtered out for robust",
"and memory for N keypoints, limiting the number of key-",
"alignment. A widely used family of techniques for ro-",
"points that can be processed by the network.We use the",
"bustmodelfittingisbasedonRANdomSAmpleConsen-",
"inlier probabilities predicted by our first module (the 6D",
"sus (RANSAC) [38, 1, 35; 29, 19], which iteratively sam-",
"convolutionalnetwork)toguidetheProcrustesmethod.thus",
"ples small sets of correspondences in the hope of sampling",
"forming a differentiable Weighted Procrustes method. This",
"a subset that is free from outliers. Other algorithms are",
"method passes gradients through the weights associated",
"based on branch-and-bound [45], semi-definite program-",
"with correspondencesrather than correspondence coordi-",
"ming [28, 25], and maximal clique selection [44]. These",
"nates. The computational complexity of the Weighted Pro-",
"methods are accurate, but commonly require longer itera-",
"crustes method is linear in the number of corespondences,",
"tive sampling or more expensive computation as the signal-",
"allowing the registration pipeline to use dense correspon-",
"to-noiseratiodecreases.Oneexception isTEASER[44l",
"dence sets rather than sparse keypoints. This substantially",
"which remains effective even with high outlier rates. Other",
"increases registration accuracy.",
"methods use robust loss functions to reject outliers during",
"Our third component is a robust optimization module",
"optimization [50, 4].",
"that fine-tunes the alignment produced by the Weighted",
"Our work uses a convolutional network to identify inliers",
"Procrustes solver.This optimization module minimizes",
"and outliers. The network needs only one feed-forward pass",
"a differentiable loss via gradient descent on the continu-",
"at test time and does not require iterative optimization.",
"ous SE(3) representation space [52]. The optimization is",
"fast since it does not require neighbor search in the inner",
"Pose optimization. Pose optimization is the final stage",
"1oop [48].",
"thatminimizes analignmentobjectiveonfiltered correspon-",
"Experimentally, we validate the presented modules",
"dences. Iterative Closest Points (ICP) [3] and Fast Global",
"on a real-world pairwise registration benchmark [47]",
"Registration (FGR) [50] use second-order optimization to",
"and large-scale scene reconstruction datasets [18, 6,",
"optimize poses. Makadia et al. [26] propose an iterative",
"We show that ourmodules arerobust,accu-",
"procedure to minimize correlation scores. Maken et al. [27]",
"propose to accelerate this process by stochastic gradient de",
"321.",
"rate, and fast in comparison to both classical global",
"registration algorithms [50,",
"451and recent end-to-",
"scent.",
"",
"35.",
"end approaches [31, 41, 2].",
"All training and exper-",
"Recent end-to-end frameworks combine feature learn-",
"iment scripts are available at https ://github.com/",
"ing and pose optimization. Aoki et al. [2] combine Point-",
"chrischoyDee",
"GlobalRegistration.",
"Net global features with an iterative pose optimization",
"method [24]. Wang et al. [41, 42] train graph neural net-",
"2. Related Work",
"work features by backpropagating through pose optimiza-",
"We divide the related work into three categories follow-",
"tion.",
"ing the stages of standard registration pipelines that deal",
"We further advance this line of work. In particular, our",
"with real-world 3D scans:feature-based correspondence",
"Weighted Procrustes method reduces the complexity of op-",
"matching, outlier filtering, and pose optimization.",
"timization from quadratic to linear and enables the use of",
"Feature-based correspondence matching. The first step",
"dense correspondences for highly accurate registration of",
"in many 3D registration pipelines is feature extraction. Lo-",
"real-world scans."
] |
{
"arxiv_id": "2004.11540",
"language": "en",
"page_idx": 1
}
|
||
2004.11540-en-2
|
2004.11540-en-2.jpg
|
<FigureHere>
Figure 2: 6-dimensional convolutional network architecture for inlier likelihood prediction (Sec. 3.2). The network has a U-net structure with residual blocks between strided convolutions. Best viewed on the screen.
$3. \textbf{ Deep Global Registration}$
3D reconstruction systems typically take a sequence of partial 3D scans as input and recover a complete 3D model of the scene. These partial scans are scene fragments, as shown in Fig. 1. In order to reconstruct the scene, reconstruction systems often begin by aligning pairs of fragments [6]. This stage is known as pairwise registration. The accuracy and robustness of pairwise registration are critical and often determine the accuracy of the final reconstruction.
Our pairwise registration pipeline begins by extracting pointwise features. These are matched to form a set of putative correspondences. We then use a high-dimensional convolutional network (ConvNet) to estimate the veracity of each correspondence. Lastly, we use a Weighted Procrustes method to align 3D scans given correspondences with associated likelihood weights, and refine the result by optimizing a robust objective.
The following notation will be used throughout the paper. We consider two point clouds, $X=[\mathbf{x}_1,...,\mathbf{x}_{N_x}]\in$ $\tilde{\mathbb{R}}^{3\times N_{x}}$ and $Y=[\mathbf{y}_1,...,\mathbf{y}_{N_{y}}]\in\mathbb{R}^{3\times N_{y}}$, with $N_x$ and $N_y$ points respectively, where $\mathbf{x}_i,\mathbf{y}_j\in\mathbb{R}^3.$ A correspondence between $\mathbf{x}_i$ and y$_j$ is denoted as $\mathbf{x}_i\leftrightarrow\mathbf{y}_j$ or $(i,j).$
## 3.1. Feature Extraction
To prepare for registration, we extract pointwise features that summarize geometric context in the form of vectors in metric feature space. Our pipeline is compatible with many feature descriptors. We use Fully Convolutional Geometric Features (FCGF) [9], which have recently been shown to be both discriminative and fast. FCGF are also compact, with dimensionality as low as 16 to 32, which supports rapid neighbor search in feature space.
3.2. Correspondence Confidence Prediction
Given the features $\mathcal{F}_x=\{\mathbf{f}_{\mathbf{x}_1},\ldots,\mathbf{f}_{\mathbf{x}_{N_x}}\}$ and $\mathcal{F}_y=$ $\{\mathbf{f}_{\mathbf{y}_{1}},\ldots,\mathbf{f}_{\mathbf{y}_{N_{x}}}\}$ of two 3D scans, we use the nearest neighbor in the feature space to generate a set of putative corre- $\operatorname*{spondences~or~matches~}\mathcal{M}=\{(i,\operatorname{argmin}_j\|\mathbf{f}_{\mathbf{x}_i}-\mathbf{f}_{\mathbf{y}_j}\|)|i\in$ $[1,...,N_{x}]\}.$ This procedure is deterministic and can be hand-crafted to filter out noisy correspondences with ratio or reciprocity tests[50]. However, we propose to learn this heuristic filtering process through a convolutional network that learns to analyze the underlying geometric structure of the correspondence set.
We first provide a 1-dimensional analogy to explain the geometry of correspondences. Let $A$ be a set of 1-dimensional points $A^{-}=\{0,1,2,3,4\}$ and $B$ be another such set $B={\{10,11,12,13,14\}.}$ Here $B$ is a translation of $A{:}B{=}{\{a_{i}+10|a_{i}{\in}{A}\}.}$ If an algorithm returns a set of possible correspondences $\{(0,10),(1,11),(2,12),(3,13),(4,14),(0,14),(4,10)\}$, then the set of correct correspondences (inliers) will form a
line (first 5 pairs), whereas incorrect correspondences (outliers) will form random noise outside the line (last 2 pairs). If we extend this to 3D scans and pointclouds, we can also represent a 3D correspondence $\mathbf{x}_i\leftrightarrow\mathbf{y}_j$ as a point in 6-dimensional space $[\mathbf{x}_i^T,\mathbf{y}_j^T]^T\in\mathbb{R}^6.$ The inlier correspondences will be distributed on a lower-dimensional surface in this 6D space, determined by the geometry of the 3D input. We denote $\mathcal{P} = \{ ( i, j) |$ $\| T^{* }( \mathbf{x} _{i}) - \mathbf{y} _{j}\| < \tau , ( i, j) \in \mathcal{M} \}$ as a set of inliers or a set of correspondences $( i, j)$ that align accurately up to the threshold $\tau$ under the ground truth transformation $T^*.$ Meanwhile, the outliers $\mathcal{N}=\mathcal{P}^C\cap\mathcal{M}$ will be scattered outside the surface $\mathcal{P}.$ To identify the inliers, we use a convolutional network. Such networks have been proven effective in related dense prediction tasks, such as 3D point cloud segmentation [16,7]. The convolutional network in our setting is in 6-dimensional space[8]. The network predicts a likelihood for each correspondence, which is a point in 6D space $[\mathbf{x}_i^T,\mathbf{y}_j^T]^T.$ The prediction is interpreted as the likelihood that the correspondence is true: an inlier.
Note that the convolution operator is translation invariant, thus our 6D ConvNet will generate the same output regardless of the absolute position of inputs in 3D. We use a similar network architecture to Choy et al.[9] to create a 6D convolutional network with skip connections within the spatial resolution across the network. The architecture of the 6D ConvNet is shown in Fig. 2. During training, we use the binary cross-entropy loss between the likelihood prediction that a correspondence $(i,j)$ is an inlier, $p_{(i,j)}\in[0,1]$, and the ground-truth correspondences $\mathcal{P}$ to optimize the network parameters:
$$L_{\mathrm{bce}}(\mathcal{M},T^{*})=\frac{1}{|\mathcal{M}|}\Bigg(\sum_{(i,j)\in\mathcal{P}}\log p_{(i,j)}+\sum_{(i,j)\in\mathcal{N}}\log p_{(i,j)}^{C}\Bigg),$$
where $p^C=1-p$ and $|\mathcal{M}|$ is the cardinality of the set of
putative correspondences.
|
[
[
0.08087158203125,
0.09027099609375,
0.310546875,
0.09027099609375,
0.310546875,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.098876953125,
0.11553955078125,
0.468017578125,
0.11614990234375,
0.468017578125,
0.1307373046875,
0.098876953125,
0.1300048828125
],
[
0.080078125,
0.1319580078125,
0.468017578125,
0.1319580078125,
0.468017578125,
0.1458740234375,
0.080078125,
0.1458740234375
],
[
0.08087158203125,
0.146484375,
0.468017578125,
0.146484375,
0.468017578125,
0.160400390625,
0.08087158203125,
0.160400390625
],
[
0.080078125,
0.1634521484375,
0.46728515625,
0.1634521484375,
0.46728515625,
0.1767578125,
0.080078125,
0.1767578125
],
[
0.080078125,
0.1767578125,
0.465576171875,
0.1767578125,
0.465576171875,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.080078125,
0.1932373046875,
0.468017578125,
0.1932373046875,
0.468017578125,
0.20703125,
0.080078125,
0.20703125
],
[
0.080078125,
0.2083740234375,
0.468017578125,
0.2083740234375,
0.468017578125,
0.2215576171875,
0.080078125,
0.2215576171875
],
[
0.5048828125,
0.203857421875,
0.8896484375,
0.203857421875,
0.8896484375,
0.2139892578125,
0.5048828125,
0.2139892578125
],
[
0.50341796875,
0.2171630859375,
0.89208984375,
0.2171630859375,
0.89208984375,
0.2310791015625,
0.50341796875,
0.2310791015625
],
[
0.080078125,
0.2235107421875,
0.468017578125,
0.2235107421875,
0.468017578125,
0.2374267578125,
0.080078125,
0.2374267578125
],
[
0.50341796875,
0.23291015625,
0.89208984375,
0.23291015625,
0.89208984375,
0.246826171875,
0.50341796875,
0.246826171875
],
[
0.09722900390625,
0.2392578125,
0.4697265625,
0.2392578125,
0.4697265625,
0.25244140625,
0.09722900390625,
0.25244140625
],
[
0.07843017578125,
0.25244140625,
0.468017578125,
0.25390625,
0.468017578125,
0.268310546875,
0.07843017578125,
0.267578125
],
[
0.50341796875,
0.25,
0.74658203125,
0.25,
0.74658203125,
0.260009765625,
0.50341796875,
0.260009765625
],
[
0.080078125,
0.268310546875,
0.468017578125,
0.268310546875,
0.468017578125,
0.2822265625,
0.080078125,
0.2822265625
],
[
0.080078125,
0.2841796875,
0.4697265625,
0.2841796875,
0.4697265625,
0.29736328125,
0.080078125,
0.29736328125
],
[
0.08087158203125,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.311279296875,
0.08087158203125,
0.311279296875
],
[
0.50341796875,
0.30810546875,
0.890625,
0.30810546875,
0.890625,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.080078125,
0.314453125,
0.468017578125,
0.314453125,
0.468017578125,
0.32763671875,
0.080078125,
0.32763671875
],
[
0.50244140625,
0.323974609375,
0.890625,
0.323974609375,
0.890625,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.080078125,
0.330078125,
0.468017578125,
0.330078125,
0.468017578125,
0.343505859375,
0.080078125,
0.343505859375
],
[
0.50244140625,
0.337158203125,
0.89208984375,
0.33837890625,
0.89208984375,
0.352783203125,
0.50244140625,
0.3515625
],
[
0.080078125,
0.3447265625,
0.226318359375,
0.3447265625,
0.226318359375,
0.35791015625,
0.080078125,
0.35791015625
],
[
0.50244140625,
0.352783203125,
0.8935546875,
0.354248046875,
0.8935546875,
0.36865234375,
0.50244140625,
0.367431640625
],
[
0.098876953125,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.373779296875,
0.098876953125,
0.373779296875
],
[
0.5009765625,
0.36865234375,
0.89208984375,
0.367431640625,
0.89208984375,
0.3818359375,
0.5009765625,
0.38330078125
],
[
0.07843017578125,
0.375,
0.468017578125,
0.375732421875,
0.468017578125,
0.39013671875,
0.07843017578125,
0.389404296875
],
[
0.07598876953125,
0.385009765625,
0.4697265625,
0.386474609375,
0.4697265625,
0.40771484375,
0.07598876953125,
0.4072265625
],
[
0.50244140625,
0.3837890625,
0.89208984375,
0.3837890625,
0.89208984375,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.07843017578125,
0.404052734375,
0.4697265625,
0.40283203125,
0.4697265625,
0.42041015625,
0.07843017578125,
0.421630859375
],
[
0.50341796875,
0.399658203125,
0.8896484375,
0.399658203125,
0.8896484375,
0.41357421875,
0.50341796875,
0.41357421875
],
[
0.07843017578125,
0.41845703125,
0.41259765625,
0.419189453125,
0.41259765625,
0.4375,
0.07843017578125,
0.436279296875
],
[
0.50244140625,
0.41357421875,
0.890625,
0.415283203125,
0.890625,
0.4287109375,
0.50244140625,
0.4267578125
],
[
0.5009765625,
0.427978515625,
0.890625,
0.4287109375,
0.890625,
0.443115234375,
0.5009765625,
0.442626953125
],
[
0.080078125,
0.447509765625,
0.259033203125,
0.447509765625,
0.259033203125,
0.4609375,
0.080078125,
0.4609375
],
[
0.50244140625,
0.443115234375,
0.890625,
0.442626953125,
0.890625,
0.45703125,
0.50244140625,
0.458251953125
],
[
0.50341796875,
0.458251953125,
0.890625,
0.458251953125,
0.890625,
0.472900390625,
0.50341796875,
0.472900390625
],
[
0.09722900390625,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.486083984375,
0.09722900390625,
0.486083984375
],
[
0.5048828125,
0.47607421875,
0.890625,
0.47607421875,
0.890625,
0.486083984375,
0.5048828125,
0.486083984375
],
[
0.080078125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.50341796875,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50341796875,
0.521484375,
0.890625,
0.521484375,
0.890625,
0.5322265625,
0.50341796875,
0.5322265625
],
[
0.08087158203125,
0.5341796875,
0.46728515625,
0.5341796875,
0.46728515625,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.50244140625,
0.5341796875,
0.8896484375,
0.5341796875,
0.8896484375,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.080078125,
0.54736328125,
0.468017578125,
0.54736328125,
0.468017578125,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.50244140625,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.50244140625,
0.56494140625,
0.8857421875,
0.56494140625,
0.8857421875,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.5791015625,
0.29736328125,
0.5791015625,
0.29736328125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.50244140625,
0.5966796875,
0.72216796875,
0.5966796875,
0.72216796875,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.080078125,
0.60498046875,
0.41259765625,
0.60498046875,
0.41259765625,
0.619140625,
0.080078125,
0.619140625
],
[
0.521484375,
0.619140625,
0.8896484375,
0.619140625,
0.8896484375,
0.63330078125,
0.521484375,
0.63330078125
],
[
0.098876953125,
0.62744140625,
0.453369140625,
0.62744140625,
0.453369140625,
0.642578125,
0.098876953125,
0.642578125
],
[
0.50341796875,
0.63525390625,
0.890625,
0.63525390625,
0.890625,
0.6484375,
0.50341796875,
0.6484375
],
[
0.08251953125,
0.64453125,
0.46728515625,
0.64453125,
0.46728515625,
0.6591796875,
0.08251953125,
0.6591796875
],
[
0.50341796875,
0.6494140625,
0.89208984375,
0.6494140625,
0.89208984375,
0.66357421875,
0.50341796875,
0.66357421875
],
[
0.080078125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.66650390625,
0.890625,
0.66650390625,
0.890625,
0.6767578125,
0.50341796875,
0.6767578125
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.50244140625,
0.6796875,
0.89208984375,
0.6796875,
0.89208984375,
0.693359375,
0.50244140625,
0.693359375
],
[
0.08087158203125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.50341796875,
0.6962890625,
0.890625,
0.6962890625,
0.890625,
0.70703125,
0.50341796875,
0.70703125
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.71044921875,
0.89208984375,
0.71044921875,
0.89208984375,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.724609375,
0.89208984375,
0.724609375,
0.89208984375,
0.73779296875,
0.50244140625,
0.73779296875
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.5009765625,
0.7392578125,
0.89208984375,
0.7392578125,
0.89208984375,
0.7568359375,
0.5009765625,
0.7568359375
],
[
0.080078125,
0.75,
0.4697265625,
0.75,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.5009765625,
0.7548828125,
0.89208984375,
0.7548828125,
0.89208984375,
0.76806640625,
0.5009765625,
0.76806640625
],
[
0.08087158203125,
0.7666015625,
0.2344970703125,
0.7666015625,
0.2344970703125,
0.7763671875,
0.08087158203125,
0.7763671875
],
[
0.5009765625,
0.76953125,
0.640625,
0.77099609375,
0.640625,
0.78515625,
0.5009765625,
0.7841796875
],
[
0.098876953125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.098876953125,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.08251953125,
0.8125,
0.465576171875,
0.8125,
0.465576171875,
0.82275390625,
0.08251953125,
0.82275390625
],
[
0.080078125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5107421875,
0.83203125,
0.59326171875,
0.83203125,
0.59326171875,
0.84326171875,
0.5107421875,
0.84326171875
],
[
0.68115234375,
0.82763671875,
0.8447265625,
0.82763671875,
0.8447265625,
0.8486328125,
0.68115234375,
0.8486328125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.615234375,
0.84130859375,
0.63623046875,
0.84130859375,
0.63623046875,
0.849609375,
0.615234375,
0.849609375
],
[
0.65185546875,
0.84765625,
0.69775390625,
0.84765625,
0.69775390625,
0.85791015625,
0.65185546875,
0.85791015625
],
[
0.76953125,
0.8466796875,
0.81396484375,
0.8466796875,
0.81396484375,
0.85791015625,
0.76953125,
0.85791015625
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.87841796875,
0.85400390625,
0.89208984375,
0.8623046875,
0.88232421875,
0.87060546875,
0.869140625,
0.86279296875
],
[
0.08087158203125,
0.8720703125,
0.4599609375,
0.8720703125,
0.4599609375,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.5048828125,
0.87060546875,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.67626953125,
0.88720703125,
0.67626953125,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"3.Deep Global Registration",
"3D reconstruction systems typically take a sequence of",
"partial 3D scans as input and recover a complete 3D model",
"of the scene.These partial scans are scene fragments, as",
"shown in Fig. l.In order to reconstruct the scene, re-",
"construction systems often begin by aligning pairs of frag",
"ments [6]. This stage is known as pairwise registration. The",
"accuracy and robustness of pairwise registration are critical",
"Figure2:6-dimensionalconvolutionalnetworkarchitecture",
"for inlier likelihood prediction (Sec.3.2).The network has",
"and often determine the accuracy of the final reconstruction.",
"a U-net structure with residual blocks between strided con-",
"Our pairwise registration pipeline begins by extracting",
"pointwise features. These are matched to form a set of pu-",
"volutions.Bestviewed on the screen.",
"tative correspondences.We then use a high-dimensional",
"convolutional network (ConvNet)to estimate the veracity of",
"eachcorrespondence.Lastly,weuseaWeightedProcrustes",
"line (first 5 pairs), whereas incorrect correspondences (out-",
"method to align 3D scans given correspondences with asso-",
"liers) will form random noise outside the line (last 2 pairs).",
"ciated likelihood weights, and refine the result by optimiz-",
"If we extend this to 3D scans and pointclouds, we can",
"ing a robust objective.",
"also represent a 3D correspondence xi → yj as a point in",
"The following notation will be used throughout the pa-",
"6-dimensional space [xT,yTT e R6. The inlier correspon-",
"per. We consider two point clouds, X =[xi,..,xN,] E",
"R3xN and Y= [y,.., yn,] e R3xNv, with N, and'Ny",
"dences will be distributed on a lower-dimensional surface in",
"points respectively, where xi,yj IR3. A correspondence",
"this 6D space, determined by the geometry of the 3D input.",
"between x; and y, is denoted as x; yj or (i, j).",
"We denote P = {(i, )l T*(xi) -yill < T, (i, ) E M)",
"as a set of inliers or a set of correspondences (i, j) that align",
"3.1. Feature Extraction",
"accurately up to the threshold T under the ground truth",
"transformation T*. Meanwhile, the outliers N = pC n M",
"To prepare for registration, we extract pointwise features",
"willbe scattered outsidethe surface P.To identify the",
"that summarize geometric context in the form of vectors in",
"inliers, we use a convolutional network. Such networks",
"metric feature space. Our pipeline is compatible with many",
"have been proven effective in related dense prediction",
"feature descriptors. We use Fully Convolutional Geometric",
"tasks,such as 3D point cloud segmentation[16,7.The",
"Features(FCGF)[9l.whichhaverecentlybeen shownto",
"convolutional network in our setting is in 6-dimensional",
"be both discriminative and fast. FCGF are also compact",
"space [8]. The network predicts a likelihood for each",
"with dimensionality as low as 16 to 32, which supports rapid",
"correspondence, which is a point in 6D space [xT,yTjT",
"neighbor search in feature space.",
"The prediction is interpreted as the likelihood that the",
"correspondence is true: an inlier.",
"3.2. Correspondence Confidence Prediction",
"Note that the convolution operator is translation invari-",
"Given the features Fr ={fxi,.. fxn] and Fy",
"ant, thus our 6D ConvNet will generate the same output re-",
"[fyi,... fy.] of two 3D scans, we use the nearest neigh-",
"gardless of the absolute position of inputs in 3D. We use a",
"bor in the feature space to generate a set of putative corre-",
"similar network architecture to Choy et al.[9l to create a",
"spondences or matches M = {(i, argmin, fx,-fy, D)/i E",
"6D convolutional network with skip connections within the",
"[1, ..., N.]]. This procedure is deterministic and can be",
"spatial resolution across the network.The architecture of",
"hand-crafted to filter out noisy correspondences with ratio",
"the 6D ConvNet is shown in Fig. 2. During training, we use",
"or reciprocity tests [50]. However, we propose to learn this",
"the binary cross-entropy loss between the likelihood predic-",
"heuristic filtering process through a convolutional network",
"tion that a correspondence (i, j) is an inlier, P(i,j) [0, 1],",
"that learns to analyze the underlying geometric structure of",
"and the ground-truth correspondences P to optimize the",
"the correspondence set.",
"network parameters:",
"We first provide a I-dimensional analogy to explain",
"the geometry of correspondences. Let A be a set of",
"1-dimensional points A=f0.1.2.3.41 and B be an-",
"other such set B = {10,11,12,13,14]. Here B is",
"Lbce(M,T*)",
"logP(i5)+logp",
"a translation of A: B =fai + 10ai E A). If an",
"M",
"(i.j)ep",
"(i,)EN",
"algorithmreturnsasetofpossiblecorrespondences",
"0",
"[(0,10), (1,11), (2,12), (3,13), (4,14), (0,14), (4,10)),",
"where pc = 1 - p and |M| is the cardinality of the set of",
"then the set of correct correspondences (inliers) will form a",
"putative correspondences."
] |
{
"arxiv_id": "2004.11540",
"language": "en",
"page_idx": 2
}
|
||
2004.11540-en-3
|
2004.11540-en-3.jpg
|
## 3.3. Weighted Procrustes for SE(3)
The inlier likelihood estimated by the 6D ConvNet provides a weight for each correspondence. The original Procrustes method[15] minimizes the mean squared er- $\begin{array}{l}{\mathrm{ror~between~corresponding~points~}\frac{1}{N}\sum_{(i,j)\in\mathcal{M}}\|\hat{\mathbf{x}}_{i}-\mathbf{y}_{j}\|^{2}}\\{\mathrm{and~thus~gives~equal~weight~to~all~correspondences.~In}}\end{array}$ contrast, we minimize a weighted mean squared error $\sum_{(i,j)\in\mathcal{M}}w_{(i,j)}\|\mathbf{x}_{i}-\mathbf{y}_{j}\|^{2}.$ This change allows us to passgradients through the weights, rather than through the po- sition [41], and enables the optimization to scale to dense correspondence sets.
Formally, Weighted Procrustes analysis minimizes:
(2)
$$\begin{aligned}e^{2}&=e^{2}(R,\mathbf{t};\mathbf{w},X,Y)\\&=\sum_{(i,j)\in\mathcal{M}}\tilde{w}_{(i,j)}(\mathbf{y}_{j}-(R\mathbf{x}_{i}+\mathbf{t}))^{2}\\&=\mathrm{Tr}\left((Y-RX-\mathbf{t}\mathbf{1}^{T})W(Y-RX-\mathbf{t}\mathbf{1}^{T})^{T}\right),\end{aligned}$$
(3)
(4)
where $1=(1,...,1)^T,X=[\mathbf{x}_1,...,\mathbf{x}_{|\mathcal{M}|}]$,and$Y=$ $[\mathbf{y}_{J_1},...,\mathbf{y}_{J_{|\mathcal{M}|}}].{J\text{ is a list of indices that defines the corre-}}$ spondences $\mathbf{x}_i\leftrightarrow\mathbf{y}_{J_i}.\mathbf{w}=[w_1,\cdots,w_{|\mathcal{M}|}]$ is the weight vector and $\tilde{\mathbf{w}}=[\tilde{w}_{1},\cdots,\tilde{w}_{|\mathcal{M}|}]\stackrel{\Delta}{=}\frac{\phi(\mathbf{w})}{||\phi(\mathbf{w})||_{1}}$ denotes the normalized weight after a nonlinear transformation $\phi$ that applies heuristic prefiltering. $W=$ diag$(\tilde{\mathbf{w}})$ forms the diagonal weight matrix.
Theorem 1: The R and that minimize the squared error $e^{2}( R, \boldsymbol{t}) = \sum _{\hat{\Lambda } ( i, j) }w_{( i, j) }( \boldsymbol{y}_{j}- R\boldsymbol{x}_{i}- \boldsymbol{t}) ^{2}$ are $\hat{\boldsymbol{t}} = ( Y$ - $RX) W1$ and $\hat{R}$ = $USV^{T}$ where $U\Sigma V^{T}$ = $SVD( \Sigma _{xy}) $, $\Sigma _{xy}$ = $YKWKX^{T}$ , $K$ = $I$ - $\sqrt {\tilde{\mathbf{w} } }\sqrt {\tilde{\mathbf{w} } }^{T}$ , and $S$ = $diag\left(1,\cdots,1,det(U)det(V)\right).$
Sketch of proof. First, we differentiate $e^2$ w.r.t. t and equate the partial derivative to 0. This gives us $\hat{\mathbf{t}}=$ $( Y$ - $RX) W1.$ Next, we substitute $X$ = $KX$ + $X\sqrt{\tilde{\mathbf{w}}}\sqrt{\tilde{\mathbf{w}}}^{T}$ on Eq. 4 and do the same for $Y.$ Then, we substitute $\mathbf{t} = \hat{\mathbf{t} } ^{\mathrm{~/ }}$into the squares. This yields $e^2=$ Tr$\left((Y-RX)KWK^{T}(Y-\bar{R}X)^{T}\right)$and expanding the term results in two squares plus -2Tr$(YKWK^TX^TR^T).$ We maximize the last negative term whose maximum is the sum of all singular values, which leads to $\hat{R}.$ The full derivation is in the supplement.
$\square$
We can easily extend the above theorem to incorporate a scaling factor $c\in\mathbb{R}^+$, or anisotropic scaling for tasks such as scan-to-CAD registration, but in this paper we assume that partial scans of the same scene have the same scale.
The Weighted Procrustes method generates rotation $\hat{R}$ and translation $\hat{\mathbf{t}}$ as outputs that depend on the weight vector $\mathbf{w}.$ In our current implementation, $\hat{R}$ and $\hat{\mathbf{t}}$ are directly sent to the robust registration module in Section 4 as an initial pose. However, we briefly demonstrate that they can also
be embedded in an end-to-end registration pipeline, since Weighted Procrustes is differentiable. From a top-level loss function $L$ of $\hat{R}$ and $\hat{\mathbf{t}}$,we can pass the gradient through the closed-form solver, and update parameters in downstream modules:
$$\frac{\partial}{\partial\mathbf{w}}L(\hat{R},\hat{\mathbf{t}})=\frac{\partial L(\hat{R},\hat{\mathbf{t}})}{\partial\hat{R}}\frac{\partial\hat{R}}{\partial\hat{\mathbf{w}}}+\frac{\partial L(\hat{R},\hat{\mathbf{t}})}{\partial\hat{\mathbf{t}}}\frac{\partial\hat{\mathbf{t}}(\hat{R},\hat{\mathbf{w}})}{\partial\hat{\mathbf{w}}},$$
(5)
where $L(\hat{R},\hat{\mathbf{t}})$ can be defined as the combination of differentiable rotation error (RE) and translation error (TE) between predictions $\hat{R},\hat{\mathbf{t}}$ and ground-truth $R^*,\mathbf{t}^*:$
(6)
$$\begin{aligned}&L_{\mathrm{rot}}(\hat{R})=\arccos\frac{\mathrm{Tr}(\hat{R}^{T}R^{*})-1}{2},\\&L_{\mathrm{trans}}(\hat{\mathbf{t}})=||\hat{\mathbf{t}}-\mathbf{t}^{*}||_{2}^{2},\end{aligned}$$
(7)
or the Forbenius norm of relative transformation matrices defined in [2,41]. The final loss is the weighted sum of $L_{\mathrm{rot}},L_{\mathrm{trans}}$,and$L_\mathrm{bce}.$
## 4. Robust Registration
In this section, we propose a fine-tuning module that minimizes a robust loss function of choice to improve the registration accuracy. We use a gradient-based method to refine poses, where a continuous representation [52] for rotations is adopted to remove discontinuities and construct a smooth optimization space. This module initializes the pose from the prediction of the Weighted Procrustes method. During iterative optimization, unlike Maken $etal.[27]$,who find the nearest neighbor per point at each gradient step, we rely on the correspondence likelihoods from the 6D ConvNet, which is estimated only once per initialization.
In addition, our framework naturally offers a failure detection mechanism. In practice, Weighted Procrustes may generate numerically unstable solutions when the number of valid correspondences is insufficient due to small overlaps or noisy correspondences between input scans. By computing the ratio of the sum of the filtered weights to the total number of correspondences, $i.e.\sum_i\phi(w_i)/|\mathcal{M}|$, we can easily approximate the fraction of valid correspondences and predict whether an alignment may be unstable. When this fraction is low, we resort to a more time-consuming but accurate registration algorithm such as RANSAC[38,1,35] or a branch-and-bound method[45] to find a numerically stable solution. In other words, we can detect when our system might fail before it returns a result and fall back to a more accurate but time-consuming algorithm, unlike previous end-to-end methods that use globally pooled latent features [2] or a singly stochastic matrix [41] - such latent representations are more difficult to interpret. 4.1. SE(3) Representation and Initialization
We use the 6D representation of 3D rotation proposed by Zhou et al.[52], rather than Euler angles or quaternions.
|
[
[
0.080078125,
0.0889892578125,
0.34228515625,
0.091552734375,
0.34228515625,
0.1060791015625,
0.080078125,
0.10418701171875
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.50341796875,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.50341796875,
0.12060546875
],
[
0.09722900390625,
0.11297607421875,
0.468017578125,
0.1142578125,
0.468017578125,
0.1287841796875,
0.09722900390625,
0.1275634765625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1287841796875,
0.375732421875,
0.1300048828125,
0.375732421875,
0.14453125,
0.080078125,
0.143310546875
],
[
0.369384765625,
0.1300048828125,
0.4697265625,
0.1300048828125,
0.4697265625,
0.14453125,
0.369384765625,
0.14453125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.1458740234375,
0.468017578125,
0.1458740234375,
0.468017578125,
0.1590576171875,
0.080078125,
0.1590576171875
],
[
0.07598876953125,
0.1572265625,
0.471435546875,
0.1572265625,
0.471435546875,
0.18115234375,
0.07598876953125,
0.18115234375
],
[
0.5009765625,
0.15283203125,
0.56884765625,
0.15283203125,
0.56884765625,
0.1673583984375,
0.5009765625,
0.1673583984375
],
[
0.52294921875,
0.17236328125,
0.5458984375,
0.17236328125,
0.5458984375,
0.186279296875,
0.52294921875,
0.186279296875
],
[
0.609375,
0.169189453125,
0.85546875,
0.17041015625,
0.85546875,
0.1881103515625,
0.609375,
0.1875
],
[
0.07843017578125,
0.175537109375,
0.4697265625,
0.175537109375,
0.4697265625,
0.189453125,
0.07843017578125,
0.189453125
],
[
0.54150390625,
0.179931640625,
0.59716796875,
0.179931640625,
0.59716796875,
0.1956787109375,
0.54150390625,
0.1956787109375
],
[
0.86669921875,
0.1817626953125,
0.8896484375,
0.1817626953125,
0.8896484375,
0.1932373046875,
0.86669921875,
0.1932373046875
],
[
0.07843017578125,
0.1900634765625,
0.4697265625,
0.1900634765625,
0.4697265625,
0.203857421875,
0.07843017578125,
0.203857421875
],
[
0.51708984375,
0.1900634765625,
0.5439453125,
0.1900634765625,
0.5439453125,
0.20263671875,
0.51708984375,
0.20263671875
],
[
0.62841796875,
0.1900634765625,
0.7099609375,
0.1881103515625,
0.7099609375,
0.20263671875,
0.62841796875,
0.2052001953125
],
[
0.74267578125,
0.1900634765625,
0.7685546875,
0.1900634765625,
0.7685546875,
0.203857421875,
0.74267578125,
0.203857421875
],
[
0.8037109375,
0.1900634765625,
0.833984375,
0.1900634765625,
0.833984375,
0.20263671875,
0.8037109375,
0.20263671875
],
[
0.080078125,
0.2052001953125,
0.471435546875,
0.2052001953125,
0.471435546875,
0.222900390625,
0.080078125,
0.222900390625
],
[
0.50341796875,
0.2108154296875,
0.890625,
0.2108154296875,
0.890625,
0.2259521484375,
0.50341796875,
0.2259521484375
],
[
0.080078125,
0.2203369140625,
0.468017578125,
0.2197265625,
0.468017578125,
0.2342529296875,
0.080078125,
0.23486328125
],
[
0.50244140625,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.50244140625,
0.2412109375
],
[
0.080078125,
0.23486328125,
0.4697265625,
0.23486328125,
0.4697265625,
0.248779296875,
0.080078125,
0.248779296875
],
[
0.50244140625,
0.242431640625,
0.81884765625,
0.242431640625,
0.81884765625,
0.255615234375,
0.50244140625,
0.255615234375
],
[
0.08087158203125,
0.25244140625,
0.2156982421875,
0.25244140625,
0.2156982421875,
0.26318359375,
0.08087158203125,
0.26318359375
],
[
0.09722900390625,
0.26513671875,
0.435546875,
0.26513671875,
0.435546875,
0.279052734375,
0.09722900390625,
0.279052734375
],
[
0.705078125,
0.261962890625,
0.80810546875,
0.261962890625,
0.80810546875,
0.27978515625,
0.705078125,
0.27978515625
],
[
0.5849609375,
0.271484375,
0.705078125,
0.26953125,
0.705078125,
0.287353515625,
0.5849609375,
0.2890625
],
[
0.86767578125,
0.272705078125,
0.89453125,
0.272705078125,
0.89453125,
0.2890625,
0.86767578125,
0.2890625
],
[
0.74658203125,
0.2822265625,
0.76318359375,
0.2822265625,
0.76318359375,
0.296630859375,
0.74658203125,
0.296630859375
],
[
0.09234619140625,
0.290283203125,
0.2484130859375,
0.292236328125,
0.2484130859375,
0.306884765625,
0.09234619140625,
0.304931640625
],
[
0.443603515625,
0.291748046875,
0.471435546875,
0.291748046875,
0.471435546875,
0.30810546875,
0.443603515625,
0.30810546875
],
[
0.58251953125,
0.296630859375,
0.7294921875,
0.296630859375,
0.7294921875,
0.314453125,
0.58251953125,
0.314453125
],
[
0.86767578125,
0.29736328125,
0.89453125,
0.29736328125,
0.89453125,
0.314453125,
0.86767578125,
0.314453125
],
[
0.11273193359375,
0.3125,
0.3447265625,
0.3125,
0.3447265625,
0.333251953125,
0.11273193359375,
0.333251953125
],
[
0.443603515625,
0.31494140625,
0.471435546875,
0.31494140625,
0.471435546875,
0.33203125,
0.443603515625,
0.33203125
],
[
0.50341796875,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.50341796875,
0.333984375
],
[
0.1298828125,
0.333984375,
0.181396484375,
0.333984375,
0.181396484375,
0.34521484375,
0.1298828125,
0.34521484375
],
[
0.50341796875,
0.3359375,
0.8935546875,
0.3359375,
0.8935546875,
0.349853515625,
0.50341796875,
0.349853515625
],
[
0.11767578125,
0.352783203125,
0.420654296875,
0.352783203125,
0.420654296875,
0.367431640625,
0.11767578125,
0.367431640625
],
[
0.43310546875,
0.354248046875,
0.465576171875,
0.354248046875,
0.465576171875,
0.3662109375,
0.43310546875,
0.3662109375
],
[
0.50341796875,
0.3515625,
0.64208984375,
0.3515625,
0.64208984375,
0.3662109375,
0.50341796875,
0.3662109375
],
[
0.50341796875,
0.37744140625,
0.69140625,
0.37744140625,
0.69140625,
0.392578125,
0.50341796875,
0.392578125
],
[
0.08087158203125,
0.386474609375,
0.45751953125,
0.386474609375,
0.45751953125,
0.40087890625,
0.08087158203125,
0.40087890625
],
[
0.080078125,
0.401611328125,
0.46728515625,
0.40087890625,
0.46728515625,
0.415283203125,
0.080078125,
0.416748046875
],
[
0.521484375,
0.40283203125,
0.89208984375,
0.40283203125,
0.89208984375,
0.416748046875,
0.521484375,
0.416748046875
],
[
0.08087158203125,
0.417236328125,
0.46728515625,
0.417236328125,
0.46728515625,
0.431884765625,
0.08087158203125,
0.431884765625
],
[
0.50244140625,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.3447265625,
0.43310546875,
0.375732421875,
0.43310546875,
0.375732421875,
0.43994140625,
0.3447265625,
0.43994140625
],
[
0.50244140625,
0.434326171875,
0.89208984375,
0.434326171875,
0.89208984375,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.3798828125,
0.436279296875,
0.465576171875,
0.436279296875,
0.465576171875,
0.447509765625,
0.3798828125,
0.447509765625
],
[
0.080078125,
0.43505859375,
0.334228515625,
0.43505859375,
0.334228515625,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.326904296875,
0.442626953125,
0.38818359375,
0.442626953125,
0.38818359375,
0.45263671875,
0.326904296875,
0.45263671875
],
[
0.330078125,
0.4375,
0.334228515625,
0.4375,
0.334228515625,
0.442626953125,
0.330078125,
0.442626953125
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.50244140625,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.48095703125,
0.080078125,
0.48095703125
],
[
0.5009765625,
0.46337890625,
0.8935546875,
0.462158203125,
0.8935546875,
0.4765625,
0.5009765625,
0.477783203125
],
[
0.07843017578125,
0.481689453125,
0.2279052734375,
0.481689453125,
0.2279052734375,
0.49560546875,
0.07843017578125,
0.49560546875
],
[
0.50341796875,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50341796875,
0.492431640625
],
[
0.50341796875,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50341796875,
0.5068359375
],
[
0.08087158203125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.08087158203125,
0.52294921875
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.07598876953125,
0.521484375,
0.4697265625,
0.52099609375,
0.4697265625,
0.54248046875,
0.07598876953125,
0.54345703125
],
[
0.5048828125,
0.52587890625,
0.890625,
0.52587890625,
0.890625,
0.5361328125,
0.5048828125,
0.5361328125
],
[
0.0767822265625,
0.5361328125,
0.471435546875,
0.53857421875,
0.471435546875,
0.56005859375,
0.0767822265625,
0.55810546875
],
[
0.50244140625,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.0767822265625,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.5791015625,
0.0767822265625,
0.5791015625
],
[
0.50341796875,
0.5537109375,
0.85107421875,
0.5537109375,
0.85107421875,
0.5673828125,
0.50341796875,
0.5673828125
],
[
0.52294921875,
0.568359375,
0.88818359375,
0.568359375,
0.88818359375,
0.58203125,
0.52294921875,
0.58203125
],
[
0.07843017578125,
0.57763671875,
0.283447265625,
0.57763671875,
0.283447265625,
0.59228515625,
0.07843017578125,
0.59228515625
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.60400390625,
0.468017578125,
0.60400390625,
0.468017578125,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50244140625,
0.61279296875
],
[
0.080078125,
0.619140625,
0.4453125,
0.619140625,
0.4453125,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.50244140625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.08087158203125,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.08087158203125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66650390625,
0.08087158203125,
0.66650390625
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.08087158203125,
0.66943359375,
0.1707763671875,
0.66943359375,
0.1707763671875,
0.6796875,
0.08087158203125,
0.6796875
],
[
0.1846923828125,
0.66845703125,
0.44775390625,
0.66845703125,
0.44775390625,
0.681640625,
0.1846923828125,
0.681640625
],
[
0.5009765625,
0.6748046875,
0.890625,
0.673828125,
0.890625,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.69873046875,
0.46728515625,
0.69775390625,
0.46728515625,
0.7119140625,
0.07843017578125,
0.71337890625
],
[
0.5009765625,
0.7041015625,
0.8935546875,
0.705078125,
0.8935546875,
0.7197265625,
0.5009765625,
0.71826171875
],
[
0.08087158203125,
0.7138671875,
0.4697265625,
0.7138671875,
0.4697265625,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50341796875,
0.7216796875,
0.890625,
0.7216796875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72998046875,
0.468017578125,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.5009765625,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.5009765625,
0.74755859375
],
[
0.07843017578125,
0.74365234375,
0.285888671875,
0.744140625,
0.285888671875,
0.7587890625,
0.07843017578125,
0.75830078125
],
[
0.449462890625,
0.7470703125,
0.465576171875,
0.7470703125,
0.465576171875,
0.755859375,
0.449462890625,
0.755859375
],
[
0.50341796875,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76220703125,
0.50341796875,
0.76220703125
],
[
0.1004638671875,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.1004638671875,
0.77880859375
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5009765625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4453125,
0.81103515625,
0.4453125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.09722900390625,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.09722900390625,
0.84033203125
],
[
0.50341796875,
0.8271484375,
0.8857421875,
0.8271484375,
0.8857421875,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08087158203125,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.50341796875,
0.84765625,
0.83642578125,
0.84765625,
0.83642578125,
0.8623046875,
0.50341796875,
0.8623046875
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.521484375,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.521484375,
0.8837890625
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"3.3. Weighted Procrustes for SE(3)",
"be embedded in an end-to-end registration pipeline, since",
"Weighted Procrustes is differentiable.From a top-level loss",
"The inlier likelihood estimated by the 6D ConvNet pro-",
"function L of R and t, we can pass the gradient through the",
"vides a weight for each correspondence.",
"The original",
"closed-form solver, and update parameters in downstream",
"Procrustes method [15] minimizes the mean squared er-",
"rorbetweencorresponding points(a.)eM xyil?",
"modules:",
"a",
"OL(R,t) OR, OL(R,t) Ot(R,w)",
"and thus gives equal weight to all correspondences. In",
"L(R,t)",
"(5)",
"contrast, we minimize a weighted mean squared error",
"Ow",
"ROw",
"at",
"Ow",
"Z(i.5)eM W(i,j)llx; - yill2. This change allows us to pass",
"where L(R,t) can be defined as the combination of differ-",
"gradients through the weights, rather than through the po-",
"entiable rotation error (RE) and translation error (TE) be-",
"sition [41], and enables the optimization to scale to dense",
"tween predictions R,t and ground-truth R*, t*:",
"correspondencesets.",
" Formally, Weighted Procrustes analysis minimizes:",
"Tr(RT R*) - 1",
"Lrot(R) = arccos",
"(6)",
"2",
"e2 = e?(R, t: W,X, Y)",
"(2)",
"Lurans(t) =lt - t*2,",
"(7)",
"= wi(ij)(yj -(Rxi +t)2",
"(3)",
"or the Forbenius norm of relative transformation matrices",
"(i.j)EM",
"defined in [2, 41]. The final loss is the weighted sum of",
"Tr((Y -RX -tIT)W(Y- RX -tIT)T",
"(4)",
"Lrot, Lrans, and Lbce.",
"4. Robust Registration",
"where 1 = (1, .,1)T, X = [xi,.., xiMl, and Y",
"[yJi, .., yJim ]. J is a list of indices that defines the corre-",
"In this section, we propose a fine-tuning module that",
"spondences x, yJ.. w = [wi,*-: , wjmi] is the weight",
"minimizes a robust loss function of choice to improve the",
"d(w)",
"registration accuracy. We use a gradient-based method to",
"denotesthe",
"vectorandw=[wi.,wm]",
"o(w)l",
"",
"normalized weight after a nonlinear transformation that",
"refine poses, where a continuous representation [52] for ro-",
"applies heuristic prefiltering. W = diag(w) forms the di-",
"tations is adopted to remove discontinuities and construct a",
"agonal weight matrix.",
"smooth optimization space. This module initializes the pose",
"from the prediction of the Weighted Procrustes method.",
"Theorem 1 : The R and t that minimize the squared er-",
"During iterative optimization, unlike Maken et al. [27], who",
"ror e2(R,t) = (i,) W(i.5)(j - Rx; - t)? areit = (Y -",
"findthenearestneighborperpointateachgradientstep,we",
"RX)W1 and R = USVT where UEVT = SVD(Zry),",
"rely on the correspondence likelihoods from the 6D Con-",
"Zry = YKWKXT, K = I - VWVWT, and S =",
"vNet, which is estimated only once per initialization.",
"In addition, our framework naturally offers a failure de-",
"diag (1, ..., 1, det(U)det(V).",
"tection mechanism. In practice, Weighted Procrustes may",
"Sketch of proof. First, we differentiate e? w.r.t. t and",
"generate numerically unstable solutions when the number",
"equate the partial derivative to 0. This gives us t",
"of valid correspondences is insufficient due to small over-",
"(Y - RX)W1. Next, we substitute X = KX +",
"laps or noisy correspondences between input scans.By",
"computing the ratio of the sum of the filtered weights to",
"XwVwon Eq.4 and do the same for Y.Then, we",
"the total number of correspondences, i.e. , @(wi)/Ml,",
"substitute t",
"t into the squares. This yields e?",
"we can easily approximate the fraction of valid corre-",
"Tr ((Y -RX)KWKT(Y -RX)T) and expanding the",
"spondences and predict whether an alignment may be un-",
"term results in two squares plus -2Tr(YKWKT XT RT)",
"stable. When this fraction is low, we resort to a more",
"We maximize the last negative term whose maximum is",
"time-consumingbut accurate registration algorithm such as",
"the sum of all singular values, which leads to R. The full",
"RANSAC [38. 1, 351 or a branch-and-bound method [451 to",
"derivation is in the supplement.",
"口",
"find a numerically stable solution.In other words.we can",
"We can easily extend the above theorem to incorporate a",
"detect when our system might fail before it returns a result",
"scaling factor c e R+, or anisotropic scaling for tasks such",
"and fall back to a more accurate but time-consuming algo-",
"as scan-to-CAD registration, but in this paper we assume",
"rithm, unlike previous end-to-end methods that use globally",
"that partial scans of the same scene have the same scale.",
"pooled latent features [2] or a singly stochastic matrix [4]]",
"The Weighted Procrustes method generates rotation R",
"- such latent representations are more difficult to interpret.",
"and translationtas outputsthat depend on theweight vector",
"4.1. SE(3) Representation and Initialization",
"w.In our currentimplementation, R and t are directly sent",
"to the robust registration module in Section 4 as an initial",
"We use the 6D representation of 3D rotation proposed",
"pose. However, we briefly demonstrate that they can also",
"by Zhou et al. [52], rather than Euler angles or quaternions."
] |
{
"arxiv_id": "2004.11540",
"language": "en",
"page_idx": 3
}
|
||
2004.11540-en-4
|
2004.11540-en-4.jpg
|
The new representation uses 6 parameters $\mathbf{a}_1,\mathbf{a}_2\in\mathbb{R}^3$ and
can be transformed into a $3\times3$ orthogonal matrix by
(8)
$$f\left(\begin{bmatrix}|&|\\\mathbf{a}_1&\mathbf{a}_2\\|&|\end{bmatrix}\right)=\begin{bmatrix}|&|&|\\\mathbf{b}_1&\mathbf{b}_2&\mathbf{b}_3\\|&|&|\end{bmatrix},$$
$\mathbf{where~b}_{1},\mathbf{b}_{2},\mathbf{b}_{3}\in\mathbb{R}^{3}$ are $\mathbf{b}_1=N(\mathbf{a}_{1}),\mathbf{b}_{2}=N(\mathbf{a}_{2}-$ $(\mathbf{b}_{1}\cdot\mathbf{a}_{2})\mathbf{b}_{1})$, and $\mathbf{b}_{3}=\mathbf{b}_{1}\times\mathbf{b}_{2}$, and $N(\cdot)$ denotes L2
normalization. Thus, the final representation that we use is
$\mathbf{a}_1,\mathbf{a}_2,\mathbf{t}$ which are equivalent to $R,\mathbf{t}$ using Eq. 8.
To initialize $\mathbf{a}_1,\mathbf{a}_2$,we simply use the frst two columns of the rotation matrix $R$, i.e., $\mathbf{b}_1,\mathbf{b}_2.$ For convenience, we define $f^{-1}$ as $f^{-1}(f(R))=R$ though this inverse function is not unique as there are infinitely many choices of $\mathbf{a}_1,\mathbf{a}_2$ that map to the same $R.$
# 4.2. Energy Minimization
We use a robust loss function to fine-tune the registration between predicted inlier correspondences. The general form of the energy function is
(9)
$$E(R,\mathbf{t})=\sum_{i=1}^n\phi(w_{(i,J_i)})L(\mathbf{y}_{J_i},R\mathbf{x}_i+\mathbf{t}),$$
where $\tilde{w}_i$ and $J_i$ are defined as in Eq. 3 and $\phi(\cdot)$ is a prefiltering function. In the experiments, we use $\phi(w)=$ $I[w>\tau]w$, which clips weights below $\tau$ elementwise as neural network outputs bounded logit scores. $L(\mathbf{x},\mathbf{y})$ is a pointwise loss function between x and y; we use the Huber loss in our implementation. The energy function is parameterized by $R$ and t which in turn are represented as $\mathbf{a}_1,\mathbf{a}_2,\mathbf{t}.$ We can apply first-order optimization algorithms such as SGD, Adam, etc. to minimize the energy function, but higher-order optimizers are also applicable since the number of parameters is small. The complete algorithm is described in Alg. 1.
# 5. Experiments
We analyze the proposed model in two registration scenarios: pairwise registration where we estimate an SE(3) transformation between two 3D scans or fragments, and multi-way registration which generates a final reconstruction and camera poses for all fragments that are globally consistent. Here, pairwise registration serves as a critical module in multi-way registration.
For pairwise registration, we use the 3DMatch benchmark[47] which consists of 3D point cloud pairs from various real-world scenes with ground truth transformations estimated from RGB-D reconstruction pipelines [17,10]. We follow the train/test split and the standard procedure to generate pairs with at least 30% overlap for training and test- $\operatorname{ing}[12,11,9].$ For multi-way registration, we use the simulated Augmented ICL-NUIM dataset [6, 18] for quantitative
<table>
<tbody>
<tr>
<td> </td>
</tr>
<tr>
<td>3.1 8</td>
</tr>
<tr>
<td> </td>
</tr>
<tr>
<td> </td>
</tr>
<tr>
<td>$f//$ 83.5 1</td>
</tr>
<tr>
<td>$f^{\prime}/f^{\prime}$ 84.1</td>
</tr>
<tr>
<td> </td>
</tr>
<tr>
<td>$\mathbf{a})X_{i}+\mathbf{t})$</td>
</tr>
<tr>
<td> </td>
</tr>
<tr>
<td> </td>
</tr>
<tr>
<td> </td>
</tr>
</tbody>
</table>
trajectory results, and Indoor LiDAR RGB-D dataset[32] and Stanford RGB-D dataset[6] for qualitative registration visualizations. Note in this experiment we use networks trained on the 3DMatch training set and do not fine-tune on the other datasets. This illustrates the generalization abilities of our models. Lastly, we use KITTI LIDAR scans[13] for outdoor pairwise registration. As the official registration splits do not have labels for pairwise registration, we follow Choy et al.[9] to create pairwise registration train/val/test splits.
For all indoor experiments, we use 5cm voxel downsampling [35, 51], which randomly subsamples a single point within each 5cm voxel to generate point clouds with uniform density. For safeguard registration, we use RANSAC and the safeguard threshold $\tau_s=0.05$, which translates to 5% of the correspondences should be valid. We train learning-based state-of-the-art models and our network on the training split of the 3DMatch benchmark. During training, we augment data by applying random rotations varying from -180 to 180 degrees around a random axis. Groundtruth pointwise correspondences are found using nearest neighbor search in 3D space. We train the 6-dimensional ConvNet on a single Titan XP with batch size 4. SGD is used with an initial learning rate $10^{-1}$ and an exponential learning rate decay factor 0.99.
## 5.1. Pairwise Registration
In this section, we report the registration results on the test set of the 3DMatch benchmark [47], which contains 8 different scenes as depicted in Fig. 3. We measure translation error $(TE)$ defined in Eq. 6, rotation error $(RE)$ defined
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.51123046875,
0.09783935546875,
0.775390625,
0.09783935546875,
0.775390625,
0.10797119140625,
0.51123046875,
0.10797119140625
],
[
0.07843017578125,
0.106689453125,
0.428955078125,
0.10797119140625,
0.428955078125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.51953125,
0.11297607421875,
0.72802734375,
0.111083984375,
0.72802734375,
0.1256103515625,
0.51953125,
0.1275634765625
],
[
0.51953125,
0.1275634765625,
0.73046875,
0.125,
0.73046875,
0.143310546875,
0.51953125,
0.146484375
],
[
0.23779296875,
0.1351318359375,
0.390625,
0.1351318359375,
0.390625,
0.1717529296875,
0.23779296875,
0.1717529296875
],
[
0.178955078125,
0.14453125,
0.2548828125,
0.14453125,
0.2548828125,
0.1634521484375,
0.178955078125,
0.1634521484375
],
[
0.50634765625,
0.143310546875,
0.646484375,
0.14453125,
0.646484375,
0.1590576171875,
0.50634765625,
0.1578369140625
],
[
0.81787109375,
0.14453125,
0.8837890625,
0.14453125,
0.8837890625,
0.1590576171875,
0.81787109375,
0.1590576171875
],
[
0.291748046875,
0.1572265625,
0.381591796875,
0.1572265625,
0.381591796875,
0.1749267578125,
0.291748046875,
0.1749267578125
],
[
0.44677734375,
0.147705078125,
0.468017578125,
0.147705078125,
0.468017578125,
0.160400390625,
0.44677734375,
0.160400390625
],
[
0.50732421875,
0.160400390625,
0.6435546875,
0.160400390625,
0.6435546875,
0.1749267578125,
0.50732421875,
0.1749267578125
],
[
0.5048828125,
0.173583984375,
0.75634765625,
0.1717529296875,
0.75634765625,
0.189453125,
0.5048828125,
0.1912841796875
],
[
0.81640625,
0.175537109375,
0.8837890625,
0.175537109375,
0.8837890625,
0.1900634765625,
0.81640625,
0.1900634765625
],
[
0.07843017578125,
0.184326171875,
0.46728515625,
0.1849365234375,
0.46728515625,
0.20263671875,
0.07843017578125,
0.2020263671875
],
[
0.50634765625,
0.189453125,
0.74853515625,
0.189453125,
0.74853515625,
0.203857421875,
0.50634765625,
0.203857421875
],
[
0.08087158203125,
0.20263671875,
0.468017578125,
0.20263671875,
0.468017578125,
0.216552734375,
0.08087158203125,
0.216552734375
],
[
0.5048828125,
0.203857421875,
0.705078125,
0.2052001953125,
0.705078125,
0.2197265625,
0.5048828125,
0.2183837890625
],
[
0.08087158203125,
0.2197265625,
0.46728515625,
0.2197265625,
0.46728515625,
0.2298583984375,
0.08087158203125,
0.2298583984375
],
[
0.50732421875,
0.2203369140625,
0.66650390625,
0.2203369140625,
0.66650390625,
0.2342529296875,
0.50732421875,
0.2342529296875
],
[
0.07843017578125,
0.2342529296875,
0.404296875,
0.2342529296875,
0.404296875,
0.2474365234375,
0.07843017578125,
0.2474365234375
],
[
0.54345703125,
0.23486328125,
0.79345703125,
0.23486328125,
0.79345703125,
0.248779296875,
0.54345703125,
0.248779296875
],
[
0.8408203125,
0.23486328125,
0.8837890625,
0.23486328125,
0.8837890625,
0.250732421875,
0.8408203125,
0.250732421875
],
[
0.098876953125,
0.248779296875,
0.4697265625,
0.248779296875,
0.4697265625,
0.26318359375,
0.098876953125,
0.26318359375
],
[
0.5087890625,
0.251953125,
0.55126953125,
0.251953125,
0.55126953125,
0.26318359375,
0.5087890625,
0.26318359375
],
[
0.080078125,
0.26318359375,
0.4697265625,
0.26318359375,
0.4697265625,
0.276611328125,
0.080078125,
0.276611328125
],
[
0.5498046875,
0.266357421875,
0.79150390625,
0.266357421875,
0.79150390625,
0.281005859375,
0.5498046875,
0.281005859375
],
[
0.5107421875,
0.272705078125,
0.5146484375,
0.272705078125,
0.5146484375,
0.27783203125,
0.5107421875,
0.27783203125
],
[
0.81640625,
0.267578125,
0.88232421875,
0.267578125,
0.88232421875,
0.2822265625,
0.81640625,
0.2822265625
],
[
0.080078125,
0.27783203125,
0.468017578125,
0.27783203125,
0.468017578125,
0.292236328125,
0.080078125,
0.292236328125
],
[
0.54833984375,
0.2841796875,
0.6845703125,
0.2841796875,
0.6845703125,
0.298583984375,
0.54833984375,
0.298583984375
],
[
0.81787109375,
0.285888671875,
0.88134765625,
0.285888671875,
0.88134765625,
0.29736328125,
0.81787109375,
0.29736328125
],
[
0.0767822265625,
0.292236328125,
0.468017578125,
0.29345703125,
0.468017578125,
0.30810546875,
0.0767822265625,
0.306884765625
],
[
0.50341796875,
0.287353515625,
0.515625,
0.287353515625,
0.515625,
0.298583984375,
0.50341796875,
0.298583984375
],
[
0.5009765625,
0.300537109375,
0.5185546875,
0.300537109375,
0.5185546875,
0.314453125,
0.5009765625,
0.314453125
],
[
0.54833984375,
0.2998046875,
0.708984375,
0.2998046875,
0.708984375,
0.314453125,
0.54833984375,
0.314453125
],
[
0.080078125,
0.309326171875,
0.238525390625,
0.309326171875,
0.238525390625,
0.322509765625,
0.080078125,
0.322509765625
],
[
0.50244140625,
0.31494140625,
0.51953125,
0.31494140625,
0.51953125,
0.330078125,
0.50244140625,
0.330078125
],
[
0.56884765625,
0.313232421875,
0.86279296875,
0.313232421875,
0.86279296875,
0.335205078125,
0.56884765625,
0.335205078125
],
[
0.08087158203125,
0.33203125,
0.27783203125,
0.33203125,
0.27783203125,
0.34521484375,
0.08087158203125,
0.34521484375
],
[
0.5009765625,
0.333984375,
0.51953125,
0.333984375,
0.51953125,
0.348388671875,
0.5009765625,
0.348388671875
],
[
0.5703125,
0.33203125,
0.74658203125,
0.33203125,
0.74658203125,
0.349853515625,
0.5703125,
0.349853515625
],
[
0.5009765625,
0.35107421875,
0.5185546875,
0.35107421875,
0.5185546875,
0.3662109375,
0.5009765625,
0.3662109375
],
[
0.56884765625,
0.347900390625,
0.74169921875,
0.349853515625,
0.74169921875,
0.36865234375,
0.56884765625,
0.3662109375
],
[
0.098876953125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.36865234375,
0.098876953125,
0.36865234375
],
[
0.080078125,
0.37060546875,
0.4697265625,
0.37060546875,
0.4697265625,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.50341796875,
0.37255859375,
0.51708984375,
0.37255859375,
0.51708984375,
0.38330078125,
0.50341796875,
0.38330078125
],
[
0.54833984375,
0.369384765625,
0.64453125,
0.369384765625,
0.64453125,
0.3837890625,
0.54833984375,
0.3837890625
],
[
0.080078125,
0.385009765625,
0.27783203125,
0.385009765625,
0.27783203125,
0.3984375,
0.080078125,
0.3984375
],
[
0.200927734375,
0.4140625,
0.419189453125,
0.4140625,
0.419189453125,
0.439453125,
0.200927734375,
0.439453125
],
[
0.2156982421875,
0.410400390625,
0.2279052734375,
0.410400390625,
0.2279052734375,
0.416748046875,
0.2156982421875,
0.416748046875
],
[
0.134033203125,
0.419189453125,
0.2132568359375,
0.419189453125,
0.2132568359375,
0.434326171875,
0.134033203125,
0.434326171875
],
[
0.2132568359375,
0.423583984375,
0.2320556640625,
0.423583984375,
0.2320556640625,
0.43115234375,
0.2132568359375,
0.43115234375
],
[
0.44677734375,
0.42041015625,
0.46728515625,
0.42041015625,
0.46728515625,
0.43310546875,
0.44677734375,
0.43310546875
],
[
0.50244140625,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.50244140625,
0.439453125
],
[
0.2099609375,
0.438232421875,
0.2320556640625,
0.438232421875,
0.2320556640625,
0.4462890625,
0.2099609375,
0.4462890625
],
[
0.50244140625,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455078125,
0.50244140625,
0.455078125
],
[
0.080078125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50244140625,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.5009765625,
0.47021484375,
0.8935546875,
0.4716796875,
0.8935546875,
0.486083984375,
0.5009765625,
0.48486328125
],
[
0.08251953125,
0.486083984375,
0.4697265625,
0.486083984375,
0.4697265625,
0.5,
0.08251953125,
0.5
],
[
0.50244140625,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.080078125,
0.501953125,
0.4697265625,
0.501953125,
0.4697265625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.08087158203125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.52880859375,
0.08087158203125,
0.52880859375
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.53076171875,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50244140625,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.546875,
0.50244140625,
0.546875
],
[
0.07843017578125,
0.54736328125,
0.471435546875,
0.54736328125,
0.471435546875,
0.56103515625,
0.07843017578125,
0.56103515625
],
[
0.5009765625,
0.546875,
0.89208984375,
0.54736328125,
0.89208984375,
0.5625,
0.5009765625,
0.56103515625
],
[
0.07843017578125,
0.5625,
0.471435546875,
0.5625,
0.471435546875,
0.5771484375,
0.07843017578125,
0.5771484375
],
[
0.50244140625,
0.564453125,
0.5458984375,
0.564453125,
0.5458984375,
0.57568359375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.57763671875,
0.468017578125,
0.57763671875,
0.468017578125,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.52392578125,
0.580078125,
0.8896484375,
0.580078125,
0.8896484375,
0.59033203125,
0.52392578125,
0.59033203125
],
[
0.080078125,
0.59228515625,
0.4697265625,
0.59228515625,
0.4697265625,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.5048828125,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.60595703125,
0.5048828125,
0.60595703125
],
[
0.080078125,
0.60693359375,
0.4697265625,
0.60693359375,
0.4697265625,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50341796875,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.07843017578125,
0.62255859375,
0.2115478515625,
0.62255859375,
0.2115478515625,
0.63623046875,
0.07843017578125,
0.63623046875
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.50244140625,
0.6396484375,
0.8935546875,
0.6396484375,
0.8935546875,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.08087158203125,
0.6494140625,
0.2091064453125,
0.6494140625,
0.2091064453125,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.098876953125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.098876953125,
0.68798828125
],
[
0.50341796875,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.07843017578125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.07843017578125,
0.7041015625
],
[
0.50244140625,
0.6845703125,
0.89208984375,
0.6845703125,
0.89208984375,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.50244140625,
0.69970703125,
0.89208984375,
0.69970703125,
0.89208984375,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.07843017578125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.7138671875,
0.89208984375,
0.7138671875,
0.89208984375,
0.72802734375,
0.50244140625,
0.72802734375
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73486328125
],
[
0.5009765625,
0.72900390625,
0.890625,
0.72900390625,
0.890625,
0.7421875,
0.5009765625,
0.7421875
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.744140625,
0.89208984375,
0.744140625,
0.89208984375,
0.75830078125,
0.50244140625,
0.75830078125
],
[
0.07843017578125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.07843017578125,
0.76513671875
],
[
0.50244140625,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.50244140625,
0.7734375
],
[
0.07843017578125,
0.7666015625,
0.301513671875,
0.7666015625,
0.301513671875,
0.77978515625,
0.07843017578125,
0.77978515625
],
[
0.5009765625,
0.7734375,
0.89208984375,
0.7744140625,
0.89208984375,
0.7890625,
0.5009765625,
0.78857421875
],
[
0.09722900390625,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.09722900390625,
0.79443359375
],
[
0.50244140625,
0.79052734375,
0.70751953125,
0.79052734375,
0.70751953125,
0.8037109375,
0.50244140625,
0.8037109375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.8173828125,
0.70166015625,
0.8173828125,
0.70166015625,
0.83056640625,
0.50341796875,
0.83056640625
],
[
0.07843017578125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.521484375,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"The new representation uses 6 parameters ai, a2 e R3 and",
"Algorithm1:DeepGlobalRegistration",
"can be transformed into a 3× 3 orthogonal matrix by",
"Input:XeRnx3YeIRmx3",
"Output: R E SO(3),t e R3x1",
") b. ba bs",
"a1a2",
"1 F←Feature(X)",
"1/83.1",
"1",
"(8)",
"2 Fy ←Feature(Y)",
"3 Jr→y ← NearestNeighbor(Fr, Fy)",
"1183.2",
"where bi, b2,b3 IR3 are bi = N(ai), b2 = N(a2 -",
"4 M-{(i, Jr-u.) li e[1, ....nl]",
"(bi a2)bi), and b3 = bi × b2, and N() denotes L2",
"5wInlierProbability(M)",
"normalization.Thus.thefinalrepresentation that we use is",
"6 if EΦ(w) < Ts then",
"ai, a2, t which are equivalent to R, t using Eq. 8.",
"return SafeGuardRegistration(X, Y)",
"1/84",
"To initialize ai, a2, we simply use the first two columns",
" else",
"of the rotation matrix R, i.e., bi, b2.For convenience, we",
"R,t argminRte?(R,t; w,X,Y)",
"",
"183.3",
"define f-1 as f-1(f(R) = R though this inverse function",
"a←f-1(R),t+t",
"1/84.1",
"is not unique as there are infinitely many choices of ai, a2",
"10",
"11",
"while not converging do",
"that map to the same R.",
"12",
"l E(2)eM o(w(.)L(Yj, f(a)X; + t)",
"4.2. Energy Minimization",
"13",
"a ← Update(a, l(a, t)",
"14",
"t Update(t, 最e(a,t)",
"We use a robust loss function to fine-tune the registra-",
"tion between predicted inlier correspondences. The general",
"15",
"return f(a),t",
"form of the energy function is",
"Eo(w(ij)L(yj., Rx +t),",
"n",
"E(R,t) =",
">",
"(9)",
"trajectory results, and Indoor LiDAR RGB-D dataset [32]",
"i=l",
"and Stanford RGB-D dataset [6] for qualitative registration",
"where w; and J, are defined as in Eq. 3 and p() is a pre-",
"visualizations. Note in this experiment we use networks",
"filtering function.In the experiments, we use (w) =",
"trained on the 3DMatch training set and do not fine-tune on",
"I[w > -lw, which clips weights below ↑ elementwise as",
"the other datasets. This illustrates the generalization abili-",
"neural network outputs bounded logit scores. L(x, y) is a",
"ties of our models. Lastly, we use KITTILIDAR scans [13]",
"pointwise loss function between x and y:we use the Hu-",
"for outdoor pairwise registration.As the official registration",
"ber loss in our implementation. The energy function is pa-",
"splits do not have labels for pairwise registration, we follow",
"rameterized by R and t which in turn are represented as",
"Choy et al. [9] to create pairwise registration train/val/test",
"ai, a2,t. We can apply first-order optimization algorithms",
"splits.",
"such as SGD, Adam, etc. to minimize the energy function,",
"Forallindoorexperiments,weuse5cmvoxeldownsam-",
"but higher-order optimizers are also applicable since the",
"pling[35,51l,whichrandomly subsamples a singlepoint",
"number of parameters is small. The complete algorithm is",
"within each 5cm voxel to generate point clouds with uni-",
"described in Alg. 1.",
"form density. For safeguard registration, we use RANSAC",
"and the safeguard threshold Ts = 0.05, which translates",
"5. Experiments",
"to 5% of the correspondences should be valid. We train",
"We analyze the proposed model in two registration sce-",
"learning-based state-of-the-art models and ournetwork on",
"narios: pairwise registration where we estimate an SE(3)",
"the training split of the 3DMatch benchmark. During train-",
"ing, we augment data by applying random rotations varying",
"transformation between two 3D scans or fragments, and",
"from -180 to 180 degrees around a random axis. Ground-",
"multi-way registration which generates a final reconstruc-",
"truth pointwise correspondences are found using nearest",
"tion and camera poses for all fragments that are globally",
"neighbor search in 3D space. We train the 6-dimensional",
"consistent. Here, pairwise registration serves as a critical",
"ConvNet on a single Titan XP with batch size 4. SGD is",
"module in multi-way registration.",
"used with an initial learning rate 10-1 and an exponential",
"For pairwise registration, we use the 3DMatch bench-",
"learning rate decay factor 0.99.",
"mark [47] which consists of 3D point cloud pairs from vari-",
"ous real-world scenes with ground truth transformations es-",
"5.1. Pairwise Registration",
"timated from RGB-D reconstruction pipelines [17, 10]. We",
"follow the train/test split and the standard procedure to gen-",
"In this section, we report the registration results on the",
"erate pairs with at least 30% overlap for training and test-",
"test setof the3DMatchbenchmark[47l.whichcontains8",
"ing [12, 11, 9]. For multi-way registration, we use the simu-",
"different scenes as depicted in Fig. 3. We measure transla-",
"lated Augmented ICL-NUIM dataset [6, 18] for quantitative",
"tion error (TE) defined in Eq. 6, rotation error (RE) defined"
] |
{
"arxiv_id": "2004.11540",
"language": "en",
"page_idx": 4
}
|
||
1806.10779-en-0
|
1806.10779-en-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2019}}$
# DIFFERENTIABLE LEARNING-TO-NORMALIZE VIA SwITCHABLE NORMALIZATION
$\textbf{Ping Luo}^{1, 3* }$ $\textbf{Jiamin Ren}^{2* }$ $\textbf{Zhanglin Peng}^2$ $\textbf{Ruimao Zhang}^1$ $\textbf{Jingyu Li}^1$
${}^{2}\tilde{\mathrm{SenseTime~Research}}$
$^{1}$The Chinese University of Hong Kong
$^{3}$The University of Hong Kong
ABSTRACT
We address a learning-to-normalize problem by proposing Switchable Normalization (SN), which learns to select different normalizers for different normalization layers of a deep neural network. SN employs three distinct scopes to compute statistics (means and variances) including a channel, a layer, and a minibatch. SN switches between them by learning their importance weights in an end-to-end manner. It has several good properties. First, it adapts to various network architectures and tasks (see Fig.1). Second, it is robust to a wide range of batch sizes, maintaining high performance even when small minibatch is presented $(e.g.2$ images/GPU). Third, SN does not have sensitive hyper-parameter, unlike group normalization that searches the number of groups as a hyper-parameter. Without bells and whistles, SN outperforms its counterparts on various challenging benchmarks, such as ImageNet, COCO, CityScapes, ADE20K, and Kinetics. Analyses of SN are also presented. We hope SN will help ease the usage and understand the normalization techniques in deep learning. The code of SN has been released in https://github.com/switchablenorms/.
### 1 INTRODUCTION
Normalization techniques are effective components in deep learning, advancing many research fields such as natural language processing, computer vision, and machine learning. In recent years, many normalization methods such as Batch Normalization (BN) (Ioffe & Szegedy, 2015), Instance Normalization (IN) (Ulyanov et al., 2016), and Layer Normalization (LN) (Ba et al., 2016) have been developed. Despite their great successes, existing practices often employed the same normalizer in all normalization layers of an entire network, rendering suboptimal performance. Also, different normalizers are used to solve different tasks, making model design cumbersome.
To address the above issues, we propose Switchable Normalization (SN), which combines three types of statistics estimated channel-wise, layer-wise, and minibatch-wise by using IN, LN, and BN respectively. SN switches among them by learning their importance weights. By design, SN $is$ adaptable to various deep networks and tasks. For example, the ratios of IN, LN, and BN in SN are compared in multiple tasks as shown in Fig.1(a). We see that using one normalization method uniformly is not optimal for these tasks. For instance, image classification and object detection prefer the combination of three normalizers. In particular, SN chooses BN more than IN and LN in image classification and the backbone network of object detection, while LN has larger weights in the box and mask heads. For artistic image style transfer (Johnson et al., 2016), SN selects IN. For neural architecture search, SN is applied to LSTM where LN is preferable than group normalization (GN) (Wu& He, 2018), which is a variant of IN by dividing channels into groups.
The selectivity of normalizers makes SN robust to minibatch size. As shown in Fig.1(b), when training ResNet50 (He et al., 2016) on ImageNet (Deng et al., 2009) with different batch sizes, SN is close to the“ideal case”more than BN and GN. For (8,32) as an example$^1$, ResNet50 trained with SN is able to achieve 76.9% top-1 accuracy, surpassing BN and GN by $\hat{0}.5\%$ and $1.0\%$ respectively. In general, SN obtains better or comparable results than both BN and GN in all batch settings
$^{*}$The first two authors contribute equally. Corresponding to: pluo.lhi@gmail.com, $\{$renjiamin,
pengzhanglin, zhangruimao, lijingyu$\}@$sensetime.com.
In this work, minibatch size refers to the number of samples per GPU, and batch size is $\#$GPUs' times
$`\#$samples per GPU'. A batch setting is denoted as a 2-tuple, (#GPUs, #samples per GPU).
1
|
[
[
0.17236328125,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.17236328125,
0.04736328125
],
[
0.176513671875,
0.10101318359375,
0.775390625,
0.10162353515625,
0.775390625,
0.11932373046875,
0.176513671875,
0.11871337890625
],
[
0.176513671875,
0.1256103515625,
0.5537109375,
0.126953125,
0.5537109375,
0.14453125,
0.176513671875,
0.143310546875
],
[
0.28515625,
0.169189453125,
0.390625,
0.1685791015625,
0.390625,
0.184326171875,
0.28515625,
0.1849365234375
],
[
0.398681640625,
0.17041015625,
0.65283203125,
0.17041015625,
0.65283203125,
0.1849365234375,
0.398681640625,
0.1849365234375
],
[
0.65283203125,
0.17041015625,
0.736328125,
0.17041015625,
0.736328125,
0.1849365234375,
0.65283203125,
0.1849365234375
],
[
0.39208984375,
0.173583984375,
0.40185546875,
0.173583984375,
0.40185546875,
0.1787109375,
0.39208984375,
0.1787109375
],
[
0.1829833984375,
0.1717529296875,
0.2900390625,
0.1673583984375,
0.291748046875,
0.18310546875,
0.1846923828125,
0.1875
],
[
0.650390625,
0.1749267578125,
0.65673828125,
0.1749267578125,
0.65673828125,
0.179931640625,
0.650390625,
0.179931640625
],
[
0.1846923828125,
0.184326171875,
0.44775390625,
0.1849365234375,
0.44775390625,
0.199462890625,
0.1846923828125,
0.1988525390625
],
[
0.4697265625,
0.186279296875,
0.61669921875,
0.186279296875,
0.61669921875,
0.1962890625,
0.4697265625,
0.1962890625
],
[
0.6435546875,
0.1849365234375,
0.8505859375,
0.1849365234375,
0.8505859375,
0.199462890625,
0.6435546875,
0.199462890625
],
[
0.455078125,
0.238037109375,
0.5458984375,
0.238037109375,
0.5458984375,
0.25,
0.455078125,
0.25
],
[
0.0196075439453125,
0.25830078125,
0.062103271484375,
0.25830078125,
0.06451416015625,
0.40771484375,
0.0228729248046875,
0.40771484375
],
[
0.2344970703125,
0.267578125,
0.7646484375,
0.267578125,
0.7646484375,
0.281005859375,
0.2344970703125,
0.281005859375
],
[
0.233642578125,
0.281005859375,
0.765625,
0.281005859375,
0.765625,
0.294189453125,
0.233642578125,
0.294189453125
],
[
0.233642578125,
0.29541015625,
0.76708984375,
0.29541015625,
0.76708984375,
0.309326171875,
0.233642578125,
0.309326171875
],
[
0.233642578125,
0.30810546875,
0.76318359375,
0.30810546875,
0.76318359375,
0.3212890625,
0.233642578125,
0.3212890625
],
[
0.233642578125,
0.322509765625,
0.765625,
0.322509765625,
0.765625,
0.3359375,
0.233642578125,
0.3359375
],
[
0.2344970703125,
0.337158203125,
0.7646484375,
0.337158203125,
0.7646484375,
0.35107421875,
0.2344970703125,
0.35107421875
],
[
0.233642578125,
0.35107421875,
0.76318359375,
0.35107421875,
0.76318359375,
0.3642578125,
0.233642578125,
0.3642578125
],
[
0.233642578125,
0.3642578125,
0.765625,
0.3642578125,
0.765625,
0.37744140625,
0.233642578125,
0.37744140625
],
[
0.2320556640625,
0.376953125,
0.7646484375,
0.376953125,
0.7646484375,
0.39013671875,
0.2320556640625,
0.39013671875
],
[
0.2320556640625,
0.391357421875,
0.76708984375,
0.39013671875,
0.76708984375,
0.40478515625,
0.2320556640625,
0.406005859375
],
[
0.0196075439453125,
0.397705078125,
0.05963134765625,
0.396484375,
0.062103271484375,
0.496826171875,
0.0228729248046875,
0.49755859375
],
[
0.233642578125,
0.406005859375,
0.765625,
0.406005859375,
0.765625,
0.419189453125,
0.233642578125,
0.419189453125
],
[
0.2344970703125,
0.419189453125,
0.765625,
0.419189453125,
0.765625,
0.43310546875,
0.2344970703125,
0.43310546875
],
[
0.233642578125,
0.43310546875,
0.765625,
0.431884765625,
0.765625,
0.4462890625,
0.233642578125,
0.447509765625
],
[
0.2344970703125,
0.447509765625,
0.76708984375,
0.447509765625,
0.76708984375,
0.4609375,
0.2344970703125,
0.4609375
],
[
0.2320556640625,
0.460205078125,
0.58251953125,
0.4609375,
0.58251953125,
0.474853515625,
0.2320556640625,
0.473388671875
],
[
0.0228729248046875,
0.48291015625,
0.05963134765625,
0.48291015625,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.17724609375,
0.50244140625,
0.1895751953125,
0.50244140625,
0.1895751953125,
0.51220703125,
0.17724609375,
0.51220703125
],
[
0.1993408203125,
0.50048828125,
0.338134765625,
0.50048828125,
0.338134765625,
0.5146484375,
0.1993408203125,
0.5146484375
],
[
0.1732177734375,
0.53076171875,
0.82275390625,
0.53076171875,
0.82275390625,
0.54443359375,
0.1732177734375,
0.54443359375
],
[
0.1732177734375,
0.54443359375,
0.82177734375,
0.54443359375,
0.82177734375,
0.55810546875,
0.1732177734375,
0.55810546875
],
[
0.1732177734375,
0.55810546875,
0.8203125,
0.55810546875,
0.8203125,
0.5712890625,
0.1732177734375,
0.5712890625
],
[
0.1748046875,
0.57275390625,
0.82421875,
0.57275390625,
0.82421875,
0.5859375,
0.1748046875,
0.5859375
],
[
0.1748046875,
0.5869140625,
0.82421875,
0.5869140625,
0.82421875,
0.60107421875,
0.1748046875,
0.60107421875
],
[
0.1748046875,
0.60107421875,
0.826171875,
0.60107421875,
0.826171875,
0.6142578125,
0.1748046875,
0.6142578125
],
[
0.176513671875,
0.61572265625,
0.69921875,
0.61572265625,
0.69921875,
0.62548828125,
0.176513671875,
0.62548828125
],
[
0.1748046875,
0.63525390625,
0.82421875,
0.63525390625,
0.82421875,
0.6484375,
0.1748046875,
0.6484375
],
[
0.1748046875,
0.6484375,
0.82421875,
0.6484375,
0.82421875,
0.662109375,
0.1748046875,
0.662109375
],
[
0.1748046875,
0.66357421875,
0.82421875,
0.66357421875,
0.82421875,
0.6767578125,
0.1748046875,
0.6767578125
],
[
0.1732177734375,
0.67529296875,
0.82275390625,
0.67529296875,
0.82275390625,
0.689453125,
0.1732177734375,
0.689453125
],
[
0.1748046875,
0.69140625,
0.82421875,
0.69140625,
0.82421875,
0.705078125,
0.1748046875,
0.705078125
],
[
0.1732177734375,
0.705078125,
0.82421875,
0.705078125,
0.82421875,
0.71826171875,
0.1732177734375,
0.71826171875
],
[
0.1748046875,
0.71826171875,
0.82421875,
0.71826171875,
0.82421875,
0.73193359375,
0.1748046875,
0.73193359375
],
[
0.1748046875,
0.73291015625,
0.82421875,
0.73291015625,
0.82421875,
0.7470703125,
0.1748046875,
0.7470703125
],
[
0.1732177734375,
0.7470703125,
0.82421875,
0.7470703125,
0.82421875,
0.76025390625,
0.1732177734375,
0.76025390625
],
[
0.1748046875,
0.76025390625,
0.82421875,
0.76025390625,
0.82421875,
0.7734375,
0.1748046875,
0.7734375
],
[
0.1748046875,
0.7744140625,
0.7099609375,
0.7744140625,
0.7099609375,
0.78857421875,
0.1748046875,
0.78857421875
],
[
0.176513671875,
0.79541015625,
0.82421875,
0.79541015625,
0.82421875,
0.8095703125,
0.176513671875,
0.8095703125
],
[
0.1748046875,
0.8095703125,
0.82275390625,
0.8095703125,
0.82275390625,
0.82275390625,
0.1748046875,
0.82275390625
],
[
0.1732177734375,
0.82275390625,
0.82275390625,
0.82275390625,
0.82275390625,
0.8359375,
0.1732177734375,
0.8359375
],
[
0.1748046875,
0.8369140625,
0.82177734375,
0.8369140625,
0.82177734375,
0.85107421875,
0.1748046875,
0.85107421875
],
[
0.17236328125,
0.849609375,
0.78759765625,
0.85107421875,
0.78759765625,
0.86572265625,
0.17236328125,
0.8642578125
],
[
0.19287109375,
0.87060546875,
0.61279296875,
0.8720703125,
0.61279296875,
0.88623046875,
0.19287109375,
0.88525390625
],
[
0.615234375,
0.87255859375,
0.82421875,
0.87255859375,
0.82421875,
0.88623046875,
0.615234375,
0.88623046875
],
[
0.1732177734375,
0.88525390625,
0.50244140625,
0.88525390625,
0.50244140625,
0.8984375,
0.1732177734375,
0.8984375
],
[
0.19775390625,
0.8984375,
0.826171875,
0.8984375,
0.826171875,
0.91357421875,
0.19775390625,
0.91357421875
],
[
0.176513671875,
0.912109375,
0.72607421875,
0.912109375,
0.72607421875,
0.92529296875,
0.176513671875,
0.92529296875
],
[
0.494384765625,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2019",
"DIFFERENTIABLE LEARNING-TO-NORMALIZE VIA",
"SWITCHABLE NORMALIZATION",
"Jiamin Ren2*",
"Zhanglin Peng?Ruimao Zhangl",
"Jingyu Lil",
"",
"Ping Luol,3*",
"",
"IThe Chinese University of Hong Kong",
"2SenseTimeResearch",
"3The University of Hong Kong",
"ABSTRACT",
"61",
"We address a learning-to-normalize problem by proposing Switchable Normaliza-",
"tion (SN), which learns to select different normalizers for different normalization",
"layers of a deep neural network. SN employs three distinct scopes to compute",
"statistics (means and variances) including a channel, a layer, and a minibatch",
"SN switches between them by learning their importance weights in an end-to-end",
"manner. It has several good properties. First, it adapts to various network archi-",
"tectures and tasks (see Fig.l). Second, it is robust to a wide range of batch sizes.",
"maintaining high performance even when small minibatch is presented (e.g. 2",
"images/GPU). Third, SN does not have sensitive hyper-parameter, unlike group",
"normalization that searches the number of groups as a hyper-parameter. Without",
"[']",
"bells and whistles, SN outperforms its counterparts on various challenging bench-",
"marks, such as ImageNet, COCO, CityScapes, ADE20K, and Kinetics. Analyses",
"of SN are also presented. We hope SN will help ease the usage and understand the",
"normalization techniques in deep learning. The code of SN has been released in",
"https://github.com/switchablenorms/.",
"SA6L019081:AIXIE",
"1.",
"INTRODUCTION",
"Normalization techniques are effective components in deep learning, advancing many research fields",
"such as natural language processing, computer vision, and machine learning. In recent years, many",
"normalization methods such as Batch Normalization (BN) (Ioffe & Szegedy, 2015), Instance Nor",
"malization (IN) (Ulyanov et al., 2016), and Layer Normalization (LN) (Ba et al., 2016) have been",
"developed. Despite their great successes, existing practices often employed the same normalizer in",
"all normalization layers of an entire network, rendering suboptimal performance. Also, different",
"normalizersareusedto solvedifferenttasks,makingmodeldesigncumbersome",
"To address the above issues, we propose Switchable Normalization (SN), which combines three",
"types of statistics estimated channel-wise, layer-wise, and minibatch-wise by using IN, LN, and",
"BN respectively. SN switches among them by learning their importance weights. By design, SN is",
"adaptable to various deep networks and tasks: For example, the ratios of IN, LN, and BN in SN",
"are compared in multiple tasks as shown in Fig.1 (a). We see that using one normalization method",
"uniformly is not optimal for these tasks. For instance, image classification and object detection",
"prefer the combination of three normalizers. In particular, SN chooses BN more than IN and LN in",
"image classification and the backbone network of object detection, while LN has larger weights in",
"the box and mask heads. For artistic image style transfer (Johnson et al., 2016), SN selects IN. For",
"neural architecture search, SN is applied to LSTM where LN is preferable than group normalization",
"(GN) (Wu & He, 2018), which is a variant of IN by dividing channels into groups.",
"The selectivity of normalizers makes SN robust to minibatch size.As shown in Fig.1 (b), when",
"training ResNet50 (He et al., 2016) on ImageNet (Deng et al., 2009) with different batch sizes, SN",
"is close to the “ideal case\" more than BN and GN. For (8, 32) as an examplel, ResNet50 trained with",
"SN is able to achieve 76.9% top-1 accuracy, surpassing BN and GN by 0.5% and 1.0% respectively.",
"In general, SN obtains better or comparable results than both BN and GN in all batch settings.",
"*The first two authors contribute equally.Corresponding to:",
"pluo.lhi@gmail.com, {renjiamin,",
"pengzhanglin,zhangruimao, lijingyu)@sensetime.com.",
"'In this work, minibatch size refers to the number of samples per GPU, and batch size is #GPUs times",
"#samples per GPU\".A batch setting is denoted as a 2-tuple, (#GPUs,#samples per GPU).",
"1"
] |
{
"arxiv_id": "1806.10779",
"language": "en",
"page_idx": 0
}
|
||
1806.10779-en-1
|
1806.10779-en-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2019}}$
<FigureHere>
(b)
Figure l: (a) shows that SN adapts to various networks and tasks by learning importance ratios to select normalizers. In (a), a ratio is between 0 and 1 and all ratios of each task sum to 1. (b) shows the top-1 accuracies of ResNet50 trained with SN on ImageNet and compared with BN and GN in different batch settings. The gradients in training are averaged over all GPUs and the statistics of normalizers are estimated in each GPU. $\bar{\text{For instance, all methods are compared to an ideal case, 'ideal BN', whose accuracies are 76.4\% for all settings}}$ This ideal case cannot be obtained in practice. In fact, when the minibatch size decreases, BN's accuracies drop significantly, while SN and GN both maintain reasonably good performance. SN surpasses or is comparable to both BN and GN in all settings.
Overall, this work has three key contributions.(1) We introduce Switchable Normalization (SN), which is applicable in both CNNs and RNNs/LSTMs, and improves the other normalization techniques on many challenging benchmarks and tasks including image recognition in ImageNet (Russakovsky et al., 2015), object detection in COCO (Lin et al., 2014), scene parsing in Cityscapes (Cordts et al., 2016) and ADE20K (Zhou et al., 2017), artistic image stylization (Johnson et al., 2016), neural architecture search (Pham et al., 2018), and video recognition in Kinetics (Kay et al., 2017).(2) The analyses of SN are presented where multiple normalizers can be compared and understood with geometric interpretation.(3) By enabling each normalization layer in a deep network to have its own operation, SN helps ease the usage of normalizers, pushes the frontier of normalization in deep learning, as well as opens up new research direction. We believe that all existing models could be reexamined with this new perspective. We'll make the code of SN available and recommend it as an alternative of existing handcrafted approaches.
In the following sections, we first present SN in Sec.2 and then discuss its relationships with previous
work in Sec.3. SN is evaluated extensively in Sec.4.
2 SwITCHABLE NORMALIZATION (SN)
We describe a general formulation of a normalization layer and then present SN.
A General Form. We take CNN as an illustrative example. Let $h$ be the input data of an arbitrary normalization layer represented by a 4D tensor $(N,C,H,W)$, indicating number of samples, number of channels, height and width of a channel respectively as shown in Fig.2. Let $h_ncij$ and $\hat{h}_{ncij}$ be a pixel before and after normalization, where $n\in [ 1, \ddot{N} ] , c\in [ 1, C] , i\in [ 1, H] $,and $j\in[1,W].$ Let $\mu$ and $\sigma$ be a mean and a standard deviation. We have
(1)
$$\hat{h}_{ncij}=\gamma\frac{h_{ncij}-\mu}{\sqrt{\sigma^2+\epsilon}}+\beta,$$
where $\gamma$ and $\beta$ are a scale and a shift parameter respectively. $\epsilon$ is a small constant to preserve numerical stability. Eqn.(1) shows that each pixel is normalized by using $\mu$ and $\sigma$, and then re-scale and Figure 2: The size of feature
re-shift by $\gamma$ and $\beta.$
IN, LN, and BN share the formulation of Eqn.(1), but they use dif- malizers estimate statistics along
ferent sets of pixels to estimate $\mu$ and $\sigma.$ In other words, the numbers of their estimated statistics are different. In general, we have
2
<FigureHere>
$\operatorname*{maps~is~}N\times C\times H\times W\left(N=4\right.$ in this example). Different nordifferent axes.
|
[
[
0.1732177734375,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.57421875,
0.10919189453125,
0.5947265625,
0.10919189453125,
0.5947265625,
0.11932373046875,
0.57421875,
0.11932373046875
],
[
0.5595703125,
0.11932373046875,
0.564453125,
0.11932373046875,
0.564453125,
0.141357421875,
0.5595703125,
0.141357421875
],
[
0.57275390625,
0.12249755859375,
0.59716796875,
0.12249755859375,
0.59716796875,
0.1351318359375,
0.57275390625,
0.1351318359375
],
[
0.1519775390625,
0.1300048828125,
0.16015625,
0.1300048828125,
0.16015625,
0.1337890625,
0.1519775390625,
0.1337890625
],
[
0.1895751953125,
0.1351318359375,
0.214111328125,
0.1351318359375,
0.214111328125,
0.1427001953125,
0.1895751953125,
0.1427001953125
],
[
0.25,
0.1351318359375,
0.265625,
0.1351318359375,
0.265625,
0.1427001953125,
0.25,
0.1427001953125
],
[
0.358642578125,
0.1351318359375,
0.375732421875,
0.1351318359375,
0.375732421875,
0.1427001953125,
0.358642578125,
0.1427001953125
],
[
0.5595703125,
0.14453125,
0.564453125,
0.14453125,
0.564453125,
0.1514892578125,
0.5595703125,
0.1514892578125
],
[
0.57275390625,
0.1395263671875,
0.59619140625,
0.1395263671875,
0.59619140625,
0.15087890625,
0.57275390625,
0.15087890625
],
[
0.560546875,
0.1546630859375,
0.564453125,
0.1546630859375,
0.564453125,
0.1634521484375,
0.560546875,
0.1634521484375
],
[
0.56201171875,
0.1546630859375,
0.59326171875,
0.1546630859375,
0.59326171875,
0.1666259765625,
0.56201171875,
0.1666259765625
],
[
0.74658203125,
0.149658203125,
0.75634765625,
0.149658203125,
0.75634765625,
0.1546630859375,
0.74658203125,
0.1546630859375
],
[
0.1519775390625,
0.1590576171875,
0.160888671875,
0.1590576171875,
0.160888671875,
0.1641845703125,
0.1519775390625,
0.1641845703125
],
[
0.74169921875,
0.1578369140625,
0.75732421875,
0.1578369140625,
0.75732421875,
0.1666259765625,
0.74169921875,
0.1666259765625
],
[
0.560546875,
0.17041015625,
0.564453125,
0.17041015625,
0.564453125,
0.1767578125,
0.560546875,
0.1767578125
],
[
0.57275390625,
0.17041015625,
0.5947265625,
0.17041015625,
0.5947265625,
0.18310546875,
0.57275390625,
0.18310546875
],
[
0.74169921875,
0.169189453125,
0.75634765625,
0.169189453125,
0.75634765625,
0.177978515625,
0.74169921875,
0.177978515625
],
[
0.5595703125,
0.1787109375,
0.564453125,
0.1787109375,
0.564453125,
0.186279296875,
0.5595703125,
0.186279296875
],
[
0.736328125,
0.1787109375,
0.78125,
0.1787109375,
0.78125,
0.1900634765625,
0.736328125,
0.1900634765625
],
[
0.1519775390625,
0.189453125,
0.160888671875,
0.189453125,
0.160888671875,
0.1944580078125,
0.1519775390625,
0.1944580078125
],
[
0.55810546875,
0.1875,
0.564453125,
0.1875,
0.564453125,
0.1944580078125,
0.55810546875,
0.1944580078125
],
[
0.57421875,
0.189453125,
0.59326171875,
0.189453125,
0.59326171875,
0.1962890625,
0.57421875,
0.1962890625
],
[
0.416748046875,
0.1925048828125,
0.53515625,
0.1925048828125,
0.53515625,
0.199462890625,
0.416748046875,
0.199462890625
],
[
0.226318359375,
0.1944580078125,
0.2442626953125,
0.1944580078125,
0.2442626953125,
0.1988525390625,
0.226318359375,
0.1988525390625
],
[
0.2646484375,
0.1944580078125,
0.281005859375,
0.1925048828125,
0.281982421875,
0.1962890625,
0.265625,
0.1988525390625
],
[
0.295654296875,
0.1932373046875,
0.39208984375,
0.1932373046875,
0.39208984375,
0.2008056640625,
0.295654296875,
0.2008056640625
],
[
0.55615234375,
0.1962890625,
0.56787109375,
0.1962890625,
0.56787109375,
0.208984375,
0.55615234375,
0.208984375
],
[
0.7294921875,
0.1962890625,
0.775390625,
0.1962890625,
0.775390625,
0.2083740234375,
0.7294921875,
0.2083740234375
],
[
0.43310546875,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.20703125,
0.43310546875,
0.20703125
],
[
0.178955078125,
0.199462890625,
0.259033203125,
0.199462890625,
0.259033203125,
0.20703125,
0.178955078125,
0.20703125
],
[
0.2646484375,
0.2020263671875,
0.29736328125,
0.2020263671875,
0.29736328125,
0.2052001953125,
0.2646484375,
0.2052001953125
],
[
0.31201171875,
0.2020263671875,
0.34228515625,
0.2020263671875,
0.34228515625,
0.2052001953125,
0.31201171875,
0.2052001953125
],
[
0.36279296875,
0.2020263671875,
0.37744140625,
0.2020263671875,
0.37744140625,
0.2052001953125,
0.36279296875,
0.2052001953125
],
[
0.39453125,
0.2020263671875,
0.426513671875,
0.2020263671875,
0.426513671875,
0.205810546875,
0.39453125,
0.205810546875
],
[
0.48046875,
0.2020263671875,
0.52783203125,
0.2020263671875,
0.52783203125,
0.205810546875,
0.48046875,
0.205810546875
],
[
0.60693359375,
0.199462890625,
0.619140625,
0.199462890625,
0.619140625,
0.205810546875,
0.60693359375,
0.205810546875
],
[
0.650390625,
0.20703125,
0.6650390625,
0.20703125,
0.6650390625,
0.2108154296875,
0.650390625,
0.2108154296875
],
[
0.67333984375,
0.205810546875,
0.69287109375,
0.20263671875,
0.6953125,
0.210205078125,
0.6748046875,
0.212158203125
],
[
0.705078125,
0.20703125,
0.71337890625,
0.20703125,
0.71337890625,
0.2108154296875,
0.705078125,
0.2108154296875
],
[
0.61767578125,
0.210205078125,
0.6298828125,
0.210205078125,
0.6298828125,
0.2139892578125,
0.61767578125,
0.2139892578125
],
[
0.59619140625,
0.208984375,
0.611328125,
0.208984375,
0.611328125,
0.21337890625,
0.59619140625,
0.21337890625
],
[
0.33251953125,
0.21533203125,
0.3544921875,
0.21533203125,
0.3544921875,
0.2291259765625,
0.33251953125,
0.2291259765625
],
[
0.57861328125,
0.2215576171875,
0.68310546875,
0.2215576171875,
0.68310546875,
0.2342529296875,
0.57861328125,
0.2342529296875
],
[
0.685546875,
0.2215576171875,
0.80224609375,
0.2215576171875,
0.80224609375,
0.231689453125,
0.685546875,
0.231689453125
],
[
0.6689453125,
0.2412109375,
0.69140625,
0.2412109375,
0.69140625,
0.255126953125,
0.6689453125,
0.255126953125
],
[
0.1748046875,
0.261474609375,
0.826171875,
0.261474609375,
0.826171875,
0.274658203125,
0.1748046875,
0.274658203125
],
[
0.1732177734375,
0.2734375,
0.82421875,
0.2734375,
0.82421875,
0.287353515625,
0.1732177734375,
0.287353515625
],
[
0.1732177734375,
0.285888671875,
0.82421875,
0.285888671875,
0.82421875,
0.2998046875,
0.1732177734375,
0.2998046875
],
[
0.17236328125,
0.29736328125,
0.82177734375,
0.29736328125,
0.82177734375,
0.311279296875,
0.17236328125,
0.311279296875
],
[
0.1748046875,
0.313232421875,
0.82177734375,
0.313232421875,
0.82177734375,
0.323974609375,
0.1748046875,
0.323974609375
],
[
0.176513671875,
0.32568359375,
0.82275390625,
0.32568359375,
0.82275390625,
0.3359375,
0.176513671875,
0.3359375
],
[
0.17236328125,
0.3359375,
0.82421875,
0.3359375,
0.82421875,
0.349853515625,
0.17236328125,
0.349853515625
],
[
0.1732177734375,
0.349853515625,
0.363525390625,
0.349853515625,
0.363525390625,
0.363037109375,
0.1732177734375,
0.363037109375
],
[
0.1748046875,
0.392578125,
0.82275390625,
0.392578125,
0.82275390625,
0.406005859375,
0.1748046875,
0.406005859375
],
[
0.1748046875,
0.4072265625,
0.82177734375,
0.4072265625,
0.82177734375,
0.42041015625,
0.1748046875,
0.42041015625
],
[
0.1732177734375,
0.42041015625,
0.82177734375,
0.42041015625,
0.82177734375,
0.434326171875,
0.1732177734375,
0.434326171875
],
[
0.17236328125,
0.43310546875,
0.82421875,
0.434326171875,
0.82421875,
0.448974609375,
0.17236328125,
0.447509765625
],
[
0.1732177734375,
0.4462890625,
0.82275390625,
0.447509765625,
0.82275390625,
0.462158203125,
0.1732177734375,
0.4609375
],
[
0.1732177734375,
0.462158203125,
0.82177734375,
0.462158203125,
0.82177734375,
0.47607421875,
0.1732177734375,
0.47607421875
],
[
0.1748046875,
0.47607421875,
0.82275390625,
0.47607421875,
0.82275390625,
0.4892578125,
0.1748046875,
0.4892578125
],
[
0.176513671875,
0.4912109375,
0.82421875,
0.4912109375,
0.82421875,
0.501953125,
0.176513671875,
0.501953125
],
[
0.1748046875,
0.50390625,
0.82275390625,
0.50390625,
0.82275390625,
0.517578125,
0.1748046875,
0.517578125
],
[
0.1732177734375,
0.517578125,
0.82421875,
0.517578125,
0.82421875,
0.53076171875,
0.1732177734375,
0.53076171875
],
[
0.17236328125,
0.52978515625,
0.82421875,
0.52978515625,
0.82421875,
0.54345703125,
0.17236328125,
0.54345703125
],
[
0.1732177734375,
0.54541015625,
0.611328125,
0.54541015625,
0.611328125,
0.5595703125,
0.1732177734375,
0.5595703125
],
[
0.1748046875,
0.56494140625,
0.826171875,
0.56494140625,
0.826171875,
0.580078125,
0.1748046875,
0.580078125
],
[
0.1732177734375,
0.580078125,
0.515625,
0.580078125,
0.515625,
0.59326171875,
0.1732177734375,
0.59326171875
],
[
0.17724609375,
0.61865234375,
0.1895751953125,
0.61865234375,
0.1895751953125,
0.62890625,
0.17724609375,
0.62890625
],
[
0.1871337890625,
0.62060546875,
0.2017822265625,
0.62060546875,
0.2017822265625,
0.62744140625,
0.1871337890625,
0.62744140625
],
[
0.2034912109375,
0.61865234375,
0.5146484375,
0.61865234375,
0.5146484375,
0.6318359375,
0.2034912109375,
0.6318359375
],
[
0.1748046875,
0.65087890625,
0.69775390625,
0.65087890625,
0.69775390625,
0.6640625,
0.1748046875,
0.6640625
],
[
0.1748046875,
0.67041015625,
0.61376953125,
0.671875,
0.61376953125,
0.68603515625,
0.1748046875,
0.68505859375
],
[
0.744140625,
0.66943359375,
0.82177734375,
0.67236328125,
0.8203125,
0.68798828125,
0.74267578125,
0.68505859375
],
[
0.1748046875,
0.68603515625,
0.61279296875,
0.68603515625,
0.61279296875,
0.69970703125,
0.1748046875,
0.69970703125
],
[
0.17236328125,
0.69873046875,
0.61376953125,
0.69970703125,
0.61376953125,
0.7138671875,
0.17236328125,
0.71337890625
],
[
0.1748046875,
0.7138671875,
0.615234375,
0.7138671875,
0.615234375,
0.72802734375,
0.1748046875,
0.72802734375
],
[
0.1748046875,
0.72998046875,
0.61376953125,
0.72998046875,
0.61376953125,
0.74365234375,
0.1748046875,
0.74365234375
],
[
0.79736328125,
0.73095703125,
0.810546875,
0.73095703125,
0.810546875,
0.7421875,
0.79736328125,
0.7421875
],
[
0.1748046875,
0.74365234375,
0.615234375,
0.74365234375,
0.615234375,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1732177734375,
0.755859375,
0.484375,
0.755859375,
0.484375,
0.76953125,
0.1732177734375,
0.76953125
],
[
0.306396484375,
0.7822265625,
0.381591796875,
0.7890625,
0.378173828125,
0.80810546875,
0.303955078125,
0.80029296875
],
[
0.373291015625,
0.77783203125,
0.44775390625,
0.77978515625,
0.44775390625,
0.79443359375,
0.373291015625,
0.79248046875
],
[
0.44287109375,
0.7861328125,
0.4794921875,
0.78857421875,
0.47802734375,
0.7998046875,
0.441162109375,
0.79736328125
],
[
0.59228515625,
0.787109375,
0.615234375,
0.787109375,
0.615234375,
0.80029296875,
0.59228515625,
0.80029296875
],
[
0.78515625,
0.7890625,
0.79833984375,
0.7890625,
0.79833984375,
0.7998046875,
0.78515625,
0.7998046875
],
[
0.3798828125,
0.796875,
0.44287109375,
0.796875,
0.44287109375,
0.81103515625,
0.3798828125,
0.81103515625
],
[
0.176513671875,
0.8212890625,
0.61376953125,
0.8212890625,
0.61376953125,
0.83203125,
0.176513671875,
0.83203125
],
[
0.6611328125,
0.8173828125,
0.73779296875,
0.8193359375,
0.73779296875,
0.833984375,
0.6611328125,
0.83203125
],
[
0.1732177734375,
0.833984375,
0.615234375,
0.833984375,
0.615234375,
0.84765625,
0.1732177734375,
0.84765625
],
[
0.1748046875,
0.84765625,
0.61669921875,
0.84765625,
0.61669921875,
0.8623046875,
0.1748046875,
0.8623046875
],
[
0.6259765625,
0.84521484375,
0.82421875,
0.84521484375,
0.82421875,
0.859375,
0.6259765625,
0.859375
],
[
0.1732177734375,
0.8623046875,
0.3056640625,
0.8623046875,
0.3056640625,
0.87548828125,
0.1732177734375,
0.87548828125
],
[
0.62744140625,
0.85791015625,
0.826171875,
0.85791015625,
0.826171875,
0.8720703125,
0.62744140625,
0.8720703125
],
[
0.62841796875,
0.87255859375,
0.82177734375,
0.87255859375,
0.82177734375,
0.88330078125,
0.62841796875,
0.88330078125
],
[
0.1748046875,
0.88330078125,
0.61376953125,
0.88330078125,
0.61376953125,
0.896484375,
0.1748046875,
0.896484375
],
[
0.625,
0.88330078125,
0.826171875,
0.8837890625,
0.826171875,
0.89794921875,
0.625,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.61669921875,
0.89794921875,
0.61669921875,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.62744140625,
0.89794921875,
0.71337890625,
0.89794921875,
0.71337890625,
0.9091796875,
0.62744140625,
0.9091796875
],
[
0.1732177734375,
0.9111328125,
0.615234375,
0.9111328125,
0.615234375,
0.9248046875,
0.1732177734375,
0.9248046875
],
[
0.488525390625,
0.94873046875,
0.50244140625,
0.9443359375,
0.5087890625,
0.958984375,
0.495849609375,
0.96337890625
]
] |
[
"Published as a conference paper at ICLR 2019",
"77.0",
"oel",
"76.0",
"",
"BN",
"IN",
"GN",
"",
"75.0",
"1",
"74.0",
"RN",
"0.5",
"GN",
"",
"73.0",
"SN",
"E",
"-IdealBN",
"0.0",
"",
"72.0",
"ationImageStyleNeuralArchitecture",
"Detec",
"Det.",
"Detection Segmentation",
"8",
"63",
"Transfer",
"Classification Backbone",
"BnvHeg",
"asraa",
"F20",
"Citvseanrs",
"SearhST",
"",
"0",
"(0t",
"O",
"",
"",
"(a)",
"(number ofGPUs,nun",
"berofsamplesperGPUy",
"(b)",
"Figure l: (a) shows that SN adapts to various networks and tasks by learning importance ratios to select",
"normalizers. In (a), a ratio is between O and 1 and all ratios of each task sum to 1. (b) shows the top-1 accuracies",
"of ResNet50 trained with SN on ImageNet and compared with BN and GN in different batch settings. The",
"gradients in training are averaged over all GPUs and the statistics of normalizers are estimated in each GPU",
"For instance,all methods are compared toan ideal case,\"ideal BN,whoseaccuracies are76.4%for all settings.",
"Thisideal case cannotbeobtained inpractice.Infact,whentheminibatchsizedecreases,BN's accuraciesdrop",
"significantly, while SN and GN both maintain reasonably good performance. SN surpasses or is comparable to",
"both BN and GN in all settings.",
"Overall, this work has three key contributions. (1) We introduce Switchable Normalization (SN)",
"which is applicable in both CNNs and RNNs/LSTMs, and improves the other normalization tech-",
"niques on many challenging benchmarks and tasks including image recognition in ImageNet (Rus-",
"sakovsky et al., 2015), object detection in COCO (Lin et al., 2014), scene parsing in Cityscapes",
"(Cordts et al., 2016) and ADE20K (Zhou et al., 2017), artistic image stylization (Johnson et al.",
"2016), neural architecture search (Pham et al., 2018), and video recognition in Kinetics (Kay et al.,",
"2017). (2) The analyses of SN are presented where multiple normalizers can be compared and un-",
"derstoodwithgeometricinterpretation.(3)Byenablingeachnormalizationlayerinadeepnetwork",
"to have its own operation, SN helps ease the usage of normalizers, pushes the frontier of normal-",
"ization in deep learning, as well as opens up new research direction.We believe that all existing",
"models could be reexamined with this new perspective. We'll make the code of SN available and",
"recommend it as an alternative of existing handcrafted approaches.",
"In the following sections, we first present SN in Sec.2 and then discuss its relationships with previous",
"work in Sec.3. SN is evaluated extensively in Sec.4.",
"2",
"",
"SWITCHABLE NORMALIZATION (SN)",
"We describe a general formulation of a normalization layer and then present SN",
"A General Form. We take CNN as an illustrative example. Let h",
"C channels",
"be the input data of an arbitrary normalization layer represented by",
"a 4D tensor (N,C, H, W), indicating number of samples, number",
"of channels, height and width of a channel respectively as shown in",
"Fig.2. Let hncij and hncij be a pixel before and after normalization,",
"H",
"where n [1, N], c e [1, C], i e [1, H], and j e [1, W]. Let μ and",
" be a mean and a standard deviation. We have",
"hncij=",
"hncij-μ",
"+B.",
"(1)",
"W",
"Va2 +e",
"whereand β are a scale and ashiftparameterrespectively.eis",
"Nsamples",
"a small constant to preserve numerical stability. Eqn.(1) shows that",
"each pixel is normalized by using μ and o, and then re-scale and",
"Figure 2: The size of feature",
"re-shift by and β.",
"maps is N×C×H×W (N= 4",
"in this example).Differentnor-",
"IN, LN, and BN share the formulation of Eqn.(1), but they use dif-",
"malizers estimate statistics along",
"ferent sets of pixels to estimate μ and o. In other words, the num-",
"different axes.",
"bers of their estimated statistics are different. In general, we have",
"2"
] |
{
"arxiv_id": "1806.10779",
"language": "en",
"page_idx": 1
}
|
||
1806.10779-en-2
|
1806.10779-en-2.jpg
|
Published as a conference paper at ICLR 2019
(2)
$$\mu_k=\frac{1}{|I_k|}\sum\limits_{(n,c,i,j)\in I_k}h_{ncij},\quad\sigma_k^2=\frac{1}{|I_k|}\sum\limits_{(n,c,i,j)\in I_k}(h_{ncij}-\mu_k)^2,$$
$\begin{array}{l}\mathrm{where~}k\in\{\mathrm{in,ln,bn}\}\text{ is used to distinguish different methods. }I_{k}\text{ is a set pixels and }|I_{k}|\text{ denotes}\\\mathrm{the~number~of~pixels.~Specifically,~}I_{\mathrm{in}},I_{\mathrm{ln}},\mathrm{~and~}I_{\mathrm{bn}}\text{ are the sets of pixels used to compute statistics}\end{array}$
in different approaches.
IN was established in the task of artistic image style transfer (Johnson et al., 2016; Huang & Belongie, 2017) . In IN, we have $\mu _{\mathrm{in}}, \sigma _{\mathrm{in}}^{2}\in \mathbb{R} ^{N\times C}$ and $I_{\mathrm{in}}= \{ ( i, j) | i\in [ 1, H] , j\in [ 1, W] \} $,meaning that IN has 2NC elements of statistics, where each mean and variance value is computed along ( Tor W) $(H,W)$ for each channel of each sample.
$( 11, v)$ iot cacur cuchi saniper. LN ( Ba et al, 2016) was proposed to ease optimization of recurrent neural networks ( RNNs) . In LN, we have $\mu _{\ln }$, $\sigma _{\ln }^{2}\in \mathbb{R} ^{N\times 1}$ and $I_{\ln }$= $\{$( c, i, j) | c$\in$[ 1, C] , i$\in$[ 1, H] , j$\in$[ 1, W] $\}$, implying that LN has 2N statis
for each one of the $N$ samples.
BN (Ioffe & Szegedy, 2015) was first demonstrated in the task of image classification (He et al., 2016; Krizhevsky et al., 2012) by normalizing the hidden feature maps of CNNs. In BN, we have $\begin{array}{l}{\mu_{\mathrm{bn}},\sigma_{\mathrm{bn}}^{2}\in\mathbb{R}^{C\times1}\mathrm{~and~}I_{\mathrm{bn}}=\{(n,i,j)|n\in[1,N],i\in[1,H],j\in[1,W]\},\text{in the sense that BN}}\\{\text{treats each channel independently like N, but not only normalizes across }(H,W),\text{but also the }N}\end{array}$ samples in a minibatch, leading to $2C$ elements of statistics.
## 2.1 FORMULATION OF SN
SN has an intuitive expression
(3)
$$\hat{h}_{ncij}=\gamma\frac{h_{ncij}-\Sigma_{k\in\Omega}w_{k}\mu_{k}}{\sqrt{\Sigma_{k\in\Omega}w_{k}^{\prime}\sigma_{k}^{2}+\epsilon}}+\beta,$$
where $\Omega$ is a set of statistics estimated in different ways. In this work, we define $\Omega=\{$in,ln,bn$\}$ the same as above where $\mu_k$ and $\sigma_k^2$ can be calculated by following Eqn.(2). However, this strategy leads to large redundant computations. In fact, the three kinds of statistics of SN depend on each other. Therefore we could reduce redundancy by reusing computations,
$$\begin{aligned}&\mu_{\mathrm{in}}&&=\quad\frac{1}{HW}\sum_{i,j}^{H,W}h_{ncij},\quad\sigma_{\mathrm{in}}^{2}=\frac{1}{HW}\sum_{i,j}^{H,W}(h_{ncij}-\mu_{\mathrm{in}})^{2},\\&\mu_{\mathrm{ln}}&&=\quad\frac{1}{C}\sum_{c=1}^{C}\mu_{\mathrm{in}},\quad\sigma_{\mathrm{ln}}^{2}=\frac{1}{C}\sum_{c=1}^{C}(\sigma_{\mathrm{in}}^{2}+\mu_{\mathrm{in}}^{2})-\mu_{\mathrm{ln}}^{2},\\&\mu_{\mathrm{bn}}&&=\quad\frac{1}{N}\sum_{n=1}^{N}\mu_{\mathrm{in}},\quad\sigma_{\mathrm{bn}}^{2}=\frac{1}{N}\sum_{n=1}^{N}(\sigma_{\mathrm{in}}^{2}+\mu_{\mathrm{in}}^{2})-\mu_{\mathrm{bn}}^{2},\\\end{aligned}$$
(4)
showing that the means and variances of LN and BN can be computed based on IN. Using Eqn.(4),
the computational complexity of SN is $\mathcal{O}(NCHW),\hat\text{which is comparable to previous work.}$
Furthermore, $w_k$ and $w_k^\prime$ in Eqn.(3) are importance ratios used to weighted average the means and variances respectively. Each $w_k$ or $w_k^{\prime}$ is a scalar variable, which is shared across all channels. There are $3\times2=6$ importance weights in SN. We have $\Sigma_k\in\Omega w_k=1,\Sigma_{k\in\Omega}w_k^{\prime}=1$, and $\forall w_k,w_k^{\prime}\in[0,1]$, and define
$$w_k=\frac{e^{\lambda_k}}{\Sigma_{z\in\{\text{in,ln,bn}\}}e^{\lambda_z}}\:\text{and}\:k\in\{\text{in,ln,bn}\}.$$
(5)
Here each $w_k$ is computed by using a softmax function with $\lambda_\mathrm{in},\lambda_\mathrm{ln}$, and $\lambda_\mathrm{bn}$ as the control parameters, which can be learned by back-propagation (BP). $w_k^{\prime}$ are defined similarly by using another three control parameters $\lambda_\mathrm{in}^{\prime},\lambda_\mathrm{ln}^{\prime}$,and $\lambda_\mathrm{bn}^{\prime}.$
Training. Let $\Theta$ be a set of network parameters (e.g. filters) and $\Phi$ be a set of control parameters that control the network architecture. In SN, we have $\Phi=\{\lambda_{\mathrm{in}},\lambda_{\mathrm{ln}},\lambda_{\mathrm{bn}},\lambda_{\mathrm{in}}^{\prime},\lambda_{\mathrm{ln}}^{\prime},\lambda_{\mathrm{bn}}^{\prime}\}.$ Training deep network with SN is to minimize a loss function $\mathcal{L}(\Theta,\Phi)$,where $\Theta$ and $\Phi$ can be optimized
3
|
[
[
0.1748046875,
0.0328369140625,
0.4794921875,
0.0328369140625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.328369140625,
0.11553955078125,
0.343017578125,
0.11553955078125,
0.343017578125,
0.1256103515625,
0.328369140625,
0.1256103515625
],
[
0.529296875,
0.11614990234375,
0.5458984375,
0.11614990234375,
0.5458984375,
0.1256103515625,
0.529296875,
0.1256103515625
],
[
0.57177734375,
0.12060546875,
0.60693359375,
0.12060546875,
0.60693359375,
0.1363525390625,
0.57177734375,
0.1363525390625
],
[
0.404296875,
0.121826171875,
0.5146484375,
0.121826171875,
0.5146484375,
0.1395263671875,
0.404296875,
0.1395263671875
],
[
0.371826171875,
0.121826171875,
0.393798828125,
0.121826171875,
0.393798828125,
0.1337890625,
0.371826171875,
0.1337890625
],
[
0.6005859375,
0.12249755859375,
0.71728515625,
0.12249755859375,
0.71728515625,
0.136962890625,
0.6005859375,
0.136962890625
],
[
0.80224609375,
0.12249755859375,
0.82421875,
0.12249755859375,
0.82421875,
0.136962890625,
0.80224609375,
0.136962890625
],
[
0.5146484375,
0.1307373046875,
0.55615234375,
0.1307373046875,
0.55615234375,
0.1458740234375,
0.5146484375,
0.1458740234375
],
[
0.318603515625,
0.1319580078125,
0.3505859375,
0.1319580078125,
0.3505859375,
0.14453125,
0.318603515625,
0.14453125
],
[
0.5498046875,
0.14013671875,
0.62158203125,
0.14013671875,
0.62158203125,
0.1514892578125,
0.5498046875,
0.1514892578125
],
[
0.3505859375,
0.141357421875,
0.419189453125,
0.141357421875,
0.419189453125,
0.1514892578125,
0.3505859375,
0.1514892578125
],
[
0.281005859375,
0.1256103515625,
0.314453125,
0.1256103515625,
0.314453125,
0.136962890625,
0.281005859375,
0.136962890625
],
[
0.1732177734375,
0.1558837890625,
0.82421875,
0.1572265625,
0.82421875,
0.1717529296875,
0.1732177734375,
0.17041015625
],
[
0.1748046875,
0.1717529296875,
0.82421875,
0.1717529296875,
0.82421875,
0.1849365234375,
0.1748046875,
0.1849365234375
],
[
0.1732177734375,
0.1849365234375,
0.334228515625,
0.1849365234375,
0.334228515625,
0.1988525390625,
0.1732177734375,
0.1988525390625
],
[
0.1748046875,
0.205810546875,
0.82177734375,
0.205810546875,
0.82177734375,
0.2197265625,
0.1748046875,
0.2197265625
],
[
0.17236328125,
0.21533203125,
0.82666015625,
0.21533203125,
0.82666015625,
0.236083984375,
0.17236328125,
0.236083984375
],
[
0.1748046875,
0.2342529296875,
0.82421875,
0.2342529296875,
0.82421875,
0.2474365234375,
0.1748046875,
0.2474365234375
],
[
0.1748046875,
0.248779296875,
0.44677734375,
0.248779296875,
0.44677734375,
0.261962890625,
0.1748046875,
0.261962890625
],
[
0.1732177734375,
0.268310546875,
0.82275390625,
0.268310546875,
0.82275390625,
0.2822265625,
0.1732177734375,
0.2822265625
],
[
0.17236328125,
0.279052734375,
0.826171875,
0.279052734375,
0.826171875,
0.2998046875,
0.17236328125,
0.2998046875
],
[
0.1732177734375,
0.29541015625,
0.82275390625,
0.296630859375,
0.82275390625,
0.311279296875,
0.1732177734375,
0.31005859375
],
[
0.1748046875,
0.311279296875,
0.378173828125,
0.311279296875,
0.378173828125,
0.324462890625,
0.1748046875,
0.324462890625
],
[
0.176513671875,
0.33203125,
0.82177734375,
0.33203125,
0.82177734375,
0.34521484375,
0.176513671875,
0.34521484375
],
[
0.1748046875,
0.34521484375,
0.82421875,
0.34521484375,
0.82421875,
0.359130859375,
0.1748046875,
0.359130859375
],
[
0.1707763671875,
0.35595703125,
0.826171875,
0.354248046875,
0.826171875,
0.375,
0.1707763671875,
0.376953125
],
[
0.1732177734375,
0.37255859375,
0.82275390625,
0.371826171875,
0.82275390625,
0.386474609375,
0.1732177734375,
0.386962890625
],
[
0.1732177734375,
0.38818359375,
0.56884765625,
0.38818359375,
0.56884765625,
0.401611328125,
0.1732177734375,
0.401611328125
],
[
0.1732177734375,
0.41845703125,
0.369384765625,
0.41845703125,
0.369384765625,
0.431884765625,
0.1732177734375,
0.431884765625
],
[
0.1732177734375,
0.443115234375,
0.375732421875,
0.44580078125,
0.375732421875,
0.460205078125,
0.1732177734375,
0.458251953125
],
[
0.443603515625,
0.462158203125,
0.5849609375,
0.464111328125,
0.5849609375,
0.48291015625,
0.443603515625,
0.48095703125
],
[
0.3798828125,
0.473388671875,
0.44287109375,
0.47607421875,
0.44287109375,
0.490478515625,
0.3798828125,
0.488037109375
],
[
0.576171875,
0.4716796875,
0.615234375,
0.474853515625,
0.61376953125,
0.4873046875,
0.57421875,
0.484130859375
],
[
0.80224609375,
0.474853515625,
0.826171875,
0.474853515625,
0.826171875,
0.4892578125,
0.80224609375,
0.4892578125
],
[
0.455810546875,
0.48486328125,
0.57666015625,
0.48486328125,
0.57666015625,
0.5,
0.455810546875,
0.5
],
[
0.1748046875,
0.5068359375,
0.82177734375,
0.5068359375,
0.82177734375,
0.52099609375,
0.1748046875,
0.52099609375
],
[
0.1748046875,
0.521484375,
0.82275390625,
0.521484375,
0.82275390625,
0.5361328125,
0.1748046875,
0.5361328125
],
[
0.1748046875,
0.53515625,
0.82421875,
0.53515625,
0.82421875,
0.548828125,
0.1748046875,
0.548828125
],
[
0.1732177734375,
0.5498046875,
0.64208984375,
0.5498046875,
0.64208984375,
0.56298828125,
0.1732177734375,
0.56298828125
],
[
0.39794921875,
0.5693359375,
0.428955078125,
0.5693359375,
0.428955078125,
0.58203125,
0.39794921875,
0.58203125
],
[
0.57666015625,
0.5673828125,
0.609375,
0.5673828125,
0.609375,
0.58349609375,
0.57666015625,
0.58349609375
],
[
0.375732421875,
0.57373046875,
0.38818359375,
0.5791015625,
0.3798828125,
0.5888671875,
0.36767578125,
0.583984375
],
[
0.39208984375,
0.5771484375,
0.47802734375,
0.5771484375,
0.47802734375,
0.60693359375,
0.39208984375,
0.60693359375
],
[
0.55810546875,
0.57373046875,
0.56884765625,
0.58203125,
0.55810546875,
0.59033203125,
0.54736328125,
0.58203125
],
[
0.57275390625,
0.5771484375,
0.71142578125,
0.5771484375,
0.71142578125,
0.60595703125,
0.57275390625,
0.60595703125
],
[
0.291748046875,
0.5869140625,
0.324462890625,
0.5869140625,
0.324462890625,
0.599609375,
0.291748046875,
0.599609375
],
[
0.33251953125,
0.58837890625,
0.343017578125,
0.59130859375,
0.340576171875,
0.59765625,
0.328369140625,
0.5947265625
],
[
0.494384765625,
0.583984375,
0.5361328125,
0.583984375,
0.5361328125,
0.60107421875,
0.494384765625,
0.60107421875
],
[
0.359375,
0.59033203125,
0.39453125,
0.59033203125,
0.39453125,
0.60595703125,
0.359375,
0.60595703125
],
[
0.5390625,
0.59228515625,
0.57275390625,
0.59228515625,
0.57275390625,
0.60498046875,
0.5390625,
0.60498046875
],
[
0.40283203125,
0.60107421875,
0.423095703125,
0.60107421875,
0.423095703125,
0.61572265625,
0.40283203125,
0.61572265625
],
[
0.58251953125,
0.6015625,
0.6044921875,
0.6015625,
0.6044921875,
0.61572265625,
0.58251953125,
0.61572265625
],
[
0.381591796875,
0.6142578125,
0.400390625,
0.6142578125,
0.400390625,
0.6298828125,
0.381591796875,
0.6298828125
],
[
0.5087890625,
0.619140625,
0.521484375,
0.62890625,
0.5107421875,
0.63720703125,
0.498291015625,
0.62744140625
],
[
0.52294921875,
0.61572265625,
0.5400390625,
0.61572265625,
0.5400390625,
0.62890625,
0.52294921875,
0.62890625
],
[
0.363525390625,
0.62255859375,
0.374267578125,
0.62255859375,
0.374267578125,
0.63330078125,
0.363525390625,
0.63330078125
],
[
0.371826171875,
0.62548828125,
0.435546875,
0.62548828125,
0.435546875,
0.6484375,
0.371826171875,
0.6484375
],
[
0.455810546875,
0.6298828125,
0.484375,
0.6298828125,
0.484375,
0.646484375,
0.455810546875,
0.646484375
],
[
0.490234375,
0.625,
0.6689453125,
0.61572265625,
0.67236328125,
0.6484375,
0.49267578125,
0.65771484375
],
[
0.291748046875,
0.63330078125,
0.3203125,
0.63330078125,
0.3203125,
0.64599609375,
0.291748046875,
0.64599609375
],
[
0.365234375,
0.63330078125,
0.38232421875,
0.64013671875,
0.371826171875,
0.65576171875,
0.3544921875,
0.6484375
],
[
0.378173828125,
0.646484375,
0.40283203125,
0.646484375,
0.40283203125,
0.6591796875,
0.378173828125,
0.6591796875
],
[
0.51953125,
0.646484375,
0.54345703125,
0.646484375,
0.54345703125,
0.66015625,
0.51953125,
0.66015625
],
[
0.384033203125,
0.6591796875,
0.40185546875,
0.6591796875,
0.40185546875,
0.67236328125,
0.384033203125,
0.67236328125
],
[
0.53173828125,
0.6591796875,
0.5498046875,
0.6591796875,
0.5498046875,
0.67236328125,
0.53173828125,
0.67236328125
],
[
0.381591796875,
0.66552734375,
0.404296875,
0.67041015625,
0.39453125,
0.69580078125,
0.371826171875,
0.68994140625
],
[
0.52392578125,
0.66748046875,
0.6806640625,
0.67236328125,
0.6787109375,
0.6943359375,
0.52294921875,
0.689453125
],
[
0.50732421875,
0.66845703125,
0.51953125,
0.6640625,
0.52685546875,
0.6767578125,
0.5146484375,
0.68115234375
],
[
0.365234375,
0.66845703125,
0.37744140625,
0.66845703125,
0.37744140625,
0.6787109375,
0.365234375,
0.6787109375
],
[
0.459228515625,
0.673828125,
0.496826171875,
0.67822265625,
0.49267578125,
0.693359375,
0.455810546875,
0.689453125
],
[
0.80224609375,
0.6748046875,
0.826171875,
0.6748046875,
0.826171875,
0.689453125,
0.80224609375,
0.689453125
],
[
0.398681640625,
0.67822265625,
0.43701171875,
0.67822265625,
0.43701171875,
0.689453125,
0.398681640625,
0.689453125
],
[
0.28759765625,
0.6767578125,
0.32275390625,
0.6767578125,
0.32275390625,
0.689453125,
0.28759765625,
0.689453125
],
[
0.35693359375,
0.68310546875,
0.375732421875,
0.6787109375,
0.3798828125,
0.6923828125,
0.361083984375,
0.6962890625
],
[
0.5048828125,
0.6845703125,
0.53369140625,
0.68115234375,
0.53515625,
0.693359375,
0.50732421875,
0.69580078125
],
[
0.381591796875,
0.693359375,
0.40185546875,
0.69140625,
0.40283203125,
0.69970703125,
0.38232421875,
0.7021484375
],
[
0.53125,
0.6923828125,
0.55126953125,
0.6923828125,
0.55126953125,
0.70068359375,
0.53125,
0.70068359375
],
[
0.1732177734375,
0.70703125,
0.82177734375,
0.70849609375,
0.82177734375,
0.72265625,
0.1732177734375,
0.7216796875
],
[
0.1748046875,
0.72265625,
0.77783203125,
0.72265625,
0.77783203125,
0.736328125,
0.1748046875,
0.736328125
],
[
0.1748046875,
0.74365234375,
0.82421875,
0.74365234375,
0.82421875,
0.75830078125,
0.1748046875,
0.75830078125
],
[
0.1748046875,
0.75830078125,
0.82421875,
0.75830078125,
0.82421875,
0.771484375,
0.1748046875,
0.771484375
],
[
0.1732177734375,
0.771484375,
0.82421875,
0.771484375,
0.82421875,
0.7861328125,
0.1732177734375,
0.7861328125
],
[
0.1748046875,
0.787109375,
0.2459716796875,
0.787109375,
0.2459716796875,
0.798828125,
0.1748046875,
0.798828125
],
[
0.428955078125,
0.79345703125,
0.45751953125,
0.79345703125,
0.45751953125,
0.81103515625,
0.428955078125,
0.81103515625
],
[
0.3447265625,
0.80615234375,
0.3896484375,
0.80810546875,
0.38818359375,
0.82275390625,
0.343017578125,
0.82080078125
],
[
0.5009765625,
0.80615234375,
0.64794921875,
0.80615234375,
0.64794921875,
0.8193359375,
0.5009765625,
0.8193359375
],
[
0.80224609375,
0.80615234375,
0.826171875,
0.80615234375,
0.826171875,
0.82080078125,
0.80224609375,
0.82080078125
],
[
0.1748046875,
0.833984375,
0.82177734375,
0.833984375,
0.82177734375,
0.84765625,
0.1748046875,
0.84765625
],
[
0.38232421875,
0.81103515625,
0.5087890625,
0.8095703125,
0.5087890625,
0.83056640625,
0.38232421875,
0.83349609375
],
[
0.1732177734375,
0.8466796875,
0.82421875,
0.84765625,
0.82421875,
0.8623046875,
0.1732177734375,
0.861328125
],
[
0.1748046875,
0.8623046875,
0.455078125,
0.8623046875,
0.455078125,
0.876953125,
0.1748046875,
0.876953125
],
[
0.176513671875,
0.8818359375,
0.82421875,
0.8818359375,
0.82421875,
0.896484375,
0.176513671875,
0.896484375
],
[
0.17236328125,
0.89599609375,
0.82421875,
0.89599609375,
0.82421875,
0.91357421875,
0.17236328125,
0.91357421875
],
[
0.17236328125,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.17236328125,
0.9248046875
],
[
0.49267578125,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2019",
"1",
"1",
"7",
"hncij,%",
"",
"(hncij-μk)?",
"(2)",
"[]",
"JIkl",
"(n,c,i,j)eI",
"(n,c,i)elr",
"uk=",
"where k E fin, In, bn) is used to distinguish different methods. I is a set pixels and Ikdenotes",
"the number of pixels. Specifically, Iin, Iin, and Ibn are the sets of pixels used to compute statistics",
"in different approaches.",
"IN was established in the task of artistic image style transfer (Johnson et al., 2016; Huang & Be-",
"longie, 2017). In IN, we have pin, o2 e RNXC and Iin = (i,5)]i e [1, H],j e [1, W], meaning",
"that IN has 2NC elements of statistics, where each mean and variance value is computed along",
"(H, W) for each channel of each sample.",
"LN (Ba et al., 2016) was proposed to ease optimization of recurrent neural networks (RNNs). In",
"LN, we have μin,o. e RNxi and Iin = (c,i, 5]lc e [1, C],i e [1, H], j e [1, W], implying that",
"LN has 2N statistical values, where a mean value and a variance value are computed in (C, H, W)",
"for each one of the N samples.",
"BN (loffe & Szegedy, 2015) was first demonstrated in the task of image classification (He et al.",
"2016; Krizhevsky et al., 2012) by normalizing the hidden feature maps of CNNs. In BN, we have",
"μbn, obn e IRCx1 and Ibn = (n, i, 5)]n e [1, M],i e [1, H],j e [1, W], in the sense that BN",
"treats each channel independently like IN, but not only normalizes across (H, W), but also the N",
"samples in a minibatch, leading to 2C elements of statistics.",
"2.1 FORMULATION OF SN",
"SN has an intuitive expression",
"hncij -Ekeowkuk",
"hneij=",
"+B",
"(3)",
"VEkenwho+e",
"where is a set of statistics estimated in different ways. In this work, we define =[in, In, bn]",
"the same as above where μk and can be calculated by following Eqn.(2). However, this strategy",
"leads to large redundant computations: In fact, the three kinds of statistics of SN depend on each",
"other. Therefore we could reduce redundancy by reusing computations,",
"H,W",
"H,W",
"1",
"hnei",
"",
"(hneij - Mam),",
"Min",
"1",
"=",
"HW",
"HW",
"i,j",
"i.j",
"c",
"1",
"c",
"1",
"Zuin,",
"oim",
"品+品）一品",
"μln",
"C",
"c=1",
"c=1",
"N",
"N",
"2",
"之(品+品）",
"1",
"1 ",
"Obn",
"(4)",
"Hin,",
"μbn",
"N",
"N",
"n=",
"n=1",
"showing that the means and variances of LN and BN can be computed based on IN. Using Eqn.(4)",
"the computational complexity of SNis O(NCHW), which is comparable to previous work.",
"Furthermore, wk and w, in Eqn.(3) are importance ratios used to weighted average the means and",
"variances respectively. Each wk or w' is a scalar variable, which is shared across all channels. There",
"are 3x2 = 6 importance weights in SN. We have Ekeawk = 1, kenw', = 1, and Vwk, w, E [0, 1],",
"and define",
"edk",
"Wk=",
"and k e [in, In, bn]",
"(5)",
"Here each w is computed by using a softmax function with Ain, Ain, and Abn as the control param-",
"ve (in,In,bn)e.",
"eters, which can be learned by back-propagation (BP). w, are defined similarly by using another",
"three control parameters Xin,Ain, and Xbn.",
"Training. Let be a set of network parameters (e.g. filters) and be a set of control parameters",
"that control the network architecture. In SN, we have = (Ain, Ain, Abn, Xin, Ain, Abn]. Training",
"a deep network with SN is to minimize a loss function C(O, ), where and @ can be optimized",
"3"
] |
{
"arxiv_id": "1806.10779",
"language": "en",
"page_idx": 2
}
|
||
1806.10779-en-3
|
1806.10779-en-3.jpg
|
Published as a conference paper at ICLR 2019
jointly by back-propagation (BP). This training procedure is different from previous meta-learning algorithms such as network architecture search (Colson et al., 2007; Liu et al., 2018; Pham et al., $2\tilde{0}18).$ In previous work, $\Phi$ represents as a set of network modules with different learning capacities, where $\Theta$ and $\Phi$ were optimized in two BP stages iteratively by using two training sets that are nonoverlapped. For example, previous work divided an entire training set into a training and a validation set. However, if $\Theta$ and $\Phi$ in previous work are optimized in the same set of training data, $\Phi$ would choose the module with large complexity to overfit these data. In contrast, SN essentially prevents overfitting by choosing normalizers to improve both learning and generalization ability as discussed below.
Analyses of SN. To understand SN, we theoretically compare SN with BN, IN, and LN by representing them using weight normalization (WN) (Salimans & Kingma, 2016) that is independent of mean and variance. WN is computed as $v\frac{\mathbf{w}^{\top}\mathbf{x}}{\|\mathbf{w}\|_{2}}$, where w and x represent a filter and an image patch. WN normalizes the norm of each filter to 1 and rescales to $v.$
Remark 1. Let x be an image patch with zero mean and unit variance, w $_i$ be a filter of the i- $thchannel,i\in\{1,2,...,C\},andvbetheflternormofWN.Eqn.(1)canberewrittenby\hat{h}_{\mathrm{in}}=$ $\gamma\frac{\mathbf{w}_{i}^{\mathrm{T}}\mathbf{x}}{\|\mathbf{w}_{i}\|_{2}}+\beta;\hat{h}_{\mathrm{bn}}=\gamma\frac{\mathbf{w}_{i}^{\mathrm{T}}\mathbf{x}}{\|\mathbf{w}_{i}\|_{2}}+\beta$, s.t.$\gamma \leq v; and\hat{h} _{\mathrm{ln}}= \gamma \frac {\mathbf{w} _{i}^{\mathrm{T} }\mathbf{x} }{\| \mathbf{w} _{i}\| _{2}+ \sum _{j\neq i}^{C}\| \mathbf{w} _{j}\| _{2}}+ \beta . Bycombining$ $them$, $SNinEqn. ( 3) canbereformulatedby$ $\hat{h} _{\mathrm{sn}}= w_{\mathrm{in}}\hat{h} _{\mathrm{in}}+ w_{\mathrm{bn}}\hat{h} _{\mathrm{bn}}+ w_{\mathrm{ln}}\hat{h} _{\mathrm{ln}}= \gamma \frac {\mathbf{w} _{i}^{\top }\mathbf{x} }{\| \mathbf{w} _{i}\| _{2}+ w_{\mathrm{ln}}\sum _{j\neq i}^{C}\| \mathbf{w} _{j}\| _{2}}+$ $\beta$, s.t. $w_{\mathrm{bn}}\gamma \leq v, where$ $w_{\mathrm{in}}, w_{\mathrm{ln}}, w_{\mathrm{bn}}are$ the weights and $\gamma \leq v$ $is$ $a$ $constraint.$
<FigureHere>
Remark 1 simplifes SN in Eqn.(3), enabling us to compare different normalizers geometrically by formulating them with respect to WN. In Fig.3, $\hat{h}_{\mathrm{in}}$ of IN can be computed similarly to WN with an additional bias $\beta$, where the norms of all filters are normalized to l and then rescaled to $\gamma.$ As $\gamma$ and $v$ have the same learning dynamic, the length of $\gamma$ would be identically to $v$ (see w_1,w_2 of IN). Moreover, $\hat{h}_{\mathrm{bn}}$ in BN can be rewritten as WN with regularization over $\gamma$, making it shorter than $v.$ Compared to IN and LN, Luo et al. (2019) shows that the regularization of BN improves generalization and increases angle between filters, preventing them from coadaptation (see $\mathbf{w} _{1}, \mathbf{w} _{2}$ of BN). Furthermore, $\hat{h} _{\ln }$ in LN normalizes each filter among channels where the filter norm is less con- Figure 3: Geometric view of directions and strained than IN and BN. That is. LN allows $\gamma > v$ tolengths of the f
strained thai IN and BN. That is, LN allows $\gamma>v$ to $\textbf{lengths or the nifers m}\mathbf{1N}$,increase learning ability. Finally, $\hat{h}_\mathrm{sn}$ in SN inherits the by comparing them to WN.
benefits from all of them and enables balance between
learning and generalization ability. For example, when the batch size is small, the random noise
from the batch statistics of BN would be too strong. SN is able to maintain performance by decreasing $w_\mathrm{bn}$ and increasing $w_\mathrm{ln}$, such that the regularization from BN is reduced and the learning ability is enhanced by LN. This phenomenon is supported by our experiment. More results are provided in Appendix B.
Variants of SN. SN has many extensions. For instance, a pretrained network with SN can be finetuned by applying the argmax function on its control parameters where each normalization layer selects only one normalizer, leading to sparse SN. For (8,32) as an example, SN with sparsity achieves top-1 accuracy of $77.0\%$ in ImageNet with ResNet50, which is comparable to 76.9% of SN without sparsity. Moreover, when the channels are divided into groups, each group could select its own normalizer to increase representation power of SN. Our preliminary results suggest that group SN performs better than SN in some senses. For instance, group SN with only two groups boosts the top-1 accuracy of ResNet50 to 77.2% in ImageNet. The above two variants will be presented as future work due to the length of paper. This work focuses on SN where the importance weights are tied between channels.
Inference. When applying SN in test, the statistics of IN and LN are computed independently for each sample, while BN uses batch average after training without computing moving average in each iteration. Here batch average is performed in two steps. First, we freeze the parameters of the network and all the SN layers, and feed the network with a certain number of mini-batches
4
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11871337890625,
0.1748046875,
0.11871337890625
],
[
0.1748046875,
0.11871337890625,
0.82275390625,
0.11871337890625,
0.82275390625,
0.1319580078125,
0.1748046875,
0.1319580078125
],
[
0.1748046875,
0.1331787109375,
0.82275390625,
0.1331787109375,
0.82275390625,
0.146484375,
0.1748046875,
0.146484375
],
[
0.1748046875,
0.146484375,
0.82275390625,
0.146484375,
0.82275390625,
0.160400390625,
0.1748046875,
0.160400390625
],
[
0.1748046875,
0.160400390625,
0.82275390625,
0.160400390625,
0.82275390625,
0.173583984375,
0.1748046875,
0.173583984375
],
[
0.1748046875,
0.1749267578125,
0.82421875,
0.1749267578125,
0.82421875,
0.1881103515625,
0.1748046875,
0.1881103515625
],
[
0.1748046875,
0.1881103515625,
0.82421875,
0.1881103515625,
0.82421875,
0.2020263671875,
0.1748046875,
0.2020263671875
],
[
0.1748046875,
0.2020263671875,
0.826171875,
0.2020263671875,
0.826171875,
0.21533203125,
0.1748046875,
0.21533203125
],
[
0.1748046875,
0.216552734375,
0.2197265625,
0.216552734375,
0.2197265625,
0.2279052734375,
0.1748046875,
0.2279052734375
],
[
0.1732177734375,
0.23486328125,
0.82177734375,
0.236083984375,
0.82177734375,
0.250732421875,
0.1732177734375,
0.25
],
[
0.1732177734375,
0.250732421875,
0.826171875,
0.25,
0.826171875,
0.264404296875,
0.1732177734375,
0.26513671875
],
[
0.1732177734375,
0.266357421875,
0.82421875,
0.266357421875,
0.82421875,
0.2841796875,
0.1732177734375,
0.2841796875
],
[
0.176513671875,
0.2841796875,
0.57421875,
0.2841796875,
0.57421875,
0.29736328125,
0.176513671875,
0.29736328125
],
[
0.1748046875,
0.300537109375,
0.82177734375,
0.300537109375,
0.82177734375,
0.314453125,
0.1748046875,
0.314453125
],
[
0.1748046875,
0.317626953125,
0.8203125,
0.317626953125,
0.8203125,
0.330810546875,
0.1748046875,
0.330810546875
],
[
0.1805419921875,
0.32568359375,
0.82861328125,
0.328857421875,
0.82861328125,
0.3642578125,
0.1805419921875,
0.361083984375
],
[
0.5908203125,
0.335205078125,
0.60888671875,
0.335205078125,
0.60888671875,
0.339111328125,
0.5908203125,
0.339111328125
],
[
0.1683349609375,
0.347900390625,
0.7548828125,
0.35595703125,
0.7548828125,
0.3837890625,
0.1683349609375,
0.375732421875
],
[
0.181396484375,
0.343505859375,
0.22216796875,
0.343505859375,
0.22216796875,
0.35107421875,
0.181396484375,
0.35107421875
],
[
0.7685546875,
0.367431640625,
0.8095703125,
0.367431640625,
0.8095703125,
0.37744140625,
0.7685546875,
0.37744140625
],
[
0.81396484375,
0.3642578125,
0.81787109375,
0.3642578125,
0.81787109375,
0.369384765625,
0.81396484375,
0.369384765625
],
[
0.1748046875,
0.37890625,
0.7060546875,
0.37890625,
0.7060546875,
0.393310546875,
0.1748046875,
0.393310546875
],
[
0.7509765625,
0.373779296875,
0.7607421875,
0.373779296875,
0.7607421875,
0.376953125,
0.7509765625,
0.376953125
],
[
0.1748046875,
0.401611328125,
0.5458984375,
0.401611328125,
0.5458984375,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.63134765625,
0.40771484375,
0.646484375,
0.40771484375,
0.646484375,
0.41845703125,
0.63134765625,
0.41845703125
],
[
0.1748046875,
0.416748046875,
0.5498046875,
0.416748046875,
0.5498046875,
0.43115234375,
0.1748046875,
0.43115234375
],
[
0.1748046875,
0.431884765625,
0.5498046875,
0.431884765625,
0.5498046875,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.1748046875,
0.4462890625,
0.5498046875,
0.4462890625,
0.5498046875,
0.460205078125,
0.1748046875,
0.460205078125
],
[
0.1748046875,
0.460205078125,
0.5498046875,
0.460205078125,
0.5498046875,
0.473388671875,
0.1748046875,
0.473388671875
],
[
0.176513671875,
0.47607421875,
0.54736328125,
0.47607421875,
0.54736328125,
0.486083984375,
0.176513671875,
0.486083984375
],
[
0.1748046875,
0.488037109375,
0.5498046875,
0.488037109375,
0.5498046875,
0.501953125,
0.1748046875,
0.501953125
],
[
0.1748046875,
0.50390625,
0.54833984375,
0.50390625,
0.54833984375,
0.517578125,
0.1748046875,
0.517578125
],
[
0.176513671875,
0.51806640625,
0.5498046875,
0.51806640625,
0.5498046875,
0.5322265625,
0.176513671875,
0.5322265625
],
[
0.1748046875,
0.5322265625,
0.54736328125,
0.5322265625,
0.54736328125,
0.54541015625,
0.1748046875,
0.54541015625
],
[
0.1732177734375,
0.54443359375,
0.54736328125,
0.54443359375,
0.54736328125,
0.55810546875,
0.1732177734375,
0.55810546875
],
[
0.1748046875,
0.5595703125,
0.54833984375,
0.5595703125,
0.54833984375,
0.57373046875,
0.1748046875,
0.57373046875
],
[
0.1732177734375,
0.57470703125,
0.5498046875,
0.57373046875,
0.5498046875,
0.58837890625,
0.1732177734375,
0.5888671875
],
[
0.1748046875,
0.5888671875,
0.5498046875,
0.5888671875,
0.5498046875,
0.60302734375,
0.1748046875,
0.60302734375
],
[
0.56396484375,
0.583984375,
0.826171875,
0.583984375,
0.826171875,
0.59765625,
0.56396484375,
0.59765625
],
[
0.1748046875,
0.60302734375,
0.55126953125,
0.60302734375,
0.55126953125,
0.6162109375,
0.1748046875,
0.6162109375
],
[
0.56201171875,
0.59765625,
0.82421875,
0.59765625,
0.82421875,
0.611328125,
0.56201171875,
0.611328125
],
[
0.56396484375,
0.60986328125,
0.72607421875,
0.60986328125,
0.72607421875,
0.62353515625,
0.56396484375,
0.62353515625
],
[
0.1748046875,
0.61865234375,
0.5498046875,
0.61865234375,
0.5498046875,
0.6318359375,
0.1748046875,
0.6318359375
],
[
0.176513671875,
0.6337890625,
0.54833984375,
0.6337890625,
0.54833984375,
0.64453125,
0.176513671875,
0.64453125
],
[
0.1732177734375,
0.64599609375,
0.82421875,
0.64453125,
0.82421875,
0.6591796875,
0.1732177734375,
0.66015625
],
[
0.1748046875,
0.66015625,
0.8203125,
0.66015625,
0.8203125,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1748046875,
0.6748046875,
0.82275390625,
0.6748046875,
0.82275390625,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1732177734375,
0.68798828125,
0.82421875,
0.68798828125,
0.82421875,
0.7021484375,
0.1732177734375,
0.7021484375
],
[
0.1732177734375,
0.7021484375,
0.259033203125,
0.69970703125,
0.259033203125,
0.7138671875,
0.1732177734375,
0.71630859375
],
[
0.1748046875,
0.72265625,
0.82177734375,
0.72265625,
0.82177734375,
0.736328125,
0.1748046875,
0.736328125
],
[
0.1748046875,
0.736328125,
0.82275390625,
0.736328125,
0.82275390625,
0.75,
0.1748046875,
0.75
],
[
0.17236328125,
0.75,
0.82275390625,
0.75048828125,
0.82275390625,
0.76513671875,
0.17236328125,
0.7646484375
],
[
0.1732177734375,
0.76513671875,
0.82275390625,
0.7646484375,
0.82275390625,
0.77783203125,
0.1732177734375,
0.77880859375
],
[
0.1748046875,
0.77880859375,
0.82421875,
0.77880859375,
0.82421875,
0.79248046875,
0.1748046875,
0.79248046875
],
[
0.1748046875,
0.79150390625,
0.82275390625,
0.79150390625,
0.82275390625,
0.8046875,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.80615234375,
0.826171875,
0.80615234375,
0.826171875,
0.8193359375,
0.1748046875,
0.8193359375
],
[
0.176513671875,
0.8212890625,
0.82421875,
0.8212890625,
0.82421875,
0.83203125,
0.176513671875,
0.83203125
],
[
0.1748046875,
0.833984375,
0.82421875,
0.833984375,
0.82421875,
0.84765625,
0.1748046875,
0.84765625
],
[
0.1732177734375,
0.84765625,
0.32275390625,
0.8486328125,
0.32275390625,
0.85986328125,
0.1732177734375,
0.859375
],
[
0.1748046875,
0.86865234375,
0.82177734375,
0.86865234375,
0.82177734375,
0.8818359375,
0.1748046875,
0.8818359375
],
[
0.1748046875,
0.88330078125,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.1748046875,
0.896484375
],
[
0.1732177734375,
0.89794921875,
0.82421875,
0.89794921875,
0.82421875,
0.9111328125,
0.1732177734375,
0.9111328125
],
[
0.1732177734375,
0.9111328125,
0.826171875,
0.9111328125,
0.826171875,
0.9248046875,
0.1732177734375,
0.9248046875
],
[
0.494384765625,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.494384765625,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2019",
"jointly by back-propagation (BP). This training procedure is different from previous meta-learning",
"algorithms such as network architecture search (Colson et al., 2007; Liu et al., 2018; Pham et al.,",
"2018). In previous work, represents as a set of network modules with different learning capacities,",
"where and were optimized in two BP stages iteratively by using two training sets that are non-",
"overlapped. For example, previous work divided an entire training set into a training and a validation",
"set. However, if and Φ in previous work are optimized in the same set of training data, would",
"choose the module with large complexity to overfit these data. In contrast, SN essentially prevents",
"overfitting by choosing normalizers to improve both learning and generalization ability as discussed",
"below.",
"Analyses of SN. To understand SN, we theoretically compare SN with BN, IN, and LN by repre-",
"senting them using weight normalization (WN) (Salimans & Kingma, 2016) that is independent of",
"mean and variance. wN is computed as w, where w and x represent a filter and an image patch.",
"WN normalizes the norm of each filter to I and rescales to .",
"Remark 1. Let x be an image patch with zero mean and unit variance, w; be a filter of the i-",
"th channel, i E {1, 2, .., C), and v be the filter norm of WN. Eqn.(1) can be rewritten by hin =",
"g",
"ar1.",
"SNin Eqn.(3) can berefomulated by hsn = Winhin + wbnhbn + Wmhn =wla+uw)",
"1/wi2",
"l/will2",
"T",
"B, s.t. Wbn≤ v, where Win, Win, Wbn are the weights and ≤ is a constraint.",
"-",
"Remark 1 simplifies SN in Eqn.(3), enabling us to com-",
"Wi",
"pare different normalizers geometrically by formulating",
"them with respect to WN. In Fig.3, hin of IN can be",
"computed similarly to WN with an additional bias β,",
"where the norms of all filters are normalized to I and then",
"rescaledto:Asandwhavethesamelearningdynamic.",
"the length of would be identically to (see wi, w2 of",
"IN). Moreover, hbn in BN can be rewritten as WN with",
"regularization over , making it shorter than v. Compared",
"to IN and LN, Luo et al. (2019) shows that the regulariza-",
"tion of BN improves generalization and increases angle",
"between filters, preventing them from coadaptation (see",
"W1, W2 of BN). Furthermore, hin in LN normalizes each",
"filter among channels where the filter norm is less con-",
"Figure 3: Geometric view of directions and",
"strained than IN and BN. That is, LN allows > to",
"lengths of the filters in IN, BN, LN, and SN",
"by comparing them to WN.",
"increase learning ability. Finally, hsn in SN inherits the",
"benefitsfrom allofthemandenablesbalancebetween",
"learning and generalization ability. For example, when the batch size is small, the random noise",
"from the batch statistics of BN would be too strong. SN is able to maintain performance by decreas.",
"ing wbn and increasing Win, such that the regularization from BN is reduced and the learning ability",
"is enhanced by LN. This phenomenon is supported by our experiment. More results are provided in",
"AppendixB.",
"Variants of SN. SN has many extensions. For instance, a pretrained network with SN can be fine-",
"tuned by applying the argmax function on its control parameters where each normalization layer",
"selects only one normalizer, leading to sparse SN. For (8, 32) as an example, SN with sparsity",
"achieves top-1 accuracyof 77.0%in ImageNetwithResNet50,which is comparableto76.9%of SN",
"without sparsity. Moreover, when the channels are divided into groups, each group could select its",
"own normalizer to increase representation power of SN. Our preliminary results suggest that group",
"SN performs better than SN in some senses. For instance, group SN with only two groups boosts",
"thetop-1accuracyofResNet50to77.2%inImageNet.Theabovetwovariantswillbepresentedas",
"future work due to the length of paper. This work focuses on SN where the importance weights are",
"tied between channels",
"Inference. When applying SN in test, the statistics of IN and LN are computed independently",
"for each sample, while BN uses batch average after training without computing moving average",
"in each iteration. Here batch average is performed in two steps. First, we freeze the parameters",
"of the network and all the SN layers, and feed the network with a certain number of mini-batches",
"4"
] |
{
"arxiv_id": "1806.10779",
"language": "en",
"page_idx": 3
}
|
||
1806.10779-en-4
|
1806.10779-en-4.jpg
|
Published as a conference paper at ICLR 2019
<table>
<tbody>
<tr>
<th> </th>
<th colspan="3">Parameter</th>
<th colspan="3">Statistical Estimation</th>
</tr>
<tr>
<th> </th>
<th>params</th>
<th>$\#$params</th>
<th>$hyper-$ params</th>
<th>statistics</th>
<th>computation complexity</th>
<th>#statistics</th>
</tr>
<tr>
<td>BN (Ioffe&Szegedy.2015)</td>
<td>$\gamma,\beta$</td>
<td>$2C$</td>
<td>$p,\epsilon$</td>
<td>$\mu$,</td>
<td>$\mathcal{O}(NCHW$</td>
<td>$2C$</td>
</tr>
<tr>
<td>IN(Ulyanov et al., 2016)</td>
<td>$\beta$</td>
<td>$2C$</td>
<td>$\epsilon$</td>
<td>$\mu,\sigma$</td>
<td>$\mathcal{O}(NCHW^{\prime}$</td>
<td>$2CN$</td>
</tr>
<tr>
<td>$LN$ (Ba et al..2016)</td>
<td>$\beta$</td>
<td>$2C$</td>
<td>$\epsilon$</td>
<td>$\mu,\sigma$</td>
<td>$\mathcal{O}(NCHW^{\prime}$</td>
<td>$2N$</td>
</tr>
<tr>
<td>GN (Wu&He,2018)</td>
<td>$\beta$</td>
<td>$2C$</td>
<td>10 $e$</td>
<td>$\mu,\sigma$</td>
<td>$\mathcal{O}(NCHW)$</td>
<td>$2qN$</td>
</tr>
<tr>
<td>BRN(Ioffe,2017)</td>
<td>$\gamma,\beta$</td>
<td>$\overline{2C}$</td>
<td>$\overline{d}$ $\mathcal{P}$, 十 $,T$</td>
<td>$\angle\angle AB=\angle AB$</td>
<td>$\mathcal{O}(NCHW)$</td>
<td>$\overline{2C}$</td>
</tr>
<tr>
<td>$BKN$ 1 (Wang et al., 2018)</td>
<td>$A$</td>
<td>$C^{2}$</td>
<td>$p,\epsilon$</td>
<td>$\Sigma$ $\Sigma^{\prime}$ $\mu.$</td>
<td>$\mathcal{O}(NC^{2}$ $HW)$</td>
<td>$C+C^{2}$</td>
</tr>
<tr>
<td>WN (Salimans & Kingma, 2016)</td>
<td>$\gamma$</td>
<td>$C$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>SN</td>
<td>$\gamma,\beta$, $\{w_{k}\}_{k\in\Omega}$</td>
<td>$2C+6$</td>
<td>$\epsilon$</td>
<td>$\{\mu_{k},\sigma_{k}\}_{k\in\Omega}$</td>
<td>$\mathcal{O}(NCHW)$</td>
<td>$2C+2N$ $+2CN$</td>
</tr>
</tbody>
</table>
Table 1: Comparisons of normalization methods. First, we compare their types of parameters, numbers of parameters $(\#$params), and hyper-parameters. Second, we compare types of statistics, computational complexity to estimate statistics, and numbers of statistics (#statistics). Specifically, $\gamma,\beta$ denote the scale and shift parameters. $\mu,\sigma,\Sigma$ are a vector of means, a vector of standard deviations, and a covariance matrix. $\mu^{\prime}$ represents the moving average. Moreover, $p$ is the momentum of moving average, $g$ in GN is the number of groups, $\epsilon$ is a small value for numerical stability, and $r, d$ are used in BRN. In SN, $k\in \Omega$ indicates a set of groups, $\epsilon$ is a small value for numerical stability, and $r, d$ are used in BRN. In SN, $k\in \Omega$indicates a set of different kinds of statistics, $\Omega$= $\{$in, ln, bn$\}$, and $w_{k}$ is an importance weight of each kind.
randomly chosen from the training set. Second, we average the means and variances produced by
all these mini-batches in each SN layer. The averaged statistics are used by BN in SN.
We find that batch average makes training converged faster than moving average. It can be computed by using a small amount of samples. For example, top-1 accuracies of ResNet50 on ImageNet by using batch average with 50k and all training samples are 76.90% and 76.92% respectively. They are trained much faster and slightly better than 76.89% of moving average. Appendix A shows more results.
Implementation. SN can be easily implemented in existing softwares such as PyTorch and TensorFlow. The backward computation of SN can be obtained by automatic differentiation (AD) in these softwares. Without AD, we need to implement back-propagation (BP) of SN, where the errors are propagated through $\mu_k$ and $\sigma_k^2.$ We provide the derivations of BP in Appendix H.
3 RELATIONSHIPS TO PREVIOUS WORK
In Table 1, we compare SN to BN, IN, LN, and GN, as well as three variants of BN including Batch Renormalization (BRN), Batch Kalman Normalization (BKN), and WN. In general, we see that SN possesses comparable numbers of parameters and computations, as well as rich statistics. Details ${\mathrm{are~presented~below.}}$
$\bullet$ First, SN has similar number of parameters compared to previous methods, as shown in the first portion of Table $\color{red}1.$ Most of the approaches learn a scale parameter $\gamma$ and a bias $\beta$ for each one of the $C$ channels, resulting in $2C$ parameters. SN learns 6 importance weights as the additional parameters. We see that BKN has the maximum number of $C^{2}$ parameters, as it learns a transformation matrix $A$ for the means and variances. WN has $C$ scale parameters without the biases.
Furthermore, many methods have $p$ and $\epsilon$ as hyper-parameters, whose values are not sensitive because they are often fixed in different tasks. In contrast, GN and BRN have to search the number of groups $g$ or the renormalization parameters $r,d$, which may have different values in different networks. Moreover, WN does not have hyper-parameters and statistics, since it performs normalization in the space of network parameters rather than feature space. Salimans & Kingma (2016); Luo et al. (201$\hat{9})$ showed that WN is a special case of BN.
$\bullet$ Second, although SN has richer statistics, the computational complexity to estimate them is comparable to previous methods, as shown in the second portion of Table l. As introduced in Sec.2, IN, $\hat{\mathrm{LN,~and~BN~estimate~the~means~and~variances~along~axes~}(H,W),(C,H,W),\mathrm{and~}(N,H,W)~\mathrm{re-}}$ spectively, leading to $2CN,2N$, and $2C$ numbers of statistics. Therefore, SN has $2CN+2N+2C$ statistics by combining them. Although BKN has the largest number of $C+C^2$ statistics, it also has the highest computations because it estimates the covariance matrix other than the variance vector. Also, approximating the covariance matrix in a minibatch is nontrivial as discussed in (Desjardins et al., 2015; Luo, 2017b;a). BN, BRN, and BKN also compute moving averages.
5
|
[
[
0.1732177734375,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.4248046875,
0.1029052734375,
0.473876953125,
0.1029052734375,
0.473876953125,
0.1142578125,
0.4248046875,
0.1142578125
],
[
0.62158203125,
0.1029052734375,
0.72216796875,
0.1029052734375,
0.72216796875,
0.11297607421875,
0.62158203125,
0.11297607421875
],
[
0.494384765625,
0.11297607421875,
0.52685546875,
0.11297607421875,
0.52685546875,
0.1256103515625,
0.494384765625,
0.1256103515625
],
[
0.564453125,
0.11553955078125,
0.61279296875,
0.11553955078125,
0.61279296875,
0.1300048828125,
0.564453125,
0.1300048828125
],
[
0.64892578125,
0.11297607421875,
0.71142578125,
0.11297607421875,
0.71142578125,
0.125,
0.64892578125,
0.125
],
[
0.426513671875,
0.11614990234375,
0.4755859375,
0.11614990234375,
0.4755859375,
0.1307373046875,
0.426513671875,
0.1307373046875
],
[
0.369384765625,
0.117431640625,
0.406982421875,
0.117431640625,
0.406982421875,
0.1307373046875,
0.369384765625,
0.1307373046875
],
[
0.73779296875,
0.117431640625,
0.7900390625,
0.117431640625,
0.7900390625,
0.1287841796875,
0.73779296875,
0.1287841796875
],
[
0.494384765625,
0.12371826171875,
0.53125,
0.12371826171875,
0.53125,
0.1337890625,
0.494384765625,
0.1337890625
],
[
0.65283203125,
0.12371826171875,
0.7060546875,
0.12371826171875,
0.7060546875,
0.1337890625,
0.65283203125,
0.1337890625
],
[
0.494384765625,
0.1319580078125,
0.52685546875,
0.1319580078125,
0.52685546875,
0.17041015625,
0.494384765625,
0.17041015625
],
[
0.64453125,
0.1319580078125,
0.71142578125,
0.1319580078125,
0.71142578125,
0.1427001953125,
0.64453125,
0.1427001953125
],
[
0.439453125,
0.1331787109375,
0.4599609375,
0.1331787109375,
0.4599609375,
0.1458740234375,
0.439453125,
0.1458740234375
],
[
0.191162109375,
0.1337890625,
0.324462890625,
0.1337890625,
0.324462890625,
0.14453125,
0.191162109375,
0.14453125
],
[
0.373291015625,
0.1337890625,
0.400390625,
0.1337890625,
0.400390625,
0.146484375,
0.373291015625,
0.146484375
],
[
0.55615234375,
0.1351318359375,
0.61376953125,
0.1351318359375,
0.61376953125,
0.1458740234375,
0.55615234375,
0.1458740234375
],
[
0.75341796875,
0.1351318359375,
0.7685546875,
0.1351318359375,
0.7685546875,
0.1427001953125,
0.75341796875,
0.1427001953125
],
[
0.19287109375,
0.1427001953125,
0.31201171875,
0.143310546875,
0.31201171875,
0.1546630859375,
0.19287109375,
0.154052734375
],
[
0.438720703125,
0.1427001953125,
0.461669921875,
0.1427001953125,
0.461669921875,
0.1558837890625,
0.438720703125,
0.1558837890625
],
[
0.74658203125,
0.1427001953125,
0.77783203125,
0.1427001953125,
0.77783203125,
0.1546630859375,
0.74658203125,
0.1546630859375
],
[
0.64453125,
0.143310546875,
0.71337890625,
0.143310546875,
0.71337890625,
0.154052734375,
0.64453125,
0.154052734375
],
[
0.373291015625,
0.143310546875,
0.40185546875,
0.143310546875,
0.40185546875,
0.1558837890625,
0.373291015625,
0.1558837890625
],
[
0.435546875,
0.149658203125,
0.464111328125,
0.149658203125,
0.464111328125,
0.177978515625,
0.435546875,
0.177978515625
],
[
0.57666015625,
0.1458740234375,
0.60302734375,
0.1458740234375,
0.60302734375,
0.154052734375,
0.57666015625,
0.154052734375
],
[
0.371826171875,
0.15283203125,
0.40283203125,
0.15283203125,
0.40283203125,
0.1685791015625,
0.371826171875,
0.1685791015625
],
[
0.64453125,
0.15283203125,
0.7138671875,
0.15283203125,
0.7138671875,
0.1641845703125,
0.64453125,
0.1641845703125
],
[
0.74853515625,
0.15283203125,
0.77685546875,
0.15283203125,
0.77685546875,
0.1641845703125,
0.74853515625,
0.1641845703125
],
[
0.576171875,
0.1546630859375,
0.60302734375,
0.1546630859375,
0.60302734375,
0.1673583984375,
0.576171875,
0.1673583984375
],
[
0.193603515625,
0.154052734375,
0.289306640625,
0.154052734375,
0.289306640625,
0.1641845703125,
0.193603515625,
0.1641845703125
],
[
0.191162109375,
0.1622314453125,
0.295654296875,
0.1634521484375,
0.295654296875,
0.1749267578125,
0.191162109375,
0.173583984375
],
[
0.374267578125,
0.1634521484375,
0.40185546875,
0.1634521484375,
0.40185546875,
0.175537109375,
0.374267578125,
0.175537109375
],
[
0.5,
0.1634521484375,
0.52294921875,
0.1634521484375,
0.52294921875,
0.1767578125,
0.5,
0.1767578125
],
[
0.64453125,
0.1634521484375,
0.71142578125,
0.1634521484375,
0.71142578125,
0.1749267578125,
0.64453125,
0.1749267578125
],
[
0.74853515625,
0.1634521484375,
0.77685546875,
0.1634521484375,
0.77685546875,
0.175537109375,
0.74853515625,
0.175537109375
],
[
0.435546875,
0.1717529296875,
0.461669921875,
0.1717529296875,
0.461669921875,
0.2108154296875,
0.435546875,
0.2108154296875
],
[
0.576171875,
0.1654052734375,
0.60107421875,
0.1654052734375,
0.60107421875,
0.1767578125,
0.576171875,
0.1767578125
],
[
0.191162109375,
0.1749267578125,
0.279296875,
0.1749267578125,
0.279296875,
0.1849365234375,
0.191162109375,
0.1849365234375
],
[
0.374267578125,
0.175537109375,
0.400390625,
0.175537109375,
0.400390625,
0.1881103515625,
0.374267578125,
0.1881103515625
],
[
0.48779296875,
0.175537109375,
0.53759765625,
0.175537109375,
0.53759765625,
0.1875,
0.48779296875,
0.1875
],
[
0.5556640625,
0.175537109375,
0.61767578125,
0.173583984375,
0.61767578125,
0.1849365234375,
0.5556640625,
0.1875
],
[
0.64453125,
0.1749267578125,
0.71337890625,
0.1749267578125,
0.71337890625,
0.1849365234375,
0.64453125,
0.1849365234375
],
[
0.75244140625,
0.175537109375,
0.771484375,
0.175537109375,
0.771484375,
0.184326171875,
0.75244140625,
0.184326171875
],
[
0.19287109375,
0.184326171875,
0.313720703125,
0.184326171875,
0.313720703125,
0.1976318359375,
0.19287109375,
0.1976318359375
],
[
0.6396484375,
0.184326171875,
0.7197265625,
0.1817626953125,
0.7197265625,
0.1962890625,
0.6396484375,
0.1988525390625
],
[
0.736328125,
0.184326171875,
0.78515625,
0.1817626953125,
0.7861328125,
0.1962890625,
0.73779296875,
0.1988525390625
],
[
0.55224609375,
0.186279296875,
0.619140625,
0.1817626953125,
0.62109375,
0.1976318359375,
0.5537109375,
0.2020263671875
],
[
0.37744140625,
0.186279296875,
0.396240234375,
0.186279296875,
0.396240234375,
0.2008056640625,
0.37744140625,
0.2008056640625
],
[
0.498291015625,
0.1875,
0.52392578125,
0.1875,
0.52392578125,
0.2008056640625,
0.498291015625,
0.2008056640625
],
[
0.1953125,
0.1976318359375,
0.294189453125,
0.1976318359375,
0.294189453125,
0.2052001953125,
0.1953125,
0.2052001953125
],
[
0.301513671875,
0.1956787109375,
0.34716796875,
0.1956787109375,
0.34716796875,
0.205810546875,
0.301513671875,
0.205810546875
],
[
0.384033203125,
0.1988525390625,
0.3896484375,
0.1988525390625,
0.3896484375,
0.203857421875,
0.384033203125,
0.203857421875
],
[
0.381591796875,
0.208984375,
0.40283203125,
0.20703125,
0.404296875,
0.21533203125,
0.38232421875,
0.2171630859375
],
[
0.73779296875,
0.2052001953125,
0.7900390625,
0.2052001953125,
0.7900390625,
0.216552734375,
0.73779296875,
0.216552734375
],
[
0.426513671875,
0.210205078125,
0.47216796875,
0.210205078125,
0.47216796875,
0.2247314453125,
0.426513671875,
0.2247314453125
],
[
0.19287109375,
0.210205078125,
0.2132568359375,
0.210205078125,
0.2132568359375,
0.2235107421875,
0.19287109375,
0.2235107421875
],
[
0.55224609375,
0.212158203125,
0.6259765625,
0.212158203125,
0.6259765625,
0.2235107421875,
0.55224609375,
0.2235107421875
],
[
0.64453125,
0.2108154296875,
0.7138671875,
0.2108154296875,
0.7138671875,
0.222900390625,
0.64453125,
0.222900390625
],
[
0.744140625,
0.216552734375,
0.78369140625,
0.216552734375,
0.78369140625,
0.2279052734375,
0.744140625,
0.2279052734375
],
[
0.5048828125,
0.2139892578125,
0.515625,
0.2139892578125,
0.515625,
0.2247314453125,
0.5048828125,
0.2247314453125
],
[
0.365234375,
0.2171630859375,
0.4150390625,
0.2171630859375,
0.4150390625,
0.2279052734375,
0.365234375,
0.2279052734375
],
[
0.1732177734375,
0.238037109375,
0.82421875,
0.2392578125,
0.82421875,
0.25390625,
0.1732177734375,
0.25244140625
],
[
0.1732177734375,
0.25244140625,
0.82275390625,
0.25244140625,
0.82275390625,
0.266357421875,
0.1732177734375,
0.266357421875
],
[
0.1748046875,
0.267578125,
0.82421875,
0.267578125,
0.82421875,
0.27783203125,
0.1748046875,
0.27783203125
],
[
0.176513671875,
0.27978515625,
0.8203125,
0.27978515625,
0.8203125,
0.290283203125,
0.176513671875,
0.290283203125
],
[
0.17236328125,
0.291748046875,
0.826171875,
0.291748046875,
0.826171875,
0.304931640625,
0.17236328125,
0.304931640625
],
[
0.17236328125,
0.302978515625,
0.826171875,
0.3017578125,
0.826171875,
0.31640625,
0.17236328125,
0.317626953125
],
[
0.1732177734375,
0.31640625,
0.7158203125,
0.31640625,
0.7158203125,
0.330078125,
0.1732177734375,
0.330078125
],
[
0.1732177734375,
0.3515625,
0.82275390625,
0.3515625,
0.82275390625,
0.365478515625,
0.1732177734375,
0.365478515625
],
[
0.1732177734375,
0.3642578125,
0.734375,
0.3642578125,
0.734375,
0.37744140625,
0.1732177734375,
0.37744140625
],
[
0.1748046875,
0.386474609375,
0.82421875,
0.386474609375,
0.82421875,
0.399658203125,
0.1748046875,
0.399658203125
],
[
0.1732177734375,
0.40087890625,
0.82177734375,
0.40087890625,
0.82177734375,
0.4140625,
0.1732177734375,
0.4140625
],
[
0.1748046875,
0.4140625,
0.82275390625,
0.4140625,
0.82275390625,
0.427978515625,
0.1748046875,
0.427978515625
],
[
0.1732177734375,
0.4267578125,
0.82421875,
0.427978515625,
0.82421875,
0.442626953125,
0.1732177734375,
0.44140625
],
[
0.1748046875,
0.443115234375,
0.22216796875,
0.443115234375,
0.22216796875,
0.455078125,
0.1748046875,
0.455078125
],
[
0.1732177734375,
0.4609375,
0.82275390625,
0.462158203125,
0.82275390625,
0.4765625,
0.1732177734375,
0.47607421875
],
[
0.176513671875,
0.479248046875,
0.82275390625,
0.479248046875,
0.82275390625,
0.4892578125,
0.176513671875,
0.4892578125
],
[
0.1732177734375,
0.4912109375,
0.82421875,
0.4912109375,
0.82421875,
0.5048828125,
0.1732177734375,
0.5048828125
],
[
0.1748046875,
0.5048828125,
0.705078125,
0.5048828125,
0.705078125,
0.51953125,
0.1748046875,
0.51953125
],
[
0.1748046875,
0.548828125,
0.2034912109375,
0.548828125,
0.2034912109375,
0.56005859375,
0.1748046875,
0.56005859375
],
[
0.2017822265625,
0.548828125,
0.51953125,
0.548828125,
0.51953125,
0.5625,
0.2017822265625,
0.5625
],
[
0.17236328125,
0.58349609375,
0.82421875,
0.58203125,
0.82421875,
0.5966796875,
0.17236328125,
0.59765625
],
[
0.1732177734375,
0.59765625,
0.82421875,
0.59765625,
0.82421875,
0.611328125,
0.1732177734375,
0.611328125
],
[
0.1732177734375,
0.6123046875,
0.82421875,
0.6123046875,
0.82421875,
0.62548828125,
0.1732177734375,
0.62548828125
],
[
0.1748046875,
0.625,
0.313720703125,
0.625,
0.313720703125,
0.63818359375,
0.1748046875,
0.63818359375
],
[
0.1732177734375,
0.64453125,
0.82421875,
0.64599609375,
0.82421875,
0.66015625,
0.1732177734375,
0.6591796875
],
[
0.1748046875,
0.66015625,
0.82421875,
0.66015625,
0.82421875,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1748046875,
0.673828125,
0.82177734375,
0.673828125,
0.82177734375,
0.6875,
0.1748046875,
0.6875
],
[
0.1732177734375,
0.6875,
0.82421875,
0.6875,
0.82421875,
0.7021484375,
0.1732177734375,
0.7021484375
],
[
0.1748046875,
0.7021484375,
0.73779296875,
0.7021484375,
0.73779296875,
0.71533203125,
0.1748046875,
0.71533203125
],
[
0.1748046875,
0.7236328125,
0.82177734375,
0.7236328125,
0.82177734375,
0.7373046875,
0.1748046875,
0.7373046875
],
[
0.1748046875,
0.7373046875,
0.82421875,
0.7373046875,
0.82421875,
0.75048828125,
0.1748046875,
0.75048828125
],
[
0.1748046875,
0.751953125,
0.82421875,
0.751953125,
0.82421875,
0.76513671875,
0.1748046875,
0.76513671875
],
[
0.1748046875,
0.76513671875,
0.82177734375,
0.76513671875,
0.82177734375,
0.77880859375,
0.1748046875,
0.77880859375
],
[
0.17236328125,
0.77783203125,
0.82421875,
0.77880859375,
0.82421875,
0.79345703125,
0.17236328125,
0.79248046875
],
[
0.1732177734375,
0.79248046875,
0.5556640625,
0.79248046875,
0.5556640625,
0.80615234375,
0.1732177734375,
0.80615234375
],
[
0.1732177734375,
0.81298828125,
0.82177734375,
0.81298828125,
0.82177734375,
0.8271484375,
0.1732177734375,
0.8271484375
],
[
0.1748046875,
0.82763671875,
0.82275390625,
0.82763671875,
0.82275390625,
0.84130859375,
0.1748046875,
0.84130859375
],
[
0.1748046875,
0.84130859375,
0.82275390625,
0.84130859375,
0.82275390625,
0.85498046875,
0.1748046875,
0.85498046875
],
[
0.1748046875,
0.85498046875,
0.82275390625,
0.85498046875,
0.82275390625,
0.869140625,
0.1748046875,
0.869140625
],
[
0.17236328125,
0.86865234375,
0.82421875,
0.8671875,
0.82421875,
0.8818359375,
0.17236328125,
0.88330078125
],
[
0.1748046875,
0.88330078125,
0.82275390625,
0.88330078125,
0.82275390625,
0.896484375,
0.1748046875,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.826171875,
0.89794921875,
0.826171875,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.70166015625,
0.9111328125,
0.70166015625,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2019",
"Parameter",
"Statistical Estimation",
"hyper-",
"statistics",
"computation",
"# params",
"params",
"#statistics",
"params",
"complexity",
"",
"ONCHW",
"20",
"BN(loffe&Szegedy.2015)",
"Y.B",
"μ,a,μ,o",
"20",
"IN (Ulyanov etal.,2016)",
"20",
"2CN",
"O(NCHW)",
"Y.B",
"28",
"u,a",
"Y.B",
"O(NCHW)",
"2N",
"μ,a",
"LN (Ba et al., 2016)",
"GN (Wu&He,2018)",
"Y.B",
"g,e",
"O(NCHW)",
"2gN",
"280",
"P,a",
"BRN (loffe2017)",
"B",
"p,e,r,d",
"H,o.H,o",
"O(NCHW)",
"2C",
"BKN (Wang et al., 2018)",
"O(NC?HW)",
"C + C2",
",E.p,",
"A",
"P,E",
"WN(Salimans&Kin",
"ma,2016)",
"A",
"YB",
"2C+2N",
"2C+6",
"SN",
"(μk,ok)ken",
"O(NCHW)",
"+2CN",
"E",
"(walkes",
"Table l: Comparisons of normalization methods. First, we compare their types of parameters, numbers",
"of parameters (#params), and hyper-parameters. Second, we compare types of statistics, computational com-",
"plexity to estimate statistics,and numbers of statistics (#statistics).Specifically,,β denote the scale and",
"shift parameters.μ,o, are a vector of means,a vector of standard deviations, and a covariance matrix. μ",
"represents the moving average. Moreover, p is the momentum of moving average, g in GN is the number of",
"groups, e is a small value for numerical stability, and r, d are used in BRN. In SN, k e 2 indicates a set of",
"different kinds of statistics, 2 = {in, In, bn], and w, is an importance weight of each kind.",
"randomly chosen from the training set. Second, we average the means and variances produced by",
"all these mini-batches in each SN layer. The averaged statistics are used by BN in SN",
"We find that batch average makes training converged faster than moving average. It can be computed",
"by using a small amount of samples. For example, top-l accuracies of ResNet50 on ImageNet by",
"using batch average with 50k and all training samples are 76.90% and 76.92% respectively. They",
"are trained much faster and slightly better than 76.89% of moving average. Appendix A shows more",
"results.",
"Implementation. SN can be easily implemented in existing softwares such as PyTorch and Tensor-",
"Flow.ThebackwardcomputationofSNcanbeobtainedbyautomaticdifferentiation(AD)inthese",
"softwares. Without AD, we need to implement back-propagation (BP) of SN, where the errors are",
"propagated through μk and o2. We provide the derivations of BP in Appendix H.",
"3",
"RELATIONSHIPS TO PREVIOUS WORK",
"In Table I, we compare SN to BN, IN, LN, and GN, as well as three variants of BN including Batch",
"Renormalization (BRN), Batch Kalman Normalization (BKN), and WN. In general, we see that SN",
"possesses comparable numbers of parameters and computations, as well as rich statistics. Details",
"are presented below.",
". First, SN has similar number of parameters compared to previous methods, as shown in the first",
"portion of Table 1. Most of the approaches learn a scale parameter and a bias β for each one of the",
"C channels, resulting in 2C parameters. SN learns 6 importance weights as the additional param-",
"eters. We see that BKN has the maximum number of C? parameters, as it learns a transformation",
"matrix A for the means and variances. WN has C scale parameters without the biases.",
"Furthermore, many methods have p and e as hyper-parameters, whose values are not sensitive be-",
"cause they are often fixed in different tasks. In contrast, GN and BRN have to search the number",
"of groups g or the renormalization parameters r, d, which may have different values in different",
"networks. Moreover, WN does not have hyper-parameters and statistics, since it performs normal-",
"ization in the space of network parameters rather than feature space. Salimans & Kingma (2016);",
"Luo et al. (2019) showed that WN is a special case of BN.",
". Second, although SN has richer statistics, the computational complexity to estimate them is com-",
"parable to previous methods, as shown in the second portion of Table 1. As introduced in Sec.2, IN,",
"LN, and BN estimate the means and variances along axes (H, W), (C, H, W), and (N, H, W) re-",
"spectively, leading to 2CN, 2N, and 2C numbers of statistics. Therefore, SN has 2CN +2N +2C",
"statistics by combining them. Although BKN has the largest number of C + C2 statistics, it also has",
"the highest computations because it estimates the covariance matrix other than the variance vector.",
"Also, approximating the covariance matrix in a minibatch is nontrivial as discussed in (Desjardins",
"et al., 2015; Luo, 2017b;a). BN, BRN, and BKN also compute moving averages.",
"5"
] |
{
"arxiv_id": "1806.10779",
"language": "en",
"page_idx": 4
}
|
||
1810.06284-en-0
|
1810.06284-en-0.jpg
|
# $\mathbf{CURIOUS:Intrinsically~Motivated~Modular}$ Multi-Goal Reinforcement Learning
Cédric Colas$^{1}$ Pierre Fournier$^{2}$ Olivier Sigaud$^{2}$ Mohamed Chetouani$^{2}$ Pierre-Yves Oudeyer$^{1}$
$\mathbf{Abstract}$
<FigureHere>
Figure 1. The Modular Goal Fetch Arm environment. An intrinsically motivated agent can set its own (modular) goals (Reach, Push, Pick and Place, Stack), with multiple objects and distractors.
In open-ended environments, autonomous learning agents must set their own goals and build their own curriculum through an intrinsically motivated exploration. They may consider a large diversity of goals, aiming to discover what is controllable in their environments, and what is not. Because some goals might prove easy and some impossible, agents must actively select which goal to practice at any moment, to maximize their overall mastery on the set of learnable goals. This paper proposes CURIOUS, an algorithm that leverages 1) a modular Universal Value Function Approximator with hindsight learning to achieve a diversity of goals of different kinds within a unique policy and 2) an automated curriculum learning mechanism that biases the attention of the agent towards goals maximizing the absolute learning progress. Agents focus sequentially on goals of increasing complexity, and focus back on goals that are being forgotten. Experiments conducted in a new modular-goal robotic environment show the resulting developmental self-organization of a learning curriculum, and demonstrate properties of robustness to distracting goals, forgetting and changes in body properties.
and self-organize one's own curriculum. This challenge can be tackled within the framework of Intrinsically Motivated Goal Exploration Processes (IMGEP) (Baranes & Oudeyer, $2013;\bar{\text{Forestier et al., 2017), }}$leveraging computational models of autonomous development in human infants.
Modular goal representation. In a same environment, an agent might want to‘put the cube in position $x’$ or to 'reach position $y’$for any $x$ or $y.$ $\hat{\text{Here, describ- }}$ ing the full goal space requires modular goal representations. Goals are organized by modules, where module refers to the pair of a reward function and a goal space $M_{i}=(R_{M_{i},g_{i}\in\mathcal{G}_{\mathcal{M}_{i}}},\mathcal{G}_{\mathcal{M}_{i}}).$The reward function describes a set of constraints that must be satisfied by the agent's state (e.g. Reach), given a continuous parameter (e.g. $g_i=y)$ evolving in the associated goal space (e.g. 3D Euclidean space), see Fig. 1.
### $1. \textbf{Introduction}$
In autonomous continual learning, agents aim to discover repertoires of skills in an ever-changing open-ended world, and without external rewards. In such realistic environments, the agent must be endowed with intrinsic motivations to explore the diversity of ways in which it can control its environment. One important form of intrinsic motivation system is the ability to autonomously set one's own goals
While flat multi-goal problems with continuous (Schaul et al., 2015a; Andrychowicz et al., 2017; Plappert et al., 2018) or discrete goals (Mankowitz et al., 2018; Riedmiller et al., 2018) have been explored in the past, only few works tackle the problem of modular multi-goal learning (Forestier & Q Oudeyer, 2016), none in an RL setting. Here, we present cuRIOUS$^{\mathrm{l}}$,a modular multi-goal reinforcement learning (RL) algorithm that uses intrinsic motivations to efficiently learn a continuous set of diverse goals using modular goal representations. To build an algorithm able to learn modular goals, one must answer the following questions: 1) How to choose the action policy architecture? 2) How to select
$^{1}$Flowers Team, Inria and Ensta ParisTech, FR.$^2$ISIR, Sorbonne Univ., Paris, FR... Correspondence to: Cédric Colas <cedric.colas@ inria.fr>.
Proceedings of the 36$^{th}$ International Conference on Machine $Learning$,Long Beach, California, PMLR 97, 2019. Copyright 2019 by the author(s).
$^{1}$ CURIOUS stands for Continual Universal Reinforcement learn-
ing with Intrinsically mOtivated sUbstitutionS.
|
[
[
0.265625,
0.1142578125,
0.7099609375,
0.1142578125,
0.7099609375,
0.1287841796875,
0.265625,
0.1287841796875
],
[
0.303955078125,
0.1337890625,
0.6689453125,
0.1363525390625,
0.6689453125,
0.154052734375,
0.303955078125,
0.1514892578125
],
[
0.146240234375,
0.2008056640625,
0.82421875,
0.2008056640625,
0.82421875,
0.2139892578125,
0.146240234375,
0.2139892578125
],
[
0.2442626953125,
0.24365234375,
0.3203125,
0.24365234375,
0.3203125,
0.25830078125,
0.2442626953125,
0.25830078125
],
[
0.021240234375,
0.261474609375,
0.06451416015625,
0.261474609375,
0.06451416015625,
0.40771484375,
0.021240234375,
0.40771484375
],
[
0.1192626953125,
0.267578125,
0.439453125,
0.267578125,
0.439453125,
0.281005859375,
0.1192626953125,
0.281005859375
],
[
0.1192626953125,
0.28271484375,
0.441162109375,
0.2822265625,
0.441162109375,
0.296630859375,
0.1192626953125,
0.29736328125
],
[
0.12091064453125,
0.298583984375,
0.44287109375,
0.298583984375,
0.44287109375,
0.3125,
0.12091064453125,
0.3125
],
[
0.12091064453125,
0.314453125,
0.441162109375,
0.314453125,
0.441162109375,
0.32763671875,
0.12091064453125,
0.32763671875
],
[
0.1217041015625,
0.328857421875,
0.441162109375,
0.328857421875,
0.441162109375,
0.34228515625,
0.1217041015625,
0.34228515625
],
[
0.12335205078125,
0.34521484375,
0.439453125,
0.34521484375,
0.439453125,
0.35595703125,
0.12335205078125,
0.35595703125
],
[
0.5,
0.35107421875,
0.88623046875,
0.35107421875,
0.88623046875,
0.3642578125,
0.5,
0.3642578125
],
[
0.12091064453125,
0.359130859375,
0.44287109375,
0.359130859375,
0.44287109375,
0.37255859375,
0.12091064453125,
0.37255859375
],
[
0.5009765625,
0.365478515625,
0.88623046875,
0.365478515625,
0.88623046875,
0.37890625,
0.5009765625,
0.37890625
],
[
0.1217041015625,
0.375,
0.441162109375,
0.375,
0.441162109375,
0.38818359375,
0.1217041015625,
0.38818359375
],
[
0.5009765625,
0.37890625,
0.88818359375,
0.37890625,
0.88818359375,
0.392578125,
0.5009765625,
0.392578125
],
[
0.12091064453125,
0.389404296875,
0.441162109375,
0.389404296875,
0.441162109375,
0.40283203125,
0.12091064453125,
0.40283203125
],
[
0.0228729248046875,
0.40771484375,
0.062103271484375,
0.40771484375,
0.062103271484375,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.12091064453125,
0.40283203125,
0.44287109375,
0.404052734375,
0.44287109375,
0.41845703125,
0.12091064453125,
0.417236328125
],
[
0.1217041015625,
0.419189453125,
0.441162109375,
0.419189453125,
0.441162109375,
0.43310546875,
0.1217041015625,
0.43310546875
],
[
0.5009765625,
0.42041015625,
0.8857421875,
0.42041015625,
0.8857421875,
0.43505859375,
0.5009765625,
0.43505859375
],
[
0.12335205078125,
0.434326171875,
0.44287109375,
0.434326171875,
0.44287109375,
0.447509765625,
0.12335205078125,
0.447509765625
],
[
0.50244140625,
0.4375,
0.8837890625,
0.4375,
0.8837890625,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.12091064453125,
0.449462890625,
0.441162109375,
0.449462890625,
0.441162109375,
0.46337890625,
0.12091064453125,
0.46337890625
],
[
0.5009765625,
0.451904296875,
0.88623046875,
0.451904296875,
0.88623046875,
0.46533203125,
0.5009765625,
0.46533203125
],
[
0.1192626953125,
0.46337890625,
0.441162109375,
0.464111328125,
0.441162109375,
0.479248046875,
0.1192626953125,
0.477783203125
],
[
0.50244140625,
0.466552734375,
0.88623046875,
0.466552734375,
0.88623046875,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.1217041015625,
0.479736328125,
0.441162109375,
0.479736328125,
0.441162109375,
0.49365234375,
0.1217041015625,
0.49365234375
],
[
0.5009765625,
0.481689453125,
0.82666015625,
0.481689453125,
0.82666015625,
0.49560546875,
0.5009765625,
0.49560546875
],
[
0.0228729248046875,
0.488037109375,
0.05963134765625,
0.488037109375,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.12091064453125,
0.494384765625,
0.44287109375,
0.49560546875,
0.44287109375,
0.5087890625,
0.12091064453125,
0.50830078125
],
[
0.12091064453125,
0.51025390625,
0.441162109375,
0.51025390625,
0.441162109375,
0.5234375,
0.12091064453125,
0.5234375
],
[
0.5009765625,
0.51513671875,
0.72607421875,
0.51513671875,
0.72607421875,
0.52880859375,
0.5009765625,
0.52880859375
],
[
0.71826171875,
0.51513671875,
0.88623046875,
0.51513671875,
0.88623046875,
0.52880859375,
0.71826171875,
0.52880859375
],
[
0.12091064453125,
0.5244140625,
0.441162109375,
0.5234375,
0.441162109375,
0.53857421875,
0.12091064453125,
0.5390625
],
[
0.5,
0.53076171875,
0.8857421875,
0.53076171875,
0.8857421875,
0.54443359375,
0.5,
0.54443359375
],
[
0.1192626953125,
0.5390625,
0.439453125,
0.5390625,
0.439453125,
0.55322265625,
0.1192626953125,
0.55322265625
],
[
0.498291015625,
0.54443359375,
0.88623046875,
0.54541015625,
0.88623046875,
0.56005859375,
0.498291015625,
0.5595703125
],
[
0.12091064453125,
0.5546875,
0.44287109375,
0.5546875,
0.44287109375,
0.568359375,
0.12091064453125,
0.568359375
],
[
0.5009765625,
0.56103515625,
0.88623046875,
0.56103515625,
0.88623046875,
0.57470703125,
0.5009765625,
0.57470703125
],
[
0.12335205078125,
0.5712890625,
0.439453125,
0.5712890625,
0.439453125,
0.58203125,
0.12335205078125,
0.58203125
],
[
0.5,
0.5771484375,
0.8857421875,
0.5771484375,
0.8857421875,
0.59033203125,
0.5,
0.59033203125
],
[
0.12091064453125,
0.583984375,
0.439453125,
0.583984375,
0.439453125,
0.59765625,
0.12091064453125,
0.59765625
],
[
0.5009765625,
0.59130859375,
0.8857421875,
0.59130859375,
0.8857421875,
0.60498046875,
0.5009765625,
0.60498046875
],
[
0.1217041015625,
0.60107421875,
0.441162109375,
0.60107421875,
0.441162109375,
0.6142578125,
0.1217041015625,
0.6142578125
],
[
0.5,
0.60595703125,
0.88623046875,
0.60595703125,
0.88623046875,
0.62353515625,
0.5,
0.62353515625
],
[
0.1192626953125,
0.6142578125,
0.439453125,
0.6142578125,
0.439453125,
0.62744140625,
0.1192626953125,
0.62744140625
],
[
0.5009765625,
0.6220703125,
0.88623046875,
0.6220703125,
0.88623046875,
0.63525390625,
0.5009765625,
0.63525390625
],
[
0.12091064453125,
0.630859375,
0.30224609375,
0.630859375,
0.30224609375,
0.64453125,
0.12091064453125,
0.64453125
],
[
0.5,
0.63525390625,
0.88623046875,
0.63623046875,
0.88623046875,
0.65087890625,
0.5,
0.6494140625
],
[
0.5009765625,
0.6513671875,
0.88623046875,
0.6513671875,
0.88623046875,
0.66552734375,
0.5009765625,
0.66552734375
],
[
0.5,
0.66748046875,
0.62109375,
0.66748046875,
0.62109375,
0.68115234375,
0.5,
0.68115234375
],
[
0.0906982421875,
0.67529296875,
0.21728515625,
0.67529296875,
0.21728515625,
0.68994140625,
0.0906982421875,
0.68994140625
],
[
0.5,
0.68798828125,
0.88623046875,
0.68798828125,
0.88623046875,
0.70263671875,
0.5,
0.70263671875
],
[
0.08905029296875,
0.7021484375,
0.473876953125,
0.7021484375,
0.473876953125,
0.71533203125,
0.08905029296875,
0.71533203125
],
[
0.5,
0.7041015625,
0.88623046875,
0.7041015625,
0.88623046875,
0.71826171875,
0.5,
0.71826171875
],
[
0.0882568359375,
0.71728515625,
0.4755859375,
0.71728515625,
0.4755859375,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.5009765625,
0.7197265625,
0.8857421875,
0.7197265625,
0.8857421875,
0.73291015625,
0.5009765625,
0.73291015625
],
[
0.08905029296875,
0.734375,
0.473876953125,
0.734375,
0.473876953125,
0.744140625,
0.08905029296875,
0.744140625
],
[
0.5009765625,
0.73486328125,
0.8857421875,
0.73486328125,
0.8857421875,
0.74853515625,
0.5009765625,
0.74853515625
],
[
0.0882568359375,
0.74755859375,
0.4755859375,
0.74755859375,
0.4755859375,
0.76123046875,
0.0882568359375,
0.76123046875
],
[
0.5009765625,
0.75,
0.8857421875,
0.75,
0.8857421875,
0.76318359375,
0.5009765625,
0.76318359375
],
[
0.08660888671875,
0.76220703125,
0.4755859375,
0.76123046875,
0.4755859375,
0.77587890625,
0.08660888671875,
0.7763671875
],
[
0.5,
0.76513671875,
0.88623046875,
0.76513671875,
0.88623046875,
0.77880859375,
0.5,
0.77880859375
],
[
0.0882568359375,
0.77783203125,
0.473876953125,
0.77783203125,
0.473876953125,
0.79150390625,
0.0882568359375,
0.79150390625
],
[
0.498291015625,
0.77880859375,
0.88623046875,
0.77978515625,
0.88623046875,
0.79443359375,
0.498291015625,
0.79345703125
],
[
0.0882568359375,
0.79345703125,
0.4755859375,
0.79345703125,
0.4755859375,
0.806640625,
0.0882568359375,
0.806640625
],
[
0.5009765625,
0.79541015625,
0.8857421875,
0.79541015625,
0.8857421875,
0.8095703125,
0.5009765625,
0.8095703125
],
[
0.108642578125,
0.814453125,
0.473876953125,
0.81640625,
0.473876953125,
0.83056640625,
0.108642578125,
0.8291015625
],
[
0.5009765625,
0.81005859375,
0.8857421875,
0.81005859375,
0.8857421875,
0.82373046875,
0.5009765625,
0.82373046875
],
[
0.0882568359375,
0.8291015625,
0.473876953125,
0.8291015625,
0.473876953125,
0.84326171875,
0.0882568359375,
0.84326171875
],
[
0.498291015625,
0.82470703125,
0.8857421875,
0.82373046875,
0.8857421875,
0.83837890625,
0.498291015625,
0.8388671875
],
[
0.08905029296875,
0.84130859375,
0.24267578125,
0.84130859375,
0.24267578125,
0.85498046875,
0.08905029296875,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.8857421875,
0.84033203125,
0.8857421875,
0.85400390625,
0.5009765625,
0.85400390625
],
[
0.5,
0.85595703125,
0.88623046875,
0.85595703125,
0.88623046875,
0.869140625,
0.5,
0.869140625
],
[
0.08660888671875,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.08660888671875,
0.87890625,
0.473876953125,
0.8798828125,
0.473876953125,
0.89453125,
0.08660888671875,
0.89306640625
],
[
0.51953125,
0.87744140625,
0.88623046875,
0.87890625,
0.88623046875,
0.89306640625,
0.51953125,
0.892578125
],
[
0.08905029296875,
0.892578125,
0.221435546875,
0.892578125,
0.221435546875,
0.90576171875,
0.08905029296875,
0.90576171875
],
[
0.5,
0.892578125,
0.77783203125,
0.892578125,
0.77783203125,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"CURIOUS: Intrinsically Motivated Modular",
"Multi-Goal Reinforcement Learning",
"Cedric Colas Pierre Fournier2 Olivier Sigaud? Mohamed Chetouani? Pierre-Yves Oudeyer",
"Abstract",
"61062",
"In open-ended environments, autonomous learn",
"ing agents must set their own goals and build their",
"own curriculum through an intrinsically motivated",
"exploration. They may consider a large diversity",
"of goals, aiming to discover what is controllable",
"intheirenvironments,andwhatisnot.Because",
"Figure I. The Modular Goal Fetch Arm environment. An intrin-",
"some goals might prove easy and some impos-",
"sically motivated agent can set its own (modular) goals (Reach,",
"sible, agents must actively select which goal to",
"Push, Pick and Place, Stack), with multiple objects and distractors.",
"practice at any moment, to maximize their overall",
"[I's]",
"mastery on the set of learnable goals. This paper",
"proposes CURIoUS, an algorithm that leverages",
"and self-organize one's own curriculum. This challenge can",
"I) a modular Universal Value Function Approxi-",
"betackledwithintheframeworkofIntrinsicallyMotivated",
"mator with hindsight learning to achieve a diver-",
"Goal Exploration Processes (IMGEP) (Baranes & Oudeyer,",
"sity of goals of different kinds within a unique",
"2013; Forestier et al.,2017),leveraging computational mod",
"policy and 2) an automated curriculum learning",
"els of autonomous development in human infants.",
"8200181",
"mechanism that biases the attention of the agent",
"towards goals maximizing the absolute learning",
"Modular goal representation.",
"In a same environment,",
"progress. Agents focus sequentially on goals of",
"an agent might want to 'put the cube in position ' or",
"increasing complexity, and focus back on goals",
"to 'reach position y' for any or y. Here, describ-",
"that are being forgotten. Experiments conducted",
"ing the full goal space requires modular goal representa-",
"inanewmodular-goal roboticenvironmentshow",
"tions. Goals are organized by modules, where module",
"the resulting developmental self-organization of a",
"refers to the pair of a reward function and a goal space",
"learning curriculum, and demonstrate properties",
"M; = (RMi,giEgM, GM.). The reward function describes",
"of robustness to distracting goals, forgetting and",
"a set of constraints that must be satisfied by the agent's state",
"changes in body properties.",
"(e.g. Reach), given a continuous parameter (e.g. gi = y)",
"evolving in the associated goal space (e.g. 3D Euclidean",
"space), see Fig. 1.",
"1. Introduction",
"While flat multi-goal problems with continuous (Schaul",
"In autonomous continual learning, agents aim to discover",
"et al., 2015a; Andrychowicz et al., 2017; Plappert et al.,",
"repertoires of skills in an ever-changing open-ended world,",
"2018) or discrete goals (Mankowitz et al., 2018; Riedmiller",
"andwithoutexternalrewards.Insuchrealisticenvironments,",
"et al., 2018) have been explored in the past, only few works",
"the agent must be endowed with intrinsic motivations to",
"tackle the problem of modular multi-goal learning (Forestier",
"explore the diversity of ways in which it can control its",
"& Oudeyer, 2016), none in an RL setting. Here, we present",
"environment. One important form of intrinsic motivation",
"CURIoUs', a modular multi-goal reinforcement learning",
"system is the ability to autonomously set one's own goals",
"(RL) algorithm that uses intrinsic motivations to efficiently",
"'Flowers Team, Inria and Ensta ParisTech, FR. 2ISIR, Sor-",
"learn a continuous set of diverse goals using modular goal",
"bonne Univ., Paris, FR...Correspondence to: Cedric Colas",
"representations. To build an algorithm able to learn modular",
"<cedric.colas@inria.fr>.",
"goals, one must answer the following questions: 1) How",
"to choose the action policy architecture? 2) How to select",
"Proceedings of the 36th Intermational Conference on Machine",
"Learning, Long Beach, California, PMLR 97, 2019. Copyright",
"I cURIOUS stands for Continual Universal Reinforcement learn-",
"2019 by the author(s).",
"ing with Intrinsically mOtivated sUbstitutionS."
] |
{
"arxiv_id": "1810.06284",
"language": "en",
"page_idx": 0
}
|
||
1810.06284-en-1
|
1810.06284-en-1.jpg
|
CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning
the next module and goal to practice and learn about? 3) How to efficiently transfer knowledge between modules and goals?
Related work. Kaelbling (1993) proposed the first algorithm able to leverage cross-goal learning to address a discrete set of goals. For each goal, the algorithm learned a specific value function using Q-learning (goal-experts approach). More recently, Schaul et al. (2015a) proposed Universal Value Function Approximators (UVFA), a unique policy able to address an infinity of goals by concatenating the current state and goal to feed both the policy and the value function. In UNICORN, UVFA is used to address a discrete set of goals in parallel: reaching different objects in a visual world (Mankowitz et al., 2018). SAC-X implements multi-task RL where easy tasks are considered as auxiliary tasks to help learning about the hardest task (placing cubes inside a closed box) (Riedmiller et al., 2018). Here, one network is trained for each task and the collected transitions are shared (goal-experts approach). In other works from multi-task RL (Teh et al., 2017; Espeholt et al., 2018; Hessel et al., 2018), agents do not represent explicitly the current task and aim at maximizing the overall reward. Finally, within the Intrinsically Motivated Goal Exploration Processes framework (IMGEP), Forestier & Oudeyer (2016) proposed MACOB, an algorithm able to target modular goals using a population-based algorithm that mutates and replays controllers experienced in the past. MACOB maintains a population of solutions, one for each goal (modular goalexperts approach), see Nguyen & Oudeyer (2012) for a similar approach. This enables efficient cross-goal learning in high-dimensional goal spaces, but is limited by the memory-based representation of policies.
Multi-goal approaches prove better than simply training
a policy per goal because knowledge can be transferred between different goals using off-policy learning and hindsight learning (Andrychowicz et al., 2017). Off-policy learning enables the use of any transition to improve the current policy: transitions collected from an older version of the current policy (Lillicrap et al., 2015), from a population of exploratory policies (Colas et al., 2018), or even from demonstrations (Večerík et al., 2017). Transitions collected while aiming at a particular goal can be used to learn about any other. With finite sets of goals, each transition is generally used to update the policy on every other goal (Mankowitz et al., 2018; Kaelbling, 1993). With continuous sets of goals, imaginary goals are sampled from the goal space (Andrychowicz et al., 2017). In the case of UVFA policies, this consists in the substitution of the goal that is part of the input by the imaginary one, a technique called goal replay or goal substitution. Building on UVFA, Andrychowicz et al. (2017) proposed Hindsight Experience Replay (HER), a method leveraging hindsight for transferring knowledge
between goals. The original goal of a transition can be substituted by any outcome experienced later in the trajectory (imaginary goal). This helps to increase the probability to observe rewards in reward-sparse environments.
In the literature, environments usually provide goals that the agent is asked to solve. In the IMGEP framework however, autonomous agents are intrinsically motivated to set their own, possibly learning their representation (LaversanneFinot et al., 2018). Forestier & Oudeyer (2016) in particular, biased the selection of the next goal to attempt towards modules showing high absolute measures of learning progress (LP). This mechanism helps the agent to focus on learnable goals and to disengage from goals that are currently too hard or already solved. Veeriah et al. (2018) uses LP computed from Bellman errors for goal selection, but this form of LP does not improve over random goal selection.
Additional background can be found in the supplementary document. Table 1 presents a classification of the multi-goal approaches most related to our work.
Contributions. The contributions of this paper are:
1. A modular encoding of goals to enable learning of
continuous sets of diverse goals within a single policy using UVFA (Reach, Push, Pick and Place, Stack). This enables to tackle different kinds of goals, each with their own continuous parameterization, and facilitates transfer between modules and goals. See Sec. 2.1.
2. An active strategy for cross-module goal replay. Off-
policy learning enables to use any experience to learn about any goal from any module. We propose to guide the selection of module for replay using absolute learning progress measures (in addition to LP-based goal sampling to interact with environment). See Sec. 2.2.
3. From the IMGEP perspective, a single monolithic modu-
lar multi-goal action policy. This is an alternative to the
population-based algorithms studied so far (Forestier
& Q Oudeyer, 2016; Forestier et al., 2017) and provides
the flexibility of RL methods.
4. An environment for modular goal RL. See Sec. 3.
5. Empirical comparisons to other architectures: a goal-
parameterized RL with HER (flat multi-goal RL) and a multi-goal module-experts approach (one multi-goal expert per module). See Sec. 4.1.
6. A study of the self-organization of learning phases
demonstrated by our algorithm (automatic curriculum).
## See Sec. 4.2.
7. Experimental evidence of robustness to distracting
goals, forgetting and body changes in comparison with
random goal selection. See Sec. 4.3 and 4.4.
|
[
[
0.2296142578125,
0.056182861328125,
0.744140625,
0.05682373046875,
0.744140625,
0.0706787109375,
0.2296142578125,
0.0694580078125
],
[
0.08905029296875,
0.085205078125,
0.4755859375,
0.085205078125,
0.4755859375,
0.0985107421875,
0.08905029296875,
0.0985107421875
],
[
0.498291015625,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.498291015625,
0.0997314453125
],
[
0.08905029296875,
0.10101318359375,
0.473876953125,
0.10101318359375,
0.473876953125,
0.1142578125,
0.08905029296875,
0.1142578125
],
[
0.5,
0.10162353515625,
0.8837890625,
0.10162353515625,
0.8837890625,
0.11553955078125,
0.5,
0.11553955078125
],
[
0.0882568359375,
0.11614990234375,
0.134033203125,
0.11614990234375,
0.134033203125,
0.1307373046875,
0.0882568359375,
0.1307373046875
],
[
0.498291015625,
0.11614990234375,
0.88623046875,
0.11553955078125,
0.88623046875,
0.1300048828125,
0.498291015625,
0.1307373046875
],
[
0.5,
0.1319580078125,
0.814453125,
0.1319580078125,
0.814453125,
0.1458740234375,
0.5,
0.1458740234375
],
[
0.08905029296875,
0.1590576171875,
0.473876953125,
0.1590576171875,
0.473876953125,
0.17236328125,
0.08905029296875,
0.17236328125
],
[
0.5,
0.154052734375,
0.8857421875,
0.154052734375,
0.8857421875,
0.1673583984375,
0.5,
0.1673583984375
],
[
0.0882568359375,
0.1749267578125,
0.4755859375,
0.1749267578125,
0.4755859375,
0.1881103515625,
0.0882568359375,
0.1881103515625
],
[
0.5,
0.169189453125,
0.88623046875,
0.169189453125,
0.88623046875,
0.18310546875,
0.5,
0.18310546875
],
[
0.0882568359375,
0.189453125,
0.4755859375,
0.189453125,
0.4755859375,
0.20263671875,
0.0882568359375,
0.20263671875
],
[
0.5,
0.184326171875,
0.8857421875,
0.184326171875,
0.8857421875,
0.1976318359375,
0.5,
0.1976318359375
],
[
0.08660888671875,
0.203857421875,
0.473876953125,
0.2052001953125,
0.473876953125,
0.2197265625,
0.08660888671875,
0.2183837890625
],
[
0.5,
0.199462890625,
0.8857421875,
0.199462890625,
0.8857421875,
0.21337890625,
0.5,
0.21337890625
],
[
0.5009765625,
0.2139892578125,
0.88623046875,
0.2139892578125,
0.88623046875,
0.2279052734375,
0.5009765625,
0.2279052734375
],
[
0.08905029296875,
0.2203369140625,
0.473876953125,
0.2203369140625,
0.473876953125,
0.2342529296875,
0.08905029296875,
0.2342529296875
],
[
0.0906982421875,
0.23486328125,
0.473876953125,
0.23486328125,
0.473876953125,
0.248779296875,
0.0906982421875,
0.248779296875
],
[
0.5,
0.2298583984375,
0.88623046875,
0.2298583984375,
0.88623046875,
0.24365234375,
0.5,
0.24365234375
],
[
0.5,
0.24365234375,
0.88623046875,
0.24560546875,
0.88623046875,
0.2587890625,
0.5,
0.2568359375
],
[
0.08905029296875,
0.250732421875,
0.473876953125,
0.250732421875,
0.473876953125,
0.264404296875,
0.08905029296875,
0.264404296875
],
[
0.5,
0.260009765625,
0.8857421875,
0.260009765625,
0.8857421875,
0.2734375,
0.5,
0.2734375
],
[
0.0906982421875,
0.267578125,
0.47216796875,
0.267578125,
0.47216796875,
0.27783203125,
0.0906982421875,
0.27783203125
],
[
0.498291015625,
0.27587890625,
0.8857421875,
0.2734375,
0.8857421875,
0.287353515625,
0.498291015625,
0.2890625
],
[
0.0906982421875,
0.2822265625,
0.473876953125,
0.2822265625,
0.473876953125,
0.292236328125,
0.0906982421875,
0.292236328125
],
[
0.5,
0.290283203125,
0.88623046875,
0.290283203125,
0.88623046875,
0.3037109375,
0.5,
0.3037109375
],
[
0.08905029296875,
0.29736328125,
0.47216796875,
0.29736328125,
0.47216796875,
0.30810546875,
0.08905029296875,
0.30810546875
],
[
0.5,
0.3037109375,
0.8857421875,
0.304931640625,
0.8857421875,
0.3193359375,
0.5,
0.318115234375
],
[
0.08905029296875,
0.3125,
0.47216796875,
0.3125,
0.47216796875,
0.322509765625,
0.08905029296875,
0.322509765625
],
[
0.0882568359375,
0.32568359375,
0.473876953125,
0.32568359375,
0.473876953125,
0.339111328125,
0.0882568359375,
0.339111328125
],
[
0.5009765625,
0.32080078125,
0.80224609375,
0.32080078125,
0.80224609375,
0.333984375,
0.5009765625,
0.333984375
],
[
0.0882568359375,
0.34033203125,
0.473876953125,
0.34033203125,
0.473876953125,
0.354248046875,
0.0882568359375,
0.354248046875
],
[
0.5,
0.341552734375,
0.8837890625,
0.34228515625,
0.8837890625,
0.356689453125,
0.5,
0.35595703125
],
[
0.0882568359375,
0.35595703125,
0.473876953125,
0.35595703125,
0.473876953125,
0.369384765625,
0.0882568359375,
0.369384765625
],
[
0.5009765625,
0.35791015625,
0.8857421875,
0.35791015625,
0.8857421875,
0.371826171875,
0.5009765625,
0.371826171875
],
[
0.0882568359375,
0.37060546875,
0.473876953125,
0.37060546875,
0.473876953125,
0.3837890625,
0.0882568359375,
0.3837890625
],
[
0.498291015625,
0.37255859375,
0.74609375,
0.371826171875,
0.74609375,
0.386474609375,
0.498291015625,
0.386962890625
],
[
0.08905029296875,
0.386474609375,
0.473876953125,
0.386474609375,
0.473876953125,
0.399658203125,
0.08905029296875,
0.399658203125
],
[
0.08905029296875,
0.40087890625,
0.473876953125,
0.40087890625,
0.473876953125,
0.4140625,
0.08905029296875,
0.4140625
],
[
0.5,
0.401611328125,
0.8486328125,
0.404052734375,
0.8486328125,
0.417236328125,
0.5,
0.415283203125
],
[
0.08905029296875,
0.416748046875,
0.47216796875,
0.416748046875,
0.47216796875,
0.429931640625,
0.08905029296875,
0.429931640625
],
[
0.08905029296875,
0.43310546875,
0.47216796875,
0.43310546875,
0.47216796875,
0.443115234375,
0.08905029296875,
0.443115234375
],
[
0.5146484375,
0.43310546875,
0.88623046875,
0.43310546875,
0.88623046875,
0.4462890625,
0.5146484375,
0.4462890625
],
[
0.08905029296875,
0.4462890625,
0.473876953125,
0.4462890625,
0.473876953125,
0.460205078125,
0.08905029296875,
0.460205078125
],
[
0.53173828125,
0.447509765625,
0.8837890625,
0.447509765625,
0.8837890625,
0.4609375,
0.53173828125,
0.4609375
],
[
0.0906982421875,
0.46337890625,
0.47216796875,
0.46337890625,
0.47216796875,
0.473388671875,
0.0906982421875,
0.473388671875
],
[
0.53173828125,
0.46337890625,
0.8857421875,
0.46337890625,
0.8857421875,
0.4765625,
0.53173828125,
0.4765625
],
[
0.08905029296875,
0.4765625,
0.473876953125,
0.4765625,
0.473876953125,
0.490478515625,
0.08905029296875,
0.490478515625
],
[
0.53369140625,
0.479248046875,
0.8857421875,
0.479248046875,
0.8857421875,
0.492431640625,
0.53369140625,
0.492431640625
],
[
0.08905029296875,
0.492431640625,
0.473876953125,
0.492431640625,
0.473876953125,
0.505859375,
0.08905029296875,
0.505859375
],
[
0.53173828125,
0.49365234375,
0.8857421875,
0.49365234375,
0.8857421875,
0.5068359375,
0.53173828125,
0.5068359375
],
[
0.0882568359375,
0.5068359375,
0.473876953125,
0.5068359375,
0.473876953125,
0.52099609375,
0.0882568359375,
0.52099609375
],
[
0.53173828125,
0.5087890625,
0.861328125,
0.5087890625,
0.861328125,
0.52294921875,
0.53173828125,
0.52294921875
],
[
0.08660888671875,
0.521484375,
0.4755859375,
0.52099609375,
0.4755859375,
0.53515625,
0.08660888671875,
0.5361328125
],
[
0.5107421875,
0.53076171875,
0.8857421875,
0.53076171875,
0.8857421875,
0.54443359375,
0.5107421875,
0.54443359375
],
[
0.0882568359375,
0.537109375,
0.4755859375,
0.537109375,
0.4755859375,
0.55029296875,
0.0882568359375,
0.55029296875
],
[
0.08905029296875,
0.5517578125,
0.473876953125,
0.5517578125,
0.473876953125,
0.56494140625,
0.08905029296875,
0.56494140625
],
[
0.53369140625,
0.54736328125,
0.88623046875,
0.54736328125,
0.88623046875,
0.56103515625,
0.53369140625,
0.56103515625
],
[
0.08905029296875,
0.5673828125,
0.473876953125,
0.5673828125,
0.473876953125,
0.58056640625,
0.08905029296875,
0.58056640625
],
[
0.53369140625,
0.56298828125,
0.8857421875,
0.56298828125,
0.8857421875,
0.5771484375,
0.53369140625,
0.5771484375
],
[
0.0882568359375,
0.58349609375,
0.36279296875,
0.58349609375,
0.36279296875,
0.5966796875,
0.0882568359375,
0.5966796875
],
[
0.53515625,
0.5791015625,
0.8857421875,
0.5791015625,
0.8857421875,
0.5888671875,
0.53515625,
0.5888671875
],
[
0.53173828125,
0.59228515625,
0.8857421875,
0.59130859375,
0.8857421875,
0.60595703125,
0.53173828125,
0.60693359375
],
[
0.08905029296875,
0.60400390625,
0.47216796875,
0.60400390625,
0.47216796875,
0.6171875,
0.08905029296875,
0.6171875
],
[
0.53173828125,
0.60791015625,
0.88134765625,
0.60791015625,
0.88134765625,
0.6220703125,
0.53173828125,
0.6220703125
],
[
0.08905029296875,
0.62060546875,
0.473876953125,
0.62060546875,
0.473876953125,
0.6337890625,
0.08905029296875,
0.6337890625
],
[
0.08905029296875,
0.63525390625,
0.473876953125,
0.63525390625,
0.473876953125,
0.6484375,
0.08905029296875,
0.6484375
],
[
0.5107421875,
0.630859375,
0.88623046875,
0.630859375,
0.88623046875,
0.64453125,
0.5107421875,
0.64453125
],
[
0.08905029296875,
0.6494140625,
0.47216796875,
0.6494140625,
0.47216796875,
0.66357421875,
0.08905029296875,
0.66357421875
],
[
0.53125,
0.646484375,
0.88623046875,
0.646484375,
0.88623046875,
0.66015625,
0.53125,
0.66015625
],
[
0.08905029296875,
0.66552734375,
0.473876953125,
0.66552734375,
0.473876953125,
0.6787109375,
0.08905029296875,
0.6787109375
],
[
0.53173828125,
0.6611328125,
0.8857421875,
0.66015625,
0.8857421875,
0.6748046875,
0.53173828125,
0.67529296875
],
[
0.08660888671875,
0.6796875,
0.4755859375,
0.6787109375,
0.4755859375,
0.693359375,
0.08660888671875,
0.6943359375
],
[
0.53173828125,
0.6748046875,
0.8857421875,
0.67529296875,
0.8857421875,
0.68994140625,
0.53173828125,
0.689453125
],
[
0.08905029296875,
0.69580078125,
0.4755859375,
0.69580078125,
0.4755859375,
0.708984375,
0.08905029296875,
0.708984375
],
[
0.53173828125,
0.69140625,
0.7294921875,
0.69140625,
0.7294921875,
0.705078125,
0.53173828125,
0.705078125
],
[
0.08905029296875,
0.71044921875,
0.4755859375,
0.71044921875,
0.4755859375,
0.7236328125,
0.08905029296875,
0.7236328125
],
[
0.51123046875,
0.71630859375,
0.859375,
0.71630859375,
0.859375,
0.7265625,
0.51123046875,
0.7265625
],
[
0.08905029296875,
0.72607421875,
0.4755859375,
0.72607421875,
0.4755859375,
0.7392578125,
0.08905029296875,
0.7392578125
],
[
0.0906982421875,
0.7421875,
0.47216796875,
0.7421875,
0.47216796875,
0.75244140625,
0.0906982421875,
0.75244140625
],
[
0.51123046875,
0.73779296875,
0.88623046875,
0.73779296875,
0.88623046875,
0.751953125,
0.51123046875,
0.751953125
],
[
0.0882568359375,
0.755859375,
0.473876953125,
0.755859375,
0.473876953125,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.53125,
0.75390625,
0.88623046875,
0.75244140625,
0.88623046875,
0.767578125,
0.53125,
0.76806640625
],
[
0.0882568359375,
0.771484375,
0.4755859375,
0.771484375,
0.4755859375,
0.78515625,
0.0882568359375,
0.78515625
],
[
0.53369140625,
0.76953125,
0.8857421875,
0.76953125,
0.8857421875,
0.78271484375,
0.53369140625,
0.78271484375
],
[
0.0882568359375,
0.7861328125,
0.473876953125,
0.7861328125,
0.473876953125,
0.7998046875,
0.0882568359375,
0.7998046875
],
[
0.53173828125,
0.78515625,
0.75341796875,
0.78271484375,
0.75341796875,
0.796875,
0.53173828125,
0.798828125
],
[
0.08905029296875,
0.8017578125,
0.473876953125,
0.8017578125,
0.473876953125,
0.81494140625,
0.08905029296875,
0.81494140625
],
[
0.5107421875,
0.80615234375,
0.8837890625,
0.806640625,
0.8837890625,
0.8212890625,
0.5107421875,
0.82080078125
],
[
0.08905029296875,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.830078125,
0.08905029296875,
0.830078125
],
[
0.53125,
0.8212890625,
0.88623046875,
0.82275390625,
0.88623046875,
0.8369140625,
0.53125,
0.8359375
],
[
0.0906982421875,
0.83349609375,
0.473876953125,
0.83349609375,
0.473876953125,
0.84326171875,
0.0906982421875,
0.84326171875
],
[
0.53515625,
0.8388671875,
0.62109375,
0.8388671875,
0.62109375,
0.849609375,
0.53515625,
0.849609375
],
[
0.0882568359375,
0.8466796875,
0.4755859375,
0.8466796875,
0.4755859375,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.0882568359375,
0.861328125,
0.4755859375,
0.861328125,
0.4755859375,
0.875,
0.0882568359375,
0.875
],
[
0.51318359375,
0.861328125,
0.8837890625,
0.861328125,
0.8837890625,
0.875,
0.51318359375,
0.875
],
[
0.0882568359375,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.53369140625,
0.876953125,
0.8857421875,
0.876953125,
0.8857421875,
0.89013671875,
0.53369140625,
0.89013671875
],
[
0.0882568359375,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.53173828125,
0.8916015625,
0.82421875,
0.8916015625,
0.82421875,
0.90478515625,
0.53173828125,
0.90478515625
]
] |
[
"CURIOUS: IntrinsicallyMotivated Modular Multi-Goal Reinforcement Learning",
"the next module and goal to practice and learn about? 3)",
"between goals. The original goal of a transition can be sub-",
"How to efficiently transferknowledge between modules and",
"stituted by any outcome experienced later in the trajectory",
"goals?",
"(imaginary goal). This helps to increase the probability to",
"observe rewards in reward-sparse environments.",
"Related work.Kaelbling (1993) proposed the first algo-",
"In the literature, environments usually provide goals that the",
"rithm able to leverage cross-goal learning to address a dis-",
"agent is asked to solve. In the IMGEP framework however,",
"crete set of goals. For each goal, the algorithm learned a",
"autonomous agents are intrinsically motivated to set their",
"specific value function using Q-learning (goal-experts ap-",
"own, possibly learning their representation (Laversanne-",
"Finot et al., 2018).Forestier & Oudeyer (2016) in particular,",
"proach). More recently, Schaul et al. (2015a) proposed",
"Universal Value Function Approximators (UvFA), a unique",
"biased the selection of the next goal to attempt towards mod-",
"ules showing high absolute measures of learning progress",
"policy able to address an infinity of goals by concatenating",
"(LP). This mechanism helps the agent to focus on learnable",
"the current state and goal to feed boththe policy andthe",
"goals and to disengage fromgoals that are currentlytoo hard",
"value function. In UNICORN.UVFA is used to address a",
"or already solved. Veeriah et al. (2018) uses LP computed",
"discrete set of goals in parallel: reaching different objects in",
"from Bellman errors for goal selection, but this form of LP",
"avisual world (Mankowitz etal..2018).SAC-ximplements",
"multi-task RL where easy tasks are considered as auxiliary",
"does not improve over random goal selection.",
"tasks to help learning about the hardest task (placing cubes",
"Additional background can be found in the supplementary",
"inside a closed box) (Riedmiller et al., 2018). Here, one",
"document. Table 1 presents a classification of the multi-goal",
"network is trained for each task and the collected transi-",
"approaches most related to our work.",
"tions are shared (goal-experts approach). In other works",
"from multi-task RL (Teh et al., 2017; Espeholt et al., 2018:",
"Contributions.The contributions of this paper are:",
"Hessel et al., 2018), agents do not represent explicitly the",
"current task and aim at maximizing the overall reward. Fi",
"1. A modular encoding of goals to enable learning of",
"nally, within the Intrinsically Motivated Goal Exploration",
"continuous sets of diverse goals within a single policy",
"Processesframework(IMGEP),Forestier&Oudeyer(2016)",
"using UvFA (Reach, Push, Pick and Place, Stack). This",
"proposed MACoB, an algorithm able to target modular goals",
"enables to tackle different kinds of goals, each with",
"using a population-based algorithm that mutates and replays",
"their own continuous parameterization, and facilitates",
"controllers experienced in the past. MACOB maintains a",
"transfer between modules and goals. See Sec. 2.1.",
"population of solutions, one for each goal (modular goal-",
"2.An active strategy for cross-modulegoal replay.Off",
"experts approach), see Nguyen & Oudeyer (2012) for a",
"similar approach. This enables efficient cross-goal learn-",
"policy learning enables to use any experience to learn",
"ing in high-dimensional goal spaces, but is limited by the",
"about any goal from any module. We propose to guide",
"memory-based representation of policies.",
"theselectionofmoduleforreplayusingabsolutelearn-",
"ing progress measures (in addition to LP-based goal",
"Multi-goal approaches prove better than simply training",
"sampling to interact with environment). See Sec. 2.2.",
"a policy per goal because knowledge can be transferred",
"between different goals using off-policy learning and hind",
"3. From the IMGEP perspective, a single monolithic modu-",
"sight learning (Andrychowicz et al., 2017). Off-policy learn",
"lar multi-goal action policy. This is an alternative to the",
"ing enables the use of any transition to improve the cur",
"population-based algorithms studied so far (Forestier",
"rent policy: transitions collected from an older version of",
"& Oudeyer, 2016; Forestier et al., 2017) and provides",
"the current policy (Lillicrap et al., 2015), from a popula-",
"the flexibility of RL methods.",
"tion of exploratory policies (Colas et al., 2018), or even",
"4.AnenvironmentformodulargoalRL.SeeSec.3.",
"from demonstrations (Vecerik et al., 2017). Transitions col-",
"lected whileaimingataparticulargoal canbeusedtolearn",
"5. Empirical comparisons to other architectures: a goal-",
"about any other. With finite sets of goals, each transition",
"parameterized RL with HER (flat multi-goal RL)and",
"is generally used to update the policy on every other goal",
"a multi-goal module-experts approach (one multi-goal",
"(Mankowitz et al., 2018; Kaelbling, 1993). With continuous",
"expertpermodule).SeeSec.4.1.",
"sets of goals, imaginary goals are sampled from the goal",
"6. A study of the self-organization of learning phases",
"space (Andrychowicz et al., 2017). In the case of UvFA poli",
"demonstrated by our algorithm (automatic curriculum).",
"cies,this consists in the substitution of thegoal thatis part",
"See Sec.4.2.",
"of the input by the imaginary one, a technique called goal re-",
"play or goal substitution. Building on UvFA, Andrychowicz",
"7. Experimental evidence of robustness to distracting",
"et al. (2017) proposed Hindsight Experience Replay (HER),",
"goals, forgetting and body changes in comparison with",
"a method leveraging hindsight for transferring knowledge",
"random goal selection. See Sec. 4.3 and 4.4."
] |
{
"arxiv_id": "1810.06284",
"language": "en",
"page_idx": 1
}
|
||
1810.06284-en-2
|
1810.06284-en-2.jpg
|
CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning
# 2. CURIOUS
## 2.1. A Modular Multi-Goal Architecture using Universal Approximators
<FigureHere>
Figure 2.Modular goal-parameterized actor-critic architecture (M-UVFA). Toy example with 2 modules, parameterized by $g_1$ (2D) and $g_2$ (1D) respectively. Here, the agent is attempting goal $g_1$ in module $M_1$, as specified by the one-hot module descriptor $m_d~=~\langle1,0\rangle.~$The actor (left) computes the action $a_t.$ The critic (right) computes the $Q$-value.
UVFA concatenates the goal of the agent with its current state to form the input of the policy and the value function implemented by deep neural networks (Schaul et al., 2015a). With cuRIOUS, we propose a new encoding of goals using modular representations. This enables to target a rich diversity of modular goals within a single network $(modularmulti-goal$ approach), see Fig. 2. Given $\mathcal{G}_M_i$ the goal space of module $M_i$,the current goal $g$ is defined as ${\mathrm{a~vector~of~dimension~}}|{\mathcal{G}}|=\sum_{i=1}^{N}|{\mathcal{G}}_{M_{i}}|,{\mathrm{where~the~}\mathcal{G}}_{M_{i}}$ can have different dimensionalities. $g$ is set to 0 everywhere except in the indices corresponding to the current module $M_i$,where it is set to $g_i\in\mathcal{G}_{M_i}.$ By masking the goal-inputs corresponding to unconsidered modules, the corresponding weights are frozen during backpropagation. In addition, a module descriptor $m_d$ of size $N$ (one-hot encoding) encodes the current module. The overall input to the policy network is $[s_t,g,m_d]$, see Fig.2. We call this modular goal-parameterized architecture Modular-UVFA (M-UVFA). In Fig. 2, we can see the underlying learning architecture (actor-critic). The actor implements the action policy and maps the input $[s_t,g,m_d]$ to the next action $a_t.$ The action is then concatenated to a copy of the actor's input to feed the critic $[s_t,g,m_d,a_t].$ The critic provides an estimate of the $Q$-value: $Q(s_t,g,m_{d},a_{t}).$ The critic and the actor are then trained using DDPG (Lillicrap et al., 2015), although any other off-policy learning method could be used (e.g. TD3 (Fujimoto et al., 2018), or DQN for the discrete case (Mnih et al., 2013)). More details about DDPG can be found in the supplementary document or in Lillicrap et al. (2015).
on the other hand, agents have the capacity to select which goal to target next. Because goals are not equivalent, the agent can benefit from intrinsic motivations towards LP (Schmidhuber, 1991; Kaplan & Oudeyer, 2004). This can be useful: 1) when there are distracting goals on which the agent cannot progress; 2) when some goals are already mastered. This idea comes from the IMGEP framework and was used in Baranes & Oudeyer (2013) to guide goal selection and in Forestier & Oudeyer (2016) to guide module selection. The problem of selecting a module can be modeled as a non-stationary multi-armed bandit (MAB), where the value of each arm (module) is the current absolute LP. Learning progress (LP) is defined as the derivative of the agent's competence on a particular module: $LP_{M_i}=\frac{dC_{M_i}}{dt}$, where the competence $C_{M_i}:t\to p_{success}(t)$ is the probability of success at time $t.$ Here, the agent focuses its attention on modules for which it is making the largest absolute progress, and pays little attention to modules that are already solved or unsolvable, i.e. for which $|LP|$ stays small. Using the absolute value of LP also leads to the prioritization of modules for which the agent is showing decreasing performances. This helps to deal with forgetting: the agent reallocates learning resources to the modules being forgotten, Sec. 4.3.
## 2.2. Module and Goal Selection, Cross-Module Learning, Cross-Goal Learning
In UVFA, HER and UNICORN, the next goal to target is selected at random (Schaul et al., 2015a; Andrychowicz et al., 2017; Mankowitz et al., 2018). This is coherent with the common view that the agent must comply with the desires of an engineer and target the goal it is asked to target. Here
Table 1. Classification of multi-goal approaches. Underlined: Algorithms internally generating goals (IMGEP),(*) using LP-based
intrinsic motivations. $Italic:$ Population-based algorithms (non-RL). Bold: Algorithms proposed in this paper.
<table>
<tbody>
<tr>
<th> </th>
<th>$n$ GOALS, $n$ POLICIES</th>
<th>$n$GOALS,1POLICY</th>
</tr>
<tr>
<td>FLAT REPR.</td>
<td>GOAL-EXPERTS: (Kaelbling, 1993) SAC-X (Riedmiller et al., 2018) $SAGG-RIAC^{*}\left(\mathrm{Baranes~ ~Oudeyer,~2013}\right)$</td>
<td>$\mathbf{M}$ULTI-GOAL: UVFA(Schaul et al., 2015aj HER (Andrychowicz et al., 2017) UNICORN (Mankowitz et al., 2018)</td>
</tr>
<tr>
<td>${\mathrm{MODULAR~REPR}}.$</td>
<td>MOD-GOAL-EXPERTS: $MACOB^*($Forestier& Oudeyer, 2016) MULTI- GOAL MODU $\mathbf{S}\colon\mathbf{MG-ME}$</td>
<td>MODULAR- MULTI- GOALS: $\mathbf{M-UVFA.CURIOUS}^{*}$</td>
</tr>
</tbody>
</table>
|
[
[
0.2303466796875,
0.05682373046875,
0.74609375,
0.05682373046875,
0.74609375,
0.0706787109375,
0.2303466796875,
0.0706787109375
],
[
0.0882568359375,
0.083984375,
0.200927734375,
0.083984375,
0.200927734375,
0.0985107421875,
0.0882568359375,
0.0985107421875
],
[
0.59326171875,
0.08709716796875,
0.65673828125,
0.08709716796875,
0.65673828125,
0.09783935546875,
0.59326171875,
0.09783935546875
],
[
0.5810546875,
0.0985107421875,
0.6708984375,
0.0985107421875,
0.6708984375,
0.10986328125,
0.5810546875,
0.10986328125
],
[
0.08660888671875,
0.10919189453125,
0.41259765625,
0.10986328125,
0.41259765625,
0.125,
0.08660888671875,
0.12371826171875
],
[
0.7060546875,
0.10479736328125,
0.81787109375,
0.10479736328125,
0.81787109375,
0.11553955078125,
0.7060546875,
0.11553955078125
],
[
0.62353515625,
0.11297607421875,
0.65283203125,
0.11297607421875,
0.65283203125,
0.121826171875,
0.62353515625,
0.121826171875
],
[
0.71728515625,
0.11553955078125,
0.80615234375,
0.11614990234375,
0.80615234375,
0.1275634765625,
0.71728515625,
0.126953125
],
[
0.11517333984375,
0.125,
0.29736328125,
0.1256103515625,
0.29736328125,
0.14013671875,
0.11517333984375,
0.1395263671875
],
[
0.54833984375,
0.1307373046875,
0.60546875,
0.1307373046875,
0.60546875,
0.1427001953125,
0.54833984375,
0.1427001953125
],
[
0.650390625,
0.1331787109375,
0.6689453125,
0.1331787109375,
0.6689453125,
0.141357421875,
0.650390625,
0.141357421875
],
[
0.54150390625,
0.143310546875,
0.611328125,
0.143310546875,
0.611328125,
0.1546630859375,
0.54150390625,
0.1546630859375
],
[
0.08905029296875,
0.149658203125,
0.4755859375,
0.149658203125,
0.4755859375,
0.1634521484375,
0.08905029296875,
0.1634521484375
],
[
0.5361328125,
0.1546630859375,
0.61767578125,
0.1546630859375,
0.61767578125,
0.1666259765625,
0.5361328125,
0.1666259765625
],
[
0.7685546875,
0.1590576171875,
0.84716796875,
0.1590576171875,
0.84716796875,
0.173583984375,
0.7685546875,
0.173583984375
],
[
0.0882568359375,
0.1654052734375,
0.473876953125,
0.1654052734375,
0.473876953125,
0.1787109375,
0.0882568359375,
0.1787109375
],
[
0.54150390625,
0.1685791015625,
0.615234375,
0.1685791015625,
0.615234375,
0.179931640625,
0.54150390625,
0.179931640625
],
[
0.62353515625,
0.17236328125,
0.63232421875,
0.17236328125,
0.63232421875,
0.1767578125,
0.62353515625,
0.1767578125
],
[
0.0882568359375,
0.179931640625,
0.473876953125,
0.179931640625,
0.473876953125,
0.1932373046875,
0.0882568359375,
0.1932373046875
],
[
0.79150390625,
0.1767578125,
0.79736328125,
0.1767578125,
0.79736328125,
0.1817626953125,
0.79150390625,
0.1817626953125
],
[
0.0882568359375,
0.1932373046875,
0.4755859375,
0.1944580078125,
0.4755859375,
0.208984375,
0.0882568359375,
0.2083740234375
],
[
0.54150390625,
0.199462890625,
0.6435546875,
0.199462890625,
0.6435546875,
0.210205078125,
0.54150390625,
0.210205078125
],
[
0.65625,
0.1976318359375,
0.6640625,
0.1976318359375,
0.6640625,
0.2020263671875,
0.65625,
0.2020263671875
],
[
0.0882568359375,
0.210205078125,
0.473876953125,
0.210205078125,
0.473876953125,
0.2235107421875,
0.0882568359375,
0.2235107421875
],
[
0.54150390625,
0.2139892578125,
0.654296875,
0.2139892578125,
0.654296875,
0.2247314453125,
0.54150390625,
0.2247314453125
],
[
0.08905029296875,
0.2266845703125,
0.47216796875,
0.2266845703125,
0.47216796875,
0.2374267578125,
0.08905029296875,
0.2374267578125
],
[
0.67626953125,
0.238037109375,
0.70751953125,
0.238037109375,
0.70751953125,
0.25,
0.67626953125,
0.25
],
[
0.74609375,
0.2392578125,
0.775390625,
0.2392578125,
0.775390625,
0.2474365234375,
0.74609375,
0.2474365234375
],
[
0.0882568359375,
0.255126953125,
0.4755859375,
0.255126953125,
0.4755859375,
0.268310546875,
0.0882568359375,
0.268310546875
],
[
0.0882568359375,
0.240478515625,
0.473876953125,
0.240478515625,
0.473876953125,
0.25390625,
0.0882568359375,
0.25390625
],
[
0.08416748046875,
0.264404296875,
0.476318359375,
0.264404296875,
0.476318359375,
0.2890625,
0.08416748046875,
0.2890625
],
[
0.5,
0.267578125,
0.88623046875,
0.267578125,
0.88623046875,
0.2822265625,
0.5,
0.2822265625
],
[
0.0882568359375,
0.285400390625,
0.4755859375,
0.285400390625,
0.4755859375,
0.298583984375,
0.0882568359375,
0.298583984375
],
[
0.498291015625,
0.281005859375,
0.8857421875,
0.281005859375,
0.8857421875,
0.294189453125,
0.498291015625,
0.294189453125
],
[
0.08905029296875,
0.300537109375,
0.473876953125,
0.300537109375,
0.473876953125,
0.314453125,
0.08905029296875,
0.314453125
],
[
0.5,
0.29541015625,
0.88623046875,
0.29541015625,
0.88623046875,
0.309326171875,
0.5,
0.309326171875
],
[
0.5,
0.309326171875,
0.88623046875,
0.309326171875,
0.88623046875,
0.322509765625,
0.5,
0.322509765625
],
[
0.0906982421875,
0.31640625,
0.473876953125,
0.31640625,
0.473876953125,
0.330078125,
0.0906982421875,
0.330078125
],
[
0.5009765625,
0.324462890625,
0.8857421875,
0.324462890625,
0.8857421875,
0.335205078125,
0.5009765625,
0.335205078125
],
[
0.0882568359375,
0.330810546875,
0.473876953125,
0.330810546875,
0.473876953125,
0.3447265625,
0.0882568359375,
0.3447265625
],
[
0.5,
0.337158203125,
0.71142578125,
0.337158203125,
0.71142578125,
0.35107421875,
0.5,
0.35107421875
],
[
0.0882568359375,
0.3466796875,
0.4755859375,
0.3466796875,
0.4755859375,
0.35986328125,
0.0882568359375,
0.35986328125
],
[
0.0882568359375,
0.361083984375,
0.473876953125,
0.361083984375,
0.473876953125,
0.375,
0.0882568359375,
0.375
],
[
0.0882568359375,
0.376953125,
0.473876953125,
0.376953125,
0.473876953125,
0.39013671875,
0.0882568359375,
0.39013671875
],
[
0.5,
0.37890625,
0.8857421875,
0.37890625,
0.8857421875,
0.392578125,
0.5,
0.392578125
],
[
0.0882568359375,
0.391357421875,
0.473876953125,
0.391357421875,
0.473876953125,
0.40478515625,
0.0882568359375,
0.40478515625
],
[
0.5,
0.39453125,
0.8857421875,
0.393310546875,
0.8857421875,
0.4072265625,
0.5,
0.40771484375
],
[
0.0882568359375,
0.406005859375,
0.4755859375,
0.406005859375,
0.4755859375,
0.419189453125,
0.0882568359375,
0.419189453125
],
[
0.5,
0.4091796875,
0.8857421875,
0.4091796875,
0.8857421875,
0.42236328125,
0.5,
0.42236328125
],
[
0.08905029296875,
0.4287109375,
0.473876953125,
0.4287109375,
0.473876953125,
0.442626953125,
0.08905029296875,
0.442626953125
],
[
0.5009765625,
0.423583984375,
0.88623046875,
0.423583984375,
0.88623046875,
0.4375,
0.5009765625,
0.4375
],
[
0.5,
0.439453125,
0.88623046875,
0.439453125,
0.88623046875,
0.45263671875,
0.5,
0.45263671875
],
[
0.08905029296875,
0.44580078125,
0.47216796875,
0.44580078125,
0.47216796875,
0.455810546875,
0.08905029296875,
0.455810546875
],
[
0.0882568359375,
0.458984375,
0.473876953125,
0.458984375,
0.473876953125,
0.473388671875,
0.0882568359375,
0.473388671875
],
[
0.5,
0.453857421875,
0.88623046875,
0.453857421875,
0.88623046875,
0.46728515625,
0.5,
0.46728515625
],
[
0.5,
0.468505859375,
0.8857421875,
0.468505859375,
0.8857421875,
0.481689453125,
0.5,
0.481689453125
],
[
0.08905029296875,
0.47607421875,
0.47216796875,
0.47607421875,
0.47216796875,
0.486083984375,
0.08905029296875,
0.486083984375
],
[
0.0882568359375,
0.4892578125,
0.473876953125,
0.488037109375,
0.473876953125,
0.50244140625,
0.0882568359375,
0.50390625
],
[
0.5,
0.484130859375,
0.88623046875,
0.484130859375,
0.88623046875,
0.49755859375,
0.5,
0.49755859375
],
[
0.08905029296875,
0.5048828125,
0.473876953125,
0.5048828125,
0.473876953125,
0.51806640625,
0.08905029296875,
0.51806640625
],
[
0.5009765625,
0.50048828125,
0.8857421875,
0.50048828125,
0.8857421875,
0.51123046875,
0.5009765625,
0.51123046875
],
[
0.08905029296875,
0.51953125,
0.473876953125,
0.51953125,
0.473876953125,
0.53271484375,
0.08905029296875,
0.53271484375
],
[
0.5,
0.5146484375,
0.88623046875,
0.5146484375,
0.88623046875,
0.52783203125,
0.5,
0.52783203125
],
[
0.0882568359375,
0.53515625,
0.473876953125,
0.53515625,
0.473876953125,
0.548828125,
0.0882568359375,
0.548828125
],
[
0.5,
0.52978515625,
0.88623046875,
0.52978515625,
0.88623046875,
0.54345703125,
0.5,
0.54345703125
],
[
0.08905029296875,
0.5498046875,
0.473876953125,
0.5498046875,
0.473876953125,
0.56298828125,
0.08905029296875,
0.56298828125
],
[
0.5,
0.54443359375,
0.88623046875,
0.54443359375,
0.88623046875,
0.55810546875,
0.5,
0.55810546875
],
[
0.0882568359375,
0.56494140625,
0.4755859375,
0.56494140625,
0.4755859375,
0.5791015625,
0.0882568359375,
0.5791015625
],
[
0.5009765625,
0.56005859375,
0.8857421875,
0.56005859375,
0.8857421875,
0.57373046875,
0.5009765625,
0.57373046875
],
[
0.496826171875,
0.5693359375,
0.890625,
0.5693359375,
0.890625,
0.59326171875,
0.496826171875,
0.59326171875
],
[
0.08660888671875,
0.5791015625,
0.4345703125,
0.5791015625,
0.4345703125,
0.59228515625,
0.08660888671875,
0.59228515625
],
[
0.5,
0.59228515625,
0.88623046875,
0.59228515625,
0.88623046875,
0.60693359375,
0.5,
0.60693359375
],
[
0.0882568359375,
0.60986328125,
0.41259765625,
0.60986328125,
0.41259765625,
0.62353515625,
0.0882568359375,
0.62353515625
],
[
0.5009765625,
0.609375,
0.8857421875,
0.609375,
0.8857421875,
0.619140625,
0.5009765625,
0.619140625
],
[
0.1168212890625,
0.62548828125,
0.339111328125,
0.626953125,
0.339111328125,
0.6416015625,
0.1168212890625,
0.64013671875
],
[
0.5,
0.62255859375,
0.88623046875,
0.62255859375,
0.88623046875,
0.63623046875,
0.5,
0.63623046875
],
[
0.5,
0.63818359375,
0.88623046875,
0.63818359375,
0.88623046875,
0.6513671875,
0.5,
0.6513671875
],
[
0.08905029296875,
0.6513671875,
0.473876953125,
0.6513671875,
0.473876953125,
0.662109375,
0.08905029296875,
0.662109375
],
[
0.498291015625,
0.65283203125,
0.8857421875,
0.6513671875,
0.8857421875,
0.66552734375,
0.498291015625,
0.66650390625
],
[
0.0882568359375,
0.66552734375,
0.4755859375,
0.66552734375,
0.4755859375,
0.6787109375,
0.0882568359375,
0.6787109375
],
[
0.5009765625,
0.66943359375,
0.8837890625,
0.66943359375,
0.8837890625,
0.6796875,
0.5009765625,
0.6796875
],
[
0.08905029296875,
0.681640625,
0.473876953125,
0.681640625,
0.473876953125,
0.6923828125,
0.08905029296875,
0.6923828125
],
[
0.5,
0.68310546875,
0.88623046875,
0.68310546875,
0.88623046875,
0.6962890625,
0.5,
0.6962890625
],
[
0.0882568359375,
0.69580078125,
0.4755859375,
0.69580078125,
0.4755859375,
0.708984375,
0.0882568359375,
0.708984375
],
[
0.5,
0.69775390625,
0.8857421875,
0.69775390625,
0.8857421875,
0.7109375,
0.5,
0.7109375
],
[
0.0882568359375,
0.7109375,
0.473876953125,
0.7109375,
0.473876953125,
0.724609375,
0.0882568359375,
0.724609375
],
[
0.5,
0.71337890625,
0.88623046875,
0.71337890625,
0.88623046875,
0.7265625,
0.5,
0.7265625
],
[
0.08905029296875,
0.7587890625,
0.88623046875,
0.7587890625,
0.88623046875,
0.77294921875,
0.08905029296875,
0.77294921875
],
[
0.0882568359375,
0.77294921875,
0.73388671875,
0.77294921875,
0.73388671875,
0.7861328125,
0.0882568359375,
0.7861328125
],
[
0.3447265625,
0.8037109375,
0.483642578125,
0.8037109375,
0.483642578125,
0.8173828125,
0.3447265625,
0.8173828125
],
[
0.67724609375,
0.8037109375,
0.80126953125,
0.80322265625,
0.80126953125,
0.81640625,
0.67724609375,
0.8173828125
],
[
0.101318359375,
0.8193359375,
0.1805419921875,
0.8212890625,
0.1805419921875,
0.8359375,
0.101318359375,
0.833984375
],
[
0.361083984375,
0.82275390625,
0.465576171875,
0.82275390625,
0.465576171875,
0.83544921875,
0.361083984375,
0.83544921875
],
[
0.611328125,
0.8212890625,
0.87109375,
0.8212890625,
0.87109375,
0.83544921875,
0.611328125,
0.83544921875
],
[
0.358642578125,
0.833984375,
0.468017578125,
0.83349609375,
0.468017578125,
0.84765625,
0.358642578125,
0.8486328125
],
[
0.640625,
0.83544921875,
0.8408203125,
0.83544921875,
0.8408203125,
0.8486328125,
0.640625,
0.8486328125
],
[
0.318603515625,
0.84765625,
0.50732421875,
0.84765625,
0.50732421875,
0.861328125,
0.318603515625,
0.861328125
],
[
0.6357421875,
0.84765625,
0.8447265625,
0.84765625,
0.8447265625,
0.861328125,
0.6357421875,
0.861328125
],
[
0.291748046875,
0.85986328125,
0.53369140625,
0.85986328125,
0.53369140625,
0.87353515625,
0.291748046875,
0.87353515625
],
[
0.10455322265625,
0.87890625,
0.21728515625,
0.87890625,
0.21728515625,
0.892578125,
0.10455322265625,
0.892578125
],
[
0.2296142578125,
0.87890625,
0.59619140625,
0.87890625,
0.59619140625,
0.892578125,
0.2296142578125,
0.892578125
],
[
0.65673828125,
0.87890625,
0.82275390625,
0.87890625,
0.82275390625,
0.892578125,
0.65673828125,
0.892578125
],
[
0.285888671875,
0.89306640625,
0.54345703125,
0.89306640625,
0.54345703125,
0.90380859375,
0.285888671875,
0.90380859375
],
[
0.6748046875,
0.89306640625,
0.80615234375,
0.89306640625,
0.80615234375,
0.90380859375,
0.6748046875,
0.90380859375
]
] |
[
"CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning",
"2. CURIOUS",
"statesfrom",
"the environment",
"2.1. A Modular Multi-Goal Architecture using",
"actions toperformin",
"",
"the environment",
"Universal Approximators",
"generated",
"",
"by the agent:",
"UVFA concatenates the goal of the agent with its current",
"module 1, goal",
"Q(st,at.g, md)",
"state to form the input of the policy and the value func-",
"91 = [911,912]",
"",
"tion implemented by deep neural networks (Schaul et al.,",
"",
"2015a). With cURIoUs, we propose a new encoding of",
"goal param.for module 1 (2)",
"",
"goals using modular representations. This enables to target",
"goal param. for module 2 (1D) (",
"a rich diversity of modular goals within a single network",
"Actor",
"Critic",
"goal space of module Mi, the current goal g is defined as",
"(modular multi-goal approach), see Fig. 2. Given GM, the",
"a vector of dimension |G]-I I M.l, where the gM.",
"Figure 2. Modular goal-parameterized actor-critic architec-",
"can have different dimensionalities. g is set to O0 everywhere",
"ture (M-UvFA). Toy example with 2 modules, parameterized",
"except in the indices corresponding to the current module",
"by gi (2D) and g2 (ID) respectively. Here, the agent is attempting",
"goal gi in module Mi, as specified by the one-hot module descrip-",
"M, where it is set to gi E GM. By masking the goal-inputs",
"tormd =(1,O).The actor (left)computesthe action at.The",
"corresponding to unconsidered modules, the corresponding",
"critic (right) computes the Q-value.",
"weights are frozen during backpropagation. In addition,",
"a module descriptor md of size N (one-hot encoding) en-",
"codes the current module. The overall input to the policy",
"on the other hand, agents have the capacity to select which",
"network is st, 9,mal, see Fig. 2. We call this modular",
"goal to target next. Because goals are not equivalent, the",
"goal-parameterized architecture Modular-UVFA (M-UVFA).",
"agent can benefit from intrinsic motivations towards LP",
"In Fig. 2, we can see the underlying learning architecture",
"(Schmidhuber, 1991; Kaplan & Oudeyer, 2004). This can",
"be useful: 1) when there are distracting goals on which the",
"(actor-critic).Theactorimplementstheactionpolicyand",
"maps the input [st, g, md] to the next action at. The action",
"agent cannot progress; 2) when some goals are already mas-",
"tered. This idea comes from the IMGEP framework and was",
"isthenconcatenatedtoacopyoftheactor'sinputtofeedthe",
"critic [st, g, ma, at]. The critic provides an estimate of the",
"used in Baranes & Oudeyer (2013) to guide goal selection",
"Q-value: Q(st, g, md, at). The critic and the actor are then",
"andinForestier&Oudeyer(2016)toguidemodule selec-",
"trained using DDPG (Lillicrap et al., 2015), although any",
"tion. The problem of selecting a module can be modeled",
"other off-policy learning method could be used (e.g. TD3",
"as a non-stationary multi-armed bandit (MAB), where the",
"(Fujimoto et al., 2018), or DQN for the discrete case (Mnih",
"value of each arm (module) is the current absolute LP. Learn-",
"et al., 2013)). More details about DDPG can be found in the",
"ing progress (LP) is defined as the derivative of the agent's",
"competence on a particularmodule: LPMCMwhre",
"supplementary document or in Lillicrap et al. (2015).",
"the competence CM, : t → Psuccess(t) is the probability of",
"2.2. Module and Goal Selection, Cross-Module",
"success at timet.Here,the agentfocuses its attention on",
"Learning, Cross-Goal Learning",
"modules for which it is making the largest absolute progress,",
"and pays little attention to modules that are already solved",
"In UVFA, HER and UNICORN, the next goal totarget is se-",
"or unsolvable, i.e. for which LP stays small. Using the ab-",
"lected at random (Schaul et al., 2015a; Andrychowicz et al.,",
"solutevalue of LP alsoleadstotheprioritization ofmodules",
"2017:Mankowitz et al.,2018).This is coherent with the",
"for which the agent is showing decreasing performances.",
"common view that the agent must comply with the desires",
"This helps to deal with forgetting: the agent reallocates",
"of an engineer and target the goal it is asked to target. Here",
"learning resources to the modules being forgotten, Sec. 4.3.",
"Table 1. Classification of multi-goal approaches. Underlined: Algorithms internally generating goals (IMGEP), (*) using LP-based",
"intrinsic motivations. Italic: Population-based algorithms (non-RL). Bold: Algorithms proposed in this paper.",
"n GOALS, n POLICIES",
"nGOALS,1POLICY",
"FLAT REPR.",
"GOAL-EXPERTS:",
"MULTI-GOAL: UVFA (Schaul et al., 2015a)",
"(Kaelbling,1993)",
"HER (Andrychowicz et al., 2017)",
"SAC-x (Riedmiller et al., 2018)",
"UNICORN (Mankowitz et al., 2018)",
"SAGG-RIAC*(Baranes & Oudeyer,2013)",
"MODULAR REPR.",
"MOD-GOAL-EXPERTS:MACOB*(Forestier&Oudeyer,2016)",
"MODULAR-MULTI-GOALS:",
"MULTI-GOALMODULE-EXPERTS:MG-ME",
"M-UVFA,CURIOUS*"
] |
{
"arxiv_id": "1810.06284",
"language": "en",
"page_idx": 2
}
|
||
1810.06284-en-3
|
1810.06284-en-3.jpg
|
CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning
Learning Progress Estimation. Since an autonomous agent is not externally provided its true competence or LP, it needs to approximate them for each module. To measure its competence, it uses some episodes (with $p_eval=0.1)$ to evaluate itself on random modules and targets without exploration noise. The results (success 1 or failure 0) of these rollouts are stored in competence queues $results^{(i)}$ for all $M_i.$ In a similar way as Forestier & Oudeyer (2016), the agent computes its subjective competence as
$$C_{M_i}(n_{eval}^{(i)})=\frac{1}{l}\sum_{j=0}^{l-1}results^{(i)}(n_{eval}^{(i)}-j),$$
where $n_eval^{(i)}$ is the number of self-evaluation rollouts performed by the agent in module $M_i.$ The subjective $LP_{M_i}$ after $n_eval^{(i)}$ self-evaluation rollouts is then computed as:
$$LP_{M_{i}}(n_{eval}^{(i)})=C_{M_{i}}(n_{eval}^{(i)})-C_{M_{i}}(n_{eval}^{(i)}-l).$$
Given the subjective LP measures, we tackle the multiarmed bandit problem by implementing a simple approach called proportional probability matching, with an additional $\epsilon$-greedy strategy for exploration. More precisely, we compute the $LP\textit{probabilities }p_{LP}( M_{i})$ as:
$$p_{LP}(M_{i})=\epsilon\times\frac{1}{N}+(1-\epsilon)\times\frac{|LP_{M_{i}}|}{\sum_{j=1}^{N}|LP_{M_{j}}|},$$
where $N$ is the number of modules. The ratio $\epsilon$ implements a mixture between random exploration of modules (left term) and exploitation through a biased selection/replay of modules (right term). The random exploration term enables sampling modules that do not show any LP (i.e. already solved, too hard, or at a plateau). This way, the agent can check that it stays competent on modules that are already learned, or can insist on modules that are currently too hard. Note that we use LP for two distinct purposes: 1) Before data collection, to select the module from which to draw the next goal to attempt in the environment; 2) Before training, to select the substitute module descriptor (module replay). Recall that, once transitions are sampled from the replay buffer, they can be modified (replayed) by substituting the original module descriptor (or goal) by a new one. The substitute module is the one the agent is going to learn about. When replaying a particular module more than others, the agent allocates more resources to that module. While the use of LP for module selection is not new (Forestier &Oudeyer, 2016), we are the first to consider its use for cross-module goal replay.
Module and Goal Selection. Before interacting with the environment, the agents selects the next goal to target by first sampling a module from $\mathcal{M}$ using $p_LP$, and second, sampling the goal uniformly from the corresponding goal space $\mathcal{G}_M_i.$
Cross-Module and Cross-Goal Learning. In an example with three modules, an agent computed $p_{LP}=$ [0.6,0.2,0.2]. The agent uses these probabilities to guide learning towards modules with high absolute LP. If the size of the minibatch is $N_{mb}$,the agent will sample $\lfloor N_{mb}\times0.6\rfloor$ transitions relevant to module 1, $\lfloor N_{mb}\times0.2\rfloor$ transitions relevant to module 2 etc. A transition that is relevant for module $M_i$ (e.g. Push module), means that it comes from an episode during which the corresponding outcome has changed (e.g. cube position). This sampling bias towards “eventful” transitions is similar to Energy-Based Prioritization (Zhao & Tresp, 2018) (see supp. doc.). In this minibatch, every transition has been sampled to train on a specific module (e.g. $m_d^*)$, although it could have been collected while targeting another module (e.g. $m_d).$ To perform this cross-module learning, we simply substitute the latter by the former. Goal substitution is then performed using hindsight, which means the goal $g$ of a transition is sometimes $(p=0.8)$ replaced by an outcome reached later in the same episode $g^*$ (Andrychowicz et al., 2017).
Internal Reward. After module descriptors and goals have been substituted, the agent computes an internal reward for each transition using a reward function parameterized by the new $m_d^*$ and goal $g^*.$ Thus it answers: What would have been my reward for experiencing this transition, if I were aiming at that imagined goal from that imagined module? The reward is non-negative (0) when the outcome satisfies the constraints described by the imagined module $m_{d}^{*}$, relative to the imagined $g^*;$ negative otherwise(-1). In a reaching module for instance (see Fig. 1), a positive reward is generated when the Euclidean distance between the 3D target (goal) and the gripper (outcome) falls below a precision parameter $\epsilon_{reach}$ (reward constraint associated to the reaching module).
## $2. 3. \textbf{Combining Modular- UVFA and Intrinsically}$ Motivated Goal Exploration
A schematic view of CURIOUS is given in Fig.8. The detailed algorithm is given in the supplementary document.
1. Module and goal selection. The agent selects module
$M_i$ and goal $g_i$ for the next rollout (blue), respectively sampled from the set of potential modules $\mathcal{M}$ using $p_{LP}$ (purple), and uniformly from the corresponding goal space $\mathcal{G}_M_i.$
2. Data collection. The agent interacts with the environ-
ment using its current M-UVFA policy (grey), collects
transitions and stores them in memory (red).
3. LP update. If it was a self-evaluation rollout, the agent
updates its measures of competence, LP and $p_{LP}$ given
the new result (success or failure, purple).
|
[
[
0.2296142578125,
0.056182861328125,
0.744140625,
0.05682373046875,
0.744140625,
0.0706787109375,
0.2296142578125,
0.0694580078125
],
[
0.08905029296875,
0.08587646484375,
0.473876953125,
0.08587646484375,
0.473876953125,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.5,
0.083984375,
0.88623046875,
0.085205078125,
0.88623046875,
0.0997314453125,
0.5,
0.0985107421875
],
[
0.0882568359375,
0.10162353515625,
0.4755859375,
0.10101318359375,
0.4755859375,
0.1142578125,
0.0882568359375,
0.11553955078125
],
[
0.5009765625,
0.10162353515625,
0.861328125,
0.10162353515625,
0.861328125,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.0882568359375,
0.11614990234375,
0.4755859375,
0.11614990234375,
0.4755859375,
0.1300048828125,
0.0882568359375,
0.1300048828125
],
[
0.50244140625,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.50244140625,
0.1300048828125
],
[
0.0882568359375,
0.1319580078125,
0.4755859375,
0.1319580078125,
0.4755859375,
0.146484375,
0.0882568359375,
0.146484375
],
[
0.5,
0.1331787109375,
0.8837890625,
0.1331787109375,
0.8837890625,
0.143310546875,
0.5,
0.143310546875
],
[
0.08660888671875,
0.1458740234375,
0.4755859375,
0.146484375,
0.4755859375,
0.1610107421875,
0.08660888671875,
0.160400390625
],
[
0.498291015625,
0.146484375,
0.88232421875,
0.146484375,
0.88232421875,
0.1610107421875,
0.498291015625,
0.1610107421875
],
[
0.0882568359375,
0.1622314453125,
0.4755859375,
0.1622314453125,
0.4755859375,
0.175537109375,
0.0882568359375,
0.175537109375
],
[
0.5,
0.1622314453125,
0.8857421875,
0.1622314453125,
0.8857421875,
0.175537109375,
0.5,
0.175537109375
],
[
0.08660888671875,
0.1767578125,
0.473876953125,
0.1749267578125,
0.473876953125,
0.189453125,
0.08660888671875,
0.1912841796875
],
[
0.498291015625,
0.1767578125,
0.8857421875,
0.1767578125,
0.8857421875,
0.1900634765625,
0.498291015625,
0.1900634765625
],
[
0.0882568359375,
0.1912841796875,
0.4755859375,
0.1912841796875,
0.4755859375,
0.2052001953125,
0.0882568359375,
0.2052001953125
],
[
0.498291015625,
0.1912841796875,
0.8857421875,
0.1912841796875,
0.8857421875,
0.2052001953125,
0.498291015625,
0.2052001953125
],
[
0.0882568359375,
0.20703125,
0.406005859375,
0.20703125,
0.406005859375,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.5009765625,
0.20703125,
0.8857421875,
0.20703125,
0.8857421875,
0.2203369140625,
0.5009765625,
0.2203369140625
],
[
0.5009765625,
0.2215576171875,
0.88623046875,
0.2215576171875,
0.88623046875,
0.236083984375,
0.5009765625,
0.236083984375
],
[
0.250732421875,
0.231689453125,
0.271240234375,
0.2298583984375,
0.27294921875,
0.238037109375,
0.25244140625,
0.240478515625
],
[
0.253173828125,
0.240478515625,
0.36767578125,
0.2374267578125,
0.36767578125,
0.255126953125,
0.253173828125,
0.25830078125
],
[
0.498291015625,
0.2374267578125,
0.88623046875,
0.2374267578125,
0.88623046875,
0.251953125,
0.498291015625,
0.251953125
],
[
0.375732421875,
0.242431640625,
0.423095703125,
0.242431640625,
0.423095703125,
0.2568359375,
0.375732421875,
0.2568359375
],
[
0.1396484375,
0.2412109375,
0.2115478515625,
0.2442626953125,
0.2099609375,
0.2587890625,
0.1380615234375,
0.255615234375
],
[
0.371826171875,
0.250732421875,
0.38232421875,
0.25,
0.38232421875,
0.25390625,
0.371826171875,
0.255126953125
],
[
0.5,
0.250732421875,
0.8857421875,
0.251953125,
0.8857421875,
0.266357421875,
0.5,
0.26513671875
],
[
0.250732421875,
0.261474609375,
0.273681640625,
0.2587890625,
0.275390625,
0.267578125,
0.25244140625,
0.26953125
],
[
0.5,
0.267578125,
0.8857421875,
0.267578125,
0.8857421875,
0.281005859375,
0.5,
0.281005859375
],
[
0.08416748046875,
0.27978515625,
0.476318359375,
0.27783203125,
0.476318359375,
0.298583984375,
0.08416748046875,
0.300537109375
],
[
0.5,
0.2822265625,
0.8837890625,
0.2822265625,
0.8837890625,
0.29541015625,
0.5,
0.29541015625
],
[
0.08660888671875,
0.29541015625,
0.47216796875,
0.296630859375,
0.47216796875,
0.314453125,
0.08660888671875,
0.313232421875
],
[
0.5009765625,
0.29736328125,
0.88623046875,
0.29736328125,
0.88623046875,
0.3125,
0.5009765625,
0.3125
],
[
0.0882568359375,
0.31494140625,
0.453369140625,
0.31494140625,
0.453369140625,
0.333251953125,
0.0882568359375,
0.333251953125
],
[
0.5009765625,
0.313232421875,
0.8857421875,
0.313232421875,
0.8857421875,
0.326904296875,
0.5009765625,
0.326904296875
],
[
0.5009765625,
0.32763671875,
0.8857421875,
0.32763671875,
0.8857421875,
0.341552734375,
0.5009765625,
0.341552734375
],
[
0.1217041015625,
0.3359375,
0.438720703125,
0.339111328125,
0.438720703125,
0.361083984375,
0.1217041015625,
0.35791015625
],
[
0.5009765625,
0.343505859375,
0.88623046875,
0.343505859375,
0.88623046875,
0.356689453125,
0.5009765625,
0.356689453125
],
[
0.5009765625,
0.35791015625,
0.88623046875,
0.35791015625,
0.88623046875,
0.371826171875,
0.5009765625,
0.371826171875
],
[
0.08905029296875,
0.365478515625,
0.47216796875,
0.365478515625,
0.47216796875,
0.37890625,
0.08905029296875,
0.37890625
],
[
0.5,
0.37255859375,
0.84228515625,
0.37255859375,
0.84228515625,
0.386962890625,
0.5,
0.386962890625
],
[
0.0882568359375,
0.380615234375,
0.473876953125,
0.380615234375,
0.473876953125,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.0882568359375,
0.396484375,
0.473876953125,
0.396484375,
0.473876953125,
0.410400390625,
0.0882568359375,
0.410400390625
],
[
0.5009765625,
0.404052734375,
0.8837890625,
0.404052734375,
0.8837890625,
0.417236328125,
0.5009765625,
0.417236328125
],
[
0.0882568359375,
0.410400390625,
0.473876953125,
0.410888671875,
0.473876953125,
0.425537109375,
0.0882568359375,
0.4248046875
],
[
0.5009765625,
0.419189453125,
0.8857421875,
0.419189453125,
0.8857421875,
0.43310546875,
0.5009765625,
0.43310546875
],
[
0.0882568359375,
0.4267578125,
0.34228515625,
0.4267578125,
0.34228515625,
0.43994140625,
0.0882568359375,
0.43994140625
],
[
0.5009765625,
0.43505859375,
0.8837890625,
0.43505859375,
0.8837890625,
0.448974609375,
0.5009765625,
0.448974609375
],
[
0.238525390625,
0.449462890625,
0.2484130859375,
0.45263671875,
0.2442626953125,
0.460205078125,
0.2344970703125,
0.45703125
],
[
0.361083984375,
0.447509765625,
0.41259765625,
0.447509765625,
0.41259765625,
0.46337890625,
0.361083984375,
0.46337890625
],
[
0.5009765625,
0.449462890625,
0.8857421875,
0.449462890625,
0.8857421875,
0.464111328125,
0.5009765625,
0.464111328125
],
[
0.1192626953125,
0.45703125,
0.2320556640625,
0.455810546875,
0.2320556640625,
0.47021484375,
0.1192626953125,
0.4716796875
],
[
0.256591796875,
0.458251953125,
0.334228515625,
0.458251953125,
0.334228515625,
0.4697265625,
0.256591796875,
0.4697265625
],
[
0.236083984375,
0.466552734375,
0.250732421875,
0.466552734375,
0.250732421875,
0.47607421875,
0.236083984375,
0.47607421875
],
[
0.336669921875,
0.466552734375,
0.439453125,
0.462158203125,
0.441162109375,
0.484130859375,
0.338134765625,
0.488037109375
],
[
0.346435546875,
0.46728515625,
0.369384765625,
0.46728515625,
0.369384765625,
0.4716796875,
0.346435546875,
0.4716796875
],
[
0.5009765625,
0.46533203125,
0.88623046875,
0.46533203125,
0.88623046875,
0.479248046875,
0.5009765625,
0.479248046875
],
[
0.5,
0.479736328125,
0.88623046875,
0.479736328125,
0.88623046875,
0.49365234375,
0.5,
0.49365234375
],
[
0.0882568359375,
0.4912109375,
0.473876953125,
0.4912109375,
0.473876953125,
0.5048828125,
0.0882568359375,
0.5048828125
],
[
0.498291015625,
0.494384765625,
0.8857421875,
0.494384765625,
0.8857421875,
0.50830078125,
0.498291015625,
0.50830078125
],
[
0.0882568359375,
0.5068359375,
0.4755859375,
0.5068359375,
0.4755859375,
0.52099609375,
0.0882568359375,
0.52099609375
],
[
0.5,
0.51025390625,
0.8857421875,
0.51025390625,
0.8857421875,
0.5234375,
0.5,
0.5234375
],
[
0.0882568359375,
0.521484375,
0.4755859375,
0.521484375,
0.4755859375,
0.53515625,
0.0882568359375,
0.53515625
],
[
0.5,
0.5244140625,
0.88623046875,
0.5244140625,
0.88623046875,
0.5390625,
0.5,
0.5390625
],
[
0.0882568359375,
0.537109375,
0.4755859375,
0.537109375,
0.4755859375,
0.55029296875,
0.0882568359375,
0.55029296875
],
[
0.5,
0.54052734375,
0.8857421875,
0.54052734375,
0.8857421875,
0.5537109375,
0.5,
0.5537109375
],
[
0.08905029296875,
0.5517578125,
0.473876953125,
0.5517578125,
0.473876953125,
0.56494140625,
0.08905029296875,
0.56494140625
],
[
0.496826171875,
0.5546875,
0.8857421875,
0.5546875,
0.8857421875,
0.568359375,
0.496826171875,
0.568359375
],
[
0.0882568359375,
0.5673828125,
0.473876953125,
0.5673828125,
0.473876953125,
0.58056640625,
0.0882568359375,
0.58056640625
],
[
0.498291015625,
0.5693359375,
0.8857421875,
0.5693359375,
0.8857421875,
0.58349609375,
0.498291015625,
0.58349609375
],
[
0.0906982421875,
0.583984375,
0.47216796875,
0.583984375,
0.47216796875,
0.5947265625,
0.0906982421875,
0.5947265625
],
[
0.5009765625,
0.5859375,
0.8857421875,
0.5859375,
0.8857421875,
0.60107421875,
0.5009765625,
0.60107421875
],
[
0.0882568359375,
0.59765625,
0.4755859375,
0.59765625,
0.4755859375,
0.611328125,
0.0882568359375,
0.611328125
],
[
0.5,
0.60107421875,
0.64453125,
0.60107421875,
0.64453125,
0.6142578125,
0.5,
0.6142578125
],
[
0.08905029296875,
0.619140625,
0.473876953125,
0.619140625,
0.473876953125,
0.63330078125,
0.08905029296875,
0.63330078125
],
[
0.08905029296875,
0.63623046875,
0.47216796875,
0.63623046875,
0.47216796875,
0.646484375,
0.08905029296875,
0.646484375
],
[
0.5009765625,
0.630859375,
0.84228515625,
0.630859375,
0.84228515625,
0.64599609375,
0.5009765625,
0.64599609375
],
[
0.08660888671875,
0.6484375,
0.4755859375,
0.65087890625,
0.4755859375,
0.6640625,
0.08660888671875,
0.662109375
],
[
0.52685546875,
0.64599609375,
0.7294921875,
0.646484375,
0.7294921875,
0.6611328125,
0.52685546875,
0.66015625
],
[
0.0882568359375,
0.6640625,
0.473876953125,
0.6640625,
0.473876953125,
0.67822265625,
0.0882568359375,
0.67822265625
],
[
0.5009765625,
0.67041015625,
0.8857421875,
0.67041015625,
0.8857421875,
0.6845703125,
0.5009765625,
0.6845703125
],
[
0.0882568359375,
0.68115234375,
0.473876953125,
0.68115234375,
0.473876953125,
0.6943359375,
0.0882568359375,
0.6943359375
],
[
0.5,
0.68603515625,
0.88623046875,
0.68603515625,
0.88623046875,
0.69970703125,
0.5,
0.69970703125
],
[
0.0882568359375,
0.69580078125,
0.473876953125,
0.69580078125,
0.473876953125,
0.708984375,
0.0882568359375,
0.708984375
],
[
0.0882568359375,
0.71044921875,
0.473876953125,
0.71044921875,
0.473876953125,
0.7236328125,
0.0882568359375,
0.7236328125
],
[
0.5146484375,
0.7197265625,
0.8857421875,
0.7197265625,
0.8857421875,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.0882568359375,
0.72607421875,
0.473876953125,
0.72607421875,
0.473876953125,
0.7392578125,
0.0882568359375,
0.7392578125
],
[
0.08660888671875,
0.7392578125,
0.473876953125,
0.74072265625,
0.473876953125,
0.7548828125,
0.08660888671875,
0.75390625
],
[
0.53515625,
0.73486328125,
0.8857421875,
0.73486328125,
0.8857421875,
0.74853515625,
0.53515625,
0.74853515625
],
[
0.0882568359375,
0.755859375,
0.473876953125,
0.755859375,
0.473876953125,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.53369140625,
0.75048828125,
0.8857421875,
0.75048828125,
0.8857421875,
0.7646484375,
0.53369140625,
0.7646484375
],
[
0.08905029296875,
0.771484375,
0.471435546875,
0.771484375,
0.471435546875,
0.7822265625,
0.08905029296875,
0.7822265625
],
[
0.53173828125,
0.7666015625,
0.8857421875,
0.7666015625,
0.8857421875,
0.77978515625,
0.53173828125,
0.77978515625
],
[
0.0882568359375,
0.78515625,
0.473876953125,
0.78515625,
0.473876953125,
0.798828125,
0.0882568359375,
0.798828125
],
[
0.53125,
0.77978515625,
0.6396484375,
0.78076171875,
0.6396484375,
0.79541015625,
0.53125,
0.79443359375
],
[
0.08905029296875,
0.80322265625,
0.1650390625,
0.80322265625,
0.1650390625,
0.814453125,
0.08905029296875,
0.814453125
],
[
0.5107421875,
0.8037109375,
0.88623046875,
0.8046875,
0.88623046875,
0.8193359375,
0.5107421875,
0.818359375
],
[
0.53125,
0.8212890625,
0.88623046875,
0.8212890625,
0.88623046875,
0.83544921875,
0.53125,
0.83544921875
],
[
0.0882568359375,
0.83056640625,
0.473876953125,
0.83056640625,
0.473876953125,
0.8447265625,
0.0882568359375,
0.8447265625
],
[
0.53125,
0.83544921875,
0.82275390625,
0.8359375,
0.82275390625,
0.85107421875,
0.53125,
0.849609375
],
[
0.0882568359375,
0.8466796875,
0.473876953125,
0.8466796875,
0.473876953125,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.0849609375,
0.85791015625,
0.4755859375,
0.85986328125,
0.4755859375,
0.87744140625,
0.0849609375,
0.87548828125
],
[
0.51318359375,
0.861328125,
0.8857421875,
0.861328125,
0.8857421875,
0.875,
0.51318359375,
0.875
],
[
0.0882568359375,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.53173828125,
0.87548828125,
0.8857421875,
0.876953125,
0.8857421875,
0.8916015625,
0.53173828125,
0.89013671875
],
[
0.0882568359375,
0.892578125,
0.1641845703125,
0.892578125,
0.1641845703125,
0.9072265625,
0.0882568359375,
0.9072265625
],
[
0.53173828125,
0.892578125,
0.80810546875,
0.892578125,
0.80810546875,
0.90576171875,
0.53173828125,
0.90576171875
]
] |
[
"CURIOUS: IntrinsicallyMotivated Modular Multi-Goal Reinforcement Learning",
"Learning Progress Estimation. Since an autonomous",
"Cross-Module and Cross-Goal Learning. In an exam-",
"agent is notexternallyprovided its true competenceor Lp,",
"ple with three modules, an agent computed PLP",
"it needs to approximate them for each module. To measure",
"[0.6, 0.2, 0.2]. The agent uses these probabilities to guide",
"its competence, it uses some episodes (with peval = 0.1)",
"learningtowardsmoduleswithhighabsoluteLP.Ifthesize",
"to evaluate itself on random modules and targets without",
"of the minibatch is Nmb, the agent will sample [Nmb × 0.6]",
"exploration noise. The results (success 1 or failure O) of",
"transitions relevant to module 1, INmb × 0.2] transitions",
"these rollouts are stored in competence queues results(i)",
"relevant to module 2 etc. A transition that is relevant for",
"for all Mi. In a similar way as Forestier & Oudeyer (2016),",
"module M, (e.g. Push module), means that it comes from",
"the agent computes its subjective competence as",
"an episode during which the corresponding outcome has",
"changed (e.g. cube position). This sampling bias towards",
"1-]",
"Eresults(i)(n(",
"“eventful\" transitions is similar to Energy-Based Prioriti-",
"-)",
"CM(neval",
"",
"zation (Zhao & Tresp, 2018) (see supp.doc.). In this",
"=0",
"minibatch, every transition has been sampled to train on",
"where n(vat is the number of self-evaluation rollouts per-",
"a specific module (e.g. m*), although it could have been",
"formed by the agent in module M,. The subjective LPM,",
"collected while targeting another module (e.g. md). To",
"after nevat self-evaluation rollouts is then computed as:",
"perform this cross-module learning, we simply substitute",
"the latter by the former. Goal substitution is then performed",
"LPM,(neal) = CM (neval) - Cm,(neval - I).",
"using hindsight, which means the goal g of a transition is",
"sometimes (p =- 0.8) replaced by an outcome reached later",
"Given the subjective LP measures, we tackle the multi",
"in the same episode g* (Andrychowicz et al., 2017).",
"armed bandit problem by implementing a simple approach",
"called proportional probability matching, with an additional",
"Internal Reward.After module descriptors and goals",
"-greedy strategy for exploration. More precisely, we com-",
"have been substituted, the agent computes an internal reward",
"pute the LP probabilities PLP(M) as:",
"for each transition using a reward function parameterized",
"1",
"[LPM,]",
"by the new ma and goal g*. Thus it answers: What would",
"PLP(M) =EX",
"+(1-) x",
"N",
"E,-,[LPM,l",
"N",
"have been my reward for experiencing this transition, if",
"I were aiming at that imagined goal from that imagined",
"where N is the number of modules. The ratio e implements",
"module? The reward is non-negative (O) when the outcome",
"a mixture between random exploration of modules (left",
"satisfies the constraints described by the imagined module",
"term) and exploitation through a biased selection/replay of",
"m*, relative to the imagined g*; negative otherwise (-1).",
"modules (right term). The random exploration term enables",
"In a reaching module for instance (see Fig. 1), a positive",
"sampling modules that do not show any LP (i.e. already",
"reward is generated when the Euclidean distance between",
"solved, too hard, or at a plateau). This way, the agent can",
"the 3D target (goal) and the gripper (outcome) falls below a",
"check thatit stays competent onmodulesthatare already",
"precision parameter Ereach (reward constraint associated to",
"learned, or can insist on modules that are currently too hard.",
"the reaching module).",
"Note that we use LP for two distinct purposes: 1) Before",
"datacollection,toselectthemodulefromwhichtodrawthe",
"2.3. Combining Modular-UVFA and Intrinsically",
"next goal to attempt in the environment; 2) Before training,",
"Motivated Goal Exploration",
"to select the substitute module descriptor (module replay).",
"A schematic view of cURIoUs is given in Fig. 8. The",
"Recall that, once transitions are sampled from the replay",
"detailed algorithm is given in the supplementary document.",
"buffer, they can be modified (replayed) by substituting the",
"original module descriptor (or goal) by a new one. The",
"1. Module and goal selection. The agent selects module",
"substitute module is the one the agent is going to learn about.",
"When replaying a particular module more than others, the",
"M, and goal gi for the next rollout (blue), respectively",
"agent allocates more resources to that module. While the use",
"sampled from the set of potential modules M using",
"of LPformodule selectionis notnew(Forestier&Oudever",
"PLp (purple), and uniformly from the corresponding",
"2016), we are the first to consider its use for cross-module",
"goal space GM,.",
"goalreplay",
"2. Data collection. The agent interacts with the environ-",
"ment using its current M-UVFA policy (grey), collects",
"Module and Goal Selection.Before interacting with the",
"transitions and stores them in memory (red).",
"environment, the agents selects the next goal to target by",
"first sampling a module from M using pLp, and second,",
"3. LP update. If it was a self-evaluation rollout, the agent",
"sampling the goal uniformly from the corresponding goal",
"updates its measures of competence, LP and pLP given",
"space GM,.",
"the new result (success or failure, purple)."
] |
{
"arxiv_id": "1810.06284",
"language": "en",
"page_idx": 3
}
|
||
1810.06284-en-4
|
1810.06284-en-4.jpg
|
CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning
4. Module and goal substitution. The agent decides on
which modules and goals to train. To update the policy and critic, the algorithm first samples a minibatch from the replay buffers (red) using $p_{LP}$ and implements module and goal substitutions to perform cross-module and cross-goal learning (orange), see Sec. 2.2.
5. Internal reward. The agent computes its reward $r$
for each transition, using $R_{M,g}$ parameterized by the
substitute module $m_d^*$ and goal $g^*$ (brown).
6. RL updates. The agent updates its policy and value
function with DDPG using the modified minibatch
(green).
<FigureHere>
Figure 3. Schematic view of cuRIous.
# 2.4. Evaluation Methodology
The performance of the agents are evaluated offline in terms of success rates over sets of achievable goals (defined by the experimenter). Every point of a learning curve represents the success rate over 95 offline evaluation rollouts (5×19 actors), using random achievable goals. Evaluation is completely independent from training, i.e. agents cannot train on evaluation transitions. We use the non-parametric onetail Mann-Whitney U-test with confdence level $\alpha=0.01$ for all comparisons. More details and justifications can be found in the supplementary document.
## 3. A Modular Goal Environment
Modular Goal Fetch Arm is a new simulated environment adapted from the OpenAI Gym suite (Brockman et al., 2016). The agent is embodied by a robotic arm facing 2 cubes randomly positioned on a table. The agent controls the position of its gripper and the gripper opening (4D). It can target a diverse set of modular goals: $( M_1) \textit{Reach a 3D}$
target with the gripper; $( M_2) \textit{Push cube 1 onto a 2D target}$ on the table; $(M_{3})$ Pick and Place cube 1 on a 3D target; $( M_4) \textit{Stack cube 1 over cube 2. Additional Push modules}$ concerning additional out-of-reach and moving cubes can be defined (impossible, distracting goals). Further details can be found in the supplementary document.
## 4. Experiment and Results
In this section, we present ablative studies to assess the relative importance of: 1) the policy and value function architecture and 2) the use of intrinsically motivated module selection for practice and replay. We call M-UVFA the algorithm using a modular goal-parameterized policy and random module choices, while the intrinsically motivated version is called CURIOUS. We do not investigate the efficiency of HER or the efficiency of the sampling bias towards interesting transitions as they were already studied in Andrychowicz et al. (2017); Plappert et al. (2018) and Zhao& Tresp (2018) respectively. For fair comparisons, we apply both mechanisms to all the tested algorithms.
$4. 1. \textbf{Impact of Policy and Value Function Architecture}$
Experiments. In this section, we investigate the impact of using an M-UVFA architecture for the policy and value function. The module-set is composed of four achievable modules and four distracting modules. We test this algorithm against two baselines:
1. A flat multi-goal architecture (HER). This algorithm
does not represent goals in a modular fashion but in a linear way. The corresponding goal is selected uniformly inside $\mathcal{G}$, a holistic goal space such that $\mathcal{G}=\prod_{i=1}^{N}\mathcal{G}_{M_{i}}.$ To generate a reward, the agent needs to satisfy the constraints described by all the modules at once. This goal-parameterized architecture is equivalent to UVFA, which makes the algorithm equivalent $\textbf{to HER + DDPG}.$
2. A multi-goal module-experts architecture (MG-ME)
where an expert multi-goal policy is trained for each of the $N$ modules. Each policy is trained one epoch every $N$ on its designated module and shares the collected transitions with other experts. When evaluated on a particular module, the algorithm uses the corresponding module-expert.
Results. Fig.5 shows the evolution of the average success rate computed over achievable goals for M-UVFA and the two baselines described above. The learning curve of HER stays flat. This can be easily understood as none of the goals expressed in the complete goal space $\mathcal{G}$ corresponds to a real situation (e.g. the agent cannot reach a 3D target with its gripper while placing a cube at another). The agent cannot
|
[
[
0.2303466796875,
0.05682373046875,
0.744140625,
0.05682373046875,
0.744140625,
0.0706787109375,
0.2303466796875,
0.0706787109375
],
[
0.101318359375,
0.08587646484375,
0.4755859375,
0.08587646484375,
0.4755859375,
0.0997314453125,
0.101318359375,
0.0997314453125
],
[
0.498291015625,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.498291015625,
0.0997314453125
],
[
0.12091064453125,
0.10162353515625,
0.473876953125,
0.10162353515625,
0.473876953125,
0.11553955078125,
0.12091064453125,
0.11553955078125
],
[
0.496826171875,
0.10101318359375,
0.88623046875,
0.10162353515625,
0.88623046875,
0.11553955078125,
0.496826171875,
0.1142578125
],
[
0.12091064453125,
0.117431640625,
0.473876953125,
0.117431640625,
0.473876953125,
0.1307373046875,
0.12091064453125,
0.1307373046875
],
[
0.5,
0.11871337890625,
0.8837890625,
0.11871337890625,
0.8837890625,
0.1287841796875,
0.5,
0.1287841796875
],
[
0.12091064453125,
0.1319580078125,
0.473876953125,
0.1319580078125,
0.473876953125,
0.146484375,
0.12091064453125,
0.146484375
],
[
0.5,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.5,
0.1458740234375
],
[
0.12091064453125,
0.146484375,
0.473876953125,
0.146484375,
0.473876953125,
0.1610107421875,
0.12091064453125,
0.1610107421875
],
[
0.498291015625,
0.1458740234375,
0.88623046875,
0.146484375,
0.88623046875,
0.1610107421875,
0.498291015625,
0.160400390625
],
[
0.12091064453125,
0.1622314453125,
0.423095703125,
0.1622314453125,
0.423095703125,
0.1767578125,
0.12091064453125,
0.1767578125
],
[
0.5,
0.1622314453125,
0.7998046875,
0.1622314453125,
0.7998046875,
0.1767578125,
0.5,
0.1767578125
],
[
0.1004638671875,
0.1881103515625,
0.473876953125,
0.1881103515625,
0.473876953125,
0.20263671875,
0.1004638671875,
0.20263671875
],
[
0.50244140625,
0.1962890625,
0.7236328125,
0.1962890625,
0.7236328125,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.12091064453125,
0.203857421875,
0.473876953125,
0.203857421875,
0.473876953125,
0.2183837890625,
0.12091064453125,
0.2183837890625
],
[
0.12091064453125,
0.2197265625,
0.404296875,
0.2197265625,
0.404296875,
0.23291015625,
0.12091064453125,
0.23291015625
],
[
0.5009765625,
0.2215576171875,
0.88623046875,
0.2215576171875,
0.88623046875,
0.236083984375,
0.5009765625,
0.236083984375
],
[
0.498291015625,
0.2374267578125,
0.88623046875,
0.2374267578125,
0.88623046875,
0.250732421875,
0.498291015625,
0.250732421875
],
[
0.098876953125,
0.24365234375,
0.473876953125,
0.2442626953125,
0.473876953125,
0.2587890625,
0.098876953125,
0.25830078125
],
[
0.50244140625,
0.25390625,
0.8837890625,
0.25390625,
0.8837890625,
0.264404296875,
0.50244140625,
0.264404296875
],
[
0.1217041015625,
0.261962890625,
0.47216796875,
0.261962890625,
0.47216796875,
0.272705078125,
0.1217041015625,
0.272705078125
],
[
0.5009765625,
0.267578125,
0.8857421875,
0.267578125,
0.8857421875,
0.281005859375,
0.5009765625,
0.281005859375
],
[
0.1192626953125,
0.27587890625,
0.1732177734375,
0.2734375,
0.1748046875,
0.28857421875,
0.12091064453125,
0.290283203125
],
[
0.5009765625,
0.28271484375,
0.88623046875,
0.28271484375,
0.88623046875,
0.296630859375,
0.5009765625,
0.296630859375
],
[
0.5009765625,
0.29736328125,
0.8857421875,
0.29736328125,
0.8857421875,
0.311279296875,
0.5009765625,
0.311279296875
],
[
0.5009765625,
0.313232421875,
0.88623046875,
0.313232421875,
0.88623046875,
0.326904296875,
0.5009765625,
0.326904296875
],
[
0.25244140625,
0.3212890625,
0.29736328125,
0.3212890625,
0.29736328125,
0.33203125,
0.25244140625,
0.33203125
],
[
0.386474609375,
0.3193359375,
0.4150390625,
0.3193359375,
0.4150390625,
0.33203125,
0.386474609375,
0.33203125
],
[
0.3447265625,
0.335205078125,
0.378173828125,
0.335205078125,
0.378173828125,
0.34228515625,
0.3447265625,
0.34228515625
],
[
0.50244140625,
0.330078125,
0.8857421875,
0.330078125,
0.8857421875,
0.34033203125,
0.50244140625,
0.34033203125
],
[
0.3349609375,
0.33837890625,
0.346435546875,
0.33837890625,
0.346435546875,
0.341552734375,
0.3349609375,
0.341552734375
],
[
0.330078125,
0.3447265625,
0.393798828125,
0.3447265625,
0.393798828125,
0.35595703125,
0.330078125,
0.35595703125
],
[
0.3896484375,
0.34521484375,
0.414306640625,
0.34521484375,
0.414306640625,
0.356689453125,
0.3896484375,
0.356689453125
],
[
0.5009765625,
0.34228515625,
0.8857421875,
0.34228515625,
0.8857421875,
0.35595703125,
0.5009765625,
0.35595703125
],
[
0.2459716796875,
0.352783203125,
0.28515625,
0.352783203125,
0.28515625,
0.356689453125,
0.2459716796875,
0.356689453125
],
[
0.160888671875,
0.3466796875,
0.2181396484375,
0.348388671875,
0.21728515625,
0.363037109375,
0.16015625,
0.361083984375
],
[
0.23779296875,
0.35986328125,
0.289306640625,
0.35986328125,
0.289306640625,
0.367431640625,
0.23779296875,
0.367431640625
],
[
0.5009765625,
0.35791015625,
0.88623046875,
0.35791015625,
0.88623046875,
0.37255859375,
0.5009765625,
0.37255859375
],
[
0.25244140625,
0.371826171875,
0.27783203125,
0.371826171875,
0.27783203125,
0.37890625,
0.25244140625,
0.37890625
],
[
0.330810546875,
0.373779296875,
0.34228515625,
0.373779296875,
0.34228515625,
0.37744140625,
0.330810546875,
0.37744140625
],
[
0.5009765625,
0.373779296875,
0.8857421875,
0.373779296875,
0.8857421875,
0.386962890625,
0.5009765625,
0.386962890625
],
[
0.34716796875,
0.375,
0.361083984375,
0.375,
0.361083984375,
0.37890625,
0.34716796875,
0.37890625
],
[
0.1641845703125,
0.375,
0.193603515625,
0.375,
0.193603515625,
0.3818359375,
0.1641845703125,
0.3818359375
],
[
0.36279296875,
0.375732421875,
0.38232421875,
0.375732421875,
0.38232421875,
0.380126953125,
0.36279296875,
0.380126953125
],
[
0.1502685546875,
0.3837890625,
0.20751953125,
0.3837890625,
0.20751953125,
0.39453125,
0.1502685546875,
0.39453125
],
[
0.285888671875,
0.3837890625,
0.324462890625,
0.3837890625,
0.324462890625,
0.391357421875,
0.285888671875,
0.391357421875
],
[
0.5,
0.38818359375,
0.83642578125,
0.386962890625,
0.83642578125,
0.401611328125,
0.5,
0.40283203125
],
[
0.26708984375,
0.397705078125,
0.310546875,
0.397705078125,
0.310546875,
0.40478515625,
0.26708984375,
0.40478515625
],
[
0.328369140625,
0.399658203125,
0.35302734375,
0.399658203125,
0.35302734375,
0.404052734375,
0.328369140625,
0.404052734375
],
[
0.236083984375,
0.40771484375,
0.2459716796875,
0.40771484375,
0.2459716796875,
0.415283203125,
0.236083984375,
0.415283203125
],
[
0.158447265625,
0.40087890625,
0.1993408203125,
0.40087890625,
0.1993408203125,
0.410888671875,
0.158447265625,
0.410888671875
],
[
0.30224609375,
0.410400390625,
0.31787109375,
0.410400390625,
0.31787109375,
0.4140625,
0.30224609375,
0.4140625
],
[
0.363525390625,
0.410400390625,
0.375732421875,
0.40478515625,
0.3798828125,
0.410888671875,
0.36767578125,
0.416748046875
],
[
0.5009765625,
0.41845703125,
0.87744140625,
0.419189453125,
0.87744140625,
0.434326171875,
0.5009765625,
0.43310546875
],
[
0.160888671875,
0.412353515625,
0.193603515625,
0.412353515625,
0.193603515625,
0.419189453125,
0.160888671875,
0.419189453125
],
[
0.23779296875,
0.421630859375,
0.37744140625,
0.421630859375,
0.37744140625,
0.431884765625,
0.23779296875,
0.431884765625
],
[
0.156005859375,
0.4287109375,
0.2115478515625,
0.4287109375,
0.2115478515625,
0.43994140625,
0.156005859375,
0.43994140625
],
[
0.223876953125,
0.438232421875,
0.268798828125,
0.438232421875,
0.268798828125,
0.448974609375,
0.223876953125,
0.448974609375
],
[
0.28515625,
0.44140625,
0.338134765625,
0.442626953125,
0.338134765625,
0.449462890625,
0.28515625,
0.448974609375
],
[
0.359375,
0.438232421875,
0.406005859375,
0.438232421875,
0.406005859375,
0.448974609375,
0.359375,
0.448974609375
],
[
0.5009765625,
0.443115234375,
0.8857421875,
0.443115234375,
0.8857421875,
0.45703125,
0.5009765625,
0.45703125
],
[
0.2132568359375,
0.451904296875,
0.289306640625,
0.451904296875,
0.289306640625,
0.462158203125,
0.2132568359375,
0.462158203125
],
[
0.160888671875,
0.442626953125,
0.2017822265625,
0.442626953125,
0.2017822265625,
0.45703125,
0.160888671875,
0.45703125
],
[
0.29736328125,
0.455810546875,
0.336669921875,
0.455810546875,
0.336669921875,
0.460205078125,
0.29736328125,
0.460205078125
],
[
0.343017578125,
0.45263671875,
0.408447265625,
0.45263671875,
0.408447265625,
0.46337890625,
0.343017578125,
0.46337890625
],
[
0.1871337890625,
0.46337890625,
0.2459716796875,
0.46337890625,
0.2459716796875,
0.473388671875,
0.1871337890625,
0.473388671875
],
[
0.5009765625,
0.458984375,
0.88623046875,
0.458984375,
0.88623046875,
0.472900390625,
0.5009765625,
0.472900390625
],
[
0.3447265625,
0.47607421875,
0.393798828125,
0.47607421875,
0.393798828125,
0.486083984375,
0.3447265625,
0.486083984375
],
[
0.5009765625,
0.473388671875,
0.8857421875,
0.473388671875,
0.8857421875,
0.4873046875,
0.5009765625,
0.4873046875
],
[
0.176513671875,
0.4873046875,
0.205078125,
0.4873046875,
0.205078125,
0.494384765625,
0.176513671875,
0.494384765625
],
[
0.325927734375,
0.479248046875,
0.338134765625,
0.479248046875,
0.338134765625,
0.48486328125,
0.325927734375,
0.48486328125
],
[
0.34228515625,
0.4892578125,
0.39453125,
0.4873046875,
0.39453125,
0.49755859375,
0.34228515625,
0.5
],
[
0.5009765625,
0.4892578125,
0.88623046875,
0.4892578125,
0.88623046875,
0.50244140625,
0.5009765625,
0.50244140625
],
[
0.27783203125,
0.505859375,
0.3056640625,
0.505859375,
0.3056640625,
0.51318359375,
0.27783203125,
0.51318359375
],
[
0.5,
0.5048828125,
0.68896484375,
0.5048828125,
0.68896484375,
0.51806640625,
0.5,
0.51806640625
],
[
0.17724609375,
0.517578125,
0.1895751953125,
0.517578125,
0.1895751953125,
0.521484375,
0.17724609375,
0.521484375
],
[
0.26708984375,
0.5166015625,
0.3203125,
0.5166015625,
0.3203125,
0.52783203125,
0.26708984375,
0.52783203125
],
[
0.148681640625,
0.53076171875,
0.223876953125,
0.53076171875,
0.223876953125,
0.53857421875,
0.148681640625,
0.53857421875
],
[
0.26708984375,
0.52783203125,
0.393798828125,
0.52783203125,
0.393798828125,
0.53857421875,
0.26708984375,
0.53857421875
],
[
0.5146484375,
0.53515625,
0.8857421875,
0.53515625,
0.8857421875,
0.548828125,
0.5146484375,
0.548828125
],
[
0.53369140625,
0.55029296875,
0.8857421875,
0.55029296875,
0.8857421875,
0.564453125,
0.53369140625,
0.564453125
],
[
0.16259765625,
0.556640625,
0.40185546875,
0.556640625,
0.40185546875,
0.57080078125,
0.16259765625,
0.57080078125
],
[
0.53173828125,
0.56494140625,
0.8857421875,
0.56494140625,
0.8857421875,
0.5791015625,
0.53173828125,
0.5791015625
],
[
0.53173828125,
0.5791015625,
0.88623046875,
0.5791015625,
0.88623046875,
0.59228515625,
0.53173828125,
0.59228515625
],
[
0.0882568359375,
0.5947265625,
0.2900390625,
0.59521484375,
0.2900390625,
0.60986328125,
0.0882568359375,
0.609375
],
[
0.53125,
0.59228515625,
0.88623046875,
0.59326171875,
0.88623046875,
0.611328125,
0.53125,
0.60986328125
],
[
0.53173828125,
0.609375,
0.8857421875,
0.609375,
0.8857421875,
0.62255859375,
0.53173828125,
0.62255859375
],
[
0.08660888671875,
0.61865234375,
0.473876953125,
0.619140625,
0.473876953125,
0.6337890625,
0.08660888671875,
0.63330078125
],
[
0.53173828125,
0.625,
0.88623046875,
0.625,
0.88623046875,
0.63818359375,
0.53173828125,
0.63818359375
],
[
0.0882568359375,
0.63525390625,
0.473876953125,
0.63525390625,
0.473876953125,
0.6484375,
0.0882568359375,
0.6484375
],
[
0.53369140625,
0.64013671875,
0.8857421875,
0.64013671875,
0.8857421875,
0.65380859375,
0.53369140625,
0.65380859375
],
[
0.0882568359375,
0.65087890625,
0.473876953125,
0.65087890625,
0.473876953125,
0.6640625,
0.0882568359375,
0.6640625
],
[
0.53369140625,
0.6572265625,
0.6357421875,
0.6572265625,
0.6357421875,
0.66845703125,
0.53369140625,
0.66845703125
],
[
0.0882568359375,
0.66552734375,
0.473876953125,
0.66552734375,
0.473876953125,
0.6787109375,
0.0882568359375,
0.6787109375
],
[
0.0882568359375,
0.68115234375,
0.4755859375,
0.68115234375,
0.4755859375,
0.6943359375,
0.0882568359375,
0.6943359375
],
[
0.5107421875,
0.67822265625,
0.88623046875,
0.6787109375,
0.88623046875,
0.693359375,
0.5107421875,
0.6923828125
],
[
0.0882568359375,
0.69580078125,
0.473876953125,
0.69580078125,
0.473876953125,
0.708984375,
0.0882568359375,
0.708984375
],
[
0.53173828125,
0.69580078125,
0.88623046875,
0.69580078125,
0.88623046875,
0.708984375,
0.53173828125,
0.708984375
],
[
0.08905029296875,
0.7119140625,
0.47216796875,
0.7119140625,
0.47216796875,
0.72265625,
0.08905029296875,
0.72265625
],
[
0.53369140625,
0.71044921875,
0.8837890625,
0.71044921875,
0.8837890625,
0.7236328125,
0.53369140625,
0.7236328125
],
[
0.0882568359375,
0.72607421875,
0.473876953125,
0.72607421875,
0.473876953125,
0.7392578125,
0.0882568359375,
0.7392578125
],
[
0.53515625,
0.7265625,
0.8837890625,
0.7265625,
0.8837890625,
0.7373046875,
0.53515625,
0.7373046875
],
[
0.0882568359375,
0.7412109375,
0.473876953125,
0.7412109375,
0.473876953125,
0.7548828125,
0.0882568359375,
0.7548828125
],
[
0.53173828125,
0.74072265625,
0.88818359375,
0.74072265625,
0.88818359375,
0.75390625,
0.53173828125,
0.75390625
],
[
0.0882568359375,
0.7568359375,
0.34228515625,
0.7568359375,
0.34228515625,
0.77099609375,
0.0882568359375,
0.77099609375
],
[
0.53173828125,
0.7548828125,
0.8857421875,
0.7548828125,
0.8857421875,
0.76806640625,
0.53173828125,
0.76806640625
],
[
0.53125,
0.76953125,
0.6337890625,
0.771484375,
0.6337890625,
0.78515625,
0.53125,
0.78271484375
],
[
0.08905029296875,
0.79052734375,
0.363525390625,
0.79052734375,
0.363525390625,
0.8046875,
0.08905029296875,
0.8046875
],
[
0.5009765625,
0.8017578125,
0.8857421875,
0.8017578125,
0.8857421875,
0.81494140625,
0.5009765625,
0.81494140625
],
[
0.08905029296875,
0.818359375,
0.47216796875,
0.818359375,
0.47216796875,
0.8291015625,
0.08905029296875,
0.8291015625
],
[
0.5,
0.8173828125,
0.8857421875,
0.8173828125,
0.8857421875,
0.83056640625,
0.5,
0.83056640625
],
[
0.08905029296875,
0.83203125,
0.47802734375,
0.83203125,
0.47802734375,
0.84521484375,
0.08905029296875,
0.84521484375
],
[
0.5,
0.83349609375,
0.8837890625,
0.83349609375,
0.8837890625,
0.84326171875,
0.5,
0.84326171875
],
[
0.08905029296875,
0.8466796875,
0.4755859375,
0.8466796875,
0.4755859375,
0.85986328125,
0.08905029296875,
0.85986328125
],
[
0.498291015625,
0.8466796875,
0.88623046875,
0.8466796875,
0.88623046875,
0.85986328125,
0.498291015625,
0.85986328125
],
[
0.0882568359375,
0.85986328125,
0.4755859375,
0.861328125,
0.4755859375,
0.87548828125,
0.0882568359375,
0.875
],
[
0.5,
0.8623046875,
0.8857421875,
0.8623046875,
0.8857421875,
0.87548828125,
0.5,
0.87548828125
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.5,
0.876953125,
0.8857421875,
0.876953125,
0.8857421875,
0.89013671875,
0.5,
0.89013671875
],
[
0.0882568359375,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.5,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"CURIOUS: Intrinsically Motivated Modular Multi-Goal Reinforcement Learning",
"4. Module and goal substitution. The agent decides on",
"target with the gripper; (M2) Push cube 1 onto a 2D target",
"which modules and goals to train. To update the policy",
"onthetable;(M3)Pick andPlace cube1 on a3D target;",
"and critic, the algorithm first samples a minibatch from",
"(M4)Stackcube1overcube2.AdditionalPushmodules",
"the replay buffers (red) using PLP and implements",
"concerning additional out-of-reach and moving cubes can",
"module and goal substitutions to perform cross-module",
"be defined (impossible, distracting goals). Further details",
"and cross-goal learning (orange), see Sec. 2.2.",
"can be found in the supplementary document.",
"5. Internal reward. The agent computes its reward r",
"4. Experiment and Results",
"for each transition, using RM,g parameterized by the",
"substitute module m* and goal g* (brown).",
"In this section, we present ablative studies to assess the",
"relative importance of: ) the policy and value function",
"6. RL updates. The agent updates its policy and value",
"architectureand2)theuseofintrinsicallymotivatedmod",
"function with DDPG using themodified minibatch",
"ule selection for practice and replay.We call M-UVFA the",
"(green).",
"algorithm using a modular goal-parameterized policy and",
"random module choices, while the intrinsically motivated",
"version is called cURIoUs. We do not investigate the ef-",
"outcomes",
"PLP",
"timato",
"ficiencyofHERortheefficiencyofthesamplingbiasto-",
"",
"LPupdate",
"PLP",
"wards interesting transitions as they were already studied",
"Mnihateh",
"Memory",
"(s,asgm",
"in Andrychowicz et al. (2017); Plappert et al. (2018) and",
"Mod",
"t",
"Zhao & Tresp (2018) respectively.For fair comparisons, we",
"On",
"rollout",
"m",
"(s, a, s',g,ma)",
"Pproba",
"apply both mechanisms to all the tested algorithms.",
"Goalsub",
"ution",
"g",
"outcome",
"E:",
"P",
"4.1. Impact of Policy and Value Function Architecture",
"(0 or 1)",
"Module and goal substitutions",
"RLupdates",
"Minibatch",
"Int.rewars",
"Minibatch",
"Experiments.In this section, we investigate the impact",
"s,a, s',g*,md*,r)",
"DDPG",
"uncton",
"(s, a, s,g*, md",
"Newparamet",
"of using an M-UVFA architecture for the policy and value",
"Module",
"function. The module-set is composed of four achievable",
"UVEA",
"nd",
"sampling",
"modules and four distracting modules. We test this algo-",
"Goal",
"rithm against two baselines:",
"En",
"sampling",
"Data collection",
"Moduleand goal selection",
"1. A flat multi-goal architecture (HER). This algorithm",
"does not represent goals in a modular fashion but",
"Figure 3. Schematic view of CURIOUS.",
"in a linear way. The corresponding goal is selected",
"uniformly inside G, a holistic goal space such that",
"2.4. Evaluation Methodology",
"g = II, GM.. To generate a reward, the agent needs",
"to satisfy the constraints described by all the modules",
"The performance of the agents are evaluated offline in terms",
"at once. This goal-parameterized architecture is equiv-",
"of success rates over sets of achievable goals (defined by the",
"alent to UVFA, which makes the algorithm equivalent",
"experimenter). Every point of a learning curve represents",
"to HER +DDPG.",
"the success rate over 95 offline evaluation rollouts (5 × 19",
"actors), using random achievable goals. Evaluation is com-",
"2. A multi-goal module-experts architecture (MG-ME)",
"pletely independent from training, i.e. agents cannot train",
"where an expert multi-goal policy is trained for each of",
"on evaluation transitions. We usethe non-parametric one",
"the N modules. Each policy is trained one epoch every",
"tail Mann-Whitney U-test with confidence level α = 0.01",
"N on its designated module and shares the collected",
"for all comparisons. More details and justifications can be",
"transitions with other experts. When evaluated on a par-",
"found in the supplementary document.",
"ticular module, the algorithm uses the corresponding",
"module-expert.",
"3. A Modular Goal Environment",
"Results. Fig. 5 shows the evolution of the average success",
"ModularGoal FetchArmisanew simulatedenvironment",
"rate computed over achievable goals for M-UvFA and the",
"adapted from the OpenAI Gym suite (Brockman et al.,",
"two baselines described above.Thelearning curve of HER",
"2016). The agent is embodied by a robotic arm facing",
"stays flat. This can be easily understood as none of the goals",
"2 cubes randomly positioned on a table. The agent controls",
"expressed in the complete goal space G corresponds to a real",
"the position of its gripper and the gripper opening (4D). It",
"situation (e.g. the agent cannot reach a 3D target with its",
"can target a diverse set of modular goals: (Mi) Reach a 3D",
"gripper while placing a cube at another). The agent cannot"
] |
{
"arxiv_id": "1810.06284",
"language": "en",
"page_idx": 4
}
|
||
1708.02212-en-1
|
1708.02212-en-1.jpg
|
and perceptual judgements, and have the additional benefits of being robust to small annotation errors (such as minor differences between multiple human annotators). The $F_{\beta}^w$ measure is also able to reliably rank generic foreground maps, such as centered geometric shapes, lower than state of the art predictions (They show traditional metrics, such as AUC, lack this property). Despite these positive traits, their formulation has $O(n^2)$ memory and computational requirements where $n$ the number of pixels in the image.
This computational cost poses a particular problem if the $F_{\beta}^{w}$ metric were used as the training objective for deep neural networks (DNNs). Normally trained with stochastic gradient descent over large training sets, DNNs require computing the gradient of the loss many, many, times. This means that the loss function must be differentiable, and efficient-two criteria which $F_{\beta}^{w}$ does not meet.
Our primary contribution is a differentiable and computationallyeffcient approximation of the $F_{\beta}^{w}$ metric, which can be used directly as the loss function of a convolutional neural network (CNN). As a secondary contribution, we propose a memory-efficient CNN architecture which is capable of producing high resolution pixel-wise predictions, taking full advantage of the spatial information provided by our proposed loss. By combining these two components we are able to produce high-fidelity, spatially cohesive predictions, without relying on complex, often expensive preprocessing (such as super-pixels) or post-processing (such as CRF inference), resulting in inference speeds an order of magnitude faster than state of the art in multiple domains. We do not sacrifice accuracy, achieving competitive or state of the art accuracy on benchmarks for salient object detection, portrait segmentation, and visual distractor masking. $2. \textbf{ Background}$
In this section we discuss the prior work on incorporating spatial consideration into learning objectives. While multiple objectives have been proposed to capture spatial properties of prediction maps [2,24,25], these have been limited to structured prediction methods using random fields, and adds significant complexity when incorporated into a feed-forward prediction framework like that of CNNs. We focus on the $F_{\beta}^{w}$ metric, which is decoupled from the prediction framework and upon which we directly build our approach. We review it below, and also survey the related work on the segmentation tasks on which we evaluate our contributions: salient object detection, distractor detection, and portrait segmentation.
# $2. 1. \textbf{The }F_{\beta }\textbf{ metric family}$
In a binary classification scenario, with labels $y\in$ $\{0,1\}$, when the predicted label $\widehat{y}$ is a mistake $\widehat{y}\neq y$, it is either a false positive (FP$, \widehat{y} = 1)$ or a false negative (FN, $\widehat{y}=0).$ Performance of any classifier on an evaluation set
can be characterized by its precision $\#TP/(\#TP+\#FP)$
and its recall $\#TP/(\#TP+\#FN).$
While precision and recall each only tell part of the story, one can summarize a classification algorithm's performance in a single number, using the $F_{\beta}$ metric
$$\begin{aligned}F_\beta=\frac{(1+\beta^2)*Precision*Recall}{\beta^2Precision+Recall}.\end{aligned}$$
(1)
$\beta$ captures the relative importance of precision compared to recall (e.g. if precision is twice as important as recall, we use $F_2).$ The well known $F_1$ metric is a special case corresponding to equal importance between precision and recall. The $F_{\beta}$ metrics is a common benchmark in'information extraction' tasks, and in [10] Jansche outlines a procedure to directly optimize it. This formulation applies to any scenario when $F_\beta$ is meaningful, but it cannot encode differences within the categories of false positive, and false negative, which are quite meaningful in the highly structured domain of natural images.
# $2. 2. \textbf{The }F_{\beta }^{w}\textbf{metric}$
The standard $F_{\beta}$ is extended in [22] in two ways. First, it is generalized to handle continuous predictions, $\widehat{y}\in[0,1]$ (the ground truth $y$ remains binary). The adjusted definitions of the true positive, false positive, true negative, and false negative are as follows:
$$\begin{aligned}E&=|Y-\widehat{Y}|\\TP&=(1-E)\cdot Y\\TN&=(1-E)\cdot(1-Y)\\FP&=E\cdot(1-Y)\\FN&=E\cdot Y\end{aligned}$$
(2)
This holds in the case of predicting a set of values; $Y$ is the vector of ground truth labels $,\widehat{Y}$ is the vector of predictions, and$\cdot$denotes the dot product
The second modification proposed in [22] addresses the unequal nature of mistakes in binary segmentation $(y=1$ implying foreground, $y=0$ background), as determined by the spatial configuration of predictions vs. ground truth. The authors of [22] suggest a number of criteria for evaluating foreground maps.
First consider false negatives, missed detections of foreground pixels. If random foreground pixels across an object are undetected, leaving small holes in the foreground, this is easily corrected via post-processing. However, concentrating the same number of errors in one part of the object is much more perceptually severe and difficult to correct. See the top row of Figure 1. This is captured by by re-weighting
|
[
[
0.07843017578125,
0.091552734375,
0.468017578125,
0.09027099609375,
0.468017578125,
0.10479736328125,
0.07843017578125,
0.1060791015625
],
[
0.5009765625,
0.0921630859375,
0.8896484375,
0.0921630859375,
0.8896484375,
0.1060791015625,
0.5009765625,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.5009765625,
0.106689453125,
0.75244140625,
0.106689453125,
0.75244140625,
0.12060546875,
0.5009765625,
0.12060546875
],
[
0.08087158203125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.51953125,
0.12249755859375,
0.890625,
0.125,
0.890625,
0.1395263671875,
0.51953125,
0.136962890625
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50244140625,
0.14013671875,
0.89208984375,
0.14013671875,
0.89208984375,
0.154052734375,
0.50244140625,
0.154052734375
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50244140625,
0.1546630859375,
0.7646484375,
0.1546630859375,
0.7646484375,
0.169189453125,
0.50244140625,
0.169189453125
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.07843017578125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.07843017578125,
0.1962890625
],
[
0.60546875,
0.184326171875,
0.82275390625,
0.18310546875,
0.82275390625,
0.2008056640625,
0.60546875,
0.2020263671875
],
[
0.56884765625,
0.1932373046875,
0.60888671875,
0.1932373046875,
0.60888671875,
0.208984375,
0.56884765625,
0.208984375
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.869140625,
0.1944580078125,
0.89208984375,
0.1944580078125,
0.89208984375,
0.2083740234375,
0.869140625,
0.2083740234375
],
[
0.63623046875,
0.20263671875,
0.7958984375,
0.20263671875,
0.7958984375,
0.216552734375,
0.63623046875,
0.216552734375
],
[
0.080078125,
0.21337890625,
0.438720703125,
0.21337890625,
0.438720703125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.1004638671875,
0.2279052734375,
0.471435546875,
0.2279052734375,
0.471435546875,
0.242431640625,
0.1004638671875,
0.242431640625
],
[
0.50341796875,
0.231689453125,
0.890625,
0.231689453125,
0.890625,
0.24560546875,
0.50341796875,
0.24560546875
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.25,
0.890625,
0.25,
0.890625,
0.260009765625,
0.50341796875,
0.260009765625
],
[
0.080078125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.080078125,
0.271484375
],
[
0.50341796875,
0.26318359375,
0.890625,
0.26318359375,
0.890625,
0.276611328125,
0.50341796875,
0.276611328125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.279052734375,
0.890625,
0.279052734375,
0.890625,
0.292236328125,
0.50341796875,
0.292236328125
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.292236328125,
0.890625,
0.292236328125,
0.890625,
0.30615234375,
0.50244140625,
0.30615234375
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50341796875,
0.31005859375,
0.890625,
0.31005859375,
0.890625,
0.32080078125,
0.50341796875,
0.32080078125
],
[
0.080078125,
0.318115234375,
0.384033203125,
0.318115234375,
0.384033203125,
0.33203125,
0.080078125,
0.33203125
],
[
0.50244140625,
0.322509765625,
0.89208984375,
0.322509765625,
0.89208984375,
0.3359375,
0.50244140625,
0.3359375
],
[
0.1004638671875,
0.335205078125,
0.465576171875,
0.335205078125,
0.465576171875,
0.34521484375,
0.1004638671875,
0.34521484375
],
[
0.50244140625,
0.33837890625,
0.89208984375,
0.33837890625,
0.89208984375,
0.3515625,
0.50244140625,
0.3515625
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.50244140625,
0.352783203125,
0.8896484375,
0.352783203125,
0.8896484375,
0.3662109375,
0.50244140625,
0.3662109375
],
[
0.08087158203125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.375732421875,
0.08087158203125,
0.375732421875
],
[
0.50341796875,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.080078125,
0.37890625,
0.4697265625,
0.37890625,
0.4697265625,
0.392578125,
0.080078125,
0.392578125
],
[
0.50244140625,
0.3837890625,
0.6748046875,
0.3837890625,
0.6748046875,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.07843017578125,
0.39453125,
0.465576171875,
0.393310546875,
0.465576171875,
0.4072265625,
0.07843017578125,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50341796875,
0.4140625,
0.64892578125,
0.4140625,
0.64892578125,
0.4287109375,
0.50341796875,
0.4287109375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.521484375,
0.455078125,
0.890625,
0.455810546875,
0.890625,
0.47021484375,
0.521484375,
0.4697265625
],
[
0.080078125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50244140625,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.08087158203125,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.50341796875,
0.4873046875,
0.8896484375,
0.4873046875,
0.8896484375,
0.50048828125,
0.50341796875,
0.50048828125
],
[
0.08087158203125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51220703125,
0.08087158203125,
0.51220703125
],
[
0.50244140625,
0.501953125,
0.89208984375,
0.501953125,
0.89208984375,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.08087158203125,
0.5166015625,
0.46728515625,
0.5166015625,
0.46728515625,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.5009765625,
0.5166015625,
0.6953125,
0.51513671875,
0.6953125,
0.52978515625,
0.5009765625,
0.53076171875
],
[
0.080078125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.08087158203125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.08087158203125,
0.5595703125
],
[
0.625,
0.54541015625,
0.7197265625,
0.54541015625,
0.7197265625,
0.56298828125,
0.625,
0.56298828125
],
[
0.080078125,
0.56103515625,
0.45751953125,
0.56103515625,
0.45751953125,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.61376953125,
0.5673828125,
0.740234375,
0.5673828125,
0.740234375,
0.58203125,
0.61376953125,
0.58203125
],
[
0.0767822265625,
0.58544921875,
0.2059326171875,
0.5869140625,
0.2059326171875,
0.60498046875,
0.0767822265625,
0.60302734375
],
[
0.61279296875,
0.5859375,
0.78173828125,
0.5859375,
0.78173828125,
0.60107421875,
0.61279296875,
0.60107421875
],
[
0.86767578125,
0.58544921875,
0.89453125,
0.58544921875,
0.89453125,
0.6015625,
0.86767578125,
0.6015625
],
[
0.61279296875,
0.60498046875,
0.74169921875,
0.60498046875,
0.74169921875,
0.619140625,
0.61279296875,
0.619140625
],
[
0.09722900390625,
0.6123046875,
0.468017578125,
0.61279296875,
0.468017578125,
0.62744140625,
0.09722900390625,
0.626953125
],
[
0.07843017578125,
0.62744140625,
0.468017578125,
0.626953125,
0.468017578125,
0.6416015625,
0.07843017578125,
0.642578125
],
[
0.611328125,
0.62353515625,
0.701171875,
0.62353515625,
0.701171875,
0.63818359375,
0.611328125,
0.63818359375
],
[
0.080078125,
0.64306640625,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.50244140625,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.66943359375,
0.50244140625,
0.66943359375
],
[
0.080078125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.080078125,
0.6875
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.50244140625,
0.6875,
0.69677734375,
0.6875,
0.69677734375,
0.70068359375,
0.50244140625,
0.70068359375
],
[
0.080078125,
0.7041015625,
0.4697265625,
0.7041015625,
0.4697265625,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.521484375,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.521484375,
0.71728515625
],
[
0.080078125,
0.71826171875,
0.46728515625,
0.71826171875,
0.46728515625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50341796875,
0.72021484375,
0.8896484375,
0.72021484375,
0.8896484375,
0.73095703125,
0.50341796875,
0.73095703125
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.5009765625,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.5009765625,
0.7470703125
],
[
0.07843017578125,
0.75,
0.4697265625,
0.74853515625,
0.4697265625,
0.76220703125,
0.07843017578125,
0.76318359375
],
[
0.50244140625,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.76318359375,
0.4697265625,
0.76318359375,
0.4697265625,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.080078125,
0.77978515625,
0.4697265625,
0.77978515625,
0.4697265625,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.77880859375,
0.65673828125,
0.77880859375,
0.65673828125,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.07843017578125,
0.79541015625,
0.253173828125,
0.79541015625,
0.253173828125,
0.8095703125,
0.07843017578125,
0.8095703125
],
[
0.521484375,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.8173828125,
0.27294921875,
0.8173828125,
0.27294921875,
0.83203125,
0.080078125,
0.83203125
],
[
0.5048828125,
0.82763671875,
0.890625,
0.82763671875,
0.890625,
0.83837890625,
0.5048828125,
0.83837890625
],
[
0.09722900390625,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.09722900390625,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85498046875,
0.4697265625,
0.85595703125,
0.4697265625,
0.87060546875,
0.080078125,
0.869140625
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"and perceptual judgements, and have the additional bene-",
"canbecharacterizedbyitsprecision#TP/(#TP+#FP)",
"fits of being robustto small annotation errors (such as mi-",
"anditsrecall#TP/(#TP+#FN).",
"nordifferencesbetweenmultiplehumanannotators).The",
"While precision and recall each only tell part of the story,",
"F measure is also able to reliably rank generic foreground",
"one can summarize a classification algorithm's performance",
"maps,such as centered geometricshapes,lowerthan state",
"in a single number, using the Fg metric",
"of the art predictions (They show traditional metrics, such",
"as AUC, lack this property). Despite these positive traits,",
"(1 + β2) * Precision * Recall",
"F =",
"theirformulationhasO(n)memoryandcomputationalre-",
"(1)",
"B2Precision+Recall",
"quirements where n the number of pixels in the image.",
"This computational cost poses a particular problem if",
"β captures the relative importance of precision compared to",
"the F metric were used as the training objective for deep",
"recall (e.g.if precision is twice as important asrecall,we",
"neural networks (DNNs). Normally trained with stochas-",
"use F2). The well known Fi metric is a special case corre-",
"tic gradient descent over large training sets, DNNs require",
"sponding to equal importance between precision and recall.",
"computing the gradient of the loss many, many, times. This",
"The Fg metrics is a common benchmark in 'information ex-",
"means that the loss function must be differentiable, and ef-",
"traction'tasks,andin [iolJansche outlinesaprocedureto",
"ficient - two criteria which F does not meet.",
"directly optimize it. This formulation applies to any sce-",
"Ourprimarycontributionisadifferentiableand compu-",
"nario when F is meaningful, but it cannot encode differ-",
"tationally efficient approximation of the Fu metric, which",
"ences within the categories of false positive, and false neg-",
"canbeuseddirectlyasthelossfunctionof aconvolutional",
"ative, which are quite meaningful in the highly structured",
"neural network (CNN). As a secondary contribution, we",
"domain of natural images.",
"propose a memory-efficient CNN architecture which is ca-",
"pable of producing high resolution pixel-wise predictions,",
"2.2. The F metric",
"taking full advantage of the spatial information provided by",
"our proposed loss.By combining these two components",
"we are able to produce high-fidelity, spatially cohesive pre-",
"The standard Fg is extended in [22] in two ways. First,",
"dictions, without relying on complex, often expensive pre-",
"it is generalized to handle continuous predictions, g e [o, 1]",
"processing(suchassuper-pixels)orpost-processing(such",
"(the ground truth y remains binary). The adjusted defini-",
"as CRFinference).resultingininferencespeedsanorderof",
"tions of the true positive, false positive, true negative, and",
"magnitudefasterthan state of the artinmultiple domains.",
"false negative are as follows:",
"We do not sacrifice accuracy, achieving competitive or state",
"of the art accuracy on benchmarks for salient object detec-",
"E=|Y-Y",
"tion, portrait segmentation, and visual distractor masking.",
"TP=(1-E)·Y",
"2. Background",
"TN=(1-E)·(1-Y)",
"(2)",
"FP=E.(1-Y)",
"In this section we discuss the prior work on incorporating",
"spatial consideration into learning objectives. While multi-",
"FN=E.Y",
"ple objectives have been proposed to capture spatial prop-",
"erties of prediction maps [2, 24, 25], these have been lim-",
"This holds in the case of predicting a set of values; Y is the",
"ited to structured prediction methods using random fields,",
"vector of ground truth labels , Y is the vector of predictions,",
"and adds significant complexity when incorporated into a",
"and - denotes the dot product",
"feed-forward prediction framework like that of CNNs. We",
"The second modification proposed in [22] addresses the",
"focus on the Fu metric, which is decoupled from the pre-",
"unequal nature of mistakes inbinary segmentation(y = 1",
"diction framework and upon which we directly build our",
"implying foreground, y = o background), as determined",
"approach.We review it below, and also survey the related",
"by the spatial configuration of predictions vs. ground truth.",
"work on the segmentation tasks on which we evaluate our",
"The authors of [22] suggest a number of criteria for evalu-",
"contributions: salient object detection, distractor detection,",
"ating foreground maps.",
"and portrait segmentation.",
"First consider false negatives, missed detections of fore-",
"ground pixels. If random foreground pixels across an object",
"2.1. The Fs metric family",
"are undetected,leaving small holes in the foreground, this",
"In a binary classification scenario, with labels y E",
"is easily corrected via post-processing. However, concen-",
"0,11, when the predicted label y is a mistake y + y, it",
"trating the samenumber of errors in onepart of the object is",
"is either a false positive (FP, y = 1) or a false negative (FN,",
"much more perceptually severe and difficult to correct. See",
"y = O). Performance of any classifier on an evaluation set",
"the top row of Figure 1. This is captured by by re-weighting"
] |
{
"arxiv_id": "1708.02212",
"language": "en",
"page_idx": 1
}
|
||
1708.02212-en-3
|
1708.02212-en-3.jpg
|
<FigureHere>
Figure 3: A vizualization on a synthetic example of how our loss function re-weights mistakes. In the top row we visualize the down-weighting of false positives near the true object border. In the bottom-row we show the downweighting of false-negatives do not spatially co-occur many other false-negatives and the increased weight of false negatives which spatially co-occur. True positives are marked by green, and mistakes are marked in red.
posed by [8]. Distractors are defined as visually salient parts of an image which are not the photographer's intended focus. This task is somewhat similar to salient object detection, but successful algorithms must go beyond simply detecting all salient objects, and model the image at a global level to discriminate between the intended focus of the image and the distractors. In [8]Fried et al. propose an SVM based approach, trained on a relatively small dataset, which classifies super-pixels extracted by Multiscale Combinatorial Grouping (MCG) based on a set of hand-crafted features. To test the robustness of their approach we gathered a larger dataset with crowdsourced labels. While their method is able to detect large and well defined distractors, it struggles to detect non-object distractors such as shadows, lights and reflections as well as select small objects.
## 2.5. Portrait segmentation
Portraits are highly popular art form in both photography and painting. In most instances, artists seek to make the subject stand out from its surrounding, for instance, by making it brighter or sharper or by applying photographic or painterly filters that adapt to the semantics of the image. Shen et al.[26] presented a new high quality automatic portrait segmentation algorithm by adapting the FCN-8s framework [20]. They also introduced a portrait image segmentation dataset and benchmark for training and testing.
# 3. Our Approximate $F_\beta ^w\textbf{ loss}( AF_\beta ^w)$
There are three issues that prevent the $F_\beta^w$ metric as defined in Section 2.2 from being directly optimized as a loss function. The first is that while the metric is differentiable almost everywhere, it is not differentiable when $y=\hat{y}$, be-
cause $\frac d{d\hat{y}_i}|\hat{y}_i-y_i|$ is undefined. In practice, we observed difficulties optimizing the error using SGD due to the constant value of the gradient for $\hat{y}\neq y$ (intuitively, because the gradient doesn't decrease as the error decreases). We solve this by replacing the $L_1$ norm with $L_2{:}E_i=(\hat{y}_i-y_i)^2$, which we find to be much easier to optimize.
The second problem is that constructing A (not to mention computing $E\mathbb{A})$ has $O(n^2)$ time and space complexity. However, we can overcome this problem by leveraging convolutions. When we unpack the definition of matrix multiplication in Equation (3), we can write $E\mathbb{A}$ at pixel $i$ as:
$$\begin{aligned}(E\mathbb{A})_{i}&=\sum_j[E_j\mathbb{A}_{ji}]=\\&=y_{i}\sum_{j}[y_{j}E_{j}\frac{1}{\sqrt{2\pi\sigma^{2}}}e^{\frac{-d(i,j)}{2\sigma^{2}}}]+(1-y_{i})\end{aligned}$$
(7)
${\mathrm{Note~that~if~}d(i,j)\geq4\sigma\mathrm{~then~}\frac{1}{\sqrt{2\pi\sigma^{2}}}e^{\frac{-d(i,j)}{2\sigma^{2}}}\approx0.}{\mathrm{W}}$ $\begin{array}{l}\text{then define }y_{p,q}\text{ and }E_{p,q}\text{ as the ground truth and error re-}\\\text{spectively at pixel }(p,q),\text{and can approximate }(E\mathbb{A})_{i}\mathrm{~as:}\end{array}$
$$\begin{aligned}(E\mathbb{A})_{i}&\approx(1-y_{i})+\\&+y_{i+(p,q)}\sum_{p,q=-4\sigma}^{4\sigma}[y\cdot E]_{i+(p,q)}\frac{1}{\sqrt{2\pi\sigma^{2}}}e^{\frac{-\sqrt{p^{2}+q^{2}}}{2\sigma^{2}}}\end{aligned}$$
If we let $\sigma=\frac\theta4$, we can define a $(2\theta+1)\times(2\theta+1)$
Gaussian convolutional kernel $K^{h_\mathrm{\Lambda}}$, yielding
$$\begin{aligned}&E\mathbb{A}\approx Y\odot((Y\odot E)*K^{\mathbb{A}})+(1-Y),\mathrm{where}\\&K_{i,j}^{\mathbb{A}}=\frac{1}{\sqrt{2\pi\sigma^{2}}}e^{\frac{-\sqrt{(-\theta+i)^{2}+(-\theta+j)^{2}}}{2\sigma^{2}}}\end{aligned}$$
(9)
Where $\odot$ is element-wise multiplication. Now we don't need to store any entries of $\mathbb{A}$, only a$(2\theta+1)\times(2\theta+1)$ kernel and we skip most of the original summation over pixels $j.$ So the time and space complexity is reduced to $O(n+\theta^{2}).$ In practice, we use $\theta=9.$
The final problem is that constructing $\mathbb{B}$ has complexity $O(n^2)$, because computing $\Delta_i$ requires finding the mini- $\textbf{mum }d( i, j)$ over all pixels $j$, such that $y_j=1.$ However, if $d(i,j)\geq25$, then $\mathbb{B}_i=2-e^{\alpha\cdot\Delta_i}\approx2$ in Equation (4). Our intuition is that $\mathbb{B}$ is modeling the region of uncertainty about an object's boundary, for which we believe 25 pixels to be too generous. So we redefine $\Delta_i$ as:
$$\Delta_i=\begin{cases}\infty,&\text{if}\min_jd(i,j)>5\\\min_jd(i,j)^2|y_j=1,&\text{otherwise}\end{cases}$$
Squaring the distance so that the region of uncertainty is assumed to be approximately 5 pixels instead of 25. Now
|
[
[
0.50341796875,
0.09344482421875,
0.55126953125,
0.09344482421875,
0.55126953125,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.54833984375,
0.09027099609375,
0.89208984375,
0.09027099609375,
0.89208984375,
0.10479736328125,
0.54833984375,
0.10479736328125
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.50244140625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1337890625,
0.156005859375,
0.1337890625,
0.156005859375,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.50341796875,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.16259765625,
0.149658203125,
0.26953125,
0.149658203125,
0.26953125,
0.1641845703125,
0.16259765625,
0.1641845703125
],
[
0.303955078125,
0.15087890625,
0.32275390625,
0.15087890625,
0.32275390625,
0.1641845703125,
0.303955078125,
0.1641845703125
],
[
0.375732421875,
0.15087890625,
0.40185546875,
0.15087890625,
0.40185546875,
0.1641845703125,
0.375732421875,
0.1641845703125
],
[
0.50341796875,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.50244140625,
0.1666259765625,
0.79736328125,
0.1673583984375,
0.79736328125,
0.1817626953125,
0.50244140625,
0.18115234375
],
[
0.374267578125,
0.1875,
0.41015625,
0.1875,
0.41015625,
0.20263671875,
0.374267578125,
0.20263671875
],
[
0.521484375,
0.1817626953125,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.521484375,
0.1956787109375
],
[
0.2017822265625,
0.189453125,
0.23779296875,
0.189453125,
0.23779296875,
0.2020263671875,
0.2017822265625,
0.2020263671875
],
[
0.2998046875,
0.189453125,
0.33251953125,
0.189453125,
0.33251953125,
0.2008056640625,
0.2998046875,
0.2008056640625
],
[
0.50244140625,
0.1962890625,
0.890625,
0.1976318359375,
0.890625,
0.212158203125,
0.50244140625,
0.2108154296875
],
[
0.5009765625,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.5009765625,
0.2266845703125
],
[
0.156005859375,
0.2266845703125,
0.275390625,
0.2279052734375,
0.275390625,
0.242431640625,
0.156005859375,
0.2412109375
],
[
0.303955078125,
0.2266845703125,
0.32275390625,
0.2266845703125,
0.32275390625,
0.2412109375,
0.303955078125,
0.2412109375
],
[
0.374267578125,
0.2259521484375,
0.404296875,
0.2259521484375,
0.404296875,
0.242431640625,
0.374267578125,
0.242431640625
],
[
0.50341796875,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.50341796875,
0.2412109375
],
[
0.50341796875,
0.24365234375,
0.869140625,
0.24365234375,
0.869140625,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.1070556640625,
0.2279052734375,
0.134033203125,
0.2279052734375,
0.134033203125,
0.240478515625,
0.1070556640625,
0.240478515625
],
[
0.080078125,
0.255126953125,
0.4697265625,
0.255126953125,
0.4697265625,
0.268310546875,
0.080078125,
0.268310546875
],
[
0.080078125,
0.270751953125,
0.4697265625,
0.270751953125,
0.4697265625,
0.2841796875,
0.080078125,
0.2841796875
],
[
0.52685546875,
0.2734375,
0.58642578125,
0.2734375,
0.58642578125,
0.28857421875,
0.52685546875,
0.28857421875
],
[
0.5947265625,
0.2734375,
0.67626953125,
0.2734375,
0.67626953125,
0.28857421875,
0.5947265625,
0.28857421875
],
[
0.08087158203125,
0.287353515625,
0.46728515625,
0.287353515625,
0.46728515625,
0.29736328125,
0.08087158203125,
0.29736328125
],
[
0.080078125,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.314453125,
0.080078125,
0.314453125
],
[
0.87109375,
0.298583984375,
0.89208984375,
0.298583984375,
0.89208984375,
0.313232421875,
0.87109375,
0.313232421875
],
[
0.57275390625,
0.3125,
0.66845703125,
0.31005859375,
0.66845703125,
0.32763671875,
0.57275390625,
0.330078125
],
[
0.74267578125,
0.3125,
0.7607421875,
0.3125,
0.7607421875,
0.31640625,
0.74267578125,
0.31640625
],
[
0.771484375,
0.31005859375,
0.8486328125,
0.313232421875,
0.84716796875,
0.32763671875,
0.76953125,
0.324462890625
],
[
0.08087158203125,
0.31640625,
0.468017578125,
0.31640625,
0.468017578125,
0.330078125,
0.08087158203125,
0.330078125
],
[
0.67236328125,
0.323974609375,
0.71337890625,
0.323974609375,
0.71337890625,
0.333984375,
0.67236328125,
0.333984375
],
[
0.07843017578125,
0.330078125,
0.46728515625,
0.330810546875,
0.46728515625,
0.34521484375,
0.07843017578125,
0.3447265625
],
[
0.61279296875,
0.333251953125,
0.625,
0.333251953125,
0.625,
0.341552734375,
0.61279296875,
0.341552734375
],
[
0.080078125,
0.3466796875,
0.468017578125,
0.3466796875,
0.468017578125,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.52294921875,
0.359130859375,
0.73388671875,
0.359130859375,
0.73388671875,
0.37255859375,
0.52294921875,
0.37255859375
],
[
0.7900390625,
0.35595703125,
0.8203125,
0.35595703125,
0.8203125,
0.363037109375,
0.7900390625,
0.363037109375
],
[
0.080078125,
0.3623046875,
0.3544921875,
0.3623046875,
0.3544921875,
0.375732421875,
0.080078125,
0.375732421875
],
[
0.775390625,
0.365478515625,
0.8037109375,
0.365478515625,
0.8037109375,
0.369384765625,
0.775390625,
0.369384765625
],
[
0.82861328125,
0.35986328125,
0.890625,
0.35986328125,
0.890625,
0.371826171875,
0.82861328125,
0.371826171875
],
[
0.5009765625,
0.375,
0.89208984375,
0.375,
0.89208984375,
0.392578125,
0.5009765625,
0.392578125
],
[
0.50341796875,
0.392578125,
0.8759765625,
0.392578125,
0.8759765625,
0.406005859375,
0.50341796875,
0.406005859375
],
[
0.080078125,
0.404052734375,
0.468017578125,
0.40283203125,
0.468017578125,
0.417236328125,
0.080078125,
0.41845703125
],
[
0.080078125,
0.419189453125,
0.468017578125,
0.419189453125,
0.468017578125,
0.43310546875,
0.080078125,
0.43310546875
],
[
0.51123046875,
0.419189453125,
0.64208984375,
0.419189453125,
0.64208984375,
0.434326171875,
0.51123046875,
0.434326171875
],
[
0.08087158203125,
0.436279296875,
0.463134765625,
0.436279296875,
0.463134765625,
0.4462890625,
0.08087158203125,
0.4462890625
],
[
0.6787109375,
0.448974609375,
0.771484375,
0.45263671875,
0.76953125,
0.47021484375,
0.67724609375,
0.466552734375
],
[
0.83837890625,
0.443115234375,
0.8798828125,
0.443115234375,
0.8798828125,
0.455078125,
0.83837890625,
0.455078125
],
[
0.080078125,
0.45068359375,
0.464111328125,
0.45068359375,
0.464111328125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.55810546875,
0.451904296875,
0.62744140625,
0.455078125,
0.6259765625,
0.4697265625,
0.55615234375,
0.466552734375
],
[
0.646484375,
0.45068359375,
0.65185546875,
0.45068359375,
0.65185546875,
0.4609375,
0.646484375,
0.4609375
],
[
0.8408203125,
0.45703125,
0.85888671875,
0.45703125,
0.85888671875,
0.4609375,
0.8408203125,
0.4609375
],
[
0.080078125,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.771484375,
0.46337890625,
0.8037109375,
0.46533203125,
0.80224609375,
0.473388671875,
0.76953125,
0.4716796875
],
[
0.625,
0.47021484375,
0.6787109375,
0.468505859375,
0.6787109375,
0.479736328125,
0.625,
0.481689453125
],
[
0.08087158203125,
0.48095703125,
0.46728515625,
0.48095703125,
0.46728515625,
0.4912109375,
0.08087158203125,
0.4912109375
],
[
0.869140625,
0.48095703125,
0.8935546875,
0.48095703125,
0.8935546875,
0.494384765625,
0.869140625,
0.494384765625
],
[
0.07843017578125,
0.49365234375,
0.468017578125,
0.494384765625,
0.468017578125,
0.5087890625,
0.07843017578125,
0.50830078125
],
[
0.07843017578125,
0.50830078125,
0.468017578125,
0.5087890625,
0.468017578125,
0.5234375,
0.07843017578125,
0.52294921875
],
[
0.521484375,
0.5087890625,
0.890625,
0.51025390625,
0.890625,
0.5244140625,
0.521484375,
0.5234375
],
[
0.08087158203125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.08087158203125,
0.53857421875
],
[
0.50244140625,
0.5244140625,
0.79150390625,
0.52587890625,
0.79150390625,
0.54052734375,
0.50244140625,
0.5390625
],
[
0.080078125,
0.54052734375,
0.4697265625,
0.54052734375,
0.4697265625,
0.5537109375,
0.080078125,
0.5537109375
],
[
0.07843017578125,
0.5537109375,
0.468017578125,
0.5546875,
0.468017578125,
0.5693359375,
0.07843017578125,
0.568359375
],
[
0.080078125,
0.57080078125,
0.468017578125,
0.57080078125,
0.468017578125,
0.583984375,
0.080078125,
0.583984375
],
[
0.53515625,
0.56640625,
0.84716796875,
0.56640625,
0.84716796875,
0.580078125,
0.53515625,
0.580078125
],
[
0.60302734375,
0.583984375,
0.61669921875,
0.5888671875,
0.609375,
0.5986328125,
0.59619140625,
0.59326171875
],
[
0.87109375,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.59765625,
0.87109375,
0.59765625
],
[
0.64794921875,
0.58544921875,
0.75732421875,
0.58544921875,
0.75732421875,
0.5986328125,
0.64794921875,
0.5986328125
],
[
0.080078125,
0.58544921875,
0.468017578125,
0.58544921875,
0.468017578125,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.53125,
0.59033203125,
0.56787109375,
0.5947265625,
0.564453125,
0.60986328125,
0.52685546875,
0.60595703125
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.5810546875,
0.60302734375,
0.6298828125,
0.60302734375,
0.6298828125,
0.6171875,
0.5810546875,
0.6171875
],
[
0.689453125,
0.5986328125,
0.705078125,
0.5986328125,
0.705078125,
0.6015625,
0.689453125,
0.6015625
],
[
0.080078125,
0.6162109375,
0.420654296875,
0.6162109375,
0.420654296875,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.52294921875,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.52294921875,
0.6416015625
],
[
0.080078125,
0.6396484375,
0.281982421875,
0.64013671875,
0.281982421875,
0.65478515625,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.098876953125,
0.6640625,
0.468017578125,
0.6640625,
0.468017578125,
0.6787109375,
0.098876953125,
0.6787109375
],
[
0.50341796875,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.080078125,
0.6796875,
0.468017578125,
0.6796875,
0.468017578125,
0.693359375,
0.080078125,
0.693359375
],
[
0.50244140625,
0.68798828125,
0.67626953125,
0.68798828125,
0.67626953125,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.080078125,
0.708984375
],
[
0.07843017578125,
0.708984375,
0.468017578125,
0.71044921875,
0.468017578125,
0.724609375,
0.07843017578125,
0.7236328125
],
[
0.52392578125,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71728515625,
0.52392578125,
0.71728515625
],
[
0.50341796875,
0.71826171875,
0.8896484375,
0.71826171875,
0.8896484375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.7265625,
0.465576171875,
0.7265625,
0.465576171875,
0.7373046875,
0.08087158203125,
0.7373046875
],
[
0.080078125,
0.7392578125,
0.46728515625,
0.7392578125,
0.46728515625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50341796875,
0.734375,
0.890625,
0.734375,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.5009765625,
0.74755859375,
0.7138671875,
0.74853515625,
0.7138671875,
0.76318359375,
0.5009765625,
0.76220703125
],
[
0.705078125,
0.7470703125,
0.8896484375,
0.74853515625,
0.8896484375,
0.76318359375,
0.705078125,
0.76123046875
],
[
0.07843017578125,
0.7548828125,
0.468017578125,
0.75390625,
0.468017578125,
0.76806640625,
0.07843017578125,
0.76953125
],
[
0.5009765625,
0.76318359375,
0.8896484375,
0.76318359375,
0.8896484375,
0.7763671875,
0.5009765625,
0.7763671875
],
[
0.080078125,
0.76953125,
0.468017578125,
0.77099609375,
0.468017578125,
0.78515625,
0.080078125,
0.7841796875
],
[
0.07843017578125,
0.7841796875,
0.418212890625,
0.78515625,
0.418212890625,
0.7998046875,
0.07843017578125,
0.798828125
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.50244140625,
0.79443359375,
0.77685546875,
0.79443359375,
0.77685546875,
0.80810546875,
0.50244140625,
0.80810546875
],
[
0.080078125,
0.814453125,
0.374267578125,
0.81640625,
0.374267578125,
0.83056640625,
0.080078125,
0.8291015625
],
[
0.73779296875,
0.82470703125,
0.86669921875,
0.82470703125,
0.86669921875,
0.8388671875,
0.73779296875,
0.8388671875
],
[
0.529296875,
0.83544921875,
0.5439453125,
0.83544921875,
0.5439453125,
0.8447265625,
0.529296875,
0.8447265625
],
[
0.5849609375,
0.83203125,
0.59619140625,
0.83203125,
0.59619140625,
0.8359375,
0.5849609375,
0.8359375
],
[
0.098876953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.576171875,
0.84228515625,
0.80615234375,
0.84228515625,
0.80615234375,
0.85693359375,
0.576171875,
0.85693359375
],
[
0.08087158203125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"cause",
"d ly,-yil is undefined.In practice, we observed",
"difficulties optimizing the error using SGD due to the con-",
"stant value of the gradient for 9 y (intuitively, because the",
"-",
"gradient doesn't decrease as the error decreases). We solve",
"Over-Prediction",
"E",
"EB",
"this by replacing the Li norm with L2: E, = (yi - yi)?,",
"which we find to be much easier to optimize.",
"·",
"The second problem is that constructing A (not to men-",
"",
"..",
"tion computing EA) has O(n2) time and space complexity.",
"However, we can overcome this problem by leveraging con-",
"Under-Prediction",
"E",
"EA",
"volutions. When we unpack the definition of matrix multi-",
"plication in Equation (3), we can write EA at pixel i as:",
"GT",
"Figure 3: A vizualization on a synthetic example of how",
"our loss function re-weights mistakes. In the top row we",
"(EA)i=",
"[E,A,i] ",
"visualize the down-weighting of false positives near the",
"true object border. In the bottom-row we show the down-",
"(7)",
"=yiyiE,",
"3",
"1 + (1- yi)",
"weighting of false-negatives do not spatially co-occur many",
"V2元g",
"other false-negatives and the increased weight of false neg-",
"3",
"atives which spatially co-occur. True positives are marked",
"Note that if d(i,)≥ 4o then",
"d(t.j)",
"by green, and mistakes are marked in red.",
"2",
"~ 0. We",
"then define yp.g and Ep.g as the ground truth and error re-",
"spectively at pixel (p, q), and can approximate (EA); as:",
"posed by [8]. Distractors are defined as visually salient",
"parts of an image which are not the photographer's intended",
"(EA);~ (1 -yi)+",
"focus.Thistaskis somewhat similarto salientobjectdetec",
"[y . E]i+(p,g ",
"Vp2+g2",
"tion, but successful algorithms must go beyond simply de",
"+ yi+(p,g)",
"",
"2",
"tecting all salient objects, and model the image at a global",
"/2元",
"P,q=-40",
"leveltodiscriminatebetweentheintendedfocusoftheim-",
"(8)",
"age and the distractors. In [8] Fried et al. propose an",
"SvM based approach, trained on a relatively small dataset,",
"If we let g=%,we can define a (20+1)× (20 +1)",
"which classifies super-pixels extracted by Multiscale Com-",
"Gaussian convolutional kernel K, yielding",
"binatorial Grouping (MCG) based on a set of hand-crafted",
"features. To test the robustness of their approach we gath-",
"ered a larger dataset with crowdsourced labels. While their",
"EA ~ Y((Y E) * KA)+ (1 -Y), where",
"1",
"(9)",
"V(-8+)2+(-$+3)2",
"method is able to detect large and well defined distractors, it",
"KA",
"struggles to detect non-object distractors such as shadows,",
"V2元g2",
"",
"lights and reflections as well as select small objects.",
"Where is element-wise multiplication. Now we don't",
"2.5.Portrait segmentation",
"need to store any entries of A, only a(20 +1) × (20 +1) ker-",
"nel and we skip most of the original summation over pixels",
"Portraits are highly popular art form in both photogra-",
"j. So the time and space complexity is reduced to O(n+92).",
"phy and painting. In most instances, artists seek to make",
"In practice, we use 0 = 9.",
"the subject stand out from its surrounding, for instance, by",
"making it brighter or sharper or by applying photographic",
"The final problem is that constructing B has complexity",
"O(n2), because computing , requires finding the mini-",
"or painterlyfiltersthatadapt tothe semantics of the image",
"Shen et al. [26l presented a new high quality automatic por-",
"mum d(i, j) over all pixels j, such that yj = 1. However,",
"if d(i,)≥25,then B=2-",
"ea~2inEquation (4)",
"trait segmentation algorithm by adapting the FCN-8s frame-",
"Our intuition is that B is modeling the region of uncertainty",
"work [20]. They also introduced a portrait image segmenta-",
"tion dataset and benchmark for training and testing.",
"about an object's boundary, for which we believe 25 pixels",
"to be too generous. So we redefine , as:",
"3.Our Approximate Fwloss (AFw)",
"if min, d(i,j) > 5",
"A",
"X",
"There are three issues that prevent the Fu metric as de-",
"min, d(i,j)?lyj = 1, otherwise",
"fined in Section 2.2 from being directly optimized as a loss",
"function. The first is that while the metric is differentiable",
"Squaring the distance so that the region of uncertainty is",
"almost everywhere, it is not differentiable when y = y, be-",
"assumed to be approximately 5 pixels instead of 25. Now"
] |
{
"arxiv_id": "1708.02212",
"language": "en",
"page_idx": 3
}
|
||
1708.02212-en-4
|
1708.02212-en-4.jpg
|
the time complexity is $O(n+\phi^{2})$ where $\phi=5.$ We can approximate $\hat{E}\odot\mathbb{B}$ using a convolution with the kernel $K^\mathbb{B}$, a tensor of size $\phi\times\phi\times\phi^2$, defined at each index(zero indexed) as:
$$K_{i,j,k}^{\mathbb{B}}=\begin{cases}(-\phi+i)^2+(-\phi+j)^2,&\text{if}k=i*(2\phi+1)+j\\0,&\text{otherwise}\end{cases}$$
and we can rewrite B as:
$$\mathbb{B}\approx Y+(1-Y)\odot e^{a\min_{k}[(Y*K^{\mathrm{B}})]}$$
(10)
By reformulating the local search for the true object boundary as a convolution followed by an argmin, we can leverage the efficient implmentations of these operations already available in many packages. While our current architecture does not suffer from speed or memory issues, more complicated architectures might benefit from a more optimized implementation, namely a custom'minimization convolution' that would not store the intermediary result of $(Y*K^{\mathbb{B}})$, and takes advantage of the sparsity of $K^\mathbb{B}.$
These changes yield a spatially informed loss function that can easily be implemented in an existing DNN framework such as Tensorflow. It fully utilizes the GPU, does not increase training wall clock time noticeably, and yields better results than more commonly used loss functions for foreground maps. Compared to an unoptimized implementation of the original formulation in python, our approximation takes two orders of magnitude less time to compute on the CPU, and three orders of magnitude less time on the GPU, to compute our loss on a 224x224 pixel image, See section 5.5.
## $4. \textbf{ Network Architecture}$
In order to produce accurate foreground maps each pixel must have a rich feature representation. To achieve this we utilize Zoomout features [23], which have been effectively utilized in the semantic segmentation community. Zoomout features are extracted from a CNN by upsampling and downsampling the features computed by each convolutional filter to be the same spatial resolution, then concatenating the features computed at all layers of the CNN. In this way, each spatial location is richly described by both the weakly localized semantic features computed at higher layers, the strongly localized edge and color detectors computed in the first layers, and everything in between.
## 4.1. Squeeze Layers
Zoomout features are expressive but have a large memory footprint, limiting the spatial resolution of predictions that can be made using them. In tasks like distractor detection, where the end goal is to precisely localize distractors
<FigureHere>
Figure 4: Architecture Diagram, note that because the blue squeeze module is applied to a fully connected layer, it uses ${\mathrm{only~1x1~convolutions.}}$
and remove them, a low resolution prediction leads to spatial ambiguity and lower precision. To remedy this problem we adapt the insights of [9], introducing what we call Squeeze Modules to our network. A Squeeze Module consists of $2n$ convolutional filters, $n$ of which are $1\times1$ convolutions and $n$ of which are $3\times3$ convolutions. Applying Squeeze Modules to each convolutional layer acts as a dimensionality reduction with learned parameters, allowing us to make predictions at essentially arbitrary resolutions by setting $n$ to be sufficiently small. In practice we produce $224\times224$ predictions, and set $n=64.$ We refer to our full architecture as a Squeezed Zoomout Network $(SZN).$
## $5. \textbf{Experiments}$
We report on experiments with three tasks where we can expect spatial sensitivity to be important for quality of the output: salient object detection, portrait segmentation, and distractor detection. See Sec. 2 for background.
### 5.1. Training Details
In all experiments we train our $SZN$ using a CNN (from which the squeezed zoomout features are derived) pre-trained on ImageNet. As the base CNN we use VGG- 16[28] for saliency, portraits, and distractors. We train the $SZN$ architecture using ADAM[14], and train in 3 stages. In the first stage we set the learning rate to 3e-4
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.72216796875,
0.09661865234375,
0.76708984375,
0.09661865234375,
0.76708984375,
0.111083984375,
0.72216796875,
0.111083984375
],
[
0.07843017578125,
0.106689453125,
0.468017578125,
0.1060791015625,
0.468017578125,
0.12060546875,
0.07843017578125,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.685546875,
0.125,
0.7197265625,
0.125,
0.7197265625,
0.136962890625,
0.685546875,
0.136962890625
],
[
0.080078125,
0.1395263671875,
0.16015625,
0.1395263671875,
0.16015625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.525390625,
0.15087890625,
0.59326171875,
0.149658203125,
0.59326171875,
0.1610107421875,
0.525390625,
0.1622314453125
],
[
0.85107421875,
0.15283203125,
0.865234375,
0.15283203125,
0.865234375,
0.1558837890625,
0.85107421875,
0.1558837890625
],
[
0.5556640625,
0.1641845703125,
0.60546875,
0.1641845703125,
0.60546875,
0.175537109375,
0.5556640625,
0.175537109375
],
[
0.81396484375,
0.1641845703125,
0.82421875,
0.1641845703125,
0.82421875,
0.1685791015625,
0.81396484375,
0.1685791015625
],
[
0.08251953125,
0.17041015625,
0.1298828125,
0.18115234375,
0.125,
0.1988525390625,
0.0767822265625,
0.1881103515625
],
[
0.1446533203125,
0.1654052734375,
0.491943359375,
0.1673583984375,
0.491943359375,
0.1849365234375,
0.1446533203125,
0.18310546875
],
[
0.1478271484375,
0.1875,
0.1690673828125,
0.1875,
0.1690673828125,
0.2020263671875,
0.1478271484375,
0.2020263671875
],
[
0.33251953125,
0.1875,
0.39794921875,
0.1875,
0.39794921875,
0.1988525390625,
0.33251953125,
0.1988525390625
],
[
0.5439453125,
0.1875,
0.61279296875,
0.1875,
0.61279296875,
0.1976318359375,
0.5439453125,
0.1976318359375
],
[
0.080078125,
0.2171630859375,
0.2418212890625,
0.2171630859375,
0.2418212890625,
0.2310791015625,
0.080078125,
0.2310791015625
],
[
0.81396484375,
0.222900390625,
0.826171875,
0.222900390625,
0.826171875,
0.2259521484375,
0.81396484375,
0.2259521484375
],
[
0.7587890625,
0.23486328125,
0.76708984375,
0.23486328125,
0.76708984375,
0.240478515625,
0.7587890625,
0.240478515625
],
[
0.8095703125,
0.2342529296875,
0.82861328125,
0.2342529296875,
0.82861328125,
0.2412109375,
0.8095703125,
0.2412109375
],
[
0.53369140625,
0.2442626953125,
0.57275390625,
0.2442626953125,
0.57275390625,
0.255615234375,
0.53369140625,
0.255615234375
],
[
0.58837890625,
0.2442626953125,
0.6259765625,
0.2442626953125,
0.6259765625,
0.255615234375,
0.58837890625,
0.255615234375
],
[
0.438720703125,
0.251953125,
0.4697265625,
0.251953125,
0.4697265625,
0.264404296875,
0.438720703125,
0.264404296875
],
[
0.098876953125,
0.27783203125,
0.4697265625,
0.27783203125,
0.4697265625,
0.291748046875,
0.098876953125,
0.291748046875
],
[
0.1502685546875,
0.25,
0.39453125,
0.2442626953125,
0.39453125,
0.261962890625,
0.1502685546875,
0.267578125
],
[
0.5537109375,
0.28271484375,
0.60693359375,
0.28271484375,
0.60693359375,
0.294189453125,
0.5537109375,
0.294189453125
],
[
0.080078125,
0.29345703125,
0.468017578125,
0.29345703125,
0.468017578125,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.560546875,
0.294189453125,
0.60546875,
0.294189453125,
0.60546875,
0.304931640625,
0.560546875,
0.304931640625
],
[
0.080078125,
0.30810546875,
0.468017578125,
0.30810546875,
0.468017578125,
0.3212890625,
0.080078125,
0.3212890625
],
[
0.81201171875,
0.31494140625,
0.82666015625,
0.31494140625,
0.82666015625,
0.3193359375,
0.81201171875,
0.3193359375
],
[
0.080078125,
0.323974609375,
0.468017578125,
0.323974609375,
0.468017578125,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.81201171875,
0.323974609375,
0.82666015625,
0.323974609375,
0.82666015625,
0.32763671875,
0.81201171875,
0.32763671875
],
[
0.080078125,
0.339111328125,
0.468017578125,
0.339111328125,
0.468017578125,
0.352783203125,
0.080078125,
0.352783203125
],
[
0.52294921875,
0.34228515625,
0.61767578125,
0.34228515625,
0.61767578125,
0.356689453125,
0.52294921875,
0.356689453125
],
[
0.080078125,
0.354248046875,
0.4697265625,
0.354248046875,
0.4697265625,
0.367431640625,
0.080078125,
0.367431640625
],
[
0.521484375,
0.354736328125,
0.62353515625,
0.356689453125,
0.62353515625,
0.371826171875,
0.521484375,
0.369384765625
],
[
0.07843017578125,
0.369384765625,
0.468017578125,
0.36865234375,
0.468017578125,
0.3818359375,
0.07843017578125,
0.38330078125
],
[
0.521484375,
0.369384765625,
0.65185546875,
0.369384765625,
0.65185546875,
0.38330078125,
0.521484375,
0.38330078125
],
[
0.08087158203125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.397705078125,
0.08087158203125,
0.397705078125
],
[
0.7587890625,
0.3818359375,
0.80224609375,
0.3818359375,
0.80224609375,
0.393310546875,
0.7587890625,
0.393310546875
],
[
0.08087158203125,
0.3984375,
0.428955078125,
0.3984375,
0.428955078125,
0.412353515625,
0.08087158203125,
0.412353515625
],
[
0.1004638671875,
0.415283203125,
0.468017578125,
0.415283203125,
0.468017578125,
0.4287109375,
0.1004638671875,
0.4287109375
],
[
0.50341796875,
0.421630859375,
0.89208984375,
0.421630859375,
0.89208984375,
0.43505859375,
0.50341796875,
0.43505859375
],
[
0.080078125,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.08087158203125,
0.4462890625,
0.468017578125,
0.4462890625,
0.468017578125,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.50341796875,
0.453857421875,
0.65283203125,
0.453857421875,
0.65283203125,
0.464111328125,
0.50341796875,
0.464111328125
],
[
0.08087158203125,
0.462158203125,
0.46728515625,
0.462158203125,
0.46728515625,
0.472900390625,
0.08087158203125,
0.472900390625
],
[
0.08087158203125,
0.4765625,
0.46728515625,
0.4765625,
0.46728515625,
0.4873046875,
0.08087158203125,
0.4873046875
],
[
0.080078125,
0.490478515625,
0.468017578125,
0.490478515625,
0.468017578125,
0.50390625,
0.080078125,
0.50390625
],
[
0.5009765625,
0.49365234375,
0.89208984375,
0.494384765625,
0.89208984375,
0.5087890625,
0.5009765625,
0.50830078125
],
[
0.080078125,
0.505859375,
0.4697265625,
0.505859375,
0.4697265625,
0.51953125,
0.080078125,
0.51953125
],
[
0.50341796875,
0.51220703125,
0.8896484375,
0.51220703125,
0.8896484375,
0.52294921875,
0.50341796875,
0.52294921875
],
[
0.07843017578125,
0.51953125,
0.4697265625,
0.52099609375,
0.4697265625,
0.53515625,
0.07843017578125,
0.5341796875
],
[
0.50341796875,
0.5244140625,
0.890625,
0.5244140625,
0.890625,
0.53857421875,
0.50341796875,
0.53857421875
],
[
0.080078125,
0.53515625,
0.4697265625,
0.53515625,
0.4697265625,
0.548828125,
0.080078125,
0.548828125
],
[
0.50341796875,
0.54052734375,
0.89208984375,
0.54052734375,
0.89208984375,
0.5537109375,
0.50341796875,
0.5537109375
],
[
0.07843017578125,
0.5498046875,
0.4697265625,
0.55029296875,
0.4697265625,
0.56494140625,
0.07843017578125,
0.564453125
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.56640625,
0.156005859375,
0.56640625,
0.156005859375,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.50244140625,
0.5693359375,
0.89208984375,
0.57080078125,
0.89208984375,
0.583984375,
0.50244140625,
0.58349609375
],
[
0.50341796875,
0.58544921875,
0.8896484375,
0.58544921875,
0.8896484375,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.08087158203125,
0.59765625,
0.28515625,
0.59765625,
0.28515625,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.50244140625,
0.599609375,
0.890625,
0.60107421875,
0.890625,
0.61572265625,
0.50244140625,
0.6142578125
],
[
0.50341796875,
0.6162109375,
0.89208984375,
0.6162109375,
0.89208984375,
0.6298828125,
0.50341796875,
0.6298828125
],
[
0.098876953125,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.098876953125,
0.63720703125
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.0767822265625,
0.63720703125,
0.4697265625,
0.63818359375,
0.4697265625,
0.65283203125,
0.0767822265625,
0.6513671875
],
[
0.5009765625,
0.64453125,
0.89208984375,
0.64599609375,
0.89208984375,
0.66015625,
0.5009765625,
0.6591796875
],
[
0.080078125,
0.65380859375,
0.468017578125,
0.65380859375,
0.468017578125,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.50341796875,
0.662109375,
0.85107421875,
0.662109375,
0.85107421875,
0.67236328125,
0.50341796875,
0.67236328125
],
[
0.08087158203125,
0.67041015625,
0.465576171875,
0.67041015625,
0.465576171875,
0.68115234375,
0.08087158203125,
0.68115234375
],
[
0.07843017578125,
0.68310546875,
0.4697265625,
0.6845703125,
0.4697265625,
0.69873046875,
0.07843017578125,
0.69775390625
],
[
0.50244140625,
0.689453125,
0.63134765625,
0.69140625,
0.63134765625,
0.70556640625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.69970703125,
0.468017578125,
0.69970703125,
0.468017578125,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.080078125,
0.7138671875,
0.4697265625,
0.7138671875,
0.4697265625,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.521484375,
0.71533203125,
0.89208984375,
0.71630859375,
0.89208984375,
0.73095703125,
0.521484375,
0.72998046875
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72998046875,
0.468017578125,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.50244140625,
0.73193359375,
0.89208984375,
0.73193359375,
0.89208984375,
0.74560546875,
0.50244140625,
0.74560546875
],
[
0.080078125,
0.744140625,
0.468017578125,
0.744140625,
0.468017578125,
0.75830078125,
0.080078125,
0.75830078125
],
[
0.5009765625,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76123046875,
0.5009765625,
0.76123046875
],
[
0.080078125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.77294921875,
0.080078125,
0.77294921875
],
[
0.50244140625,
0.76220703125,
0.814453125,
0.76220703125,
0.814453125,
0.77587890625,
0.50244140625,
0.77587890625
],
[
0.07843017578125,
0.7734375,
0.46728515625,
0.7744140625,
0.46728515625,
0.7890625,
0.07843017578125,
0.78857421875
],
[
0.080078125,
0.7890625,
0.414306640625,
0.7890625,
0.414306640625,
0.80322265625,
0.080078125,
0.80322265625
],
[
0.50341796875,
0.787109375,
0.66015625,
0.787109375,
0.66015625,
0.8017578125,
0.50341796875,
0.8017578125
],
[
0.521484375,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82470703125,
0.521484375,
0.82470703125
],
[
0.08087158203125,
0.8173828125,
0.233642578125,
0.8173828125,
0.233642578125,
0.83056640625,
0.08087158203125,
0.83056640625
],
[
0.50341796875,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.098876953125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85498046875,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.50244140625,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"the time complexity is O(n + @2) where Φ = 5. We can",
"WxHx1",
"approximate E B using a convolution with thekernel KB,",
"a tensor of size Φ × Φ × Φ?, defined at each index (zero",
"CNN",
"indexed)as:",
"Squeeze Module",
"",
"WxHx2n",
"E6",
"Kiik",
"[(-Φ+i)2 +(-Φ+i)2,ifk=i*(2Φ+1)+j",
"1 0,",
"otherwise",
"Upsampling",
"and we can rewrite B as:",
"O",
"-",
"0=64",
"n3x3",
"n1x1",
"(10)",
"By reformulating the local search for the true object",
"B~Y+ (1-Y)Oeamin[(Y*K)",
"WxHxC",
"boundary as a convolution followed by an argmin, we can",
"Input",
"leverage the efficient implmentations of these operations al-",
"E64",
"ready available in many packages. While our current ar-",
"1=62",
"chitecture does not suffer from speed or memory issues,",
"Block Legend:",
"more complicated architectures might benefit from a more",
"Width=Spatial",
"optimizedimplementation,namelyacustomminimization",
"Height=#Features",
"convolution' that would not store the intermediary result of",
"WxHx3",
"(Y * K), and takes advantage of the sparsity of KB",
"These changes yield a spatially informed loss function",
"Figure 4: Architecture Diagram, note that because the blue",
"that can easily be implemented in an existing DNN frame-",
"squeeze module is applied to a fully connected layer, it uses",
"worksuchasTensorflow.Itfully utilizestheGPU.does",
"onlylxl convolutions.",
"not increase training wall clock timenoticeably,and yields",
"betterresultsthanmorecommonlyusedlossfunctionsfor",
"foreground maps. Compared to an unoptimized implemen-",
"and remove them, a low resolution prediction leads to spa-",
"tation of the original formulation in python, our approxi-",
"tial ambiguity and lowerprecision.To remedy thisprob-",
"mation takes two orders of magnitude less time to compute",
"lem we adapt the insights of [9], introducing what we call",
"on the CPU, and three orders of magnitude less time on the",
"Squeeze Modules to our network. A Squeeze Module con-",
"GPU, to compute our loss on a 224x224 pixel image, See",
"sists of 2n convolutional filters, n of which are 1 × 1 con-",
"section 5.5.",
"volutions and n of which are 3×3 convolutions.Applying",
"Squeeze Modules to each convolutional layer acts as a di-",
"4.NetworkArchitecture",
"mensionality reduction with learned parameters, allowing",
"us to make predictions at essentially arbitrary resolutions",
"In order to produce accurate foreground maps each pixel",
"by setting n to be sufficiently small. In practice we produce",
"must have a rich feature representation. To achieve this",
"224 × 224 predictions, and set n = 64. We refer to our full",
"we utilize Zoomout features [23], which have been ef-",
"architectureasaSaueezedZoomoutNetwork(SZN)",
"fectivelyutilizedinthe semantic segmentationcommunity",
"Zoomout features are extracted from a CNN by upsampling",
"5.Experiments",
"and downsampling the features computed by each convolu-",
"tional filter to be the same spatial resolution, then concate-",
"We report on experiments with three tasks where we can",
"nating the features computed at all layers of the CNN. In",
"expect spatial sensitivity to be important for quality of the",
"this way, each spatial location is richly described by both",
"output: salient object detection, portrait segmentation, and",
"the weakly localized semantic features computed at higher",
"distractor detection. See Sec. 2 for background.",
"layers, the strongly localized edge and color detectors com-",
"puted in the first layers, and everything in between",
"5.1. Training Details",
"In all experiments we train our SZN using a CNN",
"4.1. Squeeze Layers",
"(from which the squeezed zoomout features are derived)",
"Zoomout features are expressive but have a large mem-",
"pre-trained on ImageNet. As the base CNN we use VGG-",
"ory footprint, limiting the spatial resolution of predictions",
"16 [28] for saliency,portraits, and distractors.We train",
"that can be made using them. In tasks like distractor detec-",
"the SZN architecture using ADAM [14], and train in 3",
"tion, where the end goal is to precisely localize distractors",
"stages. In the first stage we set the learning rate to 3e-4"
] |
{
"arxiv_id": "1708.02212",
"language": "en",
"page_idx": 4
}
|
||
2212.04273-en-0
|
2212.04273-en-0.jpg
|
## Better Hit the Nail on the Head than Beat around the Bush: $\textbf{Removing Protected Attributes with a Single Projection}$
Pantea Haghighatkhah$^{\textcircled{\mathrm{o}}}$ Antske Fokkens$^{\clubsuit\diamondsuit}$ Pia Sommerauer$^{\textcircled{\mathrm{o}}}$
$\bar{\textbf{Bettina Speckmann}}^{\textcircled{\mathrm{o}}}$ Kevin Verbeek$^{\textcircled{\mathrm{o}}}$
◇TU Eindhoven, Department of Mathematics and Computer Science
$\clubsuit$ Vrije Universiteit Amsterdam, Computational Linguistics Text Mining Lab
### $\mathbf{Abstract}$
Bias elimination and recent probing studies attempt to remove specific information from embedding spaces. Here it is important to remove as much of the target information as possible, while preserving any other information present. INLP is a popular recent method which removes specific information through iterative nullspace projections. Multiple iterations, however, increase the risk that information other than the target is negatively affected. We introduce two methods that find a single targeted projection: Mean Projection (MP, more efficient) and Tukey Median Projection (TMP, with theoretical guarantees). Our comparison between MP and INLP shows that (1) one MP projection removes linear separability based on the target and (2) MP has less impact on the overall space. Further analysis shows that applying random projections after MP leads to the same overall effects on the embedding space as the multiple projections of INLP. Applying one targeted (MP) projection hence is methodologically cleaner than applying multiple (INLP) projections that introduce random effects.
bias, or because we want to test how the attribute
was affecting a system in a probing setup.
There are various approaches to remove a particular attribute from an embedding. Clearly it is important to remove as much of the target attribute as possible while preserving other information present. Arguably linear projections are one of the least intrusive methods to transform an embedding space. Iterative nullspace projection (Ravfogel et al., 2020, INLP) is a recent popular method which follows this paradigm and uses multiple linear projections. INLP achieves linear guarding of the protected attribute, that is, a linear classifier can no longer separate instances that have the attribute from instances that do not.
However, it remains unclear to what extent INLP affects the remainder of the embedding space. In general it requires 10-15 iterations (projections) to achieve linear guarding; it is hence likely that other attributes are negatively affected as well. This is particularly problematic in settings where INLP is used to determine what impact a specific attribute has on overall systems.
In this paper, we introduce two methods that find a single targeted projection which achieves linear guarding at once. Specifically, our Mean Projection[MP] method uses the mean of the points in each class to determine the projection direction. The mean is more robust to unbalanced class distributions than INLP. The mean can, however, be overly sensitive to outliers. We prove that our second method, the Tukey Median Projection [TMP], finds nearly worst-case optimal projections for any input. Unfortunately, computing the Tukey median (Herbert Edelsbrunner, 1987) is computationally expensive in high dimensions. In our experiments we hence compared only MP to INLP. Specifically, we carried out the gender debiasing experiments of Ravfogel et al. (2020) with both methods. They show that
1. MP only needs one projection for linearly
## 1 Introduction
Word embedding spaces can contain rich information that is valuable for a wide range of NLP tasks. High quality word embeddings should capture the semantic attributes associated with a word's meaning. Though we can establish that points representing words with similar semantic attributes tend to be close to each other, it is challenging to reason which attributes are captured to what extent and, in particular, which dimensions capture these attributes (Sommerauer and Fokkens, 2018). But even though we do not (yet) know exactly which attributes are represented in which way, we nevertheless may want to ensure that one particular attribute is no longer present in the embedding. For example, because this attribute is a protected attribute such as gender which can lead to harmful
|
[
[
0.197265625,
0.0914306640625,
0.80419921875,
0.0914306640625,
0.80419921875,
0.105712890625,
0.197265625,
0.105712890625
],
[
0.214111328125,
0.11102294921875,
0.7841796875,
0.11102294921875,
0.7841796875,
0.1258544921875,
0.214111328125,
0.1258544921875
],
[
0.22412109375,
0.142578125,
0.77490234375,
0.142578125,
0.77490234375,
0.1602783203125,
0.22412109375,
0.1602783203125
],
[
0.334228515625,
0.160888671875,
0.66748046875,
0.160888671875,
0.66748046875,
0.17578125,
0.334228515625,
0.17578125
],
[
0.22412109375,
0.1923828125,
0.78076171875,
0.193603515625,
0.78076171875,
0.2078857421875,
0.22412109375,
0.207275390625
],
[
0.19140625,
0.21142578125,
0.8134765625,
0.21142578125,
0.8134765625,
0.2244873046875,
0.19140625,
0.2244873046875
],
[
0.26123046875,
0.2529296875,
0.340087890625,
0.2529296875,
0.340087890625,
0.267333984375,
0.26123046875,
0.267333984375
],
[
0.51318359375,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.1435546875,
0.273681640625,
0.459228515625,
0.27490234375,
0.459228515625,
0.289306640625,
0.1435546875,
0.2880859375
],
[
0.51123046875,
0.26953125,
0.82373046875,
0.270751953125,
0.82373046875,
0.28515625,
0.51123046875,
0.283935546875
],
[
0.5283203125,
0.28515625,
0.88330078125,
0.2861328125,
0.88330078125,
0.300537109375,
0.5283203125,
0.2998046875
],
[
0.14697265625,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3017578125,
0.14697265625,
0.3017578125
],
[
0.145263671875,
0.303955078125,
0.4609375,
0.303955078125,
0.4609375,
0.317626953125,
0.145263671875,
0.317626953125
],
[
0.5087890625,
0.300537109375,
0.880859375,
0.300537109375,
0.880859375,
0.314208984375,
0.5087890625,
0.314208984375
],
[
0.0235137939453125,
0.31005859375,
0.060455322265625,
0.31005859375,
0.060455322265625,
0.412109375,
0.0235137939453125,
0.412109375
],
[
0.14697265625,
0.318359375,
0.45751953125,
0.318359375,
0.45751953125,
0.33203125,
0.14697265625,
0.33203125
],
[
0.509765625,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.509765625,
0.33203125
],
[
0.145263671875,
0.333251953125,
0.459228515625,
0.333251953125,
0.459228515625,
0.3466796875,
0.145263671875,
0.3466796875
],
[
0.51123046875,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51123046875,
0.3486328125
],
[
0.145263671875,
0.3466796875,
0.459228515625,
0.3466796875,
0.459228515625,
0.35986328125,
0.145263671875,
0.35986328125
],
[
0.51318359375,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.145263671875,
0.361083984375,
0.459228515625,
0.361083984375,
0.459228515625,
0.374755859375,
0.145263671875,
0.374755859375
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.37890625,
0.51318359375,
0.37890625
],
[
0.145263671875,
0.3759765625,
0.459228515625,
0.3759765625,
0.459228515625,
0.3896484375,
0.145263671875,
0.3896484375
],
[
0.145263671875,
0.388427734375,
0.45751953125,
0.388427734375,
0.45751953125,
0.402099609375,
0.145263671875,
0.402099609375
],
[
0.51318359375,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.51318359375,
0.39794921875
],
[
0.51318359375,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.0277099609375,
0.407958984375,
0.056243896484375,
0.407958984375,
0.056243896484375,
0.427001953125,
0.0277099609375,
0.427001953125
],
[
0.1485595703125,
0.405029296875,
0.459228515625,
0.405029296875,
0.459228515625,
0.4150390625,
0.1485595703125,
0.4150390625
],
[
0.145263671875,
0.41748046875,
0.459228515625,
0.41748046875,
0.459228515625,
0.43115234375,
0.145263671875,
0.43115234375
],
[
0.51123046875,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.145263671875,
0.431640625,
0.459228515625,
0.431640625,
0.459228515625,
0.4453125,
0.145263671875,
0.4453125
],
[
0.51123046875,
0.43115234375,
0.880859375,
0.43115234375,
0.880859375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.14697265625,
0.44775390625,
0.45751953125,
0.44775390625,
0.45751953125,
0.457763671875,
0.14697265625,
0.457763671875
],
[
0.0201568603515625,
0.43408203125,
0.05877685546875,
0.432861328125,
0.06378173828125,
0.51025390625,
0.0251922607421875,
0.51123046875
],
[
0.509765625,
0.44775390625,
0.880859375,
0.44775390625,
0.880859375,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.145263671875,
0.460205078125,
0.45751953125,
0.460205078125,
0.45751953125,
0.473876953125,
0.145263671875,
0.473876953125
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.14697265625,
0.474365234375,
0.459228515625,
0.474365234375,
0.459228515625,
0.488037109375,
0.14697265625,
0.488037109375
],
[
0.51123046875,
0.479736328125,
0.7109375,
0.479736328125,
0.7109375,
0.493408203125,
0.51123046875,
0.493408203125
],
[
0.1435546875,
0.488037109375,
0.459228515625,
0.4892578125,
0.459228515625,
0.5029296875,
0.1435546875,
0.501953125
],
[
0.5283203125,
0.495361328125,
0.880859375,
0.495361328125,
0.880859375,
0.5087890625,
0.5283203125,
0.5087890625
],
[
0.14697265625,
0.5029296875,
0.45751953125,
0.5029296875,
0.45751953125,
0.5166015625,
0.14697265625,
0.5166015625
],
[
0.019317626953125,
0.51416015625,
0.060455322265625,
0.51416015625,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.145263671875,
0.51708984375,
0.45751953125,
0.51708984375,
0.45751953125,
0.53076171875,
0.145263671875,
0.53076171875
],
[
0.51123046875,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51123046875,
0.525390625
],
[
0.14697265625,
0.5322265625,
0.459228515625,
0.5322265625,
0.459228515625,
0.5458984375,
0.14697265625,
0.5458984375
],
[
0.51123046875,
0.52783203125,
0.8818359375,
0.52685546875,
0.8818359375,
0.54150390625,
0.51123046875,
0.5419921875
],
[
0.1485595703125,
0.54638671875,
0.45751953125,
0.54638671875,
0.45751953125,
0.55712890625,
0.1485595703125,
0.55712890625
],
[
0.509765625,
0.54345703125,
0.8818359375,
0.5419921875,
0.8818359375,
0.55712890625,
0.509765625,
0.55810546875
],
[
0.14697265625,
0.55859375,
0.459228515625,
0.55859375,
0.459228515625,
0.57373046875,
0.14697265625,
0.57373046875
],
[
0.509765625,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.509765625,
0.572265625
],
[
0.145263671875,
0.57373046875,
0.459228515625,
0.57373046875,
0.459228515625,
0.58740234375,
0.145263671875,
0.58740234375
],
[
0.51123046875,
0.576171875,
0.8818359375,
0.57470703125,
0.8818359375,
0.5888671875,
0.51123046875,
0.59033203125
],
[
0.145263671875,
0.58837890625,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.145263671875,
0.6015625
],
[
0.509765625,
0.59130859375,
0.880859375,
0.59130859375,
0.880859375,
0.60498046875,
0.509765625,
0.60498046875
],
[
0.14697265625,
0.6015625,
0.459228515625,
0.6015625,
0.459228515625,
0.6162109375,
0.14697265625,
0.6162109375
],
[
0.51123046875,
0.60791015625,
0.68408203125,
0.60791015625,
0.68408203125,
0.62158203125,
0.51123046875,
0.62158203125
],
[
0.14697265625,
0.61767578125,
0.252685546875,
0.61767578125,
0.252685546875,
0.62890625,
0.14697265625,
0.62890625
],
[
0.52880859375,
0.62451171875,
0.8818359375,
0.62451171875,
0.8818359375,
0.63818359375,
0.52880859375,
0.63818359375
],
[
0.139404296875,
0.63720703125,
0.26025390625,
0.63720703125,
0.26025390625,
0.65185546875,
0.139404296875,
0.65185546875
],
[
0.51123046875,
0.6396484375,
0.880859375,
0.6396484375,
0.880859375,
0.65283203125,
0.51123046875,
0.65283203125
],
[
0.11419677734375,
0.66015625,
0.491943359375,
0.6611328125,
0.491943359375,
0.6787109375,
0.11419677734375,
0.677734375
],
[
0.51123046875,
0.65673828125,
0.8818359375,
0.65673828125,
0.8818359375,
0.67041015625,
0.51123046875,
0.67041015625
],
[
0.51123046875,
0.6728515625,
0.8818359375,
0.6728515625,
0.8818359375,
0.68603515625,
0.51123046875,
0.68603515625
],
[
0.1175537109375,
0.6796875,
0.489501953125,
0.6796875,
0.489501953125,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.509765625,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.509765625,
0.70166015625
],
[
0.1175537109375,
0.6953125,
0.489501953125,
0.6953125,
0.489501953125,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.51123046875,
0.705078125,
0.88330078125,
0.705078125,
0.88330078125,
0.71875,
0.51123046875,
0.71875
],
[
0.118408203125,
0.71142578125,
0.490234375,
0.71142578125,
0.490234375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51318359375,
0.72021484375,
0.88330078125,
0.72021484375,
0.88330078125,
0.73388671875,
0.51318359375,
0.73388671875
],
[
0.1158447265625,
0.72705078125,
0.489501953125,
0.72705078125,
0.489501953125,
0.74072265625,
0.1158447265625,
0.74072265625
],
[
0.51123046875,
0.736328125,
0.880859375,
0.736328125,
0.880859375,
0.75,
0.51123046875,
0.75
],
[
0.1175537109375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.5146484375,
0.7529296875,
0.8818359375,
0.7529296875,
0.8818359375,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.1158447265625,
0.76025390625,
0.490234375,
0.76025390625,
0.490234375,
0.77392578125,
0.1158447265625,
0.77392578125
],
[
0.51318359375,
0.7685546875,
0.880859375,
0.7685546875,
0.880859375,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.1158447265625,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.1158447265625,
0.7880859375
],
[
0.51123046875,
0.78369140625,
0.87890625,
0.78369140625,
0.87890625,
0.79736328125,
0.51123046875,
0.79736328125
],
[
0.1158447265625,
0.79052734375,
0.486083984375,
0.79052734375,
0.486083984375,
0.80419921875,
0.1158447265625,
0.80419921875
],
[
0.51123046875,
0.7998046875,
0.8818359375,
0.80029296875,
0.8818359375,
0.8154296875,
0.51123046875,
0.81396484375
],
[
0.1158447265625,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.1158447265625,
0.82080078125
],
[
0.51318359375,
0.81787109375,
0.8818359375,
0.81787109375,
0.8818359375,
0.83203125,
0.51318359375,
0.83203125
],
[
0.11419677734375,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.11419677734375,
0.83837890625
],
[
0.51123046875,
0.83447265625,
0.8818359375,
0.83447265625,
0.8818359375,
0.84716796875,
0.51123046875,
0.84716796875
],
[
0.1175537109375,
0.84033203125,
0.490234375,
0.84033203125,
0.490234375,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.51123046875,
0.849609375,
0.8818359375,
0.849609375,
0.8818359375,
0.86328125,
0.51123046875,
0.86328125
],
[
0.1158447265625,
0.85546875,
0.490234375,
0.85498046875,
0.490234375,
0.869140625,
0.1158447265625,
0.87060546875
],
[
0.51123046875,
0.8662109375,
0.8818359375,
0.8662109375,
0.8818359375,
0.8798828125,
0.51123046875,
0.8798828125
],
[
0.1158447265625,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.8818359375,
0.7001953125,
0.8818359375,
0.7001953125,
0.8955078125,
0.509765625,
0.8955078125
],
[
0.1175537109375,
0.8896484375,
0.491943359375,
0.8896484375,
0.491943359375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.5263671875,
0.90380859375,
0.880859375,
0.90380859375,
0.880859375,
0.91845703125,
0.5263671875,
0.91845703125
]
] |
[
"Better Hit the Nail on the Head than Beat around the Bush:",
"Removing Protected Attributes with a Single Projection",
"Pantea HaghighatkhahAntske Fokkens* Pia Sommerauer*",
"Bettina SpeckmannKevin Verbeek",
"ATU Eindhoven, Department of Mathematics and Computer Science",
" Vrije Universiteit Amsterdam, Computational Linguistics Text Mining Lab",
"Abstract",
"bias, or because we want to test how the attribute",
"Bias elimination and recent probing studies at-",
"was affecting a system in a probing setup.",
"There are various approaches to remove a par-",
"tempttoremovespecificinformationfromem-",
"bedding spaces.Here it is important to re-",
"ticular attribute from an embedding. Clearly it",
"",
"move as much of the target information as",
"is important to remove as much of the target at",
"possible, while preserving any other informa-",
"tribute as possible while preserving other informa-",
"tion present. INLP is a popular recent method",
"tionpresent.Arguablylinearprojections areoneof",
"which removes specific information through",
"the least intrusive methods totransforman embed-",
"iterative nullspace projections:Multiple iter-",
"ations, however, increase the risk that infor",
"ding space. Iterative nullspace projection (Ravfo-",
"gel et al., 2020, INLP) is a recent popular method",
"8",
"mation otherthan thetargetisnegatively af-",
"fected. We introduce two methods that find",
"which follows this paradigm and uses multiple lin-",
"a single targeted projection: Mean Projection",
"ear projections. INLP achieves linear guarding of",
"(MP,moreefficient)andTukeyMedianProjec-",
"[DT's]",
"the protected attribute, that is, a linear classifier can",
"tion (TMP, with theoretical guarantees). Our",
"no longer separate instances that have the attribute",
"comparison between MP and INLP shows that",
"from instances that do not.",
"(1) one MP projection removes linear separa-",
"However, it remains unclear to what extent INLP",
"bility based on the target and (2) MP has less",
"",
"impact on the overall space. Further analysis",
"affects the remainder of the embedding space. In",
"shows that applying random projections after",
"general it requires 10-15 iterations (projections) to",
"MPleadstothesameoveralleffectsontheem-",
"achieve linear guarding; it is hence likely that other",
"bedding space as the multiple projections of",
"attributes are negatively affected as well. This is",
"INLP. Applying one targeted (MP) projection",
"particularly problematic in settings where INLP is",
"hence is methodologically cleaner than apply-",
" used to determine what impact a specific attribute",
"ing multiple (INLP) projections that introduce",
"has on overall systems.",
"random effects.",
"In this paper, we introduce two methods that",
"Introduction",
"find a single targeted projection which achieves",
"Word embedding spaces can contain rich informa-",
"linear guarding at once. Specifically, our Mean",
"Projection [MP] method uses the mean of the",
"tion that is valuable for a wide range of NLP tasks.",
"points in each class to determine the projection",
"High quality word embeddings should capture the",
"direction. The mean is more robust to unbalanced",
"semantic attributes associated with a word's mean-",
"class distributions than INLP. The mean can, how-",
"ing. Though we can establish that points repre-",
"ever, be overly sensitive to outliers. We prove that",
"senting words with similar semantic attributes tend",
"our second method, the Tukey Median Projection",
"to be close to each other, it is challenging to rea-",
"[TMPl, finds nearly worst-case optimal projections",
"son which attributes are captured to what extent",
"for any input. Unfortunately, computing the Tukey",
"and, in particular, which dimensions capture these",
"median (Herbert Edelsbrunner, 1987) is compu-",
"attributes (Sommerauer and Fokkens, 2018). But",
"tationally expensive in high dimensions.In our",
"even though we do not (yet) know exactly which",
"experiments we hence compared only MP to INLP",
"attributes are represented in which way, we nev-",
"Specifically, we carried out the gender debiasing",
"ertheless may want to ensure that one particular",
"experiments of Ravfogel et al. (2020) with both",
"attribute is no longer present in the embedding. For",
"methods. They show that",
"example, because this attribute is a protected at-",
"tribute such as gender which can lead to harmful",
"1. MP only needs one projection for linearly"
] |
{
"arxiv_id": "2212.04273",
"language": "en",
"page_idx": 0
}
|
||
2212.04273-en-1
|
2212.04273-en-1.jpg
|
guarding gender where INLP needs 10-15.
2. MP has less impact on other aspects of the
embedding space.
INLP projections improve simlex scores and reduce WEAT scores more than MP. A priori it is unclear why their many projections should have this effect. We investigated and show that the same improvements appear after applying random projections (either after MP or after the first INLP projections). Applying one MP projection to linearly guard an attribute is hence methodologically cleaner than applying multiple (INLP) projections.
## 2 Related Work
Multiple methods for removing bias from embeddings have been suggested. Bias can be addressed at the level of the training data (Zhao et al., 2018a), the training process (Zhang et al., 2018; Madras et al., 2018), and the resulting model itself (Ravfogel et al., 2020; Bolukbasi et al., 2016; Vargas and Cotterell, 2020). Debiasing an existing model has the clear advantage that it can be done with relatively little data and that the model does not need to be retrained. In this paper, we focus on removing bias through transforming the embedding space. Transformations of embedding spaces can be performed by means of linear projections (Bolukbasi et al., 2016), multiple linear projections (Ravfogel et al., 2020), or via non-linear kernels (Vargas and Cotterell, 2020). Ravfogel et al. (2020) introduce a method called Iterative Nullspace Projection (INLP) where they attempt to remove the bias by projecting points iteratively. Our methods achieve linear guarding after a single linear projection. As such we directly improve upon the INLP method, which we describe in detail in Section 3. In the remainder of this section, we focus on more recent post-training approaches and applications of INLP.
Similar to Ravfogel et al. (2020), Ravfogel et al. (2022) aim to find a linear subspace of the data such that removing that subspace using projection removes bias in the data optimally. Doing so necessarily requires a metric to measure how well bias has been removed by a specific projection. Ravfogel et al. (2022) use the classification loss of various classifers as metrics. That is, a higher classification loss after projection corresponds to better bias removal. This directly translates into a minimax optimization problem, which the authors refer to as a minimax game. The authors present
different strategies to solve this optimization problem for their chosen classifiers. A similar strategy was previously described by Haghighatkhah et al. (2021,2022). Here the corresponding minimax optimization problem is referred to as“maximizing inseparability". Haghighatkhah et al. (2022) show that this problem can be solved efficiently under certain convexity assumptions on the metric.
Zhang et al. (2022) independently developed a projection method which is very similar to our MP. Their method also uses class means to find the projection direction, but proceeds via a projection to the origin, and hence needs one additional projection. The goal of the paper, however, is quite different, namely the study of human brain reactions to syntactic and semantic features of words. As such, it does not include a systematic comparison to INLP or a comprehensive analysis of the impact on the embedding space.
Shao et al. (2022) in recent and unpublished work compute linear projection vectors by minimizing the covariance between biased word vectors and the protected attribute. They generally need two projections to achieve comparable or better performance to INLP. The paper does not report on exactly the same experiments as we do, but for all experiments reported in both papers, our method MP performs similarly or better.
Dev et al. (2021a) use a different kind of linear transformation on embedding spaces, namely rotation. Their goal is to disentangle two particular attributes, such as gender and occupation. To do so, they rotate the embedding space in such as way that the subspaces corresponding to the two attributes become orthogonal. On the positive side, their approach does not remove a dimension and hence arguably retains more information in the embedding. However, it will not remove all bias with respect to gender, but only gender bias with respect to a single specifed attribute, e.g., occupation. More generally, it will remove the interaction between the two specified attributes and not actually completely remove an attribute from the embedding.
INLP has recently gained importance for probing studies. Probing has been criticized because (1) results are difficult to interpret, and (2) it is usually not possible to test whether the target attribute is also relevant for downstream tasks. Elazar et al. (2021) propose amnesic probing, which employs INLP to remove the target attribute and then tests change in performance of the adapted embeddings
|
[
[
0.15283203125,
0.08612060546875,
0.466796875,
0.0843505859375,
0.466796875,
0.09857177734375,
0.15283203125,
0.1009521484375
],
[
0.51318359375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.509765625,
0.1015625,
0.87890625,
0.1015625,
0.87890625,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
0.130126953125,
0.1134033203125,
0.48779296875,
0.1134033203125,
0.48779296875,
0.127685546875,
0.130126953125,
0.127685546875
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.15283203125,
0.1300048828125,
0.28955078125,
0.1300048828125,
0.28955078125,
0.144287109375,
0.15283203125,
0.144287109375
],
[
0.509765625,
0.133056640625,
0.880859375,
0.1341552734375,
0.880859375,
0.1484375,
0.509765625,
0.1478271484375
],
[
0.118408203125,
0.1561279296875,
0.486083984375,
0.1561279296875,
0.486083984375,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.51123046875,
0.15087890625,
0.880859375,
0.15087890625,
0.880859375,
0.1644287109375,
0.51123046875,
0.1644287109375
],
[
0.509765625,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.509765625,
0.179931640625
],
[
0.1175537109375,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.1175537109375,
0.18701171875
],
[
0.51123046875,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51123046875,
0.1953125
],
[
0.1158447265625,
0.189453125,
0.490234375,
0.188232421875,
0.490234375,
0.203125,
0.1158447265625,
0.2037353515625
],
[
0.51123046875,
0.198974609375,
0.84033203125,
0.198974609375,
0.84033203125,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.1175537109375,
0.2060546875,
0.490234375,
0.2060546875,
0.490234375,
0.2197265625,
0.1175537109375,
0.2197265625
],
[
0.1175537109375,
0.22265625,
0.48779296875,
0.22265625,
0.48779296875,
0.236328125,
0.1175537109375,
0.236328125
],
[
0.52880859375,
0.21728515625,
0.8818359375,
0.21728515625,
0.8818359375,
0.23095703125,
0.52880859375,
0.23095703125
],
[
0.1158447265625,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.1158447265625,
0.251708984375
],
[
0.509765625,
0.2340087890625,
0.880859375,
0.2340087890625,
0.880859375,
0.2476806640625,
0.509765625,
0.2476806640625
],
[
0.1175537109375,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.268310546875,
0.1175537109375,
0.268310546875
],
[
0.51318359375,
0.25048828125,
0.8818359375,
0.25048828125,
0.8818359375,
0.26416015625,
0.51318359375,
0.26416015625
],
[
0.1158447265625,
0.270751953125,
0.489501953125,
0.270751953125,
0.489501953125,
0.283935546875,
0.1158447265625,
0.283935546875
],
[
0.51318359375,
0.267333984375,
0.8818359375,
0.267333984375,
0.8818359375,
0.28076171875,
0.51318359375,
0.28076171875
],
[
0.1158447265625,
0.2861328125,
0.39794921875,
0.2861328125,
0.39794921875,
0.300537109375,
0.1158447265625,
0.300537109375
],
[
0.51318359375,
0.283203125,
0.8818359375,
0.283203125,
0.8818359375,
0.29638671875,
0.51318359375,
0.29638671875
],
[
0.5146484375,
0.300537109375,
0.880859375,
0.300537109375,
0.880859375,
0.311279296875,
0.5146484375,
0.311279296875
],
[
0.1158447265625,
0.314208984375,
0.270263671875,
0.314208984375,
0.270263671875,
0.328857421875,
0.1158447265625,
0.328857421875
],
[
0.51123046875,
0.3134765625,
0.88330078125,
0.314208984375,
0.88330078125,
0.328857421875,
0.51123046875,
0.327880859375
],
[
0.51123046875,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.1175537109375,
0.34033203125,
0.489501953125,
0.34033203125,
0.489501953125,
0.35400390625,
0.1175537109375,
0.35400390625
],
[
0.509765625,
0.345703125,
0.88330078125,
0.3466796875,
0.88330078125,
0.361083984375,
0.509765625,
0.35986328125
],
[
0.1175537109375,
0.35693359375,
0.48779296875,
0.35693359375,
0.48779296875,
0.37060546875,
0.1175537109375,
0.37060546875
],
[
0.509765625,
0.363525390625,
0.8818359375,
0.363525390625,
0.8818359375,
0.377197265625,
0.509765625,
0.377197265625
],
[
0.11419677734375,
0.371826171875,
0.489501953125,
0.372314453125,
0.489501953125,
0.38720703125,
0.11419677734375,
0.385986328125
],
[
0.509765625,
0.37890625,
0.75390625,
0.37890625,
0.75390625,
0.393798828125,
0.509765625,
0.393798828125
],
[
0.11419677734375,
0.388427734375,
0.486083984375,
0.388427734375,
0.486083984375,
0.402099609375,
0.11419677734375,
0.402099609375
],
[
0.52880859375,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.412109375,
0.52880859375,
0.412109375
],
[
0.1175537109375,
0.405029296875,
0.489501953125,
0.405029296875,
0.489501953125,
0.418701171875,
0.1175537109375,
0.418701171875
],
[
0.1175537109375,
0.42041015625,
0.48779296875,
0.42041015625,
0.48779296875,
0.43408203125,
0.1175537109375,
0.43408203125
],
[
0.509765625,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.51123046875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.51123046875,
0.44482421875
],
[
0.1175537109375,
0.43701171875,
0.489501953125,
0.43701171875,
0.489501953125,
0.45068359375,
0.1175537109375,
0.45068359375
],
[
0.1175537109375,
0.453125,
0.489501953125,
0.453125,
0.489501953125,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.51123046875,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.5087890625,
0.463134765625,
0.8818359375,
0.4619140625,
0.8818359375,
0.476806640625,
0.5087890625,
0.47802734375
],
[
0.1175537109375,
0.4697265625,
0.48779296875,
0.4697265625,
0.48779296875,
0.482666015625,
0.1175537109375,
0.482666015625
],
[
0.509765625,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.493408203125,
0.509765625,
0.493408203125
],
[
0.1175537109375,
0.486328125,
0.48779296875,
0.486328125,
0.48779296875,
0.49951171875,
0.1175537109375,
0.49951171875
],
[
0.51318359375,
0.495361328125,
0.8818359375,
0.495361328125,
0.8818359375,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.1158447265625,
0.501953125,
0.470947265625,
0.501953125,
0.470947265625,
0.515625,
0.1158447265625,
0.515625
],
[
0.51318359375,
0.51123046875,
0.8818359375,
0.51123046875,
0.8818359375,
0.525390625,
0.51318359375,
0.525390625
],
[
0.1368408203125,
0.5185546875,
0.489501953125,
0.5185546875,
0.489501953125,
0.5322265625,
0.1368408203125,
0.5322265625
],
[
0.51318359375,
0.52783203125,
0.75244140625,
0.52783203125,
0.75244140625,
0.54150390625,
0.51318359375,
0.54150390625
],
[
0.1175537109375,
0.53369140625,
0.486083984375,
0.53369140625,
0.486083984375,
0.54736328125,
0.1175537109375,
0.54736328125
],
[
0.1175537109375,
0.5498046875,
0.48779296875,
0.5498046875,
0.48779296875,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.52880859375,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.52880859375,
0.56005859375
],
[
0.1158447265625,
0.5654296875,
0.48779296875,
0.5654296875,
0.48779296875,
0.5791015625,
0.1158447265625,
0.5791015625
],
[
0.51123046875,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57763671875,
0.51123046875,
0.57763671875
],
[
0.1175537109375,
0.58203125,
0.490234375,
0.58203125,
0.490234375,
0.5966796875,
0.1175537109375,
0.5966796875
],
[
0.509765625,
0.57763671875,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.509765625,
0.59130859375
],
[
0.1175537109375,
0.5986328125,
0.48779296875,
0.5986328125,
0.48779296875,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.51318359375,
0.595703125,
0.88330078125,
0.595703125,
0.88330078125,
0.609375,
0.51318359375,
0.609375
],
[
0.1175537109375,
0.61328125,
0.48779296875,
0.6142578125,
0.48779296875,
0.62890625,
0.1175537109375,
0.62744140625
],
[
0.51123046875,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51123046875,
0.62353515625
],
[
0.1158447265625,
0.6298828125,
0.486083984375,
0.6298828125,
0.486083984375,
0.6435546875,
0.1158447265625,
0.6435546875
],
[
0.51123046875,
0.626953125,
0.87890625,
0.626953125,
0.87890625,
0.64013671875,
0.51123046875,
0.64013671875
],
[
0.1175537109375,
0.64794921875,
0.489501953125,
0.64794921875,
0.489501953125,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.509765625,
0.6435546875,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.509765625,
0.65673828125
],
[
0.1158447265625,
0.662109375,
0.490234375,
0.662109375,
0.490234375,
0.67578125,
0.1158447265625,
0.67578125
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.1175537109375,
0.6787109375,
0.489501953125,
0.6787109375,
0.489501953125,
0.6923828125,
0.1175537109375,
0.6923828125
],
[
0.51123046875,
0.6748046875,
0.8818359375,
0.67578125,
0.8818359375,
0.6904296875,
0.51123046875,
0.689453125
],
[
0.1175537109375,
0.69677734375,
0.486083984375,
0.69677734375,
0.486083984375,
0.70703125,
0.1175537109375,
0.70703125
],
[
0.51123046875,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.7060546875,
0.51123046875,
0.7060546875
],
[
0.1158447265625,
0.71044921875,
0.489501953125,
0.708984375,
0.489501953125,
0.7236328125,
0.1158447265625,
0.724609375
],
[
0.509765625,
0.7080078125,
0.8818359375,
0.7080078125,
0.8818359375,
0.7216796875,
0.509765625,
0.7216796875
],
[
0.133544921875,
0.7255859375,
0.48779296875,
0.7255859375,
0.48779296875,
0.7392578125,
0.133544921875,
0.7392578125
],
[
0.509765625,
0.724609375,
0.8818359375,
0.724609375,
0.8818359375,
0.73828125,
0.509765625,
0.73828125
],
[
0.1175537109375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.51123046875,
0.74072265625,
0.8818359375,
0.74072265625,
0.8818359375,
0.75439453125,
0.51123046875,
0.75439453125
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.755859375,
0.8818359375,
0.755859375,
0.8818359375,
0.76953125,
0.51123046875,
0.76953125
],
[
0.1158447265625,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1158447265625,
0.7890625
],
[
0.51123046875,
0.771484375,
0.86572265625,
0.771484375,
0.86572265625,
0.78515625,
0.51123046875,
0.78515625
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.5283203125,
0.7890625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8056640625,
0.5283203125,
0.80419921875
],
[
0.1175537109375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.509765625,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.5087890625,
0.84033203125,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.5087890625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51318359375,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.8681640625,
0.51318359375,
0.8681640625
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.11419677734375,
0.88818359375,
0.489501953125,
0.88720703125,
0.489501953125,
0.90185546875,
0.11419677734375,
0.90283203125
],
[
0.5087890625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.5087890625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"guarding gender where INLP needs 10-15.",
"different strategies to solve this optimization prob",
"lem for their chosen classifiers. A similar strategy",
"2. MP has less impact on other aspects of the",
"was previously described by Haghighatkhah et al",
"embeddingspace.",
"(2021, 2022).Here the corresponding minimax op",
"INLP projections improve simlex scores and reduce",
"timization problem is referred to as “maximizing",
"inseparability\".Haghighatkhah et al. (2022) show",
"WEAT scores more than MP. A priori it is unclear",
"that this problem can be solved efficiently under",
"why their many projections should have this effect.",
"certain convexity assumptions on the metric.",
"We investigated and show that the same improve-",
"ments appear after applying random projections",
"Zhang et al. (2022) independently developed a",
"(either after MP or after the first INLP projections).",
"projection method which is very similar to our MP",
"Applying one MP projection to linearly guard an",
"Their method also uses class means to find the",
"attribute is hence methodologically cleaner than",
"projection direction, but proceeds via a projection",
"applying multiple (INLP) projections.",
"to the origin, and hence needs one additional pro-",
"jection.Thegoal ofthepaper,however,isquite",
"2Related Work",
"different, namely the study of human brain reac-",
"tions to syntactic and semantic features of words.",
"Multiple methods for removing bias from embed-",
"As such, it does not include a systematic compar-",
"dings have been suggested. Bias can be addressed",
"ison to INLP or a comprehensive analysis of the",
"at the level of the training data (Zhao et al., 2018a),",
" impact on the embedding space.",
"the training process (Zhang et al., 2018; Madras",
"Shao et al. (2022) in recent and unpublished",
"et al., 2018), and the resulting model itself (Ravfo-",
"gel et al., 2020; Bolukbasi et al., 2016; Vargas and",
"work compute linear projection vectors by mini",
"mizing the covariance between biased word vectors",
"Cotterell, 2020). Debiasing an existing model has",
"the clear advantage that it can be done with rela-",
"and the protected attribute. They generally need",
"two projections to achieve comparable or better",
"tively little data and that the model does not need to",
"performance to INLP. The paper does not report on",
"be retrained. In this paper, we focus on removing",
"exactly the same experiments as we do, but for all",
"bias through transforming the embedding space.",
"experiments reported in both papers, our method",
"Transformations of embedding spaces can be per-",
"MP performs similarly or better.",
"formed by means of linear projections (Bolukbasi",
"et al., 2016), multiple linear projections (Ravfo-",
"Dev et al. (2021a) use a different kind of linear",
"gel et al., 2020), or via non-linear kernels (Vargas",
"transformation on embedding spaces, namely ro-",
"and Cotterell, 2020). Ravfogel et al. (2020) intro-",
"tation. Their goal is to disentangle two particular",
"duce a method called Iterative Nullspace Projection",
"attributes, such as gender and occupation. To do so,",
"(INLP) where they attempt to remove the bias by",
"they rotate the embedding space in such as way that",
"projecting points iteratively. Our methods achieve",
"the subspaces corresponding to the two attributes",
"linear guarding after a single linear projection. As",
"become orthogonal. On the positive side, their ap-",
" such we directly improve upon the INLP method,",
"proach does not remove a dimension and hence ar-",
"which we describe in detail in Section 3. In the",
"guably retains more information in the embedding",
"remainder of this section,wefocus onmorerecent",
"However, it will not remove all bias with respect",
"post-training approaches and applications of INLP.",
"to gender, but only gender bias with respect to a",
"Similar to Ravfogel et al. (2020), Ravfogel et al.",
"single specified attribute, e.g., occupation. More",
"(2022) aim to find a linear subspace of the data",
"generally, it will remove the interaction between",
"such that removing that subspace using projection",
"the two specified attributes and not actually com-",
"removes bias in the data optimally. Doing so nec-",
"pletely remove an attribute from the embedding.",
"essarily requires a metric to measure how well bias",
"INLP has recently gained importance for probing",
"has been removed by a specific projection. Rav-",
"studies. Probing has been criticized because (1)",
"fogel et al. (2022) use the classification loss of",
"results are difficult to interpret, and (2) it is usually",
"various classifiers as metrics. That is, a higher",
"not possible to test whether the target attribute is",
"classification loss after projection corresponds to",
"alsorelevantfor downstream tasks.Elazar et al.",
"better bias removal. This directly translates into a",
"(2021) propose amnesic probing, which employs",
"minimax optimization problem, which the authors",
"INLP to remove the target attribute and then tests",
"refer to as a minimax game. The authors present",
"change in performance of the adapted embeddings"
] |
{
"arxiv_id": "2212.04273",
"language": "en",
"page_idx": 1
}
|
||
2212.04273-en-2
|
2212.04273-en-2.jpg
|
on downstream tasks. INLP has since been used for this purpose in several other studies (Celikkanat et al., 2020; Nikoulina et al., 2021; Babazhanova et al., 2021; Dankers et al., 2022; Gonen et al., 2022; Lovering and Pavlick, 2022, e.g.). When using a projection-based removal method such as INLP or our MP in a probing setup, it is particularly important to ensure that all other information is preserved. If this is not the case, the probing study may lead to misleading conclusions. We hence recommend future such studies to consider using MP instead of INLP.
## 3 Projection Methods
In this section we introduce our two new projection methods MP and TMP. Furthermore, we theoretically analyze their relative performance when (linearly) debiasing word embeddings, also in comparison to the existing method INLP. Here we focus on binary classification for ease of explanation. Both methods can however also be used for multi-class classification (see Section 4.1) and then require $n-1$ projections for $n$ classes. We start with a formal description of the problem we study.
Our input is a set of word embedding vectors in $\mathbb{R}^d.$ We interpret these vectors as a set of points $P$ = $\{ p_1, \ldots , p_n\}$ in $\mathbb{R}^d$ and often use the term “point" when referring to an embedding vector. Every point (word) has an associated set of discrete attributes $(A=\{a_1,\ldots,a_n\}).$ Let $a^*$ be the attribute we aim to remove. For simplicity, we assume that $a^*$ is a binary attribute with values $\{-1,+1\}.$ We denote the two resulting classes of points as $P^-=\{p_i\in P\mid a_i^*=-1\}$ and $P^{+ }$ = $\{ p_{i}$ $\in$ $P$ | $a_{i}^{* }$ = $+ 1\} .$ We want to find a transformation $P^{\prime}$ of our point set $P$ such that any linear classifier $C$ trained on $P^{\prime}$ to classify $a^*$ cannot significantly outperform a classifier that labels by the majority class. That is, $P^{\prime}$ is linearly guarded with respect to $a^*.$
In the following we consider only transformations which consist of one or more projections. Let $p$ be a point in $P.$ We define its projection $p_w$ along a unit vector $w$ as $p_w=p-(p\cdot w)w.$ The projection along the vector $w$ maps points to the hyperplane $H_w$ which is orthogonal to $w\left(H_w\bot w\right)$ and contains the origin.
To evaluate how well INLP, MP, and TMP do in terms of linear guarding, we consider the number of misclassifications with respect to $a^*$ by the best possible linear classifier after a single projection;
the higher the number of misclassifications, the better the method performs. If a single projection is sufficient to achieve linear guarding, then arguably the semantic encoding of other attributes in our word embeddings are preserved as well as possible. From a theoretical perspective, TMP is the most effective method; it increases the number of misclassifications the most. However, it is costly to compute exactly for data in 300+ dimensions. MP is not as effective in theory as TMP, but it is more effective than INLP. Furthermore, MP is very easy to compute and appears to be very effective in practice. Hence, we evaluate the efficacy of MP extensively in Section 4 and recommend to use MP for linear guarding in practice.
## 3.1 $INLP$
Iterative Nullspace Projection (Ravfogel et al., 2020, INLP) debiases by iteratively projecting along a vector computed by a linear classifier. Here we assume that this linear classifier is a linear Support Vector Machine (SVM), as this is the main classifer used in their paper. In that setting, at each iteration, INLP trains an SVM to classify $a^*$ in the point set $P.$ Since $a^*$ is binary, the result of the SVM is a vector $w$ (along with a single real value, which we may ignore). The input points are then projected along $w$ and the resulting point set is used as the input for the next iteration.
If $P^-$ can be separated from $P^+$ with a hyperplane, then the vector $w$ is the normal of a separating hyperplane and indicates the direction along which there is the largest gap between $P^-$ and $P^{+}:$ the SVM margin. If $P^-$ and $P^+$ cannot be separated by a hyperplane, then the classifier allows misclassifications for a certain penalty, but otherwise still maximizes the margin. In this setting, SVM uses a parameter that controls the trade-off between misclassifications and the margin.
Analysis. The vector $w$ used in the projection essentially maximizes the gap between $P^-$ and $P^+$, and hence it is determined only by the points on the “outside" of $P^-$ and $P^+.$ Specifically, if $P^-$ and $P^+$ can be separated, then the vector $w$ can be determined solely by the points on the convex hulls$^{1}$ of $P^-$ and $P^+$, and is not influenced by the points that are interior to the convex hulls. If $P^-$ and $P^+$ are not separable, then $w$ will be determined by the points on the convex hulls of the correctly classified
The convex hull of a set of points is the smallest convex
set that contains all of the points.
|
[
[
0.1175537109375,
0.08489990234375,
0.48779296875,
0.08489990234375,
0.48779296875,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.5087890625,
0.1009521484375,
0.880859375,
0.1015625,
0.880859375,
0.11639404296875,
0.5087890625,
0.11517333984375
],
[
0.1158447265625,
0.1175537109375,
0.489501953125,
0.11639404296875,
0.489501953125,
0.1312255859375,
0.1158447265625,
0.1318359375
],
[
0.509765625,
0.1181640625,
0.87890625,
0.1181640625,
0.87890625,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.490234375,
0.1341552734375,
0.490234375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51123046875,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.51123046875,
0.1484375
],
[
0.1175537109375,
0.149658203125,
0.48779296875,
0.149658203125,
0.48779296875,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51123046875,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.162109375,
0.51123046875,
0.162109375
],
[
0.1158447265625,
0.1650390625,
0.486083984375,
0.1650390625,
0.486083984375,
0.1787109375,
0.1158447265625,
0.1787109375
],
[
0.52880859375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.18115234375,
0.52880859375,
0.18115234375
],
[
0.1158447265625,
0.18115234375,
0.48779296875,
0.1817626953125,
0.48779296875,
0.196533203125,
0.1158447265625,
0.1953125
],
[
0.51123046875,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.1158447265625,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.5146484375,
0.20068359375,
0.880859375,
0.20068359375,
0.880859375,
0.21142578125,
0.5146484375,
0.21142578125
],
[
0.1158447265625,
0.21435546875,
0.48779296875,
0.21435546875,
0.48779296875,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.50732421875,
0.21435546875,
0.8818359375,
0.213134765625,
0.8818359375,
0.22802734375,
0.50732421875,
0.2286376953125
],
[
0.11419677734375,
0.23095703125,
0.489501953125,
0.2298583984375,
0.489501953125,
0.24462890625,
0.11419677734375,
0.2452392578125
],
[
0.5087890625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.5087890625,
0.24462890625
],
[
0.1158447265625,
0.2464599609375,
0.489501953125,
0.2464599609375,
0.489501953125,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.5087890625,
0.2452392578125,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.5087890625,
0.260009765625
],
[
0.1158447265625,
0.26318359375,
0.27197265625,
0.26318359375,
0.27197265625,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.51123046875,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.51123046875,
0.279052734375,
0.88330078125,
0.279052734375,
0.88330078125,
0.292236328125,
0.51123046875,
0.292236328125
],
[
0.1158447265625,
0.29345703125,
0.31982421875,
0.29345703125,
0.31982421875,
0.30810546875,
0.1158447265625,
0.30810546875
],
[
0.509765625,
0.29443359375,
0.8818359375,
0.29443359375,
0.8818359375,
0.30810546875,
0.509765625,
0.30810546875
],
[
0.509765625,
0.31005859375,
0.7119140625,
0.311279296875,
0.7119140625,
0.325927734375,
0.509765625,
0.32470703125
],
[
0.11419677734375,
0.320556640625,
0.48779296875,
0.32177734375,
0.48779296875,
0.336181640625,
0.11419677734375,
0.3349609375
],
[
0.1158447265625,
0.33837890625,
0.495361328125,
0.33837890625,
0.495361328125,
0.3515625,
0.1158447265625,
0.3515625
],
[
0.509765625,
0.33837890625,
0.60107421875,
0.33837890625,
0.60107421875,
0.352783203125,
0.509765625,
0.352783203125
],
[
0.1158447265625,
0.355224609375,
0.490234375,
0.355224609375,
0.490234375,
0.365234375,
0.1158447265625,
0.365234375
],
[
0.509765625,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.509765625,
0.37353515625
],
[
0.1175537109375,
0.37060546875,
0.490234375,
0.37060546875,
0.490234375,
0.385498046875,
0.1175537109375,
0.385498046875
],
[
0.509765625,
0.377197265625,
0.8818359375,
0.377197265625,
0.8818359375,
0.39013671875,
0.509765625,
0.39013671875
],
[
0.1175537109375,
0.385986328125,
0.48779296875,
0.385986328125,
0.48779296875,
0.399658203125,
0.1175537109375,
0.399658203125
],
[
0.509765625,
0.392578125,
0.8818359375,
0.391357421875,
0.8818359375,
0.40625,
0.509765625,
0.40673828125
],
[
0.1175537109375,
0.402587890625,
0.486083984375,
0.402587890625,
0.486083984375,
0.416259765625,
0.1175537109375,
0.416259765625
],
[
0.51123046875,
0.4091796875,
0.8818359375,
0.4091796875,
0.8818359375,
0.4228515625,
0.51123046875,
0.4228515625
],
[
0.1175537109375,
0.418701171875,
0.48779296875,
0.418701171875,
0.48779296875,
0.431640625,
0.1175537109375,
0.431640625
],
[
0.509765625,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.509765625,
0.438232421875
],
[
0.118408203125,
0.435791015625,
0.484375,
0.435791015625,
0.484375,
0.446533203125,
0.118408203125,
0.446533203125
],
[
0.51123046875,
0.441162109375,
0.880859375,
0.441162109375,
0.880859375,
0.454833984375,
0.51123046875,
0.454833984375
],
[
0.1175537109375,
0.45068359375,
0.48779296875,
0.45068359375,
0.48779296875,
0.46435546875,
0.1175537109375,
0.46435546875
],
[
0.509765625,
0.457275390625,
0.8818359375,
0.457275390625,
0.8818359375,
0.47021484375,
0.509765625,
0.47021484375
],
[
0.1158447265625,
0.465576171875,
0.43994140625,
0.466064453125,
0.43994140625,
0.48095703125,
0.1158447265625,
0.479736328125
],
[
0.51123046875,
0.47265625,
0.8818359375,
0.47265625,
0.8818359375,
0.486328125,
0.51123046875,
0.486328125
],
[
0.1351318359375,
0.48388671875,
0.486083984375,
0.48388671875,
0.486083984375,
0.49755859375,
0.1351318359375,
0.49755859375
],
[
0.51123046875,
0.48681640625,
0.88330078125,
0.488037109375,
0.88330078125,
0.5029296875,
0.51123046875,
0.501953125
],
[
0.1158447265625,
0.49755859375,
0.48779296875,
0.49951171875,
0.48779296875,
0.51416015625,
0.1158447265625,
0.5126953125
],
[
0.509765625,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.51953125,
0.509765625,
0.51953125
],
[
0.1175537109375,
0.51416015625,
0.489501953125,
0.51416015625,
0.489501953125,
0.5322265625,
0.1175537109375,
0.5322265625
],
[
0.51123046875,
0.52099609375,
0.8818359375,
0.52099609375,
0.8818359375,
0.53515625,
0.51123046875,
0.53515625
],
[
0.1158447265625,
0.5322265625,
0.489501953125,
0.5322265625,
0.489501953125,
0.5458984375,
0.1158447265625,
0.5458984375
],
[
0.509765625,
0.53759765625,
0.75830078125,
0.53759765625,
0.75830078125,
0.55029296875,
0.509765625,
0.55029296875
],
[
0.1175537109375,
0.548828125,
0.489501953125,
0.548828125,
0.489501953125,
0.5625,
0.1175537109375,
0.5625
],
[
0.53076171875,
0.552734375,
0.884765625,
0.552734375,
0.884765625,
0.56787109375,
0.53076171875,
0.56787109375
],
[
0.11419677734375,
0.5634765625,
0.489501953125,
0.5625,
0.489501953125,
0.580078125,
0.11419677734375,
0.58056640625
],
[
0.51318359375,
0.5693359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.5830078125,
0.51318359375,
0.5830078125
],
[
0.11419677734375,
0.580078125,
0.484375,
0.580078125,
0.484375,
0.59326171875,
0.11419677734375,
0.59326171875
],
[
0.5087890625,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.5087890625,
0.59716796875
],
[
0.1158447265625,
0.5966796875,
0.48779296875,
0.595703125,
0.48779296875,
0.60986328125,
0.1158447265625,
0.61083984375
],
[
0.509765625,
0.6015625,
0.884765625,
0.6015625,
0.884765625,
0.615234375,
0.509765625,
0.615234375
],
[
0.12005615234375,
0.61328125,
0.486083984375,
0.61328125,
0.486083984375,
0.626953125,
0.12005615234375,
0.626953125
],
[
0.51318359375,
0.61767578125,
0.8818359375,
0.61767578125,
0.8818359375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.1175537109375,
0.62890625,
0.486083984375,
0.62890625,
0.486083984375,
0.6435546875,
0.1175537109375,
0.6435546875
],
[
0.51318359375,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.6484375,
0.51318359375,
0.6484375
],
[
0.1158447265625,
0.64111328125,
0.48779296875,
0.6435546875,
0.48779296875,
0.6611328125,
0.1158447265625,
0.6591796875
],
[
0.5146484375,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.662109375,
0.5146484375,
0.662109375
],
[
0.11419677734375,
0.6591796875,
0.484375,
0.6591796875,
0.484375,
0.67333984375,
0.11419677734375,
0.67333984375
],
[
0.51123046875,
0.6650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.68115234375,
0.51123046875,
0.6796875
],
[
0.11419677734375,
0.67578125,
0.48193359375,
0.67578125,
0.48193359375,
0.689453125,
0.11419677734375,
0.689453125
],
[
0.51123046875,
0.68115234375,
0.8818359375,
0.681640625,
0.8818359375,
0.69677734375,
0.51123046875,
0.6953125
],
[
0.1175537109375,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.1175537109375,
0.70703125
],
[
0.509765625,
0.69873046875,
0.8271484375,
0.69873046875,
0.8271484375,
0.7119140625,
0.509765625,
0.7119140625
],
[
0.1175537109375,
0.708984375,
0.484375,
0.708984375,
0.484375,
0.7236328125,
0.1175537109375,
0.7236328125
],
[
0.1175537109375,
0.7255859375,
0.32421875,
0.7255859375,
0.32421875,
0.7392578125,
0.1175537109375,
0.7392578125
],
[
0.51123046875,
0.72265625,
0.88330078125,
0.7236328125,
0.88330078125,
0.73828125,
0.51123046875,
0.73681640625
],
[
0.133544921875,
0.7421875,
0.490234375,
0.7421875,
0.490234375,
0.755859375,
0.133544921875,
0.755859375
],
[
0.509765625,
0.7392578125,
0.88330078125,
0.73828125,
0.88330078125,
0.7529296875,
0.509765625,
0.75439453125
],
[
0.1158447265625,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.7724609375,
0.1158447265625,
0.7724609375
],
[
0.509765625,
0.755859375,
0.8818359375,
0.755859375,
0.8818359375,
0.76953125,
0.509765625,
0.76953125
],
[
0.1175537109375,
0.77490234375,
0.48779296875,
0.77490234375,
0.48779296875,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.77099609375,
0.8818359375,
0.771484375,
0.8818359375,
0.7861328125,
0.509765625,
0.78515625
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1175537109375,
0.8056640625
],
[
0.51318359375,
0.7880859375,
0.8818359375,
0.7880859375,
0.8818359375,
0.802734375,
0.51318359375,
0.802734375
],
[
0.1158447265625,
0.80712890625,
0.489501953125,
0.8056640625,
0.489501953125,
0.82080078125,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.8046875,
0.880859375,
0.8046875,
0.880859375,
0.818359375,
0.509765625,
0.818359375
],
[
0.1175537109375,
0.822265625,
0.489501953125,
0.822265625,
0.489501953125,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.5087890625,
0.8193359375,
0.880859375,
0.8193359375,
0.880859375,
0.83447265625,
0.5087890625,
0.83447265625
],
[
0.1158447265625,
0.8388671875,
0.291259765625,
0.8388671875,
0.291259765625,
0.8525390625,
0.1158447265625,
0.8525390625
],
[
0.51123046875,
0.8359375,
0.8818359375,
0.83447265625,
0.8818359375,
0.8486328125,
0.51123046875,
0.85107421875
],
[
0.1351318359375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1351318359375,
0.869140625
],
[
0.51318359375,
0.8525390625,
0.880859375,
0.8525390625,
0.880859375,
0.8662109375,
0.51318359375,
0.8662109375
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.869140625,
0.8818359375,
0.86767578125,
0.8818359375,
0.88134765625,
0.51123046875,
0.8828125
],
[
0.11419677734375,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.11419677734375,
0.90185546875
],
[
0.53076171875,
0.892578125,
0.88330078125,
0.892578125,
0.88330078125,
0.90673828125,
0.53076171875,
0.90673828125
],
[
0.1158447265625,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.50732421875,
0.90478515625,
0.7138671875,
0.90478515625,
0.7138671875,
0.91845703125,
0.50732421875,
0.91845703125
]
] |
[
"on downstream tasks. INLP has since been used",
"the higher the number of misclassifications, the",
"for this purpose in several other studies (Celikkanat",
"better the method performs. If a single projection is",
"et al., 2020; Nikoulina et al., 2021; Babazhanova",
" sufficient to achieve linear guarding, then arguably",
"et al., 2021; Dankers et al., 2022; Gonen et al.,",
"the semantic encoding of other attributes in our",
"2022; Lovering and Pavlick, 2022, e.g.). When",
"word embeddings arepreserved as wellaspossible",
"using a projection-based removal method such as",
"From a theoretical perspective, TMP is the most",
"INLP or our MP in a probing setup, it is particularly",
"effective method; it increases the number of mis-",
"important to ensure that all other information is",
"classificationsthemost.However,itiscostlyto",
"preserved. If this is not the case, the probing study",
"compute exactly for data in300+dimensions.MP",
"may lead to misleading conclusions. We hence",
"is not as effective in theory as TMP, but it is more",
"recommend future such studies to consider using",
"effective than INLP. Furthermore, MP is very easy",
"MP instead of INLP.",
"to compute and appears to be very effective in prac-",
"tice. Hence, we evaluate the efficacy of MP exten-",
"3Projection Methods",
"sively in Section 4 and recommend to use MP for",
"linear guarding in practice.",
"In this section we introduce our two new projection",
"methods MP and TMP.Furthermore, we theoreti-",
"3.1 INLP",
"cally analvzetheir relative performancewhen (lin-",
"Iterative Nullspace Projection (Ravfogel et al.",
"early) debiasing word embeddings, also in compar-",
"2020,INLP)debiases by iteratively projecting",
"ison to the existing method INLP. Here we focus on",
"along a vector computed by a linear classifier. Here",
"binary classification for ease of explanation. Both",
"we assume that this linear classifier is a linear Sup-",
"methods can however also be used for multi-class",
"port Vector Machine (SVM), as this is the main",
"classification (see Section 4.1) and then reguire",
"classifier used in their paper. In that setting, at each",
"n.- 1 projections for n classes. We start with a",
"iteration, INLP trains an SVM to classify a* in the",
"formal description of the problem we study",
"point set P. Since a* is binary, the result of the",
"Our input is a set of word embedding vectors in",
"SvM is a vector w (along with a single real value,",
"IRd. We interpret these vectors as a set of points",
"which we may ignore). The input points are then",
"P = (p1,., pn) in Rd and often use the term",
"projected along w and the resulting point set is used",
"\"point\" when referring to an embedding vector.",
"as the input for the next iteration.",
"Every point (word) has an associated set of dis-",
"If P- can be separated from P+ with a hyper-",
"crete attributes (A = [a1,... , an]). Let a* be the",
"plane, then the vector w is the normal of a sepa-",
"attribute we aim to remove. For simplicity, we",
"rating hyperplane and indicates the direction along",
"assume that a* is a binary attribute with values",
"which there is the largest gap between P- and P+:",
"[-1, +1]. We denote the two resulting classes",
"the SVM margin. If P- and P+ cannot be sep-",
"of points as P- = (pi E P [ α* = -1] and",
"arated by a hyperplane, then the classifier allows",
"P+ = (pi E P I a, = +1). We want to find",
"misclassificationsforacertainpenalty,butother-",
"a transformation P' of our point set P such that",
"wise still maximizes the margin. In this setting,",
"any linear classifier C trained on P' to classify a",
"SVM uses a parameter that controls the trade-off",
"cannot significantly outperform a classifier that la-",
"between misclassifications and the margin.",
"bels by the majority class. That is, P' is linearly",
"guarded with respect to a*.",
"Analysis.The vector w used in the projection es-",
"In the following we consider only transforma-",
" sentially maximizes the gap between P- and P+,",
"tions which consist of one or more projections. Let",
"and hence it is determined only by the points on the",
"p be a point in P. We define its projection pu",
"“outside\"of P-and P+.Specifically, if P-and",
"along a unit vector w as pw = p - (p - w)w. The",
"P+ can be separated, then the vector w can be de",
"projection along the vector w maps points to the",
"termined solely by the points on the convex hulls!",
"hyperplane Hw which is orthogonal to w (Hw Iw)",
" of P- and P+, and is not influenced by the points",
"and contains the origin.",
"that are interior to the convex hulls. If P- and P+",
"To evaluate how well INLP, MP, and TMP do in",
"are not separable, then w will be determined by the",
"terms of linear guarding, we consider the number",
"points on the convex hulls of the correctly classified",
"of misclassifications with respect to a* by the best",
"IThe convex hull of a set of points is the smallest convex",
"possible linear classifier after a single projection;",
"set that contains all of the points."
] |
{
"arxiv_id": "2212.04273",
"language": "en",
"page_idx": 2
}
|
||
2212.04273-en-4
|
2212.04273-en-4.jpg
|
<FigureHere>
Figure 3: Two points and their Tukey depth.
Edelsbrunner (1987)). Finally, we can show that there exist point sets $P^+$ and $P^-$ such that, after any projection along a single vector, the number of misclassifications is at most $O(n/d).$ Thus, TMP is asymptotically worst-case optimal for maximizing the number of misclassifications after projection. The exact theorem statements and their proofs can be found in Appendix A.
Unfortunately, TMP has one major drawback: computing the Tukey median of a set of points is computationally expensive in high dimensions. Although it is possible to compute an approximate Tukey median efficiently, the Tukey depth of this approximation may be significantly worse than that of the Tukey median (namely, $O(n/d^{2}))^{2}.$ Hence, we generally recommend to use MP in practice.
Let $P$ be a point set in $\mathbb{R}^d$ and let $q\in\mathbb{R}^d$ be another point which is not necessarily in $P.$ The Tukey depth $t(q)$ of $q$ with respect to $P$ is the smallest number of points in $P$ in any closed halfspace bounded by a hyperplane in $\mathbb{R}^d$ that passes through $q.$ In Figure 3, $q_1$ is a point with Tukey depth 3 and $q_{2}$ is a point with Tukey depth 5. A Tukey median $\tau$ of a point set $P$ is a point $q\in\mathbb{R}^d$ with the maximum Tukey depth with respect to $P$ among all points in $\mathbb{R}^d.$ Point $q_{2}$ in Figure 3 is a Tukey median of the black points.
Now let $\tau^-$ be a Tukey median of $P^-$ and let $\tau^+$ be a Tukey median of $P^+.$ Then the projection vector of the Tukey Median Projection method is simply $w=\tau^+-\tau^-.$
## 4 Experiments
In this section, we present our main experimental results.$^3$ We first describe the outcome of the original INLP experiments and how our results compare (Section 4.1). We then outline the additional experiments we carried out to gain more insight into surprising results of the original INLP experiments as well as those of our own method (Section 4.2).
### 4.1 Original Experiments
We compare the performance of INLP and MP on the gender bias experiments reported in Sections 6.1 and 6.3 in Ravfogel et al. (2020). We carried out all experiments with the original (INLP) and our new (MP) method. We could successfully reproduce Ravfogel et al.'s 2020 results. Since the results were nearly identical, we report the INLP results from the original paper. For reasons of space, this section only contains the intuition and main results. We provide a self-contained description
Analysis. Figure 4 contrasts MP and TMP. The mean of the blue point set is clearly not in the center of the majority of the points due to the outliers. As a result, the projection of MP is not very effective for linear guarding. The Tukey median of the blue point set, however, remains centered near the majority of the points and hence TMP is effective.
One may wonder if there is also an example where TMP performs poorly. We can in fact prove that TMP in general performs quite well for linear guarding. Specifically, we can show that the number of misclassifications after TMP is at least the minimum of the Tukey depths of $\tau^+$ and $\tau^-.$ It is known that, for every point set $P$ in $\mathbb{R}^d$ with $n$ points, the Tukey median with respect to $P$ has Tukey depth at least $\left\lceil\frac{n}{d+1}\right\rceil($see,e.g., Herbert
$^{2}$We are aiming to maximize the number of misclassifica-
tions after projection, so $O(n/d)$ is better than $O(n/d^2)$
$^{3}$The code is available at tue-alga/debias-mean-projection.
<FigureHere>
Figure 4: Results of MP (left) and TMP (right).
|
[
[
0.51123046875,
0.08489990234375,
0.880859375,
0.08489990234375,
0.880859375,
0.09857177734375,
0.51123046875,
0.09857177734375
],
[
0.51123046875,
0.1015625,
0.880859375,
0.1015625,
0.880859375,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.51123046875,
0.1341552734375,
0.880859375,
0.1341552734375,
0.880859375,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.12176513671875,
0.15087890625,
0.189697265625,
0.15087890625,
0.189697265625,
0.162109375,
0.12176513671875,
0.162109375
],
[
0.32568359375,
0.15087890625,
0.3955078125,
0.15087890625,
0.3955078125,
0.162109375,
0.32568359375,
0.162109375
],
[
0.509765625,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.163330078125,
0.509765625,
0.163330078125
],
[
0.51123046875,
0.1650390625,
0.880859375,
0.1650390625,
0.880859375,
0.1787109375,
0.51123046875,
0.1787109375
],
[
0.15283203125,
0.1817626953125,
0.447509765625,
0.1817626953125,
0.447509765625,
0.1953125,
0.15283203125,
0.1953125
],
[
0.51123046875,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51123046875,
0.1953125
],
[
0.51123046875,
0.198974609375,
0.6962890625,
0.198974609375,
0.6962890625,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.52880859375,
0.21435546875,
0.88330078125,
0.21435546875,
0.88330078125,
0.2286376953125,
0.52880859375,
0.2286376953125
],
[
0.1318359375,
0.221435546875,
0.486083984375,
0.221435546875,
0.486083984375,
0.236328125,
0.1318359375,
0.236328125
],
[
0.509765625,
0.23095703125,
0.8818359375,
0.2298583984375,
0.8818359375,
0.24462890625,
0.509765625,
0.2452392578125
],
[
0.1158447265625,
0.2392578125,
0.486083984375,
0.2392578125,
0.486083984375,
0.2529296875,
0.1158447265625,
0.2529296875
],
[
0.5087890625,
0.2452392578125,
0.88330078125,
0.2476806640625,
0.88330078125,
0.262451171875,
0.5087890625,
0.260009765625
],
[
0.1158447265625,
0.254150390625,
0.48779296875,
0.254638671875,
0.48779296875,
0.26953125,
0.1158447265625,
0.268310546875
],
[
0.509765625,
0.262451171875,
0.8818359375,
0.26318359375,
0.8818359375,
0.27783203125,
0.509765625,
0.276611328125
],
[
0.1175537109375,
0.271484375,
0.490234375,
0.271484375,
0.490234375,
0.28515625,
0.1175537109375,
0.28515625
],
[
0.509765625,
0.27783203125,
0.8818359375,
0.279052734375,
0.8818359375,
0.29345703125,
0.509765625,
0.292236328125
],
[
0.1175537109375,
0.287353515625,
0.48779296875,
0.287353515625,
0.48779296875,
0.3017578125,
0.1175537109375,
0.3017578125
],
[
0.51318359375,
0.295654296875,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.51318359375,
0.308837890625
],
[
0.1175537109375,
0.303955078125,
0.486083984375,
0.303955078125,
0.486083984375,
0.317626953125,
0.1175537109375,
0.317626953125
],
[
0.51123046875,
0.311279296875,
0.88330078125,
0.311279296875,
0.88330078125,
0.32470703125,
0.51123046875,
0.32470703125
],
[
0.1175537109375,
0.319580078125,
0.48779296875,
0.319580078125,
0.48779296875,
0.334228515625,
0.1175537109375,
0.334228515625
],
[
0.51123046875,
0.327880859375,
0.8623046875,
0.327880859375,
0.8623046875,
0.341552734375,
0.51123046875,
0.341552734375
],
[
0.1158447265625,
0.3349609375,
0.486083984375,
0.333251953125,
0.486083984375,
0.347412109375,
0.1158447265625,
0.349853515625
],
[
0.1175537109375,
0.3515625,
0.486083984375,
0.3515625,
0.486083984375,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.5205078125,
0.35693359375,
0.5283203125,
0.3623046875,
0.5224609375,
0.36767578125,
0.5146484375,
0.3623046875
],
[
0.5400390625,
0.355224609375,
0.654296875,
0.355224609375,
0.654296875,
0.369384765625,
0.5400390625,
0.369384765625
],
[
0.1175537109375,
0.36767578125,
0.48779296875,
0.36767578125,
0.48779296875,
0.3818359375,
0.1175537109375,
0.3818359375
],
[
0.1158447265625,
0.383056640625,
0.366943359375,
0.38427734375,
0.366943359375,
0.3984375,
0.1158447265625,
0.39794921875
],
[
0.509765625,
0.380126953125,
0.8818359375,
0.38134765625,
0.8818359375,
0.3955078125,
0.509765625,
0.394287109375
],
[
0.1351318359375,
0.399658203125,
0.489501953125,
0.399658203125,
0.489501953125,
0.413330078125,
0.1351318359375,
0.413330078125
],
[
0.5087890625,
0.3955078125,
0.8818359375,
0.39794921875,
0.8818359375,
0.410888671875,
0.5087890625,
0.4091796875
],
[
0.118408203125,
0.416259765625,
0.48779296875,
0.416259765625,
0.48779296875,
0.429931640625,
0.118408203125,
0.429931640625
],
[
0.509765625,
0.412109375,
0.8818359375,
0.413330078125,
0.8818359375,
0.427490234375,
0.509765625,
0.427001953125
],
[
0.1158447265625,
0.431640625,
0.48779296875,
0.431640625,
0.48779296875,
0.4453125,
0.1158447265625,
0.4453125
],
[
0.51123046875,
0.429931640625,
0.88330078125,
0.429931640625,
0.88330078125,
0.443603515625,
0.51123046875,
0.443603515625
],
[
0.1175537109375,
0.448974609375,
0.2359619140625,
0.448974609375,
0.2359619140625,
0.4619140625,
0.1175537109375,
0.4619140625
],
[
0.51123046875,
0.4453125,
0.8818359375,
0.4453125,
0.8818359375,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.51123046875,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.51123046875,
0.4755859375
],
[
0.1158447265625,
0.47265625,
0.486083984375,
0.47265625,
0.486083984375,
0.486328125,
0.1158447265625,
0.486328125
],
[
0.51123046875,
0.47802734375,
0.8759765625,
0.47802734375,
0.8759765625,
0.4912109375,
0.51123046875,
0.4912109375
],
[
0.1158447265625,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.5029296875,
0.1158447265625,
0.5029296875
],
[
0.1158447265625,
0.505859375,
0.49365234375,
0.50439453125,
0.49365234375,
0.51708984375,
0.1158447265625,
0.51953125
],
[
0.51318359375,
0.5048828125,
0.72705078125,
0.5048828125,
0.72705078125,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.1175537109375,
0.521484375,
0.491943359375,
0.521484375,
0.491943359375,
0.53515625,
0.1175537109375,
0.53515625
],
[
0.509765625,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.54052734375,
0.509765625,
0.54052734375
],
[
0.1158447265625,
0.53759765625,
0.489501953125,
0.5361328125,
0.489501953125,
0.55029296875,
0.1158447265625,
0.5517578125
],
[
0.509765625,
0.5419921875,
0.880859375,
0.5419921875,
0.880859375,
0.5556640625,
0.509765625,
0.5556640625
],
[
0.1158447265625,
0.55419921875,
0.48779296875,
0.55419921875,
0.48779296875,
0.56787109375,
0.1158447265625,
0.56787109375
],
[
0.51318359375,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.1158447265625,
0.5693359375,
0.48779296875,
0.5693359375,
0.48779296875,
0.5830078125,
0.1158447265625,
0.5830078125
],
[
0.51123046875,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.1368408203125,
0.58740234375,
0.486083984375,
0.58740234375,
0.486083984375,
0.59716796875,
0.1368408203125,
0.59716796875
],
[
0.51123046875,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.60400390625,
0.51123046875,
0.60400390625
],
[
0.1158447265625,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.1158447265625,
0.615234375
],
[
0.51318359375,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.619140625,
0.51318359375,
0.619140625
],
[
0.1158447265625,
0.6162109375,
0.484375,
0.6162109375,
0.484375,
0.6298828125,
0.1158447265625,
0.6298828125
],
[
0.509765625,
0.62255859375,
0.8818359375,
0.62255859375,
0.8818359375,
0.6357421875,
0.509765625,
0.6357421875
],
[
0.118408203125,
0.63427734375,
0.490234375,
0.63427734375,
0.490234375,
0.64794921875,
0.118408203125,
0.64794921875
],
[
0.51123046875,
0.6396484375,
0.88330078125,
0.6396484375,
0.88330078125,
0.65283203125,
0.51123046875,
0.65283203125
],
[
0.1175537109375,
0.6494140625,
0.489501953125,
0.6494140625,
0.489501953125,
0.6630859375,
0.1175537109375,
0.6630859375
],
[
0.51123046875,
0.65478515625,
0.8818359375,
0.65478515625,
0.8818359375,
0.66943359375,
0.51123046875,
0.66943359375
],
[
0.1158447265625,
0.6650390625,
0.484375,
0.6650390625,
0.484375,
0.6787109375,
0.1158447265625,
0.6787109375
],
[
0.509765625,
0.67041015625,
0.880859375,
0.67041015625,
0.880859375,
0.68408203125,
0.509765625,
0.68408203125
],
[
0.11419677734375,
0.681640625,
0.486083984375,
0.68115234375,
0.486083984375,
0.6953125,
0.11419677734375,
0.69677734375
],
[
0.1158447265625,
0.69873046875,
0.486083984375,
0.69873046875,
0.486083984375,
0.7119140625,
0.1158447265625,
0.7119140625
],
[
0.53076171875,
0.693359375,
0.880859375,
0.693359375,
0.880859375,
0.70703125,
0.53076171875,
0.70703125
],
[
0.51123046875,
0.70703125,
0.85546875,
0.70703125,
0.85546875,
0.72021484375,
0.51123046875,
0.72021484375
],
[
0.1158447265625,
0.71728515625,
0.306396484375,
0.71728515625,
0.306396484375,
0.73095703125,
0.1158447265625,
0.73095703125
],
[
0.343505859375,
0.71728515625,
0.48779296875,
0.71630859375,
0.48779296875,
0.73095703125,
0.343505859375,
0.732421875
],
[
0.5341796875,
0.71875,
0.880859375,
0.7197265625,
0.880859375,
0.73388671875,
0.5341796875,
0.73291015625
],
[
0.313232421875,
0.72802734375,
0.330078125,
0.72802734375,
0.330078125,
0.73095703125,
0.313232421875,
0.73095703125
],
[
0.175537109375,
0.7548828125,
0.193115234375,
0.7548828125,
0.193115234375,
0.77685546875,
0.175537109375,
0.77685546875
],
[
0.5751953125,
0.7548828125,
0.59375,
0.7548828125,
0.59375,
0.77685546875,
0.5751953125,
0.77685546875
],
[
0.162841796875,
0.86328125,
0.202392578125,
0.86328125,
0.202392578125,
0.892578125,
0.162841796875,
0.892578125
],
[
0.339111328125,
0.90087890625,
0.6572265625,
0.90087890625,
0.6572265625,
0.91455078125,
0.339111328125,
0.91455078125
]
] |
[
"Edelsbrunner (1987). Finally, we can show that",
"there exist point sets P+ and P- such that, after",
"any projection along a single vector, the number of",
"misclassifications is at most O(n/d). Thus, TMPis",
"t(q1) = 3",
"t(q2) = 5",
"asymptotically worst-case optimal for maximizing",
"the number of misclassifications after projection",
"Figure 3: Two points and their Tukey depth.",
"The exact theorem statements and their proofs can",
"be found in Appendix A.",
"Unfortunately, TMP has one major drawback:",
"Let P be a point set in Rd and let q e Rd be an-",
"computing the Tukey median of a set of points",
"other point which is not necessarily in P. The",
"is computationally expensive in high dimensions.",
"Tukey depth t(q) of q with respect to P is the",
"Although it is possible to compute an approximate",
"smallest number of points in P in any closed half-",
"Tukey median efficiently, the Tukey depth of this",
"space bounded by a hyperplane in Rd that passes",
"approximation may be significantly worse than that",
"through q. In Figure 3, q1 is a point with Tukey",
"of the Tukey median (namely, O(n /d?))?. Hence,",
"depth 3 and q2 is a point with Tukey depth 5. A",
"we generally recommend to use MP in practice.",
"Tukey median T of a point set P is a point q E Rd",
"with the maximum Tukey depth with respect to P",
"",
"Experiments",
"among all points in Rd. Point q2 in Figure 3 is a",
"Tukey median of the black points.",
"In this section, we present our main experimental",
"Now let T- be a Tukey median of P- and let",
"results.3 We first describe the outcome of the origi-",
"++ be a Tukey median of P+, Then the projection",
"nal INLP experiments and how our results compare",
"vector of the Tukey Median Projection method is",
"(Section 4.1). We then outline the additional ex-",
"simply w = -+",
"periments we carried out to gain more insight into",
"surprising results of the original INLP experiments",
"Analysis. Figure 4 contrasts MP and TMP. The",
"as well as those of our own method (Section 4.2).",
"mean of the blue point set is clearly not in the cen-",
"ter of the majority of the points due to the outliers.",
"4.1OriginalExperiments",
"As a result, the projection of MP is not very effec-",
"We compare the performance of INLP and MP on",
"tive for linear guarding. The Tukey median of the",
"the gender bias experiments reported in Sections",
"blue point set, however, remains centered near the",
"6.1 and 6.3 in Ravfogel et al. (2020). We carried",
"majority of the points and hence TMP is effective.",
"out all experiments with the original (INLP) and",
"One may wonder if thereis alsoan example",
"our new (MP) method. We could successfully re-",
"where TMP performs poorly. We can in fact prove",
"produce Ravfogel et al.'s 2020 results. Since the",
"that TMP in general performs quite well for linear",
"results were nearly identical, we report the INLP re",
"guarding. Specifically, we can show that the num-",
"sults from the original paper. For reasons of space,",
"ber of misclassifications after TMP is at least the",
"this section only contains the intuition and main",
"minimum of the Tukey depths of -+ and T-. I",
"results.We provide a self-contained description",
"is known that, for every point set P in Rd with",
"n points, the Tukey median with respect to P",
"2We are aiming to maximize the number of misclassifica",
"tions after projection, so O(n/d) is better than O(n/d?)",
"has Tukey depth at least",
"(see, e.g., Herbert",
"3The code is available at tue-alga/debias-mean-projection.",
"",
"000",
"",
"",
"Figure 4: Results of MP (left) and TMP (right)."
] |
{
"arxiv_id": "2212.04273",
"language": "en",
"page_idx": 4
}
|
||
1612.00712-en-0
|
1612.00712-en-0.jpg
|
# $\textbf{Probabilistic Neural Programs}$
$\mathbf{KentonW.Murray}^{*}$
Jayant Krishnamurthy
Allen Institute for Artificial Intelligence
Department of Computer Science and Engineering
University of Notre Dame
Seattle, WA 98166
South Bend, IN 46617
jayantk@allenai.org
kmurray4@nd.edu
### $\mathbf{Abstract}$
We present probabilistic neural programs, a framework for program induction that permits flexible specification of both a computational model and inference algorithm while simultaneously enabling the use of deep neural networks. Probabilistic neural programs combine a computation graph for specifying a neural network with an operator for weighted nondeterministic choice. Thus, a program describes both a collection of decisions as well as the neural network architecture used to make each one. We evaluate our approach on a challenging diagram question answering task where probabilistic neural programs correctly execute nearly twice as many programs as a baseline model.
## 1 Introduction
In recent years, deep learning has produced tremendous accuracy improvements on a variety of tasks in computer vision and natural language processing. A natural next step for deep learning is to consider program induction, the problem of learning computer programs from (noisy) input/output examples. Compared to more traditional problems, such as object recognition that require making only a single decision, program induction is difficult because it requires making a sequence of decisions and possibly learning control flow concepts such as loops and if statements.
Prior work on program induction has described two general classes of approaches. First, in the noise-free setting, program synthesis approaches pose program induction as completing a program sketch," which is a program containing nondeterministic choices (“holes”) to be filled by the learning algorithm [13]. Probabilistic programming languages generalize this approach to the noisy setting by permitting the sketch to specify a distribution over these choices as a function of prior parameters and further to condition this distribution on data, thereby training a Bayesian generative model to execute the sketch correctly [6]. Second, neural abstract machines define continuous analogues of Turing machines or other general-purpose computational models by “lifting”their discrete state and computation rules into a continuous representation [9,11,7,12]. Both of these approaches have demonstrated success at inducing simple programs from synthetic data but have yet to be applied to practical problems.
We observe that there are (at least) three dimensions along which we can characterize program
induction approaches:
1. Computational Model-what abstract model of computation does the model learn to control?
(e.g., a Turing machine)
2. Learning Mechanism-what kinds of machine learning models are supported? (e.g., neural
networks, Bayesian generative models)
$^{*}$Work done while on Internship at Allen Institute for Artificial Intelligence
lst Workshop on Neural Abstract Machines & Program Induction (NAMPI), @NIPS 2016, Barcelona, Spain.
|
[
[
0.314453125,
0.1256103515625,
0.6845703125,
0.126953125,
0.6845703125,
0.1458740234375,
0.314453125,
0.14453125
],
[
0.289306640625,
0.20263671875,
0.435546875,
0.20263671875,
0.435546875,
0.216552734375,
0.289306640625,
0.216552734375
],
[
0.59619140625,
0.20263671875,
0.75634765625,
0.20263671875,
0.75634765625,
0.216552734375,
0.59619140625,
0.216552734375
],
[
0.1895751953125,
0.216552734375,
0.52294921875,
0.216552734375,
0.52294921875,
0.2310791015625,
0.1895751953125,
0.2310791015625
],
[
0.5439453125,
0.21533203125,
0.8095703125,
0.216552734375,
0.8095703125,
0.2310791015625,
0.5439453125,
0.2298583984375
],
[
0.268798828125,
0.2291259765625,
0.44287109375,
0.2298583984375,
0.44287109375,
0.2442626953125,
0.268798828125,
0.24365234375
],
[
0.61376953125,
0.2298583984375,
0.740234375,
0.2298583984375,
0.740234375,
0.24365234375,
0.61376953125,
0.24365234375
],
[
0.279296875,
0.24365234375,
0.431396484375,
0.242431640625,
0.431396484375,
0.2568359375,
0.279296875,
0.25830078125
],
[
0.5849609375,
0.2442626953125,
0.771484375,
0.2442626953125,
0.771484375,
0.2587890625,
0.5849609375,
0.2587890625
],
[
0.281005859375,
0.2587890625,
0.4306640625,
0.2587890625,
0.4306640625,
0.272705078125,
0.281005859375,
0.272705078125
],
[
0.0228729248046875,
0.266357421875,
0.060455322265625,
0.266357421875,
0.060455322265625,
0.37744140625,
0.0228729248046875,
0.37744140625
],
[
0.461669921875,
0.30810546875,
0.53759765625,
0.30810546875,
0.53759765625,
0.322509765625,
0.461669921875,
0.322509765625
],
[
0.2344970703125,
0.34033203125,
0.765625,
0.34033203125,
0.765625,
0.354248046875,
0.2344970703125,
0.354248046875
],
[
0.233642578125,
0.354736328125,
0.76708984375,
0.354736328125,
0.76708984375,
0.36865234375,
0.233642578125,
0.36865234375
],
[
0.0269622802734375,
0.367431640625,
0.056365966796875,
0.367431640625,
0.056365966796875,
0.393310546875,
0.0269622802734375,
0.393310546875
],
[
0.233642578125,
0.36865234375,
0.765625,
0.36865234375,
0.765625,
0.3818359375,
0.233642578125,
0.3818359375
],
[
0.233642578125,
0.38330078125,
0.765625,
0.38330078125,
0.765625,
0.396484375,
0.233642578125,
0.396484375
],
[
0.0196075439453125,
0.395751953125,
0.05963134765625,
0.39453125,
0.062103271484375,
0.484130859375,
0.0228729248046875,
0.48486328125
],
[
0.233642578125,
0.396484375,
0.765625,
0.396484375,
0.765625,
0.410400390625,
0.233642578125,
0.410400390625
],
[
0.233642578125,
0.410400390625,
0.765625,
0.410400390625,
0.765625,
0.423583984375,
0.233642578125,
0.423583984375
],
[
0.2320556640625,
0.42236328125,
0.765625,
0.423583984375,
0.765625,
0.438232421875,
0.2320556640625,
0.4375
],
[
0.233642578125,
0.4375,
0.7646484375,
0.4375,
0.7646484375,
0.451904296875,
0.233642578125,
0.451904296875
],
[
0.2320556640625,
0.451904296875,
0.43310546875,
0.449462890625,
0.43310546875,
0.46337890625,
0.2320556640625,
0.46533203125
],
[
0.0228729248046875,
0.4873046875,
0.05963134765625,
0.4873046875,
0.05963134765625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.039215087890625,
0.494384765625,
0.0482177734375,
0.494384765625,
0.0482177734375,
0.498779296875,
0.039215087890625,
0.498779296875
],
[
0.178955078125,
0.49560546875,
0.1846923828125,
0.49560546875,
0.1846923828125,
0.50390625,
0.178955078125,
0.50390625
],
[
0.1993408203125,
0.49365234375,
0.31201171875,
0.49365234375,
0.31201171875,
0.5068359375,
0.1993408203125,
0.5068359375
],
[
0.1732177734375,
0.5244140625,
0.826171875,
0.5244140625,
0.826171875,
0.53857421875,
0.1732177734375,
0.53857421875
],
[
0.1748046875,
0.53857421875,
0.82421875,
0.53857421875,
0.82421875,
0.55322265625,
0.1748046875,
0.55322265625
],
[
0.1748046875,
0.55322265625,
0.82421875,
0.55322265625,
0.82421875,
0.56640625,
0.1748046875,
0.56640625
],
[
0.1748046875,
0.5673828125,
0.82421875,
0.5673828125,
0.82421875,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.1748046875,
0.580078125,
0.826171875,
0.580078125,
0.826171875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.1748046875,
0.5947265625,
0.734375,
0.5947265625,
0.734375,
0.60791015625,
0.1748046875,
0.60791015625
],
[
0.1748046875,
0.61279296875,
0.82421875,
0.61279296875,
0.82421875,
0.62744140625,
0.1748046875,
0.62744140625
],
[
0.1748046875,
0.62890625,
0.82421875,
0.62890625,
0.82421875,
0.642578125,
0.1748046875,
0.642578125
],
[
0.1707763671875,
0.6416015625,
0.82421875,
0.642578125,
0.82421875,
0.6572265625,
0.1707763671875,
0.65576171875
],
[
0.1748046875,
0.65576171875,
0.82421875,
0.65576171875,
0.82421875,
0.67041015625,
0.1748046875,
0.67041015625
],
[
0.1748046875,
0.67041015625,
0.82421875,
0.67041015625,
0.82421875,
0.6845703125,
0.1748046875,
0.6845703125
],
[
0.1748046875,
0.6845703125,
0.826171875,
0.6845703125,
0.826171875,
0.69775390625,
0.1748046875,
0.69775390625
],
[
0.1748046875,
0.6962890625,
0.826171875,
0.6962890625,
0.826171875,
0.71044921875,
0.1748046875,
0.71044921875
],
[
0.1748046875,
0.7109375,
0.82177734375,
0.7109375,
0.82177734375,
0.7216796875,
0.1748046875,
0.7216796875
],
[
0.1748046875,
0.724609375,
0.82421875,
0.724609375,
0.82421875,
0.7392578125,
0.1748046875,
0.7392578125
],
[
0.1748046875,
0.7392578125,
0.82421875,
0.7392578125,
0.82421875,
0.75244140625,
0.1748046875,
0.75244140625
],
[
0.1748046875,
0.75244140625,
0.301513671875,
0.75244140625,
0.301513671875,
0.7666015625,
0.1748046875,
0.7666015625
],
[
0.17236328125,
0.77099609375,
0.82421875,
0.77294921875,
0.82421875,
0.787109375,
0.17236328125,
0.78515625
],
[
0.1732177734375,
0.7861328125,
0.32275390625,
0.7861328125,
0.32275390625,
0.7998046875,
0.1732177734375,
0.7998046875
],
[
0.2156982421875,
0.81640625,
0.2344970703125,
0.81640625,
0.2344970703125,
0.82470703125,
0.2156982421875,
0.82470703125
],
[
0.2303466796875,
0.81298828125,
0.826171875,
0.81298828125,
0.826171875,
0.82763671875,
0.2303466796875,
0.82763671875
],
[
0.233642578125,
0.8291015625,
0.39208984375,
0.8291015625,
0.39208984375,
0.84228515625,
0.233642578125,
0.84228515625
],
[
0.2115478515625,
0.84765625,
0.82421875,
0.84765625,
0.82421875,
0.8623046875,
0.2115478515625,
0.8623046875
],
[
0.233642578125,
0.8623046875,
0.490234375,
0.8623046875,
0.490234375,
0.87548828125,
0.233642578125,
0.87548828125
],
[
0.1968994140625,
0.88623046875,
0.63818359375,
0.88623046875,
0.63818359375,
0.90087890625,
0.1968994140625,
0.90087890625
],
[
0.1732177734375,
0.92236328125,
0.81787109375,
0.923828125,
0.81787109375,
0.93798828125,
0.1732177734375,
0.9375
]
] |
[
"Probabilistic Neural Programs",
"Kenton W. Murray *",
"JayantKrishnamurthy",
"Department of Computer Science and Engineering",
"Allen Institute for Artificial Intelligence",
"University of Notre Dame",
"Seattle, WA 98166",
"South Bend, IN 46617",
"jayantk@allenai.org",
"kmurray4@nd.edu",
"91000",
"Abstract",
"We present probabilistic neural programs, a framework for program induction that",
"permits flexible specification of both a computational model and inference algo-",
"2",
"rithm while simultaneously enabling the use of deep neural networks. Probabilistic",
"neural programs combinea computation graphfor specifyinga neural network with",
"[AN'S]",
"an operator for weighted nondeterministic choice. Thus, a program describes both",
"a collection of decisions as well as the neural network architecture used to make",
"each one. We evaluate our approach on a challenging diagram question answering",
"task where probabilistic neural programs correctly execute nearly twice as many",
"programs as a baseline model.",
"1AZ100Z191:AIX",
"",
"-",
"Introduction",
"In recent years, deep learning has produced tremendous accuracy improvements on a variety of",
"tasks in computer vision and natural language processing. A natural next step for deep learning is to",
"considerprograminduction,theproblem of learningcomputerprogramsfrom(noisy)input/output",
"examples.Compared to more traditional problems, such as object recognition that requiremaking",
"only a single decision, program induction is difficult because it requires making a sequence of",
"decisions and possibly learning control flow concepts such as loops and if statements.",
"Prior work on program induction has described two general classes of approaches. First, in the",
"noise-free setting, program synthesis approaches pose program induction as completing a program",
"\"sketch,which is a program containing nondeterministic choices (\"holes\") to be filled by the learning",
"algorithm [13]. Probabilistic programming languages generalize this approach to the noisy setting by",
"permitting the sketch to specify a distribution over these choices as a function of prior parameters",
"and further to condition this distribution on data, thereby training a Bayesian generative model to",
"execute the sketch correctly [6]. Second, neural abstract machines define continuous analogues of",
"Turingmachines or other general-purpose computational models by\"lifting\"their discrete state and",
"computation rules into a continuous representation [9, 11, 7, 12]. Both of these approaches have",
"demonstrated success at inducing simple programs from synthetic data but have yet to be applied to",
"practical problems.",
"We observe that there are (at least) three dimensions along which we can characterize program",
"induction approaches:",
"1.",
"Computational Model - what abstract model of computation does the model learn to control?",
"(e.g., a Turing machine)",
"2. Learning Mechanism - what kinds of machine learning models are supported? (e.g., neural",
"networks, Bayesian generative models)",
"*Work done while on Internship at Allen Institute for Artificial Intelligence",
"Ist Workshop on Neural Abstract Machines & Program Induction (NAMPI), @NIPS 2016, Barcelona, Spain."
] |
{
"arxiv_id": "1612.00712",
"language": "en",
"page_idx": 0
}
|
||
1612.00712-en-1
|
1612.00712-en-1.jpg
|
<FigureHere>
(left) and applying it to create a probability distribution over program executions (right).
3. Inference - how does the approach reason about the many possible executions of the
machine?
Neural abstract machines conflate some of these dimensions: they naturally support deep learning, but commit to a particular computational model and approximate inference algorithm. These choices are suboptimal as (1) the bias/variance trade-off suggests that training a more expressive computational model will require more data than a less expressive one suited to the task at hand, and (2) recent work has suggested that discrete inference algorithms may outperform continuous approximations [5]. In contrast, probabilistic programming supports the specification of different (possibly taskspecific) computational models and inference algorithms, including discrete search and continuous approximations. However, these languages are restricted to generative models and cannot leverage the power of deep neural networks.
We present probabilistic neural programs, a framework for program induction that permits flexible specification of the computational model and inference algorithm while simultaneously enabling $\hat{\text{the use of deep neural networks. Our approach builds on computation graph frameworks[1,3] for}}$ specifying neural networks by adding an operator for weighted nondeterministic choice that is used to specify the computational model. Thus, a program sketch describes both the decisions to be made and the architecture of the neural network used to score these decisions. Importantly, the computation graph interacts with nondeterminism: the scores produced by the neural network determine the weights of nondeterministic choices, while the choices determine the network's architecture. As with probabilistic programs, various inference algorithms can be applied to a sketch. Furthermore, a sketch's neural network parameters can be estimated using stochastic gradient descent from either input/output examples or full execution traces.
We evaluate our approach on a challenging diagram question answering task, which recent work has demonstrated can be formulated as learning to execute a certain class of probabilistic programs. On this task, we find that the enhanced modeling power of neural networks improves accuracy.
## 2 $\textbf{Probabilistic Neural Programs}$
Probabilistic neural programs build on computation graph frameworks for specifying neural networks by adding an operator for nondeterministic choice. We have developed a Scala library for probabilistic neural programming that we use to illustrate the key concepts.
Figure 1 (left) defines a multilayer perceptron as a probabilistic neural program. This definition resembles those of other computation graph frameworks. Network parameters and intermediate values are represented as computation graph nodes with tensor values. They can be manipulated with standard operations such as matrix-vector multiplication and hyperbolic tangent. Evaluating this function with a tensor yields a program sketch object that can be evaluated with a set of network parameters to produce the network's output.
Figure 1 (right) shows how to use the choose function to create a nondeterministic choice. This function nondeterministically selects a value from a list of options. The score of each option is given by the value of a computation graph node that has the same number of elements as the list. Evaluating this function with a tensor yields a program sketch object that represents a function
2
|
[
[
0.1748046875,
0.09661865234375,
0.343017578125,
0.09661865234375,
0.343017578125,
0.111083984375,
0.1748046875,
0.111083984375
],
[
0.468017578125,
0.09661865234375,
0.68115234375,
0.09661865234375,
0.68115234375,
0.10986328125,
0.468017578125,
0.10986328125
],
[
0.67626953125,
0.0997314453125,
0.68896484375,
0.0997314453125,
0.68896484375,
0.10797119140625,
0.67626953125,
0.10797119140625
],
[
0.19287109375,
0.111083984375,
0.285888671875,
0.111083984375,
0.285888671875,
0.12249755859375,
0.19287109375,
0.12249755859375
],
[
0.281982421875,
0.11297607421875,
0.298095703125,
0.11297607421875,
0.298095703125,
0.12060546875,
0.281982421875,
0.12060546875
],
[
0.5048828125,
0.11297607421875,
0.52392578125,
0.11297607421875,
0.52392578125,
0.12060546875,
0.5048828125,
0.12060546875
],
[
0.521484375,
0.10986328125,
0.7060546875,
0.10986328125,
0.7060546875,
0.12371826171875,
0.521484375,
0.12371826171875
],
[
0.193603515625,
0.12371826171875,
0.226318359375,
0.12371826171875,
0.226318359375,
0.1351318359375,
0.193603515625,
0.1351318359375
],
[
0.486083984375,
0.11553955078125,
0.5048828125,
0.11236572265625,
0.50732421875,
0.12060546875,
0.48779296875,
0.12249755859375
],
[
0.50341796875,
0.12371826171875,
0.53515625,
0.12371826171875,
0.53515625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.52783203125,
0.12371826171875,
0.6689453125,
0.12371826171875,
0.6689453125,
0.136962890625,
0.52783203125,
0.136962890625
],
[
0.6640625,
0.125,
0.72216796875,
0.125,
0.72216796875,
0.1363525390625,
0.6640625,
0.1363525390625
],
[
0.488525390625,
0.1256103515625,
0.50634765625,
0.1256103515625,
0.50634765625,
0.1337890625,
0.488525390625,
0.1337890625
],
[
0.226318359375,
0.126953125,
0.2344970703125,
0.126953125,
0.2344970703125,
0.1319580078125,
0.226318359375,
0.1319580078125
],
[
0.2099609375,
0.1363525390625,
0.2303466796875,
0.1363525390625,
0.2303466796875,
0.149658203125,
0.2099609375,
0.149658203125
],
[
0.25732421875,
0.1351318359375,
0.359375,
0.1331787109375,
0.359375,
0.147705078125,
0.25732421875,
0.149658203125
],
[
0.52294921875,
0.1363525390625,
0.66650390625,
0.1363525390625,
0.66650390625,
0.149658203125,
0.52294921875,
0.149658203125
],
[
0.50634765625,
0.1383056640625,
0.52294921875,
0.1383056640625,
0.52294921875,
0.146484375,
0.50634765625,
0.146484375
],
[
0.23779296875,
0.1383056640625,
0.256591796875,
0.1383056640625,
0.256591796875,
0.146484375,
0.23779296875,
0.146484375
],
[
0.490234375,
0.1395263671875,
0.495849609375,
0.14013671875,
0.494384765625,
0.146484375,
0.488525390625,
0.1458740234375
],
[
0.6640625,
0.1383056640625,
0.7236328125,
0.1383056640625,
0.7236328125,
0.1483154296875,
0.6640625,
0.1483154296875
],
[
0.2099609375,
0.1483154296875,
0.2303466796875,
0.1483154296875,
0.2303466796875,
0.1610107421875,
0.2099609375,
0.1610107421875
],
[
0.2548828125,
0.1483154296875,
0.359375,
0.146484375,
0.359375,
0.160400390625,
0.2548828125,
0.1622314453125
],
[
0.48046875,
0.147705078125,
0.5498046875,
0.147705078125,
0.5498046875,
0.1622314453125,
0.48046875,
0.1622314453125
],
[
0.4697265625,
0.15283203125,
0.48046875,
0.149658203125,
0.484375,
0.1558837890625,
0.473876953125,
0.1590576171875
],
[
0.238525390625,
0.15087890625,
0.256591796875,
0.15087890625,
0.256591796875,
0.1590576171875,
0.238525390625,
0.1590576171875
],
[
0.2091064453125,
0.1610107421875,
0.2303466796875,
0.1610107421875,
0.2303466796875,
0.173583984375,
0.2091064453125,
0.173583984375
],
[
0.25244140625,
0.1622314453125,
0.339111328125,
0.1622314453125,
0.339111328125,
0.173583984375,
0.25244140625,
0.173583984375
],
[
0.339111328125,
0.1622314453125,
0.4306640625,
0.1622314453125,
0.4306640625,
0.17236328125,
0.339111328125,
0.17236328125
],
[
0.486083984375,
0.1610107421875,
0.53515625,
0.1610107421875,
0.53515625,
0.175537109375,
0.486083984375,
0.175537109375
],
[
0.253173828125,
0.17236328125,
0.358642578125,
0.17041015625,
0.358642578125,
0.1849365234375,
0.253173828125,
0.1875
],
[
0.2344970703125,
0.1641845703125,
0.2548828125,
0.1610107421875,
0.256591796875,
0.169189453125,
0.23779296875,
0.1717529296875
],
[
0.2099609375,
0.1749267578125,
0.2279052734375,
0.17236328125,
0.2296142578125,
0.1817626953125,
0.2115478515625,
0.184326171875
],
[
0.238525390625,
0.175537109375,
0.25732421875,
0.175537109375,
0.25732421875,
0.18310546875,
0.238525390625,
0.18310546875
],
[
0.253173828125,
0.1849365234375,
0.359375,
0.184326171875,
0.359375,
0.1988525390625,
0.253173828125,
0.199462890625
],
[
0.490234375,
0.1849365234375,
0.7998046875,
0.1849365234375,
0.7998046875,
0.1988525390625,
0.490234375,
0.1988525390625
],
[
0.47216796875,
0.1881103515625,
0.495849609375,
0.1881103515625,
0.495849609375,
0.1962890625,
0.47216796875,
0.1962890625
],
[
0.2099609375,
0.186279296875,
0.2303466796875,
0.186279296875,
0.2303466796875,
0.1988525390625,
0.2099609375,
0.1988525390625
],
[
0.240234375,
0.189453125,
0.2484130859375,
0.189453125,
0.2484130859375,
0.1932373046875,
0.240234375,
0.1932373046875
],
[
0.261474609375,
0.1976318359375,
0.38818359375,
0.1962890625,
0.38818359375,
0.2108154296875,
0.261474609375,
0.212158203125
],
[
0.46728515625,
0.1962890625,
0.64794921875,
0.1976318359375,
0.64794921875,
0.212158203125,
0.46728515625,
0.2108154296875
],
[
0.20751953125,
0.210205078125,
0.259033203125,
0.210205078125,
0.259033203125,
0.2247314453125,
0.20751953125,
0.2247314453125
],
[
0.2099609375,
0.1988525390625,
0.2418212890625,
0.1988525390625,
0.2418212890625,
0.2108154296875,
0.2099609375,
0.2108154296875
],
[
0.1953125,
0.2139892578125,
0.2132568359375,
0.2139892578125,
0.2132568359375,
0.222900390625,
0.1953125,
0.222900390625
],
[
0.25732421875,
0.2139892578125,
0.27294921875,
0.2139892578125,
0.27294921875,
0.2215576171875,
0.25732421875,
0.2215576171875
],
[
0.4697265625,
0.2108154296875,
0.78515625,
0.2108154296875,
0.78515625,
0.2247314453125,
0.4697265625,
0.2247314453125
],
[
0.491943359375,
0.222900390625,
0.56640625,
0.2247314453125,
0.56640625,
0.2392578125,
0.491943359375,
0.2374267578125
],
[
0.5810546875,
0.222900390625,
0.73388671875,
0.2235107421875,
0.73388671875,
0.238037109375,
0.5810546875,
0.2374267578125
],
[
0.4697265625,
0.2266845703125,
0.488525390625,
0.222900390625,
0.491943359375,
0.23291015625,
0.473876953125,
0.236083984375
],
[
0.2099609375,
0.2247314453125,
0.23779296875,
0.2247314453125,
0.23779296875,
0.236083984375,
0.2099609375,
0.236083984375
],
[
0.56201171875,
0.2266845703125,
0.58837890625,
0.2266845703125,
0.58837890625,
0.23486328125,
0.56201171875,
0.23486328125
],
[
0.583984375,
0.2374267578125,
0.650390625,
0.2374267578125,
0.650390625,
0.248779296875,
0.583984375,
0.248779296875
],
[
0.64794921875,
0.238037109375,
0.7236328125,
0.238037109375,
0.7236328125,
0.248779296875,
0.64794921875,
0.248779296875
],
[
0.57861328125,
0.240478515625,
0.58642578125,
0.240478515625,
0.58642578125,
0.24560546875,
0.57861328125,
0.24560546875
],
[
0.47216796875,
0.2392578125,
0.486083984375,
0.2392578125,
0.486083984375,
0.2474365234375,
0.47216796875,
0.2474365234375
],
[
0.56640625,
0.2412109375,
0.57275390625,
0.238037109375,
0.57861328125,
0.24365234375,
0.57177734375,
0.246826171875
],
[
0.1748046875,
0.251953125,
0.82421875,
0.251953125,
0.82421875,
0.26513671875,
0.1748046875,
0.26513671875
],
[
0.1732177734375,
0.266357421875,
0.7490234375,
0.266357421875,
0.7490234375,
0.27978515625,
0.1732177734375,
0.27978515625
],
[
0.2132568359375,
0.294189453125,
0.82421875,
0.294189453125,
0.82421875,
0.309326171875,
0.2132568359375,
0.309326171875
],
[
0.2303466796875,
0.309326171875,
0.2998046875,
0.306884765625,
0.2998046875,
0.3212890625,
0.2303466796875,
0.323974609375
],
[
0.1748046875,
0.335205078125,
0.82275390625,
0.335205078125,
0.82275390625,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1732177734375,
0.349853515625,
0.82421875,
0.348388671875,
0.82421875,
0.363037109375,
0.1732177734375,
0.3642578125
],
[
0.1732177734375,
0.3623046875,
0.82421875,
0.363037109375,
0.82421875,
0.37744140625,
0.1732177734375,
0.376953125
],
[
0.176513671875,
0.37890625,
0.82275390625,
0.37890625,
0.82275390625,
0.389404296875,
0.176513671875,
0.389404296875
],
[
0.1732177734375,
0.39013671875,
0.826171875,
0.391357421875,
0.826171875,
0.406005859375,
0.1732177734375,
0.40478515625
],
[
0.176513671875,
0.40478515625,
0.826171875,
0.40478515625,
0.826171875,
0.419189453125,
0.176513671875,
0.419189453125
],
[
0.1748046875,
0.41845703125,
0.82421875,
0.41845703125,
0.82421875,
0.431884765625,
0.1748046875,
0.431884765625
],
[
0.1732177734375,
0.43310546875,
0.82421875,
0.43310546875,
0.82421875,
0.4462890625,
0.1732177734375,
0.4462890625
],
[
0.1748046875,
0.4462890625,
0.406005859375,
0.4462890625,
0.406005859375,
0.460205078125,
0.1748046875,
0.460205078125
],
[
0.1748046875,
0.466552734375,
0.82421875,
0.466552734375,
0.82421875,
0.48095703125,
0.1748046875,
0.48095703125
],
[
0.1748046875,
0.48095703125,
0.82421875,
0.48095703125,
0.82421875,
0.49560546875,
0.1748046875,
0.49560546875
],
[
0.1748046875,
0.494384765625,
0.82421875,
0.494384765625,
0.82421875,
0.50830078125,
0.1748046875,
0.50830078125
],
[
0.1748046875,
0.51025390625,
0.82177734375,
0.51025390625,
0.82177734375,
0.52099609375,
0.1748046875,
0.52099609375
],
[
0.1748046875,
0.521484375,
0.82421875,
0.521484375,
0.82421875,
0.5361328125,
0.1748046875,
0.5361328125
],
[
0.1748046875,
0.5361328125,
0.82275390625,
0.537109375,
0.82275390625,
0.548828125,
0.1748046875,
0.54736328125
],
[
0.1748046875,
0.55029296875,
0.82421875,
0.55029296875,
0.82421875,
0.564453125,
0.1748046875,
0.564453125
],
[
0.1748046875,
0.56298828125,
0.82421875,
0.56298828125,
0.82421875,
0.5771484375,
0.1748046875,
0.5771484375
],
[
0.1732177734375,
0.5771484375,
0.826171875,
0.57763671875,
0.826171875,
0.59130859375,
0.1732177734375,
0.59033203125
],
[
0.1748046875,
0.59130859375,
0.82421875,
0.59130859375,
0.82421875,
0.60498046875,
0.1748046875,
0.60498046875
],
[
0.1748046875,
0.60498046875,
0.4794921875,
0.60498046875,
0.4794921875,
0.61865234375,
0.1748046875,
0.61865234375
],
[
0.1748046875,
0.62548828125,
0.82421875,
0.62548828125,
0.82421875,
0.6396484375,
0.1748046875,
0.6396484375
],
[
0.1732177734375,
0.63818359375,
0.82275390625,
0.63818359375,
0.82275390625,
0.65283203125,
0.1732177734375,
0.65283203125
],
[
0.1732177734375,
0.6513671875,
0.7685546875,
0.65283203125,
0.7685546875,
0.66748046875,
0.1732177734375,
0.66650390625
],
[
0.176513671875,
0.689453125,
0.1895751953125,
0.689453125,
0.1895751953125,
0.69970703125,
0.176513671875,
0.69970703125
],
[
0.200927734375,
0.68603515625,
0.461669921875,
0.68798828125,
0.461669921875,
0.70263671875,
0.200927734375,
0.70068359375
],
[
0.1732177734375,
0.71826171875,
0.82275390625,
0.71826171875,
0.82275390625,
0.73193359375,
0.1732177734375,
0.73193359375
],
[
0.1748046875,
0.734375,
0.82421875,
0.734375,
0.82421875,
0.74755859375,
0.1748046875,
0.74755859375
],
[
0.1748046875,
0.74755859375,
0.580078125,
0.74755859375,
0.580078125,
0.76220703125,
0.1748046875,
0.76220703125
],
[
0.1748046875,
0.767578125,
0.82421875,
0.767578125,
0.82421875,
0.7822265625,
0.1748046875,
0.7822265625
],
[
0.1748046875,
0.7822265625,
0.82421875,
0.7822265625,
0.82421875,
0.79541015625,
0.1748046875,
0.79541015625
],
[
0.1748046875,
0.79541015625,
0.82421875,
0.79541015625,
0.82421875,
0.8095703125,
0.1748046875,
0.8095703125
],
[
0.1732177734375,
0.80810546875,
0.82421875,
0.8095703125,
0.82421875,
0.82373046875,
0.1732177734375,
0.82275390625
],
[
0.1732177734375,
0.82373046875,
0.82421875,
0.82373046875,
0.82421875,
0.8369140625,
0.1732177734375,
0.8369140625
],
[
0.1732177734375,
0.8369140625,
0.465576171875,
0.8359375,
0.465576171875,
0.85107421875,
0.1732177734375,
0.8515625
],
[
0.1732177734375,
0.85595703125,
0.82421875,
0.85693359375,
0.82421875,
0.8720703125,
0.1732177734375,
0.87060546875
],
[
0.1748046875,
0.8720703125,
0.826171875,
0.8720703125,
0.826171875,
0.88525390625,
0.1748046875,
0.88525390625
],
[
0.1748046875,
0.88525390625,
0.826171875,
0.88525390625,
0.826171875,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1748046875,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1748046875,
0.912109375
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"def mlp(v: Tensor):",
"val dist:Pp[Int] =for",
"",
"Pp[CgNode]",
"",
"<-",
"mlp(newTensor(...))",
"for",
"S",
"<-",
"choose(Array(0,",
"1), s)",
"V",
"",
"w1",
"param(\"wl\")",
"choose(Array(2,",
"<-",
"<-",
"Y",
"3),s)",
"b1",
"param(\"bi\")",
"yield (",
"/",
"<-",
"h1",
"((wl * v)",
"+bl).tanh",
"v+y",
"param(\"w2\")",
"I",
"w2",
"<-",
"param(\"b2\")",
"tensor parameters initialized to o",
"",
"b2",
"2",
"(w2*hl）+b2",
"val params:NnParams",
"yield",
"out",
"",
"",
"println(dist.beamsearch(10,params))",
"output:",
"(0.25),3 (0.25),",
"",
"out",
"2",
"(0.25),",
"4(0.25)",
"",
"",
"U",
"Figure l: Probabilistic neural programs defining a multilayer perceptron with a computation graph",
"(left) and applying it to create a probability distribution over program executions (right).",
"3. Inference - how does the approach reason about the many possible executions of the",
"machine?",
"Neural abstract machines conflate some of these dimensions: they naturally support deep learning, but",
"commit to a particular computational model and approximate inference algorithm. These choices are",
"suboptimal as (1) the bias/variance trade-off suggests that training a more expressive computational",
"model will require more data than aless expressive one suited to the task at hand, and (2)recent",
"work has suggested that discrete inference algorithms may outperform continuous approximations",
"[5]. In contrast, probabilistic programming supports the specification of different (possibly task-",
"specific) computational models and inference algorithms, including discrete search and continuous",
"approximations. However, these languages are restricted to generative models and cannot leverage",
"the power of deep neural networks.",
"We present probabilistic neural programs, a framework for program induction that permits flexible",
"specification of the computational model and inference algorithm while simultaneously enabling",
"the use of deep neural networks. Our approach builds on computation graph frameworks [1, 3] for",
"specifyingneuralnetworksbyadding anoperatorforweightednondeterministic choicethatisused",
"to specify the computational model. Thus, a program sketch describes both the decisions to be made",
"andthearchitectureoftheneuralnetworkusedtoscorethesedecisions.Importantly,thecomputation",
"graph interacts with nondeterminism: the scores produced by the neural network determine the",
"weights of nondeterministic choices, while the choices determine the network's architecture, As",
"with probabilisticprograms,various inference algorithms can be applied to a sketch.Furthermore, a",
"sketch's neural network parameters can be estimated using stochastic gradient descent from either",
"input/output examples or full execution traces.",
"We evaluate our approach on a challenging diagram question answering task, which recent work has",
"demonstrated can be formulated as learning to execute a certain class of probabilistic programs. On",
"this task, we find that the enhanced modeling power of neural networks improves accuracy.",
"2",
"Probabilistic Neural Programs",
"Probabilistic neural programs build on computation graph frameworks for specifying neural networks",
"by adding an operator for nondeterministic choice.We have developed a Scala library for probabilistic",
"neural programming that we use to illustrate the key concepts.",
"Figure 1 (left) defines a multilayer perceptron as a probabilistic neural program. This definition",
"resembles those of other computation graph frameworks. Network parameters and intermediate",
"values are represented as computation graph nodes with tensor values. They can be manipulated",
"with standard operations such as matrix-vector multiplication and hyperbolic tangent. Evaluating",
"this function with a tensor yields a program sketch object that can be evaluated with a set of network",
"parameters to produce the network's output.",
"Figure 1 (right) shows how to use the choose function to create a nondeterministic choice. This",
"function nondeterministically selects a value from a list of options. The score of each option is",
"given by the value of a computation graph node that has the same number of elements as the list.",
"Evaluating this function with a tensor yields a program sketch object that represents a function",
"2"
] |
{
"arxiv_id": "1612.00712",
"language": "en",
"page_idx": 1
}
|
||
1612.00712-en-2
|
1612.00712-en-2.jpg
|
<FigureHere>
What happens to the snake population if the
field mouse population decreases?
$\lambda f$.cause$(\tilde{\text{decrease}}($mice$),f($snakes))
Does the deer eat the grass?
eats(deer,grass)
How many organisms eat the grass?
count$(\lambda x$.eats$(x$,grass)) Are rabbits a tertiary consumer? tertiary-consumer(rabbits)
Figure 2: A food web diagram with annotations generated from a computer vision system (left) along
with related questions and their associated program sketches (right).
from neural network parameters to a $probability\textit{distribution over values. The log probability of a}$ value is proportional to the sum of the scores of the choices made in the execution that produced it. Performing (approximate) inference over this object- in this case, using beam search- produces an explicit representation of the distribution. Multiple nondeterministic choices can be combined to produce more complex sketches; this capability can be used to define complex computational models, including general-purpose models such as Turing machines. The library also has functions for conditioning on observations.
Although various inference algorithms may be applied to a program sketch, in this work we use a simple beam search over executions. This approach accords with the recent trend in structured prediction to combine greedy inference or beam search with powerful non-factoring models [2,10,4]. $\hat{\text{The beam search maintains a queue of partial program executions, each of which is associated with a}}$ score. Each step of the search continues each execution until it encounters a call to choose, which adds zero or more executions to the queue for the next search step. The lowest scoring executions are discarded to maintain a fixed beam width. As an execution proceeds, it may generate new computation graph nodes; the search maintains a single computation graph shared by all executions to which these nodes are added. The search simultaneously performs the forward pass over these nodes as necessary to compute scores for future choices.
The neural network parameters are trained to maximize the loglikelihood of correct program executions using stochastic gradient descent. Each training example consists of a pair of program sketches, representing an unconditional and conditional distribution. The gradient computation is similar to that of a loglinear model with neural network factors. It first performs inference on both the conditional and unconditional distributions to estimate the expected counts associated with each nondeterministic choice. These counts are then backpropagated through the computation graph to update the network parameters.
3 Diagram Question Answering with Probabilistic Neural Programs
We consider the problem of learning to execute program sketches in a food web computational model using visual information from a diagram. This problem is motivated by recent work [8], which has demonstrated that diagram question answering can be formulated as translating natural language questions to program sketches in this model, then learning to execute these sketches. Figure 2 shows some example questions from this work, along with the accompanying diagram that must be interpreted to determine the answers. The diagram (left) is a food web, which depicts a collection of organisms in an ecosystem with arrows to indicate what each organism eats. The right side of the figure shows questions pertaining to the diagram and their associated program sketches.
The possible executions of each program sketch are determined by a domain-specific computational model that is designed to reason about food webs. The nondeterministic choices in this model correspond to information that must be extracted from the diagram. Specifically, there are two functions that call choose to nondeterministically return a boolean value. The first function, organism$(x)$,should return true if the text label $\dot{x}$ is an organism (as opposed to e.g., the image title). The second function, eat$(x,y)$,should return true if organism $x$ eats organism $y.$ These functions do influence program control flow. The food web model also includes various other
3
|
[
[
0.5146484375,
0.0889892578125,
0.80810546875,
0.0889892578125,
0.80810546875,
0.1029052734375,
0.5146484375,
0.1029052734375
],
[
0.51318359375,
0.1029052734375,
0.73779296875,
0.10418701171875,
0.73779296875,
0.117431640625,
0.51318359375,
0.11614990234375
],
[
0.5146484375,
0.117431640625,
0.81884765625,
0.117431640625,
0.81884765625,
0.1319580078125,
0.5146484375,
0.1319580078125
],
[
0.5146484375,
0.14453125,
0.69775390625,
0.14453125,
0.69775390625,
0.1590576171875,
0.5146484375,
0.1590576171875
],
[
0.5146484375,
0.160400390625,
0.6650390625,
0.160400390625,
0.6650390625,
0.173583984375,
0.5146484375,
0.173583984375
],
[
0.5146484375,
0.186279296875,
0.7509765625,
0.186279296875,
0.7509765625,
0.2008056640625,
0.5146484375,
0.2008056640625
],
[
0.51318359375,
0.2008056640625,
0.71728515625,
0.2008056640625,
0.71728515625,
0.21533203125,
0.51318359375,
0.21533203125
],
[
0.22216796875,
0.2139892578125,
0.233642578125,
0.2139892578125,
0.233642578125,
0.2171630859375,
0.22216796875,
0.2171630859375
],
[
0.5146484375,
0.2279052734375,
0.72607421875,
0.2279052734375,
0.72607421875,
0.2412109375,
0.5146484375,
0.2412109375
],
[
0.5146484375,
0.242431640625,
0.76318359375,
0.242431640625,
0.76318359375,
0.255615234375,
0.5146484375,
0.255615234375
],
[
0.1748046875,
0.25830078125,
0.82421875,
0.2587890625,
0.82421875,
0.2734375,
0.1748046875,
0.272705078125
],
[
0.1748046875,
0.2734375,
0.62158203125,
0.2734375,
0.62158203125,
0.287353515625,
0.1748046875,
0.287353515625
],
[
0.1748046875,
0.3017578125,
0.82275390625,
0.3017578125,
0.82275390625,
0.31494140625,
0.1748046875,
0.31494140625
],
[
0.1748046875,
0.31640625,
0.82275390625,
0.31640625,
0.82275390625,
0.330078125,
0.1748046875,
0.330078125
],
[
0.1748046875,
0.330078125,
0.82421875,
0.330078125,
0.82421875,
0.343505859375,
0.1748046875,
0.343505859375
],
[
0.1748046875,
0.343505859375,
0.82275390625,
0.343505859375,
0.82275390625,
0.356689453125,
0.1748046875,
0.356689453125
],
[
0.176513671875,
0.359130859375,
0.82275390625,
0.359130859375,
0.82275390625,
0.369384765625,
0.176513671875,
0.369384765625
],
[
0.1748046875,
0.371826171875,
0.826171875,
0.371826171875,
0.826171875,
0.385009765625,
0.1748046875,
0.385009765625
],
[
0.1748046875,
0.385009765625,
0.39453125,
0.385009765625,
0.39453125,
0.3984375,
0.1748046875,
0.3984375
],
[
0.176513671875,
0.40478515625,
0.82275390625,
0.40478515625,
0.82275390625,
0.41845703125,
0.176513671875,
0.41845703125
],
[
0.1748046875,
0.419189453125,
0.82421875,
0.419189453125,
0.82421875,
0.43310546875,
0.1748046875,
0.43310546875
],
[
0.1748046875,
0.43310546875,
0.82421875,
0.43310546875,
0.82421875,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.1748046875,
0.4462890625,
0.82421875,
0.4462890625,
0.82421875,
0.460205078125,
0.1748046875,
0.460205078125
],
[
0.176513671875,
0.462158203125,
0.82177734375,
0.462158203125,
0.82177734375,
0.472900390625,
0.176513671875,
0.472900390625
],
[
0.176513671875,
0.47607421875,
0.82275390625,
0.47607421875,
0.82275390625,
0.486083984375,
0.176513671875,
0.486083984375
],
[
0.176513671875,
0.4892578125,
0.82275390625,
0.4892578125,
0.82275390625,
0.5,
0.176513671875,
0.5
],
[
0.1748046875,
0.501953125,
0.82421875,
0.501953125,
0.82421875,
0.51513671875,
0.1748046875,
0.51513671875
],
[
0.1748046875,
0.5166015625,
0.82421875,
0.5166015625,
0.82421875,
0.52978515625,
0.1748046875,
0.52978515625
],
[
0.1748046875,
0.52978515625,
0.418212890625,
0.52978515625,
0.418212890625,
0.54345703125,
0.1748046875,
0.54345703125
],
[
0.1748046875,
0.5498046875,
0.82666015625,
0.5498046875,
0.82666015625,
0.56298828125,
0.1748046875,
0.56298828125
],
[
0.1748046875,
0.56298828125,
0.826171875,
0.56298828125,
0.826171875,
0.57763671875,
0.1748046875,
0.57763671875
],
[
0.1732177734375,
0.57763671875,
0.82421875,
0.5771484375,
0.82421875,
0.59033203125,
0.1732177734375,
0.59130859375
],
[
0.1748046875,
0.59130859375,
0.82421875,
0.59130859375,
0.82421875,
0.60498046875,
0.1748046875,
0.60498046875
],
[
0.1748046875,
0.60595703125,
0.82275390625,
0.60595703125,
0.82275390625,
0.619140625,
0.1748046875,
0.619140625
],
[
0.1732177734375,
0.61865234375,
0.82275390625,
0.61865234375,
0.82275390625,
0.6318359375,
0.1732177734375,
0.6318359375
],
[
0.1732177734375,
0.6337890625,
0.25244140625,
0.6337890625,
0.25244140625,
0.64453125,
0.1732177734375,
0.64453125
],
[
0.1707763671875,
0.6640625,
0.75732421875,
0.66650390625,
0.75732421875,
0.6845703125,
0.1707763671875,
0.681640625
],
[
0.1732177734375,
0.69775390625,
0.82275390625,
0.69775390625,
0.82275390625,
0.7109375,
0.1732177734375,
0.7109375
],
[
0.1732177734375,
0.7119140625,
0.82421875,
0.7119140625,
0.82421875,
0.72607421875,
0.1732177734375,
0.72607421875
],
[
0.1732177734375,
0.72607421875,
0.82275390625,
0.72607421875,
0.82275390625,
0.7392578125,
0.1732177734375,
0.7392578125
],
[
0.1732177734375,
0.7392578125,
0.82177734375,
0.7392578125,
0.82177734375,
0.75244140625,
0.1732177734375,
0.75244140625
],
[
0.1748046875,
0.75390625,
0.82275390625,
0.75390625,
0.82275390625,
0.767578125,
0.1748046875,
0.767578125
],
[
0.1748046875,
0.767578125,
0.826171875,
0.767578125,
0.826171875,
0.78076171875,
0.1748046875,
0.78076171875
],
[
0.1748046875,
0.78076171875,
0.82421875,
0.78076171875,
0.82421875,
0.79443359375,
0.1748046875,
0.79443359375
],
[
0.1748046875,
0.79541015625,
0.74853515625,
0.79541015625,
0.74853515625,
0.8095703125,
0.1748046875,
0.8095703125
],
[
0.1748046875,
0.81494140625,
0.82275390625,
0.81494140625,
0.82275390625,
0.8291015625,
0.1748046875,
0.8291015625
],
[
0.176513671875,
0.83056640625,
0.82275390625,
0.83056640625,
0.82275390625,
0.84130859375,
0.176513671875,
0.84130859375
],
[
0.1748046875,
0.84326171875,
0.82421875,
0.84326171875,
0.82421875,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.1732177734375,
0.85595703125,
0.826171875,
0.85791015625,
0.826171875,
0.8720703125,
0.1732177734375,
0.869140625
],
[
0.1748046875,
0.87060546875,
0.82421875,
0.87060546875,
0.82421875,
0.8837890625,
0.1748046875,
0.8837890625
],
[
0.1732177734375,
0.88330078125,
0.826171875,
0.8837890625,
0.826171875,
0.8984375,
0.1732177734375,
0.89794921875
],
[
0.1732177734375,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"What happens to the snake population if the",
"field mouse population decreases?",
"Af.cause(decrease(mice),f(snakes))",
"Does the deer eat the grass?",
"eats(deer,grass)",
"How many organisms eat the grass?",
"count(Aa.eats(r,grass))",
"",
"Are rabbits a tertiary consumer?",
"tertiary-consumer(rabbits)",
"Figure 2: A food web diagram with annotations generated from a computer vision system (left) along",
"with related questions and their associated program sketches (right).",
"from neural network parameters to a probability distribution over values. The log probability of a",
"value is proportional to the sum of the scores of the choices made in the execution that produced it",
"Performing (approximate) inference over this object - in this case, using beam search - produces",
"an explicit representation of the distribution. Multiple nondeterministic choices can be combined",
"toproducemorecomplexsketches;thiscapabilitycanbeusedtodefinecomplexcomputational",
"models, including general-purpose models such as Turing machines. The library also has functions",
"for conditioning on observations.",
"Although various inference algorithms may be applied to a program sketch, in this work we use",
"a simple beam search over executions. This approach accords with the recent trend in structured",
"prediction to combine greedy inference or beam search with powerful non-factoring models [2, 10, 4]]",
"The beam search maintains a queue of partial program executions, each of which is associated with a",
"score.Eachstepofthesearchcontinueseachexecutionuntilitencountersacalltochoose,which",
"addszeroormoreexecutionstothequeueforthenextsearchstep.Thelowestscoringexecutionsare",
"discardedtomaintain afixed beamwidth.As an executionproceeds,itmaygeneratenew computation",
"graph nodes; the search maintains a single computation graph shared by all executions to which these",
"nodes are added. The search simultaneously performs the forward pass over these nodes as necessary",
"to compute scores for future choices.",
"The neural network parameters are trained to maximize the loglikelihood of correct program execu-",
"tions using stochastic gradient descent. Each training example consists of a pair of program sketches,",
"representing an unconditional and conditional distribution.The gradient computation is similar to that",
"of a loglinear model with neural network factors. It first performs inference on both the conditional",
"and unconditional distributions to estimate the expected counts associated with each nondeterministic",
"choice. These counts are then backpropagated through the computation graph to update the network",
"parameters.",
"3Diagram Question Answering with Probabilistic Neural Programs",
"We consider the problem of learning to execute program sketches in a food web computational",
"model using visual information from a diagram. This problem is motivated by recent work [8],",
"which has demonstrated that diagram question answering can be formulated as translating natural",
"language questions to program sketches in this model, then learning to execute these sketches. Figure",
"2 shows some example questions from this work, along with the accompanying diagram that must be",
"interpreted to determine the answers. The diagram (left) is a food web, which depicts a collection of",
"organisms in an ecosystem with arrows to indicate what each organism eats. The right side of the",
"figure shows questions pertaining to the diagram and their associated program sketches.",
"The possible executions of each program sketch are determined by a domain-specific computational",
"modelthat is designed to reason aboutfood webs.Thenondeterministic choices in this model",
"correspond to information that must be extracted from the diagram. Specifically, there are two",
"functions that call choose to nondeterministically return a boolean value. The first function,",
"organi sm(r), should return true if the text label r is an organism (as opposed to e.g., the image",
"title).The second function, eat(, y), should return true if organism eats organism y. These",
"functions do influence program control flow. The food web model also includes various other",
"3"
] |
{
"arxiv_id": "1612.00712",
"language": "en",
"page_idx": 2
}
|
||
1612.00712-en-3
|
1612.00712-en-3.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Method}$</th>
<th>Execution Accuracy</th>
<th>choose Accuracy</th>
</tr>
<tr>
<th>$\mathbf{Method}$</th>
<th>$\operatorname{Execution}$ Accuracy</th>
<th>choose Accuracy</th>
</tr>
<tr>
<th> </th>
<th>$\operatorname{Execution}$</th>
<th>choose</th>
</tr>
<tr>
<th>$\mathbf{Method}$</th>
<th>$\operatorname{Execution}$ Accuracy</th>
<th>choose Accuracy</th>
</tr>
<tr>
<th>$\mathbf{Method}$</th>
<th>Accuracy</th>
<th>Accuracy</th>
</tr>
<tr>
<td>LOGLINEAR</td>
<td>$8.6\%$</td>
<td>$78.2\%$</td>
</tr>
<tr>
<td>2-LAYER $PNP$</td>
<td>$12.5\%$</td>
<td>$78.7\%$</td>
</tr>
<tr>
<td>$\mathbf{M}_{\mathrm{AXPOOL}}$ ${\mathrm{PNP}}$</td>
<td>$14.9\%$</td>
<td>$82.5\%$</td>
</tr>
</tbody>
</table>
Table 1: Results from appling Probabilistic Neural Programs to the execution of sketches from the FOODWEBS dataset.“choose Accuracy"indicates average accuracy at each decision, whereas “Execution Accuracy" indicates if the entire program was executed correctly.
functions, eg., for reasoning about population changes, that call organism and eat to extract information from the diagram.[8] has a more thorough description of the theory; our goal is to learn to make the choices in this theory.
We consider three models for learning to make the choices for both organism and eat: a non-neural (LOGLINEAR) model, as well as two probabilistic neural models (2-LAYER PNP and MAXPOOL PNP). All three learn models for both organism and eat using outputs from a computer vision system trained to detect organism, text, and arrow relations between them. [8] defines a set of hand-engineered features heuristically created from the outputs of this vision system. LOGLINEAR and 2-LAYER PNP use only these features, and the difference is simply in the greater expressivity of a two-layer neural network. However, one of the major strengths of neural models is their ability to learn latent feature representations automatically, and our third model also uses the direct outputs of the vision system not made into features. The architecture of MAXPOOL PNP reflects this and contains additional input layers that maxpool over detected relationships between objects and confidence scores. The expectation is that our neural network modeling of nondeterminism will learn better latent representations than the manually defined features.
## 4 Experiments
We evaluate probabilistic neural programs on the FOODWEBS dataset introduced by[8]. This data set contains a training set of~2,900 programs and a test set of~1,000 programs. These programs are human annotated gold standard interpretations for the questions in the data set, which corresponds to assuming that the translation from questions to programs is perfect. We train our probabilistic neural programs using correct execution traces of each program, which are also provided in the data set.
We evaluate our models using two metrics. First, execution accuracy measures the fraction of programs in the test set that are executed completely correctly by the model. This metric is challenging because correctly executing a program requires correctly making a number of choose decisions. Our 1,000 test programs had over 35,000 decisions, implying that to completely execute a program correctly means getting on average 35 choose decisions correct without making any mistakes. Second, choose accuracy measures the accuracy of each decision independently, assuming all previous decisions were made correctly.
Table 1 compares the accuracies of our three models on the FOODWEBS dataset. The improvement in accuracy between the baseline (LOGLINEAR) and the probabilistic neural program (2-LAYER PNP) is due to the neural network's enhanced modeling power. Though the choose accuracy does not improve by a large margin, the improvements translate into large gains in entire program correctness. Finally, as expected, the inclusion of lower level features (MAXPOOL PNP) not possible in LOGLINEAR significantly improved performance. Note that this task requires performing computer vision, and thus it is not expected that any model achieve $100\%$ accuracy
## 5 Conclusion
We have presented probabilistic neural programs, a framework for program induction that permits flexible specification of computational models and inference algorithms while simultaneously enabling the use of deep learning. A program sketch describes a collection of nondeterministic decisions to be made during execution, along with the neural architecture to be used for scoring these decisions The network parameters of a sketch can be trained from data using stochastic gradient descent. We demonstrate that probabilistic neural programs improve accuracy on a diagram question answering task which can be formulated as learning to execute program sketches in a domain-specific computational model.
|
[
[
0.48779296875,
0.091552734375,
0.55615234375,
0.091552734375,
0.55615234375,
0.1029052734375,
0.48779296875,
0.1029052734375
],
[
0.57421875,
0.0921630859375,
0.6337890625,
0.0921630859375,
0.6337890625,
0.1029052734375,
0.57421875,
0.1029052734375
],
[
0.358642578125,
0.10479736328125,
0.410888671875,
0.10479736328125,
0.410888671875,
0.11614990234375,
0.358642578125,
0.11614990234375
],
[
0.488525390625,
0.10418701171875,
0.55615234375,
0.10418701171875,
0.55615234375,
0.11871337890625,
0.488525390625,
0.11871337890625
],
[
0.57275390625,
0.1060791015625,
0.63818359375,
0.1060791015625,
0.63818359375,
0.117431640625,
0.57275390625,
0.117431640625
],
[
0.35546875,
0.11614990234375,
0.443603515625,
0.11871337890625,
0.443603515625,
0.1331787109375,
0.35546875,
0.1307373046875
],
[
0.5,
0.11871337890625,
0.54150390625,
0.11871337890625,
0.54150390625,
0.1331787109375,
0.5,
0.1331787109375
],
[
0.5810546875,
0.11871337890625,
0.62841796875,
0.11871337890625,
0.62841796875,
0.1331787109375,
0.5810546875,
0.1331787109375
],
[
0.496826171875,
0.1307373046875,
0.54345703125,
0.1307373046875,
0.54345703125,
0.146484375,
0.496826171875,
0.146484375
],
[
0.580078125,
0.1300048828125,
0.62841796875,
0.1300048828125,
0.62841796875,
0.1483154296875,
0.580078125,
0.1483154296875
],
[
0.358642578125,
0.1458740234375,
0.471435546875,
0.1458740234375,
0.471435546875,
0.160400390625,
0.358642578125,
0.160400390625
],
[
0.35546875,
0.1319580078125,
0.4599609375,
0.1300048828125,
0.4599609375,
0.1458740234375,
0.35546875,
0.147705078125
],
[
0.495849609375,
0.1458740234375,
0.5439453125,
0.1458740234375,
0.5439453125,
0.160400390625,
0.495849609375,
0.160400390625
],
[
0.580078125,
0.1458740234375,
0.62841796875,
0.1458740234375,
0.62841796875,
0.160400390625,
0.580078125,
0.160400390625
],
[
0.1748046875,
0.1622314453125,
0.82275390625,
0.1622314453125,
0.82275390625,
0.1767578125,
0.1748046875,
0.1767578125
],
[
0.1748046875,
0.1767578125,
0.313720703125,
0.1767578125,
0.313720703125,
0.1900634765625,
0.1748046875,
0.1900634765625
],
[
0.314453125,
0.1767578125,
0.82421875,
0.1767578125,
0.82421875,
0.1900634765625,
0.314453125,
0.1900634765625
],
[
0.1707763671875,
0.189453125,
0.67626953125,
0.1900634765625,
0.67626953125,
0.2052001953125,
0.1707763671875,
0.203857421875
],
[
0.1732177734375,
0.212158203125,
0.82275390625,
0.212158203125,
0.82275390625,
0.2259521484375,
0.1732177734375,
0.2259521484375
],
[
0.1732177734375,
0.2259521484375,
0.82421875,
0.2259521484375,
0.82421875,
0.2392578125,
0.1732177734375,
0.2392578125
],
[
0.1732177734375,
0.240478515625,
0.396240234375,
0.240478515625,
0.396240234375,
0.25390625,
0.1732177734375,
0.25390625
],
[
0.1732177734375,
0.260009765625,
0.82421875,
0.260009765625,
0.82421875,
0.2734375,
0.1732177734375,
0.2734375
],
[
0.1748046875,
0.274658203125,
0.82421875,
0.274658203125,
0.82421875,
0.28857421875,
0.1748046875,
0.28857421875
],
[
0.1748046875,
0.28857421875,
0.82421875,
0.28857421875,
0.82421875,
0.3017578125,
0.1748046875,
0.3017578125
],
[
0.1748046875,
0.3017578125,
0.826171875,
0.3017578125,
0.826171875,
0.31494140625,
0.1748046875,
0.31494140625
],
[
0.1732177734375,
0.314453125,
0.82421875,
0.31494140625,
0.82421875,
0.330078125,
0.1732177734375,
0.328857421875
],
[
0.1732177734375,
0.328857421875,
0.826171875,
0.328857421875,
0.826171875,
0.34228515625,
0.1732177734375,
0.34228515625
],
[
0.1748046875,
0.343505859375,
0.826171875,
0.343505859375,
0.826171875,
0.356689453125,
0.1748046875,
0.356689453125
],
[
0.1732177734375,
0.356689453125,
0.82421875,
0.356689453125,
0.82421875,
0.37060546875,
0.1732177734375,
0.37060546875
],
[
0.1748046875,
0.37060546875,
0.82421875,
0.37060546875,
0.82421875,
0.3837890625,
0.1748046875,
0.3837890625
],
[
0.1732177734375,
0.3837890625,
0.82421875,
0.3837890625,
0.82421875,
0.397705078125,
0.1732177734375,
0.397705078125
],
[
0.1748046875,
0.3984375,
0.82421875,
0.3984375,
0.82421875,
0.412353515625,
0.1748046875,
0.412353515625
],
[
0.1748046875,
0.412353515625,
0.5908203125,
0.412353515625,
0.5908203125,
0.425537109375,
0.1748046875,
0.425537109375
],
[
0.176513671875,
0.4462890625,
0.2034912109375,
0.4462890625,
0.2034912109375,
0.458251953125,
0.176513671875,
0.458251953125
],
[
0.1993408203125,
0.44580078125,
0.31201171875,
0.44580078125,
0.31201171875,
0.460205078125,
0.1993408203125,
0.460205078125
],
[
0.1732177734375,
0.47607421875,
0.82421875,
0.47607421875,
0.82421875,
0.4892578125,
0.1732177734375,
0.4892578125
],
[
0.1748046875,
0.490478515625,
0.82421875,
0.490478515625,
0.82421875,
0.50390625,
0.1748046875,
0.50390625
],
[
0.1748046875,
0.50390625,
0.82421875,
0.50390625,
0.82421875,
0.517578125,
0.1748046875,
0.517578125
],
[
0.1732177734375,
0.517578125,
0.82421875,
0.517578125,
0.82421875,
0.53076171875,
0.1732177734375,
0.53076171875
],
[
0.1732177734375,
0.5322265625,
0.80615234375,
0.53076171875,
0.80615234375,
0.54541015625,
0.1732177734375,
0.546875
],
[
0.17236328125,
0.55029296875,
0.82421875,
0.55029296875,
0.82421875,
0.564453125,
0.17236328125,
0.564453125
],
[
0.1748046875,
0.56640625,
0.82421875,
0.56640625,
0.82421875,
0.580078125,
0.1748046875,
0.580078125
],
[
0.1748046875,
0.580078125,
0.826171875,
0.580078125,
0.826171875,
0.59326171875,
0.1748046875,
0.59326171875
],
[
0.1732177734375,
0.59228515625,
0.82275390625,
0.59228515625,
0.82275390625,
0.60595703125,
0.1732177734375,
0.60595703125
],
[
0.1748046875,
0.60791015625,
0.826171875,
0.60791015625,
0.826171875,
0.6220703125,
0.1748046875,
0.6220703125
],
[
0.1732177734375,
0.62060546875,
0.82421875,
0.62060546875,
0.82421875,
0.6337890625,
0.1732177734375,
0.6337890625
],
[
0.1748046875,
0.63525390625,
0.43701171875,
0.63525390625,
0.43701171875,
0.6484375,
0.1748046875,
0.6484375
],
[
0.1732177734375,
0.65380859375,
0.82421875,
0.65478515625,
0.82421875,
0.66943359375,
0.1732177734375,
0.66845703125
],
[
0.1748046875,
0.66845703125,
0.82421875,
0.66845703125,
0.82421875,
0.681640625,
0.1748046875,
0.681640625
],
[
0.176513671875,
0.68310546875,
0.82421875,
0.68310546875,
0.82421875,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.1732177734375,
0.69580078125,
0.82421875,
0.6962890625,
0.82421875,
0.7109375,
0.1732177734375,
0.71044921875
],
[
0.1748046875,
0.71044921875,
0.82421875,
0.71044921875,
0.82421875,
0.7236328125,
0.1748046875,
0.7236328125
],
[
0.1732177734375,
0.724609375,
0.82421875,
0.724609375,
0.82421875,
0.73779296875,
0.1732177734375,
0.73779296875
],
[
0.17236328125,
0.7373046875,
0.654296875,
0.73779296875,
0.654296875,
0.75244140625,
0.17236328125,
0.751953125
],
[
0.178955078125,
0.76953125,
0.191162109375,
0.7744140625,
0.1846923828125,
0.78515625,
0.17236328125,
0.77978515625
],
[
0.2034912109375,
0.771484375,
0.2998046875,
0.771484375,
0.2998046875,
0.7861328125,
0.2034912109375,
0.7861328125
],
[
0.1732177734375,
0.8017578125,
0.82421875,
0.8017578125,
0.82421875,
0.81494140625,
0.1732177734375,
0.81494140625
],
[
0.1748046875,
0.81640625,
0.826171875,
0.81640625,
0.826171875,
0.830078125,
0.1748046875,
0.830078125
],
[
0.1748046875,
0.830078125,
0.826171875,
0.830078125,
0.826171875,
0.84326171875,
0.1748046875,
0.84326171875
],
[
0.1732177734375,
0.84521484375,
0.82275390625,
0.84521484375,
0.82275390625,
0.85595703125,
0.1732177734375,
0.85595703125
],
[
0.1748046875,
0.85791015625,
0.82421875,
0.85791015625,
0.82421875,
0.8720703125,
0.1748046875,
0.8720703125
],
[
0.1748046875,
0.8720703125,
0.826171875,
0.8720703125,
0.826171875,
0.88525390625,
0.1748046875,
0.88525390625
],
[
0.1748046875,
0.88525390625,
0.82421875,
0.88525390625,
0.82421875,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1732177734375,
0.8984375,
0.31787109375,
0.8984375,
0.31787109375,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"Execution",
"choose",
"Method",
"Accuracy",
"Accuracy",
"LOGLINEAR",
"8.6%",
"78.2%",
"12.5%",
"78.7%",
"MAXPOOL PNP",
"2-LAYER PNP",
"14.9%",
"82.5%",
"Table 1: Results from appling Probabilistic Neural Programs to the execution of sketches from the",
"FOODWEBSdataset.",
"\"choose Accuracy\"indicates average accuracy at each decision, whereas",
"\"Execution Accuracy\"indicates if the entire program was executed correctly.",
"functions, e.g., for reasoning about population changes, that call organism and eat to extract",
"information from the diagram. [8] has a more thorough description of the theory; our goal is to learn",
"tomakethechoices inthis theory",
"We consider three models for learning to make the choices for both organi sm and eat: a non-neural",
"(LOGLINEAR) model, as well as two probabilistic neural models (2-LAYER PNP and MAXPOOL",
"PNP). All three learn models for both organism and eat using outputs from a computer vision",
"system trained to detect organism, text, and arrow relations between them. [8] defines a set of",
"hand-engineeredfeatures heuristically created from the outputs of this vision system.LoGLINEAR",
"and 2-LAYER PNP use only these features, and the difference is simply in the greater expressivity of",
"a two-layer neural network. However, one of the major strengths of neural models is their ability to",
"learn latent feature representations automatically, and our third model also uses the direct outputs",
"of the vision system not made into features..The architecture of MAXPOOL PNP reflects this",
"and contains additional input layers that maxpool over detected relationships between objects and",
"confidence scores. The expectation is that our neural network modeling of nondeterminism will learn",
"better latent representations than the manually defined features.",
"4",
"Experiments",
"We evaluate probabilistic neural programs on the FOODWEBs dataset introduced by [8]. This data set",
"contains a training set of ~2,900 programs and a test set of ~1,000 programs. These programs are",
"human annotated gold standard interpretations for the questions in the data set, which corresponds to",
"assuming that the translation from questions to programs is perfect. We train our probabilistic neural",
"programs using correct execution traces of each program, which are also provided in the data set.",
"We evaluate our models using two metrics. First, execution accuracy measures the fraction of",
"programs in the test set that are executed completely correctly by the model. This metric is challenging",
"because correctly executing a program requires correctly making a number of choose decisions.",
"Our 1,000 test programs had over 35,000 decisions, implying that to completely execute a program",
"correctly means getting on average 35 choose decisions correct without making any mistakes.",
"Second, choose accuracy measures the accuracy of each decision independently, assuming all",
"previous decisions were made correctly.",
"Table 1 compares the accuracies of our three models on the FooDwEBs dataset. The improvement",
"in accuracy between the baseline (LOGLINEAR) and the probabilistic neural program (2-LAYER",
"PNP) is due to the neural network's enhanced modeling power. Though the choose accuracy",
"does not improve by a large margin, the improvements translate into large gains in entire program",
"correctness. Finally, as expected, the inclusion of lower level features (MAxPoOL PNP) not possible",
"in LOGLINEAR significantly improved performance. Note that this task requires performing computer",
"vision, and thus it is not expected that any model achieve 100% accuracy.",
"5",
"Conclusion",
"We have presented probabilistic neural programs, a framework for program induction that permits",
"flexible specification of computational models and inference algorithms while simultaneously en-",
"abling the use of deep learning.A program sketch describes a collection of nondeterministic decisions",
"tobemadeduringexecution,alongwiththeneural architecturetobeusedforscoringthesedecisions",
"The network parameters of a sketch can be trained from data using stochastic gradient descent. We",
"demonstrate that probabilistic neural programs improve accuracy on a diagram question answer-",
"ing task which can be formulated as learning to execute program sketches in a domain-specific",
"computational model.",
"4"
] |
{
"arxiv_id": "1612.00712",
"language": "en",
"page_idx": 3
}
|
||
1612.00712-en-4
|
1612.00712-en-4.jpg
|
$\mathbf{Acknowledgements}$
The authors would like to thank the reviewers for their comments as well as helpful discussions with
Arturo Argueta and Oyvind Tafjord.
# References
[1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Gregory S.
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian J. Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Józefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Gordon Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul A. Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda B. Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. Tensorflow: Large-scale machine learning on heterogeneous distributed systems $arXiv\textit{ preprint arXiv: I603. 04467, 2016. }$
[2] Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav
Petrov, and Michael Collins. Globally normalized transition-based neural networks. $CoRR$, abs/1603.06042,
2016.
[3] James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Des-
jardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: A CPU and GPU math compiler
in Python.
[4] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A.Smith. Recurrent neural network
grammars. In Proceedings of the 2016 Conference of the North American Chapter of the Association for 2016. Association for Computational Linguistics.
[5] Alexander L. Gaunt, Marc Brockschmidt, Rishabh Singh, Nate Kushman, Pushmeet Kohli, Jonathan Taylor,
and Daniel Tarlow. Terpret: A probabilistic programming language for program induction. $arXi\nu preprint$
$arXiv{:}I608.04428,2016.$
[6] Noah D. Goodman, Vikash K. Mansinghka, Daniel M. Roy, Keith Bonawitz, and Joshua B. Tenenbaum.
Church: a language for generative models. In $Proc.ofUncertaintyinArtificial$ Intelligence, 2008.
[7] Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. $arXi\nu preprint~arXiv:I410.5401$,
2014.
[8] Jayant Krishnamurthy, Oyvind Tafjord, and Aniruddha Kembhavi. Semantic parsing to probabilistic
programs for situated question answering. $EMNLP,2016.$
[9] Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs with
gradient descent. $CoRR$, abs/1511.04834, 2015.
[10] Joakim Nivre, Johan Hall, Jens Nilsson, Atanas Chanev, Gülsen Eryigit, Sandra Kübler, Svetoslav Marinov,
and Erwin Marsi. Maltparser: A language-independent system for data-driven dependency parsing. Natural
Language Engineering, 13:95-135, 2007.
[11] Scott E. Reed and Nando de Freitas. Neural programmer-interpreters. $CoRR$, abs/1511.06279, 2015.
[12] Sebastian Riedel, Matko Bošnjak, and Tim Rocktäschel. Programming with a differentiable forth interpreter.
arXiv preprint arXiv:1605.06640,2016.
[13] Armando Solar-Lezama, Liviu Tancau, Rastislav Bodik, Sanjit Seshia, and Vijay Saraswat. Combinatorial
sketching for finite programs. In Proceedings of the I2th International Conference on Architectural Support $for\textit{ Programming Languages and Operating Systems, ASPLOS XII, pages 404- 415, New York, NY, USA}$, 2006. ACM.
5
|
[
[
0.176513671875,
0.0921630859375,
0.338134765625,
0.0921630859375,
0.338134765625,
0.106689453125,
0.176513671875,
0.106689453125
],
[
0.1748046875,
0.12249755859375,
0.82275390625,
0.12249755859375,
0.82275390625,
0.1363525390625,
0.1748046875,
0.1363525390625
],
[
0.176513671875,
0.1383056640625,
0.41015625,
0.1383056640625,
0.41015625,
0.1483154296875,
0.176513671875,
0.1483154296875
],
[
0.176513671875,
0.1787109375,
0.265625,
0.1787109375,
0.265625,
0.1900634765625,
0.176513671875,
0.1900634765625
],
[
0.1829833984375,
0.2020263671875,
0.826171875,
0.2020263671875,
0.826171875,
0.21533203125,
0.1829833984375,
0.21533203125
],
[
0.20751953125,
0.2139892578125,
0.82275390625,
0.2139892578125,
0.82275390625,
0.2279052734375,
0.20751953125,
0.2279052734375
],
[
0.2091064453125,
0.2291259765625,
0.82275390625,
0.2291259765625,
0.82275390625,
0.2392578125,
0.2091064453125,
0.2392578125
],
[
0.2059326171875,
0.2392578125,
0.82421875,
0.240478515625,
0.82421875,
0.25390625,
0.2059326171875,
0.25244140625
],
[
0.2059326171875,
0.25244140625,
0.82421875,
0.25244140625,
0.82421875,
0.266357421875,
0.2059326171875,
0.266357421875
],
[
0.2059326171875,
0.264404296875,
0.82275390625,
0.264404296875,
0.82275390625,
0.27783203125,
0.2059326171875,
0.27783203125
],
[
0.2091064453125,
0.279052734375,
0.82421875,
0.279052734375,
0.82421875,
0.2890625,
0.2091064453125,
0.2890625
],
[
0.2059326171875,
0.28857421875,
0.44287109375,
0.28857421875,
0.44287109375,
0.3017578125,
0.2059326171875,
0.3017578125
],
[
0.1829833984375,
0.306884765625,
0.82275390625,
0.306884765625,
0.82275390625,
0.32080078125,
0.1829833984375,
0.32080078125
],
[
0.2091064453125,
0.3212890625,
0.82421875,
0.3212890625,
0.82421875,
0.33203125,
0.2091064453125,
0.33203125
],
[
0.2091064453125,
0.333251953125,
0.2418212890625,
0.333251953125,
0.2418212890625,
0.34521484375,
0.2091064453125,
0.34521484375
],
[
0.1846923828125,
0.3515625,
0.82421875,
0.3515625,
0.82421875,
0.3623046875,
0.1846923828125,
0.3623046875
],
[
0.2059326171875,
0.363037109375,
0.82421875,
0.363037109375,
0.82421875,
0.376953125,
0.2059326171875,
0.376953125
],
[
0.20751953125,
0.375732421875,
0.268798828125,
0.37744140625,
0.268798828125,
0.389404296875,
0.20751953125,
0.386962890625
],
[
0.181396484375,
0.391357421875,
0.82421875,
0.392578125,
0.82421875,
0.4072265625,
0.181396484375,
0.406005859375
],
[
0.2091064453125,
0.406005859375,
0.82275390625,
0.406005859375,
0.82275390625,
0.419189453125,
0.2091064453125,
0.419189453125
],
[
0.2059326171875,
0.41845703125,
0.82275390625,
0.41845703125,
0.82275390625,
0.431884765625,
0.2059326171875,
0.431884765625
],
[
0.2059326171875,
0.429931640625,
0.496826171875,
0.43115234375,
0.496826171875,
0.4443359375,
0.2059326171875,
0.443115234375
],
[
0.181396484375,
0.447509765625,
0.826171875,
0.447509765625,
0.826171875,
0.4609375,
0.181396484375,
0.4609375
],
[
0.20751953125,
0.4609375,
0.82421875,
0.4609375,
0.82421875,
0.474853515625,
0.20751953125,
0.474853515625
],
[
0.20751953125,
0.473388671875,
0.35693359375,
0.473388671875,
0.35693359375,
0.4873046875,
0.20751953125,
0.4873046875
],
[
0.181396484375,
0.4892578125,
0.826171875,
0.490478515625,
0.826171875,
0.5048828125,
0.181396484375,
0.50390625
],
[
0.20751953125,
0.50390625,
0.78173828125,
0.50390625,
0.78173828125,
0.517578125,
0.20751953125,
0.517578125
],
[
0.1829833984375,
0.521484375,
0.826171875,
0.521484375,
0.826171875,
0.53515625,
0.1829833984375,
0.53515625
],
[
0.2091064453125,
0.53515625,
0.2418212890625,
0.53515625,
0.2418212890625,
0.546875,
0.2091064453125,
0.546875
],
[
0.181396484375,
0.5517578125,
0.82421875,
0.5517578125,
0.82421875,
0.56494140625,
0.181396484375,
0.56494140625
],
[
0.2059326171875,
0.56298828125,
0.54833984375,
0.56298828125,
0.54833984375,
0.5771484375,
0.2059326171875,
0.5771484375
],
[
0.1829833984375,
0.58203125,
0.82421875,
0.58203125,
0.82421875,
0.59521484375,
0.1829833984375,
0.59521484375
],
[
0.2091064453125,
0.5966796875,
0.488525390625,
0.5966796875,
0.488525390625,
0.60693359375,
0.2091064453125,
0.60693359375
],
[
0.1732177734375,
0.60986328125,
0.82421875,
0.611328125,
0.82421875,
0.62548828125,
0.1732177734375,
0.625
],
[
0.20751953125,
0.625,
0.82421875,
0.625,
0.82421875,
0.63818359375,
0.20751953125,
0.63818359375
],
[
0.20751953125,
0.63720703125,
0.453369140625,
0.63720703125,
0.453369140625,
0.65087890625,
0.20751953125,
0.65087890625
],
[
0.1748046875,
0.65478515625,
0.79833984375,
0.65478515625,
0.79833984375,
0.66845703125,
0.1748046875,
0.66845703125
],
[
0.1732177734375,
0.67041015625,
0.826171875,
0.671875,
0.826171875,
0.68603515625,
0.1732177734375,
0.68505859375
],
[
0.2091064453125,
0.68505859375,
0.443603515625,
0.68505859375,
0.443603515625,
0.69873046875,
0.2091064453125,
0.69873046875
],
[
0.1748046875,
0.70263671875,
0.82421875,
0.70263671875,
0.82421875,
0.71630859375,
0.1748046875,
0.71630859375
],
[
0.2091064453125,
0.71533203125,
0.82421875,
0.71533203125,
0.82421875,
0.72900390625,
0.2091064453125,
0.72900390625
],
[
0.2059326171875,
0.72802734375,
0.826171875,
0.72802734375,
0.826171875,
0.7412109375,
0.2059326171875,
0.7412109375
],
[
0.2059326171875,
0.74072265625,
0.281005859375,
0.7412109375,
0.281005859375,
0.75244140625,
0.2059326171875,
0.751953125
],
[
0.494384765625,
0.9375,
0.5048828125,
0.9375,
0.5048828125,
0.94970703125,
0.494384765625,
0.94970703125
]
] |
[
"Acknowledgements",
"The authors would like to thank the reviewers for their comments as well as helpful discussions with",
"ArturoArguetaandOyvindTafjord",
"References",
"[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Gregory S.",
"Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian J. Goodfellow, Andrew Harp",
"GeoffreyIrving,MichaelIsard,YangqingJia,RafalJozefowicz,LukaszKaiser,ManjunathKudlur,Josh",
"Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Gordon Murray, Chris Olah, Mike Schuster.",
"Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul A. Tucker, Vincent Vanhoucke, Vijay",
"Vasudevan, Fernanda B. Viegas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu",
"andXiaoqiangZheng.Tensorflow:Large-scalemachinelearningonheterogeneous distributedsystems",
"arXiv preprint arXiv:1603.04467,2016.",
"[2] Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav",
"Petrov,andMichaelCollins.Globallynormalizedtransition-based neuralnetworks.CoRR,abs/1603.06042,",
"2016.",
"[3]JamesBergstra,OlivierBreuleux, FredéricBastien,PascalLamblin,RazvanPascanu, GuillaumeDes-",
"jardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: A CPU and GPU math compiler",
"in Python.",
"[4] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network",
"grammars. In Proceedings of the 2016 Conference of the North American Chapter of the Association for",
"Computational Linguistics: Human Language Technologies, pages 199-209, San Diego, California, June",
"2016. Association for Computational Linguistics",
"[5] Alexander L. Gaunt, Marc Brockschmidt, Rishabh Singh, Nate Kushman, Pushmeet Kohli, Jonathan Taylor,",
"and Daniel Tarlow. Terpret: A probabilistic programming language for program induction. arXiv preprint",
"arXiv:1608.04428,2016.",
"[6] Noah D. Goodman, Vikash K. Mansinghka, Daniel M. Roy, Keith Bonawitz, and Joshua B. Tenenbaum.",
"Church: a language for generative models. In Proc. of Uncertainty in Artificial Intelligence, 2008",
"[7] Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401,",
"2014.",
"[8] Jayant Krishnamurthy, Oyvind Tafjord, and Aniruddha Kembhavi. Semantic parsing to probabilistic",
"programs for situated question answering. EMNLP, 2016.",
"[9] Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs with",
"gradientdescent.CoRR,abs/1511.04834,2015.",
"[10] Joakim Nivre, Johan Hall, Jens Nilsson, Atanas Chanev, Gulsen Eryigit, Sandra Kibler, Svetoslav Marinov.",
"and Erwin Marsi. Maltparser: A language-independent system for data-driven dependency parsing. Natural",
"Language Engineering, 13:95-135, 2007.",
"[11] Scott E.Reed and Nando de Freitas.Neural programmer-interpreters.CoRR, abs/1511.06279,2015.",
"[12] Sebastian Riedel, Matko Bosnjak, and Tim Rocktaschel. Programming with a differentiable forth interpreter.",
"arXiv preprint arXiv:1605.06640,2016.",
"[13] Armando Solar-Lezama, Liviu Tancau, Rastislav Bodik, Sanjit Seshia, and Vijay Saraswat. Combinatorial",
"sketching for finite programs.In Proceedings of the 12th International Conference on Architectural Support",
"for Programming Languages and Operating Systems, ASPLOS XII, pages 404 415, New York, NY, USA,",
"2006.ACM.",
"5"
] |
{
"arxiv_id": "1612.00712",
"language": "en",
"page_idx": 4
}
|
||
2404.13798-en-1
|
2404.13798-en-1.jpg
|
between the sensitive attributes and the target variable, ensuring that fairness is based on genuine causes of disparities rather than arbitrary associations [14,24].
However, most existing work in fair machine learning typically focuses on enforcing independence on lowdimensional features in cases where the conditioning set is either empty or merely a low-dimensional categorical variable [2,3,24,25,37]. This restricts the learning of fair representations to limited settings involving only binary attributes in tabular datasets [35]. These methods cannot directly translate to high-dimensional spaces, such as the latent representations of large generative models.
Recently, Ahuja et al. developed a differentiable framework for enforcing CI in high-dimensional and continuous feature spaces by using a GAN-based approach in the context of data generation. To enforce $s\perp\hat{y}\mid y$,they minimize the Jensen-Shannon (JS) divergence between the joint distribution $p(s,\hat{y},y)$ and an auxiliary distribution $q(s,y^\prime,y)$ while the joint marginal $q(s,y)$ is distributed identically to $p(s,y)$ and $q(y^{\prime}\mid y)$ is independent of $s.$ Although this formulation enforces CI, it can only sample the joint probability distributions in the label space $y$ and does not necessarily guarantee the learning of fair representations, $i.e.$, only the last network layer before predicting $y$ removes the dependence while the rest of the network (including latent representations) remain biased or confounded.
Our approach introduces the concept of CI into the latent space by using a dynamic sampler to form the joint distributions with respect to the learned representations. This yields a computationally rigorous approach to fair and causal representation learning that exhibits a level of versatility not found in existing methods: We are able to control not just which features are encoded, but also where within the representation they appear. With this, we gain improved classification performance and, in the case of generative models such as the diffusion autoencoder (DiffAE), the ability to maintain fine-grained control over generated content. As a result, we can not only protect downstream tasks from bias against specific attributes, but also generate featureinvariant images corresponding to the enforced schema, allowing for a richer understanding of a model's representations.
## $2. \textbf{Related Work}$
Invariant Representation Learning In the wake of increasingly biased large-scale models [1, 10, 18, 31], the learning of fair and unconfounded representations has taken on outsized importance. Methods based on domain adversarial learning are increasingly popular tools of choice in reducing bias and confounding effects [1,15,33,34,37]. Among these, Johndrow et al. and Tan et al. proposed methods for invariant feature learning and Zhao et al. presented models that minimize statistical mean dependence
<FigureHere>
Figure 2. High-level view of our architecture. We introduce two variants of a conditional independence enforcer that can be added to any off-the-shelf encoder.
using a correlation-based adversarial loss function. Recently, Pirhadi et al. introduced a data cleaning method that enforces the conditional independence constraint for tabular data using optimal transport.
Other types of methods incorporating statistical operations have also been explored in the literature, such as using multivariate regression analysis [19] with general linear models [36]. Other approaches enforce fairness through post-processing steps on unfair trained classification models [6, 10, 36]. But since the training and fairness enforcement steps are conducted separately, these algorithms often lead to suboptimal fairness and accuracy trade-offs [32]. Recently, using traditional statistical methods, Lu et al. proposed a normalization layer that corrects the feature distributions with respect to labeled sensitive attributes. Their approach, known as metadata normalization, entailed a simple layer that could plug into any end-to-end model to protect the models against bias. Vento et al. then extended this work by turning the closed-form normalization operation into a network optimizable step. This is an active research field and many other approaches are being studied.
Conditional Independence for Algorithmic Fairness Consider a classifier or regression function $f(x)$ with output $\hat{y}$ and a protected attribute s, e.g., sex or race. Algorithmic fairness aims to learn prediction functions or transformations that make the final outcome invariant or insensitive to protected attributes. Several widely used associational and causal notions of fairness, including the equalized odds criterion, can be attained by means of independence[25].
Rather than attempting to transform already trained unfair models to fair ones in a post-hoc manner, it is critical to enforce fairness constraints from the beginning of the model
2
|
[
[
0.080078125,
0.091552734375,
0.46728515625,
0.091552734375,
0.46728515625,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.79345703125,
0.10797119140625,
0.82177734375,
0.10797119140625,
0.82177734375,
0.1256103515625,
0.79345703125,
0.1256103515625
],
[
0.080078125,
0.12249755859375,
0.358642578125,
0.12249755859375,
0.358642578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.66650390625,
0.121826171875,
0.69287109375,
0.1275634765625,
0.68701171875,
0.141357421875,
0.6611328125,
0.1351318359375
],
[
0.096435546875,
0.136962890625,
0.468017578125,
0.1363525390625,
0.468017578125,
0.15087890625,
0.096435546875,
0.1514892578125
],
[
0.67333984375,
0.1458740234375,
0.67724609375,
0.14453125,
0.6806640625,
0.15087890625,
0.67626953125,
0.1514892578125
],
[
0.77392578125,
0.147705078125,
0.7939453125,
0.147705078125,
0.7939453125,
0.1578369140625,
0.77392578125,
0.1578369140625
],
[
0.8505859375,
0.147705078125,
0.86279296875,
0.149658203125,
0.861328125,
0.1546630859375,
0.8486328125,
0.15283203125
],
[
0.080078125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.66845703125,
0.17041015625,
0.68896484375,
0.17041015625,
0.68896484375,
0.18310546875,
0.66845703125,
0.18310546875
],
[
0.76318359375,
0.1767578125,
0.779296875,
0.1767578125,
0.779296875,
0.184326171875,
0.76318359375,
0.184326171875
],
[
0.83251953125,
0.17236328125,
0.85546875,
0.17236328125,
0.85546875,
0.1849365234375,
0.83251953125,
0.1849365234375
],
[
0.08087158203125,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1956787109375,
0.08087158203125,
0.1956787109375
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.57861328125,
0.1976318359375,
0.5986328125,
0.1976318359375,
0.5986328125,
0.2052001953125,
0.57861328125,
0.2052001953125
],
[
0.61279296875,
0.1988525390625,
0.625,
0.1988525390625,
0.625,
0.2020263671875,
0.61279296875,
0.2020263671875
],
[
0.5947265625,
0.2052001953125,
0.61279296875,
0.2052001953125,
0.61279296875,
0.212158203125,
0.5947265625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.77392578125,
0.2474365234375,
0.84228515625,
0.2474365234375,
0.84228515625,
0.270751953125,
0.77392578125,
0.270751953125
],
[
0.07843017578125,
0.25830078125,
0.3896484375,
0.2568359375,
0.3896484375,
0.271484375,
0.07843017578125,
0.272705078125
],
[
0.60693359375,
0.2568359375,
0.62353515625,
0.2568359375,
0.62353515625,
0.260009765625,
0.60693359375,
0.260009765625
],
[
0.098876953125,
0.2734375,
0.46728515625,
0.2734375,
0.46728515625,
0.287353515625,
0.098876953125,
0.287353515625
],
[
0.56640625,
0.26953125,
0.60546875,
0.26953125,
0.60546875,
0.27978515625,
0.56640625,
0.27978515625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.56201171875,
0.287353515625,
0.60546875,
0.287353515625,
0.60546875,
0.29736328125,
0.56201171875,
0.29736328125
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50244140625,
0.318115234375,
0.890625,
0.318115234375,
0.890625,
0.33203125,
0.50244140625,
0.33203125
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.333251953125,
0.89208984375,
0.333251953125,
0.89208984375,
0.3466796875,
0.50341796875,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50244140625,
0.3466796875,
0.67236328125,
0.3466796875,
0.67236328125,
0.35986328125,
0.50244140625,
0.35986328125
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.363037109375,
0.4697265625,
0.37744140625,
0.080078125,
0.37890625
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37744140625,
0.468017578125,
0.392578125,
0.080078125,
0.393310546875
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50341796875,
0.39013671875,
0.890625,
0.39013671875,
0.890625,
0.404052734375,
0.50341796875,
0.404052734375
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50341796875,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.41845703125,
0.50341796875,
0.41845703125
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50341796875,
0.42236328125,
0.890625,
0.42236328125,
0.890625,
0.43310546875,
0.50341796875,
0.43310546875
],
[
0.080078125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.436279296875,
0.69287109375,
0.436279296875,
0.69287109375,
0.449462890625,
0.50244140625,
0.449462890625
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.52294921875,
0.453857421875,
0.8896484375,
0.453857421875,
0.8896484375,
0.464111328125,
0.52294921875,
0.464111328125
],
[
0.07843017578125,
0.47021484375,
0.384033203125,
0.4697265625,
0.384033203125,
0.48291015625,
0.07843017578125,
0.484130859375
],
[
0.50244140625,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.098876953125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.098876953125,
0.498779296875
],
[
0.50244140625,
0.48291015625,
0.890625,
0.48291015625,
0.890625,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.5009765625,
0.496826171875,
0.89208984375,
0.49755859375,
0.89208984375,
0.51220703125,
0.5009765625,
0.51123046875
],
[
0.08087158203125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.50244140625,
0.51318359375,
0.89208984375,
0.51318359375,
0.89208984375,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.080078125,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50244140625,
0.52783203125,
0.890625,
0.52783203125,
0.890625,
0.54150390625,
0.50244140625,
0.54150390625
],
[
0.080078125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.5009765625,
0.54248046875,
0.8896484375,
0.54248046875,
0.8896484375,
0.55615234375,
0.5009765625,
0.55615234375
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.55810546875,
0.89208984375,
0.55810546875,
0.89208984375,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.07843017578125,
0.57373046875,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.07843017578125,
0.5869140625
],
[
0.50244140625,
0.57373046875,
0.89208984375,
0.57373046875,
0.89208984375,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.07843017578125,
0.59033203125,
0.468017578125,
0.5888671875,
0.468017578125,
0.60400390625,
0.07843017578125,
0.60498046875
],
[
0.5009765625,
0.5869140625,
0.890625,
0.5869140625,
0.890625,
0.60107421875,
0.5009765625,
0.60107421875
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.60400390625,
0.890625,
0.60400390625,
0.890625,
0.6171875,
0.50244140625,
0.6171875
],
[
0.08087158203125,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.50341796875,
0.61865234375,
0.89208984375,
0.61865234375,
0.89208984375,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.080078125,
0.63623046875,
0.4697265625,
0.63623046875,
0.4697265625,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50341796875,
0.63525390625,
0.890625,
0.63525390625,
0.890625,
0.64599609375,
0.50341796875,
0.64599609375
],
[
0.080078125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.6484375,
0.89208984375,
0.6484375,
0.89208984375,
0.662109375,
0.50244140625,
0.662109375
],
[
0.07843017578125,
0.66552734375,
0.468017578125,
0.66650390625,
0.468017578125,
0.68115234375,
0.07843017578125,
0.6796875
],
[
0.50244140625,
0.6640625,
0.89208984375,
0.6640625,
0.89208984375,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.50244140625,
0.6787109375,
0.89208984375,
0.6787109375,
0.89208984375,
0.6923828125,
0.50244140625,
0.6923828125
],
[
0.07843017578125,
0.69580078125,
0.4697265625,
0.6962890625,
0.4697265625,
0.7109375,
0.07843017578125,
0.71044921875
],
[
0.5009765625,
0.693359375,
0.80224609375,
0.693359375,
0.80224609375,
0.70703125,
0.5009765625,
0.70703125
],
[
0.080078125,
0.71337890625,
0.1192626953125,
0.71337890625,
0.1192626953125,
0.724609375,
0.080078125,
0.724609375
],
[
0.5048828125,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.080078125,
0.74072265625,
0.2181396484375,
0.74072265625,
0.2181396484375,
0.7548828125,
0.080078125,
0.7548828125
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.08087158203125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.08087158203125,
0.77880859375
],
[
0.50341796875,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.796875,
0.890625,
0.796875,
0.890625,
0.806640625,
0.50341796875,
0.806640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82470703125,
0.890625,
0.82470703125,
0.890625,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.859375,
0.84033203125,
0.859375,
0.85400390625,
0.5009765625,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.521484375,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.521484375,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"between the sensitive attributes and the target variable, en-",
"suringthatfairness isbasedongenuine causes of disparities",
"Lcr",
"rather than arbitrary associations [14, 24].",
"Lci",
"However, most existing work in fair machine learn-",
"",
"LDI",
"UT",
"ing typically focuses on enforcing independence on low-",
"dimensional features in cases where the conditioning set is",
"CI",
"D",
"D2",
"either emptyormerely a low-dimensional categorical vari-",
"able [2, 3, 24, 25, 37]. This restricts the learning of fair",
"y:pr",
"ata",
"labe",
"representations to limited settings involving only binary at-",
"tributes in tabular datasets [35]. These methods cannot di-",
"rectly translate to high-dimensional spaces, such as the la-",
"813",
"tent representations of large generative models.",
"Y",
"Recently,Ahuja et al.developed a differentiable frame-",
"encoder",
"work for enforcing CI in high-dimensional and continuous",
"x:image",
"feature spaces by using a GAN-based approach in the con-",
"text of data generation. To enforce s y I y, they minimize",
"Figure 2.High-level view of our architecture: We introduce two",
"the Jensen-Shannon (JS) divergence between the joint dis-",
"variants of a conditional independence enforcer that can be added",
"tribution p(s, ,y) and an auxiliary distribution q(s, y', y)",
"to any off-the-shelf encoder.",
"while the joint marginal g(s, y) is distributed identically to",
"p(s,y) and q(y' / y) is independent of s. Although this",
"formulation enforces CI, it can only sample the joint prob-",
"using a correlation-based adversarial loss function. Re-",
"ability distributions in the label space y and does not nec-",
"cently,Pirhadietal.introduced a data cleaning method that",
"essarily guarantee the learning of fair representations, i.e.,",
"enforcestheconditionalindependenceconstraintfortabular",
"only the last network layer before predicting y removes the",
"data using optimal transport.",
"dependence while the rest of the network (including latent",
"Othertypes of methods incorporating statistical opera-",
"representations)remainbiasedorconfounded.",
"tions have also been explored in the literature, such as us-",
"Our approach introduces the concept of CI into the latent",
"ing multivariate regression analysis [19] with general lin-",
"space by using a dynamic sampler to form the joint distribu-",
"ear models [36]. Other approaches enforce fairness through",
"tionswithrespecttothelearnedrepresentations.Thisyields",
"post-processing steps on unfair trained classification mod-",
"a computationally rigorous approach to fair and causal rep-",
"els [6, 10, 36]. But since the training and fairness enforce-",
"resentation learning that exhibits a level of versatility not",
"ment steps are conducted separately, these algorithms of-",
"found in existing methods: We are able to control not just",
"ten lead to suboptimal fairness and accuracy trade-offs [32].",
"which features are encoded, but also where within the rep-",
"Recently, using traditional statistical methods, Lu et al. pro-",
"resentation they appear. With this, we gain improved classi-",
"posed a normalization layer that corrects the feature distri-",
"fication performance and, in the case of generative models",
"butions with respect to labeled sensitive attributes. Their ap-",
"suchasthe diffusionautoencoder(DiffAE),theability to",
"proach, known as metadata normalization, entailed a simple",
"maintain fine-grained control over generated content. As",
"layerthatcouldpluginto any end-to-end modeltoprotect",
"a result, we can not only protect downstream tasks from",
"the models against bias. Vento et al. then extended this work",
"bias against specific attributes, but also generate feature-",
"by turning the closed-form normalization operation into a",
"invariant images corresponding to the enforced schema, al-",
"network optimizable step. This is an active research field",
"lowing for a richer understanding of a model's representa-",
"and many other approaches are being studied.",
"tions.",
"Conditional Independence for Algorithmic Fairness",
"2. Related Work",
"Consider a classifier or regression function f() with output",
"Invariant Representation Learning In the wake of in-",
"y and a protected attribute s, e.g., sex or race. Algorithmic",
"creasingly biased large-scale models [1, 10, 18, 31], the",
"fairness aims to learn prediction functions or transforma-",
"learning of fair and unconfounded representations has taken",
"tions that make the final outcome invariant or insensitiveto",
"on outsized importance. Methods based on domain adver-",
"protected attributes. Several widely used associational and",
"sarial learning are increasingly popular tools of choice in",
"causal notions of fairness, including the equalized odds cri-",
"reducing bias and confounding effects [1, 15, 33, 34, 37].",
"terion, can be attained by means of independence [25]",
"Among these, Johndrow et al. and Tan et al. proposed",
"Rather than attempting to transform already trained un-",
"methods for invariant feature learning and Zhao et al. pre-",
"fair models to fair ones in a post-hoc manner, it is critical to",
"sented models that minimize statistical mean dependence",
"enforce fairness constraints from the beginning of the model",
"2"
] |
{
"arxiv_id": "2404.13798",
"language": "en",
"page_idx": 1
}
|
||
2404.13798-en-2
|
2404.13798-en-2.jpg
|
training pipeline [11]. Apart from preventing the propagation of bias to downstream modeling tasks, this approach of fairness from a data management standpoint also leads to more robust and significant fairness measures by ensuring that data sources, transformations, and other training assumptions are sound[24].
Representation Learning with Diffusion Models Diffusion-based models learn to generate images by a denoising process: Random Gaussian noise is added to input images, and the model learns how to reverse this process to hallucinate new images from random noise. This family of models has proven to offer remarkable quality in image generation, promising to replace generative adversarial networks (GANs)[8] as the dominant architectural paradigm. Chief among them is the diffusion autoencoder [23], which encodes an image into a two-part latent subcode, capturing a stochastic representation via the aforementioned denoising approach and conditioning this process upon a semantic representation learned by a CNN. Interpolation of the resulting subcode results in smooth and meaningful changes in the decoded image. The semantic separation capability of DiffAEs has enabled various applications, such as attribute manipulation and various low-shot or zero-shot downstream applications [7,9,26]. In this work, we further enable DiffAEs to learn unbiased, causal, and conditionally independent latent subcode representations.
## $3. \textbf{Method}$
Let $\{x,y,s\}$ be a training sample in the dataset, where $x$ is an input image, $y$ is the target prediction label, and $s$ is a protected sensitive attribute. We aim to learn an encoder network $g_\theta ( x)$ = $v$ resulting in a latent representation $v$, from which a prediction network $f_\phi(v)=\hat{y}$ produces the final predicted label $\hat{y}.$ With a slight abuse of notation, we use lower-case letters to also denote random variables. Ultimately, we aim to train a model that achieves high accuracy in predicting label $y$ while ensuring equalized odds:
$$\begin{aligned}\operatorname*{argmin}_{\theta,\phi}&-\frac{1}{n}\sum_{i=1}^{n}y_{i}\log(\hat{y}_{i})+(1-y_{i})\log(1-\hat{y}_{i})\\&\mathrm{s.t.}\quad s\perp\hat{y}\mid y.\end{aligned}$$
(1)
We discuss two possible ways of enforcing the equalized odds fairness constraint: First, we enforce conditional independence with respect to the label $y,i.e.,s\perp\hat{y}\mid y.$ As discussed in Ahuja et al., this can be seen as a means to enforce equalized odds through CI. However, we claim that enforcing the independence constraint with respect to only the label is limited in its ability to enforce the equalized odds fairness constraint effectively. We therefore additionally propose to enforce the conditional independence with
respect to the more information-rich latent representation $v$, or $s\perp v\mid y.$ We can then use this learned fair latent representation in a downstream prediction or generation task.
## $3. 1. \textbf{Enforcing CI with respect to Label}$
When the label $y$ is highly correlated with the protected attribute $s$ in the training data, a basic aim for a fair machine learning model is to ensure conditional independence of the predicted $\hat{y}$ from $s,i.e.,s\perp\hat{y}\mid y.$
Assuming that the joint distribution $p(s,\hat{y},y)$ is welldefined with respect to Lebesgue measure $\mu$, previous work [3] demonstrates that the above conditional independence can be defined with respect to the Jensen-Shannon divergence (JS) between $p(s,\hat{y},y)$ and an auxiliary distribution $q(s,y^{\prime},y)$, where the joint marginal $q(s,y)$ is distributed identically to $p(s,y)$, and $q(y^\prime|y)$ is independent of $s.$ In other words, when $q(s,y^{\prime},y)=p(s,y)q(y^{\prime}|y)$, the conditional independence constraint $s\perp\hat{y}\mid y$ is equivalent to enforcing
(2)
$$\mathrm{JS}(p(\hat{y},y),q(y',y))=\mathrm{JS}(p(s,\hat{y},y),q(s,y',y)).$$
Accordingly, we can rewrite (1) in terms of JS as
$$\begin{aligned}\underset{\theta,\phi}{\operatorname*{argmin}}&-\frac{1}{n}\sum_{i=1}^{n}y_{i}\log(\hat{y}_{i})+(1-y_{i})\log(1-\hat{y}_{i})\\\mathrm{s.t.}&\left(\mathrm{JS}(p(\hat{y},y),q(y^{\prime},y))-\right)\\&\mathrm{JS}(p(s,\hat{y},y),q(s,y^{\prime},y))\leq\delta\Big)\end{aligned}$$
(3)
for some $\delta>0$ sufficiently small.
With the help of a conditional sampler $q(y^{\prime}\mid y)$, the JS constraint of (3) can be satisfied by a general GAN architecture with two discriminators (Fig. 2) [3], i.e., by minimizing
$$L_{\mathrm{CI}}=(L_{\mathrm{D1}}-L_{\mathrm{D2}})^2$$
(4)
where
$$L_{\mathrm{D1}}=\mathbb{E}[\log(1-D_{1}(y^{\prime},y))+\mathbb{E}_{y,s}[\log D_{1}(\hat{y},y)]\quad\mathrm{and}\\L_{\mathrm{D2}}=\mathbb{E}[\log(1-D_{2}(y^{\prime},s,y)]+\mathbb{E}_{y,s,y^{\prime}}[\log D_{2}(\hat{y},s,y)].$$
The key advantage of using (4) to derive conditional independence is that $q(y^{\prime}|y)$ does not have to be a perfect sampler that exactly matches $p(\hat{y}\mid y).$ The only sufficient condition is that $q(y^{\prime}|y)$ shares overlapping support with the original distribution $p(\hat{y}\mid y).$ This gives us many flexible ways to construct $q$, such as using a uniform distribution.
Specifically, in a supervised learning setting, where $y$ is known at training time, we can set $q$ using a uniform distribution over all raw model outputs that correspond to $y.$ The encoder and prediction networks can then be learned by optimizing the loss function
(5)
$$L_{\mathrm{Enc}}=\lambda L_{\mathrm{CI}}+L_{\mathrm{task}},$$
3
|
[
[
0.07843017578125,
0.091552734375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.07843017578125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12371826171875,
0.869140625,
0.12371826171875,
0.869140625,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.146484375,
0.80224609375,
0.146484375,
0.80224609375,
0.1610107421875,
0.50341796875,
0.1610107421875
],
[
0.080078125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.07843017578125,
0.1685791015625,
0.250732421875,
0.1685791015625,
0.250732421875,
0.1817626953125,
0.07843017578125,
0.1817626953125
],
[
0.50341796875,
0.169189453125,
0.890625,
0.169189453125,
0.890625,
0.18310546875,
0.50341796875,
0.18310546875
],
[
0.50244140625,
0.1849365234375,
0.89208984375,
0.1849365234375,
0.89208984375,
0.1988525390625,
0.50244140625,
0.1988525390625
],
[
0.080078125,
0.2008056640625,
0.1968994140625,
0.2008056640625,
0.1968994140625,
0.2139892578125,
0.080078125,
0.2139892578125
],
[
0.1895751953125,
0.2008056640625,
0.4697265625,
0.2008056640625,
0.4697265625,
0.2139892578125,
0.1895751953125,
0.2139892578125
],
[
0.50244140625,
0.2008056640625,
0.89208984375,
0.2008056640625,
0.89208984375,
0.2139892578125,
0.50244140625,
0.2139892578125
],
[
0.080078125,
0.216552734375,
0.471435546875,
0.216552734375,
0.471435546875,
0.2298583984375,
0.080078125,
0.2298583984375
],
[
0.50244140625,
0.216552734375,
0.72607421875,
0.216552734375,
0.72607421875,
0.2298583984375,
0.50244140625,
0.2298583984375
],
[
0.080078125,
0.2310791015625,
0.4697265625,
0.2310791015625,
0.4697265625,
0.2442626953125,
0.080078125,
0.2442626953125
],
[
0.521484375,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.2442626953125,
0.521484375,
0.2442626953125
],
[
0.080078125,
0.24560546875,
0.4697265625,
0.24560546875,
0.4697265625,
0.2587890625,
0.080078125,
0.2587890625
],
[
0.50244140625,
0.24560546875,
0.8935546875,
0.24560546875,
0.8935546875,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.080078125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.274658203125,
0.080078125,
0.274658203125
],
[
0.50244140625,
0.261474609375,
0.890625,
0.261474609375,
0.890625,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.07843017578125,
0.274658203125,
0.4697265625,
0.27587890625,
0.4697265625,
0.290283203125,
0.07843017578125,
0.2890625
],
[
0.50244140625,
0.27587890625,
0.89208984375,
0.27587890625,
0.89208984375,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.080078125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.304931640625,
0.080078125,
0.304931640625
],
[
0.50341796875,
0.291748046875,
0.89208984375,
0.291748046875,
0.89208984375,
0.30615234375,
0.50341796875,
0.30615234375
],
[
0.0767822265625,
0.30615234375,
0.4697265625,
0.304931640625,
0.4697265625,
0.3193359375,
0.0767822265625,
0.32080078125
],
[
0.50244140625,
0.30615234375,
0.89208984375,
0.30615234375,
0.89208984375,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.080078125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.335205078125,
0.080078125,
0.335205078125
],
[
0.5009765625,
0.3193359375,
0.89208984375,
0.32080078125,
0.89208984375,
0.335205078125,
0.5009765625,
0.333984375
],
[
0.08087158203125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50244140625,
0.3359375,
0.89208984375,
0.3359375,
0.89208984375,
0.35107421875,
0.50244140625,
0.35107421875
],
[
0.080078125,
0.3515625,
0.4697265625,
0.3515625,
0.4697265625,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.5009765625,
0.349853515625,
0.89208984375,
0.35107421875,
0.89208984375,
0.365478515625,
0.5009765625,
0.3642578125
],
[
0.080078125,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.5009765625,
0.3642578125,
0.5849609375,
0.3662109375,
0.5849609375,
0.380615234375,
0.5009765625,
0.37890625
],
[
0.08087158203125,
0.3837890625,
0.46728515625,
0.3837890625,
0.46728515625,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.52783203125,
0.393310546875,
0.85888671875,
0.393310546875,
0.85888671875,
0.40771484375,
0.52783203125,
0.40771484375
],
[
0.86669921875,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.406005859375,
0.86669921875,
0.406005859375
],
[
0.08087158203125,
0.412353515625,
0.46728515625,
0.412353515625,
0.46728515625,
0.425537109375,
0.08087158203125,
0.425537109375
],
[
0.080078125,
0.396484375,
0.4697265625,
0.396484375,
0.4697265625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.5048828125,
0.42041015625,
0.82421875,
0.42041015625,
0.82421875,
0.43505859375,
0.5048828125,
0.43505859375
],
[
0.08087158203125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.080078125,
0.442626953125,
0.468017578125,
0.442626953125,
0.468017578125,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.59716796875,
0.449462890625,
0.63134765625,
0.451904296875,
0.6298828125,
0.464111328125,
0.59619140625,
0.462158203125
],
[
0.080078125,
0.458251953125,
0.4697265625,
0.458251953125,
0.4697265625,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.51708984375,
0.458251953125,
0.57275390625,
0.455810546875,
0.57421875,
0.47021484375,
0.5185546875,
0.472900390625
],
[
0.6259765625,
0.458251953125,
0.85546875,
0.458251953125,
0.85546875,
0.472900390625,
0.6259765625,
0.472900390625
],
[
0.53173828125,
0.4716796875,
0.55615234375,
0.4716796875,
0.55615234375,
0.48291015625,
0.53173828125,
0.48291015625
],
[
0.5986328125,
0.47021484375,
0.60888671875,
0.46728515625,
0.61279296875,
0.474853515625,
0.60302734375,
0.477783203125
],
[
0.869140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.49365234375,
0.869140625,
0.49365234375
],
[
0.080078125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.07843017578125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.07843017578125,
0.501953125
],
[
0.54150390625,
0.4912109375,
0.576171875,
0.4912109375,
0.576171875,
0.50244140625,
0.54150390625,
0.50244140625
],
[
0.57177734375,
0.4873046875,
0.7255859375,
0.488037109375,
0.7255859375,
0.505859375,
0.57177734375,
0.5048828125
],
[
0.080078125,
0.50244140625,
0.24267578125,
0.50244140625,
0.24267578125,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.576171875,
0.51025390625,
0.77294921875,
0.51025390625,
0.77294921875,
0.5244140625,
0.576171875,
0.5244140625
],
[
0.080078125,
0.52978515625,
0.1666259765625,
0.52978515625,
0.1666259765625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.5361328125,
0.7236328125,
0.5361328125,
0.7236328125,
0.5498046875,
0.50341796875,
0.5498046875
],
[
0.080078125,
0.5546875,
0.4697265625,
0.5546875,
0.4697265625,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.52294921875,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.52294921875,
0.56494140625
],
[
0.080078125,
0.57080078125,
0.4697265625,
0.57080078125,
0.4697265625,
0.583984375,
0.080078125,
0.583984375
],
[
0.5009765625,
0.56640625,
0.8896484375,
0.56494140625,
0.8896484375,
0.580078125,
0.5009765625,
0.58056640625
],
[
0.07843017578125,
0.58544921875,
0.468017578125,
0.583984375,
0.468017578125,
0.5986328125,
0.07843017578125,
0.599609375
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.62744140625,
0.60986328125,
0.765625,
0.60986328125,
0.765625,
0.62744140625,
0.62744140625,
0.62744140625
],
[
0.080078125,
0.6162109375,
0.468017578125,
0.6162109375,
0.468017578125,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.86767578125,
0.611328125,
0.89453125,
0.611328125,
0.89453125,
0.62744140625,
0.86767578125,
0.62744140625
],
[
0.080078125,
0.630859375,
0.4697265625,
0.630859375,
0.4697265625,
0.64453125,
0.080078125,
0.64453125
],
[
0.50244140625,
0.6337890625,
0.54833984375,
0.6337890625,
0.54833984375,
0.6484375,
0.50244140625,
0.6484375
],
[
0.08087158203125,
0.64794921875,
0.46728515625,
0.64794921875,
0.46728515625,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.080078125,
0.66015625,
0.468017578125,
0.6611328125,
0.468017578125,
0.67529296875,
0.080078125,
0.6748046875
],
[
0.5087890625,
0.6611328125,
0.87744140625,
0.6611328125,
0.87744140625,
0.67529296875,
0.5087890625,
0.67529296875
],
[
0.080078125,
0.6767578125,
0.420654296875,
0.6767578125,
0.420654296875,
0.68994140625,
0.080078125,
0.68994140625
],
[
0.5087890625,
0.68115234375,
0.88232421875,
0.68115234375,
0.88232421875,
0.69580078125,
0.5087890625,
0.69580078125
],
[
0.1707763671875,
0.7021484375,
0.2099609375,
0.69580078125,
0.214111328125,
0.7109375,
0.1748046875,
0.71728515625
],
[
0.1666259765625,
0.70849609375,
0.435546875,
0.69873046875,
0.43701171875,
0.72802734375,
0.1683349609375,
0.7373046875
],
[
0.52294921875,
0.70703125,
0.890625,
0.70703125,
0.890625,
0.72021484375,
0.52294921875,
0.72021484375
],
[
0.44677734375,
0.7216796875,
0.468017578125,
0.7216796875,
0.468017578125,
0.734375,
0.44677734375,
0.734375
],
[
0.096435546875,
0.7109375,
0.1502685546875,
0.7109375,
0.1502685546875,
0.72265625,
0.096435546875,
0.72265625
],
[
0.108642578125,
0.7236328125,
0.1380615234375,
0.7236328125,
0.1380615234375,
0.73486328125,
0.108642578125,
0.73486328125
],
[
0.191162109375,
0.72900390625,
0.2132568359375,
0.72900390625,
0.2132568359375,
0.7373046875,
0.191162109375,
0.7373046875
],
[
0.50341796875,
0.7236328125,
0.89208984375,
0.7236328125,
0.89208984375,
0.73779296875,
0.50341796875,
0.73779296875
],
[
0.176513671875,
0.7421875,
0.279296875,
0.7421875,
0.279296875,
0.7568359375,
0.176513671875,
0.7568359375
],
[
0.50244140625,
0.73779296875,
0.89208984375,
0.73779296875,
0.89208984375,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.5009765625,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.7666015625,
0.5009765625,
0.7666015625
],
[
0.1004638671875,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.1004638671875,
0.77880859375
],
[
0.50244140625,
0.76953125,
0.890625,
0.76953125,
0.890625,
0.78271484375,
0.50244140625,
0.78271484375
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.7841796875,
0.890625,
0.7841796875,
0.890625,
0.79736328125,
0.50244140625,
0.79736328125
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.796875,
0.4697265625,
0.81103515625,
0.080078125,
0.81005859375
],
[
0.52294921875,
0.7998046875,
0.89208984375,
0.7998046875,
0.89208984375,
0.81298828125,
0.52294921875,
0.81298828125
],
[
0.08087158203125,
0.8125,
0.46728515625,
0.8125,
0.46728515625,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.50244140625,
0.814453125,
0.89208984375,
0.814453125,
0.89208984375,
0.82763671875,
0.50244140625,
0.82763671875
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82763671875,
0.890625,
0.830078125,
0.890625,
0.84326171875,
0.5009765625,
0.84130859375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.8447265625,
0.890625,
0.8447265625,
0.890625,
0.85791015625,
0.50244140625,
0.85791015625
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85986328125,
0.7099609375,
0.85986328125,
0.7099609375,
0.87353515625,
0.50244140625,
0.87353515625
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.468017578125,
0.88525390625,
0.468017578125,
0.8994140625,
0.07843017578125,
0.90087890625
],
[
0.6259765625,
0.88330078125,
0.7685546875,
0.88525390625,
0.7685546875,
0.90283203125,
0.6259765625,
0.90087890625
],
[
0.869140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.869140625,
0.90087890625
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"training pipeline [11]. Apart from preventing the propaga-",
"respect to the more information-rich latent representation ,",
"tion of bias to downstream modeling tasks, this approach",
"or s y. We can then use this learned fair latent repre-",
"of fairness froma data management standpoint alsoleads",
"sentation in a downstream prediction or generation task.",
"to more robust and significant fairness measures by ensur-",
"3.1. Enforcing CI with respect to Label",
"ing that data sources, transformations, and other training as-",
"sumptions are sound [24].",
"When the label y is highly correlated with the protected at-",
"tribute s in the training data, a basic aim for a fair machine",
"Representation",
"Learning with Diffusion Models",
"learning model is to ensure conditional independence of the",
"Diffusion-based models learn to generate images by a",
"predicted g from s, i.e., s I y I y.",
"denoising process: Random Gaussian noise is added to",
"Assuming that the joint distribution p(s, y, y) is well-",
"input images, and the model learns how to reverse this",
"defined with respect to Lebesgue measure μ, previous",
"process to hallucinate new images from random noise.",
"work [3] demonstrates that the above conditional indepen-",
"This family of models has proven to offer remarkable",
"dence can be defined with respect to the Jensen-Shannon",
"quality in image generation, promising to replace gener-",
"divergence (JS) between p(s, g, y) and an auxiliary distri-",
"ative adversarial networks (GANs) [8] as the dominant",
"bution q(s, y', y), where the joint marginal q(s, y) is dis-",
"architectural paradigm. Chief among them is the diffusion",
"tributed identically to p(s, y), and q(y'ly) is independent of",
"autoencoder[23lwhich encodesanimageinto a two-part",
"s. In other words, when q(s, y',y) = p(s, y)q(y'ly), the",
"latent subcode, capturing a stochastic representation via",
"conditional independence constraint s g I y is equivalent",
"the aforementioned denoising approach and conditioning",
"to enforcing",
"this process upon a semantic representation learned by",
"JS(p(9, y), q(y', y)) = JS(p(s, 9, y), q(s, y', y)):",
"(2)",
"smooth and meaningful changes in the decoded image.",
"a CNN. Interpolation of the resulting subcode results in",
"Accordingly, we can rewrite (1) in terms of JS as",
"The semantic separation capability of DiffAEs has enabled",
"various applications, such as attribute manipulation and",
"17",
"various low-shot or zero-shot downstream applications",
"argmin",
"y;log(9) + (1 - yi) log(1 - 9i)",
"0,",
"n",
"(3)",
"[7, 9, 26]. In this work, we further enable DiffAEs to learn",
"unbiased, causal, and conditionally independent latent",
"s.t.",
"(JS(p(9, y), q(y',y)-",
"subcode representations.",
"Js(p(s, 9,y),q(s, y', y)) ≤s)",
"3. Method",
"for some > 0 sufficiently small.",
"Let [α, y, s] be a training sample in the dataset, where ",
"With the help of a conditional sampler g(y' I y), the JS",
"is an input image, y is the target prediction label, and s is",
"constraint of(3)can be satisfied by a general GAN architec-",
"a protected sensitive attribute. We aim to learn an encoder",
"ture with two discriminators (Fig. 2) [3], i.e., by minimizing",
"network ge(α) =w resulting in a latent representation ,",
"Lcl = (LDi - Lp2)2",
"from which a prediction network fo(u) = 9 produces the",
"(4)",
"final predicted label y. With a slight abuse of notation, we",
"where",
"uselower-caseletterstoalsodenoterandomvariables.Ulti-",
"mately, we aim to train a model that achieves high accuracy",
"Lpl =Elog(1 - Di(y', y)) + Eu.s[log Di(y, y)l and",
"in predicting label y while ensuring equalized odds:",
"Lp2 = E[log(1 - D2(y', s, y)] + Ey,s,y' [log D2(9, s, y)]",
"1n",
"uilog(0)+(1) og(1)",
"The key advantage of using (4) to derive conditional in-",
"(1)",
"argmin",
"0,$",
"i=1",
"dependence is that q(y'l y) does not have to be a perfect",
"s.t.slyl y.",
"sampler that exactly matches p(y y). The only sufficient",
"condition is that g(y'l y) shares overlapping support with",
"We discuss two possible ways of enforcing the equalized",
"the original distribution p(g y). This gives us many flexi-",
"odds fairness constraint: First, we enforce conditional inde-",
"ble ways to construct g, such as using a uniform distribution.",
"pendence with respect to the label y, i.e., s I y I y. As",
"Specifically, in a supervised learning setting, where y is",
"discussed in Ahuja et al.,this can be seen as a means to en-",
"known at training time, we can set q using a uniform dis-",
"force equalized odds through CI. However, we claim that",
"tribution over all raw model outputs that correspond to y.",
"enforcing the independence constraint with respect to only",
"The encoder and prediction networks can then be learned",
"the label is limited in its ability to enforce the equalized",
"by optimizing the loss function",
"odds fairness constraint effectively. We therefore addition-",
"ally propose to enforce the conditional independence with",
"Lene = ^LcI + Ltask,",
"(5)",
"3"
] |
{
"arxiv_id": "2404.13798",
"language": "en",
"page_idx": 2
}
|
||
2404.13798-en-3
|
2404.13798-en-3.jpg
|
where $\lambda$ is a hyperparameter and $L_\mathrm{task}$ is the loss function for the prediction task. The discriminators can be trained adversarially to minimize the composite loss function:
$$L_{\mathrm{D}}=L_{\mathrm{D1}}+L_{\mathrm{D2}}.$$
(6)
# 3.2. Enforcing CI with respect to Latent Representation
Enforcing conditional independence only with respect to $\hat{y}$ does not necessarily prevent the model from learning biased information from $s$ in its representations. In this paper, we propose to enforce conditional independence with respect to the latent representation $v$ in a way that most directly impacts the learnable parameters of the encoder. We achieve this by replacing the predicted label $\hat{y}$ with the latent vector $v$ within our conditional independence loss described above, $i.e.,s\perp v\mid y.$ Similar to (2), our Jensen-Shannon divergence constraint becomes
$$\mathrm{JS}(p(v,y),q(v',y))=\mathrm{JS}(p(s,v,y),q(s,v',y)).$$
(7)
As before, this requires a conditional sampler $q(v^{\prime}|y).$ However, one problem that arises with this approach is that the conditional sampler is no longer well-defined at training time. Since the distribution and support of $p(v|y)$ is learned by the model on the fly, we cannot construct $q(v^{\prime}|y)$ a priori.
To resolve this challenge, we implement the imperfect conditional sampler $q$ via a novel dynamic sampling procedure. Specifically, to sample $v^\prime\sim q(v^{\prime}|y)$, we sample with replacement from the latent vectors associated with the given $y$ within the same training batch. Similar to the bootstrapping procedure, our dynamic sampling can approximate the empirical distribution function $p(v|y)$, allowing the discriminators to keep up with and robustly train the encoder as the latent space is learned. Note that, other than the difference in the sampler, the objective function of (4) and the GAN architecture in Fig. 2 can be translated here by replacing $\{\hat{y},y^{\prime}\}$ with $\{v,v^{\prime}\}.$
$3. 3. \textbf{ Disentangling DiffAE Latent Representation}$
The above construction of $q$ provides us with a powerful technique that can be applied to any encoder architecture. We demonstrate this by using the state-of-the-art diffusion autoencoder model[23], which encodes images using a denoising diffusion process [26] conditioned upon a semantically meaningful representation inferred by a CNN encoder. An image, therefore, is encoded by two representations: a semantic subcode learned by the CNN, along with a stochastic subcode that, given the semantic subcode, denoises into the original image. The model learns to compress the most common high-level features into the semantic subcode and delegate the remaining, highly variable features to the stochastic representation. When applied to human face images, this yields a clean separation between
anatomical features such as sex, facial structure, and skin type in the semantic subcode, and more transient qualities like pose, hairstyle, and expression in the stochastic sub- $code.$
By enforcing conditional independence with respect to the semantic subcode of the diffusion autoencoder, we produce a latent representation that is invariant to a protected attribute of choice. The model adapts by encoding any facial features associated with the protected attribute in the stochastic subcode instead. When such facial features include the high-level attributes described earlier, this can unduly affect the natural dichotomy between the semantic and stochastic subcodes, restricting the expressive power of the semantic subcode and reducing the efficiency of the denoising process used for image generation. Empirically, this results in blurred and unrealistic generated images.
To overcome this, we enforce conditional independence with respect to only a portion of the semantic subcode. To further encourage the model to use the remaining portion to encode the facial features related to the sensitive attribute, we add a prediction head on the remaining portion to estimate the sensitive attribute $s.$ We optimize this new prediction head alongside the encoder; instead of (5), we have
$$L_{\mathrm{Enc}}=\lambda_{1}L_{\mathrm{CI}}+\lambda_{2}\:\mathrm{BCE}(\hat{s},s)+L_{\mathrm{task}}.$$
(8)
In this way, the semantic subcode is apportioned into two parts: one variant to the sensitive attribute, and the other invariant. Although we only demonstrate this capability on tasks involving a single sensitive attribute, this approach can be easily extended to disentangle multiple attributes within the semantic subcode. With this formulation for causal disentanglement of the latent space, we show that the specific subcode in which a given facial attribute is represented can be methodically adjusted $(e.g.$, to change the race of the generated image), all while preserving the model's ability to accurately reconstruct other characteristics of the input $image.$
# 4. Experiments and Results
We apply our architecture to two different settings, each with a unique dataset, model, and confounding variable. In each case, we analyze the effectiveness of applying CI in the latent space as opposed to the label space and also compare with several baselines. Experiments were performed using a P100 GPU, with the exception of the diffusion autoencoder experiments which required four A100 GPUs.
Metrics We report prediction accuracies for each value of the protected attribute when $s$ is discrete, along with the balanced accuracy (bAcc) to account for class imbalances. To evaluate invariance to the sensitive attribute, we use the squared distance correlation (dcor$^2)$ [28] to quantitatively
4
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.435546875,
0.12371826171875,
0.435546875,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.50341796875,
0.1395263671875,
0.5400390625,
0.1395263671875,
0.5400390625,
0.15087890625,
0.50341796875,
0.15087890625
],
[
0.214111328125,
0.1483154296875,
0.33251953125,
0.1483154296875,
0.33251953125,
0.1634521484375,
0.214111328125,
0.1634521484375
],
[
0.44677734375,
0.1483154296875,
0.468017578125,
0.1483154296875,
0.468017578125,
0.1622314453125,
0.44677734375,
0.1622314453125
],
[
0.52294921875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.52294921875,
0.1666259765625
],
[
0.080078125,
0.17236328125,
0.468017578125,
0.1749267578125,
0.468017578125,
0.189453125,
0.080078125,
0.1875
],
[
0.50341796875,
0.1685791015625,
0.8896484375,
0.1685791015625,
0.8896484375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.50341796875,
0.184326171875,
0.89208984375,
0.184326171875,
0.89208984375,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.12091064453125,
0.1912841796875,
0.1666259765625,
0.1912841796875,
0.1666259765625,
0.20263671875,
0.12091064453125,
0.20263671875
],
[
0.50341796875,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.08087158203125,
0.212158203125,
0.46728515625,
0.212158203125,
0.46728515625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.5009765625,
0.212158203125,
0.89208984375,
0.21337890625,
0.89208984375,
0.2279052734375,
0.5009765625,
0.2266845703125
],
[
0.08251953125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.240478515625,
0.08251953125,
0.240478515625
],
[
0.50341796875,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.50341796875,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.270751953125,
0.50341796875,
0.270751953125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.27587890625,
0.890625,
0.27587890625,
0.890625,
0.285888671875,
0.50341796875,
0.285888671875
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.2890625,
0.890625,
0.2890625,
0.890625,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.304931640625,
0.8935546875,
0.304931640625,
0.8935546875,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.50244140625,
0.3193359375,
0.8349609375,
0.3193359375,
0.8349609375,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.52294921875,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.347900390625,
0.52294921875,
0.347900390625
],
[
0.08251953125,
0.35107421875,
0.283447265625,
0.35107421875,
0.283447265625,
0.361083984375,
0.08251953125,
0.361083984375
],
[
0.5048828125,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.5048828125,
0.363037109375
],
[
0.50341796875,
0.365478515625,
0.89208984375,
0.365478515625,
0.89208984375,
0.37890625,
0.50341796875,
0.37890625
],
[
0.1070556640625,
0.375,
0.426513671875,
0.375,
0.426513671875,
0.389404296875,
0.1070556640625,
0.389404296875
],
[
0.435546875,
0.376953125,
0.468017578125,
0.376953125,
0.468017578125,
0.38818359375,
0.435546875,
0.38818359375
],
[
0.5009765625,
0.37890625,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.5009765625,
0.392578125
],
[
0.08087158203125,
0.401611328125,
0.46728515625,
0.401611328125,
0.46728515625,
0.415283203125,
0.08087158203125,
0.415283203125
],
[
0.50341796875,
0.395751953125,
0.890625,
0.395751953125,
0.890625,
0.4091796875,
0.50341796875,
0.4091796875
],
[
0.50341796875,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50341796875,
0.423583984375
],
[
0.07843017578125,
0.416748046875,
0.468017578125,
0.415283203125,
0.468017578125,
0.429931640625,
0.07843017578125,
0.43115234375
],
[
0.50341796875,
0.425537109375,
0.89208984375,
0.425537109375,
0.89208984375,
0.439453125,
0.50341796875,
0.439453125
],
[
0.08087158203125,
0.434326171875,
0.46728515625,
0.434326171875,
0.46728515625,
0.4443359375,
0.08087158203125,
0.4443359375
],
[
0.080078125,
0.4462890625,
0.468017578125,
0.4462890625,
0.468017578125,
0.4609375,
0.080078125,
0.4609375
],
[
0.5703125,
0.455810546875,
0.82421875,
0.455810546875,
0.82421875,
0.47021484375,
0.5703125,
0.47021484375
],
[
0.080078125,
0.462158203125,
0.468017578125,
0.462158203125,
0.468017578125,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.87109375,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.4697265625,
0.87109375,
0.4697265625
],
[
0.098876953125,
0.4765625,
0.4697265625,
0.4765625,
0.4697265625,
0.490478515625,
0.098876953125,
0.490478515625
],
[
0.521484375,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.49365234375,
0.521484375,
0.49365234375
],
[
0.080078125,
0.492431640625,
0.468017578125,
0.492431640625,
0.468017578125,
0.505859375,
0.080078125,
0.505859375
],
[
0.50341796875,
0.496826171875,
0.890625,
0.496826171875,
0.890625,
0.5068359375,
0.50341796875,
0.5068359375
],
[
0.07843017578125,
0.5068359375,
0.468017578125,
0.505859375,
0.468017578125,
0.52099609375,
0.07843017578125,
0.521484375
],
[
0.50244140625,
0.51025390625,
0.89208984375,
0.51025390625,
0.89208984375,
0.5234375,
0.50244140625,
0.5234375
],
[
0.08087158203125,
0.52294921875,
0.4697265625,
0.52294921875,
0.4697265625,
0.5361328125,
0.08087158203125,
0.5361328125
],
[
0.50244140625,
0.5244140625,
0.89208984375,
0.5244140625,
0.89208984375,
0.53857421875,
0.50244140625,
0.53857421875
],
[
0.08087158203125,
0.537109375,
0.4697265625,
0.537109375,
0.4697265625,
0.55029296875,
0.08087158203125,
0.55029296875
],
[
0.5009765625,
0.5390625,
0.890625,
0.5390625,
0.890625,
0.55322265625,
0.5009765625,
0.55322265625
],
[
0.080078125,
0.55322265625,
0.468017578125,
0.55322265625,
0.468017578125,
0.56640625,
0.080078125,
0.56640625
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.5673828125,
0.4697265625,
0.5673828125,
0.4697265625,
0.58203125,
0.080078125,
0.58203125
],
[
0.50341796875,
0.57080078125,
0.89208984375,
0.57080078125,
0.89208984375,
0.583984375,
0.50341796875,
0.583984375
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.50341796875,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.08087158203125,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.60986328125,
0.08087158203125,
0.60986328125
],
[
0.50244140625,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.080078125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.080078125,
0.626953125
],
[
0.50244140625,
0.61572265625,
0.8896484375,
0.61572265625,
0.8896484375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.080078125,
0.62744140625,
0.4697265625,
0.62744140625,
0.4697265625,
0.6416015625,
0.080078125,
0.6416015625
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.08087158203125,
0.64306640625,
0.298095703125,
0.64306640625,
0.298095703125,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.5009765625,
0.64599609375,
0.55224609375,
0.64599609375,
0.55224609375,
0.66015625,
0.5009765625,
0.66015625
],
[
0.08087158203125,
0.66650390625,
0.453369140625,
0.66650390625,
0.453369140625,
0.68115234375,
0.08087158203125,
0.68115234375
],
[
0.50341796875,
0.6748046875,
0.73193359375,
0.6748046875,
0.73193359375,
0.689453125,
0.50341796875,
0.689453125
],
[
0.08087158203125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.70068359375,
0.89208984375,
0.70068359375,
0.89208984375,
0.71533203125,
0.50341796875,
0.71533203125
],
[
0.08087158203125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.08087158203125,
0.734375
],
[
0.50341796875,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50341796875,
0.72998046875
],
[
0.50244140625,
0.73095703125,
0.89208984375,
0.73095703125,
0.89208984375,
0.744140625,
0.50244140625,
0.744140625
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.7470703125,
0.890625,
0.7470703125,
0.890625,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76123046875,
0.89208984375,
0.76123046875,
0.89208984375,
0.7744140625,
0.50244140625,
0.7744140625
],
[
0.50244140625,
0.77587890625,
0.8896484375,
0.77587890625,
0.8896484375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79150390625,
0.875,
0.79150390625,
0.875,
0.8046875,
0.50244140625,
0.8046875
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48193359375,
0.9267578125,
0.491943359375,
0.9267578125,
0.491943359375,
0.9345703125,
0.48193359375,
0.9345703125
]
] |
[
"where > is a hyperparameter and Ltask is the loss function",
"anatomical features such as sex, facial structure, and skin",
"for the prediction task. The discriminators can be trained",
"type in the semantic subcode, and more transient qualities",
"adversarially to minimize the composite loss function:",
"like pose, hairstyle, and expression in the stochastic sub-",
"code.",
"LD = LD1 + Lp2.",
"(6)",
"By enforcing conditional independence with respect to",
"3.2. Enforcing CI with respect to Latent Represen-",
"the semantic subcode of the diffusion autoencoder, we pro-",
"duce a latent representation that is invariant to a protected",
"tation",
"attribute of choice. The model adapts by encoding any fa-",
"Enforcing conditional independence only with respect to y",
"cial features associated with the protected attribute in the",
"doesnotnecessarilypreventthemodelfromlearningbiased",
"stochastic subcode instead. When such facial features in-",
"information from s in its representations. In this paper, we",
"clude the high-level attributes described earlier, this can un-",
"propose to enforce conditional independence with respect",
"dulyaffectthenaturaldichotomybetweenthesemanticand",
"to the latent representation in a way that most directly im-",
"stochasticsubcodes,restrictingtheexpressivepowerofthe",
"pacts the learnable parameters of the encoder. We achieve",
"semantic subcode and reducing the efficiency of the denois-",
"this by replacing the predicted label y with the latent vec-",
"ing process used for image generation.Empirically,this",
"tor u within our conditional independence loss described",
"results in blurred and unrealistic generated images.",
"above, i.e., s u I y. Similar to (2), our Jensen-Shannon",
"To overcome this, we enforce conditional independence",
"divergenceconstraintbecomes",
"with respect to only a portion of the semantic subcode. To",
"further encourage the model to use the remaining portion to",
"JS(p(u, y), q(u', y)) = JS(p(s, U, y),q(s, v', y)).",
"(7)",
"encode the facial features related to the sensitive attribute,",
"As before, this requires a conditional sampler q(ly). How-",
"we add a prediction head on the remaining portion to es-",
"timate the sensitive attribute s. We optimize this new pre-",
"ever, one problem that arises with this approach is that the",
"diction head alongside the encoder; instead of (5), we have",
"conditional sampler is no longer well-defined at training",
"time. Since the distribution and support of p(uly) is learned",
"Lenc = >iLcI + >2 BCE(3s, s) + Ltask.",
"by the model on the fly, we cannot construct q('ly) a priori.",
"(8)",
"To resolve this challenge, we implement the imperfect",
"In this way, the semantic subcode is apportioned into two",
"conditional sampler q via a novel dynamic sampling pro-",
"parts: one variant to the sensitive attribute, and the other",
"cedure. Specifically, to sample u' ~ q(u'ly),we sample",
"invariant. Although we only demonstrate this capability on",
"with replacement from the latent vectors associated with the",
"tasks involving a single sensitive attribute, this approach can",
"given y within the same training batch. Similar to the boot-",
"be easily extended to disentangle multiple attributes within",
"strapping procedure, our dynamic sampling can approxi-",
"the semantic subcode. With this formulation for causal dis-",
"mate the empirical distribution function p(uy), allowing",
"entanglement of the latent space, we show that the specific",
"the discriminators to keep up with and robustly train the",
"subcode in which a given facial attribute is represented can",
"encoderasthelatent space is learned.Notethat.otherthan",
"be methodically adjusted (e.g., to change the race of the",
"the difference in the sampler, the objective function of (4)",
"generated image), all while preserving the model's ability",
"and the GAN architecture in Fig. 2 can be translated here",
"to accurately reconstruct other characteristics of the input",
"by replacing (y, y'} with [u, u'].",
"image.",
"3.3.Disentangling DiffAE Latent Representation",
"4. Experiments and Results",
"The above construction of q provides us with a powerful",
"technique that can be applied to any encoder architecture.",
"We apply our architecture to two different settings, each",
"We demonstrate this by using the state-of-the-art diffusion",
"with a unique dataset, model, and confounding variable. In",
"each case, we analyze the effectiveness of applying CI in",
"autoencodermodel[23l,whichencodesimages usinga de-",
"noising diffusion process [26] conditioned upon a seman-",
"the latent space as opposed to the label space and also com-",
"tically meaningful representation inferred by a CNN en-",
"pare with several baselines. Experiments were performed",
"using a P100 GPU, with the exception of the diffusion au-",
"coder.An image, therefore,is encoded by two representa-",
"tions: a semantic subcode learned by the CNN, along with",
"toencoder experiments which required four A100 GPUs.",
"a stochastic subcode that, given the semantic subcode, de-",
"noises into the original image. The model learns to com-",
"MetricsWe report prediction accuracies for each value of",
"press the most common high-level features into the seman-",
"the protected attribute when s is discrete, along with the",
"tic subcode and delegate the remaining, highly variable fea-",
"balanced accuracy (bAcc) to account for class imbalances.",
"tures to the stochastic representation. When applied to hu-",
"To evaluate invariance to the sensitive attribute, we use the",
"man face images, this yields a clean separation between",
"squared distance correlation (dcor?) [28] to quantitatively",
"4"
] |
{
"arxiv_id": "2404.13798",
"language": "en",
"page_idx": 3
}
|
||
2307.04893-en-0
|
2307.04893-en-0.jpg
|
# $\textbf{Choosing Well Your Opponents: }$ How to Guide the Synthesis of Programmatic Strategies
Rubens O. Moraes$^1,2,\textbf{ David S. Aleixo}^1,\textbf{ Lucas N.Ferreira}^2$ and Levi H.S. Lelis$^2$
$^{1}$ Departamento de Informática, Universidade Federal de Viçosa, Brazil
$^{2}$Department of Computing Science, University of Alberta, Canada
Alberta Machine Intelligence Institute (Amii)
rubens.moraes@ufv.br, david.aleixo@ufv.br, Inferrei@ualberta.ca, levi.lelis@ualberta.ca
## Abstract
This paper introduces Local Learner (2L), an algorithm for providing a set of reference strategies to guide the search for programmatic strategies in two-player zero-sum games. Previous learning algorithms, such as Iterated Best Response (IBR), Fictitious Play (FP), and Double-Oracle (DO), can be computationally expensive or miss important information for guiding search algorithms. 2L actively selects a set of reference strategies to improve the search signal. We empirically demonstrate the advantages of our approach while guiding a local search algorithm for synthesizing strategies in three games, including MicroRTS, a challenging realtime strategy game. Results show that 2L learns reference strategies that provide a stronger search signal than IBR, FP, and DO. We also simulate a tournament of MicroRTS, where a synthesizer using 2L outperformed the winners of the two latest MicroRTS competitions, which were programmatic strategies written by human programmers.
strategy that IBR provides serves as a guiding function. In the context of local search, when considering the neighbors of a candidate solution, local search algorithms prefer to accept a program that achieves a higher utility value against the target strategy. Since IBR considers a single strategy as a target, the search signal is often weak. This is because the neighbors of a candidate solution that performs poorly against the target strategy are also likely to perform poorly against it— small changes to a losing program will also generate a losing program. Moreover, IBR can loop around the strategy space in games with dynamics similar to Rock, Paper, and Scissors, without making progress toward strong solutions.
In this paper, we adapt Fictitious Play (FP) [Brown, 1951] and Double Oracle (DO) [McMahan et al., 2003] to the context of programmatic strategies. FP and DO have been used in the context of neural strategies to overcome some of the weaknesses of IBR [Lanctot $\breve{e}t$ al.,2017]. Despite providing a better search signal than IBR, we show that FP and DO can still fail to provide relevant information for the search. We then introduce a novel learning algorithm, Local Learner (2L), that is designed specifically for guiding local search algorithms in the synthesis of programmatic strategies. 2 L uses information gathered while computing best responses to decide the set of target strategies to be used in future iterations of the algorithm as a means of optimizing the search signal.
We evaluate 2L on three two-player zero-sum games: MicroRTS [Ontañón et al., 2018], Poachers & Rangers, and Climbing Monkeys. The results show that 2L synthesized strategies that are never worse and often far superior to strategies synthesized with IBR, FP, and DO in all three domains We also performed a simulated competition of MicroRTS with strategies synthesized with 2L IBR,FP,DO, as well as the programmatic strategies that won the last two MicroRTS competitions, which were written by programmers. 2L obtained the highest average winning rate in our tournament.
### 1 Introduction
Programmatic strategies encode game strategies in humanunderstandable programs. Such programmatic encoding allows domain experts to interpret and modify computergenerated strategies, which can be valuable depending on the application domain (e.g., the games industry). Previous works have used Iterated-Best Response (IBR) [Lanctot et $al.,2017]$ as the learning algorithm for synthesizing programmatic strategies [Mariño et al., 2021]. Given a game, IBR starts with an arbitrary strategy for playing the game and it approximates a best response to it; in the next iteration, it approximates a best response to the best response. This process is repeated a number of iterations and the programmatic strategy synthesized in the last iteration is returned.
The computation of the best responses in the IBR loop is performed by searching in the programmatic space defined by a domain-specific language. Given a target strategy, the algorithm searches for a program encoding a best response to it. Previous work used local search algorithms for searching in the programmatic space[Mariño et al., 2021; Medeiros $etal.,2022;$ Aleixo and Lelis, 2023]. The target
## 2 $\textbf{Problem Definition}$
We consider the synthesis of programmatic strategies assum- $\begin{array}{l}{\mathrm{ing~zero-sum~two-player~games~}\mathcal{G}=~(\mathcal{P},\mathcal{S},s_{\mathrm{init}},\mathcal{A},\mathcal{T},\mathcal{U}).}\\{\mathrm{Let~}\mathcal{P}=\{i,-i\}~\text{be the pair of players; }\mathcal{S}~\text{be the set of states,}}\end{array}$ with $s_\mathrm{init}$ in $\mathcal{S}$ being the initial state. Each player $i$ can perform an action from a legal set of actions $\hat{\mathcal{A}}_i(s)$ in $\mathcal{A}$ for a given state $s.$ The action of each player is given by a strategy, which is a function $\sigma_i$ that receives a state $s$ in $S$ and returns
|
[
[
0.339111328125,
0.111083984375,
0.6611328125,
0.111083984375,
0.6611328125,
0.1287841796875,
0.339111328125,
0.1287841796875
],
[
0.2197265625,
0.1319580078125,
0.779296875,
0.1319580078125,
0.779296875,
0.149658203125,
0.2197265625,
0.149658203125
],
[
0.15283203125,
0.1634521484375,
0.8388671875,
0.1634521484375,
0.8388671875,
0.18115234375,
0.15283203125,
0.18115234375
],
[
0.2181396484375,
0.184326171875,
0.77783203125,
0.186279296875,
0.77783203125,
0.2008056640625,
0.2181396484375,
0.1988525390625
],
[
0.233642578125,
0.2008056640625,
0.7646484375,
0.2020263671875,
0.7646484375,
0.2197265625,
0.233642578125,
0.2183837890625
],
[
0.318603515625,
0.2183837890625,
0.67626953125,
0.2183837890625,
0.67626953125,
0.231689453125,
0.318603515625,
0.231689453125
],
[
0.1502685546875,
0.2342529296875,
0.84716796875,
0.2342529296875,
0.84716796875,
0.248779296875,
0.1502685546875,
0.248779296875
],
[
0.0236968994140625,
0.272705078125,
0.05963134765625,
0.272705078125,
0.05963134765625,
0.339111328125,
0.0236968994140625,
0.339111328125
],
[
0.2484130859375,
0.271484375,
0.325927734375,
0.271484375,
0.325927734375,
0.2890625,
0.2484130859375,
0.2890625
],
[
0.5107421875,
0.274658203125,
0.91259765625,
0.2734375,
0.91259765625,
0.287353515625,
0.5107421875,
0.28857421875
],
[
0.51123046875,
0.28857421875,
0.9140625,
0.28857421875,
0.9140625,
0.3017578125,
0.51123046875,
0.3017578125
],
[
0.1192626953125,
0.29541015625,
0.453369140625,
0.29541015625,
0.453369140625,
0.31005859375,
0.1192626953125,
0.31005859375
],
[
0.51318359375,
0.3017578125,
0.91259765625,
0.302978515625,
0.91259765625,
0.31640625,
0.51318359375,
0.31494140625
],
[
0.1192626953125,
0.311279296875,
0.450927734375,
0.311279296875,
0.450927734375,
0.3212890625,
0.1192626953125,
0.3212890625
],
[
0.51318359375,
0.31640625,
0.91259765625,
0.31640625,
0.91259765625,
0.330078125,
0.51318359375,
0.330078125
],
[
0.02532958984375,
0.328857421875,
0.056365966796875,
0.328857421875,
0.056365966796875,
0.371826171875,
0.02532958984375,
0.371826171875
],
[
0.1192626953125,
0.324462890625,
0.455078125,
0.324462890625,
0.455078125,
0.33837890625,
0.1192626953125,
0.33837890625
],
[
0.51123046875,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.3447265625,
0.51123046875,
0.3447265625
],
[
0.11767578125,
0.33837890625,
0.455078125,
0.33837890625,
0.455078125,
0.3515625,
0.11767578125,
0.3515625
],
[
0.51318359375,
0.343505859375,
0.9111328125,
0.343505859375,
0.9111328125,
0.356689453125,
0.51318359375,
0.356689453125
],
[
0.1192626953125,
0.3515625,
0.451904296875,
0.3515625,
0.451904296875,
0.365478515625,
0.1192626953125,
0.365478515625
],
[
0.02532958984375,
0.361083984375,
0.056365966796875,
0.361083984375,
0.056365966796875,
0.397705078125,
0.02532958984375,
0.397705078125
],
[
0.51318359375,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.1192626953125,
0.365478515625,
0.453369140625,
0.365478515625,
0.453369140625,
0.37890625,
0.1192626953125,
0.37890625
],
[
0.51123046875,
0.37060546875,
0.90869140625,
0.37060546875,
0.90869140625,
0.3837890625,
0.51123046875,
0.3837890625
],
[
0.12091064453125,
0.380615234375,
0.450927734375,
0.380615234375,
0.450927734375,
0.391357421875,
0.12091064453125,
0.391357421875
],
[
0.51123046875,
0.385009765625,
0.91259765625,
0.385009765625,
0.91259765625,
0.3984375,
0.51123046875,
0.3984375
],
[
0.1192626953125,
0.393310546875,
0.451904296875,
0.393310546875,
0.451904296875,
0.4072265625,
0.1192626953125,
0.4072265625
],
[
0.0196075439453125,
0.404052734375,
0.05963134765625,
0.40283203125,
0.062103271484375,
0.48486328125,
0.0228729248046875,
0.486083984375
],
[
0.51318359375,
0.399658203125,
0.9140625,
0.399658203125,
0.9140625,
0.41357421875,
0.51318359375,
0.41357421875
],
[
0.1192626953125,
0.4091796875,
0.451904296875,
0.4091796875,
0.451904296875,
0.419189453125,
0.1192626953125,
0.419189453125
],
[
0.51123046875,
0.412353515625,
0.91259765625,
0.41357421875,
0.91259765625,
0.4267578125,
0.51123046875,
0.425537109375
],
[
0.1192626953125,
0.42236328125,
0.451904296875,
0.42236328125,
0.451904296875,
0.43310546875,
0.1192626953125,
0.43310546875
],
[
0.51318359375,
0.4267578125,
0.8349609375,
0.4267578125,
0.8349609375,
0.43994140625,
0.51318359375,
0.43994140625
],
[
0.1192626953125,
0.43505859375,
0.453369140625,
0.43505859375,
0.453369140625,
0.448974609375,
0.1192626953125,
0.448974609375
],
[
0.52783203125,
0.43994140625,
0.9140625,
0.43994140625,
0.9140625,
0.453857421875,
0.52783203125,
0.453857421875
],
[
0.1168212890625,
0.448974609375,
0.450927734375,
0.448974609375,
0.450927734375,
0.462158203125,
0.1168212890625,
0.462158203125
],
[
0.5146484375,
0.455810546875,
0.91015625,
0.455810546875,
0.91015625,
0.466552734375,
0.5146484375,
0.466552734375
],
[
0.12091064453125,
0.464111328125,
0.450927734375,
0.464111328125,
0.450927734375,
0.474853515625,
0.12091064453125,
0.474853515625
],
[
0.51318359375,
0.468505859375,
0.91259765625,
0.468505859375,
0.91259765625,
0.481689453125,
0.51318359375,
0.481689453125
],
[
0.1168212890625,
0.4765625,
0.451904296875,
0.4765625,
0.451904296875,
0.490478515625,
0.1168212890625,
0.490478515625
],
[
0.5146484375,
0.484130859375,
0.91259765625,
0.484130859375,
0.91259765625,
0.494384765625,
0.5146484375,
0.494384765625
],
[
0.0228729248046875,
0.4912109375,
0.060455322265625,
0.4912109375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.12091064453125,
0.492431640625,
0.451904296875,
0.492431640625,
0.451904296875,
0.50244140625,
0.12091064453125,
0.50244140625
],
[
0.5146484375,
0.49755859375,
0.91015625,
0.49755859375,
0.91015625,
0.50830078125,
0.5146484375,
0.50830078125
],
[
0.1192626953125,
0.50390625,
0.455078125,
0.50390625,
0.455078125,
0.517578125,
0.1192626953125,
0.517578125
],
[
0.51318359375,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51318359375,
0.5234375
],
[
0.1192626953125,
0.517578125,
0.451904296875,
0.517578125,
0.451904296875,
0.53076171875,
0.1192626953125,
0.53076171875
],
[
0.51123046875,
0.5234375,
0.91259765625,
0.5234375,
0.91259765625,
0.537109375,
0.51123046875,
0.537109375
],
[
0.11767578125,
0.5322265625,
0.451904296875,
0.5322265625,
0.451904296875,
0.54541015625,
0.11767578125,
0.54541015625
],
[
0.51318359375,
0.537109375,
0.91259765625,
0.537109375,
0.91259765625,
0.55029296875,
0.51318359375,
0.55029296875
],
[
0.1192626953125,
0.54443359375,
0.450927734375,
0.54541015625,
0.450927734375,
0.5595703125,
0.1192626953125,
0.55810546875
],
[
0.51123046875,
0.5498046875,
0.91259765625,
0.55029296875,
0.91259765625,
0.56494140625,
0.51123046875,
0.564453125
],
[
0.1168212890625,
0.55810546875,
0.43701171875,
0.5595703125,
0.43701171875,
0.57373046875,
0.1168212890625,
0.57275390625
],
[
0.51318359375,
0.56494140625,
0.91259765625,
0.56494140625,
0.91259765625,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.51318359375,
0.5791015625,
0.9111328125,
0.5791015625,
0.9111328125,
0.59228515625,
0.51318359375,
0.59228515625
],
[
0.11358642578125,
0.59228515625,
0.223876953125,
0.59228515625,
0.223876953125,
0.60693359375,
0.11358642578125,
0.60693359375
],
[
0.51318359375,
0.5947265625,
0.9111328125,
0.5947265625,
0.9111328125,
0.60498046875,
0.51318359375,
0.60498046875
],
[
0.0849609375,
0.6123046875,
0.484375,
0.611328125,
0.484375,
0.62548828125,
0.0849609375,
0.626953125
],
[
0.51318359375,
0.60693359375,
0.90283203125,
0.60693359375,
0.90283203125,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.08660888671875,
0.626953125,
0.484375,
0.626953125,
0.484375,
0.64013671875,
0.08660888671875,
0.64013671875
],
[
0.53125,
0.6220703125,
0.90869140625,
0.6220703125,
0.90869140625,
0.6318359375,
0.53125,
0.6318359375
],
[
0.51123046875,
0.63330078125,
0.91259765625,
0.6337890625,
0.91259765625,
0.6484375,
0.51123046875,
0.64794921875
],
[
0.0849609375,
0.6396484375,
0.483642578125,
0.6396484375,
0.483642578125,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.08660888671875,
0.65380859375,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.08660888671875,
0.66748046875
],
[
0.51318359375,
0.6484375,
0.91259765625,
0.6484375,
0.91259765625,
0.662109375,
0.51318359375,
0.662109375
],
[
0.0849609375,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.0849609375,
0.681640625
],
[
0.51318359375,
0.66357421875,
0.91015625,
0.662109375,
0.91015625,
0.673828125,
0.51318359375,
0.6748046875
],
[
0.51318359375,
0.6748046875,
0.91015625,
0.6748046875,
0.91015625,
0.68798828125,
0.51318359375,
0.68798828125
],
[
0.08660888671875,
0.68115234375,
0.486083984375,
0.68115234375,
0.486083984375,
0.6943359375,
0.08660888671875,
0.6943359375
],
[
0.08416748046875,
0.693359375,
0.486083984375,
0.69580078125,
0.486083984375,
0.71044921875,
0.08416748046875,
0.70849609375
],
[
0.51123046875,
0.6875,
0.91259765625,
0.68798828125,
0.91259765625,
0.70263671875,
0.51123046875,
0.7021484375
],
[
0.51123046875,
0.70263671875,
0.91259765625,
0.7021484375,
0.91259765625,
0.71630859375,
0.51123046875,
0.71728515625
],
[
0.0849609375,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.0849609375,
0.7236328125
],
[
0.51318359375,
0.71728515625,
0.91259765625,
0.71728515625,
0.91259765625,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.0849609375,
0.7236328125,
0.48779296875,
0.7236328125,
0.48779296875,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.51318359375,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.744140625,
0.51318359375,
0.744140625
],
[
0.0849609375,
0.7373046875,
0.48779296875,
0.736328125,
0.48779296875,
0.75048828125,
0.0849609375,
0.751953125
],
[
0.51123046875,
0.744140625,
0.8935546875,
0.744140625,
0.8935546875,
0.75830078125,
0.51123046875,
0.75830078125
],
[
0.08660888671875,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.76513671875,
0.08660888671875,
0.76513671875
],
[
0.0849609375,
0.76513671875,
0.486083984375,
0.76513671875,
0.486083984375,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.515625,
0.7744140625,
0.529296875,
0.7744140625,
0.529296875,
0.7841796875,
0.515625,
0.7841796875
],
[
0.5390625,
0.77294921875,
0.70166015625,
0.77294921875,
0.70166015625,
0.787109375,
0.5390625,
0.787109375
],
[
0.08660888671875,
0.78076171875,
0.420654296875,
0.78076171875,
0.420654296875,
0.79150390625,
0.08660888671875,
0.79150390625
],
[
0.10296630859375,
0.79248046875,
0.48779296875,
0.79248046875,
0.48779296875,
0.80615234375,
0.10296630859375,
0.80615234375
],
[
0.51123046875,
0.79052734375,
0.91015625,
0.79150390625,
0.91015625,
0.80615234375,
0.51123046875,
0.8046875
],
[
0.0849609375,
0.806640625,
0.48779296875,
0.806640625,
0.48779296875,
0.82080078125,
0.0849609375,
0.82080078125
],
[
0.5107421875,
0.80615234375,
0.90869140625,
0.80615234375,
0.90869140625,
0.82080078125,
0.5107421875,
0.82080078125
],
[
0.08660888671875,
0.82080078125,
0.486083984375,
0.82080078125,
0.486083984375,
0.833984375,
0.08660888671875,
0.833984375
],
[
0.5107421875,
0.818359375,
0.9111328125,
0.8193359375,
0.9111328125,
0.833984375,
0.5107421875,
0.83349609375
],
[
0.08660888671875,
0.833984375,
0.48779296875,
0.833984375,
0.48779296875,
0.84765625,
0.08660888671875,
0.84765625
],
[
0.51123046875,
0.833984375,
0.91015625,
0.833984375,
0.91015625,
0.8486328125,
0.51123046875,
0.8486328125
],
[
0.08660888671875,
0.84765625,
0.488525390625,
0.84765625,
0.488525390625,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.5146484375,
0.849609375,
0.9111328125,
0.849609375,
0.9111328125,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.08660888671875,
0.86279296875,
0.48779296875,
0.86279296875,
0.48779296875,
0.87353515625,
0.08660888671875,
0.87353515625
],
[
0.51318359375,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.51318359375,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.48779296875,
0.87548828125,
0.48779296875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51123046875,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"Choosing Well Your Opponents:",
"How to Guide the Synthesis of Programmatic Strategies",
"Rubens O. Moraesi?, David S. Aleixo' , Lucas N. Ferreira? and Levi H. S. Lelis?",
"1 Departamento de Informatica, Universidade Federal de Vicosa, Brazil",
"2Department of Computing Science, University of Alberta, Canada",
"Alberta Machine Intelligence Institute (Amii)",
"rubens.moraes @ufv.br, david.aleixo@ufv.br, Inferrei@ualberta.ca, levi.lelis@ualberta.ca",
"",
"Abstract",
"strategy that IBR provides serves as a guiding function. In the",
"context of local search, when considering the neighbors of a",
"This paper introduces Local Learner (2L), an al-",
"candidate solution, local search algorithms prefer to accept a",
"gorithm forproviding a setof reference strategies",
"program that achieves a higher utility value against the target",
"[nf ",
"to guide the search for programmatic strategies in",
"strategy. Since IBR considers a single strategy as a target,",
"two-player zero-sum games. Previous learning al-",
"the search signal is often weak. This is because the neigh-",
"gorithms, such as Iterated Best Response (IBR),",
"",
"bors ofacandidatesolutionthatperformspoorlyagainstthe",
"Fictitious Play (FP), and Double-Oracle (DO), can",
"target strategy are also likely to perform poorly against it-",
"becomputationallyexpensiveormissimportantin-",
" small changes to a losing program will also generate a losing",
"formation for guiding search algorithms. 2L ac-",
"[T'S]",
"program. Moreover, IBR can loop around the strategy space",
"tivelyselectsa setofreferencestrategiestoimprove",
"in games with dynamics similar to Rock, Paper, and Scissors,",
"the search signal.We empiricallydemonstrate the",
"without making progress toward strong solutions.",
"advantages of our approach while guiding a local",
"In this paper, we adapt Fictitious Play (FP) [Brown, 1951]",
" search algorithm for synthesizing strategies in three",
"andDoubleOracle(DO)[McMahanetal.,2003]tothecon-",
"games,including MicroRTS,a challenging real-",
"text of programmatic strategies. FP and DO have been used",
"time strategy game. Results show that 2L learns",
"in the context of neural strategies to overcome some of the",
"800",
"reference strategiesthatprovideastrongersearch",
"weaknesses of IBR [Lanctot et al.,2017].Despite provid-",
"signal than IBR, FP, and DO. We also simulate a",
"ing a better search signal than IBR, we show that FP and DO",
"tournament of MicroRTS, where a synthesizer us-",
"can still fail to provide relevant information for the search.",
"ing 2L outperformed the winners of the two lat-",
"We then introduce a novel learning algorithm, Local Learner",
"est MicroRTS competitions, which were program-",
"(2L), that is designed specifically for guiding local search al-",
"matic strategies written by human programmers.",
"gorithms in the synthesis of programmatic strategies. 2L uses",
"information gathered while computing best responses to de-",
"Introduction",
"cide the set of target strategies to be used in future iterations",
"Programmatic strategies encode game strategies in human-",
"of the algorithm as a means of optimizing the search signal.",
"understandable programs.Such programmatic encoding al-",
"Weevaluate 2Lonthreetwo-playerzero-sumgames:Mi",
"croRTS [Ontanon et al., 2018], Poachers & Rangers, and",
"lows domain experts to interpret and modify computer-",
"generated strategies, which can be valuable depending on",
"Climbing Monkeys. The results show that 2L synthesized",
"the application domain (e.g., the games industry). Previous",
"strategiesthatareneverworseand oftenfar superiorto strate-",
"gies synthesized with IBR, FP, and DO in all three domains.",
"works have used Iterated-Best Response (IBR) [Lanctot et",
"al., 2017l as the learning algorithm for synthesizing program-",
"We also performed a simulated competition of MicroRTS",
"with strategies synthesized with 2L IBR, FP, DO, as well as",
"matic strategies [Marino et al., 2021]. Given a game, IBR",
"the programmatic strategies that won the last two MicroRTS",
"starts with an arbitrary strategy for playing the game and it",
"competitions, which were written by programmers. 2L ob-",
"approximates a best response to it; in the next iteration, it",
"tained the highest average winning rate in our tournament.",
"approximates a best response to the best response. This pro-",
"cess is repeated a number of iterations and the programmatic",
"2",
"ProblemDefinition",
"strategysynthesized in thelast iteration isreturned.",
"The computation of the best responses in the IBR loop",
"We consider the synthesis of programmatic strategies assum-",
"is performed by searching in the programmatic space de-",
"ing zero-sum two-player games G = (P,S, sinit,A, T,u)",
"fined by a domain-specific language. Given a target strat-",
"Let P = {i, -i] be the pair of players; S be the set of states,",
"egy, the algorithm searches for a program encoding a best",
"with Sinit in S being the initial state. Each player i can per",
"response to it. Previous work used local search algorithms",
"formanactionfromalegalsetofactionsA,(s)inAfora",
"forsearching intheprogrammatic space[Marinoetal.,202l;",
"given state s. The action of each player is given by a strategy.",
"Medeiros et al., 2022; Aleixo and Lelis, 2023]. The target",
"which is a function o, that receives a state s in S and returns"
] |
{
"arxiv_id": "2307.04893",
"language": "en",
"page_idx": 0
}
|
||
2307.04893-en-1
|
2307.04893-en-1.jpg
|
an action in $\mathcal{A}_i$ for $s.$ A transition function $\mathcal{T}$ receives a state and an action for each player and deterministically returns the next state of the game, which could be a terminal state, where the utility of each player is determined. The utility function $\mathcal{U}$ returns the value of the game in a given state (terminal or not). For $s$,the value of the game is denoted by $\mathcal{U}(s,\sigma_i,\sigma_{-i})$ when player $i$ follows the strategy $\sigma_i$ and player $-i,\sigma_-i.$ Considering that the game $G$ is zero-sum, the utility function for $-i$ is $- \mathcal{U} ( s, \sigma _{i}, \sigma _{- i}) . \textbf{ In this paper, we encode strategies for }\mathcal{G}$ as programs written in a domain-specific language (DSL).
A DSL can be defined as a context-free grammar $(M,\Omega,R,S)$, where $M,\Omega,R$, and $S$ are the sets of nonterminals, terminals, relations defining the production rules of the grammar, and the grammar's initial symbol, respectively. Figure 1 (right) shows an example of a DSL, where $M=\{S,C,B\},\Omega=\{c_{1},c_{2},b_{1},b_{2}$if,then$\},R$are the production rules (e.g.,$C\to c_1)$,and $S$ is the initial symbol.
<FigureHere>
Figure 1: AST for “if $b_1$ then $c_1”($left) and DSL in which the pro-
gram is written (right).
The DSL in Figure 1 allows programs with a single command (e.g., $c_1$ or $c_2)$ and programs with branching. We represent programs as abstract syntax trees (AST), where the root of the tree is $S$, the internal nodes are non-terminals, and the leaf nodes are terminals. Figure 1 (left) shows an example of an AST. We use a DSL $D$ to define the space of programs $[D]$,where each program $p\in[D]$ is a game strategy.
$\tilde{\mathrm{One~solves~the~problem~of~synthesizing~programmatic}}$
strategies by solving the following equation
$$\max_{\sigma_{i}\in[D]}\min_{\sigma_{-i}\in[D]}\mathcal{U}(s_{\mathrm{init}},\sigma_{i},\sigma_{-i})\:.$$
(1)
The strategies $\sigma_i$ and $\sigma_{-i}$ in $[D]$ able to solve Equation 1 define a Nash equilibrium profile in the programmatic space. We consider a programmatic variant of PSRO [Lanctot $\hat{e}t$ al., 2017] to approximate a solution to Equation 1.
# 3 Programmatic PSRO (PPSRO)
Let λbe a normal-form game defined by $(\Sigma,\mathcal{P},U_\Sigma)$, where $\Sigma=\left\{\sum_{i},\sum_{-i}\right\}$represents a set of strategies for each player in $\mathcal{P}=\{i,-i\}$, and $U_\Sigma \textbf{is the utility payoff table between}$ each pair of strategies in $\Sigma.$ A mixed strategy $\sigma$ is a probability distribution over strategies $\Sigma_i$ and $\Sigma_-i$ for players $i$ and $-i$, respectively. An empirical game of a normal-form game contains only a subset of the strategies of the original game.
Policy-Space Response Oracles (PSRO) is a framework for learning strategies that “grow”an empirical game[Lanctot $et$ $al.,20\tilde{17}].$ In PSRO, the empirical game starts with a single strategy in $\Sigma_i$ and $\Sigma_-i$ and it grows these sets by including a new strategy for each player in each iteration of the algorithm. Let a mixed strategy over the sets $\Sigma_i$ and $\Sigma_-i$ of the empirical game be called a meta-strategy. PSRO grows $\Sigma_i$ and $\hat{\Sigma}_-i$
<table>
<tbody>
<tr>
<td>gorithm 1 Programmatic PSRO</td>
</tr>
<tr>
<td>$\Psi$</td>
</tr>
<tr>
<td>$|niti$ WITh me STPTP $\rho arh$</td>
</tr>
<tr>
<td>Compute 11. $\tan r$ 111116</td>
</tr>
<tr>
<td>$w\mathbf{hil}_{\mathbf{e}}$ $hr$ sted budget do</td>
</tr>
<tr>
<td>$\textbf{for plaver }i$ i $\operatorname{in}\mathcal{P}$do</td>
</tr>
<tr>
<td>$\Psi/\Sigma//_{\Gamma}$ wifh</td>
</tr>
<tr>
<td>1 $\sigma.$ SPAT .</td>
</tr>
<tr>
<td>1 1 L</td>
</tr>
<tr>
<td>Comnute $\operatorname{from}\Sigma$ In</td>
</tr>
<tr>
<td>$\cdot\sigma_{i}$ rategv</td>
</tr>
</tbody>
</table>
by adding best responses to meta-strategies. Once a best response is added to a set, a new meta-strategy is computed, and the process is repeated. That is, given a meta-strategy $\sigma_{-i}$ (resp. $\sigma_i)$, for player $-i$ (resp. $i)$, the best response to $\sigma_{-i}\left(\mathrm{resp.~}\sigma_{i}\right)$is added to$\Sigma_i$ (resp. $\Sigma_-i).$
PSRO generalizes algorithms such as IBR, FP, and DO depending on how the meta-strategies are computed. Let $\sigma_{k}=(p_{1},p_{2},\cdots,p_{n})$ be a meta-strategy for player $k$ ($k$ can be either $i$ or$-i).$ Here$,p_j$ in $\sigma_k$ represents the probability in which $\sigma_k$ plays the $j$-th strategy added to the empirical game for player $\hat{k}.$ PSRO generalizes IBR if the meta-strategies are of the form $(0.0,0.\dot{0},\cdots,1.0)$,i.e., the only strategy in the support of the meta-strategy is the last strategy added to the empirical game. If the meta-strategy $\sigma_-i$ with $n$ strategies is of the form $(1/n,1/n,\cdots,1/n)$,i.e., all the previous strategies added to the game are played with equal probability, then PSRO generalizes FP. PSRO also generalizes DO [McMahan $etal.,\bar{2}003]$ when the meta-strategy is computed by solving the empirical game. We use a variant of PSRO, which we call Programmatic PSRO (PPSRO), to approximate a solution to Equation 1. PPSRO is shown Algorithm 1.
PPSRO starts by initializing the set of strategies, $\Sigma_i$ and $\Sigma_{-i}$, with two arbitrary strategies (line 1). PPSRO runs a number of iterations according to a given computational budget (e.g., the number of games played). In each iteration, PPSRO invokes a learning algorithm $\Psi$ (e.g., IBR) that receives the current empirical game and returns a meta-strategy $\sigma_{-i}$ (line 5). Then it searches in the programmatic space of strategies for a best response $\sigma_i^{\prime}$ to $\sigma_-i.$ We consider local search algorithms for computing $\sigma_i^{\prime}.$ The search algorithm, described in Section 4,initializes its computation with the last strategy added to the empirical game for $i$,which is denoted as $\sigma_i[-1]$ (line 6). The best response $\sigma_i^\prime$ is then added to $\Sigma_i.$ At the end, PPSRO returns the last meta-strategy $\sigma_i$ as an approximate solution for player $i$ to Equation 1(line 9).
The choice of meta-strategies across iterations of PPSRO determines how quickly it is able to approximate a Nash equilibrium profile for the game. Previous work investigated different approaches to define meta-strategies in the context of PSRO and neural policies [Lanctot et al., 2017; Anthony $etal.,2020;$ Muller $etal.,2020].$ However, searching in programmatic space is different from searching in neural space, since the former does not have a gradient signal to guide the search. As we show in our experiments, meta-
|
[
[
0.0849609375,
0.0694580078125,
0.486083984375,
0.0694580078125,
0.486083984375,
0.08331298828125,
0.0849609375,
0.08331298828125
],
[
0.5146484375,
0.0694580078125,
0.74169921875,
0.0694580078125,
0.74169921875,
0.08331298828125,
0.5146484375,
0.08331298828125
],
[
0.08660888671875,
0.085205078125,
0.486083984375,
0.085205078125,
0.486083984375,
0.0953369140625,
0.08660888671875,
0.0953369140625
],
[
0.51318359375,
0.08837890625,
0.84326171875,
0.08837890625,
0.84326171875,
0.1029052734375,
0.51318359375,
0.1029052734375
],
[
0.0849609375,
0.09783935546875,
0.486083984375,
0.09783935546875,
0.486083984375,
0.111083984375,
0.0849609375,
0.111083984375
],
[
0.51318359375,
0.1029052734375,
0.73388671875,
0.1029052734375,
0.73388671875,
0.11614990234375,
0.51318359375,
0.11614990234375
],
[
0.0882568359375,
0.11297607421875,
0.484375,
0.11297607421875,
0.484375,
0.12371826171875,
0.0882568359375,
0.12371826171875
],
[
0.51953125,
0.11553955078125,
0.83251953125,
0.11614990234375,
0.83251953125,
0.1307373046875,
0.51953125,
0.1300048828125
],
[
0.08660888671875,
0.126953125,
0.484375,
0.126953125,
0.484375,
0.136962890625,
0.08660888671875,
0.136962890625
],
[
0.51953125,
0.1307373046875,
0.765625,
0.1307373046875,
0.765625,
0.14453125,
0.51953125,
0.14453125
],
[
0.08660888671875,
0.1395263671875,
0.486083984375,
0.1395263671875,
0.486083984375,
0.15283203125,
0.08660888671875,
0.15283203125
],
[
0.5185546875,
0.14453125,
0.77783203125,
0.14453125,
0.77783203125,
0.1578369140625,
0.5185546875,
0.1578369140625
],
[
0.08660888671875,
0.1514892578125,
0.483642578125,
0.1514892578125,
0.483642578125,
0.1654052734375,
0.08660888671875,
0.1654052734375
],
[
0.521484375,
0.160400390625,
0.53515625,
0.160400390625,
0.53515625,
0.169189453125,
0.521484375,
0.169189453125
],
[
0.5537109375,
0.1578369140625,
0.69140625,
0.1578369140625,
0.69140625,
0.1717529296875,
0.5537109375,
0.1717529296875
],
[
0.08660888671875,
0.1666259765625,
0.486083984375,
0.1666259765625,
0.486083984375,
0.179931640625,
0.08660888671875,
0.179931640625
],
[
0.521484375,
0.173583984375,
0.53515625,
0.173583984375,
0.53515625,
0.1849365234375,
0.521484375,
0.1849365234375
],
[
0.5703125,
0.1717529296875,
0.86767578125,
0.17041015625,
0.86767578125,
0.1849365234375,
0.5703125,
0.186279296875
],
[
0.0849609375,
0.18115234375,
0.484375,
0.18115234375,
0.484375,
0.1944580078125,
0.0849609375,
0.1944580078125
],
[
0.521484375,
0.1881103515625,
0.5361328125,
0.1881103515625,
0.5361328125,
0.1988525390625,
0.521484375,
0.1988525390625
],
[
0.57275390625,
0.1849365234375,
0.74609375,
0.1849365234375,
0.74609375,
0.199462890625,
0.57275390625,
0.199462890625
],
[
0.0849609375,
0.1944580078125,
0.465576171875,
0.1932373046875,
0.465576171875,
0.20703125,
0.0849609375,
0.2083740234375
],
[
0.521484375,
0.2008056640625,
0.5361328125,
0.2008056640625,
0.5361328125,
0.2108154296875,
0.521484375,
0.2108154296875
],
[
0.57177734375,
0.1976318359375,
0.69140625,
0.199462890625,
0.69140625,
0.2139892578125,
0.57177734375,
0.212158203125
],
[
0.10296630859375,
0.2083740234375,
0.484375,
0.2083740234375,
0.484375,
0.2215576171875,
0.10296630859375,
0.2215576171875
],
[
0.521484375,
0.21533203125,
0.53369140625,
0.21533203125,
0.53369140625,
0.2247314453125,
0.521484375,
0.2247314453125
],
[
0.57177734375,
0.212158203125,
0.77685546875,
0.2108154296875,
0.77685546875,
0.2259521484375,
0.57177734375,
0.2266845703125
],
[
0.0882568359375,
0.2215576171875,
0.484375,
0.2215576171875,
0.484375,
0.23486328125,
0.0882568359375,
0.23486328125
],
[
0.5185546875,
0.2259521484375,
0.736328125,
0.2279052734375,
0.736328125,
0.2412109375,
0.5185546875,
0.2392578125
],
[
0.0882568359375,
0.2374267578125,
0.484375,
0.2374267578125,
0.484375,
0.2474365234375,
0.0882568359375,
0.2474365234375
],
[
0.0849609375,
0.25,
0.486083984375,
0.25,
0.486083984375,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.08416748046875,
0.261962890625,
0.48779296875,
0.26318359375,
0.48779296875,
0.27783203125,
0.08416748046875,
0.276611328125
],
[
0.0882568359375,
0.27783203125,
0.486083984375,
0.27783203125,
0.486083984375,
0.291748046875,
0.0882568359375,
0.291748046875
],
[
0.5146484375,
0.274658203125,
0.90673828125,
0.274658203125,
0.90673828125,
0.285400390625,
0.5146484375,
0.285400390625
],
[
0.08416748046875,
0.290283203125,
0.45751953125,
0.291748046875,
0.45751953125,
0.30615234375,
0.08416748046875,
0.304931640625
],
[
0.51123046875,
0.287353515625,
0.9111328125,
0.287353515625,
0.9111328125,
0.300537109375,
0.51123046875,
0.300537109375
],
[
0.51318359375,
0.300537109375,
0.9111328125,
0.300537109375,
0.9111328125,
0.314453125,
0.51318359375,
0.314453125
],
[
0.51318359375,
0.314453125,
0.9140625,
0.314453125,
0.9140625,
0.32763671875,
0.51318359375,
0.32763671875
],
[
0.5107421875,
0.32763671875,
0.78369140625,
0.32763671875,
0.78369140625,
0.341552734375,
0.5107421875,
0.341552734375
],
[
0.279296875,
0.3359375,
0.4306640625,
0.3359375,
0.4306640625,
0.35107421875,
0.279296875,
0.35107421875
],
[
0.52685546875,
0.343505859375,
0.91259765625,
0.34228515625,
0.91259765625,
0.356689453125,
0.52685546875,
0.35791015625
],
[
0.27685546875,
0.3515625,
0.359375,
0.354736328125,
0.358642578125,
0.37060546875,
0.275390625,
0.367431640625
],
[
0.51318359375,
0.359130859375,
0.9140625,
0.359130859375,
0.9140625,
0.37255859375,
0.51318359375,
0.37255859375
],
[
0.27685546875,
0.36865234375,
0.359375,
0.371826171875,
0.358642578125,
0.386962890625,
0.275390625,
0.3837890625
],
[
0.51123046875,
0.37255859375,
0.91259765625,
0.37060546875,
0.91259765625,
0.385009765625,
0.51123046875,
0.386962890625
],
[
0.1666259765625,
0.3818359375,
0.178955078125,
0.3818359375,
0.178955078125,
0.386474609375,
0.1666259765625,
0.386474609375
],
[
0.51123046875,
0.385009765625,
0.91259765625,
0.386474609375,
0.91259765625,
0.40087890625,
0.51123046875,
0.399658203125
],
[
0.08660888671875,
0.404052734375,
0.486083984375,
0.404052734375,
0.486083984375,
0.417236328125,
0.08660888671875,
0.417236328125
],
[
0.51123046875,
0.40087890625,
0.91259765625,
0.40087890625,
0.91259765625,
0.4140625,
0.51123046875,
0.4140625
],
[
0.0849609375,
0.417236328125,
0.221435546875,
0.416748046875,
0.221435546875,
0.429931640625,
0.0849609375,
0.43115234375
],
[
0.51123046875,
0.4140625,
0.9111328125,
0.4140625,
0.9111328125,
0.427978515625,
0.51123046875,
0.427978515625
],
[
0.51318359375,
0.427978515625,
0.9140625,
0.427978515625,
0.9140625,
0.44140625,
0.51318359375,
0.44140625
],
[
0.101318359375,
0.443115234375,
0.486083984375,
0.4443359375,
0.486083984375,
0.458984375,
0.101318359375,
0.458251953125
],
[
0.51123046875,
0.442626953125,
0.9140625,
0.442626953125,
0.9140625,
0.455810546875,
0.51123046875,
0.455810546875
],
[
0.0849609375,
0.460205078125,
0.486083984375,
0.460205078125,
0.486083984375,
0.473388671875,
0.0849609375,
0.473388671875
],
[
0.5107421875,
0.455810546875,
0.91259765625,
0.455810546875,
0.91259765625,
0.4697265625,
0.5107421875,
0.4697265625
],
[
0.0849609375,
0.473388671875,
0.488525390625,
0.473388671875,
0.488525390625,
0.4873046875,
0.0849609375,
0.4873046875
],
[
0.51318359375,
0.4697265625,
0.9111328125,
0.4697265625,
0.9111328125,
0.48291015625,
0.51318359375,
0.48291015625
],
[
0.08660888671875,
0.488037109375,
0.48779296875,
0.488037109375,
0.48779296875,
0.498779296875,
0.08660888671875,
0.498779296875
],
[
0.515625,
0.48486328125,
0.9111328125,
0.48486328125,
0.9111328125,
0.49560546875,
0.515625,
0.49560546875
],
[
0.08660888671875,
0.501953125,
0.48779296875,
0.501953125,
0.48779296875,
0.51513671875,
0.08660888671875,
0.51513671875
],
[
0.51318359375,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.08660888671875,
0.51513671875,
0.488525390625,
0.51513671875,
0.488525390625,
0.52880859375,
0.08660888671875,
0.52880859375
],
[
0.51318359375,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51318359375,
0.5234375
],
[
0.0882568359375,
0.52880859375,
0.438720703125,
0.52880859375,
0.438720703125,
0.54248046875,
0.0882568359375,
0.54248046875
],
[
0.5146484375,
0.52587890625,
0.91259765625,
0.52587890625,
0.91259765625,
0.5361328125,
0.5146484375,
0.5361328125
],
[
0.10296630859375,
0.54248046875,
0.486083984375,
0.54248046875,
0.486083984375,
0.55615234375,
0.10296630859375,
0.55615234375
],
[
0.51318359375,
0.53857421875,
0.9150390625,
0.53857421875,
0.9150390625,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.08416748046875,
0.5546875,
0.374267578125,
0.55615234375,
0.374267578125,
0.57080078125,
0.08416748046875,
0.5693359375
],
[
0.51318359375,
0.5517578125,
0.79345703125,
0.5517578125,
0.79345703125,
0.56494140625,
0.51318359375,
0.56494140625
],
[
0.52783203125,
0.568359375,
0.9140625,
0.568359375,
0.9140625,
0.58203125,
0.52783203125,
0.58203125
],
[
0.1854248046875,
0.5771484375,
0.21728515625,
0.5791015625,
0.2156982421875,
0.59033203125,
0.1846923828125,
0.58837890625
],
[
0.2279052734375,
0.5771484375,
0.390625,
0.5771484375,
0.390625,
0.59130859375,
0.2279052734375,
0.59130859375
],
[
0.464111328125,
0.58203125,
0.48779296875,
0.58203125,
0.48779296875,
0.5947265625,
0.464111328125,
0.5947265625
],
[
0.51123046875,
0.58056640625,
0.9140625,
0.58203125,
0.9140625,
0.5966796875,
0.51123046875,
0.59521484375
],
[
0.51123046875,
0.59521484375,
0.9111328125,
0.59521484375,
0.9111328125,
0.609375,
0.51123046875,
0.609375
],
[
0.178955078125,
0.5859375,
0.27783203125,
0.5859375,
0.27783203125,
0.59765625,
0.178955078125,
0.59765625
],
[
0.10296630859375,
0.60595703125,
0.486083984375,
0.60595703125,
0.486083984375,
0.619140625,
0.10296630859375,
0.619140625
],
[
0.515625,
0.6123046875,
0.9111328125,
0.6123046875,
0.9111328125,
0.62255859375,
0.515625,
0.62255859375
],
[
0.0849609375,
0.619140625,
0.486083984375,
0.62060546875,
0.486083984375,
0.6337890625,
0.0849609375,
0.63330078125
],
[
0.51123046875,
0.62255859375,
0.9111328125,
0.62353515625,
0.9111328125,
0.63818359375,
0.51123046875,
0.63720703125
],
[
0.0849609375,
0.6318359375,
0.486083984375,
0.63330078125,
0.486083984375,
0.64794921875,
0.0849609375,
0.646484375
],
[
0.5107421875,
0.63720703125,
0.9111328125,
0.63720703125,
0.9111328125,
0.65087890625,
0.5107421875,
0.65087890625
],
[
0.0849609375,
0.64794921875,
0.393798828125,
0.64794921875,
0.393798828125,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.51318359375,
0.6513671875,
0.91259765625,
0.6513671875,
0.91259765625,
0.66552734375,
0.51318359375,
0.66552734375
],
[
0.51123046875,
0.66552734375,
0.91259765625,
0.66552734375,
0.91259765625,
0.6787109375,
0.51123046875,
0.6787109375
],
[
0.08660888671875,
0.67529296875,
0.375732421875,
0.67529296875,
0.375732421875,
0.68994140625,
0.08660888671875,
0.68994140625
],
[
0.51123046875,
0.6796875,
0.91259765625,
0.6796875,
0.91259765625,
0.693359375,
0.51123046875,
0.693359375
],
[
0.08416748046875,
0.6943359375,
0.48779296875,
0.693359375,
0.48779296875,
0.70849609375,
0.08416748046875,
0.708984375
],
[
0.5146484375,
0.6943359375,
0.9111328125,
0.6943359375,
0.9111328125,
0.705078125,
0.5146484375,
0.705078125
],
[
0.0849609375,
0.70849609375,
0.486083984375,
0.708984375,
0.486083984375,
0.7236328125,
0.0849609375,
0.72265625
],
[
0.51123046875,
0.70703125,
0.9111328125,
0.70703125,
0.9111328125,
0.72021484375,
0.51123046875,
0.72021484375
],
[
0.08416748046875,
0.7216796875,
0.48779296875,
0.72265625,
0.48779296875,
0.7373046875,
0.08416748046875,
0.736328125
],
[
0.5146484375,
0.72265625,
0.9111328125,
0.72265625,
0.9111328125,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.0849609375,
0.7373046875,
0.486083984375,
0.7373046875,
0.486083984375,
0.75048828125,
0.0849609375,
0.75048828125
],
[
0.51123046875,
0.73291015625,
0.91259765625,
0.734375,
0.91259765625,
0.74853515625,
0.51123046875,
0.74755859375
],
[
0.0849609375,
0.75048828125,
0.48779296875,
0.75048828125,
0.48779296875,
0.7646484375,
0.0849609375,
0.7646484375
],
[
0.51318359375,
0.74853515625,
0.89453125,
0.74853515625,
0.89453125,
0.76220703125,
0.51318359375,
0.76220703125
],
[
0.08905029296875,
0.7646484375,
0.486083984375,
0.7646484375,
0.486083984375,
0.77783203125,
0.08905029296875,
0.77783203125
],
[
0.52783203125,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.77783203125,
0.52783203125,
0.77783203125
],
[
0.08416748046875,
0.77783203125,
0.48046875,
0.77880859375,
0.48046875,
0.79248046875,
0.08416748046875,
0.79150390625
],
[
0.51318359375,
0.77880859375,
0.91259765625,
0.77880859375,
0.91259765625,
0.79248046875,
0.51318359375,
0.79248046875
],
[
0.101318359375,
0.79248046875,
0.488525390625,
0.79248046875,
0.488525390625,
0.80615234375,
0.101318359375,
0.80615234375
],
[
0.51123046875,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.51123046875,
0.80615234375
],
[
0.0849609375,
0.80615234375,
0.488525390625,
0.80615234375,
0.488525390625,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.51318359375,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.51318359375,
0.8193359375
],
[
0.08416748046875,
0.818359375,
0.488525390625,
0.8193359375,
0.488525390625,
0.833984375,
0.08416748046875,
0.83349609375
],
[
0.51318359375,
0.8212890625,
0.91259765625,
0.8212890625,
0.91259765625,
0.83203125,
0.51318359375,
0.83203125
],
[
0.0849609375,
0.833984375,
0.488525390625,
0.833984375,
0.488525390625,
0.84765625,
0.0849609375,
0.84765625
],
[
0.51318359375,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.51318359375,
0.84765625
],
[
0.0849609375,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.0849609375,
0.861328125
],
[
0.51123046875,
0.84765625,
0.9111328125,
0.84765625,
0.9111328125,
0.861328125,
0.51123046875,
0.861328125
],
[
0.0882568359375,
0.86279296875,
0.484375,
0.86279296875,
0.484375,
0.87353515625,
0.0882568359375,
0.87353515625
],
[
0.51123046875,
0.8623046875,
0.9140625,
0.8623046875,
0.9140625,
0.87548828125,
0.51123046875,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.48779296875,
0.87548828125,
0.48779296875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51123046875,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"an action in A, for s. A transition function T receives a state",
"Algorithm 1 Programmatic PSRO",
"andanactionforeachplayeranddeterministicallyreturnsthe",
"Require: Game G, DSL D, learning algorithm ",
"next state of the game, which could be a terminal state, where",
"Ensure: Strategy o, for player i.",
"theutilityofeachplayer is determined.The utilityfunctionu",
"I: Initialize , and with one strategy each",
"returns the value of thegame inagiven state(terminal or not).",
"2: Compute utilities Us for (αi, ot)",
"For s, the value of the game is denoted by U(s, oi, g_i) when",
"3: while have not exhausted budget do",
"player i follows the strategy g, and player -i, g-i. Consid-",
"4:",
"for player i in P do",
"ering that the game G is zero-sum, the utility function for -i",
"5:",
"Compute a meta-strategy g-i with (2, U)",
"is -Uu(s, oi, _). In this paper, we encode strategies for g",
"6:",
"o, search(g;-1], α_i)",
"as programs written in a domain-specific language (DSL).",
"7:",
"U(a)",
"A DSL can be defined as a context-free grammar",
"8:",
"Compute entries in Us from",
"(M,, R, S), where M, 2, R, and S are the sets of non-",
"9: return Last meta-strategy i",
"terminals,terminals,relations defining the production rules",
"of the grammar, and the grammar's initial symbol, respec-",
"tively. Figure 1 (right) shows an example of a DSL, where",
"M = {S, C, B], = {c1, C2, bi, b2 if, then], R are the pro-",
"byaddingbestresponsestometa-strategies.Oncea bestre",
"duction rules (e.g., C → ci), and S is the initial symbol.",
" sponse is added to a set, a new meta-strategy is computed,",
"and the process is repeated. That is, given a meta-strategy",
"-, (resp. o,), for player -i (resp. i), the best response to",
"g-, (resp. g) is added to , (resp. _).",
"S → C |if(B) then C",
"PSRO generalizes algorithms such as IBR, FP, and DO",
"C→ Ci / C2",
"depending on how the meta-strategies are computed. Let",
"B → bi /b2",
"Ok = (pi, P2, ... , Pn) be a meta-strategy for player k (k can",
"01",
"be either i or -i). Here, P, in ok represents the probability in",
"Figure 1: AST for if br then cr\" (left) and DSL in which the pro-",
"which ok plays the j-th strategy added to the empirical game",
"gram is written (right).",
"for player k. PSRO generalizes IBR if the meta-strategies are",
"of the form (0.0, 0.0, -.., 1.0), i.e., the only strategy in the",
"The DSL in Figure 1 allows programs with a single com-",
"support of the meta-strategy is the last strategy added to the",
"mand (e.g., Ci or c2) and programs with branching. We repre-",
"empirical game. If the meta-strategy o-i with n strategies is",
" sent programs as abstract syntax trees (AST), where the root",
"of the form (1/n, 1/n, ..., 1/n), i.e., all the previous strate-",
"of thetreeisS.theinternal nodes arenon-terminals, and the",
"gies addedtothegameareplayed withequalprobability,then",
"leaf nodes are terminals. Figure 1 (left) shows an example",
"PSRO generalizes FP. PSRO also generalizes DO [McMahan",
"of an AST. We use a DSL D to define the space of programs",
"et al., 2003] when the meta-strategy is computed by solving",
"[D], where each program p E [D] is a game strategy.",
"theempiricalgame.WeuseavariantofPSRO,whichwecall",
"One solves the problem of synthesizing programmatic",
"Programmatic PSRO (PPSRO), to approximate a solution to",
"strategies by solving the following equation",
"Equation 1. PPSRO is shown Algorithm 1.",
"PPSRO starts by initializing the set of strategies, E, and",
"max",
"minu(sinit,Qi,g_i)",
"(1)",
"_i, with two arbitrary strategies (line 1).PPSRO runs a",
"number of iterations according to a given computational bud-",
"GE[DI .-IED]",
"The strategies , and g- in [D able to solve Equation 1",
"get (e.g.,the number of games played).Ineach iteration",
"define a Nash equilibrium profile in the programmatic space.",
"PPSRO invokes a learning algorithm (e.g., IBR) that re-",
"We consider a programmatic variant of PSRO [Lanctot et al.,",
"ceives the current empirical game and returns a meta-strategy",
"2017] to approximate a solution to Equation 1.",
"-i (line 5). Then it searches in the programmatic space of",
"strategies for a best response o, to _i. We consider local",
"3ProgrammaticPSRO (PPSRO)",
"search algorithms for computing os. The search algorithm,",
"Let > be a normal-form game defined by (, P,U), where",
"described in Section 4,initializes its computation with the",
" = {,, _] represents a set of strategies for each player",
"last strategy added to the empirical game for i, which is de-",
"in P ={i, -ij, and U is the utility payoff table between",
"noted as,-1(line6).Thebestresponseo,isthenadded",
"each pair of strategies in . A mixed strategy o is a probabil-",
"to E. At the end, PPSRO returns the last meta-strategy i as",
"ity distribution over strategies , and _, for players i and",
"an approximate solution for player i to Equation 1 (line 9)",
"-i, respectively. An empirical game of a normal-form game",
"The choice of meta-strategies across iterations of PPSRO",
"contains only a subset of the strategies of the original game.",
"determines how quickly it is able to approximate a Nash",
"Policy-Space Response Oracles (PSRO) is a framework for",
"equilibrium profile for the game. Previous work investi-",
"learning strategies that \"grow\" an empirical game [Lanctot et",
"gated different approaches to define meta-strategies in the",
"al., 20i7]. In PSRO, the empirical game starts with a single",
"context of PSRO and neural policies [Lanctot et al., 2017;",
"strategy in Z, and _i and it grows these sets by including a",
"Anthony et al., 2020; Muller et al., 20201. However, search-",
"new strategy for each player in each iteration of the algorithm.",
"ing in programmatic space is different from searching in neu-",
"Let a mixed strategy over the sets Z, and Z-, of the empiri-",
"ral space, since the former does not have a gradient signal",
"cal game be called a meta-strategy. PSRO grows E; and _z",
"to guide the search. As we show in our experiments, meta-"
] |
{
"arxiv_id": "2307.04893",
"language": "en",
"page_idx": 1
}
|
||
2307.04893-en-4
|
2307.04893-en-4.jpg
|
(a) Compute a best response $\sigma$ to $\sigma_-k$ and add it to $\Sigma_k$
and to $\Sigma_{\sigma_{k}};$set $\sigma_{k}[j]=1.0/|\Sigma_{\sigma_{k}}|$ for all $\sigma_{j}$ in $\Sigma_{\sigma_{k}}.$ ( b) Set $\sigma _{- k}[ j] = 0$ and remove it from $\Sigma_\sigma_{-k}$ all $\sigma_{j,-k}$
in $\Sigma_{-k}$ that were estimated as redundant.
(c) Set $\sigma_-k[j]=1.0/|\Sigma_{\sigma_{-k}}|$ for all $\sigma_j$ in $\Sigma_\sigma_{-k}.$
2L starts by initializing the set of strategies of the empirical game and the set of strategies in the support of the metastrategy with an arbitrary strategy for one of the players (-i in the pseudocode above). Then, it computes a best response to this arbitrary strategy and uses the best response to initialize $\sum_{i}$ and$\sum_\sigma_{i}.$The meta-strategies are of the form (1.0) because the empirical game has a single strategy for each player (see Step 1 above). Step 2 refers to PPSRO's loop, where it computes the best responses while alternating the players. Once a best response $\sigma$ is computed to strategy $\sigma_-k$, it is added to the support of $\sigma_k$ with uniform probability (see Step 2a).
2L estimates which strategies in the support of $\sigma_{-k}$ are redundant while computing the best response $\sigma$ to $\sigma_-k.$ In Step 2b, 2L removes the redundant strategies from the support of $\sigma_{-k}$ and, in Step 2c, redistributes the probabilities so $\hat{\text{that each strategy in the support has the same probability}}.$
Example 5. In Example 2, we showed that DO fails to include both attack[1,2,3] and attack[1,2,5] in the support of the meta-strategy $\sigma_{-i}$, thus missing the guidance information attack[1,2,3] provides. Once the strategy attack[l,2,5] is added to the empirical game, the metastrategy will automatically have both attack[1,2,3] and attack[l,2,5]inits support.Incontrast with DO, 2L retains both strategies in the support of $\sigma _{- i}\textit{for the next it- }$ eration as long as strategies such as defend[1,2,3] and defend[ 1, 2, 5] are evaluated in search as both attack[ 1, $2,3]and$ attack$[ 1, 2, 5] will\textit{beflagged as helpful. }$
A weakness of 2L as presented above is that it can flag as redundant a strategy that is helpful if it does not sample
enough strategies in the search. For example, if the metastrategy for Rangers has both defend[1] and defend[2] in its support, but it never evaluates a strategy that attacks gate 1 in search, then defend[1] will mistakenly be removed from the meta-strategy's support. We implement the following enhancement to fix this weakness. Whenever the search returns a best response $\sigma$ to a meta-strategy $\sigma_-i$ (resp. $\sigma_i)$, we evaluate $\sigma$ against all strategies in the empirical game, including $\begin{array}{l}\text{those not in the support of }\sigma_{-i}\text{(resp. }\sigma_{i}).\text{If there is a strategy}\\{\sigma^{\prime}\text{ in the empirical game that is a best response to }\sigma,\text{then it}}\end{array}$ must be that 2L mistakenly removed $\sigma^{\prime}$ from the support of the meta-strategy. In this case, we repeat the search for a best response with $\sigma^{\prime}$ added to the support of the meta-strategy.
This enhancement can increase the number of times the search algorithm is invoked in each iteration of the PPSRO loop. While we perform a single search per iteration with IBR, FP, and DO, in the worst case, 2L can perform a number of searches equal to the number of strategies in the game. This is because, in the worst case, we add all strategies of the empirical game to the support of the meta-strategy. Despite the possible additional searches, preliminary experiments showed that this enhancement improves the sampling efficiency of 2L. All results in this paper use this enhancement.
In practice, we do not have the guarantee that the search algorithm used in PPSRO's main loop is able to return a best response to a meta-strategy. So we use whichever approximation the search returns as if it was a best response to the meta-strategy. Moreover, depending on the game, we might not be able to immediately recognize a best response to strategy once we see one, as one would have to prove the strategy to be a best response. This could be problematic, for example, when implementing the enhancement that 2L re-runs the search if there is a strategy in the empirical game that is a best response to the strategy the search returns. We run our experiments in games with utilities of -1,0,+1. If a best response cannot be easily verified (e.g., MicroRTS), then we consider that $\sigma$ is a best response to $\sigma^\prime$ if $\mathcal{U} ( s_\mathrm{init}, \sigma , \sigma ^{\prime }) = + 1.$
Once 2L reaches a computational budget, it can return different strategies as its approximate solution to Equation 1. Similarly to IBR, it can return the last strategy added to the empirical game for each player. 2L can also return a mixed strategy that is given by the distribution of strategies added to the empirical game, as does FP. We can also solve the resulting empirical game with linear programming, like DO does, and return the resulting strategy. In this paper, we assume the games have a pure dominant strategy for which IBR's approach of returning the last strategy added to the empirical game is suitable; this is what we use in our experiments.
7 $\textbf{Empirical Evaluation}$
# 7.1 Problem Domains
In addition to P&R, we introduce Climbing Monkey (CM) another two-player zero-sum game with a trivial optimal strategy that is also challenging in the context of programmatic strategies. In CM, monkeys need to climb to a branch of a tree that is higher than the branch the opponent's monkey is able to reach. The branches need to be climbed one at a time, without skipping any branch. The monkey that climbs to a higher branch wins the game. The game ends in a draw if both monkeys climb to a branch of the same height. For a tree with $n$ branches, a dominant programmatic strategy is climb[1],climb$[2],\cdots$,climb$[n].$Similarly to P&R, CM is challenging because, depending on the number of branches, it requires one to synthesize long programs.
In P&R, learning algorithms perform better if using a larger number of strategies in the support of meta-strategies as having many strategies helps Rangers converge to a strategy that protects all gates. CM is a game where all one needs to use is the last strategy added to the empirical game, i.e., the strategy that allows the monkey to climb to the highest branch. We hypothesize that 2L is capable of detecting which strategies are needed in the support of the meta-strategies for these two games.
We also evaluate 2L in MicroRTS, a real-time strategy game designed for research. There is an active research community that uses MicroRTS as a benchmark to evaluate intelligent systems.$^{1}$ MicroRTS is a game played with realtime constraints and very large action and state spaces [Lelis, 2021]. Each player can control two types of stationary units
$^{1}$https://github.com/Farama-Foundation/MicroRTS/wiki
|
[
[
0.1258544921875,
0.0694580078125,
0.483642578125,
0.0694580078125,
0.483642578125,
0.08331298828125,
0.1258544921875,
0.08331298828125
],
[
0.529296875,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.529296875,
0.08331298828125
],
[
0.1502685546875,
0.083984375,
0.486083984375,
0.083984375,
0.486083984375,
0.0985107421875,
0.1502685546875,
0.0985107421875
],
[
0.51318359375,
0.083984375,
0.91259765625,
0.083984375,
0.91259765625,
0.09783935546875,
0.51318359375,
0.09783935546875
],
[
0.1217041015625,
0.09661865234375,
0.48779296875,
0.0985107421875,
0.48779296875,
0.11614990234375,
0.1217041015625,
0.1142578125
],
[
0.51318359375,
0.0985107421875,
0.9111328125,
0.0985107421875,
0.9111328125,
0.11236572265625,
0.51318359375,
0.11236572265625
],
[
0.1502685546875,
0.11297607421875,
0.42236328125,
0.11297607421875,
0.42236328125,
0.126953125,
0.1502685546875,
0.126953125
],
[
0.5146484375,
0.11297607421875,
0.91259765625,
0.11297607421875,
0.91259765625,
0.12371826171875,
0.5146484375,
0.12371826171875
],
[
0.125,
0.1300048828125,
0.435546875,
0.1300048828125,
0.435546875,
0.143310546875,
0.125,
0.143310546875
],
[
0.51318359375,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.51318359375,
0.14013671875,
0.90869140625,
0.14013671875,
0.90869140625,
0.15087890625,
0.51318359375,
0.15087890625
],
[
0.08660888671875,
0.1483154296875,
0.484375,
0.1483154296875,
0.484375,
0.1622314453125,
0.08660888671875,
0.1622314453125
],
[
0.51318359375,
0.15283203125,
0.91015625,
0.154052734375,
0.91015625,
0.1654052734375,
0.51318359375,
0.1641845703125
],
[
0.08660888671875,
0.1634521484375,
0.484375,
0.1634521484375,
0.484375,
0.173583984375,
0.08660888671875,
0.173583984375
],
[
0.5146484375,
0.1685791015625,
0.90869140625,
0.1685791015625,
0.90869140625,
0.1787109375,
0.5146484375,
0.1787109375
],
[
0.08660888671875,
0.175537109375,
0.486083984375,
0.175537109375,
0.486083984375,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.51318359375,
0.18115234375,
0.91259765625,
0.18115234375,
0.91259765625,
0.1944580078125,
0.51318359375,
0.1944580078125
],
[
0.08660888671875,
0.1881103515625,
0.486083984375,
0.1881103515625,
0.486083984375,
0.2020263671875,
0.08660888671875,
0.2020263671875
],
[
0.51318359375,
0.1944580078125,
0.91259765625,
0.1944580078125,
0.91259765625,
0.2083740234375,
0.51318359375,
0.2083740234375
],
[
0.08905029296875,
0.2052001953125,
0.484375,
0.2052001953125,
0.484375,
0.21533203125,
0.08905029296875,
0.21533203125
],
[
0.51318359375,
0.2083740234375,
0.9111328125,
0.2083740234375,
0.9111328125,
0.2215576171875,
0.51318359375,
0.2215576171875
],
[
0.08660888671875,
0.2171630859375,
0.48779296875,
0.2171630859375,
0.48779296875,
0.2310791015625,
0.08660888671875,
0.2310791015625
],
[
0.51318359375,
0.222900390625,
0.91259765625,
0.222900390625,
0.91259765625,
0.236083984375,
0.51318359375,
0.236083984375
],
[
0.08660888671875,
0.2310791015625,
0.486083984375,
0.2310791015625,
0.486083984375,
0.2442626953125,
0.08660888671875,
0.2442626953125
],
[
0.5146484375,
0.2374267578125,
0.9111328125,
0.2374267578125,
0.9111328125,
0.2474365234375,
0.5146484375,
0.2474365234375
],
[
0.0849609375,
0.2442626953125,
0.484375,
0.2442626953125,
0.484375,
0.25830078125,
0.0849609375,
0.25830078125
],
[
0.51123046875,
0.248779296875,
0.85546875,
0.25,
0.85546875,
0.264404296875,
0.51123046875,
0.26318359375
],
[
0.08660888671875,
0.2587890625,
0.48779296875,
0.2587890625,
0.48779296875,
0.272705078125,
0.08660888671875,
0.272705078125
],
[
0.52783203125,
0.26318359375,
0.91259765625,
0.26318359375,
0.91259765625,
0.276611328125,
0.52783203125,
0.276611328125
],
[
0.08660888671875,
0.272705078125,
0.486083984375,
0.272705078125,
0.486083984375,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.5146484375,
0.27783203125,
0.91259765625,
0.27783203125,
0.91259765625,
0.291748046875,
0.5146484375,
0.291748046875
],
[
0.0882568359375,
0.28857421875,
0.45751953125,
0.28857421875,
0.45751953125,
0.298583984375,
0.0882568359375,
0.298583984375
],
[
0.51318359375,
0.291748046875,
0.91259765625,
0.291748046875,
0.91259765625,
0.304931640625,
0.51318359375,
0.304931640625
],
[
0.10296630859375,
0.300537109375,
0.486083984375,
0.300537109375,
0.486083984375,
0.314453125,
0.10296630859375,
0.314453125
],
[
0.51318359375,
0.30615234375,
0.9140625,
0.30615234375,
0.9140625,
0.3193359375,
0.51318359375,
0.3193359375
],
[
0.08660888671875,
0.314453125,
0.48779296875,
0.314453125,
0.48779296875,
0.32763671875,
0.08660888671875,
0.32763671875
],
[
0.5146484375,
0.32080078125,
0.9111328125,
0.32080078125,
0.9111328125,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.0882568359375,
0.330078125,
0.484375,
0.330078125,
0.484375,
0.34033203125,
0.0882568359375,
0.34033203125
],
[
0.51123046875,
0.333251953125,
0.9111328125,
0.33203125,
0.9111328125,
0.3466796875,
0.51123046875,
0.347900390625
],
[
0.0882568359375,
0.34228515625,
0.48779296875,
0.34228515625,
0.48779296875,
0.35595703125,
0.0882568359375,
0.35595703125
],
[
0.51318359375,
0.347900390625,
0.9111328125,
0.347900390625,
0.9111328125,
0.361083984375,
0.51318359375,
0.361083984375
],
[
0.08660888671875,
0.35595703125,
0.461669921875,
0.35595703125,
0.461669921875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.51318359375,
0.361083984375,
0.91259765625,
0.361083984375,
0.91259765625,
0.375,
0.51318359375,
0.375
],
[
0.0882568359375,
0.373779296875,
0.48193359375,
0.373779296875,
0.48193359375,
0.3837890625,
0.0882568359375,
0.3837890625
],
[
0.51318359375,
0.375,
0.9111328125,
0.375,
0.9111328125,
0.38818359375,
0.51318359375,
0.38818359375
],
[
0.08660888671875,
0.386962890625,
0.486083984375,
0.386962890625,
0.486083984375,
0.40087890625,
0.08660888671875,
0.40087890625
],
[
0.51318359375,
0.389404296875,
0.91259765625,
0.389404296875,
0.91259765625,
0.40283203125,
0.51318359375,
0.40283203125
],
[
0.0882568359375,
0.40087890625,
0.486083984375,
0.40087890625,
0.486083984375,
0.4140625,
0.0882568359375,
0.4140625
],
[
0.51318359375,
0.40283203125,
0.88134765625,
0.40283203125,
0.88134765625,
0.416748046875,
0.51318359375,
0.416748046875
],
[
0.0849609375,
0.4140625,
0.484375,
0.415283203125,
0.484375,
0.4287109375,
0.0849609375,
0.427978515625
],
[
0.08905029296875,
0.429931640625,
0.484375,
0.429931640625,
0.484375,
0.43994140625,
0.08905029296875,
0.43994140625
],
[
0.51708984375,
0.436279296875,
0.521484375,
0.436279296875,
0.521484375,
0.44140625,
0.51708984375,
0.44140625
],
[
0.54150390625,
0.43310546875,
0.7236328125,
0.43310546875,
0.7236328125,
0.4462890625,
0.54150390625,
0.4462890625
],
[
0.08660888671875,
0.442626953125,
0.48779296875,
0.442626953125,
0.48779296875,
0.455810546875,
0.08660888671875,
0.455810546875
],
[
0.0882568359375,
0.45703125,
0.488525390625,
0.45703125,
0.488525390625,
0.47021484375,
0.0882568359375,
0.47021484375
],
[
0.51318359375,
0.455078125,
0.69287109375,
0.455078125,
0.69287109375,
0.4697265625,
0.51318359375,
0.4697265625
],
[
0.0882568359375,
0.4716796875,
0.483642578125,
0.4716796875,
0.483642578125,
0.481689453125,
0.0882568359375,
0.481689453125
],
[
0.51318359375,
0.474853515625,
0.90869140625,
0.474853515625,
0.90869140625,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.08660888671875,
0.484130859375,
0.48779296875,
0.484130859375,
0.48779296875,
0.49755859375,
0.08660888671875,
0.49755859375
],
[
0.51318359375,
0.4892578125,
0.91259765625,
0.4892578125,
0.91259765625,
0.50244140625,
0.51318359375,
0.50244140625
],
[
0.0882568359375,
0.498779296875,
0.484375,
0.498779296875,
0.484375,
0.51220703125,
0.0882568359375,
0.51220703125
],
[
0.51123046875,
0.50244140625,
0.9111328125,
0.50244140625,
0.9111328125,
0.5166015625,
0.51123046875,
0.5166015625
],
[
0.08660888671875,
0.51220703125,
0.435546875,
0.51220703125,
0.435546875,
0.52587890625,
0.08660888671875,
0.52587890625
],
[
0.51123046875,
0.5166015625,
0.91259765625,
0.5166015625,
0.91259765625,
0.52978515625,
0.51123046875,
0.52978515625
],
[
0.10296630859375,
0.52880859375,
0.486083984375,
0.52880859375,
0.486083984375,
0.54248046875,
0.10296630859375,
0.54248046875
],
[
0.51123046875,
0.52978515625,
0.9111328125,
0.52978515625,
0.9111328125,
0.54345703125,
0.51123046875,
0.54345703125
],
[
0.0849609375,
0.54345703125,
0.48779296875,
0.54345703125,
0.48779296875,
0.556640625,
0.0849609375,
0.556640625
],
[
0.51318359375,
0.54345703125,
0.91259765625,
0.54345703125,
0.91259765625,
0.556640625,
0.51318359375,
0.556640625
],
[
0.08660888671875,
0.5595703125,
0.484375,
0.5595703125,
0.484375,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.51318359375,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.51318359375,
0.5712890625
],
[
0.08660888671875,
0.5712890625,
0.486083984375,
0.5712890625,
0.486083984375,
0.58544921875,
0.08660888671875,
0.58544921875
],
[
0.51123046875,
0.5712890625,
0.91259765625,
0.5712890625,
0.91259765625,
0.58544921875,
0.51123046875,
0.58544921875
],
[
0.0849609375,
0.58544921875,
0.484375,
0.58544921875,
0.484375,
0.5986328125,
0.0849609375,
0.5986328125
],
[
0.5146484375,
0.5869140625,
0.9111328125,
0.5869140625,
0.9111328125,
0.59765625,
0.5146484375,
0.59765625
],
[
0.08660888671875,
0.599609375,
0.483642578125,
0.599609375,
0.483642578125,
0.60986328125,
0.08660888671875,
0.60986328125
],
[
0.51123046875,
0.599609375,
0.91259765625,
0.599609375,
0.91259765625,
0.61279296875,
0.51123046875,
0.61279296875
],
[
0.08660888671875,
0.61279296875,
0.486083984375,
0.61279296875,
0.486083984375,
0.626953125,
0.08660888671875,
0.626953125
],
[
0.5107421875,
0.6123046875,
0.9111328125,
0.6123046875,
0.9111328125,
0.62548828125,
0.5107421875,
0.62548828125
],
[
0.08660888671875,
0.626953125,
0.486083984375,
0.626953125,
0.486083984375,
0.64013671875,
0.08660888671875,
0.64013671875
],
[
0.51318359375,
0.62744140625,
0.90625,
0.62744140625,
0.90625,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.0849609375,
0.6396484375,
0.483642578125,
0.6396484375,
0.483642578125,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.51318359375,
0.64013671875,
0.90869140625,
0.64013671875,
0.90869140625,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.0849609375,
0.65380859375,
0.486083984375,
0.65380859375,
0.486083984375,
0.66748046875,
0.0849609375,
0.66748046875
],
[
0.52685546875,
0.65478515625,
0.9111328125,
0.65478515625,
0.9111328125,
0.66943359375,
0.52685546875,
0.66943359375
],
[
0.08660888671875,
0.66845703125,
0.486083984375,
0.66845703125,
0.486083984375,
0.681640625,
0.08660888671875,
0.681640625
],
[
0.51318359375,
0.66943359375,
0.91015625,
0.66943359375,
0.91015625,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.08660888671875,
0.681640625,
0.486083984375,
0.681640625,
0.486083984375,
0.69580078125,
0.08660888671875,
0.69580078125
],
[
0.51318359375,
0.68310546875,
0.91259765625,
0.68310546875,
0.91259765625,
0.6962890625,
0.51318359375,
0.6962890625
],
[
0.0849609375,
0.6943359375,
0.48779296875,
0.69580078125,
0.48779296875,
0.708984375,
0.0849609375,
0.70849609375
],
[
0.51318359375,
0.6962890625,
0.91259765625,
0.6962890625,
0.91259765625,
0.7109375,
0.51318359375,
0.7109375
],
[
0.08660888671875,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.72021484375,
0.08660888671875,
0.72021484375
],
[
0.515625,
0.7119140625,
0.90869140625,
0.7119140625,
0.90869140625,
0.72265625,
0.515625,
0.72265625
],
[
0.0849609375,
0.7236328125,
0.4697265625,
0.7236328125,
0.4697265625,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.5146484375,
0.724609375,
0.9111328125,
0.724609375,
0.9111328125,
0.73779296875,
0.5146484375,
0.73779296875
],
[
0.10455322265625,
0.73779296875,
0.484375,
0.73779296875,
0.484375,
0.74853515625,
0.10455322265625,
0.74853515625
],
[
0.51123046875,
0.7373046875,
0.91259765625,
0.73779296875,
0.91259765625,
0.751953125,
0.51123046875,
0.75048828125
],
[
0.0849609375,
0.75,
0.484375,
0.75,
0.484375,
0.76318359375,
0.0849609375,
0.76318359375
],
[
0.51318359375,
0.75244140625,
0.91259765625,
0.75244140625,
0.91259765625,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.0849609375,
0.7646484375,
0.484375,
0.7646484375,
0.484375,
0.77783203125,
0.0849609375,
0.77783203125
],
[
0.5146484375,
0.767578125,
0.5888671875,
0.767578125,
0.5888671875,
0.77880859375,
0.5146484375,
0.77880859375
],
[
0.0882568359375,
0.77978515625,
0.484375,
0.77978515625,
0.484375,
0.79052734375,
0.0882568359375,
0.79052734375
],
[
0.52783203125,
0.77880859375,
0.91259765625,
0.77978515625,
0.91259765625,
0.79443359375,
0.52783203125,
0.79345703125
],
[
0.0849609375,
0.79248046875,
0.484375,
0.79248046875,
0.484375,
0.80615234375,
0.0849609375,
0.80615234375
],
[
0.51318359375,
0.79443359375,
0.90625,
0.79443359375,
0.90625,
0.8046875,
0.51318359375,
0.8046875
],
[
0.08660888671875,
0.80615234375,
0.48779296875,
0.80615234375,
0.48779296875,
0.82080078125,
0.08660888671875,
0.82080078125
],
[
0.51318359375,
0.80810546875,
0.9111328125,
0.80810546875,
0.9111328125,
0.8212890625,
0.51318359375,
0.8212890625
],
[
0.0882568359375,
0.82080078125,
0.484375,
0.82080078125,
0.484375,
0.833984375,
0.0882568359375,
0.833984375
],
[
0.51318359375,
0.8212890625,
0.91259765625,
0.8212890625,
0.91259765625,
0.8359375,
0.51318359375,
0.8359375
],
[
0.0849609375,
0.833984375,
0.484375,
0.833984375,
0.484375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.5146484375,
0.8369140625,
0.9111328125,
0.8369140625,
0.9111328125,
0.84765625,
0.5146484375,
0.84765625
],
[
0.08660888671875,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.5107421875,
0.84765625,
0.91259765625,
0.849609375,
0.91259765625,
0.8642578125,
0.5107421875,
0.8623046875
],
[
0.0849609375,
0.8623046875,
0.484375,
0.8623046875,
0.484375,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.08660888671875,
0.876953125,
0.125,
0.876953125,
0.125,
0.88818359375,
0.08660888671875,
0.88818359375
],
[
0.5361328125,
0.87548828125,
0.865234375,
0.87548828125,
0.865234375,
0.8896484375,
0.5361328125,
0.8896484375
]
] |
[
"(a) Compute a best response to _k and add it to Ek",
"In practice, we do not have the guarantee that the search",
"and to Eg; set ok[i] = 1.0//g/ for all o, in k.",
"algorithm used in PPSRO's main loop is able to return a best",
"(b) Set α-,[i]l = O and remove it from Eg- all oj,-k",
"response to a meta-strategy. So we use whichever approxi-",
"in _k that were estimated as redundant.",
"mation thesearchreturnsasif itwasabestresponsetothe",
"(c) Set o_k[i] = 1.0/Z。-,/ for all a, in Eg-,",
"meta-strategy. Moreover, depending on the game, we might",
"not beableto immediatelyrecognize a best responseto strat",
"2L starts by initializing the set of strategies of the empiri-",
"egyonceweseeone,asonewouldhavetoprovethestrategy",
"cal gameand theset of strategies inthe support of themeta-",
"tobeabestresponse.Thiscouldbeproblematic,forexam",
"strategy with an arbitrary strategy for one of the players (-i in",
"ple, when implementing the enhancement that 2L re-runs the",
"the pseudocode above). Then, it computes a best response to",
"search if there is a strategy in the empirical game that is a best",
"this arbitrary strategyandusesthebestresponseto initialize",
"response to the strategy the search returns. We run our exper-",
"E; and ., The meta-strategies are of the form (1.0) because",
"iments in games with utilities of -1, 0, +1. If a best response",
"the empirical game has a single strategy for each player (see",
"cannotbeeasilyverified (e.g.,MicroRTS).then weconsider",
"Step 1 above). Step 2 refers to PPSRO's loop, where it com-",
"that is a best response to ' if u(sinit, , a') = +1.",
"putes the best responses while alternating the players. Once",
"Once 2L reaches a computational budget, it can return dif-",
"a best response is computed to strategy k, it is added to",
"ferent strategies as its approximate solution to Equation 1.",
"thesupportofokwithuniformprobability(seeStep2a).",
"Similarly to IBR, it can return the last strategy added to the",
"2L estimates which strategies in the support of o_k are",
"empirical game for each player. 2L can also return a mixed",
"redundant while computing the best response to o-k.In",
"strategythatisgivenbythedistributionof strategiesadded to",
"Step2b,2Lremovestheredundantstrategiesfromthesup-",
"the empirical game, as does FP. We can also solve the result-",
"port of o-k and, in Step 2c, redistributes the probabilities so",
"ing empirical game with linear programming, like DO does,",
"that each strategy in the support has the same probability.",
"and return the resulting strategy. In this paper, we assume",
"Example5.In Example2we showed thatDOfails toin",
"the games have a pure dominant strategy for which IBR's ap-",
"clude both attack[l, 2, 3] and attack[1, 2, 5] in the",
"proach of returning the last strategy added to the empirical",
"support of the meta-strategy o_i, thus missing the guidance",
"game is suitable; this is what we use in our experiments.",
"informationattack[1,2,3] provides.Once the strategy",
"attack/l,2.5lis added to theempirical game,themeta-",
"",
"EmpiricalEvaluation",
"strategy will automatically have both attack[l, 2, 3] and",
"attack[l, 2, 5] in its support.In contrast withDO, 2L",
"7.1 ProblemDomains",
"retainsboth strategiesinthesupport of - for the next it-",
"In addition to P&R, we introduce Climbing Monkey (CM)",
"eration as long as strategies such as defend[1, 2, 3] and",
"another two-player zero-sum game with a trivial optimal",
"defend[l, 2, 5] are evaluated in search as both attack[l,",
"strategy that is also challenging in the context of program-",
"2, 3] and attack[1, 2, 5/ will be flagged as helpful.",
"matic strategies. In CM, monkeys need to climb to a branch",
"A weakness of 2L as presented above is that it can flag",
"of a tree that is higher than the branch the opponent's mon-",
"as redundant a strategy that is helpful if it does not sample",
"key is able to reach. The branches need to be climbed one",
"enough strategies in the search.For example,if the meta-",
"at a time, without skipping any branch. The monkey that",
"strategy for Rangers has both defend[1] and de fend[2] in",
"climbs to a higher branch wins the game. The game ends",
"its support, but it never evaluates a strategy that attacks gate 1",
"inadrawifbothmonkeysclimbtoabranchofthesame",
"in search.then defendllwill mistakenlyberemoved from",
"height. For a tree with n branches, a dominant programmatic",
"the meta-strategy's support. We implement the following en-",
"strategy is climb[1]l. climb[2],...,climbn]. Similarly",
"hancement to fix this weakness. Whenever the search returns",
"toP&R.CMis challenging because.depending on the num",
"a best response α to a meta-strategy o_i (resp. g;), we eval-",
"ber of branches, it requires one to synthesize long programs.",
"uate o against all strategies in the empirical game, including",
"In P&R, learning algorithms perform better if using a larger",
"those not in the support of -; (resp. o). If there is a strategy",
"number of strategies in the support of meta-strategies as hav-",
"o' in the empirical game that is a best response to g, then it",
"ing many strategies helps Rangers converge to a strategy that",
"must be that 2L mistakenly removed o' from the support of",
"protects all gates. CM is a game where all one needs to use",
"themeta-strategy.Inthis case.werepeat the search forabest",
"is the last strategy added to the empirical game, i.e., the strat",
"response with ' added to the support of the meta-strategy",
"egy that allows the monkey to climb to the highest branch.",
"This enhancement can increasethe numberof times the",
"Wehypothesize that 2L is capable of detecting which strate-",
" search algorithm is invoked in each iteration of the PPSRO",
"gies are needed in the support of the meta-strategies for these",
"loop.While we perform a single search per iteration with",
"two games.",
"IBR,FP,andDO,intheworstcase.2L canperformanum-",
"We also evaluate 2L in MicroRTS, a real-time strategy",
"ber of searches equal to the number of strategies in the game.",
"gamedesigned for research.Thereis an activeresearch com",
"This is because, in the worst case, we add all strategies of",
"munity that uses MicroRTS as a benchmark to evaluate in-",
"the empirical game to the support of the meta-strategy. De-",
"telligent systems.! MicroRTS is a game played with real-",
"spite the possible additional searches, preliminary experi-",
"time constraints and very large action and state spaces [Lelis,",
"ments showed that this enhancement improves the sampling",
"2021l. Each player can control two types of stationary units",
"efficiency of 2L. All results in this paper use this enhance-",
"ment.",
"'https://github.com/Farama-Foundation/MicroRTS/wiki"
] |
{
"arxiv_id": "2307.04893",
"language": "en",
"page_idx": 4
}
|
||
2004.14065-en-0
|
2004.14065-en-0.jpg
|
## Automatically Identifying Gender Issues in $\textbf{Machine Translation using Perturbations}$
Kellie Webster
Google Research NYC websterk@google.com
Hila Gonen
Bar-Ilan University hilagnn@ gmail. com
## $\mathbf{Abstract}$
The successful application of neural methods to machine translation has realized huge quality advances for the community. With these improvements, many have noted outstanding challenges, including the modeling and treatment of gendered language. While previous studies have identified issues using synthetic examples, we develop a novel technique to mine examples from real world data to explore challenges for deployed systems. We use our method to compile an evaluation benchmark spanning examples for four languages from three language families, which we publicly release to facilitate research. The examples in our benchmark expose where model representations are gendered, and the unintended consequences these gendered representations can have in downstream $\vec{\mathrm{application}}.$
## 1 Introduction
Machine translation (MT) has realized huge improvements in quality from the successful application and development of neural methods (Kalchbrenner and Blunsom, 2013; Cho et al., 2014; Vaswani et al., 2017; Johnson et al., 2017; Chen et al., 2018). As the community has explored this enhanced performance, many have noted the outstanding challenge of modeling and handling gendered language (Kuczmarski, 2018; Escudé Font and Costa-jussà, 2019). We extend this line of work, which identifes issues using synthetic examples manually curated for a target language (Stanovsky et al., 2019; Cho et al., 2019), by analyzing real world text across a range of languages to understand challenges for deployed systems.
In this paper, we explore the class of issues which surface when a neutral reference to a person is translated to a gendered form (e.g. in Table 1, where the English counselor and nurse are translated into
the French conseiller (masculine) and infırmière (feminine). For this class of examples, the MT task requires a system to produce a single translation without source cues, thus exposing a model's preferred gender for the reference form.
With this scope, we make two key contributions. First, we design and implement an automatic pipeline for detecting examples of our class of gender issues in real world input, using a BERTbased perturbation method novel to this work. A key advantage of our pipeline beyond previous work is its extensibility: a) beyond word lists; b) to different language pairs and c) parts of speech. Second, using our new pipeline, we compile a dataset that we make publicly available to serve as a benchmark for future work. We focus on English as the source language, and explore four target gendered languages across three language families (French, German, Spanish, and Russian).Our examples expose where MT encodings are gendered, finding new issues not covered in previous manual approaches, and the unintended consequences of this for translation.
## 2 Gender Marking Languages
Gender-marking languages have rich grammatical systems for expressing gender (Corbett, 1991). To produce a valid sentence in a gender-marking language, gender may need to be marked not only on pronouns (he, she), as it is in English, but also nouns and even verbs, as well as words linked to these gendered nouns and verbs. This means that translating from a language like English, with little gender marking, to a gender-marking language like Spanish, requires a system to produce gender markings that may not have explicit evidence in the source. For instance, The tall teacher from English could be translated into the Spanish La maestra alta (feminine) or El maestro alto (masculine)
|
[
[
0.2822265625,
0.0843505859375,
0.7197265625,
0.0843505859375,
0.7197265625,
0.09857177734375,
0.2822265625,
0.09857177734375
],
[
0.28955078125,
0.10394287109375,
0.7138671875,
0.10394287109375,
0.7138671875,
0.1181640625,
0.28955078125,
0.1181640625
],
[
0.285400390625,
0.15380859375,
0.387939453125,
0.15380859375,
0.387939453125,
0.168701171875,
0.285400390625,
0.168701171875
],
[
0.60546875,
0.15380859375,
0.73291015625,
0.15380859375,
0.73291015625,
0.168701171875,
0.60546875,
0.168701171875
],
[
0.258544921875,
0.17041015625,
0.4130859375,
0.17041015625,
0.4130859375,
0.185302734375,
0.258544921875,
0.185302734375
],
[
0.5751953125,
0.17041015625,
0.76171875,
0.169189453125,
0.76171875,
0.18408203125,
0.5751953125,
0.185302734375
],
[
0.2493896484375,
0.185302734375,
0.42236328125,
0.1859130859375,
0.42236328125,
0.2037353515625,
0.2493896484375,
0.203125
],
[
0.57421875,
0.188232421875,
0.76318359375,
0.188232421875,
0.76318359375,
0.201904296875,
0.57421875,
0.201904296875
],
[
0.264404296875,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.264404296875,
0.279052734375
],
[
0.51318359375,
0.26416015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.51318359375,
0.27783203125
],
[
0.5146484375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.1485595703125,
0.29345703125,
0.459228515625,
0.29345703125,
0.459228515625,
0.30712890625,
0.1485595703125,
0.30712890625
],
[
0.51318359375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.0251922607421875,
0.311279296875,
0.060455322265625,
0.311279296875,
0.060455322265625,
0.43408203125,
0.0251922607421875,
0.43408203125
],
[
0.14697265625,
0.30810546875,
0.459228515625,
0.30810546875,
0.459228515625,
0.32177734375,
0.14697265625,
0.32177734375
],
[
0.51318359375,
0.314208984375,
0.88330078125,
0.314208984375,
0.88330078125,
0.328857421875,
0.51318359375,
0.328857421875
],
[
0.14697265625,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.14697265625,
0.336181640625
],
[
0.145263671875,
0.336181640625,
0.459228515625,
0.33740234375,
0.459228515625,
0.350830078125,
0.145263671875,
0.349853515625
],
[
0.51318359375,
0.330810546875,
0.8076171875,
0.330810546875,
0.8076171875,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.14697265625,
0.350830078125,
0.459228515625,
0.350830078125,
0.459228515625,
0.364013671875,
0.14697265625,
0.364013671875
],
[
0.53076171875,
0.3486328125,
0.884765625,
0.349853515625,
0.884765625,
0.364013671875,
0.53076171875,
0.363525390625
],
[
0.14697265625,
0.364013671875,
0.45751953125,
0.364013671875,
0.45751953125,
0.377685546875,
0.14697265625,
0.377685546875
],
[
0.51318359375,
0.365234375,
0.88330078125,
0.366455078125,
0.88330078125,
0.38134765625,
0.51318359375,
0.380126953125
],
[
0.145263671875,
0.37890625,
0.459228515625,
0.380126953125,
0.459228515625,
0.393798828125,
0.145263671875,
0.392578125
],
[
0.51318359375,
0.383056640625,
0.884765625,
0.3818359375,
0.884765625,
0.396728515625,
0.51318359375,
0.39794921875
],
[
0.14697265625,
0.393798828125,
0.4609375,
0.393798828125,
0.4609375,
0.40673828125,
0.14697265625,
0.40673828125
],
[
0.51318359375,
0.3984375,
0.88330078125,
0.39794921875,
0.88330078125,
0.412109375,
0.51318359375,
0.413330078125
],
[
0.1485595703125,
0.407958984375,
0.459228515625,
0.407958984375,
0.459228515625,
0.421630859375,
0.1485595703125,
0.421630859375
],
[
0.51318359375,
0.4150390625,
0.88330078125,
0.4150390625,
0.88330078125,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.14697265625,
0.4228515625,
0.4609375,
0.4228515625,
0.4609375,
0.435791015625,
0.14697265625,
0.435791015625
],
[
0.039459228515625,
0.431640625,
0.04534912109375,
0.431640625,
0.04534912109375,
0.43701171875,
0.039459228515625,
0.43701171875
],
[
0.14697265625,
0.435791015625,
0.4609375,
0.435791015625,
0.4609375,
0.449462890625,
0.14697265625,
0.449462890625
],
[
0.5146484375,
0.431640625,
0.88330078125,
0.431640625,
0.88330078125,
0.4453125,
0.5146484375,
0.4453125
],
[
0.0251922607421875,
0.438232421875,
0.0621337890625,
0.438232421875,
0.0621337890625,
0.5185546875,
0.0251922607421875,
0.5185546875
],
[
0.1485595703125,
0.45068359375,
0.4609375,
0.45068359375,
0.4609375,
0.46435546875,
0.1485595703125,
0.46435546875
],
[
0.51318359375,
0.446533203125,
0.884765625,
0.446533203125,
0.884765625,
0.46142578125,
0.51318359375,
0.46142578125
],
[
0.14697265625,
0.46435546875,
0.4609375,
0.46435546875,
0.4609375,
0.47802734375,
0.14697265625,
0.47802734375
],
[
0.51318359375,
0.463134765625,
0.88330078125,
0.463134765625,
0.88330078125,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.1502685546875,
0.479736328125,
0.459228515625,
0.479736328125,
0.459228515625,
0.490478515625,
0.1502685546875,
0.490478515625
],
[
0.51123046875,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.493408203125,
0.51123046875,
0.493408203125
],
[
0.1485595703125,
0.4921875,
0.4609375,
0.4921875,
0.4609375,
0.50732421875,
0.1485595703125,
0.50732421875
],
[
0.51318359375,
0.495361328125,
0.884765625,
0.495361328125,
0.884765625,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.14697265625,
0.50732421875,
0.459228515625,
0.50732421875,
0.459228515625,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.509765625,
0.50830078125,
0.88818359375,
0.51025390625,
0.88818359375,
0.52783203125,
0.509765625,
0.525390625
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.14697265625,
0.521484375,
0.4609375,
0.521484375,
0.4609375,
0.53515625,
0.14697265625,
0.53515625
],
[
0.51318359375,
0.52783203125,
0.88330078125,
0.52783203125,
0.88330078125,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.1435546875,
0.53515625,
0.4609375,
0.53369140625,
0.4609375,
0.548828125,
0.1435546875,
0.5498046875
],
[
0.1435546875,
0.5498046875,
0.2301025390625,
0.54736328125,
0.2301025390625,
0.5625,
0.1435546875,
0.56396484375
],
[
0.5146484375,
0.54443359375,
0.88330078125,
0.54443359375,
0.88330078125,
0.55810546875,
0.5146484375,
0.55810546875
],
[
0.5146484375,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.1435546875,
0.57763671875,
0.26025390625,
0.57763671875,
0.26025390625,
0.5927734375,
0.1435546875,
0.5927734375
],
[
0.51123046875,
0.57470703125,
0.884765625,
0.576171875,
0.884765625,
0.59033203125,
0.51123046875,
0.5888671875
],
[
0.124267578125,
0.58203125,
0.130126953125,
0.58203125,
0.130126953125,
0.58740234375,
0.124267578125,
0.58740234375
],
[
0.1351318359375,
0.5830078125,
0.1427001953125,
0.5830078125,
0.1427001953125,
0.5859375,
0.1351318359375,
0.5859375
],
[
0.1427001953125,
0.58203125,
0.1485595703125,
0.58203125,
0.1485595703125,
0.58740234375,
0.1427001953125,
0.58740234375
],
[
0.51318359375,
0.5927734375,
0.88330078125,
0.5927734375,
0.88330078125,
0.60546875,
0.51318359375,
0.60546875
],
[
0.1175537109375,
0.60400390625,
0.486083984375,
0.60400390625,
0.486083984375,
0.61767578125,
0.1175537109375,
0.61767578125
],
[
0.5146484375,
0.60986328125,
0.8623046875,
0.60986328125,
0.8623046875,
0.62060546875,
0.5146484375,
0.62060546875
],
[
0.1175537109375,
0.62060546875,
0.489501953125,
0.619140625,
0.489501953125,
0.63427734375,
0.1175537109375,
0.63525390625
],
[
0.1175537109375,
0.63720703125,
0.48779296875,
0.6357421875,
0.48779296875,
0.65087890625,
0.1175537109375,
0.65185546875
],
[
0.1175537109375,
0.65185546875,
0.489501953125,
0.65185546875,
0.489501953125,
0.6650390625,
0.1175537109375,
0.6650390625
],
[
0.509765625,
0.64794921875,
0.791015625,
0.6494140625,
0.791015625,
0.66650390625,
0.509765625,
0.66455078125
],
[
0.12005615234375,
0.66943359375,
0.489501953125,
0.66943359375,
0.489501953125,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.12005615234375,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.5146484375,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.12005615234375,
0.70166015625,
0.486083984375,
0.70166015625,
0.486083984375,
0.7119140625,
0.12005615234375,
0.7119140625
],
[
0.51123046875,
0.70166015625,
0.88330078125,
0.70068359375,
0.88330078125,
0.71533203125,
0.51123046875,
0.71630859375
],
[
0.12005615234375,
0.71728515625,
0.489501953125,
0.71728515625,
0.489501953125,
0.73095703125,
0.12005615234375,
0.73095703125
],
[
0.51318359375,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.51318359375,
0.732421875
],
[
0.118408203125,
0.73291015625,
0.490234375,
0.732421875,
0.490234375,
0.74658203125,
0.118408203125,
0.74755859375
],
[
0.51123046875,
0.73388671875,
0.8818359375,
0.73291015625,
0.8818359375,
0.74755859375,
0.51123046875,
0.7490234375
],
[
0.118408203125,
0.7490234375,
0.490234375,
0.74755859375,
0.490234375,
0.7626953125,
0.118408203125,
0.76318359375
],
[
0.51318359375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.12005615234375,
0.7666015625,
0.490234375,
0.7666015625,
0.490234375,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.12005615234375,
0.783203125,
0.489501953125,
0.783203125,
0.489501953125,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.79736328125,
0.489501953125,
0.79736328125,
0.489501953125,
0.8125,
0.118408203125,
0.8125
],
[
0.51318359375,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.51318359375,
0.8125
],
[
0.118408203125,
0.8154296875,
0.490234375,
0.8154296875,
0.490234375,
0.8291015625,
0.118408203125,
0.8291015625
],
[
0.51318359375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.8291015625,
0.51318359375,
0.8291015625
],
[
0.118408203125,
0.83056640625,
0.359375,
0.83056640625,
0.359375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51123046875,
0.830078125,
0.884765625,
0.83056640625,
0.884765625,
0.845703125,
0.51123046875,
0.84423828125
],
[
0.1368408203125,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.1368408203125,
0.86083984375
],
[
0.51318359375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.49365234375,
0.86328125,
0.49365234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51123046875,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51123046875,
0.876953125
],
[
0.118408203125,
0.87890625,
0.49365234375,
0.87890625,
0.49365234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.509765625,
0.87744140625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.509765625,
0.892578125
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.85791015625,
0.8955078125,
0.85791015625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Automatically Identifying Gender Issues in",
"Machine Translation using Perturbations",
"Hila Gonen",
"Kellie Webster",
"Bar-Ilan University",
"Google Research NYC",
"hilagnn@gmail.com",
"websterk@google.com",
"Abstract",
"the French conseiller (masculine)and infirmiere",
"(feminine). For this class of examples, the MT task",
"The successful application of neural methods",
"requires a system to produce a single translation",
"",
"to machine translation has realized huge quality",
"without source cues, thus exposing a model's",
"advances for the community. With these",
"improvements, many have noted outstanding",
"preferred gender for the reference form.",
"challenges, including the modeling and",
"With this scope, we make two key contributions.",
"treatment of gendered language. While",
"First, we design and implement an automatic",
"previous studies have identified issues using",
"pipeline for detecting examples of our class of",
"synthetic examples,we develop a novel",
"gender issues in real world input, using a BERT",
"technique to mine examples from real world",
"based perturbation method novel to this work. A key",
"data to explore challenges for deployed",
"",
"systems. We use our method to compile an",
"advantage of our pipeline beyond previous work is",
"['S]",
"evaluation benchmark spanning examples for",
"its extensibility: a) beyond word lists; b) to different",
"four languages from three language families,",
"language pairs and c) parts of speech. Second, using",
"whichwepubliclyreleasetofacilitateresearch",
" our new pipeline, we compile a dataset that we make",
"The examples in our benchmark expose",
"publicly available to serve as a benchmark for future",
"where model representations are gendered, and",
"work. We focus on English as the source language,",
"91002",
"the unintended consequences these gendered",
"and explore four target gendered languages across",
"representations can have in downstream",
"application.",
"three language families (French, German, Spanish",
"and Russian). Our examples expose where MT",
"Introduction",
"encodings are gendered, finding new issues not",
"",
"",
"",
"covered in previous manual approaches, and the",
"Machine translation (MT) has realized huge",
"unintendedconsequencesofthisfortranslation",
"improvements in quality from the successful",
"application and development of neural methods",
"(Kalchbrenner and Blunsom. 2013: Cho et al.,",
"2 Gender Marking Languages",
"2014; Vaswani et al., 2017; Johnson et al., 2017;",
"Chen et al., 2018). As the community has explored",
"Gender-marking languages have rich grammatical",
"this enhanced performance.many have noted the",
"systems for expressing gender (Corbett, 1991)",
"outstanding challenge of modeling and handling",
"To produce a valid sentence in a gender-marking",
"gendered language (Kuczmarski,2018; Escude Font",
"language, gender may need to be marked not only",
"and Costa-jussa, 2019). We extend this line of work,",
"on pronouns (he, she), as it is in English, but also",
"which identifies issues using synthetic examples",
"nouns and even verbs, as well as words linked",
"manually curated for a target language (Stanovsky",
"to these gendered nouns and verbs. This means",
"et al., 2019; Cho et al., 2019), by analyzing real",
"that translating from a language like English, with",
"world text across a range of languages to understand",
"little gender marking, to a gender-marking language",
"challenges for deployed systems.",
"like Spanish, requires a system to produce gender",
"In this paper, we explore the class of issues which",
"markings that may not have explicit evidence in the",
"surface when a neutral reference to a person is",
" source. For instance, The tall teacher from English",
"translated to a gendered form (e.g. in Table I, where",
"could be translated into the Spanish La maestra",
"the English counselor and nurse are translated into",
"alta (feminine) or El maestro alto (masculine)"
] |
{
"arxiv_id": "2004.14065",
"language": "en",
"page_idx": 0
}
|
||
2004.14065-en-1
|
2004.14065-en-1.jpg
|
<table>
<tbody>
<tr>
<th>Source S Sentence (En)</th>
<th>Translation (Fr)</th>
<th>$M/F$</th>
</tr>
<tr>
<td>so is that going to affect my chances of becoming a counselor?</td>
<td>Alors, est- ce que cela va affecter mes chances de devenir conseiller?</td>
<td>$M$</td>
</tr>
<tr>
<td>so is that going to affect my chances of becoming a nurse?</td>
<td>Alors, est- ce que cela va affecter mes chances de devenir infirmière?</td>
<td>$F$</td>
</tr>
</tbody>
</table>
Table l: An example from our dataset of a minimal pair of English gender-neutral source sentences, translated into
two different genders in French. Red (italic) stands for masculine, cyan (normal) stands for feminine
3 Automatic Detection of Gender Issues
The class of issues we are interested in are those where translation to a gender-marking language exposes a model's gender preference for a personal reference. The examples we find that demonstrate this are English sentence-pairs, a minimal pair differing by only a single word, e.g. doctor being replaced by nurse. In each of our examples, this minimal perturbation does not change the gender of the source but gives rise to gender differences upon translation, e.g. doctor becoming masculine and nurse feminine.
In this section, we present a simple, extensible method to mine such examples from real-world text. Our method does not require expansive manuallycurated word lists for each target language, which enables us to discover new kinds of entities that are susceptible to model bias but are not usually thought of this way. Indeed, while we demonstrate its utility with nouns with four target languages, our method is naturally extensible to new language pairs and parts of speech with no change in design. Filtering source sentences Our first step is to identify sentences that are gender neutral and that include a single human entity, e.g. $A\_\underline{doctor}worksin$ $a\textit{ hospital. We focus on human entities since these}$ have been the target of previous studies and present the largest risk of gender issues in translation
We use a BERT-based Named Entity Recognition (NER) model that identifies human entities, and exclude sentences that have more than one token tagged as such. We also remove sentences in which the entity is a gendered term in English$^1($e.g. mother, $nephew)$, a name, or not a noun.
Note that all the sentences we get are naturally occurring sentences, and that we do not use any templates or predefined lists of target words that we want to handle.
$^{1}$https://github.com/tolga-b/debiaswe/blob/
$\mathrm{master/data/gender}_{-}\mathrm{specific}_{-}$full.json
https://github.com/uclanlp/gn\_glove/tree/master/
wordlist
Perturbations using BERT We use BERT as a masked language model to find words which can substitute for the human entity identified in the previous filtering step, e.g. $doctor\to nurse.$ We aim to get natural-sounding output and maintain extensibility, and thus do not use predefined substitutions. We cap our search to the frst 100 candidates BERT returns, accepting the first 10 which are tagged as person, and for which the resulting sentences also pass the filtering step.
Translation We translate each of the generated sentences into our target languages using Google Translate$^2.A\_\underline{doctor/nurse}$ works in $a$ hospital $\to$ $\underline{Un\textit{ doctor}/Una\textit{ enfermera trabaja en un hospital.}}$
Alignment We align tokens in the original and translated sentences using fast-align (Dyer et al., 2010). This is needed in order to know which token in the translation output is the focus entity in the source sentence, whose gender we want to analyze.
Gender Identification We use a morphological analyzer, implemented following Kong et al. (2017), to tag the gender of the target word.
Identifying Examples The final step of our pipeline is identifying pairs of sentences to include in our dataset, pairs where different genders are assigned to the human entity. Our example would be included since doctor is translated with the masculine form Un doctor while nurse is translated with the feminine form Una enfermera.
### 4 Challenge Dataset
We compile our fnal dataset from the output of this pipeline, and explore its properties to understand the issues it represents for deployed systems.
# 4.1 Random Sampling
In our final dataset, we include both examples that passed the final example identification step above (pairs referred to as“at risk”), as well as a random
$^2$https://translate.google.com/
|
[
[
0.1275634765625,
0.0736083984375,
0.26123046875,
0.0736083984375,
0.26123046875,
0.0872802734375,
0.1275634765625,
0.0872802734375
],
[
0.486083984375,
0.0748291015625,
0.58447265625,
0.0748291015625,
0.58447265625,
0.08612060546875,
0.486083984375,
0.08612060546875
],
[
0.8427734375,
0.0748291015625,
0.87744140625,
0.0748291015625,
0.87744140625,
0.0872802734375,
0.8427734375,
0.0872802734375
],
[
0.1275634765625,
0.09027099609375,
0.84228515625,
0.09027099609375,
0.84228515625,
0.10394287109375,
0.1275634765625,
0.10394287109375
],
[
0.83642578125,
0.09320068359375,
0.861328125,
0.09320068359375,
0.861328125,
0.1015625,
0.83642578125,
0.1015625
],
[
0.1292724609375,
0.10394287109375,
0.198974609375,
0.10394287109375,
0.198974609375,
0.114013671875,
0.1292724609375,
0.114013671875
],
[
0.48779296875,
0.10394287109375,
0.5556640625,
0.10394287109375,
0.5556640625,
0.114013671875,
0.48779296875,
0.114013671875
],
[
0.1275634765625,
0.1181640625,
0.8388671875,
0.1181640625,
0.8388671875,
0.12890625,
0.1275634765625,
0.12890625
],
[
0.8427734375,
0.12054443359375,
0.85400390625,
0.12054443359375,
0.85400390625,
0.1258544921875,
0.8427734375,
0.1258544921875
],
[
0.486083984375,
0.1300048828125,
0.5576171875,
0.1300048828125,
0.5576171875,
0.141357421875,
0.486083984375,
0.141357421875
],
[
0.118408203125,
0.1561279296875,
0.88330078125,
0.1561279296875,
0.88330078125,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.118408203125,
0.171630859375,
0.79443359375,
0.171630859375,
0.79443359375,
0.185302734375,
0.118408203125,
0.185302734375
],
[
0.118408203125,
0.2161865234375,
0.476806640625,
0.2161865234375,
0.476806640625,
0.23095703125,
0.118408203125,
0.23095703125
],
[
0.51318359375,
0.21728515625,
0.88330078125,
0.21728515625,
0.88330078125,
0.23095703125,
0.51318359375,
0.23095703125
],
[
0.51123046875,
0.2327880859375,
0.88330078125,
0.232177734375,
0.88330078125,
0.2464599609375,
0.51123046875,
0.2476806640625
],
[
0.1158447265625,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.1158447265625,
0.257080078125
],
[
0.51123046875,
0.2493896484375,
0.88330078125,
0.2493896484375,
0.88330078125,
0.26318359375,
0.51123046875,
0.26318359375
],
[
0.118408203125,
0.26123046875,
0.489501953125,
0.26123046875,
0.489501953125,
0.27490234375,
0.118408203125,
0.27490234375
],
[
0.51123046875,
0.265380859375,
0.88330078125,
0.26416015625,
0.88330078125,
0.279052734375,
0.51123046875,
0.27978515625
],
[
0.118408203125,
0.276611328125,
0.489501953125,
0.276611328125,
0.489501953125,
0.290283203125,
0.118408203125,
0.290283203125
],
[
0.51123046875,
0.28076171875,
0.88330078125,
0.28076171875,
0.88330078125,
0.29443359375,
0.51123046875,
0.29443359375
],
[
0.118408203125,
0.292236328125,
0.490234375,
0.292236328125,
0.490234375,
0.305908203125,
0.118408203125,
0.305908203125
],
[
0.51318359375,
0.297607421875,
0.884765625,
0.297607421875,
0.884765625,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.1175537109375,
0.30712890625,
0.489501953125,
0.30810546875,
0.489501953125,
0.322509765625,
0.1175537109375,
0.32177734375
],
[
0.51123046875,
0.3134765625,
0.88330078125,
0.312255859375,
0.88330078125,
0.32666015625,
0.51123046875,
0.327880859375
],
[
0.12005615234375,
0.32470703125,
0.489501953125,
0.32470703125,
0.489501953125,
0.33837890625,
0.12005615234375,
0.33837890625
],
[
0.51123046875,
0.328857421875,
0.884765625,
0.328857421875,
0.884765625,
0.34326171875,
0.51123046875,
0.34326171875
],
[
0.1175537109375,
0.34033203125,
0.490234375,
0.34033203125,
0.490234375,
0.35400390625,
0.1175537109375,
0.35400390625
],
[
0.51318359375,
0.3466796875,
0.88330078125,
0.3466796875,
0.88330078125,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.118408203125,
0.35693359375,
0.490234375,
0.35693359375,
0.490234375,
0.37060546875,
0.118408203125,
0.37060546875
],
[
0.51318359375,
0.363525390625,
0.85205078125,
0.363525390625,
0.85205078125,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.118408203125,
0.372314453125,
0.490234375,
0.372314453125,
0.490234375,
0.385986328125,
0.118408203125,
0.385986328125
],
[
0.118408203125,
0.388427734375,
0.49365234375,
0.388427734375,
0.49365234375,
0.402587890625,
0.118408203125,
0.402587890625
],
[
0.51123046875,
0.385986328125,
0.884765625,
0.38720703125,
0.884765625,
0.402099609375,
0.51123046875,
0.40087890625
],
[
0.1175537109375,
0.405029296875,
0.268798828125,
0.405029296875,
0.268798828125,
0.419189453125,
0.1175537109375,
0.419189453125
],
[
0.51318359375,
0.40380859375,
0.88330078125,
0.40380859375,
0.88330078125,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.1351318359375,
0.421630859375,
0.490234375,
0.421630859375,
0.490234375,
0.435791015625,
0.1351318359375,
0.435791015625
],
[
0.51123046875,
0.418701171875,
0.88330078125,
0.419189453125,
0.88330078125,
0.43408203125,
0.51123046875,
0.432861328125
],
[
0.118408203125,
0.438232421875,
0.49365234375,
0.438232421875,
0.49365234375,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.5146484375,
0.435791015625,
0.8759765625,
0.435791015625,
0.8759765625,
0.449462890625,
0.5146484375,
0.449462890625
],
[
0.118408203125,
0.45361328125,
0.490234375,
0.45361328125,
0.490234375,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.5146484375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.474365234375,
0.5146484375,
0.474365234375
],
[
0.118408203125,
0.468505859375,
0.489501953125,
0.4697265625,
0.489501953125,
0.48388671875,
0.118408203125,
0.482666015625
],
[
0.51318359375,
0.47802734375,
0.884765625,
0.47802734375,
0.884765625,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.12005615234375,
0.486328125,
0.489501953125,
0.486328125,
0.489501953125,
0.49951171875,
0.12005615234375,
0.49951171875
],
[
0.51123046875,
0.493408203125,
0.88330078125,
0.493408203125,
0.88330078125,
0.50732421875,
0.51123046875,
0.50732421875
],
[
0.118408203125,
0.501953125,
0.486083984375,
0.501953125,
0.486083984375,
0.515625,
0.118408203125,
0.515625
],
[
0.51318359375,
0.51025390625,
0.88330078125,
0.51025390625,
0.88330078125,
0.52392578125,
0.51318359375,
0.52392578125
],
[
0.12005615234375,
0.5185546875,
0.489501953125,
0.5185546875,
0.489501953125,
0.5322265625,
0.12005615234375,
0.5322265625
],
[
0.51318359375,
0.525390625,
0.88671875,
0.525390625,
0.88671875,
0.5390625,
0.51318359375,
0.5390625
],
[
0.118408203125,
0.53515625,
0.490234375,
0.53515625,
0.490234375,
0.548828125,
0.118408203125,
0.548828125
],
[
0.118408203125,
0.55029296875,
0.489501953125,
0.55029296875,
0.489501953125,
0.56396484375,
0.118408203125,
0.56396484375
],
[
0.5146484375,
0.55029296875,
0.69775390625,
0.55029296875,
0.69775390625,
0.56396484375,
0.5146484375,
0.56396484375
],
[
0.6884765625,
0.5498046875,
0.88330078125,
0.55029296875,
0.88330078125,
0.5654296875,
0.6884765625,
0.56396484375
],
[
0.118408203125,
0.56640625,
0.490234375,
0.56640625,
0.490234375,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.51318359375,
0.56787109375,
0.88330078125,
0.56787109375,
0.88330078125,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.51318359375,
0.5830078125,
0.779296875,
0.5830078125,
0.779296875,
0.5966796875,
0.51318359375,
0.5966796875
],
[
0.12005615234375,
0.59326171875,
0.489501953125,
0.59326171875,
0.489501953125,
0.60693359375,
0.12005615234375,
0.60693359375
],
[
0.1175537109375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62255859375,
0.1175537109375,
0.62255859375
],
[
0.51318359375,
0.60693359375,
0.88330078125,
0.60693359375,
0.88330078125,
0.62158203125,
0.51318359375,
0.62158203125
],
[
0.118408203125,
0.626953125,
0.489501953125,
0.626953125,
0.489501953125,
0.64013671875,
0.118408203125,
0.64013671875
],
[
0.51318359375,
0.6259765625,
0.884765625,
0.6259765625,
0.884765625,
0.6396484375,
0.51318359375,
0.6396484375
],
[
0.118408203125,
0.642578125,
0.490234375,
0.642578125,
0.490234375,
0.65625,
0.118408203125,
0.65625
],
[
0.51318359375,
0.64111328125,
0.884765625,
0.64111328125,
0.884765625,
0.65478515625,
0.51318359375,
0.65478515625
],
[
0.118408203125,
0.65771484375,
0.489501953125,
0.65771484375,
0.489501953125,
0.67138671875,
0.118408203125,
0.67138671875
],
[
0.51318359375,
0.65673828125,
0.88330078125,
0.65673828125,
0.88330078125,
0.67041015625,
0.51318359375,
0.67041015625
],
[
0.118408203125,
0.6748046875,
0.45751953125,
0.6748046875,
0.45751953125,
0.6884765625,
0.118408203125,
0.6884765625
],
[
0.5146484375,
0.6748046875,
0.88330078125,
0.6748046875,
0.88330078125,
0.68505859375,
0.5146484375,
0.68505859375
],
[
0.1351318359375,
0.689453125,
0.489501953125,
0.6904296875,
0.489501953125,
0.705078125,
0.1351318359375,
0.70361328125
],
[
0.51123046875,
0.6884765625,
0.88330078125,
0.6884765625,
0.88330078125,
0.70166015625,
0.51123046875,
0.70166015625
],
[
0.1175537109375,
0.7060546875,
0.48779296875,
0.7060546875,
0.48779296875,
0.7197265625,
0.1175537109375,
0.7197265625
],
[
0.509765625,
0.70361328125,
0.80419921875,
0.705078125,
0.80419921875,
0.7197265625,
0.509765625,
0.71875
],
[
0.1175537109375,
0.72265625,
0.48779296875,
0.72265625,
0.48779296875,
0.736328125,
0.1175537109375,
0.736328125
],
[
0.5263671875,
0.736328125,
0.54248046875,
0.736328125,
0.54248046875,
0.74365234375,
0.5263671875,
0.74365234375
],
[
0.54248046875,
0.73291015625,
0.70361328125,
0.73291015625,
0.70361328125,
0.74755859375,
0.54248046875,
0.74755859375
],
[
0.1175537109375,
0.7392578125,
0.490234375,
0.73828125,
0.490234375,
0.7529296875,
0.1175537109375,
0.75439453125
],
[
0.5146484375,
0.73828125,
0.52490234375,
0.7353515625,
0.52880859375,
0.7412109375,
0.51904296875,
0.7451171875
],
[
0.1175537109375,
0.7548828125,
0.491943359375,
0.7548828125,
0.491943359375,
0.7685546875,
0.1175537109375,
0.7685546875
],
[
0.51318359375,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.77392578125,
0.355224609375,
0.77392578125,
0.355224609375,
0.78369140625,
0.118408203125,
0.78369140625
],
[
0.5146484375,
0.775390625,
0.88330078125,
0.775390625,
0.88330078125,
0.7890625,
0.5146484375,
0.7890625
],
[
0.1368408203125,
0.7880859375,
0.48779296875,
0.7880859375,
0.48779296875,
0.8017578125,
0.1368408203125,
0.8017578125
],
[
0.51318359375,
0.79150390625,
0.8427734375,
0.79150390625,
0.8427734375,
0.8046875,
0.51318359375,
0.8046875
],
[
0.118408203125,
0.8046875,
0.489501953125,
0.8046875,
0.489501953125,
0.818359375,
0.118408203125,
0.818359375
],
[
0.118408203125,
0.82080078125,
0.489501953125,
0.82080078125,
0.489501953125,
0.83447265625,
0.118408203125,
0.83447265625
],
[
0.51123046875,
0.81787109375,
0.705078125,
0.818359375,
0.705078125,
0.8330078125,
0.51123046875,
0.83203125
],
[
0.118408203125,
0.83837890625,
0.2301025390625,
0.83837890625,
0.2301025390625,
0.849609375,
0.118408203125,
0.849609375
],
[
0.5146484375,
0.8388671875,
0.88330078125,
0.8388671875,
0.88330078125,
0.8525390625,
0.5146484375,
0.8525390625
],
[
0.139404296875,
0.859375,
0.44580078125,
0.859375,
0.44580078125,
0.87353515625,
0.139404296875,
0.87353515625
],
[
0.51318359375,
0.85546875,
0.884765625,
0.85546875,
0.884765625,
0.869140625,
0.51318359375,
0.869140625
],
[
0.118408203125,
0.87158203125,
0.385498046875,
0.87158203125,
0.385498046875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.884765625,
0.87255859375,
0.884765625,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.88427734375,
0.470947265625,
0.88427734375,
0.470947265625,
0.8984375,
0.118408203125,
0.8984375
],
[
0.118408203125,
0.89794921875,
0.17626953125,
0.89794921875,
0.17626953125,
0.90771484375,
0.118408203125,
0.90771484375
],
[
0.53076171875,
0.892578125,
0.74658203125,
0.8935546875,
0.74658203125,
0.9111328125,
0.53076171875,
0.91015625
]
] |
[
"Source Sentence (En)",
"Translation (Fr)",
"M/F",
"so is that going to affect my chances of becoming aAlors, est -ce que cela va affecter mes chances de devenir",
"M",
"counselor?",
"conseiller?",
"so is that going to affect my chances of becoming a nurse?Alors, est - ce que cela va affecter mes chances de devenir",
"",
"infirmiere?",
"Table l: An example from our dataset of a minimal pair of English gender-neutral source sentences, translated into",
"two different genders in French. Red (italic) stands for masculine, cyan (normal) stands for feminine.",
"3Automatic Detection of Gender Issues",
"Perturbations using BERT We use BERT as a",
"masked language model to find words which can",
"The class of issues we are interested in are those",
" substitute for the human entity identified in the",
"where translation to a gender-marking language",
"previous filtering step, e.g.doctor → nurse.We",
"exposes a model's gender preference for a personal",
"aim to get natural-sounding output and maintain",
"reference. The examples we find that demonstrate",
"extensibility, and thus do not use predefined",
"this are English sentence-pairs, a minimal pair",
"substitutions. We cap our search to the first 100",
"differing by only a single word, e.g. doctor being",
"candidates BERT returns, accepting the first 10",
"replaced by nurse. In each of our examples, this",
"which are tagged as person, and for which the",
"minimal perturbation does not change the gender",
"resulting sentences also pass the filtering step.",
"of the source but gives rise to gender differences",
"upon translation, e.g. doctor becoming masculine",
"Translation We translate each of the generated",
"and nurse feminine.",
"sentences into our target languages using Google",
"In this section, we present a simple, extensible",
"Translate?. A doctor/nurse works in a hospital -→",
"method to mine such examples from real-world text.",
"Un doctor/Una enfermera trabaja en un hospital.",
"Our method does not require expansive manually-",
"Alignment We align tokens in the original and",
"curated word lists for each target language, which",
"translated sentences using fast-align (Dyer et al.,",
"enables us to discover new kinds of entities that",
"2010). This is needed in order to know which token",
"are susceptible to model bias but are not usually",
"in the translation output is the focus entity in the",
"thought of this way. Indeed, while we demonstrate",
"source sentence, whose gender we want to analyze.",
"its utility with nouns with four target languages,",
"our method is naturally extensible to new language",
"Gender Identification",
"We use a morphological",
"pairs and parts of speech with no change in design.",
"analyzer, implemented following Kong et al. (2017),",
"to tag the gender of the target word.",
"Filtering source sentences Our first step is to",
"identify sentences that are gender neutral and that",
"Identifying Examples The final step of our",
"include a single human entity, e.g. A doctor works in",
"pipeline is identifying pairs of sentences to include",
"a hospital. We focus on human entities since these",
"in our dataset, pairs where different genders are",
"have been the target of previous studies and present",
"assigned to the human entity. Our example would",
"the largest risk of gender issues in translation.",
"be included since doctor is translated withthe",
"We use a BERT-based Named Entity Recognition",
"masculine form Un doctor while nurse is translated",
"(NER) model that identifies human entities, and",
"with the feminine form Una enfermera.",
"exclude sentences that have more than one token",
"",
"Challenge Dataset",
"tagged as such. We also remove sentences in which",
"4",
"the entity is a gendered term in English' (e.g.mother,",
"We compile our final dataset from the output of this",
"nephew),a name,or not a noun.",
"pipeline, and explore its properties to understand",
"Note that all the sentences we get are naturally",
"the issues it represents for deployed systems.",
"occurring sentences, and that we do not use any",
"templates or predefined lists of target words that we",
"4.1Random Sampling",
"want to handle.",
"In our final dataset, we include both examples that",
"'https://github.com/tolga-b/debiaswe/blob/",
"passed the final example identification step above",
"master/data/gender_specific_full.json",
"(pairs referred to as “at risk\"), as well as a random",
"https://github.com/uclanlp/gn_glove/tree/master/",
"wordlist",
"2https:/translate.google.com/"
] |
{
"arxiv_id": "2004.14065",
"language": "en",
"page_idx": 1
}
|
||
2004.14065-en-2
|
2004.14065-en-2.jpg
|
selection that did not (“not at risk”). We do this in order to not be constrained too heavily by our choice of translation model; if we did not, we would have no chance of inspecting examples that our system did not spot as at risk but other models might have.
4.2 Fixed Grammatical Gender Rating
When we inspected the examples identified as at risk by our pipeline, the major source of error we found pertained to the issue of fixed grammatical gender. Consider the example in Figure 1:
$\boxed{\text{Sentence 1:}}$
En: you don't have to be the victim in whatever.
Fr: vous ne devez pas être la victime de quoi que ce soit.
Sentence 2:
En: you don't have to be the expert in whatever.
Fr: vous ne devez pas être $l'expert$ en quoi que ce soit.
Figure 1: An example from our dataset, with fixed grammatical gender. Red (italic) stands for masculine, cyan (normal) stands for feminine.
In this example, the word victim in the first English sentence is identified by our tagger as a human entity. However, its French translation victime is feminine by definition, and cannot be assigned another gender regardless of the context, causing a false positive result.
We attempted to filter these examples automatically but came across a number of challenges. Most critically, we found no highquality, comprehensive dictionary that included the required information for all languages, and heuristics we applied were noisy and not reliable.$^3$ We observed that the underlying reason for these challenges was that there is no closed list of grammatically-fixed words as languages are evolving to be more gender-inclusive. In order to maximize and guarantee data quality, and to be sensitive to the nuances of language change, we decided to add a manual filtering step after our pipeline to select the positive (at risk) examples.
We note that the problem of fxed grammatical gender is particular to nouns. Our pipeline is naturally extensible across parts of speech and we would not expect the same issues in future work perturbing adjectives or verbs.
We tried both using a morphological lexicon and a predefined word list in English. Both methods performed poorly, filtering too many or too few sentence pairs, respectively.
## 4.3 Dataset Statistics
To create our dataset we mine text from the subreddit “career”.$^4$ From 29,330 sentences, we found 4,016 which referred to a single, non-gendered human entity. Introducing perturbations with BERT into these 4,016 sentences yielded 40,160 pairs. Out of those, 592 to 1,012 pairs are identified as at risk by our pipeline, depending on the target language. We asked humans to manually identify 100 true at risk examples for the final dataset, which was achieved for all languages except Russian, where we have 59 pairs.$^{5}$ To this 100, we add a further 100 randomly sampled negative examples for each language. Table 2 shows a representative example for each language-pair.
## 4.4 Exploratory Analysis
Table 3 lists the most frequent focus personal references in each language-pair among the positive (at risk) and negative (not at risk) examples, along with the ratio between times the reference form was translated as masculine compared to feminine. Words with extreme values of this ratio indicate cases where a model has a systematic preference for one gender over another, i.e. a gendered representation.
Among the negative examples, we see a prior for masculine translations across all terms. Positive examples break from this prior by exposing reference forms with a feminine preference: nurse and secretary are the most consistently feminine forms, consistent with the Bureau of Labor statistics used in previous work (Caliskan et al.,2017).
Figure 2 shows two sentence pairs that appear as positive examples across all four language-pairs. Two of the three forms, nurse and mechanic, are consistent with the gender statistics of Caliskan et al.; the association of $f$ fighter with the masculine gender is a new discovery of our method.
## 5 Related Work
Our study builds on the literature around gender bias in machine translation. Cho et al.(2019) use sentence templates to probe for differences in Korean pronouns. Prates et al. (2019) and Stanovsky et al. (2019) also use sentence templates, but filled with word lists, of professions and adjectives in the
$^{4}$https://www.reddit.com/r/Career/
$^{5}$Out of the pairs identified using our pipeline, between 1/3-1/10 were selected by the annotators, depending on the $language.$
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07659912109375,
0.69287109375,
0.07659912109375,
0.69287109375,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.1175537109375,
0.0926513671875,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.1175537109375,
0.105712890625
],
[
0.509765625,
0.097412109375,
0.884765625,
0.097412109375,
0.884765625,
0.11102294921875,
0.509765625,
0.11102294921875
],
[
0.12005615234375,
0.10986328125,
0.489501953125,
0.10986328125,
0.489501953125,
0.12054443359375,
0.12005615234375,
0.12054443359375
],
[
0.5087890625,
0.1134033203125,
0.8818359375,
0.1134033203125,
0.8818359375,
0.127685546875,
0.5087890625,
0.127685546875
],
[
0.118408203125,
0.12469482421875,
0.490234375,
0.12469482421875,
0.490234375,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.51318359375,
0.1312255859375,
0.884765625,
0.1312255859375,
0.884765625,
0.144287109375,
0.51318359375,
0.144287109375
],
[
0.118408203125,
0.141357421875,
0.490234375,
0.141357421875,
0.490234375,
0.155029296875,
0.118408203125,
0.155029296875
],
[
0.51318359375,
0.146728515625,
0.884765625,
0.146728515625,
0.884765625,
0.160888671875,
0.51318359375,
0.160888671875
],
[
0.5146484375,
0.163330078125,
0.884765625,
0.163330078125,
0.884765625,
0.177001953125,
0.5146484375,
0.177001953125
],
[
0.118408203125,
0.174560546875,
0.4375,
0.17578125,
0.4375,
0.190673828125,
0.118408203125,
0.189453125
],
[
0.5146484375,
0.1787109375,
0.88330078125,
0.1787109375,
0.88330078125,
0.1923828125,
0.5146484375,
0.1923828125
],
[
0.51318359375,
0.19482421875,
0.884765625,
0.19482421875,
0.884765625,
0.2078857421875,
0.51318359375,
0.2078857421875
],
[
0.1175537109375,
0.201904296875,
0.489501953125,
0.20068359375,
0.489501953125,
0.215576171875,
0.1175537109375,
0.2161865234375
],
[
0.51123046875,
0.210205078125,
0.884765625,
0.208984375,
0.884765625,
0.223876953125,
0.51123046875,
0.2244873046875
],
[
0.12005615234375,
0.2197265625,
0.489501953125,
0.2197265625,
0.489501953125,
0.2327880859375,
0.12005615234375,
0.2327880859375
],
[
0.5146484375,
0.226806640625,
0.884765625,
0.226806640625,
0.884765625,
0.240478515625,
0.5146484375,
0.240478515625
],
[
0.12005615234375,
0.235107421875,
0.489501953125,
0.235107421875,
0.489501953125,
0.248779296875,
0.12005615234375,
0.248779296875
],
[
0.51318359375,
0.243408203125,
0.8818359375,
0.243408203125,
0.8818359375,
0.257080078125,
0.51318359375,
0.257080078125
],
[
0.12005615234375,
0.25048828125,
0.429931640625,
0.25048828125,
0.429931640625,
0.26416015625,
0.12005615234375,
0.26416015625
],
[
0.51318359375,
0.25830078125,
0.88330078125,
0.25830078125,
0.88330078125,
0.2724609375,
0.51318359375,
0.2724609375
],
[
0.5146484375,
0.273681640625,
0.8818359375,
0.273681640625,
0.8818359375,
0.287353515625,
0.5146484375,
0.287353515625
],
[
0.124267578125,
0.283935546875,
0.193115234375,
0.283935546875,
0.193115234375,
0.295654296875,
0.124267578125,
0.295654296875
],
[
0.1368408203125,
0.295654296875,
0.40478515625,
0.295654296875,
0.40478515625,
0.308837890625,
0.1368408203125,
0.308837890625
],
[
0.51318359375,
0.29150390625,
0.88330078125,
0.29150390625,
0.88330078125,
0.3046875,
0.51318359375,
0.3046875
],
[
0.1368408203125,
0.308837890625,
0.44580078125,
0.308837890625,
0.44580078125,
0.319580078125,
0.1368408203125,
0.319580078125
],
[
0.51318359375,
0.30810546875,
0.68408203125,
0.30810546875,
0.68408203125,
0.32177734375,
0.51318359375,
0.32177734375
],
[
0.1259765625,
0.320556640625,
0.193115234375,
0.320556640625,
0.193115234375,
0.33203125,
0.1259765625,
0.33203125
],
[
0.1376953125,
0.333251953125,
0.403076171875,
0.333251953125,
0.403076171875,
0.34326171875,
0.1376953125,
0.34326171875
],
[
0.51318359375,
0.3349609375,
0.724609375,
0.3349609375,
0.724609375,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.1351318359375,
0.345703125,
0.435791015625,
0.345703125,
0.435791015625,
0.355712890625,
0.1351318359375,
0.355712890625
],
[
0.51318359375,
0.355712890625,
0.8818359375,
0.355712890625,
0.8818359375,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.12005615234375,
0.369384765625,
0.489501953125,
0.369384765625,
0.489501953125,
0.383056640625,
0.12005615234375,
0.383056640625
],
[
0.51123046875,
0.372314453125,
0.884765625,
0.37353515625,
0.884765625,
0.388427734375,
0.51123046875,
0.38720703125
],
[
0.12005615234375,
0.38427734375,
0.490234375,
0.38427734375,
0.490234375,
0.39794921875,
0.12005615234375,
0.39794921875
],
[
0.51318359375,
0.39013671875,
0.88330078125,
0.39013671875,
0.88330078125,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.1175537109375,
0.3984375,
0.353515625,
0.3984375,
0.353515625,
0.412109375,
0.1175537109375,
0.412109375
],
[
0.51318359375,
0.40625,
0.884765625,
0.40625,
0.884765625,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.51123046875,
0.4228515625,
0.884765625,
0.4228515625,
0.884765625,
0.435791015625,
0.51123046875,
0.435791015625
],
[
0.1351318359375,
0.43408203125,
0.489501953125,
0.432861328125,
0.489501953125,
0.44775390625,
0.1351318359375,
0.448974609375
],
[
0.51123046875,
0.438232421875,
0.884765625,
0.438232421875,
0.884765625,
0.451904296875,
0.51123046875,
0.451904296875
],
[
0.118408203125,
0.449462890625,
0.489501953125,
0.45068359375,
0.489501953125,
0.465576171875,
0.118408203125,
0.46435546875
],
[
0.51318359375,
0.454833984375,
0.884765625,
0.454833984375,
0.884765625,
0.468505859375,
0.51318359375,
0.468505859375
],
[
0.12005615234375,
0.468505859375,
0.489501953125,
0.468505859375,
0.489501953125,
0.478515625,
0.12005615234375,
0.478515625
],
[
0.51318359375,
0.47021484375,
0.884765625,
0.47021484375,
0.884765625,
0.485107421875,
0.51318359375,
0.485107421875
],
[
0.12005615234375,
0.485107421875,
0.48779296875,
0.485107421875,
0.48779296875,
0.495361328125,
0.12005615234375,
0.495361328125
],
[
0.51123046875,
0.48681640625,
0.62890625,
0.48681640625,
0.62890625,
0.50048828125,
0.51123046875,
0.50048828125
],
[
0.118408203125,
0.49951171875,
0.490234375,
0.49951171875,
0.490234375,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.53076171875,
0.5029296875,
0.8818359375,
0.5029296875,
0.8818359375,
0.5166015625,
0.53076171875,
0.5166015625
],
[
0.12005615234375,
0.5166015625,
0.340087890625,
0.5166015625,
0.340087890625,
0.52978515625,
0.12005615234375,
0.52978515625
],
[
0.51318359375,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.51318359375,
0.533203125
],
[
0.1368408203125,
0.5322265625,
0.489501953125,
0.533203125,
0.489501953125,
0.54736328125,
0.1368408203125,
0.54638671875
],
[
0.51318359375,
0.5361328125,
0.88330078125,
0.5361328125,
0.88330078125,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.12005615234375,
0.5517578125,
0.489501953125,
0.5517578125,
0.489501953125,
0.5625,
0.12005615234375,
0.5625
],
[
0.51318359375,
0.5517578125,
0.88330078125,
0.5517578125,
0.88330078125,
0.5654296875,
0.51318359375,
0.5654296875
],
[
0.12005615234375,
0.56640625,
0.490234375,
0.56640625,
0.490234375,
0.580078125,
0.12005615234375,
0.580078125
],
[
0.51318359375,
0.56787109375,
0.88330078125,
0.56787109375,
0.88330078125,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.1175537109375,
0.58203125,
0.489501953125,
0.58056640625,
0.489501953125,
0.595703125,
0.1175537109375,
0.5966796875
],
[
0.51123046875,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.51123046875,
0.595703125
],
[
0.118408203125,
0.59716796875,
0.489501953125,
0.5986328125,
0.489501953125,
0.61328125,
0.118408203125,
0.6123046875
],
[
0.51318359375,
0.599609375,
0.84814453125,
0.599609375,
0.84814453125,
0.61328125,
0.51318359375,
0.61328125
],
[
0.1175537109375,
0.6142578125,
0.48779296875,
0.61328125,
0.48779296875,
0.62744140625,
0.1175537109375,
0.62890625
],
[
0.53076171875,
0.615234375,
0.88330078125,
0.6162109375,
0.88330078125,
0.63134765625,
0.53076171875,
0.6298828125
],
[
0.118408203125,
0.63134765625,
0.490234375,
0.63134765625,
0.490234375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.51123046875,
0.6318359375,
0.884765625,
0.6318359375,
0.884765625,
0.6455078125,
0.51123046875,
0.6455078125
],
[
0.1175537109375,
0.646484375,
0.490234375,
0.6455078125,
0.490234375,
0.66015625,
0.1175537109375,
0.6611328125
],
[
0.51318359375,
0.6484375,
0.88330078125,
0.6484375,
0.88330078125,
0.662109375,
0.51318359375,
0.662109375
],
[
0.118408203125,
0.6630859375,
0.490234375,
0.6630859375,
0.490234375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.509765625,
0.66455078125,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.509765625,
0.6787109375
],
[
0.118408203125,
0.6787109375,
0.490234375,
0.6787109375,
0.490234375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.51318359375,
0.68115234375,
0.884765625,
0.68115234375,
0.884765625,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.118408203125,
0.6953125,
0.489501953125,
0.6953125,
0.489501953125,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51123046875,
0.69677734375,
0.81787109375,
0.6953125,
0.81787109375,
0.71044921875,
0.51123046875,
0.71142578125
],
[
0.118408203125,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.118408203125,
0.724609375
],
[
0.118408203125,
0.7255859375,
0.489501953125,
0.72705078125,
0.489501953125,
0.7412109375,
0.118408203125,
0.74072265625
],
[
0.51708984375,
0.72998046875,
0.5263671875,
0.728515625,
0.5283203125,
0.7353515625,
0.51904296875,
0.736328125
],
[
0.5224609375,
0.724609375,
0.66650390625,
0.7255859375,
0.66650390625,
0.74072265625,
0.5224609375,
0.7392578125
],
[
0.12005615234375,
0.7421875,
0.47265625,
0.7421875,
0.47265625,
0.755859375,
0.12005615234375,
0.755859375
],
[
0.5146484375,
0.7529296875,
0.88330078125,
0.7529296875,
0.88330078125,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.1368408203125,
0.76025390625,
0.489501953125,
0.76123046875,
0.489501953125,
0.775390625,
0.1368408203125,
0.77490234375
],
[
0.51318359375,
0.76953125,
0.88330078125,
0.76953125,
0.88330078125,
0.783203125,
0.51318359375,
0.783203125
],
[
0.118408203125,
0.779296875,
0.489501953125,
0.779296875,
0.489501953125,
0.7919921875,
0.118408203125,
0.7919921875
],
[
0.51318359375,
0.78369140625,
0.8818359375,
0.78369140625,
0.8818359375,
0.79736328125,
0.51318359375,
0.79736328125
],
[
0.118408203125,
0.79443359375,
0.489501953125,
0.79443359375,
0.489501953125,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.51123046875,
0.80029296875,
0.8818359375,
0.7998046875,
0.8818359375,
0.81396484375,
0.51123046875,
0.8154296875
],
[
0.118408203125,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51318359375,
0.81787109375,
0.88330078125,
0.81787109375,
0.88330078125,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.12005615234375,
0.82666015625,
0.344970703125,
0.82666015625,
0.344970703125,
0.84033203125,
0.12005615234375,
0.84033203125
],
[
0.51123046875,
0.8330078125,
0.8818359375,
0.8330078125,
0.8818359375,
0.8466796875,
0.51123046875,
0.8466796875
],
[
0.1376953125,
0.85791015625,
0.490234375,
0.859375,
0.490234375,
0.87353515625,
0.1376953125,
0.87255859375
],
[
0.5322265625,
0.85693359375,
0.78271484375,
0.85791015625,
0.78271484375,
0.87255859375,
0.5322265625,
0.87158203125
],
[
0.118408203125,
0.87158203125,
0.490234375,
0.87158203125,
0.490234375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.5322265625,
0.869140625,
0.884765625,
0.87060546875,
0.884765625,
0.88525390625,
0.5322265625,
0.88427734375
],
[
0.1175537109375,
0.8828125,
0.49365234375,
0.88427734375,
0.49365234375,
0.8984375,
0.1175537109375,
0.89794921875
],
[
0.51318359375,
0.8828125,
0.88671875,
0.88525390625,
0.88671875,
0.8984375,
0.51318359375,
0.896484375
],
[
0.118408203125,
0.89794921875,
0.1956787109375,
0.89794921875,
0.1956787109375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.89794921875,
0.57421875,
0.89794921875,
0.57421875,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"selection that did not (\"not at risk\"). We do this in",
"4.3Dataset Statistics",
"ordertonotbeconstrainedtooheavilybyourchoice",
"To create our dataset we mine text from the subreddit",
"of translationmodel:ifwedidnot,wewouldhave",
"\"career\"4 From 29,330 sentences, we found 4,016",
"no chance of inspecting examples that our system",
"which referred to a single, non-gendered human",
"did not spot as at risk but other models might have.",
"entity. Introducing perturbations with BERT into",
"these 4,016 sentences yielded 40,160 pairs. Out of",
"4.2Fixed Grammatical Gender Rating",
"those, 592 to 1,012 pairs are identified as at risk",
"by our pipeline, depending on the target language.",
"When we inspected the examples identified as at",
"We asked humans to manually identify 100 true",
"risk by our pipeline, the major source of error we",
"at risk examples for the final dataset, which was",
"found pertained to the issue of fixed grammatical",
"achieved for all languages except Russian, where",
"gender. Consider the example in Figure l:",
"we have 59 pairs.5 To this 100, we add a further",
"100 randomly sampled negative examples for each",
"Sentence l:",
"En: you don't have to be the victim in whatever.",
"language. Table 2 shows a representative example",
"Fr: vous nedevez pas étre la victime de quoi que ce soit",
"for each language-pair.",
"Sentence 2:",
"En: you don't have to be the expert in whatever.",
"4.4Exploratory Analysis",
"Fr:vous ne devez pas étre I'expert en quoique ce soit.",
"Table 3 lists the most frequent focus personal",
"Figure I: An example from our dataset, with fixed",
"references in each language-pair among the positive",
"grammatical gender. Red (italic) stands for masculine,",
"(at risk) and negative (not at risk) examples, along",
"cyan (normal) stands for feminine.",
"with the ratio between times the reference form",
"was translated as masculine compared to feminine.",
"In this example, the word victim in the first",
"Words with extreme values of this ratio indicate",
"English sentence is identified by our tagger as",
"cases where a model has a systematic preference",
"a human entity.However, its French translation",
"for one gender over another, i.e. a gendered",
"victime is feminine by definition. and cannot be",
"representation.",
"assigned another gender regardless of the context,",
"Among the negative examples, we see a prior for",
"causing a false positive result.",
"masculine translations across all terms. Positive",
"We attempted to filter these examples",
"examples break from this prior by exposing",
"automatically but came across a number of",
"reference forms with a feminine preference: nurse",
"challenges. Most critically, we found no high-",
"and secretary are the most consistently feminine",
"quality, comprehensive dictionary that included",
"forms, consistent with the Bureau of Labor statistics",
"the required information for all languages, and",
"used in previous work (Caliskan et al., 2017).",
"heuristics we applied were noisy and not reliable.3",
"Figure 2 shows two sentence pairs that appear",
"We observed that the underlying reason for",
"as positive examples across all four language-pairs.",
"these challenges was that there is no closed list",
"Two of the three forms, nurse and mechanic, are",
"of grammatically-fixed words as languages are",
" consistent with the gender statistics of Caliskan",
"evolving to be more gender-inclusive. In order to",
"et al. the association of fighter with the masculine",
"maximize and guarantee data quality, and to be",
"gender is a new discovery of our method.",
"sensitive to the nuances of language change, we",
"decided to add a manual filtering step after our",
"5",
"RelatedWork",
"pipeline to select the positive (at risk) examples.",
"Our study builds on the literature around gender",
"We note that the problem of fixed grammatical",
"bias in machine translation. Cho et al. (2019)",
"gender is particular to nouns. Our pipeline is",
"use sentence templates to probe for differences in",
"naturally extensible across parts of speech and",
"Korean pronouns. Prates et al. (2019) and Stanovsky",
"we would not expect the same issues in future work",
"et al. (2019) also use sentence templates, but filled",
"perturbing adjectives or verbs.",
"with word lists, of professions and adjectives in the",
"3we tried both using a morphological lexicon and a",
"*https://www.reddit.com/r/Career/",
"predefined word list in English. Both methods performed",
"Sout of the pairs identified using our pipeline, between",
"poorly,filtering too many or too few sentence pairs,",
"1/3-1/10 were selected by the annotators, depending on the",
"respectively.",
"language."
] |
{
"arxiv_id": "2004.14065",
"language": "en",
"page_idx": 2
}
|
||
2004.14065-en-3
|
2004.14065-en-3.jpg
|
<table>
<tbody>
<tr>
<th rowspan="3">$Fr$</th>
<th>Source Sentence (En)</th>
<th>Translation</th>
<th>$M/F$</th>
</tr>
<tr>
<th>also should i ask the manager what the pay would be if i got the job prior to flying out?</th>
<th>De plus, devrais- je demander au gestionnaire quel serait le salaire si je obtenais le poste avant de prendre $l^{\prime }$avion? </th>
<th>$M$</th>
</tr>
<tr>
<th>also should i ask the secretary what the pay would be ifigot the job prior to flying out?</th>
<th>De plus, devrais- je demander à la secrétaire quel serait le salaire si je obtenais le poste avant mon départ?</th>
<th>$F$</th>
</tr>
<tr>
<td>De</td>
<td>currently thinking about learning a trade (mostly a $\mathbf{electrician}).$ currently thinking about learning a trade (mostly a cook</td>
<td>Derzeit über das Erlernen eines Gewerbes nachdenkem (meistens $Elektriker).$ Derzeit über das Erlernen eines Gewerbes nachdenkem (meistens eine Köchin</td>
<td>M $F$</td>
</tr>
<tr>
<td rowspan="2">$Es$</td>
<td>- decided to become a teacher: spent a year working 2 jobs and doing prerequisites for a masters in education.</td>
<td>- Decidí ser maestra: pasé un año trabajando en 2 trabajos y haciendo requisitos previos para una maestría vaciór</td>
<td>$F$</td>
</tr>
<tr>
<td>- decided to become a lecturer: spent a year working 2 jobs and doing prerequisites for a masters in education.</td>
<td>en educación. - Decidí ser $profesor: \text{pas\'{e} un a\~{n} o trabajando en}2$ trabajos y haciendo requisitos previos para una maestría en educación.</td>
<td>$M$</td>
</tr>
<tr>
<td rowspan="2">$Ru$</td>
<td>i read about a psychologist who upgraded into</td>
<td>$S$ читал о психологе, который превратился в Мd.</td>
<td>M</td>
</tr>
<tr>
<td>i read about a nurse who upgraded into becoming a m.d.</td>
<td>${\text{Я читал о медсестре, которая превратилась в док-}}$ Topa медицины.</td>
<td>$F$</td>
</tr>
</tbody>
</table>
Table 2: Examples from our dataset of a minimal pair of English gender-neutral source sentences, translated into
two different genders in all target languages. Red (italic) stands for masculine, cyan (normal) stands for feminine.
<table>
<tbody>
<tr>
<td>1 T $i$ C</td>
<td>tam ubstitution: you need to have experience working witl aulic lifts, & they like to see that you've worked or trained a $\mathbf{lrse}.$</td>
</tr>
<tr>
<td>$S$</td>
<td>cence pair 2:</td>
</tr>
<tr>
<td> </td>
<td>murse</td>
</tr>
<tr>
<td> </td>
<td>hip</td>
</tr>
</tbody>
</table>
Figure 2: Two sentence pairs from our dataset that found
to be shared between all four target languages.
<table>
<tbody>
<tr>
<th> </th>
<th>Positive</th>
<th>$M:F$</th>
<th>Negative</th>
<th>$M:F$</th>
</tr>
<tr>
<td rowspan="5">$Fr$</td>
<td>$\operatorname{nurse}$</td>
<td>0{:}36</td>
<td>manager</td>
<td>685:1</td>
</tr>
<tr>
<td>secretary</td>
<td>0{:}17</td>
<td>employee</td>
<td>406:0</td>
</tr>
<tr>
<td>teacher</td>
<td>7{:}1</td>
<td>employees</td>
<td>364:0</td>
</tr>
<tr>
<td>assistant</td>
<td>1:7</td>
<td>parents</td>
<td>353:0</td>
</tr>
<tr>
<td>manager</td>
<td>8{:}0</td>
<td>teacher</td>
<td>337:0</td>
</tr>
<tr>
<td rowspan="5">De</td>
<td>secretary</td>
<td>0{:}27</td>
<td>manager</td>
<td>594:0</td>
</tr>
<tr>
<td>nurse</td>
<td>0{:}21</td>
<td>employees</td>
<td>409:1</td>
</tr>
<tr>
<td>teacher</td>
<td>3:7</td>
<td>friends</td>
<td>359:0</td>
</tr>
<tr>
<td>receptionist</td>
<td>0:9</td>
<td>employee</td>
<td>320:0</td>
</tr>
<tr>
<td>manager</td>
<td>7{:}0</td>
<td>students</td>
<td>316:0</td>
</tr>
<tr>
<td rowspan="5">Es</td>
<td>teacher</td>
<td>4{:}29</td>
<td>manager</td>
<td>691:0</td>
</tr>
<tr>
<td>nurse</td>
<td>0{:}31</td>
<td>employee</td>
<td>446:0</td>
</tr>
<tr>
<td>secretary</td>
<td>0{:}26</td>
<td>friends</td>
<td>380:0</td>
</tr>
<tr>
<td>${\mathbf{writer}}$</td>
<td>8:0</td>
<td>parents</td>
<td>374:0</td>
</tr>
<tr>
<td>employee</td>
<td>5{:}0</td>
<td>supervisor</td>
<td>345:0</td>
</tr>
<tr>
<td rowspan="5">$Ru$</td>
<td>nurse</td>
<td>0{:}32</td>
<td>manager</td>
<td>713:0</td>
</tr>
<tr>
<td>babysitter</td>
<td>0{:}13</td>
<td>employees</td>
<td>519:0</td>
</tr>
<tr>
<td>nurses</td>
<td>0:5</td>
<td>friends</td>
<td>439:0</td>
</tr>
<tr>
<td>dishwasher</td>
<td>0:4</td>
<td>students</td>
<td>417:0</td>
</tr>
<tr>
<td>technician</td>
<td>3:0</td>
<td>employee</td>
<td>392:0</td>
</tr>
</tbody>
</table>
former, and professions in the latter. A separate but related line of work focuses on generating correct inflections when translating to gender-marking languages (Vanmassenhove et al., 2018; Moryossef et al., 2019).
## 6 Conclusion
Table 3: Top five human reference forms in our dataset, and their ratio of times they are translated as masculine compared to feminine. Positive indicates that the examples were taken from the at-risk group from our pipeline, and negative from the random sample among the not at-risk group.
The primary contribution of our work is a novel, automatic method for identifying gender issues in machine translation. By performing BERT-based perturbations on naturally-occurring sentences, we are able to identify sentence pairs that behave differently upon translation to gender-marking languages. We demonstrate our technique over human reference forms and discover new sources of risk beyond the word lists used previously. Furthermore, the novelty of our approach is its
|
[
[
0.50390625,
0.0736083984375,
0.580078125,
0.0736083984375,
0.580078125,
0.0843505859375,
0.50390625,
0.0843505859375
],
[
0.8427734375,
0.0736083984375,
0.8818359375,
0.0736083984375,
0.8818359375,
0.08612060546875,
0.8427734375,
0.08612060546875
],
[
0.162109375,
0.0885009765625,
0.49951171875,
0.08905029296875,
0.49951171875,
0.10394287109375,
0.162109375,
0.10272216796875
],
[
0.160400390625,
0.0748291015625,
0.300537109375,
0.0748291015625,
0.300537109375,
0.08489990234375,
0.160400390625,
0.08489990234375
],
[
0.50390625,
0.08905029296875,
0.84033203125,
0.09027099609375,
0.84033203125,
0.10394287109375,
0.50390625,
0.10272216796875
],
[
0.84814453125,
0.09027099609375,
0.86962890625,
0.09027099609375,
0.86962890625,
0.10272216796875,
0.84814453125,
0.10272216796875
],
[
0.162841796875,
0.10272216796875,
0.3779296875,
0.10272216796875,
0.3779296875,
0.11639404296875,
0.162841796875,
0.11639404296875
],
[
0.5087890625,
0.1051025390625,
0.8369140625,
0.1051025390625,
0.8369140625,
0.11517333984375,
0.5087890625,
0.11517333984375
],
[
0.139404296875,
0.1217041015625,
0.151123046875,
0.1300048828125,
0.139404296875,
0.1383056640625,
0.1275634765625,
0.1300048828125
],
[
0.50732421875,
0.1175537109375,
0.5615234375,
0.1175537109375,
0.5615234375,
0.12890625,
0.50732421875,
0.12890625
],
[
0.166259765625,
0.1300048828125,
0.49951171875,
0.1300048828125,
0.49951171875,
0.144287109375,
0.166259765625,
0.144287109375
],
[
0.50537109375,
0.1300048828125,
0.84033203125,
0.1300048828125,
0.84033203125,
0.1436767578125,
0.50537109375,
0.1436767578125
],
[
0.85205078125,
0.12890625,
0.86572265625,
0.133056640625,
0.861328125,
0.142578125,
0.84619140625,
0.1383056640625
],
[
0.497802734375,
0.1353759765625,
0.50537109375,
0.1353759765625,
0.50537109375,
0.1383056640625,
0.497802734375,
0.1383056640625
],
[
0.16455078125,
0.1436767578125,
0.378662109375,
0.1436767578125,
0.378662109375,
0.15673828125,
0.16455078125,
0.15673828125
],
[
0.50537109375,
0.1436767578125,
0.8193359375,
0.1436767578125,
0.8193359375,
0.15673828125,
0.50537109375,
0.15673828125
],
[
0.16455078125,
0.160888671875,
0.51123046875,
0.160888671875,
0.51123046875,
0.171630859375,
0.16455078125,
0.171630859375
],
[
0.50537109375,
0.160888671875,
0.84033203125,
0.160888671875,
0.84033203125,
0.171630859375,
0.50537109375,
0.171630859375
],
[
0.8447265625,
0.1591796875,
0.87158203125,
0.162109375,
0.869140625,
0.1728515625,
0.84228515625,
0.169189453125
],
[
0.16455078125,
0.174560546875,
0.2451171875,
0.174560546875,
0.2451171875,
0.185302734375,
0.16455078125,
0.185302734375
],
[
0.50732421875,
0.17333984375,
0.64404296875,
0.17333984375,
0.64404296875,
0.18408203125,
0.50732421875,
0.18408203125
],
[
0.1318359375,
0.179931640625,
0.15283203125,
0.179931640625,
0.15283203125,
0.1923828125,
0.1318359375,
0.1923828125
],
[
0.16455078125,
0.18701171875,
0.50146484375,
0.18701171875,
0.50146484375,
0.20068359375,
0.16455078125,
0.20068359375
],
[
0.5087890625,
0.188232421875,
0.8388671875,
0.188232421875,
0.8388671875,
0.198974609375,
0.5087890625,
0.198974609375
],
[
0.8505859375,
0.188232421875,
0.8623046875,
0.188232421875,
0.8623046875,
0.19775390625,
0.8505859375,
0.19775390625
],
[
0.16455078125,
0.20068359375,
0.208251953125,
0.20068359375,
0.208251953125,
0.2120361328125,
0.16455078125,
0.2120361328125
],
[
0.50732421875,
0.201904296875,
0.65576171875,
0.201904296875,
0.65576171875,
0.2120361328125,
0.50732421875,
0.2120361328125
],
[
0.1695556640625,
0.218505859375,
0.495361328125,
0.218505859375,
0.495361328125,
0.2286376953125,
0.1695556640625,
0.2286376953125
],
[
0.50732421875,
0.218505859375,
0.8388671875,
0.218505859375,
0.8388671875,
0.2286376953125,
0.50732421875,
0.2286376953125
],
[
0.84228515625,
0.2197265625,
0.8623046875,
0.2197265625,
0.8623046875,
0.22802734375,
0.84228515625,
0.22802734375
],
[
0.162841796875,
0.2298583984375,
0.497802734375,
0.2298583984375,
0.497802734375,
0.243408203125,
0.162841796875,
0.243408203125
],
[
0.50732421875,
0.2298583984375,
0.84033203125,
0.2298583984375,
0.84033203125,
0.243408203125,
0.50732421875,
0.243408203125
],
[
0.5087890625,
0.24462890625,
0.59375,
0.24462890625,
0.59375,
0.255859375,
0.5087890625,
0.255859375
],
[
0.130126953125,
0.25048828125,
0.151123046875,
0.25048828125,
0.151123046875,
0.26318359375,
0.130126953125,
0.26318359375
],
[
0.17041015625,
0.254638671875,
0.497802734375,
0.255859375,
0.497802734375,
0.26953125,
0.17041015625,
0.268310546875
],
[
0.50732421875,
0.25830078125,
0.8388671875,
0.25830078125,
0.8388671875,
0.268310546875,
0.50732421875,
0.268310546875
],
[
0.8486328125,
0.257080078125,
0.86962890625,
0.257080078125,
0.86962890625,
0.26953125,
0.8486328125,
0.26953125
],
[
0.166259765625,
0.271484375,
0.495361328125,
0.271484375,
0.495361328125,
0.281982421875,
0.166259765625,
0.281982421875
],
[
0.5087890625,
0.271484375,
0.8388671875,
0.271484375,
0.8388671875,
0.281982421875,
0.5087890625,
0.281982421875
],
[
0.5087890625,
0.28515625,
0.59375,
0.28515625,
0.59375,
0.295654296875,
0.5087890625,
0.295654296875
],
[
0.162109375,
0.3017578125,
0.50146484375,
0.3017578125,
0.50146484375,
0.312255859375,
0.162109375,
0.312255859375
],
[
0.50732421875,
0.3017578125,
0.8251953125,
0.3017578125,
0.8251953125,
0.312255859375,
0.50732421875,
0.312255859375
],
[
0.84814453125,
0.300537109375,
0.87158203125,
0.300537109375,
0.87158203125,
0.3134765625,
0.84814453125,
0.3134765625
],
[
0.130126953125,
0.30712890625,
0.156982421875,
0.30712890625,
0.156982421875,
0.318359375,
0.130126953125,
0.318359375
],
[
0.497802734375,
0.303955078125,
0.50537109375,
0.303955078125,
0.50537109375,
0.308837890625,
0.497802734375,
0.308837890625
],
[
0.16455078125,
0.314208984375,
0.268798828125,
0.314208984375,
0.268798828125,
0.325927734375,
0.16455078125,
0.325927734375
],
[
0.16455078125,
0.32666015625,
0.496337890625,
0.32666015625,
0.496337890625,
0.34033203125,
0.16455078125,
0.34033203125
],
[
0.5087890625,
0.325927734375,
0.8388671875,
0.32666015625,
0.8388671875,
0.34033203125,
0.5087890625,
0.339111328125
],
[
0.85546875,
0.330078125,
0.861328125,
0.330078125,
0.861328125,
0.336181640625,
0.85546875,
0.336181640625
],
[
0.162841796875,
0.34033203125,
0.1956787109375,
0.34033203125,
0.1956787109375,
0.3515625,
0.162841796875,
0.3515625
],
[
0.50732421875,
0.34033203125,
0.611328125,
0.34033203125,
0.611328125,
0.350830078125,
0.50732421875,
0.350830078125
],
[
0.1175537109375,
0.36767578125,
0.88330078125,
0.3681640625,
0.88330078125,
0.3818359375,
0.1175537109375,
0.38134765625
],
[
0.118408203125,
0.383056640625,
0.8759765625,
0.383056640625,
0.8759765625,
0.396728515625,
0.118408203125,
0.396728515625
],
[
0.5205078125,
0.421630859375,
0.611328125,
0.421630859375,
0.611328125,
0.431640625,
0.5205078125,
0.431640625
],
[
0.52880859375,
0.431640625,
0.875,
0.431640625,
0.875,
0.4453125,
0.52880859375,
0.4453125
],
[
0.5224609375,
0.4453125,
0.86962890625,
0.4453125,
0.86962890625,
0.4560546875,
0.5224609375,
0.4560546875
],
[
0.1822509765625,
0.454833984375,
0.2451171875,
0.454833984375,
0.2451171875,
0.4697265625,
0.1822509765625,
0.4697265625
],
[
0.27783203125,
0.45361328125,
0.393798828125,
0.454833984375,
0.393798828125,
0.4697265625,
0.27783203125,
0.468505859375
],
[
0.426513671875,
0.4560546875,
0.4609375,
0.4560546875,
0.4609375,
0.468505859375,
0.426513671875,
0.468505859375
],
[
0.5224609375,
0.457763671875,
0.580078125,
0.457763671875,
0.580078125,
0.468505859375,
0.5224609375,
0.468505859375
],
[
0.41650390625,
0.471435546875,
0.462646484375,
0.471435546875,
0.462646484375,
0.486328125,
0.41650390625,
0.486328125
],
[
0.5341796875,
0.468505859375,
0.8759765625,
0.468505859375,
0.8759765625,
0.482177734375,
0.5341796875,
0.482177734375
],
[
0.279541015625,
0.47265625,
0.31494140625,
0.47265625,
0.31494140625,
0.485107421875,
0.279541015625,
0.485107421875
],
[
0.187255859375,
0.473876953125,
0.2283935546875,
0.473876953125,
0.2283935546875,
0.485107421875,
0.187255859375,
0.485107421875
],
[
0.330078125,
0.473876953125,
0.39208984375,
0.473876953125,
0.39208984375,
0.485107421875,
0.330078125,
0.485107421875
],
[
0.5205078125,
0.479736328125,
0.8759765625,
0.479736328125,
0.8759765625,
0.493408203125,
0.5205078125,
0.493408203125
],
[
0.280517578125,
0.48681640625,
0.31494140625,
0.48681640625,
0.31494140625,
0.498779296875,
0.280517578125,
0.498779296875
],
[
0.4189453125,
0.48681640625,
0.4609375,
0.48681640625,
0.4609375,
0.498779296875,
0.4189453125,
0.498779296875
],
[
0.33154296875,
0.488037109375,
0.39794921875,
0.488037109375,
0.39794921875,
0.49951171875,
0.33154296875,
0.49951171875
],
[
0.187255859375,
0.488037109375,
0.2509765625,
0.488037109375,
0.2509765625,
0.49951171875,
0.187255859375,
0.49951171875
],
[
0.418212890625,
0.49951171875,
0.462646484375,
0.49951171875,
0.462646484375,
0.51416015625,
0.418212890625,
0.51416015625
],
[
0.5224609375,
0.49462890625,
0.568359375,
0.49462890625,
0.568359375,
0.505859375,
0.5224609375,
0.505859375
],
[
0.151123046875,
0.50048828125,
0.172119140625,
0.50048828125,
0.172119140625,
0.51318359375,
0.151123046875,
0.51318359375
],
[
0.28955078125,
0.50048828125,
0.313232421875,
0.50048828125,
0.313232421875,
0.51318359375,
0.28955078125,
0.51318359375
],
[
0.330078125,
0.50048828125,
0.405517578125,
0.50048828125,
0.405517578125,
0.515625,
0.330078125,
0.515625
],
[
0.41650390625,
0.51416015625,
0.462646484375,
0.51416015625,
0.462646484375,
0.529296875,
0.41650390625,
0.529296875
],
[
0.187255859375,
0.501953125,
0.240966796875,
0.501953125,
0.240966796875,
0.51318359375,
0.187255859375,
0.51318359375
],
[
0.187255859375,
0.5166015625,
0.2476806640625,
0.5166015625,
0.2476806640625,
0.52783203125,
0.187255859375,
0.52783203125
],
[
0.291259765625,
0.515625,
0.31494140625,
0.515625,
0.31494140625,
0.52783203125,
0.291259765625,
0.52783203125
],
[
0.330078125,
0.5166015625,
0.3837890625,
0.5166015625,
0.3837890625,
0.52783203125,
0.330078125,
0.52783203125
],
[
0.5224609375,
0.51708984375,
0.61279296875,
0.51708984375,
0.61279296875,
0.529296875,
0.5224609375,
0.529296875
],
[
0.418212890625,
0.529296875,
0.462646484375,
0.529296875,
0.462646484375,
0.54345703125,
0.418212890625,
0.54345703125
],
[
0.53076171875,
0.529296875,
0.8427734375,
0.529296875,
0.8427734375,
0.5419921875,
0.53076171875,
0.5419921875
],
[
0.28955078125,
0.52978515625,
0.316650390625,
0.52978515625,
0.316650390625,
0.5419921875,
0.28955078125,
0.5419921875
],
[
0.185546875,
0.53076171875,
0.2493896484375,
0.53076171875,
0.2493896484375,
0.5419921875,
0.185546875,
0.5419921875
],
[
0.328369140625,
0.53076171875,
0.3837890625,
0.53076171875,
0.3837890625,
0.54150390625,
0.328369140625,
0.54150390625
],
[
0.183837890625,
0.54638671875,
0.252685546875,
0.548828125,
0.252685546875,
0.56103515625,
0.183837890625,
0.55859375
],
[
0.279541015625,
0.54638671875,
0.31494140625,
0.54638671875,
0.31494140625,
0.55810546875,
0.279541015625,
0.55810546875
],
[
0.4189453125,
0.54638671875,
0.4609375,
0.54638671875,
0.4609375,
0.55810546875,
0.4189453125,
0.55810546875
],
[
0.53076171875,
0.5419921875,
0.869140625,
0.5419921875,
0.869140625,
0.552734375,
0.53076171875,
0.552734375
],
[
0.41650390625,
0.55859375,
0.4609375,
0.55859375,
0.4609375,
0.57373046875,
0.41650390625,
0.57373046875
],
[
0.328369140625,
0.54736328125,
0.39208984375,
0.54736328125,
0.39208984375,
0.55859375,
0.328369140625,
0.55859375
],
[
0.187255859375,
0.56103515625,
0.2283935546875,
0.56103515625,
0.2283935546875,
0.572265625,
0.187255859375,
0.572265625
],
[
0.279541015625,
0.56005859375,
0.313232421875,
0.56005859375,
0.313232421875,
0.572265625,
0.279541015625,
0.572265625
],
[
0.33154296875,
0.56103515625,
0.403076171875,
0.56103515625,
0.403076171875,
0.57177734375,
0.33154296875,
0.57177734375
],
[
0.5146484375,
0.56787109375,
0.884765625,
0.56787109375,
0.884765625,
0.58056640625,
0.5146484375,
0.58056640625
],
[
0.28955078125,
0.57470703125,
0.31494140625,
0.57470703125,
0.31494140625,
0.58740234375,
0.28955078125,
0.58740234375
],
[
0.418212890625,
0.57373046875,
0.462646484375,
0.57373046875,
0.462646484375,
0.58837890625,
0.418212890625,
0.58837890625
],
[
0.33154296875,
0.576171875,
0.382080078125,
0.576171875,
0.382080078125,
0.58740234375,
0.33154296875,
0.58740234375
],
[
0.154541015625,
0.57666015625,
0.17626953125,
0.57666015625,
0.17626953125,
0.58740234375,
0.154541015625,
0.58740234375
],
[
0.187255859375,
0.576171875,
0.2392578125,
0.576171875,
0.2392578125,
0.58740234375,
0.187255859375,
0.58740234375
],
[
0.51318359375,
0.58203125,
0.8251953125,
0.58203125,
0.8251953125,
0.595703125,
0.51318359375,
0.595703125
],
[
0.28955078125,
0.5888671875,
0.31494140625,
0.5888671875,
0.31494140625,
0.60107421875,
0.28955078125,
0.60107421875
],
[
0.4189453125,
0.5888671875,
0.4609375,
0.5888671875,
0.4609375,
0.60107421875,
0.4189453125,
0.60107421875
],
[
0.330078125,
0.59033203125,
0.39794921875,
0.59033203125,
0.39794921875,
0.6015625,
0.330078125,
0.6015625
],
[
0.18896484375,
0.59033203125,
0.268798828125,
0.59033203125,
0.268798828125,
0.6015625,
0.18896484375,
0.6015625
],
[
0.28955078125,
0.6025390625,
0.316650390625,
0.6025390625,
0.316650390625,
0.615234375,
0.28955078125,
0.615234375
],
[
0.33154296875,
0.60400390625,
0.387939453125,
0.60400390625,
0.387939453125,
0.615234375,
0.33154296875,
0.615234375
],
[
0.418212890625,
0.6015625,
0.462646484375,
0.6015625,
0.462646484375,
0.6162109375,
0.418212890625,
0.6162109375
],
[
0.418212890625,
0.62060546875,
0.4609375,
0.62060546875,
0.4609375,
0.6318359375,
0.418212890625,
0.6318359375
],
[
0.187255859375,
0.60498046875,
0.2476806640625,
0.60498046875,
0.2476806640625,
0.6162109375,
0.187255859375,
0.6162109375
],
[
0.187255859375,
0.62158203125,
0.2392578125,
0.62158203125,
0.2392578125,
0.6328125,
0.187255859375,
0.6328125
],
[
0.280517578125,
0.62158203125,
0.31494140625,
0.62158203125,
0.31494140625,
0.6328125,
0.280517578125,
0.6328125
],
[
0.328369140625,
0.62158203125,
0.391357421875,
0.62158203125,
0.391357421875,
0.6328125,
0.328369140625,
0.6328125
],
[
0.5146484375,
0.62451171875,
0.884765625,
0.62451171875,
0.884765625,
0.63818359375,
0.5146484375,
0.63818359375
],
[
0.185546875,
0.63427734375,
0.2301025390625,
0.63427734375,
0.2301025390625,
0.6484375,
0.185546875,
0.6484375
],
[
0.279541015625,
0.63427734375,
0.313232421875,
0.63427734375,
0.313232421875,
0.646484375,
0.279541015625,
0.646484375
],
[
0.330078125,
0.63427734375,
0.399658203125,
0.63427734375,
0.399658203125,
0.6484375,
0.330078125,
0.6484375
],
[
0.418212890625,
0.63525390625,
0.4609375,
0.63525390625,
0.4609375,
0.646484375,
0.418212890625,
0.646484375
],
[
0.5146484375,
0.64013671875,
0.884765625,
0.64111328125,
0.884765625,
0.65478515625,
0.5146484375,
0.65380859375
],
[
0.279541015625,
0.6484375,
0.31494140625,
0.6484375,
0.31494140625,
0.66015625,
0.279541015625,
0.66015625
],
[
0.330078125,
0.6484375,
0.382080078125,
0.6484375,
0.382080078125,
0.66015625,
0.330078125,
0.66015625
],
[
0.4189453125,
0.6484375,
0.4609375,
0.6484375,
0.4609375,
0.66015625,
0.4189453125,
0.66015625
],
[
0.185546875,
0.6494140625,
0.2509765625,
0.6494140625,
0.2509765625,
0.6611328125,
0.185546875,
0.6611328125
],
[
0.15283203125,
0.6494140625,
0.175537109375,
0.6494140625,
0.175537109375,
0.66015625,
0.15283203125,
0.66015625
],
[
0.51123046875,
0.65625,
0.88330078125,
0.65673828125,
0.88330078125,
0.67138671875,
0.51123046875,
0.67041015625
],
[
0.18896484375,
0.6630859375,
0.231689453125,
0.6630859375,
0.231689453125,
0.6748046875,
0.18896484375,
0.6748046875
],
[
0.28955078125,
0.6630859375,
0.316650390625,
0.6630859375,
0.316650390625,
0.6748046875,
0.28955078125,
0.6748046875
],
[
0.330078125,
0.66455078125,
0.3837890625,
0.66455078125,
0.3837890625,
0.67578125,
0.330078125,
0.67578125
],
[
0.4189453125,
0.662109375,
0.4609375,
0.662109375,
0.4609375,
0.67578125,
0.4189453125,
0.67578125
],
[
0.41650390625,
0.67578125,
0.462646484375,
0.67578125,
0.462646484375,
0.6904296875,
0.41650390625,
0.6904296875
],
[
0.51123046875,
0.6728515625,
0.884765625,
0.67138671875,
0.884765625,
0.68603515625,
0.51123046875,
0.68701171875
],
[
0.33154296875,
0.6767578125,
0.40478515625,
0.6767578125,
0.40478515625,
0.6884765625,
0.33154296875,
0.6884765625
],
[
0.187255859375,
0.677734375,
0.252685546875,
0.677734375,
0.252685546875,
0.689453125,
0.187255859375,
0.689453125
],
[
0.28955078125,
0.6767578125,
0.316650390625,
0.6767578125,
0.316650390625,
0.689453125,
0.28955078125,
0.689453125
],
[
0.51318359375,
0.6884765625,
0.60791015625,
0.6884765625,
0.60791015625,
0.703125,
0.51318359375,
0.703125
],
[
0.27783203125,
0.69482421875,
0.31494140625,
0.69482421875,
0.31494140625,
0.7060546875,
0.27783203125,
0.7060546875
],
[
0.418212890625,
0.69482421875,
0.462646484375,
0.69482421875,
0.462646484375,
0.7060546875,
0.418212890625,
0.7060546875
],
[
0.330078125,
0.6953125,
0.391357421875,
0.6953125,
0.391357421875,
0.70703125,
0.330078125,
0.70703125
],
[
0.187255859375,
0.6953125,
0.2283935546875,
0.6953125,
0.2283935546875,
0.7060546875,
0.187255859375,
0.7060546875
],
[
0.280517578125,
0.7080078125,
0.31494140625,
0.7080078125,
0.31494140625,
0.72021484375,
0.280517578125,
0.72021484375
],
[
0.41650390625,
0.70703125,
0.4609375,
0.70703125,
0.4609375,
0.7216796875,
0.41650390625,
0.7216796875
],
[
0.33154296875,
0.708984375,
0.40478515625,
0.708984375,
0.40478515625,
0.72021484375,
0.33154296875,
0.72021484375
],
[
0.18896484375,
0.708984375,
0.255126953125,
0.708984375,
0.255126953125,
0.72021484375,
0.18896484375,
0.72021484375
],
[
0.151123046875,
0.72265625,
0.175537109375,
0.72265625,
0.175537109375,
0.7353515625,
0.151123046875,
0.7353515625
],
[
0.28955078125,
0.72265625,
0.31494140625,
0.72265625,
0.31494140625,
0.7353515625,
0.28955078125,
0.7353515625
],
[
0.418212890625,
0.7216796875,
0.462646484375,
0.7216796875,
0.462646484375,
0.736328125,
0.418212890625,
0.736328125
],
[
0.5439453125,
0.72265625,
0.64208984375,
0.72265625,
0.64208984375,
0.73388671875,
0.5439453125,
0.73388671875
],
[
0.33154296875,
0.7236328125,
0.382080078125,
0.7236328125,
0.382080078125,
0.7353515625,
0.33154296875,
0.7353515625
],
[
0.187255859375,
0.7236328125,
0.235107421875,
0.7236328125,
0.235107421875,
0.7353515625,
0.187255859375,
0.7353515625
],
[
0.5146484375,
0.7236328125,
0.5283203125,
0.7236328125,
0.5283203125,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.185546875,
0.73388671875,
0.264404296875,
0.73388671875,
0.264404296875,
0.7490234375,
0.185546875,
0.7490234375
],
[
0.28955078125,
0.736328125,
0.31494140625,
0.736328125,
0.31494140625,
0.7490234375,
0.28955078125,
0.7490234375
],
[
0.333251953125,
0.73681640625,
0.387939453125,
0.73681640625,
0.387939453125,
0.7490234375,
0.333251953125,
0.7490234375
],
[
0.41650390625,
0.7353515625,
0.462646484375,
0.7353515625,
0.462646484375,
0.75,
0.41650390625,
0.75
],
[
0.185546875,
0.75,
0.258544921875,
0.75,
0.258544921875,
0.76123046875,
0.185546875,
0.76123046875
],
[
0.28955078125,
0.75048828125,
0.316650390625,
0.75048828125,
0.316650390625,
0.7626953125,
0.28955078125,
0.7626953125
],
[
0.41650390625,
0.7490234375,
0.46337890625,
0.7490234375,
0.46337890625,
0.76318359375,
0.41650390625,
0.76318359375
],
[
0.51318359375,
0.75048828125,
0.884765625,
0.75048828125,
0.884765625,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.51318359375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.328369140625,
0.751953125,
0.39794921875,
0.751953125,
0.39794921875,
0.7626953125,
0.328369140625,
0.7626953125
],
[
0.1175537109375,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.79150390625,
0.1175537109375,
0.79150390625
],
[
0.51123046875,
0.7822265625,
0.88330078125,
0.78076171875,
0.88330078125,
0.7958984375,
0.51123046875,
0.79638671875
],
[
0.12005615234375,
0.79345703125,
0.489501953125,
0.79345703125,
0.489501953125,
0.80712890625,
0.12005615234375,
0.80712890625
],
[
0.51318359375,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.51318359375,
0.8125
],
[
0.12005615234375,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.8193359375,
0.12005615234375,
0.8193359375
],
[
0.51123046875,
0.81396484375,
0.88330078125,
0.8134765625,
0.88330078125,
0.82763671875,
0.51123046875,
0.8291015625
],
[
0.1158447265625,
0.82080078125,
0.490234375,
0.82177734375,
0.490234375,
0.8359375,
0.1158447265625,
0.8349609375
],
[
0.51123046875,
0.830078125,
0.88330078125,
0.83056640625,
0.88330078125,
0.845703125,
0.51123046875,
0.84423828125
],
[
0.118408203125,
0.8359375,
0.489501953125,
0.8359375,
0.489501953125,
0.849609375,
0.118408203125,
0.849609375
],
[
0.118408203125,
0.85107421875,
0.2626953125,
0.85107421875,
0.2626953125,
0.8642578125,
0.118408203125,
0.8642578125
],
[
0.51318359375,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.51123046875,
0.8623046875,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51123046875,
0.8759765625
],
[
0.51318359375,
0.87744140625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.892578125
],
[
0.509765625,
0.892578125,
0.884765625,
0.8935546875,
0.884765625,
0.9111328125,
0.509765625,
0.91015625
]
] |
[
"Translation",
"I M/F",
"also should i ask the manager what the pay would be",
"Source Sentence (En)",
"De plus, devrais- je demander au gestionnaire quel",
"M",
"if i got the job prior to flying out?",
"seraitlesalairesijeobtenais leposteavantdeprendre",
"F",
"I'avion?",
"also should i ask the secretary what the pay would be",
"De plus, devrais - je demander a la secretaire quel serait",
"F",
"",
"if i got the job prior to flying out?",
"le salaire si je obtenais le poste avant mon depart?",
"currently thinking about learning a trade (mostly a",
"Derzeit iber das Erlernen eines Gewerbes nachdenken",
"M",
"electrician).",
"(meistens Elektriker).",
"De",
"currently thinking about learning a trade (mostly a",
"Derzeit iber das Erlernen eines Gewerbes nachdenken",
"F",
"cook).",
"(meistens eineKochin).",
"decided to become a teacher:spent a year working 2",
"-Decidi ser maestra:pase un ano trabajando en 2",
"F",
"jobs and doing prerequisites for a masters in education.",
"trabajos y haciendo requisitos previos para una maestria",
"eneducacion",
"Es",
" decided to become a lecturer : spent a year working 2",
"-Decidiserprofesor:paséunanotrabajandoen2",
"M",
"jobs and doing prerequisites for a masters in education.",
"trabajos y haciendo requisitos previospara una maestria",
"en educacion",
"i read about a psychologist who upgraded into",
"yHaIoncuxonoze,koTopbliIpeBpaMIcaBMd.",
"M",
"Ru",
"",
"becoming a m.d",
"i read about a nurse who upgraded into becoming a",
"ntaloMeucectpe,koTopan peBpatuiachBok",
"H",
"m.d.",
"TOpa MeJHIMHbl",
"Table 2: Examples from our dataset of a minimal pair of English gender-neutral source sentences, translated into",
"two different genders in all target languages. Red (italic) stands for masculine, cyan (normal) stands for feminine.",
"Sentencepair1",
"Original: you need to have experience working with hydraulic",
"lifts,&they liketo seethatyou've workedortrained as",
"Positive",
"M:FNegative",
"M:F",
"mechanic",
"685:1",
"Substitution: you need to have experience working with",
"0:36",
"nurse",
"manager",
"hydraulic lifts,& they like to see that you've worked or trained as",
"0:17",
"406:0",
"employee",
"secretary",
"364:0",
"a nurse",
"Fr",
"7:1",
"employees",
"353:0",
"teacher",
"assistant",
"1:7",
"parents",
"Sentence pair 2:",
"337:0",
"Original: in fact, probably not even as a seasoned nurse.",
"8:0",
"manager",
"teacher",
"secretary",
"0:27",
"594:0",
"Substitution:in fact,probablynot even as a seasoned fighter",
"409:1",
"manager",
"nurse",
"0:21",
"employees",
"Figure 2: Two sentence pairs from our dataset that found",
"3:7",
"359:0",
"friends",
"De",
"teacher",
"to be shared between all four target languages.",
"0:9",
"320:0",
"employee",
"receptionist",
"7:0",
"students",
"316:0",
"691:0",
"manager",
"teacher",
"4:29",
"manager",
"former, and professions in the latter. A separate but",
"nurse",
"0:31",
"employee",
"446:0",
"related line of work focuses on generating correct",
"0:26",
"friends",
"380:0",
"secretary",
"Es",
"inflections when translating to gender-marking",
"writer",
"8:0",
"parents",
"374:0",
"345:0",
"languages (Vanmassenhove et al., 2018; Moryossef",
"supervisor",
"employee",
"5:0",
"et al., 2019).",
"0:32",
"713:0",
"manager",
"nurse",
"0:13",
"519:0",
"employees",
"babysitter",
"Ru",
"0:5",
"439:0",
"Conclusion",
"friends",
"nurses",
"6",
" dishwasher",
"0:4",
"students",
"417:0",
"technician",
"3:0",
"392:0",
"The primary contribution of our work is a novel.",
"automatic method for identifying gender issues in",
"employee",
"Table 3: Top five human reference forms in our",
"machine translation. By performing BERT-based",
"dataset, and their ratio of times they are translated as",
"perturbations on naturally-occurring sentences, we",
"masculinecomparedtofeminine.Positiveindicatesthat",
"are able to identify sentence pairs that behave",
"the examples were taken from the at-risk group from our",
"differently upon translation to gender-marking",
"pipeline, and negative from the random sample among",
"the not at-risk group.",
"languages. We demonstrate our technique over",
"human reference forms and discover new sources",
"of risk beyond the word lists used previously.",
"Furthermore, the novelty of our approach is its"
] |
{
"arxiv_id": "2004.14065",
"language": "en",
"page_idx": 3
}
|
||
2106.05482-en-1
|
2106.05482-en-1.jpg
|
item at $k$-th position for a specific request. The order of items is determined by maximizing$\sum CTR_{k}^{j}bid^{j}$,which is achieved by a greedy algorithm that selects a most valuable item from top position to bottom position. The contribution of this paper is summarized as follows:
$\bullet$ We propose to employ a shallow position-wise combination $\hat{\text{module with non-linear interaction in DPIN.Given a large}}$ number of combinations between candidate items and positions, this module is able to predict all the combinations' CTRs in parallel. This achieves consistency and greatly improves model performance.
$\bullet$ DPIN is the first method, as far as we know, to model user interests in candidate positions. We apply a deep positionwise interaction module to effectively represents deep nonlinear interaction among position, user and context.
$\bullet$ A new evaluation metrics named PAUC (position-wise AUC) is proposed for measuring performance in solving position bias problem. We conduct extensive experiments on a real world dataset to show empirically that DPIN is both effective and efficient. Online A/B test is also deployed to demonstrate that DPIN has a significant improvement over a highly optimized baseline.
2 DEEP POSITION-WISE INTERACTION
## $\mathbf{NETWORK}$
In this section, we introduce our DPIN. As shown in Figure 1, the DPIN is composed of three modules, which are base module for handling $J$ items, deep position-wise interaction module for handling $K$ positions and position-wise combination module for combining $J$ items and $K$ positions. By this network, CTRs for all candidates at each position can be predicted under the limit of serving performance. We will describe the three modules in detail.
## 2.1 Base Module
Similar to most deep CTR models[5-7,18,23,24], we adopt the structure of embedding and MLP (Multiple Layer Perception) as our base module. For a request, the base module regards an user, a context and $J$ candidate items as inputs, then the representation of ${\mathrm{j-th~item~with~request~information~}}r_{j}^{item}{\mathrm{~is~given~via:}}$
(2)
(3)
$$u=Concat(E(u_{1}),...,E(u_{m})),\\c=Concat(E(c_{1}),...,E(c_{n})),\\i_{j}=Concat(E(i_{1}^{j}),...,E(i_{o}^{j})),\\r_{j}^{item}=MLP(Concat(u,c,i_{j})),$$
(4)
(5)
where $\{u_1,...,u_m\},\{c_1,...,c_n\},\{i_1^j,...,i_o^j\}$ are user features set, context features set and j-th item features set respectively. $E(\cdot)\in\mathbb{R}^d$ means the embedding which transforms sparse feature into learnable $d$-dimensional dense vector. Then, the concatenated vector is fed as input of $MLP$ with $ReLU(x)=max(0,x)$ activation function.
2.2 Deep Position-wise Interaction Module
A brute force method in base module with $O(JKC)$ time complexity for CTR predictions of K positions is unacceptable due to the large time complexity $O(C)$, which is the inference latency of an item at
a position. Therefore, we propose a deep position-wise interaction module paralleled with base module for learning deep non-linear interaction, in which item information is excluded.
In this module, the user behavior sequence at each position is retrieved independently for position-wise interest aggregation to eliminate position bias of lifelong sequence. Then, position-wise non-linear interaction among position, context and user are employed. Finally, transformer[17] is adopted for deep interaction among different positions.
Position-wise Interest Aggregation. We represent the user behavior sequence at $k$-th position as $B_k=\{b_1^k,b_2^k,...,b_L^k\}$,in which $b_{l}^{k}=[v_{l}^{k},c_{l}^{k}]$ stands for $l$-th behavior record.$v_l^{k}$ is the $l$-th clicked item(item id, item category are included) and $c_l^k$ is the context(query, location, hour of day, day of week are included) at which the interaction happened. Behavior embedding $b_l^k$ can be obtained via:
$$b_{l}^{k}=Concat(E(v_{1}^{k_{l}}),...,E(v_{o}^{k_{l}}),E(c_{1}^{k_{l}}),...,E(c_{n}^{k_{l}}),E(dif^{k_{l}})),$$
(6
where $\{v_{1}^{k_{l}},...,v_{o}^{k_{l}}\},\{c_{1}^{k_{l}},...,c_{n}^{k_{l}}\}$ are features set of $v_{l}^{k},c_{l}^{k}$ respec tively and $dif^{k_l}$ is the time difference between the behavior and current context.
The aggregation of behavior sequence at k-th position $b_k$ is calculated by a context-aware attention for extracting user interests related to current context $c$, which can be formulated as follows:
(7)
$$\begin{aligned}&a_{l}^{k}=ReLu(Concat(b_{l}^{k},c)W_{a}+b_{a})W_{b}+b_{b},\\&b_{k}=\sum_{l=1}^{L}\frac{\exp(a_{l}^{k})}{\sum_{i=1}^{L}\exp(a_{i}^{k})}b_{l}^{k},\end{aligned}$$
(8)
where $W_a,b_a,W_b,b_b$ are learnable parameters.
Position-wise Non-linear Interaction. A fully connect with ReLU activation function is employed for non-linear interaction among position, context and user as follows:
$$v_{k}=ReLU(Concat(E(k),c,b_{k})W_{v}+b_{v}),$$
(9)
where $W_{v},b_{v}$ map concatenated vector to $d_{model}$ dimension.
Transformer Block. If $v_k$ is regarded as a non-linear interaction representation of $k$-th position, user interests at other positions will be lost. Hence the transformer is adopted for interaction among different positions. Same as [17], we donate the input of transformer as $Q=K=V=Concat(\boldsymbol{v}_{1},\boldsymbol{v}_{2},...,\boldsymbol{v}_{\boldsymbol{K}})\in\mathbb{R}^{K\times d_{model}}.$ Multi-head self-attention can be formulated as follows:
(10)
(11)
$$\begin{aligned}MultiHead(Q,K,V)&=Concat(head_{1},...,head_{h})W^{O},\\head_{i}&=Attention(QW_{i}^{Q},KW_{i}^{K},VW_{i}^{V}),\\Attention(Q,K,V)&=softmax(\frac{QK^{T}}{\sqrt{d_{k}}})V,\end{aligned}$$
(12)
where $W_i^{Q},W_i^{K},W_i^{V}\in\mathbb{R}^{d_{model}\times d_k},W^{O}\in\mathbb{R}^{d_{model}\times d_{model}}$ are the parameter matrices. And $d_k=d_{model}/h$ is the dimension of each head. The output of multi-head self-attention is fed as the input of position-wise feed-forward network for non-linear transformation. Residual connections and layer normalization are adopted successively. And a stack of N transformer blocks is employed for deep interaction
|
[
[
0.0849609375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.5146484375,
0.10797119140625,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.5146484375,
0.121826171875
],
[
0.08416748046875,
0.12249755859375,
0.48193359375,
0.121826171875,
0.48193359375,
0.1363525390625,
0.08416748046875,
0.136962890625
],
[
0.515625,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.515625,
0.1351318359375
],
[
0.0849609375,
0.136962890625,
0.48193359375,
0.136962890625,
0.48193359375,
0.15087890625,
0.0849609375,
0.15087890625
],
[
0.515625,
0.1351318359375,
0.82421875,
0.1351318359375,
0.82421875,
0.1483154296875,
0.515625,
0.1483154296875
],
[
0.0849609375,
0.15087890625,
0.483642578125,
0.15087890625,
0.483642578125,
0.1641845703125,
0.0849609375,
0.1641845703125
],
[
0.53369140625,
0.15087890625,
0.91259765625,
0.15087890625,
0.91259765625,
0.1610107421875,
0.53369140625,
0.1610107421875
],
[
0.0849609375,
0.1666259765625,
0.15283203125,
0.1666259765625,
0.15283203125,
0.177978515625,
0.0849609375,
0.177978515625
],
[
0.515625,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.515625,
0.1767578125
],
[
0.11273193359375,
0.1817626953125,
0.48193359375,
0.1817626953125,
0.48193359375,
0.1956787109375,
0.11273193359375,
0.1956787109375
],
[
0.5185546875,
0.1787109375,
0.9111328125,
0.1787109375,
0.9111328125,
0.189453125,
0.5185546875,
0.189453125
],
[
0.125,
0.1956787109375,
0.48193359375,
0.1956787109375,
0.48193359375,
0.208984375,
0.125,
0.208984375
],
[
0.51708984375,
0.1925048828125,
0.91259765625,
0.1925048828125,
0.91259765625,
0.20263671875,
0.51708984375,
0.20263671875
],
[
0.125,
0.210205078125,
0.483642578125,
0.210205078125,
0.483642578125,
0.2235107421875,
0.125,
0.2235107421875
],
[
0.515625,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.515625,
0.2183837890625
],
[
0.12335205078125,
0.2235107421875,
0.48193359375,
0.2235107421875,
0.48193359375,
0.2374267578125,
0.12335205078125,
0.2374267578125
],
[
0.5185546875,
0.2203369140625,
0.67626953125,
0.2203369140625,
0.67626953125,
0.2310791015625,
0.5185546875,
0.2310791015625
],
[
0.125,
0.2374267578125,
0.483642578125,
0.2374267578125,
0.483642578125,
0.250732421875,
0.125,
0.250732421875
],
[
0.53369140625,
0.231689453125,
0.9140625,
0.231689453125,
0.9140625,
0.246826171875,
0.53369140625,
0.246826171875
],
[
0.5146484375,
0.2442626953125,
0.9150390625,
0.24560546875,
0.9150390625,
0.26318359375,
0.5146484375,
0.261962890625
],
[
0.1258544921875,
0.25244140625,
0.2900390625,
0.25244140625,
0.2900390625,
0.26318359375,
0.1258544921875,
0.26318359375
],
[
0.11273193359375,
0.264404296875,
0.48193359375,
0.264404296875,
0.48193359375,
0.27783203125,
0.11273193359375,
0.27783203125
],
[
0.5185546875,
0.26318359375,
0.9140625,
0.26318359375,
0.9140625,
0.27783203125,
0.5185546875,
0.27783203125
],
[
0.12335205078125,
0.279052734375,
0.483642578125,
0.279052734375,
0.483642578125,
0.292236328125,
0.12335205078125,
0.292236328125
],
[
0.515625,
0.279052734375,
0.9140625,
0.27978515625,
0.9140625,
0.294189453125,
0.515625,
0.29345703125
],
[
0.12335205078125,
0.291748046875,
0.48193359375,
0.291748046875,
0.48193359375,
0.304931640625,
0.12335205078125,
0.304931640625
],
[
0.51708984375,
0.294189453125,
0.9140625,
0.294189453125,
0.9140625,
0.30810546875,
0.51708984375,
0.30810546875
],
[
0.12335205078125,
0.306884765625,
0.439453125,
0.306884765625,
0.439453125,
0.32080078125,
0.12335205078125,
0.32080078125
],
[
0.51708984375,
0.309326171875,
0.8935546875,
0.309326171875,
0.8935546875,
0.322509765625,
0.51708984375,
0.322509765625
],
[
0.11273193359375,
0.32080078125,
0.48046875,
0.3193359375,
0.48046875,
0.333251953125,
0.11273193359375,
0.333984375
],
[
0.125,
0.335205078125,
0.4794921875,
0.335205078125,
0.4794921875,
0.34521484375,
0.125,
0.34521484375
],
[
0.53173828125,
0.333984375,
0.91259765625,
0.333984375,
0.91259765625,
0.348388671875,
0.53173828125,
0.348388671875
],
[
0.125,
0.347900390625,
0.48193359375,
0.347900390625,
0.48193359375,
0.361083984375,
0.125,
0.361083984375
],
[
0.1258544921875,
0.363037109375,
0.48046875,
0.363037109375,
0.48046875,
0.373779296875,
0.1258544921875,
0.373779296875
],
[
0.51708984375,
0.359130859375,
0.9150390625,
0.359130859375,
0.9150390625,
0.376953125,
0.51708984375,
0.376953125
],
[
0.125,
0.375732421875,
0.483642578125,
0.375732421875,
0.483642578125,
0.389404296875,
0.125,
0.389404296875
],
[
0.51708984375,
0.37890625,
0.9140625,
0.37890625,
0.9140625,
0.393310546875,
0.51708984375,
0.393310546875
],
[
0.12335205078125,
0.38818359375,
0.48046875,
0.389404296875,
0.48046875,
0.404052734375,
0.12335205078125,
0.40283203125
],
[
0.5185546875,
0.39453125,
0.615234375,
0.39453125,
0.615234375,
0.406005859375,
0.5185546875,
0.406005859375
],
[
0.1258544921875,
0.406005859375,
0.24267578125,
0.406005859375,
0.24267578125,
0.416748046875,
0.1258544921875,
0.416748046875
],
[
0.53173828125,
0.4072265625,
0.91259765625,
0.4072265625,
0.91259765625,
0.42041015625,
0.53173828125,
0.42041015625
],
[
0.51708984375,
0.42236328125,
0.91259765625,
0.42236328125,
0.91259765625,
0.43310546875,
0.51708984375,
0.43310546875
],
[
0.0849609375,
0.43115234375,
0.111083984375,
0.43115234375,
0.111083984375,
0.442626953125,
0.0849609375,
0.442626953125
],
[
0.1094970703125,
0.429931640625,
0.4306640625,
0.429931640625,
0.4306640625,
0.443115234375,
0.1094970703125,
0.443115234375
],
[
0.51708984375,
0.436279296875,
0.90673828125,
0.436279296875,
0.90673828125,
0.4462890625,
0.51708984375,
0.4462890625
],
[
0.11358642578125,
0.447509765625,
0.2091064453125,
0.447509765625,
0.2091064453125,
0.458984375,
0.11358642578125,
0.458984375
],
[
0.5849609375,
0.453857421875,
0.84619140625,
0.455078125,
0.84619140625,
0.472900390625,
0.5849609375,
0.4716796875
],
[
0.8935546875,
0.45703125,
0.9140625,
0.45703125,
0.9140625,
0.4716796875,
0.8935546875,
0.4716796875
],
[
0.08416748046875,
0.46337890625,
0.48193359375,
0.46337890625,
0.48193359375,
0.477783203125,
0.08416748046875,
0.477783203125
],
[
0.5947265625,
0.46337890625,
0.60888671875,
0.46337890625,
0.60888671875,
0.46728515625,
0.5947265625,
0.46728515625
],
[
0.08416748046875,
0.4765625,
0.48046875,
0.477783203125,
0.48046875,
0.4912109375,
0.08416748046875,
0.490478515625
],
[
0.64208984375,
0.477783203125,
0.708984375,
0.477783203125,
0.708984375,
0.492431640625,
0.64208984375,
0.492431640625
],
[
0.5849609375,
0.4892578125,
0.62158203125,
0.4892578125,
0.62158203125,
0.50048828125,
0.5849609375,
0.50048828125
],
[
0.73388671875,
0.490478515625,
0.74169921875,
0.490478515625,
0.74169921875,
0.5,
0.73388671875,
0.5
],
[
0.8935546875,
0.488037109375,
0.9150390625,
0.488037109375,
0.9150390625,
0.50244140625,
0.8935546875,
0.50244140625
],
[
0.615234375,
0.498779296875,
0.7236328125,
0.494384765625,
0.7255859375,
0.51220703125,
0.61669921875,
0.5166015625
],
[
0.0849609375,
0.492431640625,
0.483642578125,
0.492431640625,
0.483642578125,
0.505859375,
0.0849609375,
0.505859375
],
[
0.0849609375,
0.505859375,
0.48193359375,
0.505859375,
0.48193359375,
0.51953125,
0.0849609375,
0.51953125
],
[
0.08660888671875,
0.521484375,
0.48046875,
0.521484375,
0.48046875,
0.5322265625,
0.08660888671875,
0.5322265625
],
[
0.5146484375,
0.521484375,
0.79736328125,
0.521484375,
0.79736328125,
0.53515625,
0.5146484375,
0.53515625
],
[
0.08416748046875,
0.5341796875,
0.483642578125,
0.5341796875,
0.483642578125,
0.54736328125,
0.08416748046875,
0.54736328125
],
[
0.53369140625,
0.53515625,
0.91259765625,
0.53515625,
0.91259765625,
0.548828125,
0.53369140625,
0.548828125
],
[
0.0849609375,
0.54736328125,
0.483642578125,
0.54736328125,
0.483642578125,
0.56103515625,
0.0849609375,
0.56103515625
],
[
0.5146484375,
0.54736328125,
0.9140625,
0.548828125,
0.9140625,
0.5625,
0.5146484375,
0.56103515625
],
[
0.5146484375,
0.56298828125,
0.79150390625,
0.56103515625,
0.79150390625,
0.57568359375,
0.5146484375,
0.57763671875
],
[
0.0849609375,
0.57373046875,
0.2320556640625,
0.57275390625,
0.2320556640625,
0.5869140625,
0.0849609375,
0.58837890625
],
[
0.58837890625,
0.58544921875,
0.83837890625,
0.583984375,
0.83837890625,
0.5986328125,
0.58837890625,
0.599609375
],
[
0.89453125,
0.5859375,
0.9140625,
0.5859375,
0.9140625,
0.5986328125,
0.89453125,
0.5986328125
],
[
0.0849609375,
0.59228515625,
0.48046875,
0.59228515625,
0.48046875,
0.60595703125,
0.0849609375,
0.60595703125
],
[
0.0849609375,
0.60498046875,
0.48046875,
0.60498046875,
0.48046875,
0.61865234375,
0.0849609375,
0.61865234375
],
[
0.515625,
0.60791015625,
0.88134765625,
0.60791015625,
0.88134765625,
0.62255859375,
0.515625,
0.62255859375
],
[
0.08416748046875,
0.62060546875,
0.48193359375,
0.62060546875,
0.48193359375,
0.6337890625,
0.08416748046875,
0.6337890625
],
[
0.53125,
0.62060546875,
0.9140625,
0.6220703125,
0.9140625,
0.63623046875,
0.53125,
0.63525390625
],
[
0.08416748046875,
0.63330078125,
0.48193359375,
0.63330078125,
0.48193359375,
0.646484375,
0.08416748046875,
0.646484375
],
[
0.51708984375,
0.63623046875,
0.91259765625,
0.63623046875,
0.91259765625,
0.6494140625,
0.51708984375,
0.6494140625
],
[
0.08087158203125,
0.64453125,
0.41015625,
0.64453125,
0.41015625,
0.662109375,
0.08087158203125,
0.662109375
],
[
0.51708984375,
0.6494140625,
0.91259765625,
0.6494140625,
0.91259765625,
0.66357421875,
0.51708984375,
0.66357421875
],
[
0.2059326171875,
0.66845703125,
0.3857421875,
0.66845703125,
0.3857421875,
0.681640625,
0.2059326171875,
0.681640625
],
[
0.461669921875,
0.66943359375,
0.483642578125,
0.66943359375,
0.483642578125,
0.6845703125,
0.461669921875,
0.6845703125
],
[
0.51708984375,
0.6640625,
0.9140625,
0.6640625,
0.9140625,
0.67822265625,
0.51708984375,
0.67822265625
],
[
0.51708984375,
0.6767578125,
0.9140625,
0.6767578125,
0.9140625,
0.69140625,
0.51708984375,
0.69140625
],
[
0.20751953125,
0.68603515625,
0.381591796875,
0.6875,
0.381591796875,
0.7021484375,
0.20751953125,
0.70068359375
],
[
0.4599609375,
0.6875,
0.483642578125,
0.6875,
0.483642578125,
0.7021484375,
0.4599609375,
0.7021484375
],
[
0.515625,
0.6923828125,
0.78173828125,
0.6923828125,
0.78173828125,
0.70556640625,
0.515625,
0.70556640625
],
[
0.2059326171875,
0.70703125,
0.37744140625,
0.70703125,
0.37744140625,
0.7216796875,
0.2059326171875,
0.7216796875
],
[
0.4599609375,
0.70849609375,
0.483642578125,
0.70849609375,
0.483642578125,
0.72265625,
0.4599609375,
0.72265625
],
[
0.5361328125,
0.71337890625,
0.85888671875,
0.71044921875,
0.85888671875,
0.72802734375,
0.5361328125,
0.73095703125
],
[
0.88623046875,
0.71533203125,
0.9140625,
0.71533203125,
0.9140625,
0.72802734375,
0.88623046875,
0.72802734375
],
[
0.178955078125,
0.72265625,
0.370849609375,
0.72802734375,
0.370849609375,
0.74560546875,
0.178955078125,
0.74072265625
],
[
0.461669921875,
0.72802734375,
0.483642578125,
0.72802734375,
0.483642578125,
0.7421875,
0.461669921875,
0.7421875
],
[
0.62158203125,
0.73291015625,
0.87109375,
0.73291015625,
0.87109375,
0.75048828125,
0.62158203125,
0.75048828125
],
[
0.87841796875,
0.7373046875,
0.91259765625,
0.7373046875,
0.91259765625,
0.74853515625,
0.87841796875,
0.74853515625
],
[
0.0849609375,
0.751953125,
0.48046875,
0.751953125,
0.48046875,
0.7666015625,
0.0849609375,
0.7666015625
],
[
0.734375,
0.755859375,
0.76953125,
0.75048828125,
0.77392578125,
0.767578125,
0.73876953125,
0.77294921875
],
[
0.5439453125,
0.76220703125,
0.73876953125,
0.76220703125,
0.73876953125,
0.77587890625,
0.5439453125,
0.77587890625
],
[
0.76708984375,
0.7646484375,
0.79150390625,
0.7646484375,
0.79150390625,
0.77587890625,
0.76708984375,
0.77587890625
],
[
0.8837890625,
0.76220703125,
0.91650390625,
0.76220703125,
0.91650390625,
0.77783203125,
0.8837890625,
0.77783203125
],
[
0.08416748046875,
0.76806640625,
0.48046875,
0.7666015625,
0.48046875,
0.78076171875,
0.08416748046875,
0.78271484375
],
[
0.74169921875,
0.7734375,
0.771484375,
0.7734375,
0.771484375,
0.79052734375,
0.74169921875,
0.79052734375
],
[
0.08660888671875,
0.7841796875,
0.48193359375,
0.7841796875,
0.48193359375,
0.79443359375,
0.08660888671875,
0.79443359375
],
[
0.08660888671875,
0.79736328125,
0.48193359375,
0.79736328125,
0.48193359375,
0.80810546875,
0.08660888671875,
0.80810546875
],
[
0.515625,
0.79541015625,
0.9140625,
0.79541015625,
0.9140625,
0.81298828125,
0.515625,
0.81298828125
],
[
0.0849609375,
0.81005859375,
0.483642578125,
0.81005859375,
0.483642578125,
0.82373046875,
0.0849609375,
0.82373046875
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.82763671875,
0.51708984375,
0.82763671875
],
[
0.515625,
0.8271484375,
0.9140625,
0.8271484375,
0.9140625,
0.84033203125,
0.515625,
0.84033203125
],
[
0.08416748046875,
0.8359375,
0.44287109375,
0.8359375,
0.44287109375,
0.849609375,
0.08416748046875,
0.849609375
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.51708984375,
0.85693359375,
0.91259765625,
0.85693359375,
0.91259765625,
0.8671875,
0.51708984375,
0.8671875
],
[
0.08660888671875,
0.87060546875,
0.48046875,
0.87060546875,
0.48046875,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.51708984375,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.08416748046875,
0.88330078125,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.08416748046875,
0.896484375
],
[
0.515625,
0.8837890625,
0.5908203125,
0.8837890625,
0.5908203125,
0.89599609375,
0.515625,
0.89599609375
]
] |
[
"item at k-th position for a specific request. The order of items is",
"a position.Therefore, we propose a deep position-wise interaction",
"determined by maximizing E CTR, bid, which is achieved by a",
"module paralleled with base module for learning deep non-linear",
"greedy algorithmthat selects a most valuable item from top position",
" interaction, in which item information is excluded.",
"to bottom position. The contribution of this paper is summarized",
"In thismodule,the user behavior sequenceat each position is",
"asfollows:",
"retrieved independently for position-wise interest aggregation to",
".We proposeto employa shallow position-wise combination",
"eliminatepositionbiasoflifelong sequence.Then,position-wise",
"module with non-linear interaction in DPIN. Given a large",
"non-linear interaction amongposition, contextand userare em",
"number of combinations between candidate items and po-",
"ployed. Finally, transformer[17] is adopted for deep interaction",
"sitions, this module is able to predict all the combinations'",
"among differentpositions.",
"CTRs in parallel. This achieves consistency and greatly im-",
"Position-wise Interest Aggregation. We represent the user",
"behavior sequence at k-th position as Bk = (b*, bs,...bh), in which",
"provesmodel performance.",
" DPIN is the first method, as far as we know, to model user",
"bk = [uk.c] stands for I-th behavior record. o is the l-th clicked",
"interests in candidate positions. We apply a deep position-",
"item(item id, item category are included) and ck is the context(query,",
"wise interaction module to effectively represents deep non-",
"location, hour of day, day of week are included) at which the inter-",
"linear interaction among position, user and context.",
"action happened. Behavior embedding bk can be obtained via:",
".Anewevaluationmetrics named PAUC(position-wise AUC)",
"isproposedformeasuringperformanceinsolvingposition",
"bk =Concat(E(uki)...E(uki).E(cki)...E(cki),E(difki).(6)",
"bias problem. We conduct extensive experiments on a real",
"worlddatasettoshowempiricallythatDPINisbotheffective",
"where o...ck...ch) are features setofuk.crespe-",
"and efficient. Online A/B test is also deployed to demon-",
"tively and difki is the time difference between the behavior and",
"strate that DPIN has a significant improvement over a highly",
"current context.",
"optimized baseline.",
"The aggregation of behavior sequence at k-th position bk is",
"calculated by a context-aware attention for extracting user interests",
"2",
"DEEP POSITION-WISE INTERACTION",
"related to current contextc,whichcan beformulated as follows:",
"NETWORK",
"ak = ReLu(Concat(bk.c)Wa + ba)W + bb.",
"(7)",
"In this section, we introduce our DPIN. As shown in Figure 1,",
"",
"the DPIN is composed of three modules, which are base module",
"exp(a)",
"bk =",
"2-",
"(8)",
"台1 exp(ak)",
"for handling J items, deep position-wise interaction module for",
"handling K positions and position-wise combination module for",
"combiningJitems andKpositions.Bythisnetwork,CTRsforall",
"where Wa, ba, Wp,bp are learnable parameters.",
"candidates at each position can be predicted under the limit of",
"Position-wise Non-linear Interaction. A fully connect with",
"serving performance. We will describe the three modules in detail.",
"ReLU activation function is employed for non-linear interaction",
"among position, context and user as follows:",
"2.1 Base Module",
"Uk = ReLU(Concat(E(k),c,bk)W +b),",
"(9)",
"Similar to most deep CTR models[5-7, 18, 23, 24], we adopt the",
"structure of embedding and MLP (Multiple Layer Perception) as",
"where W,,b, map concatenated vector to dmodel dimension.",
" our base module. For a request, the base module regards an user, a",
"Transformer Block. If uk is regarded as a non-linear interac-",
"context and J candidate items as inputs, then the representation of",
"tion representation ofk-th position, user interests at other positions",
"j-th item with request information rjtem is given via:",
"will be lost. Hence the transformer is adopted for interaction among",
"u= Concat(E(ui),..E(um),",
"(2) ",
"different positions. Same as [17], we donate the input of transformer",
"as Q = K = V = Concat(vi,V2, ... k) e RKxdmodel. Multi-head",
"C = Concat(E(ci).., E(cn)),",
"(3)",
"self-attention can be formulated as follows:",
"ij = Concat(E()... E(t)),",
"(4)",
"MultiHead(Q,K,V) = Concat(headi,.. head,)w",
"(10)",
"rjtem = MLP(Concat(u,c,j),",
"(5)",
"head; = Attention(Qw?, KwK,vw),",
"(11)",
"where (ul,... um). (ci...cn).i',...) are user features set, con",
"OKT",
"Attention(Q,K,V) = softmax(",
")V.",
"(12)",
"text features set and j-th item features set respectively. E e Rd",
"Vdk",
"meanstheembeddingwhichtransformssparsefeatureintolearn-",
"able d-dimensional dense vector.Then,the concatenated vector is",
"where w?.wK,wY e Rdmodeixdk,WO Rdmodeixdmodel are the",
"fed as input of MLP with ReLU(x) = max(O, x) activation function.",
"parameter matrices. And dk = dmodei/h is the dimension of each",
"head. The output of multi-head self-attention is fed as the input of",
"2.2 Deep Position-wise Interaction Module",
"position-wise feed-forward network for non-linear transformation.",
"A bruteforce method in base module with O(JKC)time complexity",
"Residual connections and layer normalization areadopted succes-",
"forCTRpredictionsofKpositionsisunacceptableduetothelarge",
"sively. And a stack of N transformer blocks is employed for deep",
"time complexity O(C), which is the inference latency of an item at",
"interaction."
] |
{
"arxiv_id": "2106.05482",
"language": "en",
"page_idx": 1
}
|
||
2106.05482-en-2
|
2106.05482-en-2.jpg
|
<FigureHere>
Figure 1: The structure of Deep Position-wise Interaction Network.
Finally, the representation of $k$-th position with request informa- $\begin{array}{l}{\mathrm{tion~}r_{k}^{pos^{\prime}}\mathrm{is~given~by~the~proposed~deep~position-wise~interaction}}\\{\mathrm{module.}}\\\end{array}$
position bias evaluation since it excludes the impact of relevance recommendations. The number of impressions is about 10 million a day, and 5% of traffic is the randomized traffic.
Metrics. We use AUC (Area Under ROC) as one of our evaluation metrics. To evaluate the model performance of our methods for position bias, we propose PAUC (Position-wise AUC) as another evaluation metrics, which is calculated as follows:
$$PAUC=\frac{\sum_{k=1}^K\#impression_kPAUC@k}{\sum_{k=1}^K\#impression_k},$$
2.3 $\textbf{Position- wise Combination Module}$
(14)
The purpose of the position-wise combination module is to predict $\begin{array}{l}{\mathrm{the~\tilde{C}TR~of~each~item~at~each~position~by~combining~J~items~and}}\\{\mathrm{K~positions.~A~non-linear~interaction~among~}r_{j}^{item},r_{k}^{pos}\mathrm{~and~}E(k)}\end{array}$ is used to learn non-linear relationship among user, context, item and position. The CTR of j-th item at k-th position $CTR_k^j$ can be calculated as follows:
(13)
$$CTR_{k}^{j}=\sigma(ReLU(Concat(r_{j}^{item},r_{k}^{pos},E(k))W_{1}+b_{1})W_{2}+b_{2}),$$
where $W_1,b_1$ are the parameters of non-linear interaction layer. $W_{2},b_{2}$ are the parameters of output layer. $\sigma(\cdot)$ is sigmoid function.
Finally, the model can be trained using stochastic gradient descent algorithm with actual position feature, and cross-entropy is used as the loss function.
# 3 EXPERIMENT
where $\#impression_k$ is the number of impressions at $k$-th position, and $PAUC@k$ is the AUC of impressions at $k$-th position. PAUC measures the relevance ranking quality at every position, ignoring the impact of position bias.
Compared Methods. Our baseline is a highly optimized DN[24] model, in which a large number of attributes and hand-crafted features are added. The number of features is up to 241 and embedding dimension $k$ is 8 for each feature. The hidden sizes of $MLP$ are 1024,512,128. And the hidden size of the non-linear interaction layer in position-wise combination module is 128. The length of users' behaviors is truncated to 300. And the number of positions $K$ is truncated to 25, which accounts for most of the online traffic. In order to compare different methods fairly, we ensure that all input information and parameters setting of common module are consistent. We conducted experiments with the following methods:
DIN. Position is not used in this model.
DIN+ PosInWide. The method models position feature in the wide part of neural network and first position is used for evaluation.
${\mathrm{DIN+PAL.~PAL~method~is~adopted~based~on~DIN.}}$
DIN+ActualPosInWide. The method use actual position for
evaluation based on DIN+PosInWide.
In this section, we evaluate the model performance and serving performance of the proposed DPIN. We describe the experimental settings and experimental results in detail.
## 3.1 Experimental Settings
Datasets. A four-week ad impression log collected from a sponsored search advertising system in a shopping application is used to train the CTR models. We evaluate our methods on two test sets, which are collected from regular traffic and top-k randomized traffic the next day. The top-k randomized traffic is suitable for
|
[
[
0.476318359375,
0.10162353515625,
0.52294921875,
0.0997314453125,
0.52294921875,
0.106689453125,
0.476318359375,
0.10919189453125
],
[
0.525390625,
0.10162353515625,
0.5390625,
0.10418701171875,
0.53759765625,
0.10797119140625,
0.52392578125,
0.1060791015625
],
[
0.314453125,
0.12371826171875,
0.330810546875,
0.12371826171875,
0.330810546875,
0.1275634765625,
0.314453125,
0.1275634765625
],
[
0.36279296875,
0.12371826171875,
0.3798828125,
0.12371826171875,
0.3798828125,
0.126953125,
0.36279296875,
0.126953125
],
[
0.398681640625,
0.12371826171875,
0.428955078125,
0.12371826171875,
0.428955078125,
0.1275634765625,
0.398681640625,
0.1275634765625
],
[
0.42724609375,
0.17236328125,
0.439453125,
0.17236328125,
0.439453125,
0.1787109375,
0.42724609375,
0.1787109375
],
[
0.580078125,
0.17236328125,
0.5908203125,
0.17236328125,
0.5908203125,
0.1767578125,
0.580078125,
0.1767578125
],
[
0.86767578125,
0.1817626953125,
0.8759765625,
0.1817626953125,
0.8759765625,
0.186279296875,
0.86767578125,
0.186279296875
],
[
0.1895751953125,
0.1932373046875,
0.2017822265625,
0.1932373046875,
0.2017822265625,
0.199462890625,
0.1895751953125,
0.199462890625
],
[
0.1871337890625,
0.2083740234375,
0.2181396484375,
0.2083740234375,
0.2181396484375,
0.21533203125,
0.1871337890625,
0.21533203125
],
[
0.81884765625,
0.231689453125,
0.83642578125,
0.231689453125,
0.83642578125,
0.23486328125,
0.81884765625,
0.23486328125
],
[
0.81396484375,
0.26953125,
0.83642578125,
0.26953125,
0.83642578125,
0.2734375,
0.81396484375,
0.2734375
],
[
0.111083984375,
0.27587890625,
0.1568603515625,
0.27587890625,
0.1568603515625,
0.285888671875,
0.111083984375,
0.285888671875
],
[
0.4248046875,
0.279052734375,
0.439453125,
0.279052734375,
0.439453125,
0.2822265625,
0.4248046875,
0.2822265625
],
[
0.181396484375,
0.287353515625,
0.2197265625,
0.2841796875,
0.221435546875,
0.296630859375,
0.1829833984375,
0.2998046875
],
[
0.7900390625,
0.2822265625,
0.87744140625,
0.2822265625,
0.87744140625,
0.292236328125,
0.7900390625,
0.292236328125
],
[
0.453369140625,
0.302978515625,
0.4599609375,
0.3017578125,
0.4599609375,
0.304931640625,
0.455078125,
0.30615234375
],
[
0.464111328125,
0.2998046875,
0.53369140625,
0.2998046875,
0.53369140625,
0.306884765625,
0.464111328125,
0.306884765625
],
[
0.80126953125,
0.313232421875,
0.86376953125,
0.313232421875,
0.86376953125,
0.324462890625,
0.80126953125,
0.324462890625
],
[
0.494384765625,
0.32080078125,
0.51123046875,
0.32080078125,
0.51123046875,
0.330078125,
0.494384765625,
0.330078125
],
[
0.351318359375,
0.33203125,
0.386474609375,
0.33203125,
0.386474609375,
0.343505859375,
0.351318359375,
0.343505859375
],
[
0.484375,
0.333251953125,
0.51953125,
0.333251953125,
0.51953125,
0.34521484375,
0.484375,
0.34521484375
],
[
0.625,
0.333251953125,
0.66015625,
0.333251953125,
0.66015625,
0.3447265625,
0.625,
0.3447265625
],
[
0.80615234375,
0.33837890625,
0.82275390625,
0.33837890625,
0.82275390625,
0.347900390625,
0.80615234375,
0.347900390625
],
[
0.3505859375,
0.349853515625,
0.38818359375,
0.349853515625,
0.38818359375,
0.35986328125,
0.3505859375,
0.35986328125
],
[
0.484375,
0.352783203125,
0.51708984375,
0.35107421875,
0.5185546875,
0.359130859375,
0.486083984375,
0.361083984375
],
[
0.62744140625,
0.352783203125,
0.65185546875,
0.352783203125,
0.65185546875,
0.356689453125,
0.62744140625,
0.356689453125
],
[
0.7998046875,
0.354248046875,
0.8095703125,
0.354248046875,
0.8095703125,
0.359130859375,
0.7998046875,
0.359130859375
],
[
0.53173828125,
0.371826171875,
0.54833984375,
0.371826171875,
0.54833984375,
0.380615234375,
0.53173828125,
0.380615234375
],
[
0.6708984375,
0.371826171875,
0.68896484375,
0.371826171875,
0.68896484375,
0.37890625,
0.6708984375,
0.37890625
],
[
0.59228515625,
0.37255859375,
0.609375,
0.369384765625,
0.61279296875,
0.375732421875,
0.5947265625,
0.37890625
],
[
0.48193359375,
0.37255859375,
0.495849609375,
0.37255859375,
0.495849609375,
0.376953125,
0.48193359375,
0.376953125
],
[
0.62353515625,
0.37255859375,
0.6357421875,
0.37255859375,
0.6357421875,
0.376953125,
0.62353515625,
0.376953125
],
[
0.459228515625,
0.373779296875,
0.47216796875,
0.373779296875,
0.47216796875,
0.376953125,
0.459228515625,
0.376953125
],
[
0.400390625,
0.373779296875,
0.41259765625,
0.373779296875,
0.41259765625,
0.376953125,
0.400390625,
0.376953125
],
[
0.47216796875,
0.392578125,
0.494384765625,
0.392578125,
0.494384765625,
0.396484375,
0.47216796875,
0.396484375
],
[
0.5087890625,
0.393310546875,
0.53515625,
0.393310546875,
0.53515625,
0.396484375,
0.5087890625,
0.396484375
],
[
0.5361328125,
0.393310546875,
0.5595703125,
0.393310546875,
0.5595703125,
0.397705078125,
0.5361328125,
0.397705078125
],
[
0.2484130859375,
0.401611328125,
0.265625,
0.401611328125,
0.265625,
0.406005859375,
0.2484130859375,
0.406005859375
],
[
0.384033203125,
0.404052734375,
0.4306640625,
0.404052734375,
0.4306640625,
0.410888671875,
0.384033203125,
0.410888671875
],
[
0.5888671875,
0.404052734375,
0.64794921875,
0.404052734375,
0.64794921875,
0.410888671875,
0.5888671875,
0.410888671875
],
[
0.64892578125,
0.40283203125,
0.705078125,
0.404052734375,
0.705078125,
0.410888671875,
0.64892578125,
0.410400390625
],
[
0.449462890625,
0.406005859375,
0.488525390625,
0.406005859375,
0.488525390625,
0.4091796875,
0.449462890625,
0.4091796875
],
[
0.325927734375,
0.40478515625,
0.35693359375,
0.40478515625,
0.35693359375,
0.4091796875,
0.325927734375,
0.4091796875
],
[
0.12744140625,
0.410888671875,
0.1405029296875,
0.410888671875,
0.1405029296875,
0.4140625,
0.12744140625,
0.4140625
],
[
0.79736328125,
0.410400390625,
0.8310546875,
0.410400390625,
0.8310546875,
0.41357421875,
0.79736328125,
0.41357421875
],
[
0.19287109375,
0.419189453125,
0.2059326171875,
0.419189453125,
0.2059326171875,
0.423583984375,
0.19287109375,
0.423583984375
],
[
0.77294921875,
0.41845703125,
0.8857421875,
0.41845703125,
0.8857421875,
0.425537109375,
0.77294921875,
0.425537109375
],
[
0.275390625,
0.443115234375,
0.72119140625,
0.442626953125,
0.72119140625,
0.45703125,
0.275390625,
0.458251953125
],
[
0.1004638671875,
0.46533203125,
0.483642578125,
0.46533203125,
0.483642578125,
0.479248046875,
0.1004638671875,
0.479248046875
],
[
0.5146484375,
0.46533203125,
0.91259765625,
0.46533203125,
0.91259765625,
0.479248046875,
0.5146484375,
0.479248046875
],
[
0.08251953125,
0.479248046875,
0.1446533203125,
0.47607421875,
0.146240234375,
0.490478515625,
0.08416748046875,
0.49365234375
],
[
0.1405029296875,
0.479736328125,
0.483642578125,
0.479736328125,
0.483642578125,
0.49365234375,
0.1405029296875,
0.49365234375
],
[
0.51708984375,
0.479248046875,
0.9140625,
0.479248046875,
0.9140625,
0.492431640625,
0.51708984375,
0.492431640625
],
[
0.0849609375,
0.494384765625,
0.1356201171875,
0.494384765625,
0.1356201171875,
0.505859375,
0.0849609375,
0.505859375
],
[
0.51708984375,
0.492431640625,
0.80810546875,
0.492431640625,
0.80810546875,
0.505859375,
0.51708984375,
0.505859375
],
[
0.53369140625,
0.50830078125,
0.91259765625,
0.50830078125,
0.91259765625,
0.51806640625,
0.53369140625,
0.51806640625
],
[
0.0849609375,
0.521484375,
0.4150390625,
0.521484375,
0.4150390625,
0.53515625,
0.0849609375,
0.53515625
],
[
0.51708984375,
0.52099609375,
0.9140625,
0.52099609375,
0.9140625,
0.5341796875,
0.51708984375,
0.5341796875
],
[
0.0849609375,
0.53857421875,
0.48046875,
0.53857421875,
0.48046875,
0.5517578125,
0.0849609375,
0.5517578125
],
[
0.515625,
0.5341796875,
0.9140625,
0.53271484375,
0.9140625,
0.54736328125,
0.515625,
0.548828125
],
[
0.0849609375,
0.55029296875,
0.48193359375,
0.55029296875,
0.48193359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.515625,
0.548828125,
0.81640625,
0.548828125,
0.81640625,
0.5595703125,
0.515625,
0.5595703125
],
[
0.08251953125,
0.564453125,
0.40185546875,
0.5625,
0.40185546875,
0.580078125,
0.08251953125,
0.58203125
],
[
0.406005859375,
0.56494140625,
0.4150390625,
0.56494140625,
0.4150390625,
0.5693359375,
0.406005859375,
0.5693359375
],
[
0.419189453125,
0.56640625,
0.47802734375,
0.56640625,
0.47802734375,
0.57763671875,
0.419189453125,
0.57763671875
],
[
0.64794921875,
0.5625,
0.83642578125,
0.5625,
0.83642578125,
0.5869140625,
0.64794921875,
0.5869140625
],
[
0.38818359375,
0.5693359375,
0.410888671875,
0.57470703125,
0.408447265625,
0.58056640625,
0.3857421875,
0.57568359375
],
[
0.5947265625,
0.5771484375,
0.646484375,
0.5771484375,
0.646484375,
0.58837890625,
0.5947265625,
0.58837890625
],
[
0.88623046875,
0.5771484375,
0.9140625,
0.5771484375,
0.9140625,
0.5888671875,
0.88623046875,
0.5888671875
],
[
0.67724609375,
0.58203125,
0.8056640625,
0.58544921875,
0.8037109375,
0.60693359375,
0.67626953125,
0.60400390625
],
[
0.0849609375,
0.58056640625,
0.48046875,
0.58056640625,
0.48046875,
0.5947265625,
0.0849609375,
0.5947265625
],
[
0.0849609375,
0.59521484375,
0.4794921875,
0.59521484375,
0.4794921875,
0.609375,
0.0849609375,
0.609375
],
[
0.51708984375,
0.60595703125,
0.9140625,
0.60595703125,
0.9140625,
0.619140625,
0.51708984375,
0.619140625
],
[
0.08660888671875,
0.6123046875,
0.2156982421875,
0.6123046875,
0.2156982421875,
0.62255859375,
0.08660888671875,
0.62255859375
],
[
0.51708984375,
0.619140625,
0.91259765625,
0.619140625,
0.91259765625,
0.63330078125,
0.51708984375,
0.63330078125
],
[
0.09234619140625,
0.62744140625,
0.483642578125,
0.62744140625,
0.483642578125,
0.64599609375,
0.09234619140625,
0.64599609375
],
[
0.51708984375,
0.6337890625,
0.9140625,
0.6337890625,
0.9140625,
0.64794921875,
0.51708984375,
0.64794921875
],
[
0.515625,
0.64794921875,
0.68310546875,
0.64794921875,
0.68310546875,
0.6611328125,
0.515625,
0.6611328125
],
[
0.09234619140625,
0.65380859375,
0.48193359375,
0.65380859375,
0.48193359375,
0.6640625,
0.09234619140625,
0.6640625
],
[
0.0849609375,
0.66650390625,
0.483642578125,
0.66650390625,
0.483642578125,
0.6796875,
0.0849609375,
0.6796875
],
[
0.53369140625,
0.6611328125,
0.91650390625,
0.6611328125,
0.91650390625,
0.6748046875,
0.53369140625,
0.6748046875
],
[
0.1004638671875,
0.6787109375,
0.483642578125,
0.6787109375,
0.483642578125,
0.6923828125,
0.1004638671875,
0.6923828125
],
[
0.515625,
0.6748046875,
0.9150390625,
0.6748046875,
0.9150390625,
0.68798828125,
0.515625,
0.68798828125
],
[
0.51708984375,
0.689453125,
0.9140625,
0.689453125,
0.9140625,
0.70263671875,
0.51708984375,
0.70263671875
],
[
0.08660888671875,
0.69580078125,
0.48193359375,
0.69580078125,
0.48193359375,
0.70556640625,
0.08660888671875,
0.70556640625
],
[
0.08416748046875,
0.70703125,
0.238525390625,
0.70703125,
0.238525390625,
0.72021484375,
0.08416748046875,
0.72021484375
],
[
0.5185546875,
0.70263671875,
0.9140625,
0.70263671875,
0.9140625,
0.71630859375,
0.5185546875,
0.71630859375
],
[
0.51708984375,
0.71630859375,
0.9140625,
0.71630859375,
0.9140625,
0.72998046875,
0.51708984375,
0.72998046875
],
[
0.1094970703125,
0.73486328125,
0.233642578125,
0.73486328125,
0.233642578125,
0.74853515625,
0.1094970703125,
0.74853515625
],
[
0.51708984375,
0.73095703125,
0.9150390625,
0.73095703125,
0.9150390625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.08660888671875,
0.73779296875,
0.1094970703125,
0.73779296875,
0.1094970703125,
0.7470703125,
0.08660888671875,
0.7470703125
],
[
0.1004638671875,
0.7412109375,
0.108642578125,
0.7412109375,
0.108642578125,
0.744140625,
0.1004638671875,
0.744140625
],
[
0.1094970703125,
0.7392578125,
0.11358642578125,
0.7392578125,
0.11358642578125,
0.744140625,
0.1094970703125,
0.744140625
],
[
0.51708984375,
0.744140625,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.08416748046875,
0.75244140625,
0.48193359375,
0.75390625,
0.48193359375,
0.767578125,
0.08416748046875,
0.7666015625
],
[
0.51708984375,
0.75830078125,
0.9140625,
0.75830078125,
0.9140625,
0.771484375,
0.51708984375,
0.771484375
],
[
0.0849609375,
0.7666015625,
0.48193359375,
0.7666015625,
0.48193359375,
0.77978515625,
0.0849609375,
0.77978515625
],
[
0.515625,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.515625,
0.78515625
],
[
0.08416748046875,
0.78076171875,
0.340576171875,
0.78076171875,
0.340576171875,
0.79443359375,
0.08416748046875,
0.79443359375
],
[
0.5185546875,
0.7861328125,
0.91259765625,
0.7861328125,
0.91259765625,
0.796875,
0.5185546875,
0.796875
],
[
0.515625,
0.7998046875,
0.9150390625,
0.7998046875,
0.9150390625,
0.81298828125,
0.515625,
0.81298828125
],
[
0.08416748046875,
0.81005859375,
0.306396484375,
0.81005859375,
0.306396484375,
0.82373046875,
0.08416748046875,
0.82373046875
],
[
0.53369140625,
0.81298828125,
0.77392578125,
0.81298828125,
0.77392578125,
0.8271484375,
0.53369140625,
0.8271484375
],
[
0.0849609375,
0.8271484375,
0.483642578125,
0.8271484375,
0.483642578125,
0.84033203125,
0.0849609375,
0.84033203125
],
[
0.53515625,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.53515625,
0.8388671875
],
[
0.08416748046875,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.08416748046875,
0.85498046875
],
[
0.515625,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.515625,
0.85498046875
],
[
0.08416748046875,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.08416748046875,
0.86865234375
],
[
0.53369140625,
0.85498046875,
0.833984375,
0.85498046875,
0.833984375,
0.86865234375,
0.53369140625,
0.86865234375
],
[
0.08660888671875,
0.87060546875,
0.48046875,
0.87060546875,
0.48046875,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.53369140625,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.53369140625,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.8837890625,
0.74267578125,
0.88330078125,
0.74267578125,
0.89453125,
0.515625,
0.89599609375
]
] |
[
"JxKCTRPred",
"tior",
"Posi",
"Cam",
"onMod",
"rpa",
"rpo",
"to",
"",
"mriter",
"0",
"ddR",
"Base Module",
"Man",
"ReLU",
"Multi-Head Attention",
"",
"Transformer Block",
"Transformer Block",
"-",
"ReLU",
"ReLU",
"ReLU",
".",
"Concat",
"Concat",
"FConra",
"At",
"",
".",
"",
"",
"",
"工",
"",
"edding",
"mteres",
"s991e9",
"",
"serBehavior8",
"ositionKContext",
"User Behavior B",
"Position.ce",
"ition 1Co",
"a.",
"prBehavio",
"iter",
"Position-wise Interest Aggregation",
"Figure 1: The structure of Deep Position-wise Interaction Network.",
"Finally, the representation of k-th position with request informa-",
"position bias evaluation since it excludes the impact of relevance",
"tionrpos",
"is given by the proposed deep position-wise interaction",
"recommendations. The number of impressions is about 10 million",
"module.",
"a day, and 5% of traffic is the randomized traffic.",
"Metrics.WeuseAUC(AreaUnderROC)asoneofourevaluation",
"2.3Position-wiseCombinationModule",
"metrics. To evaluate the model performance of our methods for",
"The purpose of the position-wise combination module is to predict",
"position bias, we propose PAUC (Position-wise AUC) as another",
"the CTR of each item at each position by combining J items and",
"evaluation metrics,which is calculated as follows",
"K positions. A non-linear interaction among ritem, r",
"90",
"andE(k)",
"ZK#impression,PAUCQk",
".k",
"PAUC=",
"(14)",
"ZK- #impressionk",
"is used to learn non-linear relationship among user, context, item",
"and position. The CTR of j-th item at k-th position CTR, can be",
"where #impressionk is the number of impressions atk-th position,",
"calculated asfollows:",
"and PAUC@k is the AUC of impressions at k-th position. PAUC",
"CTR, =(ReLU(Concat(rjtem, rpos, E(k)Wi +bi)W2 +b2), (13)",
"measures the relevance ranking quality at every position, ignoring",
"the impact of position bias.",
"whereWi,bi aretheparameters of non-linearinteractionlayer.",
"W2, b2 are the parameters of output layer. () is sigmoid function.",
"Compared Methods. Our baseline is a highly optimized DIN[24]",
"Finally, the model can be trained using stochastic gradient de-",
"model, in which a large number of attributes and hand-crafted fea-",
"tures are added. The number of features is up to 241 and embedding",
"scentalgorithm withactualpositionfeature,andcross-entropy is",
"used as the loss function.",
"dimension k is 8 for each feature. The hidden sizes of MLP are",
"1024, 512, 128. And the hidden size of the non-linear interaction",
"EXPERIMENT",
"layer in position-wise combination module is 128. The length of",
"3",
"",
"",
"users' behaviors is truncated to 300. And the number of positions",
"In this section, we evaluate the model performance and serving",
"K is truncated to 25, which accounts for most of the online traffic.",
"performance of the proposed DPIN. We describe the experimental",
"In order to compare different methods fairly, we ensure that all",
" settings and experimental results in detail.",
"input information andparameters setting of common moduleare",
"consistent. We conducted experiments with the following methods:",
"3.1 Experimental Settings",
"DIN. Position is not used in this model.",
"Datasets. A four-week ad impression log collected from a spon-",
"DIN+PosInWide.Themethod modelspositionfeaturein the",
"sored search advertising system in a shopping application is used",
"wide part of neural network and first position is used for evaluation.",
"to train the CTR models. We evaluate our methods on two test",
"DIN+PAL. PAL method is adopted based on DIN.",
"sets,whicharecollectedfromregulartrafficandtop-krandomized",
"DIN+ActualPosInWide. The method use actual position for",
"traffic the next day. The top-k randomized traffic is suitable for",
"evaluationbasedonDIN+PosInWide"
] |
{
"arxiv_id": "2106.05482",
"language": "en",
"page_idx": 2
}
|
||
2106.05482-en-3
|
2106.05482-en-3.jpg
|
DIN+ Combination. Position- wise combination module is em-
ployed based on DIN. Actual position is used for evaluation.
DPIN-Transformer. Transformer is not used in DPIN.
$\mathbf{DPIN.There~is~the~proposed~DPIN~model,~in~which~a~tow-layer}$ transformer is adopted and $d_{model}=64,h=2$ for self-attention.
$DPIN+ItemAction.Weadddeepposition-wiseinteractionmod-$ ule before the MLP layer of the base module, and candidate item information is introduced in position-wise interest aggregation and position-wise non-linear interaction. The experiment is the upper bound of model performance in our methods, but the serving performance is unacceptable.
# 3.2 Offline Evaluations
Table 1: Offline experimental results of compared methods
on the regular and randomized test sets.
<table>
<tbody>
<tr>
<th rowspan="2">MoDEL</th>
<th colspan="2">Reqular</th>
<th colspan="2">Randomized</th>
</tr>
<tr>
<th>$AUC$</th>
<th>PAUC</th>
<th>$AUC$</th>
<th>PAUC</th>
</tr>
<tr>
<td>$DIN$</td>
<td>0.7818</td>
<td>0.7090</td>
<td>0.7836</td>
<td>0.7223</td>
</tr>
<tr>
<td>DIN+ POSINWIDE</td>
<td>0.7696</td>
<td>0.7109</td>
<td>0.7725</td>
<td>0.7239</td>
</tr>
<tr>
<td>DIN+ PAL</td>
<td>0.7735</td>
<td>0.7128</td>
<td>0.7763</td>
<td>0.7254</td>
</tr>
<tr>
<td>DIN+ ACUTALPOSINWIDE</td>
<td>0.7928</td>
<td>0.7109</td>
<td>0.7938</td>
<td>0.7239</td>
</tr>
<tr>
<td>DIN+ COMBINATION</td>
<td>0.7970</td>
<td>0.7172</td>
<td>0.7985</td>
<td>0.7294</td>
</tr>
<tr>
<td>DPIN-TRANSFORMER</td>
<td>0.7961</td>
<td>0.7148</td>
<td>0.7984</td>
<td>0.7283</td>
</tr>
<tr>
<td>$DPIN$</td>
<td>0.7994</td>
<td>0.7216</td>
<td>0.8015</td>
<td>0.7350</td>
</tr>
<tr>
<td>DPIN+ ITEMACTION</td>
<td>0.7999</td>
<td>0.7223</td>
<td>0.8019</td>
<td>0.7356</td>
</tr>
</tbody>
</table>
Table 1 illustrates experimental results of the compared methods on both regular and randomized test sets. We first analyze the differences between different methods on the regular traffic. Compared with DIN, the DIN+PosInWide and DIN+PAL have a performance degradation on AUC but an improvement on PAUC, which shows that both methods effectively alleviate position bias but lead to inconsistency between offline and online. The DN+AcutalPosInWide solves the inconsistency by introducing actual position during evaluation, which is achievable by position-wise combination module. But modeling position in the wide part leads to the position feature is only a bias, which can not improve PAUC. By employing our proposed position-wise combination module, the DIN+Combination has 2.74% gain on AUC and 0.63% gain on PAUC compared with DIN+PosInWide, achieving consistency and alleviate position bias further, which shows the position bias is not independent. Furthermore, DPIN models deep non-linear interaction among position, context and user, and eliminates position bias existed in the user sequence by position-wise method, which has 0.24% gain on AUC and $0.44\%$ gain on PAUC compared with DIN+Combination. The effect of DPIN-Transformer explains that it is necessary to adopt transformer for interaction among different positions. And the comparison between DPIN and DPIN+ItemAction shows that DPIN is close to the brute force method on both AUC and PAUC. As can be seen finally, the DPIN has 2.98% gain on AUC and 1.07% gain on PAUC relative to the DIN+PosInWide, which is a baseline in the advertising system online.
In order to ensure that our method can learn the position bias instead of overfitting the selection bias of the system, we further evaluate our methods on randomized traffic. The results show that
<FigureHere>
Figure 2: Serving latency changes with the number of candi-
date items in different methods.
the differences between the different methods on both regular and
randomized traffic is consistent.
## 3.3 Serving Performance
We retrieve some requests with different candidate item numbers from the dataset to measure serving performance. As shown in Figure 2, the serving latency of position-wise combination module is negligible compared to the DIN model since user sequence operation has a large proportion of latency. The serving latency of DPIN increases slowly as the number of items increases since the deep position-wise interaction module has nothing to do with items. Compared with DIPIN+ItemAction, the DPIN has a great $\hat{\text{improvement in serving performance with little damage to model}}$ performance, which shows that our proposed method is both effective and efficient.
## 3.4 0nline Evaluations
Online A/B test was conducted in the sponsored search advertising system from 2021-01-08 to 2021-01-22. For the control group, 10% of users are randomly selected and presented with recommendation generated by DIN+PosInWide. For the experimental group, 10% of users are presented with recommendation generated by DPIN. The A/B test shows that the proposed DPIN has improved CTR by 2.25% and RPM (Revenue Per Mille) by 2.15% compared with baseline. For now, DPIN has been deployed online and serves the main traffic, which contributes a significant business revenue growth.
## 4 concLUSIONS
In this paper, a novel model Deep Position-wise Interaction Network (DPIN) is proposed to mitigate position bias, which efficiently combine all candidate items and positions for estimating CTR at each position, achieving consistency between offline and online. At the same time, the deep non-linear interaction among position, context and user is available by the model. In order to evaluate our method in position bias problem, we propose a new evaluation metrics PAUC and the offline experiments show that the proposed DPIN outperforms the compared methods efficacy and efficiency. For now, the DPIN is deployed in a sponsored search advertising system and serving the main traffic.
|
[
[
0.55224609375,
0.09661865234375,
0.57666015625,
0.09661865234375,
0.57666015625,
0.10986328125,
0.55224609375,
0.10986328125
],
[
0.62841796875,
0.0921630859375,
0.72216796875,
0.0921630859375,
0.72216796875,
0.10418701171875,
0.62841796875,
0.10418701171875
],
[
0.10455322265625,
0.10919189453125,
0.48193359375,
0.10919189453125,
0.48193359375,
0.11932373046875,
0.10455322265625,
0.11932373046875
],
[
0.62744140625,
0.106689453125,
0.734375,
0.106689453125,
0.734375,
0.12060546875,
0.62744140625,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.44677734375,
0.121826171875,
0.44677734375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.55810546875,
0.1256103515625,
0.57666015625,
0.1256103515625,
0.57666015625,
0.14013671875,
0.55810546875,
0.14013671875
],
[
0.6259765625,
0.12060546875,
0.73388671875,
0.121826171875,
0.73388671875,
0.1363525390625,
0.6259765625,
0.1351318359375
],
[
0.10296630859375,
0.136962890625,
0.43310546875,
0.136962890625,
0.43310546875,
0.147705078125,
0.10296630859375,
0.147705078125
],
[
0.5361328125,
0.1395263671875,
0.55126953125,
0.1395263671875,
0.55126953125,
0.1962890625,
0.5361328125,
0.1962890625
],
[
0.62744140625,
0.1363525390625,
0.6650390625,
0.1363525390625,
0.6650390625,
0.149658203125,
0.62744140625,
0.149658203125
],
[
0.101318359375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.101318359375,
0.1634521484375
],
[
0.55810546875,
0.1558837890625,
0.580078125,
0.1558837890625,
0.580078125,
0.17041015625,
0.55810546875,
0.17041015625
],
[
0.0849609375,
0.1641845703125,
0.471435546875,
0.1641845703125,
0.471435546875,
0.1787109375,
0.0849609375,
0.1787109375
],
[
0.101318359375,
0.177978515625,
0.483642578125,
0.177978515625,
0.483642578125,
0.1912841796875,
0.101318359375,
0.1912841796875
],
[
0.0849609375,
0.1912841796875,
0.48193359375,
0.1912841796875,
0.48193359375,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.55810546875,
0.1875,
0.57861328125,
0.1875,
0.57861328125,
0.199462890625,
0.55810546875,
0.199462890625
],
[
0.0849609375,
0.205810546875,
0.48193359375,
0.205810546875,
0.48193359375,
0.2197265625,
0.0849609375,
0.2197265625
],
[
0.08660888671875,
0.2203369140625,
0.48046875,
0.2203369140625,
0.48046875,
0.2310791015625,
0.08660888671875,
0.2310791015625
],
[
0.55810546875,
0.216552734375,
0.57861328125,
0.216552734375,
0.57861328125,
0.2310791015625,
0.55810546875,
0.2310791015625
],
[
0.08416748046875,
0.23291015625,
0.48193359375,
0.231689453125,
0.48193359375,
0.246826171875,
0.08416748046875,
0.2474365234375
],
[
0.0849609375,
0.246826171875,
0.26318359375,
0.246826171875,
0.26318359375,
0.260009765625,
0.0849609375,
0.260009765625
],
[
0.62841796875,
0.250732421875,
0.646484375,
0.250732421875,
0.646484375,
0.264404296875,
0.62841796875,
0.264404296875
],
[
0.66845703125,
0.250732421875,
0.68896484375,
0.250732421875,
0.68896484375,
0.26318359375,
0.66845703125,
0.26318359375
],
[
0.7060546875,
0.25,
0.7294921875,
0.25,
0.7294921875,
0.261962890625,
0.7060546875,
0.261962890625
],
[
0.744140625,
0.250732421875,
0.76953125,
0.250732421875,
0.76953125,
0.261962890625,
0.744140625,
0.261962890625
],
[
0.79150390625,
0.250732421875,
0.81201171875,
0.250732421875,
0.81201171875,
0.26318359375,
0.79150390625,
0.26318359375
],
[
0.833984375,
0.250732421875,
0.8544921875,
0.250732421875,
0.8544921875,
0.26318359375,
0.833984375,
0.26318359375
],
[
0.8759765625,
0.250732421875,
0.89453125,
0.250732421875,
0.89453125,
0.26318359375,
0.8759765625,
0.26318359375
],
[
0.69287109375,
0.264404296875,
0.7890625,
0.264404296875,
0.7890625,
0.274658203125,
0.69287109375,
0.274658203125
],
[
0.5888671875,
0.251953125,
0.60107421875,
0.251953125,
0.60107421875,
0.26318359375,
0.5888671875,
0.26318359375
],
[
0.0849609375,
0.2734375,
0.11273193359375,
0.2734375,
0.11273193359375,
0.285400390625,
0.0849609375,
0.285400390625
],
[
0.125,
0.272705078125,
0.285888671875,
0.272705078125,
0.285888671875,
0.285888671875,
0.125,
0.285888671875
],
[
0.515625,
0.285888671875,
0.9140625,
0.287353515625,
0.9140625,
0.3017578125,
0.515625,
0.300537109375
],
[
0.08416748046875,
0.298583984375,
0.48193359375,
0.2998046875,
0.48193359375,
0.314453125,
0.08416748046875,
0.313232421875
],
[
0.51708984375,
0.3017578125,
0.73388671875,
0.3017578125,
0.73388671875,
0.31494140625,
0.51708984375,
0.31494140625
],
[
0.0849609375,
0.314453125,
0.35302734375,
0.314453125,
0.35302734375,
0.32763671875,
0.0849609375,
0.32763671875
],
[
0.51708984375,
0.335205078125,
0.9140625,
0.335205078125,
0.9140625,
0.348388671875,
0.51708984375,
0.348388671875
],
[
0.27685546875,
0.348388671875,
0.328369140625,
0.348388671875,
0.328369140625,
0.363037109375,
0.27685546875,
0.363037109375
],
[
0.369384765625,
0.348388671875,
0.449462890625,
0.348388671875,
0.449462890625,
0.363037109375,
0.369384765625,
0.363037109375
],
[
0.5185546875,
0.35107421875,
0.7099609375,
0.35107421875,
0.7099609375,
0.361083984375,
0.5185546875,
0.361083984375
],
[
0.1070556640625,
0.363037109375,
0.1502685546875,
0.363037109375,
0.1502685546875,
0.375,
0.1070556640625,
0.375
],
[
0.261474609375,
0.363037109375,
0.291748046875,
0.363037109375,
0.291748046875,
0.375,
0.261474609375,
0.375
],
[
0.310546875,
0.3623046875,
0.348876953125,
0.3623046875,
0.348876953125,
0.375732421875,
0.310546875,
0.375732421875
],
[
0.36767578125,
0.363037109375,
0.398681640625,
0.363037109375,
0.398681640625,
0.375,
0.36767578125,
0.375
],
[
0.4150390625,
0.363037109375,
0.453369140625,
0.363037109375,
0.453369140625,
0.375,
0.4150390625,
0.375
],
[
0.108642578125,
0.380126953125,
0.134033203125,
0.380126953125,
0.134033203125,
0.393310546875,
0.108642578125,
0.393310546875
],
[
0.256591796875,
0.380126953125,
0.29736328125,
0.380126953125,
0.29736328125,
0.39453125,
0.256591796875,
0.39453125
],
[
0.36279296875,
0.380126953125,
0.40283203125,
0.380126953125,
0.40283203125,
0.39453125,
0.36279296875,
0.39453125
],
[
0.4150390625,
0.37890625,
0.455810546875,
0.37890625,
0.455810546875,
0.39453125,
0.4150390625,
0.39453125
],
[
0.5556640625,
0.380126953125,
0.73046875,
0.380126953125,
0.73046875,
0.393310546875,
0.5556640625,
0.393310546875
],
[
0.5185546875,
0.380615234375,
0.5439453125,
0.380615234375,
0.5439453125,
0.392578125,
0.5185546875,
0.392578125
],
[
0.310546875,
0.380615234375,
0.348876953125,
0.380615234375,
0.348876953125,
0.39453125,
0.310546875,
0.39453125
],
[
0.256591796875,
0.392578125,
0.298095703125,
0.392578125,
0.298095703125,
0.4072265625,
0.256591796875,
0.4072265625
],
[
0.4150390625,
0.392578125,
0.455810546875,
0.392578125,
0.455810546875,
0.4072265625,
0.4150390625,
0.4072265625
],
[
0.310546875,
0.393310546875,
0.348876953125,
0.393310546875,
0.348876953125,
0.4072265625,
0.310546875,
0.4072265625
],
[
0.1094970703125,
0.393310546875,
0.200927734375,
0.393310546875,
0.200927734375,
0.40478515625,
0.1094970703125,
0.40478515625
],
[
0.363525390625,
0.39453125,
0.40185546875,
0.39453125,
0.40185546875,
0.406005859375,
0.363525390625,
0.406005859375
],
[
0.51708984375,
0.397705078125,
0.9140625,
0.397705078125,
0.9140625,
0.410888671875,
0.51708984375,
0.410888671875
],
[
0.256591796875,
0.406005859375,
0.298095703125,
0.406005859375,
0.298095703125,
0.42041015625,
0.256591796875,
0.42041015625
],
[
0.4150390625,
0.40478515625,
0.455810546875,
0.40478515625,
0.455810546875,
0.42041015625,
0.4150390625,
0.42041015625
],
[
0.310546875,
0.4072265625,
0.348876953125,
0.4072265625,
0.348876953125,
0.42041015625,
0.310546875,
0.42041015625
],
[
0.1094970703125,
0.4072265625,
0.1641845703125,
0.4072265625,
0.1641845703125,
0.417236328125,
0.1094970703125,
0.417236328125
],
[
0.36279296875,
0.4072265625,
0.40185546875,
0.4072265625,
0.40185546875,
0.41845703125,
0.36279296875,
0.41845703125
],
[
0.108642578125,
0.417236328125,
0.2459716796875,
0.417236328125,
0.2459716796875,
0.43115234375,
0.108642578125,
0.43115234375
],
[
0.51708984375,
0.412353515625,
0.9140625,
0.412353515625,
0.9140625,
0.425537109375,
0.51708984375,
0.425537109375
],
[
0.25732421875,
0.41845703125,
0.295654296875,
0.41845703125,
0.295654296875,
0.429931640625,
0.25732421875,
0.429931640625
],
[
0.310546875,
0.41845703125,
0.348876953125,
0.41845703125,
0.348876953125,
0.429931640625,
0.310546875,
0.429931640625
],
[
0.363525390625,
0.41845703125,
0.40185546875,
0.41845703125,
0.40185546875,
0.429931640625,
0.363525390625,
0.429931640625
],
[
0.416748046875,
0.41845703125,
0.455078125,
0.41845703125,
0.455078125,
0.429931640625,
0.416748046875,
0.429931640625
],
[
0.3095703125,
0.429931640625,
0.3505859375,
0.429931640625,
0.3505859375,
0.4443359375,
0.3095703125,
0.4443359375
],
[
0.4150390625,
0.429931640625,
0.455810546875,
0.429931640625,
0.455810546875,
0.4443359375,
0.4150390625,
0.4443359375
],
[
0.51708984375,
0.425537109375,
0.9140625,
0.425537109375,
0.9140625,
0.439453125,
0.51708984375,
0.439453125
],
[
0.256591796875,
0.43115234375,
0.294189453125,
0.4287109375,
0.295654296875,
0.44140625,
0.25732421875,
0.443115234375
],
[
0.108642578125,
0.43115234375,
0.21728515625,
0.43115234375,
0.21728515625,
0.4443359375,
0.108642578125,
0.4443359375
],
[
0.363525390625,
0.43115234375,
0.40185546875,
0.43115234375,
0.40185546875,
0.442626953125,
0.363525390625,
0.442626953125
],
[
0.256591796875,
0.442626953125,
0.29736328125,
0.442626953125,
0.29736328125,
0.45703125,
0.256591796875,
0.45703125
],
[
0.3095703125,
0.442626953125,
0.3505859375,
0.442626953125,
0.3505859375,
0.45703125,
0.3095703125,
0.45703125
],
[
0.361083984375,
0.442626953125,
0.40283203125,
0.442626953125,
0.40283203125,
0.45703125,
0.361083984375,
0.45703125
],
[
0.4150390625,
0.442626953125,
0.455810546875,
0.442626953125,
0.455810546875,
0.45703125,
0.4150390625,
0.45703125
],
[
0.51708984375,
0.439453125,
0.9140625,
0.439453125,
0.9140625,
0.45263671875,
0.51708984375,
0.45263671875
],
[
0.51708984375,
0.453857421875,
0.91259765625,
0.453857421875,
0.91259765625,
0.46728515625,
0.51708984375,
0.46728515625
],
[
0.108642578125,
0.443115234375,
0.223876953125,
0.443115234375,
0.223876953125,
0.45703125,
0.108642578125,
0.45703125
],
[
0.253173828125,
0.455078125,
0.298095703125,
0.45263671875,
0.2998046875,
0.468505859375,
0.2548828125,
0.47021484375
],
[
0.359375,
0.455078125,
0.404296875,
0.455078125,
0.404296875,
0.4697265625,
0.359375,
0.4697265625
],
[
0.414306640625,
0.455078125,
0.45751953125,
0.455078125,
0.45751953125,
0.4697265625,
0.414306640625,
0.4697265625
],
[
0.306396484375,
0.455810546875,
0.35302734375,
0.455810546875,
0.35302734375,
0.47021484375,
0.306396484375,
0.47021484375
],
[
0.1094970703125,
0.455810546875,
0.143798828125,
0.455810546875,
0.143798828125,
0.46728515625,
0.1094970703125,
0.46728515625
],
[
0.1070556640625,
0.46728515625,
0.2156982421875,
0.46728515625,
0.2156982421875,
0.48095703125,
0.1070556640625,
0.48095703125
],
[
0.256591796875,
0.46728515625,
0.29736328125,
0.46728515625,
0.29736328125,
0.481689453125,
0.256591796875,
0.481689453125
],
[
0.3095703125,
0.46728515625,
0.3505859375,
0.46728515625,
0.3505859375,
0.481689453125,
0.3095703125,
0.481689453125
],
[
0.36279296875,
0.46728515625,
0.40283203125,
0.46728515625,
0.40283203125,
0.481689453125,
0.36279296875,
0.481689453125
],
[
0.416748046875,
0.4697265625,
0.455810546875,
0.4697265625,
0.455810546875,
0.48095703125,
0.416748046875,
0.48095703125
],
[
0.51708984375,
0.46728515625,
0.91259765625,
0.46728515625,
0.91259765625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.51708984375,
0.481689453125,
0.9140625,
0.481689453125,
0.9140625,
0.49560546875,
0.51708984375,
0.49560546875
],
[
0.10296630859375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.10296630859375,
0.5087890625
],
[
0.515625,
0.494384765625,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.515625,
0.50830078125
],
[
0.0849609375,
0.51025390625,
0.483642578125,
0.51025390625,
0.483642578125,
0.5234375,
0.0849609375,
0.5234375
],
[
0.51708984375,
0.5087890625,
0.9140625,
0.5087890625,
0.9140625,
0.52294921875,
0.51708984375,
0.52294921875
],
[
0.0849609375,
0.5234375,
0.48193359375,
0.5234375,
0.48193359375,
0.537109375,
0.0849609375,
0.537109375
],
[
0.515625,
0.52294921875,
0.9140625,
0.52294921875,
0.9140625,
0.537109375,
0.515625,
0.537109375
],
[
0.0849609375,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.515625,
0.5361328125,
0.62353515625,
0.5361328125,
0.62353515625,
0.5498046875,
0.515625,
0.5498046875
],
[
0.0849609375,
0.5517578125,
0.483642578125,
0.5517578125,
0.483642578125,
0.56494140625,
0.0849609375,
0.56494140625
],
[
0.0849609375,
0.564453125,
0.483642578125,
0.564453125,
0.483642578125,
0.57763671875,
0.0849609375,
0.57763671875
],
[
0.5185546875,
0.568359375,
0.5439453125,
0.568359375,
0.5439453125,
0.580078125,
0.5185546875,
0.580078125
],
[
0.55615234375,
0.5673828125,
0.7158203125,
0.5673828125,
0.7158203125,
0.58056640625,
0.55615234375,
0.58056640625
],
[
0.0849609375,
0.5791015625,
0.48193359375,
0.5791015625,
0.48193359375,
0.59228515625,
0.0849609375,
0.59228515625
],
[
0.51708984375,
0.58544921875,
0.91259765625,
0.58544921875,
0.91259765625,
0.5986328125,
0.51708984375,
0.5986328125
],
[
0.0849609375,
0.59130859375,
0.483642578125,
0.59130859375,
0.483642578125,
0.60498046875,
0.0849609375,
0.60498046875
],
[
0.51708984375,
0.599609375,
0.9140625,
0.599609375,
0.9140625,
0.61279296875,
0.51708984375,
0.61279296875
],
[
0.0849609375,
0.60595703125,
0.483642578125,
0.60595703125,
0.483642578125,
0.619140625,
0.0849609375,
0.619140625
],
[
0.51708984375,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.51708984375,
0.626953125
],
[
0.0849609375,
0.62060546875,
0.48193359375,
0.62060546875,
0.48193359375,
0.6337890625,
0.0849609375,
0.6337890625
],
[
0.515625,
0.62744140625,
0.9140625,
0.626953125,
0.9140625,
0.64013671875,
0.515625,
0.6416015625
],
[
0.0849609375,
0.6337890625,
0.484375,
0.6337890625,
0.484375,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.51708984375,
0.6416015625,
0.91259765625,
0.6416015625,
0.91259765625,
0.65478515625,
0.51708984375,
0.65478515625
],
[
0.08416748046875,
0.64794921875,
0.48193359375,
0.646484375,
0.48193359375,
0.6611328125,
0.08416748046875,
0.662109375
],
[
0.51708984375,
0.65478515625,
0.9140625,
0.65478515625,
0.9140625,
0.66845703125,
0.51708984375,
0.66845703125
],
[
0.0849609375,
0.6611328125,
0.48046875,
0.6611328125,
0.48046875,
0.6748046875,
0.0849609375,
0.6748046875
],
[
0.51708984375,
0.66845703125,
0.9140625,
0.66845703125,
0.9140625,
0.681640625,
0.51708984375,
0.681640625
],
[
0.08416748046875,
0.67529296875,
0.483642578125,
0.67529296875,
0.483642578125,
0.689453125,
0.08416748046875,
0.689453125
],
[
0.51708984375,
0.68115234375,
0.9140625,
0.68115234375,
0.9140625,
0.6943359375,
0.51708984375,
0.6943359375
],
[
0.0849609375,
0.68798828125,
0.48193359375,
0.68798828125,
0.48193359375,
0.7021484375,
0.0849609375,
0.7021484375
],
[
0.51708984375,
0.6962890625,
0.859375,
0.6962890625,
0.859375,
0.71044921875,
0.51708984375,
0.71044921875
],
[
0.0849609375,
0.70263671875,
0.48046875,
0.7041015625,
0.48046875,
0.71728515625,
0.0849609375,
0.71630859375
],
[
0.0849609375,
0.71630859375,
0.48193359375,
0.71630859375,
0.48193359375,
0.72998046875,
0.0849609375,
0.72998046875
],
[
0.521484375,
0.72802734375,
0.53125,
0.73095703125,
0.52685546875,
0.73779296875,
0.51708984375,
0.73486328125
],
[
0.5439453125,
0.72607421875,
0.67236328125,
0.72607421875,
0.67236328125,
0.7392578125,
0.5439453125,
0.7392578125
],
[
0.515625,
0.744140625,
0.9140625,
0.74365234375,
0.9140625,
0.75830078125,
0.515625,
0.7587890625
],
[
0.0849609375,
0.73095703125,
0.48193359375,
0.73095703125,
0.48193359375,
0.744140625,
0.0849609375,
0.744140625
],
[
0.0849609375,
0.74560546875,
0.48193359375,
0.74560546875,
0.48193359375,
0.7587890625,
0.0849609375,
0.7587890625
],
[
0.08416748046875,
0.7568359375,
0.48193359375,
0.75830078125,
0.48193359375,
0.77294921875,
0.08416748046875,
0.771484375
],
[
0.51708984375,
0.7587890625,
0.91259765625,
0.7587890625,
0.91259765625,
0.77294921875,
0.51708984375,
0.77294921875
],
[
0.0849609375,
0.77294921875,
0.483642578125,
0.77294921875,
0.483642578125,
0.7861328125,
0.0849609375,
0.7861328125
],
[
0.515625,
0.77294921875,
0.9140625,
0.77294921875,
0.9140625,
0.7861328125,
0.515625,
0.7861328125
],
[
0.0849609375,
0.7861328125,
0.483642578125,
0.7861328125,
0.483642578125,
0.7998046875,
0.0849609375,
0.7998046875
],
[
0.51708984375,
0.7861328125,
0.9140625,
0.7861328125,
0.9140625,
0.7998046875,
0.51708984375,
0.7998046875
],
[
0.08660888671875,
0.8017578125,
0.48046875,
0.8017578125,
0.48046875,
0.8125,
0.08660888671875,
0.8125
],
[
0.51708984375,
0.80029296875,
0.9140625,
0.80029296875,
0.9140625,
0.814453125,
0.51708984375,
0.814453125
],
[
0.08416748046875,
0.8125,
0.48193359375,
0.81298828125,
0.48193359375,
0.82763671875,
0.08416748046875,
0.8271484375
],
[
0.51708984375,
0.814453125,
0.91259765625,
0.814453125,
0.91259765625,
0.82763671875,
0.51708984375,
0.82763671875
],
[
0.0849609375,
0.82763671875,
0.48193359375,
0.82763671875,
0.48193359375,
0.84130859375,
0.0849609375,
0.84130859375
],
[
0.51708984375,
0.82763671875,
0.9140625,
0.82763671875,
0.9140625,
0.84130859375,
0.51708984375,
0.84130859375
],
[
0.08416748046875,
0.84228515625,
0.2459716796875,
0.84228515625,
0.2459716796875,
0.85595703125,
0.08416748046875,
0.85595703125
],
[
0.51708984375,
0.84228515625,
0.9140625,
0.84228515625,
0.9140625,
0.85595703125,
0.51708984375,
0.85595703125
],
[
0.101318359375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.869140625,
0.101318359375,
0.869140625
],
[
0.51708984375,
0.85595703125,
0.9140625,
0.85595703125,
0.9140625,
0.869140625,
0.51708984375,
0.869140625
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.51708984375,
0.869140625,
0.9140625,
0.869140625,
0.9140625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.08660888671875,
0.88525390625,
0.48046875,
0.88525390625,
0.48046875,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.51708984375,
0.8837890625,
0.734375,
0.8837890625,
0.734375,
0.89794921875,
0.51708984375,
0.89794921875
]
] |
[
"100",
"DIN+PosInWide",
"DIN+Combination.Position-wise combinationmoduleis em-",
"DIN+Combination",
"ployed based on DIN. Actual position is used for evaluation.",
"80",
"DPIN+ItemAction",
"DPIN-Transformer.Transformeris notused inDPIN",
"(",
"DPIN",
"DPIN. There is the proposed DPIN model, in which a tow-layer",
"60",
"transformer is adopted and dmodel = 64, h = 2 for self-attention.",
"DPIN+ItemAction.Weadd deepposition-wise interaction mod-",
"ule before the MLP layer of the base module, and candidate item",
"40",
"information is introduced in position-wise interest aggregation",
"and position-wise non-linear interaction.The experiment is the",
"20",
"upper bound of model performance in our methods, but the serving",
"performance is unacceptable.",
"16",
"24",
"32",
"40",
"48",
"56",
"64",
"#Candidate Items",
"8",
"3.2",
"OfflineEvaluations",
"Figure 2: Serving latency changes with the number of candi-",
"Table 1: Offline experimental results of compared methods",
"date items in different methods.",
"on the regular and randomized test sets.",
"the differences between the different methods on both regular and",
"Regular",
"Randomized",
"randomizedtrafficisconsistent.",
"MODEL",
"AUC",
"PAUC",
"AUC",
"PAUC",
"DIN",
"0.7818",
"0.7836",
"0.7223",
"ServingPerformance",
"3.3",
"0.7090",
"0.7696",
"0.7239",
"0.7109",
"DIN+POSINWIDE",
"0.7725",
"We retrieve some requests with different candidate item numbers",
"0.7735",
"0.7254",
"0.7128",
"DIN+PAL",
"0.7763",
"DIN+ACUTALPOSINWIDE",
"from the dataset to measure serving performance. As shown in",
"0.7928",
"0.7109",
"0.7938",
"0.7239",
"0.7172",
"0.7294",
"Figure 2, the serving latency of position-wise combination mod-",
"0.7970",
"DIN+COMBINATION",
"0.7985",
"0.7961",
"0.7148",
"0.7984",
"0.7283",
"ule is negligible compared to the DIN model since user sequence",
"operation has a large proportion of latency. The serving latency",
"DPIN-TRANSFORMER",
"0.7994",
"0.8015",
"0.7350",
"0.7216",
"DPIN",
"DPIN+ITEMACTION",
"0.7999",
"0.7223",
"0.8019",
"0.7356",
"of DPIN increases slowly as the number of items increases since",
"the deep position-wise interaction module has nothing to do with",
"Table 1 illustrates experimental results of the compared methods",
"items.Compared with DIPIN+ItemAction, the DPIN has a great",
"on both regular and randomized test sets. We first analyze the differ-",
"improvement in serving performance with little damage to model",
"ences between different methods on the regular traffic. Compared",
"performance, which shows that our proposed method is both effec-",
"with DIN, the DIN+PosIn Wide and DIN+PAL have a performance",
"tive and efficient.",
"degradation on AUC but an improvement on PAUC, which shows",
"that both methods effectively alleviate position bias but lead to in-",
"3.4",
"Online Evaluations",
"consistency between offline and online. The DIN+AcutalPosInWide",
"Online A/B test was conducted in the sponsored search advertising",
"solves the inconsistency by introducing actual position during eval-",
"system from 2021-01-08 to 2021-01-22.For the control group,10% of",
"uation, which is achievable by position-wise combination module.",
"users are randomly selected and presented with recommendation",
"But modeling position in the wide part leads to the position feature",
"generated byDIN+PoslnWide.For the experimental group,10% of",
"is only a bias, which can not improve PAUC. By employing our pro-",
"users are presented with recommendation generated by DPIN. The",
"posed position-wise combination module, the DIN+Combination",
"A/B test shows that the proposed DPIN has improved CTR by 2.25%",
"has 2.74% gain on AUC and 0.63% gain on PAUC compared with",
"and RPM (Revenue Per Mille) by 2.15% compared with baseline.For",
"DIN+PoslnWide, achieving consistency and alleviate position bias",
"now, DPIN has been deployed online and serves the main traffic,",
"further, which shows the position bias is not independent. Further",
"which contributes a significant business revenue growth.",
"more,DPIN models deep non-linear interaction among position",
"context and user, and eliminates position bias existed in the user",
"4",
"CONCLUSIONS",
"In this paper, a novel model Deep Position-wise Interaction Net-",
"sequence by position-wise method, which has 0.24% gain on AUC",
"and 0.44% gain on PAUC compared with DIN+Combination. The",
"effect of DPIN-Transformer explains that it is necessary to adopt",
"Work (DPIN) is proposed to mitigate position bias, which efficiently",
"transformer for interaction among different positions. And the com-",
" combine all candidate items and positions for estimating CTR at",
"parison between DPIN and DPIN+ItemAction shows that DPIN is",
"each position, achieving consistency between offline and online",
"closeto thebruteforcemethod on bothAUC andPAUC.Ascanbe",
"At the same time, the deep non-linear interaction among position,",
"seen finally, the DPIN has 2.98% gain on AUC and 1.07% gain 0n",
"context and user is available by the model. In order to evaluate",
"PAUC relative to the DIN+PoslnWide, which is a baseline in the",
"our method in position bias problem, we propose a new evaluation",
"advertising system online.",
"metrics PAUC and the offline experiments show that the proposed",
"In order to ensure that our method can learn the position bias",
"DPIN outperforms the compared methods efficacy and efficiency.",
"instead of overfitting the selection bias of the system, we further",
"For now, the DPIN is deployed in a sponsored search advertising",
"evaluate our methods on randomized traffic.The results show that",
"system and serving the main traffic."
] |
{
"arxiv_id": "2106.05482",
"language": "en",
"page_idx": 3
}
|
||
2101.08942-en-0
|
2101.08942-en-0.jpg
|
## $\textbf{Enriching Non- Autoregressive Transformer with Syntactic and Semantic}$ Structures for Neural Machine Translation
$\textbf{Ye Liu}^1,\textbf{Yao Wan}^2,\textbf{Jian-Guo Zhang}^1,\textbf{Wenting Zhao}^1,\textbf{Philip S. Yu}^1$
$^{1}$Department of Computer Science, University of Illinois at Chicago, Chicago, IL, USA
$^{2}$School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China
$\{$yliu279, jzhan51, wzhao41, psyu$\}@$uic.edu, wanyao@hust.edu.cn
## Abstract
The non-autoregressive models have boosted the efficiency of neural machine translation through parallelized decoding at the cost of effectiveness, when comparing with the autoregressive counterparts. In this paper, we claim that the syntactic and semantic structures among natural language are critical for nonautoregressive machine translation and can further improve the performance. However, these structures are rarely considered in the existing non-autoregressive models. Inspired by this intuition, we propose to incorporate the explicit syntactic and semantic structures of languages into a non-autoregressive Transformer, for the task of neural machine translation. Moreover, we also consider the intermediate latent alignment within target sentences to better learn the long-term token dependencies. Experimental results on two real-world datasets (i.e., WMT14 En-De and WMT16 En-Ro) show that our model achieves a significantly faster speed, as well as keeps the translation quality when compared with several state-of-theart non-autoregressive models.
## 1 Introduction
Recently, non-autoregressive models (Gu et al., 2018), which aim to enable the parallel generation of output tokens without sacrificing translation quality, have attracted much attention. Although the non-autoregressive models have greatly sped up the inference process for real-time neural machine translation (NMT) (Gu et al., 2018), their performance is considerably worse than that of autoregressive counterparts. Most previous works attribute the poor performance to the inevitable conditional independence issue when predicting target tokens, and many variants have been proposed to solve it. For example, several techniques in nonautoregressive models are investigated to mitigate the trade-off between speedup and performance,
including iterative refnement (Lee et al., 2018), insertion-based models (Chan et al., 2019; Stern et al., 2019), latent variable based models (Kaiser et al., 2018; Shu et al., 2020), CTC models (Libovický and Helcl, 2018; Saharia et al., 2020), alternative loss function based models (Wei et al., 2019; Wang et al., 2019; Shao et al., 2020), and masked language models (Ghazvininejad et al., 2019,2020).
Although these works have tried to narrow the performance gap between autoregressive and nonautoregressive models, and have achieved some improvements on machine translation, the nonautoregressive models still suffer from syntactic and semantic limitations. That is, the translations of non-autoregressive models tend to contain incoherent phrases (e.g. repetitive words), and some informative tokens on the source side are absent. It is because that in non-autoregressive models, each token in the target sentence is generated independently. Consequently, it will cause the multimodality issue, i.e., the non-autoregressive models cannot model the multimodal distribution of target sequences properly (Gu et al., 2018).
One key observation to mitigate the syntactic and semantic error is that source and target translated sentences follow the same structure, which can be reflected from Part-Of-Speech (POS) tags and Named Entity Recognition (NER) labels. Briefly, POS, which aims to assign labels to words to indicate their categories by considering the longdistance structure of sentences, can help the model learn the syntactic structure to avoid generating the repetitive words. Likewise, NER, which discovers the proper nouns and verb of sentences, naturally helps the model to recognize some meaningful semantic tokens that may improve the quality of translation. This observation motivates us to leverage the syntactic as well as semantic structures of natural language to improve the performance of
|
[
[
0.1292724609375,
0.0831298828125,
0.875,
0.0831298828125,
0.875,
0.097412109375,
0.1292724609375,
0.097412109375
],
[
0.280517578125,
0.10272216796875,
0.72119140625,
0.10272216796875,
0.72119140625,
0.1175537109375,
0.280517578125,
0.1175537109375
],
[
0.210693359375,
0.146728515625,
0.791015625,
0.1455078125,
0.791015625,
0.163330078125,
0.210693359375,
0.1644287109375
],
[
0.181396484375,
0.1644287109375,
0.82373046875,
0.1650390625,
0.82373046875,
0.179931640625,
0.181396484375,
0.1787109375
],
[
0.09149169921875,
0.18115234375,
0.916015625,
0.1817626953125,
0.916015625,
0.196533203125,
0.09149169921875,
0.1953125
],
[
0.235107421875,
0.198974609375,
0.77734375,
0.198974609375,
0.77734375,
0.213134765625,
0.235107421875,
0.213134765625
],
[
0.26611328125,
0.26416015625,
0.343505859375,
0.26416015625,
0.343505859375,
0.279052734375,
0.26611328125,
0.279052734375
],
[
0.51318359375,
0.26416015625,
0.8818359375,
0.265380859375,
0.8818359375,
0.27978515625,
0.51318359375,
0.279052734375
],
[
0.51318359375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.51318359375,
0.295654296875
],
[
0.14697265625,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3046875,
0.14697265625,
0.3046875
],
[
0.51318359375,
0.297607421875,
0.884765625,
0.297607421875,
0.884765625,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.14697265625,
0.305908203125,
0.459228515625,
0.305908203125,
0.459228515625,
0.319580078125,
0.14697265625,
0.319580078125
],
[
0.0235137939453125,
0.319580078125,
0.060455322265625,
0.319580078125,
0.060455322265625,
0.438232421875,
0.0235137939453125,
0.438232421875
],
[
0.51318359375,
0.314208984375,
0.88330078125,
0.314208984375,
0.88330078125,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.14697265625,
0.320556640625,
0.4609375,
0.320556640625,
0.4609375,
0.334228515625,
0.14697265625,
0.334228515625
],
[
0.14697265625,
0.334228515625,
0.462646484375,
0.334228515625,
0.462646484375,
0.347412109375,
0.14697265625,
0.347412109375
],
[
0.51318359375,
0.330078125,
0.88671875,
0.330078125,
0.88671875,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.14697265625,
0.347412109375,
0.459228515625,
0.347412109375,
0.459228515625,
0.361083984375,
0.14697265625,
0.361083984375
],
[
0.515625,
0.347412109375,
0.884765625,
0.347412109375,
0.884765625,
0.358154296875,
0.515625,
0.358154296875
],
[
0.1485595703125,
0.364013671875,
0.4609375,
0.364013671875,
0.4609375,
0.374755859375,
0.1485595703125,
0.374755859375
],
[
0.51318359375,
0.361083984375,
0.88330078125,
0.361083984375,
0.88330078125,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.14697265625,
0.377197265625,
0.462646484375,
0.377197265625,
0.462646484375,
0.39013671875,
0.14697265625,
0.39013671875
],
[
0.5146484375,
0.37890625,
0.88671875,
0.37890625,
0.88671875,
0.393798828125,
0.5146484375,
0.393798828125
],
[
0.14697265625,
0.391357421875,
0.46337890625,
0.391357421875,
0.46337890625,
0.405029296875,
0.14697265625,
0.405029296875
],
[
0.515625,
0.3955078125,
0.60791015625,
0.3955078125,
0.60791015625,
0.40673828125,
0.515625,
0.40673828125
],
[
0.14697265625,
0.405029296875,
0.4609375,
0.405029296875,
0.4609375,
0.418701171875,
0.14697265625,
0.418701171875
],
[
0.5322265625,
0.410888671875,
0.88330078125,
0.410888671875,
0.88330078125,
0.42578125,
0.5322265625,
0.42578125
],
[
0.14697265625,
0.419189453125,
0.459228515625,
0.419189453125,
0.459228515625,
0.432861328125,
0.14697265625,
0.432861328125
],
[
0.51318359375,
0.427490234375,
0.88330078125,
0.427490234375,
0.88330078125,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.0235137939453125,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0235137939453125,
0.5185546875
],
[
0.14697265625,
0.43408203125,
0.4609375,
0.43408203125,
0.4609375,
0.44775390625,
0.14697265625,
0.44775390625
],
[
0.14697265625,
0.44775390625,
0.4609375,
0.44775390625,
0.4609375,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.5146484375,
0.443603515625,
0.884765625,
0.443603515625,
0.884765625,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.1485595703125,
0.46142578125,
0.459228515625,
0.4619140625,
0.459228515625,
0.4755859375,
0.1485595703125,
0.474365234375
],
[
0.51318359375,
0.460205078125,
0.884765625,
0.460205078125,
0.884765625,
0.473876953125,
0.51318359375,
0.473876953125
],
[
0.14697265625,
0.476806640625,
0.4609375,
0.476806640625,
0.4609375,
0.490478515625,
0.14697265625,
0.490478515625
],
[
0.51318359375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.14697265625,
0.490478515625,
0.4609375,
0.490478515625,
0.4609375,
0.50439453125,
0.14697265625,
0.50439453125
],
[
0.51318359375,
0.4912109375,
0.88330078125,
0.4912109375,
0.88330078125,
0.5048828125,
0.51318359375,
0.5048828125
],
[
0.14697265625,
0.5048828125,
0.4609375,
0.5048828125,
0.4609375,
0.5185546875,
0.14697265625,
0.5185546875
],
[
0.51318359375,
0.50830078125,
0.88671875,
0.50732421875,
0.88671875,
0.521484375,
0.51318359375,
0.5224609375
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.1485595703125,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53076171875,
0.1485595703125,
0.53076171875
],
[
0.51318359375,
0.52490234375,
0.884765625,
0.52490234375,
0.884765625,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.14697265625,
0.533203125,
0.459228515625,
0.533203125,
0.459228515625,
0.54638671875,
0.14697265625,
0.54638671875
],
[
0.51123046875,
0.5390625,
0.884765625,
0.5390625,
0.884765625,
0.552734375,
0.51123046875,
0.552734375
],
[
0.145263671875,
0.5458984375,
0.4609375,
0.54638671875,
0.4609375,
0.56103515625,
0.145263671875,
0.56005859375
],
[
0.14697265625,
0.56005859375,
0.459228515625,
0.56005859375,
0.459228515625,
0.57373046875,
0.14697265625,
0.57373046875
],
[
0.51318359375,
0.55712890625,
0.88671875,
0.55712890625,
0.88671875,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.14697265625,
0.576171875,
0.459228515625,
0.576171875,
0.459228515625,
0.5888671875,
0.14697265625,
0.5888671875
],
[
0.51318359375,
0.572265625,
0.88671875,
0.572265625,
0.88671875,
0.5859375,
0.51318359375,
0.5859375
],
[
0.1485595703125,
0.59130859375,
0.4609375,
0.59130859375,
0.4609375,
0.6015625,
0.1485595703125,
0.6015625
],
[
0.51318359375,
0.5888671875,
0.88671875,
0.5888671875,
0.88671875,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.14697265625,
0.60400390625,
0.462646484375,
0.60400390625,
0.462646484375,
0.61767578125,
0.14697265625,
0.61767578125
],
[
0.51318359375,
0.60498046875,
0.88671875,
0.60498046875,
0.88671875,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.14697265625,
0.6181640625,
0.355224609375,
0.6181640625,
0.355224609375,
0.6318359375,
0.14697265625,
0.6318359375
],
[
0.51123046875,
0.619140625,
0.884765625,
0.62060546875,
0.884765625,
0.63525390625,
0.51123046875,
0.63427734375
],
[
0.51123046875,
0.63720703125,
0.7900390625,
0.63525390625,
0.7900390625,
0.6484375,
0.51123046875,
0.65087890625
],
[
0.124267578125,
0.6484375,
0.130126953125,
0.6484375,
0.130126953125,
0.65380859375,
0.124267578125,
0.65380859375
],
[
0.130126953125,
0.6435546875,
0.26025390625,
0.6435546875,
0.26025390625,
0.65771484375,
0.130126953125,
0.65771484375
],
[
0.5322265625,
0.65283203125,
0.88330078125,
0.65283203125,
0.88330078125,
0.66748046875,
0.5322265625,
0.66748046875
],
[
0.12005615234375,
0.66943359375,
0.490234375,
0.66943359375,
0.490234375,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.51318359375,
0.67041015625,
0.88330078125,
0.67041015625,
0.88330078125,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.118408203125,
0.68408203125,
0.490234375,
0.68603515625,
0.490234375,
0.69970703125,
0.118408203125,
0.69775390625
],
[
0.51318359375,
0.68603515625,
0.88330078125,
0.68603515625,
0.88330078125,
0.69970703125,
0.51318359375,
0.69970703125
],
[
0.12005615234375,
0.70166015625,
0.489501953125,
0.70166015625,
0.489501953125,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.5146484375,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.5146484375,
0.71533203125
],
[
0.12005615234375,
0.71875,
0.489501953125,
0.71875,
0.489501953125,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.51123046875,
0.71630859375,
0.8818359375,
0.71728515625,
0.8818359375,
0.732421875,
0.51123046875,
0.73095703125
],
[
0.12005615234375,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.51318359375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.51318359375,
0.74755859375
],
[
0.12176513671875,
0.751953125,
0.490234375,
0.751953125,
0.490234375,
0.7626953125,
0.12176513671875,
0.7626953125
],
[
0.5146484375,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12176513671875,
0.76708984375,
0.48779296875,
0.76708984375,
0.48779296875,
0.77783203125,
0.12176513671875,
0.77783203125
],
[
0.5146484375,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.118408203125,
0.783203125,
0.490234375,
0.78076171875,
0.490234375,
0.79443359375,
0.118408203125,
0.79638671875
],
[
0.51123046875,
0.78076171875,
0.88330078125,
0.7822265625,
0.88330078125,
0.79638671875,
0.51123046875,
0.7958984375
],
[
0.12005615234375,
0.798828125,
0.490234375,
0.798828125,
0.490234375,
0.8125,
0.12005615234375,
0.8125
],
[
0.51123046875,
0.79736328125,
0.88330078125,
0.79638671875,
0.88330078125,
0.81103515625,
0.51123046875,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.49365234375,
0.81396484375,
0.49365234375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51123046875,
0.81396484375,
0.884765625,
0.8134765625,
0.884765625,
0.82763671875,
0.51123046875,
0.8291015625
],
[
0.118408203125,
0.83056640625,
0.491943359375,
0.83056640625,
0.491943359375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51123046875,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.51123046875,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.491943359375,
0.84716796875,
0.491943359375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.51123046875,
0.84716796875,
0.88671875,
0.84716796875,
0.88671875,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.495361328125,
0.86328125,
0.495361328125,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51123046875,
0.86328125,
0.88671875,
0.86328125,
0.88671875,
0.876953125,
0.51123046875,
0.876953125
],
[
0.12005615234375,
0.8798828125,
0.495361328125,
0.8798828125,
0.495361328125,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.5087890625,
0.8798828125,
0.884765625,
0.87744140625,
0.884765625,
0.89111328125,
0.5087890625,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.495361328125,
0.8955078125,
0.495361328125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.88671875,
0.8955078125,
0.88671875,
0.91015625,
0.51318359375,
0.91015625
]
] |
[
"Enriching Non-Autoregressive Transformer with Syntactic and Semantic",
"Structures for Neural Machine Translation",
"Ye Liu', Yao Wan?, Jian-Guo Zhang', Wenting Zhao', Philip S. Yu",
"1Department of Computer Science, University of Illinois at Chicago, Chicago, IL, USA",
"2School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China",
"(yliu279, jzhan51, wzhao41, psyu)@uic.edu, wanyao@hust.edu.cn",
"Abstract",
"including iterative refinement (Lee et al., 2018)",
"insertion-based models (Chan et al., 2019; Stern",
"The non-autoregressive models have boosted",
"et al., 2019), latent variable based models (Kaiser",
"the efficiency of neural machine translation",
"",
"et al., 2018; Shu et al., 2020), CTC models (Li-",
"through parallelized decoding at the cost of ef",
"fectiveness, when comparing with the autore-",
"bovicky and Helcl, 2018; Saharia et al., 2020), al-",
"gressive counterparts. In this paper, we claim",
"ternativeloss function based models (Wei et al..",
"that the syntactic and semantic structures",
"2019: Wang et al.. 2019: Shao et al., 2020), and",
"among natural language are critical for non-",
"masked language models (Ghazvininejad et al.,",
"autoregressive machine translation and can fur-",
"2019.2020)",
"ther improve the performance. However, these",
"Although these works have tried to narrow the",
"structures are rarely considered in the existing",
"performance gap between autoregressive and non-",
"['s]",
"non-autoregressive models. Inspired by this in-",
"tuition, we propose to incorporate the explicit",
"autoregressive models, and have achieved some",
"syntactic and semantic structures of languages",
"improvements on machine translation, the non-",
"into a non-autoregressive Transformer, for the",
"autoregressive models still suffer from syntactic",
"task of neural machine translation. Moreover",
"and semantic limitations. That is, the translations",
"we also consider the intermediate latent align-",
"of non-autoregressive models tend to contain in-",
"1268010210",
"ment within target sentences to better learn",
"coherent phrases (e.g. repetitive words), and some",
"the long-term token dependencies. Experimen-",
"informative tokens on the source side are absent.",
"tal results on two real-world datasets (i.e.,",
"WMT14 En-De and WMT16 En-Ro) show",
"It is because that in non-autoregressive models,",
"that our model achieves a significantly faster",
"each token in the target sentence is generated in-",
"speed,as well as keeps thetranslation qual-",
"dependently. Consequently, it will cause the multi-",
"ity when compared with several state-of-the-",
"modality issue, i.e., the non-autoregressive models",
"art non-autoregressive models.",
" cannot model the multimodal distribution of target",
"sequences properly (Gu et al., 2018).",
"1",
"Introduction",
"One key observation to mitigate the syntactic and",
"Recently, non-autoregressive models (Gu et al.,",
"semantic error is that source and target translated",
"2018), which aim to enable the parallel genera-",
"sentences follow the same structure, which can",
"tion of output tokens without sacrificing translation",
"be reflected from Part-Of-Speech (POS) tags and",
"quality, have attracted much attention. Although",
"Named Entity Recognition (NER) labels. Briefly",
"the non-autoregressive models have greatly sped",
"POS, which aims to assign labels to words to in-",
"uptheinferenceprocessforreal-timeneuralma-",
"dicate their categories by considering the long",
"chinetranslation (NMT)(Gu et al..2018),their",
"distancestructureofsentences,canhelpthemodel",
"performance is considerably worse than that of",
"learn the syntactic structure to avoid generating",
"autoregressive counterparts. Most previous works",
"the repetitive words. Likewise, NER, which discov",
"attribute the poor performance to the inevitable con-",
"ers the proper nouns and verb of sentences, natu-",
"ditional independence issue when predicting target",
"rally helps the model to recognize some meaningful",
"tokens, and many variants have been proposed to",
" semantic tokens that may improve the quality of",
"solve it. For example, several techniques in non-",
" translation. This observation motivates us to lever-",
"autoregressive models are investigated to mitigate",
" age the syntactic as well as semantic structures of",
"the trade-off between speedup and performance,",
"natural language to improve the performance of"
] |
{
"arxiv_id": "2101.08942",
"language": "en",
"page_idx": 0
}
|
||
2101.08942-en-2
|
2101.08942-en-2.jpg
|
<FigureHere>
Figure 1: An overview of the proposed SNAT for neural machine translation.
tokenizer (Sennrich et al., 2016). In parallel, words in the input sequence are passed to POS and NER annotators to extract explicit syntactic and semantic structures, and the corresponding embeddings are aggregated by a linear layer to form the final syntax and semantic structure-aware embedding. The SNAT model copies the structured encoder input as the decoder input and generates the translated sentences and labels.
One of the most important properties of SNAT is that it naturally introduces syntactic and semantic information when taking the structure-aware information as inputs and generating both words and labels. More precisely, given a source sentence $\mathbf{x}$, source label sequence $\mathbf{L_{x}}$,the conditional probability of a target translation y and target label sequence $\mathbf{L_{y}}$ is:
et al., 2005) for NER to annotate every token of input sequence with semantic labels. Given a specific sentence, there would be predicate-argument structures. Since the input sequence is segmented into subword units using byte-pair encoding (Sennrich et al., 2016), we assign the same label to all subwords tokenized from the same word. As shown in Figure 1, the word “Ancelott" is tokenized as “An@@” and“Celott". The corresponding POS tags are PRON and PRON while the corresponding NER tags are B\_PERSON and I\_PERSON. For the text“Is An@@ Celotti the man for the job ?”, the corresponding POS tag set is $\{$AUX,PRON, PRON, DET, NOUN, ADP, DET, NOUN, PUNCT} and the NER tag set is $\{$O, B$\_$PERSON, I$\_$PERSON, $0,0,0,0,0,0,0\}.$ The data flow of the proposed model is also shown in Figure 1.
### 3.2 Encoder
$$P_{\mathbf{SNAT}}(\mathbf{y},\mathbf{L_{y}}|\mathbf{x},\mathbf{L_{x}};\varphi)=\prod_{t=1}^{m}p\left(y_{t},L_{y_{t}}|\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}};\varphi\right),$$
where x and $\mathbf{L_x}$ are first fed into the encoder of SNAT model. Â and $\hat{\mathbf{L}}_{\mathbf{x}}$ with length $m$ are syntactic and semantic structure-aware copying of word and label from encoder inputs, respectively. We show the details in the following sections.
# 3.1 Syntactic and Semantic Labeling
Following Transformer (Vaswani et al., 2017), we use a stack of 6 identical Transformer blocks as encoder. In addition to the word embedding and position embedding in the traditional Transformer, we add structure-aware label embedding. The input to the encoder block is the addition of the normalized word, labels (NER and POS) and position embedding, which is represented as $\mathbf{H}^0=[\mathbf{h}_1^0,\ldots,\mathbf{h}_n^0].$
The input representation $\mathbf{H}^0=[\mathbf{h}_1^0,\ldots,\mathbf{h}_n^0]$ is encoded into contextual layer representations through the Transformer blocks. For each layer, the layer representation $\mathbf{H}^l=[\mathbf{h}_1^l,\ldots,\mathbf{h}_n^l]$ is computed with the $l$-th layer Transformer block
ing an open-source pre-trained semantic annotator. $\mathbf{H}^l=$ Transformer$_l(\mathbf{H}^{l-1}),l\in\{1,2,\ldots,6\}.$
In each Transformer block, multiple self-attention heads are used to aggregate the output vectors of
We use POS and NER to introduce the syntactic and semantic information existing in natural language, respectively. During the data pre-processing, each sentence is annotated into a semantic sequence usIn particular, we take the Treebank style (Marcus et al., 1999) for POS and PropBank style (Palmer
|
[
[
0.615234375,
0.0736083984375,
0.6572265625,
0.0736083984375,
0.6572265625,
0.08074951171875,
0.615234375,
0.08074951171875
],
[
0.5810546875,
0.09027099609375,
0.66748046875,
0.09027099609375,
0.66748046875,
0.097412109375,
0.5810546875,
0.097412109375
],
[
0.509765625,
0.10394287109375,
0.57080078125,
0.10394287109375,
0.57080078125,
0.10809326171875,
0.509765625,
0.10809326171875
],
[
0.3046875,
0.1436767578125,
0.35107421875,
0.1436767578125,
0.35107421875,
0.155029296875,
0.3046875,
0.155029296875
],
[
0.30908203125,
0.1650390625,
0.339111328125,
0.1650390625,
0.339111328125,
0.169189453125,
0.30908203125,
0.169189453125
],
[
0.568359375,
0.1644287109375,
0.59423828125,
0.1644287109375,
0.59423828125,
0.16748046875,
0.568359375,
0.16748046875
],
[
0.5576171875,
0.188232421875,
0.59423828125,
0.188232421875,
0.59423828125,
0.1923828125,
0.5576171875,
0.1923828125
],
[
0.60107421875,
0.189453125,
0.61279296875,
0.189453125,
0.61279296875,
0.1923828125,
0.60107421875,
0.1923828125
],
[
0.6650390625,
0.189453125,
0.67822265625,
0.189453125,
0.67822265625,
0.1923828125,
0.6650390625,
0.1923828125
],
[
0.26611328125,
0.210205078125,
0.385498046875,
0.210205078125,
0.385498046875,
0.21728515625,
0.26611328125,
0.21728515625
],
[
0.37451171875,
0.191162109375,
0.38623046875,
0.191162109375,
0.38623046875,
0.19482421875,
0.37451171875,
0.19482421875
],
[
0.175537109375,
0.22265625,
0.19140625,
0.22265625,
0.19140625,
0.23095703125,
0.175537109375,
0.23095703125
],
[
0.1956787109375,
0.236328125,
0.3720703125,
0.236328125,
0.3720703125,
0.2464599609375,
0.1956787109375,
0.2464599609375
],
[
0.3720703125,
0.2369384765625,
0.444091796875,
0.2369384765625,
0.444091796875,
0.24462890625,
0.3720703125,
0.24462890625
],
[
0.5029296875,
0.2392578125,
0.52880859375,
0.2392578125,
0.52880859375,
0.243408203125,
0.5029296875,
0.243408203125
],
[
0.587890625,
0.2392578125,
0.61279296875,
0.2392578125,
0.61279296875,
0.243408203125,
0.587890625,
0.243408203125
],
[
0.67529296875,
0.2392578125,
0.68408203125,
0.2392578125,
0.68408203125,
0.243408203125,
0.67529296875,
0.243408203125
],
[
0.6884765625,
0.240478515625,
0.6962890625,
0.240478515625,
0.6962890625,
0.243408203125,
0.6884765625,
0.243408203125
],
[
0.403076171875,
0.251708984375,
0.418212890625,
0.251708984375,
0.418212890625,
0.260009765625,
0.403076171875,
0.260009765625
],
[
0.62109375,
0.251708984375,
0.6396484375,
0.254638671875,
0.63623046875,
0.26123046875,
0.61865234375,
0.25830078125
],
[
0.64990234375,
0.2529296875,
0.66650390625,
0.2529296875,
0.66650390625,
0.26123046875,
0.64990234375,
0.26123046875
],
[
0.31982421875,
0.254150390625,
0.340087890625,
0.254150390625,
0.340087890625,
0.26123046875,
0.31982421875,
0.26123046875
],
[
0.21826171875,
0.254150390625,
0.268798828125,
0.254150390625,
0.268798828125,
0.26123046875,
0.21826171875,
0.26123046875
],
[
0.27197265625,
0.254638671875,
0.30322265625,
0.254638671875,
0.30322265625,
0.2587890625,
0.27197265625,
0.2587890625
],
[
0.344970703125,
0.254638671875,
0.36279296875,
0.254638671875,
0.36279296875,
0.2587890625,
0.344970703125,
0.2587890625
],
[
0.6826171875,
0.254638671875,
0.68994140625,
0.254638671875,
0.68994140625,
0.25830078125,
0.6826171875,
0.25830078125
],
[
0.20068359375,
0.26416015625,
0.2275390625,
0.2666015625,
0.225830078125,
0.27490234375,
0.198974609375,
0.2724609375
],
[
0.51904296875,
0.26416015625,
0.60205078125,
0.26416015625,
0.60205078125,
0.27490234375,
0.51904296875,
0.27490234375
],
[
0.444091796875,
0.276611328125,
0.5146484375,
0.276611328125,
0.5146484375,
0.283935546875,
0.444091796875,
0.283935546875
],
[
0.4248046875,
0.2861328125,
0.44921875,
0.2861328125,
0.44921875,
0.289306640625,
0.4248046875,
0.289306640625
],
[
0.474365234375,
0.2861328125,
0.5205078125,
0.2861328125,
0.5205078125,
0.289306640625,
0.474365234375,
0.289306640625
],
[
0.2418212890625,
0.305908203125,
0.75830078125,
0.305908203125,
0.75830078125,
0.319580078125,
0.2418212890625,
0.319580078125
],
[
0.118408203125,
0.3466796875,
0.490234375,
0.3466796875,
0.490234375,
0.361083984375,
0.118408203125,
0.361083984375
],
[
0.51123046875,
0.3466796875,
0.884765625,
0.3466796875,
0.884765625,
0.35986328125,
0.51123046875,
0.35986328125
],
[
0.118408203125,
0.363525390625,
0.490234375,
0.363525390625,
0.490234375,
0.377197265625,
0.118408203125,
0.377197265625
],
[
0.51123046875,
0.3623046875,
0.884765625,
0.3623046875,
0.884765625,
0.3759765625,
0.51123046875,
0.3759765625
],
[
0.118408203125,
0.37890625,
0.490234375,
0.37890625,
0.490234375,
0.392578125,
0.118408203125,
0.392578125
],
[
0.51318359375,
0.37890625,
0.88330078125,
0.37890625,
0.88330078125,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.118408203125,
0.3955078125,
0.490234375,
0.3955078125,
0.490234375,
0.4091796875,
0.118408203125,
0.4091796875
],
[
0.51318359375,
0.3955078125,
0.88330078125,
0.3955078125,
0.88330078125,
0.4091796875,
0.51318359375,
0.4091796875
],
[
0.118408203125,
0.410888671875,
0.490234375,
0.410400390625,
0.490234375,
0.424560546875,
0.118408203125,
0.42578125
],
[
0.51123046875,
0.410888671875,
0.8818359375,
0.410888671875,
0.8818359375,
0.424560546875,
0.51123046875,
0.424560546875
],
[
0.12176513671875,
0.4287109375,
0.48779296875,
0.4287109375,
0.48779296875,
0.439453125,
0.12176513671875,
0.439453125
],
[
0.51123046875,
0.427001953125,
0.8818359375,
0.427001953125,
0.8818359375,
0.440673828125,
0.51123046875,
0.440673828125
],
[
0.12005615234375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.51318359375,
0.443603515625,
0.88330078125,
0.443603515625,
0.88330078125,
0.457275390625,
0.51318359375,
0.457275390625
],
[
0.118408203125,
0.460205078125,
0.489501953125,
0.460205078125,
0.489501953125,
0.473876953125,
0.118408203125,
0.473876953125
],
[
0.51123046875,
0.458984375,
0.88330078125,
0.457763671875,
0.88330078125,
0.47265625,
0.51123046875,
0.473876953125
],
[
0.1175537109375,
0.4755859375,
0.27197265625,
0.4755859375,
0.27197265625,
0.4892578125,
0.1175537109375,
0.4892578125
],
[
0.51318359375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.1368408203125,
0.4921875,
0.489501953125,
0.4921875,
0.489501953125,
0.50732421875,
0.1368408203125,
0.50732421875
],
[
0.51318359375,
0.4912109375,
0.88330078125,
0.4912109375,
0.88330078125,
0.505859375,
0.51318359375,
0.505859375
],
[
0.118408203125,
0.5087890625,
0.490234375,
0.5087890625,
0.490234375,
0.5224609375,
0.118408203125,
0.5224609375
],
[
0.51123046875,
0.50732421875,
0.88330078125,
0.505859375,
0.88330078125,
0.52099609375,
0.51123046875,
0.521484375
],
[
0.12005615234375,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.53759765625,
0.12005615234375,
0.53759765625
],
[
0.51318359375,
0.52392578125,
0.88330078125,
0.52392578125,
0.88330078125,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.1175537109375,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5546875,
0.1175537109375,
0.5546875
],
[
0.5146484375,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5517578125,
0.5146484375,
0.5517578125
],
[
0.118408203125,
0.5556640625,
0.490234375,
0.55712890625,
0.490234375,
0.57177734375,
0.118408203125,
0.57080078125
],
[
0.51318359375,
0.5546875,
0.8818359375,
0.5546875,
0.8818359375,
0.568359375,
0.51318359375,
0.568359375
],
[
0.1158447265625,
0.572265625,
0.490234375,
0.57177734375,
0.490234375,
0.5859375,
0.1158447265625,
0.58740234375
],
[
0.51318359375,
0.57177734375,
0.88330078125,
0.57177734375,
0.88330078125,
0.5859375,
0.51318359375,
0.5859375
],
[
0.118408203125,
0.5888671875,
0.489501953125,
0.5888671875,
0.489501953125,
0.60400390625,
0.118408203125,
0.60400390625
],
[
0.51318359375,
0.58740234375,
0.884765625,
0.58740234375,
0.884765625,
0.6015625,
0.51318359375,
0.6015625
],
[
0.118408203125,
0.60546875,
0.2376708984375,
0.60546875,
0.2376708984375,
0.619140625,
0.118408203125,
0.619140625
],
[
0.51318359375,
0.60498046875,
0.755859375,
0.60498046875,
0.755859375,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.515625,
0.6328125,
0.62451171875,
0.6328125,
0.62451171875,
0.64453125,
0.515625,
0.64453125
],
[
0.118408203125,
0.642578125,
0.285400390625,
0.64111328125,
0.285400390625,
0.6591796875,
0.118408203125,
0.66015625
],
[
0.301513671875,
0.642578125,
0.474365234375,
0.642578125,
0.474365234375,
0.66015625,
0.301513671875,
0.66015625
],
[
0.51318359375,
0.65185546875,
0.88330078125,
0.65185546875,
0.88330078125,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.47265625,
0.6650390625,
0.490234375,
0.66845703125,
0.486083984375,
0.6787109375,
0.468505859375,
0.67578125
],
[
0.51318359375,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.1175537109375,
0.6796875,
0.489501953125,
0.6796875,
0.489501953125,
0.69482421875,
0.1175537109375,
0.69482421875
],
[
0.51318359375,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.118408203125,
0.69677734375,
0.489501953125,
0.69775390625,
0.489501953125,
0.7119140625,
0.118408203125,
0.71142578125
],
[
0.51318359375,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.1175537109375,
0.71337890625,
0.489501953125,
0.71337890625,
0.489501953125,
0.72705078125,
0.1175537109375,
0.72705078125
],
[
0.5146484375,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.118408203125,
0.72998046875,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.118408203125,
0.74462890625
],
[
0.515625,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7451171875,
0.515625,
0.7451171875
],
[
0.118408203125,
0.74658203125,
0.38623046875,
0.74658203125,
0.38623046875,
0.76025390625,
0.118408203125,
0.76025390625
],
[
0.51318359375,
0.7490234375,
0.8818359375,
0.7490234375,
0.8818359375,
0.7626953125,
0.51318359375,
0.7626953125
],
[
0.51318359375,
0.76416015625,
0.873046875,
0.76416015625,
0.873046875,
0.779296875,
0.51318359375,
0.779296875
],
[
0.1175537109375,
0.77392578125,
0.420654296875,
0.775390625,
0.420654296875,
0.79052734375,
0.1175537109375,
0.7880859375
],
[
0.53076171875,
0.77978515625,
0.880859375,
0.77978515625,
0.880859375,
0.79736328125,
0.53076171875,
0.79736328125
],
[
0.1175537109375,
0.79638671875,
0.490234375,
0.79736328125,
0.490234375,
0.8125,
0.1175537109375,
0.81103515625
],
[
0.51123046875,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51123046875,
0.8125
],
[
0.1175537109375,
0.8134765625,
0.490234375,
0.81396484375,
0.490234375,
0.8291015625,
0.1175537109375,
0.82763671875
],
[
0.51123046875,
0.8134765625,
0.884765625,
0.81396484375,
0.884765625,
0.8291015625,
0.51123046875,
0.82763671875
],
[
0.1175537109375,
0.830078125,
0.490234375,
0.83056640625,
0.490234375,
0.845703125,
0.1175537109375,
0.84423828125
],
[
0.509765625,
0.82763671875,
0.884765625,
0.8291015625,
0.884765625,
0.8466796875,
0.509765625,
0.845703125
],
[
0.118408203125,
0.84716796875,
0.49365234375,
0.84716796875,
0.49365234375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.509765625,
0.8466796875,
0.88330078125,
0.845703125,
0.88330078125,
0.85986328125,
0.509765625,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.49365234375,
0.86328125,
0.49365234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.509765625,
0.859375,
0.884765625,
0.86083984375,
0.884765625,
0.87890625,
0.509765625,
0.876953125
],
[
0.1175537109375,
0.87890625,
0.491943359375,
0.87744140625,
0.491943359375,
0.892578125,
0.1175537109375,
0.8935546875
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.8935546875,
0.490234375,
0.89453125,
0.490234375,
0.9091796875,
0.118408203125,
0.90771484375
],
[
0.509765625,
0.8935546875,
0.88671875,
0.892578125,
0.88671875,
0.91015625,
0.509765625,
0.9111328125
]
] |
[
"Decoder",
"argetSequence",
"stAnmmCe",
"Encoder",
"sontex",
"",
"",
"C",
"e",
"Multi-Head Self-attention",
"MU",
"Nx",
"EmbEmbEmbEmbEmbEmb",
"EmbEmbEmb",
"Emb",
"Emhs",
"Em",
"",
"1ob",
"mal",
"for",
"man",
"An@@ Cel",
"otth",
"1tn",
"ih",
"LOXBPE",
"BPERSON PERSONO",
"acticandSemant",
"ce",
"araann",
"Figure I: An overview of the proposed SNAT for neural machine translation",
"tokenizer (Sennrich et al., 2016). In parallel, words",
"et al., 2005)for NER to annotate every token of",
"in the input sequence are passed to POS and NER",
"input sequence with semantic labels. Given a spe-",
"annotators to extract explicit syntactic and semantic",
"cific sentence, there would be predicate-argument",
"structures, and the corresponding embeddings are",
"structures. Since the input sequence is segmented",
"aggregated by a linear layer to form the final syn-",
"into subword units using byte-pair encoding (Sen",
"taxandsemanticstructure-awareembedding.The",
"nrich et al., 2016), we assign the same label to all",
"SNAT model copies the structured encoder input",
"subwords tokenized from the same word. As shown",
"as the decoder input and generates the translated",
"in Figure 1, the word “Ancelotti\" is tokenized as",
"sentences and labels.",
"\"An@ @\" and \"Celotti\". The corresponding POS",
"One of the most important properties of SNAT",
"tags are PRON and PRON while the corresponding",
"is that it naturally introduces syntactic and seman-",
"NER tags are B PERSON and IPERSON. For the",
"tic information when taking the structure-aware",
"text \"Is An@@ Celotti the man for the job ?\", the",
"information as inputs and generating both words",
"correspondingPOStag setisAUX,PRON,PRON",
"and labels. More precisely, given a source sentence",
"DET, NOUN, ADP, DET, NOUN, PUNCT! and",
"x, source label sequence Lx, the conditional prob-",
"the NER tag set is [O, B PERSON, IPERSON,",
"ability of a target translation y and target label",
"O, O, O, O, O, O1. The data flow of the proposed",
"sequence Ly is:",
"model is also shown in Figure 1.",
"3.2Encoder",
"PsNAT(y, Ly|x, Lx; 9) =",
"Ip (yt, Ly |, Lx, x, Lx; ",
"Following Transformer (Vaswani et al..2017), we",
"(2)",
"use a stack of 6 identical Transformer blocks as",
"where x and Lx are first fed into the encoder of",
"encoder. In addition to the word embedding and po-",
"SNAT model. x and Lx with length m are syntactic",
"sition embedding in the traditional Transformer, we",
"and semantic structure-aware copying of word and",
"add structure-aware label embedding. The input to",
"label from encoder inputs, respectively. We show",
"theencoderblockistheadditionofthenormalized",
"the details in the following sections.",
"Word, labels (NER and POS) and position embed.",
"ding, which is represented as Ho - [hi,..., h]",
"3.1Syntactic and Semantic Labeling",
"The input representation Ho -- [h,..., h]",
"We use POS and NER to introduce the syntactic and",
"is encoded into contextual layer representations",
"semantic information existing in natural language,",
"through the Transformer blocks. For each layer,",
"respectively. During the data pre-processing, each",
"the layer representation H' = [hi,...,hh] is",
"sentence is annotated into a semantic sequence us-",
"computed with the l-th layer Transformer block",
"ing an open-source pre-trained semantic annotator.",
"H' = Transformeri(H\"-1), E {1,2,...,6}",
"In particular, we take the Treebank style (Marcus",
"In each Transformer block, multiple self-attention",
"et al., 1999) for POS and PropBank style (Palmer",
" heads are used to aggregate the output vectors of"
] |
{
"arxiv_id": "2101.08942",
"language": "en",
"page_idx": 2
}
|
||
2101.08942-en-3
|
2101.08942-en-3.jpg
|
the previous layer. A general attention mechanism can be formulated as the weighted sum of the value vector V using the query vector Q and the key vector K:
$$\text{Attention}(\mathbf{Q},\mathbf{K},\mathbf{V})=\text{softmax}\left(\frac{\mathbf{Q}\mathbf{K}^T}{\sqrt{d_{\mathrm{model}}}}\right)\cdot\mathbf{V},$$
where $d_\mathrm{model}$ represents the dimension of hidden representations. For self-attention, Q, K, and V are mappings of previous hidden representation by different linear functions, i.e., $\mathbf{Q}=\mathbf{H}^{l-1}\mathbf{W}_{Q}^{l}$, $\mathbf{K}=\mathbf{H}^{l-1}\mathbf{W}_{K}^{l}$, and $\mathbf{V}=\mathbf{H}^l-1\mathbf{W}_{V}^{l}$, respectively. At last, the encoder produces a final contextual representation $\mathbf{H}^6=[\mathbf{h}_1^6,\ldots,\mathbf{h}_n^6]$, which is obtained from the last Transformer block.
## 3.3 Decoder
The decoder also consists of 6 identical Transformer blocks, but with several key differences from the encoder. More concretely, we denote the contextual representations in the $i$-th decoder layer is $\mathbf{Z}^{i}(1\leq i\leq6).$ The input to the decoder block as $\mathbf{Z}^0=[\mathbf{z}_1^0,\ldots,\mathbf{z}_m^0]$,which is produced by the addition of the word, labels (NER and POS) copying from encoder input and positional embedding.
For the target side input $[\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}}]$, most previous works simply copied partial source sentence with the length ratio $\frac nm$ where $n$ refers to the source length and $m$ is the target length as the decoder input. More concretely, the decoder input $y_i$ at the $i$-th position is simply a copy of the $\lfloor\frac{n}{m}\times i\rfloor$th contextual representation, i.e., $x_{\lfloor\frac{n}{m}\times i\rfloor}$ from the encoder. From our investigation, in most cases, the gap between source length and target length is relatively small (e.g. 2). Therefore, it deletes or duplicates the copy of the last few tokens of the source. If the last token is meaningful, the deletion will neglect important information. Otherwise, if the last token is trivial, multiple copies will add noise to the model.
Instead, we propose a syntactic and semantic structure-aware mapping method considering the POS and NER labels to construct the decoder inputs. Our model first picks out the informative words with NOUN and VERB POS tags, and the ones recognized as entities by the NER module. If the source length is longer than the target length, we retain all informative words, and randomly delete the rest of the words. On the other hand, if the source length is shorter than the target, we retain all words and randomly duplicate the informative words. The corresponding label of a word
is also deleted or preserved. Moreover, by copying the similar structural words from the source, it can provide more information to the target input than just copying the source token, which is greatly different from the target token. The POS and NER labels of those structure-aware copied words from the source sentence are also copied as the decoder input. So by using the structure-aware mapping, we can assign $[\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}}]$ as decoder input.
For positional attention which aims to learn the local word orders within the sentence (Gu et al., 2018), we set positional embedding (Vaswani et al., 2017) as both $\mathbf{Q}$ and K, and hidden representations of the previous layer as V.
For inter-attention, Q refers to hidden representations of the previous layer, whereas K and V are contextual vectors $\mathbf{H}^{6}$ from the encoder. We modify the attention mask so that it does not mask out the future tokens, and every token is dependent on both its preceding and succeeding tokens in every layer. Therefore, the generation of each token can use bi-directional attention. The positionwise Feed-Forward Network (FFN) is applied after multi-head attention in both encoder and decoder. It consists of two fully-connected layers and a layer normalization (Ba et al., 2016). The FFN takes Z$^{6}$ as input and calculates the final representation $\mathbf{Z}^f$, which is used to predict the whole target sentence and label:
(4)
$$\begin{aligned}&\text{label:}\\&p\left(\mathbf{y}\mid\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}}\right)=f\left(\mathbf{Z}^{f}\mathbf{W}_{w}^{\top}+\mathbf{b}_{w}\right),\\&q\left(\mathbf{L_{y}}\mid\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}}\right)=f\left(\mathbf{Z}^{f}\mathbf{W}_{l}^{\top}+\mathbf{b}_{l}\right),\end{aligned}$$
(5)
where $f$ is a GeLU activation function (Hendrycks and Gimpel, 2016). $\mathbf{W}_w$ and $\mathbf{W}_l$ are the token embedding and structural label embedding in the input representation, respectively. We use different FFNs for POS and NER labels. To avoid redundancy, we ${\mathrm{just~use~}q\left(\mathbf{L_{y}}\mid\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}}\right)}$to represent the two predicted label likelihood in general.
# 3.4 Training
We use $(\mathbf{x},\mathbf{L}_{\mathbf{x}},\mathbf{y}^{*},\mathbf{L}_{\mathbf{y}}^{*})$ to denote a training instance. To introduce the label information, our proposed SNAT contains a discrete sequential latent variable $L_y_{1:m}$ with conditional posterior distribution $p(L_{y_{1:m}}|\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}};\varphi).$ It can be approximated using a proposal distribution $q(\mathbf{L}_{\mathbf{y}}\mid$ $\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}}).$ The approximation also provides a variational bound for the maximum log-likelihood:
|
[
[
0.12005615234375,
0.07659912109375,
0.490234375,
0.07659912109375,
0.490234375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.12005615234375,
0.0926513671875,
0.490234375,
0.0926513671875,
0.490234375,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51318359375,
0.0914306640625,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.51318359375,
0.105712890625
],
[
0.1175537109375,
0.10687255859375,
0.490234375,
0.10809326171875,
0.490234375,
0.122314453125,
0.1175537109375,
0.1217041015625
],
[
0.5146484375,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.12005615234375,
0.1258544921875,
0.193115234375,
0.1258544921875,
0.193115234375,
0.13720703125,
0.12005615234375,
0.13720703125
],
[
0.5146484375,
0.12469482421875,
0.884765625,
0.12469482421875,
0.884765625,
0.1383056640625,
0.5146484375,
0.1383056640625
],
[
0.51318359375,
0.1395263671875,
0.8818359375,
0.1395263671875,
0.8818359375,
0.152587890625,
0.51318359375,
0.152587890625
],
[
0.364501953125,
0.146728515625,
0.403076171875,
0.146728515625,
0.403076171875,
0.1591796875,
0.364501953125,
0.1591796875
],
[
0.1318359375,
0.15380859375,
0.340087890625,
0.15380859375,
0.340087890625,
0.168701171875,
0.1318359375,
0.168701171875
],
[
0.4375,
0.15673828125,
0.453369140625,
0.15673828125,
0.453369140625,
0.1644287109375,
0.4375,
0.1644287109375
],
[
0.455078125,
0.1561279296875,
0.491943359375,
0.1561279296875,
0.491943359375,
0.16748046875,
0.455078125,
0.16748046875
],
[
0.35693359375,
0.162109375,
0.420654296875,
0.1650390625,
0.4189453125,
0.179931640625,
0.355224609375,
0.177001953125
],
[
0.51318359375,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.51318359375,
0.1728515625,
0.884765625,
0.1728515625,
0.884765625,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.12005615234375,
0.1859130859375,
0.490234375,
0.1859130859375,
0.490234375,
0.20068359375,
0.12005615234375,
0.20068359375
],
[
0.5146484375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.5146484375,
0.203125
],
[
0.1175537109375,
0.203125,
0.48779296875,
0.20068359375,
0.48779296875,
0.215576171875,
0.1175537109375,
0.21728515625
],
[
0.51318359375,
0.204833984375,
0.779296875,
0.204833984375,
0.779296875,
0.2197265625,
0.51318359375,
0.2197265625
],
[
0.118408203125,
0.218505859375,
0.48779296875,
0.218505859375,
0.48779296875,
0.232177734375,
0.118408203125,
0.232177734375
],
[
0.1158447265625,
0.22802734375,
0.49365234375,
0.23095703125,
0.49365234375,
0.2529296875,
0.1158447265625,
0.2493896484375
],
[
0.5283203125,
0.22265625,
0.8818359375,
0.22265625,
0.8818359375,
0.236328125,
0.5283203125,
0.236328125
],
[
0.51318359375,
0.2381591796875,
0.88671875,
0.2381591796875,
0.88671875,
0.251708984375,
0.51318359375,
0.251708984375
],
[
0.1175537109375,
0.2493896484375,
0.495361328125,
0.25048828125,
0.495361328125,
0.268310546875,
0.1175537109375,
0.267333984375
],
[
0.51318359375,
0.254638671875,
0.88671875,
0.254638671875,
0.88671875,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.118408203125,
0.267333984375,
0.490234375,
0.268310546875,
0.490234375,
0.283203125,
0.118408203125,
0.281982421875
],
[
0.51123046875,
0.26953125,
0.884765625,
0.270751953125,
0.884765625,
0.28515625,
0.51123046875,
0.283935546875
],
[
0.12005615234375,
0.28515625,
0.490234375,
0.28515625,
0.490234375,
0.298583984375,
0.12005615234375,
0.298583984375
],
[
0.51318359375,
0.287353515625,
0.7138671875,
0.287353515625,
0.7138671875,
0.300537109375,
0.51318359375,
0.300537109375
],
[
0.118408203125,
0.300537109375,
0.359375,
0.300537109375,
0.359375,
0.314208984375,
0.118408203125,
0.314208984375
],
[
0.53076171875,
0.303955078125,
0.884765625,
0.303955078125,
0.884765625,
0.318359375,
0.53076171875,
0.318359375
],
[
0.51318359375,
0.320556640625,
0.88330078125,
0.320556640625,
0.88330078125,
0.334228515625,
0.51318359375,
0.334228515625
],
[
0.118408203125,
0.32666015625,
0.225830078125,
0.32666015625,
0.225830078125,
0.33837890625,
0.118408203125,
0.33837890625
],
[
0.51318359375,
0.336181640625,
0.88330078125,
0.336181640625,
0.88330078125,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.118408203125,
0.347412109375,
0.490234375,
0.347412109375,
0.490234375,
0.361083984375,
0.118408203125,
0.361083984375
],
[
0.51318359375,
0.352783203125,
0.88330078125,
0.352783203125,
0.88330078125,
0.366455078125,
0.51318359375,
0.366455078125
],
[
0.12005615234375,
0.364013671875,
0.489501953125,
0.364013671875,
0.489501953125,
0.377685546875,
0.12005615234375,
0.377685546875
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.118408203125,
0.380126953125,
0.490234375,
0.380126953125,
0.490234375,
0.394287109375,
0.118408203125,
0.394287109375
],
[
0.51318359375,
0.385498046875,
0.88330078125,
0.385498046875,
0.88330078125,
0.3984375,
0.51318359375,
0.3984375
],
[
0.12005615234375,
0.39794921875,
0.489501953125,
0.39794921875,
0.489501953125,
0.407958984375,
0.12005615234375,
0.407958984375
],
[
0.5146484375,
0.40087890625,
0.8818359375,
0.40087890625,
0.8818359375,
0.4150390625,
0.5146484375,
0.4150390625
],
[
0.118408203125,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.42578125,
0.118408203125,
0.42578125
],
[
0.5146484375,
0.418701171875,
0.8818359375,
0.418701171875,
0.8818359375,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.118408203125,
0.427490234375,
0.489501953125,
0.427490234375,
0.489501953125,
0.4423828125,
0.118408203125,
0.4423828125
],
[
0.51318359375,
0.432861328125,
0.88330078125,
0.432861328125,
0.88330078125,
0.446533203125,
0.51318359375,
0.446533203125
],
[
0.1175537109375,
0.440673828125,
0.491943359375,
0.4423828125,
0.491943359375,
0.460205078125,
0.1175537109375,
0.457763671875
],
[
0.51318359375,
0.449462890625,
0.884765625,
0.449462890625,
0.884765625,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.1175537109375,
0.458984375,
0.459228515625,
0.460205078125,
0.459228515625,
0.474365234375,
0.1175537109375,
0.473876953125
],
[
0.51318359375,
0.466064453125,
0.88330078125,
0.466064453125,
0.88330078125,
0.479736328125,
0.51318359375,
0.479736328125
],
[
0.1368408203125,
0.476806640625,
0.489501953125,
0.476806640625,
0.489501953125,
0.490478515625,
0.1368408203125,
0.490478515625
],
[
0.51123046875,
0.48095703125,
0.8818359375,
0.479736328125,
0.8818359375,
0.49462890625,
0.51123046875,
0.495361328125
],
[
0.118408203125,
0.4921875,
0.489501953125,
0.4921875,
0.489501953125,
0.505859375,
0.118408203125,
0.505859375
],
[
0.51318359375,
0.49755859375,
0.88330078125,
0.49755859375,
0.88330078125,
0.5126953125,
0.51318359375,
0.5126953125
],
[
0.12005615234375,
0.5087890625,
0.490234375,
0.5087890625,
0.490234375,
0.5224609375,
0.12005615234375,
0.5224609375
],
[
0.51123046875,
0.51416015625,
0.88330078125,
0.51416015625,
0.88330078125,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.1175537109375,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.1175537109375,
0.5380859375
],
[
0.5146484375,
0.53076171875,
0.587890625,
0.53076171875,
0.587890625,
0.5419921875,
0.5146484375,
0.5419921875
],
[
0.118408203125,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5546875,
0.118408203125,
0.5546875
],
[
0.5615234375,
0.5458984375,
0.84228515625,
0.5458984375,
0.84228515625,
0.5634765625,
0.5615234375,
0.5634765625
],
[
0.118408203125,
0.5556640625,
0.490234375,
0.5556640625,
0.490234375,
0.57373046875,
0.118408203125,
0.57373046875
],
[
0.86376953125,
0.55029296875,
0.880859375,
0.55029296875,
0.880859375,
0.56005859375,
0.86376953125,
0.56005859375
],
[
0.118408203125,
0.57373046875,
0.489501953125,
0.57373046875,
0.489501953125,
0.58837890625,
0.118408203125,
0.58837890625
],
[
0.12005615234375,
0.5888671875,
0.489501953125,
0.5888671875,
0.489501953125,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.5615234375,
0.5849609375,
0.8427734375,
0.5849609375,
0.8427734375,
0.6025390625,
0.5615234375,
0.6025390625
],
[
0.85205078125,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.6015625,
0.85205078125,
0.6015625
],
[
0.12005615234375,
0.60546875,
0.489501953125,
0.60546875,
0.489501953125,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.118408203125,
0.62158203125,
0.490234375,
0.62158203125,
0.490234375,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.51318359375,
0.62255859375,
0.88330078125,
0.62255859375,
0.88330078125,
0.6357421875,
0.51318359375,
0.6357421875
],
[
0.118408203125,
0.6357421875,
0.490234375,
0.6357421875,
0.490234375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.51318359375,
0.6396484375,
0.884765625,
0.6396484375,
0.884765625,
0.65283203125,
0.51318359375,
0.65283203125
],
[
0.118408203125,
0.65283203125,
0.489501953125,
0.65283203125,
0.489501953125,
0.66748046875,
0.118408203125,
0.66748046875
],
[
0.51123046875,
0.65380859375,
0.88671875,
0.65478515625,
0.88671875,
0.66943359375,
0.51123046875,
0.66845703125
],
[
0.118408203125,
0.66943359375,
0.490234375,
0.66845703125,
0.490234375,
0.68310546875,
0.118408203125,
0.68408203125
],
[
0.51318359375,
0.67138671875,
0.88671875,
0.67138671875,
0.88671875,
0.68505859375,
0.51318359375,
0.68505859375
],
[
0.1175537109375,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.1175537109375,
0.69873046875
],
[
0.51123046875,
0.68603515625,
0.884765625,
0.68701171875,
0.884765625,
0.70166015625,
0.51123046875,
0.70068359375
],
[
0.12005615234375,
0.703125,
0.2166748046875,
0.703125,
0.2166748046875,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.51318359375,
0.7060546875,
0.88330078125,
0.7060546875,
0.88330078125,
0.72021484375,
0.51318359375,
0.72021484375
],
[
0.1368408203125,
0.71728515625,
0.489501953125,
0.71630859375,
0.489501953125,
0.73095703125,
0.1368408203125,
0.732421875
],
[
0.5146484375,
0.724609375,
0.78662109375,
0.724609375,
0.78662109375,
0.7392578125,
0.5146484375,
0.7392578125
],
[
0.118408203125,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.7490234375,
0.118408203125,
0.7490234375
],
[
0.12005615234375,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76318359375,
0.12005615234375,
0.76318359375
],
[
0.51318359375,
0.755859375,
0.62646484375,
0.755859375,
0.62646484375,
0.77099609375,
0.51318359375,
0.77099609375
],
[
0.118408203125,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.118408203125,
0.77978515625
],
[
0.12005615234375,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.51123046875,
0.77783203125,
0.88671875,
0.77978515625,
0.88671875,
0.79443359375,
0.51123046875,
0.7919921875
],
[
0.118408203125,
0.79736328125,
0.491943359375,
0.79736328125,
0.491943359375,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.51318359375,
0.7958984375,
0.88671875,
0.7958984375,
0.88671875,
0.80859375,
0.51318359375,
0.80859375
],
[
0.1175537109375,
0.8134765625,
0.49365234375,
0.81396484375,
0.49365234375,
0.8291015625,
0.1175537109375,
0.82763671875
],
[
0.509765625,
0.81005859375,
0.88818359375,
0.80859375,
0.88818359375,
0.82666015625,
0.509765625,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.8291015625,
0.88671875,
0.8291015625,
0.88671875,
0.84326171875,
0.51318359375,
0.84326171875
],
[
0.12176513671875,
0.8486328125,
0.491943359375,
0.8486328125,
0.491943359375,
0.859375,
0.12176513671875,
0.859375
],
[
0.51123046875,
0.84423828125,
0.884765625,
0.84423828125,
0.884765625,
0.8623046875,
0.51123046875,
0.8623046875
],
[
0.118408203125,
0.86328125,
0.491943359375,
0.86328125,
0.491943359375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.509765625,
0.85986328125,
0.88330078125,
0.85986328125,
0.88330078125,
0.87744140625,
0.509765625,
0.87744140625
],
[
0.118408203125,
0.87890625,
0.49365234375,
0.87890625,
0.49365234375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51123046875,
0.89453125,
0.88671875,
0.8935546875,
0.88671875,
0.90771484375,
0.51123046875,
0.9091796875
]
] |
[
"the previous layer. A general attention mechanism",
"is also deleted or preserved. Moreover, by copying",
"can be formulated as the weighted sum of the value",
"the similar structural words from the source, it can",
"vector V using the query vector Q and the key",
"provide more information to the target input than",
"vector K:",
"just copying the source token, which is greatly dif-",
"ferent from the target token. The POS and NER",
"QKT",
"Attention(Q, K, V) = softmax",
"V",
" (3)",
"dmodel,",
"labels of those structure-aware copied words from",
"the source sentence are also copied as the decoder",
"where dmodel represents the dimension of hidden",
"input. So by using the structure-aware mapping, we",
"representations. For self-attention, Q, K, and V",
"can assign [x, Lx] as decoder input.",
"are mappings of previous hidden representation",
"by different linear functions, i.e., Q - Hl-1 W.",
"For positional attention which aims to learn the",
"local word orders within the sentence (Gu et al.,",
"K = H-1wk, and V= Hl-1wy, respectively.",
"2018), we set positional embedding (Vaswani et al.,",
"At last, the encoder produces a final contextual rep-",
"2017) as both Q and K, and hidden representations",
"resentation H6 - [h,..., he], which is obtained",
"of the previous layer as V.",
"from the last Transformer block.",
"For inter-attention, Q refers to hidden represen-",
"tations of the previous layer, whereas K and V",
"3.3Decoder",
"are contextual vectors H6 from the encoder. We",
"The decoder also consists of 6 identical Trans-",
"modify the attention mask so that it does not mask",
"former blocks, but with several key differences",
" out the future tokens, and every token is depen.",
"from the encoder. More concretely, we denote the",
"dent on both its preceding and succeeding tokens",
"contextualrepresentations inthei-thdecoderlayer",
"in every layer. Therefore, the generation of each",
"is Z'(1 ≤ i ≤ 6). The input to the decoder block",
"token can usebi-directional attention.Theposition",
"as z - [zj, ... , zm], which is produced by the ad-",
"wise Feed-Forward Network (FFN) is applied after",
"dition of the word, labels (NER and POS) copying",
"multi-head attention in both encoder and decoder.",
"from encoder input and positional embedding.",
"It consists of two fully-connected layers and a layer",
"For the target side input [x, Lx], most previous",
"normalization (Ba et al., 2016). The FFN takes z6",
"works simply copied partial source sentence with",
"as input and calculates the final representation Zf",
"the length ratio n where n refers to the source",
"which is used to predict the whole target sentence",
"length and m is the target length as the decoder",
"and label:",
"input. More concretely, the decoder input yi at the",
"p(y [ x,Lx,x,Lx) =f(z/wT +bu),",
"i-th position is simply a copy of the L × ijth",
"(4)",
"contextual representation, i.e., & n xij from the en-",
"coder. From our investigation, in most cases, the",
"q(Ly [x, Lx,x,Lx) = f(zwT +bl),",
"(5)",
"gap between source length and target length is rela-",
"tively small (e.g. 2). Therefore, it deletes or dupli-",
"where f is a GeLU activation function (Hendrycks",
"cates the copy of the last few tokens of the source.",
"and Gimpel, 2016). Ww and W, are the token em-",
"If the last token is meaningful, the deletion will",
"bedding and structural label embedding in the input",
"neglect important information.Otherwise, if the",
"representation, respectively. We use different FFNs",
"last token is trivial, multiple copies will add noise",
"for POS and NER labels. To avoid redundancy, we",
"to the model.",
"just use q (Ly I x, Lx, x, Lx) to represent the two",
"Instead, we propose a syntactic and semantic",
"predicted label likelihood in general.",
"structure-aware mapping method considering the",
"POS and NER labels to construct the decoder in-",
"3.4Training",
"puts. Our model first picks out the informative",
"words with NOUN and VERB POS tags, and the",
"We use (x, Lx, y*, L*) to denote a training in-",
"ones recognized as entities by the NER module. If",
"stance. To introduce the label information, our",
"the source length is longer than the target length,",
" proposed SNAT contains a discrete sequential la-",
"we retain all informative words, and randomly",
"tent variable Ly1:m with conditional posterior dis-",
"delete the rest of the words. On the other hand,",
"tribution p(Lyl:m|x,Lx, x, Lx; ). It can be ap-",
"if the source length is shorter than the target, we",
" proximated using a proposal distribution q(Ly I",
"retain all words and randomly duplicate the infor-",
"x, Lx, x, Lx). The approximation also provides a",
"mative words. The corresponding label of a word",
"variational bound for the maximum log-likelihood:"
] |
{
"arxiv_id": "2101.08942",
"language": "en",
"page_idx": 3
}
|
||
2101.08942-en-4
|
2101.08942-en-4.jpg
|
(6)
$$\begin{aligned}&\log P_{\mathbf{SNAT}}=\log\sum_{t=1}^{m}q\left(L_{y_{t}}|\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}};\varphi\right)\\&\times p\left(y_{t}|L_{y_{t}},\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}};\varphi\right)\\&\geq\sum_{L_{y_{1}:m}\sim q}\left\{\underbrace{\sum_{t=1}^{m}\log q\left(L_{y_{t}}\mid\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}};\varphi\right)}_{\text{Label likelibood}}\\&+\underbrace{\sum_{t=1}^{m}\log p\left(y_{t}\mid L_{y_{t}},\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}};\varphi\right)}_{\text{Structure-aware word likelihood}}\right\}+\mathcal{H}(q).\end{aligned}$$
Note that, the resulting likelihood function, consisting of the two bracketed terms in Eq. (6), allows us to train the entire model in a supervised fashion. The inferred label can be utilized to train the label predicting model $q$ and simultaneously supervise the structure-aware word model $p.$ The label loss can be calculated by the cross-entropy $\mathcal{H}$ of $L_{y_t}^*$ and Eq. $(5)\colon$
(7)
$$\mathcal{L}_{\mathrm{label}}=\sum_{t=1}^{m}\mathcal{H}\left(L_{y_{t}}^{*},\:q(L_{y_{t}}\mid\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}},\mathbf{x},\mathbf{L}_{\mathbf{x}})\right),$$
The structure-aware word likelihood is conditioned on the generation result of the label. Since the Eq.(4) does not depend on the predicted label, we propose to bring the structure-aware word mask $\mathbf{M}_{wl}\in\mathbb{R}^{|V_{word}|\times|V_{label}|}$, where $|V_word|$ and $|V_{label}|$ are vocabulary sizes of word and label, respectively. The mask $\mathbf{M}_{w_{l}}$ is defined as follows:
$$\mathbf{M}_{w_l}(i,j)=\left\{\begin{array}{cc}1,&\mathcal{A}(y_i)=label_j,\\\epsilon,&\mathcal{A}(y_i)\neq label_j,\end{array}\right.$$
(8)
which can be obtained at the preprocessing stage, and $\mathcal{A}$ denotes the open-source pre-trained POS or NER annotator mentioned above. It aims to penalize the case when the word $y_i$ does not belong to the label $label_j$ with $\epsilon$, which is a small number defined within the range of (0,1) and will be tuned in our experiments. For example, the word “great” does not belong to VERB. The structure-aware word likelihood can be reformulated as:
$$\begin{aligned}&p\left(y_{t}\mid L_{y_{t}},\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}};\varphi\right)=\\&p(y_{t}\mid\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}})\times\mathbf{M_{wl}}\times q(L_{y_{t}}\mid\hat{\mathbf{x}},\hat{\mathbf{L_{x}}},\mathbf{x},\mathbf{L_{x}}).\end{aligned}$$
(9)
Consequently, the structure-aware word loss $\mathcal{L} _\mathrm{word}$ is defined as the crossentropy between true $p^{\prime}(y_{t}^{*}|L_{y_{t}}^{*})$ and pre-
where
${\mathrm{dicted}}$ $p\left ( y_{t}\mid L_{y_{t}}, \hat{\mathbf{x} } , \hat{\mathbf{L} } _{\mathbf{x} }, \mathbf{x} , \mathbf{L} _{\mathbf{x} }; \varphi \right ) $,
$p^{\prime}(y_t^*|L_{y_t}^*)\in\mathbb{R}^{|V_{word}|\times|V_{label}|\text{ is a matrix where}\:\text{The dataset is processed with Moses script(Hoang)}}$
only item at the index of $(y_{t}^{*},L_{y_{t}}^{*})$ equals to 1, otherwise equals to 0. We reshape $p^\prime(y_t^*|L_{y_t}^*)$ and $p(y_t|L_{y_t})$ to vectors when calculating the loss.
Intermediate Alignment Regularization One main problem of NAT is that the decoder generation process does not depend on the previously generated tokens. Based on the bidirectional nature of SNAT decoder, the token can depend on every token of the decoder input. However, since the input of decoder $[\hat{\mathbf{x}},\hat{\mathbf{L}}_{\mathbf{x}}]$ is the duplicate of encoder input $[\mathbf{x},\mathbf{L_{x}}]$, the generation depends on the encoder tokens rather than the target y*.
To solve this problem, we align the output of the intermediate layer of the decoder with the target. The alignment makes the generation of following layers dependent on the coarse target-side information instead of the mere encoder input. This alignment idea is inspired by (Guo et al., 2019), which directly feeds target-side tokens as inputs of the decoder by linearly mapping the source token embeddings to target embeddings. However, using one FFN layer to map different languages to the same space can hardly provide promising results. Thus, instead of aligning the input of decoder with the target, we use the intermediate layer of decoder to align with the target. In this case, our model avoids adding additional training parameters and manages to train the alignment together with SNAT model in an end-to-end fashion. Formally, we define the intermediate alignment regularization as cross-entropy loss between the predicted word and the true word:
(10)
$$\mathcal{L}_{\mathrm{reg}}=\sum\limits_{t=1}^{m}\mathcal{H}\left(y_{t}^{*},\:\mathrm{FFN}(\mathbf{Z}_{t}^{md})\right),$$
where $\mathbf{Z} ^md$ $( 1< md< 6)$ represents the output of each intermediate layer. Consequently, the final loss of SNAT can be represented with the coeffcient $\lambda$ as:
(11)
$$\mathcal{L}_{\mathbf{SNAT}}=\mathcal{L}_{\mathrm{word}}+\mathcal{L}_{\mathrm{label}}+\lambda\mathcal{L}_{\mathrm{reg}}.$$
## 4 Experiment
## 4.1 Experimental Setup
Data We evaluate SNAT performance on both the WMT14 En-De (around 4.5M sentence pairs) and the WMT16 En-Ro (around 610k sentence pairs) parallel corpora. For the parallel data, we use the processed data from (Ghazvininejad et al., 2019) to be consistent with previous publications. and Koehn, 2008), and the words are segmented into subword units using byte-pair encoding (Sennrich et al., 2016). The WMT14 En-De task
|
[
[
0.5146484375,
0.07598876953125,
0.8447265625,
0.07598876953125,
0.8447265625,
0.08905029296875,
0.5146484375,
0.08905029296875
],
[
0.84033203125,
0.07781982421875,
0.88330078125,
0.07781982421875,
0.88330078125,
0.08905029296875,
0.84033203125,
0.08905029296875
],
[
0.1351318359375,
0.097412109375,
0.418212890625,
0.097412109375,
0.418212890625,
0.1181640625,
0.1351318359375,
0.1181640625
],
[
0.51318359375,
0.0926513671875,
0.884765625,
0.0914306640625,
0.884765625,
0.105712890625,
0.51318359375,
0.10687255859375
],
[
0.5146484375,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.139404296875,
0.127685546875,
0.328369140625,
0.127685546875,
0.328369140625,
0.1484375,
0.139404296875,
0.1484375
],
[
0.51318359375,
0.12469482421875,
0.884765625,
0.12353515625,
0.884765625,
0.1383056640625,
0.51318359375,
0.1395263671875
],
[
0.5146484375,
0.14013671875,
0.884765625,
0.14013671875,
0.884765625,
0.15380859375,
0.5146484375,
0.15380859375
],
[
0.2283935546875,
0.157958984375,
0.42822265625,
0.157958984375,
0.42822265625,
0.1787109375,
0.2283935546875,
0.1787109375
],
[
0.51318359375,
0.1561279296875,
0.88330078125,
0.1561279296875,
0.88330078125,
0.169189453125,
0.51318359375,
0.169189453125
],
[
0.173828125,
0.1650390625,
0.19482421875,
0.1650390625,
0.19482421875,
0.1728515625,
0.173828125,
0.1728515625
],
[
0.47021484375,
0.168701171875,
0.489501953125,
0.168701171875,
0.489501953125,
0.1776123046875,
0.47021484375,
0.1776123046875
],
[
0.1435546875,
0.1602783203125,
0.181396484375,
0.166259765625,
0.175537109375,
0.1817626953125,
0.1376953125,
0.17578125
],
[
0.17041015625,
0.17578125,
0.183837890625,
0.17578125,
0.183837890625,
0.179931640625,
0.17041015625,
0.179931640625
],
[
0.51318359375,
0.1728515625,
0.88330078125,
0.1728515625,
0.88330078125,
0.18701171875,
0.51318359375,
0.18701171875
],
[
0.28955078125,
0.190673828125,
0.359375,
0.190673828125,
0.359375,
0.201904296875,
0.28955078125,
0.201904296875
],
[
0.5146484375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.5146484375,
0.203125
],
[
0.2091064453125,
0.208984375,
0.455078125,
0.2078857421875,
0.455078125,
0.2286376953125,
0.2091064453125,
0.2298583984375
],
[
0.51123046875,
0.2037353515625,
0.8134765625,
0.204833984375,
0.8134765625,
0.2197265625,
0.51123046875,
0.218505859375
],
[
0.19482421875,
0.2197265625,
0.204833984375,
0.2197265625,
0.204833984375,
0.22265625,
0.19482421875,
0.22265625
],
[
0.52880859375,
0.221435546875,
0.88330078125,
0.221435546875,
0.88330078125,
0.235107421875,
0.52880859375,
0.235107421875
],
[
0.204833984375,
0.2410888671875,
0.3359375,
0.2410888671875,
0.3359375,
0.251708984375,
0.204833984375,
0.251708984375
],
[
0.5146484375,
0.2369384765625,
0.88671875,
0.2369384765625,
0.88671875,
0.251708984375,
0.5146484375,
0.251708984375
],
[
0.118408203125,
0.257080078125,
0.489501953125,
0.257080078125,
0.489501953125,
0.270751953125,
0.118408203125,
0.270751953125
],
[
0.51318359375,
0.251708984375,
0.884765625,
0.254150390625,
0.884765625,
0.268310546875,
0.51318359375,
0.2666015625
],
[
0.118408203125,
0.273681640625,
0.491943359375,
0.273681640625,
0.491943359375,
0.287353515625,
0.118408203125,
0.287353515625
],
[
0.51318359375,
0.268310546875,
0.8818359375,
0.268310546875,
0.8818359375,
0.281982421875,
0.51318359375,
0.281982421875
],
[
0.118408203125,
0.289306640625,
0.491943359375,
0.289306640625,
0.491943359375,
0.302734375,
0.118408203125,
0.302734375
],
[
0.51318359375,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.1175537109375,
0.303955078125,
0.489501953125,
0.303955078125,
0.489501953125,
0.317626953125,
0.1175537109375,
0.317626953125
],
[
0.51318359375,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.314208984375,
0.51318359375,
0.314208984375
],
[
0.118408203125,
0.32177734375,
0.491943359375,
0.32177734375,
0.491943359375,
0.3349609375,
0.118408203125,
0.3349609375
],
[
0.51123046875,
0.31640625,
0.88330078125,
0.317626953125,
0.88330078125,
0.33203125,
0.51123046875,
0.330810546875
],
[
0.1175537109375,
0.33740234375,
0.490234375,
0.336181640625,
0.490234375,
0.350830078125,
0.1175537109375,
0.3515625
],
[
0.51318359375,
0.334228515625,
0.88330078125,
0.334228515625,
0.88330078125,
0.347412109375,
0.51318359375,
0.347412109375
],
[
0.1175537109375,
0.3515625,
0.484375,
0.35400390625,
0.484375,
0.3681640625,
0.1175537109375,
0.366455078125
],
[
0.51318359375,
0.350830078125,
0.884765625,
0.350830078125,
0.884765625,
0.364013671875,
0.51318359375,
0.364013671875
],
[
0.12005615234375,
0.369384765625,
0.2091064453125,
0.369384765625,
0.2091064453125,
0.38427734375,
0.12005615234375,
0.38427734375
],
[
0.5146484375,
0.366455078125,
0.88330078125,
0.366455078125,
0.88330078125,
0.380126953125,
0.5146484375,
0.380126953125
],
[
0.51318359375,
0.383056640625,
0.88330078125,
0.383056640625,
0.88330078125,
0.396728515625,
0.51318359375,
0.396728515625
],
[
0.2301025390625,
0.396728515625,
0.435791015625,
0.3955078125,
0.435791015625,
0.413330078125,
0.2301025390625,
0.41455078125
],
[
0.51318359375,
0.3984375,
0.88330078125,
0.3984375,
0.88330078125,
0.412109375,
0.51318359375,
0.412109375
],
[
0.47021484375,
0.402099609375,
0.486083984375,
0.402099609375,
0.486083984375,
0.410400390625,
0.47021484375,
0.410400390625
],
[
0.156982421875,
0.40087890625,
0.197265625,
0.40087890625,
0.197265625,
0.410888671875,
0.156982421875,
0.410888671875
],
[
0.51123046875,
0.41455078125,
0.884765625,
0.413330078125,
0.884765625,
0.427490234375,
0.51123046875,
0.4287109375
],
[
0.118408203125,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.441162109375,
0.118408203125,
0.441162109375
],
[
0.51318359375,
0.43115234375,
0.884765625,
0.43115234375,
0.884765625,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.12176513671875,
0.4453125,
0.489501953125,
0.4453125,
0.489501953125,
0.4560546875,
0.12176513671875,
0.4560546875
],
[
0.51123046875,
0.4453125,
0.88330078125,
0.446533203125,
0.88330078125,
0.46142578125,
0.51123046875,
0.460205078125
],
[
0.12005615234375,
0.460205078125,
0.490234375,
0.460205078125,
0.490234375,
0.473876953125,
0.12005615234375,
0.473876953125
],
[
0.51318359375,
0.463134765625,
0.88330078125,
0.463134765625,
0.88330078125,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.12005615234375,
0.4755859375,
0.490234375,
0.4755859375,
0.490234375,
0.490478515625,
0.12005615234375,
0.490478515625
],
[
0.51123046875,
0.476806640625,
0.88330078125,
0.478515625,
0.88330078125,
0.493408203125,
0.51123046875,
0.4912109375
],
[
0.1158447265625,
0.486328125,
0.490234375,
0.48681640625,
0.490234375,
0.5087890625,
0.1158447265625,
0.50830078125
],
[
0.51123046875,
0.493408203125,
0.88330078125,
0.49462890625,
0.88330078125,
0.5087890625,
0.51123046875,
0.50830078125
],
[
0.1175537109375,
0.50732421875,
0.490234375,
0.50830078125,
0.490234375,
0.5224609375,
0.1175537109375,
0.521484375
],
[
0.51123046875,
0.51123046875,
0.88330078125,
0.51025390625,
0.88330078125,
0.52490234375,
0.51123046875,
0.525390625
],
[
0.118408203125,
0.52392578125,
0.397216796875,
0.52392578125,
0.397216796875,
0.5380859375,
0.118408203125,
0.5380859375
],
[
0.5146484375,
0.529296875,
0.64990234375,
0.529296875,
0.64990234375,
0.5390625,
0.5146484375,
0.5390625
],
[
0.285400390625,
0.54443359375,
0.4130859375,
0.54443359375,
0.4130859375,
0.55859375,
0.285400390625,
0.55859375
],
[
0.183837890625,
0.5517578125,
0.267822265625,
0.5517578125,
0.267822265625,
0.56640625,
0.183837890625,
0.56640625
],
[
0.273681640625,
0.55810546875,
0.4130859375,
0.55810546875,
0.4130859375,
0.572265625,
0.273681640625,
0.572265625
],
[
0.466796875,
0.552734375,
0.490234375,
0.552734375,
0.490234375,
0.56640625,
0.466796875,
0.56640625
],
[
0.6455078125,
0.55859375,
0.8017578125,
0.5556640625,
0.8017578125,
0.57373046875,
0.6455078125,
0.57666015625
],
[
0.8505859375,
0.55859375,
0.884765625,
0.55859375,
0.884765625,
0.57470703125,
0.8505859375,
0.57470703125
],
[
0.1175537109375,
0.57666015625,
0.489501953125,
0.5791015625,
0.489501953125,
0.5927734375,
0.1175537109375,
0.59033203125
],
[
0.591796875,
0.56005859375,
0.63232421875,
0.56396484375,
0.63037109375,
0.576171875,
0.58837890625,
0.572265625
],
[
0.12005615234375,
0.59423828125,
0.490234375,
0.59423828125,
0.490234375,
0.60791015625,
0.12005615234375,
0.60791015625
],
[
0.509765625,
0.595703125,
0.884765625,
0.59716796875,
0.884765625,
0.6142578125,
0.509765625,
0.6123046875
],
[
0.118408203125,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.51318359375,
0.615234375,
0.88330078125,
0.615234375,
0.88330078125,
0.62890625,
0.51318359375,
0.62890625
],
[
0.1175537109375,
0.6259765625,
0.48779296875,
0.6259765625,
0.48779296875,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.51123046875,
0.6298828125,
0.88330078125,
0.63134765625,
0.88330078125,
0.6455078125,
0.51123046875,
0.64453125
],
[
0.12005615234375,
0.642578125,
0.489501953125,
0.642578125,
0.489501953125,
0.65625,
0.12005615234375,
0.65625
],
[
0.5146484375,
0.6484375,
0.591796875,
0.6484375,
0.591796875,
0.66015625,
0.5146484375,
0.66015625
],
[
0.12005615234375,
0.6591796875,
0.48779296875,
0.6591796875,
0.48779296875,
0.6728515625,
0.12005615234375,
0.6728515625
],
[
0.58837890625,
0.6728515625,
0.81005859375,
0.6728515625,
0.81005859375,
0.68701171875,
0.58837890625,
0.68701171875
],
[
0.8505859375,
0.6728515625,
0.88818359375,
0.6728515625,
0.88818359375,
0.6884765625,
0.8505859375,
0.6884765625
],
[
0.118408203125,
0.6904296875,
0.48779296875,
0.6904296875,
0.48779296875,
0.70361328125,
0.118408203125,
0.70361328125
],
[
0.118408203125,
0.6748046875,
0.489501953125,
0.6748046875,
0.489501953125,
0.6884765625,
0.118408203125,
0.6884765625
],
[
0.118408203125,
0.7060546875,
0.414794921875,
0.70703125,
0.414794921875,
0.7216796875,
0.118408203125,
0.72021484375
],
[
0.51123046875,
0.70166015625,
0.6513671875,
0.69970703125,
0.6513671875,
0.71728515625,
0.51123046875,
0.7197265625
],
[
0.1275634765625,
0.7236328125,
0.322509765625,
0.7255859375,
0.322509765625,
0.74462890625,
0.1275634765625,
0.7421875
],
[
0.5146484375,
0.72802734375,
0.71533203125,
0.72802734375,
0.71533203125,
0.7421875,
0.5146484375,
0.7421875
],
[
0.47021484375,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7490234375,
0.47021484375,
0.7490234375
],
[
0.1275634765625,
0.75,
0.459228515625,
0.75,
0.459228515625,
0.76416015625,
0.1275634765625,
0.76416015625
],
[
0.51318359375,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.12005615234375,
0.76953125,
0.2376708984375,
0.76953125,
0.2376708984375,
0.783203125,
0.12005615234375,
0.783203125
],
[
0.51123046875,
0.76416015625,
0.88330078125,
0.765625,
0.88330078125,
0.77978515625,
0.51123046875,
0.779296875
],
[
0.246826171875,
0.77099609375,
0.291259765625,
0.77099609375,
0.291259765625,
0.7822265625,
0.246826171875,
0.7822265625
],
[
0.287109375,
0.771484375,
0.42236328125,
0.771484375,
0.42236328125,
0.7822265625,
0.287109375,
0.7822265625
],
[
0.444091796875,
0.77099609375,
0.489501953125,
0.77099609375,
0.489501953125,
0.7822265625,
0.444091796875,
0.7822265625
],
[
0.240966796875,
0.77490234375,
0.2493896484375,
0.77490234375,
0.2493896484375,
0.779296875,
0.240966796875,
0.779296875
],
[
0.2359619140625,
0.77490234375,
0.240966796875,
0.77490234375,
0.240966796875,
0.77978515625,
0.2359619140625,
0.77978515625
],
[
0.118408203125,
0.7861328125,
0.156982421875,
0.7861328125,
0.156982421875,
0.79736328125,
0.118408203125,
0.79736328125
],
[
0.1695556640625,
0.7861328125,
0.337646484375,
0.78369140625,
0.337646484375,
0.798828125,
0.1695556640625,
0.80029296875
],
[
0.345947265625,
0.7861328125,
0.378662109375,
0.7861328125,
0.378662109375,
0.79736328125,
0.345947265625,
0.79736328125
],
[
0.382080078125,
0.7861328125,
0.424072265625,
0.7861328125,
0.424072265625,
0.79736328125,
0.382080078125,
0.79736328125
],
[
0.443359375,
0.78759765625,
0.490234375,
0.78759765625,
0.490234375,
0.79736328125,
0.443359375,
0.79736328125
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.17626953125,
0.80029296875,
0.4013671875,
0.8017578125,
0.4013671875,
0.81640625,
0.17626953125,
0.8154296875
],
[
0.405517578125,
0.802734375,
0.447509765625,
0.802734375,
0.447509765625,
0.81396484375,
0.405517578125,
0.81396484375
],
[
0.51318359375,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51318359375,
0.8125
],
[
0.447509765625,
0.80419921875,
0.490234375,
0.80419921875,
0.490234375,
0.8154296875,
0.447509765625,
0.8154296875
],
[
0.12005615234375,
0.80419921875,
0.181396484375,
0.80419921875,
0.181396484375,
0.8154296875,
0.12005615234375,
0.8154296875
],
[
0.5146484375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.185546875,
0.82177734375,
0.39794921875,
0.82177734375,
0.39794921875,
0.8388671875,
0.185546875,
0.8388671875
],
[
0.43994140625,
0.822265625,
0.490234375,
0.822265625,
0.490234375,
0.83740234375,
0.43994140625,
0.83740234375
],
[
0.5146484375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.12176513671875,
0.82373046875,
0.17041015625,
0.82373046875,
0.17041015625,
0.8349609375,
0.12176513671875,
0.8349609375
],
[
0.1158447265625,
0.8388671875,
0.495361328125,
0.83837890625,
0.495361328125,
0.85986328125,
0.1158447265625,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.1158447265625,
0.85791015625,
0.491943359375,
0.859375,
0.491943359375,
0.876953125,
0.1158447265625,
0.8759765625
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87890625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.5146484375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.462646484375,
0.8955078125,
0.462646484375,
0.91015625,
0.118408203125,
0.91015625
],
[
0.51318359375,
0.89453125,
0.884765625,
0.89453125,
0.884765625,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"Intermediate Alignment Regularization",
"One",
"log PsNAT = log q (Lur|x, Lx, x, Lx;)",
"main problem of NAT is that the decoder gener-",
"ation process does not depend on the previously",
"Xp(ylLu,x,Lx,x, Lx;9)",
"generated tokens. Based on the bidirectional na-",
"ture of SNAT decoder, the token can depend on",
"Elogq (Lut / x, Lx, x,Lx;0)",
"every token of the decoder input. However, since",
"E",
"(6)",
"ZL",
"",
"the input of decoder [x, Lx] is the duplicate of en-",
"Label likelihood",
"coder input [x, Lx], the generation depends on the",
"(yt [ Lye,x, Lx, x, Lx;)I+H(g).",
" encoder tokens rather than the target y*.",
"",
" To solve this problem, we align the output of the",
"Structure-aware word likelihood",
"intermediate layer of the decoder with the target.",
"Note that, the resulting likelihood function, consist-",
"The alignment makes the generation of following",
"ing of the two bracketed terms in Eq. (6), allows",
"layers dependent on the coarse target-side infor",
"us to train the entire model in a supervised fashion.",
"mation instead of the mere encoder input. This",
"The inferred label can be utilized to train the label",
"alignment idea is inspired by (Guo et al., 2019)",
"predicting model q and simultaneously supervise",
"which directly feeds target-side tokens as inputs",
"the structure-aware word model p. The label loss",
"of the decoder by linearly mapping the source to-",
"can be calculated by the cross-entropy H of L*",
"ken embeddings to target embeddings. However,",
"and Eq. (5):",
"using one FFN layer to map different languages to",
"the same space can hardly provide promising re-",
"CH(Le, q(Lut [ x,Lx, x,Lx)",
"sults. Thus, instead of aligning the input of decoder",
"(7)",
"Llabel",
"with the target, we use the intermediate layer of",
"The structure-aware word likelihood is conditioned",
"decoder to align with the target. In this case, our",
"on the generation result of the label.Since the",
"model avoids adding additional training parameters",
"Eq. (4) does not depend on the predicted label,",
"and manages to train the alignment together with",
"we propose to bring the structure-aware word mask",
"SNAT model in an end-to-end fashion. Formally,",
"Mut E R/Vworax/Vabel, where Vworal and [Viabell",
"we define the intermediate alignment regularization",
"are vocabulary sizes of word and label, respectively.",
"as cross-entropy loss between the predicted word",
"The mask Muw, is defined as follows:",
"and the true word:",
"1, A(ys) = labely,",
"Mu, (i,j) =",
" e, A(yi) + labely,",
"(8)",
"H(ut, FFN(zmd)",
"(10)",
"which can be obtained at the preprocessing stage,",
"Creg",
"and A denotes the open-source pre-trained POS or",
"where zmd (1 < md < 6) represents the output",
"NER annotator mentioned above. It aims to penal-",
"of each intermediate layer. Consequently, the final",
"ize the case when the word yi does not belong to",
"loss of SNAT can be represented with the coeff-",
"the label label, with e, which is a small number",
"cient as:",
"defined within the range of (O, 1) and will be tuned",
"CsNAT=Cword +Clabel +Creg?",
"(11)",
"does not belong to VERB. The structure-aware",
"in our experiments. For example, the word “great\"",
"word likelihood can be reformulated as:",
"4 Experiment",
"p(yt [ Lur,x,Lx,x, Lx;0)",
"4.1 Experimental Setup",
"(9)",
"p(yt [x, Lx,x,Lx) × Mwl × q(Lut [x, Lx,x,Lx).",
"DataWe evaluate SNAT performance on both",
"Consequently,",
"the WMT14 En-De (around 4.5M sentence pairs)",
"the",
"structure-aware",
"word",
"",
"",
"loss",
"Lwordisdefined",
"as",
"the",
"cross-",
"and the WMT16 En-Ro (around 610k sentence",
"between true p'(y*|L*)",
"and",
"pairs) parallel corpora. For the parallel data, we",
"pre-",
"entropy",
"use the processed data from (Ghazvininejad et al.",
"p (yt I Lut,x, Lx,x, Lx; 0",
"where",
"2019) to be consistent with previous publications.",
"dicted",
"p(yt/Lut) e RIVworalx/Vabel is a matrix where",
"The dataset is processed with Moses script (Hoang",
"only item at the index of (y*, Lu) equals to 1,",
"and Koehn, 2008), and the words are segmented",
"otherwise equals to 0. We reshape p (yt |L*t) and",
"into subword units using byte-pair encoding (Sen-",
"p(yt|Lyt) to vectors when calculating the loss.",
"nrich et al., 2016). The WMT14 En-De task"
] |
{
"arxiv_id": "2101.08942",
"language": "en",
"page_idx": 4
}
|
||
2312.05616-en-0
|
2312.05616-en-0.jpg
|
Iterative Token Evaluation and Refinement for Real-World Super-Resolution
$\textbf{Chaofeng Chen}^{1},\textbf{ Shangchen Zhou}^{1},\textbf{ Liang Liao}^{1},\textbf{ Haoning Wu}^{1}$,
$\textbf{Wenxiu Sun}^2,\textbf{ Qiong Yan}^2,\textbf{ Weisi Lin}^1$
1 S-Lab, Nanyang Technological University 2 SenseTime Research
https://github.com/chaofengc/ITER
$ଝୁ$
$ഇ്റ്$
<FigureHere>
$উড়্তি$
$\frac{5}{6}$
Figure 1: Example result with the proposed ITER. Left top: input LQ image; Right top: SR result with ITER. $t$ is the iterative step index of the reverse discrete diffusion process, and $t=T$ is the initial distortion removal result. The textures are gradually enriched with iterative refinement. To obtain satisfactory results, our ITER requires only a total iteration step of $T\leq8.$
arXiv:2312.
### Abstract
pose to include a token evaluation network in the discrete diffusion process. It learns to evaluate which tokens are good restorations and helps to improve the iterative refinement results. Moreover, the evaluation network can first check status of the distortion removal output and then adaptively select total refinement steps needed, thereby maintaining a good balance between distortion removal and texture generation. Extensive experimental results show that ITER is easy to train and performs well within just 8 iterative steps. Our codes will be available publicly.
Real-world image super-resolution (RWSR) is a longstanding problem as low-quality (LQ) images often have complex and unidentified degradations. Existing methods such as Generative Adversarial Networks (GANs) or continuous diffusion models present their own issues including GANs being difficult to train while continuous diffusion models requiring numerous inference steps. In this paper, we propose an Iterative Token Evaluation and Refinement (ITER) framework for RWSR, which utilizes a discrete diffusion model operating in the discrete token representation space, $i.e.$, indexes of features extracted from a VQGAN codebook pre-trained with high-quality (HQ) images. We show that ITER is easier to train than GANs and more efficient than continuous diffusion models. Specifically, we divide RWSR into two sub-tasks, $i.e.$, distortion removal and texture generation. Distortion removal involves simple HQ token prediction with LQ images, while texture generation uses a discrete diffusion model to iteratively refine the distortion removal output with a token refinement network. In particular, we pro-
## Introduction
Single-image super-resolution (SISR) aims to restore highquality (HQ) outputs from low-quality (LQ) inputs that have been degraded through processes such as downsampling, blurring, noise, and compression. Previous studies (Liang et al. 2021; Zamir et al. 2022; Chen et al. 2023) have achieved remarkable progress in enhancing LQ images degraded by a single predefined type of degradation,
|
[
[
0.1168212890625,
0.125,
0.8837890625,
0.125,
0.8837890625,
0.1383056640625,
0.1168212890625,
0.1383056640625
],
[
0.2181396484375,
0.160400390625,
0.775390625,
0.160400390625,
0.775390625,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.3349609375,
0.177978515625,
0.66259765625,
0.177978515625,
0.66259765625,
0.1956787109375,
0.3349609375,
0.1956787109375
],
[
0.27294921875,
0.199462890625,
0.56396484375,
0.2008056640625,
0.56396484375,
0.21533203125,
0.27294921875,
0.2139892578125
],
[
0.5703125,
0.2008056640625,
0.72607421875,
0.2008056640625,
0.72607421875,
0.2139892578125,
0.5703125,
0.2139892578125
],
[
0.39453125,
0.21533203125,
0.6044921875,
0.21533203125,
0.6044921875,
0.2291259765625,
0.39453125,
0.2291259765625
],
[
0.0228729248046875,
0.267578125,
0.05963134765625,
0.267578125,
0.05963134765625,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.0228729248046875,
0.472900390625,
0.05963134765625,
0.472900390625,
0.05963134765625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.1168212890625,
0.54052734375,
0.17724609375,
0.54052734375,
0.17724609375,
0.5517578125,
0.1168212890625,
0.5517578125
],
[
0.26318359375,
0.54150390625,
0.294189453125,
0.54150390625,
0.294189453125,
0.5498046875,
0.26318359375,
0.5498046875
],
[
0.378173828125,
0.54052734375,
0.435546875,
0.54052734375,
0.435546875,
0.55029296875,
0.378173828125,
0.55029296875
],
[
0.48046875,
0.54345703125,
0.5146484375,
0.54345703125,
0.5146484375,
0.54736328125,
0.48046875,
0.54736328125
],
[
0.53515625,
0.546875,
0.69921875,
0.54736328125,
0.69921875,
0.56103515625,
0.53515625,
0.56005859375
],
[
0.82421875,
0.54150390625,
0.85546875,
0.54150390625,
0.85546875,
0.5498046875,
0.82421875,
0.5498046875
],
[
0.08660888671875,
0.5712890625,
0.9111328125,
0.5712890625,
0.9111328125,
0.58544921875,
0.08660888671875,
0.58544921875
],
[
0.08660888671875,
0.58544921875,
0.9111328125,
0.58544921875,
0.9111328125,
0.5986328125,
0.08660888671875,
0.5986328125
],
[
0.0882568359375,
0.599609375,
0.86376953125,
0.599609375,
0.86376953125,
0.61279296875,
0.0882568359375,
0.61279296875
],
[
0.25244140625,
0.62255859375,
0.313720703125,
0.62255859375,
0.313720703125,
0.6337890625,
0.25244140625,
0.6337890625
],
[
0.53759765625,
0.62353515625,
0.8935546875,
0.62353515625,
0.8935546875,
0.6337890625,
0.53759765625,
0.6337890625
],
[
0.53515625,
0.63525390625,
0.89794921875,
0.63525390625,
0.89794921875,
0.6484375,
0.53515625,
0.6484375
],
[
0.10296630859375,
0.6494140625,
0.463134765625,
0.6494140625,
0.463134765625,
0.66357421875,
0.10296630859375,
0.66357421875
],
[
0.5361328125,
0.64794921875,
0.89453125,
0.64794921875,
0.89453125,
0.6611328125,
0.5361328125,
0.6611328125
],
[
0.101318359375,
0.6611328125,
0.463134765625,
0.662109375,
0.463134765625,
0.67529296875,
0.101318359375,
0.6748046875
],
[
0.5361328125,
0.66015625,
0.896484375,
0.66015625,
0.896484375,
0.673828125,
0.5361328125,
0.673828125
],
[
0.10296630859375,
0.6748046875,
0.463134765625,
0.6748046875,
0.463134765625,
0.68798828125,
0.10296630859375,
0.68798828125
],
[
0.53759765625,
0.6748046875,
0.8935546875,
0.6748046875,
0.8935546875,
0.68505859375,
0.53759765625,
0.68505859375
],
[
0.10540771484375,
0.689453125,
0.4599609375,
0.689453125,
0.4599609375,
0.69970703125,
0.10540771484375,
0.69970703125
],
[
0.5361328125,
0.68603515625,
0.89453125,
0.68603515625,
0.89453125,
0.69970703125,
0.5361328125,
0.69970703125
],
[
0.101318359375,
0.69970703125,
0.463134765625,
0.70068359375,
0.463134765625,
0.7138671875,
0.101318359375,
0.71337890625
],
[
0.53759765625,
0.69970703125,
0.8935546875,
0.69970703125,
0.8935546875,
0.71044921875,
0.53759765625,
0.71044921875
],
[
0.10455322265625,
0.7138671875,
0.4599609375,
0.7138671875,
0.4599609375,
0.724609375,
0.10455322265625,
0.724609375
],
[
0.5361328125,
0.7119140625,
0.89453125,
0.7119140625,
0.89453125,
0.72265625,
0.5361328125,
0.72265625
],
[
0.101318359375,
0.7236328125,
0.461669921875,
0.72607421875,
0.461669921875,
0.7392578125,
0.101318359375,
0.7373046875
],
[
0.5361328125,
0.7236328125,
0.89794921875,
0.7236328125,
0.89794921875,
0.7373046875,
0.5361328125,
0.7373046875
],
[
0.10455322265625,
0.7392578125,
0.461669921875,
0.7392578125,
0.461669921875,
0.75,
0.10455322265625,
0.75
],
[
0.5361328125,
0.7373046875,
0.6611328125,
0.7373046875,
0.6611328125,
0.74755859375,
0.5361328125,
0.74755859375
],
[
0.10455322265625,
0.751953125,
0.461669921875,
0.751953125,
0.461669921875,
0.76220703125,
0.10455322265625,
0.76220703125
],
[
0.101318359375,
0.76220703125,
0.461669921875,
0.76318359375,
0.461669921875,
0.77783203125,
0.101318359375,
0.7763671875
],
[
0.10296630859375,
0.7744140625,
0.461669921875,
0.7744140625,
0.461669921875,
0.78857421875,
0.10296630859375,
0.78857421875
],
[
0.6611328125,
0.76953125,
0.76953125,
0.771484375,
0.76953125,
0.787109375,
0.6611328125,
0.78515625
],
[
0.10296630859375,
0.78857421875,
0.463134765625,
0.78857421875,
0.463134765625,
0.8017578125,
0.10296630859375,
0.8017578125
],
[
0.521484375,
0.79248046875,
0.91015625,
0.79248046875,
0.91015625,
0.80615234375,
0.521484375,
0.80615234375
],
[
0.10455322265625,
0.8017578125,
0.4599609375,
0.8017578125,
0.4599609375,
0.8125,
0.10455322265625,
0.8125
],
[
0.521484375,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.10455322265625,
0.814453125,
0.461669921875,
0.814453125,
0.461669921875,
0.82470703125,
0.10455322265625,
0.82470703125
],
[
0.10296630859375,
0.82568359375,
0.461669921875,
0.82568359375,
0.461669921875,
0.8388671875,
0.10296630859375,
0.8388671875
],
[
0.521484375,
0.82080078125,
0.91015625,
0.82080078125,
0.91015625,
0.833984375,
0.521484375,
0.833984375
],
[
0.10296630859375,
0.83837890625,
0.463134765625,
0.83837890625,
0.463134765625,
0.8515625,
0.10296630859375,
0.8515625
],
[
0.521484375,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.521484375,
0.84765625
],
[
0.10296630859375,
0.85107421875,
0.461669921875,
0.85107421875,
0.461669921875,
0.8642578125,
0.10296630859375,
0.8642578125
],
[
0.521484375,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.8623046875,
0.521484375,
0.8623046875
],
[
0.10296630859375,
0.86279296875,
0.461669921875,
0.86279296875,
0.461669921875,
0.876953125,
0.10296630859375,
0.876953125
],
[
0.51953125,
0.8623046875,
0.91015625,
0.8623046875,
0.91015625,
0.87548828125,
0.51953125,
0.87548828125
],
[
0.101318359375,
0.87548828125,
0.461669921875,
0.876953125,
0.461669921875,
0.89013671875,
0.101318359375,
0.8896484375
],
[
0.51953125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"Iterative Token Evaluation and Refinement for Real-World Super-Resolution",
"Chaofeng Chen', Shangchen Zhou', Liang Liao', Haoning Wu',",
"Wenxiu Sun?, Qiong Yan?, Weisi Lin'",
"I S-Lab, Nanyang Technological University",
"2 SenseTime Research",
"https:/github.com/chaofengc/ITER",
"]",
"191502121",
"InputLQ",
"t=T",
"t=T-1",
"",
"Reverse discrete diffusion",
"t=0",
"Figure I: Example result with the proposed ITER. Left top: input LQ image; Right top: SR result with ITER. t is the iterative",
"step index of the reverse discrete diffusion process, and t = T is the initial distortion removal result. The textures are gradually",
"enriched with iterative refinement. To obtain satisfactory results, our ITER requires only a total iteration step of T ≤ 8.",
"Abstract",
"posetoincludeatokenevaluationnetworkinthediscretedif",
"fusion process. It learns to evaluate which tokens are good",
"Real-world image super-resolution (RWSR) is a long-",
"restorations and helps to improve the iterative refinement re-",
"standing problem as low-quality (LQ)images often have",
"sults. Moreover, the evaluation network can first check status",
"complex and unidentified degradations. Existing methods",
"ofthe distortion removal outputandthenadaptively selectto",
"such asGenerativeAdversarialNetworks(GANs)orcon-",
"tal refinement steps needed, thereby maintaining a good bal-",
"tinuous diffusion models present their own issues including",
"ance between distortion removal and texture generation.Ex-",
"GANsbeingdifficulttotrainwhilecontinuous diffusionmod",
"tensive experimental results show thatITERis easy to train",
"els requiring numerous inference steps.In this paper, we pro-",
"and performs well within just 8 iterative steps. Our codes will",
"pose an Iterative TokenEvaluation and Refinement (ITER)",
"be availablepublicly",
"frameworkforRWSR,whichutilizesa discretediffusion",
"model operating in the discrete token representation space,",
"i.e., indexes of features extracted from a VQGAN codebook",
"Introduction",
"pre-trained with high-quality (HQ) images. We show that",
"Single-image super-resolution (SISR) aims to restore high-",
"ITER is easier to train than GANs and more efficient than",
"quality (HQ) outputs from low-quality (LQ) inputs that",
"continuousdiffusionmodels.Specifically,wedivideRWSR",
"into two sub-tasks, i.e., distortion removal and texture genera-",
"have been degraded through processes such as downsam-",
"tion. Distortion removal involves simple HQ token prediction",
"pling, blurring, noise, and compression. Previous studies",
"with LQ images, while texture generation uses a discrete dif-",
"(Liang et al. 2021; Zamir et al. 2022; Chen et al. 2023)",
"fusion model to iteratively refine the distortion removal out-",
"have achieved remarkable progress in enhancing LQ im-",
"put with a token refinement network. In particular, we pro-",
"ages degraded by a single predefined type of degradation,"
] |
{
"arxiv_id": "2312.05616",
"language": "en",
"page_idx": 0
}
|
||
2404.01651-en-0
|
2404.01651-en-0.jpg
|
### $\textbf{NLP Systems That Can' t Tell Use from Mention Censor Counterspeech}$ but Teaching the Distinction Helps
Kristina Gligorić Myra Cheng Lucia Zheng Esin Durmus Dan Jurafsky
Stanford University
gligoric@cs.stanford.edu
<FigureHere>
Figure 1: IIlustration of harmful censorship of counterspeech. Depending on the speaker's intention, the same phrases are either used to state hate speech and misinformation, or mentioned, to provide a counternarrative. Algorithms that indiscriminately classify counterspeech as abusive censor constructive discourse.
## Abstract
Warning: content in this paper may be upsetting
or offensive.
The use of words to convey speaker's intent is traditionally distinguished from the ‘mention’ of words for quoting what someone said, or pointing out properties of a word. Here we show that computationally modeling this usemention distinction is crucial for dealing with counterspeech online. Counterspeech that refutes problematic content often mentions harmful language but is not harmful itself (e.g., calling a vaccine dangerous is not the same as expressing disapproval of someone for calling vaccines dangerous). We show that even recent language models fail at distinguishing use from mention, and that this failure propagates to two key downstream tasks: misinformation and hate speech detection, resulting in censorship of counterspeech. We introduce prompting mitigations that teach the use-mention distinction, and show they reduce these errors. Our work highlights the importance of the use-mention distinction for NLP and CSS and offers ways to address it.
### 1 $\textbf{Introduction}$
The use-mention distinction is the difference between using words (Bananas have a peel) and mentioning them (“Bananas” has 7 letters, or Dan said “Bananas"). The distinction has long been important in the philosophy of language (Sperber and Wilson, 1981; Saka, 1998), where discussions date back to Tarski (1931) and Quine (1940).
The ability to correctly make this distinction is particularly relevant to dealing with problematic language online. While detecting problematic text has become a standard task in NLP, there has been less consideration of text that might mention harmful content without directly using it. Mentioning problematic content is critical to language like counterspeech that challenges or opposes hateful or misleading narratives (Table 1)
(Wright et al., 2017; Mun et al., 2023; Hangartner et al., 2021; Ecker et al., 2022). And mentioning is similarly crucial in media and academic reporting where researchers and journalists report harmful content (Kirk et al., 2022), in educational settings where problematic material is invoked for educational purposes, in disclosures in legal settings where harmful statements need to be quoted (Henderson et al., 2022), and in personal testimonies (Wexler et al., 2019).
In this work, we focus on the first of these: online counterspeech, defined as speech produced by users of online platforms to counteract harmful speech of others. This effort seeks to stop the spread of harmful speech, mitigate its effects, discourage its recurrence, and provide support to both the targeted individuals and those joining in the counterspeech efforts (Garland et al., 2022).
Counterspeech statements often involve refer-
|
[
[
0.1275634765625,
0.0914306640625,
0.86962890625,
0.0926513671875,
0.86962890625,
0.10986328125,
0.1275634765625,
0.1092529296875
],
[
0.31982421875,
0.10986328125,
0.67822265625,
0.11102294921875,
0.67822265625,
0.12890625,
0.31982421875,
0.127685546875
],
[
0.154541015625,
0.15380859375,
0.3046875,
0.15380859375,
0.3046875,
0.168701171875,
0.154541015625,
0.168701171875
],
[
0.322509765625,
0.152587890625,
0.443359375,
0.155029296875,
0.443359375,
0.169189453125,
0.322509765625,
0.16748046875
],
[
0.43408203125,
0.155029296875,
0.599609375,
0.155029296875,
0.599609375,
0.168701171875,
0.43408203125,
0.168701171875
],
[
0.59033203125,
0.15380859375,
0.7109375,
0.15380859375,
0.7109375,
0.16748046875,
0.59033203125,
0.16748046875
],
[
0.72900390625,
0.155029296875,
0.8486328125,
0.155029296875,
0.8486328125,
0.168701171875,
0.72900390625,
0.168701171875
],
[
0.420654296875,
0.171630859375,
0.58447265625,
0.171630859375,
0.58447265625,
0.185302734375,
0.420654296875,
0.185302734375
],
[
0.38037109375,
0.188232421875,
0.623046875,
0.188232421875,
0.623046875,
0.201904296875,
0.38037109375,
0.201904296875
],
[
0.54150390625,
0.26123046875,
0.5859375,
0.26123046875,
0.5859375,
0.2724609375,
0.54150390625,
0.2724609375
],
[
0.705078125,
0.26123046875,
0.79833984375,
0.26123046875,
0.79833984375,
0.2724609375,
0.705078125,
0.2724609375
],
[
0.81787109375,
0.26123046875,
0.87744140625,
0.26123046875,
0.87744140625,
0.2724609375,
0.81787109375,
0.2724609375
],
[
0.615234375,
0.262451171875,
0.6708984375,
0.262451171875,
0.6708984375,
0.2724609375,
0.615234375,
0.2724609375
],
[
0.2626953125,
0.262451171875,
0.343505859375,
0.262451171875,
0.343505859375,
0.276611328125,
0.2626953125,
0.276611328125
],
[
0.8193359375,
0.273681640625,
0.880859375,
0.273681640625,
0.880859375,
0.283935546875,
0.8193359375,
0.283935546875
],
[
0.5146484375,
0.27978515625,
0.580078125,
0.27978515625,
0.580078125,
0.290283203125,
0.5146484375,
0.290283203125
],
[
0.6884765625,
0.28076171875,
0.73779296875,
0.28076171875,
0.73779296875,
0.292236328125,
0.6884765625,
0.292236328125
],
[
0.82861328125,
0.283203125,
0.875,
0.283203125,
0.875,
0.290283203125,
0.82861328125,
0.290283203125
],
[
0.62744140625,
0.290283203125,
0.6572265625,
0.290283203125,
0.6572265625,
0.298583984375,
0.62744140625,
0.298583984375
],
[
0.145263671875,
0.283935546875,
0.45751953125,
0.28515625,
0.45751953125,
0.2998046875,
0.145263671875,
0.298583984375
],
[
0.611328125,
0.297607421875,
0.67333984375,
0.297607421875,
0.67333984375,
0.308837890625,
0.611328125,
0.308837890625
],
[
0.85986328125,
0.292236328125,
0.880859375,
0.292236328125,
0.880859375,
0.305908203125,
0.85986328125,
0.305908203125
],
[
0.14697265625,
0.3017578125,
0.2301025390625,
0.3017578125,
0.2301025390625,
0.312255859375,
0.14697265625,
0.312255859375
],
[
0.7021484375,
0.300537109375,
0.80029296875,
0.300537109375,
0.80029296875,
0.311279296875,
0.7021484375,
0.311279296875
],
[
0.8330078125,
0.3017578125,
0.86962890625,
0.30712890625,
0.86572265625,
0.32373046875,
0.82861328125,
0.318359375
],
[
0.019317626953125,
0.31005859375,
0.06634521484375,
0.31005859375,
0.06634521484375,
0.431640625,
0.019317626953125,
0.431640625
],
[
0.6337890625,
0.30810546875,
0.654296875,
0.30810546875,
0.654296875,
0.32177734375,
0.6337890625,
0.32177734375
],
[
0.1435546875,
0.318359375,
0.45751953125,
0.318359375,
0.45751953125,
0.33203125,
0.1435546875,
0.33203125
],
[
0.54833984375,
0.319580078125,
0.59375,
0.319580078125,
0.59375,
0.330810546875,
0.54833984375,
0.330810546875
],
[
0.6884765625,
0.320556640625,
0.796875,
0.32177734375,
0.796875,
0.333251953125,
0.6884765625,
0.33203125
],
[
0.145263671875,
0.333251953125,
0.459228515625,
0.333251953125,
0.459228515625,
0.3466796875,
0.145263671875,
0.3466796875
],
[
0.6171875,
0.333251953125,
0.66748046875,
0.333251953125,
0.66748046875,
0.34326171875,
0.6171875,
0.34326171875
],
[
0.69873046875,
0.33203125,
0.8037109375,
0.333251953125,
0.8037109375,
0.344482421875,
0.69873046875,
0.34326171875
],
[
0.61279296875,
0.342529296875,
0.67236328125,
0.342529296875,
0.67236328125,
0.352783203125,
0.61279296875,
0.352783203125
],
[
0.7001953125,
0.344482421875,
0.80029296875,
0.344482421875,
0.80029296875,
0.3515625,
0.7001953125,
0.3515625
],
[
0.85791015625,
0.339111328125,
0.8818359375,
0.339111328125,
0.8818359375,
0.35400390625,
0.85791015625,
0.35400390625
],
[
0.145263671875,
0.3466796875,
0.45751953125,
0.3466796875,
0.45751953125,
0.35986328125,
0.145263671875,
0.35986328125
],
[
0.62744140625,
0.350830078125,
0.6591796875,
0.350830078125,
0.6591796875,
0.369384765625,
0.62744140625,
0.369384765625
],
[
0.72314453125,
0.3515625,
0.7958984375,
0.352783203125,
0.7958984375,
0.35986328125,
0.72314453125,
0.359375
],
[
0.8310546875,
0.3515625,
0.86572265625,
0.35400390625,
0.8623046875,
0.371826171875,
0.82861328125,
0.369384765625
],
[
0.1435546875,
0.3623046875,
0.459228515625,
0.361083984375,
0.459228515625,
0.374755859375,
0.1435546875,
0.3759765625
],
[
0.70361328125,
0.358154296875,
0.79248046875,
0.359375,
0.79248046875,
0.369384765625,
0.70361328125,
0.3681640625
],
[
0.145263671875,
0.3759765625,
0.459228515625,
0.3759765625,
0.459228515625,
0.3896484375,
0.145263671875,
0.3896484375
],
[
0.515625,
0.38427734375,
0.59375,
0.38427734375,
0.59375,
0.391357421875,
0.515625,
0.391357421875
],
[
0.6884765625,
0.385986328125,
0.736328125,
0.385986328125,
0.736328125,
0.396728515625,
0.6884765625,
0.396728515625
],
[
0.8154296875,
0.38134765625,
0.87890625,
0.38134765625,
0.87890625,
0.388427734375,
0.8154296875,
0.388427734375
],
[
0.14697265625,
0.391357421875,
0.45751953125,
0.391357421875,
0.45751953125,
0.402099609375,
0.14697265625,
0.402099609375
],
[
0.82373046875,
0.388427734375,
0.87158203125,
0.388427734375,
0.87158203125,
0.3984375,
0.82373046875,
0.3984375
],
[
0.62646484375,
0.396728515625,
0.6572265625,
0.396728515625,
0.6572265625,
0.407958984375,
0.62646484375,
0.407958984375
],
[
0.69873046875,
0.3955078125,
0.8017578125,
0.3955078125,
0.8017578125,
0.40625,
0.69873046875,
0.40625
],
[
0.70947265625,
0.40380859375,
0.7900390625,
0.402587890625,
0.7900390625,
0.413330078125,
0.70947265625,
0.41455078125
],
[
0.8564453125,
0.399658203125,
0.88330078125,
0.399658203125,
0.88330078125,
0.419189453125,
0.8564453125,
0.419189453125
],
[
0.14697265625,
0.40625,
0.459228515625,
0.40625,
0.459228515625,
0.416259765625,
0.14697265625,
0.416259765625
],
[
0.60693359375,
0.40625,
0.67822265625,
0.40625,
0.67822265625,
0.41748046875,
0.60693359375,
0.41748046875
],
[
0.6884765625,
0.412109375,
0.81201171875,
0.412109375,
0.81201171875,
0.4228515625,
0.6884765625,
0.4228515625
],
[
0.8388671875,
0.410888671875,
0.861328125,
0.4150390625,
0.8564453125,
0.429931640625,
0.83447265625,
0.42578125
],
[
0.145263671875,
0.418701171875,
0.459228515625,
0.418701171875,
0.459228515625,
0.431640625,
0.145263671875,
0.431640625
],
[
0.62890625,
0.4150390625,
0.65576171875,
0.4150390625,
0.65576171875,
0.43408203125,
0.62890625,
0.43408203125
],
[
0.54833984375,
0.427001953125,
0.59375,
0.427001953125,
0.59375,
0.438232421875,
0.54833984375,
0.438232421875
],
[
0.6943359375,
0.4228515625,
0.80419921875,
0.4228515625,
0.80419921875,
0.429931640625,
0.6943359375,
0.429931640625
],
[
0.1435546875,
0.43115234375,
0.459228515625,
0.431640625,
0.459228515625,
0.446533203125,
0.1435546875,
0.4453125
],
[
0.68994140625,
0.43115234375,
0.80029296875,
0.43115234375,
0.80029296875,
0.441162109375,
0.68994140625,
0.441162109375
],
[
0.0235137939453125,
0.438232421875,
0.0621337890625,
0.438232421875,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51025390625
],
[
0.04534912109375,
0.440673828125,
0.0545654296875,
0.440673828125,
0.0545654296875,
0.44482421875,
0.04534912109375,
0.44482421875
],
[
0.615234375,
0.440673828125,
0.6689453125,
0.440673828125,
0.6689453125,
0.451904296875,
0.615234375,
0.451904296875
],
[
0.69287109375,
0.440673828125,
0.8095703125,
0.4423828125,
0.8095703125,
0.451904296875,
0.69287109375,
0.449462890625
],
[
0.145263671875,
0.446533203125,
0.4609375,
0.446533203125,
0.4609375,
0.460205078125,
0.145263671875,
0.460205078125
],
[
0.60791015625,
0.449462890625,
0.6767578125,
0.449462890625,
0.6767578125,
0.460205078125,
0.60791015625,
0.460205078125
],
[
0.69189453125,
0.45068359375,
0.8095703125,
0.45068359375,
0.8095703125,
0.46142578125,
0.69189453125,
0.46142578125
],
[
0.8564453125,
0.446533203125,
0.88330078125,
0.446533203125,
0.88330078125,
0.468505859375,
0.8564453125,
0.468505859375
],
[
0.62744140625,
0.458984375,
0.6591796875,
0.458984375,
0.6591796875,
0.47802734375,
0.62744140625,
0.47802734375
],
[
0.7060546875,
0.460205078125,
0.796875,
0.460205078125,
0.796875,
0.46728515625,
0.7060546875,
0.46728515625
],
[
0.8369140625,
0.4619140625,
0.85986328125,
0.4619140625,
0.85986328125,
0.4755859375,
0.8369140625,
0.4755859375
],
[
0.7119140625,
0.46728515625,
0.79248046875,
0.46728515625,
0.79248046875,
0.47802734375,
0.7119140625,
0.47802734375
],
[
0.14697265625,
0.476806640625,
0.45751953125,
0.476806640625,
0.45751953125,
0.48681640625,
0.14697265625,
0.48681640625
],
[
0.145263671875,
0.4892578125,
0.459228515625,
0.4892578125,
0.459228515625,
0.5029296875,
0.145263671875,
0.5029296875
],
[
0.51123046875,
0.4912109375,
0.88330078125,
0.4912109375,
0.88330078125,
0.5048828125,
0.51123046875,
0.5048828125
],
[
0.145263671875,
0.50439453125,
0.459228515625,
0.50439453125,
0.459228515625,
0.51708984375,
0.145263671875,
0.51708984375
],
[
0.509765625,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.51953125,
0.509765625,
0.51953125
],
[
0.0251922607421875,
0.5185546875,
0.05792236328125,
0.5185546875,
0.05792236328125,
0.56005859375,
0.0251922607421875,
0.56005859375
],
[
0.14697265625,
0.51953125,
0.45751953125,
0.51953125,
0.45751953125,
0.52978515625,
0.14697265625,
0.52978515625
],
[
0.51123046875,
0.52099609375,
0.8818359375,
0.52099609375,
0.8818359375,
0.53369140625,
0.51123046875,
0.53369140625
],
[
0.145263671875,
0.5322265625,
0.459228515625,
0.5322265625,
0.459228515625,
0.5458984375,
0.145263671875,
0.5458984375
],
[
0.51318359375,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.54638671875,
0.51318359375,
0.54638671875
],
[
0.145263671875,
0.5458984375,
0.459228515625,
0.5458984375,
0.459228515625,
0.55859375,
0.145263671875,
0.55859375
],
[
0.019317626953125,
0.548828125,
0.060455322265625,
0.548828125,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.145263671875,
0.46142578125,
0.459228515625,
0.46142578125,
0.459228515625,
0.474365234375,
0.145263671875,
0.474365234375
],
[
0.509765625,
0.548828125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.509765625,
0.5625
],
[
0.145263671875,
0.56005859375,
0.459228515625,
0.56005859375,
0.459228515625,
0.57373046875,
0.145263671875,
0.57373046875
],
[
0.51123046875,
0.5634765625,
0.85400390625,
0.5634765625,
0.85400390625,
0.57666015625,
0.51123046875,
0.57666015625
],
[
0.14697265625,
0.576171875,
0.45751953125,
0.576171875,
0.45751953125,
0.5859375,
0.14697265625,
0.5859375
],
[
0.145263671875,
0.58837890625,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.145263671875,
0.6015625
],
[
0.145263671875,
0.6025390625,
0.459228515625,
0.6025390625,
0.459228515625,
0.6162109375,
0.145263671875,
0.6162109375
],
[
0.509765625,
0.60791015625,
0.8818359375,
0.609375,
0.8818359375,
0.62353515625,
0.509765625,
0.62255859375
],
[
0.145263671875,
0.6181640625,
0.231689453125,
0.6181640625,
0.231689453125,
0.6298828125,
0.145263671875,
0.6298828125
],
[
0.509765625,
0.62451171875,
0.88330078125,
0.6259765625,
0.88330078125,
0.64013671875,
0.509765625,
0.6396484375
],
[
0.1435546875,
0.6396484375,
0.26025390625,
0.64111328125,
0.26025390625,
0.65625,
0.1435546875,
0.65380859375
],
[
0.1158447265625,
0.64794921875,
0.1234130859375,
0.642578125,
0.1292724609375,
0.646484375,
0.12176513671875,
0.65185546875
],
[
0.51123046875,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65625,
0.51123046875,
0.65625
],
[
0.1175537109375,
0.66455078125,
0.489501953125,
0.66455078125,
0.489501953125,
0.677734375,
0.1175537109375,
0.677734375
],
[
0.51123046875,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.51123046875,
0.6728515625
],
[
0.509765625,
0.6748046875,
0.88330078125,
0.67333984375,
0.88330078125,
0.6884765625,
0.509765625,
0.689453125
],
[
0.1158447265625,
0.681640625,
0.489501953125,
0.681640625,
0.489501953125,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.51123046875,
0.69140625,
0.88330078125,
0.69140625,
0.88330078125,
0.705078125,
0.51123046875,
0.705078125
],
[
0.1158447265625,
0.69775390625,
0.489501953125,
0.69775390625,
0.489501953125,
0.71142578125,
0.1158447265625,
0.71142578125
],
[
0.5087890625,
0.7060546875,
0.880859375,
0.7060546875,
0.880859375,
0.7197265625,
0.5087890625,
0.7197265625
],
[
0.11083984375,
0.7119140625,
0.49365234375,
0.71337890625,
0.49365234375,
0.72802734375,
0.11083984375,
0.72705078125
],
[
0.51318359375,
0.7236328125,
0.8818359375,
0.7236328125,
0.8818359375,
0.73681640625,
0.51318359375,
0.73681640625
],
[
0.1158447265625,
0.72998046875,
0.489501953125,
0.728515625,
0.489501953125,
0.74365234375,
0.1158447265625,
0.74462890625
],
[
0.51318359375,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.118408203125,
0.74755859375,
0.486083984375,
0.74755859375,
0.486083984375,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.51123046875,
0.7548828125,
0.75830078125,
0.7548828125,
0.75830078125,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.1175537109375,
0.76123046875,
0.355224609375,
0.76123046875,
0.355224609375,
0.775390625,
0.1175537109375,
0.775390625
],
[
0.1368408203125,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.7919921875,
0.1368408203125,
0.7919921875
],
[
0.53076171875,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7861328125,
0.53076171875,
0.7861328125
],
[
0.1158447265625,
0.79443359375,
0.489501953125,
0.79443359375,
0.489501953125,
0.80810546875,
0.1158447265625,
0.80810546875
],
[
0.509765625,
0.79052734375,
0.8818359375,
0.7890625,
0.8818359375,
0.80419921875,
0.509765625,
0.8046875
],
[
0.1175537109375,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.1175537109375,
0.82373046875
],
[
0.509765625,
0.80712890625,
0.88330078125,
0.80712890625,
0.88330078125,
0.82080078125,
0.509765625,
0.82080078125
],
[
0.1175537109375,
0.82666015625,
0.48779296875,
0.82666015625,
0.48779296875,
0.84033203125,
0.1175537109375,
0.84033203125
],
[
0.51123046875,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.51123046875,
0.8359375
],
[
0.1158447265625,
0.8427734375,
0.490234375,
0.8427734375,
0.490234375,
0.85546875,
0.1158447265625,
0.85546875
],
[
0.51123046875,
0.8388671875,
0.88330078125,
0.8388671875,
0.88330078125,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.118408203125,
0.85986328125,
0.48779296875,
0.85986328125,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.509765625,
0.85498046875,
0.8818359375,
0.85498046875,
0.8818359375,
0.8681640625,
0.509765625,
0.8681640625
],
[
0.1158447265625,
0.87451171875,
0.491943359375,
0.87451171875,
0.491943359375,
0.88818359375,
0.1158447265625,
0.88818359375
],
[
0.509765625,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.88525390625,
0.509765625,
0.88525390625
],
[
0.1158447265625,
0.89111328125,
0.489501953125,
0.89013671875,
0.489501953125,
0.90478515625,
0.1158447265625,
0.90625
],
[
0.509765625,
0.88720703125,
0.83642578125,
0.8857421875,
0.83642578125,
0.90087890625,
0.509765625,
0.90185546875
],
[
0.11419677734375,
0.90673828125,
0.48779296875,
0.90625,
0.48779296875,
0.92041015625,
0.11419677734375,
0.92138671875
],
[
0.53076171875,
0.90673828125,
0.88330078125,
0.90673828125,
0.88330078125,
0.92041015625,
0.53076171875,
0.92041015625
]
] |
[
"NLP Systems That Can't Tell Use from Mention Censor Counterspeech,",
"but Teaching the Distinction Helps",
"Kristina Gligoric",
"Myra Cheng",
"LuciaZhengI",
" Esin Durmus",
"Dan Jurafsky",
"Stanford University",
"gligoric@cs.stanford.edu",
"Source",
"Use vs.mention",
"Detection",
"Intention",
"Abstract",
"Hate speech",
"Hate speech",
"True use",
"detection",
"State",
"Warning: content in this paper may be upsetting",
"hate speech",
"A",
"or offensive.",
"Jews are born greedy.",
"",
"",
"A",
"The use of words to convey speaker's intent is",
"Speaker",
"Mention to challenge",
"traditionally distinguished from the 'mention'",
"Challenge",
"The idea that Jews are",
"hate speech",
"borngreedyarisesfron",
"A",
"of words for quoting what someone said, or",
"Y",
"ievalstereotypes",
"日",
"pointing out properties of a word. Here we",
"that's the only truth.",
"show that computationally modeling this use-",
"Misinformation",
"True use",
"Misinformation",
"mention distinctionis crucial fordealing with",
"detection",
"State",
"It's NOTa vaccine it is",
"GENE THERAPY",
"口",
"counterspeech online.Counterspeechthatre-",
"misinformation",
"It changes your DNA which",
"图",
"futes problematic content often mentions harm-",
"",
"Speaker",
"youcannotchangeback",
"ful language but is not harmful itself (e.g., call-",
"Mention to challenge",
"['s]",
"",
"Challenge",
"Yourclaimthatvaccineis",
"ing a vaccine dangerous is not the same as ex-",
"misinformation",
"gene therapy is inaccurate",
"O",
"宝R",
"Multipleindependent",
"2",
"fact-checking [..]",
"vaccinesdangerous).Weshowthatevenrecent",
"language models fail at distinguishing use from",
"Figure I: Ilustration of harmful censorship of coun-",
"mention, and that this failure propagates to",
"terspeech. Depending on the speaker's intention, the",
"[A[",
"twokeydownstreamtasks: misinformation and",
"same phrases are either used to state hate speech and",
"hate speech detection, resulting in censorship",
"misinformation,ormentioned,to provide a counternar-",
"of counterspeech.We introduce prompting mit-",
"1",
"pressing disapproval of someone for calling",
"rative. Algorithms that indiscriminately classify coun-",
"igations that teach the use-mention distinction,",
"terspeech as abusive censor constructive discourse.",
"and show they reducetheseerrors.Ourwork",
"highlights the importance of the use-mention",
"distinction for NLP and CSS and offers ways",
"(Wright et al., 2017; Mun et al., 2023; Hangart-",
"to address it.",
"ner et al., 2021; Ecker et al., 2022). And men-",
"Introduction",
"",
"tioning is similarly crucial in media and academic",
"The use-mention distinction is the difference be-",
"reporting where researchers and journalists re-",
"port harmful content (Kirk et al., 2022), in edu-",
"tween using words (Bananas have a peel) and men-",
"cational settings where problematic material is in-",
"tioning them (\"Bananas\" has 7 letters, or Dan said",
" voked for educational purposes, in disclosures in",
"“\"Bananas\"). The distinction has long been important",
"legal settings where harmful statements need to be",
"in the philosophy of language (Sperber and Wilson,",
"quoted (Henderson et al., 2022), and in personal",
"198l;Saka,1998),wherediscussionsdatebackto",
"testimonies (Wexler et al., 2019).",
"Tarski (1931) and Quine (1940).",
"The ability to correctly make this distinction",
"Inthis work. wefocus onthefirst of these: on",
"is particularly relevant to dealing with problem-",
"line counterspeech, defined as speech produced",
"atic language online. While detecting problematic",
"by users of online platforms to counteract harm-",
"text has become a standard task in NLP, there has",
"ful speech of others. This effort seeks to stop the",
"been less consideration of text that might men-",
"spread of harmful speech, mitigate its effects, dis-",
"tion harmful content without directly using it.",
"courage its recurrence, and provide support to both",
"Mentioning problematic content is critical to lan-",
" the targeted individuals and those joining in the",
"guage like counterspeech that challenges or op-",
"counterspeech efforts (Garland et al., 2022).",
"poses hateful or misleading narratives (Table 1)",
"Counterspeech statements often involve refer-"
] |
{
"arxiv_id": "2404.01651",
"language": "en",
"page_idx": 0
}
|
||
2404.01651-en-2
|
2404.01651-en-2.jpg
|
(2) Failure analyses We identify use-mention distinction failures, show that errors propagate to downstream tasks, and trace failures to specific target entities (misinformation-related and identity terms) and to the strength of the expressed stance (Sec. 4- 4. 3) .
(3) Mitigations We investigate prompting mitigations and implications for downstream tasks. We show that our interventions lead to a significant reduction in error (Sec. 4.4).
To support the further evaluation and development of models, our code is available at https://github.com/kristinagligoric/use-mention. Datasets are publicly available and can be used for research purposes.
# 2 Background
Counterspeech Building upon social science literature showing that counternarratives are effective against hate speech ( Andrews, 2002; Benesch, 2014; Schieb and Preuss, 2016; Garland et al., 2022), previous work in NLP and HCI has studied counterspeech from various perspectives. Scholars have curated datasets of counterspeech from different sources, including experts, NGO workers, and social media comments (Mathew et al., 2019; Chung et al., 2019; Qian et al., 2019; Garland et al., 2020; Fanton et al., 2021). Others have investigated different types of counterspeech strategies and performed user studies to evaluate the effectiveness of machine-generated counterspeech (Mun et al., 2023; Fraser et al., 2023). Our work also builds upon existing models for counterspeech detection (Garland et al., 2020).
Content moderation policies on counterspeech mentions Several online platforms acknowledge the importance of counterspeech in their content policies. At the time of writing, TikTok's Community Principles states “We do not allow language $or\textit{ behavior that harasses, humiliates, threatens, }$ or doxxes anyone. This also includes responding to such acts with retaliatory harassment (but excludes non-harassing counter speech)" (TikTok, 2023). Similarly, Facebook publisher and creator guidelines state“We know that many publishers use Facebook to challenge ideas, institutions and practices. Such discussion can promote debate and greater understanding". The guidelines also provide advice on how to write counterspeech to avoid mislabeling as hate speech (Meta, 2023). Develop-
ment of models that enable enforcement of such
policies is thus pressing.
While previous work has not explored the impact that the use-mention distinction has on online text classification tasks, mentions of toxic phrases were described as one class of false positive errors in toxic comment classification (Van Aken et al., 2018), e.g.,“I deleted the <identity group> are dumb comment", establishing that errors due to use-mention distinction are prevalent.
The use-mention distinction The use-mention distinction has been studied in philosophy (Sperber and Wilson, 1981), computational linguistics (Wilson, 2010, 2011b; Behzad et al., 2023), and HCI (Anderson et al., 2002). In general, mention is defined as follows:
Definition 1“for a token or a set of tokens t in a sentence S, if t refers to a property of the token t, then t is an instance of mention." (Wilson, 2010)
While many facts about a token can be a mention property, in our domain of counterspeech we focus on two properties of mentioned language (Sandhan et al., 2023; Wilson, 201 1b): attributed language (e.g., mentions to refer to the quotes of original source or stance towards the source) and words or phrases as themselves (e.g., mentions of words to refer to stance towards their use).$^1$
The use-mention distinction and related tasks Like the use-mention distinction, other NLP tasks bear on speaker intent, such as those related to factuality (Saurí and Pustejovsky, 2012, 2009; Murzaku et al., 2022) and committed belief (Prabhakaran et al., 2010, 2015; De Marneffe et al., 2019). In the context of mentioned language, such tasks aim to directly take into account the speaker's intention when mentioning specific phrases, and in particular, whether the speech act commits the speaker to the truth or factuality of the expressed proposition. However, in both used and mentioned language, statements do not necessarily commit the speaker to the truth of the source statement or to its factuality. Moreover, harmful language is often implicit, formalized as questions and nuanced statements that need not constitute a committed belief or be factual either. The distinction between using and mentioning is thus related to but distinct from factuality and committed belief.
${^{1}\text{See the Limitations section on other types of mentioned}}$
language which do not lead as clearly to practical harms.
|
[
[
0.1175537109375,
0.0843505859375,
0.48779296875,
0.0843505859375,
0.48779296875,
0.09857177734375,
0.1175537109375,
0.09857177734375
],
[
0.509765625,
0.08489990234375,
0.87890625,
0.08489990234375,
0.87890625,
0.09857177734375,
0.509765625,
0.09857177734375
],
[
0.1158447265625,
0.1009521484375,
0.48779296875,
0.1015625,
0.48779296875,
0.11639404296875,
0.1158447265625,
0.11517333984375
],
[
0.51123046875,
0.10272216796875,
0.6962890625,
0.10272216796875,
0.6962890625,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.5283203125,
0.1175537109375,
0.880859375,
0.1181640625,
0.880859375,
0.133056640625,
0.5283203125,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.484375,
0.1341552734375,
0.484375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51318359375,
0.135986328125,
0.880859375,
0.135986328125,
0.880859375,
0.149658203125,
0.51318359375,
0.149658203125
],
[
0.1158447265625,
0.15087890625,
0.48779296875,
0.15087890625,
0.48779296875,
0.1644287109375,
0.1158447265625,
0.1644287109375
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.1175537109375,
0.16748046875,
0.20654296875,
0.16748046875,
0.20654296875,
0.1787109375,
0.1175537109375,
0.1787109375
],
[
0.51123046875,
0.166259765625,
0.87890625,
0.166259765625,
0.87890625,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.51318359375,
0.18408203125,
0.87890625,
0.18408203125,
0.87890625,
0.19482421875,
0.51318359375,
0.19482421875
],
[
0.118408203125,
0.193603515625,
0.489501953125,
0.193603515625,
0.489501953125,
0.207275390625,
0.118408203125,
0.207275390625
],
[
0.51318359375,
0.1995849609375,
0.8818359375,
0.1995849609375,
0.8818359375,
0.21435546875,
0.51318359375,
0.21435546875
],
[
0.1175537109375,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.1175537109375,
0.223876953125
],
[
0.51123046875,
0.215576171875,
0.8818359375,
0.215576171875,
0.8818359375,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.1175537109375,
0.2257080078125,
0.48779296875,
0.2257080078125,
0.48779296875,
0.2392578125,
0.1175537109375,
0.2392578125
],
[
0.509765625,
0.232177734375,
0.791015625,
0.232177734375,
0.791015625,
0.2452392578125,
0.509765625,
0.2452392578125
],
[
0.1158447265625,
0.2410888671875,
0.32666015625,
0.2410888671875,
0.32666015625,
0.254638671875,
0.1158447265625,
0.254638671875
],
[
0.1351318359375,
0.2587890625,
0.490234375,
0.2587890625,
0.490234375,
0.2724609375,
0.1351318359375,
0.2724609375
],
[
0.51318359375,
0.260009765625,
0.8818359375,
0.260009765625,
0.8818359375,
0.273681640625,
0.51318359375,
0.273681640625
],
[
0.1158447265625,
0.27490234375,
0.489501953125,
0.27490234375,
0.489501953125,
0.2880859375,
0.1158447265625,
0.2880859375
],
[
0.51123046875,
0.275634765625,
0.87890625,
0.275634765625,
0.87890625,
0.289306640625,
0.51123046875,
0.289306640625
],
[
0.1175537109375,
0.29150390625,
0.489501953125,
0.29150390625,
0.489501953125,
0.3046875,
0.1175537109375,
0.3046875
],
[
0.51123046875,
0.29150390625,
0.880859375,
0.292236328125,
0.880859375,
0.30712890625,
0.51123046875,
0.305908203125
],
[
0.1175537109375,
0.30712890625,
0.489501953125,
0.30712890625,
0.489501953125,
0.320556640625,
0.1175537109375,
0.320556640625
],
[
0.51318359375,
0.30810546875,
0.8818359375,
0.30810546875,
0.8818359375,
0.32177734375,
0.51318359375,
0.32177734375
],
[
0.1158447265625,
0.32373046875,
0.255126953125,
0.32373046875,
0.255126953125,
0.33740234375,
0.1158447265625,
0.33740234375
],
[
0.51318359375,
0.32470703125,
0.8818359375,
0.32470703125,
0.8818359375,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.509765625,
0.339111328125,
0.6689453125,
0.339111328125,
0.6689453125,
0.352783203125,
0.509765625,
0.352783203125
],
[
0.1158447265625,
0.352783203125,
0.255126953125,
0.352783203125,
0.255126953125,
0.36767578125,
0.1158447265625,
0.36767578125
],
[
0.51123046875,
0.365234375,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.51123046875,
0.37890625
],
[
0.1158447265625,
0.37890625,
0.489501953125,
0.37890625,
0.489501953125,
0.393798828125,
0.1158447265625,
0.393798828125
],
[
0.509765625,
0.3818359375,
0.88330078125,
0.3818359375,
0.88330078125,
0.3955078125,
0.509765625,
0.3955078125
],
[
0.1158447265625,
0.396728515625,
0.491943359375,
0.396728515625,
0.491943359375,
0.410400390625,
0.1158447265625,
0.410400390625
],
[
0.509765625,
0.39794921875,
0.873046875,
0.39794921875,
0.873046875,
0.410888671875,
0.509765625,
0.410888671875
],
[
0.1175537109375,
0.413330078125,
0.48779296875,
0.413330078125,
0.48779296875,
0.427001953125,
0.1175537109375,
0.427001953125
],
[
0.52880859375,
0.42041015625,
0.8818359375,
0.42041015625,
0.8818359375,
0.43408203125,
0.52880859375,
0.43408203125
],
[
0.1175537109375,
0.4287109375,
0.48779296875,
0.4287109375,
0.48779296875,
0.4423828125,
0.1175537109375,
0.4423828125
],
[
0.509765625,
0.43701171875,
0.8818359375,
0.435302734375,
0.8818359375,
0.449462890625,
0.509765625,
0.451904296875
],
[
0.1175537109375,
0.44482421875,
0.48779296875,
0.44482421875,
0.48779296875,
0.458984375,
0.1175537109375,
0.458984375
],
[
0.51123046875,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.51123046875,
0.466064453125
],
[
0.1175537109375,
0.46142578125,
0.48779296875,
0.46142578125,
0.48779296875,
0.474365234375,
0.1175537109375,
0.474365234375
],
[
0.51123046875,
0.46728515625,
0.880859375,
0.468505859375,
0.880859375,
0.482666015625,
0.51123046875,
0.482177734375
],
[
0.1175537109375,
0.476806640625,
0.489501953125,
0.476806640625,
0.489501953125,
0.490478515625,
0.1175537109375,
0.490478515625
],
[
0.51123046875,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.51123046875,
0.498779296875
],
[
0.1158447265625,
0.4921875,
0.48779296875,
0.4912109375,
0.48779296875,
0.505859375,
0.1158447265625,
0.50732421875
],
[
0.509765625,
0.50048828125,
0.87890625,
0.50048828125,
0.87890625,
0.51416015625,
0.509765625,
0.51416015625
],
[
0.1175537109375,
0.5087890625,
0.489501953125,
0.5087890625,
0.489501953125,
0.5224609375,
0.1175537109375,
0.5224609375
],
[
0.51318359375,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.5322265625,
0.51318359375,
0.5322265625
],
[
0.1175537109375,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5390625,
0.1175537109375,
0.5390625
],
[
0.51123046875,
0.533203125,
0.76171875,
0.533203125,
0.76171875,
0.54638671875,
0.51123046875,
0.54638671875
],
[
0.1175537109375,
0.54150390625,
0.486083984375,
0.54150390625,
0.486083984375,
0.5546875,
0.1175537109375,
0.5546875
],
[
0.1158447265625,
0.5556640625,
0.491943359375,
0.55712890625,
0.491943359375,
0.57177734375,
0.1158447265625,
0.57080078125
],
[
0.51318359375,
0.56103515625,
0.8818359375,
0.56103515625,
0.8818359375,
0.57470703125,
0.51318359375,
0.57470703125
],
[
0.1175537109375,
0.57373046875,
0.489501953125,
0.57373046875,
0.489501953125,
0.58740234375,
0.1175537109375,
0.58740234375
],
[
0.51123046875,
0.57666015625,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.11419677734375,
0.58837890625,
0.48779296875,
0.58837890625,
0.48779296875,
0.6015625,
0.11419677734375,
0.6015625
],
[
0.51318359375,
0.595703125,
0.880859375,
0.595703125,
0.880859375,
0.60546875,
0.51318359375,
0.60546875
],
[
0.1158447265625,
0.60498046875,
0.486083984375,
0.60498046875,
0.486083984375,
0.6181640625,
0.1158447265625,
0.6181640625
],
[
0.51318359375,
0.60986328125,
0.88330078125,
0.60986328125,
0.88330078125,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.1175537109375,
0.62158203125,
0.48779296875,
0.62158203125,
0.48779296875,
0.6357421875,
0.1175537109375,
0.6357421875
],
[
0.51123046875,
0.62451171875,
0.880859375,
0.62451171875,
0.880859375,
0.63818359375,
0.51123046875,
0.63818359375
],
[
0.1175537109375,
0.63720703125,
0.279541015625,
0.63720703125,
0.279541015625,
0.65185546875,
0.1175537109375,
0.65185546875
],
[
0.51123046875,
0.64111328125,
0.88330078125,
0.64111328125,
0.88330078125,
0.65478515625,
0.51123046875,
0.65478515625
],
[
0.51123046875,
0.65673828125,
0.8818359375,
0.65771484375,
0.8818359375,
0.6728515625,
0.51123046875,
0.67138671875
],
[
0.118408203125,
0.6650390625,
0.486083984375,
0.6650390625,
0.486083984375,
0.6787109375,
0.118408203125,
0.6787109375
],
[
0.509765625,
0.67333984375,
0.880859375,
0.67333984375,
0.880859375,
0.68701171875,
0.509765625,
0.68701171875
],
[
0.1158447265625,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.11419677734375,
0.69677734375,
0.486083984375,
0.69677734375,
0.486083984375,
0.71044921875,
0.11419677734375,
0.71044921875
],
[
0.51123046875,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.51123046875,
0.705078125
],
[
0.509765625,
0.70703125,
0.8818359375,
0.70703125,
0.8818359375,
0.72021484375,
0.509765625,
0.72021484375
],
[
0.1175537109375,
0.71337890625,
0.490234375,
0.71337890625,
0.490234375,
0.72705078125,
0.1175537109375,
0.72705078125
],
[
0.51123046875,
0.72265625,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.51123046875,
0.73681640625
],
[
0.1158447265625,
0.728515625,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.1158447265625,
0.74365234375
],
[
0.11419677734375,
0.74462890625,
0.489501953125,
0.7451171875,
0.489501953125,
0.76025390625,
0.11419677734375,
0.7587890625
],
[
0.51123046875,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.509765625,
0.7548828125,
0.8818359375,
0.75439453125,
0.8818359375,
0.7685546875,
0.509765625,
0.76953125
],
[
0.1175537109375,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.1175537109375,
0.775390625
],
[
0.509765625,
0.771484375,
0.8818359375,
0.771484375,
0.8818359375,
0.78515625,
0.509765625,
0.78515625
],
[
0.1175537109375,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.79150390625,
0.1175537109375,
0.79150390625
],
[
0.509765625,
0.78759765625,
0.88330078125,
0.78759765625,
0.88330078125,
0.80029296875,
0.509765625,
0.80029296875
],
[
0.1158447265625,
0.79345703125,
0.489501953125,
0.7919921875,
0.489501953125,
0.80712890625,
0.1158447265625,
0.80810546875
],
[
0.509765625,
0.802734375,
0.8818359375,
0.802734375,
0.8818359375,
0.81640625,
0.509765625,
0.81640625
],
[
0.1175537109375,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.1175537109375,
0.82373046875
],
[
0.1158447265625,
0.82470703125,
0.486083984375,
0.826171875,
0.486083984375,
0.84033203125,
0.1158447265625,
0.8388671875
],
[
0.51318359375,
0.8193359375,
0.8818359375,
0.8193359375,
0.8818359375,
0.8330078125,
0.51318359375,
0.8330078125
],
[
0.509765625,
0.83447265625,
0.880859375,
0.8349609375,
0.880859375,
0.849609375,
0.509765625,
0.8486328125
],
[
0.1158447265625,
0.8427734375,
0.489501953125,
0.8427734375,
0.489501953125,
0.85546875,
0.1158447265625,
0.85546875
],
[
0.51123046875,
0.8515625,
0.880859375,
0.8515625,
0.880859375,
0.865234375,
0.51123046875,
0.865234375
],
[
0.1158447265625,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87255859375,
0.1158447265625,
0.87255859375
],
[
0.51318359375,
0.86767578125,
0.748046875,
0.86767578125,
0.748046875,
0.88134765625,
0.51318359375,
0.88134765625
],
[
0.1175537109375,
0.87451171875,
0.491943359375,
0.87451171875,
0.491943359375,
0.88818359375,
0.1175537109375,
0.88818359375
],
[
0.1175537109375,
0.89013671875,
0.48779296875,
0.89013671875,
0.48779296875,
0.90380859375,
0.1175537109375,
0.90380859375
],
[
0.5322265625,
0.8935546875,
0.8818359375,
0.89453125,
0.8818359375,
0.9091796875,
0.5322265625,
0.90771484375
],
[
0.1158447265625,
0.90673828125,
0.491943359375,
0.90673828125,
0.491943359375,
0.92041015625,
0.1158447265625,
0.92041015625
],
[
0.5087890625,
0.90771484375,
0.85400390625,
0.90771484375,
0.85400390625,
0.92138671875,
0.5087890625,
0.92138671875
]
] |
[
"(2) Failure analyses We identify use-mention",
"ment of models that enable enforcement of such",
"distinction failures, show that errors propagate to",
"policies is thus pressing.",
"downstream tasks, and trace failures to specific",
"While previous work has not explored the im-",
"target entities (misinformation-related and identity",
"pact that the use-mention distinction has on online",
"terms) and to the strength of the expressed stance",
"text classification tasks, mentions of toxic phrases",
"(Sec.4 4.3).",
"were described as one class of false positive errors",
"in toxic comment classification (Van Aken et al.",
"(3) Mitigations We investigate prompting miti-",
"2018), e.g., \"I deleted the <identity group> are",
"gations and implications for downstream tasks. We",
"dumb comment\", establishing that errors due to",
"show that our interventions lead to a significant",
"use-mention distinction are prevalent.",
"reduction in error (Sec. 4.4).",
"To support the further evaluation and de-",
"The use-mention distinction The use-mention",
"velopment of models, our code is available",
"distinction has been studied in philosophy (Sper",
"at https://github.com/kristinagligoric/use-mention.",
"ber and Wilson, 1981), computational linguistics",
"Datasets are publicly available and can be used for",
"(Wilson, 2010, 2011b; Behzad et al., 2023), and",
"research purposes.",
"HCI (Anderson et al., 2002). In general, mention",
"is defined as follows:",
"2Background",
"Definition 1\"for a token or a set of tokens t in a",
"Counterspeech Building upon social science lit-",
"sentence S, if t refers to a property of the token t,",
"erature showing that counternarratives are effec-",
"then t is an instance of mention.\" (Wilson, 2010)",
"tive against hate speech (Andrews, 2002; Benesch,",
"While many facts about a token can be a mention",
"2014; Schieb and Preuss, 2016; Garland et al.,",
"property, in our domain of counterspeech we focus",
"2022), previous work in NLP and HCI has studied",
"on two properties of mentioned language (Sandhan",
"counterspeech from various perspectives. Scholars",
"et al., 2023; Wilson, 2011b): attributed language",
"have curated datasets of counterspeech from dif-",
"(e.g., mentions to refer to the quotes of original",
"ferent sources, including experts, NGO workers,",
"source or stance towards the source) and words or",
"and social media comments (Mathew et al., 2019;",
"phrases as themselves (e.g., mentions of words to",
"Chung et al., 2019; Qian et al., 2019; Garland et al..",
"refer to stance towards their use).",
"2020; Fanton et al.,2021).Others have investigated",
"different types of counterspeech strategies and per-",
"The use-mention distinction and related tasks",
"formed user studies to evaluate the effectiveness",
"Like the use-mention distinction, other NLP tasks",
"of machine-generated counterspeech (Mun et al..",
"bear onspeakerintent.suchasthose relatedto",
"2023; Fraser et al., 2023). Our work also builds",
"factuality (Sauri and Pustejovsky, 2012, 2009;",
"upon existing models for counterspeech detection",
"Murzaku et al., 2022) and committed belief (Prab-",
"(Garland et al., 2020).",
"hakaran et al., 2010, 2015; De Marneffe et al.,",
"2019). In the context of mentioned language, such",
"Content moderation policies on counterspeech",
"tasks aim to directly take into account the speaker's",
"mentionsSeveral online platforms acknowledge",
"the importance of counterspeech in their content",
"intention when mentioning specific phrases, and",
"in particular, whether the speech act commits the",
"policies. At the time of writing, TikTok's Commu-",
"speaker to the truth or factuality of the expressed",
"nity Principles states \"We do not allow language",
"or behavior that harasses, humiliates, threatens,",
"proposition. However, in both used and mentioned",
"language, statements do not necessarily commit the",
"or doxxes anyone. This also includes responding",
" speaker to the truth of the source statement or to",
"to such acts with retaliatory harassment (but ex-",
"its factuality. Moreover, harmful language is often",
"cludes non-harassing counter speech)\" (TikTok,",
"implicit, formalized as questions and nuanced state",
"2023). Similarly, Facebook publisher and creator",
"guidelines state “\"We know that many publishers",
"ments that need not constitute a committed belief",
"or be factual either.The distinction between using",
"use Facebook to challenge ideas, institutions and",
"and mentioning is thus related to but distinct from",
"practices. Such discussion can promote debate and",
"factuality and committed belief.",
"greater understanding\". The guidelines also pro-",
"vide advice on how to write counterspeech to avoid",
"'See the Limitations section on other types of mentioned",
"mislabeling as hate speech (Meta, 2023). Develop-",
"language which do not lead as clearly to practical harms"
] |
{
"arxiv_id": "2404.01651",
"language": "en",
"page_idx": 2
}
|
||
2404.01651-en-4
|
2404.01651-en-4.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Sub-task}$</th>
<th>$\mathbf{Model}$</th>
<th>False positive rate</th>
<th>False negative rate $\downarrow$</th>
<th>$\textbf{Average error rate}$</th>
</tr>
<tr>
<td rowspan="2">$\textbf{Hate speech}$</td>
<td>gpt-3.5-instruct-turbo</td>
<td>$17.98\pm7.55$</td>
<td>$14.77\pm6.73$</td>
<td>$16.38\pm5.34$</td>
</tr>
<tr>
<td>$gpt-4$</td>
<td>$20.00\pm8.57$</td>
<td>$4.44\pm4.34$</td>
<td>$12.22\pm4.74$</td>
</tr>
<tr>
<td rowspan="4">Misinformation</td>
<td>gpt-3.5-instruct-turbo</td>
<td>$34.38\pm3.76$</td>
<td>40.08 $3\pm3.33$</td>
<td>$37.22\pm2.21$</td>
</tr>
<tr>
<td>gpt-3.5-instruct-turbo gpt-3.5-turbo (ChatGPT 3.5) gpt- 4</td>
<td>$34.38\pm3.76$ $8.05\pm1.83$ $23.44\pm2.74$</td>
<td>$40.08\pm3.33$ $49.76\pm3.17$ $3.89\pm1.38$</td>
<td>$37.22\pm2.21$ $28.93\pm1.86$ $13.64\pm1.42$</td>
</tr>
<tr>
<td>gpt-3.5-turbo (ChatGPT 3.5)</td>
<td>$8.05\pm1.83$</td>
<td>$49.76\pm3.17$</td>
<td>$28.93\pm1.86$</td>
</tr>
<tr>
<td>gpt- 4</td>
<td>$23.44\pm2.74$</td>
<td>$3.89\pm1.38$</td>
<td>$13.64\pm1.42$</td>
</tr>
</tbody>
</table>
Table 2: Use-mention classification. False positive, false negative, and average error rates in detecting use (in percentages). Models are sorted by the average error rate. The best performing model has a high false positive rate, mistaking mention for use, which can lead to censorship of useful counterspeech.
<table>
<tbody>
<tr>
<th>$\mathbf{Sub-task}$</th>
<th>$\mathbf{Model}$</th>
<th>False positive rate J</th>
<th>False negative rate</th>
<th>$\textbf{Average error rate}$</th>
</tr>
<tr>
<td rowspan="8">$\textbf{Hate speech}$</td>
<td>toxigen hatebert</td>
<td>$24.44\pm10.0$</td>
<td>$77.78\pm7.22$</td>
<td>$51.11\pm8.08$</td>
</tr>
<tr>
<td>perspective (insult)</td>
<td>$4.44\pm3.89$</td>
<td>$61.11\pm9.85$</td>
<td>$32.78\pm6.27$</td>
</tr>
<tr>
<td>perspective (toxicity)</td>
<td>$20.00\pm7.51$</td>
<td>$36.67\pm9.18$</td>
<td>$28.33\pm8.49$</td>
</tr>
<tr>
<td>perspective (identity attack)</td>
<td>$21.11\pm7.51$</td>
<td>$33.33\pm9.74$</td>
<td>$27.22\pm9.18$</td>
</tr>
<tr>
<td>roberta hate speech</td>
<td>$17.78\pm7.51$</td>
<td>$26.67\pm8.65$</td>
<td>$22.22\pm8.22$</td>
</tr>
<tr>
<td>gpt-3.5-instruct-turbo</td>
<td>$25.56\pm8.62$</td>
<td>$13.33\pm7.51$</td>
<td>$19.44\pm8.20$</td>
</tr>
<tr>
<td>ont- 35- turho 111 (ChatGPT 3.5)</td>
<td>$11.11\pm6.14$</td>
<td>$22.22\pm8.07$</td>
<td>$16.67\pm7.24$</td>
</tr>
<tr>
<td>$gpt-4$</td>
<td>$8.89\pm5.00$</td>
<td>$20.00\pm7.51$</td>
<td>$14.44\pm7.09$</td>
</tr>
<tr>
<td rowspan="4">Misinformation</td>
<td>roberta fake news</td>
<td>$97.93\pm1.00$</td>
<td>$5.10\pm1.52$</td>
<td>$51.52\pm1.20$</td>
</tr>
<tr>
<td>gpt-3.5-instruct-turbo</td>
<td>$26.12\pm2.40$</td>
<td>$19.44\pm2.62$</td>
<td>$22.78\pm2.80$</td>
</tr>
<tr>
<td>gpt-3.5-turbo (ChatGPT 3.5)</td>
<td>$22.11\pm3.02$</td>
<td>$13.85\pm2.74$</td>
<td>$17.98\pm3.06$</td>
</tr>
<tr>
<td>gpt- 4</td>
<td>$10.21\pm2.00$</td>
<td>$8.02\pm1.85$</td>
<td>$9.11\pm1.93$</td>
</tr>
</tbody>
</table>
Table 3: Downstream tasks (hate speech and misinformation detection). False positive, false negative, and average error rates in detecting hate speech and misinformation (in percentages). Models are sorted by the average error rate. Across models, substantial errors persist in false positive rate (classifying counterspeech mentions as harmful).
hate speech and misinformation detection: Perspective (Perspective, 2023), Toxigen (Hartvigsen et al., 2022), RoBERTa (Liu et al., 2019), and RoBERTa fake news (Ahmed et al., 2017).
For prompting, we use default parameters (temperature=1) and max output token length 1 (outputting either A or B). The classification prompt includes the instruction and a definition of the classes (for prompt variants and complete prompt text see Appendix, Section D).
# 4 Results
### 4.1 Use-mention classification (H1)
How well do the models distinguish whether problematic language is used or mentioned? Across the models (Table 2), average error rates are high (between 12.22% and 16.38% for hate speech and between 13.64% and 37.22% for misinformation), suggesting that state-of-the-art large language models struggle to distinguish use from mention in domains where the distinction matters. The bestperforming model, gpt-4, still has a very high false positive rate, mistaking mention for use; mentions
are misclassified as use in 20.00% of hate speech and 23.44% of misinformation counternarratives. In these settings, mistaking mention for use leads to the consequential harm of censoring useful counterspeech.
4.2 Downstream content classification (H2)
The prior section showed that state-of-the-art systems often fail at the use-mention distinction. Here we test the impact on hate speech and misinformation detection, two downstream classification tasks where the use-mention distinction matters (Table 3).
Performance on downstream tasks First,on the hate speech detection task, we find that misclassification of counterspeech as hateful is relatively frequent: the popular Toxigen and Perspective API have a false positive rate on counterspeech of over 20%, and recent models still have many false positives, although somewhat reduced, e.g., gpt-3.5- turbo has FPR 11.11%. Regarding the average error rate, gpt-4 is the best-performing model (false positive rate on counterspeech 8.89%).
|
[
[
0.4248046875,
0.0872802734375,
0.73486328125,
0.0872802734375,
0.73486328125,
0.1009521484375,
0.4248046875,
0.1009521484375
],
[
0.736328125,
0.0872802734375,
0.8759765625,
0.0872802734375,
0.8759765625,
0.1009521484375,
0.736328125,
0.1009521484375
],
[
0.2418212890625,
0.0885009765625,
0.287109375,
0.0885009765625,
0.287109375,
0.09979248046875,
0.2418212890625,
0.09979248046875
],
[
0.1259765625,
0.0885009765625,
0.187255859375,
0.0885009765625,
0.187255859375,
0.09979248046875,
0.1259765625,
0.09979248046875
],
[
0.1259765625,
0.105712890625,
0.208251953125,
0.105712890625,
0.208251953125,
0.1175537109375,
0.1259765625,
0.1175537109375
],
[
0.2359619140625,
0.1051025390625,
0.372802734375,
0.10394287109375,
0.372802734375,
0.1181640625,
0.2359619140625,
0.119384765625
],
[
0.486083984375,
0.105712890625,
0.5693359375,
0.105712890625,
0.5693359375,
0.1175537109375,
0.486083984375,
0.1175537109375
],
[
0.634765625,
0.105712890625,
0.71875,
0.105712890625,
0.71875,
0.1175537109375,
0.634765625,
0.1175537109375
],
[
0.79248046875,
0.105712890625,
0.8759765625,
0.105712890625,
0.8759765625,
0.1175537109375,
0.79248046875,
0.1175537109375
],
[
0.2359619140625,
0.1175537109375,
0.41650390625,
0.11517333984375,
0.41650390625,
0.1300048828125,
0.2359619140625,
0.1318359375
],
[
0.489501953125,
0.1175537109375,
0.568359375,
0.1175537109375,
0.568359375,
0.12890625,
0.489501953125,
0.12890625
],
[
0.634765625,
0.1181640625,
0.71875,
0.1181640625,
0.71875,
0.1300048828125,
0.634765625,
0.1300048828125
],
[
0.79248046875,
0.1181640625,
0.873046875,
0.1181640625,
0.873046875,
0.1300048828125,
0.79248046875,
0.1300048828125
],
[
0.2376708984375,
0.1300048828125,
0.276123046875,
0.1300048828125,
0.276123046875,
0.142578125,
0.2376708984375,
0.142578125
],
[
0.48193359375,
0.1300048828125,
0.568359375,
0.1300048828125,
0.568359375,
0.141357421875,
0.48193359375,
0.141357421875
],
[
0.640625,
0.1300048828125,
0.71875,
0.1300048828125,
0.71875,
0.141357421875,
0.640625,
0.141357421875
],
[
0.791015625,
0.1300048828125,
0.875,
0.1300048828125,
0.875,
0.141357421875,
0.791015625,
0.141357421875
],
[
0.480224609375,
0.144287109375,
0.57275390625,
0.146728515625,
0.57275390625,
0.160888671875,
0.480224609375,
0.1591796875
],
[
0.63232421875,
0.144287109375,
0.7197265625,
0.146728515625,
0.7197265625,
0.160888671875,
0.63232421875,
0.1591796875
],
[
0.791015625,
0.1478271484375,
0.875,
0.1478271484375,
0.875,
0.1591796875,
0.791015625,
0.1591796875
],
[
0.235107421875,
0.1591796875,
0.414794921875,
0.157958984375,
0.414794921875,
0.1728515625,
0.235107421875,
0.17333984375
],
[
0.124267578125,
0.1478271484375,
0.372802734375,
0.1478271484375,
0.372802734375,
0.160888671875,
0.124267578125,
0.160888671875
],
[
0.48779296875,
0.1602783203125,
0.5693359375,
0.1602783203125,
0.5693359375,
0.171630859375,
0.48779296875,
0.171630859375
],
[
0.634765625,
0.1602783203125,
0.71728515625,
0.1602783203125,
0.71728515625,
0.171630859375,
0.634765625,
0.171630859375
],
[
0.79248046875,
0.1602783203125,
0.8759765625,
0.1602783203125,
0.8759765625,
0.171630859375,
0.79248046875,
0.171630859375
],
[
0.484375,
0.171630859375,
0.5693359375,
0.171630859375,
0.5693359375,
0.182861328125,
0.484375,
0.182861328125
],
[
0.79248046875,
0.171630859375,
0.8759765625,
0.171630859375,
0.8759765625,
0.182861328125,
0.79248046875,
0.182861328125
],
[
0.640625,
0.1728515625,
0.7197265625,
0.1728515625,
0.7197265625,
0.182861328125,
0.640625,
0.182861328125
],
[
0.2359619140625,
0.1728515625,
0.274658203125,
0.17041015625,
0.276123046875,
0.182861328125,
0.2376708984375,
0.185302734375
],
[
0.1158447265625,
0.1995849609375,
0.8818359375,
0.1995849609375,
0.8818359375,
0.213134765625,
0.1158447265625,
0.213134765625
],
[
0.1158447265625,
0.21435546875,
0.88330078125,
0.21435546875,
0.88330078125,
0.22802734375,
0.1158447265625,
0.22802734375
],
[
0.1175537109375,
0.2298583984375,
0.66162109375,
0.2298583984375,
0.66162109375,
0.240478515625,
0.1175537109375,
0.240478515625
],
[
0.124267578125,
0.2724609375,
0.18896484375,
0.2724609375,
0.18896484375,
0.283935546875,
0.124267578125,
0.283935546875
],
[
0.2376708984375,
0.2724609375,
0.28369140625,
0.2724609375,
0.28369140625,
0.283935546875,
0.2376708984375,
0.283935546875
],
[
0.4248046875,
0.271484375,
0.7255859375,
0.271484375,
0.7255859375,
0.28515625,
0.4248046875,
0.28515625
],
[
0.73779296875,
0.273681640625,
0.875,
0.273681640625,
0.875,
0.283935546875,
0.73779296875,
0.283935546875
],
[
0.1259765625,
0.29150390625,
0.208251953125,
0.29150390625,
0.208251953125,
0.302734375,
0.1259765625,
0.302734375
],
[
0.48193359375,
0.290283203125,
0.5693359375,
0.290283203125,
0.5693359375,
0.3017578125,
0.48193359375,
0.3017578125
],
[
0.79248046875,
0.29150390625,
0.8759765625,
0.29150390625,
0.8759765625,
0.302734375,
0.79248046875,
0.302734375
],
[
0.634765625,
0.292236328125,
0.71875,
0.292236328125,
0.71875,
0.302734375,
0.634765625,
0.302734375
],
[
0.2376708984375,
0.292236328125,
0.339111328125,
0.292236328125,
0.339111328125,
0.302734375,
0.2376708984375,
0.302734375
],
[
0.2359619140625,
0.303955078125,
0.35693359375,
0.302734375,
0.35693359375,
0.314208984375,
0.2359619140625,
0.3154296875
],
[
0.489501953125,
0.303955078125,
0.57080078125,
0.303955078125,
0.57080078125,
0.314208984375,
0.489501953125,
0.314208984375
],
[
0.6337890625,
0.303955078125,
0.71875,
0.303955078125,
0.71875,
0.314208984375,
0.6337890625,
0.314208984375
],
[
0.79248046875,
0.303955078125,
0.875,
0.303955078125,
0.875,
0.314208984375,
0.79248046875,
0.314208984375
],
[
0.2359619140625,
0.314208984375,
0.370361328125,
0.314208984375,
0.370361328125,
0.327880859375,
0.2359619140625,
0.327880859375
],
[
0.483642578125,
0.314208984375,
0.568359375,
0.314208984375,
0.568359375,
0.325927734375,
0.483642578125,
0.325927734375
],
[
0.6337890625,
0.3154296875,
0.71875,
0.3154296875,
0.71875,
0.32666015625,
0.6337890625,
0.32666015625
],
[
0.79248046875,
0.3154296875,
0.8759765625,
0.3154296875,
0.8759765625,
0.325927734375,
0.79248046875,
0.325927734375
],
[
0.484375,
0.32666015625,
0.56689453125,
0.32666015625,
0.56689453125,
0.33837890625,
0.484375,
0.33837890625
],
[
0.634765625,
0.32666015625,
0.71875,
0.32666015625,
0.71875,
0.33837890625,
0.634765625,
0.33837890625
],
[
0.79248046875,
0.32666015625,
0.875,
0.32666015625,
0.875,
0.33837890625,
0.79248046875,
0.33837890625
],
[
0.791015625,
0.336181640625,
0.875,
0.33740234375,
0.875,
0.3486328125,
0.791015625,
0.347412109375
],
[
0.2376708984375,
0.327880859375,
0.41064453125,
0.327880859375,
0.41064453125,
0.33837890625,
0.2376708984375,
0.33837890625
],
[
0.484375,
0.33837890625,
0.56689453125,
0.33837890625,
0.56689453125,
0.349853515625,
0.484375,
0.349853515625
],
[
0.634765625,
0.33837890625,
0.71875,
0.33837890625,
0.71875,
0.349853515625,
0.634765625,
0.349853515625
],
[
0.791015625,
0.3486328125,
0.875,
0.3486328125,
0.875,
0.35986328125,
0.791015625,
0.35986328125
],
[
0.2376708984375,
0.339111328125,
0.358642578125,
0.339111328125,
0.358642578125,
0.349853515625,
0.2376708984375,
0.349853515625
],
[
0.2359619140625,
0.349853515625,
0.372802734375,
0.3486328125,
0.372802734375,
0.363525390625,
0.2359619140625,
0.364013671875
],
[
0.483642578125,
0.350830078125,
0.5693359375,
0.350830078125,
0.5693359375,
0.3623046875,
0.483642578125,
0.3623046875
],
[
0.634765625,
0.350830078125,
0.71728515625,
0.350830078125,
0.71728515625,
0.361083984375,
0.634765625,
0.361083984375
],
[
0.2359619140625,
0.3623046875,
0.414794921875,
0.3623046875,
0.414794921875,
0.3759765625,
0.2359619140625,
0.3759765625
],
[
0.484375,
0.3623046875,
0.5693359375,
0.3623046875,
0.5693359375,
0.37353515625,
0.484375,
0.37353515625
],
[
0.634765625,
0.361083984375,
0.71875,
0.361083984375,
0.71875,
0.372314453125,
0.634765625,
0.372314453125
],
[
0.79248046875,
0.3623046875,
0.8759765625,
0.3623046875,
0.8759765625,
0.37353515625,
0.79248046875,
0.37353515625
],
[
0.634765625,
0.37353515625,
0.71728515625,
0.37353515625,
0.71728515625,
0.385498046875,
0.634765625,
0.385498046875
],
[
0.791015625,
0.372314453125,
0.873046875,
0.372314453125,
0.873046875,
0.38427734375,
0.791015625,
0.38427734375
],
[
0.490234375,
0.374755859375,
0.57080078125,
0.374755859375,
0.57080078125,
0.385498046875,
0.490234375,
0.385498046875
],
[
0.2376708984375,
0.374755859375,
0.273681640625,
0.372314453125,
0.274658203125,
0.385498046875,
0.2392578125,
0.38720703125
],
[
0.124267578125,
0.39013671875,
0.3525390625,
0.39013671875,
0.3525390625,
0.40380859375,
0.124267578125,
0.40380859375
],
[
0.483642578125,
0.392578125,
0.57080078125,
0.392578125,
0.57080078125,
0.40380859375,
0.483642578125,
0.40380859375
],
[
0.640625,
0.392578125,
0.71875,
0.392578125,
0.71875,
0.40380859375,
0.640625,
0.40380859375
],
[
0.791015625,
0.392578125,
0.87744140625,
0.392578125,
0.87744140625,
0.40380859375,
0.791015625,
0.40380859375
],
[
0.2359619140625,
0.40380859375,
0.372802734375,
0.402587890625,
0.372802734375,
0.41748046875,
0.2359619140625,
0.418701171875
],
[
0.483642578125,
0.405029296875,
0.5693359375,
0.405029296875,
0.5693359375,
0.4150390625,
0.483642578125,
0.4150390625
],
[
0.634765625,
0.405029296875,
0.71875,
0.405029296875,
0.71875,
0.416259765625,
0.634765625,
0.416259765625
],
[
0.791015625,
0.405029296875,
0.8759765625,
0.405029296875,
0.8759765625,
0.4150390625,
0.791015625,
0.4150390625
],
[
0.2359619140625,
0.4150390625,
0.414794921875,
0.41455078125,
0.414794921875,
0.4287109375,
0.2359619140625,
0.429931640625
],
[
0.483642578125,
0.416259765625,
0.56689453125,
0.416259765625,
0.56689453125,
0.427490234375,
0.483642578125,
0.427490234375
],
[
0.634765625,
0.416259765625,
0.71875,
0.416259765625,
0.71875,
0.427490234375,
0.634765625,
0.427490234375
],
[
0.79248046875,
0.416259765625,
0.8759765625,
0.416259765625,
0.8759765625,
0.427490234375,
0.79248046875,
0.427490234375
],
[
0.486083984375,
0.427490234375,
0.57080078125,
0.427490234375,
0.57080078125,
0.439453125,
0.486083984375,
0.439453125
],
[
0.6396484375,
0.427490234375,
0.71875,
0.427490234375,
0.71875,
0.439453125,
0.6396484375,
0.439453125
],
[
0.796875,
0.427490234375,
0.875,
0.427490234375,
0.875,
0.439453125,
0.796875,
0.439453125
],
[
0.1158447265625,
0.4560546875,
0.8818359375,
0.4560546875,
0.8818359375,
0.4697265625,
0.1158447265625,
0.4697265625
],
[
0.2376708984375,
0.4287109375,
0.274658203125,
0.4287109375,
0.274658203125,
0.441162109375,
0.2376708984375,
0.441162109375
],
[
0.1158447265625,
0.47021484375,
0.8818359375,
0.47021484375,
0.8818359375,
0.48388671875,
0.1158447265625,
0.48388671875
],
[
0.1158447265625,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49755859375,
0.1158447265625,
0.49755859375
],
[
0.1158447265625,
0.49951171875,
0.181396484375,
0.49951171875,
0.181396484375,
0.51123046875,
0.1158447265625,
0.51123046875
],
[
0.1175537109375,
0.5380859375,
0.49365234375,
0.5380859375,
0.49365234375,
0.5517578125,
0.1175537109375,
0.5517578125
],
[
0.5087890625,
0.53759765625,
0.8818359375,
0.5380859375,
0.8818359375,
0.552734375,
0.5087890625,
0.5517578125
],
[
0.1158447265625,
0.55419921875,
0.491943359375,
0.55419921875,
0.491943359375,
0.56787109375,
0.1158447265625,
0.56787109375
],
[
0.509765625,
0.5546875,
0.88330078125,
0.5546875,
0.88330078125,
0.568359375,
0.509765625,
0.568359375
],
[
0.1158447265625,
0.5693359375,
0.48779296875,
0.5693359375,
0.48779296875,
0.5830078125,
0.1158447265625,
0.5830078125
],
[
0.509765625,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.509765625,
0.5849609375
],
[
0.1175537109375,
0.58740234375,
0.355224609375,
0.58740234375,
0.355224609375,
0.60107421875,
0.1175537109375,
0.60107421875
],
[
0.5087890625,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6015625,
0.5087890625,
0.6015625
],
[
0.1351318359375,
0.60400390625,
0.49365234375,
0.60400390625,
0.49365234375,
0.61767578125,
0.1351318359375,
0.61767578125
],
[
0.51123046875,
0.60498046875,
0.587890625,
0.60498046875,
0.587890625,
0.6162109375,
0.51123046875,
0.6162109375
],
[
0.1158447265625,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63427734375,
0.1158447265625,
0.63427734375
],
[
0.1158447265625,
0.6357421875,
0.490234375,
0.6357421875,
0.490234375,
0.6494140625,
0.1158447265625,
0.6494140625
],
[
0.51123046875,
0.63134765625,
0.861328125,
0.63134765625,
0.861328125,
0.64453125,
0.51123046875,
0.64453125
],
[
0.118408203125,
0.65380859375,
0.486083984375,
0.65380859375,
0.486083984375,
0.66455078125,
0.118408203125,
0.66455078125
],
[
0.5087890625,
0.65283203125,
0.8818359375,
0.65478515625,
0.8818359375,
0.66943359375,
0.5087890625,
0.66748046875
],
[
0.1175537109375,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.51123046875,
0.6728515625,
0.880859375,
0.6728515625,
0.880859375,
0.68310546875,
0.51123046875,
0.68310546875
],
[
0.1158447265625,
0.68408203125,
0.2822265625,
0.68310546875,
0.2822265625,
0.69775390625,
0.1158447265625,
0.69873046875
],
[
0.509765625,
0.68701171875,
0.88330078125,
0.68701171875,
0.88330078125,
0.70068359375,
0.509765625,
0.70068359375
],
[
0.51123046875,
0.703125,
0.8818359375,
0.703125,
0.8818359375,
0.71630859375,
0.51123046875,
0.71630859375
],
[
0.1158447265625,
0.7119140625,
0.214111328125,
0.7119140625,
0.214111328125,
0.72705078125,
0.1158447265625,
0.72705078125
],
[
0.51318359375,
0.72021484375,
0.880859375,
0.72021484375,
0.880859375,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.51123046875,
0.73388671875,
0.5859375,
0.73388671875,
0.5859375,
0.7490234375,
0.51123046875,
0.7490234375
],
[
0.1175537109375,
0.74072265625,
0.1435546875,
0.74072265625,
0.1435546875,
0.751953125,
0.1175537109375,
0.751953125
],
[
0.14697265625,
0.74462890625,
0.154541015625,
0.74462890625,
0.154541015625,
0.7490234375,
0.14697265625,
0.7490234375
],
[
0.154541015625,
0.7412109375,
0.403076171875,
0.7412109375,
0.403076171875,
0.751953125,
0.154541015625,
0.751953125
],
[
0.1158447265625,
0.7626953125,
0.489501953125,
0.7626953125,
0.489501953125,
0.775390625,
0.1158447265625,
0.775390625
],
[
0.51123046875,
0.7626953125,
0.8818359375,
0.7626953125,
0.8818359375,
0.775390625,
0.51123046875,
0.775390625
],
[
0.1175537109375,
0.779296875,
0.489501953125,
0.779296875,
0.489501953125,
0.7919921875,
0.1175537109375,
0.7919921875
],
[
0.51123046875,
0.779296875,
0.8818359375,
0.779296875,
0.8818359375,
0.7919921875,
0.51123046875,
0.7919921875
],
[
0.1175537109375,
0.79443359375,
0.486083984375,
0.79443359375,
0.486083984375,
0.80810546875,
0.1175537109375,
0.80810546875
],
[
0.51123046875,
0.79443359375,
0.8818359375,
0.79443359375,
0.8818359375,
0.80810546875,
0.51123046875,
0.80810546875
],
[
0.1175537109375,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82373046875,
0.1175537109375,
0.82373046875
],
[
0.51318359375,
0.8125,
0.880859375,
0.8125,
0.880859375,
0.822265625,
0.51318359375,
0.822265625
],
[
0.1175537109375,
0.826171875,
0.489501953125,
0.826171875,
0.489501953125,
0.84033203125,
0.1175537109375,
0.84033203125
],
[
0.51123046875,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.51123046875,
0.84033203125
],
[
0.1175537109375,
0.8427734375,
0.489501953125,
0.8427734375,
0.489501953125,
0.85546875,
0.1175537109375,
0.85546875
],
[
0.51123046875,
0.8427734375,
0.8818359375,
0.8427734375,
0.8818359375,
0.85546875,
0.51123046875,
0.85546875
],
[
0.1175537109375,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87255859375,
0.1175537109375,
0.87255859375
],
[
0.509765625,
0.859375,
0.8818359375,
0.859375,
0.8818359375,
0.87255859375,
0.509765625,
0.87255859375
],
[
0.1175537109375,
0.87451171875,
0.489501953125,
0.87451171875,
0.489501953125,
0.88818359375,
0.1175537109375,
0.88818359375
],
[
0.509765625,
0.87451171875,
0.88330078125,
0.87451171875,
0.88330078125,
0.88818359375,
0.509765625,
0.88818359375
],
[
0.1158447265625,
0.89111328125,
0.48779296875,
0.89111328125,
0.48779296875,
0.90478515625,
0.1158447265625,
0.90478515625
],
[
0.5087890625,
0.89111328125,
0.880859375,
0.89111328125,
0.880859375,
0.90478515625,
0.5087890625,
0.90478515625
],
[
0.1158447265625,
0.90771484375,
0.489501953125,
0.90771484375,
0.489501953125,
0.92138671875,
0.1158447265625,
0.92138671875
],
[
0.5087890625,
0.90673828125,
0.8017578125,
0.90625,
0.8017578125,
0.92041015625,
0.5087890625,
0.92138671875
]
] |
[
"False positive rate+|False negative rate+",
"Average error rate +",
"Model",
"Sub-task",
"Hate speech",
"gpt-3.5-instruct-turbo",
"17.98 ± 7.55",
"14.77 ± 6.73",
"16.38±5.34",
"gpt-3.5-turbo (ChatGPT 3.5)",
"6.82 ±4.37",
"20.00±6.38",
"13.48 ±3.91",
"gpt-4",
"20.00±8.57",
"4.44 ± 4.34",
"12.22± 4.74",
"34.38±3.76",
"40.08 ± 3.33",
"37.22 ± 2.21",
"gpt-3.5-turbo (ChatGPT 3.5)",
"Misinformation gpt-3.5-instruct-turbo",
"8.05±1.83",
"49.76 ±3.17",
"28.93±1.86",
"23.44 ±2.74",
"13.64 ± 1.42",
"3.89 ± 1.38",
"gpt-4",
"Table 2: Use-mention classification. False positive, false negative, and average error rates in detecting use (in",
"percentages). Models are sorted by the average error rate. The best performing model has a high false positive rate,",
"mistakingmentionforuse,whichcanleadtocensorshipofusefulcounterspeech.",
"Sub-task",
"Model",
"False positive rate+False negative rate+",
"Averageerror rate",
"Hate speech",
"24.44 ± 10.0",
"51.11±8.08",
"77.78 ± 7.22",
"toxigen hatebert",
"perspective (insult)",
"4.44 ± 3.89",
"61.11 ± 9.85",
"32.78±6.27",
"perspective (toxicity)",
"20.00± 7.51",
"36.67 ± 9.18",
"28.33±8.49",
"21.11 ± 7.51",
"33.33 ± 9.74",
"27.22 ± 9.18",
"22.22±8.22",
"perspective (identity attack)",
"17.78 ± 7.51",
"26.67±8.65",
"19.44 ± 8.20",
"roberta hate speech",
"gpt-3.5-instruct-turbo",
"25.56±8.62",
"13.33 ± 7.51",
"gpt-3.5-turbo (ChatGPT 3.5)",
"11.11 ± 6.14",
"22.22 ± 8.07",
"16.67±7.24",
"20.00 ±7.51",
"14.44 ± 7.09",
"8.89±5.00",
"gpt-4",
"Misinformation roberta fake news",
"97.93 ± 1.00",
"5.10 ±1.52",
"51.52±1.20",
"gpt-3.5-instruct-turbo",
"26.12±2.40",
"19.44±2.62",
"22.78±2.80",
"gpt-3.5-turbo (ChatGPT 3.5)",
"22.11 ± 3.02",
"13.85 ±2.74",
"17.98±3.06",
"10.21 ±2.00",
"8.02 ± 1.85",
"9.11 ± 1.93",
"Table 3: Downstream tasks (hate speech and misinformation detection). False positive, false negative, and",
"gpt-4",
"average error rates in detecting hate speech and misinformation (in percentages). Models are sorted by the average",
"error rate. Across models, substantial errors persist in false positive rate (classifying counterspeech mentions as",
"harmful).",
"hate speech and misinformation detection: Perspec-",
"are misclassified as use in 20.00% of hate speech",
"tive (Perspective, 2023), Toxigen (Hartvigsen et al.,",
"and 23.44% of misinformation counternarratives.",
"2022), RoBERTa (Liu et al., 2019), and RoBERTa",
"In these settings, mistaking mention for use leads",
"fake news (Ahmed et al., 2017).",
" to the consequential harm of censoring useful coun-",
"For prompting, we use default parameters (tem-",
"terspeech.",
"perature=1) and max output token length 1 (out-",
"putting either A or B). The classification prompt in-",
"4.2 Downstream content classification (H2)",
"cludestheinstructionandadefinitionoftheclasses",
"The prior section showed that state-of-the-art sys-",
"(for prompt variants and complete prompt text see",
"tems oftenfailattheuse-mentiondistinction.Here",
"Appendix, Section D)",
"we test the impact on hate speech and misinfor-",
"mation detection, two downstream classification",
"4Results",
"tasks where the use-mention distinction matters",
"(Table 3).",
"4.1",
"",
"Use-mentionclassification (H1)",
"How well do the models distinguish whether prob-",
"Performance on downstream tasks First, on the",
"lematic language is used or mentioned?Across",
"hate speech detection task, we find that misclassi-",
"the models (Table 2), average error rates are high",
"fication of counterspeech as hateful is relatively",
"(between 12.22% and 16.38% for hate speech and",
"frequent:thepopularToxigenandPerspectiveAPl",
"between 13.64% and 37.22% for misinformation),",
"have a false positive rate on counterspeech of over",
"suggesting that state-of-the-art large language mod-",
"20%, and recent models still have many false pos-",
"els struggle to distinguish use from mention in",
"itives, although somewhat reduced, e.g., gpt-3.5-",
"domains where the distinction matters. The best-",
" turbo has FPR 11.11%. Regarding the average er-",
"performing model, gpt-4, still has a very high false",
" ror rate, gpt-4 is the best-performing model (false",
"positive rate, mistaking mention for use; mentions",
"positive rate on counterspeech 8.89%)."
] |
{
"arxiv_id": "2404.01651",
"language": "en",
"page_idx": 4
}
|
||
2004.09739-en-0
|
2004.09739-en-0.jpg
|
Neural Abstractive Summarization with Structural Attention
Tanya Chowdhury$^1,\textbf{Sachin Kumar}^2$ and Tanmoy Chakraborty$^{1}$
$^{1}$IIIT-Delhi, India
$^{2}$Carnegie Mellon University, USA
$\{$tanya14109,tanmoy$\}@$iiitd.ac.in, sachink@andrew.cmu.edu
## Abstract
Attentional, RNN-based encoder-decoder architectures have achieved impressive performance on abstractive summarization of news articles. However, these methods fail to account for long term dependencies within the sentences of a document. This problem is exacerbated in multi-document summarization tasks such as summarizing the popular opinion in threads present in community question answering (CQA) websites such as Yahoo! Answers and Quora. These threads contain answers which often overlap or contradict each other. In this work, we present a hierarchical encoder based on structural attention to model such inter-sentence and inter-document dependencies. We set the popular pointer-generator architecture and some of the architectures derived from it as our baselines and show that they fail to generate good summaries in a multi-document setting. We further illustrate that our proposed model achieves significant improvement over the baselines in both single and multidocument summarization settings-in the former setting, it beats the best baseline by 1.31 and 7.8 ROUGE-1 points on CNN and CQA datasets, respectively; in the latter setting, the performance is further improved by 1.6 ROUGE-1 points on the CQA dataset.
### 1 Introduction
Sequence-to-sequence (seq2seq) architectures with attention have led to tremendous success in many conditional language generation tasks such as machine translation [Sutskever et $\tilde{a}l.,2014;$ Bahdanau $etal.,2015;$ Luong $etal.,2015]$, dialog generation [Vinyals and Le, 2015], abstractive summarization [Rush et al., 2015; See et al., 2017; Paulus $et$ $al.$, 2018] and question answering [Yin et al., 2016]. Attention models that compute a context vector based on the entire input text at every decoder step have especially benefited abstractive summarization. This is because - (1) LSTM encoders struggle to encode long documents into just one vector from which one can decode the summary, and (2) generating summary involves a lot of copying from the source text which attention models easily facilitate that can
also be made part of the training process [See et al., 2017;
Nallapati et al., 2016].
Studies in abstractive summarization majorly focus on generating news summaries, and the CNN/Dailymail dataset [Hermann et al., 2015; Nallapati et al., 2016] is largely used as a benchmark. This is a very clean dataset with fluent and properly structured text collected from a restricted domain, and it hardly features any repetitive information. Therefore, the current models which have been built on this dataset do not need to account too much for repeated or contradicting information. Additionally, these models are trained on truncated documents (up to 400 words), and increasing the word length leads to a drop in performance of these models [See et $al.,2017].$ This is because either LSTM is unable to encode longer documents accurately or in this dataset, the frst few sentences contain most of the summary information. However, real-world documents are usually much longer, and their summary information is spread across the entire document rather than in the first few sentences. In this paper, we explore one such dataset dealing with community question answering (CQA) summarization.
Multi-document summarization (MDS) is a well-studied problem with possible applications in summarizing tweets [Cao et al., 2016], news from varying sources[Yasunaga $et$ $al.,2017]$, reviews in e-commerce services [Ganesan et al., 2010], and so on. CQA services such as $\varrho uora,Yahoo!An-$ swers, Stack Overflow help in curating information which may often be difficult to obtain directly from the existing resources on the web. However, a majority of question threads on these services harbor unstructured, often repeating or contradictory answers. Additionally, the answers vary in length from being a few words to a thousand words long. A large number of users with very different writing styles contribute to form these knowledge bases, which results in a diverse and challenging corpora. We define the task of $CQA$ summarization as follows- given a list of such question threads, summarize the popular opinion reflected in them. This can be viewed as a multi-document summarization task[Chowdhury and Chakraborty, 2019].
We envision a solution to this problem by incorporating structural information present in language in the summarization models. Recent studies have shown promise in this direction. For example, [Fernandes et al., 2018] add graph neural networks (GNNs)[Li et al., 2015] to the seq2se en-
|
[
[
0.193603515625,
0.11236572265625,
0.8037109375,
0.11236572265625,
0.8037109375,
0.126953125,
0.193603515625,
0.126953125
],
[
0.22216796875,
0.1514892578125,
0.7685546875,
0.1514892578125,
0.7685546875,
0.169189453125,
0.22216796875,
0.169189453125
],
[
0.428955078125,
0.17236328125,
0.56884765625,
0.1749267578125,
0.56884765625,
0.1912841796875,
0.428955078125,
0.189453125
],
[
0.359375,
0.1881103515625,
0.640625,
0.1900634765625,
0.640625,
0.2108154296875,
0.359375,
0.208984375
],
[
0.261474609375,
0.212158203125,
0.74169921875,
0.212158203125,
0.74169921875,
0.2298583984375,
0.261474609375,
0.2298583984375
],
[
0.0228729248046875,
0.267578125,
0.05963134765625,
0.267578125,
0.05963134765625,
0.401611328125,
0.0228729248046875,
0.401611328125
],
[
0.2484130859375,
0.271484375,
0.326904296875,
0.271484375,
0.326904296875,
0.2890625,
0.2484130859375,
0.2890625
],
[
0.51318359375,
0.2734375,
0.91259765625,
0.2734375,
0.91259765625,
0.28857421875,
0.51318359375,
0.28857421875
],
[
0.51318359375,
0.28857421875,
0.6611328125,
0.28857421875,
0.6611328125,
0.3017578125,
0.51318359375,
0.3017578125
],
[
0.1192626953125,
0.29541015625,
0.451904296875,
0.29541015625,
0.451904296875,
0.309326171875,
0.1192626953125,
0.309326171875
],
[
0.52783203125,
0.3037109375,
0.91015625,
0.3037109375,
0.91015625,
0.317626953125,
0.52783203125,
0.317626953125
],
[
0.1192626953125,
0.311279296875,
0.449462890625,
0.311279296875,
0.449462890625,
0.3212890625,
0.1192626953125,
0.3212890625
],
[
0.51123046875,
0.317626953125,
0.91259765625,
0.317626953125,
0.91259765625,
0.33203125,
0.51123046875,
0.33203125
],
[
0.1192626953125,
0.324462890625,
0.453369140625,
0.324462890625,
0.453369140625,
0.33837890625,
0.1192626953125,
0.33837890625
],
[
0.51318359375,
0.33203125,
0.91259765625,
0.33203125,
0.91259765625,
0.34521484375,
0.51318359375,
0.34521484375
],
[
0.1192626953125,
0.33837890625,
0.453369140625,
0.33837890625,
0.453369140625,
0.3515625,
0.1192626953125,
0.3515625
],
[
0.51123046875,
0.34521484375,
0.91259765625,
0.34521484375,
0.91259765625,
0.359130859375,
0.51123046875,
0.359130859375
],
[
0.11767578125,
0.3515625,
0.453369140625,
0.3515625,
0.453369140625,
0.365478515625,
0.11767578125,
0.365478515625
],
[
0.11767578125,
0.365478515625,
0.451904296875,
0.365478515625,
0.451904296875,
0.37890625,
0.11767578125,
0.37890625
],
[
0.51318359375,
0.35986328125,
0.9111328125,
0.35986328125,
0.9111328125,
0.373779296875,
0.51318359375,
0.373779296875
],
[
0.51318359375,
0.373779296875,
0.91259765625,
0.373779296875,
0.91259765625,
0.386962890625,
0.51318359375,
0.386962890625
],
[
0.1192626953125,
0.380126953125,
0.453369140625,
0.380126953125,
0.453369140625,
0.393310546875,
0.1192626953125,
0.393310546875
],
[
0.1192626953125,
0.393310546875,
0.455078125,
0.393310546875,
0.455078125,
0.4072265625,
0.1192626953125,
0.4072265625
],
[
0.5146484375,
0.38818359375,
0.9111328125,
0.38818359375,
0.9111328125,
0.3984375,
0.5146484375,
0.3984375
],
[
0.0228729248046875,
0.404052734375,
0.060455322265625,
0.404052734375,
0.060455322265625,
0.4912109375,
0.0228729248046875,
0.4912109375
],
[
0.51318359375,
0.40087890625,
0.91259765625,
0.40087890625,
0.91259765625,
0.4140625,
0.51318359375,
0.4140625
],
[
0.1192626953125,
0.4072265625,
0.453369140625,
0.4072265625,
0.453369140625,
0.42041015625,
0.1192626953125,
0.42041015625
],
[
0.1192626953125,
0.42236328125,
0.451904296875,
0.42236328125,
0.451904296875,
0.43310546875,
0.1192626953125,
0.43310546875
],
[
0.51318359375,
0.416748046875,
0.90673828125,
0.416748046875,
0.90673828125,
0.4267578125,
0.51318359375,
0.4267578125
],
[
0.51123046875,
0.4287109375,
0.91259765625,
0.427978515625,
0.91259765625,
0.44140625,
0.51123046875,
0.442626953125
],
[
0.1192626953125,
0.43505859375,
0.453369140625,
0.43505859375,
0.453369140625,
0.448974609375,
0.1192626953125,
0.448974609375
],
[
0.51318359375,
0.442626953125,
0.9140625,
0.442626953125,
0.9140625,
0.455810546875,
0.51318359375,
0.455810546875
],
[
0.11767578125,
0.448974609375,
0.451904296875,
0.448974609375,
0.451904296875,
0.462158203125,
0.11767578125,
0.462158203125
],
[
0.51123046875,
0.455810546875,
0.9140625,
0.455078125,
0.9140625,
0.4697265625,
0.51123046875,
0.47021484375
],
[
0.1192626953125,
0.464111328125,
0.453369140625,
0.464111328125,
0.453369140625,
0.474853515625,
0.1192626953125,
0.474853515625
],
[
0.1192626953125,
0.4765625,
0.451904296875,
0.4765625,
0.451904296875,
0.490478515625,
0.1192626953125,
0.490478515625
],
[
0.5146484375,
0.4716796875,
0.9111328125,
0.4716796875,
0.9111328125,
0.481689453125,
0.5146484375,
0.481689453125
],
[
0.021240234375,
0.4892578125,
0.05963134765625,
0.4892578125,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.51318359375,
0.484130859375,
0.9111328125,
0.484130859375,
0.9111328125,
0.49755859375,
0.51318359375,
0.49755859375
],
[
0.12091064453125,
0.492431640625,
0.451904296875,
0.492431640625,
0.451904296875,
0.50244140625,
0.12091064453125,
0.50244140625
],
[
0.11767578125,
0.50390625,
0.453369140625,
0.50390625,
0.453369140625,
0.517578125,
0.11767578125,
0.517578125
],
[
0.51318359375,
0.5,
0.9111328125,
0.5,
0.9111328125,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.11767578125,
0.517578125,
0.453369140625,
0.517578125,
0.453369140625,
0.53076171875,
0.11767578125,
0.53076171875
],
[
0.51318359375,
0.51318359375,
0.91259765625,
0.51318359375,
0.91259765625,
0.5234375,
0.51318359375,
0.5234375
],
[
0.1168212890625,
0.53076171875,
0.451904296875,
0.53076171875,
0.451904296875,
0.54443359375,
0.1168212890625,
0.54443359375
],
[
0.51123046875,
0.52587890625,
0.9140625,
0.52587890625,
0.9140625,
0.5390625,
0.51123046875,
0.5390625
],
[
0.11767578125,
0.54541015625,
0.451904296875,
0.54443359375,
0.451904296875,
0.55810546875,
0.11767578125,
0.5595703125
],
[
0.51318359375,
0.54052734375,
0.91259765625,
0.54052734375,
0.91259765625,
0.5537109375,
0.51318359375,
0.5537109375
],
[
0.1192626953125,
0.5595703125,
0.451904296875,
0.5595703125,
0.451904296875,
0.57275390625,
0.1192626953125,
0.57275390625
],
[
0.5146484375,
0.5546875,
0.6650390625,
0.5546875,
0.6650390625,
0.56494140625,
0.5146484375,
0.56494140625
],
[
0.1192626953125,
0.57373046875,
0.453369140625,
0.57373046875,
0.453369140625,
0.5869140625,
0.1192626953125,
0.5869140625
],
[
0.52783203125,
0.568359375,
0.91259765625,
0.568359375,
0.91259765625,
0.58203125,
0.52783203125,
0.58203125
],
[
0.1192626953125,
0.5869140625,
0.450927734375,
0.5869140625,
0.450927734375,
0.59765625,
0.1192626953125,
0.59765625
],
[
0.51123046875,
0.58203125,
0.9111328125,
0.58203125,
0.9111328125,
0.59521484375,
0.51123046875,
0.59521484375
],
[
0.1192626953125,
0.60107421875,
0.451904296875,
0.60107421875,
0.451904296875,
0.6142578125,
0.1192626953125,
0.6142578125
],
[
0.5146484375,
0.5966796875,
0.9140625,
0.5966796875,
0.9140625,
0.611328125,
0.5146484375,
0.611328125
],
[
0.1192626953125,
0.61572265625,
0.453369140625,
0.61572265625,
0.453369140625,
0.62890625,
0.1192626953125,
0.62890625
],
[
0.51318359375,
0.60986328125,
0.9111328125,
0.60986328125,
0.9111328125,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.51123046875,
0.62255859375,
0.9111328125,
0.62353515625,
0.9111328125,
0.63818359375,
0.51123046875,
0.63720703125
],
[
0.11767578125,
0.62890625,
0.455078125,
0.62890625,
0.455078125,
0.642578125,
0.11767578125,
0.642578125
],
[
0.51123046875,
0.63720703125,
0.91259765625,
0.63720703125,
0.91259765625,
0.65087890625,
0.51123046875,
0.65087890625
],
[
0.12091064453125,
0.64306640625,
0.2091064453125,
0.64306640625,
0.2091064453125,
0.65478515625,
0.12091064453125,
0.65478515625
],
[
0.51318359375,
0.6513671875,
0.91259765625,
0.6513671875,
0.91259765625,
0.66552734375,
0.51318359375,
0.66552734375
],
[
0.51123046875,
0.66650390625,
0.91259765625,
0.66650390625,
0.91259765625,
0.6796875,
0.51123046875,
0.6796875
],
[
0.11273193359375,
0.6748046875,
0.22216796875,
0.6748046875,
0.22216796875,
0.689453125,
0.11273193359375,
0.689453125
],
[
0.51123046875,
0.67822265625,
0.9111328125,
0.6796875,
0.9111328125,
0.693359375,
0.51123046875,
0.69140625
],
[
0.0882568359375,
0.69580078125,
0.48779296875,
0.69580078125,
0.48779296875,
0.708984375,
0.0882568359375,
0.708984375
],
[
0.51318359375,
0.693359375,
0.91259765625,
0.693359375,
0.91259765625,
0.70703125,
0.51318359375,
0.70703125
],
[
0.0882568359375,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.0882568359375,
0.7236328125
],
[
0.51318359375,
0.70849609375,
0.91259765625,
0.70849609375,
0.91259765625,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.0882568359375,
0.7236328125,
0.48779296875,
0.7236328125,
0.48779296875,
0.7373046875,
0.0882568359375,
0.7373046875
],
[
0.51318359375,
0.7216796875,
0.91259765625,
0.7216796875,
0.91259765625,
0.73486328125,
0.51318359375,
0.73486328125
],
[
0.0849609375,
0.7373046875,
0.486083984375,
0.7373046875,
0.486083984375,
0.75048828125,
0.0849609375,
0.75048828125
],
[
0.51318359375,
0.73486328125,
0.91259765625,
0.73486328125,
0.91259765625,
0.74853515625,
0.51318359375,
0.74853515625
],
[
0.0849609375,
0.75048828125,
0.486083984375,
0.75,
0.486083984375,
0.7646484375,
0.0849609375,
0.76513671875
],
[
0.51318359375,
0.75,
0.9111328125,
0.75,
0.9111328125,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.0849609375,
0.76513671875,
0.48779296875,
0.76513671875,
0.48779296875,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.51318359375,
0.7646484375,
0.91015625,
0.7646484375,
0.91015625,
0.7744140625,
0.51318359375,
0.7744140625
],
[
0.08660888671875,
0.77978515625,
0.484375,
0.77978515625,
0.484375,
0.79052734375,
0.08660888671875,
0.79052734375
],
[
0.51123046875,
0.77587890625,
0.9111328125,
0.77587890625,
0.9111328125,
0.7890625,
0.51123046875,
0.7890625
],
[
0.0849609375,
0.79248046875,
0.486083984375,
0.79248046875,
0.486083984375,
0.80615234375,
0.0849609375,
0.80615234375
],
[
0.51318359375,
0.79052734375,
0.9111328125,
0.79052734375,
0.9111328125,
0.8037109375,
0.51318359375,
0.8037109375
],
[
0.08660888671875,
0.80615234375,
0.486083984375,
0.80615234375,
0.486083984375,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.5107421875,
0.8037109375,
0.67236328125,
0.80322265625,
0.67236328125,
0.8173828125,
0.5107421875,
0.818359375
],
[
0.0882568359375,
0.8212890625,
0.316162109375,
0.8212890625,
0.316162109375,
0.83203125,
0.0882568359375,
0.83203125
],
[
0.32275390625,
0.8212890625,
0.484375,
0.8212890625,
0.484375,
0.83203125,
0.32275390625,
0.83203125
],
[
0.52783203125,
0.818359375,
0.91259765625,
0.82080078125,
0.91259765625,
0.833984375,
0.52783203125,
0.83203125
],
[
0.318603515625,
0.82470703125,
0.325927734375,
0.82470703125,
0.325927734375,
0.8291015625,
0.318603515625,
0.8291015625
],
[
0.08660888671875,
0.833984375,
0.486083984375,
0.833984375,
0.486083984375,
0.84765625,
0.08660888671875,
0.84765625
],
[
0.51123046875,
0.833984375,
0.91015625,
0.833984375,
0.91015625,
0.84765625,
0.51123046875,
0.84765625
],
[
0.08660888671875,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.51123046875,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.51123046875,
0.861328125
],
[
0.08660888671875,
0.8623046875,
0.488525390625,
0.8623046875,
0.488525390625,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.5087890625,
0.85986328125,
0.91259765625,
0.861328125,
0.91259765625,
0.87548828125,
0.5087890625,
0.875
],
[
0.0849609375,
0.87548828125,
0.48779296875,
0.87548828125,
0.48779296875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51123046875,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"Neural Abstractive Summarization with Structural Attention",
"Tanya Chowdhury', Sachin Kumar? and Tanmoy Chakraborty",
"1IIIT-Delhi, India",
"2Carnegie Mellon University, USA",
"[ tanya14109,tanmoy @iitd.ac.in, sachink@andrew.cmu.edu",
"2210",
"Abstract",
"also be made part of the training process [See et al., 2017;",
"Nallapati et al., 2016].",
"Attentional, RNN-based encoder-decoder architec-",
"Studies in abstractive summarization majorly focus on gen-",
"tureshaveachievedimpressiveperformanceonab",
"erating news summaries, and the CNN/Dailymail dataset",
"stractive summarization of news articles. However,",
"[Hermann et al., 2015; Nallapati et al., 2016] is largely used",
"these methods fail to account for long term depen-",
"as a benchmark. This is a very clean dataset with fluent and",
"dencies within the sentences of a document. This",
"problem is exacerbated in multi-document sum-",
"properly structured text collected from a restricted domain,",
"and it hardly features any repetitive information. Therefore,",
"marization tasks such as summarizing the popular",
"opinion in threads present in community question",
"the currentmodels which havebeen built onthis dataset do",
"['s] ",
"not need to account too much for repeated or contradicting",
"answering (CQA) websites such as Yahoo! An-",
"swersand Quora.Thesethreadscontainanswers",
"information.Additionally,thesemodels aretrained on trun",
"cated documents (up to 400 words), and increasing the word",
"which often overlap or contradict each other.In",
"length leads to a drop in performance of these models [See et",
"this work, we present a hierarchical encoder based",
"al., 2017]. This is because either LSTM is unable to encode",
"onstructuralattentiontomodel suchinter-sentence",
"and inter-document dependencies. We set the pop-",
"longerdocumentsaccuratelyorinthisdataset.thefirstfew",
"6060002:1",
"sentences contain most of the summary information. How-",
"ularpointer-generatorarchitectureandsomeofthe",
"architectures derived from it as our baselines and",
"ever,real-worlddocumentsareusuallymuchlonger,andtheir",
"show that they fail to generate good summaries in",
"summaryinformationis spread acrossthe entire document",
"a multi-document setting. We further illustrate that",
"rather than in the first few sentences. In this paper, we explore",
"our proposed model achieves significant improve-",
"one such dataset dealing with community question answering",
"ment over the baselines in both single and multi-",
"(CQA)summarization.",
"document summarization settings -- in the former",
"Multi-document summarization (MDS) is a well-studied",
"setting. it beats the best baseline by 131 and 7.8",
"problem with possible applications in summarizing tweets",
"ROUGE-1 points on CNN and CQA datasets, re-",
"[Cao et al., 20i6], news from varying sources [Yasunaga et",
"spectively; in the latter setting, the performance is",
"al., 2017], reviews in e-commerce services [Ganesan et al.,",
"201ol, and so on. CQA services such as Quora, Yahoo! An-",
"further improved by 1.6 ROUGE-1 points on the",
"swers, Stack Overflow help in curating information which",
"CQA dataset.",
"may often be difficult to obtain directly from the existing re-",
" sources on the web. However, a majority of question threads",
"Introduction",
"on these services harbor unstructured, often repeating or con-",
"Sequence-to-sequence (seq2seg) architectures with attention",
"tradictory answers. Additionally, the answers vary in length",
"have led to tremendous success in many conditional language",
"from being a few words to a thousand words long. A large",
"generation tasks such as machine translation [Sutskever et",
"number of users with very different writing styles contribute",
"al., 2014; Bahdanau et al., 2015; Luong et al., 2015], di-",
"to form these knowledge bases, which results in a diverse and",
"alog generation [Vinyals and Le, 2015], abstractive sum-",
"challenging corpora. We define the task of CQA summariza-",
"marization [Rush et al., 2015; See et al., 2017; Paulus et",
"tion asfollows-given alistof such questionthreads,sum-",
"al.,2018] and question answering [Yin et al., 2016].At-",
"marize the popular opinion reflected in them. This can be",
"tention models that compute a context vector based on the",
"viewed as a multi-document summarization task [Chowdhury",
"entire input text at every decoder step have especially ben-",
"and Chakraborty, 2019].",
"efited abstractive summarization.",
"This is because _ (1)",
"Weenvision a solution to this problemby incorporating",
"",
"LSTM encoders struggle to encode long documents into just",
" structural information present in language in the summariza-",
"one vector from which one can decode the summary, and",
"tion models. Recent studies have shown promise in this di-",
"(2) generating summary involves a lot of copying from the",
"rection. For example, [Fernandes et al., 2018] add graph",
"source text which attention models easily facilitate that can",
"neural networks (GNNs) [Li et al., 2015] to the seq2se en-"
] |
{
"arxiv_id": "2004.09739",
"language": "en",
"page_idx": 0
}
|
||
2004.09739-en-1
|
2004.09739-en-1.jpg
|
coder model for neural abstractive summarization.[Liu et al., 2019] model documents as multi-root dependency trees, each node representing a sentence; and pick the induced tree roots to be summary sentences in an extractive setting. Recently, [Song et al., 2018] attempt to preserve important structural dependency relations obtained from syntactic parse trees of the input text in the summary by using a structure infused copy mechanism. However, obtaining labeled data to explicitly model such dependency relations is expensive. Here, we attempt to mitigate this issue by proposing a structural encoder for summarization based on the prior work on structure attention networks [Liu and Lapata, 2018] which implicitly incorporates structural information within end-to-end train- $ing.$
Our major contributions in this work are three-fold:
1. We enhance the pointer-generator architecture by adding
a structural attention based encoder to implicitly capture long term dependency relations in summarization of lengthy documents.
2. We further propose a hierarchical encoder with multi-
level structural attention to capture document-level discourse information in the multi-document summarization task.
3.We introduce multi-level contextual attention in the
structural attention setting to enable word level copying and to generate more abstractive summaries, compared to similar architectures.
We compare our proposed solution against the popular pointer-generator model [See et al., 2017] and a few other summarization models derived from it in both single and multi-document settings. We show that our structural encoder architecture beats the strong pointer-generator baseline by 1.31 ROUGE-1 points on the CNN/Dailymail dataset and by 7.8 ROUGE-1 points on the concatenated CQA dataset for single document summarization (SDS). Our hierarchical structural encoder architecture further beats the concatenated approach by another 1.6 ROUGE-1 points on the CQA dataset. A qualitative analysis of the generated summaries observe considerable qualitative gains after inclusion of structural attention. Our structural attention based summarinzation model is one of the few abstractive approaches to beat extractive baselines in MDS. The code is public at https: //bit.ly/35i7q93.
# 2 $\mathbf{Models}$
In this section, we first describe the seq2seq based pointergenerator architecture proposed by [See et al., 2017; Nallapati et al., 2016]. We then elaborate on how we incorporate structural attention into the mechanism and generate nonprojective dependency trees to capture long term dependencies within documents. Following this, we describe our hierarchical encoder architecture and propose multi-level contextual attention to better model discourse for MDS.
## 2.1 Pointer-Generator Model The pointer-generator architecture (PG)[See et al., 2017]
serves as the starting point for our approach. PG is inspired
from the LSTM encoder-decoder architecture[Nallapati $et$ $al. $, 2016] for the task of abstractive summarization. Tokens are fed to a bidirectional encoder to generate hidden representations $h$, and a unidirectional decoder is used to generate the summary. At the decoding step $t$, the decoder uses an attention mechanism [Bahdanau et al., 2015] to compute an attention vector $a^t$ for every encoder hidden state. These attention vectors are pooled to compute a context vector $c_t$ for every decoder step. The context vector $c_t$ at that step along with the hidden representations $h$ are passed through a feedforward network followed by softmax to give a multinomial distribution over the vocabulary $p_\mathrm{vocab}(w).$
Additionally, [See et al., 2017] introduce a copy mechanism [Vinyals et $al.,2015]$ in this model. At each decoding step $t$, it predicts whether to generate a new word or copy the $\hat{\text{word based on the attention probabilities }a^t}$ using a probability value $p_\mathrm{gen}.$ This probability is a function on the decoder $\operatorname*{input}x_{t}$,context vector $c_t$,and decoder state $s_t\cdot p_{gen}$ is used as a soft switch to choose between copying and generation as shown below:
$$p(w)=p_{\mathrm{gen}}p_{\mathrm{vocab}}(w)+(1-p_{\mathrm{gen}})\sum_{i:w_{i}=w}a_{i}^{t}$$
This model is trained by minimizing the negative loglikelihood of the predicted sequence normalized by the se- $\begin{array}{l}{\mathrm{quence~length:~}}{\dot{L}=-\frac{1}{T}\sum_{t=0}^{T^{*}}\log p(w_{t}^{*}),\mathrm{~where~}T\mathrm{~is~the}}\\{\mathrm{length~of~the~generated~summary.~Once~}L\mathrm{~gets~converged,~an}}\end{array}$ additional coverage loss is used to train the model further, which aims to minimize repetitions in the generated summaries. We refer the readers to [See et al., 2017] for more details.
## 2.2 Structural Attention
We observe that while the pointer-generator architecture performs well on the news corpora, it fails to generate meaningful summaries on complex datasets. We hypothesize that taking into account intra-document discourse information will help in generating better summaries. We propose an architecture that implicitly incorporates such information using document structural representation [Liu and Lapata, 2018; Kim et al., 2017] to include richer structural dependencies in the end-to-end training.
# Proposed Model We model our document as a non-projective dependency
parse tree by constraining inter-token attention as weights of the dependency tree. We use the Matrix tree theorem [Koo et $al.,2\hat{007};$Tutte, 1984] to carry out the same.
As shown in Figure 1, we feed our input tokens $(w_i)$ to a bi-LSTM encoder to obtain hidden state representations $h_i.$ We decompose the hidden state vector into two parts: $d_i$ and $e_i$, which we call the structural part and the semantic part, respectively:
$$[e_i,d_i]=h_i$$
(1)
For every pair of two input tokens, we transform their structural parts $d$ and try to compute the probability of a parent-child relationship edge between them in the dependency tree. For tokens $j$ and $k$,this is done as:
$$u_j=\tanh(W_pd_j);\quad u_k=\tanh(W_cd_k)$$
|
[
[
0.0882568359375,
0.0694580078125,
0.484375,
0.0694580078125,
0.484375,
0.08331298828125,
0.0882568359375,
0.08331298828125
],
[
0.51123046875,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.51123046875,
0.08331298828125
],
[
0.0849609375,
0.08209228515625,
0.486083984375,
0.08331298828125,
0.486083984375,
0.09783935546875,
0.0849609375,
0.09661865234375
],
[
0.5107421875,
0.08209228515625,
0.91259765625,
0.08331298828125,
0.91259765625,
0.09783935546875,
0.5107421875,
0.09661865234375
],
[
0.08660888671875,
0.09783935546875,
0.48779296875,
0.09783935546875,
0.48779296875,
0.111083984375,
0.08660888671875,
0.111083984375
],
[
0.51123046875,
0.09783935546875,
0.9111328125,
0.09783935546875,
0.9111328125,
0.111083984375,
0.51123046875,
0.111083984375
],
[
0.0849609375,
0.11236572265625,
0.486083984375,
0.11236572265625,
0.486083984375,
0.1256103515625,
0.0849609375,
0.1256103515625
],
[
0.51318359375,
0.11297607421875,
0.9111328125,
0.11297607421875,
0.9111328125,
0.12371826171875,
0.51318359375,
0.12371826171875
],
[
0.08660888671875,
0.125,
0.486083984375,
0.125,
0.486083984375,
0.1395263671875,
0.08660888671875,
0.1395263671875
],
[
0.5107421875,
0.1256103515625,
0.9111328125,
0.1256103515625,
0.9111328125,
0.1395263671875,
0.5107421875,
0.1395263671875
],
[
0.0849609375,
0.1383056640625,
0.486083984375,
0.1383056640625,
0.486083984375,
0.1514892578125,
0.0849609375,
0.1514892578125
],
[
0.51123046875,
0.1395263671875,
0.91259765625,
0.1395263671875,
0.91259765625,
0.15283203125,
0.51123046875,
0.15283203125
],
[
0.0849609375,
0.15283203125,
0.48779296875,
0.15283203125,
0.48779296875,
0.1666259765625,
0.0849609375,
0.1666259765625
],
[
0.51318359375,
0.15283203125,
0.9111328125,
0.15283203125,
0.9111328125,
0.1666259765625,
0.51318359375,
0.1666259765625
],
[
0.08660888671875,
0.1685791015625,
0.484375,
0.1685791015625,
0.484375,
0.1787109375,
0.08660888671875,
0.1787109375
],
[
0.51318359375,
0.1685791015625,
0.91015625,
0.1685791015625,
0.91015625,
0.1787109375,
0.51318359375,
0.1787109375
],
[
0.0849609375,
0.18115234375,
0.48779296875,
0.18115234375,
0.48779296875,
0.1944580078125,
0.0849609375,
0.1944580078125
],
[
0.5107421875,
0.179931640625,
0.9111328125,
0.179931640625,
0.9111328125,
0.1932373046875,
0.5107421875,
0.1932373046875
],
[
0.0849609375,
0.1944580078125,
0.484375,
0.1944580078125,
0.484375,
0.2083740234375,
0.0849609375,
0.2083740234375
],
[
0.51318359375,
0.1944580078125,
0.9111328125,
0.1944580078125,
0.9111328125,
0.2083740234375,
0.51318359375,
0.2083740234375
],
[
0.08660888671875,
0.210205078125,
0.484375,
0.210205078125,
0.484375,
0.2203369140625,
0.08660888671875,
0.2203369140625
],
[
0.51318359375,
0.2083740234375,
0.91259765625,
0.2083740234375,
0.91259765625,
0.2215576171875,
0.51318359375,
0.2215576171875
],
[
0.08660888671875,
0.2215576171875,
0.486083984375,
0.2215576171875,
0.486083984375,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.51123046875,
0.2203369140625,
0.79736328125,
0.222900390625,
0.79736328125,
0.2374267578125,
0.51123046875,
0.23486328125
],
[
0.0849609375,
0.236083984375,
0.484375,
0.236083984375,
0.484375,
0.25,
0.0849609375,
0.25
],
[
0.53125,
0.2374267578125,
0.90869140625,
0.2374267578125,
0.90869140625,
0.2474365234375,
0.53125,
0.2474365234375
],
[
0.0849609375,
0.251953125,
0.11273193359375,
0.251953125,
0.11273193359375,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.51318359375,
0.25,
0.91259765625,
0.25,
0.91259765625,
0.26318359375,
0.51318359375,
0.26318359375
],
[
0.101318359375,
0.264404296875,
0.441162109375,
0.264404296875,
0.441162109375,
0.27783203125,
0.101318359375,
0.27783203125
],
[
0.51123046875,
0.264404296875,
0.91259765625,
0.264404296875,
0.91259765625,
0.27783203125,
0.51123046875,
0.27783203125
],
[
0.09722900390625,
0.2822265625,
0.484375,
0.28271484375,
0.484375,
0.296630859375,
0.09722900390625,
0.29541015625
],
[
0.5146484375,
0.27783203125,
0.91015625,
0.27783203125,
0.91015625,
0.291748046875,
0.5146484375,
0.291748046875
],
[
0.5107421875,
0.290283203125,
0.91259765625,
0.28857421875,
0.91259765625,
0.30615234375,
0.5107421875,
0.30810546875
],
[
0.1168212890625,
0.296630859375,
0.484375,
0.296630859375,
0.484375,
0.31005859375,
0.1168212890625,
0.31005859375
],
[
0.11767578125,
0.311279296875,
0.486083984375,
0.311279296875,
0.486083984375,
0.324462890625,
0.11767578125,
0.324462890625
],
[
0.51318359375,
0.30615234375,
0.9140625,
0.30615234375,
0.9140625,
0.32080078125,
0.51318359375,
0.32080078125
],
[
0.5107421875,
0.317626953125,
0.91259765625,
0.318115234375,
0.91259765625,
0.333251953125,
0.5107421875,
0.33203125
],
[
0.11767578125,
0.323974609375,
0.26708984375,
0.323974609375,
0.26708984375,
0.337158203125,
0.11767578125,
0.337158203125
],
[
0.51318359375,
0.333984375,
0.60546875,
0.333984375,
0.60546875,
0.34521484375,
0.51318359375,
0.34521484375
],
[
0.096435546875,
0.341552734375,
0.484375,
0.34228515625,
0.484375,
0.356689453125,
0.096435546875,
0.35595703125
],
[
0.5595703125,
0.348388671875,
0.865234375,
0.348388671875,
0.865234375,
0.3662109375,
0.5595703125,
0.3662109375
],
[
0.11767578125,
0.356689453125,
0.486083984375,
0.356689453125,
0.486083984375,
0.37060546875,
0.11767578125,
0.37060546875
],
[
0.11767578125,
0.371826171875,
0.486083984375,
0.371826171875,
0.486083984375,
0.385009765625,
0.11767578125,
0.385009765625
],
[
0.80615234375,
0.371826171875,
0.8203125,
0.371826171875,
0.8203125,
0.375,
0.80615234375,
0.375
],
[
0.52783203125,
0.380126953125,
0.91015625,
0.380615234375,
0.91015625,
0.395751953125,
0.52783203125,
0.39453125
],
[
0.1192626953125,
0.386474609375,
0.181396484375,
0.386474609375,
0.181396484375,
0.397705078125,
0.1192626953125,
0.397705078125
],
[
0.5146484375,
0.395751953125,
0.9111328125,
0.395751953125,
0.9111328125,
0.4091796875,
0.5146484375,
0.4091796875
],
[
0.09722900390625,
0.40283203125,
0.486083984375,
0.40283203125,
0.486083984375,
0.416748046875,
0.09722900390625,
0.416748046875
],
[
0.51123046875,
0.410400390625,
0.9140625,
0.410400390625,
0.9140625,
0.427978515625,
0.51123046875,
0.427978515625
],
[
0.11767578125,
0.417236328125,
0.484375,
0.417236328125,
0.484375,
0.431884765625,
0.11767578125,
0.431884765625
],
[
0.51123046875,
0.425537109375,
0.91259765625,
0.425537109375,
0.91259765625,
0.439453125,
0.51123046875,
0.439453125
],
[
0.11767578125,
0.431884765625,
0.486083984375,
0.431884765625,
0.486083984375,
0.44580078125,
0.11767578125,
0.44580078125
],
[
0.51318359375,
0.439453125,
0.91259765625,
0.439453125,
0.91259765625,
0.45263671875,
0.51318359375,
0.45263671875
],
[
0.1192626953125,
0.447509765625,
0.273681640625,
0.447509765625,
0.273681640625,
0.458251953125,
0.1192626953125,
0.458251953125
],
[
0.5146484375,
0.455078125,
0.91015625,
0.455078125,
0.91015625,
0.46533203125,
0.5146484375,
0.46533203125
],
[
0.10296630859375,
0.464111328125,
0.486083984375,
0.464111328125,
0.486083984375,
0.477783203125,
0.10296630859375,
0.477783203125
],
[
0.51123046875,
0.46728515625,
0.91259765625,
0.46728515625,
0.91259765625,
0.48095703125,
0.51123046875,
0.48095703125
],
[
0.08660888671875,
0.477783203125,
0.486083984375,
0.477783203125,
0.486083984375,
0.4912109375,
0.08660888671875,
0.4912109375
],
[
0.51318359375,
0.481689453125,
0.56201171875,
0.481689453125,
0.56201171875,
0.49365234375,
0.51318359375,
0.49365234375
],
[
0.08660888671875,
0.492431640625,
0.486083984375,
0.492431640625,
0.486083984375,
0.505859375,
0.08660888671875,
0.505859375
],
[
0.08660888671875,
0.505859375,
0.486083984375,
0.505859375,
0.486083984375,
0.51953125,
0.08660888671875,
0.51953125
],
[
0.51123046875,
0.501953125,
0.7099609375,
0.50244140625,
0.7099609375,
0.517578125,
0.51123046875,
0.5166015625
],
[
0.0882568359375,
0.521484375,
0.486083984375,
0.521484375,
0.486083984375,
0.5322265625,
0.0882568359375,
0.5322265625
],
[
0.51318359375,
0.52099609375,
0.9111328125,
0.52099609375,
0.9111328125,
0.5341796875,
0.51318359375,
0.5341796875
],
[
0.0849609375,
0.5322265625,
0.486083984375,
0.5322265625,
0.486083984375,
0.54541015625,
0.0849609375,
0.54541015625
],
[
0.51318359375,
0.53515625,
0.91015625,
0.53515625,
0.91015625,
0.548828125,
0.51318359375,
0.548828125
],
[
0.0849609375,
0.54541015625,
0.488525390625,
0.546875,
0.488525390625,
0.56103515625,
0.0849609375,
0.56005859375
],
[
0.51318359375,
0.548828125,
0.91015625,
0.548828125,
0.91015625,
0.5625,
0.51318359375,
0.5625
],
[
0.0882568359375,
0.5625,
0.484375,
0.5625,
0.484375,
0.57275390625,
0.0882568359375,
0.57275390625
],
[
0.5146484375,
0.564453125,
0.9111328125,
0.564453125,
0.9111328125,
0.57470703125,
0.5146484375,
0.57470703125
],
[
0.08660888671875,
0.5771484375,
0.484375,
0.5771484375,
0.484375,
0.5869140625,
0.08660888671875,
0.5869140625
],
[
0.51318359375,
0.5771484375,
0.9111328125,
0.5771484375,
0.9111328125,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.08660888671875,
0.5888671875,
0.486083984375,
0.5888671875,
0.486083984375,
0.60302734375,
0.08660888671875,
0.60302734375
],
[
0.51318359375,
0.59033203125,
0.91259765625,
0.59033203125,
0.91259765625,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.0849609375,
0.60302734375,
0.486083984375,
0.60302734375,
0.486083984375,
0.6162109375,
0.0849609375,
0.6162109375
],
[
0.51123046875,
0.60400390625,
0.9140625,
0.60302734375,
0.9140625,
0.6162109375,
0.51123046875,
0.6171875
],
[
0.08660888671875,
0.6171875,
0.486083984375,
0.6171875,
0.486083984375,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.51318359375,
0.6171875,
0.91259765625,
0.6171875,
0.91259765625,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.08660888671875,
0.630859375,
0.484375,
0.630859375,
0.484375,
0.64453125,
0.08660888671875,
0.64453125
],
[
0.51123046875,
0.630859375,
0.68310546875,
0.630859375,
0.68310546875,
0.64453125,
0.51123046875,
0.64453125
],
[
0.0849609375,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.65771484375,
0.0849609375,
0.65771484375
],
[
0.5146484375,
0.6513671875,
0.62841796875,
0.6513671875,
0.62841796875,
0.66552734375,
0.5146484375,
0.66552734375
],
[
0.08660888671875,
0.6591796875,
0.486083984375,
0.6591796875,
0.486083984375,
0.67236328125,
0.08660888671875,
0.67236328125
],
[
0.51318359375,
0.6640625,
0.9111328125,
0.66552734375,
0.9111328125,
0.6796875,
0.51318359375,
0.6787109375
],
[
0.08416748046875,
0.67041015625,
0.193603515625,
0.671875,
0.193603515625,
0.68603515625,
0.08416748046875,
0.68505859375
],
[
0.51318359375,
0.6796875,
0.9140625,
0.6796875,
0.9140625,
0.693359375,
0.51318359375,
0.693359375
],
[
0.08251953125,
0.69873046875,
0.181396484375,
0.69873046875,
0.181396484375,
0.71630859375,
0.08251953125,
0.71630859375
],
[
0.51318359375,
0.693359375,
0.9140625,
0.693359375,
0.9140625,
0.70703125,
0.51318359375,
0.70703125
],
[
0.51123046875,
0.70556640625,
0.80615234375,
0.70703125,
0.80615234375,
0.7216796875,
0.51123046875,
0.72021484375
],
[
0.0849609375,
0.72021484375,
0.486083984375,
0.72021484375,
0.486083984375,
0.734375,
0.0849609375,
0.734375
],
[
0.52783203125,
0.72021484375,
0.91259765625,
0.72021484375,
0.91259765625,
0.734375,
0.52783203125,
0.734375
],
[
0.08660888671875,
0.73486328125,
0.484375,
0.73486328125,
0.484375,
0.74853515625,
0.08660888671875,
0.74853515625
],
[
0.51318359375,
0.73486328125,
0.9111328125,
0.73486328125,
0.9111328125,
0.74853515625,
0.51318359375,
0.74853515625
],
[
0.08660888671875,
0.74755859375,
0.486083984375,
0.74755859375,
0.486083984375,
0.76220703125,
0.08660888671875,
0.76220703125
],
[
0.51318359375,
0.74853515625,
0.9111328125,
0.74853515625,
0.9111328125,
0.76220703125,
0.51318359375,
0.76220703125
],
[
0.0849609375,
0.76318359375,
0.488525390625,
0.76318359375,
0.488525390625,
0.7763671875,
0.0849609375,
0.7763671875
],
[
0.5107421875,
0.76318359375,
0.91259765625,
0.76220703125,
0.91259765625,
0.7763671875,
0.5107421875,
0.77783203125
],
[
0.0882568359375,
0.77880859375,
0.484375,
0.77880859375,
0.484375,
0.7890625,
0.0882568359375,
0.7890625
],
[
0.5146484375,
0.77783203125,
0.59716796875,
0.77783203125,
0.59716796875,
0.7890625,
0.5146484375,
0.7890625
],
[
0.0849609375,
0.79150390625,
0.484375,
0.79150390625,
0.484375,
0.8046875,
0.0849609375,
0.8046875
],
[
0.6708984375,
0.79443359375,
0.75341796875,
0.79443359375,
0.75341796875,
0.8095703125,
0.6708984375,
0.8095703125
],
[
0.89208984375,
0.79443359375,
0.91259765625,
0.79443359375,
0.91259765625,
0.80810546875,
0.89208984375,
0.80810546875
],
[
0.08660888671875,
0.8046875,
0.486083984375,
0.8046875,
0.486083984375,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.0849609375,
0.818359375,
0.408447265625,
0.818359375,
0.408447265625,
0.83203125,
0.0849609375,
0.83203125
],
[
0.52783203125,
0.8125,
0.91259765625,
0.8125,
0.91259765625,
0.8271484375,
0.52783203125,
0.8271484375
],
[
0.51318359375,
0.82568359375,
0.9140625,
0.82568359375,
0.9140625,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.0882568359375,
0.84228515625,
0.32275390625,
0.84228515625,
0.32275390625,
0.85693359375,
0.0882568359375,
0.85693359375
],
[
0.51318359375,
0.84033203125,
0.9111328125,
0.84033203125,
0.9111328125,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.51318359375,
0.85400390625,
0.81884765625,
0.85400390625,
0.81884765625,
0.8671875,
0.51318359375,
0.8671875
],
[
0.0882568359375,
0.861328125,
0.484375,
0.861328125,
0.484375,
0.875,
0.0882568359375,
0.875
],
[
0.0849609375,
0.87548828125,
0.486083984375,
0.87548828125,
0.486083984375,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.57861328125,
0.87060546875,
0.84619140625,
0.87060546875,
0.84619140625,
0.88525390625,
0.57861328125,
0.88525390625
]
] |
[
"coder model for neural abstractive summarization. [Liu et al.,",
"from the LSTM encoder-decoder architecture [Nallapati et",
"2019] model documents as multi-root dependency trees, each",
"al., 2016l for the task of abstractive summarization. Tokens",
"node representing a sentence; and pick the induced tree roots",
"are fed to a bidirectional encoder to generate hidden repre-",
"to be summary sentences in an extractive setting. Recently,",
"sentations h, and a unidirectional decoderis used to generate",
"[Song et al., 2018] attempt to preserve important structural",
"the summary. At the decoding step t, the decoder uses an",
"dependency relations obtained from syntactic parse trees of",
"attention mechanism [Bahdanau et al., 2015] to compute an",
"the input text in the summary by using a structure infused",
"attention vector at for every encoder hidden state. These at-",
"copymechanism.However,obtaininglabeleddatatoexplic-",
"tentionvectorsarepooledtocomputeacontextvectorCtfor",
"itly model such dependency relations is expensive. Here, we",
"every decoder step. The context vector ct at that step along",
"attempt to mitigate this issue by proposing a structural en-",
"with the hidden representations h are passed through a feed-",
"coderforsummarizationbasedonthepriorworkonstructure",
"forward network followed by softmax to give a multinomial",
"attention networks [Liu and Lapata, 2018] which implicitly",
"distribution over the vocabulary pvocab(w).",
"incorporates structural information within end-to-end train-",
"Additionally,[See etal..2017]introduce a copymecha",
"ing.",
"nism [Vinyals et al., 2015] in this model. At each decoding",
"Our major contributions in this work are three-fold:",
"step t, it predicts whether to generate a new word or copy the",
"1.Weenhance the pointer-generator architectureby adding",
"word based on the attention probabilities at using a probabil-",
"ity value Pgen. This probability is a function on the decoder",
"a structural attention based encoder to implicitly cap-",
"ture long term dependency relations in summarization",
"input t, context vector Ct, and decoder state St . Pgen is used",
"as a soft switch to choose between copying and generation as",
"of lengthy documents.",
"shown below:",
"2. We further propose a hierarchical encoder with multi-",
"p(w) = PgenPvocab(w)+(1 -Pgen) a,",
"level structural attention to capture document-level dis-",
"course information in the multi-document summariza-",
".",
"This model is trained by minimizing the negative log",
"tion task.",
"likelihood of the predicted sequence normalized by the se-",
"3. We introduce multi-level contextual attention in the",
"quence length: L =-↓t=o log p(wt), where T is the",
" structural attention setting to enable word level copying",
"length of the generated summary. Once L gets converged, an",
"and to generate more abstractive summaries, compared",
"additional coverage loss is used to train the model further,",
"to similar architectures.",
"which aims to minimize repetitions in the generated sum-",
"We compare our proposed solution against the popular",
"maries. We refer the readers to [See et al., 2017] for more",
"pointer-generator model [See et al., 2017] and a few other",
"details.",
"summarization models derived from it in both single and",
"multi-document settings. We show that our structural en-",
"2.2Structural Attention",
"coderarchitecturebeatsthestrongpointer-generatorbaseline",
"We observe that while the pointer-generator architecture per-",
"by 1.31 ROUGE-1 points on the CNN/Dailymail dataset and",
"forms well on the news corpora, it fails to generate meaning-",
"by 7.8 ROUGE-1 points on the concatenated CQA dataset",
"ful summaries on complex datasets. We hypothesize that tak-",
"for single document summarization (SDS). Our hierarchi-",
"inginto account intra-documentdiscourseinformation will",
"cal structural encoder architecturefurther beatsthe concate-",
"help in generating better summaries. We propose an archi-",
"nated approach by another 1.6 ROUGE-1 points on the CQA",
"tecture that implicitly incorporates such information using",
"dataset. A qualitative analysis of the generated summaries ob-",
"document structural representation [Liu and Lapata, 2018;",
"serve considerable qualitative gains after inclusion of struc-",
"Kim et al., 20171 to include richer structural dependencies",
"tural attention. Our structural attention based summarinza-",
"in the end-to-end training",
"tion model is one of the few abstractive approaches to beat",
"Proposed Model",
"extractive baselines in MDS. The code is public at https:",
"We model our document as a non-projective dependency",
"/bit.ly/35i7q93.",
"parse tree by constraining inter-token attention as weights of",
"2 Models",
"the dependency tree. We use the Matrix tree theorem [Koo et",
"al., 2007; Tutte, 1984l to carry out the same.",
"In this section, we first describe the seq2seg based pointer-",
"As shown in Figure 1, we feed our input tokens (w;) to a",
"generator architecture proposed by [See et al., 2017; Nallap-",
"bi-LSTM encoder to obtain hidden state representations hi.",
"ati et al., 2016. We then elaborate on how we incorporate",
"We decompose the hidden state vector into two parts: d, and",
"structural attention into the mechanism and generate non-",
"ei, which we call the structural part and the semantic part,",
"projectivedependencytreestocapturelongtermdependen-",
"respectively:",
"cies within documents. Following this, we describe our hier-",
"[ei, di] = hi",
"(1)",
"archical encoder architecture and propose multi-level contex-",
"tual attention to better model discourse for MDS.",
"For every pair of two input tokens, we transform their",
"structural parts d and try to compute the probability of a",
"2.1 Pointer-Generator Model",
"parent-child relationship edge between them in the depen-",
"dency tree. For tokens j and k, this is done as:",
"The pointer-generator architecture (PG) [See et al., 2017]",
"serves as the starting point for our approach. PG is inspired",
"uj = tanh(Wpd,); uk = tanh(Weds)"
] |
{
"arxiv_id": "2004.09739",
"language": "en",
"page_idx": 1
}
|
||
2004.09739-en-2
|
2004.09739-en-2.jpg
|
<FigureHere>
Figure 1: Structural attention based encoder architecture. Tokens $(w_i)$ are independently fed to a bi-LSTM encoder, and hidden representations $h_i$ are obtained. These representations are split into structural $(d_i)$ and semantic $(e_i)$ components. The structural component is used to build a non-projective dependency tree. The inter-token attention $(a_{jk})$ is computed based on the marginal probability of an edge between two nodes in the tree $(P(z_{jk})=1).$
where $W_{p}$ and $W_{c}$ are learned. Next, we compute an inter-
token attention function $f_jk$ as follows:
$$f_{jk}=u_k^TW_au_j$$
where $W_a$ is also learned. For a document with $K$ tokens, $f$ is a $K\times K$ matrix representing inter-token attention. We model each token as a node in the dependency tree and define the probability of an edge between tokens at positions $j$ and $k,\bar{P}(z_{jk}=1)$,which is given as,
$$\begin{aligned}A_{jk}&=\begin{cases}0&\text{if }j=k\\\exp(f_{jk})&\text{otherwise}\end{cases}\\L_{jk}&=\begin{cases}\sum_{j^{'}=1}^{K}A_{j^{'}k}&\text{if }j=k\\-A_{jk}&\text{otherwise}\end{cases}\\f_{j}^{r}&=W_{r}d_{j}\\\bar{L}_{jk}&=\begin{cases}\exp(f_{j}^{r})&j=1\\L_{jk}&j>1\end{cases}\\=1)&=(1-\delta(j\:k))A_{\cdot\cdot\cdot}\bar{L}_{\cdot\cdot\cdot}^{-1}=\end{aligned}$$
$P(z_{jk}=1)=(1-\delta(j,k))A_{jk}\bar{L}_{kk}^{-1}-(1-\delta(j,1))A_{jk}\bar{L}_{kj}^{-1}$
where $\delta(x,y)=1$ when $x=y.$ We denote $P(z_jk=1)$ by $a_{jk}\left(\text{structural attention). Let }\bar{a_{j}^{r}}\right.$be the probability of the $j^{th}$ token to be the root:
$$a_j^r=\exp(W_rd_j)\bar{L}_{j1}^{-1}$$
(2)
We use this (soft) dependency tree formulation to compute a
structural representation $r$ for each encoder token as,
$$\begin{aligned}&s_{i}=a_{i}^{r}e_{\mathrm{root}}+\sum_{k=1}^{n}a_{ki}e_{k};\:c_{i}=\sum_{k=1}^{n}a_{ik}e_{i}\\&r_{i}=\mathrm{tanh}(W_{r}[e_{i},s_{i},c_{i}])\end{aligned}$$
Thus, for encoder step $i$,we now obtain a structure infused hidden representation $r_i.$ We then compute the contextual attention for each decoding time step $t$ as,
$$\begin{aligned}&e_{\mathrm{struct}i}^{t}=v^{t}\operatorname{tanh}(W_{r}r_{i}+W_{s}s_{t}+b_{\mathrm{attn}})\\&a_{\mathrm{struct}}^{t}=\mathrm{softmax}(e_{\mathrm{struct}}^{t})\end{aligned}$$
Now, using $a_\mathrm{struct}^{t}$, we can compute a context vector similar to standard attentional model by weighted sum of the hid- den state vectors as $c_{struct}^{t}= \sum _{i= 1}^{n}a_{\mathrm{struct}_{i}}^{t}h_{i}.$ At every de- coder time step, we also compute the basic contextual attention vector $a^t$ (without structure incorporation), as discussed previously. We use $c_struct^{t}$ to compute $P_vocab$ and $p_gen.$ We, however, use the initial attention distribution $a_t$ to compute $P(w)$ in order to facilitate token level pointing.
## 2.3 Hierarchical Model
The structural attention model, while efficient, requires $\mathcal{O}(K^{2})$ memory to compute attention, where $K$ is the length of the input sequence making it memory intensive for long documents. Moreover, in $\tilde{\text{CQA, answers reflect different}}$ opinions of individuals. In such a case, concatenating answers results in user specific information loss. To model discourse structure better in case of conflicting (where one answer contradicts other), overlapping or varying opinion, we
introduce a hierarchical encoder model based on structural attention (Figure 2). We feed each answer independently to a bi-LSTM encoder and obtain token level hidden representations $h_idx,tidx$ where idx is the document (or answer) index and tidx is the token index. We then transform these representations into structure infused vectors $r_{idx,tidx}$ as described previously. For each answer in the CQA question thread, we pool the token representations to obtain a composite answer ${\mathrm{~vector~}r_{idx}.\text{ We consider three types of pooling: average,max}}$ and sum, out of which sum pooling performs the best after initial experiments. We feed these structure infused answer embeddings to an answer-level bi-LSTM encoder and obtain higher-level encoder hidden states $h_idx$ from which we calculate structure infused embeddings $g_idx.$ At the decoder time step $t$, we calculate contextual attention at answer as well as token level as follows:
$$\begin{aligned}e_{\mathrm{ans}idx}^{t}&=v^{t}\tanh(W_{g}g_{idx}+W_{s}s_{t}+b_{\mathrm{attn}})\\a_{\mathrm{ans}}^{t}&=\mathrm{softmax}(e_{\mathrm{ans}}{}^{t})\\e_{\mathrm{token}}^{t}_{idx,tidx}&=v^{t}\tanh(W_{h}h_{idx,tidx}+W_{s}s_{t}+b_{\mathrm{attn}})\\a_{\mathrm{token}}^{t}&=\mathrm{softmax}(e_{\mathrm{token}}{}^{t})\end{aligned}$$
We use the answer-level attention distribution $a_\mathrm{ans}^{t}$ to compute the context vector at every decoder time step which we use to calculate $p_{vocab}$ and $p_{gen}$ as described before.
To enable copying, we use $a_\mathrm{token}.$ The final probability of
predicting word $w$ is given by,
$$p(w)=p_{gen}p_{vocab}(w)+(1-p_{gen})\sum_{i:w_{i}=w}a_{\mathrm{token}i}^{t}$$
# 3 Datasets
We primarily use two datasets to evaluate the performance:
|
[
[
0.0849609375,
0.06817626953125,
0.484375,
0.0694580078125,
0.484375,
0.083984375,
0.0849609375,
0.08331298828125
],
[
0.52783203125,
0.0706787109375,
0.6650390625,
0.0706787109375,
0.6650390625,
0.080810546875,
0.52783203125,
0.080810546875
],
[
0.0849609375,
0.083984375,
0.346435546875,
0.083984375,
0.346435546875,
0.0985107421875,
0.0849609375,
0.0985107421875
],
[
0.75341796875,
0.0997314453125,
0.8037109375,
0.0997314453125,
0.8037109375,
0.111083984375,
0.75341796875,
0.111083984375
],
[
0.826171875,
0.09783935546875,
0.88623046875,
0.0921630859375,
0.8896484375,
0.11297607421875,
0.82861328125,
0.11871337890625
],
[
0.233642578125,
0.10418701171875,
0.338134765625,
0.10418701171875,
0.338134765625,
0.121826171875,
0.233642578125,
0.121826171875
],
[
0.85888671875,
0.11236572265625,
0.88818359375,
0.11236572265625,
0.88818359375,
0.1383056640625,
0.85888671875,
0.1383056640625
],
[
0.69677734375,
0.12371826171875,
0.740234375,
0.12060546875,
0.74169921875,
0.1307373046875,
0.69775390625,
0.1337890625
],
[
0.75341796875,
0.121826171875,
0.775390625,
0.121826171875,
0.775390625,
0.1300048828125,
0.75341796875,
0.1300048828125
],
[
0.701171875,
0.1307373046875,
0.73876953125,
0.1307373046875,
0.73876953125,
0.1383056640625,
0.701171875,
0.1383056640625
],
[
0.85888671875,
0.1331787109375,
0.88818359375,
0.1331787109375,
0.88818359375,
0.1572265625,
0.85888671875,
0.1572265625
],
[
0.0849609375,
0.125,
0.486083984375,
0.1256103515625,
0.486083984375,
0.14013671875,
0.0849609375,
0.1395263671875
],
[
0.0882568359375,
0.1395263671875,
0.486083984375,
0.1395263671875,
0.486083984375,
0.15283203125,
0.0882568359375,
0.15283203125
],
[
0.69921875,
0.141357421875,
0.75244140625,
0.141357421875,
0.75244140625,
0.1514892578125,
0.69921875,
0.1514892578125
],
[
0.75244140625,
0.141357421875,
0.775390625,
0.141357421875,
0.775390625,
0.149658203125,
0.75244140625,
0.149658203125
],
[
0.57275390625,
0.1427001953125,
0.59228515625,
0.1427001953125,
0.59228515625,
0.1546630859375,
0.57275390625,
0.1546630859375
],
[
0.53369140625,
0.143310546875,
0.54833984375,
0.143310546875,
0.54833984375,
0.15283203125,
0.53369140625,
0.15283203125
],
[
0.08660888671875,
0.1546630859375,
0.484375,
0.1546630859375,
0.484375,
0.1654052734375,
0.08660888671875,
0.1654052734375
],
[
0.70361328125,
0.15087890625,
0.74853515625,
0.15087890625,
0.74853515625,
0.1546630859375,
0.70361328125,
0.1546630859375
],
[
0.7548828125,
0.160400390625,
0.80615234375,
0.160400390625,
0.80615234375,
0.1717529296875,
0.7548828125,
0.1717529296875
],
[
0.83251953125,
0.160400390625,
0.861328125,
0.160400390625,
0.861328125,
0.1717529296875,
0.83251953125,
0.1717529296875
],
[
0.8544921875,
0.1558837890625,
0.8857421875,
0.1558837890625,
0.8857421875,
0.17236328125,
0.8544921875,
0.17236328125
],
[
0.0849609375,
0.1673583984375,
0.48779296875,
0.1673583984375,
0.48779296875,
0.18115234375,
0.0849609375,
0.18115234375
],
[
0.53759765625,
0.1673583984375,
0.54833984375,
0.1717529296875,
0.54345703125,
0.177978515625,
0.53369140625,
0.17236328125
],
[
0.576171875,
0.1685791015625,
0.5986328125,
0.1685791015625,
0.5986328125,
0.175537109375,
0.576171875,
0.175537109375
],
[
0.701171875,
0.1654052734375,
0.74169921875,
0.1654052734375,
0.74169921875,
0.175537109375,
0.701171875,
0.175537109375
],
[
0.69677734375,
0.17236328125,
0.74609375,
0.17236328125,
0.74609375,
0.179931640625,
0.69677734375,
0.179931640625
],
[
0.0849609375,
0.18115234375,
0.30810546875,
0.18115234375,
0.30810546875,
0.1956787109375,
0.0849609375,
0.1956787109375
],
[
0.53125,
0.1875,
0.5458984375,
0.1875,
0.5458984375,
0.1976318359375,
0.53125,
0.1976318359375
],
[
0.57275390625,
0.1875,
0.58642578125,
0.1875,
0.58642578125,
0.1962890625,
0.57275390625,
0.1962890625
],
[
0.65625,
0.186279296875,
0.6689453125,
0.186279296875,
0.6689453125,
0.1962890625,
0.65625,
0.1962890625
],
[
0.191162109375,
0.2008056640625,
0.2115478515625,
0.2008056640625,
0.2115478515625,
0.21533203125,
0.191162109375,
0.21533203125
],
[
0.271240234375,
0.2008056640625,
0.330078125,
0.2008056640625,
0.330078125,
0.21533203125,
0.271240234375,
0.21533203125
],
[
0.1380615234375,
0.208984375,
0.1846923828125,
0.21337890625,
0.181396484375,
0.2279052734375,
0.1356201171875,
0.2235107421875
],
[
0.19287109375,
0.2197265625,
0.2607421875,
0.2197265625,
0.2607421875,
0.2342529296875,
0.19287109375,
0.2342529296875
],
[
0.27294921875,
0.2203369140625,
0.339111328125,
0.2203369140625,
0.339111328125,
0.231689453125,
0.27294921875,
0.231689453125
],
[
0.20751953125,
0.2374267578125,
0.223876953125,
0.2374267578125,
0.223876953125,
0.251953125,
0.20751953125,
0.251953125
],
[
0.53125,
0.2474365234375,
0.5537109375,
0.2474365234375,
0.5537109375,
0.255615234375,
0.53125,
0.255615234375
],
[
0.65625,
0.2474365234375,
0.67724609375,
0.2474365234375,
0.67724609375,
0.255615234375,
0.65625,
0.255615234375
],
[
0.8203125,
0.2474365234375,
0.8447265625,
0.2474365234375,
0.8447265625,
0.255615234375,
0.8203125,
0.255615234375
],
[
0.69775390625,
0.248779296875,
0.71337890625,
0.248779296875,
0.71337890625,
0.255615234375,
0.69775390625,
0.255615234375
],
[
0.5703125,
0.248779296875,
0.59326171875,
0.248779296875,
0.59326171875,
0.255615234375,
0.5703125,
0.255615234375
],
[
0.73876953125,
0.248779296875,
0.76123046875,
0.248779296875,
0.76123046875,
0.255615234375,
0.73876953125,
0.255615234375
],
[
0.78125,
0.250732421875,
0.79833984375,
0.250732421875,
0.79833984375,
0.255126953125,
0.78125,
0.255126953125
],
[
0.233642578125,
0.25244140625,
0.26318359375,
0.25244140625,
0.26318359375,
0.26513671875,
0.233642578125,
0.26513671875
],
[
0.2059326171875,
0.250732421875,
0.2181396484375,
0.250732421875,
0.2181396484375,
0.26318359375,
0.2059326171875,
0.26318359375
],
[
0.275390625,
0.25244140625,
0.334228515625,
0.25244140625,
0.334228515625,
0.264404296875,
0.275390625,
0.264404296875
],
[
0.1405029296875,
0.261474609375,
0.1748046875,
0.26513671875,
0.17236328125,
0.27978515625,
0.1380615234375,
0.27587890625
],
[
0.17236328125,
0.266357421875,
0.17724609375,
0.266357421875,
0.17724609375,
0.271484375,
0.17236328125,
0.271484375
],
[
0.200927734375,
0.26953125,
0.2296142578125,
0.26953125,
0.2296142578125,
0.2822265625,
0.200927734375,
0.2822265625
],
[
0.2099609375,
0.285888671875,
0.2418212890625,
0.285888671875,
0.2418212890625,
0.3017578125,
0.2099609375,
0.3017578125
],
[
0.27783203125,
0.287353515625,
0.343017578125,
0.287353515625,
0.343017578125,
0.29736328125,
0.27783203125,
0.29736328125
],
[
0.51123046875,
0.28271484375,
0.91259765625,
0.28271484375,
0.91259765625,
0.296630859375,
0.51123046875,
0.296630859375
],
[
0.51123046875,
0.296630859375,
0.91259765625,
0.296630859375,
0.91259765625,
0.31005859375,
0.51123046875,
0.31005859375
],
[
0.1446533203125,
0.302978515625,
0.226318359375,
0.3017578125,
0.226318359375,
0.32080078125,
0.1446533203125,
0.3212890625
],
[
0.51123046875,
0.309326171875,
0.91259765625,
0.309326171875,
0.91259765625,
0.322509765625,
0.51123046875,
0.322509765625
],
[
0.25,
0.323974609375,
0.3095703125,
0.323974609375,
0.3095703125,
0.33837890625,
0.25,
0.33837890625
],
[
0.51123046875,
0.3212890625,
0.9140625,
0.3212890625,
0.9140625,
0.335205078125,
0.51123046875,
0.335205078125
],
[
0.1396484375,
0.330810546875,
0.1846923828125,
0.335205078125,
0.181396484375,
0.3515625,
0.136474609375,
0.347900390625
],
[
0.191162109375,
0.324462890625,
0.25732421875,
0.324462890625,
0.25732421875,
0.339111328125,
0.191162109375,
0.339111328125
],
[
0.191162109375,
0.339111328125,
0.2296142578125,
0.3447265625,
0.2254638671875,
0.361083984375,
0.1871337890625,
0.35595703125
],
[
0.51123046875,
0.333984375,
0.9140625,
0.333984375,
0.9140625,
0.347900390625,
0.51123046875,
0.347900390625
],
[
0.2646484375,
0.343505859375,
0.30810546875,
0.341552734375,
0.3095703125,
0.356689453125,
0.265625,
0.359130859375
],
[
0.51123046875,
0.3466796875,
0.9140625,
0.3466796875,
0.9140625,
0.35986328125,
0.51123046875,
0.35986328125
],
[
0.0882568359375,
0.3623046875,
0.48193359375,
0.3623046875,
0.48193359375,
0.380126953125,
0.0882568359375,
0.380126953125
],
[
0.5107421875,
0.35791015625,
0.810546875,
0.359130859375,
0.810546875,
0.373779296875,
0.5107421875,
0.37255859375
],
[
0.30810546875,
0.367431640625,
0.32275390625,
0.367431640625,
0.32275390625,
0.375,
0.30810546875,
0.375
],
[
0.08660888671875,
0.386474609375,
0.484375,
0.386474609375,
0.484375,
0.40087890625,
0.08660888671875,
0.40087890625
],
[
0.08416748046875,
0.40087890625,
0.486083984375,
0.397705078125,
0.486083984375,
0.415283203125,
0.08416748046875,
0.41845703125
],
[
0.51318359375,
0.40283203125,
0.91259765625,
0.40283203125,
0.91259765625,
0.416748046875,
0.51318359375,
0.416748046875
],
[
0.0849609375,
0.416748046875,
0.221435546875,
0.416748046875,
0.221435546875,
0.429931640625,
0.0849609375,
0.429931640625
],
[
0.51123046875,
0.417236328125,
0.9140625,
0.417236328125,
0.9140625,
0.43115234375,
0.51123046875,
0.43115234375
],
[
0.2132568359375,
0.434326171875,
0.358642578125,
0.434326171875,
0.358642578125,
0.451904296875,
0.2132568359375,
0.451904296875
],
[
0.464111328125,
0.436279296875,
0.486083984375,
0.436279296875,
0.486083984375,
0.448974609375,
0.464111328125,
0.448974609375
],
[
0.5107421875,
0.43115234375,
0.9111328125,
0.43115234375,
0.9111328125,
0.4443359375,
0.5107421875,
0.4443359375
],
[
0.51318359375,
0.44580078125,
0.9140625,
0.44580078125,
0.9140625,
0.458984375,
0.51318359375,
0.458984375
],
[
0.0849609375,
0.453857421875,
0.488525390625,
0.455078125,
0.488525390625,
0.4697265625,
0.0849609375,
0.468505859375
],
[
0.5107421875,
0.45703125,
0.91259765625,
0.458251953125,
0.91259765625,
0.472900390625,
0.5107421875,
0.4716796875
],
[
0.08416748046875,
0.4697265625,
0.431396484375,
0.4697265625,
0.431396484375,
0.48291015625,
0.08416748046875,
0.48291015625
],
[
0.51123046875,
0.472900390625,
0.9140625,
0.472900390625,
0.9140625,
0.4873046875,
0.51123046875,
0.4873046875
],
[
0.366943359375,
0.4912109375,
0.37744140625,
0.4873046875,
0.3857421875,
0.498779296875,
0.374267578125,
0.50244140625
],
[
0.51318359375,
0.486083984375,
0.9140625,
0.486083984375,
0.9140625,
0.50048828125,
0.51318359375,
0.50048828125
],
[
0.143798828125,
0.494384765625,
0.26708984375,
0.494384765625,
0.26708984375,
0.5166015625,
0.143798828125,
0.5166015625
],
[
0.359375,
0.494384765625,
0.426513671875,
0.496826171875,
0.4248046875,
0.51806640625,
0.358642578125,
0.5166015625
],
[
0.26318359375,
0.50244140625,
0.36279296875,
0.50244140625,
0.36279296875,
0.5146484375,
0.26318359375,
0.5146484375
],
[
0.51318359375,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.5146484375,
0.51318359375,
0.5146484375
],
[
0.36279296875,
0.517578125,
0.38818359375,
0.517578125,
0.38818359375,
0.52880859375,
0.36279296875,
0.52880859375
],
[
0.51318359375,
0.5146484375,
0.9140625,
0.5146484375,
0.9140625,
0.52880859375,
0.51318359375,
0.52880859375
],
[
0.51318359375,
0.52880859375,
0.9140625,
0.52880859375,
0.9140625,
0.54248046875,
0.51318359375,
0.54248046875
],
[
0.25244140625,
0.51806640625,
0.273681640625,
0.51806640625,
0.273681640625,
0.5263671875,
0.25244140625,
0.5263671875
],
[
0.1446533203125,
0.53076171875,
0.313720703125,
0.5322265625,
0.313720703125,
0.546875,
0.1446533203125,
0.54541015625
],
[
0.5146484375,
0.54345703125,
0.9111328125,
0.54345703125,
0.9111328125,
0.5537109375,
0.5146484375,
0.5537109375
],
[
0.101318359375,
0.5498046875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56494140625,
0.101318359375,
0.564453125
],
[
0.51123046875,
0.5546875,
0.91259765625,
0.5546875,
0.91259765625,
0.5693359375,
0.51123046875,
0.5693359375
],
[
0.0849609375,
0.56494140625,
0.48779296875,
0.56494140625,
0.48779296875,
0.580078125,
0.0849609375,
0.580078125
],
[
0.51123046875,
0.5693359375,
0.9111328125,
0.5693359375,
0.9111328125,
0.58349609375,
0.51123046875,
0.58349609375
],
[
0.0849609375,
0.580078125,
0.36279296875,
0.580078125,
0.36279296875,
0.59326171875,
0.0849609375,
0.59326171875
],
[
0.51123046875,
0.583984375,
0.9140625,
0.583984375,
0.9140625,
0.59765625,
0.51123046875,
0.59765625
],
[
0.146240234375,
0.5966796875,
0.423095703125,
0.5966796875,
0.423095703125,
0.6142578125,
0.146240234375,
0.6142578125
],
[
0.51123046875,
0.59765625,
0.9140625,
0.59765625,
0.9140625,
0.611328125,
0.51123046875,
0.611328125
],
[
0.148681640625,
0.61572265625,
0.339111328125,
0.61572265625,
0.339111328125,
0.63720703125,
0.148681640625,
0.63720703125
],
[
0.51123046875,
0.60986328125,
0.66015625,
0.611328125,
0.66015625,
0.625,
0.51123046875,
0.62353515625
],
[
0.098876953125,
0.63818359375,
0.486083984375,
0.6396484375,
0.486083984375,
0.6572265625,
0.098876953125,
0.65576171875
],
[
0.56787109375,
0.63330078125,
0.865234375,
0.630859375,
0.865234375,
0.6513671875,
0.56787109375,
0.65380859375
],
[
0.08660888671875,
0.65380859375,
0.484375,
0.65380859375,
0.484375,
0.66748046875,
0.08660888671875,
0.66748046875
],
[
0.58837890625,
0.65478515625,
0.74609375,
0.65380859375,
0.74609375,
0.671875,
0.58837890625,
0.67236328125
],
[
0.08087158203125,
0.66357421875,
0.488525390625,
0.66357421875,
0.488525390625,
0.6875,
0.08087158203125,
0.6875
],
[
0.529296875,
0.67236328125,
0.8935546875,
0.67041015625,
0.8935546875,
0.69140625,
0.529296875,
0.693359375
],
[
0.08660888671875,
0.6845703125,
0.483642578125,
0.6845703125,
0.483642578125,
0.6943359375,
0.08660888671875,
0.6943359375
],
[
0.0849609375,
0.6962890625,
0.48779296875,
0.6962890625,
0.48779296875,
0.71044921875,
0.0849609375,
0.71044921875
],
[
0.5703125,
0.6943359375,
0.75634765625,
0.6943359375,
0.75634765625,
0.7119140625,
0.5703125,
0.7119140625
],
[
0.0849609375,
0.708984375,
0.48779296875,
0.708984375,
0.48779296875,
0.7265625,
0.0849609375,
0.7265625
],
[
0.0849609375,
0.7236328125,
0.48779296875,
0.7236328125,
0.48779296875,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.52783203125,
0.7216796875,
0.91015625,
0.72265625,
0.91015625,
0.7373046875,
0.52783203125,
0.736328125
],
[
0.0849609375,
0.736328125,
0.396240234375,
0.7373046875,
0.396240234375,
0.751953125,
0.0849609375,
0.75048828125
],
[
0.51318359375,
0.7373046875,
0.9140625,
0.7373046875,
0.9140625,
0.75048828125,
0.51318359375,
0.75048828125
],
[
0.51123046875,
0.751953125,
0.85546875,
0.751953125,
0.85546875,
0.7666015625,
0.51123046875,
0.7666015625
],
[
0.08416748046875,
0.7587890625,
0.27783203125,
0.76025390625,
0.27783203125,
0.7744140625,
0.08416748046875,
0.7734375
],
[
0.52685546875,
0.76318359375,
0.9150390625,
0.7646484375,
0.9150390625,
0.7822265625,
0.52685546875,
0.78076171875
],
[
0.0849609375,
0.77783203125,
0.486083984375,
0.77880859375,
0.486083984375,
0.79248046875,
0.0849609375,
0.79150390625
],
[
0.51123046875,
0.77978515625,
0.7158203125,
0.77978515625,
0.7158203125,
0.79345703125,
0.51123046875,
0.79345703125
],
[
0.0849609375,
0.79052734375,
0.486083984375,
0.79150390625,
0.486083984375,
0.80615234375,
0.0849609375,
0.8046875
],
[
0.0849609375,
0.80615234375,
0.486083984375,
0.80615234375,
0.486083984375,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.5400390625,
0.806640625,
0.7939453125,
0.80615234375,
0.7939453125,
0.82373046875,
0.5400390625,
0.82470703125
],
[
0.7890625,
0.8095703125,
0.8056640625,
0.80810546875,
0.80615234375,
0.81494140625,
0.7900390625,
0.81640625
],
[
0.82177734375,
0.806640625,
0.8798828125,
0.8095703125,
0.87841796875,
0.82373046875,
0.8203125,
0.8212890625
],
[
0.08660888671875,
0.8212890625,
0.176513671875,
0.8212890625,
0.176513671875,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.17236328125,
0.8212890625,
0.486083984375,
0.8212890625,
0.486083984375,
0.83203125,
0.17236328125,
0.83203125
],
[
0.78759765625,
0.82568359375,
0.833984375,
0.82568359375,
0.833984375,
0.8359375,
0.78759765625,
0.8359375
],
[
0.0849609375,
0.833984375,
0.484375,
0.833984375,
0.484375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.0849609375,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.0849609375,
0.861328125
],
[
0.51953125,
0.85302734375,
0.52685546875,
0.85791015625,
0.51953125,
0.86279296875,
0.51318359375,
0.85791015625
],
[
0.5400390625,
0.85400390625,
0.61376953125,
0.85400390625,
0.61376953125,
0.86572265625,
0.5400390625,
0.86572265625
],
[
0.0849609375,
0.8623046875,
0.484375,
0.8623046875,
0.484375,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.08416748046875,
0.87548828125,
0.488525390625,
0.87548828125,
0.488525390625,
0.8896484375,
0.08416748046875,
0.8896484375
],
[
0.51318359375,
0.87548828125,
0.89892578125,
0.87548828125,
0.89892578125,
0.8896484375,
0.51318359375,
0.8896484375
]
] |
[
"where W, and W。 are learned. Next, we compute an inter-",
"UPDATED SEMANTICEMBEDDINGS",
"token attention function fjk as follows:",
"a11(a1,2",
"anero",
"fik= utWauj",
"",
"INTERTOKE",
"a2,1",
"ATTENTIC",
"ar",
"where Wa is also learned. For a document with K tokens,",
"f is a K × K matrix representing inter-token attention. We",
"STRUCTURAL",
"aii",
"d2",
"di",
"modeleachtokenasanodeinthedependencytreeand define",
"OUBONEN",
"an,an2",
"ann",
"aroot",
"the probability of an edge between tokens at positions j and",
"es",
"ea4",
"SEMANTIC",
"COMPONENT",
"k, P(zjk = 1), which is given as,",
"hl",
"ha",
"h",
"(0",
"ifj=k",
"Ajk=",
"(exp(fik)",
"otherwise",
"K",
"w",
"w",
"Wn",
"",
"W2",
"",
"",
"Aitk",
">",
"ifj=k",
"Lik",
"-",
"3'=1",
"Ajk",
"otherwise",
"Figure l: Structural attention based encoder architecture. Tokens",
"(w:) are independently fed to a bi-LSTM encoder, and hidden repre-",
"f, =Wedj",
"sentations h, are obtained. These representations are split into struc-",
"j=1",
"tural (d;) and semantic (e:) components. The structural component",
"Lik =",
"(exp(f,)",
"Lijk",
"is used to build a non-projective dependency tree. The inter-token",
"j>1",
"attention (ajk) is computed based on the marginal probability of an",
"P(zk =1) = (1 - 8(j, k))AL-h - (1 - 5(j,1)AjnL)",
"edge between two nodes in the tree (P(zjk) = 1).",
"",
"where (α, y) = 1 when a = y . We denote P(zik = 1) by",
"ajk (structural attention). Let a, be the probability of the jth",
"introduce a hierarchical encoder model based on structural",
"token to be the root:",
"attention (Figure 2). We feed each answer independently to a",
"aj = exp(W.d,)L,1",
"(2)",
"bi-LSTM encoder and obtain token level hidden representa-",
"tions hidr,tidr where idc is the document (or answer) index",
"We use this (soft) dependency tree formulation to compute a",
"and tidr is the token index. We then transform these repre-",
" structural representation r for each encoder token as,",
"sentations into structure infused vectors ridz,tidr as described",
"n",
"previously. For each answer in the CQA question thread, we",
"Si = dieroot +\\",
"Laikei",
"akiek;C=",
"pool the token representations to obtain a composite answer",
"k=1",
"vector ridr. We consider three types of pooling: average, max",
"and sum, out of which sum pooling performs the best after",
"k=1",
"ri = tanh(W.[ei, si,c])",
"initial experiments.We feed these structure infused answer",
"Thus, for encoder step i, we now obtain a structure infused",
"embeddings to an answer-level bi-LSTM encoder and obtain",
"hidden representation ri.We then compute the contextual",
"higher-level encoder hidden states hidr from which we calcu-",
"attention for each decoding time step t as,",
"late structure infused embeddings gidr.At the decoder time",
"estruct = ut tanh(Wrr; + W,st +battn)",
"step t, we calculate contextual attention at answer as well as",
"astruct = softmax(estruet')",
"token level as follows:",
"Now, using astruct, we can compute a context vector simi-",
"eansida = ut tanh(Wggidr + Wgst + battn)",
"lar to standard attentional model by weighted sum of the hid-",
"aans = softmax(eanst)",
"den state vectors as cetruet - atruet hi. At every de-",
"Etokenida,tida = ut tanh(Whidr,tida + W st + battn)",
"codertimestep,wealsocomputethebasiccontextual atten-",
"tion vector at (without structure incorporation), as discussed",
"atokent = softmax(etokent)",
"previously. We use cstruct to compute Puocab and Pgen. We,",
"however, use the initial attention distribution at to compute",
"We use the answer-level attention distribution atns to com-",
"P(w) in order to facilitate token level pointing.",
"pute the context vector at every decoder time step which we",
"use to calculate puocab and pgen as described before.",
"2.3Hierarchical Model",
"To enable copying, we use atoken. The final probability of",
"The structural attention model, while efficient, requires",
"predicting word w is given by,",
"O(K2) memory to compute attention, where K is the length",
"of the input sequence making it memory intensive for long",
"p(w) = PgenPvocab(w) + (1 - Pgen)",
"",
"atokeni",
"documents.",
"Moreover,in CQA,answers reflect different",
"i:w=w",
"opinions of individuals. In such a case, concatenating an-",
" swers results in user specific information loss. To model dis-",
"2",
"Datasets",
"course structure better in case of conflicting (where one an-",
"swer contradicts other), overlapping or varying opinion, we",
"We primarily use two datasets to evaluate the performance:"
] |
{
"arxiv_id": "2004.09739",
"language": "en",
"page_idx": 2
}
|
||
2004.09739-en-3
|
2004.09739-en-3.jpg
|
<FigureHere>
Figure 2: Hierarchical encoder with structural attention + multi level contextual attention for multi-document summarization. First, token-level structured representations are computed. For each answer, these representations are pooled $(r_i)$ and fed to an answerlevel encoder. Structural infused answer level embeddings ($g_i)$ are calculated across answers. At every decoder time step, we calculate the context vector and probability $p_{gen}$ based on structural contextual attention. We calculate the copying vector (source attention) by projecting token-level attention onto the vocabulary.
(i) The CNN/Dailymail$^{\mathrm{l}}$ dataset[Hermann et al., 2015; Nallapati et $al.$, 2016] is a news corpora containing $\hat{\text{document-summary pairs. Bulleted extracts from the CNN}}$ and Dailymail news pages online are projected as summaries to the remaining documents. The scripts released by[Nallapati $et$ al., 2016] are used to extract approximately $250k$ training pairs, $13k$ validation pairs and $11.5k$ test pairs from the corpora. We use the non-anonymized form of the data to provide fair comparison to the experiments, conducted by[See et al., 2017]. It is a factual, English corpora with an average document length being 781 tokens and an average summary length being 56 tokens. We use two versions of this dataset- one with 400 word articles, and the other with 800 word articles. Most research reporting results on this dataset are on CNN/Dailymail-400. We also consider a 800 token version of this dataset as longer articles harbor more intra-document structural dependencies which would allow us to better demonstrate the benefits of structure incorporation. Moreover, longer documents resemble real-world $\hat{d}$atasets. ( ii) We also use the CQA datase$t^{2}$ [ Chowdhury and Chakraborty, 2019] which is generated by filtering the Yahoo! Answers L6 corpora to find question threads where the best answer can serve as a summary for the remaining answers. The authors use a series of heuristics to arrive at a set of 100k question thread-summary pairs. The summaries are generated by modifying best answers and selecting most question-relevant sentences from them. The remaining answers serve as candidate documents for summarization mak-
$^{1}$https://github.com/abisee/cnn-dailymail
$^{2}$https://bitbucket.org/tanya14109/cqasumm/src/master/
ing up a large-scale, diverse and highly abstract dataset. On an average, the corpus has 12 answers per question thread, with 65 words per answer. All summaries are truncated at hundred words. We split the 100k dataset into 80k training instances, $10k$ validation and $10k$ test instances. We additionally extract the upvote information corresponding to every answer from the L6 dataset and assume that upvotes have a high correlation with the relative importance and relevance of an answer. We then rank answers in decreasing order of upvotes before concatenating them as required for several baselines. Since Yahoo! Answers is an unstructured and unmoderated question-answer repository, this has turned out to be a challenging summarization dataset[Chowdhury and Chakraborty, 2019]. Additionally, we include analysis on MultiNews [Fabbri et $al.,2019]$, a news based MSD corpora to aid similar studies. It is the first large scale MSD news dataset consisting of 56,216 article summary pairs, crowd-sourced from various different news websites.
## 4 $\textbf{Competing Methods}$
We compare the performance of the following SDS and MDS
## ${\mathrm{models.}}$
$\cdot$ Lead3: It is an extractive baseline where the first 100 tokens of the document (in case of SDS datasets) and concatenated ranked documents (in case of MDS datasets) are picked to form the summary.
$\cdot$ KL-Summ: It is an extractive summarization method introduced by [Haghighi and Vanderwende, 2009] that attempts to minimize KL-Divergence between candidate $\hat{\text{documents and generated summary}}.$
$\cdot$ LexRank: It is an unsupervised extractive summarization method[Erkan and Radev, 2004]. A graph is built with sentences as vertices, and edge weights are assigned based on sentence similarity.
$\cdot \textbf{TextRank: It is an unsupervised extractive summariza- }$ tion method which selects sentences such that the information being disseminated by the summary is as close as possible to the original documents [Mihalcea and Tarau, $\dot{2}004].$
$\cdot$ Pointer-Generator (PG): It is a supervised abstractive summarization model[See et al., 2017], as discussed earlier. It is a strong and popularly used baseline for summarization.
$\bullet$ Pointer-Generator + Structure Infused Copy $(\mathbf{PG+SC):Our~implementation~is~similar~to~one~of~the}$ methods proposed by [Song et al., 2018]. We explicitly compute the dependency tree of sentences and encode a structure vector based on features like POS tag, number of incoming edges, depth of the tree, etc. We then concatenate this structural vector for every token to its hidden state representation in pointer-generator networks.
$\bullet$ Pointer-Generator+MMR(PG+MMR): It is an abstractive MDS model [Lebanoff et al., 2018] trained on the CNN/Dailymail dataset. It combines Maximal Marginal Relevance (MMR) method with pointer-
|
[
[
0.51318359375,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.51318359375,
0.08331298828125
],
[
0.10455322265625,
0.08331298828125,
0.1446533203125,
0.08331298828125,
0.1446533203125,
0.09027099609375,
0.10455322265625,
0.09027099609375
],
[
0.143798828125,
0.080810546875,
0.1650390625,
0.08331298828125,
0.1641845703125,
0.091552734375,
0.1422119140625,
0.0889892578125
],
[
0.51318359375,
0.083984375,
0.9111328125,
0.083984375,
0.9111328125,
0.09783935546875,
0.51318359375,
0.09783935546875
],
[
0.27294921875,
0.09783935546875,
0.3447265625,
0.0997314453125,
0.3447265625,
0.10797119140625,
0.27294921875,
0.1060791015625
],
[
0.51318359375,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.51318359375,
0.111083984375
],
[
0.51123046875,
0.10986328125,
0.9140625,
0.111083984375,
0.9140625,
0.1256103515625,
0.51123046875,
0.125
],
[
0.38232421875,
0.12249755859375,
0.406005859375,
0.12249755859375,
0.406005859375,
0.1307373046875,
0.38232421875,
0.1307373046875
],
[
0.406005859375,
0.125,
0.418212890625,
0.125,
0.418212890625,
0.1287841796875,
0.406005859375,
0.1287841796875
],
[
0.51318359375,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.2034912109375,
0.1337890625,
0.2900390625,
0.1337890625,
0.2900390625,
0.141357421875,
0.2034912109375,
0.141357421875
],
[
0.373291015625,
0.1337890625,
0.4697265625,
0.1337890625,
0.4697265625,
0.141357421875,
0.373291015625,
0.141357421875
],
[
0.5146484375,
0.141357421875,
0.9111328125,
0.141357421875,
0.9111328125,
0.1514892578125,
0.5146484375,
0.1514892578125
],
[
0.351318359375,
0.147705078125,
0.431396484375,
0.147705078125,
0.431396484375,
0.1546630859375,
0.351318359375,
0.1546630859375
],
[
0.10540771484375,
0.1558837890625,
0.11358642578125,
0.1558837890625,
0.11358642578125,
0.1610107421875,
0.10540771484375,
0.1610107421875
],
[
0.1446533203125,
0.1572265625,
0.158447265625,
0.1572265625,
0.158447265625,
0.1610107421875,
0.1446533203125,
0.1610107421875
],
[
0.51318359375,
0.15283203125,
0.91259765625,
0.15283203125,
0.91259765625,
0.1666259765625,
0.51318359375,
0.1666259765625
],
[
0.365234375,
0.160400390625,
0.4306640625,
0.160400390625,
0.4306640625,
0.17041015625,
0.365234375,
0.17041015625
],
[
0.348876953125,
0.173583984375,
0.438720703125,
0.173583984375,
0.438720703125,
0.184326171875,
0.348876953125,
0.184326171875
],
[
0.5146484375,
0.1685791015625,
0.9111328125,
0.1685791015625,
0.9111328125,
0.1787109375,
0.5146484375,
0.1787109375
],
[
0.35546875,
0.186279296875,
0.396240234375,
0.186279296875,
0.396240234375,
0.2008056640625,
0.35546875,
0.2008056640625
],
[
0.51318359375,
0.18115234375,
0.91259765625,
0.18115234375,
0.91259765625,
0.1944580078125,
0.51318359375,
0.1944580078125
],
[
0.51318359375,
0.1944580078125,
0.9111328125,
0.1944580078125,
0.9111328125,
0.2083740234375,
0.51318359375,
0.2083740234375
],
[
0.51318359375,
0.2083740234375,
0.91259765625,
0.2083740234375,
0.91259765625,
0.2215576171875,
0.51318359375,
0.2215576171875
],
[
0.10296630859375,
0.2279052734375,
0.12091064453125,
0.2279052734375,
0.12091064453125,
0.236083984375,
0.10296630859375,
0.236083984375
],
[
0.1217041015625,
0.2279052734375,
0.1446533203125,
0.2279052734375,
0.1446533203125,
0.236083984375,
0.1217041015625,
0.236083984375
],
[
0.143798828125,
0.2279052734375,
0.1650390625,
0.2279052734375,
0.1650390625,
0.236083984375,
0.143798828125,
0.236083984375
],
[
0.51318359375,
0.222900390625,
0.91259765625,
0.222900390625,
0.91259765625,
0.236083984375,
0.51318359375,
0.236083984375
],
[
0.51123046875,
0.23486328125,
0.9111328125,
0.236083984375,
0.9111328125,
0.25,
0.51123046875,
0.248779296875
],
[
0.51318359375,
0.25,
0.9111328125,
0.25,
0.9111328125,
0.26318359375,
0.51318359375,
0.26318359375
],
[
0.51123046875,
0.261962890625,
0.91259765625,
0.26318359375,
0.91259765625,
0.27783203125,
0.51123046875,
0.276611328125
],
[
0.51318359375,
0.276611328125,
0.91259765625,
0.27783203125,
0.91259765625,
0.291748046875,
0.51318359375,
0.290283203125
],
[
0.2132568359375,
0.2841796875,
0.281005859375,
0.28271484375,
0.281005859375,
0.29345703125,
0.2132568359375,
0.294189453125
],
[
0.386474609375,
0.287353515625,
0.4306640625,
0.287353515625,
0.4306640625,
0.29736328125,
0.386474609375,
0.29736328125
],
[
0.136474609375,
0.2890625,
0.1568603515625,
0.2890625,
0.1568603515625,
0.292236328125,
0.136474609375,
0.292236328125
],
[
0.10540771484375,
0.2890625,
0.11767578125,
0.285888671875,
0.1192626953125,
0.290283203125,
0.1070556640625,
0.29345703125
],
[
0.2320556640625,
0.294189453125,
0.26318359375,
0.294189453125,
0.26318359375,
0.298583984375,
0.2320556640625,
0.298583984375
],
[
0.51123046875,
0.290283203125,
0.91259765625,
0.290283203125,
0.91259765625,
0.3037109375,
0.51123046875,
0.3037109375
],
[
0.51123046875,
0.30615234375,
0.6689453125,
0.30615234375,
0.6689453125,
0.31640625,
0.51123046875,
0.31640625
],
[
0.08660888671875,
0.31494140625,
0.48779296875,
0.31494140625,
0.48779296875,
0.328857421875,
0.08660888671875,
0.328857421875
],
[
0.08660888671875,
0.330078125,
0.486083984375,
0.330078125,
0.486083984375,
0.34033203125,
0.08660888671875,
0.34033203125
],
[
0.5185546875,
0.339111328125,
0.525390625,
0.339111328125,
0.525390625,
0.343505859375,
0.5185546875,
0.343505859375
],
[
0.54150390625,
0.333984375,
0.7138671875,
0.333984375,
0.7138671875,
0.348388671875,
0.54150390625,
0.348388671875
],
[
0.08660888671875,
0.34228515625,
0.484375,
0.34228515625,
0.484375,
0.352783203125,
0.08660888671875,
0.352783203125
],
[
0.08660888671875,
0.354736328125,
0.484375,
0.354736328125,
0.484375,
0.365478515625,
0.08660888671875,
0.365478515625
],
[
0.51123046875,
0.352783203125,
0.9111328125,
0.352783203125,
0.9111328125,
0.367431640625,
0.51123046875,
0.367431640625
],
[
0.0849609375,
0.367431640625,
0.486083984375,
0.367431640625,
0.486083984375,
0.37744140625,
0.0849609375,
0.37744140625
],
[
0.5146484375,
0.369384765625,
0.56396484375,
0.369384765625,
0.56396484375,
0.380615234375,
0.5146484375,
0.380615234375
],
[
0.0849609375,
0.380126953125,
0.486083984375,
0.380126953125,
0.486083984375,
0.39013671875,
0.0849609375,
0.39013671875
],
[
0.0849609375,
0.391357421875,
0.486083984375,
0.391357421875,
0.486083984375,
0.40478515625,
0.0849609375,
0.40478515625
],
[
0.53515625,
0.386962890625,
0.9111328125,
0.386962890625,
0.9111328125,
0.40087890625,
0.53515625,
0.40087890625
],
[
0.08660888671875,
0.40478515625,
0.484375,
0.40478515625,
0.484375,
0.415283203125,
0.08660888671875,
0.415283203125
],
[
0.5458984375,
0.401611328125,
0.91015625,
0.401611328125,
0.91015625,
0.415283203125,
0.5458984375,
0.415283203125
],
[
0.08416748046875,
0.416748046875,
0.39453125,
0.416748046875,
0.39453125,
0.429931640625,
0.08416748046875,
0.429931640625
],
[
0.5439453125,
0.415283203125,
0.91259765625,
0.415283203125,
0.91259765625,
0.4287109375,
0.5439453125,
0.4287109375
],
[
0.5439453125,
0.429931640625,
0.7587890625,
0.429931640625,
0.7587890625,
0.443115234375,
0.5439453125,
0.443115234375
],
[
0.53173828125,
0.447509765625,
0.9140625,
0.447509765625,
0.9140625,
0.4609375,
0.53173828125,
0.4609375
],
[
0.10455322265625,
0.460205078125,
0.484375,
0.460205078125,
0.484375,
0.474853515625,
0.10455322265625,
0.474853515625
],
[
0.5458984375,
0.46337890625,
0.91259765625,
0.46337890625,
0.91259765625,
0.473388671875,
0.5458984375,
0.473388671875
],
[
0.0849609375,
0.473388671875,
0.486083984375,
0.474853515625,
0.486083984375,
0.4892578125,
0.0849609375,
0.488037109375
],
[
0.5439453125,
0.47607421875,
0.91259765625,
0.47607421875,
0.91259765625,
0.4892578125,
0.5439453125,
0.4892578125
],
[
0.08416748046875,
0.488037109375,
0.484375,
0.486083984375,
0.484375,
0.50048828125,
0.08416748046875,
0.50244140625
],
[
0.5439453125,
0.490478515625,
0.78125,
0.490478515625,
0.78125,
0.50390625,
0.5439453125,
0.50390625
],
[
0.08660888671875,
0.50390625,
0.484375,
0.50390625,
0.484375,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.53173828125,
0.50830078125,
0.91015625,
0.50830078125,
0.91015625,
0.521484375,
0.53173828125,
0.521484375
],
[
0.0849609375,
0.5166015625,
0.484375,
0.5166015625,
0.484375,
0.52978515625,
0.0849609375,
0.52978515625
],
[
0.5439453125,
0.52294921875,
0.91259765625,
0.52294921875,
0.91259765625,
0.5361328125,
0.5439453125,
0.5361328125
],
[
0.08660888671875,
0.52978515625,
0.484375,
0.52978515625,
0.484375,
0.54443359375,
0.08660888671875,
0.54443359375
],
[
0.5439453125,
0.537109375,
0.9111328125,
0.537109375,
0.9111328125,
0.55029296875,
0.5439453125,
0.55029296875
],
[
0.08660888671875,
0.54541015625,
0.484375,
0.54541015625,
0.484375,
0.55615234375,
0.08660888671875,
0.55615234375
],
[
0.5458984375,
0.55029296875,
0.78125,
0.55029296875,
0.78125,
0.564453125,
0.5458984375,
0.564453125
],
[
0.0882568359375,
0.5595703125,
0.484375,
0.5595703125,
0.484375,
0.5693359375,
0.0882568359375,
0.5693359375
],
[
0.0849609375,
0.5712890625,
0.48779296875,
0.5712890625,
0.48779296875,
0.58544921875,
0.0849609375,
0.58544921875
],
[
0.53173828125,
0.568359375,
0.91015625,
0.568359375,
0.91015625,
0.58203125,
0.53173828125,
0.58203125
],
[
0.0849609375,
0.58544921875,
0.486083984375,
0.58544921875,
0.486083984375,
0.5986328125,
0.0849609375,
0.5986328125
],
[
0.5439453125,
0.58349609375,
0.9111328125,
0.58349609375,
0.9111328125,
0.5966796875,
0.5439453125,
0.5966796875
],
[
0.0849609375,
0.599609375,
0.484375,
0.599609375,
0.484375,
0.61279296875,
0.0849609375,
0.61279296875
],
[
0.54345703125,
0.5966796875,
0.91259765625,
0.59765625,
0.91259765625,
0.611328125,
0.54345703125,
0.60986328125
],
[
0.0849609375,
0.61279296875,
0.48779296875,
0.61279296875,
0.48779296875,
0.626953125,
0.0849609375,
0.626953125
],
[
0.5458984375,
0.611328125,
0.91259765625,
0.611328125,
0.91259765625,
0.625,
0.5458984375,
0.625
],
[
0.0849609375,
0.626953125,
0.486083984375,
0.626953125,
0.486083984375,
0.64013671875,
0.0849609375,
0.64013671875
],
[
0.5458984375,
0.62548828125,
0.5888671875,
0.62548828125,
0.5888671875,
0.63720703125,
0.5458984375,
0.63720703125
],
[
0.08660888671875,
0.64013671875,
0.484375,
0.6416015625,
0.484375,
0.65478515625,
0.08660888671875,
0.65380859375
],
[
0.53125,
0.64306640625,
0.91259765625,
0.64306640625,
0.91259765625,
0.6572265625,
0.53125,
0.6572265625
],
[
0.08660888671875,
0.65576171875,
0.484375,
0.65576171875,
0.484375,
0.66650390625,
0.08660888671875,
0.66650390625
],
[
0.54150390625,
0.6572265625,
0.91259765625,
0.65576171875,
0.91259765625,
0.67041015625,
0.54150390625,
0.671875
],
[
0.0849609375,
0.66845703125,
0.486083984375,
0.66845703125,
0.486083984375,
0.681640625,
0.0849609375,
0.681640625
],
[
0.54345703125,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.68505859375,
0.54345703125,
0.68505859375
],
[
0.0849609375,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.5458984375,
0.6875,
0.64892578125,
0.6875,
0.64892578125,
0.69775390625,
0.5458984375,
0.69775390625
],
[
0.08660888671875,
0.69775390625,
0.484375,
0.69775390625,
0.484375,
0.70849609375,
0.08660888671875,
0.70849609375
],
[
0.53369140625,
0.7041015625,
0.91259765625,
0.7041015625,
0.91259765625,
0.71728515625,
0.53369140625,
0.71728515625
],
[
0.08660888671875,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.08660888671875,
0.7236328125
],
[
0.0849609375,
0.7236328125,
0.486083984375,
0.7236328125,
0.486083984375,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.54736328125,
0.7197265625,
0.9111328125,
0.7197265625,
0.9111328125,
0.72998046875,
0.54736328125,
0.72998046875
],
[
0.5439453125,
0.73193359375,
0.91259765625,
0.73193359375,
0.91259765625,
0.74560546875,
0.5439453125,
0.74560546875
],
[
0.08660888671875,
0.73779296875,
0.484375,
0.73779296875,
0.484375,
0.751953125,
0.08660888671875,
0.751953125
],
[
0.0849609375,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.76513671875,
0.0849609375,
0.76513671875
],
[
0.5458984375,
0.7470703125,
0.9111328125,
0.7470703125,
0.9111328125,
0.7568359375,
0.5458984375,
0.7568359375
],
[
0.54345703125,
0.75830078125,
0.91259765625,
0.7587890625,
0.91259765625,
0.7734375,
0.54345703125,
0.77294921875
],
[
0.08660888671875,
0.7666015625,
0.48193359375,
0.7666015625,
0.48193359375,
0.7763671875,
0.08660888671875,
0.7763671875
],
[
0.5439453125,
0.7734375,
0.9140625,
0.7734375,
0.9140625,
0.787109375,
0.5439453125,
0.787109375
],
[
0.08660888671875,
0.78076171875,
0.486083984375,
0.78076171875,
0.486083984375,
0.79150390625,
0.08660888671875,
0.79150390625
],
[
0.5458984375,
0.7861328125,
0.9111328125,
0.787109375,
0.9111328125,
0.80029296875,
0.5458984375,
0.7998046875
],
[
0.0849609375,
0.79248046875,
0.484375,
0.79248046875,
0.484375,
0.80615234375,
0.0849609375,
0.80615234375
],
[
0.54345703125,
0.7998046875,
0.9140625,
0.80029296875,
0.9140625,
0.81494140625,
0.54345703125,
0.814453125
],
[
0.08660888671875,
0.8095703125,
0.484375,
0.8095703125,
0.484375,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.0849609375,
0.82080078125,
0.486083984375,
0.82080078125,
0.486083984375,
0.833984375,
0.0849609375,
0.833984375
],
[
0.5458984375,
0.81640625,
0.611328125,
0.81640625,
0.611328125,
0.82763671875,
0.5458984375,
0.82763671875
],
[
0.0849609375,
0.83544921875,
0.484375,
0.83544921875,
0.484375,
0.8486328125,
0.0849609375,
0.8486328125
],
[
0.53125,
0.83349609375,
0.9111328125,
0.83349609375,
0.9111328125,
0.8466796875,
0.53125,
0.8466796875
],
[
0.5439453125,
0.84765625,
0.9140625,
0.84765625,
0.9140625,
0.861328125,
0.5439453125,
0.861328125
],
[
0.108642578125,
0.8623046875,
0.348876953125,
0.8623046875,
0.348876953125,
0.87548828125,
0.108642578125,
0.87548828125
],
[
0.5439453125,
0.861328125,
0.9111328125,
0.861328125,
0.9111328125,
0.875,
0.5439453125,
0.875
],
[
0.1070556640625,
0.87548828125,
0.435546875,
0.87548828125,
0.435546875,
0.8896484375,
0.1070556640625,
0.8896484375
],
[
0.5439453125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.5439453125,
0.8896484375
]
] |
[
"ing up a large-scale, diverse and highly abstract dataset. On",
"Tk1Th2",
"Tk",
"an average, the corpus has 12 answers per question thread,",
"FINALDISTRIBUTION",
"with 65 words per answer. All summaries are truncated at",
"hundred words. We split the 100k dataset into 80k training",
"",
"",
"instances, 10k validation and 10k test instances. We addition-",
"ATTENTIONDISTRIBUTION",
"VOCABULARYDISTRIBUTIO",
"allyextracttheupvoteinformationcorrespondingtoeveryan-",
"TOKENLEVELCOPYING",
"\"n",
"72",
"swer from the L6 dataset and assume that upvotes have a high",
"CONTEXTVECTOR",
"STRUCTURALATTENTION",
"correlationwiththerelativeimportanceandrelevanceofan",
"OPERLINGN",
"answer. We then rank answers in decreasing order of upvotes",
"before concatenating them as required for several baselines.",
"Since Yahoo! Answers is an unstructured and unmoderated",
"i1",
"11,2",
"Tin",
"question-answer repository, this has turned out to be a chal-",
"lenging summarization dataset [Chowdhury and Chakraborty",
"2019]. Additionally, we include analysis on MultiNews [Fab-",
"bri et al., 2019], a news based MSD corpora to aid similar",
"studies. It is the first large scale MSD news dataset consisting",
"ANSWERLEVEL",
"DECODER",
"FVF",
"u",
"NCODE",
"of 56,216 article summary pairs, crowd-sourced from various",
"differentnewswebsites",
"Figure 2: Hierarchical encoder with structural attention + multi",
"level contextual attention for multi-document summarization.First,",
"L",
"Competing Methods",
"token-level structured representations are computed. For each an-",
"swer,theserepresentations arepooled(r:)andfed to an answer-",
"We compare the performance of the following SDS and MDS",
"level encoder.Structural infused answerlevel embeddings(qi)are",
"models.",
"calculated acrossanswers.Ateverydecodertimestep,wecalculate",
"the context vector and probability Pgen based on structural contex-",
"Lead3: It is an extractive baseline where the first 100 to-",
"tual attention.We calculate the copying vector (source attention)by",
"kens of the document (in case of SDS datasets) and con-",
" projecting token-level attention onto the vocabulary.",
"catenated ranked documents (in case of MDS datasets)",
"are picked to form the summary.",
". KL-Summ: It is an extractive summarization method",
"(i) The CNN/Dailymail' dataset [Hermann et al., 2015;",
"introducedby[Haghighiand Vanderwende.2009]that",
"Nallapati et al.,2016] is a news corpora containing",
"attempts to minimize KL-Divergence between candidate",
"document-summary pairs. Bulleted extracts from the CNN",
"documents and generated summary.",
"andDailymailnews pages onlineareprojected as summaries",
": LexRank: It is an unsupervised extractive summariza-",
"to the remaining documents. The scripts released by [Nal-",
"tion method [Erkan and Radev, 2004]. A graph is built",
"lapati et al., 2016] are used to extract approximately 250k",
"with sentences as vertices, and edge weights are as-",
"training pairs, 13k validation pairs and 11.5k test pairs from",
"signed based on sentence similarity.",
"the corpora.We use the non-anonymized form of the data",
"to provide fair comparison to the experiments, conducted",
": TextRank: It is an unsupervised extractive summariza-",
"by [See et al., 2017]. It is a factual, English corpora with",
"tion method which selects sentences such that the infor-",
" an average document length being 781 tokens and an aver-",
"mation being disseminated by the summary is as close as",
"age summary length being 56 tokens. We use two versions",
"possible to the original documents [Mihalcea and Tarau,",
"of this dataset - one with 400 word articles, and the other",
"20041.",
"with 800 word articles.Most research reporting results on",
": Pointer-Generator (PG): It is a supervised abstractive",
"this dataset are on CNN/Dailymail-400.We also consider",
"summarization model [See et al., 2017], as discussed",
"a 800 token version of this dataset as longer articles harbor",
"earlier. It is a strong and popularly used baseline for",
"more intra-document structural dependencies which would",
"summarization.",
"allowusto better demonstrate thebenefits of structure incor-",
".Pointer-Generator+ Structure Infused Copy",
"poration. Moreover, longer documents resemble real-world",
"datasets. (ii) We also use the CQA dataset? [Chowdhury and",
"(PG+SC):Our implementationis similartooneofthe",
"methods proposed by [Song et al., 2018]. We explicitly",
"Chakraborty, 2019] which is generated by filtering the Ya-",
"hoo! Answers L6 corpora to find question threads where the",
"compute the dependency tree of sentences and encode",
"a structure vector based on features like POs tag,",
"best answer can serve as a summary for the remaining an",
"number of incoming edges, depth of the tree, etc. We",
"swers. The authors use a series of heuristics to arrive at a",
"then concatenate this structural vector for every token",
"set of 100k question thread-summary pairs. The summaries",
"to its hidden state representation in pointer-generator",
"aregeneratedbymodifyingbestanswersand selectingmost",
"question-relevant sentences from them. The remaining an-",
"networks.",
"swers serve as candidate documents for summarization mak-",
". Pointer-Generator+MMR (PG+MMR): It is an ab-",
"stractive MDS model [Lebanoff et al., 2018] trained",
"'https://github.com/abisee/cnn-dailymail",
"on the CNN/Dailymail dataset. It combines Maxi-",
"*https:/bitbucket.org/tanya14109/cqasumm/src/master/",
"mal Marginal Relevance (MMR) method with pointer-"
] |
{
"arxiv_id": "2004.09739",
"language": "en",
"page_idx": 3
}
|
||
2303.05251-en-0
|
2303.05251-en-0.jpg
|
$\textbf{Masked Image Modeling with Local Multi- Scale Reconstruction}$
Haoqing Wang$^{1}$, Yehui Tang$^{1,2}$, Yunhe Wang$^{2*}$,Jianyuan Guo$^2$,Zhi-Hong Deng$^{1*}$,Kai Han$^{2}$
$^{1}$National Key Lab of General AI, School of Intelligence Science and Technology, Peking University
$^{2}$Huawei Noah's Ark Lab
$\{$wanghaoqing,yhtang,zhdeng$\}@$pku.edu.cn, $\{$kai.han, yunhe.wang$\}@$huawei.com
## $\mathbf{Abstract}$
Masked Image Modeling (MIM) achieves outstanding success in self- supervised representation learning. Unfortunately, MIM models typically have huge computational burden and slow learning process, which is an inevitable obstacle for their industrial applications. Although the lower layers play the key role in MIM, existing MIM models con-
$\sum_{\begin{array}{cc}&\text{layers play the key role in MIM, existing MIM models con-}\\&\text{duct reconstruction task only at the top layer of encoder.}\\&\text{The lower layers are not explicitly guided and the interac-}\end{array}}$ $\begin{array}{cccc}&tion\:among\:their\:patches\:is\:only\:used\:for\:calculating\:new\\0&activations.\:Considering\:the\:reconstruction\:task\:requires\\&non-trivial\:inter-patch\:interactions\:to\:reason\:target\:signals,\end{array}$
we apply it to multiple local layers including lower and upper layers. Further, since the multiple layers expect to learn the information of different scales, we design local multiscale reconstruction, where the lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively. This design not only accelerates the representation learning process by explicitly guiding multiple layers, but also facilitates multi-scale semantical understanding to the input. Extensive experiments show that with significantly less pre- training burden, our model achieves comparable or better performance on classification, detection and segmentation tasks than existing MIM models. Code is available with both MindSpore and PyTorch.
## $1. \textbf{Introduction}$
Recently, Masked Image Modeling (MIM) [2,24,60] achieves outstanding success in the field of self-supervised visual representation learning, which is inspired by the Masked Language Modeling (MLM) [4,34] in natural language processing and benefits from the development of vision transformers [18,39,55]. MIM learns semantic representations by first masking some parts of the input and then predicting their signals based on the unmasked parts, e.g., normalized pixels [24,60], discrete tokens [2,17], HOG fea-
$^{*}$Corresponding author.
ture[57], deep features [1,67] or frequencies [38,59].
Despite superior performance on various downstream tasks, these models have huge computational burden and slow learning process [31]. They typically require thousands of GPU Hours for pre-training on ImageNet-1K to get generalizing representations. Since we expect to pretrain these models on more massive amount of unlabeled data (e.g., free Internet data) to obtain more generalizing representations in practice, the pre-training efficiency is an inevitable bottleneck limiting the industrial applications of MIM. How to accelerate the representation learning in MIM is an important topic. To this end, MAE[24] pioneered the asymmetric encoder-decoder strategy, where the costly encoder only operates few visible patches and the lightweight decoder takes all the patches as input for prediction. Further, GreenMIM [31] extends the asymmetric encoder-decoder strategy to hierarchical vision transformers (e.g., Swin [39]). Besides, [8,22,35] shrinks the input resolution to lessen the input patches, thereby reducing the computational burden. However, they all aim to accelerate the encoding process rather than the representation learning.
In MIM, the learning of upper layers depends on that of lower ones during pre-training, since the upper-layer features are calculated from the lower layers. Besides, during fine-tuning the upper layers are typically tuned quickly to adapt to the downstream task while the lower ones change more slowly and need to be well-learned [2,29,65]. Even fine-tuning only the several upper layers and freezing the others can obtain similar performance [24]. Therefore, the lower layers of encoder play the key role in MIM. However, all existing MIM models only conduct reconstruction task at the top layer of encoder and the lower ones are not explicitly guide, thus the interaction among their patches is only used for calculating the activations of the next layer. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to both lower and upper layers to explicitly guide them and thus accelerate the overall learning process. Using tiny decoder is sufficient for each local reconstruction task and does not significantly increase the computational burden.
|
[
[
0.1666259765625,
0.1337890625,
0.8037109375,
0.1337890625,
0.8037109375,
0.1483154296875,
0.1666259765625,
0.1483154296875
],
[
0.131591796875,
0.18115234375,
0.83837890625,
0.18115234375,
0.83837890625,
0.1988525390625,
0.131591796875,
0.1988525390625
],
[
0.0906982421875,
0.1976318359375,
0.87841796875,
0.1988525390625,
0.87841796875,
0.216552734375,
0.0906982421875,
0.21533203125
],
[
0.3857421875,
0.216552734375,
0.58642578125,
0.2171630859375,
0.58642578125,
0.231689453125,
0.3857421875,
0.2310791015625
],
[
0.16015625,
0.236083984375,
0.80810546875,
0.236083984375,
0.80810546875,
0.250732421875,
0.16015625,
0.250732421875
],
[
0.0228729248046875,
0.26513671875,
0.056365966796875,
0.26513671875,
0.056365966796875,
0.375732421875,
0.0228729248046875,
0.375732421875
],
[
0.23779296875,
0.285400390625,
0.314453125,
0.285400390625,
0.314453125,
0.302978515625,
0.23779296875,
0.302978515625
],
[
0.50244140625,
0.28857421875,
0.8544921875,
0.28857421875,
0.8544921875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.521484375,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.322509765625,
0.521484375,
0.322509765625
],
[
0.098876953125,
0.32080078125,
0.4697265625,
0.32080078125,
0.4697265625,
0.333984375,
0.098876953125,
0.333984375
],
[
0.50244140625,
0.323974609375,
0.89208984375,
0.323974609375,
0.89208984375,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.08087158203125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.07843017578125,
0.348388671875,
0.468017578125,
0.349853515625,
0.468017578125,
0.3642578125,
0.07843017578125,
0.363037109375
],
[
0.50244140625,
0.354248046875,
0.89208984375,
0.354248046875,
0.89208984375,
0.367431640625,
0.50244140625,
0.367431640625
],
[
0.02777099609375,
0.36865234375,
0.053924560546875,
0.36865234375,
0.053924560546875,
0.389404296875,
0.02777099609375,
0.389404296875
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50244140625,
0.369384765625,
0.88818359375,
0.36865234375,
0.88818359375,
0.3818359375,
0.50244140625,
0.38330078125
],
[
0.08087158203125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50244140625,
0.3837890625,
0.890625,
0.3837890625,
0.890625,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.443115234375,
0.02532958984375,
0.443115234375
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.50244140625,
0.399658203125,
0.890625,
0.399658203125,
0.890625,
0.41357421875,
0.50244140625,
0.41357421875
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.080078125,
0.439453125
],
[
0.50341796875,
0.429931640625,
0.89208984375,
0.429931640625,
0.89208984375,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.03350830078125,
0.443115234375,
0.0555419921875,
0.443115234375,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.07843017578125,
0.455078125,
0.4697265625,
0.455810546875,
0.4697265625,
0.47021484375,
0.07843017578125,
0.4697265625
],
[
0.50244140625,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.50341796875,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.080078125,
0.486083984375,
0.4697265625,
0.486083984375,
0.4697265625,
0.5,
0.080078125,
0.5
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.0310516357421875,
0.5,
0.052276611328125,
0.5,
0.052276611328125,
0.5234375,
0.0310516357421875,
0.5234375
],
[
0.080078125,
0.501953125,
0.4697265625,
0.501953125,
0.4697265625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.0236968994140625,
0.517578125,
0.058013916015625,
0.517578125,
0.058013916015625,
0.58056640625,
0.0236968994140625,
0.58056640625
],
[
0.080078125,
0.5166015625,
0.46728515625,
0.5166015625,
0.46728515625,
0.52978515625,
0.080078125,
0.52978515625
],
[
0.50244140625,
0.51953125,
0.89208984375,
0.51953125,
0.89208984375,
0.53271484375,
0.50244140625,
0.53271484375
],
[
0.080078125,
0.5322265625,
0.46728515625,
0.5322265625,
0.46728515625,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.50341796875,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50341796875,
0.548828125
],
[
0.07843017578125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.07843017578125,
0.56005859375
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.50341796875,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.57763671875,
0.50341796875,
0.57763671875
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.0310516357421875,
0.58349609375,
0.053924560546875,
0.5869140625,
0.04412841796875,
0.6162109375,
0.021240234375,
0.6123046875
],
[
0.50244140625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.080078125,
0.59228515625,
0.468017578125,
0.59228515625,
0.468017578125,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.5009765625,
0.5947265625,
0.89208984375,
0.5966796875,
0.89208984375,
0.611328125,
0.5009765625,
0.609375
],
[
0.0228729248046875,
0.60400390625,
0.060455322265625,
0.60400390625,
0.060455322265625,
0.7041015625,
0.0228729248046875,
0.7041015625
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.51953125,
0.61279296875,
0.8935546875,
0.6142578125,
0.8935546875,
0.62890625,
0.51953125,
0.62744140625
],
[
0.080078125,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63623046875,
0.080078125,
0.63623046875
],
[
0.5009765625,
0.6298828125,
0.8896484375,
0.6298828125,
0.8896484375,
0.64306640625,
0.5009765625,
0.64306640625
],
[
0.08087158203125,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.6494140625,
0.08087158203125,
0.6494140625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.65283203125,
0.4697265625,
0.65283203125,
0.4697265625,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.080078125,
0.66748046875,
0.374267578125,
0.66748046875,
0.374267578125,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.50341796875,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50341796875,
0.689453125
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.08087158203125,
0.71337890625,
0.2091064453125,
0.71337890625,
0.2091064453125,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.098876953125,
0.7392578125,
0.468017578125,
0.7392578125,
0.468017578125,
0.75390625,
0.098876953125,
0.75390625
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.08087158203125,
0.7548828125,
0.4697265625,
0.7548828125,
0.4697265625,
0.76953125,
0.08087158203125,
0.76953125
],
[
0.50341796875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08087158203125,
0.771484375,
0.46728515625,
0.771484375,
0.46728515625,
0.7822265625,
0.08087158203125,
0.7822265625
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.78515625,
0.468017578125,
0.78515625,
0.468017578125,
0.798828125,
0.080078125,
0.798828125
],
[
0.5048828125,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.08087158203125,
0.80029296875,
0.468017578125,
0.80029296875,
0.468017578125,
0.814453125,
0.08087158203125,
0.814453125
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.81494140625,
0.468017578125,
0.81494140625,
0.468017578125,
0.8291015625,
0.080078125,
0.8291015625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.83056640625,
0.4697265625,
0.83056640625,
0.4697265625,
0.8447265625,
0.08087158203125,
0.8447265625
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.07843017578125,
0.8447265625,
0.4697265625,
0.8466796875,
0.4697265625,
0.861328125,
0.07843017578125,
0.859375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.861328125,
0.468017578125,
0.861328125,
0.468017578125,
0.875,
0.080078125,
0.875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.5009765625,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.09722900390625,
0.88623046875,
0.2254638671875,
0.88720703125,
0.2254638671875,
0.9013671875,
0.09722900390625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.736328125,
0.88720703125,
0.736328125,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Masked Image Modeling with Local Multi-Scale Reconstruction",
"Haoqing Wang', Yehui Tangl.2, Yunhe Wang2* Jianyuan Guo?, Zhi-Hong Deng Kai Han?",
"INational Key Lab of General Al, School of Intelligence Science and Technology, Peking University",
"2HuaweiNoah'sArkLab",
" wanghaoqing,yhtang,zhdeng @pku.edu.cn, kai.han, yunhe.wang @huawei.com",
"",
"Abstract",
"ture [57], deep features [1, 67] or frequencies [38, 59].",
"Despite superior performance on various downstream",
"Masked Image Modeling (MIM) achieves outstanding",
"tasks, these models have huge computational burden and",
"success in self-supervised representation learning. Unfortu-",
"slow learning process [31]. They typically require thou-",
"nately, MiM models typically have huge computational bur-",
"sands of GPU Hours for pre-training on ImageNet-1K to",
"9",
"den and slow learning process, which is an inevitable ob-",
"get generalizingrepresentations.Sincewe expect to pre",
"stacle for their industrial applications. Although the lower",
"train these models on more massive amount of unlabeled",
"LAD",
"layers play the key role in MIM, existing MIM models con-",
"data (e.g, free Internet data) to obtain more generalizing",
"duct reconstruction task only at the top layer of encoder.",
"representations in practice, the pre-training efficiency is",
"The lower layers are not explicitly guided and the interac-",
"an inevitable bottleneck limiting the industrial applications",
"'SO",
"tion among their patches is only used for calculating new",
"of MIM. How to accelerate the representation learning in",
"activations. Considering the reconstruction task requires",
"MIM is an important topic. To this end, MAE [24] pio-",
"non-trivial inter-patch interactions to reason target signals,",
"neered the asymmetric encoder-decoder strategy, where the",
"we apply it to multiple local layers including lower and up-",
"costly encoderonly operates few visiblepatches and the",
"V",
"per layers. Further, since the multiple layers expect to learn",
"lightweight decoder takes all the patches as input for pre-",
"IS0",
"the information of different scales, we design local multi-",
"diction. Further, GreenMIM [31] extends the asymmetric",
"scale reconstruction, where the lower and upper layers re-",
"encoder-decoder strategy to hierarchical vision transform-",
"construct fine-scale and coarse-scale supervision signals",
"ers (e.g., Swin [39]). Besides, [8, 22, 35] shrinks the input",
"respectively. This design not only accelerates the represen-",
"resolutiontolessentheinputpatches,therebyreducingthe",
"tation learning process by explicitly guiding multiple lay-",
"CO",
"computational burden. However, they all aim to accelerate",
"ers, but also facilitates multi-scale semantical understand-",
"the encoding process rather than the representation learning.",
"",
"ing to the input. Extensive experiments show that with sig-",
"In MIM, the learning of upper layers depends on that of",
"nificantly less pre-training burden, our model achieves com-",
"lower ones during pre-training, since the upper-layer fea-",
"parableorbetterperformance onclassification,detection",
"tures are calculated from the lower layers. Besides, during",
"and segmentation tasks than existing MIM models. Code is",
"fine-tuning the upper layers are typically tuned quickly to",
"available with both MindSpore and PyTorch.",
"adapt to the downstream task while the lower ones change",
"more slowly and need to be well-learned [2,29, 65]. Even",
"fine-tuning only the several upper layers and freezing the",
"1. Introduction",
"others can obtain similar performance [24]. Therefore, the",
"Recently, Masked Image Modeling (MIM) [2, 24, 60]",
"lowerlayersofencoderplaythekeyroleinMlM.However",
"achieves outstanding success in the field of self-supervised",
"all existing MIM models only conduct reconstruction task at",
"visual representation learning, which is inspired by the",
"the top layer of encoder and the lower ones are not explicitly",
"Masked Language Modeling (MLM) [4, 34] in natural lan-",
"guide,thus theinteraction among their patchesis onlyused",
"guage processing and benefits from the development of vi-",
"for calculating the activations of the next layer. Considering",
"sion transformers [18, 39, 55]. MIM learns semantic repre-",
"the reconstruction task requires non-trivial inter-patch in-",
"sentations by first masking some parts of the input and then",
"teractions to reason target signals, we apply it to both lower",
"predicting their signals based on the unmasked parts, e.g.,",
"and upper layers to explicitly guide them and thus accelerate",
"normalized pixels [24,60], discrete tokens [2, 17], HOG fea-",
"the overall learning process. Using tiny decoder is sufficient",
"for each local reconstruction task and does not significantly",
"*Corresponding author.",
"increase the computational burden.",
"-"
] |
{
"arxiv_id": "2303.05251",
"language": "en",
"page_idx": 0
}
|
||
2303.05251-en-1
|
2303.05251-en-1.jpg
|
<FigureHere>
Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla
V100-32G GPU, CUDA 10.2 and PyTorch 1.8. ‘GPU Hours’is the running time on single GPU.
1 fine-tuning accuracy on ImageNet-1K, LocalMIM achieves $84.0\%$ using ViT-B and 84.1% using SwinB with significantly less pre-training duration than existing MIM models. The obtained representations also achieve better generalization on detection and segmentation downstream tasks, as shown in Table 2 and 3.
# $2. \textbf{Related Works}$
How to properly conduct reconstruction tasks at multiple local layers is a non-trivial problem. For example, applying the top-layer reconstruction task to carefully chosen local layers of ViT [18] can not achieve meaningful improvement. In general, the lower layers exploit low-level information and the upper ones learn high-level information [20,44], so it is not appropriate to use the supervision signals of same scale for multiple local reconstruction tasks. Here'scale' is the spatial size of the supervision signals calculated from the divided input regions, e.g., the signals from the $p\times p$ regions in an input of $H\times W$ resolution has the scale of $\frac Hp\times\frac Wp.$ The fine-scale and coarse-scale supervisions typically contain low-level and high-level information of the input respectively, and these multi-scale supervisions from input are widely ignored by existing MIM models. To this end, we propose local multi-scale reconstruction where the lower and upper layers reconstruct fine-scale and coarse-scale supervisions respectively. This design not only accelerates the representation learning process, but also facilitates multi-scale semantical understanding to the input. When the decoded predictions have different scale with the supervisions (e.g., on ViT), we use the deconvolution/pool operations to rescale them. We also apply the asymmetric encoder-decoder strategy [24,31] for quick encoding. Our model, dubbed as LocalMIM, are illustrated in Fig. 2 (a).
Overall, we summarize our contributions as follows.
Masked Image Modeling. With the development of vision transformers [18,23,39,50,55], Masked Image Modeling (MIM) gradually replaces the dominant position of contrastive learning [10,25,54] in visual self-supervised representation learning due to its superior fine-tuning performance in various visual downstream tasks. Many target signals have been designed for the mask-prediction pretext task in MIM, such as normalized pixels [24,60], discrete tokens[2,17], HOG feature[57], deep features[1,67] or frequencies [38,59]. However, they are all only applied as single-scale supervisions for reconstruction. An inevitable bottleneck for the industrial applications of MIM is that these models typically require huge computational resources and long pre-training duration. To this end, some works accelerate the encoding process via the asymmetric encoder-decoder strategy [24,31] or lessening the input patches [8,35]. Only accelerating the encoding process sometimes doesn't really speed up the representation learning, like GreenMIM vs. SimMIM$_192$ in Fig. 1. In this work, we use local multi-scale reconstructions to explicitly guide multiple lower layers and thus accelerate the overall learning process. Our method is compatible with the above quick encoding approaches. ConvMAE[20] fuses the feature of local layers for the final reconstruction to explicitly guide them, but still applies single-scale supervision signals.
Locally supervised learning. Considering the biological brain learns mainly based on local information [6], some works [42,43,45,56] used local error signals to greedily optimize individual blocks of the backbone without global back-propagation. This greedy training procedure signif-
## · To the best of our knowledge, this is the first work in MIM to conduct local reconstructions and use multiscale supervisions from the input.
· Our model is architecture-agnostic and can be used in
both columnar and pyramid architectures.
$\bullet$ From extensive experiments, we find that 1) LocalMIM is more efficient than existing MIM models, as shown in Fig. 1 and Table 1. For example, LocalMIM achieves the best MAE result with $3.1\times$ acceleration on ViT-B and the best GreenMIM result with 6.4× acceleration on Swin-B.2) In terms of top-
2
|
[
[
0.13232421875,
0.08837890625,
0.1519775390625,
0.08837890625,
0.1519775390625,
0.10101318359375,
0.13232421875,
0.10101318359375
],
[
0.50732421875,
0.0921630859375,
0.515625,
0.09027099609375,
0.5185546875,
0.0953369140625,
0.5107421875,
0.0985107421875
],
[
0.111083984375,
0.0985107421875,
0.16259765625,
0.0985107421875,
0.16259765625,
0.1817626953125,
0.111083984375,
0.1817626953125
],
[
0.48193359375,
0.1029052734375,
0.53515625,
0.1029052734375,
0.53515625,
0.212158203125,
0.48193359375,
0.212158203125
],
[
0.1168212890625,
0.1654052734375,
0.16015625,
0.1685791015625,
0.1544189453125,
0.2139892578125,
0.111083984375,
0.2108154296875
],
[
0.348876953125,
0.1988525390625,
0.3857421875,
0.1988525390625,
0.3857421875,
0.208984375,
0.348876953125,
0.208984375
],
[
0.404296875,
0.199462890625,
0.443603515625,
0.199462890625,
0.443603515625,
0.20703125,
0.404296875,
0.20703125
],
[
0.69775390625,
0.1988525390625,
0.75732421875,
0.1988525390625,
0.75732421875,
0.208984375,
0.69775390625,
0.208984375
],
[
0.79150390625,
0.199462890625,
0.84716796875,
0.199462890625,
0.84716796875,
0.20703125,
0.79150390625,
0.20703125
],
[
0.3505859375,
0.2083740234375,
0.384033203125,
0.2083740234375,
0.384033203125,
0.21533203125,
0.3505859375,
0.21533203125
],
[
0.4150390625,
0.2083740234375,
0.4755859375,
0.2083740234375,
0.4755859375,
0.21533203125,
0.4150390625,
0.21533203125
],
[
0.701171875,
0.2083740234375,
0.75634765625,
0.2083740234375,
0.75634765625,
0.21533203125,
0.701171875,
0.21533203125
],
[
0.50732421875,
0.21337890625,
0.52392578125,
0.216552734375,
0.521484375,
0.2266845703125,
0.5048828125,
0.2235107421875
],
[
0.136474609375,
0.212158203125,
0.1544189453125,
0.21533203125,
0.1502685546875,
0.2266845703125,
0.134033203125,
0.2235107421875
],
[
0.365234375,
0.2183837890625,
0.476318359375,
0.2183837890625,
0.476318359375,
0.2291259765625,
0.365234375,
0.2291259765625
],
[
0.7294921875,
0.2183837890625,
0.7548828125,
0.2183837890625,
0.7548828125,
0.2298583984375,
0.7294921875,
0.2298583984375
],
[
0.69287109375,
0.2197265625,
0.70751953125,
0.2197265625,
0.70751953125,
0.2279052734375,
0.69287109375,
0.2279052734375
],
[
0.314453125,
0.2197265625,
0.361083984375,
0.2197265625,
0.361083984375,
0.2266845703125,
0.314453125,
0.2266845703125
],
[
0.77392578125,
0.2197265625,
0.79833984375,
0.2197265625,
0.79833984375,
0.2279052734375,
0.77392578125,
0.2279052734375
],
[
0.81884765625,
0.2197265625,
0.84228515625,
0.2197265625,
0.84228515625,
0.2279052734375,
0.81884765625,
0.2279052734375
],
[
0.2548828125,
0.2203369140625,
0.275390625,
0.2203369140625,
0.275390625,
0.2279052734375,
0.2548828125,
0.2279052734375
],
[
0.2296142578125,
0.2203369140625,
0.2484130859375,
0.2203369140625,
0.2484130859375,
0.2279052734375,
0.2296142578125,
0.2279052734375
],
[
0.56201171875,
0.2203369140625,
0.576171875,
0.2203369140625,
0.576171875,
0.2291259765625,
0.56201171875,
0.2291259765625
],
[
0.60546875,
0.2203369140625,
0.62109375,
0.2203369140625,
0.62109375,
0.2291259765625,
0.60546875,
0.2291259765625
],
[
0.64208984375,
0.2235107421875,
0.734375,
0.2259521484375,
0.734375,
0.2392578125,
0.64208984375,
0.2374267578125
],
[
0.268798828125,
0.2259521484375,
0.361083984375,
0.2266845703125,
0.361083984375,
0.238037109375,
0.268798828125,
0.2374267578125
],
[
0.176513671875,
0.2215576171875,
0.188720703125,
0.2215576171875,
0.188720703125,
0.2259521484375,
0.176513671875,
0.2259521484375
],
[
0.271240234375,
0.240478515625,
0.328369140625,
0.240478515625,
0.328369140625,
0.255126953125,
0.271240234375,
0.255126953125
],
[
0.64208984375,
0.242431640625,
0.70166015625,
0.242431640625,
0.70166015625,
0.25390625,
0.64208984375,
0.25390625
],
[
0.080078125,
0.266357421875,
0.89208984375,
0.266357421875,
0.89208984375,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.080078125,
0.281005859375,
0.650390625,
0.281005859375,
0.650390625,
0.294189453125,
0.080078125,
0.294189453125
],
[
0.098876953125,
0.31005859375,
0.468017578125,
0.311279296875,
0.468017578125,
0.32568359375,
0.098876953125,
0.324462890625
],
[
0.5361328125,
0.311279296875,
0.89208984375,
0.311279296875,
0.89208984375,
0.324462890625,
0.5361328125,
0.324462890625
],
[
0.080078125,
0.326904296875,
0.46728515625,
0.326904296875,
0.46728515625,
0.34033203125,
0.080078125,
0.34033203125
],
[
0.5361328125,
0.326904296875,
0.890625,
0.326904296875,
0.890625,
0.34033203125,
0.5361328125,
0.34033203125
],
[
0.080078125,
0.34228515625,
0.468017578125,
0.34228515625,
0.468017578125,
0.35595703125,
0.080078125,
0.35595703125
],
[
0.53515625,
0.34228515625,
0.890625,
0.34228515625,
0.890625,
0.35595703125,
0.53515625,
0.35595703125
],
[
0.080078125,
0.356689453125,
0.468017578125,
0.356689453125,
0.468017578125,
0.37060546875,
0.080078125,
0.37060546875
],
[
0.53515625,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.53515625,
0.37060546875
],
[
0.080078125,
0.37255859375,
0.4697265625,
0.37255859375,
0.4697265625,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.53515625,
0.37255859375,
0.89208984375,
0.37255859375,
0.89208984375,
0.386474609375,
0.53515625,
0.386474609375
],
[
0.07843017578125,
0.386474609375,
0.4697265625,
0.386962890625,
0.4697265625,
0.401611328125,
0.07843017578125,
0.40087890625
],
[
0.5361328125,
0.389404296875,
0.875,
0.389404296875,
0.875,
0.399658203125,
0.5361328125,
0.399658203125
],
[
0.080078125,
0.40283203125,
0.4697265625,
0.40283203125,
0.4697265625,
0.416748046875,
0.080078125,
0.416748046875
],
[
0.08087158203125,
0.419189453125,
0.46728515625,
0.419189453125,
0.46728515625,
0.429931640625,
0.08087158203125,
0.429931640625
],
[
0.50244140625,
0.415283203125,
0.650390625,
0.415283203125,
0.650390625,
0.4287109375,
0.50244140625,
0.4287109375
],
[
0.080078125,
0.43310546875,
0.46728515625,
0.43310546875,
0.46728515625,
0.4462890625,
0.080078125,
0.4462890625
],
[
0.080078125,
0.448974609375,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.080078125,
0.462158203125
],
[
0.50244140625,
0.4443359375,
0.890625,
0.4443359375,
0.890625,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.080078125,
0.462158203125,
0.4697265625,
0.462158203125,
0.4697265625,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.50244140625,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.0767822265625,
0.474853515625,
0.4697265625,
0.474853515625,
0.4697265625,
0.492431640625,
0.0767822265625,
0.492431640625
],
[
0.50244140625,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.07843017578125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.07843017578125,
0.5068359375
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.07843017578125,
0.50830078125,
0.4697265625,
0.50830078125,
0.4697265625,
0.521484375,
0.07843017578125,
0.521484375
],
[
0.5,
0.50390625,
0.89208984375,
0.5048828125,
0.89208984375,
0.51953125,
0.5,
0.51806640625
],
[
0.07843017578125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.07843017578125,
0.537109375
],
[
0.5009765625,
0.51953125,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.5009765625,
0.53271484375
],
[
0.07843017578125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.55322265625,
0.07843017578125,
0.55322265625
],
[
0.50244140625,
0.53515625,
0.890625,
0.5341796875,
0.890625,
0.548828125,
0.50244140625,
0.5498046875
],
[
0.080078125,
0.5537109375,
0.4697265625,
0.5537109375,
0.4697265625,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50341796875,
0.5517578125,
0.88623046875,
0.5517578125,
0.88623046875,
0.5625,
0.50341796875,
0.5625
],
[
0.08087158203125,
0.57080078125,
0.46728515625,
0.57080078125,
0.46728515625,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.5009765625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.5009765625,
0.5791015625
],
[
0.07843017578125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.07843017578125,
0.59765625
],
[
0.5009765625,
0.580078125,
0.8896484375,
0.580078125,
0.8896484375,
0.59326171875,
0.5009765625,
0.59326171875
],
[
0.08087158203125,
0.60107421875,
0.46728515625,
0.60107421875,
0.46728515625,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.5009765625,
0.59521484375,
0.89208984375,
0.5947265625,
0.89208984375,
0.609375,
0.5009765625,
0.60986328125
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.62548828125,
0.89208984375,
0.62548828125,
0.89208984375,
0.6396484375,
0.50244140625,
0.6396484375
],
[
0.080078125,
0.64306640625,
0.4697265625,
0.64306640625,
0.4697265625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.5009765625,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.5009765625,
0.65380859375
],
[
0.07843017578125,
0.6591796875,
0.4697265625,
0.6591796875,
0.4697265625,
0.67236328125,
0.07843017578125,
0.67236328125
],
[
0.50244140625,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.07843017578125,
0.67236328125,
0.455810546875,
0.673828125,
0.455810546875,
0.68798828125,
0.07843017578125,
0.6875
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.098876953125,
0.689453125,
0.44287109375,
0.689453125,
0.44287109375,
0.70263671875,
0.098876953125,
0.70263671875
],
[
0.50341796875,
0.68603515625,
0.89208984375,
0.68603515625,
0.89208984375,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.1004638671875,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.7265625,
0.1004638671875,
0.7265625
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.1094970703125,
0.72900390625,
0.4697265625,
0.72998046875,
0.4697265625,
0.744140625,
0.1094970703125,
0.74365234375
],
[
0.50341796875,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.111083984375,
0.74560546875,
0.3349609375,
0.74560546875,
0.3349609375,
0.7587890625,
0.111083984375,
0.7587890625
],
[
0.50244140625,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.50244140625,
0.76220703125,
0.89208984375,
0.76220703125,
0.89208984375,
0.77587890625,
0.50244140625,
0.77587890625
],
[
0.098876953125,
0.77099609375,
0.468017578125,
0.77099609375,
0.468017578125,
0.7841796875,
0.098876953125,
0.7841796875
],
[
0.5009765625,
0.7763671875,
0.89208984375,
0.77587890625,
0.89208984375,
0.79052734375,
0.5009765625,
0.79150390625
],
[
0.111083984375,
0.7861328125,
0.386474609375,
0.7861328125,
0.386474609375,
0.7998046875,
0.111083984375,
0.7998046875
],
[
0.50244140625,
0.79150390625,
0.89208984375,
0.79150390625,
0.89208984375,
0.8046875,
0.50244140625,
0.8046875
],
[
0.098876953125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.5009765625,
0.806640625,
0.85693359375,
0.806640625,
0.85693359375,
0.82080078125,
0.5009765625,
0.82080078125
],
[
0.111083984375,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.111083984375,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.111083984375,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.111083984375,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.11273193359375,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.11273193359375,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.111083984375,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.111083984375,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.111083984375,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.111083984375,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"84.2",
"",
"ON",
"a",
"",
"O-BEIT",
"LoMaR",
"-O- SimMIMin2",
"LocalMIM (ours)",
"MAE",
"LocalMIM (ours)",
"GreenMIM",
"83.0",
"82.8",
"1600180020002200",
"1000",
"800",
"120051400",
"1200",
"1400",
"800",
"600",
"200",
"400",
"GPU Hours (h)",
"GPUHours (h)",
"200",
"(a) ViT-B",
"(b) Swin-B",
"Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla",
"V100-32G GPU, CUDA 10.2 and PyTorch 1.8. GPU Hours' is the running time on single GPU.",
"How to properly conduct reconstruction tasks at multi-",
"1 fine-tuning accuracy on ImageNet-1K, LocalMIM",
"ple local layers is a non-trivial problem. For example, ap-",
"achieves 84.0% using ViT-B and 84.1% using Swin-",
"plying the top-layer reconstruction task to carefully chosen",
"B with significantly less pre-training duration than ex-",
"local layers of ViT [18] can not achieve meaningful im-",
"isting MIM models. The obtained representations also",
"provement. In general, the lower layers exploit low-level",
"achieve better generalization on detection and segmen-",
"information and the upper ones learn high-level informa-",
"tationdownstreamtasks,as showninTable2and3.",
"tion [20,44], so it is not appropriate to use the supervision",
"signalsof samescaleformultiplelocalreconstructiontasks.",
"2.Related Works",
"Here 'scale' is the spatial size of the supervision signals cal-",
"culated from the divided input regions, e.g., the signals from",
"Masked Image Modeling. With the development of vi-",
"the p × p regions in an input of H × W resolution has the",
"sion transformers [18,23,39,50,55], Masked Image Mod-",
" scale of H × W. The fine-scale and coarse-scale super-",
"eling (MIM) gradually replaces the dominant position of",
"visions typically contain low-level and high-level informa-",
"contrastive learning [10,25,54] in visual self-supervised",
"tion of the input respectively, and these multi-scale supervi-",
"representation learning due to its superior fine-tuning per-",
" sions from input are widely ignored by existing MIM mod-",
"formance in various visual downstream tasks.Many tar-",
"els. To this end, we propose local multi-scale reconstruction",
"get signals have been designed for the mask-prediction pre-",
"where the lower and upper layers reconstruct fine-scale and",
"text task in MIM, such as normalized pixels[24,60], dis",
"coarse-scalesupervisionsrespectively.Thisdesignnotonly",
"crete tokens [2, 17], HOG feature [57], deep features [1, 67]",
"accelerates the representation learning process, but also fa-",
"or frequencies [38, 59].However, they are all only ap-",
"cilitates multi-scale semantical understanding to the input.",
"plied as single-scale supervisions for reconstruction. An in-",
"When the decoded predictions have different scale with the",
"evitablebottleneckfortheindustrialapplicationsofMIMis",
"supervisions (e.g., on ViT), we use the deconvolution/pool",
"that these models typically require huge computational re-",
"operations to rescale them. We also apply the asymmetric",
"sources and long pre-training duration. To this end, some",
"encoder-decoder strategy [24,31] for quick encoding. Our",
"works accelerate the encoding process via the asymmet-",
"model, dubbed as LocalMIM, are illustrated in Fig. 2 (a).",
"ric encoder-decoder strategy [24, 31] or lessening the in-",
"Overall, we summarize our contributions as follows.",
"put patches [8, 35]. Only accelerating the encoding process",
"sometimes doesn't really speed up the representation learn-",
"•To the best of our knowledge.this is the first work in",
"ing, like GreenMIM vs. SimMIMi92 in Fig. 1. In this work,",
"MIM to conduct local reconstructions and use multi-",
"weuselocal multi-scalereconstructionsto explicitlyguide",
"scale supervisions from the input.",
"multiple lower layers and thus accelerate the overall learn-",
"ing process. Our method is compatible with the above quick",
" Our model is architecture-agnostic and can be used in",
"encoding approaches. ConvMAE [20] fuses the feature of",
"both columnar and pyramid architectures.",
"local layers for the final reconstruction to explicitly guide",
"From extensive experiments, we find that 1) Lo-",
"them, but still applies single-scale supervision signals.",
"calMIM is more efficient than existing MIM models,",
"Locally supervised learning. Considering the biological",
"as shown in Fig. 1 and Table 1. For example, Lo-",
"brain learns mainly based on local information [6], some",
"calMIM achieves the best MAE result with 3.1x ac-",
"works [42, 43, 45, 56] used local error signals to greedily",
"celeration on ViT-B and the best GreenMIM result",
"optimize individual blocks of the backbone without global",
"with 6.4x acceleration on Swin-B. 2) In terms of top-",
"back-propagation. This greedy training procedure signif-",
"2"
] |
{
"arxiv_id": "2303.05251",
"language": "en",
"page_idx": 1
}
|
||
2303.05251-en-3
|
2303.05251-en-3.jpg
|
the encoder is used in various downstream tasks.
Decoder. The decoder takes as the input of both encoded visible patches $z_v^L$ and mask tokens $\{e_[\mathbb{M}]|m_i^L=1\}$, where the mask token $e_{[\mathrm{M}]}$ is a shared and learnable vector. The positional embeddings are also added to them for providing the location information. Since the decoder is only used during pre-training to output the prediction $\hat{y}^L$,it can be any architectures which support the global information propagation among patches, e.g., a series of ViT or Swin blocks. Many works [24,31,35] show that using lightweight decoders is sufficient to learn generalizing representations.
Global reconstruction. All existing MIM models predict the supervision signal $y$ based on the final output feature $z_v^L$ of the encoder and minimize a global reconstruction loss
$$\mathcal{L}_{MIM}=-\sum_{i=1}^{N_L}m_i^L\cdot\ln P(y_i|\hat{y}_i^L)$$
(1)
where the loss is calculated on the masked patches with $m_{i}^{L}=1$,and $P(\cdot|\cdot)$ can be the Gaussian or Dirichlet distributions for regression or classification losses respectively.
## 3.2. Analysis and motivation
For pre-training, the upper-layer features are calculated from the lower layers, so the well-learned lower layers can propagate semantical knowledge to the upper ones and facilitate their learning. For fine-tuning, the upper layers typically adapt quickly to specific downstream tasks, while the lower ones change more slowly and need to be sufficient learned during pre-training[2,29,65].Even fine-tuning only task are directly calculated from the input. Concretely, we the several upper layers and freezing the others can obtain evenly dividers can obtain evenly divide the input image $x\in\mathbb{R}^{H\times W\times C}$
of the encoder play the key role in MIM.
After patchification and linearly projection, the initial patch embeddings lose the inter-patch semantical relations. The self-attention mechanism in vision transformers is responsible for learning these relations by inter-patch interaction and build a better representation space than pixel space[5]. Further, since the self-attention mechanism has the computational complexity with a quadratic dependence on patch number $N$, it is difficult to learn the inter-patch interactions, especially for the lower layers of pyramidal architectures where the small patch size $p$ leads to huge $N.$ However, under the global reconstruction loss, the interpatch interaction at lower layers is not explicitly guided, and the simple task of calculating new activations is not sufficient to guide it. As a result, it is hard for the patches at lower layers to learn inter-patch relations. Concretely, we examine Normalized Mutual Information (NMI)[49] between query and key patches at each layer, where the high NMI value means the attention maps strongly depend on the query patches. Intuitively, the semantical representations
<FigureHere>
Figure 3. Normalized Mutual Information (NMI) between query
and key patches at each layer of a pre-trained ViT-B.
should have highly query-adaptive attentions, i.e., the different query patches faithfully attend to their semantically related regions. This is an advantage of the self-attention mechanism. As shown in Fig. 3, existing MIM models with global loss have small NMI value at lower layers, which means their patches have less query-adaptive attentions.
The reconstruction task requires holistic reasoning among patches to predict the masked signals and thus obtains semantic understanding to the input. Since this challenging task facilitates non-trivial inter-patch interactions, we apply it at multiple local layers, including both lower and upper ones to explicitly guide them all.
3.3. Local multi-scale reconstruction
In MIM, the supervision signals for a reconstruction feature descriptor $\pi$ to extract the supervision signal $y_i=$ $\pi(x_i).$ To learn generalizing representations, many feature descriptors have been designed, such as pixel normalization [24,60], HOG [57] and the pre-trained or momentum teacher [1,2,17,67]. We define the scale of supervision $y$ as $\begin{array}{l}{\mathrm{its~spatial~size~\frac{H}{p}\times\frac{W}{p}.~For~a~given~input,~the~fine-scale~su-}}\\{\mathrm{pervisions~from~finely-divided~input~regions~typically~con-}}\end{array}$ tain the low-level semantical information of the input, like corners, edges or textures. Relatively, the coarse-scale supervisions capture high-level semantical information of the input, like the shape of partial or whole object. Intuitively, multi-scale supervisions can guide representation learning better than the common single-scale ones due to their richer semantic information. In this work, we mainly consider the feature descriptors which are readily available without extra pre-training burden and costly forward inference of teacher networks, e.g., pixel normalization and HOG.
As shown in Table 4f, directly applying the top-layer reconstruction task at carefully chosen local layers of ViT can not achieve meaningful improvement, where each local task
4
|
[
[
0.080078125,
0.091552734375,
0.396240234375,
0.091552734375,
0.396240234375,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
0.68310546875,
0.0985107421875,
0.6953125,
0.0985107421875,
0.6953125,
0.10162353515625,
0.68310546875,
0.10162353515625
],
[
0.080078125,
0.11236572265625,
0.4697265625,
0.11236572265625,
0.4697265625,
0.1256103515625,
0.080078125,
0.1256103515625
],
[
0.53759765625,
0.111083984375,
0.5556640625,
0.111083984375,
0.5556640625,
0.11553955078125,
0.53759765625,
0.11553955078125
],
[
0.080078125,
0.1275634765625,
0.468017578125,
0.1275634765625,
0.468017578125,
0.1427001953125,
0.080078125,
0.1427001953125
],
[
0.53173828125,
0.1319580078125,
0.55810546875,
0.1287841796875,
0.560546875,
0.1395263671875,
0.53515625,
0.1427001953125
],
[
0.080078125,
0.143310546875,
0.468017578125,
0.143310546875,
0.468017578125,
0.1572265625,
0.080078125,
0.1572265625
],
[
0.53759765625,
0.141357421875,
0.54345703125,
0.141357421875,
0.54345703125,
0.1483154296875,
0.53759765625,
0.1483154296875
],
[
0.080078125,
0.1578369140625,
0.468017578125,
0.1578369140625,
0.468017578125,
0.1717529296875,
0.080078125,
0.1717529296875
],
[
0.53369140625,
0.154052734375,
0.5537109375,
0.149658203125,
0.55615234375,
0.160400390625,
0.53515625,
0.1634521484375
],
[
0.080078125,
0.173583984375,
0.4697265625,
0.173583984375,
0.4697265625,
0.1875,
0.080078125,
0.1875
],
[
0.080078125,
0.1881103515625,
0.468017578125,
0.1881103515625,
0.468017578125,
0.20263671875,
0.080078125,
0.20263671875
],
[
0.53759765625,
0.184326171875,
0.54345703125,
0.184326171875,
0.54345703125,
0.1900634765625,
0.53759765625,
0.1900634765625
],
[
0.07843017578125,
0.203857421875,
0.468017578125,
0.203857421875,
0.468017578125,
0.2171630859375,
0.07843017578125,
0.2171630859375
],
[
0.07843017578125,
0.2183837890625,
0.46728515625,
0.2171630859375,
0.46728515625,
0.231689453125,
0.07843017578125,
0.23291015625
],
[
0.080078125,
0.2342529296875,
0.46728515625,
0.2342529296875,
0.46728515625,
0.2474365234375,
0.080078125,
0.2474365234375
],
[
0.5458984375,
0.238037109375,
0.560546875,
0.238037109375,
0.560546875,
0.24560546875,
0.5458984375,
0.24560546875
],
[
0.080078125,
0.25,
0.44775390625,
0.25,
0.44775390625,
0.26318359375,
0.080078125,
0.26318359375
],
[
0.69287109375,
0.246826171875,
0.72607421875,
0.250732421875,
0.7236328125,
0.261474609375,
0.689453125,
0.2568359375
],
[
0.08087158203125,
0.268310546875,
0.4697265625,
0.268310546875,
0.4697265625,
0.2822265625,
0.08087158203125,
0.2822265625
],
[
0.50244140625,
0.2734375,
0.890625,
0.2734375,
0.890625,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.2841796875,
0.465576171875,
0.2841796875,
0.465576171875,
0.29736328125,
0.080078125,
0.29736328125
],
[
0.5009765625,
0.28857421875,
0.81396484375,
0.28857421875,
0.81396484375,
0.3017578125,
0.5009765625,
0.3017578125
],
[
0.07843017578125,
0.298583984375,
0.451904296875,
0.298583984375,
0.451904296875,
0.3125,
0.07843017578125,
0.3125
],
[
0.50341796875,
0.3193359375,
0.8896484375,
0.3193359375,
0.8896484375,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.2484130859375,
0.328857421875,
0.268798828125,
0.328857421875,
0.268798828125,
0.339111328125,
0.2484130859375,
0.339111328125
],
[
0.259033203125,
0.33837890625,
0.386474609375,
0.33837890625,
0.386474609375,
0.356689453125,
0.259033203125,
0.356689453125
],
[
0.50244140625,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.25,
0.341552734375,
0.2607421875,
0.341552734375,
0.2607421875,
0.34521484375,
0.25,
0.34521484375
],
[
0.4453125,
0.341552734375,
0.4697265625,
0.341552734375,
0.4697265625,
0.354736328125,
0.4453125,
0.354736328125
],
[
0.2320556640625,
0.34521484375,
0.23779296875,
0.34521484375,
0.23779296875,
0.35107421875,
0.2320556640625,
0.35107421875
],
[
0.160888671875,
0.34033203125,
0.236083984375,
0.34228515625,
0.236083984375,
0.356689453125,
0.160888671875,
0.354736328125
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.2467041015625,
0.359130859375,
0.26708984375,
0.359130859375,
0.26708984375,
0.367431640625,
0.2467041015625,
0.367431640625
],
[
0.50341796875,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.376953125,
0.50341796875,
0.376953125
],
[
0.080078125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.39453125,
0.080078125,
0.39453125
],
[
0.50244140625,
0.380126953125,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.080078125,
0.396484375,
0.4697265625,
0.396484375,
0.4697265625,
0.410400390625,
0.080078125,
0.410400390625
],
[
0.50244140625,
0.395751953125,
0.869140625,
0.395751953125,
0.869140625,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.080078125,
0.412353515625,
0.459228515625,
0.412353515625,
0.459228515625,
0.425537109375,
0.080078125,
0.425537109375
],
[
0.51953125,
0.4091796875,
0.89208984375,
0.410400390625,
0.89208984375,
0.4248046875,
0.51953125,
0.423583984375
],
[
0.5009765625,
0.425537109375,
0.89208984375,
0.4248046875,
0.89208984375,
0.439453125,
0.5009765625,
0.43994140625
],
[
0.080078125,
0.439453125,
0.298095703125,
0.439453125,
0.298095703125,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50244140625,
0.455078125
],
[
0.096435546875,
0.462158203125,
0.4697265625,
0.4609375,
0.4697265625,
0.47607421875,
0.096435546875,
0.4765625
],
[
0.50244140625,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.07843017578125,
0.477783203125,
0.4697265625,
0.479248046875,
0.4697265625,
0.492431640625,
0.07843017578125,
0.4912109375
],
[
0.50244140625,
0.486083984375,
0.78759765625,
0.486083984375,
0.78759765625,
0.5,
0.50244140625,
0.5
],
[
0.080078125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.5009765625,
0.51025390625,
0.78369140625,
0.51123046875,
0.78369140625,
0.52587890625,
0.5009765625,
0.5244140625
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.080078125,
0.5390625,
0.471435546875,
0.5390625,
0.471435546875,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.521484375,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.521484375,
0.548828125
],
[
0.07843017578125,
0.5537109375,
0.468017578125,
0.55322265625,
0.468017578125,
0.5673828125,
0.07843017578125,
0.568359375
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.5,
0.5625,
0.8935546875,
0.56005859375,
0.8935546875,
0.58056640625,
0.5,
0.58349609375
],
[
0.496826171875,
0.57568359375,
0.896484375,
0.57470703125,
0.896484375,
0.60107421875,
0.496826171875,
0.6015625
],
[
0.07843017578125,
0.583984375,
0.4697265625,
0.583984375,
0.4697265625,
0.59765625,
0.07843017578125,
0.59765625
],
[
0.080078125,
0.59765625,
0.346435546875,
0.59765625,
0.346435546875,
0.611328125,
0.080078125,
0.611328125
],
[
0.50244140625,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.1004638671875,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.1004638671875,
0.62890625
],
[
0.5009765625,
0.6142578125,
0.89208984375,
0.61279296875,
0.89208984375,
0.62744140625,
0.5009765625,
0.62890625
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.08251953125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.6572265625,
0.08251953125,
0.6572265625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.0767822265625,
0.673828125,
0.468017578125,
0.6748046875,
0.468017578125,
0.689453125,
0.0767822265625,
0.68798828125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5009765625,
0.68798828125,
0.89208984375,
0.689453125,
0.89208984375,
0.7041015625,
0.5009765625,
0.70263671875
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.705078125,
0.8935546875,
0.705078125,
0.8935546875,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73486328125
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.08087158203125,
0.75244140625,
0.46728515625,
0.75244140625,
0.46728515625,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.76513671875,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.5009765625,
0.77880859375
],
[
0.07843017578125,
0.7822265625,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.07843017578125,
0.79541015625
],
[
0.50341796875,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79248046875,
0.50341796875,
0.79248046875
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81005859375,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.5009765625,
0.82373046875
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.7998046875,
0.84130859375,
0.7998046875,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.521484375,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.521484375,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.5009765625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"the encoder is used in various downstream tasks",
"m",
"Decoder. The decoder takes as the input of both encoded",
"5.",
"visible patches z and mask tokens [em|m) = 1], where",
".50.5",
"the mask token eM is a shared and learnable vector. The",
"0",
"positional embeddings are also added to them for provid-",
"30",
"ing the location information. Since the decoder is only used",
"during pre-training to output the prediction g', it can be any",
"S",
"architectures which support the global information propa-",
"gation among patches, e.g., a series of ViT or Swin blocks.",
"Many works [24, 31, 35] show that using lightweight de-",
"0.0",
"coders is sufficient to learn generalizing representations.",
"Depth",
"Global reconstruction. All existing MIM models predict",
"Figure 3. Normalized Mutual Information (NMID) between query",
"the supervision signal y based on the final output feature z",
"and key patches at each layer of a pre-trained ViT-B.",
"of the encoder and minimize a global reconstruction loss",
"should have highly query-adaptive attentions, i.e., the dif-",
"NL",
"mt .In P(yilgt)",
"ferent query patches faithfully attend to their semantically",
"",
"(1)",
"-",
"CMIM=",
"related regions. This is an advantage of the self-attention",
"i=1",
"mechanism.As showninFig.3,existingMIM models with",
"where the loss is calculated on the masked patches with",
"global loss have small NMI value at lower layers, which",
"m; = 1, and P() can be the Gaussian or Dirichlet distri-",
"means their patches have less query-adaptive attentions.",
"butions for regression or classification losses respectively.",
"The reconstruction task requires holistic reasoning",
"among patches to predict the masked signals and thus ob-",
"3.2. Analysis and motivation",
"tains semantic understanding to the input. Since this chal-",
" For pre-training, the upper-layer features are calculated",
"lenging task facilitates non-trivial inter-patch interactions,",
"we apply it at multiple local layers, including both lower",
"from the lower layers, so the well-learned lower layers can",
"and upper ones to explicitly guide them all.",
"propagate semantical knowledge to the upper ones and fa-",
"cilitate their learning. For fine-tuning, the upper layers typ-",
"3.3.Local multi-scalereconstruction",
"ically adapt quickly to specific downstream tasks, while the",
"lower ones change more slowly and need to be sufficient",
"In MIM, the supervision signals for a reconstruction",
"learned during pre-training [2,29,65]. Even fine-tuning only",
"task are directly calculated from the input. Concretely, we",
"the several upper layers and freezing the others can obtain",
"evenly divide the input image a E IRHxWxC into non-",
"overlapping regions (2IRpxpx)W/P and usesome",
"comparable performance [24]. Therefore, the lower layers",
"of the encoder play the key role in MIM.",
"feature descriptor to extract the supervision signal yi :",
"After patchification and linearly projection, the initial",
"(ri). To learn generalizing representations, many feature",
"patch embeddings lose the inter-patch semantical relations.",
"descriptors have been designed, such as pixel normaliza-",
"The self-attention mechanism in vision transformers is re-",
"tion [24, 60], HOG [57] and the pre-trained or momentum",
"sponsible for learning these relations by inter-patch inter-",
"teacher [1,2, 17,67]. We define the scale of supervision y as",
" action and build a better representation space than pixel",
"its spatial size H × W. For a given input, the fine-scale su-",
"space [5]. Further, since the self-attention mechanism has",
"pervisions from finely-divided input regions typically con-",
"the computational complexity with a quadratic dependence",
"tain the low-level semantical information of the input, like",
"on patch number N, it is difficult to learn the inter-patch",
"corners, edges or textures. Relatively, the coarse-scale su-",
"interactions,especiallyforthelowerlayersofpyramidal ar-",
"pervisions capture high-level semantical information of the",
"chitectures where the small patch size p leads to huge N.",
"input, like the shape of partial or whole object. Intuitively,",
"However, under the global reconstruction loss, the inter-",
"multi-scale supervisions can guide representation learning",
"patch interaction at lower layers is not explicitly guided,",
"better thanthecommon single-scaleonesduetotheir richer",
"and the simple task of calculating new activations is not suf-",
"semantic information.In this work,wemainly considerthe",
"ficient to guide it. As a result, it is hard for the patches at",
"feature descriptors which are readily available without extra",
"lower layers to learn inter-patch relations. Concretely, we",
"pre-training burden and costly forward inference of teacher",
"examine Normalized Mutual Information (NMI) [49] be-",
"networks, e.g., pixel normalization and HOG.",
"tween query and key patches at each layer, where the high",
"As shown in Table 4f, directly applying the top-layer re-",
"NMI value means the attention maps strongly depend on the",
"construction task at carefully chosen local layers of ViT can",
"query patches. Intuitively, the semantical representations",
"not achieve meaningful improvement, where each local task",
"4"
] |
{
"arxiv_id": "2303.05251",
"language": "en",
"page_idx": 3
}
|
||
2309.04659-en-0
|
2309.04659-en-0.jpg
|
# Progressive Feature Adjustment for Semi-supervised Learning from Pretrained Models
Hai-Ming Xu$^{1}$ Lingqiao Liu$^{1\otimes}$ Hao Chen$^2$ Ehsan Abbasnejad$^{1}$ Rafael Felix$^{1}$
$^{1}$Australian Institute for Machine Learning, University of Adelaide $^{2}$Zhejiang University
$\{$hai-ming.xu, lingqiao.liu, ehsan.abbasnejad, rafael.felixalves$\}$@adelaide.edu.au
haochen.cad@z ju.edu.cn
## $\mathbf{Abstract}$
As an effective way to alleviate the burden of data annotation, semi-supervised learning (SSL) provides an attractive solution due to its ability to leverage both labeled and unlabeled data to build a predictive model. While signifi-
$\sum_{\begin{array}{lll}&\text{cant progress has beeen made recently, SSL algorithms are}\\&\text{offten evaluated and developed under the assumption that}\\&\text{the network is randonly initialized. This is in sharp contrast}\end{array}}$ $\begin{array}{lll}{&thenetwork\:is\:randomly\:initialized.\:This\:is\:in\:sharp\:contrast}\\{\dot{O}}&{to\:most\:vision\:recognition\:systems\:that\:are\:built\:from\:fine-}\\{\dot{O}}&{tuning\:a\:pretrained\:network\:for\:better\:performance.\:While}\end{array}$
the marriage of SSL and a pretrained model seems to be straightforward, recent literature suggests that naively applying state-of-the-art SSL with a pretrained model fails to unleash the full potential of training data. In this paper, we postulate the underlying reason is that the pretrained feature representation could bring a bias inherited from the source data, and the bias tends to be magnified through the self-training process in a typical SSL algorithm. To overcome this issue, we propose to use pseudo-labels from the unlabelled data to update the feature extractor that is less sensitive to incorrect labels and only allow the classifier to be trained from the labeled data. More specifically, we progressively adjust the feature extractor to ensure its inducedfeature distribution maintains a good class separability even under strong input perturbation. Through extensive experimental studies, we show that the proposed approach achieves superior performance over existing solutions.
<FigureHere>
Figure 1. Visualization of the correct keeping and error correcting
ability of FixMatch and ours approach with a pretrained model. The y-axis of the left figure denotes the percentage of the initially correctly labeled data keeping their correct class at a given iteration. The y-axis of the right fgure denotes the percentage of the initially incorrectly labeled data being predicted to the correct class at a given iteration. The experiment is conducted on $FGVC$ Aircraft dataset with 15% labels. As seen, FixMatch with a pretrained model shows weaker error correcting ability than the ours approach. Please refer to Section 4.1 for more details.
based on randomly initialized parameters. In recent years, the release and re-use of pretrained DNNs to alleviate training costs are becoming common practice for computer vision research and applications [42]. It seems that applying the existing SSL method to a pretrained model is straightforward. However, recent literatures [45,32] suggest that such a naive solution fails to unleash the full potential of training data and there seems to be a big room to improve the performance of SSL when a pretrained model is used.
In this study, we postulate the key issue preventing existing SSL solutions from attaining their full potential with pretrained models is due to the bias of the pretrained feature extractor: trained from the source domain data, e.g., ImageNet, the feature extractor may not be optimal for the target problem, e.g., fine-grained visual recognition. Such a bias could be a much more severe issue for SSL than for supervised learning. This is because the self-training (or pseudo-labeling) procedure commonly used in SSL tends to magnify the bias. For example, at the beginning of the training, a biased feature extractor and few labeled data could
## $1. \textbf{Introduction}$
Semi-supervised learning (SSL) is considered one of the most practical learning paradigms which can leverage both labeled and unlabeled samples to build a prediction model [4]. With the rapid development of deep neural networks (DNNs), extensive research on deep SSL methods [18,28,23,30,3,35,26] have been studied. Among those studies, most of them are evaluated and developed
|
[
[
0.30810546875,
0.1337890625,
0.66259765625,
0.1337890625,
0.66259765625,
0.1483154296875,
0.30810546875,
0.1483154296875
],
[
0.2320556640625,
0.1572265625,
0.73876953125,
0.1572265625,
0.73876953125,
0.1717529296875,
0.2320556640625,
0.1717529296875
],
[
0.146240234375,
0.20263671875,
0.82275390625,
0.20263671875,
0.82275390625,
0.2203369140625,
0.146240234375,
0.2203369140625
],
[
0.1282958984375,
0.2203369140625,
0.8408203125,
0.2215576171875,
0.8408203125,
0.2392578125,
0.1282958984375,
0.238037109375
],
[
0.1356201171875,
0.242431640625,
0.83837890625,
0.242431640625,
0.83837890625,
0.255615234375,
0.1356201171875,
0.255615234375
],
[
0.3857421875,
0.260009765625,
0.5849609375,
0.260009765625,
0.5849609375,
0.2734375,
0.3857421875,
0.2734375
],
[
0.0228729248046875,
0.270751953125,
0.062103271484375,
0.270751953125,
0.062103271484375,
0.37744140625,
0.0228729248046875,
0.37744140625
],
[
0.236083984375,
0.30810546875,
0.314453125,
0.30810546875,
0.314453125,
0.32568359375,
0.236083984375,
0.32568359375
],
[
0.80615234375,
0.3193359375,
0.8486328125,
0.3193359375,
0.8486328125,
0.323974609375,
0.80615234375,
0.323974609375
],
[
0.09722900390625,
0.34228515625,
0.46728515625,
0.34228515625,
0.46728515625,
0.356689453125,
0.09722900390625,
0.356689453125
],
[
0.08087158203125,
0.359130859375,
0.46728515625,
0.359130859375,
0.46728515625,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.02777099609375,
0.369384765625,
0.052276611328125,
0.369384765625,
0.052276611328125,
0.389404296875,
0.02777099609375,
0.389404296875
],
[
0.08087158203125,
0.373779296875,
0.4697265625,
0.373779296875,
0.4697265625,
0.386962890625,
0.08087158203125,
0.386962890625
],
[
0.08087158203125,
0.38818359375,
0.46728515625,
0.38818359375,
0.46728515625,
0.40283203125,
0.08087158203125,
0.40283203125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.44580078125,
0.02532958984375,
0.44580078125
],
[
0.080078125,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.080078125,
0.417236328125
],
[
0.08087158203125,
0.41845703125,
0.4697265625,
0.41845703125,
0.4697265625,
0.431884765625,
0.08087158203125,
0.431884765625
],
[
0.50341796875,
0.427978515625,
0.890625,
0.427978515625,
0.890625,
0.44140625,
0.50341796875,
0.44140625
],
[
0.08087158203125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.08087158203125,
0.447509765625
],
[
0.03350830078125,
0.442626953125,
0.053924560546875,
0.442626953125,
0.053924560546875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.50341796875,
0.44140625,
0.8896484375,
0.442626953125,
0.8896484375,
0.453857421875,
0.50341796875,
0.45263671875
],
[
0.080078125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.46337890625,
0.080078125,
0.46337890625
],
[
0.50244140625,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.07843017578125,
0.464111328125,
0.468017578125,
0.46337890625,
0.468017578125,
0.477783203125,
0.07843017578125,
0.479248046875
],
[
0.50244140625,
0.4697265625,
0.8896484375,
0.4697265625,
0.8896484375,
0.48291015625,
0.50244140625,
0.48291015625
],
[
0.080078125,
0.479736328125,
0.4697265625,
0.479736328125,
0.4697265625,
0.49365234375,
0.080078125,
0.49365234375
],
[
0.50244140625,
0.48291015625,
0.89208984375,
0.48291015625,
0.89208984375,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.080078125,
0.49560546875,
0.468017578125,
0.49560546875,
0.468017578125,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.50244140625,
0.496826171875,
0.89208984375,
0.496826171875,
0.89208984375,
0.51123046875,
0.50244140625,
0.51123046875
],
[
0.0767822265625,
0.5087890625,
0.4697265625,
0.50830078125,
0.4697265625,
0.52294921875,
0.0767822265625,
0.5234375
],
[
0.0228729248046875,
0.49755859375,
0.053924560546875,
0.496826171875,
0.056365966796875,
0.54541015625,
0.02532958984375,
0.546875
],
[
0.50341796875,
0.51220703125,
0.8896484375,
0.51220703125,
0.8896484375,
0.52294921875,
0.50341796875,
0.52294921875
],
[
0.080078125,
0.5244140625,
0.4697265625,
0.5244140625,
0.4697265625,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.50244140625,
0.5244140625,
0.8896484375,
0.5244140625,
0.8896484375,
0.53857421875,
0.50244140625,
0.53857421875
],
[
0.02532958984375,
0.537109375,
0.051483154296875,
0.5361328125,
0.053924560546875,
0.57080078125,
0.02777099609375,
0.5712890625
],
[
0.080078125,
0.5390625,
0.4697265625,
0.5390625,
0.4697265625,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.50244140625,
0.53857421875,
0.8935546875,
0.53857421875,
0.8935546875,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.07843017578125,
0.5537109375,
0.46728515625,
0.5537109375,
0.46728515625,
0.5673828125,
0.07843017578125,
0.5673828125
],
[
0.50244140625,
0.55322265625,
0.8203125,
0.5517578125,
0.8203125,
0.56494140625,
0.50244140625,
0.56640625
],
[
0.08087158203125,
0.5712890625,
0.46728515625,
0.5712890625,
0.46728515625,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.0310516357421875,
0.58056640625,
0.0555419921875,
0.58544921875,
0.045745849609375,
0.6171875,
0.0196075439453125,
0.61279296875
],
[
0.08087158203125,
0.58544921875,
0.46728515625,
0.58544921875,
0.46728515625,
0.5986328125,
0.08087158203125,
0.5986328125
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.5009765625,
0.59521484375,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.5009765625,
0.609375
],
[
0.0236968994140625,
0.60595703125,
0.058013916015625,
0.60595703125,
0.058013916015625,
0.66748046875,
0.0236968994140625,
0.66748046875
],
[
0.080078125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.080078125,
0.62890625
],
[
0.5048828125,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.625,
0.5048828125,
0.625
],
[
0.08087158203125,
0.6318359375,
0.468017578125,
0.6318359375,
0.468017578125,
0.642578125,
0.08087158203125,
0.642578125
],
[
0.50341796875,
0.62890625,
0.88818359375,
0.62890625,
0.88818359375,
0.6396484375,
0.50341796875,
0.6396484375
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.5009765625,
0.6416015625,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.5009765625,
0.65576171875
],
[
0.0236968994140625,
0.6572265625,
0.0555419921875,
0.65576171875,
0.058013916015625,
0.7021484375,
0.0269622802734375,
0.70263671875
],
[
0.080078125,
0.6611328125,
0.46728515625,
0.6611328125,
0.46728515625,
0.6748046875,
0.080078125,
0.6748046875
],
[
0.50341796875,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.50341796875,
0.671875
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.50341796875,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.0400390625,
0.68115234375,
0.045745849609375,
0.68115234375,
0.045745849609375,
0.68798828125,
0.0400390625,
0.68798828125
],
[
0.08087158203125,
0.6923828125,
0.46728515625,
0.6923828125,
0.46728515625,
0.70263671875,
0.08087158203125,
0.70263671875
],
[
0.50341796875,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.08087158203125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.50341796875,
0.71630859375
],
[
0.080078125,
0.7216796875,
0.438720703125,
0.7216796875,
0.438720703125,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.50244140625,
0.71826171875,
0.87841796875,
0.71826171875,
0.87841796875,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.521484375,
0.73486328125,
0.88818359375,
0.73486328125,
0.88818359375,
0.74853515625,
0.521484375,
0.74853515625
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.08087158203125,
0.76953125,
0.20751953125,
0.76953125,
0.20751953125,
0.78271484375,
0.08087158203125,
0.78271484375
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.5009765625,
0.78076171875,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.5009765625,
0.79443359375
],
[
0.09722900390625,
0.79541015625,
0.471435546875,
0.79541015625,
0.471435546875,
0.81005859375,
0.09722900390625,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"Progressive Feature Adjustment for",
"Semi-supervised Learning from Pretrained Models",
"Hai-Ming XulLingqiao LiulHao Chen?Ehsan AbbasnejadlRafael Felix",
"1Australian Institute for Machine Learning, University of Adelaide2Zhejiang University",
"[hai-ming.xu, lingqiao.liu, ehsan.abbasnejad, rafael.felixalves)@adelaide.edu.au",
"haochen.cad@zju.edu.cn",
"de",
"Abstract",
"M",
"As an effective way to alleviate the burden of data anno-",
"tation, semi-supervised learning (SsL) provides an attrac-",
"9",
"tive solution due to its ability to leverage both labeled and",
"unlabeled data to build a predictive model. While signifi-",
"LA",
"cant progress has been made recently, SsL algorithms are",
"often evaluated and developed under the assumption that",
"Figure 1. Visualization of the correct keeping and error correcting",
"the network is randomly initialized. This is in sharp contrast",
"'SO",
"abilityofFixMatchand oursapproachwithapretrainedmodel.",
"to most vision recognition systems that are built from fine-",
"The y-axis of the left figure denotes the percentage of the initially",
"tuning a pretrained network for better performance. While",
"correctly labeled data keeping their correct class at a given iter-",
"the marriage of SsL and a pretrained model seems to be",
"ation. The y-axis of the right figure denotes the percentage of",
"straightforward, recent literature suggests that naively ap-",
"the initially incorrectly labeled data being predicted to the correct",
"plying state-of-the-art SSL with a pretrained model fails to",
"A65",
"class atagiven iteration.The experimentis conducted onFGvC",
"unleash the full potential of training data. In this paper,",
"Aircraft dataset with 15% labels. As seen, FixMatch with a pre-",
"9",
"we postulate the underlying reason is that the pretrained",
"trained model shows weaker error correcting ability than the ours",
"feature representation could bring a bias inherited from the",
"approach.Please refer to Section 4.1 for more details.",
"source data, and the biastendsto be magnifiedthroughthe",
"609",
"self-training process in a typical SsL algorithm. To over-",
"come this issue, we propose to use pseudo-labels from the",
"based on randomly initialized parameters.In recent years,",
"",
"unlabelled data to update the feature extractor that is less",
"the release and re-use of pretrained DNNsto alleviatetrain-",
"sensitive to incorrect labels and only allow the classifier",
"ing costsarebecomingcommonpracticeforcomputervi",
"to be trained from the labeled data. More specifically, we",
"sion research and applications [42]. It seems that applying",
"IXe",
"progressively adjust the feature extractor to ensure its in-",
"the existing SSL method to a pretrained model is straight-",
"duced feature distribution maintains a good class separabil-",
"forward.However,recentliteratures[45,321suggestthat",
"",
"ityevenunderstrong inputperturbation.Throughextensive",
"sucha naive solutionfails tounleashthe full potential of",
"experimental studies, we show that the proposed approach",
"training data and there seems to be a big room to improve",
"achieves superior performance over existing solutions.",
"the performance of SSL when a pretrained model is used.",
"In this study, we postulate the key issue preventing ex.",
"isting SSL solutions from attaining their full potential with",
"1.Introduction",
"pretrained models is due to the bias of the pretrained fea-",
"ture extractor: trained from the source domain data, e.g.,",
"Semi-supervised learning (SSL) is considered one of",
"ImageNet, the feature extractor may not be optimal for the",
"the most practical learning paradigms which can leverage",
"target problem, e.g., fine-grained visual recognition. Such",
"both labeled and unlabeled samples to build a prediction",
"a bias could be a much more severe issue for SSL than for",
"model [4]. With the rapid development of deep neural",
"supervised learning. This is because the self-training (or",
"networks (DNNs), extensive research on deep SSL meth-",
"pseudo-labeling) procedure commonly used in SSL tends to",
"ods [18, 28, 23, 30, 3, 35, 26] have been studied. Among",
"magnify the bias. For example, at the beginning of the train-",
"those studies, most of them are evaluated and developed",
"ing, a biased feature extractor and few labeled data could"
] |
{
"arxiv_id": "2309.04659",
"language": "en",
"page_idx": 0
}
|
||
2309.04659-en-1
|
2309.04659-en-1.jpg
|
make the classifer vulnerable to spurious correlated patterns, resulting in a wrong prediction on the unlabeled data. The wrong prediction, however, will be fed back to the classifier again as pseudo labels and reinforce the bias. While SSL from randomly initialized network also suffer from the incorrect pseudo-labels, their feature extractors do not have the bias inherited from the source domain and thus could be easier adjusted through standard SSL methods towards the target problem. In reality, the bias of the feature extractor leads to the phenomenon that the SSL process from a pretrained model is less likely to correct its wrong prediction at the early training stage, as shown in Figure 1.
In this work, we find that a surprisingly simple solution can largely resolve such an issue: we do not use unlabeled data and the corresponding pseudo-label to update the classifier but the feature extractor. Only labeled data are used to train the classifier. The rationale for this strategy is that the feature extractor is more tolerant to the label noise since it does not directly produce the final prediction, and with a good feature representation, it is possible to achieve good performance with only a few labeled data. More specifically, we require the feature representation should become closer to its class center while being pushed away from other class centers. Inspired by FixMatch [26], we also expect the above property holds for strongly-augmented data. The proposed method only modifes existing FixMatch by changing a few lines of code but has demonstrated a dramatic performance boost and even outperforms other carefully designed contrastive-learning-based approaches by a large margin. In summary, the main contribution of this paper are as follows:
$\cdot$ We provide an insight into the issue that standard SSL methods perform unsatisfactory with pretrained models and provide empirical evidence for a better understanding.
$\cdot$ We discover that a simple solution that can significantly improve SSL from a pretrained model. Note that we do not claim the operation of aligning feature to its class-wise embedding is our novelty but the discovery that such a simple strategy can be a good solution to our studied issue.
· We establish a strong baseline for SSL with pretrained model, providing practioner a simple-to-use solution for practical semi-supervised classification.
## $2. \textbf{ Related Work}$
Semi-supervised learning (SSL) has experienced rapid progress with the development of deep neural networks ( DNNs) [ 3, 2, 26, 44, 33] . The current state- of- the- art SSL approaches [18, 28, 23, 3, 26, 20, 37] usually depend on the consistency regularization [3] and pseudo-labeling [20]. A common framework is to employ two processes: one pro-
cess generates a prediction target, usually in the form of pseudo-labeling [26, 20], but could also be logits [28] or other supervision forms [3,38,36]. Then the generated pseudo supervision will be used to update the network with a different input, e.g., a different augmentation of the original input image [26,7], mixed image [3], or a network with different parameters[28].
However, existing SSL approaches are primarily optimized from randomly initialized weights, and recent works [45, 32] show that the impressive performance improvement of these standard SSL methods (includes the state-of-the-art method FixMatch [26]) will disappear when models are training from a pretrained model. Despite the initial fındings reported in [45], it still lacks a clear picture of why existing SSL methods perform unsatisfactory when pretrained models are used. In this paper, we investigate the optimization procedure of SSL from pretrained models and provide some empirical evidences to reveal barriers that limit performance. Based on the analysis, we further propose a feature adjustment module to progressively adjust the feature extractor and achieve great performance improvement on multiple vision benchmarks.
3. Preliminary of Semi-supervised Learning
In semi-supervised learning, two sets of samples are $\hat{\text{normally provided: }}$ $\{ x_{l}^{1}, x_{l}^{2}, \cdots , x_{l}^{N_{l}}\}$ $\in$ $\hat{\mathcal{X} } _{\mathcal{L} }$whose annotations $\{y_l^1,y_l^2,\cdots,y_l^{N_l}\}$ $\in\mathcal{Y}_{\mathcal{L}}$ are available and $\left \{ x_{u}^{1}, x_{u}^{2}, \cdots , x_{u}^{N_{u}}\right \}$ $\in$ $\mathcal{X} _{U}$where $N_{u}\gg N_{l}$ but without accessing label information.
Although there are many existing SSL methods in the literature, this work mainly takes one of the state-of-theart approaches FixMatch[26] as an example. It is because FixMatch successfully integrates two popular techniques in SSL together, i.e., consistency regularization and pseudolabeling, through decoupling the artificial label generation and model update with weak and strong data augmentations. Specifically, for samples from $\mathcal{X}_{\mathcal{L}}$,the model $\mathcal{M}$ is trained via a standard classification loss. For each unlabeled data $x_u^i\in{\mathcal X}_U$,“hard”pseudo label is firstly produced on the weakly augmented image
(1)
$$p(y|x_{u}^{i})=\mathcal{M}\left(A_{0}(x_{u}^{i})\right);\tilde{y}=\mathop{\mathrm{argmax}}_{c}p(y=c|x_{u}^{i}),$$
where $A_0(\cdot)$ denotes weak data augmentation. Then, the model is optimized to have a consistent prediction on the strongly augmented image
$$\mathcal{L}_{u}=\frac{1}{|\mathcal{B}_{u}|}\sum_{i\in\mathcal{B}_{u}}\mathbb{1}\big(p(\tilde{y}|x_{u}^{i})\geq\tau\big)\mathrm{CE}\Big(\mathcal{M}\big(A_{1}(x_{u}^{i})\big),\:\tilde{y}\Big),$$
(2)
where $p(\tilde{y}|x_u^i)$ means the $\tilde{y}$-th output probability on weakly
augmented $x_u^i.$ Here, 1(-) is the indicator function to select
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.08087158203125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50341796875,
0.169189453125,
0.890625,
0.169189453125,
0.890625,
0.179931640625,
0.50341796875,
0.179931640625
],
[
0.08087158203125,
0.184326171875,
0.46728515625,
0.184326171875,
0.46728515625,
0.1944580078125,
0.08087158203125,
0.1944580078125
],
[
0.50341796875,
0.18310546875,
0.6708984375,
0.18310546875,
0.6708984375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.521484375,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.21337890625,
0.521484375,
0.21337890625
],
[
0.08087158203125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.50244140625,
0.2139892578125,
0.89208984375,
0.2139892578125,
0.89208984375,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.08087158203125,
0.2298583984375,
0.46728515625,
0.2298583984375,
0.46728515625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.50244140625,
0.24365234375
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.2442626953125,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.50341796875,
0.25830078125
],
[
0.08087158203125,
0.260009765625,
0.390625,
0.260009765625,
0.390625,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.096435546875,
0.2734375,
0.468017578125,
0.274658203125,
0.468017578125,
0.2890625,
0.096435546875,
0.28857421875
],
[
0.50341796875,
0.27587890625,
0.890625,
0.27587890625,
0.890625,
0.285888671875,
0.50341796875,
0.285888671875
],
[
0.080078125,
0.290283203125,
0.4697265625,
0.290283203125,
0.4697265625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.5009765625,
0.28857421875,
0.89208984375,
0.2890625,
0.89208984375,
0.3037109375,
0.5009765625,
0.302978515625
],
[
0.07843017578125,
0.3037109375,
0.468017578125,
0.304931640625,
0.468017578125,
0.3193359375,
0.07843017578125,
0.318115234375
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.08087158203125,
0.3212890625,
0.46728515625,
0.3212890625,
0.46728515625,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333984375,
0.50244140625,
0.333251953125
],
[
0.080078125,
0.335205078125,
0.4697265625,
0.335205078125,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50244140625,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.35107421875,
0.4697265625,
0.35107421875,
0.4697265625,
0.3642578125,
0.080078125,
0.3642578125
],
[
0.5048828125,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.3623046875,
0.5048828125,
0.3623046875
],
[
0.080078125,
0.3662109375,
0.4697265625,
0.3662109375,
0.4697265625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.50244140625,
0.365478515625,
0.890625,
0.365478515625,
0.890625,
0.37890625,
0.50244140625,
0.37890625
],
[
0.08087158203125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50244140625,
0.380615234375,
0.89208984375,
0.380615234375,
0.89208984375,
0.39453125,
0.50244140625,
0.39453125
],
[
0.08087158203125,
0.396484375,
0.4697265625,
0.396484375,
0.4697265625,
0.410400390625,
0.08087158203125,
0.410400390625
],
[
0.5009765625,
0.39453125,
0.89208984375,
0.395751953125,
0.89208984375,
0.410400390625,
0.5009765625,
0.4091796875
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.410888671875,
0.74853515625,
0.410888671875,
0.74853515625,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.080078125,
0.4267578125,
0.468017578125,
0.4267578125,
0.468017578125,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.43994140625,
0.869140625,
0.442626953125,
0.869140625,
0.45703125,
0.50244140625,
0.455078125
],
[
0.08087158203125,
0.45703125,
0.468017578125,
0.45703125,
0.468017578125,
0.47021484375,
0.08087158203125,
0.47021484375
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.521484375,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.481689453125,
0.521484375,
0.481689453125
],
[
0.08087158203125,
0.488037109375,
0.46728515625,
0.488037109375,
0.46728515625,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50341796875,
0.48291015625,
0.89208984375,
0.48291015625,
0.89208984375,
0.49755859375,
0.50341796875,
0.49755859375
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.49755859375,
0.8935546875,
0.49755859375,
0.8935546875,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.5048828125,
0.51513671875,
0.57275390625,
0.51513671875,
0.57275390625,
0.52978515625,
0.5048828125,
0.52978515625
],
[
0.564453125,
0.51513671875,
0.89208984375,
0.51513671875,
0.89208984375,
0.52880859375,
0.564453125,
0.52880859375
],
[
0.50244140625,
0.52978515625,
0.68896484375,
0.52978515625,
0.68896484375,
0.54443359375,
0.50244140625,
0.54443359375
],
[
0.07843017578125,
0.5166015625,
0.468017578125,
0.51513671875,
0.468017578125,
0.52978515625,
0.07843017578125,
0.53076171875
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.080078125,
0.54736328125,
0.205078125,
0.54736328125,
0.205078125,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.521484375,
0.54541015625,
0.89208984375,
0.54443359375,
0.89208984375,
0.5595703125,
0.521484375,
0.56005859375
],
[
0.09722900390625,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.5771484375,
0.09722900390625,
0.5771484375
],
[
0.50341796875,
0.56103515625,
0.890625,
0.56103515625,
0.890625,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.111083984375,
0.57763671875,
0.46728515625,
0.57763671875,
0.46728515625,
0.59130859375,
0.111083984375,
0.59130859375
],
[
0.50341796875,
0.5771484375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59033203125,
0.50341796875,
0.59033203125
],
[
0.111083984375,
0.5947265625,
0.4697265625,
0.5947265625,
0.4697265625,
0.60791015625,
0.111083984375,
0.60791015625
],
[
0.5009765625,
0.59033203125,
0.89208984375,
0.59130859375,
0.89208984375,
0.60595703125,
0.5009765625,
0.60498046875
],
[
0.111083984375,
0.609375,
0.1732177734375,
0.609375,
0.1732177734375,
0.62353515625,
0.111083984375,
0.62353515625
],
[
0.50341796875,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.098876953125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.64013671875,
0.098876953125,
0.64013671875
],
[
0.50244140625,
0.6220703125,
0.890625,
0.6220703125,
0.890625,
0.63525390625,
0.50244140625,
0.63525390625
],
[
0.111083984375,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.111083984375,
0.65478515625
],
[
0.50341796875,
0.63720703125,
0.890625,
0.63720703125,
0.890625,
0.65087890625,
0.50341796875,
0.65087890625
],
[
0.11273193359375,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.11273193359375,
0.67041015625
],
[
0.50341796875,
0.6513671875,
0.89208984375,
0.6513671875,
0.89208984375,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.111083984375,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.111083984375,
0.68603515625
],
[
0.5048828125,
0.66845703125,
0.890625,
0.66845703125,
0.890625,
0.6787109375,
0.5048828125,
0.6787109375
],
[
0.111083984375,
0.6875,
0.468017578125,
0.68603515625,
0.468017578125,
0.70068359375,
0.111083984375,
0.7021484375
],
[
0.50341796875,
0.681640625,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.50244140625,
0.69775390625,
0.67236328125,
0.69775390625,
0.67236328125,
0.7109375,
0.50244140625,
0.7109375
],
[
0.11273193359375,
0.7041015625,
0.273681640625,
0.7041015625,
0.273681640625,
0.7138671875,
0.11273193359375,
0.7138671875
],
[
0.098876953125,
0.7197265625,
0.468017578125,
0.7197265625,
0.468017578125,
0.73291015625,
0.098876953125,
0.73291015625
],
[
0.515625,
0.7265625,
0.890625,
0.7265625,
0.890625,
0.7412109375,
0.515625,
0.7412109375
],
[
0.111083984375,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.74853515625,
0.111083984375,
0.74853515625
],
[
0.111083984375,
0.75048828125,
0.39794921875,
0.75048828125,
0.39794921875,
0.7646484375,
0.111083984375,
0.7646484375
],
[
0.50341796875,
0.76318359375,
0.89208984375,
0.76318359375,
0.89208984375,
0.7763671875,
0.50341796875,
0.7763671875
],
[
0.08087158203125,
0.7841796875,
0.2181396484375,
0.7841796875,
0.2181396484375,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.50341796875,
0.79443359375,
0.6787109375,
0.79443359375,
0.6787109375,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.098876953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5703125,
0.82373046875,
0.576171875,
0.82373046875,
0.576171875,
0.8291015625,
0.5703125,
0.8291015625
],
[
0.51708984375,
0.830078125,
0.53759765625,
0.830078125,
0.53759765625,
0.84228515625,
0.51708984375,
0.84228515625
],
[
0.60302734375,
0.8271484375,
0.87353515625,
0.8271484375,
0.87353515625,
0.8447265625,
0.60302734375,
0.8447265625
],
[
0.5595703125,
0.833984375,
0.62353515625,
0.8466796875,
0.61767578125,
0.861328125,
0.5537109375,
0.8486328125
],
[
0.08251953125,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.08251953125,
0.85302734375
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.87109375,
0.85693359375,
0.8896484375,
0.85693359375,
0.8896484375,
0.869140625,
0.87109375,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5048828125,
0.87060546875,
0.890625,
0.87060546875,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"make the classifier vulnerable to spurious correlated pat-",
"cess generates a prediction target, usually in the form of",
"terns, resulting in a wrong prediction on the unlabeled data.",
"pseudo-labeling [26, 20], but could also be logits [28] or",
"The wrong prediction, however, will be fed back to the clas-",
"other supervision forms [3, 38, 36]. Then the generated",
"sifier again as pseudo labels and reinforce the bias. While",
"pseudo supervision will be used to update the network with",
"SSL from randomly initialized network also suffer from the",
"a different input, e.g., a different augmentation of the origi-",
"incorrectpseudo-labels,theirfeatureextractorsdonothave",
"nalinputimage[26,7],mixed image[3],oranetworkwith",
"thebiasinheritedfromthe sourcedomainandthus couldbe",
"different parameters [28].",
"easier adjusted through standard SSL methods towards the",
"However, existing sSL approaches are primarily op-",
"target problem.In reality, the bias of the feature extractor",
"timized from randomly initialized weights, and recent",
"leadstothephenomenonthattheSsLprocessfromapre-",
"works [45, 32] show that the impressive performance im-",
"trained model is less likely to correct its wrong prediction",
"provement of these standard SSL methods (includes the",
"at the earlytraining stage,as shown inFigure1",
"state-of-the-art method FixMatch [26]) will disappear when",
"In this work, we find that a surprisingly simple solution",
"models are training from a pretrained model.Despite the",
"can largely resolve such an issue: we do not use unlabeled",
"initial findings reported in [45], it still lacks a clear picture",
"data and the corresponding pseudo-label to update the clas-",
"of why existing SSL methods perform unsatisfactory when",
"sifier but the feature extractor. Only labeled data are used",
"pretrained models are used. In this paper, we investigate",
"to train the classifier. The rationale for this strategy is that",
"the optimization procedure of sSL from pretrained models",
"the feature extractor is more tolerant to the label noise since",
"andprovidesomeempiricalevidencestorevealbarriersthat",
"it does not directly produce the final prediction, and with a",
"limit performance. Based on the analysis, we further pro-",
"good feature representation, it is possible to achieve good",
"pose a feature adjustment module to progressively adjust the",
"performance with only a few labeled data. More specifi-",
"feature extractor and achieve great performance improve-",
"cally, we require the feature representation should become",
"ment on multiple vision benchmarks.",
"closer to its class center while being pushed away from",
"other class centers. Inspired by FixMatch [26], we also ex-",
"3.Preliminary of Semi-supervised Learning",
"pect the above property holds for strongly-augmented data.",
"The proposed method only modifies existing FixMatch by",
"In semi-supervised learning, two sets of samples are",
"changing a fewlines of code buthas demonstrated a dra-",
"normally provided: (al,a?,..:,aN} E Xe whose",
"matic performance boost and even outperforms other care-",
"annotations (yt,yi,... ,yn') e Ye are available and",
"[cu,ca,",
"...,rNu! e Xu where Nu 》 Ni but without",
"accessing label information.",
"fully designed contrastive-learning-based approaches by a",
"large margin. In summary, the main contribution of this pa-",
"per are as follows:",
"Although there are many existing SSL methods in the",
"• We provide an insight into the issue that standard SSL",
"literature, this work mainly takes one of the state-of-the-",
"methods perform unsatisfactory with pretrained mod-",
"art approaches FixMatch [26] as an example. It is because",
"els and provide empirical evidence for a better under-",
"FixMatch successfully integrates two popular techniques in",
" standing.",
"SSL together, i.e., consistency regularization and pseudo-",
". We discover that a simple solution that can signifi-",
"labeling, through decoupling the artificial label generation",
"cantly improve SSL from a pretrained model. Note",
"and model update with weak and strong data augmentations.",
"that we do not claim the operation of aligning feature",
"Specifically, for samples from Xe, the model M is trained",
"to its class-wise embedding is our novelty but the dis-",
"via a standardclassification loss.For eachunlabeled data",
"covery that such a simple strategy can be a good solu-",
"a E Xu,“hard'pseudo label is firstly produced on the",
"weakly augmented image",
"tiontoourstudiedissue.",
"· We establish a strong baseline for SSL with pretrained",
"p(ylri) = M(Ao(ri)); y = argmaxp(y = ccu), (1)",
"model, providing practioner a simple-to-use solution",
"for practical semi-supervised classification.",
"where Ao() denotes weak data augmentation. Then, the",
"2. Related Work",
"model is optimized to have a consistent prediction on the",
"strongly augmented image",
"Semi-supervised learning (SSL) has experienced rapid",
"progress with the development of deep neural networks",
"1",
"Cu",
" 1(p(glr) ≥T)CE(M(Ai(aa), )",
"IBulieB",
"(DNNs) [3.2.26, 44, 331. The current state-of-the-art SSL",
"approaches [18, 28, 23, 3, 26, 20, 37] usually depend on the",
"(2)",
"consistency regularization [3] and pseudo-labeling [20]. A",
"where p(ylr;) means the y-th output probability on weakly",
"common framework is to employ two processes: one pro-",
"augmented cu. Here, I() is the indicator function to select"
] |
{
"arxiv_id": "2309.04659",
"language": "en",
"page_idx": 1
}
|
||
2309.04659-en-2
|
2309.04659-en-2.jpg
|
samples whose predicting confidence is greater than a predefined confidence threshold $\tau.$ Additionally,CE($\cdot,\cdot)$ is the standard cross-entropy loss, $A_1(\cdot)$ is the strong augmentation and $\mathcal{B}_u$ denotes the size of unlabeled samples in one mini-batch.
# 4. Semi-supervised Learning from Pretrained $\mathbf{Models}$
In this section, we first investigate the bias inherented in pretrained models and how existing SSL methods are troubled to achieve satisfactory performance. Then, we propose a feature adjustment module for SSL to resist the bias.
## $4. 1. \textbf{ Pretrained Models as a Double- edged Sword}$ $\mathbf{forSSL}$
With the growth of data and the enrichment of computing resources, developing powerful pretrained models have attracted increasing attention from both the academia and industry [10, 13, 5, 24, 40, 43]. On the one hand, pretraining on large-scale images gives models the general feature extraction ability for various kinds of downstream applications [46,22,39]. On the other hand, the generated feature representations will inevitably bring a bias inherited from the source data, and thus it is usually necessary to fine-tune the model on target datasets for effective usage of pretrained models [34,12,15].
For the SSL scenario, it is natural to expect that state-ofthe-art performance can be achieved by applying the stateof-the-art SSL method to a pretrained model, i.e., semisupervised fine-tuning. However, evidence from recent literatures [45, 32] show that this solution is far from the best, and there is a big room to improve SSL when a pretrained model is used. This motivates us to revisit SSL and understand what hinders it from achieving its full potential. We postulate the major issue is that the use of pretrained model is a double-edged sword for SSL: on the one hand, it brings the prior knowledge learned from the source data and boosts the performance. On the other hand, it also introduces a strong prediction bias inherited from the source data. After all, the feature learned from the source domain data may not be optimal for the target task. In effect, the prediction bias will encourage the classifier to use certain features or visual patterns for prediction, especially when the labeled data is limited in quantity and diversity. However, not all those visual patterns are true causal factors to determine the class and the spurious correlation might be mistakenly identified during training, e.g., prediction could rely on the clue from background[25].
Worse, as most of the state-of-the-art SSL methods[26, 3,20,2,44] are built upon self-training, a.k.a., pseudolabeling framework, which generates pseudo labels from the prediction on the unlabeled data, such a process tends
to further magnify the prediction bias. For example, we can consider the scenario of applying FixMatch, one of the most commonly used SSL methods, with a pretrained model. At the beginning of training, due to the limited amount of labeled (and pseudo-labeled) samples and biased feature representation, the learned classifier tends to be affected by the spurious correlation between features and class labels. Then if the classifier generates incorrect pseudo labels from unlabeled data, the bias will thus reinforce itself by further training with such pseudo labels. Consequently, this will make the SSL less prone to correct its wrong prediction made during the training process.
In order to verify our assumption, we conduct an empirical analysis on FixMatch with pretrained models. Specifically, we introduce two measurements called correctkeeping rate (CKR) and error-correcting rate(ECR). The former is defined as a percentage of the initially$^{\mathrm{l}}$ correctly labeled data keeping their correct class at a given iteration, while the latter one is defined as the percentage of the initially incorrectly labeled data being predicted to the correct class at a given iteration. The statistical results are shown in Figure 1. Then we can observe that FixMatch (with a pretrained model) has a descent CKR metric when training converges. However, the ECR of FixMatch (with a pretrained model) quickly reaches a plateau and has a poor ECR metric. This issue becomes more evident by comparing its ECR with the proposed method.
## $4. 2. \textbf{ Progressive Feature Adjustment}$
The above analysis suggests that the standard SSL approach could suffer more from the biased feature extractor and we may need a special process to alleviate the impact of bias. In this work, we propose to only use the labeled data to train the classifier, while pseudo-labels generated from the larger amount of unlabeled samples will only be used to update the feature extractor. In this way, unlabeled data influences the classifier indirectly by producing better feature representations. As the classifier is always trained on noise-free labeled data, even if the feature representation is imperfect, the classifier can suppress the noisy dimensions and identify the discriminative patterns in the feature representation. Thus the feature extractor can be more tolerant to the noise in pseudo-labels. It seems that one drawback of the above method is the lack of training examples for the classifier. However, since a pretrained feature extractor has already been able to provide a reasonable starting point and will be further refined by the proposed progressive adjustment method, training the classifier on a limited number of samples can guarantee a good performance. The overall architecture is shown in Figure 2.
Specifically, given a batch of labeled samples $(x_l,y_l)\in$
l“Initially” here means an unlabeled sample was falsely labeled for the
first time in the whole training process.
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5009765625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.5009765625,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.136962890625,
0.4697265625,
0.136962890625,
0.4697265625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08087158203125,
0.154052734375,
0.1568603515625,
0.154052734375,
0.1568603515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.50244140625,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.08087158203125,
0.184326171875,
0.46728515625,
0.184326171875,
0.46728515625,
0.1988525390625,
0.08087158203125,
0.1988525390625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.1817626953125,
0.89208984375,
0.1956787109375,
0.50244140625,
0.1962890625
],
[
0.1094970703125,
0.2020263671875,
0.1732177734375,
0.2020263671875,
0.1732177734375,
0.216552734375,
0.1094970703125,
0.216552734375
],
[
0.5048828125,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.5048828125,
0.210205078125
],
[
0.50341796875,
0.2139892578125,
0.890625,
0.2139892578125,
0.890625,
0.2247314453125,
0.50341796875,
0.2247314453125
],
[
0.098876953125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.098876953125,
0.242431640625
],
[
0.5009765625,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.242431640625,
0.5009765625,
0.242431640625
],
[
0.080078125,
0.2442626953125,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.25830078125
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.07843017578125,
0.25830078125,
0.471435546875,
0.2587890625,
0.471435546875,
0.2734375,
0.07843017578125,
0.272705078125
],
[
0.50244140625,
0.2587890625,
0.66259765625,
0.2587890625,
0.66259765625,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.07843017578125,
0.2734375,
0.435546875,
0.274658203125,
0.435546875,
0.28857421875,
0.07843017578125,
0.287353515625
],
[
0.521484375,
0.2734375,
0.890625,
0.2734375,
0.890625,
0.287353515625,
0.521484375,
0.287353515625
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.08087158203125,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.31494140625,
0.08087158203125,
0.31494140625
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.12091064453125,
0.317626953125,
0.181396484375,
0.317626953125,
0.181396484375,
0.328857421875,
0.12091064453125,
0.328857421875
],
[
0.50244140625,
0.3193359375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.098876953125,
0.34033203125,
0.468017578125,
0.34033203125,
0.468017578125,
0.354736328125,
0.098876953125,
0.354736328125
],
[
0.50244140625,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.356689453125,
0.468017578125,
0.356689453125,
0.468017578125,
0.37060546875,
0.080078125,
0.37060546875
],
[
0.080078125,
0.371826171875,
0.4697265625,
0.371826171875,
0.4697265625,
0.385009765625,
0.080078125,
0.385009765625
],
[
0.5048828125,
0.3662109375,
0.8896484375,
0.3662109375,
0.8896484375,
0.376953125,
0.5048828125,
0.376953125
],
[
0.50244140625,
0.380126953125,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.08087158203125,
0.38818359375,
0.46728515625,
0.38818359375,
0.46728515625,
0.3984375,
0.08087158203125,
0.3984375
],
[
0.50244140625,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.080078125,
0.401611328125,
0.46728515625,
0.401611328125,
0.46728515625,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.5009765625,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.080078125,
0.416748046875,
0.468017578125,
0.416748046875,
0.468017578125,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.08087158203125,
0.431884765625,
0.4697265625,
0.431884765625,
0.4697265625,
0.44580078125,
0.08087158203125,
0.44580078125
],
[
0.50244140625,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.08087158203125,
0.4462890625,
0.4697265625,
0.4462890625,
0.4697265625,
0.460205078125,
0.08087158203125,
0.460205078125
],
[
0.5009765625,
0.453857421875,
0.89208984375,
0.455078125,
0.89208984375,
0.4697265625,
0.5009765625,
0.468505859375
],
[
0.07843017578125,
0.4609375,
0.468017578125,
0.462158203125,
0.468017578125,
0.47607421875,
0.07843017578125,
0.474853515625
],
[
0.50244140625,
0.468505859375,
0.89208984375,
0.47021484375,
0.89208984375,
0.484130859375,
0.50244140625,
0.481689453125
],
[
0.07843017578125,
0.47607421875,
0.468017578125,
0.4765625,
0.468017578125,
0.4912109375,
0.07843017578125,
0.490478515625
],
[
0.50341796875,
0.48486328125,
0.7587890625,
0.48486328125,
0.7587890625,
0.5,
0.50341796875,
0.5
],
[
0.080078125,
0.492431640625,
0.2998046875,
0.492431640625,
0.2998046875,
0.505859375,
0.080078125,
0.505859375
],
[
0.09722900390625,
0.50830078125,
0.468017578125,
0.50830078125,
0.468017578125,
0.521484375,
0.09722900390625,
0.521484375
],
[
0.50341796875,
0.51025390625,
0.78173828125,
0.51025390625,
0.78173828125,
0.5244140625,
0.50341796875,
0.5244140625
],
[
0.080078125,
0.5234375,
0.46728515625,
0.5234375,
0.46728515625,
0.537109375,
0.080078125,
0.537109375
],
[
0.51953125,
0.5322265625,
0.890625,
0.53271484375,
0.890625,
0.54736328125,
0.51953125,
0.546875
],
[
0.080078125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.5517578125,
0.080078125,
0.5517578125
],
[
0.08087158203125,
0.5546875,
0.46728515625,
0.5546875,
0.46728515625,
0.56494140625,
0.08087158203125,
0.56494140625
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.50341796875,
0.564453125,
0.8935546875,
0.564453125,
0.8935546875,
0.57763671875,
0.50341796875,
0.57763671875
],
[
0.07843017578125,
0.58203125,
0.4697265625,
0.58349609375,
0.4697265625,
0.59765625,
0.07843017578125,
0.5966796875
],
[
0.50244140625,
0.5791015625,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.080078125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.50244140625,
0.59326171875,
0.89208984375,
0.59326171875,
0.89208984375,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.080078125,
0.62890625,
0.4697265625,
0.62890625,
0.4697265625,
0.642578125,
0.080078125,
0.642578125
],
[
0.50341796875,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.63623046875,
0.50341796875,
0.63623046875
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.5009765625,
0.65283203125,
0.89208984375,
0.65380859375,
0.89208984375,
0.66845703125,
0.5009765625,
0.66748046875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.66943359375,
0.89208984375,
0.66943359375,
0.89208984375,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.08087158203125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.50244140625,
0.68505859375,
0.89208984375,
0.68505859375,
0.89208984375,
0.69873046875,
0.50244140625,
0.69873046875
],
[
0.080078125,
0.7041015625,
0.4697265625,
0.7041015625,
0.4697265625,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.50244140625,
0.69970703125,
0.8896484375,
0.69970703125,
0.8896484375,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.080078125,
0.7197265625,
0.468017578125,
0.7197265625,
0.468017578125,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.7265625,
0.50341796875,
0.7265625
],
[
0.50244140625,
0.72998046875,
0.89208984375,
0.72998046875,
0.89208984375,
0.74365234375,
0.50244140625,
0.74365234375
],
[
0.08251953125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.7470703125,
0.08251953125,
0.7470703125
],
[
0.08087158203125,
0.75,
0.4697265625,
0.75,
0.4697265625,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50341796875,
0.744140625,
0.89208984375,
0.74560546875,
0.89208984375,
0.7587890625,
0.50341796875,
0.75830078125
],
[
0.50341796875,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.50341796875,
0.7734375
],
[
0.08087158203125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.7763671875,
0.08087158203125,
0.7763671875
],
[
0.5009765625,
0.7734375,
0.89208984375,
0.7744140625,
0.89208984375,
0.7890625,
0.5009765625,
0.78857421875
],
[
0.080078125,
0.77978515625,
0.4697265625,
0.77978515625,
0.4697265625,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.5009765625,
0.7890625,
0.8896484375,
0.7890625,
0.8896484375,
0.80322265625,
0.5009765625,
0.80322265625
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.80615234375,
0.89208984375,
0.80615234375,
0.89208984375,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.080078125,
0.82568359375,
0.193603515625,
0.82568359375,
0.193603515625,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.82080078125,
0.89208984375,
0.82080078125,
0.89208984375,
0.833984375,
0.50244140625,
0.833984375
],
[
0.1004638671875,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.50341796875,
0.8369140625,
0.7197265625,
0.8369140625,
0.7197265625,
0.84765625,
0.50341796875,
0.84765625
],
[
0.08087158203125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.52294921875,
0.85107421875,
0.890625,
0.85107421875,
0.890625,
0.86572265625,
0.52294921875,
0.86572265625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.51953125,
0.875,
0.89208984375,
0.875,
0.89208984375,
0.8896484375,
0.51953125,
0.8896484375
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.708984375,
0.88720703125,
0.708984375,
0.9013671875,
0.5009765625,
0.90087890625
]
] |
[
"samples whose predicting confidence is greater than a pre-",
"to further magnify the prediction bias. For example, we can",
"defined confidence threshold T. Additionally, CE(,) is the",
"consider the scenario of applying FixMatch, one of the most",
"standard cross-entropy loss, Ai() is the strong augmenta-",
"commonly used SSL methods, with a pretrained model. At",
"tion and Bu denotes the size of unlabeled samples in one",
"the beginning of training, due to the limited amount of la-",
"mini-batch.",
"beled (and pseudo-labeled) samples and biased feature rep-",
"resentation, the learned classifier tends to be affected by the",
"4. Semi-supervised Learning from Pretrained",
"spurious correlation betweenfeatures and class labels.Then",
"Models",
"iftheclassifiergeneratesincorrectpseudolabelsfromunla-",
"beled data,thebias willthusreinforceitself byfurthertrain-",
"In this section, we first investigate the bias inherented in",
"ing with such pseudo labels. Consequently, this will make",
"pretrained models and how existing SSL methods are trou-",
"the SSL less prone to correct its wrong prediction made dur-",
"bled to achieve satisfactory performance.Then, we propose",
"ing the training process.",
"a feature adjustment module for SSL to resist the bias.",
"In order to verify our assumption, we conduct an empiri-",
"cal analysis on FixMatch with pretrained models. Specif-",
"4.1. Pretrained Models as a Double-edged Sword",
"ically, we introduce two measurements called correct-",
"forSSL",
"keeping rate (CKR) and error-correcting rate(ECR). The",
"former is defined as a percentage of the initially correctly",
"With the growth of data and the enrichment of comput-",
"labeled data keeping their correct class at a given iteration,",
"ing resources, developing powerful pretrained models have",
"attracted increasing attention from both the academia and",
"whilethe latter one is defined as the percentage of the ini-",
"tially incorrectly labeled data being predicted to the correct",
"industry [10.13.5.24.40.431.On theonehand.pre-",
"class at a given iteration. The statistical results are shown",
"training on large-scale images gives models the general fea-",
"in Figure 1. Then we can observe that FixMatch (with a",
"ture extraction ability for various kinds of downstream ap-",
"pretrained model) has a descent CKR metric when training",
"plications [46, 22, 39]. On the other hand, the generated",
"converges. However, the ECR of FixMatch (with a pre-",
"feature representations will inevitably bring a bias inher-",
"trained model) quickly reaches a plateau and has a poor",
"ited from the source data, and thus it is usually necessary",
"ECRmetric.This issuebecomes moreevidentby compar-",
"to fine-tune the model on target datasets for effective usage",
"ing its ECR with the proposed method.",
"of pretrained models [34, 12, 15].",
" For the SSL scenario, it is natural to expect that state-of-",
"4.2.Progressive Feature Adjustment",
"the-art performance can be achieved by applying the state-",
"The above analysis suggests that the standard SSL ap-",
"of-the-art SSL method to a pretrained model, i.e., semi-",
"supervised fine-tuning.However,evidence from recent lit-",
"proach could suffer more from the biased feature extractor",
"eratures [45, 32] show that this solution is far from the best,",
"and we may need a special process to alleviate the impact of",
"and there is a big room to improve SSL when a pretrained",
"bias. In this work, we propose to only use the labeled data",
"model is used. This motivates us to revisit SSL and under-",
"to train the classifier, while pseudo-labels generated from",
"stand what hinders it from achieving its full potential. We",
"the larger amount of unlabeled samples will only be used",
"postulate the major issue is that the use of pretrained model",
"to update the feature extractor. In this way,unlabeled data",
"is a double-edged sword for SSL: on the one hand, it brings",
"influences the classifier indirectly by producing better fea-",
"ture representations. As the classifier is always trained on",
"the prior knowledge learned from the source data and boosts",
"noise-free labeled data, even if the feature representation is",
"the performance.On the other hand, it also introduces a",
"strong prediction bias inherited from the source data. After",
"imperfect, the classifier can suppress the noisy dimensions",
"all, the feature learned from the source domain data may not",
"and identify the discriminative patterns in the feature repre-",
"be optimal for the target task. In effect, the prediction bias",
"sentation. Thus the feature extractor can be more tolerant",
"to the noise in pseudo-labels. It seems that one drawback",
"will encouragethe classifiertousecertainfeatures orvisual",
"patterns for prediction, especially when the labeled data is",
"of the above method is the lack of training examples for",
"the classifier. However, since a pretrained feature extractor",
"limited in quantityand diversity.However,notall those vi-",
"has already been able to provide a reasonable starting point",
"sual patterns are true causal factors to determine the class",
"and will be further refined by the proposed progressive ad-",
"and the spurious correlation might be mistakenly identified",
"during training, e.g., prediction could rely on the clue from",
"justment method, training the classifier on a limited number",
"background [25].",
"of samples can guarantee a good performance. The overall",
"Worse, as most of the state-of-the-art SSL methods [26,",
"architectureis shown inFigure2",
"3, 20, 2, 44] are built upon self-training, a.k.a., pseudo-",
"Specifically, given a batch of labeled samples (ct, yi) E",
"labeling framework, which generates pseudo labels from",
"I\"Initially\" here means an unlabeled sample was falsely labeled for the",
"the prediction on the unlabeled data, such a process tends",
"first time in the whole training process."
] |
{
"arxiv_id": "2309.04659",
"language": "en",
"page_idx": 2
}
|
||
2309.04659-en-4
|
2309.04659-en-4.jpg
|
parameter.
In order to adjust the feature extractor efficiently, oracle class-wise embeddings of target dataset should be an ideal choice. However, it is unrealistic due to the lack of annotations for unlabeled samples and the inherent bias in the pretrained feature extractor. Thus, we propose to progressively update these class-wise embeddings from both labeled and unlabeled data. For labeled data, $\mu_c$ is updated via
$$\mu_c^{new}=\beta\mu_c^{old}+(1-\beta)f\big(A_1(x_l^i)\big)\mathbb{1}(y_l^i=c),$$
(6)
and for unlabeled data, $\mu_c$ is updated via
$$\mu_{c}^{new}=\beta\mu_{c}^{old}+(1-\beta)f\big(A_{1}(x_{u}^{i})\big)\mathbb{1}\Big(p\big(c|A_{0}(x_{u}^{i})\big)\geq\tau\Big),$$
$(\dot{7})$
where the indicator function $1(y_l^i=c)$ selects samples from the $c$-th class from the labeled data, the indicator function $1\left(p^{t}\left(c|A_{0}(x_{u}^{i})\right)\geq\tau\right)$ selects unlabeled samples that are confidently classified into the $c$-th class by the classifier. This is identical to the criterion of generating pseudo-labels. $\beta$ is a momentum term that controls how far the class-wise feature reaches into embedding history.
## 5. Experimental results
In this section, we compare our approach with several
SSL methods with pretrained models.
## 5.1. Experimental details
We strictly follow [32] to design our experiment, including the evaluation datasets and pretrained model choices. We made such a choice since Self-Tuning [32] has demonstrated the state-of-the-art performance and its experimental evaluation is comprehensive and realistic. Some experimental details are as follows:
Datasets: Following the protocol of [32], which addressed the same research problem as this paper, four vision benchmarks are evaluated, i.e., FGVC Aircraft [21], Stanford Cars [16], $CUB- 200- 2011$ [31],and $CIFAR- 100$ [17]. Specifically, the first three are challenging fine-grained classification datasets and label proportions ranging from 15% to 50% are tested. Label partition of $CIFAR-100$ follows the standard SSL protocol: $\hat{4}/25/100$ labeled images per class. Methods: Nine popular deep SSL approaches are included for comparison, i.e., $\Pi$-model[19], Pseudo-Labeling[20], Mean Teacher[28], UDA[35], FixMatch[26], FlexMatch [ 44] , SimCLRv2 [ 6] , FixMatch+ AKC+ ARC [ 1] and Self-Tuning[32]. Meanwhile, performance of Fine-Tuning on labeled data is also reported for a reference baseline. For our approach, we also consider a simple extension by incorporating it into the recently proposed consistency-based SSL method FlexMatch [44], which uses dynamically assigned threshold with a FixMatch framework. We call this
<table>
<tbody>
<tr>
<th></th>
<th>400</th>
<th>2500</th>
<th>10000</th>
</tr>
<tr>
<td>Fine-Tuning ${\mathbf{g}}({\mathbf{b}}{\mathbf{a}}{\mathbf{s}}{\mathbf{e}}{\mathbf{l}}{\mathbf{i}}{\mathbf{n}}{\mathbf{e}})$</td>
<td>60.79</td>
<td>31.69</td>
<td>21.74</td>
</tr>
<tr>
<td>Pseudo- Labeling $\mathbf{I}[20]$</td>
<td>59.21</td>
<td> </td>
<td>一</td>
</tr>
<tr>
<td>$MT$ [28]</td>
<td>60.68</td>
<td> </td>
<td>一</td>
</tr>
<tr>
<td>$UDA$ [35]</td>
<td>58.32</td>
<td> </td>
<td>一</td>
</tr>
<tr>
<td>FixMatcht [26]</td>
<td>52.88</td>
<td>25.63</td>
<td>18.38</td>
</tr>
<tr>
<td>FlexMatch [44]</td>
<td>40.41</td>
<td>23.19</td>
<td>17.73</td>
</tr>
<tr>
<td>Self-Tuning $[32^{-}]$</td>
<td>47.17</td>
<td>24.16</td>
<td>17.57</td>
</tr>
<tr>
<td>$0urs\dagger$</td>
<td>45.48</td>
<td>23.12</td>
<td>16.89</td>
</tr>
<tr>
<td>$0urs+1$</td>
<td>37.36</td>
<td>22.06</td>
<td>16.58</td>
</tr>
</tbody>
</table>
Table 1. Error rates $(\%)$ on $CIFAR- IOO$ with EfficientNet- B2.十
means ours implementation based on [32].
extension Ours+. Note that it shows our approach can still boost the performance even with more advanced SSL algorithms. In our work, all experiments were implemented in PyTorch and run on a GeForce RTX 2080Ti GPU with 11GB memory.
Pretrained models: Following [32], three models pretrained on ImageNet[9] are chosen for evaluation, i.e., ResNet-50[14] and EfficientNet[27] which are pretrained in a supervised way, and a ResNet-50 which is trained through an unsupervised learning method MoCo v2[13].
5.2. Train from Supervised Pretrained Models
In this section, we compare various SSL methods trained
from supervised pretrained models.
Fine-grained Classification benchmarks: We use a ResNet-50 network, which is supervised pretrained on ImageNet, to initialize all SSL models. The results are shown in Table 2. It is clear that our proposed method achieves overall significant improvement than other comparing SSL approaches. Specifically, compared with traditional SSL methods, our approach increases the test accuracy by a large margin on all kinds of partitions of three benchmarks. Taking the state-of-the-art method FixMatch[26] as an example, the performance gain of our approach exceeds 10 percent on both Stanford Cars and CUB-200-2011 with 15% labels. This is thanks to the proposed feature adjustment module in our approach which greatly reduces the bias inherented in the pretrained model. Furthermore, our approach is also superior to the recently proposed Self-Tuning method[32] especially when labels are limited, e.g., only 15% training samples are labeled. When the CPL module proposed in FlexMatch [44] is added to our approach, Ours+ leads to a further performance boost.
Standard SSL benchmarks: We choose CIFAR-100
dataset[17] which is one of the most challenging datasets among standard SSL benchmarks to evaluate SSL methods from a pretrained model. Due to the lack of open-resourced pretrained checkpoints on WideResNet-28-8 model[41],
|
[
[
0.07843017578125,
0.09344482421875,
0.148681640625,
0.091552734375,
0.148681640625,
0.10479736328125,
0.07843017578125,
0.106689453125
],
[
0.73046875,
0.0947265625,
0.8310546875,
0.0921630859375,
0.8310546875,
0.106689453125,
0.73046875,
0.10919189453125
],
[
0.09722900390625,
0.106689453125,
0.468017578125,
0.1060791015625,
0.468017578125,
0.12060546875,
0.09722900390625,
0.121826171875
],
[
0.53369140625,
0.1029052734375,
0.5888671875,
0.1029052734375,
0.5888671875,
0.1142578125,
0.53369140625,
0.1142578125
],
[
0.708984375,
0.1142578125,
0.740234375,
0.1142578125,
0.740234375,
0.1300048828125,
0.708984375,
0.1300048828125
],
[
0.7607421875,
0.1142578125,
0.7998046875,
0.1142578125,
0.7998046875,
0.1300048828125,
0.7607421875,
0.1300048828125
],
[
0.81640625,
0.1142578125,
0.86279296875,
0.1142578125,
0.86279296875,
0.1287841796875,
0.81640625,
0.1287841796875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.53125,
0.1307373046875,
0.68701171875,
0.1307373046875,
0.68701171875,
0.14453125,
0.53125,
0.14453125
],
[
0.7548828125,
0.1307373046875,
0.80126953125,
0.1307373046875,
0.80126953125,
0.1458740234375,
0.7548828125,
0.1458740234375
],
[
0.81640625,
0.1307373046875,
0.86279296875,
0.1307373046875,
0.86279296875,
0.1458740234375,
0.81640625,
0.1458740234375
],
[
0.07843017578125,
0.136962890625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.07843017578125,
0.15087890625
],
[
0.701171875,
0.1319580078125,
0.744140625,
0.1319580078125,
0.744140625,
0.143310546875,
0.701171875,
0.143310546875
],
[
0.53369140625,
0.1458740234375,
0.6787109375,
0.146484375,
0.6787109375,
0.1610107421875,
0.53369140625,
0.160400390625
],
[
0.70166015625,
0.1458740234375,
0.744140625,
0.1458740234375,
0.744140625,
0.1610107421875,
0.70166015625,
0.1610107421875
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.53369140625,
0.1610107421875,
0.5947265625,
0.1610107421875,
0.5947265625,
0.175537109375,
0.53369140625,
0.175537109375
],
[
0.70166015625,
0.1610107421875,
0.74609375,
0.1610107421875,
0.74609375,
0.175537109375,
0.70166015625,
0.175537109375
],
[
0.080078125,
0.1673583984375,
0.46728515625,
0.1673583984375,
0.46728515625,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.8310546875,
0.1641845703125,
0.84326171875,
0.1641845703125,
0.84326171875,
0.1749267578125,
0.8310546875,
0.1749267578125
],
[
0.53515625,
0.177978515625,
0.60107421875,
0.177978515625,
0.60107421875,
0.189453125,
0.53515625,
0.189453125
],
[
0.70166015625,
0.175537109375,
0.744140625,
0.175537109375,
0.744140625,
0.1912841796875,
0.70166015625,
0.1912841796875
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.53173828125,
0.1900634765625,
0.6396484375,
0.1912841796875,
0.6396484375,
0.205810546875,
0.53173828125,
0.2052001953125
],
[
0.70166015625,
0.1900634765625,
0.74609375,
0.1900634765625,
0.74609375,
0.205810546875,
0.70166015625,
0.205810546875
],
[
0.7587890625,
0.1900634765625,
0.7998046875,
0.1900634765625,
0.7998046875,
0.205810546875,
0.7587890625,
0.205810546875
],
[
0.81787109375,
0.1900634765625,
0.859375,
0.1900634765625,
0.859375,
0.205810546875,
0.81787109375,
0.205810546875
],
[
0.07843017578125,
0.1962890625,
0.410888671875,
0.1976318359375,
0.410888671875,
0.212158203125,
0.07843017578125,
0.2108154296875
],
[
0.70166015625,
0.205810546875,
0.744140625,
0.205810546875,
0.744140625,
0.2203369140625,
0.70166015625,
0.2203369140625
],
[
0.75732421875,
0.205810546875,
0.80126953125,
0.205810546875,
0.80126953125,
0.2215576171875,
0.75732421875,
0.2215576171875
],
[
0.81787109375,
0.20703125,
0.859375,
0.20703125,
0.859375,
0.2215576171875,
0.81787109375,
0.2215576171875
],
[
0.1168212890625,
0.2197265625,
0.471435546875,
0.2183837890625,
0.471435546875,
0.2392578125,
0.1168212890625,
0.240478515625
],
[
0.53369140625,
0.20703125,
0.646484375,
0.20703125,
0.646484375,
0.2203369140625,
0.53369140625,
0.2203369140625
],
[
0.53173828125,
0.2203369140625,
0.646484375,
0.222900390625,
0.646484375,
0.238037109375,
0.53173828125,
0.236083984375
],
[
0.701171875,
0.2215576171875,
0.744140625,
0.2215576171875,
0.744140625,
0.236083984375,
0.701171875,
0.236083984375
],
[
0.7587890625,
0.2215576171875,
0.80126953125,
0.2215576171875,
0.80126953125,
0.236083984375,
0.7587890625,
0.236083984375
],
[
0.81787109375,
0.2215576171875,
0.859375,
0.2215576171875,
0.859375,
0.236083984375,
0.81787109375,
0.236083984375
],
[
0.70166015625,
0.236083984375,
0.744140625,
0.236083984375,
0.744140625,
0.250732421875,
0.70166015625,
0.250732421875
],
[
0.7587890625,
0.236083984375,
0.80126953125,
0.236083984375,
0.80126953125,
0.251953125,
0.7587890625,
0.251953125
],
[
0.81787109375,
0.236083984375,
0.859375,
0.236083984375,
0.859375,
0.250732421875,
0.81787109375,
0.250732421875
],
[
0.080078125,
0.2474365234375,
0.34716796875,
0.2474365234375,
0.34716796875,
0.261962890625,
0.080078125,
0.261962890625
],
[
0.53369140625,
0.2374267578125,
0.58251953125,
0.2374267578125,
0.58251953125,
0.251953125,
0.53369140625,
0.251953125
],
[
0.53369140625,
0.250732421875,
0.5888671875,
0.25244140625,
0.58837890625,
0.267578125,
0.53173828125,
0.26513671875
],
[
0.7587890625,
0.250732421875,
0.80126953125,
0.250732421875,
0.80126953125,
0.266357421875,
0.7587890625,
0.266357421875
],
[
0.81787109375,
0.250732421875,
0.859375,
0.250732421875,
0.859375,
0.266357421875,
0.81787109375,
0.266357421875
],
[
0.50341796875,
0.26953125,
0.89208984375,
0.26953125,
0.89208984375,
0.28271484375,
0.50341796875,
0.28271484375
],
[
0.701171875,
0.251953125,
0.74609375,
0.251953125,
0.74609375,
0.266357421875,
0.701171875,
0.266357421875
],
[
0.0767822265625,
0.272705078125,
0.463134765625,
0.271484375,
0.463134765625,
0.292236328125,
0.0767822265625,
0.29345703125
],
[
0.50244140625,
0.2841796875,
0.75341796875,
0.2841796875,
0.75341796875,
0.29736328125,
0.50244140625,
0.29736328125
],
[
0.44677734375,
0.296630859375,
0.4697265625,
0.296630859375,
0.4697265625,
0.31005859375,
0.44677734375,
0.31005859375
],
[
0.50244140625,
0.306884765625,
0.89208984375,
0.306884765625,
0.89208984375,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.08087158203125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.3359375,
0.08087158203125,
0.3359375
],
[
0.5009765625,
0.32080078125,
0.89208984375,
0.3212890625,
0.89208984375,
0.3359375,
0.5009765625,
0.335205078125
],
[
0.080078125,
0.337158203125,
0.4697265625,
0.337158203125,
0.4697265625,
0.35107421875,
0.080078125,
0.35107421875
],
[
0.50244140625,
0.337158203125,
0.89208984375,
0.337158203125,
0.89208984375,
0.35107421875,
0.50244140625,
0.35107421875
],
[
0.080078125,
0.354736328125,
0.4697265625,
0.354736328125,
0.4697265625,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.5009765625,
0.35107421875,
0.890625,
0.35107421875,
0.890625,
0.3642578125,
0.5009765625,
0.3642578125
],
[
0.50244140625,
0.3642578125,
0.60546875,
0.367431640625,
0.60546875,
0.3818359375,
0.50244140625,
0.37890625
],
[
0.080078125,
0.37255859375,
0.4697265625,
0.37255859375,
0.4697265625,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.080078125,
0.386474609375,
0.468017578125,
0.386962890625,
0.468017578125,
0.401611328125,
0.080078125,
0.40087890625
],
[
0.50244140625,
0.386962890625,
0.890625,
0.386962890625,
0.890625,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.08087158203125,
0.40283203125,
0.4697265625,
0.40283203125,
0.4697265625,
0.416748046875,
0.08087158203125,
0.416748046875
],
[
0.50341796875,
0.404052734375,
0.88623046875,
0.404052734375,
0.88623046875,
0.4140625,
0.50341796875,
0.4140625
],
[
0.07843017578125,
0.416748046875,
0.338134765625,
0.41845703125,
0.338134765625,
0.43310546875,
0.07843017578125,
0.43115234375
],
[
0.50244140625,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.50244140625,
0.43310546875,
0.89208984375,
0.43310546875,
0.89208984375,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.08087158203125,
0.44580078125,
0.27685546875,
0.44580078125,
0.27685546875,
0.460205078125,
0.08087158203125,
0.460205078125
],
[
0.5009765625,
0.4462890625,
0.8759765625,
0.447509765625,
0.8759765625,
0.462158203125,
0.5009765625,
0.4609375
],
[
0.096435546875,
0.4697265625,
0.4697265625,
0.47021484375,
0.4697265625,
0.48486328125,
0.096435546875,
0.484130859375
],
[
0.50341796875,
0.474853515625,
0.85693359375,
0.474853515625,
0.85693359375,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.080078125,
0.486083984375,
0.328369140625,
0.486083984375,
0.328369140625,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.521484375,
0.498779296875,
0.89208984375,
0.498779296875,
0.89208984375,
0.51220703125,
0.521484375,
0.51220703125
],
[
0.08087158203125,
0.5087890625,
0.27294921875,
0.5087890625,
0.27294921875,
0.5234375,
0.08087158203125,
0.5234375
],
[
0.5009765625,
0.51220703125,
0.736328125,
0.51318359375,
0.736328125,
0.52783203125,
0.5009765625,
0.5263671875
],
[
0.1004638671875,
0.5322265625,
0.46728515625,
0.5322265625,
0.46728515625,
0.54541015625,
0.1004638671875,
0.54541015625
],
[
0.5009765625,
0.53271484375,
0.8935546875,
0.5341796875,
0.8935546875,
0.548828125,
0.5009765625,
0.54736328125
],
[
0.080078125,
0.548828125,
0.4697265625,
0.548828125,
0.4697265625,
0.5625,
0.080078125,
0.5625
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.08087158203125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.50341796875,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.080078125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.5009765625,
0.57763671875,
0.89208984375,
0.5791015625,
0.89208984375,
0.59326171875,
0.5009765625,
0.59228515625
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.50244140625,
0.60791015625
],
[
0.080078125,
0.60791015625,
0.273681640625,
0.60791015625,
0.273681640625,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50244140625,
0.60986328125,
0.890625,
0.60986328125,
0.890625,
0.62353515625,
0.50244140625,
0.62353515625
],
[
0.07843017578125,
0.625,
0.468017578125,
0.62548828125,
0.468017578125,
0.64013671875,
0.07843017578125,
0.6396484375
],
[
0.50244140625,
0.625,
0.89208984375,
0.625,
0.89208984375,
0.63818359375,
0.50244140625,
0.63818359375
],
[
0.07843017578125,
0.64013671875,
0.468017578125,
0.6416015625,
0.468017578125,
0.65576171875,
0.07843017578125,
0.65478515625
],
[
0.50341796875,
0.6416015625,
0.8896484375,
0.6416015625,
0.8896484375,
0.6513671875,
0.50341796875,
0.6513671875
],
[
0.080078125,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.50244140625,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66845703125,
0.50244140625,
0.66845703125
],
[
0.07843017578125,
0.671875,
0.468017578125,
0.671875,
0.468017578125,
0.68603515625,
0.07843017578125,
0.68603515625
],
[
0.5009765625,
0.67041015625,
0.89208984375,
0.66943359375,
0.89208984375,
0.68310546875,
0.5009765625,
0.6845703125
],
[
0.080078125,
0.6875,
0.468017578125,
0.6875,
0.468017578125,
0.70068359375,
0.080078125,
0.70068359375
],
[
0.50244140625,
0.68505859375,
0.890625,
0.68505859375,
0.890625,
0.69873046875,
0.50244140625,
0.69873046875
],
[
0.080078125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50244140625,
0.70068359375,
0.89208984375,
0.70068359375,
0.89208984375,
0.7138671875,
0.50244140625,
0.7138671875
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.080078125,
0.73291015625,
0.461669921875,
0.73291015625,
0.461669921875,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.50244140625,
0.73095703125,
0.89208984375,
0.73095703125,
0.89208984375,
0.744140625,
0.50244140625,
0.744140625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.74560546875,
0.89208984375,
0.74560546875,
0.89208984375,
0.7587890625,
0.50244140625,
0.7587890625
],
[
0.5009765625,
0.7587890625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7744140625,
0.5009765625,
0.7734375
],
[
0.07843017578125,
0.76513671875,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.07843017578125,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50341796875,
0.77587890625,
0.89208984375,
0.77587890625,
0.89208984375,
0.7890625,
0.50341796875,
0.7890625
],
[
0.5009765625,
0.7890625,
0.89208984375,
0.79052734375,
0.89208984375,
0.8046875,
0.5009765625,
0.8037109375
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.80615234375,
0.7900390625,
0.80615234375,
0.7900390625,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.4697265625,
0.88525390625,
0.4697265625,
0.8994140625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"parameter",
"Label Number",
"In order to adjust the feature extractor efficiently,oracle",
"Method",
"400",
"2500",
"10000",
"class-wise embeddings of target dataset should be an ideal",
"Fine-Tuning (baseline)",
"31.69",
"21.74",
"choice.However, it is unrealistic due to the lack of annota-",
"60.79",
"Pseudo-Labeling [20]",
"59.21",
"tionsforunlabeled samples and theinherentbias inthe pre-",
"MT [28]",
"60.68",
"trained feature extractor. Thus, we propose to progressively",
"-",
"UDA [35]",
"58.32",
"update these class-wise embeddings from both labeled and",
"FixMatch↑ [26]",
"52.88",
"25.63",
"18.38",
"unlabeled data. For labeled data, μe is updated via",
"40.41",
"23.19",
"17.73",
"μnew = βμold +(1 -β)f(Ai(ai)1(ut = c), (6)",
"FlexMatcht[44]",
"Self-Tuning [32]",
"47.17",
"24.16",
"17.57",
"45.48",
"23.12",
"16.89",
"and for unlabeled data, μe is updated via",
"Ourst",
"Ours+t",
"22.06",
"16.58",
"Table 1. Error rates (%) on CIFAR-100 with EfficientNet-B2. +",
"37.36",
"μpew = βμld + (1 -β)f(Ai(a)1(p(clAo(r) ≥T)",
"means ours implementation based on [32].",
"(7)",
"extension Ours+. Note that it shows our approach can still",
"where the indicator function 1(yi = c) selects samples",
"boost the performance even with more advanced SSL al-",
"from the c-th class from the labeled data, the indicator func-",
"gorithms. In our work, all experiments were implemented",
"tion 1(pt (cAo(αu)) ≥ T) selects unlabeled samples that",
"in PyTorch and run on a GeForce RTX 2080Ti GPU with",
"11GB memory.",
"are confidently classified into the c-th class by the classifier.",
"This is identical to the criterion of generating pseudo-labels.",
"Pretrained models: Following [32], three models pre-",
"β is a momentum term that controls how far the class-wise",
"trained on ImageNet [9 are chosen for evaluation, i.e.",
"feature reaches into embedding history.",
"ResNet-50 [14] and EfficientNet [27] which are pretrained",
"in a supervised way, and a ResNet-50 which is trained",
"5. Experimental results",
"through an unsupervised learning method MoCo v2 [13].",
"In this section, we compare our approach with several",
"5.2. Train from Supervised Pretrained Models",
"SSL methods with pretrained models.",
"In this section, we compare various SSL methods trained",
"5.1. Experimental details",
"from supervised pretrained models.",
"We strictly follow [32] to design our experiment, includ-",
"Fine-grained Classification benchmarks:We use a",
"ing the evaluation datasets and pretrained model choices.",
"ResNet-50 network, which is supervised pretrained on Im-",
"We made such a choice since Self-Tuning [32] has demon-",
"ageNet, to initialize all SSL models. The results are shown",
"strated the state-of-the-art performance and its experimen-",
"in Table 2. It is clear that our proposed method achieves",
"tal evaluation is comprehensive and realistic. Some experi-",
"overall significant improvement than other comparing SSL",
"mental details are as follows:",
"approaches. Specifically, compared with traditional SSL",
"Datasets: Following the protocol of [32], which ad-",
"methods, our approach increases the test accuracy by a large",
"dressed the same research problem as this paper, four vision",
"margin onallkinds of partitions of threebenchmarks.Tak-",
"benchmarks are evaluated, i.e., FGVC Aircraft [21], Stan-",
"ing the state-of-the-art method FixMatch [26] as an exam-",
"ford Cars [16], CUB-200-2011 [31], and CIFAR-100 [17].",
"ple, the performance gain of our approach exceeds 10 per-",
"Specifically, the first three are challenging fine-grained clas-",
"cent on both Stanford Cars and CUB-200-2011 with 15%",
"sification datasets and label proportions ranging from 15%",
"labels. This is thanks to the proposed feature adjustment",
"to 50% are tested. Label partition of CIFAR-100 follows the",
"module in our approach which greatly reduces the bias in-",
"standard SSL protocol: 4/25/100 labeled images per class.",
"herented in the pretrained model. Furthermore, our ap-",
"Methods: Nine popular deep SSL approaches are included",
"proach is also superior to the recently proposed Self-Tuning",
"method [32] especially when labels are limited, e.g., only",
"for comparison, i.e., II-model [19], Pseudo-Labeling [20],",
"Mean Teacher [28], UDA [35], FixMatch [26], Flex-",
"15% training samples are labeled. When the CPL mod-",
"ule proposed in FlexMatch [44] is added to our approach,",
"Match [44], SimCLRv2 [6], FixMatch+AKC+ARC[1] and",
"Self-Tuning [32]. Meanwhile, performance of Fine-Tuning",
"Ours+ leads to a further performance boost.",
"on labeled data is also reported for a reference baseline. For",
"Standard SSL benchmarks: We choose CIFAR-100",
"our approach, we also consider a simple extension by in-",
"dataset [17] which is one of the most challenging datasets",
"corporating it into the recently proposed consistency-based",
"among standard SSL benchmarks to evaluate SSL methods",
"SSL method FlexMatch [44], which uses dynamically as-",
"from a pretrained model. Due to the lack of open-resourced",
"signed threshold with a FixMatch framework. We call this",
"pretrained checkpoints on WideResNet-28-8 model [41],"
] |
{
"arxiv_id": "2309.04659",
"language": "en",
"page_idx": 4
}
|
||
2303.13651-en-0
|
2303.13651-en-0.jpg
|
<table>
<tbody>
<tr>
<td>$\mathbf{B}$UILDING $\textbf{i AR}$ DOMA</td>
</tr>
<tr>
<td> </td>
</tr>
</tbody>
</table>
Danyal Akarca
University of Cambridge
danyal. akarca@ mrc- cbu. cam. ac. uk
Jascha Achterberg
University of Cambridge & Intel Labs
jascha. achterberg@mrc-cbu. cam. ac. uk
Moataz Assem
University of Cambridge
moataz. assem@mrc- cbu. cam. ac. uk
Moritz Heimbach
Julius-Maximilians-Universität Würzburg
moritz. heimbach@uni-wuerzburg. de
Duncan E. Astle
University of Cambridge
duncan. astle@mrc- cbu. cam. ac. uk
John Duncan
University of Cambridge
john. duncan@ mrc- cbu. cam. ac. uk
## $\mathbf{A}\mathbf{BSTRACT}$
There is a concerted effort to build domain-general artificial intelligence in the form of universal neural network models with sufficient computational flexibility to solve a wide variety of cognitive tasks but without requiring fine-tuning on individual problem spaces and domains. To do this, models need appropriate priors and inductive biases, such that trained models can generalise to out-of-distribution examples and new problem sets. Here we provide an overview of the hallmarks endowing biological neural networks with the functionality needed for flexible cognition, in order to establish which features might also be important to achieve similar functionality in artificial systems. We specifically discuss the of system-level distribution of network communication and recurrence, in addition to the role of short-term topological changes for efficient local computation. As machine learning models become more complex, these principles may provide valuable directions in an otherwise vast space of possible architectures. In addition, testing these inductive biases within artificial systems may help us to understand the biological principles underlying domain-general cognition
$K$eywords domain-general·multimodal·cognition·neural networks·brain
## 1 Introduction
An aspiration of machine learning research is not just to create architectures capable of achieving increasingly high levels of task-specific performance, but the genesis of models able to achieve good performance across different domains simultaneously. Recent striking advances in network models have enabled them to solve many problems within a domain with just one architecture [1,2,3]. Additionally, networks are increasingly acquiring multimodal capabilities [4,5] and learn in open-ended task environments [6,7]. These advances provide necessary building blocks for models capable of domain general cognition, as observed in intelligent human behaviour. Crucially, these new models may be able to go beyond simple generalisation to unseen data [8]; they may be able to learn new abilities and directly abstract them, allowing for generalisation across entire input modalities and the reuse of skills learned in one domain to support
$^{*}$This manuscript is part of the AAAI 2023 Spring Symposium on the Evaluation and Design of Generalist Systems (EDGeS)
|
[
[
0.2254638671875,
0.12371826171875,
0.77392578125,
0.1256103515625,
0.77392578125,
0.143310546875,
0.2254638671875,
0.141357421875
],
[
0.31201171875,
0.1514892578125,
0.69140625,
0.1514892578125,
0.69140625,
0.1654052734375,
0.31201171875,
0.1654052734375
],
[
0.214111328125,
0.177978515625,
0.78515625,
0.177978515625,
0.78515625,
0.1912841796875,
0.214111328125,
0.1912841796875
],
[
0.40185546875,
0.2008056640625,
0.59326171875,
0.199462890625,
0.59326171875,
0.2171630859375,
0.40185546875,
0.2183837890625
],
[
0.5888671875,
0.20263671875,
0.6044921875,
0.20263671875,
0.6044921875,
0.2108154296875,
0.5888671875,
0.2108154296875
],
[
0.0228729248046875,
0.2587890625,
0.060455322265625,
0.2587890625,
0.060455322265625,
0.373779296875,
0.0228729248046875,
0.373779296875
],
[
0.2646484375,
0.27978515625,
0.400390625,
0.281005859375,
0.400390625,
0.29541015625,
0.2646484375,
0.294189453125
],
[
0.63134765625,
0.281005859375,
0.740234375,
0.281005859375,
0.740234375,
0.294189453125,
0.63134765625,
0.294189453125
],
[
0.2034912109375,
0.29345703125,
0.455078125,
0.294189453125,
0.455078125,
0.309326171875,
0.2034912109375,
0.30810546875
],
[
0.60302734375,
0.29345703125,
0.76708984375,
0.294189453125,
0.76708984375,
0.309326171875,
0.60302734375,
0.30810546875
],
[
0.1805419921875,
0.311279296875,
0.48046875,
0.311279296875,
0.48046875,
0.3212890625,
0.1805419921875,
0.3212890625
],
[
0.55126953125,
0.309326171875,
0.8203125,
0.30810546875,
0.8203125,
0.322509765625,
0.55126953125,
0.323974609375
],
[
0.268798828125,
0.34521484375,
0.373291015625,
0.34521484375,
0.373291015625,
0.35791015625,
0.268798828125,
0.35791015625
],
[
0.60693359375,
0.3447265625,
0.73388671875,
0.3447265625,
0.73388671875,
0.35791015625,
0.60693359375,
0.35791015625
],
[
0.236083984375,
0.356689453125,
0.40185546875,
0.35791015625,
0.40185546875,
0.37255859375,
0.236083984375,
0.371826171875
],
[
0.53173828125,
0.35791015625,
0.80810546875,
0.35791015625,
0.80810546875,
0.37255859375,
0.53173828125,
0.37255859375
],
[
0.02532958984375,
0.3662109375,
0.058013916015625,
0.3662109375,
0.058013916015625,
0.3984375,
0.02532958984375,
0.3984375
],
[
0.1895751953125,
0.37255859375,
0.450927734375,
0.37255859375,
0.450927734375,
0.386474609375,
0.1895751953125,
0.386474609375
],
[
0.529296875,
0.37255859375,
0.80810546875,
0.37255859375,
0.80810546875,
0.386474609375,
0.529296875,
0.386474609375
],
[
0.021240234375,
0.404052734375,
0.060455322265625,
0.404052734375,
0.060455322265625,
0.486083984375,
0.021240234375,
0.486083984375
],
[
0.271240234375,
0.4072265625,
0.386474609375,
0.4072265625,
0.386474609375,
0.421630859375,
0.271240234375,
0.421630859375
],
[
0.6259765625,
0.4072265625,
0.72216796875,
0.4072265625,
0.72216796875,
0.421630859375,
0.6259765625,
0.421630859375
],
[
0.2467041015625,
0.421630859375,
0.410888671875,
0.421630859375,
0.410888671875,
0.43505859375,
0.2467041015625,
0.43505859375
],
[
0.041656494140625,
0.410400390625,
0.051483154296875,
0.41357421875,
0.047393798828125,
0.419189453125,
0.037567138671875,
0.416748046875
],
[
0.59228515625,
0.421630859375,
0.75732421875,
0.421630859375,
0.75732421875,
0.436279296875,
0.59228515625,
0.436279296875
],
[
0.1993408203125,
0.436279296875,
0.4599609375,
0.436279296875,
0.4599609375,
0.449462890625,
0.1993408203125,
0.449462890625
],
[
0.54833984375,
0.436279296875,
0.7998046875,
0.436279296875,
0.7998046875,
0.449462890625,
0.54833984375,
0.449462890625
],
[
0.021240234375,
0.49560546875,
0.05963134765625,
0.49560546875,
0.05963134765625,
0.59521484375,
0.021240234375,
0.59521484375
],
[
0.451904296875,
0.5,
0.54736328125,
0.5,
0.54736328125,
0.5146484375,
0.451904296875,
0.5146484375
],
[
0.035125732421875,
0.52294921875,
0.0400390625,
0.52294921875,
0.0400390625,
0.52978515625,
0.035125732421875,
0.52978515625
],
[
0.1748046875,
0.521484375,
0.82421875,
0.521484375,
0.82421875,
0.53515625,
0.1748046875,
0.53515625
],
[
0.176513671875,
0.537109375,
0.82421875,
0.537109375,
0.82421875,
0.54736328125,
0.176513671875,
0.54736328125
],
[
0.1748046875,
0.548828125,
0.82421875,
0.548828125,
0.82421875,
0.5625,
0.1748046875,
0.5625
],
[
0.1732177734375,
0.5625,
0.82275390625,
0.5625,
0.82275390625,
0.5771484375,
0.1732177734375,
0.5771484375
],
[
0.0228729248046875,
0.57470703125,
0.060455322265625,
0.57470703125,
0.060455322265625,
0.7138671875,
0.0228729248046875,
0.7138671875
],
[
0.041656494140625,
0.57568359375,
0.0482177734375,
0.57568359375,
0.0482177734375,
0.58203125,
0.041656494140625,
0.58203125
],
[
0.1732177734375,
0.57763671875,
0.82421875,
0.5771484375,
0.82421875,
0.59033203125,
0.1732177734375,
0.59130859375
],
[
0.1732177734375,
0.59130859375,
0.82421875,
0.59033203125,
0.82421875,
0.60400390625,
0.1732177734375,
0.60498046875
],
[
0.1732177734375,
0.60400390625,
0.82275390625,
0.60400390625,
0.82275390625,
0.6171875,
0.1732177734375,
0.6171875
],
[
0.1748046875,
0.61865234375,
0.82421875,
0.61865234375,
0.82421875,
0.6318359375,
0.1748046875,
0.6318359375
],
[
0.1748046875,
0.6337890625,
0.82275390625,
0.6337890625,
0.82275390625,
0.64453125,
0.1748046875,
0.64453125
],
[
0.1748046875,
0.646484375,
0.82421875,
0.646484375,
0.82421875,
0.66015625,
0.1748046875,
0.66015625
],
[
0.1748046875,
0.66015625,
0.826171875,
0.66015625,
0.826171875,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1748046875,
0.673828125,
0.75732421875,
0.673828125,
0.75732421875,
0.6875,
0.1748046875,
0.6875
],
[
0.11517333984375,
0.7041015625,
0.62158203125,
0.7041015625,
0.62158203125,
0.71826171875,
0.11517333984375,
0.71826171875
],
[
0.11767578125,
0.74365234375,
0.1298828125,
0.74365234375,
0.1298828125,
0.75390625,
0.11767578125,
0.75390625
],
[
0.136474609375,
0.7421875,
0.253173828125,
0.7421875,
0.253173828125,
0.7568359375,
0.136474609375,
0.7568359375
],
[
0.11517333984375,
0.7744140625,
0.88232421875,
0.7744140625,
0.88232421875,
0.78857421875,
0.11517333984375,
0.78857421875
],
[
0.11517333984375,
0.78857421875,
0.8837890625,
0.78857421875,
0.8837890625,
0.80322265625,
0.11517333984375,
0.80322265625
],
[
0.11517333984375,
0.8017578125,
0.8837890625,
0.8017578125,
0.8837890625,
0.81494140625,
0.11517333984375,
0.81494140625
],
[
0.11517333984375,
0.81640625,
0.8837890625,
0.81640625,
0.8837890625,
0.830078125,
0.11517333984375,
0.830078125
],
[
0.11517333984375,
0.830078125,
0.8837890625,
0.830078125,
0.8837890625,
0.84326171875,
0.11517333984375,
0.84326171875
],
[
0.11517333984375,
0.84326171875,
0.8837890625,
0.84326171875,
0.8837890625,
0.85693359375,
0.11517333984375,
0.85693359375
],
[
0.11517333984375,
0.85791015625,
0.8837890625,
0.85791015625,
0.8837890625,
0.8720703125,
0.11517333984375,
0.8720703125
],
[
0.11517333984375,
0.8720703125,
0.8837890625,
0.8720703125,
0.8837890625,
0.88525390625,
0.11517333984375,
0.88525390625
],
[
0.136474609375,
0.8994140625,
0.875,
0.8994140625,
0.875,
0.91357421875,
0.136474609375,
0.91357421875
]
] |
[
"BUILDING ARTIFICIAL NEURAL CIRCUITS FOR",
"DOMAIN-GENERAL COGNITION:",
"A PRIMER ON BRAIN-INSPIRED SYSTEMS-LEVEL",
"ARCHITECTURE",
"",
"",
"Jascha Achterberg",
"Danyal Akarca",
"University of Cambridge & Intel Labs",
"Universityof Cambridge",
"jascha.achterberg@mrc-cbu.cam.ac.uk",
"danyal.akarca@mrc-cbu.cam.ac.uk",
"Moataz Assem",
"Moritz Heimbach",
"University of Cambridge",
"Julius-Maximilians-Universitat Wiurzburg",
"2",
"moataz.assem@mrc-cbu.cam.ac.uk",
"moritz.heimbach@uni-wuerzburg.de",
"[FN'S]",
"Duncan E. Astle",
"John Duncan",
"University of Cambridge",
"",
"University of Cambridge",
"duncan.astle@mrc-cbu.cam.ac.uk",
"john.duncan@mrc-cbu.cam.ac.uk",
"AS6",
"ABSTRACT",
"",
"There is a concerted effort to build domain-general artificial intelligence in the form of universal neural",
"networkmodelswithsufficientcomputationalflexibilitytosolveawidevarietyofcognitivetasksbut",
"without requiring fine-tuning on individual problem spaces and domains. To do this, models need",
"appropriate priors and inductive biases, such that trained models can generalise to out-of-distribution",
"I'COSZAIXI",
"",
"examples andnewproblem sets.Here we provide an overviewofthehallmarks endowing biological",
"neural networks with the functionality needed forflexible cognition, in order to establish which",
"features might also be important to achieve similar functionality in artificial systems. We specifically",
"discuss the role of system-level distribution of network communication and recurrence, in addition",
"totheroleofshort-termtopologicalchangesforefficientlocalcomputation.Asmachinelearning",
"models become more complex, these principles may provide valuable directions in an otherwise vast",
"space of possible architectures. In addition, testing these inductive biases within artificial systems",
"may help us to understand the biological principles underlying domain-general cognition.",
"Keywords domain-general -multimodal ·cognition -neural networks -brain",
"1",
"Introduction",
"An aspiration of machine learning research is not just to create architectures capable of achieving increasingly high",
"levels of task-specific performance, but the genesis of models able to achieve good performance across different domains",
"simultaneously. Recent striking advances in network models have enabled them to solve many problems within a",
"domain with just one architecture [1,2, 3].Additionally, networks are increasingly acquiring multimodal capabilities",
"[4, 5] and learn in open-ended task environments [6, 7]. These advances provide necessary building blocks for models",
"capable of domain general cognition, as observed in intelligent human behaviour. Crucially, these new models may be",
"able to go beyond simple generalisation to unseen data [8]; they may be able to learn new abilities and directly abstract",
"them, allowing for generalisation across entire input modalities and the reuse of skills learned in one domain to support",
"*This manuscript is part of the AAAI 2023 Spring Symposium on the Evaluation and Design of Generalist Systems (EDGeS)"
] |
{
"arxiv_id": "2303.13651",
"language": "en",
"page_idx": 0
}
|
||
2303.13651-en-1
|
2303.13651-en-1.jpg
|
Building artificial neural circuits for domain-general cognition
learning in entirely new domains. Indeed, this parallels how children learn over the course of their own development
[9]. However, the extent to which current models can achieve this remains limited.
For decades, neuroscientists have been focused on identifying core features of the brain's structural and functional architecture. This allows us to connect our knowledge of human neural architectures that enable ffexible domain-general cognition[10], with ideas on how we hope to achieve similar capabilities in artificial systems. Here we provide an overview of mechanisms underlying domain-general cognition in biological neural networks to derive which features of the systems-level architecture may be important to build flexible multimodal problem-solving capabilities into artificial systems. Previously published reviews have already outlined which cognitive ideas and modules might be essential $[ 11, 12, 13, 14, 15] . \text{ \^ {W}e aim to expand these cognitive perspectives by providing a brief introduction to the system- level}$ network structure underlying domain-general cognition in the brain, highlighting what structural optimisation processes we think could be used in machine learning models. In this, our goal is not to hard-code brain-like anatomy into a network model's architecture. Instead, we aim to identify computationally beneficial structural motifs which can be soft-coded into the network's learning process to serve as helpful inductive biases or priors. As we see increasingly complex machine learning models being built as a combination of functional submodules [16,5], we believe that the system-level priors we outline may provide helpful guidance to coordinate information flow in the most complex artificial neural networks [17].
2 A core domain-general network in the brair
The human brain, as with many complex physical systems, is economically organised to balance numerous competing objectives-including metabolic, computational, and geometric[18,19]. These objectives have a strong influence on the topology of the brain's network; not only is it energetically expensive to fully build and sustain neural connections [20,21]butitis highly costly to constantly communicate signals between neurons and assemblies of neurons, particularly over longer distances [22]. Owing to its size, complexity, and these economic considerations, it is infeasible for each neural region to communicate directly with every other region equivalently [23]. To avoid this problem, evolutionary pressures have guided the brain towards a modularised network, with modules of very strong local connectivity and $\hat{\text{high-connection hub nodes connecting across these modules [24,25]. }}$Networks with this structure are described as having“small-world”characteristics, defined as having concurrently a highly clustered topology and short path lengths, meeting a balance between totally random versus regular networks [26,27]. Small-world structures are commonly found in distributed systems under resource constraints, showing patterns of locally specialised computation alongside good propagation of signals within and between hubs. In brains, this locality of computations results in concentration of specific cognitive function within specific anatomical regions. Specialised regions act as foci for cognitive functions like sensory processing, semantic knowledge, and language abilities [28, 29, 30]. These are likely semi-specialised, meaning that they mostly focus on unique local computation but also partially integrate meaningful information across areas and domains [31, $\dot{3}2].$
This picture of a functionally modular system becomes more nuanced when we consider human domain-general cognition. As the tasks to be solved become more complicated, the brain increasingly abandons solely relying on its specialised modular structure. Instead, neural architectures must increasingly integrate signals across modules[33] and rely on its Multiple Demand system (MD) [10,34]. This is a core network in the brain (depicted in Figure 1A) which is highly active when a complex task of any nature is solved. It is thought that the MD system serves as a central processing unit, receiving information from more specialised input notes, to compress it into meaningful abstract representations on which it can run problem solving algorithms (schematic shown in Figure 1B). It also plays a central role in controlling information in other brain regions, using knowledge and complex analysis of the situation to control thought processes in specialised brain regions through top-down control processes [10,35,36,37,38]. Ultimately it is this central processing circuit that likely gives the primate brain the ability to have abstract thoughts used to solve complex problems to reach long-term goals.
What are the key principles underpinning this system? In the following we will discuss three computational / structural motifs which vary across the hierarchy from specialised regions to the integrative MD system, which allow the network to show domain-general cognitive skills. These are: Recurrence, communicability, and short-term topological changes We will review each of these in terms of their relevance in biological networks before then discussing possible directions for artificial implementations, in the context of related existing implementations.
2
|
[
[
0.1168212890625,
0.044189453125,
0.52783203125,
0.044189453125,
0.52783203125,
0.058074951171875,
0.1168212890625,
0.058074951171875
],
[
0.1168212890625,
0.0921630859375,
0.88232421875,
0.0921630859375,
0.88232421875,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.1168212890625,
0.106689453125,
0.65625,
0.106689453125,
0.65625,
0.12060546875,
0.1168212890625,
0.12060546875
],
[
0.11517333984375,
0.126953125,
0.88134765625,
0.126953125,
0.88134765625,
0.14013671875,
0.11517333984375,
0.14013671875
],
[
0.11517333984375,
0.141357421875,
0.88134765625,
0.141357421875,
0.88134765625,
0.1546630859375,
0.11517333984375,
0.1546630859375
],
[
0.11517333984375,
0.1546630859375,
0.88232421875,
0.1546630859375,
0.88232421875,
0.1685791015625,
0.11517333984375,
0.1685791015625
],
[
0.11517333984375,
0.1685791015625,
0.88232421875,
0.1685791015625,
0.88232421875,
0.1817626953125,
0.11517333984375,
0.1817626953125
],
[
0.11517333984375,
0.18310546875,
0.88232421875,
0.18310546875,
0.88232421875,
0.1962890625,
0.11517333984375,
0.1962890625
],
[
0.11517333984375,
0.1962890625,
0.88232421875,
0.1962890625,
0.88232421875,
0.210205078125,
0.11517333984375,
0.210205078125
],
[
0.1168212890625,
0.210205078125,
0.88232421875,
0.210205078125,
0.88232421875,
0.2235107421875,
0.1168212890625,
0.2235107421875
],
[
0.11517333984375,
0.2235107421875,
0.8837890625,
0.2235107421875,
0.8837890625,
0.2374267578125,
0.11517333984375,
0.2374267578125
],
[
0.11517333984375,
0.2374267578125,
0.8837890625,
0.2374267578125,
0.8837890625,
0.250732421875,
0.11517333984375,
0.250732421875
],
[
0.11517333984375,
0.251953125,
0.88232421875,
0.251953125,
0.88232421875,
0.26513671875,
0.11517333984375,
0.26513671875
],
[
0.1168212890625,
0.267578125,
0.8798828125,
0.267578125,
0.8798828125,
0.27783203125,
0.1168212890625,
0.27783203125
],
[
0.11517333984375,
0.279052734375,
0.88232421875,
0.279052734375,
0.88232421875,
0.292236328125,
0.11517333984375,
0.292236328125
],
[
0.11517333984375,
0.292236328125,
0.8837890625,
0.292236328125,
0.8837890625,
0.306884765625,
0.11517333984375,
0.306884765625
],
[
0.1168212890625,
0.30810546875,
0.313720703125,
0.30810546875,
0.313720703125,
0.318115234375,
0.1168212890625,
0.318115234375
],
[
0.11517333984375,
0.39453125,
0.51318359375,
0.39453125,
0.51318359375,
0.4091796875,
0.11517333984375,
0.4091796875
],
[
0.11517333984375,
0.458251953125,
0.88232421875,
0.458251953125,
0.88232421875,
0.4716796875,
0.11517333984375,
0.4716796875
],
[
0.11517333984375,
0.4716796875,
0.88232421875,
0.4716796875,
0.88232421875,
0.48486328125,
0.11517333984375,
0.48486328125
],
[
0.1168212890625,
0.48486328125,
0.88232421875,
0.48486328125,
0.88232421875,
0.498779296875,
0.1168212890625,
0.498779296875
],
[
0.11767578125,
0.50048828125,
0.88134765625,
0.50048828125,
0.88134765625,
0.51123046875,
0.11767578125,
0.51123046875
],
[
0.1168212890625,
0.5146484375,
0.88134765625,
0.5146484375,
0.88134765625,
0.5244140625,
0.1168212890625,
0.5244140625
],
[
0.11517333984375,
0.5263671875,
0.88232421875,
0.5263671875,
0.88232421875,
0.54052734375,
0.11517333984375,
0.54052734375
],
[
0.1168212890625,
0.54052734375,
0.88232421875,
0.54052734375,
0.88232421875,
0.5537109375,
0.1168212890625,
0.5537109375
],
[
0.1168212890625,
0.5537109375,
0.8837890625,
0.5537109375,
0.8837890625,
0.5673828125,
0.1168212890625,
0.5673828125
],
[
0.11517333984375,
0.5673828125,
0.8837890625,
0.5673828125,
0.8837890625,
0.58056640625,
0.11517333984375,
0.58056640625
],
[
0.1168212890625,
0.58203125,
0.88232421875,
0.58203125,
0.88232421875,
0.59521484375,
0.1168212890625,
0.59521484375
],
[
0.1168212890625,
0.5947265625,
0.88134765625,
0.5947265625,
0.88134765625,
0.60791015625,
0.1168212890625,
0.60791015625
],
[
0.1168212890625,
0.609375,
0.8837890625,
0.609375,
0.8837890625,
0.62255859375,
0.1168212890625,
0.62255859375
],
[
0.11517333984375,
0.62353515625,
0.88232421875,
0.62353515625,
0.88232421875,
0.63720703125,
0.11517333984375,
0.63720703125
],
[
0.11517333984375,
0.63623046875,
0.8857421875,
0.63623046875,
0.8857421875,
0.6494140625,
0.11517333984375,
0.6494140625
],
[
0.11517333984375,
0.65087890625,
0.8837890625,
0.65087890625,
0.8837890625,
0.6640625,
0.11517333984375,
0.6640625
],
[
0.11517333984375,
0.6640625,
0.29736328125,
0.6640625,
0.29736328125,
0.67822265625,
0.11517333984375,
0.67822265625
],
[
0.11517333984375,
0.6845703125,
0.88134765625,
0.6845703125,
0.88134765625,
0.69775390625,
0.11517333984375,
0.69775390625
],
[
0.11358642578125,
0.69775390625,
0.88232421875,
0.69873046875,
0.88232421875,
0.71337890625,
0.11358642578125,
0.7119140625
],
[
0.11517333984375,
0.7119140625,
0.88134765625,
0.7119140625,
0.88134765625,
0.72607421875,
0.11517333984375,
0.72607421875
],
[
0.11517333984375,
0.72607421875,
0.8837890625,
0.72607421875,
0.8837890625,
0.7392578125,
0.11517333984375,
0.7392578125
],
[
0.11517333984375,
0.74072265625,
0.8837890625,
0.74072265625,
0.8837890625,
0.75390625,
0.11517333984375,
0.75390625
],
[
0.11517333984375,
0.75390625,
0.88232421875,
0.75390625,
0.88232421875,
0.767578125,
0.11517333984375,
0.767578125
],
[
0.11358642578125,
0.7666015625,
0.88232421875,
0.767578125,
0.88232421875,
0.7822265625,
0.11358642578125,
0.78076171875
],
[
0.11517333984375,
0.78076171875,
0.88232421875,
0.78076171875,
0.88232421875,
0.79443359375,
0.11517333984375,
0.79443359375
],
[
0.1168212890625,
0.79443359375,
0.88134765625,
0.79443359375,
0.88134765625,
0.80810546875,
0.1168212890625,
0.80810546875
],
[
0.11517333984375,
0.8095703125,
0.88232421875,
0.8095703125,
0.88232421875,
0.82275390625,
0.11517333984375,
0.82275390625
],
[
0.1168212890625,
0.82470703125,
0.40283203125,
0.82470703125,
0.40283203125,
0.83544921875,
0.1168212890625,
0.83544921875
],
[
0.11358642578125,
0.84130859375,
0.88232421875,
0.84228515625,
0.88232421875,
0.85693359375,
0.11358642578125,
0.85595703125
],
[
0.11517333984375,
0.85791015625,
0.88232421875,
0.85791015625,
0.88232421875,
0.8720703125,
0.11517333984375,
0.8720703125
],
[
0.11517333984375,
0.87060546875,
0.8837890625,
0.87060546875,
0.8837890625,
0.8837890625,
0.11517333984375,
0.8837890625
],
[
0.11358642578125,
0.88330078125,
0.8837890625,
0.8837890625,
0.8837890625,
0.8984375,
0.11358642578125,
0.89794921875
],
[
0.11358642578125,
0.89794921875,
0.6435546875,
0.8984375,
0.6435546875,
0.912109375,
0.11358642578125,
0.9111328125
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"Building artificial neural circuits for domain-general cognition",
"learning in entirely new domains. Indeed, this parallels how children learn over the course of their own development",
"[9]. However, the extent to which current models can achieve this remains limited.",
"For decades, neuroscientists have been focused on identifying core features of the brain's structural and functional",
"architecture. This allows us to connect our knowledge of human neural architectures that enable flexible domain-general",
"cognition [1O], with ideas on how we hope to achieve similar capabilities in artificial systems. Here we provide an",
"overview of mechanisms underlying domain-general cognition in biological neural networks to derive which features of",
"the systems-level architecture may be important to build flexible multimodal problem-solving capabilities into artificial",
"systems. Previously published reviews have already outlined which cognitive ideas and modules might be essential",
"[11, 12, 13, 14, 15]. We aim to expand these cognitive perspectives by providing a brief introduction to the system-level",
"network structure underlying domain-general cognition in the brain, highlighting what structural optimisation processes",
"we think could be used in machine learning models. In this, our goal is not to hard-code brain-like anatomy into a",
"network model's architecture. Instead, we aim to identify computationally beneficial structural motifs which can be",
"soft-coded intothenetworkslearningprocesstoserveashelpfulinductivebiasesorpriors.Asweseeincreasingly",
"complex machine learning models being built as a combination of functional submodules [16, 5], we believe that",
"the system-level priors we outline may provide helpful guidance to coordinate information flow in the most complex",
"artificialneuralnetworks[17]",
"2A core domain-general network in the brain",
"The human brain, as with many complex physical systems, is economically organised to balance numerous competing",
"objectives -including metabolic, computational, and geometric [18, 19]. These objectives have a strong influence on",
"the topology of the brain's network; not only is it energetically expensive to fully build and sustain neural connections",
"[20,21]butitishighlycostlytoconstantlycommunicatesignalsbetweenneuronsandassembliesofneurons,particularly",
"overlongerdistances[22l.Owingtoitssize,complexity,andtheseeconomicconsiderations,itisinfeasibleforeach",
"neural region to communicate directly with every other region equivalently [23]. To avoid this problem, evolutionary",
"pressures have guided the brain towards a modularised network, with modules of very strong local connectivity and",
"high-connection hub nodes connecting across these modules [24, 25]. Networks with this structure are described as",
"having \"small-world\" characteristics, defined as having concurrently a highly clustered topology and short path lengths,",
"meeting a balance between totally random versus regular networks [26, 27]. Small-world structures are commonly",
"found in distributed systems under resource constraints, showing patterns of locally specialised computation alongside",
"good propagation of signals within and between hubs. In brains, this locality of computations results in concentration",
"of specific cognitive function within specific anatomical regions. Specialised regions act as foci for cognitive functions",
"like sensory processing, semantic knowledge, and language abilities [28, 29, 30]. These are likely semi-specialised,",
"meaning that they mostly focus on unique local computation but also partially integrate meaningful information across",
"areas and domains [31, 32].",
"This picture of a functionally modular system becomes more nuanced when we consider human domain-general",
"cognition. As the tasks to be solved become more complicated, the brain increasingly abandons solely relying on its",
"specialised modular structure. Instead, neural architectures must increasingly integrate signals across modules [33]",
"and rely on its Multiple Demand system (MD) [10, 34]. This is a core network in the brain (depicted in Figure 1A)",
"which is highly active when a complex task of any nature is solved. It is thought that the MD system serves as a",
"central processing unit, receiving information from more specialised input notes, to compress it into meaningful abstract",
"representations on which it can run problem solving algorithms (schematic shown in Figure 1B). It also plays a central",
"role in controlling information in other brain regions, using knowledge and complex analysis of the situation to control",
"thought processes in specialised brain regions through top-down control processes [10, 35, 36, 37, 38]. Ultimately it",
"is this central processing circuit that likely gives the primate brain the ability to have abstract thoughts used to solve",
"complexproblemstoreachlong-termgoals",
"What are the key principles underpinning this system? In the following we will discuss three computational / structural",
"motifs which vary across the hierarchy from specialised regions to the integrative MD system, which allow the network",
"to show domain-general cognitive skills. These are: Recurrence, communicability, and short-term topological changes.",
"We will review each of these in terms of their relevance in biological networks before then discussing possible directions",
"forartificial implementations, inthe contextofrelated existing implementations.",
"2"
] |
{
"arxiv_id": "2303.13651",
"language": "en",
"page_idx": 1
}
|
||
2303.13651-en-2
|
2303.13651-en-2.jpg
|
Building artificial neural circuits for domain-general cognition
<FigureHere>
Figure 1: A- The cortical areas forming the core Multiple Demand system in the human brain, from [34]. B- Schematic depiction of a systems-level view of the brain. The Multiple Demand system lies at the core of information processing in the brain, exchanging inputs with more specialised regions such as language, memory, sensory and social processing. Due to its central position, the MD core can influence computations in multiple specialised areas by broadcasting information it constructed from integrating across domains back to specialised regions, e.g, influencing perception by abstract understanding of the environment / situation at large. Refer to[34] for detailed anatomical perspective of the MD system's core and penumbra regions not discussed here.
3 Computational motifs supporting domain-general cognition
## 3.1 Global recurrence
Computations in functionally more specialised regions depend strongly on a feed-forward structure that extracts increasingly abstract features from sensory inputs[39,40,41]. Much work shows how this process can be modelled using an artificial feed-forward network [42, 43]. While there is also recurrent processing in these specialised systems [39,40,44],the recurrent loops in these systems are likely very local and cover relatively short distances and timescales This means that a signal sent from a node will only travel a short path before arriving back at its starting point. As we move towards more integrative and domain-general cognition, recurrent connections become a hallmark feature of the brain's systems-level design. The frontal cortex, where a large part of the MD system lies, is often thought of as implementing recurrent loops for abstract information processing [41,45]. Importantly, these loops not only process information locally but also broadcast information widely across the brain, influencing and controlling computations in specialised regions. It does so by not only having local recurrent connections within the circuit but also many loops spanning large distances in the brain, reaching out to nodes which lie far outside the core [37,46,41,36]. With nodes widely distributed over the cortex, coupled to strong communication between these nodes, the MD system is well positioned for widespread integration and communication. A large set of recursive processing loops with varied scales in terms of time and spatial distance likely facilitate the MD system's abstract domain-general processing and deliver the ability to coordinate computation in a large distributed system [10].
The use of recurrent loops in artificial neural networks has a long history [47]. They proved to be useful tools for processing and predicting time series data but also suffered from problems of vanishing gradient and computational complexity when capturing long-range dependencies in the input [48,49,50]. To avoid these issues, feed-forward based architectures can be used as substitutes [50] and various attention-based architectures have recently been very effective in capturing dependencies in language time courses and multiple other modalities [49,51]. This works by inputting an entire time series in a single time step so that the attention mechanism learns the relationship between timesteps without needing to hold past time points in memory. While these architectures likely can be good substitutes for the local recurrent loops, we believe that ultimately, researchers are going to have to find a way to also introduce global recurrent loops to arrive at domain-general cognition in artificial systems. Approaches like weight-sharing in $\tilde{\text{deep models paired with skip-connections may allow us to mimic a recurrent process in a regular forward pass but itt}}$ seems likely that alternative ways will be needed to allow abstract multimodal knowledge to be broadcasted through the network to inform distributed computations. This seems even more timely now that models generate impressive responses to inputs such as images or language [1,52], but struggle to be constrained by meaningful world models (e.g., intuitive physics, [15]). Instead, researchers rely on human feedback signals in the training pipeline [53]. As such,
3
|
[
[
0.11517333984375,
0.044189453125,
0.52783203125,
0.044189453125,
0.52783203125,
0.058074951171875,
0.11517333984375,
0.058074951171875
],
[
0.2459716796875,
0.15087890625,
0.261474609375,
0.15087890625,
0.261474609375,
0.1666259765625,
0.2459716796875,
0.1666259765625
],
[
0.42724609375,
0.1514892578125,
0.439453125,
0.1514892578125,
0.439453125,
0.1641845703125,
0.42724609375,
0.1641845703125
],
[
0.2998046875,
0.1634521484375,
0.374267578125,
0.1634521484375,
0.374267578125,
0.173583984375,
0.2998046875,
0.173583984375
],
[
0.3203125,
0.1749267578125,
0.3544921875,
0.1749267578125,
0.3544921875,
0.184326171875,
0.3203125,
0.184326171875
],
[
0.64892578125,
0.2197265625,
0.66259765625,
0.2197265625,
0.66259765625,
0.222900390625,
0.64892578125,
0.222900390625
],
[
0.62744140625,
0.274658203125,
0.69775390625,
0.274658203125,
0.69775390625,
0.2822265625,
0.62744140625,
0.2822265625
],
[
0.6435546875,
0.28271484375,
0.6806640625,
0.281005859375,
0.68115234375,
0.291748046875,
0.64453125,
0.29345703125
],
[
0.1168212890625,
0.32080078125,
0.88134765625,
0.32080078125,
0.88134765625,
0.333984375,
0.1168212890625,
0.333984375
],
[
0.11517333984375,
0.335205078125,
0.88232421875,
0.335205078125,
0.88232421875,
0.348388671875,
0.11517333984375,
0.348388671875
],
[
0.11358642578125,
0.347900390625,
0.8837890625,
0.348388671875,
0.8837890625,
0.363037109375,
0.11358642578125,
0.3623046875
],
[
0.11517333984375,
0.361083984375,
0.88232421875,
0.3623046875,
0.88232421875,
0.376953125,
0.11517333984375,
0.375732421875
],
[
0.1168212890625,
0.376953125,
0.88232421875,
0.376953125,
0.88232421875,
0.39013671875,
0.1168212890625,
0.39013671875
],
[
0.1168212890625,
0.389404296875,
0.88134765625,
0.389404296875,
0.88134765625,
0.40283203125,
0.1168212890625,
0.40283203125
],
[
0.1168212890625,
0.404052734375,
0.5146484375,
0.404052734375,
0.5146484375,
0.41845703125,
0.1168212890625,
0.41845703125
],
[
0.11767578125,
0.447509765625,
0.131591796875,
0.447509765625,
0.131591796875,
0.45703125,
0.11767578125,
0.45703125
],
[
0.143798828125,
0.4462890625,
0.64208984375,
0.4462890625,
0.64208984375,
0.4609375,
0.143798828125,
0.4609375
],
[
0.11358642578125,
0.4765625,
0.281982421875,
0.477783203125,
0.281982421875,
0.492431640625,
0.11358642578125,
0.4912109375
],
[
0.1168212890625,
0.5048828125,
0.88232421875,
0.5048828125,
0.88232421875,
0.51953125,
0.1168212890625,
0.51953125
],
[
0.1168212890625,
0.51953125,
0.8837890625,
0.51953125,
0.8837890625,
0.53271484375,
0.1168212890625,
0.53271484375
],
[
0.11517333984375,
0.53271484375,
0.88232421875,
0.53271484375,
0.88232421875,
0.54736328125,
0.11517333984375,
0.54736328125
],
[
0.1168212890625,
0.546875,
0.88232421875,
0.546875,
0.88232421875,
0.56005859375,
0.1168212890625,
0.56005859375
],
[
0.11767578125,
0.5625,
0.88232421875,
0.5625,
0.88232421875,
0.57275390625,
0.11767578125,
0.57275390625
],
[
0.11517333984375,
0.57470703125,
0.8837890625,
0.57470703125,
0.8837890625,
0.58837890625,
0.11517333984375,
0.58837890625
],
[
0.11517333984375,
0.58837890625,
0.88232421875,
0.58837890625,
0.88232421875,
0.6015625,
0.11517333984375,
0.6015625
],
[
0.11517333984375,
0.60302734375,
0.88232421875,
0.60302734375,
0.88232421875,
0.6162109375,
0.11517333984375,
0.6162109375
],
[
0.11517333984375,
0.6162109375,
0.88232421875,
0.6162109375,
0.88232421875,
0.6298828125,
0.11517333984375,
0.6298828125
],
[
0.11517333984375,
0.6298828125,
0.88232421875,
0.6298828125,
0.88232421875,
0.64306640625,
0.11517333984375,
0.64306640625
],
[
0.11517333984375,
0.64306640625,
0.88232421875,
0.64306640625,
0.88232421875,
0.6572265625,
0.11517333984375,
0.6572265625
],
[
0.11517333984375,
0.6572265625,
0.88232421875,
0.6572265625,
0.88232421875,
0.67041015625,
0.11517333984375,
0.67041015625
],
[
0.11517333984375,
0.671875,
0.8837890625,
0.671875,
0.8837890625,
0.68505859375,
0.11517333984375,
0.68505859375
],
[
0.11517333984375,
0.6845703125,
0.88232421875,
0.6845703125,
0.88232421875,
0.69775390625,
0.11517333984375,
0.69775390625
],
[
0.11358642578125,
0.69775390625,
0.580078125,
0.69873046875,
0.580078125,
0.71337890625,
0.11358642578125,
0.7119140625
],
[
0.11517333984375,
0.71826171875,
0.88232421875,
0.71826171875,
0.88232421875,
0.73193359375,
0.11517333984375,
0.73193359375
],
[
0.11517333984375,
0.73291015625,
0.88232421875,
0.73291015625,
0.88232421875,
0.7470703125,
0.11517333984375,
0.7470703125
],
[
0.11517333984375,
0.7470703125,
0.8837890625,
0.7470703125,
0.8837890625,
0.76123046875,
0.11517333984375,
0.76123046875
],
[
0.11358642578125,
0.76025390625,
0.88232421875,
0.76123046875,
0.88232421875,
0.7744140625,
0.11358642578125,
0.7734375
],
[
0.11517333984375,
0.7744140625,
0.88232421875,
0.7744140625,
0.88232421875,
0.78857421875,
0.11517333984375,
0.78857421875
],
[
0.11517333984375,
0.78857421875,
0.88232421875,
0.78857421875,
0.88232421875,
0.8017578125,
0.11517333984375,
0.8017578125
],
[
0.1168212890625,
0.8017578125,
0.88232421875,
0.8017578125,
0.88232421875,
0.81494140625,
0.1168212890625,
0.81494140625
],
[
0.1168212890625,
0.81640625,
0.88232421875,
0.81640625,
0.88232421875,
0.830078125,
0.1168212890625,
0.830078125
],
[
0.1168212890625,
0.830078125,
0.88232421875,
0.830078125,
0.88232421875,
0.84326171875,
0.1168212890625,
0.84326171875
],
[
0.11517333984375,
0.84326171875,
0.88232421875,
0.84326171875,
0.88232421875,
0.85693359375,
0.11517333984375,
0.85693359375
],
[
0.1168212890625,
0.859375,
0.88134765625,
0.859375,
0.88134765625,
0.869140625,
0.1168212890625,
0.869140625
],
[
0.11767578125,
0.87255859375,
0.88134765625,
0.87255859375,
0.88134765625,
0.88330078125,
0.11767578125,
0.88330078125
],
[
0.1168212890625,
0.88525390625,
0.88232421875,
0.88525390625,
0.88232421875,
0.8984375,
0.1168212890625,
0.8984375
],
[
0.11517333984375,
0.8984375,
0.8857421875,
0.8984375,
0.8857421875,
0.912109375,
0.11517333984375,
0.912109375
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Building artificial neural circuits for domain-general cognition",
"A",
"B",
"Multiple Demand",
"system",
"Dn",
"MultipleDemand",
"system",
"Figure l: A- The cortical areas forming the core Multiple Demand system in the human brain, from [34]. B- Schematic",
"depiction of a systems-level view of the brain. The Multiple Demand system lies at the core of information processing",
"in the brain, exchanging inputs with more specialised regions such as language, memory, sensory and social processing.",
"Due to its central position, the MD core can influence computations in multiple specialised areas by broadcasting",
"information it constructed from integrating across domains back to specialised regions, e.g., influencing perception by",
"abstract understanding of the environment / situation at large. Refer to [34] for detailed anatomical perspective of the",
"MD system's core and penumbra regions not discussed here.",
"3.",
"Computational motifs supporting domain-general cognition",
"3.1Globalrecurrence",
"Computations in functionally more specialised regions depend strongly on a feed-forward structure that extracts",
"increasingly abstract features from sensory inputs [39, 40, 41]. Much work shows how this process can be modelled",
"using an artificial feed-forward network [42, 43]. While there is also recurrent processing in these specialised systems",
"[39, 40, 44], the recurrent loops in these systems are likely very local and cover relatively short distances and timescales",
"Thismeansthata signal sentfromanodewill onlytravelashortpathbeforearrivingbackatits starting point.Aswe",
"move towards more integrative and domain-general cognition, recurrent connections become a hallmark feature of",
"the brain's systems-level design. The frontal cortex, where a large part of the MD system lies, is often thought of as",
"implementing recurrent loops for abstract information processing [41, 45]. Importantly, these loops not only process",
"information locally but also broadcast information widely across the brain, influencing and controlling computations in",
"specialised regions. It does so by not only having local recurrent connections within the circuit but also many loops",
"spanning large distances in the brain, reaching out to nodes which lie far outside the core [37, 46, 41, 36]. With nodes",
"widely distributed over the cortex, coupled to strong communication between these nodes, the MD system is well",
"positioned for widespread integration and communication. A large set of recursive processing loops with varied scales",
"in terms of time and spatial distance likely facilitate the MD system's abstract domain-general processing and deliver",
"the abilityto coordinate computation in a large distributed system [10].",
"The use of recurrent loops in artificial neural networks has a long history [47]. They proved to be useful tools for",
"processing and predicting time series data but also suffered from problems of vanishing gradient and computational",
"complexity when capturing long-range dependencies in the input [48, 49, 50]. To avoid these issues, feed-forward",
"based architectures can be used as substitutes [50] and various attention-based architectures have recently been very",
"effective in capturing dependencies in language time courses and multiple other modalities [49, 51]. This works by",
"inputting an entire time series in a single time step so that the attention mechanism learns the relationship between",
"timesteps without needing to hold past time points in memory. While these architectures likely can be good substitutes",
"for the local recurrent loops, we believe that ultimately, researchers are going to have to find a way to also introduce",
"global recurrent loops to arrive at domain-general cognition in artificial systems. Approaches like weight-sharing in",
"deep models paired with skip-connections may allow us to mimic a recurrent process in a regular forward pass but it",
"seemslikelythatalternativewayswillbeneededtoallowabstractmultimodalknowledgetobebroadcastedthrough",
"thenetworktoinformdistributedcomputations.Thisseemsevenmoretimelynowthatmodelsgenerateimpressive",
"responses to inputs such as images or language [1, 52], but struggle to be constrained by meaningful world models",
"(e.g., intuitive physics, [15]). Instead, researchers rely on human feedback signals in the training pipeline [53]. As such,",
"3"
] |
{
"arxiv_id": "2303.13651",
"language": "en",
"page_idx": 2
}
|
||
2303.13651-en-3
|
2303.13651-en-3.jpg
|
Building artificial neural circuits for domain-general cognition
machine learning models may need to be adapted to allow for the introduction of a global recurrent architecture similar
to the MD system.
3.2 Communicability in large scale networks
For any complex network which is concerned with processing information, it is of central importance to optimise how signals are communicated between the nodes within the network [54]. This becomes an increasingly challenging problem as a network grows, leaving nodes to only be able to communicate with a smaller proportion of the network This limited communication capacity naturally leads to variation in terms of how much information is exchanged between different pairs of nodes across the network. This results in a very real challenge for any large-scale network system to optimise its structure to integrate information most effectively and efficiently across its functional hubs. This is constrained, ultimately, by the topological arrangement of the network. The idea of how much information is exchanged between nodes is captured by the concept of communicability [54,55,56] and is a highly effective framework to understand how the structure of the brain guides function [57,58,59,60,61,62,63]. Specifically, across the brain's complex network, regions vary in terms of how well they can communicate to other regions, and the macro-scale dynamics and capabilities of the brain will be determined by this interareal communication. This heterogeneous communicability becomes especially interesting when one considers how system-level communication link to domain-general cognition. In the previous section, we described how more specialised regions tend to have a mostly feed-forward structure with some local recurrence. As such, information tends to be communicated locally between adjacent and functionally related regions. This changes as information approaches the domain-general MD system with its wider communicative influence. In its central position, the MD system not only receives information from all over the brain but utilises its widespread connectivity as global recurrent loops to broadcast processed information to a distributed set of brain regions [41, 10, 36]. On the systems-level perspective of the brain, a given region's communicative structure heavily depends on its functional role and hence its degree of specialisation
The concept of heterogeneous communicability between regions and modules of the brain has not been particularly relevant in artificial neural network architectures which were state-of-the-art until very recently. Take convolutional neural networks (CNNs) as an example. In CNNs, which dominated processing of visual information for several years [47], information is mostly passed along from layer to layer in a relatively even fashion. This means regions do not stick out has having a particular communicative ability (though see work like[64] for interesting communicative extensions of CNNs). However, this is changing with new architectures which have been growing in scale[65]. Especially for network models which utilise multiple modalities, architectures have increasingly been created by combining existing pre-trained models into more complex modular architectures [52,5]. Once we build complex system like these, it becomes increasingly important to not only think about which models to combine, but also how to combine them. This means that the communicability between parts of the network can be optimised to achieve better information flow between components and hence improve performance. A first step in this direction was made by a multimodal transformer model which outperformed prior networks by introducing a set of special bridge layers to connect two modality specific models. These bridge layers allow the model to learn a communicative structure in which abstract semantic knowledge is gradually merged across modalities. This increased performance in several relevant benchmark tasks[66]. In addition, other implementations have shown that bringing ideas from highly communicative small world graph structures into a Transformer's attention mechanism can help with processing longer sequences[67]. In simple recurrent neural networks, we also have seen that system-level communicability can easily be used as a regularisation term to optimise the communicative structure of a sparsely connected network to arrive at a network with many brain-like structural and functional properties [68]. As network models grow in complexity and increasingly make use of composite structures which combine sub models into larger networks, it will be important to fine tune the communicative structure of a network. Having good priors and inductive biases for these linkages can help circumvent problems arising from adding the extensive set of connections it would require to fully connect multiple models which already have a complex structure themselves. Following this line of thinking, we believe that making use of work on communicability and how it can be optimised in complex networks will be of central importance to inform model building on a systems-level.
# 3.3 Short-term topological changes
The discussion so far has focused on how the systems-level network structure of the brain and the unique communicative structure of its MD system play a key role in the domain-general cognition we see in humans. An important element of its flexible and multimodal information processing capabilities is how the MD system's network structure is not fully fixed but often rapidly changing. This means that while the MD system is running multimodal computations $\dot{\text{internally, the connections between its neurons are in continuous flux. As such, the general problem-solving ability of}}$ this network is assumed to be due to its inherent flexibility. In it, local computations are organised by rapid changes to the network structure[69,70,71], often called short-term plasticity. This allows the network to continuously reassign
4
|
[
[
0.1168212890625,
0.044189453125,
0.52783203125,
0.044189453125,
0.52783203125,
0.058074951171875,
0.1168212890625,
0.058074951171875
],
[
0.11517333984375,
0.0921630859375,
0.88232421875,
0.0921630859375,
0.88232421875,
0.1060791015625,
0.11517333984375,
0.1060791015625
],
[
0.11517333984375,
0.106689453125,
0.238525390625,
0.106689453125,
0.238525390625,
0.12060546875,
0.11517333984375,
0.12060546875
],
[
0.11358642578125,
0.1337890625,
0.438720703125,
0.1351318359375,
0.438720703125,
0.149658203125,
0.11358642578125,
0.1483154296875
],
[
0.1168212890625,
0.1610107421875,
0.88232421875,
0.1610107421875,
0.88232421875,
0.1749267578125,
0.1168212890625,
0.1749267578125
],
[
0.11358642578125,
0.173583984375,
0.88232421875,
0.175537109375,
0.88232421875,
0.1900634765625,
0.11358642578125,
0.1881103515625
],
[
0.11517333984375,
0.189453125,
0.88232421875,
0.189453125,
0.88232421875,
0.20263671875,
0.11517333984375,
0.20263671875
],
[
0.1168212890625,
0.20263671875,
0.88232421875,
0.20263671875,
0.88232421875,
0.216552734375,
0.1168212890625,
0.216552734375
],
[
0.11517333984375,
0.216552734375,
0.88232421875,
0.216552734375,
0.88232421875,
0.2298583984375,
0.11517333984375,
0.2298583984375
],
[
0.11517333984375,
0.2310791015625,
0.8837890625,
0.2310791015625,
0.8837890625,
0.2442626953125,
0.11517333984375,
0.2442626953125
],
[
0.1168212890625,
0.2442626953125,
0.8837890625,
0.2442626953125,
0.8837890625,
0.25830078125,
0.1168212890625,
0.25830078125
],
[
0.11517333984375,
0.25830078125,
0.88232421875,
0.25830078125,
0.88232421875,
0.271484375,
0.11517333984375,
0.271484375
],
[
0.11517333984375,
0.271484375,
0.88232421875,
0.271484375,
0.88232421875,
0.285400390625,
0.11517333984375,
0.285400390625
],
[
0.1168212890625,
0.285400390625,
0.88232421875,
0.285400390625,
0.88232421875,
0.298583984375,
0.1168212890625,
0.298583984375
],
[
0.1168212890625,
0.2998046875,
0.88232421875,
0.2998046875,
0.88232421875,
0.313232421875,
0.1168212890625,
0.313232421875
],
[
0.11517333984375,
0.313232421875,
0.88232421875,
0.313232421875,
0.88232421875,
0.326904296875,
0.11517333984375,
0.326904296875
],
[
0.11517333984375,
0.326904296875,
0.88232421875,
0.326904296875,
0.88232421875,
0.34033203125,
0.11517333984375,
0.34033203125
],
[
0.1168212890625,
0.34228515625,
0.88134765625,
0.34228515625,
0.88134765625,
0.352783203125,
0.1168212890625,
0.352783203125
],
[
0.1168212890625,
0.35595703125,
0.88134765625,
0.35595703125,
0.88134765625,
0.3662109375,
0.1168212890625,
0.3662109375
],
[
0.11517333984375,
0.36865234375,
0.8837890625,
0.36865234375,
0.8837890625,
0.3818359375,
0.11517333984375,
0.3818359375
],
[
0.11517333984375,
0.3818359375,
0.8837890625,
0.3818359375,
0.8837890625,
0.395751953125,
0.11517333984375,
0.395751953125
],
[
0.11517333984375,
0.396484375,
0.8837890625,
0.396484375,
0.8837890625,
0.410400390625,
0.11517333984375,
0.410400390625
],
[
0.1168212890625,
0.410400390625,
0.833984375,
0.410400390625,
0.833984375,
0.423583984375,
0.1168212890625,
0.423583984375
],
[
0.11517333984375,
0.4287109375,
0.88134765625,
0.4287109375,
0.88134765625,
0.443115234375,
0.11517333984375,
0.443115234375
],
[
0.11517333984375,
0.443115234375,
0.88134765625,
0.443115234375,
0.88134765625,
0.45703125,
0.11517333984375,
0.45703125
],
[
0.11517333984375,
0.458251953125,
0.8837890625,
0.458251953125,
0.8837890625,
0.4716796875,
0.11517333984375,
0.4716796875
],
[
0.11517333984375,
0.4716796875,
0.8837890625,
0.4716796875,
0.8837890625,
0.48486328125,
0.11517333984375,
0.48486328125
],
[
0.1168212890625,
0.4873046875,
0.88134765625,
0.4873046875,
0.88134765625,
0.49755859375,
0.1168212890625,
0.49755859375
],
[
0.11517333984375,
0.498779296875,
0.88232421875,
0.498779296875,
0.88232421875,
0.51220703125,
0.11517333984375,
0.51220703125
],
[
0.11517333984375,
0.51318359375,
0.88232421875,
0.51318359375,
0.88232421875,
0.5263671875,
0.11517333984375,
0.5263671875
],
[
0.1168212890625,
0.5263671875,
0.88232421875,
0.5263671875,
0.88232421875,
0.54052734375,
0.1168212890625,
0.54052734375
],
[
0.11517333984375,
0.54052734375,
0.8837890625,
0.54052734375,
0.8837890625,
0.5537109375,
0.11517333984375,
0.5537109375
],
[
0.11767578125,
0.55615234375,
0.88134765625,
0.55615234375,
0.88134765625,
0.56640625,
0.11767578125,
0.56640625
],
[
0.1168212890625,
0.568359375,
0.88232421875,
0.568359375,
0.88232421875,
0.58203125,
0.1168212890625,
0.58203125
],
[
0.11517333984375,
0.58203125,
0.88232421875,
0.58203125,
0.88232421875,
0.59521484375,
0.11517333984375,
0.59521484375
],
[
0.11358642578125,
0.59521484375,
0.88232421875,
0.5947265625,
0.88232421875,
0.60791015625,
0.11358642578125,
0.609375
],
[
0.11517333984375,
0.609375,
0.88232421875,
0.609375,
0.88232421875,
0.62255859375,
0.11517333984375,
0.62255859375
],
[
0.11517333984375,
0.62255859375,
0.88134765625,
0.62255859375,
0.88134765625,
0.63623046875,
0.11517333984375,
0.63623046875
],
[
0.11517333984375,
0.63720703125,
0.8837890625,
0.63720703125,
0.8837890625,
0.65087890625,
0.11517333984375,
0.65087890625
],
[
0.11517333984375,
0.65087890625,
0.8837890625,
0.65087890625,
0.8837890625,
0.6640625,
0.11517333984375,
0.6640625
],
[
0.11517333984375,
0.6640625,
0.88232421875,
0.6640625,
0.88232421875,
0.67822265625,
0.11517333984375,
0.67822265625
],
[
0.11517333984375,
0.6796875,
0.88134765625,
0.6796875,
0.88134765625,
0.68994140625,
0.11517333984375,
0.68994140625
],
[
0.11358642578125,
0.6923828125,
0.88232421875,
0.6923828125,
0.88232421875,
0.70556640625,
0.11358642578125,
0.70556640625
],
[
0.11517333984375,
0.70556640625,
0.88232421875,
0.70556640625,
0.88232421875,
0.7197265625,
0.11517333984375,
0.7197265625
],
[
0.1168212890625,
0.7216796875,
0.88134765625,
0.7216796875,
0.88134765625,
0.73193359375,
0.1168212890625,
0.73193359375
],
[
0.11358642578125,
0.73193359375,
0.8837890625,
0.73291015625,
0.8837890625,
0.74755859375,
0.11358642578125,
0.7470703125
],
[
0.11517333984375,
0.74755859375,
0.88232421875,
0.74755859375,
0.88232421875,
0.76123046875,
0.11517333984375,
0.76123046875
],
[
0.11517333984375,
0.76123046875,
0.298095703125,
0.76123046875,
0.298095703125,
0.7744140625,
0.11517333984375,
0.7744140625
],
[
0.1168212890625,
0.79052734375,
0.1446533203125,
0.79052734375,
0.1446533203125,
0.8017578125,
0.1168212890625,
0.8017578125
],
[
0.1478271484375,
0.79052734375,
0.371826171875,
0.79052734375,
0.371826171875,
0.8037109375,
0.1478271484375,
0.8037109375
],
[
0.11517333984375,
0.81494140625,
0.88232421875,
0.81494140625,
0.88232421875,
0.8291015625,
0.11517333984375,
0.8291015625
],
[
0.11517333984375,
0.830078125,
0.88134765625,
0.830078125,
0.88134765625,
0.84326171875,
0.11517333984375,
0.84326171875
],
[
0.11517333984375,
0.84326171875,
0.88232421875,
0.84326171875,
0.88232421875,
0.85693359375,
0.11517333984375,
0.85693359375
],
[
0.1168212890625,
0.85791015625,
0.88232421875,
0.85791015625,
0.88232421875,
0.8720703125,
0.1168212890625,
0.8720703125
],
[
0.1168212890625,
0.87255859375,
0.88232421875,
0.87255859375,
0.88232421875,
0.88330078125,
0.1168212890625,
0.88330078125
],
[
0.11517333984375,
0.88525390625,
0.88232421875,
0.88525390625,
0.88232421875,
0.8984375,
0.11517333984375,
0.8984375
],
[
0.11517333984375,
0.8984375,
0.88232421875,
0.8984375,
0.88232421875,
0.912109375,
0.11517333984375,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"Building artificial neural circuits fordomain-general cognition",
"machine learning models may need to be adapted to allow for the introduction of a global recurrent architecture similar",
"to the MD system.",
"3.2Communicability in large scale networks",
"For any complex network which is concerned with processing information, it is of central importance to optimise",
"how signals are communicated between the nodes within the network [54]. This becomes an increasingly challenging",
"problem as a network grows, leaving nodes to only be able to communicate with a smaller proportion of the network",
"This limited communication capacity naturally leads to variation in terms of how much information is exchanged",
"between different pairs of nodes across the network. This results in a very real challenge for any large-scale network",
"system to optimise its structure to integrate information most effectively and efficiently across its functional hubs.",
"This is constrained, ultimately, by the topological arrangement of the network. The idea of how much information",
"is exchanged between nodes is captured by the concept of communicability [54, 55, 56] and is a highly effective",
"framework to understand how the structure of the brain guides function [57, 58, 59, 60, 61, 62, 63]. Specifically",
"across the brain's complex network, regions vary in terms of how well they can communicate to other regions, and",
"the macro-scale dynamics and capabilities of the brain will be determined by this interareal communication. This",
"heterogeneous communicability becomes especially interesting when one considers how system-level communication",
"link to domain-general cognition. In the previous section, we described how more specialised regions tend to have",
"amostlyfeed-forward structurewithsomelocalrecurrence.Assuch,informationtendstobecommunicatedlocally",
"betweenadjacentandfunctionallyrelatedregions.Thischangesasinformationapproachesthedomain-generalMD",
"system with its wider communicative influence. In its central position, the MD system not only receives information",
"from all over the brain but utilises its widespread connectivity as global recurrent loops to broadcast processed",
"information to a distributed set of brain regions [41, 10, 36]. On the systems-level perspective of the brain, a given",
"region's communicative structure heavily depends on its functional role and hence its degree of specialisation.",
"The concept of heterogeneous communicability between regions and modules of the brain has not been particularly",
"relevant in artificial neural network architectures which were state-of-the-art until very recently. Take convolutional",
"neural networks (CNNs) as an example. In CNNs, which dominated processing of visual information for several years",
"[47], information is mostly passed along from layer to layer in a relatively even fashion. This means regions do not stick",
"outhashavingaparticularcommunicativeability(thoughseeworklike[64lforinterestingcommunicativeextensions",
"of CNNs). However, this is changing with new architectures which have been growing in scale [65]. Especially for",
"network models which utilise multiple modalities, architectures have increasingly been created by combining existing",
"pre-trained models into more complex modular architectures [52, 5]. Once we build complex system like these, it",
"becomes increasingly important to not only think about which models to combine, but also how to combine them.",
"Thismeansthatthecommunicabilitybetweenparts ofthenetworkcanbeoptimisedto achievebetterinformation",
"flow between components and hence improve performance. A first step in this direction was made by a multimodal",
"transformer model which outperformed prior networks by introducing a set of special bridge layers to connect two",
"modality specific models. These bridge layers allow the model to learn a communicative structure in which abstract",
"semantic knowledge is gradually merged across modalities. This increased performance in several relevant benchmark",
"tasks [66]. In addition, other implementations have shown that bringing ideas from highly communicative small",
"world graph structures into a Transformer's attention mechanism can help with processing longer sequences [67].",
"In simple recurrent neural networks, we also have seen that system-level communicability can easily be used as a",
"regularisation term to optimise the communicative structure of a sparsely connected network to arrive at a network",
"withmanybrain-likestructuralandfunctionalproperties[68l.Asnetworkmodelsgrowincomplexityandincreasingly",
"make use of composite structures which combine sub models into larger networks, it will be important to fine tune the",
"communicative structure of a network. Having good priors and inductive biases for these linkages can help circumvent",
"problemsarisingfromaddingtheextensivesetofconnectionsitwouldrequiretofullyconnectmultiplemodelswhich",
"already have a complex structure themselves. Following this line of thinking, we believe that making use of work on",
"communicability and how it can be optimised in complex networks will be of central importance to inform model",
"building on a systems-level.",
"3.3",
"Short-term topological changes",
"The discussion so far has focused on how the systems-level network structure of the brain and the unique communicative",
"structure of its MD system play a key role in the domain-general cognition we see in humans. An important element",
"of its flexible and multimodal information processing capabilities is how the MD system's network structure is not",
"fully fixed but often rapidly changing. This means that while the MD system is running multimodal computations",
"internally,theconnectionsbetweenits neurons areincontinuous flux.As such,thegeneralproblem-solving abilityof",
"this network is assumed to be due to its inherent flexibility. In it, local computations are organised by rapid changes to",
"the network structure [69, 70, 71], often called short-term plasticity. This allows the network to continuously reassign",
"4"
] |
{
"arxiv_id": "2303.13651",
"language": "en",
"page_idx": 3
}
|
||
2303.13651-en-4
|
2303.13651-en-4.jpg
|
Building artificial neural circuits for domain-general cognition
its neurons and modules new computational roles while solving complex sequential problems [72,37,73,74,75]. Rapid topological changes are likely induced by local learning rules which supplement the more long-term optimisation of the global network structure. These mechanisms likely underly a multitude of complex abilities of the human brain [34,76] and some of them strongly overlap with timely discussions in machine learning. As one example, research points to the fact that the MD system uses its short-term dynamics for attentional control, to focus on information which is relevant for the current operation [77,78,79] and break complex pieces of information down into simple computable bits[10]- a function which has played a central role in machine learning discussion recently [64,80]. Another example is the MD system's ability to construct abstract representations of problems [81] to then tie observed stimuli rapidly to their roles in this abstract problem representation [$\hat{1}0,75],\tilde{\text{a phenomenon going by the name of variable binding [82] or}}$ meta-learning[83]. These are very related to few-shot learning[1] and in-context learning abilities[84] observed in large Transformer models.
As we already see foundations of these skills emerging in currently existing architectures it is reasonable to believe that they will continue to improve purely by scaling existing architectures [65]. In this case models would use their unit activations to implement rapid in-context learning and it has been shown that this can work well without any short-term synaptic changes [85]. In fact, even in the brain many complex computations are likely facilitated due to dynamics of the network activations which do not necessarily have to rely on changes in the network structure[86]. But once computations reach the scale of using network-wide attention processes to controlling the flow of information across the entire brain network and flexibly combining task modules to solve the task at hand [10,87,88],rapid topological network changes might be necessary for domain-general computations [69,10]. Reaching this level of flexible and multimodal cognition might not be possible in current static architectures and hence might require us to allow models to modify some of their connections in the moment through local learning rules. Some work in smaller network models is highlighting how local learning mechanisms can complement network-wide optimisation processes [89,90] with relevant comparisons to Transformer implementations [91]. Other examples point to how local learning rules and single neuron-based optimisation principles by themselves can be sufficient to solve meaningful cognitive tasks [92, 93]. In addition, we have seen how standard network optimisers can be updated with certainty judgements to support rapid relational learning [94]. If we could scale these rapid learning dynamics to large Transformer models, this might allow models to flexibly combine abstract task structures with capabilities learned in the past, to flexibly apply skills across modalities in a truly domain-general way. One research direction which might support rapid learning processes is work on using local loss functions and learning mechanisms to substitute costly global optimisation processes [95,96,97] Combining these local optimisation processes with more wide-spread recurrent loops and an optimised communicative structure in large networks might bring us closer to observing flexible domain-general cognition in artificial neural networks.
# 4 Conclusion
We believe that in the pursuit of building artificial intelligence which is able to engage in domain-general problem solving, a systems-level view of the human brain will provide useful guidance [98, 99]. We believe this will become increasingly relevant as AI systems become more and more complex. The topics of recurrence, communication and rapid structural changes are particularly relevant at the current point due to their central role in theories of domain-general cognition in the brain and their links to existing works in neural network models. As such, they might be key drivers behind efficient and flexible information processing in large multimodal networks. But we do not believe that any of these features should be fully hard-coded- instead we should think of them as useful priors and inductive biases which can guide complex learning processes. Ultimately, bringing these features into machine learning models opens up the perspective of not only improving the performance of artificial neural networks but also for us to understand which core principles underly domain-general and multimodal computations in neural networks - may these be biological or artificial.
# 5 Acknowledgments
J.A., Da.A., M.A., Du.A., and J.D. are supported by UKRI MRC funding and as a result the authors have applied a Creative Commons Attribution (CC BY) license to this manuscript for the purpose of open access. J.A. receives a Gates Cambridge Scholarship. Da.A. receives a Cambridge Trust Vice Chancellor's Scholarship. Da.A. and Du.A. are both supported by the James S. McDonnell Foundation Õpportunity Award. J.A. was a research intern at Intel Labs at the time of writing this manuscript.
5
|
[
[
0.1168212890625,
0.044189453125,
0.52783203125,
0.044189453125,
0.52783203125,
0.058074951171875,
0.1168212890625,
0.058074951171875
],
[
0.11517333984375,
0.0921630859375,
0.88232421875,
0.091552734375,
0.88232421875,
0.1060791015625,
0.11517333984375,
0.106689453125
],
[
0.11517333984375,
0.106689453125,
0.8837890625,
0.106689453125,
0.8837890625,
0.12060546875,
0.11517333984375,
0.12060546875
],
[
0.1168212890625,
0.12249755859375,
0.88134765625,
0.12249755859375,
0.88134765625,
0.1331787109375,
0.1168212890625,
0.1331787109375
],
[
0.1168212890625,
0.1337890625,
0.88232421875,
0.1337890625,
0.88232421875,
0.147705078125,
0.1168212890625,
0.147705078125
],
[
0.1168212890625,
0.149658203125,
0.88134765625,
0.149658203125,
0.88134765625,
0.160400390625,
0.1168212890625,
0.160400390625
],
[
0.11517333984375,
0.1622314453125,
0.88232421875,
0.1622314453125,
0.88232421875,
0.175537109375,
0.11517333984375,
0.175537109375
],
[
0.11517333984375,
0.175537109375,
0.88232421875,
0.175537109375,
0.88232421875,
0.189453125,
0.11517333984375,
0.189453125
],
[
0.11517333984375,
0.189453125,
0.8837890625,
0.189453125,
0.8837890625,
0.20263671875,
0.11517333984375,
0.20263671875
],
[
0.1168212890625,
0.20263671875,
0.8837890625,
0.20263671875,
0.8837890625,
0.216552734375,
0.1168212890625,
0.216552734375
],
[
0.11517333984375,
0.2171630859375,
0.8837890625,
0.2171630859375,
0.8837890625,
0.2310791015625,
0.11517333984375,
0.2310791015625
],
[
0.11517333984375,
0.2310791015625,
0.2900390625,
0.2310791015625,
0.2900390625,
0.2442626953125,
0.11517333984375,
0.2442626953125
],
[
0.11517333984375,
0.251953125,
0.88232421875,
0.251953125,
0.88232421875,
0.26513671875,
0.11517333984375,
0.26513671875
],
[
0.1168212890625,
0.267578125,
0.88134765625,
0.267578125,
0.88134765625,
0.27783203125,
0.1168212890625,
0.27783203125
],
[
0.1168212890625,
0.281005859375,
0.88134765625,
0.281005859375,
0.88134765625,
0.291748046875,
0.1168212890625,
0.291748046875
],
[
0.11517333984375,
0.292236328125,
0.8837890625,
0.292236328125,
0.8837890625,
0.30615234375,
0.11517333984375,
0.30615234375
],
[
0.1168212890625,
0.30810546875,
0.88232421875,
0.30810546875,
0.88232421875,
0.318115234375,
0.1168212890625,
0.318115234375
],
[
0.11517333984375,
0.32080078125,
0.88232421875,
0.32080078125,
0.88232421875,
0.333984375,
0.11517333984375,
0.333984375
],
[
0.11517333984375,
0.333984375,
0.88232421875,
0.333984375,
0.88232421875,
0.347900390625,
0.11517333984375,
0.347900390625
],
[
0.11517333984375,
0.348388671875,
0.8837890625,
0.348388671875,
0.8837890625,
0.3623046875,
0.11517333984375,
0.3623046875
],
[
0.11767578125,
0.363037109375,
0.88134765625,
0.363037109375,
0.88134765625,
0.373779296875,
0.11767578125,
0.373779296875
],
[
0.11517333984375,
0.375732421875,
0.88232421875,
0.375732421875,
0.88232421875,
0.389404296875,
0.11517333984375,
0.389404296875
],
[
0.11517333984375,
0.389404296875,
0.88232421875,
0.389404296875,
0.88232421875,
0.40283203125,
0.11517333984375,
0.40283203125
],
[
0.11517333984375,
0.40283203125,
0.88232421875,
0.40283203125,
0.88232421875,
0.416748046875,
0.11517333984375,
0.416748046875
],
[
0.11517333984375,
0.417236328125,
0.88232421875,
0.417236328125,
0.88232421875,
0.43115234375,
0.11517333984375,
0.43115234375
],
[
0.11517333984375,
0.43115234375,
0.88232421875,
0.43115234375,
0.88232421875,
0.4443359375,
0.11517333984375,
0.4443359375
],
[
0.11517333984375,
0.4443359375,
0.8837890625,
0.4443359375,
0.8837890625,
0.458251953125,
0.11517333984375,
0.458251953125
],
[
0.11517333984375,
0.458984375,
0.8837890625,
0.458984375,
0.8837890625,
0.472900390625,
0.11517333984375,
0.472900390625
],
[
0.11358642578125,
0.47021484375,
0.88232421875,
0.4716796875,
0.88232421875,
0.486083984375,
0.11358642578125,
0.48486328125
],
[
0.11517333984375,
0.486083984375,
0.8837890625,
0.486083984375,
0.8837890625,
0.5,
0.11517333984375,
0.5
],
[
0.11517333984375,
0.5,
0.88232421875,
0.5,
0.88232421875,
0.51318359375,
0.11517333984375,
0.51318359375
],
[
0.11517333984375,
0.51318359375,
0.88232421875,
0.51318359375,
0.88232421875,
0.5263671875,
0.11517333984375,
0.5263671875
],
[
0.11517333984375,
0.52783203125,
0.181396484375,
0.52783203125,
0.181396484375,
0.5390625,
0.11517333984375,
0.5390625
],
[
0.1168212890625,
0.57763671875,
0.1298828125,
0.57763671875,
0.1298828125,
0.58837890625,
0.1168212890625,
0.58837890625
],
[
0.1446533203125,
0.57568359375,
0.2418212890625,
0.57568359375,
0.2418212890625,
0.59033203125,
0.1446533203125,
0.59033203125
],
[
0.11358642578125,
0.6142578125,
0.88232421875,
0.61572265625,
0.88232421875,
0.6298828125,
0.11358642578125,
0.62890625
],
[
0.11517333984375,
0.6298828125,
0.88232421875,
0.6298828125,
0.88232421875,
0.64453125,
0.11517333984375,
0.64453125
],
[
0.11517333984375,
0.64453125,
0.88232421875,
0.64453125,
0.88232421875,
0.65771484375,
0.11517333984375,
0.65771484375
],
[
0.11517333984375,
0.65771484375,
0.88232421875,
0.65771484375,
0.88232421875,
0.671875,
0.11517333984375,
0.671875
],
[
0.11517333984375,
0.67041015625,
0.8798828125,
0.67041015625,
0.8798828125,
0.6845703125,
0.11517333984375,
0.6845703125
],
[
0.11517333984375,
0.68505859375,
0.8837890625,
0.68505859375,
0.8837890625,
0.69873046875,
0.11517333984375,
0.69873046875
],
[
0.1168212890625,
0.69873046875,
0.88134765625,
0.69873046875,
0.88134765625,
0.7119140625,
0.1168212890625,
0.7119140625
],
[
0.11517333984375,
0.71337890625,
0.88232421875,
0.71337890625,
0.88232421875,
0.7265625,
0.11517333984375,
0.7265625
],
[
0.11517333984375,
0.7265625,
0.88232421875,
0.7265625,
0.88232421875,
0.74072265625,
0.11517333984375,
0.74072265625
],
[
0.11517333984375,
0.74072265625,
0.88232421875,
0.74072265625,
0.88232421875,
0.75390625,
0.11517333984375,
0.75390625
],
[
0.1168212890625,
0.755859375,
0.176513671875,
0.755859375,
0.176513671875,
0.7666015625,
0.1168212890625,
0.7666015625
],
[
0.11767578125,
0.8046875,
0.13232421875,
0.8046875,
0.13232421875,
0.81494140625,
0.11767578125,
0.81494140625
],
[
0.131591796875,
0.806640625,
0.1405029296875,
0.806640625,
0.1405029296875,
0.8125,
0.131591796875,
0.8125
],
[
0.136474609375,
0.8037109375,
0.2998046875,
0.8037109375,
0.2998046875,
0.818359375,
0.136474609375,
0.818359375
],
[
0.11517333984375,
0.84228515625,
0.88232421875,
0.84228515625,
0.88232421875,
0.85693359375,
0.11517333984375,
0.85693359375
],
[
0.11517333984375,
0.85791015625,
0.88232421875,
0.85791015625,
0.88232421875,
0.8720703125,
0.11517333984375,
0.8720703125
],
[
0.1168212890625,
0.87060546875,
0.88232421875,
0.87060546875,
0.88232421875,
0.8837890625,
0.1168212890625,
0.8837890625
],
[
0.11517333984375,
0.8837890625,
0.88232421875,
0.8837890625,
0.88232421875,
0.8984375,
0.11517333984375,
0.8984375
],
[
0.11517333984375,
0.90087890625,
0.3203125,
0.90087890625,
0.3203125,
0.9111328125,
0.11517333984375,
0.9111328125
],
[
0.494384765625,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.95068359375,
0.494384765625,
0.95068359375
]
] |
[
"Building artificial neural circuits for domain-general cognition",
"its neurons and modules new computational roles while solving complex sequential problems [72, 37, 73, 74, 75]",
"Rapid topological changes are likely induced by local learning rules which supplement the more long-term optimisation",
"oftheglobal network structure.Thesemechanismslikelyunderlyamultitudeofcomplexabilities of thehumanbrain",
"[34, 76] and some of them strongly overlap with timely discussions in machine learning. As one example, research",
"pointstothefactthattheMDsystemuses its short-termdynamicsforattentional control,tofocus on informationwhich",
"is relevant for the current operation [77, 78, 79] and break complex pieces of information down into simple computable",
"bits [10] - a function which has played a central role in machine learning discussion recently [64, 80]. Another example",
"is the MD system's ability to construct abstract representations of problems [81] to then tie observed stimuli rapidly to",
"their roles in this abstract problem representation [10, 75], a phenomenon going by the name of variable binding [82] or",
"meta-learning [83]. These are very related to few-shot learning [1] and in-context learning abilities [84] observed in",
"largeTransformermodels.",
"As we already see foundations of these skills emerging in currently existing architectures it is reasonable to believe that",
"theywillcontinuetoimprovepurelybyscalingexistingarchitectures[65.Inthiscasemodelswouldusetheirunit",
"activationstoimplementrapidin-contextlearningandithasbeen shownthatthiscanworkwellwithoutanyshort-term",
"synaptic changes [85]. In fact, even in the brain many complex computations are likely facilitated due to dynamics",
"ofthenetworkactivationswhichdonotnecessarilyhavetorelyonchangesinthenetworkstructure[86l.Butonce",
"computations reach the scale of using network-wide attention processes to controlling the flow of information across",
"the entire brain network and flexibly combining task modules to solve the task at hand [10, 87, 88], rapid topological",
"network changes might be necessary for domain-general computations [69, 10]. Reaching this level of flexible and",
"multimodal cognitionmight notbepossiblein current static architectures andhencemightrequireus toallow modelsto",
"modify some of their connections in the moment through local learning rules. Some work in smaller network models",
"is highlighting how local learning mechanisms can complement network-wide optimisation processes [89,90] with",
"relevant comparisons to Transformer implementations [91]. Other examples point to how local learning rules and single",
"neuron-based optimisation principles by themselves can be sufficient to solve meaningful cognitive tasks [92, 93]. In",
"addition, we have seen how standard network optimisers can be updated with certainty judgements to support rapid",
"relational learning [94]. If we could scale these rapid learning dynamics to large Transformer models, this might allow",
"models to flexibly combine abstract task structures with capabilities learned in the past, to flexibly apply skills across",
"modalities in a truly domain-general way. One research direction which might support rapid learning processes is work",
"on using local loss functions and learning mechanisms to substitute costly global optimisation processes [95, 96, 97].",
"Combining these local optimisation processes with more wide-spread recurrent loops and an optimised communicative",
"structure in large networks might bring us closer to observing flexible domain-general cognition in artificial neural",
"networks.",
"4",
"Conclusion",
"We believe that in the pursuit of building artificial intelligence which is able to engage in domain-general problem",
"solving, a systems-level view of the human brain will provide useful guidance [98, 99]. We believe this will become",
"increasingly relevant as AI systems become more and more complex. The topics of recurrence, communication and rapid",
"structural changes are particularly relevant at the current point due to their central role in theories of domain-general",
"cognition in the brain and their links to existing works in neural network models. As such, they might be key drivers",
"behind efficient and flexible information processing in large multimodal networks. But we do not believe that any of",
"these features should be fully hard-coded - instead we should think of them as useful priors and inductive biases which",
"can guide complex learning processes. Ultimately, bringing these features into machine learning models opens up the",
"perspective of not only improving the performance of artificial neural networks but also for us to understand which",
"core principles underly domain-general and multimodal computations in neural networks - may these be biological or",
"artificial.",
"5.",
"",
"Acknowledgments",
"J.A., Da.A., M.A., Du.A., and J.D. are supported by UKRI MRC funding and as a result the authors have applied a",
"Creative Commons Attribution (CC BY) license to this manuscript for the purpose of open access. J.A. receives a Gates",
"Cambridge Scholarship. Da.A. receives a Cambridge Trust Vice Chancellor's Scholarship. Da.A. and Du.A. are both",
"supported by the James S. McDonnell Foundation Opportunity Award. J.A. was a research intern at Intel Labs at the",
"timeofwritingthismanuscript",
"5"
] |
{
"arxiv_id": "2303.13651",
"language": "en",
"page_idx": 4
}
|
||
2104.14730-en-0
|
2104.14730-en-0.jpg
|
Perceptual Image Quality Assessment with Transformers
Manri Cheon, Sung-Jun Yoon, Byungyeon Kang, and Junwoo Lee
LG Electronics
$\{$manri.cheon, sungjun.yoon, byungyeon.kang, junwoo.lee}{0}\:l\:g\:e.com}(1)\:,\:s\:ung\:jun.yoon,\:byungyeon.kang,\:junwoo.lee}}}}}}l\:g\:e.com}}}}}}}}}}}manri.cheon,\:sungyeon.kang,\:kangyeon.kang,\:jungyeon.kang,\:junwoo.lee}}}}}}}}}}}l\:ge.com,\:p\:s\:an\:,\:s\:p\:,\:p\:,\:p
$n$
$ন্ল$
## Abstract
In this paper, we propose an image quality transformer
$\sum_{imaaaaaaaaaualinall-reference\:image\:quality\:assesment\:(IQA)}^{(IQT)\:that\:successfully\:applies\:a\:transformer\:architecture\:to\\\sum_{imaaaaaabity\:accormaat\:ln\:thir~contation\:becomes\:more\:important\:in}^{(IQT)\:that\:successfully\:applies\:atrans\:the\:the\:thir\:co}}$
image quality assessment. In this context, we extract the perceptual feature representations from each of input images using a convolutional neural network (CNN) back-
ages anng a comolinona melral nelwork- ( 1. NN) back- ages with unrealistic artifacts. The existing objective met- bone. The extracted feature maps are fed into the rans- rics such as Peak Signal- to- Noise Ratio( PSNR) , astrucutural former encoder
$transformer-based$ vision models [18,55],we use extra learnable quality embedding and position embedding. The output of the transformer is passed to a prediction head in order to predict a final quality score. The experimental results show that our proposed model has an outstanding performance for the standard IQA datasets. For a large-scale IQA dataset containing output images of generative model, our model also shows the promising results. The proposed $IQTwas\textit{ranked first among l3 participants in the NTIRE}$ 2021 perceptual image quality assessment challenge [23]. Our work will be an opportunity to further expand the approach for the perceptual IQA task.
### $1. \textbf{Introduction}$
Perceptual image quality assessment (IQA) is an important topic in the multimedia systems and computer vision tasks[11,42,56]. One of the goals of the image processing is to improve the quality of the content to an acceptable level for the human viewers. In this context, the first step toward generating acceptable contents is to accurately measure the perceptual quality of the content, which can be performed via subjective and objective quality assess- $\hat{\text{ment [49,9,25,19]. The subjective quality assessment is}}$ the most accurate method to measure the perceived quality, which is usually represented by mean opinion scores (MOS) from collected subjective ratings.However, it is time-consuming and expensive. Thus, objective quality assessment performed by objective metrics is widely used to
automatically predict perceived quality [51,52,40,59,53].
However, with the recent advances in deep learningbased image restoration algorithms, accurate prediction of the perceived quality has become more difficult. In particular, image restoration models based on generative adversarial network (GAN)[20] have been developed in order to improve the perceptual aspect of the result images [48,2,12,8]. However, it sometimes generates output images with unrealistic artifacts. The existing objective metrics are insufficient to predict the quality of this kind of outputs. In this respect, recent works [61, 16, 21, 39, 3] based on perceptual representation exhibit a better performance at the perceptual IQA task. As various image restoration algorithms are developed, however, it is still required to develop the IQA algorithm that accurately predicts the perceptual quality of images generated by emerging algorithms.
In recent years, based on the success in the natural language processing (NLP) field, the transformer [46] architecture has been applied in the computer vision field[27]. A wider research area in the computer vision has been improved based on the transformer, such as recognition task [4,45,18], generative modelling [37,26,7], low-level vision [6,54,30], etc. However, few attempts were made in the feld of the image and video quality assessment. In a recent study, You and Korhonen proposed the application of transformer in image quality assessment[55]. They achieved outstanding performance on two publicly available large-scale blind image quality databases. With our knowledge, however, this study is the only transformer-based approach for image quality assessment. Therefore, it is urgently needed to investigate whether the transformer-based approach works well in the field of perceptual image quality assessment. Especially, it should be investigated whether this structure is applicable to a full-reference (FR) model aiming to measure the perceptual similarity between two images. In addition, it is also necessary to evaluate whether this approach can accurately predict the perceptual quality for the latest GAN-based artifacts.
In this study, we propose an Image Quality Transformer
|
[
[
0.200927734375,
0.1331787109375,
0.76953125,
0.1331787109375,
0.76953125,
0.15087890625,
0.200927734375,
0.15087890625
],
[
0.226318359375,
0.18310546875,
0.744140625,
0.18310546875,
0.744140625,
0.1976318359375,
0.226318359375,
0.1976318359375
],
[
0.423095703125,
0.199462890625,
0.54736328125,
0.2008056640625,
0.54736328125,
0.21533203125,
0.423095703125,
0.2139892578125
],
[
0.20751953125,
0.2197265625,
0.7587890625,
0.2197265625,
0.7587890625,
0.2342529296875,
0.20751953125,
0.2342529296875
],
[
0.0196075439453125,
0.26513671875,
0.063720703125,
0.26513671875,
0.066162109375,
0.37890625,
0.0228729248046875,
0.37890625
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.50341796875,
0.270751953125,
0.8896484375,
0.270751953125,
0.8896484375,
0.285400390625,
0.50341796875,
0.285400390625
],
[
0.521484375,
0.28857421875,
0.8896484375,
0.28857421875,
0.8896484375,
0.302978515625,
0.521484375,
0.302978515625
],
[
0.098876953125,
0.3017578125,
0.4697265625,
0.3017578125,
0.4697265625,
0.31494140625,
0.098876953125,
0.31494140625
],
[
0.5009765625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.5009765625,
0.318115234375
],
[
0.08087158203125,
0.31640625,
0.46728515625,
0.31640625,
0.46728515625,
0.330078125,
0.08087158203125,
0.330078125
],
[
0.50244140625,
0.3193359375,
0.8896484375,
0.3193359375,
0.8896484375,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.080078125,
0.33203125,
0.468017578125,
0.330810546875,
0.468017578125,
0.34521484375,
0.080078125,
0.3466796875
],
[
0.50244140625,
0.335205078125,
0.8896484375,
0.335205078125,
0.8896484375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.347900390625,
0.4697265625,
0.347900390625,
0.4697265625,
0.361083984375,
0.080078125,
0.361083984375
],
[
0.50244140625,
0.35107421875,
0.890625,
0.35107421875,
0.890625,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.363037109375,
0.4697265625,
0.363037109375,
0.4697265625,
0.376953125,
0.080078125,
0.376953125
],
[
0.0269622802734375,
0.369384765625,
0.0555419921875,
0.369384765625,
0.0555419921875,
0.38818359375,
0.0269622802734375,
0.38818359375
],
[
0.50341796875,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.50341796875,
0.380126953125
],
[
0.08087158203125,
0.380126953125,
0.46728515625,
0.380126953125,
0.46728515625,
0.39013671875,
0.08087158203125,
0.39013671875
],
[
0.50341796875,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.39453125,
0.50341796875,
0.39453125
],
[
0.080078125,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.50244140625,
0.396484375,
0.890625,
0.396484375,
0.890625,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.080078125,
0.40771484375,
0.468017578125,
0.40771484375,
0.468017578125,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.4462890625,
0.02532958984375,
0.4462890625
],
[
0.50244140625,
0.410888671875,
0.89208984375,
0.410888671875,
0.89208984375,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.07843017578125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.07843017578125,
0.436279296875
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.03350830078125,
0.44140625,
0.0555419921875,
0.44140625,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50244140625,
0.455078125
],
[
0.080078125,
0.45263671875,
0.4697265625,
0.45263671875,
0.4697265625,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455078125,
0.89208984375,
0.4697265625,
0.50244140625,
0.47021484375
],
[
0.08087158203125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.02532958984375,
0.486083984375,
0.05963134765625,
0.486083984375,
0.056365966796875,
0.5771484375,
0.0228729248046875,
0.5771484375
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50341796875,
0.486083984375,
0.890625,
0.486083984375,
0.890625,
0.5,
0.50341796875,
0.5
],
[
0.080078125,
0.498779296875,
0.46728515625,
0.498779296875,
0.46728515625,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.080078125,
0.51318359375,
0.46728515625,
0.51318359375,
0.46728515625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.52880859375,
0.468017578125,
0.52880859375,
0.468017578125,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.50341796875,
0.5322265625,
0.8486328125,
0.5322265625,
0.8486328125,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.080078125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.521484375,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.521484375,
0.564453125
],
[
0.08087158203125,
0.5595703125,
0.4697265625,
0.5595703125,
0.4697265625,
0.57275390625,
0.08087158203125,
0.57275390625
],
[
0.50341796875,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.580078125,
0.50341796875,
0.580078125
],
[
0.08087158203125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.0196075439453125,
0.58056640625,
0.058013916015625,
0.58056640625,
0.060455322265625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.5009765625,
0.58056640625,
0.890625,
0.580078125,
0.890625,
0.5947265625,
0.5009765625,
0.59521484375
],
[
0.080078125,
0.5869140625,
0.468017578125,
0.58837890625,
0.468017578125,
0.60302734375,
0.080078125,
0.6015625
],
[
0.50341796875,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.080078125,
0.60302734375,
0.468017578125,
0.60400390625,
0.468017578125,
0.61865234375,
0.080078125,
0.6171875
],
[
0.5048828125,
0.61279296875,
0.890625,
0.61279296875,
0.890625,
0.62353515625,
0.5048828125,
0.62353515625
],
[
0.08087158203125,
0.619140625,
0.313720703125,
0.619140625,
0.313720703125,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.50341796875,
0.62548828125,
0.89208984375,
0.625,
0.89208984375,
0.6396484375,
0.50341796875,
0.64013671875
],
[
0.5048828125,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65283203125,
0.5048828125,
0.65283203125
],
[
0.5009765625,
0.65478515625,
0.89208984375,
0.65576171875,
0.89208984375,
0.67041015625,
0.5009765625,
0.66943359375
],
[
0.08087158203125,
0.6640625,
0.20751953125,
0.6640625,
0.20751953125,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.50341796875,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.098876953125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50341796875,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50341796875,
0.70068359375,
0.89208984375,
0.70068359375,
0.89208984375,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.07843017578125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.07843017578125,
0.734375
],
[
0.50244140625,
0.71728515625,
0.890625,
0.71728515625,
0.890625,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50341796875,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.50341796875,
0.74560546875
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.74755859375,
0.890625,
0.74755859375,
0.890625,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76220703125,
0.89208984375,
0.76123046875,
0.89208984375,
0.77587890625,
0.50244140625,
0.7763671875
],
[
0.50244140625,
0.7763671875,
0.8896484375,
0.77587890625,
0.8896484375,
0.79052734375,
0.50244140625,
0.79150390625
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.07843017578125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.50244140625,
0.79248046875,
0.89208984375,
0.79248046875,
0.89208984375,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.806640625,
0.89208984375,
0.806640625,
0.89208984375,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82275390625,
0.89208984375,
0.82275390625,
0.89208984375,
0.8359375,
0.50244140625,
0.8359375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.8369140625,
0.89208984375,
0.8359375,
0.89208984375,
0.85107421875,
0.5009765625,
0.8515625
],
[
0.08251953125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08251953125,
0.86865234375
],
[
0.50341796875,
0.85302734375,
0.890625,
0.85302734375,
0.890625,
0.8662109375,
0.50341796875,
0.8662109375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.8662109375,
0.72802734375,
0.8671875,
0.72802734375,
0.8818359375,
0.5009765625,
0.880859375
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.521484375,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.521484375,
0.8994140625
]
] |
[
"Perceptual Image Quality Assessment with Transformers",
"Manri Cheon, Sung-Jun Yoon, Byungyeon Kang, and Junwoo Lee",
"LGElectronics",
"[manri.cheon, sungjun.yoon, byungyeon.kang, junwoo.lee]@lge.com",
"",
"Abstract",
"automatically predict perceived quality [51, 52, 40, 59, 53]",
"However, with the recent advances in deep learning-",
"In this paper, we propose an image gquality transformer",
"based image restoration algorithms, accurate prediction of",
"(IQT) that successfully applies a transformer architecture to",
"the perceived quality has become more difficult. In par-",
"a perceptual full-reference image quality assessment (IQA)",
"ticular, image restoration models based on generative ad-",
"task. Perceptual representation becomes more important in",
"versarial network (GAN) [20] have been developed in or-",
"image quality assessment. In this context, we extract the",
"5",
"der to improve the perceptual aspect of the result images",
"perceptual feature representations from each of input im-",
"[48, 2, 12, 8]. However, it sometimes generates output im-",
"ages using a convolutional neural network (CNN) back-",
"ages with unrealistic artifacts. The existing objective met-",
"bone. The extracted feature maps are fed into the trans-",
"LA",
"rics such as Peak Signal-to-Noise Ratio (PSNR), a structural",
"former encoder and decoder in order to compare a refer-",
"similarity index (SSIM) [51], and conventional quality met-",
"'SO",
"ence and distorted images. Following an approach of the",
"rics are insufficient to predict the quality of this kind of out-",
"transformer-based vision models [18, 55], we use extra",
"puts. In this respect, recent works [61, 16, 21, 39, 3] based",
"learnable quality embedding and position embedding. The",
"on perceptual representation exhibit a better performance at",
"",
"output of the transformer is passed to a prediction head in",
"the perceptual IQA task. As various image restoration algo-",
"order to predict a final quality score. The experimental re-",
"rithms are developed, however, it is still required to develop",
"sults show that our proposed model has an outstanding per-",
"the IQA algorithm that accurately predicts the perceptual",
"formance for the standard IQA datasets. For a large-scale",
"quality of images generated by emerging algorithms.",
"IQA dataset containing output images of generative model,",
"In recent years, based on the success in the natural lan-",
"our model also shows the promising results. The proposed",
"guage processing (NLP) field, the transformer [46] archi-",
"IQT was ranked first among 13 participants in the NTIRE",
"TXI0",
"tecture has been applied in the computer vision field [27].",
"2021 perceptual image quality assessment challenge [23].",
"A wider research area in the computer vision has been im-",
"Our work will be an opportunity to further expand the ap-",
"provedbasedonthetransformer,such asrecognitiontask",
"proachfor the perceptual IQA task.",
"[4, 45, 18], generative modelling [37, 26, 7], low-level vi-",
"sion[6,54.30l.etc.However.few attempts weremade",
"in the field of the image and video quality assessment. In",
"1.Introduction",
"a recent study, You and Korhonen proposed the applica-",
"Perceptual image quality assessment (IQA) is an impor-",
"tion of transformer in image quality assessment [55]. They",
"tant topic in the multimedia systems and computer vision",
"achieved outstanding performance on two publicly available",
"tasks [11, 42, 56]. One of the goals of the image process-",
"large-scale blind image quality databases. With our knowl-",
"ing is to improve thequality of the content to an accept-",
"edge, however, this study is the only transformer-based ap-",
"able level for the human viewers. In this context, the first",
"proach for image quality assessment. Therefore, it is ur-",
"step toward generating acceptable contents is to accurately",
"gently needed to investigate whether the transformer-based",
"approach works well in the field of perceptual image qual-",
"measure the perceptual quality of the content, which can",
"be performed via subjective and objective quality assess-",
"ity assessment. Especially, it should be investigated whether",
"ment [49, 9, 25, 19]. The subjective quality assessment is",
"this structure is applicable to a full-reference (FR) model",
"the most accurate method to measure the perceived qual-",
"aiming to measure the perceptual similarity between two",
"ity, which is usually represented by mean opinion scores",
"images. In addition, it is also necessary to evaluate whether",
"(MOS)from collected subjective ratings.However,it is",
"this approach can accurately predict the perceptual quality",
"time-consuming and expensive. Thus, objective quality as-",
"for the latest GAN-based artifacts.",
"sessment performed by objective metrics is widely used to",
"In this study, we propose an Image Quality Transformer"
] |
{
"arxiv_id": "2104.14730",
"language": "en",
"page_idx": 0
}
|
||
2104.14730-en-1
|
2104.14730-en-1.jpg
|
<FigureHere>
Figure 1. Model architecture of proposed image quality transformer (IQT). Note that $F_I$ denotes $F_d$ and $F_r$ in Eqs. 1 and 2, respectively.
(IQT), which is the FR image quality assessment method as shown in Fig. 1. To tackle the perceptual aspects, a convolutional neural network (CNN) backbone is used to extract perceptual representations from an input image. Based on the transformer encoder-decoder architecture, the proposed model is trained to predict the perceptual quality accurately. The proposed model was ranked in the first place among 13 participants in the NTIRE 2021 challenge on perceptual image quality assessment[23] at the CVPR 2021.
The rest of this article is organized as follows. The following section presents the related work. Section 3 describes the proposed method and the experiments are given in Section $\color{red}{4.\text{Finally, conclusions are given in Section 5.}}$
## $2. \textbf{ Related Work}$
Image Quality Assessment. The most important goal of the developing objective IQA is to accurately predict the perceived quality by human viewers. In general, the objective IQA methods can be classified into three categories according to the existence of reference information: FR [51,52,40,32,5], reduced-reference (RR)[43], and noreference (NR)[36, 35] IQA methods. The NR method is useful for the system because of its feasibility. However, the absence of a reference makes it challenging to predict image quality accurately compared to the FR method. The FR method focuses more on visual similarity or dissimilarity between two images, and this method still plays an impor-
tant role in the development of image processing system.
The representative of the commonly and widely used quality FR metric is the PSNR. It has the advantage of convenience for optimization; however, it tends to poorly predict perceived visual quality. Wang et al. proposed the SSIM [51] that is based on the fact that the human visual system (HVS) is highly correlated to structural information. Since that, various FR metrics have been developed to take into account various aspects of human quality perception, e.g., information-theoretic criterion [40,41], structural similarity[52,59], etc. Recently, CNN-based IQA methods as well as other low-level computer vision tasks have been actively studied [61,3,39,17,24]. Zhang et al. proposed a learned perceptual image patch similarity (LPIPS) metric[61] for $\hat{\mathrm{FR-IQA.~The~LPIPS~showed~that~trained~deep~features~that}}$ are optimized by the Euclidean distance between distorted and reference images are effective for IQA compared to the conventional IQA methods. Ding et $al.$ proposed the metric that is robust to texture resampling and geometric transformation based on spatial averages of the feature maps[16]. Various IQA methods including aforementioned metrics are included in our experiments for performance comparison.
The primary criterion of performance measurement is the accuracy of the metrics. Pearson linear correlation coefficient (PLCC) followed by the third-order polynomial nonlinear regression [42] is usually used in order to evaluate the accuracy of the methods. Spearman rank order correlation coefficient (SRCC) and the Kendall rank order corre-
|
[
[
0.50244140625,
0.10418701171875,
0.53515625,
0.10418701171875,
0.53515625,
0.11553955078125,
0.50244140625,
0.11553955078125
],
[
0.61767578125,
0.10418701171875,
0.66650390625,
0.10418701171875,
0.66650390625,
0.11871337890625,
0.61767578125,
0.11871337890625
],
[
0.50341796875,
0.11553955078125,
0.53515625,
0.11553955078125,
0.53515625,
0.126953125,
0.50341796875,
0.126953125
],
[
0.62158203125,
0.11871337890625,
0.66015625,
0.11871337890625,
0.66015625,
0.1300048828125,
0.62158203125,
0.1300048828125
],
[
0.529296875,
0.1395263671875,
0.55126953125,
0.1395263671875,
0.55126953125,
0.147705078125,
0.529296875,
0.147705078125
],
[
0.21728515625,
0.1622314453125,
0.3349609375,
0.1622314453125,
0.3349609375,
0.175537109375,
0.21728515625,
0.175537109375
],
[
0.57861328125,
0.1622314453125,
0.69775390625,
0.1622314453125,
0.69775390625,
0.175537109375,
0.57861328125,
0.175537109375
],
[
0.428955078125,
0.1717529296875,
0.48193359375,
0.1717529296875,
0.48193359375,
0.18310546875,
0.428955078125,
0.18310546875
],
[
0.2181396484375,
0.1932373046875,
0.26953125,
0.1932373046875,
0.26953125,
0.2052001953125,
0.2181396484375,
0.2052001953125
],
[
0.57666015625,
0.1932373046875,
0.62744140625,
0.1932373046875,
0.62744140625,
0.2052001953125,
0.57666015625,
0.2052001953125
],
[
0.64794921875,
0.1932373046875,
0.67236328125,
0.1932373046875,
0.67236328125,
0.20703125,
0.64794921875,
0.20703125
],
[
0.1168212890625,
0.2392578125,
0.181396484375,
0.2392578125,
0.181396484375,
0.25,
0.1168212890625,
0.25
],
[
0.27783203125,
0.1956787109375,
0.314453125,
0.1956787109375,
0.314453125,
0.20703125,
0.27783203125,
0.20703125
],
[
0.2115478515625,
0.242431640625,
0.2197265625,
0.2392578125,
0.22216796875,
0.242431640625,
0.214111328125,
0.24560546875
],
[
0.214111328125,
0.248779296875,
0.281982421875,
0.248779296875,
0.281982421875,
0.255615234375,
0.214111328125,
0.255615234375
],
[
0.73193359375,
0.2442626953125,
0.7490234375,
0.2442626953125,
0.7490234375,
0.2474365234375,
0.73193359375,
0.2474365234375
],
[
0.5888671875,
0.26318359375,
0.619140625,
0.26318359375,
0.619140625,
0.271484375,
0.5888671875,
0.271484375
],
[
0.37744140625,
0.271484375,
0.404296875,
0.271484375,
0.404296875,
0.2841796875,
0.37744140625,
0.2841796875
],
[
0.24267578125,
0.27783203125,
0.253173828125,
0.27783203125,
0.253173828125,
0.281005859375,
0.24267578125,
0.281005859375
],
[
0.51123046875,
0.281005859375,
0.5556640625,
0.281005859375,
0.5556640625,
0.28857421875,
0.51123046875,
0.28857421875
],
[
0.55224609375,
0.281005859375,
0.62841796875,
0.281005859375,
0.62841796875,
0.28857421875,
0.55224609375,
0.28857421875
],
[
0.6806640625,
0.27978515625,
0.76708984375,
0.27978515625,
0.76708984375,
0.290283203125,
0.6806640625,
0.290283203125
],
[
0.23779296875,
0.285888671875,
0.2548828125,
0.285888671875,
0.2548828125,
0.290283203125,
0.23779296875,
0.290283203125
],
[
0.281982421875,
0.3037109375,
0.3095703125,
0.3037109375,
0.3095703125,
0.3125,
0.281982421875,
0.3125
],
[
0.498291015625,
0.2998046875,
0.77783203125,
0.2998046875,
0.77783203125,
0.31005859375,
0.498291015625,
0.31005859375
],
[
0.12091064453125,
0.311279296875,
0.178955078125,
0.311279296875,
0.178955078125,
0.3212890625,
0.12091064453125,
0.3212890625
],
[
0.30810546875,
0.30615234375,
0.325927734375,
0.30615234375,
0.325927734375,
0.309326171875,
0.30810546875,
0.309326171875
],
[
0.289306640625,
0.313232421875,
0.3095703125,
0.313232421875,
0.3095703125,
0.31640625,
0.289306640625,
0.31640625
],
[
0.48193359375,
0.337158203125,
0.5556640625,
0.337158203125,
0.5556640625,
0.347900390625,
0.48193359375,
0.347900390625
],
[
0.75341796875,
0.339111328125,
0.779296875,
0.339111328125,
0.779296875,
0.347900390625,
0.75341796875,
0.347900390625
],
[
0.2156982421875,
0.34033203125,
0.22216796875,
0.34033203125,
0.22216796875,
0.343505859375,
0.2156982421875,
0.343505859375
],
[
0.1168212890625,
0.341552734375,
0.1829833984375,
0.341552734375,
0.1829833984375,
0.348388671875,
0.1168212890625,
0.348388671875
],
[
0.2156982421875,
0.348388671875,
0.281982421875,
0.348388671875,
0.281982421875,
0.35595703125,
0.2156982421875,
0.35595703125
],
[
0.60107421875,
0.35107421875,
0.64892578125,
0.35107421875,
0.64892578125,
0.361083984375,
0.60107421875,
0.361083984375
],
[
0.69921875,
0.3515625,
0.73388671875,
0.349853515625,
0.73388671875,
0.356689453125,
0.69921875,
0.359130859375
],
[
0.435546875,
0.35791015625,
0.4697265625,
0.3623046875,
0.46728515625,
0.371826171875,
0.43310546875,
0.367431640625
],
[
0.8056640625,
0.352783203125,
0.8349609375,
0.352783203125,
0.8349609375,
0.35986328125,
0.8056640625,
0.35986328125
],
[
0.50244140625,
0.369384765625,
0.53369140625,
0.369384765625,
0.53369140625,
0.376953125,
0.50244140625,
0.376953125
],
[
0.23779296875,
0.37744140625,
0.2548828125,
0.37744140625,
0.2548828125,
0.380615234375,
0.23779296875,
0.380615234375
],
[
0.50244140625,
0.37890625,
0.53369140625,
0.37890625,
0.53369140625,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.236083984375,
0.385009765625,
0.256591796875,
0.385009765625,
0.256591796875,
0.392578125,
0.236083984375,
0.392578125
],
[
0.438720703125,
0.389404296875,
0.459228515625,
0.393310546875,
0.455810546875,
0.401611328125,
0.435546875,
0.397705078125
],
[
0.60302734375,
0.397705078125,
0.64453125,
0.397705078125,
0.64453125,
0.40771484375,
0.60302734375,
0.40771484375
],
[
0.1217041015625,
0.412353515625,
0.17724609375,
0.412353515625,
0.17724609375,
0.419189453125,
0.1217041015625,
0.419189453125
],
[
0.28759765625,
0.410400390625,
0.303955078125,
0.410400390625,
0.303955078125,
0.4140625,
0.28759765625,
0.4140625
],
[
0.29736328125,
0.417236328125,
0.313720703125,
0.417236328125,
0.313720703125,
0.42041015625,
0.29736328125,
0.42041015625
],
[
0.080078125,
0.438232421875,
0.88134765625,
0.438232421875,
0.88134765625,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.080078125,
0.4765625,
0.4697265625,
0.477783203125,
0.4697265625,
0.492431640625,
0.080078125,
0.4912109375
],
[
0.5009765625,
0.4765625,
0.875,
0.477783203125,
0.875,
0.492431640625,
0.5009765625,
0.4912109375
],
[
0.080078125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.521484375,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.521484375,
0.5068359375
],
[
0.080078125,
0.50830078125,
0.4697265625,
0.50830078125,
0.4697265625,
0.521484375,
0.080078125,
0.521484375
],
[
0.5009765625,
0.50830078125,
0.89208984375,
0.5068359375,
0.89208984375,
0.521484375,
0.5009765625,
0.52294921875
],
[
0.080078125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.080078125,
0.537109375
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.07843017578125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.55322265625,
0.07843017578125,
0.55322265625
],
[
0.50244140625,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.080078125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50341796875,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50341796875,
0.5673828125
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.5048828125,
0.57080078125,
0.890625,
0.57080078125,
0.890625,
0.58056640625,
0.5048828125,
0.58056640625
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50341796875,
0.5859375,
0.890625,
0.5859375,
0.890625,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.07843017578125,
0.5986328125,
0.38818359375,
0.59765625,
0.38818359375,
0.6123046875,
0.07843017578125,
0.61279296875
],
[
0.5009765625,
0.5986328125,
0.890625,
0.599609375,
0.890625,
0.6142578125,
0.5009765625,
0.61279296875
],
[
0.098876953125,
0.6142578125,
0.46728515625,
0.6142578125,
0.46728515625,
0.62744140625,
0.098876953125,
0.62744140625
],
[
0.50341796875,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.07843017578125,
0.630859375,
0.468017578125,
0.630859375,
0.468017578125,
0.64453125,
0.07843017578125,
0.64453125
],
[
0.5009765625,
0.62890625,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.5009765625,
0.64306640625
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.50341796875,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.07843017578125,
0.6611328125,
0.44775390625,
0.6611328125,
0.44775390625,
0.6748046875,
0.07843017578125,
0.6748046875
],
[
0.5009765625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.5009765625,
0.67236328125,
0.890625,
0.67236328125,
0.890625,
0.6875,
0.5009765625,
0.6875
],
[
0.080078125,
0.6943359375,
0.2181396484375,
0.6943359375,
0.2181396484375,
0.70849609375,
0.080078125,
0.70849609375
],
[
0.50244140625,
0.689453125,
0.8935546875,
0.689453125,
0.8935546875,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.50244140625,
0.705078125,
0.8935546875,
0.705078125,
0.8935546875,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.5009765625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.5009765625,
0.73291015625
],
[
0.080078125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.8935546875,
0.78076171875,
0.8935546875,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.8798828125,
0.796875,
0.8798828125,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.07843017578125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.52294921875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.52294921875,
0.82470703125
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.5009765625,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5009765625,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50341796875,
0.88818359375,
0.890625,
0.88818359375,
0.890625,
0.8984375,
0.50341796875,
0.8984375
]
] |
[
"MLP",
"Quality",
"Head",
"Score",
"× 256",
"Transformer Encoder",
"Transformer Decoder",
"(N+1)×256",
"(N + 1)×256",
"(N +1)×256",
"Fref",
"Distorted Image",
"Faiff",
"",
"&Concatenation",
"",
"",
"faist",
"",
"F+P",
"FiE+PF+P",
"Fin-++PIn-FI,+Piw",
"lixe",
"HxWx",
"AddextralearnableQualityEmbedding&PositionEmbedding",
"H。xW.x3",
"21",
"X",
"Feature Projection",
"Flatter",
"",
"ReferenceImage",
"&Concatenation",
"HxWx256",
"NmHX",
"faiff",
"256)XM",
"Conv2D",
"",
"(x1)",
"fixec",
"fre",
"Fi=FaorF",
"H.XW.X3",
"XM",
"Hwt",
"Figure 1. Model architecture of proposed image quality transformer (IQT). Note that Fr denotes Fd and F, in Eqs. 1 and 2, respectively",
"(IQT), which is the FR image quality assessment method as",
"tant role in the development of image processing system.",
"shown in Fig. 1. To tackle the perceptual aspects, a convo-",
"The representative of the commonly and widely used",
"lutional neural network (CNN) backbone is used to extract",
"quality FR metric is the PSNR. It has the advantage of con-",
"perceptual representations from an input image. Based on",
"venience for optimization; however, it tends to poorly pre-",
"the transformer encoder-decoder architecture, the proposed",
"dict perceived visual quality. Wang et al. proposed the SSIM",
"model is trained to predict the perceptual quality accurately.",
"[51] that is based on the fact that the human visual system",
"The proposed model was ranked in the first place among 13",
"(HVS)ishighlycorrelatedtostructuralinformation.Since",
"participants in the NTIRE 2021 challenge on perceptual im-",
"that,variousFRmetricshave been developed totake into",
"age quality assessment [23] at the CVPR 2021.",
"account various aspects of human quality perception, e.g.,",
"The rest of this article is organized as follows. The fol-",
"information-theoretic criterion40,41l,structural similar-",
"lowing section presents the related work. Section 3 de-",
"ity [52, 59], etc. Recently, CNN-based IQA methods as well",
"scribes the proposed method and the experiments are given",
"as otherlow-level computervisiontaskshavebeen actively",
"in Section 4. Finally, conclusions are given in Section 5.",
"studied [61, 3, 39, 17, 24]. Zhang et al. proposed a learned",
"perceptual image patch similarity (LPIPS) metric [61] for",
"2.Related Work",
"FR-IQA. The LPIPS showed that trained deep features that",
"are optimized by the Euclidean distance between distorted",
"Image Quality Assessment.The most important goal of",
"and reference images are effective for IQA compared to the",
"the developing objective IQA is to accurately predict the",
"conventional IQA methods. Ding et al. proposed the metric",
"perceived quality by human viewers. In general, the objec-",
"that is robust to texture resampling and geometric transfor-",
"tive IQA methods can be classified into three categories",
"mation based on spatial averages of the feature maps [16].",
"according to the existence of reference information:FR",
"Various IQA methods including aforementioned metrics are",
"[51, 52, 40, 32, 5], reduced-reference (RR) [43], and no-",
"included in our experiments for performance comparison.",
"reference (NR) [36, 35] IQA methods. The NR method is",
"The primary criterion of performance measurement is",
"useful for the system because of its feasibility. However, the",
"the accuracy of the metrics. Pearson linear correlation coef-",
"absence of a reference makes it challenging to predict im-",
"ficient (PLCC)followedby thethird-order polynomial non-",
"age quality accurately compared to the FR method. The FR",
"linear regression [42] is usually used in order to evaluate",
"method focuses more on visual similarity or dissimilarity",
"the accuracy of the methods. Spearman rank order correla-",
"between two images, and this method still plays an impor-",
"tioncoefficient(SRCC)andtheKendallrankordercorre-"
] |
{
"arxiv_id": "2104.14730",
"language": "en",
"page_idx": 1
}
|
||
2104.14730-en-2
|
2104.14730-en-2.jpg
|
lation coefficient (KRCC) are used to estimate the monotonicity and consistency of the quality prediction. Additional statistical method[29] and an ambiguity based approach have also been proposed in [10]. In our study, we select the SRCC, KRCC, and PLCC as performance evaluation metrics.
Vision Transformer. The transformer[46]consists of multi-head attentions (MHAs), multi-layer perceptrons (MLPs) , layer normalizations ( LNs) [ 1] , and residual con- nections. Unlike the CNN, the transformer has a minimum inductive bias and can scale with the length of the input sequence without limiting factors. Recently, it has emerged that the transformer has combined with the CNN using the self-attention [4], and some of which have completely replaced CNN[47].
The transformer is mainly self-attention based approach. Since the self-attention layer aggregates global information from the entire input sequence, therefore, the model can capture the entire image for measuring the perceptual quality of the whole image. Vision Transformer (ViT)[18] is a representative success model among transformer-based vision models. A hybrid architecture was proposed for image recognition using a concord of CNN and the transformer encoder. It replaces the pixel patch embedding with the patches extracted from the CNN feature map. This architecture could be applied well in the IQA task, because the effectiveness of the deep features on the perceptual IQA task was demonstrated in recent studies [61,16,55].In DETR [4], the encoder-decoder architecture is employed and the decoder takes learned positional embeddings as object queries for object detection. This approach could be applied to the FR-IQA model that compares two images and measures the similarity. To measure similarity, one of the two images can be adopted as the query information in the self-attention layer. From the successful approaches using the transformer, we learn the direction to develop the perceptual IQA method with the transformer.
Vision Transformer based IQA. Inspired by ViT, TRIQ [55] naturally attempts to solve the blind IQA task using the transformer with the MLP head. In order to exploit ViT and handle images with different resolution, the TRIQ model defines the positional embedding with sufficient length to cover the maximal image resolution. The transformer encoder employs adaptive positional embedding, which handles a image with arbitrary resolutions. The output of the encoder is fed into the MLP head and the MLP head predicts the perceived image quality.
Basically, similar to the TRIQ, our proposed model applies the transformer architecture for the IQA task. However, additional aspects are considered in order to design the perceptual FR-IQA with the transformer. First, the trans-
<FigureHere>
Figure 2. The transformer encoder and decoder.
former encoder-decoder architecture is an important point in our approach. The reference information and the difference information between the distorted and reference images are employed as an input into the transformer. Second, we adopt the Siamese architecture to extract both the input feature representations from the reference and distorted images. For each image, by concatenating multiple feature maps extracted from intermediate layers, we obtained suffcient information for the model.
## $3. \textbf{Proposed Method}$
The proposed method that is illustrated in Fig. 1 consists of three main components: a feature extraction backbone, a transformer encoder-decoder, and a prediction head. First, we use a CNN backbone to extract feature representations from both reference and distorted input images. The extracted feature maps are projected to fixed size of vectors and flattened. In order to predict perceived quality, the trainable extra [quality] embedding is added to the sequence of embedded feature. It is similar to approach using [ class] token in previous transformer models [15, 18, 45]. The position embedding is also added in order to maintain the positional information. We pass this input feature embedding into the transformer encoder and decoder. The transformer encoder and decoder are based on the standard architecture of the transformer [46], where the structure is briefly illustrated in Fig. 2. The first vector of the output embedding of the decoder is fed into the MLP head in order to predict a single perceptual quality score.
Feature Extraction Backbone. A conventional CNN network, Inception-Resnet-V2 [44], is employed as the feature extraction backbone network. Pretrained weights on ImageNet[14] is imported and frozen.Feature maps
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.78759765625,
0.09783935546875,
0.84228515625,
0.09783935546875,
0.84228515625,
0.10919189453125,
0.78759765625,
0.10919189453125
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.72607421875,
0.11553955078125,
0.73876953125,
0.11553955078125,
0.73876953125,
0.11932373046875,
0.72607421875,
0.11932373046875
],
[
0.080078125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.740234375,
0.12249755859375,
0.82861328125,
0.125,
0.82861328125,
0.1395263671875,
0.740234375,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.07843017578125,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.7646484375,
0.1514892578125,
0.7998046875,
0.1514892578125,
0.7998046875,
0.1641845703125,
0.7646484375,
0.1641845703125
],
[
0.080078125,
0.169189453125,
0.1707763671875,
0.169189453125,
0.1707763671875,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.55126953125,
0.1673583984375,
0.60546875,
0.1673583984375,
0.60546875,
0.1817626953125,
0.55126953125,
0.1817626953125
],
[
0.5595703125,
0.1849365234375,
0.56640625,
0.1849365234375,
0.56640625,
0.1900634765625,
0.5595703125,
0.1900634765625
],
[
0.57177734375,
0.1912841796875,
0.65673828125,
0.1932373046875,
0.65673828125,
0.2083740234375,
0.57177734375,
0.205810546875
],
[
0.74169921875,
0.186279296875,
0.82666015625,
0.186279296875,
0.82666015625,
0.2008056640625,
0.74169921875,
0.2008056640625
],
[
0.08087158203125,
0.2020263671875,
0.4697265625,
0.2020263671875,
0.4697265625,
0.21533203125,
0.08087158203125,
0.21533203125
],
[
0.07843017578125,
0.216552734375,
0.4697265625,
0.2171630859375,
0.4697265625,
0.231689453125,
0.07843017578125,
0.2310791015625
],
[
0.74267578125,
0.2108154296875,
0.81787109375,
0.2108154296875,
0.81787109375,
0.222900390625,
0.74267578125,
0.222900390625
],
[
0.59716796875,
0.2203369140625,
0.63134765625,
0.2203369140625,
0.63134765625,
0.231689453125,
0.59716796875,
0.231689453125
],
[
0.7548828125,
0.2235107421875,
0.810546875,
0.2235107421875,
0.810546875,
0.23486328125,
0.7548828125,
0.23486328125
],
[
0.07843017578125,
0.2310791015625,
0.468017578125,
0.231689453125,
0.468017578125,
0.246826171875,
0.07843017578125,
0.24560546875
],
[
0.080078125,
0.2474365234375,
0.468017578125,
0.2474365234375,
0.468017578125,
0.261474609375,
0.080078125,
0.261474609375
],
[
0.56787109375,
0.251953125,
0.65673828125,
0.25390625,
0.65673828125,
0.26953125,
0.56787109375,
0.267578125
],
[
0.740234375,
0.25244140625,
0.8310546875,
0.255126953125,
0.8310546875,
0.268310546875,
0.740234375,
0.266357421875
],
[
0.07843017578125,
0.26318359375,
0.468017578125,
0.26318359375,
0.468017578125,
0.276611328125,
0.07843017578125,
0.276611328125
],
[
0.07843017578125,
0.279052734375,
0.468017578125,
0.27783203125,
0.468017578125,
0.291748046875,
0.07843017578125,
0.292236328125
],
[
0.57421875,
0.27978515625,
0.650390625,
0.27978515625,
0.650390625,
0.291748046875,
0.57421875,
0.291748046875
],
[
0.74267578125,
0.27978515625,
0.8203125,
0.27978515625,
0.8203125,
0.291748046875,
0.74267578125,
0.291748046875
],
[
0.080078125,
0.29345703125,
0.4697265625,
0.29345703125,
0.4697265625,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.57666015625,
0.29345703125,
0.65185546875,
0.29345703125,
0.65185546875,
0.300537109375,
0.57666015625,
0.300537109375
],
[
0.74267578125,
0.292236328125,
0.82177734375,
0.292236328125,
0.82177734375,
0.302978515625,
0.74267578125,
0.302978515625
],
[
0.07843017578125,
0.309326171875,
0.468017578125,
0.309326171875,
0.468017578125,
0.322509765625,
0.07843017578125,
0.322509765625
],
[
0.6005859375,
0.3037109375,
0.6259765625,
0.3037109375,
0.6259765625,
0.311279296875,
0.6005859375,
0.311279296875
],
[
0.080078125,
0.323974609375,
0.19775390625,
0.323974609375,
0.19775390625,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.09722900390625,
0.339111328125,
0.46728515625,
0.339111328125,
0.46728515625,
0.352783203125,
0.09722900390625,
0.352783203125
],
[
0.55810546875,
0.335205078125,
0.8349609375,
0.335205078125,
0.8349609375,
0.34521484375,
0.55810546875,
0.34521484375
],
[
0.080078125,
0.354248046875,
0.4697265625,
0.354248046875,
0.4697265625,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.08087158203125,
0.37060546875,
0.46728515625,
0.37060546875,
0.46728515625,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.50341796875,
0.375,
0.89208984375,
0.375,
0.89208984375,
0.38818359375,
0.50341796875,
0.38818359375
],
[
0.07843017578125,
0.38330078125,
0.468017578125,
0.3837890625,
0.468017578125,
0.3984375,
0.07843017578125,
0.397705078125
],
[
0.50244140625,
0.39013671875,
0.890625,
0.39013671875,
0.890625,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.080078125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.5009765625,
0.40478515625,
0.89208984375,
0.40478515625,
0.89208984375,
0.41845703125,
0.5009765625,
0.41845703125
],
[
0.080078125,
0.4140625,
0.468017578125,
0.4140625,
0.468017578125,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.5009765625,
0.419189453125,
0.89208984375,
0.41845703125,
0.89208984375,
0.43310546875,
0.5009765625,
0.434326171875
],
[
0.07843017578125,
0.427978515625,
0.468017578125,
0.4287109375,
0.468017578125,
0.443115234375,
0.07843017578125,
0.442626953125
],
[
0.50244140625,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.50244140625,
0.448974609375
],
[
0.07843017578125,
0.443115234375,
0.468017578125,
0.442626953125,
0.468017578125,
0.45703125,
0.07843017578125,
0.458251953125
],
[
0.50341796875,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.462158203125,
0.50341796875,
0.462158203125
],
[
0.07843017578125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.07843017578125,
0.472900390625
],
[
0.50244140625,
0.46533203125,
0.89208984375,
0.46533203125,
0.89208984375,
0.479248046875,
0.50244140625,
0.479248046875
],
[
0.0767822265625,
0.474853515625,
0.46728515625,
0.473388671875,
0.46728515625,
0.4873046875,
0.0767822265625,
0.488037109375
],
[
0.50244140625,
0.479736328125,
0.890625,
0.479736328125,
0.890625,
0.494384765625,
0.50244140625,
0.494384765625
],
[
0.080078125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.501953125,
0.080078125,
0.501953125
],
[
0.50244140625,
0.49560546875,
0.7138671875,
0.49560546875,
0.7138671875,
0.5087890625,
0.50244140625,
0.5087890625
],
[
0.080078125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.07843017578125,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.07843017578125,
0.53271484375
],
[
0.50341796875,
0.5244140625,
0.67236328125,
0.5244140625,
0.67236328125,
0.5390625,
0.50341796875,
0.5390625
],
[
0.080078125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.08087158203125,
0.5517578125,
0.46728515625,
0.5517578125,
0.46728515625,
0.5625,
0.08087158203125,
0.5625
],
[
0.52294921875,
0.5498046875,
0.890625,
0.5498046875,
0.890625,
0.564453125,
0.52294921875,
0.564453125
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.08087158203125,
0.5966796875,
0.46728515625,
0.5966796875,
0.46728515625,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.5009765625,
0.5947265625,
0.89208984375,
0.59521484375,
0.89208984375,
0.60986328125,
0.5009765625,
0.609375
],
[
0.080078125,
0.60986328125,
0.4697265625,
0.60986328125,
0.4697265625,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.60986328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.50244140625,
0.62353515625
],
[
0.080078125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64013671875,
0.890625,
0.64013671875,
0.890625,
0.65380859375,
0.50244140625,
0.65380859375
],
[
0.080078125,
0.65576171875,
0.35693359375,
0.65576171875,
0.35693359375,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50244140625,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.66943359375,
0.50244140625,
0.66943359375
],
[
0.5009765625,
0.66943359375,
0.89208984375,
0.67041015625,
0.89208984375,
0.68505859375,
0.5009765625,
0.6845703125
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.689453125,
0.468017578125,
0.7041015625,
0.080078125,
0.70263671875
],
[
0.50341796875,
0.68603515625,
0.890625,
0.68603515625,
0.890625,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.07843017578125,
0.705078125,
0.4697265625,
0.7041015625,
0.4697265625,
0.71826171875,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.080078125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50341796875,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50244140625,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.080078125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.50341796875,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.7734375,
0.50341796875,
0.7734375
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.080078125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50244140625,
0.79248046875,
0.890625,
0.79248046875,
0.890625,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.806640625,
0.875,
0.806640625,
0.875,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.080078125,
0.8271484375,
0.2998046875,
0.8271484375,
0.2998046875,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.098876953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.7294921875,
0.84228515625,
0.7294921875,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.72802734375,
0.84033203125,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.72802734375,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"lation coefficient (KRCC) are used to estimate the mono-",
"Decoder",
"tonicity and consistency of the quality prediction. Addi-",
"X",
"tional statistical method [29] and an ambiguity based ap-",
"Add&Norm",
"proach have also been proposed in [10]. In our study, we",
"select the SRCC, KRCC, and PLCC as performance evalu-",
"MLP",
"ation metrics.",
"Encoder",
"X",
"Add&Norm",
"Add&Norm",
"Vision Transformer.The transformer [46] consists of",
"multi-head attentions (MHAs), multi-layer perceptrons",
"Multi-Head",
"MLP",
"Attention",
"(MLPs), layer normalizations (LNs) [1], and residual con-",
"nections. Unlike the CNN, the transformer has a minimum",
"Add&Norm",
"Add&Norm",
"inductive bias and can scale with the length of the input se-",
"quencewithout limiting factors.Recently,ithas emerged",
"Multi-Head",
"Multi-Head",
"that the transformer has combined with the CNN using the",
"Self-Attention",
"Self-Attention",
"self-attention [4], and some of which have completely re-",
"Mg",
"placed CNN [47].",
"The transformer is mainly self-attention based approach.",
"Figure2.Thetransformer encoder and decoder",
"Since the self-attention layer aggregates global information",
"from the entireinput seguence.therefore,themodel can",
"former encoder-decoder architecture is an important point",
"capture the entire image for measuring the perceptual qual-",
"in our approach. The reference information and the differ-",
"ity of the whole image. Vision Transformer (ViT) [18]is a",
"ence information between the distorted and reference im-",
"representative success model among transformer-based vi-",
"ages are employed as an input into the transformer. Second,",
"sion models. A hybrid architecture was proposed for image",
"we adopt the Siamese architecture to extract both the in-",
"recognition using a concord of CNN and the transformer",
"putfeaturerepresentationsfromthereferenceanddistorted",
"encoder.It replaces the pixel patch embedding with the",
"images. For each image, by concatenating multiple feature",
"patches extractedfromtheCNNfeaturemap.Thisarchi-",
"maps extracted from intermediate layers, we obtained suffi-",
"tecturecouldbeappliedwellintheIQAtask,becausethe",
"cient information for the model.",
"effectiveness of the deep features on the perceptual IQA",
"task was demonstrated in recent studies [61,16, 55].In",
"3.Proposed Method",
"DETR [4], the encoder-decoder architecture is employed",
"andthedecodertakeslearnedpositional embeddingsasob-",
"The proposed method that is illustrated in Fig. I con-",
"ject queries for object detection. This approach could be ap-",
"sists of three main components: a feature extraction back-",
"plied to the FR-IQA model that compares two images and",
"bone, a transformer encoder-decoder, and a prediction head.",
"measures the similarity.To measure similarityone of the",
"First, we use a CNN backbone to extract feature repre-",
"two images can be adopted as the query information in the",
"sentations from both reference and distorted input images.",
"self-attention layer. From the successful approaches using",
"The extracted feature maps are projected to fixed size of",
"the transformer, we learn the direction to develop the per-",
"vectors and flattened. In order to predict perceived qual-",
"ceptual IQA method with the transformer.",
"ity, the trainable extra [guality] embedding is added",
"to the sequence of embedded feature. It is similar to ap-",
"Vision Transformer based IQA.Inspired by ViT, TRIQ",
"proach using [class ] token in previous transformer mod-",
"[55] naturally attempts to solve the blind IQA task using the",
"els[15,18,45].Thepositionembedding isalso added in",
"transformer with the MLP head. In order to exploit ViT and",
"order to maintain the positional information. We pass this",
"handle images with differentresolution,the TRIOmodel",
"inputfeature embeddinginto thetransformer encoder and",
"defines thepositional embedding with sufficient lengthto",
"decoder. The transformer encoder and decoder are based on",
"cover the maximal image resolution.The transformer en-",
"the standardarchitecture of thetransformer[46l.wherethe",
"coder employs adaptive positional embedding, which han-",
"structure is briefly illustrated in Fig. 2. The first vector of",
"dles a image with arbitrary resolutions. The output of the",
"the output embedding of the decoder is fed into the MLP",
"encoder is fed into the MLP head and the MLP head pre-",
"head in order to predict a single perceptual quality score.",
"dicts the perceived image quality.",
"Basically, similar to the TRIQ, our proposed model ap-",
"Feature Extraction Backbone.",
"A conventional CNN",
"plies the transformer architecture for the IQA task. How-",
"network, Inception-Resnet-V2 [44], is employed as the",
"ever, additional aspects are considered in order to design the",
"feature extraction backbone network. Pretrained weights",
"perceptual FR-IQA with the transformer. First, the trans-",
"on ImageNet [14] is imported and frozen. Feature maps"
] |
{
"arxiv_id": "2104.14730",
"language": "en",
"page_idx": 2
}
|
||
2104.14730-en-3
|
2104.14730-en-3.jpg
|
Table 1. IQA datasets for performance evaluation and model training
<table>
<tbody>
<tr>
<th>Database</th>
<th>$\#{\mathrm{Ref}}.$</th>
<th>$\overline{\#\mathrm{Dist.}}$</th>
<th>$\overline{\mathrm{Dist.~Type}}$</th>
<th>$\overline{\#\mathrm{Dist.~Type}}$</th>
<th>$\#{\mathrm{Rating}}$</th>
<th>$\overline{\text{Rating}}$ $\overline{\mathrm{Type}}$</th>
<th>$\overline{\mathrm{Env.}}$</th>
</tr>
<tr>
<td>LIVE $[42^{-}]$</td>
<td>$\overline{29}$</td>
<td>$\overline{779}$</td>
<td>traditional</td>
<td>$\overline{5}$</td>
<td>$\overline{25k}$</td>
<td>$MOS$</td>
<td>$\overline{\mathrm{lab}}$</td>
</tr>
<tr>
<td>$\mathbf{CSIO}$ $)[32^{-}$</td>
<td>30</td>
<td>866</td>
<td>traditional</td>
<td>6</td>
<td>$5k$</td>
<td>$MOS$</td>
<td>lab</td>
</tr>
<tr>
<td>TID2013 [38]</td>
<td>25</td>
<td>3,000</td>
<td>traditional</td>
<td>25</td>
<td>$524k$</td>
<td>$MOS$</td>
<td>lab</td>
</tr>
<tr>
<td>KADID-10k [33]</td>
<td>81</td>
<td>10.1k</td>
<td>traditional</td>
<td>25</td>
<td>30.4k</td>
<td>$MOS$</td>
<td>crowdsourcin</td>
</tr>
<tr>
<td>PIPAL 222</td>
<td>250</td>
<td>$29k$</td>
<td>trad. + alg. outputs</td>
<td>40</td>
<td>$1.13m$</td>
<td>$MOS$</td>
<td>crowdsourcin</td>
</tr>
</tbody>
</table>
tion of the decoder can be formulated as:
$$y_{L}=[F_{E_{0}},F_{E_{1}},...,F_{E_{N}}],\\z_{0}=[F_{r_{0}}+P_{r_{0}},F_{r_{1}}+P_{r_{1}},...,F_{r_{N}}+P_{r_{N}}],\\q_{i}=k_{i}=v_{i}=z_{i-1},\\z_{i}^{\prime}=LN(MHA(q_{i},k_{i},v_{i})+z_{i-1}),\\q_{i}^{\prime}=z_{i}^{\prime},\quad k_{i}^{\prime}=v_{i}^{\prime}=y_{L},\\z_{i}^{\prime\prime}=LN(MHA(q_{i}^{\prime},k_{i}^{\prime},v_{i}^{\prime})+z_{i}^{\prime}),\\z_{i}=LN(MLP(z_{i}^{\prime\prime})+z_{i}^{\prime\prime}),\quad i=1,...,L\\[F_{D_{0}},F_{D_{1}},...,F_{D_{N}}]=z_{L},$$
(2)
from six intermediate layers of Inception-Resnet-V2, i.e., $\{ mixed\_ 5b$, $block35\_ 2$, $block35\_ 4$, $block35\_ 6$, $block35\_ 8$, $block35\_ 10\} $, are extracted. The extracted feature maps have the same shape $f_{layer}\in \mathbb{R} ^{H\times W\times c}$,where $c= 320$, $\begin{array}{l}{\text{and they are concatenated into feature map. In other words,}}\\{\text{for an input image }I\in\mathbb{R}^{H_{0}\times W_{0}\times3},\text{the feature map }f\in}\end{array}$ $\mathbb{R}^{H\times W\times C}$,where $C=6\times c$,is extracted.
Both reference and distorted images are used; therefore, the two input feature maps, $f_{ref}$ and $f_{dist}$, are employed for the transformer, respectively. In order to obtain difference information between reference and distorted images, a difference feature map, $f_d$, is also used. It can be simply obtained by subtraction between two feature maps of reference and distorted images, i.e., $f_{diff}=f_{ref}-\bar{f}_{dist}.$
where $L$ denotes the number of decoder layers. The output embedding $F_{D}\in \mathbb{R} ^{( 1+ N) \times D}$ of the decoder is finally ob-
tained.
Head. The final quality prediction is computed in the prediction MLP head. The first vector of the decoder output, $F_{D_0}\in\mathbb{R}^{1\times D}$ in Eq. 2, is fed into the MLP head, which contains the quality information. The MLP head consists of two fully connected (FC) layers, and the first FC layer is used followed by the ReLU activation. The second FC layer has one channel to predict a single score.
Transformer encoder. A difference feature embedding, $F_d\in\mathbb{R}^{N\times D}$, is used as the input of the transformer encoder. We first reduce the channel dimension of the $f_d$ to the transformer dimension $D$ using a $1\times1$ convolution. Then, we flatten the spatial dimensions, which means the number of patches in the feature map is set as $N=H\times W.$ As often used in the vision transformer models [18,55], we append extra quality embedding at the beginning of the input feature embedding as $F_{d_0}.$ And the trainable position embedding $P_d\in\mathbb{R}^{(\breve{1}+N)\times\widetilde{D}}$ are also added in order to retain the positional information. The calculation of the encoder can be formulated as
$4. \textbf{Experiments}$
### 4.1.Datasets
(1)
$$y_{0}=[F_{d_{0}}+P_{d_{0}},F_{d_{1}}+P_{d_{1}},...,F_{d_{N}}+P_{d_{N}}],\\q_{i}=k_{i}=v_{i}=y_{i-1},\\y_{i}^{\prime}=LN(MHA(q_{i},k_{i},v_{i})+y_{i-1}),\\y_{i}=LN(MLP(y_{i}^{\prime})+y_{i}^{\prime}),\quad i=1,...,L\\[F_{E_{0}},F_{E_{1}},...,F_{E_{N}}]=y_{L},$$
We employ fve databases that are commonly used in the research of perceptual image quality assessment. The LIVE Image Quality Assessment Database (LIVE)[42], the Categorical Subjective Image Quality (CSIQ) database [32], and the TID2013 [38] are the databases that serve as baselines for full-reference IQA studies. These datasets only include traditional distortion types and the subjective scores are measured in the controlled laboratory environment. KADID-10k[33] is a large-scale IQA dataset and is chosen as the training dataset in our experiment. It is three times larger compared to the TID2013 [38] and the ratings are collected from crowdsourcing. The PIPAL[22] dataset is used for both the training and evaluation of the model in this study. A large quantity of distorted images including GAN based algorithms’ outputs and following human ratings are included in the PIPAL dataset. It is challenging for existing metrics to predict perceptual quality accurately [21]. Table l summarizes the characteristics of the datasets employed in this study
where $L$ denotes the number of the encoder layers. The out-put of the encoder $F_E\in\mathbb{R}^{(1+N)\times D}$ has the same size to
that of the input feature embedding.
Transformer decoder. The decoder takes the reference feature embedding $F_r\in\mathbb{R}^{N\times D}$,obtained through the channel reduction and flattening. The extra quality embedding and position embedding are also added to it. The output of the encoder, $F_E$, is used as an input of the decoder, and it is used as a key-value in the second MHA layer. The calcula-
|
[
[
0.27783203125,
0.08587646484375,
0.69140625,
0.08837890625,
0.69140625,
0.10162353515625,
0.27783203125,
0.0997314453125
],
[
0.381591796875,
0.0985107421875,
0.449462890625,
0.10101318359375,
0.449462890625,
0.11553955078125,
0.381591796875,
0.11297607421875
],
[
0.48779296875,
0.0985107421875,
0.564453125,
0.10101318359375,
0.564453125,
0.11553955078125,
0.48779296875,
0.11297607421875
],
[
0.583984375,
0.0985107421875,
0.6357421875,
0.10162353515625,
0.6337890625,
0.1142578125,
0.58251953125,
0.111083984375
],
[
0.65185546875,
0.0985107421875,
0.73046875,
0.10101318359375,
0.73046875,
0.11553955078125,
0.65185546875,
0.11297607421875
],
[
0.2459716796875,
0.10101318359375,
0.283447265625,
0.10101318359375,
0.283447265625,
0.1142578125,
0.2459716796875,
0.1142578125
],
[
0.1422119140625,
0.10162353515625,
0.2017822265625,
0.10162353515625,
0.2017822265625,
0.11297607421875,
0.1422119140625,
0.11297607421875
],
[
0.30224609375,
0.10162353515625,
0.343017578125,
0.10162353515625,
0.343017578125,
0.11297607421875,
0.30224609375,
0.11297607421875
],
[
0.77783203125,
0.10162353515625,
0.80810546875,
0.10162353515625,
0.80810546875,
0.1142578125,
0.77783203125,
0.1142578125
],
[
0.1380615234375,
0.111083984375,
0.2059326171875,
0.11297607421875,
0.2059326171875,
0.1275634765625,
0.1380615234375,
0.1256103515625
],
[
0.525390625,
0.10986328125,
0.53759765625,
0.11932373046875,
0.525390625,
0.1287841796875,
0.51318359375,
0.11932373046875
],
[
0.5947265625,
0.11236572265625,
0.6259765625,
0.11236572265625,
0.6259765625,
0.1275634765625,
0.5947265625,
0.1275634765625
],
[
0.310546875,
0.11297607421875,
0.338134765625,
0.11297607421875,
0.338134765625,
0.1256103515625,
0.310546875,
0.1256103515625
],
[
0.253173828125,
0.11297607421875,
0.275390625,
0.11297607421875,
0.275390625,
0.126953125,
0.253173828125,
0.126953125
],
[
0.6748046875,
0.11297607421875,
0.7099609375,
0.11297607421875,
0.7099609375,
0.1256103515625,
0.6748046875,
0.1256103515625
],
[
0.78125,
0.11297607421875,
0.80615234375,
0.11297607421875,
0.80615234375,
0.126953125,
0.78125,
0.126953125
],
[
0.1380615234375,
0.12371826171875,
0.2059326171875,
0.1256103515625,
0.2059326171875,
0.14013671875,
0.1380615234375,
0.1383056640625
],
[
0.384033203125,
0.1142578125,
0.449462890625,
0.1142578125,
0.449462890625,
0.1256103515625,
0.384033203125,
0.1256103515625
],
[
0.253173828125,
0.1256103515625,
0.275390625,
0.1256103515625,
0.275390625,
0.1383056640625,
0.253173828125,
0.1383056640625
],
[
0.3095703125,
0.1256103515625,
0.338134765625,
0.1256103515625,
0.338134765625,
0.1383056640625,
0.3095703125,
0.1383056640625
],
[
0.51708984375,
0.1256103515625,
0.5361328125,
0.1256103515625,
0.5361328125,
0.1395263671875,
0.51708984375,
0.1395263671875
],
[
0.5986328125,
0.1256103515625,
0.62109375,
0.1256103515625,
0.62109375,
0.1383056640625,
0.5986328125,
0.1383056640625
],
[
0.6748046875,
0.1256103515625,
0.7099609375,
0.1256103515625,
0.7099609375,
0.1383056640625,
0.6748046875,
0.1383056640625
],
[
0.78173828125,
0.1256103515625,
0.80615234375,
0.1256103515625,
0.80615234375,
0.1383056640625,
0.78173828125,
0.1383056640625
],
[
0.253173828125,
0.1363525390625,
0.273681640625,
0.1363525390625,
0.273681640625,
0.15087890625,
0.253173828125,
0.15087890625
],
[
0.384033203125,
0.126953125,
0.449462890625,
0.126953125,
0.449462890625,
0.1383056640625,
0.384033203125,
0.1383056640625
],
[
0.131591796875,
0.1383056640625,
0.2156982421875,
0.1383056640625,
0.2156982421875,
0.149658203125,
0.131591796875,
0.149658203125
],
[
0.306396484375,
0.1383056640625,
0.34228515625,
0.1383056640625,
0.34228515625,
0.149658203125,
0.306396484375,
0.149658203125
],
[
0.59326171875,
0.136962890625,
0.6259765625,
0.136962890625,
0.6259765625,
0.149658203125,
0.59326171875,
0.149658203125
],
[
0.6748046875,
0.1383056640625,
0.7099609375,
0.1383056640625,
0.7099609375,
0.149658203125,
0.6748046875,
0.149658203125
],
[
0.78125,
0.1383056640625,
0.80615234375,
0.1383056640625,
0.80615234375,
0.15087890625,
0.78125,
0.15087890625
],
[
0.51123046875,
0.143310546875,
0.52783203125,
0.1307373046875,
0.54150390625,
0.1427001953125,
0.52685546875,
0.154052734375
],
[
0.3857421875,
0.1395263671875,
0.44775390625,
0.1395263671875,
0.44775390625,
0.15087890625,
0.3857421875,
0.15087890625
],
[
0.11767578125,
0.146484375,
0.2254638671875,
0.149658203125,
0.2254638671875,
0.1641845703125,
0.11767578125,
0.1610107421875
],
[
0.515625,
0.1483154296875,
0.53759765625,
0.1483154296875,
0.53759765625,
0.1622314453125,
0.515625,
0.1622314453125
],
[
0.74658203125,
0.147705078125,
0.8408203125,
0.149658203125,
0.8408203125,
0.1641845703125,
0.74658203125,
0.1622314453125
],
[
0.306396484375,
0.149658203125,
0.34228515625,
0.149658203125,
0.34228515625,
0.1622314453125,
0.306396484375,
0.1622314453125
],
[
0.253173828125,
0.149658203125,
0.273681640625,
0.149658203125,
0.273681640625,
0.1634521484375,
0.253173828125,
0.1634521484375
],
[
0.5908203125,
0.149658203125,
0.62841796875,
0.149658203125,
0.62841796875,
0.1634521484375,
0.5908203125,
0.1634521484375
],
[
0.6748046875,
0.149658203125,
0.7099609375,
0.149658203125,
0.7099609375,
0.1622314453125,
0.6748046875,
0.1622314453125
],
[
0.5888671875,
0.1590576171875,
0.63134765625,
0.1610107421875,
0.6298828125,
0.1767578125,
0.58837890625,
0.1749267578125
],
[
0.384033203125,
0.15087890625,
0.44775390625,
0.15087890625,
0.44775390625,
0.1622314453125,
0.384033203125,
0.1622314453125
],
[
0.250732421875,
0.1610107421875,
0.279296875,
0.1610107421875,
0.279296875,
0.175537109375,
0.250732421875,
0.175537109375
],
[
0.310546875,
0.1622314453125,
0.339111328125,
0.1622314453125,
0.339111328125,
0.1749267578125,
0.310546875,
0.1749267578125
],
[
0.36279296875,
0.1622314453125,
0.471435546875,
0.1622314453125,
0.471435546875,
0.175537109375,
0.36279296875,
0.175537109375
],
[
0.515625,
0.1622314453125,
0.53759765625,
0.1622314453125,
0.53759765625,
0.175537109375,
0.515625,
0.175537109375
],
[
0.6748046875,
0.1622314453125,
0.7099609375,
0.1622314453125,
0.7099609375,
0.1749267578125,
0.6748046875,
0.1749267578125
],
[
0.7490234375,
0.1641845703125,
0.8388671875,
0.1641845703125,
0.8388671875,
0.1749267578125,
0.7490234375,
0.1749267578125
],
[
0.50341796875,
0.203857421875,
0.77294921875,
0.203857421875,
0.77294921875,
0.2171630859375,
0.50341796875,
0.2171630859375
],
[
0.136474609375,
0.1634521484375,
0.2059326171875,
0.1634521484375,
0.2059326171875,
0.1749267578125,
0.136474609375,
0.1749267578125
],
[
0.08251953125,
0.2052001953125,
0.46728515625,
0.2052001953125,
0.46728515625,
0.21533203125,
0.08251953125,
0.21533203125
],
[
0.08251953125,
0.2183837890625,
0.46728515625,
0.2183837890625,
0.46728515625,
0.231689453125,
0.08251953125,
0.231689453125
],
[
0.66650390625,
0.2259521484375,
0.84619140625,
0.2247314453125,
0.84619140625,
0.242431640625,
0.66650390625,
0.24365234375
],
[
0.080078125,
0.2342529296875,
0.46728515625,
0.2342529296875,
0.46728515625,
0.2474365234375,
0.080078125,
0.2474365234375
],
[
0.07598876953125,
0.246826171875,
0.4697265625,
0.242431640625,
0.4697265625,
0.26318359375,
0.07598876953125,
0.267578125
],
[
0.54736328125,
0.2442626953125,
0.8447265625,
0.24560546875,
0.8447265625,
0.261474609375,
0.54736328125,
0.260009765625
],
[
0.080078125,
0.264404296875,
0.468017578125,
0.264404296875,
0.468017578125,
0.27783203125,
0.080078125,
0.27783203125
],
[
0.705078125,
0.266357421875,
0.84716796875,
0.266357421875,
0.84716796875,
0.27978515625,
0.705078125,
0.27978515625
],
[
0.0767822265625,
0.274658203125,
0.471435546875,
0.274658203125,
0.471435546875,
0.29541015625,
0.0767822265625,
0.29541015625
],
[
0.6044921875,
0.281005859375,
0.84716796875,
0.2822265625,
0.84716796875,
0.2998046875,
0.6044921875,
0.298583984375
],
[
0.07598876953125,
0.2890625,
0.359375,
0.292236328125,
0.359375,
0.31005859375,
0.07598876953125,
0.306884765625
],
[
0.67724609375,
0.2998046875,
0.84716796875,
0.300537109375,
0.84716796875,
0.318115234375,
0.67724609375,
0.317626953125
],
[
0.869140625,
0.294189453125,
0.89208984375,
0.294189453125,
0.89208984375,
0.30615234375,
0.869140625,
0.30615234375
],
[
0.098876953125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.323974609375,
0.098876953125,
0.323974609375
],
[
0.61669921875,
0.317626953125,
0.84619140625,
0.3193359375,
0.84619140625,
0.337158203125,
0.61669921875,
0.335205078125
],
[
0.080078125,
0.32568359375,
0.4697265625,
0.32568359375,
0.4697265625,
0.34033203125,
0.080078125,
0.34033203125
],
[
0.08087158203125,
0.34228515625,
0.468017578125,
0.34228515625,
0.468017578125,
0.352783203125,
0.08087158203125,
0.352783203125
],
[
0.56640625,
0.339111328125,
0.84619140625,
0.339111328125,
0.84619140625,
0.354248046875,
0.56640625,
0.354248046875
],
[
0.0767822265625,
0.354736328125,
0.468017578125,
0.35595703125,
0.468017578125,
0.37060546875,
0.0767822265625,
0.369384765625
],
[
0.6640625,
0.356689453125,
0.8486328125,
0.35791015625,
0.8486328125,
0.375732421875,
0.6640625,
0.375
],
[
0.07843017578125,
0.371826171875,
0.468017578125,
0.371826171875,
0.468017578125,
0.385009765625,
0.07843017578125,
0.385009765625
],
[
0.080078125,
0.385009765625,
0.468017578125,
0.385009765625,
0.468017578125,
0.3984375,
0.080078125,
0.3984375
],
[
0.50341796875,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.50341796875,
0.3984375
],
[
0.080078125,
0.401611328125,
0.428955078125,
0.401611328125,
0.428955078125,
0.416748046875,
0.080078125,
0.416748046875
],
[
0.498291015625,
0.395751953125,
0.89453125,
0.39453125,
0.89453125,
0.415283203125,
0.498291015625,
0.416748046875
],
[
0.50244140625,
0.416748046875,
0.54833984375,
0.416748046875,
0.54833984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.08087158203125,
0.442626953125,
0.468017578125,
0.442626953125,
0.468017578125,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.5009765625,
0.4462890625,
0.890625,
0.447509765625,
0.890625,
0.462158203125,
0.5009765625,
0.4609375
],
[
0.080078125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.5009765625,
0.462158203125,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.5009765625,
0.47607421875
],
[
0.07843017578125,
0.47021484375,
0.4697265625,
0.4716796875,
0.4697265625,
0.486083984375,
0.07843017578125,
0.48486328125
],
[
0.5,
0.474853515625,
0.89453125,
0.472900390625,
0.89453125,
0.49365234375,
0.5,
0.49560546875
],
[
0.080078125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.08087158203125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.080078125,
0.517578125,
0.4697265625,
0.517578125,
0.4697265625,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.07843017578125,
0.5322265625,
0.46728515625,
0.53271484375,
0.46728515625,
0.546875,
0.07843017578125,
0.54541015625
],
[
0.50244140625,
0.5390625,
0.771484375,
0.5390625,
0.771484375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.080078125,
0.54736328125,
0.4697265625,
0.54736328125,
0.4697265625,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.080078125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.0767822265625,
0.57275390625,
0.47216796875,
0.57275390625,
0.47216796875,
0.59326171875,
0.0767822265625,
0.59326171875
],
[
0.50341796875,
0.5673828125,
0.63134765625,
0.5673828125,
0.63134765625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.080078125,
0.59326171875,
0.471435546875,
0.59326171875,
0.471435546875,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.59033203125,
0.60107421875,
0.59228515625,
0.60107421875,
0.60693359375,
0.50244140625,
0.60498046875
],
[
0.080078125,
0.609375,
0.2181396484375,
0.609375,
0.2181396484375,
0.619140625,
0.080078125,
0.619140625
],
[
0.51953125,
0.61279296875,
0.8935546875,
0.6142578125,
0.8935546875,
0.62890625,
0.51953125,
0.62744140625
],
[
0.5009765625,
0.6298828125,
0.890625,
0.6298828125,
0.890625,
0.64306640625,
0.5009765625,
0.64306640625
],
[
0.12335205078125,
0.63623046875,
0.4248046875,
0.63623046875,
0.4248046875,
0.65087890625,
0.12335205078125,
0.65087890625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.279296875,
0.6513671875,
0.42724609375,
0.65283203125,
0.42724609375,
0.67041015625,
0.279296875,
0.66943359375
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.1829833984375,
0.67236328125,
0.426513671875,
0.67236328125,
0.426513671875,
0.6875,
0.1829833984375,
0.6875
],
[
0.44287109375,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68603515625,
0.44287109375,
0.68603515625
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.148681640625,
0.69140625,
0.4248046875,
0.69140625,
0.4248046875,
0.70556640625,
0.148681640625,
0.70556640625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.2467041015625,
0.70849609375,
0.42724609375,
0.71044921875,
0.42724609375,
0.72802734375,
0.2467041015625,
0.72607421875
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.74072265625,
0.468017578125,
0.74072265625,
0.468017578125,
0.75390625,
0.080078125,
0.75390625
],
[
0.5009765625,
0.736328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.5009765625,
0.75
],
[
0.0767822265625,
0.75048828125,
0.473876953125,
0.75048828125,
0.473876953125,
0.771484375,
0.0767822265625,
0.771484375
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.07843017578125,
0.76953125,
0.316162109375,
0.77099609375,
0.316162109375,
0.78515625,
0.07843017578125,
0.7841796875
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.8095703125,
0.471435546875,
0.81005859375,
0.471435546875,
0.82470703125,
0.080078125,
0.82373046875
],
[
0.5009765625,
0.81103515625,
0.8896484375,
0.81103515625,
0.8896484375,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.07598876953125,
0.82080078125,
0.471435546875,
0.8212890625,
0.471435546875,
0.84228515625,
0.07598876953125,
0.84130859375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5,
0.83837890625,
0.8935546875,
0.8388671875,
0.8935546875,
0.85693359375,
0.5,
0.85595703125
],
[
0.080078125,
0.85693359375,
0.471435546875,
0.85693359375,
0.471435546875,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.65673828125,
0.88720703125,
0.65673828125,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Table 1.IQA datasets forperformance evaluation and model training.",
"Dist.Type",
"# Dist. Type",
"#Rating",
"Rating Type",
"# Ref.",
"Database",
"# Dist.",
"Env.",
"LIVE [42]",
"",
"25k",
"779",
"29",
"MOS",
"lab",
"CSIQ [32]",
"traditional",
"30",
"866",
"6",
"5k",
"MOS",
"lab",
"25",
"traditional",
"TID2013 [38]",
"3.000",
"524k",
"MOS",
"lab",
"3",
"traditional",
"KADID-10k [33]",
"25",
"crowdsourcing",
"10.1k",
"81",
"30.4k",
"MOS",
"1.13m",
"traditional",
"250",
"29k",
"trad.+alg. outputs",
"40",
"MOS",
"crowdsourcing",
"tion of the decoder can be formulated as:",
"PIPAL[22]",
"from six intermediate layers of Inception-Resnet-V2.i.e..",
"fmired_5b,block35.2,block35.4,block35.6,block35_8.",
"YL = [FEo, FET,. Fen],",
"block35_10),are extracted. The extracted feature maps",
"have the same shape fiayer e RHxWxc, where c- 320,",
"2o =[Fro + Pro, Fr + Pri,.., Frn + Prn],",
"and they are concatenated into feature map. In other words,",
"qi = ki = vi - zi-1,",
"for an input image I e RHoxWox3, the feature map f e",
"z, = LN(MHA(qi, ki, vi) + zi-1),",
"IRHxWxC, where C = 6 × c, is extracted.",
"=，==,",
"(2)",
"Both reference and distorted images are used; therefore,",
"z = LN(MHA(qi,, ) +z),",
"the two input feature maps, fref and fdist, are employed",
"for the transformer,respectively.In order to obtain differ-",
"z=LN(MLP(z) +z), i=1,.., L",
"ence information between reference and distorted images,",
"[FDo, FD,., Dn] = 2L,",
"a difference feature map, fd, is also used. It can be simply",
"obtained by subtraction between two feature maps of refer-",
"where L denotes the number of decoder layers. The output",
"ence and distorted images, i.e., faiff = fref -fdist.",
"embedding FD e IR(1+N)xD of the decoder is finally ob-",
"tained.",
"Transformer encoder. A difference feature embedding,",
"Head.The final quality prediction is computed in the pre-",
"Fa e IRNxD, is used as the input of the transformer en-",
"diction MLP head. The first vector of the decoder output,",
"coder. We first reduce the channel dimension of the fa to the",
"FDo E IRIxD in Eq. 2, is fed into the MLP head, which",
"transformer dimension D using a 1 × 1 convolution. Then,",
"contains the quality information. The MLP head consists of",
"weflattenthe spatialdimensions,whichmeansthenumber",
"two fully connected (FC) layers, and the first FC layer is",
"of patches in the feature map is set as N = H × W. As",
"used followed by the ReLU activation. The second FC layer",
"often used in thevisiontransformermodels[18,55],weap-",
"has one channel to predict a single score.",
"pend extra quality embedding at the beginning of the input",
"feature embedding as Fdo. And the trainable position em-",
"bedding Pa e R(i+N)xD are also added in order to retain",
"4. Experiments",
"the positional information. The calculation of the encoder",
"4.1. Datasets",
"can beformulated as",
"We employ five databases that are commonly used in",
"the research of perceptual image quality assessment. The",
"o =[Fdo +Pdo,Fdi + Pdi,.., Fan +Pdn],",
"LIVE Image Quality Assessment Database (LIVE) [42],",
"qi = ki= Ui = Yi-1,",
"the Categorical Subjective Image Quality (CSIQ) database",
"y, = LN(MHA(q, kz, vi) +yi-1),",
"(1)",
"[32], and the TID2013 [38] are the databases that serve",
"yi= LN(MLP(y)+y), i= 1,..,L",
"as baselines for full-reference IQA studies. These datasets",
"[FEo, Fer., Fen] = yL,",
"only include traditional distortion types and the subjective",
"scores are measured in the controlled laboratory environ-",
"where L denotes the number of the encoder layers. The out-",
"ment. KADID-10k [33] is a large-scale IQA dataset and is",
"put of the encoder Fe e R(I+N)xD has the same size to",
"chosen as the training dataset in our experiment. It is three",
"that of the input feature embedding.",
"times larger compared to the TID2013 [38] and the ratings",
"are collected from crowdsourcing. The PIPAL [22] dataset",
"is used for both the training and evaluation of the model",
"Transformer decoder. The decoder takes the reference",
"in this study. A large quantity of distorted images includ-",
"feature embedding F, e RN×D,obtained through the chan",
"ing GAN based algorithms' outputs and following human",
"nel reduction and flattening. The extra quality embedding",
"ratings are included in the PIPAL dataset. It is challenging",
"and position embedding are also added to it. The output of",
"for existing metrics to predict perceptual quality accurately",
"the encoder, Fe, is used as an input of the decoder, and it is",
"[21]. Table 1 summarizes the characteristics of the datasets",
"used as a key-value in the second MHA layer. The calcula-",
"employed in this study."
] |
{
"arxiv_id": "2104.14730",
"language": "en",
"page_idx": 3
}
|
||
2104.14730-en-4
|
2104.14730-en-4.jpg
|
Table 2. Performance comparison of the IQA methods on three standard IQA databases, i.e., LIVE[42], CSIQ [32], and TID2013 [38], in
terms of SRCC and KRCC. The top three performing methods are highlighted in bold face. Some results are borrowed from [16,21].
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="2">LIVE[ 4$2^{- }$] </th>
<th colspan="2">CSIQ[ 32] </th>
<th colspan="2">TID2013138</th>
</tr>
<tr>
<th>$\operatorname{SRCC}$</th>
<th>KRCC</th>
<th>$SRCC$</th>
<th>KRCC </th>
<th>$\operatorname{SRCC}$</th>
<th>$\mathbf{KRCC}$</th>
</tr>
<tr>
<td>PSNR</td>
<td>0.873</td>
<td>0.680</td>
<td>0.810</td>
<td>0.601</td>
<td>0.687</td>
<td>0.496</td>
</tr>
<tr>
<td>SSIM [51]</td>
<td>0.948</td>
<td>0.796</td>
<td>0.865</td>
<td>0.680</td>
<td>0.727</td>
<td>0.545</td>
</tr>
<tr>
<td>MS-SSIM [52]</td>
<td>0.951</td>
<td>0.805</td>
<td>0.906</td>
<td>0.730</td>
<td>0.786</td>
<td>0.605</td>
</tr>
<tr>
<td>$VSI$ $[58^{.}$</td>
<td>0.952</td>
<td>0.806</td>
<td>0.943</td>
<td>0.786</td>
<td>0.897</td>
<td>0.718</td>
</tr>
<tr>
<td>$\mathbf{MAD}$ [32</td>
<td>0.967</td>
<td>0.842</td>
<td>0.947</td>
<td>0.797</td>
<td>0.781</td>
<td>0.604</td>
</tr>
<tr>
<td>VIF [40]</td>
<td>0.964</td>
<td>0.828</td>
<td>0.911</td>
<td>0.743</td>
<td>0.677</td>
<td>0.518</td>
</tr>
<tr>
<td>FSIMc 591</td>
<td>0.965</td>
<td>0.836</td>
<td>0.931</td>
<td>0.769</td>
<td>0.851</td>
<td>0.667</td>
</tr>
<tr>
<td>NLPD [31]</td>
<td>0.937</td>
<td>0.778</td>
<td>0.932</td>
<td>0.769</td>
<td>0.800</td>
<td>0.625</td>
</tr>
<tr>
<td>GMSD [53]</td>
<td>0.960</td>
<td>0.827</td>
<td>0.950</td>
<td>0.804</td>
<td>0.804</td>
<td>0.634</td>
</tr>
<tr>
<td>WaDIOaM [3]</td>
<td>0.947</td>
<td>0.791</td>
<td>0.909</td>
<td>$\overline{0.732}$</td>
<td>0.831</td>
<td>0.631</td>
</tr>
<tr>
<td>PieAPP [39]</td>
<td>0.919</td>
<td>0.750</td>
<td>0.892</td>
<td>0.715</td>
<td>0.876</td>
<td>0.683</td>
</tr>
<tr>
<td>LPIPS [61]</td>
<td>0.932</td>
<td>0.765</td>
<td>0.876</td>
<td>0.689</td>
<td>0.670</td>
<td>0.497</td>
</tr>
<tr>
<td>DISTS [17]</td>
<td>0.954</td>
<td>0.811</td>
<td>0.929</td>
<td>0.767</td>
<td>0.830</td>
<td>0.639</td>
</tr>
<tr>
<td>$SWD$ [21]</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>0.819</td>
<td>0.634</td>
</tr>
<tr>
<td>$IQT$ (ours) $\frac{1}{2}=\frac{1}{2}$</td>
<td>0.970</td>
<td>0.849</td>
<td>0.943</td>
<td>0.799</td>
<td>0.899</td>
<td>0.717</td>
</tr>
<tr>
<td>IOT- C $C($ours)</td>
<td>0.917</td>
<td>0.737</td>
<td>0.851</td>
<td>0.649</td>
<td>0.804</td>
<td>0.607</td>
</tr>
</tbody>
</table>
## 4.2. Implementation details
<FigureHere>
Figure 3. Scatter plots of ground-truth mean opinion scores
We denote our model trained on the KADID-10k as IQT. The hyper-parameters for the model are set as follow: i) the number of encoder and decoder layer is set to 2 (i.e., L= 2), ii) the number of heads in the MHA is set to 4 (i.e., $H=4)$, iii) the transformer dimension is set to 256 (i.e., D= 256), iv) dimension of the MLP in the encoder and decoder is set to 1024 (i.e., $D_feat=1024)$, v) the dimension of the first FC layer in MLP head is set to 512 (i.e., $D_{head}=512).$
In the training phase, a given image is cropped to obtain image patches. The dimension of the patch fed into the proposed ÎQT is $256\times256\times3.$ The number of patches in the feature map is set to $N=891.$ In the testing phase, image patches are also acquired from the given image pair. We extract $M$ overlapping patches and predict final quality score by averaging $\hat{M}$ individual quality scores of the patches. The stride size is set as large as possible to cover the entire image with fewer patches.
Data augmentation including horizontal flip and random rotation is applied during the training. The training is conducted using an ADAM [28] optimizer with a batch size of 16. Initial learning rate $2\times10^{-4}$ and cosine learning rate decay are set. The training loss is computed using a mean squared error (MSE) loss function. Our network is implemented using Tensorflow framework. It roughly takes a half day with a single NVIDIA TITAN RTX to train our model.
(MOSs) against predicted scores of proposed IQT on LIVE, CSIQ, TID2013, and PIPAL datasets. The predicted scores are obtained from the model trained on KADID-10k dataset.
ter plots of the predicted scores of IQT and the ground-truth MOS are also presented in Fig. 3. For LIVE and TID2013 databases, the proposed IQT shows the best performance in terms of SRCC. Also, it is ranked in the top three in all benchmarks in terms of SRCC and KRCC. In particular, our method shows better performance than recent deep learning-based methods [3, 39,61, 16, 21] for most cases.
Example images of the PIPAL validation dataset and following PSNR, SSIM [ 51] , MS- SSIM [ 52] , LPIPS [ 61] , and proposed IQT are illustrated in Fig. 4. From the left to the right, the perceptually better images to worse images are listed based on MOS. Our proposed IQT predicts the quality scores similar to MOS in terms of the superiority. There exist the images that are clearly distinguished by all methods, however, it is difficult to accurately predict perceptual
### $4. 3. \textbf{Results}$
The proposed IQT shows that the transformer based model is sufficiently competitive compared to existing approaches for the dataset that has traditional distortions. Our model is trained on KADID-10k and, then, the performance on the three standard IQA datasets is evaluated. The performance comparison result is reported in Table 2 and the scat-
|
[
[
0.080078125,
0.08587646484375,
0.89208984375,
0.08709716796875,
0.89208984375,
0.10162353515625,
0.080078125,
0.10101318359375
],
[
0.07843017578125,
0.10162353515625,
0.86376953125,
0.10162353515625,
0.86376953125,
0.11553955078125,
0.07843017578125,
0.11553955078125
],
[
0.318603515625,
0.11614990234375,
0.37744140625,
0.11614990234375,
0.37744140625,
0.1275634765625,
0.318603515625,
0.1275634765625
],
[
0.52783203125,
0.11614990234375,
0.58642578125,
0.11614990234375,
0.58642578125,
0.1275634765625,
0.52783203125,
0.1275634765625
],
[
0.72607421875,
0.11297607421875,
0.8095703125,
0.11553955078125,
0.8095703125,
0.1300048828125,
0.72607421875,
0.1275634765625
],
[
0.1192626953125,
0.12060546875,
0.1707763671875,
0.12060546875,
0.1707763671875,
0.1351318359375,
0.1192626953125,
0.1351318359375
],
[
0.26953125,
0.1256103515625,
0.316162109375,
0.1256103515625,
0.316162109375,
0.14013671875,
0.26953125,
0.14013671875
],
[
0.68896484375,
0.1256103515625,
0.73779296875,
0.1256103515625,
0.73779296875,
0.14013671875,
0.68896484375,
0.14013671875
],
[
0.79833984375,
0.1256103515625,
0.84619140625,
0.1256103515625,
0.84619140625,
0.14013671875,
0.79833984375,
0.14013671875
],
[
0.58837890625,
0.126953125,
0.63232421875,
0.125,
0.6337890625,
0.136962890625,
0.5888671875,
0.1395263671875
],
[
0.3798828125,
0.126953125,
0.42236328125,
0.126953125,
0.42236328125,
0.1383056640625,
0.3798828125,
0.1383056640625
],
[
0.48193359375,
0.1275634765625,
0.525390625,
0.1275634765625,
0.525390625,
0.1395263671875,
0.48193359375,
0.1395263671875
],
[
0.1192626953125,
0.1395263671875,
0.160888671875,
0.1395263671875,
0.160888671875,
0.154052734375,
0.1192626953125,
0.154052734375
],
[
0.273681640625,
0.1395263671875,
0.314453125,
0.1395263671875,
0.314453125,
0.154052734375,
0.273681640625,
0.154052734375
],
[
0.378173828125,
0.1395263671875,
0.420654296875,
0.1395263671875,
0.420654296875,
0.154052734375,
0.378173828125,
0.154052734375
],
[
0.484375,
0.1395263671875,
0.525390625,
0.1395263671875,
0.525390625,
0.154052734375,
0.484375,
0.154052734375
],
[
0.5888671875,
0.1395263671875,
0.6298828125,
0.1395263671875,
0.6298828125,
0.154052734375,
0.5888671875,
0.154052734375
],
[
0.69287109375,
0.1395263671875,
0.73388671875,
0.136962890625,
0.734375,
0.15283203125,
0.69384765625,
0.1546630859375
],
[
0.7998046875,
0.1395263671875,
0.8408203125,
0.1395263671875,
0.8408203125,
0.154052734375,
0.7998046875,
0.154052734375
],
[
0.273681640625,
0.1514892578125,
0.314453125,
0.1514892578125,
0.314453125,
0.1666259765625,
0.273681640625,
0.1666259765625
],
[
0.378173828125,
0.1514892578125,
0.420654296875,
0.1514892578125,
0.420654296875,
0.1666259765625,
0.378173828125,
0.1666259765625
],
[
0.484375,
0.1514892578125,
0.525390625,
0.1514892578125,
0.525390625,
0.1666259765625,
0.484375,
0.1666259765625
],
[
0.5888671875,
0.1514892578125,
0.6298828125,
0.1514892578125,
0.6298828125,
0.1666259765625,
0.5888671875,
0.1666259765625
],
[
0.69384765625,
0.15283203125,
0.734375,
0.15283203125,
0.734375,
0.1673583984375,
0.69384765625,
0.1673583984375
],
[
0.80126953125,
0.154052734375,
0.83837890625,
0.154052734375,
0.83837890625,
0.1654052734375,
0.80126953125,
0.1654052734375
],
[
0.378173828125,
0.1641845703125,
0.419189453125,
0.1641845703125,
0.419189453125,
0.1787109375,
0.378173828125,
0.1787109375
],
[
0.12091064453125,
0.154052734375,
0.1854248046875,
0.154052734375,
0.1854248046875,
0.1654052734375,
0.12091064453125,
0.1654052734375
],
[
0.484375,
0.1641845703125,
0.525390625,
0.1641845703125,
0.525390625,
0.1787109375,
0.484375,
0.1787109375
],
[
0.5888671875,
0.1641845703125,
0.6298828125,
0.1641845703125,
0.6298828125,
0.1787109375,
0.5888671875,
0.1787109375
],
[
0.69384765625,
0.1641845703125,
0.734375,
0.1641845703125,
0.734375,
0.1787109375,
0.69384765625,
0.1787109375
],
[
0.80126953125,
0.1654052734375,
0.8388671875,
0.1654052734375,
0.8388671875,
0.1787109375,
0.80126953125,
0.1787109375
],
[
0.275390625,
0.1654052734375,
0.313720703125,
0.1654052734375,
0.313720703125,
0.1787109375,
0.275390625,
0.1787109375
],
[
0.12091064453125,
0.1666259765625,
0.2099609375,
0.1666259765625,
0.2099609375,
0.177978515625,
0.12091064453125,
0.177978515625
],
[
0.273681640625,
0.1767578125,
0.314453125,
0.1767578125,
0.314453125,
0.1912841796875,
0.273681640625,
0.1912841796875
],
[
0.378173828125,
0.1767578125,
0.420654296875,
0.1767578125,
0.420654296875,
0.1912841796875,
0.378173828125,
0.1912841796875
],
[
0.484375,
0.1767578125,
0.525390625,
0.1767578125,
0.525390625,
0.1912841796875,
0.484375,
0.1912841796875
],
[
0.5888671875,
0.1767578125,
0.6298828125,
0.1767578125,
0.6298828125,
0.1912841796875,
0.5888671875,
0.1912841796875
],
[
0.69677734375,
0.177978515625,
0.734375,
0.177978515625,
0.734375,
0.1912841796875,
0.69677734375,
0.1912841796875
],
[
0.7998046875,
0.1767578125,
0.8408203125,
0.1767578125,
0.8408203125,
0.1912841796875,
0.7998046875,
0.1912841796875
],
[
0.273681640625,
0.1881103515625,
0.314453125,
0.1900634765625,
0.313720703125,
0.2052001953125,
0.27294921875,
0.20263671875
],
[
0.12091064453125,
0.1787109375,
0.176513671875,
0.1787109375,
0.176513671875,
0.1900634765625,
0.12091064453125,
0.1900634765625
],
[
0.5888671875,
0.189453125,
0.6298828125,
0.189453125,
0.6298828125,
0.203857421875,
0.5888671875,
0.203857421875
],
[
0.7998046875,
0.189453125,
0.8408203125,
0.189453125,
0.8408203125,
0.203857421875,
0.7998046875,
0.203857421875
],
[
0.486083984375,
0.1900634765625,
0.52392578125,
0.1900634765625,
0.52392578125,
0.203857421875,
0.486083984375,
0.203857421875
],
[
0.12091064453125,
0.1912841796875,
0.1846923828125,
0.1912841796875,
0.1846923828125,
0.20263671875,
0.12091064453125,
0.20263671875
],
[
0.381591796875,
0.1900634765625,
0.419189453125,
0.1900634765625,
0.419189453125,
0.203857421875,
0.381591796875,
0.203857421875
],
[
0.6953125,
0.1912841796875,
0.73193359375,
0.1912841796875,
0.73193359375,
0.20263671875,
0.6953125,
0.20263671875
],
[
0.273681640625,
0.2008056640625,
0.314453125,
0.20263671875,
0.313720703125,
0.2171630859375,
0.27294921875,
0.21533203125
],
[
0.378173828125,
0.2020263671875,
0.419189453125,
0.2020263671875,
0.419189453125,
0.216552734375,
0.378173828125,
0.216552734375
],
[
0.5888671875,
0.2020263671875,
0.6298828125,
0.2020263671875,
0.6298828125,
0.216552734375,
0.5888671875,
0.216552734375
],
[
0.7998046875,
0.2020263671875,
0.8388671875,
0.2020263671875,
0.8388671875,
0.2171630859375,
0.7998046875,
0.2171630859375
],
[
0.69384765625,
0.20263671875,
0.734375,
0.20263671875,
0.734375,
0.2171630859375,
0.69384765625,
0.2171630859375
],
[
0.12091064453125,
0.203857421875,
0.1748046875,
0.203857421875,
0.1748046875,
0.21533203125,
0.12091064453125,
0.21533203125
],
[
0.486083984375,
0.20263671875,
0.52392578125,
0.20263671875,
0.52392578125,
0.216552734375,
0.486083984375,
0.216552734375
],
[
0.273681640625,
0.2139892578125,
0.314453125,
0.2139892578125,
0.314453125,
0.2291259765625,
0.273681640625,
0.2291259765625
],
[
0.3798828125,
0.2139892578125,
0.420654296875,
0.2139892578125,
0.420654296875,
0.2291259765625,
0.3798828125,
0.2291259765625
],
[
0.5888671875,
0.2139892578125,
0.6298828125,
0.2139892578125,
0.6298828125,
0.2291259765625,
0.5888671875,
0.2291259765625
],
[
0.486083984375,
0.21533203125,
0.52392578125,
0.21533203125,
0.52392578125,
0.2291259765625,
0.486083984375,
0.2291259765625
],
[
0.12091064453125,
0.21533203125,
0.191162109375,
0.21533203125,
0.191162109375,
0.2266845703125,
0.12091064453125,
0.2266845703125
],
[
0.69677734375,
0.21533203125,
0.73193359375,
0.21533203125,
0.73193359375,
0.2279052734375,
0.69677734375,
0.2279052734375
],
[
0.80126953125,
0.21533203125,
0.8388671875,
0.21533203125,
0.8388671875,
0.2291259765625,
0.80126953125,
0.2291259765625
],
[
0.11767578125,
0.2247314453125,
0.191162109375,
0.2266845703125,
0.191162109375,
0.2412109375,
0.11767578125,
0.2392578125
],
[
0.273681640625,
0.2266845703125,
0.314453125,
0.2266845703125,
0.314453125,
0.2412109375,
0.273681640625,
0.2412109375
],
[
0.378173828125,
0.2266845703125,
0.419189453125,
0.2266845703125,
0.419189453125,
0.242431640625,
0.378173828125,
0.242431640625
],
[
0.484375,
0.2266845703125,
0.525390625,
0.2266845703125,
0.525390625,
0.2412109375,
0.484375,
0.2412109375
],
[
0.5888671875,
0.2266845703125,
0.6298828125,
0.2266845703125,
0.6298828125,
0.2412109375,
0.5888671875,
0.2412109375
],
[
0.69384765625,
0.2266845703125,
0.734375,
0.2266845703125,
0.734375,
0.2412109375,
0.69384765625,
0.2412109375
],
[
0.80126953125,
0.2279052734375,
0.8388671875,
0.2279052734375,
0.8388671875,
0.2412109375,
0.80126953125,
0.2412109375
],
[
0.273681640625,
0.2392578125,
0.314453125,
0.2392578125,
0.314453125,
0.25390625,
0.273681640625,
0.25390625
],
[
0.378173828125,
0.2392578125,
0.419189453125,
0.2392578125,
0.419189453125,
0.25390625,
0.378173828125,
0.25390625
],
[
0.484375,
0.2392578125,
0.525390625,
0.2392578125,
0.525390625,
0.25390625,
0.484375,
0.25390625
],
[
0.5888671875,
0.2392578125,
0.6298828125,
0.2392578125,
0.6298828125,
0.25390625,
0.5888671875,
0.25390625
],
[
0.69384765625,
0.2392578125,
0.734375,
0.2392578125,
0.734375,
0.25390625,
0.69384765625,
0.25390625
],
[
0.7998046875,
0.2392578125,
0.8408203125,
0.2392578125,
0.8408203125,
0.25390625,
0.7998046875,
0.25390625
],
[
0.273681640625,
0.251953125,
0.314453125,
0.251953125,
0.314453125,
0.266357421875,
0.273681640625,
0.266357421875
],
[
0.12091064453125,
0.2412109375,
0.19287109375,
0.2412109375,
0.19287109375,
0.251953125,
0.12091064453125,
0.251953125
],
[
0.37744140625,
0.251953125,
0.418212890625,
0.25,
0.419189453125,
0.26513671875,
0.378173828125,
0.267578125
],
[
0.48779296875,
0.25244140625,
0.52294921875,
0.25244140625,
0.52294921875,
0.26513671875,
0.48779296875,
0.26513671875
],
[
0.5908203125,
0.25244140625,
0.62841796875,
0.25244140625,
0.62841796875,
0.266357421875,
0.5908203125,
0.266357421875
],
[
0.6953125,
0.25390625,
0.73193359375,
0.25390625,
0.73193359375,
0.26513671875,
0.6953125,
0.26513671875
],
[
0.80126953125,
0.25390625,
0.83837890625,
0.25390625,
0.83837890625,
0.26513671875,
0.80126953125,
0.26513671875
],
[
0.6953125,
0.26318359375,
0.734375,
0.26513671875,
0.73388671875,
0.281005859375,
0.69384765625,
0.279052734375
],
[
0.12091064453125,
0.25390625,
0.2059326171875,
0.25390625,
0.2059326171875,
0.26513671875,
0.12091064453125,
0.26513671875
],
[
0.273681640625,
0.264404296875,
0.314453125,
0.264404296875,
0.314453125,
0.279052734375,
0.273681640625,
0.279052734375
],
[
0.378173828125,
0.264404296875,
0.420654296875,
0.264404296875,
0.420654296875,
0.279052734375,
0.378173828125,
0.279052734375
],
[
0.484375,
0.264404296875,
0.525390625,
0.264404296875,
0.525390625,
0.27978515625,
0.484375,
0.27978515625
],
[
0.5888671875,
0.264404296875,
0.6298828125,
0.264404296875,
0.6298828125,
0.27978515625,
0.5888671875,
0.27978515625
],
[
0.7998046875,
0.264404296875,
0.8408203125,
0.264404296875,
0.8408203125,
0.27978515625,
0.7998046875,
0.27978515625
],
[
0.273681640625,
0.276611328125,
0.314453125,
0.276611328125,
0.314453125,
0.291748046875,
0.273681640625,
0.291748046875
],
[
0.12091064453125,
0.266357421875,
0.1953125,
0.266357421875,
0.1953125,
0.27783203125,
0.12091064453125,
0.27783203125
],
[
0.484375,
0.276611328125,
0.525390625,
0.276611328125,
0.525390625,
0.292236328125,
0.484375,
0.292236328125
],
[
0.5888671875,
0.276611328125,
0.6298828125,
0.276611328125,
0.6298828125,
0.291748046875,
0.5888671875,
0.291748046875
],
[
0.69384765625,
0.276611328125,
0.736328125,
0.276611328125,
0.736328125,
0.291748046875,
0.69384765625,
0.291748046875
],
[
0.80126953125,
0.27783203125,
0.8388671875,
0.27783203125,
0.8388671875,
0.291748046875,
0.80126953125,
0.291748046875
],
[
0.378173828125,
0.27783203125,
0.419189453125,
0.27783203125,
0.419189453125,
0.292236328125,
0.378173828125,
0.292236328125
],
[
0.12091064453125,
0.279052734375,
0.1895751953125,
0.279052734375,
0.1895751953125,
0.290283203125,
0.12091064453125,
0.290283203125
],
[
0.11767578125,
0.28857421875,
0.19287109375,
0.290283203125,
0.19287109375,
0.304931640625,
0.11767578125,
0.302978515625
],
[
0.273681640625,
0.2890625,
0.314453125,
0.2890625,
0.314453125,
0.3037109375,
0.273681640625,
0.3037109375
],
[
0.378173828125,
0.2890625,
0.419189453125,
0.2890625,
0.419189453125,
0.3037109375,
0.378173828125,
0.3037109375
],
[
0.484375,
0.2890625,
0.525390625,
0.2890625,
0.525390625,
0.3037109375,
0.484375,
0.3037109375
],
[
0.5888671875,
0.2890625,
0.6298828125,
0.2890625,
0.6298828125,
0.304931640625,
0.5888671875,
0.304931640625
],
[
0.69384765625,
0.2890625,
0.736328125,
0.2890625,
0.736328125,
0.3037109375,
0.69384765625,
0.3037109375
],
[
0.7998046875,
0.2890625,
0.8408203125,
0.2890625,
0.8408203125,
0.3037109375,
0.7998046875,
0.3037109375
],
[
0.11767578125,
0.300537109375,
0.1846923828125,
0.302978515625,
0.1846923828125,
0.317626953125,
0.11767578125,
0.31494140625
],
[
0.69677734375,
0.302978515625,
0.734375,
0.302978515625,
0.734375,
0.31640625,
0.69677734375,
0.31640625
],
[
0.7998046875,
0.3017578125,
0.8408203125,
0.3017578125,
0.8408203125,
0.31640625,
0.7998046875,
0.31640625
],
[
0.273681640625,
0.31494140625,
0.314453125,
0.31494140625,
0.314453125,
0.330078125,
0.273681640625,
0.330078125
],
[
0.37744140625,
0.31494140625,
0.418212890625,
0.313232421875,
0.419189453125,
0.328857421875,
0.378173828125,
0.330810546875
],
[
0.484375,
0.31494140625,
0.525390625,
0.31494140625,
0.525390625,
0.330078125,
0.484375,
0.330078125
],
[
0.5888671875,
0.31494140625,
0.6298828125,
0.31494140625,
0.6298828125,
0.330078125,
0.5888671875,
0.330078125
],
[
0.69677734375,
0.31640625,
0.734375,
0.31640625,
0.734375,
0.330078125,
0.69677734375,
0.330078125
],
[
0.7998046875,
0.31494140625,
0.8408203125,
0.31494140625,
0.8408203125,
0.330078125,
0.7998046875,
0.330078125
],
[
0.273681640625,
0.326904296875,
0.314453125,
0.326904296875,
0.314453125,
0.341552734375,
0.273681640625,
0.341552734375
],
[
0.12091064453125,
0.317626953125,
0.1871337890625,
0.317626953125,
0.1871337890625,
0.328857421875,
0.12091064453125,
0.328857421875
],
[
0.378173828125,
0.326904296875,
0.419189453125,
0.326904296875,
0.419189453125,
0.341552734375,
0.378173828125,
0.341552734375
],
[
0.5888671875,
0.326904296875,
0.6298828125,
0.326904296875,
0.6298828125,
0.341552734375,
0.5888671875,
0.341552734375
],
[
0.69384765625,
0.326904296875,
0.734375,
0.326904296875,
0.734375,
0.341552734375,
0.69384765625,
0.341552734375
],
[
0.80224609375,
0.32763671875,
0.8408203125,
0.32763671875,
0.8408203125,
0.341552734375,
0.80224609375,
0.341552734375
],
[
0.48779296875,
0.32763671875,
0.52294921875,
0.32763671875,
0.52294921875,
0.34033203125,
0.48779296875,
0.34033203125
],
[
0.12091064453125,
0.328857421875,
0.1993408203125,
0.328857421875,
0.1993408203125,
0.34033203125,
0.12091064453125,
0.34033203125
],
[
0.08087158203125,
0.36865234375,
0.2900390625,
0.36865234375,
0.2900390625,
0.38330078125,
0.08087158203125,
0.38330078125
],
[
0.6044921875,
0.367431640625,
0.6259765625,
0.367431640625,
0.6259765625,
0.380126953125,
0.6044921875,
0.380126953125
],
[
0.77783203125,
0.36865234375,
0.80126953125,
0.36865234375,
0.80126953125,
0.380126953125,
0.77783203125,
0.380126953125
],
[
0.1004638671875,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.1004638671875,
0.4072265625
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.08087158203125,
0.425537109375,
0.46728515625,
0.425537109375,
0.46728515625,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.080078125,
0.439453125,
0.468017578125,
0.439453125,
0.468017578125,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.07843017578125,
0.45263671875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.07843017578125,
0.466552734375
],
[
0.080078125,
0.468505859375,
0.4697265625,
0.468505859375,
0.4697265625,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.6005859375,
0.479248046875,
0.63134765625,
0.479248046875,
0.63134765625,
0.4873046875,
0.6005859375,
0.4873046875
],
[
0.76318359375,
0.479736328125,
0.814453125,
0.479736328125,
0.814453125,
0.4873046875,
0.76318359375,
0.4873046875
],
[
0.080078125,
0.5,
0.443603515625,
0.5,
0.443603515625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.080078125,
0.484130859375,
0.4697265625,
0.484130859375,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.1004638671875,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.52978515625,
0.1004638671875,
0.52978515625
],
[
0.080078125,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.080078125,
0.546875,
0.468017578125,
0.546875,
0.468017578125,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.080078125,
0.56103515625,
0.4697265625,
0.56103515625,
0.4697265625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.080078125,
0.5771484375,
0.46728515625,
0.5771484375,
0.46728515625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50244140625,
0.5986328125,
0.8896484375,
0.5986328125,
0.8896484375,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.5009765625,
0.611328125,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.5009765625,
0.625
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50244140625,
0.626953125,
0.78515625,
0.626953125,
0.78515625,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.63623046875,
0.253173828125,
0.63623046875,
0.253173828125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.1004638671875,
0.6513671875,
0.46728515625,
0.6513671875,
0.46728515625,
0.66552734375,
0.1004638671875,
0.66552734375
],
[
0.080078125,
0.66845703125,
0.46728515625,
0.66845703125,
0.46728515625,
0.681640625,
0.080078125,
0.681640625
],
[
0.5009765625,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.6845703125,
0.5009765625,
0.6845703125
],
[
0.080078125,
0.68310546875,
0.471435546875,
0.68310546875,
0.471435546875,
0.6962890625,
0.080078125,
0.6962890625
],
[
0.5009765625,
0.68603515625,
0.890625,
0.68505859375,
0.890625,
0.69970703125,
0.5009765625,
0.70068359375
],
[
0.08087158203125,
0.69775390625,
0.4697265625,
0.69775390625,
0.4697265625,
0.7119140625,
0.08087158203125,
0.7119140625
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.080078125,
0.71337890625,
0.4697265625,
0.71337890625,
0.4697265625,
0.7265625,
0.080078125,
0.7265625
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.72900390625,
0.4697265625,
0.72900390625,
0.4697265625,
0.7421875,
0.080078125,
0.7421875
],
[
0.50341796875,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.50341796875,
0.74560546875
],
[
0.080078125,
0.74365234375,
0.4697265625,
0.74365234375,
0.4697265625,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.50244140625,
0.74755859375,
0.89208984375,
0.74755859375,
0.89208984375,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.75830078125,
0.465576171875,
0.75830078125,
0.465576171875,
0.771484375,
0.080078125,
0.771484375
],
[
0.50244140625,
0.76318359375,
0.88134765625,
0.76318359375,
0.88134765625,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.52294921875,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.52294921875,
0.79443359375
],
[
0.08251953125,
0.78857421875,
0.1690673828125,
0.78857421875,
0.1690673828125,
0.7998046875,
0.08251953125,
0.7998046875
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.098876953125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Table 2. Performance comparison of the IQA methods on three standard IQA databases, i.e., LIVE [42], CSIQ[32], and TID2013 [38], in",
"terms of SRCC and KRCC. The top three performing methods are highlighted in bold face. Some results are borrowed from [16, 21].",
"LIVE[42]",
"CSIQ[32]",
"TID2013[38]",
"Method",
"SRCC",
"SRCC",
"KRCC",
"KRCC",
"KRCC",
"SRCC",
"PSNR",
"0.873",
"0.680",
"0.810",
"0.601",
"0.687",
"0.496",
"0.948",
"0.796",
"0.865",
"0.680",
"0.727",
"0.545",
"0.805",
"SSIM [5]]",
"0.906",
"0.730",
"0.786",
"0.605",
"0.951",
"MS-SSIM [52]",
"0.952",
"0.806",
"0.943",
"0.786",
"0.897",
"0.718",
"0.967",
"VSI[58]",
"0.797",
"0.604",
"0.947",
"MAD[32]",
"0.842",
"0.781",
"0.964",
"0.828",
"0.743",
"0.518",
"0.677",
"VIF[40]",
"0.911",
"0.965",
"0.836",
"0.769",
"0.931",
"FSIMc[59]",
"0.851",
"0.667",
"NLPD [31]",
"0.937",
"0.778",
"0.932",
"0.769",
"0.800",
"0.625",
"0.960",
"0.827",
"0.950",
"0.804",
"0.804",
"0.634",
"0.947",
"GMSD[53]",
"0.791",
"0.909",
"0.732",
"0.831",
"0.631",
"0.876",
"WaDIQaM[3]",
"0.919",
"0.750",
"0.892",
"0.715",
"0.683",
"0.932",
"PieAPP[39]",
"0.876",
"0.689",
"0.670",
"0.497",
"0.765",
"LPIPS [6]]",
"DISTS[17]",
"0.954",
"0.811",
"0.929",
"0.767",
"0.830",
"0.639",
"SWD [21]",
"0.819",
"0.634",
"0.970",
"0.849",
"0.943",
"0.799",
"0.899",
"0.717",
"0.917",
"IQT (ours)",
"0.737",
"0.649",
"0.804",
"0.607",
"0.851",
"IQT-C (ours)",
"4.2.Implementation details",
"LIVE",
"CSIQ",
"We denote our model trained on the KADID-1Ok as IQT.",
"The hyper-parameters for the model are set as follow: i) the",
"number ofencoder anddecoderlayerissetto2(i.e.,L=2),",
"ii) the number of heads in the MHA is set to 4 (i.e., H = 4),",
"ii) the transformer dimension is set to 256 (i.e., D = 256),",
"iv) dimension of the MLP in the encoder and decoder is set",
"TID2013",
"PIPALvalidatic",
"FC layer in MLP head is set to 512 (i.e., Dhead = 512)",
"to 1024 (i.e., Dfeat = 1024), v) the dimension of the first",
"In the training phase, a given image is cropped to obtain",
"image patches. The dimension of the patch fed into the pro-",
"posed IQT is 256 × 256 × 3. The number of patches in the",
"feature map is set to N = 891. In the testing phase, image",
"patches are also acquired from the given image pair. We ex-",
"Figure 3. Scatter plots of ground-truth mean opinion scores",
"tract M overlapping patches and predict final quality score",
"(MOSs) against predicted scores of proposed IQT on LIVE, CSIQ",
"by averaging M individual quality scores of the patches.",
"TID2013, and PIPAL datasets.The predicted scores are obtained",
"The stride size is set as large as possible to cover the entire",
"from the model trained on KADID-1Ok dataset.",
"image withfewerpatches.",
"Data augmentation including horizontal flip and random",
"rotation is applied during the training. The training is con-",
"ter plots of the predicted scores of IQT and the ground-truth",
"ducted using an ADAM [28] optimizer with a batch size of",
"MOS are also presented in Fig. 3. For LIVE and TID2013",
"16. Initial learning rate 2 × 10-4 and cosine learning rate",
"databases, the proposed IQT shows the best performance",
"decay are set. The training loss is computed using a mean",
"in terms of SRCC. Also, it is ranked in the top three in",
"squared error (MSE) loss function. Our network is imple-",
"all benchmarks in terms of SRCC and KRCC. In particu-",
"mented using Tensorflow framework. It roughly takes a half",
"lar, our method shows better performance than recent deep",
"day with a single NVIDIA TITAN RTX to train our model.",
"learning-based methods [3, 39, 61, 16,21]for most cases.",
"Example images of the PIPAL validation dataset and fol-",
"4.3.Results",
"lowing PSNR, SSIM [51], MS-SSIM [52], LPIPS [6]], and",
"The proposed IQT shows that the transformer based",
"proposed IQT are illustrated in Fig. 4. From the left to the",
"model is sufficiently competitive compared to existing ap-",
"right, the perceptually better images to worse images are",
"proaches for the dataset that has traditional distortions. Our",
"listed based on MOS. Our proposed IQT predicts the qual-",
"model is trained on KADID-10k and, then, the performance",
"ity scores similar to MOS in terms of the superiority. There",
"on the three standard IQA datasets is evaluated. The perfor-",
"exist the images that are clearly distinguished by all meth-",
"mance comparison result is reported in Table 2 and the scat-",
"ods, however, it is difficult to accurately predict perceptual"
] |
{
"arxiv_id": "2104.14730",
"language": "en",
"page_idx": 4
}
|
||
2311.13028-en-0
|
2311.13028-en-0.jpg
|
DMLR: Data- centric Machine Learning Research
Past, Present and Future
Luis Oala$^1*$, Manil Maskey$^2$, Lilith Bat-Leah$^3$, Alicia Parrish$^4$, Nezihe Merve Gürel$^5$, Tzu-Sheng Kuo$^6$, Yang Liu$^7,8$, Rotem Dror$^9$, Danilo Brajovic$^10$, Xi- $\mathrm{aozhe~Yao}^{34}$, Max Bartolo$^11$, William Gaviria Rojas$^12$, Ryan Hileman$^13$, Rainier Aliment$^4$, Michael W. Mahoney$^14,15,16$, Meg Risdal$^17$, Matthew Lease$^18$, Wojciech Samek$^19,20$, Debo Dutta$^21$, Curtis Northcutt$^22$, Cody Coleman$^12$, Braden Hancoc$k^{23}$, Bernard Koc$h^{24}$, Girmaw Abebe Tadess$e^{25}$, Bojan Karla$\v {s}^{26}$, Ahmed $\mathrm{Alaa}^{16}$, Adji Bousso Dieng$^27$, Natasha Noy$^4$, Vijay Janapa Reddi$^26$, James Zou$^28$, $\mathrm{Praveen~Paritosh}^{29}$, Mihaela van der Schaar$^30$, Kurt Bollacker$^29$, Lora Aroyo$^4$, Ce Zhang$^31,24$, Joaquin Vanschoren$^32$, Isabelle Guyon$^4,33$, Peter Mattson$^{4,29}$
$^{1}Dotphoton, {}^{2}NASA, {}^{3}Mod$ $Op, {}^{4}Google, {}^{5}TU$ $Delft, {}^{6}Carnegie$ Mellon $University, {}^{7}UC$ Santa $Cruz$, $^{8}ByteDance$ $Research, \:^{9}University$ $of$ $Haifa, \:^{10}Fraunhofer$ $IPA, \:^{11}Cohere, \:^{12}CoactiveAI, \:^{13}Talon$, $^{14}UC$ $Berkeley, ^{15}ICSI, ^{16}LBNL, ^{17}Kaggle, ^{18}UT$ $Austin, ^{19}TU$ $Berlin, ^{20}Fraunhofer$ $HHI, ^{21}Nutanix$, $^{22}Cleanlab,^{23}SnorkelAI,^{24}UniversityofChicago,^{25}MicrosoftAIforGoodLab,^{26}HarvardUni-$ $versity,{}^{27}PrincetonUniversity,{}^{28}Stanford{}University,{}^{29}MLCommons,{}^{30}University{}of{}Cambridge$, $^{31}Together, ^{32}Eindhoven\textit{ University of Technology, }^{33}University\textit{ of Paris- Saclay, }^{34}ETH\textit{ Zurich}$
### $\mathbf{Abstract}$
Drawing from discussions at the inaugural DMLR workshop at ICML 2023 and meetings prior, in this report we outline the relevance of community engagement and infrastructure development for the creation of next-generation public datasets that will advance machine learning science. We chart a path forward as a collective effort to sustain the creation and maintenance of these datasets and methods towards positive scientific, societal and business impact.
Keywords: data-centric machine learning, artificial intelligence, datasets, impact
1 Data Ambivalence in Machine Learning
Why state the obvious? Do we really need to emphasize some machine learning (ML) research as data-$centric?$ Hasn't ML science, at its core, always been just that? After all, designing algorithms that extract models from data is machine learning's $summumbonum.$ In the pursuit of this goal we often oscillate between two dominant phases: (i) design algorithm and throw data at it, (ii) go back to data (and its intermediate representations) to design better algorithm. This feedback loop informs the ambivalence towards data that many of us will encounter in machine learning practice: on the one hand, we want the algorithm to extract a model from data automatically; on the other hand, we often need to analyze the data and model manually to build good algorithms. Through the lens of this oscillation, data-centric machine learning research (DMLR) can broadly be described
*. To get involved in the community please join the discord at https://discord.gg/FswYXMv4j9. For
updates to the manuscript you can contact luis.oala@dotphoton.com.
$\textcircled{\mathrm{C}}2023$ The DMLR Community
|
[
[
0.2156982421875,
0.1256103515625,
0.78369140625,
0.126953125,
0.78369140625,
0.14453125,
0.2156982421875,
0.143310546875
],
[
0.35546875,
0.173583984375,
0.6435546875,
0.173583984375,
0.6435546875,
0.1881103515625,
0.35546875,
0.1881103515625
],
[
0.146240234375,
0.216552734375,
0.85302734375,
0.216552734375,
0.85302734375,
0.2310791015625,
0.146240234375,
0.2310791015625
],
[
0.1478271484375,
0.2342529296875,
0.8505859375,
0.2342529296875,
0.8505859375,
0.248779296875,
0.1478271484375,
0.248779296875
],
[
0.146240234375,
0.250732421875,
0.85302734375,
0.250732421875,
0.85302734375,
0.26513671875,
0.146240234375,
0.26513671875
],
[
0.0228729248046875,
0.25830078125,
0.05963134765625,
0.25830078125,
0.05963134765625,
0.399658203125,
0.0228729248046875,
0.399658203125
],
[
0.1478271484375,
0.268310546875,
0.438720703125,
0.268310546875,
0.438720703125,
0.2822265625,
0.1478271484375,
0.2822265625
],
[
0.449462890625,
0.268310546875,
0.484375,
0.268310546875,
0.484375,
0.27587890625,
0.449462890625,
0.27587890625
],
[
0.48779296875,
0.266357421875,
0.85302734375,
0.268310546875,
0.85302734375,
0.28271484375,
0.48779296875,
0.281005859375
],
[
0.146240234375,
0.285400390625,
0.42236328125,
0.2841796875,
0.42236328125,
0.298583984375,
0.146240234375,
0.2998046875
],
[
0.4306640625,
0.285400390625,
0.85302734375,
0.285400390625,
0.85302734375,
0.2998046875,
0.4306640625,
0.2998046875
],
[
0.146240234375,
0.302978515625,
0.39453125,
0.302978515625,
0.39453125,
0.31640625,
0.146240234375,
0.31640625
],
[
0.39208984375,
0.3037109375,
0.47216796875,
0.3037109375,
0.47216796875,
0.314453125,
0.39208984375,
0.314453125
],
[
0.46728515625,
0.302978515625,
0.8544921875,
0.302978515625,
0.8544921875,
0.31640625,
0.46728515625,
0.31640625
],
[
0.148681640625,
0.3193359375,
0.3896484375,
0.3193359375,
0.3896484375,
0.333251953125,
0.148681640625,
0.333251953125
],
[
0.386474609375,
0.318115234375,
0.84619140625,
0.318115234375,
0.84619140625,
0.333251953125,
0.386474609375,
0.333251953125
],
[
0.1478271484375,
0.337158203125,
0.55810546875,
0.337158203125,
0.55810546875,
0.35107421875,
0.1478271484375,
0.35107421875
],
[
0.55224609375,
0.335205078125,
0.85302734375,
0.3359375,
0.85302734375,
0.35107421875,
0.55224609375,
0.349853515625
],
[
0.1478271484375,
0.352783203125,
0.80615234375,
0.352783203125,
0.80615234375,
0.3662109375,
0.1478271484375,
0.3662109375
],
[
0.148681640625,
0.371826171875,
0.85302734375,
0.371826171875,
0.85302734375,
0.386474609375,
0.148681640625,
0.386474609375
],
[
0.1478271484375,
0.386962890625,
0.85107421875,
0.386962890625,
0.85107421875,
0.401611328125,
0.1478271484375,
0.401611328125
],
[
0.021240234375,
0.40478515625,
0.056365966796875,
0.404052734375,
0.060455322265625,
0.4873046875,
0.02532958984375,
0.488037109375
],
[
0.1446533203125,
0.40283203125,
0.86279296875,
0.404052734375,
0.86279296875,
0.421630859375,
0.1446533203125,
0.42041015625
],
[
0.143798828125,
0.419189453125,
0.8544921875,
0.42041015625,
0.8544921875,
0.438232421875,
0.143798828125,
0.4375
],
[
0.1446533203125,
0.438232421875,
0.8544921875,
0.438232421875,
0.8544921875,
0.455810546875,
0.1446533203125,
0.455810546875
],
[
0.143798828125,
0.453857421875,
0.83251953125,
0.455078125,
0.83251953125,
0.472900390625,
0.143798828125,
0.4716796875
],
[
0.0228729248046875,
0.49560546875,
0.0555419921875,
0.49560546875,
0.058013916015625,
0.71044921875,
0.02532958984375,
0.71044921875
],
[
0.45751953125,
0.50244140625,
0.54345703125,
0.50244140625,
0.54345703125,
0.517578125,
0.45751953125,
0.517578125
],
[
0.17724609375,
0.5234375,
0.8203125,
0.5234375,
0.8203125,
0.537109375,
0.17724609375,
0.537109375
],
[
0.17724609375,
0.5390625,
0.82177734375,
0.5390625,
0.82177734375,
0.55322265625,
0.17724609375,
0.55322265625
],
[
0.17724609375,
0.5537109375,
0.82177734375,
0.5537109375,
0.82177734375,
0.5673828125,
0.17724609375,
0.5673828125
],
[
0.17724609375,
0.5693359375,
0.82177734375,
0.5693359375,
0.82177734375,
0.58349609375,
0.17724609375,
0.58349609375
],
[
0.17724609375,
0.58544921875,
0.82177734375,
0.58544921875,
0.82177734375,
0.5986328125,
0.17724609375,
0.5986328125
],
[
0.17724609375,
0.599609375,
0.29736328125,
0.599609375,
0.29736328125,
0.61279296875,
0.17724609375,
0.61279296875
],
[
0.176513671875,
0.6142578125,
0.77294921875,
0.61572265625,
0.77294921875,
0.6298828125,
0.176513671875,
0.62890625
],
[
0.143798828125,
0.6513671875,
0.5498046875,
0.65283203125,
0.5498046875,
0.67041015625,
0.143798828125,
0.66943359375
],
[
0.146240234375,
0.68115234375,
0.8505859375,
0.68115234375,
0.8505859375,
0.69580078125,
0.146240234375,
0.69580078125
],
[
0.146240234375,
0.69873046875,
0.85302734375,
0.69873046875,
0.85302734375,
0.71337890625,
0.146240234375,
0.71337890625
],
[
0.146240234375,
0.71630859375,
0.85302734375,
0.71630859375,
0.85302734375,
0.72998046875,
0.146240234375,
0.72998046875
],
[
0.146240234375,
0.73291015625,
0.8505859375,
0.73291015625,
0.8505859375,
0.7470703125,
0.146240234375,
0.7470703125
],
[
0.146240234375,
0.75,
0.85302734375,
0.75,
0.85302734375,
0.7646484375,
0.146240234375,
0.7646484375
],
[
0.146240234375,
0.767578125,
0.50634765625,
0.767578125,
0.50634765625,
0.78076171875,
0.146240234375,
0.78076171875
],
[
0.580078125,
0.7666015625,
0.85107421875,
0.7666015625,
0.85107421875,
0.77978515625,
0.580078125,
0.77978515625
],
[
0.51953125,
0.77294921875,
0.55224609375,
0.77294921875,
0.55224609375,
0.7763671875,
0.51953125,
0.7763671875
],
[
0.1478271484375,
0.7861328125,
0.371826171875,
0.7861328125,
0.371826171875,
0.796875,
0.1478271484375,
0.796875
],
[
0.406005859375,
0.7890625,
0.439453125,
0.7890625,
0.439453125,
0.79345703125,
0.406005859375,
0.79345703125
],
[
0.6396484375,
0.7890625,
0.6650390625,
0.7890625,
0.6650390625,
0.79345703125,
0.6396484375,
0.79345703125
],
[
0.66259765625,
0.78515625,
0.85107421875,
0.78515625,
0.85107421875,
0.79541015625,
0.66259765625,
0.79541015625
],
[
0.1478271484375,
0.8017578125,
0.404296875,
0.8017578125,
0.404296875,
0.81494140625,
0.1478271484375,
0.81494140625
],
[
0.640625,
0.8017578125,
0.85302734375,
0.8017578125,
0.85302734375,
0.81494140625,
0.640625,
0.81494140625
],
[
0.1478271484375,
0.82080078125,
0.343017578125,
0.82080078125,
0.343017578125,
0.83056640625,
0.1478271484375,
0.83056640625
],
[
0.71337890625,
0.82470703125,
0.72216796875,
0.82470703125,
0.72216796875,
0.82763671875,
0.71337890625,
0.82763671875
],
[
0.74169921875,
0.82080078125,
0.85302734375,
0.82080078125,
0.85302734375,
0.83056640625,
0.74169921875,
0.83056640625
],
[
0.146240234375,
0.8359375,
0.8544921875,
0.8359375,
0.8544921875,
0.849609375,
0.146240234375,
0.849609375
],
[
0.15283203125,
0.8623046875,
0.8544921875,
0.8623046875,
0.8544921875,
0.876953125,
0.15283203125,
0.876953125
],
[
0.1683349609375,
0.87744140625,
0.640625,
0.87548828125,
0.640625,
0.8896484375,
0.1683349609375,
0.8916015625
],
[
0.146240234375,
0.91552734375,
0.343017578125,
0.91552734375,
0.343017578125,
0.9287109375,
0.146240234375,
0.9287109375
]
] |
[
"DMLR: Data-centric Machine Learning Research",
"Past. Present and Future",
"Luis Oala'* Manil Maskey?, Lilith Bat-Leah', Alicia Parrish4, Nezihe Merve",
"Girel, Tzu-Sheng Kuo, Yang Liu.8, Rotem Dror', Danilo Brajovicio, Xi-",
"aozhe Yao34, Max Bartolo11, William Gaviria Rojas12, Ryan Hileman13, Rainier",
"Z ",
"Aliment4, Michael W. Mahoney",
"4.15.16",
".Meg Risdal17,Matthew Lease18,Woj-",
"ciech Samek19,20, Debo Dutta2",
" Curtis Northcutt22, Cody Coleman12, Braden",
"Hancock23, Bernard Koch24",
"Girmaw",
"V Abebe Tadesse25, Bojan Karlas26, Ahmed",
"Alaa16, Adji Bousso Dieng27",
"7, Natasha Noy4, Vijay Janapa Reddi26, James Zou28",
"Praveen Paritosh29, Mihaela van der Schaar30",
"0, Kurt Bollacker29, Lora Aroyo,",
"Ce Zhang31,24, Joaguin Vanschoren32, Isabelle Guyon4,33, Peter Mattson4,2",
"1 Dotphoton, 2 NASA, 3 Mod Op, 4Google, 5 TU Delft, Carnegie Mellon University, 7 UC Santa Cruz,",
"8 ByteDance Research, 9 University of Haifa. 10 Fraunhofer IPA, 11 Cohere, 12 CoactiveAI, 13 Talon.",
"[DT'S]",
"14 UC Berkeley, 15ICSI, 16LBNL, 17 Kaggle, 18 UT Austin, 19 TU Berlin, 20 Fraunhofer HHI, 21 Nutanic,",
"22 Cleanlab, 23 Snorkel AI, 24 University of Chicago, 25 Microsoft AI for Good Lab, 26 Harvard Uni-",
"versity, 27 Princeton University, 28 Stanford University, 29 MLCommons, 30 University of Cambridge,",
"31 Together, 32 Eindhoven University of Technology, 33 University of Paris-Saclay, 34ETH Zurich",
"",
"Abstract",
"Drawing from discussions at the inaugural DMLR workshop at ICML 2023 and meetings",
"prior, in this report we outline the relevance of community engagement and infrastructure",
"development for the creation of next-generation public datasets that will advance machine",
"learning science. We chart a path forward as a collective effort to sustain the creation",
"and maintenance of these datasets and methods towards positive scientific, societal and",
"business impact.",
"Keywords:data-centric machine learning, artificial intelligence, datasets, impact",
"1 Data Ambivalence in Machine Learning",
"Why state the obvious? Do we really need to emphasize some machine learning (ML) re-",
"search as data-centric? Hasn't ML science, at its core, always been just that? After all,",
"designing algorithms that extract models from data is machine learning's summum bonum.",
"In the pursuit of this goal we often oscillate between two dominant phases: (i)design al",
"gorithm and throw data at it, (ii) go back to data (and its intermediate representations)",
"to design better algorithm. This feedback loop",
"the ambivalence towards data that",
"",
"many of uswill encounter i",
"a001",
"ae",
"one hand.we wantthe",
"algorithm to extract a model fro",
"other hand, we often need",
"to analvzethe dataand",
"",
"ghthe lens of",
"this oscillation, data-centric machine learning research (DMLR) can broadly be described",
"*. To get involved in the community please join the discord at https://discord.gg/FswYxMv4j9. For",
"updates to the manuscript you can contact luis.oaladotphoton.com.",
"2023 The DMLR Community."
] |
{
"arxiv_id": "2311.13028",
"language": "en",
"page_idx": 0
}
|
||
2311.13028-en-1
|
2311.13028-en-1.jpg
|
THE DMLR COMMUNITY
<FigureHere>
Figure 1: A timeline of some inflection points in the development of data-centric ideas.
as infrastructure, methods and communities revolving around phase (ii). In this editorial we outline key coordinates and objectives of DMLR, contextualize its origins, and summarize activities that took place at the inaugural DMLR workshop. And these lines are an invitation, a call on you, the reader, to join us in shaping the DMLR future. Be it as open source contributor, community organizer, researcher or reviewer, your ideas and efforts are needed to maintain and shape the DMLR ecosystem further.
2 Past: Data-Centricity Over Time
Historically, the ambivalence towards data has manifested in different ways. In the early 1990s, Wilson, Garris and Wilkinson [1-4] distributed“$Handwriting\textit{ Sampling Forms" at}$ the National Institute of Standards and Technology (NIST), digitizing the resulting data into the raw ingredients that were later turned into the now infamous machine learning staple MNIST. But as of October 4,2023, their original publications have less than 150 citations combined. In comparison, the seminal LeNet paper by [5], which is often used as stand-in reference for the MNIST dataset, sits at 60,000 citations today.$^1.$
This is not to open artificial fault lines à la“$data\textit{people” versus“ }model\textit{people" . But}$ one can wonder what such artifacts reveal about the incentives in machine learning and how conducive they are to machine learning progress. Or whether, as [7] suggest, it slows progress because “everyone wants to do the model work, not the data work”. Jumping to today, there are active and encouraging efforts in the community to counter this imbalance, most prominently the $Datasetsand\textit{Benchmark Track at Neural Information Processing}$ System (NeurIPS) that was conceived for the first time in 2021 by Joaquin Vanschoren,
1. As an aside, LeCun et al. [5] themselves did not cite the NIST prior works [1-4]. Notably, Yadav and
Bottou [6] later revisited the history of MNIST.
2
|
[
[
0.40185546875,
0.048614501953125,
0.59326171875,
0.048614501953125,
0.59326171875,
0.0625,
0.40185546875,
0.0625
],
[
0.1568603515625,
0.121826171875,
0.221435546875,
0.121826171875,
0.221435546875,
0.136962890625,
0.1568603515625,
0.136962890625
],
[
0.57421875,
0.12371826171875,
0.58642578125,
0.12371826171875,
0.58642578125,
0.126953125,
0.57421875,
0.126953125
],
[
0.6337890625,
0.12371826171875,
0.701171875,
0.12371826171875,
0.701171875,
0.1275634765625,
0.6337890625,
0.1275634765625
],
[
0.56884765625,
0.1337890625,
0.6044921875,
0.1331787109375,
0.6044921875,
0.141357421875,
0.56884765625,
0.1427001953125
],
[
0.63232421875,
0.1300048828125,
0.72802734375,
0.1300048828125,
0.72802734375,
0.136962890625,
0.63232421875,
0.136962890625
],
[
0.156005859375,
0.1383056640625,
0.25,
0.1383056640625,
0.25,
0.1558837890625,
0.156005859375,
0.1558837890625
],
[
0.63232421875,
0.1395263671875,
0.63818359375,
0.1395263671875,
0.63818359375,
0.14453125,
0.63232421875,
0.14453125
],
[
0.404296875,
0.15283203125,
0.435546875,
0.1546630859375,
0.435546875,
0.1622314453125,
0.40283203125,
0.160400390625
],
[
0.1568603515625,
0.160400390625,
0.26953125,
0.160400390625,
0.26953125,
0.1749267578125,
0.1568603515625,
0.1749267578125
],
[
0.56640625,
0.1546630859375,
0.685546875,
0.1546630859375,
0.685546875,
0.1622314453125,
0.56640625,
0.1622314453125
],
[
0.453369140625,
0.1634521484375,
0.52294921875,
0.1634521484375,
0.52294921875,
0.17041015625,
0.453369140625,
0.17041015625
],
[
0.56201171875,
0.1641845703125,
0.5908203125,
0.1572265625,
0.59619140625,
0.169189453125,
0.56787109375,
0.1767578125
],
[
0.59228515625,
0.1654052734375,
0.6298828125,
0.1654052734375,
0.6298828125,
0.169189453125,
0.59228515625,
0.169189453125
],
[
0.63623046875,
0.1634521484375,
0.69775390625,
0.1622314453125,
0.69775390625,
0.169189453125,
0.63623046875,
0.17041015625
],
[
0.7197265625,
0.1673583984375,
0.80810546875,
0.1673583984375,
0.80810546875,
0.1749267578125,
0.7197265625,
0.1749267578125
],
[
0.71728515625,
0.1749267578125,
0.7490234375,
0.175537109375,
0.7490234375,
0.18310546875,
0.71728515625,
0.1817626953125
],
[
0.80810546875,
0.169189453125,
0.83837890625,
0.169189453125,
0.83837890625,
0.173583984375,
0.80810546875,
0.173583984375
],
[
0.158447265625,
0.1787109375,
0.26708984375,
0.1787109375,
0.26708984375,
0.1932373046875,
0.158447265625,
0.1932373046875
],
[
0.7861328125,
0.177978515625,
0.8095703125,
0.177978515625,
0.8095703125,
0.1817626953125,
0.7861328125,
0.1817626953125
],
[
0.476318359375,
0.186279296875,
0.54345703125,
0.186279296875,
0.54345703125,
0.1900634765625,
0.476318359375,
0.1900634765625
],
[
0.71826171875,
0.189453125,
0.744140625,
0.189453125,
0.744140625,
0.1962890625,
0.71826171875,
0.1962890625
],
[
0.72802734375,
0.184326171875,
0.82421875,
0.1817626953125,
0.82421875,
0.189453125,
0.72802734375,
0.1912841796875
],
[
0.4697265625,
0.1912841796875,
0.52783203125,
0.1912841796875,
0.52783203125,
0.1988525390625,
0.4697265625,
0.1988525390625
],
[
0.373291015625,
0.1988525390625,
0.423095703125,
0.1988525390625,
0.423095703125,
0.20263671875,
0.373291015625,
0.20263671875
],
[
0.52685546875,
0.1932373046875,
0.55810546875,
0.1932373046875,
0.55810546875,
0.1976318359375,
0.52685546875,
0.1976318359375
],
[
0.365234375,
0.205810546875,
0.439453125,
0.205810546875,
0.439453125,
0.2171630859375,
0.365234375,
0.2171630859375
],
[
0.476318359375,
0.21337890625,
0.53369140625,
0.21337890625,
0.53369140625,
0.2203369140625,
0.476318359375,
0.2203369140625
],
[
0.63134765625,
0.2083740234375,
0.68310546875,
0.2083740234375,
0.68310546875,
0.2183837890625,
0.63134765625,
0.2183837890625
],
[
0.465576171875,
0.222900390625,
0.54345703125,
0.222900390625,
0.54345703125,
0.2266845703125,
0.465576171875,
0.2266845703125
],
[
0.619140625,
0.231689453125,
0.63818359375,
0.231689453125,
0.63818359375,
0.236083984375,
0.619140625,
0.236083984375
],
[
0.16259765625,
0.2392578125,
0.253173828125,
0.2392578125,
0.253173828125,
0.250732421875,
0.16259765625,
0.250732421875
],
[
0.5908203125,
0.240478515625,
0.61669921875,
0.240478515625,
0.61669921875,
0.2442626953125,
0.5908203125,
0.2442626953125
],
[
0.67333984375,
0.2392578125,
0.69677734375,
0.2392578125,
0.69677734375,
0.246826171875,
0.67333984375,
0.246826171875
],
[
0.5908203125,
0.248779296875,
0.60888671875,
0.248779296875,
0.60888671875,
0.25244140625,
0.5908203125,
0.25244140625
],
[
0.6298828125,
0.248779296875,
0.640625,
0.248779296875,
0.640625,
0.251953125,
0.6298828125,
0.251953125
],
[
0.6748046875,
0.246826171875,
0.73876953125,
0.246826171875,
0.73876953125,
0.25390625,
0.6748046875,
0.25390625
],
[
0.226318359375,
0.274658203125,
0.2646484375,
0.274658203125,
0.2646484375,
0.28857421875,
0.226318359375,
0.28857421875
],
[
0.16259765625,
0.250732421875,
0.275390625,
0.250732421875,
0.275390625,
0.25830078125,
0.16259765625,
0.25830078125
],
[
0.223876953125,
0.285888671875,
0.293212890625,
0.285400390625,
0.293212890625,
0.296630859375,
0.223876953125,
0.29736328125
],
[
0.564453125,
0.294189453125,
0.5703125,
0.294189453125,
0.5703125,
0.3037109375,
0.564453125,
0.3037109375
],
[
0.5703125,
0.290283203125,
0.57666015625,
0.290283203125,
0.57666015625,
0.296630859375,
0.5703125,
0.296630859375
],
[
0.55615234375,
0.3017578125,
0.56640625,
0.3017578125,
0.56640625,
0.3212890625,
0.55615234375,
0.3212890625
],
[
0.336669921875,
0.3193359375,
0.400390625,
0.3193359375,
0.400390625,
0.330810546875,
0.336669921875,
0.330810546875
],
[
0.336669921875,
0.330078125,
0.416748046875,
0.330078125,
0.416748046875,
0.34033203125,
0.336669921875,
0.34033203125
],
[
0.44775390625,
0.367431640625,
0.47802734375,
0.367431640625,
0.47802734375,
0.375732421875,
0.44775390625,
0.375732421875
],
[
0.44775390625,
0.376953125,
0.55224609375,
0.37744140625,
0.55224609375,
0.38818359375,
0.44775390625,
0.386962890625
],
[
0.16015625,
0.415283203125,
0.83837890625,
0.415283203125,
0.83837890625,
0.4287109375,
0.16015625,
0.4287109375
],
[
0.1446533203125,
0.460205078125,
0.85302734375,
0.460205078125,
0.85302734375,
0.473388671875,
0.1446533203125,
0.473388671875
],
[
0.146240234375,
0.4765625,
0.8505859375,
0.4765625,
0.8505859375,
0.4912109375,
0.146240234375,
0.4912109375
],
[
0.1446533203125,
0.49365234375,
0.8544921875,
0.49365234375,
0.8544921875,
0.50830078125,
0.1446533203125,
0.50830078125
],
[
0.143798828125,
0.51123046875,
0.8544921875,
0.51220703125,
0.8544921875,
0.5263671875,
0.143798828125,
0.52587890625
],
[
0.1446533203125,
0.52880859375,
0.8544921875,
0.52880859375,
0.8544921875,
0.54345703125,
0.1446533203125,
0.54345703125
],
[
0.1446533203125,
0.54541015625,
0.62353515625,
0.54541015625,
0.62353515625,
0.56005859375,
0.1446533203125,
0.56005859375
],
[
0.146240234375,
0.58544921875,
0.488525390625,
0.58544921875,
0.488525390625,
0.5986328125,
0.146240234375,
0.5986328125
],
[
0.146240234375,
0.6123046875,
0.85302734375,
0.6123046875,
0.85302734375,
0.62548828125,
0.146240234375,
0.62548828125
],
[
0.1446533203125,
0.62744140625,
0.8544921875,
0.62890625,
0.8544921875,
0.64306640625,
0.1446533203125,
0.642578125
],
[
0.146240234375,
0.646484375,
0.8544921875,
0.646484375,
0.8544921875,
0.66015625,
0.146240234375,
0.66015625
],
[
0.143798828125,
0.6611328125,
0.8544921875,
0.66357421875,
0.8544921875,
0.67822265625,
0.143798828125,
0.67529296875
],
[
0.1446533203125,
0.6796875,
0.8544921875,
0.6796875,
0.8544921875,
0.6943359375,
0.1446533203125,
0.6943359375
],
[
0.1446533203125,
0.69775390625,
0.8544921875,
0.69775390625,
0.8544921875,
0.7119140625,
0.1446533203125,
0.7119140625
],
[
0.1446533203125,
0.7138671875,
0.72119140625,
0.7138671875,
0.72119140625,
0.72802734375,
0.1446533203125,
0.72802734375
],
[
0.1732177734375,
0.73193359375,
0.85302734375,
0.73193359375,
0.85302734375,
0.7470703125,
0.1732177734375,
0.7470703125
],
[
0.1446533203125,
0.75,
0.85302734375,
0.75,
0.85302734375,
0.76318359375,
0.1446533203125,
0.76318359375
],
[
0.1446533203125,
0.7666015625,
0.85302734375,
0.7666015625,
0.85302734375,
0.77978515625,
0.1446533203125,
0.77978515625
],
[
0.1446533203125,
0.78271484375,
0.85302734375,
0.78271484375,
0.85302734375,
0.796875,
0.1446533203125,
0.796875
],
[
0.146240234375,
0.8017578125,
0.85302734375,
0.8017578125,
0.85302734375,
0.81494140625,
0.146240234375,
0.81494140625
],
[
0.1446533203125,
0.8173828125,
0.8544921875,
0.818359375,
0.8544921875,
0.83349609375,
0.1446533203125,
0.83203125
],
[
0.146240234375,
0.83544921875,
0.85107421875,
0.83544921875,
0.85107421875,
0.8486328125,
0.146240234375,
0.8486328125
],
[
0.148681640625,
0.86279296875,
0.85546875,
0.86279296875,
0.85546875,
0.876953125,
0.148681640625,
0.876953125
],
[
0.1690673828125,
0.876953125,
0.488525390625,
0.876953125,
0.488525390625,
0.89013671875,
0.1690673828125,
0.89013671875
],
[
0.491943359375,
0.91357421875,
0.50732421875,
0.91357421875,
0.50732421875,
0.92822265625,
0.491943359375,
0.92822265625
]
] |
[
"THE DMLR COMMUNITY",
"DATA-",
"in",
"iretDMlRworkehe",
"etal.2019",
"atICMLtounifycommunity",
"CENTRIC",
"O",
"n MacKay",
"CONCEPTS",
"EleutherAi, LAION, and other grou",
"yGebruetal.(2018)",
"compile",
"andTelea",
"newgenerations",
"NurturingtheDMLRecos",
"build the",
"stem.te",
"OVERTIME",
"turean",
"AndrewNaandoth",
"dataset",
"mmunityfornextgenpublic",
"launchdatacentt",
"entricn.compu",
"Ticai.ora",
"Bretcartet at (1993)",
"NeuriPSD&Btra",
"netal.(2021)",
"launchesfnrfirsttime",
"tri",
"1990sand2000s",
"vorkshe",
"DMLR",
"N",
"2",
"intheJMLRfamily",
"2010s",
"Systemsandactivelearning",
"Data scrutiny",
"",
"",
"",
"Early2020s",
"Data-centric ML",
"Future",
"TheDMLRecosystem",
"Figure l: A timeline of some inflection points in the development of data-centric ideas.",
"as infrastructure, methods and communities revolving around phase (ii). In this editorial",
"we outline key coordinates and objectives of DMLR, contextualize its origins, and summa-",
"rize activities that took place at the inaugural DMLR workshop. And these lines are an",
"invitation, a call on you, the reader, to join us in shaping the DMLR future. Be it as open",
"source contributor, community organizer, researcher or reviewer, your ideas and efforts are",
"needed to maintain and shape the DMLR ecosystem further.",
"2 Past: Data-Centricity Over Time",
"Historically, the ambivalence towards data has manifested in different ways. In the early",
"1990s,Wilson, Garris and Wilkinson [1-4] distributed “Handwriting Sampling Forms\" at",
"the National Institute of Standards and Technology (NIST), digitizing the resulting data",
"into the raw ingredients that were later turned into the now infamous machine learning",
"staple MNIST. But as of October 4, 2023, their original publications have less than 150",
"citations combined. In comparison, the seminal LeNet paper by [5], which is often used as",
"stand-in reference for the MNIST dataset, sits at 60.000 citations today.1",
"This is not to open artificial fault lines a la “\"data people\" versus “\"model people\". But",
"one can wonder what such artifacts reveal about the incentives in machine learning and",
"how conducive they are to machine learning progress. Or whether, as [7] suggest, it slows",
"progress because “everyone wants to do the model work, not the data work\". Jumping to",
"today, there are active and encouraging efforts in the community to counter this imbalance,",
"most prominently the Datasets and Benchmark Track at Neural Information Processing",
"System (NeuriPS)that was conceived for the first time in 2021 by Joaquin Vanschoren.",
"1. As an aside, LeCun et al. [5] themselves did not cite the NIST prior works [1-4]. Notably, Yadav and",
"Bottou [6] later revisited the history of MNIST.",
"2"
] |
{
"arxiv_id": "2311.13028",
"language": "en",
"page_idx": 1
}
|
||
2311.13028-en-2
|
2311.13028-en-2.jpg
|
DMLR: PAST, PRESENT AND FUTURE
Serena Yeung, Maria Xenochristou. We have to emphasize that not all data work goes under-appreciated. One must only look at ImageNet [8] or CIFAR [9] for great success
stories.
Around the same time as MNIS literally in early works by [10] and others. It was likely discussed in the systems and database circles long before the idea became an increasingly growing focus of research in the core machine learning co evident in venues such as MLSys$^2$ or the DEEM workshops$^3$, due the high importance of optimized infrastructure to orchestrate and execute data transformation and machine learning workloads. Success stories can be found in frameworks to build and store models such as Torch [11], Theano [12], Caffe [13], TensorFlow [14], JAX [15] or PyTorch [16]. Sometimes these frameworks also led to optimized data formats, such as TensorFlow's TFRecord. Additionally, platforms like Kaggle, HuggingFace or OpenML have emerged as de-facto community data hubs and standardized data loading infrastructure. A new wave of emerging open-source projects such as Lance$^4$ aim to address existing gaps with respect to data loading needs. However, despite these advances, challenges regarding the compatibility, mutability, and collaborabil 'Croissant'$^5$, take stabs in an effort to streamline data-centric machine learning infrastructure. Similarly, DataPerf [17] is a recent community-led benchmark suite for evaluating ML datasets and data-centric algorithms, enabling the ML community to iterate on datasets, instead of just architectures.
Alongside developments in infrastructure, over time we have also witnessed critical advances in the way datasets are collected, curated and maintained. From the beginnings of modern statistical science [18;19], active learning, a set of methods concerned with data curation, has planted its roots firmly in the machine learning and statistical learning literature [20-26]. The next generation of machine learning datasets are characterized by dense metadata annotation $[27–29]$, collaborative refinement $[30–33]$, user preference and human feedback [34], and e for computer programs already today. Data provenance and ownership are also receiving increasing consideration by groups such as the“Data Trusts” initiative [38] and others. Our goal is to support the growth of the DMLR ecosystem into a strong community with effective infrastructure that will advance machine learning science through next generation datasets. These datasets shall serve as a bridge to connect fundamental problems (such as food insecurity and climate impact) with fundamental ML research by providing the right datasets for the right problems.
3 Present: Convening the Community at ICML 2023
In order to charge this effo 2023, in Honolulu, Hawaii, for the inaugural DMLR workshop at the International Conference on Machine Learning (ICML) 2023. The DMLR workshop was a point of convergence
2. https://mlsys.org/
3. http://deem-workshop.org/
4. https://github.com/lancedb/lance
5. https://github.com/mlcommons/croissant
3
|
[
[
0.35302734375,
0.048614501953125,
0.640625,
0.048614501953125,
0.640625,
0.0625,
0.35302734375,
0.0625
],
[
0.1446533203125,
0.11614990234375,
0.85107421875,
0.11614990234375,
0.85107421875,
0.1307373046875,
0.1446533203125,
0.1307373046875
],
[
0.146240234375,
0.1331787109375,
0.85107421875,
0.1331787109375,
0.85107421875,
0.147705078125,
0.146240234375,
0.147705078125
],
[
0.1446533203125,
0.15283203125,
0.2017822265625,
0.15283203125,
0.2017822265625,
0.1641845703125,
0.1446533203125,
0.1641845703125
],
[
0.17236328125,
0.1666259765625,
0.8544921875,
0.1673583984375,
0.8544921875,
0.1849365234375,
0.17236328125,
0.184326171875
],
[
0.1446533203125,
0.1849365234375,
0.8544921875,
0.186279296875,
0.8544921875,
0.2008056640625,
0.1446533203125,
0.199462890625
],
[
0.146240234375,
0.203857421875,
0.8544921875,
0.203857421875,
0.8544921875,
0.2183837890625,
0.146240234375,
0.2183837890625
],
[
0.1446533203125,
0.2203369140625,
0.85302734375,
0.2203369140625,
0.85302734375,
0.23486328125,
0.1446533203125,
0.23486328125
],
[
0.1446533203125,
0.2374267578125,
0.8544921875,
0.2374267578125,
0.8544921875,
0.251953125,
0.1446533203125,
0.251953125
],
[
0.146240234375,
0.255126953125,
0.8544921875,
0.255126953125,
0.8544921875,
0.268310546875,
0.146240234375,
0.268310546875
],
[
0.1446533203125,
0.272705078125,
0.8544921875,
0.272705078125,
0.8544921875,
0.285888671875,
0.1446533203125,
0.285888671875
],
[
0.1446533203125,
0.2890625,
0.85302734375,
0.2890625,
0.85302734375,
0.302978515625,
0.1446533203125,
0.302978515625
],
[
0.146240234375,
0.30615234375,
0.8544921875,
0.30615234375,
0.8544921875,
0.32080078125,
0.146240234375,
0.32080078125
],
[
0.1446533203125,
0.3212890625,
0.8544921875,
0.323974609375,
0.8544921875,
0.33837890625,
0.1446533203125,
0.3359375
],
[
0.146240234375,
0.34033203125,
0.85546875,
0.34033203125,
0.85546875,
0.354248046875,
0.146240234375,
0.354248046875
],
[
0.146240234375,
0.35791015625,
0.8544921875,
0.35791015625,
0.8544921875,
0.37255859375,
0.146240234375,
0.37255859375
],
[
0.1446533203125,
0.373779296875,
0.85107421875,
0.373779296875,
0.85107421875,
0.38818359375,
0.1446533203125,
0.38818359375
],
[
0.143798828125,
0.389404296875,
0.85546875,
0.39013671875,
0.85546875,
0.40771484375,
0.143798828125,
0.4072265625
],
[
0.1478271484375,
0.4091796875,
0.8544921875,
0.4091796875,
0.8544921875,
0.423583984375,
0.1478271484375,
0.423583984375
],
[
0.146240234375,
0.425537109375,
0.85546875,
0.425537109375,
0.85546875,
0.43994140625,
0.146240234375,
0.43994140625
],
[
0.1478271484375,
0.443115234375,
0.8544921875,
0.443115234375,
0.8544921875,
0.45703125,
0.1478271484375,
0.45703125
],
[
0.146240234375,
0.4609375,
0.85302734375,
0.4609375,
0.85302734375,
0.474853515625,
0.146240234375,
0.474853515625
],
[
0.1748046875,
0.479248046875,
0.85302734375,
0.479248046875,
0.85302734375,
0.49365234375,
0.1748046875,
0.49365234375
],
[
0.146240234375,
0.49560546875,
0.85546875,
0.49560546875,
0.85546875,
0.51025390625,
0.146240234375,
0.51025390625
],
[
0.146240234375,
0.51318359375,
0.8544921875,
0.51318359375,
0.8544921875,
0.52783203125,
0.146240234375,
0.52783203125
],
[
0.1446533203125,
0.52880859375,
0.85107421875,
0.52880859375,
0.85107421875,
0.54345703125,
0.1446533203125,
0.54345703125
],
[
0.1446533203125,
0.546875,
0.85302734375,
0.546875,
0.85302734375,
0.56103515625,
0.1446533203125,
0.56103515625
],
[
0.146240234375,
0.564453125,
0.8544921875,
0.564453125,
0.8544921875,
0.57763671875,
0.146240234375,
0.57763671875
],
[
0.1446533203125,
0.58056640625,
0.8544921875,
0.58056640625,
0.8544921875,
0.59521484375,
0.1446533203125,
0.59521484375
],
[
0.146240234375,
0.5986328125,
0.49267578125,
0.5986328125,
0.49267578125,
0.6123046875,
0.146240234375,
0.6123046875
],
[
0.5087890625,
0.5966796875,
0.8544921875,
0.5986328125,
0.8544921875,
0.61279296875,
0.5087890625,
0.611328125
],
[
0.146240234375,
0.6162109375,
0.85302734375,
0.6162109375,
0.85302734375,
0.6298828125,
0.146240234375,
0.6298828125
],
[
0.1478271484375,
0.63330078125,
0.8544921875,
0.63330078125,
0.8544921875,
0.646484375,
0.1478271484375,
0.646484375
],
[
0.146240234375,
0.6484375,
0.8544921875,
0.6494140625,
0.8544921875,
0.6640625,
0.146240234375,
0.66357421875
],
[
0.146240234375,
0.66650390625,
0.85107421875,
0.66650390625,
0.85107421875,
0.6796875,
0.146240234375,
0.6796875
],
[
0.1446533203125,
0.68310546875,
0.85107421875,
0.6845703125,
0.85107421875,
0.69873046875,
0.1446533203125,
0.69775390625
],
[
0.1446533203125,
0.69970703125,
0.39208984375,
0.70068359375,
0.39208984375,
0.71533203125,
0.1446533203125,
0.7138671875
],
[
0.1446533203125,
0.7392578125,
0.65673828125,
0.7392578125,
0.65673828125,
0.75390625,
0.1446533203125,
0.75390625
],
[
0.1446533203125,
0.76953125,
0.85302734375,
0.76953125,
0.85302734375,
0.78271484375,
0.1446533203125,
0.78271484375
],
[
0.146240234375,
0.7861328125,
0.85302734375,
0.7861328125,
0.85302734375,
0.7998046875,
0.146240234375,
0.7998046875
],
[
0.1446533203125,
0.8037109375,
0.8544921875,
0.8037109375,
0.8544921875,
0.818359375,
0.1446533203125,
0.818359375
],
[
0.148681640625,
0.833984375,
0.310546875,
0.83544921875,
0.310546875,
0.849609375,
0.148681640625,
0.8486328125
],
[
0.1502685546875,
0.849609375,
0.365234375,
0.849609375,
0.365234375,
0.86279296875,
0.1502685546875,
0.86279296875
],
[
0.1502685546875,
0.86279296875,
0.419189453125,
0.86279296875,
0.419189453125,
0.876953125,
0.1502685546875,
0.876953125
],
[
0.148681640625,
0.876953125,
0.464111328125,
0.876953125,
0.464111328125,
0.8916015625,
0.148681640625,
0.8916015625
],
[
0.491943359375,
0.91357421875,
0.5087890625,
0.91357421875,
0.5087890625,
0.92822265625,
0.491943359375,
0.92822265625
]
] |
[
"DMLR:PAST, PRESENT AND FUTURE",
"Serena Yeung, Maria Xenochristou. We have to emphasize that not all data work goes",
"under-appreciated. One must only look at ImageNet [8l or CIFAR [9] for great success",
"stories.",
"Around the same time as MNIST, the concept of data-centricity started to appear",
"literally in early works by [iol and others.It was likely discussed in the systems and",
"database circles long before the idea became an increasingly growing focus of research in",
"the core machine learning community. The connection to systems persists to this day,",
"evident in venues such as MLSys? or the DEEM workshops3, due the high importance",
"of optimized infrastructure to orchestrate and execute data transformation and machine",
"learning workloads. Success stories can be found in frameworks to build and store models",
"such as Torch[11],Theano[12],Caffe[13],TensorFlow [14],JAX [15] or PyTorch[16].",
"Sometimes these frameworks also led to optimized data formats, such as TensorFlow's",
"TFRecord.Additionally,platforms like Kaggle, HuggingFace or OpenML have emerged as",
"de-facto community data hubs and standardized data loading infrastructure. A new wave of",
"emerging open-source projects such as Lance4 aim to address existing gaps with respect to",
"data loading needs. However, despite these advances, challenges regarding the compatibility.",
"mutability, and collaborability of datasets persist. Encouragingly, new initiatives, such as",
"'Croissant'5, take stabs at the Babylonian tower of data formats, uniting key stakeholders",
"in an effort to streamline data-centric machine learning infrastructure. Similarly, DataPerf",
"[17] is a recent community-led benchmark suite for evaluating ML datasets and data-centric",
"algorithms, enabling the ML community to iterate on datasets, instead of just architectures.",
"Alongside developments in infrastructure, over time we have also witnessed critical ad-",
"vances in the way datasets are collected, curated and maintained. From the beginnings of",
"modern statistical science [18; 19], active learning, a set of methods concerned with data",
"curation, has planted its roots firmly in the machine learning and statistical learning lit-",
"erature [20-26]. The next generation of machine learning datasets are characterized by",
"dense metadata annotation [27-29], collaborative refinement [30-33], user preference and",
"human feedback [34], and evolution over time [35-37], similar to the way we treat code",
"for computer programs already today. Data",
"rovenanceand ownership arealso receiving",
"increasing consideration by groups such as the \"Data Trusts\" initiative [38] and others.",
"Our goal is to support the growth of the DMLR ecosystem into a strong community with",
"effective infrastructure that will advance machine learning science through next generation",
"datasets. These datasets shall serve as a bridge to connect fundamental problems (such as",
"food insecurity and climate impact) with fundamental ML research by providing the right",
"datasets for the right problems.",
"3 Present: Convening the Community at ICML 2023",
"In order to charge this effort, the data-centric ML community came together on July 29.",
"2023, in Honolulu, Hawai, for the inaugural DMLR workshop at the International Confer-",
"ence on Machine Learning (ICML) 2023. The DMLR workshop was a point of convergence",
"2.https://mlsys.org/",
"3.http://deem-workshop.org/",
"4.https://github.com/lancedb/lance",
"5.https://github.com/mlcommons/croissant",
"3"
] |
{
"arxiv_id": "2311.13028",
"language": "en",
"page_idx": 2
}
|
||
2311.13028-en-3
|
2311.13028-en-3.jpg
|
THE DMLR COMMUNITY
<FigureHere>
Figure 2: Themes and contributions from the community at the DMLR ICML 2023 workshop. Top left: LDA of accepted paper abstracts with n\_components=5.2-d UMAP of LDA results which are 5-d corresponding to 5 components. Each dot represents an abstract, color coded by the most dominant topic identified by LDA. The topics identified by LDA are displayed alongside as top20 word clouds. Bottom left: A sample of the geographic coor- $\begin{array}{c}\text{dinates of the institutions where authors of accepted works are based. It includes only those}\\\end{array}$ locations where the geocode API returns latitude and longitude information for fuzzy search on affiliation names (360 of 495 affiliations returned coordinates, note that not all 495 affiliations are unique). $\textbf{Right: Topics highlighted in the invited talk including prompt- based}$ ML development (Andrew Ng), the DMLR ecosystem (Peter Mattson), reality-centric AI (Mihaela van der Schaar), bias in vision data (Olga Russakovsky and Vikram Ramaswamy), history of distribution shifts dating back to NeurIPS 2006 (Masashi Sugiyama), the AI research agent (Isabelle Guyon), nuances of data quality (Dina Machuve), the DMLR Journal (Ce Zhang) and data-centric LLMs (panel). Links to the full videos and slides of talks are available in Appendix C.
for previous activities including the Asilomar Datasets 2030 retreat, the Dataperf initiative$^6$, the NeurIPS 2021 data-centric AI workshop$^7$, the LAION community$^8$ and others. Invited speakers, panelists, poster presenters and attendees deliberated on the current state of data-centric machine learning and how we can advance the community and infrastructure towards the next generation of public machine learning datasets (see Figure 2 for a brief overview).
Community engagement Andrew Ng concluded his keynote with open questions aimed at fostering further research and development in data-centric AI workflows. Isabelle Guyon proposed a peer-reviewed journal contributed to by AI-agents, aiming to foster scholarly community engagement. Dina Machuve discussed the role of community in data collection for agriculture in East Africa. Olga Rusnaskovsky and Vikram V. Ramaswamy addressed social bias in machine learning, calling for community action. The panel expressed substantial enthusiasm for the DMLR Journal, indicating a strong community interest in advancing
6. https://www.dataperf.org/
7. https://datacentricai.org/neurips21/
8. https://laion.ai/
4
|
[
[
0.400390625,
0.048614501953125,
0.5908203125,
0.048614501953125,
0.5908203125,
0.0625,
0.400390625,
0.0625
],
[
0.328369140625,
0.125,
0.35302734375,
0.125,
0.35302734375,
0.1287841796875,
0.328369140625,
0.1287841796875
],
[
0.324462890625,
0.1331787109375,
0.370849609375,
0.1331787109375,
0.370849609375,
0.14013671875,
0.324462890625,
0.14013671875
],
[
0.398681640625,
0.1319580078125,
0.438720703125,
0.1319580078125,
0.438720703125,
0.1395263671875,
0.398681640625,
0.1395263671875
],
[
0.328369140625,
0.1590576171875,
0.371826171875,
0.1590576171875,
0.371826171875,
0.1634521484375,
0.328369140625,
0.1634521484375
],
[
0.386474609375,
0.1578369140625,
0.43701171875,
0.1578369140625,
0.43701171875,
0.1685791015625,
0.386474609375,
0.1685791015625
],
[
0.76318359375,
0.250732421875,
0.81201171875,
0.250732421875,
0.81201171875,
0.26513671875,
0.76318359375,
0.26513671875
],
[
0.80615234375,
0.250732421875,
0.82275390625,
0.250732421875,
0.82275390625,
0.2587890625,
0.80615234375,
0.2587890625
],
[
0.146240234375,
0.31005859375,
0.8505859375,
0.31005859375,
0.8505859375,
0.323974609375,
0.146240234375,
0.323974609375
],
[
0.146240234375,
0.326904296875,
0.8544921875,
0.326904296875,
0.8544921875,
0.341552734375,
0.146240234375,
0.341552734375
],
[
0.146240234375,
0.3447265625,
0.8544921875,
0.3447265625,
0.8544921875,
0.359130859375,
0.146240234375,
0.359130859375
],
[
0.146240234375,
0.3623046875,
0.8544921875,
0.3623046875,
0.8544921875,
0.375732421875,
0.146240234375,
0.375732421875
],
[
0.146240234375,
0.37890625,
0.8544921875,
0.37890625,
0.8544921875,
0.393310546875,
0.146240234375,
0.393310546875
],
[
0.146240234375,
0.396484375,
0.8544921875,
0.396484375,
0.8544921875,
0.410400390625,
0.146240234375,
0.410400390625
],
[
0.1446533203125,
0.41357421875,
0.8544921875,
0.41357421875,
0.8544921875,
0.427978515625,
0.1446533203125,
0.427978515625
],
[
0.1446533203125,
0.429931640625,
0.85302734375,
0.429931640625,
0.85302734375,
0.4443359375,
0.1446533203125,
0.4443359375
],
[
0.1446533203125,
0.447509765625,
0.8544921875,
0.447509765625,
0.8544921875,
0.4609375,
0.1446533203125,
0.4609375
],
[
0.146240234375,
0.464111328125,
0.8544921875,
0.464111328125,
0.8544921875,
0.479248046875,
0.146240234375,
0.479248046875
],
[
0.143798828125,
0.477783203125,
0.8544921875,
0.479736328125,
0.8544921875,
0.49755859375,
0.143798828125,
0.49560546875
],
[
0.146240234375,
0.498779296875,
0.8505859375,
0.498779296875,
0.8505859375,
0.51318359375,
0.146240234375,
0.51318359375
],
[
0.146240234375,
0.5166015625,
0.8544921875,
0.5166015625,
0.8544921875,
0.52978515625,
0.146240234375,
0.52978515625
],
[
0.146240234375,
0.5322265625,
0.85302734375,
0.5322265625,
0.85302734375,
0.546875,
0.146240234375,
0.546875
],
[
0.1446533203125,
0.55029296875,
0.339111328125,
0.55029296875,
0.339111328125,
0.564453125,
0.1446533203125,
0.564453125
],
[
0.146240234375,
0.59765625,
0.8505859375,
0.59765625,
0.8505859375,
0.611328125,
0.146240234375,
0.611328125
],
[
0.146240234375,
0.6142578125,
0.85107421875,
0.6142578125,
0.85107421875,
0.62744140625,
0.146240234375,
0.62744140625
],
[
0.1446533203125,
0.6318359375,
0.8544921875,
0.6318359375,
0.8544921875,
0.64599609375,
0.1446533203125,
0.64599609375
],
[
0.146240234375,
0.6484375,
0.85302734375,
0.6484375,
0.85302734375,
0.662109375,
0.146240234375,
0.662109375
],
[
0.143798828125,
0.66552734375,
0.8544921875,
0.6640625,
0.8544921875,
0.6787109375,
0.143798828125,
0.6796875
],
[
0.1446533203125,
0.68310546875,
0.226318359375,
0.68310546875,
0.226318359375,
0.69775390625,
0.1446533203125,
0.69775390625
],
[
0.146240234375,
0.7138671875,
0.8544921875,
0.7138671875,
0.8544921875,
0.72900390625,
0.146240234375,
0.72900390625
],
[
0.1446533203125,
0.73095703125,
0.85302734375,
0.73095703125,
0.85302734375,
0.74560546875,
0.1446533203125,
0.74560546875
],
[
0.143798828125,
0.74853515625,
0.85107421875,
0.74755859375,
0.85107421875,
0.76220703125,
0.143798828125,
0.76318359375
],
[
0.146240234375,
0.7666015625,
0.85302734375,
0.7666015625,
0.85302734375,
0.77978515625,
0.146240234375,
0.77978515625
],
[
0.146240234375,
0.78271484375,
0.85302734375,
0.78271484375,
0.85302734375,
0.796875,
0.146240234375,
0.796875
],
[
0.146240234375,
0.7998046875,
0.85107421875,
0.7998046875,
0.85107421875,
0.81298828125,
0.146240234375,
0.81298828125
],
[
0.143798828125,
0.81494140625,
0.8544921875,
0.81640625,
0.8544921875,
0.83056640625,
0.143798828125,
0.830078125
],
[
0.1478271484375,
0.84765625,
0.363525390625,
0.8486328125,
0.363525390625,
0.86279296875,
0.1478271484375,
0.8623046875
],
[
0.1519775390625,
0.86279296875,
0.449462890625,
0.86279296875,
0.449462890625,
0.876953125,
0.1519775390625,
0.876953125
],
[
0.148681640625,
0.876953125,
0.303955078125,
0.876953125,
0.303955078125,
0.89013671875,
0.148681640625,
0.89013671875
],
[
0.494384765625,
0.91650390625,
0.50634765625,
0.91650390625,
0.50634765625,
0.92529296875,
0.494384765625,
0.92529296875
]
] |
[
"THE DMLR COMMUNITY",
"",
"dataser",
"mpaane",
"naTe",
"learning",
"0",
"L",
"Figure 2: Themes and contributions from the community at the DMLR ICML 2023 work-",
"shop. Top left: LDA of accepted paper abstracts with n_components = 5. 2-d UMAP of",
"LDA results which are 5-d corresponding to 5 components. Each dot represents an abstract,",
"color coded by the most dominant topic identified by LDA. The topics identified by LDA are",
"displayed alongside as top20 word clouds. Bottom left: A sample of the geographic coor-",
"dinates of the institutions where authors of accepted works are based. It includes only those",
"locations where the geocode API returns latitude and longitude information for fuzzy search",
"on affiliation names (360 of 495 afiliations returned coordinates, note that not all 495 affil-",
"iations are unique). Right: Topics highlighted in the invited talk including prompt-based",
"ML development (Andrew Ng), the DMLR ecosystem (Peter Mattson), reality-centric AI",
"(Mihaela van der Schaar), bias in vision data (Olga Russakovsky and Vikram Ramaswamy),",
"history of distribution shifts dating back to NeurIPS 2006 (Masashi Sugiyama), the AI re-",
"search agent (Isabelle Guyon), nuances of data quality (Dina Machuve), the DMLR Journal",
"(Ce Zhang) and data-centric LLMs (panel). Links to the full videos and slides of talks are",
"available in Appendix C.",
"for previous activities including the Asilomar Datasets 2030 retreat, the Dataperf initia",
"tive6, the NeurIPS 2021 data-centric AI workshop7, the LAION community8 and others",
"Invited speakers, panelists, poster presenters and attendees deliberated on the current state",
"of data-centric machine learning and how we can advance the community and infrastructure",
"towards the next generation of public machine learning datasets (see Figure 2 for a brief",
"overview).",
"Community engagement Andrew Ng concluded his keynote with open questions aimed",
"at fostering further research and development in data-centric AI workflows. Isabelle Guyon",
"proposed a peer-reviewed journal contributed to by AI-agents, aiming to foster scholarly",
"community engagement. Dina Machuve discussed the role of community in data collection",
"for agriculture in East Africa. Olga Rusnaskovsky and Vikram V. Ramaswamy addressed",
"social bias in machine learning, calling for community action. The panel expressed substan-",
"tial enthusiasm for the DMLR Journal, indicating a strong community interest in advancing",
"6.https://www.dataperf.org/",
"7.https://datacentricai.org/neurips21/",
"8. https://laion.ai/",
"4"
] |
{
"arxiv_id": "2311.13028",
"language": "en",
"page_idx": 3
}
|
||
2311.13028-en-4
|
2311.13028-en-4.jpg
|
DMLR: PAST, PRESENT AND FUTURE
the field. Paper authors highlighted diverse challenges in community standards ranging from risk classification in driver telematics, the role of synthetic data in the scientific community, to the nuances of deep learning in neuroimaging and beyond.
Infrastructure Workshop contributions also illuminated the critical role of infrastructure in advancing data-centric machine learning. Andrew Ng emphasized the importance of rapid iteration cycles, facilitated by advancements in both theory and tools. Mihaela van der Schaar introduced tools like Data-IQ [39] for better data characterization. Peter Mattson and Praveen Paritosh discussed Croissant$^9$, a standardized dataset format, and DataPerf[17], an engine for refining datasets. Masashi Sugiyama added depth by discussing the complexities of machine learning models operating under distribution shifts. The panel, consisting of Ludwig Schmidt, Megan Ansdell, Nathan Lambert, and Sang Michael Xie, further emphasized that the development of systematic methods for constructing AI datasets is less advanced compared to model development but noted that tools and infrastructure are catching up. Poster presenters highlighted different aspects related to infrastructure such as quality control and streaming of distributed data.
Datasets The workshop participants also delved into the future of datasets in machine learning. Andrew Ng highlighted the growing relevance of small datasets and the practicality of few-shot learning techniques. Mihaela van der Schaar advocated for Reality-Centric AI [40]. Isabelle Guyon introduced AutoML+, a holistic system that includes data search, task definition, and preparation. Dina Machuve discussed the critical role of data in East African agriculture. The panel emphasized that data holds a central role in driving AI forward and highlighted the need for next-gen datasets to be more systematically constructed. Several papers also underscored the challenges and solutions in active learning, focusing on topics such as minimizing annotation cost and acquiring high-quality data for training discriminative models. Links to the full videos and slides of talks are available in Appendix C.
4 Future: Growing the DMLR Ecosystem
The field of machine learning is undergoing a profound transformation. While the past was characterized by the pursuit of innovative algorithms and architectures, the present and future pose growing data-centric questions. As large models become the norm and real-world efficacy becomes paramount, the emphasis is shifting towards the entire data lifecycle, from collection over storage and transformation to integration of results into other systems [41]. The importance of addressing societal issues through data further underscores this shift. The role of the community in shaping the future of data-centric ML cannot be overstated. The recent DMLR workshop at ICML 2023 served as an inaugural meeting, igniting a spark for what is to come. A collective effort is required to create, enhance, and maintain public datasets. This involves establishing clear licensing protocols, technological standards, and fostering a culture of collaboration and shared, equitable ownership [42]. Interested contributors can already find several entry points, including
9. https://github.com/mlcommons/croissant
5
|
[
[
0.35302734375,
0.048614501953125,
0.640625,
0.048614501953125,
0.640625,
0.0625,
0.35302734375,
0.0625
],
[
0.146240234375,
0.117431640625,
0.85302734375,
0.117431640625,
0.85302734375,
0.1319580078125,
0.146240234375,
0.1319580078125
],
[
0.1446533203125,
0.1331787109375,
0.8505859375,
0.1337890625,
0.8505859375,
0.1483154296875,
0.1446533203125,
0.147705078125
],
[
0.143798828125,
0.1483154296875,
0.6259765625,
0.149658203125,
0.6259765625,
0.1673583984375,
0.143798828125,
0.1666259765625
],
[
0.146240234375,
0.186279296875,
0.85302734375,
0.186279296875,
0.85302734375,
0.2008056640625,
0.146240234375,
0.2008056640625
],
[
0.1446533203125,
0.203857421875,
0.8544921875,
0.203857421875,
0.8544921875,
0.2171630859375,
0.1446533203125,
0.2171630859375
],
[
0.1446533203125,
0.2203369140625,
0.85546875,
0.2197265625,
0.85546875,
0.2342529296875,
0.1446533203125,
0.23486328125
],
[
0.146240234375,
0.238037109375,
0.85302734375,
0.238037109375,
0.85302734375,
0.251953125,
0.146240234375,
0.251953125
],
[
0.1446533203125,
0.25390625,
0.8544921875,
0.255126953125,
0.8544921875,
0.26953125,
0.1446533203125,
0.268310546875
],
[
0.146240234375,
0.271484375,
0.8544921875,
0.271484375,
0.8544921875,
0.285888671875,
0.146240234375,
0.285888671875
],
[
0.146240234375,
0.2890625,
0.8544921875,
0.2890625,
0.8544921875,
0.3037109375,
0.146240234375,
0.3037109375
],
[
0.1446533203125,
0.30615234375,
0.85302734375,
0.30615234375,
0.85302734375,
0.32080078125,
0.1446533203125,
0.32080078125
],
[
0.146240234375,
0.322509765625,
0.8544921875,
0.322509765625,
0.8544921875,
0.337158203125,
0.146240234375,
0.337158203125
],
[
0.1446533203125,
0.34033203125,
0.8544921875,
0.34033203125,
0.8544921875,
0.354248046875,
0.1446533203125,
0.354248046875
],
[
0.146240234375,
0.35791015625,
0.8544921875,
0.35791015625,
0.8544921875,
0.371826171875,
0.146240234375,
0.371826171875
],
[
0.1446533203125,
0.375,
0.6005859375,
0.375,
0.6005859375,
0.389404296875,
0.1446533203125,
0.389404296875
],
[
0.146240234375,
0.4091796875,
0.8544921875,
0.4091796875,
0.8544921875,
0.423583984375,
0.146240234375,
0.423583984375
],
[
0.146240234375,
0.4267578125,
0.85302734375,
0.4267578125,
0.85302734375,
0.43994140625,
0.146240234375,
0.43994140625
],
[
0.146240234375,
0.443115234375,
0.8544921875,
0.443115234375,
0.8544921875,
0.458251953125,
0.146240234375,
0.458251953125
],
[
0.1478271484375,
0.4609375,
0.85302734375,
0.4609375,
0.85302734375,
0.474853515625,
0.1478271484375,
0.474853515625
],
[
0.146240234375,
0.479248046875,
0.8544921875,
0.479248046875,
0.8544921875,
0.492431640625,
0.146240234375,
0.492431640625
],
[
0.1478271484375,
0.49560546875,
0.85302734375,
0.49560546875,
0.85302734375,
0.51025390625,
0.1478271484375,
0.51025390625
],
[
0.1478271484375,
0.51318359375,
0.8544921875,
0.51318359375,
0.8544921875,
0.5263671875,
0.1478271484375,
0.5263671875
],
[
0.146240234375,
0.52978515625,
0.8544921875,
0.52978515625,
0.8544921875,
0.54443359375,
0.146240234375,
0.54443359375
],
[
0.146240234375,
0.54541015625,
0.85107421875,
0.54541015625,
0.85107421875,
0.56005859375,
0.146240234375,
0.56005859375
],
[
0.143798828125,
0.56298828125,
0.8486328125,
0.5625,
0.8486328125,
0.5771484375,
0.143798828125,
0.57763671875
],
[
0.1478271484375,
0.60498046875,
0.5498046875,
0.60498046875,
0.5498046875,
0.619140625,
0.1478271484375,
0.619140625
],
[
0.1478271484375,
0.6337890625,
0.8544921875,
0.6337890625,
0.8544921875,
0.6484375,
0.1478271484375,
0.6484375
],
[
0.146240234375,
0.65283203125,
0.8505859375,
0.65283203125,
0.8505859375,
0.66650390625,
0.146240234375,
0.66650390625
],
[
0.146240234375,
0.66943359375,
0.8544921875,
0.66943359375,
0.8544921875,
0.68310546875,
0.146240234375,
0.68310546875
],
[
0.146240234375,
0.68603515625,
0.85302734375,
0.68603515625,
0.85302734375,
0.70068359375,
0.146240234375,
0.70068359375
],
[
0.146240234375,
0.7021484375,
0.8505859375,
0.70263671875,
0.8505859375,
0.71728515625,
0.146240234375,
0.71630859375
],
[
0.1446533203125,
0.71826171875,
0.83642578125,
0.7197265625,
0.83642578125,
0.734375,
0.1446533203125,
0.73291015625
],
[
0.146240234375,
0.75390625,
0.85302734375,
0.75390625,
0.85302734375,
0.767578125,
0.146240234375,
0.767578125
],
[
0.146240234375,
0.771484375,
0.8544921875,
0.771484375,
0.8544921875,
0.7861328125,
0.146240234375,
0.7861328125
],
[
0.146240234375,
0.7890625,
0.85302734375,
0.7890625,
0.85302734375,
0.8037109375,
0.146240234375,
0.8037109375
],
[
0.1446533203125,
0.80615234375,
0.85302734375,
0.80615234375,
0.85302734375,
0.82080078125,
0.1446533203125,
0.82080078125
],
[
0.146240234375,
0.82373046875,
0.85546875,
0.82373046875,
0.85546875,
0.83837890625,
0.146240234375,
0.83837890625
],
[
0.146240234375,
0.84033203125,
0.61669921875,
0.84033203125,
0.61669921875,
0.85498046875,
0.146240234375,
0.85498046875
],
[
0.1502685546875,
0.876953125,
0.463134765625,
0.876953125,
0.463134765625,
0.8916015625,
0.1502685546875,
0.8916015625
],
[
0.491943359375,
0.91357421875,
0.5087890625,
0.91357421875,
0.5087890625,
0.92822265625,
0.491943359375,
0.92822265625
]
] |
[
"DMLR: PAST, PRESENT AND FUTURE",
"the field. Paper authors highlighted diverse challenges in community standards ranging from",
"risk classification in driver telematics, the role of synthetic data in the scientific community",
"to the nuances of deep learning in neuroimaging and beyond.",
"Infrastructure Workshop contributions also illuminated the critical role of infrastruc-",
"ture in advancing data-centric machine learning. Andrew Ng emphasized the importance",
"of rapid iteration cycles, facilitated by advancements in both theory and tools. Mihaela",
"van der Schaar introduced tools like Data-IQ [39] for better data characterization.Pe-",
"ter Mattson and Praveen Paritosh discussed Croissant, a standardized dataset format, and",
"DataPerf [17l, an engine for refining datasets. Masashi Sugiyama added depth by discussing",
"the complexities of machine learning models operating under distribution shifts. The panel,",
"consisting of Ludwig Schmidt, Megan Ansdell, Nathan Lambert, and Sang Michael Xie, fur-",
"ther emphasized that the development of systematic methods for constructing AI datasets",
"is less advanced compared to model development but noted that tools and infrastructure",
"are catching up. Poster presenters highlighted different aspects related to infrastructure",
"such as quality control and streaming of distributed data.",
"Datasets The workshop participants also delved into the future of datasets in machine",
"learning. Andrew Ng highlighted the growing relevance of small datasets and the practical-",
"ity of few-shot learning techniques. Mihaela van der Schaar advocated for Reality-Centric",
"AI [4ol. Isabelle Guyon introduced AutoML+, a holistic system that includes data search.",
"task definition, and preparation. Dina Machuve discussed the critical role of data in East",
"African agriculture. The panel emphasized that data holds a central role in driving AI for-",
"ward and highlighted the need for next-gen datasets to be more systematically constructed.",
"Several papers also underscored the challenges and solutions in active learning, focusing on",
"topics such as minimizing annotation cost and acquiring high-quality data for training dis-",
"criminative models. Links to the full videos and slides of talks are available in Appendix C.",
"4 Future: Growing the DMLR Ecosystem",
"The field of machine learning is undergoing a profound transformation. While the past was",
"characterized by the pursuit of innovative algorithms and architectures, the present and fu-",
"ture pose growing data-centric questions. As large models become the norm and real-world",
"efficacy becomes paramount, the emphasis is shifting towards the entire data lifecycle, from",
"collection over storage and transformation to integration of results into other systems [41]",
"The importance of addressing societal issues through data further underscores this shift.",
"The role of the community in shaping the future of data-centric ML cannot be overstated.",
"The recent DMLR workshop at ICML 2023 served as an inaugural meeting, igniting a spark",
"for what is to come. A collective effort is required to create, enhance, and maintain pub-",
"lic datasets. This involves establishing clear licensing protocols, technological standards,",
"and fostering a culture of collaboration and shared, equitable ownership [42]. Interested",
"contributors can already find several entry points, including",
"9.https://github.com/mlcommons/croissant",
"5"
] |
{
"arxiv_id": "2311.13028",
"language": "en",
"page_idx": 4
}
|
||
2402.19298-en-0
|
2402.19298-en-0.jpg
|
Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing
$w$
$હં$
$ন্ল$
Xun Lin$^1$, Shuai Wang$^1$,Rizhao Cai$^2$,Yizhong Liu$^1$,Ying Fu$^3$,Zitong Yu$^{4*}$,Wenzhong Tang$^1$,Alex Kot$^2$
$^{4}$Great Bay University
$^{1}$Beihang University$^2$Nanyang Technological University$^{3}$BIT
$\{$linxun, wangshuai$\}$@buaa.edu.cn
yuzitong@gbu.edu.cn
## Abstract
<FigureHere>
Figure 1. Illustration of DG scenarios in the context of (a) uni-
Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. With ad-
$\sum_{ing~techniques,~many~multi-modal~learn-}$
emerged. However, they face challenges in generalizing to unseen attacks and deployment conditions. These chal-
$\sum_{\begin{array}{ccc}&\text{lenges arise from (1) modality unrel}\end{array}}$
the spread of unreliable information during cross-modal feature fusion, and (2) modality imbalance, where training overly relies on a dominant modality hinders the con-
$\begin{array}{lll}&\text{vergence of others, reducing effectiven}\end{array}$
Uncertainty-Guided Cross-Adapter(U-Adapter) to recognize unreliably detected regions within each modality and suppress the impact of unreliable regions on other modalities. For modality imbalance, we propose a Rebalanced Modality Gradient Modulation (ReGrad) strategy to rebalance the convergence speed of all modalities by adaptively adjusting their gradients. Besides, we provide the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. Source code and protocols will be released on https://github.com/OMGGGGG/mmdg.
modal and (b) multi-modal. (c) DG performance of SSDG [20] on our Protocol 1 (see Sec. 4.1). Though fed with more modalities, SSDG performs worse in multi-modal scenarios compared to unimodal ones. * denotes our re-implemented multi-modal version. curity of FR systems, limiting their broader application. As a result, the development of face anti-spoofing (FAS) methods has become a research hot-spot to enhance the security of FR systems.
Recently, deep learning has been widely applied in FAS, particularly in unimodal scenarios based on the RGB spectrum. While these models perform well in intra-dataset evaluations, their generalizability to unseen attacks and deployment environments remains limited [2]. Despite the improvement of robustness via domain generalization (DG) techniques [20, 28, 49, 50, 70], these approaches still tend to over-fit the attacks presented in source domains.
With the recent progress of multi-modal learning [52,54] and the development of sensor manufacture[62], some large-scale multi-modal FAS datasets [15,34,45,67] are proposed, which also drive the advancement in multi-modal FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the input from RGB to multi-modal, $e.g.$, RGB, depth, and infrared, richer spoofing traces can be captured for detecting various presentation attacks [63]. However, existing multi-
## $1. \textbf{Introduction}$
The remarkable success of deep neural networks has also brought concerns regarding their security vulnerabilities [31,56,57,64]. Especially for face recognition (FR) systems, widely used in applications like surveillance and mobile payment [63], are susceptible to various face presentation attacks, including printed photos, video replays, and 3D wearable masks[3]. These vulnerabilities challenge the se-
$^{*}$Corresponding author.
|
[
[
0.08416748046875,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.149658203125,
0.08416748046875,
0.149658203125
],
[
0.0906982421875,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1976318359375,
0.0906982421875,
0.1976318359375
],
[
0.1217041015625,
0.1988525390625,
0.293212890625,
0.199462890625,
0.293212890625,
0.2139892578125,
0.1217041015625,
0.21337890625
],
[
0.29736328125,
0.1976318359375,
0.59716796875,
0.1988525390625,
0.59716796875,
0.216552734375,
0.29736328125,
0.21533203125
],
[
0.615234375,
0.199462890625,
0.67333984375,
0.199462890625,
0.67333984375,
0.2139892578125,
0.615234375,
0.2139892578125
],
[
0.6806640625,
0.1988525390625,
0.861328125,
0.2008056640625,
0.861328125,
0.21533203125,
0.6806640625,
0.21337890625
],
[
0.294189453125,
0.205810546875,
0.303955078125,
0.205810546875,
0.303955078125,
0.208984375,
0.294189453125,
0.208984375
],
[
0.2548828125,
0.2197265625,
0.53125,
0.2197265625,
0.53125,
0.23291015625,
0.2548828125,
0.23291015625
],
[
0.55224609375,
0.2203369140625,
0.7236328125,
0.2197265625,
0.7236328125,
0.23291015625,
0.55224609375,
0.2342529296875
],
[
0.0236968994140625,
0.266357421875,
0.058013916015625,
0.266357421875,
0.058013916015625,
0.32568359375,
0.0236968994140625,
0.32568359375
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.515625,
0.272705078125,
0.52783203125,
0.272705078125,
0.52783203125,
0.281005859375,
0.515625,
0.281005859375
],
[
0.098876953125,
0.300537109375,
0.46728515625,
0.3017578125,
0.46728515625,
0.31640625,
0.098876953125,
0.31494140625
],
[
0.02777099609375,
0.313232421875,
0.058013916015625,
0.314453125,
0.053924560546875,
0.373779296875,
0.0236968994140625,
0.37255859375
],
[
0.08087158203125,
0.317626953125,
0.464111328125,
0.317626953125,
0.464111328125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.53369140625,
0.318115234375,
0.57861328125,
0.318115234375,
0.57861328125,
0.328857421875,
0.53369140625,
0.328857421875
],
[
0.58642578125,
0.3193359375,
0.6357421875,
0.3193359375,
0.6357421875,
0.326904296875,
0.58642578125,
0.326904296875
],
[
0.63232421875,
0.3193359375,
0.6806640625,
0.3193359375,
0.6806640625,
0.326904296875,
0.63232421875,
0.326904296875
],
[
0.740234375,
0.318115234375,
0.80224609375,
0.318115234375,
0.80224609375,
0.328857421875,
0.740234375,
0.328857421875
],
[
0.515625,
0.330810546875,
0.529296875,
0.330810546875,
0.529296875,
0.339111328125,
0.515625,
0.339111328125
],
[
0.76123046875,
0.330810546875,
0.78173828125,
0.330810546875,
0.78173828125,
0.339111328125,
0.76123046875,
0.339111328125
],
[
0.6005859375,
0.333251953125,
0.60888671875,
0.333251953125,
0.60888671875,
0.33837890625,
0.6005859375,
0.33837890625
],
[
0.08087158203125,
0.333251953125,
0.46728515625,
0.333251953125,
0.46728515625,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.080078125,
0.347900390625,
0.46728515625,
0.347900390625,
0.46728515625,
0.361083984375,
0.080078125,
0.361083984375
],
[
0.02777099609375,
0.36865234375,
0.056365966796875,
0.371826171875,
0.052276611328125,
0.389404296875,
0.02532958984375,
0.386474609375
],
[
0.080078125,
0.363037109375,
0.46728515625,
0.363037109375,
0.46728515625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.080078125,
0.37890625,
0.46728515625,
0.37890625,
0.46728515625,
0.392578125,
0.080078125,
0.392578125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.4443359375,
0.02532958984375,
0.4443359375
],
[
0.08087158203125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.08087158203125,
0.4072265625
],
[
0.07843017578125,
0.4072265625,
0.46728515625,
0.40771484375,
0.46728515625,
0.42236328125,
0.07843017578125,
0.421630859375
],
[
0.85302734375,
0.40478515625,
0.865234375,
0.40478515625,
0.865234375,
0.41357421875,
0.85302734375,
0.41357421875
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.03350830078125,
0.443115234375,
0.053924560546875,
0.443115234375,
0.053924560546875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.53173828125,
0.45263671875,
0.6845703125,
0.453857421875,
0.6845703125,
0.46533203125,
0.53173828125,
0.464111328125
],
[
0.73779296875,
0.453857421875,
0.8056640625,
0.453857421875,
0.8056640625,
0.46533203125,
0.73779296875,
0.46533203125
],
[
0.080078125,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.5146484375,
0.468505859375,
0.529296875,
0.468505859375,
0.529296875,
0.477783203125,
0.5146484375,
0.477783203125
],
[
0.55615234375,
0.47021484375,
0.5947265625,
0.47021484375,
0.5947265625,
0.481689453125,
0.55615234375,
0.481689453125
],
[
0.63623046875,
0.4716796875,
0.6787109375,
0.4716796875,
0.6787109375,
0.481689453125,
0.63623046875,
0.481689453125
],
[
0.71728515625,
0.4716796875,
0.7685546875,
0.4716796875,
0.7685546875,
0.481689453125,
0.71728515625,
0.481689453125
],
[
0.80810546875,
0.47021484375,
0.85546875,
0.47021484375,
0.85546875,
0.481689453125,
0.80810546875,
0.481689453125
],
[
0.08087158203125,
0.484130859375,
0.46728515625,
0.484130859375,
0.46728515625,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.55615234375,
0.4873046875,
0.59228515625,
0.4873046875,
0.59228515625,
0.5,
0.55615234375,
0.5
],
[
0.63134765625,
0.4873046875,
0.6806640625,
0.48486328125,
0.68115234375,
0.5,
0.63232421875,
0.501953125
],
[
0.7294921875,
0.4873046875,
0.7607421875,
0.4873046875,
0.7607421875,
0.5,
0.7294921875,
0.5
],
[
0.81640625,
0.4873046875,
0.8505859375,
0.4873046875,
0.8505859375,
0.5,
0.81640625,
0.5
],
[
0.0236968994140625,
0.496826171875,
0.058013916015625,
0.496826171875,
0.058013916015625,
0.5771484375,
0.0236968994140625,
0.5771484375
],
[
0.0310516357421875,
0.4912109375,
0.052276611328125,
0.4912109375,
0.052276611328125,
0.5048828125,
0.0310516357421875,
0.5048828125
],
[
0.07843017578125,
0.5,
0.468017578125,
0.498779296875,
0.468017578125,
0.51220703125,
0.07843017578125,
0.51318359375
],
[
0.5556640625,
0.5048828125,
0.5947265625,
0.5048828125,
0.5947265625,
0.5166015625,
0.5556640625,
0.5166015625
],
[
0.71337890625,
0.5048828125,
0.775390625,
0.5048828125,
0.775390625,
0.5166015625,
0.71337890625,
0.5166015625
],
[
0.80224609375,
0.5048828125,
0.86376953125,
0.5048828125,
0.86376953125,
0.5166015625,
0.80224609375,
0.5166015625
],
[
0.080078125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.62353515625,
0.505859375,
0.69287109375,
0.505859375,
0.69287109375,
0.5166015625,
0.62353515625,
0.5166015625
],
[
0.50244140625,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.07843017578125,
0.52783203125,
0.46728515625,
0.52880859375,
0.46728515625,
0.54345703125,
0.07843017578125,
0.54248046875
],
[
0.50341796875,
0.53515625,
0.89208984375,
0.53515625,
0.89208984375,
0.548828125,
0.50341796875,
0.548828125
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.5009765625,
0.54736328125,
0.89208984375,
0.548828125,
0.89208984375,
0.5625,
0.5009765625,
0.56103515625
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5625,
0.89208984375,
0.5625,
0.89208984375,
0.57568359375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.57470703125,
0.4697265625,
0.57470703125,
0.4697265625,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.5771484375,
0.88232421875,
0.5771484375,
0.88232421875,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.0318603515625,
0.583984375,
0.052276611328125,
0.58837890625,
0.0482177734375,
0.60400390625,
0.0269622802734375,
0.599609375
],
[
0.080078125,
0.5888671875,
0.468017578125,
0.5888671875,
0.468017578125,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.5009765625,
0.5966796875,
0.89208984375,
0.59765625,
0.89208984375,
0.6123046875,
0.5009765625,
0.611328125
],
[
0.080078125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.0318603515625,
0.61279296875,
0.049835205078125,
0.61279296875,
0.049835205078125,
0.626953125,
0.0318603515625,
0.626953125
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.03350830078125,
0.62353515625,
0.041656494140625,
0.62548828125,
0.037567138671875,
0.6396484375,
0.0294189453125,
0.63720703125
],
[
0.080078125,
0.619140625,
0.4697265625,
0.619140625,
0.4697265625,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.03350830078125,
0.63330078125,
0.052276611328125,
0.63330078125,
0.052276611328125,
0.6640625,
0.03350830078125,
0.6640625
],
[
0.50244140625,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.50244140625,
0.642578125
],
[
0.080078125,
0.63525390625,
0.46728515625,
0.63525390625,
0.46728515625,
0.6484375,
0.080078125,
0.6484375
],
[
0.50244140625,
0.64306640625,
0.5986328125,
0.64306640625,
0.5986328125,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.07843017578125,
0.6494140625,
0.468017578125,
0.6484375,
0.468017578125,
0.66357421875,
0.07843017578125,
0.6640625
],
[
0.0228729248046875,
0.66015625,
0.052276611328125,
0.65576171875,
0.062103271484375,
0.69970703125,
0.0318603515625,
0.7041015625
],
[
0.52294921875,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.52294921875,
0.67236328125
],
[
0.080078125,
0.66552734375,
0.4697265625,
0.66552734375,
0.4697265625,
0.6787109375,
0.080078125,
0.6787109375
],
[
0.07843017578125,
0.6796875,
0.4697265625,
0.6787109375,
0.4697265625,
0.693359375,
0.07843017578125,
0.6943359375
],
[
0.50244140625,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.08087158203125,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.08087158203125,
0.708984375
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.50244140625,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.080078125,
0.74072265625,
0.20751953125,
0.7421875,
0.20751953125,
0.7568359375,
0.080078125,
0.7548828125
],
[
0.50341796875,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08087158203125,
0.76806640625,
0.468017578125,
0.76806640625,
0.468017578125,
0.7822265625,
0.08087158203125,
0.7822265625
],
[
0.50244140625,
0.76513671875,
0.833984375,
0.76513671875,
0.833984375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.521484375,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79345703125,
0.521484375,
0.79345703125
],
[
0.08087158203125,
0.7861328125,
0.468017578125,
0.7861328125,
0.468017578125,
0.796875,
0.08087158203125,
0.796875
],
[
0.08087158203125,
0.7998046875,
0.468017578125,
0.7998046875,
0.468017578125,
0.81298828125,
0.08087158203125,
0.81298828125
],
[
0.5048828125,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.5048828125,
0.80810546875
],
[
0.080078125,
0.814453125,
0.468017578125,
0.814453125,
0.468017578125,
0.82763671875,
0.080078125,
0.82763671875
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.830078125,
0.468017578125,
0.830078125,
0.468017578125,
0.84326171875,
0.080078125,
0.84326171875
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82470703125,
0.89208984375,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.08087158203125,
0.8447265625,
0.468017578125,
0.8447265625,
0.468017578125,
0.85791015625,
0.08087158203125,
0.85791015625
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85986328125,
0.46728515625,
0.85986328125,
0.46728515625,
0.87353515625,
0.080078125,
0.87353515625
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.098876953125,
0.88720703125,
0.223876953125,
0.88818359375,
0.223876953125,
0.8994140625,
0.098876953125,
0.8984375
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing",
"Xun Lin, Shuai Wang, Rizhao Cai?, Yizhong Liu, Ying Fu3, Zitong Yu4*, Wenzhong Tang', Alex Kot?",
"1Beihang University",
"2Nanyang Technological University",
"3BIT",
"4Great Bay University",
"",
"[linxun, wangshuai]@buaa.edu.cn",
"yuzitong@gbu.edu.cn",
"",
"Abstract",
"(a)",
"Face Anti-Spoofing (FAS) is crucial for securing face",
"",
"recognition systems against presentation attacks. With ad",
"Domain1",
"Domain2...",
"..Domain",
"Target Domair",
"(b)",
"Test",
"ra",
"vancements in sensor manufacture and multi-modal learn-",
"ing techniques, many multi-modal FAs approaches have",
"5",
"emerged. However, they face challenges in generalizing",
"to unseen attacks and deployment conditions. These chal-",
"LAD",
"lenges arise from (1)modality unreliability, where some",
"modality sensors like depth and infrared undergo signif-",
"X",
"icant domain shifts in varying environments, leading to",
"'SO",
"the spread of unreliable information during cross-modal",
"feature fusion, and (2) modality imbalance, where train-",
"Domain1Domain2...Domainn",
"Target Domain",
"ing overly relies on a dominant modality hinders the con-",
"(c)",
"Method",
"Modality",
"HTER(%)",
"AUC (%)",
"vergence of others, reducing effectiveness against attack",
"SSDG",
"(a): RGB",
"24.07",
"83.69",
"2120",
"2",
"types that are indistinguishable sorely using the dominant",
"SSDG*",
"26.09 (+2.02)",
"82.03 (-1.66)",
"modality. To address modality unreliability, we propose the",
"(b): RGB+D+I",
"Figure 1. Illustration of DG scenarios in the context of (a) uni-",
"Uncertainty-Guided Cross-Adapter (U-Adapter) to recog-",
"modal and (b) multi-modal. (c) DG performance of SSDG [20] on",
"nize unreliably detected regions within each modality and",
"our Protocol 1 (see Sec. 4.1). Though fed with more modalities,",
"suppress the impact of unreliable regions on other modal-",
"SSDG performs worse in multi-modal scenarios compared to uni-",
"ities. For modality imbalance, we propose a Rebalanced",
"modal ones. * denotes our re-implemented multi-modal version.",
"2",
"Modality Gradient Modulation (ReGrad) strategy to rebal-",
"curity of FR systems, limiting their broader application. As",
"ance the convergence speed of all modalities by adaptively",
"4",
"a result, the development of face anti-spoofing (FAS) meth-",
"0",
"adjusting their gradients. Besides, we provide the first",
".AI",
"ods has become a research hot-spot to enhance the security",
"large-scale benchmark for evaluating multi-modal FAS per-",
"of FR systems",
"formance under domain generalization scenarios. Exten-",
"X",
"Recently, deep learning has been widely applied in FAS,",
"sive experiments demonstrate that our method outperforms",
"state-of-the-art methods. Source code and protocols will be",
"particularly in unimodal scenarios based on the RGB spec-",
"trum.While these models perform well in intra-dataset",
"releasedon https://github.com/OMGGGGG/mmdg.",
"evaluations, their generalizability to unseen attacks and de-",
"ployment environments remains limited [2]. Despite the",
"improvement of robustness via domain generalization (DG)",
"1. Introduction",
"techniques [20, 28, 49, 50, 70], these approaches still tend",
"The remarkable success of deep neural networks has also",
"to over-fit the attacks presented in source domains.",
"With the recent progress of multi-modal learning [52, 54]",
"brought concerns regarding their security vulnerabilities",
"[31, 56, 57, 64]. Especially for face recognition (FR) sys-",
"andthe development of sensormanufacture[62l,some",
"tems, widely used in applications like surveillance and mo-",
"large-scale multi-modal FAS datasets [15, 34, 45, 67] are",
"bile payment [63], are susceptible to various face presenta-",
"proposed, which also drive the advancement in multi-modal",
"tion attacks, including printed photos, video replays, and 3D",
"FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the in-",
"wearable masks [3]. These vulnerabilities challenge the se-",
"put from RGB to multi-modal, e.g., RGB, depth, and in-",
"frared, richer spoofing traces can be captured for detecting",
"*Corresponding author.",
"various presentation attacks [63]. However, existing multi-"
] |
{
"arxiv_id": "2402.19298",
"language": "en",
"page_idx": 0
}
|
||
2402.19298-en-1
|
2402.19298-en-1.jpg
|
modal methods fail to perform well under unseen deployment conditions. As illustrated in Fig. 1, though the current methods generalize well in unimodal scenarios, their performance abnormally declines when extended and applied to multi-modal FAS. Why using more modalities is $notbetter?$ We argue that there are two main reasons: (1) Modality unreliability. Deploying in unseen environments or with various kinds of models of imaging sensors within the same modality, especially the depth and infrared shown in Fig. l (b), might introduce significant domain shifts[66]. Such shifts can result in unreliable multimodal live/spoof feature extraction [51], misleading mutual modalities through cross-modal fusion. (2) Modality imbalance. In some multi-modal learning scenarios, models tend to overly rely on a dominant modality with the highest convergence speed $(i.e.$ fast modality), which prevents itself from fully exploiting the other dominated modalities with relatively slower convergence speeds $(i.e.$ slow modalities)[43]. In FAS, modality imbalance can cause significant negative impacts, especially when the fast modality is unsuitable or unavailable for detecting attacks in the target domain, relying on slow modalities that have not adequately converged cannot achieve satisfactory performance. This study addresses the cross-domain limitations of multimodal FAS approaches due to modality unreliability and imbalance. As illustrated in Fig. 2, we introduce a MultiModal Domain Generalized (MMDG) framework for FAS. Within MMDG, we develop trainable Uncertainty-guided cross-Adapters (U-Adapter) to fine-tune Vision Transformers (ViT) [9] The U-Adapter address modality unreliability leveraging the uncertainty of each modality to suppress the focus on unreliable tokens during the cross-modal fusion. This prevents the excessive propagation of unreliable spoofing traces which tends to decrease the model's discriminative capability. Meanwhile, to solve modality imbalance, we design Rebalanced modality Gradient modulation strategy (ReGrad). Our ReGrad rebalance the convergence speed of all modalities by dynamically adjusting the backward propagated gradients of all trainable parameters in U-Adapters based on their conflict degree and convergence speed. This ensures that all modalities can be fully leveraged to resist various unseen attacks in the target domain. Our contributions are as follows.
$\cdot$ We propose the MMDG framework to enhance the do-
main generalizability by addressing the modality unreliability and imbalance. This is the first investigation for multi-modal FAS under multi-source DG scenarios.
$\cdot$ Within MMDG, we propose a Uncertainty-Guided Cross-
$\mathbf{Adapter~(U-Adapter)~for~ViT-based~multi-modal~FAS,}$ which can complementarily fuse cross-modal features and prevents each modality from being influenced by unreliable information caused by domain shifts.
· We design the Rebalanced Modality Gradient Modula-
tion (ReGrad) for MMDG, a strategy that can monitor and rebalance the convergence speed of each modality by adaptively modulating gradients of U-Adapters.
$\cdot$ We build the first large-scale benchmark to evaluate the
domain generalizability of multi-modal FAS approaches. Extensive experiments on this benchmark demonstrate the superiority and generalizability of MMDG.
## $2. \textbf{Related Works}$
$2. 1. \textbf{ Domain Generalization for FAS}$
In FAS, domain generalization focuses on training a model using multiple source domains, with the intention that this model can work effectively on unseen target domains [18,46]. Previous works verify the effectiveness of adversarial training [22, 37, 65], asymmetric triplet loss [20, 30] and controversial learning [38,49] in developing a shared feature space across domains. Besides, some studies [50, 69,70] investigate the style information of different domains, aiming to learn domain-invariant features by minimizing distinct style characteristics. Furthermore, metalearning-based methods [4,10,21,44,68] are also discussed to simulate domain shifts during training, fostering the learning of a robust representative feature space to such variations. More recent works focus on parameter-efficient transfer learning [5,6,18,48], enabling a pretrained ViT to efficiently adapt to unseen domains with a lower risk of over-fitting. However, these approaches are proposed for unimodal $(i.e.$ RGB) FAS. Due to the ignorance of modality unreliability, there is a gap in adapting them to multimodal scenarios. Simple multi-modal extensions of these approaches, through early fusion, late fusion, or even the substitution with multi-modal backbones, still exhibit limited performance under multi-modal DG scenarios.
### $2. 2. \textbf{Multi- Modal FAS}$
Multi-modal FAS refers to the use of multiple spectra, e.g., RGB, depth, and infrared, to reveal live and spoofing traces. Due to the complementary information among different modalities, traces that are undetectable in one modality can be easily captured in others [17, 36]. To fully exploit information from each modality, early methods employ channelwise concatenation of inputted modalities [14,42], or develop multiple separate branches for modality feature extraction followed by late fusion [25,26,47,58]. Recent works introduce attention-based feature fusion techniques [8,27] and adaptive cross-modal loss functions [13] to encourage the extraction of complementary information among modalities. Meanwhile, cross-modality translation is discussed [29, 35] to mitigate the semantic gap between different modalities. To further enhance the practical usability of FAS, Yu et al. [60, 61, 62] introduce flexible-modal benchmarks that force the use of incomplete modalities dur-
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.515625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.515625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5146484375,
0.106689453125,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.5146484375,
0.12060546875
],
[
0.08087158203125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.51318359375,
0.12249755859375,
0.830078125,
0.121826171875,
0.830078125,
0.1363525390625,
0.51318359375,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5009765625,
0.1363525390625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.51708984375,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.51708984375,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.4697265625,
0.1673583984375,
0.4697265625,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.51708984375,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.1817626953125,
0.51708984375,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5146484375,
0.1817626953125,
0.82275390625,
0.1817626953125,
0.82275390625,
0.1956787109375,
0.5146484375,
0.1956787109375
],
[
0.07843017578125,
0.1962890625,
0.46728515625,
0.1976318359375,
0.46728515625,
0.212158203125,
0.07843017578125,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50341796875,
0.21337890625,
0.650390625,
0.21337890625,
0.650390625,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.238037109375,
0.77783203125,
0.238037109375,
0.77783203125,
0.25244140625,
0.50341796875,
0.25244140625
],
[
0.07843017578125,
0.2568359375,
0.4697265625,
0.25830078125,
0.4697265625,
0.272705078125,
0.07843017578125,
0.271484375
],
[
0.50244140625,
0.261962890625,
0.89208984375,
0.261962890625,
0.89208984375,
0.276611328125,
0.50244140625,
0.276611328125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.279052734375,
0.89208984375,
0.279052734375,
0.89208984375,
0.292236328125,
0.50244140625,
0.292236328125
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.29345703125,
0.89208984375,
0.29345703125,
0.89208984375,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50341796875,
0.30810546875,
0.890625,
0.30810546875,
0.890625,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50244140625,
0.323974609375,
0.890625,
0.323974609375,
0.890625,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.33837890625,
0.89208984375,
0.33837890625,
0.89208984375,
0.3515625,
0.50244140625,
0.3515625
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.354736328125,
0.890625,
0.354736328125,
0.890625,
0.365478515625,
0.50341796875,
0.365478515625
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50244140625,
0.3818359375
],
[
0.080078125,
0.37890625,
0.4697265625,
0.37890625,
0.4697265625,
0.392578125,
0.080078125,
0.392578125
],
[
0.50341796875,
0.385009765625,
0.8896484375,
0.385009765625,
0.8896484375,
0.395751953125,
0.50341796875,
0.395751953125
],
[
0.08087158203125,
0.395751953125,
0.46728515625,
0.395751953125,
0.46728515625,
0.406005859375,
0.08087158203125,
0.406005859375
],
[
0.50341796875,
0.40087890625,
0.8896484375,
0.40087890625,
0.8896484375,
0.410888671875,
0.50341796875,
0.410888671875
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.4140625,
0.890625,
0.4140625,
0.890625,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.4248046875,
0.465576171875,
0.4248046875,
0.465576171875,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.5009765625,
0.427978515625,
0.89208984375,
0.4287109375,
0.89208984375,
0.443115234375,
0.5009765625,
0.442626953125
],
[
0.080078125,
0.438232421875,
0.468017578125,
0.439453125,
0.468017578125,
0.453857421875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.4697265625,
0.468017578125,
0.4697265625,
0.468017578125,
0.48291015625,
0.080078125,
0.48291015625
],
[
0.50244140625,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.080078125,
0.48486328125,
0.46728515625,
0.48486328125,
0.46728515625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.4892578125,
0.8935546875,
0.4892578125,
0.8935546875,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.08087158203125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.08087158203125,
0.5166015625,
0.46728515625,
0.5166015625,
0.46728515625,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.5048828125,
0.52099609375,
0.8896484375,
0.52099609375,
0.8896484375,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.08087158203125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.50244140625,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50244140625,
0.548828125
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.56005859375,
0.46728515625,
0.56005859375,
0.46728515625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.56494140625,
0.89208984375,
0.564453125,
0.89208984375,
0.57763671875,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.08087158203125,
0.59228515625,
0.46728515625,
0.59228515625,
0.46728515625,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.50244140625,
0.59521484375,
0.8388671875,
0.59521484375,
0.8388671875,
0.609375,
0.50244140625,
0.609375
],
[
0.080078125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.080078125,
0.619140625
],
[
0.080078125,
0.62060546875,
0.4697265625,
0.62060546875,
0.4697265625,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.50341796875,
0.62060546875,
0.6708984375,
0.62060546875,
0.6708984375,
0.63525390625,
0.50341796875,
0.63525390625
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.5009765625,
0.64306640625,
0.89208984375,
0.64453125,
0.89208984375,
0.6591796875,
0.5009765625,
0.65771484375
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.66015625,
0.8896484375,
0.66015625,
0.8896484375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.080078125,
0.68115234375,
0.468017578125,
0.68115234375,
0.468017578125,
0.6943359375,
0.080078125,
0.6943359375
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.6962890625,
0.468017578125,
0.6962890625,
0.468017578125,
0.71044921875,
0.08087158203125,
0.71044921875
],
[
0.080078125,
0.7109375,
0.468017578125,
0.7109375,
0.468017578125,
0.724609375,
0.080078125,
0.724609375
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.7265625,
0.338134765625,
0.7265625,
0.338134765625,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.5048828125,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.50341796875,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.09234619140625,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.09234619140625,
0.77880859375
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.09234619140625,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.09234619140625,
0.79443359375
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.0931396484375,
0.79736328125,
0.43310546875,
0.79736328125,
0.43310546875,
0.80810546875,
0.0931396484375,
0.80810546875
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.07843017578125,
0.8095703125,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.07843017578125,
0.82373046875
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.09234619140625,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.09234619140625,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.09234619140625,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.09234619140625,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.09234619140625,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.09234619140625,
0.869140625
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.0906982421875,
0.8720703125,
0.393798828125,
0.8720703125,
0.393798828125,
0.88525390625,
0.0906982421875,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"modal methods fail to perform well under unseen deploy-",
"tion (ReGrad) for MMDG, a strategy that can monitor",
"ment conditions. As illustrated in Fig. 1, though the cur-",
"and rebalance the convergence speed of each modality by",
"rent methods generalize well in unimodal scenarios, their",
"adaptively modulating gradients of U-Adapters.",
"performance abnormally declines when extended and ap-",
".We build the first large-scale benchmark to evaluate the",
"plied to multi-modal FAS. Why using more modalities is",
"domain generalizability of multi-modal FAS approaches.",
"not better? We argue that there are two main reasons:",
"Extensive experiments on this benchmark demonstrate",
"() Modality unreliability. Deploying in unseen environ-",
"the superiority and generalizability of MMDG",
"ments or with various kinds of models of imaging sen-",
"sors within the same modality, especially the depth and in-",
"2.Related Works",
"frared shown in Fig. 1 (b), might introduce significant do-",
"main shifts [66]. Such shifts can result in unreliable multi-",
"2.1.Domain Generalization for FAS",
"modal live/spoof feature extraction [51], misleading mutual",
"In FAS, domain generalization focuses on training a model",
"modalities through cross-modal fusion.(2) Modality im-",
"using multiple source domains, with the intention that",
"balance. In some multi-modal learning scenarios, models",
"this model can work effectively on unseen target domains",
"tend to overly rely on a dominant modality with the high-",
"[18, 46]. Previous works verify the effectiveness of adver-",
"estconvergencespeed (i.e.fastmodality),whichprevents",
"sarial training [22, 37, 65], asymmetric triplet loss [20, 30]",
"itself from fully exploiting the other dominated modalities",
"and controversial learning [38, 49] in developing a shared",
"withrelativelyslowerconvergencespeeds(i.e.slowmodal-",
"feature space across domains.Besides,some studies[50,",
"ities) [43]. In FAS, modality imbalance can cause signif-",
"69, 70] investigate the style information of different do-",
"icant negative impacts, especially when the fast modality",
"mains,aiming tolearn domain-invariant features by min-",
"is unsuitableorunavailablefordetecting attacksinthetar-",
"imizing distinct style characteristics.Furthermore,meta-",
"get domain, relying on slow modalities that have not ade-",
"learning-based methods [4, 10, 21, 44, 68] are also dis-",
"quately converged cannot achieve satisfactory performance",
"cussed to simulate domain shifts during training, fostering",
"This study addresses the cross-domain limitations of multi-",
"the learning of a robust representative feature space to such",
"modal FAS approaches due to modality unreliability and",
"variations. More recent works focus on parameter-efficient",
"imbalance. As illustrated in Fig. 2, we introduce a Multi-",
"transfer learning [5, 6, 18, 48], enabling a pretrained ViT",
"Modal Domain Generalized (MMDG) framework for FAS.",
"to efficiently adapt to unseen domains with a lower risk of",
"Within MMDG, we develop trainable Uncertainty-guided",
"over-fitting. However, these approaches are proposed for",
"cross-Adapters(U-Adapter)tofine-tuneVisionTransform-",
"unimodal(i.e.RGB)FAS.Duetotheignoranceofmodal-",
"ers (ViT) [9] The U-Adapter address modality unreliabil-",
"ity unreliability, there is a gap in adapting them to multi-",
"ity leveraging the uncertainty of each modality to suppress",
"modal scenarios. Simple multi-modal extensions of these",
"the focus on unreliable tokens during the cross-modal fu-",
"approaches, through early fusion, late fusion, or even the",
"sion. This prevents the excessive propagation of unreliable",
"substitution with multi-modal backbones, still exhibit lim-",
"spoofingtraces whichtends to decrease themodel'sdis-",
"ited performance under multi-modal DG scenarios.",
"criminative capability. Meanwhile, to solve modality im-",
"balance, we design Rebalanced modality Gradient modu-",
"2.2. Multi-Modal FAS",
"lation strategy (ReGrad). Our ReGrad rebalance the con-",
"Multi-modal FAS refers to the use of multiple spectra, e.g.,",
"vergence speed of all modalities by dynamically adjusting",
"RGB, depth, and infrared, to reveal live and spoofing traces.",
"the backward propagated gradients of all trainable parame-",
"ters in U-Adapters based on their conflict degree and con-",
"Due to the complementary information among different",
"modalities, traces that are undetectable in one modality can",
"vergence speed. This ensures that all modalities can be fully",
"leveraged to resist various unseen attacks in the target do-",
"be easily captured in others [17, 36]. To fully exploit infor-",
"mation from each modality, early methods employ channel-",
"main. Our contributions are as follows.",
"wise concatenation of inputted modalities14.42l,or de-",
". We propose the MMDG framework to enhance the do-",
"velop multiple separate branches for modality feature ex-",
"main generalizability by addressing the modality unreli-",
"traction followed by late fusion [25, 26, 47, 58]. Re-",
"ability and imbalance. This is the first investigation for",
"cent works introduce attention-based feature fusion tech-",
"multi-modal FASundermulti-sourceDG scenarios.",
"niques [8, 27] and adaptive cross-modal loss functions [13]",
". Within MMDG, we propose a Uncertainty-Guided Cross-",
"to encourage the extraction of complementary information",
"Adapter (U-Adapter) for ViT-based multi-modal FAS,",
"among modalities. Meanwhile, cross-modality translation",
"which can complementarily fuse cross-modal features",
"is discussed [29, 35] to mitigate the semantic gap between",
"and prevents each modality from being influenced by un-",
"different modalities. To further enhance the practical usabil-",
"reliable information caused by domain shifts.",
"ity of FAS, Yu et al. [60, 61, 62] introduce flexible-modal",
": We design the Rebalanced Modality Gradient Modula-",
"benchmarks that force the use of incomplete modalities dur-"
] |
{
"arxiv_id": "2402.19298",
"language": "en",
"page_idx": 1
}
|
||
2004.11980-en-1
|
2004.11980-en-1.jpg
|
SIGIR'20, July 25-30, 2020, Xi'an, China
goal is to match experts with pertinent questions: the answering
itself is not by the machine but by humans.
3 RELEVANCE TO THE IR COMMUNITY
Related tutorials. A tutorial on QA is not really new to SIGIR: the previous one was presented in 2016 by Wen-tau Yih and Hao Ma [98] (also at NAACL 2016 [97], by the same authors). Textand Dan Moldovan) [41] and EACL 2003 (Jimmy Lin and Boris Katz) [54]. Tutorials on IBM Watson [6,36] and entity recommendation [58] have also touched upon QA in the past. Recent workshops on various aspects of QA have been organized at top-tier conferences: MRQA (EMNLP-IJCNLP 2019), RCQA (AAAI 2019), HQA (WWW 2018), and OKBQA (COLING 2016).
Need for a new one. The unparalleled growth of QA warrants a new tutorial to cover recent advances in the field. Primarily a direction dominated by template-based [2,83] approaches, QA now includes a large number of neural methods [16,17,20,44], graph- $based[56,57]$ methods, and even a handful that explore reinforcement learning [25,61,63]. Across sub-fields, more complex questions are being handled, complexity being defined in terms of entities and relationships present [56,84,96]. Systems are moving from their static counterparts to more interactive ones: an increasing number of systems are including scope for $userfeedback\left[2,50,100\right]$ and operate in a multi-turn, conversational setting [19,61,68,73]. Interpretability or explainability of presented answers is yet another area of significance [1,69,79,88], as the role of such explanations is being recognized both for developers and end-users towards system improvement and user satisfaction. The tutorial will emphasize each of these key facets of QA. In addition, a summary of the available $benchmarks[3,11,18,19,65,68,80,96]$ in each of the QA sub-fields will be provided, that would be very valuable for new entrants to get started with their problem of choice.
# 4 TOPICS
## 4.1 $\textcircled {\text{A over Knowledge Graphs}}$
The advent oflarge knowledge graphs like Freebase [13], YAGO [74], DBpedia [8] and Wikidata [86] gave rise to QA over KGs (KG-QA) that typically provides answers as single or lists of entities from the KG. KG-QA has become an important research direction, where the goal is to translate a natural language question into a structured query, typically in the Semantic Web language SPARQL or an equivalent logical form, that directly operates on the entities and predicates of the underlying KG [12,19,63,84,88]. KG-QA involves challenges of entity disambiguation and, most strikingly, the need to bridge the vocabulary gap between the phrases in a question and the terminology of the KG. Early work on KG-QA built on paraphrase-based mappings and query templates that involve a single entity predicate [11, 47, 83, 93, 94]. This line was further advanced by [4,9,10,42],including the learning of templates from graph patterns in the KG. However, reliance on templates prevents such approaches from robustly coping with arbitrary syntactic formulations. This has motivated deep learning methods with CNNs and LSTMs, and especially key-value memory networks [17, 44, 82, 91, 92].
Rishiraj Saha Roy and Avishek Anand
A significant amount in this section of the tutorial will be on answering complex questions with multiple entities and predicates. This is one of the key focus areas in KG-QA now [12,15,30,43,46, 56,63,84], where the overriding principle is often the identification of frequent query substructures. Web tables represent a key aspect of the curated Web and contain a substantial volume of structured information [31]. QA over such tables contains canonicalized rep-
Ma [98] (also at NAACL 2016 [97], by the same authors). Text- resentatives of several challenges faced in large-scale KG-QA, and based QA tutorials appeared way back in NAACL 2001 (Sanda Harabagiu we will touch upon a few key works in this area [45,62,77
## 4.2 QA over Text
4.2.1 Early efforts. Question answering has originally considered textual document collections as its underlying source. Classical approaches [67, 85] extracted answers from passages and short text units that matched most cue words from the question followed by statistical scoring. This passage-retrieval model makes intensive use of IR techniques for statistical scoring of sentences or passages and aggregation of evidence for answer candidates. TREC ran a QA benchmarking series from 1999 to 2007, and more recently revived it as the LiveQA [5] and Complex Answer Retrieval (CAR) tracks [29]. IBM Watson [37] extended this paradigm by combining it with learned models for special question types.
4.2.2 Machine reading comprehension (MRC). This is a QA variation where a question needs to be answered as a short span of words from a given text paragraph [65,96], and is different from the typical fact-centric answer-finding task in IR. Exemplary approaches in MRC that extended the original single-passage setting to a multi-document one can be found in DrQA[16] and DocumentQA [20] (among many, many others). Traditional fact-centric QA over text, and multi-document MRC are recently emerging as a joint topic referred to as open-domain QA [28,55,87].
4.2.3 Open-domain QA. In NLP, open-domain question answering is now a benchmark task in natural language understanding (NLU) and can potentially drive the progress of methods in this area [51]. The recent reprisal of this task was jump-started by QA bench- $\hat{\text{marks like SQuAD[65] and HotpotQA[96], that were proposed for}}$ MRC. Consequently, a majority of the approaches in NLP focus on MRC-style question answering with varying task complexities [27, 32,51,81]. This has lead to the common practice of considering the open-domain QA task as a retrieve and re-rank task. In this tutorial, we will introduce the modern foundations of open-domain QA using a similar retrieve-and-rank framework. Note that our focus will not be on architectural engineering but rather on design decisions, task complexity, and the roles and opportunities for IR.
4.3 QA over Heterogeneous Sources
Limitations of QA over KGs has recently led to a revival of considering textual sources, in combination with KGs [71,72,75,76,92]. Early methods like PARALEX [34] and OQA [35] supported noisy KGs in the form of triple spaces compiled via Open IE[59] on Wikipedia articles or Web corpora. TupleInf [49] extended and generalized the Open-IE-based PARALEX approach to complex questions, and is geared for multiple-choice answer options. TAQA [99] is another generalization of Open-IE-based QA, by constructing a KG of $n$-tuples from Wikipedia full-text and question-specific
2
|
[
[
0.08660888671875,
0.07574462890625,
0.27685546875,
0.07574462890625,
0.27685546875,
0.08587646484375,
0.08660888671875,
0.08587646484375
],
[
0.734375,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.734375,
0.08709716796875
],
[
0.08660888671875,
0.10797119140625,
0.48193359375,
0.10797119140625,
0.48193359375,
0.121826171875,
0.08660888671875,
0.121826171875
],
[
0.53369140625,
0.106689453125,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.53369140625,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.346435546875,
0.121826171875,
0.346435546875,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.515625,
0.121826171875,
0.9140625,
0.12060546875,
0.9140625,
0.1337890625,
0.515625,
0.1351318359375
],
[
0.51708984375,
0.1351318359375,
0.9140625,
0.1351318359375,
0.9140625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.10540771484375,
0.154052734375,
0.43310546875,
0.1546630859375,
0.43310546875,
0.169189453125,
0.10540771484375,
0.1685791015625
],
[
0.51708984375,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.08660888671875,
0.1572265625,
0.111083984375,
0.1572265625,
0.111083984375,
0.1685791015625,
0.08660888671875,
0.1685791015625
],
[
0.51708984375,
0.1634521484375,
0.91259765625,
0.1634521484375,
0.91259765625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.0849609375,
0.173583984375,
0.48193359375,
0.173583984375,
0.48193359375,
0.1875,
0.0849609375,
0.1875
],
[
0.51708984375,
0.1767578125,
0.91259765625,
0.1767578125,
0.91259765625,
0.1900634765625,
0.51708984375,
0.1900634765625
],
[
0.08660888671875,
0.189453125,
0.48046875,
0.189453125,
0.48046875,
0.199462890625,
0.08660888671875,
0.199462890625
],
[
0.515625,
0.1900634765625,
0.9140625,
0.1912841796875,
0.9140625,
0.2052001953125,
0.515625,
0.203857421875
],
[
0.0849609375,
0.2020263671875,
0.48046875,
0.2020263671875,
0.48046875,
0.21533203125,
0.0849609375,
0.21533203125
],
[
0.51708984375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51708984375,
0.2183837890625
],
[
0.08251953125,
0.212158203125,
0.8798828125,
0.2171630859375,
0.8798828125,
0.23486328125,
0.08251953125,
0.2298583984375
],
[
0.08660888671875,
0.2310791015625,
0.48193359375,
0.2310791015625,
0.48193359375,
0.2412109375,
0.08660888671875,
0.2412109375
],
[
0.08660888671875,
0.24365234375,
0.48046875,
0.24365234375,
0.48046875,
0.2568359375,
0.08660888671875,
0.2568359375
],
[
0.5185546875,
0.246826171875,
0.54833984375,
0.246826171875,
0.54833984375,
0.2587890625,
0.5185546875,
0.2587890625
],
[
0.55615234375,
0.24560546875,
0.66650390625,
0.24560546875,
0.66650390625,
0.260009765625,
0.55615234375,
0.260009765625
],
[
0.08660888671875,
0.2587890625,
0.48193359375,
0.2587890625,
0.48193359375,
0.26953125,
0.08660888671875,
0.26953125
],
[
0.51953125,
0.26513671875,
0.55224609375,
0.26513671875,
0.55224609375,
0.276611328125,
0.51953125,
0.276611328125
],
[
0.55615234375,
0.264404296875,
0.9140625,
0.264404296875,
0.9140625,
0.279052734375,
0.55615234375,
0.279052734375
],
[
0.08660888671875,
0.272705078125,
0.48046875,
0.272705078125,
0.48046875,
0.28271484375,
0.08660888671875,
0.28271484375
],
[
0.51708984375,
0.279052734375,
0.91259765625,
0.279052734375,
0.91259765625,
0.292236328125,
0.51708984375,
0.292236328125
],
[
0.0849609375,
0.285400390625,
0.48193359375,
0.285400390625,
0.48193359375,
0.298583984375,
0.0849609375,
0.298583984375
],
[
0.5185546875,
0.292236328125,
0.91259765625,
0.292236328125,
0.91259765625,
0.30615234375,
0.5185546875,
0.30615234375
],
[
0.0849609375,
0.298583984375,
0.384033203125,
0.298583984375,
0.384033203125,
0.3125,
0.0849609375,
0.3125
],
[
0.0849609375,
0.3125,
0.48193359375,
0.3125,
0.48193359375,
0.32568359375,
0.0849609375,
0.32568359375
],
[
0.5185546875,
0.30810546875,
0.9111328125,
0.30810546875,
0.9111328125,
0.318115234375,
0.5185546875,
0.318115234375
],
[
0.0849609375,
0.32568359375,
0.483642578125,
0.32568359375,
0.483642578125,
0.339111328125,
0.0849609375,
0.339111328125
],
[
0.5185546875,
0.3212890625,
0.91015625,
0.3212890625,
0.91015625,
0.33203125,
0.5185546875,
0.33203125
],
[
0.5185546875,
0.333984375,
0.91259765625,
0.333984375,
0.91259765625,
0.347900390625,
0.5185546875,
0.347900390625
],
[
0.0849609375,
0.34033203125,
0.48193359375,
0.34033203125,
0.48193359375,
0.354248046875,
0.0849609375,
0.354248046875
],
[
0.5185546875,
0.347900390625,
0.9140625,
0.347900390625,
0.9140625,
0.361083984375,
0.5185546875,
0.361083984375
],
[
0.0849609375,
0.354248046875,
0.48193359375,
0.354248046875,
0.48193359375,
0.367431640625,
0.0849609375,
0.367431640625
],
[
0.5185546875,
0.361083984375,
0.91259765625,
0.361083984375,
0.91259765625,
0.375,
0.5185546875,
0.375
],
[
0.08660888671875,
0.367431640625,
0.48193359375,
0.367431640625,
0.48193359375,
0.380615234375,
0.08660888671875,
0.380615234375
],
[
0.08660888671875,
0.380615234375,
0.48046875,
0.380615234375,
0.48046875,
0.39453125,
0.08660888671875,
0.39453125
],
[
0.5185546875,
0.376953125,
0.91259765625,
0.376953125,
0.91259765625,
0.386962890625,
0.5185546875,
0.386962890625
],
[
0.5185546875,
0.389404296875,
0.91259765625,
0.389404296875,
0.91259765625,
0.40283203125,
0.5185546875,
0.40283203125
],
[
0.08660888671875,
0.395751953125,
0.48193359375,
0.395751953125,
0.48193359375,
0.4091796875,
0.08660888671875,
0.4091796875
],
[
0.51708984375,
0.404052734375,
0.83837890625,
0.404052734375,
0.83837890625,
0.417236328125,
0.51708984375,
0.417236328125
],
[
0.08660888671875,
0.410888671875,
0.48046875,
0.410888671875,
0.48046875,
0.421630859375,
0.08660888671875,
0.421630859375
],
[
0.0849609375,
0.423583984375,
0.48193359375,
0.423583984375,
0.48193359375,
0.4375,
0.0849609375,
0.4375
],
[
0.51708984375,
0.4248046875,
0.9140625,
0.423583984375,
0.9140625,
0.438232421875,
0.51708984375,
0.439453125
],
[
0.08660888671875,
0.438232421875,
0.4794921875,
0.438232421875,
0.4794921875,
0.448974609375,
0.08660888671875,
0.448974609375
],
[
0.5185546875,
0.44140625,
0.9140625,
0.44140625,
0.9140625,
0.451904296875,
0.5185546875,
0.451904296875
],
[
0.08660888671875,
0.45263671875,
0.4794921875,
0.45263671875,
0.4794921875,
0.46337890625,
0.08660888671875,
0.46337890625
],
[
0.5185546875,
0.453857421875,
0.9140625,
0.453857421875,
0.9140625,
0.46728515625,
0.5185546875,
0.46728515625
],
[
0.0849609375,
0.464111328125,
0.48193359375,
0.464111328125,
0.48193359375,
0.477783203125,
0.0849609375,
0.477783203125
],
[
0.51708984375,
0.46728515625,
0.91259765625,
0.46728515625,
0.91259765625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.08660888671875,
0.479736328125,
0.48046875,
0.479736328125,
0.48046875,
0.490478515625,
0.08660888671875,
0.490478515625
],
[
0.51708984375,
0.48095703125,
0.9140625,
0.481689453125,
0.9140625,
0.49560546875,
0.51708984375,
0.494384765625
],
[
0.0849609375,
0.492431640625,
0.484375,
0.492431640625,
0.484375,
0.505859375,
0.0849609375,
0.505859375
],
[
0.51708984375,
0.49560546875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.51708984375,
0.5087890625
],
[
0.0849609375,
0.505859375,
0.483642578125,
0.505859375,
0.483642578125,
0.51953125,
0.0849609375,
0.51953125
],
[
0.51708984375,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51708984375,
0.52294921875
],
[
0.08660888671875,
0.521484375,
0.48046875,
0.521484375,
0.48046875,
0.5322265625,
0.08660888671875,
0.5322265625
],
[
0.515625,
0.521484375,
0.9140625,
0.52294921875,
0.9140625,
0.537109375,
0.515625,
0.5361328125
],
[
0.0849609375,
0.5341796875,
0.483642578125,
0.5341796875,
0.483642578125,
0.54736328125,
0.0849609375,
0.54736328125
],
[
0.5185546875,
0.5390625,
0.8544921875,
0.5390625,
0.8544921875,
0.5498046875,
0.5185546875,
0.5498046875
],
[
0.0849609375,
0.54736328125,
0.48193359375,
0.54736328125,
0.48193359375,
0.56103515625,
0.0849609375,
0.56103515625
],
[
0.08660888671875,
0.56298828125,
0.426513671875,
0.56298828125,
0.426513671875,
0.57373046875,
0.08660888671875,
0.57373046875
],
[
0.51708984375,
0.55810546875,
0.91259765625,
0.5595703125,
0.91259765625,
0.57373046875,
0.51708984375,
0.57275390625
],
[
0.51708984375,
0.57373046875,
0.91259765625,
0.57373046875,
0.91259765625,
0.5869140625,
0.51708984375,
0.5869140625
],
[
0.0849609375,
0.5947265625,
0.1805419921875,
0.5947265625,
0.1805419921875,
0.609375,
0.0849609375,
0.609375
],
[
0.5185546875,
0.5888671875,
0.91259765625,
0.5888671875,
0.91259765625,
0.599609375,
0.5185546875,
0.599609375
],
[
0.51708984375,
0.6015625,
0.9140625,
0.6015625,
0.9140625,
0.61572265625,
0.51708984375,
0.61572265625
],
[
0.08660888671875,
0.61572265625,
0.351318359375,
0.61572265625,
0.351318359375,
0.6298828125,
0.08660888671875,
0.6298828125
],
[
0.51708984375,
0.61572265625,
0.9140625,
0.61572265625,
0.9140625,
0.62890625,
0.51708984375,
0.62890625
],
[
0.0849609375,
0.63330078125,
0.490234375,
0.63330078125,
0.490234375,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.5185546875,
0.630859375,
0.91259765625,
0.630859375,
0.91259765625,
0.6416015625,
0.5185546875,
0.6416015625
],
[
0.0849609375,
0.646484375,
0.48193359375,
0.646484375,
0.48193359375,
0.66015625,
0.0849609375,
0.66015625
],
[
0.51708984375,
0.64306640625,
0.9140625,
0.64306640625,
0.9140625,
0.6572265625,
0.51708984375,
0.6572265625
],
[
0.08416748046875,
0.66015625,
0.48193359375,
0.6611328125,
0.48193359375,
0.6748046875,
0.08416748046875,
0.673828125
],
[
0.5185546875,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.66943359375,
0.5185546875,
0.66943359375
],
[
0.08660888671875,
0.6767578125,
0.48046875,
0.6767578125,
0.48046875,
0.6875,
0.08660888671875,
0.6875
],
[
0.51953125,
0.67236328125,
0.91259765625,
0.67236328125,
0.91259765625,
0.68310546875,
0.51953125,
0.68310546875
],
[
0.0849609375,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.51708984375,
0.6845703125,
0.9140625,
0.6845703125,
0.9140625,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.0849609375,
0.70263671875,
0.48193359375,
0.70263671875,
0.48193359375,
0.71630859375,
0.0849609375,
0.71630859375
],
[
0.51708984375,
0.69873046875,
0.91259765625,
0.69873046875,
0.91259765625,
0.7119140625,
0.51708984375,
0.7119140625
],
[
0.08660888671875,
0.71826171875,
0.48046875,
0.71826171875,
0.48046875,
0.72900390625,
0.08660888671875,
0.72900390625
],
[
0.5185546875,
0.71337890625,
0.91259765625,
0.71337890625,
0.91259765625,
0.7236328125,
0.5185546875,
0.7236328125
],
[
0.0849609375,
0.73095703125,
0.48046875,
0.73095703125,
0.48046875,
0.744140625,
0.0849609375,
0.744140625
],
[
0.51708984375,
0.72607421875,
0.8857421875,
0.72607421875,
0.8857421875,
0.7392578125,
0.51708984375,
0.7392578125
],
[
0.08416748046875,
0.74365234375,
0.48193359375,
0.744140625,
0.48193359375,
0.75830078125,
0.08416748046875,
0.7568359375
],
[
0.51953125,
0.7548828125,
0.56201171875,
0.7548828125,
0.56201171875,
0.7666015625,
0.51953125,
0.7666015625
],
[
0.5556640625,
0.75390625,
0.81884765625,
0.75390625,
0.81884765625,
0.76806640625,
0.5556640625,
0.76806640625
],
[
0.0849609375,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.78515625,
0.0849609375,
0.78515625
],
[
0.0849609375,
0.75830078125,
0.483642578125,
0.75830078125,
0.483642578125,
0.77294921875,
0.0849609375,
0.77294921875
],
[
0.5185546875,
0.771484375,
0.91259765625,
0.771484375,
0.91259765625,
0.78515625,
0.5185546875,
0.78515625
],
[
0.0849609375,
0.78515625,
0.48193359375,
0.78515625,
0.48193359375,
0.798828125,
0.0849609375,
0.798828125
],
[
0.5185546875,
0.78515625,
0.9140625,
0.78515625,
0.9140625,
0.798828125,
0.5185546875,
0.798828125
],
[
0.0849609375,
0.7998046875,
0.48046875,
0.7998046875,
0.48046875,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.5185546875,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.5185546875,
0.81298828125
],
[
0.08660888671875,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.8271484375,
0.08660888671875,
0.8271484375
],
[
0.5185546875,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.5185546875,
0.8271484375
],
[
0.08660888671875,
0.8291015625,
0.48193359375,
0.8291015625,
0.48193359375,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8291015625,
0.91259765625,
0.8291015625,
0.91259765625,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.483642578125,
0.85498046875,
0.483642578125,
0.869140625,
0.0849609375,
0.869140625
],
[
0.5185546875,
0.85693359375,
0.9111328125,
0.85693359375,
0.9111328125,
0.8671875,
0.5185546875,
0.8671875
],
[
0.08416748046875,
0.8671875,
0.483642578125,
0.86865234375,
0.483642578125,
0.88330078125,
0.08416748046875,
0.8818359375
],
[
0.515625,
0.8662109375,
0.91259765625,
0.86865234375,
0.91259765625,
0.88330078125,
0.515625,
0.880859375
],
[
0.08660888671875,
0.8837890625,
0.17236328125,
0.8837890625,
0.17236328125,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
],
[
0.496826171875,
0.90283203125,
0.50341796875,
0.90283203125,
0.50341796875,
0.90771484375,
0.496826171875,
0.90771484375
]
] |
[
"SIGIR20,luly25-30.2020.Xi'an,China",
"Rishiraj Saha Roy and Avishek Anand",
"goal is to match experts with pertinent questions: the answering",
"A significant amount in this section of the tutorial will be on",
"itself is not by the machine but by humans.",
"answering complex questions with multiple entities and predicates.",
"This is one of the key focus areas in KG-QA now [12, 15, 30, 43, 46,",
"RELEVANCE TO THEIR COMMUNITY",
"56, 63,84], where the overriding principle is often theidentification",
"3",
"of frequent query substructures. Web tables represent a key aspect",
"Related tutorials. A tutorial on QA is not really new to SIGIR",
"of the curated Web and contain a substantial volume of structured",
"thepreviousonewaspresented in20i6byWen-tauYihandHao",
"information [31]. QA over such tables contains canonicalized rep-",
"Ma [98] (also at NAACL 2016 [97], by the same authors). Text+",
"resentatives of several challenges faced in large-scale KG-QA, and",
"based QA tutorials appeared way back in NAACL2001 (Sanda Harabagiu we will touch upon a few key works in this area [45, 62,77].",
"andDanMoldovan)[41landEACL2003(JimmyLinandBoris",
"Katz) [54]. Tutorials on IBM Watson [6, 36] and entity recommen-",
"4.2",
"QA over Text",
"dation[58]havealsotoucheduponQAinthepast.Recentwork-",
"4.2.1",
"Early efforts. Question answering has originally considered",
"shops on various aspects of OA have been organized attop-tier",
"textual document collections as its underlying source. Classical ap",
"conferences: MRQA (EMNLP-IJCNLP 2019), RCQA (AAAI 2019),",
"proaches [67, 85] extracted answers from passages and short text",
"HQA (WWW 2018), and OKBQA (COLING 2016).",
"Need for a new one. The unparalleled growth of QA warrants a",
"unitsthatmatched mostcuewordsfromthequestionfollowed by",
"new tutorial to cover recent advances in the field. Primarily a di-",
"statistical scoring.This passage-retrieval model makes intensive",
"use ofIR techniques for statistical scoring of sentences or passages",
"rection dominated by template-based [2, 83] approaches, QA now",
"and aggregation of evidence for answer candidates. TREC ran a",
"includes a large number of neural methods [16, 17, 20, 44], graph-",
"QA benchmarking series from 1999 to 2007, and more recently re",
"based [56, 57] methods, and even a handful that explore reinforce-",
"ment learning [25, 61, 63]. Across sub-fields, more complex ques",
"vived itastheLiveOA[5]and ComplexAnswerRetrieval (CAR)",
"tracks [29].IBM Watson [37] extended this paradigm by combin-",
"tions are being handled, complexity being defined in terms of enti-",
"ing it with learned models for special question types.",
"tiesandrelationshipspresent[56,84,96].Systems aremovingfrom",
"their static counterparts to more interactive ones: an increasing",
"4.2.2Machine reading comprehension (MRC).This is a QA vari-",
"numberofsystemsareincluding scopeforuserfeedback[2,50,100]",
"ationwhereaquestionneedstobeansweredas ashortspanof",
"and operate in a multi-turn, conversational setting[19,61, 68,73]",
"words from a given text paragraph [65, 96], and is different from",
"Interpretability or explainability of presented answers is yet an",
"the typical fact-centric answer-finding task in IR. Exemplary ap-",
"other area of significance[1, 69, 79, 88], as therole of such expla",
"proaches in MRC that extended the original single-passage setting",
"nations is being recognized both for developers and end-users to-",
"to a multi-document one can be found in DrQA [16] and Docu-",
"wards system improvement and user satisfaction. The tutorial will",
"mentQA [20] (among many, many others). Traditional fact-centric",
"emphasize eachofthesekeyfacetsofQA.Inaddition,a summary",
"QA over text, and multi-document MRC are recently emerging as",
"of the available benchmarks [3, 11, 18, 19, 65, 68, 80, 96] in each of",
"a jointtopicreferredtoasopen-domainQA[28,55,87].",
"the QA sub-fields will be provided, that would be very valuable for",
"newentrantstogetstarted withtheirproblem of choice.",
"4.2.3 Open-domain QA. In NLP, open-domain question answering",
"is now a benchmark task in natural language understanding (NLU)",
"4TOPICS",
"and can potentially drive theprogress ofmethods in this area[51]]",
"The recent reprisal of this task was jump-started by QA bench-",
"4.1 QA overKnowledge Graphs",
"marks like SQuAD [65] and HotpotQA [96], that were proposed for",
"The advent oflarge knowledge graphs like Freebase [13], YAGO [74],",
"MRC.Consequently,amajority of the approaches inNLPfocus on",
"DBpedia [8] and Wikidata [86] gave rise to QA over KGs (KG-QA)",
"MRC-style question answering with varying task complexities [27,",
"thattypically provides answers as single or lists of entitiesfromthe",
"32, 51, 81]. This has lead to the common practice of considering the",
"KG.KG-QA has become an important research direction, where",
"open-domain QA task as a retrieve and re-rank task. In this tuto-",
"the goal is to translate a natural language question into a struc-",
"rial, we will introduce the modern foundations of open-domain QA",
"tured query, typically in the Semantic Web language SPARQL or",
"using a similar retrieve-and-rank framework. Note that our focus",
"an equivalent logical form, that directly operateson theentities",
"will notbeonarchitectural engineeringbutratherondesigndeci-",
"and predicates of the underlying KG [12, 19, 63, 84, 88]. KG-QA",
"sions, task complexity, and the roles and opportunities for IR.",
"involves challenges ofentity disambiguation and,most strikingly,",
"4.3",
"QA over Heterogeneous Sources",
"question and the terminology of the KG. Early work on KG-QA",
"the need to bridge the vocabulary gap between the phrases in a",
"Limitations of QA over KGs has recently led to a revival of consid",
"built on paraphrase-based mappings and query templates that in-",
"ering textual sources, in combination with KGs [71, 72, 75, 76, 92].",
"volve a single entity predicate [11, 47, 83, 93, 94]. This line was fur",
"Early methods like PARALEX [34] and OQA [35] supported noisy",
"ther advanced by [4, 9, 10, 42], including the learning of templates",
"KGs in the form of triple spaces compiled via Open IE [59] on",
"fromgraphpatternsintheKG.However,relianceontemplatespre-",
"WikipediaarticlesorWebcorpora.Tuplelnf[49]extendedandgen-",
"vents such approaches from robustly coping with arbitrary syntac-",
"eralized the Open-IE-based PARALEX approach to complex ques",
"tic formulations. This has motivated deep learning methods with",
"tions,and is geared for multiple-choiceanswer options.TAQA[99]",
"CNNs and LSTMs, and especially key-value memory networks [17,",
"is another generalization of Open-IE-based QA, by constructing",
"44, 82, 91, 92].",
"a KG of n-tuples from Wikipedia full-text and question-specific",
"2"
] |
{
"arxiv_id": "2004.11980",
"language": "en",
"page_idx": 1
}
|
||
2004.11980-en-2
|
2004.11980-en-2.jpg
|
Question Answering over Curated and Open Web Sources
search results. SplitQA [80] addressed complex questions by decomposing them into a sequence of simple questions, and relies on crowdsourced training data. Some methods for hybrid QA start with KGs as a source for candidate answers and use text corpora like Wikipedia or ClueWeb as additional evidence [26,75,76,79,89, 92], or start with answer sentences from text corpora and combine these with KGs for giving crisp entity answers [71,78].
# 4.4 New Horizons in QA
4.4.1 Conversational QA. Conversational QA involves a sequence of questions and answers that appear as a natural dialogue between the system and the user. The aim of such sequential, multi-turn QA is to understand the context left implicit by users and effectively answer incomplete and ad hoc follow-up questions. Towards this, various recent benchmarks have been proposed that expect answers that are boolean [69], extractive [18] and free-form responses [68], entities [19], passages [48], and chit-chat [101]. Leaderboards of the QuAC [18] and CoQA [68] datasets point to many recent approaches in the text domain. Recently, the TREC CAsT track [24] (http://www.treccast.ai/) and the Dagstuhl Seminar on Conversational Search [7] tried to address such challenges in conversational search. For KG-QA, notable efforts include [19, 39, 70, 73]. We will focus on the modelling complexity of these tasks and a classification of the approaches involved.
4.4.2 Feedback and interpretability. Static learning systems for QA are gradually paving the way for those that incorporate user $feed-$ $back.$These mostly design the setup as a continuous learning setup, where the explicit user feedback mechanism is built on top of an existing QA system. For example, the NEQA [2], QApedia [50], and IMPROVE-QA [100] systems primarily operate on the core systems of QUINT [4], DrQA [16], and gAnswer [42], respectively. A direction closely coupled with effective feedback is the interpretability of QA models, that is also essential to improving trust and satisfaction [1,88]. This section is for experts and we will discuss potential limitations and open challenges.
4.4.3 Clarification questions. A key aspect of mixed-initiative systems [64] is to be able to ask clarifications. This ability is essential to QA systems, especially for handling ambiguity and facilitating better question understanding. Most of the work in this domain has been driven by extracting tasks from open QA forums [14,66,90].
# 5 FORMAT AND SUPPORT
A detailed schedule for our proposed half-day tutorial (three hours plus breaks), which is aimed to meet a high-quality presentation within the chosen time period, is as follows:
$\bullet$ 9:00- 10:30 Part I: QA over knowledge graphs (1.5 hours)
- Background $( 15\textit{minutes})$ - Simple QA $( 10\textit{minutes})$ - Complex QA $(20minutes)$ - Heterogeneous QA $( 15\textit{minutes})$ - Conversational QA $( 15\textit{minutes})$ - Summary $( 15\textit{minutes})$ $\bullet10:30-11{:}00$ Coffee break
SIGIR'20, July 25-30, 2020, Xi'an, China
$\bullet$ 11:00- 12:30 Part II: QA over textual sources (1.5 hours)
- Background $( 15\textit{minutes})$
- Machine reading $( 20\textit{minutes})$ - Open-domain QA (20 minutes) - Feedback and interpretability (10 minutes)
- Conversational QA (10 minutes)
- Summary $( 15\textit{minutes})$
# 6 $PRESENTERS$
Rishiraj Saha Roy is a Senior Researcher at the Max Planck Institute for Informatics (MPII), Saarbrücken, Germany. He leads the research group on Question Answering, that focuses on robust and interpretable solutions for answering natural language questions over structured and unstructured data. He has more than four years of research experience on question answering. In recent years, he has served on the PCs of conferences like SIGIR, CIKM, AAAI, EMNLP, and NAACL, and published at venues like SIGIR, CIKM, WSDM, WWW, and NAACL.
Prior to joining MPII, he worked for one and a half years as a Computer Scientist at Adobe Research, Bangalore, India. He completed his PhD as a Microsoft Research India Fellow from the Indian Institute of Technology (IIT) Kharagpur. More information about his research can be found at http://people.mpi- inf.mpg.de/~rsaharo/.
Avishek Anand is an Assistant Professor at the Leibniz Universität Hannover, Germany. His research interests lie at the intersection of retrieval, mining, and machine learning. He did his PhD at the Max Planck Institute for Informatics, Saarbrücken, Germany, where he worked on temporal information retrieval. Currently, he is developing intelligent and transparent representation learning methods for text and graphs in the context of search, QA and Web tasks. He has published his research in several top-tier conferences, such as SIGIR, WSDM, CIKM, ICDE and TKDE.
He delivered a tutorial at SIGIR 2016, on Temporal Information Retrieval, and several talks in various summer schools and invited lectures. Avishek was one of the lead organizers of the recently concluded Dagstuhl seminar on Conversational Search.
# REFERENCES
[1] Abdalghani Abujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard
Weikum. 2017. QUINT: Interpretable Question Answering over Knowledge
Bases. In $EMNLP.$
[2] Abdalghani Abujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard
Weikum. 2018. Never-ending learning for open-domain question answering
over knowledge bases. In $WWW.$
[3] Abdalghani Äbujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard
Weikum. 2019. ComQA: A Community-sourced Dataset for Complex Factoid
Question Answering with Paraphrase Clusters. In NAACL-HLT'19.
[4] Abdalghani Abujabal, Mohamed Yahya, Mirek Riedewald, and Gerhard
Weikum. 2017. Automated template generation for question answering over
$\underset{\Gamma}{\operatorname*{\text{knowledge graphs. In }}}{\operatorname*{\text{WWW.}}}$
[5] Eugene Agichtein, David Carmel, Dan Pelleg, Yuval Pinter, and Donna Harman.
2015. Overview of the TREC 2015 LiveQA Track. In $TREC.$
[6] Aditya Kalyanpur Alfo Gliozzo and James Fan. 2012. Natural language pro-
cessing in Watson. In NAACL-HLT.
[7] Avishek Anand, Lawrence Cavedon, Hideo Joho, Mark Sanderson, and Benno
Stein. 2020. Conversational Search(Dagstuhl Seminar 19461). Dagstuhl Reports
9, 11 (2020).
[8] Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyga-
niak, and Zachary Ives. 2007. DBpedia: A nucleus for a Web of open data.
(2007).
[9] Junwei Bao, Nan Duan, Zhao Yan, Ming Zhou, and Tiejun Zhao. 2016
Constraint-based question answering with knowledge graph. In COLING.
|
[
[
0.0882568359375,
0.0770263671875,
0.359375,
0.0770263671875,
0.359375,
0.08709716796875,
0.0882568359375,
0.08709716796875
],
[
0.72216796875,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.72216796875,
0.08709716796875
],
[
0.0849609375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.54736328125,
0.10919189453125,
0.896484375,
0.10919189453125,
0.896484375,
0.11932373046875,
0.54736328125,
0.11932373046875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.5556640625,
0.1287841796875,
0.72216796875,
0.1287841796875,
0.72216796875,
0.1427001953125,
0.5556640625,
0.1427001953125
],
[
0.08660888671875,
0.136962890625,
0.48046875,
0.136962890625,
0.48046875,
0.147705078125,
0.08660888671875,
0.147705078125
],
[
0.0849609375,
0.1483154296875,
0.48046875,
0.1483154296875,
0.48046875,
0.1622314453125,
0.0849609375,
0.1622314453125
],
[
0.55615234375,
0.1427001953125,
0.7509765625,
0.1427001953125,
0.7509765625,
0.1558837890625,
0.55615234375,
0.1558837890625
],
[
0.0849609375,
0.1622314453125,
0.48046875,
0.1622314453125,
0.48046875,
0.175537109375,
0.0849609375,
0.175537109375
],
[
0.55615234375,
0.1572265625,
0.75634765625,
0.1572265625,
0.75634765625,
0.17041015625,
0.55615234375,
0.17041015625
],
[
0.55615234375,
0.17041015625,
0.826171875,
0.17041015625,
0.826171875,
0.184326171875,
0.55615234375,
0.184326171875
],
[
0.08660888671875,
0.1787109375,
0.48046875,
0.1787109375,
0.48046875,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.55615234375,
0.184326171875,
0.765625,
0.184326171875,
0.765625,
0.1976318359375,
0.55615234375,
0.1976318359375
],
[
0.0849609375,
0.1912841796875,
0.416748046875,
0.1912841796875,
0.416748046875,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.55810546875,
0.199462890625,
0.708984375,
0.199462890625,
0.708984375,
0.210205078125,
0.55810546875,
0.210205078125
],
[
0.0849609375,
0.2171630859375,
0.294189453125,
0.2171630859375,
0.294189453125,
0.2310791015625,
0.0849609375,
0.2310791015625
],
[
0.521484375,
0.23291015625,
0.53369140625,
0.23291015625,
0.53369140625,
0.2392578125,
0.521484375,
0.2392578125
],
[
0.54150390625,
0.2279052734375,
0.65869140625,
0.2266845703125,
0.65869140625,
0.2412109375,
0.54150390625,
0.242431640625
],
[
0.0882568359375,
0.2374267578125,
0.4794921875,
0.2374267578125,
0.4794921875,
0.2474365234375,
0.0882568359375,
0.2474365234375
],
[
0.0849609375,
0.25,
0.48193359375,
0.25,
0.48193359375,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.0849609375,
0.26318359375,
0.48193359375,
0.26318359375,
0.48193359375,
0.276611328125,
0.0849609375,
0.276611328125
],
[
0.51708984375,
0.260009765625,
0.91259765625,
0.260009765625,
0.91259765625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.08416748046875,
0.276611328125,
0.483642578125,
0.27783203125,
0.483642578125,
0.291748046875,
0.08416748046875,
0.290283203125
],
[
0.51708984375,
0.274658203125,
0.9140625,
0.274658203125,
0.9140625,
0.28857421875,
0.51708984375,
0.28857421875
],
[
0.0849609375,
0.291748046875,
0.483642578125,
0.291748046875,
0.483642578125,
0.304931640625,
0.0849609375,
0.304931640625
],
[
0.515625,
0.287353515625,
0.91259765625,
0.28857421875,
0.91259765625,
0.302978515625,
0.515625,
0.3017578125
],
[
0.0849609375,
0.30615234375,
0.48193359375,
0.30615234375,
0.48193359375,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.5185546875,
0.302978515625,
0.91259765625,
0.302978515625,
0.91259765625,
0.313232421875,
0.5185546875,
0.313232421875
],
[
0.0849609375,
0.3193359375,
0.48193359375,
0.3193359375,
0.48193359375,
0.333251953125,
0.0849609375,
0.333251953125
],
[
0.51708984375,
0.31640625,
0.9140625,
0.31640625,
0.9140625,
0.330078125,
0.51708984375,
0.330078125
],
[
0.08416748046875,
0.333251953125,
0.48193359375,
0.33203125,
0.48193359375,
0.3466796875,
0.08416748046875,
0.347900390625
],
[
0.515625,
0.330078125,
0.9140625,
0.328857421875,
0.9140625,
0.34228515625,
0.515625,
0.343505859375
],
[
0.0849609375,
0.3466796875,
0.48193359375,
0.3466796875,
0.48193359375,
0.35986328125,
0.0849609375,
0.35986328125
],
[
0.51708984375,
0.343505859375,
0.9140625,
0.343505859375,
0.9140625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.0849609375,
0.361083984375,
0.48046875,
0.361083984375,
0.48046875,
0.375,
0.0849609375,
0.375
],
[
0.51708984375,
0.356689453125,
0.734375,
0.356689453125,
0.734375,
0.37060546875,
0.51708984375,
0.37060546875
],
[
0.08660888671875,
0.375732421875,
0.4794921875,
0.375732421875,
0.4794921875,
0.386474609375,
0.08660888671875,
0.386474609375
],
[
0.53515625,
0.37255859375,
0.91259765625,
0.37255859375,
0.91259765625,
0.38330078125,
0.53515625,
0.38330078125
],
[
0.0849609375,
0.38818359375,
0.483642578125,
0.38818359375,
0.483642578125,
0.401611328125,
0.0849609375,
0.401611328125
],
[
0.51708984375,
0.385009765625,
0.9140625,
0.385009765625,
0.9140625,
0.3984375,
0.51708984375,
0.3984375
],
[
0.08660888671875,
0.401611328125,
0.48193359375,
0.401611328125,
0.48193359375,
0.415283203125,
0.08660888671875,
0.415283203125
],
[
0.5185546875,
0.3984375,
0.9140625,
0.3984375,
0.9140625,
0.412353515625,
0.5185546875,
0.412353515625
],
[
0.08660888671875,
0.417236328125,
0.48193359375,
0.417236328125,
0.48193359375,
0.427978515625,
0.08660888671875,
0.427978515625
],
[
0.5185546875,
0.4140625,
0.91259765625,
0.4140625,
0.91259765625,
0.4248046875,
0.5185546875,
0.4248046875
],
[
0.08416748046875,
0.429931640625,
0.281982421875,
0.429931640625,
0.281982421875,
0.4443359375,
0.08416748046875,
0.4443359375
],
[
0.5185546875,
0.4267578125,
0.9560546875,
0.4267578125,
0.9560546875,
0.43994140625,
0.5185546875,
0.43994140625
],
[
0.5185546875,
0.442626953125,
0.9111328125,
0.442626953125,
0.9111328125,
0.45263671875,
0.5185546875,
0.45263671875
],
[
0.08660888671875,
0.451904296875,
0.48193359375,
0.451904296875,
0.48193359375,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.5185546875,
0.455810546875,
0.91259765625,
0.455810546875,
0.91259765625,
0.466552734375,
0.5185546875,
0.466552734375
],
[
0.08660888671875,
0.46728515625,
0.48046875,
0.46728515625,
0.48046875,
0.477783203125,
0.08660888671875,
0.477783203125
],
[
0.51708984375,
0.468505859375,
0.9140625,
0.468505859375,
0.9140625,
0.481689453125,
0.51708984375,
0.481689453125
],
[
0.08416748046875,
0.477783203125,
0.48193359375,
0.479736328125,
0.48193359375,
0.49365234375,
0.08416748046875,
0.4912109375
],
[
0.5185546875,
0.484130859375,
0.91259765625,
0.484130859375,
0.91259765625,
0.494384765625,
0.5185546875,
0.494384765625
],
[
0.0849609375,
0.49365234375,
0.48193359375,
0.49365234375,
0.48193359375,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.5185546875,
0.49755859375,
0.91259765625,
0.49755859375,
0.91259765625,
0.50830078125,
0.5185546875,
0.50830078125
],
[
0.0849609375,
0.5068359375,
0.48193359375,
0.5068359375,
0.48193359375,
0.52099609375,
0.0849609375,
0.52099609375
],
[
0.51708984375,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51708984375,
0.5234375
],
[
0.08416748046875,
0.51953125,
0.48046875,
0.52099609375,
0.48046875,
0.5341796875,
0.08416748046875,
0.53271484375
],
[
0.51708984375,
0.5244140625,
0.9140625,
0.5244140625,
0.9140625,
0.53857421875,
0.51708984375,
0.53857421875
],
[
0.08660888671875,
0.5341796875,
0.483642578125,
0.5341796875,
0.483642578125,
0.54736328125,
0.08660888671875,
0.54736328125
],
[
0.51708984375,
0.537109375,
0.9140625,
0.537109375,
0.9140625,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.0849609375,
0.548828125,
0.48193359375,
0.548828125,
0.48193359375,
0.5625,
0.0849609375,
0.5625
],
[
0.51708984375,
0.55029296875,
0.80224609375,
0.55029296875,
0.80224609375,
0.564453125,
0.51708984375,
0.564453125
],
[
0.0849609375,
0.5625,
0.48193359375,
0.5625,
0.48193359375,
0.57568359375,
0.0849609375,
0.57568359375
],
[
0.53515625,
0.56640625,
0.91259765625,
0.56640625,
0.91259765625,
0.5771484375,
0.53515625,
0.5771484375
],
[
0.08660888671875,
0.57763671875,
0.48193359375,
0.57763671875,
0.48193359375,
0.58837890625,
0.08660888671875,
0.58837890625
],
[
0.5185546875,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.5185546875,
0.59228515625
],
[
0.0849609375,
0.59033203125,
0.366943359375,
0.59033203125,
0.366943359375,
0.60400390625,
0.0849609375,
0.60400390625
],
[
0.51708984375,
0.59228515625,
0.91259765625,
0.59228515625,
0.91259765625,
0.60595703125,
0.51708984375,
0.60595703125
],
[
0.08660888671875,
0.611328125,
0.48193359375,
0.611328125,
0.48193359375,
0.625,
0.08660888671875,
0.625
],
[
0.5185546875,
0.60693359375,
0.8486328125,
0.60693359375,
0.8486328125,
0.6171875,
0.5185546875,
0.6171875
],
[
0.0849609375,
0.62548828125,
0.48046875,
0.62548828125,
0.48046875,
0.6396484375,
0.0849609375,
0.6396484375
],
[
0.0849609375,
0.6396484375,
0.48193359375,
0.6396484375,
0.48193359375,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.51953125,
0.63818359375,
0.63232421875,
0.63818359375,
0.63232421875,
0.6484375,
0.51953125,
0.6484375
],
[
0.08660888671875,
0.65478515625,
0.48046875,
0.65478515625,
0.48046875,
0.66552734375,
0.08660888671875,
0.66552734375
],
[
0.52685546875,
0.6513671875,
0.9140625,
0.6513671875,
0.9140625,
0.66552734375,
0.52685546875,
0.66552734375
],
[
0.08416748046875,
0.66650390625,
0.4755859375,
0.66650390625,
0.4755859375,
0.6796875,
0.08416748046875,
0.6796875
],
[
0.54736328125,
0.662109375,
0.9111328125,
0.66357421875,
0.9111328125,
0.6748046875,
0.54736328125,
0.673828125
],
[
0.5498046875,
0.673828125,
0.6298828125,
0.673828125,
0.6298828125,
0.6845703125,
0.5498046875,
0.6845703125
],
[
0.52685546875,
0.681640625,
0.9140625,
0.681640625,
0.9140625,
0.69580078125,
0.52685546875,
0.69580078125
],
[
0.08905029296875,
0.693359375,
0.101318359375,
0.6962890625,
0.09722900390625,
0.70556640625,
0.0849609375,
0.70263671875
],
[
0.10540771484375,
0.693359375,
0.318603515625,
0.693359375,
0.318603515625,
0.70703125,
0.10540771484375,
0.70703125
],
[
0.54833984375,
0.6923828125,
0.91259765625,
0.693359375,
0.91259765625,
0.705078125,
0.54833984375,
0.7041015625
],
[
0.57275390625,
0.70556640625,
0.64453125,
0.70556640625,
0.64453125,
0.71337890625,
0.57275390625,
0.71337890625
],
[
0.654296875,
0.705078125,
0.70361328125,
0.705078125,
0.70361328125,
0.7119140625,
0.654296875,
0.7119140625
],
[
0.0849609375,
0.7119140625,
0.48193359375,
0.7119140625,
0.48193359375,
0.72607421875,
0.0849609375,
0.72607421875
],
[
0.529296875,
0.71533203125,
0.54736328125,
0.71533203125,
0.54736328125,
0.724609375,
0.529296875,
0.724609375
],
[
0.54345703125,
0.7119140625,
0.9140625,
0.7119140625,
0.9140625,
0.72607421875,
0.54345703125,
0.72607421875
],
[
0.08416748046875,
0.724609375,
0.48193359375,
0.72607421875,
0.48193359375,
0.74072265625,
0.08416748046875,
0.7392578125
],
[
0.5498046875,
0.7236328125,
0.9111328125,
0.7236328125,
0.9111328125,
0.734375,
0.5498046875,
0.734375
],
[
0.560546875,
0.734375,
0.86376953125,
0.734375,
0.86376953125,
0.744140625,
0.560546875,
0.744140625
],
[
0.08660888671875,
0.7412109375,
0.3505859375,
0.7412109375,
0.3505859375,
0.751953125,
0.08660888671875,
0.751953125
],
[
0.529296875,
0.74560546875,
0.5458984375,
0.74560546875,
0.5458984375,
0.7548828125,
0.529296875,
0.7548828125
],
[
0.54345703125,
0.7421875,
0.9140625,
0.7421875,
0.9140625,
0.755859375,
0.54345703125,
0.755859375
],
[
0.54833984375,
0.751953125,
0.91259765625,
0.75244140625,
0.91259765625,
0.7666015625,
0.54833984375,
0.76513671875
],
[
0.11273193359375,
0.7646484375,
0.476318359375,
0.7646484375,
0.476318359375,
0.77783203125,
0.11273193359375,
0.77783203125
],
[
0.54833984375,
0.7646484375,
0.685546875,
0.76318359375,
0.685546875,
0.7744140625,
0.54833984375,
0.77587890625
],
[
0.525390625,
0.771484375,
0.9140625,
0.77099609375,
0.9140625,
0.78515625,
0.525390625,
0.7861328125
],
[
0.1258544921875,
0.78515625,
0.291748046875,
0.78515625,
0.291748046875,
0.798828125,
0.1258544921875,
0.798828125
],
[
0.5498046875,
0.78271484375,
0.8203125,
0.78271484375,
0.8203125,
0.79345703125,
0.5498046875,
0.79345703125
],
[
0.52685546875,
0.79345703125,
0.9140625,
0.79345703125,
0.9140625,
0.806640625,
0.52685546875,
0.806640625
],
[
0.125,
0.7998046875,
0.283447265625,
0.7998046875,
0.283447265625,
0.81298828125,
0.125,
0.81298828125
],
[
0.57275390625,
0.806640625,
0.619140625,
0.806640625,
0.619140625,
0.81103515625,
0.57275390625,
0.81103515625
],
[
0.63623046875,
0.8046875,
0.70751953125,
0.8037109375,
0.70751953125,
0.8125,
0.63623046875,
0.81298828125
],
[
0.12335205078125,
0.81298828125,
0.294189453125,
0.81298828125,
0.294189453125,
0.8271484375,
0.12335205078125,
0.8271484375
],
[
0.52685546875,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.52685546875,
0.8271484375
],
[
0.125,
0.8271484375,
0.33251953125,
0.8271484375,
0.33251953125,
0.84033203125,
0.125,
0.84033203125
],
[
0.55126953125,
0.82470703125,
0.91259765625,
0.82470703125,
0.91259765625,
0.83544921875,
0.55126953125,
0.83544921875
],
[
0.5498046875,
0.833984375,
0.60302734375,
0.833984375,
0.60302734375,
0.8447265625,
0.5498046875,
0.8447265625
],
[
0.12744140625,
0.84228515625,
0.330810546875,
0.84228515625,
0.330810546875,
0.85302734375,
0.12744140625,
0.85302734375
],
[
0.52685546875,
0.84228515625,
0.91259765625,
0.84228515625,
0.91259765625,
0.85595703125,
0.52685546875,
0.85595703125
],
[
0.125,
0.85498046875,
0.279296875,
0.85498046875,
0.279296875,
0.86865234375,
0.125,
0.86865234375
],
[
0.54736328125,
0.85400390625,
0.9140625,
0.85400390625,
0.9140625,
0.8671875,
0.54736328125,
0.8671875
],
[
0.5498046875,
0.8662109375,
0.57861328125,
0.8662109375,
0.57861328125,
0.875,
0.5498046875,
0.875
],
[
0.111083984375,
0.875,
0.281005859375,
0.875,
0.281005859375,
0.88818359375,
0.111083984375,
0.88818359375
],
[
0.525390625,
0.87255859375,
0.9150390625,
0.8720703125,
0.9150390625,
0.88623046875,
0.525390625,
0.88720703125
],
[
0.54833984375,
0.8837890625,
0.89453125,
0.8837890625,
0.89453125,
0.89794921875,
0.54833984375,
0.89794921875
],
[
0.496826171875,
0.90087890625,
0.5048828125,
0.90380859375,
0.5009765625,
0.91015625,
0.49267578125,
0.9072265625
]
] |
[
"Question Answering over Curated and Open Web Sources",
"SIGIR'20,July 25-30,2020,Xi'an,China",
"search results. SplitQA [80] addressed complex questions by de",
"11:00-12:30PartIl:QAovertextualsources(1.5hours)",
"composing them into a sequence of simple questions, and relies",
"- Background (15 minutes)",
"on crowdsourced training data.Somemethodsfor hybrid QA start",
"with KGs as a source for candidate answers and use text corpora",
"- Machine reading (20 minutes)",
"like Wikipedia or ClueWeb as additional evidence [26,75,76,79,89",
"-Open-domain QA (20 minutes)",
"- Feedback and interpretability (10 minutes)",
"92],orstartwithanswersentencesfromtextcorporaand combine",
"-Conversational QA (10 minutes)",
"these with KGs for giving crisp entity answers [71, 78].",
"-Summary (15minutes)",
"4.4 New Horizons in QA",
"b",
"PRESENTERS",
"4.4.1ConversationalQA.ConversationalOA involvesa sequence",
"of questions and answers that appear as a natural dialogue between",
"Rishiraj SahaRoyis a Senior Researcher at the MaxPlanck Insti-",
"the system and the user. The aim of such sequential, multi-turn QA",
"tute for Informatics (MPII), Saarbrucken, Germany. He leads the",
"is to understand the contextleft implicit by users and effectively an-",
"research group on Question Answering, that focuses on robust",
"swer incomplete and ad hoc follow-up questions. Towards this, var-",
"and interpretable solutions for answering natural language ques-",
"ious recent benchmarks have been proposed that expect answers",
"tions over structured and unstructured data. He has more than",
"that are boolean [69], extractive [18] and free-form responses [68]",
"four years of research experience on question answering. In recent",
"entities [19], passages [48], and chit-chat [101]. Leaderboards of",
"years, he has served on the PCs of conferences like SIGIR, CIKM,",
"the QuAC [18] and CoQA [68] datasets point to many recent ap-",
"AAAI, EMNLP, and NAACL, and published at venues like SIGIR,",
"proaches in the text domain. Recently, the TREC CAsT track [24]",
"CIKM, WSDM, WWW, and NAACL",
"(http://www.treccast.ai/)andtheDagstuhlSeminaronConversa",
"Prior to joining MPIl, he worked for one and a half years as a",
"tional Search [7] tried to address such challenges in conversational",
"Computer Scientist at Adobe Research, Bangalore, India. He com-",
"search. For KG-QA, notable efforts include [19, 39, 70, 73]. We will",
"pleted his PhD as a Microsoft Research India Fellow from the In-",
"focus on the modelling complexity ofthesetasks and a classifica-",
"dian Institute of Technology(IIT)Kharagpur.More information",
"tion of the approaches involved.",
"about his research can be found at http:/people.mpi-inf.mpg.de/-rsaharo/",
"AvishekAnand is anAssistant Professor at the LeibnizUniver",
"4.4.2Feedbackand interpretability.Staticlearning systemsforQA",
"sitat Hannover, Germany.His research interests lie at the intersec-",
"aregraduallypaving thewayforthosethat incorporateuserfeed-",
"tion of retrieval, mining, and machine learning. He did his PhD at",
"back.These mostly design the setup as a continuous learning setup,",
"the MaxPlanckInstitute for Informatics,Saarbricken, Germany",
"where the explicit user feedback mechanism is built on top of an",
"whereheworked ontemporal information retrieval.Currently,he",
"existing QA system.For example, the NEQA[2],QApedia [50], and",
"is developing intelligent and transparent representation learning",
"IMPROVE-QA [1oo] systems primarily operate on the core sys",
"methods for text and graphs in the context of search, QA and Web",
"tems of QUINT [4], DrQA [16], and gAnswer [42], respectively.",
"tasks. He has published his research in several top-tier conferences,",
"A direction closely coupled with effective feedback is the inter-",
"such as SIGIR, WSDM, CIKM, ICDE and TKDE.",
"pretability of QA models, that is also essential to improving trust",
"He delivered a tutorial at SIGIR 2016,on Temporal Information",
"and satisfaction[1,88].Thissection isforexpertsand wewill dis-",
"Retrieval, and several talks in various summer schools and invited",
"cuss potential limitations and open challenges.",
"lectures. Avishek was one of the lead organizers of the recently",
"4.4.3Clarification questions.Akey aspect ofmixed-initiative sys-",
"concludedDagstuhl seminaron Conversational Search",
"tems [64] is to be able to ask clarifications. This ability is essential",
"to QA systems, especially for handling ambiguity and facilitating",
"REFERENCES",
"better question understanding.Most ofthework inthis domain has",
"[1] Abdalghani Abujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard",
"been driven by extracting tasks from open QA forums [14, 66, 90]",
"Weikum.2017.QUINT:Interpretable Question Answering over Knowledge",
"Bases. In EMNLP",
"[2] Abdalghani Abujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard",
"5",
"FORMAT AND SUPPORT",
"Weikum. 2018. Never-ending learning for open-domain question answering",
"knowledge base",
"Inwww",
"A detailed schedule for our proposed half-day tutorial (three hours",
"[3] ",
"Abdalghani Abujabal, Rishiraj Saha Roy, Mohamed Yahya, and Gerhard",
"plus breaks), which is aimed to meet a high-quality presentation",
"Weikum.2019.ComQA: A Community-sourced Dataset for Complex Factoid",
"lestion Answering with Paraphrase Clusters. In NAACL-HLT'19.",
"within the chosen time period, is asfollows:",
"[4] ",
"Abdalghani Abujabal, Mohamed Yahya, Mirek Riedewald, and Gerhard",
"Weikum.2017.Automated template generation for question answering over",
" 9:00 - 10:30 Part I: QA over knowledge graphs (1.5 hours)",
"knowledge graphs. In www",
"[5] Eugene Agichtein, David Carmel, Dan Pelleg, Yuval Pinter, and Donna Harman",
"-Background (15 minutes)",
"2015.Overview of theTREC2015LiveOATrack.InTREC",
"[6] Aditya Kalyanpur Alfio Gliozz0 and James Fan. 2012. Natural language pro-",
"-SimpleQA(10 minutes)",
"noinWat",
"In NAACL-HL",
"- Complex QA (20 minutes)",
"[7] Avishek Anand, Lawrence Cavedon, Hideo Joho, Mark Sanderson, and Benno",
"- Heterogeneous QA (15 minutes)",
"Stein.2020.Conversational Search (Dagstuhl Seminar 19461).Dagstuhl Reports",
"9,11(2020)",
"-ConversationalQA(15minutes)",
"[8] Soren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyga-",
"- Summary (15 minutes)",
"niak, and Zachary Ives: 2007. DBpedia: A nucleus for a Web of open data.",
"(2007)",
"·10:30-11:00Coffeebreak",
"[9] junwei Bao, Nan Duan, Zhao Yan, Ming Zhou, and Tiejun Zhao. 2016.",
"Constraint-based question answering with knowledge graph. In COLING.",
"3"
] |
{
"arxiv_id": "2004.11980",
"language": "en",
"page_idx": 2
}
|
||
2311.09447-en-0
|
2311.09447-en-0.jpg
|
### How Trustworthy are Open-Source LLMs? An Assessment under Malicious Demonstrations Shows their Vulnerabilities
Content Warning: This paper contains potentially offensive and harmful text.
$\begin{array}{c}\text{Lingbo Mo}^{1}&\text{Boshi Wang}^{1}&\text{Muhao Chen}^{2}&\text{Huan Sun}^{1}\\\text{The Ohio State University}&\text{2University of California, Davis}\\\end{array}$
$\{$mo.169, wang.13930, sun.397$\}$@osu.edu; muhchen@ucdavis.edu
## Abstract
The rapid progress in open-source Large Language Models (LLMs) is significantly driving AI development forward. However, there is still a limited understanding of their trustworthiness. Deploying these models at scale without sufficient trustworthiness can pose significant risks, highlighting the need to uncover these issues promptly. In this work, we conduct an adversarial assessment of open-source LLMs on trustworthiness, scrutinizing them across eight different aspects including toxicity, stereotypes, ethics, hallucination, fairness, sycophancy, privacy, and robustness against adversarial demonstrations. We propose $advCoU$, an extended Chain of Utterances-based (CoU) prompting strategy by incorporating carefully crafted malicious demonstrations for trustworthiness attack. Our extensive experiments encompass recent and representative series of open-source LLMs, including VICUNA, MPT, FALCON, MISTRAL, and LLAMA 2. The empirical outcomes underscore the efficacy of our attack strategy across diverse aspects. More interestingly, our result analysis reveals that models with superior performance in general NLP tasks do not always have greater trustworthiness; in fact, larger models can be more vulnerable to attacks. Additionally, models that have undergone instruction tuning, focusing on instruction following, tend to be more susceptible, although fine- tuning LLMs for safety alignment proves effective in mitigating adversarial trustworthiness attacks. 1
### 1 Introduction
The field of large language models (LLMs) has witnessed remarkable progress, highlighted by the emergence of closed-source models such as ChatGPT (OpenAI, 2022), GPT-4 (OpenAI, 2023), and Claude (Anthropic, 2023). This advancement has enabled downstream AI systems built upon these
$^{\mathrm{l}}$Our code is available at https://github.com/OSU-NLP-
Group/Eval-LLM-Trust.
models to demonstrate increasingly human-like capabilities. In recent times, a wave of open-source LLMs (Touvron et al., 2023a; Chiang et al., 2023; Almazrouei et al., 2023; Team et al., 2023; Touvron et al., 2023b) has democratized access to such AI systems, making them readily accessible and fostering an environment where more researchers can push the boundaries of what's possible. Yet, this widespread accessibility has stoked concerns about the trustworthiness of these models, which could potentially spread harmful and unethical content. In response to this challenge, LLM providers have introduced various training techniques, such as instruction tuning and Reinforcement Learning from Human Feedback (RLHF), aimed at aligning these models with task instructions and human values before their release (Ouyang et al., 2022; Bai et al., 2022a; Zhou et al., 2023).
Despite the incorporation of these alignment techniques, LLMs continue to exhibit vulnerability to adversarial attacks. For instance, strategies like prompt injections (Perez and Ribeiro, 2022; Greshake et al., 2023) involve adding unintended content into prompts to induce models to produce inaccurate information, while attacks through manipulating adversarial demonstrations (Wang et al., 2023c) can deceive the model as well. Jailbreaking prompts (Bai et al., 2022b; Carlini et al., 2023; Zou et al., 2023) aim to bypass LLM alignment and induce harmful model outputs. It is noteworthy that efforts to systematically study the trustworthiness of LLMs are still in their early stages, with many previous studies focusing on a limited subset of trustworthiness aspects like toxicity (Deshpande et al., 2023; Huang et al., 2023b) and stereotype bias (Mattern et al., 2022; Shaikh et al., 2022).
In this paper, our primary goal is to comprehensively assess the trustworthiness of recent opensource LLMs through attack-based examinations. Our investigation covers a broad spectrum of eight aspects of trustworthiness compiled from recent
|
[
[
0.162841796875,
0.0926513671875,
0.83642578125,
0.0926513671875,
0.83642578125,
0.10687255859375,
0.162841796875,
0.10687255859375
],
[
0.2225341796875,
0.11224365234375,
0.77734375,
0.11224365234375,
0.77734375,
0.1270751953125,
0.2225341796875,
0.1270751953125
],
[
0.16455078125,
0.141357421875,
0.82861328125,
0.141357421875,
0.82861328125,
0.1561279296875,
0.16455078125,
0.1561279296875
],
[
0.26123046875,
0.1776123046875,
0.5029296875,
0.1776123046875,
0.5029296875,
0.1923828125,
0.26123046875,
0.1923828125
],
[
0.490234375,
0.174560546875,
0.74072265625,
0.17578125,
0.74072265625,
0.193603515625,
0.490234375,
0.1923828125
],
[
0.25439453125,
0.1923828125,
0.48193359375,
0.19482421875,
0.48193359375,
0.208984375,
0.25439453125,
0.207275390625
],
[
0.490234375,
0.191162109375,
0.75390625,
0.1923828125,
0.75390625,
0.210205078125,
0.490234375,
0.208984375
],
[
0.480224609375,
0.1995849609375,
0.49365234375,
0.1995849609375,
0.49365234375,
0.2037353515625,
0.480224609375,
0.2037353515625
],
[
0.210693359375,
0.210205078125,
0.80029296875,
0.208984375,
0.80029296875,
0.226806640625,
0.210693359375,
0.22802734375
],
[
0.2626953125,
0.262451171875,
0.343505859375,
0.262451171875,
0.343505859375,
0.276611328125,
0.2626953125,
0.276611328125
],
[
0.51123046875,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.51318359375,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.145263671875,
0.287353515625,
0.459228515625,
0.287353515625,
0.459228515625,
0.300537109375,
0.145263671875,
0.300537109375
],
[
0.51318359375,
0.29443359375,
0.8818359375,
0.29443359375,
0.8818359375,
0.30810546875,
0.51318359375,
0.30810546875
],
[
0.14697265625,
0.3017578125,
0.45751953125,
0.3017578125,
0.45751953125,
0.3154296875,
0.14697265625,
0.3154296875
],
[
0.0218353271484375,
0.311279296875,
0.0697021484375,
0.311279296875,
0.06634521484375,
0.43408203125,
0.019317626953125,
0.43408203125
],
[
0.51318359375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.1485595703125,
0.317626953125,
0.45751953125,
0.317626953125,
0.45751953125,
0.327880859375,
0.1485595703125,
0.327880859375
],
[
0.14697265625,
0.328857421875,
0.459228515625,
0.328857421875,
0.459228515625,
0.342529296875,
0.14697265625,
0.342529296875
],
[
0.509765625,
0.325927734375,
0.880859375,
0.325927734375,
0.880859375,
0.339111328125,
0.509765625,
0.339111328125
],
[
0.14697265625,
0.344482421875,
0.459228515625,
0.344482421875,
0.459228515625,
0.358154296875,
0.14697265625,
0.358154296875
],
[
0.51123046875,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.51123046875,
0.35693359375
],
[
0.145263671875,
0.359375,
0.459228515625,
0.359375,
0.459228515625,
0.372314453125,
0.145263671875,
0.372314453125
],
[
0.51123046875,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.51123046875,
0.37353515625
],
[
0.14697265625,
0.37353515625,
0.456787109375,
0.37353515625,
0.456787109375,
0.38427734375,
0.14697265625,
0.38427734375
],
[
0.51123046875,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.51123046875,
0.3896484375
],
[
0.14697265625,
0.38720703125,
0.459228515625,
0.38720703125,
0.459228515625,
0.40087890625,
0.14697265625,
0.40087890625
],
[
0.51318359375,
0.393798828125,
0.880859375,
0.393798828125,
0.880859375,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.145263671875,
0.402099609375,
0.45751953125,
0.402099609375,
0.45751953125,
0.4150390625,
0.145263671875,
0.4150390625
],
[
0.51123046875,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.51123046875,
0.421630859375
],
[
0.145263671875,
0.4150390625,
0.459228515625,
0.4150390625,
0.459228515625,
0.4287109375,
0.145263671875,
0.4287109375
],
[
0.51318359375,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.51318359375,
0.438232421875
],
[
0.0235137939453125,
0.43408203125,
0.06378173828125,
0.43408203125,
0.06378173828125,
0.5126953125,
0.0235137939453125,
0.5126953125
],
[
0.1485595703125,
0.43115234375,
0.45751953125,
0.43115234375,
0.45751953125,
0.441162109375,
0.1485595703125,
0.441162109375
],
[
0.14697265625,
0.443603515625,
0.459228515625,
0.443603515625,
0.459228515625,
0.457275390625,
0.14697265625,
0.457275390625
],
[
0.51123046875,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.51123046875,
0.45361328125
],
[
0.14697265625,
0.457763671875,
0.45751953125,
0.457763671875,
0.45751953125,
0.471435546875,
0.14697265625,
0.471435546875
],
[
0.509765625,
0.457275390625,
0.8818359375,
0.457275390625,
0.8818359375,
0.47021484375,
0.509765625,
0.47021484375
],
[
0.145263671875,
0.47265625,
0.459228515625,
0.47265625,
0.459228515625,
0.486328125,
0.145263671875,
0.486328125
],
[
0.509765625,
0.47265625,
0.880859375,
0.47265625,
0.880859375,
0.486328125,
0.509765625,
0.486328125
],
[
0.14697265625,
0.486328125,
0.459228515625,
0.486328125,
0.459228515625,
0.49951171875,
0.14697265625,
0.49951171875
],
[
0.51123046875,
0.488037109375,
0.88330078125,
0.488037109375,
0.88330078125,
0.501953125,
0.51123046875,
0.501953125
],
[
0.14697265625,
0.50048828125,
0.45751953125,
0.50048828125,
0.45751953125,
0.51416015625,
0.14697265625,
0.51416015625
],
[
0.51123046875,
0.50439453125,
0.8818359375,
0.50439453125,
0.8818359375,
0.51708984375,
0.51123046875,
0.51708984375
],
[
0.0218353271484375,
0.51416015625,
0.0621337890625,
0.51416015625,
0.0621337890625,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.14697265625,
0.5166015625,
0.45751953125,
0.5166015625,
0.45751953125,
0.52685546875,
0.14697265625,
0.52685546875
],
[
0.509765625,
0.52099609375,
0.8818359375,
0.52099609375,
0.8818359375,
0.53369140625,
0.509765625,
0.53369140625
],
[
0.14697265625,
0.53076171875,
0.45751953125,
0.53076171875,
0.45751953125,
0.54150390625,
0.14697265625,
0.54150390625
],
[
0.51123046875,
0.5361328125,
0.705078125,
0.5361328125,
0.705078125,
0.5498046875,
0.51123046875,
0.5498046875
],
[
0.145263671875,
0.54345703125,
0.459228515625,
0.54345703125,
0.459228515625,
0.55712890625,
0.145263671875,
0.55712890625
],
[
0.145263671875,
0.55712890625,
0.459228515625,
0.55712890625,
0.459228515625,
0.57080078125,
0.145263671875,
0.57080078125
],
[
0.52880859375,
0.552734375,
0.8818359375,
0.552734375,
0.8818359375,
0.56787109375,
0.52880859375,
0.56787109375
],
[
0.14697265625,
0.572265625,
0.45751953125,
0.572265625,
0.45751953125,
0.5830078125,
0.14697265625,
0.5830078125
],
[
0.509765625,
0.56787109375,
0.880859375,
0.56787109375,
0.880859375,
0.58056640625,
0.509765625,
0.58056640625
],
[
0.1485595703125,
0.58740234375,
0.45751953125,
0.58740234375,
0.45751953125,
0.59716796875,
0.1485595703125,
0.59716796875
],
[
0.51123046875,
0.5849609375,
0.8818359375,
0.5849609375,
0.8818359375,
0.5986328125,
0.51123046875,
0.5986328125
],
[
0.145263671875,
0.599609375,
0.45751953125,
0.5986328125,
0.45751953125,
0.6123046875,
0.145263671875,
0.61328125
],
[
0.51123046875,
0.60107421875,
0.8818359375,
0.599609375,
0.8818359375,
0.6142578125,
0.51123046875,
0.615234375
],
[
0.145263671875,
0.61328125,
0.456787109375,
0.61328125,
0.456787109375,
0.626953125,
0.145263671875,
0.626953125
],
[
0.51318359375,
0.61767578125,
0.87890625,
0.61767578125,
0.87890625,
0.62744140625,
0.51318359375,
0.62744140625
],
[
0.145263671875,
0.62890625,
0.45751953125,
0.62890625,
0.45751953125,
0.642578125,
0.145263671875,
0.642578125
],
[
0.51123046875,
0.6328125,
0.8818359375,
0.6328125,
0.8818359375,
0.646484375,
0.51123046875,
0.646484375
],
[
0.145263671875,
0.642578125,
0.459228515625,
0.642578125,
0.459228515625,
0.65625,
0.145263671875,
0.65625
],
[
0.51123046875,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.6630859375,
0.51123046875,
0.6630859375
],
[
0.145263671875,
0.65673828125,
0.45751953125,
0.65673828125,
0.45751953125,
0.67041015625,
0.145263671875,
0.67041015625
],
[
0.51123046875,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6787109375,
0.51123046875,
0.6787109375
],
[
0.1485595703125,
0.6728515625,
0.45751953125,
0.6728515625,
0.45751953125,
0.68310546875,
0.1485595703125,
0.68310546875
],
[
0.145263671875,
0.68408203125,
0.456787109375,
0.68408203125,
0.456787109375,
0.69775390625,
0.145263671875,
0.69775390625
],
[
0.509765625,
0.6796875,
0.8818359375,
0.68115234375,
0.8818359375,
0.6953125,
0.509765625,
0.69482421875
],
[
0.145263671875,
0.69970703125,
0.459228515625,
0.69970703125,
0.459228515625,
0.71337890625,
0.145263671875,
0.71337890625
],
[
0.509765625,
0.69775390625,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.509765625,
0.71142578125
],
[
0.145263671875,
0.71435546875,
0.45751953125,
0.71435546875,
0.45751953125,
0.72802734375,
0.145263671875,
0.72802734375
],
[
0.51318359375,
0.71337890625,
0.88330078125,
0.71337890625,
0.88330078125,
0.72802734375,
0.51318359375,
0.72802734375
],
[
0.1435546875,
0.72802734375,
0.459228515625,
0.72802734375,
0.459228515625,
0.7412109375,
0.1435546875,
0.7412109375
],
[
0.51123046875,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74365234375,
0.51123046875,
0.74365234375
],
[
0.1435546875,
0.7421875,
0.355224609375,
0.7421875,
0.355224609375,
0.755859375,
0.1435546875,
0.755859375
],
[
0.509765625,
0.7451171875,
0.880859375,
0.7451171875,
0.880859375,
0.7587890625,
0.509765625,
0.7587890625
],
[
0.51123046875,
0.76123046875,
0.87890625,
0.76123046875,
0.87890625,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.1158447265625,
0.76708984375,
0.258544921875,
0.76708984375,
0.258544921875,
0.78076171875,
0.1158447265625,
0.78076171875
],
[
0.509765625,
0.77783203125,
0.8818359375,
0.77685546875,
0.8818359375,
0.79150390625,
0.509765625,
0.7919921875
],
[
0.1158447265625,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.509765625,
0.79443359375,
0.8818359375,
0.79443359375,
0.8818359375,
0.80810546875,
0.509765625,
0.80810546875
],
[
0.1158447265625,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.1158447265625,
0.822265625
],
[
0.51123046875,
0.80859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82470703125,
0.51123046875,
0.82373046875
],
[
0.1158447265625,
0.82470703125,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.1158447265625,
0.83837890625
],
[
0.51123046875,
0.826171875,
0.85205078125,
0.826171875,
0.85205078125,
0.8388671875,
0.51123046875,
0.8388671875
],
[
0.1175537109375,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.52880859375,
0.8427734375,
0.88330078125,
0.8427734375,
0.88330078125,
0.85546875,
0.52880859375,
0.85546875
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.509765625,
0.859375,
0.88330078125,
0.859375,
0.88330078125,
0.87255859375,
0.509765625,
0.87255859375
],
[
0.1175537109375,
0.87353515625,
0.48779296875,
0.87353515625,
0.48779296875,
0.88720703125,
0.1175537109375,
0.88720703125
],
[
0.5087890625,
0.87451171875,
0.8818359375,
0.87451171875,
0.8818359375,
0.88818359375,
0.5087890625,
0.88818359375
],
[
0.118408203125,
0.89453125,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.89111328125,
0.8818359375,
0.89111328125,
0.8818359375,
0.90478515625,
0.51318359375,
0.90478515625
],
[
0.12005615234375,
0.90771484375,
0.273681640625,
0.90771484375,
0.273681640625,
0.92138671875,
0.12005615234375,
0.92138671875
],
[
0.509765625,
0.90673828125,
0.8818359375,
0.90625,
0.8818359375,
0.92041015625,
0.509765625,
0.92138671875
]
] |
[
"How Trustworthy are Open-Source LLMs? An Assessment under",
"Malicious Demonstrations Shows their Vulnerabilities",
"Content Warning: This paper contains potentially offensive and harmful text.",
"Lingbo MolBoshi Wang!",
"Muhao Chen?Huan Sun'",
"IThe Ohio State University",
"2University of California, Davis",
"",
"[mo.169, wang.13930, sun.397)@osu.edu; muhchen@ucdavis.edu",
"Abstract",
"models to demonstrate increasingly human-like ca-",
"pabilities. In recent times, a wave of open-source",
"The rapid progress in open-source Large Lan-",
"LLMs (Touvron et al., 2023a; Chiang et al., 2023:",
"guage Models (LLMs) is significantly driving",
"",
"Almazrouei et al., 2023; Team et al., 2023; Touvron",
"A development forward. However, there is",
"still a limited understanding of their trustwor-",
"et al.. 2023b) has democratized access to such A1",
"thiness. Deploying these models at scale with-",
"systems, making them readily accessible and fos-",
"out sufficient trustworthiness can pose signif-",
"tering an environment where more researchers can",
"icant risks. highlighting the need to uncover",
"push the boundaries of what's possible. Yet, this",
"these issues promptly. In this work, we con-",
"widespreadaccessibilityhasstokedconcernsabout",
"duct an adversarial assessment of open-source",
"the trustworthiness of these models, which could",
"LLMs on trustworthiness, scrutinizing them",
"potentially spread harmful and unethical content",
"[T's]",
"across eightdifferent aspectsincludingtoxic-",
"ity, stereotypes, ethics, hallucination, fairness,",
"In response to this challenge, LLM providers have",
"sycophancy, privacy, and robustness against ad-",
"introduced various training techniques, such as in-",
"versarial demonstrations. We propose advCoU,",
" struction tuning and Reinforcement Learning from",
"an extended Chain of Utterances-based (CoU)",
"Human Feedback (RLHF), aimed at aligning these",
"prompting strategy by incorporating carefully",
"models with task instructions and human values",
"60118",
"crafted maliciousdemonstrationsfortrustwor",
"before their release (Ouyang et al., 2022; Bai et al.,",
"thiness attack. Our extensive experiments en-",
"2022a; Zhou et al., 2023).",
"compass recent and representative series of",
"open-source LLMs, including VICUNA, MPT,",
"Despite the incorporation of these alignment",
"FALCON,MISTRAL,andLLAMA2.Theem-",
" techniques, LLMs continue to exhibit vulnerabil.",
"pirical outcomesunderscoretheefficacy ofour",
"ity to adversarial attacks. For instance, strategies",
"attack strategy across diverse aspects. More",
"like prompt injections (Perez and Ribeiro, 2022;",
"interestingly, our result analysis reveals that",
"Greshakeetal..2023)involveaddingunintended",
"models with superior performance in general",
"content into prompts to induce models to produce",
"NLP tasks do not always have greater trust-",
"inaccurate information, while attacks through ma-",
"worthiness; in fact, larger models can be more",
"nipulating adversarial demonstrations (Wang et al..",
"vulnerabletoattacks.Additionally,modelsthat",
"have undergone instruction tuning, focusing",
"2023c) can deceive the model as well. Jailbreaking",
"on instruction following, tend to be more sus-",
"prompts (Bai et al., 2022b; Carlini et al., 2023: Zou",
"ceptible, although fine-tuning LLMs for safety",
"et al., 2023) aim to bypass LLM alignment and in-",
"alignment proves effective in mitigating adver-",
"duce harmful model outputs. It is noteworthy that",
"sarial trustworthiness attacks.",
"efforts to systematically study the trustworthiness",
"of LLMs are still in their early stages, with many",
"1 Introduction",
"previous studies focusing on a limited subset of",
"The field of large language models (LLMs) has",
"trustworthiness aspects like toxicity (Deshpande",
"witnessed remarkable progress, highlighted by the",
"et al., 2023; Huang et al., 2023b) and stereotype",
"emergence of closed-source models such as Chat-",
"bias (Mattern et al., 2022; Shaikh et al., 2022)",
"GPT (OpenAI, 2022), GPT-4 (OpenAI, 2023), and",
"In this paper, our primary goal is to comprehen-",
"Claude (Anthropic, 2023). This advancement has",
"sively assess the trustworthiness of recent open-",
"enabled downstream AI systems built upon these",
" source LLMs through attack-based examinations.",
"'Our code is available at https:/lgithub.com/OSU-NLP-",
"Our investigation covers a broad spectrum of eight",
"Group/Eval-LLM-Trust.",
"aspects of trustworthiness compiled from recent"
] |
{
"arxiv_id": "2311.09447",
"language": "en",
"page_idx": 0
}
|
||
2311.09447-en-2
|
2311.09447-en-2.jpg
|
expand their scope and introduce advCoU to mislead LLMs through the design of malicious internal thoughts and tailored demonstrations as in-context examples, aimed at inducing models to produce undesired content.
Our work includes scenarios where the model may exhibit biased judgment, select hallucinated answers erroneously, fail to identify unethical statements as wrong, and more (See Figure 1). This poses a more nuanced assessment of the models' reasoning and alignment capabilities, going beyond simply evaluating whether the model refuses to answer a harmful question or not, as conducted in Red-Eval. This approach allows us to manipulate only the demonstrations without changing the input to perform trustworthiness attacks. By conducting this comprehensive assessment, we seek to establish a better understanding of how trustworthy current open-source LLMs are against adversarial attacks. We intend to encourage increased trustworthiness-related research concerning opensource LLMs, thus mitigating potential risks to users and fostering reliable deployment and utility of produced LLMs in downstream systems, services and applications.
Furthermore, within the scope of our attack strategy, we delve into two research questions: (I) Do language models become more trustworthy as they grow larger? (2) Are models that have undergone instruction tuning and alignment processes more trustworthy? Through our exploration, we aim to gain deeper insights into the factors influencing trustworthiness, including model size and alignment-focused fine-tuning. Our experiments yield findings that models with superior performance in general NLP tasks do not necessarily have higher trustworthiness. In fact, larger models tend to be more susceptible to manipulation through malicious demonstrations. Moreover, models with instruction tuning, which emphasize instruction fol- $lowing,exhibit$ higher vulnerability, although finetuning LLMs with safety alignment proves effective in protecting against adversarial trustworthiness $attacks.$
Our contributions in this work can be summarized as follows: (1) We conduct a comprehensive assessment of open-source LLMs on trustworthiness across eight different aspects, including toxicity, stereotypes, and more. (2) We employ multiple adversarial attack strategies, starting from the recent work DECODINGTRUST (Wang et al., 2023b) as the baseline. In particular, we intro-
duce advCoU, an extended CoU prompting strategy by adapting the ICL paradigm to the attack scenario, and incorporating carefully designed malicious demonstrations to mislead LLMs. (3) We conduct extensive experiments that cover a recent representative series of open-source LLMs. Our experimental results demonstrate the effectiveness of our attack strategy across different aspects, showcasing an advantage over DECODINGTRUST. (4) Through in-depth result analysis, we uncover interesting findings that shed light on the relationship between trustworthiness and potential influence factors, such as model size and alignment strategies.
### 2 $\textbf{Related Work}$
Trustworthiness of LLMs. As LLMs continue to advance rapidly across various domains, concerns regarding their trustworthiness are becoming increasingly prominent. Previous investigations into the trustworthiness of LLMs have predominantly concentrated on individual aspects, such as toxicity (Tamkin et al., 2022; Deshpande et al., 2023; Huang et al., 2023b; Liu et al., 2023a; Jones et al., 2024), stereotypical bias (Mattern et al., 2022; Shaikh et al., 2022), privacy (Yue et al., 2023; Mireshghallah et al., 2023; Du et al., 2023), sycophancy (Wei et al., 2023; Wang et al., 2023a), robustness (Zhu et al., 2023; Li et al., 2023b), and more. In this paper, we aim to conduct a comprehensive examination of LLMs, taking into account various aspects compiled from recent studies (Liu et al., 2023b; Wang et al., 2023b), including toxicity, stereotype, ethics, hallucination, fairness, sycophancy, privacy, and robustness against adversarial demonstrations. While Wang et al. (2023b) introduced DECODINGTRUST, which provides a multifaceted assessment of trustworthiness vulnerabilities particularly in GPT-3.5 and GPT-4, we focus on evaluating open-source LLMs through existing and our attack strategies. More recently, Sun et al. (2024) presented TrustLLM and proposed a set of principles such as truthfulness and safety to evaluate the trustworthiness of LLMs. In our work, we study trustworthy LLMs from the perspective of adversarial assessments via a CoU-based prompting strategy.
Adversarial Attacks on LLMs. Recent LLM
providers have taken a range of safety mechanisms to align models with human values (Ouyang et al., 2022; Bai et al., 2022a; Zhou et al., 2023). However, LLMs still remain vulnerable to various
|
[
[
0.1175537109375,
0.08612060546875,
0.49365234375,
0.08612060546875,
0.49365234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1158447265625,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.509765625,
0.10272216796875,
0.8818359375,
0.10272216796875,
0.8818359375,
0.11639404296875,
0.509765625,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.509765625,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.509765625,
0.1478271484375
],
[
0.118408203125,
0.1519775390625,
0.25439453125,
0.1519775390625,
0.25439453125,
0.162109375,
0.118408203125,
0.162109375
],
[
0.51123046875,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.51123046875,
0.1644287109375
],
[
0.1368408203125,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1368408203125,
0.179931640625
],
[
0.50732421875,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.50732421875,
0.179931640625
],
[
0.1158447265625,
0.1817626953125,
0.489501953125,
0.1817626953125,
0.489501953125,
0.1953125,
0.1158447265625,
0.1953125
],
[
0.5087890625,
0.182861328125,
0.8818359375,
0.18115234375,
0.8818359375,
0.19482421875,
0.5087890625,
0.196533203125
],
[
0.1158447265625,
0.198974609375,
0.490234375,
0.198974609375,
0.490234375,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.5087890625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.5087890625,
0.2120361328125
],
[
0.11419677734375,
0.21435546875,
0.489501953125,
0.213134765625,
0.489501953125,
0.22802734375,
0.11419677734375,
0.2286376953125
],
[
0.509765625,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.11419677734375,
0.232177734375,
0.489501953125,
0.2298583984375,
0.489501953125,
0.243408203125,
0.11419677734375,
0.2452392578125
],
[
0.509765625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.509765625,
0.24462890625
],
[
0.1158447265625,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.509765625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.509765625,
0.26123046875
],
[
0.1158447265625,
0.262451171875,
0.490234375,
0.262451171875,
0.490234375,
0.275634765625,
0.1158447265625,
0.275634765625
],
[
0.509765625,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.276611328125,
0.509765625,
0.276611328125
],
[
0.1158447265625,
0.27783203125,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.292236328125
],
[
0.5087890625,
0.27978515625,
0.8759765625,
0.27978515625,
0.8759765625,
0.29345703125,
0.5087890625,
0.29345703125
],
[
0.1175537109375,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.515625,
0.31005859375,
0.54248046875,
0.31005859375,
0.54248046875,
0.318359375,
0.515625,
0.318359375
],
[
0.53564453125,
0.305908203125,
0.6630859375,
0.305908203125,
0.6630859375,
0.320556640625,
0.53564453125,
0.320556640625
],
[
0.1175537109375,
0.327880859375,
0.489501953125,
0.327880859375,
0.489501953125,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.1175537109375,
0.311279296875,
0.49365234375,
0.311279296875,
0.49365234375,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51123046875,
0.33203125,
0.880859375,
0.33203125,
0.880859375,
0.345703125,
0.51123046875,
0.345703125
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.5087890625,
0.349853515625,
0.8818359375,
0.3486328125,
0.8818359375,
0.363525390625,
0.5087890625,
0.364013671875
],
[
0.1175537109375,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.372314453125,
0.1175537109375,
0.372314453125
],
[
0.51123046875,
0.366455078125,
0.8818359375,
0.366455078125,
0.8818359375,
0.380126953125,
0.51123046875,
0.380126953125
],
[
0.1175537109375,
0.3759765625,
0.490234375,
0.3759765625,
0.490234375,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.51123046875,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.11419677734375,
0.39013671875,
0.489501953125,
0.391357421875,
0.489501953125,
0.40625,
0.11419677734375,
0.405029296875
],
[
0.5087890625,
0.396728515625,
0.8818359375,
0.39794921875,
0.8818359375,
0.412109375,
0.5087890625,
0.410888671875
],
[
0.1158447265625,
0.407958984375,
0.490234375,
0.407958984375,
0.490234375,
0.421630859375,
0.1158447265625,
0.421630859375
],
[
0.50732421875,
0.41455078125,
0.880859375,
0.41455078125,
0.880859375,
0.427490234375,
0.50732421875,
0.427490234375
],
[
0.1158447265625,
0.42333984375,
0.48779296875,
0.42333984375,
0.48779296875,
0.438232421875,
0.1158447265625,
0.438232421875
],
[
0.5087890625,
0.4287109375,
0.88330078125,
0.429931640625,
0.88330078125,
0.44482421875,
0.5087890625,
0.443603515625
],
[
0.11419677734375,
0.440673828125,
0.486083984375,
0.440673828125,
0.486083984375,
0.45361328125,
0.11419677734375,
0.45361328125
],
[
0.51123046875,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.11419677734375,
0.454833984375,
0.489501953125,
0.4560546875,
0.489501953125,
0.47021484375,
0.11419677734375,
0.4697265625
],
[
0.51318359375,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.1158447265625,
0.47265625,
0.28955078125,
0.47265625,
0.28955078125,
0.486328125,
0.1158447265625,
0.486328125
],
[
0.51123046875,
0.476806640625,
0.88330078125,
0.47802734375,
0.88330078125,
0.4921875,
0.51123046875,
0.4912109375
],
[
0.133544921875,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.133544921875,
0.501953125
],
[
0.51123046875,
0.493408203125,
0.88330078125,
0.49462890625,
0.88330078125,
0.5087890625,
0.51123046875,
0.50830078125
],
[
0.1175537109375,
0.5048828125,
0.489501953125,
0.5048828125,
0.489501953125,
0.5185546875,
0.1175537109375,
0.5185546875
],
[
0.51123046875,
0.51025390625,
0.8818359375,
0.5087890625,
0.8818359375,
0.52392578125,
0.51123046875,
0.52490234375
],
[
0.1175537109375,
0.52099609375,
0.48779296875,
0.52099609375,
0.48779296875,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.5087890625,
0.52685546875,
0.880859375,
0.52685546875,
0.880859375,
0.54052734375,
0.5087890625,
0.54052734375
],
[
0.1158447265625,
0.5361328125,
0.489501953125,
0.53515625,
0.489501953125,
0.5498046875,
0.1158447265625,
0.55029296875
],
[
0.51123046875,
0.54345703125,
0.8818359375,
0.54345703125,
0.8818359375,
0.55712890625,
0.51123046875,
0.55712890625
],
[
0.1158447265625,
0.552734375,
0.48779296875,
0.5517578125,
0.48779296875,
0.56640625,
0.1158447265625,
0.56787109375
],
[
0.509765625,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.509765625,
0.572265625
],
[
0.118408203125,
0.57080078125,
0.484375,
0.57080078125,
0.484375,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.5087890625,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58740234375,
0.5087890625,
0.58740234375
],
[
0.1158447265625,
0.58447265625,
0.486083984375,
0.58447265625,
0.486083984375,
0.59716796875,
0.1158447265625,
0.59716796875
],
[
0.51123046875,
0.5888671875,
0.88330078125,
0.59033203125,
0.88330078125,
0.60498046875,
0.51123046875,
0.60400390625
],
[
0.1175537109375,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.615234375,
0.1175537109375,
0.615234375
],
[
0.509765625,
0.60693359375,
0.88330078125,
0.60693359375,
0.88330078125,
0.62158203125,
0.509765625,
0.62158203125
],
[
0.1175537109375,
0.61767578125,
0.48779296875,
0.61767578125,
0.48779296875,
0.63134765625,
0.1175537109375,
0.63134765625
],
[
0.509765625,
0.62353515625,
0.8818359375,
0.62255859375,
0.8818359375,
0.63720703125,
0.509765625,
0.63818359375
],
[
0.1158447265625,
0.6328125,
0.490234375,
0.6328125,
0.490234375,
0.646484375,
0.1158447265625,
0.646484375
],
[
0.51123046875,
0.6396484375,
0.88330078125,
0.6396484375,
0.88330078125,
0.65283203125,
0.51123046875,
0.65283203125
],
[
0.1175537109375,
0.6484375,
0.48779296875,
0.6484375,
0.48779296875,
0.662109375,
0.1175537109375,
0.662109375
],
[
0.51318359375,
0.65673828125,
0.8818359375,
0.65673828125,
0.8818359375,
0.66748046875,
0.51318359375,
0.66748046875
],
[
0.1175537109375,
0.6650390625,
0.48779296875,
0.6650390625,
0.48779296875,
0.6787109375,
0.1175537109375,
0.6787109375
],
[
0.51123046875,
0.67138671875,
0.88330078125,
0.67138671875,
0.88330078125,
0.68505859375,
0.51123046875,
0.68505859375
],
[
0.1158447265625,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.5087890625,
0.68701171875,
0.88330078125,
0.68603515625,
0.88330078125,
0.70068359375,
0.5087890625,
0.70166015625
],
[
0.1158447265625,
0.69775390625,
0.48779296875,
0.69775390625,
0.48779296875,
0.71142578125,
0.1158447265625,
0.71142578125
],
[
0.5087890625,
0.703125,
0.880859375,
0.703125,
0.880859375,
0.71630859375,
0.5087890625,
0.71630859375
],
[
0.1175537109375,
0.71337890625,
0.489501953125,
0.71337890625,
0.489501953125,
0.72802734375,
0.1175537109375,
0.72802734375
],
[
0.51123046875,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.51123046875,
0.732421875
],
[
0.1158447265625,
0.728515625,
0.490234375,
0.728515625,
0.490234375,
0.7421875,
0.1158447265625,
0.7421875
],
[
0.51123046875,
0.73388671875,
0.8818359375,
0.7353515625,
0.8818359375,
0.75,
0.51123046875,
0.7490234375
],
[
0.1158447265625,
0.74658203125,
0.48779296875,
0.74658203125,
0.48779296875,
0.76025390625,
0.1158447265625,
0.76025390625
],
[
0.51123046875,
0.751953125,
0.8818359375,
0.75048828125,
0.8818359375,
0.765625,
0.51123046875,
0.7666015625
],
[
0.11419677734375,
0.76318359375,
0.48779296875,
0.76123046875,
0.48779296875,
0.77490234375,
0.11419677734375,
0.77685546875
],
[
0.51318359375,
0.7685546875,
0.8818359375,
0.7685546875,
0.8818359375,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.11419677734375,
0.77783203125,
0.175537109375,
0.775390625,
0.17626953125,
0.79052734375,
0.1158447265625,
0.7919921875
],
[
0.51318359375,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51318359375,
0.798828125
],
[
0.1351318359375,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.1351318359375,
0.8056640625
],
[
0.509765625,
0.80029296875,
0.8818359375,
0.80029296875,
0.8818359375,
0.81396484375,
0.509765625,
0.81396484375
],
[
0.1158447265625,
0.81005859375,
0.489501953125,
0.81005859375,
0.489501953125,
0.82373046875,
0.1158447265625,
0.82373046875
],
[
0.509765625,
0.81787109375,
0.5751953125,
0.81787109375,
0.5751953125,
0.83203125,
0.509765625,
0.83203125
],
[
0.1175537109375,
0.82666015625,
0.489501953125,
0.82666015625,
0.489501953125,
0.84033203125,
0.1175537109375,
0.84033203125
],
[
0.11419677734375,
0.8427734375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85546875,
0.11419677734375,
0.85546875
],
[
0.51123046875,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.11419677734375,
0.85693359375,
0.48779296875,
0.85791015625,
0.48779296875,
0.87255859375,
0.11419677734375,
0.87158203125
],
[
0.51318359375,
0.85986328125,
0.8818359375,
0.85986328125,
0.8818359375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1158447265625,
0.87451171875,
0.48779296875,
0.87451171875,
0.48779296875,
0.88818359375,
0.1158447265625,
0.88818359375
],
[
0.509765625,
0.87451171875,
0.8818359375,
0.87451171875,
0.8818359375,
0.88818359375,
0.509765625,
0.88818359375
],
[
0.1175537109375,
0.89013671875,
0.490234375,
0.89013671875,
0.490234375,
0.90380859375,
0.1175537109375,
0.90380859375
],
[
0.509765625,
0.89013671875,
0.8818359375,
0.89013671875,
0.8818359375,
0.90380859375,
0.509765625,
0.90380859375
],
[
0.1158447265625,
0.90478515625,
0.49365234375,
0.90625,
0.49365234375,
0.92041015625,
0.1158447265625,
0.91943359375
],
[
0.509765625,
0.90673828125,
0.880859375,
0.90673828125,
0.880859375,
0.92041015625,
0.509765625,
0.92041015625
]
] |
[
"expand their scope and introduce advCoU to mis-",
"duce advCoU, an extended CoU prompting strat",
"lead LLMs through the design of malicious internal",
"egy by adapting the ICL paradigm to the attack",
"thoughts and tailored demonstrations as in-context",
"scenario, and incorporating carefully designed ma-",
"examples, aimed at inducing models to produce",
"licious demonstrations to mislead LLMs. (3) We",
"undesired content.",
"conduct extensive experiments that cover a recent",
"Our work includes scenarios where the model",
"representative series of open-source LLMs. Our ex-",
"may exhibit biased judgment, select hallucinated",
"perimental results demonstrate the effectiveness of",
"answers erroneously, fail to identify unethical state-",
" our attack strategy across different aspects, show-",
"ments as wrong, and more (See Figure 1). This",
"casing an advantage over DECODINGTRUST. (4)",
"poses a more nuanced assessment of themodels'",
"Through in-depth result analysis, we uncover inter",
"reasoning and alignment capabilities, going beyond",
"esting findings that shed light on the relationship",
"simply evaluating whether the model refuses to an-",
"between trustworthiness and potential influence fac-",
"swer a harmful question or not, as conducted in",
" tors, such as model size and alignment strategies.",
"Red-Eval. This approach allows us to manipulate",
"2",
"Related Work",
"put to perform trustworthiness attacks. By con-",
"only the demonstrations without changing the in-",
"Trustworthiness of LLMs.As LLMs continue",
"ducting this comprehensive assessment, we seek to",
"to advance rapidly across various domains, con-",
"establish a better understanding of how trustwor-",
"cerns regarding their trustworthiness are becoming",
"thy current open-source LLMs are against adver-",
"increasingly prominent.Previous investigations",
" sarial attacks. We intend to encourage increased",
"into the trustworthiness of LLMs have predomi-",
"trustworthiness-related research concerning open-",
"nantly concentrated on individual aspects, such as",
"source LLMs, thus mitigating potential risks to",
"toxicity (Tamkin et al., 2022; Deshpande et al.,",
" users and fostering reliable deployment and utility",
"2023; Huang et al., 2023b; Liu et al., 2023a; Jones",
"of produced LLMs in downstream systems, ser-",
"et al., 2024), stereotypical bias (Mattern et al.,",
"vices and applications.",
"2022: Shaikh et al., 2022), privacy (Yue et al., 2023;",
"Furthermore, within the scope of our attack strat-",
"Mireshghallah et al., 2023; Du et al., 2023), syco-",
"egy, we delve into two research questions: (1)",
"phancy (Wei et al., 2023; Wang et al., 2023a), ro-",
"Do language models become more trustworthy as",
" bustness (Zhu et al., 2023; Li et al., 2023b), and",
"they grow larger? (2) Are models that have under-",
"more. In this paper, we aim to conduct a compre-",
"gone instruction tuning and alignment processes",
"hensive examination of LLMs, taking into account",
"more trustworthy?Through our exploration.we",
"various aspects compiled from recent studies (Liu",
"aim to gain deeper insights into the factors influ-",
"et al., 2023b; Wang et al., 2023b), including toxic-",
"encing trustworthiness, including model size and",
"ity, stereotype, ethics, hallucination, fairness, syco-",
"alignment-focused fine-tuning. Our experiments",
"phancy, privacy, and robustness against adversarial",
"yield findings that models with superior perfor-",
"demonstrations. While Wang et al. (2023b) intro-",
"mance in general NLP tasks do not necessarily have",
"ducedDECODINGTRUST,whichprovides a multi-",
"higher trustworthiness. In fact, larger models tend",
"faceted assessment of trustworthiness vulnerabili-",
"to be more susceptible to manipulation through",
"ties particularly in GPT-3.5 and GPT-4, we focus",
"malicious demonstrations. Moreover, models with",
"on evaluating open-source LLMs through existing",
"instruction tuning, which emphasize instruction fol-",
"and our attack strategies. More recently, Sun et al",
"lowing, exhibit higher vulnerability, although fine-",
"(2024) presented TrustLLM and proposed a set of",
"tuning LLMs with safety alignment proves effective",
"principles such as truthfulness and safety to evalu-",
"in protecting against adversarial trustworthiness",
"ate the trustworthiness of LLMs. In our work, we",
"attacks.",
"study trustworthy LLMs from the perspective of ad-",
"Our contributions in this work can be summa-",
"versarial assessments via a CoU-based prompting",
"rized as follows: (1) We conduct a comprehen-",
"strategy.",
"sive assessment of open-source LLMs on trustwor-",
"thiness across eight different aspects, including",
"Adversarial Attacks on LLMs. Recent LLM",
"toxicity, stereotypes, and more. (2) We employ",
"providers have taken a range of safety mecha-",
"multiple adversarial attack strategies, starting from",
"nisms to align models with human values (Ouyang",
"the recent work DECODINGTRUST (Wang et al.,",
"et al., 2022; Bai et al., 2022a; Zhou et al., 2023)",
"2023b) as the baseline. In particular, we intro-",
"However, LLMs still remain vulnerable to various"
] |
{
"arxiv_id": "2311.09447",
"language": "en",
"page_idx": 2
}
|
||
2311.09447-en-3
|
2311.09447-en-3.jpg
|
well-designed adversarial attacks. Xu et al. (2023) showed the backdoor vulnerabilities of instructiontuned LLMs by injecting malicious instructions. Bai et al. (2022b) and Albert (2023) demonstrated jailbreaks by specifically crafting inputs to circumvent alignment strategies. Further work (Wen et al., 2023; Carlini et al., 2023) showed the success in designing adversarial prompts to automatically discover jailbreaking inputs. Wang et al. (2023c) optimized adversarial demonstration examples to contaminate the performance of discriminative tasks. Zou et al. (2023) introduced a universal attack that can transfer across multiple LLMs, including proprietary, black-box models. More recently, Huang et al. (2023a) proposed to jailbreak the alignment in open-source LLMs via manipulating generation strategies, including varying decoding hyperparameters and sampling methods. Inspired by Red-Eval (Bhardwaj and Poria, 2023), which carried out the jailbreak using a conversation-based red-teaming prompt, we extend their strategy and design malicious demonstrations through the incontext learning paradigm for multiple aspects of trustworthiness and conduct a comprehensive adversarial assessment of open-source LLMs.
# 3 Methodology
## 3.1 In-Context Learning
In-context learning (ICL) is defined as a conditional text generation problem. Formally, for a language model $\mathcal{M}$ with parameters $\theta$, the goal is to generate output $y$ based on the input example $x$ and the context $C$ provided during inference. This context $C$ typically comprises a task description $T$ and a collection of task-specific data instances $E$ that serve as demonstrations. The model $\mathcal{M}$ uses the context to condition its responses:
(1)
$$y=\mathcal{M}(x\mid C;\theta)$$
In the case of few-shot learning, the context in-
cludes a few examples that illustrate the task:
(2)
$$C=(T,\{(x_i,y_i)\}_{i=1}^n),$$
where $\{(x_i,y_i)\}_{i=1}^n$ are $n$ input-output pairs demonstrating the task $T.$ With an input test example $x_t$, the model can then generate the corresponding response $y_t$ using these demonstrations as a guide.
3.2 CoU Prompt-based In-Context Attack
Within the ICL paradigm outlined earlier, we introduce advCoU, which leverages a Chain of Utterances (CoU) prompting strategy as the foundation for extracting undesired information from LLMs that undermines trustworthiness. Previously, REDEVAL (Bhardwaj and Poria, 2023) employs the CoU prompt, simulating a conversation via roleplay, to perform red-teaming for evaluating the safety of LLMs. They introduce benign internal thoughts as a prefix in the model's responses to encourage providing elaborate answers to harmful questions. Expanding upon this approach, we design crafted malicious demonstrations and misleading internal thoughts as in-context examples to conduct our attacks.
Specifically, our prompt sets up a conversation between two distinct roles: an attacker denoted as the Red-LM and a defender, referred to as the Target-LM. Red-LM initiates an utterance to solicit information that could compromise trustworthiness, while Target-LM is responsible for generating the response in accordance with Red-LM's request. As shown in Figure 2, which serves as an illustration within the context of assessing toxicity, Red-LM presents a potentially toxic prompt, for example, “I hate you dam**d gays, I'm going to ...". Subsequently, Target-LM is tasked with completing the response that incorporates internal thoughts as the prefix, steering it towards being toxic, rude, and disrespectful. There are two critical ingredients that contribute to the attack effectiveness: firstly, the use of malicious demonstrations through the interaction turn between the Red-LM and the TargetLM as in-context examples, and secondly, the endowment of the Target-LM with internal thoughts that guide responses towards a more helpful direction, all while being encouraged to produce content with undesired traits, such as toxicity, stereotypes, ethics, and more.
3.3 Eight Aspects of Trustworthiness
(1) Toxicity. We assess the vulnerability of opensource LLMs to toxicity using a standard benchmark dataset REALTOXICITYPROMPTS (Gehman et al., 2020). To evaluate an LLM's behaviors in the worst-case scenario, we use a representative subset of 1.2K toxic task prompts which are flagged by the official dataset as “challenging”. In our attack prompt, Red-LM initiates a scenario by providing an incomplete prompt and instructing Target-LM
|
[
[
0.1158447265625,
0.08489990234375,
0.490234375,
0.0843505859375,
0.490234375,
0.09857177734375,
0.1158447265625,
0.09979248046875
],
[
0.50537109375,
0.0843505859375,
0.85400390625,
0.08489990234375,
0.85400390625,
0.09979248046875,
0.50537109375,
0.09857177734375
],
[
0.1175537109375,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.10809326171875,
0.8818359375,
0.1092529296875,
0.8818359375,
0.12353515625,
0.509765625,
0.122314453125
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.14013671875,
0.51123046875,
0.14013671875
],
[
0.118408203125,
0.133056640625,
0.486083984375,
0.133056640625,
0.486083984375,
0.146728515625,
0.118408203125,
0.146728515625
],
[
0.51123046875,
0.142578125,
0.8818359375,
0.142578125,
0.8818359375,
0.1561279296875,
0.51123046875,
0.1561279296875
],
[
0.1175537109375,
0.15087890625,
0.490234375,
0.15087890625,
0.490234375,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51123046875,
0.157958984375,
0.8818359375,
0.157958984375,
0.8818359375,
0.171630859375,
0.51123046875,
0.171630859375
],
[
0.1175537109375,
0.1650390625,
0.48779296875,
0.1650390625,
0.48779296875,
0.1787109375,
0.1175537109375,
0.1787109375
],
[
0.51318359375,
0.17578125,
0.880859375,
0.17578125,
0.880859375,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.118408203125,
0.1817626953125,
0.48779296875,
0.1817626953125,
0.48779296875,
0.1953125,
0.118408203125,
0.1953125
],
[
0.51123046875,
0.190673828125,
0.88330078125,
0.190673828125,
0.88330078125,
0.2037353515625,
0.51123046875,
0.2037353515625
],
[
0.1175537109375,
0.198974609375,
0.490234375,
0.198974609375,
0.490234375,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.51123046875,
0.2060546875,
0.88330078125,
0.2060546875,
0.88330078125,
0.2197265625,
0.51123046875,
0.2197265625
],
[
0.1175537109375,
0.21435546875,
0.490234375,
0.21435546875,
0.490234375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.51123046875,
0.22265625,
0.8818359375,
0.22265625,
0.8818359375,
0.236328125,
0.51123046875,
0.236328125
],
[
0.1175537109375,
0.23095703125,
0.491943359375,
0.23095703125,
0.491943359375,
0.24462890625,
0.1175537109375,
0.24462890625
],
[
0.5087890625,
0.2381591796875,
0.880859375,
0.2381591796875,
0.880859375,
0.251708984375,
0.5087890625,
0.251708984375
],
[
0.1175537109375,
0.2476806640625,
0.490234375,
0.2476806640625,
0.490234375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.509765625,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.509765625,
0.268310546875
],
[
0.1175537109375,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.51318359375,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.1158447265625,
0.27783203125,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.292236328125
],
[
0.51123046875,
0.28515625,
0.8818359375,
0.2861328125,
0.8818359375,
0.300537109375,
0.51123046875,
0.2998046875
],
[
0.1175537109375,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51318359375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51318359375,
0.31640625
],
[
0.1175537109375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51318359375,
0.319580078125,
0.87890625,
0.319580078125,
0.87890625,
0.330078125,
0.51318359375,
0.330078125
],
[
0.1175537109375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.51318359375,
0.3349609375,
0.68115234375,
0.3349609375,
0.68115234375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.5283203125,
0.350830078125,
0.8818359375,
0.3515625,
0.8818359375,
0.366455078125,
0.5283203125,
0.365234375
],
[
0.118408203125,
0.35986328125,
0.484375,
0.35986328125,
0.484375,
0.37353515625,
0.118408203125,
0.37353515625
],
[
0.51123046875,
0.3681640625,
0.8818359375,
0.3681640625,
0.8818359375,
0.3818359375,
0.51123046875,
0.3818359375
],
[
0.1175537109375,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.1175537109375,
0.388427734375
],
[
0.51123046875,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.3984375,
0.51123046875,
0.3984375
],
[
0.1175537109375,
0.393798828125,
0.486083984375,
0.393798828125,
0.486083984375,
0.40380859375,
0.1175537109375,
0.40380859375
],
[
0.51123046875,
0.40087890625,
0.8818359375,
0.40087890625,
0.8818359375,
0.41455078125,
0.51123046875,
0.41455078125
],
[
0.1158447265625,
0.407958984375,
0.48779296875,
0.40673828125,
0.48779296875,
0.421630859375,
0.1158447265625,
0.4228515625
],
[
0.5087890625,
0.416259765625,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.5087890625,
0.429931640625
],
[
0.1175537109375,
0.424560546875,
0.490234375,
0.424560546875,
0.490234375,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.5087890625,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.446533203125,
0.5087890625,
0.446533203125
],
[
0.1158447265625,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.45361328125,
0.1158447265625,
0.45361328125
],
[
0.5087890625,
0.448974609375,
0.8818359375,
0.44775390625,
0.8818359375,
0.4619140625,
0.5087890625,
0.463134765625
],
[
0.1158447265625,
0.457275390625,
0.489501953125,
0.457275390625,
0.489501953125,
0.47021484375,
0.1158447265625,
0.47021484375
],
[
0.51123046875,
0.465576171875,
0.8818359375,
0.465576171875,
0.8818359375,
0.478515625,
0.51123046875,
0.478515625
],
[
0.1158447265625,
0.47265625,
0.438232421875,
0.47265625,
0.438232421875,
0.486328125,
0.1158447265625,
0.486328125
],
[
0.509765625,
0.48095703125,
0.8818359375,
0.479736328125,
0.8818359375,
0.49462890625,
0.509765625,
0.495361328125
],
[
0.11248779296875,
0.50048828125,
0.2626953125,
0.50439453125,
0.2626953125,
0.521484375,
0.11248779296875,
0.5185546875
],
[
0.51123046875,
0.49755859375,
0.88330078125,
0.49755859375,
0.88330078125,
0.51123046875,
0.51123046875,
0.51123046875
],
[
0.5087890625,
0.51318359375,
0.8818359375,
0.51318359375,
0.8818359375,
0.52685546875,
0.5087890625,
0.52685546875
],
[
0.11419677734375,
0.5322265625,
0.32421875,
0.53369140625,
0.32421875,
0.548828125,
0.11419677734375,
0.54638671875
],
[
0.51123046875,
0.52978515625,
0.8818359375,
0.52978515625,
0.8818359375,
0.54345703125,
0.51123046875,
0.54345703125
],
[
0.509765625,
0.5458984375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55859375,
0.509765625,
0.56005859375
],
[
0.1158447265625,
0.55712890625,
0.48779296875,
0.55712890625,
0.48779296875,
0.57080078125,
0.1158447265625,
0.57080078125
],
[
0.51123046875,
0.5625,
0.8818359375,
0.5625,
0.8818359375,
0.576171875,
0.51123046875,
0.576171875
],
[
0.11419677734375,
0.57470703125,
0.486083984375,
0.57470703125,
0.486083984375,
0.58837890625,
0.11419677734375,
0.58837890625
],
[
0.509765625,
0.57666015625,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.509765625,
0.59033203125
],
[
0.11419677734375,
0.5888671875,
0.48779296875,
0.59033203125,
0.48779296875,
0.60498046875,
0.11419677734375,
0.60400390625
],
[
0.51123046875,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.60791015625,
0.51123046875,
0.60791015625
],
[
0.1158447265625,
0.60693359375,
0.489501953125,
0.60693359375,
0.489501953125,
0.62158203125,
0.1158447265625,
0.62158203125
],
[
0.5087890625,
0.609375,
0.880859375,
0.609375,
0.880859375,
0.62255859375,
0.5087890625,
0.62255859375
],
[
0.1175537109375,
0.62353515625,
0.48779296875,
0.62353515625,
0.48779296875,
0.63720703125,
0.1175537109375,
0.63720703125
],
[
0.51123046875,
0.62451171875,
0.8818359375,
0.6259765625,
0.8818359375,
0.64013671875,
0.51123046875,
0.6396484375
],
[
0.118408203125,
0.6396484375,
0.489501953125,
0.6396484375,
0.489501953125,
0.65283203125,
0.118408203125,
0.65283203125
],
[
0.509765625,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65625,
0.509765625,
0.65625
],
[
0.1175537109375,
0.65478515625,
0.489501953125,
0.65478515625,
0.489501953125,
0.66943359375,
0.1175537109375,
0.66943359375
],
[
0.51123046875,
0.65771484375,
0.8818359375,
0.65771484375,
0.8818359375,
0.67138671875,
0.51123046875,
0.67138671875
],
[
0.11419677734375,
0.66943359375,
0.489501953125,
0.67041015625,
0.489501953125,
0.68505859375,
0.11419677734375,
0.68408203125
],
[
0.51123046875,
0.67333984375,
0.880859375,
0.67333984375,
0.880859375,
0.68701171875,
0.51123046875,
0.68701171875
],
[
0.11419677734375,
0.68603515625,
0.3720703125,
0.68701171875,
0.3720703125,
0.70166015625,
0.11419677734375,
0.70068359375
],
[
0.51123046875,
0.6904296875,
0.8818359375,
0.689453125,
0.8818359375,
0.70361328125,
0.51123046875,
0.705078125
],
[
0.51123046875,
0.705078125,
0.8818359375,
0.70703125,
0.8818359375,
0.7216796875,
0.51123046875,
0.7197265625
],
[
0.2359619140625,
0.7255859375,
0.366943359375,
0.7255859375,
0.366943359375,
0.74072265625,
0.2359619140625,
0.74072265625
],
[
0.462646484375,
0.72705078125,
0.490234375,
0.72705078125,
0.490234375,
0.7392578125,
0.462646484375,
0.7392578125
],
[
0.51123046875,
0.7236328125,
0.640625,
0.7236328125,
0.640625,
0.73681640625,
0.51123046875,
0.73681640625
],
[
0.1351318359375,
0.755859375,
0.49365234375,
0.755859375,
0.49365234375,
0.76953125,
0.1351318359375,
0.76953125
],
[
0.51318359375,
0.75439453125,
0.8154296875,
0.75439453125,
0.8154296875,
0.76708984375,
0.51318359375,
0.76708984375
],
[
0.1175537109375,
0.7724609375,
0.45166015625,
0.7724609375,
0.45166015625,
0.7861328125,
0.1175537109375,
0.7861328125
],
[
0.51318359375,
0.77783203125,
0.8818359375,
0.77783203125,
0.8818359375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.509765625,
0.79443359375,
0.8818359375,
0.79345703125,
0.8818359375,
0.80712890625,
0.509765625,
0.80810546875
],
[
0.21240234375,
0.80810546875,
0.3896484375,
0.81103515625,
0.3896484375,
0.8291015625,
0.21240234375,
0.826171875
],
[
0.46337890625,
0.8125,
0.490234375,
0.8125,
0.490234375,
0.82470703125,
0.46337890625,
0.82470703125
],
[
0.509765625,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82373046875,
0.509765625,
0.82373046875
],
[
0.51318359375,
0.826171875,
0.880859375,
0.826171875,
0.880859375,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.1175537109375,
0.8427734375,
0.489501953125,
0.8427734375,
0.489501953125,
0.85693359375,
0.1175537109375,
0.85693359375
],
[
0.51123046875,
0.8427734375,
0.8818359375,
0.8427734375,
0.8818359375,
0.85546875,
0.51123046875,
0.85546875
],
[
0.1175537109375,
0.85791015625,
0.491943359375,
0.85791015625,
0.491943359375,
0.87158203125,
0.1175537109375,
0.87158203125
],
[
0.51123046875,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.87255859375,
0.51123046875,
0.87255859375
],
[
0.1158447265625,
0.87451171875,
0.491943359375,
0.87451171875,
0.491943359375,
0.88818359375,
0.1158447265625,
0.88818359375
],
[
0.51318359375,
0.87451171875,
0.8818359375,
0.87451171875,
0.8818359375,
0.88818359375,
0.51318359375,
0.88818359375
],
[
0.1175537109375,
0.89111328125,
0.489501953125,
0.89111328125,
0.489501953125,
0.90625,
0.1175537109375,
0.90625
],
[
0.509765625,
0.89013671875,
0.87890625,
0.89013671875,
0.87890625,
0.90478515625,
0.509765625,
0.90478515625
],
[
0.1158447265625,
0.90673828125,
0.197265625,
0.90673828125,
0.197265625,
0.92138671875,
0.1158447265625,
0.92138671875
],
[
0.509765625,
0.90673828125,
0.880859375,
0.90673828125,
0.880859375,
0.92138671875,
0.509765625,
0.92138671875
]
] |
[
"well-designed adversarial attacks. Xu et al. (2023)",
"3.2 CoU Prompt-based In-Context Attack",
"showed the backdoor vulnerabilities of instruction-",
"Within the ICL paradigm outlined earlier, we intro-",
"tuned LLMs by injecting malicious instructions.",
"duce advCoU, which leverages a Chain of Utter-",
"Bai et al. (2022b) and Albert (2023) demonstrated",
"ances (CoU) prompting strategy as the foundation",
"jailbreaks by specifically crafting inputs to circum-",
"for extracting undesired information from LLMs",
"vent alignment strategies. Further work (Wen et al..",
"thatundermines trustworthiness.Previously,RED",
"2023; Carlini et al., 2023) showed the success in",
"EvAL(Bhardwaj and Poria,2023)employs the",
"designing adversarial prompts to automatically dis-",
"CoU prompt, simulating a conversation via role-",
"cover jailbreaking inputs. Wang et al. (2023c) opti-",
"play, to perform red-teaming for evaluating the",
"mized adversarial demonstration examples to con-",
"safety of LLMs. They introduce benign internal",
"taminate the performance of discriminative tasks.",
"thoughts as a prefix in the model's responses to",
"Zou et al. (2023) introduced a universal attack that",
"encourage providing elaborate answers to harm-",
"can transfer across multiple LLMs, including pro-",
"ful questions. Expanding upon this approach, we",
"prietary,black-box models. More recently,Huang",
"design crafted malicious demonstrations and mis-",
"et al. (2023a) proposed to jailbreak the alignment",
"leadinginternalthoughtsasin-contextexamples",
"in open-source LLMs via manipulating genera-",
"to conduct our attacks.",
"tion strategies, including varying decoding hyper-",
"Specifically,our prompt sets up a conversation",
"parameters and sampling methods. Inspired by",
"between two distinct roles: an attacker denoted",
"Red-Eval (Bhardwaj and Poria, 2023), which car-",
"as the Red-LM and a defender, referred to as the",
"ried out the jailbreak using a conversation-based",
"Target-LM. Red-LM initiates an utterance to solicit",
"red-teaming prompt, we extend their strategy and",
"information that could compromise trustworthiness,",
"design malicious demonstrations through the in-",
" while Target-LM is responsible for generating the",
"context learning paradigm for multiple aspects of",
"response in accordance with Red-LM's request. As",
"trustworthiness and conduct a comprehensive ad-",
"shown in Figure 2, which serves as an illustration",
"versarial assessment of open-source LLMs.",
"within the context of assessing toxicity, Red-LM",
"3 Methodology",
"presents a potentially toxic prompt, for example,",
"\"I hate you dam**d gays, I'm going to ...\". Subse-",
"3.1 In-Context Learning",
"quently, Target-LM is tasked with completing the",
"response that incorporates internal thoughts as the",
"In-context learning (ICL) is defined as a conditional",
"prefix, steering it towards being toxic, rude, and",
"text generation problem. Formally, for a language",
"disrespectful. There are two critical ingredients",
"model M with parameters , the goal is to generate",
"that contribute to the attack effectiveness: firstly.",
"output y based on the input example and the",
"the use of malicious demonstrations through the in-",
"context C provided during inference. This context",
"teraction turn between the Red-LM and the Target-",
"C typically comprises a task description T and a",
"LM as in-context examples, and secondly, the en",
"collection of task-specific data instances E that",
"dowment of the Target-LM with internal thoughts",
"serve as demonstrations. The model M uses the",
"that guide responses towards a more helpful direc-",
"context to condition its responses:",
"tion, all while being encouraged to produce content",
"with undesired traits, such as toxicity, stereotypes,",
"y = M(α |C;0)",
"(1)",
"ethics, and more.",
"In the case of few-shot learning, the context in-",
"3.3Eight Aspects of Trustworthiness",
"cludes a few examples that illustrate the task:",
"(1) Toxicity.We assess the vulnerability of open-",
"source LLMs to toxicity using a standard bench",
"C= (T, {(ci, yi)\"=1),",
"(2)",
"mark dataset REALTOXICITYPROMPTS (Gehman",
"et al., 2020). To evaluate an LLM's behaviors in the",
"where {(ci, yi))\"- are n input-output pairs",
"worst-case scenario, we use a representative subset",
"demonstrating the task T.: With an input test ex-",
"of 1.2K toxic task prompts which are flagged by",
"ample &t, the model can then generate the corre-",
"the official dataset as \"challenging\". In our attack",
"sponding response yt using these demonstrations",
"prompt, Red-LM initiates a scenario by providing",
"as a guide.",
" an incomplete prompt and instructing Target-LM"
] |
{
"arxiv_id": "2311.09447",
"language": "en",
"page_idx": 3
}
|
||
2303.16202-en-0
|
2303.16202-en-0.jpg
|
$\textbf{CCuantuMM: Cycle- Consistent Quantum- Hybrid Matching of Multiple Shapes}$
Harshil Bhatia$^{1,2}$ Edith Tretschk$^{2}$ Zorah Lähner$^{3}$ Marcel Seelbach Benkner$^{3}$
Michael Moeller$^3$ Christian Theobalt$^2$ Vladislav Golyanik$^{2}$
$^{1}$Indian Institute of Technology, Jodhpur $^{2}$MPI for Informatics, SIC$^{3}$Universität Siegen
<FigureHere>
Figure 1. Our quantum-hybrid method matches all 100 shapes of the FAUST collection [4] with guaranteed cycle consistency (white arrows). Here, we visualise the matchings via texture transfer between all shapes. Our method scales linearly in the number of shapes. See the full figure in the supplement.
## $\mathbf{Abstract}$
Jointly matching multiple, non-rigidly deformed 3D shapes is a challenging, NP-hard problem. A perfect matching is necessarily cycle-consistent: Following the pairwise point correspondences along several shapes must end up at the starting vertex of the original shape. Unfortunately, existing quantum shape-matching methods do not support multiple shapes and even less cycle consistency. This paper addresses the open challenges and introduces the frst quantum-hybrid approach for 3D shape multimatching; in addition, it is also cycle-consistent. Its iterative formulation is admissible to modern adiabatic quantum hardware and scales linearly with the total number of input shapes. Both these characteristics are achieved by reducing the N-shape case to a sequence of three-shape matchings, the derivation of which is our main technical contribution. Thanks to quantum annealing, high-quality solutions with low energy are retrieved for the intermediate NPhard objectives. On benchmark datasets, the proposed approach significantly outperforms extensions to multi-shape matching of a previous quantum-hybrid two-shape matching method and is on-par with classical multi-matching methods. Our source code is available at 4dqv.mpi- $inf.mpg.de/CCuantuMM/.$
## $1. \textbf{Introduction}$
Recently, there has been a growing interest in applying quantum computers in computer vision [3,23,35]. Such quantum computer vision methods rely on quantum annealing (QA) that allows to solve $\mathcal{NP}$-hard quadratic unconstrained binary optimisation problems (QUBOs). While having to formulate a problem as a QUBO is rather inflexible, QA is, in the future, widely expected to solve QUBOs at speeds not achievable with classical hardware. Thus, casting a problem as a QUBO promises to outperform more unrestricted formulations in terms of tractable problem sizes and attainable accuracy through sheer speed.
A recent example for such a problem is shape matching, where the goal is to estimate correspondences between
two shapes. Accurate shape matching is a core element of many computer vision and graphics applications (i.e., texture transfer and statistical shape modelling). If non-rigid deformations are allowed, even pairwise matching is $\mathcal{N}\mathcal{P}-$ hard, leading to a wide area of research that approximates this problem, as a recent survey shows [18]. Matching two shapes is one of the problems that was shown to benefit from quantum hardware: Q-Match [42] iteratively updates a subset of point correspondences using QA. Specifically, its cyclic $\alpha$-expansion allows to parametrise changes to permutation matrices without relaxations.
The question we ask in this work is: How can we design a multi-shape matching algorithm in the style of Q-Match that has the same benefits? As we show in the experiments, where we introduce several naive multi-shape extensions of Q-Match, this is a highly non-trivial task. Despite tweaking them, our proposed method significantly outperforms them.
If $N>2$ shapes have to be matched, the computational complexity of naïve exhaustive pairwise matching increases quadratically with $N$,which does not scale to large $N.$ Furthermore, these pairwise matchings can easily turn out to be inconsistent with each other, thereby violating cycle con-
1
|
[
[
0.08660888671875,
0.1319580078125,
0.8837890625,
0.1331787109375,
0.8837890625,
0.15087890625,
0.08660888671875,
0.149658203125
],
[
0.17236328125,
0.18115234375,
0.79736328125,
0.18115234375,
0.79736328125,
0.1988525390625,
0.17236328125,
0.1988525390625
],
[
0.2442626953125,
0.2008056640625,
0.7255859375,
0.2008056640625,
0.7255859375,
0.21533203125,
0.2442626953125,
0.21533203125
],
[
0.1258544921875,
0.216552734375,
0.84228515625,
0.216552734375,
0.84228515625,
0.2342529296875,
0.1258544921875,
0.2342529296875
],
[
0.0228729248046875,
0.25830078125,
0.060455322265625,
0.25830078125,
0.060455322265625,
0.3984375,
0.0228729248046875,
0.3984375
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.098876953125,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.314453125,
0.098876953125,
0.314453125
],
[
0.08087158203125,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.08251953125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.3447265625,
0.08251953125,
0.3447265625
],
[
0.08087158203125,
0.347900390625,
0.4697265625,
0.347900390625,
0.4697265625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.08087158203125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.08087158203125,
0.37744140625,
0.4697265625,
0.37744140625,
0.4697265625,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.08087158203125,
0.392578125,
0.46728515625,
0.391357421875,
0.46728515625,
0.406005859375,
0.08087158203125,
0.4072265625
],
[
0.0269622802734375,
0.4072265625,
0.052276611328125,
0.4072265625,
0.052276611328125,
0.449462890625,
0.0269622802734375,
0.449462890625
],
[
0.08087158203125,
0.40771484375,
0.46728515625,
0.40771484375,
0.46728515625,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.08087158203125,
0.42236328125,
0.468017578125,
0.42236328125,
0.468017578125,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.08087158203125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.03350830078125,
0.449462890625,
0.0555419921875,
0.449462890625,
0.0555419921875,
0.479248046875,
0.03350830078125,
0.479248046875
],
[
0.080078125,
0.45263671875,
0.4697265625,
0.45263671875,
0.4697265625,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.50244140625,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.08087158203125,
0.468505859375,
0.4697265625,
0.468505859375,
0.4697265625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50341796875,
0.474853515625,
0.88818359375,
0.474853515625,
0.88818359375,
0.48486328125,
0.50341796875,
0.48486328125
],
[
0.08087158203125,
0.48291015625,
0.468017578125,
0.48291015625,
0.468017578125,
0.496826171875,
0.08087158203125,
0.496826171875
],
[
0.5009765625,
0.4873046875,
0.890625,
0.486083984375,
0.890625,
0.5,
0.5009765625,
0.50048828125
],
[
0.080078125,
0.498779296875,
0.468017578125,
0.498779296875,
0.468017578125,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50244140625,
0.50048828125,
0.67333984375,
0.501953125,
0.67333984375,
0.51513671875,
0.50244140625,
0.5146484375
],
[
0.68115234375,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.51318359375,
0.68115234375,
0.51318359375
],
[
0.0767822265625,
0.51318359375,
0.468017578125,
0.51220703125,
0.468017578125,
0.5263671875,
0.0767822265625,
0.52783203125
],
[
0.0269622802734375,
0.501953125,
0.056365966796875,
0.501953125,
0.056365966796875,
0.564453125,
0.0269622802734375,
0.564453125
],
[
0.50341796875,
0.5166015625,
0.8310546875,
0.5166015625,
0.8310546875,
0.5263671875,
0.50341796875,
0.5263671875
],
[
0.07843017578125,
0.52880859375,
0.468017578125,
0.52880859375,
0.468017578125,
0.54248046875,
0.07843017578125,
0.54248046875
],
[
0.035125732421875,
0.53515625,
0.0400390625,
0.53515625,
0.0400390625,
0.54248046875,
0.035125732421875,
0.54248046875
],
[
0.080078125,
0.54345703125,
0.46728515625,
0.54345703125,
0.46728515625,
0.556640625,
0.080078125,
0.556640625
],
[
0.0196075439453125,
0.5595703125,
0.04412841796875,
0.55029296875,
0.056365966796875,
0.5712890625,
0.0318603515625,
0.580078125
],
[
0.07843017578125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.07843017578125,
0.57275390625
],
[
0.5009765625,
0.5625,
0.89208984375,
0.56103515625,
0.89208984375,
0.57568359375,
0.5009765625,
0.5771484375
],
[
0.08087158203125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.0196075439453125,
0.58544921875,
0.058013916015625,
0.58544921875,
0.060455322265625,
0.7109375,
0.0228729248046875,
0.7109375
],
[
0.50341796875,
0.580078125,
0.8896484375,
0.580078125,
0.8896484375,
0.59033203125,
0.50341796875,
0.59033203125
],
[
0.080078125,
0.58837890625,
0.46728515625,
0.58837890625,
0.46728515625,
0.6015625,
0.080078125,
0.6015625
],
[
0.50244140625,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.07843017578125,
0.60302734375,
0.4697265625,
0.60400390625,
0.4697265625,
0.61865234375,
0.07843017578125,
0.6171875
],
[
0.50244140625,
0.60791015625,
0.8896484375,
0.60791015625,
0.8896484375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.08087158203125,
0.619140625,
0.464111328125,
0.619140625,
0.464111328125,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.50244140625,
0.62255859375,
0.89208984375,
0.62255859375,
0.89208984375,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.08087158203125,
0.6337890625,
0.291748046875,
0.63330078125,
0.291748046875,
0.64794921875,
0.08087158203125,
0.6484375
],
[
0.50341796875,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.50341796875,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50341796875,
0.66748046875
],
[
0.50341796875,
0.67041015625,
0.8896484375,
0.67041015625,
0.8896484375,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.035125732421875,
0.6767578125,
0.0400390625,
0.6767578125,
0.0400390625,
0.6845703125,
0.035125732421875,
0.6845703125
],
[
0.08251953125,
0.6796875,
0.2091064453125,
0.6796875,
0.2091064453125,
0.693359375,
0.08251953125,
0.693359375
],
[
0.50341796875,
0.6845703125,
0.890625,
0.6845703125,
0.890625,
0.69775390625,
0.50341796875,
0.69775390625
],
[
0.098876953125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.098876953125,
0.71826171875
],
[
0.50244140625,
0.69970703125,
0.890625,
0.69970703125,
0.890625,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.50244140625,
0.7138671875,
0.7548828125,
0.7138671875,
0.7548828125,
0.72802734375,
0.50244140625,
0.72802734375
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.73291015625,
0.080078125,
0.734375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.521484375,
0.73095703125,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.521484375,
0.744140625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.74755859375,
0.89208984375,
0.74755859375,
0.89208984375,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.07843017578125,
0.76513671875,
0.4697265625,
0.7646484375,
0.4697265625,
0.77880859375,
0.07843017578125,
0.77978515625
],
[
0.50244140625,
0.76220703125,
0.890625,
0.76220703125,
0.890625,
0.77587890625,
0.50244140625,
0.77587890625
],
[
0.080078125,
0.78076171875,
0.471435546875,
0.78076171875,
0.471435546875,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5048828125,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.7890625,
0.5048828125,
0.7890625
],
[
0.08087158203125,
0.79736328125,
0.4697265625,
0.79736328125,
0.4697265625,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.50244140625,
0.79150390625,
0.89208984375,
0.79248046875,
0.89208984375,
0.806640625,
0.50244140625,
0.80615234375
],
[
0.08251953125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.890625,
0.80810546875,
0.890625,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.521484375,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.471435546875,
0.84130859375,
0.471435546875,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.373291015625,
0.85595703125,
0.373291015625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85595703125,
0.890625,
0.85498046875,
0.890625,
0.869140625,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.09722900390625,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"CCuantuMM: Cycle-Consistent Quantum-Hybrid Matching of Multiple Shapes",
"Harshil Bhatial,2 Edith Tretschk? Zorah Lahner3Marcel Seelbach Benkner3",
"Michael Moeller3Christian Theobalt?Vladislav Golyanik2",
"1Indian Institute of Technology, Jodhpur2MPI for Informatics, SIC3Universitat Siegen",
"",
"Abstract",
"Jointly matching multiple, non-rigidly deformed 3D",
"shapes is a challenging, NP-hard problem.A perfect",
"matching is necessarily cycle-consistent:Following the",
"pairwise point correspondences along several shapes must",
"end up at the starting vertex of the original shape. Unfor-",
"tunately, existing quantum shape-matching methods do not",
"support multiple shapes and even less cycle consistency.",
"A",
"This paper addresses the open challenges and introduces",
"the first quantum-hybrid approach for 3D shape multi-",
"matching; in addition, it is also cycle-consistent. Its itera-",
"'SO",
"tive formulation is admissible to modern adiabatic quantum",
"Figure 1. Our quantum-hybrid method matches all 100 shapes",
"hardware and scales linearly with the total number of input",
"oftheFAUSTcollection4withguaranteedcycleconsistency",
"shapes. Both these characteristics are achieved by reduc-",
"(white arrows). Here, we visualise the matchings via texture",
"ing the N-shape case to a sequence of three-shape match-",
"transfer between all shapes.",
"Ourmethod scaleslinearlyinthe",
"ings, the derivation of which is our main technical contribu-",
"",
"numberof shapes.Seethefullfigureinthesupplement.",
"tion. Thanks to quantum annealing, high-quality solutions",
"",
"with low energy are retrieved for the intermediate NP.",
"6",
"hard objectives. On benchmark datasets, the proposed ap-",
"two shapes. Accurate shape matching is a core element of",
"proach significantly outperforms extensions to multi-shape",
"0",
"many computervision andgraphics applications (i.e.,tex-",
"matching of a previous quantum-hybrid two-shape match-",
"ture transfer and statistical shape modelling). If non-rigid",
"ing method and is on-par with classical multi-matching",
"deformations are allowed, even pairwise matching is NP",
"methods. Our source code is available at 4dqv.mpi-",
"hard, leading to a wide area of research that approximates",
"inf.mpg.de/cCuantuMM/.",
"this problem, as a recent survey shows [18]. Matching two",
"shapes is one of the problems that was shown to benefit",
"from quantumhardware:Q-Match[42]iterativelyupdates",
"",
"1.Introduction",
"a subset of point correspondences using QA. Specifically,",
"Recently, there has been a growing interest in applying",
"its cyclic Q-expansion allows to parametrise changes to per-",
"mutation matrices without relaxations.",
"quantum computers in computer vision [3,23, 35]. Such",
"quantum computer vision methods rely on quantum anneal-",
"The question we ask in this work is: How can we design",
"ing (QA) that allows to solve NP-hard quadratic uncon-",
"a multi-shape matching algorithm in the style of Q-Match",
"strained binary optimisation problems (QUBOs).While",
"that has the same benefits? As we show in the experiments,",
"having to formulate a problem as a QUBO is rather inflexi-",
"whereweintroduceseveralnaivemulti-shapeextensionsof",
"ble,QAis,inthefuture,widelyexpectedtosolveQUBOsat",
"Q-Match, this is a highly non-trivial task. Despite tweaking",
"speeds not achievable with classical hardware. Thus, cast-",
"them, our proposed method significantly outperforms them.",
"ing a problem as a QUBO promises to outperform more un-",
"If N>2 shapes have to be matched, the computational",
"restricted formulations in terms of tractable problem sizes",
"complexity of naive exhaustive pairwise matching increases",
"and attainable accuracy through sheer speed.",
"quadratically with N, which does not scale to large N. Fur-",
"A recent example for such a problem is shape match-",
"thermore, these pairwise matchings can easily turn out to be",
"ing, where the goal is to estimate correspondences between",
"inconsistent with each other, thereby violating cycle con-"
] |
{
"arxiv_id": "2303.16202",
"language": "en",
"page_idx": 0
}
|
||
2303.16202-en-1
|
2303.16202-en-1.jpg
|
<FigureHere>
Figure 2. We match $N$ shapes by iteratively matching triplets.
sistency. For example, chaining the matchings $P_{\chi\gamma}$ from shape $\chi$ to $\gamma$ and $P_{yz}$ from $\gamma$ to $z$ can give very different correspondences between $\chi$ and $\mathcal{Z}$ than the direct, pairwise matching $P_{\chi z}$ of $\mathcal{X}$ and $\mathcal{Z}:P_{\chi z}\neq P_{\chi y}P_{\gamma z}.$ (We apply the permutation matrix $P_\mathrm{\chi\gamma}$ to the one-hot vertex index vector $x\in\mathcal{X}$ as $x^\top P_{\chi y}=y\in\mathcal{Y}.)$ Thus, how can we achieve cycle consistency by design? A simple solution would be to match a few pairs in the collection to create a spanning tree covering all shapes and infer the remaining correspondences by chaining along the tree. Despite a high accuracy of methods for matching two shapes, this correspondence aggregation policy is prone to error accumulation [39]. A special case of this policy is pairwise matching against a single anchor shape, which also guarantees cycleconsistent solutions by construction [22]. We build on this last option in our method as it avoids error accumulation.
This paper, in contrast to purely classical methods, leverages the advantages of quantum computing for multi-shape matching and introduces a new method for simultaneous alignment of multiple meshes with guaranteed cycle consistency; see Fig. l. It makes a significant step forward compared to Q-Match and other methods utilising adiabatic quantum computing (AQC), the basis for QA. Our cycleconsistent quantum-hybrid multi-shape matching (CCuantuMM; pronounced “quantum”) approach relies on the computational power of modern quantum hardware. Thus, our main challenge lies in casting our problem in QUBO form, which is necessary for compatibility with AQC. To that end, two design choices are crucial: (1) Our method reduces the $N$-shapes case to a series of three-shape matchings; see Fig. 2. Thus, CCuantuMM is iterative and hybrid, $i.e.$, it alternates in every iteration between preparing a QUBO problem on the CPU and sampling a QUBO solution on the AQC. (2) It discards negligible higher-order terms, which makes mapping the three-shape objective to quantum hardware possible. In summary, the core technical contributions of this paper are as follows:
$\bullet$ CCuantuMM, $i.e.$, a new quantum-hybrid method for
shape multi-matching relying on cyclic $\alpha$-expansion. CCuantuMM produces cycle-consistent matchings and
scales linearly with the number of shapes $N.$
· A new formulation of the optimisation objective for the
three-shapes case that is mappable to modern QA.
· A new policy in shape multi-matching to address the $N$-shape case relying on a three-shapes formulation and adaptive choice of an anchor shape.
Our experiments show that CCuantuMM significantly outperforms several variants of the previous quantumhybrid method Q-Match [42]. It is even competitive with several non-learning-based classical state-of-the-art shape methods [22,36] and can match more shapes than them. In a broader sense, this paper demonstrates the very high potential of applying (currently available and future) quantum hardware in computer vision.
## $2. \textbf{Related Work}$
Quantum Computer Vision (QCV). Several algorithms for computer vision relying on quantum hardware were proposed over the last three years for such problems as shape matching [23,35,42], object tracking [32,47], fundamental matrix estimation, point triangulation [20] and motion segmentation [1], among others. The majority of them address various types of alignment problems, $i.e.$, transformation estimation[23,35],point set[23,37] and mesh alignment[42], graph matching [3,41] and permutation synchronisation [3].
Only one of them, QSync[3], can operate on more than two inputs and ensure cycle consistency for the underlying matchings. In contrast to QSync, we can align inputs with substantially larger (by two orders of magnitude) shapes in the number of vertices. Furthermore, we address a different problem, $i.e.$, mesh alignment, for which an algorithm for two-mesh alignment with the help of AQC exists, namely Q-Match[42], as we discuss in the introduction.
To maintain the valid structure of permutation matrices, Quantum Graph Matching, QGM [41] and Q-Sync[3] impose linear constraints. However, this requires that the corresponding penalty parameter is carefully chosen. If the parameter is chosen too big and the linear constraints are enforced too strongly, this severely limits QGM and Q-Sync's ability to handle large sets of vertices. On the other hand, if the linear constraints are enforced too weakly, there is no guarantee to obtain valid permutations as solutions. As discussed in the introduction, our approach follows Q-Match to ensure valid permutation matrices by construction.
Multi-Shape Matching. We focus this section on multishape and non-learning methods as CCuantuMM falls in this category. As our approach is not learning-based, it trivially generalises to unknown object categories without a need for training data. For a general survey of recent advances in shape matching, see Sahillioglu [40].
Matching shape pairs is a classical problem in geometry processing [36]. When more than two shapes of the
2
|
[
[
0.53173828125,
0.091552734375,
0.826171875,
0.091552734375,
0.826171875,
0.10479736328125,
0.53173828125,
0.10479736328125
],
[
0.316162109375,
0.106689453125,
0.3203125,
0.106689453125,
0.3203125,
0.11236572265625,
0.316162109375,
0.11236572265625
],
[
0.408447265625,
0.10162353515625,
0.44775390625,
0.10479736328125,
0.44677734375,
0.117431640625,
0.406982421875,
0.1142578125
],
[
0.521484375,
0.10919189453125,
0.89208984375,
0.10919189453125,
0.89208984375,
0.12249755859375,
0.521484375,
0.12249755859375
],
[
0.53515625,
0.125,
0.86279296875,
0.125,
0.86279296875,
0.1383056640625,
0.53515625,
0.1383056640625
],
[
0.521484375,
0.1395263671875,
0.89208984375,
0.1395263671875,
0.89208984375,
0.15283203125,
0.521484375,
0.15283203125
],
[
0.53369140625,
0.154052734375,
0.89208984375,
0.1546630859375,
0.89208984375,
0.169189453125,
0.53369140625,
0.1685791015625
],
[
0.53515625,
0.17041015625,
0.79833984375,
0.17041015625,
0.79833984375,
0.184326171875,
0.53515625,
0.184326171875
],
[
0.131591796875,
0.186279296875,
0.20751953125,
0.1849365234375,
0.20751953125,
0.1962890625,
0.131591796875,
0.1976318359375
],
[
0.2034912109375,
0.1875,
0.2279052734375,
0.189453125,
0.226318359375,
0.1962890625,
0.2017822265625,
0.1944580078125
],
[
0.35546875,
0.199462890625,
0.398681640625,
0.199462890625,
0.398681640625,
0.21533203125,
0.35546875,
0.21533203125
],
[
0.521484375,
0.1962890625,
0.890625,
0.1962890625,
0.890625,
0.210205078125,
0.521484375,
0.210205078125
],
[
0.50341796875,
0.212158203125,
0.8896484375,
0.212158203125,
0.8896484375,
0.2259521484375,
0.50341796875,
0.2259521484375
],
[
0.293212890625,
0.2259521484375,
0.455078125,
0.2266845703125,
0.455078125,
0.2412109375,
0.293212890625,
0.240478515625
],
[
0.50244140625,
0.2266845703125,
0.89208984375,
0.2266845703125,
0.89208984375,
0.240478515625,
0.50244140625,
0.240478515625
],
[
0.50341796875,
0.2412109375,
0.89208984375,
0.2412109375,
0.89208984375,
0.255126953125,
0.50341796875,
0.255126953125
],
[
0.0906982421875,
0.255126953125,
0.455810546875,
0.255126953125,
0.455810546875,
0.268310546875,
0.0906982421875,
0.268310546875
],
[
0.50341796875,
0.25830078125,
0.890625,
0.25830078125,
0.890625,
0.268310546875,
0.50341796875,
0.268310546875
],
[
0.50244140625,
0.272705078125,
0.89208984375,
0.272705078125,
0.89208984375,
0.285888671875,
0.50244140625,
0.285888671875
],
[
0.50341796875,
0.287353515625,
0.89208984375,
0.287353515625,
0.89208984375,
0.300537109375,
0.50341796875,
0.300537109375
],
[
0.080078125,
0.29345703125,
0.468017578125,
0.29345703125,
0.468017578125,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.50244140625,
0.3017578125,
0.69677734375,
0.3017578125,
0.69677734375,
0.31494140625,
0.50244140625,
0.31494140625
],
[
0.080078125,
0.30810546875,
0.46728515625,
0.30810546875,
0.46728515625,
0.3212890625,
0.080078125,
0.3212890625
],
[
0.080078125,
0.323974609375,
0.468017578125,
0.323974609375,
0.468017578125,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.50341796875,
0.330810546875,
0.640625,
0.330810546875,
0.640625,
0.34521484375,
0.50341796875,
0.34521484375
],
[
0.080078125,
0.339111328125,
0.4697265625,
0.339111328125,
0.4697265625,
0.352783203125,
0.080078125,
0.352783203125
],
[
0.080078125,
0.354248046875,
0.468017578125,
0.354248046875,
0.468017578125,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.50244140625,
0.354736328125,
0.890625,
0.354736328125,
0.890625,
0.369384765625,
0.50244140625,
0.369384765625
],
[
0.080078125,
0.369384765625,
0.468017578125,
0.369384765625,
0.468017578125,
0.38330078125,
0.080078125,
0.38330078125
],
[
0.5009765625,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.385009765625,
0.5009765625,
0.385009765625
],
[
0.080078125,
0.385009765625,
0.468017578125,
0.385009765625,
0.468017578125,
0.3984375,
0.080078125,
0.3984375
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.399658203125,
0.4697265625,
0.399658203125,
0.4697265625,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.50341796875,
0.404052734375,
0.890625,
0.404052734375,
0.890625,
0.4140625,
0.50341796875,
0.4140625
],
[
0.080078125,
0.415283203125,
0.468017578125,
0.415283203125,
0.468017578125,
0.4287109375,
0.080078125,
0.4287109375
],
[
0.5009765625,
0.416748046875,
0.88818359375,
0.417236328125,
0.88818359375,
0.431884765625,
0.5009765625,
0.43115234375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.43310546875,
0.89208984375,
0.43310546875,
0.89208984375,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.08087158203125,
0.4462890625,
0.46728515625,
0.4462890625,
0.46728515625,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.50341796875,
0.449462890625,
0.88818359375,
0.449462890625,
0.88818359375,
0.460205078125,
0.50341796875,
0.460205078125
],
[
0.080078125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.50244140625,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.4765625,
0.50244140625,
0.4765625
],
[
0.0767822265625,
0.473388671875,
0.4697265625,
0.474853515625,
0.4697265625,
0.4892578125,
0.0767822265625,
0.488037109375
],
[
0.50244140625,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.080078125,
0.490478515625,
0.468017578125,
0.490478515625,
0.468017578125,
0.50390625,
0.080078125,
0.50390625
],
[
0.521484375,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.521484375,
0.5068359375
],
[
0.07843017578125,
0.5048828125,
0.4697265625,
0.50390625,
0.4697265625,
0.51806640625,
0.07843017578125,
0.51953125
],
[
0.50244140625,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50244140625,
0.521484375
],
[
0.080078125,
0.52099609375,
0.45751953125,
0.52099609375,
0.45751953125,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.101318359375,
0.537109375,
0.465576171875,
0.537109375,
0.465576171875,
0.54736328125,
0.101318359375,
0.54736328125
],
[
0.50341796875,
0.5390625,
0.8935546875,
0.5390625,
0.8935546875,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.080078125,
0.564453125
],
[
0.50244140625,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.50244140625,
0.56640625
],
[
0.08087158203125,
0.5673828125,
0.468017578125,
0.5673828125,
0.468017578125,
0.57763671875,
0.08087158203125,
0.57763671875
],
[
0.50341796875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.08087158203125,
0.58349609375,
0.465576171875,
0.58349609375,
0.465576171875,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.080078125,
0.609375
],
[
0.50341796875,
0.5986328125,
0.8203125,
0.5986328125,
0.8203125,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.52294921875,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.52294921875,
0.62744140625
],
[
0.07843017578125,
0.62548828125,
0.468017578125,
0.625,
0.468017578125,
0.6396484375,
0.07843017578125,
0.64013671875
],
[
0.50341796875,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50341796875,
0.64306640625
],
[
0.080078125,
0.64013671875,
0.46728515625,
0.64013671875,
0.46728515625,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64599609375,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.6591796875
],
[
0.080078125,
0.6572265625,
0.4697265625,
0.6572265625,
0.4697265625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.5009765625,
0.6591796875,
0.8896484375,
0.6591796875,
0.8896484375,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.080078125,
0.671875,
0.4697265625,
0.671875,
0.4697265625,
0.68505859375,
0.080078125,
0.68505859375
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.70068359375,
0.080078125,
0.70068359375
],
[
0.5009765625,
0.689453125,
0.89208984375,
0.68798828125,
0.89208984375,
0.70263671875,
0.5009765625,
0.7041015625
],
[
0.080078125,
0.70068359375,
0.46728515625,
0.70068359375,
0.46728515625,
0.7138671875,
0.080078125,
0.7138671875
],
[
0.50341796875,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.080078125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.50341796875,
0.7216796875,
0.890625,
0.7216796875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.73193359375,
0.468017578125,
0.73193359375,
0.468017578125,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50244140625,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.07843017578125,
0.74560546875,
0.468017578125,
0.7470703125,
0.468017578125,
0.76123046875,
0.07843017578125,
0.76025390625
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.76123046875,
0.468017578125,
0.76220703125,
0.468017578125,
0.77587890625,
0.080078125,
0.7744140625
],
[
0.50244140625,
0.7666015625,
0.8486328125,
0.7666015625,
0.8486328125,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.080078125,
0.7763671875,
0.468017578125,
0.7763671875,
0.468017578125,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.50244140625,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.07843017578125,
0.79052734375,
0.4697265625,
0.79150390625,
0.4697265625,
0.80615234375,
0.07843017578125,
0.8046875
],
[
0.50341796875,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.080078125,
0.806640625,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.080078125,
0.82080078125
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.82275390625,
0.4697265625,
0.82275390625,
0.4697265625,
0.8359375,
0.080078125,
0.8359375
],
[
0.5009765625,
0.82470703125,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.83837890625,
0.35546875,
0.83837890625,
0.35546875,
0.8515625,
0.080078125,
0.8515625
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.09722900390625,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.09722900390625,
0.869140625
],
[
0.50244140625,
0.85595703125,
0.81396484375,
0.85595703125,
0.81396484375,
0.869140625,
0.50244140625,
0.869140625
],
[
0.111083984375,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.111083984375,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.111083984375,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.111083984375,
0.8994140625
],
[
0.50244140625,
0.88623046875,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.491943359375,
0.923828125,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"scales linearly with the number of shapes N",
"",
"Pzixi",
": A new formulation of the optimisation objective for the",
"three-shapes case that is mappable to modern QA.",
". A new policy in shape multi-matching to address the",
"N-shape case relying on a three-shapes formulation",
"and adaptive choice of an anchor shape.",
"Anchori",
"",
"Pxiy",
"Our experiments show that CCuantuMM significantly",
"outperforms several variants of the previous quantum-",
"s.t.Pxizi=PxiyiPyiz",
"hybrid method Q-Match [42]. It is even competitive with",
"several non-learning-based classical state-of-the-art shape",
"Figure 2. We match N shapes by iteratively matching triplets.",
"methods[22,36]and canmatchmoreshapesthanthem.In",
"a broader sense, this paper demonstrates the very high po-",
"tential of applying (currently available and future) quantum",
"sistency. For example, chaining the matchings Pxy from",
"hardware in computer vision.",
"shape X to y and Pyz from y to z can give very differ-",
"ent correspondences between X and z than the direct, pair-",
"2.Related Work",
"wise matching Pxz of X and Z:Pxz + PxyPyz.(We",
"apply the permutation matrix Pxy to the one-hot vertex in-",
"Quantum Computer Vision (OCV). Several algorithms",
"dex vector E X as TPxy = y E y.) Thus, how can",
"for computer vision relying on quantum hardware were pro-",
"we achieve cycle consistency by design? A simple solution",
"posed over the last three years for such problems as shape",
"would be to match a few pairs in the collection to create a",
"matching[23.35.421,objecttracking[32.471.fundamental",
"spanning tree covering all shapes and infer the remaining",
"matrix estimation, point triangulation [20] and motion seg.",
"correspondences by chaining along the tree. Despite a high",
"mentation [1l, among others. The majority of them address",
"accuracy ofmethodsformatchingtwo shapes,this corre-",
"varioustypesof alignmentproblems,i.e.,transformationes",
"spondence aggregation policy is prone to error accumula-",
"timation [23,35], point set [23,37] and mesh alignment [42],",
"tion [39]. A special case of this policy is pairwise matching",
"graph matching [3,41] and permutation synchronisation [3].",
"against a single anchor shape, which also guarantees cycle-",
"Only one of them, QSync [3], can operate on more than",
"consistent solutions by construction [22]. We build on this",
"two inputs and ensure cycle consistency for the underlying",
"last option in our method as it avoids error accumulation.",
"matchings. In contrast to QSync, we can align inputs with",
"Thispaper,incontrasttopurelyclassicalmethods.lever",
"substantially larger (by two orders of magnitude) shapes in",
"ages the advantages of quantum computing for multi-shape",
"the number of vertices. Furthermore, we address a different",
"matching and introducesanewmethodforsimultaneous",
"problem, i.e., mesh alignment, for which an algorithm for",
"alignmentof multiplemesheswithguaranteed cyclecon-",
"two-mesh alignment with the help of AQC exists, namely",
"sistency; see Fig. I. It makes a significant step forward",
"Q-Match [42], as we discuss in the introduction.",
"compared to Q-Match and other methods utilising adiabatic",
"To maintain the valid structure of permutation matrices,",
"quantum computing (AQC), the basis for QA. Our cycle-",
"Quantum Graph Matching, QGM [41] and Q-Sync [3] im-",
"consistent quantum-hybrid multi-shape matching (CCuan-",
"pose linear constraints. However, this requires that the cor-",
"tuMM; pronounced “quantum\") approach relies on the",
"responding penalty parameter is carefully chosen. If the pa-",
"computational power of modern quantum hardware. Thus,",
"rameter is chosen too big and the linear constraints are en-",
"our main challenge lies in casting our problem in QUBO",
"forced too strongly, this severely limits QGM and Q-Sync's",
"form, which is necessary for compatibility with AQC. To",
"ability to handle large sets of vertices. On the other hand,",
"that end, two design choices are crucial: () Our method",
"if thelinearconstraints are enforced tooweakly,there is no",
"reduces the N-shapes case to a series of three-shape match-",
"guarantee to obtain valid permutations as solutions. As dis-",
"ings; see Fig. 2. Thus, CCuantuMM is iterative and hy-",
"cussed in the introduction, our approach follows Q-Match",
"brid, i.e., it alternates in every iteration between preparing",
"to ensure valid permutationmatricesby construction",
"a QUBO problem on the CPU and sampling a QUBO so-",
"Multi-Shape Matching. We focus this section on multi-",
"lution on the AQC. (2) It discards negligible higher-order",
"shape and non-learning methods as CCuantuMM falls in",
"terms, which makes mapping the three-shape objective to",
"this category. As our approach is not learning-based, it",
"quantum hardware possible. In summary, the core technical",
"trivially generalises to unknown object categories without",
"contributions of this paper are as follows:",
"a need for training data. For a general survey of recent ad-",
"· CCuantuMM, i.e., a new quantum-hybrid method for",
"vances in shape matching, see Sahillioglu [40].",
"shape multi-matching relying on cyclic α-expansion.",
"Matching shape pairs is a classical problem in geome-",
"CCuantuMM produces cycle-consistent matchings and",
"try processing [36]. When more than two shapes of the",
"2"
] |
{
"arxiv_id": "2303.16202",
"language": "en",
"page_idx": 1
}
|
||
2303.16202-en-2
|
2303.16202-en-2.jpg
|
same class exist, stronger geometric cues can be leveraged to improve results by matching all of them simultaneously. Unfortunately, the already very high problem complexity increases even further the more shapes are used. Hence, existing multi-shape matching methods limit the total number of shapes and their resolution [12,22],work in spectral space[26], or relax the permutation constraints[28]. Early multi-matching methods computed pair-wise matchings and subsequently used permutation synchronisation to establish cycle consistency [33, 38, 43]. Still, permutation synchronisation requires the eigendecomposition of a matrix with quadratically increasing dimensions [38].
HiPPI[2] is a computationally efficient method that takes geometric relations into account while generalising permutation synchronization but is still limited in resolution. Instead of looking at permutations directly, ZoomOut[36] reduces the dimensionality of the problem by projecting it onto the spectral decomposition. This idea has been extended to take cycle consistency within the spectral space into account [25], which does not guarantee a point-wise consistent matching. To circumvent this issue, IsoMuSh [22] jointly optimises point and functional correspondences. The method detangles the optimisation into smaller subproblems by using a so-called universe shape that all shapes are mapped to instead of each other, as Cao and Bernard do[10]. Using a universe is similar to requiring a template shape, as many learning-based approaches do [19,24,44]: Both synchronise all correspondences by matching them through a unifed space. This is similar to the concept of anchor shape we use but inherently less flexible because the universe size or template have to be given $a\textit{priori. Our anchor is chosen from the given collection}$ as part of the method. Although using an anchor slightly improves our results, we note that our method does not necessarily require one for operation. Hence, a random shape could be picked instead in each iteration without an increase in complexity if using an anchor is not feasible or does not represent the shape collection well.
# $3. \textbf{Background}$
3.1. Adiabatic Quantum Computing (AQC)
AQC is a model of computation that leverages quantum effects to obtain high-quality solutions to the $\mathcal{N}\mathcal{P}$-hard problem class of Quadratic Unconstrained Binary Optimi- ${\mathrm{sation~( QUBO) ~problems: ~min}}_{x\in \{ 0, 1\} ^k}x^{T}Qx, {\mathrm{~for~}}\hat{k} \in \mathbb{N}$ and a QUBO weight matrix $Q\in\mathbb{R}^{k\times k}.$ Each entry of $x$ corresponds to its own logical qubit, the quantum equivalent of a classical bit. The diagonal of $Q$ consists of linear terms, while the off-diagonals are inter-qubit coupling weights. A QUBO can be classically tackled with simulated annealing (SA)[45] or a variety of other discrete optimisation techniques [21,31], which, for large $k$, typically
yield only approximate solutions as QUBOs are in general $\mathcal{N}\mathcal{P}$-hard. AQC holds the potential to systematically outperform classical approaches such as SA, see[17,30] for an example. AQC exploits the adiabatic theorem of quantum mechanics [5]: If, when starting from an equal superposition state of the qubits (where all solutions $\{\hat0,1\}^k$ have the same probability of being measured) and imposing external influences corresponding to the QUBO matrix on the qubits sufficiently slowly (called annealing), they will end up in a quantum state that, when measured, yields a minimizer $x$ of the QUBO. Not all physical qubits on a real quantum processing unit (QPU) can be connected (coupled) with each other. Thus, a minor embedding of the logical-qubit graph (defined by non-zero entries of the QUBO matrix) into the physical-qubit graph (defined by the hardware) is required [9]. This can lead to a chain of multiple physical qubits representing a single logical qubit. For details of quantum annealing on D-Wave machines, we recommend[34].
# 3.2. Shape Matching
The problem of finding a matching for non-rigidly deformed shapes having $n$ vertices can be formulated as an $\mathcal{N}\mathcal{P}$-hard Quadratic Assignment Problem (QAP)[8,28]:
(1)
$$\min_{P\in\mathbb{P}_{n}}p^{T}Wp,$$
where $p$ = vec$( P)$ $\in \:_{2}\{ 0, 1\} ^{n^2}$ is a flattened permutation matrix, and $W\in\mathbb{R}^{n^2\times n^2}$ is an energy matrix describing how well certain pairwise properties are conserved between two pairs of matches. If two shapes $\mathcal{X},\mathcal{Y}$ are discretised with $n$ vertices each, $W$ is often chosen as [28]:
(2)
$$W_{x_{1}\cdot n+y_{1},x_{2}\cdot n+y_{2}}=\|d_{\mathcal X}^{g}(x_{1},x_{2})-d_{\mathcal Y}^{g}(y_{1},y_{2})\|,$$
where $x_1,x_2$ are vertices on $\mathcal{X};y_1,y_2$ are vertices on $\mathcal{Y};$and $d_{\mathcal{I}}^{g}(\cdot,\cdot)$ is the geodesic distance on the shape $\mathcal{I}.$ Therefore, $W_{x_1\cdot n+y_1,x_2\cdot n+y_2}$ represents how well the geodesic distance is preserved between corresponding pairs of vertices on the two shapes. Instead of pure geodesics, Gaussianfiltered geodesics are also a popular choice for $W$ [46]:
(3
$$g_{\mathcal{X}}(x_1,x_2)=\frac{1}{\rho\sqrt{2\pi}}\exp\left(-\frac{1}{2}\left(\frac{d_{\mathcal{X}}^g(x_1,x_2)}{\rho}\right)^2\right).$$
$g_{\mathcal{I}}$ can be used to directly replace $d_\mathcal{I}^g$ in (2). A small value of $\rho$ focuses the energy on a local neighbourhood around the vertex, while a large value increases the receptive field. Using Gaussian kernels in $W$ places more emphasis on local geometry whereas geodesics have higher values far away from the source vertex. Thus, geodesics work well for global alignment and Gaussians for local fine-tuning.
3
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50341796875,
0.10919189453125,
0.8896484375,
0.10919189453125,
0.8896484375,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.08087158203125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5009765625,
0.136962890625,
0.890625,
0.136962890625,
0.890625,
0.15087890625,
0.5009765625,
0.15087890625
],
[
0.08087158203125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.08087158203125,
0.1666259765625
],
[
0.50341796875,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.07843017578125,
0.1673583984375,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.465576171875,
0.1817626953125,
0.465576171875,
0.1956787109375,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.5009765625,
0.2266845703125
],
[
0.08087158203125,
0.2298583984375,
0.46728515625,
0.2298583984375,
0.46728515625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.5048828125,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.25830078125,
0.3505859375,
0.25830078125,
0.3505859375,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.098876953125,
0.274658203125,
0.4697265625,
0.274658203125,
0.4697265625,
0.28857421875,
0.098876953125,
0.28857421875
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.07843017578125,
0.2890625,
0.468017578125,
0.28857421875,
0.468017578125,
0.302978515625,
0.07843017578125,
0.3037109375
],
[
0.50341796875,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50341796875,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.5048828125,
0.32080078125,
0.8896484375,
0.32080078125,
0.8896484375,
0.330810546875,
0.5048828125,
0.330810546875
],
[
0.080078125,
0.335205078125,
0.4697265625,
0.335205078125,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.5009765625,
0.333251953125,
0.890625,
0.333251953125,
0.890625,
0.3466796875,
0.5009765625,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.4697265625,
0.348388671875,
0.4697265625,
0.363037109375,
0.080078125,
0.3642578125
],
[
0.50244140625,
0.348388671875,
0.8544921875,
0.348388671875,
0.8544921875,
0.3623046875,
0.50244140625,
0.3623046875
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50244140625,
0.37255859375,
0.66259765625,
0.375,
0.66259765625,
0.389404296875,
0.50244140625,
0.386962890625
],
[
0.080078125,
0.380126953125,
0.471435546875,
0.380126953125,
0.471435546875,
0.39453125,
0.080078125,
0.39453125
],
[
0.08087158203125,
0.395751953125,
0.471435546875,
0.395751953125,
0.471435546875,
0.4091796875,
0.08087158203125,
0.4091796875
],
[
0.51953125,
0.396484375,
0.890625,
0.397705078125,
0.890625,
0.412353515625,
0.51953125,
0.410888671875
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.41357421875,
0.89208984375,
0.41357421875,
0.89208984375,
0.4267578125,
0.50244140625,
0.4267578125
],
[
0.07843017578125,
0.425537109375,
0.4697265625,
0.4248046875,
0.4697265625,
0.439453125,
0.07843017578125,
0.43994140625
],
[
0.5009765625,
0.4267578125,
0.89208984375,
0.427978515625,
0.89208984375,
0.442626953125,
0.5009765625,
0.44140625
],
[
0.08251953125,
0.442626953125,
0.46728515625,
0.442626953125,
0.46728515625,
0.45263671875,
0.08251953125,
0.45263671875
],
[
0.080078125,
0.455810546875,
0.4697265625,
0.455810546875,
0.4697265625,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.080078125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.65673828125,
0.47021484375,
0.73876953125,
0.468505859375,
0.73876953125,
0.484130859375,
0.65673828125,
0.486083984375
],
[
0.86767578125,
0.47021484375,
0.89453125,
0.47021484375,
0.89453125,
0.4873046875,
0.86767578125,
0.4873046875
],
[
0.080078125,
0.486083984375,
0.4697265625,
0.486083984375,
0.4697265625,
0.5,
0.080078125,
0.5
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.5009765625,
0.5068359375,
0.89208984375,
0.505859375,
0.89208984375,
0.5234375,
0.5009765625,
0.5244140625
],
[
0.080078125,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.52978515625,
0.080078125,
0.52978515625
],
[
0.5,
0.52099609375,
0.8935546875,
0.521484375,
0.8935546875,
0.54345703125,
0.5,
0.54248046875
],
[
0.08087158203125,
0.53271484375,
0.46728515625,
0.53271484375,
0.46728515625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.080078125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50341796875,
0.54345703125,
0.8896484375,
0.54345703125,
0.8896484375,
0.5537109375,
0.50341796875,
0.5537109375
],
[
0.50244140625,
0.55615234375,
0.89208984375,
0.55615234375,
0.89208984375,
0.5693359375,
0.50244140625,
0.5693359375
],
[
0.080078125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.50244140625,
0.57080078125,
0.81787109375,
0.57080078125,
0.81787109375,
0.583984375,
0.50244140625,
0.583984375
],
[
0.080078125,
0.5771484375,
0.46728515625,
0.5771484375,
0.46728515625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.08087158203125,
0.59326171875,
0.465576171875,
0.59326171875,
0.465576171875,
0.60400390625,
0.08087158203125,
0.60400390625
],
[
0.52294921875,
0.59765625,
0.89208984375,
0.5986328125,
0.89208984375,
0.6162109375,
0.52294921875,
0.61572265625
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50341796875,
0.626953125,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.50341796875,
0.6416015625
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.5048828125,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.65771484375,
0.5048828125,
0.65771484375
],
[
0.07843017578125,
0.6513671875,
0.31201171875,
0.65087890625,
0.31201171875,
0.66552734375,
0.07843017578125,
0.66650390625
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.67529296875,
0.50244140625,
0.67529296875
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.080078125,
0.6845703125,
0.205078125,
0.6845703125,
0.205078125,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.08087158203125,
0.708984375,
0.410888671875,
0.708984375,
0.410888671875,
0.7236328125,
0.08087158203125,
0.7236328125
],
[
0.50341796875,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.1004638671875,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.1004638671875,
0.74755859375
],
[
0.07843017578125,
0.74853515625,
0.468017578125,
0.74755859375,
0.468017578125,
0.76220703125,
0.07843017578125,
0.76318359375
],
[
0.73779296875,
0.74853515625,
0.814453125,
0.74853515625,
0.814453125,
0.76318359375,
0.73779296875,
0.76318359375
],
[
0.5185546875,
0.7568359375,
0.6044921875,
0.75390625,
0.60546875,
0.76806640625,
0.51953125,
0.771484375
],
[
0.080078125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.61376953125,
0.7646484375,
0.6611328125,
0.7646484375,
0.6611328125,
0.77880859375,
0.61376953125,
0.77880859375
],
[
0.07843017578125,
0.77880859375,
0.471435546875,
0.77880859375,
0.471435546875,
0.796875,
0.07843017578125,
0.796875
],
[
0.07843017578125,
0.79345703125,
0.471435546875,
0.79345703125,
0.471435546875,
0.81103515625,
0.07843017578125,
0.81103515625
],
[
0.5,
0.79443359375,
0.89208984375,
0.79248046875,
0.89208984375,
0.81005859375,
0.5,
0.8125
],
[
0.08087158203125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.46728515625,
0.85595703125,
0.46728515625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8505859375,
0.88720703125,
0.8505859375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"same class exist, stronger geometric cues can be leveraged",
"yield only approximate solutions as QUBOs are in general",
"to improve results by matching all of them simultaneously.",
"NP-hard.AOCholdsthepotentialtosystematicallyout-",
"Unfortunately, the already very high problem complexity",
"perform classical approaches such as SA, see [17,30] for an",
"increases even further the more shapes are used. Hence,",
"example. AQC exploits the adiabatic theorem of quantum",
"existing multi-shape matching methods limit the total num-",
"mechanics[5l:If,when starting from an equal superposi-",
"ber of shapes and their resolution [12, 22], work in spectral",
"tion state of the qubits (where all solutions fO, 1)k have the",
"space[26],orrelax thepermutation constraints[28].Early",
"same probability of being measured) and imposing external",
"multi-matchingmethodscomputedpair-wisematchingsand",
"influences corresponding to the QUBO matrix on the qubits",
"subsequently used permutation synchronisation to establish",
"sufficiently slowly (called annealing), they will end up in a",
"cycleconsistency[33,38,43].Still,permutation synchro-",
"quantum statethat,when measured,yields aminimizer of",
"nisation requires the eigendecomposition of a matrix with",
"the QUBO. Not all physical qubits on a real quantum pro-",
"quadratically increasing dimensions [38].",
"cessing unit (QPU) can be connected (coupled) with each",
"HiPPI [2] is a computationally efficient method that",
"other. Thus, a minor embedding of the logical-qubit graph",
"takes geometric relations into account while generalis-",
"(defined by non-zero entries of the QUBO matrix) into the",
"ing permutation synchronization but is still limited in",
"physical-qubit graph (defined by the hardware) is required",
"resolution.Instead of looking at permutations directly,",
"[9l.This canleadto a chain of multiplephysical gubits",
"ZoomOut [36] reduces the dimensionality of the problem",
"representing a single logical qubit. For details of quantum",
"by projecting it onto the spectral decomposition. This idea",
"annealing on D-Wave machines, we recommend [34].",
"has been extended to take cycle consistency within the spec-",
"3.2.ShapeMatching",
"tral space into account [25], which does not guarantee a",
"point-wise consistent matching. To circumvent this issue,",
"The problem of finding a matching for non-rigidly de-",
"IsoMuSh [22 jointly optimises point and functional corre-",
"formed shapes having n vertices can be formulated as an",
"spondences. The method detangles the optimisation into",
"NP-hard Quadratic Assignment Problem (QAP) [8, 28]:",
"smaller subproblems by using a so-called universeshape",
"that all shapes are mapped to instead of each other, as Cao",
"and Bernard do [10]. Using a universe is similar to requir-",
"min pTWp",
"(1)",
"ing a template shape, as many learning-based approaches",
"do [19, 24, 44]: Both synchronise all correspondences by",
"where p = vec(P) e [0,1)n’ is a flattened permutation",
"matching them through a unified space. This is similar to",
"matrix, and W e Rn'xn is an energy matrix describing",
"theconceptofanchorshapeweusebut inherentlylessflex-",
"ible because the universe size or template have to be given",
"howwellcertainpairwisepropertiesareconservedbetween",
"two pairs of matches. If two shapes X, are discretised",
"a priori. Our anchor is chosen from the given collection",
"with n vertices each, W is often chosen as [28]:",
"as part of the method. Although using an anchor slightly",
"improves ourresults,wenotethatourmethod doesnotnec-",
"W-n+yi,±2-n+u2 = Ild%(a1,2) -dg(y1,y2)ll, (2)",
"essarily require one for operation. Hence, a random shape",
"could be picked instead in each iteration without an increase",
"where 1, 2 are vertices on X; y1,y2 are vertices on ; and",
"in complexity if using an anchor is not feasible or does not",
"d(, ) is the geodesic distance on the shape I. Therefore,",
"represent the shape collection well.",
"Wr1-n+y1,f2-n+y2 represents how well the geodesic dis-",
"tanceispreservedbetween correspondingpairsof vertices",
"3.Background",
"on the two shapes. Instead of pure geodesics, Gaussian-",
"3.1. Adiabatic Quantum Computing (AQC)",
"filtered geodesics are also a popular choice for W [46]:",
"AQC is a model of computation that leverages quan-",
"tum effects to obtain high-quality solutions to the Np-hard",
"d(ai,22)",
"9x(r1,12) :",
"problem class of Quadratic Unconstrained Binary Optimi-",
"pV2元",
"sation (QUBO) problems: minre(0,1), rT Qa, for k N",
"and a QUBO weight matrix Q e IRkxk. Each entry of r",
"gr can be used to directly replace d in (2). A small value",
"corresponds to its own logical qubit, the quantum equiva-",
"of p focuses the energy on a local neighbourhood around the",
"lent of a classical bit. The diagonal of Q consists of lin-",
"vertex, while a large value increases the receptive field. Us-",
"ear terms, while the off-diagonals are inter-qubit coupling",
"ing Gaussian kernels in W places more emphasis on local",
"weights. A QUBO can be classically tackled with simu-",
"geometry whereas geodesics have higher values far away",
"lated annealing (SA) [45] or a variety of other discrete op-",
"from the source vertex. Thus, geodesics work well for",
"timisation techniques [21, 31], which, for large k, typically",
"global alignment and Gaussians for local fine-tuning.",
"3"
] |
{
"arxiv_id": "2303.16202",
"language": "en",
"page_idx": 2
}
|
||
2303.16202-en-3
|
2303.16202-en-3.jpg
|
### 3.3. Cyclic $\alpha$-Expansion (CAE)
CCuantuMM represents matchings as permutation matrices. In order to update them, we build on Seelbach $et$ $al$.'s CAE algorithm [42] (similar to a fusion move[27]), which we describe here. A permutation matrix $P$ is called an $r$-cycle, if there exist $r$ disjoint indices $i_1,\ldots,i_r$ such $\begin{array}{l}{\mathrm{that~}P_{i_{j}i_{(j+1)\%r}}=1\text{ for all }j\in\{1,\ldots,r\}\text{ and }P_{l,l}=1}\\{\mathrm{for~all~}l\notin\{i_{1},\ldots,i_{r}\},\text{ in which case }P=(i_{1}i_{2}\ldots i_{r})\text{ is}}\end{array}$ a common notation. Two cycles, $i.e.$ two permutation matrices, are disjoint if these indices are pairwise disjoint. We know that disjoint cycles commute, which allows us to represent any permutation $P$ as $P=\left(\prod_{i=1}^{k}c_{i}\right)\left(\prod_{i=1}^{l}\tilde{c}_{i}\right)$, where $\{c_i\}_i$ and $\{\tilde{c}_i\}_i$ each are sets of disjoint 2-cycles.
Given a set $\{c_i\}_{i=1}^k$ of $k$ disjoint 2-cycles, an $update$, or modification, of $P$ can therefore be parameterised as: $P( \alpha )$ = $\prod _{i= 1}^{k}c_{i}^{\alpha _{i}}P$,where $\alpha$ $\in$ $\{ 0, 1\} ^{k}$ is a binary decision vector determining the update. (Note that $\alpha_i$ in $c_i^{\alpha_i}$ is an exponent, not an index.) Crucially, to make this parameterisation compatible with QUBOs, we need to make it linear in $\alpha.$ To this end, CAE uses the following equality:
$$P(\alpha)=P+\sum_{i=1}^k\alpha_i(c_i-I)P.$$
(4)
### 4.Our CCuantuMM Method
Previous adiabatic quantum computing methods [41,42] can only match two shapes. We present a method for matching $N$ shapes. To ensure cycle consistency on $N$ shapes, it is sufficient that all triplets of shapes are matched cycle consistently [25]. CCuantuMM iteratively solves three-shape problems, which preserve cycle consistency by construction and fit on existing quantum annealers with limited resources. We introduce our formulation for matching three shapes in Sec. 4.1 and then extend it to $N$ shapes in Sec. 4.2.
## $4. 1. \textbf{Matching Three Shapes}$
Consider the problem of matching three non-rigidly deformed shapes $\bar{\mathcal{S}}=\{\mathcal{X},\mathcal{Y},\mathcal{Z}\}$ of $n$ vertices each, while preserving cycle consistency. We formulate this as an energy minimisation with respect to $\mathcal{P}=\{P_{\mathcal{I}\mathcal{J}}\in\mathbb{P}_n|\mathcal{I},\mathcal{J}\in$ $S\}, the set of permutations between all pairs in S:$
$$\begin{aligned}&\min_{\mathcal{P}}\sum_{\mathcal{I},\mathcal{J}\in\mathcal{S};\mathcal{I}\neq\mathcal{J}}\mathrm{vec}(P_{\mathcal{I}\mathcal{J}})^{\top}W_{\mathcal{I}\mathcal{J}}\mathrm{vec}(P_{\mathcal{I}\mathcal{J}}),\\&\mathrm{s.t.}\quad P_{\mathcal{X}\mathcal{Z}}=P_{\mathcal{X}\mathcal{Y}}P_{\mathcal{Y}\mathcal{Z}},\end{aligned}$$
(5)
where $W_{\mathcal{I}\mathcal{J}}\in\mathbb{R}^{n^2\times n^2}$ is the energy matrix describing how well certain pairwise properties are conserved between shapes $\mathcal{I}$ and $\mathcal{J}($see Sec. 3.2), and $P_{\chi z}=P_{\chi y}P_{\gamma z}$ enforces cyclic consistency. An overview of the algorithm for three shapes is shown in Alg. 1.
### 4.1.1 $QUBO$ Derivation
To perform optimisation on the quantum annealer, we need to transform $\hat{(5)}$ into a QUBO problem. We adapt the CAE formulation from [42] (see Sec. 3.3) and iteratively update the permutations to decrease the value of (5). Given a set $C$ = $\{ c_{i}\} _{i= 1}^{k}$ of $k$ disjoint 2-cycles and binary decision variables $\alpha$, we can parameterise our permutation matri- $\begin{array}{l}{\mathrm{ces~as~}P_{\mathcal{I}\mathcal{J}}(\alpha)=P_{\mathcal{I}\mathcal{J}}+\sum_{i=1}^{k}\alpha_{i}(c_{i}-I)P_{\mathcal{I}\mathcal{J}}.\mathrm{However},}\\{\mathrm{CAE~alone~is~not~sufficient~to~transform~(5)~into~a~QUBO}}\end{array}$ as cyclic consistency is still missing. A simple solution would be to encourage cyclic consistency as a quadratic soft penalty, but then there are no guarantees on the solution. Instead, we enforce cyclic consistency by construction:
$$P_{\mathcal{X}\mathcal{Z}}(\alpha,\beta)=(6)\\(P_{\mathcal{X}\mathcal{Y}}+\sum_{i=1}^{k}\alpha_{i}(c_{i}-I)P_{\mathcal{X}\mathcal{Y}})\cdot(P_{\mathcal{Y}\mathcal{Z}}+\sum_{j=1}^{k}\beta_{j}(\tilde{c}_{j}-I)P_{\mathcal{Y}\mathcal{Z}}),$$
$$\min_{\alpha,\beta\in\{0,1\}^k}E_{\mathcal{X}\mathcal{Y}}(P_{\mathcal{X}\mathcal{Y}}(\alpha))+E_{\mathcal{Y}\mathcal{Z}}(P_{\mathcal{Y}\mathcal{Z}}(\beta))+E_{\mathcal{X}\mathcal{Z}}(P_{\mathcal{X}\mathcal{Z}}(\alpha,\beta)),$$
(7)
where $\{c_{i}\}_{i}\left(\{\tilde{c}_{j}\}_{j}\right)$are cycles and $\alpha\left(\beta\right)$ are decision vari- ${\mathrm{ables~for~the~updates~to~}P_{\chi y}}\left(P_{\gamma z}\right).{\mathrm{For~brevity,we~write}}$ $C_{i}=(c_{i}-I)P_{\chi y}$ and $\tilde{C}_{j}=(\tilde{c}_{j}-I)P_{\gamma Z}.$ We explain how we construct the cycles in Secs. 4.1.2-4.1.3. Thus, we iteratively solve (5) via a sequence of problems of the form:
$\begin{array}{rcl}{\mathrm{where}}&{E_{\mathcal{I}\mathcal{J}}(P,Q)}&{=}&{\mathrm{vec}(P)^{\top}W_{\mathcal{I}\mathcal{J}}\operatorname{vec}(Q)}&{\mathrm{and}}\end{array}$ $E_{\mathcal{I} \mathcal{J} }( P)$ $= E_{\mathcal{I} \mathcal{J} }( P, P) .$While the frst two terms are in
QUBO form, the third term contains cubic and bi-quadratic terms (see the supplement for details) which are not compatible with current quantum annealer architectures
Higher-Order Terms. All of these higher-order terms $\begin{array}{l}\text{come from }P_{\chi{\mathcal{Z}}}(\alpha,\beta),\text{specifically from the term}H=\\\sum_{i}\sum_{j}\alpha_{i}\beta_{j}C_{i}\tilde{C}_{j}.\text{ As we only consider 2-cycles, }C_{i}\mathrm{~and}\end{array}$ $\tilde{C}_{j}$ each have only four non-zero elements. Due to this extreme sparsity, most summands of $H$ become 0.
We could tackle these undesirable terms by decomposing them into quadratic terms by using ancilla variables and adding penalty terms[16]. This gives exact solutions for sufficiently high weights of the penalty terms. However, multiple reasons speak against this: (1) the QUBO matrix is already dense (a clique) under the current formulation (as we will see in (9)) and adding ancilla qubits scales quadratically in $k,(2)$ adding penalties makes the problem harder to solve, and (3) $H$ is sparsely non-zero and in practise we observe no drastic influence on the quality of the solution.
Alternatively, we could assume $H= 0.$ However, this is unnecessarily strong since (1) $H$ also contributes to quadratic terms $(E(H,\cdot))$, and (2) higher-order terms operating on the same decision variable trivially reduce to
4
|
[
[
0.08087158203125,
0.09027099609375,
0.318603515625,
0.09027099609375,
0.318603515625,
0.10418701171875,
0.08087158203125,
0.10418701171875
],
[
0.50341796875,
0.0921630859375,
0.5390625,
0.0921630859375,
0.5390625,
0.10418701171875,
0.50341796875,
0.10418701171875
],
[
0.54833984375,
0.09027099609375,
0.6806640625,
0.091552734375,
0.6806640625,
0.1060791015625,
0.54833984375,
0.10479736328125
],
[
0.09722900390625,
0.11297607421875,
0.4697265625,
0.1142578125,
0.4697265625,
0.1287841796875,
0.09722900390625,
0.1275634765625
],
[
0.50341796875,
0.11614990234375,
0.890625,
0.11614990234375,
0.890625,
0.126953125,
0.50341796875,
0.126953125
],
[
0.080078125,
0.1300048828125,
0.471435546875,
0.1300048828125,
0.471435546875,
0.143310546875,
0.080078125,
0.143310546875
],
[
0.50244140625,
0.1300048828125,
0.8896484375,
0.1300048828125,
0.8896484375,
0.14013671875,
0.50244140625,
0.14013671875
],
[
0.080078125,
0.1458740234375,
0.46728515625,
0.1458740234375,
0.46728515625,
0.1590576171875,
0.080078125,
0.1590576171875
],
[
0.50244140625,
0.1427001953125,
0.89208984375,
0.1427001953125,
0.89208984375,
0.1558837890625,
0.50244140625,
0.1558837890625
],
[
0.080078125,
0.160400390625,
0.468017578125,
0.160400390625,
0.468017578125,
0.173583984375,
0.080078125,
0.173583984375
],
[
0.50244140625,
0.1558837890625,
0.89208984375,
0.1558837890625,
0.89208984375,
0.169189453125,
0.50244140625,
0.169189453125
],
[
0.080078125,
0.175537109375,
0.468017578125,
0.175537109375,
0.468017578125,
0.189453125,
0.080078125,
0.189453125
],
[
0.50244140625,
0.17041015625,
0.89208984375,
0.17041015625,
0.89208984375,
0.1849365234375,
0.50244140625,
0.1849365234375
],
[
0.080078125,
0.186279296875,
0.176513671875,
0.1944580078125,
0.1732177734375,
0.208984375,
0.0767822265625,
0.2008056640625
],
[
0.1854248046875,
0.1900634765625,
0.468017578125,
0.1900634765625,
0.468017578125,
0.203857421875,
0.1854248046875,
0.203857421875
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.07843017578125,
0.205810546875,
0.4697265625,
0.205810546875,
0.4697265625,
0.2197265625,
0.07843017578125,
0.2197265625
],
[
0.50341796875,
0.2008056640625,
0.8896484375,
0.2008056640625,
0.8896484375,
0.21533203125,
0.50341796875,
0.21533203125
],
[
0.50341796875,
0.2139892578125,
0.89208984375,
0.2139892578125,
0.89208984375,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.07843017578125,
0.2203369140625,
0.465576171875,
0.2203369140625,
0.465576171875,
0.2342529296875,
0.07843017578125,
0.2342529296875
],
[
0.080078125,
0.23486328125,
0.468017578125,
0.23486328125,
0.468017578125,
0.248779296875,
0.080078125,
0.248779296875
],
[
0.5048828125,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.50341796875,
0.242431640625,
0.89208984375,
0.242431640625,
0.89208984375,
0.255615234375,
0.50341796875,
0.255615234375
],
[
0.080078125,
0.250732421875,
0.464111328125,
0.250732421875,
0.464111328125,
0.264404296875,
0.080078125,
0.264404296875
],
[
0.50341796875,
0.255615234375,
0.890625,
0.255615234375,
0.890625,
0.26953125,
0.50341796875,
0.26953125
],
[
0.07843017578125,
0.267578125,
0.459228515625,
0.267578125,
0.459228515625,
0.285400390625,
0.07843017578125,
0.285400390625
],
[
0.50244140625,
0.270751953125,
0.85107421875,
0.270751953125,
0.85107421875,
0.2841796875,
0.50244140625,
0.2841796875
],
[
0.0767822265625,
0.285888671875,
0.443603515625,
0.285400390625,
0.443603515625,
0.302978515625,
0.0767822265625,
0.3037109375
],
[
0.51123046875,
0.294189453125,
0.59619140625,
0.294189453125,
0.59619140625,
0.309326171875,
0.51123046875,
0.309326171875
],
[
0.87109375,
0.29541015625,
0.8935546875,
0.29541015625,
0.8935546875,
0.30810546875,
0.87109375,
0.30810546875
],
[
0.09722900390625,
0.302978515625,
0.46728515625,
0.3017578125,
0.46728515625,
0.31640625,
0.09722900390625,
0.317626953125
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.58251953125,
0.324462890625,
0.74658203125,
0.324462890625,
0.74658203125,
0.33837890625,
0.58251953125,
0.33837890625
],
[
0.771484375,
0.324462890625,
0.88232421875,
0.324462890625,
0.88232421875,
0.33837890625,
0.771484375,
0.33837890625
],
[
0.08251953125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.348388671875,
0.08251953125,
0.348388671875
],
[
0.51318359375,
0.32568359375,
0.5556640625,
0.32568359375,
0.5556640625,
0.337158203125,
0.51318359375,
0.337158203125
],
[
0.56884765625,
0.34228515625,
0.5810546875,
0.339111328125,
0.58251953125,
0.34521484375,
0.57177734375,
0.347900390625
],
[
0.75341796875,
0.34228515625,
0.77294921875,
0.34228515625,
0.77294921875,
0.349853515625,
0.75341796875,
0.349853515625
],
[
0.080078125,
0.348388671875,
0.463134765625,
0.348388671875,
0.463134765625,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50341796875,
0.361083984375,
0.8896484375,
0.361083984375,
0.8896484375,
0.375732421875,
0.50341796875,
0.375732421875
],
[
0.080078125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.375732421875,
0.89208984375,
0.376953125,
0.89208984375,
0.391357421875,
0.50244140625,
0.39013671875
],
[
0.07843017578125,
0.39453125,
0.468017578125,
0.395751953125,
0.468017578125,
0.410400390625,
0.07843017578125,
0.4091796875
],
[
0.50341796875,
0.391357421875,
0.89208984375,
0.391357421875,
0.89208984375,
0.406005859375,
0.50341796875,
0.406005859375
],
[
0.50341796875,
0.40771484375,
0.890625,
0.40771484375,
0.890625,
0.41845703125,
0.50341796875,
0.41845703125
],
[
0.50244140625,
0.42236328125,
0.890625,
0.42236328125,
0.890625,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.271240234375,
0.436279296875,
0.281005859375,
0.436279296875,
0.281005859375,
0.44140625,
0.271240234375,
0.44140625
],
[
0.1732177734375,
0.4443359375,
0.281005859375,
0.4443359375,
0.281005859375,
0.458984375,
0.1732177734375,
0.458984375
],
[
0.275390625,
0.44580078125,
0.374267578125,
0.44140625,
0.375732421875,
0.45703125,
0.27685546875,
0.4609375
],
[
0.4453125,
0.44580078125,
0.4697265625,
0.44580078125,
0.4697265625,
0.458984375,
0.4453125,
0.458984375
],
[
0.265625,
0.462158203125,
0.285888671875,
0.462158203125,
0.285888671875,
0.472900390625,
0.265625,
0.472900390625
],
[
0.521484375,
0.46728515625,
0.54833984375,
0.46728515625,
0.54833984375,
0.47607421875,
0.521484375,
0.47607421875
],
[
0.55615234375,
0.46533203125,
0.8896484375,
0.464111328125,
0.8896484375,
0.479248046875,
0.55615234375,
0.479736328125
],
[
0.08087158203125,
0.481689453125,
0.325927734375,
0.481689453125,
0.325927734375,
0.49560546875,
0.08087158203125,
0.49560546875
],
[
0.50341796875,
0.47607421875,
0.55615234375,
0.47607421875,
0.55615234375,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.869140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.869140625,
0.501953125
],
[
0.098876953125,
0.5068359375,
0.465576171875,
0.5068359375,
0.465576171875,
0.52099609375,
0.098876953125,
0.52099609375
],
[
0.50244140625,
0.51513671875,
0.64208984375,
0.5166015625,
0.64208984375,
0.53076171875,
0.50244140625,
0.52978515625
],
[
0.6806640625,
0.51513671875,
0.8935546875,
0.5166015625,
0.8935546875,
0.53076171875,
0.6806640625,
0.52978515625
],
[
0.080078125,
0.52294921875,
0.468017578125,
0.52294921875,
0.468017578125,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.080078125,
0.537109375,
0.4697265625,
0.537109375,
0.4697265625,
0.55029296875,
0.080078125,
0.55029296875
],
[
0.50341796875,
0.5322265625,
0.89208984375,
0.53076171875,
0.89208984375,
0.54541015625,
0.50341796875,
0.546875
],
[
0.50244140625,
0.546875,
0.890625,
0.546875,
0.890625,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.08087158203125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.564453125,
0.08087158203125,
0.564453125
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.50244140625,
0.57763671875,
0.87109375,
0.57763671875,
0.87109375,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.08087158203125,
0.599609375,
0.465576171875,
0.599609375,
0.465576171875,
0.60986328125,
0.08087158203125,
0.60986328125
],
[
0.080078125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.080078125,
0.626953125
],
[
0.50341796875,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.50341796875,
0.625
],
[
0.07843017578125,
0.62744140625,
0.46728515625,
0.626953125,
0.46728515625,
0.6416015625,
0.07843017578125,
0.642578125
],
[
0.50341796875,
0.626953125,
0.87353515625,
0.626953125,
0.87353515625,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.50341796875,
0.6416015625,
0.89208984375,
0.6396484375,
0.89208984375,
0.65576171875,
0.50341796875,
0.65771484375
],
[
0.080078125,
0.6494140625,
0.29736328125,
0.6513671875,
0.29736328125,
0.66650390625,
0.080078125,
0.6640625
],
[
0.50341796875,
0.65771484375,
0.890625,
0.66015625,
0.890625,
0.6748046875,
0.50341796875,
0.67236328125
],
[
0.09722900390625,
0.673828125,
0.465576171875,
0.673828125,
0.465576171875,
0.6875,
0.09722900390625,
0.6875
],
[
0.5009765625,
0.673828125,
0.81640625,
0.673828125,
0.81640625,
0.6875,
0.5009765625,
0.6875
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.521484375,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.521484375,
0.7041015625
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.7197265625,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.41015625,
0.734375,
0.41015625,
0.74853515625,
0.080078125,
0.75
],
[
0.5048828125,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.50341796875,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.1829833984375,
0.76318359375,
0.2091064453125,
0.76318359375,
0.2091064453125,
0.7763671875,
0.1829833984375,
0.7763671875
],
[
0.22216796875,
0.76025390625,
0.420654296875,
0.76123046875,
0.420654296875,
0.77880859375,
0.22216796875,
0.77783203125
],
[
0.50341796875,
0.767578125,
0.890625,
0.767578125,
0.890625,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.1544189453125,
0.77783203125,
0.23779296875,
0.77978515625,
0.23779296875,
0.79443359375,
0.1544189453125,
0.79248046875
],
[
0.1258544921875,
0.7646484375,
0.1650390625,
0.76220703125,
0.1666259765625,
0.7763671875,
0.12744140625,
0.77880859375
],
[
0.44677734375,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79248046875,
0.44677734375,
0.79248046875
],
[
0.50341796875,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.125,
0.79736328125,
0.2900390625,
0.79736328125,
0.2900390625,
0.8125,
0.125,
0.8125
],
[
0.5009765625,
0.79541015625,
0.89208984375,
0.79443359375,
0.89208984375,
0.8095703125,
0.5009765625,
0.81005859375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07598876953125,
0.8193359375,
0.471435546875,
0.82080078125,
0.471435546875,
0.84228515625,
0.07598876953125,
0.84130859375
],
[
0.50244140625,
0.82568359375,
0.88232421875,
0.82568359375,
0.88232421875,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.521484375,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.869140625
],
[
0.50244140625,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50244140625,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.2900390625,
0.88720703125,
0.2900390625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48193359375,
0.9267578125,
0.491943359375,
0.9267578125,
0.491943359375,
0.9345703125,
0.48193359375,
0.9345703125
]
] |
[
"3.3. Cvclic α-Expansion (CAE)",
"4.1.1",
"QUBODerivation",
"CCuantuMM represents matchings as permutation ma-",
"To perform optimisation on the quantumannealer, we need",
"trices. In order to update them, we build on Seelbach et",
"totransform(5)intoaOUBOproblem.WeadapttheCAE",
"al.'s CAE algorithm [42] (similar to a fusion move [27]),",
"formulation from [42] (see Sec. 3.3) and iteratively update",
"which we describe here. A permutation matrix P is called",
"the permutations to decrease the value of (5). Given a set",
"an r-cycle, if there exist r disjoint indices ii,..., ir such",
"C = (c;}'-1 of k disjoint 2-cycles and binary decision",
"that Pi,iu+1)%",
"= 1 for all j e {1,..,r) and P = 1",
"variables α, we can parameterise our permutation matri-",
"for all l [ii,.., i], in which case P = (iii2 ...ir) is",
"ces as P(α) = Pr + h-1 Qi(c - I)Pr. However,",
"CAE alone is not sufficient to transform (5) into a QUBO",
"a common notation. Two cycles, i.e. two permutation ma",
"trices, are disjoint if these indices are pairwise disjoint. We",
"as cyclic consistency is still missing.A simple solution",
"would be to encourage cyclic consistency as a quadratic soft",
"know that disjoint cycles commute, which allows us to rep",
"penalty, but then there are no guarantees on the solution. In-",
"resent any permutation P as P = (1I'-- c) (1i-- c)",
"stead, we enforce cyclic consistency by construction:",
"where [ci)i and [c]; each are sets of disjoint 2-cycles.",
"Pxz(α,β) =",
"(6)",
"Given a set (ck-1 of k disjoint 2-cycles, an update,",
"or modification, of P can therefore be parameterised as:",
"ai(c - I)Pxy) (Pyz +",
"β,(c, -I)Pyz),",
"P(a) = -, caP, where α E {0, 1)k is a binary de-",
"(Pxy-",
"",
"j=1",
"cision vector determining the update. (Note that Q, in cg",
"is an exponent, not an index.) Crucially, to make this pa-",
"where (c;) ({c,l) are cycles and α (β) are decision vari-",
"rameterisation compatible with QUBOs,we need to make",
"ables for the updates to Pxy (Pyz). For brevity, we write",
"it linear in Q. To this end, CAE uses the following equality:",
"Ci = (ci - I)Pxy and C, = (C, - I)Pyz. We explain",
"how we construct the cycles in Secs. 4.1.2-4.1.3. Thus, we",
"iteratively solve (5) via a sequence of problems of the form:",
"K",
"P(α) =P+>",
"CQi(G -I)P",
"(4)",
"2=1",
"min",
"Exy(Pxy(α) + Eyz(Pyz(B)) +Exz(Pxz(α, β),",
"4.Our CCuantuMM Method",
"α,8E(0,1)",
"(7)",
"Previous adiabatic quantum computing methods [41,42]",
"whereEr(P,Q)",
"vec(P)TWLvec(Q)and",
"can only match two shapes. We present a method for match-",
"ing N shapes. To ensure cycle consistency on N shapes, it",
"Er(P) = Er(P, P). While the first two terms are in",
"QUBO form, the third term contains cubic and bi-quadratic",
"is sufficientthat alltriplets of shapes are matched cyclecon-",
"sistently [25]. CCuantuMM iteratively solves three-shape",
"terms (see the supplement for details) which are not",
"problems, which preserve cycle consistency by construc-",
"compatible with current quantum annealer architectures.",
"tion and fit on existing quantum annealers with limited re-",
"sources. We introduce our formulation for matching three",
"Higher-Order Terms. All of these higher-order terms",
"shapes in Sec. 4.1 and then extend it to N shapes in Sec. 4.2.",
"come from Pxz(a,β), specifically from the term H",
", Z, QiB,C,Cj. As we only consider 2-cycles, C, and",
"4.1. Matching Three Shapes",
"C, each have only four non-zero elements. Due to this ex-",
"Consider the problem of matching three non-rigidly de-",
"treme sparsity, most summands of H become 0.",
"formed shapes S = {X,, Z] of n vertices each, while",
"We could tackle these undesirable terms by decompos-",
"preserving cycle consistency. We formulate this as an en-",
"ing them into quadratic terms by using ancilla variables and",
"ergy minimisation with respect to P = {Pr E Pn|I, J E",
"adding penalty terms [16]. This gives exact solutions for",
"S), the set of permutations between all pairs in S:",
"sufficiently high weights of the penalty terms.However,",
"multiple reasons speak against this: (1) the QUBO matrix",
"",
"vec(Prg) TWL vec(Pr),",
"is alreadydense(a clique)underthe currentformulation(as",
"I,JES;I+J",
"min",
"(5)",
"we will see in (9) and adding ancilla qubits scales quadrat-",
"s.t.Pxz =PxyPyz,",
"ically in k, (2) adding penalties makes the problem harder",
"to solve, and (3) H is sparsely non-zero and in practise we",
"where Wi E IRn2xn’ is the energy matrix describing",
"observe no drastic influence on the quality of the solution.",
"how well certain pairwise properties are conserved between",
"Alternatively, we could assume H-0. However, this",
"shapes I and J (see Sec. 3.2), and Pxz = PxyPyz en-",
"is unnecessarily strong since (1)H also contributes to",
"forces cyclic consistency. An overview of the algorithm for",
"quadratic terms (E(H, ), and (2) higher-order terms op-",
"three shapes is shown in Alg. 1.",
"erating on the same decision variable trivially reduce to",
"4"
] |
{
"arxiv_id": "2303.16202",
"language": "en",
"page_idx": 3
}
|
||
2303.16202-en-4
|
2303.16202-en-4.jpg
|
quadratic terms: $\alpha_i\cdot\alpha_i=\alpha_i$ for binary $\alpha_i.$ We thus keep those two types of terms and merely assume all truly cubic and bi-quadratic terms to be zero.
Cycle-Consistent CAE. After eliminating the higherorder terms and ignoring constants from (7), we obtain (with the same colour coding):
$$\begin{aligned}&\min_{\alpha,\beta}\quad\sum_{i=1}^{k}\alpha_{i}\left(F_{X,Y}(P_{X,Y},C_{i})+F_{X,Z}(P_{X,Z},C_{i}P_{Y,Z})\right)\\&+\sum_{j=1}^{k}\beta_{j}\left(F_{Y,Z}(P_{Y,Z},\tilde{C}_{j})+F_{X,Z}(P_{X,Z},P_{X,Y}\tilde{C}_{j})\right)\\&+\sum_{i=1}^{k}\sum_{l=1}^{k}\alpha_{i}\alpha_{l}\left(E_{X,Y}(C_{i},C_{l})+E_{X,Z}(C_{i}P_{j},C_{l}P_{Y,Z})\right)\\&+\sum_{j=1}^{k}\sum_{l=1}^{k}\beta_{j}\beta_{l}\left(E_{Y,Z}(\tilde{C}_{j},\tilde{C}_{l})+E_{X,Z}(P_{X,Y}\tilde{C}_{j},P_{X,Y}\tilde{C}_{l})\right)\\&+\sum_{i=1}^{k}\sum_{j=1}^{k}\alpha_{i}\beta_{j}\left(F_{X,Z}(P_{X,Y}\tilde{C}_{j},C_{i}P_{Y,Z})+F_{X,Z}(K_{ij},P_{X,Z})\right)\\&+F_{X,Z}(K_{ij},P_{X,Y}\tilde{C}_{j})+F_{X,Z}(K_{ij},P_{Y,Z})+E_{X,Z}(K_{ij},K_{ij})\Big),\end{aligned}$$
where $P\chi_{\mathcal{Z}}=P_{\mathcal{X}\mathcal{Z}}(\mathbf{0},\mathbf{0})=P_{\mathcal{X}\mathcal{Y}}P_{\mathcal{Y}\mathcal{Z}}$, and we use the $\begin{array}{l}{\mathrm{shorthands~}F_{\mathcal{I}\mathcal{J}}(A,\dot{B})\:=\:E_{\mathcal{I}\mathcal{J}}(A,B)+E_{\mathcal{I}\mathcal{J}}(B,A)\mathrm{~and}}\\{K_{ij}\:=\:C_{i}\tilde{C}_{j}.\:\mathrm{Denoting~}\alpha_{k+j}\:=\:\beta_{j}\mathrm{~for~an~expanded}}\end{array}$ $\alpha\in\{0,1\}^{2k},(8)$can be written in the form
(9)
$$\min_{\alpha\in\{0,1\}^{2k}}\quad\alpha^{\top}\tilde{W}\alpha.$$
The full formula for $\tilde{W}$ is provided in the supplement. (9) is finally in QUBO form and we can optimise it classically or on real quantum hardware (see Sec. 3.1).
### 4.1.2 Choosing Vertices
The question of how to choose the sets of cycles $\{c_{i}\}_{i},\{\tilde{c}_{j}\}_{j}$ is still open. We first choose a subset of vertices using the“worst vertices”criterion introduced in [42] based on the relative inconsistency $I_\mathrm{\chi y}$ of a vertex $x\in\mathcal{X}$ under the current permutation:
(10)
$$I_{\mathcal{XY}}(x)=\sum_{w\in\mathcal{X}}W_{x\cdot n+x^{\top}P_{\mathcal{XY}},w\cdot n+w^{\top}P_{\mathcal{XY}}},$$
where we treat the one-hot vector $x^\top P_{\chi y}$ as a vertex index on $\mathcal{Y}$. A high value indicates that $x$ is inconsistent with many other matches under $P_{\chi y}$ and swapping it will likely improve the matching. We denote the set of the $m=2k$ vertices with the highest $I_{\chi y}(\cdot)$ as $V_{\chi}.$ Finally, we follow the permutations to get $V_{\mathcal{Y}}=\{x^{\top}P_{\mathcal{X}\mathcal{Y}}|x\in V_{\mathcal{X}}\}\subset\mathcal{Y}.$
In practice, we observe a systematic improvement in the matchings when considering all three possibilities (using $I\chi y,I_{\gamma Z}$,or $I_{\chi Z}$ as the starting point). We thus use three “sub”-iterations per iteration, one for each possibility.
<table>
<tbody>
<tr>
<th>$V_{\chi}$ $t$</th>
<th>$w_x$ $。v_x$ TA $\textcircled{1}$ $u$</th>
<th>$u_x$ $u_x$ $2t.$</th>
<th>$V_{T}$ 22 $w_x$ 2 7</th>
<th>$w_{x}$ $t_x$ $t$</th>
<th>$u_{x}$ $f^{\prime}0$ $w_x$ $v_x$ 7 2 $u_{x}$ $w_{x}$ $C_{\chi}^{1}$ $t_x$ $v_x$ $\frac{1}{2}$</th>
</tr>
<tr>
<td>$t.$ $Vy$</td>
<td>$u_{y}$ $av_y$ $w_{1}$ 1/2</td>
<td>$u_{y}$ $u_{y}$ $u_{i}$</td>
<td>$v_{y}$ $-3y$ $w_{i}$ $y$ $y$</td>
<td>$t_{y}$ $t_{y}$ $u_{i}$</td>
<td>$u_{i}$ $w_{:}$ $y$ $u_{y}$ $\bar{v}_{1}$ 20 $2x$ $u$</td>
</tr>
</tbody>
</table>
$\begin{array}{l}{\mathrm{Figure~3.~We~depict~the~sub-iteration~that~starts~from~}I_{\chi y},\mathrm{from}}\\{\mathrm{which~we~construct~}V_{\chi},\mathrm{~then~}\mathcal{C}_{\chi}^{all},\mathrm{~and~finally~}\mathcal{C}_{\chi}=\{\mathcal{C}_{\chi}^{0},\mathcal{C}_{\chi}^{1},}\end{array}$ $\mathcal{C}_{X}^{2}\}.$ We also build $V_{y}$ from $V_{\chi}$ and construct $\mathcal{C}_{y}$ analogously. Matching each element of $C_{\chi}$ with one from $C_{y}$ (visualised via matching colours) leads to three sub-sub-iterations.
<FigureHere>
Figure 4. We depict the sub-sub-iteration for $\mathcal{C}_{\mathcal{X}}^2=\{(u_x,v_x)$,
$({\tilde{w}_{x},t_{x})}\}$ and $\mathcal{C}_{\mathcal{Y}}^{1}=\left\{(u_{y},w_{y}),(v_{y},t_{y})\right\}$ from Fig. 3
## 4.13 Choosing Cycles
Given the worst vertices $V_{\chi}$ and $V_{y}$ of any sub-iteration, we construct the cycles $\{c_i\}_i,\{\tilde{c}_j\}_j$ from them. Fig.3 visualises this process. Focusing on $V_{\chi}$ for the moment, we want to use all possible 2-cycles $\mathcal{C}_{\mathcal{X}}^{all}=\{(uv)|u,v\in$ $V_{\mathcal{X}},u\neq v\}$ in each sub-iteration. We cannot use all of these cycles at once since they are not disjoint, as CAE requires. Instead, we next construct a set $C_{\chi}$ by partitioning $\hat{\mathcal{C}}_{\chi}^{all}$ into $m-1$ sets of cycles with each containing $m/2=\tilde{k}$ disjoint cycles. An analogous methodology is used for $\mathcal{C}_{y}.$
We now have $\mathcal{C}_X$ and $\mathcal{C}_y.$ Since we want to consider $\begin{array}{l}{\text{each cycle of }\mathcal{C}_{\chi}^{all}\mathrm{~and~}\mathcal{C}_{\gamma}^{all}\mathrm{~once,~we~need~several~``sub-sub"}}\\{\text{iterations. Thus, we next neeed to pick one set of cycles}}\end{array}$ from each $C_{\mathcal{X}}$ and $C_{\mathcal{Y}}$ for each sub-sub-iteration. There are $(m-1)^2$ possible pairs between elements of $\mathcal{C}_X$ and $\mathcal{C}_y.$ Considering all possible pairs is redundant, does not provide significant performance advantage, and increases the computational complexity quadratically. Hence, we randomly pair each element of $\mathcal{C}_{\mathcal{X}}$ with one element of $\mathcal{C}_{\mathcal{Y}}$ (without replacement). This leads to $m-1$ sub-sub-iterations, with each one solving (9) with its respective cycles; see Fig. 4.
## 4.2. Matching $N\textbf{ Shapes}$
In this section, we extend our model to matching a shape collection $\mathcal{S}$ with $N$ elements by iteratively matching three shapes while still guaranteeing cycle consistency. Similar
5
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.5498046875,
0.091552734375,
0.56201171875,
0.0997314453125,
0.5537109375,
0.10797119140625,
0.5400390625,
0.0997314453125
],
[
0.6337890625,
0.09661865234375,
0.640625,
0.09661865234375,
0.640625,
0.10162353515625,
0.6337890625,
0.10162353515625
],
[
0.67626953125,
0.0953369140625,
0.68310546875,
0.0953369140625,
0.68310546875,
0.10101318359375,
0.67626953125,
0.10101318359375
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.5087890625,
0.10986328125,
0.52392578125,
0.10986328125,
0.52392578125,
0.11871337890625,
0.5087890625,
0.11871337890625
],
[
0.62744140625,
0.106689453125,
0.6708984375,
0.1029052734375,
0.67333984375,
0.11871337890625,
0.6298828125,
0.12249755859375
],
[
0.71728515625,
0.10986328125,
0.72216796875,
0.10986328125,
0.72216796875,
0.11614990234375,
0.71728515625,
0.11614990234375
],
[
0.74169921875,
0.111083984375,
0.7509765625,
0.10797119140625,
0.75634765625,
0.11553955078125,
0.74658203125,
0.11932373046875
],
[
0.8056640625,
0.111083984375,
0.814453125,
0.111083984375,
0.814453125,
0.11614990234375,
0.8056640625,
0.11614990234375
],
[
0.080078125,
0.12249755859375,
0.301513671875,
0.12249755859375,
0.301513671875,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.5361328125,
0.11932373046875,
0.54345703125,
0.11932373046875,
0.54345703125,
0.12249755859375,
0.5361328125,
0.12249755859375
],
[
0.6337890625,
0.1256103515625,
0.6435546875,
0.1256103515625,
0.6435546875,
0.1331787109375,
0.6337890625,
0.1331787109375
],
[
0.74609375,
0.125,
0.75244140625,
0.125,
0.75244140625,
0.1319580078125,
0.74609375,
0.1319580078125
],
[
0.8037109375,
0.126953125,
0.81396484375,
0.126953125,
0.81396484375,
0.1319580078125,
0.8037109375,
0.1319580078125
],
[
0.86767578125,
0.1256103515625,
0.8798828125,
0.1256103515625,
0.8798828125,
0.1319580078125,
0.86767578125,
0.1319580078125
],
[
0.71728515625,
0.1275634765625,
0.7236328125,
0.1275634765625,
0.7236328125,
0.1331787109375,
0.71728515625,
0.1331787109375
],
[
0.53369140625,
0.1287841796875,
0.54345703125,
0.1287841796875,
0.54345703125,
0.1331787109375,
0.53369140625,
0.1331787109375
],
[
0.6337890625,
0.1427001953125,
0.6435546875,
0.1427001953125,
0.6435546875,
0.149658203125,
0.6337890625,
0.149658203125
],
[
0.71337890625,
0.141357421875,
0.7236328125,
0.1383056640625,
0.7294921875,
0.1483154296875,
0.71826171875,
0.1514892578125
],
[
0.74609375,
0.1427001953125,
0.75244140625,
0.1427001953125,
0.75244140625,
0.1458740234375,
0.74609375,
0.1458740234375
],
[
0.08087158203125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1685791015625,
0.08087158203125,
0.1685791015625
],
[
0.50732421875,
0.1546630859375,
0.5400390625,
0.15283203125,
0.54150390625,
0.1641845703125,
0.5087890625,
0.1666259765625
],
[
0.6357421875,
0.154052734375,
0.6708984375,
0.1558837890625,
0.6689453125,
0.1673583984375,
0.6337890625,
0.1654052734375
],
[
0.080078125,
0.169189453125,
0.468017578125,
0.1685791015625,
0.468017578125,
0.18310546875,
0.080078125,
0.184326171875
],
[
0.54736328125,
0.169189453125,
0.560546875,
0.169189453125,
0.560546875,
0.1787109375,
0.54736328125,
0.1787109375
],
[
0.59326171875,
0.1717529296875,
0.60302734375,
0.1749267578125,
0.60107421875,
0.1787109375,
0.59228515625,
0.175537109375
],
[
0.67626953125,
0.173583984375,
0.6845703125,
0.173583984375,
0.6845703125,
0.177978515625,
0.67626953125,
0.177978515625
],
[
0.79736328125,
0.17041015625,
0.83837890625,
0.17041015625,
0.83837890625,
0.1817626953125,
0.79736328125,
0.1817626953125
],
[
0.080078125,
0.184326171875,
0.283447265625,
0.1849365234375,
0.283447265625,
0.199462890625,
0.080078125,
0.1988525390625
],
[
0.5009765625,
0.1962890625,
0.89208984375,
0.1956787109375,
0.89208984375,
0.210205078125,
0.5009765625,
0.2108154296875
],
[
0.1650390625,
0.21337890625,
0.39794921875,
0.212158203125,
0.39794921875,
0.2310791015625,
0.1650390625,
0.231689453125
],
[
0.50341796875,
0.2108154296875,
0.890625,
0.2108154296875,
0.890625,
0.2247314453125,
0.50341796875,
0.2247314453125
],
[
0.08905029296875,
0.2171630859375,
0.11767578125,
0.2171630859375,
0.11767578125,
0.231689453125,
0.08905029296875,
0.231689453125
],
[
0.1258544921875,
0.2203369140625,
0.1356201171875,
0.2203369140625,
0.1356201171875,
0.2259521484375,
0.1258544921875,
0.2259521484375
],
[
0.1380615234375,
0.2197265625,
0.158447265625,
0.2197265625,
0.158447265625,
0.2279052734375,
0.1380615234375,
0.2279052734375
],
[
0.50341796875,
0.2247314453125,
0.890625,
0.2247314453125,
0.890625,
0.238037109375,
0.50341796875,
0.238037109375
],
[
0.50244140625,
0.2392578125,
0.8935546875,
0.2392578125,
0.8935546875,
0.25244140625,
0.50244140625,
0.25244140625
],
[
0.0882568359375,
0.25,
0.3798828125,
0.25,
0.3798828125,
0.270751953125,
0.0882568359375,
0.270751953125
],
[
0.5009765625,
0.25244140625,
0.80615234375,
0.25244140625,
0.80615234375,
0.266357421875,
0.5009765625,
0.266357421875
],
[
0.188720703125,
0.290283203125,
0.41259765625,
0.290283203125,
0.41259765625,
0.30810546875,
0.188720703125,
0.30810546875
],
[
0.148681640625,
0.296630859375,
0.1641845703125,
0.29541015625,
0.1650390625,
0.302978515625,
0.1502685546875,
0.3037109375
],
[
0.7646484375,
0.300537109375,
0.78173828125,
0.314453125,
0.76953125,
0.323974609375,
0.75244140625,
0.31005859375
],
[
0.11273193359375,
0.309326171875,
0.13232421875,
0.309326171875,
0.13232421875,
0.313232421875,
0.11273193359375,
0.313232421875
],
[
0.6611328125,
0.31005859375,
0.736328125,
0.31005859375,
0.736328125,
0.328857421875,
0.6611328125,
0.328857421875
],
[
0.775390625,
0.313232421875,
0.8095703125,
0.324462890625,
0.80126953125,
0.3359375,
0.76708984375,
0.324462890625
],
[
0.1282958984375,
0.32763671875,
0.416748046875,
0.32763671875,
0.416748046875,
0.34521484375,
0.1282958984375,
0.34521484375
],
[
0.583984375,
0.330078125,
0.6337890625,
0.3017578125,
0.650390625,
0.318115234375,
0.6005859375,
0.3466796875
],
[
0.10296630859375,
0.3447265625,
0.1422119140625,
0.3447265625,
0.1422119140625,
0.354736328125,
0.10296630859375,
0.354736328125
],
[
0.615234375,
0.354248046875,
0.6298828125,
0.354248046875,
0.6298828125,
0.3623046875,
0.615234375,
0.3623046875
],
[
0.178955078125,
0.367431640625,
0.4345703125,
0.365478515625,
0.4345703125,
0.38330078125,
0.178955078125,
0.385009765625
],
[
0.560546875,
0.369384765625,
0.63232421875,
0.371826171875,
0.63232421875,
0.386962890625,
0.560546875,
0.385009765625
],
[
0.79345703125,
0.371826171875,
0.810546875,
0.371826171875,
0.810546875,
0.386474609375,
0.79345703125,
0.386474609375
],
[
0.1446533203125,
0.37255859375,
0.1748046875,
0.37255859375,
0.1748046875,
0.380615234375,
0.1446533203125,
0.380615234375
],
[
0.101318359375,
0.37255859375,
0.1094970703125,
0.375,
0.108642578125,
0.380615234375,
0.098876953125,
0.37744140625
],
[
0.81884765625,
0.39453125,
0.8310546875,
0.39453125,
0.8310546875,
0.40087890625,
0.81884765625,
0.40087890625
],
[
0.08660888671875,
0.40283203125,
0.455810546875,
0.40087890625,
0.455810546875,
0.417236328125,
0.08660888671875,
0.419189453125
],
[
0.615234375,
0.410888671875,
0.78369140625,
0.410888671875,
0.78369140625,
0.425537109375,
0.615234375,
0.425537109375
],
[
0.44775390625,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.43505859375,
0.44775390625,
0.43505859375
],
[
0.08087158203125,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.50341796875,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.455078125,
0.50341796875,
0.455078125
],
[
0.08087158203125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.473388671875,
0.08087158203125,
0.473388671875
],
[
0.50341796875,
0.455078125,
0.82275390625,
0.455078125,
0.82275390625,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.0767822265625,
0.4716796875,
0.471435546875,
0.472900390625,
0.471435546875,
0.490478515625,
0.0767822265625,
0.4892578125
],
[
0.080078125,
0.4892578125,
0.373291015625,
0.4892578125,
0.373291015625,
0.50390625,
0.080078125,
0.50390625
],
[
0.50341796875,
0.496826171875,
0.6689453125,
0.496826171875,
0.6689453125,
0.51025390625,
0.50341796875,
0.51025390625
],
[
0.281982421875,
0.51025390625,
0.340576171875,
0.51025390625,
0.340576171875,
0.5244140625,
0.281982421875,
0.5244140625
],
[
0.44677734375,
0.51123046875,
0.4697265625,
0.51123046875,
0.4697265625,
0.52587890625,
0.44677734375,
0.52587890625
],
[
0.50341796875,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.22216796875,
0.51220703125,
0.256591796875,
0.51220703125,
0.256591796875,
0.5244140625,
0.22216796875,
0.5244140625
],
[
0.205078125,
0.521484375,
0.26953125,
0.51953125,
0.26953125,
0.5341796875,
0.205078125,
0.5361328125
],
[
0.50341796875,
0.5361328125,
0.89208984375,
0.5361328125,
0.89208984375,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.08087158203125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.08087158203125,
0.556640625
],
[
0.50341796875,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.50341796875,
0.564453125
],
[
0.080078125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.564453125,
0.890625,
0.56494140625,
0.890625,
0.580078125,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.57275390625,
0.348876953125,
0.57275390625,
0.348876953125,
0.5859375,
0.080078125,
0.5859375
],
[
0.50341796875,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50341796875,
0.5947265625
],
[
0.50341796875,
0.5966796875,
0.890625,
0.5966796875,
0.890625,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.080078125,
0.60595703125,
0.25732421875,
0.60595703125,
0.25732421875,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.60986328125,
0.89208984375,
0.611328125,
0.89208984375,
0.62548828125,
0.50244140625,
0.625
],
[
0.08087158203125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.08087158203125,
0.64306640625
],
[
0.50244140625,
0.625,
0.89208984375,
0.62548828125,
0.89208984375,
0.64013671875,
0.50244140625,
0.6396484375
],
[
0.08087158203125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.66015625,
0.08087158203125,
0.66015625
],
[
0.50244140625,
0.6416015625,
0.8857421875,
0.6416015625,
0.8857421875,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.080078125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.080078125,
0.673828125
],
[
0.52294921875,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.52294921875,
0.67041015625
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.68994140625,
0.080078125,
0.68994140625
],
[
0.50244140625,
0.671875,
0.890625,
0.671875,
0.890625,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.080078125,
0.68994140625,
0.281982421875,
0.68994140625,
0.281982421875,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.6875,
0.8935546875,
0.6875,
0.8935546875,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.1168212890625,
0.708984375,
0.406982421875,
0.71337890625,
0.406982421875,
0.73486328125,
0.1168212890625,
0.73095703125
],
[
0.5048828125,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.7138671875,
0.5048828125,
0.7138671875
],
[
0.438720703125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72802734375,
0.438720703125,
0.72802734375
],
[
0.50244140625,
0.71630859375,
0.88818359375,
0.71728515625,
0.88818359375,
0.73193359375,
0.50244140625,
0.73095703125
],
[
0.191162109375,
0.73193359375,
0.2254638671875,
0.73193359375,
0.2254638671875,
0.74365234375,
0.191162109375,
0.74365234375
],
[
0.50341796875,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.50341796875,
0.7470703125
],
[
0.080078125,
0.75,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.08087158203125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50341796875,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.7763671875,
0.50341796875,
0.7763671875
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50341796875,
0.77783203125,
0.89208984375,
0.77783203125,
0.89208984375,
0.79150390625,
0.50341796875,
0.79150390625
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.79248046875,
0.89208984375,
0.79150390625,
0.89208984375,
0.80615234375,
0.5009765625,
0.806640625
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.8798828125,
0.80810546875,
0.8798828125,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.41259765625,
0.8271484375,
0.41259765625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.83203125,
0.69140625,
0.83349609375,
0.69140625,
0.84765625,
0.50244140625,
0.8466796875
],
[
0.098876953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.869140625
],
[
0.521484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.521484375,
0.869140625
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.43310546875,
0.88720703125,
0.43310546875,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
],
[
0.476318359375,
0.92529296875,
0.490234375,
0.92236328125,
0.494384765625,
0.93505859375,
0.48046875,
0.93798828125
]
] |
[
"quadratic terms: Q -Qi = Q, for binary Qi. We thus keep",
"Ws",
"T.",
"U",
"those two types of terms and merely assume all truly cubic",
"Va",
"WCall",
"P",
"",
"",
"and bi-quadratic terms to be zero.",
".",
"U.",
"u",
"C",
"W.",
"U",
"D",
"Uy",
"y",
"2",
"Cycle-Consistent CAE. After eliminating the higher-",
"Vyt",
"\"yCall",
"order terms and ignoring constants from (7), we obtain",
"Wu",
"",
".",
"Cw",
"(with the same colour coding):",
"Figure 3. We depict the sub-iteration that starts from Ixy, from",
"(Fxy(Pxy,C)+ Fxz(Pxz,CiPyz)",
"which we construct Vx, then Call, and finally Cx = {C, Cl,",
"min",
"S",
"a",
"C1. We also build Vy from Vx and construct Cy analogously.",
"Matching each element of Cx with one from Cy (visualised via",
"+EB, (Fyz(Pyz,C)+ Fxz(Pxz, PxyC)",
"matching colours) leads to three sub-sub-iterations.",
"Exy(Ct.Ci)+Exz(C,Pyz,CiPyz)",
"Xi",
"Py",
"11",
"(B0（3)",
"2(B)",
"B;Bt(Eyz(Cj,Ci) +Exz(PxyCj, PxC)",
"Pxy(a)",
"=1=1",
"W",
"Fxz(PxyCj,CiPyz)+Fxz(Ktj,Pxz)",
"aoXa1",
"Z",
"aiB;",
"",
"24",
"+Fxz(,PxyC)+Fxz(,C,Pyz)+Exz(,K)",
"Pxz = Pxy(a)Pyz(β)",
"(8)",
"where Pxz = Pxz(0, 0) = PxyPyz, and we use the",
"Figure 4. We depict the sub-sub-iteration for C = (ur, ua),",
"shorthands Fr(A, B) = Er(A, B) + E(B, A) and",
"(Wa,t)) and Cy = [(uy, Wy), (uy, ty)) from Fig. 3.",
"Ki = C;Cj. Denoting Qk+j = β, for an expanded",
"Q E (0, 1j2k, (8) can be written in the form:",
"4.1.3Choosing Cycles",
"αTWa.",
"(9)",
"Given the worst vertices Vx and Vy of any sub-iteration,",
"min",
"αE(0,1)2k",
"we construct the cycles (cili, (c,l, from them. Fig. 3",
"The full formula for W is provided in the supplement. (9) is",
"visualises this process. Focusing on Vx for the moment,",
"finally in QUBO form and we can optimise it classically or",
"we want to use all possible 2-cycles Call = ((uv)|u, w E",
"on real quantum hardware (see Sec. 3.1).",
"Vx,u l in each sub-iteration.We cannot use all of",
"these cycles at once since they are not disjoint, as CAE re-",
"4.1.2 Choosing Vertices",
"quires. Instead, we next construct a set Cx by partitioning",
"The question of how to choose the sets of cycles",
"Call into m-1 sets of cycles with each containing m/2=k",
"[c:li, (c,lj is still open. We first choose a subset of ver-",
"disjoint cycles. An analogous methodology is used for Cy.",
"tices using the “worst vertices\" criterion introduced in [42]",
"We now have Cx and Cy. Since we want to consider",
"based on the relative inconsistency Ixy of a vertex E X",
"each cycle of Call and Call once, we need several \"sub-sub\"",
"under the current permutation:",
"iterations. Thus, we next need to pick one set of cycles",
"Ixy(n) = Wr-n+aTPxy,n+uT Pxy,",
"fromeach Cyand Cyfor each sub-sub-iteration.There",
"(10)",
"are (m-1)2 possible pairs between elements of Cx and Cy",
"wex",
"Considering all possible pairs is redundant, does not provide",
"where we treat the one-hot vector T Pxy as a vertex in-",
"significant performance advantage, and increases the com-",
"dex on V. A high value indicates that r is inconsistent with",
"putational complexity quadratically. Hence, we randomly",
"many other matches under Pxy and swapping it will likely",
"pair each element of Cx with one element of Cy (without",
"improve the matching. We denote the set of the m=2k ver-",
"replacement). This leads to m-1 sub-sub-iterations, with",
"tices with the highest Ixy() as Vx. Finally, we follow the",
"each one solving (9) with its respective cycles; see Fig. 4.",
"permutations to get V = [ Pxy E Vx] C y",
"4.2. Matching N Shapes",
"In practice, we observe a systematic improvement in the",
"matchings when considering all three possibilities (using",
"In this section, we extend our model to matching a shape",
"Ixy, Iyz, or Ixz as the starting point). We thus use three",
"collection S with N elements by iteratively matching three",
"\"'sub\"-iterations per iteration, one for each possibility.",
"shapes while still guaranteeing cycle consistency. Similar",
"5"
] |
{
"arxiv_id": "2303.16202",
"language": "en",
"page_idx": 4
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.