id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2008.04115-en-1
|
2008.04115-en-1.jpg
|
T-GD: Transferable GAN-generated Images Detection Framework
<FigureHere>
Figure 2. Overview of our T-GD network. For efficient transfer learning, our network uses $L^2$-Starting Point (red) and Self-training (blue).
solution for transfer learning. As the regularization strength changes, the control of the regularization can easily be lost. Our approach overcomes this issue through self-training, where the teacher model automatically helps control the strength of regularization when the student model learns from the target dataset.
However, above methods are ineffective for improving transfer learning performance. That is, when the CNN classifier is trained on one dataset, it shows poor performance on other datasets. ForensicTransfer (Cozzolino et al., 2018) introduced an autoencoder for the GAN-image detection. They apply autoencoder and detect GAN-images through reconstruction error. This learning method has advantages regarding lower data usage, when the model is well trained. Although the ForensicTransfer showed promise for model transferability, its performance remains mediocre. In previous research, either artifacts, patterns, or augmentations were utilized individually for successful transfer learning, yet it is possible to combine them to transfer knowledge of GAN-image detection.
In addition, we introduce a novel augmentation method to solve the over-fitting problem by transforming Cutmix (Yun et al., 2019), which randomly mixes up a rectangle patch of training images. Note that the original Cutmix mixes up inter-class images; our experiment showed that this renders the learning process highly unstable for a binary classifer, and thus we transform the inter-class Cutmix to an intra-class Cutmix to increase the stability of the learning process. Also, our method combines Gaussian blur (Xuan et al., 2019) and Joint Photograph Experts Group (JPEG) compression (Wang et al., 2019), which were previously studied for GAN-image detection.
For transfer learning, we apply learning strategies, such as Weight Standardization (Qiao et al., 2019) (WS), Group Normalization (Wu&He,2018) (GN), and the tuning of learning rate and momentum rate (Li et al., 2020). WS and GN achieved comparable performance on image classification, not depending on batch size statistics. Also, we implement transfer learning using low learning and momentum rates for stochastic gradient descent (SGD) inspired by (Li et al., 2020). These have been experimented on object detection transfer, but we demonstrate that these strategies also work well for transfer learning across different domains. Fi-
In this paper, our objective is to maintain a high detection performance during transfer learning on the source and target datasets, without suffering catastrophic forget- $ting.$ While many studies on transfer learning have already shown impressive performance, they have not applied for GAN-image detection. Therefore, in this work, we propose a novel regularization method with self-training for transfer learning by combining and transforming regularization, augmentation, self-training, and learning strategies to improve transferability of GAN-image detection. In particular, our approach is inspired by starting point as the reference (SPAR), $L^{2}$-$SP$ (Li et al., 2018), which regularizes the weight variation of a target model by referring to the weights pre-trained on the source dataset. The limitation of the latter method is that it cannot provide an optimal
|
[
[
0.27783203125,
0.05682373046875,
0.69677734375,
0.05682373046875,
0.69677734375,
0.0706787109375,
0.27783203125,
0.0706787109375
],
[
0.583984375,
0.1337890625,
0.68310546875,
0.1337890625,
0.68310546875,
0.1458740234375,
0.583984375,
0.1458740234375
],
[
0.43310546875,
0.15087890625,
0.52392578125,
0.15087890625,
0.52392578125,
0.1610107421875,
0.43310546875,
0.1610107421875
],
[
0.74169921875,
0.147705078125,
0.75341796875,
0.147705078125,
0.75341796875,
0.1514892578125,
0.74169921875,
0.1514892578125
],
[
0.1446533203125,
0.1558837890625,
0.191162109375,
0.1558837890625,
0.191162109375,
0.1634521484375,
0.1446533203125,
0.1634521484375
],
[
0.2181396484375,
0.154052734375,
0.26318359375,
0.154052734375,
0.26318359375,
0.1641845703125,
0.2181396484375,
0.1641845703125
],
[
0.56201171875,
0.1610107421875,
0.6259765625,
0.1634521484375,
0.6259765625,
0.1749267578125,
0.56201171875,
0.17236328125
],
[
0.3544921875,
0.173583984375,
0.420654296875,
0.173583984375,
0.420654296875,
0.184326171875,
0.3544921875,
0.184326171875
],
[
0.56787109375,
0.173583984375,
0.625,
0.173583984375,
0.625,
0.184326171875,
0.56787109375,
0.184326171875
],
[
0.348876953125,
0.18310546875,
0.42236328125,
0.18310546875,
0.42236328125,
0.1932373046875,
0.348876953125,
0.1932373046875
],
[
0.346435546875,
0.1932373046875,
0.4248046875,
0.1932373046875,
0.4248046875,
0.203857421875,
0.346435546875,
0.203857421875
],
[
0.48779296875,
0.1932373046875,
0.5009765625,
0.189453125,
0.5048828125,
0.1988525390625,
0.491943359375,
0.2020263671875
],
[
0.443603515625,
0.2008056640625,
0.50341796875,
0.2008056640625,
0.50341796875,
0.2052001953125,
0.443603515625,
0.2052001953125
],
[
0.76708984375,
0.1956787109375,
0.7998046875,
0.1956787109375,
0.7998046875,
0.20703125,
0.76708984375,
0.20703125
],
[
0.463134765625,
0.216552734375,
0.53173828125,
0.216552734375,
0.53173828125,
0.2266845703125,
0.463134765625,
0.2266845703125
],
[
0.52294921875,
0.2310791015625,
0.611328125,
0.23291015625,
0.611328125,
0.2442626953125,
0.52294921875,
0.242431640625
],
[
0.75634765625,
0.2291259765625,
0.810546875,
0.2291259765625,
0.810546875,
0.2392578125,
0.75634765625,
0.2392578125
],
[
0.55615234375,
0.24365234375,
0.58251953125,
0.24365234375,
0.58251953125,
0.251953125,
0.55615234375,
0.251953125
],
[
0.490234375,
0.272705078125,
0.5498046875,
0.272705078125,
0.5498046875,
0.27978515625,
0.490234375,
0.27978515625
],
[
0.17724609375,
0.27978515625,
0.25,
0.27978515625,
0.25,
0.290283203125,
0.17724609375,
0.290283203125
],
[
0.488525390625,
0.281005859375,
0.51953125,
0.281005859375,
0.51953125,
0.28857421875,
0.488525390625,
0.28857421875
],
[
0.5185546875,
0.281005859375,
0.55810546875,
0.281005859375,
0.55810546875,
0.28857421875,
0.5185546875,
0.28857421875
],
[
0.76708984375,
0.3037109375,
0.80126953125,
0.3037109375,
0.80126953125,
0.31494140625,
0.76708984375,
0.31494140625
],
[
0.58251953125,
0.322509765625,
0.6845703125,
0.322509765625,
0.6845703125,
0.333984375,
0.58251953125,
0.333984375
],
[
0.4345703125,
0.33837890625,
0.52392578125,
0.33837890625,
0.52392578125,
0.349853515625,
0.4345703125,
0.349853515625
],
[
0.346435546875,
0.352783203125,
0.410888671875,
0.352783203125,
0.410888671875,
0.363037109375,
0.346435546875,
0.363037109375
],
[
0.340576171875,
0.3623046875,
0.418212890625,
0.3623046875,
0.418212890625,
0.37255859375,
0.340576171875,
0.37255859375
],
[
0.338134765625,
0.371826171875,
0.420654296875,
0.371826171875,
0.420654296875,
0.38330078125,
0.338134765625,
0.38330078125
],
[
0.348876953125,
0.3818359375,
0.41259765625,
0.3818359375,
0.41259765625,
0.392578125,
0.348876953125,
0.392578125
],
[
0.0882568359375,
0.410888671875,
0.88623046875,
0.410888671875,
0.88623046875,
0.425537109375,
0.0882568359375,
0.425537109375
],
[
0.08905029296875,
0.451904296875,
0.4755859375,
0.451904296875,
0.4755859375,
0.46533203125,
0.08905029296875,
0.46533203125
],
[
0.5009765625,
0.451904296875,
0.8857421875,
0.451904296875,
0.8857421875,
0.46533203125,
0.5009765625,
0.46533203125
],
[
0.0882568359375,
0.46728515625,
0.4755859375,
0.46728515625,
0.4755859375,
0.48095703125,
0.0882568359375,
0.48095703125
],
[
0.5,
0.46728515625,
0.88623046875,
0.46728515625,
0.88623046875,
0.48095703125,
0.5,
0.48095703125
],
[
0.0882568359375,
0.481689453125,
0.473876953125,
0.481689453125,
0.473876953125,
0.49560546875,
0.0882568359375,
0.49560546875
],
[
0.5,
0.481689453125,
0.88623046875,
0.481689453125,
0.88623046875,
0.496826171875,
0.5,
0.496826171875
],
[
0.0882568359375,
0.49755859375,
0.4755859375,
0.49755859375,
0.4755859375,
0.51123046875,
0.0882568359375,
0.51123046875
],
[
0.498291015625,
0.49755859375,
0.8857421875,
0.49755859375,
0.8857421875,
0.51123046875,
0.498291015625,
0.51123046875
],
[
0.0882568359375,
0.51123046875,
0.4755859375,
0.51220703125,
0.4755859375,
0.5263671875,
0.0882568359375,
0.52587890625
],
[
0.5,
0.51318359375,
0.88623046875,
0.51318359375,
0.88623046875,
0.5263671875,
0.5,
0.5263671875
],
[
0.08905029296875,
0.52783203125,
0.473876953125,
0.52783203125,
0.473876953125,
0.54150390625,
0.08905029296875,
0.54150390625
],
[
0.5,
0.52783203125,
0.65185546875,
0.52783203125,
0.65185546875,
0.54150390625,
0.5,
0.54150390625
],
[
0.0882568359375,
0.54345703125,
0.473876953125,
0.54345703125,
0.473876953125,
0.556640625,
0.0882568359375,
0.556640625
],
[
0.5,
0.55029296875,
0.88623046875,
0.55029296875,
0.88623046875,
0.564453125,
0.5,
0.564453125
],
[
0.0882568359375,
0.55810546875,
0.473876953125,
0.55810546875,
0.473876953125,
0.5712890625,
0.0882568359375,
0.5712890625
],
[
0.498291015625,
0.564453125,
0.8857421875,
0.56494140625,
0.8857421875,
0.580078125,
0.498291015625,
0.5791015625
],
[
0.08905029296875,
0.57470703125,
0.47216796875,
0.57470703125,
0.47216796875,
0.58544921875,
0.08905029296875,
0.58544921875
],
[
0.5,
0.58056640625,
0.88623046875,
0.58056640625,
0.88623046875,
0.5947265625,
0.5,
0.5947265625
],
[
0.0882568359375,
0.58837890625,
0.476318359375,
0.58837890625,
0.476318359375,
0.6015625,
0.0882568359375,
0.6015625
],
[
0.5,
0.59521484375,
0.88623046875,
0.59521484375,
0.88623046875,
0.60986328125,
0.5,
0.60986328125
],
[
0.0882568359375,
0.60302734375,
0.473876953125,
0.60302734375,
0.473876953125,
0.6162109375,
0.0882568359375,
0.6162109375
],
[
0.5009765625,
0.60986328125,
0.8857421875,
0.60986328125,
0.8857421875,
0.62353515625,
0.5009765625,
0.62353515625
],
[
0.0882568359375,
0.61865234375,
0.473876953125,
0.61865234375,
0.473876953125,
0.6318359375,
0.0882568359375,
0.6318359375
],
[
0.5,
0.62548828125,
0.88623046875,
0.62548828125,
0.88623046875,
0.6396484375,
0.5,
0.6396484375
],
[
0.0882568359375,
0.63330078125,
0.4755859375,
0.63330078125,
0.4755859375,
0.646484375,
0.0882568359375,
0.646484375
],
[
0.5009765625,
0.64013671875,
0.88623046875,
0.64013671875,
0.88623046875,
0.65380859375,
0.5009765625,
0.65380859375
],
[
0.08660888671875,
0.646484375,
0.23779296875,
0.646484375,
0.23779296875,
0.66015625,
0.08660888671875,
0.66015625
],
[
0.5,
0.65576171875,
0.8857421875,
0.65576171875,
0.8857421875,
0.66943359375,
0.5,
0.66943359375
],
[
0.08905029296875,
0.67041015625,
0.4755859375,
0.67041015625,
0.4755859375,
0.68505859375,
0.08905029296875,
0.68505859375
],
[
0.5,
0.671875,
0.88623046875,
0.671875,
0.88623046875,
0.68505859375,
0.5,
0.68505859375
],
[
0.0882568359375,
0.68603515625,
0.473876953125,
0.68603515625,
0.473876953125,
0.69970703125,
0.0882568359375,
0.69970703125
],
[
0.5,
0.68603515625,
0.8857421875,
0.68603515625,
0.8857421875,
0.69970703125,
0.5,
0.69970703125
],
[
0.08905029296875,
0.7021484375,
0.473876953125,
0.7021484375,
0.473876953125,
0.71533203125,
0.08905029296875,
0.71533203125
],
[
0.5009765625,
0.70068359375,
0.8857421875,
0.70068359375,
0.8857421875,
0.71533203125,
0.5009765625,
0.71533203125
],
[
0.0882568359375,
0.71630859375,
0.473876953125,
0.71630859375,
0.473876953125,
0.72998046875,
0.0882568359375,
0.72998046875
],
[
0.5,
0.71630859375,
0.72607421875,
0.71630859375,
0.72607421875,
0.72998046875,
0.5,
0.72998046875
],
[
0.08905029296875,
0.73193359375,
0.4755859375,
0.73193359375,
0.4755859375,
0.74560546875,
0.08905029296875,
0.74560546875
],
[
0.08660888671875,
0.744140625,
0.476318359375,
0.7470703125,
0.476318359375,
0.76123046875,
0.08660888671875,
0.7587890625
],
[
0.5009765625,
0.7392578125,
0.8857421875,
0.7392578125,
0.8857421875,
0.75244140625,
0.5009765625,
0.75244140625
],
[
0.5,
0.75244140625,
0.88623046875,
0.75390625,
0.88623046875,
0.76806640625,
0.5,
0.767578125
],
[
0.08905029296875,
0.76220703125,
0.4755859375,
0.76220703125,
0.4755859375,
0.77587890625,
0.08905029296875,
0.77587890625
],
[
0.5009765625,
0.76806640625,
0.8857421875,
0.76806640625,
0.8857421875,
0.7822265625,
0.5009765625,
0.7822265625
],
[
0.08905029296875,
0.7763671875,
0.473876953125,
0.7763671875,
0.473876953125,
0.79052734375,
0.08905029296875,
0.79052734375
],
[
0.5,
0.7841796875,
0.88623046875,
0.7841796875,
0.88623046875,
0.79736328125,
0.5,
0.79736328125
],
[
0.0882568359375,
0.79150390625,
0.4755859375,
0.79150390625,
0.4755859375,
0.8046875,
0.0882568359375,
0.8046875
],
[
0.5,
0.79736328125,
0.88623046875,
0.798828125,
0.88623046875,
0.81298828125,
0.5,
0.8125
],
[
0.08905029296875,
0.80810546875,
0.47216796875,
0.80810546875,
0.47216796875,
0.818359375,
0.08905029296875,
0.818359375
],
[
0.5,
0.814453125,
0.88818359375,
0.814453125,
0.88818359375,
0.82763671875,
0.5,
0.82763671875
],
[
0.0882568359375,
0.8212890625,
0.473876953125,
0.8212890625,
0.473876953125,
0.8359375,
0.0882568359375,
0.8359375
],
[
0.08660888671875,
0.83544921875,
0.473876953125,
0.8359375,
0.473876953125,
0.85107421875,
0.08660888671875,
0.849609375
],
[
0.5,
0.830078125,
0.8857421875,
0.830078125,
0.8857421875,
0.84326171875,
0.5,
0.84326171875
],
[
0.5,
0.8447265625,
0.88623046875,
0.8447265625,
0.88623046875,
0.85791015625,
0.5,
0.85791015625
],
[
0.0906982421875,
0.85400390625,
0.47216796875,
0.85400390625,
0.47216796875,
0.8642578125,
0.0906982421875,
0.8642578125
],
[
0.5009765625,
0.85986328125,
0.8857421875,
0.85986328125,
0.8857421875,
0.87353515625,
0.5009765625,
0.87353515625
],
[
0.0882568359375,
0.8671875,
0.473876953125,
0.8671875,
0.473876953125,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.5,
0.875,
0.88623046875,
0.875,
0.88623046875,
0.88818359375,
0.5,
0.88818359375
],
[
0.0882568359375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4755859375,
0.896484375,
0.0882568359375,
0.896484375
],
[
0.5009765625,
0.8916015625,
0.8857421875,
0.8916015625,
0.8857421875,
0.9013671875,
0.5009765625,
0.9013671875
]
] |
[
"T-GD: Transferable GAN-generated Images Detection Framework",
"Cross-EntropyLoss",
"TeacherClassifier",
"Rsal",
"RealImag",
"GAN-Ima",
"Self-training",
"Pre-Trained",
"Feedback",
"Teacher Model",
"onSourceData",
"ona",
"Regularizatior",
"Predict",
"L?Starting Poin",
"Regualarization",
"Final Loss",
"Loss",
"Fully-Cohnecte",
"Target dataset",
"L2regt",
"lhrization",
"Target",
"Cross-EntropyLoss",
"Student Classifier",
"Pre-Trained",
"Student Model",
"to be Trained on",
"Target Data",
"Figure 2. Overview of our T-GD network. For efficient transfer learning, our network uses L?-Starting Point (red) and Self-training (blue)",
"However, above methods are ineffective for improving trans-",
"solution for transfer learning. As the regularization strength",
"fer learning performance. That is, when the CNN classifier",
"changes, the control of the regularization can easily be lost.",
"is trained on one dataset, it shows poor performance on",
"Our approach overcomes this issue through self-training,",
"other datasets. ForensicTransfer (Cozzolino et al., 2018)",
"where the teacher model automatically helps control the",
"introduced an autoencoder for the GAN-image detection.",
"strength of regularization when the student model learns",
"They apply autoencoder and detect GAN-images through",
"from the target dataset.",
"reconstruction error. This learning method has advantages",
"In addition, we introduce a novel augmentation method to",
"regarding lower data usage, when the model is well trained.",
"solve the over-fitting problem by transforming Cutmix (Yun",
"AlthoughtheForensicTransfershowedpromiseformodel",
"et al., 2019), which randomly mixes up a rectangle patch",
"transferability, its performance remains mediocre. In pre-",
"of training images. Note that the original Cutmix mixes up",
"vious research, either artifacts, patterns, or augmentations",
"inter-class images; our experiment showed that this renders",
"were utilized individually for successful transfer learning,",
"the learning process highly unstable for a binary classi-",
"yet it is possible to combine them to transfer knowledge of",
"fier, and thus we transform the inter-class Cutmix to an",
"GAN-image detection.",
"intra-class Cutmix to increase the stability of the learning",
"In this paper, our objective is to maintain a high detec-",
"process. Also, our method combines Gaussian blur (Xuan",
"tion performance during transfer learning on the source",
"et al., 2019) and Joint Photograph Experts Group (JPEG)",
"and target datasets, without suffering catastrophic forget-",
"compression (Wang et al., 2019), which were previously",
"ting. While many studies on transfer learning have already",
"studied for GAN-image detection.",
"shown impressive performance, they have not applied for",
"GAN-image detection. Therefore, in this work, we pro-",
"For transfer learning, we apply learning strategies, such as",
"Weight Standardization (Qiao et al., 2019) (WS), Group",
"pose a novel regularization method with self-training for",
"Normalization (Wu & He, 2018) (GN), and the tuning of",
"transfer learning by combining and transforming regular-",
"learning rate and momentum rate (Li et al., 2020). WS and",
"ization, augmentation, self-training, and learning strategies",
"GN achieved comparable performance on image classifica-",
"toimprovetransferabilityofGAN-imagedetection.Inpar",
"tion, not depending on batch size statistics. Also, we imple-",
"ticular, our approach is inspired by starting point as the",
"reference (SPAR), L?-SP (Li et al., 2018), which regular-",
"ment transfer learning using low learning and momentum",
"rates for stochastic gradient descent (SGD) inspired by (Li",
"izestheweightvariationofatargetmodelbyreferringto",
"et al., 2020). These have been experimented on object detec-",
"the weights pre-trained on the source dataset. The limitation",
"tion transfer, but we demonstrate that these strategies also",
"of the latter method is that it cannot provide an optimal",
"workwellfortransferlearningacrossdifferentdomains.Fi-"
] |
{
"arxiv_id": "2008.04115",
"language": "en",
"page_idx": 1
}
|
||
2008.04115-en-2
|
2008.04115-en-2.jpg
|
T-GD: Transferable GAN-generated Images Detection Framework
nally, we integrate these approaches into one framework, the Transferable GAN-generated image Detection framework (T-GD). Compared to general methods of transfer learning and recent GAN-image detection, we show that T-GD is equipped with robust transferability and achieves high performance.
## $2. \textbf{Related Work}$
GAN-image detection. Widely used approaches for the detection of GAN-images include the addition of a learning method, the transformation of GAN-images, the application of data augmentation, and the use of metadata. Another technique, based on multi-task incremental learning (Marra et al., 2019b), shows great promise for transferability within different types of GAN-images, changing the existing learning method (autoencoder) and loss function (incremental learning). However, a lingering issue is the need for a large amount of data; our work directly alleviates this problem, maintaining the model performance with a smaller amount of data. (Nataraj et al., 2019) proposed using a combination of co-occurrence matrices and JPEG compression to transform GAN-images for data augmentation. Co-occurrence matrices are extracted from three color channels in the pixel domain following JPEG compression and are used to train the CNN. In their approach, the JPEG compression contributed to the improvement of performance, but the transformation of the input data into co-occurrence matrices caused over-fitting and reduced generalizability.
Some methods identify a unique artifact spectrum caused by the up-sampling component (Zhang et al., 2019), while others use the photo-response non-uniformity (PRNU) pattern as the input of CNN classifiers (Marra et al., 2019a). Augmentation techniques requiring domain knowledge of GAN-image detection, such as Gaussian blur and Gaussian noise (Xuan et al., 2019), were also studied and the combination of Gaussian blur and JPEG compression was shown to achieve high performance (Wang et al., 2019). However, employing a single data augmentation method achieved limited transfer learning performance. Therefore, we use the combination of JPEG compression and Gaussian blur to achieve better transfer learning results. Similar to prior digital fingerprint techniques, GAN fingerprints (i.e., image and model fingerprints) are used to differentiate real and GAN-images using metadata (Yu et al., 2019). They assume white-box attack scenarios, where detectors possess knowledge of the data and the model (metadata information) of attackers. Our approach differs from theirs, where ours is constrained to $a$ non-adaptive black-box approach scenario; detectors only possess knowledge of the training data.
Transfer learning. Several transfer learning methods have been explored in terms of meta-learning, self-supervised learning, domain adaptation (Zamir et al., 2018), knowl-
edge distillation, and continual learning. In particular, starting point as the reference (SPAR) (Li et al.,2018) and Deep Learning Transfer using Feature Map with Attention $(\hat{\text{DELTA) (Li et al., 2019) use }L^{2}}\hat{\text{regularization as a starting}}$ point to maintain the source dataset as the inductive learning method during domain adaptation. The difference from ours is that we use self-training to control the regularization effect, which has the advantage of preventing either excessive or minor regularization.
(Yim et al., 2017) proposed two additional layers to calculate the flow of solution procedure matrix for knowledge distillation, but our T-GD shows robust performance in transfer learning without these layers and expensive computation. Learning with continual tasks (Zenke et al., 2017), where the node weight is regularized based on the importance of previous tasks, is similar to our method. However, we provide the following differences: First, the source and target datasets differ in size (see Table 2), resulting in relatively low computational cost in $L^2$-$SP.$ Second, in our task, achieving generalizability using a small amount of target data is as essential as the prevention of forgetting; to address this trade-off, we chose $L^2$-$SP$ using all weights. Third, unlike continual learning on each independent task, our T-GD focuses on transfer learning within the GAN-image domains.
Self-training. Self-training methods (Yalniz et al., 2019;
Xie et al., 2019) were used to increase the state-of-the-art top-1 accuracy of ImageNet (Russakovsky et al., 2015). The difference from our work is that our objective is to increase transferability and not the single model performance. We use a teacher-student structure, and inject noise to the student model and the input data to prevent over-fitting by effective techniques, such as dropout (Srivastava et al., 2014), stochastic depth (Huang et al., 2016), and intra-class Cutmix (data augmentation)). Also, self-training was used for domain adaptation by (RoyChowdhury et al., 2019). Their self-training model is video-specific, applying a teacher model to the target domain, which is different from our image classification task.
## 3.0ur Method
The first step of T-GD is to train the pre-trained models, namely the binary classifiers predicting whether an image is GAN-generated or not.
CNN binary classifier. We chose CNN binary classifiers as classifers for the source dataset. This choice has three advantages: (1) it is easy to reuse pre-trained models, (2) many pre-studied CNN architectures can be utilized, and (3) it shows a more stable performance in binary classification than other methods such as autoencoders. We pre-train the CNN binary classifier on the source dataset and trans-
|
[
[
0.27783203125,
0.05682373046875,
0.6953125,
0.05682373046875,
0.6953125,
0.0706787109375,
0.27783203125,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.4755859375,
0.08587646484375,
0.4755859375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.498291015625,
0.085205078125,
0.8837890625,
0.085205078125,
0.8837890625,
0.0985107421875,
0.498291015625,
0.0985107421875
],
[
0.08905029296875,
0.1029052734375,
0.473876953125,
0.1029052734375,
0.473876953125,
0.11297607421875,
0.08905029296875,
0.11297607421875
],
[
0.498291015625,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.498291015625,
0.11553955078125
],
[
0.08905029296875,
0.11871337890625,
0.473876953125,
0.11871337890625,
0.473876953125,
0.1287841796875,
0.08905029296875,
0.1287841796875
],
[
0.5009765625,
0.11871337890625,
0.8837890625,
0.11871337890625,
0.8837890625,
0.1287841796875,
0.5009765625,
0.1287841796875
],
[
0.0882568359375,
0.1319580078125,
0.4755859375,
0.1319580078125,
0.4755859375,
0.1458740234375,
0.0882568359375,
0.1458740234375
],
[
0.5,
0.1307373046875,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.5,
0.14453125
],
[
0.0882568359375,
0.147705078125,
0.4755859375,
0.147705078125,
0.4755859375,
0.1610107421875,
0.0882568359375,
0.1610107421875
],
[
0.5,
0.146484375,
0.8857421875,
0.146484375,
0.8857421875,
0.1610107421875,
0.5,
0.1610107421875
],
[
0.08905029296875,
0.1634521484375,
0.1568603515625,
0.1634521484375,
0.1568603515625,
0.1749267578125,
0.08905029296875,
0.1749267578125
],
[
0.5009765625,
0.1622314453125,
0.8857421875,
0.1622314453125,
0.8857421875,
0.175537109375,
0.5009765625,
0.175537109375
],
[
0.50244140625,
0.1787109375,
0.8857421875,
0.1787109375,
0.8857421875,
0.189453125,
0.50244140625,
0.189453125
],
[
0.498291015625,
0.1900634765625,
0.88623046875,
0.1912841796875,
0.88623046875,
0.205810546875,
0.498291015625,
0.2052001953125
],
[
0.08905029296875,
0.1962890625,
0.226318359375,
0.1962890625,
0.226318359375,
0.210205078125,
0.08905029296875,
0.210205078125
],
[
0.5,
0.20703125,
0.65869140625,
0.20703125,
0.65869140625,
0.2203369140625,
0.5,
0.2203369140625
],
[
0.08905029296875,
0.2215576171875,
0.473876953125,
0.2215576171875,
0.473876953125,
0.23486328125,
0.08905029296875,
0.23486328125
],
[
0.08660888671875,
0.23486328125,
0.473876953125,
0.2374267578125,
0.473876953125,
0.251953125,
0.08660888671875,
0.25
],
[
0.5,
0.2291259765625,
0.88623046875,
0.2291259765625,
0.88623046875,
0.242431640625,
0.5,
0.242431640625
],
[
0.5009765625,
0.2442626953125,
0.88623046875,
0.2442626953125,
0.88623046875,
0.25830078125,
0.5009765625,
0.25830078125
],
[
0.0882568359375,
0.251953125,
0.473876953125,
0.251953125,
0.473876953125,
0.26513671875,
0.0882568359375,
0.26513671875
],
[
0.5,
0.260009765625,
0.88623046875,
0.260009765625,
0.88623046875,
0.2734375,
0.5,
0.2734375
],
[
0.0882568359375,
0.267578125,
0.473876953125,
0.267578125,
0.473876953125,
0.281005859375,
0.0882568359375,
0.281005859375
],
[
0.498291015625,
0.2734375,
0.88623046875,
0.274658203125,
0.88623046875,
0.2890625,
0.498291015625,
0.28857421875
],
[
0.0882568359375,
0.28271484375,
0.473876953125,
0.28271484375,
0.473876953125,
0.296630859375,
0.0882568359375,
0.296630859375
],
[
0.5,
0.290283203125,
0.8857421875,
0.290283203125,
0.8857421875,
0.3037109375,
0.5,
0.3037109375
],
[
0.08905029296875,
0.29736328125,
0.473876953125,
0.29736328125,
0.473876953125,
0.3125,
0.08905029296875,
0.3125
],
[
0.496826171875,
0.3037109375,
0.88623046875,
0.304931640625,
0.88623046875,
0.3193359375,
0.496826171875,
0.318115234375
],
[
0.08905029296875,
0.313232421875,
0.473876953125,
0.313232421875,
0.473876953125,
0.326904296875,
0.08905029296875,
0.326904296875
],
[
0.5,
0.32080078125,
0.88623046875,
0.32080078125,
0.88623046875,
0.333984375,
0.5,
0.333984375
],
[
0.08905029296875,
0.32763671875,
0.473876953125,
0.32763671875,
0.473876953125,
0.341552734375,
0.08905029296875,
0.341552734375
],
[
0.5009765625,
0.335205078125,
0.88623046875,
0.335205078125,
0.88623046875,
0.348388671875,
0.5009765625,
0.348388671875
],
[
0.0882568359375,
0.343505859375,
0.473876953125,
0.343505859375,
0.473876953125,
0.356689453125,
0.0882568359375,
0.356689453125
],
[
0.498291015625,
0.349853515625,
0.8837890625,
0.349853515625,
0.8837890625,
0.363037109375,
0.498291015625,
0.363037109375
],
[
0.0882568359375,
0.35791015625,
0.473876953125,
0.35791015625,
0.473876953125,
0.371826171875,
0.0882568359375,
0.371826171875
],
[
0.498291015625,
0.365478515625,
0.88623046875,
0.365478515625,
0.88623046875,
0.37890625,
0.498291015625,
0.37890625
],
[
0.0882568359375,
0.373779296875,
0.473876953125,
0.373779296875,
0.473876953125,
0.386962890625,
0.0882568359375,
0.386962890625
],
[
0.498291015625,
0.380126953125,
0.8857421875,
0.380126953125,
0.8857421875,
0.393310546875,
0.498291015625,
0.393310546875
],
[
0.08660888671875,
0.38818359375,
0.4755859375,
0.38818359375,
0.4755859375,
0.401611328125,
0.08660888671875,
0.401611328125
],
[
0.5,
0.39453125,
0.8837890625,
0.39453125,
0.8837890625,
0.40771484375,
0.5,
0.40771484375
],
[
0.0882568359375,
0.40283203125,
0.473876953125,
0.40283203125,
0.473876953125,
0.416748046875,
0.0882568359375,
0.416748046875
],
[
0.5009765625,
0.410400390625,
0.88623046875,
0.410400390625,
0.88623046875,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.0882568359375,
0.41845703125,
0.47216796875,
0.41845703125,
0.47216796875,
0.431884765625,
0.0882568359375,
0.431884765625
],
[
0.5,
0.425537109375,
0.8857421875,
0.425537109375,
0.8857421875,
0.439453125,
0.5,
0.439453125
],
[
0.08905029296875,
0.43505859375,
0.47216796875,
0.43505859375,
0.47216796875,
0.44580078125,
0.08905029296875,
0.44580078125
],
[
0.5,
0.44140625,
0.8857421875,
0.44140625,
0.8857421875,
0.455078125,
0.5,
0.455078125
],
[
0.0882568359375,
0.448974609375,
0.47216796875,
0.448974609375,
0.47216796875,
0.462158203125,
0.0882568359375,
0.462158203125
],
[
0.08660888671875,
0.46337890625,
0.4755859375,
0.462158203125,
0.4755859375,
0.4765625,
0.08660888671875,
0.477783203125
],
[
0.5,
0.458251953125,
0.56201171875,
0.458251953125,
0.56201171875,
0.4697265625,
0.5,
0.4697265625
],
[
0.08905029296875,
0.479248046875,
0.47802734375,
0.479248046875,
0.47802734375,
0.492431640625,
0.08905029296875,
0.492431640625
],
[
0.5,
0.477783203125,
0.8857421875,
0.477783203125,
0.8857421875,
0.492431640625,
0.5,
0.492431640625
],
[
0.08905029296875,
0.49560546875,
0.473876953125,
0.49560546875,
0.473876953125,
0.505859375,
0.08905029296875,
0.505859375
],
[
0.498291015625,
0.49365234375,
0.8857421875,
0.49365234375,
0.8857421875,
0.5068359375,
0.498291015625,
0.5068359375
],
[
0.0882568359375,
0.51025390625,
0.40283203125,
0.51025390625,
0.40283203125,
0.5234375,
0.0882568359375,
0.5234375
],
[
0.5,
0.5087890625,
0.88623046875,
0.5087890625,
0.88623046875,
0.52294921875,
0.5,
0.52294921875
],
[
0.5,
0.5234375,
0.88623046875,
0.5234375,
0.88623046875,
0.537109375,
0.5,
0.537109375
],
[
0.08905029296875,
0.5322265625,
0.4755859375,
0.5322265625,
0.4755859375,
0.54541015625,
0.08905029296875,
0.54541015625
],
[
0.5,
0.5390625,
0.88623046875,
0.5390625,
0.88623046875,
0.55322265625,
0.5,
0.55322265625
],
[
0.08660888671875,
0.546875,
0.473876953125,
0.54541015625,
0.473876953125,
0.56005859375,
0.08660888671875,
0.56103515625
],
[
0.498291015625,
0.5537109375,
0.88623046875,
0.5537109375,
0.88623046875,
0.5673828125,
0.498291015625,
0.5673828125
],
[
0.08660888671875,
0.5625,
0.4755859375,
0.56005859375,
0.4755859375,
0.57373046875,
0.08660888671875,
0.57568359375
],
[
0.5009765625,
0.57080078125,
0.8857421875,
0.57080078125,
0.8857421875,
0.58056640625,
0.5009765625,
0.58056640625
],
[
0.0882568359375,
0.5771484375,
0.473876953125,
0.5771484375,
0.473876953125,
0.59033203125,
0.0882568359375,
0.59033203125
],
[
0.5,
0.583984375,
0.88623046875,
0.583984375,
0.88623046875,
0.59765625,
0.5,
0.59765625
],
[
0.08905029296875,
0.59326171875,
0.47216796875,
0.59326171875,
0.47216796875,
0.60400390625,
0.08905029296875,
0.60400390625
],
[
0.5,
0.599609375,
0.88623046875,
0.599609375,
0.88623046875,
0.61279296875,
0.5,
0.61279296875
],
[
0.0882568359375,
0.60693359375,
0.4755859375,
0.60693359375,
0.4755859375,
0.62060546875,
0.0882568359375,
0.62060546875
],
[
0.5009765625,
0.6162109375,
0.8857421875,
0.6162109375,
0.8857421875,
0.626953125,
0.5009765625,
0.626953125
],
[
0.0882568359375,
0.62255859375,
0.473876953125,
0.62255859375,
0.473876953125,
0.63623046875,
0.0882568359375,
0.63623046875
],
[
0.5,
0.6298828125,
0.8857421875,
0.6298828125,
0.8857421875,
0.64306640625,
0.5,
0.64306640625
],
[
0.0882568359375,
0.63720703125,
0.4755859375,
0.63720703125,
0.4755859375,
0.65087890625,
0.0882568359375,
0.65087890625
],
[
0.498291015625,
0.64453125,
0.8857421875,
0.64453125,
0.8857421875,
0.65771484375,
0.498291015625,
0.65771484375
],
[
0.0882568359375,
0.6513671875,
0.4755859375,
0.6513671875,
0.4755859375,
0.66552734375,
0.0882568359375,
0.66552734375
],
[
0.5,
0.66015625,
0.88623046875,
0.66015625,
0.88623046875,
0.673828125,
0.5,
0.673828125
],
[
0.0882568359375,
0.66650390625,
0.473876953125,
0.66748046875,
0.473876953125,
0.681640625,
0.0882568359375,
0.68115234375
],
[
0.498291015625,
0.67529296875,
0.66845703125,
0.67529296875,
0.66845703125,
0.689453125,
0.498291015625,
0.689453125
],
[
0.0882568359375,
0.68310546875,
0.473876953125,
0.68310546875,
0.473876953125,
0.6962890625,
0.0882568359375,
0.6962890625
],
[
0.0882568359375,
0.69775390625,
0.473876953125,
0.69775390625,
0.473876953125,
0.7109375,
0.0882568359375,
0.7109375
],
[
0.08905029296875,
0.7138671875,
0.4755859375,
0.7138671875,
0.4755859375,
0.724609375,
0.08905029296875,
0.724609375
],
[
0.498291015625,
0.70849609375,
0.6259765625,
0.70849609375,
0.6259765625,
0.72265625,
0.498291015625,
0.72265625
],
[
0.08905029296875,
0.72802734375,
0.4755859375,
0.72802734375,
0.4755859375,
0.7412109375,
0.08905029296875,
0.7412109375
],
[
0.498291015625,
0.73486328125,
0.88623046875,
0.73486328125,
0.88623046875,
0.74853515625,
0.498291015625,
0.74853515625
],
[
0.08905029296875,
0.74365234375,
0.473876953125,
0.74365234375,
0.473876953125,
0.7568359375,
0.08905029296875,
0.7568359375
],
[
0.5,
0.75048828125,
0.88623046875,
0.75048828125,
0.88623046875,
0.7646484375,
0.5,
0.7646484375
],
[
0.0882568359375,
0.75830078125,
0.473876953125,
0.75830078125,
0.473876953125,
0.771484375,
0.0882568359375,
0.771484375
],
[
0.5,
0.76513671875,
0.65625,
0.76513671875,
0.65625,
0.77880859375,
0.5,
0.77880859375
],
[
0.0882568359375,
0.7734375,
0.473876953125,
0.7734375,
0.473876953125,
0.787109375,
0.0882568359375,
0.787109375
],
[
0.0882568359375,
0.78857421875,
0.4755859375,
0.78857421875,
0.4755859375,
0.8017578125,
0.0882568359375,
0.8017578125
],
[
0.5,
0.787109375,
0.88623046875,
0.787109375,
0.88623046875,
0.8017578125,
0.5,
0.8017578125
],
[
0.0882568359375,
0.8037109375,
0.473876953125,
0.8037109375,
0.473876953125,
0.8173828125,
0.0882568359375,
0.8173828125
],
[
0.5,
0.8046875,
0.8837890625,
0.8046875,
0.8837890625,
0.81494140625,
0.5,
0.81494140625
],
[
0.0882568359375,
0.818359375,
0.4755859375,
0.818359375,
0.4755859375,
0.83203125,
0.0882568359375,
0.83203125
],
[
0.5,
0.818359375,
0.8857421875,
0.818359375,
0.8857421875,
0.83203125,
0.5,
0.83203125
],
[
0.0882568359375,
0.833984375,
0.441162109375,
0.833984375,
0.441162109375,
0.84765625,
0.0882568359375,
0.84765625
],
[
0.5,
0.833984375,
0.8857421875,
0.833984375,
0.8857421875,
0.84765625,
0.5,
0.84765625
],
[
0.08905029296875,
0.85595703125,
0.473876953125,
0.85595703125,
0.473876953125,
0.869140625,
0.08905029296875,
0.869140625
],
[
0.50244140625,
0.85107421875,
0.8857421875,
0.85107421875,
0.8857421875,
0.861328125,
0.50244140625,
0.861328125
],
[
0.5,
0.8642578125,
0.88623046875,
0.8642578125,
0.88623046875,
0.87744140625,
0.5,
0.87744140625
],
[
0.0882568359375,
0.8720703125,
0.473876953125,
0.8720703125,
0.473876953125,
0.88623046875,
0.0882568359375,
0.88623046875
],
[
0.5,
0.8798828125,
0.88623046875,
0.8798828125,
0.88623046875,
0.89306640625,
0.5,
0.89306640625
],
[
0.0882568359375,
0.88623046875,
0.4755859375,
0.88623046875,
0.4755859375,
0.90087890625,
0.0882568359375,
0.90087890625
]
] |
[
"T-GD: Transferable GAN-generated Images Detection Framework",
"nally,we integrate these approaches into one framework, the",
"edge distillation, and continual learning. In particular, start",
"TransferableGAN-generatedimageDetectionframework",
"ing point as the reference (SPAR) (Li et al., 2018) and",
"(T-GD).Compared togeneralmethods oftransferlearning",
"DeepLearningTransferusingFeatureMapwithAttention",
"and recent GAN-image detection, we show that T-GD is",
"(DELTA) (Li et al., 2019) use L? regularization as a starting",
"equipped with robust transferability and achieves high per-",
"point to maintain the source dataset as the inductive learning",
"formance.",
"method during domain adaptation. The difference from ours",
"is that weuse self-trainingto control theregularization ef-",
"fect, which has the advantage of preventing either excessive",
"2. Related Work",
"or minor regularization.",
"GAN-image detection.Widely used approaches for the",
"detection of GAN-images include the addition of a learning",
"(Yim et al., 2017) proposed two additional layers to calcu-",
"late the flow of solution procedure matrix for knowledge",
"method, the transformation of GAN-images, the application",
"distillation, but our T-GD shows robust performance in trans-",
"of data augmentation, and the use of metadata. Another",
"fer learning without these layers and expensive computation.",
"technique, based on multi-task incremental learning (Marra",
"Learning with continual tasks (Zenke et al., 2017), where",
"et al., 2019b), shows great promise for transferability within",
"the node weight is regularized based on the importance of",
"different types of GAN-images, changing the existing learn",
"previous tasks, is similar to our method. However, we pro-",
"ing method (autoencoder) and loss function (incremental",
"vide the following differences: First, the source and target",
"learning). However, a lingering issue is the need for a large",
"datasets differ in size (see Table 2), resulting in relatively",
"amount of data; our work directly alleviates this problem,",
"low computational cost in L?-SP. Second, in our task,",
"maintaining the model performance with a smaller amount",
"achieving generalizability using a small amount of target",
"of data. (Nataraj et al., 2019) proposed using a combination",
"data is as essential as the prevention of forgetting; to address",
"of co-occurrence matrices and JPEG compression to trans-",
"this trade-off, we chose L?-SP using all weights. Third,",
"form GAN-images for data augmentation. Co-occurrence",
"unlike continual learning on each independent task, our",
"matricesareextractedfromthree color channels inthepixel",
"T-GD focuses on transfer learning within the GAN-image",
"domain following JPEG compression and are used to train",
"the CNN. In their approach, the JPEG compression con-",
"domains.",
"tributed to the improvement of performance, but the trans-",
"Self-training. Self-training methods (Yalniz et al., 2019;",
"formation of the input data into co-occurrence matrices",
"Xie et al., 2019) were used to increase the state-of-the-art",
"caused over-fitting and reduced generalizability.",
"top-1 accuracy of ImageNet (Russakovsky et al., 2015). The",
"difference from our work is that our objective is to increase",
"Some methods identify a unique artifact spectrum caused",
"transferability and not the single model performance. We",
"by the up-sampling component (Zhang et al., 2019), while",
"use a teacher-student structure, and inject noise to the stu-",
"othersusethephoto-responsenon-uniformity(PRNU)pat-",
"dentmodel and the inputdatatoprevent over-fitting byef-",
"tern as the input of CNN classifiers (Marra et al.. 2019a)",
"fective techniques, such as dropout (Srivastava et al., 2014),",
"Augmentationtechniquesrequiringdomainknowledgeof",
"stochastic depth (Huang et al., 2016), and intra-class Cut-",
"GAN-image detection, such as Gaussian blur and Gaus-",
"mix (data augmentation)).Also, self-training was used for",
"sian noise (Xuan et al., 2019), were also studied and the",
"domain adaptation by (RoyChowdhury et al., 2019). Their",
"combination of Gaussian blur and JPEG compression was",
"self-training model is video-specific, applying a teacher",
"shown to achieve high performance (Wang et al., 2019).",
"model to the target domain, which is different from our",
"However, employing a single data augmentation method",
"image classification task.",
"achieved limited transfer learning performance.Therefore,",
"we use the combination of JPEG compression and Gaussian",
"blurtoachieve bettertransferlearning results.Similarto",
"3. Our Method",
"prior digital fingerprint techniques, GAN fingerprints (i.e.,",
"The first step of T-GD is to train the pre-trained models,",
"image and model fingerprints) are used to differentiate real",
"namely the binary classifiers predicting whether an image is",
"and GAN-images using metadata (Yu et al., 2019). They",
"GAN-generated or not.",
"assume white-box attack scenarios, where detectors possess",
"knowledge of the data and the model (metadata information)",
"CNN binary classifier. We chose CNN binary classifiers",
"of attackers. Our approach differs from theirs, where ours is",
"as classifiers forthe sourcedataset.This choicehas three",
"constrained to a non-adaptive black-box approach scenario;",
"advantages: (1) it is easy to reuse pre-trained models, (2)",
"detectors only possess knowledge of the training data.",
"many pre-studied CNN architectures can be utilized, and",
"Transfer learning.Several transfer learning methods have",
"(3)itshowsamorestableperformanceinbinaryclassifica-",
"tion than other methods such as autoencoders. We pre-train",
"been explored in terms of meta-learning, self-supervised",
"the CNN binary classifier on the source dataset and trans-",
"learning, domain adaptation (Zamir et al., 2018), knowl-"
] |
{
"arxiv_id": "2008.04115",
"language": "en",
"page_idx": 2
}
|
||
2008.04115-en-3
|
2008.04115-en-3.jpg
|
T-GD: Transferable GAN-generated Images Detection Framework
fer (fine-tune) this pre-trained model to the target dataset. For instance, EfficientNet (the CNN classifier) is trained on the PGGAN-dataset (the source) and fine-tuned on the StyleGAN-dataset (the target).
EfficientNet. We implemented EfficientNet-B0 (Tan & Le, 2019) and used it as the CNN classifier. Although EfficientNet-B0 has the lowest number of parameters (about four million) among the EfficientNets, it performs well in GAN-image detection in our experiment, compared to Inception-V3 (Szegedy et al., 2017) and Xception (Chollet, 2017). Another change we make to the model is the use of WS (Qiao et al., 2019) and GN (Wu &He, 2018), instead of batch normalization (BN) (Ioffe&Szegedy,2015), due to their superior efficiency regarding transfer learning to that of batch statistics.
ResNext. We implemented ResNext32×4d (Xie et al., 2017) and used it as the CNN classifier, where ResNext32×4d has more parameters (about twenty million) than EfficientNet-B0. We also replace BN with WS and GN.
## $3.1.L^{2}.SP$
The next step is transfer learning. The weight of the pretrained model from the source dataset is used as the SPAR. In particular, we use $L^2$-$SP$ for transfer learning. Regularization can lead to a better optimization by preventing over-fitting when learning from scratch; $L^2$-$\tilde{SP}$ differs in that the starting point from a well pre-trained source dataset guides the learning process by referring to the information of the pre-trained source dataset. This method does not require freezing the weights of the pre-trained model nor using weight decay. Our method regularizes convolution layers and fully-connected (FC) layers independently.
General form of regularization. Let $w$ be the weight parameters, and $J(\hat{y}_i,y_i)$ be the loss function of the neural networks, where $\hat{y}_{i}$ is the $i^th$ score predicted by the models and $y_i$ is the $i^th$ label. And $\Omega(w)$ is the $p$-norm function of the weight $w$ as a general form of regularization loss, $f_w(x_i)$ is the neural network function with the $i^{th}$ data $x_i$, and $n$ is the dataset size. Equation 1 indicates the general form of the loss function with a weight regularization component:
(1)
$$\begin{aligned}&\min_{w}\frac{1}{n}\sum_{i=1}^{n}J(\hat{y_{i}},y_{i})+\lambda\cdot\Omega(w),\\&\hat{y_{i}}=f_{w}(x_{i}),\end{aligned}$$
where $\lambda$ balances the regularization and the loss function, $J$ is the cross-entropy function, and $\Omega$ is the $L^1$ or $L^2$-norm of the parameter $w.$
$L^2$ regularization. $L^2$ regularization is used in transfer learning to avoid over-fitting and to overcome the forgetting
of the learned information or catastrophic forgetting.
$$\Omega_{l2}(w)=\|w\|_2^2.$$
(2)
Equation 2 is the $\Omega$ function, namely the $L^2$-norm of $w.$
$$\begin{aligned}\min_w\frac{1}{n}\sum_{i=1}^nJ(\hat{y}_i,y_i)+\beta\cdot\Omega_{l2}(w_{fc}).\end{aligned}$$
(3)
In Eq. 3, the first term is the same as in Eq. 1, representing the cross-entropy loss function. The second term is the $\Omega_l2$ function or the $L^2$ regularization term (Eq. 2) of $w_{fc}$,the weights of the FC layers, scaled by $\beta$,which is equivalent to $\lambda$ in Eq. 1. Note that the $L^2$ regularization is applied solely to the FC layers since over-fitting and forgetting are delayed, but not completely prevented in the course of learning.
$L^2$-$SP.$ Let $w^\prime$ be the pre-trained weights from the source dataset, as shown in Section 3.1, serving as the starting point (SP) as the reference, as well as a regularization point that provides guidance for transfer learning when fine-tuning. ${\hat{\text{Using L2-norm, we define }L^{2}-SP{\mathrm{~as~follows:}}}}$
$$\Omega_{sp}(w,w')=\|w_{conv}-w'_{conv}\|_2^2,$$
(4)
where $w_conv$ denotes the weights of the convolution layers, excluding those of the FC layers. Equation 4 indicates that $L^{2}$-$SP$ is a one-to-one mapping between the convolution layers of the source and target datasets, e.g., the PGGANclassifier (the source) to the StyleGAN-classifier (the target). Loss function. We combine Eq. 4, sharing the architecture of the source and target models, with the second term of Eq. 3, accounting for the FC layer (final layer) as follows:
$$\operatorname*{min}_{w}\frac{1}{n}\sum_{i=1}^{n}J(\hat{y_{i}},y_{i})+\alpha\cdot\Omega_{sp}(w,w^{\prime})+\beta\cdot\Omega_{l2}(w_{fc}),\\J(\hat{y_{i}},y_{i})=-y_{i}\log(\hat{y_{i}})-(1-y_{i})\log(1-\hat{y_{i}}),$$
(5)
where $J$ is the negative log-likelihood loss function, and $\alpha$ and $\beta$ are tunable hyperparameters of which (Li et al., 2019) use values in the range from 0.1 to 0.01. The difference from $L^2$-$SP$ is that we transform $\alpha$ and $\beta$ into $\gamma$, a parameter which adjusts itself according to the learning situation. More details about the transformed parameters are provided in Section 3.3.
# 3.2. Self-training for $L^2$-$SP$
We transform the transfer learning framework into a selftraining framework. In other words, the source/target model is changed into a teacher/student model. In addition to the role of a typical source model, which serves as SPAR and a regularizer to guide the learning process, the teacher model has the role of adjusting the parameters based on the learned
|
[
[
0.27783203125,
0.058074951171875,
0.6953125,
0.058074951171875,
0.6953125,
0.06817626953125,
0.27783203125,
0.06817626953125
],
[
0.08905029296875,
0.08587646484375,
0.47802734375,
0.08587646484375,
0.47802734375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.496826171875,
0.083984375,
0.84619140625,
0.08587646484375,
0.84619140625,
0.10101318359375,
0.496826171875,
0.0985107421875
],
[
0.08905029296875,
0.10101318359375,
0.473876953125,
0.10101318359375,
0.473876953125,
0.1142578125,
0.08905029296875,
0.1142578125
],
[
0.6337890625,
0.11236572265625,
0.74609375,
0.111083984375,
0.74609375,
0.126953125,
0.6337890625,
0.1275634765625
],
[
0.86279296875,
0.1142578125,
0.88623046875,
0.1142578125,
0.88623046875,
0.1275634765625,
0.86279296875,
0.1275634765625
],
[
0.0882568359375,
0.1319580078125,
0.2900390625,
0.1319580078125,
0.2900390625,
0.1458740234375,
0.0882568359375,
0.1458740234375
],
[
0.08905029296875,
0.11614990234375,
0.473876953125,
0.11614990234375,
0.473876953125,
0.1300048828125,
0.08905029296875,
0.1300048828125
],
[
0.5009765625,
0.14013671875,
0.865234375,
0.14013671875,
0.865234375,
0.154052734375,
0.5009765625,
0.154052734375
],
[
0.08905029296875,
0.154052734375,
0.473876953125,
0.154052734375,
0.473876953125,
0.1673583984375,
0.08905029296875,
0.1673583984375
],
[
0.0882568359375,
0.169189453125,
0.473876953125,
0.169189453125,
0.473876953125,
0.18310546875,
0.0882568359375,
0.18310546875
],
[
0.60302734375,
0.17236328125,
0.62158203125,
0.1685791015625,
0.62744140625,
0.1817626953125,
0.60888671875,
0.186279296875
],
[
0.62841796875,
0.169189453125,
0.63818359375,
0.169189453125,
0.63818359375,
0.1749267578125,
0.62841796875,
0.1749267578125
],
[
0.61669921875,
0.1767578125,
0.81640625,
0.1749267578125,
0.81640625,
0.1932373046875,
0.61669921875,
0.1956787109375
],
[
0.62353515625,
0.177978515625,
0.63232421875,
0.177978515625,
0.63232421875,
0.1817626953125,
0.62353515625,
0.1817626953125
],
[
0.86376953125,
0.177978515625,
0.88623046875,
0.177978515625,
0.88623046875,
0.1912841796875,
0.86376953125,
0.1912841796875
],
[
0.08905029296875,
0.184326171875,
0.473876953125,
0.184326171875,
0.473876953125,
0.1976318359375,
0.08905029296875,
0.1976318359375
],
[
0.580078125,
0.1787109375,
0.611328125,
0.1787109375,
0.611328125,
0.1912841796875,
0.580078125,
0.1912841796875
],
[
0.61279296875,
0.1875,
0.619140625,
0.1875,
0.619140625,
0.1944580078125,
0.61279296875,
0.1944580078125
],
[
0.0882568359375,
0.199462890625,
0.473876953125,
0.199462890625,
0.473876953125,
0.21337890625,
0.0882568359375,
0.21337890625
],
[
0.62353515625,
0.1956787109375,
0.6435546875,
0.1956787109375,
0.6435546875,
0.203857421875,
0.62353515625,
0.203857421875
],
[
0.08905029296875,
0.2139892578125,
0.473876953125,
0.2139892578125,
0.473876953125,
0.2279052734375,
0.08905029296875,
0.2279052734375
],
[
0.498291015625,
0.21533203125,
0.8857421875,
0.216552734375,
0.8857421875,
0.2310791015625,
0.498291015625,
0.2298583984375
],
[
0.08905029296875,
0.2298583984375,
0.4755859375,
0.2298583984375,
0.4755859375,
0.24365234375,
0.08905029296875,
0.24365234375
],
[
0.498291015625,
0.231689453125,
0.88232421875,
0.231689453125,
0.88232421875,
0.24560546875,
0.498291015625,
0.24560546875
],
[
0.08905029296875,
0.2442626953125,
0.473876953125,
0.2442626953125,
0.473876953125,
0.25830078125,
0.08905029296875,
0.25830078125
],
[
0.498291015625,
0.24560546875,
0.8857421875,
0.246826171875,
0.8857421875,
0.261474609375,
0.498291015625,
0.260009765625
],
[
0.08905029296875,
0.261474609375,
0.473876953125,
0.261474609375,
0.473876953125,
0.271484375,
0.08905029296875,
0.271484375
],
[
0.5,
0.261962890625,
0.8857421875,
0.261962890625,
0.8857421875,
0.27587890625,
0.5,
0.27587890625
],
[
0.0882568359375,
0.274658203125,
0.473876953125,
0.274658203125,
0.473876953125,
0.28857421875,
0.0882568359375,
0.28857421875
],
[
0.5,
0.276611328125,
0.8837890625,
0.276611328125,
0.8837890625,
0.290283203125,
0.5,
0.290283203125
],
[
0.0882568359375,
0.290283203125,
0.473876953125,
0.290283203125,
0.473876953125,
0.3037109375,
0.0882568359375,
0.3037109375
],
[
0.5,
0.292236328125,
0.88623046875,
0.292236328125,
0.88623046875,
0.30615234375,
0.5,
0.30615234375
],
[
0.0882568359375,
0.30615234375,
0.2099609375,
0.30615234375,
0.2099609375,
0.3193359375,
0.0882568359375,
0.3193359375
],
[
0.5,
0.306884765625,
0.85693359375,
0.306884765625,
0.85693359375,
0.3212890625,
0.5,
0.3212890625
],
[
0.0882568359375,
0.32763671875,
0.473876953125,
0.32763671875,
0.473876953125,
0.341552734375,
0.0882568359375,
0.341552734375
],
[
0.5,
0.32568359375,
0.88818359375,
0.32763671875,
0.88818359375,
0.34521484375,
0.5,
0.343505859375
],
[
0.0882568359375,
0.341552734375,
0.473876953125,
0.34228515625,
0.473876953125,
0.356689453125,
0.0882568359375,
0.35595703125
],
[
0.5,
0.343505859375,
0.88623046875,
0.3447265625,
0.88623046875,
0.359130859375,
0.5,
0.35791015625
],
[
0.08905029296875,
0.35791015625,
0.473876953125,
0.35791015625,
0.473876953125,
0.371826171875,
0.08905029296875,
0.371826171875
],
[
0.5009765625,
0.35986328125,
0.88623046875,
0.35986328125,
0.88623046875,
0.373779296875,
0.5009765625,
0.373779296875
],
[
0.0882568359375,
0.37255859375,
0.473876953125,
0.37255859375,
0.473876953125,
0.386474609375,
0.0882568359375,
0.386474609375
],
[
0.5009765625,
0.373779296875,
0.88623046875,
0.375,
0.88623046875,
0.389404296875,
0.5009765625,
0.38818359375
],
[
0.0882568359375,
0.389404296875,
0.1446533203125,
0.389404296875,
0.1446533203125,
0.40087890625,
0.0882568359375,
0.40087890625
],
[
0.5009765625,
0.39013671875,
0.8037109375,
0.39013671875,
0.8037109375,
0.404052734375,
0.5009765625,
0.404052734375
],
[
0.57861328125,
0.416748046875,
0.80615234375,
0.416748046875,
0.80615234375,
0.434326171875,
0.57861328125,
0.434326171875
],
[
0.86279296875,
0.417236328125,
0.88623046875,
0.417236328125,
0.88623046875,
0.43115234375,
0.86279296875,
0.43115234375
],
[
0.0882568359375,
0.443115234375,
0.473876953125,
0.443115234375,
0.473876953125,
0.45703125,
0.0882568359375,
0.45703125
],
[
0.08660888671875,
0.41845703125,
0.17236328125,
0.416748046875,
0.17236328125,
0.431884765625,
0.08660888671875,
0.434326171875
],
[
0.5,
0.4443359375,
0.88623046875,
0.4443359375,
0.88623046875,
0.458984375,
0.5,
0.458984375
],
[
0.0882568359375,
0.458251953125,
0.473876953125,
0.458251953125,
0.473876953125,
0.4716796875,
0.0882568359375,
0.4716796875
],
[
0.5009765625,
0.460205078125,
0.88623046875,
0.460205078125,
0.88623046875,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.0882568359375,
0.473388671875,
0.4755859375,
0.473388671875,
0.4755859375,
0.4873046875,
0.0882568359375,
0.4873046875
],
[
0.5,
0.473388671875,
0.8857421875,
0.474853515625,
0.8857421875,
0.4892578125,
0.5,
0.488037109375
],
[
0.0882568359375,
0.4892578125,
0.473876953125,
0.4892578125,
0.473876953125,
0.50244140625,
0.0882568359375,
0.50244140625
],
[
0.5009765625,
0.490478515625,
0.8857421875,
0.490478515625,
0.8857421875,
0.50390625,
0.5009765625,
0.50390625
],
[
0.0882568359375,
0.50244140625,
0.473876953125,
0.501953125,
0.473876953125,
0.5166015625,
0.0882568359375,
0.517578125
],
[
0.5009765625,
0.505859375,
0.88623046875,
0.505859375,
0.88623046875,
0.51953125,
0.5009765625,
0.51953125
],
[
0.0882568359375,
0.51806640625,
0.473876953125,
0.51806640625,
0.473876953125,
0.5322265625,
0.0882568359375,
0.5322265625
],
[
0.5,
0.52587890625,
0.8857421875,
0.52783203125,
0.8857421875,
0.54248046875,
0.5,
0.54052734375
],
[
0.08905029296875,
0.5341796875,
0.473876953125,
0.5341796875,
0.473876953125,
0.54736328125,
0.08905029296875,
0.54736328125
],
[
0.50244140625,
0.54443359375,
0.8857421875,
0.54443359375,
0.8857421875,
0.5546875,
0.50244140625,
0.5546875
],
[
0.08905029296875,
0.55029296875,
0.47216796875,
0.55029296875,
0.47216796875,
0.56103515625,
0.08905029296875,
0.56103515625
],
[
0.0882568359375,
0.564453125,
0.473876953125,
0.564453125,
0.473876953125,
0.57763671875,
0.0882568359375,
0.57763671875
],
[
0.5009765625,
0.5595703125,
0.88134765625,
0.5595703125,
0.88134765625,
0.57275390625,
0.5009765625,
0.57275390625
],
[
0.0882568359375,
0.580078125,
0.473876953125,
0.580078125,
0.473876953125,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.5498046875,
0.58837890625,
0.56640625,
0.59326171875,
0.560546875,
0.60302734375,
0.5439453125,
0.59765625
],
[
0.08905029296875,
0.5947265625,
0.439453125,
0.5947265625,
0.439453125,
0.60791015625,
0.08905029296875,
0.60791015625
],
[
0.51708984375,
0.5966796875,
0.55126953125,
0.5966796875,
0.55126953125,
0.609375,
0.51708984375,
0.609375
],
[
0.5810546875,
0.59521484375,
0.87109375,
0.59521484375,
0.87109375,
0.60986328125,
0.5810546875,
0.60986328125
],
[
0.5458984375,
0.60791015625,
0.56396484375,
0.60791015625,
0.56396484375,
0.6162109375,
0.5458984375,
0.6162109375
],
[
0.08905029296875,
0.6162109375,
0.47216796875,
0.6162109375,
0.47216796875,
0.6298828125,
0.08905029296875,
0.6298828125
],
[
0.560546875,
0.611328125,
0.5849609375,
0.611328125,
0.5849609375,
0.62353515625,
0.560546875,
0.62353515625
],
[
0.0882568359375,
0.6318359375,
0.473876953125,
0.630859375,
0.473876953125,
0.64599609375,
0.0882568359375,
0.646484375
],
[
0.5595703125,
0.626953125,
0.87109375,
0.626953125,
0.87109375,
0.6416015625,
0.5595703125,
0.6416015625
],
[
0.08905029296875,
0.646484375,
0.473876953125,
0.646484375,
0.473876953125,
0.6611328125,
0.08905029296875,
0.6611328125
],
[
0.86376953125,
0.64453125,
0.88623046875,
0.64453125,
0.88623046875,
0.6572265625,
0.86376953125,
0.6572265625
],
[
0.08905029296875,
0.662109375,
0.473876953125,
0.662109375,
0.473876953125,
0.67529296875,
0.08905029296875,
0.67529296875
],
[
0.5009765625,
0.67041015625,
0.8837890625,
0.67041015625,
0.8837890625,
0.6845703125,
0.5009765625,
0.6845703125
],
[
0.08905029296875,
0.67822265625,
0.473876953125,
0.67822265625,
0.473876953125,
0.69140625,
0.08905029296875,
0.69140625
],
[
0.0882568359375,
0.68994140625,
0.473876953125,
0.69140625,
0.473876953125,
0.70556640625,
0.0882568359375,
0.705078125
],
[
0.5009765625,
0.68505859375,
0.8857421875,
0.68505859375,
0.8857421875,
0.69873046875,
0.5009765625,
0.69873046875
],
[
0.5009765625,
0.70068359375,
0.8857421875,
0.70068359375,
0.8857421875,
0.7138671875,
0.5009765625,
0.7138671875
],
[
0.0882568359375,
0.70703125,
0.473876953125,
0.70703125,
0.473876953125,
0.72021484375,
0.0882568359375,
0.72021484375
],
[
0.5009765625,
0.7138671875,
0.8857421875,
0.71533203125,
0.8857421875,
0.72900390625,
0.5009765625,
0.72802734375
],
[
0.0882568359375,
0.72265625,
0.46728515625,
0.72265625,
0.46728515625,
0.736328125,
0.0882568359375,
0.736328125
],
[
0.5009765625,
0.73095703125,
0.8857421875,
0.73095703125,
0.8857421875,
0.744140625,
0.5009765625,
0.744140625
],
[
0.5,
0.744140625,
0.88623046875,
0.74560546875,
0.88623046875,
0.76025390625,
0.5,
0.7587890625
],
[
0.2059326171875,
0.755859375,
0.2197265625,
0.75390625,
0.22216796875,
0.7646484375,
0.2091064453125,
0.7666015625
],
[
0.226318359375,
0.76123046875,
0.386474609375,
0.7587890625,
0.386474609375,
0.77587890625,
0.226318359375,
0.77783203125
],
[
0.5009765625,
0.76025390625,
0.5810546875,
0.76025390625,
0.5810546875,
0.7744140625,
0.5009765625,
0.7744140625
],
[
0.2254638671875,
0.76953125,
0.233642578125,
0.76953125,
0.233642578125,
0.7734375,
0.2254638671875,
0.7734375
],
[
0.17724609375,
0.76318359375,
0.2132568359375,
0.76318359375,
0.2132568359375,
0.77587890625,
0.17724609375,
0.77587890625
],
[
0.19775390625,
0.7744140625,
0.2181396484375,
0.7744140625,
0.2181396484375,
0.7822265625,
0.19775390625,
0.7822265625
],
[
0.2181396484375,
0.77783203125,
0.24267578125,
0.77783203125,
0.24267578125,
0.79150390625,
0.2181396484375,
0.79150390625
],
[
0.450927734375,
0.7734375,
0.473876953125,
0.7734375,
0.473876953125,
0.787109375,
0.450927734375,
0.787109375
],
[
0.1871337890625,
0.79248046875,
0.27783203125,
0.79248046875,
0.27783203125,
0.81005859375,
0.1871337890625,
0.81005859375
],
[
0.5009765625,
0.79150390625,
0.69921875,
0.79150390625,
0.69921875,
0.80615234375,
0.5009765625,
0.80615234375
],
[
0.5009765625,
0.81640625,
0.88623046875,
0.81640625,
0.88623046875,
0.830078125,
0.5009765625,
0.830078125
],
[
0.0882568359375,
0.82373046875,
0.473876953125,
0.82373046875,
0.473876953125,
0.8369140625,
0.0882568359375,
0.8369140625
],
[
0.5,
0.83056640625,
0.88623046875,
0.83056640625,
0.88623046875,
0.8447265625,
0.5,
0.8447265625
],
[
0.0882568359375,
0.8388671875,
0.473876953125,
0.8388671875,
0.473876953125,
0.85302734375,
0.0882568359375,
0.85302734375
],
[
0.5,
0.8466796875,
0.8857421875,
0.8466796875,
0.8857421875,
0.85986328125,
0.5,
0.85986328125
],
[
0.0882568359375,
0.85400390625,
0.221435546875,
0.85400390625,
0.221435546875,
0.8671875,
0.0882568359375,
0.8671875
],
[
0.498291015625,
0.861328125,
0.8857421875,
0.85986328125,
0.8857421875,
0.875,
0.498291015625,
0.87548828125
],
[
0.0906982421875,
0.87548828125,
0.473876953125,
0.87548828125,
0.473876953125,
0.8896484375,
0.0906982421875,
0.8896484375
],
[
0.5,
0.876953125,
0.8857421875,
0.876953125,
0.8857421875,
0.89013671875,
0.5,
0.89013671875
],
[
0.0882568359375,
0.892578125,
0.4755859375,
0.892578125,
0.4755859375,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.5009765625,
0.89306640625,
0.8837890625,
0.89306640625,
0.8837890625,
0.90380859375,
0.5009765625,
0.90380859375
]
] |
[
"T-GD:TransferableGAN-generatedImagesDetectionFramework",
"fer (fine-tune) this pre-trained model to the target dataset.",
"of thelearned information orcatastrophic forgetting.",
"For instance, EfficientNet (the CNN classifier) is trained",
"212(w)= [wl3",
"(2)",
"StyleGAN-dataset (the target).",
"on the PGGAN-dataset (the source) and fine-tuned on the",
"Equation 2 is the 2 function, namely the L?-norm of w.",
"EfficientNet.We implemented EfficientNet-BO (Tan &",
"Le, 2019) and used it as the CNN classifier. Although",
"1.",
"n",
"Z J(gi, yi)+ B. 2u2(wfe).",
"",
"(3)",
"EfficientNet-BO has the lowest number of parameters (about",
"nin-",
"",
"four million) among the EfficientNets, it performs well",
"2-1",
"in GAN-image detection in our experiment, compared to",
"In Eg. 3, the first term is the same as in Eq. 1, representing",
"Inception-V3 (Szegedy et al., 2017) and Xception (Chollet,",
"the cross-entropy loss function. The second term is the u2",
"2017). Another change we make to the model is the use of",
"function or the L? regularization term (Eq. 2) of w fc, the",
"WS(Oiaoetal.,2019)andGN(Wu&He.2018),insteadof",
"weights of the FC layers, scaled by β, which is equivalent to",
"batch normalization (BN) (loffe & Szegedy,2015), due to",
"A in Eq. 1. Note that the L2 regularization is applied solely",
"their superior efficiency regarding transfer learning to that",
"to the FC layers since over-fitting and forgetting are delayed,",
"of batch statistics.",
"but not completely prevented in the course of learning.",
"ResNext.We implemented ResNext32x4d (Xie et al.,",
"L2-SP. Let w' be the pre-trained weights from the source",
"2017) and used it as the CNN classifier, where",
"dataset, as shown in Section 3.1, serving as the starting point",
"ResNext32x4d has more parameters (about twenty mil-",
"(SP) as the reference, as well as a regularization point that",
"lion) than EfficientNet-BO. We also replace BN with WS",
"provides guidance for transfer learning when fine-tuning",
"and GN.",
"Using L2-norm, we define L?-SP as follows:",
"Nsp(w, w') = llwconu - wconvll2,",
"(4)",
"The next step is transfer learning. The weight of the pre-",
"3.1. L2-SP",
"where wcon denotes the weights of the convolution layers,",
"trained model from the source dataset is used as the SPAR.",
"excluding those of the FC layers. Equation 4 indicates that",
"In particular, we use L?-SP for transfer learning. Regu-",
"L?-sP is a one-to-one mapping between the convolution",
"larization can lead to a better optimization by preventing",
"layers of the source and target datasets, e.g., the PGGAN-",
"over-fitting when learning from scratch; L?-SP differs in",
"classifier (the source) to the StyleGAN-classifier (the target).",
"that the starting point from a well pre-trained source dataset",
"Loss function. We combine Eq. 4, sharing the architecture",
"guides the learning process by referring to the information",
"ofthe sourceand targetmodels.with the second term of",
"of the pre-trained source dataset.This method does not",
"require freezing the weights of the pre-trained model nor",
"Eq. 3, accounting for the FC layer (final layer) as follows:",
"using weight decay. Our method regularizes convolution",
"1",
"layers and fully-connected (FC) layers independently",
"min",
"J(yi, yi)+Q-sp(w, w') +β.Q12(wfe),",
"n",
"General form of regularization. Let w be the weight",
"i=1",
"parameters, and J(yi, yi) be the loss function of the neural",
"J(yi, ya) =-yi log(yi) - (1 -yi) log(1 -yi),",
"networks, where y; is the ith score predicted by the models",
"(5)",
"and y; is the ith label. And 2(w) is the p-norm function of",
"where J is the negative log-likelihood loss function, and α",
"the weight w as a general form of regularization loss, fw(ri)",
"is the neural network function with the ith data Ci, and n is",
"and β are tunable hyperparameters of which (Li et al., 2019)",
"use values in the range from 0.1 to 0.01. The difference from",
"the dataset size. Equation 1 indicates the general form of",
"L?-SP is that we transform α and β into , a parameter",
"the loss function with a weight regularization component:",
"which adjusts itself according to the learning situation. More",
"details about the transformed parameters are provided in",
"1",
"J(gi,yi) +^.2(w),",
"Section 3.3.",
"2",
"min-",
"",
"i=1",
"(1)",
"gi = fu(ri),",
"3.2. Self-training for L2-SP",
"We transform the transfer learning framework into a self-",
"where balances the regularization and the loss function, J",
"training framework. In other words, the source/target model",
"is the cross-entropy function, and 2 is the LI or L2-norm",
"is changed into a teacher/student model. In addition to the",
"of the parameter w.",
"role of a typical source model, which serves as SPAR and a",
"L? regularization. L? regularization is used in transfer",
"regularizer to guide the learning process, the teacher model",
"learning to avoid over-fitting and to overcome the forgetting",
"has theroleof adjustingtheparametersbased onthelearned"
] |
{
"arxiv_id": "2008.04115",
"language": "en",
"page_idx": 3
}
|
||
2002.00412-en-1
|
2002.00412-en-1.jpg
|
# 2 Background
We study the problem of training agents to follow natural language instructions grounded in a partially-observable environment using adversarial imitation learning. In particular, we assume that $N$ pairs $(c_i,\tau_i)$ of instructions $c_i$ and their respective trajectories $\tau_i=\left((o_{1}^{i},a_{1}^{i}),\ldots(o_{T(i)}^{i},a_{T(i)}^{i})\right)$ are provided for the agent to learn from. Here, $o_t^i$ and $a_t^i$ are the observation and action at the time step $t$, and $T(i)$ is the $i$-th episode length.
GAIL (Ho and Ermon 2016) is an imitation learning framework which combines ideas from Inverse Reinforcement Learning (Ng and Russell 2000; Abbeel and Ng 2004a) and Generative Adversarial Networks (GANs) (Goodfellow et al. 2014). In GAIL, the actor network represents the agent's policy $\pi$ while the discriminator network $D$ serves as a local reward function differentiating between the expert policy $\pi_E$ and the actor using observation-action pairs $(o,a).$ Ho and Ermon (2016) define the optimization objective of GAIL as follows:
(1)
$$\begin{aligned}\max_{\pi}\operatorname*{min}_{D}\Big(\mathbb{E}_{\pi}[-\log(1-D(o,a))]+\\&+\mathbb{E}_{\pi_{E}}[-\log D(o,a)]+\lambda H(\pi)\Big),\end{aligned}$$
where function $H$ stands for the maximum entropy introduced by Ziebart et al. (2008), ensuring adequate exploration of the environment. GAIL was originally proposed for fully observable environments where the observation $o$ in Equation 1 must capture the complete state of the environment.
Similarly to Fu et al. (2019), we extend GAIL to learn instruction-conditioned policies by conditioning both the generator (policy) $\pi$ and the discriminator $D$ on the instruction $c.$ The resulting objective is
(2)
$$\begin{aligned}\max_{\pi}\operatorname*{min}_{D}\Big(\mathbb{E}_{\pi}[-\log(1-D(o,c,a))]+\\&+\mathbb{E}_{\pi_{E}}[-\log D(o,c,a)]+\lambda H(\pi)\Big),\end{aligned}$$
where $\mathbb{E}_\pi$ stands for sampling $(o,c,a)$ by acting with the policy $\pi(a|o,c)$ conditioned on both the observation $o$ and the instruction $c.$
2.1 Conditioned Recurrent Discriminator
For partially-observable environments it is necessary to equip the discriminator with a memory to model the true environment reward (see Figure 1(b) for example). To this end, we implement the discriminator as a recurrent neural network. Its input is not a single observation-action pair $(o_{i},a_{i})$, as in the original GAIL formulation, but a full tra- $\mathop{\text{jectory}\tau}=((o_{1},a_{1}),\ldots,(o_{T},a_{T})).$To address the goalconditioned nature of the problem, the discriminator is still conditioned on the instruction $c$, as described in the previous section. Since our recurrent discriminator receives only full trajectories as the training signal, its predictions for incomplete sub-trajectories are not reliable, and hence its outputs should not be used to reward the policy at each intermediate step like Equation 2 would require. For this reason, we
reward the policy, which is also implemented as a recurrent network, only at the last step of the episode. The resulting training objective is
(3)
$$\begin{aligned}\max_{\pi}\operatorname*{min}_{D}\Big(\mathbb{E}_{\pi}[-\log(1-D(c,\tau))]+\\&+\mathbb{E}_{\pi_{E}}[-\log D(c,\tau)]+\lambda H(\pi)\Big).\end{aligned}$$
An alternative approach to adapt GAIL to partially observable environments would be to train the discriminator to distinguish not just between complete trajectories but also between incomplete sub-trajectories. The policy could then be rewarded at each step like it is done in the original GAIL. Such a dense reward formulation may be necessary when the discriminator does not have memory, but with memory in place the dense rewards become optional. We have chosen the full trajectory approach to avoid some of the known issues of the dense reward GAIL formulation, such as the survival bias (Kostrikov et al. 2018). In Section 7.1, we present additional experiments in which incomplete sub-trajectories are used to train the discriminator
To compute a batched version of the discriminator loss we use two finite buffers $B_{agent}$ and $B_{expert}$ that contain agent trajectories and expert demonstrations respectively:
(4)
$$\begin{aligned}L_{D}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{agent}}-\log(1-D_{\theta}(c,\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c,\tau)).\end{aligned}$$
# 3 False Negatives
The main premise of GAIL is that discriminator-based rewards will be high for the episodes that manifest expected behaviour. In practice, however, the discriminator is trained to output higher rewards for the expert and lower for the agent which, unfortunately, is not the same.
In the starting phase of the learning procedure, the discriminator can reliably assign a high value to a successful episode. Indeed, such episode must be an expert demonstration as the agent always fails. However, the streams of positive and negative examples that the discriminator receives can become very similar as the agent gets better and more likely to succeed. The discriminator can no longer assume that all successful trajectories are coming from the expert and has to detect idiosyncratic features in expert demonstrations that are not necessarily related to solving a given task. In the extreme case of the perfect agent, the discriminator has to overfit the nuisances in expert demonstrations to minimize its loss, because the agent and expert behaviours are indistinguishable. Discriminator-based rewards for successful agent episodes can potentially become as low as for unsuccessful ones.
The culprit of the aforementioned problem is a naive labeling of successful agent episodes as negative examples for the discriminator. We would call such examples False Negatives (FN).
# 3.1 Oracle Filtering We propose an approach, which we call Oracle Filtering
(OF), to diagnose if the false negatives hinder the performance of adversarial imitation learning. In particular, OF
|
[
[
0.2115478515625,
0.06500244140625,
0.3505859375,
0.06756591796875,
0.3505859375,
0.085205078125,
0.2115478515625,
0.08331298828125
],
[
0.51953125,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.51953125,
0.08331298828125
],
[
0.0882568359375,
0.08837890625,
0.4794921875,
0.08837890625,
0.4794921875,
0.1029052734375,
0.0882568359375,
0.1029052734375
],
[
0.5185546875,
0.083984375,
0.91259765625,
0.083984375,
0.91259765625,
0.09783935546875,
0.5185546875,
0.09783935546875
],
[
0.0849609375,
0.1029052734375,
0.47802734375,
0.1029052734375,
0.47802734375,
0.11614990234375,
0.0849609375,
0.11614990234375
],
[
0.5185546875,
0.09783935546875,
0.65283203125,
0.09783935546875,
0.65283203125,
0.111083984375,
0.5185546875,
0.111083984375
],
[
0.0849609375,
0.117431640625,
0.47802734375,
0.117431640625,
0.47802734375,
0.1307373046875,
0.0849609375,
0.1307373046875
],
[
0.654296875,
0.121826171875,
0.8095703125,
0.12371826171875,
0.8095703125,
0.1383056640625,
0.654296875,
0.1363525390625
],
[
0.890625,
0.12371826171875,
0.9140625,
0.12371826171875,
0.9140625,
0.136962890625,
0.890625,
0.136962890625
],
[
0.0849609375,
0.1300048828125,
0.47802734375,
0.1300048828125,
0.47802734375,
0.143310546875,
0.0849609375,
0.143310546875
],
[
0.57275390625,
0.1256103515625,
0.619140625,
0.1256103515625,
0.619140625,
0.1363525390625,
0.57275390625,
0.1363525390625
],
[
0.08416748046875,
0.1483154296875,
0.47802734375,
0.1483154296875,
0.47802734375,
0.1666259765625,
0.08416748046875,
0.1666259765625
],
[
0.6337890625,
0.149658203125,
0.85546875,
0.1483154296875,
0.85546875,
0.1666259765625,
0.6337890625,
0.1673583984375
],
[
0.08660888671875,
0.1685791015625,
0.47802734375,
0.1685791015625,
0.47802734375,
0.1817626953125,
0.08660888671875,
0.1817626953125
],
[
0.0849609375,
0.18310546875,
0.4794921875,
0.18310546875,
0.4794921875,
0.1962890625,
0.0849609375,
0.1962890625
],
[
0.5361328125,
0.177978515625,
0.91015625,
0.177978515625,
0.91015625,
0.1912841796875,
0.5361328125,
0.1912841796875
],
[
0.08660888671875,
0.1976318359375,
0.1871337890625,
0.1976318359375,
0.1871337890625,
0.208984375,
0.08660888671875,
0.208984375
],
[
0.52294921875,
0.1932373046875,
0.9111328125,
0.1932373046875,
0.9111328125,
0.203857421875,
0.52294921875,
0.203857421875
],
[
0.10296630859375,
0.210205078125,
0.47802734375,
0.210205078125,
0.47802734375,
0.2235107421875,
0.10296630859375,
0.2235107421875
],
[
0.521484375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.521484375,
0.2183837890625
],
[
0.0882568359375,
0.2259521484375,
0.4755859375,
0.2259521484375,
0.4755859375,
0.236083984375,
0.0882568359375,
0.236083984375
],
[
0.52294921875,
0.2203369140625,
0.9111328125,
0.2203369140625,
0.9111328125,
0.2310791015625,
0.52294921875,
0.2310791015625
],
[
0.08660888671875,
0.238037109375,
0.476318359375,
0.238037109375,
0.476318359375,
0.251953125,
0.08660888671875,
0.251953125
],
[
0.51953125,
0.23291015625,
0.91015625,
0.23291015625,
0.91015625,
0.246826171875,
0.51953125,
0.246826171875
],
[
0.521484375,
0.246826171875,
0.91259765625,
0.246826171875,
0.91259765625,
0.260009765625,
0.521484375,
0.260009765625
],
[
0.08660888671875,
0.25390625,
0.4755859375,
0.25390625,
0.4755859375,
0.264404296875,
0.08660888671875,
0.264404296875
],
[
0.08660888671875,
0.266357421875,
0.47802734375,
0.266357421875,
0.47802734375,
0.27978515625,
0.08660888671875,
0.27978515625
],
[
0.52294921875,
0.261962890625,
0.9111328125,
0.261962890625,
0.9111328125,
0.272705078125,
0.52294921875,
0.272705078125
],
[
0.5185546875,
0.274658203125,
0.91259765625,
0.2734375,
0.91259765625,
0.287353515625,
0.5185546875,
0.28857421875
],
[
0.0882568359375,
0.281005859375,
0.47802734375,
0.281005859375,
0.47802734375,
0.291748046875,
0.0882568359375,
0.291748046875
],
[
0.08660888671875,
0.29345703125,
0.476318359375,
0.29345703125,
0.476318359375,
0.306884765625,
0.08660888671875,
0.306884765625
],
[
0.51953125,
0.28857421875,
0.91259765625,
0.28857421875,
0.91259765625,
0.302978515625,
0.51953125,
0.302978515625
],
[
0.51953125,
0.3017578125,
0.91015625,
0.3017578125,
0.91015625,
0.31494140625,
0.51953125,
0.31494140625
],
[
0.0882568359375,
0.309326171875,
0.476318359375,
0.309326171875,
0.476318359375,
0.3193359375,
0.0882568359375,
0.3193359375
],
[
0.51953125,
0.31494140625,
0.91259765625,
0.31494140625,
0.91259765625,
0.328857421875,
0.51953125,
0.328857421875
],
[
0.08660888671875,
0.3212890625,
0.476318359375,
0.3212890625,
0.476318359375,
0.335205078125,
0.08660888671875,
0.335205078125
],
[
0.0849609375,
0.335205078125,
0.2484130859375,
0.335205078125,
0.2484130859375,
0.348388671875,
0.0849609375,
0.348388671875
],
[
0.521484375,
0.330078125,
0.9140625,
0.330078125,
0.9140625,
0.343505859375,
0.521484375,
0.343505859375
],
[
0.521484375,
0.34521484375,
0.74169921875,
0.34521484375,
0.74169921875,
0.35595703125,
0.521484375,
0.35595703125
],
[
0.223876953125,
0.35791015625,
0.374267578125,
0.35791015625,
0.374267578125,
0.37255859375,
0.223876953125,
0.37255859375
],
[
0.45751953125,
0.35986328125,
0.4794921875,
0.35986328125,
0.4794921875,
0.373779296875,
0.45751953125,
0.373779296875
],
[
0.53515625,
0.35791015625,
0.91259765625,
0.35791015625,
0.91259765625,
0.371826171875,
0.53515625,
0.371826171875
],
[
0.51953125,
0.371826171875,
0.9140625,
0.371826171875,
0.9140625,
0.386474609375,
0.51953125,
0.386474609375
],
[
0.136474609375,
0.361083984375,
0.2034912109375,
0.361083984375,
0.2034912109375,
0.37255859375,
0.136474609375,
0.37255859375
],
[
0.200927734375,
0.386962890625,
0.41259765625,
0.385009765625,
0.41259765625,
0.40283203125,
0.200927734375,
0.40478515625
],
[
0.51953125,
0.3837890625,
0.85693359375,
0.3837890625,
0.85693359375,
0.397705078125,
0.51953125,
0.397705078125
],
[
0.56884765625,
0.404052734375,
0.86376953125,
0.404052734375,
0.86376953125,
0.4248046875,
0.56884765625,
0.4248046875
],
[
0.890625,
0.4072265625,
0.9140625,
0.4072265625,
0.9140625,
0.419189453125,
0.890625,
0.419189453125
],
[
0.0849609375,
0.4140625,
0.4755859375,
0.415283203125,
0.4755859375,
0.429931640625,
0.0849609375,
0.4287109375
],
[
0.61767578125,
0.423583984375,
0.84228515625,
0.423583984375,
0.84228515625,
0.4443359375,
0.61767578125,
0.4443359375
],
[
0.0849609375,
0.429931640625,
0.47802734375,
0.429931640625,
0.47802734375,
0.443115234375,
0.0849609375,
0.443115234375
],
[
0.08660888671875,
0.443115234375,
0.4794921875,
0.443115234375,
0.4794921875,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.6337890625,
0.45068359375,
0.79736328125,
0.45263671875,
0.79736328125,
0.46728515625,
0.6337890625,
0.46533203125
],
[
0.0849609375,
0.45703125,
0.4794921875,
0.45703125,
0.4794921875,
0.47021484375,
0.0849609375,
0.47021484375
],
[
0.0849609375,
0.4716796875,
0.4794921875,
0.4716796875,
0.4794921875,
0.48486328125,
0.0849609375,
0.48486328125
],
[
0.51953125,
0.47021484375,
0.91015625,
0.4716796875,
0.91015625,
0.486083984375,
0.51953125,
0.48486328125
],
[
0.0849609375,
0.486083984375,
0.143798828125,
0.486083984375,
0.143798828125,
0.49755859375,
0.0849609375,
0.49755859375
],
[
0.52294921875,
0.488037109375,
0.9111328125,
0.488037109375,
0.9111328125,
0.498779296875,
0.52294921875,
0.498779296875
],
[
0.10296630859375,
0.49755859375,
0.47802734375,
0.49755859375,
0.47802734375,
0.51220703125,
0.10296630859375,
0.51220703125
],
[
0.51953125,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.5146484375,
0.51953125,
0.5146484375
],
[
0.0849609375,
0.51220703125,
0.4794921875,
0.51220703125,
0.4794921875,
0.52587890625,
0.0849609375,
0.52587890625
],
[
0.51953125,
0.5146484375,
0.91259765625,
0.5146484375,
0.91259765625,
0.52783203125,
0.51953125,
0.52783203125
],
[
0.08660888671875,
0.5263671875,
0.476318359375,
0.5263671875,
0.476318359375,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.521484375,
0.52978515625,
0.80615234375,
0.52978515625,
0.80615234375,
0.54052734375,
0.521484375,
0.54052734375
],
[
0.0849609375,
0.5390625,
0.2998046875,
0.5390625,
0.2998046875,
0.55322265625,
0.0849609375,
0.55322265625
],
[
0.53515625,
0.54248046875,
0.91015625,
0.54248046875,
0.91015625,
0.55615234375,
0.53515625,
0.55615234375
],
[
0.51953125,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.5693359375,
0.51953125,
0.5693359375
],
[
0.193603515625,
0.564453125,
0.384033203125,
0.56494140625,
0.384033203125,
0.58349609375,
0.193603515625,
0.58203125
],
[
0.45751953125,
0.5673828125,
0.4794921875,
0.5673828125,
0.4794921875,
0.58056640625,
0.45751953125,
0.58056640625
],
[
0.51953125,
0.5693359375,
0.91015625,
0.5693359375,
0.91015625,
0.58349609375,
0.51953125,
0.58349609375
],
[
0.521484375,
0.58544921875,
0.91015625,
0.58544921875,
0.91015625,
0.59521484375,
0.521484375,
0.59521484375
],
[
0.131591796875,
0.568359375,
0.2034912109375,
0.568359375,
0.2034912109375,
0.580078125,
0.131591796875,
0.580078125
],
[
0.1953125,
0.59521484375,
0.42236328125,
0.59521484375,
0.42236328125,
0.60986328125,
0.1953125,
0.60986328125
],
[
0.51953125,
0.59765625,
0.91259765625,
0.59765625,
0.91259765625,
0.611328125,
0.51953125,
0.611328125
],
[
0.51953125,
0.611328125,
0.91259765625,
0.611328125,
0.91259765625,
0.625,
0.51953125,
0.625
],
[
0.0882568359375,
0.62353515625,
0.47802734375,
0.62353515625,
0.47802734375,
0.63818359375,
0.0882568359375,
0.63818359375
],
[
0.51953125,
0.625,
0.91259765625,
0.625,
0.91259765625,
0.63818359375,
0.51953125,
0.63818359375
],
[
0.08660888671875,
0.63818359375,
0.4794921875,
0.63818359375,
0.4794921875,
0.6513671875,
0.08660888671875,
0.6513671875
],
[
0.5185546875,
0.63720703125,
0.9140625,
0.63818359375,
0.9140625,
0.65283203125,
0.5185546875,
0.6513671875
],
[
0.0849609375,
0.6513671875,
0.1993408203125,
0.6513671875,
0.1993408203125,
0.66552734375,
0.0849609375,
0.66552734375
],
[
0.521484375,
0.65380859375,
0.90869140625,
0.65380859375,
0.90869140625,
0.6640625,
0.521484375,
0.6640625
],
[
0.521484375,
0.66845703125,
0.91015625,
0.66845703125,
0.91015625,
0.6787109375,
0.521484375,
0.6787109375
],
[
0.08660888671875,
0.6767578125,
0.416748046875,
0.6767578125,
0.416748046875,
0.68994140625,
0.08660888671875,
0.68994140625
],
[
0.521484375,
0.68115234375,
0.91259765625,
0.68115234375,
0.91259765625,
0.6943359375,
0.521484375,
0.6943359375
],
[
0.08660888671875,
0.69580078125,
0.4794921875,
0.69580078125,
0.4794921875,
0.708984375,
0.08660888671875,
0.708984375
],
[
0.521484375,
0.69580078125,
0.90869140625,
0.69580078125,
0.90869140625,
0.70556640625,
0.521484375,
0.70556640625
],
[
0.0849609375,
0.708984375,
0.47802734375,
0.708984375,
0.47802734375,
0.72265625,
0.0849609375,
0.72265625
],
[
0.51953125,
0.70849609375,
0.91259765625,
0.70849609375,
0.91259765625,
0.7216796875,
0.51953125,
0.7216796875
],
[
0.0849609375,
0.7236328125,
0.47802734375,
0.7236328125,
0.47802734375,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.52294921875,
0.7236328125,
0.9111328125,
0.7236328125,
0.9111328125,
0.734375,
0.52294921875,
0.734375
],
[
0.08416748046875,
0.736328125,
0.47802734375,
0.7373046875,
0.47802734375,
0.75048828125,
0.08416748046875,
0.75
],
[
0.51953125,
0.73486328125,
0.9111328125,
0.73486328125,
0.9111328125,
0.74853515625,
0.51953125,
0.74853515625
],
[
0.08660888671875,
0.751953125,
0.4794921875,
0.751953125,
0.4794921875,
0.76513671875,
0.08660888671875,
0.76513671875
],
[
0.51953125,
0.75048828125,
0.6259765625,
0.75,
0.6259765625,
0.76123046875,
0.51953125,
0.76220703125
],
[
0.0849609375,
0.76513671875,
0.476318359375,
0.7646484375,
0.476318359375,
0.77783203125,
0.0849609375,
0.77880859375
],
[
0.53515625,
0.76220703125,
0.9140625,
0.76123046875,
0.9140625,
0.77587890625,
0.53515625,
0.7763671875
],
[
0.0849609375,
0.77880859375,
0.47802734375,
0.77783203125,
0.47802734375,
0.79150390625,
0.0849609375,
0.79248046875
],
[
0.51953125,
0.7763671875,
0.9111328125,
0.7763671875,
0.9111328125,
0.79052734375,
0.51953125,
0.79052734375
],
[
0.08660888671875,
0.79345703125,
0.47802734375,
0.79345703125,
0.47802734375,
0.8037109375,
0.08660888671875,
0.8037109375
],
[
0.521484375,
0.79150390625,
0.91259765625,
0.79150390625,
0.91259765625,
0.8046875,
0.521484375,
0.8046875
],
[
0.0849609375,
0.80615234375,
0.476318359375,
0.80615234375,
0.476318359375,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.51953125,
0.8046875,
0.66650390625,
0.8046875,
0.66650390625,
0.818359375,
0.51953125,
0.818359375
],
[
0.08660888671875,
0.8212890625,
0.476318359375,
0.8212890625,
0.476318359375,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.0849609375,
0.833984375,
0.476318359375,
0.833984375,
0.476318359375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.5185546875,
0.8291015625,
0.685546875,
0.830078125,
0.685546875,
0.8447265625,
0.5185546875,
0.84326171875
],
[
0.08660888671875,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.521484375,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.521484375,
0.861328125
],
[
0.0849609375,
0.8623046875,
0.47802734375,
0.8623046875,
0.47802734375,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.52294921875,
0.86279296875,
0.91015625,
0.86279296875,
0.91015625,
0.87353515625,
0.52294921875,
0.87353515625
],
[
0.0849609375,
0.87548828125,
0.4794921875,
0.87548828125,
0.4794921875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51953125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"2Background",
"reward the policy, which is also implemented as a recurrent",
"We study the problem of training agents to follow natural",
"network, only at the last step of the episode. The resulting",
"language instructions grounded in a partially-observable en-",
"training objective is",
"vironment using adversarial imitation learning. In particular,",
"元[- log(1 - D(c, T)))+",
"(3)",
"we assume that N pairs (ci, T;) of instructions c, and their",
"maxm",
"respective trajectories T; = ((oi,ai), ... (or(i),ar()) are",
"+ E元[-1og D(c, T)) + ^H())",
"provided for the agent to learn from. Here, o and ai are the",
"observation and action at the time step t, and T(i) is the i-th",
"An alternative approach to adapt GAIL to partially ob-",
"episode length.",
"servableenvironmentswouldbetotrainthediscriminatorto",
"GAIL (Ho and Ermon 2016) is an imitation learning",
"distinguish not just between complete trajectories but also",
"frameworkwhich combinesideasfrom Inverse Reinforce-",
"betweenincompletesub-trajectories.Thepolicycouldthen",
"ment Learning (Ng and Russell 2000; Abbeel and Ng 2004a)",
"be rewarded at each step like it is done in the original GAIL",
"Such a dense reward formulation may be necessary when",
"andGenerativeAdversarialNetworks(GANs)(Goodfellow",
"et al. 2014). In GAIL, the actor network represents the",
"thediscriminatordoesnothavememory,butwithmemory",
"in place the dense rewards become optional. We have chosen",
"agent'spolicyπ whilethe discriminatornetworkDserves",
"as a local reward function differentiating between the ex-",
"the full trajectory approach to avoid some of the known is-",
"sues of the dense reward GAIL formulation, such as the sur-",
"pertpolicy Eand the actorusing observation-actionpairs",
"vival bias (Kostrikov et al. 2018). In Section 7.1, we present",
"(o, a). Ho and Ermon (2016) define the optimization objec-",
"tive of GAIL as follows:",
"additional experiments in which incomplete sub-trajectories",
"areused totrain thediscriminator",
"[- log(1 - D(o, a))I+",
"(1)",
"To compute a batched version of the discriminator loss we",
"use two finite buffers Bagent and Berpert that contain agent",
"maxmin",
"+ E[-log D(0, a)] +入H(π)",
"trajectories and expert demonstrations respectively:",
"LD(0) =E(c,r)~Bagent - log(1 -De(c, T)",
"(4)",
"where function H stands for the maximum entropy intro-",
"+E(ce,t)~Beapert - log(De(c, T).",
"duced by Ziebart et al. (2008), ensuring adequate explo-",
"ration of the environment. GAIL was originally proposed",
"3False Negatives",
"for fully observable environments where the observation o",
"in Equation 1 must capture the complete state of the envi-",
"The main premise of GAIL is that discriminator-based re-",
"ronment",
"wards will behighfor theepisodes that manifest expected",
"Similarly to Fu et al. (2019), we extend GAIL to learn",
"behaviour. In practice, however, the discriminator is trained",
"instruction-conditioned policies by conditioning both the",
"to output higher rewards for the expert and lower for the",
"generator (policy) and the discriminator D on the instruc-",
"agent which, unfortunately, is not the same.",
"tion c. The resulting objective is",
"In the starting phase of the learning procedure, the dis-",
"criminator can reliably assign a high value to a successful",
"(E[- log(1 - D(o, c, a)]+",
"(2)",
"episode. Indeed, such episode must be an expert demonstra-",
"tion as the agentalways fails.However,the streams of pos-",
"max min()",
"+ E元E/log D(o, C, a)l + ^H(元)",
"itive and negative examples that the discriminator receives",
"can become very similar as the agent gets better and more",
"where E, stands for sampling (o, c, a) by acting with the",
"likely to succeed. The discriminator can no longer assume",
"policy π(ao,c) conditioned on both the observation o and",
"that all successful trajectories are coming from the expert",
"the instruction c.",
"andhastodetectidiosyncraticfeaturesinexpertdemonstra",
"tionsthatarenotnecessarilyrelatedtosolvingagiventask",
"2.1 Conditioned Recurrent Discriminator",
"In the extreme case of the perfect agent, the discriminator",
"For partially-observable environments it is necessary to",
"hastooverfitthenuisances inexpertdemonstrationstomin",
"equip the discriminator with a memory to model the true",
"imize its loss, because the agent and expert behaviours are",
"environment reward (see Figure 1(b) for example). To this",
"indistinguishable.Discriminator-basedrewardsforsuccess-",
"end, we implement the discriminator as a recurrent neu-",
"ful agent episodes can potentially become as low as for un-",
"ral network. Its input is not a single observation-action pair",
"successful ones",
"(oi, ai), as in the original GAIL formulation, but a full tra-",
"The culprit of the aforementioned problem is a naive",
"jectory T = ((o1,ai),...,(oT,at)).To address the goal-",
"labeling of successful agent episodes as negative exam-",
"conditionednatureof theproblemthediscriminatoris still",
"ples for the discriminator. We would call such examples",
"conditioned on the instruction c, as described in the previous",
"False Negatives (FN).",
"section. Since ourrecurrent discriminatorreceives onlyfull",
"trajectories as the training signal, its predictions for incom-",
"3.1 OracleFiltering",
"plete sub-trajectories are not reliable, and hence its outputs",
"We propose an approach, which we call Oracle Filtering",
" should not be used to reward the policy at each intermedi-",
"(OF),to diagnose if the false negatives hinderthe perfor-",
"ate step like Equation 2 would require. For this reason, we",
"mance of adversarial imitation learning. In particular, OF"
] |
{
"arxiv_id": "2002.00412",
"language": "en",
"page_idx": 1
}
|
||
2002.00412-en-2
|
2002.00412-en-2.jpg
|
uses the environment's true reward signal to identify the successful trajectories generated by the agent and filter them $\begin{array}{l}\mathrm{out.~In~this~setting,~the~buffer~of~negative~examples~}B_{agent}\\\mathrm{contains~only~unsuccessful~trajectories.~The~discriminator's}\end{array}$ loss is given as in Equation 4 but instead of $B_{agent}$ we use $\begin{array}{l}B_{agent}^{oracle}\text{which is a subset of }B_{agent}\text{ that contains unsuccess-}\\\text{ful trajectories only:}\end{array}$
(5)
$$\begin{aligned}L_{D}^{O}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{agent}^{oracle}}-\log(1-D_{\theta}(c,\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c,\tau)).\end{aligned}$$
Note that when the OF method is applied, the environment rewards are used solely to filter out successful agent trajectories-the agent does not have access to them.
We found that using OF significantly improves the performance, and reduces sample complexity by at least an order of magnitude. This confirms our hypothesis that false negatives have a detrimental effect on GAIL training.
# 3.2 Fake Conditioning
Since the Oracle Filtering technique requires access to environment rewards, it is only a diagnostic method suited to assess the impact of false negatives. We propose a technique that does not need environment rewards to address the FN problem. Our technique is aimed at tasks where the policy is goal-conditioned. In our case, we assume that the policy is conditioned on language instructions but the technique is general and can be applied in any multi-goal setup.
First, we maintain a set of possible language instructions $S$ initialized from all the unique instructions in the expert demonstrations and, optionally, updated with the instructions collected when the agent interacts with the environment. Secondly, for each instruction-trajectory pair $(c,\tau)$, we replace its original instruction $c$ with $\tilde{c}$ that is randomly $\mathop{\mathrm{sampled~from~}}S\backslash\{c\}.$
We call this technique Fake Conditioning (FC). It is motivated by the fact that the success of a trajectory is conditioned on the instruction. Therefore, for each trajectory, by replacing the instruction with a new one, we produce a new instruction-trajectory pair that is very likely to be unsuccessful, even when it is successfully conditioned on the original instruction. The FC technique can be used to prepare discriminator training data with greatly limited number of the false negatives.
Expert FC We call Expert FC a technique whereby the discriminator is trained to distinguish the expert trajectories with the original instructions from fake conditioned expert trajectories. In this case, the agent trajectories are not used at all, and hence the training is no longer adversarial. The discriminator's loss is given as follows:
(6)
$$\begin{aligned}L_{D}^{E}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{expert},\:\tilde{c}\sim S\setminus\{c\}}-\log(1-D_{\theta}(\tilde{c},\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c,\tau)),\end{aligned}$$
where $(\tilde{c},\tau)$ is fake conditioned instruction-trajectory pair. Agent FC The FC technique can also be used with agent trajectories. In this case, which we call Agent FC, all agent trajectories are fake conditioned while expert ones are left
unaltered. It means that positive examples stay as in Expert FC formulation but negatives are constructed from the fake conditioned agent trajectories instead of expert ones. In this case, the discriminator's loss is given as follows:
$$\begin{aligned}L_{D}^{A}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{agent},\:\tilde{c}\sim S\setminus\{c\}}-\log(1-D_{\theta}(\tilde{c},\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c,\tau)).\end{aligned}$$
We note that the policy that generates agent trajectories is always conditioned on the original instruction, for both Expert FC and Agent FC. The fake instruction-trajectory pairs are used to train the discriminator only.
# 3.3 Auxiliary Rewards
The FC technique greatly reduces the percentage of FN. Unfortunately, the discriminator is no longer trained on agent trajectories with the original instructions, and hence the rewards that the discriminator issues to the policy can potentially be less accurate. These reward inaccuracies could be exploited by the agent.
Another potential problem can occur when the fakeness of an instruction-trajectory pair can be inferred solely from the initial observation (without the agent's action). This could be used by the discriminator to surely identify negative examples since only these are false conditioned. For example, when the instruction requires reaching an object that is not present in the scene, the discriminator could spot that and reward the agent lowly regardless of its actions.
To address the aforementioned problems, we rely on auxiliary rewards provided by additional discriminators. We design these trainable auxiliary rewards to discourage degenerate behaviours that may arise from using the FC technique.
In this section we propose two possible ways of training an additional discriminator, which are $Blank\textit{Condition- }$ ing and Done Detector. Together with previously proposed Agent FC and Expert FC techniques, four different discriminators can be trained. Each discriminator is trained separately to minimize its own loss and provides rewards of the form $-\log(1-D_{\theta}(\cdot)).$ In practice, a small $\epsilon>0$ is added to the discriminator's predictions to make the rewards bounded.
Blank Conditioning When Expert FC is used, the dis-
criminator's training distribution includes only expert trajectories. As result, its predictions on agent-generated trajectories may be less accurate which can be exploited by the policy. To address this problem, we propose to train an extra discriminator which distinguishes between agent and expert trajectories but is not conditioned. It means that the agent is additionally rewarded for generating trajectories which resemble the expert (for $a$ goal/instruction, not the given goal/instruction). To get the highest rewards, its trajectories have to match the distribution of expert trajectories on which the main discriminator is trained. The auxiliary reward for Blank Conditioning is based on the discriminator trained with the following loss:
(8)
$$\begin{aligned}L_{D}^{B}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{agent}}-\log(1-D_{\theta}(c_{\emptyset},\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c_{\emptyset},\tau)),\end{aligned}$$
|
[
[
0.08660888671875,
0.0694580078125,
0.4755859375,
0.0694580078125,
0.4755859375,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.521484375,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.521484375,
0.08331298828125
],
[
0.0849609375,
0.08331298828125,
0.476318359375,
0.08331298828125,
0.476318359375,
0.09661865234375,
0.0849609375,
0.09661865234375
],
[
0.5185546875,
0.08209228515625,
0.9111328125,
0.08331298828125,
0.9111328125,
0.09783935546875,
0.5185546875,
0.09661865234375
],
[
0.0849609375,
0.09661865234375,
0.47802734375,
0.09783935546875,
0.47802734375,
0.11236572265625,
0.0849609375,
0.111083984375
],
[
0.51953125,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.51953125,
0.111083984375
],
[
0.08905029296875,
0.11297607421875,
0.476318359375,
0.11297607421875,
0.476318359375,
0.12371826171875,
0.08905029296875,
0.12371826171875
],
[
0.51953125,
0.11236572265625,
0.8408203125,
0.11236572265625,
0.8408203125,
0.1256103515625,
0.51953125,
0.1256103515625
],
[
0.08660888671875,
0.1256103515625,
0.4794921875,
0.1256103515625,
0.4794921875,
0.14013671875,
0.08660888671875,
0.14013671875
],
[
0.53125,
0.1319580078125,
0.9140625,
0.1331787109375,
0.9140625,
0.1546630859375,
0.53125,
0.154052734375
],
[
0.0882568359375,
0.1383056640625,
0.476318359375,
0.141357421875,
0.476318359375,
0.1558837890625,
0.0882568359375,
0.15283203125
],
[
0.08660888671875,
0.1546630859375,
0.2254638671875,
0.1546630859375,
0.2254638671875,
0.1685791015625,
0.08660888671875,
0.1685791015625
],
[
0.583984375,
0.1572265625,
0.68115234375,
0.1590576171875,
0.68115234375,
0.17041015625,
0.583984375,
0.1685791015625
],
[
0.689453125,
0.1546630859375,
0.80224609375,
0.1546630859375,
0.80224609375,
0.169189453125,
0.689453125,
0.169189453125
],
[
0.13232421875,
0.1749267578125,
0.4345703125,
0.175537109375,
0.4345703125,
0.1976318359375,
0.13232421875,
0.1962890625
],
[
0.45751953125,
0.179931640625,
0.47802734375,
0.179931640625,
0.47802734375,
0.1925048828125,
0.45751953125,
0.1925048828125
],
[
0.5390625,
0.1787109375,
0.9111328125,
0.1787109375,
0.9111328125,
0.1925048828125,
0.5390625,
0.1925048828125
],
[
0.521484375,
0.1925048828125,
0.91015625,
0.1925048828125,
0.91015625,
0.205810546875,
0.521484375,
0.205810546875
],
[
0.1829833984375,
0.199462890625,
0.406005859375,
0.199462890625,
0.406005859375,
0.2171630859375,
0.1829833984375,
0.2171630859375
],
[
0.51953125,
0.20703125,
0.91259765625,
0.20703125,
0.91259765625,
0.2203369140625,
0.51953125,
0.2203369140625
],
[
0.0849609375,
0.222900390625,
0.4794921875,
0.222900390625,
0.4794921875,
0.236083984375,
0.0849609375,
0.236083984375
],
[
0.51953125,
0.2203369140625,
0.77783203125,
0.2215576171875,
0.77783203125,
0.23291015625,
0.51953125,
0.231689453125
],
[
0.08416748046875,
0.236083984375,
0.47802734375,
0.2374267578125,
0.47802734375,
0.250732421875,
0.08416748046875,
0.25
],
[
0.0849609375,
0.25244140625,
0.393798828125,
0.25244140625,
0.393798828125,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.51953125,
0.246826171875,
0.7060546875,
0.246826171875,
0.7060546875,
0.261474609375,
0.51953125,
0.261474609375
],
[
0.10455322265625,
0.266357421875,
0.476318359375,
0.266357421875,
0.476318359375,
0.276611328125,
0.10455322265625,
0.276611328125
],
[
0.521484375,
0.26513671875,
0.90869140625,
0.26513671875,
0.90869140625,
0.279052734375,
0.521484375,
0.279052734375
],
[
0.08416748046875,
0.27783203125,
0.47802734375,
0.279052734375,
0.47802734375,
0.292236328125,
0.08416748046875,
0.291748046875
],
[
0.52294921875,
0.281005859375,
0.91259765625,
0.281005859375,
0.91259765625,
0.291748046875,
0.52294921875,
0.291748046875
],
[
0.08416748046875,
0.291748046875,
0.476318359375,
0.292236328125,
0.476318359375,
0.306884765625,
0.08416748046875,
0.30615234375
],
[
0.51953125,
0.29345703125,
0.9111328125,
0.29345703125,
0.9111328125,
0.306884765625,
0.51953125,
0.306884765625
],
[
0.0849609375,
0.306884765625,
0.404296875,
0.306884765625,
0.404296875,
0.32080078125,
0.0849609375,
0.32080078125
],
[
0.521484375,
0.309326171875,
0.91015625,
0.309326171875,
0.91015625,
0.3193359375,
0.521484375,
0.3193359375
],
[
0.521484375,
0.322509765625,
0.91259765625,
0.322509765625,
0.91259765625,
0.333251953125,
0.521484375,
0.333251953125
],
[
0.0849609375,
0.330810546875,
0.271240234375,
0.33203125,
0.271240234375,
0.3466796875,
0.0849609375,
0.34521484375
],
[
0.51953125,
0.335205078125,
0.6689453125,
0.335205078125,
0.6689453125,
0.348388671875,
0.51953125,
0.348388671875
],
[
0.08660888671875,
0.35107421875,
0.47802734375,
0.35107421875,
0.47802734375,
0.3642578125,
0.08660888671875,
0.3642578125
],
[
0.5361328125,
0.348388671875,
0.9140625,
0.348388671875,
0.9140625,
0.3623046875,
0.5361328125,
0.3623046875
],
[
0.0882568359375,
0.3662109375,
0.47802734375,
0.3662109375,
0.47802734375,
0.376953125,
0.0882568359375,
0.376953125
],
[
0.51953125,
0.363037109375,
0.91259765625,
0.363037109375,
0.91259765625,
0.376953125,
0.51953125,
0.376953125
],
[
0.0882568359375,
0.37890625,
0.4794921875,
0.37890625,
0.4794921875,
0.392578125,
0.0882568359375,
0.392578125
],
[
0.521484375,
0.375732421875,
0.91259765625,
0.375732421875,
0.91259765625,
0.389404296875,
0.521484375,
0.389404296875
],
[
0.08660888671875,
0.392578125,
0.476318359375,
0.392578125,
0.476318359375,
0.406005859375,
0.08660888671875,
0.406005859375
],
[
0.521484375,
0.392578125,
0.90869140625,
0.392578125,
0.90869140625,
0.40283203125,
0.521484375,
0.40283203125
],
[
0.08660888671875,
0.4072265625,
0.476318359375,
0.4072265625,
0.476318359375,
0.42041015625,
0.08660888671875,
0.42041015625
],
[
0.521484375,
0.406005859375,
0.9111328125,
0.406005859375,
0.9111328125,
0.416748046875,
0.521484375,
0.416748046875
],
[
0.08660888671875,
0.42041015625,
0.476318359375,
0.419189453125,
0.476318359375,
0.43310546875,
0.08660888671875,
0.434326171875
],
[
0.521484375,
0.41845703125,
0.91259765625,
0.41845703125,
0.91259765625,
0.431884765625,
0.521484375,
0.431884765625
],
[
0.08660888671875,
0.434326171875,
0.47802734375,
0.434326171875,
0.47802734375,
0.447509765625,
0.08660888671875,
0.447509765625
],
[
0.52294921875,
0.434326171875,
0.91259765625,
0.434326171875,
0.91259765625,
0.4443359375,
0.52294921875,
0.4443359375
],
[
0.0849609375,
0.448974609375,
0.419189453125,
0.448974609375,
0.419189453125,
0.462158203125,
0.0849609375,
0.462158203125
],
[
0.51953125,
0.4462890625,
0.8349609375,
0.4462890625,
0.8349609375,
0.460205078125,
0.51953125,
0.460205078125
],
[
0.10296630859375,
0.462158203125,
0.476318359375,
0.462158203125,
0.476318359375,
0.47607421875,
0.10296630859375,
0.47607421875
],
[
0.53759765625,
0.460205078125,
0.9111328125,
0.460205078125,
0.9111328125,
0.473388671875,
0.53759765625,
0.473388671875
],
[
0.0882568359375,
0.47607421875,
0.47802734375,
0.47607421875,
0.47802734375,
0.4892578125,
0.0882568359375,
0.4892578125
],
[
0.521484375,
0.473388671875,
0.9111328125,
0.473388671875,
0.9111328125,
0.4873046875,
0.521484375,
0.4873046875
],
[
0.0882568359375,
0.4912109375,
0.4755859375,
0.4912109375,
0.4755859375,
0.501953125,
0.0882568359375,
0.501953125
],
[
0.521484375,
0.4892578125,
0.91015625,
0.4892578125,
0.91015625,
0.5,
0.521484375,
0.5
],
[
0.0882568359375,
0.5048828125,
0.476318359375,
0.5048828125,
0.476318359375,
0.51513671875,
0.0882568359375,
0.51513671875
],
[
0.51953125,
0.501953125,
0.90673828125,
0.501953125,
0.90673828125,
0.51513671875,
0.51953125,
0.51513671875
],
[
0.08660888671875,
0.517578125,
0.476318359375,
0.517578125,
0.476318359375,
0.53076171875,
0.08660888671875,
0.53076171875
],
[
0.53759765625,
0.517578125,
0.90869140625,
0.517578125,
0.90869140625,
0.52783203125,
0.53759765625,
0.52783203125
],
[
0.08660888671875,
0.5322265625,
0.47802734375,
0.5322265625,
0.47802734375,
0.54541015625,
0.08660888671875,
0.54541015625
],
[
0.52294921875,
0.53076171875,
0.91015625,
0.53076171875,
0.91015625,
0.54150390625,
0.52294921875,
0.54150390625
],
[
0.0849609375,
0.54541015625,
0.226318359375,
0.546875,
0.226318359375,
0.55810546875,
0.0849609375,
0.556640625
],
[
0.52294921875,
0.54345703125,
0.91259765625,
0.54345703125,
0.91259765625,
0.556640625,
0.52294921875,
0.556640625
],
[
0.10296630859375,
0.5595703125,
0.476318359375,
0.5595703125,
0.476318359375,
0.57275390625,
0.10296630859375,
0.57275390625
],
[
0.52392578125,
0.5595703125,
0.90869140625,
0.5595703125,
0.90869140625,
0.5693359375,
0.52392578125,
0.5693359375
],
[
0.0882568359375,
0.57470703125,
0.476318359375,
0.57470703125,
0.476318359375,
0.58544921875,
0.0882568359375,
0.58544921875
],
[
0.52294921875,
0.57275390625,
0.9111328125,
0.57275390625,
0.9111328125,
0.58349609375,
0.52294921875,
0.58349609375
],
[
0.08660888671875,
0.5869140625,
0.4755859375,
0.58837890625,
0.4755859375,
0.599609375,
0.08660888671875,
0.5986328125
],
[
0.521484375,
0.58544921875,
0.9140625,
0.58544921875,
0.9140625,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.0849609375,
0.60107421875,
0.476318359375,
0.599609375,
0.476318359375,
0.61279296875,
0.0849609375,
0.6142578125
],
[
0.51953125,
0.599609375,
0.9140625,
0.599609375,
0.9140625,
0.61279296875,
0.51953125,
0.61279296875
],
[
0.08660888671875,
0.6142578125,
0.4755859375,
0.6142578125,
0.4755859375,
0.62744140625,
0.08660888671875,
0.62744140625
],
[
0.521484375,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.521484375,
0.626953125
],
[
0.0882568359375,
0.6298828125,
0.476318359375,
0.6298828125,
0.476318359375,
0.64013671875,
0.0882568359375,
0.64013671875
],
[
0.51953125,
0.625,
0.5849609375,
0.626953125,
0.5849609375,
0.6416015625,
0.51953125,
0.6396484375
],
[
0.08660888671875,
0.64306640625,
0.4755859375,
0.64306640625,
0.4755859375,
0.65380859375,
0.08660888671875,
0.65380859375
],
[
0.521484375,
0.64599609375,
0.91015625,
0.64599609375,
0.91015625,
0.66015625,
0.521484375,
0.66015625
],
[
0.08660888671875,
0.65576171875,
0.47802734375,
0.65576171875,
0.47802734375,
0.66943359375,
0.08660888671875,
0.66943359375
],
[
0.52294921875,
0.662109375,
0.90869140625,
0.662109375,
0.90869140625,
0.67236328125,
0.52294921875,
0.67236328125
],
[
0.0882568359375,
0.671875,
0.1871337890625,
0.671875,
0.1871337890625,
0.68310546875,
0.0882568359375,
0.68310546875
],
[
0.521484375,
0.6748046875,
0.91259765625,
0.6748046875,
0.91259765625,
0.68798828125,
0.521484375,
0.68798828125
],
[
0.08660888671875,
0.689453125,
0.47802734375,
0.689453125,
0.47802734375,
0.70263671875,
0.08660888671875,
0.70263671875
],
[
0.51953125,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.51953125,
0.7021484375
],
[
0.0882568359375,
0.705078125,
0.4755859375,
0.705078125,
0.4755859375,
0.71533203125,
0.0882568359375,
0.71533203125
],
[
0.51953125,
0.7021484375,
0.91259765625,
0.7021484375,
0.91259765625,
0.71533203125,
0.51953125,
0.71533203125
],
[
0.08660888671875,
0.71728515625,
0.47802734375,
0.71728515625,
0.47802734375,
0.73095703125,
0.08660888671875,
0.73095703125
],
[
0.52294921875,
0.71728515625,
0.9111328125,
0.71728515625,
0.9111328125,
0.72802734375,
0.52294921875,
0.72802734375
],
[
0.0849609375,
0.73193359375,
0.47802734375,
0.73193359375,
0.47802734375,
0.74560546875,
0.0849609375,
0.74560546875
],
[
0.52294921875,
0.73095703125,
0.9111328125,
0.73095703125,
0.9111328125,
0.7412109375,
0.52294921875,
0.7412109375
],
[
0.0849609375,
0.744140625,
0.4794921875,
0.74560546875,
0.4794921875,
0.7587890625,
0.0849609375,
0.75830078125
],
[
0.51953125,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.51953125,
0.7568359375
],
[
0.0849609375,
0.75830078125,
0.346435546875,
0.7587890625,
0.346435546875,
0.7734375,
0.0849609375,
0.77294921875
],
[
0.521484375,
0.7587890625,
0.9111328125,
0.7587890625,
0.9111328125,
0.76953125,
0.521484375,
0.76953125
],
[
0.52294921875,
0.77294921875,
0.9111328125,
0.77294921875,
0.9111328125,
0.78271484375,
0.52294921875,
0.78271484375
],
[
0.09722900390625,
0.78076171875,
0.4794921875,
0.78076171875,
0.4794921875,
0.8017578125,
0.09722900390625,
0.8017578125
],
[
0.51953125,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.51953125,
0.798828125
],
[
0.146240234375,
0.80029296875,
0.370849609375,
0.80029296875,
0.370849609375,
0.8212890625,
0.146240234375,
0.8212890625
],
[
0.52294921875,
0.80029296875,
0.90869140625,
0.80029296875,
0.90869140625,
0.81103515625,
0.52294921875,
0.81103515625
],
[
0.521484375,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.521484375,
0.8271484375
],
[
0.08660888671875,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84130859375,
0.08660888671875,
0.84130859375
],
[
0.51953125,
0.8271484375,
0.72607421875,
0.8271484375,
0.72607421875,
0.84033203125,
0.51953125,
0.84033203125
],
[
0.0882568359375,
0.8466796875,
0.4794921875,
0.8466796875,
0.4794921875,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.564453125,
0.8466796875,
0.86767578125,
0.8486328125,
0.86767578125,
0.87060546875,
0.564453125,
0.86865234375
],
[
0.890625,
0.8515625,
0.91259765625,
0.8515625,
0.91259765625,
0.8642578125,
0.890625,
0.8642578125
],
[
0.0882568359375,
0.86279296875,
0.47802734375,
0.86279296875,
0.47802734375,
0.87353515625,
0.0882568359375,
0.87353515625
],
[
0.615234375,
0.86865234375,
0.84619140625,
0.8671875,
0.84619140625,
0.88623046875,
0.615234375,
0.88720703125
],
[
0.0849609375,
0.87548828125,
0.4794921875,
0.87548828125,
0.4794921875,
0.8896484375,
0.0849609375,
0.8896484375
]
] |
[
"uses the environment's true reward signal to identify the suc.",
"unaltered. It means that positive examples stay as in Expert",
"cessful trajectories generated by the agent and filter them",
"FC formulation but negatives are constructed from the fake",
"out. In this setting, the buffer of negative examples Bagent",
"conditioned agent trajectories instead of expert ones. In this",
"contains only unsuccessful trajectories.Thediscriminator's",
"case, the discriminator's loss is given as follows:",
"loss is given as in Equation 4 but instead of Bagent we use",
"LB(0) =E(c,T)~Bagent,c~SI(c) -log(1 - De(c,T) (7)",
"Boracle which is a subset of Bagent that contains unsuccess-",
"ful trajectories only:",
"+E(c.T)~Berper",
"- log(De(c, T)",
"LB(0) =E(c,T)~Berete - log(1 - Do(c, T)",
"(5)",
"We note that the policy that generates agent trajectories is",
"always conditioned on the original instruction, for both Ex-",
"+E(ce,T)~Bezpert log(De(c, T)",
"pert FC and Agent FC. The fake instruction-trajectory pairs",
"Note that when the OF method is applied, the environment",
"are used to train the discriminator only.",
"rewards are used solely to filter out successful agent trajec-",
"tories-the agent does not have access to them",
"3.3Auxiliary Rewards",
"WefoundthatusingOFsignificantlyimprovestheperfor-",
"The FC technique greatly reduces the percentage of FN. Un.",
"mance,and reduces sample complexityby atleast an order",
"fortunately,thediscriminatoris nolonger trained on agent",
"of magnitude. This confirms our hypothesis that false nega-",
"trajectories with the original instructions, and hence the re-",
"tives have a detrimental effect on GAIL training.",
"wardsthatthe discriminatorissuestothepolicy canpoten-",
"tiallybeless accurate.These reward inaccuracies could be",
"3.2Fake Conditioning",
"exploited by the agent.",
"Since the Oracle Filtering technique requires access to en-",
"Another potential problem can occur when the fakeness of",
"vironmentrewards, itis only a diagnostic method suited to",
"an instruction-trajectory pair can be inferred solely from the",
"assess the impact of false negatives. We propose a technique",
"initial observation (without the agent's action). This could",
"that does not need environment rewards to address the FN",
"be used bythe discriminator to surelyidentifynegative ex",
"problem. Our technique is aimed at tasks where the policy",
"amples sinceonlythesearefalseconditioned.Forexample",
"is goal-conditioned. In our case, we assume that the policy",
"when the instruction requires reaching an object that is not",
"is conditioned on language instructions but the technique is",
"present inthe scene,the discriminator could spot that and",
" general and can be applied in any multi-goal setup.",
"reward the agent lowly regardless of its actions.",
"First, we maintain a set of possible language instructions",
"To address the aforementioned problems, we rely on aux-",
"S initialized from all the unique instructions in the expert",
"iliary rewards provided by additional discriminators. We de-",
"demonstrations and,optionally,updated with the instruc-",
"sign these trainable auxiliaryrewards todiscourage degener-",
"tions collected when the agent interacts with the environ-",
"ate behaviours that may arise from using the FC technique.",
"ment. Secondly, for each instruction-trajectory pair (c, T)",
"Inthissectionweproposetwopossiblewaysoftrain",
"we replace its original instruction c with that is randomly",
"ing an additional discriminator,whichareBlankCondition-",
"sampledfromSfcl",
"ing and Done Detector.Together with previously proposed",
"We call this technique Fake Conditioning (FC). It is mo",
"AgentFCandExpertFCtechnigues,fourdifferentdiscrim",
"tivated by thefact that the success of a trajectory is condi-",
"inators can be trained.Each discriminator is trained sep-",
"tioned ontheinstruction.Therefore,foreachtrajectory,by",
"arately to minimize its own loss and provides rewards of",
"replacing the instruction with a new one, we produce a new",
"the form -log(1 - De(). In practice, a small e > 0 is",
"instruction-trajectory pair that is very likely to be unsuccess-",
"added to the discriminator's predictions to make the rewards",
"ful,evenwhenitis successfullyconditioned ontheoriginal",
"bounded.",
"instruction.The FC technique canbeused to prepare dis-",
"Blank Conditioning When Expert FC is used, the dis-",
"criminator training data with greatly limited number of the",
"criminator's training distribution includes only expert tra",
"false negatives.",
"jectories. As result, its predictions on agent-generated tra-",
"Expert FCWe call Expert FC a technique whereby the",
"jectories may be less accurate which can be exploited by the",
"discriminatoris trained to distinguishthe experttrajectories",
"policy. To address this problem, we propose to train an extra",
"with the original instructions from fake conditioned expert",
"discriminator whichdistinguishes between agentand expert",
"trajectories. In this case, the agent trajectories are not used",
"trajectories but is not conditioned.It means that the agent",
"at all, and hence the training is no longer adversarial. The",
"is additionally rewarded for generating trajectories which",
"discriminator's loss is given as follows:",
"resembletheexpert(foragoal/instruction,notthegiven",
"goal/instruction). To get the highest rewards.its trajecto-",
"LB(0) =E(c,T)~Beapert, ~S(c) - 1og(1 - De(c, T) (6)",
"ries have to match the distribution of expert trajectories on",
"+E(e,T)~Berpert - log(De(c, T),",
"which the main discriminator is trained. The auxiliary re",
"ward for Blank Conditioning is based on the discriminator",
"where (c, T) is fake conditioned instruction-trajectory pair.",
"trained with the following loss:",
"Agent FC The FC technique can also be used with agent",
"LB(0) =E(c,T)~Bagent - log(1 - De(co, T)",
"(8)",
"trajectories. In this case, which we call Agent FC, all agent",
"+E(c,T)~Berpert - log(De(co, T),",
"trajectories are fake conditioned while expert ones are left"
] |
{
"arxiv_id": "2002.00412",
"language": "en",
"page_idx": 2
}
|
||
2002.00412-en-3
|
2002.00412-en-3.jpg
|
Table 1: Comparison of positives and negatives used for different type of discriminators.
Blank
<table>
<tbody>
<tr>
<th rowspan="1">$\mathbf{P}_{0}\mathbf{sitives}$ M 4. $\Delta c$</th>
<th rowspan="1">trajectories instructions trajectories</th>
<th colspan="3">expert original</th>
<th> </th>
<th rowspan="1">$\exp\exp(x)$ blank agent</th>
</tr>
<tr>
<td>Negatives</td>
<td>trajectories instructions</td>
<td>agent original</td>
<td>agent fake</td>
<td>$\exp$ert fake</td>
<td>incomplete expert original</td>
<td>$\operatorname{agent}$ blank</td>
</tr>
</tbody>
</table>
$\mathbf{Conventional}$
$FC$
$FC$
Done Detector
### 4 $\textbf{Related Work}$
where $c_{\emptyset}$ is a fixed blank instruction that masks the original
ones$^{\mathrm{l} }.$ We call this technique Blank Conditioning.
Unfortunately, Blank Conditioning can be impeded by the FN problem as well. However, we hypothesize that the repercussions will be less severe because only an auxiliary reward is affected. In particular, even when the agent masters the task and generates a lot of FN in the auxiliary discriminator's training data, the main reward will still be informative since the main discriminator uses training data without FN. Done Detector The second auxiliary discriminator is trained to detect if a given trajectory is finished. Its negative examples are unfinished expert demonstrations (with the original instruction), while the finished ones are positive examples. In both positive and negative examples the original instruction is used. Hence, the loss forces the discriminator to focus on both the trajectory and the instruction to understand if the goal is reached. The loss is the following:
Instruction Following with Natural Language We consider the problem of following natural language instructions. Recently, reinforcement learning methods have been applied to make progress in this area across a variety of environments (Chaplot et al. 2017; Hermann et al. 2017; Mirowski et al. 2018). In these approaches, an agent is rewarded when it successfully follows an instruction. However, designing the respective reward function is non-trivial (Luketina et al. 2019). Formalizing the completion semantics, implicit in natural language instruction, has been an open research problem since early efforts (Winograd 1972). Human handcrafting of the reward function becomes increasingly infeasible as the number of instructions and complexity of the environment scale. Applying IL methods in this area is therefore promising (Duvallet, Kollar, and Stentz 2013).
(9)
$$\begin{aligned}L_{D}^{D}(\theta)=&\mathbb{E}_{(c,\tau)\sim B_{expert}^{sub}}-\log(1-D_{\theta}(c,\tau))\\&+\mathbb{E}_{(c,\tau)\sim B_{expert}}-\log(D_{\theta}(c,\tau)),\end{aligned}$$
Imitation Learning As mentioned before, IL alleviates the need to handcraft reward functions by learning policies from demonstrations. In this work we consider Behavioral Cloning (BC, (Pomerleau 1989)) as a baseline and our method improves Generative Adversarial Imitation Learning (GAIL) (Ho and Ermon 2016) which is closely related to Inverse Reinforcement Learning (Abbeel and Ng 2004b). While all imitation learning methods can be used for learning instruction-conditioned policies (Mei, Bansal, and Walter 2016; Fu et al. 2019; Bahdanau et al. 2019), prior literature only features instructions that can be verified by the final observation and does not discuss the topic of converging to near-perfect performance (with the exception of (Bahdanau et al. 2019), see more on that below).
False Negatives We found that the False Negative problem significantly impedes adversarial imitation learning. As such, the issue was also found problematic in prior work (Zolna et al. 2019). However, prior work lacked a diagnostic tool to assess the impact of the FN problem and estimate how much performance can be improved if the problem is addressed. Zolna et al. (2019) proposed a heuristic method to address the FN problem called Actor Early Stopping. This method terminates well performing episodes to limit the number of successful agent states used to train the discriminator. AGILE proposed by Bahdanau et al.(2019) is another heuristic solution which deals with False Negatives in the very similar setup of training a reward model. The method filters out states that were most highly rewarded by the discriminator to not train on them. Finally, Xu and Denil (2019) reformulate discriminator training as PositiveUnlabeled (PU) learning which can be seen as a way of combating False Negatives.
where $B_expert$ is built out of $B_expert$ and contains all pos- $\begin{array}{l}\text{sible incomplete sub-trajectories only (i.e. each trajectory}\\\mathrm{from~}B_{expert}^{sub}\text{comes from the element of }B_{expert}\text{ but is cut}\\\mathrm{before~they~have~finished).~We~call~this~discriminator~}Done\end{array}$ Detector.
We note that the loss for Done Detector does not depend on the agent's performance, and hence the extra discriminator is not trained adversarially. There are also no false negatives since all negative examples are constructed in a controlled way and are never successful.
Blank Conditioning is supposed to address the main limitation of Expert FC (lack of agent trajectories in discriminator training data), while Done Detector is for Agent FC (fake instructions with agent trajectories may be too easy to identify). However, both auxiliary rewards can be used with any FC method. It means that we can train all four discriminators simultaneously (one for each loss introduced in this subsection) to get four different rewards which can be mixed together to train the agent.
In Table 1, we provide the overview of positives and negatives used to train discriminators presented in this work. OF is skipped as it is only a diagnostic tool.
In theory the conditioning on the instruction could be skipped but using $c_\emptyset$ makes sharing the architecture straightforward-the extra reward model can be implemented as the extra discriminator head.
|
[
[
0.2115478515625,
0.0662841796875,
0.7861328125,
0.0662841796875,
0.7861328125,
0.0802001953125,
0.2115478515625,
0.0802001953125
],
[
0.31201171875,
0.09661865234375,
0.5849609375,
0.09661865234375,
0.5849609375,
0.10986328125,
0.31201171875,
0.10986328125
],
[
0.5986328125,
0.09783935546875,
0.6953125,
0.09783935546875,
0.6953125,
0.10797119140625,
0.5986328125,
0.10797119140625
],
[
0.70361328125,
0.0947265625,
0.84619140625,
0.09661865234375,
0.84619140625,
0.111083984375,
0.70361328125,
0.10919189453125
],
[
0.2320556640625,
0.12060546875,
0.29736328125,
0.12060546875,
0.29736328125,
0.1307373046875,
0.2320556640625,
0.1307373046875
],
[
0.49267578125,
0.12060546875,
0.53173828125,
0.12060546875,
0.53173828125,
0.1319580078125,
0.49267578125,
0.1319580078125
],
[
0.7607421875,
0.12060546875,
0.80126953125,
0.12060546875,
0.80126953125,
0.1319580078125,
0.7607421875,
0.1319580078125
],
[
0.486083984375,
0.1307373046875,
0.5390625,
0.1307373046875,
0.5390625,
0.1458740234375,
0.486083984375,
0.1458740234375
],
[
0.7587890625,
0.1307373046875,
0.7998046875,
0.1307373046875,
0.7998046875,
0.1458740234375,
0.7587890625,
0.1458740234375
],
[
0.2320556640625,
0.1331787109375,
0.3056640625,
0.1331787109375,
0.3056640625,
0.14453125,
0.2320556640625,
0.14453125
],
[
0.2320556640625,
0.1546630859375,
0.2998046875,
0.1546630859375,
0.2998046875,
0.1666259765625,
0.2320556640625,
0.1666259765625
],
[
0.34716796875,
0.1546630859375,
0.3857421875,
0.1546630859375,
0.3857421875,
0.1666259765625,
0.34716796875,
0.1666259765625
],
[
0.439453125,
0.1546630859375,
0.4755859375,
0.1546630859375,
0.4755859375,
0.1673583984375,
0.439453125,
0.1673583984375
],
[
0.521484375,
0.1546630859375,
0.560546875,
0.1546630859375,
0.560546875,
0.1666259765625,
0.521484375,
0.1666259765625
],
[
0.59326171875,
0.1558837890625,
0.69775390625,
0.1558837890625,
0.69775390625,
0.1666259765625,
0.59326171875,
0.1666259765625
],
[
0.76123046875,
0.1558837890625,
0.79833984375,
0.1558837890625,
0.79833984375,
0.1673583984375,
0.76123046875,
0.1673583984375
],
[
0.1519775390625,
0.160400390625,
0.2156982421875,
0.160400390625,
0.2156982421875,
0.1717529296875,
0.1519775390625,
0.1717529296875
],
[
0.1519775390625,
0.1256103515625,
0.2099609375,
0.1256103515625,
0.2099609375,
0.136962890625,
0.1519775390625,
0.136962890625
],
[
0.339111328125,
0.1654052734375,
0.393798828125,
0.1654052734375,
0.393798828125,
0.179931640625,
0.339111328125,
0.179931640625
],
[
0.521484375,
0.1654052734375,
0.55810546875,
0.1654052734375,
0.55810546875,
0.18115234375,
0.521484375,
0.18115234375
],
[
0.619140625,
0.1654052734375,
0.67333984375,
0.1654052734375,
0.67333984375,
0.179931640625,
0.619140625,
0.179931640625
],
[
0.76123046875,
0.1666259765625,
0.7998046875,
0.1666259765625,
0.7998046875,
0.179931640625,
0.76123046875,
0.179931640625
],
[
0.441162109375,
0.1666259765625,
0.473876953125,
0.1666259765625,
0.473876953125,
0.1787109375,
0.441162109375,
0.1787109375
],
[
0.2303466796875,
0.1673583984375,
0.30224609375,
0.1673583984375,
0.30224609375,
0.177978515625,
0.2303466796875,
0.177978515625
],
[
0.0882568359375,
0.205810546875,
0.4794921875,
0.205810546875,
0.4794921875,
0.2197265625,
0.0882568359375,
0.2197265625
],
[
0.6689453125,
0.2052001953125,
0.7890625,
0.2052001953125,
0.7890625,
0.2197265625,
0.6689453125,
0.2197265625
],
[
0.6396484375,
0.212158203125,
0.65185546875,
0.20263671875,
0.6611328125,
0.210205078125,
0.64892578125,
0.2197265625
],
[
0.08416748046875,
0.2183837890625,
0.41259765625,
0.2203369140625,
0.41259765625,
0.23486328125,
0.08416748046875,
0.23291015625
],
[
0.51953125,
0.2259521484375,
0.9111328125,
0.2259521484375,
0.9111328125,
0.240478515625,
0.51953125,
0.240478515625
],
[
0.10455322265625,
0.2374267578125,
0.4755859375,
0.2374267578125,
0.4755859375,
0.2474365234375,
0.10455322265625,
0.2474365234375
],
[
0.521484375,
0.240478515625,
0.9111328125,
0.240478515625,
0.9111328125,
0.25390625,
0.521484375,
0.25390625
],
[
0.08660888671875,
0.25,
0.476318359375,
0.25,
0.476318359375,
0.26318359375,
0.08660888671875,
0.26318359375
],
[
0.521484375,
0.25390625,
0.91259765625,
0.25390625,
0.91259765625,
0.267578125,
0.521484375,
0.267578125
],
[
0.0849609375,
0.264404296875,
0.47802734375,
0.264404296875,
0.47802734375,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.51953125,
0.267578125,
0.9111328125,
0.267578125,
0.9111328125,
0.281005859375,
0.51953125,
0.281005859375
],
[
0.08660888671875,
0.279052734375,
0.476318359375,
0.279052734375,
0.476318359375,
0.2890625,
0.08660888671875,
0.2890625
],
[
0.521484375,
0.28271484375,
0.91259765625,
0.28271484375,
0.91259765625,
0.29345703125,
0.521484375,
0.29345703125
],
[
0.0849609375,
0.291748046875,
0.47802734375,
0.291748046875,
0.47802734375,
0.304931640625,
0.0849609375,
0.304931640625
],
[
0.51953125,
0.29541015625,
0.91259765625,
0.29541015625,
0.91259765625,
0.309326171875,
0.51953125,
0.309326171875
],
[
0.08660888671875,
0.306884765625,
0.476318359375,
0.306884765625,
0.476318359375,
0.317626953125,
0.08660888671875,
0.317626953125
],
[
0.521484375,
0.309326171875,
0.91259765625,
0.309326171875,
0.91259765625,
0.322509765625,
0.521484375,
0.322509765625
],
[
0.08660888671875,
0.32080078125,
0.47216796875,
0.32080078125,
0.47216796875,
0.330810546875,
0.08660888671875,
0.330810546875
],
[
0.51953125,
0.323974609375,
0.9140625,
0.323974609375,
0.9140625,
0.337158203125,
0.51953125,
0.337158203125
],
[
0.08660888671875,
0.34033203125,
0.47802734375,
0.34033203125,
0.47802734375,
0.354248046875,
0.08660888671875,
0.354248046875
],
[
0.521484375,
0.337158203125,
0.9111328125,
0.33837890625,
0.9111328125,
0.349853515625,
0.521484375,
0.348388671875
],
[
0.08416748046875,
0.352783203125,
0.47802734375,
0.354736328125,
0.47802734375,
0.36865234375,
0.08416748046875,
0.3662109375
],
[
0.521484375,
0.35107421875,
0.91259765625,
0.35107421875,
0.91259765625,
0.3642578125,
0.521484375,
0.3642578125
],
[
0.08660888671875,
0.36865234375,
0.47802734375,
0.36865234375,
0.47802734375,
0.3818359375,
0.08660888671875,
0.3818359375
],
[
0.52294921875,
0.3662109375,
0.90869140625,
0.3662109375,
0.90869140625,
0.376953125,
0.52294921875,
0.376953125
],
[
0.08660888671875,
0.3818359375,
0.4755859375,
0.3818359375,
0.4755859375,
0.395751953125,
0.08660888671875,
0.395751953125
],
[
0.521484375,
0.37890625,
0.91015625,
0.380126953125,
0.91015625,
0.391357421875,
0.521484375,
0.39013671875
],
[
0.08660888671875,
0.397705078125,
0.476318359375,
0.397705078125,
0.476318359375,
0.40771484375,
0.08660888671875,
0.40771484375
],
[
0.521484375,
0.392578125,
0.91259765625,
0.392578125,
0.91259765625,
0.406005859375,
0.521484375,
0.406005859375
],
[
0.08660888671875,
0.410400390625,
0.4794921875,
0.410400390625,
0.4794921875,
0.423583984375,
0.08660888671875,
0.423583984375
],
[
0.52294921875,
0.40771484375,
0.91015625,
0.40771484375,
0.91015625,
0.41845703125,
0.52294921875,
0.41845703125
],
[
0.0849609375,
0.423583984375,
0.476318359375,
0.423583984375,
0.476318359375,
0.4375,
0.0849609375,
0.4375
],
[
0.521484375,
0.42041015625,
0.85693359375,
0.42041015625,
0.85693359375,
0.434326171875,
0.521484375,
0.434326171875
],
[
0.08416748046875,
0.436279296875,
0.438720703125,
0.4375,
0.438720703125,
0.451904296875,
0.08416748046875,
0.45068359375
],
[
0.521484375,
0.439453125,
0.91259765625,
0.439453125,
0.91259765625,
0.45263671875,
0.521484375,
0.45263671875
],
[
0.1356201171875,
0.4609375,
0.43310546875,
0.4609375,
0.43310546875,
0.48291015625,
0.1356201171875,
0.48291015625
],
[
0.521484375,
0.455078125,
0.91015625,
0.455078125,
0.91015625,
0.46533203125,
0.521484375,
0.46533203125
],
[
0.45751953125,
0.464111328125,
0.47802734375,
0.464111328125,
0.47802734375,
0.4765625,
0.45751953125,
0.4765625
],
[
0.51953125,
0.46728515625,
0.9111328125,
0.46728515625,
0.9111328125,
0.48095703125,
0.51953125,
0.48095703125
],
[
0.181396484375,
0.48291015625,
0.406982421875,
0.48291015625,
0.406982421875,
0.50390625,
0.181396484375,
0.50390625
],
[
0.51953125,
0.481689453125,
0.91259765625,
0.481689453125,
0.91259765625,
0.49560546875,
0.51953125,
0.49560546875
],
[
0.521484375,
0.496826171875,
0.91015625,
0.496826171875,
0.91015625,
0.5068359375,
0.521484375,
0.5068359375
],
[
0.08251953125,
0.50830078125,
0.48046875,
0.50830078125,
0.48046875,
0.5322265625,
0.08251953125,
0.5322265625
],
[
0.51953125,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51953125,
0.52294921875
],
[
0.0849609375,
0.52783203125,
0.476318359375,
0.52783203125,
0.476318359375,
0.54150390625,
0.0849609375,
0.54150390625
],
[
0.521484375,
0.5244140625,
0.9111328125,
0.5244140625,
0.9111328125,
0.53515625,
0.521484375,
0.53515625
],
[
0.160888671875,
0.54052734375,
0.4794921875,
0.54150390625,
0.4794921875,
0.55615234375,
0.160888671875,
0.5546875
],
[
0.52294921875,
0.53857421875,
0.91015625,
0.53857421875,
0.91015625,
0.548828125,
0.52294921875,
0.548828125
],
[
0.146240234375,
0.55029296875,
0.16259765625,
0.55029296875,
0.16259765625,
0.5537109375,
0.146240234375,
0.5537109375
],
[
0.0849609375,
0.54150390625,
0.156005859375,
0.5390625,
0.156005859375,
0.55322265625,
0.0849609375,
0.5546875
],
[
0.0849609375,
0.5537109375,
0.4794921875,
0.5546875,
0.4794921875,
0.5693359375,
0.0849609375,
0.568359375
],
[
0.521484375,
0.55029296875,
0.91259765625,
0.55029296875,
0.91259765625,
0.564453125,
0.521484375,
0.564453125
],
[
0.0882568359375,
0.57080078125,
0.148681640625,
0.57080078125,
0.148681640625,
0.58056640625,
0.0882568359375,
0.58056640625
],
[
0.51953125,
0.56494140625,
0.91015625,
0.56494140625,
0.91015625,
0.5791015625,
0.51953125,
0.5791015625
],
[
0.10296630859375,
0.583984375,
0.4794921875,
0.583984375,
0.4794921875,
0.59765625,
0.10296630859375,
0.59765625
],
[
0.521484375,
0.580078125,
0.9111328125,
0.580078125,
0.9111328125,
0.59033203125,
0.521484375,
0.59033203125
],
[
0.51953125,
0.59228515625,
0.9140625,
0.59228515625,
0.9140625,
0.60595703125,
0.51953125,
0.60595703125
],
[
0.08660888671875,
0.5986328125,
0.476318359375,
0.5986328125,
0.476318359375,
0.6123046875,
0.08660888671875,
0.6123046875
],
[
0.08416748046875,
0.611328125,
0.47802734375,
0.6123046875,
0.47802734375,
0.62548828125,
0.08416748046875,
0.625
],
[
0.51953125,
0.60595703125,
0.91259765625,
0.60595703125,
0.91259765625,
0.619140625,
0.51953125,
0.619140625
],
[
0.08660888671875,
0.626953125,
0.47802734375,
0.626953125,
0.47802734375,
0.64013671875,
0.08660888671875,
0.64013671875
],
[
0.521484375,
0.6220703125,
0.7646484375,
0.6220703125,
0.7646484375,
0.6318359375,
0.521484375,
0.6318359375
],
[
0.0849609375,
0.6396484375,
0.330078125,
0.6396484375,
0.330078125,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.51953125,
0.63818359375,
0.91015625,
0.6396484375,
0.91015625,
0.65380859375,
0.51953125,
0.65283203125
],
[
0.10296630859375,
0.65478515625,
0.47802734375,
0.65478515625,
0.47802734375,
0.66845703125,
0.10296630859375,
0.66845703125
],
[
0.521484375,
0.65283203125,
0.91259765625,
0.65283203125,
0.91259765625,
0.66650390625,
0.521484375,
0.66650390625
],
[
0.0849609375,
0.66845703125,
0.476318359375,
0.66845703125,
0.476318359375,
0.681640625,
0.0849609375,
0.681640625
],
[
0.521484375,
0.66845703125,
0.91259765625,
0.66845703125,
0.91259765625,
0.681640625,
0.521484375,
0.681640625
],
[
0.0849609375,
0.681640625,
0.47802734375,
0.681640625,
0.47802734375,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.521484375,
0.681640625,
0.91015625,
0.681640625,
0.91015625,
0.69580078125,
0.521484375,
0.69580078125
],
[
0.0882568359375,
0.69775390625,
0.47802734375,
0.69775390625,
0.47802734375,
0.70849609375,
0.0882568359375,
0.70849609375
],
[
0.51953125,
0.69580078125,
0.9111328125,
0.69580078125,
0.9111328125,
0.708984375,
0.51953125,
0.708984375
],
[
0.08660888671875,
0.708984375,
0.476318359375,
0.708984375,
0.476318359375,
0.72265625,
0.08660888671875,
0.72265625
],
[
0.521484375,
0.7109375,
0.9111328125,
0.7109375,
0.9111328125,
0.7216796875,
0.521484375,
0.7216796875
],
[
0.08660888671875,
0.72607421875,
0.476318359375,
0.72607421875,
0.476318359375,
0.736328125,
0.08660888671875,
0.736328125
],
[
0.51953125,
0.7236328125,
0.91259765625,
0.7236328125,
0.91259765625,
0.7373046875,
0.51953125,
0.7373046875
],
[
0.08660888671875,
0.7392578125,
0.476318359375,
0.7392578125,
0.476318359375,
0.75,
0.08660888671875,
0.75
],
[
0.51953125,
0.7373046875,
0.91259765625,
0.7373046875,
0.91259765625,
0.75048828125,
0.51953125,
0.75048828125
],
[
0.08660888671875,
0.751953125,
0.47802734375,
0.751953125,
0.47802734375,
0.76513671875,
0.08660888671875,
0.76513671875
],
[
0.521484375,
0.75,
0.9140625,
0.75,
0.9140625,
0.7646484375,
0.521484375,
0.7646484375
],
[
0.08416748046875,
0.7666015625,
0.259033203125,
0.76513671875,
0.259033203125,
0.77880859375,
0.08416748046875,
0.77978515625
],
[
0.521484375,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.521484375,
0.77880859375
],
[
0.101318359375,
0.78076171875,
0.476318359375,
0.78076171875,
0.476318359375,
0.79443359375,
0.101318359375,
0.79443359375
],
[
0.52294921875,
0.77978515625,
0.91015625,
0.77978515625,
0.91015625,
0.79052734375,
0.52294921875,
0.79052734375
],
[
0.0849609375,
0.79443359375,
0.47802734375,
0.79345703125,
0.47802734375,
0.806640625,
0.0849609375,
0.80810546875
],
[
0.521484375,
0.79248046875,
0.91015625,
0.79248046875,
0.91015625,
0.80615234375,
0.521484375,
0.80615234375
],
[
0.0849609375,
0.8095703125,
0.348876953125,
0.8095703125,
0.348876953125,
0.82275390625,
0.0849609375,
0.82275390625
],
[
0.51953125,
0.80615234375,
0.91015625,
0.80615234375,
0.91015625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.521484375,
0.82080078125,
0.91259765625,
0.82080078125,
0.91259765625,
0.833984375,
0.521484375,
0.833984375
],
[
0.1094970703125,
0.83837890625,
0.47802734375,
0.83837890625,
0.47802734375,
0.8515625,
0.1094970703125,
0.8515625
],
[
0.51953125,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.51953125,
0.84765625
],
[
0.0849609375,
0.85107421875,
0.47802734375,
0.85107421875,
0.47802734375,
0.8642578125,
0.0849609375,
0.8642578125
],
[
0.51953125,
0.84765625,
0.9111328125,
0.84765625,
0.9111328125,
0.861328125,
0.51953125,
0.861328125
],
[
0.08660888671875,
0.86572265625,
0.47802734375,
0.86572265625,
0.47802734375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51953125,
0.85986328125,
0.9111328125,
0.861328125,
0.9111328125,
0.87548828125,
0.51953125,
0.875
],
[
0.0849609375,
0.876953125,
0.12091064453125,
0.876953125,
0.12091064453125,
0.88818359375,
0.0849609375,
0.88818359375
],
[
0.51953125,
0.87548828125,
0.67626953125,
0.87548828125,
0.67626953125,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"Table 1: Comparison of positives and negatives used for different type of discriminators.",
"ConventionalAgentFCExpertFC",
"DoneDetector",
"Blank Conditioning",
"trajectories",
"expert",
"expert",
"original",
"blank",
"instructions",
"trajectories",
"agent",
"agent",
"expert",
"incompleteexpert",
"agent",
"Negatives",
"Positives",
"original",
"fake",
"original",
"blank",
"fake",
"instructions",
"where Co is a fixed blank instruction that masks the original",
"Related Work",
"X",
"ones'. We call this technique Blank Conditioning.",
"Instruction Following with Natural Language We con-",
"Unfortunately.Blank Conditioning can beimpeded by",
"sider the problem of following natural language instructions",
"the FN problem as well. However, we hypothesize that the",
"Recently, reinforcement learning methods have been applied",
"repercussions will be less severe because only an auxiliary",
"to make progress in this area across a variety of environ-",
"rewardisaffected.Inparticular,evenwhentheagentmasters",
"ments (Chaplot et al.2017: Hermann et al.2017:Mirowski",
"the task and generates a lot of FN in the auxiliary discrimina-",
"et al. 2018). In these approaches, an agent is rewarded when",
"tor's training data, the main reward will still be informative",
"it successfully follows an instruction. However, designing",
"sincethemaindiscriminator usestraining data without FN.",
"the respective reward function is non-trivial (Luketina et",
"Done DetectorThe second auxiliary discriminator is",
"al.2019).Formalizing the completion semantics,implicit",
"trained to detect if a given trajectory is finished. Its nega-",
"in natural language instruction, has been an open research",
"tive examples are unfinished expert demonstrations (with the",
"problem sinceearlyefforts(Winograd 1972).Humanhand",
"original instruction), while the finished ones are positive ex-",
"craftingof therewardfunctionbecomes increasinglyinfea-",
"amples.Inbothpositiveandnegativeexamplesthe original",
"sible as the number of instructions and complexity of the en-",
"instruction is used. Hence, the loss forces the discriminator",
"vironment scale.ApplyingIL methods in this area is there-",
"to focus on both the trajectory and the instruction to under-",
"fore promising (Duvallet, Kollar, and Stentz 2013).",
" stand if the goal is reached. The loss is the following:",
"Imitation Learning As mentioned before, IL alleviates",
"LB(0) =E(c,T)~Bapert - log(1 - De(c, T)",
"the need to handcraft reward functions by learningpoli-",
"(9)",
"cies from demonstrations. In this work we consider Behav-",
"+E(c,T)~Berpert -log(De(c, T),",
"ioral Cloning (BC, (Pomerleau 1989)) as a baseline and our",
"method improves Generative AdversarialImitationLearn-",
"where Beubert is built out of Berpert and contains all pos-",
"ing (GAIL) (Ho and Ermon 2016) which is closely related",
"sible incomplete sub-trajectories only (i.e. each trajectory",
"toInverseReinforcementLearning(AbbeelandNg2004b)",
"rt comes from the element of Berpert but is cut",
"Whileallimitationlearningmethodscanbeusedforlearn-",
"Te",
"fromBsub",
"before they have finished). We call this discriminator Done",
"ing instruction-conditioned policies (Mei, Bansal, and Wal-",
"Detector.",
"ter 2016; Fu et al. 2019; Bahdanau et al. 2019), prior litera-",
"We note that the loss for Done Detector does not depend",
"tureonlyfeaturesinstructionsthat canbeverifiedbythefinal",
"observation and does not discuss the topic of converging to",
"on the agent's performance, and hence the extra discrimina-",
"tor is not trained adversarially.There are also no false neg-",
"near-perfect performance (with the exception of (Bahdanau",
"atives since all negative examples are constructed in a con-",
"et al.2019),see more on that below).",
"trolled way and are never successful.",
"False NegativesWe found that the False Negative prob.",
"Blank Conditioning is supposed to address the main lim-",
"lem significantly impedes adversarial imitation learning.As",
"itation of Expert FC (lack of agent trajectories in discrim-",
"such, the issue was also found problematic in prior work",
"inator training data), while Done Detector is for Agent FC",
"(Zolna et al. 2019). However, prior work lacked a diagnos-",
"(fakeinstructionswithagent trajectories maybetoo easyto",
"tic tool to assess the impact of the FN problem and esti-",
"identify). However, both auxiliary rewards can be used with",
"matehowmuchperformancecanbeimprovedif theprob-",
"anyFC method.It means thatwe cantrain all four discrim-",
"lem is addressed. Zolna et al. (2019) proposed a heuristic",
"inatorssimultaneously(oneforeachlossintroducedinthis",
"method to address the FN problem called Actor Early Stop-",
"subsection) to get four different rewards which can be mixed",
"ping. This method terminates well performing episodes to",
"together to train the agent.",
"limit the number of successful agent states used to train the",
"In Table I, we provide the overview of positives and nega-",
"discriminator.AGILEproposedbyBahdanau etal.(2019)",
"tives used to train discriminators presented in this work.OF",
"is another heuristic solution which deals with False Nega-",
"is skipped as it is only a diagnostic tool.",
"tives in the very similar setup of training a reward model",
"The method filters out states that were most highly rewarded",
"'In theory the conditioning on the instruction could be skipped",
"by the discriminator to not train on them. Finally, Xu and",
"but using co makes sharing the architecture straightforward - the",
"Denil (2019) reformulate discriminator training as Positive-",
"extrareward model can be implemented as the extradiscriminator",
"Unlabeled (PU) learning which can be seen as a way of com-",
"head.",
"bating False Negatives."
] |
{
"arxiv_id": "2002.00412",
"language": "en",
"page_idx": 3
}
|
||
2002.00412-en-4
|
2002.00412-en-4.jpg
|
<FigureHere>
Figure 2: Model architectures. Best viewed in color.
## 5 $\textbf{Experiment setup}$
5.1 The BabyAI Environment
BabyAI is a deterministic, partially-observable 2D gridworld based on MiniGrid. The natural-looking language instruction is supplied as a string. At each time step, the agent receives a visual input of the $7\times7$ cells in front of it. Precise details of the visual input are described in (ChevalierBoisvert et al. 2018). We report performances on the following four single-room levels.
GoToRedBall is the simplest level among the four and can be solved purely from visual inputs. An agent is tasked with go to the/a red ball in the presence of other distracting objects of other colors and shapes.
GoToLocal extends GoToRedBall. An agent is tasked with instructions of the form go to the <color> <object>,where <color> and <object> are no longer limited to red and ball. As a consequence, this level can no longer be solved by using just visual inputs-a given instruction has to be parsed and understood.
PickupLoc and PutNextLocal are harder tasks which additionally require to equip the reward model with memory. In PickupLoc the instructions refer to objects not just by their type and color but also by their location relative to the initial position of the agent, e.g. go to the red ball in front of you.In the case of PutNextLocal, the instruction requires putting an object next to another object, each described with a type and color, e.g. put the blue ball next to the blue key. Since putting object1 next to object2 is not the same as putting object2 next to ob ject1, the reward model needs to remember the trajectory in order to give a correct reward at the end.
## 5.2 $\textbf{Episode Termination}$
In the BabyAI platform the episode is terminated when the task is solved, or after a maximum allowed number of steps is reached, which is a standard practice for similar RL setups. However, in the GAIL setup, no environment reward is available. The naive approach would be to always run every episode for the maximum number of steps.
We let the agent perform a special Done action. At the same time, all expert demonstrations are padded with this action as the last one (i.e. just before termination of the episode). It makes the agent perform the Done action when it considers the task to be done and it can be used to terminate the episode during training. It turns out that it significantly speeds up the training procedure as it collects richer data $\hat{(\text{more episodes for the same number of frames experienced}}$ by the agent). Done termination is used for all methods, including baselines.
The approach is related to actor early stopping introduced by Zolna et al. (2019). The difference is that we terminate episodes earlier based on the policy's predictions while the cited method terminates based on the discriminator's predictions.
## 53 Architecture and Training
In our experiments, as mentioned in Section 2.1, the agent is provided with a non-zero reward only when it finishes an episode by performing a Done action described in the previous section. This makes rewards sparse and similar to the original rewards used in BabyAI platform to train RL algorithms. This choice allows us to use the agent architecture and the hyperparameters from the original BabyAI paper (Chevalier-Boisvert et al. 2018) without extra tuning.
Actor-critic The model underlying the RL agent is presented in Figure 2(a). It consists of standard components to predict the next action based on the current observation, the memory of the past observations and the instruction. It uses GRU to encode the instruction and a convolutional network with two batch-normalized FiLM layers to jointly process the observation and the instruction. An LSTM memory is used to integrate representations produced by the FiLM module at each step. It uses a memory of 128 units and encodes the instruction with a unidirectional GRU.
Discriminator The discriminator architecture (see Figure 2(b)) is similar to the actor-critic model for symmetry and efficiency reasons. We just replace the final actor and critic layers by 3-layers MLP, placed after the FiLM block. The LSTM here takes as input not only the FiLM embedding but also a one-hot action vector (both concatenated).
Training In all our experiments, we use the same hyperparameters for actor training as used for RL training in (Chevalier-Boisvert et al.2018). For both models, we used the Adam optimizer with the hyperparameters $\alpha=$ $10^{- 4}, \beta _{1}$ = $0. 9, \dot{\beta } _{2}$ = 0.999 and $\epsilon$ = $10^{- 5}.$ We used the Proximal Policy Optimization algorithm with parallelized data collection: we performed 4 epochs of PPO using 64 rollouts of length 40 collected with multiple processes. We
|
[
[
0.188720703125,
0.06500244140625,
0.373291015625,
0.06756591796875,
0.373291015625,
0.085205078125,
0.188720703125,
0.08331298828125
],
[
0.79833984375,
0.0738525390625,
0.833984375,
0.0738525390625,
0.833984375,
0.08587646484375,
0.79833984375,
0.08587646484375
],
[
0.54150390625,
0.09344482421875,
0.57421875,
0.09344482421875,
0.57421875,
0.10479736328125,
0.54150390625,
0.10479736328125
],
[
0.619140625,
0.0921630859375,
0.68896484375,
0.0921630859375,
0.68896484375,
0.1029052734375,
0.619140625,
0.1029052734375
],
[
0.8056640625,
0.09661865234375,
0.82861328125,
0.09661865234375,
0.82861328125,
0.10919189453125,
0.8056640625,
0.10919189453125
],
[
0.55615234375,
0.111083984375,
0.560546875,
0.111083984375,
0.560546875,
0.12249755859375,
0.55615234375,
0.12249755859375
],
[
0.0849609375,
0.09661865234375,
0.328369140625,
0.09783935546875,
0.328369140625,
0.11236572265625,
0.0849609375,
0.111083984375
],
[
0.0882568359375,
0.12371826171875,
0.47802734375,
0.12371826171875,
0.47802734375,
0.136962890625,
0.0882568359375,
0.136962890625
],
[
0.8056640625,
0.12060546875,
0.82666015625,
0.12060546875,
0.82666015625,
0.1331787109375,
0.8056640625,
0.1331787109375
],
[
0.56884765625,
0.1307373046875,
0.58251953125,
0.1307373046875,
0.58251953125,
0.14013671875,
0.56884765625,
0.14013671875
],
[
0.64794921875,
0.1319580078125,
0.65673828125,
0.1319580078125,
0.65673828125,
0.136962890625,
0.64794921875,
0.136962890625
],
[
0.08660888671875,
0.1383056640625,
0.4794921875,
0.1383056640625,
0.4794921875,
0.1514892578125,
0.08660888671875,
0.1514892578125
],
[
0.8056640625,
0.14453125,
0.82861328125,
0.14453125,
0.82861328125,
0.1572265625,
0.8056640625,
0.1572265625
],
[
0.08660888671875,
0.154052734375,
0.47802734375,
0.154052734375,
0.47802734375,
0.1641845703125,
0.08660888671875,
0.1641845703125
],
[
0.08660888671875,
0.1673583984375,
0.476318359375,
0.1673583984375,
0.476318359375,
0.177978515625,
0.08660888671875,
0.177978515625
],
[
0.59228515625,
0.1654052734375,
0.62841796875,
0.1654052734375,
0.62841796875,
0.175537109375,
0.59228515625,
0.175537109375
],
[
0.6396484375,
0.169189453125,
0.65673828125,
0.169189453125,
0.65673828125,
0.173583984375,
0.6396484375,
0.173583984375
],
[
0.80224609375,
0.1673583984375,
0.8310546875,
0.1673583984375,
0.8310546875,
0.1787109375,
0.80224609375,
0.1787109375
],
[
0.76953125,
0.17236328125,
0.78515625,
0.17236328125,
0.78515625,
0.175537109375,
0.76953125,
0.175537109375
],
[
0.55810546875,
0.17041015625,
0.56640625,
0.1685791015625,
0.56787109375,
0.17236328125,
0.5595703125,
0.173583984375
],
[
0.0849609375,
0.179931640625,
0.4794921875,
0.179931640625,
0.4794921875,
0.1932373046875,
0.0849609375,
0.1932373046875
],
[
0.08660888671875,
0.1932373046875,
0.47802734375,
0.1932373046875,
0.47802734375,
0.20703125,
0.08660888671875,
0.20703125
],
[
0.57861328125,
0.1900634765625,
0.63818359375,
0.1900634765625,
0.63818359375,
0.2020263671875,
0.57861328125,
0.2020263671875
],
[
0.74658203125,
0.1944580078125,
0.775390625,
0.1944580078125,
0.775390625,
0.20263671875,
0.74658203125,
0.20263671875
],
[
0.0849609375,
0.2083740234375,
0.26953125,
0.2083740234375,
0.26953125,
0.2215576171875,
0.0849609375,
0.2215576171875
],
[
0.79150390625,
0.210205078125,
0.85546875,
0.210205078125,
0.85546875,
0.2247314453125,
0.79150390625,
0.2247314453125
],
[
0.101318359375,
0.2235107421875,
0.4794921875,
0.2235107421875,
0.4794921875,
0.2374267578125,
0.101318359375,
0.2374267578125
],
[
0.640625,
0.2298583984375,
0.67236328125,
0.2298583984375,
0.67236328125,
0.2412109375,
0.640625,
0.2412109375
],
[
0.0849609375,
0.238037109375,
0.47802734375,
0.238037109375,
0.47802734375,
0.251953125,
0.0849609375,
0.251953125
],
[
0.08660888671875,
0.251953125,
0.47802734375,
0.251953125,
0.47802734375,
0.26513671875,
0.08660888671875,
0.26513671875
],
[
0.564453125,
0.251953125,
0.58837890625,
0.251953125,
0.58837890625,
0.260009765625,
0.564453125,
0.260009765625
],
[
0.6435546875,
0.251953125,
0.6689453125,
0.251953125,
0.6689453125,
0.260009765625,
0.6435546875,
0.260009765625
],
[
0.8544921875,
0.251953125,
0.88232421875,
0.251953125,
0.88232421875,
0.26318359375,
0.8544921875,
0.26318359375
],
[
0.0849609375,
0.266357421875,
0.36767578125,
0.266357421875,
0.36767578125,
0.27978515625,
0.0849609375,
0.27978515625
],
[
0.5458984375,
0.26953125,
0.6044921875,
0.271484375,
0.6044921875,
0.28271484375,
0.5458984375,
0.281005859375
],
[
0.640625,
0.271484375,
0.6708984375,
0.271484375,
0.6708984375,
0.28271484375,
0.640625,
0.28271484375
],
[
0.77294921875,
0.2734375,
0.7998046875,
0.2734375,
0.7998046875,
0.2822265625,
0.77294921875,
0.2822265625
],
[
0.8544921875,
0.274658203125,
0.88134765625,
0.274658203125,
0.88134765625,
0.2822265625,
0.8544921875,
0.2822265625
],
[
0.10455322265625,
0.28271484375,
0.4794921875,
0.28271484375,
0.4794921875,
0.29345703125,
0.10455322265625,
0.29345703125
],
[
0.57421875,
0.29345703125,
0.619140625,
0.29345703125,
0.619140625,
0.300537109375,
0.57421875,
0.300537109375
],
[
0.63134765625,
0.291748046875,
0.6845703125,
0.291748046875,
0.6845703125,
0.3017578125,
0.63134765625,
0.3017578125
],
[
0.75634765625,
0.294189453125,
0.81201171875,
0.294189453125,
0.81201171875,
0.304931640625,
0.75634765625,
0.304931640625
],
[
0.8544921875,
0.29541015625,
0.88134765625,
0.29541015625,
0.88134765625,
0.3037109375,
0.8544921875,
0.3037109375
],
[
0.5361328125,
0.294189453125,
0.5703125,
0.294189453125,
0.5703125,
0.298583984375,
0.5361328125,
0.298583984375
],
[
0.08660888671875,
0.296630859375,
0.47802734375,
0.296630859375,
0.47802734375,
0.31005859375,
0.08660888671875,
0.31005859375
],
[
0.0882568359375,
0.309326171875,
0.4794921875,
0.309326171875,
0.4794921875,
0.322509765625,
0.0882568359375,
0.322509765625
],
[
0.65625,
0.304931640625,
0.66259765625,
0.304931640625,
0.66259765625,
0.30810546875,
0.65625,
0.30810546875
],
[
0.74658203125,
0.31494140625,
0.78759765625,
0.31494140625,
0.78759765625,
0.322509765625,
0.74658203125,
0.322509765625
],
[
0.78515625,
0.31494140625,
0.830078125,
0.31494140625,
0.830078125,
0.322509765625,
0.78515625,
0.322509765625
],
[
0.84326171875,
0.313232421875,
0.896484375,
0.313232421875,
0.896484375,
0.323974609375,
0.84326171875,
0.323974609375
],
[
0.0849609375,
0.323974609375,
0.4794921875,
0.323974609375,
0.4794921875,
0.337158203125,
0.0849609375,
0.337158203125
],
[
0.86767578125,
0.326904296875,
0.875,
0.326904296875,
0.875,
0.33203125,
0.86767578125,
0.33203125
],
[
0.0849609375,
0.3359375,
0.47802734375,
0.3359375,
0.47802734375,
0.349853515625,
0.0849609375,
0.349853515625
],
[
0.56884765625,
0.34228515625,
0.650390625,
0.34228515625,
0.650390625,
0.354248046875,
0.56884765625,
0.354248046875
],
[
0.77294921875,
0.34228515625,
0.87744140625,
0.34228515625,
0.87744140625,
0.354736328125,
0.77294921875,
0.354736328125
],
[
0.0882568359375,
0.352783203125,
0.41259765625,
0.352783203125,
0.41259765625,
0.363037109375,
0.0882568359375,
0.363037109375
],
[
0.10296630859375,
0.367431640625,
0.47802734375,
0.367431640625,
0.47802734375,
0.380615234375,
0.10296630859375,
0.380615234375
],
[
0.5458984375,
0.3662109375,
0.88623046875,
0.3662109375,
0.88623046875,
0.380126953125,
0.5458984375,
0.380126953125
],
[
0.08660888671875,
0.380615234375,
0.47802734375,
0.380615234375,
0.47802734375,
0.39453125,
0.08660888671875,
0.39453125
],
[
0.08660888671875,
0.396484375,
0.47802734375,
0.396484375,
0.47802734375,
0.4072265625,
0.08660888671875,
0.4072265625
],
[
0.0849609375,
0.4091796875,
0.47802734375,
0.4091796875,
0.47802734375,
0.42236328125,
0.0849609375,
0.42236328125
],
[
0.5185546875,
0.4072265625,
0.7646484375,
0.40771484375,
0.7646484375,
0.42236328125,
0.5185546875,
0.421630859375
],
[
0.0849609375,
0.423583984375,
0.4794921875,
0.423583984375,
0.4794921875,
0.4375,
0.0849609375,
0.4375
],
[
0.51953125,
0.4267578125,
0.9140625,
0.4267578125,
0.9140625,
0.43994140625,
0.51953125,
0.43994140625
],
[
0.08660888671875,
0.438232421875,
0.476318359375,
0.438232421875,
0.476318359375,
0.448974609375,
0.08660888671875,
0.448974609375
],
[
0.5185546875,
0.44140625,
0.9140625,
0.44140625,
0.9140625,
0.455078125,
0.5185546875,
0.455078125
],
[
0.0849609375,
0.45068359375,
0.476318359375,
0.45068359375,
0.476318359375,
0.464111328125,
0.0849609375,
0.464111328125
],
[
0.51953125,
0.455078125,
0.9140625,
0.455078125,
0.9140625,
0.468505859375,
0.51953125,
0.468505859375
],
[
0.0849609375,
0.46533203125,
0.47802734375,
0.46533203125,
0.47802734375,
0.479248046875,
0.0849609375,
0.479248046875
],
[
0.521484375,
0.47021484375,
0.9111328125,
0.47021484375,
0.9111328125,
0.48095703125,
0.521484375,
0.48095703125
],
[
0.0849609375,
0.479248046875,
0.47802734375,
0.479736328125,
0.47802734375,
0.4912109375,
0.0849609375,
0.490478515625
],
[
0.51953125,
0.48291015625,
0.9111328125,
0.48291015625,
0.9111328125,
0.496826171875,
0.51953125,
0.496826171875
],
[
0.08660888671875,
0.492431640625,
0.4794921875,
0.492431640625,
0.4794921875,
0.505859375,
0.08660888671875,
0.505859375
],
[
0.51953125,
0.496826171875,
0.9140625,
0.496826171875,
0.9140625,
0.51025390625,
0.51953125,
0.51025390625
],
[
0.0882568359375,
0.50830078125,
0.476318359375,
0.50830078125,
0.476318359375,
0.51806640625,
0.0882568359375,
0.51806640625
],
[
0.51953125,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51953125,
0.5234375
],
[
0.0849609375,
0.51953125,
0.4794921875,
0.51953125,
0.4794921875,
0.53271484375,
0.0849609375,
0.53271484375
],
[
0.51953125,
0.52294921875,
0.86767578125,
0.5234375,
0.86767578125,
0.53857421875,
0.51953125,
0.537109375
],
[
0.0849609375,
0.5341796875,
0.1168212890625,
0.5341796875,
0.1168212890625,
0.546875,
0.0849609375,
0.546875
],
[
0.51953125,
0.54345703125,
0.91015625,
0.54345703125,
0.91015625,
0.556640625,
0.51953125,
0.556640625
],
[
0.51953125,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.51953125,
0.5712890625
],
[
0.0849609375,
0.56640625,
0.289306640625,
0.56640625,
0.289306640625,
0.58056640625,
0.0849609375,
0.58056640625
],
[
0.51953125,
0.57275390625,
0.91259765625,
0.57275390625,
0.91259765625,
0.5859375,
0.51953125,
0.5859375
],
[
0.521484375,
0.5869140625,
0.9111328125,
0.5869140625,
0.9111328125,
0.59765625,
0.521484375,
0.59765625
],
[
0.08660888671875,
0.5947265625,
0.476318359375,
0.5947265625,
0.476318359375,
0.60498046875,
0.08660888671875,
0.60498046875
],
[
0.51953125,
0.599609375,
0.9111328125,
0.599609375,
0.9111328125,
0.61279296875,
0.51953125,
0.61279296875
],
[
0.08660888671875,
0.609375,
0.476318359375,
0.609375,
0.476318359375,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.5185546875,
0.6123046875,
0.91259765625,
0.6142578125,
0.91259765625,
0.62744140625,
0.5185546875,
0.62548828125
],
[
0.0849609375,
0.6220703125,
0.476318359375,
0.6220703125,
0.476318359375,
0.63525390625,
0.0849609375,
0.63525390625
],
[
0.0849609375,
0.6337890625,
0.476318359375,
0.6337890625,
0.476318359375,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.521484375,
0.62890625,
0.9111328125,
0.62890625,
0.9111328125,
0.6396484375,
0.521484375,
0.6396484375
],
[
0.5185546875,
0.6416015625,
0.91259765625,
0.6416015625,
0.91259765625,
0.65478515625,
0.5185546875,
0.65478515625
],
[
0.0849609375,
0.6494140625,
0.47802734375,
0.6494140625,
0.47802734375,
0.66357421875,
0.0849609375,
0.66357421875
],
[
0.5185546875,
0.65478515625,
0.91259765625,
0.65478515625,
0.91259765625,
0.66943359375,
0.5185546875,
0.66943359375
],
[
0.08416748046875,
0.662109375,
0.365234375,
0.662109375,
0.365234375,
0.67529296875,
0.08416748046875,
0.67529296875
],
[
0.521484375,
0.67041015625,
0.83642578125,
0.67041015625,
0.83642578125,
0.68115234375,
0.521484375,
0.68115234375
],
[
0.10296630859375,
0.6787109375,
0.4794921875,
0.6787109375,
0.4794921875,
0.6923828125,
0.10296630859375,
0.6923828125
],
[
0.0849609375,
0.693359375,
0.47802734375,
0.693359375,
0.47802734375,
0.70703125,
0.0849609375,
0.70703125
],
[
0.5185546875,
0.6875,
0.9111328125,
0.68798828125,
0.9111328125,
0.70263671875,
0.5185546875,
0.7021484375
],
[
0.0849609375,
0.70556640625,
0.47802734375,
0.70556640625,
0.47802734375,
0.7197265625,
0.0849609375,
0.7197265625
],
[
0.51953125,
0.70263671875,
0.91259765625,
0.70263671875,
0.91259765625,
0.71630859375,
0.51953125,
0.71630859375
],
[
0.08660888671875,
0.72265625,
0.47802734375,
0.72265625,
0.47802734375,
0.73291015625,
0.08660888671875,
0.73291015625
],
[
0.51953125,
0.71728515625,
0.91259765625,
0.71728515625,
0.91259765625,
0.73095703125,
0.51953125,
0.73095703125
],
[
0.0849609375,
0.73486328125,
0.4794921875,
0.73486328125,
0.4794921875,
0.74853515625,
0.0849609375,
0.74853515625
],
[
0.51953125,
0.72998046875,
0.91259765625,
0.72998046875,
0.91259765625,
0.74365234375,
0.51953125,
0.74365234375
],
[
0.0849609375,
0.74853515625,
0.47802734375,
0.74853515625,
0.47802734375,
0.76220703125,
0.0849609375,
0.76220703125
],
[
0.51953125,
0.74365234375,
0.91259765625,
0.74560546875,
0.91259765625,
0.7587890625,
0.51953125,
0.7568359375
],
[
0.0849609375,
0.76220703125,
0.4794921875,
0.76220703125,
0.4794921875,
0.77587890625,
0.0849609375,
0.77587890625
],
[
0.51953125,
0.7587890625,
0.861328125,
0.7587890625,
0.861328125,
0.76953125,
0.51953125,
0.76953125
],
[
0.0849609375,
0.7744140625,
0.47802734375,
0.7744140625,
0.47802734375,
0.78857421875,
0.0849609375,
0.78857421875
],
[
0.521484375,
0.77880859375,
0.9111328125,
0.77880859375,
0.9111328125,
0.79248046875,
0.521484375,
0.79248046875
],
[
0.0849609375,
0.79052734375,
0.47802734375,
0.79052734375,
0.47802734375,
0.8037109375,
0.0849609375,
0.8037109375
],
[
0.51953125,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.8037109375,
0.2091064453125,
0.8046875,
0.2091064453125,
0.81640625,
0.0849609375,
0.81494140625
],
[
0.51953125,
0.80615234375,
0.9140625,
0.80615234375,
0.9140625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.10455322265625,
0.82080078125,
0.476318359375,
0.82080078125,
0.476318359375,
0.833984375,
0.10455322265625,
0.833984375
],
[
0.5185546875,
0.8193359375,
0.896484375,
0.82080078125,
0.896484375,
0.833984375,
0.5185546875,
0.83349609375
],
[
0.0849609375,
0.833984375,
0.4794921875,
0.833984375,
0.4794921875,
0.84765625,
0.0849609375,
0.84765625
],
[
0.51953125,
0.83349609375,
0.9140625,
0.83349609375,
0.9140625,
0.84765625,
0.51953125,
0.84765625
],
[
0.08660888671875,
0.849609375,
0.47802734375,
0.849609375,
0.47802734375,
0.85986328125,
0.08660888671875,
0.85986328125
],
[
0.51953125,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.51953125,
0.861328125
],
[
0.0849609375,
0.8623046875,
0.47802734375,
0.8623046875,
0.47802734375,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.51953125,
0.861328125,
0.9140625,
0.861328125,
0.9140625,
0.875,
0.51953125,
0.875
],
[
0.0849609375,
0.876953125,
0.125,
0.876953125,
0.125,
0.88818359375,
0.0849609375,
0.88818359375
],
[
0.5185546875,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.5185546875,
0.8896484375
]
] |
[
"5Experiment setup",
"Reward",
"Action",
"Value estimation",
"FC3",
"+",
"5.1 TheBabyAIEnvironment",
"BabyAI is a deterministic, partially-observable 2D grid-",
"FC2",
"FC",
"FC",
"world based on MiniGrid. The natural-looking language in-",
"FC1",
"struction is supplied asa string.Ateachtime step,the agent",
"receives a visual input of the 7 × 7 cells in front of it. Pre-",
"LSTM",
"",
"LSTM",
"",
"",
"cise details of the visual input are described in (Chevalier-",
"Boisvert et al. 2018). We report performances on the follow-",
"FiLMBlock",
"Action",
"ing four single-room levels.",
"FiLMBlock",
"GoToRedBall is the simplest level among the four and can",
"Conv3",
"be solved purely from visual inputs. An agent is tasked with",
"go to the/a red ball in the presence of other dis-",
"GRU",
"Conv2",
"Conv3",
"tracting objects of other colors and shapes.",
"Embedding",
"Conv1",
"GRU",
"Conv2",
"GoToLocal extends GoToRedBall.An agent is tasked",
"instruction",
"Observation",
"Embedding",
"Conv1",
"anguage",
"with instructions of the form go to the <color>",
"<object>,where <color> and <object> are no",
"",
"Languagei",
"instruction",
"Observation",
"longer limited to red and ball. As a consequence, this",
"S.",
"level can no longer be solved by using just visual inputs -- a",
"(a) Generator",
"(b)Discriminator",
"giveninstructionhastobeparsedandunderstood.",
"PickupLoc and PutNextLocal are harder tasks which ad-",
"Figure 2: Model architectures. Best viewed in color.",
"ditionally require to equip the reward model with mem-",
"ory.InPickupLoctheinstructionsrefertoobjects notjust",
"by their type and color but also by their location relative",
"5.3Architecture and Training",
"to the initial position of the agent, e.g. go to the red",
"In our experiments, as mentioned in Section 2.1, the agent",
"ballin front ofyou.InthecaseofPutNextLocal.",
"is provided with a non-zero reward only when it finishes an",
"the instruction requires putting an object next to another ob-",
"episode by performing a Done action described in the pre-",
"ject, each described with a type and color, e.g- put the",
"vious section. This makes rewards sparse and similar to the",
"blue ball next to the blue key.Sinceputting",
"original rewards used in BabyAI platform to train RL algo-",
"object1 next to object2 is not the same as putting",
"rithms. This choice allows us to use the agent architecture",
"object2nexttoobject1,therewardmodelneedstore-",
"and the hyperparameters from the original BabyAI paper",
"member the trajectory in order to give a correct reward at the",
"(Chevalier-Boisvert et al. 2018) without extra tuning.",
"end.",
"Actor-critic The model underlying the RL agent is pre-",
"sented in Figure 2(a). It consists of standard components",
"5.2EpisodeTermination",
"to predict the next action based on the current observation,",
"thememory ofthepast observationsand the instruction.It",
"IntheBabyAIplatformtheepisodeisterminated whenthe",
"uses GRU to encode the instruction and a convolutional net-",
"taskissolved,orafteramaximumallowednumberof steps",
"work with two batch-normalized FiLM layers to jointly pro-",
"is reached, which is a standard practice for similar RL se-",
"tups. However, in the GAIL setup, no environment reward is",
"cesstheobservationand theinstruction.AnLSTMmemory",
"is used to integrate representations produced by the FiLM",
"available. The naive approach would be to always run every",
"module at each step. It uses a memory of 128 units and en-",
"episodeforthemaximum number of steps.",
"codes the instruction with a unidirectional GRU.",
"We let the agent perform a special Done action. At the",
"same time, all expert demonstrations are padded with this",
"Discriminator The discriminator architecture (see Fig-",
"action as the last one (i.e. just before termination of the",
"ure 2(b)) is similar to the actor-critic model for symmetry",
"episode).ItmakestheagentperformtheDoneactionwhenit",
"and efficiency reasons. We just replace the final actor and",
"considers the task to be done and it can be used to terminate",
"critic layers by 3-layers MLP, placed after the FiLM block.",
"the episode during training. It turns out that it significantly",
"The LSTM here takes as input not only the FiLM embedding",
"speeds up the training procedure as it collects richer data",
"butalsoaone-hotactionvector (bothconcatenated)",
"(more episodes for the same number of frames experienced",
"Training In all our experiments, we use the same hy-",
"by the agent). Done termination is used for all methods, in-",
"perparameters for actor training as used for RL training",
"cludingbaselines.",
"in (Chevalier-Boisvert et al. 2018). For both models, we",
"The approach is related to actor early stopping introduced",
"used the Adam optimizer withthe hyperparameters Q",
"by Zolna et al. (2019). The difference is that we terminate",
"10-4,β1 = 0.9,β2 = 0.999 and e = 10-5. We used the",
"episodes earlierbased onthepolicy'spredictions whilethe",
"Proximal Policy Optimization algorithm with parallelized",
"cited method terminates based on the discriminator's predic-",
"data collection: we performed 4 epochs of PPO using 64",
"tions.",
"rollouts of length 40 collected with multiple processes. We"
] |
{
"arxiv_id": "2002.00412",
"language": "en",
"page_idx": 4
}
|
||
2111.12341-en-0
|
2111.12341-en-0.jpg
|
## EvDistill: Asynchronous Events to End-task Learning via Bidirectional $\textbf{Reconstruction- guided Cross- modal Knowledge Distillation}$
Lin Wang$^{1}$, Yujeong Chae$^{1*}$,Sung-Hoon Yoon$^{1*}$, Tae-Kyun Kim$^2$, and Kuk-Jin Yoon$^{1}$
$^{1}$Visual Intelligence Lab., KAIST, Korea
$^2$ICVL Lab., KAIST, Korea and Imperial College London, UK
$\{$wanglin, yujeong, yoon307, kimtaekyun, kjyoon$\}$@kaist.ac.kr
## $\mathbf{Abstract}$
Event cameras sense per-pixel intensity changes and produce asynchronous event streams with high dynamic range and less motion blur, showing advantages over the conventional cameras. A hurdle of training event-based models is the lack of large qualitative labeled data. Prior
$\sum_{\begin{array}{cc}&works&learning&end-tasks&mostly&rely&on&labeled&or&pseudo-\\&&labeled&datasets&obtainedfrom&the&active&pixel&sensor&(APS)\\&&frames;&however,&such&datasets’&qualitv&is&far&from&rival-\end{array}}$
$frames;however,such$ datasets'quality is far from rival-
$\begin{array}{ccc}{\dot{o}}&{ing~those~based~on~the~canonical~images.~In~this~paper,}\\{\underline{C}}&{we~propose~a~novel~approach,~called~EvDistill,~to~learn}\end{array}$
a student network on the unlabeled and unpaired event data ( target modality) via knowledge distillation ( KD) from a teacher network trained with large-scale, labeled image data (source modality). To enable KD across the unpaired modalities, we first propose a bidirectional modality reconstruction (BMR) module to bridge both modalities and simultaneously exploit them to distill knowledge via the $craftedpairs,causing$ no extra computation in the inference. The BMR is improved by the end-tasks and KD losses in an end-to-end manner. Second, we leverage the structural similarities of both modalities and adapt the knowledge by matching their distributions. Moreover, as most prior feature KD methods are uni-modality and less applicable to our problem, we propose to leverage an affinity graph KD loss to boost the distillation. Our extensive experiments on semantic segmentation and object recognition demonstrate that EvDistill achieves significantly better results than the prior works and KD with only events and APS frames.
## $1. \textbf{Introduction}$
Event cameras have recently received much attention in the computer vision and robotics community for their distinctive advantages, such as high dynamic range (HDR) and much less motion blur. Event cameras sense the intensity changes at each pixel asynchronously and produce event streams encoding time, pixel location, and polarity
$^{*}$These two authors contributed equally.
<FigureHere>
Figure 1: EvDistill distills knowledge from a teacher network trained with large labeled images to a student network learning unpaired and unlabeled events for the end-tasks. To distill knowledge, a bidirectional modality reconstruction and distribution adaptation schemes, with the novel KD losses, are proposed
(sign) of intensity changes. Recently, deep neural network (DNN)-based methods with large-scale, labeled image data have shown significant performance gains on many tasks. However, learning effective event-based DNNs has been impeded by the lack of large pixel-level labeled event data. Prior works learning event-based high-level tasks have resorted to the manually annotated task-specific datasets in a supervised manner[1,4,7,19,20,39,42,56,64,71]. Although some labeled event datasets [31,47,85] have been collected, the quantity and quality are far less favorable compared to those based on the canonical images. Some works [1, 19] have made pseudo labels using the active pixel sensor (APS) images; however, these labels are less accurate due to the low quality of APS images and considerable domain gap with the source data. While [85,87] have explored unsupervised learning, they only focus on the pixellevel prediction tasks, $e.g.$, depth estimation. Another line of research has reconstructed intensity images from events [43,56,61,65,67,68], and these images have been used to learn DNNs on end-tasks, $e.g.$, object recognition [56]; however, annotated labels are still needed, and extra latency is introduced in the inference time.
We explore to leverage large labeled image data (a.k.a. source modality) and the learned models, and aim to learn a model on the unpaired and unlabeled events (a.k.a. target
1
|
[
[
0.131591796875,
0.1337890625,
0.8408203125,
0.1337890625,
0.8408203125,
0.1483154296875,
0.131591796875,
0.1483154296875
],
[
0.1895751953125,
0.1572265625,
0.78173828125,
0.1572265625,
0.78173828125,
0.1717529296875,
0.1895751953125,
0.1717529296875
],
[
0.15283203125,
0.2052001953125,
0.81640625,
0.2052001953125,
0.81640625,
0.2197265625,
0.15283203125,
0.2197265625
],
[
0.325927734375,
0.222900390625,
0.646484375,
0.222900390625,
0.646484375,
0.2374267578125,
0.325927734375,
0.2374267578125
],
[
0.238525390625,
0.2374267578125,
0.73193359375,
0.238037109375,
0.73193359375,
0.255615234375,
0.238525390625,
0.255126953125
],
[
0.0236968994140625,
0.260009765625,
0.058013916015625,
0.260009765625,
0.058013916015625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.2254638671875,
0.260009765625,
0.74169921875,
0.260009765625,
0.74169921875,
0.274658203125,
0.2254638671875,
0.274658203125
],
[
0.236083984375,
0.30810546875,
0.314453125,
0.30810546875,
0.314453125,
0.32568359375,
0.236083984375,
0.32568359375
],
[
0.5556640625,
0.30810546875,
0.6357421875,
0.30810546875,
0.6357421875,
0.3193359375,
0.5556640625,
0.3193359375
],
[
0.66650390625,
0.30810546875,
0.7138671875,
0.30810546875,
0.7138671875,
0.3193359375,
0.66650390625,
0.3193359375
],
[
0.57421875,
0.3193359375,
0.62109375,
0.3193359375,
0.62109375,
0.330810546875,
0.57421875,
0.330810546875
],
[
0.5146484375,
0.33203125,
0.53759765625,
0.33203125,
0.53759765625,
0.4072265625,
0.5146484375,
0.4072265625
],
[
0.66845703125,
0.333251953125,
0.7138671875,
0.333251953125,
0.7138671875,
0.3447265625,
0.66845703125,
0.3447265625
],
[
0.75244140625,
0.333251953125,
0.78369140625,
0.333251953125,
0.78369140625,
0.34033203125,
0.75244140625,
0.34033203125
],
[
0.5703125,
0.335205078125,
0.62158203125,
0.335205078125,
0.62158203125,
0.3466796875,
0.5703125,
0.3466796875
],
[
0.1004638671875,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.348388671875,
0.1004638671875,
0.348388671875
],
[
0.08087158203125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.58642578125,
0.347900390625,
0.60693359375,
0.347900390625,
0.60693359375,
0.35986328125,
0.58642578125,
0.35986328125
],
[
0.5595703125,
0.359130859375,
0.6337890625,
0.35986328125,
0.6337890625,
0.371826171875,
0.5595703125,
0.37060546875
],
[
0.84619140625,
0.361083984375,
0.8544921875,
0.361083984375,
0.8544921875,
0.37060546875,
0.84619140625,
0.37060546875
],
[
0.65869140625,
0.363037109375,
0.7197265625,
0.365478515625,
0.7197265625,
0.376953125,
0.65869140625,
0.375
],
[
0.0236968994140625,
0.3623046875,
0.052276611328125,
0.361083984375,
0.0555419921875,
0.39453125,
0.0269622802734375,
0.395751953125
],
[
0.08251953125,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.376953125,
0.08251953125,
0.376953125
],
[
0.55810546875,
0.37060546875,
0.63818359375,
0.37060546875,
0.63818359375,
0.380615234375,
0.55810546875,
0.380615234375
],
[
0.08087158203125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.58642578125,
0.380615234375,
0.60888671875,
0.380615234375,
0.60888671875,
0.39453125,
0.58642578125,
0.39453125
],
[
0.75244140625,
0.386962890625,
0.810546875,
0.38818359375,
0.810546875,
0.3984375,
0.75244140625,
0.397705078125
],
[
0.080078125,
0.39453125,
0.468017578125,
0.393310546875,
0.468017578125,
0.40771484375,
0.080078125,
0.4091796875
],
[
0.6689453125,
0.399658203125,
0.71337890625,
0.399658203125,
0.71337890625,
0.410888671875,
0.6689453125,
0.410888671875
],
[
0.75244140625,
0.3984375,
0.81201171875,
0.3984375,
0.81201171875,
0.4091796875,
0.75244140625,
0.4091796875
],
[
0.56201171875,
0.40087890625,
0.63232421875,
0.40087890625,
0.63232421875,
0.412353515625,
0.56201171875,
0.412353515625
],
[
0.02532958984375,
0.404052734375,
0.0555419921875,
0.404052734375,
0.0555419921875,
0.449462890625,
0.02532958984375,
0.449462890625
],
[
0.08087158203125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.5556640625,
0.41357421875,
0.63818359375,
0.41357421875,
0.63818359375,
0.4248046875,
0.5556640625,
0.4248046875
],
[
0.80810546875,
0.410400390625,
0.87353515625,
0.410400390625,
0.87353515625,
0.421630859375,
0.80810546875,
0.421630859375
],
[
0.08087158203125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.4375,
0.08087158203125,
0.4375
],
[
0.57275390625,
0.423583984375,
0.62158203125,
0.423583984375,
0.62158203125,
0.434326171875,
0.57275390625,
0.434326171875
],
[
0.041656494140625,
0.436279296875,
0.049835205078125,
0.436279296875,
0.049835205078125,
0.44140625,
0.041656494140625,
0.44140625
],
[
0.07843017578125,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.07843017578125,
0.453857421875
],
[
0.50244140625,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.03350830078125,
0.448974609375,
0.05963134765625,
0.45068359375,
0.0555419921875,
0.48095703125,
0.0294189453125,
0.479248046875
],
[
0.07843017578125,
0.455078125,
0.4697265625,
0.453857421875,
0.4697265625,
0.468505859375,
0.07843017578125,
0.4697265625
],
[
0.50341796875,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.46533203125,
0.50341796875,
0.46533203125
],
[
0.07843017578125,
0.47021484375,
0.4697265625,
0.468505859375,
0.4697265625,
0.481689453125,
0.07843017578125,
0.484130859375
],
[
0.50244140625,
0.46728515625,
0.89208984375,
0.46728515625,
0.89208984375,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.48095703125,
0.89208984375,
0.48095703125,
0.89208984375,
0.494384765625,
0.50244140625,
0.494384765625
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.49560546875,
0.88623046875,
0.49560546875,
0.88623046875,
0.5087890625,
0.50244140625,
0.5087890625
],
[
0.035125732421875,
0.5087890625,
0.051483154296875,
0.5087890625,
0.051483154296875,
0.5244140625,
0.035125732421875,
0.5244140625
],
[
0.080078125,
0.51513671875,
0.46728515625,
0.51513671875,
0.46728515625,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.50341796875,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50341796875,
0.537109375
],
[
0.0294189453125,
0.53515625,
0.051483154296875,
0.53515625,
0.051483154296875,
0.548828125,
0.0294189453125,
0.548828125
],
[
0.080078125,
0.53076171875,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.53857421875,
0.8935546875,
0.53857421875,
0.8935546875,
0.5517578125,
0.50341796875,
0.5517578125
],
[
0.0269622802734375,
0.54736328125,
0.051483154296875,
0.54443359375,
0.0555419921875,
0.55810546875,
0.0310516357421875,
0.5625
],
[
0.0767822265625,
0.54541015625,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.0767822265625,
0.56005859375
],
[
0.02777099609375,
0.5595703125,
0.0555419921875,
0.5595703125,
0.0555419921875,
0.57568359375,
0.02777099609375,
0.57568359375
],
[
0.50341796875,
0.5546875,
0.890625,
0.5546875,
0.890625,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.07843017578125,
0.56103515625,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.07843017578125,
0.57470703125
],
[
0.50341796875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.07843017578125,
0.60400390625,
0.4697265625,
0.60498046875,
0.4697265625,
0.619140625,
0.07843017578125,
0.61865234375
],
[
0.50244140625,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.080078125,
0.62060546875,
0.4697265625,
0.62060546875,
0.4697265625,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.02777099609375,
0.62744140625,
0.0555419921875,
0.62744140625,
0.0555419921875,
0.64453125,
0.02777099609375,
0.64453125
],
[
0.50341796875,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.50341796875,
0.642578125
],
[
0.080078125,
0.63525390625,
0.46728515625,
0.63525390625,
0.46728515625,
0.6484375,
0.080078125,
0.6484375
],
[
0.035125732421875,
0.646484375,
0.051483154296875,
0.646484375,
0.051483154296875,
0.6640625,
0.035125732421875,
0.6640625
],
[
0.080078125,
0.6513671875,
0.46728515625,
0.6513671875,
0.46728515625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.5048828125,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.65576171875,
0.5048828125,
0.65576171875
],
[
0.5009765625,
0.6591796875,
0.89208984375,
0.65771484375,
0.89208984375,
0.67236328125,
0.5009765625,
0.673828125
],
[
0.07843017578125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.68115234375,
0.07843017578125,
0.68115234375
],
[
0.0318603515625,
0.6787109375,
0.052276611328125,
0.67822265625,
0.0555419921875,
0.70556640625,
0.035125732421875,
0.70703125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.07843017578125,
0.68115234375,
0.468017578125,
0.6796875,
0.468017578125,
0.6943359375,
0.07843017578125,
0.69580078125
],
[
0.080078125,
0.6962890625,
0.4697265625,
0.6962890625,
0.4697265625,
0.71044921875,
0.080078125,
0.71044921875
],
[
0.50341796875,
0.69140625,
0.890625,
0.69140625,
0.890625,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.50341796875,
0.70556640625,
0.8896484375,
0.70556640625,
0.8896484375,
0.71630859375,
0.50341796875,
0.71630859375
],
[
0.08087158203125,
0.7119140625,
0.468017578125,
0.7119140625,
0.468017578125,
0.72607421875,
0.08087158203125,
0.72607421875
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.080078125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.080078125,
0.7421875,
0.4345703125,
0.7421875,
0.4345703125,
0.755859375,
0.080078125,
0.755859375
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76806640625,
0.20751953125,
0.76806640625,
0.20751953125,
0.78271484375,
0.080078125,
0.78271484375
],
[
0.50341796875,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.098876953125,
0.7890625,
0.4697265625,
0.7890625,
0.4697265625,
0.80322265625,
0.098876953125,
0.80322265625
],
[
0.5048828125,
0.796875,
0.890625,
0.796875,
0.890625,
0.806640625,
0.5048828125,
0.806640625
],
[
0.080078125,
0.8046875,
0.4697265625,
0.8046875,
0.4697265625,
0.8193359375,
0.080078125,
0.8193359375
],
[
0.50244140625,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.07843017578125,
0.8193359375,
0.468017578125,
0.818359375,
0.468017578125,
0.83349609375,
0.07843017578125,
0.833984375
],
[
0.5009765625,
0.82568359375,
0.8896484375,
0.82568359375,
0.8896484375,
0.8388671875,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.83544921875,
0.468017578125,
0.83544921875,
0.468017578125,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.50341796875,
0.84228515625,
0.73193359375,
0.84228515625,
0.73193359375,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.08087158203125,
0.8515625,
0.46728515625,
0.8515625,
0.46728515625,
0.8623046875,
0.08087158203125,
0.8623046875
],
[
0.51953125,
0.85498046875,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.51953125,
0.869140625
],
[
0.080078125,
0.86572265625,
0.468017578125,
0.86572265625,
0.468017578125,
0.87890625,
0.080078125,
0.87890625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88720703125,
0.30810546875,
0.88720703125,
0.30810546875,
0.90087890625,
0.09722900390625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.488525390625,
0.92822265625,
0.488525390625,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"EvDistill: Asynchronous Events to End-task Learning via Bidirectional",
"Reconstruction-guided Cross-modal Knowledge Distillation",
"Lin Wangl, Yujeong Chael* Sung-Hoon Yoonl*, Tae-Kyun Kim?, and Kuk-Jin Yoon",
"1Visual Intelligence Lab., KAIST, Korea",
"2ICVL Lab., KAIST, Korea and Imperial College London, UK",
"2 A",
"[wanglin,yujeong, yoon3o7,kimtaekyun, kjyoon]@kaist.ac.kr",
"Abstract",
"Eventmodality",
"End-task",
"(Target)",
"eies parien",
"Student",
"Classe",
"Ino labels",
"Event cameras sense per-pixel intensity changes and",
"produce asynchronous event streams with high dynamic",
"金",
"Bidirectional",
"KD",
"Knowledge",
"",
"range and less motion blurshowing advantages over the",
"Reconstruction",
"conventional cameras.A hurdle of training event-based",
"+",
"Distributior",
"models is the lack of large qualitative labeled data. Prior",
"Teacher",
"Adaptation",
"with labels",
"LAD",
"works learning end-tasks mostly rely on labeled or pseudo-",
"Imagemodality",
"+Predictions",
"labeleddatasetsobtainedfromtheactivepixelsensor(APs)",
"(Source)",
"中",
"frames; however, such datasets'quality is far from rival-",
"Figure 1: EvDistill distills knowledge from a teacher network",
"SO",
"ing those based on the canonical images.In this paper,",
"trained with large labeled images to a student network learn-",
"we propose a novel approach, called EvDistill, to learn",
"ing unpaired and unlabeled events for the end-tasks. To distill",
"a student network on the unlabeled and unpaired event",
"knowledge, a bidirectional modality reconstruction and distribu-",
"data (target modality) via knowledge distillation (KD) from",
"tion adaptation schemes, with the novel KD losses, are proposed.",
"V",
"a teacher network trained with large-scale, labeled im-",
"(sign) of intensity changes. Recently, deep neural network",
"4",
"age data (source modality). To enable KD across the un-",
"(DNN)-based methods with large-scale, labeled image data",
"3",
"paired modalities, we first propose a bidirectional modal-",
"2",
"haveshownsignificantperformance gains on manytasks.",
"ityreconstruction(BMR)moduletobridgebothmodalities",
"However, learning effective event-based DNNs has been",
"and simultaneously exploit them to distill knowledge via the",
"impeded by the lack of large pixel-level labeled event data.",
"crafted pairs, causing no extra computation in the inference.",
"The BMR is improved by the end-tasks and KD losses in",
"Prior works learning event-based high-level tasks have re-",
"sorted to the manually annotated task-specific datasets in a",
"an end-to-end manner: Second, we leverage the structural",
"2",
"supervised manner [1, 4, 7, 19, 20, 39, 42, 56, 64, 71]. Al-",
"similarities of both modalities and adapt the knowledge by",
"V",
"matching their distributions. Moreover, as most prior fea-",
"thoughsomelabeled eventdatasets[3l,47.851havebeen",
"collected, the quantity and quality are far less favorable",
"ture KD methods are uni-modality and less applicable to",
"e",
"compared to those based on the canonical images. Some",
"our problem, we propose to leverage an affinity graph KD",
"loss to boost the distillation. Our extensive experiments on",
"works[1,19havemadepseudolabelsusingtheactivepixel",
"sensor (APS)images;however,these labels are less accu-",
"semantic segmentation and object recognition demonstrate",
"rate due to the low quality of APS images and considerable",
"that EvDistill achieves significantly better results than the",
"domain gap with the source data. While [85, 87] have ex-",
"prior works and KD with only events and APS frames.",
"plored unsupervised learning, they only focus on the pixel-",
"1.Introduction",
"level prediction tasks, e.g., depth estimation. Another line",
"of research has reconstructed intensity images from events",
"Event cameras have recently received much attention in",
"[43,56.61.65.67.68l.and these imageshavebeen used",
"the computer vision and robotics community for their dis-",
"to learn DNNs on end-tasks, e.g., object recognition [56];",
"tinctive advantages, such as high dynamic range (HDR)",
"however, annotated labels are still needed, and extra latency",
"and much less motion blur. Event cameras sense the in-",
"is introduced in the inference time.",
"tensity changes at eachpixel asynchronouslyand produce",
"We explore to leverage large labeled image data (a.k.a.",
"event streams encoding time, pixel location, and polarity",
"source modality) and the learned models, and aim to learn",
"*These two authors contributed equally.",
"a model on the unpaired and unlabeled events (a.k.a. target",
"1"
] |
{
"arxiv_id": "2111.12341",
"language": "en",
"page_idx": 0
}
|
||
2111.12341-en-1
|
2111.12341-en-1.jpg
|
modality) via cross-modal learning [23,83] and knowledge distillation (KD) [23,63,70]. Most existing cross-modal learning methods have relied on paired data(e.g., image and depth) with the same labels [23,24,36,48,69,73,78,83] or extra information $(e.g.$, data or labels) [2, 18,52,79] or grafting networks between modalities $(e.g.$, image to thermal) [30] for learning the end-tasks. Some works have explored the unpaired multi-modality data[14,36]; however, it is assumed that labels for both modalities are available, which is difficult to achieve for the event data.
To overcome these limitations, we propose a novel method, called EvDistill, to efficiently learn a student network on the unpaired and unlabeled event data by distilling the knowledge from a robust teacher network trained with large labeled image data, as shown in Fig. 1. Firstly, we propose a bidirectional modality reconstruction (BMR) module to bridge both modalities, and then simultaneously exploit them to distill knowledge via the crafted pairs, adding no extra computation cost during inference (Sec.3.2). Importantly, the BMR is improved by the end-task and the KD losses in an end-to-end manner. That is, BMR produces the crafted pairs of both modalities to distill knowledge to the student network in the forward pass, and KD facilitates the learning of the BMR in the backward pass. Secondly, as the feature representations of two modalities extracted from the task networks could suffer from distribution mismatch, we leverage the structural similarities and adapt knowledge by matching the class distributions based on the BMR module (Sec. 3.3). Moreover, as most existing methods are limited to uni-modality [34,38,57,80], we leverage a graph affinity KD loss and other losses to learn a better model on the event data (Sec. 3.4). We evaluate the performance of the proposed framework on three datasets in semantic segmentation (Sec. 4.1) and one dataset in object recognition (Sec. 4.2). The experiments show that our approach achieves significantly better performance than the prior works for both end-tasks and the naive setting, KD with only events and the APS frames (when APS frames are available). The validation code and trained models are available at https://github.com/ addisonwang2013/evdistill.
## $2. \textbf{Related Works}$
DNNs for event-based end-tasks. DNNs with event data was first explored for the classification task[46] and for robot control [42]. [39] then trained a DNN for steering angle prediction on DDD17 dataset [5]. This dataset has been utilized by [1, 19] to perform semantic segmentation using pseudo labels obtained from the APS frames. Moreover, DNNs have been applied to some high-level prediction tasks, such as object detection and tracking $\hat{[7,30,40,53]}$, human pose estimation [6,71,75],motion estimation [33,41,60,72,76], object recognition [4, 19, 56]
on N-Caltech [47] and other datasets [4,35,59].
DNNs for event-based low-level vision. Meanwhile, another line of research focuses on the low-level prediction tasks, such as optical flow estimation [17,20,61,87], stereo depth estimation [64, 87] on MVSEC dataset [85]. In addition, [44, 50, 56, 58, 61, 68] attempted to reconstruct intensity image/video from events using camera simulator [45,54],and [43,65,67] tried to reconstruct high-resolution images. In contrast to image/video reconstruction from events, [19] proposed to generate events from video frames. Some other works also explored the potential of events for image deblurring [26, 32, 67], HDR imaging [25, 81], and event denoising [3]. For more details about event-based vision, refer to a survey [16]. Differently, we propose EvDistill, learning event-based end-tasks on the unpaired and unlabeled events via cross modal KD, in which a BMR moudle is proposed to bridge both modalities and is learned with the end-task networks in an end-to-end manner.
Knowledge Distillation. KD aims to build a smaller (student) model with the softmax labels of a larger (teacher) model [28,57,70]. Most KD methods learning end-tasks have been focused on uni-modality $(e.g.$, image) data and distill knowledge using logits [9,74,77,82] or features [27, 34, 51, 57, 80]. Cross-modal KD aims to transfer knowledge across different modalities. Most prior crossmodal KD methods[8, 14,23,24,29,30,36,48,49,69,73, 78, 83] relied on the paired data with the common labels, while some works utilized extra information (e.g., data) [2, 18, 52, 79] or grafting networks [30] to transfer knowledge. Although[14, 36] explored the unpaired multi-modal KD, the labels for both modalities are needed. For more details about KD, refer to[70]. Learning from event data is more challenging as no paired labels for two modalities exist; we thus propose EvDistill, where we exploit a BMR module to connect them and simultaneously distill knowledge by adapting distribution with novel KD losses.
$3. \textbf{ The Proposed Method: EvDistill}$
Event Representation An event $e$ is interpreted as a tuple $(\mathbf{u},t,\mathbf{\hat{p}})$,where $\mathbf{u}=(x,y)$ is the pixel coordinate, $t$ is the timestamp, and $p$ is the polarity indicating the sign of brightness change. An event occurs whenever a change in log-scale intensity exceeds a threshold. A natural choice is to encode events in a spatial-temporal 3D volume to a voxel grid [56, 86, 87] or event frame [20, 55] or multi-channel images [37,65,68]. In this paper, we represent events to multi-channel event images as the inputs to the DNNs. Details are provided in the suppl. material.
## 3.1.0verview
We describe the proposed EvDistill framework for learning end-tasks from events, as shown in Fig.2. For event cameras, $e.g.$, DAVIS346 [62] with APS frames, assume that we are given the target modality data $\mathcal{X}_T=\{e,x_{aps}\}_i$
2
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.8203125,
0.091552734375,
0.8203125,
0.10479736328125,
0.50244140625,
0.10479736328125
],
[
0.07843017578125,
0.1060791015625,
0.468017578125,
0.106689453125,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.50244140625,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.50244140625,
0.12060546875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.149658203125,
0.50341796875,
0.149658203125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.07843017578125,
0.1673583984375,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
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
],
[
0.50341796875,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.2108154296875,
0.50341796875,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.08087158203125,
0.2298583984375,
0.38232421875,
0.2298583984375,
0.38232421875,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50244140625,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.098876953125,
0.24560546875,
0.4697265625,
0.24560546875,
0.4697265625,
0.260009765625,
0.098876953125,
0.260009765625
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.274658203125,
0.080078125,
0.274658203125
],
[
0.50244140625,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.271484375,
0.50244140625,
0.271484375
],
[
0.080078125,
0.276611328125,
0.4697265625,
0.276611328125,
0.4697265625,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.290283203125,
0.46728515625,
0.290283203125,
0.46728515625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.07843017578125,
0.3212890625,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.07843017578125,
0.335205078125
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50244140625,
0.33203125
],
[
0.080078125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.080078125,
0.349853515625
],
[
0.50244140625,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.50341796875,
0.35107421875,
0.7900390625,
0.35107421875,
0.7900390625,
0.361083984375,
0.50341796875,
0.361083984375
],
[
0.080078125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.50341796875,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.376953125,
0.50341796875,
0.376953125
],
[
0.08087158203125,
0.38330078125,
0.46728515625,
0.38330078125,
0.46728515625,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.50244140625,
0.37890625,
0.89208984375,
0.37890625,
0.89208984375,
0.392578125,
0.50244140625,
0.392578125
],
[
0.080078125,
0.396484375,
0.468017578125,
0.396484375,
0.468017578125,
0.410400390625,
0.080078125,
0.410400390625
],
[
0.50341796875,
0.395751953125,
0.8896484375,
0.395751953125,
0.8896484375,
0.406005859375,
0.50341796875,
0.406005859375
],
[
0.08087158203125,
0.41357421875,
0.46728515625,
0.41357421875,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.50244140625,
0.410400390625,
0.89208984375,
0.410400390625,
0.89208984375,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4267578125,
0.468017578125,
0.4267578125,
0.468017578125,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.442626953125,
0.46728515625,
0.442626953125,
0.46728515625,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.50341796875,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50341796875,
0.45263671875
],
[
0.08087158203125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.080078125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.50244140625,
0.4697265625,
0.8896484375,
0.4697265625,
0.8896484375,
0.48291015625,
0.50244140625,
0.48291015625
],
[
0.08087158203125,
0.488037109375,
0.46728515625,
0.488037109375,
0.46728515625,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.08087158203125,
0.50390625,
0.46728515625,
0.50390625,
0.46728515625,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50341796875,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.50341796875,
0.5146484375
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.5048828125,
0.5166015625,
0.88818359375,
0.5166015625,
0.88818359375,
0.5263671875,
0.5048828125,
0.5263671875
],
[
0.08087158203125,
0.5341796875,
0.46728515625,
0.5341796875,
0.46728515625,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.50244140625,
0.52978515625,
0.89208984375,
0.52978515625,
0.89208984375,
0.54345703125,
0.50244140625,
0.54345703125
],
[
0.07843017578125,
0.54736328125,
0.4697265625,
0.546875,
0.4697265625,
0.56103515625,
0.07843017578125,
0.5625
],
[
0.5048828125,
0.546875,
0.890625,
0.546875,
0.890625,
0.556640625,
0.5048828125,
0.556640625
],
[
0.080078125,
0.564453125,
0.468017578125,
0.564453125,
0.468017578125,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.50244140625,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.07843017578125,
0.57763671875,
0.4697265625,
0.5771484375,
0.4697265625,
0.59130859375,
0.07843017578125,
0.59228515625
],
[
0.50244140625,
0.57568359375,
0.89208984375,
0.57568359375,
0.89208984375,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.5009765625,
0.59033203125,
0.890625,
0.5888671875,
0.890625,
0.60400390625,
0.5009765625,
0.60498046875
],
[
0.080078125,
0.60791015625,
0.4697265625,
0.60791015625,
0.4697265625,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50244140625,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.619140625,
0.50244140625,
0.619140625
],
[
0.080078125,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.6220703125,
0.84326171875,
0.6220703125,
0.84326171875,
0.63525390625,
0.50244140625,
0.63525390625
],
[
0.07843017578125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.07843017578125,
0.6513671875
],
[
0.5009765625,
0.642578125,
0.7958984375,
0.64306640625,
0.7958984375,
0.6611328125,
0.5009765625,
0.66015625
],
[
0.07843017578125,
0.65283203125,
0.4697265625,
0.65283203125,
0.4697265625,
0.66650390625,
0.07843017578125,
0.66650390625
],
[
0.08087158203125,
0.66943359375,
0.46728515625,
0.66943359375,
0.46728515625,
0.6796875,
0.08087158203125,
0.6796875
],
[
0.50341796875,
0.66650390625,
0.890625,
0.66650390625,
0.890625,
0.6796875,
0.50341796875,
0.6796875
],
[
0.50244140625,
0.67822265625,
0.890625,
0.67822265625,
0.890625,
0.69140625,
0.50244140625,
0.69140625
],
[
0.07843017578125,
0.6845703125,
0.46728515625,
0.6845703125,
0.46728515625,
0.69775390625,
0.07843017578125,
0.69775390625
],
[
0.50341796875,
0.693359375,
0.89208984375,
0.693359375,
0.89208984375,
0.70703125,
0.50341796875,
0.70703125
],
[
0.080078125,
0.70068359375,
0.330810546875,
0.70068359375,
0.330810546875,
0.7109375,
0.080078125,
0.7109375
],
[
0.50341796875,
0.71044921875,
0.890625,
0.71044921875,
0.890625,
0.72021484375,
0.50341796875,
0.72021484375
],
[
0.080078125,
0.72802734375,
0.226318359375,
0.72802734375,
0.226318359375,
0.7421875,
0.080078125,
0.7421875
],
[
0.50244140625,
0.7236328125,
0.89208984375,
0.7236328125,
0.89208984375,
0.7373046875,
0.50244140625,
0.7373046875
],
[
0.50244140625,
0.73779296875,
0.89208984375,
0.73779296875,
0.89208984375,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.080078125,
0.75,
0.46728515625,
0.75,
0.46728515625,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.50341796875,
0.75390625,
0.89208984375,
0.75390625,
0.89208984375,
0.767578125,
0.50341796875,
0.767578125
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76806640625,
0.89208984375,
0.76806640625,
0.89208984375,
0.7822265625,
0.50244140625,
0.7822265625
],
[
0.080078125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.7841796875,
0.890625,
0.7841796875,
0.890625,
0.79736328125,
0.50244140625,
0.79736328125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.7998046875,
0.7646484375,
0.7998046875,
0.7646484375,
0.81298828125,
0.50244140625,
0.81298828125
],
[
0.07843017578125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8212890625,
0.609375,
0.8212890625,
0.609375,
0.83544921875,
0.50244140625,
0.83544921875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.521484375,
0.84033203125,
0.89208984375,
0.84033203125,
0.89208984375,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.07843017578125,
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
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5009765625,
0.88623046875
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.5,
0.8837890625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90380859375,
0.5,
0.9013671875
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"modality) via cross-modal learning [23, 83] and knowledge",
"on N-Caltech [471 and other datasets [4, 35, 59].",
"distillation (KD) [23, 63, 70]. Most existing cross-modal",
"DNNs for event-based low-level vision. Meanwhile, an-",
"learning methods have relied on paired data (e.g., image and",
"other line of research focuses on the low-level prediction",
"depth) with the same labels [23, 24, 36, 48, 69, 73, 78, 83]",
"tasks,such as optical flowestimation[17.20,61.871,stereo",
"or extra information (e.g., data or labels) [2, 18, 52, 79] or",
"depth estimation [64, 87] on MVSEC dataset [85]. In ad-",
"grafting networks between modalities (e.g., image to ther-",
"dition, [44, 50, 56, 58, 61, 68] attempted to reconstruct",
"mal) [30] for learning the end-tasks. Some works have ex-",
"intensity image/video from events using camera simulator",
"plored theunpaired multi-modality data [14,36l:however.",
"[45, 54], and [43, 65,67] tried to reconstruct high-resolution",
"it is assumed that labels for both modalities are available,",
"images.In contrast to image/video reconstruction from",
"which is difficultto achievefortheevent data.",
"events, [19] proposed to generate events from video frames.",
"To overcome these limitations, we propose a novel",
"Some other works also explored the potential of events for",
"method, called EvDistill, to efficiently learn a student net-",
"image deblurring [26, 32, 67], HDR imaging [25, 81], and",
"work on the unpaired and unlabeled event data by distilling",
"event denoising [3]. For more details about event-based vi-",
"the knowledge from a robust teacher network trained with",
"sion, refer to a survey [16]. Differently, we propose EvDis-",
"large labeled image data, as shown in Fig. 1. Firstly, we pro-",
"till, learning event-based end-tasks on the unpaired and un-",
"pose a bidirectional modality reconstruction (BMR)module",
"labeled events via cross modal KD, in which a BMR moudle",
"to bridge both modalities, and then simultaneously exploit",
"is proposed to bridge both modalities and is learned with the",
"them to distill knowledge via the crafted pairs, adding no",
"end-tasknetworksinan end-to-endmanner.",
"extra computation cost during inference (Sec. 3.2). Im-",
"Knowledge Distillation. KD aims to build a smaller (stu-",
"portantly.the BMR is improvedby the end-task and the",
"dent) model with the softmax labels of a larger (teacher)",
"KD losses in an end-to-end manner. That is, BMR pro-",
"model [28.57.70]:Most KD methods learning end-tasks",
"duces the crafted pairs of bothmodalitiestodistillknowl-",
"have been focused on uni-modality (e.g., image) data and",
"edge to the student network in the forward pass, and KD",
"distill knowledge using logits [9, 74, 77, 82] or features",
"facilitates the learning of the BMR in the backward pass.",
"[27, 34, 51, 57, 80]. Cross-modal KD aims to transfer",
"Secondly,as the feature representations of two modalities",
"knowledge across different modalities. Most prior cross-",
"extracted from the task networks could suffer from distri-",
"modal KD methods [8, 14, 23,24, 29, 30, 36, 48, 49, 69, 73",
"bution mismatch,we leveragethe structural similarities and",
"78, 83] relied on the paired data with the common labels,",
"adaptknowledge by matching the class distributions based",
"while some works utilized extra information (e.g., data)",
"on the BMR module (Sec. 3.3). Moreover, as most exist-",
"[2.18.52.79]orgraftingnetworks[30]totransferknowl-",
"ing methods are limited touni-modality[34,38,57.80].",
"edge. Although [14, 36] explored the unpaired multi-modal",
"we leverage a graph affinity KD loss and other losses to",
"KD.thelabelsforbothmodalities areneeded.Formore",
"learn a better model on the event data (Sec. 3.4).We",
"details about KD, refer to [70]. Learning from event data",
"evaluate the performance of the proposed framework on",
"is more challenging as no paired labels for two modalities",
"three datasets in semantic segmentation (Sec. 4.1) and one",
"exist; we thus propose EvDistill, where we exploit a BMR",
"dataset in object recognition (Sec. 4.2). The experiments",
"module to connect them and simultaneously distill knowl-",
"show that our approach achieves significantly better per-",
"edge by adapting distribution with novel KD losses.",
"formance than the prior works for both end-tasks and the",
"3. The Proposed Method: EvDistill",
"naive setting, KD with only events and the APS frames",
"(when APSframes areavailable).Thevalidationcodeand",
"Event Representation An event e is interpreted as a tu-",
"ple (u,t,p), where u = (c, y) is the pixel coordinate, t is",
"trained models are available at https : / /github.con/",
"the timestamp, and p is the polarity indicating the sign of",
"addisonwang2013/evdistill.",
"brightness change.An eventoccurs whenever a change in",
"2.Related Works",
"log-scale intensity exceeds a threshold. A natural choice is",
"to encode events in a spatial-temporal 3D volume to a voxel",
"DNNs for event-based end-tasks. DNNs with event data",
"grid [56, 86, 87] or event frame [20, 55] or multi-channel",
"was first explored for the classification task [46] and for",
"images [37, 65, 68]. In this paper, we represent events to",
"robot control [42]. [39] then trained a DNN for steer-",
"multi-channel event images as the inputs to the DNNs. De-",
"ing angle prediction on DDD17 dataset [5]. This dataset",
"tails are provided in the suppl. material.",
"has been utilized by [1, 19] to perform semantic segmen-",
"tation using pseudo labels obtained from the APS frames.",
"3.1.0verview",
"Moreover, DNNs have been applied to some high-level",
"We describe the proposed EvDistill framework for learn-",
"prediction tasks, such as object detection and tracking",
"ing end-tasks from events, as shown in Fig. 2. For event",
"[7, 30, 40, 53], human pose estimation [6, 71, 75], motion",
"cameras, e.g., DAVIS346 [62] with APS frames, assume",
"estimation [33,41, 60, 72, 76], object recognition [4, 19, 56]",
"that we are given the target modality data XT = fe, Caps i",
"2"
] |
{
"arxiv_id": "2111.12341",
"language": "en",
"page_idx": 1
}
|
||
2111.12341-en-4
|
2111.12341-en-4.jpg
|
<FigureHere>
Figure 5: IIlustration of the affinity graph distillation loss between
the event and image modalities.
two modalities, we denote $A_uv^{\mathcal{T}}$ and $A_uv^{\mathcal{S}}$ are the affinities between the $u$-th node and the $v$-th node obtained from the teacher and student, respectively, which is formulated as:
(7)
$$\mathcal{L}_{AG}=\frac{1}{H\times W\times\sigma}\sum_{u\sim R}\sum_{v\sim\sigma}\|A_{uv}^{\mathcal{T}}-A_{uv}^{\mathcal{S}}\|_{2}^{2}$$
where $R=\{1,2,\cdots,H\times W\}$ indicates all the nodes in the graph. The similarity between two nodes is calculated from the aggregated features $F_u$ and $F_v$ as $A_uv=\frac{F_u^\mathsf{T}F_v}{||F_u\text{T}||_2||F_v||_2}$, where $F_u^\mathrm{T}$ is the transposed feature vector of $F_u$ [13,38]. Mutual Distillation (MD). When APS frames exist, $\mathcal{S}_{aps}$ indeed can facilitate the learning of $\mathcal{S}_ev.$ Since $\mathcal{S}_aps$ and $\mathcal{S}_ev$ start from different initial conditions, they learn different representations, and consequently, their prediction of probabilities can be an effective regularization to each other. We thus let $\mathcal{S}_{aps}$ and $\mathcal{S}_ev$ learn from each other's predictions via the KL divergence losses with a temperature parameter $\tau$[82]. This helps $\mathcal{S}_ev$ converge to better minima for better generalization to test data. The MD loss is formulated as:
(8)
$$\mathcal{L}_{MD}=\mathbb{E}_{e,x_{aps}\sim\mathcal{X}_{T}}KL\left[\mathcal{S}_{ev}(e)|\mathcal{S}_{aps}(x_{aps}),\tau\right]+\\\mathbb{E}_{e,x_{aps}\sim\mathcal{X}_{T}}KL\left[\mathcal{S}_{aps}(x_{aps})|\mathcal{S}_{ev}(e),\tau\right].$$
In summary, the objective $\mathcal{L}$ of EvDistill is as follows:
(9)
$$\mathcal{L}=\mathcal{L}_{CE}+\mathcal{L}_{BMR}+\lambda_{1}\mathcal{L}_{DA}+\lambda_{2}\mathcal{L}_{AG}+\lambda_{3}\mathcal{L}_{MD}$$
where $\lambda_1,\lambda_2$ and $\lambda_3$ are the hyper-parameters.
4. Experiment and Evaluation
4.1. Event-based semantic segmentation
Semantic segmentation is a task that aims to assign a semantic label, $e.g.$, road, car, in a given scene to each pixel. Unlike image data, annotating events requires special processing for the raw event streams. Besides, it is challenging to correctly label the pixels due to the sparsity of events and lacking information $(e.g.$,material and textures).
Datasets. We use the publicly available driving scene dataset DDD17 [5], which includes both events and APS frames recorded by a DAVIS346 event camera. In [l], 19,840 APS frames are utilized to generate pseudo annotations(6 classes) based on a pretrained network for events (15,950 for training and 3,890 for test). However, such a way leads to less precise segmentation labels because there is a considerable domain gap between the source data and
APS images of low resolution and quality. Note that our method does not rely on any annotations of events in training, and the pseudo annotations by [1] for test are only used for evaluation and comparison. As the events in the DDD17 dataset are very sparse and noisy, we show more results on the driving sequences in the MVSEC dataset [87], collected for the stereo purpose. Moreover, we show qualitative results on the E2Vid driving scene dataset [56] captured by using a Samsung event camera (with higher resolution).
Implementation details. For each label of DDD17 dataset provided by[1], we use the events occurred in a 50ms time window before a label for prediction, as done in [1, 19]. We also consider the event representation method in[1]for semantic segmentation on the DDD17 dataset, in addition to that described in Sec. 3. For the event representation on MVSEC and E2VID datasets, we use the method in Sec. 3. For the teacher $\mathcal{T}$ and students $\mathcal{S}_ev$ and $\mathcal{S}_{aps}$, we adopt a segmentation network [11]. We use the following metric to evaluate the performance. The intersection of union (IoU) score is calculated as the ratio of intersection and union between the GT mask and the predicted segmentation mask for each class. We use the mean IoU (MIoU) to measure the effectiveness, as done in [10, 11]. The segmentation maps are with 6 classes, as done in [1]. For more implementation details $(e.g.$, training), refer to the suppl. material.
## 4.1.1 Evaluation on DDD17 dataset
Comparison. We first present the experimental results on the DDD17 dataset [l]. We evaluate our method on the test set and vary the window size of events between 10,50, and 250ms, as done in $[\color{red}{\color{red}{1}}].$ The quantitative and qualitative results are shown in Table l and Fig. 6. We compare our method with two existing methods, EvSegNet[l] and Vid2E[19] that uses synthetic version of DDD17 data. It turns out that, without using labels, EvDistill significantly improves the segmentation results on events and surpasses the existing methods with around $7\%$ increase in MIoU using a multi-channel event representation. Segmentation using the voxel grid representation is slightly less effective than that of using the multi-channel representation. Meanwhile, on the time interval of 10ms and 250ms, EvDistill also shows a significant increase of MIoU by around $7.5\%$ and 9.5% than those of EvSegNet, respectively.
We also demonstrate that our method significantly enhances the semantic segmentation performance on the APS frames in Table 1. Quantitatively, without resorting to the pseudo labels, our method surpasses EvSegNet by a large margin with around $12\%$ increase of MIoU. This reflects that our method dramatically minimizes the domain gap between APS frames and labeled source data and distills the knowledge from the teacher network to learn a student network $\mathcal{S}_{aps}$ showing better segmentation capability. The results on both events and APS frames indicate that our
5
|
[
[
0.08416748046875,
0.0889892578125,
0.1650390625,
0.0889892578125,
0.1650390625,
0.0997314453125,
0.08416748046875,
0.0997314453125
],
[
0.37744140625,
0.091552734375,
0.459228515625,
0.091552734375,
0.459228515625,
0.0985107421875,
0.37744140625,
0.0985107421875
],
[
0.50341796875,
0.09027099609375,
0.8896484375,
0.09027099609375,
0.8896484375,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.4306640625,
0.10418701171875,
0.439453125,
0.10418701171875,
0.439453125,
0.10919189453125,
0.4306640625,
0.10919189453125
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.240234375,
0.11932373046875,
0.295654296875,
0.11932373046875,
0.295654296875,
0.1300048828125,
0.240234375,
0.1300048828125
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.340576171875,
0.1383056640625,
0.34716796875,
0.1383056640625,
0.34716796875,
0.143310546875,
0.340576171875,
0.143310546875
],
[
0.50244140625,
0.136962890625,
0.890625,
0.136962890625,
0.890625,
0.15087890625,
0.50244140625,
0.15087890625
],
[
0.250732421875,
0.154052734375,
0.2900390625,
0.154052734375,
0.2900390625,
0.1641845703125,
0.250732421875,
0.1641845703125
],
[
0.50341796875,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.2467041015625,
0.1673583984375,
0.283447265625,
0.1673583984375,
0.283447265625,
0.1749267578125,
0.2467041015625,
0.1749267578125
],
[
0.5009765625,
0.1673583984375,
0.89208984375,
0.1666259765625,
0.89208984375,
0.18115234375,
0.5009765625,
0.1817626953125
],
[
0.080078125,
0.1767578125,
0.4697265625,
0.1767578125,
0.4697265625,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.50341796875,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.080078125,
0.1900634765625,
0.26708984375,
0.1900634765625,
0.26708984375,
0.203857421875,
0.080078125,
0.203857421875
],
[
0.5009765625,
0.1962890625,
0.890625,
0.1976318359375,
0.890625,
0.212158203125,
0.5009765625,
0.2108154296875
],
[
0.080078125,
0.2183837890625,
0.468017578125,
0.2183837890625,
0.468017578125,
0.231689453125,
0.080078125,
0.231689453125
],
[
0.50341796875,
0.21337890625,
0.86767578125,
0.21337890625,
0.86767578125,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.08087158203125,
0.236083984375,
0.46728515625,
0.236083984375,
0.46728515625,
0.246826171875,
0.08087158203125,
0.246826171875
],
[
0.50341796875,
0.2310791015625,
0.89208984375,
0.2310791015625,
0.89208984375,
0.2442626953125,
0.50341796875,
0.2442626953125
],
[
0.080078125,
0.25,
0.455810546875,
0.25,
0.455810546875,
0.26318359375,
0.080078125,
0.26318359375
],
[
0.50244140625,
0.246826171875,
0.89208984375,
0.246826171875,
0.89208984375,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.50244140625,
0.260009765625,
0.890625,
0.261474609375,
0.890625,
0.27587890625,
0.50244140625,
0.274658203125
],
[
0.214111328125,
0.271484375,
0.2254638671875,
0.271484375,
0.2254638671875,
0.281005859375,
0.214111328125,
0.281005859375
],
[
0.12335205078125,
0.27783203125,
0.1732177734375,
0.27978515625,
0.17236328125,
0.294189453125,
0.1217041015625,
0.292236328125
],
[
0.265625,
0.27783203125,
0.431396484375,
0.27783203125,
0.431396484375,
0.292236328125,
0.265625,
0.292236328125
],
[
0.4306640625,
0.27978515625,
0.46728515625,
0.27978515625,
0.46728515625,
0.291748046875,
0.4306640625,
0.291748046875
],
[
0.50341796875,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.290283203125,
0.50341796875,
0.290283203125
],
[
0.1732177734375,
0.285400390625,
0.268798828125,
0.28857421875,
0.268798828125,
0.302978515625,
0.1732177734375,
0.2998046875
],
[
0.2646484375,
0.29541015625,
0.32275390625,
0.29541015625,
0.32275390625,
0.30615234375,
0.2646484375,
0.30615234375
],
[
0.50244140625,
0.291748046875,
0.89208984375,
0.291748046875,
0.89208984375,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.50244140625,
0.306884765625,
0.89208984375,
0.306884765625,
0.89208984375,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.08087158203125,
0.313232421875,
0.4697265625,
0.313232421875,
0.4697265625,
0.326904296875,
0.08087158203125,
0.326904296875
],
[
0.5009765625,
0.32080078125,
0.89208984375,
0.3212890625,
0.89208984375,
0.3359375,
0.5009765625,
0.335205078125
],
[
0.07843017578125,
0.328857421875,
0.4697265625,
0.32763671875,
0.4697265625,
0.34228515625,
0.07843017578125,
0.343505859375
],
[
0.5009765625,
0.3359375,
0.89208984375,
0.337158203125,
0.89208984375,
0.3515625,
0.5009765625,
0.35107421875
],
[
0.080078125,
0.3466796875,
0.371826171875,
0.3466796875,
0.371826171875,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.406005859375,
0.343505859375,
0.4306640625,
0.343505859375,
0.4306640625,
0.35107421875,
0.406005859375,
0.35107421875
],
[
0.381591796875,
0.354248046875,
0.46728515625,
0.352783203125,
0.46728515625,
0.363037109375,
0.381591796875,
0.3642578125
],
[
0.50244140625,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.08087158203125,
0.363037109375,
0.453369140625,
0.363037109375,
0.453369140625,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.50244140625,
0.367431640625,
0.89208984375,
0.367431640625,
0.89208984375,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.0767822265625,
0.376953125,
0.4697265625,
0.37890625,
0.4697265625,
0.396484375,
0.0767822265625,
0.39453125
],
[
0.50244140625,
0.3837890625,
0.8896484375,
0.3837890625,
0.8896484375,
0.39453125,
0.50244140625,
0.39453125
],
[
0.07843017578125,
0.393310546875,
0.468017578125,
0.395751953125,
0.468017578125,
0.410400390625,
0.07843017578125,
0.40771484375
],
[
0.5009765625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.5009765625,
0.410888671875
],
[
0.08087158203125,
0.412353515625,
0.468017578125,
0.412353515625,
0.468017578125,
0.42236328125,
0.08087158203125,
0.42236328125
],
[
0.50244140625,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.08087158203125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.4375,
0.08087158203125,
0.4375
],
[
0.50244140625,
0.427978515625,
0.89208984375,
0.427978515625,
0.89208984375,
0.44140625,
0.50244140625,
0.44140625
],
[
0.07843017578125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.07843017578125,
0.453857421875
],
[
0.50341796875,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.455078125,
0.50341796875,
0.455078125
],
[
0.080078125,
0.455810546875,
0.4697265625,
0.455810546875,
0.4697265625,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50244140625,
0.458251953125,
0.8310546875,
0.458251953125,
0.8310546875,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.07843017578125,
0.4697265625,
0.4697265625,
0.47021484375,
0.4697265625,
0.48486328125,
0.07843017578125,
0.484130859375
],
[
0.08087158203125,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.50341796875,
0.4873046875,
0.76123046875,
0.4873046875,
0.76123046875,
0.50048828125,
0.50341796875,
0.50048828125
],
[
0.07843017578125,
0.50048828125,
0.45751953125,
0.5,
0.45751953125,
0.5146484375,
0.07843017578125,
0.51513671875
],
[
0.50341796875,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50341796875,
0.521484375
],
[
0.101318359375,
0.52294921875,
0.4306640625,
0.52294921875,
0.4306640625,
0.54052734375,
0.101318359375,
0.54052734375
],
[
0.5009765625,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.537109375,
0.5009765625,
0.537109375
],
[
0.4453125,
0.5341796875,
0.46728515625,
0.5341796875,
0.46728515625,
0.546875,
0.4453125,
0.546875
],
[
0.1666259765625,
0.54248046875,
0.4306640625,
0.54248046875,
0.4306640625,
0.56005859375,
0.1666259765625,
0.56005859375
],
[
0.50341796875,
0.54052734375,
0.890625,
0.54052734375,
0.890625,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.08087158203125,
0.5673828125,
0.439453125,
0.5673828125,
0.439453125,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50341796875,
0.57080078125,
0.890625,
0.57080078125,
0.890625,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.50244140625,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.0931396484375,
0.59130859375,
0.46728515625,
0.59130859375,
0.46728515625,
0.60595703125,
0.0931396484375,
0.60595703125
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.611328125
],
[
0.080078125,
0.61572265625,
0.3857421875,
0.61572265625,
0.3857421875,
0.62890625,
0.080078125,
0.62890625
],
[
0.50341796875,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.5048828125,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.6416015625,
0.5048828125,
0.6416015625
],
[
0.08087158203125,
0.64306640625,
0.33251953125,
0.64306640625,
0.33251953125,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.5048828125,
0.64599609375,
0.8896484375,
0.64599609375,
0.8896484375,
0.65576171875,
0.5048828125,
0.65576171875
],
[
0.50341796875,
0.6611328125,
0.8896484375,
0.6611328125,
0.8896484375,
0.671875,
0.50341796875,
0.671875
],
[
0.08087158203125,
0.66748046875,
0.3857421875,
0.66748046875,
0.3857421875,
0.68115234375,
0.08087158203125,
0.68115234375
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.09722900390625,
0.6875,
0.46728515625,
0.6875,
0.46728515625,
0.70068359375,
0.09722900390625,
0.70068359375
],
[
0.50341796875,
0.69140625,
0.8896484375,
0.69140625,
0.8896484375,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.080078125,
0.70263671875,
0.46728515625,
0.70263671875,
0.46728515625,
0.71630859375,
0.080078125,
0.71630859375
],
[
0.5009765625,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71728515625,
0.5009765625,
0.71728515625
],
[
0.080078125,
0.71826171875,
0.46728515625,
0.71826171875,
0.46728515625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73486328125,
0.81396484375,
0.73486328125,
0.81396484375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.74755859375,
0.468017578125,
0.74755859375,
0.468017578125,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.52294921875,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.080078125,
0.7646484375,
0.39794921875,
0.7646484375,
0.39794921875,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.5009765625,
0.76513671875,
0.890625,
0.7646484375,
0.890625,
0.77880859375,
0.5009765625,
0.77978515625
],
[
0.08087158203125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.5009765625,
0.77978515625,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81005859375,
0.46728515625,
0.81005859375,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.080078125,
0.8388671875
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.08087158203125,
0.85791015625,
0.4697265625,
0.85791015625,
0.4697265625,
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
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88818359375,
0.89208984375,
0.88818359375,
0.89208984375,
0.8984375,
0.50244140625,
0.8984375
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"Eventmodality",
"Imagemodality",
"APs images of low resolution and quality. Note that our",
"CE",
"method does not rely on any annotations ofevents intrain-",
"Similarity",
"ing, and the pseudo annotations by [1] for test are only used",
"a",
"for evaluation and comparison. As the events in the DDD17",
"Affinity",
"dataset are very sparse and noisy, we show more results on",
"grapr",
"the driving sequences in the MVSEC dataset [87], collected",
"Figure 5: Illustration of the affinity graph distillation loss between",
"for the stereo purpose. Moreover, we show qualitative re-",
"the event and image modalities.",
"sults on the E2Vid driving scene dataset [56] captured by",
"two modalities, we denote AT, and As, are the affinities",
"using a Samsung event camera (with higher resolution).",
"between the u-thnode and the w-th node obtained from the",
"Implementation details. For each label of DDD17 dataset",
"teacher and student, respectively, which is formulated as:",
"provided by [1], we use the events occurred in a 5Oms time",
"window before a label for prediction, as done in [l, 19].",
"1",
"CAG=",
"AT,-ASJ2",
"(7)",
"We also consider the event representation method in [1] for",
"HxWxo",
"u~RU~g",
"semantic segmentation on the DDD17 dataset, in addition",
"to that described in Sec. 3. For the event representation on",
"where R = {1,2, . , Hx W indicates all the nodes in the",
"MVSEC and E2VID datasets, we use the method in Sec. 3.",
"graph. The similarity between two nodes is calculated from",
"For the teacher T and students Seu and Saps, we adopt a",
"the aggregated features Fu and F, as Auy =-",
"FTF",
"FuT||2F2",
"segmentation network [11]. We use the following metric to",
"where F1 is the transposed feature vector of Fu [13, 38].",
"evaluate the performance. The intersection of union (loU)",
"Mutual Distillation (MD). When APS frames exist, Saps",
"score is calculated as the ratio of intersection and union be-",
"indeed can facilitate the learning of Seu. Since Saps and Se",
"tween the GT mask and the predicted segmentation mask",
"start from different initialconditions.they learn different",
"for each class. We use the mean IoU (MloU) to measure the",
"representations,andconsequently,theirprediction ofprob-",
"effectiveness, as done in [10, 11]. The segmentation maps",
"abilities can be an effective regularization to each other. We",
"arewith6classes,asdonein[1].Formoreimplementation",
"thus let Saps and Seu learn from each other's predictions",
"details (e.g., training), refer to the suppl. material.",
"via the KL divergence losses with a temperature parameter",
"T[82l.ThishelpsSeconvergetobetterminimaforbetter",
"4.1.1EvaluationonDDD17dataset",
"generalization to test data. The MD loss is formulated as:",
"Comparison. We first present the experimental results on",
"CMD = Ee,rapa~Xr KL[Seu(e)]Saps(aps), T] +",
"the DDD17 dataset [1]. We evaluate our method on the",
"(8)",
"Ee,rapa~Xr KL[Saps (raps)|Se(e), T] .",
"testsetandvarythe windowsize of eventsbetween 10, 50",
"and 250ms, as done in [1]. The quantitative and qualita-",
"In summary, the objective C of EvDistill is as follows :",
"tive results are shown in Table 1 and Fig.6.We compare",
"our method with two existing methods, EvSegNet [1] and",
"C=LCE+CBMR+AiCDA+^2CAG+A3CMD(9)",
"Vid2E [19] that uses synthetic version of DDD17 data. It",
"where A1, A2 and A3 are the hyper-parameters.",
"turns out that, without using labels, EvDistill significantly",
"improvesthe segmentation results on events and surpasses",
"4.Experiment and Evaluation",
"theexisting methodswitharound7%increaseinMloUus-",
"ing a multi-channel event representation. Segmentation us-",
"4.1.Event-based semantic segmentation",
"ing the voxel grid representation is slightly less effective",
" Semantic segmentation is a task that aims to assign a se-",
"than that of using the multi-channel representation. Mean-",
"mantic label, e.g., road, car, in a given scene to each pixel.",
"while, on the time interval of 10ms and 25Oms, EvDistill",
"Unlike image data, annotating events requires special pro-",
"also shows a significant increase of MloU by around 7.5%",
"cessing for the raw event streams. Besides, it is challenging",
"and 9.5% than those of EvSegNet, respectively.",
"to correctly label the pixels due to the sparsity of events and",
"We also demonstrate that our method significantly en-",
"lacking information (e.g., material and textures).",
"hances the semantic segmentation performance on the APS",
"Datasets. We use the publicly available driving scene",
"frames in Table 1. Quantitatively, without resorting to the",
"dataset DDD17 [5], which includes both events and APS",
"pseudo labels, our method surpasses EvSegNet by a large",
"frames recorded by a DAVIS346 event camera. In [1],",
"margin with around 12% increase of MloU. This reflects",
"19,840 APS frames are utilized to generate pseudo anno-",
"that our method dramatically minimizes the domain gap",
"tations(6 classes) based on a pretrained network for events",
"between APSframes andlabeled sourcedata and distills",
"(15.950 fortraining and3.890fortest).However, such a",
"theknowledgefromthe teachernetworkto learn a student",
"way leads to less precise segmentation labels because there",
"network Saps showing better segmentation capability. The",
"is a considerable domain gap between the source data and",
"results on both events and APS frames indicate that our",
"5"
] |
{
"arxiv_id": "2111.12341",
"language": "en",
"page_idx": 4
}
|
||
2309.03453-en-0
|
2309.03453-en-0.jpg
|
Published as a conference paper at ICLR 2024
SYNCDREAMER: GENERATING MULTIVIEW-
CONSISTENT IMAGES FROM A SINGLE-VIEW IMAGE
Yuan Liu$^{1,2*}$ Cheng Lin$^{2*}$ Zijiao Zeng$^2$ Xiaoxiao Long$^{1+}$ Lingjie Liu$^{3}$
Taku Komura$^{1}$ $\bar{\textbf{Wenping Wang}^{4+ }}$
The university of Hong Kong $^2$Tencent Games$\:^3$University of Pennsylvania
$^{4}$Texas A&M University
zijiaogtencent.com
$\{$yuanly, chlin, xxlong$\}$gconnect.hku.hk
lingjie.liu@seas.upenn.edu wenping@tamu.edu
### ABSTRACT
In this paper, we present a novel diffusion model called SyncDreamer that generates multiview-consistent images from a single-view image. Using pretrained large-scale 2D diffusion models, recent work Zero123 (Liu et al., 2023b) demonstrates the ability to generate plausible novel views from a single-view image of an object. However, maintaining consistency in geometry and colors for the generated images remains a challenge. To address this issue, we propose a synchronized multiview diffusion model that models the joint probability distribution of multiview images, enabling the generation of multiview-consistent images in a single reverse process. SyncDreamer synchronizes the intermediate states of all the generated images at every step of the reverse process through a 3D-aware feature attention mechanism that correlates the corresponding features across different views. Experiments show that SyncDreamer generates images with high consistency across different views, thus making it well-suited for various 3D generation tasks such as novel-view-synthesis, text-to-3D, and image-to-3D. Project page: https://liuyuan-pal.github.io/SyncDreamer/.
## 1 INTRODUCTION
Humans possess a remarkable ability to perceive 3D structures from a single image. When presented with an image of an object, humans can easily imagine the other views of the object. Despite great progress (Yao et al., 2018; Tewari et al., 2020; Wang et al., 2021; Mildenhall et al., 2020; Åie et al., 2022) brought by neural networks in computer vision or graphics fields for extracting 3D information from images, generating multiview-consistent images from a single-view image of an object is still a challenging problem due to the limited 3D information available in an image.
Recently, diffusion models (Rombach et al., 2022; Ho et al., 2020) have demonstrated huge success in 2D image generation, which unlocks new potential for 3D generation tasks. However, directly training a generalizable 3D diffusion model (Wang et al., 2023b; Jun & Nichol, 2023; Nichol et al., 2022; Müller et al., 2023) usually requires a large amount of 3D data while existing 3D datasets are insufficient for capture the complexity of arbitrary 3D shapes. Therefore, recent methods (Poole et al., 2023; Wang et al., 2023a;d; Lin et al., 2023; Chen et al., 2023b) resort to distilling pretrained text-to-image diffusion models for creating 3D models from texts, which shows impressive results on this text-to-3D task. Some works (Tang et al., 2023a; Melas-Kyriazi et al., 2023; Xu et al., 2022; Raj et al., 2023) extend such a distillation process to train a neural radiance field (Mildenhall et al., 2020) (NeRF) for the image-to-3D task. In order to utilize pretrained text-to-image models, these methods have to perform textual inversion (Gal et al., 2022) to find a suitable text description of the input image. However, the distillation process along with the textual inversion usually takes a long time to generate a single shape and requires tedious parameter tuning for satisfactory quality. Moreover, due to the abundance of specific details in an image, such as object category, appearance,
$^*$Equal contribution.
$^{\dagger}$Corresponding Authors.
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.176513671875,
0.10101318359375,
0.70751953125,
0.10162353515625,
0.70751953125,
0.11932373046875,
0.176513671875,
0.11871337890625
],
[
0.1748046875,
0.126953125,
0.79345703125,
0.1275634765625,
0.79345703125,
0.1458740234375,
0.1748046875,
0.14453125
],
[
0.404296875,
0.169189453125,
0.5,
0.169189453125,
0.5,
0.184326171875,
0.404296875,
0.184326171875
],
[
0.51708984375,
0.17041015625,
0.64453125,
0.17041015625,
0.64453125,
0.1849365234375,
0.51708984375,
0.1849365234375
],
[
0.64892578125,
0.17041015625,
0.73876953125,
0.17041015625,
0.73876953125,
0.1849365234375,
0.64892578125,
0.1849365234375
],
[
0.295654296875,
0.1717529296875,
0.390625,
0.169189453125,
0.390625,
0.184326171875,
0.295654296875,
0.186279296875
],
[
0.1871337890625,
0.1717529296875,
0.27783203125,
0.1685791015625,
0.279296875,
0.184326171875,
0.188720703125,
0.1875
],
[
0.39453125,
0.1749267578125,
0.40185546875,
0.1749267578125,
0.40185546875,
0.1787109375,
0.39453125,
0.1787109375
],
[
0.1829833984375,
0.186279296875,
0.2900390625,
0.1849365234375,
0.2900390625,
0.199462890625,
0.1829833984375,
0.2008056640625
],
[
0.310546875,
0.186279296875,
0.43310546875,
0.1849365234375,
0.43310546875,
0.199462890625,
0.310546875,
0.2008056640625
],
[
0.1846923828125,
0.2008056640625,
0.386474609375,
0.2020263671875,
0.386474609375,
0.216552734375,
0.1846923828125,
0.21533203125
],
[
0.39453125,
0.2020263671875,
0.515625,
0.2020263671875,
0.515625,
0.21533203125,
0.39453125,
0.21533203125
],
[
0.53173828125,
0.199462890625,
0.7236328125,
0.2008056640625,
0.7236328125,
0.21533203125,
0.53173828125,
0.2139892578125
],
[
0.1829833984375,
0.2139892578125,
0.346435546875,
0.216552734375,
0.346435546875,
0.2310791015625,
0.1829833984375,
0.2291259765625
],
[
0.1846923828125,
0.2298583984375,
0.56396484375,
0.2298583984375,
0.56396484375,
0.2442626953125,
0.1846923828125,
0.2442626953125
],
[
0.56396484375,
0.2342529296875,
0.576171875,
0.2342529296875,
0.576171875,
0.238037109375,
0.56396484375,
0.238037109375
],
[
0.57177734375,
0.2291259765625,
0.75634765625,
0.2298583984375,
0.75634765625,
0.2442626953125,
0.57177734375,
0.24365234375
],
[
0.1854248046875,
0.2442626953125,
0.4599609375,
0.2442626953125,
0.4599609375,
0.25830078125,
0.1854248046875,
0.25830078125
],
[
0.451904296875,
0.246826171875,
0.615234375,
0.246826171875,
0.615234375,
0.2568359375,
0.451904296875,
0.2568359375
],
[
0.021240234375,
0.260009765625,
0.06451416015625,
0.260009765625,
0.06451416015625,
0.406005859375,
0.021240234375,
0.406005859375
],
[
0.453369140625,
0.294189453125,
0.54736328125,
0.294189453125,
0.54736328125,
0.309326171875,
0.453369140625,
0.309326171875
],
[
0.233642578125,
0.32763671875,
0.7646484375,
0.32763671875,
0.7646484375,
0.341552734375,
0.233642578125,
0.341552734375
],
[
0.233642578125,
0.34228515625,
0.76708984375,
0.34228515625,
0.76708984375,
0.35595703125,
0.233642578125,
0.35595703125
],
[
0.233642578125,
0.35595703125,
0.7646484375,
0.35595703125,
0.7646484375,
0.369384765625,
0.233642578125,
0.369384765625
],
[
0.2320556640625,
0.37060546875,
0.765625,
0.369384765625,
0.765625,
0.38330078125,
0.2320556640625,
0.3837890625
],
[
0.233642578125,
0.3837890625,
0.765625,
0.3837890625,
0.765625,
0.397705078125,
0.233642578125,
0.397705078125
],
[
0.021240234375,
0.39453125,
0.06451416015625,
0.395751953125,
0.062103271484375,
0.49560546875,
0.0187835693359375,
0.494384765625
],
[
0.233642578125,
0.397705078125,
0.76708984375,
0.397705078125,
0.76708984375,
0.412353515625,
0.233642578125,
0.412353515625
],
[
0.2344970703125,
0.41357421875,
0.76123046875,
0.41357421875,
0.76123046875,
0.423583984375,
0.2344970703125,
0.423583984375
],
[
0.233642578125,
0.425537109375,
0.765625,
0.425537109375,
0.765625,
0.439453125,
0.233642578125,
0.439453125
],
[
0.2320556640625,
0.438232421875,
0.7646484375,
0.439453125,
0.7646484375,
0.453857421875,
0.2320556640625,
0.45263671875
],
[
0.2344970703125,
0.455078125,
0.7646484375,
0.455078125,
0.7646484375,
0.46533203125,
0.2344970703125,
0.46533203125
],
[
0.233642578125,
0.46728515625,
0.765625,
0.46728515625,
0.765625,
0.48095703125,
0.233642578125,
0.48095703125
],
[
0.2320556640625,
0.48095703125,
0.76318359375,
0.48095703125,
0.76318359375,
0.494384765625,
0.2320556640625,
0.494384765625
],
[
0.0228729248046875,
0.4892578125,
0.05963134765625,
0.4892578125,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.233642578125,
0.494384765625,
0.765625,
0.494384765625,
0.765625,
0.50830078125,
0.233642578125,
0.50830078125
],
[
0.2320556640625,
0.5068359375,
0.765625,
0.50830078125,
0.765625,
0.52294921875,
0.2320556640625,
0.521484375
],
[
0.2320556640625,
0.521484375,
0.51953125,
0.52099609375,
0.51953125,
0.53515625,
0.2320556640625,
0.5361328125
],
[
0.17724609375,
0.5693359375,
0.1895751953125,
0.5693359375,
0.1895751953125,
0.5791015625,
0.17724609375,
0.5791015625
],
[
0.2017822265625,
0.5673828125,
0.338134765625,
0.5673828125,
0.338134765625,
0.58056640625,
0.2017822265625,
0.58056640625
],
[
0.1732177734375,
0.599609375,
0.82421875,
0.599609375,
0.82421875,
0.61279296875,
0.1732177734375,
0.61279296875
],
[
0.1748046875,
0.6142578125,
0.82421875,
0.6142578125,
0.82421875,
0.62744140625,
0.1748046875,
0.62744140625
],
[
0.1732177734375,
0.62744140625,
0.82421875,
0.626953125,
0.82421875,
0.6416015625,
0.1732177734375,
0.642578125
],
[
0.1748046875,
0.642578125,
0.82421875,
0.642578125,
0.82421875,
0.65576171875,
0.1748046875,
0.65576171875
],
[
0.1732177734375,
0.65576171875,
0.826171875,
0.65576171875,
0.826171875,
0.67041015625,
0.1732177734375,
0.67041015625
],
[
0.1732177734375,
0.67041015625,
0.69287109375,
0.67041015625,
0.69287109375,
0.6845703125,
0.1732177734375,
0.6845703125
],
[
0.1732177734375,
0.689453125,
0.826171875,
0.68994140625,
0.826171875,
0.705078125,
0.1732177734375,
0.7041015625
],
[
0.1732177734375,
0.7041015625,
0.82421875,
0.7041015625,
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,
0.73193359375,
0.826171875,
0.73193359375,
0.826171875,
0.7470703125,
0.1748046875,
0.7470703125
],
[
0.176513671875,
0.74755859375,
0.82275390625,
0.74755859375,
0.82275390625,
0.75830078125,
0.176513671875,
0.75830078125
],
[
0.1748046875,
0.76025390625,
0.82421875,
0.76025390625,
0.82421875,
0.7734375,
0.1748046875,
0.7734375
],
[
0.1748046875,
0.7734375,
0.826171875,
0.7734375,
0.826171875,
0.787109375,
0.1748046875,
0.787109375
],
[
0.1748046875,
0.78857421875,
0.82421875,
0.78857421875,
0.82421875,
0.8017578125,
0.1748046875,
0.8017578125
],
[
0.1748046875,
0.8017578125,
0.82421875,
0.8017578125,
0.82421875,
0.81494140625,
0.1748046875,
0.81494140625
],
[
0.1748046875,
0.81494140625,
0.82421875,
0.81494140625,
0.82421875,
0.8291015625,
0.1748046875,
0.8291015625
],
[
0.176513671875,
0.83056640625,
0.82421875,
0.83056640625,
0.82421875,
0.84130859375,
0.176513671875,
0.84130859375
],
[
0.1748046875,
0.84326171875,
0.82421875,
0.84326171875,
0.82421875,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.1732177734375,
0.85791015625,
0.82275390625,
0.85791015625,
0.82275390625,
0.8720703125,
0.1732177734375,
0.8720703125
],
[
0.1732177734375,
0.8720703125,
0.82421875,
0.8720703125,
0.82421875,
0.88525390625,
0.1732177734375,
0.88525390625
],
[
0.1953125,
0.89794921875,
0.316162109375,
0.89794921875,
0.316162109375,
0.912109375,
0.1953125,
0.912109375
],
[
0.1953125,
0.91015625,
0.343017578125,
0.9111328125,
0.343017578125,
0.92529296875,
0.1953125,
0.9248046875
],
[
0.494384765625,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2024",
"SYNCDREAMER: GENERATING MULTIVIEW-",
"CONSISTENT IMAGES FROM A SINGLE-VIEW IMAGE",
"Zijiao Zeng?",
"Xiaoxiao Long't",
"Lingjie Liu3",
"Cheng Lin?*",
"Yuan Liul,2*",
"",
"Taku Komura'",
"Wenping Wang4",
"IThe university of Hong Kong",
"2Tencent Games",
"3University of Pennsylvania",
"4Texas A&MUniversity",
"(yuanly,chlin,xxlong)@connect.hku.hk",
"",
"zijiao@tencent.com",
"lingjie.liu@seas.upenn.edu",
"wenping@tamu.edu",
" ",
"ABSTRACT",
"In this paper, we present a novel diffusion model called SyncDreamer that gen-",
"erates multiview-consistent images from a single-view image. Using pretrained",
"large-scale 2D diffusion models, recent work Zero123 (Liu et al., 2023b) demon-",
"stratesthe abilitytogenerateplausiblenovelviewsfroma single-view imageofan",
"object. However, maintaining consistency in geometry and colors for the gener-",
"['s]",
"ated images remains a challenge. To address this issue, we propose a synchronized",
"multiviewdiffusionmodelthatmodelsthejointprobabilitydistributionofmulti",
"view images, enabling the generation of multiview-consistent images in a single",
"reverse process. SyncDreamer synchronizes the intermediate states of all the gen-",
"eratedimagesat every stepof thereverseprocessthrougha3D-awarefeature",
"attention mechanism that correlates the corresponding features across different",
"views. Experiments show that SyncDreamer generates images with high consis-",
"0",
"tency across different views, thus making it well-suited for various 3D generation",
"tasks such as novel-view-synthesis, text-to-3D, and image-to-3D. Project page:",
"https:/liuyuan-pal.github.io/SyncDreamer/.",
"1",
"INTRODUCTION",
"Humans possess a remarkable ability to perceive 3D structures from a single image. When presented",
"with an image of an object, humans can easily imagine the other views of the object. Despite great",
"progress (Yao et al., 2018; Tewari et al., 2020; Wang et al., 2021; Mildenhall et al., 2020; Xie et al.,",
"2022)broughtbyneural networks incomputer vision orgraphicsfieldsforextracting3D information",
"from images, generating multiview-consistent images from a single-view image of an object is still",
"a challenging problem due to the limited 3D information available in an image.",
"Recently, diffusion models (Rombach et al., 2022; Ho et al., 2020) have demonstrated huge success",
"in 2D image generation, which unlocks new potential for 3D generation tasks. However, directly",
"training a generalizable 3D diffusion model (Wang et al., 2023b; Jun & Nichol, 2023; Nichol et al.,",
"2022; Miller et al., 2023) usually requires a large amount of 3D data while existing 3D datasets",
"areinsufficientfor capturethecomplexity ofarbitrary3D shapes.Therefore,recentmethods(Poole",
"et al., 2023; Wang et al., 2023a;d; Lin et al., 2023; Chen et al., 2023b) resort to distilling pretrained",
"text-to-image diffusion models for creating 3D models from texts, which shows impressive results",
"on this text-to-3D task. Some works (Tang et al., 2023a; Melas-Kyriazi et al., 2023; Xu et al., 2022;",
"Raj et al., 2023) extend such a distillation process to train a neural radiance field (Mildenhall et al.,",
"2020) (NeRF) for the image-to-3D task. In order to utilize pretrained text-to-image models, these",
"methodshavetoperformtextual inversion(Gal et al.,2022)tofind a suitabletextdescription of",
"the input image. However, the distillation process along with the textual inversion usually takes a",
"long time to generate a single shape and requires tedious parameter tuning for satisfactory quality.",
"Moreover, due to the abundance of specific details in an image, such as object category, appearance,",
"*Equal contribution.",
"fCorresponding Authors",
"1"
] |
{
"arxiv_id": "2309.03453",
"language": "en",
"page_idx": 0
}
|
||
2309.03453-en-1
|
2309.03453-en-1.jpg
|
Published as a conference paper at ICLR 2024
<FigureHere>
Figure 1: SyncDreamer is able to generate multiview-consistent images from a single-view input
image of arbitrary objects. The generated multiview images can be used for mesh reconstruction by
reconstruction methods like NeuS (Wang et al., 2021) without using SDS (Poole et al., 2023) loss.
and pose, it is challenging to accurately represent an image using a single word embedding, which
results in a decrease in the quality of 3D shapes reconstructed by the distillation method.
Instead of distillation, some recent works (Watson et al., 2022; Gu et al., 2023b; Deng et al., 2023a; Zhou & Tulsiani, 2023; Tseng et al., 2023; Yu et al., 2023b; Chan et al., 2023; Tewari et al., 2023; Zhang et al., 2023b; Xiang et al., 2023) apply 2D diffusion models to directly generate multiview images for the 3D reconstruction task. The key problem is how to maintain the multiview consistency when generating images of the same object. To improve the multiview consistency, these methods allow the diffusion model to condition on the input images (Zhou & Tulsiani, 2023; Tseng et al., 2023; Watson et al., 2022; Liu et al., 2023b; Yu et al.., 2023b), previously generated images (Tewari et al., 2023; Chan et al., 2023) or renderings from a neural field (Gu et al., 2023b). Although some impressive results are achieved for specific object categories from ShapeNet (Chang et al., 2015) or $\hat{\text{Co3D (Reizenstein et al., 2021), how to design a diffusion model to generate multiview-consistent}}$ images for arbitrary objects still remains unsolved.
In this paper, we propose a simple yet effective framework to generate multiview-consistent images for the single-view 3D reconstruction of arbitrary objects. The key idea is to extend the diffusion framework (Ho et al., 2020) to model the joint probability distribution of multiview images. We show that modeling the joint distribution can be achieved by introducing a synchronized multiview diffusion model. Specifically, for $N$ target views to be generated, we construct $N$ shared noise predictors respectively. The reverse diffusion process simultaneously generates $N$ images by $N$ corresponding noise predictors, where information across different images is shared among noise predictors by attention layers on every denoising step. Thus, we name our framework SyncDreamer $\dot{\text{which synchronizes intermediate states of all noise predictors on every step in the reverse process.}}$ SyncDreamer has the following characteristics that make it a competitive tool for lifting 2D singleview images to 3D. First, SyncDreamer retains strong generalization ability by initializing its weights from the pretrained Zero123 (Liu et al., 2023b) model which is finetuned from the Stable $\tilde{\text{Diffusion model(Rombach et al.,2022)on the Objaverse (Deitke et al., 2023b) dataset. Thus,}}$ SyncDreamer is able to reconstruct shapes from both photorealistic images and hand drawings as shown in Fig. 1. Second, SyncDreamer makes the single-view reconstruction easier than the distillation methods. Because the generated images are consistent in both geometry and appearance, we can simply run a vanilla NeRF (Mildenhall et al., 2020) or a vanilla NeuS (Wang et al., 2021) without using any special losses for reconstruction. Given the generated images, one can easily reckon the final reconstruction quality while it is hard for distillation methods to know the output reconstruction quality beforehand. Third, SyncDreamer maintains creativity and diversity when inferring 3D information, which enables generating multiple reasonable objects from a given image as shown in Fig. 4. In comparison, previous distillation methods can only converge to one single shape.
We quantitatively compare SyncDreamer with baseline methods on the Google Scanned Object (Downs et al., 2022) dataset. The results show that, in comparison with baseline methods, SyncDreamer is able to generate more consistent images and reconstruct better shapes from input single-view images. We further demonstrate that SyncDreamer supports various styles of 2D input
2
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.41015625,
0.270751953125,
0.44775390625,
0.270751953125,
0.44775390625,
0.281005859375,
0.41015625,
0.281005859375
],
[
0.74853515625,
0.270751953125,
0.771484375,
0.270751953125,
0.771484375,
0.27783203125,
0.74853515625,
0.27783203125
],
[
0.451904296875,
0.274658203125,
0.483642578125,
0.274658203125,
0.483642578125,
0.27783203125,
0.451904296875,
0.27783203125
],
[
0.2115478515625,
0.272705078125,
0.2607421875,
0.272705078125,
0.2607421875,
0.27978515625,
0.2115478515625,
0.27978515625
],
[
0.1732177734375,
0.292236328125,
0.82275390625,
0.292236328125,
0.82275390625,
0.30615234375,
0.1732177734375,
0.30615234375
],
[
0.176513671875,
0.309326171875,
0.82177734375,
0.309326171875,
0.82177734375,
0.3193359375,
0.176513671875,
0.3193359375
],
[
0.1732177734375,
0.3212890625,
0.81884765625,
0.3212890625,
0.81884765625,
0.335205078125,
0.1732177734375,
0.335205078125
],
[
0.1732177734375,
0.352783203125,
0.82275390625,
0.354248046875,
0.82275390625,
0.36865234375,
0.1732177734375,
0.367431640625
],
[
0.1748046875,
0.36865234375,
0.75341796875,
0.36865234375,
0.75341796875,
0.3818359375,
0.1748046875,
0.3818359375
],
[
0.1748046875,
0.389404296875,
0.82275390625,
0.389404296875,
0.82275390625,
0.40283203125,
0.1748046875,
0.40283203125
],
[
0.1748046875,
0.40283203125,
0.82275390625,
0.40283203125,
0.82275390625,
0.416748046875,
0.1748046875,
0.416748046875
],
[
0.1748046875,
0.417236328125,
0.82177734375,
0.417236328125,
0.82177734375,
0.431884765625,
0.1748046875,
0.431884765625
],
[
0.1748046875,
0.431884765625,
0.82275390625,
0.431884765625,
0.82275390625,
0.44580078125,
0.1748046875,
0.44580078125
],
[
0.1748046875,
0.44580078125,
0.82275390625,
0.44580078125,
0.82275390625,
0.458984375,
0.1748046875,
0.458984375
],
[
0.176513671875,
0.460205078125,
0.82177734375,
0.460205078125,
0.82177734375,
0.47021484375,
0.176513671875,
0.47021484375
],
[
0.176513671875,
0.472900390625,
0.82421875,
0.472900390625,
0.82421875,
0.486083984375,
0.176513671875,
0.486083984375
],
[
0.17724609375,
0.488037109375,
0.82275390625,
0.488037109375,
0.82275390625,
0.498779296875,
0.17724609375,
0.498779296875
],
[
0.1748046875,
0.50048828125,
0.82421875,
0.50048828125,
0.82421875,
0.5146484375,
0.1748046875,
0.5146484375
],
[
0.1748046875,
0.5146484375,
0.82421875,
0.5146484375,
0.82421875,
0.52783203125,
0.1748046875,
0.52783203125
],
[
0.1748046875,
0.52880859375,
0.5087890625,
0.52880859375,
0.5087890625,
0.54248046875,
0.1748046875,
0.54248046875
],
[
0.1748046875,
0.5498046875,
0.82421875,
0.5498046875,
0.82421875,
0.56298828125,
0.1748046875,
0.56298828125
],
[
0.1748046875,
0.56298828125,
0.82421875,
0.56298828125,
0.82421875,
0.57763671875,
0.1748046875,
0.57763671875
],
[
0.176513671875,
0.5791015625,
0.82275390625,
0.5791015625,
0.82275390625,
0.5888671875,
0.176513671875,
0.5888671875
],
[
0.1748046875,
0.59130859375,
0.82275390625,
0.59130859375,
0.82275390625,
0.60498046875,
0.1748046875,
0.60498046875
],
[
0.1748046875,
0.60498046875,
0.82421875,
0.60498046875,
0.82421875,
0.61865234375,
0.1748046875,
0.61865234375
],
[
0.1748046875,
0.61865234375,
0.82275390625,
0.61865234375,
0.82275390625,
0.6318359375,
0.1748046875,
0.6318359375
],
[
0.1748046875,
0.63330078125,
0.82275390625,
0.63330078125,
0.82275390625,
0.646484375,
0.1748046875,
0.646484375
],
[
0.1748046875,
0.646484375,
0.82421875,
0.646484375,
0.82421875,
0.66015625,
0.1748046875,
0.66015625
],
[
0.1748046875,
0.66015625,
0.814453125,
0.66015625,
0.814453125,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1732177734375,
0.6796875,
0.82177734375,
0.68115234375,
0.82177734375,
0.69580078125,
0.1732177734375,
0.6943359375
],
[
0.1748046875,
0.69580078125,
0.82421875,
0.69580078125,
0.82421875,
0.708984375,
0.1748046875,
0.708984375
],
[
0.1748046875,
0.708984375,
0.82275390625,
0.708984375,
0.82275390625,
0.72265625,
0.1748046875,
0.72265625
],
[
0.1732177734375,
0.7216796875,
0.82275390625,
0.72265625,
0.82275390625,
0.7373046875,
0.1732177734375,
0.736328125
],
[
0.1748046875,
0.7373046875,
0.82421875,
0.7373046875,
0.82421875,
0.75048828125,
0.1748046875,
0.75048828125
],
[
0.1748046875,
0.751953125,
0.82275390625,
0.751953125,
0.82275390625,
0.76513671875,
0.1748046875,
0.76513671875
],
[
0.1748046875,
0.76513671875,
0.82421875,
0.76513671875,
0.82421875,
0.77880859375,
0.1748046875,
0.77880859375
],
[
0.1748046875,
0.77880859375,
0.82177734375,
0.77880859375,
0.82177734375,
0.79248046875,
0.1748046875,
0.79248046875
],
[
0.1748046875,
0.79248046875,
0.82421875,
0.79248046875,
0.82421875,
0.80615234375,
0.1748046875,
0.80615234375
],
[
0.1748046875,
0.80615234375,
0.82177734375,
0.80615234375,
0.82177734375,
0.8193359375,
0.1748046875,
0.8193359375
],
[
0.1732177734375,
0.8193359375,
0.82421875,
0.82080078125,
0.82421875,
0.833984375,
0.1732177734375,
0.83349609375
],
[
0.176513671875,
0.833984375,
0.826171875,
0.833984375,
0.826171875,
0.84765625,
0.176513671875,
0.84765625
],
[
0.1748046875,
0.84765625,
0.78369140625,
0.84765625,
0.78369140625,
0.861328125,
0.1748046875,
0.861328125
],
[
0.1748046875,
0.86865234375,
0.8203125,
0.86865234375,
0.8203125,
0.8818359375,
0.1748046875,
0.8818359375
],
[
0.1748046875,
0.88330078125,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.1748046875,
0.896484375
],
[
0.17236328125,
0.89599609375,
0.82421875,
0.896484375,
0.82421875,
0.9111328125,
0.17236328125,
0.91015625
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
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 2024",
"Generate",
"Mesh",
"",
"InputImag",
"Figure l: SyncDreamer is able to generate multiview-consistent images from a single-view input",
"imageofarbitraryobjects.Thegeneratedmultiviewimagescanbeusedformeshreconstructionby",
"reconstruction methods like NeuS (Wang et al., 2021) without using SDS (Poole et al., 2023) loss.",
"and pose, it is challenging to accurately represent an image using a single word embedding, which",
"results in a decrease in the quality of 3D shapes reconstructed by the distillation method.",
"Instead of distillation, some recent works (Watson et al., 2022; Gu et al., 2023b; Deng et al., 2023a",
"Zhou & Tulsiani, 2023: Tseng et al., 2023; Yu et al., 2023b; Chan et al., 2023; Tewari et al., 2023:",
"Zhang et al., 2023b; Xiang et al., 2023) apply 2D diffusion models to directly generate multiview im-",
"ages for the 3D reconstruction task. The key problem is how to maintain the multiview consistency",
"when generating images of the same object. To improve the multiview consistency, these methods",
"allow thediffusion model to condition on theinputimages (Zhou&Tulsiani,2023:Tseng etal..",
"2023; Watson et al., 2022; Liu et al., 2023b; Yu et al., 2023b), previously generated images (Tewari",
"etal.,2023:Chanetal.,2023)orrenderingsfromaneuralfield(Guetal.,2023b).Although some",
"impressive results are achieved for specific object categories from ShapeNet (Chang et al., 2015) or",
"Co3D (Reizenstein et al., 2021), how to design a diffusion model to generate multiview-consistent",
"images forarbitrary objects still remains unsolved.",
"In this paper, we propose a simple yet effective framework to generate multiview-consistent images",
"for the single-view 3D reconstruction of arbitrary objects. The key idea is to extend the diffusion",
"framework (Hoetal..202O)tomodel thejointprobability distribution ofmultiview images.We",
"show that modeling the joint distribution can be achieved by introducing a synchronized multiview",
"diffusion model. Specifically, for N target views to be generated, we construct N shared noise",
"predictors respectively. The reverse diffusion process simultaneously generates N images by N",
"corresponding noise predictors, where information across different images is shared among noise",
"predictors by attention layers on every denoising step. Thus, we name our framework SyncDreamer",
"which synchronizes intermediate states of all noise predictors on every step in the reverse process.",
"SyncDreamer has the following characteristics that make it a competitive tool for lifting 2D single-",
"view images to 3D. First, SyncDreamer retains strong generalization ability by initializing its",
"weights from the pretrained Zero123 (Liu et al., 2023b) model which is finetuned from the Sta-",
"ble Diffusion model (Rombach et al., 2022) on the Objaverse (Deitke et al., 2023b) dataset. Thus,",
"SyncDreamer is able to reconstruct shapes from both photorealistic images and hand drawings as",
"shown in Fig. 1. Second, SyncDreamer makes the single-view reconstruction easier than the distil-",
"lation methods. Because the generated images are consistent in both geometry and appearance, we",
"can simply run a vanilla NeRF (Mildenhall et al., 2020) or a vanilla NeuS (Wang et al., 2021) with-",
"out using any special losses for reconstruction. Given the generated images, one can easily reckon",
"the final reconstruction quality while it is hard for distillation methods to know the output recon-",
"struction quality beforehand.Third, SyncDreamer maintains creativity and diversity when inferring",
"3D information, which enables generating multiple reasonable objects from a given image as shown",
"in Fig. 4. In comparison, previous distillation methods can only converge to one single shape.",
"We quantitatively compare SyncDreamer with baseline methods on the Google Scanned Ob-",
"ject (Downs et al., 2022) dataset. The results show that, in comparison with baseline methods",
"SyncDreamer is able to generate more consistent images and reconstruct better shapes from input",
"single-view images. We further demonstrate that SyncDreamer supports various styles of 2D input",
"2"
] |
{
"arxiv_id": "2309.03453",
"language": "en",
"page_idx": 1
}
|
||
2309.03453-en-2
|
2309.03453-en-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
like cartoons, sketches, ink paintings, and oil paintings for generating consistent views and reconstructing 3D shapes, which verifies the effectiveness of SyncDreamer in lifting 2D images to 3D. 2 RELATED WORK
## 2.1 DIFFUSION MODELS
Diffusion models (Ho et al., 2020; Rombach et al., 2022; Croitoru et al., 2023) have shown impressive results on 2D image generation. Concurrent work MVDiffusion (Tang et al., 2023b) also adopts the multiview diffusion formulation to synthesize textures or panoramas with known geometry. We propose similar formulations in SyncDreamer but with unknown geometry. MultiDiffusion (Bar-Tal et al., 2023) and SyncDiffusion (Lee et al., 2023) correlate multiple diffusion models for different regions of a 2D image. Many recent works (Nichol et al., 2022; Jun & Nichol, 2023; Müller et al., 2023; Zhang et al., 2023a; Liu et al., 2023d; Wang et al., 2023b; Gupta et al., 2023; Cheng et al., 2023; Karnewar et al., 2023b; Anciukevičius et al., 2023; Zeng et al., 2022; Erkoç et al., 2023; Chen et al., 2023a; Kim et al., 2023; Ntavelis et al., 2023; Gu et al., 2023a; Karnewar et al., 2023a) try to repeat the success of diffusion models on the 3D generation task. However, the scarcity of 3D data makes it difficult to directly train diffusion models on 3D and the resulting generation quality is still much worse and less generalizable than the counterpart image generation models, though some works (Anciukevičius et al., 2023; Chen et al., 2023a; Karnewar et al., 2023b) are trying to only use 2D images for training 3D diffusion models.
2.2 USING 2D DIFFUSION MODELS FOR 3D
Instead of directly learning a 3D diffusion model, many works resort to using high-quality 2D diffusion models (Rombach et al., 2022; Saharia et al., 2022) for 3D tasks. Pioneer works DreamFusion (Poole et al., 2023) and SJC (Wang et al., 2023a) propose to distill a 2D text-to-image generation model to generate 3D shapes from texts. Follow-up works (Chen et al., 2023b; Wang et al., 2023d; Seo et al., 2023a; Yu et al., 2023a; Lin et al., 2023; Seo et al., 2023b; Tsalicoglou et al., 2023; Zhu & & Zhuang, 2023; Huang et al., 2023; Armandpour et al., 2023; Wu et al., 2023; Chen et al., 2023c) improve such text-to-3D distillation methods in various aspects. Many works (Tang et al., 2023a; Melas-Kyriazi et al., 2023; Qian et al., 2023; Xu et al., 2022; Raj et al., 2023; Shen et al., 2023) also apply such a distillation pipeline in the single-view reconstruction task. Though some impressive results are achieved, these methods usually require a long time for textual inversion (Liu et al., 2023a) and NeRF optimization and they do not guarantee to get satisfactory results.
Other works (Watson et al., 2022; Gu et al., 2023b; Deng et al., 2023a; Zhou " Zhou " Tseng et al., 2023; Chan et al., 2023; Yu et al., 2023b; Tewari et al., 2023; Yoo et al., 2023; Szymanowicz et al., 2023; Tang et al., 2023b; Xiang et al., 2023; Liu et al., 2023c; Lei et al., 2022) directly apply the 2D diffusion models to generate multiview images for 3D reconstruction. (Tseng et al., 2023; Yu et al., 2023b) are conditioned on the input image by attention layers for novel-view synthesis in indoor scenes. Our method also uses attention layers but is intended for object reconstruction. (Xiang et al., 2023; Zhang et al., 2023b) resort to estimated depth maps to warp and inpaint for novel-view image generation, which strongly relies on the performance of the external single-view depth estimator. Two concurrent works (Chan et al., 2023; Tewari et al., 2023) generate new images in an autoregressive render-and-generate manner, which demonstrates good performances on specific object categories or scenes. In comparison, SyncDreamer is targeted to reconstruct arbitrary objects and generates all images in one reverse process. The concurrent work Viewset Diffusion (Szymanowicz et al., 2023) shares a similar idea to generate a set of images. The differences betweem SyncDreamer and Viewset Diffusion are that SyncDreamer does not require predicting a radiance field like Viewset Diffusion but only uses attention to synchronize the states among views and SyncDreamer fixes the viewpoints of generated views for better convergence. Another concurrent work MVDream (Shi et al., 2023) also proposes multiview generation for the text-to-3D task while our work aims to reconstruct shapes from single-view images.
2.3 OTHER SINGLE-VIEW RECONSTRUCTION METHODS
Single-view reconstruction is a challenging ill-posed problem. Before the prosperity of generative models used in 3D reconstruction, there are many works (Tatarchenko et al., 2019; Fu et al., 2021; Kato& Harada, 2019; Li et al., 2020; Fahim et al., 2021) that reconstruct 3D shapes from single-view images by regression (Li et al., 2020) or retrieval (Tatarchenko et al., 2019), which have difficulty in generalizing to new categories. Recent NeRF-GAN methods (Niemeyer&Geiger,2021;Chan et al., 2022; Gu et al., 2021; Schwarz et al., 2020; Gao et al., 2022; Deng et al., 2023b) learn to generate
3
|
[
[
0.1732177734375,
0.0328369140625,
0.48046875,
0.032196044921875,
0.48046875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11932373046875,
0.1748046875,
0.11932373046875
],
[
0.1732177734375,
0.11932373046875,
0.8056640625,
0.11932373046875,
0.8056640625,
0.1331787109375,
0.1732177734375,
0.1331787109375
],
[
0.17236328125,
0.1395263671875,
0.34716796875,
0.14013671875,
0.34716796875,
0.1578369140625,
0.17236328125,
0.1572265625
],
[
0.1732177734375,
0.1666259765625,
0.3544921875,
0.1666259765625,
0.3544921875,
0.179931640625,
0.1732177734375,
0.179931640625
],
[
0.1748046875,
0.1875,
0.82177734375,
0.1875,
0.82177734375,
0.2008056640625,
0.1748046875,
0.2008056640625
],
[
0.1748046875,
0.2020263671875,
0.82421875,
0.2020263671875,
0.82421875,
0.21533203125,
0.1748046875,
0.21533203125
],
[
0.1748046875,
0.21533203125,
0.82275390625,
0.21533203125,
0.82275390625,
0.2291259765625,
0.1748046875,
0.2291259765625
],
[
0.1732177734375,
0.2298583984375,
0.82421875,
0.2291259765625,
0.82421875,
0.242431640625,
0.1732177734375,
0.24365234375
],
[
0.1748046875,
0.24365234375,
0.82421875,
0.24365234375,
0.82421875,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.1748046875,
0.2568359375,
0.82275390625,
0.2568359375,
0.82275390625,
0.270751953125,
0.1748046875,
0.270751953125
],
[
0.1748046875,
0.26953125,
0.82275390625,
0.26953125,
0.82275390625,
0.28271484375,
0.1748046875,
0.28271484375
],
[
0.1748046875,
0.2841796875,
0.82421875,
0.2841796875,
0.82421875,
0.298583984375,
0.1748046875,
0.298583984375
],
[
0.1732177734375,
0.29736328125,
0.82177734375,
0.29736328125,
0.82177734375,
0.311279296875,
0.1732177734375,
0.311279296875
],
[
0.1732177734375,
0.3125,
0.82421875,
0.311279296875,
0.82421875,
0.324462890625,
0.1732177734375,
0.32568359375
],
[
0.1732177734375,
0.32568359375,
0.826171875,
0.32568359375,
0.826171875,
0.339111328125,
0.1732177734375,
0.339111328125
],
[
0.1748046875,
0.34033203125,
0.82421875,
0.34033203125,
0.82421875,
0.354248046875,
0.1748046875,
0.354248046875
],
[
0.1748046875,
0.354248046875,
0.82421875,
0.354248046875,
0.82421875,
0.367431640625,
0.1748046875,
0.367431640625
],
[
0.1748046875,
0.367431640625,
0.465576171875,
0.367431640625,
0.465576171875,
0.380615234375,
0.1748046875,
0.380615234375
],
[
0.1748046875,
0.386474609375,
0.486083984375,
0.386474609375,
0.486083984375,
0.399658203125,
0.1748046875,
0.399658203125
],
[
0.1732177734375,
0.4072265625,
0.82275390625,
0.4072265625,
0.82275390625,
0.42041015625,
0.1732177734375,
0.42041015625
],
[
0.1732177734375,
0.42041015625,
0.82275390625,
0.42041015625,
0.82275390625,
0.434326171875,
0.1732177734375,
0.434326171875
],
[
0.1748046875,
0.43505859375,
0.82421875,
0.43505859375,
0.82421875,
0.448974609375,
0.1748046875,
0.448974609375
],
[
0.1748046875,
0.448974609375,
0.82421875,
0.448974609375,
0.82421875,
0.462158203125,
0.1748046875,
0.462158203125
],
[
0.1748046875,
0.462158203125,
0.82421875,
0.462158203125,
0.82421875,
0.47607421875,
0.1748046875,
0.47607421875
],
[
0.1748046875,
0.47607421875,
0.82275390625,
0.47607421875,
0.82275390625,
0.4892578125,
0.1748046875,
0.4892578125
],
[
0.1748046875,
0.490478515625,
0.82275390625,
0.490478515625,
0.82275390625,
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,
0.82177734375,
0.5166015625,
0.82177734375,
0.53076171875,
0.1732177734375,
0.5322265625
],
[
0.1748046875,
0.5322265625,
0.82421875,
0.5322265625,
0.82421875,
0.54541015625,
0.1748046875,
0.54541015625
],
[
0.1732177734375,
0.54443359375,
0.72216796875,
0.54541015625,
0.72216796875,
0.56005859375,
0.1732177734375,
0.5595703125
],
[
0.1732177734375,
0.56494140625,
0.82275390625,
0.56640625,
0.82275390625,
0.58056640625,
0.1732177734375,
0.580078125
],
[
0.1748046875,
0.580078125,
0.82421875,
0.580078125,
0.82421875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.1732177734375,
0.59228515625,
0.82275390625,
0.5947265625,
0.82275390625,
0.609375,
0.1732177734375,
0.60693359375
],
[
0.1748046875,
0.609375,
0.82275390625,
0.609375,
0.82275390625,
0.62255859375,
0.1748046875,
0.62255859375
],
[
0.1732177734375,
0.62060546875,
0.82275390625,
0.6220703125,
0.82275390625,
0.63623046875,
0.1732177734375,
0.63525390625
],
[
0.1748046875,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.1748046875,
0.6494140625,
0.82421875,
0.6494140625,
0.82421875,
0.66357421875,
0.1748046875,
0.66357421875
],
[
0.1732177734375,
0.66357421875,
0.82275390625,
0.662109375,
0.82275390625,
0.6767578125,
0.1732177734375,
0.67822265625
],
[
0.1748046875,
0.67822265625,
0.82421875,
0.67822265625,
0.82421875,
0.69140625,
0.1748046875,
0.69140625
],
[
0.1732177734375,
0.69140625,
0.82275390625,
0.69140625,
0.82275390625,
0.705078125,
0.1732177734375,
0.705078125
],
[
0.17236328125,
0.705078125,
0.826171875,
0.705078125,
0.826171875,
0.7197265625,
0.17236328125,
0.7197265625
],
[
0.1748046875,
0.7197265625,
0.82177734375,
0.7197265625,
0.82177734375,
0.73291015625,
0.1748046875,
0.73291015625
],
[
0.1748046875,
0.73291015625,
0.82275390625,
0.73291015625,
0.82275390625,
0.7470703125,
0.1748046875,
0.7470703125
],
[
0.1732177734375,
0.74560546875,
0.826171875,
0.7470703125,
0.826171875,
0.76123046875,
0.1732177734375,
0.76025390625
],
[
0.1748046875,
0.76123046875,
0.82177734375,
0.76123046875,
0.82177734375,
0.7744140625,
0.1748046875,
0.7744140625
],
[
0.1732177734375,
0.7734375,
0.82177734375,
0.7734375,
0.82177734375,
0.787109375,
0.1732177734375,
0.787109375
],
[
0.1732177734375,
0.787109375,
0.82275390625,
0.787109375,
0.82275390625,
0.80029296875,
0.1732177734375,
0.80029296875
],
[
0.1732177734375,
0.80029296875,
0.5556640625,
0.8017578125,
0.5556640625,
0.81640625,
0.1732177734375,
0.81494140625
],
[
0.176513671875,
0.8212890625,
0.56884765625,
0.8212890625,
0.56884765625,
0.83544921875,
0.176513671875,
0.83544921875
],
[
0.1748046875,
0.84130859375,
0.82421875,
0.84130859375,
0.82421875,
0.85498046875,
0.1748046875,
0.85498046875
],
[
0.1748046875,
0.85498046875,
0.82421875,
0.85498046875,
0.82421875,
0.869140625,
0.1748046875,
0.869140625
],
[
0.1748046875,
0.869140625,
0.82421875,
0.869140625,
0.82421875,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.17236328125,
0.88330078125,
0.82421875,
0.8818359375,
0.82421875,
0.896484375,
0.17236328125,
0.89794921875
],
[
0.1748046875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1732177734375,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.1732177734375,
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 2024",
"like cartoons, sketches, ink paintings, and oil paintings for generating consistent views and recon-",
"structing 3D shapes, which verifies the effectiveness of SyncDreamer in lifting 2D images to 3D",
"2RELATEDWORK",
"2.1DIFFUSIONMODELS",
"Diffusion models (Ho et al., 2020; Rombach et al., 2022: Croitoru et al., 2023) have shown impres-",
"sive results on 2D image generation. Concurrent work MVDiffusion (Tang et al., 2023b) also adopts",
"the multiview diffusion formulation to synthesize textures or panoramas with known geometry. We",
"propose similarformulations in SyncDreamer but with unknown geometry.MultiDiffusion (Bar-Tal",
"et al., 2023) and SyncDiffusion (Lee et al., 2023) correlate multiple diffusion models for different",
"regions of a 2D image. Many recent works (Nichol et al., 2022; Jun & Nichol, 2023; Miller et al.",
"2023; Zhang et al., 2023a; Liu et al., 2023d: Wang et al., 2023b; Gupta et al., 2023: Cheng et al.",
"2023; Karnewar et al., 2023b; Anciukevicius et al., 2023; Zeng et al., 2022; Erkog et al., 2023; Chen",
"et al., 2023a; Kim et al., 2023; Ntavelis et al., 2023; Gu et al., 2023a; Karnewar et al., 2023a) try",
"to repeat the success of diffusion models on the 3D generation task. However, the scarcity of 3D",
"data makes it difficult to directly train diffusion models on 3D and the resulting generation quality is",
"still much worse and less generalizable than the counterpart image generation models, though some",
"works (Anciukevicius et al., 2023; Chen et al., 2023a; Karnewar et al., 2023b) are trying to only use",
"2D images for training 3D diffusion models.",
"2.2USING2DDIFFUSIONMODELSFOR3D",
"Instead of directly learning a 3D diffusion model, many works resort to using high-quality 2D dif-",
"fusion models (Rombach et al., 2022; Saharia et al., 2022) for 3D tasks. Pioneer works DreamFu-",
"sion (Poole et al., 2023) and SJC (Wang et al., 2023a) propose to distill a 2D text-to-image generation",
"model to generate 3D shapes from texts. Follow-up works (Chen et al., 2023b; Wang et al., 2023d;",
"Seo et al., 2023a; Yu et al., 2023a; Lin et al., 2023; Seo et al., 2023b; Tsalicoglou et al., 2023; Zhu",
"& Zhuang, 2023; Huang et al., 2023; Armandpour et al., 2023; Wu et al., 2023; Chen et al., 2023c)",
"improve such text-to-3D distillation methods in various aspects. Many works (Tang et al., 2023a;",
"Melas-Kyriazi et al., 2023; Qian et al., 2023; Xu et al., 2022; Raj et al., 2023; Shen et al., 2023)",
"also apply such a distillation pipeline in the single-view reconstruction task. Though some impres-",
"sive results are achieved, these methods usually require a long time for textual inversion (Liu et al.,",
"2023a) and NeRF optimization and they do not guarantee to get satisfactory results.",
"Other works (Watson et al., 2022; Gu et al., 2023b; Deng et al., 2023a; Zhou & Tulsiani, 2023; Tseng",
"et al., 2023; Chan et al., 2023; Yu et al., 2023b; Tewari et al., 2023; Yoo et al., 2023; Szymanowicz",
"et al., 2023; Tang et al., 2023b; Xiang et al., 2023; Liu et al., 2023c; Lei et al., 2022) directly apply",
"the 2D diffusion models to generate multiview images for 3D reconstruction. (Tseng et al., 2023; Yu",
"et al., 2023b) are conditioned on the input image by attention layers for novel-view synthesis in in-",
"door scenes. Our method also uses attention layers but is intended for object reconstruction. (Xiang",
"et al., 2023; Zhang et al., 2023b) resort to estimated depth maps to warp and inpaint for novel-view",
"image generation, which strongly relies on the performance of the external single-view depth esti-",
"mator. Two concurrent works (Chan et al., 2023; Tewari et al., 2023) generate new images in an",
"autoregressive render-and-generate manner, which demonstrates good performances on specific ob-",
"ject categories or scenes. In comparison, SyncDreamer is targeted to reconstruct arbitrary objects",
"and generates all images in one reverse process. The concurrent work Viewset Diffusion (Szy-",
"manowicz et al., 2023) shares a similar idea to generate a set of images: The differences between",
"SyncDreamer and Viewset Diffusion are that SyncDreamer does not require predicting a radiance",
"field like Viewset Diffusion but only uses attention to synchronize the states among views and Sync-",
"Dreamer fixes the viewpoints of generated views for better convergence.Another concurrent work",
"MVDream (Shi et al., 2023) also proposes multiview generation for the text-to-3D task while our",
"work aims to reconstruct shapes from single-view images.",
"2.3OTHER SINGLE-VIEWRECONSTRUCTION METHODS",
"Single-view reconstruction is a challenging ill-posed problem. Before the prosperity of generative",
"models used in 3D reconstruction, there are many works (Tatarchenko et al., 2019; Fu et al., 2021;",
"Kato & Harada, 2019; Li et al., 2020; Fahim et al., 2021) that reconstruct 3D shapes from single-view",
"images by regression (Li et al., 2020) or retrieval (Tatarchenko et al., 2019), which have difficulty in",
"generalizing to new categories. Recent NeRF-GAN methods (Niemeyer & Geiger, 2021; Chan et al.,",
"2022; Gu et al., 2021; Schwarz et al., 2020; Gao et al., 2022; Deng et al., 2023b) learn to generate",
"3"
] |
{
"arxiv_id": "2309.03453",
"language": "en",
"page_idx": 2
}
|
||
2309.03453-en-3
|
2309.03453-en-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
NeRFs for specific categories like human or cat faces. These NeRF-GANs achieve impressive results on single-view image reconstruction but fail to generalize to arbitrary objects. Although some recent works also attempt to generalize NeRF-GAN to ImageNet (Skorokhodov et al., 2023; Sargent et al., 2023), training NeRF-GANs for arbitrary objects is still challenging.
## 3 METHOD
Given an input view y of an object, our target is to generate multiview images of the object. We assume that the object is located at the origin and is normalized inside a cube of length l. The target images are generated on Nfxed viewpoints looking at the object with azimuths evenly ranging from $0^\circ$ to $360^\circ$ and elevations of $30^\circ.$ To improve the multiview consistency of generated images, we formulate this generation process as a multiview diffusion model. In the following, we begin with a review of diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020).
### 3.1 DIFFUSION
Diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020) aim to learn a probability model $p_{\theta }( \mathbf{x} _{0})$ $= \int p_{\theta }( \mathbf{x} _{0: T}) d\mathbf{x} _{1: T}$ where $\mathbf{x}_0$ is the data and $\mathbf{x}_{1:T}:=\mathbf{x}_{1},...,\mathbf{x}_{T}$ are latent variables. The joint distribution is characterized by a Markov Chain (reverse process)
(1)
$$p_\theta(\mathbf{x}_{0:T})=p(\mathbf{x}_T)\prod_{t=1}^Tp_\theta(\mathbf{x}_{t-1}|\mathbf{x}_t),$$
$$\begin{array}{l}{t=1}\\{{\mathrm{nd}\:p_{\theta}(\mathbf{x}_{t-1}|\mathbf{x}_{t})\:=\:\mathcal{N}(\mathbf{x}_{t-1};\mu_{\theta}(\mathbf{x}_{t},t),}}\end{array}$$
target is to learn the $\mu_\theta$ for the generation. To learn $\mu_\theta$, a Markov chain called forward process is
constructed as
$$q(\mathbf{x}_{1:T}|\mathbf{x}_0)=\prod\limits_{t=1}^Tq(\mathbf{x}_t|\mathbf{x}_{t-1}),$$
(2)
where $q( \mathbf{x} _t| \mathbf{x} _{t- 1}) = \mathcal{N} ( \mathbf{x} _t; \sqrt {1- \beta _t}\mathbf{x} _{t- 1}, \beta _t\mathbf{I} )$ and $\beta _t$ are all constants. DDPM (Ho et al. , 2020)
shows that by defining
$$\mu_\theta(\mathbf{x}_t,t)=\frac{1}{\sqrt{\alpha}_t}\left(\mathbf{x}_t-\frac{\beta_t}{\sqrt{1-\bar{\alpha}_t}}\epsilon_\theta(\mathbf{x}_t,t)\right),$$
(3)
where $\alpha_t$ and $\bar{\alpha}_t$ are constants derived from $\beta_t$ and $\epsilon_\theta$ is a noise predictor, we can learn $\epsilon_\theta$ by
$$\ell=\mathbb{E}_{t,\mathbf{x}_{0},\epsilon}\left[\|\epsilon-\epsilon_{\theta}(\sqrt{\bar{\alpha}_{t}}\mathbf{x}_{0}+\sqrt{1-\bar{\alpha}_{t}}\epsilon,t)\|_{2}\right],$$
where $\epsilon$ is a random variable sampled from $\mathcal{N}(\mathbf{0},\mathbf{I}).$
(4)
## 3.2 MULTIVIEW DIFFUSION
Applying the vanilla DDPM model to generate novel-view images separately would lead to difficulty in maintaining multiview consistency across different views. To address this problem, we formulate the generation process as a multiview diffusion model that correlates the generation of each view. Let us denote the $N$ images that we want to generate on the predefined viewpoints as $\{\mathbf{x}_0^{(1)},...,\mathbf{x}_0^{(N)}\}$ where suffix 0 means the time step 0. We want to learn the $joint\textit{distribution of all these views}$ $p_{\theta }( \mathbf{x} _{0}^{( 1: N) }| \mathbf{y} ) : = p_{\theta }( \mathbf{x} _{0}^{( 1) }, . . . , \mathbf{x} _{0}^{( N) }| \mathbf{y} ) .$ In the following discussion, all the probability functions are conditioned on the input view y so we omit y for simplicity.
The forward process of the multiview diffusion model is a direct extension of the vanilla DDPM in
Eq. 2, where noises are added to every view independently by
$$q(\mathbf{x}_{1:T}^{(1:N)}|\mathbf{x}_{0}^{(1:N)})=\prod_{t=1}^Tq(\mathbf{x}_t^{(1:N)}|\mathbf{x}_{t-1}^{(1:N)})=\prod_{t=1}^T\prod_{n=1}^Nq(\mathbf{x}_t^{(n)}|\mathbf{x}_{t-1}^{(n)}),$$
(5)
where $q(\mathbf{x}_t^{(n)}|\mathbf{x}_{t-1}^{(n)})=\mathcal{N}(\mathbf{x}_t^{(n)};\sqrt{1-\beta_t}\mathbf{x}_{t-1}^{(n)},\beta_t\mathbf{I}).$ Similarly, following Eq. 1, the reverse process
is constructed as
$$p_\theta(\mathbf{x}_{0:T}^{(1:N)})=p(\mathbf{x}_T^{(1:N)})\prod_{t=1}^Tp_\theta(\mathbf{x}_{t-1}^{(1:N)}|\mathbf{x}_t^{(1:N)})=p(\mathbf{x}_T^{(1:N)})\prod_{t=1}^T\prod_{n=1}^Np_\theta(\mathbf{x}_{t-1}^{(n)}|\mathbf{x}_t^{(1:N)}),$$
(6)
4
|
[
[
0.1748046875,
0.0328369140625,
0.4794921875,
0.0328369140625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11871337890625,
0.1748046875,
0.11871337890625
],
[
0.176513671875,
0.11932373046875,
0.82275390625,
0.11932373046875,
0.82275390625,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.176513671875,
0.1331787109375,
0.82275390625,
0.1331787109375,
0.82275390625,
0.146484375,
0.176513671875,
0.146484375
],
[
0.1732177734375,
0.1458740234375,
0.625,
0.146484375,
0.625,
0.1610107421875,
0.1732177734375,
0.160400390625
],
[
0.1732177734375,
0.1673583984375,
0.281982421875,
0.1673583984375,
0.281982421875,
0.1817626953125,
0.1732177734375,
0.1817626953125
],
[
0.1732177734375,
0.1925048828125,
0.82177734375,
0.1925048828125,
0.82177734375,
0.205810546875,
0.1732177734375,
0.205810546875
],
[
0.1748046875,
0.20703125,
0.8203125,
0.20703125,
0.8203125,
0.2203369140625,
0.1748046875,
0.2203369140625
],
[
0.1748046875,
0.2215576171875,
0.82177734375,
0.2215576171875,
0.82177734375,
0.23486328125,
0.1748046875,
0.23486328125
],
[
0.1748046875,
0.23486328125,
0.82275390625,
0.23486328125,
0.82275390625,
0.248779296875,
0.1748046875,
0.248779296875
],
[
0.1748046875,
0.248779296875,
0.82421875,
0.248779296875,
0.82421875,
0.261962890625,
0.1748046875,
0.261962890625
],
[
0.1748046875,
0.26318359375,
0.64892578125,
0.26318359375,
0.64892578125,
0.2734375,
0.1748046875,
0.2734375
],
[
0.1732177734375,
0.29345703125,
0.293212890625,
0.29345703125,
0.293212890625,
0.306884765625,
0.1732177734375,
0.306884765625
],
[
0.1748046875,
0.318115234375,
0.82275390625,
0.318115234375,
0.82275390625,
0.33203125,
0.1748046875,
0.33203125
],
[
0.238525390625,
0.333984375,
0.57177734375,
0.333984375,
0.57177734375,
0.347900390625,
0.238525390625,
0.347900390625
],
[
0.625,
0.333984375,
0.82275390625,
0.333984375,
0.82275390625,
0.347900390625,
0.625,
0.347900390625
],
[
0.5888671875,
0.33837890625,
0.61279296875,
0.33837890625,
0.61279296875,
0.34521484375,
0.5888671875,
0.34521484375
],
[
0.1748046875,
0.3359375,
0.240234375,
0.333251953125,
0.2418212890625,
0.3447265625,
0.176513671875,
0.347900390625
],
[
0.1732177734375,
0.34521484375,
0.63818359375,
0.34521484375,
0.63818359375,
0.359130859375,
0.1732177734375,
0.359130859375
],
[
0.3798828125,
0.37744140625,
0.61767578125,
0.37744140625,
0.61767578125,
0.395751953125,
0.3798828125,
0.395751953125
],
[
0.80126953125,
0.37744140625,
0.82861328125,
0.37744140625,
0.82861328125,
0.39453125,
0.80126953125,
0.39453125
],
[
0.176513671875,
0.412353515625,
0.82275390625,
0.412353515625,
0.82275390625,
0.4267578125,
0.176513671875,
0.4267578125
],
[
0.176513671875,
0.4267578125,
0.82275390625,
0.4267578125,
0.82275390625,
0.43994140625,
0.176513671875,
0.43994140625
],
[
0.1748046875,
0.43994140625,
0.82421875,
0.43994140625,
0.82421875,
0.453857421875,
0.1748046875,
0.453857421875
],
[
0.176513671875,
0.455810546875,
0.26953125,
0.455810546875,
0.26953125,
0.466552734375,
0.176513671875,
0.466552734375
],
[
0.50244140625,
0.4697265625,
0.5107421875,
0.4697265625,
0.5107421875,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.400390625,
0.4765625,
0.5986328125,
0.4765625,
0.5986328125,
0.494384765625,
0.400390625,
0.494384765625
],
[
0.80126953125,
0.4765625,
0.82666015625,
0.4765625,
0.82666015625,
0.494384765625,
0.80126953125,
0.494384765625
],
[
0.1748046875,
0.5087890625,
0.82177734375,
0.5087890625,
0.82177734375,
0.5234375,
0.1748046875,
0.5234375
],
[
0.1732177734375,
0.5234375,
0.32275390625,
0.5234375,
0.32275390625,
0.537109375,
0.1732177734375,
0.537109375
],
[
0.53759765625,
0.54443359375,
0.55126953125,
0.54443359375,
0.55126953125,
0.55615234375,
0.53759765625,
0.55615234375
],
[
0.57275390625,
0.55029296875,
0.62744140625,
0.548828125,
0.62841796875,
0.564453125,
0.57421875,
0.56640625
],
[
0.80126953125,
0.5498046875,
0.82666015625,
0.5498046875,
0.82666015625,
0.5673828125,
0.80126953125,
0.5673828125
],
[
0.51318359375,
0.56103515625,
0.57177734375,
0.56103515625,
0.57177734375,
0.57275390625,
0.51318359375,
0.57275390625
],
[
0.3505859375,
0.55322265625,
0.414306640625,
0.55322265625,
0.414306640625,
0.564453125,
0.3505859375,
0.564453125
],
[
0.43701171875,
0.56298828125,
0.463134765625,
0.56298828125,
0.463134765625,
0.5712890625,
0.43701171875,
0.5712890625
],
[
0.1748046875,
0.5771484375,
0.78173828125,
0.5791015625,
0.78173828125,
0.59326171875,
0.1748046875,
0.59130859375
],
[
0.339111328125,
0.59765625,
0.65625,
0.59765625,
0.65625,
0.6123046875,
0.339111328125,
0.6123046875
],
[
0.80224609375,
0.5986328125,
0.826171875,
0.5986328125,
0.826171875,
0.61279296875,
0.80224609375,
0.61279296875
],
[
0.1748046875,
0.619140625,
0.51318359375,
0.619140625,
0.51318359375,
0.63330078125,
0.1748046875,
0.63330078125
],
[
0.1732177734375,
0.642578125,
0.37744140625,
0.64306640625,
0.37744140625,
0.65771484375,
0.1732177734375,
0.6572265625
],
[
0.176513671875,
0.662109375,
0.82177734375,
0.662109375,
0.82177734375,
0.6767578125,
0.176513671875,
0.6767578125
],
[
0.1732177734375,
0.67822265625,
0.82177734375,
0.6767578125,
0.82177734375,
0.68994140625,
0.1732177734375,
0.69140625
],
[
0.1748046875,
0.69140625,
0.82177734375,
0.69140625,
0.82177734375,
0.705078125,
0.1748046875,
0.705078125
],
[
0.1732177734375,
0.70849609375,
0.75244140625,
0.70556640625,
0.75244140625,
0.72021484375,
0.1732177734375,
0.72265625
],
[
0.75341796875,
0.708984375,
0.8203125,
0.705078125,
0.82177734375,
0.72021484375,
0.7548828125,
0.724609375
],
[
0.1732177734375,
0.7216796875,
0.82177734375,
0.7216796875,
0.82177734375,
0.73486328125,
0.1732177734375,
0.73486328125
],
[
0.1690673828125,
0.73291015625,
0.82861328125,
0.73193359375,
0.82861328125,
0.755859375,
0.1690673828125,
0.7568359375
],
[
0.1732177734375,
0.751953125,
0.57275390625,
0.75244140625,
0.57275390625,
0.767578125,
0.1732177734375,
0.7666015625
],
[
0.1748046875,
0.77294921875,
0.82275390625,
0.7734375,
0.82275390625,
0.78857421875,
0.1748046875,
0.787109375
],
[
0.1748046875,
0.78857421875,
0.580078125,
0.78857421875,
0.580078125,
0.8017578125,
0.1748046875,
0.8017578125
],
[
0.303955078125,
0.8125,
0.369384765625,
0.818359375,
0.365234375,
0.83837890625,
0.301513671875,
0.83203125
],
[
0.361083984375,
0.8193359375,
0.3857421875,
0.8193359375,
0.3857421875,
0.8271484375,
0.361083984375,
0.8271484375
],
[
0.375732421875,
0.814453125,
0.7158203125,
0.814453125,
0.7158203125,
0.83837890625,
0.375732421875,
0.83837890625
],
[
0.28515625,
0.8212890625,
0.330078125,
0.8212890625,
0.330078125,
0.8359375,
0.28515625,
0.8359375
],
[
0.80224609375,
0.8193359375,
0.82666015625,
0.8193359375,
0.82666015625,
0.8369140625,
0.80224609375,
0.8369140625
],
[
0.5087890625,
0.82763671875,
0.52685546875,
0.82763671875,
0.52685546875,
0.83056640625,
0.5087890625,
0.83056640625
],
[
0.57275390625,
0.8388671875,
0.61376953125,
0.8388671875,
0.61376953125,
0.8466796875,
0.57275390625,
0.8466796875
],
[
0.1732177734375,
0.85302734375,
0.82421875,
0.85498046875,
0.82421875,
0.87255859375,
0.1732177734375,
0.87060546875
],
[
0.1748046875,
0.87060546875,
0.285888671875,
0.87060546875,
0.285888671875,
0.8837890625,
0.1748046875,
0.8837890625
],
[
0.221435546875,
0.89599609375,
0.775390625,
0.89453125,
0.775390625,
0.91650390625,
0.221435546875,
0.91748046875
],
[
0.8037109375,
0.9013671875,
0.82275390625,
0.9013671875,
0.82275390625,
0.9140625,
0.8037109375,
0.9140625
],
[
0.386474609375,
0.921875,
0.400390625,
0.921875,
0.400390625,
0.9248046875,
0.386474609375,
0.9248046875
],
[
0.61669921875,
0.921875,
0.65625,
0.921875,
0.65625,
0.92529296875,
0.61669921875,
0.92529296875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2024",
"NeRFs for specific categories like human or cat faces. These NeRF-GANs achieve impressive results",
"on single-view image reconstruction but fail to generalize to arbitrary objects. Although some recent",
"works also attempt to generalize NeRF-GAN to ImageNet (Skorokhodov et al., 2023; Sargent et al..",
"2023), training NeRF-GANs for arbitrary objects is still challenging.",
"3METHOD",
"Given an input view y of an object, our target is to generate multiview images of the object. We",
"assume that the object is located at the origin and is normalized inside a cube of length 1. The targei",
"images are generated on N fixed viewpoints looking at the object with azimuths evenly ranging from",
"0° to 360° and elevations of 30°. To improve the multiview consistency of generated images, we",
"formulate this generation process as a multiview diffusion model. In the following, we begin with a",
"review of diffusion models (Sohl-Dickstein et al..2015:Ho et al..2020).",
"3.1DIFFUSION",
"Diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020) aim to learn a probability model",
"J pe(xo:T)dx1:T where Xo is the data and x1:T :",
"-,XT are latent variables. The",
"X1.",
"pe(xo) =",
"joint distribution is characterized by a Markov Chain (reverse process)",
"pe(Xo:T) = p(xT) II pe(xt-1xt),",
"(1)",
"where p(xT) = N(xT;0, I) and pe(xt-1xt) = N(xt-1; μe(xt,t), o2I). μe(xt,t) is a trainable",
"component while the variance ? is untrained time-dependent constants (Ho et al., 2020). The",
"target is to learn the μg for the generation. To learn μo, a Markov chain called forward process is",
"constructedas",
"1.",
"q(x1:T|xo) = IIq(xtxt-1),",
"(2)",
"where q(xt|xt-1) = N(xt; V1-βtxt-1, βtI) and βt are all constants. DDPM (Ho et al., 2020)",
"shows that by defining",
"Bt",
"Eo(xt,t)",
"(3)",
"V1-at",
"uo(Xt,t)",
"or",
"where t and &t are constants derived from βt and eg is a noise predictor, we can learn ee by",
"l =Et,xo.e [le- Ee(Vatxo + V1- ate,t)l2] ",
"(4)",
"where e is a random variable sampled from N(o, I)",
"3.2MULTIVIEWDIFFUSION",
"Applying the vanilla DDPM model to generate novel-view images separately would lead to difficulty",
"in maintaining multiview consistency across different views. To address this problem, we formulate",
"the generation process as a multiview diffusion model that correlates the generation of each view. Let",
"us denote the N images that we want to generate on the predefined viewpoints as fx(1)",
"... x(N))",
"where suffix O means the time step O. We want to learn the joint distribution of ali these views",
"pe)y)pe(\" ()Inthefollowingdiscussiona theprobabilityfunctionsare",
"conditioned on the input view y so we omit y for simplicity.",
"Theforward process of the multiview diffusionmodel is a directextension of thevanilla DDPM in",
"Eq.2, where noises are added to every view independently by",
"N)",
"1:N",
"V) = IIg(x(:N)x(N) =I11Ig(x(\")x(),",
"q(xiT",
"(5)",
"",
"t=1n=",
"where g(x(n) |x()) = N(x(n); V1- βtx(m), βtI). Similarly, following Eq. 1, the reverse process",
"is constructed as",
"pe(x(:N) = p(x(:N) II pe(x(N |x(1:N) = p(x(:N) I pe(x() /x(1:N),",
"(6)",
"-",
"",
"4"
] |
{
"arxiv_id": "2309.03453",
"language": "en",
"page_idx": 3
}
|
||
2309.03453-en-4
|
2309.03453-en-4.jpg
|
Published as a conference paper at ICLR 2024
<FigureHere>
Figure 2: The pipeline of a synchronized multiview noise predictor to denoise the target view $\mathbf{x}_{t}^{(n)}$ for one step. First, a spatial feature volume is constructed from all the noisy target views $\mathbf{x}_{t}^{(1:N)}.$ Then, we construct a view frustum feature volume for $\mathbf{x}_t^{(n)}$ by interpolating the features of spatial feature volume. The input view y, current target view $\mathbf{x}_t^{(n)}$ and viewpoint difference $\Delta\mathbf{v}^{(n)}$ are fed into the backbone UNet initialized from Zero123 (Liu et al., 2023b). On the intermediate feature maps of the UNet, new depth-wise attention layers are applied to extract features from the view frustum feature volume. Finally, the output of the UNet is used to denoise $\mathbf{x}_t^{(n)}$ to obtain $\mathbf{x}_t-1^{(n)}.$
where $p_{\theta}(\mathbf{x}_{t-1}^{(n)}|\mathbf{x}_{t}^{(1:N)})=\mathcal{N}(\mathbf{x}_{t-1}^{(n)};\mu_{\theta}^{(n)}(\mathbf{x}_{t}^{(1:N)},t),\sigma_{t}^{2}\mathbf{I}).$ Note that the second equation in Eq. holds because we assume a diagonal variance matrix. However, the mean $\mu_{\theta}^{(n)}$ of $n$-th view $\mathbf{x}_{t-1}^{(n)}$ depends on the states of all the views $\mathbf{x}_t^{(1:N)}.$ Similar to Eq. 3, we define $\mu_\theta^{(n)}$ and the loss by
(7)
$$\mu_\theta^{(n)}(\mathbf{x}_t^{(1:N)},t)=\frac{1}{\sqrt{\alpha}_t}\left(\mathbf{x}_t^{(n)}-\frac{\beta_t}{\sqrt{1-\bar{\alpha}_t}}\epsilon_\theta^{(n)}(\mathbf{x}_t^{(1:N)},t)\right).$$
$$\ell=\mathbb{E}_{t,\mathbf{x}_0^{(1:N)},n,\epsilon^{(1:N)}}\left[\|\epsilon^{(n)}-\epsilon_\theta^{(n)}(\mathbf{x}_t^{(1:N)},t)\|_2\right],$$
(8)
where $\epsilon^{(1:N)}$ is the standard Gaussian noise of size $N\times H\times W$ added to all $N$ views, $\epsilon^{(n)}$ is the
noise added to the $n$-th view, and $\epsilon_\theta^{(n)}$ is the noise predictor on the $n$-th view.
Training procedure. In one training step, we first obtain $N$ images $\mathbf{x}_0^{(1:N)}$ of the same object from the dataset. Then, we sample a timestep $t$ and the noise $\epsilon ^{( 1: N) }$ which is added to all the images $\mathbf{x} _{0}^{( 1: N) }$to obtain $\mathbf{x} _{t}^{( 1: N) }\cdot$After that, we randomly select a view $n$ and apply the corresponding noise predictor $\epsilon_\theta^{(n)}$ on the selected view to predict the noise. Finally, the L2 distance between the sampled noise $\epsilon^{(n)}$ and the predicted noise is computed as the loss for the training.
Synchronized $N\textbf{- view noise predictor. The proposed multiview diffusion model can be regarded}$ as $N$ synchronized noise predictors $\{\epsilon_\theta^{(n)}|n=1,...,N\}.$ On each time step $t$, each noise predictor $\epsilon^{(n)}$ is in charge of predicting noise on its corresponding view $\mathbf{x}_t^{(n)}$ to get $\mathbf{x}_t^{(n)}.$ Meanwhile, these noise predictors are synchronized because, on every denoising step, every noise predictor exchanges $informationwitheachotherbycorrelatingthestatesx_{t}^{(1:N)}$of all the other views. In practical implementation, we use a shared UNet for all $\tilde{N}$ noise predictors and put the viewpoint difference between the input view and the $n$-th target view $\Delta\mathbf{v}^{(n)}$, and the states $\mathbf{x}_t^{(1:N)}$ of all views as conditions to this shared noise predictor, i.e., $\epsilon_{\theta}^{(n)}(\mathbf{x}_{t}^{(1:N)},t)=\epsilon_{\theta}(\mathbf{x}_{t}^{(n)};t,\Delta\mathbf{v}^{(n)},\mathbf{x}_{t}^{(1:N)}).$ The detailed computation of the viewpoint difference can be found in the supplementary material.
3.3 3D-AWARE FEATURE ATTENTION FOR DENOISING
$\begin{array}{lll}\mathrm{In}&\mathrm{this}&\mathrm{section,~we}&\mathrm{discuss}&\mathrm{how}&\mathrm{to}&\mathrm{implement}&\mathrm{the}&\mathrm{synchronized}&\mathrm{noise}&\mathrm{predictor}\\\epsilon_{\theta}(\mathbf{x}^{(n)};t,\Delta\mathbf{v}^{(n)},\mathbf{x}_{t}^{(1:N)},\mathbf{y})&\mathrm{by}&\mathrm{correlating}&\mathrm{the}&\mathrm{multiview}&\mathrm{features}&\mathrm{using}&\mathrm{3D-aware}&\mathrm{attention}\\\text{scheme The overvi}\end{array}$
scheme. The overview is shown in Fig. 2.
5
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.38232421875,
0.10479736328125,
0.6396484375,
0.10418701171875,
0.6396484375,
0.117431640625,
0.38232421875,
0.11871337890625
],
[
0.5537109375,
0.1275634765625,
0.6259765625,
0.126953125,
0.6259765625,
0.1383056640625,
0.5537109375,
0.1395263671875
],
[
0.6640625,
0.1256103515625,
0.7646484375,
0.1256103515625,
0.7646484375,
0.1363525390625,
0.6640625,
0.1363525390625
],
[
0.373291015625,
0.1351318359375,
0.420654296875,
0.1351318359375,
0.420654296875,
0.1458740234375,
0.373291015625,
0.1458740234375
],
[
0.560546875,
0.1395263671875,
0.619140625,
0.1395263671875,
0.619140625,
0.15087890625,
0.560546875,
0.15087890625
],
[
0.685546875,
0.136962890625,
0.74658203125,
0.136962890625,
0.74658203125,
0.147705078125,
0.685546875,
0.147705078125
],
[
0.47802734375,
0.17041015625,
0.53369140625,
0.17041015625,
0.53369140625,
0.1849365234375,
0.47802734375,
0.1849365234375
],
[
0.253173828125,
0.212158203125,
0.31201171875,
0.212158203125,
0.31201171875,
0.222900390625,
0.253173828125,
0.222900390625
],
[
0.324462890625,
0.2108154296875,
0.338134765625,
0.2108154296875,
0.338134765625,
0.2139892578125,
0.324462890625,
0.2139892578125
],
[
0.406005859375,
0.2291259765625,
0.450927734375,
0.2291259765625,
0.450927734375,
0.2392578125,
0.406005859375,
0.2392578125
],
[
0.4453125,
0.2310791015625,
0.46728515625,
0.2291259765625,
0.468017578125,
0.236083984375,
0.44677734375,
0.238037109375
],
[
0.268798828125,
0.285888671875,
0.281982421875,
0.285888671875,
0.281982421875,
0.2890625,
0.268798828125,
0.2890625
],
[
0.57275390625,
0.2998046875,
0.60302734375,
0.2998046875,
0.60302734375,
0.3125,
0.57275390625,
0.3125
],
[
0.318603515625,
0.31005859375,
0.3798828125,
0.31005859375,
0.3798828125,
0.3212890625,
0.318603515625,
0.3212890625
],
[
0.41259765625,
0.3125,
0.48779296875,
0.309326171875,
0.48779296875,
0.32080078125,
0.41259765625,
0.323974609375
],
[
0.69384765625,
0.3125,
0.77392578125,
0.31005859375,
0.77392578125,
0.324462890625,
0.69384765625,
0.326904296875
],
[
0.1748046875,
0.34228515625,
0.80615234375,
0.34228515625,
0.80615234375,
0.35595703125,
0.1748046875,
0.35595703125
],
[
0.80615234375,
0.343505859375,
0.81787109375,
0.343505859375,
0.81787109375,
0.347900390625,
0.80615234375,
0.347900390625
],
[
0.1748046875,
0.35986328125,
0.79150390625,
0.35986328125,
0.79150390625,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.78515625,
0.356689453125,
0.81640625,
0.354736328125,
0.81787109375,
0.3662109375,
0.7861328125,
0.36865234375
],
[
0.17236328125,
0.37255859375,
0.826171875,
0.37255859375,
0.826171875,
0.393310546875,
0.17236328125,
0.393310546875
],
[
0.1707763671875,
0.39013671875,
0.82666015625,
0.389404296875,
0.82666015625,
0.410400390625,
0.1707763671875,
0.410888671875
],
[
0.1748046875,
0.40771484375,
0.82275390625,
0.40771484375,
0.82275390625,
0.421630859375,
0.1748046875,
0.421630859375
],
[
0.1748046875,
0.42236328125,
0.82421875,
0.42236328125,
0.82421875,
0.4375,
0.1748046875,
0.4375
],
[
0.1690673828125,
0.43310546875,
0.7900390625,
0.43310546875,
0.7900390625,
0.458251953125,
0.1690673828125,
0.458251953125
],
[
0.1690673828125,
0.458984375,
0.47216796875,
0.460205078125,
0.47216796875,
0.48486328125,
0.1690673828125,
0.484130859375
],
[
0.441162109375,
0.460205078125,
0.82666015625,
0.464111328125,
0.82666015625,
0.48486328125,
0.441162109375,
0.48095703125
],
[
0.17236328125,
0.484130859375,
0.8203125,
0.481689453125,
0.8203125,
0.498779296875,
0.17236328125,
0.50048828125
],
[
0.4248046875,
0.49755859375,
0.67724609375,
0.50244140625,
0.67724609375,
0.5234375,
0.4248046875,
0.51806640625
],
[
0.1732177734375,
0.5048828125,
0.438720703125,
0.5048828125,
0.438720703125,
0.51806640625,
0.1732177734375,
0.51806640625
],
[
0.6708984375,
0.50244140625,
0.78173828125,
0.50390625,
0.78173828125,
0.51806640625,
0.6708984375,
0.517578125
],
[
0.29736328125,
0.52880859375,
0.3857421875,
0.5244140625,
0.386474609375,
0.54736328125,
0.2998046875,
0.5517578125
],
[
0.48046875,
0.52978515625,
0.5048828125,
0.52978515625,
0.5048828125,
0.53857421875,
0.48046875,
0.53857421875
],
[
0.5439453125,
0.5263671875,
0.5595703125,
0.5263671875,
0.5595703125,
0.5361328125,
0.5439453125,
0.5361328125
],
[
0.59326171875,
0.53271484375,
0.60302734375,
0.53271484375,
0.60302734375,
0.537109375,
0.59326171875,
0.537109375
],
[
0.62841796875,
0.53076171875,
0.64794921875,
0.53076171875,
0.64794921875,
0.53857421875,
0.62841796875,
0.53857421875
],
[
0.80224609375,
0.53271484375,
0.826171875,
0.53271484375,
0.826171875,
0.54736328125,
0.80224609375,
0.54736328125
],
[
0.51708984375,
0.54052734375,
0.5849609375,
0.54248046875,
0.5849609375,
0.556640625,
0.51708984375,
0.5546875
],
[
0.381591796875,
0.5341796875,
0.41259765625,
0.5341796875,
0.41259765625,
0.54541015625,
0.381591796875,
0.54541015625
],
[
0.423095703125,
0.54248046875,
0.4697265625,
0.54052734375,
0.471435546875,
0.5546875,
0.4248046875,
0.556640625
],
[
0.3349609375,
0.56298828125,
0.47802734375,
0.5693359375,
0.476318359375,
0.5869140625,
0.334228515625,
0.58056640625
],
[
0.484375,
0.564453125,
0.53369140625,
0.564453125,
0.53369140625,
0.57568359375,
0.484375,
0.57568359375
],
[
0.57861328125,
0.56298828125,
0.609375,
0.56298828125,
0.609375,
0.5712890625,
0.57861328125,
0.5712890625
],
[
0.80224609375,
0.56494140625,
0.826171875,
0.56494140625,
0.826171875,
0.580078125,
0.80224609375,
0.580078125
],
[
0.60888671875,
0.56640625,
0.65625,
0.56640625,
0.65625,
0.58056640625,
0.60888671875,
0.58056640625
],
[
0.5498046875,
0.56640625,
0.5595703125,
0.56640625,
0.5595703125,
0.57080078125,
0.5498046875,
0.57080078125
],
[
0.5361328125,
0.57373046875,
0.55126953125,
0.57373046875,
0.55126953125,
0.57763671875,
0.5361328125,
0.57763671875
],
[
0.56396484375,
0.57373046875,
0.57861328125,
0.57373046875,
0.57861328125,
0.57763671875,
0.56396484375,
0.57763671875
],
[
0.1732177734375,
0.5859375,
0.826171875,
0.5859375,
0.826171875,
0.60400390625,
0.1732177734375,
0.60400390625
],
[
0.1748046875,
0.60498046875,
0.67724609375,
0.60498046875,
0.67724609375,
0.619140625,
0.1748046875,
0.619140625
],
[
0.1748046875,
0.62255859375,
0.63818359375,
0.625,
0.63818359375,
0.6396484375,
0.1748046875,
0.63720703125
],
[
0.61767578125,
0.61865234375,
0.826171875,
0.62353515625,
0.826171875,
0.6416015625,
0.61767578125,
0.63623046875
],
[
0.1732177734375,
0.63720703125,
0.82666015625,
0.63720703125,
0.82666015625,
0.65478515625,
0.1732177734375,
0.65478515625
],
[
0.1748046875,
0.65087890625,
0.82666015625,
0.65283203125,
0.82666015625,
0.673828125,
0.1748046875,
0.671875
],
[
0.17724609375,
0.6611328125,
0.193603515625,
0.6611328125,
0.193603515625,
0.66943359375,
0.17724609375,
0.66943359375
],
[
0.17236328125,
0.67236328125,
0.826171875,
0.671875,
0.826171875,
0.689453125,
0.17236328125,
0.68994140625
],
[
0.17236328125,
0.68798828125,
0.654296875,
0.689453125,
0.654296875,
0.70703125,
0.17236328125,
0.70556640625
],
[
0.1732177734375,
0.7109375,
0.82275390625,
0.7109375,
0.82275390625,
0.72607421875,
0.1732177734375,
0.72607421875
],
[
0.1707763671875,
0.7265625,
0.82421875,
0.7265625,
0.82421875,
0.744140625,
0.1707763671875,
0.744140625
],
[
0.1732177734375,
0.74560546875,
0.826171875,
0.74560546875,
0.826171875,
0.76318359375,
0.1732177734375,
0.76318359375
],
[
0.1748046875,
0.76220703125,
0.82421875,
0.76220703125,
0.82421875,
0.77587890625,
0.1748046875,
0.77587890625
],
[
0.1707763671875,
0.77294921875,
0.82421875,
0.7734375,
0.82421875,
0.79541015625,
0.1707763671875,
0.79443359375
],
[
0.1748046875,
0.79345703125,
0.82421875,
0.79345703125,
0.82421875,
0.806640625,
0.1748046875,
0.806640625
],
[
0.1707763671875,
0.80322265625,
0.82861328125,
0.80322265625,
0.82861328125,
0.8271484375,
0.1707763671875,
0.8271484375
],
[
0.1690673828125,
0.8193359375,
0.82861328125,
0.82080078125,
0.82861328125,
0.84521484375,
0.1690673828125,
0.8447265625
],
[
0.1748046875,
0.84228515625,
0.646484375,
0.84228515625,
0.646484375,
0.85595703125,
0.1748046875,
0.85595703125
],
[
0.1748046875,
0.85986328125,
0.5595703125,
0.85986328125,
0.5595703125,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.1732177734375,
0.87890625,
0.82421875,
0.87890625,
0.82421875,
0.89306640625,
0.1732177734375,
0.89306640625
],
[
0.17236328125,
0.892578125,
0.31787109375,
0.892578125,
0.31787109375,
0.91357421875,
0.17236328125,
0.91357421875
],
[
0.2998046875,
0.892578125,
0.339111328125,
0.892578125,
0.339111328125,
0.90283203125,
0.2998046875,
0.90283203125
],
[
0.330810546875,
0.896484375,
0.82421875,
0.896484375,
0.82421875,
0.9111328125,
0.330810546875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.44775390625,
0.9111328125,
0.44775390625,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9609375,
0.491943359375,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2024",
"Synchronized Multiview Noise Predictor",
"Spatial volume",
"View frustum volume",
"Unproject",
"(F,V.V,V)",
"(E,D,H,W)",
"3DCNN",
"Target views",
"1-N",
"CLIP",
"+A",
"rtr",
"UNet",
"Input view y",
"Target view x(n",
"Target view x(n)",
"Figure 2: The pipeline of a synchronized multiview noise predictor to denoise the target view x",
"72",
"for one step. First, a spatial feature volume is constructed from all the noisy target views x",
"(1:N)",
"Then, we construct a view frustum feature volume for x(n) by interpolating the features of spatial",
"feature volume. The input view y, current target view x(n) and viewpoint difference Av(n) are fed",
"into the backbone UNet initialized from Zero123 (Liu et al., 2023b). On the intermediate feature",
"maps of the UNet, new depth-wise attention layers are applied to extract features from the view",
"rusmaumFinathuutthtiued",
"where po(x(|x(:N) = N(x();μ\"(x",
"(x(1:N),t), o?I). Note that the second equation in Eq. 6",
"holds because we assume a diagonal variance matrix. However, the mean μ(n) of n-th view x(n)",
"x(1:)Smilaro3, wedfine",
"depends on the states of all the views xt",
"2)and the loss by",
"μ\"(x(1:N),",
"(n)",
"Bt",
"72",
"1:N",
"(7)",
"V1-at",
",t)=",
"Va(",
"l =Et.x(N),n,e(1:N)",
"Ile(n)",
"(1:N)",
"(8)",
",t)[2",
"H",
"Cn",
"",
"where e(1:N) is the standard Gaussian noise of size N × H × W added to all N views, e(n) is the",
"noise added to the n-th view, and egn) is the noise predictor on the n-th view.",
"Training procedure. In one training step, we first obtain N images xo",
"x(1:N) of the same object from",
"the dataset. Then, we sample a timestep t and the noise e(1:N) which is added to all the images",
"x(1:N) to obtain x(1:N). After that, we randomly select a view n and apply the corresponding noise",
"Xo",
"predictor en) on the selected view to predict the noise. Finally, the L2 distance between the sampled",
"noise e(n) and the predicted noise is computed as the loss for the training.",
"Synchronized N-view noise predictor. The proposed multiview diffusion model can be regarded",
"as N synchronized noise predictors (e(n'|n = 1,., N). On each time step t, each noise predictor",
"e(n) is in charge of predicting noise on its corresponding view x(n) to get x(n). Meanwhile, these",
"noise predictors are synchronized because, on every denoising step,every noise predictor exchanges",
"information with each other by correlating thestatesx(1N)ofall theother views. In practicalimple-",
"mentation, we use a shared UNet for all N noise predictors and put the viewpoint difference between",
"the input view and then-thtarget viewv(n),and thestatesx(i)ofall viewsasconditions tothis",
"shared noise predicto, ie. (x(),t) = o(x();t,(m),), he detailed computation",
"of the viewpoint difference can be found in the supplementary material.",
"3.33D-AWAREFEATUREATTENTIONFORDENOISING",
"In thissection,wediscuss how to implementthe synchronized noisepredictor",
"Eo(x(n);t, Av(n), x(",
"(1:N)",
"',y) by correlating the multiview features using a 3D-aware attention",
"scheme. The overview is shown in Fig. 2.",
"5"
] |
{
"arxiv_id": "2309.03453",
"language": "en",
"page_idx": 4
}
|
||
2401.08038-en-0
|
2401.08038-en-0.jpg
|
## Calpric: Inclusive and Fine-grained Labeling of Privacy Policies with $\textbf{Crowdsourcing and Active Learning}$
Wenjun Qiu $\begin{array}{cc}\text{David Lie}&\text{Lisa Austin}\\&University\textit{ of Toronto}\end{array}$
## $\mathbf{Abstract}$
A significant challenge to training accurate deep learning models on privacy policies is the cost and difficulty of obtaining a large and comprehensive set of training data. To address these challenges, we present Calpric, which combines automatic text selection and segmentation, active learning and the use of crowdsourced annotators to generate a large, balanced training set for Android privacy policies at low cost. Automated text selection and segmentation simpliffes the labeling task, enabling untrained annotators from crowdsourcing platforms, like Amazon's Mechanical Turk, to be competitive with trained annotators, such as law students, and also reduces inter-annotator agreement, which decreases labeling cost. Having reliable labels for training enables the use of active learning, which uses fewer training samples to efficiently cover the input space, further reducing cost and improving class and data category balance in the data set.
The combination of these techniques allows Calpric to produce models that are accurate over a wider range of data categories, and provide more detailed, fine-grained labels than previous work. Our crowdsourcing process enables Calpric to attain reliable labeled data at a cost of roughly $0.92-$1.71 per labeled text segment. Calpric's training process also generates a labeled data set of 16K privacy policy text segments across 9 data categories with balanced assertions and denials
## 1 Introduction
Privacy policies are legal documents that disclose how a party collects, uses, and shares personally identifiable information (PII). Privacy legislation, such as the California Online Privacy Protection Act (CalOPPA) and the General Data Protection Regulation (GDPR), require digital services to use privacy policies to obtain consent for the collection and use of PII. In addition to consent, privacy policies also serve as a mechanism for transparency and accountability to ensure that digital services comply with privacy legislation [26]. With the massive growth of privacy policies, there has been significant
Data Category: Contact
We|will not share your email address or phone number with any outside parties
Action Mode: Denial
Data Action: 3rd party sharing
Figure 1: A fully labeled privacy policy segment with data category (contact information), data action (third party sharing), and its corresponding action mode (does not share).
interest in building tools to automate the analysis of privacy policies for compliance [49,50], as well as to help end users understand them [19].
However, previous attempts have been limited in inclusiveness and granularity$^1.$ For example, [49] and [50] are limited to only the most common categories of private data—contacts, device identifiers and location—but do not include less common data categories such as private health, financial and demographic information. Similarly, [19] trains hierarchical models that classify text segments by various privacy-related properties, such as the type of data action (i.e. collection, storage) and the data category (location, health). However, their models cannot distinguish fine-grained attributes, such as an explicit denial of a data action from the lack of a disclosure of a data action. In addition, they also cannot guarantee uniform performance on rare combinations (i.e.“We do not store your health information", since explicit denials and storage are both rarer classes). PolicyLint [2] and PoliCheck[3] use a rules-based ontology to organize information types as a hierarchy, but also do not explicitly address minority data categories. These shortcomings exist because 1) the training set in previous approaches is not sufficiently large to cover less common cases, and 2) the imbalanced distribution of cases in privacy policies means that there will be insufficient training samples to provide good performance at fine-granularity.
In this paper, we address both challenges with a new training method for privacy policy classifiers, which we call
lWe use inclusiveness to denote accurate classification for minority data
categories and fine/coarse to denote the number of possible labels a model
can predict at its output.
|
[
[
0.1568603515625,
0.154052734375,
0.84228515625,
0.154052734375,
0.84228515625,
0.1685791015625,
0.1568603515625,
0.1685791015625
],
[
0.314453125,
0.1749267578125,
0.6845703125,
0.1767578125,
0.6845703125,
0.1956787109375,
0.314453125,
0.1932373046875
],
[
0.306396484375,
0.2279052734375,
0.406005859375,
0.2298583984375,
0.406005859375,
0.2474365234375,
0.306396484375,
0.24560546875
],
[
0.4599609375,
0.2291259765625,
0.5439453125,
0.2291259765625,
0.5439453125,
0.246826171875,
0.4599609375,
0.246826171875
],
[
0.59619140625,
0.2291259765625,
0.69384765625,
0.2291259765625,
0.69384765625,
0.246826171875,
0.59619140625,
0.246826171875
],
[
0.414306640625,
0.24560546875,
0.5849609375,
0.246826171875,
0.5849609375,
0.264404296875,
0.414306640625,
0.26318359375
],
[
0.0228729248046875,
0.26953125,
0.060455322265625,
0.26953125,
0.060455322265625,
0.3984375,
0.0228729248046875,
0.3984375
],
[
0.2484130859375,
0.318115234375,
0.32275390625,
0.318115234375,
0.32275390625,
0.333251953125,
0.2484130859375,
0.333251953125
],
[
0.65185546875,
0.32080078125,
0.76953125,
0.32080078125,
0.76953125,
0.333984375,
0.65185546875,
0.333984375
],
[
0.0849609375,
0.341552734375,
0.48046875,
0.34228515625,
0.48046875,
0.356689453125,
0.0849609375,
0.35595703125
],
[
0.52392578125,
0.337158203125,
0.91015625,
0.33837890625,
0.91015625,
0.352783203125,
0.52392578125,
0.3515625
],
[
0.08660888671875,
0.359130859375,
0.483642578125,
0.359130859375,
0.483642578125,
0.37255859375,
0.08660888671875,
0.37255859375
],
[
0.53369140625,
0.3662109375,
0.6337890625,
0.3662109375,
0.6337890625,
0.376953125,
0.53369140625,
0.376953125
],
[
0.67724609375,
0.3662109375,
0.830078125,
0.365478515625,
0.830078125,
0.376953125,
0.67724609375,
0.37744140625
],
[
0.08660888671875,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.386474609375,
0.08660888671875,
0.386474609375
],
[
0.08660888671875,
0.38818359375,
0.484375,
0.38818359375,
0.484375,
0.401611328125,
0.08660888671875,
0.401611328125
],
[
0.51708984375,
0.386474609375,
0.9111328125,
0.386962890625,
0.9111328125,
0.40087890625,
0.51708984375,
0.399658203125
],
[
0.0882568359375,
0.40478515625,
0.4794921875,
0.40478515625,
0.4794921875,
0.415283203125,
0.0882568359375,
0.415283203125
],
[
0.515625,
0.399658203125,
0.9140625,
0.40087890625,
0.9140625,
0.415283203125,
0.515625,
0.4140625
],
[
0.02532958984375,
0.4091796875,
0.058013916015625,
0.410400390625,
0.0555419921875,
0.45068359375,
0.0228729248046875,
0.449462890625
],
[
0.08660888671875,
0.41845703125,
0.48046875,
0.41845703125,
0.48046875,
0.431884765625,
0.08660888671875,
0.431884765625
],
[
0.51708984375,
0.415283203125,
0.7509765625,
0.415283203125,
0.7509765625,
0.4287109375,
0.51708984375,
0.4287109375
],
[
0.0849609375,
0.434326171875,
0.48193359375,
0.434326171875,
0.48193359375,
0.447509765625,
0.0849609375,
0.447509765625
],
[
0.0318603515625,
0.44580078125,
0.05963134765625,
0.4462890625,
0.056365966796875,
0.481689453125,
0.0294189453125,
0.48095703125
],
[
0.0849609375,
0.448974609375,
0.48046875,
0.448974609375,
0.48046875,
0.462158203125,
0.0849609375,
0.462158203125
],
[
0.515625,
0.455078125,
0.9111328125,
0.455810546875,
0.9111328125,
0.47021484375,
0.515625,
0.4697265625
],
[
0.08660888671875,
0.464111328125,
0.483642578125,
0.464111328125,
0.483642578125,
0.477783203125,
0.08660888671875,
0.477783203125
],
[
0.51708984375,
0.4716796875,
0.91259765625,
0.4716796875,
0.91259765625,
0.486083984375,
0.51708984375,
0.486083984375
],
[
0.08660888671875,
0.479248046875,
0.48193359375,
0.479248046875,
0.48193359375,
0.492431640625,
0.08660888671875,
0.492431640625
],
[
0.51708984375,
0.4873046875,
0.6640625,
0.4873046875,
0.6640625,
0.50048828125,
0.51708984375,
0.50048828125
],
[
0.0236968994140625,
0.50048828125,
0.058013916015625,
0.50048828125,
0.058013916015625,
0.5888671875,
0.0236968994140625,
0.5888671875
],
[
0.0882568359375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.505859375,
0.0882568359375,
0.505859375
],
[
0.53369140625,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.53369140625,
0.51513671875
],
[
0.0849609375,
0.50830078125,
0.48046875,
0.5087890625,
0.48046875,
0.5234375,
0.0849609375,
0.52294921875
],
[
0.51708984375,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.51708984375,
0.53076171875
],
[
0.08660888671875,
0.5244140625,
0.483642578125,
0.5244140625,
0.483642578125,
0.53857421875,
0.08660888671875,
0.53857421875
],
[
0.51708984375,
0.53271484375,
0.9140625,
0.53271484375,
0.9140625,
0.546875,
0.51708984375,
0.546875
],
[
0.0849609375,
0.54052734375,
0.48046875,
0.54052734375,
0.48046875,
0.5537109375,
0.0849609375,
0.5537109375
],
[
0.51708984375,
0.54736328125,
0.9140625,
0.54736328125,
0.9140625,
0.56103515625,
0.51708984375,
0.56103515625
],
[
0.08660888671875,
0.5546875,
0.48193359375,
0.5546875,
0.48193359375,
0.568359375,
0.08660888671875,
0.568359375
],
[
0.51708984375,
0.56298828125,
0.9140625,
0.56298828125,
0.9140625,
0.5771484375,
0.51708984375,
0.5771484375
],
[
0.0882568359375,
0.57080078125,
0.3896484375,
0.57080078125,
0.3896484375,
0.58056640625,
0.0882568359375,
0.58056640625
],
[
0.03594970703125,
0.58349609375,
0.051483154296875,
0.58349609375,
0.051483154296875,
0.59228515625,
0.03594970703125,
0.59228515625
],
[
0.51708984375,
0.57763671875,
0.91259765625,
0.57763671875,
0.91259765625,
0.59130859375,
0.51708984375,
0.59130859375
],
[
0.0196075439453125,
0.59033203125,
0.058013916015625,
0.59033203125,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.10296630859375,
0.58544921875,
0.48046875,
0.58544921875,
0.48046875,
0.5986328125,
0.10296630859375,
0.5986328125
],
[
0.041656494140625,
0.59228515625,
0.0482177734375,
0.59228515625,
0.0482177734375,
0.59521484375,
0.041656494140625,
0.59521484375
],
[
0.51708984375,
0.59228515625,
0.91259765625,
0.59228515625,
0.91259765625,
0.60595703125,
0.51708984375,
0.60595703125
],
[
0.08660888671875,
0.60107421875,
0.48046875,
0.60107421875,
0.48046875,
0.6142578125,
0.08660888671875,
0.6142578125
],
[
0.51708984375,
0.609375,
0.9140625,
0.609375,
0.9140625,
0.62255859375,
0.51708984375,
0.62255859375
],
[
0.0849609375,
0.61572265625,
0.48193359375,
0.61572265625,
0.48193359375,
0.62890625,
0.0849609375,
0.62890625
],
[
0.51708984375,
0.62353515625,
0.9140625,
0.62353515625,
0.9140625,
0.63720703125,
0.51708984375,
0.63720703125
],
[
0.08660888671875,
0.630859375,
0.48193359375,
0.630859375,
0.48193359375,
0.64453125,
0.08660888671875,
0.64453125
],
[
0.51708984375,
0.63818359375,
0.9140625,
0.63818359375,
0.9140625,
0.6513671875,
0.51708984375,
0.6513671875
],
[
0.0849609375,
0.64599609375,
0.48046875,
0.64599609375,
0.48046875,
0.6591796875,
0.0849609375,
0.6591796875
],
[
0.51708984375,
0.65283203125,
0.91259765625,
0.65283203125,
0.91259765625,
0.66650390625,
0.51708984375,
0.66650390625
],
[
0.08660888671875,
0.662109375,
0.4794921875,
0.662109375,
0.4794921875,
0.67236328125,
0.08660888671875,
0.67236328125
],
[
0.51708984375,
0.66845703125,
0.91259765625,
0.66845703125,
0.91259765625,
0.681640625,
0.51708984375,
0.681640625
],
[
0.08416748046875,
0.6748046875,
0.48193359375,
0.67529296875,
0.48193359375,
0.68994140625,
0.08416748046875,
0.689453125
],
[
0.0849609375,
0.68994140625,
0.441162109375,
0.68994140625,
0.441162109375,
0.7041015625,
0.0849609375,
0.7041015625
],
[
0.51708984375,
0.6845703125,
0.91259765625,
0.6845703125,
0.91259765625,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.51708984375,
0.69873046875,
0.9140625,
0.69873046875,
0.9140625,
0.7119140625,
0.51708984375,
0.7119140625
],
[
0.515625,
0.71337890625,
0.91259765625,
0.7119140625,
0.91259765625,
0.7265625,
0.515625,
0.72802734375
],
[
0.0849609375,
0.73095703125,
0.223876953125,
0.73095703125,
0.223876953125,
0.74560546875,
0.0849609375,
0.74560546875
],
[
0.515625,
0.72802734375,
0.91259765625,
0.72900390625,
0.91259765625,
0.74365234375,
0.515625,
0.7421875
],
[
0.51708984375,
0.744140625,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.08660888671875,
0.76220703125,
0.48193359375,
0.76220703125,
0.48193359375,
0.7763671875,
0.08660888671875,
0.7763671875
],
[
0.51708984375,
0.7587890625,
0.91259765625,
0.7587890625,
0.91259765625,
0.7734375,
0.51708984375,
0.7734375
],
[
0.0849609375,
0.77783203125,
0.48193359375,
0.77783203125,
0.48193359375,
0.79150390625,
0.0849609375,
0.79150390625
],
[
0.51708984375,
0.7744140625,
0.91259765625,
0.7744140625,
0.91259765625,
0.78857421875,
0.51708984375,
0.78857421875
],
[
0.0849609375,
0.79248046875,
0.483642578125,
0.79248046875,
0.483642578125,
0.80615234375,
0.0849609375,
0.80615234375
],
[
0.51708984375,
0.79052734375,
0.9140625,
0.79052734375,
0.9140625,
0.8037109375,
0.51708984375,
0.8037109375
],
[
0.0849609375,
0.806640625,
0.48193359375,
0.806640625,
0.48193359375,
0.82080078125,
0.0849609375,
0.82080078125
],
[
0.51708984375,
0.8046875,
0.9140625,
0.8046875,
0.9140625,
0.818359375,
0.51708984375,
0.818359375
],
[
0.0849609375,
0.82275390625,
0.48193359375,
0.82275390625,
0.48193359375,
0.8359375,
0.0849609375,
0.8359375
],
[
0.51708984375,
0.8193359375,
0.88818359375,
0.8193359375,
0.88818359375,
0.83349609375,
0.51708984375,
0.83349609375
],
[
0.0849609375,
0.8369140625,
0.483642578125,
0.8359375,
0.483642578125,
0.85107421875,
0.0849609375,
0.8515625
],
[
0.53369140625,
0.83544921875,
0.9111328125,
0.83544921875,
0.9111328125,
0.8486328125,
0.53369140625,
0.8486328125
],
[
0.0849609375,
0.85302734375,
0.48046875,
0.85302734375,
0.48046875,
0.8662109375,
0.0849609375,
0.8662109375
],
[
0.515625,
0.85107421875,
0.91259765625,
0.85107421875,
0.91259765625,
0.8642578125,
0.515625,
0.8642578125
],
[
0.08660888671875,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.53369140625,
0.87255859375,
0.9111328125,
0.875,
0.9111328125,
0.8896484375,
0.53369140625,
0.88720703125
],
[
0.0849609375,
0.88330078125,
0.4794921875,
0.88330078125,
0.4794921875,
0.896484375,
0.0849609375,
0.896484375
],
[
0.51708984375,
0.88818359375,
0.91259765625,
0.88818359375,
0.91259765625,
0.8984375,
0.51708984375,
0.8984375
],
[
0.0849609375,
0.8984375,
0.48193359375,
0.8984375,
0.48193359375,
0.912109375,
0.0849609375,
0.912109375
],
[
0.51708984375,
0.90087890625,
0.64453125,
0.90087890625,
0.64453125,
0.9111328125,
0.51708984375,
0.9111328125
]
] |
[
"Calpric: Inclusive and Fine-grained Labeling of Privacy Policies with",
"Crowdsourcing and Active Learning",
"Wenjun Qiu",
"David Lie",
"Lisa Austin",
"University of Toronto",
"",
"Abstract",
"Data Category:Contact",
"A significant challenge to training accurate deep learning",
"Wewillnotshareyouremailaddressorphonenumberwithanyoutsideparties",
"models on privacy policies is the cost and difficulty of obtain-",
"ActionMode:Denial",
"Data Action:3rd party sharing",
"ing a large and comprehensive set of training data. To address",
"these challenges, we present Calpric, which combines auto-",
"Figure I:Afully labeled privacy policy segment with data category",
"matictext selectionand segmentation,activelearning andthe",
"(contact information), data action (third party sharing), and its corre-",
"[T",
"use of crowdsourced annotators to generate a large, balanced",
"sponding action mode (does not share).",
"training set for Android privacy policies at low cost. Auto-",
"S",
"mated text selection and segmentation simplifies the labeling",
"interest in building tools to automate the analysis of privacy",
"task, enabling untrained annotators from crowdsourcing plat-",
"policies for compliance [49, 50], as well as to help end users",
"forms, like Amazon's Mechanical Turk, to be competitive",
"understand them [19].",
"1180080",
"with trained annotators.such aslaw students,and also re-",
"However,previous attempts have been limited in inclusive-",
"duces inter-annotator agreement, which decreases labeling",
"ness and granularity'.For example, [49] and [50] are limited",
"cost. Having reliable labels for training enables the use of ac-",
"to only the most common categories of private data-contacts,",
"tive learning, which uses fewer training samples to efficiently",
"device identifiers and location-but do not include less com-",
"cover the input space, further reducing cost and improving",
"mon data categories such as private health, financial and de-",
"class and data category balance in the data set.",
"",
"mographic information. Similarly,[19] trains hierarchical",
"10:0",
"The combination of these techniques allows Calpric to",
"",
"models that classify text segments by various privacy-related",
"produce models that are accurate over a wider range of data",
"properties, such as the type of data action (i.e. collection,",
"categories, and provide more detailed, fine-grained labels than",
"storage) and the data category (location, health). However,",
"previous work. Our crowdsourcing process enables Calpric to",
"their models cannot distinguish fine-grained attributes, such",
" attain reliable labeled data at a cost of roughly $0.92-$1.71 per",
"as an explicit denial of a data action from the lack of a disclo-",
"labeledtextsegment.Calpric'straining processalso generates",
"sure of a data action. In addition, they also cannot guarantee",
"a labeled data set of 16K privacy policy text segments across",
"9 data categories with balanced assertions and denials.",
"uniform performance on rare combinations (i.e.\"We do not",
"store your health information\", since explicit denials and stor-",
"age are both rarer classes). PolicyLint [2] and PoliCheck [3]",
"1 Introduction",
"use a rules-based ontology to organize information types as a",
"hierarchy, but also do not explicitly address minority data cat-",
"Privacy policies are legal documents that disclose how a party",
"egories. These shortcomings exist because 1) the training set",
"collects, uses, and shares personally identifiable information",
"in previous approaches is not sufficiently large to cover less",
"(PIl). Privacy legislation, such as the California Online Pri-",
"common cases, and 2) the imbalanced distribution of cases in",
"vacy Protection Act (CalOPPA) and the General Data Pro",
"privacy policies means that there will be insufficient training",
"tection Regulation (GDPR), require digital services to use",
"samples to provide good performance at fine-granularity.",
"privacy policies to obtain consent for the collection and use",
"In this paper, we address both challenges with a new",
"of PII. In addition to consent, privacy policies also serve as a",
"training method for privacy policy classifiers, which we call",
"mechanism for transparency and accountability to ensure that",
"Iwe use inclusiveness to denote accurate classification for minority data",
"digital services comply with privacy legislation [26]. With the",
"categories and fine/coarsetodenote the number of possiblelabelsamodel",
"massive growth of privacy policies, there has been significant",
"can predict at its output."
] |
{
"arxiv_id": "2401.08038",
"language": "en",
"page_idx": 0
}
|
||
2401.08038-en-1
|
2401.08038-en-1.jpg
|
Calpric (Crowdsourcing Active Learning PRIvacy Policy Classifer). Calpric's training approach simultaneously gathers data and trains on it, synergistically using crowdsourced annotators and active learning to produce a larger and distributionally balanced dataset. Calpric overcomes a number of challenges that arise from crowdsourcing, such as untrained annotators and unreliable annotations, through a data categoryspecific text segmentation algorithm. The construction of previous data sets, such as OPP-115 and APP-350[42,49], required the annotators to both select the relevant text segments in the privacy policy and label them, a complex annotation task for crowdsourced workers. Calpric automatically extracts segments comparable to the ones selected by trained annotators by determining whether a segment is relevant to the label being sought, identifying the correct segment boundary and capturing the necessary local context. However, simply acquiring more training samples would not guarantee that sufficient samples of rare categories are found. To overcome distributional imbalance, Calpric uses active learning, which uses a partially trained model to select unlabeled samples that are likely to contain information that the model currently lacks. The use of active learning itself is not without pitfalls— because active learning strategically selects samples for which it currently has little information, there tends to be little redundancy in the training samples it selects, and so an incorrect label provided by an annotator can have a large impact on model accuracy. Unfortunately, crowdsourced annotators do not provide correct labels consistently. Calpric's segmentation algorithm overcomes this challenge by increasing interannotator agreement and annotation accuracy, thus providing benefits to both the crowdsourcing and active learning components.
Building on these innovations, we use Calpric to build the Calpric Privacy Policy Corpus (CPPS). Training on CPPS, Calpric produces models that can provide accuracy over a greater breadth of data categories and with finer-grained labels than previous work. Calpric achieves an overall classification accuracy of 93.0% on the sentence/segment level, and 92.2% on the document level, which exceeds that of prior models trained on smaller data sets generated by trained annotators [19,42,49]. Simultaneously, Calpric's training methods enable the use of crowdsourced annotators labeled text segment for $0.92-$1.71 at an annotation quality similar to previous studies that used used more expensive trained annotators [49,50]. This enables Calpric to generate a training set set of 16856 labeled text segments, which is 3× larger than the largest previous human-annotated data set and contains significantly more information on minority data categories and classes.
In summary, this paper makes the following contributions: 1. We present Calpric, which synergistically combines
crowdsourcing and active learning to produce Calpric Privacy Policy Corpus (CPPS), which contains 16856
labeled text segments extracted from 52K Android privacy policies covering 9 data categories and 3 data actions. CPPS is balanced across data categories and action modes. As far as we are aware, CPPS is the largest corpus of privacy policy text segments to date. The Calpric project is openly available https://github.com/ dlgroupuoft/calpric.
2. While creating CPPS, Calpric trains models that provide
more inclusive and finer-grained classification performance on privacy policies. Calpric's models cover 9 data categories and can differentiate denials (i.e. explicit denial of a data action) from the absence of disclosure about a data action.
3. We perform measurements and find that Calpric's au-
tomated text segmentation increases the number of usable labels produced by 65%, enables the use of crowdsourced annotators, which reduces labeling costs by 9×, and results in a model that is more accurate than prior work, especially on minority data categories and classes.
4. We conduct a study on rare data categories and actions
in 1,800 Android app privacy policies. We find that some categories are not as rare as previously believed, and that with the exception of the most popular applications, the number of explicit denials and use of controllable data actions increases with application popularity.
## 2 $\textbf{Challenges}$
In this section we better characterize the challenges of inclusive and fine-grained privacy policy classification and the availability of training data, which covers data imbalance, labeling cost, and segmentation challenges.
2.1 $\textbf{Inclusive and Fine- grained Classification}$
While previous work has demonstrated that it is possible to train models that achieve high classification accuracy on privacy policy text, that classification has been limited in inclusiveness and granularity. We summarize the limitations of previous approaches as compared to Calpric in Table 1. Limitations in breadth mean that model predication accuracy may suffer for minority (rare) classes, as machine trained models tend to optimize for accuracy in majority (common) classes.[49] and[50] only cover 3 common categories and do not cover rare data categories such as health, financial, demographic, social media, personal identifer, or survey data.[19] covers 11$^2$ rare categories, but has low accuracy in them due to the limited access to training data in those categories.
To alleviate some of these issues [19] combines different data actions when training models to predict certain private
${^{2}\text{Polisis covers non-mobile privacy policies, we only train Calpric on}}$
categories that are applicable to mobile privacy policies.
|
[
[
0.0882568359375,
0.0921630859375,
0.48046875,
0.0921630859375,
0.48046875,
0.1060791015625,
0.0882568359375,
0.1060791015625
],
[
0.5498046875,
0.0921630859375,
0.9140625,
0.0921630859375,
0.9140625,
0.1060791015625,
0.5498046875,
0.1060791015625
],
[
0.0882568359375,
0.10797119140625,
0.483642578125,
0.10797119140625,
0.483642578125,
0.121826171875,
0.0882568359375,
0.121826171875
],
[
0.5498046875,
0.10797119140625,
0.9140625,
0.10797119140625,
0.9140625,
0.121826171875,
0.5498046875,
0.121826171875
],
[
0.0849609375,
0.12249755859375,
0.483642578125,
0.121826171875,
0.483642578125,
0.1363525390625,
0.0849609375,
0.136962890625
],
[
0.54833984375,
0.12249755859375,
0.9140625,
0.12249755859375,
0.9140625,
0.1363525390625,
0.54833984375,
0.1363525390625
],
[
0.08660888671875,
0.136962890625,
0.48046875,
0.136962890625,
0.48046875,
0.15087890625,
0.08660888671875,
0.15087890625
],
[
0.54833984375,
0.136962890625,
0.9140625,
0.136962890625,
0.9140625,
0.15087890625,
0.54833984375,
0.15087890625
],
[
0.08660888671875,
0.15283203125,
0.48193359375,
0.15283203125,
0.48193359375,
0.1666259765625,
0.08660888671875,
0.1666259765625
],
[
0.5498046875,
0.15283203125,
0.9140625,
0.15283203125,
0.9140625,
0.1666259765625,
0.5498046875,
0.1666259765625
],
[
0.0882568359375,
0.169189453125,
0.48046875,
0.169189453125,
0.48046875,
0.179931640625,
0.0882568359375,
0.179931640625
],
[
0.5498046875,
0.1685791015625,
0.9111328125,
0.1685791015625,
0.9111328125,
0.1817626953125,
0.5498046875,
0.1817626953125
],
[
0.0849609375,
0.18310546875,
0.48193359375,
0.18310546875,
0.48193359375,
0.1962890625,
0.0849609375,
0.1962890625
],
[
0.5498046875,
0.18310546875,
0.7138671875,
0.18310546875,
0.7138671875,
0.1962890625,
0.5498046875,
0.1962890625
],
[
0.0849609375,
0.1988525390625,
0.48193359375,
0.1988525390625,
0.48193359375,
0.212158203125,
0.0849609375,
0.212158203125
],
[
0.52783203125,
0.2052001953125,
0.91259765625,
0.20703125,
0.91259765625,
0.2215576171875,
0.52783203125,
0.2197265625
],
[
0.08660888671875,
0.21337890625,
0.483642578125,
0.21337890625,
0.483642578125,
0.2266845703125,
0.08660888671875,
0.2266845703125
],
[
0.54833984375,
0.222900390625,
0.9140625,
0.222900390625,
0.9140625,
0.236083984375,
0.54833984375,
0.236083984375
],
[
0.08660888671875,
0.2291259765625,
0.48193359375,
0.2291259765625,
0.48193359375,
0.242431640625,
0.08660888671875,
0.242431640625
],
[
0.08660888671875,
0.24365234375,
0.48193359375,
0.24365234375,
0.48193359375,
0.2568359375,
0.08660888671875,
0.2568359375
],
[
0.54833984375,
0.238037109375,
0.9140625,
0.238037109375,
0.9140625,
0.251953125,
0.54833984375,
0.251953125
],
[
0.5498046875,
0.25244140625,
0.91259765625,
0.25244140625,
0.91259765625,
0.266357421875,
0.5498046875,
0.266357421875
],
[
0.08660888671875,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.270751953125,
0.08660888671875,
0.270751953125
],
[
0.0849609375,
0.2734375,
0.483642578125,
0.2734375,
0.483642578125,
0.287353515625,
0.0849609375,
0.287353515625
],
[
0.5498046875,
0.268310546875,
0.91259765625,
0.268310546875,
0.91259765625,
0.2822265625,
0.5498046875,
0.2822265625
],
[
0.08660888671875,
0.2890625,
0.483642578125,
0.2890625,
0.483642578125,
0.302978515625,
0.08660888671875,
0.302978515625
],
[
0.5498046875,
0.285400390625,
0.68115234375,
0.285400390625,
0.68115234375,
0.29541015625,
0.5498046875,
0.29541015625
],
[
0.08660888671875,
0.3037109375,
0.48193359375,
0.3037109375,
0.48193359375,
0.317626953125,
0.08660888671875,
0.317626953125
],
[
0.52783203125,
0.304931640625,
0.91259765625,
0.306884765625,
0.91259765625,
0.3212890625,
0.52783203125,
0.3193359375
],
[
0.08660888671875,
0.3193359375,
0.48046875,
0.3193359375,
0.48046875,
0.333251953125,
0.08660888671875,
0.333251953125
],
[
0.5498046875,
0.322509765625,
0.9140625,
0.322509765625,
0.9140625,
0.3359375,
0.5498046875,
0.3359375
],
[
0.08660888671875,
0.333984375,
0.48046875,
0.333984375,
0.48046875,
0.347900390625,
0.08660888671875,
0.347900390625
],
[
0.54833984375,
0.337158203125,
0.9140625,
0.337158203125,
0.9140625,
0.35107421875,
0.54833984375,
0.35107421875
],
[
0.0882568359375,
0.35107421875,
0.4794921875,
0.35107421875,
0.4794921875,
0.361083984375,
0.0882568359375,
0.361083984375
],
[
0.54833984375,
0.352783203125,
0.91259765625,
0.3515625,
0.91259765625,
0.365478515625,
0.54833984375,
0.3662109375
],
[
0.08660888671875,
0.3642578125,
0.48193359375,
0.3642578125,
0.48193359375,
0.37744140625,
0.08660888671875,
0.37744140625
],
[
0.55126953125,
0.369384765625,
0.9111328125,
0.369384765625,
0.9111328125,
0.380126953125,
0.55126953125,
0.380126953125
],
[
0.0849609375,
0.37744140625,
0.48193359375,
0.37890625,
0.48193359375,
0.393310546875,
0.0849609375,
0.392578125
],
[
0.54833984375,
0.3818359375,
0.9140625,
0.3818359375,
0.9140625,
0.395751953125,
0.54833984375,
0.395751953125
],
[
0.08660888671875,
0.39453125,
0.48046875,
0.39453125,
0.48046875,
0.40771484375,
0.08660888671875,
0.40771484375
],
[
0.0882568359375,
0.410888671875,
0.471435546875,
0.410888671875,
0.471435546875,
0.421630859375,
0.0882568359375,
0.421630859375
],
[
0.53125,
0.4072265625,
0.91259765625,
0.4072265625,
0.91259765625,
0.42041015625,
0.53125,
0.42041015625
],
[
0.08660888671875,
0.4248046875,
0.48046875,
0.4248046875,
0.48046875,
0.438232421875,
0.08660888671875,
0.438232421875
],
[
0.54833984375,
0.42236328125,
0.91259765625,
0.42236328125,
0.91259765625,
0.4375,
0.54833984375,
0.4375
],
[
0.08660888671875,
0.43994140625,
0.48193359375,
0.43994140625,
0.48193359375,
0.453857421875,
0.08660888671875,
0.453857421875
],
[
0.54833984375,
0.438232421875,
0.9140625,
0.438232421875,
0.9140625,
0.451904296875,
0.54833984375,
0.451904296875
],
[
0.08660888671875,
0.455078125,
0.48193359375,
0.455078125,
0.48193359375,
0.4697265625,
0.08660888671875,
0.4697265625
],
[
0.5498046875,
0.455078125,
0.9111328125,
0.455078125,
0.9111328125,
0.46533203125,
0.5498046875,
0.46533203125
],
[
0.08660888671875,
0.4697265625,
0.48193359375,
0.47021484375,
0.48193359375,
0.484130859375,
0.08660888671875,
0.48291015625
],
[
0.5498046875,
0.468505859375,
0.9140625,
0.468505859375,
0.9140625,
0.481689453125,
0.5498046875,
0.481689453125
],
[
0.0882568359375,
0.48486328125,
0.48193359375,
0.48486328125,
0.48193359375,
0.498779296875,
0.0882568359375,
0.498779296875
],
[
0.54833984375,
0.48291015625,
0.8447265625,
0.48291015625,
0.8447265625,
0.496826171875,
0.54833984375,
0.496826171875
],
[
0.08660888671875,
0.50048828125,
0.483642578125,
0.50048828125,
0.483642578125,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.0849609375,
0.51513671875,
0.48193359375,
0.51513671875,
0.48193359375,
0.52880859375,
0.0849609375,
0.52880859375
],
[
0.51708984375,
0.52099609375,
0.53173828125,
0.51806640625,
0.53515625,
0.52783203125,
0.51953125,
0.53076171875
],
[
0.5458984375,
0.51806640625,
0.640625,
0.51806640625,
0.640625,
0.53271484375,
0.5458984375,
0.53271484375
],
[
0.0849609375,
0.52880859375,
0.48046875,
0.52978515625,
0.48046875,
0.54443359375,
0.0849609375,
0.54345703125
],
[
0.08660888671875,
0.54541015625,
0.483642578125,
0.54541015625,
0.483642578125,
0.5595703125,
0.08660888671875,
0.5595703125
],
[
0.51708984375,
0.54736328125,
0.9140625,
0.54736328125,
0.9140625,
0.56103515625,
0.51708984375,
0.56103515625
],
[
0.08660888671875,
0.5625,
0.143798828125,
0.5625,
0.143798828125,
0.57373046875,
0.08660888671875,
0.57373046875
],
[
0.51708984375,
0.56298828125,
0.9140625,
0.56298828125,
0.9140625,
0.57763671875,
0.51708984375,
0.57763671875
],
[
0.101318359375,
0.57568359375,
0.48046875,
0.57568359375,
0.48046875,
0.59033203125,
0.101318359375,
0.59033203125
],
[
0.5185546875,
0.580078125,
0.91259765625,
0.580078125,
0.91259765625,
0.59033203125,
0.5185546875,
0.59033203125
],
[
0.08660888671875,
0.59033203125,
0.4794921875,
0.59033203125,
0.4794921875,
0.60400390625,
0.08660888671875,
0.60400390625
],
[
0.51708984375,
0.59326171875,
0.80126953125,
0.59326171875,
0.80126953125,
0.60693359375,
0.51708984375,
0.60693359375
],
[
0.08660888671875,
0.60693359375,
0.48193359375,
0.60693359375,
0.48193359375,
0.62060546875,
0.08660888671875,
0.62060546875
],
[
0.08660888671875,
0.6220703125,
0.483642578125,
0.6220703125,
0.483642578125,
0.63525390625,
0.08660888671875,
0.63525390625
],
[
0.5185546875,
0.62744140625,
0.90283203125,
0.62744140625,
0.90283203125,
0.642578125,
0.5185546875,
0.642578125
],
[
0.0882568359375,
0.63623046875,
0.48046875,
0.63623046875,
0.48046875,
0.6494140625,
0.0882568359375,
0.6494140625
],
[
0.08660888671875,
0.6513671875,
0.48193359375,
0.6513671875,
0.48193359375,
0.66552734375,
0.08660888671875,
0.66552734375
],
[
0.51708984375,
0.6513671875,
0.9111328125,
0.6513671875,
0.9111328125,
0.66552734375,
0.51708984375,
0.66552734375
],
[
0.0849609375,
0.66552734375,
0.48046875,
0.66650390625,
0.48046875,
0.6796875,
0.0849609375,
0.6787109375
],
[
0.5185546875,
0.66943359375,
0.9111328125,
0.66943359375,
0.9111328125,
0.6796875,
0.5185546875,
0.6796875
],
[
0.0849609375,
0.681640625,
0.48046875,
0.681640625,
0.48046875,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.5185546875,
0.68310546875,
0.9140625,
0.68310546875,
0.9140625,
0.6962890625,
0.5185546875,
0.6962890625
],
[
0.0849609375,
0.6962890625,
0.483642578125,
0.6962890625,
0.483642578125,
0.71044921875,
0.0849609375,
0.71044921875
],
[
0.51953125,
0.69970703125,
0.91259765625,
0.69970703125,
0.91259765625,
0.71044921875,
0.51953125,
0.71044921875
],
[
0.0882568359375,
0.71337890625,
0.48046875,
0.71337890625,
0.48046875,
0.7236328125,
0.0882568359375,
0.7236328125
],
[
0.51708984375,
0.71337890625,
0.9140625,
0.71337890625,
0.9140625,
0.7265625,
0.51708984375,
0.7265625
],
[
0.08660888671875,
0.7265625,
0.48046875,
0.7265625,
0.48046875,
0.74072265625,
0.08660888671875,
0.74072265625
],
[
0.51953125,
0.72998046875,
0.9111328125,
0.72998046875,
0.9111328125,
0.74072265625,
0.51953125,
0.74072265625
],
[
0.0882568359375,
0.7421875,
0.48193359375,
0.7421875,
0.48193359375,
0.755859375,
0.0882568359375,
0.755859375
],
[
0.5185546875,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.08660888671875,
0.7587890625,
0.4794921875,
0.7587890625,
0.4794921875,
0.76953125,
0.08660888671875,
0.76953125
],
[
0.51708984375,
0.7587890625,
0.91259765625,
0.7587890625,
0.91259765625,
0.77294921875,
0.51708984375,
0.77294921875
],
[
0.08660888671875,
0.77294921875,
0.48193359375,
0.77294921875,
0.48193359375,
0.7861328125,
0.08660888671875,
0.7861328125
],
[
0.51708984375,
0.7734375,
0.91259765625,
0.7734375,
0.91259765625,
0.787109375,
0.51708984375,
0.787109375
],
[
0.08660888671875,
0.7890625,
0.4794921875,
0.7890625,
0.4794921875,
0.7998046875,
0.08660888671875,
0.7998046875
],
[
0.51708984375,
0.78857421875,
0.9140625,
0.78857421875,
0.9140625,
0.8017578125,
0.51708984375,
0.8017578125
],
[
0.0849609375,
0.80322265625,
0.48046875,
0.80322265625,
0.48046875,
0.81640625,
0.0849609375,
0.81640625
],
[
0.51708984375,
0.8037109375,
0.91259765625,
0.8037109375,
0.91259765625,
0.8173828125,
0.51708984375,
0.8173828125
],
[
0.0849609375,
0.818359375,
0.1641845703125,
0.818359375,
0.1641845703125,
0.830078125,
0.0849609375,
0.830078125
],
[
0.51708984375,
0.818359375,
0.91259765625,
0.818359375,
0.91259765625,
0.83203125,
0.51708984375,
0.83203125
],
[
0.0849609375,
0.8359375,
0.468017578125,
0.8369140625,
0.468017578125,
0.8515625,
0.0849609375,
0.85107421875
],
[
0.51708984375,
0.833984375,
0.88134765625,
0.833984375,
0.88134765625,
0.84765625,
0.51708984375,
0.84765625
],
[
0.5361328125,
0.85107421875,
0.9111328125,
0.85107421875,
0.9111328125,
0.861328125,
0.5361328125,
0.861328125
],
[
0.098876953125,
0.8662109375,
0.48193359375,
0.8671875,
0.48193359375,
0.8818359375,
0.098876953125,
0.880859375
],
[
0.51708984375,
0.8642578125,
0.91259765625,
0.8642578125,
0.91259765625,
0.87744140625,
0.51708984375,
0.87744140625
],
[
0.1192626953125,
0.88330078125,
0.4794921875,
0.88330078125,
0.4794921875,
0.896484375,
0.1192626953125,
0.896484375
],
[
0.53369140625,
0.88525390625,
0.91259765625,
0.88623046875,
0.91259765625,
0.8994140625,
0.53369140625,
0.8984375
],
[
0.1192626953125,
0.89794921875,
0.48193359375,
0.89794921875,
0.48193359375,
0.912109375,
0.1192626953125,
0.912109375
],
[
0.51708984375,
0.90087890625,
0.81396484375,
0.90087890625,
0.81396484375,
0.9111328125,
0.51708984375,
0.9111328125
]
] |
[
"Calpric (Crowdsourcing Active Learning PRIvacy Policy",
"labeled text segments extracted from 52K Android pri-",
"Classifier). Calpric's training approach simultaneously gath-",
"vacy policies covering 9 data categories and 3 data ac-",
"ers data and trains on it, synergistically using crowdsourced",
"tions. CPPS is balanced across data categories and ac-",
"annotators and active learning to produce a larger and distri",
"tion modes. As far as we are aware, CPPS is the largest",
"butionally balanced dataset. Calpric overcomes a number of",
"corpus of privacy policy text segments to date. The Cal-",
"challenges that arisefrom crowdsourcing,such as untrained",
"pric project is openly available https: //github.com/",
"annotators and unreliable annotations, through a data category-",
"dlgroupuoft/calpric.",
"specific text segmentation algorithm. The construction of",
"2. While creating CPPS, Calpric trains models that provide",
"previous data sets, such as OPP-115 and APP-350 [42,49], re-",
"more inclusive and finer-grained classification perfor-",
"quired the annotators to both select the relevant text segments",
"in the privacy policy and label them, a complex annotation",
"mance on privacy policies. Calpric's models cover 9",
"data categories and can differentiate denials (i.e. explicit",
"taskforcrowdsourcedworkers.Calpricautomaticallyextracts",
"segments comparable to the ones selected by trained anno-",
"denial of a data action) from the absence of disclosure",
"tators by determining whether a segment is relevant to the",
"about a data action.",
"label being sought, identifying the correct segment boundary",
"3. We perform measurements and find that Calpric's au-",
"and capturing the necessary local context. However, simply",
"tomated text segmentation increases the number of us-",
"acquiring more training samples would not guarantee that",
"able labels produced by 65%, enables the use of crowd-",
"sufficientsamples ofrarecategoriesarefound.Toovercome",
"sourced annotators, which reduces labeling costs by9x",
"distributional imbalance, Calpric uses active learning, which",
"and results in amodel that is moreaccurate thanprior",
"uses a partially trained model to select unlabeled samples",
"work, especially on minority data categories and classes.",
"that are likely to contain information that the model currently",
"lacks.Theuseof activelearningitselfisnotwithoutpitfalls-",
"4. We conduct a study on rare data categories and actions",
"because active learning strategically selects samples for which",
"in 1,800 Android app privacy policies. We find that some",
"it currently has little information, there tends to be little redun-",
" categories are not as rare as previously believed, and that",
"dancy in the training samples it selects, and so an incorrect",
"withtheexceptionofthemostpopularapplications,the",
"label provided byan annotator canhave a large impact on",
"number of explicit denials and use of controllable data",
"model accuracy. Unfortunately, crowdsourced annotators do",
"actions increases with application popularity.",
"not provide correct labels consistently. Calpric's segmenta-",
"tion algorithm overcomes this challenge by increasing inter",
"2",
"Challenges",
"annotator agreement and annotation accuracy, thus providing",
"benefits to both the crowdsourcing and active learning com-",
"In this section we better characterize the challenges of in-",
"ponents.",
"clusive and fine-grained privacy policy classification and the",
"Building on these innovations, we use Calpric to build the",
"availability of training data,which covers data imbalance.",
"Calpric Privacy Policy Corpus (CPPS). Training on CPPS",
"labeling cost, and segmentation challenges.",
"Calpric produces models that can provide accuracy over a",
"greater breadth of data categories and with finer-grained la-",
"2.1 Inclusive and Fine-grained Classification",
"bels than previous work. Calpric achieves an overall classifi",
"cation accuracy of 93.0% on the sentence/segment level, and",
"While previous work has demonstrated that it is possible to",
"92.2% on the document level, which exceeds that of prior",
"trainmodelsthatachievehighclassification accuracy onpri",
"models trained on smaller data sets generated by trained an-",
"vacy policy text, that classification has been limited in in-",
"notators [19, 42, 49]. Simultaneously, Calpric's training meth-",
"clusiveness and granularity.We summarize the limitations",
"odsenabletheuseofcrowdsourcedannotatorslabeledtext",
"of previous approaches as compared to Calpric in Table 1.",
"segment for $0.92-$1.71 at an annotation quality similar to",
"Limitationsinbreadthmeanthatmodelpredicationaccuracy",
"previous studies that used used more expensive trained anno-",
"may suffer for minority (rare) classes, as machine trained",
"tators[49,50l.ThisenablesCalprictogenerateatraining set",
"models tend to optimize for accuracy in majority (common)",
"set of 16856 labeled text segments, which is 3× larger than",
"classes. [49] and [50] only cover 3 common categories and do",
"thelargestprevious human-annotated data set and contains",
"not cover rare data categories such as health, financial, demo-",
"significantly more information on minority data categories",
"graphic, social media, personal identifier, or survey data. [19]",
"and classes.",
"covers 11? rare categories, but has low accuracy in them due",
"In summary, this paper makes the following contributions:",
"to the limited access to training data in those categories.",
"ToalleviatesomeoftheseissuesT19lcombinesdifferent",
"1. We present Calpric, which synergistically combines",
"data actions when training models to predict certain private",
"crowdsourcing and active learning to produce Calpric",
"2polisis covers non-mobile privacy policies,we only train Calpric on",
"Privacy Policy Corpus (CPPS), which contains 16856",
"categories that are applicable to mobile privacy policies."
] |
{
"arxiv_id": "2401.08038",
"language": "en",
"page_idx": 1
}
|
||
2401.08038-en-3
|
2401.08038-en-3.jpg
|
on each privacy policy. The speed, availability and cost of trained annotators limits the size of the data set. Recognizing this challenge, we propose a crowdsourcing solution reduces the cost of labeling by 9×.
One obvious question that arises with the use of crowdsourced annotators for privacy policies is whether such untrained annotators can sufficiently understand privacy policy language to produce accurate labels [43]. We note however, that since privacy policies are the primary method of obtaining consent from individuals for the collection of information, to obtain meaningful consent, they must be understandable by a representative target individual from whom the information collection would likely happen. For example, Canada's privacy legislation, PIPEDA, states that“the consent of an individual is only valid if it is reasonable to expect that an individual to whom the organization's activities are directed would understand" and the GDPR defines consent as “any freely given, specific, informed and unambiguous indication of the data subject's wishes"—in both cases, the user must “understand" or be “informed.”While it is true that many privacy policy documents fall short of this idealized level of transparency, we argue that the interpretation that an average person has for a privacy policy intended for them (were they to actually read the policy) is still relevant, and is what Calpric tries to predict.
Nevertheless, one cannot simply use crowdsourced annotators and expect to achieve accurate classification, especially in combination with active learning. First, crowdsourced annotators acquired through platforms such as Amazon's mTurk service, generally do best on short, simple tasks, such as single-sentence translation [1] and sentiment analysis [7]. Further, reducing the complexity and length of tasks is beneficial for crowdsourced annotators [8,23]. Second, active learning methods generally assume reliable oracles that always return correct labels, which is not the case for crowdsourced annotators. Previous privacy policy studies presented several crowdsourced annotators with the entire privacy policy, asked them each to assign a label, and then attempt to increase label reliability by only accepting the label if the inter-annotator agreement is above some threshold [43]. A drawback to this is that different annotators may select different text segments to label, and the text segments themselves may not align perfectly, decreasing inter-annotator agreement.
Calpric addresses these challenges by extracting relevant text segments from a privacy policy and presenting these to the crowdsourced annotators so that all of them label the same segments. This both decreases the time they must spend on the task and increases inter-annotator agreement rate, which both decrease the cost of acquiring labels. Calpric re-requests labels for text segments that did not achieve an inter-annotator agreement above an acceptable threshold— a mechanism called relabeling. However, some segments may suffer from chronically poor agreement—that is they do not achieve sufficient agreement even after several attempts.
There can be several reasons for the chronically poor agreement for a segment. First, Calpric may mistakenly extract a segment that is not relevant to the labeling task. Our crowdsourced task also has workers annotate whether a segment is relevant or not, which is then used as a label to retrain the model used to classify if a text segment contains relevant information about a data category. Second, the text segment may be relevant but inherently ambiguous—that is the text is not sufficiently clear about whether collection is occurring or not[32]. For example, in
“ if you log- in using a third party social media account, we may not collect any personal or account
$information\textit{ from that social media provider" }$
the“may” makes it ambiguous whether information is collected or not, and users have no control over the usage of private data. In addition, using the ambiguous term “personal information" without further details makes it impossible to infer which data category the segment refers to and undermines the purpose and value of privacy policies [32]. To prevent wasting more crowdsourcer resources on such text segments, Calpric stops the relabeling attempt after a preset number of iterations, and labels such segments as ambiguous and does not use them for training. While classifying all text segments that do not achieve an acceptable level of inter-annotator agreement as ambiguous necessarily over-approximates the true set of ambiguous text segments, we show that Calpric is still able to achieve high classification accuracy despite the loss of potentially usable text segments due to this over-approximation.
Our current implementation of Calpric uses crowdsourced annotators from Amazon's mTurk service to provide labels used for training. While mTurk captures a large portion of the mobile app user population, who are the target individuals of mobile app privacy policies in our data set, we note that mTurk cannot capture every subgroup of mobile app user. For example, children are not permitted to use the mTurk service, so our current Calpric models cannot predict how children might interpret privacy policies. Nonetheless, this is a limitation of mTurk, not Calpric\_we believe that our current implementation of Calpric can be extended in a straightforward manner to predict for any subgroup of individuals so long as annotators from that subgroup can be recruited to provide labels from which Calpric can learn.
# 2.3 Segmentation
As mentioned earlier, extracting text segments is crucial to Calpric's ability to provide accurate, broad classification of privacy policies at a lower annotation cost. While previous work, such as Polisis, used existing off-the-shelf segmentation tools [17], Calpric requires text segments for annotation, not just classification. For accurate annotation by crowdsourced workers, Calpric must extract text segments that not only semantically coherent to a topic, but relevant to the specific data
|
[
[
0.0849609375,
0.0921630859375,
0.48193359375,
0.0921630859375,
0.48193359375,
0.1060791015625,
0.0849609375,
0.1060791015625
],
[
0.515625,
0.09027099609375,
0.9140625,
0.091552734375,
0.9140625,
0.1060791015625,
0.515625,
0.10479736328125
],
[
0.0849609375,
0.106689453125,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.0849609375,
0.12060546875
],
[
0.515625,
0.10797119140625,
0.9140625,
0.10797119140625,
0.9140625,
0.121826171875,
0.515625,
0.121826171875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.12249755859375,
0.48193359375,
0.136962890625,
0.0849609375,
0.1363525390625
],
[
0.515625,
0.12371826171875,
0.9140625,
0.12249755859375,
0.9140625,
0.1363525390625,
0.515625,
0.136962890625
],
[
0.0849609375,
0.1383056640625,
0.265625,
0.1383056640625,
0.265625,
0.1514892578125,
0.0849609375,
0.1514892578125
],
[
0.51708984375,
0.1383056640625,
0.9140625,
0.1383056640625,
0.9140625,
0.1514892578125,
0.51708984375,
0.1514892578125
],
[
0.101318359375,
0.1546630859375,
0.484375,
0.1546630859375,
0.484375,
0.1685791015625,
0.101318359375,
0.1685791015625
],
[
0.51708984375,
0.1514892578125,
0.9140625,
0.1514892578125,
0.9140625,
0.1654052734375,
0.51708984375,
0.1654052734375
],
[
0.0849609375,
0.169189453125,
0.483642578125,
0.169189453125,
0.483642578125,
0.18310546875,
0.0849609375,
0.18310546875
],
[
0.51953125,
0.169189453125,
0.9111328125,
0.169189453125,
0.9111328125,
0.179931640625,
0.51953125,
0.179931640625
],
[
0.08660888671875,
0.184326171875,
0.4794921875,
0.184326171875,
0.4794921875,
0.1976318359375,
0.08660888671875,
0.1976318359375
],
[
0.51708984375,
0.1817626953125,
0.91259765625,
0.1817626953125,
0.91259765625,
0.1956787109375,
0.51708984375,
0.1956787109375
],
[
0.08660888671875,
0.199462890625,
0.483642578125,
0.199462890625,
0.483642578125,
0.21337890625,
0.08660888671875,
0.21337890625
],
[
0.5185546875,
0.199462890625,
0.9111328125,
0.199462890625,
0.9111328125,
0.210205078125,
0.5185546875,
0.210205078125
],
[
0.08660888671875,
0.21533203125,
0.483642578125,
0.21533203125,
0.483642578125,
0.2291259765625,
0.08660888671875,
0.2291259765625
],
[
0.51708984375,
0.21337890625,
0.91259765625,
0.21337890625,
0.91259765625,
0.2266845703125,
0.51708984375,
0.2266845703125
],
[
0.0849609375,
0.2298583984375,
0.48193359375,
0.2298583984375,
0.48193359375,
0.24365234375,
0.0849609375,
0.24365234375
],
[
0.51708984375,
0.2291259765625,
0.6845703125,
0.2291259765625,
0.6845703125,
0.242431640625,
0.51708984375,
0.242431640625
],
[
0.0849609375,
0.24560546875,
0.48193359375,
0.24560546875,
0.48193359375,
0.2587890625,
0.0849609375,
0.2587890625
],
[
0.55126953125,
0.2474365234375,
0.8798828125,
0.248779296875,
0.8798828125,
0.26318359375,
0.55126953125,
0.261962890625
],
[
0.08660888671875,
0.260009765625,
0.483642578125,
0.260009765625,
0.483642578125,
0.2734375,
0.08660888671875,
0.2734375
],
[
0.54833984375,
0.26318359375,
0.88134765625,
0.26318359375,
0.88134765625,
0.276611328125,
0.54833984375,
0.276611328125
],
[
0.0849609375,
0.27587890625,
0.48193359375,
0.27587890625,
0.48193359375,
0.2890625,
0.0849609375,
0.2890625
],
[
0.54736328125,
0.279052734375,
0.84619140625,
0.279052734375,
0.84619140625,
0.292236328125,
0.54736328125,
0.292236328125
],
[
0.0849609375,
0.2890625,
0.48193359375,
0.290283203125,
0.48193359375,
0.304931640625,
0.0849609375,
0.3037109375
],
[
0.51708984375,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.3125,
0.51708984375,
0.3125
],
[
0.0849609375,
0.30615234375,
0.48193359375,
0.30615234375,
0.48193359375,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.5185546875,
0.313232421875,
0.9140625,
0.313232421875,
0.9140625,
0.326904296875,
0.5185546875,
0.326904296875
],
[
0.08660888671875,
0.322509765625,
0.48046875,
0.322509765625,
0.48046875,
0.333251953125,
0.08660888671875,
0.333251953125
],
[
0.51708984375,
0.328857421875,
0.91259765625,
0.32763671875,
0.91259765625,
0.341552734375,
0.51708984375,
0.34228515625
],
[
0.0882568359375,
0.335205078125,
0.48046875,
0.335205078125,
0.48046875,
0.348388671875,
0.0882568359375,
0.348388671875
],
[
0.51708984375,
0.343505859375,
0.91259765625,
0.343505859375,
0.91259765625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.08416748046875,
0.35107421875,
0.48046875,
0.349853515625,
0.48046875,
0.3642578125,
0.08416748046875,
0.365478515625
],
[
0.5185546875,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.37255859375,
0.5185546875,
0.37255859375
],
[
0.0882568359375,
0.367431640625,
0.48046875,
0.367431640625,
0.48046875,
0.37744140625,
0.0882568359375,
0.37744140625
],
[
0.08660888671875,
0.380615234375,
0.48046875,
0.380615234375,
0.48046875,
0.39453125,
0.08660888671875,
0.39453125
],
[
0.51708984375,
0.375,
0.91259765625,
0.375,
0.91259765625,
0.38818359375,
0.51708984375,
0.38818359375
],
[
0.08660888671875,
0.396484375,
0.48193359375,
0.396484375,
0.48193359375,
0.410400390625,
0.08660888671875,
0.410400390625
],
[
0.5185546875,
0.391357421875,
0.91259765625,
0.391357421875,
0.91259765625,
0.401611328125,
0.5185546875,
0.401611328125
],
[
0.5185546875,
0.40478515625,
0.9140625,
0.40478515625,
0.9140625,
0.41845703125,
0.5185546875,
0.41845703125
],
[
0.08660888671875,
0.410888671875,
0.48193359375,
0.410888671875,
0.48193359375,
0.4248046875,
0.08660888671875,
0.4248046875
],
[
0.51708984375,
0.419189453125,
0.91259765625,
0.419189453125,
0.91259765625,
0.43310546875,
0.51708984375,
0.43310546875
],
[
0.08660888671875,
0.4267578125,
0.48193359375,
0.4267578125,
0.48193359375,
0.43994140625,
0.08660888671875,
0.43994140625
],
[
0.51708984375,
0.43505859375,
0.91259765625,
0.43505859375,
0.91259765625,
0.448974609375,
0.51708984375,
0.448974609375
],
[
0.0849609375,
0.44140625,
0.48046875,
0.44140625,
0.48046875,
0.455810546875,
0.0849609375,
0.455810546875
],
[
0.51708984375,
0.449462890625,
0.91259765625,
0.449462890625,
0.91259765625,
0.46337890625,
0.51708984375,
0.46337890625
],
[
0.08660888671875,
0.458251953125,
0.1871337890625,
0.458251953125,
0.1871337890625,
0.4697265625,
0.08660888671875,
0.4697265625
],
[
0.5185546875,
0.466552734375,
0.91259765625,
0.466552734375,
0.91259765625,
0.4765625,
0.5185546875,
0.4765625
],
[
0.101318359375,
0.473388671875,
0.483642578125,
0.473388671875,
0.483642578125,
0.4873046875,
0.101318359375,
0.4873046875
],
[
0.5185546875,
0.48095703125,
0.9111328125,
0.48095703125,
0.9111328125,
0.4912109375,
0.5185546875,
0.4912109375
],
[
0.08416748046875,
0.4873046875,
0.483642578125,
0.488037109375,
0.483642578125,
0.50244140625,
0.08416748046875,
0.501953125
],
[
0.5185546875,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.5068359375,
0.5185546875,
0.5068359375
],
[
0.0849609375,
0.50390625,
0.483642578125,
0.50390625,
0.483642578125,
0.517578125,
0.0849609375,
0.517578125
],
[
0.51708984375,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51708984375,
0.5234375
],
[
0.0849609375,
0.51806640625,
0.48046875,
0.51806640625,
0.48046875,
0.5322265625,
0.0849609375,
0.5322265625
],
[
0.53369140625,
0.52587890625,
0.9140625,
0.52587890625,
0.9140625,
0.5390625,
0.53369140625,
0.5390625
],
[
0.0849609375,
0.5341796875,
0.48046875,
0.5341796875,
0.48046875,
0.54736328125,
0.0849609375,
0.54736328125
],
[
0.51708984375,
0.54150390625,
0.91259765625,
0.54150390625,
0.91259765625,
0.5546875,
0.51708984375,
0.5546875
],
[
0.0849609375,
0.548828125,
0.48193359375,
0.548828125,
0.48193359375,
0.5625,
0.0849609375,
0.5625
],
[
0.51708984375,
0.556640625,
0.91259765625,
0.556640625,
0.91259765625,
0.57080078125,
0.51708984375,
0.57080078125
],
[
0.08660888671875,
0.564453125,
0.48193359375,
0.564453125,
0.48193359375,
0.57763671875,
0.08660888671875,
0.57763671875
],
[
0.51708984375,
0.5712890625,
0.91259765625,
0.5712890625,
0.91259765625,
0.5859375,
0.51708984375,
0.5859375
],
[
0.08416748046875,
0.57763671875,
0.48046875,
0.5791015625,
0.48046875,
0.59326171875,
0.08416748046875,
0.59228515625
],
[
0.51708984375,
0.5869140625,
0.9140625,
0.5869140625,
0.9140625,
0.6015625,
0.51708984375,
0.6015625
],
[
0.08416748046875,
0.59326171875,
0.48046875,
0.5947265625,
0.48046875,
0.60791015625,
0.08416748046875,
0.60693359375
],
[
0.5185546875,
0.6015625,
0.91259765625,
0.6015625,
0.91259765625,
0.61572265625,
0.5185546875,
0.61572265625
],
[
0.08660888671875,
0.611328125,
0.4794921875,
0.611328125,
0.4794921875,
0.6220703125,
0.08660888671875,
0.6220703125
],
[
0.5185546875,
0.6171875,
0.9140625,
0.6171875,
0.9140625,
0.630859375,
0.5185546875,
0.630859375
],
[
0.0849609375,
0.625,
0.48046875,
0.625,
0.48046875,
0.63818359375,
0.0849609375,
0.63818359375
],
[
0.5185546875,
0.6337890625,
0.9111328125,
0.6337890625,
0.9111328125,
0.64453125,
0.5185546875,
0.64453125
],
[
0.0849609375,
0.64013671875,
0.48193359375,
0.64013671875,
0.48193359375,
0.65380859375,
0.0849609375,
0.65380859375
],
[
0.51708984375,
0.646484375,
0.9140625,
0.64599609375,
0.9140625,
0.66015625,
0.51708984375,
0.6611328125
],
[
0.08660888671875,
0.65576171875,
0.48046875,
0.65576171875,
0.48046875,
0.66650390625,
0.08660888671875,
0.66650390625
],
[
0.0849609375,
0.66943359375,
0.48046875,
0.66943359375,
0.48046875,
0.68310546875,
0.0849609375,
0.68310546875
],
[
0.51953125,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.673828125,
0.51953125,
0.673828125
],
[
0.51708984375,
0.6767578125,
0.91259765625,
0.6767578125,
0.91259765625,
0.68994140625,
0.51708984375,
0.68994140625
],
[
0.0849609375,
0.68505859375,
0.48046875,
0.6845703125,
0.48046875,
0.69775390625,
0.0849609375,
0.69873046875
],
[
0.51708984375,
0.6923828125,
0.9140625,
0.6923828125,
0.9140625,
0.70556640625,
0.51708984375,
0.70556640625
],
[
0.0849609375,
0.69970703125,
0.48046875,
0.69970703125,
0.48046875,
0.71337890625,
0.0849609375,
0.71337890625
],
[
0.51708984375,
0.70849609375,
0.9140625,
0.70849609375,
0.9140625,
0.7216796875,
0.51708984375,
0.7216796875
],
[
0.0849609375,
0.71533203125,
0.48193359375,
0.71533203125,
0.48193359375,
0.72900390625,
0.0849609375,
0.72900390625
],
[
0.51708984375,
0.72265625,
0.81201171875,
0.72265625,
0.81201171875,
0.736328125,
0.51708984375,
0.736328125
],
[
0.0849609375,
0.73095703125,
0.378173828125,
0.73095703125,
0.378173828125,
0.744140625,
0.0849609375,
0.744140625
],
[
0.101318359375,
0.74560546875,
0.483642578125,
0.7470703125,
0.483642578125,
0.76123046875,
0.101318359375,
0.76025390625
],
[
0.0849609375,
0.76123046875,
0.48193359375,
0.76220703125,
0.48193359375,
0.7763671875,
0.0849609375,
0.77587890625
],
[
0.51708984375,
0.7646484375,
0.67724609375,
0.7646484375,
0.67724609375,
0.77880859375,
0.51708984375,
0.77880859375
],
[
0.08660888671875,
0.77880859375,
0.4794921875,
0.77880859375,
0.4794921875,
0.7890625,
0.08660888671875,
0.7890625
],
[
0.0849609375,
0.79248046875,
0.48193359375,
0.79248046875,
0.48193359375,
0.806640625,
0.0849609375,
0.806640625
],
[
0.5185546875,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.5185546875,
0.80615234375
],
[
0.0849609375,
0.80810546875,
0.48046875,
0.80810546875,
0.48046875,
0.8212890625,
0.0849609375,
0.8212890625
],
[
0.5185546875,
0.80810546875,
0.9140625,
0.80810546875,
0.9140625,
0.8212890625,
0.5185546875,
0.8212890625
],
[
0.0849609375,
0.82275390625,
0.483642578125,
0.82275390625,
0.483642578125,
0.8359375,
0.0849609375,
0.8359375
],
[
0.51708984375,
0.82275390625,
0.91259765625,
0.82275390625,
0.91259765625,
0.8359375,
0.51708984375,
0.8359375
],
[
0.08660888671875,
0.8388671875,
0.48046875,
0.8388671875,
0.48046875,
0.849609375,
0.08660888671875,
0.849609375
],
[
0.51708984375,
0.8369140625,
0.91259765625,
0.8369140625,
0.91259765625,
0.85107421875,
0.51708984375,
0.85107421875
],
[
0.08660888671875,
0.85400390625,
0.47802734375,
0.85400390625,
0.47802734375,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.51708984375,
0.85302734375,
0.91259765625,
0.85302734375,
0.91259765625,
0.8662109375,
0.51708984375,
0.8662109375
],
[
0.08660888671875,
0.86865234375,
0.48193359375,
0.86865234375,
0.48193359375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.5185546875,
0.869140625,
0.9111328125,
0.869140625,
0.9111328125,
0.8798828125,
0.5185546875,
0.8798828125
],
[
0.0849609375,
0.88330078125,
0.48193359375,
0.88330078125,
0.48193359375,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.515625,
0.896484375
],
[
0.0849609375,
0.8984375,
0.483642578125,
0.8984375,
0.483642578125,
0.912109375,
0.0849609375,
0.912109375
],
[
0.515625,
0.8984375,
0.91259765625,
0.8984375,
0.91259765625,
0.912109375,
0.515625,
0.912109375
]
] |
[
"on each privacy policy. The speed, availability and cost of",
"There can be several reasons for the chronically poor agree-",
"trained annotators limits the size of the data set. Recognizing",
"ment for a segment. First, Calpric may mistakenly extract a",
"this challenge, we propose a crowdsourcing solution reduces",
"segment that isnot relevant tothelabeling task.Our crowd-",
"the cost of labeling by 9 ×.",
"sourced task also has workers annotate whether a segment",
"One obvious question that arises with the use of crowd-",
"is relevant or not, which is then used as a label to retrain the",
"sourced annotators for privacy policies is whether such un-",
"model usedto classifyifatext segment contains relevantin",
"trained annotators can sufficiently understand privacy policy",
"formation about a data category. Second, the text segment",
"language to produce accurate labels [43]. We note however,",
"mayberelevant butinherentlyambiguous-that isthetextis",
"that since privacy policies are the primary method of obtain-",
"not sufficiently clear about whether collection is occurring or",
"ing consent from individuals for the collection of information,",
"not [32]. For example, in",
"to obtain meaningful consent, they must be understandable",
"\"if you log-in using a third party social media ac-",
"by a representative target individual from whom the informa-",
"count,we may not collectany personal or account",
"tion collection would likely happen. For example, Canada's",
"informationfrom that social mediaprovider\"",
"privacy legislation, PIPEDA, states that “the consent of an",
"the \"may\" makes it ambiguous whether information is col-",
"individual is only valid if it is reasonable to expect that an",
"lected or not, and users have no control over the usage of",
"individual to whomthe organization's activities are directed",
"private data. In addition,using the ambiguous term“personal",
"would understand\"and the GDPR defines consent as “any",
"information without further details makes it impossible to in-",
"freely given, specific, informed and unambiguous indication",
"fer which data category the segment refers to and undermines",
"ofthe datasubject'swishes\"-inboth cases,theusermust",
"\"understand\"or be \"informed.\"While it is true that many",
"the purpose and value of privacy policies [32]. To prevent",
"privacy policy documents fall short of this idealized level of",
"wastingmorecrowdsourcerresourcesonsuchtextsegments.",
"Calpric stops the relabeling attempt after a preset number of it-",
"transparency, we argue that the interpretation that an average",
"erations, and labels such segments as ambiguous and does not",
"person has for a privacy policy intended for them (were they",
"use them for training. While classifying all text segments that",
"to actually read the policy) is still relevant, and is what Calpric",
"do not achieve an acceptable level of inter-annotator agree-",
"triestopredict.",
"mentas ambiguousnecessarilyover-approximatesthetrueset",
"Nevertheless, one cannot simply use crowdsourced annota-",
"ofambiguoustextsegments,weshowthatCalpricisstillable",
"tors and expect to achieve accurate classification, especially in",
"toachievehighclassificationaccuracydespitethelossofpo-",
"combination with active learning. First, crowdsourced annota-",
"tentially usable text segments due to this over-approximation",
"tors acquired through platforms such as Amazon's mTurk",
"Our current implementation of Calpric uses crowdsourced",
" service, generally do best on short, simple tasks, such as",
"annotators from Amazon's mTurk service to provide labels",
" single-sentence translation [1] and sentiment analysis [7]. Fur",
"used for training. While mTurk captures a large portion of the",
"ther, reducing the complexity and length of tasks is beneficial",
"mobile app user population, who are the target individuals",
"for crowdsourced annotators [8, 23]. Second, active learning",
"of mobile app privacy policies in our data set, we note that",
"methods generally assume reliable oracles that always return",
"mTurk cannot capture every subgroup of mobile app user. For",
"correct labels, which is not the case for crowdsourced an.",
"example, children are not permitted to use the mTurk service,",
"notators. Previous privacy policy studies presented several",
"so our currentCalpricmodels cannotpredicthow children",
"crowdsourced annotators with the entire privacy policy, asked",
"might interpret privacy policies. Nonetheless, this is a lim-",
"them each to assign a label, and then attemptto increaselabel",
"reliability by only accepting the label if the inter-annotator",
"itation of mTurk.notCalpric--webelievethatour current",
"implementation of Calpric can be extended in a straightfor-",
"agreement is above some threshold [43].A drawback to this",
"ward manner to predict for any subgroup of individuals so",
"is that different annotators may select different text segments",
"long as annotators from that subgroup can be recruited to",
"to label, and the text segments themselves may not align per",
"provide labels from which Calpric can learn.",
"fectly, decreasing inter-annotator agreement.",
"Calpric addresses these challenges by extracting relevant",
"text segments from a privacy policy and presenting these",
"2.3Segmentation",
"to the crowdsourced annotators so that all of them label",
"the same segments. This both decreases the time they must",
"As mentioned earlier, extracting text segments is crucial to",
" spend on the task and increases inter-annotator agreement",
"Calpric's ability to provide accurate, broad classification of",
"rate, which both decrease the cost of acquiring labels. Calpric",
"privacy policies at a lower annotation cost. While previous",
"re-requests labels for text segments that did not achieve an",
"work, such as Polisis, used existing off-the-shelf segmentation",
"inter-annotatoragreementaboveanacceptablethreshold-",
"tools [17], Calpric requires text segments for annotation, not",
"a mechanism called relabeling. However, some segments",
"just classification.For accurateannotation by crowdsourced",
"may suffer from chronically poor agreement-that is they do",
"workers, Calpric must extract text segments that not only se-",
"not achieve sufficient agreement even after several attempts.",
" mantically coherent to a topic, but relevant to the specific data"
] |
{
"arxiv_id": "2401.08038",
"language": "en",
"page_idx": 3
}
|
||
1910.14599-en-0
|
1910.14599-en-0.jpg
|
### $\textbf{Adversarial NLI: A New Benchmark}$ for Natural Language Understanding
Yixin Nie$^*$,Adina Williams$^\dagger$,Emily Dinan$^\dagger,\textbf{Mohit Bansal}^*,\textbf{Jason Weston}^\dagger,\textbf{Douwe Kiela}^\dagger$
$^{*}\mathbf{UNC}$ Chapel Hill
$^{\dagger}$Facebook AI Research
### $\mathbf{Abstract}$
We introduce a new large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure. We show that training models on this new dataset leads to state-of-the-art performance on a variety of popular NLI benchmarks, while posing a more difficult challenge with its new test set. Our analysis sheds light on the shortcomings of current state-of-theart models, and shows that non-expert annotators are successful at finding their weaknesses. The data collection method can be applied in a never-ending learning scenario, becoming a moving target for NLU, rather than a static benchmark that will quickly saturate.
### 1 Introduction
Progress in AI has been driven by, among other things, the development of challenging large-scale benchmarks like ImageNet (Russakovsky et al., 2015) in computer vision, and SNLI (Bowman et al., 2015), SQuAD (Rajpurkar et al., 2016), and others in natural language processing (NLP). Recently, for natural language understanding (NLU) in particular, the focus has shifted to combined benchmarks like SentEval (Conneau and Kiela, 2018) and GLUE (Wang et al., 2018), which track model performance on multiple tasks and provide a unified platform for analysis.
With the rapid pace of advancement in AI, however, NLU benchmarks struggle to keep up with model improvement. Whereas it took around 15 years to achieve “near-human performance”on MNIST (LeCun et al., 1998; Cireșan et al., 2012; Wan et al., 2013) and approximately 7 years to surpass humans on ImageNet (Deng et al., 2009; Russakovsky et al., 2015; He et al., 2016), the GLUE benchmark did not last as long as we would have hoped after the advent of BERT (Devlin et al.,
2018), and rapidly had to be extended into SuperGLUE (Wang et al., 2019). This raises an important question: Can we collect a large benchmark dataset that can last longer?
The speed with which benchmarks become obsolete raises another important question: are current NLU models genuinely as good as their high performance on benchmarks suggests? A growing body of evidence shows that state-of-the-art models learn to exploit spurious statistical patterns in datasets (Gururangan et al., 2018; Poliak et al., 2018; Tsuchiya, 2018; Glockner et al., 2018; Geva et al., 2019; McCoy et al., 2019), instead of learning meaning in the flexible and generalizable way that humans do. Given this, human annotators—be they seasoned NLP researchers or non-experts— might easily be able to construct examples that expose model brittleness.
We propose an iterative, adversarial human-andmodel-in-the-loop solution for NLU dataset collection that addresses both benchmark longevity and robustness issues. In the first stage, human annotators devise examples that our current best models cannot determine the correct label for. These resulting hard examples—which should expose additional model weaknesses—can be added to the training set and used to train a stronger model. We then subject the strengthened model to the same procedure and collect weaknesses over several rounds. After each round, we train a new model and set aside a new test set. The process can be iteratively repeated in a never-ending learning (Mitchell et al., 2018) setting, with the model getting stronger and the test set getting harder in each new round. Thus, not only is the resultant dataset harder than existing benchmarks, but this process also yields a“moving post” dynamic target for NLU systems, rather than a static benchmark that will eventually saturate.
Our approach draws inspiration from recent ef-
|
[
[
0.313232421875,
0.0831298828125,
0.68994140625,
0.0831298828125,
0.68994140625,
0.097412109375,
0.313232421875,
0.097412109375
],
[
0.307373046875,
0.1009521484375,
0.69287109375,
0.1015625,
0.69287109375,
0.119384765625,
0.307373046875,
0.1181640625
],
[
0.12005615234375,
0.15087890625,
0.88330078125,
0.1519775390625,
0.88330078125,
0.169189453125,
0.12005615234375,
0.168701171875
],
[
0.426513671875,
0.169189453125,
0.580078125,
0.17041015625,
0.580078125,
0.185302734375,
0.426513671875,
0.18408203125
],
[
0.405517578125,
0.18408203125,
0.60107421875,
0.185302734375,
0.60107421875,
0.203125,
0.405517578125,
0.201904296875
],
[
0.26611328125,
0.265380859375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.26611328125,
0.27978515625
],
[
0.5146484375,
0.265380859375,
0.88330078125,
0.265380859375,
0.88330078125,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.5146484375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.14697265625,
0.29638671875,
0.4609375,
0.29638671875,
0.4609375,
0.31005859375,
0.14697265625,
0.31005859375
],
[
0.51318359375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.0235137939453125,
0.305908203125,
0.06634521484375,
0.305908203125,
0.06378173828125,
0.432861328125,
0.0201568603515625,
0.432861328125
],
[
0.1485595703125,
0.3134765625,
0.459228515625,
0.3134765625,
0.459228515625,
0.32373046875,
0.1485595703125,
0.32373046875
],
[
0.51318359375,
0.314208984375,
0.6650390625,
0.314208984375,
0.6650390625,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.14697265625,
0.325927734375,
0.4609375,
0.325927734375,
0.4609375,
0.339111328125,
0.14697265625,
0.339111328125
],
[
0.52880859375,
0.330078125,
0.8818359375,
0.330078125,
0.8818359375,
0.34326171875,
0.52880859375,
0.34326171875
],
[
0.14697265625,
0.339111328125,
0.4609375,
0.339111328125,
0.4609375,
0.352783203125,
0.14697265625,
0.352783203125
],
[
0.51318359375,
0.347412109375,
0.88671875,
0.347412109375,
0.88671875,
0.361083984375,
0.51318359375,
0.361083984375
],
[
0.14697265625,
0.35400390625,
0.462646484375,
0.35400390625,
0.462646484375,
0.36767578125,
0.14697265625,
0.36767578125
],
[
0.14697265625,
0.3681640625,
0.4609375,
0.3681640625,
0.4609375,
0.3818359375,
0.14697265625,
0.3818359375
],
[
0.51318359375,
0.363525390625,
0.88330078125,
0.363525390625,
0.88330078125,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.51318359375,
0.377685546875,
0.88330078125,
0.377685546875,
0.88330078125,
0.392578125,
0.51318359375,
0.392578125
],
[
0.1485595703125,
0.38427734375,
0.459228515625,
0.38427734375,
0.459228515625,
0.394287109375,
0.1485595703125,
0.394287109375
],
[
0.14697265625,
0.396728515625,
0.4609375,
0.396728515625,
0.4609375,
0.410400390625,
0.14697265625,
0.410400390625
],
[
0.51318359375,
0.394287109375,
0.884765625,
0.393798828125,
0.884765625,
0.407958984375,
0.51318359375,
0.4091796875
],
[
0.1485595703125,
0.412109375,
0.459228515625,
0.412109375,
0.459228515625,
0.4228515625,
0.1485595703125,
0.4228515625
],
[
0.5146484375,
0.412109375,
0.88330078125,
0.412109375,
0.88330078125,
0.42578125,
0.5146484375,
0.42578125
],
[
0.14697265625,
0.42578125,
0.4609375,
0.42578125,
0.4609375,
0.439453125,
0.14697265625,
0.439453125
],
[
0.5146484375,
0.427490234375,
0.88330078125,
0.427490234375,
0.88330078125,
0.441162109375,
0.5146484375,
0.441162109375
],
[
0.0235137939453125,
0.43408203125,
0.06378173828125,
0.43408203125,
0.06378173828125,
0.5126953125,
0.0235137939453125,
0.5126953125
],
[
0.14697265625,
0.439453125,
0.459228515625,
0.439453125,
0.459228515625,
0.453125,
0.14697265625,
0.453125
],
[
0.5146484375,
0.443603515625,
0.884765625,
0.443603515625,
0.884765625,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.145263671875,
0.453125,
0.4609375,
0.45361328125,
0.4609375,
0.46728515625,
0.145263671875,
0.466064453125
],
[
0.51318359375,
0.457763671875,
0.88330078125,
0.458984375,
0.88330078125,
0.473876953125,
0.51318359375,
0.47265625
],
[
0.1485595703125,
0.468505859375,
0.4609375,
0.468505859375,
0.4609375,
0.482177734375,
0.1485595703125,
0.482177734375
],
[
0.5146484375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4892578125,
0.5146484375,
0.4892578125
],
[
0.1485595703125,
0.482177734375,
0.4609375,
0.482177734375,
0.4609375,
0.495361328125,
0.1485595703125,
0.495361328125
],
[
0.14697265625,
0.49462890625,
0.39794921875,
0.495361328125,
0.39794921875,
0.5087890625,
0.14697265625,
0.50830078125
],
[
0.5146484375,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50439453125,
0.5146484375,
0.50439453125
],
[
0.51318359375,
0.50732421875,
0.87158203125,
0.505859375,
0.87158203125,
0.52099609375,
0.51318359375,
0.521484375
],
[
0.019317626953125,
0.51318359375,
0.060455322265625,
0.51318359375,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.130126953125,
0.52783203125,
0.26123046875,
0.52783203125,
0.26123046875,
0.54150390625,
0.130126953125,
0.54150390625
],
[
0.51123046875,
0.52392578125,
0.884765625,
0.5224609375,
0.884765625,
0.53759765625,
0.51123046875,
0.5380859375
],
[
0.124267578125,
0.5322265625,
0.130126953125,
0.5322265625,
0.130126953125,
0.53759765625,
0.124267578125,
0.53759765625
],
[
0.51123046875,
0.54052734375,
0.70361328125,
0.5380859375,
0.70361328125,
0.5517578125,
0.51123046875,
0.55419921875
],
[
0.1175537109375,
0.5546875,
0.489501953125,
0.5546875,
0.489501953125,
0.568359375,
0.1175537109375,
0.568359375
],
[
0.5322265625,
0.5556640625,
0.884765625,
0.5556640625,
0.884765625,
0.5693359375,
0.5322265625,
0.5693359375
],
[
0.1175537109375,
0.57080078125,
0.489501953125,
0.57177734375,
0.489501953125,
0.5859375,
0.1175537109375,
0.5849609375
],
[
0.51123046875,
0.572265625,
0.884765625,
0.572265625,
0.884765625,
0.5859375,
0.51123046875,
0.5859375
],
[
0.1175537109375,
0.5859375,
0.491943359375,
0.5859375,
0.491943359375,
0.599609375,
0.1175537109375,
0.599609375
],
[
0.51123046875,
0.58740234375,
0.88671875,
0.58837890625,
0.88671875,
0.6025390625,
0.51123046875,
0.6015625
],
[
0.12005615234375,
0.60400390625,
0.490234375,
0.60400390625,
0.490234375,
0.61767578125,
0.12005615234375,
0.61767578125
],
[
0.51123046875,
0.60498046875,
0.88671875,
0.60498046875,
0.88671875,
0.6181640625,
0.51123046875,
0.6181640625
],
[
0.118408203125,
0.619140625,
0.489501953125,
0.619140625,
0.489501953125,
0.6328125,
0.118408203125,
0.6328125
],
[
0.51318359375,
0.62158203125,
0.88818359375,
0.62158203125,
0.88818359375,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.1175537109375,
0.63525390625,
0.49365234375,
0.63427734375,
0.49365234375,
0.65185546875,
0.1175537109375,
0.65283203125
],
[
0.51123046875,
0.63720703125,
0.884765625,
0.63720703125,
0.884765625,
0.65087890625,
0.51123046875,
0.65087890625
],
[
0.118408203125,
0.65283203125,
0.490234375,
0.65283203125,
0.490234375,
0.66650390625,
0.118408203125,
0.66650390625
],
[
0.5146484375,
0.65478515625,
0.8818359375,
0.65478515625,
0.8818359375,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.1175537109375,
0.66748046875,
0.48779296875,
0.66748046875,
0.48779296875,
0.68115234375,
0.1175537109375,
0.68115234375
],
[
0.51123046875,
0.66845703125,
0.8818359375,
0.66845703125,
0.8818359375,
0.681640625,
0.51123046875,
0.681640625
],
[
0.12176513671875,
0.68603515625,
0.489501953125,
0.68603515625,
0.489501953125,
0.69677734375,
0.12176513671875,
0.69677734375
],
[
0.51318359375,
0.68505859375,
0.884765625,
0.68505859375,
0.884765625,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.12005615234375,
0.70068359375,
0.489501953125,
0.70068359375,
0.489501953125,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.51123046875,
0.69970703125,
0.88330078125,
0.70068359375,
0.88330078125,
0.71435546875,
0.51123046875,
0.71337890625
],
[
0.12005615234375,
0.71630859375,
0.489501953125,
0.71630859375,
0.489501953125,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.51318359375,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.118408203125,
0.73291015625,
0.345947265625,
0.73291015625,
0.345947265625,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.51123046875,
0.732421875,
0.880859375,
0.732421875,
0.880859375,
0.7451171875,
0.51123046875,
0.7451171875
],
[
0.1368408203125,
0.7490234375,
0.490234375,
0.75,
0.490234375,
0.76416015625,
0.1368408203125,
0.76318359375
],
[
0.5146484375,
0.75,
0.884765625,
0.75,
0.884765625,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.5146484375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.12005615234375,
0.7822265625,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.1175537109375,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.1175537109375,
0.8125
],
[
0.51318359375,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51318359375,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.490234375,
0.81396484375,
0.490234375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51318359375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.845703125,
0.12005615234375,
0.845703125
],
[
0.51318359375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.118408203125,
0.8466796875,
0.49365234375,
0.845703125,
0.49365234375,
0.85986328125,
0.118408203125,
0.86083984375
],
[
0.51318359375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.12005615234375,
0.8623046875,
0.490234375,
0.8623046875,
0.490234375,
0.8759765625,
0.12005615234375,
0.8759765625
],
[
0.51318359375,
0.8623046875,
0.884765625,
0.8623046875,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.876953125,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.118408203125,
0.89111328125
],
[
0.5146484375,
0.87890625,
0.724609375,
0.87890625,
0.724609375,
0.892578125,
0.5146484375,
0.892578125
],
[
0.118408203125,
0.89453125,
0.490234375,
0.8935546875,
0.490234375,
0.90771484375,
0.118408203125,
0.9091796875
],
[
0.53076171875,
0.89453125,
0.884765625,
0.8935546875,
0.884765625,
0.90771484375,
0.53076171875,
0.9091796875
]
] |
[
"Adversarial NLI: A New Benchmark",
"for Natural Language Understanding",
"Yixin Nie*, Adina Williamst, Emily Dinant, Mohit Bansal*, Jason Westont, Douwe Kielat",
"*UNC Chapel Hill",
"tFacebook AI Research",
"Abstract",
"2018), and rapidly had to be extended into Super-",
"GLUE (Wang et al., 2019). This raises an important",
"We introduce a new large-scale NLI bench-",
"question: Can we collect a large benchmark dataset",
"2 9",
"mark dataset, collected via an iterative, ad-",
"that can last longer?",
"versarial human-and-model-in-the-loop proce-",
" The speed with which benchmarks become ob.",
"dure. We show that training models on this",
"solete raises another important question: are cur-",
"new dataset leads to state-of-the-art perfor-",
"mance on a variety of popular NLI bench-",
"rent NLU models genuinely as good as their high",
"performance on benchmarks suggests? A grow-",
"marks,whileposingamoredifficultchallenge",
"with its new test set. Our analysis sheds light",
"ing body of evidence shows that state-of-the-art",
"on the shortcomings of current state-of-the",
"models learn to exploit spurious statistical patterns",
"art models, and shows that non-expert annota-",
"in datasets (Gururangan et al., 2018; Poliak et al.,",
"[T's]",
"tors are successful at finding their weaknesses",
"2018; Tsuchiya, 2018; Glockner et al., 2018; Geva",
"The data collection method can be applied in",
"et al., 2019; McCoy et al., 2019), instead of learn-",
"a never-ending learning scenario, becoming a",
"ing meaning in the flexible and generalizable way",
"moving target for NLU, rather than a static",
"benchmark that will quickly saturate.",
"that humans do.Given this, human annotators-be",
"they seasoned NLP researchers or non-experts-",
"60161:11",
"Introduction",
"might easily be able to construct examples that",
"",
"expose model brittleness.",
"Progress in AI has been driven by, among other",
"We propose an iterative, adversarial human-and-",
"things, the development of challenging large-scale",
"model-in-the-loop solution for NLU dataset collec-",
"benchmarks like ImageNet (Russakovsky et al.,",
"tion that addresses both benchmark longevity and",
"2015) in computer vision, and SNLI (Bowman",
"robustness issues. In the first stage, human anno-",
"et al., 2015), SOuAD (Rajpurkar et al.. 2016), and",
"tators devise examples that our current best mod-",
"others in natural language processing (NLP). Re-",
"els cannot determine the correct label for. These",
"cently, for natural language understanding (NLU)",
"resultinghardexamples-whichshouldexposead",
"in particular, the focus has shifted to combined",
"ditional model weaknesses-can be added to the",
"benchmarkslikeSentEval(ConneauandKiela,",
"training set and used to train a stronger model.",
"2018) and GLUE (Wang et al., 2018), which track",
"We then subject the strengthened model to the",
"model performance on multiple tasks and provide",
"same procedure and collect weaknesses over sev",
"a unified platform for analysis.",
"eral rounds.After each round, we train a new",
"With the rapid pace of advancement in AI, how-",
"model and set aside a new test set. The process",
"ever, NLU benchmarks struggle to keep up with",
"canbeiterativelyrepeatedinanever-endinglearn-",
"model improvement. Whereas it took around 15",
"ing (Mitchell et al., 2018) setting, with the model",
"years to achieve “near-human performance\" on",
"getting stronger and the test set getting harder in",
"MNIST (LeCun et al., 1998; Ciresan et al., 2012;",
"each new round. Thus, not only is the resultant",
"Wan et al., 2013) and approximately 7 years to",
"dataset harder than existing benchmarks, but this",
"surpass humans on ImageNet (Deng et al., 2009;",
"process also yields a “moving post\" dynamic target",
"Russakovsky et al., 2015; He et al., 2016), the",
"for NLU systems, rather than a static benchmark",
"GLUE benchmark did not last as long as we would",
"that will eventually saturate.",
"have hoped after the advent of BERT (Devlin et al.,",
"Our approach draws inspiration from recent ef-"
] |
{
"arxiv_id": "1910.14599",
"language": "en",
"page_idx": 0
}
|
||
1910.14599-en-3
|
1910.14599-en-3.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{Dataset}$</th>
<th rowspan="2">Genre</th>
<th rowspan="2">$\mathbf{Context}$</th>
<th rowspan="2">$\mathbf{Train/Dev/Test}$</th>
<th colspan="2">$\mathbf{Modele}$ error rate</th>
<th>Tries</th>
<th>Time $(\mathbf{sec}.$</th>
</tr>
<tr>
<th>${\mathrm{Unverified}}$</th>
<th>Verified</th>
<th colspan="2">mean/median per verified ex.</th>
</tr>
<tr>
<td>Al</td>
<td>Wiki</td>
<td>2,080</td>
<td>16,946/1,000/1,000</td>
<td>$29.68\%$</td>
<td>$18.33\%$</td>
<td>3.4/2.0</td>
<td>199.2/125.2</td>
</tr>
<tr>
<td>$A2$</td>
<td>Wiki</td>
<td>2,694</td>
<td>45,460/1,000/1,000</td>
<td>$16.59\%$</td>
<td>$8.07\%$</td>
<td>6.4/4.0</td>
<td>355.3/189.1</td>
</tr>
<tr>
<td rowspan="2">$A3$</td>
<td>Various</td>
<td>6,002</td>
<td>100,459/1,200/1,200</td>
<td>$17.47\%$</td>
<td>$8.60\%$</td>
<td>6.4/4.0</td>
<td>284.0/157.0</td>
</tr>
<tr>
<td>$(\mathbf{Wiki}$ i subset)</td>
<td>1,000</td>
<td>19,920/200/200</td>
<td>$14.79\%$</td>
<td>$6.92\%$</td>
<td>7.4/5.0</td>
<td>337.3/189.6</td>
</tr>
<tr>
<td>$ANLI$</td>
<td>Various</td>
<td>10,776</td>
<td>162,865/3,200/3,200</td>
<td>$18.54\%$</td>
<td>$9.52\%$</td>
<td>5.7/3.0</td>
<td>282.9/156.3</td>
</tr>
</tbody>
</table>
Table 2: Dataset statistics: "Model error rate’is the percentage of examples that the model got wrong; `unverified’
is the overall percentage, while `verified’ is the percentage that was verified by at least 2 human annotators.
selected the model with the best performance on the A1 development set. Then, using the hyperparameters selected from this search, we created a final set of models by training several models with different random seeds. During annotation, we constructed an ensemble by randomly picking a model from the model set as the adversary each turn. This helps us avoid annotators exploiting vulnerabilities in one single model. A new non-overlapping set of contexts was again constructed from Wikipedia via HotpotQA using the same method as Round 1.
## 2.5 Round 3
namely that its contexts (or premises) are very short, because they were selected from the image captioning domain. We believe longer contexts should naturally lead to harder examples, and so we constructed ANLI contexts from longer, multisentence source material.
Following previous observations that models might exploit spurious biases in NLI hypotheses, (Gururangan et al., 2018; Poliak et al., 2018), we conduct a study of the performance of hypothesisonly models on our dataset. We show that such models perform poorly on our test sets.
With respect to data generation with naïve annotators, Geva et al. (2019) noted that models can pick up on annotator bias, modelling annotator artefacts rather than the intended reasoning phenomenon. To counter this, we selected a subset of annotators (i.e., the“exclusive” workers) whose data would only be included in the test set. This enables us to avoid overfitting to the writing style biases of particular annotators, and also to determine how much individual annotator bias is present for the main portion of the data. Examples from each round of dataset collection are provided in Table 1.
Furthermore, our dataset poses new challenges to the community that were less relevant for previous work, such as: can we improve performance online without having to train a new model from scratch every round, how can we overcome catastrophic forgetting, how do we deal with mixed model biases, etc. Because the training set includes examples that the model got right but were not verified, learning from noisy and potentially unverified data becomes an additional interesting challenge.
For the third round, we selected a more diverse set of contexts, in order to explore robustness under domain transfer. In addition to contexts from Wikipedia for Round 3, we also included contexts from the following domains: News (extracted from Common Crawl), fiction (extracted from StoryCloze (Mostafazadeh et al., 2016) and CBT (Hill et al., 2015)), formal spoken text (excerpted from court and presidential debate transcripts in the Manually Annotated Sub-Corpus (MASC) of the Open American National Corpus$^{3}$), and causal or procedural text, which describes sequences of events or actions, extracted from WikiHow. Finally, we also collected annotations using the longer contexts present in the GLUE RTE training data, which came from the RTE5 dataset (Bentivogli et al., 2009). We trained an even stronger RoBERTa ensemble by adding the training set from the second round (A2) to the training data.
# 2.6 Comparing with other datasets
### 3 $\textbf{Dataset statistics}$
The ANLI dataset, comprising three rounds, improves upon previous work in several ways. First, and most obviously, the dataset is collected to be more difficult than previous datasets, by design. Second, it remedies a problem with SNLI,
The dataset statistics can be found in Table 2. The number of examples we collected increases per round, starting with approximately 19k examples for Round 1, to around 47k examples for Round 2,
$^{3}$anc.org/data/masc/corpus/
|
[
[
0.1427001953125,
0.07598876953125,
0.2032470703125,
0.07598876953125,
0.2032470703125,
0.09027099609375,
0.1427001953125,
0.09027099609375
],
[
0.2376708984375,
0.07781982421875,
0.2822265625,
0.07781982421875,
0.2822265625,
0.08905029296875,
0.2376708984375,
0.08905029296875
],
[
0.318115234375,
0.07781982421875,
0.37451171875,
0.07781982421875,
0.37451171875,
0.08905029296875,
0.318115234375,
0.08905029296875
],
[
0.420654296875,
0.07781982421875,
0.5322265625,
0.07781982421875,
0.5322265625,
0.08905029296875,
0.420654296875,
0.08905029296875
],
[
0.56103515625,
0.07781982421875,
0.67529296875,
0.07781982421875,
0.67529296875,
0.09027099609375,
0.56103515625,
0.09027099609375
],
[
0.71728515625,
0.07781982421875,
0.755859375,
0.07781982421875,
0.755859375,
0.08905029296875,
0.71728515625,
0.08905029296875
],
[
0.78076171875,
0.07781982421875,
0.85400390625,
0.07781982421875,
0.85400390625,
0.08905029296875,
0.78076171875,
0.08905029296875
],
[
0.5498046875,
0.08905029296875,
0.61865234375,
0.08905029296875,
0.61865234375,
0.1009521484375,
0.5498046875,
0.1009521484375
],
[
0.62744140625,
0.08905029296875,
0.68408203125,
0.08905029296875,
0.68408203125,
0.1009521484375,
0.62744140625,
0.1009521484375
],
[
0.705078125,
0.08905029296875,
0.84619140625,
0.08905029296875,
0.84619140625,
0.10272216796875,
0.705078125,
0.10272216796875
],
[
0.1435546875,
0.10687255859375,
0.172119140625,
0.10687255859375,
0.172119140625,
0.12054443359375,
0.1435546875,
0.12054443359375
],
[
0.2392578125,
0.10394287109375,
0.279541015625,
0.105712890625,
0.27783203125,
0.1217041015625,
0.2376708984375,
0.119384765625
],
[
0.32568359375,
0.10809326171875,
0.365234375,
0.10809326171875,
0.365234375,
0.119384765625,
0.32568359375,
0.119384765625
],
[
0.397216796875,
0.10687255859375,
0.53564453125,
0.10687255859375,
0.53564453125,
0.12054443359375,
0.397216796875,
0.12054443359375
],
[
0.56494140625,
0.10809326171875,
0.6171875,
0.10809326171875,
0.6171875,
0.119384765625,
0.56494140625,
0.119384765625
],
[
0.634765625,
0.10809326171875,
0.68603515625,
0.10809326171875,
0.68603515625,
0.119384765625,
0.634765625,
0.119384765625
],
[
0.69873046875,
0.10809326171875,
0.755859375,
0.10809326171875,
0.755859375,
0.119384765625,
0.69873046875,
0.119384765625
],
[
0.77099609375,
0.10809326171875,
0.8564453125,
0.10809326171875,
0.8564453125,
0.119384765625,
0.77099609375,
0.119384765625
],
[
0.1435546875,
0.12353515625,
0.175537109375,
0.12353515625,
0.175537109375,
0.14013671875,
0.1435546875,
0.14013671875
],
[
0.56689453125,
0.12469482421875,
0.61865234375,
0.12469482421875,
0.61865234375,
0.1395263671875,
0.56689453125,
0.1395263671875
],
[
0.6396484375,
0.12469482421875,
0.6884765625,
0.12469482421875,
0.6884765625,
0.1395263671875,
0.6396484375,
0.1395263671875
],
[
0.6962890625,
0.12469482421875,
0.75830078125,
0.12469482421875,
0.75830078125,
0.1395263671875,
0.6962890625,
0.1395263671875
],
[
0.76904296875,
0.12469482421875,
0.8564453125,
0.12469482421875,
0.8564453125,
0.1395263671875,
0.76904296875,
0.1395263671875
],
[
0.32666015625,
0.1258544921875,
0.365234375,
0.1258544921875,
0.365234375,
0.1383056640625,
0.32666015625,
0.1383056640625
],
[
0.2418212890625,
0.1258544921875,
0.27783203125,
0.1258544921875,
0.27783203125,
0.1383056640625,
0.2418212890625,
0.1383056640625
],
[
0.397216796875,
0.1258544921875,
0.53564453125,
0.1258544921875,
0.53564453125,
0.1395263671875,
0.397216796875,
0.1395263671875
],
[
0.2333984375,
0.142578125,
0.287109375,
0.142578125,
0.287109375,
0.15673828125,
0.2333984375,
0.15673828125
],
[
0.391357421875,
0.1436767578125,
0.5341796875,
0.1436767578125,
0.5341796875,
0.15673828125,
0.391357421875,
0.15673828125
],
[
0.56689453125,
0.142578125,
0.6171875,
0.142578125,
0.6171875,
0.15673828125,
0.56689453125,
0.15673828125
],
[
0.640625,
0.144287109375,
0.68603515625,
0.144287109375,
0.68603515625,
0.1561279296875,
0.640625,
0.1561279296875
],
[
0.69873046875,
0.144287109375,
0.755859375,
0.144287109375,
0.755859375,
0.1561279296875,
0.69873046875,
0.1561279296875
],
[
0.77099609375,
0.144287109375,
0.85546875,
0.144287109375,
0.85546875,
0.1561279296875,
0.77099609375,
0.1561279296875
],
[
0.32568359375,
0.144287109375,
0.365234375,
0.144287109375,
0.365234375,
0.1561279296875,
0.32568359375,
0.1561279296875
],
[
0.145263671875,
0.149658203125,
0.172119140625,
0.149658203125,
0.172119140625,
0.163330078125,
0.145263671875,
0.163330078125
],
[
0.32568359375,
0.155029296875,
0.366943359375,
0.155029296875,
0.366943359375,
0.17041015625,
0.32568359375,
0.17041015625
],
[
0.56689453125,
0.1561279296875,
0.6171875,
0.1561279296875,
0.6171875,
0.17041015625,
0.56689453125,
0.17041015625
],
[
0.420654296875,
0.15673828125,
0.5322265625,
0.15673828125,
0.5322265625,
0.168701171875,
0.420654296875,
0.168701171875
],
[
0.21826171875,
0.15673828125,
0.301513671875,
0.15673828125,
0.301513671875,
0.168701171875,
0.21826171875,
0.168701171875
],
[
0.640625,
0.15673828125,
0.68408203125,
0.15673828125,
0.68408203125,
0.168701171875,
0.640625,
0.168701171875
],
[
0.69873046875,
0.15673828125,
0.755859375,
0.15673828125,
0.755859375,
0.168701171875,
0.69873046875,
0.168701171875
],
[
0.77099609375,
0.15673828125,
0.85400390625,
0.15673828125,
0.85400390625,
0.168701171875,
0.77099609375,
0.168701171875
],
[
0.2333984375,
0.17333984375,
0.287109375,
0.17333984375,
0.287109375,
0.188232421875,
0.2333984375,
0.188232421875
],
[
0.32080078125,
0.17333984375,
0.3720703125,
0.17333984375,
0.3720703125,
0.188232421875,
0.32080078125,
0.188232421875
],
[
0.387939453125,
0.17333984375,
0.53564453125,
0.17333984375,
0.53564453125,
0.18701171875,
0.387939453125,
0.18701171875
],
[
0.63818359375,
0.1728515625,
0.68701171875,
0.1728515625,
0.68701171875,
0.18701171875,
0.63818359375,
0.18701171875
],
[
0.568359375,
0.174560546875,
0.6171875,
0.174560546875,
0.6171875,
0.1859130859375,
0.568359375,
0.1859130859375
],
[
0.1435546875,
0.174560546875,
0.189697265625,
0.174560546875,
0.189697265625,
0.1859130859375,
0.1435546875,
0.1859130859375
],
[
0.69873046875,
0.174560546875,
0.755859375,
0.174560546875,
0.755859375,
0.1859130859375,
0.69873046875,
0.1859130859375
],
[
0.76904296875,
0.174560546875,
0.8564453125,
0.174560546875,
0.8564453125,
0.1859130859375,
0.76904296875,
0.1859130859375
],
[
0.118408203125,
0.201904296875,
0.88330078125,
0.201904296875,
0.88330078125,
0.215576171875,
0.118408203125,
0.215576171875
],
[
0.1175537109375,
0.2161865234375,
0.84033203125,
0.215576171875,
0.84033203125,
0.2298583984375,
0.1175537109375,
0.23095703125
],
[
0.118408203125,
0.257080078125,
0.490234375,
0.257080078125,
0.490234375,
0.270751953125,
0.118408203125,
0.270751953125
],
[
0.51123046875,
0.255859375,
0.88330078125,
0.257080078125,
0.88330078125,
0.271484375,
0.51123046875,
0.270751953125
],
[
0.118408203125,
0.273681640625,
0.49365234375,
0.273681640625,
0.49365234375,
0.287353515625,
0.118408203125,
0.287353515625
],
[
0.51318359375,
0.273681640625,
0.88330078125,
0.273681640625,
0.88330078125,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.118408203125,
0.289306640625,
0.490234375,
0.289306640625,
0.490234375,
0.302734375,
0.118408203125,
0.302734375
],
[
0.51123046875,
0.289306640625,
0.884765625,
0.2880859375,
0.884765625,
0.302734375,
0.51123046875,
0.303955078125
],
[
0.118408203125,
0.305908203125,
0.490234375,
0.305908203125,
0.490234375,
0.319580078125,
0.118408203125,
0.319580078125
],
[
0.51318359375,
0.305908203125,
0.884765625,
0.305908203125,
0.884765625,
0.319580078125,
0.51318359375,
0.319580078125
],
[
0.1175537109375,
0.320556640625,
0.491943359375,
0.320556640625,
0.491943359375,
0.334228515625,
0.1175537109375,
0.334228515625
],
[
0.51123046875,
0.32177734375,
0.88671875,
0.32177734375,
0.88671875,
0.3349609375,
0.51123046875,
0.3349609375
],
[
0.1175537109375,
0.33740234375,
0.490234375,
0.336181640625,
0.490234375,
0.350830078125,
0.1175537109375,
0.3515625
],
[
0.51123046875,
0.33740234375,
0.705078125,
0.33740234375,
0.705078125,
0.350830078125,
0.51123046875,
0.350830078125
],
[
0.118408203125,
0.352783203125,
0.489501953125,
0.3515625,
0.489501953125,
0.366455078125,
0.118408203125,
0.36767578125
],
[
0.53076171875,
0.35400390625,
0.884765625,
0.35400390625,
0.884765625,
0.3681640625,
0.53076171875,
0.3681640625
],
[
0.118408203125,
0.369384765625,
0.489501953125,
0.369384765625,
0.489501953125,
0.383056640625,
0.118408203125,
0.383056640625
],
[
0.51318359375,
0.37060546875,
0.88671875,
0.37060546875,
0.88671875,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.118408203125,
0.385986328125,
0.490234375,
0.385986328125,
0.490234375,
0.399658203125,
0.118408203125,
0.399658203125
],
[
0.51123046875,
0.385986328125,
0.884765625,
0.385498046875,
0.884765625,
0.399658203125,
0.51123046875,
0.40087890625
],
[
0.118408203125,
0.402099609375,
0.490234375,
0.402099609375,
0.490234375,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.51318359375,
0.402587890625,
0.88330078125,
0.402587890625,
0.88330078125,
0.416259765625,
0.51318359375,
0.416259765625
],
[
0.118408203125,
0.418701171875,
0.46337890625,
0.418701171875,
0.46337890625,
0.431640625,
0.118408203125,
0.431640625
],
[
0.51318359375,
0.418701171875,
0.88330078125,
0.418701171875,
0.88330078125,
0.431640625,
0.51318359375,
0.431640625
],
[
0.51318359375,
0.435302734375,
0.8037109375,
0.435302734375,
0.8037109375,
0.448974609375,
0.51318359375,
0.448974609375
],
[
0.12005615234375,
0.44482421875,
0.231689453125,
0.44482421875,
0.231689453125,
0.458984375,
0.12005615234375,
0.458984375
],
[
0.53076171875,
0.449462890625,
0.884765625,
0.45068359375,
0.884765625,
0.465576171875,
0.53076171875,
0.46435546875
],
[
0.118408203125,
0.46728515625,
0.489501953125,
0.46728515625,
0.489501953125,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.51318359375,
0.46728515625,
0.88330078125,
0.46728515625,
0.88330078125,
0.48095703125,
0.51318359375,
0.48095703125
],
[
0.118408203125,
0.48388671875,
0.490234375,
0.48388671875,
0.490234375,
0.49755859375,
0.118408203125,
0.49755859375
],
[
0.5146484375,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.495361328125,
0.5146484375,
0.495361328125
],
[
0.12005615234375,
0.49951171875,
0.490234375,
0.49951171875,
0.490234375,
0.51318359375,
0.12005615234375,
0.51318359375
],
[
0.51318359375,
0.49951171875,
0.88671875,
0.49951171875,
0.88671875,
0.51318359375,
0.51318359375,
0.51318359375
],
[
0.12005615234375,
0.515625,
0.491943359375,
0.515625,
0.491943359375,
0.529296875,
0.12005615234375,
0.529296875
],
[
0.51318359375,
0.515625,
0.884765625,
0.515625,
0.884765625,
0.529296875,
0.51318359375,
0.529296875
],
[
0.118408203125,
0.5322265625,
0.490234375,
0.5322265625,
0.490234375,
0.5458984375,
0.118408203125,
0.5458984375
],
[
0.51318359375,
0.5322265625,
0.88330078125,
0.5322265625,
0.88330078125,
0.5458984375,
0.51318359375,
0.5458984375
],
[
0.12005615234375,
0.54736328125,
0.49365234375,
0.54736328125,
0.49365234375,
0.56103515625,
0.12005615234375,
0.56103515625
],
[
0.509765625,
0.548828125,
0.88330078125,
0.548828125,
0.88330078125,
0.5625,
0.509765625,
0.5625
],
[
0.1175537109375,
0.5634765625,
0.489501953125,
0.5625,
0.489501953125,
0.57666015625,
0.1175537109375,
0.57763671875
],
[
0.51123046875,
0.5634765625,
0.884765625,
0.56396484375,
0.884765625,
0.5791015625,
0.51123046875,
0.57763671875
],
[
0.1175537109375,
0.5791015625,
0.490234375,
0.580078125,
0.490234375,
0.59423828125,
0.1175537109375,
0.59326171875
],
[
0.51318359375,
0.58056640625,
0.884765625,
0.58056640625,
0.884765625,
0.59423828125,
0.51318359375,
0.59423828125
],
[
0.118408203125,
0.5966796875,
0.495361328125,
0.5966796875,
0.495361328125,
0.60986328125,
0.118408203125,
0.60986328125
],
[
0.51318359375,
0.5966796875,
0.884765625,
0.5966796875,
0.884765625,
0.60986328125,
0.51318359375,
0.60986328125
],
[
0.1175537109375,
0.61083984375,
0.490234375,
0.6123046875,
0.490234375,
0.626953125,
0.1175537109375,
0.6259765625
],
[
0.51123046875,
0.6123046875,
0.884765625,
0.61083984375,
0.884765625,
0.6259765625,
0.51123046875,
0.626953125
],
[
0.12005615234375,
0.62890625,
0.49365234375,
0.62890625,
0.49365234375,
0.642578125,
0.12005615234375,
0.642578125
],
[
0.51123046875,
0.62744140625,
0.8212890625,
0.626953125,
0.8212890625,
0.64111328125,
0.51123046875,
0.642578125
],
[
0.1175537109375,
0.64453125,
0.489501953125,
0.64453125,
0.489501953125,
0.65771484375,
0.1175537109375,
0.65771484375
],
[
0.5322265625,
0.6455078125,
0.884765625,
0.6455078125,
0.884765625,
0.6591796875,
0.5322265625,
0.6591796875
],
[
0.1175537109375,
0.6591796875,
0.490234375,
0.66015625,
0.490234375,
0.6748046875,
0.1175537109375,
0.67333984375
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.1175537109375,
0.67578125,
0.48779296875,
0.67578125,
0.48779296875,
0.689453125,
0.1175537109375,
0.689453125
],
[
0.51318359375,
0.677734375,
0.884765625,
0.677734375,
0.884765625,
0.69140625,
0.51318359375,
0.69140625
],
[
0.118408203125,
0.69140625,
0.48779296875,
0.69140625,
0.48779296875,
0.7060546875,
0.118408203125,
0.7060546875
],
[
0.5146484375,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.118408203125,
0.7080078125,
0.49365234375,
0.7080078125,
0.49365234375,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.51123046875,
0.708984375,
0.8818359375,
0.708984375,
0.8818359375,
0.72265625,
0.51123046875,
0.72265625
],
[
0.12005615234375,
0.724609375,
0.49365234375,
0.724609375,
0.49365234375,
0.73828125,
0.12005615234375,
0.73828125
],
[
0.51123046875,
0.7255859375,
0.884765625,
0.724609375,
0.884765625,
0.7392578125,
0.51123046875,
0.74072265625
],
[
0.118408203125,
0.7412109375,
0.490234375,
0.7412109375,
0.490234375,
0.7548828125,
0.118408203125,
0.7548828125
],
[
0.51318359375,
0.7421875,
0.884765625,
0.7421875,
0.884765625,
0.755859375,
0.51318359375,
0.755859375
],
[
0.1175537109375,
0.75830078125,
0.35107421875,
0.75830078125,
0.35107421875,
0.771484375,
0.1175537109375,
0.771484375
],
[
0.5146484375,
0.75830078125,
0.884765625,
0.75830078125,
0.884765625,
0.771484375,
0.5146484375,
0.771484375
],
[
0.51318359375,
0.7724609375,
0.88330078125,
0.77392578125,
0.88330078125,
0.7880859375,
0.51318359375,
0.78759765625
],
[
0.118408203125,
0.78369140625,
0.4072265625,
0.78369140625,
0.4072265625,
0.798828125,
0.118408203125,
0.798828125
],
[
0.51123046875,
0.7890625,
0.875,
0.7890625,
0.875,
0.802734375,
0.51123046875,
0.802734375
],
[
0.1175537109375,
0.8046875,
0.490234375,
0.8056640625,
0.490234375,
0.82080078125,
0.1175537109375,
0.8193359375
],
[
0.1175537109375,
0.822265625,
0.49365234375,
0.82177734375,
0.49365234375,
0.8359375,
0.1175537109375,
0.83740234375
],
[
0.515625,
0.822265625,
0.54248046875,
0.822265625,
0.54248046875,
0.83056640625,
0.515625,
0.83056640625
],
[
0.5380859375,
0.8193359375,
0.69189453125,
0.8193359375,
0.69189453125,
0.8330078125,
0.5380859375,
0.8330078125
],
[
0.118408203125,
0.8388671875,
0.490234375,
0.8388671875,
0.490234375,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.51318359375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.85498046875,
0.490234375,
0.85498046875,
0.490234375,
0.8681640625,
0.118408203125,
0.8681640625
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87060546875,
0.490234375,
0.87060546875,
0.490234375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.14111328125,
0.8955078125,
0.37451171875,
0.8955078125,
0.37451171875,
0.9091796875,
0.14111328125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Dataset",
"Genre",
"Context",
"Train/Dev/Test",
"Model error rate",
"Tries",
"Time (sec.)",
"Unverified",
"Verified",
"mean/median per verified ex.",
"A1",
"Wiki",
"2,080",
"16,946/1,000 /1,000",
"29.68%",
"18.33%",
"3.4/2.0",
"199.2/125.2",
"A2",
"16.59%",
"8.07%",
"6.4/4.0",
"355.3/189.1",
"2,694",
"Wiki",
"45,460/1,000/1,000",
"Various",
"100,459 /1,200 /1,200",
"17.47%",
"8.60%",
"6.4/4.0",
"284.0/157.0",
"6,002",
"A3",
"1,000",
"14.79%",
"19,920/200/200",
"(Wiki subset)",
"6.92%",
"7.4/5.0",
"337.3/189.6",
"Various",
"10.776",
"162,865/3,200/3,200",
"9.52%",
"18.54%",
"ANLI",
"5.7/3.0",
"282.9/156.3",
"Table 2: Dataset statistics: Model error rate' is the percentage of examples that the model got wrong; unverified\"",
"is the overall percentage, while 'verified' is the percentage that was verified by at least 2 human annotators.",
"selected the model with the best performance on",
"namely that its contexts (or premises) are very",
"the A1 development set. Then, using the hyperpa-",
"short, because they were selected from the image",
"rameters selected from this search, we created a",
"captioning domain.We believe longer contexts",
"final set of models by training several models with",
"should naturally lead to harder examples, and so",
"different random seeds. During annotation, we con-",
"we constructed ANLI contexts from longer, multi-",
"structed an ensemble by randomly picking a model",
" sentence source material.",
"from the model set as the adversary each turn. This",
"Following previous observations that models",
"helps us avoid annotators exploiting vulnerabilities",
"might exploit spurious biases in NLI hypotheses,",
"in one single model. A new non-overlapping set of",
"(Gururangan et al., 2018; Poliak et al., 2018), we",
"contexts was again constructed from Wikipedia via",
"conduct a study of the performance of hypothesis-",
"HotpotQA using the same method as Round 1.",
"only models on our dataset. We show that such",
"models perform poorly on our test sets.",
"2.5Round3",
"With respect to data generation with naive anno-",
"For the third round, we selected a more diverse",
"tators, Geva et al. (2019) noted that models can pick",
"set of contexts, in order to explore robustness un-",
"uponannotatorbias,modellingannotatorartefacts",
"der domain transfer. In addition to contexts from",
"rather than the intended reasoning phenomenon.",
"Wikipedia for Round 3, we also included con-",
"To counter this, we selected a subset of annotators",
"texts from the following domains: News (extracted",
"(i.e., the “\"exclusive\" workers) whose data would",
"from Common Crawl),fiction (extracted from Sto-",
" only be included in the test set. This enables us to",
"ryCloze (Mostafazadeh et al., 2016) and CBT (Hill",
"avoid overfitting to the writing style biases of par-",
"et al., 2015)), formal spoken text (excerpted from",
"ticular annotators, and also to determine how much",
"court and presidential debate transcripts in the Man-",
"individual annotator bias is present for the main",
"ually Annotated Sub-Corpus (MASC) of the Open",
"portion of the data. Examples from each round of",
"American National Corpus'), and causal or pro-",
"dataset collection are provided in Table 1.",
"cedural text, which describes sequences of events",
"Furthermore, our dataset poses new challenges",
"or actions,extracted from WikiHow. Finally,we",
"to the community that were less relevant for previ",
"also collected annotations using the longer contexts",
"ous work, such as: can we improve performance",
"present in the GLUE RTE training data, which",
"onlinewithouthavingtotrainanewmodelfrom",
"came from the RTE5 dataset (Bentivogli et al.,",
"scratch every round, how can we overcome catas.",
"2009). We trained an even stronger RoBERTa en-",
"trophic forgetting, how do we deal with mixed",
"semble by adding the training set from the second",
"model biases, etc. Because the training set includes",
"round (A2) to the training data.",
"examples that the model got right but were not veri-",
"fied, learning from noisy and potentially unverified",
"2.6Comparing with other datasets",
"data becomes an additional interesting challenge.",
"The ANLI dataset, comprising three rounds, im-",
"proves upon previous work in several ways. First,",
"3",
"Dataset statistics",
"and most obviously, the dataset is collected to",
"The dataset statistics can be found in Table 2. The",
"be more difficult than previous datasets, by de-",
"number of examples we collected increases per",
"sign. Second, it remedies a problem with SNLI,",
"round, starting with approximately 19k examples",
"3anc.org/data/masc/corpus/",
"for Round 1, to around 47k examples for Round 2."
] |
{
"arxiv_id": "1910.14599",
"language": "en",
"page_idx": 3
}
|
||
1910.14599-en-4
|
1910.14599-en-4.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Model}$</th>
<th>Training Data</th>
<th>Al</th>
<th>$A2$</th>
<th>$A3$</th>
<th>$ANLI$</th>
<th>ANLI- E</th>
<th>$SNLI$</th>
<th>MNLI- m/ - mm</th>
</tr>
<tr>
<td rowspan="5">$\mathbf{BERT}$</td>
<td>$S.M^{\star1}$</td>
<td>00.0</td>
<td>28.9</td>
<td>28.8</td>
<td>19.8</td>
<td>19.9</td>
<td>91.3</td>
<td>86.7/86.4</td>
</tr>
<tr>
<td>+Al</td>
<td>44.2</td>
<td>32.6</td>
<td>29.3</td>
<td>35.0</td>
<td>34.2</td>
<td>91.3</td>
<td>86.3/86.5</td>
</tr>
<tr>
<td>+Al+A2</td>
<td>57.3</td>
<td>45.2</td>
<td>33.4</td>
<td>44.6</td>
<td>43.2</td>
<td>90.9</td>
<td>86.3/86.3</td>
</tr>
<tr>
<td>+Al+A2+A3</td>
<td>57.2</td>
<td>49.0</td>
<td>46.1</td>
<td>50.5</td>
<td>46.3</td>
<td>90.9</td>
<td>85.6/85.4</td>
</tr>
<tr>
<td>$\mathbf{S,M,F,ANLI}$</td>
<td>57.4</td>
<td>48.3</td>
<td>43.5</td>
<td>49.3</td>
<td>44.2</td>
<td>90.4</td>
<td>86.0/85.8</td>
</tr>
<tr>
<td>XLNet</td>
<td>$\mathbf{S,M,F,ANLI}$</td>
<td>67.6</td>
<td>50.7</td>
<td>48.3</td>
<td>55.1</td>
<td>52.0</td>
<td>91.8</td>
<td>89.6/89.4</td>
</tr>
<tr>
<td rowspan="5">RoBERTa</td>
<td>$S,M$</td>
<td>47.6</td>
<td>25.4</td>
<td>22.1</td>
<td>31.1</td>
<td>31.4</td>
<td>92.6</td>
<td>90.8/90.6</td>
</tr>
<tr>
<td>$+F$</td>
<td>54.0</td>
<td>24.2</td>
<td>22.4</td>
<td>32.8</td>
<td>33.7</td>
<td>92.7</td>
<td>90.6/90.5</td>
</tr>
<tr>
<td>$+\mathrm{F+Al}^{\star2}$</td>
<td>68.7</td>
<td>19.3</td>
<td>22.0</td>
<td>35.8</td>
<td>36.8</td>
<td>92.8</td>
<td>90.9/90.7</td>
</tr>
<tr>
<td>$+\mathrm{F+Al+A2}^{\star3}$</td>
<td>71.2</td>
<td>44.3</td>
<td>20.4</td>
<td>43.7</td>
<td>41.4</td>
<td>92.9</td>
<td>91.0/90.7</td>
</tr>
<tr>
<td>$\mathbf{S,M,F,ANLI}$</td>
<td>73.8</td>
<td>48.9</td>
<td>44.4</td>
<td>53.7</td>
<td>49.7</td>
<td>92.6</td>
<td>91.0/90.6</td>
</tr>
</tbody>
</table>
Table 3: Model Performance. 'S’refers to SNLI, "M’to MNLI dev (-m=matched, -mm=mismatched), and ‘F’ to FEVER;“A1–A3’refer to the rounds respectively and‘ANL’ refers to A1+A2+A3,‘-E’refers to test set examples written by annotators exclusive to the test set. Datasets marked $`\star n$, were used to train the base model for round $n$, and their performance on that round is underlined (A2 and A3 used ensembles, and hence have non-zero scores).
tries is exceeded), and the average time this took (in seconds). Again, these metrics are useful for evaluating model quality: observe that the average number of tries and average time per verified error both go up with later rounds. This demonstrates that the rounds are getting increasingly more difficult. Further dataset statistics and inter-annotator agreement are reported in Appendix C.
### 4 Results
Table 3 reports the main results. In addition to BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019b), we also include XLNet (Yang et al., 2019) as an example of a strong, but different, model architecture. We show test set performance on the ANLI test sets per round, the total ANLI test set, and the exclusive test subset (examples from test-set-exclusive workers). We also show accuracy on the SNLI test set and the MNLI development set (for the purpose of comparing between different model configurations across table rows). In what follows, we discuss our observations.
to over 103k examples for Round 3. We collected more data for later rounds not only because that data is likely to be more interesting, but also simply because the base model is better and so annotation took longer to collect good, verified correct examples of model vulnerabilities.
For each round, we report the model error rate, both on verified and unverified examples. The unverified model error rate captures the percentage of examples where the model disagreed with the writer's target label, but where we are not (yet) sure if the example is correct. The verified model error rate is the percentage of model errors from example pairs that other annotators confirmed the correct label for. Note that error rate is a useful way to evaluate model quality: the lower the model error rate\_assuming constant annotator quality and context-difficulty—the better the model.
We observe that model error rates decrease as we progress through rounds. In Round 3, where we included a more diverse range of contexts from various domains, the overall error rate went slightly up compared to the preceding round, but for Wikipedia contexts the error rate decreased substantially. While for the first round roughly 1 in every 5 examples were verified model errors, this quickly dropped over consecutive rounds, and the overall model error rate is less than 1 in 10. On the one hand, this is impressive, and shows how far we have come with just three rounds. On the other hand, it shows that we still have a long way to go if even untrained annotators can fool ensembles of state-of-the-art models with relative ease.
Table 2 also reports the average number of “tries", i.e., attempts made for each context until a model error was found (or the number of possible
Base model performance is low. Notice that the base model for each round performs very poorly on that round's test set. This is the expected outcome: For round 1, the base model gets the entire test set wrong, by design. For rounds 2 and 3, we used an ensemble, so performance is not necessarily zero. However, as it turns out, performance still falls well below chance$^{4}$, indicating that workers did not find vulnerabilities specific to a single model, but generally applicable ones for that model class.
$^{4}$Chance is at 33%, since the test set labels are balanced
|
[
[
0.444091796875,
0.0736083984375,
0.47021484375,
0.07781982421875,
0.465087890625,
0.0926513671875,
0.43994140625,
0.0885009765625
],
[
0.490234375,
0.07659912109375,
0.5146484375,
0.07659912109375,
0.5146484375,
0.09027099609375,
0.490234375,
0.09027099609375
],
[
0.280517578125,
0.07781982421875,
0.372802734375,
0.07781982421875,
0.372802734375,
0.08905029296875,
0.280517578125,
0.08905029296875
],
[
0.1956787109375,
0.07781982421875,
0.2435302734375,
0.07781982421875,
0.2435302734375,
0.08905029296875,
0.1956787109375,
0.08905029296875
],
[
0.397216796875,
0.07781982421875,
0.4189453125,
0.07781982421875,
0.4189453125,
0.09027099609375,
0.397216796875,
0.09027099609375
],
[
0.52880859375,
0.07781982421875,
0.5693359375,
0.07781982421875,
0.5693359375,
0.08905029296875,
0.52880859375,
0.08905029296875
],
[
0.5859375,
0.07781982421875,
0.69287109375,
0.07781982421875,
0.69287109375,
0.08905029296875,
0.5859375,
0.08905029296875
],
[
0.70947265625,
0.07781982421875,
0.80029296875,
0.07781982421875,
0.80029296875,
0.08905029296875,
0.70947265625,
0.08905029296875
],
[
0.273681640625,
0.0955810546875,
0.31982421875,
0.0914306640625,
0.322509765625,
0.10687255859375,
0.276123046875,
0.11102294921875
],
[
0.387939453125,
0.0955810546875,
0.42236328125,
0.0955810546875,
0.42236328125,
0.11102294921875,
0.387939453125,
0.11102294921875
],
[
0.435791015625,
0.0968017578125,
0.466796875,
0.0968017578125,
0.466796875,
0.1092529296875,
0.435791015625,
0.1092529296875
],
[
0.48193359375,
0.0968017578125,
0.51318359375,
0.0968017578125,
0.51318359375,
0.1092529296875,
0.48193359375,
0.1092529296875
],
[
0.5400390625,
0.0968017578125,
0.5693359375,
0.0968017578125,
0.5693359375,
0.1092529296875,
0.5400390625,
0.1092529296875
],
[
0.611328125,
0.0968017578125,
0.640625,
0.0968017578125,
0.640625,
0.1092529296875,
0.611328125,
0.1092529296875
],
[
0.66162109375,
0.0968017578125,
0.69287109375,
0.0968017578125,
0.69287109375,
0.1092529296875,
0.66162109375,
0.1092529296875
],
[
0.7314453125,
0.097412109375,
0.8017578125,
0.097412109375,
0.8017578125,
0.1092529296875,
0.7314453125,
0.1092529296875
],
[
0.435791015625,
0.10809326171875,
0.468505859375,
0.10809326171875,
0.468505859375,
0.12054443359375,
0.435791015625,
0.12054443359375
],
[
0.480224609375,
0.10809326171875,
0.51318359375,
0.10809326171875,
0.51318359375,
0.12054443359375,
0.480224609375,
0.12054443359375
],
[
0.5380859375,
0.10809326171875,
0.5693359375,
0.10809326171875,
0.5693359375,
0.12054443359375,
0.5380859375,
0.12054443359375
],
[
0.609375,
0.10809326171875,
0.6396484375,
0.10809326171875,
0.6396484375,
0.12054443359375,
0.609375,
0.12054443359375
],
[
0.66162109375,
0.10809326171875,
0.69287109375,
0.10809326171875,
0.69287109375,
0.12054443359375,
0.66162109375,
0.12054443359375
],
[
0.3896484375,
0.1092529296875,
0.420654296875,
0.1092529296875,
0.420654296875,
0.1217041015625,
0.3896484375,
0.1217041015625
],
[
0.279541015625,
0.1092529296875,
0.310546875,
0.1092529296875,
0.310546875,
0.12054443359375,
0.279541015625,
0.12054443359375
],
[
0.7314453125,
0.1092529296875,
0.80029296875,
0.1092529296875,
0.80029296875,
0.12054443359375,
0.7314453125,
0.12054443359375
],
[
0.387939453125,
0.119384765625,
0.42236328125,
0.119384765625,
0.42236328125,
0.1353759765625,
0.387939453125,
0.1353759765625
],
[
0.48193359375,
0.1181640625,
0.51318359375,
0.12054443359375,
0.51123046875,
0.133056640625,
0.480224609375,
0.1312255859375
],
[
0.53662109375,
0.119384765625,
0.57275390625,
0.119384765625,
0.57275390625,
0.1353759765625,
0.53662109375,
0.1353759765625
],
[
0.435791015625,
0.12054443359375,
0.466796875,
0.12054443359375,
0.466796875,
0.133056640625,
0.435791015625,
0.133056640625
],
[
0.198974609375,
0.12054443359375,
0.240966796875,
0.12054443359375,
0.240966796875,
0.1341552734375,
0.198974609375,
0.1341552734375
],
[
0.609375,
0.12054443359375,
0.640625,
0.12054443359375,
0.640625,
0.133056640625,
0.609375,
0.133056640625
],
[
0.66162109375,
0.12054443359375,
0.69287109375,
0.12054443359375,
0.69287109375,
0.1318359375,
0.66162109375,
0.1318359375
],
[
0.7314453125,
0.1217041015625,
0.80029296875,
0.1217041015625,
0.80029296875,
0.133056640625,
0.7314453125,
0.133056640625
],
[
0.5380859375,
0.1312255859375,
0.5693359375,
0.1312255859375,
0.5693359375,
0.1436767578125,
0.5380859375,
0.1436767578125
],
[
0.279541015625,
0.1217041015625,
0.337646484375,
0.1217041015625,
0.337646484375,
0.133056640625,
0.279541015625,
0.133056640625
],
[
0.279541015625,
0.1318359375,
0.36279296875,
0.1312255859375,
0.36279296875,
0.142578125,
0.279541015625,
0.1436767578125
],
[
0.3896484375,
0.1318359375,
0.420654296875,
0.1318359375,
0.420654296875,
0.1436767578125,
0.3896484375,
0.1436767578125
],
[
0.435791015625,
0.1318359375,
0.468505859375,
0.1318359375,
0.468505859375,
0.1436767578125,
0.435791015625,
0.1436767578125
],
[
0.480224609375,
0.1318359375,
0.51318359375,
0.1318359375,
0.51318359375,
0.1436767578125,
0.480224609375,
0.1436767578125
],
[
0.60791015625,
0.1312255859375,
0.640625,
0.1312255859375,
0.640625,
0.1436767578125,
0.60791015625,
0.1436767578125
],
[
0.66162109375,
0.1318359375,
0.6943359375,
0.1318359375,
0.6943359375,
0.1436767578125,
0.66162109375,
0.1436767578125
],
[
0.7314453125,
0.1318359375,
0.8017578125,
0.1318359375,
0.8017578125,
0.1436767578125,
0.7314453125,
0.1436767578125
],
[
0.387939453125,
0.141357421875,
0.4248046875,
0.144287109375,
0.42236328125,
0.1591796875,
0.385498046875,
0.1561279296875
],
[
0.480224609375,
0.142578125,
0.5146484375,
0.142578125,
0.5146484375,
0.157958984375,
0.480224609375,
0.157958984375
],
[
0.53662109375,
0.142578125,
0.57080078125,
0.142578125,
0.57080078125,
0.157958984375,
0.53662109375,
0.157958984375
],
[
0.60693359375,
0.142578125,
0.64404296875,
0.142578125,
0.64404296875,
0.157958984375,
0.60693359375,
0.157958984375
],
[
0.66015625,
0.142578125,
0.6943359375,
0.142578125,
0.6943359375,
0.157958984375,
0.66015625,
0.157958984375
],
[
0.73046875,
0.1436767578125,
0.8037109375,
0.1436767578125,
0.8037109375,
0.157958984375,
0.73046875,
0.157958984375
],
[
0.435791015625,
0.1436767578125,
0.466796875,
0.1436767578125,
0.466796875,
0.1561279296875,
0.435791015625,
0.1561279296875
],
[
0.279541015625,
0.144287109375,
0.359375,
0.144287109375,
0.359375,
0.1561279296875,
0.279541015625,
0.1561279296875
],
[
0.1956787109375,
0.160888671875,
0.246826171875,
0.160888671875,
0.246826171875,
0.17578125,
0.1956787109375,
0.17578125
],
[
0.27783203125,
0.160888671875,
0.359375,
0.160888671875,
0.359375,
0.17578125,
0.27783203125,
0.17578125
],
[
0.387939453125,
0.160888671875,
0.424072265625,
0.160888671875,
0.424072265625,
0.177001953125,
0.387939453125,
0.177001953125
],
[
0.478515625,
0.160888671875,
0.5146484375,
0.160888671875,
0.5146484375,
0.177001953125,
0.478515625,
0.177001953125
],
[
0.53662109375,
0.160888671875,
0.57080078125,
0.160888671875,
0.57080078125,
0.177001953125,
0.53662109375,
0.177001953125
],
[
0.6591796875,
0.160888671875,
0.6962890625,
0.160888671875,
0.6962890625,
0.177001953125,
0.6591796875,
0.177001953125
],
[
0.73046875,
0.160888671875,
0.8037109375,
0.160888671875,
0.8037109375,
0.17578125,
0.73046875,
0.17578125
],
[
0.60791015625,
0.162109375,
0.64990234375,
0.162109375,
0.64990234375,
0.17578125,
0.60791015625,
0.17578125
],
[
0.43408203125,
0.162109375,
0.466796875,
0.162109375,
0.466796875,
0.174560546875,
0.43408203125,
0.174560546875
],
[
0.276123046875,
0.1787109375,
0.310546875,
0.1787109375,
0.310546875,
0.19482421875,
0.276123046875,
0.19482421875
],
[
0.38623046875,
0.1787109375,
0.424072265625,
0.1787109375,
0.424072265625,
0.19482421875,
0.38623046875,
0.19482421875
],
[
0.53662109375,
0.1787109375,
0.57080078125,
0.1787109375,
0.57080078125,
0.19482421875,
0.53662109375,
0.19482421875
],
[
0.60693359375,
0.1787109375,
0.64404296875,
0.1787109375,
0.64404296875,
0.19482421875,
0.60693359375,
0.19482421875
],
[
0.66015625,
0.1787109375,
0.6962890625,
0.1787109375,
0.6962890625,
0.19482421875,
0.66015625,
0.19482421875
],
[
0.480224609375,
0.179931640625,
0.51318359375,
0.179931640625,
0.51318359375,
0.1923828125,
0.480224609375,
0.1923828125
],
[
0.435791015625,
0.179931640625,
0.466796875,
0.179931640625,
0.466796875,
0.1923828125,
0.435791015625,
0.1923828125
],
[
0.7314453125,
0.18115234375,
0.8017578125,
0.18115234375,
0.8017578125,
0.1923828125,
0.7314453125,
0.1923828125
],
[
0.48193359375,
0.189453125,
0.51318359375,
0.191162109375,
0.51123046875,
0.204833984375,
0.480224609375,
0.203125
],
[
0.53662109375,
0.190673828125,
0.57080078125,
0.190673828125,
0.57080078125,
0.2060546875,
0.53662109375,
0.2060546875
],
[
0.66162109375,
0.191162109375,
0.69287109375,
0.191162109375,
0.69287109375,
0.2037353515625,
0.66162109375,
0.2037353515625
],
[
0.3896484375,
0.1923828125,
0.42236328125,
0.1923828125,
0.42236328125,
0.204833984375,
0.3896484375,
0.204833984375
],
[
0.27783203125,
0.1923828125,
0.30322265625,
0.1923828125,
0.30322265625,
0.204833984375,
0.27783203125,
0.204833984375
],
[
0.435791015625,
0.1923828125,
0.466796875,
0.1923828125,
0.466796875,
0.204833984375,
0.435791015625,
0.204833984375
],
[
0.60791015625,
0.1923828125,
0.6396484375,
0.1923828125,
0.6396484375,
0.204833984375,
0.60791015625,
0.204833984375
],
[
0.7314453125,
0.1923828125,
0.80029296875,
0.1923828125,
0.80029296875,
0.2037353515625,
0.7314453125,
0.2037353515625
],
[
0.480224609375,
0.20068359375,
0.515625,
0.2037353515625,
0.51318359375,
0.218505859375,
0.476806640625,
0.215576171875
],
[
0.66015625,
0.203125,
0.6962890625,
0.203125,
0.6962890625,
0.218505859375,
0.66015625,
0.218505859375
],
[
0.3896484375,
0.204833984375,
0.420654296875,
0.204833984375,
0.420654296875,
0.21728515625,
0.3896484375,
0.21728515625
],
[
0.197265625,
0.204833984375,
0.2626953125,
0.204833984375,
0.2626953125,
0.2161865234375,
0.197265625,
0.2161865234375
],
[
0.435791015625,
0.204833984375,
0.466796875,
0.204833984375,
0.466796875,
0.21728515625,
0.435791015625,
0.21728515625
],
[
0.5380859375,
0.204833984375,
0.5693359375,
0.204833984375,
0.5693359375,
0.21728515625,
0.5380859375,
0.21728515625
],
[
0.60791015625,
0.204833984375,
0.640625,
0.204833984375,
0.640625,
0.21728515625,
0.60791015625,
0.21728515625
],
[
0.7314453125,
0.204833984375,
0.80029296875,
0.204833984375,
0.80029296875,
0.2161865234375,
0.7314453125,
0.2161865234375
],
[
0.480224609375,
0.2161865234375,
0.5146484375,
0.2161865234375,
0.5146484375,
0.232177734375,
0.480224609375,
0.232177734375
],
[
0.274658203125,
0.2060546875,
0.339111328125,
0.20068359375,
0.341796875,
0.21435546875,
0.27783203125,
0.2197265625
],
[
0.53662109375,
0.2161865234375,
0.57080078125,
0.2161865234375,
0.57080078125,
0.232177734375,
0.53662109375,
0.232177734375
],
[
0.60693359375,
0.2161865234375,
0.64208984375,
0.2161865234375,
0.64208984375,
0.232177734375,
0.60693359375,
0.232177734375
],
[
0.3896484375,
0.21728515625,
0.420654296875,
0.21728515625,
0.420654296875,
0.2298583984375,
0.3896484375,
0.2298583984375
],
[
0.27783203125,
0.21728515625,
0.365234375,
0.21435546875,
0.366943359375,
0.2286376953125,
0.279541015625,
0.232177734375
],
[
0.435791015625,
0.21728515625,
0.466796875,
0.21728515625,
0.466796875,
0.2298583984375,
0.435791015625,
0.2298583984375
],
[
0.66162109375,
0.21728515625,
0.69287109375,
0.21728515625,
0.69287109375,
0.2298583984375,
0.66162109375,
0.2298583984375
],
[
0.7314453125,
0.218505859375,
0.80029296875,
0.218505859375,
0.80029296875,
0.2298583984375,
0.7314453125,
0.2298583984375
],
[
0.66015625,
0.22802734375,
0.6962890625,
0.22802734375,
0.6962890625,
0.243408203125,
0.66015625,
0.243408203125
],
[
0.73046875,
0.2286376953125,
0.8017578125,
0.23095703125,
0.8017578125,
0.2423095703125,
0.73046875,
0.240478515625
],
[
0.3896484375,
0.2298583984375,
0.42236328125,
0.2298583984375,
0.42236328125,
0.2423095703125,
0.3896484375,
0.2423095703125
],
[
0.279541015625,
0.2298583984375,
0.359375,
0.2298583984375,
0.359375,
0.2410888671875,
0.279541015625,
0.2410888671875
],
[
0.435791015625,
0.2298583984375,
0.466796875,
0.2298583984375,
0.466796875,
0.2410888671875,
0.435791015625,
0.2410888671875
],
[
0.480224609375,
0.2298583984375,
0.5146484375,
0.2298583984375,
0.5146484375,
0.2423095703125,
0.480224609375,
0.2423095703125
],
[
0.5380859375,
0.2298583984375,
0.5693359375,
0.2298583984375,
0.5693359375,
0.2423095703125,
0.5380859375,
0.2423095703125
],
[
0.60791015625,
0.2298583984375,
0.64208984375,
0.2298583984375,
0.64208984375,
0.2423095703125,
0.60791015625,
0.2423095703125
],
[
0.1175537109375,
0.25048828125,
0.88330078125,
0.2493896484375,
0.88330078125,
0.26416015625,
0.1175537109375,
0.265380859375
],
[
0.118408203125,
0.26416015625,
0.88330078125,
0.2666015625,
0.88330078125,
0.27978515625,
0.118408203125,
0.27783203125
],
[
0.1158447265625,
0.276611328125,
0.88671875,
0.27783203125,
0.88671875,
0.295654296875,
0.1158447265625,
0.29443359375
],
[
0.118408203125,
0.29443359375,
0.8759765625,
0.29443359375,
0.8759765625,
0.30810546875,
0.118408203125,
0.30810546875
],
[
0.118408203125,
0.32470703125,
0.489501953125,
0.32470703125,
0.489501953125,
0.33837890625,
0.118408203125,
0.33837890625
],
[
0.51318359375,
0.325927734375,
0.88330078125,
0.325927734375,
0.88330078125,
0.339111328125,
0.51318359375,
0.339111328125
],
[
0.12005615234375,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35400390625,
0.12005615234375,
0.35400390625
],
[
0.51318359375,
0.342529296875,
0.88330078125,
0.342529296875,
0.88330078125,
0.355712890625,
0.51318359375,
0.355712890625
],
[
0.118408203125,
0.358154296875,
0.489501953125,
0.358154296875,
0.489501953125,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51318359375,
0.358154296875,
0.884765625,
0.358154296875,
0.884765625,
0.371826171875,
0.51318359375,
0.371826171875
],
[
0.118408203125,
0.37353515625,
0.490234375,
0.37353515625,
0.490234375,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.51318359375,
0.374755859375,
0.884765625,
0.374755859375,
0.884765625,
0.388427734375,
0.51318359375,
0.388427734375
],
[
0.118408203125,
0.39013671875,
0.490234375,
0.39013671875,
0.490234375,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.51318359375,
0.39013671875,
0.884765625,
0.39013671875,
0.884765625,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.118408203125,
0.40625,
0.337646484375,
0.40625,
0.337646484375,
0.419189453125,
0.118408203125,
0.419189453125
],
[
0.51318359375,
0.40625,
0.88671875,
0.40625,
0.88671875,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.1351318359375,
0.4228515625,
0.490234375,
0.42333984375,
0.490234375,
0.438232421875,
0.1351318359375,
0.43701171875
],
[
0.51318359375,
0.4228515625,
0.884765625,
0.4228515625,
0.884765625,
0.435791015625,
0.51318359375,
0.435791015625
],
[
0.118408203125,
0.440673828125,
0.490234375,
0.440673828125,
0.490234375,
0.45361328125,
0.118408203125,
0.45361328125
],
[
0.51318359375,
0.438232421875,
0.8037109375,
0.438232421875,
0.8037109375,
0.451904296875,
0.51318359375,
0.451904296875
],
[
0.118408203125,
0.457275390625,
0.489501953125,
0.457275390625,
0.489501953125,
0.47021484375,
0.118408203125,
0.47021484375
],
[
0.1175537109375,
0.47265625,
0.48779296875,
0.47265625,
0.48779296875,
0.486328125,
0.1175537109375,
0.486328125
],
[
0.51708984375,
0.47802734375,
0.5283203125,
0.47802734375,
0.5283203125,
0.485107421875,
0.51708984375,
0.485107421875
],
[
0.54150390625,
0.4755859375,
0.609375,
0.4755859375,
0.609375,
0.48681640625,
0.54150390625,
0.48681640625
],
[
0.118408203125,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.118408203125,
0.501953125
],
[
0.1175537109375,
0.50439453125,
0.490234375,
0.5029296875,
0.490234375,
0.51708984375,
0.1175537109375,
0.5185546875
],
[
0.51123046875,
0.505859375,
0.884765625,
0.505859375,
0.884765625,
0.51953125,
0.51123046875,
0.51953125
],
[
0.12005615234375,
0.52099609375,
0.490234375,
0.52099609375,
0.490234375,
0.53369140625,
0.12005615234375,
0.53369140625
],
[
0.51318359375,
0.521484375,
0.884765625,
0.521484375,
0.884765625,
0.53515625,
0.51318359375,
0.53515625
],
[
0.12005615234375,
0.5361328125,
0.489501953125,
0.5361328125,
0.489501953125,
0.5498046875,
0.12005615234375,
0.5498046875
],
[
0.51318359375,
0.5380859375,
0.88671875,
0.5380859375,
0.88671875,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.118408203125,
0.552734375,
0.489501953125,
0.552734375,
0.489501953125,
0.56640625,
0.118408203125,
0.56640625
],
[
0.51318359375,
0.5546875,
0.88671875,
0.5546875,
0.88671875,
0.568359375,
0.51318359375,
0.568359375
],
[
0.12005615234375,
0.57080078125,
0.489501953125,
0.57080078125,
0.489501953125,
0.58056640625,
0.12005615234375,
0.58056640625
],
[
0.51318359375,
0.57177734375,
0.884765625,
0.57177734375,
0.884765625,
0.5849609375,
0.51318359375,
0.5849609375
],
[
0.12005615234375,
0.58740234375,
0.181396484375,
0.58740234375,
0.181396484375,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.177978515625,
0.5849609375,
0.489501953125,
0.5849609375,
0.489501953125,
0.5986328125,
0.177978515625,
0.5986328125
],
[
0.51318359375,
0.58740234375,
0.884765625,
0.58740234375,
0.884765625,
0.60107421875,
0.51318359375,
0.60107421875
],
[
0.118408203125,
0.60107421875,
0.41650390625,
0.60107421875,
0.41650390625,
0.6142578125,
0.118408203125,
0.6142578125
],
[
0.51318359375,
0.6025390625,
0.884765625,
0.6025390625,
0.884765625,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1376953125,
0.619140625,
0.489501953125,
0.619140625,
0.489501953125,
0.6328125,
0.1376953125,
0.6328125
],
[
0.51123046875,
0.6181640625,
0.88330078125,
0.619140625,
0.88330078125,
0.6328125,
0.51123046875,
0.6318359375
],
[
0.1175537109375,
0.63525390625,
0.489501953125,
0.63427734375,
0.489501953125,
0.6484375,
0.1175537109375,
0.6494140625
],
[
0.51318359375,
0.63525390625,
0.88330078125,
0.63525390625,
0.88330078125,
0.6484375,
0.51318359375,
0.6484375
],
[
0.118408203125,
0.65185546875,
0.489501953125,
0.65185546875,
0.489501953125,
0.6650390625,
0.118408203125,
0.6650390625
],
[
0.5146484375,
0.65185546875,
0.88330078125,
0.65185546875,
0.88330078125,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.118408203125,
0.66650390625,
0.48779296875,
0.66650390625,
0.48779296875,
0.6796875,
0.118408203125,
0.6796875
],
[
0.51318359375,
0.66748046875,
0.884765625,
0.66748046875,
0.884765625,
0.68115234375,
0.51318359375,
0.68115234375
],
[
0.12005615234375,
0.68408203125,
0.489501953125,
0.68408203125,
0.489501953125,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.51318359375,
0.68310546875,
0.791015625,
0.68310546875,
0.791015625,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.12005615234375,
0.69970703125,
0.489501953125,
0.69970703125,
0.489501953125,
0.71337890625,
0.12005615234375,
0.71337890625
],
[
0.1175537109375,
0.71435546875,
0.489501953125,
0.71533203125,
0.489501953125,
0.72998046875,
0.1175537109375,
0.728515625
],
[
0.5146484375,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.72802734375,
0.5146484375,
0.72802734375
],
[
0.118408203125,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.74462890625,
0.118408203125,
0.74462890625
],
[
0.51318359375,
0.73291015625,
0.88330078125,
0.73291015625,
0.88330078125,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.12005615234375,
0.7490234375,
0.490234375,
0.7490234375,
0.490234375,
0.7626953125,
0.12005615234375,
0.7626953125
],
[
0.5146484375,
0.75048828125,
0.884765625,
0.75048828125,
0.884765625,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.118408203125,
0.76416015625,
0.490234375,
0.76416015625,
0.490234375,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51318359375,
0.765625,
0.88330078125,
0.765625,
0.88330078125,
0.779296875,
0.51318359375,
0.779296875
],
[
0.118408203125,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.79638671875,
0.491943359375,
0.79638671875,
0.491943359375,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.51318359375,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51318359375,
0.8125
],
[
0.118408203125,
0.8134765625,
0.490234375,
0.8134765625,
0.490234375,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.51318359375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.118408203125,
0.8291015625,
0.490234375,
0.8291015625,
0.490234375,
0.8427734375,
0.118408203125,
0.8427734375
],
[
0.51123046875,
0.830078125,
0.884765625,
0.830078125,
0.884765625,
0.84326171875,
0.51123046875,
0.84326171875
],
[
0.1175537109375,
0.84326171875,
0.4248046875,
0.84423828125,
0.4248046875,
0.859375,
0.1175537109375,
0.85791015625
],
[
0.51318359375,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.85986328125,
0.51318359375,
0.85986328125
],
[
0.1351318359375,
0.86083984375,
0.490234375,
0.8623046875,
0.490234375,
0.876953125,
0.1351318359375,
0.8759765625
],
[
0.51318359375,
0.8623046875,
0.8564453125,
0.86083984375,
0.8564453125,
0.8759765625,
0.51318359375,
0.876953125
],
[
0.1158447265625,
0.87744140625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.1158447265625,
0.892578125
],
[
0.1175537109375,
0.8955078125,
0.48779296875,
0.8955078125,
0.48779296875,
0.9091796875,
0.1175537109375,
0.9091796875
],
[
0.53662109375,
0.8955078125,
0.8759765625,
0.8955078125,
0.8759765625,
0.9091796875,
0.53662109375,
0.9091796875
]
] |
[
"A2",
"A3",
"Training Data",
"Model",
"A1",
"ANLI",
"ANLI-E/SNLI",
"MNLI-m/-mm",
"S,M*1",
"00.0",
"28.9",
"28.8",
"19.8",
"19.9",
"91.3",
"86.7 /86.4",
"32.6",
"29.3",
"35.0",
"34.2",
"91.3",
"44.2",
"+A1",
"86.3/86.5",
"57.3",
"33.4",
"44.6",
"45.2",
"BERT",
"43.2",
"90.9",
"86.3/86.3",
"50.5",
"+A1+A2",
"+A1+A2+A3",
"57.2",
"49.0",
"46.1",
"46.3",
"90.9",
"85.6/85.4",
"57.4",
"43.5",
"49.3",
"44.2",
"90.4",
"86.0/85.8",
"48.3",
"S,M,F,ANLI",
"XLNet",
"S,M,F,ANLI",
"67.6",
"48.3",
"55.1",
"91.8",
"89.6/89.4",
"52.01",
"50.7",
"S,M",
"47.6",
"31.1",
"31.4",
"92.6",
"22.1",
"25.4",
"90.8 /90.6",
"22.4",
"32.8",
"92.7",
"54.0",
"+F",
"24.2",
"33.7",
"90.6/90.5",
"22.0",
"92.8",
"68.7",
"RoBERTa",
"19.3",
"35.8",
"36.8",
"90.9 /90.7",
"20.4",
"+F+A1*2",
"43.7",
"41.4",
"71.2",
"+F+A1+A2*3",
"44.3",
"92.9",
"91.0/90.7",
"92.6",
"91.0/90.6",
"73.8",
"S,M,F,ANLI",
"48.9",
"44.4",
"53.7",
"49.7",
"Table 3: Model Performance. \"S' refers to SNLI, \"M' to MNLI dev (-m=matched, -mm=mismatched), and F' to",
"FEVER;'A1-A3'refer to the rounds respectively and ‘ANLI' refers toA1+A2+A3,-E'refers to test set examples",
"written by annotators exclusive to the test set. Datasets marked *n' were used to train the base model for round n,",
"and their performance on that round is underlined (A2 and A3 used ensembles, and hence have non-zero scores).",
"to over 103k examples for Round 3. We collected",
"tries is exceeded), and the average time this took",
"moredata for laterrounds not onlybecause that",
"(in seconds). Again, these metrics are useful for",
"data is likely to be more interesting, but also simply",
"evaluating model quality: observe that the average",
"because the base model is better and so annotation",
"number of tries and average time per verified error",
"took longer to collect good, verified correct exam-",
"both go up with later rounds. This demonstrates",
"ples of model vulnerabilities.",
"that the rounds are getting increasingly more diffi-",
"For each round, we report the model error rate,",
"cult. Further dataset statistics and inter-annotator",
"both on verified and unverified examples. The un-",
"agreement are reported in Appendix C.",
"verified model error rate captures the percentage",
"of examples where the model disagreed with the",
"4",
"Results",
"writer's target label, but where we are not (yet)",
"sure if the example is correct. The verified model",
"Table 3 reports the main results. In addition to",
"error rate is the percentage of model errors from",
"BERT (Devlin et al., 2018) and RoBERTa (Liu",
"example pairs that other annotators confirmed the",
"et al., 2019b), we also include XLNet (Yang et al.,",
"correct label for. Note that error rate is a useful way",
"2019) as an example of a strong, but different,",
"toevaluatemodelquality:thelowerthemodeler-",
"model architecture. We show test set performance",
"rorrate",
"-assuming constant annotator quality and",
"on the ANLI test sets per round, the total ANLI test",
"context-difficulty-the better the model.",
"set, and the exclusive test subset (examples from",
"We observe that model error rates decrease as",
"test-set-exclusive workers). We also show accuracy",
"we progress through rounds. In Round 3, where",
"on the SNLI test set and the MNLI development",
"we included a more diverse range of contexts",
"set (for the purpose of comparing between different",
"from various domains, the overall error rate went",
"model configurations across table rows). In what",
"slightly up compared to the preceding round, but",
"follows, we discuss our observations.",
"for Wikipedia contexts the error rate decreased sub-",
"stantially. While for the first round roughly 1 in",
"Basemodel performanceislow.Noticethatthe",
"every 5 examples were verified model errors, this",
"base model for each round performs very poorly on",
"quickly dropped over consecutive rounds, and the",
"that round's test set. This is the expected outcome:",
"overall model error rate is less than 1 in 10. On",
"For round 1, the base model gets the entire test set",
"the one hand, this is impressive, and shows how far",
"wrong, by design. For rounds 2 and 3, we used an",
"we have come with just three rounds. On the other",
"ensemble, so performance is not necessarily zero",
"hand, it shows that we still have a long way to go",
"However, as it turns out, performance still falls",
"if even untrained annotators can fool ensembles of",
" well below chance4, indicating that workers did not",
"state-of-the-art models with relative ease.",
"find vulnerabilities specific to a single model, but",
"Table 2 also reports the average number of",
"generally applicable ones for that model class.",
"\"tries\", i.e., attempts made for each context until a",
"model error was found (or the number of possible",
"4Chance is at 33%, since the test set labels are balanced."
] |
{
"arxiv_id": "1910.14599",
"language": "en",
"page_idx": 4
}
|
||
1912.04421-en-1
|
1912.04421-en-1.jpg
|
denoising kernels, and thus benefit from wider aggregation, while simultaneously limiting output dimensionality at each pixel, making the prediction network easier to train and compute- and memory-efficient. This approach is motivated by a long history of successful image restoration methods that have leveraged internal structure and self-similarity in natural images [5,2,35,22,40]. In burst denoising, selfsimilarity is particularly strong because we expect both spatial structure in the form of similar patterns that recur within a frame and across frames of the same scene, and temporal structure caused by consistency in scene and camera motion. Given the expected self-similarity and structure in the image intensities themselves, we argue that the corresponding denoising kernels must also have similar structure. Specifically, while allowing for individual per-pixel denoising kernels to be large, we assume that all the kernels for a given image span a lower-dimensional subspace.
Based on this observation, we introduce a new method for kernel-based burst denoising that achieves better accuracy and efficiency. Our contributions are:
$\cdot$ We introduce basis prediction networks (BPNs): given an input noisy burst, these networks predict a global low-dimensional basis set for large denoising kernels, and per-pixel coefficient vectors relative to this basis. BPN outputs are thus significantly lower-dimensional than the per-pixel arbitrary kernels from a regular KPN. $\bullet$ Enforcing this structure on the denoising kernels acts as a form of regularization, and our experiments show this leads to state-of-the-art denoising performance with significantly higher quality (>1 dB PSNR) than KPNs. $\bullet$ Beyond reducing memory usage and computational burden at the output layer, the structure of BPN's output enables the final kernel filtering step to be performed much more efficiently in the Fourier domain. We show this in terms of required number of FLOPs, as well as experimentally with actual run-times.
$2. \textbf{Related Work}$
Single-image and video denoising. Single-image denoising has been studied extensively. To overcome the illposed nature of the problem, classical approaches [32,33, 42] developed regularization schemes that model the local statistics of natural images. The most successful approaches [5, 2] exploit non-local self-similarity within the image and denoise pixels by aggregating similar pixels or patches from distant regions. These methods have been extended to denoise videos [19, 25], where the search for similar patches proceeds not only within frames, but also across different frames. Recent work has improved image denoising performance using convolutional networks trained on large datasets [3, 38, 39, 22, 40, 35, 36]. Other works use hand-crafted features to select per-pixel filters from learned filter sets for efficient image enhancement[30,8].
Burst denoising. Single-image denoising is a fundamentally under-constrained problem. Burst processing can reduce this ambiguity by using multiple observations (the frames of the burst) to recover a noise-free depiction of the scene. Burst denoising algorithms are now extensively used in commercial smartphone cameras [12] and can produce compelling results even in extreme low light scenarios [4,21]. Like in video denoising, a significant challenge for burst processing is the robustness to inter-frame motion. Many methods explicitly estimate this motion to align and denoise the frames [12,24,13,14,18]. Current state-ofthe-art burst denoising techniques [27,18,11,26] are based on deep neural networks. Many of them only require coarse registration, relying on the network to account for the small residual misalignments [27,11,26].
Kernel Prediction Networks. Given a burst sequence, Mildenhall et al.[27] predict per-pixel kernels that are then applied to the input burst to produce the denoised output. They demonstrate that KPNs outperform direct pixel synthesis that produces oversmooth results. This idea has been extended to use multiple varying size kernels at every pixel[26]. KPNs have also been used in other applications, including denoising Monte Carlo renderings [1,34,9], video super-resolution [16] and deblurring [41], frame interpolation [28, 29, 23] and video prediction [15,6, 23, 37].
Given their high-dimensional output (per-pixel kernels, that are three-dimensional in the case of burst denoising), KPNs have significant memory and compute requirements. Marinčet $al.[26]$ and Niklaus $etal.[29]$ ameliorate this by predicting spatially separable kernels: forming spatial kernels as outer products of predicted horizontal and vertical kernels. However, this makes a strong a-priori assumption about kernel structure, and still requires constructing and filtering with different per-pixel kernels. In contrast, our approach assumes that the set of per-pixel kernels for a scene span a low-dimensional sub-space, and predicts a basis for this sub-space based on the burst input. This approach also allows us to benefit from fast filtering in the Fourier domain.
## $3. \textbf{Method}$
In burst denoising, we are given an input noisy burst of images $I[n,t]$, where $n$ indexes spatial locations and $t\in\{1,\ldots,T\}$ the different frames in the burst. Using a heteroscedastic Gaussian noise model[7], which accounts for both read and shot noise, we relate this to the corresponding noise-free frames $R[n,t]$ as:
$$I[n,t]\sim\mathcal{N}(R[n,t],\sigma_{r}^{2}+\sigma_{s}^{2}R[n,t]),$$
(1)
where $\sigma_r^2$ and $\sigma_s^2$ are the read- and shot-noise parameters. Choosing the first frame as reference, our goal is to produce a single denoised image $\hat{R}[n]$ as an estimate of the first noise-free frame $R[n,1].$
2
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.080078125,
0.106689453125
],
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.080078125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50341796875,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.149658203125,
0.50341796875,
0.149658203125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.5009765625,
0.1514892578125,
0.890625,
0.154052734375,
0.890625,
0.1673583984375,
0.5009765625,
0.1654052734375
],
[
0.07843017578125,
0.1666259765625,
0.4697265625,
0.1673583984375,
0.4697265625,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.5009765625,
0.1666259765625,
0.890625,
0.1673583984375,
0.890625,
0.1817626953125,
0.5009765625,
0.18115234375
],
[
0.080078125,
0.1817626953125,
0.46728515625,
0.1817626953125,
0.46728515625,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.50341796875,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.07843017578125,
0.2279052734375,
0.471435546875,
0.2291259765625,
0.471435546875,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.5009765625,
0.2266845703125,
0.89208984375,
0.2279052734375,
0.89208984375,
0.242431640625,
0.5009765625,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.5009765625,
0.2568359375
],
[
0.080078125,
0.25830078125,
0.4697265625,
0.25830078125,
0.4697265625,
0.271484375,
0.080078125,
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
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.5009765625,
0.3037109375,
0.73876953125,
0.3037109375,
0.73876953125,
0.317626953125,
0.5009765625,
0.317626953125
],
[
0.08087158203125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.080078125,
0.333984375,
0.400390625,
0.333984375,
0.400390625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.335205078125,
0.890625,
0.3359375,
0.890625,
0.349853515625,
0.50244140625,
0.348388671875
],
[
0.098876953125,
0.348388671875,
0.4697265625,
0.348388671875,
0.4697265625,
0.3623046875,
0.098876953125,
0.3623046875
],
[
0.5009765625,
0.349853515625,
0.89208984375,
0.35107421875,
0.89208984375,
0.365478515625,
0.5009765625,
0.3642578125
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.367431640625,
0.890625,
0.367431640625,
0.890625,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.08087158203125,
0.380615234375,
0.35546875,
0.380615234375,
0.35546875,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.08905029296875,
0.392578125,
0.468017578125,
0.393310546875,
0.468017578125,
0.40771484375,
0.08905029296875,
0.4072265625
],
[
0.50244140625,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.10455322265625,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.10455322265625,
0.423583984375
],
[
0.50244140625,
0.412353515625,
0.890625,
0.412353515625,
0.890625,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.10455322265625,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.10455322265625,
0.438232421875
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.10455322265625,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.10455322265625,
0.453857421875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.10455322265625,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.10455322265625,
0.468505859375
],
[
0.5009765625,
0.45703125,
0.89208984375,
0.455810546875,
0.89208984375,
0.47021484375,
0.5009765625,
0.4716796875
],
[
0.10455322265625,
0.4697265625,
0.46728515625,
0.4697265625,
0.46728515625,
0.48291015625,
0.10455322265625,
0.48291015625
],
[
0.5009765625,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.486083984375,
0.5009765625,
0.486083984375
],
[
0.0906982421875,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.0906982421875,
0.498779296875
],
[
0.52294921875,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.52294921875,
0.50048828125
],
[
0.10296630859375,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.10296630859375,
0.5146484375
],
[
0.50244140625,
0.50048828125,
0.890625,
0.501953125,
0.890625,
0.5166015625,
0.50244140625,
0.51513671875
],
[
0.10455322265625,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.10455322265625,
0.52880859375
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.10296630859375,
0.52978515625,
0.463134765625,
0.52880859375,
0.463134765625,
0.54345703125,
0.10296630859375,
0.54443359375
],
[
0.5009765625,
0.5322265625,
0.890625,
0.53271484375,
0.890625,
0.546875,
0.5009765625,
0.54541015625
],
[
0.0906982421875,
0.54345703125,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.0906982421875,
0.55810546875
],
[
0.50244140625,
0.54736328125,
0.8896484375,
0.546875,
0.8896484375,
0.56103515625,
0.50244140625,
0.5625
],
[
0.10455322265625,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.10455322265625,
0.57373046875
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.10455322265625,
0.57568359375,
0.4697265625,
0.57568359375,
0.4697265625,
0.5888671875,
0.10455322265625,
0.5888671875
],
[
0.50244140625,
0.57763671875,
0.89208984375,
0.57763671875,
0.89208984375,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.10455322265625,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.10455322265625,
0.60400390625
],
[
0.50244140625,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.10455322265625,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.10455322265625,
0.61865234375
],
[
0.50244140625,
0.60693359375,
0.890625,
0.60791015625,
0.890625,
0.62255859375,
0.50244140625,
0.6220703125
],
[
0.10455322265625,
0.6220703125,
0.351318359375,
0.6220703125,
0.351318359375,
0.63525390625,
0.10455322265625,
0.63525390625
],
[
0.50341796875,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.5009765625,
0.63818359375,
0.890625,
0.63818359375,
0.890625,
0.6513671875,
0.5009765625,
0.6513671875
],
[
0.08087158203125,
0.6494140625,
0.2181396484375,
0.6494140625,
0.2181396484375,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.5048828125,
0.65478515625,
0.890625,
0.65478515625,
0.890625,
0.66552734375,
0.5048828125,
0.66552734375
],
[
0.50341796875,
0.66845703125,
0.890625,
0.66845703125,
0.890625,
0.681640625,
0.50341796875,
0.681640625
],
[
0.080078125,
0.6748046875,
0.46728515625,
0.6748046875,
0.46728515625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.6943359375,
0.5908203125,
0.6943359375,
0.5908203125,
0.708984375,
0.50244140625,
0.708984375
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.08087158203125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.08087158203125,
0.734375
],
[
0.51953125,
0.71826171875,
0.890625,
0.72021484375,
0.890625,
0.73486328125,
0.51953125,
0.73291015625
],
[
0.08087158203125,
0.7373046875,
0.465576171875,
0.7373046875,
0.465576171875,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.08087158203125,
0.79736328125,
0.46728515625,
0.79736328125,
0.46728515625,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.50244140625,
0.79443359375,
0.75244140625,
0.79443359375,
0.75244140625,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.57177734375,
0.81494140625,
0.82177734375,
0.81640625,
0.82177734375,
0.833984375,
0.57177734375,
0.83349609375
],
[
0.869140625,
0.818359375,
0.89208984375,
0.818359375,
0.89208984375,
0.83203125,
0.869140625,
0.83203125
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5048828125,
0.84033203125,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.5048828125,
0.85400390625
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.5048828125,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.5048828125,
0.86865234375
],
[
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,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.406982421875,
0.88720703125,
0.406982421875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88330078125,
0.701171875,
0.88525390625,
0.701171875,
0.90283203125,
0.5009765625,
0.90087890625
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"denoising kernels, and thus benefit from wider aggregation,",
"Burst denoising.Single-image denoising is a fundamen-",
"while simultaneously limiting output dimensionality at each",
"tally under-constrained problem. Burst processing can re-",
"pixel, making the prediction network easier to train and",
"duce this ambiguity by using multiple observations (the",
"compute-and memory-efficient. This approach is motivated",
"frames of the burst)to recover a noise-free depictionof",
"by a long history of successful image restoration methods",
"the scene.Burst denoising algorithms are now extensively",
"that have leveraged internal structure and self-similarity in",
"used in commercial smartphone cameras [12] and can pro-",
"natural images [5, 2, 35, 22, 40]. In burst denoising, self-",
"duce compelling results even in extreme low light scenar-",
"similarityisparticularlystrongbecauseweexpectbothspa-",
"ios [4, 21]. Like in video denoising, a significant challenge",
"tial structure in the form of similar patterns that recur within",
"for burst processing is the robustness to inter-frame motion",
"a frame and across frames of the same scene, and temporal",
"Many methods explicitly estimate this motion to align and",
"structure caused by consistency in scene and camera mo-",
"denoise the frames [12, 24, 13, 14, 18]. Current state-of-",
"tion. Given the expected self-similarity and structure in",
"the-art burst denoising techniques [27, 18, 11, 26] are based",
"the image intensities themselves, we argue that the corre-",
"on deep neural networks. Many of them only require coarse",
"sponding denoising kernels must also have similar structure.",
"registration, relying on the network to account for the small",
"Specifically,while allowing for individual per-pixel denois-",
"residual misalignments [27, 11, 26].",
"ingkernels to belarge,weassumethat all thekernels for a",
"given image span a lower-dimensional subspace.",
"Kernel Prediction Networks.Given a burst sequence,",
"Based on this observation, we introduce a new method",
"Mildenhall et al. [27] predict per-pixel kernels that are then",
"for kernel-based burst denoising that achieves better accu-",
"applied to the input burst to produce the denoised out-",
"racy and efficiency.Ourcontributions are:",
"put. They demonstrate that KPNs outperform direct pixel",
". We introduce basis prediction networks (BPNs): given",
"synthesis that produces oversmooth results. This idea has",
"an input noisy burst, these networks predict a global",
"been extended to use multiple varying size kernels at every",
"low-dimensional basis set for large denoising kernels,",
"pixel [26]. KPNs have also been used in other applications,",
"and per-pixel coefficient vectors relative to this basis.",
"including denoising Monte Carlo renderings [l, 34, 9],",
"BPN outputs are thus significantly lower-dimensional",
"video super-resolution [16] and deblurring [41], frame in-",
"than the per-pixel arbitrary kernels from a regular KPN.",
"terpolation [28, 29, 23] and video prediction [15, 6, 23, 37].",
": Enforcing this structure on the denoising kernels acts",
"Given their high-dimensional output (per-pixel kernels,",
"as a form of regularization, and our experiments show",
"that are three-dimensional in the case of burst denoising),",
"this leads to state-of-the-art denoising performance with",
"KPNs have significant memory and compute requirements.",
"significantly higher quality (>1 dB PSNR) than KPNs.",
"Marincet al.[26] and Niklaus et al.[29] amelioratethis by",
".Beyond reducing memory usage and computational bur-",
"predicting spatially separable kernels: forming spatial ker-",
"den at the output layer, the structure of BPN's output",
"nels as outer products of predicted horizontal and vertical",
"enables the final kernel filtering step to be performed",
"kernels. However, this makes a strong a-priori assumption",
"much more efficiently in the Fourier domain. We show",
"about kernel structure, and still requires constructing and",
"this in terms of required number of FLOPs, as well as",
"filtering with different per-pixel kernels. In contrast, our ap-",
"experimentally with actual run-times.",
"proach assumes that the set of per-pixel kernels for a scene",
"span a low-dimensional sub-space, and predicts a basis for",
"2. Related Work",
"this sub-spacebased on theburstinput.This approachalso",
"allows us to benefit from fast filtering in the Fourier domain.",
"Single-image and video denoising. Single-image de-",
"noising has been studied extensively. To overcome the ill-",
"3. Method",
"posed nature of the problem, classical approaches [32, 33,",
"421 developed regularization schemes that model the lo-",
"In burst denoising, we are given an input noisy burst",
"cal statistics of natural images.The most successful ap-",
"of images I n,tl, where n indexes spatial locations and",
"proaches [5, 2] exploit non-local self-similarity within the",
"t E [1,..., T] the different frames in the burst. Using a",
"image and denoise pixels by aggregating similar pixels or",
"heteroscedastic Gaussian noise model [7], which accounts",
"patches from distant regions. These methods have been ex-",
"for both read and shot noise, we relate this to the corre-",
"tended to denoise videos[19.251.where the search for sim-",
"sponding noise-free frames R[n, t] as:",
"ilar patches proceeds not only within frames, but also across",
"I[n,t) ~ N(R[n, t],o? +o?R[n, t),",
"(1)",
"different frames. Recent work has improved image denois-",
"ing performance using convolutional networks trained on",
"where α? and ? are the read- and shot-noise parameters.",
"large datasets[3.38,39.22, 40,35, 361.Other works use",
"Choosing the first frame as reference, our goal is to pro-",
"hand-crafted features to select per-pixel filters from learned",
"duce a single denoised image R[n] as an estimate of the",
"filter sets for efficient image enhancement [30, 8].",
"first noise-free frame R[n, 1].",
"2"
] |
{
"arxiv_id": "1912.04421",
"language": "en",
"page_idx": 1
}
|
||
1912.04421-en-2
|
1912.04421-en-2.jpg
|
<FigureHere>
Figure 2: Our basis prediction network takes as input a burst of noisy input frames (a) together with the noise parameters (b). The frames are encoded into a shared feature space (c). These features are then decoded by two decoders with skip connections into a burst-specific basis of 3D kernels (e) and a set of per-pixel mixing coefficients (d). Both the coefficients and basis kernels are individually unit-normalized. Finally, we obtain per-pixel kernels by mixing the basis elements according to the coefficients and we apply them to the input burst to produce the final denoised image (f).
# 3.1. Kernel-based burst denoising
Rather than train a network to regress $\hat{R}$ directly, Kernel Prediction Networks output a field of denoising kernels $w_n[\delta,t]$, one for each pixel $n$ at each frame $t.$ The kernels have a spatial support $K\times K$, indexed by $\delta$, with separate weights for each frame. Given these predicted kernels, the denoised estimate $\hat{R}$ is formed as:
(2)
$$\hat{R}[n]=\sum_t\sum_\delta w_n[\delta,t]I[n-\delta,t].$$
A key bottleneck in this pipeline is the prediction of this dense kernel field $w$, which requires producing $K^2T$ numbers at every pixel of the output. Since networks with highdimensional outputs are both expensive and require learning a large number of parameters in their last layer, KPNs have typically been used only with small kernels $(K=5$ in [27]).
We designed our network following an encoder-decoder architecture with skip connections [31]. Our model, however, has two decoder branches, one for the basis, the other for the coefficients (Figure 2). The encoder is shared between the two branches because the meaning of the coefficients $c$ is dependent on the predicted basis $v$ in Equation (3), so the two outputs need to be co-ordinated. This encoder takes the noisy burst and noisy parameters as input, and through multiple levels of downsampling and global average pooling at the end, yields a single global feature vector as its encoding of the image. The per-pixel coefficients $c$ are then decoded from the encoder bottleneck to the full image resolution $W\times H$,with $B$ channels as output. The common basis $v$ is decoded up to distinct spatial dimensions -that of the kernels $K\times K-$with $B\times T$ output channels.
Since the basis branch decodes to a different spatial resolution, we need a careful treatment of the skip connections. Unlike a usual U-Net, the encoder and decoder feature size do not match. Specifically, a pixel $\delta$ in the basis kernel $v_b[\delta,\cdot]$ has no meaningful relation to a pixel $n$ in the input frames $I[n,\cdot].$ Therefore, in the skip connections from the shared encoder to the basis decoder, we apply a global spatial average pooling of the encoder's activations, and replicate the average vector to the resolution of the decoder layer. This mechanism ensures the encoder information is globally aggregated without creating nonsensical correspondences between kernel and image locations, while allowing features at multiple scales of the encoder to inform the basis decoder.
We ensure each of the reconstructed kernel $w$ has positive weights that sum to one, to represent averaging. We implement this constraint using soft-max normalizations on both the coefficient and basis decoder outputs. So every 3D kernel of the basis $v_b[\cdot,\cdot]$ and every coefficient vector $c_n[\cdot]$ is normalized individually. A more detailed description of the architecture is provided in the supplementary.
## 3.2. Basis Prediction Networks
Instead of directly predicting unconstrained kernels for each spatial location, we designed a network that outputs: (1) a global kernel basis $v_b[\tilde{\delta},t]$, of size $K^2T\times B$ with $b\in\{1,\ldots B\};$ and (2) a $B$ dimensional coefficient vector $c_n[b]$ at each spatial location.
(3)
$$w_n[\delta,t]=\sum_bv_b[\delta,t]c_n[b].$$
Note that we typically choose the number of basis kernels $B\ll K^{2}T.$ This implies that all the kernels for a given burst lie in a low-dimensional subspace, but this subspace will be different for different bursts, $i.e.$ the basis is burstspecific. This procedure allows us to recreate a full kernel field with far fewer predictions. Assuming a $W\times H$ resolution image, we need only make $WHB+\tilde{K}^2TB$ predictions to effectively recreate a kernel feld of size $WHK^2T.$
3
|
[
[
0.0882568359375,
0.10162353515625,
0.2017822265625,
0.10162353515625,
0.2017822265625,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.68115234375,
0.126953125,
0.7490234375,
0.126953125,
0.7490234375,
0.136962890625,
0.68115234375,
0.136962890625
],
[
0.44775390625,
0.136962890625,
0.48779296875,
0.136962890625,
0.48779296875,
0.14453125,
0.44775390625,
0.14453125
],
[
0.6748046875,
0.136962890625,
0.7587890625,
0.1395263671875,
0.7587890625,
0.15087890625,
0.6748046875,
0.1483154296875
],
[
0.44775390625,
0.146484375,
0.484375,
0.146484375,
0.484375,
0.154052734375,
0.44775390625,
0.154052734375
],
[
0.2320556640625,
0.1673583984375,
0.26708984375,
0.1673583984375,
0.26708984375,
0.1749267578125,
0.2320556640625,
0.1749267578125
],
[
0.51123046875,
0.169189453125,
0.65673828125,
0.169189453125,
0.65673828125,
0.18310546875,
0.51123046875,
0.18310546875
],
[
0.2344970703125,
0.177978515625,
0.2548828125,
0.177978515625,
0.2548828125,
0.1817626953125,
0.2344970703125,
0.1817626953125
],
[
0.313720703125,
0.1875,
0.420654296875,
0.1875,
0.420654296875,
0.1976318359375,
0.313720703125,
0.1976318359375
],
[
0.77392578125,
0.1962890625,
0.88232421875,
0.1962890625,
0.88232421875,
0.210205078125,
0.77392578125,
0.210205078125
],
[
0.11273193359375,
0.2108154296875,
0.17236328125,
0.2108154296875,
0.17236328125,
0.2259521484375,
0.11273193359375,
0.2259521484375
],
[
0.459228515625,
0.205810546875,
0.490234375,
0.205810546875,
0.490234375,
0.21337890625,
0.459228515625,
0.21337890625
],
[
0.324462890625,
0.2139892578125,
0.374267578125,
0.2139892578125,
0.374267578125,
0.2215576171875,
0.324462890625,
0.2215576171875
],
[
0.459228515625,
0.21533203125,
0.494384765625,
0.21533203125,
0.494384765625,
0.2197265625,
0.459228515625,
0.2197265625
],
[
0.324462890625,
0.2247314453125,
0.358642578125,
0.2247314453125,
0.358642578125,
0.2291259765625,
0.324462890625,
0.2291259765625
],
[
0.08660888671875,
0.2310791015625,
0.19775390625,
0.2310791015625,
0.19775390625,
0.2412109375,
0.08660888671875,
0.2412109375
],
[
0.53173828125,
0.2298583984375,
0.6396484375,
0.2298583984375,
0.6396484375,
0.24365234375,
0.53173828125,
0.24365234375
],
[
0.08087158203125,
0.255126953125,
0.890625,
0.255126953125,
0.890625,
0.268310546875,
0.08087158203125,
0.268310546875
],
[
0.080078125,
0.270751953125,
0.89208984375,
0.270751953125,
0.89208984375,
0.2841796875,
0.080078125,
0.2841796875
],
[
0.080078125,
0.285400390625,
0.89208984375,
0.285400390625,
0.89208984375,
0.298583984375,
0.080078125,
0.298583984375
],
[
0.0767822265625,
0.2998046875,
0.89208984375,
0.300537109375,
0.89208984375,
0.31494140625,
0.0767822265625,
0.314453125
],
[
0.080078125,
0.31640625,
0.701171875,
0.31640625,
0.701171875,
0.330078125,
0.080078125,
0.330078125
],
[
0.07843017578125,
0.354736328125,
0.338134765625,
0.356689453125,
0.338134765625,
0.371826171875,
0.07843017578125,
0.369384765625
],
[
0.52294921875,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.52294921875,
0.37060546875
],
[
0.50244140625,
0.37255859375,
0.89208984375,
0.37255859375,
0.89208984375,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.098876953125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.39453125,
0.098876953125,
0.39453125
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.50244140625,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.50244140625,
0.416748046875
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.080078125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.5009765625,
0.431884765625,
0.890625,
0.434326171875,
0.890625,
0.447509765625,
0.5009765625,
0.44580078125
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.455810546875,
0.3056640625,
0.455810546875,
0.3056640625,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50244140625,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.16259765625,
0.484130859375,
0.38818359375,
0.484130859375,
0.38818359375,
0.498779296875,
0.16259765625,
0.498779296875
],
[
0.4453125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.4453125,
0.498779296875
],
[
0.5009765625,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.5009765625,
0.492431640625
],
[
0.50341796875,
0.494384765625,
0.890625,
0.494384765625,
0.890625,
0.50830078125,
0.50341796875,
0.50830078125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.08251953125,
0.5244140625,
0.4697265625,
0.5244140625,
0.4697265625,
0.53857421875,
0.08251953125,
0.53857421875
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.07843017578125,
0.5390625,
0.468017578125,
0.53857421875,
0.468017578125,
0.55322265625,
0.07843017578125,
0.5537109375
],
[
0.5009765625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.5009765625,
0.5517578125
],
[
0.07843017578125,
0.5546875,
0.468017578125,
0.5537109375,
0.468017578125,
0.568359375,
0.07843017578125,
0.5693359375
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.07843017578125,
0.5693359375,
0.4697265625,
0.57080078125,
0.4697265625,
0.583984375,
0.07843017578125,
0.58349609375
],
[
0.521484375,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.521484375,
0.58203125
],
[
0.08087158203125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.5966796875,
0.08087158203125,
0.5966796875
],
[
0.521484375,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.521484375,
0.59765625
],
[
0.080078125,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.50341796875,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.611328125,
0.50341796875,
0.611328125
],
[
0.50244140625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.62548828125,
0.316162109375,
0.62548828125,
0.316162109375,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.098876953125,
0.6494140625,
0.468017578125,
0.6494140625,
0.468017578125,
0.66357421875,
0.098876953125,
0.66357421875
],
[
0.5048828125,
0.646484375,
0.890625,
0.646484375,
0.890625,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.080078125,
0.66552734375,
0.4697265625,
0.66552734375,
0.4697265625,
0.6787109375,
0.080078125,
0.6787109375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.5009765625,
0.673828125,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.5009765625,
0.68798828125
],
[
0.08087158203125,
0.6796875,
0.4697265625,
0.6796875,
0.4697265625,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.080078125,
0.69580078125,
0.4697265625,
0.69580078125,
0.4697265625,
0.708984375,
0.080078125,
0.708984375
],
[
0.50341796875,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08087158203125,
0.7109375,
0.27294921875,
0.7109375,
0.27294921875,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.268798828125,
0.7392578125,
0.365234375,
0.7392578125,
0.365234375,
0.7568359375,
0.268798828125,
0.7568359375
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.44677734375,
0.7412109375,
0.4697265625,
0.7412109375,
0.4697265625,
0.7548828125,
0.44677734375,
0.7548828125
],
[
0.1829833984375,
0.74072265625,
0.26318359375,
0.73779296875,
0.26318359375,
0.75390625,
0.1829833984375,
0.755859375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.2646484375,
0.76123046875,
0.273681640625,
0.76123046875,
0.273681640625,
0.7666015625,
0.2646484375,
0.7666015625
],
[
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,
0.6689453125,
0.78076171875,
0.6689453125,
0.79443359375,
0.5009765625,
0.79443359375
],
[
0.08087158203125,
0.79443359375,
0.471435546875,
0.79541015625,
0.471435546875,
0.81005859375,
0.08087158203125,
0.8095703125
],
[
0.521484375,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.521484375,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.07843017578125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.50341796875,
0.82763671875,
0.890625,
0.82763671875,
0.890625,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84033203125,
0.468017578125,
0.85400390625,
0.07843017578125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.88818359375,
0.85693359375,
0.88818359375,
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
],
[
0.080078125,
0.88623046875,
0.43310546875,
0.88623046875,
0.43310546875,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.82666015625,
0.88720703125,
0.82666015625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"(a) Noise Input Burst",
"Kernel-basec",
"oefficient",
"Reconstruction",
"Decoder",
"Common",
"(d) Basis Coefficients Maps",
"Enco",
"(c)EncoderFeatures",
"(f) Denoised Output",
"Or,s",
"Basis",
"Pooled Skip",
"Deroder",
"Connect",
"(b)NoiseParameters",
"(e) 3D Kermel Basis",
"Figure 2: Our basis prediction network takes as input a burst of noisy input frames (a) together with the noise parameters",
"(b). The frames are encoded into a shared feature space (c). These features are then decoded by two decoders with skip",
"connections into a burst-specific basis of 3D kernels (e) and a set of per-pixel mixing coefficients (d). Both the coefficients and",
"basis kernels are individually unit-normalized. Finally, we obtain per-pixel kernels by mixing the basis elements according",
"to the coefficients and we apply them to the input burst to produce the final denoised image (f).",
"3.1.Kernel-based burst denoising",
"We designed our network following an encoder-decoder",
"architecture with skip connections [31]. Our model, how-",
"Rather than train a network to regress R directly, Ker-",
"ever, has two decoder branches, one for the basis, the other",
"nel Prediction Networks output a field of denoising kernels",
"for the coefficients (Figure 2). The encoder is shared be-",
"Wn[o, tl, one for each pixel n at each frame t. The kernels",
"tween the two branches because the meaning of the coef-",
"have a spatial support K × K, indexed by , with separate",
"ficients c is dependent on the predicted basis in Equa-",
"weights for each frame. Given these predicted kernels, the",
"tion (3), so the two outputs need to be co-ordinated. This",
"denoised estimate R is formed as:",
"encoder takes the noisy burst and noisy parameters as input,",
"R[n] =wn[8,t]I[n -, t].",
"(2)",
" and through multiple levels of downsampling and global av-",
"erage pooling at the end, yields a single global feature vec-",
"tor as its encoding of the image. The per-pixel coefficients",
"A key bottleneck in this pipeline is the prediction of this",
"c are then decoded from the encoder bottleneck to the full",
"dense kernel field w, which requires producing K?T num-",
"image resolution W x H, with B channels as output. The",
"bers at every pixel of the output. Since networks with high-",
"common basis v is decoded up to distinct spatial dimensions",
"dimensional outputs are both expensive and require learning",
"that of the kernels K ×x K with B ×T output channels.",
"alargenumberofparametersintheirlastlayer,KPNshave",
"Since the basis branch decodes to a different spatial res-",
"typically been used only with small kernels (K = 5 in [27]).",
"olution,weneed a careful treatment ofthe skip connec-",
"tions. Unlike a usual U-Net, the encoder and decoder fea-",
"3.2. Basis Prediction Networks",
"ture size do not match. Specifically, a pixel in the ba-",
"Instead of directly predicting unconstrained kernels for",
"sis kernel vns,-has no meaningful relation to a pixel n",
"each spatial location, we designed a network that outputs:",
"in the input frames I n, -l. Therefore, in the skip connec-",
"tions from the shared encoder to the basis decoder, we ap-",
"(1) a global kernel basis vb[s,t], of size K2T × B with",
"b E (1, ... B); and (2) a B dimensional coefficient vector",
"ply a global spatial average pooling of the encoder's acti-",
"Cn [b] at each spatial location.",
"vations, and replicate the average vector to the resolution",
"of the decoder layer. This mechanism ensures the encoder",
"vr[8, t]cn [6].",
"informationisgloballyaggregatedwithoutcreatingnonsen-",
"(3)",
"wn[8,t] =",
"sical correspondences between kernel and image locations,",
"h",
"while allowing features at multiple scales of the encoder to",
"Note that we typically choose the number of basis kernels",
"inform the basis decoder.",
"B < K?T. This implies that all the kernels for a given",
"We ensure each of the reconstructed kernel w has pos-",
"burst lie in a low-dimensional subspace, but this subspace",
"itive weights that sum to one, to represent averaging. We",
"will be different for different bursts, i.e. the basis is burst-",
"implement this constraint using soft-max normalizations on",
"specific. This procedure allows us to recreate a full kernel",
"both the coefficient and basis decoder outputs. So every 3D",
"field with far fewer predictions. Assuming a W × H resolu-",
"kernel of the basis vb[, I and every coefficient vector cn :",
"tion image, we need only make WHB+K?TB predictions",
"is normalized individually. A more detailed description of",
"to effectively recreate a kernel field of size W HK?T.",
"the architecture is provided in the supplementary.",
"3"
] |
{
"arxiv_id": "1912.04421",
"language": "en",
"page_idx": 2
}
|
||
1912.04421-en-4
|
1912.04421-en-4.jpg
|
<FigureHere>
Figure 3: We illustrate denoising performance on a benchmark synthetic grayscale test set [27] for our method, a direct
prediction network (which directly regresses denoised pixels), and two KPN variants [27,26] with the same kernel size $K=15$ as our method. The numbers in inset refer to the PSNR (dB) on the full image. In addition to better quantitative performance, our method does better at reproducing perceptual details like textures, edges, and text.
mented a version ourselves (that we use in ablations in the next section) and also report its performance in Table l. We find it closely matches the performance reported in [27]. Additionally, we train a network to directly regress the denoised pixel values from the input burst $(i.e.$, without kernels), as well as our implementation of [26] with a larger kernel size of $K=15$ for fair comparison.
We find that our method outperforms KPN [27] by a significant margin, over 1 dB PSNR at all noise levels. Our implementation of [26] also does well, but remains inferior to our model. We show qualitative results for a subset of methods in Figure 3. Our have fewer artifacts, especially in textured regions and around thin structures like printed text.
work's output dimensionality and number of learnable parameters. To tease apart the contributions of kernel size and the structure of our kernel decomposition, we conduct an ablation study on our validation set. The results can be found in Table 2. The performance gap between the test and validation set results (Table 1 and 2) comes from differences in the datasets themselves.
Kernel Size. As a baseline, we consider using KPN directly with our larger kernel size of $K=15.$ We also consider predicting a single separable kernel at that size ([26] predicts separable kernels at multiple sizes, and adds them $\hat{\text{together). }}$We find that our network outperforms the large kernel KPN variant at all noise levels—suggesting that simply increasing the kernel size is not enough. It also outperforms separable kernel prediction, suggesting that a lowdimensional subspace constraint better captures the structure of natural images than spatial separability.
## 4.2. Ablation and analysis
Our approach leads to better denoising quality because it enables larger kernels without vastly increasing the net-
5
|
[
[
0.5498046875,
0.3125,
0.560546875,
0.3125,
0.560546875,
0.322509765625,
0.5498046875,
0.322509765625
],
[
0.64794921875,
0.311279296875,
0.65625,
0.317626953125,
0.64892578125,
0.322509765625,
0.640625,
0.31640625
],
[
0.73876953125,
0.3125,
0.7509765625,
0.3125,
0.7509765625,
0.322509765625,
0.73876953125,
0.322509765625
],
[
0.833984375,
0.3125,
0.8447265625,
0.3125,
0.8447265625,
0.322509765625,
0.833984375,
0.322509765625
],
[
0.77392578125,
0.317626953125,
0.80126953125,
0.317626953125,
0.80126953125,
0.324462890625,
0.77392578125,
0.324462890625
],
[
0.455078125,
0.314453125,
0.464111328125,
0.314453125,
0.464111328125,
0.3212890625,
0.455078125,
0.3212890625
],
[
0.50244140625,
0.3193359375,
0.51123046875,
0.31640625,
0.5146484375,
0.3212890625,
0.5048828125,
0.324462890625
],
[
0.5908203125,
0.318115234375,
0.609375,
0.318115234375,
0.609375,
0.3212890625,
0.5908203125,
0.3212890625
],
[
0.6611328125,
0.323974609375,
0.7060546875,
0.323974609375,
0.7060546875,
0.333984375,
0.6611328125,
0.333984375
],
[
0.685546875,
0.3193359375,
0.70166015625,
0.313232421875,
0.70751953125,
0.3212890625,
0.69140625,
0.32763671875
],
[
0.2034912109375,
0.32568359375,
0.256591796875,
0.32568359375,
0.256591796875,
0.333251953125,
0.2034912109375,
0.333251953125
],
[
0.301513671875,
0.32568359375,
0.32275390625,
0.32568359375,
0.32275390625,
0.333251953125,
0.301513671875,
0.333251953125
],
[
0.3798828125,
0.32568359375,
0.420654296875,
0.32568359375,
0.420654296875,
0.333251953125,
0.3798828125,
0.333251953125
],
[
0.476318359375,
0.32568359375,
0.5146484375,
0.32568359375,
0.5146484375,
0.333251953125,
0.476318359375,
0.333251953125
],
[
0.57861328125,
0.32568359375,
0.61279296875,
0.32568359375,
0.61279296875,
0.335205078125,
0.57861328125,
0.335205078125
],
[
0.7587890625,
0.32568359375,
0.80126953125,
0.32568359375,
0.80126953125,
0.333251953125,
0.7587890625,
0.333251953125
],
[
0.67626953125,
0.333251953125,
0.705078125,
0.333251953125,
0.705078125,
0.341552734375,
0.67626953125,
0.341552734375
],
[
0.57421875,
0.326904296875,
0.580078125,
0.326904296875,
0.580078125,
0.33203125,
0.57421875,
0.33203125
],
[
0.488525390625,
0.335205078125,
0.51318359375,
0.330810546875,
0.515625,
0.339111328125,
0.491943359375,
0.343505859375
],
[
0.59716796875,
0.3359375,
0.60693359375,
0.3359375,
0.60693359375,
0.34033203125,
0.59716796875,
0.34033203125
],
[
0.7685546875,
0.333984375,
0.79833984375,
0.333984375,
0.79833984375,
0.341552734375,
0.7685546875,
0.341552734375
],
[
0.50341796875,
0.3447265625,
0.5107421875,
0.3447265625,
0.5107421875,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.68701171875,
0.3447265625,
0.69775390625,
0.34228515625,
0.69921875,
0.3466796875,
0.68896484375,
0.348388671875
],
[
0.81640625,
0.354248046875,
0.8203125,
0.354248046875,
0.8203125,
0.359130859375,
0.81640625,
0.359130859375
],
[
0.4306640625,
0.36865234375,
0.4599609375,
0.36865234375,
0.4599609375,
0.380126953125,
0.4306640625,
0.380126953125
],
[
0.525390625,
0.36865234375,
0.55615234375,
0.36865234375,
0.55615234375,
0.380126953125,
0.525390625,
0.380126953125
],
[
0.62109375,
0.36865234375,
0.65283203125,
0.36865234375,
0.65283203125,
0.380615234375,
0.62109375,
0.380615234375
],
[
0.71728515625,
0.36865234375,
0.74853515625,
0.367431640625,
0.74853515625,
0.37890625,
0.71728515625,
0.380126953125
],
[
0.15283203125,
0.5322265625,
0.236083984375,
0.53271484375,
0.236083984375,
0.54443359375,
0.15283203125,
0.54345703125
],
[
0.28515625,
0.5322265625,
0.366943359375,
0.5322265625,
0.366943359375,
0.54248046875,
0.28515625,
0.54248046875
],
[
0.49267578125,
0.53271484375,
0.5439453125,
0.53271484375,
0.5439453125,
0.54443359375,
0.49267578125,
0.54443359375
],
[
0.580078125,
0.53271484375,
0.640625,
0.53271484375,
0.640625,
0.54443359375,
0.580078125,
0.54443359375
],
[
0.7646484375,
0.53271484375,
0.833984375,
0.53271484375,
0.833984375,
0.54345703125,
0.7646484375,
0.54345703125
],
[
0.68896484375,
0.5341796875,
0.71337890625,
0.5341796875,
0.71337890625,
0.54248046875,
0.68896484375,
0.54248046875
],
[
0.404296875,
0.5341796875,
0.435546875,
0.5341796875,
0.435546875,
0.54248046875,
0.404296875,
0.54248046875
],
[
0.08087158203125,
0.55615234375,
0.89208984375,
0.55615234375,
0.89208984375,
0.5693359375,
0.08087158203125,
0.5693359375
],
[
0.08087158203125,
0.57080078125,
0.89208984375,
0.57080078125,
0.89208984375,
0.58544921875,
0.08087158203125,
0.58544921875
],
[
0.08087158203125,
0.5859375,
0.89208984375,
0.5859375,
0.89208984375,
0.599609375,
0.08087158203125,
0.599609375
],
[
0.08087158203125,
0.6015625,
0.73193359375,
0.6015625,
0.73193359375,
0.61572265625,
0.08087158203125,
0.61572265625
],
[
0.080078125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.65576171875,
0.080078125,
0.65576171875
],
[
0.5009765625,
0.6416015625,
0.8896484375,
0.642578125,
0.8896484375,
0.6572265625,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.080078125,
0.671875
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50244140625,
0.671875
],
[
0.080078125,
0.673828125,
0.46728515625,
0.673828125,
0.46728515625,
0.6875,
0.080078125,
0.6875
],
[
0.50341796875,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.08087158203125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.70263671875,
0.08087158203125,
0.70263671875
],
[
0.50341796875,
0.68798828125,
0.8935546875,
0.68798828125,
0.8935546875,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.080078125,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.50341796875,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.080078125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.734375,
0.359375,
0.734375,
0.359375,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50341796875,
0.73486328125,
0.7158203125,
0.73486328125,
0.7158203125,
0.74560546875,
0.50341796875,
0.74560546875
],
[
0.098876953125,
0.74755859375,
0.46728515625,
0.74853515625,
0.46728515625,
0.76318359375,
0.098876953125,
0.76220703125
],
[
0.50341796875,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79150390625,
0.08087158203125,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79443359375,
0.4697265625,
0.79443359375,
0.4697265625,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.080078125,
0.8095703125,
0.4697265625,
0.8095703125,
0.4697265625,
0.82275390625,
0.080078125,
0.82275390625
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.82470703125,
0.46728515625,
0.82470703125,
0.46728515625,
0.83837890625,
0.080078125,
0.83837890625
],
[
0.50244140625,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50244140625,
0.8388671875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.8388671875,
0.890625,
0.85400390625,
0.50244140625,
0.85595703125
],
[
0.080078125,
0.8486328125,
0.279296875,
0.8486328125,
0.279296875,
0.8623046875,
0.080078125,
0.8623046875
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.09722900390625,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.09722900390625,
0.88623046875
],
[
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,
0.88720703125,
0.80810546875,
0.88720703125,
0.80810546875,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"S",
"O",
"'S",
"'S",
"attack",
"S",
"",
"",
"tovietory",
"tact",
"thebirthplace",
"n",
"oviutar",
"viutary",
"vetuny.",
"pvictory",
"Rove",
"1",
"Rn",
"",
"23.0Vg",
"",
"",
"-",
"28.4",
"28.0",
"28.9",
"30.7",
"Full Input Image",
"Noisy Reference",
"KPN [27]",
"MKPN [26]",
"Ground Truth",
"Ours",
"Direct",
"Figure 3: We illustrate denoising performance on a benchmark synthetic grayscale test set [27] for our method, a direct",
"prediction network (which directly regresses denoised pixels), and two KPN variants [27, 26] with the same kernel size",
"K = 15 as our method. The numbers in inset refer to the PSNR (dB) on the full image. In addition to better quantitative",
"performance, our method does better at reproducing perceptual details like textures, edges, and text.",
"mented a version ourselves (that we use in ablations in the",
"work's output dimensionality and number of learnable pa-",
"next section) and also report its performance in Table 1. We",
"rameters. To tease apart the contributions of kernel size and",
"find it closely matches the performance reported in [27].",
"thestructureofourkerneldecomposition,weconductan",
"Additionally, we train a network to directly regress the de-",
"ablation study on our validation set. The results can be",
"noised pixel values from the input burst (i.e., without ker-",
"found in Table 2. The performance gap between the test",
"nels), as well as our implementation of [26] with a larger",
"and validation set results (Table 1 and 2) comes from differ-",
"kernel size of K = 15 for fair comparison.",
"encesinthedatasetsthemselves.",
"We find that our method outperforms KPN [27] by a sig-",
"Kernel Size. As a baseline, we consider using KPN di-",
"nificant margin, over 1 dB PSNR at all noise levels. Our",
"rectly with our larger kernel size of K = 15. We also con-",
"implementationof[26lalsodoeswell,butremainsinferior",
"sider predictinga single separablekernel atthat size([26)",
"to our model. We show qualitative results for a subset of",
"predicts separable kernels at multiple sizes, and adds them",
"methods in Figure 3.Our have fewer artifacts, especially in",
"together). We find that our network outperforms the large",
"textured regions and around thin structures like printed text.",
"kernel KPN variant at all noise levels-suggesting that sim-",
"ply increasing the kernel size is not enough. It also outper-",
"4.2. Ablation and analysis",
"forms separable kernel prediction, suggesting that a low-",
"Our approach leads to better denoising quality because",
"dimensional subspace constraint better captures the struc-",
"it enables larger kernels without vastly increasing the net-",
"ture of natural images than spatial separability.",
"5"
] |
{
"arxiv_id": "1912.04421",
"language": "en",
"page_idx": 4
}
|
||
1806.05559-en-0
|
1806.05559-en-0.jpg
|
## Extracting Parallel Sentences with Bidirectional Recurrent Neural $\textbf{Networks to Improve Machine Translation}$
Francis Grégoire
Philippe Langlais
Université de Montréal
Université de Montréal
francis.gregoire@rd.mila.quebec felipe@iro.umontreal.ca
## Abstract
Parallel sentence extraction is a task addressing the data sparsity problem found in multilingual natural language processing applications. We propose a bidirectional recurrent neural network based approach to extract parallel sentences from collections of multilingual texts. Our experiments with noisy parallel corpora show that we can achieve promising results against a competitive baseline by removing the need of specific feature engineering or additional external resources. To justify the utility of our approach, we extract sentence pairs from Wikipedia articles to train machine translation systems and show significant improvements in translation performance.
## 1 Introduction
Parallel corpora are a prerequisite for many multilingual natural language processing (NLP) applications. Unfortunately, parallel data is available for a relatively small number of language pairs and for few specific domains. With the increasing amount of comparable corpora on the World Wide Web, a potential solution to alleviate the parallel data sparsity issue is to extract parallel sentences from this more abundant source of information. Therefore, the objective of parallel sentence extraction is to extract parallel sentences from such comparable corpora to increase the quantity of parallel data and the range of the covered domains. Comparable corpora can be defined as collections of topic-aligned but non-sentence-aligned multilingual texts.
Recent advances in deep learning architectures with recurrent neural networks (RNNs) have shown that they can successfully learn complex mappings from variable-length sequences to continuous vector representations. While numerous natural language processing tasks have successfully applied those models, ranging from handwriting generation (Graves, 2013) to machine comprehension (Hermann et al., 2015), most of the multilingual efforts have been devoted to machine translation (Sutskever et al., 2014; Cho et al., 2014), although more research interests have been recently devoted to multilingual semantic textual similarity.$^{1}$
In this paper, we propose a parallel sentence extraction system to measure the translational equivalence between sentences in two languages. Our system is based on bidirectional recurrent neural networks that can learn sentence representations in a shared vector space by explicitly maximizing the similarity between parallel sentences. In contrast to previous approaches, by leveraging these continuous vector representation of sentences we remove the need to rely on multiple models and specific feature engineering. Experiments on noisy parallel corpora show that our approach outperforms a competitive baseline. To justify the utility of our approach, we add the sentence pairs extracted from Wikipedia articles to a parallel corpus to train machine translation systems and show improvements in translation performance. Our experimental results lead us to believe that our system is a promising tool to create new aligned multilingual resources.
This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://
creativecommons.org/licenses/by/4.0/
$^{1}$http://alt.qcri.org/semeval2017/task2/
|
[
[
0.162841796875,
0.08197021484375,
0.8388671875,
0.08197021484375,
0.8388671875,
0.0968017578125,
0.162841796875,
0.0968017578125
],
[
0.2822265625,
0.10394287109375,
0.7197265625,
0.10394287109375,
0.7197265625,
0.1175537109375,
0.2822265625,
0.1175537109375
],
[
0.26123046875,
0.146728515625,
0.408935546875,
0.146728515625,
0.408935546875,
0.160888671875,
0.26123046875,
0.160888671875
],
[
0.59033203125,
0.1455078125,
0.74462890625,
0.1455078125,
0.74462890625,
0.163330078125,
0.59033203125,
0.163330078125
],
[
0.2392578125,
0.162109375,
0.429931640625,
0.163330078125,
0.429931640625,
0.1776123046875,
0.2392578125,
0.177001953125
],
[
0.5751953125,
0.163330078125,
0.76416015625,
0.163330078125,
0.76416015625,
0.1776123046875,
0.5751953125,
0.1776123046875
],
[
0.145263671875,
0.179931640625,
0.8076171875,
0.18115234375,
0.8076171875,
0.1953125,
0.145263671875,
0.19482421875
],
[
0.4609375,
0.243408203125,
0.5439453125,
0.2452392578125,
0.5439453125,
0.26318359375,
0.4609375,
0.26123046875
],
[
0.14697265625,
0.28076171875,
0.85205078125,
0.28076171875,
0.85205078125,
0.295654296875,
0.14697265625,
0.295654296875
],
[
0.019317626953125,
0.2998046875,
0.07135009765625,
0.2998046875,
0.07135009765625,
0.440673828125,
0.019317626953125,
0.440673828125
],
[
0.145263671875,
0.295654296875,
0.85400390625,
0.29443359375,
0.85400390625,
0.312255859375,
0.145263671875,
0.3134765625
],
[
0.1485595703125,
0.3134765625,
0.85205078125,
0.3134765625,
0.85205078125,
0.32666015625,
0.1485595703125,
0.32666015625
],
[
0.1485595703125,
0.330078125,
0.8564453125,
0.330078125,
0.8564453125,
0.34326171875,
0.1485595703125,
0.34326171875
],
[
0.1485595703125,
0.345703125,
0.85400390625,
0.345703125,
0.85400390625,
0.35986328125,
0.1485595703125,
0.35986328125
],
[
0.14697265625,
0.3623046875,
0.85546875,
0.3623046875,
0.85546875,
0.3759765625,
0.14697265625,
0.3759765625
],
[
0.14697265625,
0.377685546875,
0.85546875,
0.377685546875,
0.85546875,
0.391357421875,
0.14697265625,
0.391357421875
],
[
0.1485595703125,
0.3955078125,
0.246826171875,
0.3955078125,
0.246826171875,
0.40673828125,
0.1485595703125,
0.40673828125
],
[
0.139404296875,
0.427001953125,
0.26025390625,
0.427001953125,
0.26025390625,
0.440673828125,
0.139404296875,
0.440673828125
],
[
0.0218353271484375,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.12005615234375,
0.454833984375,
0.8818359375,
0.454833984375,
0.8818359375,
0.4697265625,
0.12005615234375,
0.4697265625
],
[
0.12005615234375,
0.471435546875,
0.88671875,
0.471435546875,
0.88671875,
0.486328125,
0.12005615234375,
0.486328125
],
[
0.118408203125,
0.488037109375,
0.884765625,
0.488037109375,
0.884765625,
0.501953125,
0.118408203125,
0.501953125
],
[
0.12005615234375,
0.50439453125,
0.88671875,
0.50439453125,
0.88671875,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.019317626953125,
0.5185546875,
0.060455322265625,
0.5185546875,
0.06378173828125,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.12005615234375,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.12005615234375,
0.533203125
],
[
0.12005615234375,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.54736328125,
0.12005615234375,
0.54736328125
],
[
0.12005615234375,
0.5517578125,
0.88330078125,
0.5517578125,
0.88330078125,
0.5654296875,
0.12005615234375,
0.5654296875
],
[
0.118408203125,
0.568359375,
0.414794921875,
0.568359375,
0.414794921875,
0.58203125,
0.118408203125,
0.58203125
],
[
0.1368408203125,
0.5849609375,
0.88330078125,
0.5849609375,
0.88330078125,
0.5986328125,
0.1368408203125,
0.5986328125
],
[
0.12005615234375,
0.60107421875,
0.87890625,
0.60107421875,
0.87890625,
0.6142578125,
0.12005615234375,
0.6142578125
],
[
0.12005615234375,
0.61767578125,
0.8818359375,
0.61767578125,
0.8818359375,
0.6318359375,
0.12005615234375,
0.6318359375
],
[
0.12005615234375,
0.63427734375,
0.88330078125,
0.63427734375,
0.88330078125,
0.64794921875,
0.12005615234375,
0.64794921875
],
[
0.12005615234375,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.6630859375,
0.12005615234375,
0.6630859375
],
[
0.118408203125,
0.66455078125,
0.88330078125,
0.6650390625,
0.88330078125,
0.6796875,
0.118408203125,
0.6787109375
],
[
0.1175537109375,
0.68115234375,
0.32421875,
0.68115234375,
0.32421875,
0.69482421875,
0.1175537109375,
0.69482421875
],
[
0.1368408203125,
0.69970703125,
0.88330078125,
0.69970703125,
0.88330078125,
0.71337890625,
0.1368408203125,
0.71337890625
],
[
0.118408203125,
0.71533203125,
0.88330078125,
0.71533203125,
0.88330078125,
0.728515625,
0.118408203125,
0.728515625
],
[
0.118408203125,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.118408203125,
0.7490234375,
0.88330078125,
0.7490234375,
0.88330078125,
0.7626953125,
0.118408203125,
0.7626953125
],
[
0.12005615234375,
0.76318359375,
0.880859375,
0.76318359375,
0.880859375,
0.77685546875,
0.12005615234375,
0.77685546875
],
[
0.12005615234375,
0.77978515625,
0.88330078125,
0.77978515625,
0.88330078125,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.118408203125,
0.79638671875,
0.88330078125,
0.79638671875,
0.88330078125,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.12005615234375,
0.8125,
0.880859375,
0.8125,
0.880859375,
0.826171875,
0.12005615234375,
0.826171875
],
[
0.118408203125,
0.82763671875,
0.88330078125,
0.82666015625,
0.88330078125,
0.84130859375,
0.118408203125,
0.8427734375
],
[
0.12005615234375,
0.845703125,
0.2880859375,
0.845703125,
0.2880859375,
0.85546875,
0.12005615234375,
0.85546875
],
[
0.124267578125,
0.87060546875,
0.8818359375,
0.87060546875,
0.8818359375,
0.88427734375,
0.124267578125,
0.88427734375
],
[
0.12176513671875,
0.88525390625,
0.447509765625,
0.88525390625,
0.447509765625,
0.8955078125,
0.12176513671875,
0.8955078125
],
[
0.14111328125,
0.896484375,
0.37451171875,
0.8955078125,
0.37451171875,
0.90673828125,
0.14111328125,
0.90771484375
]
] |
[
"Extracting Parallel Sentences with Bidirectional Recurrent Neural",
"Networks to ImproveMachine Translation",
"Francis Gregoire",
"Philippe Langlais",
"Université de Montréal",
"Université de Montreal",
"francis.gregoire@rd.mila.guebec felipe@iro.umontreal.ca",
"Abstract",
"Parallel sentence extraction is a task addressing the data sparsity problem found in multilin-",
"8102 ",
"gual natural language processing applications. We propose a bidirectional recurrent neural net-",
"work based approach to extract parallel sentences from collections of multilingual texts. Our",
"experiments with noisy parallel corpora show that we can achieve promising results against a",
"competitive baseline by removing the need of specific feature engineering or additional exter-",
"nal resources. To justify the utility of our approach, we extract sentence pairs from Wikipedia",
"articles to train machine translation systems and show significant improvements in translation",
"performance.",
"Introduction",
"[T']",
"Parallel corpora are a prerequisite for many multilingual natural language processing (NLP) applica-",
"tions. Unfortunately, parallel data is available for a relatively small number of language pairs and for",
"few specific domains. With the increasing amount of comparable corpora on the World Wide Web, a",
"potential solution to alleviate the parallel data sparsity issue is to extract parallel sentences from this",
"60509081A0",
"more abundant source of information. Therefore, the objective of parallel sentence extraction is to ex-",
"tractparallel sentencesfrom suchcomparablecorporato increasetheguantity of paralleldata andthe",
"range of the covered domains. Comparable corpora can be defined as collections of topic-aligned but",
"non-sentence-aligned multilingual texts.",
"Recent advances in deep learning architectures with recurrent neural networks (RNNs) have shown",
"that they can successfully learn complex mappings from variable-length sequences to continuous vec-",
"tor representations. While numerous natural language processing tasks have successfully applied those",
"models, ranging from handwriting generation (Graves, 2013) to machine comprehension (Hermann et",
"al., 2015), most of the multilingual efforts have been devoted to machine translation (Sutskever et al.,",
"2014; Cho et al., 2014), although more research interests have been recently devoted to multilingual",
" semantic textual similarity.!",
"In this paper, we propose a parallel sentence extraction system to measure the translational equivalence",
"between sentences in two languages. Our system is based on bidirectional recurrent neural networks that",
"can learn sentence representations in a shared vector space by explicitly maximizing the similarity be-",
"tween parallel sentences. In contrast to previous approaches, by leveraging these continuous vector",
"representation of sentences we remove the need to rely on multiple models and specific feature engineer-",
"ing. Experiments on noisy parallel corpora show that our approach outperforms a competitive baseline.",
"To justify the utility of our approach, we add the sentence pairs extracted from Wikipedia articles to a",
"parallel corpus to train machine translation systems and show improvements in translation performance",
"Our experimental results lead us to believe that our system is a promising tool to create new aligned",
"multilingualresources.",
"This work is licensed under a Creative Commons Attribution 4.0 International License.License details: http: //",
"creativecommons.org/licenses/by/4.0/",
"'http://alt.qcri.org/semeval2017/task2/"
] |
{
"arxiv_id": "1806.05559",
"language": "en",
"page_idx": 0
}
|
||
1806.05559-en-1
|
1806.05559-en-1.jpg
|
## 2 $\textbf{Related work}$
Traditional systems developed to extract parallel sentences from comparable corpora typically rely on multiples models or metadata from articles structure. Munteanu and Marcu (2005) present a complete system based on statistical word alignment models and a maximum entropy classifier to automatically extract sentence pairs from newspaper collections. The authors evaluate the quality of the extracted sentences by showing that they improve the performance of statistical machine translation (SMT) systems The overall structure of this approach is still considered state-of-the-art. Adafre and Rijke (2006) is the first work to observe that articles from Wikipedia are likely to generate parallel corpora useful for machine translation. These two approaches are extended by Smith et al. (2010) where the authors introduce several new features by exploiting the structure and metadata of Wikipedia article pairs. They use their augmented set of features in a conditional random field and obtain state-of-the-art results on a small set of 20 manually annotated Wikipedia article pairs. The work of Abdul-Rauf and Schwenk (2009) proposes a different approach, in which they use an SMT system built from a small parallel corpus. Instead of using a classifier, they translate the source language side of a comparable corpus to find candidate sentences on the target language side. They determine if a translated source sentence and a candidate target sentence are parallel by measuring the word error rate and the translation error rate. A simplification of these systems is proposed in (Azpeitia et al., 2017), where the similarity between two sentences is defined as the average of the Jaccard similarity coefficients obtained between sentence token sets and lexical translations determined by IBM models (Brown et al., 1993). Their approach also combine various features, such as longest common prefix matching, numbers and capitalized truecased tokens.
Chu et al. (2016) train a neural machine translation (NMT) system to generate sentence representations with the encoder, which are then used as additional features to the system of Munteanu and Marcu (2005). Similarly, Cristina et al.(2017) study sentence representations obtained from the encoder of an NMT system to detect new parallel sentence pairs. By comparing cosine similarities, they show that they can distinguish parallel and non-parallel sentences. A different approach exploiting continuous vector representations is proposed in Grover and Mitra (2017). After learning word representations using the bilingual word embeddings model of (Luong et al., 2015), they use a convolutional neural network on a similarity matrix to classify if a pair of sentences is aligned or not. These approaches are different from ours, where we use a single end-to-end model to estimate the conditional probability distribution that two sentences are parallel.
# 3 Approach
# 3.1 Negative sampling
As positive examples, we use a parallel corpus $C$ consisting of $n$ parallel sentence pairs $\left\{(\mathbf{s}_k^S,\mathbf{s}_k^T)\right\}_{k=1}^n$, where $S$ and $T$ denote the source sentences and target sentences. Since we want a model that learns differentiable vector representations to distinguish parallel from non-parallel sentences, we use negative sampling to generate negative examples. Therefore, we randomly sample $m$ non-parallel target sentences for every positive source sentence, such that $(\mathbf{s}_k^S,\mathbf{s}_j^T)$ for $j\neq k.$ This process is repeated at the beginning of each training epoch to allow the model to learn on a larger number of non-parallel sentence pairs. Hence, our training set contains $n(1+m)$ triples $(\mathbf{s}_i^S,\mathbf{s}_i^T,y_i)$, where $\mathbf{s}_i^S=(w_{i,1}^S,\ldots,w_{i,N}^S)$ is a source sentence of $N$ tokens, $\mathbf{s}_i^T=(w_{i,1}^T,\ldots,w_{i,M}^T)$ is a target sentence of $M$ tokens, and $y_i$ is the label representing the translation relationship between $\mathbf{s}_i^S$ and $\mathbf{s}_i^T$, so that $y_i=1$ if $(\mathbf{s}_i^S,\mathbf{s}_i^T)\in C$ and $y_i=0$ otherwise.
The advantage of negative sampling is its simplicity. However, relying only on randomness to select negative sentence pairs makes most of the examples very non-parallel and easy to classify. An interesting way to generate negative examples would be to replace only a segment of a sentence. Similarly, we could replace a sentence from a parallel pair with another sentence that is close to it in the vector space. This would make the problem harder, but potentially could make the classifier stronger. We leave such investigations as future work.
|
[
[
0.1158447265625,
0.0736083984375,
0.26611328125,
0.07598876953125,
0.26611328125,
0.09027099609375,
0.1158447265625,
0.0885009765625
],
[
0.118408203125,
0.105712890625,
0.88330078125,
0.105712890625,
0.88330078125,
0.119384765625,
0.118408203125,
0.119384765625
],
[
0.12005615234375,
0.122314453125,
0.88330078125,
0.122314453125,
0.88330078125,
0.135986328125,
0.12005615234375,
0.135986328125
],
[
0.118408203125,
0.1383056640625,
0.88330078125,
0.1383056640625,
0.88330078125,
0.152587890625,
0.118408203125,
0.152587890625
],
[
0.12005615234375,
0.15380859375,
0.8818359375,
0.15380859375,
0.8818359375,
0.16748046875,
0.12005615234375,
0.16748046875
],
[
0.12005615234375,
0.169189453125,
0.8818359375,
0.169189453125,
0.8818359375,
0.18408203125,
0.12005615234375,
0.18408203125
],
[
0.12005615234375,
0.18701171875,
0.884765625,
0.18701171875,
0.884765625,
0.20068359375,
0.12005615234375,
0.20068359375
],
[
0.12176513671875,
0.2037353515625,
0.87890625,
0.2037353515625,
0.87890625,
0.21435546875,
0.12176513671875,
0.21435546875
],
[
0.12005615234375,
0.218505859375,
0.88330078125,
0.218505859375,
0.88330078125,
0.232177734375,
0.12005615234375,
0.232177734375
],
[
0.12005615234375,
0.235107421875,
0.88330078125,
0.235107421875,
0.88330078125,
0.248779296875,
0.12005615234375,
0.248779296875
],
[
0.12005615234375,
0.25048828125,
0.88330078125,
0.25048828125,
0.88330078125,
0.26416015625,
0.12005615234375,
0.26416015625
],
[
0.12176513671875,
0.268310546875,
0.8818359375,
0.268310546875,
0.8818359375,
0.279052734375,
0.12176513671875,
0.279052734375
],
[
0.12005615234375,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.12005615234375,
0.2998046875,
0.8818359375,
0.2998046875,
0.8818359375,
0.3134765625,
0.12005615234375,
0.3134765625
],
[
0.12005615234375,
0.31640625,
0.88330078125,
0.31640625,
0.88330078125,
0.330078125,
0.12005615234375,
0.330078125
],
[
0.12005615234375,
0.33203125,
0.88330078125,
0.33203125,
0.88330078125,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.118408203125,
0.347412109375,
0.88330078125,
0.3466796875,
0.88330078125,
0.361083984375,
0.118408203125,
0.3623046875
],
[
0.118408203125,
0.364013671875,
0.88330078125,
0.364013671875,
0.88330078125,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.118408203125,
0.380126953125,
0.88330078125,
0.380126953125,
0.88330078125,
0.393798828125,
0.118408203125,
0.393798828125
],
[
0.118408203125,
0.3955078125,
0.7421875,
0.3955078125,
0.7421875,
0.410400390625,
0.118408203125,
0.410400390625
],
[
0.1351318359375,
0.412109375,
0.88330078125,
0.413330078125,
0.88330078125,
0.427490234375,
0.1351318359375,
0.427001953125
],
[
0.118408203125,
0.4287109375,
0.8818359375,
0.429931640625,
0.8818359375,
0.44482421875,
0.118408203125,
0.443603515625
],
[
0.12005615234375,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.12005615234375,
0.460205078125
],
[
0.118408203125,
0.4619140625,
0.88330078125,
0.4619140625,
0.88330078125,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.12005615234375,
0.478515625,
0.88330078125,
0.478515625,
0.88330078125,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.12005615234375,
0.49462890625,
0.88330078125,
0.49462890625,
0.88330078125,
0.50830078125,
0.12005615234375,
0.50830078125
],
[
0.118408203125,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.118408203125,
0.52392578125
],
[
0.12005615234375,
0.52685546875,
0.8818359375,
0.52685546875,
0.8818359375,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.118408203125,
0.5419921875,
0.88330078125,
0.5419921875,
0.88330078125,
0.5556640625,
0.118408203125,
0.5556640625
],
[
0.1175537109375,
0.55859375,
0.316650390625,
0.55859375,
0.316650390625,
0.572265625,
0.1175537109375,
0.572265625
],
[
0.1158447265625,
0.59033203125,
0.2376708984375,
0.59033203125,
0.2376708984375,
0.60791015625,
0.1158447265625,
0.60791015625
],
[
0.118408203125,
0.62255859375,
0.307373046875,
0.62255859375,
0.307373046875,
0.63720703125,
0.118408203125,
0.63720703125
],
[
0.11419677734375,
0.6435546875,
0.8818359375,
0.6455078125,
0.8818359375,
0.66650390625,
0.11419677734375,
0.66455078125
],
[
0.12005615234375,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.12005615234375,
0.6767578125
],
[
0.118408203125,
0.68115234375,
0.884765625,
0.68115234375,
0.884765625,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.12005615234375,
0.69677734375,
0.884765625,
0.69677734375,
0.884765625,
0.71044921875,
0.12005615234375,
0.71044921875
],
[
0.12005615234375,
0.7119140625,
0.88330078125,
0.7119140625,
0.88330078125,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.12005615234375,
0.728515625,
0.88330078125,
0.728515625,
0.88330078125,
0.7421875,
0.12005615234375,
0.7421875
],
[
0.11419677734375,
0.7392578125,
0.88671875,
0.74365234375,
0.88671875,
0.76318359375,
0.11419677734375,
0.7587890625
],
[
0.1158447265625,
0.76025390625,
0.88671875,
0.76025390625,
0.88671875,
0.78076171875,
0.1158447265625,
0.78076171875
],
[
0.118408203125,
0.77978515625,
0.88330078125,
0.77978515625,
0.88330078125,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.118408203125,
0.79736328125,
0.198974609375,
0.79736328125,
0.198974609375,
0.80859375,
0.118408203125,
0.80859375
],
[
0.1368408203125,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.8291015625,
0.1368408203125,
0.8291015625
],
[
0.1175537109375,
0.83056640625,
0.8818359375,
0.830078125,
0.8818359375,
0.84423828125,
0.1175537109375,
0.845703125
],
[
0.12005615234375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.118408203125,
0.87744140625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.118408203125,
0.892578125
],
[
0.1175537109375,
0.8955078125,
0.3359375,
0.8955078125,
0.3359375,
0.9091796875,
0.1175537109375,
0.9091796875
]
] |
[
"2Related work",
"Traditional systems developed to extract parallel sentences from comparable corpora typically rely on",
"multiples models or metadata from articles structure. Munteanu and Marcu (2005) present a complete",
"system based on statistical word alignment models and a maximum entropy classifier to automatically",
"extract sentence pairs from newspaper collections. The authors evaluate the quality of the extracted sen-",
"tences by showing that they improve the performance of statistical machine translation (SMT) systems.",
"The overall structure of this approach is still considered state-of-the-art. Adafre and Rijke (2006) is the",
"first work to observe that articles from Wikipedia are likely to generate parallel corpora useful for ma-",
"chine translation. These two approaches are extended by Smith et al. (2010) where the authors introduce",
"several new features by exploiting the structure and metadata of Wikipedia article pairs. They use their",
"augmented set of features in a conditional random field and obtain state-of-the-art results on a small set of",
"20manuallyannotatedWikipediaarticlepairs.TheworkofAbdul-RaufandSchwenk(2O09)proposesa",
"different approach, in which they use an SMT system built from a small parallel corpus. Instead of using",
"a classifier, they translate the source language side of a comparable corpus to find candidate sentences on",
"the target language side. They determine if a translated source sentence and a candidate target sentence",
"are parallel by measuring the word error rate and the translation error rate. A simplification of these",
"systems is proposed in (Azpeitia et al., 2017), where the similarity between two sentences is defined as",
"the average of the Jaccard similarity coefficients obtained between sentence token sets and lexical trans-",
"lations determined by IBM models (Brown et al., 1993). Their approach also combine various features,",
"such as longest common prefix matching, numbers and capitalized truecased tokens.",
"Chu et al. (2016) train a neural machine translation (NMT) system to generate sentence representations",
"with the encoder, which are then used as additional features to the system of Munteanu and Marcu (2005).",
"Similarly, Cristina et al. (2017) study sentence representations obtained from the encoder of an NMT",
"system to detect new parallel sentence pairs. By comparing cosine similarities, they show that they",
"can distinguish parallel and non-parallel sentences. A different approach exploiting continuous vector",
"representations is proposed in Grover and Mitra (2017). After learning word representations using the",
"bilingual word embeddings model of (Luong et al., 2015), they use a convolutional neural network on a",
"similarity matrix to classify if a pair of sentences is aligned or not. These approaches are different from",
"ours, where we use a single end-to-end model to estimate the conditional probability distribution that",
"two sentences are parallel.",
"3Approach",
"3.1 Negative sampling",
"As positive examples, we use a parallel corpus C consisting of n parllel sentence pairs (s, sT))\"=1",
"where S and T denote the source sentences and target sentences. Since we want a model that learns",
"differentiable vector representations to distinguish parallel from non-parallel sentences, we use negative",
"sampling to generate negative examples. Therefore, we randomly sample m non-parallel target sentences",
"for every positive source sentence, such that (s, s) for j + k. This process is repeated at the beginning",
"of each training epoch to allow the model to learn on a larger number of non-parallel sentence pairs.",
"Hence, our training set contains n(1 + m) triples (s,sT , i), where s? = (we,1,..., we,N) is a source",
"sentence of N tokens, sT = (wf,i,..., wl,M) is a target sentence of M tokens, and yi is the label",
"representing the translation relationship between ss and sT, so that yi = 1 if (s?, sT) E C and yi = O",
"otherwise.",
"The advantage of negative sampling is its simplicity. However, relying only on randomness to select",
"negative sentence pairs makes most of the examples very non-parallel and easy to classify. An interesting",
"way to generate negative examples would be to replace only a segment of a sentence. Similarly, we",
"could replace a sentence from a parallel pair with another sentence that is close to it in the vector space.",
"This would make the problem harder, but potentially could make the classifier stronger. We leave such",
"investigations as future work."
] |
{
"arxiv_id": "1806.05559",
"language": "en",
"page_idx": 1
}
|
||
1806.05559-en-2
|
1806.05559-en-2.jpg
|
<FigureHere>
Figure 1: Graphical illustration of the siamese bidirectional recurrent neural networks. The fınal recurrent state of the forward and backward RNNs are concatenated and then fed into a feed-forward neural network culminating in a sigmoid output layer.
### 3.2 $\mathbf{Model}$
Our idea is to use RNNs to learn cross-language semantics between sentence pairs to estimate the probability that they are translations of each other. The proposed model architecture consists of a bidirectional RNNs (BiRNN) (Schuster and Paliwal, 1997) sentence encoder with recurrent activation functions such as long short-term memory units (LSTM) (Hochreiter and Schmidhuber, 1997) or gated recurrent units (GRU) ( Cho et al. , 2014) . Since we want vector representations in a shared vector space we use a siamese network (Bromley et al., 1993) with tied weights, which is equivalent to using a single BiRNN to encode a pair of sentences into two continuous vector representations. The source and target sentence representations are then fed into a feed-forward neural network with a sigmoid output layer which calculates the probability that they are parallel. The architecture of our approach is illustrated in Figure 1.
To avoid repetition and for clarity, we only define equations of the BiRNN encoding the source sentence. For the target sentence, simply substitute $S$ for $T.$ At each time step $t$, the token in the $i$-th sentence, $w_i, t^S$,defined by its integer index $k$ in the vocabulary $V^S$, is represented as a one-hot vector, $\mathbf{x}_k^S\in\{0,1\}^{|V|}.$ This one-hot vector is multiplied with an embedding matrix, $\mathbf{E}^S\in\mathbb{R}^{|V^S|\times d_e}$, to get a continuous vector representation of this token, $\mathbf{w}_i,t^S\in\mathbb{R}^{d_e}$, which serves as an input for the forward and backward recurrent states of the BiRNN encoder, $\overrightarrow{\mathbf{h}}_{i,t}^S$ and $\overleftarrow{\mathbf{h}}_{i,t}^S.$ The forward RNN reads the variable-length sentence and updates its recurrent state from the first token until the last one to create a fixed-size continuous vector representation of the sentence, $\mathbf{h}_{i,N}^S\in\mathbb{R}^{d_h}.$ The backward RNN processes the sentence in reverse. We use the concatenation of the last recurrent state in both directions as a final representation $\mathbf{h}_i^S=[\overrightarrow{\mathbf{h}}_{i,N}^S;\overline{\mathbf{h}}_{i,1}^S].^2$ The steps we described to encode a sentence are defined as:
(1)
$$\mathbf{w}_{i,t}^{S}=\mathbf{E}^{S\top}\mathbf{w}_{k}^{S},\\\overrightarrow{\mathbf{h}}_{i,t}^{S}=\phi(\overrightarrow{\mathbf{h}}_{i,t-1}^{S},\mathbf{w}_{i,t}^{S}),\\\overleftarrow{\mathbf{h}}_{i,t}^{S}=\phi(\overleftarrow{\mathbf{h}}_{i,t+1}^{S},\mathbf{w}_{i,t}^{S}),$$
(2)
(3)
where $\phi(\cdot)$ is an LSTM or GRU.
After both source and target sentences have been encoded, we capture their matching information by using their element-wise product and absolute element-wise difference. We estimate the conditional ${^{2}\text{We consider}}$ed combining the recurrent states with mean pooling and max pooling to obtain a fixed-size vector representa-
tion, but obtained inferior performance.
|
[
[
0.43994140625,
0.11517333984375,
0.5634765625,
0.11517333984375,
0.5634765625,
0.1258544921875,
0.43994140625,
0.1258544921875
],
[
0.12005615234375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.12005615234375,
0.350830078125
],
[
0.118408203125,
0.352783203125,
0.88330078125,
0.352783203125,
0.88330078125,
0.366455078125,
0.118408203125,
0.366455078125
],
[
0.118408203125,
0.37060546875,
0.46337890625,
0.37060546875,
0.46337890625,
0.38427734375,
0.118408203125,
0.38427734375
],
[
0.118408203125,
0.40625,
0.2149658203125,
0.40625,
0.2149658203125,
0.42041015625,
0.118408203125,
0.42041015625
],
[
0.118408203125,
0.427001953125,
0.880859375,
0.427490234375,
0.880859375,
0.4423828125,
0.118408203125,
0.441162109375
],
[
0.118408203125,
0.44482421875,
0.88330078125,
0.44482421875,
0.88330078125,
0.457763671875,
0.118408203125,
0.457763671875
],
[
0.118408203125,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.118408203125,
0.473876953125
],
[
0.118408203125,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.490478515625,
0.118408203125,
0.490478515625
],
[
0.12005615234375,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.505859375,
0.12005615234375,
0.505859375
],
[
0.12005615234375,
0.5087890625,
0.88330078125,
0.5087890625,
0.88330078125,
0.5224609375,
0.12005615234375,
0.5224609375
],
[
0.118408203125,
0.52490234375,
0.8818359375,
0.52490234375,
0.8818359375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.118408203125,
0.54052734375,
0.8818359375,
0.54052734375,
0.8818359375,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.118408203125,
0.55712890625,
0.791015625,
0.55712890625,
0.791015625,
0.57177734375,
0.118408203125,
0.57177734375
],
[
0.1351318359375,
0.572265625,
0.8818359375,
0.57177734375,
0.8818359375,
0.5859375,
0.1351318359375,
0.58740234375
],
[
0.118408203125,
0.58837890625,
0.8818359375,
0.58837890625,
0.8818359375,
0.6015625,
0.118408203125,
0.6015625
],
[
0.12005615234375,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.118408203125,
0.62158203125,
0.884765625,
0.62158203125,
0.884765625,
0.6396484375,
0.118408203125,
0.6396484375
],
[
0.1175537109375,
0.64013671875,
0.884765625,
0.63818359375,
0.884765625,
0.65625,
0.1175537109375,
0.65771484375
],
[
0.118408203125,
0.65771484375,
0.884765625,
0.65771484375,
0.884765625,
0.67578125,
0.118408203125,
0.67578125
],
[
0.12005615234375,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.12005615234375,
0.6904296875
],
[
0.118408203125,
0.693359375,
0.88330078125,
0.6923828125,
0.88330078125,
0.70703125,
0.118408203125,
0.7080078125
],
[
0.118408203125,
0.71044921875,
0.88330078125,
0.71044921875,
0.88330078125,
0.7236328125,
0.118408203125,
0.7236328125
],
[
0.1175537109375,
0.724609375,
0.8369140625,
0.724609375,
0.8369140625,
0.7421875,
0.1175537109375,
0.7421875
],
[
0.46337890625,
0.75,
0.55908203125,
0.7529296875,
0.5576171875,
0.771484375,
0.462646484375,
0.7685546875
],
[
0.444091796875,
0.7587890625,
0.47021484375,
0.75732421875,
0.470947265625,
0.765625,
0.44580078125,
0.76708984375
],
[
0.8564453125,
0.7548828125,
0.88671875,
0.7548828125,
0.88671875,
0.771484375,
0.8564453125,
0.771484375
],
[
0.41064453125,
0.7724609375,
0.591796875,
0.775390625,
0.591796875,
0.79736328125,
0.41064453125,
0.79443359375
],
[
0.8564453125,
0.77783203125,
0.88671875,
0.77783203125,
0.88671875,
0.7958984375,
0.8564453125,
0.7958984375
],
[
0.41064453125,
0.79638671875,
0.591796875,
0.7998046875,
0.591796875,
0.82177734375,
0.41064453125,
0.818359375
],
[
0.8564453125,
0.8017578125,
0.88671875,
0.8017578125,
0.88671875,
0.818359375,
0.8564453125,
0.818359375
],
[
0.118408203125,
0.8291015625,
0.359375,
0.8291015625,
0.359375,
0.8427734375,
0.118408203125,
0.8427734375
],
[
0.1376953125,
0.845703125,
0.880859375,
0.845703125,
0.880859375,
0.859375,
0.1376953125,
0.859375
],
[
0.118408203125,
0.86083984375,
0.88330078125,
0.86083984375,
0.88330078125,
0.87451171875,
0.118408203125,
0.87451171875
],
[
0.139404296875,
0.8818359375,
0.8818359375,
0.8828125,
0.8818359375,
0.89794921875,
0.139404296875,
0.896484375
],
[
0.118408203125,
0.8955078125,
0.35693359375,
0.896484375,
0.35693359375,
0.90771484375,
0.118408203125,
0.90673828125
]
] |
[
"feed-forward neural network",
"Figure l: Graphical illustration of the siamese bidirectional recurrent neural networks. The final recur-",
"rent state of the forward and backward RNNs are concatenated and then fed into a feed-forward neural",
"network culminating in a sigmoid output layer.",
"3.2Model",
"Our idea is to use RNNs to learn cross-language semantics between sentence pairs to estimate the proba-",
"bility that they are translations of each other. The proposed model architecture consists of a bidirectional",
"RNNs (BiRNN) (Schuster and Paliwal, 1997) sentence encoder with recurrent activation functions such",
"as long short-term memory units (LSTM) (Hochreiter and Schmidhuber, 1997) or gated recurrent units",
"(GRU) (Cho et al., 2014). Since we want vector representations in a shared vector space we use a siamese",
"network (Bromley et al., 1993) with tied weights, which is equivalent to using a single BiRNN to encode",
"a pair of sentences into two continuous vector representations. The source and target sentence represen-",
"tations are then fed into a feed-forward neural network with a sigmoid output layer which calculates the",
"probability that they are parallel. The architecture of our approach is illustrated in Figure 1.",
"To avoid repetition and for clarity, we only define equations of the BiRNN encoding the source sen-",
"tence. For the target sentence, simply substitute S for T. At each time step t, the token in the i-th",
"sentence, wst, defined by its integer index k in the vocabulary Vs, is represented as a one-hot vec-",
"tor, x e [0, 1]lVl. This one-hot vector is multiplied with an embedding matrix, Es e R/vs[xde, to",
"get a continuous vector representation of this token, w2t e IRde, which serves as an input for the for-",
"ward and backward recurrent states of the BiRNN encoder, $ and hSt. The forward RNN reads the",
"variable-length sentence and updates its recurrent state from the first token until the last one to create a",
"fixed-size continuous vector representation of the sentence, hn E Rdh.The backward RNN processes",
"the sentence in reverse. We use the concatenation of the last recurrent state in both directions as a final",
"representation hs = [h ; h]? The steps we described to encode a sentence are defined as:",
"=ESTwS",
"wS",
"(1)",
"ht = o(hst-1, wst),",
"(2)",
"hst = s(hst+1, wet),",
"(3)",
"where @() is an LSTM or GRU.",
"After both source and target sentences have been encoded, we capture their matching information",
"by using their element-wise product and absolute element-wise difference. We estimate the conditional",
"2we considered combining the recurrent states with mean pooling and max pooling to obtain a fixed-size vector representa-",
"tion,but obtained inferiorperformance"
] |
{
"arxiv_id": "1806.05559",
"language": "en",
"page_idx": 2
}
|
||
1806.05559-en-3
|
1806.05559-en-3.jpg
|
probability that the sentences are parallel by feeding the matching vectors into a feed-forward neural
network with a sigmoid output layer:
(4)
(5)
$$\begin{aligned}\mathbf{h}_{i}^{(1)}&=\mathbf{h}_{i}^{S}\odot\mathbf{h}_{i}^{T},\\\mathbf{h}_{i}^{(2)}&=|\mathbf{h}_{i}^{S}-\mathbf{h}_{i}^{T}|,\\\mathbf{h}_{i}&=\mathrm{tanh}(\mathbf{W}^{(1)}\mathbf{h}_{i}^{(1)}+\mathbf{W}^{(2)}\mathbf{h}_{i}^{(2)}+\mathbf{b}),\\&p(y_{i}=1|\mathbf{h}_{i})=\sigma(\mathbf{v}\mathbf{h}_{i}+b),\end{aligned}$$
(6)
(7)
where $\sigma(\cdot)$ is the sigmoid function, $\mathbf{W}^{(1)}\in\mathbb{R}^{d_f\times d_h},\mathbf{W}^{(2)}\in\mathbb{R}^{d_f\times d_h},\mathbf{v}\in\mathbb{R}^{d_f},\mathbf{b}\in\mathbb{R}^{d_f}$ and $b$ are
model parameters. The value $d_f$ is the size of the hidden layer in the feed-forward neural network.
The model is trained by minimizing the cross entropy of our labeled sentence pairs:
$$\mathcal{L}=-\sum_{i=1}^{n(1+m)}\Big(y_i\log\sigma(\mathbf{vh}_i+b)+(1-y_i)\log(1-\sigma(\mathbf{vh}_i+b))\Big).$$
(8)
For prediction, a sentence pair is classified as parallel if its probability is greater than or equal to a
decision threshold $\rho$ that we need to fix:
$$\hat{y}_i=\begin{cases}1&\text{if}p(y_i=1|\mathbf{h}_i)\ge\rho,\\0&\text{otherwise.}\end{cases}$$
(9)
The two embedding matrices $\mathbf{E}^S$ and $\mathbf{E}^T$ are parameters of the model that we must learn. As the size of the vocabulary and the dimension $d_e$ increase, the number of parameters can become considerably expensive to estimate. For that reason, it is common to initialize the embedding matrices using word embeddings pre-trained on a large collection of texts.$^{3}$
## 4 Experiments and Results
To assess the effectiveness of our approach, we compare it in multiple settings. In Section 4.1, we measure its capacity to extract parallel sentences found in parallel corpora in which we inserted nonparallel sentences. We extract sentence pairs from real comparable corpora in Section 4.2 and validate their utility by measuring their impact on SMT and NMT systems.
## 4.1 System comparison
4.1.1 Data
To create our training and test sets we use the WMT’15 English-French datasets.$^{4}$ The positive examples of our training set consist of 500,000 parallel sentence pairs randomly selected from the Europarl corpus (Koehn, 2005). To generate negative examples, at the beginning of each training epoch we sample $m$ non-parallel sentences per positive example, as described in Section 3.1. The vocabulary size is 69,381 for English and 84,182 for French.
The most reliable way to create test sets to compare different approaches would be to have professional translators manually annotate parallel sentences from comparable corpora. However, this option is expensive and impractical. Therefore, it is common practice to compare parallel sentence extraction systems on artificially-created noisy parallel data by inserting non-parallel sentences into a parallel corpus. Thus, to create our test sets we use parallel sentences from the newstest2012 and Europarl corpora and insert artificial noise by substituting target sentences with non-parallel target sentences from their respective held-out dataset. We sample 1,000 parallel sentences from the newstest2012 corpus and create ten test sets with a noise ratio $r\in\{0\%,10\%,\ldots,80\%,90\%\}$, where $r$ is the ratio of artificial
$^{3}$We pre-trained bilingual word embeddings as proposed by (Gouws et al., 2015) and (Smith et al., 2017). While we observed faster learning and a small improvement in the performance of our model, we prefer to learn the embedding matrices from scratch instead of relying on additional external resources.
$^{4}$http://www.statmt.org/wmt15/translation-task.html
|
[
[
0.118408203125,
0.07659912109375,
0.88330078125,
0.07598876953125,
0.88330078125,
0.09027099609375,
0.118408203125,
0.0914306640625
],
[
0.12005615234375,
0.09320068359375,
0.393798828125,
0.09320068359375,
0.393798828125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.43408203125,
0.1175537109375,
0.56494140625,
0.119384765625,
0.56494140625,
0.13720703125,
0.43408203125,
0.1353759765625
],
[
0.8564453125,
0.1217041015625,
0.88671875,
0.1217041015625,
0.88671875,
0.1383056640625,
0.8564453125,
0.1383056640625
],
[
0.429931640625,
0.141357421875,
0.57275390625,
0.141357421875,
0.57275390625,
0.1602783203125,
0.429931640625,
0.1602783203125
],
[
0.8564453125,
0.144287109375,
0.88671875,
0.144287109375,
0.88671875,
0.162109375,
0.8564453125,
0.162109375
],
[
0.34765625,
0.163330078125,
0.6513671875,
0.1644287109375,
0.6513671875,
0.185302734375,
0.34765625,
0.18408203125
],
[
0.8564453125,
0.168701171875,
0.88671875,
0.168701171875,
0.88671875,
0.185302734375,
0.8564453125,
0.185302734375
],
[
0.393798828125,
0.18701171875,
0.60546875,
0.1859130859375,
0.60546875,
0.2037353515625,
0.393798828125,
0.204833984375
],
[
0.8564453125,
0.188232421875,
0.88671875,
0.188232421875,
0.88671875,
0.204833984375,
0.8564453125,
0.204833984375
],
[
0.118408203125,
0.215576171875,
0.884765625,
0.215576171875,
0.884765625,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.12005615234375,
0.235107421875,
0.84033203125,
0.235107421875,
0.84033203125,
0.248779296875,
0.12005615234375,
0.248779296875
],
[
0.1368408203125,
0.2493896484375,
0.75244140625,
0.25048828125,
0.75244140625,
0.265380859375,
0.1368408203125,
0.26416015625
],
[
0.3046875,
0.27978515625,
0.3525390625,
0.27978515625,
0.3525390625,
0.287353515625,
0.3046875,
0.287353515625
],
[
0.35693359375,
0.29150390625,
0.74462890625,
0.290283203125,
0.74462890625,
0.30810546875,
0.35693359375,
0.308837890625
],
[
0.8564453125,
0.29150390625,
0.88818359375,
0.29150390625,
0.88818359375,
0.30810546875,
0.8564453125,
0.30810546875
],
[
0.1351318359375,
0.330810546875,
0.88330078125,
0.33203125,
0.88330078125,
0.3466796875,
0.1351318359375,
0.345703125
],
[
0.2493896484375,
0.292236328125,
0.306396484375,
0.292236328125,
0.306396484375,
0.30712890625,
0.2493896484375,
0.30712890625
],
[
0.118408203125,
0.3486328125,
0.414794921875,
0.3486328125,
0.414794921875,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.42822265625,
0.3759765625,
0.6171875,
0.3759765625,
0.6171875,
0.393798828125,
0.42822265625,
0.393798828125
],
[
0.38623046875,
0.3896484375,
0.405517578125,
0.3896484375,
0.405517578125,
0.3984375,
0.38623046875,
0.3984375
],
[
0.8564453125,
0.385498046875,
0.88671875,
0.385498046875,
0.88671875,
0.402099609375,
0.8564453125,
0.402099609375
],
[
0.429931640625,
0.396728515625,
0.54248046875,
0.396728515625,
0.54248046875,
0.410888671875,
0.429931640625,
0.410888671875
],
[
0.1351318359375,
0.42333984375,
0.88330078125,
0.424560546875,
0.88330078125,
0.4423828125,
0.1351318359375,
0.441162109375
],
[
0.12005615234375,
0.443603515625,
0.880859375,
0.443603515625,
0.880859375,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.12005615234375,
0.458984375,
0.88330078125,
0.458984375,
0.88330078125,
0.47265625,
0.12005615234375,
0.47265625
],
[
0.1175537109375,
0.474365234375,
0.51708984375,
0.473876953125,
0.51708984375,
0.488037109375,
0.1175537109375,
0.4892578125
],
[
0.1175537109375,
0.5029296875,
0.365234375,
0.5029296875,
0.365234375,
0.51708984375,
0.1175537109375,
0.51708984375
],
[
0.118408203125,
0.52783203125,
0.88330078125,
0.52783203125,
0.88330078125,
0.5419921875,
0.118408203125,
0.5419921875
],
[
0.118408203125,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.118408203125,
0.55810546875
],
[
0.118408203125,
0.56005859375,
0.88330078125,
0.56005859375,
0.88330078125,
0.57373046875,
0.118408203125,
0.57373046875
],
[
0.118408203125,
0.576171875,
0.609375,
0.576171875,
0.609375,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.1175537109375,
0.60107421875,
0.31494140625,
0.6015625,
0.31494140625,
0.61767578125,
0.1175537109375,
0.6162109375
],
[
0.118408203125,
0.62255859375,
0.2149658203125,
0.62255859375,
0.2149658203125,
0.63720703125,
0.118408203125,
0.63720703125
],
[
0.11419677734375,
0.64013671875,
0.884765625,
0.6396484375,
0.884765625,
0.65673828125,
0.11419677734375,
0.65771484375
],
[
0.118408203125,
0.65771484375,
0.8818359375,
0.65771484375,
0.8818359375,
0.6728515625,
0.118408203125,
0.6728515625
],
[
0.118408203125,
0.6748046875,
0.8818359375,
0.67333984375,
0.8818359375,
0.6884765625,
0.118408203125,
0.689453125
],
[
0.118408203125,
0.6904296875,
0.8818359375,
0.689453125,
0.8818359375,
0.70361328125,
0.118408203125,
0.705078125
],
[
0.118408203125,
0.70703125,
0.37451171875,
0.70703125,
0.37451171875,
0.72021484375,
0.118408203125,
0.72021484375
],
[
0.1368408203125,
0.7216796875,
0.880859375,
0.7216796875,
0.880859375,
0.736328125,
0.1368408203125,
0.736328125
],
[
0.12005615234375,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.12005615234375,
0.7529296875
],
[
0.118408203125,
0.7548828125,
0.8818359375,
0.7548828125,
0.8818359375,
0.76953125,
0.118408203125,
0.76953125
],
[
0.118408203125,
0.771484375,
0.88671875,
0.771484375,
0.88671875,
0.78515625,
0.118408203125,
0.78515625
],
[
0.118408203125,
0.7861328125,
0.8818359375,
0.7861328125,
0.8818359375,
0.7998046875,
0.118408203125,
0.7998046875
],
[
0.118408203125,
0.80419921875,
0.88330078125,
0.80419921875,
0.88330078125,
0.81787109375,
0.118408203125,
0.81787109375
],
[
0.12005615234375,
0.8193359375,
0.88330078125,
0.8193359375,
0.88330078125,
0.8330078125,
0.12005615234375,
0.8330078125
],
[
0.1175537109375,
0.8349609375,
0.88330078125,
0.83447265625,
0.88330078125,
0.8486328125,
0.1175537109375,
0.849609375
],
[
0.139404296875,
0.85693359375,
0.88330078125,
0.85791015625,
0.88330078125,
0.87255859375,
0.139404296875,
0.87158203125
],
[
0.12005615234375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8828125,
0.12005615234375,
0.8828125
],
[
0.118408203125,
0.8828125,
0.5087890625,
0.8828125,
0.5087890625,
0.896484375,
0.118408203125,
0.896484375
],
[
0.1427001953125,
0.89794921875,
0.456787109375,
0.89794921875,
0.456787109375,
0.90771484375,
0.1427001953125,
0.90771484375
]
] |
[
"probability that the sentences are parallel by feeding the matching vectors into a feed-forward neural",
"network with a sigmoid output layer:",
"h(1) = hs o hT,",
"(4)",
"h(2) = [h - hT],",
"(5)",
"h; = tanh(W(1)h(1) + W(2)h(2) + b),",
"(6)",
"p(yi = 1|hi) = o(vh; + b),",
"(7)",
"where o() is the sigmoid function, W(1) e IRdy xdh, W(2) e IRdy xd, v e Rdr, b e Rdf and b are",
"model parameters. The value df is the size of the hidden layer in the feed-forward neural network.",
"The model is trained by minimizing the cross entropy of our labeled sentence pairs:",
"n(1+m",
"(yi log a(vhi + b) + (1 - yi) log(1 - o(vhi + b))",
"(8) ",
"For prediction, a sentence pair is classified as parallel if its probability is greater than or equal to a",
"L=-",
"decision threshold p that we need to fix:",
"[1 if p(yi= 1|hi) ≥ p,",
"Ui",
"(9)",
"[0otherwise.",
"The two embedding matrices ES and ET are parameters of the model that we must learn. As the size",
"of the vocabulary and the dimension de increase, the number of parameters can become considerably",
"expensive to estimate. For that reason, it is common to initialize the embedding matrices using word",
"embeddings pre-trained on a large collection of texts.3",
"4 Experiments and Results",
"To assess the effectiveness of our approach, we compare it in multiple settings. In Section 4.1, we",
"measure its capacity to extract parallel sentences found in parallel corpora in which we inserted non-",
"parallel sentences. We extract sentence pairs from real comparable corpora in Section 4.2 and validate",
"their utility by measuring their impact on SMT and NMT systems.",
"4.1 System comparison",
"4.1.1 Data",
"To create our training and test sets we use the WMT'15 English-French datasets.4 The positive examples",
"of our training set consist of 500,000 parallel sentence pairs randomly selected from the Europarl cor-",
"pus (Koehn, 2005). To generate negative examples, at the beginning of each training epoch we sample m",
"non-parallel sentences per positive example, as described in Section 3.1. The vocabulary size is 69,381",
"for English and 84,182 for French.",
"The most reliable way to create test sets to compare different approaches would be to have profes-",
"sional translators manually annotate parallel sentences from comparable corpora. However, this option",
"is expensive and impractical. Therefore, it is common practice to compare parallel sentence extrac-",
"tion systems on artificially-created noisy parallel data by inserting non-parallel sentences into a parallel",
"corpus. Thus, to create our test sets we use parallel sentences from the newstest2012 and Europarl cor-",
"pora and insert artificial noise by substituting target sentences with non-parallel target sentences from",
"their respective held-out dataset. We sample 1,000 parallel sentences from the newstest2012 corpus and",
"create ten test sets with a noise ratio r e [0%, 10%,..:,80%, 90%], where r is the ratio of artificial",
"3we pre-trained bilingual word embeddings as proposed by (Gouws et al., 2015) and (Smith et al., 2017). While we",
"observed faster learning and a small improvement in the performance of ourmodel,we prefer to learn the embedding matrices",
"from scratch instead of relying on additional external resources.",
"*http://www.statmt.org/wmt15/translation-task.html"
] |
{
"arxiv_id": "1806.05559",
"language": "en",
"page_idx": 3
}
|
||
1806.05559-en-4
|
1806.05559-en-4.jpg
|
non-parallel sentences. The same process is repeated with the Europarl corpus to create three test sets with $r\in\{0\%,50\%,90\%\}.$ Each final test set consists of 1,000,0000 sentence pairs generated from the Cartesian product between the sentences in both languages. For example, 400 out of the 1,000 sentence pairs are parallel if $r=60\%$, in which only $0.04\%$ of the 1,000,000 sentence pairs generated from the Cartesian product are truly parallel.
We tokenize all datasets with the scripts from Moses (Koehn et al., 2007).$^{5}$ The maximum sentence
length is set to 80 tokens. Each out-of-vocabulary word is mapped to a special UNK token.
## 4.1.2 Evaluation metrics
For evaluating the performance of our models, a sentence pair predicted as parallel is correct if it is present in the set of parallel sentences of the test set. Precision is the proportion of truly parallel sentence pairs among all extracted sentence pairs. Recall is the proportion of truly parallel extracted sentence pairs among all parallel sentence pairs in the test set. The $\mathcal{F}_1$ score is the harmonic mean of precision and recall.
## 4.1.3 Baseline
For comparison, we use a parallel sentence extraction system developed in-house based on the works of (Munteanu and Marcu, 2005) and (Smith et al., 2010). The system consists of a candidate sentence pair filtering process and three models; two word alignment models and a maximum entropy classifier. The word alignment models are trained on both language directions using our training set of 500,000 parallel sentence pairs. To train the classifier, we select another 100,000 parallel sentence pairs from the held-out Europarl dataset. To generate negative examples, we use 100,000 non-parallel sentence pairs that have successfully passed the candidate sentence pair selection process.$^{6}$
Candidate sentence pair selection A sentence pair filtering process is used to select a fixed number of negative sentence pairs to train the maximum entropy classifier. The objective is to select similar non-parallel sentence pairs to make the classifier more robust to noise. It is also used during prediction to filter out the unlikely sentence pairs of the Cartesian product. Since most of these sentence pairs are not parallel, the filtering process significantly reduces the number of sentence pairs to evaluate. The filtering process consists of a two-step procedure. First, it verifies that the ratio of the lengths between two sentences is not greater than $2.^{7}$ It then uses a word-overlap filter to check for both sentences that at least 50% of their words have a translation in the other sentence. Every pair that does not fulfill these two conditions is discarded. The filtering process is only applied to the baseline model.
Word alignment models The translation and alignment tables are estimated using the HMM alignment model of (Vogel et al., 1996). These probability tables are required to calculate the value of many alignment features used in the classifier.
Maximum entropy classifier The classifier relies on word-level alignment features between two sentences, such as the number of connected words, top three largest fertilities, length of the longest connected substring, log probability of the alignments, and also general features, such as the lengths of the sentences, length difference and the percentage of words on each side that have a translation on the other side. For each sentence pair, a total of 31 features must be calculated and it is classified as parallel if the classifier outputs a probability score greater than or equal to a decision threshold $\rho$ which needs to be fixed.
## 4.1.4 Training settings
The models of our approach are implemented in TensorFlow (Abadi et al., 2016). We use a BiRNN with a single layer in each direction with 512-dimensional word embeddings and 512-dimensional recurrent states. We use LSTM as recurrent activation functions. The hidden layer of the feed-forward neural network has 256 hidden units. To train our models, we use Adam (Kingma and Ba, 2014) with a learning rate of 0.0002 and a minibatch of 128 examples. Models are trained for a total of 15 epochs. To avoid
$^{5}$https://github.com/moses-smt/mosesdecoder
6During our experiments, we did not observe any significant gain by using more than 100, 000 parallel sentence pairs 7
$^{7}$Most parallel sentence extraction systems use a length ratio value equals to 2.
|
[
[
0.12005615234375,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.12005615234375,
0.0926513671875,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.12005615234375,
0.10809326171875,
0.88330078125,
0.10809326171875,
0.88330078125,
0.122314453125,
0.12005615234375,
0.122314453125
],
[
0.12005615234375,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.12005615234375,
0.141357421875,
0.38037109375,
0.141357421875,
0.38037109375,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.1368408203125,
0.15673828125,
0.88330078125,
0.15673828125,
0.88330078125,
0.171630859375,
0.1368408203125,
0.171630859375
],
[
0.118408203125,
0.1728515625,
0.78662109375,
0.1728515625,
0.78662109375,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.12005615234375,
0.1995849609375,
0.32568359375,
0.1995849609375,
0.32568359375,
0.213134765625,
0.12005615234375,
0.213134765625
],
[
0.118408203125,
0.2197265625,
0.8818359375,
0.2197265625,
0.8818359375,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.12005615234375,
0.236328125,
0.88330078125,
0.236328125,
0.88330078125,
0.2493896484375,
0.12005615234375,
0.2493896484375
],
[
0.12005615234375,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.2666015625,
0.12005615234375,
0.2666015625
],
[
0.118408203125,
0.268310546875,
0.88330078125,
0.268310546875,
0.88330078125,
0.281982421875,
0.118408203125,
0.281982421875
],
[
0.118408203125,
0.28515625,
0.16796875,
0.28515625,
0.16796875,
0.29638671875,
0.118408203125,
0.29638671875
],
[
0.118408203125,
0.308837890625,
0.2418212890625,
0.308837890625,
0.2418212890625,
0.32373046875,
0.118408203125,
0.32373046875
],
[
0.118408203125,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.344482421875,
0.118408203125,
0.344482421875
],
[
0.118408203125,
0.3466796875,
0.88330078125,
0.3466796875,
0.88330078125,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.12005615234375,
0.3623046875,
0.8818359375,
0.3623046875,
0.8818359375,
0.377197265625,
0.12005615234375,
0.377197265625
],
[
0.118408203125,
0.377685546875,
0.88330078125,
0.377197265625,
0.88330078125,
0.391357421875,
0.118408203125,
0.392578125
],
[
0.12005615234375,
0.3955078125,
0.88330078125,
0.3955078125,
0.88330078125,
0.4091796875,
0.12005615234375,
0.4091796875
],
[
0.118408203125,
0.410888671875,
0.88330078125,
0.410888671875,
0.88330078125,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.12005615234375,
0.427001953125,
0.67822265625,
0.427001953125,
0.67822265625,
0.441162109375,
0.12005615234375,
0.441162109375
],
[
0.12005615234375,
0.443603515625,
0.88330078125,
0.443603515625,
0.88330078125,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.12005615234375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.473876953125,
0.12005615234375,
0.473876953125
],
[
0.118408203125,
0.474365234375,
0.88330078125,
0.4755859375,
0.88330078125,
0.490478515625,
0.118408203125,
0.4892578125
],
[
0.1158447265625,
0.490478515625,
0.88330078125,
0.4912109375,
0.88330078125,
0.505859375,
0.1158447265625,
0.5048828125
],
[
0.12005615234375,
0.50830078125,
0.88330078125,
0.50830078125,
0.88330078125,
0.5224609375,
0.12005615234375,
0.5224609375
],
[
0.12005615234375,
0.52490234375,
0.88330078125,
0.52490234375,
0.88330078125,
0.5380859375,
0.12005615234375,
0.5380859375
],
[
0.118408203125,
0.54052734375,
0.884765625,
0.54052734375,
0.884765625,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.12005615234375,
0.5556640625,
0.88330078125,
0.5556640625,
0.88330078125,
0.57080078125,
0.12005615234375,
0.57080078125
],
[
0.118408203125,
0.572265625,
0.759765625,
0.572265625,
0.759765625,
0.5859375,
0.118408203125,
0.5859375
],
[
0.12005615234375,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.1175537109375,
0.6025390625,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.1175537109375,
0.61767578125
],
[
0.12005615234375,
0.619140625,
0.418212890625,
0.619140625,
0.418212890625,
0.6328125,
0.12005615234375,
0.6328125
],
[
0.12005615234375,
0.63720703125,
0.88330078125,
0.63720703125,
0.88330078125,
0.65185546875,
0.12005615234375,
0.65185546875
],
[
0.12005615234375,
0.65283203125,
0.8818359375,
0.65283203125,
0.8818359375,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.118408203125,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.118408203125,
0.68310546875
],
[
0.12005615234375,
0.68505859375,
0.8818359375,
0.68505859375,
0.8818359375,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.12176513671875,
0.70361328125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71435546875,
0.12176513671875,
0.71435546875
],
[
0.12005615234375,
0.71875,
0.88330078125,
0.71875,
0.88330078125,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.1175537109375,
0.73388671875,
0.16455078125,
0.73388671875,
0.16455078125,
0.7490234375,
0.1175537109375,
0.7490234375
],
[
0.1175537109375,
0.75830078125,
0.30908203125,
0.76025390625,
0.30908203125,
0.77490234375,
0.1175537109375,
0.7724609375
],
[
0.118408203125,
0.779296875,
0.8818359375,
0.779296875,
0.8818359375,
0.79345703125,
0.118408203125,
0.79345703125
],
[
0.118408203125,
0.79638671875,
0.884765625,
0.79638671875,
0.884765625,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.118408203125,
0.8125,
0.88330078125,
0.8125,
0.88330078125,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.118408203125,
0.8291015625,
0.8818359375,
0.8291015625,
0.8818359375,
0.8427734375,
0.118408203125,
0.8427734375
],
[
0.12005615234375,
0.84423828125,
0.88330078125,
0.84423828125,
0.88330078125,
0.859375,
0.12005615234375,
0.859375
],
[
0.14111328125,
0.87060546875,
0.41650390625,
0.87060546875,
0.41650390625,
0.88427734375,
0.14111328125,
0.88427734375
],
[
0.14111328125,
0.8828125,
0.8486328125,
0.8828125,
0.8486328125,
0.896484375,
0.14111328125,
0.896484375
],
[
0.14111328125,
0.89453125,
0.6171875,
0.8955078125,
0.6171875,
0.9091796875,
0.14111328125,
0.90771484375
]
] |
[
"non-parallel sentences. The same process is repeated with the Europarl corpus to create three test sets",
"with r E [0%, 50%, 90%]. Each final test set consists of 1,000,000 sentence pairs generated from the",
"Cartesian product between the sentences in both languages. For example, 400 out of the 1,000 sentence",
"pairs are parallel if r = 60%, in which only 0.04% of the 1,000,000 sentence pairs generated from the",
"Cartesian product are truly parallel.",
"We tokenize all datasets with the scripts from Moses (Koehn et al., 2007).5 The maximum sentence",
"length is set to 80 tokens. Each out-of-vocabulary word is mapped to a special UNK token.",
"4.1.2Evaluation metrics",
"For evaluating the performance of our models, a sentence pair predicted as parallel is correct if it is",
"present in the set of parallel sentences of the test set. Precision is the proportion of truly parallel sentence",
"pairs among all extracted sentence pairs. Recall is the proportion of truly parallel extracted sentence",
"pairs among all parallel sentence pairs in the test set. The Fi score is the harmonic mean of precision and",
"recall.",
"4.1.3Baseline",
"For comparison, we use a parallel sentence extraction system developed in-house based on the works",
"of (Munteanu and Marcu, 2005) and (Smith et al., 2010). The system consists of a candidate sentence",
"pair filtering process and three models; two word alignment models and a maximum entropy classifier.",
"The word alignment models are trained on both language directions using our training set of 500,000",
"parallel sentence pairs. To train the classifier, we select another 100,000 parallel sentence pairs from the",
"held-out Europarl dataset. To generate negative examples, we use 100,000 non-parallel sentence pairs",
"that have successfully passed the candidate sentence pair selection process.6",
"Candidate sentence pair selection A sentence pair filtering process is used to select a fixed number",
"of negative sentence pairs to train the maximum entropy classifier. The objective is to select similar",
"non-parallel sentence pairs to make the classifier more robust to noise. It is also used during prediction",
"to filter out the unlikely sentence pairs of the Cartesian product. Since most of these sentence pairs are",
"not parallel, the filtering process significantly reduces the number of sentence pairs to evaluate. The",
"filtering process consists of a two-step procedure. First, it verifies that the ratio of the lengths between",
"two sentences is not greater than 2.7 It then uses a word-overlap filter to check for both sentences that",
"at least 50% of their words have a translation in the other sentence. Every pair that does not fulfill these",
"two conditions is discarded. The filtering process is only applied to the baseline model.",
"Word alignment models The translation and alignment tables are estimated using the HMM alignment",
"model of (Vogel et al., 1996). These probability tables are required to calculate the value of many",
"alignment features used in the classifier.",
"Maximum entropy classifier The classifier relies on word-level alignment features between two sen-",
"tences, such as the number of connected words, top three largest fertilities, length of the longest con-",
"nected substring, log probability of the alignments, and also general features, such as the lengths of the",
"sentences, length difference and the percentage of words on each side that have a translation on the other",
"side.For eachsentencepair,atotal of3lfeaturesmustbecalculated anditis classified asparallel if the",
"classifier outputs a probability score greater than or equal to a decision threshold p which needs to be",
"fixed.",
"4.1.4 Training settings",
"The models of our approach are implemented in TensorFlow (Abadi et al., 2016). We use a BiRNN with",
"a single layer in each direction with 512-dimensional word embeddings and 512-dimensional recurrent",
"states. We use LSTM as recurrent activation functions. The hidden layer of the feed-forward neural",
"network has 256 hidden units. To train our models, we use Adam (Kingma and Ba, 2014) with a learning",
"rate of 0.0002 and a minibatch of 128 examples. Models are trained for a total of 15 epochs. To avoid",
"https:/lgithub.com/moses-smt/mosesdecoder",
"6During our experiments, we did not observe any significant gain by using more than 100,000 parallel sentence pairs.",
"7Most parallel sentence extraction systems use a length ratio value equals to 2."
] |
{
"arxiv_id": "1806.05559",
"language": "en",
"page_idx": 4
}
|
||
2004.09329-en-0
|
2004.09329-en-0.jpg
|
$\textbf{Robust Partial Matching for Person Search in the Wild}$
Yingji Zhong$^1$, Xiaoyu Wang$^{2,3}$, Shiliang Zhang$^{1}$
$^{1}$Peking University, $^2\text{Intellifusion, }^3$The Chinese University of Hong Kong, Shenzhen
$^{1}\{$zhongyj, slzhang.jdl$\}$@pku.edu.cn,$^2,3$fanghuaxue@gmail.com
$సి$
$ଟ୍ରି$
## Abstract
Various factors like occlusions, backgrounds, etc., would
$\begin{array}{ll}{}lead\:to\:misaligned\:detected\:bounding\:boxes\:,\:e.g,\:ones\:cov-\\{}&ering\:only\:portions\:of\:human\:body.\:This\:issue\:is\:common\\{}&but\:overlooked\:bv\:previous\:person\:search\:works.\:To\:allevi-\end{array} \\$
but overlooked by previous person search works. To alleviate this issue, this paper proposes an Align-to-Part Network $( APNet) \textit{for person detection and re- Identification ( reID) . }$ APNet refines detected bounding boxes to cover the esti-
$\sum_{\begin{array}{ccc}\text{mated holistic body regions,from which discriminative part}\\\text{已}&\text{features can be extracted and aligned. Aligned part features}\\\text{naturally formulate reID as a partial feature matching pro-}\end{array}}$ $\begin{array}{ccc}&\text{nanarany jormaue reib as a parnat jeanare macting pro- cedure, where valid part features are selected for similarity}\\&\text{computation, while part features on occluded or noisy re- computation, while part features on occluded or noisy re- compos on}\end{array}$
gions are discarded. This design enhances the robustness of person search to real-world challenges with marginal computation overhead. This paper also contributes a Large- $Scale\textit{dataset for Person Search in the wild ( LSPS) , which is}$ by far the largest and the most challenging dataset for person search. Experiments show that APNet brings considerable performance improvement on LSPS. Meanwhile, it achieves competitive performance on existing person search $benchmarks\textit{like CUHK- SYSU and PRW}.$
<FigureHere>
Figure 1. Illustration of misaligned person bounding boxes and issues of person matching with global feature. Misalignments commonly occur in person search and can not be eliminated by training stronger detectors. This paper addresses this issue with bounding box alignment and partial matching.
erally helpful for reID, they may still produce misaligned person bounding boxes. As shown in Fig. 1, occlusions and limited camera viewing field lead to many accurate but misaligned bounding boxes covering portions of pedestrians. Most of existing person search methods extracts global feature from the detected bounding boxes, no matter the joint models[24, 15] or the separate models[1,28,3]. As shown in Fig. 1, misalignment degrades the performance of global features, because it is not reasonable to match partial features against global features. More detailed reviews to person search works can be found in Sec. 2.
This paper targets to design a unified person search framework robust to the misaligned bounding boxes. As shown in Fig. 1, the basic idea is to refine detected bounding boxes to cover the estimated holistic body regions to eliminate misalignment errors. Then, aligned part features could be extracted from refined bounding boxes with simple region division. Part features hence allow for robust partial matching across bounding boxes, where features on the mutual visible regions are matched for similarity computation, and features on noisy regions are not considered for matching. In other words, our strategy improves both the feature resolution and robustness to misalignment errors.
The above idea leads to the Align-to-Part Network (APNet) for person search. Built on OIM[24], APNet consists of the additional Bounding Box Aligner (BBA) and Region
### $1. \textbf{Introduction}$
Thanks to recent research efforts [27,17,10,8,25,22, 30, 34, 31], the performance of person re-Identification (reID) has been significantly improved. However, one potential issue with current reID setting is that, it treats person detection and reID as two isolated steps. This issue can be compensated by person search [24,33,1,28,3,15] to jointly accomplish person detection and reID. Being able to jointly optimize person detection and reID, person search is attracting more and more attention, and is potential to present advantages in flexibility, efficiency, and accuracy.
Compared with reID, person search needs to design pedestrian detectors robust to variances of scales, illuminations, backgrounds, occlusions, etc. Some person search works refine detectors by jointly training detectors and reID models [24,15,4,3]. Although strong detectors are gen-
|
[
[
0.2132568359375,
0.1337890625,
0.7587890625,
0.1337890625,
0.7587890625,
0.1483154296875,
0.2132568359375,
0.1483154296875
],
[
0.291748046875,
0.18115234375,
0.6748046875,
0.18115234375,
0.6748046875,
0.1988525390625,
0.291748046875,
0.1988525390625
],
[
0.1502685546875,
0.1962890625,
0.8203125,
0.1988525390625,
0.8203125,
0.216552734375,
0.1502685546875,
0.2139892578125
],
[
0.2344970703125,
0.2197265625,
0.5439453125,
0.2197265625,
0.5439453125,
0.2342529296875,
0.2344970703125,
0.2342529296875
],
[
0.54345703125,
0.2197265625,
0.740234375,
0.2197265625,
0.740234375,
0.23291015625,
0.54345703125,
0.23291015625
],
[
0.0196075439453125,
0.260009765625,
0.05963134765625,
0.2587890625,
0.06451416015625,
0.37255859375,
0.02532958984375,
0.373779296875
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.58837890625,
0.26513671875,
0.62841796875,
0.26513671875,
0.62841796875,
0.27978515625,
0.58837890625,
0.27978515625
],
[
0.7548828125,
0.274658203125,
0.83251953125,
0.276611328125,
0.83251953125,
0.28857421875,
0.7548828125,
0.285888671875
],
[
0.85546875,
0.290283203125,
0.88134765625,
0.292236328125,
0.8798828125,
0.2998046875,
0.8544921875,
0.29736328125
],
[
0.101318359375,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.314453125,
0.101318359375,
0.314453125
],
[
0.080078125,
0.31494140625,
0.468017578125,
0.317626953125,
0.468017578125,
0.33203125,
0.080078125,
0.330078125
],
[
0.51123046875,
0.31640625,
0.619140625,
0.313232421875,
0.619140625,
0.330810546875,
0.51123046875,
0.333984375
],
[
0.08251953125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.3447265625,
0.08251953125,
0.3447265625
],
[
0.07843017578125,
0.34521484375,
0.468017578125,
0.3466796875,
0.468017578125,
0.361083984375,
0.07843017578125,
0.35986328125
],
[
0.035125732421875,
0.3515625,
0.0400390625,
0.3515625,
0.0400390625,
0.359130859375,
0.035125732421875,
0.359130859375
],
[
0.0228729248046875,
0.363037109375,
0.052276611328125,
0.361083984375,
0.056365966796875,
0.39453125,
0.0269622802734375,
0.396484375
],
[
0.08087158203125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.53369140625,
0.367431640625,
0.66259765625,
0.367431640625,
0.66259765625,
0.380615234375,
0.53369140625,
0.380615234375
],
[
0.685546875,
0.36865234375,
0.7138671875,
0.36865234375,
0.7138671875,
0.380126953125,
0.685546875,
0.380126953125
],
[
0.08087158203125,
0.37744140625,
0.46728515625,
0.37744140625,
0.46728515625,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.55224609375,
0.3837890625,
0.65283203125,
0.3837890625,
0.65283203125,
0.395751953125,
0.55224609375,
0.395751953125
],
[
0.08251953125,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.406005859375,
0.08251953125,
0.406005859375
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.08087158203125,
0.40771484375,
0.4697265625,
0.40771484375,
0.4697265625,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.50244140625,
0.412353515625,
0.890625,
0.412353515625,
0.890625,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.03350830078125,
0.447509765625,
0.0555419921875,
0.447509765625,
0.0555419921875,
0.479248046875,
0.03350830078125,
0.479248046875
],
[
0.080078125,
0.45263671875,
0.468017578125,
0.45263671875,
0.468017578125,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.5048828125,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.466552734375,
0.5048828125,
0.466552734375
],
[
0.08087158203125,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50244140625,
0.468505859375,
0.71728515625,
0.468505859375,
0.71728515625,
0.481689453125,
0.50244140625,
0.481689453125
],
[
0.08251953125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.49560546875,
0.08251953125,
0.49560546875
],
[
0.02532958984375,
0.50048828125,
0.058013916015625,
0.50048828125,
0.058013916015625,
0.5888671875,
0.02532958984375,
0.5888671875
],
[
0.080078125,
0.498779296875,
0.468017578125,
0.498779296875,
0.468017578125,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50244140625,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.51318359375,
0.468017578125,
0.51318359375,
0.468017578125,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.5048828125,
0.521484375,
0.890625,
0.521484375,
0.890625,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.080078125,
0.52880859375,
0.4697265625,
0.52880859375,
0.4697265625,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.50244140625,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50244140625,
0.548828125
],
[
0.08087158203125,
0.54541015625,
0.465576171875,
0.54541015625,
0.465576171875,
0.55615234375,
0.08087158203125,
0.55615234375
],
[
0.5048828125,
0.5517578125,
0.8896484375,
0.5517578125,
0.8896484375,
0.5625,
0.5048828125,
0.5625
],
[
0.080078125,
0.5595703125,
0.46728515625,
0.5595703125,
0.46728515625,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.5869140625,
0.080078125,
0.5869140625
],
[
0.5009765625,
0.5791015625,
0.89208984375,
0.580078125,
0.89208984375,
0.5947265625,
0.5009765625,
0.59326171875
],
[
0.0228729248046875,
0.58544921875,
0.056365966796875,
0.58544921875,
0.05963134765625,
0.71044921875,
0.02532958984375,
0.71044921875
],
[
0.035125732421875,
0.58544921875,
0.049835205078125,
0.58544921875,
0.049835205078125,
0.59228515625,
0.035125732421875,
0.59228515625
],
[
0.080078125,
0.5888671875,
0.471435546875,
0.5888671875,
0.471435546875,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50341796875,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.50341796875,
0.609375
],
[
0.035125732421875,
0.60595703125,
0.0400390625,
0.60595703125,
0.0400390625,
0.61572265625,
0.035125732421875,
0.61572265625
],
[
0.080078125,
0.60400390625,
0.348876953125,
0.60400390625,
0.348876953125,
0.6171875,
0.080078125,
0.6171875
],
[
0.50341796875,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.50244140625,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.6396484375,
0.50244140625,
0.6396484375
],
[
0.5009765625,
0.6416015625,
0.890625,
0.64013671875,
0.890625,
0.65380859375,
0.5009765625,
0.65478515625
],
[
0.08087158203125,
0.6484375,
0.2091064453125,
0.6484375,
0.2091064453125,
0.66357421875,
0.08087158203125,
0.66357421875
],
[
0.50341796875,
0.65771484375,
0.7685546875,
0.65771484375,
0.7685546875,
0.66845703125,
0.50341796875,
0.66845703125
],
[
0.1004638671875,
0.6748046875,
0.46728515625,
0.6748046875,
0.46728515625,
0.68798828125,
0.1004638671875,
0.68798828125
],
[
0.521484375,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.6875,
0.521484375,
0.6875
],
[
0.5009765625,
0.6875,
0.8935546875,
0.68798828125,
0.8935546875,
0.70263671875,
0.5009765625,
0.7021484375
],
[
0.035125732421875,
0.6767578125,
0.0400390625,
0.6767578125,
0.0400390625,
0.6845703125,
0.035125732421875,
0.6845703125
],
[
0.08087158203125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.08087158203125,
0.70263671875
],
[
0.08087158203125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.08087158203125,
0.71826171875
],
[
0.50341796875,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.50244140625,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.080078125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.5009765625,
0.76318359375,
0.89208984375,
0.76220703125,
0.89208984375,
0.7763671875,
0.5009765625,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.08087158203125,
0.79736328125,
0.4697265625,
0.79736328125,
0.4697265625,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.50244140625,
0.79345703125,
0.890625,
0.79345703125,
0.890625,
0.806640625,
0.50244140625,
0.806640625
],
[
0.080078125,
0.81103515625,
0.455078125,
0.81103515625,
0.455078125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.8095703125,
0.89208984375,
0.8095703125,
0.89208984375,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.09722900390625,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.09722900390625,
0.84033203125
],
[
0.50244140625,
0.82470703125,
0.89208984375,
0.82470703125,
0.89208984375,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5,
0.83837890625,
0.826171875,
0.8388671875,
0.826171875,
0.85400390625,
0.5,
0.85302734375
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.521484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.521484375,
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,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.5009765625,
0.8994140625
]
] |
[
"Robust Partial Matching for Person Search in the Wild",
"Yingji Zhong', Xiaoyu Wang2.3, Shiliang Zhang",
"IPeking University, 2Intellifusion, 3The Chinese University of Hong Kong, Shenzhen",
"I[zhongyj, slzhang.jdl)@pku.edu.cn,",
"23fanghuaxue@gmail.com",
" d (",
"Abstract",
"Gallery",
"Global Matching",
"Ouer",
"Various factors like occlusions, backgrounds, etc., would",
"lead to misaligned detected bounding boxes,e.g., ones cov-",
"EO",
"ering onlyportions ofhumanbodyThis issue is common",
"but overlooked by previous person search works. To allevi-",
"",
"2",
"ate this issue, this paper proposes an Align-to-Part Network",
"CompleteBodyRegion",
"Align",
"(APNet) for person detection and re-Identification (relD).",
"Partial Body Region",
"APNet refines detected bounding boxes to cover the esti-",
"LAD",
"mated holistic body regions, from which discriminative part",
"Figure 1. Ilustration of misaligned person bounding boxes and is-",
"featurescanbeextractedandaligned.Alignedpartfeatures",
"sues of person matching with global feature. Misalignments com-",
"naturally formulate relD as a partial feature matching pro-",
"monly occur in person search and can not be eliminated by training",
"'S",
"cedure, where valid part features are selected for similarity",
"stronger detectors.This paper addresses this issuewithbounding",
"computation, while part features on occluded or noisy re-",
"box alignment and partial matching.",
"gions are discarded.This design enhances the robustness of",
"10",
"person search to real-world challenges with marginal com-",
"erally helpful for relD, they may still produce misaligned",
"putation overhead. This paper also contributes a Large-",
"personboundingboxes.As showninFig.l,occlusionsand",
"Scale dataset for Person Search in the wild (LSPS), which is",
"limited camera viewing field lead to many accurate but mis-",
"byfarthe largest and the most challenging datasetforper",
"alignedboundingboxes coveringportions ofpedestrians.",
"son search. Experiments show that APNet brings consid-",
"Most of existing person search methods extracts global fea-",
"erable performance improvement on LSPS. Meanwhile, it",
"ture from the detected bounding boxes, no matter the joint",
"",
"",
"achieves competitive performance on existing person search",
"models [24, 15] or the separate models [1, 28, 3]. As shown",
"",
"benchmarks like CUHK-SYSU and PRW.",
"inFig.l.misalignmentdegradestheperformanceof global",
"features, because it is not reasonable to match partial fea-",
"tures against global features.More detailed reviews to per-",
"1. Introduction",
"son search workscan befoundin Sec.2.",
"Thanks to recent research efforts [27, 17, 10, 8, 25, 22,",
"This paper targets to design a unified person search",
"framework robust to the misaligned bounding boxes.As",
"",
"30, 34, 31], the performance of person re-Identification",
"(relD) has been significantly improved. However, one po-",
"shown in Fig. I, the basic idea is to refine detected bound-",
"tential issue with current relD setting is that, it treats person",
"ing boxes to cover the estimated holistic body regions to",
"detection and relD as two isolated steps. This issue can",
"eliminate misalignment errors. Then, aligned part features",
"be compensated by person search [24, 33, 1, 28, 3, 15] to",
"could be extracted from refined bounding boxes with simple",
"jointly accomplish person detection and reID. Being able to",
"region division.Part features hence allow for robust partial",
"jointly optimize person detection and relD, person search",
"matching across bounding boxes, where features on the mu-",
"is attracting more and more attention, and is potential to",
"tual visible regions are matched for similarity computation,",
"present advantages in flexibility, efficiency, and accuracy.",
"and features on noisy regions are not considered for match-",
"Compared with relD, person search needs to design",
"ing. In other words, our strategy improves both the feature",
"pedestrian detectors robust to variances of scales, illumi-",
"resolution and robustness to misalignment errors.",
"nations, backgrounds, occlusions, etc. Some person search",
"The above idea leads to the Align-to-Part Network (AP-",
"works refine detectors by jointly training detectors and reID",
"Net) for person search. Built on OIM [24], APNet consists",
"models [24, 15, 4, 3]. Although strong detectors are gen-",
"oftheadditionalBoundingBoxAligner(BBA)andRegion"
] |
{
"arxiv_id": "2004.09329",
"language": "en",
"page_idx": 0
}
|
||
2004.09329-en-1
|
2004.09329-en-1.jpg
|
Sensitive Feature Extractor (RSFE).The BBA module is implemented by predicting 4 bounding box offset values. It can be trained by automatic data augmentation without requiring manual annotations. Because of enlarged receptive fields of neurons in feature extraction CNN layer, noisy or occluded parts would affect features of their adjacent parts in the same bounding box. We further design the RSFE module to extract part features. RSFE reinforces local cues in each part feature, hence effectively alleviates the negative effects of adjacent noises in part feature learning.
We test our approach on current person search benchmarks, $i.e.$, PRW [33] and CUHK-SYSU [24]. Experiment results show that, our APNet achieves competitive performance compared with recent works. To test our approach in more challenging scenarios, we contribute a new person search dataset named as Large-Scale dataset for Person Search in the wild (LSPS). Compared with PRW and CUHK-SYSU, LSPS presents several new features that would encourage the research towards more realistic person search: 1) large scale: it contains a large number of pedestrians and bounding boxes; 2) automatically detected bounding boxes, which differ with the manually annotated ones in PRW and CUHK-SYSU; 3) collected in scenarios with occlusions and crowdings. On this challenging LSPS, our method exhibits substantial advantages.
Compared with reID, person search is still relatively under-explored. It is critical to deal with the challenge occuring in both person detection and reID. To the best of our knowledge, the proposed APNet is an early attempt in boosting the robustness to misalignment issue in person search. As shown in our experiments, the APNet brings considerable performance gains. The proposed LSPS dataset simulates realistic scenarios and presents a more challenging person search task than existing datasets. We believe that, the proposed methods and the contributed LSPS can inspire more research efforts in person search.
## $2. \textbf{Related Work}$
This paper is closely related to person reID and person search. The following parts review those two lines of works.
Person reID extracts robust and discriminative features from given person bounding boxes. Current fully supervised reID methods can be briefly summarized into the following categories. 1) Learn local features of regions utilizing offline detector to split body into several regions [27,17,10,11],or uniformly split the body into several stripes [31, 22, 26]; 2) Learn the robust feature utilizing additional attribute annotation [19, 18, 20, 23]; 3) Enhance the discriminativeness of feature by attention mechanism [12,9,25]; 4) Impose constraints on feature space with loss functions, like verification loss [36] and triplet loss [8]. Among the above methods, PCB [22] and its variants [31, 26] dominate the state-of-the-art performance.
PCB uniformly splits the feature map into stripes and supervises them to ensure their discriminative power. However, PCB requires a strict alignment of the input images since misalignment breaks the correspondence of stripes on the same spatial position. The performance of PCB thus degrades substantially when misalignments exist[21].
There are some reID works focusing on partial reID. Zheng et al.[35] propose AMC+SWM to solve the partial reID. AMC collects patches from all gallery images and set up a dictionary for patch level matching. He et al.[7] propose DSR to reconstruct the partial image by holistic image and utilize the reconstruction error as the similarity. Both AMC[35] and DSR[7] are optimisation-based thus it is expensive to get the similarity of each query-gallery pair. PGFA[14] applies partial matching scheme with the aligned input images, whose occluded parts are indicated by the expensive offline keypoint detector. Sun et al.[21] propose a VPM to extract local features while simutaneously being aware of the visibility of each body part. However, VPM does not consider the deformation mentioned in DSR. Our method address the deformation by an explicit alignment step and gets rid of the time-consuming comparsion of each query-gallery pair during the inference stage.
Person search considers the detection stage in raw video frames before reID. Current methods of person search can be divided into two categories. One integrates detection and reID into a unifed framework [24,15,4,29]. Based on Faster-RCNN[16], OIM proposed by Xiao et al.[24] inserts an additional feature extraction branch on top of the detector head, which is the first end-to-end learning framework for person search. Similar to NPSM of Liu et al.[4], Munjal $etal.[15]$ apply a query-guided method for person search while it is end-to-end optimized compared to NPSM. The other category solves detection and re-identification with two separate models, $i.e.$, performing person search in a sequential manner [1,28,3]. Chen et $\hat{al}.[1]$ propose a mask-guided feature learning method to make the reID network focus more on foregrounds. Lan et al.[28] propose CLSA which utilizes multi-level features from reID network to solve the multi-scale matching problem. Han et al.[3] claim that bounding box from detector might not be optimal for re-identification, they thus refine the detector results driven by re-identification loss.
Our work belongs to the first category of person search, integrating detection and reID into a unified framework. Different from existing person search works, this work further addresses the misalignment issue of detected bounding boxes and optimizes aligned part feature learning to achieve robust partial matching in person search.
## $3. \textbf{Problem Formulation}$
Given a query person image $q$ and a gallery set $\mathcal{G}=$ $\{g_i\}_{i=1:N}$ containing $N$ frames, person search aims at
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
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,
0.1383056640625,
0.471435546875,
0.1383056640625,
0.471435546875,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.8388671875,
0.1673583984375,
0.8388671875,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.1817626953125,
0.46728515625,
0.1817626953125,
0.46728515625,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.521484375,
0.184326171875,
0.88818359375,
0.184326171875,
0.88818359375,
0.1976318359375,
0.521484375,
0.1976318359375
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.50244140625,
0.1988525390625,
0.8896484375,
0.1988525390625,
0.8896484375,
0.21337890625,
0.50244140625,
0.21337890625
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2139892578125,
0.89208984375,
0.2139892578125,
0.89208984375,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.080078125,
0.2291259765625,
0.40185546875,
0.2291259765625,
0.40185546875,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.50244140625,
0.24365234375
],
[
0.09722900390625,
0.242431640625,
0.468017578125,
0.2412109375,
0.468017578125,
0.255615234375,
0.09722900390625,
0.2568359375
],
[
0.5009765625,
0.2442626953125,
0.89208984375,
0.24365234375,
0.89208984375,
0.25830078125,
0.5009765625,
0.2587890625
],
[
0.0767822265625,
0.25830078125,
0.468017578125,
0.2568359375,
0.468017578125,
0.271484375,
0.0767822265625,
0.272705078125
],
[
0.50244140625,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.2734375,
0.50244140625,
0.2734375
],
[
0.080078125,
0.2734375,
0.471435546875,
0.2734375,
0.471435546875,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.290283203125,
0.890625,
0.290283203125,
0.890625,
0.3037109375,
0.50244140625,
0.3037109375
],
[
0.07843017578125,
0.302978515625,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.07843017578125,
0.3193359375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.07843017578125,
0.333251953125
],
[
0.5048828125,
0.3212890625,
0.8896484375,
0.3212890625,
0.8896484375,
0.33203125,
0.5048828125,
0.33203125
],
[
0.080078125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50244140625,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.5009765625,
0.365478515625,
0.89208984375,
0.365478515625,
0.89208984375,
0.37890625,
0.5009765625,
0.37890625
],
[
0.07843017578125,
0.37890625,
0.4697265625,
0.37744140625,
0.4697265625,
0.392578125,
0.07843017578125,
0.393310546875
],
[
0.50341796875,
0.380126953125,
0.890625,
0.380126953125,
0.890625,
0.393310546875,
0.50341796875,
0.393310546875
],
[
0.07843017578125,
0.39453125,
0.4697265625,
0.393310546875,
0.4697265625,
0.4072265625,
0.07843017578125,
0.40771484375
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.410888671875,
0.89208984375,
0.410888671875,
0.89208984375,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.07843017578125,
0.423583984375,
0.471435546875,
0.42236328125,
0.471435546875,
0.4375,
0.07843017578125,
0.438232421875
],
[
0.50244140625,
0.425537109375,
0.8505859375,
0.425537109375,
0.8505859375,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.521484375,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455810546875,
0.521484375,
0.455810546875
],
[
0.07843017578125,
0.455078125,
0.366943359375,
0.455078125,
0.366943359375,
0.468505859375,
0.07843017578125,
0.468505859375
],
[
0.50244140625,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.098876953125,
0.4697265625,
0.468017578125,
0.4697265625,
0.468017578125,
0.48291015625,
0.098876953125,
0.48291015625
],
[
0.5009765625,
0.47021484375,
0.89208984375,
0.4716796875,
0.89208984375,
0.486083984375,
0.5009765625,
0.48486328125
],
[
0.080078125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50341796875,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.501953125,
0.8896484375,
0.501953125,
0.8896484375,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.07843017578125,
0.51513671875,
0.468017578125,
0.5146484375,
0.468017578125,
0.52880859375,
0.07843017578125,
0.52978515625
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.5341796875,
0.8896484375,
0.5341796875,
0.8896484375,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.080078125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.080078125,
0.57470703125,
0.4697265625,
0.57470703125,
0.4697265625,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57763671875,
0.890625,
0.57763671875,
0.890625,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.07843017578125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.07843017578125,
0.60498046875
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.07843017578125,
0.60400390625,
0.4697265625,
0.60498046875,
0.4697265625,
0.619140625,
0.07843017578125,
0.61865234375
],
[
0.50244140625,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.07843017578125,
0.619140625,
0.450927734375,
0.62060546875,
0.450927734375,
0.63525390625,
0.07843017578125,
0.6337890625
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.50244140625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.080078125,
0.6494140625,
0.2181396484375,
0.6494140625,
0.2181396484375,
0.6640625,
0.080078125,
0.6640625
],
[
0.50341796875,
0.65380859375,
0.890625,
0.65380859375,
0.890625,
0.66748046875,
0.50341796875,
0.66748046875
],
[
0.50341796875,
0.66845703125,
0.890625,
0.66845703125,
0.890625,
0.681640625,
0.50341796875,
0.681640625
],
[
0.098876953125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.098876953125,
0.68798828125
],
[
0.5009765625,
0.681640625,
0.89208984375,
0.68310546875,
0.89208984375,
0.69775390625,
0.5009765625,
0.6962890625
],
[
0.07843017578125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.07843017578125,
0.7041015625
],
[
0.50341796875,
0.69873046875,
0.89208984375,
0.69873046875,
0.89208984375,
0.7119140625,
0.50341796875,
0.7119140625
],
[
0.09722900390625,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.09722900390625,
0.7197265625
],
[
0.080078125,
0.7197265625,
0.46728515625,
0.7197265625,
0.46728515625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.71533203125,
0.890625,
0.71533203125,
0.890625,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.50244140625,
0.72900390625,
0.7587890625,
0.72900390625,
0.7587890625,
0.7421875,
0.50244140625,
0.7421875
],
[
0.07843017578125,
0.736328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.74853515625,
0.07843017578125,
0.75
],
[
0.0767822265625,
0.75,
0.46728515625,
0.75048828125,
0.46728515625,
0.76513671875,
0.0767822265625,
0.7646484375
],
[
0.521484375,
0.744140625,
0.890625,
0.744140625,
0.890625,
0.7587890625,
0.521484375,
0.7587890625
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.50244140625,
0.76025390625,
0.890625,
0.76025390625,
0.890625,
0.7734375,
0.50244140625,
0.7734375
],
[
0.5009765625,
0.7744140625,
0.89208984375,
0.7734375,
0.89208984375,
0.78857421875,
0.5009765625,
0.7890625
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.79052734375,
0.89208984375,
0.79052734375,
0.89208984375,
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
],
[
0.50244140625,
0.80615234375,
0.8935546875,
0.80615234375,
0.8935546875,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.5009765625,
0.8193359375,
0.76953125,
0.82080078125,
0.76953125,
0.83544921875,
0.5009765625,
0.833984375
],
[
0.07843017578125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.0767822265625,
0.84033203125,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.0767822265625,
0.85400390625
],
[
0.50244140625,
0.8447265625,
0.705078125,
0.8447265625,
0.705078125,
0.859375,
0.50244140625,
0.859375
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.521484375,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.8935546875,
0.88623046875,
0.8935546875,
0.90087890625,
0.50341796875,
0.9013671875
]
] |
[
"Sensitive Feature Extractor (RSFE). The BBA module is",
"PCB uniformly splits thefeature map into stripes and super-",
"implemented by predicting 4 bounding box offset values. It",
"vises them to ensure their discriminative power. However,",
"can be trained by automatic data augmentation without re-",
"PCB requires a strict alignment of the input images since",
"quiring manual annotations. Because of enlarged receptive",
"misalignment breaks the correspondence of stripes on the",
"fields of neurons in feature extraction CNN layer, noisy or",
"same spatial position.The performance of PCB thus de-",
"occluded parts would affect features of their adjacent parts",
"grades substantially when misalignments exist [21]",
"in the same bounding box.We further design the RSFE",
"There are some relD works focusing on partial reID",
"module to extract part features.RSFE reinforces local cues",
"Zheng et al. [35] propose AMC+SWM to solve the par-",
"in each part feature, hence effectively alleviates the negative",
"tial reID. AMC collects patches from all gallery images and",
"effects of adjacent noises in part feature learning.",
"set up a dictionary for patch level matching. He et al. [7]",
"We test our approach on current person search bench-",
"propose DSR to reconstruct the partial image by holistic",
"marks, i.e., PRW [33] and CUHK-SYSU [24]. Experi-",
"image and utilize the reconstruction error as the similarity.",
"ment results show that, our APNet achieves competitive",
"Both AMC [35] and DSR [7] are optimisation-based thus",
"performance compared with recent works. To test our ap-",
"it is expensive to get the similarity of each query-gallery",
"proach in more challenging scenarios, we contribute a new",
"pair. PGFA [14] applies partial matching scheme with the",
"person search dataset named as Large-Scale dataset for",
"aligned inputimages,whoseoccludedpartsareindicatedby",
"Person Search in the wild (LSPS). Compared with PRW",
"the expensive offline keypoint detector. Sun et al. [21] pro-",
"andCUHK-SYSU.LSPSpresentsseveralnewfeaturesthat",
"pose a VPM to extract local features while simutaneously",
"would encourage the research towards more realistic per-",
"being aware of the visibility of each body part. However,",
"son search: 1) large scale: it contains a large number of",
"VPM does not consider the deformation mentioned in DSR.",
"pedestriansandboundingboxes;2)automaticallydetected",
"Our method address the deformation by an explicit align-",
"bounding boxes, which differ with the manually annotated",
"ment step and gets rid of the time-consuming comparsion",
"ones in PRW and CUHK-SYSU; 3) collected in scenarios",
"of each query-gallery pair during the inference stage.",
"with occlusions and crowdings. On this challenging LSPS,",
"Person search considers the detection stage in raw video",
" our method exhibits substantial advantages.",
"frames before relD. Current methods of person search can",
"Compared with relD, person search is still relatively",
"be divided into two categories. One integrates detection and",
"under-explored.It is critical to deal with the challenge",
"reID into a unified framework [24.15,4.29l.Based on",
"occuring in both person detection and reID. To the best",
"Faster-RCNN [16], OIM proposed by Xiao et al. [24] in-",
"of our knowledge, the proposed APNet is an early at-",
"serts an additional feature extraction branch on top of the",
"tempt in boosting the robustness to misalignment issue in",
"detectorhead,whichisthe firstend-to-end learning frame-",
"person search. As shown in our experiments, the AP-",
"work for person search. Similar to NPSM of Liu et al. [4]",
"Net brings considerable performance gains. The proposed",
"Munjal et al. [15] apply a query-guided method for person",
"LSPS dataset simulates realistic scenarios and presents a",
"search while it is end-to-end optimized compared to NPSM",
"more challenging person search task than existing datasets.",
"The other category solves detection and re-identification",
"We believe that, the proposed methods and the contributed",
"with two separate models, i.e., performing person search",
"LSPS can inspire more research efforts in person search.",
"in a sequential manner [1, 28, 3]. Chen et al. [1l propose",
"a mask-guided feature learning method to make the reID",
"2.Related Work",
"network focus more on foregrounds. Lan et al. [28] pro-",
"pose CLSA which utilizes multi-level features from reID",
"This paper is closely related to person reID and person",
"network to solve the multi-scale matching problem.Han",
"search. The following parts review those two lines of works.",
"et al. [3] claim that bounding box from detector might not",
" Person relD extracts robust and discriminative features",
"from given person bounding boxes.Current fully su-",
"beoptimalforre-identification,theythusrefinethedetector",
"results driven by re-identification loss.",
"pervised reID methods can be briefly summarized into",
"the following categories. 1) Learn local features of re-",
"Our work belongs to the first category of person search,",
"gions utilizing offline detector to split body into several re-",
"integrating detection and relD into a unified framework.",
"Different from existing person search works, this work fur-",
"gions [27, 17, 10, 11], or uniformly split the body into sev-",
"ther addresses the misalignment issue of detected bounding",
"eral stripes [31, 22, 26]; 2) Learn the robust feature utiliz-",
"ing additional attribute annotation [19, 18, 20, 23]; 3] En-",
"boxes and optimizes aligned part feature learning to achieve",
"robust partial matching in person search.",
"hance the discriminativeness of feature by attention mech-",
"anism [12,9,25]; 4)Impose constraints on feature space",
"3.ProblemFormulation",
"with loss functions, like verification loss [36] and triplet",
"loss [8]. Among the above methods, PCB [22] and its",
"Given a query person image q and a gallery set g =",
"variants [31, 26] dominate the state-of-the-art performance.",
"giji=l:N containing N frames, person search aims at"
] |
{
"arxiv_id": "2004.09329",
"language": "en",
"page_idx": 1
}
|
||
2004.09329-en-2
|
2004.09329-en-2.jpg
|
<FigureHere>
Figure 2. Architecture of the proposed Align-to-Part Network ( APNet), which consists of person detector, Bounding Box Aligner ( BBA), and Region Sensitive Feature Extractor (RSFE). The detection branch returns bounding boxes $\mathcal{B}$ and global feature $F.$ BBA refines detected boxes $\tilde{B}$ to aligned ones, $i.e.\hat{B}.$ Red and green bounding boxes denotes detected boxes $\mathcal{B}$ and refined boxes $\hat{\mathcal{B}}$, respectively. RSFE extracts part features $\{f^{(k)}\}_{k=1:K}$ from refined boxes and alleviates negative effects of adjacent noises. RAP and GAP denote Regional Average Pooling and Global Average Pooling, respectively. $F_{\mathrm{RSM}}$ is the global feature from Region Sensitive Map, which is only used for training $RSFE.$
objective as
(2)
$$\mathcal{L}=\mathcal{L}_{D}+\mathcal{L}_{P}+\sum_{k=1}^{K}\mathcal{L}_{f^{(k)}},$$
where $\mathcal{L}_D$ denotes loss of the detector, which is optimized with both bounding box locations and person reID. $\mathcal{L}_P$ denotes the loss of part identification. $\mathcal{L} _{f}^{( k) }$ evaluates the dis- criminative power of the $k$- th part feature, which can be implemented with reID loss. The following section presents details of our implementation to person detector, part identifier, part feature extraction, as well as the network optimization.
detecting a collection of person bounding boxes $\mathcal{B}=$ $\{b_i\}_{i=1:M}$ from $\mathcal{G}$, then matching $q$ against the bounding boxes in $\mathcal{B}.$ Suppose a person can be divided into $K$ body parts, we could represent the bounding box $b_i$ containing a complete person as a collection of $K$ parts, $i.e.$, ${\mathcal P}_{i}=\{p_{i}^{(k)}\}_{k=1:K}$,where $p_{i}^{(k)}$ denotes the $k$-th part.
With the above formulation, person detection is expected to return bounding boxes containing complete persons with $K$ parts. Person reID targets to extract a discriminative feature to identify the same identity of $q$ in $\mathcal{B}.$ Most of existing works extract a global feature $F$ and performs reID by L2-distance. Due to the occlusions or background clutters, certain bounding boxes only contain portions of body parts, making the global feature degrades to a partial feature, $i.e.$, the global feature is extracted from $l,\bar{l}<K$ visible parts. Such partial features lead to inaccurate person matching when compared with global features.
Our solution is to introduce a part identifer E(-) to identify visible parts in each detected bounding box, $i.e.$, $\mathcal{P}_{i}=$E$( b_i) . \textbf{ With identified parts, the person image match- }$ ing could be treated as a part feature matching task, where features on mutual visible parts of two bounding boxes are matched. The distance between $q$ and $b_i$ can be denoted as, $i.e.$,
$$\operatorname{dist}_P(q,b_i)=\frac{\sum\limits_{k\in\mathcal{P}_q\cap\mathcal{P}_i}\operatorname{D}(f_q^{(k)},f_i^{(k)})}{|\mathcal{P}_q\cap\mathcal{P}_i|},$$
## $4. \textbf{Proposed Methods}$
We propose the Align-to-Part Network (APNet) to implement the formulation in Sec. 3. The architecture of APNet is shown in Fig. 2. APNet consists of a person detector, Bounding Box Aligner (BBA), and the Region Sensitive Feature Extractor (RSFE), respectively. The following parts present details of those components.
The detector is built upon OIM [24], which is an end-toend person detector returning bounding boxes $\mathcal{B}$ as well as their corresponding global feature $F.$ As shown in Fig. 2, the detector is trained with RPN loss $(\mathcal{L}_\mathrm{rpncls},\mathcal{L}_\mathrm{rpnreg})[16]$, ROI Head loss $(\mathcal{L}_{\mathrm{cls}},\mathcal{L}_{\mathrm{reg}})$, as well as reID loss. We denote the detector training loss as
(1)
$$\mathcal{L}_{D}=\mathcal{L}_{\mathrm{cls}}+\mathcal{L}_{\mathrm{reg}}+\mathcal{L}_{\mathrm{rpncls}}+\mathcal{L}_{\mathrm{rpnreg}}+\mathcal{L}_{\mathrm{ID}},$$
(3)
where $f_q^{(k)}$ and $f_i^{(k)}$ are features extracted from $k$-th part
from $q$ and $b_i$, and D(-) refers the L2-distance.
The training of our person search model should guarantee an accurate person detector, the reliable part identifier, and discriminative part features. We formulate our training
where $\mathcal{L}_{\mathrm{cls}}$ and $\mathcal{L}_\mathrm{reg}$ denotes the person classification loss and bounding box regression loss in ROI Head. $\mathcal{L}_\mathrm{rpncls}$ and $\mathcal{L}_\mathrm{rpnreg}$ denote the objectiveness loss and proposal regression loss in RPN. $\mathcal{L}_\mathrm{ID}$ is the reID loss computed on the
|
[
[
0.34228515625,
0.0921630859375,
0.464111328125,
0.09344482421875,
0.464111328125,
0.10479736328125,
0.34228515625,
0.10418701171875
],
[
0.615234375,
0.0889892578125,
0.80126953125,
0.0889892578125,
0.80126953125,
0.1029052734375,
0.615234375,
0.1029052734375
],
[
0.2900390625,
0.11297607421875,
0.35302734375,
0.11297607421875,
0.35302734375,
0.12371826171875,
0.2900390625,
0.12371826171875
],
[
0.5458984375,
0.11553955078125,
0.61669921875,
0.11553955078125,
0.61669921875,
0.126953125,
0.5458984375,
0.126953125
],
[
0.63623046875,
0.10986328125,
0.69921875,
0.10986328125,
0.69921875,
0.121826171875,
0.63623046875,
0.121826171875
],
[
0.451904296875,
0.117431640625,
0.468017578125,
0.117431640625,
0.468017578125,
0.1337890625,
0.451904296875,
0.1337890625
],
[
0.650390625,
0.121826171875,
0.685546875,
0.121826171875,
0.685546875,
0.1337890625,
0.650390625,
0.1337890625
],
[
0.7939453125,
0.1300048828125,
0.87744140625,
0.1363525390625,
0.875,
0.15283203125,
0.79150390625,
0.146484375
],
[
0.293212890625,
0.136962890625,
0.3544921875,
0.136962890625,
0.3544921875,
0.147705078125,
0.293212890625,
0.147705078125
],
[
0.310546875,
0.1483154296875,
0.3447265625,
0.1483154296875,
0.3447265625,
0.160400390625,
0.310546875,
0.160400390625
],
[
0.72607421875,
0.1610107421875,
0.765625,
0.1610107421875,
0.765625,
0.1717529296875,
0.72607421875,
0.1717529296875
],
[
0.143798828125,
0.17041015625,
0.191162109375,
0.17041015625,
0.191162109375,
0.1849365234375,
0.143798828125,
0.1849365234375
],
[
0.2197265625,
0.17236328125,
0.2320556640625,
0.1817626953125,
0.221435546875,
0.1900634765625,
0.2091064453125,
0.18115234375
],
[
0.28515625,
0.169189453125,
0.358642578125,
0.169189453125,
0.358642578125,
0.184326171875,
0.28515625,
0.184326171875
],
[
0.810546875,
0.18115234375,
0.84228515625,
0.18115234375,
0.84228515625,
0.1925048828125,
0.810546875,
0.1925048828125
],
[
0.464111328125,
0.2083740234375,
0.476318359375,
0.2083740234375,
0.476318359375,
0.2183837890625,
0.464111328125,
0.2183837890625
],
[
0.7236328125,
0.210205078125,
0.765625,
0.210205078125,
0.765625,
0.2203369140625,
0.7236328125,
0.2203369140625
],
[
0.82421875,
0.212158203125,
0.84326171875,
0.212158203125,
0.84326171875,
0.2203369140625,
0.82421875,
0.2203369140625
],
[
0.3349609375,
0.2247314453125,
0.38232421875,
0.2247314453125,
0.38232421875,
0.23486328125,
0.3349609375,
0.23486328125
],
[
0.5595703125,
0.2279052734375,
0.61669921875,
0.2259521484375,
0.61669921875,
0.2374267578125,
0.5595703125,
0.2392578125
],
[
0.65283203125,
0.2266845703125,
0.70361328125,
0.2266845703125,
0.70361328125,
0.238037109375,
0.65283203125,
0.238037109375
],
[
0.343017578125,
0.236083984375,
0.37744140625,
0.236083984375,
0.37744140625,
0.2474365234375,
0.343017578125,
0.2474365234375
],
[
0.178955078125,
0.2291259765625,
0.3056640625,
0.2291259765625,
0.3056640625,
0.242431640625,
0.178955078125,
0.242431640625
],
[
0.62841796875,
0.240478515625,
0.73388671875,
0.240478515625,
0.73388671875,
0.25244140625,
0.62841796875,
0.25244140625
],
[
0.12335205078125,
0.25830078125,
0.17236328125,
0.25830078125,
0.17236328125,
0.26953125,
0.12335205078125,
0.26953125
],
[
0.080078125,
0.290283203125,
0.890625,
0.290283203125,
0.890625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.080078125,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.080078125,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.080078125,
0.33203125
],
[
0.07843017578125,
0.330810546875,
0.8935546875,
0.330810546875,
0.8935546875,
0.348388671875,
0.07843017578125,
0.348388671875
],
[
0.080078125,
0.3466796875,
0.89208984375,
0.3466796875,
0.89208984375,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.07843017578125,
0.359130859375,
0.1217041015625,
0.359130859375,
0.1217041015625,
0.373779296875,
0.07843017578125,
0.373779296875
],
[
0.080078125,
0.39013671875,
0.44775390625,
0.39013671875,
0.44775390625,
0.404052734375,
0.080078125,
0.404052734375
],
[
0.44287109375,
0.392578125,
0.465576171875,
0.392578125,
0.465576171875,
0.399658203125,
0.44287109375,
0.399658203125
],
[
0.50244140625,
0.39013671875,
0.5849609375,
0.39013671875,
0.5849609375,
0.401611328125,
0.50244140625,
0.401611328125
],
[
0.08251953125,
0.406005859375,
0.468017578125,
0.406005859375,
0.468017578125,
0.42041015625,
0.08251953125,
0.42041015625
],
[
0.7255859375,
0.41357421875,
0.74609375,
0.41357421875,
0.74609375,
0.425537109375,
0.7255859375,
0.425537109375
],
[
0.080078125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.60107421875,
0.421630859375,
0.79150390625,
0.423583984375,
0.79150390625,
0.4443359375,
0.60107421875,
0.442626953125
],
[
0.869140625,
0.427978515625,
0.8935546875,
0.427978515625,
0.8935546875,
0.44140625,
0.869140625,
0.44140625
],
[
0.080078125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.72119140625,
0.4443359375,
0.7490234375,
0.4443359375,
0.7490234375,
0.455810546875,
0.72119140625,
0.455810546875
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.0767822265625,
0.46337890625,
0.426513671875,
0.462158203125,
0.426513671875,
0.484130859375,
0.0767822265625,
0.48486328125
],
[
0.50341796875,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50341796875,
0.477783203125
],
[
0.1004638671875,
0.48486328125,
0.46728515625,
0.48486328125,
0.46728515625,
0.498779296875,
0.1004638671875,
0.498779296875
],
[
0.50244140625,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.498291015625,
0.4892578125,
0.896484375,
0.4892578125,
0.896484375,
0.51318359375,
0.498291015625,
0.51318359375
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.08087158203125,
0.5146484375,
0.468017578125,
0.5146484375,
0.468017578125,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.50244140625,
0.51123046875,
0.890625,
0.51123046875,
0.890625,
0.5244140625,
0.50244140625,
0.5244140625
],
[
0.080078125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.50244140625,
0.5263671875,
0.89208984375,
0.5263671875,
0.89208984375,
0.54052734375,
0.50244140625,
0.54052734375
],
[
0.07843017578125,
0.54443359375,
0.468017578125,
0.54345703125,
0.468017578125,
0.55810546875,
0.07843017578125,
0.5595703125
],
[
0.50244140625,
0.54150390625,
0.89208984375,
0.54150390625,
0.89208984375,
0.5546875,
0.50244140625,
0.5546875
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.556640625,
0.890625,
0.556640625,
0.890625,
0.57080078125,
0.50244140625,
0.57080078125
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.57373046875,
0.56787109375,
0.57373046875,
0.56787109375,
0.58544921875,
0.50244140625,
0.58544921875
],
[
0.07843017578125,
0.59033203125,
0.468017578125,
0.5888671875,
0.468017578125,
0.60400390625,
0.07843017578125,
0.60498046875
],
[
0.50341796875,
0.5986328125,
0.6806640625,
0.5986328125,
0.6806640625,
0.61279296875,
0.50341796875,
0.61279296875
],
[
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,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.52294921875,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63818359375,
0.52294921875,
0.63818359375
],
[
0.080078125,
0.63623046875,
0.386474609375,
0.63623046875,
0.386474609375,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50244140625,
0.6396484375,
0.890625,
0.63818359375,
0.890625,
0.65283203125,
0.50244140625,
0.65380859375
],
[
0.098876953125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.098876953125,
0.66552734375
],
[
0.50244140625,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66845703125,
0.50244140625,
0.66845703125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66748046875,
0.468017578125,
0.681640625,
0.080078125,
0.68115234375
],
[
0.50244140625,
0.66943359375,
0.89208984375,
0.66943359375,
0.89208984375,
0.68310546875,
0.50244140625,
0.68310546875
],
[
0.080078125,
0.68115234375,
0.468017578125,
0.681640625,
0.468017578125,
0.6962890625,
0.080078125,
0.69580078125
],
[
0.5009765625,
0.6845703125,
0.89208984375,
0.68505859375,
0.89208984375,
0.69970703125,
0.5009765625,
0.69873046875
],
[
0.080078125,
0.69873046875,
0.4697265625,
0.69873046875,
0.4697265625,
0.7119140625,
0.080078125,
0.7119140625
],
[
0.5009765625,
0.69970703125,
0.740234375,
0.69970703125,
0.740234375,
0.71337890625,
0.5009765625,
0.71337890625
],
[
0.080078125,
0.71337890625,
0.4697265625,
0.71337890625,
0.4697265625,
0.7265625,
0.080078125,
0.7265625
],
[
0.52294921875,
0.71533203125,
0.890625,
0.71533203125,
0.890625,
0.72900390625,
0.52294921875,
0.72900390625
],
[
0.080078125,
0.72900390625,
0.4697265625,
0.72900390625,
0.4697265625,
0.7421875,
0.080078125,
0.7421875
],
[
0.50244140625,
0.73095703125,
0.89208984375,
0.73095703125,
0.89208984375,
0.744140625,
0.50244140625,
0.744140625
],
[
0.07843017578125,
0.74365234375,
0.1070556640625,
0.74560546875,
0.10540771484375,
0.75830078125,
0.0767822265625,
0.755859375
],
[
0.5009765625,
0.74560546875,
0.89208984375,
0.744140625,
0.89208984375,
0.7587890625,
0.5009765625,
0.76025390625
],
[
0.30224609375,
0.76025390625,
0.351318359375,
0.7666015625,
0.34716796875,
0.7822265625,
0.298095703125,
0.77587890625
],
[
0.5,
0.7568359375,
0.89208984375,
0.7587890625,
0.89208984375,
0.7763671875,
0.5,
0.7744140625
],
[
0.268798828125,
0.7646484375,
0.2900390625,
0.7646484375,
0.2900390625,
0.7763671875,
0.268798828125,
0.7763671875
],
[
0.343017578125,
0.7666015625,
0.35302734375,
0.7666015625,
0.35302734375,
0.77099609375,
0.343017578125,
0.77099609375
],
[
0.366943359375,
0.7646484375,
0.39208984375,
0.7646484375,
0.39208984375,
0.77294921875,
0.366943359375,
0.77294921875
],
[
0.346435546875,
0.7734375,
0.37744140625,
0.7734375,
0.37744140625,
0.7763671875,
0.346435546875,
0.7763671875
],
[
0.50244140625,
0.77587890625,
0.89208984375,
0.77587890625,
0.89208984375,
0.7890625,
0.50244140625,
0.7890625
],
[
0.2998046875,
0.76806640625,
0.303955078125,
0.76806640625,
0.303955078125,
0.7734375,
0.2998046875,
0.7734375
],
[
0.2467041015625,
0.77783203125,
0.3095703125,
0.77783203125,
0.3095703125,
0.7890625,
0.2467041015625,
0.7890625
],
[
0.148681640625,
0.7861328125,
0.24267578125,
0.7861328125,
0.24267578125,
0.79736328125,
0.148681640625,
0.79736328125
],
[
0.44677734375,
0.7861328125,
0.4697265625,
0.7861328125,
0.4697265625,
0.798828125,
0.44677734375,
0.798828125
],
[
0.291748046875,
0.79345703125,
0.3544921875,
0.79345703125,
0.3544921875,
0.80810546875,
0.291748046875,
0.80810546875
],
[
0.50244140625,
0.79150390625,
0.685546875,
0.79150390625,
0.685546875,
0.8046875,
0.50244140625,
0.8046875
],
[
0.54150390625,
0.81298828125,
0.86669921875,
0.81494140625,
0.86669921875,
0.83349609375,
0.54150390625,
0.83056640625
],
[
0.859375,
0.8173828125,
0.890625,
0.8173828125,
0.890625,
0.8291015625,
0.859375,
0.8291015625
],
[
0.080078125,
0.8388671875,
0.386474609375,
0.8388671875,
0.386474609375,
0.85302734375,
0.080078125,
0.85302734375
],
[
0.07843017578125,
0.818359375,
0.47216796875,
0.818359375,
0.47216796875,
0.8388671875,
0.07843017578125,
0.8388671875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.1004638671875,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.1004638671875,
0.86865234375
],
[
0.5009765625,
0.85498046875,
0.89208984375,
0.85693359375,
0.89208984375,
0.8720703125,
0.5009765625,
0.869140625
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.8935546875,
0.87060546875,
0.8935546875,
0.88818359375,
0.5009765625,
0.88818359375
],
[
0.0767822265625,
0.88525390625,
0.4697265625,
0.88720703125,
0.4697265625,
0.9013671875,
0.0767822265625,
0.8994140625
],
[
0.50244140625,
0.88623046875,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Bounding BoxAligner",
"Region Sensitive Feature Extractor",
"0,O%01o",
"Rol Align",
"Extractor",
"B",
"Head",
"(f(k)/ k=1:K",
"Estimator",
"Head",
"RAP",
"Stem",
"",
"RolAlign",
"Kxd",
"B",
"GAP",
"RSM",
"Detector",
"HxWxD",
"Hxwxd",
"Head",
"RPNRolAlign",
"Region Sensitive Map",
"Detector",
"Figure 2. Architecture of the proposed Align-to-Part Network (APNet), which consists of person detector, Bounding Box Aligner (BBA),",
"and Region Sensitive Feature Extractor (RSFE). The detection branch returns bounding boxes B and global feature F. BBA refines detected",
"boxes B to aligned ones, i.e., B. Red and green bounding boxes denotes detected boxes B and refined boxes B, respectively. RSFE extracts",
"part features ( f(k) k=1:k from refined boxes and alleviates negative effects of adjacent noises. RAP and GAP denote Regional Average",
"Pooling and Global Average Pooling, respectively. FrsM is the global feature from Region Sensitive Map, which is only used for training",
"RSFE.",
"detecting a collection of person bounding boxes B",
"",
"objective as.",
"[bi]i-i:M from G, then matching q against the bound-",
"K",
"ing boxes in B. Suppose a person can be divided into K",
"C=CD+ Cp+C(n),",
"(2)",
"body parts, we could represent the bounding box b, con-",
"k=1",
"taining a complete person as a collection of K parts, i.e.,",
"P; = (p()k=1:K, where p(*) denotes the k-th part.",
"where LD denotes loss of the detector, which is optimized",
"With the above formulation,person detection is expected",
"with both bounding box locations and person reID. Lp de-",
"notes the loss of part identification. C)evaluates the dis-",
"to return bounding boxes containing complete persons with",
"K parts. Person reID targets to extract a discriminative fea-",
"criminative power of the k-th part feature, which can be im-",
"ture to identify the same identity of q in B. Most of ex-",
"plemented with relD loss.The following section presents",
"isting works extract a global feature F and performs reID",
"details of our implementation to person detector, part iden-",
"by L2-distance. Due to the occlusions or background clut-",
"tifier, part feature extraction, as well as the network opti-",
"ters, certain bounding boxes only contain portions of body",
"mization.",
"parts, making the global feature degrades to a partial fea-",
"4. Proposed Methods",
"ture, i.e., the global feature is extracted from l, l < K vis-",
"ible parts. Such partial features lead to inaccurate person",
"We propose the Align-to-Part Network (APNet) to im-",
"matching when compared with global features.",
"plement the formulation in Sec. 3. The architecture of AP-",
"Our solution is to introduce a part identifier E() to",
"Net is shown in Fig. 2. APNet consists of a person detec-",
"identify visible parts in each detected bounding box, i.e.,",
"tor, Bounding Box Aligner (BBA), and the Region Sensitive",
"P, =E(b,).With identified parts, the person image match-",
"Feature Extractor (RSFE), respectively. The following parts",
"ing could be treated as a part feature matching task, where",
"present details of those components.",
"features on mutual visible parts of two bounding boxes are",
"The detector is built upon OIM [24], which is an end-to-",
"matched. The distance between q and b; can be denoted as,",
"end person detector returning bounding boxes B as well as",
"i.e.,",
"their corresponding global feature F. As shown in Fig. 2,",
"D(fq",
"the detector is trained with RPN loss (Crpncls, Crpnreg) [16],",
"7",
"",
"e(k)",
"",
"ROI Head loss (Lels, Lreg ), as well as relD loss. We denote",
"",
"keP.nP,",
"distp(q, b,) =",
"(1)",
"IPg n Pil",
"the detector training loss as",
"CD = Lels + Creg + Crpncls + Crpnreg + LID,",
"(3)",
"from q and bi, and D() refers the L2-distance.",
"where f(k) and f(k) are features extracted from k-th part",
"where Lcls and Lreg denotes the person classification loss",
"Thetraining of ourperson searchmodel should guaran-",
"and bounding box regression loss in ROI Head. Crpncls",
"tee an accurate person detector, the reliable part identifier,",
"and Lrpnreg denote the objectiveness loss and proposal re-",
"and discriminative part features. We formulate our training",
"gression loss in RPN. LiD is the reID loss computed on the"
] |
{
"arxiv_id": "2004.09329",
"language": "en",
"page_idx": 2
}
|
||
2004.09329-en-3
|
2004.09329-en-3.jpg
|
<FigureHere>
Figure 3. Illustration of BBA for bounding box refinement. Red and green boxes refer to detected boxes and refined boxes, respectively. BBA predicts four bounding box offset values $o_t,o_b,o_l$, and $o_r$, which shift the detected bounding boxes to cover the estimated holistic body region. 0,1 indicate the validity of each part. global feature. We refer readers to OIM [24] for more details of the loss computation.
As shown in Fig. 1, a well trained detector may produce misaligned person bounding boxes. We hence design a part estimator with BBA to estimate visible body parts of each detected box.
## $4. 1. \textbf{Bounding Box Aligner}$
BBA implements part identifer E(·) in Sec. 3 to identify visible parts in each bounding box. This could be achieved with various methods, e.g., through segmenting person foregrounds[13] or estimating body keypoints[2]. However, those methods require extra annotations and considerable computations. We implement BBA with a more efficient solution.
Because most pedestrians show upright posture in surveillance videos, aligned body parts can be extracted by dividing a holistic body region into horizontal and vertical stripes. This operation generates aligned part regions, e.g., top and bottom horizontal stripes correspond to head and foot, respectively. With this intuition, BBA first refines detected bounding boxes, then extracts horizontal and vertical stripes as body parts. We illustrate this procedure in Fig. 3.
$\hat{\text{BBA estimates an offset vector}O}=\{o_{t},o_{b},o_{l},o_{r}\}$ to refine each detected bounding box to cover the holistic body region, where each offset value is in the range of [-1,1]. The four offset values move the top, bottom, left, and right boundaries of each bounding box. Fig. 3 shows examples of moving the boundaries by $o_{t}$ and $o_{b}.$ Suppose the coordinate of a bounding box $b$ is $\{x_min,y_{min},x_{max},y_{max}\}$, its coordinate after refinement with $O$ can be denoted as $\{\hat{x}_{min},\hat{y}_{min},\hat{x}_{max},\hat{y}_{max}\},e.g.$,
$$\hat{y}_{min}=y_{min}-\frac{h\cdot o_{t}}{1-o_{t}-o_{b}}\\\hat{y}_{max}=y_{max}+\frac{h\cdot o_{b}}{1-o_{t}-o_{b}},$$
(4)
where $h$ is the height of detected bounding box, computed
as $y_{max}-y_{min}.$ Similar computation can be applied to $\begin{array}{c}{\text{compute the }\hat{x}_{min}\mathrm{~and~}\hat{x}_{max}\mathrm{~with~}o_{l}\mathrm{~and~}o_{r}.}\\{\text{We denote the refined bounding box as }\hat{b},\mathrm{~which~is~used}}\\\end{array}$ to extracted horizontal and vertical stripes as parts. As shown in Fig. 3, refinements may introduce occluded parts and noises into $\hat{b}.$ To extract $K$ horizontal stripes, we introduce a $K$-dim validity vector $v$ to record the visibility of each stripe. The $k$-th stripe is considered as valid, $i.e.$, $v[k]=1$,if
$$\lceil K\cdot\max(0,o_t)\rceil\leq k\leq K-\lfloor K\cdot\max(0,o_b)\rfloor.$$
(5)
Similar computation is applied to extract valid vertical stripes. The final valid part collection $\mathcal{P}$ of each bounding box collects valid stripes.
As shown in Fig. 2, BBA predicts O based on the bounding box feature extracted with ROIAlign. BBA can be trained by automatically generating training data. We first crop bounding boxes from frames by their groundtruth coordinates, which we denote as $\mathcal{B}_gt.$ AlphaPose[2] is utilized to estimate the keypoints of each bounding box, which provide cues about missing body parts. We then trans- $\mathop{\mathrm{form~}}\mathcal{B}_{gt}$ to $\hat{\mathcal{B}}_{gt}$ to cover the holistic body region. Comparing $\mathcal{B}_{gt}$ and $\hat{\mathcal{B}}_{gt}$ generates the groundtruth offset labels $O_{gt}=\{o_{gt-t},o_{gt-b},o_{gt-l},o_{gt-r}\}.$The training of BBA can thus be supervised by following loss
(6)
$$\begin{aligned}\mathcal{L}_P&=\sum_{i\in\{t,b,l,r\}}\sinh_{l1}(o_i,o_{gt-i}),\end{aligned}$$
where smooth$_l1$ computes differences between predicted offset values and the groundtruth values. More details of smooth$_l1$ can be found in [16].
## 4.2. Region Sensitive Feature Extractor
The part collection $\mathcal{P}$ makes it possible to extract part features for partial matching. The following section shows our method to extract horizontal stripe features. Vertical stripe features can be extracted with similar way.
Vanilla Part Feature Extractor: Part features can be extracted by applying Region Average Pooling (RAP) on the feature map of video frame. As shown in Fig. 2, for a refined bounding box $\hat{b}_i$, we first extract its feature map from the frame feature map $\mathcal{M}$ with ROIAlign, which is $\begin{array}{l}\text{then input into a convolutional block to generate a feature}\\\text{map }T\in\mathbb{R}^{H\times W\times D}.&\text{We denote part features of }\hat{b}_{i}\end{array}$ tracted by RAP as,
$$\{\bar{f}_i^{(k)}\}_{k=1:l}=\mathrm{RAP}(T,\mathcal{P}_i),l=|\mathcal{P}_i|,$$
(7)
where $l$ denotes the number of valid horizontal stripes of
bounding box $\hat{b_i}.$
Part feature learning can be achieved by computing the
reID loss on each valid part feature, $i.e.$,
(8)
$$\mathcal{L}_{f}^{(k)}=\mathcal{L}_{\mathrm{ID}}(\bar{f}^{(k)},y),$$
|
[
[
0.5009765625,
0.0921630859375,
0.89208984375,
0.09027099609375,
0.89208984375,
0.10479736328125,
0.5009765625,
0.106689453125
],
[
0.50244140625,
0.10797119140625,
0.7900390625,
0.10797119140625,
0.7900390625,
0.12249755859375,
0.50244140625,
0.12249755859375
],
[
0.521484375,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.521484375,
0.1363525390625
],
[
0.50244140625,
0.1383056640625,
0.8935546875,
0.1383056640625,
0.8935546875,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.50341796875,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1944580078125,
0.50341796875,
0.1944580078125
],
[
0.400390625,
0.199462890625,
0.44677734375,
0.199462890625,
0.44677734375,
0.2139892578125,
0.400390625,
0.2139892578125
],
[
0.5009765625,
0.1976318359375,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.5009765625,
0.2108154296875
],
[
0.30810546875,
0.2008056640625,
0.348876953125,
0.1988525390625,
0.3505859375,
0.212158203125,
0.3095703125,
0.2139892578125
],
[
0.12091064453125,
0.2020263671875,
0.160888671875,
0.199462890625,
0.16259765625,
0.2108154296875,
0.1217041015625,
0.21337890625
],
[
0.2156982421875,
0.2008056640625,
0.2548828125,
0.1988525390625,
0.256591796875,
0.2108154296875,
0.21728515625,
0.21337890625
],
[
0.5048828125,
0.2139892578125,
0.58251953125,
0.2139892578125,
0.58251953125,
0.2259521484375,
0.5048828125,
0.2259521484375
],
[
0.08087158203125,
0.2279052734375,
0.4697265625,
0.2279052734375,
0.4697265625,
0.2412109375,
0.08087158203125,
0.2412109375
],
[
0.529296875,
0.236083984375,
0.85693359375,
0.236083984375,
0.85693359375,
0.250732421875,
0.529296875,
0.250732421875
],
[
0.86279296875,
0.2412109375,
0.869140625,
0.2412109375,
0.869140625,
0.24560546875,
0.86279296875,
0.24560546875
],
[
0.86767578125,
0.238037109375,
0.890625,
0.238037109375,
0.890625,
0.25,
0.86767578125,
0.25
],
[
0.07843017578125,
0.255126953125,
0.468017578125,
0.255615234375,
0.468017578125,
0.270751953125,
0.07843017578125,
0.26953125
],
[
0.080078125,
0.242431640625,
0.46728515625,
0.242431640625,
0.46728515625,
0.255615234375,
0.080078125,
0.255615234375
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.260009765625,
0.89208984375,
0.274658203125,
0.50244140625,
0.2734375
],
[
0.08087158203125,
0.271484375,
0.46728515625,
0.271484375,
0.46728515625,
0.2822265625,
0.08087158203125,
0.2822265625
],
[
0.50341796875,
0.27587890625,
0.89208984375,
0.27587890625,
0.89208984375,
0.290283203125,
0.50341796875,
0.290283203125
],
[
0.07843017578125,
0.28271484375,
0.464111328125,
0.28271484375,
0.464111328125,
0.296630859375,
0.07843017578125,
0.296630859375
],
[
0.50244140625,
0.291748046875,
0.67236328125,
0.291748046875,
0.67236328125,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.51953125,
0.30615234375,
0.890625,
0.304931640625,
0.890625,
0.3193359375,
0.51953125,
0.32080078125
],
[
0.08087158203125,
0.31640625,
0.46728515625,
0.31640625,
0.46728515625,
0.330078125,
0.08087158203125,
0.330078125
],
[
0.50244140625,
0.3212890625,
0.8935546875,
0.3212890625,
0.8935546875,
0.335205078125,
0.50244140625,
0.335205078125
],
[
0.08087158203125,
0.333251953125,
0.268798828125,
0.333251953125,
0.268798828125,
0.343505859375,
0.08087158203125,
0.343505859375
],
[
0.50244140625,
0.337158203125,
0.8935546875,
0.337158203125,
0.8935546875,
0.35107421875,
0.50244140625,
0.35107421875
],
[
0.098876953125,
0.3466796875,
0.4697265625,
0.3466796875,
0.4697265625,
0.361083984375,
0.098876953125,
0.361083984375
],
[
0.50244140625,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.080078125,
0.361083984375,
0.468017578125,
0.361083984375,
0.468017578125,
0.375,
0.080078125,
0.375
],
[
0.5009765625,
0.3662109375,
0.890625,
0.365478515625,
0.890625,
0.380126953125,
0.5009765625,
0.380615234375
],
[
0.080078125,
0.376953125,
0.4697265625,
0.376953125,
0.4697265625,
0.39013671875,
0.080078125,
0.39013671875
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.080078125,
0.393310546875,
0.1683349609375,
0.393310546875,
0.1683349609375,
0.40478515625,
0.080078125,
0.40478515625
],
[
0.50341796875,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410888671875,
0.50341796875,
0.410888671875
],
[
0.08087158203125,
0.416748046875,
0.281982421875,
0.416748046875,
0.281982421875,
0.429931640625,
0.08087158203125,
0.429931640625
],
[
0.50341796875,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50341796875,
0.425537109375
],
[
0.50341796875,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.09722900390625,
0.439453125,
0.46728515625,
0.439453125,
0.46728515625,
0.45263671875,
0.09722900390625,
0.45263671875
],
[
0.5009765625,
0.442626953125,
0.8935546875,
0.44140625,
0.8935546875,
0.458984375,
0.5009765625,
0.460205078125
],
[
0.08087158203125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.5009765625,
0.458984375,
0.771484375,
0.458984375,
0.771484375,
0.472900390625,
0.5009765625,
0.472900390625
],
[
0.080078125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.576171875,
0.484130859375,
0.62353515625,
0.484130859375,
0.62353515625,
0.498779296875,
0.576171875,
0.498779296875
],
[
0.66650390625,
0.48291015625,
0.81884765625,
0.486083984375,
0.81884765625,
0.50048828125,
0.66650390625,
0.49755859375
],
[
0.869140625,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.869140625,
0.498779296875
],
[
0.6396484375,
0.490478515625,
0.654296875,
0.490478515625,
0.654296875,
0.494384765625,
0.6396484375,
0.494384765625
],
[
0.08087158203125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.619140625,
0.50390625,
0.68115234375,
0.50390625,
0.68115234375,
0.51513671875,
0.619140625,
0.51513671875
],
[
0.08087158203125,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.50341796875,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.537109375,
0.50341796875,
0.537109375
],
[
0.080078125,
0.53076171875,
0.1380615234375,
0.53076171875,
0.1380615234375,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.098876953125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.098876953125,
0.5595703125
],
[
0.50244140625,
0.53857421875,
0.8935546875,
0.53857421875,
0.8935546875,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.5009765625,
0.5517578125,
0.71337890625,
0.5517578125,
0.71337890625,
0.56640625,
0.5009765625,
0.56640625
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50341796875,
0.57568359375,
0.8056640625,
0.57568359375,
0.8056640625,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.07843017578125,
0.5888671875,
0.468017578125,
0.59033203125,
0.468017578125,
0.60498046875,
0.07843017578125,
0.60400390625
],
[
0.521484375,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.521484375,
0.6123046875
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.62060546875,
0.46728515625,
0.62060546875,
0.46728515625,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.50341796875,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.08087158203125,
0.63720703125,
0.46728515625,
0.63720703125,
0.46728515625,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.5009765625,
0.64306640625,
0.82421875,
0.64599609375,
0.82421875,
0.6591796875,
0.5009765625,
0.6572265625
],
[
0.080078125,
0.65087890625,
0.464111328125,
0.65087890625,
0.464111328125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.521484375,
0.65771484375,
0.89208984375,
0.6591796875,
0.89208984375,
0.673828125,
0.521484375,
0.67236328125
],
[
0.098876953125,
0.6640625,
0.468017578125,
0.66650390625,
0.468017578125,
0.68115234375,
0.098876953125,
0.6787109375
],
[
0.5009765625,
0.6748046875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.08087158203125,
0.68310546875,
0.465576171875,
0.68310546875,
0.465576171875,
0.693359375,
0.08087158203125,
0.693359375
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.6962890625,
0.46728515625,
0.6962890625,
0.46728515625,
0.71044921875,
0.080078125,
0.71044921875
],
[
0.08087158203125,
0.7109375,
0.468017578125,
0.7109375,
0.468017578125,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.080078125,
0.72607421875,
0.4697265625,
0.72607421875,
0.4697265625,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.07843017578125,
0.74072265625,
0.468017578125,
0.7412109375,
0.468017578125,
0.755859375,
0.07843017578125,
0.7548828125
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.5,
0.74560546875,
0.8935546875,
0.74560546875,
0.8935546875,
0.7666015625,
0.5,
0.7666015625
],
[
0.07843017578125,
0.7548828125,
0.468017578125,
0.7568359375,
0.468017578125,
0.771484375,
0.07843017578125,
0.76953125
],
[
0.07843017578125,
0.77099609375,
0.471435546875,
0.771484375,
0.471435546875,
0.7861328125,
0.07843017578125,
0.78515625
],
[
0.50244140625,
0.76513671875,
0.62841796875,
0.76513671875,
0.62841796875,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.08087158203125,
0.78857421875,
0.29736328125,
0.78857421875,
0.29736328125,
0.8017578125,
0.08087158203125,
0.8017578125
],
[
0.60302734375,
0.78857421875,
0.82666015625,
0.78857421875,
0.82666015625,
0.80615234375,
0.60302734375,
0.80615234375
],
[
0.869140625,
0.79150390625,
0.89208984375,
0.79150390625,
0.89208984375,
0.8046875,
0.869140625,
0.8046875
],
[
0.310546875,
0.81005859375,
0.35302734375,
0.8125,
0.351318359375,
0.8271484375,
0.3095703125,
0.82470703125
],
[
0.57275390625,
0.79248046875,
0.609375,
0.787109375,
0.61376953125,
0.80322265625,
0.57666015625,
0.80810546875
],
[
0.1805419921875,
0.82080078125,
0.279296875,
0.82275390625,
0.279296875,
0.8369140625,
0.1805419921875,
0.83544921875
],
[
0.50341796875,
0.81494140625,
0.89208984375,
0.81494140625,
0.89208984375,
0.8291015625,
0.50341796875,
0.8291015625
],
[
0.291748046875,
0.8291015625,
0.373291015625,
0.83056640625,
0.373291015625,
0.84521484375,
0.291748046875,
0.84326171875
],
[
0.5,
0.8291015625,
0.61767578125,
0.830078125,
0.61767578125,
0.8447265625,
0.5,
0.84326171875
],
[
0.3056640625,
0.84228515625,
0.348876953125,
0.8447265625,
0.34716796875,
0.85986328125,
0.303955078125,
0.85791015625
],
[
0.44677734375,
0.8369140625,
0.4697265625,
0.8369140625,
0.4697265625,
0.85107421875,
0.44677734375,
0.85107421875
],
[
0.51953125,
0.8447265625,
0.89208984375,
0.84521484375,
0.89208984375,
0.85986328125,
0.51953125,
0.859375
],
[
0.1748046875,
0.85498046875,
0.27783203125,
0.85498046875,
0.27783203125,
0.869140625,
0.1748046875,
0.869140625
],
[
0.28759765625,
0.8623046875,
0.371826171875,
0.8623046875,
0.371826171875,
0.876953125,
0.28759765625,
0.876953125
],
[
0.5009765625,
0.859375,
0.765625,
0.85986328125,
0.765625,
0.875,
0.5009765625,
0.87353515625
],
[
0.62744140625,
0.8818359375,
0.76708984375,
0.8837890625,
0.76708984375,
0.90283203125,
0.62744140625,
0.90087890625
],
[
0.86767578125,
0.88525390625,
0.89453125,
0.88525390625,
0.89453125,
0.9013671875,
0.86767578125,
0.9013671875
],
[
0.08087158203125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
]
] |
[
"as mar -Ymin:Similar computation can be applied to",
"compute the &min and mar with oi and or.",
"We denote the refined bounding box as b, which is used",
"to extracted horizontal and vertical stripes as parts., As",
"shown in Fig. 3, refinements may introduce occluded parts",
"and noises into b. To extract K horizontal stripes, we in-",
"troduce a K-dim validity vector u torecord thevisibility",
"0b<0",
"of each stripe. The k-th stripe is considered as valid, i.e.,",
"0b>0",
"0,>0",
"01<0",
"u[] = 1, if",
"Figure 3. Illustration of BBA for bounding box refinement. Red",
"[K·max(O,ot)| ≤k≤K-K-max(0,ob)].",
"",
"(5)",
"tively. BBA predicts four bounding box offset values ot, Ob, ot,",
"and green boxes refer to detected boxes and refined boxes, respec-",
"Similar computation is applied to extract valid vertical",
"and or,which shiftthe detected bounding boxesto cover the esti-",
"stripes. The final valid part collection P of each bounding",
"mated holistic body region. O, Iindicate the validity of each part.",
"box collects valid stripes.",
"As shown in Fig. 2, BBA predicts O based on the bound-",
"global feature. We refer readers to OIM [24] for more de-",
"ing box feature extracted with ROIAlign. BBA can be",
"tails of theloss computation.",
"trained by automatically generating training data. We first",
"As shown in Fig. 1, a well trained detector may produce",
"crop bounding boxes from frames by their groundtruth co-",
"misaligned person bounding boxes. We hence design a part",
"ordinates, which we denote as Bgt. AlphaPose [2] is uti-",
"estimator with BBA to estimate visible body parts of each",
"lized to estimate the keypoints of each bounding box, which",
"detected box.",
"provide cues about missing body parts. We then trans-",
"4.1.Bounding Box Aligner",
"form Bat to Bqt to cover the holistic body region. Com-",
"paring Bgt and Bqt generates the groundtruth offset labels",
"BBA implements part identifier E() in Sec.3 to identify",
"Ogt = (ogt-t, Ogt-b, Ogt-1,Ogt-r]. The training of BBA",
"visible parts in each bounding box. This could be achieved",
"can thus be supervised by following loss",
"with various methods, e.g., through segmenting person fore-",
"Cp=",
"smoothui(oi,Ogt-i),",
"(6)",
"",
"grounds [13] or estimating body keypoints [2]. However,",
"those methods require extra annotations and considerable",
"ie(t,b,l,r]",
"computations.WeimplementBBAwithamoreefficient",
"where smoothi computes differences between predicted",
"solution.",
"Because most pedestrians show upright posture in",
"offset values and the groundtruth values. More details of",
"smoothu can be found in [16].",
"surveillance videos, aligned body parts can be extracted by",
"dividing a holistic body region into horizontal and vertical",
"4.2. Region Sensitive Feature Extractor",
"stripes. This operation generates aligned part regions, e.g.,",
"The part collection P makes it possible to extract part",
"top and bottom horizontal stripes correspond to head and",
"features for partial matching. The following section shows",
"foot, respectively. With this intuition, BBA first refines de-",
"our method to extract horizontal stripe features.Vertical",
"tectedboundingboxes.thenextractshorizontal and vertical",
"stripe features can be extracted with similar way.",
"stripes as body parts. We illustrate this procedure in Fig. 3",
"Vanilla Part Feature Extractor: Part features can be",
"BBA estimates an offset vector O = {ot, Ob, Ot, O,] to",
"extracted by applying Region Average Pooling (RAP) on",
"refineeachdetectedboundingboxtocovertheholisticbody",
"the feature map of video frame. As shown in Fig. 2, for",
"region, where each offset value is in the range of [-1, 1]",
"The four offset values move the top, bottom, left, and right",
"a refined bounding box bi, we first extract its feature map",
"from the frame feature map M with ROIAlign, which is",
"boundaries of each bounding box. Fig. 3 shows examples",
"of moving the boundaries by ot and ob. Suppose the co-",
"then input into a convolutional block to generate a feature",
"map T e RHxWxD, We denote part features of b; ex-",
"ordinate of a bounding box b is (αmin, Ymin, mar, Ymar],",
"its coordinate after refinement with O can be denoted as",
"tracted by RAP as,",
"[min, Ymin,Emar,Ymar],e.g.,",
"}k=1:l = RAP(T, P), = [Pl,",
"(7)",
"h-Ot",
"(f(k)",
"Ymin = Ymin",
"where l denotes the number of valid horizontal stripes of",
"1-0t-0b",
"bounding box bi.",
"h.Ob",
"(4)",
"Part feature learning can be achieved by computing the",
"9mar =Ymar",
"1-0t-0b",
"relD loss on each valid part feature, i.e.,",
"C(k) = LID(f(k), y),",
"(8)",
"where h is the height of detected bounding box, computed"
] |
{
"arxiv_id": "2004.09329",
"language": "en",
"page_idx": 3
}
|
||
2004.09329-en-4
|
2004.09329-en-4.jpg
|
<FigureHere>
Figure 4. Illustration of detected boxes (red) and refined boxes (green), as well as their feature maps. (b) and (c) show feature maps of refined boxes trained with vanilla part feature extractor and our RSFE. It is clear that, RSFE depresses noises on noisy and invalid regions.
where $\mathcal{L}_\mathrm{ID}$ refers to reID loss which is implemented with
OIM loss [24]. $y$ is the ground truth person ID label.
Fig. 4 (b) shows feature maps of aligned bounding boxes trained by the above feature extractor. It is clear that, noisy or invisible regions still have strong responses. This could be because the training procedure in Eq. (8) focuses on visible body parts, and can not tune the features on noisy parts. Because of enlarged receptive felds of neurons in feature extraction CNN layer, the strong CNN activations on noisy or occluded parts would affect features of their adjacent valid parts. This may degrades the validity of the above feature extractors. Therefore, part feature extractors robust to occlusions and noises are required
Feature Extraction with Region Sensitive Map: Our RSFE introduces a Region Sensitive Map (RSM) to suppress responses of invalid regions. To tune the responses of each feature stripe, we generate the RSM $M\in\mathbb{R}^H\times W\times d$ by applying 1 × 1 convolution on each horizontal feature $\begin{array}{l}{\mathrm{stripe~}T[i]\in\mathbb{R}^{W\times D},i\:=\:1\::\:H.\:\mathrm{The~computation~of}}\\{M[i]\in\mathbb{R}^{W\times d}\text{can be denoted as},}\end{array}$
(9)
$$M[i]=\mathrm{conv}_{1\times1}^{(\mathrm{i})}(T[i]),i=1:H,$$
where conv$_{1\times1}^{(\mathrm{i})}$refers to the $i$-th convolution kernel. Different from a global $1\times1$ convolution, $H$ convolutional kernels in Eq. (9) do not share parameters to deal with complicated occlusions on different spatial locations.
To suppress noisy regions, we supervise $M$ with a global feature. The loss can be computed by applying Average Pooling on $M$ and supervise the resulted feature $F_\mathrm{RSM},i.e.$,
$$\mathcal{L}_{\mathrm{RSM}}=\mathcal{L}_{\mathrm{ID}}(F_{\mathrm{RSM}},y).$$
(10)
This loss enforces the network to depress activations on noisy regions to generate a better $F_\mathrm{RSM}$,which is then used for stripe feature extraction. Fig. 4(c) shows feature maps
after training with Eq. (10), where invalid regions are effec-
tively suppressed
From $M$,the stripe features can be extracted by RAP as,
$$\{f_i^{(k)}\}_{k=1:l}=\mathrm{RAP}(M,\mathcal{P}_i),l=|\mathcal{P}_i|,$$
(11)
where each part feature can be trained with part feature loss
$\operatorname{in}\operatorname{Eq.}(7).$
Eq. (9) applies different $1\times1$ convolution kernels on spatial locations of $T.$ This enables more specific refinements for each part feature. Besides that, different spatial locations show varied probabilities of occlusion, e.g., the foot area is more likely to be occluded. This property makes Eq. (9) more effective in depressing occlusions and noises than learning a global $1\times1$ convolution kernel
The methods above extracts features from horizontal stripes. Similar procedure can be applied to extract features from vertical stripes. Given a query person image $q$ and a detected gallery bounding boxes $b_i$ to be compared, we utilize both global and part features for person reID. The part feature distance can be computed with Eq.(1). We also utilize the global feature $F$ from detector branch for distance computation. The overall distance between $q$ and $b_i$ can be computed as:
(12)
$$\mathrm{dist}(q,b_i)=\mathrm{dist}_P(q,b_i)+\lambda\cdot\mathrm{D}(F_q,F_i),$$
where $F_q$ and $F_i$ correspond to the global features of $q$ and $b_i$ respectively. We set $\lambda$ to 1 in the following experiments.
## $5. \textbf{LSPS Dataset}$
5.1. Previous Datasets
CUHK-SYSU[24] consists of 5,694 frames captured from movie snapshots, and 12,490 frames from street snap. Both bounding boxes and identities are labeled manually The dataset provides 8,432 labeled identities, and 23,430 labeled bounding boxes. 96,143 bounding boxes are provided in total. 11,206 frames with 5,532 identities make up the training set. CUHK-SYSU does not annotate bounding boxes with partial bodies.
PRW [33] is captured with six cameras deployed at a campus, containing 11,816 frames with 932 identities in total. 5,134 frames are selected as training set with 432 identities, while the rest 6,112 frames make up the test set. A total of 34,304 bounding boxes are annotated with identity. Similar to CUHK-SYSU, bounding box locations in PRW are also manually labeled. PRW includes some misaligned bounding boxes in both queries and galleries.
## 5.2. Description to LSPS
This paper contributes a new Large-Scale dataset for Person Search in the wild (LSPS), which shows the following characteristics:
|
[
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.50341796875,
0.10797119140625,
0.62353515625,
0.10797119140625,
0.62353515625,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.521484375,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.521484375,
0.1363525390625
],
[
0.57275390625,
0.15087890625,
0.830078125,
0.15087890625,
0.830078125,
0.1654052734375,
0.57275390625,
0.1654052734375
],
[
0.861328125,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1641845703125,
0.861328125,
0.1641845703125
],
[
0.5048828125,
0.1787109375,
0.89208984375,
0.1787109375,
0.89208984375,
0.1925048828125,
0.5048828125,
0.1925048828125
],
[
0.50244140625,
0.1944580078125,
0.56787109375,
0.1944580078125,
0.56787109375,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.521484375,
0.208984375,
0.890625,
0.2083740234375,
0.890625,
0.222900390625,
0.521484375,
0.2235107421875
],
[
0.5048828125,
0.2259521484375,
0.890625,
0.2259521484375,
0.890625,
0.236083984375,
0.5048828125,
0.236083984375
],
[
0.50341796875,
0.2392578125,
0.890625,
0.2392578125,
0.890625,
0.25244140625,
0.50341796875,
0.25244140625
],
[
0.1650390625,
0.250732421875,
0.1895751953125,
0.250732421875,
0.1895751953125,
0.26318359375,
0.1650390625,
0.26318359375
],
[
0.2279052734375,
0.250732421875,
0.25,
0.250732421875,
0.25,
0.26318359375,
0.2279052734375,
0.26318359375
],
[
0.30224609375,
0.250732421875,
0.324462890625,
0.250732421875,
0.324462890625,
0.26318359375,
0.30224609375,
0.26318359375
],
[
0.36279296875,
0.25,
0.38232421875,
0.25,
0.38232421875,
0.26318359375,
0.36279296875,
0.26318359375
],
[
0.423095703125,
0.250732421875,
0.443603515625,
0.250732421875,
0.443603515625,
0.26318359375,
0.423095703125,
0.26318359375
],
[
0.50244140625,
0.255126953125,
0.89208984375,
0.255126953125,
0.89208984375,
0.26953125,
0.50244140625,
0.26953125
],
[
0.50244140625,
0.268310546875,
0.89208984375,
0.26953125,
0.89208984375,
0.2841796875,
0.50244140625,
0.28271484375
],
[
0.108642578125,
0.251953125,
0.1282958984375,
0.251953125,
0.1282958984375,
0.264404296875,
0.108642578125,
0.264404296875
],
[
0.08251953125,
0.279052734375,
0.46728515625,
0.279052734375,
0.46728515625,
0.2890625,
0.08251953125,
0.2890625
],
[
0.50341796875,
0.285400390625,
0.8935546875,
0.285400390625,
0.8935546875,
0.298583984375,
0.50341796875,
0.298583984375
],
[
0.08087158203125,
0.291748046875,
0.46728515625,
0.290283203125,
0.46728515625,
0.3037109375,
0.08087158203125,
0.304931640625
],
[
0.080078125,
0.30615234375,
0.4697265625,
0.30615234375,
0.4697265625,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.50244140625,
0.300537109375,
0.81640625,
0.300537109375,
0.81640625,
0.314453125,
0.50244140625,
0.314453125
],
[
0.07843017578125,
0.3193359375,
0.46728515625,
0.32080078125,
0.46728515625,
0.33203125,
0.07843017578125,
0.330810546875
],
[
0.52294921875,
0.31494140625,
0.89208984375,
0.31494140625,
0.89208984375,
0.328857421875,
0.52294921875,
0.328857421875
],
[
0.080078125,
0.335205078125,
0.19775390625,
0.335205078125,
0.19775390625,
0.34521484375,
0.080078125,
0.34521484375
],
[
0.50341796875,
0.330810546875,
0.89208984375,
0.330810546875,
0.89208984375,
0.3447265625,
0.50341796875,
0.3447265625
],
[
0.50244140625,
0.3447265625,
0.89208984375,
0.34521484375,
0.89208984375,
0.35986328125,
0.50244140625,
0.359130859375
],
[
0.080078125,
0.359130859375,
0.4697265625,
0.359130859375,
0.4697265625,
0.37255859375,
0.080078125,
0.37255859375
],
[
0.50341796875,
0.35986328125,
0.890625,
0.35986328125,
0.890625,
0.373779296875,
0.50341796875,
0.373779296875
],
[
0.08087158203125,
0.375,
0.4248046875,
0.375,
0.4248046875,
0.38818359375,
0.08087158203125,
0.38818359375
],
[
0.5009765625,
0.375,
0.890625,
0.375,
0.890625,
0.38818359375,
0.5009765625,
0.38818359375
],
[
0.098876953125,
0.391357421875,
0.468017578125,
0.391357421875,
0.468017578125,
0.401611328125,
0.098876953125,
0.401611328125
],
[
0.50244140625,
0.39013671875,
0.890625,
0.39013671875,
0.890625,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.080078125,
0.40478515625,
0.468017578125,
0.40478515625,
0.468017578125,
0.41845703125,
0.080078125,
0.41845703125
],
[
0.50244140625,
0.406005859375,
0.8935546875,
0.406005859375,
0.8935546875,
0.419189453125,
0.50244140625,
0.419189453125
],
[
0.08087158203125,
0.421630859375,
0.468017578125,
0.421630859375,
0.468017578125,
0.431884765625,
0.08087158203125,
0.431884765625
],
[
0.50244140625,
0.421630859375,
0.8935546875,
0.421630859375,
0.8935546875,
0.43505859375,
0.50244140625,
0.43505859375
],
[
0.080078125,
0.43505859375,
0.465576171875,
0.43505859375,
0.465576171875,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.50341796875,
0.4375,
0.59326171875,
0.4375,
0.59326171875,
0.448974609375,
0.50341796875,
0.448974609375
],
[
0.07843017578125,
0.448974609375,
0.468017578125,
0.449462890625,
0.468017578125,
0.464111328125,
0.07843017578125,
0.46337890625
],
[
0.5439453125,
0.462158203125,
0.82666015625,
0.462158203125,
0.82666015625,
0.4765625,
0.5439453125,
0.4765625
],
[
0.859375,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.47607421875,
0.859375,
0.47607421875
],
[
0.07843017578125,
0.479248046875,
0.468017578125,
0.479736328125,
0.468017578125,
0.494384765625,
0.07843017578125,
0.49365234375
],
[
0.080078125,
0.464111328125,
0.4697265625,
0.46533203125,
0.4697265625,
0.479248046875,
0.080078125,
0.477783203125
],
[
0.080078125,
0.49560546875,
0.468017578125,
0.49560546875,
0.468017578125,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.5048828125,
0.490478515625,
0.890625,
0.490478515625,
0.890625,
0.50390625,
0.5048828125,
0.50390625
],
[
0.08087158203125,
0.51025390625,
0.4697265625,
0.51025390625,
0.4697265625,
0.5234375,
0.08087158203125,
0.5234375
],
[
0.50244140625,
0.505859375,
0.88818359375,
0.505859375,
0.88818359375,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.52587890625,
0.4697265625,
0.52587890625,
0.4697265625,
0.5390625,
0.080078125,
0.5390625
],
[
0.50244140625,
0.52978515625,
0.640625,
0.5322265625,
0.640625,
0.5498046875,
0.50244140625,
0.54736328125
],
[
0.08087158203125,
0.54248046875,
0.324462890625,
0.54248046875,
0.324462890625,
0.55322265625,
0.08087158203125,
0.55322265625
],
[
0.096435546875,
0.5546875,
0.4697265625,
0.55615234375,
0.4697265625,
0.57080078125,
0.096435546875,
0.5693359375
],
[
0.5009765625,
0.556640625,
0.67333984375,
0.55810546875,
0.67333984375,
0.57275390625,
0.5009765625,
0.5712890625
],
[
0.080078125,
0.5712890625,
0.46728515625,
0.5712890625,
0.46728515625,
0.58544921875,
0.080078125,
0.58544921875
],
[
0.080078125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.599609375,
0.080078125,
0.599609375
],
[
0.521484375,
0.580078125,
0.89208984375,
0.58056640625,
0.89208984375,
0.59521484375,
0.521484375,
0.5947265625
],
[
0.0767822265625,
0.599609375,
0.47216796875,
0.5966796875,
0.47216796875,
0.6142578125,
0.0767822265625,
0.6171875
],
[
0.50341796875,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.611328125,
0.50341796875,
0.611328125
],
[
0.080078125,
0.6162109375,
0.468017578125,
0.6162109375,
0.468017578125,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.50244140625,
0.6123046875,
0.890625,
0.6123046875,
0.890625,
0.626953125,
0.50244140625,
0.626953125
],
[
0.080078125,
0.6298828125,
0.471435546875,
0.6298828125,
0.471435546875,
0.64453125,
0.080078125,
0.64453125
],
[
0.50341796875,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.07843017578125,
0.642578125,
0.303955078125,
0.64306640625,
0.303955078125,
0.662109375,
0.07843017578125,
0.6611328125
],
[
0.50244140625,
0.6416015625,
0.8896484375,
0.642578125,
0.8896484375,
0.6572265625,
0.50244140625,
0.65576171875
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.16259765625,
0.67529296875,
0.386474609375,
0.67529296875,
0.386474609375,
0.68994140625,
0.16259765625,
0.68994140625
],
[
0.50244140625,
0.671875,
0.89208984375,
0.67236328125,
0.89208984375,
0.6875,
0.50244140625,
0.68603515625
],
[
0.25244140625,
0.68115234375,
0.261474609375,
0.68115234375,
0.261474609375,
0.68505859375,
0.25244140625,
0.68505859375
],
[
0.44677734375,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.68994140625,
0.44677734375,
0.68994140625
],
[
0.50341796875,
0.68603515625,
0.67626953125,
0.68603515625,
0.67626953125,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.52294921875,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.52294921875,
0.71630859375
],
[
0.08087158203125,
0.708984375,
0.46728515625,
0.708984375,
0.46728515625,
0.72265625,
0.08087158203125,
0.72265625
],
[
0.50244140625,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.080078125,
0.724609375,
0.468017578125,
0.724609375,
0.468017578125,
0.73779296875,
0.080078125,
0.73779296875
],
[
0.50341796875,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.50341796875,
0.7470703125
],
[
0.080078125,
0.7392578125,
0.4697265625,
0.7392578125,
0.4697265625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50341796875,
0.74853515625,
0.89208984375,
0.74853515625,
0.89208984375,
0.76220703125,
0.50341796875,
0.76220703125
],
[
0.080078125,
0.7548828125,
0.346435546875,
0.7548828125,
0.346435546875,
0.76806640625,
0.080078125,
0.76806640625
],
[
0.5009765625,
0.76220703125,
0.890625,
0.76318359375,
0.890625,
0.77783203125,
0.5009765625,
0.7763671875
],
[
0.098876953125,
0.76953125,
0.468017578125,
0.76953125,
0.468017578125,
0.78271484375,
0.098876953125,
0.78271484375
],
[
0.50341796875,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.79248046875,
0.50341796875,
0.79248046875
],
[
0.08087158203125,
0.78515625,
0.4697265625,
0.78515625,
0.4697265625,
0.798828125,
0.08087158203125,
0.798828125
],
[
0.50341796875,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50341796875,
0.806640625
],
[
0.080078125,
0.80029296875,
0.4697265625,
0.80029296875,
0.4697265625,
0.814453125,
0.080078125,
0.814453125
],
[
0.50244140625,
0.8095703125,
0.80224609375,
0.8095703125,
0.80224609375,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.191162109375,
0.82568359375,
0.35693359375,
0.8271484375,
0.35693359375,
0.8447265625,
0.191162109375,
0.84326171875
],
[
0.43701171875,
0.8271484375,
0.471435546875,
0.8271484375,
0.471435546875,
0.84228515625,
0.43701171875,
0.84228515625
],
[
0.50341796875,
0.83203125,
0.69287109375,
0.83203125,
0.69287109375,
0.8466796875,
0.50341796875,
0.8466796875
],
[
0.08087158203125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.08087158203125,
0.869140625
],
[
0.521484375,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.521484375,
0.869140625
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.63232421875,
0.88720703125,
0.63232421875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"after training with Eq. (10), where invalid regions are effec-",
"tively suppressed.",
"From M, the stripe features can be extracted by RAP as,",
"(f()1k=1il = RAP(M,P),l = Pil,",
"(11)",
"where each part feature can be trained with part feature loss",
"in Eq. (7)",
"Eq. (9) applies different 1 ×1 convolution kernels on spa-",
"tial locations ofT.This enables more specificrefinements",
"for each part feature. Besides that, different spatial loca-",
"(b)",
"(c)",
"(a)",
"(b)",
"(c)",
"tions show varied probabilities of occlusion, e.g., the foot",
"area is more likely to be occluded. This property makes",
"(a)",
"Figure4.Illustration of detectedboxes(red)andrefinedboxes",
"Eq. (9) more effective in depressing occlusions and noises",
"(green), as well as their feature maps. (b) and (c) show feature",
"maps of refined boxes trained with vanilla part feature extractor",
"than learning a global 1 × 1 convolution kernel.",
"andourRSFE.Itisclearthat,RSFEdepressesnoisesonnoisy",
"The methods above extracts features from horizontal",
"and invalid regions.",
"stripes. Similar procedure can be applied to extract features",
"from vertical stripes. Given a query person image q and a",
"where CrD refers to reID loss which is implemented with",
"detected gallery bounding boxes b, to be compared, we uti-",
"OIM loss [24]. y is the ground truth person ID label.",
"lize both global and part features for person reID. The part",
"Fig.4(b)showsfeaturemapsofalignedboundingboxes",
"feature distance can be computed with Eq. (1). We also uti-",
"trained by the above feature extractor. It is clear that, noisy",
"lize the global feature F from detector branch for distance",
"orinvisibleregions stillhave strongresponses.This could",
"computation. The overall distance between q and b, can be",
"be because the training procedure in Eq.(8) focuses on vis-",
"computed as:",
"ible body parts, and can not tune thefeatures on noisy parts.",
"dist(q, bi) = distp(g, b) + >. D(Fa, F),",
"(12)",
"extraction CNN layer, the strong CNN activations on noisy",
"Because of enlarged receptive fields of neurons in feature",
"or occluded parts would affect features of their adjacent",
"where F. and F, correspond to the global features of g and",
"valid parts.This may degrades the validity of the above",
"b, respectively. We set A to 1 in the following experiments.",
"feature extractors. Therefore, part feature extractors robust",
"5. LSPS Dataset",
"to occlusions and noises are required.",
"Feature Extraction with Region Sensitive Map: Our",
"5.1.Previous Datasets",
"RSFE introduces a Region Sensitive Map (RSM) to sup-",
"press responses of invalid regions. To tune the responses of",
"CUHK-SYSU [24] consists of 5,694 frames captured",
"each feature stripe, we generate the RSM M e RHxWxd",
"from movie snapshots, and 12,490 frames from street snap.",
"by applying 1 × 1 convolution on each horizontal feature",
"Both bounding boxes and identities are labeled manually.",
"stripe T[il e RWxD,i = 1 : H. The computation of",
"The dataset provides 8,432 labeled identities, and 23,430",
"M[ij e R iw xd can be denoted as,",
"labeled bounding boxes.96,143 bounding boxes are pro-",
"vided in total. 11,206 frames with 5,532 identities make up",
"M[il = conv@,(Ti),i= 1 : H.",
"the training set. CUHK-SYSU does not annotate bounding",
"",
"(9)",
"boxes with partial bodies.",
"PRW [33] is captured with six cameras deployed at a",
"where convx refers to the i-th convolution kernel. Differ-",
"campus, containing 11,816 frames with 932 identities in to-",
"ent from a global 1 x 1 convolution, H convolutional kernels",
"tal. 5,134 frames are selected as training set with 432 iden-",
"in Eq.(9)do not share parameters to deal with complicated",
"tities, while the rest 6,112 frames make up the test set. A",
"occlusions on different spatial locations.",
"total of 34,304 bounding boxes are annotated with identity.",
"To suppress noisy regions, we supervise M with a global",
"Similar to CUHK-SYSU, bounding box locations in PRW",
"feature.The loss can be computed by applying Average",
"are also manually labeled.PRW includes some misaligned",
"Pooling on M and supervise the resulted feature FrsM, i.e.,",
"bounding boxes in both queries and galleries.",
"CRSM = CID(FRSM, y).",
"(10)",
"5.2. Description to LSPS",
"This loss enforces the network to depress activations on",
"This paper contributes a new Large-Scale dataset for",
"noisy regions to generate a better FrsM, which is then used",
"Person Search in the wild (LSPS), which shows the follow-",
"for stripe feature extraction. Fig. 4 (c) shows feature maps",
"ing characteristics:"
] |
{
"arxiv_id": "2004.09329",
"language": "en",
"page_idx": 4
}
|
||
2308.10445-en-1
|
2308.10445-en-1.jpg
|
lem is effectively alleviated since the backbone is fixed; (2) learning prompts instead of backbone significantly reduces training costs and improves learning efficiency; (3) promptbased methods are free from keeping exemplars. To employ prompts for task-agnostic class-incremental learning, a crucial step is to select task-specific prompts given any input images. Existing methods maintain a prompt pool and retrieve prompts by directly computing the similarity between the image feature extracted by the pretrained model and the prompts in the pool, which is simple yet effective with a strong pretrained model. However, as the pretrained model dominates the retrieval, such a non-learnable retrieval process will be problematic because the future tasks are unknown and the gap between the pretraining task and the unknown future tasks could be large. As in Fig. 1(c), when the first task in incremental learning is used for pretraining, the classes in the pretraining task are totally different from other tasks, which we refer to as a semantic gap that degenerates existing models. Although the semantic gap between domains is also studied in some works such as transfer learning [14,44,53], their works do not consider the forgetting problem in sequential tasks. It is necessary to emphasize that the intention of this work is NOT to refuse pretraining but to propose a more general method that does not rely heavily on strong pretraining and can benefit from it if task-related pretraining is available. For more experiments regarding the necessity of our work, please refer to the supplementary materials.
In this work, we develop a learnable Adaptive Prompt Generator ( APG) to effectively bridge the potential gap between pretraining tasks and unknown future tasks. The core of our method is to unify the prompt retrieval and the prompt learning process into a learnable prompt generator. In this way, the whole prompting process can be optimized to reduce the negative effects of the gap between tasks effectively. Besides, rather than retrieving prompts from a fixed-size prompt pool, learning to generate prompts enhances the expression ability of prompts. As a result, the APG can be applied to a model without strong pretraining, and notably, the employment of APG does not discount the effort on overcoming forgetting since the backbone is still fixed.
For incremental learning, APG holds an extendable prompt candidate list for aggregating knowledge from seen tasks into a group of prompts. To adaptively prompt the backbone, the knowledge aggregation in APG is conditioned on the immediate feature from the backbone. In addition, we form a knowledge pool to summarize the knowledge encoded in the feature space. The summarized knowledge is further used to regularize the AGP to prevent it from learning ineffective knowledge.
In summary, our contributions are as follows. (1) We propose a learnable adaptive prompt generator (APG) to re-
duce the negative effects of the gap between the pretraining task and unknown future tasks, which is critical but ignored by previous work. Our adaptive prompting eases the reliance on intensive pretraining.(2) To regularize APG, we propose the knowledge pool, which retains the knowledge effectively with only the statistics of each class. (3) The extensive experiments show that our method significantly outperforms advanced exemplar-free incremental learning methods without pretraining. Besides, under strong pretraining, our method also achieves comparable satisfactory performance to existing prompt-based models.
## $2. \textbf{Related Work}$
Non-pretrained Class-incremental Learning. Rehearsalbased class incremental learning methods [2,3,8,18,19,22, 38,54,55] have access to a handful of old samples. With the retained samples, several works propose to distill old knowledge to the current network [2,8,18,19,54] or to maintain the old feature space [22,55]. Different from rehearsalbased methods, we propose to get rid of the memory buffer and bridge the gap between old and new tasks with a learnable module. A straightforward idea to handle incremental learning is to dynamically expand the network for each task[1,23,39,40,46,47,50,58,63]. Although these methods are more intuitive, their methods often require careful design of network architecture and often require a memory bank like rehearsal-based methods. There also exists exemplar-free methods aiming to learn sequentially tasks without saving any images. These works propose to estimate the semantic drift [64], or to keep knowledge in class prototypes [66,67]. Although exemplar-free methods proposed an appealing prospect for incremental learning, they typically introduce manually designed priors to the learning of new tasks, and human-designed priors are less generalizable, which makes their results unsatisfactory.
Prompt-based Class-incremental Learning. Inspired by
prompting in natural language processing, some methods investigate prompt-based incremental learning and achieve great success [59-61]. These methods are generally inherited directly from NLP, utilizing a pretrained transformer as a base. S-prompt[59] focuses on domain-incremental learning and learns different prompts across domains. For class incremental learning, L2P[61] and DualPrompt[60] first propose to learn a pool of prompts and query the prompts based on the feature extracted by the pretrained backbone. DualPrompt[60] further proposes to attach complementary prompts to divide old knowledge into the general one and the expert one. The main idea of prompt-based methods is to encode knowledge from old tasks into sets of vectors $(i.e.$ prompts) and retrieve them to instruct the backbone when the old data is not accessible. These methods are in need of a strong pretrained backbone to assist such a
|
[
[
0.080078125,
0.091552734375,
0.46728515625,
0.091552734375,
0.46728515625,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
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
],
[
0.5009765625,
0.10797119140625,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.5009765625,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.5009765625,
0.15283203125,
0.89208984375,
0.1514892578125,
0.89208984375,
0.1666259765625,
0.5009765625,
0.1673583984375
],
[
0.080078125,
0.1685791015625,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.1817626953125,
0.8896484375,
0.1817626953125,
0.8896484375,
0.1956787109375,
0.5009765625,
0.1956787109375
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.2108154296875,
0.890625,
0.21337890625,
0.890625,
0.2279052734375,
0.5009765625,
0.2259521484375
],
[
0.080078125,
0.2279052734375,
0.4697265625,
0.2279052734375,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.24365234375,
0.8095703125,
0.242431640625,
0.8095703125,
0.2568359375,
0.5009765625,
0.25830078125
],
[
0.08087158203125,
0.260009765625,
0.46728515625,
0.260009765625,
0.46728515625,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.279052734375,
0.640625,
0.279052734375,
0.640625,
0.29345703125,
0.50341796875,
0.29345703125
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.080078125,
0.302978515625,
0.468017578125,
0.302978515625,
0.468017578125,
0.31640625,
0.080078125,
0.31640625
],
[
0.50244140625,
0.30615234375,
0.890625,
0.30615234375,
0.890625,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50244140625,
0.322509765625,
0.890625,
0.322509765625,
0.890625,
0.3359375,
0.50244140625,
0.3359375
],
[
0.07843017578125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.07843017578125,
0.347900390625
],
[
0.50341796875,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.35107421875,
0.50341796875,
0.35107421875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.5009765625,
0.35107421875,
0.890625,
0.35107421875,
0.890625,
0.3642578125,
0.5009765625,
0.3642578125
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.367431640625,
0.890625,
0.367431640625,
0.890625,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50244140625,
0.3818359375,
0.890625,
0.3818359375,
0.890625,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.080078125,
0.4091796875,
0.4697265625,
0.4091796875,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.0767822265625,
0.4248046875,
0.4697265625,
0.423583984375,
0.4697265625,
0.4375,
0.0767822265625,
0.438232421875
],
[
0.5048828125,
0.4287109375,
0.8896484375,
0.4287109375,
0.8896484375,
0.439453125,
0.5048828125,
0.439453125
],
[
0.080078125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.5009765625,
0.443115234375,
0.890625,
0.443115234375,
0.890625,
0.45703125,
0.5009765625,
0.45703125
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50244140625,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.080078125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50341796875,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.48486328125,
0.50341796875,
0.48486328125
],
[
0.080078125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.488037109375,
0.8935546875,
0.488037109375,
0.8935546875,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50048828125,
0.26953125,
0.50048828125,
0.26953125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.50390625,
0.8935546875,
0.50390625,
0.8935546875,
0.517578125,
0.50244140625,
0.517578125
],
[
0.098876953125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.098876953125,
0.53076171875
],
[
0.50244140625,
0.51806640625,
0.89208984375,
0.51806640625,
0.89208984375,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.53271484375,
0.468017578125,
0.54736328125,
0.080078125,
0.546875
],
[
0.5009765625,
0.5322265625,
0.890625,
0.53271484375,
0.890625,
0.54736328125,
0.5009765625,
0.546875
],
[
0.080078125,
0.548828125,
0.4697265625,
0.548828125,
0.4697265625,
0.5625,
0.080078125,
0.5625
],
[
0.50244140625,
0.548828125,
0.890625,
0.548828125,
0.890625,
0.5625,
0.50244140625,
0.5625
],
[
0.0767822265625,
0.5625,
0.4697265625,
0.56298828125,
0.4697265625,
0.57763671875,
0.0767822265625,
0.5771484375
],
[
0.5009765625,
0.56298828125,
0.890625,
0.5625,
0.890625,
0.5771484375,
0.5009765625,
0.57763671875
],
[
0.080078125,
0.57763671875,
0.468017578125,
0.57763671875,
0.468017578125,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.50244140625,
0.5791015625,
0.890625,
0.5791015625,
0.890625,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.50244140625,
0.60791015625
],
[
0.07843017578125,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62255859375,
0.07843017578125,
0.62255859375
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.07843017578125,
0.62255859375,
0.471435546875,
0.62353515625,
0.471435546875,
0.63818359375,
0.07843017578125,
0.63720703125
],
[
0.50244140625,
0.625,
0.80810546875,
0.625,
0.80810546875,
0.63818359375,
0.50244140625,
0.63818359375
],
[
0.080078125,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.65283203125,
0.080078125,
0.65283203125
],
[
0.50341796875,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.07843017578125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.07843017578125,
0.66845703125
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.080078125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.080078125,
0.681640625
],
[
0.5009765625,
0.6748046875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.080078125,
0.6845703125,
0.4697265625,
0.6845703125,
0.4697265625,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.69970703125,
0.468017578125,
0.69970703125,
0.468017578125,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.5009765625,
0.7041015625,
0.89208984375,
0.705078125,
0.89208984375,
0.7197265625,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.71533203125,
0.11767578125,
0.71533203125,
0.11767578125,
0.7265625,
0.080078125,
0.7265625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.09722900390625,
0.73095703125,
0.468017578125,
0.73193359375,
0.468017578125,
0.7470703125,
0.09722900390625,
0.74560546875
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.74853515625,
0.4697265625,
0.74853515625,
0.4697265625,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.76220703125,
0.468017578125,
0.76220703125,
0.468017578125,
0.77587890625,
0.080078125,
0.77587890625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.07843017578125,
0.77783203125,
0.468017578125,
0.7763671875,
0.468017578125,
0.79150390625,
0.07843017578125,
0.79248046875
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79345703125,
0.468017578125,
0.79345703125,
0.468017578125,
0.806640625,
0.080078125,
0.806640625
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.080078125,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.080078125,
0.8212890625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.82373046875,
0.4697265625,
0.82373046875,
0.4697265625,
0.8369140625,
0.080078125,
0.8369140625
],
[
0.5009765625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.83837890625,
0.4697265625,
0.83837890625,
0.4697265625,
0.8515625,
0.080078125,
0.8515625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.8515625,
0.291748046875,
0.85302734375,
0.291748046875,
0.8671875,
0.07843017578125,
0.8662109375
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.09722900390625,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"lem is effectively alleviated since the backbone is fixed; (2)",
"duce the negative effects of the gap between the pretraining",
"learning prompts instead of backbone significantly reduces",
"task and unknown future tasks, which is critical but ignored",
"training costs and improves learning efficiency; (3)prompt-",
"by previous work. Our adaptive prompting eases the re-",
"based methods are free from keeping exemplars. To employ",
"liance on intensive pretraining.(2)To regularize APG, we",
"promptsfortask-agnosticclass-incrementallearning,a cru-",
"propose theknowledge pool, which retains theknowledge",
"cial step is to select task-specific prompts given any input",
"effectively with only the statistics of each class. (3) The",
"images. Existing methods maintain a prompt pool and re-",
"extensive experiments show that our method significantly",
"trieve prompts by directly computing the similarity between",
"outperforms advanced exemplar-free incremental learning",
"the image feature extracted by the pretrained model and the",
"methods without pretraining.Besides, under strong pre-",
"prompts in the pool, which is simple yet effective with a",
"training, our method also achieves comparable satisfactory",
"strong pretrained model. However, as the pretrained model",
"performance to existing prompt-based models.",
"dominates the retrieval,such a non-learnableretrieval pro-",
"cess will be problematic because the future tasks are un-",
"2.Related Work",
"known and the gap between the pretraining task and the un-",
"known future tasks could be large. As in Fig. 1 (c), when",
"Non-pretrained Class-incremental Learning. Rehearsal-",
"the first task in incremental learning is usedfor pretrain-",
"based class incremental learning methods [2,3, 8, 18, 19,22,",
"ing, the classes in the pretraining task are totally different",
"38, 54, 55J have access to a handful of old samples. With",
"from other tasks, which we refer to as a semantic gap that",
"the retained samples, several works propose to distill old",
"degenerates existing models. Although the semantic gap",
"knowledge to the current network [2,8,18, 19,54]or to main-",
"between domains isalso studied in someworks such as",
"tain the old feature space [22, 55]. Different from rehearsal-",
"transfer learning [14, 44,53], their works do not consider",
"based methods, we propose to get rid of the memory buffer",
"the forgetting problem in sequential tasks. It is necessary to",
"and bridge the gap between old and new tasks with a learn-",
"emphasizethat the intentionof thiswork is NOT to refuse",
"ablemodule.A straightforward idea to handle incremen-",
"pretraining but to propose a more general method that does",
"tal learning is to dynamically expand the network for each",
"not rely heavily on strong pretraining and can benefit from",
"task [1,23,39, 40, 46, 47, 50, 58, 63]. Although these meth-",
"it if task-related pretraining is available. For more experi-",
"ods are more intuitive. their methods often reguire careful",
"ments regarding the necessity of our work, please refer to",
"design of network architecture and often require a mem-",
"the supplementary materials.",
"ory bank like rehearsal-based methods. There also exists",
"In this work, we develop a learnable Adaptive Prompt",
"exemplar-free methods aiming to learn sequentially tasks",
"Generator (APG) to effectively bridge the potential gap be-",
"without saving any images. These works propose to esti-",
"tween pretraining tasks and unknown future tasks.\"The",
"mate the semantic drift [64], or to keep knowledge in class",
" core of our method is to unify the prompt retrieval and the",
"prototypes [66, 67]. Although exemplar-free methods pro-",
"prompt learning process into a learnable prompt generator.",
"posed an appealing prospect for incremental learning, they",
"In this way, the whole prompting process can be optimized",
"typically introduce manually designed priors to the learning",
"to reduce the negative effects of the gap between tasks ef-",
"of new tasks, and human-designed priors are less generaliz-",
"fectively. Besides, rather than retrieving prompts from a",
"able, which makes their results unsatisfactory.",
"fixed-size prompt pool, learning to generate prompts en-",
"Prompt-based Class-incremental Learning. Inspired by",
"hances the expression ability of prompts. As a result, the",
"prompting in natural language processing, some methods",
"APG can be applied to a model without strong pretraining,",
"investigate prompt-based incremental learning and achieve",
"and notably, the employment of APG does not discount the",
"great success [59-61]. These methods are generally inher-",
"effort on overcoming forgetting since the backbone is still",
"ited directly from NLP, utilizing a pretrained transformer",
"fixed.",
"as a base. S-prompt [59] focuses on domain-incremental",
"For incremental learning,APG holds an extendable",
"learning andlearns differentprompts across domains.For",
"prompt candidate list for aggregating knowledge from seen",
"class incremental learning, L2P [6]] and DualPrompt [60]",
"tasks into a group of prompts.To adaptively prompt the",
"first propose to learn a pool of prompts and query the",
"backbone, the knowledge aggregation in APG is condi-",
"prompts based on the feature extracted bythe pretrained",
"tioned on the immediate feature from the backbone. In ad-",
"backbone. DualPrompt [60] further proposes to attach com-",
"dition, we form a knowledge pool to summarize the knowl-",
"plementary prompts to divide old knowledge into the gen-",
"edge encoded in the feature space.The summarized knowl-",
"eral one and the expert one. The main idea of prompt-based",
"edge is further used to regularize the AGP to prevent it from",
"methods is to encode knowledge from old tasks into sets of",
"learningineffectiveknowledge.",
"vectors (i.e. prompts) and retrieve them to instruct the back-",
"In summary, our contributions are as follows. (1) We",
"bone when the old data is not accessible. These methods",
"propose a learnable adaptive prompt generator (APG) to re-",
"are in need of a strong pretrained backbone to assist such a"
] |
{
"arxiv_id": "2308.10445",
"language": "en",
"page_idx": 1
}
|
||
2308.10445-en-2
|
2308.10445-en-2.jpg
|
retrieval process. Different from their methods, we propose a unified prompting scheme that eases the reliance on intensive pretraining and makes it suitable for both pretrained and non-pretrained scenarios.
## $3. \textbf{ Adaptive- prompting for Incremental Learn- }$ ing
$3. 1. \textbf{Preliminary}$
Class-incremental learning. For a deep model $\Phi(f(\cdot))$ consisting of a classifer $\Phi(\cdot)$ and the backbone $f(\cdot)$, the goal of class-incremental learning is to train the model $\Phi(f(\cdot))$ on tasks $\{\mathcal{T}_t\}_t=1^n$ sequentially so that the model can classify testing samples from any task. Specifically, at the $t$-th task $\mathcal{T}_t$, the training set is $\mathcal{D}_t=\{x_m^{t},y_m^{t}\}_{m=1}^{N_{t}}$,where $x_{m}^{t}$ is the $m$-th image at task $\mathcal{T}_t$ with label $y_m^t.$ And the label space $\mathcal{Y}_t$ of task $\mathcal{T}_t$ is disjoint with other tasks, $i.e.$ $\bigcap_{t=1}^{n}\hat{\mathcal{Y}}_{t}=\emptyset.$ Once the model finishes learning on task $\mathcal{T}_t$, the corresponding training set $\mathcal{D}_t$ will be dropped and becomes inaccessible when learning from $\mathcal{T}_t+1$, which raises a risk of forgetting old tasks when learning new ones.
Vision transformers and prompting. Different from convolutional neural networks, Vision Transformers (ViTs) treat the image as a sequence of tokens. Specifically, the image $x_m^t$ will be converted into a sequence E $\in\mathbb{R}^N_E\times d$ through patch embedding layers, where $N_E$ is the total number of image patches $(i.e.$ tokens) and $d$ is the embedding dimension. An extra token called class token $e_cls\in$ $\mathbb{R}^d$ will be concatenated to E to gather information from patches. The resultant token sequence $\mathbf{X}=[e_{cls};\mathbf{E}]$ will be fed into transformer layers for feature extraction, and the core is self-attention operation:
$$\mathrm{SelfAttn}(\mathbf{X})=\mathrm{Attn}(\mathbf{XW}^{Q},\mathbf{XW}^{K},\mathbf{XW}^{V}),\\\mathrm{Attn}(Q,K,V)=\mathrm{Softmax}(\frac{QK^{T}}{\sqrt{d}})V,$$
(1)
where $\mathbf{W}^Q,\mathbf{W}^K$ and $\mathbf{W}^V$ are learnable projections.
To facilitate visual feature extraction, a common prompting [21,30,32,33,35,36] technique is to insert a set of learnable extra tokens $\mathbf{P}\in\mathbb{R}^{N_P\times d}$ into the original sequence as $[e_{cls};\mathbf{P};\mathbf{E}].$ In this work, we only discuss vanilla ViTs[7,56] due to their simplicity and versatility. We detail our method as follows, and an overview is in Fig. 2.
## 3.2. Adaptive Prompt Generation
Adaptive prompt generator. We propose an adaptive prompt generator (APG) for adaptively aggregating old and new knowledge. Assume there are $N_L$ transformer layers denoted as $\{L_l\}_{l=1}^{N_L}$,and the output of the $l$-th transformer layer is $\mathbf{X}_{l}=L_{l}(L_{l-1}(\cdots L_{1}(x)))\in\mathbb{R}^{(N_{E}+1)\times d}$,where
$x^1$ is the input image. The APG takes the intermediate feature as input, which is denoted as $v_l=\mathbf{X}_l[0,:]$, and outputs the class-specific prompts to facilitate feature extraction in deeper layers:
$$\mathbf{P}=\mathrm{APG}(v_{l})\in\mathbb{R}^{N_{P}\times d}.$$
(2)
Specifically, in the APG, we keep updating a prompt candidate list to maintain different types of knowledge. Starting from an empty list denoted as $\mathbf{I}_0=[]$, the list is extended at each task, and the list at task $\mathcal{T}_t$ is denoted as:
(3)
$$\mathbf{I}_{t}=\begin{bmatrix}\mathbf{I}_{t-1};\widehat{\mathbf{P}}_{1},\dots,\widehat{\mathbf{P}}_{c},\dots,\widehat{\mathbf{P}}_{|\mathcal{Y}_{t}|}\end{bmatrix},$$
where $\widehat{\mathbf{P}}_c\in\mathbb{R}^{N_g\times d}$ represents a group of prompts candi-
dates, and there are totally $N_g\times|\tilde{\mathcal{Y}}_t|$ prompt candidates are extended at $\mathcal{T}_t$ to form I$_t.$ In order to generate prompts with class-specific knowledge adaptively for $v_l$, we firstly apply a projection module $M_{in}$ to $v_l$, denoted as $z=M_{in}(v_l)$, and then adopt a cross-attention operation between the projection $z$ and the candidate list $\mathbf{I}_t:$
$$\widetilde{\mathbf{P}}=\mathrm{CrossAttn}(z,\mathbf{I}_{t})=\mathrm{MMHA}(z,\mathbf{I}_{t},\mathbf{I}_{t}),$$
(4)
where the MMHA denotes Multi-output Multi-head Attention, an extension of the Multi-head attention (MHA)[57]. While the original MHA outputs the same number of tokens with the input query, in our case, the query is a single token $z$ and we want to obtain $N_P$ prompts after cross-attention. To avoid restricting $N_P=1$, we directly extend the MHA and form the MMHA as follows. Firstly, with the input $z$ and $\mathbf{I}_t$, we define the $j$-th token computed from the $h$-th attention head as:
(5)
$$r_{h}^{j}=\mathrm{Attn}(z\mathbf{W}_{h,j}^{Q},\mathbf{I}_{t}\mathbf{W}_{h,j}^{K},\mathbf{I}_{t}\mathbf{W}_{h,j}^{V}).$$
Assuming the number of heads is $n_h$,we concatenate results from every head as the $j$-th token $\mathbf{R}_j$, and pass the token through a projection module as the output of MMHA:
$$\mathbf{R}_{j}=\mathrm{Concat}(r_{1}^{j},\ldots,r_{h}^{j},\ldots,r_{n_{h}}^{j}),\\\widetilde{\mathbf{P}}=\mathrm{MMHA}(z,\mathbf{I}_{t},\mathbf{I}_{t})=[\mathbf{R}_{1},\ldots,\mathbf{R}_{N_{P}}]\:\mathbf{W}^{o}.$$
(6)
We further put the $\widetilde{\mathbf{P}}$ into an output projection module $M_{out}$, and the obtained tokens $\mathbf{P}=M_{out}(\widetilde{\mathbf{P}})\in\mathbb{R}^{N_{P}\times d}$ are the adaptive prompts for current image $x.$
The generated prompts P are further integrated with the original output $\mathbf{X}_l$ of the $l$-th layer, and the token sequence for the next layer becomes $\mathbf{X}_l=[e_{cls,l};\mathbf{P};\mathbf{E}_l].$ we feed $\mathbf{X}_l$ into the remaining deeper layers $\{L_i\}_{i=l+1}^{N_L}.$ The final output of the network with APG is denoted as $\mathbf{X} _{N_{L}}$ = $[ e_{cls, N_{L}}; \mathbf{P} _{N_{L}}; \mathbf{E} _{N_{L}}] .$ We take the class token
In the later part of Sec. 3.2, we will omit the task label $t$ and the index
$m$ of image $x_m^t$ for simplicity.
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50341796875,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
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,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.12371826171875,
0.8935546875,
0.12371826171875,
0.8935546875,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.275390625,
0.1383056640625,
0.275390625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.1383056640625,
0.59619140625,
0.1383056640625,
0.59619140625,
0.149658203125,
0.50341796875,
0.149658203125
],
[
0.609375,
0.1610107421875,
0.78173828125,
0.1578369140625,
0.78173828125,
0.1787109375,
0.609375,
0.1817626953125
],
[
0.869140625,
0.1654052734375,
0.8935546875,
0.1654052734375,
0.8935546875,
0.1787109375,
0.869140625,
0.1787109375
],
[
0.08087158203125,
0.1717529296875,
0.46728515625,
0.1717529296875,
0.46728515625,
0.186279296875,
0.08087158203125,
0.186279296875
],
[
0.108642578125,
0.1875,
0.1396484375,
0.189453125,
0.1380615234375,
0.2083740234375,
0.10540771484375,
0.2052001953125
],
[
0.521484375,
0.1912841796875,
0.890625,
0.1932373046875,
0.890625,
0.20703125,
0.521484375,
0.2052001953125
],
[
0.5048828125,
0.2083740234375,
0.890625,
0.2083740234375,
0.890625,
0.2215576171875,
0.5048828125,
0.2215576171875
],
[
0.080078125,
0.21533203125,
0.205078125,
0.21533203125,
0.205078125,
0.2298583984375,
0.080078125,
0.2298583984375
],
[
0.50341796875,
0.222900390625,
0.89208984375,
0.222900390625,
0.89208984375,
0.236083984375,
0.50341796875,
0.236083984375
],
[
0.08087158203125,
0.2412109375,
0.46728515625,
0.2412109375,
0.46728515625,
0.255126953125,
0.08087158203125,
0.255126953125
],
[
0.50244140625,
0.238037109375,
0.80615234375,
0.238037109375,
0.80615234375,
0.251953125,
0.50244140625,
0.251953125
],
[
0.07843017578125,
0.255126953125,
0.46728515625,
0.255615234375,
0.46728515625,
0.270751953125,
0.07843017578125,
0.26953125
],
[
0.5810546875,
0.260009765625,
0.80615234375,
0.261962890625,
0.80615234375,
0.281005859375,
0.5810546875,
0.279052734375
],
[
0.87109375,
0.26513671875,
0.89208984375,
0.26513671875,
0.89208984375,
0.279052734375,
0.87109375,
0.279052734375
],
[
0.080078125,
0.271484375,
0.468017578125,
0.271484375,
0.468017578125,
0.285400390625,
0.080078125,
0.285400390625
],
[
0.08087158203125,
0.287353515625,
0.468017578125,
0.287353515625,
0.468017578125,
0.300537109375,
0.08087158203125,
0.300537109375
],
[
0.50244140625,
0.290283203125,
0.89208984375,
0.292236328125,
0.89208984375,
0.3125,
0.50244140625,
0.31005859375
],
[
0.08087158203125,
0.3017578125,
0.468017578125,
0.3017578125,
0.468017578125,
0.31494140625,
0.08087158203125,
0.31494140625
],
[
0.07843017578125,
0.31494140625,
0.4697265625,
0.31494140625,
0.4697265625,
0.333251953125,
0.07843017578125,
0.333251953125
],
[
0.50341796875,
0.3125,
0.89208984375,
0.3125,
0.89208984375,
0.32568359375,
0.50341796875,
0.32568359375
],
[
0.08087158203125,
0.33203125,
0.4697265625,
0.33203125,
0.4697265625,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.50341796875,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.08087158203125,
0.3466796875,
0.468017578125,
0.3466796875,
0.468017578125,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.34228515625,
0.890625,
0.34228515625,
0.890625,
0.35595703125,
0.50341796875,
0.35595703125
],
[
0.50244140625,
0.35595703125,
0.89208984375,
0.356689453125,
0.89208984375,
0.371826171875,
0.50244140625,
0.37060546875
],
[
0.080078125,
0.3623046875,
0.468017578125,
0.361083984375,
0.468017578125,
0.375732421875,
0.080078125,
0.376953125
],
[
0.07843017578125,
0.376953125,
0.46728515625,
0.375732421875,
0.46728515625,
0.39013671875,
0.07843017578125,
0.391357421875
],
[
0.50244140625,
0.37255859375,
0.890625,
0.37255859375,
0.890625,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.080078125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.5009765625,
0.386962890625,
0.72607421875,
0.386474609375,
0.72607421875,
0.40087890625,
0.5009765625,
0.401611328125
],
[
0.080078125,
0.40771484375,
0.431396484375,
0.40771484375,
0.431396484375,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.55224609375,
0.410400390625,
0.84326171875,
0.412353515625,
0.84326171875,
0.429931640625,
0.55224609375,
0.427978515625
],
[
0.869140625,
0.415283203125,
0.89208984375,
0.415283203125,
0.89208984375,
0.427978515625,
0.869140625,
0.427978515625
],
[
0.08087158203125,
0.427978515625,
0.34716796875,
0.427978515625,
0.34716796875,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.351318359375,
0.43310546875,
0.363525390625,
0.43310546875,
0.363525390625,
0.4375,
0.351318359375,
0.4375
],
[
0.359375,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.359375,
0.44140625
],
[
0.080078125,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.45703125,
0.080078125,
0.45703125
],
[
0.50341796875,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50341796875,
0.455078125
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.50341796875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.0767822265625,
0.4716796875,
0.4697265625,
0.468505859375,
0.4697265625,
0.486083984375,
0.0767822265625,
0.4892578125
],
[
0.50341796875,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.07843017578125,
0.488037109375,
0.4697265625,
0.488037109375,
0.4697265625,
0.501953125,
0.07843017578125,
0.501953125
],
[
0.50244140625,
0.488037109375,
0.8935546875,
0.488037109375,
0.8935546875,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50390625,
0.46728515625,
0.50390625,
0.46728515625,
0.517578125,
0.080078125,
0.517578125
],
[
0.50341796875,
0.50244140625,
0.88818359375,
0.50244140625,
0.88818359375,
0.5166015625,
0.50341796875,
0.5166015625
],
[
0.08087158203125,
0.51806640625,
0.468017578125,
0.51806640625,
0.468017578125,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.50244140625,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.07843017578125,
0.5322265625,
0.468017578125,
0.53271484375,
0.468017578125,
0.54736328125,
0.07843017578125,
0.546875
],
[
0.50341796875,
0.5341796875,
0.8896484375,
0.5341796875,
0.8896484375,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.07843017578125,
0.548828125,
0.468017578125,
0.54736328125,
0.468017578125,
0.5625,
0.07843017578125,
0.56298828125
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.564453125,
0.4697265625,
0.564453125,
0.4697265625,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.50244140625,
0.56298828125,
0.62353515625,
0.56298828125,
0.62353515625,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.080078125,
0.580078125,
0.28759765625,
0.580078125,
0.28759765625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.57421875,
0.58837890625,
0.8203125,
0.58837890625,
0.8203125,
0.60595703125,
0.57421875,
0.60595703125
],
[
0.6787109375,
0.59228515625,
0.69775390625,
0.59228515625,
0.69775390625,
0.599609375,
0.6787109375,
0.599609375
],
[
0.87109375,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.60400390625,
0.87109375,
0.60400390625
],
[
0.125,
0.60791015625,
0.416748046875,
0.60595703125,
0.416748046875,
0.62353515625,
0.125,
0.62548828125
],
[
0.50341796875,
0.61865234375,
0.890625,
0.61865234375,
0.890625,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.330078125,
0.62890625,
0.371826171875,
0.62890625,
0.371826171875,
0.64453125,
0.330078125,
0.64453125
],
[
0.1519775390625,
0.63818359375,
0.330810546875,
0.63818359375,
0.330810546875,
0.65283203125,
0.1519775390625,
0.65283203125
],
[
0.50341796875,
0.63330078125,
0.89208984375,
0.63330078125,
0.89208984375,
0.646484375,
0.50341796875,
0.646484375
],
[
0.371826171875,
0.6396484375,
0.39453125,
0.6396484375,
0.39453125,
0.6513671875,
0.371826171875,
0.6513671875
],
[
0.44677734375,
0.6396484375,
0.4697265625,
0.6396484375,
0.4697265625,
0.65283203125,
0.44677734375,
0.65283203125
],
[
0.343017578125,
0.6484375,
0.36767578125,
0.6484375,
0.36767578125,
0.66015625,
0.343017578125,
0.66015625
],
[
0.50244140625,
0.6484375,
0.85546875,
0.6484375,
0.85546875,
0.662109375,
0.50244140625,
0.662109375
],
[
0.07843017578125,
0.6748046875,
0.42724609375,
0.67529296875,
0.42724609375,
0.693359375,
0.07843017578125,
0.6923828125
],
[
0.615234375,
0.673828125,
0.8544921875,
0.67529296875,
0.8544921875,
0.693359375,
0.615234375,
0.69140625
],
[
0.5400390625,
0.6943359375,
0.85107421875,
0.69580078125,
0.85107421875,
0.71337890625,
0.5400390625,
0.7119140625
],
[
0.87158203125,
0.68994140625,
0.88818359375,
0.68994140625,
0.88818359375,
0.69873046875,
0.87158203125,
0.69873046875
],
[
0.080078125,
0.71044921875,
0.4697265625,
0.71044921875,
0.4697265625,
0.7236328125,
0.080078125,
0.7236328125
],
[
0.1004638671875,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.1004638671875,
0.708984375
],
[
0.080078125,
0.724609375,
0.468017578125,
0.724609375,
0.468017578125,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.50341796875,
0.72802734375,
0.890625,
0.72802734375,
0.890625,
0.7421875,
0.50341796875,
0.7421875
],
[
0.080078125,
0.74072265625,
0.468017578125,
0.7392578125,
0.468017578125,
0.75390625,
0.080078125,
0.7548828125
],
[
0.5009765625,
0.7421875,
0.89208984375,
0.7392578125,
0.89208984375,
0.75830078125,
0.5009765625,
0.76123046875
],
[
0.08087158203125,
0.755859375,
0.468017578125,
0.755859375,
0.468017578125,
0.76953125,
0.08087158203125,
0.76953125
],
[
0.50341796875,
0.7587890625,
0.79833984375,
0.7587890625,
0.79833984375,
0.77294921875,
0.50341796875,
0.77294921875
],
[
0.080078125,
0.771484375,
0.419189453125,
0.771484375,
0.419189453125,
0.78515625,
0.080078125,
0.78515625
],
[
0.521484375,
0.7734375,
0.89208984375,
0.77294921875,
0.89208984375,
0.787109375,
0.521484375,
0.78857421875
],
[
0.50341796875,
0.7890625,
0.890625,
0.7890625,
0.890625,
0.80322265625,
0.50341796875,
0.80322265625
],
[
0.07843017578125,
0.7998046875,
0.338134765625,
0.80029296875,
0.338134765625,
0.81494140625,
0.07843017578125,
0.814453125
],
[
0.50244140625,
0.8046875,
0.890625,
0.8046875,
0.890625,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.08087158203125,
0.82568359375,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50341796875,
0.82080078125,
0.865234375,
0.82080078125,
0.865234375,
0.83544921875,
0.50341796875,
0.83544921875
],
[
0.86279296875,
0.8291015625,
0.8798828125,
0.8291015625,
0.8798828125,
0.83349609375,
0.86279296875,
0.83349609375
],
[
0.08087158203125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.8359375,
0.89208984375,
0.8359375,
0.89208984375,
0.849609375,
0.50341796875,
0.849609375
],
[
0.50244140625,
0.8515625,
0.89208984375,
0.8515625,
0.89208984375,
0.8662109375,
0.50244140625,
0.8662109375
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.2034912109375,
0.8720703125,
0.2034912109375,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.1993408203125,
0.8720703125,
0.471435546875,
0.869140625,
0.471435546875,
0.8837890625,
0.1993408203125,
0.88623046875
],
[
0.51708984375,
0.87353515625,
0.89208984375,
0.875,
0.89208984375,
0.8896484375,
0.51708984375,
0.88818359375
],
[
0.07598876953125,
0.8837890625,
0.47216796875,
0.880859375,
0.47216796875,
0.9013671875,
0.07598876953125,
0.90478515625
],
[
0.5009765625,
0.88623046875,
0.66650390625,
0.88720703125,
0.66650390625,
0.9013671875,
0.5009765625,
0.90087890625
]
] |
[
"retrieval process.Different from theirmethods,we propose",
"r' is the input image. The APG takes the intermediate fea-",
"a unified prompting scheme that eases the reliance on in-",
"ture as input, which is denoted as ui = X,[0, :, and outputs",
"tensive pretraining and makes it suitable for both pretrained",
"the class-specific prompts to facilitate feature extraction in",
"and non-pretrained scenarios.",
"deeper layers:",
"P = APG(u) E IRNp×d.",
"(2)",
"3.Adaptive-prompting for Incremental Learn-",
"ing",
"Specifically,in theAPG,wekeepupdatinga prompt can-",
"didate list to maintain different types of knowledge. Starting",
"3.1. Preliminary",
"from an empty list denoted as Io = l, the list is extended at",
"Class-incremental learning. For a deep model (f()",
"each task, and the list at task Tt is denoted as:",
"consisting of a classifier () and the backbone f(), the",
"I= [I;Pi.., Pe..,Py]",
"(3)",
"goal of class-incremental learning is to train the model",
"@(f() on tasks [Tt )n-, sequentially so that the model can",
"where P。 E RNgxd represents a group of prompts candi-",
"classify testing samples from any task. Specifically, at the",
"t-th task Tt, the training set is Dt = (am, ym)m-1, where",
"dates, and there are totally Ng × t| prompt candidates are",
"ctm is the m-th image at task Tt with label ytm. And the",
"extended at Tt to form It. In order to generate prompts with",
"label space Yt of task Tt is disjoint with other tasks, i.e.",
"class-specific knowledge adaptively for ui, we firstly apply",
"a projection module Min to vi, denoted as z = Min(vi),",
"Nt=1 Jt = 0. Once the model finishes learning on task Tt,",
"the corresponding training set Dt will be dropped and be-",
"and then adopt a cross-attention operation between the pro-",
"comes inaccessible when learning from Tt+1, which raises",
"jection z and the candidate list It:",
"a risk of forgetting old tasks when learning new ones.",
"P = CrossAttn(z, It) = MMHA(z, It, It),",
"(4)",
"Vision transformers and prompting.",
"",
"Different from",
"convolutional neural networks, Vision Transformers (ViTs)",
"where the MMHA denotes Multi-output Multi-head Atten-",
"treat the image as a sequence of tokens. Specifically, the",
"tion, an extension of theMulti-head attention (MHA)[57].",
"image atm will be converted into a sequence E e IRNe xd",
"While the original MHA outputs the same number of tokens",
"through patch embedding layers, where Ne is the total",
"with the input query, in our case, the query is a single token",
"number of image patches (i.e. tokens) and d is the embed-",
"z and we want to obtain Np prompts after cross-attention",
"ding dimension. An extra token called class token ecls E",
"To avoid restricting Np = 1, we directly extend the MHA",
"IRd will be concatenated to E to gather information from",
"and form theMMHA asfollows.Firstly,with theinput z",
"patches. The resultant token sequence X = [ecls; E] will",
"and It, we define the j-th token computed from the h-th",
"be fed into transformer layers for feature extraction, and the",
"attention head as:",
"core is self-attention operation:",
"rh = Attn(2wh., I wk,I wh.,)",
"",
"(5)",
"SelfAttn(X) = Att(Xw,XwK,Xw))",
"Assuming the number of heads is nh, we concatenate results",
"QKT",
"Attn(Q, K,V) = Softmax(",
"from every head as the j-th token R,, and pass the token",
")V",
"(1)",
"Va",
"through a projection module as the output of MMHA:",
"where WQ, WK and WV are learnable projections.",
"R, = Concat(ri,....,h,..., rh.),",
"P = MMHA(z,It,It) = [Ri,..., RNp]Wo.",
"(6",
"ing [21, 30, 32, 33, 35, 36] technique is to insert a set of",
"To facilitate visual feature extraction, a common prompt-",
"learnable extra tokens P e RNpxd into the original se-",
"We further put the P into an output projection module",
"quence as [ecls; P; E]. In this work, we only discuss vanilla",
"Mout, and the obtained tokens P = Mout(P) e RNpxd",
"ViTs [7, 56] due to their simplicity and versatility. We detail",
"are the adaptive prompts for current image r.",
"our method as follows, and an overview is in Fig. 2.",
"The generated prompts P are further integrated with",
"the original output Xy of the l-th layer, and the token se-",
"3.2. Adaptive Prompt Generation",
"quence for the next layer becomes X, = [ecls,i; P;Ei].",
"Adaptive prompt generator.We propose an adaptive",
"we feed X, into the remaining deeper layers [L,]",
"71",
"prompt generator (APG) for adaptively aggregating old and",
"The final output of the network with APG is denoted as",
"XN = [ecls,Ni;PNi;ENz]. We take the class token",
"newknowledge.Assume there are N transformer layers",
"denoted as {L)-",
", and the output of the l-th transformer",
"'In thelater partof Sec.3.2,we willomit thetask labelt and the index",
"layer is X, = Li(Li-1(... Li(a)) e IR(Ne+1)xd, where",
"m of image rtm, for simplicity."
] |
{
"arxiv_id": "2308.10445",
"language": "en",
"page_idx": 2
}
|
||
2308.10445-en-4
|
2308.10445-en-4.jpg
|
(10)
$$\mathcal{L}_{tri}=[d_p-d_n+\alpha]_+,$$
where $[\cdot]_+$ denotes the hinge loss, and $\alpha$ is the margin.
Discussion. The proposed APG has the following advantages. First, APG continuously learns to generate diverse prompts conditioned on the input of the network. Hence, APG breaks the limitation of selecting prompts from a fixed-size prompt pool and alleviates the dependence on the pretrained model that is used for querying discrete prompts ${\hat{\text{from a pool. Second, the prompt candidate list is extendable}}}$ to handle growing knowledge. This enables the model capable of long incremental learning (see Sec. 4.2 and Sec. 4.3) as the APG can aggregate knowledge from different tasks by the cross-attention operation (Eq. 5).
$3. 3. \textbf{Anti- forgetting Learning: A Knowledge Pool}$
Construction of the knowledge pool. Any trainable module faces the forgetting $(i.e.$ old knowledge degradation) problem when learning from sequential tasks, including the proposed APG and the classifier $\Phi(\cdot).$ Hence, we develop a knowledge pool to better maintain old knowledge.
Once finishing learning at task $\mathcal{T}_t$, we summarize the knowledge in the training set $\mathcal{D}_t$ into the knowledge pool before dropping the dataset. Specifically, denote all the images in class $c$ as $\mathcal{D}_t^c=\{x_m^t,y_m^t|x_m^t\in\mathcal{D}_t,y_m^t=c\}.$ We feed these images into the backbone and get a set of features $V_{c}^{l}=\{v_{m}^{l}|x_{m}^{t}\in\mathcal{D}_{t}^{c}\}$ from the $l$-th transformer layer. Then the class centroid $\mu_c^l$ is then calculated by averaging features and the and $\Sigma_c^l$ is a matrix where each element $\Sigma_c^{l,(i,k)}$ is the covariance between the $i$-th feature and the $k$-th feature. We use the class-specific statistics to form a multivariate normal distribution $\hat{\mathcal{N}}_c^l=\mathcal{N}(\mu_c^l,\Sigma_c^l)$ for each class. Besides, to better constrain the APG, the feature centroid $\mu_c^l$ is used to extract corresponding prompts by feeding the centroid into the APG, $i.e.{\overset{.}{\operatorname*{\overline{P}}}c}={\overset{.}{\operatorname*{APG}}(\mu_{c})}.$
Finally, we form the knowledge pool for class $c$ as $\mathcal{S}_c=$ $(\mathcal{N}_{c}^{l},\mathcal{N}_{c}^{N_{L}},\overline{\mathbf{P}^{c}})$,where $\mathcal{N}_{c}^{N_{L}}$ is computed from the features from the last layer. Generally, we store the statistics for each class $\left\{\mathcal{S}_{c}\right\}_{c=1}^{N_{t}^{cls}}$,where $N_{t}^{cls}=\sum_{t}\left|\mathcal{Y}_{t}\right|$ is the number of seen classes so far.
Anti-forgetting learning. For exemplar-free incremental learning, we can only access the current training set $\mathcal{D}_t$ during task $\mathcal{T}_{t}.$ The knowledge pool $\left\{\mathcal{S}_{c}\right\}_{c=1}^{N_{t-1}^{cls}}$ described above is used to maintain old knowledge. To alleviate forgetting in the APG, we sample from $\mathcal{N}_c^l$, and get intermediate knowledge vector $\widetilde{v}^{c}\sim\mathcal{N}_{\underline{c}.}^{l}$With the vector and the corresponding prompt centroid $\overline{\mathbf{P}^j}$, we apply a constraint on APG:
$$\mathcal{L}_{ConA}=\phi(\mathrm{APG}(\widetilde{v}^{c}),\overline{{\mathbf{P}^{c}}}),$$
(11)
where $\phi$ is a distance metric and we find that the $L1$ loss is useful. With the constraint, APG can be reminded of old knowledge and perform aggregation on them.
The classifier $\Phi(\cdot)$ is extended after each task, and needs to be updated during every task. Thus forgetting also happens at the classifier. Similar to the constraint on the APG, we sample the knowledge vector from $\widetilde{v}^{c}\sim\mathcal{N}_{c}^{N_{L}}.$ Then a simple but effective cross-entropy loss is adopted:
(12)
$$\mathcal{L}_{ConC}=-\log\Phi(\widetilde{v}_{m}^{c})^{(c)},$$
where $\Phi(\widetilde{v}^{c})^{(c)}$ is the $c$-th element of $\Phi(\widetilde{v}^{c}).$
Training objective for the whole model. The objective function for optimizing the model and APG is summarized as follows:
$$\mathcal{L}=\mathcal{L}_{attn}+\mathcal{L}_{tri}+\mathcal{L}_{cls}+\mathcal{L}_{conA}+\mathcal{L}_{conC}.$$
(13)
For the backbone $f(\cdot)$,it is only trained on the first task and then is fxed in the case of without pretraining, and otherwise it is fixed all the time. The classifier $\Phi(\cdot)$ and APG are optimized across all tasks.
## $4. \textbf{Experiments}$
$4. 1. \textbf{Experiment Setup}$
Datasets. We conduct experiments on three datasets: CIFAR100[25], ImageNet-Subset[6], and ImageNet-R[17]. CIFAR100[25] contains 100 classes. There are 50,000 images for training and 10,000 images for evaluation. ImageNet[6] is a large-scale dataset containing 1,000 classes, 1300 images per class. We follow prior works [8,18,19,26, 54,58,62,63,65] and use ImageNet-Subset which contains 100 classes. ImageNet-R [17] dataset contains newly collected data that are different styles of the original ImageNet classes. We follow prior work[60] to split the dataset with $80\%$ samples for training and the rest 20% samples are used for evaluation.
Training and testing protocols. For the training and test-
ing protocols, we follow previous works [8,18,19,54,58, 60,61,63] to conduct class-incremental learning. After every training stage, we evaluate the model by testing on the union of all testing sets of seen tasks. The classification accuracies in every task are averaged, and we report the $aver-$ age accuracy in all experiments [8,19,48,54,60-63]. When comparing with L2P[61] and DualPrompt[60], we also report the forgetting metric following their papers.
We do not save any images as a memory buffer and only keep the class statics of old classes on the feature level, following the exemplar-free class-incremental learning proposed in PASS[66] and SSRE[67].
Implementation details. For all experiments below, we use the vanilla ViT as our backbone. For the standard nonpretrained setting, we adjust the embedding dimension and number of heads to match the number of parameters of the widely used ResNet-18[15]. For more information about the adjusted backbone, please refer to the appendix. We
|
[
[
0.521484375,
0.09027099609375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.521484375,
0.10479736328125
],
[
0.1968994140625,
0.106689453125,
0.3544921875,
0.106689453125,
0.3544921875,
0.121826171875,
0.1968994140625,
0.121826171875
],
[
0.438720703125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.12060546875,
0.438720703125,
0.12060546875
],
[
0.50244140625,
0.106689453125,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.12060546875
],
[
0.08087158203125,
0.126953125,
0.435546875,
0.126953125,
0.435546875,
0.14013671875,
0.08087158203125,
0.14013671875
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.5009765625,
0.1363525390625,
0.89208984375,
0.1363525390625,
0.89208984375,
0.15087890625,
0.5009765625,
0.15087890625
],
[
0.08251953125,
0.1458740234375,
0.468017578125,
0.1458740234375,
0.468017578125,
0.160400390625,
0.08251953125,
0.160400390625
],
[
0.5009765625,
0.15283203125,
0.82861328125,
0.15283203125,
0.82861328125,
0.1666259765625,
0.5009765625,
0.1666259765625
],
[
0.080078125,
0.1610107421875,
0.468017578125,
0.1610107421875,
0.468017578125,
0.1749267578125,
0.080078125,
0.1749267578125
],
[
0.08087158203125,
0.1767578125,
0.468017578125,
0.1767578125,
0.468017578125,
0.1900634765625,
0.08087158203125,
0.1900634765625
],
[
0.609375,
0.175537109375,
0.78515625,
0.175537109375,
0.78515625,
0.1932373046875,
0.609375,
0.1932373046875
],
[
0.859375,
0.1767578125,
0.8935546875,
0.1767578125,
0.8935546875,
0.1925048828125,
0.859375,
0.1925048828125
],
[
0.080078125,
0.1900634765625,
0.468017578125,
0.1925048828125,
0.468017578125,
0.205810546875,
0.080078125,
0.203857421875
],
[
0.080078125,
0.20703125,
0.468017578125,
0.20703125,
0.468017578125,
0.2203369140625,
0.080078125,
0.2203369140625
],
[
0.50341796875,
0.20263671875,
0.79345703125,
0.20263671875,
0.79345703125,
0.2171630859375,
0.50341796875,
0.2171630859375
],
[
0.080078125,
0.2215576171875,
0.4697265625,
0.2215576171875,
0.4697265625,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.50341796875,
0.222900390625,
0.89208984375,
0.222900390625,
0.89208984375,
0.236083984375,
0.50341796875,
0.236083984375
],
[
0.08087158203125,
0.238037109375,
0.46728515625,
0.238037109375,
0.46728515625,
0.248779296875,
0.08087158203125,
0.248779296875
],
[
0.50244140625,
0.238037109375,
0.89208984375,
0.238037109375,
0.89208984375,
0.25244140625,
0.50244140625,
0.25244140625
],
[
0.080078125,
0.251953125,
0.46728515625,
0.251953125,
0.46728515625,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50341796875,
0.255126953125,
0.57861328125,
0.255126953125,
0.57861328125,
0.266357421875,
0.50341796875,
0.266357421875
],
[
0.080078125,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.27978515625,
0.080078125,
0.27978515625
],
[
0.5361328125,
0.279052734375,
0.833984375,
0.279052734375,
0.833984375,
0.29345703125,
0.5361328125,
0.29345703125
],
[
0.859375,
0.27978515625,
0.89208984375,
0.27978515625,
0.89208984375,
0.292236328125,
0.859375,
0.292236328125
],
[
0.080078125,
0.29736328125,
0.34228515625,
0.29736328125,
0.34228515625,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.080078125,
0.2822265625,
0.4697265625,
0.2822265625,
0.4697265625,
0.29541015625,
0.080078125,
0.29541015625
],
[
0.50341796875,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.080078125,
0.3193359375,
0.45751953125,
0.3193359375,
0.45751953125,
0.333984375,
0.080078125,
0.333984375
],
[
0.50244140625,
0.32080078125,
0.890625,
0.32080078125,
0.890625,
0.333984375,
0.50244140625,
0.333984375
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.08087158203125,
0.34228515625,
0.46728515625,
0.34228515625,
0.46728515625,
0.35595703125,
0.08087158203125,
0.35595703125
],
[
0.50244140625,
0.35107421875,
0.67724609375,
0.35107421875,
0.67724609375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.08087158203125,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.37060546875,
0.08087158203125,
0.37060546875
],
[
0.080078125,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.50244140625,
0.376953125,
0.63232421875,
0.376953125,
0.63232421875,
0.39453125,
0.50244140625,
0.39453125
],
[
0.080078125,
0.38818359375,
0.4697265625,
0.38818359375,
0.4697265625,
0.401611328125,
0.080078125,
0.401611328125
],
[
0.07843017578125,
0.401611328125,
0.410888671875,
0.40283203125,
0.410888671875,
0.417236328125,
0.07843017578125,
0.416748046875
],
[
0.50341796875,
0.40283203125,
0.6748046875,
0.40283203125,
0.6748046875,
0.417236328125,
0.50341796875,
0.417236328125
],
[
0.098876953125,
0.41845703125,
0.468017578125,
0.41845703125,
0.468017578125,
0.431884765625,
0.098876953125,
0.431884765625
],
[
0.50244140625,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.50244140625,
0.439453125
],
[
0.07843017578125,
0.431884765625,
0.468017578125,
0.43310546875,
0.468017578125,
0.447509765625,
0.07843017578125,
0.4462890625
],
[
0.50244140625,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50244140625,
0.455078125
],
[
0.080078125,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.462158203125,
0.080078125,
0.462158203125
],
[
0.50341796875,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50341796875,
0.47021484375
],
[
0.080078125,
0.464111328125,
0.4697265625,
0.464111328125,
0.4697265625,
0.479248046875,
0.080078125,
0.479248046875
],
[
0.5009765625,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.48486328125,
0.5009765625,
0.48486328125
],
[
0.08087158203125,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.50341796875,
0.486083984375,
0.89208984375,
0.486083984375,
0.89208984375,
0.5,
0.50341796875,
0.5
],
[
0.08087158203125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.50244140625,
0.5,
0.8896484375,
0.50048828125,
0.8896484375,
0.51513671875,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.5234375,
0.080078125,
0.5234375
],
[
0.0767822265625,
0.52099609375,
0.471435546875,
0.52099609375,
0.471435546875,
0.54150390625,
0.0767822265625,
0.54150390625
],
[
0.5048828125,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.52880859375,
0.5048828125,
0.52880859375
],
[
0.50244140625,
0.52978515625,
0.890625,
0.53076171875,
0.890625,
0.54541015625,
0.50244140625,
0.54443359375
],
[
0.080078125,
0.54052734375,
0.46728515625,
0.54052734375,
0.46728515625,
0.5537109375,
0.080078125,
0.5537109375
],
[
0.50244140625,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.08087158203125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.568359375,
0.08087158203125,
0.568359375
],
[
0.50244140625,
0.5625,
0.890625,
0.5625,
0.890625,
0.57568359375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.5712890625,
0.471435546875,
0.5712890625,
0.471435546875,
0.5859375,
0.080078125,
0.5859375
],
[
0.50341796875,
0.57763671875,
0.89208984375,
0.57763671875,
0.89208984375,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.07843017578125,
0.58544921875,
0.4697265625,
0.5859375,
0.4697265625,
0.60107421875,
0.07843017578125,
0.599609375
],
[
0.50341796875,
0.5947265625,
0.6005859375,
0.5947265625,
0.6005859375,
0.60595703125,
0.50341796875,
0.60595703125
],
[
0.080078125,
0.6015625,
0.4697265625,
0.6015625,
0.4697265625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.07843017578125,
0.61572265625,
0.27685546875,
0.6162109375,
0.27685546875,
0.630859375,
0.07843017578125,
0.6298828125
],
[
0.50341796875,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.50341796875,
0.625
],
[
0.1004638671875,
0.6318359375,
0.465576171875,
0.6318359375,
0.465576171875,
0.64599609375,
0.1004638671875,
0.64599609375
],
[
0.50244140625,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.08251953125,
0.646484375,
0.4697265625,
0.646484375,
0.4697265625,
0.6611328125,
0.08251953125,
0.6611328125
],
[
0.50244140625,
0.6416015625,
0.890625,
0.6416015625,
0.890625,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.07843017578125,
0.6611328125,
0.4697265625,
0.662109375,
0.4697265625,
0.6767578125,
0.07843017578125,
0.67529296875
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.080078125,
0.68115234375,
0.193603515625,
0.68115234375,
0.193603515625,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.200927734375,
0.68115234375,
0.4697265625,
0.68115234375,
0.4697265625,
0.69580078125,
0.200927734375,
0.69580078125
],
[
0.50341796875,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.69873046875,
0.50341796875,
0.69873046875
],
[
0.07843017578125,
0.69580078125,
0.223876953125,
0.6962890625,
0.223876953125,
0.7109375,
0.07843017578125,
0.71044921875
],
[
0.5009765625,
0.7021484375,
0.89208984375,
0.69970703125,
0.89208984375,
0.7138671875,
0.5009765625,
0.71630859375
],
[
0.08087158203125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72900390625,
0.08087158203125,
0.72900390625
],
[
0.50244140625,
0.71728515625,
0.890625,
0.71728515625,
0.890625,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.080078125,
0.744140625
],
[
0.5009765625,
0.73095703125,
0.81884765625,
0.73193359375,
0.81884765625,
0.7470703125,
0.5009765625,
0.74560546875
],
[
0.0767822265625,
0.744140625,
0.47216796875,
0.744140625,
0.47216796875,
0.76513671875,
0.0767822265625,
0.76513671875
],
[
0.51953125,
0.74560546875,
0.890625,
0.7470703125,
0.890625,
0.76123046875,
0.51953125,
0.76025390625
],
[
0.080078125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.50341796875,
0.76220703125,
0.890625,
0.76220703125,
0.890625,
0.77587890625,
0.50341796875,
0.77587890625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.7763671875,
0.890625,
0.77783203125,
0.890625,
0.79150390625,
0.50244140625,
0.79052734375
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.08087158203125,
0.8095703125
],
[
0.50244140625,
0.79150390625,
0.73876953125,
0.79248046875,
0.73876953125,
0.806640625,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.81005859375,
0.443603515625,
0.81005859375,
0.443603515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.181396484375,
0.83203125,
0.365234375,
0.83056640625,
0.365234375,
0.8486328125,
0.181396484375,
0.849609375
],
[
0.43701171875,
0.83203125,
0.4697265625,
0.83203125,
0.4697265625,
0.84765625,
0.43701171875,
0.84765625
],
[
0.50341796875,
0.82763671875,
0.8896484375,
0.82763671875,
0.8896484375,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
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,
0.4697265625,
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,
0.381591796875,
0.88720703125,
0.381591796875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"The classifier () is extended after each task, and needs",
"Ctri = [dp- dn +a]+,",
"(10)",
"to be updated during every task. Thus forgetting also hap-",
"where [l+ denotes the hinge loss, and α is the margin.",
"pens at the classifier. Similar to the constraint on the APG,",
"we sample theknowledge vector from uc ~ NN.Then a",
"Discussion. The proposed APG has the following advan-",
"simple but effective cross-entropy loss is adopted:",
"tages. First, APG continuously learns to generate diverse",
"prompts conditioned on the input of the network: Hence,",
"Cconc = - log (um)(c),",
"(12)",
"APGbreaks thelimitation of selectingprompts from a",
"fixed-size prompt pool and alleviates the dependence on the",
"where (uc)(c) is the c-th element of (uc).",
"pretrained model that is used for querying discrete prompts",
"Training objective for the whole model. The objective",
"fromapool.Second,thepromptcandidatelistisextendable",
"function for optimizing the model and APG is summarized",
"to handle growing knowledge. This enables the model capa-",
"as follows:",
"ble of long incremental learning (see Sec. 4.2 and Sec. 4.3)",
"C= Lattn + Ltri +Cels + LconA+LconC.",
"(13)",
"by the cross-attention operation (Eq. 5).",
"as the APG can aggregate knowledge from different tasks",
"For the backbone f(), it is only trained on the first task and",
"3.3. Anti-forgetting Learning: A Knowledge Pool",
"then is fixed in the case of without pretraining, and other-",
"wise it is fixed all the time. The classifier () and APG are",
"Construction of the knowledge pool. Any trainable mod-",
"optimized across all tasks.",
"ule facestheforgetting (i.e.oldknowledgedegradation)",
"problem when learning from sequential tasks, including the",
"4. Experiments",
"proposed APG and the classifier (). Hence, we develop a",
"knowledge pool to better maintain old knowledge.",
"4.1. Experiment Setup",
"Once finishing learning at task Tt, we summarize the",
"Datasets. We conduct experiments on three datasets: CI-",
"knowledge in the training set Dt into the knowledge pool",
"FAR100 [25], ImageNet-Subset [6], and ImageNet-R [17]",
"before dropping the dataset. Specifically, denote all the im-",
"CIFAR100 [25] contains 100 classes. There are 50,000 im-",
"ages in class c as De = (rtm, ymlctm E Dt, ym = c]. We",
"ages for training and 10,000 images for evaluation. Ima-",
"feed these images into the backbone and get a set of features",
"geNet [6] is a large-scale dataset containing 1,000 classes,",
"V! =- [um [rtm E Ds] from the I-th transformer layer. Then",
"1300 images per class.We follow prior works [8,18,19,26,",
"the class centroid μ, is then calculated by averaging features",
"and the and , is a matrix where each element f(c.k) is the",
"54.58,62,63,651and useImageNet-Subsetwhichcontains",
"100 classes. ImageNet-R [17] dataset contains newly col-",
"covariance between the i-th feature and the k-th feature. We",
"lected data that are different styles of the original ImageNet",
"usetheclass-specificstatisticstoformamultivariatenormal",
"classes. We follow prior work [60] to split the dataset with",
"distribution N! = N(μ,) for each class. Besides, to",
"80% samples for training and the rest 20% samples are used",
"better constrain the APG, the feature centroid μ is used to",
"for evaluation.",
"extract corresponding prompts byfeeding the centroid into",
"the APG, i.e. Pc =APG(μc).",
"Training and testing protocols. For the training and test-",
"Finally, we form the knowledge pool for class c as S.",
"ing protocols, we follow previous works [8, 18, 19,54, 58,",
"(N,NN,P),whereNNis computedfrom thefeatures",
"60,61, 63] to conduct class-incremental learning. After ev-",
"from the last layer. Generally, we store the statistics for",
"ery training stage, we evaluate the model by testing on the",
"union of all testing sets of seen tasks. The classification ac-",
"each class [S.]",
"1, where Nels = t /Vt is the number",
"curacies in everytaskareaveraged,and wereporttheaver-",
"of seen classes so far.",
"age accuracy in all experiments [8, 19,48,54,60-63]. When",
"Anti-forgetting learning. For exemplar-free incremental",
"comparing with L2P [61] and DualPrompt [60], we also re-",
"learning, we can only access the current training set Dt dur-",
"port the forgetting metric following their papers",
"ing task T. The knowledge pool (Se) described above",
"We do not save any images as a memory buffer and only",
"is used to maintain old knowledge. To alleviate forgetting in",
"keep the class statics of old classes on the feature level",
"the APG, we sample from N, and get intermediate knowl-",
"following the exemplar-free class-incremental learning pro-",
"edge vector uc ~ Nl. With the vector and the correspond-",
"posed in PASS [66] and SSRE [67].",
"ing prompt centroid Pj, we apply a constraint on APG:",
"Implementation details. For all experiments below, we",
"LconA = Φ(APG(u),P),",
"(11)",
"usethe vanilla ViT as our backbone.For the standard non-",
"pretrained setting, we adjust the embedding dimension and",
"where is a distance metric and we find that the L1 loss",
"number of heads to match the number of parameters of the",
"is useful. With the constraint, APG can be reminded of old",
"widely used ResNet-18 [15]. For more information about",
"knowledge and perform aggregation on them.",
"the adjusted backbone, please refer to the appendix. We"
] |
{
"arxiv_id": "2308.10445",
"language": "en",
"page_idx": 4
}
|
||
2007.10449-en-0
|
2007.10449-en-0.jpg
|
# Sinkhorn Barycenter via Functional Gradient $\textbf{Descent}$
Zebang Shen* Zhenfu Wangt Alejandro Ribeiro
Hamed Hassani$^{*}$
$^*$Department of Electrical and Systems Engineering $^\dagger$Department of Mathematics
University of Pennsylvania
$\{$zebang®seas,zwang423@math,aribeiro@seas,hassani@seas$\}.$upenn.edu
### $\mathbf{Abstract}$
In this paper, we consider the problem of computing the barycenter of a set of probability distributions under the Sinkhorn divergence. This problem has recently found applications across various domains, including graphics, learning, and vision, as it provides a meaningful mechanism to aggregate knowledge. Unlike previous $\hat{\text{approaches which directly operate in the space of probability measures, we recast}}$ the Sinkhorn barycenter problem as an instance of unconstrained functional optimization and develop a novel functional gradient descent method named Sinkhorn Descent (SD). We prove that SD converges to a stationary point at a sublinear rate, and under reasonable assumptions, we further show that it asymptotically finds a global minimizer of the Sinkhorn barycenter problem. Moreover, by providing a mean-field analysis, we show that SD preserves the weak convergence of empirical measures. Importantly, the computational complexity of SD scales linearly in the dimension $\hat{d\text{ and we demonstrate its scalability by solving a 100-dimensional}}$ Sinkhorn barycenter problem.
### 1 Introduction
Computing a nonlinear interpolation between a set of probability measures is a foundational task across many disciplines. This problem is typically referred as the barycenter problem and, as it provides a meaningful metric to aggregate knowledge, it has found numerous applications. Examples $\hat{\text{include distribution clustering[Ye et al., 2017], Bayesian inference[Srivastavâ et al., 2015], texture}}$ mixing [Rabin et al., 2011], and graphics [Solomon et al., 2015], etc. The barycenter problem can be naturally cast as minimization of the average distance between the target measure (barycenter) and the source measures; and the choice of the distance metric can significantly impact the quality of the barycenter [Feydy et al., 2019]. In this regard, the Optimal Transport (OT) distance (a.k.a. the Wasserstein distance) and its entropy regularized variant (a.k.a. the Sinkhorn divergence) are the most suitable geometrically-faithful metrics, while the latter is more computational friendly. In this paper, we provide efficient and provable methods for the Sinkhorn barycenter problem.
The prior work in this domain has mainly focused on finding the barycenter by optimizing directly in the space of (discrete) probability measures. We can divide these previous methods into three broad classes depending on how the support of the barycenter is determined:
(i) The first class assumes a fixed and prespecified support set for the barycenter and only optimizes the corresponding weights[Staib et al., 2017, Dvurechenskii et al., 2018, Kroshnin et al., 2019]. Accordingly, the problem reduces to minimizing a convex objective subject to a simplex constraint. However, fixing the support without any prior knowledge creates undesired bias and affects the quality of the final solution. $\hat{\text{While increasing the support size(possibly exponentially in the dimension }d})$ can help to mitigate the bias, it renders the procedure computationally prohibitive as $d$ grows.
(ii) To reduce the bias, the second class considers optimizing the support and the weights through an alternating procedure [Cuturi and Doucet, 2014, Claici et al., 2018]. Since the barycenter objective is Preprint. Under review.
|
[
[
0.2254638671875,
0.126953125,
0.77685546875,
0.126953125,
0.77685546875,
0.14453125,
0.2254638671875,
0.14453125
],
[
0.450927734375,
0.15087890625,
0.54833984375,
0.15283203125,
0.54833984375,
0.1717529296875,
0.450927734375,
0.169189453125
],
[
0.2418212890625,
0.222900390625,
0.346435546875,
0.2215576171875,
0.346435546875,
0.236083984375,
0.2418212890625,
0.2374267578125
],
[
0.358642578125,
0.222900390625,
0.46728515625,
0.222900390625,
0.46728515625,
0.2374267578125,
0.358642578125,
0.2374267578125
],
[
0.46728515625,
0.222900390625,
0.7548828125,
0.222900390625,
0.7548828125,
0.2374267578125,
0.46728515625,
0.2374267578125
],
[
0.2254638671875,
0.238037109375,
0.56787109375,
0.238037109375,
0.56787109375,
0.251953125,
0.2254638671875,
0.251953125
],
[
0.5595703125,
0.238037109375,
0.77392578125,
0.238037109375,
0.77392578125,
0.251953125,
0.5595703125,
0.251953125
],
[
0.41015625,
0.25244140625,
0.5888671875,
0.25244140625,
0.5888671875,
0.266357421875,
0.41015625,
0.266357421875
],
[
0.223876953125,
0.26513671875,
0.77392578125,
0.266357421875,
0.77392578125,
0.281005859375,
0.223876953125,
0.27978515625
],
[
0.0236968994140625,
0.2734375,
0.058013916015625,
0.2734375,
0.058013916015625,
0.339111328125,
0.0236968994140625,
0.339111328125
],
[
0.461669921875,
0.31640625,
0.5390625,
0.31640625,
0.5390625,
0.330810546875,
0.461669921875,
0.330810546875
],
[
0.0236968994140625,
0.32763671875,
0.058013916015625,
0.32763671875,
0.058013916015625,
0.40087890625,
0.0236968994140625,
0.40087890625
],
[
0.233642578125,
0.3466796875,
0.76708984375,
0.3466796875,
0.76708984375,
0.35986328125,
0.233642578125,
0.35986328125
],
[
0.2344970703125,
0.35986328125,
0.765625,
0.35986328125,
0.765625,
0.373779296875,
0.2344970703125,
0.373779296875
],
[
0.233642578125,
0.375,
0.76708984375,
0.375,
0.76708984375,
0.38818359375,
0.233642578125,
0.38818359375
],
[
0.233642578125,
0.38818359375,
0.765625,
0.38818359375,
0.765625,
0.40283203125,
0.233642578125,
0.40283203125
],
[
0.0228729248046875,
0.401611328125,
0.062103271484375,
0.401611328125,
0.062103271484375,
0.4912109375,
0.0228729248046875,
0.4912109375
],
[
0.233642578125,
0.40283203125,
0.76708984375,
0.40283203125,
0.76708984375,
0.416748046875,
0.233642578125,
0.416748046875
],
[
0.233642578125,
0.416748046875,
0.76708984375,
0.416748046875,
0.76708984375,
0.429931640625,
0.233642578125,
0.429931640625
],
[
0.233642578125,
0.429931640625,
0.765625,
0.429931640625,
0.765625,
0.443115234375,
0.233642578125,
0.443115234375
],
[
0.233642578125,
0.443115234375,
0.76708984375,
0.443115234375,
0.76708984375,
0.45703125,
0.233642578125,
0.45703125
],
[
0.233642578125,
0.458251953125,
0.765625,
0.458251953125,
0.765625,
0.4716796875,
0.233642578125,
0.4716796875
],
[
0.233642578125,
0.4716796875,
0.76708984375,
0.4716796875,
0.76708984375,
0.48486328125,
0.233642578125,
0.48486328125
],
[
0.041656494140625,
0.481689453125,
0.052276611328125,
0.481689453125,
0.052276611328125,
0.486083984375,
0.041656494140625,
0.486083984375
],
[
0.0228729248046875,
0.4892578125,
0.05963134765625,
0.4892578125,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.2320556640625,
0.484130859375,
0.765625,
0.484130859375,
0.765625,
0.498779296875,
0.2320556640625,
0.498779296875
],
[
0.233642578125,
0.498779296875,
0.765625,
0.498779296875,
0.765625,
0.51220703125,
0.233642578125,
0.51220703125
],
[
0.233642578125,
0.51220703125,
0.76708984375,
0.51220703125,
0.76708984375,
0.52587890625,
0.233642578125,
0.52587890625
],
[
0.233642578125,
0.5263671875,
0.4306640625,
0.5263671875,
0.4306640625,
0.54052734375,
0.233642578125,
0.54052734375
],
[
0.17724609375,
0.56640625,
0.19775390625,
0.56640625,
0.19775390625,
0.5771484375,
0.17724609375,
0.5771484375
],
[
0.2017822265625,
0.564453125,
0.31201171875,
0.564453125,
0.31201171875,
0.5791015625,
0.2017822265625,
0.5791015625
],
[
0.1732177734375,
0.59326171875,
0.82275390625,
0.59326171875,
0.82275390625,
0.60693359375,
0.1732177734375,
0.60693359375
],
[
0.1748046875,
0.609375,
0.826171875,
0.609375,
0.826171875,
0.62255859375,
0.1748046875,
0.62255859375
],
[
0.1748046875,
0.62255859375,
0.82421875,
0.62255859375,
0.82421875,
0.63623046875,
0.1748046875,
0.63623046875
],
[
0.1748046875,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.1748046875,
0.6494140625,
0.82421875,
0.6494140625,
0.82421875,
0.66357421875,
0.1748046875,
0.66357421875
],
[
0.1748046875,
0.66357421875,
0.826171875,
0.66357421875,
0.826171875,
0.6767578125,
0.1748046875,
0.6767578125
],
[
0.1748046875,
0.67822265625,
0.82275390625,
0.67822265625,
0.82275390625,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1748046875,
0.69140625,
0.82421875,
0.69140625,
0.82421875,
0.705078125,
0.1748046875,
0.705078125
],
[
0.1748046875,
0.705078125,
0.82421875,
0.705078125,
0.82421875,
0.7197265625,
0.1748046875,
0.7197265625
],
[
0.1732177734375,
0.71728515625,
0.82666015625,
0.71826171875,
0.82666015625,
0.73291015625,
0.1732177734375,
0.73193359375
],
[
0.1748046875,
0.73291015625,
0.69775390625,
0.73291015625,
0.69775390625,
0.7470703125,
0.1748046875,
0.7470703125
],
[
0.1748046875,
0.75244140625,
0.82421875,
0.75244140625,
0.82421875,
0.7666015625,
0.1748046875,
0.7666015625
],
[
0.1748046875,
0.767578125,
0.826171875,
0.767578125,
0.826171875,
0.78076171875,
0.1748046875,
0.78076171875
],
[
0.1748046875,
0.78076171875,
0.6337890625,
0.78076171875,
0.6337890625,
0.79443359375,
0.1748046875,
0.79443359375
],
[
0.1732177734375,
0.79443359375,
0.826171875,
0.79443359375,
0.826171875,
0.80810546875,
0.1732177734375,
0.80810546875
],
[
0.1748046875,
0.80810546875,
0.826171875,
0.80810546875,
0.826171875,
0.8212890625,
0.1748046875,
0.8212890625
],
[
0.1732177734375,
0.8212890625,
0.82421875,
0.8212890625,
0.82421875,
0.83544921875,
0.1732177734375,
0.83544921875
],
[
0.1748046875,
0.8359375,
0.82421875,
0.8359375,
0.82421875,
0.849609375,
0.1748046875,
0.849609375
],
[
0.1748046875,
0.849609375,
0.826171875,
0.849609375,
0.826171875,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.1748046875,
0.86279296875,
0.78369140625,
0.86279296875,
0.78369140625,
0.876953125,
0.1748046875,
0.876953125
],
[
0.17236328125,
0.876953125,
0.82421875,
0.87548828125,
0.82421875,
0.89013671875,
0.17236328125,
0.8916015625
],
[
0.1748046875,
0.8916015625,
0.82421875,
0.8916015625,
0.82421875,
0.90478515625,
0.1748046875,
0.90478515625
],
[
0.1748046875,
0.92529296875,
0.310546875,
0.92529296875,
0.310546875,
0.93603515625,
0.1748046875,
0.93603515625
]
] |
[
"Sinkhorn Barycenter via Functional Gradient",
"Descent",
"Zebang Shen*",
"Zhenfu Wangt",
"Alejandro Ribeiro*Hamed Hassani*",
"*Department of Electrical and Systems Engineering",
"gDepartmentofMathematics",
"Universityof Pennsylvania",
"(zebang@seas,zwang423@math,aribeiro@seas,hassani@seas.upenn.edu",
"2",
"Abstract",
"[ z",
"In this paper, we consider the problem of computing the barycenter of a set of",
"probability distributions under the Sinkhorn divergence. This problem has recently",
"found applications across various domains, including graphics, learning, and vision,",
"as it provides a meaningful mechanism to aggregate knowledge. Unlike previous",
"[T'S]",
"approaches which directly operate in the space of probability measures, we recast",
"the Sinkhorn barycenter problem as an instance of unconstrained functional opti-",
"mization and develop a novel functional gradient descent method named Sinkhorn",
"Descent (SD). We prove that SD converges to a stationary point at a sublinear rate,",
"and under reasonable assumptions, we further show that it asymptotically finds a",
"global minimizer of the Sinkhorn barycenter problem.Moreover, by providing a",
"",
"",
"mean-field analysis, we show that SD preserves the weak convergence of empiri-",
"cal measures. Importantly, the computational complexity of sD scales linearly in",
"the dimension d and we demonstrate its scalability by solving a 100-dimensional",
"Sinkhorn barycenter problem.",
"1.",
"Introduction",
"Computing a nonlinear interpolation between a set of probability measures is a foundational task",
"across many disciplines.Thisproblem is typicallyreferred as the barycenter problem and,as it",
"provides a meaningful metric to aggregate knowledge, it has found numerous applications. Examples",
"include distribution clustering [Ye et al., 2017], Bayesian inference [Srivastava et al., 2015], texture",
"mixing [Rabin et al., 2011], and graphics [Solomon et al., 2015], etc. The barycenter problem can",
"be naturally cast as minimization of the average distance between the target measure (barycenter)",
"andthesourcemeasures:andthechoiceofthedistancemetriccansignificantlyimpactthequalityof",
"the barycenter [Feydy et al., 2019]. In this regard, the Optimal Transport (OT) distance (a.k.a. the",
"Wasserstein distance) and its entropy regularized variant (a.k.a. the Sinkhorn divergence) are the most",
"suitable geometrically-faithful metrics, while the latter is more computational friendly. In this paper,",
"we provide efficient and provable methods for the Sinkhorn barycenter problem.",
"The prior work in this domain has mainly focused on finding the barycenter by optimizing directly in",
"the space of (discrete) probability measures. We can divide these previous methods into three broad",
"classes depending on how the support of the barycenter is determined:",
"(i) The first class assumes a fixed and prespecified support set for the barycenter and only optimizes",
"the corresponding weights [Staib et al., 2017, Dvurechenskii et al., 2018, Kroshnin et al., 2019].",
"Accordingly, the problem reduces to minimizing a convex objective subject to a simplex constraint.",
"However, fixing the support without any prior knowledge creates undesired bias and affects the quality",
"of the final solution. While increasing the support size (possibly exponentially in the dimension d)",
"can help to mitigate the bias, it renders the procedure computationally prohibitive as d grows.",
"(ii) To reduce the bias, the second class considers optimizing the support and the weights through an",
"alternating procedure [Cuturi and Doucet, 2014, Claici et al., 2018]. Since the barycenter objective is",
"Preprint.Under review"
] |
{
"arxiv_id": "2007.10449",
"language": "en",
"page_idx": 0
}
|
||
2007.10449-en-1
|
2007.10449-en-1.jpg
|
not jointly convex with respect to the support and the weights, these methods in general only converge
to a stationary point, which can be far from the true minimizers.
(iii) Unlike the aforementioned classes, Luise et al.[2019] recently proposed a conditional gradient
$\begin{array}{c}\text{method with a growing support set. This method enjoys sublinear convergence to the global optimum}\\\text{under the premise that a }d\text{-dimensional nonconvex subproblem can be globally minimized per-}\end{array}$
iteration. However, nonconvex optimization is generally intractable in high dimensional problems
(large $d$) and only stationary points can be efficiently reached. Hence, the guarantee of [Luise et al.,
2019] has limited applicability as the dimension grows.
In this paper, we provide a new perspective on the Sinkhorn barycenter problem: Instead of operating
in the space of probability measures, we view the barycenter as the push-forward measure of a
given initial measure under an unknown mapping. We thus recast the barycenter problem as an
unconstrained functional optimization over the space of mappings. Equipped with this perspective,
we make the following contributions:
$\bullet$ We develop a novel functional gradient descent method, called Sinkhorn Descent (SD),
which operates by finding the push-forward mapping in a Reproducing Kernel Hilbert Space
that allows the fastest descent, and consequently solves the Sinkhorn barycenter problem
iteratively. We then define the Kernelized Sinkhorn Barycenter Discrepancy (KSBD) to
characterize the non-asymptotic convergence of SD. In particular, we prove that KSBD
vanishes under the SD iterates at the rate of $\mathcal{O}(\frac1t)$, where $\hat{t}$ is the iteration number.
$\bullet$ We prove that SD preserves the weak convergence of empirical measures. Concretely, use
SD$^t\hat{(}\cdot)$ to denote the output of SD after $t$ iterations and let $\alpha_N$ be an empirical measure of $\alpha$
with $N$ samples. We have $\lim_N\to\infty\mathbf{SD}^t(\alpha_N)=\mathbf{SD}^t(\alpha).$ Such asymptotic analysis allows
us to jointly study the behavior of SD under either discrete or continuous initialization.
$\bullet$ Under a mild assumption, we prove that KSBD is a valid discrepancy to characterize the
optimality of the solution, i.e. the vanishing of KSBD implies the output measure of SD
converges to the global optimal solution set of the Sinkhorn barycenter problem.
Further, we show the efficiency and efficacy of SD by comparing it with prior art on several problems.
We note that the computation complexity of SD depends linearly on the dimension $d. \text{ \^ {W}ehence}$
validate the scalability of SD by solving a 100-dimensional barycenter problem, which cannot be
handled by previous methods due to their exponential dependence on the problem dimension.
Notations. Let $\mathcal{X}\subseteq\mathbb{R}^d$ be a compact ground set, endowed with a symmetric ground metric $c:\mathcal{X}\times\mathcal{X}\to\mathbb{R}_+.$ Without loss of generality, we assume $c(x,y)=\infty$ if $x\notin\mathcal{X}$ or $y\notin\mathcal{X}.$ We use $\nabla_{1}c(\cdot,\cdot):\mathcal{X}^{2}\to\mathcal{X}$ to denote its gradient w.r.t. its frst argument. Let $\mathcal{M}_1^{+}(\mathcal{X})$ and $\mathcal{C}(\mathcal{X})$ be the space of probability measures and continuous functions on $\tilde{\chi}.$ We denote the support for a probability measure $\alpha\in\mathcal{M}_1^{\prime}(\mathcal{X})$ by supp$(\alpha)$ and we use $\alpha-a.e.$ to denote "almost everywhere w.r.t. $\alpha$". For a vector $\mathbf{a}\in\mathbb{R}^d$, we denote its $\ell_2$ norm by $\|\mathbf{a}\|.$ For a function $f:\mathcal{X}\to\mathbb{R}$, we denote its $L^\infty$ $\operatorname{norm}$by$\left\|f\right\|_\infty:=\max_{x\in\mathcal{X}}\left|f(x)\right|$and denote its gradient by $\nabla f(\cdot):\mathcal{X}\to\mathbb{R}^d.$ For a vector function $f:\mathcal{X}\to\mathbb{R}^d$, we denote its $(2,\infty)$ norm by $\|f\|_2,\infty:=\max_{x\in\mathcal{X}}\|f(x)\|.$ For an integer $n$,denote $[n]{:}=\{1,\cdots,n\}.$
Given an Reproducing Kernel Hilbert Space (RKHS) $\mathcal{H}$ with a kernel function $k:\mathcal{X}\times\mathcal{X}\to\mathbb{R}_+$, we say a vector function $\psi=[[\psi]_1,\cdots,[\psi]_d]\in\mathcal{H}^d$ if each component $[\psi]_i$ is in $\mathcal{H}.$ The space $\mathcal{H}$ has a natural inner product structure and an induced norm, and so does $\mathcal{H}^d$, i.e. $\langle f,g\rangle_\mathcal{H}^d=$ $\sum_{i=1}^d\langle[f]_i,[g]_i\rangle_{\mathcal{H}},\forall f,g\in\mathcal{H}^d$ and the norm $\|f\|_{\mathcal{H}^d}^2=\langle f,f\rangle_{\mathcal{H}^d}.$ The reproducing property of the RKHS $\mathcal{H}$ reads that given $f\in\mathcal{H}^d$, one has $[f]_i(x)=\langle[f]_i,k_x\rangle_{\mathcal{H}}$ with $k_x(y)=k(x,y)$, which by Cauchy-Schwarz inequality implies that there exists some constant $M_{\mathcal{H}}>0$ such that
$$\|f\|_{2,\infty}\leq M_{\mathcal H}\|f\|_{\mathcal H^d},\forall f\in\mathcal H^d.$$
(1)
Additionally, for a functional $F:\mathcal{H}^d\to\mathbb{R}$, the Fréchet derivative of $F$ is defined as follows.
Definition 1.1 (Fréchet derivative in RKHS). For afunctional $F: \mathcal{H} ^d\to \mathbb{R} , its\textit{Fr\'{e} chet derivative}$
$DF[\psi]$ at $\psi\in{\grave{\mathcal{H}}}^d$ is afunction in $\mathcal{H}^d$ satisfying the following: For any $\xi\in{\dot{\mathcal{H}}}^d$ with $\|\xi\|_{\mathcal{H}^d}<\infty$,
$$\begin{aligned}\lim_{\epsilon\to0}\frac{F[\psi+\epsilon\xi]-F[\psi]}{\epsilon}=\langle DF[\psi],\xi\rangle_{\mathcal{H}^{d}}.\end{aligned}$$
Note that the Fréchet derivative at $\psi$,i.e. $DF[\psi]$, is a bounded linear operator from $\mathcal{H}^d$ to $\mathbb{R}.$ It can be written in the form $DF[\psi](\xi)=\langle DF[\psi],\xi\rangle_{\mathcal{H}^{d}}$ due to the Riesz-Fréchet representation theorem.
2
|
[
[
0.17236328125,
0.091552734375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.106689453125,
0.17236328125,
0.1060791015625
],
[
0.1732177734375,
0.106689453125,
0.59326171875,
0.106689453125,
0.59326171875,
0.12060546875,
0.1732177734375,
0.12060546875
],
[
0.176513671875,
0.121826171875,
0.82177734375,
0.121826171875,
0.82177734375,
0.1319580078125,
0.176513671875,
0.1319580078125
],
[
0.1748046875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.1748046875,
0.147705078125
],
[
0.1732177734375,
0.146484375,
0.82421875,
0.146484375,
0.82421875,
0.160400390625,
0.1732177734375,
0.160400390625
],
[
0.176513671875,
0.1634521484375,
0.82275390625,
0.1634521484375,
0.82275390625,
0.173583984375,
0.176513671875,
0.173583984375
],
[
0.1748046875,
0.175537109375,
0.826171875,
0.175537109375,
0.826171875,
0.189453125,
0.1748046875,
0.189453125
],
[
0.1748046875,
0.189453125,
0.5361328125,
0.189453125,
0.5361328125,
0.20263671875,
0.1748046875,
0.20263671875
],
[
0.1732177734375,
0.208984375,
0.82177734375,
0.208984375,
0.82177734375,
0.2235107421875,
0.1732177734375,
0.2235107421875
],
[
0.176513671875,
0.2247314453125,
0.82275390625,
0.2247314453125,
0.82275390625,
0.23486328125,
0.176513671875,
0.23486328125
],
[
0.1732177734375,
0.2374267578125,
0.82275390625,
0.236083984375,
0.82275390625,
0.25,
0.1732177734375,
0.250732421875
],
[
0.1748046875,
0.251953125,
0.82666015625,
0.251953125,
0.82666015625,
0.26513671875,
0.1748046875,
0.26513671875
],
[
0.1732177734375,
0.26513671875,
0.42236328125,
0.26513671875,
0.42236328125,
0.279052734375,
0.1732177734375,
0.279052734375
],
[
0.2181396484375,
0.28857421875,
0.826171875,
0.28857421875,
0.826171875,
0.3017578125,
0.2181396484375,
0.3017578125
],
[
0.233642578125,
0.3037109375,
0.82177734375,
0.3037109375,
0.82177734375,
0.314453125,
0.233642578125,
0.314453125
],
[
0.2344970703125,
0.31640625,
0.82275390625,
0.31640625,
0.82275390625,
0.330078125,
0.2344970703125,
0.330078125
],
[
0.233642578125,
0.330078125,
0.82421875,
0.330078125,
0.82421875,
0.343505859375,
0.233642578125,
0.343505859375
],
[
0.233642578125,
0.34228515625,
0.82275390625,
0.34228515625,
0.82275390625,
0.35595703125,
0.233642578125,
0.35595703125
],
[
0.233642578125,
0.35791015625,
0.76708984375,
0.35791015625,
0.76708984375,
0.371826171875,
0.233642578125,
0.371826171875
],
[
0.2181396484375,
0.375,
0.82275390625,
0.375,
0.82275390625,
0.38818359375,
0.2181396484375,
0.38818359375
],
[
0.2320556640625,
0.386962890625,
0.82421875,
0.38818359375,
0.82421875,
0.40283203125,
0.2320556640625,
0.401611328125
],
[
0.233642578125,
0.40283203125,
0.82421875,
0.40283203125,
0.82421875,
0.416748046875,
0.233642578125,
0.416748046875
],
[
0.2344970703125,
0.41845703125,
0.7958984375,
0.41845703125,
0.7958984375,
0.4287109375,
0.2344970703125,
0.4287109375
],
[
0.21728515625,
0.434326171875,
0.82275390625,
0.434326171875,
0.82275390625,
0.447509765625,
0.21728515625,
0.447509765625
],
[
0.233642578125,
0.447509765625,
0.82421875,
0.447509765625,
0.82421875,
0.4609375,
0.233642578125,
0.4609375
],
[
0.2320556640625,
0.462158203125,
0.75732421875,
0.462158203125,
0.75732421875,
0.47607421875,
0.2320556640625,
0.47607421875
],
[
0.1748046875,
0.48486328125,
0.826171875,
0.48486328125,
0.826171875,
0.498779296875,
0.1748046875,
0.498779296875
],
[
0.1748046875,
0.5,
0.82275390625,
0.5,
0.82275390625,
0.51318359375,
0.1748046875,
0.51318359375
],
[
0.1748046875,
0.51318359375,
0.82421875,
0.51318359375,
0.82421875,
0.5263671875,
0.1748046875,
0.5263671875
],
[
0.1732177734375,
0.5263671875,
0.7861328125,
0.5263671875,
0.7861328125,
0.54052734375,
0.1732177734375,
0.54052734375
],
[
0.1748046875,
0.5546875,
0.82275390625,
0.5546875,
0.82275390625,
0.5693359375,
0.1748046875,
0.5693359375
],
[
0.1748046875,
0.5693359375,
0.82275390625,
0.5693359375,
0.82275390625,
0.58349609375,
0.1748046875,
0.58349609375
],
[
0.176513671875,
0.583984375,
0.82275390625,
0.583984375,
0.82275390625,
0.5986328125,
0.176513671875,
0.5986328125
],
[
0.1732177734375,
0.59765625,
0.82275390625,
0.5966796875,
0.82275390625,
0.611328125,
0.1732177734375,
0.6123046875
],
[
0.1748046875,
0.6123046875,
0.82275390625,
0.6123046875,
0.82275390625,
0.626953125,
0.1748046875,
0.626953125
],
[
0.1732177734375,
0.626953125,
0.82177734375,
0.626953125,
0.82177734375,
0.64013671875,
0.1732177734375,
0.64013671875
],
[
0.17236328125,
0.6416015625,
0.826171875,
0.6396484375,
0.826171875,
0.65576171875,
0.17236328125,
0.65771484375
],
[
0.1748046875,
0.65478515625,
0.826171875,
0.65576171875,
0.826171875,
0.673828125,
0.1748046875,
0.67236328125
],
[
0.176513671875,
0.671875,
0.291748046875,
0.671875,
0.291748046875,
0.68505859375,
0.176513671875,
0.68505859375
],
[
0.1732177734375,
0.68310546875,
0.82177734375,
0.68310546875,
0.82177734375,
0.69775390625,
0.1732177734375,
0.69775390625
],
[
0.1732177734375,
0.69873046875,
0.82275390625,
0.69775390625,
0.82275390625,
0.7119140625,
0.1732177734375,
0.71337890625
],
[
0.1732177734375,
0.7119140625,
0.81201171875,
0.7138671875,
0.81201171875,
0.72900390625,
0.1732177734375,
0.7265625
],
[
0.1829833984375,
0.72900390625,
0.82421875,
0.72900390625,
0.82421875,
0.7470703125,
0.1829833984375,
0.7470703125
],
[
0.1732177734375,
0.74365234375,
0.826171875,
0.744140625,
0.826171875,
0.76220703125,
0.1732177734375,
0.76123046875
],
[
0.1732177734375,
0.7587890625,
0.73876953125,
0.76025390625,
0.73876953125,
0.7744140625,
0.1732177734375,
0.7734375
],
[
0.3896484375,
0.77880859375,
0.60888671875,
0.7763671875,
0.60888671875,
0.79443359375,
0.3896484375,
0.796875
],
[
0.80224609375,
0.77978515625,
0.826171875,
0.77978515625,
0.826171875,
0.79345703125,
0.80224609375,
0.79345703125
],
[
0.176513671875,
0.79541015625,
0.77783203125,
0.79541015625,
0.77783203125,
0.8095703125,
0.176513671875,
0.8095703125
],
[
0.176513671875,
0.81103515625,
0.82421875,
0.81103515625,
0.82421875,
0.82568359375,
0.176513671875,
0.82568359375
],
[
0.17236328125,
0.82373046875,
0.81640625,
0.82470703125,
0.81640625,
0.84228515625,
0.17236328125,
0.84130859375
],
[
0.38818359375,
0.83349609375,
0.640625,
0.8486328125,
0.63818359375,
0.87353515625,
0.3857421875,
0.859375
],
[
0.365234375,
0.85302734375,
0.3896484375,
0.85302734375,
0.3896484375,
0.86572265625,
0.365234375,
0.86572265625
],
[
0.365234375,
0.8642578125,
0.390625,
0.8642578125,
0.390625,
0.8720703125,
0.365234375,
0.8720703125
],
[
0.449462890625,
0.86572265625,
0.455810546875,
0.86572265625,
0.455810546875,
0.8720703125,
0.449462890625,
0.8720703125
],
[
0.1748046875,
0.88330078125,
0.82275390625,
0.88330078125,
0.82275390625,
0.89794921875,
0.1748046875,
0.89794921875
],
[
0.1732177734375,
0.8984375,
0.81884765625,
0.8984375,
0.81884765625,
0.91357421875,
0.1732177734375,
0.91357421875
],
[
0.490234375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94873046875,
0.490234375,
0.94873046875
]
] |
[
"not jointly convex with respect to the support and the weights, these methods in general only converge",
"to a stationary point, which can be far from the true minimizers.",
"(iii)Unliketheaforementionedclasses.Luiseetal.[2019lrecentlyproposedaconditionalgradient",
"method with a growing support set. This method enjoys sublinear convergence to the global optimum",
"under the premise that a d-dimensional nonconvex subproblem can be globally minimized per-",
"iteration.However,nonconvexoptimizationisgenerallyintractableinhighdimensionalproblems",
"(large d) and only stationary points can be efficiently reached. Hence, the guarantee of [Luise et al.,",
"2019] has limited applicabilityas the dimension grows.",
"In this paper, we provide a new perspective on the Sinkhorn barycenter problem: Instead of operating",
"inthespace of probabilitymeasures,weviewthebarycenterasthepush-forward measure of a",
"given initial measureunderanunknown mapping.Wethusrecastthebarycenterproblemas an",
"unconstrained functional optimization over the space of mappings. Equipped with this perspective,",
"wemakethefollowing contributions:",
" We develop a novel functional gradient descent method, called sinkhorn Descent (sD),",
"whichoperatesbyfindingthepush-forwardmappinginaReproducingKernelHilbertSpace",
"that allows the fastest descent, and consequently solves the Sinkhorn barycenter problem",
"iteratively. We then define the Kernelized Sinkhorn Barycenter Discrepancy (KSBD) to",
"characterize the non-asymptotic convergence of sD. In particular, we prove that KSBD",
"vanishes under the SD iterates at the rate of O(), where t is the iteration number.",
" We prove that sD preserves the weak convergence of empirical measures. Concretely, use",
"SDt () to denote the output of sD after t iterations and let Q be an empirical measure of α",
"with N samples. We have limN SDt(QN) = SDt(α). Such asymptotic analysis allows",
"us to jointly studythebehaviorof sD under either discrete or continuous initialization.",
". Under a mild assumption, we prove that KSBD is a valid discrepancy to characterize the",
"optimality of the solution, i.e. the vanishing of KSBD implies the output measure of sD",
"converges to the global optimal solution set of the Sinkhorn barycenter problem.",
"Further, we show the efficiency and efficacy of sD by comparing it with prior art on several problems.",
"We note that the computation complexity of sD depends linearly on the dimension d. We hence",
"validate the scalability of sD by solving a 100-dimensional barycenter problem, which cannot be",
"handled by previous methods due to their exponential dependence on the problem dimension.",
"Notations. Let X Rd be a compact ground set, endowed with a symmetric ground metric",
"c : X × X -→ R+. Without loss of generality, we assume c(a, y) = oo if r X or y $ X. We use",
"Vic(, ) : X? → X to denote its gradient w.r.t. its first argument. Let M(X) and C(X) be the",
"space of probability measures and continuous functions on X. We denote the support for a probability",
"measure α E Mf(x) by supp(α) and we use α - a.e. to denote \"almost everywhere w.r.t. α\"",
"Fora vectora E Rd, we denote its l2 norm byall.For a function f : X→R,we denote its Loo",
"norm by llfll:= maxrex If(r)| and denote its gradient by f() : X → Rd. For a vector function",
"f : x → Rd, we denote its (2, oo) norm by Ilfl2,co:= maxaex llf(a)l. For an integer n, denote",
"[n]:=[1,- ,n].",
"Given an Reproducing Kernel Hilbert Space (RKHS) H with a kernel function k : X × X -→ R+",
"we say a vector function =[]i,-.., []d] E Hd if each component [], is in H.The space",
"H has a natural inner product structure and an induced norm, and so does Hd, i.e.(f,g)ad",
"d-([fl, [gli)H, Vf,g E Hd and the norm Ilflla = (f,f)Hd. The reproducing property of the",
"RKHS H reads that given f E Hd, one has [f];() = <[f]e, ka)n with ka(y) = k(r, y), which by",
"Cauchy-Schwarz inequality implies that there exists some constant M > O such that",
"Ilfll2,≤MnllfllHd,Vf eHd",
"(1)",
"Additionally, for a functional F : Hd → R, the Frechet derivative of F is defined as follows.",
"Definition 1.1 (Frechet derivative in RKHS). For a functional F : Hd → IR, its Frechet derivative",
"DF[y] at b E Hd is a function in Hd satisfying the following: For any E E Hd with Ilsllha < o0,",
"Flb + eE] -F[a] = (DF[4], ) d.",
"lim",
"0",
"6",
"Note that the Frechet derivative at , i.e. DF[l, is a bounded linear operator from Hd to R. It can",
"be written in the form DF[]($) = (DF[], $)Ha due to the Riesz-Frechet representation theorem.",
"2"
] |
{
"arxiv_id": "2007.10449",
"language": "en",
"page_idx": 1
}
|
||
2007.10449-en-2
|
2007.10449-en-2.jpg
|
1.1 Related Work on Functional Gradient Descent
A related functional gradient descent type method is the Stein Variation Gradient Descent (SVGD) method by Liu and Wang [2016]. SVGD considers the problem of minimizing the Kullback-Leibler (KL) divergence between a variable distribution and a posterior $p.$ Note that SVGD updates the positions of a set of $N$ particles using the score function of the posterior $p$,i.e. $\nabla\log p.$ Consequently, it requires the access to the target distribution function. Later, Liu [2017] prove that SVGD has convergence guarantee in its continuous-time limit (taking infinitesimal step size) using infinite number of particles $(N\to\infty).$ In comparison, SD is designed to solve the significantly more complicated Sinkhon barycenter problem and has a stronger convergence guarantee. More precisely, while SD updates the measure using only a sampling machinery of the target measures (no score functions), Ît is guaranteed to converge sub-linearly to a stationary point when $\alpha$ is a discrete measure using discrete time steps. This is in sharp contrast to the results for SVGD.
In another work, Mroueh et al.[2019] considers minimizing the Maximum Mean Discrepancy (MMD) between a source measure and a variable measure. They solve this problem by incrementally following a Sobolev critic function and propose the Sobolev Descent (SoD) method. To show the global convergence of the measure sequence generated by SoD, Mroueh et al.[2019] assumes the entire sequence satisfies certain spectral properties, which is in general difficult to verify. Later, Arbel et al.[2019] consider the same MMD minimization problem from a gradient flow perspective They propose two assumptions that if either one holds, the MMD gradient flow converges to the global solution. However, similar to[Mroueh et al., 2019], these assumptions have to be satisfied for the entire measure sequence. We note that the Sinkhorn barycenter is a strict generalization of the above MMD minimization problem and is hence much more challenging: By setting the number of source measures $n=1$ and setting the entropy regularization parameter $\gamma=\infty$, problem (4) degenerates to the special case of MMD. Further, the MMD between two probability measures has a closed form expression while the Sinkhorn Divergence can only be described via a set of optimization problems. Consequently, the Sinkhorn barycenter is significantly more challenging. To guarantee global convergence, the proposed SD algorithm only requires one of accumulation points of the measure sequence to be fully supported on $\chi$ with no restriction on the entire sequence.
## 2 Sinkhorn Barycenter
We first introduce the entropy-regularized optimal transport distance and its debiased version, a.k.a the Sinkhorn divergence. Given two probability measures $\alpha,\beta\in\mathcal{M}_1^+(\mathcal{X})$,use $\Pi(\alpha,\beta)$ to denote the set of joint distributions over $\mathcal{X}^2$ with marginais $\alpha$ and $\beta.$ For $\pi\in\Pi$, use $\langle c,\pi\rangle$ to denote the integral $\langle c,\pi\rangle=\int_{\mathcal{X}^{2}}c(x,y)\mathbf{d}\pi(x,y)$ and use K$\tilde{\mathcal{L}}(\pi||\alpha\otimes\beta)$ to denote the Kullback-Leibler divergence between the candidate transport plan $\pi$ and the product measure $\alpha\otimes\beta.$ The entropy-regularized optimal transport distance
is defined as
$$:\hat{\mathrm{OT}}_{\gamma}(\hat{\alpha},\beta):\mathcal{M}_{1}^{+}(\mathcal{X})^{*}\times\mathcal{M}_{1}^{+}(\mathcal{X})\to\mathbb{R}_{+}$$
(2)
$$\mathrm{OT}_{\gamma}(\alpha,\beta)=\min_{\pi\in\Pi(\alpha,\beta)}\langle c,\pi\rangle+\gamma\mathrm{KL}(\pi||\alpha\otimes\beta).$$
Here, $\gamma>0$ is a regularization parameter. Note that OT$_\gamma(\alpha,\beta)$ is not a valid metric as there exists $\alpha\in\mathcal{M}_{1}^{+}(\mathcal{X})$ such that OT$_\gamma(\alpha,\alpha)\neq0$ when $\gamma\neq0.$ To remove this bias, Peyré et al.[2019] introduced the Sinkhorn divergence S$_\gamma(\alpha,\beta):\mathcal{M}_1^+(\mathcal{X})\times\mathcal{M}_1^+(\mathcal{X})\to\mathbb{R}_+:$
$$\mathbb{S}_{\gamma}(\alpha,\beta):=\mathrm{OT}_{\gamma}(\alpha,\beta)-\frac{1}{2}\mathrm{OT}_{\gamma}(\alpha,\alpha)-\frac{1}{2}\mathrm{OT}_{\gamma}(\beta,\beta),$$
(3)
which is a debiased version of OT$_\gamma(\alpha,\beta).$ It is further proved that $\mathcal{S}_\gamma(\alpha,\beta)$ is nonnegative, biconvex and metrizes the convergence in law when the ground set $\chi$ is compact and the metric $c$ is Lipschitz. Now given a set of probability measures $\{\beta_i\}_{i=1}^n$,the Sinkhorn barycenter is the measure $\alpha\in\mathcal{M}_1^+(\mathcal{X})$ that minimizes the average of Sinkhorn divergences
$$\min_{\alpha\in\mathcal{M}_{1}^{+}(\mathcal{X})}\Big(\mathcal{S}_{\gamma}(\alpha){:}=\frac{1}{n}\sum_{i=1}^{n}\mathbb{S}_{\gamma}(\alpha,\beta_{i})\Big).$$
(4)
We will next focus on the properties of OT$_\gamma$ since $\mathcal S_\gamma(\alpha)$ is the linear combination of these terms.
3
|
[
[
0.178955078125,
0.0947265625,
0.20751953125,
0.0947265625,
0.20751953125,
0.1029052734375,
0.178955078125,
0.1029052734375
],
[
0.200927734375,
0.091552734375,
0.5361328125,
0.091552734375,
0.5361328125,
0.10479736328125,
0.200927734375,
0.10479736328125
],
[
0.1748046875,
0.11932373046875,
0.82421875,
0.11932373046875,
0.82421875,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.1748046875,
0.1331787109375,
0.82421875,
0.1331787109375,
0.82421875,
0.147705078125,
0.1748046875,
0.147705078125
],
[
0.1748046875,
0.147705078125,
0.82421875,
0.147705078125,
0.82421875,
0.1610107421875,
0.1748046875,
0.1610107421875
],
[
0.176513671875,
0.1622314453125,
0.82421875,
0.1622314453125,
0.82421875,
0.17236328125,
0.176513671875,
0.17236328125
],
[
0.1748046875,
0.1749267578125,
0.826171875,
0.1749267578125,
0.826171875,
0.1881103515625,
0.1748046875,
0.1881103515625
],
[
0.1732177734375,
0.1881103515625,
0.82421875,
0.1881103515625,
0.82421875,
0.2020263671875,
0.1732177734375,
0.2020263671875
],
[
0.1748046875,
0.2020263671875,
0.82421875,
0.2020263671875,
0.82421875,
0.21533203125,
0.1748046875,
0.21533203125
],
[
0.1748046875,
0.216552734375,
0.826171875,
0.216552734375,
0.826171875,
0.2298583984375,
0.1748046875,
0.2298583984375
],
[
0.1748046875,
0.2298583984375,
0.82421875,
0.2298583984375,
0.82421875,
0.24365234375,
0.1748046875,
0.24365234375
],
[
0.1748046875,
0.24365234375,
0.82421875,
0.24365234375,
0.82421875,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.1748046875,
0.2568359375,
0.65283203125,
0.2568359375,
0.65283203125,
0.270751953125,
0.1748046875,
0.270751953125
],
[
0.1732177734375,
0.27587890625,
0.82421875,
0.276611328125,
0.82421875,
0.291748046875,
0.1732177734375,
0.290283203125
],
[
0.1732177734375,
0.290283203125,
0.82421875,
0.291748046875,
0.82421875,
0.30615234375,
0.1732177734375,
0.304931640625
],
[
0.1748046875,
0.30615234375,
0.82421875,
0.30615234375,
0.82421875,
0.3193359375,
0.1748046875,
0.3193359375
],
[
0.1732177734375,
0.318115234375,
0.82421875,
0.317626953125,
0.82421875,
0.33203125,
0.1732177734375,
0.333251953125
],
[
0.1748046875,
0.333251953125,
0.826171875,
0.333251953125,
0.826171875,
0.3466796875,
0.1748046875,
0.3466796875
],
[
0.1748046875,
0.3466796875,
0.82666015625,
0.3466796875,
0.82666015625,
0.35986328125,
0.1748046875,
0.35986328125
],
[
0.176513671875,
0.3623046875,
0.82275390625,
0.3623046875,
0.82275390625,
0.37255859375,
0.176513671875,
0.37255859375
],
[
0.1748046875,
0.375,
0.826171875,
0.375,
0.826171875,
0.38818359375,
0.1748046875,
0.38818359375
],
[
0.1732177734375,
0.386962890625,
0.82421875,
0.386962890625,
0.82421875,
0.40087890625,
0.1732177734375,
0.40087890625
],
[
0.1748046875,
0.401611328125,
0.826171875,
0.401611328125,
0.826171875,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.176513671875,
0.417236328125,
0.701171875,
0.417236328125,
0.701171875,
0.427978515625,
0.176513671875,
0.427978515625
],
[
0.6953125,
0.416748046875,
0.82421875,
0.415283203125,
0.82421875,
0.4287109375,
0.6953125,
0.429931640625
],
[
0.176513671875,
0.43115234375,
0.82275390625,
0.43115234375,
0.82275390625,
0.44140625,
0.176513671875,
0.44140625
],
[
0.1748046875,
0.443115234375,
0.82421875,
0.443115234375,
0.82421875,
0.45703125,
0.1748046875,
0.45703125
],
[
0.1748046875,
0.45703125,
0.82421875,
0.45703125,
0.82421875,
0.47021484375,
0.1748046875,
0.47021484375
],
[
0.176513671875,
0.472900390625,
0.82275390625,
0.472900390625,
0.82275390625,
0.48291015625,
0.176513671875,
0.48291015625
],
[
0.1748046875,
0.484130859375,
0.74853515625,
0.484130859375,
0.74853515625,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.17724609375,
0.52294921875,
0.191162109375,
0.52294921875,
0.191162109375,
0.53271484375,
0.17724609375,
0.53271484375
],
[
0.200927734375,
0.52099609375,
0.381591796875,
0.521484375,
0.381591796875,
0.5361328125,
0.200927734375,
0.53515625
],
[
0.1732177734375,
0.5537109375,
0.826171875,
0.55322265625,
0.826171875,
0.5673828125,
0.1732177734375,
0.568359375
],
[
0.1748046875,
0.568359375,
0.82421875,
0.568359375,
0.82421875,
0.58203125,
0.1748046875,
0.58203125
],
[
0.1748046875,
0.58203125,
0.82421875,
0.58203125,
0.82421875,
0.5966796875,
0.1748046875,
0.5966796875
],
[
0.1748046875,
0.5966796875,
0.82421875,
0.5966796875,
0.82421875,
0.611328125,
0.1748046875,
0.611328125
],
[
0.1748046875,
0.60986328125,
0.826171875,
0.60986328125,
0.826171875,
0.62353515625,
0.1748046875,
0.62353515625
],
[
0.1748046875,
0.62353515625,
0.70361328125,
0.62353515625,
0.70361328125,
0.63720703125,
0.1748046875,
0.63720703125
],
[
0.340576171875,
0.64794921875,
0.65673828125,
0.64794921875,
0.65673828125,
0.662109375,
0.340576171875,
0.662109375
],
[
0.80224609375,
0.6484375,
0.826171875,
0.6484375,
0.826171875,
0.66357421875,
0.80224609375,
0.66357421875
],
[
0.435546875,
0.6611328125,
0.491943359375,
0.6611328125,
0.491943359375,
0.671875,
0.435546875,
0.671875
],
[
0.1748046875,
0.6845703125,
0.82421875,
0.6845703125,
0.82421875,
0.69775390625,
0.1748046875,
0.69775390625
],
[
0.1748046875,
0.69970703125,
0.82421875,
0.69970703125,
0.82421875,
0.71337890625,
0.1748046875,
0.71337890625
],
[
0.1732177734375,
0.71533203125,
0.67236328125,
0.71533203125,
0.67236328125,
0.72900390625,
0.1732177734375,
0.72900390625
],
[
0.313720703125,
0.74560546875,
0.59619140625,
0.7470703125,
0.59619140625,
0.7646484375,
0.313720703125,
0.76318359375
],
[
0.5888671875,
0.74560546875,
0.6806640625,
0.74755859375,
0.6806640625,
0.7646484375,
0.5888671875,
0.76220703125
],
[
0.80224609375,
0.74853515625,
0.826171875,
0.74853515625,
0.826171875,
0.76220703125,
0.80224609375,
0.76220703125
],
[
0.59326171875,
0.76123046875,
0.60107421875,
0.76123046875,
0.60107421875,
0.76513671875,
0.59326171875,
0.76513671875
],
[
0.1748046875,
0.77880859375,
0.826171875,
0.77880859375,
0.826171875,
0.79345703125,
0.1748046875,
0.79345703125
],
[
0.1748046875,
0.79345703125,
0.82421875,
0.79345703125,
0.82421875,
0.806640625,
0.1748046875,
0.806640625
],
[
0.1748046875,
0.80810546875,
0.82421875,
0.80810546875,
0.82421875,
0.82275390625,
0.1748046875,
0.82275390625
],
[
0.1732177734375,
0.82275390625,
0.60693359375,
0.82275390625,
0.60693359375,
0.8359375,
0.1732177734375,
0.8359375
],
[
0.50341796875,
0.8515625,
0.5400390625,
0.84765625,
0.54345703125,
0.861328125,
0.50634765625,
0.86572265625
],
[
0.44287109375,
0.859375,
0.50244140625,
0.859375,
0.50244140625,
0.875,
0.44287109375,
0.875
],
[
0.54345703125,
0.859375,
0.60693359375,
0.859375,
0.60693359375,
0.875,
0.54345703125,
0.875
],
[
0.80224609375,
0.85986328125,
0.826171875,
0.85986328125,
0.826171875,
0.87353515625,
0.80224609375,
0.87353515625
],
[
0.370849609375,
0.87255859375,
0.43701171875,
0.87255859375,
0.43701171875,
0.88330078125,
0.370849609375,
0.88330078125
],
[
0.38818359375,
0.85986328125,
0.420654296875,
0.85986328125,
0.420654296875,
0.87255859375,
0.38818359375,
0.87255859375
],
[
0.5107421875,
0.87353515625,
0.51953125,
0.87255859375,
0.521484375,
0.87890625,
0.51123046875,
0.8798828125
],
[
0.521484375,
0.87548828125,
0.54736328125,
0.87548828125,
0.54736328125,
0.88818359375,
0.521484375,
0.88818359375
],
[
0.17236328125,
0.896484375,
0.81396484375,
0.89794921875,
0.81396484375,
0.912109375,
0.17236328125,
0.9111328125
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"1.1",
"Related WorkonFunctional GradientDescent",
"A related functional gradient descent type method is the Stein Variation Gradient Descent (SvGD)",
"method by Liu and Wang [2016]. SvGD considers the problem of minimizing the Kullback-Leibler",
"(KL) divergence between a variable distribution and a posterior p. Note that SvGD updates the",
"positionsofasetofNparticlesusingthescorefunctionoftheposteriorp,i.e.logp.Consequently",
"it requires the access to the target distribution function. Later, Liu [2017] prove that SvGD has",
"convergence guarantee in its continuous-time limit (taking infinitesimal step size) using infinite",
"number of particles (N → oo). In comparison, sD is designed to solve the significantly more",
"complicated Sinkhon barycenter problem and has a stronger convergence guarantee. More precisely,",
"while sD updates the measure using only a sampling machinery of the target measures (no score",
"functions), it is guaranteed to converge sub-linearly to a stationary point when Q is a discrete measure",
"using discrete time steps. This is in sharp contrast to the results for SVGD.",
"In another work, Mroueh et al. [2019] considers minimizing the Maximum Mean Discrepancy",
"(MMD) between a source measure and a variable measure. They solve this problem by incrementally",
"following a Sobolev critic function and propose the Sobolev Descent (SoD) method. To show the",
"global convergence of the measure sequence generated by SoD, Mroueh et al. [2019] assumes the",
"entire sequence satisfies certain spectral properties, which is in general difficult to verify. Later,",
"Arbel et al. [2019] consider the same MMD minimization problem from a gradient flow perspective.",
"Theyproposetwoassumptionsthatifeitheroneholds,theMMDgradientflowconvergestothe",
"global solution. However, similar to [Mroueh et al., 2019], these assumptions have to be satisfied",
"for the entire measure sequence. We note that the Sinkhorn barycenter is a strict generalization of",
"the above MMD minimization problem and is hence much more challenging: By setting the number",
"of sourcemeasuresn =1 andsettingtheentropyregularizationparameter",
"=8o,problem(4)",
"degeneratestothespecialcaseofMMD.Further,theMMDbetweentwoprobabilitymeasureshasa",
"closed form expression while the Sinkhorn Divergence can only be described via a set of optimization",
"problems. Consequently, the Sinkhorn barycenter is significantly more challenging. To guarantee",
"globalconvergence,theproposedsDalgorithmonlyrequiresoneofaccumulationpointsofthe",
"measure sequence to be fully supported on X with no restriction on the entire sequence.",
"2",
"Sinkhorn Barycenter",
"We first introduce the entropy-regularized optimal transport distance and its debiased version, a.k.a.",
"the Sinkhorn divergence. Given two probability measures α, β E M(X), use II(α, β) to denote the",
"set of joint distributions over 2 with marginals Q and β. For T E II, use (c, ) to denote the integral",
"(c, ) = Jx2 c(α, y)d(α, y) and use KL(lα @ β) to denote the Kullback-Leibler divergence",
"between the candidate transport plan and the product measure α β. The entropy-regularized",
"optimal transport distance OT(α,β) : M(α) × M(α)→R+ is defined as",
"OT(α,β)=min《c,元)+KL(元αβ)",
"(2)",
"TEII(α.B)",
"Here, > O is a regularization parameter. Note that OT(α, β) is not a valid metric as there exists",
"Q E M(X) such that OT(α,α) / 0 when + 0. To remove this bias, Peyre et al. [2019]",
"introduced the Sinkhorn divergence S(α, β) : M(X) × M+(α) -→ R+:",
"S(α,β):=OT,(α,β) -OT(α,α) -",
"OT(B,β),",
"(3)",
"",
"which is a debiased version of OT,(α, β). It is further proved that S(α, β) is nonnegative, bi-",
"convex and metrizes the convergence in law when the ground set X is compact and the metric c is",
"Lipschitz. Now given a set of probability measures (B; J\"-1, the Sinkhorn barycenter is the measure",
"α E M(X) that minimizes the average of Sinkhorn divergences",
"1n",
"(S(α):",
"S(a,Bi",
"(4)",
"αeM+(X)",
"min",
"n",
"i=1",
"We will next focus on the properties of OT since S, (α) is the linear combination of these terms.",
"3"
] |
{
"arxiv_id": "2007.10449",
"language": "en",
"page_idx": 2
}
|
||
2007.10449-en-3
|
2007.10449-en-3.jpg
|
<table>
<tbody>
<tr>
<th>$\overline{\mathbf{Al\xi}}$ </th>
<th>rithm1 Sinkhorn $\textbf{n}$Descent(SD)</th>
</tr>
<tr>
<td>$\textbf{t}$ $a_{2}=\frac{1}{2}$ $a_{2}=\frac{1}{2}$</td>
<td>$d$ $\textbf{lfor}$</td>
</tr>
</tbody>
</table>
The Dual Formulation of OT$_\gamma.$ As a convex program, the entropy-regularized optimal transport
problem OT$_\gamma(2)$ has a equivalent dual formulation, which is given as follows:
(5)
$$\mathrm{OT}_{\gamma}(\alpha,\beta)=\max_{f,g\in\mathcal{C}(\mathcal{X})}\langle f,\alpha\rangle+\langle g,\beta\rangle-\gamma\langle\exp((f\oplus g-c)/\gamma)-1,\alpha\otimes\beta\rangle,$$
where we denote $[f\oplus g](x,y)=f(x)+g(y).$ The maximizers $f_\alpha,\beta$ and $g_\alpha,\beta$ of (5) are called the Sinkhorn potentials of OT$_\gamma(\alpha,\beta).$ Define the Sinkhorn mapping $\mathcal{A}:\mathcal{C}(\mathcal{X})\times\mathcal{M}_1^+(\mathcal{X})\to\mathcal{C}(\mathcal{X})$ by
(6)
$$\mathcal{A}(f,\alpha)(y)=-\gamma\log\int_{\mathcal{X}}\exp\big((f(x)-c(x,y))/\gamma\big)\mathbf{d}\alpha(x).$$
The following lemma states the optimality condition for the Sinkhorn potentials $f_\alpha,\beta$ and $g_\alpha,\beta.$
Lemma 2.1 (Optimality Peyré et al.[2019]). The pair (f,g) are the Sinkhorn potentials of the
entropy-regularized optimal transport problem (5) if they satisfy
$$f=\mathcal A(g,\beta),\alpha-a.e.\quad and\quad g=\mathcal A(f,\alpha),\beta-a.e..$$
(7)
The Sinkhorn potential is the cornerstone of the entropy regularized OT problem. In the discrete case it can be computed by a standard method in Genevay et al. [2016]. In particular, when $\alpha$ is discrete, $f$ can be simply represented by a finite dimensional vector since only its values on supp$(\alpha)$ matter. We describe such method in Appendix A.1 for completeness. In the following, we treat the computation of Sinkhorn potentials as a blackbox, and refer to it as $\mathcal{S}\mathcal{P}_\gamma(\alpha,\beta).$
# 3 Methodology
We present the Sinkhorn Descent (SD) algorithm for the Sinkhorn barycenter problem (4) in two steps: We first reformulate (4) as an unconstrained functional minimization problem and then derive the descent direction as the negative functional gradient over a RKHS $\mathcal{H}^d.$ Operating in RKHS allows us to measure the quality of the iterates using a so-called kernelized discrepancy which we introduce in Defnition 4.1. This quantity will be crucial for our convergence analysis. The restriction of a functional optimization problem to RKHS is common in the literature as discussed in Remark 3.1. Alternative Formulation. Instead of directly solving the Sinkhorn barycenter problem in the probability space $\mathcal{M}_1^+(\mathcal{X})$, we reformulate it as a functional minimization over all mappings on $\mathcal{X}:$
(8)
$$\min_{\mathcal{P}}\Big(\mathcal{S}_{\gamma}(\mathcal{P}_{\sharp}\alpha_{0}):=\dfrac{1}{n}\sum_{i=1}^{n}\mathbb{S}_{\gamma}(\mathcal{P}_{\sharp}\alpha_{0},\beta_{i})\Big),$$
where $\alpha_0\in\mathcal{M}_1^+(\mathcal{X})$ is some given initial measure, and $\mathcal{P}_\sharp\alpha$ is the push-forward measure of $\alpha\in\mathcal{M}_1^+(\mathcal{X})$ under the mapping $\mathcal{P}:\mathcal{X}\to\mathcal{X}.$ When $\alpha_0$ is sufficiently regular, e.g. absolutely continuous, for any $\alpha\in\mathcal{M}_1^+(\mathcal{X})$ there always exists a mapping $\mathcal{P}$ such that $\alpha=\mathcal{P}_\sharp\alpha_0$ (see Theorem 1.33 of[Ambrosio and Gigli, 2013]). Consequently, problems (8) and (4) are equivalent with appropriate initialization.
Algorithm Derivation. For a probability measure $\alpha$, define the functional $\mathcal{S}_\alpha:\mathcal{H}^d\to\mathbb{R}$
(9)
$$\mathcal{S}_{\alpha}[\psi]=\mathcal{S}_{\gamma}\big((\mathcal{I}+\psi)_{\sharp}\alpha\big),\psi\in\mathcal{H}^{d}.$$
Here$\mathcal{I}$ is the identity mapping and $\mathcal{S}_\gamma$ is defined in (4). Let $\alpha^t$ be the estimation of the Sinkhorn barycenter in the $t^{th}$ iteration. Sinkhorn Descent (SD) iteratively updates the measure $\alpha^{t+1}$ as
$$\alpha^{t+1}=\mathcal{T}[\alpha^t]_\sharp\alpha^t,$$
(10)
4
|
[
[
0.176513671875,
0.0921630859375,
0.43701171875,
0.0921630859375,
0.43701171875,
0.1060791015625,
0.176513671875,
0.1060791015625
],
[
0.188720703125,
0.111083984375,
0.82275390625,
0.10986328125,
0.82275390625,
0.1275634765625,
0.188720703125,
0.1287841796875
],
[
0.191162109375,
0.126953125,
0.708984375,
0.1287841796875,
0.708984375,
0.143310546875,
0.191162109375,
0.141357421875
],
[
0.191162109375,
0.1427001953125,
0.336669921875,
0.1427001953125,
0.336669921875,
0.1558837890625,
0.191162109375,
0.1558837890625
],
[
0.205078125,
0.154052734375,
0.51318359375,
0.154052734375,
0.51318359375,
0.1717529296875,
0.205078125,
0.1717529296875
],
[
0.1895751953125,
0.169189453125,
0.24267578125,
0.169189453125,
0.24267578125,
0.184326171875,
0.1895751953125,
0.184326171875
],
[
0.1748046875,
0.210205078125,
0.82421875,
0.2108154296875,
0.82421875,
0.2259521484375,
0.1748046875,
0.2247314453125
],
[
0.1748046875,
0.2259521484375,
0.69140625,
0.2259521484375,
0.69140625,
0.2392578125,
0.1748046875,
0.2392578125
],
[
0.2467041015625,
0.24365234375,
0.75244140625,
0.2442626953125,
0.75244140625,
0.261962890625,
0.2467041015625,
0.261474609375
],
[
0.80126953125,
0.2442626953125,
0.82666015625,
0.2442626953125,
0.82666015625,
0.261962890625,
0.80126953125,
0.261962890625
],
[
0.346435546875,
0.261474609375,
0.3896484375,
0.261474609375,
0.3896484375,
0.26513671875,
0.346435546875,
0.26513671875
],
[
0.1732177734375,
0.2734375,
0.82421875,
0.2734375,
0.82421875,
0.291748046875,
0.1732177734375,
0.291748046875
],
[
0.1732177734375,
0.291748046875,
0.8203125,
0.291748046875,
0.8203125,
0.30615234375,
0.1732177734375,
0.30615234375
],
[
0.306396484375,
0.32080078125,
0.449462890625,
0.32080078125,
0.449462890625,
0.335205078125,
0.306396484375,
0.335205078125
],
[
0.47216796875,
0.32080078125,
0.689453125,
0.32080078125,
0.689453125,
0.335205078125,
0.47216796875,
0.335205078125
],
[
0.80224609375,
0.32080078125,
0.826171875,
0.32080078125,
0.826171875,
0.335205078125,
0.80224609375,
0.335205078125
],
[
0.1707763671875,
0.3466796875,
0.7958984375,
0.347900390625,
0.7958984375,
0.365478515625,
0.1707763671875,
0.3642578125
],
[
0.1748046875,
0.367431640625,
0.82275390625,
0.367431640625,
0.82275390625,
0.3818359375,
0.1748046875,
0.3818359375
],
[
0.1748046875,
0.3818359375,
0.59228515625,
0.3818359375,
0.59228515625,
0.395751953125,
0.1748046875,
0.395751953125
],
[
0.32177734375,
0.401611328125,
0.67724609375,
0.401611328125,
0.67724609375,
0.416748046875,
0.32177734375,
0.416748046875
],
[
0.80224609375,
0.40283203125,
0.826171875,
0.40283203125,
0.826171875,
0.416748046875,
0.80224609375,
0.416748046875
],
[
0.1732177734375,
0.4287109375,
0.82421875,
0.4287109375,
0.82421875,
0.443115234375,
0.1732177734375,
0.443115234375
],
[
0.1732177734375,
0.443115234375,
0.81787109375,
0.4443359375,
0.81787109375,
0.458251953125,
0.1732177734375,
0.45703125
],
[
0.81396484375,
0.449462890625,
0.81787109375,
0.449462890625,
0.81787109375,
0.455078125,
0.81396484375,
0.455078125
],
[
0.1748046875,
0.458251953125,
0.82275390625,
0.458251953125,
0.82275390625,
0.4716796875,
0.1748046875,
0.4716796875
],
[
0.1748046875,
0.4716796875,
0.82275390625,
0.4716796875,
0.82275390625,
0.48486328125,
0.1748046875,
0.48486328125
],
[
0.1732177734375,
0.48291015625,
0.609375,
0.48486328125,
0.609375,
0.5,
0.1732177734375,
0.49755859375
],
[
0.1707763671875,
0.517578125,
0.314453125,
0.51953125,
0.314453125,
0.5341796875,
0.1707763671875,
0.5322265625
],
[
0.17236328125,
0.54736328125,
0.82275390625,
0.548828125,
0.82275390625,
0.56298828125,
0.17236328125,
0.5625
],
[
0.1748046875,
0.56298828125,
0.82421875,
0.56298828125,
0.82421875,
0.5771484375,
0.1748046875,
0.5771484375
],
[
0.1748046875,
0.5771484375,
0.82421875,
0.5771484375,
0.82421875,
0.59033203125,
0.1748046875,
0.59033203125
],
[
0.176513671875,
0.59228515625,
0.82177734375,
0.59228515625,
0.82177734375,
0.60302734375,
0.176513671875,
0.60302734375
],
[
0.1732177734375,
0.60400390625,
0.82177734375,
0.60400390625,
0.82177734375,
0.6171875,
0.1732177734375,
0.6171875
],
[
0.17236328125,
0.6171875,
0.81640625,
0.6171875,
0.81640625,
0.630859375,
0.17236328125,
0.630859375
],
[
0.176513671875,
0.646484375,
0.82275390625,
0.646484375,
0.82275390625,
0.6611328125,
0.176513671875,
0.6611328125
],
[
0.1748046875,
0.6611328125,
0.82177734375,
0.6611328125,
0.82177734375,
0.67529296875,
0.1748046875,
0.67529296875
],
[
0.365234375,
0.69140625,
0.48779296875,
0.69140625,
0.48779296875,
0.708984375,
0.365234375,
0.708984375
],
[
0.51953125,
0.69140625,
0.6298828125,
0.69140625,
0.6298828125,
0.708984375,
0.51953125,
0.708984375
],
[
0.80224609375,
0.693359375,
0.826171875,
0.693359375,
0.826171875,
0.70849609375,
0.80224609375,
0.70849609375
],
[
0.1748046875,
0.72802734375,
0.826171875,
0.72802734375,
0.826171875,
0.7421875,
0.1748046875,
0.7421875
],
[
0.1748046875,
0.744140625,
0.82275390625,
0.744140625,
0.82275390625,
0.75830078125,
0.1748046875,
0.75830078125
],
[
0.1732177734375,
0.75830078125,
0.82177734375,
0.75830078125,
0.82177734375,
0.77294921875,
0.1732177734375,
0.77294921875
],
[
0.1748046875,
0.7734375,
0.82421875,
0.7734375,
0.82421875,
0.78857421875,
0.1748046875,
0.78857421875
],
[
0.1748046875,
0.78857421875,
0.375732421875,
0.78857421875,
0.375732421875,
0.8017578125,
0.1748046875,
0.8017578125
],
[
0.1732177734375,
0.81298828125,
0.7685546875,
0.8125,
0.7685546875,
0.830078125,
0.1732177734375,
0.83056640625
],
[
0.381591796875,
0.8359375,
0.61376953125,
0.83544921875,
0.61376953125,
0.85302734375,
0.381591796875,
0.85400390625
],
[
0.8037109375,
0.83837890625,
0.826171875,
0.83837890625,
0.826171875,
0.85302734375,
0.8037109375,
0.85302734375
],
[
0.1732177734375,
0.861328125,
0.82177734375,
0.861328125,
0.82177734375,
0.87548828125,
0.1732177734375,
0.87548828125
],
[
0.1732177734375,
0.876953125,
0.81201171875,
0.87548828125,
0.81201171875,
0.89013671875,
0.1732177734375,
0.8916015625
],
[
0.43701171875,
0.89453125,
0.560546875,
0.896484375,
0.560546875,
0.9140625,
0.43701171875,
0.912109375
],
[
0.79345703125,
0.89794921875,
0.826171875,
0.89794921875,
0.826171875,
0.91357421875,
0.79345703125,
0.91357421875
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"Algorithm1Sinkhorn Descent (SD)",
"Input: measures [B:)n-1, a discrete initial measure 0, a step size n, and number of iterations S;",
"Output: A measure αs that approximates the Sinkhorn barycenter of (βB;Jn-1;",
"fort=0to S-1 do",
"αt+1 := T[αtl:αt, with T[αt] defined in (11);",
"end for",
"The Dual Formulation of OT.As a convex program, the entropy-regularized optimal transport",
"problem OT (2) has a equivalent dual formulation, which is given as follows:",
"OT(α,β) =max<f,α) +<g, β) -(exp((f @g -c)/) -1,α@β),",
"(5)",
"ToF",
"where we denote [f @ gl(a, y) = f(r) + g(y). The maximizers fa,β and ga,β of (5) are called the",
"Sinkhorn potentials of OT(α, β). Define the Sinkhorn mapping A : C(X) × M(X) → C(X) by",
"A(f,α)(y) = -log",
"exp ((f(r) -c(c, y)/)da(r)",
"(6)",
"The following lemma states the optimality condition for the Sinkhorn potentials fa,β and ga,β.",
"Lemma 2.1 (Optimality Peyre et al. [2019]). The pair (f, g) are the Sinkhorn potentials of the",
"entropy-regularizedoptimal transportproblem(5)if theysatisfy",
"f =A(g,β),α-a.e. and g=A(f,α),β-a.e..",
"(7)",
"The Sinkhorn potential is the cornerstone of the entropy regularized OT problem. In the discrete case,",
"it can be computed by a standard method in Genevay et al. [2016]. In particular, when Q is discrete, .",
"",
"can be simply represented by a finite dimensional vector since only its values on supp(α) matter.We",
"describe such method in Appendix A.1 for completeness. In the following, we treat the computation",
"of Sinkhorn potentials as a blackbox, and refer to it as SP(α, β).",
"3Methodology",
"We present the Sinkhorn Descent (SD) algorithm for the Sinkhorn barycenter problem (4) in two",
"steps: We first reformulate (4) as an unconstrained functional minimization problem and then derive",
"the descent direction as the negative functional gradient over a RKHS Hd. Operating in RKHS allows",
"ustomeasurethequality oftheiteratesusinga so-calledkernelized discrepancywhichweintroduce",
"in Definition 4.1. This quantity will be crucial for our convergence analysis. The restriction of a",
"functional optimization problem to RKHS is common in the literature as discussed in Remark 3.1.",
"Alternative Formulation.Instead of directly solving the Sinkhorn barycenter problem in the",
"probability space M (X), we reformulate it as a functional minimization over all mappings on X:",
"min (s(Ptao):=",
"Cs(Prao, B:))",
"(8)",
"where Qo E M(X) is some given initial measure, and PgQ is the push-forward measure of",
"Q E Mf(α) under the mapping P : X → X. When ao is sufficiently regular, e.g. absolutely",
"continuous, for any α E M(α) there always exists a mapping P such that α = Prao (see",
"Theorem 1.33 of [Ambrosio and Gigli, 2013]). Consequently, problems (8) and (4) are equivalent",
"with appropriate initialization.",
"Algorithm Derivation. For a probability measure α, define the functional Sα : Hd R",
"Sa[] = S (I +),a), E Hd",
"(9)",
"Here I is the identity mapping and S, is defined in (4). Let αt be the estimation of the Sinkhorn",
"barycenter in the tth iteration. Sinkhorn Descent (SD) iteratively updates the measure at+1 as",
"Qt+1 = T[αt]at,",
"(10)",
"4"
] |
{
"arxiv_id": "2007.10449",
"language": "en",
"page_idx": 3
}
|
||
2007.10449-en-4
|
2007.10449-en-4.jpg
|
via the push-forward mapping (with $\eta>0$ being a step-size)
$$\mathcal{T}[\alpha^t](x)=x-\eta\cdot D\mathcal{S}_{\alpha^t}[0](x).$$
(11)
Recall that $D\mathcal{S}_\alpha[0]$ is the Fréchet derivative of $\mathcal{S}_\alpha$ at $\psi=0($see Definition 1.1). Note that $(\mathcal{I}+\psi)_\sharp\alpha=$
$\alpha$ when $\psi=0.$ Öur choice of the negative Fréchet derivative in $\mathcal{T}[\alpha^t]$ allows the objective $\mathcal{S}_\gamma(\alpha)$ to
have the fastest descent at the current measure $\alpha=\alpha^t.$ We our line the details of SD in Algorithm 1.
Consequently, a solution of (8) will be found by finite-step compositions and then formally passing
to the limit $\tilde{\mathcal{P}}=\lim_{t\to\infty}\left(\mathcal{P}^{t}:=\mathcal{T}[\alpha^{t}]\circ\cdots\circ\tilde{\mathcal{T}}[\alpha^{0}]\right).$
Remark 3.1. We restrict ψ in (9) to the space H$^d$ to avoid the inherent difficulty when the perturbation
ofSinkhorn potentials introduced by the mapping $(\mathcal{I}+\psi)$ can no longer be properly bounded (for
$\psi\in\mathcal{H}^d,we$ always have the upper bound (1) which is necessary in our comvergence analysis). This
restriction will potentially introduce error to the minimization of (8). However, this restriction is a
$commonpracticeforgeneralfunctionaloptimizationproblems:BothSVGD[LiuandWang,2016]$
and SoD[Mroueh et al., 2019] explicitly make such RKHS restriction on their transport mappings.
$[Arbeletal.,2019]constructsthe$ transport mapping using the witness function of the Maximum
Mean Discrepancy (MMD) which also lies in an RKHS.
In what follows, we first derive a formula for the Fréchet derivative $DS_{\alpha^t}[0]$ (see (13)) and then
explain how it is efficiently computed. The proof of the next proposition requires additional continuity
study of the Sinkhorn potentials and is deferred to Appendix C.5.
Proposition 3.1. Recall the Fréchet derivative in Definition $l. l. Given$ $\alpha , \beta \in \mathcal{M} _1^+ ( \mathcal{X} ) , for$ $\psi \in \mathcal{H} ^{d}$
denote $F_{1}[ \psi ] = \mathrm{OT}_{\gamma }\left ( ( \mathcal{I} + \psi ) _{\sharp }\alpha , \beta \right )$ and $F_{2}[ \psi ] = \mathrm{OT}_{\gamma }\left ( ( \mathcal{I} + \psi ) _{\sharp }\alpha , ( \mathcal{I} + \psi ) _{\sharp }\alpha \right ) .$ Under Assump
tions 4.1 and 4.2 (described below), we can compute
$$DF_{1}[0](y)=\int_{\mathcal{X}}\nabla f_{\alpha,\beta}(x)k(x,y)\mathbf{d}\alpha(x),\quad DF_{2}[0](y)=2\int_{\mathcal{X}}\nabla f_{\alpha,\alpha}(x)k(x,y)\mathbf{d}\alpha(x),$$
where $\nabla f_{\alpha , \beta }$ and $\nabla f_{\alpha , \alpha }$ are the gradients $of$ the Sinkhorn potentials $of$ $\mathrm{OT}_{\gamma }( \alpha , \beta )$ and $\mathrm{OT}_{\gamma }( \alpha , \alpha )$
Consequently the Fréchet derivative of the Sinkhorn Barycenter problem (9) can be computed by
(13)
$$D\mathcal{S}_{\alpha}[0](y)=\int_{\mathcal{X}}\dfrac{1}{n}[\sum_{i=1}^{n}\nabla f_{\alpha,\beta_{i}}(x)-\nabla f_{\alpha,\alpha}(x)]k(x,y)\mathbf{d}\alpha(x).$$
This quantity can be computed efficiently when $\alpha$ is discrete: Consider an individual term $\nabla f_{\alpha,\beta}.$
${\textbf{Define }h(x,y):=\exp\left(\frac{1}{\gamma}(f_{\alpha,\beta}(x)+\mathcal{A}[f_{\alpha,\beta},\alpha](y)-c(x,y))\right).}$Lemma2.1 implies
$$\int h(x,y)\mathbf{d}\beta(y)=1.$$
Taking derivative with respect to $x$ on both sides and rearranging terms, we have
$$\nabla f_{\alpha,\beta}(x)=\frac{\int_{\mathcal{X}}h(x,y)\nabla_xc(x,y)\mathbf{d}\beta(y)}{\int h(x,y)\mathbf{d}\beta(y)}=\int_{\mathcal{X}}h(x,y)\nabla_xc(x,y)\mathbf{d}\beta(y),$$
(14)
which itself is an expectation. Note that to evaluate (13), we only need $\nabla f_\alpha,\beta(x)$ on supp$(\alpha).$ Using $\mathcal{SP}_{\gamma}(\alpha,\beta)$ (see the end of Section 2), the function value of $f_{\alpha,\beta}$ on supp$(\alpha)$ can be efficiently computed. Together with the expression in (14), the gradients $\nabla f_{\alpha,\beta}^{\prime\prime}(x)$ at $x\in\operatorname{supp}(\alpha)$ can also be $\dot{\text{obtained by a simple Monte-Carlo integration with respect to }\beta}.$
## 4 Analysis
In this section, we analyze the finite time convergence and the mean field limit of SD under the
following assumptions on the ground cost function $c$ and the kernel function $k$ of the RKHS $\mathcal{H}^d.$
Assumption 4.1. The ground cost function c$(x,y)$ is bounded, i.e. $\forall x,y\in\mathcal{X},c(x,y)\leq M_{c};G_{c}$- Lipschitz continuous, i.e. $\forall x,x^{\prime},y\in\mathcal{X},|c(x,y)-c(x^{\prime},y)|\leq G_{c}\|x-x^{\prime}\|;and~L_{c}-Lipschitz~smooth$, $i. e.$ $\forall x, x^{\prime }, y\in \mathcal{X} , \| \nabla _{1}c( x, y) - \nabla _{1}c( x^{\prime }, y) \| \leq L_{c}\| x- x^{\prime }\| .$
Assumption 4.2. The kernel function $k(x,y)$ is bounded, i.e. $\forall x,y\in\mathcal{X},k(x,y)\leq D_k;G_k$-
Lipschitz $continuous$, $i. e.$ $\forall x, \dot{x^{\prime }} , y\in \mathcal{X} , | \dot{k} ( x, y) - k( x^{\prime }, y) | \leq G_{c}\| x- x^{\prime }\| .$
5
|
[
[
0.1748046875,
0.0921630859375,
0.57177734375,
0.0921630859375,
0.57177734375,
0.106689453125,
0.1748046875,
0.106689453125
],
[
0.39208984375,
0.11236572265625,
0.6044921875,
0.11236572265625,
0.6044921875,
0.126953125,
0.39208984375,
0.126953125
],
[
0.79345703125,
0.11236572265625,
0.82666015625,
0.11236572265625,
0.82666015625,
0.1275634765625,
0.79345703125,
0.1275634765625
],
[
0.1748046875,
0.1319580078125,
0.81787109375,
0.1319580078125,
0.81787109375,
0.1458740234375,
0.1748046875,
0.1458740234375
],
[
0.176513671875,
0.146484375,
0.82275390625,
0.146484375,
0.82275390625,
0.160400390625,
0.176513671875,
0.160400390625
],
[
0.1748046875,
0.160400390625,
0.491943359375,
0.160400390625,
0.491943359375,
0.173583984375,
0.1748046875,
0.173583984375
],
[
0.494384765625,
0.160400390625,
0.826171875,
0.160400390625,
0.826171875,
0.173583984375,
0.494384765625,
0.173583984375
],
[
0.1748046875,
0.1749267578125,
0.82275390625,
0.1749267578125,
0.82275390625,
0.1881103515625,
0.1748046875,
0.1881103515625
],
[
0.1732177734375,
0.1881103515625,
0.52685546875,
0.1881103515625,
0.52685546875,
0.20263671875,
0.1732177734375,
0.20263671875
],
[
0.176513671875,
0.2083740234375,
0.82275390625,
0.2083740234375,
0.82275390625,
0.2215576171875,
0.176513671875,
0.2215576171875
],
[
0.1748046875,
0.222900390625,
0.82275390625,
0.222900390625,
0.82275390625,
0.236083984375,
0.1748046875,
0.236083984375
],
[
0.1748046875,
0.2374267578125,
0.82275390625,
0.2374267578125,
0.82275390625,
0.250732421875,
0.1748046875,
0.250732421875
],
[
0.1748046875,
0.250732421875,
0.82177734375,
0.250732421875,
0.82177734375,
0.264404296875,
0.1748046875,
0.264404296875
],
[
0.1748046875,
0.264404296875,
0.826171875,
0.264404296875,
0.826171875,
0.27783203125,
0.1748046875,
0.27783203125
],
[
0.1748046875,
0.27783203125,
0.826171875,
0.279052734375,
0.826171875,
0.292236328125,
0.1748046875,
0.291748046875
],
[
0.176513671875,
0.292236328125,
0.82421875,
0.292236328125,
0.82421875,
0.30615234375,
0.176513671875,
0.30615234375
],
[
0.1732177734375,
0.304931640625,
0.5400390625,
0.304931640625,
0.5400390625,
0.318115234375,
0.1732177734375,
0.318115234375
],
[
0.17236328125,
0.328857421875,
0.82275390625,
0.330078125,
0.82275390625,
0.3447265625,
0.17236328125,
0.343505859375
],
[
0.1748046875,
0.3447265625,
0.82275390625,
0.3447265625,
0.82275390625,
0.35791015625,
0.1748046875,
0.35791015625
],
[
0.1748046875,
0.35791015625,
0.60302734375,
0.35791015625,
0.60302734375,
0.371826171875,
0.1748046875,
0.371826171875
],
[
0.1732177734375,
0.375,
0.82177734375,
0.373779296875,
0.82177734375,
0.38818359375,
0.1732177734375,
0.389404296875
],
[
0.176513671875,
0.39013671875,
0.82421875,
0.39013671875,
0.82421875,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.17236328125,
0.40478515625,
0.521484375,
0.406005859375,
0.521484375,
0.42041015625,
0.17236328125,
0.419189453125
],
[
0.19775390625,
0.43310546875,
0.279296875,
0.43310546875,
0.279296875,
0.447509765625,
0.19775390625,
0.447509765625
],
[
0.303955078125,
0.43310546875,
0.5947265625,
0.43310546875,
0.5947265625,
0.447509765625,
0.303955078125,
0.447509765625
],
[
0.60546875,
0.429931640625,
0.82421875,
0.431884765625,
0.82421875,
0.449462890625,
0.60546875,
0.447509765625
],
[
0.1748046875,
0.460205078125,
0.82421875,
0.4609375,
0.82421875,
0.47607421875,
0.1748046875,
0.474853515625
],
[
0.1748046875,
0.47607421875,
0.54736328125,
0.47607421875,
0.54736328125,
0.490478515625,
0.1748046875,
0.490478515625
],
[
0.1732177734375,
0.498779296875,
0.80615234375,
0.5,
0.80615234375,
0.5146484375,
0.1732177734375,
0.51318359375
],
[
0.293212890625,
0.52880859375,
0.390625,
0.52880859375,
0.390625,
0.546875,
0.293212890625,
0.546875
],
[
0.423095703125,
0.53076171875,
0.7060546875,
0.53076171875,
0.7060546875,
0.54541015625,
0.423095703125,
0.54541015625
],
[
0.79345703125,
0.52978515625,
0.826171875,
0.52978515625,
0.826171875,
0.546875,
0.79345703125,
0.546875
],
[
0.431396484375,
0.548828125,
0.44775390625,
0.548828125,
0.44775390625,
0.55615234375,
0.431396484375,
0.55615234375
],
[
0.1732177734375,
0.56103515625,
0.82666015625,
0.5625,
0.82666015625,
0.580078125,
0.1732177734375,
0.5791015625
],
[
0.1707763671875,
0.580078125,
0.72802734375,
0.5791015625,
0.72802734375,
0.599609375,
0.1707763671875,
0.60107421875
],
[
0.439453125,
0.6142578125,
0.57177734375,
0.6142578125,
0.57177734375,
0.6318359375,
0.439453125,
0.6318359375
],
[
0.1748046875,
0.642578125,
0.701171875,
0.642578125,
0.701171875,
0.6572265625,
0.1748046875,
0.6572265625
],
[
0.3505859375,
0.662109375,
0.54150390625,
0.6611328125,
0.54150390625,
0.6787109375,
0.3505859375,
0.6796875
],
[
0.26318359375,
0.67236328125,
0.3447265625,
0.67236328125,
0.3447265625,
0.6875,
0.26318359375,
0.6875
],
[
0.5703125,
0.67236328125,
0.736328125,
0.67236328125,
0.736328125,
0.6875,
0.5703125,
0.6875
],
[
0.79345703125,
0.671875,
0.82666015625,
0.671875,
0.82666015625,
0.68798828125,
0.79345703125,
0.68798828125
],
[
0.3896484375,
0.681640625,
0.494384765625,
0.681640625,
0.494384765625,
0.6962890625,
0.3896484375,
0.6962890625
],
[
0.1748046875,
0.7021484375,
0.82275390625,
0.7021484375,
0.82275390625,
0.71533203125,
0.1748046875,
0.71533203125
],
[
0.1748046875,
0.71533203125,
0.82177734375,
0.71533203125,
0.82177734375,
0.72900390625,
0.1748046875,
0.72900390625
],
[
0.176513671875,
0.72900390625,
0.82275390625,
0.72900390625,
0.82275390625,
0.74365234375,
0.176513671875,
0.74365234375
],
[
0.1748046875,
0.74365234375,
0.5947265625,
0.74365234375,
0.5947265625,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1732177734375,
0.77783203125,
0.27783203125,
0.77783203125,
0.27783203125,
0.79248046875,
0.1732177734375,
0.79248046875
],
[
0.1748046875,
0.806640625,
0.82275390625,
0.806640625,
0.82275390625,
0.82080078125,
0.1748046875,
0.82080078125
],
[
0.17236328125,
0.82080078125,
0.80224609375,
0.8193359375,
0.80224609375,
0.833984375,
0.17236328125,
0.83544921875
],
[
0.1748046875,
0.8388671875,
0.82275390625,
0.8388671875,
0.82275390625,
0.85302734375,
0.1748046875,
0.85302734375
],
[
0.1748046875,
0.85302734375,
0.82666015625,
0.85302734375,
0.82666015625,
0.8662109375,
0.1748046875,
0.8662109375
],
[
0.1732177734375,
0.8671875,
0.57275390625,
0.8671875,
0.57275390625,
0.880859375,
0.1732177734375,
0.880859375
],
[
0.1732177734375,
0.88330078125,
0.826171875,
0.8837890625,
0.826171875,
0.8984375,
0.1732177734375,
0.89794921875
],
[
0.1732177734375,
0.8984375,
0.6708984375,
0.8984375,
0.6708984375,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"via the push-forward mapping (with n > 0 being a step-size)",
"T[αt]() = r -n· DSα[0](α)",
"(11)",
"Recall that DS[o] is the Fréchet derivative of S at y = O (see Definition 1.1). Note that (I+)Q =",
"α when b = O. Our choice of the negative Frechet derivative in T[αt] allows the objective S(α) to",
"have the fastest descent at the current measure α",
"= αt. We our line the details of sD in Algorithm 1.",
"Consequently, a solution of (8) will be found by finite-step compositions and then formally passing",
"to the limit P = limt-→ (Pt:=T[αt] o ... o T[α]).",
"Remark 3.1. We restrict in (9)to the space Hd to avoid the inherent difficulty when the perturbation",
"of Sinkhorn potentials introduced by the mapping (I+ ) can no longer be properly bounded (for",
" e Hd, we always have the upper bound (1) which is necessary in our convergence analysis). This",
"restriction will potentially introduce error to the minimization of (8). However, this restriction is a",
"common practice for general functional optimization problems:Both SvGD [Liu and Wang, 2016]",
"and SoD [Mroueh et al., 2019] explicitly make such RKHS restriction on their transport mappings.",
"[Arbel et al., 2019] constructs the transport mapping using the witness function of the Maximum",
"MeanDiscrepancy (MMD)whichalso lies inanRKHS.",
"In what follows, we first derive a formula for the Frechet derivative DSα [0] (see (13)) and then",
"explain how it is efficiently computed. The proof of the next proposition requires additional continuity",
"study of the Sinkhorn potentials and is deferred to Appendix C.5.",
"Proposition 3.1. Recall the Frechet derivative in Definition 1.1. Given Q, β M(),for b E Hd",
"denote Fi[l = OT((I +)α, β) and F2[] = OT(( + ),a, (I + )μa). Under Assump-",
"tions 4.1 and 4.2 (described below), we can compute",
"DFi[0]() :",
"fα,β(r)k(c, y)da(r), DF2[0](y) = 2",
"Vfa,a(r)k(r,y)da(a),(12)",
"wherefa.sand faaare thegradients oftheSinkhornpotentials ofOT,(a,β)andOT(Q,a)",
"respectively, and k is the kernel function of the RKHS H.",
"Consequently the Frechet derivative of the Sinkhorn Barycenter problem (9) can be computed by",
"DSa[0](y) =",
"[Vfa,B,(r) -fa,a(r)]k(, y)da(α).",
"(13)",
"-1",
"This quantity can be computed efficiently when α is discrete: Consider an individual term V fa,β.",
"Define h(a, ):=exp ((fa,B(a) + A[fa,B,al(u) -c(r, ). Lemma 2.1 implies",
"h(r, y)dp(y) = 1.",
"Taking derivative with respect to on both sides and rearranging terms, we have",
"Jx h(a, y)Vc(r, y)dB(y)",
"Vfa,β(r) =",
"h(c,y)V,c(r,y)dp(y),",
"(14)",
"J h(r, y)dβ(y)",
"which itself is an expectation. Note that to evaluate (13), we only need V fα.(r) on supp(α). Using",
"SP(α, β) (see the end of Section 2), the function value of fα,β on supp(α) can be efficiently",
"computed. Together with the expression in (14), the gradients V fα,B(α) at r E supp(α) can also be",
"obtained by a simple Monte-Carlo integration with respect to β.",
"4Analysis",
"In this section, we analyze the finite time convergence and the mean field limit of sD under the",
"following assumptions on the ground cost function c and the kernel function k of the RKHS Hd",
"Assumption 4.1. The ground cost function c(r, y) is bounded, i.e. Va, y E X,c(r, y) ≤ Me; Ge",
"Lipschitz continuous, i.e. Ve, r', y E X, Ic(, y) - c(α', y)l ≤ Gella -'ll; and Le-Lipschitz smooth,",
"i.e. Va, r', y E X, /Vic(a,y) - Vic(a',y)ll ≤ Lell -a'll.",
"Assumption 4.2. The kernel function k(,y) is bounded, i.e. Vr, y X,k(a,y) ≤ Dk; Gk-",
"Lipschitz continuous, i.e. Vr, a', y E X, [k(, y) -k(r', y)l ≤Gellr - r'll.",
"5"
] |
{
"arxiv_id": "2007.10449",
"language": "en",
"page_idx": 4
}
|
||
1605.07717-en-0
|
1605.07717-en-0.jpg
|
$\textbf{Deep Structured Energy Based Models for Anomaly Detection}$
Shuangfei Zhai$^{*\dagger}$
$YuCheng^{*\S}$
Weining Lu$^{\dagger}$
$\mathbf{Zhongfei~(Mark)~Zhang}^{\dagger}$
$^{\dagger}$Binghamton University, Vestal, NY 13902, USA.
$^{6}$IBM T. J. Watson Research Center, Yorktown Heights, NY 10598, USA.
$^{\ddagger}$Tsinghua University, Beijing 10084, China.
$^*$Equal contribution
SZHAI2@ BINGHAMTON.EDU
CHENGYU@ US.IBM.COM
LUWN14@ MAILS.TSINGHUA.EDU.CN ZHONGFEI@ CS. BINGHAMTON. EDU
## Abstract
In this paper, we attack the anomaly detection problem by directly modeling the data distribution with deep architectures. We propose deep structured energy based models (DSEBMs), where the energy function is the output of a deterministic deep neural network with structure. We develop novel model architectures to integrate EBMs with different types of data such as static data, sequential data, and spatial data, and apply appropriate model architectures to adapt to the data structure. Our training algorithm is built upon the recent development of score matching (Hyvärinen, 2005), which connects an EBM with a regularized autoencoder, eliminating the need for complicated sampling method. Statistically sound decision criterion can be derived for anomaly detection purpose from the perspective of the energy landscape of the data distribution. We investigate two decision criteria for performing anomaly detection: the energy score and the reconstruction error. Extensive empirical studies on benchmark tasks demonstrate that our proposed model consistently matches or outperforms all the competing methods.
## $1. \textbf{Introduction}$
Anomaly detection (also called novelty or outlier detection) is to identify patterns that do not conform to the expected normal patterns (Chandola et al., 2009). Existing methods for outlier detection either construct a profile for normal data examples and then identify the examples not conform-
$\overline{Proceedingsof}$ the 33$^rd$ International Conference on Machine $Learning$, New York, NY, USA, 2016. JMLR: W& volume 48. Copyright 2016 by the author(s).
ing to the normal profile as outliers, or explicitly isolate outliers based on statistical or geometric measures of abnormality. A variety of methods can be found in the survey (Zimek et al., 2012). Anomaly detection is to correctly characterize data distribution in nature, so the normality of the data characteristic can be characterized as a distribution and any future data can be benchmarked against the normality. Apparently, the statistical power and accuracy of the anomaly detection methods depend on the capacity of the model that is used to characterize the data distribution. Our work is inspired by the extraordinary capacity of deep models which are able to capture the complex distributions in real-world applications. Recent empirical and theoretical studies indicate that deep architectures are able to achieve better generalization ability compared to the shallow counterparts on challenging recognition tasks (Bengio, 2009). The key ingredient to the success of deep learning is its ability to learn multiple levels of representations with increasing abstraction. For example, it is shown that properly regularized autoencoders (Vincent et al., 2010; Rifai et al., 2011) are able to effectively characterize the data distribution and learn useful representations, which are not achieved by shallow methods such as PCA or K-Means.
However, deep models have not been systematically studied and developed for anomaly detection. Central questions critical to this development include: 1) how to effectively model the data generating distribution with a deep model? 2) how to generalize a model to a range of data structures such as static data, sequential data, and spatial data? 3) how to develop computationally efficient training algorithms to make them scalable? and 4) how to derive statistically sound decision criteria for anomaly detection purpose?
To answer these questions in a systematic manner, in this work, we propose deep structured energy based models (DSEBMs). Our approach falls into the category of en-
|
[
[
0.17724609375,
0.1142578125,
0.79736328125,
0.1142578125,
0.79736328125,
0.1287841796875,
0.17724609375,
0.1287841796875
],
[
0.0882568359375,
0.177978515625,
0.2091064453125,
0.177978515625,
0.2091064453125,
0.1925048828125,
0.0882568359375,
0.1925048828125
],
[
0.67236328125,
0.1787109375,
0.87353515625,
0.1787109375,
0.87353515625,
0.1925048828125,
0.67236328125,
0.1925048828125
],
[
0.08905029296875,
0.1932373046875,
0.1748046875,
0.1932373046875,
0.1748046875,
0.2083740234375,
0.08905029296875,
0.2083740234375
],
[
0.69921875,
0.1944580078125,
0.87158203125,
0.1944580078125,
0.87158203125,
0.2052001953125,
0.69921875,
0.2052001953125
],
[
0.08905029296875,
0.2083740234375,
0.178955078125,
0.2083740234375,
0.178955078125,
0.222900390625,
0.08905029296875,
0.222900390625
],
[
0.61376953125,
0.208984375,
0.87353515625,
0.208984375,
0.87353515625,
0.222900390625,
0.61376953125,
0.222900390625
],
[
0.08905029296875,
0.2235107421875,
0.2607421875,
0.2235107421875,
0.2607421875,
0.238037109375,
0.08905029296875,
0.238037109375
],
[
0.625,
0.2247314453125,
0.869140625,
0.2247314453125,
0.869140625,
0.238037109375,
0.625,
0.238037109375
],
[
0.08905029296875,
0.2412109375,
0.4150390625,
0.2412109375,
0.4150390625,
0.255615234375,
0.08905029296875,
0.255615234375
],
[
0.08905029296875,
0.2568359375,
0.5703125,
0.2568359375,
0.5703125,
0.270751953125,
0.08905029296875,
0.270751953125
],
[
0.0228729248046875,
0.266357421875,
0.058013916015625,
0.266357421875,
0.058013916015625,
0.330810546875,
0.0228729248046875,
0.330810546875
],
[
0.0882568359375,
0.270751953125,
0.381591796875,
0.271484375,
0.381591796875,
0.285888671875,
0.0882568359375,
0.285400390625
],
[
0.035125732421875,
0.27978515625,
0.047393798828125,
0.27978515625,
0.047393798828125,
0.2841796875,
0.035125732421875,
0.2841796875
],
[
0.08905029296875,
0.287353515625,
0.221435546875,
0.287353515625,
0.221435546875,
0.3017578125,
0.08905029296875,
0.3017578125
],
[
0.0236968994140625,
0.3193359375,
0.058013916015625,
0.3193359375,
0.058013916015625,
0.389404296875,
0.0236968994140625,
0.389404296875
],
[
0.24267578125,
0.3359375,
0.32177734375,
0.3359375,
0.32177734375,
0.354248046875,
0.24267578125,
0.354248046875
],
[
0.5009765625,
0.339111328125,
0.8857421875,
0.339111328125,
0.8857421875,
0.352783203125,
0.5009765625,
0.352783203125
],
[
0.12091064453125,
0.359130859375,
0.44287109375,
0.359130859375,
0.44287109375,
0.373779296875,
0.12091064453125,
0.373779296875
],
[
0.5009765625,
0.354736328125,
0.8857421875,
0.354736328125,
0.8857421875,
0.36865234375,
0.5009765625,
0.36865234375
],
[
0.1217041015625,
0.375,
0.441162109375,
0.375,
0.441162109375,
0.38818359375,
0.1217041015625,
0.38818359375
],
[
0.5009765625,
0.369384765625,
0.8837890625,
0.369384765625,
0.8837890625,
0.38330078125,
0.5009765625,
0.38330078125
],
[
0.041656494140625,
0.38330078125,
0.0482177734375,
0.38330078125,
0.0482177734375,
0.386962890625,
0.041656494140625,
0.386962890625
],
[
0.12091064453125,
0.39013671875,
0.3505859375,
0.39013671875,
0.3505859375,
0.404052734375,
0.12091064453125,
0.404052734375
],
[
0.50244140625,
0.385009765625,
0.88232421875,
0.385009765625,
0.88232421875,
0.3984375,
0.50244140625,
0.3984375
],
[
0.343017578125,
0.391357421875,
0.439453125,
0.391357421875,
0.439453125,
0.40283203125,
0.343017578125,
0.40283203125
],
[
0.12091064453125,
0.40478515625,
0.441162109375,
0.40478515625,
0.441162109375,
0.419189453125,
0.12091064453125,
0.419189453125
],
[
0.50244140625,
0.401611328125,
0.8857421875,
0.401611328125,
0.8857421875,
0.412353515625,
0.50244140625,
0.412353515625
],
[
0.5009765625,
0.415283203125,
0.8857421875,
0.415283203125,
0.8857421875,
0.4287109375,
0.5009765625,
0.4287109375
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.448974609375,
0.02532958984375,
0.448974609375
],
[
0.1217041015625,
0.42041015625,
0.441162109375,
0.42041015625,
0.441162109375,
0.434326171875,
0.1217041015625,
0.434326171875
],
[
0.1217041015625,
0.43505859375,
0.439453125,
0.43505859375,
0.439453125,
0.448974609375,
0.1217041015625,
0.448974609375
],
[
0.5,
0.429931640625,
0.8857421875,
0.429931640625,
0.8857421875,
0.443115234375,
0.5,
0.443115234375
],
[
0.02777099609375,
0.44140625,
0.05963134765625,
0.44140625,
0.05963134765625,
0.484130859375,
0.02777099609375,
0.484130859375
],
[
0.1217041015625,
0.45068359375,
0.441162109375,
0.45068359375,
0.441162109375,
0.464111328125,
0.1217041015625,
0.464111328125
],
[
0.5009765625,
0.44580078125,
0.88623046875,
0.44580078125,
0.88623046875,
0.458984375,
0.5009765625,
0.458984375
],
[
0.1217041015625,
0.46533203125,
0.441162109375,
0.46533203125,
0.441162109375,
0.479248046875,
0.1217041015625,
0.479248046875
],
[
0.5009765625,
0.460205078125,
0.88623046875,
0.460205078125,
0.88623046875,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.12091064453125,
0.48095703125,
0.441162109375,
0.48095703125,
0.441162109375,
0.494384765625,
0.12091064453125,
0.494384765625
],
[
0.5009765625,
0.47607421875,
0.88134765625,
0.47607421875,
0.88134765625,
0.4892578125,
0.5009765625,
0.4892578125
],
[
0.0196075439453125,
0.490478515625,
0.058013916015625,
0.490478515625,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.12091064453125,
0.49560546875,
0.439453125,
0.49560546875,
0.439453125,
0.5087890625,
0.12091064453125,
0.5087890625
],
[
0.498291015625,
0.496826171875,
0.88623046875,
0.49755859375,
0.88623046875,
0.51220703125,
0.498291015625,
0.51123046875
],
[
0.1217041015625,
0.51220703125,
0.441162109375,
0.51220703125,
0.441162109375,
0.52294921875,
0.1217041015625,
0.52294921875
],
[
0.5,
0.51318359375,
0.88623046875,
0.51318359375,
0.88623046875,
0.5263671875,
0.5,
0.5263671875
],
[
0.1217041015625,
0.52587890625,
0.441162109375,
0.52587890625,
0.441162109375,
0.5390625,
0.1217041015625,
0.5390625
],
[
0.5009765625,
0.52880859375,
0.88623046875,
0.52880859375,
0.88623046875,
0.54248046875,
0.5009765625,
0.54248046875
],
[
0.12091064453125,
0.54052734375,
0.441162109375,
0.54052734375,
0.441162109375,
0.5537109375,
0.12091064453125,
0.5537109375
],
[
0.5009765625,
0.54345703125,
0.8857421875,
0.54345703125,
0.8857421875,
0.556640625,
0.5009765625,
0.556640625
],
[
0.1192626953125,
0.5546875,
0.441162109375,
0.5546875,
0.441162109375,
0.568359375,
0.1192626953125,
0.568359375
],
[
0.50244140625,
0.56005859375,
0.8837890625,
0.56005859375,
0.8837890625,
0.57080078125,
0.50244140625,
0.57080078125
],
[
0.12091064453125,
0.57080078125,
0.441162109375,
0.57080078125,
0.441162109375,
0.58544921875,
0.12091064453125,
0.58544921875
],
[
0.5,
0.57373046875,
0.88232421875,
0.57373046875,
0.88232421875,
0.5869140625,
0.5,
0.5869140625
],
[
0.12091064453125,
0.5859375,
0.441162109375,
0.5859375,
0.441162109375,
0.599609375,
0.12091064453125,
0.599609375
],
[
0.5009765625,
0.5888671875,
0.8857421875,
0.5888671875,
0.8857421875,
0.60302734375,
0.5009765625,
0.60302734375
],
[
0.1217041015625,
0.6015625,
0.439453125,
0.6015625,
0.439453125,
0.61572265625,
0.1217041015625,
0.61572265625
],
[
0.50244140625,
0.60498046875,
0.88232421875,
0.60498046875,
0.88232421875,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.12091064453125,
0.6162109375,
0.441162109375,
0.61572265625,
0.441162109375,
0.6298828125,
0.12091064453125,
0.630859375
],
[
0.5009765625,
0.61865234375,
0.88232421875,
0.61865234375,
0.88232421875,
0.6318359375,
0.5009765625,
0.6318359375
],
[
0.1192626953125,
0.6298828125,
0.441162109375,
0.630859375,
0.441162109375,
0.64599609375,
0.1192626953125,
0.64453125
],
[
0.5,
0.63330078125,
0.8857421875,
0.6318359375,
0.8857421875,
0.646484375,
0.5,
0.64794921875
],
[
0.12091064453125,
0.64599609375,
0.439453125,
0.64599609375,
0.439453125,
0.6591796875,
0.12091064453125,
0.6591796875
],
[
0.5009765625,
0.6484375,
0.8857421875,
0.6484375,
0.8857421875,
0.662109375,
0.5009765625,
0.662109375
],
[
0.1217041015625,
0.662109375,
0.441162109375,
0.662109375,
0.441162109375,
0.67529296875,
0.1217041015625,
0.67529296875
],
[
0.5,
0.66357421875,
0.8857421875,
0.66357421875,
0.8857421875,
0.6767578125,
0.5,
0.6767578125
],
[
0.12091064453125,
0.6767578125,
0.441162109375,
0.6767578125,
0.441162109375,
0.68994140625,
0.12091064453125,
0.68994140625
],
[
0.5009765625,
0.68115234375,
0.865234375,
0.68115234375,
0.865234375,
0.69140625,
0.5009765625,
0.69140625
],
[
0.1217041015625,
0.68994140625,
0.439453125,
0.68994140625,
0.439453125,
0.705078125,
0.1217041015625,
0.705078125
],
[
0.12091064453125,
0.70703125,
0.34228515625,
0.70703125,
0.34228515625,
0.72021484375,
0.12091064453125,
0.72021484375
],
[
0.5009765625,
0.7021484375,
0.8837890625,
0.7021484375,
0.8837890625,
0.71533203125,
0.5009765625,
0.71533203125
],
[
0.5009765625,
0.71728515625,
0.88232421875,
0.71728515625,
0.88232421875,
0.73095703125,
0.5009765625,
0.73095703125
],
[
0.50244140625,
0.734375,
0.88232421875,
0.734375,
0.88232421875,
0.744140625,
0.50244140625,
0.744140625
],
[
0.08905029296875,
0.75,
0.2156982421875,
0.75,
0.2156982421875,
0.7646484375,
0.08905029296875,
0.7646484375
],
[
0.5009765625,
0.74755859375,
0.8857421875,
0.74755859375,
0.8857421875,
0.76123046875,
0.5009765625,
0.76123046875
],
[
0.5009765625,
0.76220703125,
0.8857421875,
0.76220703125,
0.8857421875,
0.77587890625,
0.5009765625,
0.77587890625
],
[
0.08905029296875,
0.77587890625,
0.473876953125,
0.77587890625,
0.473876953125,
0.7890625,
0.08905029296875,
0.7890625
],
[
0.5009765625,
0.7763671875,
0.8857421875,
0.7763671875,
0.8857421875,
0.79052734375,
0.5009765625,
0.79052734375
],
[
0.0882568359375,
0.79150390625,
0.473876953125,
0.79150390625,
0.473876953125,
0.8046875,
0.0882568359375,
0.8046875
],
[
0.5,
0.79248046875,
0.8857421875,
0.79248046875,
0.8857421875,
0.80615234375,
0.5,
0.80615234375
],
[
0.0882568359375,
0.80615234375,
0.4755859375,
0.80615234375,
0.4755859375,
0.8193359375,
0.0882568359375,
0.8193359375
],
[
0.5,
0.806640625,
0.88623046875,
0.80615234375,
0.88623046875,
0.82080078125,
0.5,
0.8212890625
],
[
0.0882568359375,
0.82080078125,
0.4755859375,
0.82080078125,
0.4755859375,
0.83544921875,
0.0882568359375,
0.83544921875
],
[
0.5,
0.82275390625,
0.88623046875,
0.82275390625,
0.88623046875,
0.8359375,
0.5,
0.8359375
],
[
0.08905029296875,
0.8359375,
0.473876953125,
0.8359375,
0.473876953125,
0.849609375,
0.08905029296875,
0.849609375
],
[
0.5,
0.83837890625,
0.56396484375,
0.83837890625,
0.56396484375,
0.85302734375,
0.5,
0.85302734375
],
[
0.0882568359375,
0.859375,
0.47802734375,
0.859375,
0.47802734375,
0.87353515625,
0.0882568359375,
0.87353515625
],
[
0.498291015625,
0.85986328125,
0.8857421875,
0.85986328125,
0.8857421875,
0.87353515625,
0.498291015625,
0.87353515625
],
[
0.0882568359375,
0.87353515625,
0.4794921875,
0.87353515625,
0.4794921875,
0.88720703125,
0.0882568359375,
0.88720703125
],
[
0.5,
0.87548828125,
0.8857421875,
0.87548828125,
0.8857421875,
0.8896484375,
0.5,
0.8896484375
],
[
0.0882568359375,
0.88623046875,
0.30810546875,
0.88623046875,
0.30810546875,
0.8994140625,
0.0882568359375,
0.8994140625
],
[
0.5,
0.89013671875,
0.8837890625,
0.89013671875,
0.8837890625,
0.90478515625,
0.5,
0.90478515625
]
] |
[
"Deep Structured Energy Based Models for Anomaly Detection",
"Shuangfei Zhai*",
"SZHAI2 @BINGHAMTON.EDU",
"Yu Cheng*$",
"CHENGYU@US.IBM.COM",
"Weining Lut",
"LUWN14@MAILS.TSINGHUA.EDU.CN",
"Zhongfei (Mark) Zhang",
"ZHONGFEI@CS.BINGHAMTON.EDU",
"tBinghamton University, Vestal, NY 13902, USA",
"$IBM T. J. Watson Research Center, Yorktown Heights, NY 10598, USA.",
"910",
"+Tsinghua University, Beijing 10084, China.",
"",
"*Equal contribution",
"unr91",
"Abstract",
"ing to the normal profile as outliers, or explicitly isolate",
"In this paper, we attack the anomaly detection",
"outliers based on statistical or geometric measures of ab-",
"problem by directly modeling the data distri-",
"normality. A variety of methods can be found in the sur-",
"",
"bution with deep architectures.",
"vey (Zimek et al., 2012). Anomaly detection is to correctly",
"Wepropose",
"deep structured energy based models (DSEBMs),",
"characterizedata distribution innature,sothe normality of",
"the data characteristic can be characterized as a distribution",
"[DT",
"where the energy function is the output of a de-",
"terministic deep neural network with structure.",
"and any future data can be benchmarked against the nor-",
"I'so",
"We develop novel model architectures to inte-",
"mality.Apparently, the statistical power and accuracy of",
"grate EBMs with different types of data such as",
"the anomaly detection methods depend on the capacity of",
"static data, sequential data, and spatial data, and",
"the model that is used to characterize the data distribution.",
"ZA1LLOS091AI",
"apply appropriate model architectures to adapt to",
"Our work is inspired by the extraordinary capacity of deep",
"the data structure. Our training algorithm is built",
"models which are able to capture the complex distributions",
"upon the recent development of score match-",
"in real-world applications. Recent empirical and theoretical",
"ing (Hyvarinen,2005), which connects an EBM",
"studies indicate that deep architectures are able to achieve",
"with a regularized autoencoder, eliminating the",
"bettergeneralizationabilitycomparedtotheshallow coun-",
"need for complicated sampling method. Statis-",
"terparts on challenging recognition tasks (Bengio, 2009)",
"tically sound decision criterion can be derived",
"The key ingredient to the success of deep learning is its",
"for anomaly detection purpose from the perspec-",
"abilitytolearn multiple levelsof representations within-",
"tive of the energy landscape of the data distri-",
"creasing abstraction. For example, it is shown that prop-",
"bution.We investigate two decision criteria for",
"erly regularized autoencoders (Vincent et al., 2010; Rifai",
"performing anomaly detection: the energy score",
"et al., 2011) are able to effectively characterize the data",
"and the reconstruction error. Extensive empirical",
"distribution and learn useful representations, which are not",
"studies on benchmark tasks demonstrate that our",
"achievedbyshallowmethods suchasPCAorK-Means.",
"proposed model consistently matches or outper-",
"forms all the competing methods.",
"However, deep models have not been systematically stud-",
"ied and developed for anomaly detection.Central ques-",
"tionscriticaltothisdevelopmentinclude:1)howtoeffec-",
"1. Introduction",
"tively model the data generating distribution with a deep",
"model? 2) how to generalize a model to a range of data",
"Anomaly detection (also called novelty or outlier detection)",
"structures such as static data, sequential data, and spatial",
"is to identify patterns that do not conform to the expected",
"data? 3) how to develop computationally efficient training",
"normal patterns (Chandola et al., 2009). Existing methods",
"algorithms to make them scalable? and 4) how to derive",
"for outlier detection either construct a profile for normal",
"statistically sound decision criteria for anomaly detection",
"data examples and then identify the examples not conform-",
"purpose?",
"Proceedings of the 3ord International Conference on Machine",
"To answer these questions in a systematic manner, in this",
"Learning,New York, NY, USA, 2016. JMLR: W&CP volume",
"work, we propose deep structured energy based models",
"48. Copyright 2016 by the author(s).",
"(DSEBMs). Our approach falls into the category of en-"
] |
{
"arxiv_id": "1605.07717",
"language": "en",
"page_idx": 0
}
|
||
1605.07717-en-2
|
1605.07717-en-2.jpg
|
Deep Structured Energy Based Models for Anomaly Detection
where $\epsilon\sim\mathcal{N}(0,\sigma^{2}I)$ is an isotropic Gaussian noise. DAEs are easy to train with standard stochastic gradient descent (SGD) and perform significantly better than unregularized autoencoders.
While RBM and DAEs are typically considered as two alternative unsupervised deep models, it is recently shown they are closely related to each other. In particular, (Vincent, 2011) shows that training an RBM with score matching (SM) (Hyvärinen, 2005) is equivalent to a one-layer DAE. SM is an alternative method to MLE, which is especially suitable for estimating non-normalized density functions such as EBM. Instead of trying to directly maximize the probability of training instances, SM minimizes the following objective function:
$$J(\theta)=\dfrac{1}{2}\int_xp_x(x)\|\psi(x;\theta)-\psi_x(x)\|_2^2dx,$$
(5)
where $p_x(x)$ is the true data distribution which is unknown; $\psi ( x; \theta )$ = $\nabla _{x}\log p( x; \theta )$ = $- \nabla _{x}E( x; \theta )$ and$\psi _x( x)$ = $\nabla_{x}\log p_{x}(x)$ are the score function of the model and the true density function, respectively. In words, J(θ) measures the expected distance of the scores between the model density and the true density. (Vincent, 2011) shows that by $\underset{N}{\operatorname*{\operatorname*{approximating~the~}}}p_{x}(x)$ with the Parzen window density $\frac{1}{N}\sum_{i=1}^{N}\mathcal{N}(x_{i},\sigma^{2}\mathbf{I})$, minimizing Equation 5 yields an objective function in the same form as that of an autoencoder in Equation 4, with a reconstruction function defined as:
(6)
$$f(x;\theta)=x-\nabla_xE(x;\theta).$$
Substituting $E(x;\theta)$ with Equation 2 yields a typical one-
layer DAE (up to a constant factor):
$$f(x;\theta)=W\sigma(W^Tx+b)+b^{\prime},$$
(7)
where $\sigma(x)$ is the sigmoid function $\frac1{1+e^{-x}}.$
Equation 6 plays a key role in this work, as it allows one to efficiently train an arbitrary EBM in a similar way to that of a DAE, as long as the energy function is differentiable w.r.t. $x$ and $\theta.$ As will be demonstrated below, this makes training a deep EBM with various underlying structures in an end-to-end fashion, without the need of resorting to sophisticated sampling procedures (Ngiam et al., 2011) or layer wise pretraining (Hinton, 2010).
# 3. Deep Structured EBMs
Deep architectures allow one to model complicated patterns efficiently, which makes it especially suitable for high dimensional data. On the other hand, it is often necessary to adapt the architecture of a deep model to the structure of data. For example, recurrent neural networks (RNNs) have been shown to work very well at modeling sequential data;
convolutional neural networks (CNNs) are very effective at modeling data with spatial structure. In this work, we extend EBMs further to deep structured EBMs (DSEBMs), where we allow the underlying deep neural network encoding the energy function to take architectures varying from fully connected, recurrent, and convolutional. This generalizes EBMs (LeCun et al., 2006) as well as deep EBMs (Ngiam et al., 2011) as it makes our model applicable to a much wider spectrum of applications, including static data, sequential data and spatial data. Moreover, we show that by deriving the proper reconstruction functions with Equation 6, DSEBMs can be easily trained with SGD, regardless of the type of the underlying architecture. In the following, we will elaborate the formulation of DSEBMs for the three cases.
## $3. 1. \textbf{Fully Connected EBMs}$
This case is conceptually the same as the deep EBMs proposed in (Ngiam et al., 2011). Without loss of generality, we express the energy function of an L-layer fully connected EBM as:
(8)
$$E(x;\theta)=\frac{1}{2}\|x-b^{\prime}\|_{2}^{2}-\sum_{j=1}^{K_{L}}h_{L,j}\\s.t.\:h_{l}=g(W_{l}^{T}h_{l-1}+b_{1}),\:l\in[1,L]$$
where $W_l\in R^{K_{l-1}\times K_l},b_l\in R^{K_l}$ are the parameters for the $l$th layer; $K_l$ is the dimensionality of the $l$th layer. The 0th layer is defined as the input itself; and thus we have $K_{0}$ = $d, h_{0}$ = $x.$ We have explicitly included the term $\|x-b^{\prime}\|_2^2$ which acts as a prior, punishing the probability of the inputs that are far away from $b^\prime\in\bar{R^d}.$
Following the chain rule of gradient computation, one can
derive the reconstruction function as follows:
(9)
$$\begin{aligned}&f(x;\theta)=x-\nabla_{x}E(x;\theta)=h_{0}^{\prime}+b^{\prime}\\&h_{l-1}^{\prime}=\frac{\partial h_{l}}{\partial h_{l-1}}h_{l}^{\prime}=\sigma(W_{l}^{T}h_{l-1}+b_{l})\cdot(W_{l}h_{l}^{\prime}),\\&for\:l\in[1,L-1],\end{aligned}$$
where $h_L^{\prime}=1$ with $1\in R^K_L$ denoting a column vector of all ones; - is the element-wise product between vectors. One can then plug in the resulting $f(x;\theta)$ into Equation 4 and train it as a regular L-layer DAE.
## 3.2. Recurrent EBMs
Our formulation of recurrent EBMs is similar to (Boulanger-Lewandowski et al., 2012), where an EBM is built at each time step, with parameters determined by an underlying RNN. Formally, given a sequence of length $T$ $\mathbf{x}=[x^{1},...,x^{T}],x^{t}\in R^{d}$,we factorize the joint probability as $p(\mathbf{x})=\prod_{t=1}^{T}p(x^{t}|x^{1,\ldots,t-1})$ with the chain rule of
|
[
[
0.291748046875,
0.05682373046875,
0.68310546875,
0.05682373046875,
0.68310546875,
0.0706787109375,
0.291748046875,
0.0706787109375
],
[
0.0882568359375,
0.083984375,
0.4755859375,
0.085205078125,
0.4755859375,
0.0997314453125,
0.0882568359375,
0.0985107421875
],
[
0.5,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.0906982421875,
0.1029052734375,
0.47216796875,
0.1029052734375,
0.47216796875,
0.11297607421875,
0.0906982421875,
0.11297607421875
],
[
0.5,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.5,
0.11553955078125
],
[
0.08905029296875,
0.11614990234375,
0.473876953125,
0.11614990234375,
0.473876953125,
0.1300048828125,
0.08905029296875,
0.1300048828125
],
[
0.5,
0.11614990234375,
0.8837890625,
0.11614990234375,
0.8837890625,
0.1300048828125,
0.5,
0.1300048828125
],
[
0.08905029296875,
0.1331787109375,
0.1805419921875,
0.1331787109375,
0.1805419921875,
0.14453125,
0.08905029296875,
0.14453125
],
[
0.5009765625,
0.1319580078125,
0.88232421875,
0.1319580078125,
0.88232421875,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.5009765625,
0.146484375,
0.88623046875,
0.146484375,
0.88623046875,
0.1610107421875,
0.5009765625,
0.1610107421875
],
[
0.08905029296875,
0.154052734375,
0.473876953125,
0.154052734375,
0.473876953125,
0.1685791015625,
0.08905029296875,
0.1685791015625
],
[
0.5009765625,
0.1622314453125,
0.8857421875,
0.1622314453125,
0.8857421875,
0.175537109375,
0.5009765625,
0.175537109375
],
[
0.08905029296875,
0.169189453125,
0.473876953125,
0.169189453125,
0.473876953125,
0.18310546875,
0.08905029296875,
0.18310546875
],
[
0.5009765625,
0.1767578125,
0.8857421875,
0.1767578125,
0.8857421875,
0.1900634765625,
0.5009765625,
0.1900634765625
],
[
0.0882568359375,
0.184326171875,
0.473876953125,
0.184326171875,
0.473876953125,
0.1976318359375,
0.0882568359375,
0.1976318359375
],
[
0.5009765625,
0.1912841796875,
0.8857421875,
0.1912841796875,
0.8857421875,
0.2052001953125,
0.5009765625,
0.2052001953125
],
[
0.08905029296875,
0.2008056640625,
0.471435546875,
0.2008056640625,
0.471435546875,
0.2108154296875,
0.08905029296875,
0.2108154296875
],
[
0.498291015625,
0.20703125,
0.8837890625,
0.20703125,
0.8837890625,
0.2203369140625,
0.498291015625,
0.2203369140625
],
[
0.08905029296875,
0.2139892578125,
0.473876953125,
0.2139892578125,
0.473876953125,
0.2279052734375,
0.08905029296875,
0.2279052734375
],
[
0.5,
0.222900390625,
0.8857421875,
0.222900390625,
0.8857421875,
0.236083984375,
0.5,
0.236083984375
],
[
0.08905029296875,
0.2298583984375,
0.473876953125,
0.2298583984375,
0.473876953125,
0.24365234375,
0.08905029296875,
0.24365234375
],
[
0.5009765625,
0.2374267578125,
0.8857421875,
0.2374267578125,
0.8857421875,
0.250732421875,
0.5009765625,
0.250732421875
],
[
0.0882568359375,
0.2442626953125,
0.473876953125,
0.2442626953125,
0.473876953125,
0.25830078125,
0.0882568359375,
0.25830078125
],
[
0.5009765625,
0.251953125,
0.88623046875,
0.251953125,
0.88623046875,
0.26513671875,
0.5009765625,
0.26513671875
],
[
0.08905029296875,
0.260009765625,
0.473876953125,
0.260009765625,
0.473876953125,
0.2734375,
0.08905029296875,
0.2734375
],
[
0.5,
0.267578125,
0.8837890625,
0.267578125,
0.8837890625,
0.281005859375,
0.5,
0.281005859375
],
[
0.0882568359375,
0.274658203125,
0.473876953125,
0.274658203125,
0.473876953125,
0.28857421875,
0.0882568359375,
0.28857421875
],
[
0.50244140625,
0.2841796875,
0.8857421875,
0.2841796875,
0.8857421875,
0.294189453125,
0.50244140625,
0.294189453125
],
[
0.0882568359375,
0.290283203125,
0.26318359375,
0.290283203125,
0.26318359375,
0.3037109375,
0.0882568359375,
0.3037109375
],
[
0.498291015625,
0.29736328125,
0.54345703125,
0.29736328125,
0.54345703125,
0.3125,
0.498291015625,
0.3125
],
[
0.1422119140625,
0.322509765625,
0.1829833984375,
0.322509765625,
0.1829833984375,
0.337158203125,
0.1422119140625,
0.337158203125
],
[
0.2296142578125,
0.322509765625,
0.426513671875,
0.322509765625,
0.426513671875,
0.337158203125,
0.2296142578125,
0.337158203125
],
[
0.443603515625,
0.324462890625,
0.47216796875,
0.324462890625,
0.47216796875,
0.337158203125,
0.443603515625,
0.337158203125
],
[
0.5,
0.326904296875,
0.69384765625,
0.32763671875,
0.69384765625,
0.34228515625,
0.5,
0.341552734375
],
[
0.0882568359375,
0.356689453125,
0.473876953125,
0.35595703125,
0.473876953125,
0.37060546875,
0.0882568359375,
0.371826171875
],
[
0.5,
0.35107421875,
0.88232421875,
0.3515625,
0.88232421875,
0.3662109375,
0.5,
0.365478515625
],
[
0.5,
0.367431640625,
0.88232421875,
0.3662109375,
0.88232421875,
0.380615234375,
0.5,
0.3818359375
],
[
0.0906982421875,
0.375,
0.471435546875,
0.375,
0.471435546875,
0.385009765625,
0.0906982421875,
0.385009765625
],
[
0.08905029296875,
0.386962890625,
0.473876953125,
0.386474609375,
0.473876953125,
0.40087890625,
0.08905029296875,
0.401611328125
],
[
0.5,
0.38330078125,
0.8837890625,
0.38330078125,
0.8837890625,
0.396484375,
0.5,
0.396484375
],
[
0.08905029296875,
0.40283203125,
0.47216796875,
0.40283203125,
0.47216796875,
0.416748046875,
0.08905029296875,
0.416748046875
],
[
0.5,
0.397705078125,
0.60888671875,
0.397705078125,
0.60888671875,
0.410888671875,
0.5,
0.410888671875
],
[
0.08905029296875,
0.41845703125,
0.473876953125,
0.41845703125,
0.473876953125,
0.431884765625,
0.08905029296875,
0.431884765625
],
[
0.7646484375,
0.421630859375,
0.78173828125,
0.421630859375,
0.78173828125,
0.43115234375,
0.7646484375,
0.43115234375
],
[
0.08905029296875,
0.43505859375,
0.471435546875,
0.43505859375,
0.471435546875,
0.44580078125,
0.08905029296875,
0.44580078125
],
[
0.59619140625,
0.43310546875,
0.64794921875,
0.43310546875,
0.64794921875,
0.447509765625,
0.59619140625,
0.447509765625
],
[
0.7138671875,
0.436279296875,
0.73388671875,
0.436279296875,
0.73388671875,
0.4443359375,
0.7138671875,
0.4443359375
],
[
0.78369140625,
0.43505859375,
0.81396484375,
0.4375,
0.81201171875,
0.448974609375,
0.78173828125,
0.4462890625
],
[
0.08905029296875,
0.448974609375,
0.47216796875,
0.448974609375,
0.47216796875,
0.462158203125,
0.08905029296875,
0.462158203125
],
[
0.86376953125,
0.4443359375,
0.88623046875,
0.4443359375,
0.88623046875,
0.458984375,
0.86376953125,
0.458984375
],
[
0.76953125,
0.455078125,
0.78125,
0.455078125,
0.78125,
0.458984375,
0.76953125,
0.458984375
],
[
0.09234619140625,
0.464111328125,
0.1356201171875,
0.462158203125,
0.136474609375,
0.477783203125,
0.0931396484375,
0.479736328125
],
[
0.1298828125,
0.464111328125,
0.47216796875,
0.46337890625,
0.47216796875,
0.477783203125,
0.1298828125,
0.479248046875
],
[
0.56640625,
0.46728515625,
0.81884765625,
0.46728515625,
0.81884765625,
0.481689453125,
0.56640625,
0.481689453125
],
[
0.08660888671875,
0.479736328125,
0.4755859375,
0.479248046875,
0.4755859375,
0.49365234375,
0.08660888671875,
0.494384765625
],
[
0.0882568359375,
0.49560546875,
0.4599609375,
0.49560546875,
0.4599609375,
0.5087890625,
0.0882568359375,
0.5087890625
],
[
0.5,
0.492431640625,
0.88623046875,
0.492431640625,
0.88623046875,
0.51025390625,
0.5,
0.51025390625
],
[
0.5009765625,
0.51025390625,
0.8857421875,
0.51025390625,
0.8857421875,
0.5244140625,
0.5009765625,
0.5244140625
],
[
0.193603515625,
0.5234375,
0.369384765625,
0.5234375,
0.369384765625,
0.53857421875,
0.193603515625,
0.53857421875
],
[
0.450927734375,
0.5244140625,
0.473876953125,
0.52294921875,
0.476318359375,
0.5361328125,
0.451904296875,
0.5390625
],
[
0.5009765625,
0.52587890625,
0.8857421875,
0.52587890625,
0.8857421875,
0.5390625,
0.5009765625,
0.5390625
],
[
0.50244140625,
0.54150390625,
0.8857421875,
0.54150390625,
0.8857421875,
0.5546875,
0.50244140625,
0.5546875
],
[
0.08905029296875,
0.5537109375,
0.473876953125,
0.5537109375,
0.473876953125,
0.568359375,
0.08905029296875,
0.568359375
],
[
0.5009765625,
0.5546875,
0.8837890625,
0.55615234375,
0.8837890625,
0.57080078125,
0.5009765625,
0.5693359375
],
[
0.08905029296875,
0.5693359375,
0.325927734375,
0.5693359375,
0.325927734375,
0.58349609375,
0.08905029296875,
0.58349609375
],
[
0.5,
0.57080078125,
0.7900390625,
0.5693359375,
0.7900390625,
0.583984375,
0.5,
0.58544921875
],
[
0.1732177734375,
0.59521484375,
0.386474609375,
0.5947265625,
0.386474609375,
0.6123046875,
0.1732177734375,
0.61279296875
],
[
0.451904296875,
0.5986328125,
0.4755859375,
0.5986328125,
0.4755859375,
0.6123046875,
0.451904296875,
0.6123046875
],
[
0.5,
0.59326171875,
0.88623046875,
0.59326171875,
0.88623046875,
0.60693359375,
0.5,
0.60693359375
],
[
0.5009765625,
0.609375,
0.79736328125,
0.609375,
0.79736328125,
0.62255859375,
0.5009765625,
0.62255859375
],
[
0.0849609375,
0.6220703125,
0.351318359375,
0.625,
0.351318359375,
0.64599609375,
0.0849609375,
0.642578125
],
[
0.52783203125,
0.6318359375,
0.76708984375,
0.6298828125,
0.76708984375,
0.64453125,
0.52783203125,
0.646484375
],
[
0.08905029296875,
0.6494140625,
0.473876953125,
0.6494140625,
0.473876953125,
0.66357421875,
0.08905029296875,
0.66357421875
],
[
0.580078125,
0.64599609375,
0.8486328125,
0.65478515625,
0.84716796875,
0.68310546875,
0.57861328125,
0.673828125
],
[
0.52783203125,
0.65478515625,
0.62158203125,
0.66650390625,
0.61767578125,
0.68505859375,
0.52392578125,
0.673828125
],
[
0.0882568359375,
0.66552734375,
0.473876953125,
0.66552734375,
0.473876953125,
0.6787109375,
0.0882568359375,
0.6787109375
],
[
0.85888671875,
0.66015625,
0.8857421875,
0.66015625,
0.8857421875,
0.67236328125,
0.85888671875,
0.67236328125
],
[
0.0882568359375,
0.68115234375,
0.473876953125,
0.68115234375,
0.473876953125,
0.6943359375,
0.0882568359375,
0.6943359375
],
[
0.52685546875,
0.68505859375,
0.65185546875,
0.68505859375,
0.65185546875,
0.69970703125,
0.52685546875,
0.69970703125
],
[
0.0882568359375,
0.69580078125,
0.4755859375,
0.69580078125,
0.4755859375,
0.708984375,
0.0882568359375,
0.708984375
],
[
0.0882568359375,
0.71044921875,
0.473876953125,
0.71044921875,
0.473876953125,
0.7236328125,
0.0882568359375,
0.7236328125
],
[
0.5,
0.71337890625,
0.8857421875,
0.71337890625,
0.8857421875,
0.72802734375,
0.5,
0.72802734375
],
[
0.0882568359375,
0.72607421875,
0.473876953125,
0.72607421875,
0.473876953125,
0.7392578125,
0.0882568359375,
0.7392578125
],
[
0.5009765625,
0.73095703125,
0.88232421875,
0.73095703125,
0.88232421875,
0.7412109375,
0.5009765625,
0.7412109375
],
[
0.08660888671875,
0.74072265625,
0.473876953125,
0.7392578125,
0.473876953125,
0.75390625,
0.08660888671875,
0.7548828125
],
[
0.5,
0.74365234375,
0.88623046875,
0.7421875,
0.88623046875,
0.7568359375,
0.5,
0.75830078125
],
[
0.0849609375,
0.755859375,
0.35693359375,
0.7548828125,
0.35693359375,
0.76953125,
0.0849609375,
0.77099609375
],
[
0.5,
0.7587890625,
0.74609375,
0.7587890625,
0.74609375,
0.77294921875,
0.5,
0.77294921875
],
[
0.08905029296875,
0.79052734375,
0.306396484375,
0.79052734375,
0.306396484375,
0.8046875,
0.08905029296875,
0.8046875
],
[
0.5009765625,
0.79052734375,
0.65185546875,
0.79052734375,
0.65185546875,
0.8046875,
0.5009765625,
0.8046875
],
[
0.08905029296875,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.830078125,
0.08905029296875,
0.830078125
],
[
0.5,
0.814453125,
0.88623046875,
0.814453125,
0.88623046875,
0.82763671875,
0.5,
0.82763671875
],
[
0.0882568359375,
0.83056640625,
0.473876953125,
0.83056640625,
0.473876953125,
0.8447265625,
0.0882568359375,
0.8447265625
],
[
0.5009765625,
0.830078125,
0.88623046875,
0.830078125,
0.88623046875,
0.84326171875,
0.5009765625,
0.84326171875
],
[
0.0882568359375,
0.8466796875,
0.4755859375,
0.8466796875,
0.4755859375,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.498291015625,
0.84326171875,
0.8857421875,
0.8447265625,
0.8857421875,
0.859375,
0.498291015625,
0.85791015625
],
[
0.0906982421875,
0.86279296875,
0.473876953125,
0.86279296875,
0.473876953125,
0.87353515625,
0.0906982421875,
0.87353515625
],
[
0.5,
0.85986328125,
0.8857421875,
0.85986328125,
0.8857421875,
0.87353515625,
0.5,
0.87353515625
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.498291015625,
0.875,
0.88232421875,
0.87353515625,
0.88232421875,
0.88818359375,
0.498291015625,
0.8896484375
],
[
0.0882568359375,
0.892578125,
0.4755859375,
0.892578125,
0.4755859375,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.495849609375,
0.8896484375,
0.8896484375,
0.88623046875,
0.8896484375,
0.9072265625,
0.495849609375,
0.91015625
]
] |
[
"Deep Structured Energy Based Models for Anomaly Detection",
"where E ~ N(O, 2I) is an isotropic Gaussian noise. DAEs",
"convolutional neural networks (CNNs) are very effective",
"areeasytotrainwithstandard stochasticgradientdescent",
"at modeling data with spatial structure. In this work, we",
"(SGD) and perform significantly better than unregularized",
"extend EBMs further to deep structured EBMs (DSEBMs)",
"autoencoders.",
"where we allow the underlying deep neural network encod-",
"ing the energy function to take architectures varying from",
"While RBM and DAEs are typically considered as two al-",
"fully connected, recurrent, and convolutional. This gener-",
"ternative unsupervised deep models, it is recently shown",
"alizes EBMs (LeCun et al., 2006) as well as deep EBMs",
"they are closely related to each other. In particular, (Vin-",
"(Ngiam et al., 2011) as it makes our model applicable to a",
"cent,20ll)showsthattraininganRBMwithscorematch-",
"much wider spectrum of applications, including static data,",
"ing (SM) (Hyvarinen, 2005) is equivalent to a one-layer",
"sequential data and spatial data. Moreover, we show that by",
"DAE. SM is an alternative method to MLE, which is espe-",
"deriving the proper reconstruction functions with Equation",
"cially suitable for estimating non-normalized density func-",
"6, DSEBMs can be easily trained with SGD, regardless of",
"tions such as EBM. Instead of trying to directly maximize",
"the type of the underlying architecture. In the following,",
"the probability of training instances, SM minimizes the fol-",
"wewillelaboratetheformulationofDSEBMsforthethree",
"lowing objective function:",
"cases.",
"J(0) :",
"pr()(; 0) -()da,",
"(5)",
"3.1. Fully Connected EBMs",
"where p(r) is the true data distribution which is unknown;",
"This case is conceptually the same as the deep EBMs pro",
"posed in (Ngiam et al., 2011).. Without loss of general",
"b(:0)=Vrlogp(a:0)=-VrE(a:0)andr()=",
"V log Pr(α) are the score function of the model and the",
"ity, we express the energy function of an L-layer fully con-",
"true density function, respectively. In words, J(0) mea-",
"nected EBM as:",
"sures the expected distance of the scores between the model",
"Ki",
"density and the true density.(Vincent,2011)shows thatby",
"E(c;0)",
"b",
"hL,",
"approximating the Pr() with the Parzen window density",
"(8)",
"-",
"EN",
",N(ci,21),minimizing Equation 5yields an ob",
"s.t. hi = g(WThi-1 +bi), l e [1, L]",
"jective function in the same form as that of an autoencoder",
"in Equation 4, with a reconstruction function defined as:",
"where Wi e RKi-ixKi,b E RKi are the parameters for",
"the Ith layer; K, is the dimensionality of the lth layer. The",
"f(α; 0) = - VE(r; 0).",
"(6)",
"Oth layer is defined as the input itself; and thus we have",
"Ko = d, ho = c. We have explicitly included the term",
"Substituting E(α; 0) with Equation 2 yields a typical one-",
"-b'2 which acts as a prior,punishingtheprobability",
"layer DAE (up to a constant factor):",
"of the inputs that are far away from b'eRd",
"f(r; 0) = Wa(WTr+ b) +b,",
"(7)",
"Following the chain rule of gradient computation, one can",
"derive the reconstruction function as follows:",
"where o(a) is the sigmoid function +e",
"f(a;0)=-VE(;0)=h+b)",
"Equation 6 plays a key role in this work, as it allows one",
"Shh=o(WT hu-1+b) (Wh),",
"hi-1=Shi-",
"to efficiently train an arbitrary EBM in a similar way to",
"(9)",
"that of a DAE, as long as the energy function is differen-",
"for l e [1, L- 1],",
"tiable w.r.t. and 0. As will be demonstrated below, this",
"makes training a deep EBM with various underlying struc-",
"where h',= 1 with 1 E RKi denoting a column vector",
"tures in an end-to-end fashion, without the need of resorting",
"ofall ones::istheelement-wiseproductbetween vectors",
"to sophisticated sampling procedures (Ngiam et al., 2011)",
"One can then plug in the resulting f(α; ) into Equation 4",
"or layer wise pretraining (Hinton, 2010).",
"and train it as a regular L-layer DAE.",
"3. Deep Structured EBMs",
"3.2. Recurrent EBMs",
"Deep architectures allow one to model complicated pat-",
"Ourformulationof recurrentEBMsis similarto",
"terns efficiently, which makes it especially suitable for high",
"(Boulanger-Lewandowski et al., 2012), where an EBM is",
"dimensional data. On the other hand, it is often necessary",
"built at each time step, with parameters determined by an",
"to adapt the architecture of a deep model tothe structure of",
"underlying RNN. Formally, given a sequence of length T",
"data. For example, recurrent neural networks (RNNs) have",
"X = [',.., T], art Rd, we factorize the joint probabil-",
"been shown to work very well at modeling sequential data;",
"ity as p(x) = IIT1 p(ar*[a..t-1) withthe chain rule of"
] |
{
"arxiv_id": "1605.07717",
"language": "en",
"page_idx": 2
}
|
||
1605.07717-en-3
|
1605.07717-en-3.jpg
|
Deep Structured Energy Based Models for Anomaly Detection
probability. For each time step $t,p(x^t|x^{1,\ldots,t-1})$ is modeled as an EBM with energy $E( x^t| \theta ^t) . \textbf{In contrast to the}$ conventional formulation of EBM, the parameters of the EMB $\theta^t$ is a function of the inputs from all the previous time steps $x^1,\ldots,t-1.$ A natural choice of $\theta^t$ is to let it be the output of an RNN. As a concrete example, consider the energy function at each step follows that in Equation 2, by replacing $W,b,b^{\prime}$ with $W^t,b^t,b^{\prime t}.$ Directly letting the RNN to update all the parameters at each step requires a large RNN with lots of parameters. As a remedy, (BoulangerLewandowski et al., 2012) proposes to fix $W^t=W$ for all time steps, only letting $b^t$ and $b^{\prime t}$ to be updated by the $RNN:$
$$\begin{aligned}h^t&=g(W_{hh}h^{t-1}+W_{hx}x^t+b_h)\\b^t&=W_{bh}h^t+b,\:b^{\prime t}=W_{b^{\prime}h}h^t+b^{\prime},\end{aligned}$$
(10)
$\begin{array}{cccc}{\mathrm{where~}W_{hh}}&{\in}&{R^{K_{rnn}\times K_{rnn}},W_{hx}}&{\in}&{R^{K_{rnn}\times d}},{b_{h}}&{\in}\end{array}$
$R_{rnn}^{K_{rnn}}$ are parameters of the RNN; $W_bh$
$\in$
$R^{K_{ebm}\times K_{rnn}},\hat{b}\in R^{K_{ebm}},W_{b^{\prime}h}\in R^{d\times K_{rnn}},b^{\prime}\in R^{d}$ are the weights with which to transform the hidden state of the RNN to the adaptive biases.
Training the recurrent EBM with score matching is simiis the partition function for the $t$th step. Plugging $p(\mathbf{x})$ $\operatorname*{into}_{[-T]}$Equation 5, we have $\psi(\mathbf{x})=\nabla_{\mathbf{x}}\log p(\mathbf{x})\approx$ $-[\nabla_{x^{1}}^{T}\hat{E}(x^{1};\theta^{1}),...,\nabla_{x^{T}}^{T}E(x^{T};\theta^{T})]^{T}.$In the last step, we have made a simplification by omitting the gradient term of $\nabla_{x^{j}}E(x^{i};\theta^{i}),\tilde{\mathrm{for~}j}<i^{\:1}.\tilde{\mathrm{Accordingly,Equation~6~is~mod-}}$ $\begin{array}{c}{\mathrm{iffed~to~}}f(\mathbf{x})=\mathbf{x}-[\nabla_{x^{1}}^{T}E(x^{1};\theta^{1}),...,\hat{\nabla}_{x^{T}}^{T}E(x^{T};\theta^{T})]^{T}.\end{array}$ One is then able to train the recurrent EBM by plugging $f(\mathbf{x})$ into Equation 4 and perform the standard SGD. Note that in this case the standard backpropagation is replaced with backpropagation through time, due to the need of updating the RNN parameters.
### 3.3. Convolutional EBMs
Previously, the combination of CNN and RBM have been proposed in (Lee et al.,2009),where several layers of RBMs are alternately convolved with an image then stacked on top of each other. In this paper, we take a significantly different approach by directly building deep EBMs with convolution operators (with optional pooling layers or fully connected layers), simply by replacing $h_L$ in Equation 8 with the output of a CNN. Using a deterministic deep convolutional EBM allows one to directly train the model end-to-end with score matching, thus significantly simplifies the training procedure compared with (Lee et al., 2009). Formally, consider the input of the $(l-1)$th layer
lWhile one can also choose to use the full gradient, we find this simplification works well in practice, yielding an objective in a much more succinct form.
$h_{l-1}\in R^{K_{l-1}\times d_{l-1}\times d_{l-1}}$ is a $d_l-1\times d_{l-1}$ image with $K_l-1$ channels. We define $h_l$ as the output of a convolution layer:
$$h_{l,j}=g(\sum\limits_{k=1}^{K_{l-1}}\tilde{W}_{l,j,k}*h_{l-1,k}+b_{l,j}),\:j\in[1,K_{l}].$$
Here $W\in R^{K_l\times K_{l-1}\times d_{w,l}\times d_{w,l}}$ are the $K_l$ convolutional filters of size $d_w,l\times d_{w,l};b_l\in R^{K_l}$ is the bias for each filter. We denote the tilde operator $(\tilde{A})$ as flipping a matrix $A$ horizontally and vertically; * is the "valid" convolution operator, where convolution is only conducted where the input and the filter fully overlap. $d_l=d_{l-1}-d_{w,l}+1$ is thus the size of the output image following the valid convolution. In order to compute the reconstruction function following equation 9, we modify the recurrence equation from $h_l^{\prime}$ to $h_l-1^{\prime}$ for a convolutional operator as
$$h'_{l-1,k}=\sum_{j=1}^{K_{l}}[\sigma(\sum_{k=1}^{K_{l-1}}\tilde{W}_{l,j,k}*h_{l-1,k}+b_{l,j})\cdot(W_{l,j,k}\odot h_{l,j}^{\prime})].$$
Here we have denoted $\odot$ as the "full" convolution operator,
where convolution is conducted whenever the input and the
ically follows a convolution layer. Denote $d_p\times d_p$ as the pooling window size, passing $h_l-1$ through a max pooling layer gives an output as:
(13)
$$\begin{aligned}&h_{l,k,p,q},\:x_{l,k,p,q},y_{l,k,p,q}\\&=\max_{(p-1)d_p+1\leq i\leq pd_p,(q-1)d_p+1\leq j\leq qd_p}h_{l-1,k,i,j}.\end{aligned}$$
Here the max operation returns the maximum value as the first term and the corresponding maximum coordinates as the second and third. Rewriting the recurrence equation corresponding to Equation 8 yields:
(14)
$$h_{l-1,k,x_{l,k,p,q},y_{l,k,p,q}}^{\prime}=h_{l,k,p,q}^{\prime},\:p,q\in[1,d_{l}],$$
where the unassigned entries of $h_l-1^{\prime}$ are all set as zero.
The derivation above shows that one is able to compute the reconstruction function $f(x)$ for a deep convolutional EBM consisting of convolution, max pooling and fully connected layers. Other types of layers such as mean pooling can also be managed in a similar way, which is omitted due to the space limit.
## $4. \textbf{Deep Structured EBMs for Anomaly}$ Detection
Performing anomaly detection given a trained EBM naturally corresponds to identifying data points that are assigned low probability by the model. With a trained DSEBM, we can then select samples that are assigned
|
[
[
0.291748046875,
0.05682373046875,
0.68310546875,
0.05682373046875,
0.68310546875,
0.0706787109375,
0.291748046875,
0.0706787109375
],
[
0.08660888671875,
0.083984375,
0.4755859375,
0.08209228515625,
0.4755859375,
0.0985107421875,
0.08660888671875,
0.10101318359375
],
[
0.496826171875,
0.0802001953125,
0.88623046875,
0.08209228515625,
0.88623046875,
0.10162353515625,
0.496826171875,
0.0997314453125
],
[
0.08905029296875,
0.10101318359375,
0.473876953125,
0.10101318359375,
0.473876953125,
0.11553955078125,
0.08905029296875,
0.11553955078125
],
[
0.5009765625,
0.10101318359375,
0.8857421875,
0.10101318359375,
0.8857421875,
0.1142578125,
0.5009765625,
0.1142578125
],
[
0.08905029296875,
0.11614990234375,
0.473876953125,
0.11614990234375,
0.473876953125,
0.1300048828125,
0.08905029296875,
0.1300048828125
],
[
0.57666015625,
0.12371826171875,
0.60546875,
0.1256103515625,
0.6044921875,
0.1383056640625,
0.576171875,
0.1363525390625
],
[
0.08905029296875,
0.1307373046875,
0.473876953125,
0.1307373046875,
0.473876953125,
0.14453125,
0.08905029296875,
0.14453125
],
[
0.5087890625,
0.136962890625,
0.88818359375,
0.1351318359375,
0.88818359375,
0.1558837890625,
0.5087890625,
0.1578369140625
],
[
0.0849609375,
0.14453125,
0.4755859375,
0.143310546875,
0.4755859375,
0.1610107421875,
0.0849609375,
0.1622314453125
],
[
0.08905029296875,
0.1610107421875,
0.473876953125,
0.1610107421875,
0.473876953125,
0.1749267578125,
0.08905029296875,
0.1749267578125
],
[
0.57861328125,
0.1572265625,
0.60546875,
0.1572265625,
0.60546875,
0.1685791015625,
0.57861328125,
0.1685791015625
],
[
0.0882568359375,
0.1767578125,
0.473876953125,
0.1767578125,
0.473876953125,
0.1900634765625,
0.0882568359375,
0.1900634765625
],
[
0.498291015625,
0.1749267578125,
0.8896484375,
0.1749267578125,
0.8896484375,
0.1956787109375,
0.498291015625,
0.1956787109375
],
[
0.0882568359375,
0.1912841796875,
0.47216796875,
0.1912841796875,
0.47216796875,
0.2052001953125,
0.0882568359375,
0.2052001953125
],
[
0.5,
0.1944580078125,
0.8857421875,
0.1932373046875,
0.8857421875,
0.2083740234375,
0.5,
0.208984375
],
[
0.0882568359375,
0.20703125,
0.473876953125,
0.20703125,
0.473876953125,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.5,
0.210205078125,
0.8857421875,
0.2108154296875,
0.8857421875,
0.2259521484375,
0.5,
0.2247314453125
],
[
0.08905029296875,
0.2215576171875,
0.47216796875,
0.2215576171875,
0.47216796875,
0.236083984375,
0.08905029296875,
0.236083984375
],
[
0.5009765625,
0.2266845703125,
0.8857421875,
0.2266845703125,
0.8857421875,
0.240478515625,
0.5009765625,
0.240478515625
],
[
0.0882568359375,
0.236083984375,
0.473876953125,
0.23486328125,
0.473876953125,
0.25,
0.0882568359375,
0.250732421875
],
[
0.498291015625,
0.242431640625,
0.8857421875,
0.240478515625,
0.8857421875,
0.25390625,
0.498291015625,
0.255615234375
],
[
0.08660888671875,
0.25,
0.4755859375,
0.248779296875,
0.4755859375,
0.266357421875,
0.08660888671875,
0.267578125
],
[
0.5,
0.2568359375,
0.88623046875,
0.2568359375,
0.88623046875,
0.270751953125,
0.5,
0.270751953125
],
[
0.0882568359375,
0.266357421875,
0.13232421875,
0.266357421875,
0.13232421875,
0.281005859375,
0.0882568359375,
0.281005859375
],
[
0.5,
0.272705078125,
0.8857421875,
0.272705078125,
0.8857421875,
0.285888671875,
0.5,
0.285888671875
],
[
0.160888671875,
0.287353515625,
0.3857421875,
0.287353515625,
0.3857421875,
0.3017578125,
0.160888671875,
0.3017578125
],
[
0.5009765625,
0.287353515625,
0.88623046875,
0.287353515625,
0.88623046875,
0.300537109375,
0.5009765625,
0.300537109375
],
[
0.443603515625,
0.298583984375,
0.473876953125,
0.298583984375,
0.473876953125,
0.311279296875,
0.443603515625,
0.311279296875
],
[
0.158447265625,
0.30615234375,
0.400390625,
0.30615234375,
0.400390625,
0.32080078125,
0.158447265625,
0.32080078125
],
[
0.5009765625,
0.302978515625,
0.88623046875,
0.302978515625,
0.88623046875,
0.31640625,
0.5009765625,
0.31640625
],
[
0.5,
0.31640625,
0.81396484375,
0.317626953125,
0.81396484375,
0.33203125,
0.5,
0.330810546875
],
[
0.0849609375,
0.326904296875,
0.47802734375,
0.326904296875,
0.47802734375,
0.35107421875,
0.0849609375,
0.35107421875
],
[
0.1396484375,
0.3466796875,
0.438720703125,
0.347900390625,
0.438720703125,
0.3623046875,
0.1396484375,
0.361083984375
],
[
0.56884765625,
0.3466796875,
0.58837890625,
0.3466796875,
0.58837890625,
0.354736328125,
0.56884765625,
0.354736328125
],
[
0.60888671875,
0.3447265625,
0.63623046875,
0.3447265625,
0.63623046875,
0.35595703125,
0.60888671875,
0.35595703125
],
[
0.463134765625,
0.3515625,
0.4697265625,
0.3515625,
0.4697265625,
0.356689453125,
0.463134765625,
0.356689453125
],
[
0.496826171875,
0.354736328125,
0.8837890625,
0.352783203125,
0.8837890625,
0.375,
0.496826171875,
0.376953125
],
[
0.0906982421875,
0.347900390625,
0.1422119140625,
0.347900390625,
0.1422119140625,
0.359130859375,
0.0906982421875,
0.359130859375
],
[
0.08660888671875,
0.356689453125,
0.476318359375,
0.35791015625,
0.476318359375,
0.380126953125,
0.08660888671875,
0.37890625
],
[
0.564453125,
0.375732421875,
0.5947265625,
0.375732421875,
0.5947265625,
0.386962890625,
0.564453125,
0.386962890625
],
[
0.60888671875,
0.376953125,
0.6357421875,
0.376953125,
0.6357421875,
0.385009765625,
0.60888671875,
0.385009765625
],
[
0.85546875,
0.386962890625,
0.8857421875,
0.386962890625,
0.8857421875,
0.3984375,
0.85546875,
0.3984375
],
[
0.08905029296875,
0.37744140625,
0.473876953125,
0.37744140625,
0.473876953125,
0.391357421875,
0.08905029296875,
0.391357421875
],
[
0.08905029296875,
0.393310546875,
0.273681640625,
0.393310546875,
0.273681640625,
0.4072265625,
0.08905029296875,
0.4072265625
],
[
0.5009765625,
0.401611328125,
0.8837890625,
0.401611328125,
0.8837890625,
0.415283203125,
0.5009765625,
0.415283203125
],
[
0.08905029296875,
0.415283203125,
0.473876953125,
0.415283203125,
0.473876953125,
0.429931640625,
0.08905029296875,
0.429931640625
],
[
0.5009765625,
0.417236328125,
0.8857421875,
0.417236328125,
0.8857421875,
0.43115234375,
0.5009765625,
0.43115234375
],
[
0.0882568359375,
0.43115234375,
0.473876953125,
0.43115234375,
0.473876953125,
0.4443359375,
0.0882568359375,
0.4443359375
],
[
0.5009765625,
0.431884765625,
0.88232421875,
0.431884765625,
0.88232421875,
0.44580078125,
0.5009765625,
0.44580078125
],
[
0.1846923828125,
0.4462890625,
0.273681640625,
0.442626953125,
0.273681640625,
0.453857421875,
0.1846923828125,
0.458251953125
],
[
0.27294921875,
0.448974609375,
0.476318359375,
0.4443359375,
0.476318359375,
0.46533203125,
0.27294921875,
0.4697265625
],
[
0.498291015625,
0.447509765625,
0.88232421875,
0.447509765625,
0.88232421875,
0.4609375,
0.498291015625,
0.4609375
],
[
0.2059326171875,
0.458984375,
0.2548828125,
0.45703125,
0.2548828125,
0.468505859375,
0.2059326171875,
0.47021484375
],
[
0.08660888671875,
0.45068359375,
0.1871337890625,
0.448974609375,
0.1871337890625,
0.46337890625,
0.08660888671875,
0.46533203125
],
[
0.498291015625,
0.4609375,
0.8857421875,
0.462158203125,
0.8857421875,
0.4765625,
0.498291015625,
0.47607421875
],
[
0.08905029296875,
0.4697265625,
0.47216796875,
0.4697265625,
0.47216796875,
0.484130859375,
0.08905029296875,
0.484130859375
],
[
0.5009765625,
0.477783203125,
0.8857421875,
0.477783203125,
0.8857421875,
0.4912109375,
0.5009765625,
0.4912109375
],
[
0.08905029296875,
0.48486328125,
0.3203125,
0.48486328125,
0.3203125,
0.498779296875,
0.08905029296875,
0.498779296875
],
[
0.351318359375,
0.484130859375,
0.44677734375,
0.486083984375,
0.44677734375,
0.50048828125,
0.351318359375,
0.498779296875
],
[
0.4453125,
0.488037109375,
0.471435546875,
0.488037109375,
0.471435546875,
0.496826171875,
0.4453125,
0.496826171875
],
[
0.09234619140625,
0.498779296875,
0.473876953125,
0.498779296875,
0.473876953125,
0.51318359375,
0.09234619140625,
0.51318359375
],
[
0.5009765625,
0.49365234375,
0.6611328125,
0.49365234375,
0.6611328125,
0.5068359375,
0.5009765625,
0.5068359375
],
[
0.08905029296875,
0.51513671875,
0.473876953125,
0.51513671875,
0.473876953125,
0.52880859375,
0.08905029296875,
0.52880859375
],
[
0.51953125,
0.51806640625,
0.6806640625,
0.52099609375,
0.6806640625,
0.53515625,
0.51953125,
0.53271484375
],
[
0.0906982421875,
0.52978515625,
0.47216796875,
0.52978515625,
0.47216796875,
0.54443359375,
0.0906982421875,
0.54443359375
],
[
0.76708984375,
0.53515625,
0.83837890625,
0.54150390625,
0.8349609375,
0.5546875,
0.7646484375,
0.548828125
],
[
0.85546875,
0.5322265625,
0.8857421875,
0.5322265625,
0.8857421875,
0.54443359375,
0.85546875,
0.54443359375
],
[
0.08905029296875,
0.54443359375,
0.471435546875,
0.54443359375,
0.471435546875,
0.5595703125,
0.08905029296875,
0.5595703125
],
[
0.6435546875,
0.54052734375,
0.6689453125,
0.54052734375,
0.6689453125,
0.54736328125,
0.6435546875,
0.54736328125
],
[
0.54150390625,
0.546875,
0.76953125,
0.546875,
0.76953125,
0.56103515625,
0.54150390625,
0.56103515625
],
[
0.08660888671875,
0.5595703125,
0.473876953125,
0.56005859375,
0.473876953125,
0.57470703125,
0.08660888671875,
0.57373046875
],
[
0.0882568359375,
0.57470703125,
0.473876953125,
0.57373046875,
0.473876953125,
0.58837890625,
0.0882568359375,
0.5888671875
],
[
0.5009765625,
0.57080078125,
0.8857421875,
0.57080078125,
0.8857421875,
0.583984375,
0.5009765625,
0.583984375
],
[
0.0882568359375,
0.59033203125,
0.473876953125,
0.59033203125,
0.473876953125,
0.60400390625,
0.0882568359375,
0.60400390625
],
[
0.5009765625,
0.5859375,
0.88623046875,
0.5859375,
0.88623046875,
0.599609375,
0.5009765625,
0.599609375
],
[
0.0882568359375,
0.60595703125,
0.47216796875,
0.60595703125,
0.47216796875,
0.619140625,
0.0882568359375,
0.619140625
],
[
0.5009765625,
0.6015625,
0.88623046875,
0.6015625,
0.88623046875,
0.61572265625,
0.5009765625,
0.61572265625
],
[
0.08660888671875,
0.62060546875,
0.275390625,
0.619140625,
0.275390625,
0.6337890625,
0.08660888671875,
0.63525390625
],
[
0.5009765625,
0.6162109375,
0.73779296875,
0.6162109375,
0.73779296875,
0.6298828125,
0.5009765625,
0.6298828125
],
[
0.529296875,
0.64013671875,
0.66015625,
0.646484375,
0.65869140625,
0.66552734375,
0.52783203125,
0.6591796875
],
[
0.70166015625,
0.64306640625,
0.83251953125,
0.64013671875,
0.83251953125,
0.65576171875,
0.70166015625,
0.6591796875
],
[
0.85302734375,
0.642578125,
0.8857421875,
0.642578125,
0.8857421875,
0.65478515625,
0.85302734375,
0.65478515625
],
[
0.0882568359375,
0.6513671875,
0.265625,
0.6513671875,
0.265625,
0.66552734375,
0.0882568359375,
0.66552734375
],
[
0.68115234375,
0.646484375,
0.70166015625,
0.646484375,
0.70166015625,
0.65380859375,
0.68115234375,
0.65380859375
],
[
0.0882568359375,
0.67529296875,
0.4755859375,
0.67529296875,
0.4755859375,
0.689453125,
0.0882568359375,
0.689453125
],
[
0.5009765625,
0.67041015625,
0.861328125,
0.67041015625,
0.861328125,
0.6845703125,
0.5009765625,
0.6845703125
],
[
0.0882568359375,
0.68994140625,
0.47216796875,
0.68994140625,
0.47216796875,
0.705078125,
0.0882568359375,
0.705078125
],
[
0.5009765625,
0.6923828125,
0.8857421875,
0.6923828125,
0.8857421875,
0.70556640625,
0.5009765625,
0.70556640625
],
[
0.0882568359375,
0.70556640625,
0.473876953125,
0.70556640625,
0.473876953125,
0.7197265625,
0.0882568359375,
0.7197265625
],
[
0.5,
0.70849609375,
0.8857421875,
0.70849609375,
0.8857421875,
0.7216796875,
0.5,
0.7216796875
],
[
0.0882568359375,
0.7216796875,
0.473876953125,
0.7216796875,
0.473876953125,
0.73486328125,
0.0882568359375,
0.73486328125
],
[
0.5,
0.7236328125,
0.8857421875,
0.7236328125,
0.8857421875,
0.7373046875,
0.5,
0.7373046875
],
[
0.08660888671875,
0.736328125,
0.473876953125,
0.73486328125,
0.473876953125,
0.75,
0.08660888671875,
0.75048828125
],
[
0.498291015625,
0.7373046875,
0.88623046875,
0.73779296875,
0.88623046875,
0.75244140625,
0.498291015625,
0.751953125
],
[
0.08905029296875,
0.751953125,
0.473876953125,
0.751953125,
0.473876953125,
0.76513671875,
0.08905029296875,
0.76513671875
],
[
0.5009765625,
0.75390625,
0.88623046875,
0.75390625,
0.88623046875,
0.767578125,
0.5009765625,
0.767578125
],
[
0.0882568359375,
0.767578125,
0.47216796875,
0.767578125,
0.47216796875,
0.78076171875,
0.0882568359375,
0.78076171875
],
[
0.5,
0.76806640625,
0.580078125,
0.7666015625,
0.580078125,
0.78076171875,
0.5,
0.78271484375
],
[
0.0882568359375,
0.7822265625,
0.47216796875,
0.7822265625,
0.47216796875,
0.79541015625,
0.0882568359375,
0.79541015625
],
[
0.08660888671875,
0.79541015625,
0.473876953125,
0.796875,
0.473876953125,
0.81103515625,
0.08660888671875,
0.81005859375
],
[
0.50244140625,
0.80322265625,
0.82421875,
0.80322265625,
0.82421875,
0.8173828125,
0.50244140625,
0.8173828125
],
[
0.0882568359375,
0.81103515625,
0.47216796875,
0.81298828125,
0.47216796875,
0.8271484375,
0.0882568359375,
0.82470703125
],
[
0.52294921875,
0.82275390625,
0.60107421875,
0.82275390625,
0.60107421875,
0.833984375,
0.52294921875,
0.833984375
],
[
0.0906982421875,
0.8291015625,
0.47216796875,
0.8291015625,
0.47216796875,
0.8388671875,
0.0906982421875,
0.8388671875
],
[
0.08660888671875,
0.84130859375,
0.473876953125,
0.84228515625,
0.473876953125,
0.85693359375,
0.08660888671875,
0.85595703125
],
[
0.5,
0.84521484375,
0.8857421875,
0.84521484375,
0.8857421875,
0.85986328125,
0.5,
0.85986328125
],
[
0.108642578125,
0.8642578125,
0.473876953125,
0.86572265625,
0.473876953125,
0.8798828125,
0.108642578125,
0.87890625
],
[
0.5,
0.861328125,
0.8837890625,
0.85986328125,
0.8837890625,
0.875,
0.5,
0.87548828125
],
[
0.0882568359375,
0.8798828125,
0.4755859375,
0.8798828125,
0.4755859375,
0.89306640625,
0.0882568359375,
0.89306640625
],
[
0.5009765625,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.08660888671875,
0.892578125,
0.25732421875,
0.892578125,
0.25732421875,
0.90576171875,
0.08660888671875,
0.90576171875
],
[
0.5,
0.8916015625,
0.88623046875,
0.8916015625,
0.88623046875,
0.90478515625,
0.5,
0.90478515625
]
] |
[
"Deep Structured Energy Based Models for Anomaly Detection",
"probability. For each time step t, p(ct|l...t-1) is mod-",
"ht-1 E RKi-1xdi-1xdi-1 is adl-1xdi-1 image with Ki-1",
"eled as an EBM with energy E(αtjot). In contrast to the",
"channels. We define hi as the output of a convolution layer:",
"conventional formulation of EBM, the parameters of the",
"Ki-1",
"EMB ot is a function of the inputs from all the previous",
"hl,j = g( Wi,j,k * hu-1,k + b.,), je [1, Ki]: (11)",
"time steps rl...-1. A natural choice of ot is to let it be",
"the output of an RNN. As a concrete example, consider the",
"k=1",
"energy function at each step follows that in Equation 2, by",
"Here W e RKixKi-1xdw,ixdw,l are the Ki convolutional",
"replacing W, b, b' with Wt, bt, b't. Directly letting the RNN",
"filters of size duw, × dwt; bu E RKi is the bias for each",
"to update all the parameters at each step requires a large",
"filter. We denote the tilde operator (A) as flipping a matrix",
"RNN with lots of parameters. As a remedy, (Boulanger-",
"A horizontally and vertically; * is the \"valid\"convolution",
"Lewandowski et al., 2012) proposes to fix wt = W for",
"operator, where convolution is only conducted where the",
"all time steps, only letting bt and b't to be updated by the",
"input and the filter fully overlap. di = di-1 - dw.t + 1 is",
"RNN:",
"thus the size of the output image following the valid con-",
"ht = g(Whhht-1 + Whrct + bh)",
"volution. In order to compute the reconstruction function",
"(10)",
"bt = Whhht + b. b't =Whhht + b'",
"following equation 9, we modify the recurrence equation",
"from h, to h,-, for a convolutional operator as:",
"Where Wh E RKrnnxKrmn,Wh E RKrnnxd, bh E",
"areparameters of theRNN;Wbh",
"Ki",
"Ki-",
"a",
"hi-1, =[o(W1,3,h*h-1,h+bl,3)-(Wi,j,koht,)]",
"RKrnn",
"RKebmxKrn,be RKebm,Wyh e RdxKnn,b' e Rdare",
"j=1",
"k=1",
"(12)",
"the weights with which to transform the hidden state of the",
"RNN to the adaptive biases.",
"Here we have denoted as the\"full\" convolution operator,",
"Training the recurrent EBM with score matching is simi-",
"where convolution is conducted whenever the input and the",
"lar to that of a fully connected EBM. To see this, we now",
"filter overlap by at least one position. Besides the convolu",
"e-T-)(at:et",
"2, where Zt =Zre-E(z;0\")",
"tion layer, we can also use a max pooling layer which typ",
"I-,Z",
"have p(x)=",
"ically follows a convolution layer. Denote dp × d, as the",
"is the partition function for the tth step. Plugging p(x)",
"pooling window size, passing hi-1 through a max pooling",
"into Equation 5, we have (x)",
"Vxlogp(x)",
"~",
"-/VT E(αl: o'), ...TE(T: oT))T. In the last step, we",
"layer gives an output as:",
"have made a simplification by omitting the gradient term of",
"hi,k,p,g,Ti,k,p,q,yl,k,p.q",
"VE(; o'), for j<.Accordingly,Equation 6 is mod-",
"hi-1,k,i.j.",
"(13)",
"ified to f(x) = x - [VTIE(rl; 01),., VT-E(αT;OT))T",
"max",
"(p-1)dp+1≤i≤pdp.(q-1)dp+1≤j≤qdp",
"One is then able to train the recurrent EBM by plugging",
"f(x) into Equation 4 and perform the standard SGD. Note",
"Here the max operation returns the maximum value as the",
"that in this case the standard backpropagation is replaced",
"first term and the corresponding maximum coordinates as",
"with backpropagation through time, due to the need of up-",
"the second and third. Rewriting the recurrence equation",
"dating the RNN parameters.",
"corresponding to Equation 8 yields:",
"h'1,k,TL,,p.q.y.k.p.",
"k,p.g, p,q e[1, di].",
"(14)",
"3.3. Convolutional EBMs",
"hi",
"Previously, the combination of CNN and RBM have",
"where the unassigned entries of h,-, are all set as zero",
"been proposed in (Lee et al., 2009), where several lay-",
"The derivation above shows that one is able to compute the",
"ers of RBMs are alternately convolved with an image then",
"reconstructionfunctionf(α)foradeepconvolutionalEBM",
"stacked on top of each other. In this paper, we take a signif-",
"consisting of convolution, max pooling and fully connected",
"icantly different approach by directly building deep EBMs",
"layers. Other types of layers such as mean pooling can also",
"with convolution operators (with optional pooling layers or",
"be managed in a similar way, which is omitted due to the",
"fully connected layers), simply by replacing h in Equa-",
"space limit.",
"tion 8 with the output of a CNN. Using a deterministic",
"deep convolutional EBM allows one to directly train the",
"4.Deep Structured EBMs for Anomaly",
"model end-to-end with score matching,thus significantly",
"Detection",
"simplifiesthetraining procedurecompared with (Lee etal.",
"2009). Formally, consider the input of the (1 - 1)th layer",
"Performing anomaly detection given a trained EBM nat-",
"Iwhile one can also choose to use the full gradient, we find",
"urally corresponds to identifying data points that are as-",
"this simplification works well in practice, yielding an objective in",
"signed low probability by the model.With a trained",
"a much more succinct form.",
"DSEBM, we can then select samples that are assigned"
] |
{
"arxiv_id": "1605.07717",
"language": "en",
"page_idx": 3
}
|
||
0710.4852-en-0
|
0710.4852-en-0.jpg
|
An Assembler Driven Verification Methodology (ADVM)
John S. MacBeth Verilab GmbH
Lichtenbergerstrasse 8, Garching
85748 Munich, Germany john.macbeth@verilab.com
Dietmar Heinz
Infineon Technologies AG
St. Martin Strasse 7681609 Munich, Germany dietmar.heinz@ infineon.com
Ken Gray
Verilab GmbH
Lichtenbergerstrasse 8, Garching
85748 Munich, Germany ken.gray@verilab.com
# Abstract
This paper presents an overview of an assembler driven verification methodology (ADVM) that was created and implemented for a chip card project at Infineon Technologies $AG\left[2\right].The$ primary advantage of this methodology is that it enables rapid porting of directed tests to new targets and derivatives, with only a minimum amount of code re- $factoring.Asaconsequence,considerableverificationde-$ velopment time and effort was saved.
## $1. \textbf{Introduction}$
Presently, the methods used for verifying complex microprocessor based System-on-Chip (SoC) devices can take many forms [1]. Such verification methods vary from high-level verification languages or tools (such as Verisity's Specman tool and 'e' language or Synopsys's Vera) to running compiled test code on the device. The advantage of the latter method, running code on the device, is that the code can be run on any simulation or emulation platform. This means that the same suite of assembler tests can be used to perform functional verification of each of the following development platforms:
## Golden Reference Model The software simulator that is supplied to the customer for software development
HDL-RTL Simulation The HDL design for silicon
HDL-Gate Level Simulation The post-synthesis HDL de-
sign for silicon
Hardware Accelerator The hardware emulator used by
both the customer and the embedded software teams
for fırmware sign-off (e.g Quickturn and IKOS)
Bondout Silicon The software development platform with
silicon performance. It is also enhanced to include ex-
tra hardware debugging capabilities
Product Silicon The final product silicon for the customer
Although this method has a clear advantage in terms of crossing many simulation/emulation domains, it is restricted by the short comings of directed testing.
To verify a complex device using directed testing, many tests must be written in an attempt to cover as many functional modes of operation as possible. This means that over time a large collection of directed test code will be developed and will require re-factoring (i.e. maintenance) with each change in the specification or when migrating the test code to new derivatives.
Writing test code that crosses simulation domains should not present many problems as most platforms should execute the code in the same way. If they don't then a bug or issue has been found in that particular simulation domain. A challenge emerges however, if the aim is to re-target existing test code to a new derivative and it is this challenge that the ADVM primarily overcomes. The following section describes the test environment required to accomplish this goal.
## 2. Test Environment Structure
Abstraction is the key to dealing with change. Creating an abstraction layer between the test and the physical implementation provides a layer that can absorb any changes.
61
$\mathrm{COMPUTER}^{\text{1125年}}$ $SOCIETY$
Proceedings of the Design, Automation and Test in Europe Conference and Exhibition (DATE'05)
$1530-1591/05\$20.00$ IEEE
|
[
[
0.2303466796875,
0.1029052734375,
0.79736328125,
0.1029052734375,
0.79736328125,
0.117431640625,
0.2303466796875,
0.117431640625
],
[
0.301513671875,
0.1514892578125,
0.438720703125,
0.1514892578125,
0.438720703125,
0.1666259765625,
0.301513671875,
0.1666259765625
],
[
0.61376953125,
0.1514892578125,
0.73193359375,
0.1514892578125,
0.73193359375,
0.1666259765625,
0.61376953125,
0.1666259765625
],
[
0.313720703125,
0.169189453125,
0.428955078125,
0.169189453125,
0.428955078125,
0.184326171875,
0.313720703125,
0.184326171875
],
[
0.56787109375,
0.169189453125,
0.77685546875,
0.1685791015625,
0.77685546875,
0.18310546875,
0.56787109375,
0.184326171875
],
[
0.238525390625,
0.186279296875,
0.5,
0.1881103515625,
0.5,
0.20263671875,
0.238525390625,
0.2008056640625
],
[
0.58837890625,
0.1881103515625,
0.75634765625,
0.1881103515625,
0.75634765625,
0.2020263671875,
0.58837890625,
0.2020263671875
],
[
0.26953125,
0.2052001953125,
0.468017578125,
0.2052001953125,
0.468017578125,
0.2197265625,
0.26953125,
0.2197265625
],
[
0.57177734375,
0.203857421875,
0.76953125,
0.205810546875,
0.76953125,
0.2203369140625,
0.57177734375,
0.2183837890625
],
[
0.2607421875,
0.222900390625,
0.4794921875,
0.222900390625,
0.4794921875,
0.2374267578125,
0.2607421875,
0.2374267578125
],
[
0.5595703125,
0.222900390625,
0.78759765625,
0.222900390625,
0.78759765625,
0.2374267578125,
0.5595703125,
0.2374267578125
],
[
0.47216796875,
0.2442626953125,
0.55126953125,
0.246826171875,
0.55126953125,
0.261962890625,
0.47216796875,
0.260009765625
],
[
0.455810546875,
0.264404296875,
0.57177734375,
0.264404296875,
0.57177734375,
0.279052734375,
0.455810546875,
0.279052734375
],
[
0.381591796875,
0.281005859375,
0.64208984375,
0.28271484375,
0.64208984375,
0.29736328125,
0.381591796875,
0.29541015625
],
[
0.414306640625,
0.300537109375,
0.611328125,
0.300537109375,
0.611328125,
0.31494140625,
0.414306640625,
0.31494140625
],
[
0.420654296875,
0.31494140625,
0.6044921875,
0.31640625,
0.6044921875,
0.333984375,
0.420654296875,
0.333251953125
],
[
0.2646484375,
0.365478515625,
0.340576171875,
0.365478515625,
0.340576171875,
0.380126953125,
0.2646484375,
0.380126953125
],
[
0.529296875,
0.365478515625,
0.91845703125,
0.3662109375,
0.91845703125,
0.380615234375,
0.529296875,
0.380126953125
],
[
0.56396484375,
0.3837890625,
0.91845703125,
0.3837890625,
0.91845703125,
0.39453125,
0.56396484375,
0.39453125
],
[
0.12744140625,
0.3984375,
0.496826171875,
0.3984375,
0.496826171875,
0.41357421875,
0.12744140625,
0.41357421875
],
[
0.56396484375,
0.397705078125,
0.8759765625,
0.397705078125,
0.8759765625,
0.412353515625,
0.56396484375,
0.412353515625
],
[
0.1070556640625,
0.4140625,
0.495849609375,
0.4140625,
0.495849609375,
0.427978515625,
0.1070556640625,
0.427978515625
],
[
0.53125,
0.417236328125,
0.9189453125,
0.417236328125,
0.9189453125,
0.43115234375,
0.53125,
0.43115234375
],
[
0.1070556640625,
0.429931640625,
0.495849609375,
0.429931640625,
0.495849609375,
0.443115234375,
0.1070556640625,
0.443115234375
],
[
0.56396484375,
0.43310546875,
0.91845703125,
0.43310546875,
0.91845703125,
0.4462890625,
0.56396484375,
0.4462890625
],
[
0.1070556640625,
0.4443359375,
0.495849609375,
0.4443359375,
0.495849609375,
0.458251953125,
0.1070556640625,
0.458251953125
],
[
0.560546875,
0.447509765625,
0.79833984375,
0.4462890625,
0.79833984375,
0.4609375,
0.560546875,
0.462158203125
],
[
0.1070556640625,
0.460205078125,
0.495849609375,
0.460205078125,
0.495849609375,
0.473388671875,
0.1070556640625,
0.473388671875
],
[
0.53125,
0.468505859375,
0.9189453125,
0.468505859375,
0.9189453125,
0.481689453125,
0.53125,
0.481689453125
],
[
0.108642578125,
0.474853515625,
0.494384765625,
0.474853515625,
0.494384765625,
0.488037109375,
0.108642578125,
0.488037109375
],
[
0.10540771484375,
0.4892578125,
0.495849609375,
0.4892578125,
0.495849609375,
0.50244140625,
0.10540771484375,
0.50244140625
],
[
0.53125,
0.490478515625,
0.9208984375,
0.490478515625,
0.9208984375,
0.50390625,
0.53125,
0.50390625
],
[
0.1070556640625,
0.5048828125,
0.351318359375,
0.5048828125,
0.351318359375,
0.51806640625,
0.1070556640625,
0.51806640625
],
[
0.53173828125,
0.50830078125,
0.91650390625,
0.50830078125,
0.91650390625,
0.51806640625,
0.53173828125,
0.51806640625
],
[
0.52783203125,
0.51953125,
0.84716796875,
0.52099609375,
0.84716796875,
0.53515625,
0.52783203125,
0.5341796875
],
[
0.5498046875,
0.53515625,
0.91845703125,
0.537109375,
0.91845703125,
0.55029296875,
0.5498046875,
0.548828125
],
[
0.529296875,
0.5517578125,
0.91650390625,
0.5517578125,
0.91650390625,
0.56494140625,
0.529296875,
0.56494140625
],
[
0.108642578125,
0.5625,
0.2459716796875,
0.5625,
0.2459716796875,
0.57568359375,
0.108642578125,
0.57568359375
],
[
0.53125,
0.56640625,
0.9189453125,
0.56640625,
0.9189453125,
0.580078125,
0.53125,
0.580078125
],
[
0.53125,
0.58349609375,
0.91650390625,
0.58349609375,
0.91650390625,
0.59326171875,
0.53125,
0.59326171875
],
[
0.12744140625,
0.59228515625,
0.495849609375,
0.59228515625,
0.495849609375,
0.60595703125,
0.12744140625,
0.60595703125
],
[
0.53125,
0.5966796875,
0.9189453125,
0.5966796875,
0.9189453125,
0.60986328125,
0.53125,
0.60986328125
],
[
0.108642578125,
0.60791015625,
0.494384765625,
0.60791015625,
0.494384765625,
0.6220703125,
0.108642578125,
0.6220703125
],
[
0.53125,
0.6123046875,
0.9189453125,
0.6123046875,
0.9189453125,
0.62548828125,
0.53125,
0.62548828125
],
[
0.1070556640625,
0.62353515625,
0.494384765625,
0.62353515625,
0.494384765625,
0.63720703125,
0.1070556640625,
0.63720703125
],
[
0.529296875,
0.626953125,
0.68896484375,
0.626953125,
0.68896484375,
0.64013671875,
0.529296875,
0.64013671875
],
[
0.1070556640625,
0.63818359375,
0.496826171875,
0.63818359375,
0.496826171875,
0.6513671875,
0.1070556640625,
0.6513671875
],
[
0.5498046875,
0.6416015625,
0.9189453125,
0.6416015625,
0.9189453125,
0.65576171875,
0.5498046875,
0.65576171875
],
[
0.1070556640625,
0.65380859375,
0.495849609375,
0.65380859375,
0.495849609375,
0.66748046875,
0.1070556640625,
0.66748046875
],
[
0.529296875,
0.6572265625,
0.91845703125,
0.6572265625,
0.91845703125,
0.67041015625,
0.529296875,
0.67041015625
],
[
0.1070556640625,
0.66845703125,
0.495849609375,
0.66845703125,
0.495849609375,
0.681640625,
0.1070556640625,
0.681640625
],
[
0.53125,
0.67236328125,
0.9189453125,
0.67236328125,
0.9189453125,
0.68603515625,
0.53125,
0.68603515625
],
[
0.10540771484375,
0.681640625,
0.495849609375,
0.68310546875,
0.495849609375,
0.69775390625,
0.10540771484375,
0.6962890625
],
[
0.53125,
0.68798828125,
0.9189453125,
0.68798828125,
0.9189453125,
0.7021484375,
0.53125,
0.7021484375
],
[
0.108642578125,
0.69970703125,
0.49267578125,
0.69970703125,
0.49267578125,
0.71044921875,
0.108642578125,
0.71044921875
],
[
0.53173828125,
0.7041015625,
0.9150390625,
0.7041015625,
0.9150390625,
0.7138671875,
0.53173828125,
0.7138671875
],
[
0.1070556640625,
0.71337890625,
0.496826171875,
0.71337890625,
0.496826171875,
0.7265625,
0.1070556640625,
0.7265625
],
[
0.52783203125,
0.71630859375,
0.9189453125,
0.71728515625,
0.9189453125,
0.73193359375,
0.52783203125,
0.73095703125
],
[
0.1070556640625,
0.72900390625,
0.494384765625,
0.72900390625,
0.494384765625,
0.7421875,
0.1070556640625,
0.7421875
],
[
0.53125,
0.73193359375,
0.9150390625,
0.73193359375,
0.9150390625,
0.7470703125,
0.53125,
0.7470703125
],
[
0.10540771484375,
0.7421875,
0.495849609375,
0.74365234375,
0.495849609375,
0.75830078125,
0.10540771484375,
0.7568359375
],
[
0.52783203125,
0.74755859375,
0.9189453125,
0.7470703125,
0.9189453125,
0.76123046875,
0.52783203125,
0.76220703125
],
[
0.108642578125,
0.76025390625,
0.176513671875,
0.76025390625,
0.176513671875,
0.771484375,
0.108642578125,
0.771484375
],
[
0.52783203125,
0.76318359375,
0.59326171875,
0.76318359375,
0.59326171875,
0.77783203125,
0.52783203125,
0.77783203125
],
[
0.1070556640625,
0.7822265625,
0.496826171875,
0.7822265625,
0.496826171875,
0.79541015625,
0.1070556640625,
0.79541015625
],
[
0.1396484375,
0.79736328125,
0.473876953125,
0.79736328125,
0.473876953125,
0.81103515625,
0.1396484375,
0.81103515625
],
[
0.53125,
0.79541015625,
0.7939453125,
0.79541015625,
0.7939453125,
0.81005859375,
0.53125,
0.81005859375
],
[
0.108642578125,
0.818359375,
0.451904296875,
0.818359375,
0.451904296875,
0.83203125,
0.108642578125,
0.83203125
],
[
0.54833984375,
0.82373046875,
0.9189453125,
0.82470703125,
0.9189453125,
0.8388671875,
0.54833984375,
0.83837890625
],
[
0.10540771484375,
0.83837890625,
0.494384765625,
0.8388671875,
0.494384765625,
0.85400390625,
0.10540771484375,
0.85302734375
],
[
0.53125,
0.84033203125,
0.91845703125,
0.84033203125,
0.91845703125,
0.85400390625,
0.53125,
0.85400390625
],
[
0.1396484375,
0.85693359375,
0.2418212890625,
0.85693359375,
0.2418212890625,
0.86865234375,
0.1396484375,
0.86865234375
],
[
0.529296875,
0.85595703125,
0.91845703125,
0.85595703125,
0.91845703125,
0.869140625,
0.529296875,
0.869140625
],
[
0.90185546875,
0.9345703125,
0.93115234375,
0.9345703125,
0.93115234375,
0.9521484375,
0.90185546875,
0.9521484375
],
[
0.0310516357421875,
0.955078125,
0.59716796875,
0.955078125,
0.59716796875,
0.96826171875,
0.0310516357421875,
0.96826171875
],
[
0.86279296875,
0.9541015625,
0.96337890625,
0.95703125,
0.96337890625,
0.9716796875,
0.86279296875,
0.96826171875
],
[
0.0310516357421875,
0.96728515625,
0.1953125,
0.96728515625,
0.1953125,
0.98095703125,
0.0310516357421875,
0.98095703125
],
[
0.88134765625,
0.9697265625,
0.95166015625,
0.9697265625,
0.95166015625,
0.97998046875,
0.88134765625,
0.97998046875
]
] |
[
"An Assembler Driven Verification Methodology (ADVM)",
"John S. MacBeth",
"Dietmar Heinz",
"Verilab GmbH",
"Infineon Technologies AG",
"Lichtenbergerstrasse 8, Garching",
"St. Martin Strasse 76",
"85748 Munich, Germany",
"81609 Munich, Germany",
"john.macbeth@verilab.com",
"dietmar.heinz@infineon.com",
"Ken Gray",
"Verilab GmbH",
"Lichtenbergerstrasse 8, Garching",
"85748 Munich, Germany",
"ken.gray@ verilab.com",
"Abstract",
"Hardware Accelerator The hardware emulator used by",
"boththe customer and the embedded software teams",
"This paper presents an overview of an assembler driven",
"for firmware sign-off (e.g Quickturn and IKOS)",
"verification methodology (ADVM) that was created and im-",
"Bondout Silicon The software development platform with",
"plemented for a chip card project at Infineon Technologies",
"silicon performance. It is also enhanced to include ex-",
"AG [2]. The primary advantage of this methodology is that",
"tra hardware debugging capabilities",
"it enables rapid porting of directed tests to new targets",
"Product Silicon The final product silicon for the customer",
"and derivatives, with only a minimum amount of code re-",
"factoring. As a consequence, considerable verification de-",
"Although this method has a clear advantage in terms of",
"velopment time and effort was saved.",
"crossing many simulation/emulation domains,it is re-",
"stricted by the short comings of directed testing.",
"To verify a complex device using directed testing, many",
"tests must be written in an attempt to cover as many func-",
"1.Introduction",
"tional modes of operation as possible. This means that over",
"timealargecollectionof directed testcodewill bedevel-",
"Presently, the methods used for verifying com-",
"oped and will require re-factoring (i.e. maintenance) with",
"plex microprocessor based System-on-Chip (SoC) de-",
"each change in the specification or when migrating the test",
"vices can take many forms [1]. Such verification meth-",
"code to new derivatives.",
"ods vary from high-level verification languages or tools",
"Writing test code that crosses simulation domains should",
"(such as Verisity's Specman tool and 'e' language or Syn-",
"not present many problems as most platforms should exe-",
"opsys's Vera) to running compiled test code on the de-",
"cute the code in the same way. If they don't then a bug or",
"vice. The advantage of the latter method, running code",
"issue has been found in that particular simulation domain.",
"onthe device,isthatthecode canbe run on any simu-",
"Achallengeemergeshowever.iftheaimistore-targetex",
"lation or emulation platform. This means that the same",
"isting test code to a new derivative and it is this challenge",
"suite of assembler tests can be used to perform func-",
"that the ADVM primarily overcomes. The following sec",
"tional verification of each of the following development",
"tion describes the test environment required to accomplish",
"platforms:",
"this goal.",
"Golden Reference Model The software simulator that is",
"supplied to the customer for software development",
"2. Test Environment Structure",
"HDL-RTL Simulation The HDL design for silicon",
"Abstraction is the key to dealing with change. Creating",
"HDL-Gate Level Simulation The post-synthesis HDL de-",
"an abstraction layer between the test and the physical im-",
"sign for silicon",
"plementation provides a layer that can absorb any changes.",
"@",
"Proceedings of the Design, Automation and Test in Europe Conference and Exhibition (DATE'o5)",
"COMPUTER",
"1530-1591/05$20.00IEEE",
"SOCIETY"
] |
{
"arxiv_id": "0710.4852",
"language": "en",
"page_idx": 0
}
|
||
0710.4852-en-4
|
0710.4852-en-4.jpg
|
Another simple code example is shown in Figure 7.In this case a base functions file (base-functions.asm) is also shown in the abstraction layer and is really a library of functions required by the tests. For this particular example a function that has global scope (i.e. not under the module test environment control) is required by the test. Normally, this code would be included and called directly, but this solution would bypass the abstraction layer and would cause problems if it ever changed (remember that it is not under your control). To solve this issue the function is wrapped by the abstraction layer then this encapsulated function is called from the test. Some examples of typical changes that could occur are; the function name, the input and output parameters or the actual code itself. The abstraction layer does not stop the change taking place but provides a single point to handle it. The solution might be as simple as breaking the link to the global layer and inserting a specific version of the function or it could just be a re-map of the inputs and outputs respectively.
# 5. Conclusions
This paper provided an overview of the assembler driven verification methodology (ADVM) that was deployed as a
fected tests are known). The next type of change that can occur is that a new derivative is produced but this version of the module is now capable of handling more pages. To handle these extra pages the page control field size has increased by one bit. This is not a problem because the globals file can be used to specify derivative specific information (allowed only in the abstraction layer) so such a change can be handled without a re-write of the tests (i.e. the PAGE-FILE-SIZE define can be changed from 5 to 6 for this derivative).
<FigureHere>
Figure 7. Simple Code Example 2
solution and enhancement to Infineon's existing verification
environment.
As with any form of software development (complex test environments are just software projects) the code must be written with a focus on the present requirements but also with an awareness of possible future needs. This requires some up-front thinking about which aspects of the test may change. Such changes may include; the derivative, the target simulation platform, the specification or the focus of the test (control for corner cases). The need to code in this way can be viewed as a disadvantage as it is not possible to get all of this correct first time. As a result the test environment will require some iterative development between the test and abstraction layers.
Clearly, this style of coding introduces some overhead and possibly some redundancy (not all of the predicted changes will happen) but with more readable and controllable code this overhead is acceptable. Consequently, there is an initial time penalty associated with the development of the test environment. However, this time is easily recovered on first reuse with a new target platform or derivative.
Any disadvantages are outweighed by the advantages
that the ADVM offers, namely:
$\bullet$ All changes are resolved by the abstraction layer
$\bullet$ Rapid porting to new derivatives is achieved since the
abstraction layer is inherited by all tests
· Provides a consistent method for the creation of tests
that enables easier training of new personnel and eas-
ier debugging of tests written by another author
$\bullet$ Provides a better method of test control that allows specific corner cases to be investigated (as the need arises)
· Once the base functions for each environment have
been created the test development time is significantly
reduced
$\bullet$ The existing test environment is not lost, but can be re-
placed gradually
$\bullet$ Provides the possibility of generating more complex
test scenarios using high level languages
In conclusion, the application of the ADVM has greatly
improved our existing and future verification efforts.
## 6. Acknowledgements
The authors acknowledge the support given to this work by the SLE88 verification team, namely: Steffen Heinkel, $\dot{\text{Hans-Hubert Kallus, Josef Huber, Wolfgang Gaertner and}}$ Ulrike Pfannkuchen.
66
$\mathrm{COMPUTER}^{10.58}$ $\operatorname{SOCIETY}$
Proceedings of the Design, Automation and Test in Europe Conference and Exhibition (DATE'05)
1530-1591/05 $\$$ 20.00 $IEEE$
|
[
[
0.1070556640625,
0.059356689453125,
0.495849609375,
0.059967041015625,
0.495849609375,
0.07452392578125,
0.1070556640625,
0.0738525390625
],
[
0.529296875,
0.059967041015625,
0.9189453125,
0.059967041015625,
0.9189453125,
0.0738525390625,
0.529296875,
0.0738525390625
],
[
0.1070556640625,
0.07574462890625,
0.495849609375,
0.07574462890625,
0.495849609375,
0.09027099609375,
0.1070556640625,
0.09027099609375
],
[
0.53125,
0.07763671875,
0.61767578125,
0.07763671875,
0.61767578125,
0.0889892578125,
0.53125,
0.0889892578125
],
[
0.1070556640625,
0.091552734375,
0.495849609375,
0.091552734375,
0.495849609375,
0.10479736328125,
0.1070556640625,
0.10479736328125
],
[
0.5498046875,
0.0921630859375,
0.9189453125,
0.0921630859375,
0.9189453125,
0.1060791015625,
0.5498046875,
0.1060791015625
],
[
0.108642578125,
0.1060791015625,
0.495849609375,
0.1060791015625,
0.495849609375,
0.12060546875,
0.108642578125,
0.12060546875
],
[
0.53125,
0.10797119140625,
0.9189453125,
0.10797119140625,
0.9189453125,
0.121826171875,
0.53125,
0.121826171875
],
[
0.108642578125,
0.121826171875,
0.495849609375,
0.121826171875,
0.495849609375,
0.1351318359375,
0.108642578125,
0.1351318359375
],
[
0.53125,
0.125,
0.91845703125,
0.125,
0.91845703125,
0.1351318359375,
0.53125,
0.1351318359375
],
[
0.1070556640625,
0.136962890625,
0.495849609375,
0.136962890625,
0.495849609375,
0.15087890625,
0.1070556640625,
0.15087890625
],
[
0.529296875,
0.1383056640625,
0.9189453125,
0.1383056640625,
0.9189453125,
0.1514892578125,
0.529296875,
0.1514892578125
],
[
0.1070556640625,
0.1514892578125,
0.494384765625,
0.1514892578125,
0.494384765625,
0.1654052734375,
0.1070556640625,
0.1654052734375
],
[
0.529296875,
0.1514892578125,
0.9189453125,
0.15283203125,
0.9189453125,
0.1673583984375,
0.529296875,
0.1666259765625
],
[
0.1070556640625,
0.1673583984375,
0.495849609375,
0.1673583984375,
0.495849609375,
0.18115234375,
0.1070556640625,
0.18115234375
],
[
0.529296875,
0.1666259765625,
0.91845703125,
0.1673583984375,
0.91845703125,
0.1817626953125,
0.529296875,
0.18115234375
],
[
0.1070556640625,
0.1817626953125,
0.494384765625,
0.1817626953125,
0.494384765625,
0.1956787109375,
0.1070556640625,
0.1956787109375
],
[
0.53125,
0.18310546875,
0.9189453125,
0.18310546875,
0.9189453125,
0.1962890625,
0.53125,
0.1962890625
],
[
0.10540771484375,
0.1956787109375,
0.443603515625,
0.1962890625,
0.443603515625,
0.2108154296875,
0.10540771484375,
0.210205078125
],
[
0.52783203125,
0.1976318359375,
0.91845703125,
0.1988525390625,
0.91845703125,
0.212158203125,
0.52783203125,
0.2108154296875
],
[
0.529296875,
0.21337890625,
0.9189453125,
0.21337890625,
0.9189453125,
0.2266845703125,
0.529296875,
0.2266845703125
],
[
0.529296875,
0.2266845703125,
0.91845703125,
0.2279052734375,
0.91845703125,
0.242431640625,
0.529296875,
0.2412109375
],
[
0.111083984375,
0.2392578125,
0.181396484375,
0.2392578125,
0.181396484375,
0.250732421875,
0.111083984375,
0.250732421875
],
[
0.2156982421875,
0.242431640625,
0.3095703125,
0.242431640625,
0.3095703125,
0.24560546875,
0.2156982421875,
0.24560546875
],
[
0.2303466796875,
0.248779296875,
0.26318359375,
0.248779296875,
0.26318359375,
0.255615234375,
0.2303466796875,
0.255615234375
],
[
0.529296875,
0.24365234375,
0.9189453125,
0.24365234375,
0.9189453125,
0.2568359375,
0.529296875,
0.2568359375
],
[
0.2646484375,
0.250732421875,
0.291748046875,
0.250732421875,
0.291748046875,
0.25390625,
0.2646484375,
0.25390625
],
[
0.291748046875,
0.250732421875,
0.298095703125,
0.250732421875,
0.298095703125,
0.25390625,
0.291748046875,
0.25390625
],
[
0.2197265625,
0.25830078125,
0.2467041015625,
0.25830078125,
0.2467041015625,
0.266357421875,
0.2197265625,
0.266357421875
],
[
0.221435546875,
0.25244140625,
0.2279052734375,
0.25,
0.2303466796875,
0.25244140625,
0.22216796875,
0.255615234375
],
[
0.52783203125,
0.2587890625,
0.7060546875,
0.2587890625,
0.7060546875,
0.272705078125,
0.52783203125,
0.272705078125
],
[
0.261474609375,
0.274658203125,
0.2998046875,
0.274658203125,
0.2998046875,
0.2822265625,
0.261474609375,
0.2822265625
],
[
0.54833984375,
0.274658203125,
0.9189453125,
0.274658203125,
0.9189453125,
0.28857421875,
0.54833984375,
0.28857421875
],
[
0.226318359375,
0.28271484375,
0.25732421875,
0.28271484375,
0.25732421875,
0.290283203125,
0.226318359375,
0.290283203125
],
[
0.2303466796875,
0.27587890625,
0.25,
0.27587890625,
0.25,
0.279052734375,
0.2303466796875,
0.279052734375
],
[
0.529296875,
0.290283203125,
0.9189453125,
0.290283203125,
0.9189453125,
0.3037109375,
0.529296875,
0.3037109375
],
[
0.11273193359375,
0.309326171875,
0.2254638671875,
0.309326171875,
0.2254638671875,
0.322509765625,
0.11273193359375,
0.322509765625
],
[
0.53125,
0.3037109375,
0.9189453125,
0.3037109375,
0.9189453125,
0.317626953125,
0.53125,
0.317626953125
],
[
0.295654296875,
0.317626953125,
0.400390625,
0.317626953125,
0.400390625,
0.324462890625,
0.295654296875,
0.324462890625
],
[
0.53125,
0.3193359375,
0.9189453125,
0.3193359375,
0.9189453125,
0.333251953125,
0.53125,
0.333251953125
],
[
0.29736328125,
0.326904296875,
0.33251953125,
0.326904296875,
0.33251953125,
0.333984375,
0.29736328125,
0.333984375
],
[
0.336669921875,
0.326904296875,
0.39794921875,
0.326904296875,
0.39794921875,
0.333984375,
0.336669921875,
0.333984375
],
[
0.529296875,
0.335205078125,
0.9189453125,
0.335205078125,
0.9189453125,
0.348388671875,
0.529296875,
0.348388671875
],
[
0.316162109375,
0.3359375,
0.3505859375,
0.3359375,
0.3505859375,
0.34033203125,
0.316162109375,
0.34033203125
],
[
0.125,
0.330078125,
0.193603515625,
0.33203125,
0.193603515625,
0.34228515625,
0.125,
0.34033203125
],
[
0.12744140625,
0.33837890625,
0.2320556640625,
0.33837890625,
0.2320556640625,
0.348388671875,
0.12744140625,
0.348388671875
],
[
0.325927734375,
0.343505859375,
0.408447265625,
0.343505859375,
0.408447265625,
0.35107421875,
0.325927734375,
0.35107421875
],
[
0.338134765625,
0.352783203125,
0.374267578125,
0.352783203125,
0.374267578125,
0.356689453125,
0.338134765625,
0.356689453125
],
[
0.53125,
0.3515625,
0.9150390625,
0.3515625,
0.9150390625,
0.3623046875,
0.53125,
0.3623046875
],
[
0.3056640625,
0.35986328125,
0.3349609375,
0.35986328125,
0.3349609375,
0.367431640625,
0.3056640625,
0.367431640625
],
[
0.529296875,
0.365478515625,
0.9140625,
0.365478515625,
0.9140625,
0.37890625,
0.529296875,
0.37890625
],
[
0.5498046875,
0.380615234375,
0.91845703125,
0.380615234375,
0.91845703125,
0.39453125,
0.5498046875,
0.39453125
],
[
0.11767578125,
0.386962890625,
0.1968994140625,
0.386962890625,
0.1968994140625,
0.397705078125,
0.11767578125,
0.397705078125
],
[
0.23779296875,
0.389404296875,
0.273681640625,
0.389404296875,
0.273681640625,
0.393310546875,
0.23779296875,
0.393310546875
],
[
0.221435546875,
0.3984375,
0.2467041015625,
0.3984375,
0.2467041015625,
0.401611328125,
0.221435546875,
0.401611328125
],
[
0.2646484375,
0.3984375,
0.275390625,
0.3984375,
0.275390625,
0.401611328125,
0.2646484375,
0.401611328125
],
[
0.52783203125,
0.395751953125,
0.73779296875,
0.396484375,
0.73779296875,
0.410888671875,
0.52783203125,
0.410400390625
],
[
0.22216796875,
0.40478515625,
0.2459716796875,
0.40478515625,
0.2459716796875,
0.412353515625,
0.22216796875,
0.412353515625
],
[
0.253173828125,
0.4072265625,
0.279296875,
0.4072265625,
0.279296875,
0.410400390625,
0.253173828125,
0.410400390625
],
[
0.3544921875,
0.4072265625,
0.381591796875,
0.4072265625,
0.381591796875,
0.410888671875,
0.3544921875,
0.410888671875
],
[
0.5498046875,
0.423583984375,
0.88134765625,
0.423583984375,
0.88134765625,
0.4375,
0.5498046875,
0.4375
],
[
0.2459716796875,
0.412353515625,
0.34716796875,
0.412353515625,
0.34716796875,
0.42236328125,
0.2459716796875,
0.42236328125
],
[
0.54736328125,
0.447509765625,
0.9189453125,
0.447509765625,
0.9189453125,
0.4609375,
0.54736328125,
0.4609375
],
[
0.1732177734375,
0.45703125,
0.431396484375,
0.45703125,
0.431396484375,
0.4716796875,
0.1732177734375,
0.4716796875
],
[
0.560546875,
0.4609375,
0.826171875,
0.462158203125,
0.826171875,
0.4765625,
0.560546875,
0.47607421875
],
[
0.54833984375,
0.48486328125,
0.91845703125,
0.48486328125,
0.91845703125,
0.498779296875,
0.54833984375,
0.498779296875
],
[
0.56201171875,
0.50048828125,
0.91845703125,
0.50048828125,
0.91845703125,
0.5146484375,
0.56201171875,
0.5146484375
],
[
0.12744140625,
0.5068359375,
0.494384765625,
0.5068359375,
0.494384765625,
0.52099609375,
0.12744140625,
0.52099609375
],
[
0.1070556640625,
0.52099609375,
0.496826171875,
0.521484375,
0.496826171875,
0.5361328125,
0.1070556640625,
0.53515625
],
[
0.56201171875,
0.5166015625,
0.8759765625,
0.5166015625,
0.8759765625,
0.52978515625,
0.56201171875,
0.52978515625
],
[
0.1070556640625,
0.537109375,
0.495849609375,
0.537109375,
0.495849609375,
0.55029296875,
0.1070556640625,
0.55029296875
],
[
0.54833984375,
0.53857421875,
0.91845703125,
0.53857421875,
0.91845703125,
0.5517578125,
0.54833984375,
0.5517578125
],
[
0.108642578125,
0.55322265625,
0.496826171875,
0.55322265625,
0.496826171875,
0.56640625,
0.108642578125,
0.56640625
],
[
0.56201171875,
0.5546875,
0.9189453125,
0.5546875,
0.9189453125,
0.568359375,
0.56201171875,
0.568359375
],
[
0.1070556640625,
0.568359375,
0.495849609375,
0.568359375,
0.495849609375,
0.58203125,
0.1070556640625,
0.58203125
],
[
0.56201171875,
0.5693359375,
0.611328125,
0.5693359375,
0.611328125,
0.583984375,
0.56201171875,
0.583984375
],
[
0.1070556640625,
0.58203125,
0.494384765625,
0.58203125,
0.494384765625,
0.59521484375,
0.1070556640625,
0.59521484375
],
[
0.1070556640625,
0.59765625,
0.494384765625,
0.59765625,
0.494384765625,
0.611328125,
0.1070556640625,
0.611328125
],
[
0.54736328125,
0.59228515625,
0.9189453125,
0.59228515625,
0.9189453125,
0.60595703125,
0.54736328125,
0.60595703125
],
[
0.560546875,
0.60693359375,
0.91845703125,
0.60791015625,
0.91845703125,
0.62255859375,
0.560546875,
0.6220703125
],
[
0.1070556640625,
0.61279296875,
0.495849609375,
0.61279296875,
0.495849609375,
0.626953125,
0.1070556640625,
0.626953125
],
[
0.1070556640625,
0.62744140625,
0.495849609375,
0.62744140625,
0.495849609375,
0.6416015625,
0.1070556640625,
0.6416015625
],
[
0.56396484375,
0.62353515625,
0.61767578125,
0.62353515625,
0.61767578125,
0.63525390625,
0.56396484375,
0.63525390625
],
[
0.108642578125,
0.64306640625,
0.495849609375,
0.64306640625,
0.495849609375,
0.6572265625,
0.108642578125,
0.6572265625
],
[
0.5458984375,
0.64599609375,
0.91650390625,
0.646484375,
0.91650390625,
0.6611328125,
0.5458984375,
0.66015625
],
[
0.108642578125,
0.65771484375,
0.496826171875,
0.65771484375,
0.496826171875,
0.671875,
0.108642578125,
0.671875
],
[
0.56201171875,
0.662109375,
0.67236328125,
0.662109375,
0.67236328125,
0.67529296875,
0.56201171875,
0.67529296875
],
[
0.1070556640625,
0.673828125,
0.495849609375,
0.673828125,
0.495849609375,
0.6875,
0.1070556640625,
0.6875
],
[
0.10540771484375,
0.689453125,
0.495849609375,
0.689453125,
0.495849609375,
0.70263671875,
0.10540771484375,
0.70263671875
],
[
0.54833984375,
0.68505859375,
0.91845703125,
0.68505859375,
0.91845703125,
0.69873046875,
0.54833984375,
0.69873046875
],
[
0.108642578125,
0.705078125,
0.495849609375,
0.705078125,
0.495849609375,
0.71533203125,
0.108642578125,
0.71533203125
],
[
0.560546875,
0.70068359375,
0.830078125,
0.70068359375,
0.830078125,
0.7138671875,
0.560546875,
0.7138671875
],
[
0.108642578125,
0.72021484375,
0.494384765625,
0.72021484375,
0.494384765625,
0.73095703125,
0.108642578125,
0.73095703125
],
[
0.54736328125,
0.7236328125,
0.91845703125,
0.724609375,
0.91845703125,
0.7392578125,
0.54736328125,
0.73779296875
],
[
0.1070556640625,
0.734375,
0.495849609375,
0.734375,
0.495849609375,
0.74755859375,
0.1070556640625,
0.74755859375
],
[
0.53125,
0.74072265625,
0.875,
0.74072265625,
0.875,
0.75390625,
0.53125,
0.75390625
],
[
0.1070556640625,
0.75,
0.495849609375,
0.75,
0.495849609375,
0.76318359375,
0.1070556640625,
0.76318359375
],
[
0.1070556640625,
0.7646484375,
0.495849609375,
0.7646484375,
0.495849609375,
0.77783203125,
0.1070556640625,
0.77783203125
],
[
0.10540771484375,
0.77880859375,
0.22216796875,
0.77880859375,
0.22216796875,
0.79248046875,
0.10540771484375,
0.79248046875
],
[
0.53173828125,
0.77880859375,
0.72216796875,
0.77880859375,
0.72216796875,
0.79345703125,
0.53173828125,
0.79345703125
],
[
0.1070556640625,
0.81005859375,
0.240234375,
0.81005859375,
0.240234375,
0.82470703125,
0.1070556640625,
0.82470703125
],
[
0.5498046875,
0.81005859375,
0.91845703125,
0.81005859375,
0.91845703125,
0.82470703125,
0.5498046875,
0.82470703125
],
[
0.53125,
0.8271484375,
0.9140625,
0.8271484375,
0.9140625,
0.8369140625,
0.53125,
0.8369140625
],
[
0.12744140625,
0.8388671875,
0.496826171875,
0.8388671875,
0.496826171875,
0.85400390625,
0.12744140625,
0.85400390625
],
[
0.529296875,
0.84033203125,
0.91845703125,
0.84033203125,
0.91845703125,
0.85400390625,
0.529296875,
0.85400390625
],
[
0.10540771484375,
0.85400390625,
0.498291015625,
0.85498046875,
0.498291015625,
0.869140625,
0.10540771484375,
0.86865234375
],
[
0.52783203125,
0.85400390625,
0.66845703125,
0.85498046875,
0.66845703125,
0.869140625,
0.52783203125,
0.86865234375
],
[
0.904296875,
0.9375,
0.92724609375,
0.9375,
0.92724609375,
0.94970703125,
0.904296875,
0.94970703125
],
[
0.0310516357421875,
0.955078125,
0.59716796875,
0.955078125,
0.59716796875,
0.96826171875,
0.0310516357421875,
0.96826171875
],
[
0.86376953125,
0.955078125,
0.96142578125,
0.95849609375,
0.96142578125,
0.9697265625,
0.86376953125,
0.96630859375
],
[
0.0318603515625,
0.96728515625,
0.1953125,
0.96728515625,
0.1953125,
0.98095703125,
0.0318603515625,
0.98095703125
],
[
0.88134765625,
0.9697265625,
0.95166015625,
0.9697265625,
0.95166015625,
0.97998046875,
0.88134765625,
0.97998046875
]
] |
[
"fected tests are known). The next type of change that can",
"solution and enhancement to Infineon's existing verification",
"occur is that a new derivative is produced but this ver-",
"environment.",
"sion of the module is now capable of handling more",
"As with any form of software development (complex test",
"pages.To handle these extra pages the page control",
"environments are just software projects) the code must be",
"field size has increased by one bit. This is not a prob-",
"written with a focus on the present requirements but also",
"lem because the globals file can be used to specify",
"with an awareness of possible future needs. This requires",
"derivative specific information (allowed only in the ab",
"some up-front thinking about which aspects of the test may",
"straction layer) so such a change can be handled without",
"change. Such changes may include; the derivative, the tar-",
"a re-Write of the tests (i.e.the PAGEFILE_SIZE de-",
"get simulation platform, the specification or the focus of the",
"fine can be changed from 5 to 6 for this derivative).",
"test (control for corner cases).The need to code in this way",
"can be viewed as a disadvantage as it is not possible to get",
"all of this correct first time. As a result the test environ-",
"TestLayer",
".omon+oo+",
"ICLUDE",
"ment will require some iterative development between the",
"nn",
"-",
"main",
"",
"test and abstraction layers.",
"CallAdds",
"Clearly, this style of coding introduces some overhead",
"RETURM",
"AATT",
"and possibly some redundancy (not all of the predicted",
"Abstraction Layer",
"changes will happen) but with more readable and control-",
":Base Functions.as",
"lable code this overhead is acceptable. Consequently, there",
"INCLUD",
"Globals.ir",
"is an initial time penalty associated with the development",
"Tni+r",
"Giobals.ir",
".DEFINE CallAddr A12",
"CallAddr,Es",
"1Ha",
"of thetest environment. However,this time is easily recov",
"RETURN",
"ered on first reuse with a new target platform or derivative.",
"Any disadvantages are outweighed by the advantages",
"Global Layer",
"haddan",
"Tn+",
"",
"that the ADVM offers, namely:",
"LOAD",
"=111e",
"TTE",
"All changes are resolved by the abstraction layer",
"[ADDR],ValueForRe",
". Rapid porting to new derivatives is achieved since the",
"Figure7.SimpleCodeExample2",
"abstraction layer is inherited by all tests",
". Provides a consistent method for the creation of tests",
"that enables easier training of new personnel and eas-",
"Another simple code example is shown in Figure 7. In",
"this case a base functions file (base_functions.asm) is also",
"ier debugging of tests written by another author",
"shown in the abstraction layer and is really a library of func-",
". Provides a better method of test control that allows",
"tions required by the tests. For this particular example a",
"specific corner cases to be investigated (as the need",
"function that has global scope (i.e. not under the module",
"arises)",
"test environment control) is required by the test. Normally",
"this code would be included and called directly, but this so-",
". Once the base functions for each environment have",
"been created the test development time is significantly",
"lution would bypass the abstraction layer and would cause",
"problems if it ever changed (remember that it is not under",
"reduced",
"your control). To solve this issue the function is wrapped",
". The existing test environment is not lost, but can be re-",
"by the abstraction layer then this encapsulated function is",
"placed gradually",
"called from the test. Some examples of typical changes that",
"could occur are; the function name, the input and output pa-",
". Provides the possibility of generating more complex",
"rametersortheactualcodeitself.Theabstractionlayerdoes",
"test scenarios using high level languages",
"not stopthechangetakingplacebutprovidesa singlepoint",
"In conclusion, the application of the ADVM has greatly",
"to handle it. The solution might be as simple as breaking the",
"improved our existing and future verification efforts.",
"link to the global layer and inserting a specific version of the",
"function or it could just be a re-map of the inputs and out-",
"puts respectively.",
"6.Acknowledgements",
"5. Conclusions",
"The authors acknowledge the support given to this work",
"bytheSLE88 verificationteam.namely:SteffenHeinkel",
"This paper provided an overview of the assembler driven",
"Hans-Hubert Kallus, Josef Huber, Wolfgang Gaertner and",
"verification methodology (ADVM) that was deployed as a",
"Ulrike Pfannkuchen.",
"",
"Proceedings of the Design, Automation and Test in Europe Conference and Exhibition (DATE'o5)",
"COMPUTER",
"1530-1591/05$20.00IEEE",
"SOCIETY"
] |
{
"arxiv_id": "0710.4852",
"language": "en",
"page_idx": 4
}
|
||
2308.05320-en-0
|
2308.05320-en-0.jpg
|
GENERATING TRANSFERABLE AND STEALTHY ADVERSARIAL PATCH VIA ATTENTION-GUIDED ADVERSARIAL INPAINTING
Yanjie Li$^1$, Mingxing Duan$^2$, Xuelong Dai $^3,\textbf{Bin Xiao}^{1*}$ 1. Computing Department, Hong Kong Polytechnic University 2. School of Information Science and Engineering, Hunan University
yanjie.li@connect.polyu.hk, xuelong.dai@connect.polyu.hk,
duanmingxing@hnu.edu.cn, b.xiao@polyu.edu.hk
### ABSTRACT
Adversarial patch attacks can fool the face recognition (FR) models via small patches. However, previous adversarial patch attacks often result in unnatural patterns that are easily noticeable. Generating transferable and stealthy adversarial patches that can efficiently deceive the black-box FR models while having good camouflage is challenging because of the huge stylistic difference between the source and target images. To generate transferable, natural-looking, and stealthy adversarial patches, we propose an innovative two-stage attack called $Ad\nu-Inpainting$, which extracts style features and identity features from the attacker and target faces, respectively and then fills the patches with misleading and inconspicuous content guided by attention maps. In the first stage, we extract multi-scale style embeddings by a pyramid-like network and identity embeddings by a pretrained FR model and propose a novel Attention-guided Adaptive Instance Normalization layer (AAIN) to merge them via background-patch cross-attention maps. The proposed layer can adaptively fuse identity and style embeddings by fully exploiting priority contextual information. In the second stage, we design an Adversarial Patch Refinement Network ( APR-Net) with a novel boundary variance loss, a spatial discounted reconstruction loss, and a perceptual loss to boost the stealthiness further. Experiments demonstrate that our attack can generate adversarial patches with improved visual quality, better stealthiness, and stronger transferability than state-of-the-art adversarial patch attacks and semantic attacks.
1 INTRODUCTION
Deep neural networks are widely used in many applications, such as image classification and face recognition (FR). Previous research has shown thât they are vulnerable to adversarial examples that are created by adding elaborate noises, which are invisible to human eyes but can deceive the neural network (Madry et al., 2017; Carlini & Wagner, 2017). However, such attacks are impractical in real-world implementation because pixel-wise manipulation is usually infeasible. To attack realworld applications like face recognition systems, most researchers use adversarial patches rather than pixel-wise noises, which limits adversarial noises to a small area to be printable and can be physically used to attack FR systems (Komkov &Petiushko, 2021; Yang et al., 2020).
Stealthiness and transferability to black-box models are two crucial requirements for practical adversarial patch attacks. Although some methods have been proposed to enhance the transferability of adversarial patches (Wang et al., 2021; Dong et al., 2018; Zhao et al., 2021; Ma et al., 2023), they often produce unnatural styles and noticeable boundaries, making the patches easily detectable (see Figure 1(c-g)). Xiao et al. (2021) found that sampling from the latent space of generative models can enhance transferability. However, their method also did not take the camouflage into consideration, making the generated patches conspicuous (see Figure 1(f)).
${^*}{\mathrm{Bin~Xiao~is~the~corresponding~author}}.$
1
|
[
[
0.17724609375,
0.10162353515625,
0.82177734375,
0.10162353515625,
0.82177734375,
0.11932373046875,
0.17724609375,
0.11932373046875
],
[
0.1748046875,
0.1275634765625,
0.82177734375,
0.126953125,
0.82177734375,
0.14453125,
0.1748046875,
0.1458740234375
],
[
0.176513671875,
0.1546630859375,
0.4150390625,
0.1546630859375,
0.4150390625,
0.169189453125,
0.176513671875,
0.169189453125
],
[
0.1854248046875,
0.1925048828125,
0.61279296875,
0.1912841796875,
0.61279296875,
0.212158203125,
0.1854248046875,
0.21337890625
],
[
0.1846923828125,
0.2108154296875,
0.5947265625,
0.2108154296875,
0.5947265625,
0.2247314453125,
0.1846923828125,
0.2247314453125
],
[
0.1829833984375,
0.2259521484375,
0.63623046875,
0.2259521484375,
0.63623046875,
0.2392578125,
0.1829833984375,
0.2392578125
],
[
0.1829833984375,
0.2392578125,
0.73046875,
0.238037109375,
0.73046875,
0.25244140625,
0.1829833984375,
0.25390625
],
[
0.1871337890625,
0.255126953125,
0.60888671875,
0.255126953125,
0.60888671875,
0.26513671875,
0.1871337890625,
0.26513671875
],
[
0.01715087890625,
0.26953125,
0.05963134765625,
0.26953125,
0.063720703125,
0.4912109375,
0.021240234375,
0.4912109375
],
[
0.453369140625,
0.302978515625,
0.54736328125,
0.302978515625,
0.54736328125,
0.317626953125,
0.453369140625,
0.317626953125
],
[
0.2344970703125,
0.337158203125,
0.765625,
0.337158203125,
0.765625,
0.35107421875,
0.2344970703125,
0.35107421875
],
[
0.2344970703125,
0.3515625,
0.765625,
0.3515625,
0.765625,
0.365478515625,
0.2344970703125,
0.365478515625
],
[
0.233642578125,
0.365478515625,
0.7646484375,
0.365478515625,
0.7646484375,
0.37890625,
0.233642578125,
0.37890625
],
[
0.2344970703125,
0.37890625,
0.7646484375,
0.37890625,
0.7646484375,
0.392578125,
0.2344970703125,
0.392578125
],
[
0.233642578125,
0.393310546875,
0.7646484375,
0.393310546875,
0.7646484375,
0.4072265625,
0.233642578125,
0.4072265625
],
[
0.2320556640625,
0.406005859375,
0.7646484375,
0.4072265625,
0.7646484375,
0.421630859375,
0.2320556640625,
0.42041015625
],
[
0.236083984375,
0.42236328125,
0.7646484375,
0.42236328125,
0.7646484375,
0.43310546875,
0.236083984375,
0.43310546875
],
[
0.233642578125,
0.43505859375,
0.7646484375,
0.43505859375,
0.7646484375,
0.448974609375,
0.233642578125,
0.448974609375
],
[
0.233642578125,
0.448974609375,
0.765625,
0.448974609375,
0.765625,
0.462158203125,
0.233642578125,
0.462158203125
],
[
0.2344970703125,
0.462158203125,
0.76708984375,
0.462158203125,
0.76708984375,
0.47607421875,
0.2344970703125,
0.47607421875
],
[
0.0228729248046875,
0.479736328125,
0.05963134765625,
0.479736328125,
0.05963134765625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.233642578125,
0.47607421875,
0.765625,
0.47607421875,
0.765625,
0.4892578125,
0.233642578125,
0.4892578125
],
[
0.233642578125,
0.4892578125,
0.765625,
0.4892578125,
0.765625,
0.50244140625,
0.233642578125,
0.50244140625
],
[
0.2344970703125,
0.50390625,
0.765625,
0.50390625,
0.765625,
0.517578125,
0.2344970703125,
0.517578125
],
[
0.233642578125,
0.517578125,
0.765625,
0.5166015625,
0.765625,
0.53076171875,
0.233642578125,
0.5322265625
],
[
0.2320556640625,
0.52978515625,
0.765625,
0.53076171875,
0.765625,
0.54541015625,
0.2320556640625,
0.54443359375
],
[
0.236083984375,
0.546875,
0.76318359375,
0.546875,
0.76318359375,
0.556640625,
0.236083984375,
0.556640625
],
[
0.2344970703125,
0.56103515625,
0.765625,
0.56103515625,
0.765625,
0.5712890625,
0.2344970703125,
0.5712890625
],
[
0.2320556640625,
0.5712890625,
0.765625,
0.57275390625,
0.765625,
0.5869140625,
0.2320556640625,
0.5859375
],
[
0.2344970703125,
0.5869140625,
0.765625,
0.5869140625,
0.765625,
0.60107421875,
0.2344970703125,
0.60107421875
],
[
0.2344970703125,
0.60107421875,
0.7646484375,
0.60107421875,
0.7646484375,
0.6142578125,
0.2344970703125,
0.6142578125
],
[
0.178955078125,
0.6513671875,
0.1846923828125,
0.6513671875,
0.1846923828125,
0.65771484375,
0.178955078125,
0.65771484375
],
[
0.19775390625,
0.646484375,
0.338134765625,
0.6484375,
0.338134765625,
0.66357421875,
0.19775390625,
0.6611328125
],
[
0.1748046875,
0.6796875,
0.82421875,
0.6796875,
0.82421875,
0.6943359375,
0.1748046875,
0.6943359375
],
[
0.176513671875,
0.6962890625,
0.82275390625,
0.6962890625,
0.82275390625,
0.70703125,
0.176513671875,
0.70703125
],
[
0.1748046875,
0.708984375,
0.82421875,
0.708984375,
0.82421875,
0.72265625,
0.1748046875,
0.72265625
],
[
0.1748046875,
0.72265625,
0.82421875,
0.72265625,
0.82421875,
0.736328125,
0.1748046875,
0.736328125
],
[
0.1748046875,
0.736328125,
0.82275390625,
0.736328125,
0.82275390625,
0.75,
0.1748046875,
0.75
],
[
0.1748046875,
0.75048828125,
0.82275390625,
0.75048828125,
0.82275390625,
0.7646484375,
0.1748046875,
0.7646484375
],
[
0.1748046875,
0.7646484375,
0.82421875,
0.7646484375,
0.82421875,
0.77783203125,
0.1748046875,
0.77783203125
],
[
0.1732177734375,
0.77783203125,
0.736328125,
0.7763671875,
0.736328125,
0.79150390625,
0.1732177734375,
0.79248046875
],
[
0.1748046875,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.1748046875,
0.81298828125
],
[
0.1748046875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.8271484375,
0.1748046875,
0.8271484375
],
[
0.1748046875,
0.8271484375,
0.82275390625,
0.8271484375,
0.82275390625,
0.84033203125,
0.1748046875,
0.84033203125
],
[
0.176513671875,
0.84228515625,
0.82275390625,
0.84228515625,
0.82275390625,
0.85302734375,
0.176513671875,
0.85302734375
],
[
0.1748046875,
0.85498046875,
0.82421875,
0.85498046875,
0.82421875,
0.86865234375,
0.1748046875,
0.86865234375
],
[
0.1748046875,
0.86865234375,
0.82275390625,
0.86865234375,
0.82275390625,
0.8818359375,
0.1748046875,
0.8818359375
],
[
0.176513671875,
0.8837890625,
0.56787109375,
0.8837890625,
0.56787109375,
0.89453125,
0.176513671875,
0.89453125
],
[
0.193603515625,
0.91015625,
0.423095703125,
0.9111328125,
0.423095703125,
0.92529296875,
0.193603515625,
0.9248046875
],
[
0.494384765625,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"GENERATING TRANSFERABLE AND STEALTHY AD-",
"VERSARIAL PATCH VIA ATTENTION-GUIDED ADVER-",
"SARIAL INPAINTING",
"Yanjie Li', Mingxing Duan2, Xuelong Dai3, Bin Xiao 1*",
"1. Computing Department, Hong Kong Polytechnic University",
"2. School of Information Science and Engineering,Hunan University",
"yanjie.li@connect.polyu.hk,xuelong.dai@connect.polyu.hk,",
"duanmingxing@hnu.edu.cn,b.xiao@polyu.edu.hk",
"[S]",
"ABSTRACT",
"Adversarial patch attacks can fool the face recognition (FR) models via small",
"patches. However, previous adversarial patch attacks often result in unnatural",
"patterns that are easily noticeable. Generating transferable and stealthy adver-",
"sarial patches that can efficiently deceive the black-box FR models while hav-",
"ing good camouflage is challenging because of the huge stylistic difference be-",
"tween the source and target images. To generate transferable, natural-looking,",
"andstealthyadversarialpatches,weproposeaninnovativetwo-stageattackcalled",
"Adv-Inpainting, which extracts style features and identity features from the at-",
"tacker and target faces, respectively and then fills the patches with misleading and",
"inconspicuous content guided by attention maps. In the first stage, we extract",
"25801",
"multi-scale style embeddings by a pyramid-like network and identity embeddings",
"by a pretrained FR model and propose a novel Attention-guided Adaptive Instance",
"Normalization layer (AAIN) to merge them via background-patch cross-attention",
"maps. The proposed layer can adaptively fuse identity and style embeddings by",
"fully exploiting priority contextual information. In the second stage, we design",
"anAdversarialPatchRefinementNetwork(APR-Net)withanovelboundaryvari-",
"anceloss,a spatialdiscountedreconstructionloss,andaperceptuallosstoboost",
"the stealthiness further. Experiments demonstrate that our attack can generate ad-",
"versarial patches with improved visual quality, better stealthiness, and stronger",
"transferability than state-of-the-art adversarial patch attacks and semantic attacks.",
"1",
"INTRODUCTION",
"Deep neural networks are widely used in many applications, such as image classification and face",
"recognition(FR).Previousresearchhas shown that they arevulnerableto adversarial examplesthat",
"are created by adding elaborate noises, which are invisible to human eyes but can deceive the neural",
"network (Madry et al., 2017; Carlini & Wagner, 2017).However, such attacks are impractical in",
"real-world implementation because pixel-wise manipulation is usually infeasible. To attack real-",
"world applications like face recognition systems, most researchers use adversarial patches rather",
"than pixel-wise noises, which limits adversarial noises to a small area to be printable and can be",
"physically used to attack FR systems (Komkov & Petiushko, 2021; Yang et al., 2020).",
"Stealthiness and transferability to black-box models are two crucial requirements for practical ad-",
"versarial patch attacks. Although some methods have been proposed to enhance the transferability",
"of adversarial patches (Wang et al., 2021; Dong et al., 2018; Zhao et al., 2021; Ma et al., 2023), they",
"oftenproduceunnaturalstylesandnoticeableboundaries,makingthepatcheseasilydetectable(see",
"Figure 1(c-g). Xiao et al. (2021) found that sampling from the latent space of generative models can",
"enhance transferability. However, their method also did not take the camouflage into consideration,",
"makingthegeneratedpatchesconspicuous(seeFigure1(f))",
"*BinXiao isthecorresponding author.",
"1"
] |
{
"arxiv_id": "2308.05320",
"language": "en",
"page_idx": 0
}
|
||
2308.05320-en-1
|
2308.05320-en-1.jpg
|
<FigureHere>
$10urs$
(a)
Target (c)Adv-Glasses (d)Adv-Hat (e)
(f)GenAP (g)
Figure 1: IIlustration of different adversarial patch attacks, including Adv-Glasses (Sharif et al., 2016), Adv-Hat (Komkov & Petiushko, 2021), TIDIM (Xie et al., 2019b)), GenAP (Xiao et al. 2021)), SMAP (Ma et al., 2023) and ours. The adversarial patches generated by Adv-Inpainting have higher transferability and stealthiness than previous state-of-the-art adversarial patch attacks.
Is there an inherent contradiction between stealthiness and transferability? Recently, researchers have found that editing deep features can boost transferability (Inkawhich et al., 2019; Jia et al., 2022). Based on these findings, we propose that the paradox between stealthiness and transferability can be effectively addressed by deep feature manipulation. However, because of the huge property difference between the source and target image, like age, gender, skin color, and expressions, blending the stylistic difference seamlessly while deceiving the FR model through a small patch remains a significant challenge. To improve attack transferability and patch stealthiness simultaneously, we propose a novel attack called $\hat{A}d\nu$-Inpainting by employing two-stage attention-guided deep feature manipulation. Our method merges the deep features of the original and target face images while considering the context through an attention-guided adaptive instance normalization (AAIN) layer, which can fuse the identity features of the target image into the source image while maintaining the features not related to the identity, such as expressions and postures. Furthermore, compared with previous feature space attacks (Jia et al., 2022), the stealthiness of our attack is measured by the perceptual similarity (Zhang et al., 2018), which is more consistent with human perception and can be optimized simultaneously with adversary loss, making the training process more stable
Adv-Inpainting consists of two stages. The first stage contains an attention-guided StyleGAN-based generative model (Att-StyleGAN) with a new AAIN layer to generate reasonable and transferable adversarial patches. To further improve the camouflage of the adversarial patches, we append a second stage named Adversarial Patch Refinement Network (APR-Net) to blend the adversarial patch with the source image seamlessly while maintaining the adversarial attributes. An example of the generated patch is shown in Figure 1(h). In summary, our contributions are as follows:
· We are the first to realize a transferable and stealthy adversarial patch attack. Our research indicates that filling reasonable content into the masked region through attention-guided deep feature manipulation can significantly enhance the transferability of the adversarial patches. Compared with previous attack methods, Adv-Inpainting is superior in generating photorealistic, stealthy, and highly transferable adversarial patches.
·We design a novel two-stage coarse-to-fine attack framework. The first stage employs a novel AAIN layer, which has a new mechanism to combine identity and style features. It can fully exploit prior information of the source image to generate transferable and stylistically consistent adversarial patches. In the second stage, dual spatial attention layers are utilized to fine-tune the adversarial patch to improve concealment further.
$\bullet$ Our research pioneers the adoption of perception distance to enhance the stealthiness of the generated patches and stabilize the training process of adversarial attacks. Additionally, we introduce a novel boundary variance loss that facilitates the fine-tuning of the patch, ensuring its coherency with the background.
$\bullet$ Experiments on various FR models demonstrate our method's effectiveness. Our approach significantly improves transferability and stealthiness and surpasses the state-of-the-art attack by a large margin. Moreover, our trained model can directly generate adversarial patches end-to-end, which is more efficient than previous generative attacks.
2
|
[
[
0.1805419921875,
0.1634521484375,
0.50341796875,
0.1634521484375,
0.50341796875,
0.1767578125,
0.1805419921875,
0.1767578125
],
[
0.498291015625,
0.1654052734375,
0.57275390625,
0.1654052734375,
0.57275390625,
0.175537109375,
0.498291015625,
0.175537109375
],
[
0.5849609375,
0.1654052734375,
0.650390625,
0.1654052734375,
0.650390625,
0.1767578125,
0.5849609375,
0.1767578125
],
[
0.6650390625,
0.1654052734375,
0.7294921875,
0.1654052734375,
0.7294921875,
0.1767578125,
0.6650390625,
0.1767578125
],
[
0.75634765625,
0.1654052734375,
0.810546875,
0.1654052734375,
0.810546875,
0.1767578125,
0.75634765625,
0.1767578125
],
[
0.1748046875,
0.1944580078125,
0.8203125,
0.1944580078125,
0.8203125,
0.2083740234375,
0.1748046875,
0.2083740234375
],
[
0.1748046875,
0.2083740234375,
0.82275390625,
0.208984375,
0.82275390625,
0.2235107421875,
0.1748046875,
0.222900390625
],
[
0.1748046875,
0.2235107421875,
0.82275390625,
0.2235107421875,
0.82275390625,
0.2374267578125,
0.1748046875,
0.2374267578125
],
[
0.1732177734375,
0.236083984375,
0.81640625,
0.238037109375,
0.81640625,
0.251953125,
0.1732177734375,
0.25
],
[
0.1748046875,
0.2890625,
0.82275390625,
0.2890625,
0.82275390625,
0.302978515625,
0.1748046875,
0.302978515625
],
[
0.1748046875,
0.3037109375,
0.82275390625,
0.3037109375,
0.82275390625,
0.317626953125,
0.1748046875,
0.317626953125
],
[
0.1732177734375,
0.31640625,
0.82275390625,
0.317626953125,
0.82275390625,
0.33203125,
0.1732177734375,
0.330810546875
],
[
0.1732177734375,
0.330810546875,
0.82177734375,
0.33203125,
0.82177734375,
0.34521484375,
0.1732177734375,
0.3447265625
],
[
0.17724609375,
0.3466796875,
0.8203125,
0.3466796875,
0.8203125,
0.356689453125,
0.17724609375,
0.356689453125
],
[
0.1748046875,
0.359130859375,
0.82275390625,
0.359130859375,
0.82275390625,
0.37255859375,
0.1748046875,
0.37255859375
],
[
0.176513671875,
0.375,
0.82275390625,
0.375,
0.82275390625,
0.385009765625,
0.176513671875,
0.385009765625
],
[
0.1732177734375,
0.386962890625,
0.82275390625,
0.386474609375,
0.82275390625,
0.399658203125,
0.1732177734375,
0.40087890625
],
[
0.1732177734375,
0.40087890625,
0.82177734375,
0.40087890625,
0.82177734375,
0.4140625,
0.1732177734375,
0.4140625
],
[
0.17236328125,
0.4140625,
0.82275390625,
0.415283203125,
0.82275390625,
0.4287109375,
0.17236328125,
0.427978515625
],
[
0.1748046875,
0.429931640625,
0.8203125,
0.429931640625,
0.8203125,
0.43994140625,
0.1748046875,
0.43994140625
],
[
0.1748046875,
0.442626953125,
0.82275390625,
0.442626953125,
0.82275390625,
0.455810546875,
0.1748046875,
0.455810546875
],
[
0.1732177734375,
0.45703125,
0.82421875,
0.455810546875,
0.82421875,
0.4697265625,
0.1732177734375,
0.47021484375
],
[
0.17724609375,
0.4716796875,
0.82275390625,
0.4716796875,
0.82275390625,
0.481689453125,
0.17724609375,
0.481689453125
],
[
0.1748046875,
0.484130859375,
0.7646484375,
0.484130859375,
0.7646484375,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.176513671875,
0.50390625,
0.82275390625,
0.50390625,
0.82275390625,
0.517578125,
0.176513671875,
0.517578125
],
[
0.1748046875,
0.51806640625,
0.82275390625,
0.51806640625,
0.82275390625,
0.5322265625,
0.1748046875,
0.5322265625
],
[
0.176513671875,
0.53271484375,
0.82275390625,
0.53271484375,
0.82275390625,
0.546875,
0.176513671875,
0.546875
],
[
0.1748046875,
0.546875,
0.82421875,
0.546875,
0.82421875,
0.56005859375,
0.1748046875,
0.56005859375
],
[
0.1748046875,
0.56005859375,
0.826171875,
0.56005859375,
0.826171875,
0.57373046875,
0.1748046875,
0.57373046875
],
[
0.1748046875,
0.57470703125,
0.765625,
0.57470703125,
0.765625,
0.58837890625,
0.1748046875,
0.58837890625
],
[
0.2197265625,
0.60986328125,
0.82275390625,
0.60986328125,
0.82275390625,
0.62353515625,
0.2197265625,
0.62353515625
],
[
0.233642578125,
0.625,
0.82421875,
0.625,
0.82421875,
0.63818359375,
0.233642578125,
0.63818359375
],
[
0.233642578125,
0.63818359375,
0.82421875,
0.63818359375,
0.82421875,
0.6513671875,
0.233642578125,
0.6513671875
],
[
0.2320556640625,
0.65087890625,
0.82275390625,
0.6513671875,
0.82275390625,
0.66650390625,
0.2320556640625,
0.66552734375
],
[
0.2320556640625,
0.66650390625,
0.67333984375,
0.66650390625,
0.67333984375,
0.6796875,
0.2320556640625,
0.6796875
],
[
0.2181396484375,
0.693359375,
0.82275390625,
0.6943359375,
0.82275390625,
0.708984375,
0.2181396484375,
0.70849609375
],
[
0.233642578125,
0.71044921875,
0.82275390625,
0.71044921875,
0.82275390625,
0.7236328125,
0.233642578125,
0.7236328125
],
[
0.2344970703125,
0.724609375,
0.8203125,
0.724609375,
0.8203125,
0.73486328125,
0.2344970703125,
0.73486328125
],
[
0.2344970703125,
0.7373046875,
0.82275390625,
0.73779296875,
0.82275390625,
0.75,
0.2344970703125,
0.74853515625
],
[
0.233642578125,
0.75048828125,
0.7158203125,
0.75048828125,
0.7158203125,
0.7646484375,
0.233642578125,
0.7646484375
],
[
0.221435546875,
0.77978515625,
0.82275390625,
0.77978515625,
0.82275390625,
0.79443359375,
0.221435546875,
0.79443359375
],
[
0.2344970703125,
0.79443359375,
0.82275390625,
0.79443359375,
0.82275390625,
0.80810546875,
0.2344970703125,
0.80810546875
],
[
0.2344970703125,
0.8095703125,
0.82275390625,
0.8095703125,
0.82275390625,
0.82275390625,
0.2344970703125,
0.82275390625
],
[
0.233642578125,
0.82275390625,
0.525390625,
0.82275390625,
0.525390625,
0.8359375,
0.233642578125,
0.8359375
],
[
0.2197265625,
0.8515625,
0.82177734375,
0.8515625,
0.82177734375,
0.86572265625,
0.2197265625,
0.86572265625
],
[
0.233642578125,
0.8662109375,
0.82177734375,
0.8662109375,
0.82177734375,
0.8798828125,
0.233642578125,
0.8798828125
],
[
0.2320556640625,
0.8798828125,
0.82275390625,
0.8798828125,
0.82275390625,
0.89306640625,
0.2320556640625,
0.89306640625
],
[
0.233642578125,
0.89306640625,
0.73388671875,
0.89306640625,
0.73388671875,
0.9072265625,
0.233642578125,
0.9072265625
],
[
0.494384765625,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.494384765625,
0.96044921875
]
] |
[
"(a) Attacker (b) Target (c) Adv-Glasses (d) Adv-Hat",
"(e)TIDIM",
"(f)GenAP",
"(g) SMAP",
"(h)Ours",
"Figure I: Illustration of different adversarial patch attacks, including Adv-Glasses (Sharif et al.",
"2016), Adv-Hat (Komkov & Petiushko, 2021), TIDIM (Xie et al., 2019b), GenAP (Xiao et al.,",
"2021), SMAP (Ma et al., 2023) and ours. The adversarial patches generated by Adv-Inpainting",
"have higher transferability and stealthiness than previous state-of-the-art adversarial patch attacks.",
"Is there an inherent contradiction between stealthiness and transferability? Recently, researchers",
"have found that editing deep features can boost transferability (Inkawhich et al., 2019; Jia et al.,",
"2022). Based on these findings, we propose that the paradox between stealthiness and transferability",
"can be effectively addressed by deep feature manipulation.However, because of the huge property",
"differencebetweenthesourceandtargetimage,likeage,gender,skincolor.and expressions.blend",
"ing the stylistic difference seamlessly while deceiving the FR model through a small patch remains",
"a significantchallenge.Toimproveattacktransferabilityandpatchstealthiness simultaneously,we",
"propose a novel attack called Adv-Inpainting by employing two-stage attention-guided deep feature",
"manipulation. Our method merges the deep features of the original and target face images while",
"considering the context through an attention-guided adaptive instance normalization (AAIN) layer",
"whichcanfusetheidentityfeaturesofthetargetimageintothe source imagewhilemaintaining the",
"features not related to the identity, such as expressions and postures. Furthermore, compared with",
"previous featurespaceattacks (Jia et al.,2022),the stealthinessof ourattackis measured bythe",
"perceptualsimilarity(Zhangetal.,2o18).whichismoreconsistentwithhumanperceptionandcan",
"be optimized simultaneously with adversary loss, making the training process more stable.",
"Adv-Inpainting consists of two stages. The first stage contains an attention-guided StyleGAN-based",
"generative model (Att-StyleGAN) with a new AAIN layer to generate reasonable and transferable",
"adversarial patches. To further improve the camouflage of the adversarial patches, we append a",
"second stage named Adversarial Patch Refinement Network (APR-Net) to blend the adversarial",
"patch with the source image seamlessly while maintaining the adversarial attributes.An example of",
"the generated patch is shown in Figure 1(h). In summary, our contributions are as follows:",
": We are the first to realize a transferable and stealthy adversarial patch attack. Our research",
"indicates that filling reasonable content into the masked region through attention-guided",
"deep feature manipulation can significantly enhance the transferability of the adversarial",
"patches.Compared with previous attack methods,Adv-Inpainting is superior in generating",
"photorealistic, stealthy, and highly transferable adversarial patches.",
". We design a novel two-stage coarse-to-fine attack framework. The first stage employs a",
"novel AAIN layer, which has a new mechanism to combine identity and style features. It",
"canfullyexploitpriorinformationofthesourceimagetogeneratetransferableandstylis-",
"ticallyconsistentadversarialpatches.Inthesecondstage,dual spatial attentionlayersare",
"utilized to fine-tune the adversarial patch to improve concealment further.",
": Our research pioneers the adoption of perception distance to enhance the stealthiness of the",
"generated patches and stabilize the training process of adversarial attacks. Additionally,",
"we introduce a novel boundary variance loss that facilitates the fine-tuning of the patch,",
"ensuring its coherency with the background.",
": Experiments on various FR models demonstrate our method's effectiveness. Our approach",
"significantly improves transferability and stealthiness and surpasses the state-of-the-art at-",
"tack by a large margin. Moreover, our trained model can directly generate adversarial",
"patches end-to-end, which is more efficient than previous generative attacks.",
"2"
] |
{
"arxiv_id": "2308.05320",
"language": "en",
"page_idx": 1
}
|
||
2308.05320-en-2
|
2308.05320-en-2.jpg
|
## 2 RELATED WORK
Transferable Adversarial Patch Attack. Different from the traditional strategy of finding adversarial examples, adversarial patch attacks confine the perturbation in a small area, enabling them to be printable and affixed onto the original object to deceive the classification model. For example, Sharif et al. (2016) employed meticulously crafted eyeglass frames to attack the face recognition model. Komkov & Petiushko (2021) confined the perturbation in a hat area and initialized the adversarial patch to the target face. However, these early attacks focus on white-box settings and have low transferability to black-box models. Xie et al. (2019b) proposed TIDIM attack, which combines FGSM (Goodfellow et al., 2014) with random transformations to improve transferability. However, TIDIM did not consider the stealthiness of the patches. Xiao et al. (2021) shows that regularizing the adversarial patches on the latent space of StyleGAN can improve the transferability. However, GenAP did not consider the contextual information of the image either, resulting in conspicuous patches with evident stylistic differences from the original image. Furthermore, iterative sampling from the latent space of a face generator is very time-consuming.
Generative Adversarial Attack. Our attack is also related to generative adversarial attacks. To enhance the transferability of adversarial examples, some attacks perturb the high-level semantics using generative models, such as SemanticAdv (Qiu et al., 2020), Adv-Makeup (Yin et al., 2021) and Adv-attribute (Jia et al., 2022). While these attacks have demonstrated improved visual quality, their computational demands are high. For example, Adv-Makeup needs to generate a makeup dataset from a makeup generation model and train the attack model on multiple face recognition models. Adv-attribute needs to select the most important attributes and define the stealthiness loss and adversary loss in conflicting ways, making the training process unstable and need to adjust weights carefully during the training. Furthermore, SemanticAdv and Adv-attribute are difficult to realize in the physical world as accuirately reproducing attributes like hairstyle and facial expression proves difficult.
Image Inpainting and Image Composition. Our work draws inspiration from state-of-the-art image inpainting and composition technologies. Yu et al.(2018) first proposed a contextual attention layer to generate consistent patches. Building upon this, Zeng et al.(2019) combined the attention layer and pyramid-context encoder to get the masked image's embedding. Xie et al.(2019a) presented learnable bidirectional attention maps to produce visually plausible inpainting content. Zhou et al. (2020) enhance the patch quality further through a dual spatial attention module, leveraging background-foreground cross attention and foreground-self attention. Additionally, our work also draws inspiration from image composition, particularly the module proposed by Ling et al.(2021) that explicitly captures the visual style from the background and applies it to the foreground. However, our work addresses a more challenging scenario than previous inpainting and composition tasks. While filling the masked region with coherent content remains important, we aim to ensure that the resulting inpainted image is correctly classified as the desired target label. This requirement is crucial for effective adversarial patch attacks in face recognition systems. Due to the great difference in age, gender, posture, skin color, and other characteristics between the source and the target image, simultaneously bridging the gap between the source and target image and fooling the face recognition model with a small patch is very challenging.
## 3 METHODOLOGY
We propose a two-stage attack called Adv-Inpainting to generate transferable and stealthy patches. In the first stage, we propose a novel framework consisting of a style encoder, an identity encoder, and an attention-guided StyleGAN (Att-StyleGAN) to generate transferable and natural patches. In the second stage, we propose an Adversarial Patch Refinement Network (APR-Net) to fine-tune the adversarial patch to improve its camouflage further while maintaining adversarial attributes.
3.1 ATTENTION-GUIDED STYLEGAN FOR ADVERSARIAL PATCH GENERATION
StyleGAN (Karras et al., 2019) has shown excellent ability to generate photorealistic images from noises. Recently, Xiao et al. (2021) utilized it to generate transferable adversarial patches. However, because StyleGAN is not designed for adversarial attacks, direct sampling noises from the latent
3
|
[
[
0.176513671875,
0.10479736328125,
0.1895751953125,
0.10479736328125,
0.1895751953125,
0.11614990234375,
0.176513671875,
0.11614990234375
],
[
0.205078125,
0.10418701171875,
0.346435546875,
0.10418701171875,
0.346435546875,
0.11871337890625,
0.205078125,
0.11871337890625
],
[
0.1748046875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.1748046875,
0.147705078125
],
[
0.176513671875,
0.15087890625,
0.82177734375,
0.15087890625,
0.82177734375,
0.1610107421875,
0.176513671875,
0.1610107421875
],
[
0.176513671875,
0.1641845703125,
0.82177734375,
0.1641845703125,
0.82177734375,
0.1749267578125,
0.176513671875,
0.1749267578125
],
[
0.1748046875,
0.1767578125,
0.82421875,
0.1767578125,
0.82421875,
0.1900634765625,
0.1748046875,
0.1900634765625
],
[
0.1748046875,
0.1900634765625,
0.82275390625,
0.1900634765625,
0.82275390625,
0.203857421875,
0.1748046875,
0.203857421875
],
[
0.1748046875,
0.203857421875,
0.82421875,
0.203857421875,
0.82421875,
0.2171630859375,
0.1748046875,
0.2171630859375
],
[
0.1748046875,
0.2171630859375,
0.82421875,
0.2171630859375,
0.82421875,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.1748046875,
0.231689453125,
0.82421875,
0.231689453125,
0.82421875,
0.24560546875,
0.1748046875,
0.24560546875
],
[
0.1748046875,
0.24365234375,
0.82421875,
0.24560546875,
0.82421875,
0.260009765625,
0.1748046875,
0.25830078125
],
[
0.1748046875,
0.2587890625,
0.82421875,
0.2587890625,
0.82421875,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1732177734375,
0.271484375,
0.82421875,
0.272705078125,
0.82421875,
0.287353515625,
0.1732177734375,
0.285888671875
],
[
0.1732177734375,
0.285400390625,
0.82421875,
0.287353515625,
0.82421875,
0.3017578125,
0.1732177734375,
0.2998046875
],
[
0.17236328125,
0.2998046875,
0.6005859375,
0.300537109375,
0.6005859375,
0.31494140625,
0.17236328125,
0.314453125
],
[
0.1748046875,
0.330078125,
0.82275390625,
0.330078125,
0.82275390625,
0.343505859375,
0.1748046875,
0.343505859375
],
[
0.1732177734375,
0.343505859375,
0.82275390625,
0.3447265625,
0.82275390625,
0.35791015625,
0.1732177734375,
0.356689453125
],
[
0.1732177734375,
0.35791015625,
0.82421875,
0.35791015625,
0.82421875,
0.371826171875,
0.1732177734375,
0.371826171875
],
[
0.17236328125,
0.37060546875,
0.82275390625,
0.371826171875,
0.82275390625,
0.386474609375,
0.17236328125,
0.385009765625
],
[
0.1732177734375,
0.3837890625,
0.82421875,
0.385009765625,
0.82421875,
0.399658203125,
0.1732177734375,
0.3984375
],
[
0.1732177734375,
0.399658203125,
0.82421875,
0.399658203125,
0.82421875,
0.41357421875,
0.1732177734375,
0.41357421875
],
[
0.1748046875,
0.4140625,
0.82275390625,
0.4140625,
0.82275390625,
0.4248046875,
0.1748046875,
0.4248046875
],
[
0.1748046875,
0.4267578125,
0.82421875,
0.4267578125,
0.82421875,
0.43994140625,
0.1748046875,
0.43994140625
],
[
0.1748046875,
0.44140625,
0.82421875,
0.44140625,
0.82421875,
0.455078125,
0.1748046875,
0.455078125
],
[
0.1732177734375,
0.455078125,
0.82421875,
0.455078125,
0.82421875,
0.468505859375,
0.1732177734375,
0.468505859375
],
[
0.1748046875,
0.4697265625,
0.279296875,
0.4697265625,
0.279296875,
0.481689453125,
0.1748046875,
0.481689453125
],
[
0.17236328125,
0.49755859375,
0.4794921875,
0.496826171875,
0.4794921875,
0.51123046875,
0.17236328125,
0.51220703125
],
[
0.47802734375,
0.496826171875,
0.82177734375,
0.49755859375,
0.82177734375,
0.51220703125,
0.47802734375,
0.51123046875
],
[
0.1748046875,
0.51123046875,
0.82275390625,
0.51123046875,
0.82275390625,
0.5244140625,
0.1748046875,
0.5244140625
],
[
0.1748046875,
0.52587890625,
0.82421875,
0.52587890625,
0.82421875,
0.5390625,
0.1748046875,
0.5390625
],
[
0.1748046875,
0.5390625,
0.82177734375,
0.5390625,
0.82177734375,
0.55322265625,
0.1748046875,
0.55322265625
],
[
0.1748046875,
0.5546875,
0.82275390625,
0.5546875,
0.82275390625,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.1748046875,
0.5673828125,
0.82421875,
0.5673828125,
0.82421875,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.1748046875,
0.58056640625,
0.82421875,
0.58056640625,
0.82421875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.1748046875,
0.5947265625,
0.82421875,
0.5947265625,
0.82421875,
0.60791015625,
0.1748046875,
0.60791015625
],
[
0.1732177734375,
0.609375,
0.82275390625,
0.609375,
0.82275390625,
0.62255859375,
0.1732177734375,
0.62255859375
],
[
0.1748046875,
0.62255859375,
0.82421875,
0.62255859375,
0.82421875,
0.63623046875,
0.1748046875,
0.63623046875
],
[
0.1748046875,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.17236328125,
0.6494140625,
0.82275390625,
0.6494140625,
0.82275390625,
0.66357421875,
0.17236328125,
0.66357421875
],
[
0.1748046875,
0.6640625,
0.82177734375,
0.6640625,
0.82177734375,
0.67822265625,
0.1748046875,
0.67822265625
],
[
0.1748046875,
0.67822265625,
0.82421875,
0.67822265625,
0.82421875,
0.69140625,
0.1748046875,
0.69140625
],
[
0.1732177734375,
0.69140625,
0.82421875,
0.69140625,
0.82421875,
0.705078125,
0.1732177734375,
0.705078125
],
[
0.1707763671875,
0.70263671875,
0.5498046875,
0.7041015625,
0.5498046875,
0.7216796875,
0.1707763671875,
0.72021484375
],
[
0.17236328125,
0.7412109375,
0.191162109375,
0.7373046875,
0.1953125,
0.75,
0.176513671875,
0.75390625
],
[
0.19775390625,
0.73779296875,
0.34228515625,
0.74072265625,
0.34228515625,
0.7548828125,
0.19775390625,
0.75244140625
],
[
0.1748046875,
0.77099609375,
0.82177734375,
0.77099609375,
0.82177734375,
0.78515625,
0.1748046875,
0.78515625
],
[
0.1748046875,
0.78515625,
0.82275390625,
0.78515625,
0.82275390625,
0.798828125,
0.1748046875,
0.798828125
],
[
0.1748046875,
0.798828125,
0.82421875,
0.798828125,
0.82421875,
0.8125,
0.1748046875,
0.8125
],
[
0.1748046875,
0.81298828125,
0.82421875,
0.81298828125,
0.82421875,
0.8271484375,
0.1748046875,
0.8271484375
],
[
0.17236328125,
0.8271484375,
0.77294921875,
0.8271484375,
0.77294921875,
0.84033203125,
0.17236328125,
0.84033203125
],
[
0.1748046875,
0.85791015625,
0.2017822265625,
0.85791015625,
0.2017822265625,
0.869140625,
0.1748046875,
0.869140625
],
[
0.2132568359375,
0.85693359375,
0.73193359375,
0.85693359375,
0.73193359375,
0.87060546875,
0.2132568359375,
0.87060546875
],
[
0.1748046875,
0.88330078125,
0.82421875,
0.88330078125,
0.82421875,
0.896484375,
0.1748046875,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.82421875,
0.89794921875,
0.82421875,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"2",
"RELATED WORK",
"Transferable Adversarial Patch Attack.Different from the traditional strategy of finding adver-",
"sarialexamples,adversarialpatchattacksconfinetheperturbationinasmallarea,enablingthemto",
"beprintableandaffixedontotheoriginalobjecttodeceivetheclassificationmodel.Forexample",
"Sharif et al. (2016) employed meticulously crafted eyeglass frames to attack the face recognition",
"model. Komkov & Petiushko (2021) confined the perturbation in a hat area and initialized the ad-",
"versarial patch to the target face. However, these early attacks focus on white-box settings and have",
"low transferability to black-box models. Xie et al. (2019b) proposed TIDIM attack, which combines",
"FGSM (Goodfellow et al., 2014) with random transformations to improve transferability. However,",
"TIDIM did not consider the stealthiness of the patches. Xiao et al. (2021) shows that regularizing",
"the adversarial patches on the latent space of StyleGAN can improve the transferability. However,",
"GenAP did not consider the contextual information of the image either, resulting in conspicuous",
"patches with evident stylistic differences from the original image. Furthermore, iterative sampling",
"from the latent space of a face generator is very time-consuming.",
"Generative Adversarial Attack.Our attack is also related to generative adversarial attacks. To",
"enhance the transferability of adversarial examples, some attacks perturb the high-level semantics",
"using generative models, such as SemanticAdv (Qiu et al., 2020), Adv-Makeup (Yin et al., 2021)",
"and Adv-attribute (Jia et al., 2022). While these attacks have demonstrated improved visual quality.",
"their computational demands are high.For example, Adv-Makeup needs to generate a makeup",
"dataset from a makeup generation model and train the attack model on multiple face recognition",
"models. Adv-attribute needs to select the most important attributes and define the stealthiness loss",
"and adversary loss in conflicting ways, making the training process unstable and need to adjust",
"weights carefully during the training. Furthermore, SemanticAdv and Adv-attribute are difficult to",
"realize in the physical world as accurately reproducing attributes like hairstyle and facial expression",
"proves difficult.",
"ImageInpaintingand ImageComposition.",
"Our work draws inspiration from state-of-the-art im-",
"age inpainting and composition technologies. Yu et al. (2o18) first proposed a contextual attention",
"layer to generate consistent patches. Building upon this, Zeng et al. (2019) combined the attention",
"layer and pyramid-context encoder to get the masked image's embedding. Xie et al. (2019a) pre-",
"sentedlearnablebidirectional attentionmapstoproducevisuallyplausibleinpaintingcontent.Zhou",
"et al. (2020) enhance the patch quality further through a dual spatial attention module, leveraging",
"background-foreground cross attention and foreground-self attention. Additionally, our work also",
"draws inspiration from image composition, particularly the module proposed by Ling et al. (2021)",
"that explicitly captures the visual style from the background and applies it to the foreground. How-",
"ever, our work addresses a more challenging scenario than previous inpainting and composition",
"tasks. While filling the masked region with coherent content remains important, we aim to ensure",
"that the resulting inpainted image is correctly classified as the desired target label. This requirement",
"is crucial for effective adversarial patch attacks in face recognition systems. Due to the great differ-",
"ence in age, gender, posture, skin color, and other characteristics between the source and the target",
"image, simultaneously bridging the gap between the source and target image and fooling the face",
"recognition model with a small patch is very challenging.",
"3",
"METHODOLOGY",
"We propose a two-stage attack called Adv-Inpainting to generate transferable and stealthy patches.",
"In the first stage, we propose a novel framework consisting of a style encoder, an identity encoder,",
"and an attention-guided StyleGAN (Att-StyleGAN) to generate transferable and natural patches. In",
"the second stage, we propose an Adversarial Patch Refinement Network (APR-Net) to fine-tune the",
"adversarial patch to improve its camouflage further while maintaining adversarial attributes.",
"3.1",
"ATTENTION-GUIDED STYLEGAN FOR ADVERSARIAL PATCH GENERATION",
"StyleGAN (Karras et al., 2019) has shown excellent ability to generate photorealistic images from",
"noises. Recently, Xiao et al. (2021) utilized it to generate transferable adversarial patches. However,",
"because StyleGAN is not designed for adversarial attacks, direct sampling noises from the latent",
"3"
] |
{
"arxiv_id": "2308.05320",
"language": "en",
"page_idx": 2
}
|
||
2308.05320-en-3
|
2308.05320-en-3.jpg
|
<FigureHere>
Figure 2: (a) Overview of the first stage of Adv-Inpainting, called Att-StyleGAN, which contains a pSp encoder, an identity encoder, and an attention-guided StyleGAN generator.(b) Illustration of the proposed AAIN layer, which fuses the style embedding and identity embedding through a contextual attention map to generate transferable and stealthy patches.
space of the StyleGAN will generate unnatural and conspicuous patches (Figure 1(f)). To solve this problem, we propose a novel end-to-end adversarial patch generation framework. We will first introduce the architecture of the proposed framework and then introduce the design of loss functions.
The architecture of Att-StyleGAN As shown in Figure 2(a), Att-StyleGAN consists of a style encoder to extract style-related features, a pretrained identity encoder to extract identity features, and an attention-guided StyleGAN generator to generate transferable and photorealistic images. The style encode utilizes the the state-of-the-art style embedding framework, called pixel-to-Style-pixel $(\dot{\text{pSp) (Richardson et al., 2021) network, which adopts pyramid-like architectures to extract multi-}}$ scale features. It was initially designed for style transfer tasks. In this paper, we use it to extract features related to the posture and stylistic information, then map the features to style embeddings $w_{sty}$ in the $\mathcal{W} +$space ( Abdal et al. , 2019) of the StyleGAN. Also, we use a pretrained face recogni- tion model, e. g, Arcface ( Deng et al. , 2019) , to extract the identity feature $z_{id}$ from the target image. We then map $z_{id}$ to $w_{}$ $w_{sty}$ and identity embeddings $w_id$ are fed into the AAIN layers. Let $x_syn$ be the synthesized image of the generator, $x_s$ be the source image, and $M$ be the mask, then the final output is
$$x_{out}=x_{syn}\cdot(1-M)+x_s\cdot M.$$
(1)
Attention-guided Adaptive Instance Normorlization Adaptive instance normalization (AdaIN) (Huang & $\bar{\text{Belongie,2017)can realize real-time style transfer and is widely used in image-to-image}}$ translation. The state-of-the-art face generative model, StyleGAN adopts AdaIN to generate highfidelity images:
$$AdaIN(F_i,w)=\gamma_w(\frac{F_i-\mu(F_i)}{\sigma(F_i)})+\mu_w,$$
(2)
where $F_i$ is the $i^th$ block's output of StyleGAN and is normalized over its spatial locations. $\gamma_w$ and $\mu_w$ are affine transformations of latent vector $w$ to stylize the feature map $F_i.$ However, AdaIN is unsuitable for our attack because the context information of the source image is not considered.
To merge the identity embeddings of the target image and the style embeddings of the source image, such as expression and posture features, we propose a new attention-guided adaptive instance normalization layer (AAIN). The architecture of AAIN is shown in Figure 2(b). We first use two linear layers to get the attacker-related style vectors $\beta_t,\gamma_t$ and fused style vectors $\beta_f,\gamma_f$ from the style embedding $w_{sty}$ and the identity embedding $w_id$,
$$[\beta_{t},\gamma_{t}]=Lin_{tex}(w_{sty}),\quad[\beta_{f},\gamma_{f}]=Lin_{fuse}([w_{sty},w_{id}]),$$
(3)
Although the fuse layer $Lin_fuse$ can mix the feature embeddings of the source and target images, it does not consider the contextual coherence with the background. To improve the coherence and
4
|
[
[
0.26708984375,
0.1319580078125,
0.365234375,
0.1458740234375,
0.35546875,
0.186279296875,
0.25732421875,
0.17236328125
],
[
0.281982421875,
0.15087890625,
0.293212890625,
0.15087890625,
0.293212890625,
0.154052734375,
0.281982421875,
0.154052734375
],
[
0.316162109375,
0.279052734375,
0.4150390625,
0.279052734375,
0.4150390625,
0.290283203125,
0.316162109375,
0.290283203125
],
[
0.6748046875,
0.276611328125,
0.76123046875,
0.276611328125,
0.76123046875,
0.28857421875,
0.6748046875,
0.28857421875
],
[
0.1748046875,
0.302978515625,
0.82275390625,
0.302978515625,
0.82275390625,
0.31640625,
0.1748046875,
0.31640625
],
[
0.1748046875,
0.317626953125,
0.82275390625,
0.317626953125,
0.82275390625,
0.330810546875,
0.1748046875,
0.330810546875
],
[
0.1748046875,
0.330810546875,
0.82421875,
0.330810546875,
0.82421875,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.1748046875,
0.34521484375,
0.6337890625,
0.34521484375,
0.6337890625,
0.359130859375,
0.1748046875,
0.359130859375
],
[
0.176513671875,
0.38818359375,
0.8203125,
0.38818359375,
0.8203125,
0.3984375,
0.176513671875,
0.3984375
],
[
0.1748046875,
0.40087890625,
0.82275390625,
0.40087890625,
0.82275390625,
0.4140625,
0.1748046875,
0.4140625
],
[
0.1748046875,
0.416748046875,
0.82177734375,
0.416748046875,
0.82177734375,
0.4267578125,
0.1748046875,
0.4267578125
],
[
0.1748046875,
0.443115234375,
0.82177734375,
0.443115234375,
0.82177734375,
0.45703125,
0.1748046875,
0.45703125
],
[
0.1748046875,
0.458251953125,
0.82275390625,
0.458251953125,
0.82275390625,
0.4716796875,
0.1748046875,
0.4716796875
],
[
0.1748046875,
0.4716796875,
0.82275390625,
0.4716796875,
0.82275390625,
0.48486328125,
0.1748046875,
0.48486328125
],
[
0.176513671875,
0.4873046875,
0.82177734375,
0.4873046875,
0.82177734375,
0.49755859375,
0.176513671875,
0.49755859375
],
[
0.1748046875,
0.5,
0.82177734375,
0.5,
0.82177734375,
0.51318359375,
0.1748046875,
0.51318359375
],
[
0.1748046875,
0.51318359375,
0.82421875,
0.51318359375,
0.82421875,
0.5263671875,
0.1748046875,
0.5263671875
],
[
0.1748046875,
0.5263671875,
0.82421875,
0.5263671875,
0.82421875,
0.54052734375,
0.1748046875,
0.54052734375
],
[
0.1748046875,
0.54150390625,
0.82177734375,
0.54052734375,
0.82177734375,
0.5537109375,
0.1748046875,
0.5546875
],
[
0.1732177734375,
0.5537109375,
0.82177734375,
0.5537109375,
0.82177734375,
0.5673828125,
0.1732177734375,
0.5673828125
],
[
0.1748046875,
0.568359375,
0.82421875,
0.568359375,
0.82421875,
0.58203125,
0.1748046875,
0.58203125
],
[
0.176513671875,
0.58349609375,
0.82421875,
0.58349609375,
0.82421875,
0.5966796875,
0.176513671875,
0.5966796875
],
[
0.1732177734375,
0.59521484375,
0.72607421875,
0.59521484375,
0.72607421875,
0.609375,
0.1732177734375,
0.609375
],
[
0.3857421875,
0.61865234375,
0.611328125,
0.6171875,
0.611328125,
0.6318359375,
0.3857421875,
0.63330078125
],
[
0.80224609375,
0.61865234375,
0.82666015625,
0.61865234375,
0.82666015625,
0.6318359375,
0.80224609375,
0.6318359375
],
[
0.176513671875,
0.65576171875,
0.82275390625,
0.65576171875,
0.82275390625,
0.66943359375,
0.176513671875,
0.66943359375
],
[
0.1732177734375,
0.66943359375,
0.82275390625,
0.67041015625,
0.82275390625,
0.6845703125,
0.1732177734375,
0.68310546875
],
[
0.1748046875,
0.6845703125,
0.82275390625,
0.6845703125,
0.82275390625,
0.69873046875,
0.1748046875,
0.69873046875
],
[
0.17236328125,
0.6962890625,
0.279296875,
0.69775390625,
0.279296875,
0.7119140625,
0.17236328125,
0.7109375
],
[
0.5107421875,
0.7119140625,
0.5810546875,
0.7119140625,
0.5810546875,
0.7236328125,
0.5107421875,
0.7236328125
],
[
0.359375,
0.71826171875,
0.486083984375,
0.71826171875,
0.486083984375,
0.73291015625,
0.359375,
0.73291015625
],
[
0.48193359375,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.72998046875,
0.48193359375,
0.72998046875
],
[
0.48779296875,
0.724609375,
0.496826171875,
0.724609375,
0.496826171875,
0.72998046875,
0.48779296875,
0.72998046875
],
[
0.80224609375,
0.72021484375,
0.826171875,
0.72021484375,
0.826171875,
0.734375,
0.80224609375,
0.734375
],
[
0.52685546875,
0.72998046875,
0.56787109375,
0.72998046875,
0.56787109375,
0.7412109375,
0.52685546875,
0.7412109375
],
[
0.17236328125,
0.744140625,
0.82666015625,
0.74560546875,
0.82666015625,
0.76318359375,
0.17236328125,
0.76220703125
],
[
0.1732177734375,
0.76220703125,
0.826171875,
0.76123046875,
0.826171875,
0.7744140625,
0.1732177734375,
0.77587890625
],
[
0.1748046875,
0.77587890625,
0.79736328125,
0.77587890625,
0.79736328125,
0.7890625,
0.1748046875,
0.7890625
],
[
0.1748046875,
0.796875,
0.82275390625,
0.796875,
0.82275390625,
0.81103515625,
0.1748046875,
0.81103515625
],
[
0.1748046875,
0.8125,
0.82275390625,
0.81103515625,
0.82275390625,
0.82275390625,
0.1748046875,
0.82373046875
],
[
0.1748046875,
0.82470703125,
0.82421875,
0.82470703125,
0.82421875,
0.83837890625,
0.1748046875,
0.83837890625
],
[
0.1748046875,
0.8388671875,
0.826171875,
0.8388671875,
0.826171875,
0.85400390625,
0.1748046875,
0.85400390625
],
[
0.1748046875,
0.85302734375,
0.53759765625,
0.85302734375,
0.53759765625,
0.8671875,
0.1748046875,
0.8671875
],
[
0.30224609375,
0.875,
0.6953125,
0.875,
0.6953125,
0.8896484375,
0.30224609375,
0.8896484375
],
[
0.8037109375,
0.875,
0.826171875,
0.875,
0.826171875,
0.8896484375,
0.8037109375,
0.8896484375
],
[
0.1748046875,
0.896484375,
0.8203125,
0.896484375,
0.8203125,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"",
"",
"(a)Att-StyleGAN",
"(b) AAIN layer",
"Figure 2: (a) Overview of the first stage of Adv-Inpainting, called Att-StyleGAN, which contains",
"a pSp encoder, an identity encoder, and an attention-guided StyleGAN generator. (b) Illustration",
"of the proposed AAIN layer, which fuses the style embedding and identity embedding through a",
"contextual attention map to generate transferable and stealthy patches.",
"space of the StyleGAN will generate unnatural and conspicuous patches (Figure l(f)).To solve",
"this problem, we propose a novel end-to-end adversarial patch generation framework. We will firs",
"introducethearchitectureoftheproposedframeworkandthenintroducethedesignoflossfunctions.",
"The architecture of Att-StyleGANAs shown in Figure 2(a), Att-StyleGAN consists of a style en-",
"coder to extract style-related features, a pretrained identity encoder to extract identity features, and",
"an attention-guided StyleGAN generator to generate transferable and photorealistic images. The",
"styleencodeutilizesthethestate-of-the-artstyleembeddingframework,calledpixel-to-Style-pixel",
"(pSp) (Richardson et al., 2021) network, which adopts pyramid-like architectures to extract multi-",
"scale features. It was initially designed for style transfer tasks. In this paper, we use it to extract",
"features related to the posture and stylistic information, then map the features to style embeddings",
"Wsty in theW+ space (Abdal et al., 2019)of the StyleGAN.Also, we use a pretrained face recogni-",
"tion model, e.g., Arcface (Deng et al., 2019), to extract the identity feature zid from the target image.",
"We then map zid to wid in the W+ space through a mapping network Gm. At last, style embeddings",
"Wsty and identity embeddings wid are fed into the AAIN layers. Let syn be the synthesized image",
"of the generator, g be the source image, and M be the mask, then the final output is",
"Tout=asyn·(1-M)+rsM.",
"(1)",
"Attention-guided Adaptive Instance NormorlizationAdaptive instance normalization (AdalN)",
"(Huang & Belongie, 2017) can realize real-time style transfer and is widely used in image-to-image",
"translation. The state-of-the-art face generative model, StyleGAN adopts AdalN to generate high-",
"fidelity images:",
"F-H(F)",
"AdaIN(Fi,w)=",
"A",
"V.",
"(2)",
"α(F)",
"where F, is the ith block's output of StyleGAN and is normalized over its spatial locations. w and",
"μw, are affine transformations of latent vector w to stylize the feature map F.However, AdalN is",
"unsuitable for our attack because the context information of the source image is not considered.",
"To merge the identity embeddings of the target image and the style embeddings of the source im-",
"age, such as expression and posture features,wepropose a new attention-guided adaptiveinstance",
"normalization layer (AAIN). The architecture of AAIN is shown in Figure 2(b). We first use two",
"linear layers to get the attacker-related style vectors βt, t and fused style vectors βf, Yf from the",
"style embedding wWsty and the identity embedding Wid,",
"[βt, t] = Linter(Wsty), [βf,f] =Linfuse([Wsty,Widl),",
"(3)",
"Although the fuse layer Linf use can mix the feature embeddings of the source and target images",
"it does not consider the contextual coherence with the background. To improve the coherence and",
"4"
] |
{
"arxiv_id": "2308.05320",
"language": "en",
"page_idx": 3
}
|
||
2308.05320-en-4
|
2308.05320-en-4.jpg
|
camouflage of the adversarial patch, we compute the background-patch cross-attention map to fuse the styles further. As shown in Figure 2(b), We use the feature map $F_p$ of the source image, the $\tilde{\text{normalized feature map }F_{in}}$ from the previous block and the mask $M$ as the prior information to get the background-patch cross-attention map $D_{h}:$
$$D_h=\sigma(Conv([F^p,F^{in}])\cdot(1-M)+M,$$
(4)
where the hole of $M$ is set to 0, and the background of $M$ is set to 1. The Conv is a convolution layer, and $\sigma$ is a sigmoid activation function. The synthesized style vectors for the patch area are :
$$\gamma_{syn}=D_{h}\gamma_{t}+(1-D_{h})\gamma_{f},\quad\beta_{syn}=D_{h}\beta_{t}+(1-D_{h})\beta_{f}.$$
(5)
The total styles to stylize the normalized feature map $F_{in}$ are:
$$\gamma_i=M\gamma_t+(1-M)\gamma_{syn},\quad\beta_i=M\gamma_t+(1-M)\gamma_{syn}.$$
(6)
Finally, the output of AAIN layer can be expressed as:
$$AAIN(F_i,F_p,w_{sty},w_{id})=\gamma_i(\frac{F_i-\mu(F_i)}{\sigma(F_i)})+\mu_i.$$
(7)
The loss functions of Att-StyleGAN In our attack, the goal is to make the source image with the adversarial patch be identified as the target image $x_t.$ Therefore, we use the cosine similarity loss as the adversarial loss:
$$\mathcal{L}_{adv}=1-cossim(f(x_{out}),f(x_{t})),$$
(8)
where $f(\cdot)$ is the white-box substitute model to be attacked.
We also use the semi-supervised learning strategy to train the Att-StyleGAN. Let $id(\cdot)$ be the identity
function to map the face image to its labels. The recovery loss is expressed as:
$$\mathcal{L}_{rec}=\begin{cases}\dfrac{1}{2}\|x_{syn}-x_s\|_2^2,&if\:id(x_s)=id(x_t)\\[2ex]\dfrac{1}{2}\|x_{syn}\cdot M_d-x_s\cdot M_d\|_2^2,&if\:id(x_s)\neq id(x_t)\end{cases},$$
(9)
where $M_d$ is a discounted mask, where the value outside the patch area is 1, and the value in the patch area is set to $\frac1{\alpha\cdot e^l}$,where $l$ is the distance from the pixel to the patch boundary, and $\alpha$ is set to 0.15 in our experiments. We use $M_d$ to force the pixel values near the patch boundary to be more consistent with the source image while giving more freedom to the inner pixels.
Furthermore, we include the Learned Perceptual Image Patch Similarity (LPIPS) loss (Zhang et al., 2018) to improve perceptual similarity between $x_syn$ and $x_s$, which computes the cosine distance of deep features via a pretrained VGG-Net. LPIPS loss can make the adversarial patches' quality and style more similar to the original content and is consistent with human perception, which enables a further blend between the source image and the patches to boost stealthiness:
(10)
$$\mathcal{L}_{\mathrm{LPIPS}}=\Sigma_{l}\frac{1}{H_{l}W_{l}}\Sigma_{h,w}\|w_{l}\odot(\hat{y}_{hw}^{l}-\hat{y}_{0}^{l}hw)\|_{2}^{2},$$
where $\hat{y}^l,\hat{y}_0^l$ are features of $x_syn$ and $x_s$ extracted from the pretrained VGG-Net's layer $l.$ Finally,
we include the discriminator loss $\mathcal{L}_dis$, which is the same as the StyleGAN. The fnal loss is:
$$L_{total}=\lambda_{adv}\mathcal{L}_{adv}+\lambda_{rec}\mathcal{L}_{rec}+\lambda_{\mathrm{LPIPS}}\mathcal{L}_{\mathrm{LPIPS}}+\lambda_{dis}\mathcal{L}_{dis},$$
(11)
where $\lambda_adv,\lambda_{rec},\lambda_{\mathrm{LPIPS}}$ and $\lambda_dis$ are balance weights. During the training process, the pretrained FR model is fixed, and the style encoder, the mapping network $G_m$ and the generator are updated. 3.2 ADVERSARIAL PATCH REFINEMENT NETWORK
In this paper, we propose an adversarial patch refinement network (APR-Net) to refine the generated adversarial patch further to make it more stealthy and coherent with the source image while maintaining the adversarial properties. As shown in Figure 3, APR-Net consists of a U-Net-like generator with dual spatial attention layers (Zhou et al., 2020) and a PatchGAN (Isola et al., 2017) discriminator. Dual spatial attention can greatly improve the naturalness of image inpainting through background-foreground cross attention and foreground-self attention. PatchGAN discriminator tries
5
|
[
[
0.176513671875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11871337890625,
0.176513671875,
0.11871337890625
],
[
0.176513671875,
0.12060546875,
0.82177734375,
0.12060546875,
0.82177734375,
0.1307373046875,
0.176513671875,
0.1307373046875
],
[
0.1748046875,
0.1331787109375,
0.82421875,
0.1331787109375,
0.82421875,
0.146484375,
0.1748046875,
0.146484375
],
[
0.1748046875,
0.1458740234375,
0.48046875,
0.1458740234375,
0.48046875,
0.1590576171875,
0.1748046875,
0.1590576171875
],
[
0.35693359375,
0.1673583984375,
0.64208984375,
0.1673583984375,
0.64208984375,
0.1817626953125,
0.35693359375,
0.1817626953125
],
[
0.80224609375,
0.1685791015625,
0.826171875,
0.1685791015625,
0.826171875,
0.1817626953125,
0.80224609375,
0.1817626953125
],
[
0.17724609375,
0.1912841796875,
0.82275390625,
0.1912841796875,
0.82275390625,
0.2020263671875,
0.17724609375,
0.2020263671875
],
[
0.1748046875,
0.203857421875,
0.81396484375,
0.203857421875,
0.81396484375,
0.2171630859375,
0.1748046875,
0.2171630859375
],
[
0.293212890625,
0.2247314453125,
0.70166015625,
0.222900390625,
0.70166015625,
0.240478515625,
0.293212890625,
0.242431640625
],
[
0.80126953125,
0.2235107421875,
0.82666015625,
0.2235107421875,
0.82666015625,
0.2412109375,
0.80126953125,
0.2412109375
],
[
0.1748046875,
0.24560546875,
0.580078125,
0.246826171875,
0.580078125,
0.261474609375,
0.1748046875,
0.260009765625
],
[
0.3095703125,
0.268310546875,
0.68701171875,
0.268310546875,
0.68701171875,
0.28271484375,
0.3095703125,
0.28271484375
],
[
0.80224609375,
0.268310546875,
0.82666015625,
0.268310546875,
0.82666015625,
0.2822265625,
0.80224609375,
0.2822265625
],
[
0.17236328125,
0.28857421875,
0.53125,
0.2890625,
0.53125,
0.3037109375,
0.17236328125,
0.302978515625
],
[
0.5400390625,
0.3125,
0.62841796875,
0.311279296875,
0.62841796875,
0.326904296875,
0.5400390625,
0.32763671875
],
[
0.330078125,
0.3193359375,
0.5458984375,
0.3193359375,
0.5458984375,
0.333984375,
0.330078125,
0.333984375
],
[
0.62744140625,
0.322509765625,
0.6640625,
0.322509765625,
0.6640625,
0.333984375,
0.62744140625,
0.333984375
],
[
0.7998046875,
0.318115234375,
0.82861328125,
0.318115234375,
0.82861328125,
0.3359375,
0.7998046875,
0.3359375
],
[
0.55615234375,
0.32763671875,
0.6044921875,
0.32763671875,
0.6044921875,
0.34228515625,
0.55615234375,
0.34228515625
],
[
0.1748046875,
0.35595703125,
0.826171875,
0.35595703125,
0.826171875,
0.369384765625,
0.1748046875,
0.369384765625
],
[
0.1748046875,
0.37060546875,
0.826171875,
0.37060546875,
0.826171875,
0.3837890625,
0.1748046875,
0.3837890625
],
[
0.17236328125,
0.38330078125,
0.3095703125,
0.3837890625,
0.3095703125,
0.3984375,
0.17236328125,
0.397705078125
],
[
0.373291015625,
0.395751953125,
0.62353515625,
0.396484375,
0.62353515625,
0.4140625,
0.373291015625,
0.41357421875
],
[
0.80224609375,
0.3984375,
0.826171875,
0.3984375,
0.826171875,
0.41357421875,
0.80224609375,
0.41357421875
],
[
0.1732177734375,
0.416748046875,
0.56640625,
0.415283203125,
0.56640625,
0.429931640625,
0.1732177734375,
0.43115234375
],
[
0.1748046875,
0.4375,
0.82275390625,
0.4375,
0.82275390625,
0.451904296875,
0.1748046875,
0.451904296875
],
[
0.1732177734375,
0.451904296875,
0.68896484375,
0.451904296875,
0.68896484375,
0.46533203125,
0.1732177734375,
0.46533203125
],
[
0.418212890625,
0.479248046875,
0.4755859375,
0.479248046875,
0.4755859375,
0.494384765625,
0.418212890625,
0.494384765625
],
[
0.560546875,
0.479736328125,
0.69140625,
0.479736328125,
0.69140625,
0.494384765625,
0.560546875,
0.494384765625
],
[
0.301513671875,
0.492431640625,
0.32275390625,
0.496826171875,
0.318603515625,
0.50830078125,
0.29736328125,
0.50390625
],
[
0.359375,
0.48095703125,
0.419189453125,
0.477783203125,
0.420654296875,
0.49560546875,
0.36279296875,
0.5
],
[
0.80126953125,
0.492431640625,
0.82861328125,
0.492431640625,
0.82861328125,
0.5087890625,
0.80126953125,
0.5087890625
],
[
0.378173828125,
0.51025390625,
0.689453125,
0.50830078125,
0.689453125,
0.52294921875,
0.378173828125,
0.5244140625
],
[
0.1748046875,
0.5361328125,
0.82275390625,
0.5361328125,
0.82275390625,
0.5498046875,
0.1748046875,
0.5498046875
],
[
0.1748046875,
0.55029296875,
0.82275390625,
0.55029296875,
0.82275390625,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.1748046875,
0.564453125,
0.82421875,
0.564453125,
0.82421875,
0.57763671875,
0.1748046875,
0.57763671875
],
[
0.1748046875,
0.5791015625,
0.69677734375,
0.5791015625,
0.69677734375,
0.59228515625,
0.1748046875,
0.59228515625
],
[
0.1748046875,
0.5986328125,
0.8203125,
0.5986328125,
0.8203125,
0.61279296875,
0.1748046875,
0.61279296875
],
[
0.176513671875,
0.61279296875,
0.82421875,
0.61279296875,
0.82421875,
0.62744140625,
0.176513671875,
0.62744140625
],
[
0.1748046875,
0.626953125,
0.82421875,
0.626953125,
0.82421875,
0.6416015625,
0.1748046875,
0.6416015625
],
[
0.1732177734375,
0.6396484375,
0.82421875,
0.6416015625,
0.82421875,
0.65478515625,
0.1732177734375,
0.65283203125
],
[
0.1732177734375,
0.65478515625,
0.68115234375,
0.65478515625,
0.68115234375,
0.66845703125,
0.1732177734375,
0.66845703125
],
[
0.461669921875,
0.681640625,
0.65673828125,
0.681640625,
0.65673828125,
0.6962890625,
0.461669921875,
0.6962890625
],
[
0.79345703125,
0.681640625,
0.826171875,
0.681640625,
0.826171875,
0.69775390625,
0.79345703125,
0.69775390625
],
[
0.419189453125,
0.69140625,
0.464111328125,
0.69140625,
0.464111328125,
0.70556640625,
0.419189453125,
0.70556640625
],
[
0.340576171875,
0.68310546875,
0.42236328125,
0.68310546875,
0.42236328125,
0.69775390625,
0.340576171875,
0.69775390625
],
[
0.1748046875,
0.7109375,
0.82421875,
0.7109375,
0.82421875,
0.72900390625,
0.1748046875,
0.72900390625
],
[
0.1748046875,
0.72802734375,
0.78369140625,
0.72802734375,
0.78369140625,
0.7421875,
0.1748046875,
0.7421875
],
[
0.3056640625,
0.74853515625,
0.69287109375,
0.74853515625,
0.69287109375,
0.76318359375,
0.3056640625,
0.76318359375
],
[
0.79345703125,
0.74755859375,
0.826171875,
0.74755859375,
0.826171875,
0.76318359375,
0.79345703125,
0.76318359375
],
[
0.1748046875,
0.76953125,
0.82275390625,
0.76953125,
0.82275390625,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1748046875,
0.7841796875,
0.81396484375,
0.7841796875,
0.81396484375,
0.79736328125,
0.1748046875,
0.79736328125
],
[
0.17236328125,
0.81494140625,
0.54150390625,
0.81494140625,
0.54150390625,
0.8291015625,
0.17236328125,
0.8291015625
],
[
0.1732177734375,
0.84130859375,
0.8203125,
0.84130859375,
0.8203125,
0.85498046875,
0.1732177734375,
0.85498046875
],
[
0.1732177734375,
0.85498046875,
0.82421875,
0.85498046875,
0.82421875,
0.86865234375,
0.1732177734375,
0.86865234375
],
[
0.176513671875,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.880859375,
0.176513671875,
0.880859375
],
[
0.1748046875,
0.88330078125,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.1748046875,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.826171875,
0.9111328125,
0.826171875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.494384765625,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.494384765625,
0.96044921875
]
] |
[
"camouflage of the adversarial patch, we compute the background-patch cross-attention map to fuse",
"the styles further. As shown in Figure 2(b),We use thefeature map Fn of the source image. the",
"normalized feature map Fin from the previous block and the mask M as the prior information to get",
"the background-patch cross-attention map Dh:",
"Dh = α(Conv([FP, Fin) - (1 - M) + M,",
"(4)",
"where the hole of M is set to O,and the background of M is setto l.The Conv is a convolution",
"layer, and is a sigmoid activation function. The synthesized style vectors for the patch area are :",
"syn=Dht +(1-Dh)f, βsyn =Dnβt+(1-Dh)βf.",
"(5)",
"The total styles to stylize the normalized feature map Fin are:",
"%= Mat+(1-M)syn, B, =Mt+(1-M)syn",
"(6)",
"Finally, the output of AAIN layer can be expressed as:",
"F-μ(F)",
"AAIN(Fi, Fp, Wsty, Wid) = i(",
"+",
"(7)",
"α(F)",
"The loss functions of Att-StyleGANIn our attack, the goal is to make the source image with the",
"adversarial patch be identified as the target image &t. Therefore, we use the cosine similarity loss as",
"the adversarial loss:",
"Ladw = 1 - cossim(f(aout), f(rt),",
"(8)",
"where f() is the white-box substitute model to be attacked.",
"We also use the semi-supervised learning strategy to train the Att-StyleGAN. Let id() be the identity",
"function to map the face image to its labels. The recovery loss is expressed as:",
"-Cs/2,",
"if id(as) = id(ct)",
"Cr",
"llasyn",
"(9)",
"IlasynMd-&s·Mall2,if id(a)手id(at)",
"where Md is a discounted mask, where the value outside the patch area is l, and the value in the",
"patch area is set to er, where I is the distance from the pixel to the patch boundary, and α is set to",
"0.15 in our experiments. We use Ma to force the pixel values near the patch boundary to be more",
"consistent with the source image while giving more freedom to the inner pixels.",
"Furthermore, we include the Learned Perceptual Image Patch Similarity (LPIPS) loss (Zhang et al.",
"2018) to improve perceptual similarity between syn and c s, which computes the cosine distance of",
"deep features via a pretrained VGG-Net. LPIPS loss can make the adversarial patches' quality and",
"style more similar to the original content and is consistent with human perception, which enables a",
"further blend between the source image and the patches to boost stealthiness:",
"Dh.wwi(gh-gohw)l2",
"(10)",
"H,Wi",
"CLPIPS =Z",
"where g', 96 are features of E syn and &s extracted from the pretrained VGG-Net's layer l. Finally,",
"we include the discriminator loss Cdis , which is the same as the StyleGAN. The final loss is:",
"Ltotal=AadyCad+ArecCrec+ALPIPsCLPIPS+AdisCdis",
"(11)",
"where Aady,Arec,Lpips and Adis are balance weights. During the training process, the pretrained",
"FR model is fixed, and the style encoder, the mapping network Gm and the generator are updated.",
"3.2ADVERSARIALPATCHREFINEMENTNETWORK",
"In this paper, we propose an adversarial patch refinement network (APR-Net) to refine the gener",
"ated adversarial patch further to make it more stealthy and coherent with the source image while",
"maintainingtheadversarial properties.As showninFigure3.APR-Net consists of a U-Net-like",
"generator with dual spatial attention layers (Zhou et al., 2020) and a PatchGAN (Isola et al., 2017)",
"discriminator. Dual spatial attention can greatly improve the naturalness of image inpainting through",
"background-foreground cross attention and foreground-self attention. PatchGAN discriminator tries",
"5"
] |
{
"arxiv_id": "2308.05320",
"language": "en",
"page_idx": 4
}
|
||
2008.08823-en-0
|
2008.08823-en-0.jpg
|
DronePose: Photorealistic UAV- Assistant
Dataset Synthesis for 3D Pose Estimation via a
Smooth Silhouette Loss
Georgios Albanis$^\star1$, Nikolaos Zioulis$^*1$, Anastasios Dimou$^1$, Dimitrios
$\mathrm{Zarpalas}^{1}$,and Petros Daras$^1$
Visual Computing Lab, Information Technologies Institute, Centre for Research and
Technology Hellas, Greece
$\{$galbanis, nzioulis, dimou, zarpalas, daras$\}$@iti.gr
vcl.iti.gr
Abstract. In this work we consider UAVs as cooperative agents supporting human users in their operations. In this context, the 3D localisation of the UAV assistant is an important task that can facilitate the exchange of spatial information between the user and the UAV. To address this in a data-driven manner, we design a data synthesis pipeline to create a realistic multimodal dataset that includes both the exocentric user view, and the egocentric UAV view. We then exploit the joint availability of photorealistic and synthesized inputs to train a single-shot monoculan pose estimation model. During training we leverage differentiable rendering to supplement a state-of-the-art direct regression objective with a novel smooth silhouette loss. Our results demonstrate its qualitative and quantitative performance gains over traditional silhouette objectives. Our data and code are available at https://vcl3d.github.io/DronePose.
Keywords: 3D Pose Estimation, Dataset Generation, UAV, Differen-
tiable Rendering
## 1 Introduction
Advances in robotics, their autonomy, and computer vision are constituting Unmanned Aerial Vehicles (UAVs)-or otherwise known as drones- an emerging ubiquitous technology. Indeed, 'the potential that autonomous or piloted UAVs can provide has already been acknowledged in numerous application fields like filming [69], emergency response [34], animal conservation [30], and infrastructure inspection [17]. In addition, a lot of research has focused on the detection of drones in the context of anti-drone systems [59] and their accurate identification [12].
Nevertheless, the miniaturization and commoditisation of drones open up new types of applications for mini-UAVs. Moving beyond the role of remote, and/or malicious agents, personal - friendly - drones can operate in tandem $^{\star}$Equal contribution
|
[
[
0.2607421875,
0.143310546875,
0.74267578125,
0.14453125,
0.74267578125,
0.1622314453125,
0.2607421875,
0.1610107421875
],
[
0.2296142578125,
0.1685791015625,
0.77685546875,
0.1685791015625,
0.77685546875,
0.18310546875,
0.2296142578125,
0.18310546875
],
[
0.366943359375,
0.1912841796875,
0.6396484375,
0.1912841796875,
0.6396484375,
0.205810546875,
0.366943359375,
0.205810546875
],
[
0.2548828125,
0.236083984375,
0.7509765625,
0.236083984375,
0.7509765625,
0.250732421875,
0.2548828125,
0.250732421875
],
[
0.0196075439453125,
0.255615234375,
0.06781005859375,
0.255615234375,
0.06781005859375,
0.4091796875,
0.0196075439453125,
0.4091796875
],
[
0.39794921875,
0.25244140625,
0.60693359375,
0.25244140625,
0.60693359375,
0.266357421875,
0.39794921875,
0.266357421875
],
[
0.223876953125,
0.279052734375,
0.78173828125,
0.279052734375,
0.78173828125,
0.292236328125,
0.223876953125,
0.292236328125
],
[
0.41259765625,
0.291748046875,
0.5908203125,
0.292236328125,
0.5908203125,
0.306884765625,
0.41259765625,
0.30615234375
],
[
0.3056640625,
0.30615234375,
0.69775390625,
0.306884765625,
0.69775390625,
0.3212890625,
0.3056640625,
0.32080078125
],
[
0.473876953125,
0.322509765625,
0.53125,
0.322509765625,
0.53125,
0.333984375,
0.473876953125,
0.333984375
],
[
0.2646484375,
0.37060546875,
0.73876953125,
0.371826171875,
0.73876953125,
0.386474609375,
0.2646484375,
0.385009765625
],
[
0.2646484375,
0.386962890625,
0.73779296875,
0.386474609375,
0.73779296875,
0.397705078125,
0.2646484375,
0.3984375
],
[
0.0228729248046875,
0.397705078125,
0.062103271484375,
0.397705078125,
0.062103271484375,
0.498779296875,
0.0228729248046875,
0.498779296875
],
[
0.2646484375,
0.399658203125,
0.73876953125,
0.399658203125,
0.73876953125,
0.41357421875,
0.2646484375,
0.41357421875
],
[
0.2646484375,
0.41357421875,
0.73876953125,
0.41357421875,
0.73876953125,
0.4267578125,
0.2646484375,
0.4267578125
],
[
0.2646484375,
0.427978515625,
0.73876953125,
0.427978515625,
0.73876953125,
0.44140625,
0.2646484375,
0.44140625
],
[
0.265625,
0.443115234375,
0.73876953125,
0.443115234375,
0.73876953125,
0.453857421875,
0.265625,
0.453857421875
],
[
0.2646484375,
0.455078125,
0.73876953125,
0.455078125,
0.73876953125,
0.468505859375,
0.2646484375,
0.468505859375
],
[
0.265625,
0.47021484375,
0.73876953125,
0.47021484375,
0.73876953125,
0.48095703125,
0.265625,
0.48095703125
],
[
0.021240234375,
0.4873046875,
0.058013916015625,
0.4873046875,
0.058013916015625,
0.7119140625,
0.021240234375,
0.7119140625
],
[
0.265625,
0.48291015625,
0.73876953125,
0.48291015625,
0.73876953125,
0.496826171875,
0.265625,
0.496826171875
],
[
0.2646484375,
0.496826171875,
0.73876953125,
0.496826171875,
0.73876953125,
0.51025390625,
0.2646484375,
0.51025390625
],
[
0.265625,
0.51220703125,
0.73876953125,
0.51220703125,
0.73876953125,
0.52294921875,
0.265625,
0.52294921875
],
[
0.2646484375,
0.5244140625,
0.736328125,
0.5244140625,
0.736328125,
0.53857421875,
0.2646484375,
0.53857421875
],
[
0.2646484375,
0.53857421875,
0.73388671875,
0.53857421875,
0.73388671875,
0.5517578125,
0.2646484375,
0.5517578125
],
[
0.2646484375,
0.56494140625,
0.736328125,
0.56494140625,
0.736328125,
0.5791015625,
0.2646484375,
0.5791015625
],
[
0.261474609375,
0.5771484375,
0.37744140625,
0.580078125,
0.37744140625,
0.5947265625,
0.261474609375,
0.59130859375
],
[
0.2197265625,
0.6220703125,
0.2344970703125,
0.6220703125,
0.2344970703125,
0.6318359375,
0.2197265625,
0.6318359375
],
[
0.2484130859375,
0.62060546875,
0.374267578125,
0.62060546875,
0.374267578125,
0.63525390625,
0.2484130859375,
0.63525390625
],
[
0.2181396484375,
0.6513671875,
0.78173828125,
0.6513671875,
0.78173828125,
0.66552734375,
0.2181396484375,
0.66552734375
],
[
0.21728515625,
0.66650390625,
0.7861328125,
0.66748046875,
0.7861328125,
0.681640625,
0.21728515625,
0.68115234375
],
[
0.21728515625,
0.68115234375,
0.7861328125,
0.6796875,
0.7861328125,
0.6943359375,
0.21728515625,
0.69580078125
],
[
0.2181396484375,
0.6962890625,
0.7861328125,
0.6962890625,
0.7861328125,
0.71044921875,
0.2181396484375,
0.71044921875
],
[
0.2181396484375,
0.7119140625,
0.78515625,
0.7119140625,
0.78515625,
0.72607421875,
0.2181396484375,
0.72607421875
],
[
0.2181396484375,
0.7265625,
0.78759765625,
0.7265625,
0.78759765625,
0.74072265625,
0.2181396484375,
0.74072265625
],
[
0.2181396484375,
0.7421875,
0.78515625,
0.7421875,
0.78515625,
0.755859375,
0.2181396484375,
0.755859375
],
[
0.2197265625,
0.75830078125,
0.2484130859375,
0.75830078125,
0.2484130859375,
0.77099609375,
0.2197265625,
0.77099609375
],
[
0.240234375,
0.771484375,
0.7861328125,
0.77294921875,
0.7861328125,
0.787109375,
0.240234375,
0.7861328125
],
[
0.21728515625,
0.787109375,
0.78515625,
0.787109375,
0.78515625,
0.8017578125,
0.21728515625,
0.8017578125
],
[
0.2181396484375,
0.80322265625,
0.7861328125,
0.80322265625,
0.7861328125,
0.81640625,
0.2181396484375,
0.81640625
],
[
0.2197265625,
0.82470703125,
0.361083984375,
0.82568359375,
0.361083984375,
0.84033203125,
0.2197265625,
0.8388671875
]
] |
[
"DronePose: Photorealistic UAV-Assistant",
"Dataset Synthesis for 3D Pose Estimation via a",
"Smooth Silhouette Loss",
"Georgios Albanis*1, Nikolaos Zioulis*l, Anastasios Dimoul, Dimitrios",
" ",
"Zarpalasl, and Petros Daras!",
"Visual Computing Lab, Information Technologies Institute, Centre for Research and",
"Technology Hellas, Greece",
"(galbanis, nzioulis, dimou, zarpalas, daras)@iti.gr",
"vcl.iti.gr",
"Abstract. In this work we consider UAVs as cooperative agents sup-",
"porting human users in their operations.In this context,the 3D localisa-",
"['S]",
"tion of the UAV assistant is an important task that can facilitate the ex-",
"change of spatial information between the user and the UAV. To address",
"this in a data-driven manner, we design a data synthesis pipeline to cre-",
"ate a realistic multimodal dataset that includes both the exocentric user",
"view, and the egocentric UAV view. We then exploit the joint availability",
"of photorealistic and synthesized inputs to train a single-shot monocular",
"2280802",
"pose estimation model. During training we leverage differentiable ren-",
"dering to supplement a state-of-the-art direct regression objective with",
"anovel smooth silhouetteloss.Our results demonstrate its qualitative",
"and quantitative performance gains over traditional silhouette objectives",
"Our data and code are available at https://vcl3d.github.io/DronePose.",
"Keywords: 3D Pose Estimation, Dataset Generation, UAV, Differen",
"tiable Rendering",
"1.",
"Introduction",
"Advances in robotics, their autonomy, and computer vision are constituting Un",
"manned Aerial Vehicles (UAVs) - or otherwise known as drones - an emerging",
"ubiquitous technology. Indeed, the potential that autonomous or piloted UAVs",
"can provide has already been acknowledged in numerous application fields like",
"filming [69], emergency response [34], animal conservation [30], and infrastruc-",
"ture inspection [17]. In addition, a lot of research has focused on the detection of",
"drones in the context of anti-drone systems [59] and their accurate identification",
"[12]",
"Nevertheless, the miniaturization and commoditisation of drones open up",
"new types of applications for mini-UAVs. Moving beyond the role of remote,",
"and/or malicious agents, personal - friendly - drones can operate in tandem",
"* Equal contribution"
] |
{
"arxiv_id": "2008.08823",
"language": "en",
"page_idx": 0
}
|
||
2008.08823-en-1
|
2008.08823-en-1.jpg
|
2 G. Albanis and N. Zioulis et al.
with human users, supporting their activities and acting as external assistants [4]. Yet, human operators can also be augmented with modern sensing capabilities $(e.g.$ HoloLens), enabling advanced cooperation schemes like XRay vision [16]. Under such conditions, the registration of the human operator and the peer drone is an important task as it facilitates the exchange of spatial information, enabling functionalities like gaze-driven navigation [70].
To address friendly UAV 3D pose estimation in a data-driven manner, we first need to overcome the lack of data. All UAV related datasets either target remote sensing applications, and therefore, offering purely drone egocentric views, or consider drones as malicious agents and offer (multi-modal) sensor exocentric views. However, datasets created for anti-UAV applications cannot be re-used in UAV-Assistant (UAVA) settings mainly because of the very large distance between the sensors and the drones. This is not only out of context, but also suffers from low-resolution modalities $(i.e.$ thermal),tiny object image sizes, and a serious lack of fner-grained discriminating information for 3D pose estimation.
The availability of UAV simulators integrated with high-end rendering inside game engines, like AirSim [58] and Sim4CV [43], offers a promising direction for synthesizing UAVA data. This synthetic dataset generation approach has been used for spacecraft pose estimation [51] and generating low-altitude flight data [30]. Still, the synthetic to real domain gap will hinder performance when applied to actual real-world cases. While approaches like FlightGoggles [20] mitigate this gap through photorealistic sensor simulation, actual real-world data are still preferable. Towards that end, we follow a principled approach to create a realistic multimodal UAVA dataset. We employ real-world 3D scanned datasets, physically-based shading, a gamified simulator for realistic drone navigation trajectory collection and randomized sampling, to generate multimodal data both from the user's exocentric view of the drone, as well as the drone's egocentric view. This allows us to train a CNN for single-shot drone 3D pose estimation from monocular (exocentric) input. Given the availability of projected drone silhouettes aligned with the exocentric view, we complement our model with an enhanced smooth silhouette consistency loss. Overall, our contributions are the following:
- We create a realistic multimodal UAVA dataset that offers aligned exocentric
(user) and egocentric (drone) data along with the ground-truth poses. We also consider an inter-frame motion for both views, with the ego-motion supported by the scene's optical flow, while the exocentric view provides the drone's optical flow.
- We develop an end-to-end model that combines direct pose regression and differentiable rendering for learning 6DOF pose estimation and train it with complex color images to explore their combined potential.
- We propose a smooth silhouette loss that improves performance compared
to traditional alternatives as well as model robustness.
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.271240234375,
0.11553955078125,
0.48193359375,
0.11553955078125,
0.48193359375,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.2181396484375,
0.1483154296875,
0.78173828125,
0.1483154296875,
0.78173828125,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2197265625,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2197265625,
0.177978515625
],
[
0.2197265625,
0.1787109375,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.2197265625,
0.1925048828125
],
[
0.21728515625,
0.1925048828125,
0.7861328125,
0.1932373046875,
0.7861328125,
0.2083740234375,
0.21728515625,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.21728515625,
0.222900390625,
0.61376953125,
0.2235107421875,
0.61376953125,
0.238037109375,
0.21728515625,
0.2374267578125
],
[
0.24267578125,
0.2442626953125,
0.78369140625,
0.2442626953125,
0.78369140625,
0.2587890625,
0.24267578125,
0.2587890625
],
[
0.2181396484375,
0.260009765625,
0.78515625,
0.260009765625,
0.78515625,
0.2734375,
0.2181396484375,
0.2734375
],
[
0.2181396484375,
0.27587890625,
0.7861328125,
0.27587890625,
0.7861328125,
0.2890625,
0.2181396484375,
0.2890625
],
[
0.2181396484375,
0.290283203125,
0.78515625,
0.290283203125,
0.78515625,
0.3037109375,
0.2181396484375,
0.3037109375
],
[
0.21728515625,
0.304931640625,
0.78759765625,
0.304931640625,
0.78759765625,
0.3193359375,
0.21728515625,
0.3193359375
],
[
0.2181396484375,
0.32080078125,
0.7861328125,
0.32080078125,
0.7861328125,
0.333984375,
0.2181396484375,
0.333984375
],
[
0.2181396484375,
0.335205078125,
0.7861328125,
0.335205078125,
0.7861328125,
0.348388671875,
0.2181396484375,
0.348388671875
],
[
0.2181396484375,
0.35107421875,
0.7861328125,
0.35107421875,
0.7861328125,
0.3642578125,
0.2181396484375,
0.3642578125
],
[
0.21728515625,
0.365478515625,
0.78173828125,
0.365478515625,
0.78173828125,
0.37890625,
0.21728515625,
0.37890625
],
[
0.24267578125,
0.385009765625,
0.78515625,
0.385009765625,
0.78515625,
0.3984375,
0.24267578125,
0.3984375
],
[
0.2197265625,
0.40283203125,
0.7861328125,
0.40283203125,
0.7861328125,
0.416748046875,
0.2197265625,
0.416748046875
],
[
0.2181396484375,
0.417236328125,
0.7861328125,
0.417236328125,
0.7861328125,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.2181396484375,
0.431884765625,
0.78515625,
0.431884765625,
0.78515625,
0.44580078125,
0.2181396484375,
0.44580078125
],
[
0.2181396484375,
0.4462890625,
0.7861328125,
0.4462890625,
0.7861328125,
0.4609375,
0.2181396484375,
0.4609375
],
[
0.2181396484375,
0.462158203125,
0.78515625,
0.462158203125,
0.78515625,
0.47607421875,
0.2181396484375,
0.47607421875
],
[
0.2181396484375,
0.477783203125,
0.7861328125,
0.477783203125,
0.7861328125,
0.4912109375,
0.2181396484375,
0.4912109375
],
[
0.2181396484375,
0.492431640625,
0.7861328125,
0.492431640625,
0.7861328125,
0.505859375,
0.2181396484375,
0.505859375
],
[
0.2181396484375,
0.50830078125,
0.78515625,
0.50830078125,
0.78515625,
0.521484375,
0.2181396484375,
0.521484375
],
[
0.2181396484375,
0.52294921875,
0.78369140625,
0.52294921875,
0.78369140625,
0.537109375,
0.2181396484375,
0.537109375
],
[
0.214111328125,
0.53515625,
0.78759765625,
0.5361328125,
0.78759765625,
0.5537109375,
0.214111328125,
0.55322265625
],
[
0.2181396484375,
0.55322265625,
0.78515625,
0.55322265625,
0.78515625,
0.56640625,
0.2181396484375,
0.56640625
],
[
0.2197265625,
0.568359375,
0.7861328125,
0.568359375,
0.7861328125,
0.58203125,
0.2197265625,
0.58203125
],
[
0.2181396484375,
0.583984375,
0.78515625,
0.583984375,
0.78515625,
0.59765625,
0.2181396484375,
0.59765625
],
[
0.2181396484375,
0.59765625,
0.78515625,
0.59765625,
0.78515625,
0.6123046875,
0.2181396484375,
0.6123046875
],
[
0.2181396484375,
0.61279296875,
0.7861328125,
0.61279296875,
0.7861328125,
0.626953125,
0.2181396484375,
0.626953125
],
[
0.21728515625,
0.626953125,
0.289306640625,
0.62890625,
0.289306640625,
0.64306640625,
0.21728515625,
0.6416015625
],
[
0.238525390625,
0.68115234375,
0.2467041015625,
0.68115234375,
0.2467041015625,
0.68603515625,
0.238525390625,
0.68603515625
],
[
0.240234375,
0.67822265625,
0.78515625,
0.67822265625,
0.78515625,
0.69140625,
0.240234375,
0.69140625
],
[
0.2467041015625,
0.693359375,
0.7861328125,
0.693359375,
0.7861328125,
0.70703125,
0.2467041015625,
0.70703125
],
[
0.2459716796875,
0.708984375,
0.7861328125,
0.708984375,
0.7861328125,
0.72265625,
0.2459716796875,
0.72265625
],
[
0.2459716796875,
0.7236328125,
0.78515625,
0.72265625,
0.78515625,
0.7373046875,
0.2459716796875,
0.73779296875
],
[
0.2459716796875,
0.7392578125,
0.38818359375,
0.7392578125,
0.38818359375,
0.75244140625,
0.2459716796875,
0.75244140625
],
[
0.2442626953125,
0.75830078125,
0.7861328125,
0.7587890625,
0.7861328125,
0.7734375,
0.2442626953125,
0.77294921875
],
[
0.2467041015625,
0.7744140625,
0.7861328125,
0.7744140625,
0.7861328125,
0.78857421875,
0.2467041015625,
0.78857421875
],
[
0.2459716796875,
0.79052734375,
0.6611328125,
0.79052734375,
0.6611328125,
0.8037109375,
0.2459716796875,
0.8037109375
],
[
0.2296142578125,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82568359375,
0.2296142578125,
0.82568359375
],
[
0.2459716796875,
0.8271484375,
0.6357421875,
0.8271484375,
0.6357421875,
0.84033203125,
0.2459716796875,
0.84033203125
]
] |
[
"2",
"G. Albanis and N. Zioulis et al.",
"with human users, supporting their activities and acting as external assistants [4]",
"Yet, human operators can also be augmented with modern sensing capabilities",
"(e.g. HoloLens), enabling advanced cooperation schemes like XRay vision [16]]",
"Under such conditions, the registration of the human operator and the peer",
"drone is an important task as it facilitates the exchange of spatial information.",
"enabling functionalities like gaze-driven navigation [70]",
"To address friendly UAV 3D pose estimation in a data-driven manner, we first",
"need to overcome the lack of data. All UAV related datasets either target remote",
"sensing applications, and therefore, offering purely drone egocentric views, or",
"consider drones as malicious agents and offer (multi-modal) sensor exocentric",
"views. However, datasets created for anti-UAV applications cannot be re-used",
"in UAV-Assistant (UAVA) settings mainly because of the very large distance",
"between the sensors and the drones. This is not only out of context, but also",
"suffers from low-resolution modalities (i.e. thermal), tiny object image sizes, and",
"a serious lack of finer-grained discriminating information for 3D pose estimation",
"The availability of UAV simulators integrated with high-end rendering inside",
"game engines, like AirSim [58] and Sim4CV [43], offers a promising direction",
"for synthesizing UAVA data. This synthetic dataset generation approach has",
"been used for spacecraft pose estimation [51] and generating low-altitude flight",
"data [30]. Still, the synthetic to real domain gap will hinder performance when",
"applied to actual real-world cases. While approaches like FlightGoggles [20] mit-",
"igate this gap through photorealistic sensor simulation, actual real-world data",
"are still preferable. Towards that end, we follow a principled approach to create a",
"realistic multimodal UAVA dataset. We employ real-world 3D scanned datasets,",
"physically-based shading, a gamified simulator for realistic drone navigation tra",
"jectory collection and randomized sampling, to generate multimodal data both",
"from the user's exocentric view of the drone, as well as the drone's egocentric",
"view. This allows us to train a CNN for single-shot drone 3D pose estimation",
"from monocular (exocentric) input. Given the availability of projected drone sil-",
"houettes aligned with the exocentric view, we complement our model with an",
"enhanced smooth silhouette consistency loss. Overall, our contributions are the",
"following:",
"",
"We create a realistic multimodal UAVA dataset that offers aligned exocentric",
"(user) and egocentric (drone) data along with the ground-truth poses. We",
"also consider an inter-frame motion for both views, with the ego-motion",
"supported by the scene's optical flow, while the exocentric view provides the",
"drone's optical flow",
"We develop an end-to-end model that combines direct pose regression and",
"differentiable rendering for learning 6DOF pose estimation and train it with",
"complex color images to explore their combined potential.",
"- We propose a smooth silhouette loss that improves performance compared",
"to traditional alternatives as well as model robustness."
] |
{
"arxiv_id": "2008.08823",
"language": "en",
"page_idx": 1
}
|
||
2008.08823-en-2
|
2008.08823-en-2.jpg
|
DronePose 3
## 2 Related Work
UAV Datasets: Most existing UAV datasets can be categorized into two main categories, datasets for remote sensing and anti-UAV datasets. The former utilize drones as remote eyes observing remote scenes and aim at either applying real-time computer vision algorithms on live data streams, or recording them for offline analysis. The VisDrone2019 dataset [74] contains diverse data captured from drone-mounted cameras and is annotated with bounding boxes for object detection and tracking tasks, focused on the surveillance aspect of UAVs. Similar data is included in the Stanford Drone Dataset [55] with various bounding box annotations of different classes found in a campus, with the distinct difference being its strict top-down view. The VIRAT video surveillance dataset [44] also contains UAV captured scenes which, apart from object annotations, also includes activity and event annotations. Apart from these close-range egocentric view datasets, typical remote sensing drone datasets involve ground observing aerial images. Examples include the urban settlement dataset of [39], or the Kuzikus dataset used in [30] which contains wildlife detection annotations. Similar aerial top-down view datasets are used for object counting [25] and tracking [42]. Summarizing, datasets that view UAVs as remote sensors, offer the drone's egocentric view, usually with object $(i.e.$ bounding box) or image segment $(i.e.$ mask) annotations.
The second UAV dataset category, considers UAVs as hostile agents, a topic that is gaining significant traction with the research community lately, mainly because of the challenges it entails. Anti-UAV systems need to operate in-thewild, in adverse weather conditions, during the whole day-night cycle, and offer timely and accurate information for any neutralizing actions to take place. Towards that end, the available datasets address tiny object detection [56] and adversarial conditions that involve birds [12]. Recent organized challenges like the Anti-UAV challenge [1] include multimodal $(i.e.$ color and thermal) inputs to increase detection robustness under different lighting and weather conditions.
All the aforementioned datasets are captured with traditional sensors and are manually annotated. Nonetheless, recent advances in UAV simulators, a prominent one being AirSim [58], enable the generation of data via synthesis, and thus, the availability of information that is traditionally very hard to acquire with traditional means $(i.e$ depth, surface orientation, semantics, 3D pose, etc.). This was the case of the AimSim-w dataset [5] which synthesized remote sensing data for wildlife observation, in addition to modelling the thermal imaging process to synthesise infrared data. Similarly, the Mid-Air dataset [18] used AirSim to synthesize low-altitude flight multimodal data and illustrated its efficacy in drone depth estimation.
Still, the domain gap is an important obstacle in using fully synthetic datasets that employ purely computer-generated models and imagery. While domain adaptation itself is another important task, with the reader referred to a survey [66] for more details, we can also find other intermediary approaches in the literature. FlightGoggles [20] is a photorealistic sensor simulator that infuses realism in a data generation pipeline for UAVs. It was used when creating the
|
[
[
0.66015625,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1275634765625,
0.66015625,
0.1275634765625
],
[
0.77294921875,
0.11553955078125,
0.7890625,
0.11553955078125,
0.7890625,
0.1300048828125,
0.77294921875,
0.1300048828125
],
[
0.2197265625,
0.1483154296875,
0.2344970703125,
0.1483154296875,
0.2344970703125,
0.1590576171875,
0.2197265625,
0.1590576171875
],
[
0.2418212890625,
0.146484375,
0.386474609375,
0.146484375,
0.386474609375,
0.1610107421875,
0.2418212890625,
0.1610107421875
],
[
0.2197265625,
0.1767578125,
0.78515625,
0.1767578125,
0.78515625,
0.1900634765625,
0.2197265625,
0.1900634765625
],
[
0.2181396484375,
0.1925048828125,
0.78515625,
0.1925048828125,
0.78515625,
0.205810546875,
0.2181396484375,
0.205810546875
],
[
0.2156982421875,
0.20703125,
0.7861328125,
0.2083740234375,
0.7861328125,
0.222900390625,
0.2156982421875,
0.2215576171875
],
[
0.21728515625,
0.222900390625,
0.78515625,
0.222900390625,
0.78515625,
0.236083984375,
0.21728515625,
0.236083984375
],
[
0.2181396484375,
0.2374267578125,
0.7861328125,
0.2374267578125,
0.7861328125,
0.251953125,
0.2181396484375,
0.251953125
],
[
0.21728515625,
0.25244140625,
0.7861328125,
0.25244140625,
0.7861328125,
0.266357421875,
0.21728515625,
0.266357421875
],
[
0.2181396484375,
0.268310546875,
0.78369140625,
0.268310546875,
0.78369140625,
0.2822265625,
0.2181396484375,
0.2822265625
],
[
0.2181396484375,
0.28271484375,
0.7861328125,
0.28271484375,
0.7861328125,
0.296630859375,
0.2181396484375,
0.296630859375
],
[
0.2181396484375,
0.29736328125,
0.78515625,
0.29736328125,
0.78515625,
0.3125,
0.2181396484375,
0.3125
],
[
0.2181396484375,
0.313232421875,
0.7861328125,
0.313232421875,
0.7861328125,
0.326904296875,
0.2181396484375,
0.326904296875
],
[
0.2197265625,
0.32763671875,
0.7861328125,
0.32763671875,
0.7861328125,
0.34228515625,
0.2197265625,
0.34228515625
],
[
0.21728515625,
0.34228515625,
0.78369140625,
0.343505859375,
0.78369140625,
0.35791015625,
0.21728515625,
0.356689453125
],
[
0.21728515625,
0.359130859375,
0.78369140625,
0.359130859375,
0.78369140625,
0.37255859375,
0.21728515625,
0.37255859375
],
[
0.2181396484375,
0.373779296875,
0.78369140625,
0.373779296875,
0.78369140625,
0.386962890625,
0.2181396484375,
0.386962890625
],
[
0.21728515625,
0.38818359375,
0.78369140625,
0.38818359375,
0.78369140625,
0.401611328125,
0.21728515625,
0.401611328125
],
[
0.2181396484375,
0.404052734375,
0.78515625,
0.404052734375,
0.78515625,
0.417236328125,
0.2181396484375,
0.417236328125
],
[
0.2181396484375,
0.41845703125,
0.78515625,
0.41845703125,
0.78515625,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.2181396484375,
0.434326171875,
0.7861328125,
0.434326171875,
0.7861328125,
0.447509765625,
0.2181396484375,
0.447509765625
],
[
0.2197265625,
0.45068359375,
0.451904296875,
0.45068359375,
0.451904296875,
0.4609375,
0.2197265625,
0.4609375
],
[
0.24267578125,
0.46337890625,
0.7861328125,
0.46337890625,
0.7861328125,
0.477783203125,
0.24267578125,
0.477783203125
],
[
0.2181396484375,
0.479248046875,
0.7861328125,
0.479248046875,
0.7861328125,
0.49365234375,
0.2181396484375,
0.49365234375
],
[
0.2181396484375,
0.494384765625,
0.78369140625,
0.494384765625,
0.78369140625,
0.50830078125,
0.2181396484375,
0.50830078125
],
[
0.2181396484375,
0.51025390625,
0.7861328125,
0.51025390625,
0.7861328125,
0.5234375,
0.2181396484375,
0.5234375
],
[
0.2181396484375,
0.5244140625,
0.78369140625,
0.5244140625,
0.78369140625,
0.53857421875,
0.2181396484375,
0.53857421875
],
[
0.2181396484375,
0.54052734375,
0.78759765625,
0.54052734375,
0.78759765625,
0.5537109375,
0.2181396484375,
0.5537109375
],
[
0.2181396484375,
0.5546875,
0.7861328125,
0.5546875,
0.7861328125,
0.568359375,
0.2181396484375,
0.568359375
],
[
0.21728515625,
0.568359375,
0.78515625,
0.568359375,
0.78515625,
0.58203125,
0.21728515625,
0.58203125
],
[
0.2181396484375,
0.58544921875,
0.78369140625,
0.58544921875,
0.78369140625,
0.5986328125,
0.2181396484375,
0.5986328125
],
[
0.2459716796875,
0.599609375,
0.7861328125,
0.599609375,
0.7861328125,
0.61279296875,
0.2459716796875,
0.61279296875
],
[
0.21728515625,
0.61572265625,
0.78515625,
0.61572265625,
0.78515625,
0.62890625,
0.21728515625,
0.62890625
],
[
0.21728515625,
0.6298828125,
0.78759765625,
0.6298828125,
0.78759765625,
0.64306640625,
0.21728515625,
0.64306640625
],
[
0.21728515625,
0.64453125,
0.78515625,
0.64453125,
0.78515625,
0.65771484375,
0.21728515625,
0.65771484375
],
[
0.2181396484375,
0.66015625,
0.78515625,
0.66015625,
0.78515625,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.2181396484375,
0.67529296875,
0.78759765625,
0.67529296875,
0.78759765625,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.21728515625,
0.68994140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.21728515625,
0.7041015625
],
[
0.21728515625,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.7197265625,
0.21728515625,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2197265625,
0.7373046875,
0.3896484375,
0.7373046875,
0.3896484375,
0.74755859375,
0.2197265625,
0.74755859375
],
[
0.2418212890625,
0.75048828125,
0.78759765625,
0.75048828125,
0.78759765625,
0.7646484375,
0.2418212890625,
0.7646484375
],
[
0.21728515625,
0.76513671875,
0.78515625,
0.76513671875,
0.78515625,
0.77880859375,
0.21728515625,
0.77880859375
],
[
0.2181396484375,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.78759765625,
0.796875,
0.78759765625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.21728515625,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82568359375,
0.21728515625,
0.82568359375
],
[
0.21728515625,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"DronePose",
"3",
"2",
"Related Work",
"UAV Datasets: Most existing UAV datasets can be categorized into two main",
"categories, datasets for remote sensing and anti-UAV datasets. The former uti-",
"lize drones as remote eyes observing remote scenes and aim at either applying",
"real-time computer vision algorithms on live data streams, or recording them for",
"offline analysis. The VisDrone2019 dataset [74] contains diverse data captured",
"from drone-mounted cameras and is annotated with bounding boxes for object",
"detection and tracking tasks, focused on the surveillance aspect of UAVs. Sim-",
"ilar data is included in the Stanford Drone Dataset [55] with various bounding",
"box annotations of different classes found in a campus, with the distinct dif-",
"ference being its strict top-down view. The VIRAT video surveillance dataset",
"[44] also contains UAV captured scenes which, apart from object annotations,",
"also includes activity and event annotations. Apart from these close-range ego-",
"centric view datasets, typical remote sensing drone datasets involve ground ob-",
"serving aerial images. Examples include the urban settlement dataset of [39]",
"or the Kuzikus dataset used in [3ol which contains wildlife detection annota-",
"tions. Similar aerial top-down view datasets are used for object counting [25]",
"and tracking [42]. Summarizing, datasets that view UAVs as remote sensors, of-",
"fer the drone's egocentric view, usually with object (i.e. bounding box) or image",
"segment (z.e.mask)annotations",
"The second UAV dataset category, considers UAVs as hostile agents, a topic",
"that is gaining significant traction with the research community lately, mainly",
"because of the challenges it entails. Anti-UAV systems need to operate in-the-",
"wild, in adverse weather conditions, during the whole day-night cycle, and offer",
"timely and accurate information for any neutralizing actions to take place. To-",
"wards that end, the available datasets address tiny object detection [56] and",
"adversarial conditions that involve birds [12]. Recent organized challenges like",
"the Anti-UAV challenge [1l include multimodal (i.e. color and thermal) inputs",
"to increase detection robustness under different lighting and weather conditions.",
"All the aforementioned datasets are captured with traditional sensors and are",
"manually annotated. Nonetheless, recent advances in UAV simulators, a promi-",
"nent one being AirSim [5sl, enable the generation of data via synthesis, and",
"thus, the availability of information that is traditionally very hard to acquire",
"with traditional means (i.e depth, surface orientation, semantics, 3D pose, etc.).",
"This was the case of the AimSim-w dataset [5] which synthesized remote sensing",
"data for wildlife observation, in addition to modelling the thermal imaging pro-",
"cess to synthesise infrared data. Similarly, the Mid-Air dataset [18] used AirSim",
"to synthesize low-altitude flight multimodal data and illustrated its efficacy in",
"drone depth estimation.",
"Still, the domain gap is an important obstacle in using fully synthetic datasets",
"that employ purely computer-generated models and imagery. While domain",
"adaptation itself is another important task, with the reader referred to a sur-",
"vey [66] for more details, we can also find other intermediary approaches in the",
"literature. FlightGoggles [20] is a photorealistic sensor simulator that infuses",
"realism in a data generation pipeline for UAVs. It was used when creating the"
] |
{
"arxiv_id": "2008.08823",
"language": "en",
"page_idx": 2
}
|
||
2008.08823-en-3
|
2008.08823-en-3.jpg
|
4 G. Albanis and N. Zioulis et al
aggressive drone flights Blackbird dataset [3], which does not fit in the two aforementioned categories in the broader sense, as it is used for visual-inertial SLAM benchmarking. Still, the synthesis of realistic data, as encompassed by FlightGoggles and the Blackbird dataset is highly related to our approach. We also use scanned 3D models like the photogrammetry models used in FlightGoggles and generate external trajectories like those used the Blackbird dataset. Contrary to those though, we collect UAV flight trajectories with significantly higher variance than the periodic Blackbird ones, and utilize a vastly larger 3D scene corpus than FlightGoggles. Moreover, as it is apparent by the above dataset analysis (with more details available in the supplement), none are suitable for UAVA applications as the exocentric anti-UAV datasets image drones from very far distances, similar to how egocentric remote sensing drone datasets image (very) far scenes. In the UAVA setting, the user and the drone work under a cooperative context that allows for the availability of both the egocentric and exocentric views simultaneously, and will typically be close in close proximity.
Data-driven 6DOF Pose Estimation: Even before the establishment of deep models, data-driven methods relying on random forests [14,61,6] or deformable part models [49] started producing high-quality results for the task of 3D pose estimation. Early deep models [41,21,60,57] focused on learning rotations $(i.e.$ viewpoints) and performed classification after binning them, using either normal maps [21],synthesized [60] or traditional [41] images, or even employed colorized depth inputs [57]. Follow-up works transitioned from binned rotation classification to direct regression of the 6DOF camera pose using varying rotation representations. PoseCNN [68] used a quaternion representation and employed learned segmentation predictions to handle cluttered object scenes, while [40] demonstrated results for an axis-angle representation in addition to the quaternion one. Currently, the state-of-the-art has been focusing on addressing the discontinuity of rotations by disentangling the rotation manifold [37] or exploring continuous rotation representations [73]. Single-shot variants predict the 3D bounding boxes from monocular inputs [62], or formulate multiple pose hypotheses from 2D bounding box detections [29], solving either a Perspectiven-Point (PnP) problem or locally optimizing the pose afterwards respectively. A slightly different approach has been introduced in [52] where the image is cropped around the object by frst predicting a segmentation mask and then the 2D bounding box is predicted. While most works are trained on real-world images, in [ 53] it was shown that using synthetic images and single channel edge information can lead to high-quality results, even when directly regressing the 6DOF pose. Even though direct pose regression has demonstrated high-quality results, especially under iterative frameworks [64], most recent research has resorted to correspondence-based pose estimation. Deep keypoint prediction was originally used for 3D pose estimation in [63] and has since spun to a variety of approaches even in drone pose estimation[28]. PVNet [48] votes for keypoints, which are then used to estimate the object's 6DOF pose using PnP. Following a dense correspondence approach, various works directly regress either 2D image to 3D model correspondences [71], normalized object coordinates [65], or leverage
|
[
[
0.223876953125,
0.11553955078125,
0.2320556640625,
0.121826171875,
0.223876953125,
0.1275634765625,
0.2156982421875,
0.121826171875
],
[
0.271240234375,
0.11553955078125,
0.48193359375,
0.11553955078125,
0.48193359375,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.2156982421875,
0.147705078125,
0.78369140625,
0.146484375,
0.78369140625,
0.1610107421875,
0.2156982421875,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.1767578125,
0.2181396484375,
0.1767578125
],
[
0.2181396484375,
0.177978515625,
0.78515625,
0.177978515625,
0.78515625,
0.1912841796875,
0.2181396484375,
0.1912841796875
],
[
0.21728515625,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.21728515625,
0.20703125
],
[
0.21728515625,
0.20703125,
0.78515625,
0.2083740234375,
0.78515625,
0.222900390625,
0.21728515625,
0.2215576171875
],
[
0.2181396484375,
0.2235107421875,
0.78369140625,
0.2235107421875,
0.78369140625,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.21728515625,
0.238037109375,
0.7861328125,
0.238037109375,
0.7861328125,
0.25244140625,
0.21728515625,
0.25244140625
],
[
0.21728515625,
0.25390625,
0.78759765625,
0.25390625,
0.78759765625,
0.267578125,
0.21728515625,
0.267578125
],
[
0.2197265625,
0.270751953125,
0.78515625,
0.270751953125,
0.78515625,
0.281005859375,
0.2197265625,
0.281005859375
],
[
0.2181396484375,
0.2841796875,
0.7861328125,
0.2841796875,
0.7861328125,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.2181396484375,
0.29736328125,
0.78515625,
0.2998046875,
0.78515625,
0.313232421875,
0.2181396484375,
0.311279296875
],
[
0.2181396484375,
0.314453125,
0.78515625,
0.314453125,
0.78515625,
0.32763671875,
0.2181396484375,
0.32763671875
],
[
0.2181396484375,
0.328857421875,
0.7861328125,
0.328857421875,
0.7861328125,
0.343505859375,
0.2181396484375,
0.343505859375
],
[
0.2197265625,
0.34521484375,
0.78515625,
0.34521484375,
0.78515625,
0.35595703125,
0.2197265625,
0.35595703125
],
[
0.2181396484375,
0.359130859375,
0.775390625,
0.359130859375,
0.775390625,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.24267578125,
0.386962890625,
0.78759765625,
0.386962890625,
0.78759765625,
0.40087890625,
0.24267578125,
0.40087890625
],
[
0.2197265625,
0.40478515625,
0.78369140625,
0.40478515625,
0.78369140625,
0.415283203125,
0.2197265625,
0.415283203125
],
[
0.2181396484375,
0.41845703125,
0.7861328125,
0.41845703125,
0.7861328125,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.21728515625,
0.434326171875,
0.78369140625,
0.434326171875,
0.78369140625,
0.447509765625,
0.21728515625,
0.447509765625
],
[
0.21728515625,
0.447509765625,
0.7861328125,
0.448974609375,
0.7861328125,
0.46337890625,
0.21728515625,
0.462158203125
],
[
0.2181396484375,
0.46337890625,
0.78515625,
0.46337890625,
0.78515625,
0.477783203125,
0.2181396484375,
0.477783203125
],
[
0.2181396484375,
0.479248046875,
0.78369140625,
0.479248046875,
0.78369140625,
0.492431640625,
0.2181396484375,
0.492431640625
],
[
0.2156982421875,
0.49365234375,
0.7861328125,
0.494384765625,
0.7861328125,
0.50830078125,
0.2156982421875,
0.5068359375
],
[
0.2181396484375,
0.5087890625,
0.7861328125,
0.5087890625,
0.7861328125,
0.5234375,
0.2181396484375,
0.5234375
],
[
0.2181396484375,
0.5244140625,
0.78515625,
0.5244140625,
0.78515625,
0.53857421875,
0.2181396484375,
0.53857421875
],
[
0.21728515625,
0.5390625,
0.7861328125,
0.53857421875,
0.7861328125,
0.55322265625,
0.21728515625,
0.5537109375
],
[
0.2197265625,
0.55615234375,
0.78369140625,
0.55615234375,
0.78369140625,
0.56640625,
0.2197265625,
0.56640625
],
[
0.2181396484375,
0.5693359375,
0.7861328125,
0.5693359375,
0.7861328125,
0.58349609375,
0.2181396484375,
0.58349609375
],
[
0.21728515625,
0.583984375,
0.78369140625,
0.583984375,
0.78369140625,
0.59765625,
0.21728515625,
0.59765625
],
[
0.2181396484375,
0.599609375,
0.7861328125,
0.599609375,
0.7861328125,
0.61279296875,
0.2181396484375,
0.61279296875
],
[
0.2156982421875,
0.61279296875,
0.78173828125,
0.6142578125,
0.78173828125,
0.62890625,
0.2156982421875,
0.62744140625
],
[
0.21728515625,
0.6298828125,
0.78515625,
0.6298828125,
0.78515625,
0.64306640625,
0.21728515625,
0.64306640625
],
[
0.2181396484375,
0.64599609375,
0.78759765625,
0.64599609375,
0.78759765625,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.2181396484375,
0.66015625,
0.7861328125,
0.66015625,
0.7861328125,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.2156982421875,
0.6748046875,
0.7861328125,
0.6748046875,
0.7861328125,
0.68798828125,
0.2156982421875,
0.68798828125
],
[
0.21728515625,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.21728515625,
0.705078125
],
[
0.21728515625,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.7197265625,
0.21728515625,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2197265625,
0.7373046875,
0.78173828125,
0.7373046875,
0.78173828125,
0.74755859375,
0.2197265625,
0.74755859375
],
[
0.2181396484375,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.78759765625,
0.7666015625,
0.78759765625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.78076171875,
0.78515625,
0.77978515625,
0.78515625,
0.79443359375,
0.21728515625,
0.79541015625
],
[
0.2156982421875,
0.79443359375,
0.7861328125,
0.79541015625,
0.7861328125,
0.81005859375,
0.2156982421875,
0.8095703125
],
[
0.21728515625,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82470703125,
0.21728515625,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"M",
"G. Albanis and N. Zioulis et al.",
"aggressive drone flights Blackbird dataset 3l, which does not fit in the two afore",
"mentioned categories in the broader sense, as it is used for visual-inertial SLAM",
"benchmarking. Still, the synthesis of realistic data, as encompassed by Flight-",
"Goggles and the Blackbird dataset is highly related to our approach. We also",
"use scanned 3D models like the photogrammetry models used in FlightGoggles",
"and generate external trajectories like those used the Blackbird dataset. Con-",
"trary to those though, we collect UAV flight trajectories with significantly higher",
"variance than the periodic Blackbird ones, and utilize a vastly larger 3D scene",
"corpusthan FlightGoggles.Moreover,asitis apparentbythe above dataset",
"analysis (with more details available in the supplement), none are suitable for",
"UAVA applications as the exocentric anti-UAV datasets image drones from very",
"far distances, similar to how egocentric remote sensing drone datasets image",
"(very) far scenes. In the UAVA setting, the user and the drone work under a",
"cooperative context that allows for the availability of both theegocentric and",
"exocentric views simultaneously, and will typically be close in close proximity.",
"Data-driven 6DOF Pose Estimation: Even before the establishment of",
"deep models, data-driven methods relying on random forests14,61.6l or de-",
"formable part models [49] started producing high-quality results for the task",
"of 3D pose estimation. Early deep models [41,21,60,57] focused on learning ro-",
"tations (i.e. viewpoints) and performed classification after binning them, using",
"either normal maps [21], synthesized [60] or traditional [41] images, or even em-",
"ployed colorized depth inputs [57]. Follow-up works transitioned from binned ro-",
"tation classification to direct regression of the 6DOF camera pose using varying",
"rotation representations. PoseCNN [68] used a quaternion representation and",
"employed learned segmentation predictions to handle cluttered object scenes,",
"while [40] demonstrated results for an axis-angle representation in addition to",
"the quaternion one.Currently,the state-of-the-arthas been focusing on address-",
"ing the discontinuity of rotations by disentangling the rotation manifold [37] or",
"exploring continuous rotation representations [73]. Single-shot variants predict",
"the 3D bounding boxes from monocular inputs [62l, or formulate multiple pose",
"hypotheses from 2D bounding box detections [29], solving either a Perspective-",
"n-Point (PnP) problem or locally optimizing the pose afterwards respectively.",
"A slightly different approach has been introduced in [52] where the image is",
"cropped around the object by first predicting a segmentation mask and then",
"the 2D bounding box is predicted. While most works are trained on real-world",
"images, in [53] it was shown that using synthetic images and single channel edge",
"information can lead to high-quality results, even when directly regressing the",
"6DOF pose. Even though direct pose regression has demonstrated high-quality",
"results,especially underiterativeframeworks64most recent researchhas re",
"sorted to correspondence-based pose estimation. Deep keypoint prediction was",
"originally used for 3D pose estimation in [63] and has since spun to a variety of",
"approaches even in drone pose estimation[28]. PVNet [4s] votes for keypoints,",
"which are then used to estimate the object's 6DOF pose using PnP. Following a",
"dense correspondence approach, various works directly regress either 2D image",
"to 3D model correspondences [71], normalized object coordinates [65], or leverage"
] |
{
"arxiv_id": "2008.08823",
"language": "en",
"page_idx": 3
}
|
||
2008.08823-en-4
|
2008.08823-en-4.jpg
|
DronePose 5
deep Hough voting as in the case of PVN3D [23]. Regarding coordinates based regression, separate branches for the translation and rotational components were found to be more robust in predict pose [36],especially for occluded and textureless objects. More recently, the various advances in differentiable rendering are enabling end-to-end 3D vision, and naturally, their application to 6DOF pose estimation is being explored. A differentiable point cloud projection function was employed [26] that used two simplistic views $(i.e.$ white background, silhouettelike images) of an object to jointly estimate its shape and pose, using an MSE silhouette loss. Similarly, a rendering-like occlusion removal operator was added in [67] to handle point cloud viewpoint-based self-occlusions and drive supervision solely from depth images. Given that only the pose needs to be optimized, a lightweight abstract rendering mechanism based on OpenDR [38] was developed in [50] to compare learned feature representations from meshes and the observed $\begin{array}{c}\text{images, and then estimate their poses in the observed scenes. LatentFusion [46]}\end{array}$ learns a reconstructed feature representation that can be rendered in order to optimize for an object's pose as it is observed within an image, and has been shown to also generalize well to unseen objects. Finally, a rendered silhouette error is backpropagated in [45] to refine the CNN estimated pose. However, it was only applied in simplistic CAD renderings and was not trained in an end-toend fashion. This was also the case for [26] which did not use monocular input and only learned pose estimation using simplistic images. Instead, we leverage end-to-end differentiable rendering using complex imagery for boosting 6DOF pose estimation.
## 3 Dataset
Our goal is to create a realistic UAVA dataset for applications in which users and drones are considered cooperative agents. To achieve this, we need to overcome a set of challenges associated with the photorealism of the visual data, the plausibility of the drone-to-user spatial relation, the variety/plurality of the generated views, and the drone's trajectories. Contrary to similarly oriented datasets like URSO [51] and Blackbird [3] we are not confined to only a single ego- [3] or exocentric view [51], and thus, we need to consider both viewpoints and synthesize temporally and spatially aligned data. While both [51] and [3] render purely computer-generated scenes, we rely on a recent 3D scanned dataset, Matterport3D [9] that offers us scene photorealism.
The Matterport3D dataset was created after capturing multiple RGB-D panoramas within large buildings. As the poses of each panorama are available, we can leverage them as stationary points $(i.e.$ anchors) to drive our UAVA oriented data generation. Aiming to create realistic drone flight trajectories and simultaneously position user viewpoints in a plausible way, we employ a gamified approach. More specifically, we designed and developed a Unity3D game where collectible cube“coins”are placed at each of the known panorama positions (anchors) and use AirSim's drone flight simulator to let players navigate a drone within each building to collect the pre-placed“coins”using a game controller.
|
[
[
0.6611328125,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1275634765625,
0.6611328125,
0.1275634765625
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.77392578125,
0.1287841796875
],
[
0.2181396484375,
0.147705078125,
0.78759765625,
0.147705078125,
0.78759765625,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2197265625,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.177978515625,
0.2197265625,
0.177978515625
],
[
0.2197265625,
0.1787109375,
0.78515625,
0.1787109375,
0.78515625,
0.1925048828125,
0.2197265625,
0.1925048828125
],
[
0.2181396484375,
0.1932373046875,
0.78759765625,
0.1932373046875,
0.78759765625,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.7861328125,
0.208984375,
0.7861328125,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.7861328125,
0.2235107421875,
0.7861328125,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.2181396484375,
0.2392578125,
0.78515625,
0.2392578125,
0.78515625,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.7861328125,
0.25390625,
0.7861328125,
0.268310546875,
0.2181396484375,
0.268310546875
],
[
0.2181396484375,
0.26953125,
0.78759765625,
0.26953125,
0.78759765625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2181396484375,
0.2841796875,
0.78369140625,
0.2841796875,
0.78369140625,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.2181396484375,
0.314453125,
0.7861328125,
0.314453125,
0.7861328125,
0.32763671875,
0.2181396484375,
0.32763671875
],
[
0.21728515625,
0.330078125,
0.7861328125,
0.330078125,
0.7861328125,
0.343505859375,
0.21728515625,
0.343505859375
],
[
0.2181396484375,
0.3447265625,
0.78515625,
0.3447265625,
0.78515625,
0.35791015625,
0.2181396484375,
0.35791015625
],
[
0.21728515625,
0.35986328125,
0.7861328125,
0.35986328125,
0.7861328125,
0.373779296875,
0.21728515625,
0.373779296875
],
[
0.2181396484375,
0.375,
0.7861328125,
0.375,
0.7861328125,
0.389404296875,
0.2181396484375,
0.389404296875
],
[
0.2181396484375,
0.39013671875,
0.7861328125,
0.39013671875,
0.7861328125,
0.404052734375,
0.2181396484375,
0.404052734375
],
[
0.2181396484375,
0.40478515625,
0.7861328125,
0.40478515625,
0.7861328125,
0.41845703125,
0.2181396484375,
0.41845703125
],
[
0.2181396484375,
0.42041015625,
0.78515625,
0.42041015625,
0.78515625,
0.434326171875,
0.2181396484375,
0.434326171875
],
[
0.2181396484375,
0.43505859375,
0.78515625,
0.43505859375,
0.78515625,
0.448974609375,
0.2181396484375,
0.448974609375
],
[
0.21728515625,
0.449462890625,
0.78515625,
0.45068359375,
0.78515625,
0.46533203125,
0.21728515625,
0.464111328125
],
[
0.2181396484375,
0.46533203125,
0.78515625,
0.46533203125,
0.78515625,
0.479248046875,
0.2181396484375,
0.479248046875
],
[
0.21728515625,
0.48095703125,
0.3349609375,
0.48095703125,
0.3349609375,
0.49560546875,
0.21728515625,
0.49560546875
],
[
0.21728515625,
0.52099609375,
0.236083984375,
0.52099609375,
0.236083984375,
0.53515625,
0.21728515625,
0.53515625
],
[
0.25,
0.52099609375,
0.326904296875,
0.52099609375,
0.326904296875,
0.53515625,
0.25,
0.53515625
],
[
0.21728515625,
0.5537109375,
0.7861328125,
0.55322265625,
0.7861328125,
0.5673828125,
0.21728515625,
0.568359375
],
[
0.2181396484375,
0.5693359375,
0.7861328125,
0.5693359375,
0.7861328125,
0.58349609375,
0.2181396484375,
0.58349609375
],
[
0.2181396484375,
0.58544921875,
0.78369140625,
0.58544921875,
0.78369140625,
0.5986328125,
0.2181396484375,
0.5986328125
],
[
0.2181396484375,
0.60107421875,
0.7861328125,
0.60107421875,
0.7861328125,
0.6142578125,
0.2181396484375,
0.6142578125
],
[
0.2181396484375,
0.61572265625,
0.7861328125,
0.61572265625,
0.7861328125,
0.62890625,
0.2181396484375,
0.62890625
],
[
0.2181396484375,
0.6298828125,
0.7861328125,
0.6298828125,
0.7861328125,
0.64306640625,
0.2181396484375,
0.64306640625
],
[
0.2181396484375,
0.64453125,
0.78369140625,
0.64453125,
0.78369140625,
0.65771484375,
0.2181396484375,
0.65771484375
],
[
0.2181396484375,
0.66015625,
0.78515625,
0.66015625,
0.78515625,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.21728515625,
0.6748046875,
0.78515625,
0.673828125,
0.78515625,
0.68798828125,
0.21728515625,
0.689453125
],
[
0.21728515625,
0.68994140625,
0.53759765625,
0.68994140625,
0.53759765625,
0.7041015625,
0.21728515625,
0.7041015625
],
[
0.240234375,
0.7041015625,
0.78369140625,
0.7041015625,
0.78369140625,
0.71826171875,
0.240234375,
0.71826171875
],
[
0.21728515625,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.734375,
0.21728515625,
0.734375
],
[
0.2181396484375,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.78076171875,
0.7861328125,
0.77978515625,
0.7861328125,
0.79443359375,
0.21728515625,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.81005859375,
0.7861328125,
0.81103515625,
0.7861328125,
0.82568359375,
0.2181396484375,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84130859375,
0.2181396484375,
0.84130859375
]
] |
[
"DronePose",
"5",
"deep Hough voting as in the case of PVN3D [23]. Regarding coordinates based",
"regression, separate branches for the translation and rotational components were",
"found to be more robust in predict pose [36], especially for occluded and texture-",
"less objects. More recently, the various advances in differentiable rendering are",
"enabling end-to-end 3D vision, and naturally, their application to 6DOF pose",
"estimation is being explored. A differentiable point cloud projection function was",
"employed [26] that used two simplistic views (i.e. white background, silhouette-",
"like images) of an object to jointly estimate its shape and pose, using an MSE",
"silhouette loss. Similarly, a rendering-like occlusion removal operator was added",
"in [67] to handle point cloud viewpoint-based self-occlusions and drive supervi-",
"sion solely from depth images. Given that only the pose needs to be optimized, a",
"lightweight abstract rendering mechanism based on OpenDR [38l was developed",
"in [50] to compare learned feature representations from meshes and the observed",
"images, and then estimate their poses in the observed scenes. LatentFusion [46]",
"learns a reconstructed feature representation that can be rendered in order to",
"optimize for an object's pose as it is observed within an image, and has been",
"shown to also generalize well to unseen objects. Finally, a rendered silhouette",
"error is backpropagated in [45] to refine the CNN estimated pose.However, it",
"was only applied in simplistic CAD renderings and was not trained in an end-to-",
"end fashion. This was also the case for [26] which did not use monocular input",
"and only learned pose estimation using simplistic images. Instead, we leverage",
"end-to-end differentiable rendering using complex imagery for boosting 6DOF",
"pose estimation.",
"3",
"Dataset",
"Our goal is to create a realistic UAVA dataset for applications in which users and",
"drones are considered cooperative agents. To achieve this, we need to overcome a",
"set of challenges associated with the photorealism of the visual data, the plausi-",
"bility of the drone-to-user spatial relation, the variety/plurality of the generated",
"views, and the drone's trajectories. Contrary to similarly oriented datasets like",
"URSO [5]] and Blackbird [3] we are not confined to only a single ego- [3] or",
"exocentric view [51], and thus, we need to consider both viewpoints and synthe-",
"size temporally and spatially aligned data. While both [51] and [3] render purely",
"computer-generated scenes, we rely on a recent 3D scanned dataset, Matter-",
"port3D [9] that offers us scene photorealism.",
"The Matterport3D dataset was created after capturing multiple RGB-D",
"panoramas within large buildings. As the poses of each panorama are available,",
"we can leverage them as stationary points (i.e. anchors) to drive our UAVA",
"oriented data generation. Aiming to create realistic drone flight trajectories and",
"simultaneously position user viewpoints in a plausible way, we employ a gamified",
"approach. More specifically, we designed and developed a Unity3D game where",
"collectible cube “coins\" are placed at each of the known panorama positions",
"(anchors) and use AirSim's drone flight simulator to let players navigate a drone",
"within each building to collect the pre-placed “coins\" using a game controller."
] |
{
"arxiv_id": "2008.08823",
"language": "en",
"page_idx": 4
}
|
||
2305.17369-en-0
|
2305.17369-en-0.jpg
|
$\textbf{Modularized Zero- shot VQA with Pre- trained Models}$
Rui Cao and Jing Jiang
School of Computing and Information Systems
Singapore Management University
ruicao.2020@phdcs.smu.edu.sg, jingjiang@smu.edu.sg
## $\mathbf{Abstract}$
Large-scale pre-trained models (PTMs) show great zero-shot capabilities. In this paper, we study how to leverage them for zero-shot visual question answering (VQA).Our approach is motivated by a few observations. First, VQA questions often require multiple steps of reasoning, which is still a capability that most PTMs lack. Second, different steps in VQA reasoning chains require different skills such as object detection and relational reasoning, but a single PTM may not possess all these skills. Third, recent work on zero-shot VQA does not explicitly consider multi-step reasoning chains, which makes them less interpretable compared with a decomposition-based approach. We propose a modularized zero-shot network that explicitly decomposes questions into sub reasoning steps and is highly interpretable. We convert sub reasoning tasks to acceptable objectives of PTMs and assign tasks to proper PTMs without any adaptation. Our experiments on two VQA benchmarks under the zero-shot setting demonstrate the effectiveness of our method and better interpretability compared with several baselines.
## 1 Introduction
Visual Question Answering (VQA), the task of answering textual queries based on information contained in an image, is a multimodal task that requires comprehension and reasoning of both visual and textual content (Agrawal et al.,2017; Hudson and Manning, 2019). Most previous work on VQA either trains VQA models from scratch (e.g., Fukui et al. (2016); Anderson et al. (2018)) or fine-tunes pre-trained vision-language models for VQA (e.g., Li et al. (2019); Lu et al. (2019)). Thus, they rely heavily on labeled VQA data, which are expensive to obtain. VQA models based on supervised learning are also hard to generalize to new domains or new datasets (Xu et al., 2020; Chao et al., 2018; Zhang et al., 2021).
Recently, large-scale pre-trained models (PTMs) have demonstrated strong transferability to different downstream tasks under zero-shot settings, i.e., without any training data for the downstream tasks (Brown et al., 2020; Radford et al., 2021). With increased pre-training data size, these models show strong zero-shot performance on various down-stream tasks, such as image classification and face detection with the CLIP model (Radford et al., 2021) and sentiment analysis and commonsense question answering with the GPT-3 model (Brown et al., 2020). However, few studies have focused on zero-shot VQA from pre-trained models.
Despite the power of these PTMs, it is not straightforward to directly apply them to VQA under zero-shot settings, because they are not pretrained with the same objective as VQA. Some recent work converts images to tokens that pretrained language models can understand so that VQA can be converted to text-based QA (Yang et al., 2022b; Tiong et al., 2022; Tsimpoukelli et al., 2021; Jin et al., 2022; Dai et al., 2022). However, this approach requires either a strong pre-trained image captioning model that can capture sufficient visual details or auxiliary training to obtain such a captioning model. Some other work converts VQA into a multimodal matching problem so that pretrained vision-language models (PT-VLMs) such as CLIP can be used (Song et al., 2022; Shen et al., 2022). However, complex VQA questions such as those found in the GQA dataset (Hudson and Manning, 2019) often require spatial reasoning and/or multi-step reasoning, which PT-VLMs may not be strong at (Subramanian et al., 2022; Thrush et al., 2022).
VQA questions can be complicated and often require different reasoning steps such as object detection and spatial reasoning, as the example question in Figure 1 illustrates. Previously, people proposed Neural Module Networks (Andreas et al., 2016; Hu et al., 2017), which are modularized net-
|
[
[
0.2216796875,
0.0926513671875,
0.77490234375,
0.0926513671875,
0.77490234375,
0.10687255859375,
0.2216796875,
0.10687255859375
],
[
0.397216796875,
0.1383056640625,
0.60791015625,
0.1383056640625,
0.60791015625,
0.152587890625,
0.397216796875,
0.152587890625
],
[
0.30908203125,
0.15380859375,
0.69287109375,
0.155029296875,
0.69287109375,
0.169189453125,
0.30908203125,
0.168701171875
],
[
0.359375,
0.171630859375,
0.64208984375,
0.171630859375,
0.64208984375,
0.1859130859375,
0.359375,
0.1859130859375
],
[
0.252685546875,
0.18701171875,
0.75,
0.188232421875,
0.75,
0.203125,
0.252685546875,
0.201904296875
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.52880859375,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.52880859375,
0.268310546875
],
[
0.51123046875,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.1435546875,
0.281982421875,
0.45751953125,
0.28076171875,
0.45751953125,
0.295654296875,
0.1435546875,
0.29638671875
],
[
0.51123046875,
0.28515625,
0.88330078125,
0.2861328125,
0.88330078125,
0.300537109375,
0.51123046875,
0.2998046875
],
[
0.1435546875,
0.295654296875,
0.459228515625,
0.29638671875,
0.459228515625,
0.311279296875,
0.1435546875,
0.31005859375
],
[
0.51123046875,
0.303955078125,
0.880859375,
0.303955078125,
0.880859375,
0.314208984375,
0.51123046875,
0.314208984375
],
[
0.0235137939453125,
0.31005859375,
0.060455322265625,
0.31005859375,
0.060455322265625,
0.43408203125,
0.0235137939453125,
0.43408203125
],
[
0.145263671875,
0.311279296875,
0.45751953125,
0.311279296875,
0.45751953125,
0.32470703125,
0.145263671875,
0.32470703125
],
[
0.509765625,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.509765625,
0.33203125
],
[
0.145263671875,
0.325927734375,
0.459228515625,
0.325927734375,
0.459228515625,
0.339111328125,
0.145263671875,
0.339111328125
],
[
0.51123046875,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51123046875,
0.3486328125
],
[
0.14697265625,
0.341552734375,
0.456787109375,
0.341552734375,
0.456787109375,
0.3515625,
0.14697265625,
0.3515625
],
[
0.145263671875,
0.352783203125,
0.45751953125,
0.352783203125,
0.45751953125,
0.366455078125,
0.145263671875,
0.366455078125
],
[
0.51318359375,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.14697265625,
0.3681640625,
0.45751953125,
0.3681640625,
0.45751953125,
0.3818359375,
0.14697265625,
0.3818359375
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.1435546875,
0.3818359375,
0.45751953125,
0.383056640625,
0.45751953125,
0.396728515625,
0.1435546875,
0.3955078125
],
[
0.51123046875,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.14697265625,
0.39794921875,
0.45751953125,
0.39794921875,
0.45751953125,
0.407958984375,
0.14697265625,
0.407958984375
],
[
0.51123046875,
0.39794921875,
0.8818359375,
0.3984375,
0.8818359375,
0.413330078125,
0.51123046875,
0.412109375
],
[
0.14697265625,
0.410400390625,
0.45751953125,
0.410400390625,
0.45751953125,
0.42333984375,
0.14697265625,
0.42333984375
],
[
0.509765625,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.145263671875,
0.424560546875,
0.459228515625,
0.424560546875,
0.459228515625,
0.438232421875,
0.145263671875,
0.438232421875
],
[
0.51318359375,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.0218353271484375,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.1435546875,
0.439453125,
0.456787109375,
0.439453125,
0.456787109375,
0.453125,
0.1435546875,
0.453125
],
[
0.145263671875,
0.453125,
0.45751953125,
0.453125,
0.45751953125,
0.466064453125,
0.145263671875,
0.466064453125
],
[
0.51123046875,
0.44775390625,
0.8369140625,
0.44775390625,
0.8369140625,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.145263671875,
0.46728515625,
0.459228515625,
0.46728515625,
0.459228515625,
0.48095703125,
0.145263671875,
0.48095703125
],
[
0.52880859375,
0.466064453125,
0.8818359375,
0.466064453125,
0.8818359375,
0.479736328125,
0.52880859375,
0.479736328125
],
[
0.145263671875,
0.482177734375,
0.456787109375,
0.482177734375,
0.456787109375,
0.495361328125,
0.145263671875,
0.495361328125
],
[
0.51123046875,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.51123046875,
0.496337890625
],
[
0.14697265625,
0.495361328125,
0.45751953125,
0.496337890625,
0.45751953125,
0.50830078125,
0.14697265625,
0.50732421875
],
[
0.51123046875,
0.49951171875,
0.880859375,
0.49951171875,
0.880859375,
0.51318359375,
0.51123046875,
0.51318359375
],
[
0.145263671875,
0.51025390625,
0.456787109375,
0.51025390625,
0.456787109375,
0.52392578125,
0.145263671875,
0.52392578125
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.51123046875,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.51123046875,
0.529296875
],
[
0.145263671875,
0.52490234375,
0.459228515625,
0.52490234375,
0.459228515625,
0.5380859375,
0.145263671875,
0.5380859375
],
[
0.509765625,
0.52978515625,
0.8818359375,
0.53076171875,
0.8818359375,
0.5458984375,
0.509765625,
0.54443359375
],
[
0.145263671875,
0.5380859375,
0.459228515625,
0.5380859375,
0.459228515625,
0.552734375,
0.145263671875,
0.552734375
],
[
0.51123046875,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.51123046875,
0.56103515625
],
[
0.14697265625,
0.55419921875,
0.45751953125,
0.55419921875,
0.45751953125,
0.56396484375,
0.14697265625,
0.56396484375
],
[
0.51123046875,
0.5625,
0.880859375,
0.5625,
0.880859375,
0.576171875,
0.51123046875,
0.576171875
],
[
0.14697265625,
0.568359375,
0.456787109375,
0.568359375,
0.456787109375,
0.5791015625,
0.14697265625,
0.5791015625
],
[
0.145263671875,
0.58056640625,
0.45751953125,
0.58056640625,
0.45751953125,
0.59423828125,
0.145263671875,
0.59423828125
],
[
0.51123046875,
0.57763671875,
0.8818359375,
0.57763671875,
0.8818359375,
0.5927734375,
0.51123046875,
0.5927734375
],
[
0.14697265625,
0.5966796875,
0.45751953125,
0.5966796875,
0.45751953125,
0.60693359375,
0.14697265625,
0.60693359375
],
[
0.51318359375,
0.595703125,
0.88330078125,
0.595703125,
0.88330078125,
0.609375,
0.51318359375,
0.609375
],
[
0.145263671875,
0.60791015625,
0.459228515625,
0.60791015625,
0.459228515625,
0.62158203125,
0.145263671875,
0.62158203125
],
[
0.51318359375,
0.6123046875,
0.8818359375,
0.6123046875,
0.8818359375,
0.6259765625,
0.51318359375,
0.6259765625
],
[
0.145263671875,
0.62451171875,
0.2435302734375,
0.62451171875,
0.2435302734375,
0.6357421875,
0.145263671875,
0.6357421875
],
[
0.51123046875,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.51318359375,
0.6455078125,
0.880859375,
0.6455078125,
0.880859375,
0.65625,
0.51318359375,
0.65625
],
[
0.118408203125,
0.65478515625,
0.1318359375,
0.65478515625,
0.1318359375,
0.6630859375,
0.118408203125,
0.6630859375
],
[
0.1435546875,
0.65625,
0.1485595703125,
0.65625,
0.1485595703125,
0.662109375,
0.1435546875,
0.662109375
],
[
0.145263671875,
0.65380859375,
0.26025390625,
0.65380859375,
0.26025390625,
0.66650390625,
0.145263671875,
0.66650390625
],
[
0.51318359375,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51318359375,
0.67333984375
],
[
0.509765625,
0.67578125,
0.8818359375,
0.67578125,
0.8818359375,
0.6904296875,
0.509765625,
0.6904296875
],
[
0.133544921875,
0.65673828125,
0.1427001953125,
0.65673828125,
0.1427001953125,
0.662109375,
0.133544921875,
0.662109375
],
[
0.1175537109375,
0.6787109375,
0.489501953125,
0.6787109375,
0.489501953125,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.1158447265625,
0.6953125,
0.48779296875,
0.6953125,
0.48779296875,
0.708984375,
0.1158447265625,
0.708984375
],
[
0.509765625,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.509765625,
0.705078125
],
[
0.1175537109375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.509765625,
0.7080078125,
0.88330078125,
0.7080078125,
0.88330078125,
0.7216796875,
0.509765625,
0.7216796875
],
[
0.1175537109375,
0.72802734375,
0.490234375,
0.72802734375,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.51123046875,
0.72265625,
0.8818359375,
0.7236328125,
0.8818359375,
0.73828125,
0.51123046875,
0.73681640625
],
[
0.1158447265625,
0.7421875,
0.48779296875,
0.7421875,
0.48779296875,
0.755859375,
0.1158447265625,
0.755859375
],
[
0.51318359375,
0.74072265625,
0.88330078125,
0.74072265625,
0.88330078125,
0.75439453125,
0.51318359375,
0.75439453125
],
[
0.1175537109375,
0.7587890625,
0.48779296875,
0.7587890625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.75732421875,
0.88330078125,
0.75732421875,
0.88330078125,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.11419677734375,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.11419677734375,
0.7880859375
],
[
0.509765625,
0.7724609375,
0.8818359375,
0.7724609375,
0.8818359375,
0.78759765625,
0.509765625,
0.78759765625
],
[
0.1175537109375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.509765625,
0.7880859375,
0.884765625,
0.7880859375,
0.884765625,
0.802734375,
0.509765625,
0.802734375
],
[
0.509765625,
0.8017578125,
0.5615234375,
0.80419921875,
0.56103515625,
0.8193359375,
0.5087890625,
0.81787109375
],
[
0.1158447265625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.1158447265625,
0.822265625,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.52880859375,
0.82373046875,
0.880859375,
0.82373046875,
0.880859375,
0.83740234375,
0.52880859375,
0.83740234375
],
[
0.118408203125,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.11419677734375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.1175537109375,
0.8896484375,
0.490234375,
0.8896484375,
0.490234375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.88671875,
0.88720703125,
0.88671875,
0.90185546875,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90283203125,
0.444091796875,
0.90380859375,
0.444091796875,
0.91845703125,
0.1158447265625,
0.91748046875
],
[
0.51318359375,
0.90478515625,
0.884765625,
0.90478515625,
0.884765625,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Modularized Zero-shot VOA with Pre-trained Models",
"Rui Cao and Jing Jiang",
"School of Computing and Information Systems",
"Singapore Management University",
"ruicao.2020@phdcs.smu.edu.sg,jingjiang@smu.edu.sg",
"Abstract",
"Recently, large-scale pre-trained models (PTMs)",
"have demonstrated strong transferability to dif-",
"Large-scale pre-trained models (PTMs) show",
"ferent downstream tasks under zero-shot settings,",
"great zero-shot capabilities. In this paper, we",
"i.e.,without any training datafor the downstream",
"",
"study how to leverage them for zero-shot visual",
"tasks (Brown et al., 2020; Radford et al., 2021)",
"question answering (VQA). Our approach is",
"With increased pre-training data size, these mod-",
"motivated by afew observations.First,VQA",
"questions often require multiple steps of reason",
"els show strong zero-shot performance on various",
"ing, which is still a capability that most PTMs",
"down-stream tasks, such as image classification and",
"lack. Second, different steps in VQA reasoning",
"face detection with the CLIP model (Radford et al..",
"chains require different skills such as object",
"2021) and sentiment analysis and commonsense",
"detection and relational reasoning, but a single",
"question answering with the GPT-3 model (Brown",
"PTM may not possess all these skills. Third, re-",
"et al., 2020). However, few studies have focused",
"[A']",
"cent work on zero-shot VQA does not explicitly",
"consider multi-step reasoning chains, which",
"on zero-shot VQA from pre-trained models.",
"makes them less interpretable compared with",
"Despite the power of these PTMs, it is not",
"a decomposition-based approach. We propose",
"straightforward to directly apply them to VQA un-",
"a modularized zero-shot network that explic",
"der zero-shot settings, because they are not pre-",
"itly decomposes questions into sub reasoning",
"096115002110",
"trained with the same objective as VQA. Some",
"steps and is highly interpretable. We convert",
"recent work converts images to tokens that pre-",
"sub reasoning tasks to acceptable objectives of",
"trained language models can understand so that",
"PTMs and assigntasks toproperPTMs with-",
"VQA can be converted to text-based QA (Yang",
"outany adaptation.Our experiments on two",
"VQA benchmarks under the zero-shot setting",
"et al., 2022b; Tiong et al., 2022; Tsimpoukelli et al.",
"demonstrate the effectiveness of our method",
"2021; Jin et al., 2022; Dai et al., 2022). However,",
"and better interpretability compared with sev-",
"this approach requires either a strong pre-trained",
"eral baselines.",
"image captioning model that can capture sufficient",
"visual details or auxiliary trainingto obtain such a",
"1",
"",
"Introduction",
"captioning model.Some other work converts VQA",
"into a multimodal matching problem so that pre",
"",
"Visual Question Answering (VQA), the task of",
"answering textual queries based on information",
"trained vision-language models (PT-VLMs) such",
"contained in an image, is a multimodal task that",
"as CLIP can be used (Song et al., 2022; Shen et al.,",
"requires comprehension and reasoning of both vi-",
"2022). However, complex VQA questions such as",
"sual and textual content (Agrawal et al., 2017:",
"those found in the GQA dataset (Hudson and Man-",
"Hudson and Manning, 2019). Most previous",
"ning, 2019) often require spatial reasoning and/or",
"work on VQA either trains VOA models from",
"multi-step reasoning, which PT-VLMs may not be",
"scratch (e.g., Fukui et al. (2016); Anderson et al.",
" strong at (Subramanian et al., 2022; Thrush et al.,",
"2022).",
"(2018)) or fine-tunes pre-trained vision-language",
"models for VQA (e.g., Li et al. (2019); Lu et al.",
"VQA questions can be complicated and often",
"(2019)). Thus, they rely heavily on labeled VQA",
"require different reasoning steps such as object",
"data, which are expensive to obtain. VQA models",
"detection and spatial reasoning, as the example",
" based on supervised learning are also hard to gen-",
"question in Figure 1 illustrates. Previously, people",
"eralize to new domains or new datasets (Xu et al.,",
"proposed Neural Module Networks (Andreas et al.,",
"2020; Chao et al., 2018; Zhang et al., 2021).",
"2016; Hu et al., 2017), which are modularized net-"
] |
{
"arxiv_id": "2305.17369",
"language": "en",
"page_idx": 0
}
|
||
2305.17369-en-1
|
2305.17369-en-1.jpg
|
<FigureHere>
Figure 1: An overview of our proposed method. Instead of training modules in NMN, we propose a modularized
zero-shot VQA method leveraging pre-trained models to perform different reasoning tasks.
The contributions of our work can be summarized as follows:(1) We propose a novel modularized zero-shot VQA method that utilizes different pre-trained models for different reasoning steps; (2) We design rules to map different VQA reasoning steps to suitable PTMs so that we can leverage these PTMs without any adaptation; 3) Experiment results show the effectiveness of the proposed method, especially when questions consist of multiple steps of reasoning.
### 2 Background
works where each pre-defined module performs a specific reasoning task. These pre-defined modules are trained end-to-end from labeled VQA data. Motivated by the idea of modularization, in this paper, we propose a modularized zero-shot network for VQA (Mod-Zero-VQA) by decomposing questions into sub-tasks and assigning appropriate sub-tasks to PTMs without any adaptation. Given a question, we first parse the question into basic reasoning steps explicitly. These reasoning steps will then be reconfigured and mapped to different PTMs based on a set of rules we define. Specifically, we consider the following PTMs: $OWL$ (Minderer et al., 2022) as the object detector, $MDETR$ (Kamath et al., 2021) for reference expression localization (including several skills such as relational and spatial reasoning) and $CLIP$ (Radford et al., 2021) as the answer generator for open-ended questions. Considering the limited capabilities of current pretrained vision-language models in spatial relation understanding (Subramanian et al., 2022), we also define simple and general heuristics to aid spatial reasoning. Note that only when we decompose questions and reasoning chains step by step can we insert human heuristics for spatial reasoning, because we have the intermediate outputs such as objects' bounding boxes from previous steps.
We evaluate the proposed method on the GQA
Task Definition. Given an image $I$ and a question $Q$,a VQA system is expected to return an answer $a.$ Traditional fully supervised VQA relies on a training set consisting of (image, question, answer) triplets. For zero-shot VQA, no such training data is given. However, in this paper we assume that we can use pre-trained models (PTMs) to help us with zero-shot VQA.
Existing Zero-shot VQA Methods. Work on zero-shot VQA is very limited. We can organize existing work into the following categories. One line of work leverages the question answering capability in pre-trained language model (LMs). Some of them adopt prefix language modeling with weakly-supervised data other than VQA data (i.e., image-text pairs) to convert visual information into discrete tokens (prefix) that LMs can understand. Frozen (Tsimpoukelli et al., 2021), VLKD (Dai et al., 2022) and FewVLM (Jin et al., 2022) fall under this category. Some directly convert VQA images into textual descriptions so that the task of VQA changes to text-based QA and LMs can be applied. Methods in this category include PICa (Yang et al., 2022b) and PnP-VQA (Tiong et al., 2022). Recent work (Song et al., 2022; Shen et al., 2022)
dataset (Hudson and Manning, 2019) where questions are compositional and require multi-step reasoning. The experiment result shows that the proposed model surpasses the baselines significantly on GQA, with near $13\%$ of relative improvement over the strongest baseline (from 41.9 to 47.3). The results confirm the benefit of modualarization when using PTMs for zero-shot VQA. In addition, our method is interpretable because of the explicit reasoning steps generated.
|
[
[
0.1259765625,
0.08074951171875,
0.2216796875,
0.08074951171875,
0.2216796875,
0.0914306640625,
0.1259765625,
0.0914306640625
],
[
0.340087890625,
0.0831298828125,
0.86962890625,
0.0831298828125,
0.86962890625,
0.0968017578125,
0.340087890625,
0.0968017578125
],
[
0.124267578125,
0.09027099609375,
0.2333984375,
0.0914306640625,
0.2333984375,
0.1015625,
0.124267578125,
0.1009521484375
],
[
0.268798828125,
0.08905029296875,
0.30322265625,
0.08905029296875,
0.30322265625,
0.09320068359375,
0.268798828125,
0.09320068359375
],
[
0.124267578125,
0.09979248046875,
0.231689453125,
0.1009521484375,
0.231689453125,
0.11224365234375,
0.124267578125,
0.11102294921875
],
[
0.74072265625,
0.1009521484375,
0.85205078125,
0.1009521484375,
0.85205078125,
0.11224365234375,
0.74072265625,
0.11224365234375
],
[
0.45166015625,
0.10687255859375,
0.5517578125,
0.10687255859375,
0.5517578125,
0.1175537109375,
0.45166015625,
0.1175537109375
],
[
0.1275634765625,
0.1134033203125,
0.1695556640625,
0.1134033203125,
0.1695556640625,
0.12054443359375,
0.1275634765625,
0.12054443359375
],
[
0.71533203125,
0.1134033203125,
0.8076171875,
0.1134033203125,
0.8076171875,
0.12353515625,
0.71533203125,
0.12353515625
],
[
0.71533203125,
0.1270751953125,
0.7421875,
0.1270751953125,
0.7421875,
0.1353759765625,
0.71533203125,
0.1353759765625
],
[
0.755859375,
0.1270751953125,
0.84619140625,
0.1270751953125,
0.84619140625,
0.13720703125,
0.755859375,
0.13720703125
],
[
0.45751953125,
0.1341552734375,
0.5380859375,
0.1341552734375,
0.5380859375,
0.144287109375,
0.45751953125,
0.144287109375
],
[
0.5517578125,
0.1341552734375,
0.6826171875,
0.1341552734375,
0.6826171875,
0.144287109375,
0.5517578125,
0.144287109375
],
[
0.7138671875,
0.13720703125,
0.869140625,
0.13720703125,
0.869140625,
0.1478271484375,
0.7138671875,
0.1478271484375
],
[
0.58447265625,
0.15087890625,
0.60107421875,
0.15087890625,
0.60107421875,
0.15380859375,
0.58447265625,
0.15380859375
],
[
0.276123046875,
0.1318359375,
0.43994140625,
0.133056640625,
0.43994140625,
0.144287109375,
0.276123046875,
0.1436767578125
],
[
0.7138671875,
0.1519775390625,
0.73876953125,
0.1519775390625,
0.73876953125,
0.1602783203125,
0.7138671875,
0.1602783203125
],
[
0.75830078125,
0.15087890625,
0.8310546875,
0.15087890625,
0.8310546875,
0.160888671875,
0.75830078125,
0.160888671875
],
[
0.5576171875,
0.163330078125,
0.603515625,
0.163330078125,
0.603515625,
0.174560546875,
0.5576171875,
0.174560546875
],
[
0.6484375,
0.1644287109375,
0.68408203125,
0.1644287109375,
0.68408203125,
0.168701171875,
0.6484375,
0.168701171875
],
[
0.73876953125,
0.166259765625,
0.84228515625,
0.166259765625,
0.84228515625,
0.177001953125,
0.73876953125,
0.177001953125
],
[
0.56103515625,
0.174560546875,
0.60205078125,
0.174560546875,
0.60205078125,
0.1817626953125,
0.56103515625,
0.1817626953125
],
[
0.6455078125,
0.179931640625,
0.66162109375,
0.179931640625,
0.66162109375,
0.18701171875,
0.6455078125,
0.18701171875
],
[
0.66015625,
0.1817626953125,
0.68603515625,
0.1817626953125,
0.68603515625,
0.1859130859375,
0.66015625,
0.1859130859375
],
[
0.748046875,
0.1817626953125,
0.77490234375,
0.1817626953125,
0.77490234375,
0.190673828125,
0.748046875,
0.190673828125
],
[
0.73779296875,
0.1923828125,
0.78857421875,
0.1923828125,
0.78857421875,
0.203125,
0.73779296875,
0.203125
],
[
0.55908203125,
0.182861328125,
0.623046875,
0.182861328125,
0.623046875,
0.19482421875,
0.55908203125,
0.19482421875
],
[
0.56103515625,
0.19482421875,
0.60205078125,
0.19482421875,
0.60205078125,
0.204833984375,
0.56103515625,
0.204833984375
],
[
0.6455078125,
0.201904296875,
0.68408203125,
0.201904296875,
0.68408203125,
0.208984375,
0.6455078125,
0.208984375
],
[
0.73779296875,
0.207275390625,
0.78271484375,
0.207275390625,
0.78271484375,
0.21728515625,
0.73779296875,
0.21728515625
],
[
0.80029296875,
0.203125,
0.85205078125,
0.203125,
0.85205078125,
0.21435546875,
0.80029296875,
0.21435546875
],
[
0.75390625,
0.223876953125,
0.767578125,
0.223876953125,
0.767578125,
0.226806640625,
0.75390625,
0.226806640625
],
[
0.1175537109375,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.1175537109375,
0.262451171875,
0.72705078125,
0.262451171875,
0.72705078125,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.1175537109375,
0.302734375,
0.489501953125,
0.302734375,
0.489501953125,
0.31640625,
0.1175537109375,
0.31640625
],
[
0.52880859375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.52880859375,
0.31640625
],
[
0.1158447265625,
0.318359375,
0.509765625,
0.317626953125,
0.509765625,
0.33203125,
0.1158447265625,
0.333251953125
],
[
0.50537109375,
0.317626953125,
0.8818359375,
0.318359375,
0.8818359375,
0.333251953125,
0.50537109375,
0.33203125
],
[
0.1175537109375,
0.3349609375,
0.490234375,
0.3349609375,
0.490234375,
0.3486328125,
0.1175537109375,
0.3486328125
],
[
0.509765625,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.509765625,
0.3486328125
],
[
0.1175537109375,
0.350830078125,
0.489501953125,
0.350830078125,
0.489501953125,
0.364013671875,
0.1175537109375,
0.364013671875
],
[
0.51123046875,
0.349853515625,
0.880859375,
0.350830078125,
0.880859375,
0.365234375,
0.51123046875,
0.364013671875
],
[
0.1175537109375,
0.36767578125,
0.490234375,
0.36767578125,
0.490234375,
0.38134765625,
0.1175537109375,
0.38134765625
],
[
0.509765625,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.509765625,
0.38134765625
],
[
0.1175537109375,
0.3818359375,
0.48779296875,
0.3818359375,
0.48779296875,
0.3955078125,
0.1175537109375,
0.3955078125
],
[
0.5087890625,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.5087890625,
0.3955078125
],
[
0.1175537109375,
0.399658203125,
0.489501953125,
0.399658203125,
0.489501953125,
0.413330078125,
0.1175537109375,
0.413330078125
],
[
0.5087890625,
0.3984375,
0.880859375,
0.3984375,
0.880859375,
0.413330078125,
0.5087890625,
0.413330078125
],
[
0.1175537109375,
0.4150390625,
0.490234375,
0.4150390625,
0.490234375,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.51318359375,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.1175537109375,
0.431640625,
0.491943359375,
0.431640625,
0.491943359375,
0.4453125,
0.1175537109375,
0.4453125
],
[
0.509765625,
0.431640625,
0.88330078125,
0.431640625,
0.88330078125,
0.4453125,
0.509765625,
0.4453125
],
[
0.11419677734375,
0.44775390625,
0.48779296875,
0.446533203125,
0.48779296875,
0.46142578125,
0.11419677734375,
0.4619140625
],
[
0.50732421875,
0.448974609375,
0.76318359375,
0.448974609375,
0.76318359375,
0.4619140625,
0.50732421875,
0.4619140625
],
[
0.11419677734375,
0.4619140625,
0.48779296875,
0.463134765625,
0.48779296875,
0.47802734375,
0.11419677734375,
0.476806640625
],
[
0.1158447265625,
0.479736328125,
0.489501953125,
0.479736328125,
0.489501953125,
0.493408203125,
0.1158447265625,
0.493408203125
],
[
0.51318359375,
0.47802734375,
0.6513671875,
0.47802734375,
0.6513671875,
0.4921875,
0.51318359375,
0.4921875
],
[
0.1158447265625,
0.495361328125,
0.48779296875,
0.495361328125,
0.48779296875,
0.5087890625,
0.1158447265625,
0.5087890625
],
[
0.1175537109375,
0.51123046875,
0.490234375,
0.51123046875,
0.490234375,
0.52490234375,
0.1175537109375,
0.52490234375
],
[
0.51123046875,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.51953125,
0.51123046875,
0.51953125
],
[
0.5087890625,
0.521484375,
0.8818359375,
0.5224609375,
0.8818359375,
0.53759765625,
0.5087890625,
0.5361328125
],
[
0.1158447265625,
0.52783203125,
0.489501953125,
0.52783203125,
0.489501953125,
0.54150390625,
0.1158447265625,
0.54150390625
],
[
0.1158447265625,
0.5419921875,
0.486083984375,
0.5419921875,
0.486083984375,
0.5556640625,
0.1158447265625,
0.5556640625
],
[
0.5087890625,
0.5380859375,
0.8818359375,
0.53759765625,
0.8818359375,
0.5517578125,
0.5087890625,
0.552734375
],
[
0.1158447265625,
0.55859375,
0.486083984375,
0.55859375,
0.486083984375,
0.572265625,
0.1158447265625,
0.572265625
],
[
0.5087890625,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.5087890625,
0.56787109375
],
[
0.5087890625,
0.5693359375,
0.8818359375,
0.57080078125,
0.8818359375,
0.5849609375,
0.5087890625,
0.58447265625
],
[
0.1175537109375,
0.576171875,
0.48779296875,
0.576171875,
0.48779296875,
0.5888671875,
0.1175537109375,
0.5888671875
],
[
0.51123046875,
0.5849609375,
0.8818359375,
0.5859375,
0.8818359375,
0.60107421875,
0.51123046875,
0.599609375
],
[
0.1158447265625,
0.59130859375,
0.489501953125,
0.5927734375,
0.489501953125,
0.60693359375,
0.1158447265625,
0.60546875
],
[
0.51318359375,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1158447265625,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62255859375,
0.1158447265625,
0.62255859375
],
[
0.51123046875,
0.61767578125,
0.6884765625,
0.61767578125,
0.6884765625,
0.63134765625,
0.51123046875,
0.63134765625
],
[
0.11419677734375,
0.62353515625,
0.48779296875,
0.62353515625,
0.48779296875,
0.63720703125,
0.11419677734375,
0.63720703125
],
[
0.1158447265625,
0.64013671875,
0.48779296875,
0.6396484375,
0.48779296875,
0.65380859375,
0.1158447265625,
0.65478515625
],
[
0.51318359375,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.66015625,
0.51318359375,
0.66015625
],
[
0.1158447265625,
0.65673828125,
0.48779296875,
0.65673828125,
0.48779296875,
0.67041015625,
0.1158447265625,
0.67041015625
],
[
0.5087890625,
0.662109375,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.5087890625,
0.6767578125
],
[
0.1175537109375,
0.67333984375,
0.48779296875,
0.67333984375,
0.48779296875,
0.68701171875,
0.1175537109375,
0.68701171875
],
[
0.51123046875,
0.6787109375,
0.88330078125,
0.6787109375,
0.88330078125,
0.693359375,
0.51123046875,
0.693359375
],
[
0.11419677734375,
0.68701171875,
0.489501953125,
0.689453125,
0.489501953125,
0.70361328125,
0.11419677734375,
0.70166015625
],
[
0.51123046875,
0.69482421875,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.51123046875,
0.708984375
],
[
0.11419677734375,
0.70361328125,
0.48779296875,
0.705078125,
0.48779296875,
0.7197265625,
0.11419677734375,
0.71875
],
[
0.51123046875,
0.71142578125,
0.8818359375,
0.71044921875,
0.8818359375,
0.724609375,
0.51123046875,
0.7255859375
],
[
0.1175537109375,
0.7216796875,
0.45166015625,
0.7216796875,
0.45166015625,
0.7353515625,
0.1175537109375,
0.7353515625
],
[
0.51318359375,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51318359375,
0.7412109375
],
[
0.1368408203125,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1368408203125,
0.75732421875
],
[
0.509765625,
0.7421875,
0.88330078125,
0.74365234375,
0.88330078125,
0.75830078125,
0.509765625,
0.75732421875
],
[
0.1175537109375,
0.7587890625,
0.490234375,
0.7587890625,
0.490234375,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.1175537109375,
0.775390625,
0.490234375,
0.775390625,
0.490234375,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.79345703125,
0.490234375,
0.79345703125,
0.490234375,
0.80419921875,
0.1175537109375,
0.80419921875
],
[
0.51318359375,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51318359375,
0.8046875
],
[
0.1158447265625,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.509765625,
0.82080078125
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.509765625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.11419677734375,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.11419677734375,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.5087890625,
0.85498046875,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.5087890625,
0.8681640625
],
[
0.11419677734375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.49365234375,
0.88818359375,
0.49365234375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.51318359375,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.51318359375,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.28955078125,
0.90478515625,
0.28955078125,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.51318359375,
0.90380859375,
0.8818359375,
0.90380859375,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Question:Theman",
"Find(controller) Relocate([1],holding) Filter([2], man)-Find(woman) -Choose(3], [4],[left, right)",
"holding controlleris to",
"parser",
"the leftor right of the",
"Pre-trained Models",
"Layout Mapping",
"woman?",
"Model Function",
"OWL",
"ObjectDetection",
"owL(woman)",
"CLIP([1], [2], [left, right]",
"MDETRModulatedDetection",
"",
"MDETR(manholdingcontroller)",
"CLIP",
"MMMatching",
"Theman",
"uaht.o",
"Traditional NMNs",
"istothe",
"Left",
"00.74",
"Find",
"Relocate",
"[ANS] of the",
"woman",
"Ans.Vog",
"Choose",
"VQAData",
"E",
"Figure I: An overview of our proposed method. Instead of training modules in NMN, we propose a modularized",
"zero-shot VQA method leveraging pre-trained models to perform different reasoning tasks.",
"works where each pre-defined module performs a",
"The contributions of our work can be summa-",
"specific reasoning task.Thesepre-defined mod-",
"rized as follows: (1)We propose a novel mod-",
"ules are trained end-to-end from labeled VQA data.",
"ularized zero-shot VQA method that utilizes dif",
"Motivated by the idea of modularization, in this",
"ferent pre-trained models for different reasoning",
"paper, we propose a modularized zero-shot net-",
"steps; (2) We design rules to map different VQA",
"work for VOA (Mod-Zero-VQA) by decomposing",
" reasoning steps to suitable PTMs so that we can",
"questions into sub-tasks and assigning appropriate",
"leverage these PTMs without any adaptation; 3)",
"sub-tasks to PTMs without any adaptation. Given a",
"Experiment results show the effectiveness of the",
"question, we first parse the question into basic rea-",
"proposed method, especially when questions con-",
"soning steps explicitly.These reasoning steps will",
"sist of multiple steps of reasoning",
"then be reconfigured and mapped to different PTMs",
"based on a set of rules we define. Specifically,",
"2Background",
"we consider the following PTMs: OWL (Minderer",
"et al., 2022) as the object detector, MDETR (Ka-",
"Task Definition. Given an image I and a ques-",
"tion Q, a VQA system is expected to return an",
"math et al., 2021) for reference expression localiza-",
"tion (including several skills such as relational and",
"answer a. Traditional fully supervised VQA relies",
"spatial reasoning) and CLIP (Radford et al., 2021)",
" on a training set consisting of (image, question, an-",
"swer) triplets. For zero-shot VQA, no such training",
"as the answer generator for open-ended questions.",
"data is given. However, in this paper we assume",
"Considering the limited capabilities of current pre-",
"that we can use pre-trained models (PTMs) to help",
"trainedvision-languagemodels in spatialrelation",
"us with zero-shot VQA.",
"understanding (Subramanian et al., 2022), we also",
"define simple and general heuristics to aid spatial",
"Existing Zero-shot VQA Methods.Work on",
"reasoning. Note that only when we decompose",
"zero-shot VQA is very limited. We can orga-",
"questions and reasoning chains step by step can",
"nize existing work into the following categories.",
"we insert human heuristics for spatial reasoning,",
"One line of work leverages the question answering",
" because we have the intermediate outputs such as",
"capability in pre-trained language model (LMs)",
"objects' bounding boxes from previous steps.",
"Some of them adopt prefix language modeling with",
"We evaluate the proposed method on the GQA",
"weakly-supervised data other than VQA data (i.e.,",
"dataset (Hudson and Manning, 2019) where ques-",
"image-text pairs) to convert visual information into",
"tions are compositional and require multi-step rea-",
"discrete tokens (prefix) that LMs can understand.",
"soning.Theexperimentresult showsthatthepro-",
"Frozen (Tsimpoukelli et al., 202I), VLKD (Dai",
"posed model surpasses the baselines significantly",
"et al., 2022) and FewVLM (Jin et al., 2022) fall",
"on GQA, with near 13% of relative improvement",
"under this category. Some directly convert VQA",
"over the strongest baseline (from 41.9 to 47.3). The",
"images into textual descriptions so that the task of",
"resultsconfirmthebenefitofmodualarizationwhen",
"VOAchangestotext-basedQAand LMs can beap-",
"using PTMs for zero-shot VQA. In addition, our",
"plied. Methods in this category include PICa (Yang",
"method is interpretable because of the explicit rea-",
"et al., 2022b) and PnP-VQA (Tiong et al., 2022).",
"soning steps generated.",
"Recent work (Song et al., 2022; Shen et al., 2022)"
] |
{
"arxiv_id": "2305.17369",
"language": "en",
"page_idx": 1
}
|
||
2305.17369-en-2
|
2305.17369-en-2.jpg
|
converts VQA to an image-text matching problem and prompts the CLIP model (Radford et al., 2021), a large-scale vision-language model pre-trained on the image-text matching task. The prompts can be either question irrelevant such as $\varrho uesion:$ [Ques]; Answer: IMASK/ (QIP by Shen et al. (2022)) or question-related by converting questions into a masked statement (TAC-P by Song et al. (2022)).
However, a limitation with these methods is that several of them still require training, although the training data is not in the form of VQA. Besides, converting images to captions and leveraging textbased QA may lose important visual details during the caption generation step. The two methods above using CLIP do not address the issue that CLIP model lacks compositional and spatial reasoning abilities, which has been observed in previous work (Subramanian et al., 2022; Thrush et al., 2022).
## 3 Modularized Zero-shot VQA
Our method is motivated by Neural Module Network (NMN) based VQA, which decomposes questions into reasoning steps, where each module in the NMN is pre-defined to perform a specific reasoning task. The idea allows us to select appropriate pre-trained models to handle different reasoning tasks in a question. Specifically, in NMN-based VQA, we first manually define a set of reasoning steps such as object detection and spatial reasoning, each represented by a module. A question is then explicitly decomposed and converted into a layout of modules, which is an executable program showing the reasoning chain to reach the final answer. The top section of Figure 1 shows the layout corresponding to the sample question. To train an NMNbased VQA system, usually a layout generator is separately built first, which either uses hand-crafted rules over dependency parses of questions or is a trained seq2seq model. Then, the parameters of the various VQA modules are learned from VQA training data.
For our work, we do not want to use VQA data for training. But we observe that many modules in NMN-based VQA can be supported by pre-trained models that have already acquired the capabilities needed by these modules. The key component of our method is therefore to map a layout of modules produced by traditional NMN-based VQA to a simplified layout of zero-shot components that can be implemented directly using pre-trained models.
## 3.1 Traditional VQA Modules
There is not any standard set of modules for VQA We largely adopt the design of modules introduced by Hu et al. (2017) with some minor changes. We assume that the image has been pre-processed and $N$ bounding boxes have been detected, each represented as an embedding vector, collectively denoted as $\mathbf{V}=(\mathbf{v}_1,\mathbf{v}_2,\ldots,\mathbf{v}_N).$ An attention map $\alpha$ is defined to be a distribution over the $N$ bounding boxes.
Table 1 lists the most important traditional VQA modules that we will replace with pre-trained models. The full list of modules can be found in Table 7 in the appendices. It is worth explaining that besides taking in V and $\alpha$ as either input or output, many modules also take in the word embeddings of some text description extracted from the question. These text embeddings are arguments to control the behaviors of the modules. For example, the Find module's objective is to locate an object among all the bounding boxes given. The textual input goBJ is therefore the word embedding of the name of the object to be found. Similarly, gaELA ,gaTTR and goueRY are word embeddings for the description of relation (e.g., to the left of), attribute (e.g., red) and aspect to query (e.g., querying name).
<table>
<tbody>
<tr>
<th>$\mathbf{Module}$</th>
<th>Inputs</th>
</tr>
<tr>
<td>Find</td>
<td>$V,g_{\mathrm{OBJ}}$</td>
</tr>
<tr>
<td>Relocate</td>
<td>$\alpha , \mathbf{V} , \mathbf{g} _{\mathrm{RELA}}$</td>
</tr>
<tr>
<td>Filter</td>
<td>$\alpha , \mathbf{V} , \mathbf{g} _{\mathrm{CONDI}}$</td>
</tr>
<tr>
<td>Choose</td>
<td>$\alpha _{1}, \alpha _{2}, \mathbf{V} , \mathbf{g} _{\mathrm{RF}}$ $\mathrm{A}^{1},\mathbf{g}_{\mathrm{REI.A}}$</td>
</tr>
<tr>
<td>Query</td>
<td>$\alpha , \mathbf{V} , \mathbf{g} _{\mathrm{QUERY}}$</td>
</tr>
</tbody>
</table>
Table 1: A subset of the modules in traditional NMN that we replace with pre-trained models. Modules in the first block output an attention map and those in the second block generate an answer.
Traditionally, the parameters of the modules in Table 1 need to be learned from VQA training data. In other words, these modules' underlying capabilities such as object recognition and relational reasoning need to be acquired from VQA data. However, we hypothesize that recently developed pre-trained models may already have some of these capabilities and can therefore directly equip these modules with such capabilities. For example, the Find module is mainly responsible for object recognition, and previously the parameters of Find have to be learned from scratch using VQA data. Now with
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.08489990234375,
0.759765625,
0.08489990234375,
0.759765625,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.1158447265625,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.5087890625,
0.10687255859375,
0.880859375,
0.10687255859375,
0.880859375,
0.12054443359375,
0.5087890625,
0.12054443359375
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.12353515625,
0.8818359375,
0.12353515625,
0.8818359375,
0.13720703125,
0.51123046875,
0.13720703125
],
[
0.1175537109375,
0.1341552734375,
0.48779296875,
0.1341552734375,
0.48779296875,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51318359375,
0.141357421875,
0.880859375,
0.141357421875,
0.880859375,
0.1519775390625,
0.51318359375,
0.1519775390625
],
[
0.1175537109375,
0.15087890625,
0.489501953125,
0.15087890625,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51123046875,
0.1561279296875,
0.88330078125,
0.1561279296875,
0.88330078125,
0.169189453125,
0.51123046875,
0.169189453125
],
[
0.1158447265625,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1158447265625,
0.179931640625
],
[
0.51123046875,
0.171630859375,
0.88330078125,
0.171630859375,
0.88330078125,
0.1859130859375,
0.51123046875,
0.1859130859375
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.509765625,
0.188232421875,
0.884765625,
0.188232421875,
0.884765625,
0.203125,
0.509765625,
0.203125
],
[
0.11419677734375,
0.19775390625,
0.480224609375,
0.196533203125,
0.480224609375,
0.21142578125,
0.11419677734375,
0.2120361328125
],
[
0.509765625,
0.203125,
0.8818359375,
0.2037353515625,
0.8818359375,
0.218505859375,
0.509765625,
0.21728515625
],
[
0.1351318359375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.1351318359375,
0.2286376953125
],
[
0.51123046875,
0.2203369140625,
0.88330078125,
0.2203369140625,
0.88330078125,
0.2340087890625,
0.51123046875,
0.2340087890625
],
[
0.1158447265625,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1158447265625,
0.2452392578125
],
[
0.51318359375,
0.2369384765625,
0.59033203125,
0.2369384765625,
0.59033203125,
0.248779296875,
0.51318359375,
0.248779296875
],
[
0.1158447265625,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.1158447265625,
0.262451171875
],
[
0.53076171875,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.2666015625,
0.53076171875,
0.2666015625
],
[
0.1175537109375,
0.265380859375,
0.489501953125,
0.265380859375,
0.489501953125,
0.279052734375,
0.1175537109375,
0.279052734375
],
[
0.51123046875,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.28076171875,
0.51123046875,
0.28076171875
],
[
0.1158447265625,
0.27978515625,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.29443359375
],
[
0.51318359375,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.51318359375,
0.298583984375
],
[
0.1158447265625,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.1158447265625,
0.31005859375
],
[
0.51123046875,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.51123046875,
0.3154296875
],
[
0.1175537109375,
0.312255859375,
0.489501953125,
0.312255859375,
0.489501953125,
0.325927734375,
0.1175537109375,
0.325927734375
],
[
0.509765625,
0.317626953125,
0.8818359375,
0.317626953125,
0.8818359375,
0.330810546875,
0.509765625,
0.330810546875
],
[
0.1158447265625,
0.32666015625,
0.489501953125,
0.327880859375,
0.489501953125,
0.342529296875,
0.1158447265625,
0.341552734375
],
[
0.51318359375,
0.333251953125,
0.8818359375,
0.333251953125,
0.8818359375,
0.3466796875,
0.51318359375,
0.3466796875
],
[
0.11419677734375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.11419677734375,
0.35693359375
],
[
0.5146484375,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.1158447265625,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.372314453125,
0.1158447265625,
0.372314453125
],
[
0.51318359375,
0.365234375,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.51318359375,
0.37890625
],
[
0.11419677734375,
0.3759765625,
0.16796875,
0.3759765625,
0.16796875,
0.39013671875,
0.11419677734375,
0.39013671875
],
[
0.509765625,
0.38134765625,
0.880859375,
0.38134765625,
0.880859375,
0.394287109375,
0.509765625,
0.394287109375
],
[
0.51123046875,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.412109375,
0.51123046875,
0.412109375
],
[
0.1158447265625,
0.407958984375,
0.3955078125,
0.407958984375,
0.3955078125,
0.4228515625,
0.1158447265625,
0.4228515625
],
[
0.509765625,
0.412109375,
0.880859375,
0.41455078125,
0.880859375,
0.4287109375,
0.509765625,
0.427001953125
],
[
0.509765625,
0.429931640625,
0.880859375,
0.429931640625,
0.880859375,
0.443603515625,
0.509765625,
0.443603515625
],
[
0.1175537109375,
0.435791015625,
0.489501953125,
0.435791015625,
0.489501953125,
0.449462890625,
0.1175537109375,
0.449462890625
],
[
0.51318359375,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.46142578125,
0.51318359375,
0.46142578125
],
[
0.1175537109375,
0.453125,
0.489501953125,
0.453125,
0.489501953125,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.509765625,
0.4619140625,
0.8818359375,
0.46142578125,
0.8818359375,
0.4755859375,
0.509765625,
0.476806640625
],
[
0.1175537109375,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482177734375,
0.1175537109375,
0.482177734375
],
[
0.509765625,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.509765625,
0.4921875
],
[
0.1175537109375,
0.485107421875,
0.48779296875,
0.485107421875,
0.48779296875,
0.498779296875,
0.1175537109375,
0.498779296875
],
[
0.51318359375,
0.49462890625,
0.82275390625,
0.49462890625,
0.82275390625,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.1175537109375,
0.501953125,
0.489501953125,
0.501953125,
0.489501953125,
0.515625,
0.1175537109375,
0.515625
],
[
0.1175537109375,
0.51708984375,
0.486083984375,
0.51708984375,
0.486083984375,
0.53076171875,
0.1175537109375,
0.53076171875
],
[
0.54736328125,
0.52685546875,
0.611328125,
0.52685546875,
0.611328125,
0.54150390625,
0.54736328125,
0.54150390625
],
[
0.1175537109375,
0.533203125,
0.486083984375,
0.533203125,
0.486083984375,
0.54638671875,
0.1175537109375,
0.54638671875
],
[
0.6396484375,
0.52783203125,
0.69775390625,
0.52783203125,
0.69775390625,
0.5419921875,
0.6396484375,
0.5419921875
],
[
0.1175537109375,
0.5498046875,
0.486083984375,
0.5498046875,
0.486083984375,
0.5634765625,
0.1175537109375,
0.5634765625
],
[
0.5439453125,
0.5498046875,
0.59033203125,
0.5498046875,
0.59033203125,
0.5654296875,
0.5439453125,
0.5654296875
],
[
0.63818359375,
0.5458984375,
0.7080078125,
0.5517578125,
0.705078125,
0.57080078125,
0.634765625,
0.56396484375
],
[
0.1158447265625,
0.56396484375,
0.486083984375,
0.5654296875,
0.486083984375,
0.580078125,
0.1158447265625,
0.5791015625
],
[
0.5439453125,
0.5654296875,
0.62451171875,
0.5654296875,
0.62451171875,
0.580078125,
0.5439453125,
0.580078125
],
[
0.62744140625,
0.56640625,
0.73779296875,
0.568359375,
0.73779296875,
0.58203125,
0.62744140625,
0.580078125
],
[
0.1175537109375,
0.58203125,
0.486083984375,
0.58203125,
0.486083984375,
0.595703125,
0.1175537109375,
0.595703125
],
[
0.5458984375,
0.58056640625,
0.60546875,
0.5830078125,
0.603515625,
0.5986328125,
0.5439453125,
0.5966796875
],
[
0.63818359375,
0.58203125,
0.74658203125,
0.58447265625,
0.74658203125,
0.5986328125,
0.63818359375,
0.5966796875
],
[
0.1175537109375,
0.59716796875,
0.486083984375,
0.59716796875,
0.486083984375,
0.61083984375,
0.1175537109375,
0.61083984375
],
[
0.5458984375,
0.60498046875,
0.60693359375,
0.60498046875,
0.60693359375,
0.62060546875,
0.5458984375,
0.62060546875
],
[
0.63818359375,
0.60400390625,
0.84619140625,
0.60546875,
0.84619140625,
0.62353515625,
0.63818359375,
0.62158203125
],
[
0.1158447265625,
0.6123046875,
0.489501953125,
0.61328125,
0.489501953125,
0.62744140625,
0.1158447265625,
0.626953125
],
[
0.5458984375,
0.62158203125,
0.59765625,
0.62158203125,
0.59765625,
0.63720703125,
0.5458984375,
0.63720703125
],
[
0.63818359375,
0.62158203125,
0.748046875,
0.62353515625,
0.748046875,
0.63818359375,
0.63818359375,
0.6357421875
],
[
0.1175537109375,
0.6298828125,
0.489501953125,
0.6298828125,
0.489501953125,
0.6435546875,
0.1175537109375,
0.6435546875
],
[
0.1175537109375,
0.6455078125,
0.489501953125,
0.6455078125,
0.489501953125,
0.66015625,
0.1175537109375,
0.66015625
],
[
0.509765625,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.509765625,
0.6650390625
],
[
0.1175537109375,
0.662109375,
0.489501953125,
0.662109375,
0.489501953125,
0.67578125,
0.1175537109375,
0.67578125
],
[
0.5087890625,
0.6650390625,
0.880859375,
0.6650390625,
0.880859375,
0.6787109375,
0.5087890625,
0.6787109375
],
[
0.1158447265625,
0.6767578125,
0.489501953125,
0.677734375,
0.489501953125,
0.6923828125,
0.1158447265625,
0.69140625
],
[
0.509765625,
0.68115234375,
0.8818359375,
0.68115234375,
0.8818359375,
0.69482421875,
0.509765625,
0.69482421875
],
[
0.1175537109375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.509765625,
0.6953125,
0.74072265625,
0.6953125,
0.74072265625,
0.708984375,
0.509765625,
0.708984375
],
[
0.118408203125,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.1175537109375,
0.7255859375,
0.486083984375,
0.7255859375,
0.486083984375,
0.74072265625,
0.1175537109375,
0.74072265625
],
[
0.52880859375,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.7412109375,
0.52880859375,
0.7412109375
],
[
0.1158447265625,
0.7421875,
0.48779296875,
0.7421875,
0.48779296875,
0.755859375,
0.1158447265625,
0.755859375
],
[
0.5087890625,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.5087890625,
0.75732421875
],
[
0.1175537109375,
0.76025390625,
0.2166748046875,
0.76025390625,
0.2166748046875,
0.771484375,
0.1175537109375,
0.771484375
],
[
0.51318359375,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.133544921875,
0.77490234375,
0.489501953125,
0.77490234375,
0.489501953125,
0.7890625,
0.133544921875,
0.7890625
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1158447265625,
0.79052734375,
0.486083984375,
0.79052734375,
0.486083984375,
0.80419921875,
0.1158447265625,
0.80419921875
],
[
0.51123046875,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.1158447265625,
0.8056640625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82080078125
],
[
0.5087890625,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.5087890625,
0.82177734375
],
[
0.1158447265625,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.88330078125,
0.82470703125,
0.88330078125,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85498046875,
0.1158447265625,
0.85498046875
],
[
0.509765625,
0.8388671875,
0.8818359375,
0.84033203125,
0.8818359375,
0.85498046875,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.509765625,
0.85498046875,
0.88330078125,
0.85546875,
0.88330078125,
0.87060546875,
0.509765625,
0.869140625
],
[
0.1175537109375,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.466796875,
0.90478515625,
0.466796875,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"converts VQA to an image-text matching problem",
"3.1 Traditional VQA Modules",
"and prompts the CLIP model (Radford et al., 2021),",
" There is not any standard set of modules for VQA",
"a large-scale vision-language model pre-trained on",
"We largely adopt the design of modules introduced",
"the image-text matching task. The prompts can be",
"byHu etal.(2017)with someminorchanges.We",
"either question irrelevant such as Quesion: [Ques];",
"assume that the image has been pre-processed and",
"Answer: [MASK] (QIP by Shen et al. (2022) or",
"N bounding boxes have been detected, each rep-",
"question-related by converting questions into a",
"resented as an embedding vector, collectively de-",
"masked statement (TAC-P by Song et al. (2022).",
"noted as V- (v1,V2,...,V).An attention map",
"However, a limitation with these methods is that",
"α is defined to be a distribution over the N bound-",
"several of them still require training, although the",
"ing boxes.",
"training data is not in the form of VQA. Besides,",
"Table 1 lists the most important traditional VQA",
"converting images to captions and leveraging text-",
"modules that we will replace withpre-trained mod-",
"based QA may lose important visual details dur-",
"els. The full list of modules can be found in Table 7",
"ing the caption generation step. The two methods",
"in the appendices. It is worth explaining that be-",
"above using CLIP do not address the issue that",
"sides taking in V and α as either input or output,",
"CLIP model lacks compositional and spatial rea-",
"many modules also take in the word embeddings of",
" soning abilities, which has been observed in previ-",
"sometextdescriptionextractedfromtheguestion",
"ous work (Subramanian et al., 2022: Thrush et al..",
"These text embeddings are arguments to control the",
"2022).",
"behaviors of the modules. For example, the Find",
"module's objective is to locate an object among all",
"3Modularized Zero-shot VQA",
"the bounding boxes given. The textual input goBJ",
"is therefore the word embedding of the name of the",
"Our method is motivated by Neural Module Net-",
"object to be found. Similarly, gRELA ,gATTR and",
"work (NMN) based VQA, which decomposes ques-",
"gQUERY are word embeddings for the description",
"tions into reasoning steps, where each module in",
"of relation (e.g., to the left of), attribute (e.g., red)",
"the NMN is pre-defined to perform a specific rea-",
"and aspect to query (e.g., querying name).",
"soning task. The idea allows us to select appropri-",
"ate pre-trained models to handle different reasoning",
"Module",
"tasks in a question. Specifically, in NMN-based",
"Inputs",
"VQA, we first manually define a set of reasoning",
"Find",
"V,goBJ",
"steps such as object detection and spatial reasoning:",
"Relocate",
"Q, V,gRELA",
"each represented by a module. A question is then",
"Filter",
"α, V, gcONDI",
"explicitly decomposed and converted into a layout",
"Choose",
"Q1, Q2, V, gRELAI, gRELA?",
"of modules, which is an executable program show-",
"Query",
"Q, V, gQUERY",
"ing the reasoning chain to reach the final answer.",
"The top section of Figure 1 shows the layout corre-",
"Table 1: A subset of the modules in traditional NMN",
"sponding to the sample question. To train an NMN-",
"that we replace with pre-trained models. Modules in",
"based VQA system, usually a layout generator is",
" the first block output an attention map and those in the",
"separately built first, which either uses hand-crafted",
" second block generate an answer.",
"rules over dependency parses of guestions or is a",
"trained seq2seq model. Then, the parameters of",
"Traditionally, the parameters of the modules in",
"the various VQA modules are learned from VQA",
" Table 1 need to be learned from VQA training data.",
"training data.",
"In other words, these modules' underlying capabili-",
"For our work, we do not want to use VQA data",
"ties such as object recognition and relational reason-",
"for training. But we observe that many modules in",
"ing need tobeacquiredfrom VQA data.However",
"NMN-based VQA can be supported by pre-trained",
" we hypothesize that recently developed pre-trained",
"models that have already acquired the capabilities",
"models may already have some of these capabili-",
"needed by these modules. The key component of",
"ties and can therefore directly equip these modules",
"our method is therefore to map a layout of modules",
"with such capabilities. For example, the Find mod-",
"produced by traditional NMN-based VQA to a sim-",
"ule is mainly responsible for object recognition,",
"plified layout of zero-shot components that can be",
"and previously the parameters of Find have to be",
"implemented directly using pre-trained models.",
"learned from scratch using VQA data. Now with"
] |
{
"arxiv_id": "2305.17369",
"language": "en",
"page_idx": 2
}
|
||
2305.17369-en-3
|
2305.17369-en-3.jpg
|
a powerful pre-trained model such as OWL (Minderer et al., 2022) that can recognize a wide range of objects, we can presumably directly use a model like OWL to replace the traditional Find module.
## 3.2 Pre-trained Models
We utilize three pre-trained models that we believe
are highly relevant to VQA.
$\mathbf{OWL.}$ The Vision Transformer for Open-World Localization (OWL) model (Minderer et al., 2022) is a model for open-vocabulary object detection. It is first pre-trained on large-scale image-text pairs and then fine-tuned with added detection heads and medium-sized detection data. Given the category name of an object and an image, the model is able to locate bounding box(es) in the image containing the object together with a confidence score for each $box.$
MDETR. The modulated DETER (DEtection TRansformer) model (Kamath et al., 2021) is an end-to-end detector that can detect an object in an image conditioned on a piece of textual description of the object such as its attributes and its relation with another object in the image. The model is pre-trained on image-text pairs with explicit alignment between phrases in the text and bounding boxes of objects in the image. Given an image and the description of an object, MDETR is able to locate the bounding box(es) in the image containing the object satisfying the description. Note that different from OWL, MDETR is able to understand textual descriptions that may contain attribute information and/or complex visual relations. For example, given the description $a$ man holding a yellow cup is talking, MDETR will detect the bounding box containing the man holding a yellow cup in the given image, whereas OWL is not able to use the description and will only recognize all bounding boxes containing a man. Note that we use the version of MDETR pre-trained on general modulated detection without fine-tuning for any downstream tasks.
CLIP. CLIP is a well-known large-scale visionlanguage model by OpenAI. It is pre-trained with 400M image-caption pairs through contrastive learning. Given an (image, text) pair, CLIP uses its separate image encoder and text encoder to turn the image and the text each into a vector, and the cosine similarity between the two vectors directly measures the compatibility of the two. Recent work
has shown that CLIP can be directly used for VQA in a zero-shot setting, if we can come up with a set of candidate answers and transform each (question, answer) pair into a statement (Song et al., 2022).
3.3 Zero-shot NMN using Pre-trained Models Based on the descriptions of the traditional VQA modules in Section 3.1 and of the three PTMs we consider in Section 3.2, we can see that there are obvious connections between the capabilities desired by the traditional modules and the capabilities that these PTMs have already acquired.
However, the mapping between them is not trivial. First of all, there is no simple one-to-one mapping from traditional VQA modules to the PTMs. For example, the MDETR model can already perform multiple steps of reasoning to locate the desired object, so it can be used to cover a sequence of modules in an NMN layout. Second, there may be capabilities required when applying PTMs but not captured by modules defined in NMN-based VQA. In particular, the MDETR model always assumes that the object to be grounded exists in the given image, but for those questions asking for the existence of a specified object, we cannot directly use MDETR
To address these challenges, we carefully design a mapping mechanism that can map an NMN-based module layout to a simplified layout consisting of a few zero-shot modules. Three of these zero-shot modules (OWL, MDETR and CL IP) correspond exactly to the three PTMs introduced earlier. The rest of the zero-shot modules are defined by simple heuristic rules. We list these zero-shot modules in Table 2.
<table>
<tbody>
<tr>
<th>$\mathbf{Module}$</th>
<th>Inputs</th>
<th>$\mathbf{Output}$</th>
</tr>
<tr>
<td>$OWL$</td>
<td>$I,OBJ$</td>
<td>$\mathcal{B},\mathbf{s}$</td>
</tr>
<tr>
<td>MDETR</td>
<td>$I$,SENT</td>
<td>$B,\mathbf{s}$</td>
</tr>
<tr>
<td>$CLIP$</td>
<td>$\mathcal{B},I,V$</td>
<td>$Ans.$</td>
</tr>
<tr>
<td>Count</td>
<td>$\textcircled{B}$</td>
<td>$Num.$</td>
</tr>
<tr>
<td>$\mathsf{Exist}$</td>
<td>$\mathcal{B} $, (ATTR/RELA</td>
<td>$Yes/No$</td>
</tr>
<tr>
<td>And</td>
<td>$\mathsf{Exist}_{1}$, $,Exist_{2}$</td>
<td>$Yes/No$</td>
</tr>
<tr>
<td>Or</td>
<td>$\mathsf{Exist}_{1}$, $,Exist_{2}$</td>
<td>$Yes/No$</td>
</tr>
</tbody>
</table>
Table 2: Zero-shot modules with either pre-trained models or heuristics. The $I$ is the VQA image, $\nu$ is the answer vocabulary and $B$ is the set of bounding boxes.
We now give a high-level summary of the mapping mechanism below. We first look at the last module in the NMN layout. If the last module
|
[
[
0.1158447265625,
0.08489990234375,
0.496337890625,
0.0843505859375,
0.496337890625,
0.09857177734375,
0.1158447265625,
0.09979248046875
],
[
0.5087890625,
0.0843505859375,
0.87890625,
0.0843505859375,
0.87890625,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.1175537109375,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.51123046875,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.51123046875,
0.114013671875
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.509765625,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.480224609375,
0.133056640625,
0.480224609375,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.509765625,
0.1341552734375,
0.86962890625,
0.133056640625,
0.86962890625,
0.1478271484375,
0.509765625,
0.1484375
],
[
0.1158447265625,
0.1591796875,
0.313232421875,
0.1602783203125,
0.313232421875,
0.174560546875,
0.1158447265625,
0.17333984375
],
[
0.51318359375,
0.162109375,
0.8818359375,
0.162109375,
0.8818359375,
0.17578125,
0.51318359375,
0.17578125
],
[
0.11419677734375,
0.18115234375,
0.489501953125,
0.182861328125,
0.489501953125,
0.19775390625,
0.11419677734375,
0.1953125
],
[
0.51123046875,
0.182861328125,
0.87890625,
0.182861328125,
0.87890625,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.1158447265625,
0.1995849609375,
0.32421875,
0.1995849609375,
0.32421875,
0.213134765625,
0.1158447265625,
0.213134765625
],
[
0.51123046875,
0.20068359375,
0.8818359375,
0.20068359375,
0.8818359375,
0.21435546875,
0.51123046875,
0.21435546875
],
[
0.51318359375,
0.218505859375,
0.880859375,
0.218505859375,
0.880859375,
0.2286376953125,
0.51318359375,
0.2286376953125
],
[
0.1175537109375,
0.2244873046875,
0.48779296875,
0.2244873046875,
0.48779296875,
0.2381591796875,
0.1175537109375,
0.2381591796875
],
[
0.509765625,
0.2327880859375,
0.8818359375,
0.2327880859375,
0.8818359375,
0.2464599609375,
0.509765625,
0.2464599609375
],
[
0.11419677734375,
0.240478515625,
0.48779296875,
0.2392578125,
0.48779296875,
0.254150390625,
0.11419677734375,
0.254638671875
],
[
0.509765625,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.509765625,
0.26318359375
],
[
0.1175537109375,
0.2587890625,
0.486083984375,
0.2587890625,
0.486083984375,
0.26953125,
0.1175537109375,
0.26953125
],
[
0.509765625,
0.265380859375,
0.8037109375,
0.265380859375,
0.8037109375,
0.279052734375,
0.509765625,
0.279052734375
],
[
0.11419677734375,
0.2724609375,
0.486083984375,
0.2724609375,
0.486083984375,
0.2861328125,
0.11419677734375,
0.2861328125
],
[
0.5283203125,
0.27978515625,
0.88330078125,
0.28076171875,
0.88330078125,
0.295654296875,
0.5283203125,
0.29443359375
],
[
0.1158447265625,
0.289306640625,
0.48779296875,
0.289306640625,
0.48779296875,
0.302734375,
0.1158447265625,
0.302734375
],
[
0.509765625,
0.29638671875,
0.8818359375,
0.297607421875,
0.8818359375,
0.312255859375,
0.509765625,
0.311279296875
],
[
0.1158447265625,
0.305908203125,
0.486083984375,
0.305908203125,
0.486083984375,
0.319580078125,
0.1158447265625,
0.319580078125
],
[
0.509765625,
0.3134765625,
0.88330078125,
0.312255859375,
0.88330078125,
0.32666015625,
0.509765625,
0.327880859375
],
[
0.1158447265625,
0.322509765625,
0.48779296875,
0.322509765625,
0.48779296875,
0.336181640625,
0.1158447265625,
0.336181640625
],
[
0.51123046875,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.1158447265625,
0.33837890625,
0.48779296875,
0.33837890625,
0.48779296875,
0.3515625,
0.1158447265625,
0.3515625
],
[
0.51123046875,
0.3466796875,
0.88330078125,
0.3466796875,
0.88330078125,
0.35986328125,
0.51123046875,
0.35986328125
],
[
0.1158447265625,
0.35400390625,
0.48779296875,
0.35400390625,
0.48779296875,
0.3681640625,
0.1158447265625,
0.3681640625
],
[
0.51123046875,
0.361083984375,
0.88330078125,
0.3623046875,
0.88330078125,
0.377197265625,
0.51123046875,
0.3759765625
],
[
0.1158447265625,
0.37060546875,
0.151123046875,
0.37060546875,
0.151123046875,
0.383056640625,
0.1158447265625,
0.383056640625
],
[
0.51123046875,
0.377197265625,
0.8818359375,
0.377685546875,
0.8818359375,
0.392578125,
0.51123046875,
0.391357421875
],
[
0.1158447265625,
0.393798828125,
0.489501953125,
0.394287109375,
0.489501953125,
0.4091796875,
0.1158447265625,
0.407958984375
],
[
0.51123046875,
0.394287109375,
0.8818359375,
0.394287109375,
0.8818359375,
0.407958984375,
0.51123046875,
0.407958984375
],
[
0.1175537109375,
0.412109375,
0.48779296875,
0.412109375,
0.48779296875,
0.42578125,
0.1175537109375,
0.42578125
],
[
0.51123046875,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.118408203125,
0.429931640625,
0.486083984375,
0.429931640625,
0.486083984375,
0.440673828125,
0.118408203125,
0.440673828125
],
[
0.51318359375,
0.427001953125,
0.8818359375,
0.427001953125,
0.8818359375,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.1158447265625,
0.44482421875,
0.486083984375,
0.44482421875,
0.486083984375,
0.457763671875,
0.1158447265625,
0.457763671875
],
[
0.51123046875,
0.4423828125,
0.8818359375,
0.4423828125,
0.8818359375,
0.4560546875,
0.51123046875,
0.4560546875
],
[
0.1175537109375,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.1175537109375,
0.473876953125
],
[
0.51318359375,
0.458984375,
0.8818359375,
0.458984375,
0.8818359375,
0.47265625,
0.51318359375,
0.47265625
],
[
0.1158447265625,
0.476806640625,
0.486083984375,
0.476806640625,
0.486083984375,
0.490478515625,
0.1158447265625,
0.490478515625
],
[
0.51123046875,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.488037109375,
0.51123046875,
0.488037109375
],
[
0.1175537109375,
0.4921875,
0.489501953125,
0.4921875,
0.489501953125,
0.505859375,
0.1175537109375,
0.505859375
],
[
0.5087890625,
0.490478515625,
0.611328125,
0.4892578125,
0.611328125,
0.50439453125,
0.5087890625,
0.5048828125
],
[
0.1175537109375,
0.5087890625,
0.48779296875,
0.5087890625,
0.48779296875,
0.5224609375,
0.1175537109375,
0.5224609375
],
[
0.5283203125,
0.505859375,
0.8818359375,
0.50732421875,
0.8818359375,
0.521484375,
0.5283203125,
0.52099609375
],
[
0.1158447265625,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.1158447265625,
0.5380859375
],
[
0.5087890625,
0.52392578125,
0.8818359375,
0.5224609375,
0.8818359375,
0.53759765625,
0.5087890625,
0.5380859375
],
[
0.11419677734375,
0.54052734375,
0.489501953125,
0.5390625,
0.489501953125,
0.55419921875,
0.11419677734375,
0.5546875
],
[
0.509765625,
0.54052734375,
0.88330078125,
0.54052734375,
0.88330078125,
0.55419921875,
0.509765625,
0.55419921875
],
[
0.11419677734375,
0.5556640625,
0.491943359375,
0.5556640625,
0.491943359375,
0.5693359375,
0.11419677734375,
0.5693359375
],
[
0.509765625,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.509765625,
0.5693359375
],
[
0.118408203125,
0.57470703125,
0.48779296875,
0.57470703125,
0.48779296875,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.509765625,
0.572265625,
0.8818359375,
0.572265625,
0.8818359375,
0.5859375,
0.509765625,
0.5859375
],
[
0.1175537109375,
0.58837890625,
0.491943359375,
0.58837890625,
0.491943359375,
0.6015625,
0.1175537109375,
0.6015625
],
[
0.509765625,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.60107421875,
0.509765625,
0.60107421875
],
[
0.1175537109375,
0.60693359375,
0.49365234375,
0.60693359375,
0.49365234375,
0.61767578125,
0.1175537109375,
0.61767578125
],
[
0.509765625,
0.60400390625,
0.8818359375,
0.60400390625,
0.8818359375,
0.61767578125,
0.509765625,
0.61767578125
],
[
0.1158447265625,
0.62158203125,
0.49365234375,
0.62158203125,
0.49365234375,
0.63525390625,
0.1158447265625,
0.63525390625
],
[
0.5087890625,
0.619140625,
0.875,
0.6181640625,
0.875,
0.6328125,
0.5087890625,
0.63427734375
],
[
0.1175537109375,
0.63720703125,
0.48779296875,
0.63720703125,
0.48779296875,
0.65185546875,
0.1175537109375,
0.65185546875
],
[
0.54736328125,
0.65087890625,
0.615234375,
0.65087890625,
0.615234375,
0.6650390625,
0.54736328125,
0.6650390625
],
[
0.62744140625,
0.65283203125,
0.68408203125,
0.65283203125,
0.68408203125,
0.66455078125,
0.62744140625,
0.66455078125
],
[
0.78271484375,
0.65087890625,
0.84814453125,
0.65087890625,
0.84814453125,
0.6650390625,
0.78271484375,
0.6650390625
],
[
0.1175537109375,
0.66943359375,
0.486083984375,
0.66943359375,
0.486083984375,
0.68310546875,
0.1175537109375,
0.68310546875
],
[
0.1158447265625,
0.65380859375,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.1158447265625,
0.66748046875
],
[
0.54736328125,
0.67333984375,
0.5810546875,
0.67333984375,
0.5810546875,
0.689453125,
0.54736328125,
0.689453125
],
[
0.62890625,
0.67333984375,
0.68701171875,
0.67333984375,
0.68701171875,
0.689453125,
0.62890625,
0.689453125
],
[
0.78271484375,
0.67138671875,
0.8212890625,
0.67333984375,
0.8193359375,
0.6904296875,
0.779296875,
0.68701171875
],
[
0.1158447265625,
0.68505859375,
0.484375,
0.68408203125,
0.484375,
0.69873046875,
0.1158447265625,
0.69970703125
],
[
0.62744140625,
0.689453125,
0.69775390625,
0.68701171875,
0.69873046875,
0.703125,
0.62890625,
0.705078125
],
[
0.78271484375,
0.6884765625,
0.82275390625,
0.69140625,
0.8193359375,
0.70703125,
0.78076171875,
0.705078125
],
[
0.1175537109375,
0.70166015625,
0.484375,
0.70166015625,
0.484375,
0.71533203125,
0.1175537109375,
0.71533203125
],
[
0.54736328125,
0.6904296875,
0.60107421875,
0.6904296875,
0.60107421875,
0.705078125,
0.54736328125,
0.705078125
],
[
0.54736328125,
0.7060546875,
0.591796875,
0.7060546875,
0.591796875,
0.72021484375,
0.54736328125,
0.72021484375
],
[
0.62744140625,
0.7060546875,
0.69189453125,
0.7060546875,
0.69189453125,
0.72021484375,
0.62744140625,
0.72021484375
],
[
0.78271484375,
0.7060546875,
0.82275390625,
0.7060546875,
0.82275390625,
0.7216796875,
0.78271484375,
0.7216796875
],
[
0.1175537109375,
0.71728515625,
0.486083984375,
0.71728515625,
0.486083984375,
0.73095703125,
0.1175537109375,
0.73095703125
],
[
0.54736328125,
0.72802734375,
0.60107421875,
0.72998046875,
0.599609375,
0.7451171875,
0.5458984375,
0.74365234375
],
[
0.62744140625,
0.728515625,
0.6484375,
0.728515625,
0.6484375,
0.74365234375,
0.62744140625,
0.74365234375
],
[
0.78271484375,
0.728515625,
0.82861328125,
0.728515625,
0.82861328125,
0.74462890625,
0.78271484375,
0.74462890625
],
[
0.54736328125,
0.74462890625,
0.60107421875,
0.74462890625,
0.60107421875,
0.76025390625,
0.54736328125,
0.76025390625
],
[
0.1158447265625,
0.73388671875,
0.484375,
0.73388671875,
0.484375,
0.74755859375,
0.1158447265625,
0.74755859375
],
[
0.62890625,
0.7451171875,
0.76904296875,
0.7451171875,
0.76904296875,
0.7587890625,
0.62890625,
0.7587890625
],
[
0.7841796875,
0.74658203125,
0.8369140625,
0.74658203125,
0.8369140625,
0.75830078125,
0.7841796875,
0.75830078125
],
[
0.62646484375,
0.75830078125,
0.75244140625,
0.76025390625,
0.75244140625,
0.77783203125,
0.62646484375,
0.775390625
],
[
0.1158447265625,
0.7490234375,
0.484375,
0.75,
0.484375,
0.76416015625,
0.1158447265625,
0.76318359375
],
[
0.54736328125,
0.76123046875,
0.58447265625,
0.76123046875,
0.58447265625,
0.77685546875,
0.54736328125,
0.77685546875
],
[
0.7841796875,
0.76123046875,
0.84033203125,
0.76123046875,
0.84033203125,
0.775390625,
0.7841796875,
0.775390625
],
[
0.5458984375,
0.77392578125,
0.57666015625,
0.77783203125,
0.57275390625,
0.79443359375,
0.54150390625,
0.79052734375
],
[
0.1158447265625,
0.7666015625,
0.255126953125,
0.7666015625,
0.255126953125,
0.77978515625,
0.1158447265625,
0.77978515625
],
[
0.62744140625,
0.77685546875,
0.75146484375,
0.77783203125,
0.75146484375,
0.7919921875,
0.62744140625,
0.79150390625
],
[
0.7841796875,
0.77685546875,
0.84033203125,
0.77685546875,
0.84033203125,
0.7919921875,
0.7841796875,
0.7919921875
],
[
0.118408203125,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.118408203125,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.509765625,
0.82080078125
],
[
0.1158447265625,
0.822265625,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.509765625,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.509765625,
0.8359375
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.8388671875,
0.489501953125,
0.85400390625,
0.1158447265625,
0.85498046875
],
[
0.51123046875,
0.83740234375,
0.87890625,
0.83740234375,
0.87890625,
0.85107421875,
0.51123046875,
0.85107421875
],
[
0.1158447265625,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.53076171875,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.53076171875,
0.8857421875
],
[
0.1175537109375,
0.89013671875,
0.486083984375,
0.89013671875,
0.486083984375,
0.90087890625,
0.1175537109375,
0.90087890625
],
[
0.509765625,
0.88818359375,
0.88330078125,
0.88720703125,
0.88330078125,
0.90185546875,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"a powerful pre-trained model such as OWL (Min-",
" has shown that CLIP can be directly used for VQA",
"derer et al., 2022) that can recognize a wide range",
"in a zero-shot setting. if we can come up with a set",
"of objects, we can presumably directly use a model",
"of candidate answers and transform each (question",
"like OWL to replace the traditional Find module.",
"answer) pair into a statement (Song et al., 2022).",
"3.2Pre-trained Models",
"3.3Zero-shot NMN using Pre-trained Models",
"We utilize three pre-trained models that we believe",
"Based on the descriptions of the traditional VOA",
"are highly relevant to VQA.",
"modules in Section 3.1 and of the three PTMs we",
"consider in Section 3.2. we can see that there are",
"oWL.TheVision Transformerfor Open-World",
"obvious connections between the capabilities de",
"Localization (OWL) model (Minderer et al., 2022)",
" sired by the traditional modules and the capabilities",
"is amodelfor open-vocabularyobject detection.It",
" that these PTMs have already acquired.",
"is first pre-trained on large-scale image-text pairs",
"However, the mapping between them is not triv-",
"and then fine-tuned with added detection heads and",
"ial. First of all, there is no simple one-to-one map-",
"medium-sized detection data. Given the category",
"ping from traditional VQA modules to the PTMs.",
"name of an object and an image, the model is able",
"For example, the MDETR model can already per-",
"to locate bounding box(es) in the image containing",
"form multiple steps of reasoning to locate the de-",
"the object together with a confidence score for each",
"sired object, so it can be used to cover a sequence",
"box.",
"of modules in an NMN layout. Second, there may",
"MDETR. The modulated DETER (DEtection",
"be capabilities required when applying PTMs but",
"TRansformer) model (Kamath et al., 2021) is an",
"not captured by modules defined in NMN-based",
"end-to-end detectorthat can detectan objectinan",
"VQA. In particular, the MDETR model always as",
"image conditioned on a piece of textual description",
"sumes that the object to be grounded exists in the",
"of the object such as its attributes and its relation",
"given image, but for those questions asking for the",
"with another object in the image. The model is",
"existence of a specified object, we cannot directly",
"pre-trained on image-text pairs with explicit align-",
"use MDETR.",
"ment between phrases in the text and bounding",
"To address these challenges, we carefully design",
"boxes of objects in the image. Given an image",
"a mapping mechanism that can map an NMN-based",
"and the description of an object, MDETR is able",
" module layout to a simplified layout consisting of",
"to locate the bounding box(es) in the image con-",
"a few zero-shot modules. Three of these zero-shot",
"taining the object satisfying the description.Note",
"modules (OWL, MDETR and CLIP) correspond exactly",
"that different from OWL, MDETR is able to un-",
"to the three PTMs introduced earlier. The rest of the",
"derstand textualdescriptions thatmay contain at-",
"zero-shot modules are defined by simple heuristic",
"tribute information and/or complex visual relations.",
"rules. We list these zero-shot modules in Table 2.",
"For example, given the description a man holding",
"Module",
"Inputs",
"Output",
"bounding box containing the man holding a yellow",
"a yellow cup is talking, MDETR will detect the",
"OWL",
"I, OBJ",
"B,s",
"cup in the given image, whereas OWL is not able",
"I, SENT",
"B,s",
"to use the description and will only recognize all",
"MDETR",
"CLIP",
"B, I, V",
"Ans.",
"bounding boxes containing a man. Note that we",
"Count",
"B",
"Num.",
"Exist",
"use the version of MDETR pre-trained on general",
"B, (ATTR/RELA)",
"Yes/No",
"Existi,Exist2",
"modulated detection without fine-tuning for any",
"And",
"Yes/No",
"or",
"downstream tasks.",
"Existi,Exist2",
"Yes/No",
"CLIP. CLIP is a well-known large-scale vision-",
"language model by OpenAI. It is pre-trained with",
"Table 2: Zero-shot modules with either pre-trained mod",
"400M image-caption pairs through contrastive",
"els or heuristics. The I is the VQA image, V is the",
"learning. Given an (image, text) pair, CLIP uses",
"answer vocabulary and B is the set of bounding boxes",
"its separate image encoder and text encoder to turn",
"the image and the text each into a vector, and the",
"We now give a high-level summary of the map-",
"cosine similaritybetweenthetwovectorsdirectly",
"ping mechanism below. We first look at the last",
"measures the compatibility of the two. Recent work",
"module in the NMN layout. If the last module"
] |
{
"arxiv_id": "2305.17369",
"language": "en",
"page_idx": 3
}
|
||
2305.17369-en-4
|
2305.17369-en-4.jpg
|
is one of Choose, Compare and Query, we know that the input to this last module is either a single attention map or two attention maps, where each attention map essentially tries to capture an object matching some textual descriptions. By tracing the path in the layout leading to the attention map, we choose either the zero-shot OWL module (when the path has a length of 1) or the zero-shot MDETR module (when the path is longer than 1 hop). This is because when the path length equals to one, it involves only object detection (corresponding to a single Find module in the NMN layout for generation of the attention map). When the path length is more than one, it indicates the generation of the attention map in the NMN layout involves other modules such as Filter and Relocate, which calls for the other abilities than object detection, such as language understanding, attribute recognition and relational reasoning. Different from NMN modules which takes in image features and object embeddings to generate an attention map, our zero-shot OWL and zero-shot MDETR takes in the raw image and raw texts to locate (OBJ for OWL and SENT for MDETR) to generate a set of detected bounding boxes $\mathcal{B}=\{\mathbf{b}_{n}\}_{n=1}^{N}$ together with their confident scores $\mathbf{s}\in\mathbb{R}^N$, where $\mathbf{b}_n\in\mathbb{R}^4$ represents the relative position and size of the detected bounding box in the image. We keep only the bounding box from either OWL or MDETR with the highest confident score and feed it to CLIP. We generate an answer by leveraging the capability of multimodal matching of CLIP. Specifically, given $\mathcal{B}$, we generate an input image (which we refer to as $I^\mathrm{in})$ by either masking regions not containing those detected boxes (|B|=2) or cropping the image so that only the part containing the box remains $(|\beta|=1).$ If the final NMN module is Choose, we generate a masked template by question conversion as in (Song et al., 2022); otherwise the masked template will be a simple “[MASK]”. Then we match the image $I^\mathrm{in}$ with the template where the[MASK] token is replaced by each of the answer candidates in $V.$ We then select the answer that, when placed inside the template, best matches the image.
If the module is Exist, we trace back the path leading to Exist to determine whether the module is asking for the existence of an object, an attribute or a relation. For object existence (e.g., is there $a\textit{ car) , we use the zero- shot OWL module. For at- }$ tribute existence and relation existence, we first verify whether all mentioned nouns (objects) de-
tected by a POS tagger in the question exist with the OWL module. Once we detect an object that does not exist, the predicted answer will be no. If all objects exist, then we generate corresponding bounding boxes leveraging either OWL or MDETR following the method described in the paragraph above. For attribute existence, we generate a pair of a positive and a negative descriptions: (ATTR, not ATTR), e.g., $( red, not\textit{red}) .$ We then find which description aligns better with the cropped image according to b. If the image aligns better with the positive statement, then the answer will be yes; otherwise, $no.$ For relation existence, we generate the masked image $I^\mathrm{in}$ according to b$_1$ and $\mathbf{b}_2$ (the bounding boxes of the two objects whose relation is to be checked) and a pair of opposite statements regarding the relation to be checked, following (Song et al., 2022). For example, if the question is to check whether $A$ is holding $B$, the two opposite statements will be A is holding B and A is not holding $B.$ For both attribute and relation existence, we use zero-shot CLIP for the alignment between the input image and the statements. More details and the work flows of existence-related questions are provided in Appendix C.
If the module is Count, we directly count the number of bounding boxes in $\beta$ returned either from OWL or MDETR. Finally, if the last module is a logical AND or logical OR, we further trace to the inputs of this module, which should both be an Exist module. We then use the same mechanism described above for Exist to process the module. By receiving the outputs from the Exist modules, logical operations will be applied to determine the output. The deterministic logical operations can be found in Appendix B.
### 3.4 Spatial Heuristics
As mentioned in (Subramanian et al., 2022), CLIP is less capable of spatial reasoning. Using CLIP for answer generation may not be enough when it involves spatial relation understanding. Following (Subramanian et al., 2022), we define simple and general heuristics to perform certain types of spatial reasoning. Note that only when we decompose questions explicitly can we insert the spatial heuristics into CLIP-based answer generation because we have the intermediate outputs from previous reasoning steps.
First of all, given the coordinates and the size of a bounding box, we use manual rules (named
|
[
[
0.1158447265625,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.1158447265625,
0.09979248046875
],
[
0.509765625,
0.08489990234375,
0.880859375,
0.08489990234375,
0.880859375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.118408203125,
0.10394287109375,
0.48779296875,
0.10394287109375,
0.48779296875,
0.114013671875,
0.118408203125,
0.114013671875
],
[
0.509765625,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
0.1158447265625,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.118408203125,
0.135986328125,
0.48779296875,
0.135986328125,
0.48779296875,
0.146728515625,
0.118408203125,
0.146728515625
],
[
0.5087890625,
0.133056640625,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1484375,
0.5087890625,
0.1478271484375
],
[
0.1175537109375,
0.15087890625,
0.489501953125,
0.15087890625,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.509765625,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.509765625,
0.1644287109375
],
[
0.1175537109375,
0.166259765625,
0.490234375,
0.166259765625,
0.490234375,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.509765625,
0.1650390625,
0.880859375,
0.166259765625,
0.880859375,
0.18115234375,
0.509765625,
0.179931640625
],
[
0.118408203125,
0.18408203125,
0.48779296875,
0.18408203125,
0.48779296875,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.51123046875,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.21435546875,
0.489501953125,
0.21435546875,
0.489501953125,
0.22802734375,
0.1175537109375,
0.22802734375
],
[
0.509765625,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.11419677734375,
0.2298583984375,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.11419677734375,
0.24462890625
],
[
0.51123046875,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.51123046875,
0.2452392578125
],
[
0.1158447265625,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.509765625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.509765625,
0.26123046875
],
[
0.1158447265625,
0.26318359375,
0.49365234375,
0.26318359375,
0.49365234375,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.5087890625,
0.262451171875,
0.88330078125,
0.26318359375,
0.88330078125,
0.27783203125,
0.5087890625,
0.276611328125
],
[
0.1175537109375,
0.279052734375,
0.490234375,
0.279052734375,
0.490234375,
0.29345703125,
0.1175537109375,
0.29345703125
],
[
0.509765625,
0.279052734375,
0.8818359375,
0.279052734375,
0.8818359375,
0.292236328125,
0.509765625,
0.292236328125
],
[
0.1175537109375,
0.295654296875,
0.491943359375,
0.295654296875,
0.491943359375,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51318359375,
0.29443359375,
0.8818359375,
0.29443359375,
0.8818359375,
0.308837890625,
0.51318359375,
0.308837890625
],
[
0.1175537109375,
0.311279296875,
0.490234375,
0.311279296875,
0.490234375,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.509765625,
0.31005859375,
0.880859375,
0.31005859375,
0.880859375,
0.32373046875,
0.509765625,
0.32373046875
],
[
0.1158447265625,
0.325927734375,
0.48779296875,
0.325927734375,
0.48779296875,
0.339111328125,
0.1158447265625,
0.339111328125
],
[
0.51123046875,
0.328857421875,
0.880859375,
0.328857421875,
0.880859375,
0.339111328125,
0.51123046875,
0.339111328125
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.51318359375,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.1175537109375,
0.35986328125,
0.483642578125,
0.35986328125,
0.483642578125,
0.37060546875,
0.1175537109375,
0.37060546875
],
[
0.509765625,
0.358154296875,
0.8818359375,
0.359375,
0.8818359375,
0.37353515625,
0.509765625,
0.372314453125
],
[
0.1158447265625,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.1158447265625,
0.388427734375
],
[
0.509765625,
0.37353515625,
0.8818359375,
0.374755859375,
0.8818359375,
0.3896484375,
0.509765625,
0.388427734375
],
[
0.1158447265625,
0.39013671875,
0.489501953125,
0.39013671875,
0.489501953125,
0.40380859375,
0.1158447265625,
0.40380859375
],
[
0.51123046875,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.51123046875,
0.405029296875
],
[
0.1175537109375,
0.407958984375,
0.490234375,
0.407958984375,
0.490234375,
0.421630859375,
0.1175537109375,
0.421630859375
],
[
0.51123046875,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.51123046875,
0.421630859375
],
[
0.1158447265625,
0.42333984375,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.1158447265625,
0.43701171875
],
[
0.51123046875,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.51123046875,
0.43701171875
],
[
0.11419677734375,
0.439453125,
0.48779296875,
0.439453125,
0.48779296875,
0.453125,
0.11419677734375,
0.453125
],
[
0.509765625,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.509765625,
0.45361328125
],
[
0.1175537109375,
0.4560546875,
0.489501953125,
0.4560546875,
0.489501953125,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.509765625,
0.457275390625,
0.8818359375,
0.457275390625,
0.8818359375,
0.47021484375,
0.509765625,
0.47021484375
],
[
0.11248779296875,
0.4697265625,
0.491943359375,
0.4697265625,
0.491943359375,
0.490478515625,
0.11248779296875,
0.490478515625
],
[
0.509765625,
0.47265625,
0.7001953125,
0.47265625,
0.7001953125,
0.486328125,
0.509765625,
0.486328125
],
[
0.1158447265625,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.1158447265625,
0.501953125
],
[
0.52880859375,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.50439453125,
0.52880859375,
0.50439453125
],
[
0.1175537109375,
0.50439453125,
0.489501953125,
0.50439453125,
0.489501953125,
0.51708984375,
0.1175537109375,
0.51708984375
],
[
0.509765625,
0.505859375,
0.88330078125,
0.505859375,
0.88330078125,
0.51953125,
0.509765625,
0.51953125
],
[
0.1158447265625,
0.52099609375,
0.489501953125,
0.52099609375,
0.489501953125,
0.53369140625,
0.1158447265625,
0.53369140625
],
[
0.5087890625,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.5361328125,
0.5087890625,
0.5361328125
],
[
0.1158447265625,
0.53515625,
0.486083984375,
0.53515625,
0.486083984375,
0.548828125,
0.1158447265625,
0.548828125
],
[
0.50732421875,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.50732421875,
0.5517578125
],
[
0.1175537109375,
0.552734375,
0.490234375,
0.552734375,
0.490234375,
0.56640625,
0.1175537109375,
0.56640625
],
[
0.5087890625,
0.55419921875,
0.880859375,
0.55419921875,
0.880859375,
0.56787109375,
0.5087890625,
0.56787109375
],
[
0.1175537109375,
0.568359375,
0.490234375,
0.568359375,
0.490234375,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.51318359375,
0.57080078125,
0.8818359375,
0.57080078125,
0.8818359375,
0.58447265625,
0.51318359375,
0.58447265625
],
[
0.118408203125,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.118408203125,
0.5986328125
],
[
0.5146484375,
0.58837890625,
0.880859375,
0.58837890625,
0.880859375,
0.5986328125,
0.5146484375,
0.5986328125
],
[
0.1158447265625,
0.599609375,
0.48779296875,
0.60107421875,
0.48779296875,
0.615234375,
0.1158447265625,
0.6142578125
],
[
0.51318359375,
0.6025390625,
0.88330078125,
0.6025390625,
0.88330078125,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1175537109375,
0.61767578125,
0.489501953125,
0.61767578125,
0.489501953125,
0.63134765625,
0.1175537109375,
0.63134765625
],
[
0.51123046875,
0.62060546875,
0.880859375,
0.62060546875,
0.880859375,
0.63134765625,
0.51123046875,
0.63134765625
],
[
0.1158447265625,
0.6328125,
0.489501953125,
0.6318359375,
0.489501953125,
0.646484375,
0.1158447265625,
0.64794921875
],
[
0.51123046875,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.51123046875,
0.6484375
],
[
0.1175537109375,
0.6484375,
0.48779296875,
0.6484375,
0.48779296875,
0.662109375,
0.1175537109375,
0.662109375
],
[
0.509765625,
0.65087890625,
0.67529296875,
0.65087890625,
0.67529296875,
0.6650390625,
0.509765625,
0.6650390625
],
[
0.1158447265625,
0.6650390625,
0.48779296875,
0.6650390625,
0.48779296875,
0.6787109375,
0.1158447265625,
0.6787109375
],
[
0.1158447265625,
0.68115234375,
0.489501953125,
0.68115234375,
0.489501953125,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.509765625,
0.68408203125,
0.6962890625,
0.68408203125,
0.6962890625,
0.69775390625,
0.509765625,
0.69775390625
],
[
0.11419677734375,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.71044921875,
0.11419677734375,
0.71044921875
],
[
0.11248779296875,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.11248779296875,
0.7255859375
],
[
0.509765625,
0.7080078125,
0.87890625,
0.7080078125,
0.87890625,
0.7216796875,
0.509765625,
0.7216796875
],
[
0.11419677734375,
0.728515625,
0.486083984375,
0.728515625,
0.486083984375,
0.7421875,
0.11419677734375,
0.7421875
],
[
0.5087890625,
0.724609375,
0.87890625,
0.724609375,
0.87890625,
0.73828125,
0.5087890625,
0.73828125
],
[
0.1158447265625,
0.74462890625,
0.48779296875,
0.74462890625,
0.48779296875,
0.75830078125,
0.1158447265625,
0.75830078125
],
[
0.509765625,
0.7421875,
0.8818359375,
0.7421875,
0.8818359375,
0.755859375,
0.509765625,
0.755859375
],
[
0.1175537109375,
0.7626953125,
0.489501953125,
0.7626953125,
0.489501953125,
0.775390625,
0.1175537109375,
0.775390625
],
[
0.509765625,
0.75830078125,
0.88330078125,
0.75830078125,
0.88330078125,
0.771484375,
0.509765625,
0.771484375
],
[
0.51123046875,
0.7724609375,
0.880859375,
0.7724609375,
0.880859375,
0.78759765625,
0.51123046875,
0.78759765625
],
[
0.1158447265625,
0.779296875,
0.295654296875,
0.779296875,
0.295654296875,
0.7919921875,
0.1158447265625,
0.7919921875
],
[
0.51318359375,
0.79052734375,
0.8818359375,
0.79052734375,
0.8818359375,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.1351318359375,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.1351318359375,
0.82080078125
],
[
0.51123046875,
0.8056640625,
0.8818359375,
0.8046875,
0.8818359375,
0.8193359375,
0.51123046875,
0.82080078125
],
[
0.1158447265625,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.509765625,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.509765625,
0.8359375
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.509765625,
0.83837890625,
0.88330078125,
0.83837890625,
0.88330078125,
0.8515625,
0.509765625,
0.8515625
],
[
0.11419677734375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.11419677734375,
0.87060546875
],
[
0.5087890625,
0.8525390625,
0.88330078125,
0.85400390625,
0.88330078125,
0.8681640625,
0.5087890625,
0.86767578125
],
[
0.1175537109375,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.87060546875,
0.6650390625,
0.87060546875,
0.6650390625,
0.88427734375,
0.509765625,
0.88427734375
],
[
0.11419677734375,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.11419677734375,
0.90185546875
],
[
0.5283203125,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90185546875,
0.5283203125,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.491943359375,
0.90478515625,
0.491943359375,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.5087890625,
0.91845703125
]
] |
[
"is one of Choose, Compare and Query, we know",
"tected by a POS tagger in the question exist with",
"that the input to this last module is either a single",
"the OwL module. Once we detect an object that",
"attention map or two attention maps, where each",
"does not exist, the predicted answer will be no.",
"attentionmapessentiallytriesto captureanobject",
"If all objects exist, then we generate correspond-",
"matching some textual descriptions. By tracing",
"ing bounding boxes leveraging either OWL or MDETR",
"the path in the layout leading to the attention map,",
"following the method described in the paragraph",
"wechooseeitherthezero-shotOwLmodule(when",
"above. For attribute existence, we generate a pair",
"the path has a length of 1) or the zero-shot MDETR",
"of a positive and a negative descriptions: (ATTR,",
"module (when the path is longer than 1 hop). This",
"not ATTR), e.g., (red, not red). We then find which",
"is because when the path length equals to one, it",
"description aligns better with the cropped image",
"involves only object detection (corresponding to a",
"according to b. If the image aligns better with",
"single Find module in the NMN layout for genera-",
"the positive statement, then the answer will be yes;",
"tion of the attention map). When the path length is ",
"otherwise, no. For relation existence, we generate",
"more than one, it indicates the generation of the at-",
"the masked image fin according to bi and b2 (the",
"tention map in the NMN layout involves other mod-",
"bounding boxes of the two objects whose relation is",
"ules such as Filter and Relocate, which calls for",
"tobechecked)andapairofoppositestatementsre",
"the other abilities than object detection, such as",
"garding the relation to be checked, following (Song",
"languageunderstanding.attribute recognition and",
"et al., 2022). For example, if the question is to",
"relational reasoning. Different from NMN modules",
"check whether A is holding B, the two opposite",
"which takes in image features and object embed-",
"statements will be A is holding B and A is not hold-",
"dings to generate an attention map, our zero-shot",
"ing B. For both attribute and relation existence, we",
"OwL and zero-shot MDETR takes in the raw image",
"use zero-shot CLIP for the alignment between the",
" and raw texts to locate (OBJ for OwL and SENT for",
"input image and the statements. More details and",
"MDETR) to generate a set of detected bounding boxes",
"the work flows of existence-related questions are",
"B = [bnJn=1 together with their confident scores",
"provided in Appendix C.",
"s E RN, where bn E R4 represents the relative po-",
"If the module is Count, we directly count the",
"sition and size of the detected bounding box in the",
"number of bounding boxes in B returned either",
"image. We keep only the bounding box from either",
"from OWL or MDETR. Finally, if the last module is",
"OWL or MDETR with the highest confident score and",
" a logical AND or logical oR, we further trace to the",
"feed it to CLIP. We generate an answer by leverag-",
"inputs of this module, which should both be an",
"ing the capability of multimodal matching of CLIP.",
"Exist module. We then use the same mechanism",
"Specifically, given B, we generate an input image",
"describedaboveforExist toprocessthemodule",
"(which we refer to as Tin) by either masking regions",
"By receiving the outputs from the Exist modules,",
"not containing those detected boxes (B| = 2) or",
"logical operationswillbeappliedtodeterminethe",
"cropping the image so that only the part contain-",
"output. The deterministic logical operations can be",
"ing the box remains (|Bl = 1). If the final NMN",
"found in Appendix B.",
"module is Choose, we generate a masked template",
"by question conversion as in (Song et al., 2022);",
"3.4 Spatial Heuristics",
" otherwise the masked template will be a simple",
"\"[MASK]\". Then we match the image Tin with the",
"As mentioned in (Subramanian et al.. 2022), CLIP",
"template where the [MASK] token is replaced by",
"is less capable of spatial reasoning. Using CLIP",
"each of the answer candidates in V. We then select",
"for answer generation may not be enough when",
"the answer that, when placed inside the template,",
"it involves spatial relation understanding. Follow-",
"ing (Subramanian et al., 2022), we define simple",
"best matches the image.",
"and general heuristics to perform certain types of",
"If the module is Exist, we trace back the path",
"spatial reasoning. Note that only when we decom-",
"leading to Exist to determine whether the module",
"pose questions explicitly can we insert the spatial",
"is asking for the existence of an object, an attribute",
"heuristics into CLIP-based answer generation be-",
"or a relation. For object existence (e.g, is there",
"cause we have the intermediate outputs from previ-",
"a car), we use the zero-shot owL module. For at-",
"ous reasoning steps.",
"tribute existence and relation existence, we first",
"First of all, given the coordinates and the size",
"verify whether all mentioned nouns (objects) de-",
" of a bounding box, we use manual rules (named"
] |
{
"arxiv_id": "2305.17369",
"language": "en",
"page_idx": 4
}
|
||
2402.17200-en-1
|
2402.17200-en-1.jpg
|
role of the compression domain in artifact reduction. To delve into the ramification of this oversight, we conduct a comprehensive investigation into these methods. Our findings highlight a pervasive enhancement bias towards the compression domain in two aspects. 1) A deficiency in discriminating between the compression and raw domains: Despite the presence of compression artifacts, these methods perceive the compression domain as more realistic than the raw domain. 2) A tendency to generate an enhancement domain more akin to the compression domain than to the raw domain: For existing methods, enhancementdomain images are more aligned with compression-domain images than with raw-domain images. Given these insights, it becomes clear that the compression domain should not be overlooked; instead, it is vital to harness its potential for enhancing image realism and quality.
Building upon these insights, we propose a simple yet effective method to mitigate the enhancement bias and enhance the quality of compressed images. On one hand, our method discerns the compression domain by employing a conditional discriminator [29] with the compressed image as a key condition. We theoretically prove that the domain discrimination in GAN-based quality enhancement can be facilitated by our improved discriminator. On the other hand, our method actively distinguishes the enhancement domain from the compression domain. This distancing is achieved by a domain-divergence regularization that promotes the disparity between the enhancement and compression domains. Through this dual strategy, our method enables the discrimination against the compression domain, and ultimately brings the enhancement domain closer to the raw domain to achieve superior image quality, as illustrated in Fig. 1. Finally, to verify the effectiveness of our proposed method, we conduct extensive experiments and affirm that our method proficiently addresses the problems highlighted in our observations, culminating in a substantial enhancement in image quality.
The main contributions of this paper are as follows.
1) We revisits the compression domain, an aspect often
overlooked by existing methods. Through this exploration, we identify a significant enhancement bias towards the compression domain, which adversely affects the performance of quality enhancement.
2) We introduce a new method aimed at mitigating this
bias. This method, characterized by its simplicity and effectiveness, is model-agnostic and yields a marked improvement in image quality without incurring any inference overhead to existing methods.
3) We conduct a series of comprehensive experiments in-
cluding a thorough set of quality evaluations, demon-
strating the effectiveness of the proposed method.
## $2. \textbf{Related Works}$
In this section, we provide an overview of existing methods developed for enhancing the quality of compressed images. Most of these methods [11,15,42,45,47,48,54] have focused on improving compression fidelity using convolutional neural networks (CNNs). For example, Dong $et$ $al.[11]$ introduced a pioneering CNN-based method for enhancing the quality of JPEG-compressed images through a shallow four-layer artifacts reduction-CNN (AR-CNN). Following this, Zhang et $al.$ [54] developed the denoising CNN (DnCNN) that combines a 20-layer deep network with cutting-edge techniques such as residual learning [17] and batch normalization [3]. Wang et al.[42] designed a 10-layer deep CNN-based auto decoder (DCAD), signifying the first CNN-based quality enhancement method for HEVC-compressed images. Then, Xing $etal.[47]$ proposed the resource-efficient blind quality enhancement (RBQE) method suitable for both JPEG and HEVC-compressed images, which uniquely incorporates a dynamic inference structure for blind yet effective quality enhancement. Most recently, the defocus-aware quality enhancement (DAQE) method[48] emerged to initially discern region-wise quality differences via defocus estimation, followed by a regionwise divide-and-conquer enhancement. However, these fidelity-oriented methods, despite their advancements, still face challenges in generating realistic textures and often produce over-smoothed images [6].
To address the challenges of fidelity-oriented methods, Ghosh et al.[12] and Kim et al.[23] have focused on enhancing the perceptual quality of compressed images. They both leverage super-resolution (SR) methods [25,43] as baselines by incorporating different GANs[13]. Specifically, Ghosh et al.[12] proposed the image enhancement GAN (IEGAN), a method grounded in the SRGAN method[25] with a standard GAN, for enhancing the quality of JPEG-compressed images. In [23], Kim et al. leveraged the enhanced SRGAN (ESRGAN) method [43] for quality enhancement by inheriting the relativistic average GAN (RaGAN) [21] and all loss functions. Recently, the RealESRGAN method [44] with a vanilla GAN regularized by the spectral normalization [31] has demonstrated its capability in enhancing real-world images that contain compression artifacts.
Unfortunately, the aforementioned quality enhancement methods neglect the compression domain for reducing artifacts, since they focus intently on aligning the enhancement domain with the raw domain, in a manner akin to the prevalent SR methods. Consequently, their enhanced images exhibit a strong resemblance to the compressed counterparts, leading to an enhancement bias towards the compression domain. We shall delve into this bias in the subsequent section.
2
|
[
[
0.07843017578125,
0.091552734375,
0.4697265625,
0.09027099609375,
0.4697265625,
0.10479736328125,
0.07843017578125,
0.1060791015625
],
[
0.50244140625,
0.09027099609375,
0.64892578125,
0.09027099609375,
0.64892578125,
0.10479736328125,
0.50244140625,
0.10479736328125
],
[
0.07843017578125,
0.106689453125,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.5009765625,
0.12060546875,
0.890625,
0.12060546875,
0.890625,
0.1351318359375,
0.5009765625,
0.1351318359375
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1483154296875,
0.50341796875,
0.1483154296875
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.1514892578125,
0.890625,
0.1514892578125,
0.890625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50244140625,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1944580078125,
0.50341796875,
0.1944580078125
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.1976318359375,
0.890625,
0.1976318359375,
0.890625,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2108154296875,
0.89208984375,
0.2108154296875,
0.89208984375,
0.2259521484375,
0.50244140625,
0.2259521484375
],
[
0.080078125,
0.2291259765625,
0.471435546875,
0.2291259765625,
0.471435546875,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2266845703125,
0.890625,
0.2266845703125,
0.890625,
0.240478515625,
0.50244140625,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.242431640625,
0.8896484375,
0.242431640625,
0.8896484375,
0.255615234375,
0.5009765625,
0.255615234375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.2568359375,
0.89208984375,
0.2568359375,
0.89208984375,
0.270751953125,
0.50244140625,
0.270751953125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.272705078125,
0.89208984375,
0.272705078125,
0.89208984375,
0.285888671875,
0.50244140625,
0.285888671875
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.287353515625,
0.8935546875,
0.287353515625,
0.8935546875,
0.300537109375,
0.50244140625,
0.300537109375
],
[
0.07843017578125,
0.302978515625,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.3017578125,
0.890625,
0.302978515625,
0.890625,
0.31640625,
0.50244140625,
0.31494140625
],
[
0.080078125,
0.3193359375,
0.325927734375,
0.3193359375,
0.325927734375,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50244140625,
0.31640625,
0.89208984375,
0.317626953125,
0.89208984375,
0.33203125,
0.50244140625,
0.330810546875
],
[
0.50341796875,
0.333251953125,
0.89208984375,
0.333251953125,
0.89208984375,
0.3466796875,
0.50341796875,
0.3466796875
],
[
0.098876953125,
0.34521484375,
0.4697265625,
0.34521484375,
0.4697265625,
0.35986328125,
0.098876953125,
0.35986328125
],
[
0.50341796875,
0.349853515625,
0.8896484375,
0.349853515625,
0.8896484375,
0.35986328125,
0.50341796875,
0.35986328125
],
[
0.08087158203125,
0.361083984375,
0.468017578125,
0.361083984375,
0.468017578125,
0.375,
0.08087158203125,
0.375
],
[
0.50244140625,
0.363037109375,
0.89208984375,
0.363037109375,
0.89208984375,
0.376953125,
0.50244140625,
0.376953125
],
[
0.080078125,
0.375732421875,
0.46728515625,
0.375732421875,
0.46728515625,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.50244140625,
0.37744140625,
0.89208984375,
0.37744140625,
0.89208984375,
0.391357421875,
0.50244140625,
0.391357421875
],
[
0.07843017578125,
0.39013671875,
0.4697265625,
0.391357421875,
0.4697265625,
0.406005859375,
0.07843017578125,
0.40478515625
],
[
0.5009765625,
0.392578125,
0.890625,
0.392578125,
0.890625,
0.406005859375,
0.5009765625,
0.406005859375
],
[
0.080078125,
0.4072265625,
0.4697265625,
0.4072265625,
0.4697265625,
0.42041015625,
0.080078125,
0.42041015625
],
[
0.50244140625,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.07843017578125,
0.421630859375,
0.4697265625,
0.421630859375,
0.4697265625,
0.43505859375,
0.07843017578125,
0.43505859375
],
[
0.50244140625,
0.423583984375,
0.890625,
0.423583984375,
0.890625,
0.4375,
0.50244140625,
0.4375
],
[
0.080078125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.5009765625,
0.4375,
0.89208984375,
0.438232421875,
0.89208984375,
0.45263671875,
0.5009765625,
0.451904296875
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.5048828125,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.46533203125,
0.5048828125,
0.46533203125
],
[
0.08087158203125,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.479248046875,
0.08087158203125,
0.479248046875
],
[
0.50341796875,
0.468505859375,
0.89208984375,
0.468505859375,
0.89208984375,
0.481689453125,
0.50341796875,
0.481689453125
],
[
0.080078125,
0.481689453125,
0.468017578125,
0.481689453125,
0.468017578125,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.50244140625,
0.484130859375,
0.89208984375,
0.484130859375,
0.89208984375,
0.49755859375,
0.50244140625,
0.49755859375
],
[
0.080078125,
0.49755859375,
0.4697265625,
0.49755859375,
0.4697265625,
0.51123046875,
0.080078125,
0.51123046875
],
[
0.50244140625,
0.5,
0.736328125,
0.5,
0.736328125,
0.51318359375,
0.50244140625,
0.51318359375
],
[
0.080078125,
0.51220703125,
0.4697265625,
0.51220703125,
0.4697265625,
0.52587890625,
0.080078125,
0.52587890625
],
[
0.521484375,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.521484375,
0.5322265625
],
[
0.080078125,
0.52783203125,
0.4697265625,
0.52783203125,
0.4697265625,
0.54150390625,
0.080078125,
0.54150390625
],
[
0.50341796875,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.080078125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.080078125,
0.556640625
],
[
0.50244140625,
0.5498046875,
0.890625,
0.5498046875,
0.890625,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.5595703125,
0.46728515625,
0.5595703125,
0.46728515625,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.50341796875,
0.56640625,
0.8896484375,
0.56640625,
0.8896484375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.08087158203125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.58544921875,
0.08087158203125,
0.58544921875
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.080078125,
0.5888671875,
0.46728515625,
0.5888671875,
0.46728515625,
0.599609375,
0.080078125,
0.599609375
],
[
0.50244140625,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.50244140625,
0.60791015625
],
[
0.08087158203125,
0.60400390625,
0.468017578125,
0.60400390625,
0.468017578125,
0.6142578125,
0.08087158203125,
0.6142578125
],
[
0.5009765625,
0.60791015625,
0.890625,
0.609375,
0.890625,
0.62353515625,
0.5009765625,
0.62255859375
],
[
0.08087158203125,
0.619140625,
0.46728515625,
0.619140625,
0.46728515625,
0.6298828125,
0.08087158203125,
0.6298828125
],
[
0.50244140625,
0.625,
0.890625,
0.625,
0.890625,
0.63818359375,
0.50244140625,
0.63818359375
],
[
0.080078125,
0.63330078125,
0.468017578125,
0.63330078125,
0.468017578125,
0.646484375,
0.080078125,
0.646484375
],
[
0.50341796875,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50341796875,
0.65380859375
],
[
0.07843017578125,
0.6484375,
0.2296142578125,
0.6484375,
0.2296142578125,
0.662109375,
0.07843017578125,
0.662109375
],
[
0.50244140625,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66845703125,
0.50244140625,
0.66845703125
],
[
0.1004638671875,
0.6748046875,
0.43701171875,
0.6748046875,
0.43701171875,
0.689453125,
0.1004638671875,
0.689453125
],
[
0.5048828125,
0.671875,
0.890625,
0.671875,
0.890625,
0.681640625,
0.5048828125,
0.681640625
],
[
0.50341796875,
0.68505859375,
0.890625,
0.68505859375,
0.890625,
0.69873046875,
0.50341796875,
0.69873046875
],
[
0.08087158203125,
0.7021484375,
0.46728515625,
0.7021484375,
0.46728515625,
0.71533203125,
0.08087158203125,
0.71533203125
],
[
0.50244140625,
0.70068359375,
0.890625,
0.70068359375,
0.890625,
0.7138671875,
0.50244140625,
0.7138671875
],
[
0.101318359375,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.101318359375,
0.73193359375
],
[
0.50341796875,
0.71728515625,
0.8896484375,
0.71728515625,
0.8896484375,
0.72802734375,
0.50341796875,
0.72802734375
],
[
0.101318359375,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.101318359375,
0.74755859375
],
[
0.50244140625,
0.73095703125,
0.890625,
0.73095703125,
0.890625,
0.744140625,
0.50244140625,
0.744140625
],
[
0.101318359375,
0.74853515625,
0.4697265625,
0.74853515625,
0.4697265625,
0.76220703125,
0.101318359375,
0.76220703125
],
[
0.50341796875,
0.74755859375,
0.5947265625,
0.74755859375,
0.5947265625,
0.7587890625,
0.50341796875,
0.7587890625
],
[
0.101318359375,
0.76318359375,
0.373291015625,
0.76318359375,
0.373291015625,
0.7763671875,
0.101318359375,
0.7763671875
],
[
0.521484375,
0.7646484375,
0.89208984375,
0.76513671875,
0.89208984375,
0.77978515625,
0.521484375,
0.77880859375
],
[
0.08087158203125,
0.77783203125,
0.468017578125,
0.77783203125,
0.468017578125,
0.79150390625,
0.08087158203125,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.101318359375,
0.79248046875,
0.46728515625,
0.79248046875,
0.46728515625,
0.80615234375,
0.101318359375,
0.80615234375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.101318359375,
0.80810546875,
0.4697265625,
0.80810546875,
0.4697265625,
0.8212890625,
0.101318359375,
0.8212890625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.101318359375,
0.82373046875,
0.468017578125,
0.82373046875,
0.468017578125,
0.8369140625,
0.101318359375,
0.8369140625
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.1004638671875,
0.8388671875,
0.3349609375,
0.8388671875,
0.3349609375,
0.85302734375,
0.1004638671875,
0.85302734375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85400390625,
0.468017578125,
0.85400390625,
0.468017578125,
0.8671875,
0.080078125,
0.8671875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.10296630859375,
0.87060546875,
0.46728515625,
0.87060546875,
0.46728515625,
0.880859375,
0.10296630859375,
0.880859375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.101318359375,
0.8837890625,
0.428955078125,
0.8837890625,
0.428955078125,
0.8984375,
0.101318359375,
0.8984375
],
[
0.50244140625,
0.88720703125,
0.5361328125,
0.88720703125,
0.5361328125,
0.8994140625,
0.50244140625,
0.8994140625
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"role of the compression domain in artifact reduction.To",
"2. Related Works",
"delve into the ramification of this oversight, we conduct a",
"comprehensive investigation into these methods. Our find-",
"In this section, we provide an overview of existing meth-",
"ings highlight a pervasive enhancement bias towards the",
"odsdevelopedforenhancingthequalityofcompressedim-",
"compression domain in two aspects.1) A deficiency in",
"ages.Most of these methods [11,15,42,45,47,48,54]",
"discriminating between the compression and raw domains:",
"have focused on improving compression fidelity using con-",
"Despite the presence of compression artifacts, these meth-",
"volutional neural networks(CNNs).Forexample,Dong et",
"odsperceivethecompressiondomainasmorerealisticthan",
"al. [11] introduced a pioneering CNN-based method for en-",
"the raw domain., 2) A tendency to generate an enhance-",
"hancing the quality of JPEG-compressed images through",
"ment domain more akin to the compression domain than",
"a shallow four-layer artifacts reduction-CNN (AR-CNN).",
"to the raw domain: For existing methods, enhancement-",
"Following this, Zhang et al. [54] developed the denois-",
"domain images are more aligned with compression-domain",
"ing CNN (DnCNN) that combines a 20-layer deep network",
"images than with raw-domain images. Given these insights,",
"with cutting-edge techniques such as residual learning [17]",
"it becomes clear that the compression domain should not",
"and batch normalization [3]. Wang et al. [42] designed a",
"be overlooked; instead, it is vital to harness its potential for",
"10-layer deep CNN-based auto decoder (DCAD), signify-",
"enhancing image realism and quality.",
"ing the first CNN-based quality enhancement method for",
"HEVC-compressed images. Then, Xing et al. [47] proposed",
"Building upon these insights, we propose a simple yet",
"theresource-efficientblindqualityenhancement(RBOE)",
"effective method to mitigate the enhancement bias and en-",
"method suitable for both JPEG and HEVC-compressed im-",
"hance the quality of compressed images. On one hand",
"ages, which uniquely incorporates a dynamic inference",
"our method discerns the compression domain by employ-",
"structure for blind yet effective quality enhancement. Most",
"ing a conditional discriminator [29] with the compressed",
"recently, the defocus-aware quality enhancement (DAQE)",
"image as a key condition. We theoretically prove that the",
"method [48] emerged to initially discern region-wise qual-",
"domain discrimination in GAN-based quality enhancement",
"ity differences via defocus estimation, followed by a region-",
"can be facilitated by our improved discriminator. On the",
"wise divide-and-conguer enhancement.However.these",
"otherhand, our method actively distinguishes the enhance-",
"fidelity-oriented methods, despite their advancements, still",
"ment domain from the compression domain. This distanc-",
"face challenges in generating realistic textures and often",
"ing is achieved by a domain-divergence regularization that",
"produce over-smoothed images [6].",
"promotes the disparity between the enhancement and com-",
"To address the challenges of fidelity-oriented methods,",
"pression domains. Through this dual strategy, our method",
"Ghosh et al. [12] and Kim et al. [23] have focused on en-",
"enables the discrimination against the compression domain,",
"hancing the perceptual quality of compressed images. They",
"andultimatelybringstheenhancementdomainclosertothe",
"both leverage super-resolution(SR)methods[25,43]as",
"raw domain to achieve superior image quality,asillustrated",
"baselines by incorporating different GANs [13]. Specif-",
"inFig.1.Finally,toverifytheeffectiveness of ourproposed",
"ically, Ghosh et al. [12] proposed the image enhance-",
"method,weconductextensiveexperiments andaffirmthat",
"ment GAN (IEGAN), a method grounded in the SRGAN",
"ourmethodproficientlyaddressestheproblemshighlighted",
"method [25] with a standard GAN, for enhancing the quality",
"in our observations, culminating in a substantial enhance-",
"of JPEG-compressed images. In [23], Kim et al. leveraged",
"ment in image quality.",
"the enhanced SRGAN (ESRGAN) method [43] for qual-",
"The main contributions of this paper are as follows.",
"ityenhancementbyinheritingtherelativisticaverageGAN",
"(RaGAN) [21] and all loss functions. Recently, the Real-",
"1) We revisits the compression domain, an aspect often",
"ESRGAN method [44] with a vanilla GAN regularized by",
"overlooked by existing methods.Through this explo-",
"the spectral normalization[31|has demonstrated its capa-",
"ration, we identify a significant enhancement bias to-",
"bility in enhancing real-world images that contain compres-",
"wards the compression domain, which adversely affects",
"sion artifacts.",
"the performance of quality enhancement.",
"Unfortunately, the aforementioned quality enhancement",
"2) We introduce a new method aimed at mitigating this",
"methods neglectthe compression domainforreducing arti-",
"bias. This method, characterized by its simplicity and",
"facts, since they focus intently on aligning the enhancement",
"effectiveness, is model-agnostic and yields a marked im-",
"domain with the raw domain, in a manner akin to the preva-",
"provement in image quality without incurring any infer-",
"lent SR methods. Consequently, their enhanced images ex-",
"ence overhead to existing methods.",
"hibit a strong resemblance to the compressed counterparts,",
"3) We conduct a series of comprehensive experiments in-",
"leading to an enhancement bias towardsthe compression",
"cluding a thorough set of quality evaluations, demon-",
"domain. We shall delve into this bias in the subsequent sec-",
"strating the effectiveness of the proposed method.",
"tion.",
"2"
] |
{
"arxiv_id": "2402.17200",
"language": "en",
"page_idx": 1
}
|
||
2402.17200-en-2
|
2402.17200-en-2.jpg
|
<table>
<tbody>
<tr>
<th>Meth. </th>
<th>$Data.$</th>
<th>Raw</th>
<th>$Enh.$</th>
<th>$\Delta$ to raw</th>
<th>Comp. </th>
<th>$\Delta$ to raw</th>
</tr>
<tr>
<td rowspan="2">[44]</td>
<td>DIV2K</td>
<td>0.76</td>
<td>0.28</td>
<td>-0.48</td>
<td>0.84</td>
<td>+0.08</td>
</tr>
<tr>
<td>${\mathrm{Flickr2K}}$</td>
<td>0.74</td>
<td>0.27</td>
<td>-0.47</td>
<td>0.84</td>
<td>+0.10</td>
</tr>
<tr>
<td rowspan="2">[43]</td>
<td>DIV2K</td>
<td>0.73</td>
<td>0.27</td>
<td>-0.46</td>
<td>0.76</td>
<td>+0.03</td>
</tr>
<tr>
<td>${\mathrm{Flickr2K}}$</td>
<td>0.74</td>
<td>0.26</td>
<td>-0.48</td>
<td>0.78</td>
<td>+0.04</td>
</tr>
</tbody>
</table>
Table 1. Discriminator-evaluated realism scores for raw, enhanced, and compressed images. Higher scores indicate greater perceived realism. The results reveal that existing methods regard compressed images as more realistic than raw images.
## 3. Enhancement Bias Towards Compression Domain
In this section, we thoroughly examine the enhancement bias towards the compression domain, a phenomenon we consistently observe across existing perception-driven quality enhancement methods for compressed images. As aforementioned, these methods predominantly leverage
We thus
perception-driven SR methods as baselines.
present the analysis results based on [43,44], given their prominence in recent perception-driven quality enhancement endeavors. Note that these methods are substantially different: The former method [43] utilizes a VGG [35]- based RaGAN[21], while the latter method [44] adopts a UNet[32]-based vanilla GAN[13] with spectral normalization regularization [31]. We re-train these methods by their default settings for quality enhancement using the images from the DIV2K training set [1]. Subsequently, these methods are evaluated using the images from the DIV2K validation and Flickr2K [39] datasets. All images are compressed by better portable graphics (BPG)[4] and JPEG [40].$^{2}$
$\hat{\textbf{Observation 1: Existing quality enhancement methods}}$ fail to discriminate against compressed images, regarding them as more realistic than raw images.
Analysis. We measure the realism scores of enhanced, compressed, and raw images evaluated by discriminators. Specifically, we crop all images into patches with the size of 128×128 (in accordance with the patch size during training), which are then sent to the discriminators for evaluating realism scores. As shown in Tab. 1, the discriminators effectively discriminate between enhanced and raw images by assigning higher realism scores to the raw images. However, these discriminators struggle to distinguish between compressed and raw images; more surprisingly, the average realism score for compressed images even exceeds that of raw images. This reveals that the compression domain is perceived as more realistic, undermining the efforts to reduce compression artifacts and leading to the enhance-
2BPG implements the HEVC intra-frame compression [37]. Results
for BPG-compressed images with quality parameter (QP) set to 37 are presented, while other results are comprehensively provided in the supplementary material.
<FigureHere>
Figure 2. Visual demonstrations of realism scores for images as evaluated by [44]. The results highlight that existing methods overlook significant compression artifacts in realism evaluation.
ment bias towards the compression domain. Fig. 2 provides visual demonstrations to further illustrate our observation that the severe compression artifacts $(e.g.$,blurring) are neglected when evaluating image realism. We argue that these discriminators predominantly base their judgment on generative artifacts; as a result, the compressed image appears more realistic since the introduced compression artifacts differ significantly from generative artifacts. This insight wraps up our analysis for Observation 1.
Observation 2: For existing methods, enhancementdomain images are more aligned with compression-domain images than with raw-domain images.
Analysis. We analyze the similarity scores among compressed, raw, and enhanced images via utilizing the datasets and re-trained models above. Here, the similarity is measured by two widely-adopted metrics for assessing perceptual quality enhancement: the Frechet inception distance (FID) [ 18] and the learned perceptual image patch similarity (LPIPS)[55]. As illustrated in Fig. 3, there is a consistent trend where the enhancement domain exhibits closer alignment with the compression domain, compared to the raw domain. To quantitatively measure this trend, we calculate the horizontal deviation of each vertex relative to the centroid of its base, which can be equivalently expressed as:
$$\mathrm{Deviation}:=\frac{S_{\mathrm{C,E}}^{2}-S_{\mathrm{R,E}}^{2}}{S_{\mathrm{C,R}}^{2}}\times100\%,$$
(1)
where $S_{\mathrm{C,E}},S_{\mathrm{C,R}}$ and $S_{\mathrm{R,E}}$ represent the similarity (in terms of either FID or LPIPS) between compression and enhancement domains, between compression and raw domains, and between raw and enhancement domains, respectively. The obtained results range from -5.71% to -98.43%, demonstrating the pervasive bias towards the compression domain, as observed consistently across datasets and similarity metrics.
3
|
[
[
0.56201171875,
0.09027099609375,
0.5888671875,
0.09027099609375,
0.5888671875,
0.1029052734375,
0.56201171875,
0.1029052734375
],
[
0.66845703125,
0.091552734375,
0.72607421875,
0.091552734375,
0.72607421875,
0.10162353515625,
0.66845703125,
0.10162353515625
],
[
0.78173828125,
0.091552734375,
0.85546875,
0.091552734375,
0.85546875,
0.1029052734375,
0.78173828125,
0.1029052734375
],
[
0.43310546875,
0.0947265625,
0.450927734375,
0.0947265625,
0.450927734375,
0.10797119140625,
0.43310546875,
0.10797119140625
],
[
0.318603515625,
0.0953369140625,
0.33251953125,
0.0953369140625,
0.33251953125,
0.1060791015625,
0.318603515625,
0.1060791015625
],
[
0.08660888671875,
0.10162353515625,
0.12335205078125,
0.10162353515625,
0.12335205078125,
0.11297607421875,
0.08660888671875,
0.11297607421875
],
[
0.1396484375,
0.10162353515625,
0.1748046875,
0.10162353515625,
0.1748046875,
0.1142578125,
0.1396484375,
0.1142578125
],
[
0.2115478515625,
0.10162353515625,
0.2442626953125,
0.10162353515625,
0.2442626953125,
0.11297607421875,
0.2115478515625,
0.11297607421875
],
[
0.256591796875,
0.10162353515625,
0.289306640625,
0.10162353515625,
0.289306640625,
0.11297607421875,
0.256591796875,
0.11297607421875
],
[
0.361083984375,
0.0997314453125,
0.404296875,
0.10162353515625,
0.40283203125,
0.11614990234375,
0.359375,
0.1142578125
],
[
0.3056640625,
0.10919189453125,
0.34716796875,
0.10919189453125,
0.34716796875,
0.12060546875,
0.3056640625,
0.12060546875
],
[
0.419189453125,
0.10797119140625,
0.463134765625,
0.10797119140625,
0.463134765625,
0.12249755859375,
0.419189453125,
0.12249755859375
],
[
0.1405029296875,
0.1287841796875,
0.1871337890625,
0.1287841796875,
0.1871337890625,
0.14013671875,
0.1405029296875,
0.14013671875
],
[
0.2099609375,
0.1287841796875,
0.24267578125,
0.1287841796875,
0.24267578125,
0.14013671875,
0.2099609375,
0.14013671875
],
[
0.256591796875,
0.1287841796875,
0.289306640625,
0.1287841796875,
0.289306640625,
0.14013671875,
0.256591796875,
0.14013671875
],
[
0.3095703125,
0.1287841796875,
0.3447265625,
0.1287841796875,
0.3447265625,
0.14013671875,
0.3095703125,
0.14013671875
],
[
0.36767578125,
0.1287841796875,
0.398681640625,
0.1287841796875,
0.398681640625,
0.14013671875,
0.36767578125,
0.14013671875
],
[
0.423095703125,
0.1287841796875,
0.461669921875,
0.1287841796875,
0.461669921875,
0.14013671875,
0.423095703125,
0.14013671875
],
[
0.1405029296875,
0.141357421875,
0.19775390625,
0.141357421875,
0.19775390625,
0.1558837890625,
0.1405029296875,
0.1558837890625
],
[
0.08416748046875,
0.1337890625,
0.1168212890625,
0.1337890625,
0.1168212890625,
0.149658203125,
0.08416748046875,
0.149658203125
],
[
0.2132568359375,
0.1427001953125,
0.24267578125,
0.1427001953125,
0.24267578125,
0.1546630859375,
0.2132568359375,
0.1546630859375
],
[
0.25732421875,
0.1427001953125,
0.289306640625,
0.1427001953125,
0.289306640625,
0.1546630859375,
0.25732421875,
0.1546630859375
],
[
0.3095703125,
0.1427001953125,
0.343017578125,
0.1427001953125,
0.343017578125,
0.1546630859375,
0.3095703125,
0.1546630859375
],
[
0.369384765625,
0.1427001953125,
0.398681640625,
0.1427001953125,
0.398681640625,
0.1546630859375,
0.369384765625,
0.1546630859375
],
[
0.42236328125,
0.141357421875,
0.463134765625,
0.141357421875,
0.463134765625,
0.1558837890625,
0.42236328125,
0.1558837890625
],
[
0.1405029296875,
0.1610107421875,
0.188720703125,
0.1610107421875,
0.188720703125,
0.175537109375,
0.1405029296875,
0.175537109375
],
[
0.25732421875,
0.1622314453125,
0.289306640625,
0.1622314453125,
0.289306640625,
0.1749267578125,
0.25732421875,
0.1749267578125
],
[
0.369384765625,
0.1622314453125,
0.398681640625,
0.1622314453125,
0.398681640625,
0.1749267578125,
0.369384765625,
0.1749267578125
],
[
0.3095703125,
0.1634521484375,
0.343017578125,
0.1634521484375,
0.343017578125,
0.1749267578125,
0.3095703125,
0.1749267578125
],
[
0.2115478515625,
0.1634521484375,
0.2418212890625,
0.1634521484375,
0.2418212890625,
0.1749267578125,
0.2115478515625,
0.1749267578125
],
[
0.42236328125,
0.1634521484375,
0.461669921875,
0.1634521484375,
0.461669921875,
0.1749267578125,
0.42236328125,
0.1749267578125
],
[
0.1405029296875,
0.1767578125,
0.1968994140625,
0.1767578125,
0.1968994140625,
0.1881103515625,
0.1405029296875,
0.1881103515625
],
[
0.0849609375,
0.1685791015625,
0.1168212890625,
0.1685791015625,
0.1168212890625,
0.184326171875,
0.0849609375,
0.184326171875
],
[
0.2115478515625,
0.1767578125,
0.24267578125,
0.1767578125,
0.24267578125,
0.1881103515625,
0.2115478515625,
0.1881103515625
],
[
0.256591796875,
0.1767578125,
0.2900390625,
0.1767578125,
0.2900390625,
0.1881103515625,
0.256591796875,
0.1881103515625
],
[
0.3095703125,
0.1767578125,
0.343017578125,
0.1767578125,
0.343017578125,
0.1881103515625,
0.3095703125,
0.1881103515625
],
[
0.369384765625,
0.1767578125,
0.398681640625,
0.1767578125,
0.398681640625,
0.189453125,
0.369384765625,
0.189453125
],
[
0.423095703125,
0.1767578125,
0.4599609375,
0.1767578125,
0.4599609375,
0.1881103515625,
0.423095703125,
0.1881103515625
],
[
0.521484375,
0.1925048828125,
0.6298828125,
0.1912841796875,
0.6298828125,
0.20263671875,
0.521484375,
0.203857421875
],
[
0.66845703125,
0.1925048828125,
0.72607421875,
0.1925048828125,
0.72607421875,
0.20263671875,
0.66845703125,
0.20263671875
],
[
0.78759765625,
0.1925048828125,
0.8505859375,
0.1925048828125,
0.8505859375,
0.203857421875,
0.78759765625,
0.203857421875
],
[
0.080078125,
0.205810546875,
0.46728515625,
0.205810546875,
0.46728515625,
0.2197265625,
0.080078125,
0.2197265625
],
[
0.080078125,
0.2215576171875,
0.468017578125,
0.2215576171875,
0.468017578125,
0.231689453125,
0.080078125,
0.231689453125
],
[
0.07843017578125,
0.2342529296875,
0.468017578125,
0.2342529296875,
0.468017578125,
0.2474365234375,
0.07843017578125,
0.2474365234375
],
[
0.08087158203125,
0.25,
0.371826171875,
0.25,
0.371826171875,
0.260009765625,
0.08087158203125,
0.260009765625
],
[
0.080078125,
0.272705078125,
0.465576171875,
0.272705078125,
0.465576171875,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.108642578125,
0.2890625,
0.178955078125,
0.291748046875,
0.178955078125,
0.30615234375,
0.108642578125,
0.3037109375
],
[
0.52294921875,
0.296630859375,
0.6298828125,
0.296630859375,
0.6298828125,
0.306884765625,
0.52294921875,
0.306884765625
],
[
0.66845703125,
0.298583984375,
0.72607421875,
0.298583984375,
0.72607421875,
0.30615234375,
0.66845703125,
0.30615234375
],
[
0.7861328125,
0.296630859375,
0.85107421875,
0.296630859375,
0.85107421875,
0.306884765625,
0.7861328125,
0.306884765625
],
[
0.080078125,
0.31494140625,
0.468017578125,
0.31494140625,
0.468017578125,
0.328857421875,
0.080078125,
0.328857421875
],
[
0.50244140625,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50244140625,
0.335205078125
],
[
0.07843017578125,
0.330810546875,
0.4697265625,
0.330810546875,
0.4697265625,
0.3447265625,
0.07843017578125,
0.3447265625
],
[
0.50341796875,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.080078125,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.50244140625,
0.349853515625,
0.8798828125,
0.349853515625,
0.8798828125,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.361083984375,
0.4697265625,
0.361083984375,
0.4697265625,
0.375,
0.080078125,
0.375
],
[
0.07843017578125,
0.375,
0.4697265625,
0.375732421875,
0.4697265625,
0.39013671875,
0.07843017578125,
0.389404296875
],
[
0.50244140625,
0.37744140625,
0.890625,
0.37744140625,
0.890625,
0.391357421875,
0.50244140625,
0.391357421875
],
[
0.07843017578125,
0.39013671875,
0.396240234375,
0.389404296875,
0.396240234375,
0.404052734375,
0.07843017578125,
0.40478515625
],
[
0.406005859375,
0.391357421875,
0.4697265625,
0.391357421875,
0.4697265625,
0.40283203125,
0.406005859375,
0.40283203125
],
[
0.50244140625,
0.393310546875,
0.89208984375,
0.393310546875,
0.89208984375,
0.4072265625,
0.50244140625,
0.4072265625
],
[
0.080078125,
0.406005859375,
0.4697265625,
0.406005859375,
0.4697265625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.50244140625,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.080078125,
0.42041015625,
0.4697265625,
0.42041015625,
0.4697265625,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.50341796875,
0.423583984375,
0.89208984375,
0.423583984375,
0.89208984375,
0.4375,
0.50341796875,
0.4375
],
[
0.080078125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.50244140625,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.50244140625,
0.453857421875,
0.89208984375,
0.453857421875,
0.89208984375,
0.46728515625,
0.50244140625,
0.46728515625
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50244140625,
0.468505859375,
0.89208984375,
0.468505859375,
0.89208984375,
0.481689453125,
0.50244140625,
0.481689453125
],
[
0.08087158203125,
0.48291015625,
0.465576171875,
0.48291015625,
0.465576171875,
0.49365234375,
0.08087158203125,
0.49365234375
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.49560546875,
0.50341796875,
0.49560546875
],
[
0.080078125,
0.496826171875,
0.4697265625,
0.496826171875,
0.4697265625,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.5009765625,
0.498779296875,
0.7685546875,
0.496826171875,
0.7685546875,
0.51025390625,
0.5009765625,
0.51220703125
],
[
0.08087158203125,
0.51220703125,
0.471435546875,
0.51220703125,
0.471435546875,
0.52587890625,
0.08087158203125,
0.52587890625
],
[
0.521484375,
0.5146484375,
0.890625,
0.5146484375,
0.890625,
0.52783203125,
0.521484375,
0.52783203125
],
[
0.080078125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.080078125,
0.54052734375
],
[
0.5009765625,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54248046875,
0.5009765625,
0.54248046875
],
[
0.08087158203125,
0.54345703125,
0.46728515625,
0.54345703125,
0.46728515625,
0.5537109375,
0.08087158203125,
0.5537109375
],
[
0.50244140625,
0.54443359375,
0.75244140625,
0.54443359375,
0.75244140625,
0.55810546875,
0.50244140625,
0.55810546875
],
[
0.08087158203125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.568359375,
0.08087158203125,
0.568359375
],
[
0.521484375,
0.5595703125,
0.890625,
0.5595703125,
0.890625,
0.57275390625,
0.521484375,
0.57275390625
],
[
0.080078125,
0.57080078125,
0.435546875,
0.57080078125,
0.435546875,
0.583984375,
0.080078125,
0.583984375
],
[
0.50341796875,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.5869140625,
0.50341796875,
0.5869140625
],
[
0.098876953125,
0.58544921875,
0.468017578125,
0.5859375,
0.468017578125,
0.60107421875,
0.098876953125,
0.599609375
],
[
0.50244140625,
0.59033203125,
0.890625,
0.59033203125,
0.890625,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.07843017578125,
0.60107421875,
0.4697265625,
0.6015625,
0.4697265625,
0.6162109375,
0.07843017578125,
0.61572265625
],
[
0.5009765625,
0.60400390625,
0.89208984375,
0.60498046875,
0.89208984375,
0.619140625,
0.5009765625,
0.61865234375
],
[
0.080078125,
0.6162109375,
0.340576171875,
0.6162109375,
0.340576171875,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.5009765625,
0.619140625,
0.890625,
0.619140625,
0.890625,
0.63330078125,
0.5009765625,
0.63330078125
],
[
0.098876953125,
0.6318359375,
0.468017578125,
0.6318359375,
0.468017578125,
0.64599609375,
0.098876953125,
0.64599609375
],
[
0.50341796875,
0.63525390625,
0.890625,
0.63525390625,
0.890625,
0.6484375,
0.50341796875,
0.6484375
],
[
0.080078125,
0.64794921875,
0.468017578125,
0.64794921875,
0.468017578125,
0.6611328125,
0.080078125,
0.6611328125
],
[
0.50341796875,
0.65087890625,
0.89208984375,
0.65087890625,
0.89208984375,
0.6640625,
0.50341796875,
0.6640625
],
[
0.080078125,
0.66357421875,
0.4697265625,
0.66357421875,
0.4697265625,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.50244140625,
0.66552734375,
0.890625,
0.66552734375,
0.890625,
0.6787109375,
0.50244140625,
0.6787109375
],
[
0.080078125,
0.67822265625,
0.468017578125,
0.67822265625,
0.468017578125,
0.69140625,
0.080078125,
0.69140625
],
[
0.50244140625,
0.68115234375,
0.89208984375,
0.68115234375,
0.89208984375,
0.6943359375,
0.50244140625,
0.6943359375
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50244140625,
0.69580078125,
0.89208984375,
0.69580078125,
0.89208984375,
0.708984375,
0.50244140625,
0.708984375
],
[
0.080078125,
0.70849609375,
0.471435546875,
0.70849609375,
0.471435546875,
0.7216796875,
0.080078125,
0.7216796875
],
[
0.50341796875,
0.7119140625,
0.8896484375,
0.7119140625,
0.8896484375,
0.72265625,
0.50341796875,
0.72265625
],
[
0.07843017578125,
0.7216796875,
0.4697265625,
0.72265625,
0.4697265625,
0.7373046875,
0.07843017578125,
0.736328125
],
[
0.50244140625,
0.72607421875,
0.89208984375,
0.72607421875,
0.89208984375,
0.7392578125,
0.50244140625,
0.7392578125
],
[
0.08087158203125,
0.7392578125,
0.46728515625,
0.7392578125,
0.46728515625,
0.75,
0.08087158203125,
0.75
],
[
0.66845703125,
0.744140625,
0.82177734375,
0.75830078125,
0.81787109375,
0.7841796875,
0.6640625,
0.77099609375
],
[
0.080078125,
0.75244140625,
0.468017578125,
0.75244140625,
0.468017578125,
0.7666015625,
0.080078125,
0.7666015625
],
[
0.57861328125,
0.76318359375,
0.66015625,
0.76318359375,
0.66015625,
0.77783203125,
0.57861328125,
0.77783203125
],
[
0.869140625,
0.76318359375,
0.8935546875,
0.76318359375,
0.8935546875,
0.7763671875,
0.869140625,
0.7763671875
],
[
0.69384765625,
0.77099609375,
0.7294921875,
0.7734375,
0.72607421875,
0.7890625,
0.69140625,
0.7861328125
],
[
0.080078125,
0.78515625,
0.468017578125,
0.78515625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.080078125,
0.76806640625,
0.46728515625,
0.76806640625,
0.46728515625,
0.7822265625,
0.080078125,
0.7822265625
],
[
0.080078125,
0.798828125,
0.4697265625,
0.798828125,
0.4697265625,
0.8125,
0.080078125,
0.8125
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81103515625,
0.50341796875,
0.81103515625
],
[
0.07843017578125,
0.814453125,
0.471435546875,
0.814453125,
0.471435546875,
0.82763671875,
0.07843017578125,
0.82763671875
],
[
0.5009765625,
0.81103515625,
0.88623046875,
0.81103515625,
0.88623046875,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.080078125,
0.8291015625,
0.4697265625,
0.8291015625,
0.4697265625,
0.84228515625,
0.080078125,
0.84228515625
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.09722900390625,
0.85107421875,
0.4697265625,
0.85107421875,
0.4697265625,
0.8642578125,
0.09722900390625,
0.8642578125
],
[
0.07843017578125,
0.86279296875,
0.471435546875,
0.86279296875,
0.471435546875,
0.876953125,
0.07843017578125,
0.876953125
],
[
0.50341796875,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.87548828125,
0.468017578125,
0.87548828125,
0.468017578125,
0.8896484375,
0.080078125,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.8896484375,
0.19287109375,
0.8896484375,
0.19287109375,
0.8994140625,
0.080078125,
0.8994140625
],
[
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,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"Raw",
"Enhanced",
"Compressed",
"△",
"A",
"Meth.",
"Data.",
"Raw",
"Enh.",
"Comp",
"to raw",
"to raw",
"DIV2K",
"0.76",
"0.28",
"-0.48",
"0.84",
"+0.08",
"Flickr2K",
"[44]",
"0.74",
"0.27",
"-0.47",
"0.84",
"+0.10",
"DIV2K",
"0.27",
"0.76",
"-0.46",
"0.73",
"+0.03",
"Flickr2K",
"[43]",
"0.74",
"0.26",
"-0.48",
"0.78",
"+0.04",
"Realismscore:0.85",
"0.28(-0.57)",
"0.87 (+0.02)",
"Table 1. Discriminator-evaluated realism scores for raw, enhanced",
"and compressed images.Higherscores indicategreater perceived",
"realism.The results reveal that existing methods regard com-",
"pressed images as morerealisticthan rawimages",
"3.Enhancement Bias Towards Compression",
"Domain",
"Realism score:0.66",
"0.13(-0.53)",
"0.85 (+0.19)",
"In this section, we thoroughly examine the enhancement",
"Figure 2. Visual demonstrations of realism scores for images as",
"bias towards the compression domain, a phenomenon we",
"evaluated by[44l.The results highlight that existing methods",
"consistently observe across existing perception-driven qual-",
"overlook significant compression artifacts in realism evaluation.",
"ity enhancement methods for compressed images.As",
"aforementioned, these methods predominantly leverage",
"ment bias towards the compression domain. Fig. 2 provides",
"perception-driven SR methods as baselines.",
"We thus",
"visual demonstrations to further illustrate our observation",
"present the analysis results based on [43, 44], given their",
"that the severe compression artifacts (e.g., blurring) are ne-",
"prominence in recent perception-driven quality enhance-",
"glected when evaluating image realism. We argue that these",
"ment endeavors. Note that these methods are substantially",
"discriminators predominantly base their judgment on gen-",
"different: The former method [43] utilizes a VGG [35]-",
"erative artifacts; as a result, the compressed image appears",
"based RaGAN [21], while the latter method [44] adopts a U-",
"more realistic since the introduced compression artifacts",
"Net[32]-based vanillaGAN[13]withspectral normaliza-",
"differ significantly from generative artifacts.This insight",
"tion regularization [31]. We re-train these methods by their",
"wraps up our analysis for Observation 1.",
"default settings for quality enhancement using the images",
"Observation 2: For existing methods, enhancement-",
"from the DIV2K training set [1]. Subsequently, these meth-",
"domain images are more aligned with compression-domain",
"odsareevaluatedusingtheimagesfromtheDIV2Kvalida-",
"images than with raw-domain images.",
"tionandFlickr2K[39datasets.Allimagesarecompressed",
"Analysis. We analyze the similarity scores among com-",
"by better portable graphics (BPG) [4] and JPEG[40]?",
"pressed,raw,and enhancedimages viautilizing the datasets",
"Observation1:Existing quality enhancement methods",
"and re-trained models above. Here, the similarity is mea-",
"fail to discriminate against compressed images, regarding",
"sured by two widely-adopted metrics for assessing percep-",
"them as more realistic than raw images.",
"tual quality enhancement: the Frechet inception distance",
"Analysis.We measure the realism scores of enhanced,",
"(FID) [18] and the learned perceptual image patch similarity",
"compressed, and raw images evaluated by discriminators.",
"(LPIPS) [55]. As illustrated in Fig. 3, there is a consistent",
"Specifically, we crop all images into patches with the size",
"trend where the enhancement domain exhibits closer align-",
"of 128 × 128 (in accordance with the patch size during train-",
"ment with the compression domain, compared to the raw",
"ing), which are then sent to the discriminators for evaluat-",
"domain. To quantitatively measure this trend, we calculate",
"ing realism scores. As shown in Tab.1, the discriminators",
"the horizontal deviation of each vertex relative to the cen-",
"effectively discriminate between enhanced and raw images",
"troid of its base, which can be equivalently expressed as:",
"byassigninghigherrealism scorestotherawimages.How-",
"SE.E - SRE × 100%,",
"ever, these discriminators struggle to distinguish between",
"Deviation :=",
"(1)",
"S2.R",
"agerealismscoreforcompressedimagesevenexceedsthat",
"compressed and raw images; more surprisingly, the aver-",
"of raw images. This reveals that the compression domain",
"where Sc,E, Sc,R and SR,E represent the similarity (in terms",
"is perceived as more realistic, undermining the efforts to",
"of either FID or LPIPS) between compression and enhance",
"reduce compression artifacts and leading to the enhance-",
"ment domains, between compression and raw domains, and",
"between raw and enhancement domains, respectively. The",
"2BPG implements the HEVC intra-frame compression [37]. Results",
"for BPG-compressed images with quality parameter (QP) set to 37 are",
"obtainedresultsrangefrom-5.71%to-98.43%,demonstrat-",
"presented, while other results are comprehensively provided in the sup-",
"ing the pervasive bias towards the compression domain, as",
"plementarymaterial.",
"observed consistently across datasets and similarity metrics.",
"3"
] |
{
"arxiv_id": "2402.17200",
"language": "en",
"page_idx": 2
}
|
||
2005.00673-en-1
|
2005.00673-en-1.jpg
|
that vehicle-based ReID is more difficult. Although license plates could potentially be useful to identify each vehicle, they often cannot be read from traffic cameras due to occlusion, oblique viewpoint, or low image resolution, and they present privacy concerns.
Recent methods to vehicle ReID exploit feature learning [37,45,46] and/or distance metric learning [3,10,13] to train deep neural networks (DNNs) to distinguish between vehicle pairs, but the current state-of-the-art performance is still far from its counterpart in person ReID [43]. Moreover, it has been shown [30] that directly using state-ofthe-art person ReID methods for vehicles does not close this gap, indicating fundamental differences between the two tasks. We believe the key to vehicle ReID is to exploit viewpoint-invariant information such as color, type, and deformable shape models encoding pose. To jointly learn these attributes along with pose information, we propose to use synthetic data to overcome the prohibitive cost of manually labeling real training images with such detailed information.
In this work, we propose a novel framework named PAMTRI, for Pose-Aware Multi-Task Re-Identification. Our major contribution is threefold:
1. PAMTRI embeds keypoints, heatmaps and segments
from pose estimation into the multi-task learning pipeline for vehicle ReID, which guides the network to pay attention to viewpoint-related information.
2. PAMTRI is trained with large-scale synthetic data that
include randomized vehicle models, color and orientation under different backgrounds, lighting conditions and occlusion. Annotations of vehicle identity, color, type and 2D pose are automatically generated for train- $ing.$
3. Our proposed method achieves significant improve-
ment over the state-of-the-art on two mainstream benchmarks: VeRi[14] and CityFlow-ReID [30]. Additional experiments validate that our unique architecture exploiting explicit pose information, along with our use of randomized synthetic data for training, are key to the method's success.
## $2. \textbf{ Related work}$
Vehicle ReID. Among the earliest attempts for vehicle ReID that involve deep learning, Liu et $al.[\hat{14},15]$ propose a progressive framework that employs a Siamese neural network with contrastive loss for training, and they also introduced VeRi[14] as the first large-scale benchmark specifically for vehicle ReID. Bai $etal.[3]$ and Kumar $etal.[10]$ also take advantage of distance metric learning by extending the success of triplet embedding in person ReID [6] to the vehicle-based task. Especially, the batch-sampling
variant from Kumar et al. is the current state-of-the-art on both VeRi and CityFlow-ReID [30], the latter being a subset of a recent multi-target multi-camera vehicle tracking benchmark. On the other hand, some methods focus on exploiting viewpoint-invariant features, $e.g.$, the approach by Wang et al.[37] that embeds local region features from extracted vehicle keypoints for training with cross-entropy loss. Similarly, Zhou $etal.[45,46]$ use a generative adversarial network (GAN) to generate multi-view features to be selected by a viewpoint-aware attention model, in which attribute classification is also trained through the discriminative net. In addition, Yan et $al.[41]$ apply multi-task learning to address multi-grain ranking and attribute classification simultaneously, but the search for visually similar vehicles is different from our goal of ReID. To our knowledge, none of the methods jointly embody pose information and multi-task learning to address vehicle ReID
Vehicle pose estimation. Vehicle pose estimation via deformable $(i.e.$, keypoint-based) modeling is a promising approach to deal with viewpoint information. In [31], Tang $etal.$ propose to use a 16-keypoint-based car model generated from evolutionary optimization to build multiple kernels for 3D tracking. Ansari et $al.[2]$ designed a more complex vehicle model with 36 keypoints for 3D localization and shape estimation from a dash camera. The ReID method by $\hat{\text{Wang et al.[37] also employs a20-keypoint}}$ model to extract orientation-based features for region proposal. However, instead of explicitly locating keypoint coordinates, their network is trained for estimating response maps only, and semantic attributes are not exploited in their framework. Other methods can directly regress to the car pose with 6 degrees of freedom (DoF) [11,16,18,39],but they are limited for our purposes as detailed vehicle shape modeling via keypoints is not provided.
Synthetic data. To generate sufficiently detailed labels on training images, our approach leverages synthetic data. Our method is trained on a mix of rendered and real images. This places our work in the context of other research on using simulated data to train DNNs. A popular approach to overcome the so-called reality gap is domain randomization [34, 35], in which a model is trained with extreme visual variety so that when presented with a real-world image the model treats it as just another variation. Synthetic data have been successfully applied to a variety of problems, such as optical flow [17], car detection [22], object pose estimation [26, 36], vision-based robotic manipulation [8,34],and robotic control [4,29]. We extend this research to ReID and semantic attribute understanding.
3. Methodology
In this section, we describe the algorithmic design of our proposed PAMTRI framework. An overview flow diagram of the proposed system is presented in Fig. 2.
|
[
[
0.080078125,
0.0921630859375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
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,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5009765625,
0.136962890625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.154052734375,
0.2459716796875,
0.15283203125,
0.2459716796875,
0.1666259765625,
0.080078125,
0.1673583984375
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.098876953125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.098876953125,
0.18115234375
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.1817626953125,
0.8896484375,
0.1817626953125,
0.8896484375,
0.1956787109375,
0.5009765625,
0.1956787109375
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.1976318359375,
0.890625,
0.1976318359375,
0.890625,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50341796875,
0.21337890625,
0.8935546875,
0.21337890625,
0.8935546875,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.5009765625,
0.2266845703125,
0.89208984375,
0.2279052734375,
0.89208984375,
0.2412109375,
0.5009765625,
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
],
[
0.07843017578125,
0.25830078125,
0.4697265625,
0.2568359375,
0.4697265625,
0.271484375,
0.07843017578125,
0.272705078125
],
[
0.5009765625,
0.2568359375,
0.89208984375,
0.25830078125,
0.89208984375,
0.272705078125,
0.5009765625,
0.271484375
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.07843017578125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.07843017578125,
0.302978515625
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.07843017578125,
0.302978515625,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.07843017578125,
0.317626953125,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.07843017578125,
0.330810546875
],
[
0.50341796875,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.07843017578125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.07843017578125,
0.347900390625
],
[
0.50244140625,
0.333984375,
0.81884765625,
0.333984375,
0.81884765625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.521484375,
0.347900390625,
0.89208984375,
0.348388671875,
0.89208984375,
0.363037109375,
0.521484375,
0.3623046875
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.5009765625,
0.3623046875,
0.89208984375,
0.363037109375,
0.89208984375,
0.37744140625,
0.5009765625,
0.376953125
],
[
0.080078125,
0.380615234375,
0.1641845703125,
0.380615234375,
0.1641845703125,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.50244140625,
0.380126953125,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.096435546875,
0.393310546875,
0.4697265625,
0.392578125,
0.4697265625,
0.4072265625,
0.096435546875,
0.40771484375
],
[
0.50244140625,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4248046875,
0.316162109375,
0.4248046875,
0.316162109375,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.50244140625,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.09234619140625,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.09234619140625,
0.46533203125
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.11273193359375,
0.46728515625,
0.468017578125,
0.46728515625,
0.468017578125,
0.48095703125,
0.11273193359375,
0.48095703125
],
[
0.50244140625,
0.47021484375,
0.89208984375,
0.47021484375,
0.89208984375,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.11273193359375,
0.481689453125,
0.4697265625,
0.481689453125,
0.4697265625,
0.49560546875,
0.11273193359375,
0.49560546875
],
[
0.50244140625,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.111083984375,
0.496826171875,
0.438720703125,
0.496826171875,
0.438720703125,
0.51025390625,
0.111083984375,
0.51025390625
],
[
0.5009765625,
0.50048828125,
0.8896484375,
0.50048828125,
0.8896484375,
0.5146484375,
0.5009765625,
0.5146484375
],
[
0.09234619140625,
0.521484375,
0.4697265625,
0.521484375,
0.4697265625,
0.53515625,
0.09234619140625,
0.53515625
],
[
0.5048828125,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.52783203125,
0.5048828125,
0.52783203125
],
[
0.5009765625,
0.52978515625,
0.89208984375,
0.52880859375,
0.89208984375,
0.54345703125,
0.5009765625,
0.54443359375
],
[
0.11273193359375,
0.53857421875,
0.468017578125,
0.53857421875,
0.468017578125,
0.548828125,
0.11273193359375,
0.548828125
],
[
0.50244140625,
0.54541015625,
0.89208984375,
0.54541015625,
0.89208984375,
0.5595703125,
0.50244140625,
0.5595703125
],
[
0.11358642578125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.564453125,
0.11358642578125,
0.564453125
],
[
0.50341796875,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.111083984375,
0.56640625,
0.46728515625,
0.56640625,
0.46728515625,
0.580078125,
0.111083984375,
0.580078125
],
[
0.5009765625,
0.57470703125,
0.890625,
0.57470703125,
0.890625,
0.58837890625,
0.5009765625,
0.58837890625
],
[
0.11273193359375,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.59521484375,
0.11273193359375,
0.59521484375
],
[
0.50244140625,
0.59033203125,
0.7646484375,
0.59033203125,
0.7646484375,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.11273193359375,
0.5986328125,
0.1396484375,
0.5986328125,
0.1396484375,
0.611328125,
0.11273193359375,
0.611328125
],
[
0.521484375,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.619140625,
0.521484375,
0.619140625
],
[
0.09234619140625,
0.62060546875,
0.46728515625,
0.62060546875,
0.46728515625,
0.63525390625,
0.09234619140625,
0.63525390625
],
[
0.50341796875,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63525390625,
0.50341796875,
0.63525390625
],
[
0.11273193359375,
0.63720703125,
0.4697265625,
0.63720703125,
0.4697265625,
0.65087890625,
0.11273193359375,
0.65087890625
],
[
0.50341796875,
0.63623046875,
0.890625,
0.63623046875,
0.890625,
0.6494140625,
0.50341796875,
0.6494140625
],
[
0.11358642578125,
0.65380859375,
0.46728515625,
0.65380859375,
0.46728515625,
0.6640625,
0.11358642578125,
0.6640625
],
[
0.5048828125,
0.65283203125,
0.890625,
0.65283203125,
0.890625,
0.66357421875,
0.5048828125,
0.66357421875
],
[
0.111083984375,
0.66650390625,
0.46728515625,
0.66650390625,
0.46728515625,
0.6796875,
0.111083984375,
0.6796875
],
[
0.5009765625,
0.66552734375,
0.89208984375,
0.66650390625,
0.89208984375,
0.68115234375,
0.5009765625,
0.6796875
],
[
0.111083984375,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.111083984375,
0.69580078125
],
[
0.50244140625,
0.681640625,
0.890625,
0.681640625,
0.890625,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.111083984375,
0.6962890625,
0.468017578125,
0.6962890625,
0.468017578125,
0.71044921875,
0.111083984375,
0.71044921875
],
[
0.50244140625,
0.6962890625,
0.89208984375,
0.6962890625,
0.89208984375,
0.71044921875,
0.50244140625,
0.71044921875
],
[
0.111083984375,
0.71337890625,
0.301513671875,
0.71337890625,
0.301513671875,
0.7265625,
0.111083984375,
0.7265625
],
[
0.5048828125,
0.71337890625,
0.890625,
0.71337890625,
0.890625,
0.7236328125,
0.5048828125,
0.7236328125
],
[
0.50341796875,
0.7265625,
0.89208984375,
0.7265625,
0.89208984375,
0.74072265625,
0.50341796875,
0.74072265625
],
[
0.080078125,
0.74072265625,
0.214111328125,
0.74072265625,
0.214111328125,
0.7548828125,
0.080078125,
0.7548828125
],
[
0.50244140625,
0.7412109375,
0.89208984375,
0.7412109375,
0.89208984375,
0.755859375,
0.50244140625,
0.755859375
],
[
0.50244140625,
0.7568359375,
0.89208984375,
0.7568359375,
0.89208984375,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.1004638671875,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.1004638671875,
0.77880859375
],
[
0.50244140625,
0.771484375,
0.89208984375,
0.771484375,
0.89208984375,
0.78515625,
0.50244140625,
0.78515625
],
[
0.07843017578125,
0.77978515625,
0.4697265625,
0.78076171875,
0.4697265625,
0.79541015625,
0.07843017578125,
0.79443359375
],
[
0.5048828125,
0.78857421875,
0.890625,
0.78857421875,
0.890625,
0.798828125,
0.5048828125,
0.798828125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5,
0.80029296875,
0.8037109375,
0.8017578125,
0.8037109375,
0.81640625,
0.5,
0.81494140625
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.83056640625,
0.63232421875,
0.83056640625,
0.63232421875,
0.84521484375,
0.50244140625,
0.84521484375
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.465576171875,
0.85791015625,
0.465576171875,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.521484375,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.521484375,
0.869140625
],
[
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,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.7998046875,
0.88720703125,
0.7998046875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"that vehicle-based ReID is more difficult. Although license",
"variant from Kumar et al. is the current state-of-the-art on",
"plates could potentially be useful to identify each vehicle,",
"both VeRi and CityFlow-ReID [30], the latter being a sub-",
"theyoftencannotbereadfromtrafficcamerasduetoocclu-",
"set of a recent multi-target multi-camera vehicle tracking",
"sion, oblique viewpoint, or low image resolution, and they",
"benchmark.Onthe otherhand, somemethods focus on",
"present privacy concerns.",
"exploiting viewpoint-invariant features, e.g., the approach",
"Recent methods to vehicle RelD exploit feature learn-",
"by Wang et al. [37] that embeds local region features from",
"ing [37, 45, 46] and/or distance metric learning [3, 10, 13] to",
"extracted vehicle keypoints for training with cross-entropy",
"traindeepneuralnetworks(DNNs)todistinguishbetween",
"loss. Similarly, Zhou et al.[45, 46] use a generative ad-",
"vehicle pairs, but the current state-of-the-art performance is",
"versarial network (GAN) to generate multi-view features to",
"still far from its counterpart in person ReiD [43]. More-",
"be selected by a viewpoint-aware attention model, in which",
"over, it has been shown [30] that directly using state-of-",
"attribute classification is also trained through the discrim-",
"the-art person ReID methods for vehicles does not close",
"inative net. In addition, Yan et al. [41] apply multi-task",
"this gap, indicating fundamental differences between the",
"learning to address multi-grain ranking and attribute classi-",
"two tasks. We believe the key to vehicle RelD is to ex-",
"fication simultaneously, but the search for visually similar",
"ploit viewpoint-invariant information such as color, type,",
"vehicles is different from our goal of ReID. To our knowl-",
"anddeformableshapemodels encodingpose.Tojointly",
"edgenoneofthemethodsjointlyembodyposeinformation",
"learn these attributes along with pose information, we pro-",
"and multi-task learning to address vehicle RelD.",
"posetouse syntheticdatatoovercometheprohibitivecost",
"Vehicle pose estimation.Vehicle pose estimation via",
"of manually labeling real training images with such detailed",
"deformable (i.e., keypoint-based) modeling is a promising",
"information.",
"approach to deal with viewpoint information. In [31], Tang",
"In this work,we propose a novel framework named",
"et al. propose to use a 16-keypoint-based car model gener-",
"PAMTRI, for Pose-Aware Multi-Task Re-Identification.",
"ated from evolutionary optimization to build multiple ker-",
"Our major contribution is threefold:",
"nels for 3D tracking. Ansari et al. [2] designed a more",
"complex vehicle model with 36 keypoints for 3D localiza-",
"1. PAMTRI embeds keypoints, heatmaps and segments",
"tion and shape estimation from a dash camera. The ReD",
"from pose estimation into the multi-task learning",
"method by Wang et al. [371 also employs a 20-keypoint",
"pipeline for vehicle RelD, which guides the network",
"model to extract orientation-based features for region pro-",
"to pay attention to viewpoint-related information.",
"posal. However, instead of explicitly locating keypoint co-",
"2. PAMTRI is trained with large-scale synthetic data that",
"ordinates,theirnetworkistrainedforestimatingresponse",
"maps only, and semantic attributes are not exploited in their",
"includerandomizedvehiclemodels.colorandorienta-",
"framework. Other methods can directly regress to the car",
"tion underdifferentbackgrounds,lighting conditions",
"pose with 6 degrees of freedom (DoF) [11, 16, 18, 39], but",
"and occlusion. Annotations of vehicle identity, color.",
"they are limited for our purposes as detailed vehicle shape",
"type and 2D pose are automatically generated for train-",
"modeling via keypoints is not provided.",
"ing.",
"Synthetic data. To generate sufficiently detailed labels",
"3. Our proposed method achieves significant improve-",
"on training images, our approach leverages synthetic data.",
"ment over the state-of-the-art on two mainstream",
"Our method is trained on a mix of rendered and real im-",
"benchmarks:VeRi[14]andCityFlow-ReID[30l.Ad-",
"ages.Thisplacesourworkinthecontextofotherresearch",
"ditional experiments validate that our unique architec-",
"on using simulated data to train DNNs. A popular approach",
"ture exploiting explicit pose information, along with",
"to overcome the so-called reality gap is domain random-",
"our use of randomized synthetic data for training, are",
"ization [34, 35], in which a model is trained with extreme",
"key to the method's success.",
"visualvariety sothatwhenpresentedwithareal-worldim-",
"age the model treats it as just another variation. Synthetic",
"2. Related work",
"data have been successfully applied to a variety of prob-",
"lems, such as optical flow [17], car detection [22], object",
"Vehicle RelD. Among the earliest attempts for vehicle",
"pose estimation [26, 36], vision-based robotic manipulation",
"ReID that involve deep learning, Liu et al. [14, 15] propose",
"[8,34].androboticcontrol [4.29].Weextendthisresearch",
"a progressive framework that employs a Siamese neural net-",
"to RelD and semantic attribute understanding.",
"work with contrastive loss for training, and they also intro-",
"duced VeRi [14] as the first large-scale benchmark specifi-",
"3.Methodology",
"cally for vehicle ReID. Bai et al. [3] and Kumar et al. [l0]",
"also take advantage of distance metric learningbyextend-",
"In this section, we describe the algorithmic design of our",
"ing the success of triplet embedding in person RelD [6]",
"proposed PAMTRI framework. An overview flow diagram",
"to the vehicle-based task. Especially, the batch-sampling",
"of the proposed system is presented in Fig. 2."
] |
{
"arxiv_id": "2005.00673",
"language": "en",
"page_idx": 1
}
|
||
2005.00673-en-2
|
2005.00673-en-2.jpg
|
<FigureHere>
Figure 2. Overview of the proposed method. Each training batch includes both real and synthesized images. To embed pose information for multi-task learning, the heatmaps or segments output by a pre-trained network are stacked with the original RGB channels as input. The estimated keypoint coordinates and confidence scores are also concatenated with deep learning features for ReID and attribute (color and type) classification. The pose estimation network (top, blue) is based on HRNet [25], while the multi-task learming network (bottom, orange) is based on DenseNet121[7]. Best viewed in color.
## 3.1. Randomized synthetic dataset
## $3. 2. \textbf{ Vehicle pose estimation}$
Besides vehicle identities, our approach requires additional labels of vehicle attributes and locations of keypoints. These values, particularly the keypoints, would require considerable, even prohibitive effort, if annotated manually. To overcome this problem, we generated a large-scale synthetic dataset by employing our deep learning dataset synthesizer ( NDDS) [ 33] to create a randomized environment in Unreal Engine 4 (UE4), into which 3D vehicle meshes from [22] were imported. We added to NDDS the ability to label and export specific 3D locations, $i.e.$, keypoints (denoted as sockets in UE4), on a CAD model. As such we manually annotated each vehicle model with the 36 3D keypoints defined by Ansari $etal.[2];$ the projected 2D locations were then output with the synthesized images. For randomization we used 42 vehicle 3D CAD models with 10 body colors. To train the data for ReID, we define a unique identity for each combination of vehicle model with a particular color. The final generated dataset consists of 41,000 unique images with 402 identities,$^{1}$ including the following annotations: keypoints, orientation, and vehicle attributes (color and type). When generating the dataset, background images were sampled from CityFlow [30], and we also randomized the vehicle position and intensity of light. Additionally, during training we perform randomized post-processing such as scaling, cropping, horizontal flipping, and adding occlusion. Some examples are shown in Fig. 3.
To leverage viewpoint-aware information for multi-task learning, we train a robust DNN for extracting pose-related representations. Similar to Tremblay et $al.[35]$ we mix real and synthetic data to bridge the reality gap. More specifically, in each dataset, we utilize the pre-trained model[2] to process sampled images and manually select about 10,000 successful samples as real training data.
Instead of using the stacked hourglass network[21] as the backbone like previous approaches [2,37],we modify the state-of-the-art DNN for human pose estimation, HRNet[25], for our purposes. Compared to the stacked hourglass architecture and other methods that recover highresolution representations from low-resolution representations, HRNet maintains high-resolution representations and gradually add high-to-low resolution sub-networks with multi-scale fusion. As a result, the predicted keypoints and heatmaps are more accurate and spatially more precise, which benefit our embedding for multi-task learning.
We propose two ways to embed the vehicle pose information as additional channels at the input layer of the multi-task network, based on heatmaps and segments, respectively. In one approach, after the final deconvolutional layer, we extract the 36 heatmaps for each of the keypoints used to capture the vehicle shape and pose. In the other approach, the predicted keypoint coordinates from the final fully-connected (FC) layer are used to segment the vehicle body. For example, in Fig. 3 the keypoints #16,#17,#35 and #34 from the deformable model form a segment that represents the car hood. Accordingly, we define 13 segmentation masks for each vehicle (see Fig. 3 TOP-LEFT), where those formed by keypoint(s) with low confidence are
l The concrete mixer truck and the school bus did not get color variation and as such we exported 500 unique images for each of them. 100 images were generated for each of the remaining identities.
|
[
[
0.2017822265625,
0.09027099609375,
0.236083984375,
0.09027099609375,
0.236083984375,
0.09783935546875,
0.2017822265625,
0.09783935546875
],
[
0.67236328125,
0.0953369140625,
0.71728515625,
0.0953369140625,
0.71728515625,
0.1060791015625,
0.67236328125,
0.1060791015625
],
[
0.408447265625,
0.10479736328125,
0.496826171875,
0.10479736328125,
0.496826171875,
0.11553955078125,
0.408447265625,
0.11553955078125
],
[
0.63232421875,
0.10479736328125,
0.64892578125,
0.10479736328125,
0.64892578125,
0.11236572265625,
0.63232421875,
0.11236572265625
],
[
0.67236328125,
0.10479736328125,
0.73779296875,
0.10479736328125,
0.73779296875,
0.11553955078125,
0.67236328125,
0.11553955078125
],
[
0.6708984375,
0.1142578125,
0.736328125,
0.11236572265625,
0.736328125,
0.12371826171875,
0.6708984375,
0.1256103515625
],
[
0.67236328125,
0.12371826171875,
0.71142578125,
0.12371826171875,
0.71142578125,
0.1351318359375,
0.67236328125,
0.1351318359375
],
[
0.439453125,
0.136962890625,
0.5400390625,
0.136962890625,
0.5400390625,
0.1483154296875,
0.439453125,
0.1483154296875
],
[
0.468017578125,
0.149658203125,
0.5087890625,
0.149658203125,
0.5087890625,
0.160400390625,
0.468017578125,
0.160400390625
],
[
0.701171875,
0.146484375,
0.72119140625,
0.146484375,
0.72119140625,
0.1546630859375,
0.701171875,
0.1546630859375
],
[
0.775390625,
0.15087890625,
0.78759765625,
0.15087890625,
0.78759765625,
0.1546630859375,
0.775390625,
0.1546630859375
],
[
0.56640625,
0.1673583984375,
0.63232421875,
0.169189453125,
0.63232421875,
0.18115234375,
0.56640625,
0.1787109375
],
[
0.58837890625,
0.184326171875,
0.62841796875,
0.1875,
0.62744140625,
0.1988525390625,
0.58642578125,
0.1956787109375
],
[
0.70751953125,
0.1849365234375,
0.7509765625,
0.1875,
0.7490234375,
0.1988525390625,
0.7060546875,
0.1962890625
],
[
0.7197265625,
0.1962890625,
0.80224609375,
0.1962890625,
0.80224609375,
0.2083740234375,
0.7197265625,
0.2083740234375
],
[
0.414306640625,
0.189453125,
0.52685546875,
0.189453125,
0.52685546875,
0.199462890625,
0.414306640625,
0.199462890625
],
[
0.2034912109375,
0.2008056640625,
0.2344970703125,
0.2008056640625,
0.2344970703125,
0.208984375,
0.2034912109375,
0.208984375
],
[
0.32275390625,
0.208984375,
0.346435546875,
0.208984375,
0.346435546875,
0.216552734375,
0.32275390625,
0.216552734375
],
[
0.70751953125,
0.2203369140625,
0.74853515625,
0.2203369140625,
0.74853515625,
0.231689453125,
0.70751953125,
0.231689453125
],
[
0.443603515625,
0.236083984375,
0.5458984375,
0.236083984375,
0.5458984375,
0.246826171875,
0.443603515625,
0.246826171875
],
[
0.72119140625,
0.2342529296875,
0.80126953125,
0.2342529296875,
0.80126953125,
0.2442626953125,
0.72119140625,
0.2442626953125
],
[
0.473876953125,
0.2474365234375,
0.515625,
0.2474365234375,
0.515625,
0.2587890625,
0.473876953125,
0.2587890625
],
[
0.72119140625,
0.242431640625,
0.8037109375,
0.24365234375,
0.8037109375,
0.255126953125,
0.72119140625,
0.25390625
],
[
0.708984375,
0.2587890625,
0.74658203125,
0.2587890625,
0.74658203125,
0.270751953125,
0.708984375,
0.270751953125
],
[
0.64892578125,
0.266357421875,
0.6689453125,
0.266357421875,
0.6689453125,
0.2734375,
0.64892578125,
0.2734375
],
[
0.7197265625,
0.26953125,
0.80224609375,
0.26953125,
0.80224609375,
0.281005859375,
0.7197265625,
0.281005859375
],
[
0.32177734375,
0.270751953125,
0.3505859375,
0.270751953125,
0.3505859375,
0.27783203125,
0.32177734375,
0.27783203125
],
[
0.2099609375,
0.2734375,
0.223876953125,
0.2734375,
0.223876953125,
0.27783203125,
0.2099609375,
0.27783203125
],
[
0.3203125,
0.27978515625,
0.338134765625,
0.27978515625,
0.338134765625,
0.287353515625,
0.3203125,
0.287353515625
],
[
0.64794921875,
0.274658203125,
0.68701171875,
0.274658203125,
0.68701171875,
0.285400390625,
0.64794921875,
0.285400390625
],
[
0.08087158203125,
0.3125,
0.890625,
0.3125,
0.890625,
0.322509765625,
0.08087158203125,
0.322509765625
],
[
0.080078125,
0.324462890625,
0.89208984375,
0.324462890625,
0.89208984375,
0.33837890625,
0.080078125,
0.33837890625
],
[
0.080078125,
0.33837890625,
0.89208984375,
0.33837890625,
0.89208984375,
0.3515625,
0.080078125,
0.3515625
],
[
0.07843017578125,
0.352783203125,
0.89208984375,
0.352783203125,
0.89208984375,
0.3662109375,
0.07843017578125,
0.3662109375
],
[
0.080078125,
0.36865234375,
0.435546875,
0.36865234375,
0.435546875,
0.37890625,
0.080078125,
0.37890625
],
[
0.08087158203125,
0.406005859375,
0.343017578125,
0.406005859375,
0.343017578125,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50341796875,
0.40478515625,
0.71337890625,
0.40478515625,
0.71337890625,
0.419189453125,
0.50341796875,
0.419189453125
],
[
0.098876953125,
0.43505859375,
0.46728515625,
0.43505859375,
0.46728515625,
0.448974609375,
0.098876953125,
0.448974609375
],
[
0.521484375,
0.429931640625,
0.89208984375,
0.429931640625,
0.89208984375,
0.4443359375,
0.521484375,
0.4443359375
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.5009765625,
0.44580078125,
0.890625,
0.44580078125,
0.890625,
0.458984375,
0.5009765625,
0.458984375
],
[
0.080078125,
0.464111328125,
0.468017578125,
0.46533203125,
0.468017578125,
0.479736328125,
0.080078125,
0.479248046875
],
[
0.5009765625,
0.4609375,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.5009765625,
0.474853515625
],
[
0.080078125,
0.48095703125,
0.4697265625,
0.48095703125,
0.4697265625,
0.494384765625,
0.080078125,
0.494384765625
],
[
0.50244140625,
0.47607421875,
0.890625,
0.47607421875,
0.890625,
0.4892578125,
0.50244140625,
0.4892578125
],
[
0.080078125,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.50244140625,
0.4912109375,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.50244140625,
0.5048828125
],
[
0.080078125,
0.51123046875,
0.468017578125,
0.51123046875,
0.468017578125,
0.5244140625,
0.080078125,
0.5244140625
],
[
0.5009765625,
0.505859375,
0.89208984375,
0.5048828125,
0.89208984375,
0.51953125,
0.5009765625,
0.52099609375
],
[
0.08087158203125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.08087158203125,
0.54052734375
],
[
0.5009765625,
0.521484375,
0.7685546875,
0.521484375,
0.7685546875,
0.53515625,
0.5009765625,
0.53515625
],
[
0.080078125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.080078125,
0.5546875
],
[
0.521484375,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.521484375,
0.55029296875
],
[
0.080078125,
0.55615234375,
0.468017578125,
0.55615234375,
0.468017578125,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.5009765625,
0.5517578125,
0.8896484375,
0.5517578125,
0.8896484375,
0.56494140625,
0.5009765625,
0.56494140625
],
[
0.07843017578125,
0.5712890625,
0.4697265625,
0.5712890625,
0.4697265625,
0.58544921875,
0.07843017578125,
0.58544921875
],
[
0.50244140625,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50244140625,
0.58203125
],
[
0.07843017578125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.599609375,
0.07843017578125,
0.599609375
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.6015625,
0.46728515625,
0.6015625,
0.46728515625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.5009765625,
0.59765625,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.5009765625,
0.611328125
],
[
0.07843017578125,
0.6162109375,
0.468017578125,
0.61572265625,
0.468017578125,
0.6298828125,
0.07843017578125,
0.630859375
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.6318359375,
0.468017578125,
0.6318359375,
0.468017578125,
0.64599609375,
0.080078125,
0.64599609375
],
[
0.5009765625,
0.626953125,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.5009765625,
0.6416015625
],
[
0.07843017578125,
0.646484375,
0.4697265625,
0.646484375,
0.4697265625,
0.66015625,
0.07843017578125,
0.66015625
],
[
0.5009765625,
0.64306640625,
0.89208984375,
0.642578125,
0.89208984375,
0.6572265625,
0.5009765625,
0.65771484375
],
[
0.0767822265625,
0.66015625,
0.4697265625,
0.6611328125,
0.4697265625,
0.67529296875,
0.0767822265625,
0.6748046875
],
[
0.50244140625,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.08087158203125,
0.6787109375,
0.468017578125,
0.6787109375,
0.468017578125,
0.689453125,
0.08087158203125,
0.689453125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.07843017578125,
0.6923828125,
0.4697265625,
0.6923828125,
0.4697265625,
0.70556640625,
0.07843017578125,
0.70556640625
],
[
0.50244140625,
0.689453125,
0.8505859375,
0.689453125,
0.8505859375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.07843017578125,
0.70703125,
0.468017578125,
0.70849609375,
0.468017578125,
0.7216796875,
0.07843017578125,
0.72021484375
],
[
0.521484375,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.521484375,
0.7197265625
],
[
0.07843017578125,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.736328125,
0.07843017578125,
0.736328125
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.07843017578125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75048828125,
0.07843017578125,
0.75048828125
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.08087158203125,
0.75390625,
0.468017578125,
0.75390625,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.5009765625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.5009765625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.796875,
0.080078125,
0.796875
],
[
0.50244140625,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.798828125,
0.468017578125,
0.798828125,
0.468017578125,
0.8125,
0.080078125,
0.8125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81298828125,
0.35302734375,
0.81298828125,
0.35302734375,
0.8271484375,
0.080078125,
0.8271484375
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.86279296875,
0.4697265625,
0.86279296875,
0.4697265625,
0.876953125,
0.09722900390625,
0.876953125
],
[
0.080078125,
0.87548828125,
0.4697265625,
0.87548828125,
0.4697265625,
0.8896484375,
0.080078125,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88818359375,
0.35546875,
0.88818359375,
0.35546875,
0.9013671875,
0.080078125,
0.9013671875
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"Heatm",
"Keypoint",
"Poseestimation",
"FC",
"2Dcoordinates",
"&confidence",
"(36×3)",
"Convolutional neural",
"network",
"",
"",
"Deconvolution",
"Pooling",
"+Color",
"Cross-entropyloss",
"Multi-task learning",
"Batch",
"mple",
"+RelD",
"Convolutional neural",
"Cross-entropyloss",
"network",
"& Hard triplet loss",
"+Type",
"Deer",
"Cross-entropy loss",
"sample",
"nnl",
"(real",
"learning",
"Figure2.Overviewof theproposed method.Eachtrainingbatch includesbothreal and synthesized images.Toembedposeinformation",
"for multi-task learning, the heatmaps or segments output by a pre-trained network are stacked with the original RGB channels as input.",
"The estimated keypoint coordinates and confidence scores are also concatenated with deep learning features for ReID and attribute (color",
"and type) classification. The pose estimation network (top, blue) is based on HRNet [25], while the multi-task learning network (bottom,",
"orange)isbasedonDenseNet12l [7].Bestviewedincolor.",
"3.1.Randomized synthetic dataset",
"3.2. Vehicle pose estimation",
"Besides vehicle identities, our approach requires addi-",
"To leverage viewpoint-aware information for multi-task",
"tional labels of vehicle attributes and locations of keypoints.",
"learning, we train a robust DNN for extracting pose-related",
"These values, particularly the keypoints, would require con-",
"representations.Similar toTremblay et al.[35] wemixreal",
"siderable, even prohibitive effort, if annotated manually.To",
"and synthetic data to bridge the reality gap. More specifi-",
"overcome this problem, we generated a large-scale synthetic",
"cally, in each dataset, we utilize the pre-trained model [2] to",
"dataset by employing our deep learning dataset synthesizer",
"process sampled images and manually select about 10,000",
"(NDDS) [33] to create a randomized environment in Unreal",
"successful samples as real training data.",
"Engine 4 (UE4), into which 3D vehicle meshes from [22]",
"Instead of using the stacked hourglass network [2]] as",
"were imported. We added to NDDS the ability to label",
"the backbone like previous approaches [2, 37l, we mod-",
"and export specific 3D locations, i.e., keypoints (denoted",
"ify the state-of-the-art DNN for human pose estimation,",
"as sockets in UE4), on a CAD model. As such we manually",
"HRNet [25], for our purposes. Compared to the stacked",
"annotated each vehicle model with the 36 3D keypoints de-",
"hourglass architecture and other methods that recover high-",
"fined by Ansari et al. [2]; the projected 2D locations were",
"resolution representations from low-resolution representa-",
"then output with the synthesized images. For randomiza-",
"tions,HRNet maintains high-resolution representations and",
"tion we used 42 vehicle 3D CAD models with 10 body col-",
"gradually add high-to-low resolution sub-networks with",
" ors.To train the data for RelD, we define a unique iden-",
"multi-scale fusion.As a result, the predicted keypoints",
"tityforeachcombinationofvehiclemodelwithaparticular",
"and heatmaps are more accurate and spatially more precise,",
"color. The final generated dataset consists of 41,000 unique",
"which benefit our embedding for multi-task learning.",
"images with 402 identities,including the following anno-",
"We propose two ways to embed the vehicle pose in-",
"tations: keypoints, orientation, and vehicle attributes (color",
"formation as additional channels at the input layer of the",
" and type). When generating the dataset, background images",
"multi-task network, based on heatmaps and segments, re-",
"weresampledfromCityFlow[30l,andwealsorandomized",
"spectively. In one approach, after the final deconvolutional",
"the vehicle position and intensity of light. Additionally, dur-",
"layer, we extract the 36 heatmaps for each of the keypoints",
"ing training we perform randomized post-processing such",
"used to capture the vehicle shape and pose. In the other",
"as scaling, cropping, horizontal flipping, and adding occlu-",
"approach, the predicted keypoint coordinates from the final",
"sion. Some examples are shown in Fig. 3.",
"fully-connected (FC) layer are used to segment the vehicle",
"body. For example, in Fig. 3 the keypoints #16, #17, #35",
"and #34 from the deformable model form a segment that",
"represents the car hood. Accordingly, we define 13 seg-",
"IThe concrete mixer truck and the school bus did not get color variation",
"and as such we exported 500 unique images for each of them. 100 images",
"mentation masks for each vehicle (see Fig. 3 ToP-LEFT),",
"were generated for each of the remaining identities.",
"where those formed by keypoint(s) with low confidence are"
] |
{
"arxiv_id": "2005.00673",
"language": "en",
"page_idx": 2
}
|
||
2005.00673-en-4
|
2005.00673-en-4.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>$\#$total IDs</th>
<th>$\#$train IDs</th>
<th>井 test IDs</th>
<th>$\#$query images</th>
<th>$\#total$ images</th>
</tr>
<tr>
<td>VeRi</td>
<td>776</td>
<td>576</td>
<td>200</td>
<td>1,678</td>
<td>51,038</td>
</tr>
<tr>
<td>CityFlow- ReID</td>
<td>666</td>
<td>333</td>
<td>333</td>
<td>1,052</td>
<td>56,277</td>
</tr>
<tr>
<td>Synthetic</td>
<td>402</td>
<td>402</td>
<td> </td>
<td> </td>
<td>41,000</td>
</tr>
</tbody>
</table>
Table 1. Statistics of the datasets used for training and evaluation.
<table>
<tbody>
<tr>
<th>Method</th>
<th>mAP</th>
<th>Rank- 1</th>
<th>Rank- 5</th>
<th>Rank- 20</th>
</tr>
<tr>
<td>FACT $^{\prime}[13]$</td>
<td>18.73</td>
<td>51.85</td>
<td>67.16</td>
<td>79.56</td>
</tr>
<tr>
<td>PROVID* [15]</td>
<td>48.47</td>
<td>76.76</td>
<td>91.40</td>
<td> </td>
</tr>
<tr>
<td>$OIFE$ [37]</td>
<td>48.00</td>
<td>65.92</td>
<td>87.66</td>
<td>96.63</td>
</tr>
<tr>
<td>PathLSTM* [24]</td>
<td>58.27</td>
<td>83.49</td>
<td>90.04</td>
<td>97.16</td>
</tr>
<tr>
<td>GSTE [3]</td>
<td>59.47</td>
<td>96.24</td>
<td>98.97</td>
<td> </td>
</tr>
<tr>
<td>VAMI [45]</td>
<td>50.13</td>
<td>77.03</td>
<td>90.82</td>
<td>97.16</td>
</tr>
<tr>
<td>$VAMI^{*}$ [45]</td>
<td>61.32</td>
<td>85.92</td>
<td>91.84</td>
<td>97.70</td>
</tr>
<tr>
<td>ABLN [46]</td>
<td>24.92</td>
<td>60.49</td>
<td>77.33</td>
<td>88.27</td>
</tr>
<tr>
<td>$BA$ $\Lambda[10]$ </td>
<td>66.91</td>
<td>90.11</td>
<td>96.01</td>
<td>98.27</td>
</tr>
<tr>
<td>$BS[10]$</td>
<td>67.55</td>
<td>90.23</td>
<td>96.42</td>
<td>98.63</td>
</tr>
<tr>
<td>$RS$</td>
<td>63.76</td>
<td>90.70</td>
<td>94.40</td>
<td>97.47</td>
</tr>
<tr>
<td>RS+ MT</td>
<td>66.18</td>
<td>91.90</td>
<td>96.90</td>
<td>98.99</td>
</tr>
<tr>
<td>RS+ MT+ K</td>
<td>68.64</td>
<td>91.60</td>
<td>96.78</td>
<td>98.75</td>
</tr>
<tr>
<td>$RS+MT+K+H$</td>
<td>71.16</td>
<td>92.74</td>
<td>96.68</td>
<td>98.40</td>
</tr>
<tr>
<td>RS+ MT+ K+ S</td>
<td>71.88</td>
<td>92.86</td>
<td>96.97</td>
<td>98.23</td>
</tr>
<tr>
<td>RS w/ Xent only</td>
<td>56.52</td>
<td>83.41</td>
<td>92.07</td>
<td>97.02</td>
</tr>
<tr>
<td>RS w/ Htri only</td>
<td>47.50</td>
<td>73.54</td>
<td>87.25</td>
<td>96.01</td>
</tr>
<tr>
<td>RS+MT w/ DN201</td>
<td>64.42</td>
<td>90.58</td>
<td>96.36</td>
<td>98.81</td>
</tr>
<tr>
<td>R+ MT+ K</td>
<td>65.44</td>
<td>90.94</td>
<td>96.72</td>
<td>99.11</td>
</tr>
</tbody>
</table>
Table 2. Experimental comparison of state-of-the-art in vehicle ReID on VeRi[13]. All values are shown as percentages. For our proposed method, MT, K, H, S, RS and R respectively denote multi-task learning, explicit keypoints embedded, heatmaps embedded, segments embedded, training with both real and synthetic data, and training with real data only. Xent, Htri and DN201 stand for cross-entropy loss, hard triplet loss and DenseNet201, respectively. (*) indicates the usage of spatio-temporal information.
# 4. Evaluation
In this section, we present the datasets used for evaluating our proposed approach, the implementation details, experimental results showing state-of-the-art performance, and a detailed analysis of the effects of various components of PAMTRI.
## 4.1. Datasets and evaluation protocol
Our PAMTRI system was evaluated on two mainstream large-scale vehicle ReID benchmarks, namely, VeRi[14] and CityFlow-ReID [30], whose statistics are summarized in Tab. l together with the details of the synthetic data we generated for training.
VeRi[14] has been widely used by most recent research in vehicle ReID, as it provides multiple views of vehicles captured from 20 cameras. CityFlow-ReID [30] is a subset of the recent multi-target multi-camera vehicle tracking benchmark, CityFlow, which has been adopted for the AI City Challenge[20] at CVPR 2019. The latter is significantly more challenging, as the footage is captured with more cameras (40) in more diverse environments (residential areas and highways). Unlike VeRi, the original videos are available in CityFlow, which enable us to extract background images for randomization to generate realistic synthetic data. Whereas the color and type information is available with the VeRi dataset, such attribute annotation is not provided by CityFlow. Hence, another contribution of this work is that we manually labeled vehicle attributes (color and type) for all the 666 identities in CityFlow-ReID.
In our experiments, we strictly follow the evaluation protocol proposed in Market1501 [42] measuring the mean Average Precision (mAP) and the rank-$K$ hit rates. For mAP, we compute the mean of all queries' average precision, $i.e.$, the area under the Precision-Recall curve. The rank-$K$ hit rate denotes the possibility that at least one true positive is ranked within the top $K$ positions. When all the rank- $K$ hit rates are plotted against $K$,we have the Cumulative Matching Characteristic (CMC). In addition, rank-$K$ mAP is introduced in [30] that measures the mean of average precision for each query considering only the top $K$ matches.
## 4.2. Implementation details
Training for multi-task learning. Leveraging the offthe-shelf implementation in [44], we use DenseNet121[7]
as our backbone CNN for multi-task learning, whose initial weights are from the model pre-trained on ImageNet[5]. The input images are resized to $256\times256$ and the training batch size is set as 32. We utilize the Adam optimizer [9] to train the base model for 60 maximum epochs. The initial learning rate was set to 3e-4, which decays to 3e-5 and 3e-6 at the 20$^\mathrm{th}$ and 40$^\mathrm{th}$ epochs, respectively. For multitask learning the dimension of the last FC layer for ReID is 1,024, whereas the two FC layers for attribute classification share the size of 512 each. For all the final FC layers, we adopt the leaky rectifed linear unit (Leaky ReLU)[40] as the activation function.
Training for pose estimation. The state-of-the-art HRNet[25] for human pose estimation is used as our backbone for vehicle pose estimation, which is built upon the original implementation by Sun et al. Again we adopt the pre-trained weights on ImageNet[5] for initialization. Each input image is also resized to $256\times256$ and the size of the heatmap/segment output is $64\times64.$ We set the training batch size to be 32, and the maximum number of epochs is 210 with learning rate of le-3. The fnal FC layer is adjusted to output a 108-dimensional vector, as our vehicle model consists of 36 keypoints in 2D whose visibility (indicated by confidence scores) is also computed.
|
[
[
0.096435546875,
0.0947265625,
0.1405029296875,
0.0947265625,
0.1405029296875,
0.1060791015625,
0.096435546875,
0.1060791015625
],
[
0.19775390625,
0.0947265625,
0.236083984375,
0.0947265625,
0.236083984375,
0.1060791015625,
0.19775390625,
0.1060791015625
],
[
0.250732421875,
0.0947265625,
0.289306640625,
0.0947265625,
0.289306640625,
0.1060791015625,
0.250732421875,
0.1060791015625
],
[
0.303955078125,
0.0947265625,
0.336669921875,
0.0947265625,
0.336669921875,
0.1060791015625,
0.303955078125,
0.1060791015625
],
[
0.3505859375,
0.0947265625,
0.39453125,
0.0947265625,
0.39453125,
0.1060791015625,
0.3505859375,
0.1060791015625
],
[
0.410888671875,
0.0947265625,
0.450927734375,
0.0947265625,
0.450927734375,
0.1060791015625,
0.410888671875,
0.1060791015625
],
[
0.52685546875,
0.09344482421875,
0.576171875,
0.09344482421875,
0.576171875,
0.10797119140625,
0.52685546875,
0.10797119140625
],
[
0.65185546875,
0.0947265625,
0.68310546875,
0.0947265625,
0.68310546875,
0.106689453125,
0.65185546875,
0.106689453125
],
[
0.69775390625,
0.0947265625,
0.74169921875,
0.0947265625,
0.74169921875,
0.1060791015625,
0.69775390625,
0.1060791015625
],
[
0.75634765625,
0.0947265625,
0.7998046875,
0.0947265625,
0.7998046875,
0.1060791015625,
0.75634765625,
0.1060791015625
],
[
0.81640625,
0.0947265625,
0.86669921875,
0.0947265625,
0.86669921875,
0.1060791015625,
0.81640625,
0.1060791015625
],
[
0.1993408203125,
0.10479736328125,
0.2320556640625,
0.10479736328125,
0.2320556640625,
0.12060546875,
0.1993408203125,
0.12060546875
],
[
0.253173828125,
0.10479736328125,
0.28515625,
0.10479736328125,
0.28515625,
0.12060546875,
0.253173828125,
0.12060546875
],
[
0.3056640625,
0.1060791015625,
0.334228515625,
0.1060791015625,
0.334228515625,
0.11871337890625,
0.3056640625,
0.11871337890625
],
[
0.351318359375,
0.1060791015625,
0.39453125,
0.10797119140625,
0.393798828125,
0.11932373046875,
0.3505859375,
0.117431640625
],
[
0.41015625,
0.106689453125,
0.451904296875,
0.106689453125,
0.451904296875,
0.11871337890625,
0.41015625,
0.11871337890625
],
[
0.529296875,
0.11236572265625,
0.5888671875,
0.11236572265625,
0.5888671875,
0.12371826171875,
0.529296875,
0.12371826171875
],
[
0.650390625,
0.11236572265625,
0.6845703125,
0.11236572265625,
0.6845703125,
0.125,
0.650390625,
0.125
],
[
0.70361328125,
0.11236572265625,
0.73779296875,
0.11236572265625,
0.73779296875,
0.12371826171875,
0.70361328125,
0.12371826171875
],
[
0.7607421875,
0.11236572265625,
0.7958984375,
0.11236572265625,
0.7958984375,
0.125,
0.7607421875,
0.125
],
[
0.82275390625,
0.11236572265625,
0.85693359375,
0.11236572265625,
0.85693359375,
0.125,
0.82275390625,
0.125
],
[
0.096435546875,
0.12371826171875,
0.1282958984375,
0.12371826171875,
0.1282958984375,
0.1363525390625,
0.096435546875,
0.1363525390625
],
[
0.205078125,
0.12371826171875,
0.2296142578125,
0.12371826171875,
0.2296142578125,
0.1363525390625,
0.205078125,
0.1363525390625
],
[
0.256591796875,
0.12371826171875,
0.28515625,
0.12371826171875,
0.28515625,
0.1383056640625,
0.256591796875,
0.1383056640625
],
[
0.306396484375,
0.12371826171875,
0.334228515625,
0.12371826171875,
0.334228515625,
0.1363525390625,
0.306396484375,
0.1363525390625
],
[
0.35693359375,
0.125,
0.390625,
0.125,
0.390625,
0.1363525390625,
0.35693359375,
0.1363525390625
],
[
0.41015625,
0.125,
0.451904296875,
0.125,
0.451904296875,
0.1363525390625,
0.41015625,
0.1363525390625
],
[
0.52783203125,
0.12371826171875,
0.61279296875,
0.125,
0.61279296875,
0.1363525390625,
0.52783203125,
0.1351318359375
],
[
0.650390625,
0.125,
0.6845703125,
0.125,
0.6845703125,
0.1363525390625,
0.650390625,
0.1363525390625
],
[
0.70361328125,
0.125,
0.73779296875,
0.125,
0.73779296875,
0.1363525390625,
0.70361328125,
0.1363525390625
],
[
0.7607421875,
0.125,
0.79736328125,
0.125,
0.79736328125,
0.1363525390625,
0.7607421875,
0.1363525390625
],
[
0.098876953125,
0.1363525390625,
0.1829833984375,
0.1363525390625,
0.1829833984375,
0.147705078125,
0.098876953125,
0.147705078125
],
[
0.2034912109375,
0.1363525390625,
0.2303466796875,
0.1363525390625,
0.2303466796875,
0.1483154296875,
0.2034912109375,
0.1483154296875
],
[
0.256591796875,
0.1351318359375,
0.28515625,
0.1351318359375,
0.28515625,
0.149658203125,
0.256591796875,
0.149658203125
],
[
0.3056640625,
0.1363525390625,
0.334228515625,
0.1363525390625,
0.334228515625,
0.1483154296875,
0.3056640625,
0.1483154296875
],
[
0.35546875,
0.1363525390625,
0.390625,
0.1363525390625,
0.390625,
0.1483154296875,
0.35546875,
0.1483154296875
],
[
0.41015625,
0.1363525390625,
0.451904296875,
0.1363525390625,
0.451904296875,
0.147705078125,
0.41015625,
0.147705078125
],
[
0.53125,
0.1363525390625,
0.58837890625,
0.1363525390625,
0.58837890625,
0.147705078125,
0.53125,
0.147705078125
],
[
0.650390625,
0.1363525390625,
0.685546875,
0.1363525390625,
0.685546875,
0.1483154296875,
0.650390625,
0.1483154296875
],
[
0.70361328125,
0.1363525390625,
0.73876953125,
0.1363525390625,
0.73876953125,
0.1483154296875,
0.70361328125,
0.1483154296875
],
[
0.7607421875,
0.1363525390625,
0.79833984375,
0.1363525390625,
0.79833984375,
0.149658203125,
0.7607421875,
0.149658203125
],
[
0.82275390625,
0.1363525390625,
0.85693359375,
0.1363525390625,
0.85693359375,
0.1483154296875,
0.82275390625,
0.1483154296875
],
[
0.2034912109375,
0.147705078125,
0.2296142578125,
0.147705078125,
0.2296142578125,
0.1610107421875,
0.2034912109375,
0.1610107421875
],
[
0.256591796875,
0.147705078125,
0.281982421875,
0.147705078125,
0.281982421875,
0.1610107421875,
0.256591796875,
0.1610107421875
],
[
0.41015625,
0.1483154296875,
0.451904296875,
0.1483154296875,
0.451904296875,
0.160400390625,
0.41015625,
0.160400390625
],
[
0.53125,
0.1483154296875,
0.62353515625,
0.1483154296875,
0.62353515625,
0.160400390625,
0.53125,
0.160400390625
],
[
0.650390625,
0.1483154296875,
0.6845703125,
0.1483154296875,
0.6845703125,
0.160400390625,
0.650390625,
0.160400390625
],
[
0.70361328125,
0.1483154296875,
0.73876953125,
0.1483154296875,
0.73876953125,
0.160400390625,
0.70361328125,
0.160400390625
],
[
0.76123046875,
0.1483154296875,
0.7958984375,
0.1483154296875,
0.7958984375,
0.160400390625,
0.76123046875,
0.160400390625
],
[
0.82275390625,
0.1483154296875,
0.85693359375,
0.1483154296875,
0.85693359375,
0.1610107421875,
0.82275390625,
0.1610107421875
],
[
0.31201171875,
0.15283203125,
0.32275390625,
0.1483154296875,
0.326904296875,
0.1546630859375,
0.316162109375,
0.1590576171875
],
[
0.09722900390625,
0.1483154296875,
0.1519775390625,
0.1483154296875,
0.1519775390625,
0.160400390625,
0.09722900390625,
0.160400390625
],
[
0.529296875,
0.1572265625,
0.58837890625,
0.1590576171875,
0.58642578125,
0.173583984375,
0.52783203125,
0.1717529296875
],
[
0.650390625,
0.160400390625,
0.6845703125,
0.160400390625,
0.6845703125,
0.1717529296875,
0.650390625,
0.1717529296875
],
[
0.70361328125,
0.160400390625,
0.73779296875,
0.160400390625,
0.73779296875,
0.17236328125,
0.70361328125,
0.17236328125
],
[
0.76123046875,
0.160400390625,
0.7958984375,
0.160400390625,
0.7958984375,
0.1717529296875,
0.76123046875,
0.1717529296875
],
[
0.53125,
0.17236328125,
0.59228515625,
0.17236328125,
0.59228515625,
0.184326171875,
0.53125,
0.184326171875
],
[
0.650390625,
0.17236328125,
0.6845703125,
0.17236328125,
0.6845703125,
0.184326171875,
0.650390625,
0.184326171875
],
[
0.70361328125,
0.17236328125,
0.73876953125,
0.17236328125,
0.73876953125,
0.184326171875,
0.70361328125,
0.184326171875
],
[
0.76123046875,
0.17236328125,
0.7958984375,
0.17236328125,
0.7958984375,
0.184326171875,
0.76123046875,
0.184326171875
],
[
0.82275390625,
0.17236328125,
0.85693359375,
0.17236328125,
0.85693359375,
0.1849365234375,
0.82275390625,
0.1849365234375
],
[
0.52783203125,
0.18115234375,
0.5986328125,
0.18310546875,
0.5986328125,
0.1976318359375,
0.52783203125,
0.1956787109375
],
[
0.650390625,
0.184326171875,
0.6845703125,
0.184326171875,
0.6845703125,
0.1962890625,
0.650390625,
0.1962890625
],
[
0.080078125,
0.175537109375,
0.46728515625,
0.175537109375,
0.46728515625,
0.189453125,
0.080078125,
0.189453125
],
[
0.70361328125,
0.184326171875,
0.73779296875,
0.184326171875,
0.73779296875,
0.1962890625,
0.70361328125,
0.1962890625
],
[
0.7607421875,
0.184326171875,
0.7958984375,
0.184326171875,
0.7958984375,
0.1962890625,
0.7607421875,
0.1962890625
],
[
0.82275390625,
0.184326171875,
0.85693359375,
0.184326171875,
0.85693359375,
0.1962890625,
0.82275390625,
0.1962890625
],
[
0.53125,
0.1962890625,
0.5947265625,
0.1962890625,
0.5947265625,
0.2083740234375,
0.53125,
0.2083740234375
],
[
0.650390625,
0.1962890625,
0.6845703125,
0.1962890625,
0.6845703125,
0.2083740234375,
0.650390625,
0.2083740234375
],
[
0.70361328125,
0.1962890625,
0.73876953125,
0.1962890625,
0.73876953125,
0.2083740234375,
0.70361328125,
0.2083740234375
],
[
0.76123046875,
0.1956787109375,
0.7958984375,
0.1956787109375,
0.7958984375,
0.2083740234375,
0.76123046875,
0.2083740234375
],
[
0.82275390625,
0.1962890625,
0.85693359375,
0.1962890625,
0.85693359375,
0.208984375,
0.82275390625,
0.208984375
],
[
0.529296875,
0.205810546875,
0.57861328125,
0.2083740234375,
0.57666015625,
0.222900390625,
0.52783203125,
0.2203369140625
],
[
0.650390625,
0.2083740234375,
0.68310546875,
0.2083740234375,
0.68310546875,
0.2203369140625,
0.650390625,
0.2203369140625
],
[
0.70361328125,
0.2083740234375,
0.73779296875,
0.2083740234375,
0.73779296875,
0.2203369140625,
0.70361328125,
0.2203369140625
],
[
0.7607421875,
0.2083740234375,
0.7939453125,
0.2083740234375,
0.7939453125,
0.2203369140625,
0.7607421875,
0.2203369140625
],
[
0.82275390625,
0.2083740234375,
0.85693359375,
0.2083740234375,
0.85693359375,
0.2203369140625,
0.82275390625,
0.2203369140625
],
[
0.529296875,
0.2197265625,
0.57666015625,
0.2197265625,
0.57666015625,
0.2342529296875,
0.529296875,
0.2342529296875
],
[
0.7607421875,
0.2197265625,
0.79736328125,
0.2197265625,
0.79736328125,
0.23486328125,
0.7607421875,
0.23486328125
],
[
0.650390625,
0.2203369140625,
0.6845703125,
0.2203369140625,
0.6845703125,
0.23291015625,
0.650390625,
0.23291015625
],
[
0.70361328125,
0.2203369140625,
0.73779296875,
0.2203369140625,
0.73779296875,
0.23291015625,
0.70361328125,
0.23291015625
],
[
0.82275390625,
0.2203369140625,
0.85693359375,
0.2203369140625,
0.85693359375,
0.23291015625,
0.82275390625,
0.23291015625
],
[
0.52783203125,
0.2374267578125,
0.55126953125,
0.2374267578125,
0.55126953125,
0.25,
0.52783203125,
0.25
],
[
0.650390625,
0.238037109375,
0.6845703125,
0.238037109375,
0.6845703125,
0.250732421875,
0.650390625,
0.250732421875
],
[
0.705078125,
0.238037109375,
0.73779296875,
0.238037109375,
0.73779296875,
0.250732421875,
0.705078125,
0.250732421875
],
[
0.76123046875,
0.238037109375,
0.7958984375,
0.238037109375,
0.7958984375,
0.25,
0.76123046875,
0.25
],
[
0.82421875,
0.238037109375,
0.85693359375,
0.238037109375,
0.85693359375,
0.250732421875,
0.82421875,
0.250732421875
],
[
0.098876953125,
0.2412109375,
0.468017578125,
0.2412109375,
0.468017578125,
0.255126953125,
0.098876953125,
0.255126953125
],
[
0.080078125,
0.2139892578125,
0.193603515625,
0.21533203125,
0.193603515625,
0.2298583984375,
0.080078125,
0.2291259765625
],
[
0.53125,
0.250732421875,
0.57861328125,
0.250732421875,
0.57861328125,
0.261962890625,
0.53125,
0.261962890625
],
[
0.650390625,
0.250732421875,
0.6845703125,
0.250732421875,
0.6845703125,
0.26318359375,
0.650390625,
0.26318359375
],
[
0.70361328125,
0.250732421875,
0.73876953125,
0.250732421875,
0.73876953125,
0.26318359375,
0.70361328125,
0.26318359375
],
[
0.7607421875,
0.25,
0.7958984375,
0.25,
0.7958984375,
0.261962890625,
0.7607421875,
0.261962890625
],
[
0.82275390625,
0.250732421875,
0.85693359375,
0.250732421875,
0.85693359375,
0.26318359375,
0.82275390625,
0.26318359375
],
[
0.080078125,
0.2568359375,
0.46728515625,
0.2568359375,
0.46728515625,
0.270751953125,
0.080078125,
0.270751953125
],
[
0.53125,
0.261962890625,
0.59326171875,
0.261962890625,
0.59326171875,
0.2734375,
0.53125,
0.2734375
],
[
0.650390625,
0.261474609375,
0.6845703125,
0.26318359375,
0.68310546875,
0.27587890625,
0.64892578125,
0.2734375
],
[
0.70361328125,
0.261962890625,
0.73876953125,
0.261962890625,
0.73876953125,
0.274658203125,
0.70361328125,
0.274658203125
],
[
0.76123046875,
0.261962890625,
0.7958984375,
0.261962890625,
0.7958984375,
0.274658203125,
0.76123046875,
0.274658203125
],
[
0.82275390625,
0.261962890625,
0.85693359375,
0.261962890625,
0.85693359375,
0.274658203125,
0.82275390625,
0.274658203125
],
[
0.080078125,
0.271484375,
0.468017578125,
0.271484375,
0.468017578125,
0.285400390625,
0.080078125,
0.285400390625
],
[
0.53125,
0.2734375,
0.611328125,
0.2734375,
0.611328125,
0.285400390625,
0.53125,
0.285400390625
],
[
0.650390625,
0.2734375,
0.685546875,
0.2734375,
0.685546875,
0.285888671875,
0.650390625,
0.285888671875
],
[
0.70361328125,
0.2734375,
0.73876953125,
0.2734375,
0.73876953125,
0.285888671875,
0.70361328125,
0.285888671875
],
[
0.76123046875,
0.2734375,
0.7958984375,
0.2734375,
0.7958984375,
0.285888671875,
0.76123046875,
0.285888671875
],
[
0.82275390625,
0.2734375,
0.85693359375,
0.2734375,
0.85693359375,
0.285888671875,
0.82275390625,
0.285888671875
],
[
0.0767822265625,
0.285400390625,
0.4697265625,
0.285888671875,
0.4697265625,
0.300537109375,
0.0767822265625,
0.2998046875
],
[
0.529296875,
0.285400390625,
0.609375,
0.285400390625,
0.609375,
0.2998046875,
0.529296875,
0.2998046875
],
[
0.650390625,
0.285888671875,
0.6845703125,
0.285888671875,
0.6845703125,
0.298583984375,
0.650390625,
0.298583984375
],
[
0.70361328125,
0.285888671875,
0.73876953125,
0.285888671875,
0.73876953125,
0.298583984375,
0.70361328125,
0.298583984375
],
[
0.76123046875,
0.285888671875,
0.7958984375,
0.285888671875,
0.7958984375,
0.298583984375,
0.76123046875,
0.298583984375
],
[
0.82275390625,
0.285888671875,
0.85693359375,
0.285888671875,
0.85693359375,
0.298583984375,
0.82275390625,
0.298583984375
],
[
0.080078125,
0.302978515625,
0.16259765625,
0.302978515625,
0.16259765625,
0.314453125,
0.080078125,
0.314453125
],
[
0.52783203125,
0.3037109375,
0.62109375,
0.3037109375,
0.62109375,
0.31494140625,
0.52783203125,
0.31494140625
],
[
0.650390625,
0.3037109375,
0.6845703125,
0.3037109375,
0.6845703125,
0.31640625,
0.650390625,
0.31640625
],
[
0.705078125,
0.3037109375,
0.73779296875,
0.3037109375,
0.73779296875,
0.31640625,
0.705078125,
0.31640625
],
[
0.7607421875,
0.3037109375,
0.7939453125,
0.3037109375,
0.7939453125,
0.31640625,
0.7607421875,
0.31640625
],
[
0.82275390625,
0.3037109375,
0.85693359375,
0.3037109375,
0.85693359375,
0.31494140625,
0.82275390625,
0.31494140625
],
[
0.53125,
0.31640625,
0.61669921875,
0.31640625,
0.61669921875,
0.32763671875,
0.53125,
0.32763671875
],
[
0.650390625,
0.31640625,
0.685546875,
0.31640625,
0.685546875,
0.32763671875,
0.650390625,
0.32763671875
],
[
0.70361328125,
0.31640625,
0.73876953125,
0.31640625,
0.73876953125,
0.328857421875,
0.70361328125,
0.328857421875
],
[
0.7607421875,
0.31640625,
0.7958984375,
0.31640625,
0.7958984375,
0.32763671875,
0.7607421875,
0.32763671875
],
[
0.82177734375,
0.31640625,
0.85546875,
0.31640625,
0.85546875,
0.32763671875,
0.82177734375,
0.32763671875
],
[
0.08087158203125,
0.326904296875,
0.36279296875,
0.326904296875,
0.36279296875,
0.341552734375,
0.08087158203125,
0.341552734375
],
[
0.53125,
0.32763671875,
0.6357421875,
0.32763671875,
0.6357421875,
0.34033203125,
0.53125,
0.34033203125
],
[
0.64892578125,
0.32763671875,
0.685546875,
0.32763671875,
0.685546875,
0.339111328125,
0.64892578125,
0.339111328125
],
[
0.70361328125,
0.326904296875,
0.73876953125,
0.328857421875,
0.73779296875,
0.341552734375,
0.70166015625,
0.339111328125
],
[
0.76123046875,
0.32763671875,
0.79736328125,
0.32763671875,
0.79736328125,
0.34033203125,
0.76123046875,
0.34033203125
],
[
0.82177734375,
0.32763671875,
0.85546875,
0.32763671875,
0.85546875,
0.34033203125,
0.82177734375,
0.34033203125
],
[
0.529296875,
0.339111328125,
0.58837890625,
0.339111328125,
0.58837890625,
0.354248046875,
0.529296875,
0.354248046875
],
[
0.650390625,
0.34033203125,
0.6845703125,
0.34033203125,
0.6845703125,
0.352783203125,
0.650390625,
0.352783203125
],
[
0.70361328125,
0.34033203125,
0.73876953125,
0.34033203125,
0.73876953125,
0.352783203125,
0.70361328125,
0.352783203125
],
[
0.76123046875,
0.34033203125,
0.7958984375,
0.34033203125,
0.7958984375,
0.352783203125,
0.76123046875,
0.352783203125
],
[
0.82177734375,
0.33837890625,
0.85693359375,
0.33837890625,
0.85693359375,
0.354248046875,
0.82177734375,
0.354248046875
],
[
0.098876953125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.098876953125,
0.365478515625
],
[
0.080078125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.080078125,
0.380615234375
],
[
0.50244140625,
0.367431640625,
0.89208984375,
0.367431640625,
0.89208984375,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.080078125,
0.3818359375,
0.4697265625,
0.3818359375,
0.4697265625,
0.395751953125,
0.080078125,
0.395751953125
],
[
0.50244140625,
0.3818359375,
0.8935546875,
0.3818359375,
0.8935546875,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.080078125,
0.397705078125,
0.471435546875,
0.397705078125,
0.471435546875,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.080078125,
0.41357421875,
0.2279052734375,
0.41357421875,
0.2279052734375,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.42236328125,
0.50244140625,
0.42236328125
],
[
0.1004638671875,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.1004638671875,
0.44140625
],
[
0.50244140625,
0.423583984375,
0.89208984375,
0.423583984375,
0.89208984375,
0.4375,
0.50244140625,
0.4375
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.50244140625,
0.45068359375,
0.89208984375,
0.45068359375,
0.89208984375,
0.464111328125,
0.50244140625,
0.464111328125
],
[
0.080078125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.50244140625,
0.464111328125,
0.86376953125,
0.464111328125,
0.86376953125,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.080078125,
0.473388671875,
0.468017578125,
0.473388671875,
0.468017578125,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.080078125,
0.488037109375,
0.4697265625,
0.488037109375,
0.4697265625,
0.501953125,
0.080078125,
0.501953125
],
[
0.07843017578125,
0.50244140625,
0.46728515625,
0.501953125,
0.46728515625,
0.5166015625,
0.07843017578125,
0.517578125
],
[
0.07843017578125,
0.51806640625,
0.468017578125,
0.517578125,
0.468017578125,
0.5322265625,
0.07843017578125,
0.53271484375
],
[
0.50341796875,
0.5263671875,
0.89208984375,
0.5263671875,
0.89208984375,
0.54150390625,
0.50341796875,
0.54150390625
],
[
0.080078125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50244140625,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.5498046875,
0.471435546875,
0.5498046875,
0.471435546875,
0.56298828125,
0.080078125,
0.56298828125
],
[
0.5009765625,
0.556640625,
0.890625,
0.556640625,
0.890625,
0.57080078125,
0.5009765625,
0.57080078125
],
[
0.080078125,
0.56298828125,
0.46728515625,
0.56298828125,
0.46728515625,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.50244140625,
0.57373046875,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.080078125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.07843017578125,
0.59326171875,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.07843017578125,
0.60693359375
],
[
0.50244140625,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.6015625,
0.50244140625,
0.6015625
],
[
0.080078125,
0.609375,
0.4697265625,
0.609375,
0.4697265625,
0.62255859375,
0.080078125,
0.62255859375
],
[
0.50244140625,
0.60400390625,
0.89208984375,
0.60400390625,
0.89208984375,
0.6171875,
0.50244140625,
0.6171875
],
[
0.50244140625,
0.6171875,
0.890625,
0.6171875,
0.890625,
0.6318359375,
0.50244140625,
0.6318359375
],
[
0.07843017578125,
0.625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.07843017578125,
0.63818359375
],
[
0.080078125,
0.6396484375,
0.4697265625,
0.6396484375,
0.4697265625,
0.65283203125,
0.080078125,
0.65283203125
],
[
0.50244140625,
0.6337890625,
0.89208984375,
0.6337890625,
0.89208984375,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.50341796875,
0.6484375,
0.89208984375,
0.6484375,
0.89208984375,
0.662109375,
0.50341796875,
0.662109375
],
[
0.07843017578125,
0.65478515625,
0.428955078125,
0.65478515625,
0.428955078125,
0.66845703125,
0.07843017578125,
0.66845703125
],
[
0.50244140625,
0.6640625,
0.89208984375,
0.6640625,
0.89208984375,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.098876953125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.098876953125,
0.6845703125
],
[
0.50244140625,
0.6787109375,
0.89208984375,
0.6787109375,
0.89208984375,
0.6923828125,
0.50244140625,
0.6923828125
],
[
0.080078125,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.080078125,
0.70068359375,
0.468017578125,
0.70068359375,
0.468017578125,
0.7138671875,
0.080078125,
0.7138671875
],
[
0.50341796875,
0.69580078125,
0.65283203125,
0.69580078125,
0.65283203125,
0.70556640625,
0.50341796875,
0.70556640625
],
[
0.080078125,
0.71630859375,
0.4697265625,
0.71630859375,
0.4697265625,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.52392578125,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.52392578125,
0.73291015625
],
[
0.080078125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.080078125,
0.744140625
],
[
0.5009765625,
0.73486328125,
0.88818359375,
0.73486328125,
0.88818359375,
0.74853515625,
0.5009765625,
0.74853515625
],
[
0.080078125,
0.7470703125,
0.4697265625,
0.7470703125,
0.4697265625,
0.76025390625,
0.080078125,
0.76025390625
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.080078125,
0.76123046875,
0.468017578125,
0.76123046875,
0.468017578125,
0.7744140625,
0.080078125,
0.7744140625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.77587890625,
0.46728515625,
0.77587890625,
0.46728515625,
0.7890625,
0.08087158203125,
0.7890625
],
[
0.5009765625,
0.7822265625,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.5009765625,
0.79541015625
],
[
0.080078125,
0.79052734375,
0.468017578125,
0.79052734375,
0.468017578125,
0.8037109375,
0.080078125,
0.8037109375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.0767822265625,
0.8046875,
0.46728515625,
0.80615234375,
0.46728515625,
0.82080078125,
0.0767822265625,
0.8193359375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8212890625,
0.463134765625,
0.8212890625,
0.463134765625,
0.83544921875,
0.080078125,
0.83544921875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.84765625,
0.291748046875,
0.84765625,
0.291748046875,
0.8623046875,
0.08087158203125,
0.8623046875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.098876953125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.81396484375,
0.88720703125,
0.81396484375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Dataset",
"# total",
"# train",
"# test",
"# query",
"# total",
"Method",
"mAP",
"Rank-1",
"Rank-5",
"Rank-20",
"IDs",
"IDs",
"IDs",
"images",
"images",
"FACT [13]",
"18.73",
"51.85",
"67.16",
"79.56",
"VeRi",
"776",
"576",
"200",
"1,678",
"51,038",
"PROVID*[15]",
"48.47",
"76.76",
"91.40",
"CityFlow-ReID",
"666",
"333",
"333",
"1,052",
"56,277",
"OIFE [37]",
"48.00",
"65.92",
"87.66",
"96.63",
"402",
"402",
"41.000",
"PathLSTM*[24]",
"58.27",
"83.49",
"90.04",
"97.16",
"/",
"Synthetic",
"GSTE[3]",
"59.47",
"96.24",
"98.97",
"VAMI [45]",
"50.13",
"77.03",
"90.82",
"97.16",
"VAMI* [45]",
"61.32",
"Table 1. Statistics of the datasets used for training and evaluation.",
"85.92",
"91.84",
"97.70",
"ABLN [46]",
"24.92",
"60.49",
"77.33",
"88.27",
"BA [10]",
"66.91",
"90.11",
"96.01",
"98.27",
"BS [10]",
"96.42",
"67.55",
"90.23",
"98.63",
"RS",
"63.76",
"90.70",
"94.40",
"97.47",
"In this section, we present the datasets used for evalu-",
"4.Evaluation",
"RS+MT",
"66.18",
"91.90",
"96.90",
"98.99",
"ating our proposed approach, the implementation details,",
"RS+MT+K",
"68.64",
"91.60",
"96.78",
"98.75",
"experimental results showing state-of-the-art performance,",
"RS+MT+K+H",
"71.16",
"92.74",
"96.68",
"98.40",
" and a detailed analysis of the effects of various components",
"RS+MT+K+S",
"71.88",
"92.86",
"96.97",
"98.23",
"of PAMTRI.",
"RSw/Xentonly",
"56.52",
"83.41",
"92.07",
"97.02",
"RS w/ Htri only",
"47.50",
"73.54",
"87.25",
"96.01",
"4.1. Datasets and evaluation protocol",
"RS+MTw/DN201",
"64.42",
"90.58",
"96.36",
"98.81",
"R+MT+K",
"65.44",
"90.94",
"96.72",
"99.11",
"Our PAMTRI system was evaluated on two mainstream",
"large-scale vehicle RelD benchmarks, namely, VeRi [l4]",
"Table 2. Experimental comparison of state-of-the-art in vehicle",
"and CityFlow-RelD [30], whose statistics are summarized",
"RelD on VeRi [13]. All values are shown as percentages. For",
"in Tab. 1 together with the details of the synthetic data we",
"our proposed method, MT, K, H, S, RS and R respectively denote",
"generated for training.",
"multi-task learning, explicit keypoints embedded, heatmaps em-",
"VeRi [14] has been widely used by most recent research",
"bedded, segments embedded, training with both real and synthetic",
"in vehicle ReID, as it provides multiple views of vehicles",
"data, and training with real data only. Xent, Htri and DN201 stand",
"for cross-entropy loss, hard triplet loss and DenseNet201, respec-",
"captured from 20 cameras. CityFlow-ReID [30] is a sub-",
"tively. (*) indicates the usage of spatio-temporal information.",
"set of the recent multi-target multi-camera vehicle tracking",
"benchmark, CityFlow, which has been adopted for the AI",
"City Challenge [20] at CVPR 2019. The latter is signif-",
"icantly more challenging, as the footage is captured with",
"as our backbone CNN for multi-task learning, whose initial",
"more cameras (40) in more diverse environments (residen-",
"weights are from the model pre-trained on ImageNet [5].",
"tial areas and highways). Unlike VeRi, the original videos",
"The input images are resized to 256 × 256 and the training",
"are available in CityFlow, which enable us to extract back-",
"batch size is set as 32. We utilize the Adam optimizer [9]",
"ground images for randomization to generate realistic syn-",
"thetic data. Whereas the color and type information is avail-",
"to train the base model for 60 maximum epochs. The ini-",
"able with the VeRi dataset, such attribute annotation is not",
"tial learning rate was set to 3e-4, which decays to 3e-5 and",
"3e-6 at the 20th and 4oth epochs, respectively. For multi-",
"provided by CityFlow. Hence, another contribution of this",
"work is that we manually labeled vehicle attributes (color",
"task learning the dimension of the last FC layer for ReID is",
"1,024, whereas the two FC layers for attribute classification",
"and type) for all the 666 identities in CityFlow-ReID.",
"share the size of 512 each. For all the final FC layers, we",
"In our experiments, we strictly follow the evaluation pro-",
"adopt the leaky rectified linear unit (Leaky ReLU) [40] as",
"tocol proposed in Market1501 [42] measuring the mean Av-",
"erage Precision (mAP) and the rank-K hit rates. For mAP,",
"theactivationfunction",
"we compute the mean of all queries' average precision, i.e.,",
"Training for pose estimation. The state-of-the-art HR-",
"the area under the Precision-Recall curve. The rank-K hit",
"Net [25] for human pose estimation is used as our back",
"rate denotes the possibility that at least one true positive",
"bone for vehicle pose estimation, which is built upon the",
"is ranked within the top K positions. When all the rank-",
"original implementation by Sun et al. Again we adopt the",
"K hit rates are plotted against K, we have the Cumulative",
"pre-trained weights on ImageNet [5] for initialization. Each",
"Matching Characteristic (CMC). In addition, rank-K mAP",
"input image is also resized to 256 × 256 and the size of",
"is introduced in [30] that measures the mean of average pre-",
"the heatmap/segment output is 64 × 64. We set the training",
"cision for each query considering only the top K matches.",
"batch size to be 32, and the maximum number of epochs",
"is 210 with learning rate of le-3. The final FC layer is ad-",
"4.2. Implementation details",
"justed to output a 108-dimensional vector, as our vehicle",
"Training for multi-task learning. Leveraging the off-",
"model consists of 36 keypoints in 2D whose visibility (in-",
"the-shelf implementation in [44], we use DenseNet121 [7]",
"dicated by confidence scores) is also computed."
] |
{
"arxiv_id": "2005.00673",
"language": "en",
"page_idx": 4
}
|
||
2205.12219-en-0
|
2205.12219-en-0.jpg
|
Aerial Vision-and-Dialog Navigation
Yue Fan, Winson Chen, Tongzhou Jiang, Chun Zhou, Yi Zhang, Xin Eric Wang
University of California, Santa Cruz
$\{$yfan71, wchen157, tojiang, czhou43, yiz, xwang366}@ucsc.edu
## $\mathbf{Abstract}$
The ability to converse with humans and follow natural language commands is crucial for intelligent unmanned aerial vehicles (a.k.a. drones). It can relieve people's burden of holding a controller all the time, allow multitasking, and make drone control more accessible for people with disabilities or with their hands occupied. To this end, we introduce Aerial Vision-andDialog Navigation (AVDN), to navigate a drone via natural language conversation. We build a drone simulator with a continuous photorealistic environment and collect a new AVDN dataset of over 3k recorded navigation trajectories with asynchronous human-human dialogs between commanders and followers. The commander provides initial navigation instruction and further guidance by request, while the follower navigates the drone in the simulator and asks questions when needed. During data collection, followers' attention on the drone's visual observation is also recorded. Based on the AVDN dataset, we study the tasks of aerial navigation from (full) dialog history and propose an effective Human Attention Aided Transformer model (HAA-Transformer), which learns to predict both navigation waypoints and human attention. Dataset and code are released: https://sites.google.com/view/ aerial-vision-and-dialog/home.
## 1 Introduction
Drones have been widely adopted for many applications in our daily life, from personal entertainment to professional use. It has the advantage of mobility and observing large areas over ground robots. However, compared with ground robots, the control of the aerial robot is more complex because an extra degree of freedom, altitude, is involved. To control a drone, people often need to hold a controller all the time, so it is essential to create a hands-free control experience for drone users and develop an intelligent drone that can complete tasks simply by talking to humans. It can lower the barrier of
<FigureHere>
Figure 1: An example of Aerial Vision-and-Dialog Nav-
igation (AVDN). The user instructs the agent to fly to a destination. During the navigation, the agent can ask questions while showing the images of past visual observations and relative trajectory. The user will talk back at a convenient time to provide further guidance to the agent without having to monitor the agent all the time.
drone control for users with some disabilities and who have their hands occupied by activities such as taking photos, writing, etc.
Therefore, this work introduces Aerial Visionand-Dialog Navigation (AVDN), aiming to develop an intelligent drone that can converse with its user to fly to the expected destination. As shown in Figure 1, the user (commander) provides instructions, and the aerial agent (follower) follows the instruction and asks questions when needed. The past visual trajectories are also provided along with the question, which frees the commander from monitoring the drone all the time and minimizes the burden of drone control. In this free-form dialog, potential ambiguities in the instruction can be gradually resolved through the further instructions provided by the commanders upon request.
To implement and evaluate the AVDN task, we
|
[
[
0.30908203125,
0.0914306640625,
0.68701171875,
0.0926513671875,
0.68701171875,
0.10986328125,
0.30908203125,
0.1092529296875
],
[
0.162109375,
0.1436767578125,
0.8427734375,
0.144287109375,
0.8427734375,
0.1591796875,
0.162109375,
0.157958984375
],
[
0.353515625,
0.1602783203125,
0.64990234375,
0.160888671875,
0.64990234375,
0.17578125,
0.353515625,
0.174560546875
],
[
0.240966796875,
0.1776123046875,
0.765625,
0.1776123046875,
0.765625,
0.1923828125,
0.240966796875,
0.1923828125
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.6171875,
0.26123046875,
0.771484375,
0.26123046875,
0.771484375,
0.268310546875,
0.6171875,
0.268310546875
],
[
0.62109375,
0.270751953125,
0.67822265625,
0.270751953125,
0.67822265625,
0.27490234375,
0.62109375,
0.27490234375
],
[
0.145263671875,
0.279052734375,
0.45751953125,
0.279052734375,
0.45751953125,
0.292236328125,
0.145263671875,
0.292236328125
],
[
0.6708984375,
0.27490234375,
0.78271484375,
0.27490234375,
0.78271484375,
0.28515625,
0.6708984375,
0.28515625
],
[
0.615234375,
0.283935546875,
0.68994140625,
0.283935546875,
0.68994140625,
0.29150390625,
0.615234375,
0.29150390625
],
[
0.145263671875,
0.29345703125,
0.4609375,
0.29345703125,
0.4609375,
0.30712890625,
0.145263671875,
0.30712890625
],
[
0.64990234375,
0.292236328125,
0.7421875,
0.292236328125,
0.7421875,
0.302734375,
0.64990234375,
0.302734375
],
[
0.53564453125,
0.2998046875,
0.5751953125,
0.2998046875,
0.5751953125,
0.311279296875,
0.53564453125,
0.311279296875
],
[
0.8212890625,
0.298583984375,
0.85400390625,
0.298583984375,
0.85400390625,
0.31005859375,
0.8212890625,
0.31005859375
],
[
0.14697265625,
0.308837890625,
0.45751953125,
0.308837890625,
0.45751953125,
0.319580078125,
0.14697265625,
0.319580078125
],
[
0.52880859375,
0.311279296875,
0.580078125,
0.311279296875,
0.580078125,
0.322509765625,
0.52880859375,
0.322509765625
],
[
0.8076171875,
0.312255859375,
0.8671875,
0.312255859375,
0.8671875,
0.322509765625,
0.8076171875,
0.322509765625
],
[
0.145263671875,
0.320556640625,
0.459228515625,
0.320556640625,
0.459228515625,
0.334228515625,
0.145263671875,
0.334228515625
],
[
0.81689453125,
0.32666015625,
0.8486328125,
0.3349609375,
0.8369140625,
0.35986328125,
0.80419921875,
0.352783203125
],
[
0.0235137939453125,
0.3154296875,
0.060455322265625,
0.3154296875,
0.060455322265625,
0.439453125,
0.0235137939453125,
0.439453125
],
[
0.8369140625,
0.32666015625,
0.8623046875,
0.330078125,
0.85400390625,
0.369384765625,
0.82861328125,
0.366455078125
],
[
0.145263671875,
0.3349609375,
0.45751953125,
0.3349609375,
0.45751953125,
0.3486328125,
0.145263671875,
0.3486328125
],
[
0.1435546875,
0.349853515625,
0.459228515625,
0.350830078125,
0.459228515625,
0.364013671875,
0.1435546875,
0.363525390625
],
[
0.60791015625,
0.352783203125,
0.77099609375,
0.352783203125,
0.77099609375,
0.363525390625,
0.60791015625,
0.363525390625
],
[
0.603515625,
0.3623046875,
0.73876953125,
0.3623046875,
0.73876953125,
0.369384765625,
0.603515625,
0.369384765625
],
[
0.8076171875,
0.361083984375,
0.86572265625,
0.361083984375,
0.86572265625,
0.407958984375,
0.8076171875,
0.407958984375
],
[
0.61279296875,
0.37353515625,
0.76416015625,
0.37353515625,
0.76416015625,
0.38427734375,
0.61279296875,
0.38427734375
],
[
0.145263671875,
0.364013671875,
0.459228515625,
0.364013671875,
0.459228515625,
0.377685546875,
0.145263671875,
0.377685546875
],
[
0.14697265625,
0.380126953125,
0.45751953125,
0.380126953125,
0.45751953125,
0.39013671875,
0.14697265625,
0.39013671875
],
[
0.615234375,
0.3818359375,
0.78857421875,
0.38134765625,
0.78857421875,
0.391357421875,
0.615234375,
0.392578125
],
[
0.145263671875,
0.392578125,
0.459228515625,
0.392578125,
0.459228515625,
0.40625,
0.145263671875,
0.40625
],
[
0.615234375,
0.391357421875,
0.64990234375,
0.391357421875,
0.64990234375,
0.3984375,
0.615234375,
0.3984375
],
[
0.654296875,
0.40380859375,
0.66650390625,
0.40380859375,
0.66650390625,
0.40673828125,
0.654296875,
0.40673828125
],
[
0.8154296875,
0.3984375,
0.86572265625,
0.402587890625,
0.85791015625,
0.4453125,
0.8076171875,
0.441162109375
],
[
0.145263671875,
0.40673828125,
0.459228515625,
0.40673828125,
0.459228515625,
0.42041015625,
0.145263671875,
0.42041015625
],
[
0.0218353271484375,
0.42578125,
0.06378173828125,
0.42578125,
0.06378173828125,
0.51123046875,
0.0218353271484375,
0.51123046875
],
[
0.14697265625,
0.4228515625,
0.45751953125,
0.4228515625,
0.45751953125,
0.432861328125,
0.14697265625,
0.432861328125
],
[
0.145263671875,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.448974609375,
0.145263671875,
0.448974609375
],
[
0.145263671875,
0.449462890625,
0.459228515625,
0.449462890625,
0.459228515625,
0.463134765625,
0.145263671875,
0.463134765625
],
[
0.609375,
0.46142578125,
0.67919921875,
0.46142578125,
0.67919921875,
0.468505859375,
0.609375,
0.468505859375
],
[
0.67919921875,
0.457763671875,
0.75732421875,
0.460205078125,
0.75732421875,
0.47021484375,
0.67919921875,
0.468505859375
],
[
0.145263671875,
0.46435546875,
0.459228515625,
0.46435546875,
0.459228515625,
0.47802734375,
0.145263671875,
0.47802734375
],
[
0.1435546875,
0.476806640625,
0.459228515625,
0.47802734375,
0.459228515625,
0.4912109375,
0.1435546875,
0.490478515625
],
[
0.6533203125,
0.471435546875,
0.73876953125,
0.471435546875,
0.73876953125,
0.478515625,
0.6533203125,
0.478515625
],
[
0.61376953125,
0.482177734375,
0.70947265625,
0.482177734375,
0.70947265625,
0.4892578125,
0.61376953125,
0.4892578125
],
[
0.145263671875,
0.4921875,
0.459228515625,
0.4921875,
0.459228515625,
0.505859375,
0.145263671875,
0.505859375
],
[
0.51123046875,
0.49951171875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.51123046875,
0.51318359375
],
[
0.019317626953125,
0.51318359375,
0.060455322265625,
0.51318359375,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.14697265625,
0.50830078125,
0.45751953125,
0.50830078125,
0.45751953125,
0.5185546875,
0.14697265625,
0.5185546875
],
[
0.51123046875,
0.51416015625,
0.8818359375,
0.51416015625,
0.8818359375,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.145263671875,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53369140625,
0.145263671875,
0.53369140625
],
[
0.509765625,
0.52783203125,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.509765625,
0.54150390625
],
[
0.145263671875,
0.53515625,
0.45751953125,
0.53515625,
0.45751953125,
0.548828125,
0.145263671875,
0.548828125
],
[
0.51123046875,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.51123046875,
0.5556640625
],
[
0.145263671875,
0.548828125,
0.45751953125,
0.548828125,
0.45751953125,
0.5625,
0.145263671875,
0.5625
],
[
0.509765625,
0.55712890625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57080078125,
0.509765625,
0.57080078125
],
[
0.145263671875,
0.56396484375,
0.45751953125,
0.56396484375,
0.45751953125,
0.57470703125,
0.145263671875,
0.57470703125
],
[
0.51123046875,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.51123046875,
0.5849609375
],
[
0.145263671875,
0.5791015625,
0.45751953125,
0.5791015625,
0.45751953125,
0.5888671875,
0.145263671875,
0.5888671875
],
[
0.145263671875,
0.59033203125,
0.45751953125,
0.59033203125,
0.45751953125,
0.60400390625,
0.145263671875,
0.60400390625
],
[
0.51123046875,
0.5859375,
0.87744140625,
0.5859375,
0.87744140625,
0.599609375,
0.51123046875,
0.599609375
],
[
0.145263671875,
0.60546875,
0.459228515625,
0.60546875,
0.459228515625,
0.619140625,
0.145263671875,
0.619140625
],
[
0.14697265625,
0.62158203125,
0.456787109375,
0.62158203125,
0.456787109375,
0.6318359375,
0.14697265625,
0.6318359375
],
[
0.51123046875,
0.626953125,
0.880859375,
0.626953125,
0.880859375,
0.64013671875,
0.51123046875,
0.64013671875
],
[
0.1435546875,
0.6328125,
0.459228515625,
0.63427734375,
0.459228515625,
0.6484375,
0.1435546875,
0.64794921875
],
[
0.145263671875,
0.64794921875,
0.459228515625,
0.6484375,
0.459228515625,
0.662109375,
0.145263671875,
0.6611328125
],
[
0.51123046875,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.51123046875,
0.65771484375
],
[
0.145263671875,
0.6630859375,
0.459228515625,
0.6630859375,
0.459228515625,
0.6767578125,
0.145263671875,
0.6767578125
],
[
0.51123046875,
0.6611328125,
0.7314453125,
0.6611328125,
0.7314453125,
0.6748046875,
0.51123046875,
0.6748046875
],
[
0.145263671875,
0.6767578125,
0.39794921875,
0.6767578125,
0.39794921875,
0.6904296875,
0.145263671875,
0.6904296875
],
[
0.53076171875,
0.677734375,
0.88330078125,
0.677734375,
0.88330078125,
0.69140625,
0.53076171875,
0.69140625
],
[
0.51123046875,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.11419677734375,
0.70068359375,
0.26025390625,
0.70166015625,
0.26025390625,
0.71630859375,
0.11419677734375,
0.71533203125
],
[
0.51318359375,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.72265625,
0.51318359375,
0.72265625
],
[
0.1175537109375,
0.72705078125,
0.490234375,
0.72705078125,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.7392578125,
0.509765625,
0.7392578125
],
[
0.1175537109375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.5087890625,
0.7412109375,
0.88330078125,
0.7421875,
0.88330078125,
0.75732421875,
0.5087890625,
0.755859375
],
[
0.1175537109375,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.1158447265625,
0.775390625,
0.490234375,
0.775390625,
0.490234375,
0.7890625,
0.1158447265625,
0.7890625
],
[
0.509765625,
0.77490234375,
0.8818359375,
0.77490234375,
0.8818359375,
0.7880859375,
0.509765625,
0.7880859375
],
[
0.1175537109375,
0.7919921875,
0.489501953125,
0.7919921875,
0.489501953125,
0.8056640625,
0.1175537109375,
0.8056640625
],
[
0.51123046875,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.802734375,
0.51123046875,
0.802734375
],
[
0.1175537109375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.80712890625,
0.88330078125,
0.80712890625,
0.88330078125,
0.82080078125,
0.509765625,
0.82080078125
],
[
0.1175537109375,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.509765625,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.509765625,
0.8359375
],
[
0.1158447265625,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.1158447265625,
0.85498046875
],
[
0.509765625,
0.8388671875,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1158447265625,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.5087890625,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.869140625,
0.5087890625,
0.869140625
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.5087890625,
0.87158203125,
0.880859375,
0.87158203125,
0.880859375,
0.88525390625,
0.5087890625,
0.88525390625
],
[
0.1158447265625,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.7421875,
0.88818359375,
0.7421875,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.52880859375,
0.90380859375,
0.8818359375,
0.90283203125,
0.8818359375,
0.91748046875,
0.52880859375,
0.91845703125
]
] |
[
"Aerial Vision-and-Dialog Navigation",
"Yue Fan, Winson Chen, Tongzhou Jiang, Chun Zhou, Yi Zhang, Xin Eric Wang",
"University of California, Santa Cruz",
"(yfan71, wchen157, tojiang, czhou43, yiz, xwang366)@ucsc.edu",
"Abstract",
"Heydrone.Godirectlyeastacross on",
"oadthennass",
"The ability to converse with humans and follow",
"nd. Destination is a building",
"withayellowroof.",
"natural language commands is crucial for intel-",
"Agent beginsnavigation",
"Agent",
"User",
"ligent unmanned aerial vehicles (a.k.a. drones)",
"(Follower)",
"(Commander)",
"It can relieve people's burden of holding a con-",
"1",
"",
"",
"troller all the time, allow multitasking, and",
"make drone control more accessible for people",
"I see multiple yellow buildings. Can I see",
"the destination?Howtoget here",
"亨骨",
"Yes,The destination is next adjacent",
"with disabilities or with their hands occupied",
"To this end, we introduce Aerial Vision-and",
"building at your position.Please go a ltte",
"Dialog Navigation (AVDN), to navigate a drone",
"at4o'cl",
"Han",
"i",
"via natural language conversation. We build",
"['S]",
"a drone simulator with a continuous photore-",
"alistic environment and collect a new AVDN",
"dataset of over 3k recorded navigation trajecto-",
"NowIthinkIhave",
"arrived.AmIright",
"ries with asynchronous human-human dialogs",
"between commanders and followers. The com-",
"Agentstopsnavigation",
"Yes.Youhavearrived!",
"mander provides initial navigation instruction",
"Figure 1: An example of Aerial Vision-and-Dialog Nav-",
"C612152:A1",
"and further guidance by request, while the",
"igation (AVDN). The user instructs the agent to fly to",
"follower navigates the drone in the simulator",
"a destination. During the navigation, the agent can ask",
"and asks questions when needed. During data",
"questions while showing the images of past visual obser-",
"collection, followers' attention on the drone's",
"vations and relative trajectory. The user will talk back",
"visual observation is also recorded.Based",
"at a convenient time to provide further guidance to the",
"on the AVDN dataset,we studythe tasks of",
"aerial navigation from (full) dialog history and",
"agent without having to monitor the agent all the time.",
"propose an effective Human Attention Aided",
"Transformermodel(HAA-Transformer),which",
"drone control for users with some disabilities and",
"learns to predict both navigation waypoints",
"and human attention.Dataset and code are re-",
"who have their hands occupied by activities such",
"leased:https://sites.google.com/view/",
"as taking photos, writing, etc.",
"aerial-vision-and-dialog/home.",
"Therefore, this work introduces Aerial Vision-",
"and-Dialog Navigation (AVDN), aiming to develop",
"1 Introduction",
"anintelligent drone that can converse with its user",
"Drones have been widely adopted for many applica-",
"to fly to the expected destination. As shown in Fig",
"tions in our daily life, from personal entertainment",
"ure 1, the user (commander) provides instructions,",
"to professional use. It has the advantage of mobility",
"and the aerial agent (follower) follows the instruc-",
"and observing large areas over ground robots. How-",
"tion and asks questions when needed. The past",
"ever, compared with ground robots, the control of",
"visual trajectoriesare also provided along withthe",
"the aerial robot is more complex because an extra",
"question, which frees the commander from monitor-",
"degree of freedom, altitude, is involved. To control",
"ing the drone all the time and minimizes the burden",
"a drone, people often need to hold a controller all",
" of drone control. In this free-form dialog, potential",
"the time, so it is essential to create a hands-free",
" ambiguities in the instruction can be gradually re-",
"control experience for drone users and develop an",
" solved through the further instructions provided by",
"intelligent drone that can complete tasks simply",
"the commanders upon request.",
"by talking to humans. It can lower the barrier of",
"To implement and evaluate the AVDN task, we"
] |
{
"arxiv_id": "2205.12219",
"language": "en",
"page_idx": 0
}
|
||
2205.12219-en-1
|
2205.12219-en-1.jpg
|
build a photorealistic simulator with continuous state space to simulate a drone flying with its onboard camera pointing straight downward. Then we collect an AVDN dataset of 3,064 aerial navigation trajectories with human-human dialogs, where crowd-sourcing workers play the commander role and drone experts play the follower role, as illustrated in Figure 1. Moreover, we also collect the attention of human followers over the aerial scenes for a better understanding of where humans ground navigation instructions.
Based on our AVDN dataset, we introduce two challenging navigation tasks, Aerial Navigation from Dialog History (ANDH) and Aerial Navigation from Full Dialog History (ANDH-Full). Both tasks focus on predicting navigation actions that can lead the agent to the destination area, whereas the difference is that ANDH-Full presents the agent with full dialog and requires it to reach the final destination (Kim et al., 2021), while ANDH evaluates the agent's completion of the sub-trajectory within a dialog round given the previous dialog information (Thomason et al., 2020).
The proposed tasks open new challenges of sequential action prediction in a large continuous space and natural language grounding on photorealistic aerial scenes. We propose a sequenceto-sequence Human Attention Aided Transformer model (HAA-Transformer) for both tasks. The HAA-Transformer model predicts waypoints to reduce the complexity of the search space and learns to stop at the desired location. More importantly, it is jointly trained to predict human attention from the input dialog and visual observations and learns where to look during inference. Experiments on our AVDN dataset show that multitask learning is beneficial and human attention prediction improves navigation performance. The main contributions are concluded as follows:
$\cdot$ We create a new dataset and simulator for aerial vision-and-dialog navigation. The dataset includes over 3K aerial navigation trajectories with human-human dialogs.
· We introduce ANDH and ANDH-Full tasks to evaluate the agent's ability to understand natural language dialog, reason about aerial scenes, and navigate to the target location in a continuous photorealistic aerial environment.
· We propose an HAA-Transformer model as
the baseline for ANDH and ANDH-Full. Be-
sides predicting the waypoint navigation actions, HAA-Transformer also learns to predict the attention of the human follower along the navigation trajectory. Experiments on our AVDN dataset validate the effectiveness of the HAA-Transformer model.
## 2 $\textbf{Related work}$
Vision-and-Language Navigation Vision-andLanguage Navigation (VLN) is an emerging multimodal task that studies the problem of using both language instructions and visual observation to predict navigation actions. We compare some of the works with our AVDN dataset in Table 1. Early VLN datasets such as Anderson et al. (2018); Ku et al. (2020) start with the indoor house environments in the Matterport3D simulator (Chang et al., 2017), where the visual scenes are connected on a navigation graph. To simulate continuous state change as in the real world, Krantz et al. (2020) built a 3D continuous environment by reconstructing the scene based on topological connections where the agent uses continuous actions during the navigation. Some other VLN studies focus on language instructions. Nguyen et al. (2019); Nguyen and Daumé III (2019); Thomason et al. (2020) created datasets where the agent can interact with the user by sending fixed signals or having dialogs. There are also works on synthetic indoor environments, such as Shridhar et al. (2020b); Padmakumar et al. (2021) that use an interactive simulation environment with synthetic views named ALFRED, where the agent needs to follow language instructions or dialogs to finish household tasks. Besides the indoor environment, some VLN datasets work on the more complex outdoor environment, such as the Touchdown dataset (Chen et al., 2019) and the modified LANI dataset (Misra et al., 2018). Blukis et al. (2019) is similar to ours for both using drones. However, the synthetic environment used has a gap from the realistic scene, and they ignored the control of the drone's altitude, where such navigation is oversimplified and has a large gap towards navigation in the real world in terms of language and vision aspects. Our work absorbs the advantage from previous works where we have continuous environments and dialog instructions to better approximate the real-world scenario.
Aerial Navigation Using both vision and lan-
guage for aerial navigation is a less studied topic, whereas vision-only aerial navigation for drones is
|
[
[
0.1158447265625,
0.08489990234375,
0.486083984375,
0.08489990234375,
0.486083984375,
0.09857177734375,
0.1158447265625,
0.09857177734375
],
[
0.5498046875,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.5498046875,
0.09979248046875
],
[
0.1158447265625,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.1158447265625,
0.11639404296875
],
[
0.54833984375,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.54833984375,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.5498046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.5498046875,
0.1318359375
],
[
0.1158447265625,
0.1341552734375,
0.48779296875,
0.1341552734375,
0.48779296875,
0.1478271484375,
0.1158447265625,
0.1478271484375
],
[
0.54833984375,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.54833984375,
0.1478271484375
],
[
0.1175537109375,
0.15087890625,
0.483642578125,
0.15087890625,
0.483642578125,
0.160888671875,
0.1175537109375,
0.160888671875
],
[
0.54833984375,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.163330078125,
0.54833984375,
0.163330078125
],
[
0.1158447265625,
0.166259765625,
0.486083984375,
0.1650390625,
0.486083984375,
0.179931640625,
0.1158447265625,
0.18115234375
],
[
0.54833984375,
0.166259765625,
0.744140625,
0.166259765625,
0.744140625,
0.179931640625,
0.54833984375,
0.179931640625
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.51708984375,
0.198974609375,
0.52490234375,
0.198974609375,
0.52490234375,
0.2037353515625,
0.51708984375,
0.2037353515625
],
[
0.5400390625,
0.193603515625,
0.66015625,
0.193603515625,
0.66015625,
0.2078857421875,
0.5400390625,
0.2078857421875
],
[
0.11419677734375,
0.21435546875,
0.489501953125,
0.213134765625,
0.489501953125,
0.22802734375,
0.11419677734375,
0.2286376953125
],
[
0.51318359375,
0.2197265625,
0.8818359375,
0.2197265625,
0.8818359375,
0.2327880859375,
0.51318359375,
0.2327880859375
],
[
0.1175537109375,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.5087890625,
0.2340087890625,
0.8818359375,
0.235107421875,
0.8818359375,
0.2493896484375,
0.5087890625,
0.248779296875
],
[
0.1158447265625,
0.2476806640625,
0.28955078125,
0.2476806640625,
0.28955078125,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.509765625,
0.251708984375,
0.880859375,
0.251708984375,
0.880859375,
0.265380859375,
0.509765625,
0.265380859375
],
[
0.1351318359375,
0.26318359375,
0.48779296875,
0.26318359375,
0.48779296875,
0.276611328125,
0.1351318359375,
0.276611328125
],
[
0.51318359375,
0.268310546875,
0.880859375,
0.268310546875,
0.880859375,
0.281982421875,
0.51318359375,
0.281982421875
],
[
0.1175537109375,
0.279052734375,
0.48779296875,
0.279052734375,
0.48779296875,
0.29345703125,
0.1175537109375,
0.29345703125
],
[
0.509765625,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.509765625,
0.297607421875
],
[
0.1175537109375,
0.29443359375,
0.490234375,
0.29443359375,
0.490234375,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51123046875,
0.298583984375,
0.880859375,
0.2998046875,
0.880859375,
0.314208984375,
0.51123046875,
0.3134765625
],
[
0.1175537109375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51123046875,
0.3154296875,
0.880859375,
0.3154296875,
0.880859375,
0.328857421875,
0.51123046875,
0.328857421875
],
[
0.1158447265625,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.1158447265625,
0.341552734375
],
[
0.51318359375,
0.33203125,
0.88330078125,
0.33203125,
0.88330078125,
0.345703125,
0.51318359375,
0.345703125
],
[
0.1158447265625,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.1158447265625,
0.355712890625
],
[
0.509765625,
0.3486328125,
0.88330078125,
0.3486328125,
0.88330078125,
0.3623046875,
0.509765625,
0.3623046875
],
[
0.1158447265625,
0.358154296875,
0.48779296875,
0.359375,
0.48779296875,
0.37353515625,
0.1158447265625,
0.372314453125
],
[
0.5087890625,
0.363525390625,
0.88330078125,
0.365234375,
0.88330078125,
0.37890625,
0.5087890625,
0.377197265625
],
[
0.1158447265625,
0.374755859375,
0.486083984375,
0.374755859375,
0.486083984375,
0.388427734375,
0.1158447265625,
0.388427734375
],
[
0.509765625,
0.38134765625,
0.8818359375,
0.38134765625,
0.8818359375,
0.394287109375,
0.509765625,
0.394287109375
],
[
0.1175537109375,
0.391357421875,
0.490234375,
0.391357421875,
0.490234375,
0.405029296875,
0.1175537109375,
0.405029296875
],
[
0.509765625,
0.396728515625,
0.880859375,
0.396728515625,
0.880859375,
0.410400390625,
0.509765625,
0.410400390625
],
[
0.11419677734375,
0.40673828125,
0.48779296875,
0.407958984375,
0.48779296875,
0.4228515625,
0.11419677734375,
0.421630859375
],
[
0.51123046875,
0.413330078125,
0.88330078125,
0.413330078125,
0.88330078125,
0.427001953125,
0.51123046875,
0.427001953125
],
[
0.1175537109375,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.509765625,
0.4287109375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.509765625,
0.4423828125
],
[
0.1158447265625,
0.440673828125,
0.3896484375,
0.440673828125,
0.3896484375,
0.45361328125,
0.1158447265625,
0.45361328125
],
[
0.5146484375,
0.446533203125,
0.880859375,
0.446533203125,
0.880859375,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.1351318359375,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.4697265625,
0.1351318359375,
0.4697265625
],
[
0.51123046875,
0.46142578125,
0.8818359375,
0.46142578125,
0.8818359375,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.1175537109375,
0.47265625,
0.486083984375,
0.47265625,
0.486083984375,
0.486328125,
0.1175537109375,
0.486328125
],
[
0.51318359375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.1175537109375,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.5029296875,
0.1175537109375,
0.5029296875
],
[
0.51123046875,
0.493408203125,
0.880859375,
0.493408203125,
0.880859375,
0.50732421875,
0.51123046875,
0.50732421875
],
[
0.1158447265625,
0.5048828125,
0.490234375,
0.5048828125,
0.490234375,
0.5185546875,
0.1158447265625,
0.5185546875
],
[
0.51123046875,
0.51025390625,
0.880859375,
0.51025390625,
0.880859375,
0.52392578125,
0.51123046875,
0.52392578125
],
[
0.1175537109375,
0.521484375,
0.48779296875,
0.521484375,
0.48779296875,
0.5322265625,
0.1175537109375,
0.5322265625
],
[
0.509765625,
0.525390625,
0.88330078125,
0.525390625,
0.88330078125,
0.5390625,
0.509765625,
0.5390625
],
[
0.1175537109375,
0.5361328125,
0.489501953125,
0.5361328125,
0.489501953125,
0.5498046875,
0.1175537109375,
0.5498046875
],
[
0.509765625,
0.54150390625,
0.88330078125,
0.54150390625,
0.88330078125,
0.5546875,
0.509765625,
0.5546875
],
[
0.1158447265625,
0.5517578125,
0.489501953125,
0.552734375,
0.489501953125,
0.56787109375,
0.1158447265625,
0.56640625
],
[
0.51318359375,
0.55712890625,
0.88330078125,
0.55712890625,
0.88330078125,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.118408203125,
0.57080078125,
0.484375,
0.57080078125,
0.484375,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.51318359375,
0.57373046875,
0.88330078125,
0.57373046875,
0.88330078125,
0.58740234375,
0.51318359375,
0.58740234375
],
[
0.1175537109375,
0.58447265625,
0.484375,
0.58447265625,
0.484375,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.51123046875,
0.59130859375,
0.8818359375,
0.59130859375,
0.8818359375,
0.6015625,
0.51123046875,
0.6015625
],
[
0.118408203125,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.61328125,
0.118408203125,
0.61328125
],
[
0.509765625,
0.60546875,
0.88330078125,
0.60546875,
0.88330078125,
0.619140625,
0.509765625,
0.619140625
],
[
0.1158447265625,
0.6162109375,
0.486083984375,
0.6162109375,
0.486083984375,
0.6298828125,
0.1158447265625,
0.6298828125
],
[
0.5087890625,
0.62158203125,
0.8818359375,
0.62060546875,
0.8818359375,
0.63525390625,
0.5087890625,
0.6357421875
],
[
0.1158447265625,
0.6318359375,
0.486083984375,
0.6318359375,
0.486083984375,
0.6455078125,
0.1158447265625,
0.6455078125
],
[
0.509765625,
0.63818359375,
0.8818359375,
0.63818359375,
0.8818359375,
0.65185546875,
0.509765625,
0.65185546875
],
[
0.1158447265625,
0.64794921875,
0.489501953125,
0.6484375,
0.489501953125,
0.6630859375,
0.1158447265625,
0.662109375
],
[
0.51123046875,
0.65625,
0.880859375,
0.65625,
0.880859375,
0.66650390625,
0.51123046875,
0.66650390625
],
[
0.11419677734375,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.11419677734375,
0.677734375
],
[
0.5087890625,
0.66943359375,
0.880859375,
0.66943359375,
0.880859375,
0.68310546875,
0.5087890625,
0.68310546875
],
[
0.1158447265625,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.51123046875,
0.68505859375,
0.880859375,
0.68505859375,
0.880859375,
0.69873046875,
0.51123046875,
0.69873046875
],
[
0.11419677734375,
0.69677734375,
0.306396484375,
0.69677734375,
0.306396484375,
0.71044921875,
0.11419677734375,
0.71044921875
],
[
0.51123046875,
0.70068359375,
0.8818359375,
0.703125,
0.8818359375,
0.71630859375,
0.51123046875,
0.71435546875
],
[
0.1368408203125,
0.72265625,
0.489501953125,
0.7216796875,
0.489501953125,
0.736328125,
0.1368408203125,
0.73681640625
],
[
0.509765625,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.509765625,
0.732421875
],
[
0.15283203125,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.15283203125,
0.7529296875
],
[
0.509765625,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.509765625,
0.7490234375
],
[
0.51123046875,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.51123046875,
0.76318359375
],
[
0.15283203125,
0.755859375,
0.491943359375,
0.755859375,
0.491943359375,
0.76953125,
0.15283203125,
0.76953125
],
[
0.151123046875,
0.77099609375,
0.429931640625,
0.771484375,
0.429931640625,
0.7861328125,
0.151123046875,
0.78515625
],
[
0.5146484375,
0.7685546875,
0.87890625,
0.7685546875,
0.87890625,
0.779296875,
0.5146484375,
0.779296875
],
[
0.51123046875,
0.783203125,
0.8818359375,
0.78076171875,
0.8818359375,
0.7958984375,
0.51123046875,
0.79736328125
],
[
0.1376953125,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.81103515625,
0.1376953125,
0.81103515625
],
[
0.51318359375,
0.80029296875,
0.880859375,
0.80029296875,
0.880859375,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.151123046875,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.151123046875,
0.82763671875
],
[
0.509765625,
0.81640625,
0.880859375,
0.81640625,
0.880859375,
0.830078125,
0.509765625,
0.830078125
],
[
0.15283203125,
0.830078125,
0.48779296875,
0.830078125,
0.48779296875,
0.84423828125,
0.15283203125,
0.84423828125
],
[
0.5087890625,
0.83203125,
0.8818359375,
0.83203125,
0.8818359375,
0.845703125,
0.5087890625,
0.845703125
],
[
0.151123046875,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.85986328125,
0.151123046875,
0.85986328125
],
[
0.509765625,
0.84716796875,
0.82958984375,
0.84716796875,
0.82958984375,
0.8623046875,
0.509765625,
0.8623046875
],
[
0.15283203125,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.15283203125,
0.876953125
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1368408203125,
0.88525390625,
0.490234375,
0.8857421875,
0.490234375,
0.90380859375,
0.1368408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.15283203125,
0.90380859375,
0.491943359375,
0.90380859375,
0.491943359375,
0.91748046875,
0.15283203125,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"build a photorealistic simulator with continuous",
"sides predicting the waypoint navigation ac-",
" state space to simulate a drone flying with its on-",
"tions, HAA-Transformer also learns to pre-",
"board camera pointing straight downward. Then",
"dict the attention of the human follower along",
"we collect an AVDN dataset of 3064 aerial naviga-",
"the navigation trajectory. Experiments on our",
"tion traiectories with human-human dialogs. where",
"AVDN dataset validate the effectiveness of the",
"crowd-sourcing workers play the commander role",
"HAA-Transformer model.",
"and drone experts play the follower role, as illus-",
"trated in Figure 1. Moreover, we also collect the",
"",
"Related work",
"attention of human followers over the aerial scenes",
"Vision-and-Language Navigation Vision-and-",
"for a better understanding of where humans ground",
"Language Navigation (VLN) is an emerging multi-",
"navigation instructions.",
"modal task that studies the problem of using both",
"Based on our AVDN dataset, we introduce two",
"language instructions and visual observation to pre",
"challenging navigation tasks, Aerial Navigation",
"dict navigation actions. We compare some of the",
"from Dialog History (ANDH) and Aerial Naviga-",
"works with our AVDN dataset in Table 1. Early",
"tion from Full Dialog History (ANDH-Full). Both",
"VLN datasets such as Anderson et al. (2018): Ku",
"tasks focus on predicting navigation actions that",
"et al. (2020) start with the indoor house environ-",
"can lead the agent to the destination area, whereas",
"ments in the Matterport3D simulator (Chang et al.,",
"the difference is that ANDH-Full presents the agent",
"2017), where the visual scenes are connected on",
"with full dialog and requires it to reach the final",
"a navigation graph. To simulate continuous state",
"destination (Kim et al., 2021), while ANDH eval-",
"change as in the real world, Krantz et al. (2020)",
"uates the agent's completion of the sub-trajectory",
"built a 3D continuous environment by reconstruct-",
"within a dialog round given the previous dialog",
"ing the scene based on topological connections",
"information (Thomason et al., 2020).",
"where the agent uses continuous actions during the",
"The proposed tasks open new challenges of se-",
"navigation. Some other VLN studies focus on lan",
"quential action prediction in a large continuous",
"guage instructions. Nguyen et al. (2019); Nguyen",
"space and natural language grounding on photo-",
"and Daumé I1I (2019); Thomason et al. (2020) cre-",
"realistic aerial scenes. We propose a sequence-",
"ated datasets where the agent can interact with the",
"to-sequenceHumanAttention AidedTransformer",
"user by sending fixed signals or having dialogs.",
"model (HAA-Transformer) for both tasks. The",
"There arealso works on synthetic indoor environ-",
"HAA-Transformer model predicts waypoints to re-",
"ments, such as Shridhar et al. (2020b); Padmaku-",
"ducethecomplexityofthe searchspaceandlearns",
"mar et al. (2021) that use an interactive simulation",
"to stop at the desired location. More importantly, it",
"environmentwithsyntheticviewsnamedALFRED.",
"is jointly trained to predict human attentionfrom",
" where the agent needs to follow language instruc-",
"the input dialog and visual observations and learns",
"tions or dialogs to finish household tasks. Besides",
"where to look during inference. Experiments on",
"the indoor environment, some VLN datasets work",
"our AVDN dataset show that multitask learning is",
"onthemore complex outdoorenvironment.such",
" beneficial and human attention prediction improves",
"as the Touchdown dataset (Chen et al., 2019) and",
"navigation performance. The main contributions",
"the modified LANI dataset (Misra et al., 2018)",
"are concluded as follows:",
"Blukis et al. (2019) is similar to ours for both us-",
": We create a new dataset and simulator for",
"ing drones. However, the synthetic environment",
"aerial vision-and-dialog navigation. The",
"used has a gap from the realistic scene, and they",
"ignored the control of the drone's altitude, where",
"dataset includes over 3K aerial navigation tra-",
"jectories with human-human dialogs.",
"suchnavigation isoversimplifiedandhasalarge",
"gap towards navigation in the real world in terms",
". We introduce ANDH and ANDH-Full tasks",
"of languageandvisionaspects.Ourworkabsorbs",
"to evaluate the agent's ability to understand",
"the advantage from previous works where we have",
"natural language dialog, reason about aerial",
" continuous environments and dialog instructions to",
"scenes, and navigate to the target location in a",
"better approximate the real-world scenario.",
"continuous photorealistic aerial environment.",
"Aerial Navigation Using both vision and lan-",
". We propose an HAA-Transformer model as",
"guage for aerial navigation is a less studied topic,",
"the baseline for ANDH and ANDH-Full. Be-",
"whereas vision-only aerial navigation for drones is"
] |
{
"arxiv_id": "2205.12219",
"language": "en",
"page_idx": 1
}
|
||
2205.12219-en-2
|
2205.12219-en-2.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Env</th>
<th>Photorealistic</th>
<th>Continuous Space</th>
<th>Dialog</th>
<th>Free Form</th>
</tr>
<tr>
<td>$\mathbf{R}2\mathbf{R},\mathbf{R}$xR</td>
<td>indoor</td>
<td>$√$</td>
<td>$X$</td>
<td>$X$</td>
<td>$√$</td>
</tr>
<tr>
<td>VNLA, HANNA</td>
<td>indoor</td>
<td>$√$</td>
<td>$X$</td>
<td>$√$</td>
<td>$x$</td>
</tr>
<tr>
<td>$\mathbf{VLN-CE}$</td>
<td>indoor</td>
<td>$X$</td>
<td>$√$</td>
<td>$X$</td>
<td>$\checkmark$</td>
</tr>
<tr>
<td>CVDN</td>
<td>indoor</td>
<td>$√$</td>
<td>$X$</td>
<td>$√$</td>
<td>$\checkmark$</td>
</tr>
<tr>
<td>ALFRED</td>
<td>indoor</td>
<td>$X$</td>
<td>$X$</td>
<td>$X$</td>
<td>$\checkmark$</td>
</tr>
<tr>
<td>DialFRED</td>
<td>indoor</td>
<td>$X$</td>
<td>$X$</td>
<td>$√$</td>
<td>$x$</td>
</tr>
<tr>
<td>TEACh</td>
<td>indoor</td>
<td>$X$</td>
<td>$√$</td>
<td>$√$</td>
<td>$\checkmark$</td>
</tr>
<tr>
<td>TouchDown modified LANI</td>
<td>street-view aerial</td>
<td>$√$ $V$ $X$</td>
<td>$X$ $√$</td>
<td>$X$ $X$</td>
<td>$\checkmark$ $√$</td>
</tr>
<tr>
<td>AVDN( ours) </td>
<td>aerial</td>
<td>$√$</td>
<td>$√$</td>
<td>$√$</td>
<td>$\checkmark$</td>
</tr>
</tbody>
</table>
Table 1: Example Vision-and-Language Navigation
Datasets. R2R (Anderson et al., 2018), RxR (Ku et al., 2020), VNLA (Nguyen et al., 2019), HANNA (Nguyen and Daumé III, 2019), VLN-CE (Krantz et al., 2020), CVDN (Thomason et al., 2020), ALFRED (Shridhar et al., 2020a), DialFRED (Gao et al., 2022), TEACh (Padmakumar et al., 2021), TouchDown (Chen et al., 2019), modifed LANI (Blukis et al., 2019).
<FigureHere>
Figure 2: Example of a trajectory in our AVDN dataset.
On the left, the commander's turn and the follower's turn alternate in chronological order. In each turn, dialog utterances are shown, and the follower's turn also shows the navigation process that spans from time step $T$ to $T+$ 1, including the follower's observation and attention. On the right, there are trajectory overviews at different time steps. More examples can be found in the Appendix
already an active topic in the field. Some inspiring works (Loquercio et al., 2018; Giusti et al., 2015; Smolyanskiy et al., 2017; Fan et al., 2020; Bozcan and Kayacan, 2020; Majdik et al., 2017; Kang et al., 2019) worked on using pre-collected realworld drone data to tackle aerial vision navigation problems. Due to the hardness of collecting data and the risk of crashes, some other works applied simulation for aerial navigation (Chen et al., 2018; Shah et al., 2017; Chen et al., 2020), where rich ground truths are provided without the need for annotation. However, the modality of language is missing in these prior works and as a result, the navigation tasks only contain simple goals. As for the aerial vision-and-language navigation task in this work, the navigation is guided by natural dialog. As a result, it allows more diverse and complex navigation and also resolves ambiguities during complicated navigation.
## 3 Dataset
The AVDN dataset includes dialogs, navigation trajectories, and the drone's visual observation with human attention, where an example is shown in Figure 2. With the help of a newly proposed simulator, we record the AVDN trajectories created by two groups of humans interacting with each other, playing either the commander role or the follower role. Our AVDN dataset is the first aerial navigation dataset based on dialogs to the best of our knowledge.
# 3.1 Simulator
We build a simulator to simulate the drone with a top-down view area. Our simulation environment is a continuous space so that the simulated drone can move continuously to any point within the environment. The drone's visual observations are square images generated corresponding to the drone's view area by cropping from high-resolution satellite images in the xView dataset (Lam et al., 2018), an open-source large-scale satellite image object detection dataset. In this way, our simulator is capable of providing continuous frames with rich visual features. We also design an interface for our simulator, where the simulated drone can be controlled with a keyboard and the drone's visual observation will be displayed in real-time with a digital compass. During the control, users can also
|
[
[
0.33154296875,
0.07781982421875,
0.483642578125,
0.09442138671875,
0.478515625,
0.1134033203125,
0.32666015625,
0.0968017578125
],
[
0.447509765625,
0.08612060546875,
0.47607421875,
0.08612060546875,
0.47607421875,
0.097412109375,
0.447509765625,
0.097412109375
],
[
0.51904296875,
0.08612060546875,
0.60107421875,
0.08612060546875,
0.60107421875,
0.09320068359375,
0.51904296875,
0.09320068359375
],
[
0.2216796875,
0.0914306640625,
0.2509765625,
0.0914306640625,
0.2509765625,
0.10272216796875,
0.2216796875,
0.10272216796875
],
[
0.12005615234375,
0.0914306640625,
0.16455078125,
0.0914306640625,
0.16455078125,
0.10272216796875,
0.12005615234375,
0.10272216796875
],
[
0.26025390625,
0.0914306640625,
0.3359375,
0.0914306640625,
0.3359375,
0.1015625,
0.26025390625,
0.1015625
],
[
0.3525390625,
0.0968017578125,
0.3896484375,
0.0968017578125,
0.3896484375,
0.1092529296875,
0.3525390625,
0.1092529296875
],
[
0.54248046875,
0.09979248046875,
0.55908203125,
0.09979248046875,
0.55908203125,
0.10394287109375,
0.54248046875,
0.10394287109375
],
[
0.5751953125,
0.097412109375,
0.61376953125,
0.097412109375,
0.61376953125,
0.1051025390625,
0.5751953125,
0.1051025390625
],
[
0.12005615234375,
0.11102294921875,
0.175537109375,
0.11102294921875,
0.175537109375,
0.122314453125,
0.12005615234375,
0.122314453125
],
[
0.214111328125,
0.11224365234375,
0.252685546875,
0.11224365234375,
0.252685546875,
0.12353515625,
0.214111328125,
0.12353515625
],
[
0.291259765625,
0.1134033203125,
0.3046875,
0.1134033203125,
0.3046875,
0.12469482421875,
0.291259765625,
0.12469482421875
],
[
0.364501953125,
0.1134033203125,
0.376220703125,
0.1134033203125,
0.376220703125,
0.12469482421875,
0.364501953125,
0.12469482421875
],
[
0.418212890625,
0.1134033203125,
0.429931640625,
0.1134033203125,
0.429931640625,
0.12469482421875,
0.418212890625,
0.12469482421875
],
[
0.45751953125,
0.11517333984375,
0.46337890625,
0.11517333984375,
0.46337890625,
0.1217041015625,
0.45751953125,
0.1217041015625
],
[
0.291259765625,
0.122314453125,
0.3046875,
0.122314453125,
0.3046875,
0.1341552734375,
0.291259765625,
0.1341552734375
],
[
0.418212890625,
0.122314453125,
0.429931640625,
0.122314453125,
0.429931640625,
0.1341552734375,
0.418212890625,
0.1341552734375
],
[
0.359375,
0.1258544921875,
0.372802734375,
0.1217041015625,
0.38037109375,
0.1318359375,
0.366943359375,
0.135986328125
],
[
0.12005615234375,
0.12353515625,
0.25439453125,
0.12353515625,
0.25439453125,
0.1341552734375,
0.12005615234375,
0.1341552734375
],
[
0.45751953125,
0.1258544921875,
0.465087890625,
0.1258544921875,
0.465087890625,
0.133056640625,
0.45751953125,
0.133056640625
],
[
0.12005615234375,
0.1341552734375,
0.1695556640625,
0.1341552734375,
0.1695556640625,
0.1455078125,
0.12005615234375,
0.1455078125
],
[
0.2149658203125,
0.1341552734375,
0.25439453125,
0.1341552734375,
0.25439453125,
0.1455078125,
0.2149658203125,
0.1455078125
],
[
0.5205078125,
0.12890625,
0.58447265625,
0.12890625,
0.58447265625,
0.135986328125,
0.5205078125,
0.135986328125
],
[
0.291259765625,
0.1353759765625,
0.3046875,
0.1353759765625,
0.3046875,
0.1455078125,
0.291259765625,
0.1455078125
],
[
0.364501953125,
0.1353759765625,
0.3779296875,
0.1353759765625,
0.3779296875,
0.1455078125,
0.364501953125,
0.1455078125
],
[
0.418212890625,
0.1353759765625,
0.4306640625,
0.1353759765625,
0.4306640625,
0.1455078125,
0.418212890625,
0.1455078125
],
[
0.45166015625,
0.1353759765625,
0.46337890625,
0.1318359375,
0.47021484375,
0.1436767578125,
0.45751953125,
0.146728515625
],
[
0.5380859375,
0.1395263671875,
0.57275390625,
0.1395263671875,
0.57275390625,
0.146728515625,
0.5380859375,
0.146728515625
],
[
0.59033203125,
0.141357421875,
0.60107421875,
0.141357421875,
0.60107421875,
0.1455078125,
0.59033203125,
0.1455078125
],
[
0.28955078125,
0.144287109375,
0.306396484375,
0.144287109375,
0.306396484375,
0.168701171875,
0.28955078125,
0.168701171875
],
[
0.118408203125,
0.1455078125,
0.160400390625,
0.1455078125,
0.160400390625,
0.1561279296875,
0.118408203125,
0.1561279296875
],
[
0.364501953125,
0.1455078125,
0.3779296875,
0.1455078125,
0.3779296875,
0.157958984375,
0.364501953125,
0.157958984375
],
[
0.418212890625,
0.1455078125,
0.429931640625,
0.1455078125,
0.429931640625,
0.15673828125,
0.418212890625,
0.15673828125
],
[
0.455078125,
0.1455078125,
0.466796875,
0.1455078125,
0.466796875,
0.15673828125,
0.455078125,
0.15673828125
],
[
0.12005615234375,
0.1561279296875,
0.173828125,
0.1561279296875,
0.173828125,
0.16748046875,
0.12005615234375,
0.16748046875
],
[
0.2166748046875,
0.146728515625,
0.25439453125,
0.146728515625,
0.25439453125,
0.1561279296875,
0.2166748046875,
0.1561279296875
],
[
0.2149658203125,
0.1561279296875,
0.25439453125,
0.1561279296875,
0.25439453125,
0.16748046875,
0.2149658203125,
0.16748046875
],
[
0.364501953125,
0.1561279296875,
0.3779296875,
0.1561279296875,
0.3779296875,
0.16748046875,
0.364501953125,
0.16748046875
],
[
0.418212890625,
0.1561279296875,
0.4306640625,
0.1561279296875,
0.4306640625,
0.168701171875,
0.418212890625,
0.168701171875
],
[
0.455078125,
0.1561279296875,
0.466796875,
0.1561279296875,
0.466796875,
0.166259765625,
0.455078125,
0.166259765625
],
[
0.118408203125,
0.166259765625,
0.17626953125,
0.166259765625,
0.17626953125,
0.1776123046875,
0.118408203125,
0.1776123046875
],
[
0.291259765625,
0.166259765625,
0.3046875,
0.166259765625,
0.3046875,
0.1787109375,
0.291259765625,
0.1787109375
],
[
0.418212890625,
0.166259765625,
0.429931640625,
0.166259765625,
0.429931640625,
0.1787109375,
0.418212890625,
0.1787109375
],
[
0.456787109375,
0.166259765625,
0.468505859375,
0.166259765625,
0.468505859375,
0.1787109375,
0.456787109375,
0.1787109375
],
[
0.364501953125,
0.168701171875,
0.3779296875,
0.168701171875,
0.3779296875,
0.1787109375,
0.364501953125,
0.1787109375
],
[
0.2166748046875,
0.168701171875,
0.2509765625,
0.168701171875,
0.2509765625,
0.177001953125,
0.2166748046875,
0.177001953125
],
[
0.118408203125,
0.1776123046875,
0.162841796875,
0.1776123046875,
0.162841796875,
0.189453125,
0.118408203125,
0.189453125
],
[
0.214111328125,
0.1776123046875,
0.25439453125,
0.1776123046875,
0.25439453125,
0.189453125,
0.214111328125,
0.189453125
],
[
0.41650390625,
0.1776123046875,
0.429931640625,
0.1776123046875,
0.429931640625,
0.190673828125,
0.41650390625,
0.190673828125
],
[
0.456787109375,
0.1776123046875,
0.466796875,
0.1776123046875,
0.466796875,
0.190673828125,
0.456787109375,
0.190673828125
],
[
0.35693359375,
0.1817626953125,
0.3720703125,
0.174560546875,
0.3837890625,
0.1859130859375,
0.36865234375,
0.193603515625
],
[
0.291259765625,
0.1787109375,
0.3046875,
0.1787109375,
0.3046875,
0.191162109375,
0.291259765625,
0.191162109375
],
[
0.361083984375,
0.188232421875,
0.378662109375,
0.188232421875,
0.378662109375,
0.2120361328125,
0.361083984375,
0.2120361328125
],
[
0.453369140625,
0.188232421875,
0.47021484375,
0.188232421875,
0.47021484375,
0.21142578125,
0.453369140625,
0.21142578125
],
[
0.2032470703125,
0.189453125,
0.2626953125,
0.189453125,
0.2626953125,
0.20068359375,
0.2032470703125,
0.20068359375
],
[
0.12005615234375,
0.189453125,
0.185546875,
0.189453125,
0.185546875,
0.1995849609375,
0.12005615234375,
0.1995849609375
],
[
0.291259765625,
0.189453125,
0.30322265625,
0.189453125,
0.30322265625,
0.20068359375,
0.291259765625,
0.20068359375
],
[
0.418212890625,
0.189453125,
0.429931640625,
0.189453125,
0.429931640625,
0.20068359375,
0.418212890625,
0.20068359375
],
[
0.54833984375,
0.193603515625,
0.57275390625,
0.19775390625,
0.5693359375,
0.204833984375,
0.5458984375,
0.20068359375
],
[
0.12005615234375,
0.1995849609375,
0.20068359375,
0.1995849609375,
0.20068359375,
0.210205078125,
0.12005615234375,
0.210205078125
],
[
0.2166748046875,
0.1995849609375,
0.2509765625,
0.1995849609375,
0.2509765625,
0.21142578125,
0.2166748046875,
0.21142578125
],
[
0.291259765625,
0.1995849609375,
0.30322265625,
0.1995849609375,
0.30322265625,
0.21142578125,
0.291259765625,
0.21142578125
],
[
0.418212890625,
0.1995849609375,
0.429931640625,
0.1995849609375,
0.429931640625,
0.21142578125,
0.418212890625,
0.21142578125
],
[
0.5634765625,
0.204833984375,
0.57861328125,
0.204833984375,
0.57861328125,
0.2078857421875,
0.5634765625,
0.2078857421875
],
[
0.370361328125,
0.21142578125,
0.382080078125,
0.2197265625,
0.370361328125,
0.2286376953125,
0.358642578125,
0.2203369140625
],
[
0.45751953125,
0.213134765625,
0.47021484375,
0.21728515625,
0.462646484375,
0.226806640625,
0.450927734375,
0.22265625
],
[
0.2166748046875,
0.21435546875,
0.2509765625,
0.21435546875,
0.2509765625,
0.226806640625,
0.2166748046875,
0.226806640625
],
[
0.12176513671875,
0.21435546875,
0.189697265625,
0.21435546875,
0.189697265625,
0.2257080078125,
0.12176513671875,
0.2257080078125
],
[
0.28955078125,
0.215576171875,
0.306396484375,
0.215576171875,
0.306396484375,
0.2257080078125,
0.28955078125,
0.2257080078125
],
[
0.411376953125,
0.2203369140625,
0.4248046875,
0.2120361328125,
0.4375,
0.2203369140625,
0.424072265625,
0.2286376953125
],
[
0.60205078125,
0.2244873046875,
0.623046875,
0.2244873046875,
0.623046875,
0.232177734375,
0.60205078125,
0.232177734375
],
[
0.6689453125,
0.223876953125,
0.6943359375,
0.223876953125,
0.6943359375,
0.232177734375,
0.6689453125,
0.232177734375
],
[
0.1175537109375,
0.2327880859375,
0.486083984375,
0.2327880859375,
0.486083984375,
0.2464599609375,
0.1175537109375,
0.2464599609375
],
[
0.53662109375,
0.235107421875,
0.623046875,
0.235107421875,
0.623046875,
0.2464599609375,
0.53662109375,
0.2464599609375
],
[
0.1158447265625,
0.2476806640625,
0.486083984375,
0.2476806640625,
0.486083984375,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.5380859375,
0.2464599609375,
0.6591796875,
0.2464599609375,
0.6591796875,
0.260009765625,
0.5380859375,
0.260009765625
],
[
0.6826171875,
0.2529296875,
0.70361328125,
0.2529296875,
0.70361328125,
0.255859375,
0.6826171875,
0.255859375
],
[
0.1158447265625,
0.260009765625,
0.48779296875,
0.26123046875,
0.48779296875,
0.275634765625,
0.1158447265625,
0.27490234375
],
[
0.1175537109375,
0.275634765625,
0.48779296875,
0.275634765625,
0.48779296875,
0.289306640625,
0.1175537109375,
0.289306640625
],
[
0.59375,
0.273681640625,
0.6484375,
0.273681640625,
0.6484375,
0.276611328125,
0.59375,
0.276611328125
],
[
0.118408203125,
0.29150390625,
0.486083984375,
0.29150390625,
0.486083984375,
0.3017578125,
0.118408203125,
0.3017578125
],
[
0.5205078125,
0.295654296875,
0.60205078125,
0.295654296875,
0.60205078125,
0.302734375,
0.5205078125,
0.302734375
],
[
0.118408203125,
0.305908203125,
0.484375,
0.305908203125,
0.484375,
0.31640625,
0.118408203125,
0.31640625
],
[
0.5439453125,
0.305908203125,
0.62060546875,
0.305908203125,
0.62060546875,
0.3134765625,
0.5439453125,
0.3134765625
],
[
0.7841796875,
0.30712890625,
0.8037109375,
0.30712890625,
0.8037109375,
0.31005859375,
0.7841796875,
0.31005859375
],
[
0.8193359375,
0.30712890625,
0.8330078125,
0.30712890625,
0.8330078125,
0.31005859375,
0.8193359375,
0.31005859375
],
[
0.1158447265625,
0.317626953125,
0.484375,
0.317626953125,
0.484375,
0.330810546875,
0.1158447265625,
0.330810546875
],
[
0.6171875,
0.3154296875,
0.64208984375,
0.3154296875,
0.64208984375,
0.318359375,
0.6171875,
0.318359375
],
[
0.5751953125,
0.32177734375,
0.59619140625,
0.32177734375,
0.59619140625,
0.328857421875,
0.5751953125,
0.328857421875
],
[
0.59033203125,
0.322509765625,
0.6455078125,
0.322509765625,
0.6455078125,
0.334228515625,
0.59033203125,
0.334228515625
],
[
0.7900390625,
0.32373046875,
0.8076171875,
0.32373046875,
0.8076171875,
0.33837890625,
0.7900390625,
0.33837890625
],
[
0.1158447265625,
0.33203125,
0.411376953125,
0.33203125,
0.411376953125,
0.345703125,
0.1158447265625,
0.345703125
],
[
0.51904296875,
0.34326171875,
0.587890625,
0.34326171875,
0.587890625,
0.35400390625,
0.51904296875,
0.35400390625
],
[
0.5341796875,
0.35400390625,
0.634765625,
0.35400390625,
0.634765625,
0.364013671875,
0.5341796875,
0.364013671875
],
[
0.1175537109375,
0.385498046875,
0.486083984375,
0.385498046875,
0.486083984375,
0.3984375,
0.1175537109375,
0.3984375
],
[
0.1175537109375,
0.40087890625,
0.48779296875,
0.40087890625,
0.48779296875,
0.41455078125,
0.1175537109375,
0.41455078125
],
[
0.1175537109375,
0.416259765625,
0.489501953125,
0.416259765625,
0.489501953125,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.5517578125,
0.4150390625,
0.56494140625,
0.4150390625,
0.56494140625,
0.418701171875,
0.5517578125,
0.418701171875
],
[
0.1158447265625,
0.43115234375,
0.486083984375,
0.431640625,
0.486083984375,
0.446533203125,
0.1158447265625,
0.4453125
],
[
0.7841796875,
0.429931640625,
0.796875,
0.429931640625,
0.796875,
0.432861328125,
0.7841796875,
0.432861328125
],
[
0.72314453125,
0.441162109375,
0.79833984375,
0.4423828125,
0.79833984375,
0.449462890625,
0.72314453125,
0.448974609375
],
[
0.8076171875,
0.440673828125,
0.86572265625,
0.440673828125,
0.86572265625,
0.451904296875,
0.8076171875,
0.451904296875
],
[
0.6767578125,
0.44482421875,
0.68408203125,
0.44482421875,
0.68408203125,
0.448974609375,
0.6767578125,
0.448974609375
],
[
0.5341796875,
0.446533203125,
0.6962890625,
0.453125,
0.6943359375,
0.471435546875,
0.5322265625,
0.465576171875
],
[
0.60546875,
0.443603515625,
0.61865234375,
0.443603515625,
0.61865234375,
0.44775390625,
0.60546875,
0.44775390625
],
[
0.72705078125,
0.453125,
0.85205078125,
0.453125,
0.85205078125,
0.460205078125,
0.72705078125,
0.460205078125
],
[
0.74462890625,
0.4619140625,
0.8251953125,
0.4619140625,
0.8251953125,
0.47265625,
0.74462890625,
0.47265625
],
[
0.1158447265625,
0.448974609375,
0.486083984375,
0.448974609375,
0.486083984375,
0.4619140625,
0.1158447265625,
0.4619140625
],
[
0.1175537109375,
0.465576171875,
0.486083984375,
0.465576171875,
0.486083984375,
0.478515625,
0.1175537109375,
0.478515625
],
[
0.53662109375,
0.468505859375,
0.60107421875,
0.47021484375,
0.60107421875,
0.482177734375,
0.53662109375,
0.479736328125
],
[
0.56689453125,
0.465576171875,
0.62060546875,
0.463134765625,
0.62060546875,
0.473876953125,
0.56689453125,
0.4755859375
],
[
0.72900390625,
0.473876953125,
0.8251953125,
0.473876953125,
0.8251953125,
0.48095703125,
0.72900390625,
0.48095703125
],
[
0.1175537109375,
0.482177734375,
0.486083984375,
0.482177734375,
0.486083984375,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.51123046875,
0.4892578125,
0.88330078125,
0.4892578125,
0.88330078125,
0.5029296875,
0.51123046875,
0.5029296875
],
[
0.1158447265625,
0.49755859375,
0.48779296875,
0.49755859375,
0.48779296875,
0.51123046875,
0.1158447265625,
0.51123046875
],
[
0.51318359375,
0.5048828125,
0.880859375,
0.5048828125,
0.880859375,
0.515625,
0.51318359375,
0.515625
],
[
0.1175537109375,
0.51318359375,
0.489501953125,
0.51318359375,
0.489501953125,
0.52685546875,
0.1175537109375,
0.52685546875
],
[
0.509765625,
0.51708984375,
0.8818359375,
0.5185546875,
0.8818359375,
0.5322265625,
0.509765625,
0.53076171875
],
[
0.1158447265625,
0.529296875,
0.486083984375,
0.529296875,
0.486083984375,
0.5419921875,
0.1158447265625,
0.5419921875
],
[
0.509765625,
0.5322265625,
0.8818359375,
0.5322265625,
0.8818359375,
0.5458984375,
0.509765625,
0.5458984375
],
[
0.1158447265625,
0.54443359375,
0.48779296875,
0.54443359375,
0.48779296875,
0.55810546875,
0.1158447265625,
0.55810546875
],
[
0.509765625,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.509765625,
0.56005859375
],
[
0.1158447265625,
0.5625,
0.489501953125,
0.5625,
0.489501953125,
0.576171875,
0.1158447265625,
0.576171875
],
[
0.51318359375,
0.5625,
0.880859375,
0.5625,
0.880859375,
0.572265625,
0.51318359375,
0.572265625
],
[
0.118408203125,
0.5791015625,
0.48779296875,
0.5791015625,
0.48779296875,
0.5888671875,
0.118408203125,
0.5888671875
],
[
0.5087890625,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58740234375,
0.5087890625,
0.58740234375
],
[
0.1175537109375,
0.59326171875,
0.486083984375,
0.59326171875,
0.486083984375,
0.60693359375,
0.1175537109375,
0.60693359375
],
[
0.51123046875,
0.5888671875,
0.869140625,
0.5888671875,
0.869140625,
0.6025390625,
0.51123046875,
0.6025390625
],
[
0.1175537109375,
0.60986328125,
0.486083984375,
0.60986328125,
0.486083984375,
0.62353515625,
0.1175537109375,
0.62353515625
],
[
0.1158447265625,
0.6259765625,
0.486083984375,
0.6259765625,
0.486083984375,
0.6396484375,
0.1158447265625,
0.6396484375
],
[
0.51318359375,
0.6357421875,
0.5498046875,
0.6357421875,
0.5498046875,
0.64794921875,
0.51318359375,
0.64794921875
],
[
0.5498046875,
0.63720703125,
0.634765625,
0.63720703125,
0.634765625,
0.6484375,
0.5498046875,
0.6484375
],
[
0.118408203125,
0.6435546875,
0.486083984375,
0.6435546875,
0.486083984375,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.1158447265625,
0.6591796875,
0.48779296875,
0.6591796875,
0.48779296875,
0.6728515625,
0.1158447265625,
0.6728515625
],
[
0.509765625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.1158447265625,
0.6748046875,
0.2939453125,
0.6748046875,
0.2939453125,
0.6884765625,
0.1158447265625,
0.6884765625
],
[
0.51123046875,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.51123046875,
0.693359375
],
[
0.509765625,
0.6953125,
0.8818359375,
0.69482421875,
0.8818359375,
0.708984375,
0.509765625,
0.71044921875
],
[
0.51318359375,
0.7119140625,
0.8818359375,
0.7119140625,
0.8818359375,
0.7255859375,
0.51318359375,
0.7255859375
],
[
0.1158447265625,
0.72021484375,
0.145263671875,
0.72021484375,
0.145263671875,
0.73291015625,
0.1158447265625,
0.73291015625
],
[
0.1435546875,
0.7197265625,
0.2149658203125,
0.7197265625,
0.2149658203125,
0.73388671875,
0.1435546875,
0.73388671875
],
[
0.1435546875,
0.7236328125,
0.1485595703125,
0.7236328125,
0.1485595703125,
0.728515625,
0.1435546875,
0.728515625
],
[
0.509765625,
0.72705078125,
0.880859375,
0.72705078125,
0.880859375,
0.74072265625,
0.509765625,
0.74072265625
],
[
0.51123046875,
0.74462890625,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75830078125
],
[
0.11419677734375,
0.75830078125,
0.489501953125,
0.7587890625,
0.489501953125,
0.77392578125,
0.11419677734375,
0.7724609375
],
[
0.51123046875,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.1175537109375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.775390625,
0.88330078125,
0.775390625,
0.88330078125,
0.7890625,
0.509765625,
0.7890625
],
[
0.1175537109375,
0.79345703125,
0.486083984375,
0.79345703125,
0.486083984375,
0.80419921875,
0.1175537109375,
0.80419921875
],
[
0.51123046875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.1175537109375,
0.80712890625,
0.491943359375,
0.80712890625,
0.491943359375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1158447265625,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.5087890625,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5087890625,
0.83837890625
],
[
0.11419677734375,
0.84033203125,
0.490234375,
0.8388671875,
0.490234375,
0.85400390625,
0.11419677734375,
0.85498046875
],
[
0.509765625,
0.84033203125,
0.88330078125,
0.84033203125,
0.88330078125,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90380859375,
0.2032470703125,
0.90380859375,
0.2032470703125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Continuous Dialog Form",
"Free",
"Commander'stum",
"Env",
"Dataset",
"Photorealistic",
"Space",
"Itte",
"ceat T=(",
"R2R, RxR",
"indoor",
"1",
"x",
"x",
">",
"1",
"V",
"+",
"VNLA,HANNA indoor",
"x",
"VLN-CE",
"indoor",
"Follower'stum",
"X",
"V",
"x",
"V",
"Followe",
"na",
">X",
"CVDN",
"x",
"1",
"1",
"ALFRED",
"indoor",
"indoor",
"x",
"x",
"V",
"DialFRED",
"x",
"I",
"X",
"x",
"indoor",
"TEACh",
"indoor",
"1",
">",
"-",
"x",
"xy",
"",
"street-view",
"TouchDown",
"V",
"x",
"tent",
"modified LANI",
"aerial",
"x",
"x",
"",
"1",
"J",
"aerial",
"AVDN (ours)",
"V",
"-",
"T=O",
"T=1",
"Table l: Example Vision-and-Language Navigation",
"Utterance at T=1",
"Datasets. R2R (Anderson et al., 2018), RxR (Ku et al.",
"u",
"kenn",
"2020), VNLA (Nguyen et al., 2019), HANNA (Nguyen",
"and Daume III, 2019), VLN-CE (Krantz et al., 2020),",
"HOa",
"CVDN (Thomason et al.. 2020). ALFRED (Shridhar",
"Commander'stum",
"et al.,2020a),DialFRED (Gao et al..2022).TEACh",
"UtteranceatT=1",
"nua",
"Fa",
"(Padmakumar et al., 2021), TouchDown (Chen et al.",
"ot",
"n, hea",
"warehousen fice",
"1",
"2019), modified LANI (Blukis et al., 2019).",
"Follower's tum:",
"Follower'snavigation",
"already an active topic in the field. Some inspiring",
"works (Loquercio et al., 2018; Giusti et al., 2015;",
"Smolyanskiy et al., 2017; Fan et al., 2020; Boz-",
"ta",
"can and Kayacan, 2020; Majdik et al., 2017; Kang",
"Ow",
"Startingpositio",
"ODestinatior",
"=",
"eaeah",
"二1",
"TFollower'sposition-anddirectic",
"Follower's trajectory",
"et al., 2019) worked on using pre-collected real-",
"world drone data to tackle aerial vision navigation",
"Auto-hint Yes yc",
"tion.Am Inight?",
"Follower'sviewarea",
"problems. Due to the hardness of collecting data",
"Figure 2: Example of a trajectory in our AVDN dataset",
"and the risk of crashes, some other works applied",
"Ontheleft,thecommander'sturn andthefollower'sturn",
"simulation for aerial navigation (Chen et al., 2018;",
"alternate in chronological order. In each turn, dialog",
"Shah et al., 2017; Chen et al., 2020), where rich",
"utterances are shown, and the follower's turn also shows",
"ground truths are provided without the need for",
"the navigation process that spans from time step T to T+",
"annotation. However, the modality of language is",
"I,includingthefollower'sobservationand attention.On",
"missing in these prior works and as a result, the nav-",
"the right, there are trajectory overviews at different time",
"igation tasks only contain simple goals. As for the",
"steps. More examples can be found in the Appendix.",
"aerial vision-and-language navigation task in this",
"work, the navigation is guided by natural dialog",
"3.1",
" Simulator",
"As aresult,it allows more diverse and complex",
"navigation and also resolves ambiguities during",
"We build a simulator to simulate the drone with",
"complicated navigation.",
"a top-down view area. Our simulation environ-",
"ment is a continuous space so that the simulated",
"drone can move continuously to any point within",
"3",
"Dataset",
"",
"the environment. The drone's visual observations",
"are square images generated corresponding to the",
"The AVDN dataset includes dialogs, navigation tra-",
"drone's view area by cropping from high-resolution",
"jectories, and the drone's visual observation with",
"satellite images in the x View dataset (Lam et al.,",
"human attention, where an example is shown in",
"2018), an open-source large-scale satellite image",
"Figure 2. With the help of a newly proposed simu-",
"object detection dataset. In this way, our simula",
"lator, we record the AVDN trajectories created by",
" tor is capable of providing continuous frames with",
" two groups of humans interacting with each other,",
"rich visual features. We also design an interface for",
"playing either the commander role or the follower",
"our simulator, where the simulated drone can be",
"role. Our AVDN dataset is the first aerial navi-",
"controlled with a keyboard and the drone's visual",
"gation dataset based on dialogs to the best of our",
"observation will be displayed in real-time with a",
"knowledge.",
"digital compass. During the control, users can also"
] |
{
"arxiv_id": "2205.12219",
"language": "en",
"page_idx": 2
}
|
||
2205.12219-en-3
|
2205.12219-en-3.jpg
|
provide their attention over the displayed images on the interface by clicking the region they attend to. Last but not least, our simulator is capable of generating trajectory overviews, i.e. commander's view, showing the starting positions, destination areas, current view area and past trajectory (if exists) as in Figure 2.
## 3.2 Dataset Structure
In our AVDN dataset, each navigation trajectory includes time steps $T=0,1,\ldots,M$,where $M\geq1.$ At $T=0$, an initial instruction is provided by the commander. Between adjacent time steps, there is a corresponding navigation sub-trajectory. At every time step of $0<T<M$,there are questions from the follower and the corresponding answers from the commander. At $T=M$, the navigation trajectory ends because the destination area Des is reached and claimed by the follower. For details about when a trajectory ends, please refer to Section 3.3 Success Condition.
There are $M$ follower's view area sequences sequence, where the view area's center coordinate $c_i^T$ always falls on the trajectory. Therefore, based on each view area,we could retrieve not only the simulated drone's location $c_i$,but also direction $d_i$ and altitude $h_i.$ Last but not least, for each view area $u$, there is a corresponding binary human attention mask with the same size. The area in $u$ that corresponds to the white area on the mask is where the follower attended.
## 33 Dataset Collection
We collect our dataset with the help of Amazon Mechanical Turk (AMT) workers and drone experts, where AMT workers play the commander role to provide instructions and drone experts play the follower role to control a simulated drone and carry out the instruction. We pay the workers with wages no less than $15/h, and the data collection lasts for 90 days. We adopt an asynchronous data collection method, where the followers and commanders work in turns rather than simultaneously. This not only lowers the cost of data collection but also simulates how aerial vision-and-dialog navigation would work in practice, where the commanders will not monitor the follower's actions all the time. Pipeline Before the start of data collection, we
first sample objects in the xView dataset (Lam et al., 2018) as the destination areas and pair them
with randomly selected initial follower's view areas within 1.5km distance. Then, using our simulator, we generate the trajectory overview at time step $T=0$, as shown in Figure 2, which becomes the initial commander's view.
During data collection, the initial commander's view is presented to AMT workers for creating the initial instructions. We instruct the AMT workers to write instructions as if they are talking to a drone pilot based on the marked satellite images. Next, we let human drone experts play the follower role, i.e. controlling the simulated drone through our simulator interface, following the instructions and asking questions if they cannot find the destination area. When the experts stop the current navigation, they can either enter questions into a chatbox, claim the destination with a template sentence or reject the instruction for bad quality. If the destination is falsely claimed, the simulator will generate an auto-hint to let the follower ask some questions. For questions asked, AMT workers will provide further instructions accordingly based on
$<u_{0}^{T},u_{1}^{T},\ldots,u_{N_{T}}^{T}>,N_{T}$ is the length of $T$-th given navigation information and dialog history.
Then, the same drone experts will continue playing the follower role and again asking questions when necessary. We iterate the process until the destination is successfully reached and claimed by the follower.
Success Condition The navigation trajectory is successful only when the destination is reached at the time the follower claims it. We determine that the destination is reached in view area $u_j$ by checking the center $c_j$ and computing the Intersection over Union (IoU) between $u_j$ and $Des.$ If $c_j$ is inside Des and the IoU of $u_j$ and Des is larger than 0.4, the destination is regarded in $u_j.$
## 3.4 Data Analysis
Our AVDN dataset includes 3,064 aerial navigation trajectories, each with multi-round natural language dialog. There are two rounds of dialog on average per trajectory, where the number of dialog rounds in a trajectory equals to the maximum time step $M.$ The most frequent words are shown in Figure 3a. The recorded AVDN trajectory path length has an average of 287m, and its distribution is shown in Figure 3b. The trajectories and dialogs can be further separated into 6,269 sub-trajectories corresponding to the dialog rounds.
We split our dataset into training, seenvalidation, unseen-validation, and unseen-testing
|
[
[
0.1158447265625,
0.08489990234375,
0.489501953125,
0.08489990234375,
0.489501953125,
0.09857177734375,
0.1158447265625,
0.09857177734375
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.1015625,
0.88330078125,
0.1015625,
0.88330078125,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
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,
0.1353759765625,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.1175537109375,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.118408203125,
0.1519775390625,
0.489501953125,
0.1519775390625,
0.489501953125,
0.162109375,
0.118408203125,
0.162109375
],
[
0.51123046875,
0.1519775390625,
0.70361328125,
0.1519775390625,
0.70361328125,
0.162109375,
0.51123046875,
0.162109375
],
[
0.1175537109375,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.52880859375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.18115234375,
0.52880859375,
0.18115234375
],
[
0.1175537109375,
0.18408203125,
0.22412109375,
0.18408203125,
0.22412109375,
0.19482421875,
0.1175537109375,
0.19482421875
],
[
0.51123046875,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.51123046875,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.11419677734375,
0.210205078125,
0.300537109375,
0.210205078125,
0.300537109375,
0.223876953125,
0.11419677734375,
0.223876953125
],
[
0.51123046875,
0.215576171875,
0.8818359375,
0.215576171875,
0.8818359375,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.1158447265625,
0.23095703125,
0.49365234375,
0.232177734375,
0.49365234375,
0.2464599609375,
0.1158447265625,
0.2452392578125
],
[
0.509765625,
0.2298583984375,
0.88330078125,
0.23095703125,
0.88330078125,
0.2452392578125,
0.509765625,
0.24462890625
],
[
0.1158447265625,
0.2476806640625,
0.49365234375,
0.2464599609375,
0.49365234375,
0.26123046875,
0.1158447265625,
0.262451171875
],
[
0.5087890625,
0.2464599609375,
0.8818359375,
0.2476806640625,
0.8818359375,
0.262451171875,
0.5087890625,
0.26123046875
],
[
0.1158447265625,
0.26318359375,
0.48779296875,
0.26318359375,
0.48779296875,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.509765625,
0.26416015625,
0.8818359375,
0.26416015625,
0.8818359375,
0.27783203125,
0.509765625,
0.27783203125
],
[
0.1175537109375,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.1175537109375,
0.29443359375
],
[
0.509765625,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1158447265625,
0.297607421875,
0.489501953125,
0.297607421875,
0.489501953125,
0.311279296875,
0.1158447265625,
0.311279296875
],
[
0.5087890625,
0.295654296875,
0.88330078125,
0.29443359375,
0.88330078125,
0.308837890625,
0.5087890625,
0.31005859375
],
[
0.1158447265625,
0.312255859375,
0.489501953125,
0.312255859375,
0.489501953125,
0.325927734375,
0.1158447265625,
0.325927734375
],
[
0.509765625,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.325927734375,
0.509765625,
0.325927734375
],
[
0.1158447265625,
0.327880859375,
0.489501953125,
0.328857421875,
0.489501953125,
0.34326171875,
0.1158447265625,
0.342529296875
],
[
0.509765625,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.509765625,
0.341552734375
],
[
0.1158447265625,
0.34326171875,
0.489501953125,
0.344482421875,
0.489501953125,
0.359375,
0.1158447265625,
0.358154296875
],
[
0.509765625,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.355712890625,
0.509765625,
0.355712890625
],
[
0.1158447265625,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1158447265625,
0.37353515625
],
[
0.509765625,
0.359375,
0.87890625,
0.359375,
0.87890625,
0.372314453125,
0.509765625,
0.372314453125
],
[
0.11419677734375,
0.3759765625,
0.489501953125,
0.3759765625,
0.489501953125,
0.3896484375,
0.11419677734375,
0.3896484375
],
[
0.509765625,
0.374755859375,
0.880859375,
0.374755859375,
0.880859375,
0.388427734375,
0.509765625,
0.388427734375
],
[
0.1158447265625,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.1158447265625,
0.40625
],
[
0.51123046875,
0.391357421875,
0.8818359375,
0.39013671875,
0.8818359375,
0.405029296875,
0.51123046875,
0.40625
],
[
0.1175537109375,
0.4091796875,
0.34765625,
0.4091796875,
0.34765625,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.51123046875,
0.407958984375,
0.8818359375,
0.40673828125,
0.8818359375,
0.421630859375,
0.51123046875,
0.4228515625
],
[
0.1368408203125,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.1368408203125,
0.439453125
],
[
0.51123046875,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.51123046875,
0.438232421875
],
[
0.1158447265625,
0.441162109375,
0.489501953125,
0.439453125,
0.489501953125,
0.4560546875,
0.1158447265625,
0.457763671875
],
[
0.51318359375,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.1158447265625,
0.457763671875,
0.486083984375,
0.457763671875,
0.486083984375,
0.471435546875,
0.1158447265625,
0.471435546875
],
[
0.509765625,
0.457275390625,
0.8818359375,
0.457275390625,
0.8818359375,
0.47021484375,
0.509765625,
0.47021484375
],
[
0.1158447265625,
0.473876953125,
0.486083984375,
0.473876953125,
0.486083984375,
0.488037109375,
0.1158447265625,
0.488037109375
],
[
0.509765625,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.509765625,
0.48681640625
],
[
0.118408203125,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.5029296875,
0.118408203125,
0.5029296875
],
[
0.509765625,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.5029296875,
0.509765625,
0.5029296875
],
[
0.1158447265625,
0.50732421875,
0.486083984375,
0.50732421875,
0.486083984375,
0.52099609375,
0.1158447265625,
0.52099609375
],
[
0.509765625,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.509765625,
0.5185546875
],
[
0.1158447265625,
0.5224609375,
0.490234375,
0.5224609375,
0.490234375,
0.5361328125,
0.1158447265625,
0.5361328125
],
[
0.51123046875,
0.5224609375,
0.580078125,
0.5224609375,
0.580078125,
0.53369140625,
0.51123046875,
0.53369140625
],
[
0.1158447265625,
0.5390625,
0.490234375,
0.5390625,
0.490234375,
0.552734375,
0.1158447265625,
0.552734375
],
[
0.509765625,
0.5458984375,
0.880859375,
0.5458984375,
0.880859375,
0.56005859375,
0.509765625,
0.56005859375
],
[
0.1158447265625,
0.5546875,
0.489501953125,
0.5546875,
0.489501953125,
0.568359375,
0.1158447265625,
0.568359375
],
[
0.1158447265625,
0.57080078125,
0.48779296875,
0.57080078125,
0.48779296875,
0.58447265625,
0.1158447265625,
0.58447265625
],
[
0.51318359375,
0.5654296875,
0.8818359375,
0.5654296875,
0.8818359375,
0.576171875,
0.51318359375,
0.576171875
],
[
0.51123046875,
0.580078125,
0.88330078125,
0.580078125,
0.88330078125,
0.59326171875,
0.51123046875,
0.59326171875
],
[
0.1158447265625,
0.5859375,
0.280517578125,
0.5859375,
0.280517578125,
0.599609375,
0.1158447265625,
0.599609375
],
[
0.509765625,
0.5966796875,
0.88330078125,
0.5966796875,
0.88330078125,
0.60986328125,
0.509765625,
0.60986328125
],
[
0.1158447265625,
0.615234375,
0.3046875,
0.615234375,
0.3046875,
0.62890625,
0.1158447265625,
0.62890625
],
[
0.51123046875,
0.61328125,
0.88330078125,
0.61328125,
0.88330078125,
0.626953125,
0.51123046875,
0.626953125
],
[
0.51123046875,
0.626953125,
0.8818359375,
0.62744140625,
0.8818359375,
0.642578125,
0.51123046875,
0.64111328125
],
[
0.1158447265625,
0.6357421875,
0.489501953125,
0.6357421875,
0.489501953125,
0.6494140625,
0.1158447265625,
0.6494140625
],
[
0.51318359375,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.51318359375,
0.65771484375
],
[
0.1175537109375,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.1175537109375,
0.66650390625
],
[
0.5087890625,
0.6591796875,
0.8212890625,
0.6611328125,
0.8212890625,
0.67578125,
0.5087890625,
0.67333984375
],
[
0.1175537109375,
0.66845703125,
0.486083984375,
0.66845703125,
0.486083984375,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.1175537109375,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.1175537109375,
0.69873046875
],
[
0.51318359375,
0.6884765625,
0.66650390625,
0.6884765625,
0.66650390625,
0.703125,
0.51318359375,
0.703125
],
[
0.1158447265625,
0.69970703125,
0.484375,
0.70068359375,
0.484375,
0.71533203125,
0.1158447265625,
0.71435546875
],
[
0.51123046875,
0.708984375,
0.8818359375,
0.71044921875,
0.8818359375,
0.724609375,
0.51123046875,
0.7236328125
],
[
0.1175537109375,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.1175537109375,
0.73095703125
],
[
0.51123046875,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.74072265625,
0.51123046875,
0.74072265625
],
[
0.118408203125,
0.7353515625,
0.486083984375,
0.7353515625,
0.486083984375,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51123046875,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.51123046875,
0.755859375
],
[
0.1175537109375,
0.75,
0.490234375,
0.75,
0.490234375,
0.76318359375,
0.1175537109375,
0.76318359375
],
[
0.51318359375,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.76708984375,
0.486083984375,
0.76708984375,
0.486083984375,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.509765625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.509765625,
0.7890625
],
[
0.1175537109375,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.1175537109375,
0.7958984375
],
[
0.51318359375,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51318359375,
0.8046875
],
[
0.1175537109375,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.81103515625,
0.1175537109375,
0.81103515625
],
[
0.5087890625,
0.8056640625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.5087890625,
0.82080078125
],
[
0.1158447265625,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.1158447265625,
0.82763671875
],
[
0.51123046875,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83740234375,
0.51123046875,
0.83740234375
],
[
0.1158447265625,
0.830078125,
0.489501953125,
0.830078125,
0.489501953125,
0.84326171875,
0.1158447265625,
0.84326171875
],
[
0.1158447265625,
0.845703125,
0.489501953125,
0.845703125,
0.489501953125,
0.859375,
0.1158447265625,
0.859375
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.51123046875,
0.85546875,
0.880859375,
0.85546875,
0.880859375,
0.869140625,
0.51123046875,
0.869140625
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.77685546875,
0.87255859375,
0.77685546875,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.5283203125,
0.88720703125,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.5283203125,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"provide their attention over the displayed images",
"with randomly selected initial follower's view areas",
"on the interface by clicking the region they attend",
"within 1.5km distance. Then, using our simulator,",
"to. Last but not least, our simulator is capable of",
"we generate the trajectory overview at time step",
"generating trajectory overviews,i.e.commander's",
"T = 0, as shown in Figure 2, which becomes the",
"view,showingthe startingpositions,destination ar-",
"initial commander's view.",
"eas, current view area and past trajectory (if exists)",
"During data collection, the initial commander's",
"as in Figure 2.",
"view is presented to AMT workers for creating the",
"initial instructions. We instruct the AMT work",
"3.2 Dataset Structure",
"ers to write instructions as if they are talking to a",
"In our AVDN dataset, each navigation trajectory in-",
"drone pilot based on the marked satellite images.",
"cludes time steps T = 0, 1,..., M, where M ≥ 1.",
"Next, we let human drone experts play the follower",
"At T = O, an initial instruction is provided by the",
"role, i.e. controlling the simulated drone through",
"commander. Between adjacent time steps, there",
"our simulator interface, following the instructions",
"is a corresponding navigation sub-trajectory. At",
"and asking questions if they cannot find the des-",
"every time step of 0 < T < M, there are questions",
"tination area. When the experts stop the current",
"from the follower and the corresponding answers",
"navigation, they can either enter questions into a",
"from the commander. At T = M, the navigation",
"chatbox, claim the destination with a template sen-",
"trajectory ends because the destination area Des",
"tence or reject the instruction for bad quality. If the",
"is reached and claimed by the follower. For de-",
"destination is falsely claimed, the simulator will",
"tails about when a trajectory ends, please refer to",
"generate an auto-hint to let the follower ask some",
"Section 3.3 Success Condition.",
"questions. For questions asked, AMT workers will",
"There are M follower's view area sequences",
"provide further instructions accordingly based on",
"< uT, uT,..., uN >, NT is the length of T-th",
"given navigation information and dialog history",
"sequence, where the view area's center coordinate",
"Then, the same drone experts will continue playing",
"cT always falls on the trajectory. Therefore, based",
"the follower role and again asking questions when",
"on eachviewarea,wecouldretrievenotonlythe",
"necessary. We iterate the process until the desti",
" simulated drone's location ci, but also direction d;",
"nation is successfully reached and claimed by the",
"and altitude hi. Last but not least, for each view",
"follower.",
"area u, there is a corresponding binary human at-",
"Success Condition The navigation trajectory is",
"tention mask with the same size. The area in u that",
"corresponds to the white area on the mask is where",
"successful only when the destination is reached at",
"the time the follower claims it. We determine that",
"the follower attended.",
"the destination is reached in view area u, by check-",
"3.3Dataset Collection",
"ing the center Ci and computing the Intersection",
"over Union (oU) between uj and Des. If cj is",
"We collect our dataset with the help of Amazon Me-",
"inside Des and the IoU of u and Des is larger",
"chanical Turk (AMT) workers and drone experts,",
"than 0.4, the destination is regarded in uj.",
"where AMT workers play the commander role to",
"provide instructions and drone experts play the fol-",
"3.4 Data Analysis",
"lower role to control a simulated drone and carry",
"Our AVDN dataset includes 3,064 aerial naviga-",
"out the instruction. We pay the workers with wages",
"tion trajectories, each with multi-round natural lan-",
"nolessthan$15/h.andthedatacollectionlastsfor",
"guage dialog. There are two rounds of dialog on",
"90 days. We adopt an asynchronous data collec-",
"average per trajectory, where the number of dia-",
"tionmethod,wherethefollowersand commanders",
"log rounds in a trajectory equals to the maximum",
"work in turns rather than simultaneously. This not",
"time step M. The most frequent words are shown",
"only lowers the cost of data collection but also",
"in Figure 3a. The recorded AVDN trajectory path",
" simulates how aerial vision-and-dialog navigation",
"length has an average of 287m, and its distribution",
"would work in practice, where the commanders",
"will not monitor the follower's actions all the time.",
"is shown in Figure 3b. The trajectories and dialogs",
"can be further separated into 6,269 sub-trajectories",
"Pipeline Before the start of data collection, we",
" corresponding to the dialog rounds.",
"first sample objects in the xView dataset (Lam",
"We split our dataset into training, seen-",
"et al., 2018) as the destination areas and pair them",
"validation, unseen-validation, and unseen-testing"
] |
{
"arxiv_id": "2205.12219",
"language": "en",
"page_idx": 3
}
|
||
2205.12490-en-0
|
2205.12490-en-0.jpg
|
# Learning from a Friend: Improving Event Extraction via Self-Training with Feedback from Abstract Meaning Representation
$\begin{array}{l}\textbf{Zhiyang Xu}^{\spadesuit}&\textbf{Jay-Yoon Lee}^{\clubsuit*}&\textbf{Lifu Huang}^{\spadesuit}\\\spadesuit_{\text{Department of Computer Science, Virginia Tech}}\end{array}$
Graduate School of Data Science, Seoul National University
$\{$zhiyangx,lifuh$\}$@vt.edu
lee.jayyoon@snu.ac.kr
### $\mathbf{Abstract}$
Data scarcity has been the main factor that hinders the progress of event extraction. To overcome this issue, we propose a Self-Training with Feedback (STF) framework that leverages the large-scale unlabeled data and acquires feedback for each new event prediction from the unlabeled data by comparing it to the Abstract Meaning Representation (AMR) graph of the same sentence. Specifically, STF consists of (1) a base event extraction model trained on existing event annotations and then applied to large-scale unlabeled corpora to predict new event mentions as pseudo training samples, and (2) a novel scoring model that takes in each new predicted event trigger, an argument, its argument role, as well as their paths in the AMR graph to estimate a compatibility score indicating the correctness of the pseudo label. The compatibility scores further act as feedback to encourage or discourage the model learning on the pseudo labels during self-training. Experimental results on three benchmark datasets, including ACE05-E, ACE05-E$^+$, and ERE, demonstrate the effectiveness of the STF framework on event extraction, especially event argument extraction, with significant performance gain over the base event extraction models and strong baselines. Our experimental analysis further shows that STF is a generic framework as it can be applied to improve most, if not all, event extraction models by leveraging largescale unlabeled data, even when high-quality AMR graph annotations are not available.$^{1}$
## 1 Introduction
Event extraction (EE), which aims to identify and classify event triggers and arguments, has been a long-stand challenging problem in natural language processing. Despite the large performance leap brought by advances in deep learning, recent
$^{*}$corresponding authors
The source code and model checkpoints are publicly available at https://github.com/VT-NLP/Event\_ $\operatorname{Extraction}_{\mathrm{with}}\underline{\mathrm{Self}}\_$Training
studies (Deng et al., 2021; Wang et al., 2021b) have shown that the data scarcity of existing event annotations has been the major issue that hinders the progress of EE. For example, in ACE-052, one of the most popular event extraction benchmark datasets, 10 of the 33 event types have less than 80 annotations. However, creating event annotations is extremely expensive and time-consuming, e.g., it takes several linguists over one year to annotate 500 documents with about 5000 event mentions for ACE- 05.
To overcome the data scarcity issue of EE, previous studies (Chen and Ji, 2009; Liao and Grishman, 2011a; Ferguson et al., 2018a) develop self-training methods that allow the trained EE model to learn further by regarding its own predictions on large-scale unlabeled corpora as pseudo labels. However, simply adding the high-confidence event predictions to the training set inevitably introduces noises (Liu et al., 2021; Arazo et al., 2020; Jiang et al., 2018), especially given that the current state-of-the-art performance of event argument extraction is still less than 60% F-score. To tackle this challenge, we introduce a Self-Training with Feedback framework, named STF, which consists of an event extraction model that is firstly trained on the existing event annotations and then continually updated on the unlabeled corpus with selftraining, and a scoring model that is to evaluate the correctness of the new event predictions (pseudo labels) from the unlabeled corpus, and the scores further act as feedback to encourage or discourage the learning of the event extraction model on the pseudo labels during self-training, inspired by the REINFORCE algorithms (Williams, 1992).
Specifically, the event extraction model of our STF framework can be based on any state-of-the-art architecture. In this paper, we choose OneIE (Lin et al., 2020) and AMR-IE (Zhang and Ji, 2021), due
$^2$https://ww.ldc.upenn.edu/collaborations/
past-projects/ace
|
[
[
0.1318359375,
0.0885009765625,
0.86572265625,
0.09027099609375,
0.86572265625,
0.10809326171875,
0.1318359375,
0.105712890625
],
[
0.21826171875,
0.11102294921875,
0.78076171875,
0.11102294921875,
0.78076171875,
0.1258544921875,
0.21826171875,
0.1258544921875
],
[
0.295654296875,
0.141357421875,
0.7109375,
0.14013671875,
0.7109375,
0.1591796875,
0.295654296875,
0.1602783203125
],
[
0.297119140625,
0.157958984375,
0.705078125,
0.1591796875,
0.705078125,
0.177001953125,
0.297119140625,
0.17578125
],
[
0.2493896484375,
0.174560546875,
0.75244140625,
0.177001953125,
0.75244140625,
0.19482421875,
0.2493896484375,
0.1923828125
],
[
0.382080078125,
0.19482421875,
0.623046875,
0.19482421875,
0.623046875,
0.208984375,
0.382080078125,
0.208984375
],
[
0.408935546875,
0.21142578125,
0.59619140625,
0.21142578125,
0.59619140625,
0.2257080078125,
0.408935546875,
0.2257080078125
],
[
0.2626953125,
0.2529296875,
0.343505859375,
0.2529296875,
0.343505859375,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51123046875,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51123046875,
0.267333984375
],
[
0.51123046875,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.145263671875,
0.27783203125,
0.459228515625,
0.27783203125,
0.459228515625,
0.29150390625,
0.145263671875,
0.29150390625
],
[
0.51123046875,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.51123046875,
0.2998046875
],
[
0.1435546875,
0.292236328125,
0.4609375,
0.292236328125,
0.4609375,
0.305908203125,
0.1435546875,
0.305908203125
],
[
0.1435546875,
0.30712890625,
0.45751953125,
0.30712890625,
0.45751953125,
0.320556640625,
0.1435546875,
0.320556640625
],
[
0.51123046875,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.31640625,
0.51123046875,
0.31640625
],
[
0.0235137939453125,
0.308837890625,
0.060455322265625,
0.308837890625,
0.060455322265625,
0.36767578125,
0.0235137939453125,
0.36767578125
],
[
0.145263671875,
0.322509765625,
0.45751953125,
0.322509765625,
0.45751953125,
0.333251953125,
0.145263671875,
0.333251953125
],
[
0.509765625,
0.318359375,
0.8818359375,
0.317626953125,
0.8818359375,
0.33203125,
0.509765625,
0.333251953125
],
[
0.14697265625,
0.33740234375,
0.456787109375,
0.33740234375,
0.456787109375,
0.347412109375,
0.14697265625,
0.347412109375
],
[
0.51318359375,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.145263671875,
0.349853515625,
0.459228515625,
0.349853515625,
0.459228515625,
0.363525390625,
0.145263671875,
0.363525390625
],
[
0.03106689453125,
0.35693359375,
0.0679931640625,
0.35986328125,
0.05792236328125,
0.431640625,
0.0201568603515625,
0.4287109375
],
[
0.51318359375,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.145263671875,
0.364013671875,
0.459228515625,
0.364013671875,
0.459228515625,
0.377685546875,
0.145263671875,
0.377685546875
],
[
0.509765625,
0.365234375,
0.88330078125,
0.36767578125,
0.88330078125,
0.3818359375,
0.509765625,
0.380126953125
],
[
0.145263671875,
0.37890625,
0.459228515625,
0.37890625,
0.459228515625,
0.392578125,
0.145263671875,
0.392578125
],
[
0.51123046875,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.14697265625,
0.393798828125,
0.45751953125,
0.393798828125,
0.45751953125,
0.40380859375,
0.14697265625,
0.40380859375
],
[
0.509765625,
0.39794921875,
0.88330078125,
0.3984375,
0.88330078125,
0.413330078125,
0.509765625,
0.412109375
],
[
0.1435546875,
0.40625,
0.45751953125,
0.40625,
0.45751953125,
0.419189453125,
0.1435546875,
0.419189453125
],
[
0.145263671875,
0.42041015625,
0.45751953125,
0.42041015625,
0.45751953125,
0.43408203125,
0.145263671875,
0.43408203125
],
[
0.51318359375,
0.416259765625,
0.57666015625,
0.416259765625,
0.57666015625,
0.427490234375,
0.51318359375,
0.427490234375
],
[
0.145263671875,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.448974609375,
0.145263671875,
0.448974609375
],
[
0.52880859375,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.52880859375,
0.4453125
],
[
0.0235137939453125,
0.438232421875,
0.0621337890625,
0.438232421875,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51025390625
],
[
0.14697265625,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.509765625,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.145263671875,
0.463134765625,
0.459228515625,
0.463134765625,
0.459228515625,
0.476806640625,
0.145263671875,
0.476806640625
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.145263671875,
0.47802734375,
0.459228515625,
0.47802734375,
0.459228515625,
0.4912109375,
0.145263671875,
0.4912109375
],
[
0.509765625,
0.479736328125,
0.8818359375,
0.478515625,
0.8818359375,
0.493408203125,
0.509765625,
0.49462890625
],
[
0.145263671875,
0.4912109375,
0.456787109375,
0.4912109375,
0.456787109375,
0.5048828125,
0.145263671875,
0.5048828125
],
[
0.51123046875,
0.496337890625,
0.88330078125,
0.496337890625,
0.88330078125,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.14697265625,
0.505859375,
0.459228515625,
0.505859375,
0.459228515625,
0.51953125,
0.14697265625,
0.51953125
],
[
0.019317626953125,
0.51318359375,
0.060455322265625,
0.51318359375,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.509765625,
0.51318359375,
0.8818359375,
0.51318359375,
0.8818359375,
0.52685546875,
0.509765625,
0.52685546875
],
[
0.145263671875,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53369140625,
0.145263671875,
0.53369140625
],
[
0.145263671875,
0.53369140625,
0.459228515625,
0.53369140625,
0.459228515625,
0.54736328125,
0.145263671875,
0.54736328125
],
[
0.51123046875,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.51123046875,
0.5419921875
],
[
0.145263671875,
0.548828125,
0.459228515625,
0.548828125,
0.459228515625,
0.5625,
0.145263671875,
0.5625
],
[
0.51123046875,
0.54443359375,
0.88330078125,
0.54345703125,
0.88330078125,
0.55810546875,
0.51123046875,
0.55859375
],
[
0.145263671875,
0.5625,
0.4609375,
0.5625,
0.4609375,
0.576171875,
0.145263671875,
0.576171875
],
[
0.51318359375,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.145263671875,
0.57666015625,
0.459228515625,
0.57666015625,
0.459228515625,
0.59033203125,
0.145263671875,
0.59033203125
],
[
0.51123046875,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.145263671875,
0.59033203125,
0.459228515625,
0.59033203125,
0.459228515625,
0.60400390625,
0.145263671875,
0.60400390625
],
[
0.51123046875,
0.59326171875,
0.88330078125,
0.59326171875,
0.88330078125,
0.60693359375,
0.51123046875,
0.60693359375
],
[
0.1435546875,
0.60498046875,
0.4609375,
0.60498046875,
0.4609375,
0.6181640625,
0.1435546875,
0.6181640625
],
[
0.51123046875,
0.60791015625,
0.8818359375,
0.60791015625,
0.8818359375,
0.62158203125,
0.51123046875,
0.62158203125
],
[
0.145263671875,
0.619140625,
0.459228515625,
0.619140625,
0.459228515625,
0.6328125,
0.145263671875,
0.6328125
],
[
0.509765625,
0.62451171875,
0.8818359375,
0.62353515625,
0.8818359375,
0.63818359375,
0.509765625,
0.6396484375
],
[
0.145263671875,
0.63427734375,
0.45751953125,
0.63427734375,
0.45751953125,
0.64794921875,
0.145263671875,
0.64794921875
],
[
0.51123046875,
0.64013671875,
0.880859375,
0.64013671875,
0.880859375,
0.65380859375,
0.51123046875,
0.65380859375
],
[
0.145263671875,
0.64794921875,
0.459228515625,
0.64794921875,
0.459228515625,
0.6611328125,
0.145263671875,
0.6611328125
],
[
0.145263671875,
0.6611328125,
0.456787109375,
0.6611328125,
0.456787109375,
0.6748046875,
0.145263671875,
0.6748046875
],
[
0.51318359375,
0.6591796875,
0.880859375,
0.6591796875,
0.880859375,
0.66943359375,
0.51318359375,
0.66943359375
],
[
0.1435546875,
0.6748046875,
0.459228515625,
0.67578125,
0.459228515625,
0.6904296875,
0.1435546875,
0.689453125
],
[
0.51123046875,
0.67333984375,
0.88330078125,
0.67333984375,
0.88330078125,
0.68701171875,
0.51123046875,
0.68701171875
],
[
0.14697265625,
0.6923828125,
0.45751953125,
0.6923828125,
0.45751953125,
0.703125,
0.14697265625,
0.703125
],
[
0.51123046875,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.70361328125,
0.51123046875,
0.70361328125
],
[
0.1435546875,
0.70361328125,
0.459228515625,
0.70361328125,
0.459228515625,
0.71728515625,
0.1435546875,
0.71728515625
],
[
0.5087890625,
0.705078125,
0.880859375,
0.705078125,
0.880859375,
0.71875,
0.5087890625,
0.71875
],
[
0.145263671875,
0.71875,
0.45751953125,
0.71875,
0.45751953125,
0.732421875,
0.145263671875,
0.732421875
],
[
0.51123046875,
0.7216796875,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.51123046875,
0.736328125
],
[
0.1435546875,
0.732421875,
0.4306640625,
0.732421875,
0.4306640625,
0.7451171875,
0.1435546875,
0.7451171875
],
[
0.51123046875,
0.73828125,
0.8818359375,
0.73828125,
0.8818359375,
0.751953125,
0.51123046875,
0.751953125
],
[
0.1368408203125,
0.755859375,
0.26025390625,
0.75830078125,
0.26025390625,
0.7724609375,
0.1368408203125,
0.77099609375
],
[
0.51123046875,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76708984375,
0.51123046875,
0.76708984375
],
[
0.12176513671875,
0.76123046875,
0.1318359375,
0.76123046875,
0.1318359375,
0.7685546875,
0.12176513671875,
0.7685546875
],
[
0.1318359375,
0.76123046875,
0.1427001953125,
0.76123046875,
0.1427001953125,
0.7666015625,
0.1318359375,
0.7666015625
],
[
0.51318359375,
0.76953125,
0.8818359375,
0.76953125,
0.8818359375,
0.783203125,
0.51318359375,
0.783203125
],
[
0.1158447265625,
0.7822265625,
0.486083984375,
0.7822265625,
0.486083984375,
0.7958984375,
0.1158447265625,
0.7958984375
],
[
0.51123046875,
0.78515625,
0.8818359375,
0.7861328125,
0.8818359375,
0.80029296875,
0.51123046875,
0.7998046875
],
[
0.1175537109375,
0.7998046875,
0.489501953125,
0.7998046875,
0.489501953125,
0.8134765625,
0.1175537109375,
0.8134765625
],
[
0.51123046875,
0.8017578125,
0.8330078125,
0.8017578125,
0.8330078125,
0.8154296875,
0.51123046875,
0.8154296875
],
[
0.1158447265625,
0.81640625,
0.490234375,
0.81640625,
0.490234375,
0.830078125,
0.1158447265625,
0.830078125
],
[
0.52880859375,
0.82080078125,
0.87890625,
0.82080078125,
0.87890625,
0.83056640625,
0.52880859375,
0.83056640625
],
[
0.1175537109375,
0.83203125,
0.48779296875,
0.83203125,
0.48779296875,
0.845703125,
0.1175537109375,
0.845703125
],
[
0.51318359375,
0.8349609375,
0.8818359375,
0.8349609375,
0.8818359375,
0.8486328125,
0.51318359375,
0.8486328125
],
[
0.1158447265625,
0.8466796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.8623046875,
0.1158447265625,
0.86083984375
],
[
0.5087890625,
0.85107421875,
0.8818359375,
0.849609375,
0.8818359375,
0.8642578125,
0.5087890625,
0.865234375
],
[
0.14697265625,
0.8681640625,
0.291259765625,
0.8681640625,
0.291259765625,
0.8818359375,
0.14697265625,
0.8818359375
],
[
0.509765625,
0.8662109375,
0.8818359375,
0.865234375,
0.8818359375,
0.8798828125,
0.509765625,
0.88134765625
],
[
0.139404296875,
0.88134765625,
0.489501953125,
0.88134765625,
0.489501953125,
0.89453125,
0.139404296875,
0.89453125
],
[
0.1158447265625,
0.8935546875,
0.48779296875,
0.8935546875,
0.48779296875,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.53076171875,
0.89111328125,
0.86572265625,
0.892578125,
0.86572265625,
0.90673828125,
0.53076171875,
0.90625
],
[
0.11419677734375,
0.90380859375,
0.353515625,
0.90625,
0.353515625,
0.91943359375,
0.11419677734375,
0.91748046875
],
[
0.51123046875,
0.90625,
0.6533203125,
0.90625,
0.6533203125,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"Learning from a Friend: Improving Event Extraction via Self-Training",
"with Feedback fromAbstract Meaning Representation",
"Zhiyang Xu* Jay-Yoon Lee** Lifu Huang**",
"*Department of Computer Science, Virginia Tech",
"*Graduate School of Data Science, Seoul National University",
"(zhiyangx,lifuh}@vt.edu",
"lee.jayyoon @ snu.ac.kr",
"Abstract",
"studies (Deng et al., 2021; Wang et al., 2021b)",
"have shown that the data scarcity of existing event",
"Data scarcity has been the main factor that hin-",
"annotations has been the major issue that hinders",
"ders the progress of event extraction. To over-",
"come this issue, we propose a Self-Training",
"the progress of EE. For example, in ACE-05?, one",
"",
"with Feedback (STF) framework that lever-",
"of the most popular event extraction benchmark",
"ages thelarge-scaleunlabeled data and acquires",
"datasets, 10 of the 33 event types have less than 80",
"feedback for each new event prediction from",
"o",
"annotations. However, creating event annotations",
"the unlabeled data by comparing it to the Ab-",
"is extremely expensive and time-consuming, e.g.,",
"stract Meaning Representation (AMR) graph of",
"it takes several linguists over one year to annotate",
"the same sentence.Specifically,STFconsists",
"500 documents with about 5000 event mentions for",
"of (1) a base event extraction model trained on",
"existing event annotations and then applied to",
"ACE-05.",
"large-scale unlabeled corpora to predict new",
"To overcome the data scarcity issue of EE, pre-",
"['s]",
"eventmentions aspseudotrainingsamples,and",
"vious studies (Chen and Ji, 2009; Liao and Gr",
"(2) a novel scoring model that takes in each",
"ishman, 201la; Ferguson et al., 2018a) develop",
"new predicted event trigger, an argument, its ar-",
"self-training methods that allow the trained EE",
"gument role, as well as their paths in the AMR",
"model to learn further by regarding its own predic-",
"graph to estimate a compatibility score indicat-",
"652:",
"tions on large-scale unlabeled corpora as pseudo la-",
"ing the correctness of the pseudo label. The",
"compatibility scores further act as feedback to",
"bels. However, simply adding the high-confidence",
"encourage or discourage the model learning on",
"event predictions to the training set inevitably intro-",
"the pseudo labels during self-training. Exper-",
"duces noises (Liu et al., 2021; Arazo et al., 2020;",
"imental results on three benchmark datasets,",
"Jiang et al., 2018), especially given that the current",
"including ACE05-E, ACE05-E+,and ERE,",
"state-of-the-art performance of event argument ex-",
"demonstrate the effectiveness of the STF frame-",
"traction is still less than 60% F-score. To tackle",
"work on event extraction, especially event argu-",
"this challenge, we introduce a Self-Training with",
"ment extraction, with significant performance",
"Feedback framework, named STF, which consists",
"gain over the base event extraction models and",
"strong baselines. Our experimental analysis",
"of an eventextractionmodel thatisfirstlytrained",
"further shows that STF is a generic framework",
"on the existing event annotations and then contin-",
"as it can be applied to improve most, if not all",
"ually updated on the unlabeled corpus with self-",
"event extraction models by leveraging large-",
" training, and a scoring model that is to evaluate the",
"scale unlabeled data, even when high-quality",
"correctness of the new event predictions (pseudo",
"AMR graph annotations are not available.",
"labels) from the unlabeled corpus, and the scores",
"Introduction",
"further act as feedback to encourage or discourage",
"1",
"",
"the learning of the event extraction model on the",
"Event extraction (EE), which aims to identify and",
"pseudo labels during self-training, inspired by the",
"classify event triggers and arguments, has been",
"REINFORCE algorithms (Williams, 1992).",
"a long-stand challenging problem in natural lan-",
"Specifically,theeventextractionmodel of our",
"guage processing. Despite the large performance",
"STF framework can be based on any state-of-the-art",
"leap brought by advances in deep learning, recent",
"architecture. In this paper, we choose OneIE (Lin",
"* corresponding authors",
"et al., 2020) and AMR-IE (Zhang and Ji, 2021), due",
"IThe source code and model checkpoints are pub-",
"licly available at https://github.com/VT-NLP/Event-",
"https://www.ldc.upenn.edu/collaborations/",
"Extraction_with_Self_Training.",
"past-projects/ace"
] |
{
"arxiv_id": "2205.12490",
"language": "en",
"page_idx": 0
}
|
||
2205.12490-en-1
|
2205.12490-en-1.jpg
|
to their superior performance and publicly available source code. The scoring model leverages the Abstract Meaning Representation (AMR) (Banarescu et al., 2013) which has been proven to be able to provide rich semantic and structural signals to map AMR structures to event predictions (Huang et al., 2016,2018; Wang et al., 2021b) and thus their compatibility can indicate the correctness of each event prediction. The scoring model is a self-attention network that takes in a predicted event trigger, a candidate argument and its argument role, as well as their path in the AMR graph of the whole sentence, and computes a score ranging in [-1, 1] based on the compatibility between the AMR and the predicted event structure: -1 means incompatible, 1 means compatible, and 0 means uncertain. Inspired by the REINFORCE algorithm (Williams, 1992), we multiply the compatibility scores and the gradient of the EE model computed on the pseudo event labels during self-training, so as to (1) encourage the event extraction model to follow the gradient and hence maximize the likelihood of the pseudo label when it is compatible with the AMR structure; (2) negate the gradient and minimize the likelihood of the pseudo label when it is incompatible with the AMR structure; and (3) reduce the magnitude of the gradient when the scoring model is uncertain about the correctness of the pseudo label.
We take AMR 3.03 and part of the New York Times (NYT) 2004 corpus$^{4}$ as additional unlabeled corpora to enhance the event extraction model with STF, and evaluate the event extraction performance on three public benchmark datasets: ACE05-E$^{5}$, ACE05-E$^{+6}$, and ERE-EN$^7.$ The experimental results demonstrate that: (1) the vanilla self-training barely improves event extraction due to the noise introduced by the pseudo examples, while the proposed STF framework leverages the compatibility scores from the scoring model as feedback and thus makes more robust and efficient use of the pseudo labels; (2) STF is a generic framework and can be applied to improve most, if not all, of the event extraction models optimized by gradient descent algorithm and achieves significant improvement over the base event extraction models and strong baselines on event argument extraction on the three public benchmark datasets; (3) By exploiting dif-
$^{3}$https://catalog.ldc.upenn.edu/LDC2020T02. $_{\mathrm{e}}^{4}$https://catalog.ldc.upenn.edu/LDC2008T19 $^{5}$https://catalog.ldc.upenn.edu/LDC2006T06 $^{6}$https://catalog.ldc.upenn.edu/LDC2006T06 $^{7}$Deep Exploration and Filtering of Test (DEFT) program.
ferent unlabeled corpora with gold or system-based AMR parsing, STF always improves the base event extraction models, demonstrating that it can work with various qualities of AMR parsing. Notably, different from previous studies (Huang et al., 2018; Zhang and Ji, 2021; Wang et al., 2021b) that require high-quality AMR graphs as input to the model during both training and inference, STF does not require any AMR graphs during inference, making it more computationally efficient and free from the potential errors propagated from AMR parsing.
## 2 STF for Event Extraction
The event extraction task consists of three subtasks: event detection, argument identification and argument role classification. Given an input sentence $W$ = $[ w_1, w_2, . . . , w_N] $, event detection aims to identify the span of an event trigger $\tau_i$ in $W$ and assign a label $l_{\tau_i}\in\mathcal{T}$ where $\mathcal{T}$ denotes the set of target event types. Argument identification aims to find the span of an argument $\varepsilon_j$ in $W$, and $ar-$ gument role classification further predicts a role $\alpha_{ij}\in\mathcal{A}$ that the argument $\varepsilon_j$ plays in an event $\tau_i$ given the set of target argument roles $\mathcal{A}.$
Figure 1 shows the overview of our STF framework which consists of two training stages. At the first stage, a base event extraction model (Section 2.1) is trained on a labeled dataset. At the second stage, we apply the trained event extraction model to an unlabeled corpus to predict new event mentions. Instead of directly taking the new event predictions as pseudo training examples like the vanilla self-training, we propose a novel scoring model (Section 2.2) to estimate the correctness of each event prediction by measuring its compatibility to the corresponding AMR graph, and then take both event predictions and their compatibility scores to continue to train the base event extraction model while the scores update the gradient computed on pseudo labels (Section 2.3). After the training of the second stage, we get a new event extraction model and evaluate it on the test set.
### 2.1 Base Event Extraction Model
Our proposed framework can be applied to most, if not all, event extraction models. We select OneIE (Lin et al., 2020) and AMR-IE (Zhang and Ji, 2021) as base models given their state-of-theart performance on the event extraction task and publicly available source code. Next, we briefly describe the common architectures in the two models
|
[
[
0.1158447265625,
0.08489990234375,
0.48779296875,
0.08489990234375,
0.48779296875,
0.09857177734375,
0.1158447265625,
0.09857177734375
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.490234375,
0.1015625,
0.490234375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.5087890625,
0.1015625,
0.880859375,
0.1015625,
0.880859375,
0.11517333984375,
0.5087890625,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
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,
0.1484375,
0.509765625,
0.1484375
],
[
0.1158447265625,
0.1484375,
0.48779296875,
0.149658203125,
0.48779296875,
0.1644287109375,
0.1158447265625,
0.163330078125
],
[
0.509765625,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.160888671875,
0.509765625,
0.160888671875
],
[
0.118408203125,
0.166259765625,
0.49365234375,
0.166259765625,
0.49365234375,
0.179931640625,
0.118408203125,
0.179931640625
],
[
0.51123046875,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.1175537109375,
0.1817626953125,
0.491943359375,
0.1817626953125,
0.491943359375,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51123046875,
0.18115234375,
0.8818359375,
0.1817626953125,
0.8818359375,
0.196533203125,
0.51123046875,
0.1953125
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1158447265625,
0.215576171875,
0.489501953125,
0.215576171875,
0.489501953125,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.50537109375,
0.215576171875,
0.87890625,
0.215576171875,
0.87890625,
0.2286376953125,
0.50537109375,
0.2286376953125
],
[
0.1158447265625,
0.232177734375,
0.489501953125,
0.232177734375,
0.489501953125,
0.2452392578125,
0.1158447265625,
0.2452392578125
],
[
0.5087890625,
0.23095703125,
0.880859375,
0.23095703125,
0.880859375,
0.24462890625,
0.5087890625,
0.24462890625
],
[
0.1158447265625,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.5087890625,
0.2476806640625,
0.85986328125,
0.2476806640625,
0.85986328125,
0.262451171875,
0.5087890625,
0.262451171875
],
[
0.1158447265625,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.1158447265625,
0.27978515625,
0.489501953125,
0.27978515625,
0.489501953125,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.51123046875,
0.27978515625,
0.54248046875,
0.27978515625,
0.54248046875,
0.2880859375,
0.51123046875,
0.2880859375
],
[
0.53564453125,
0.275634765625,
0.76171875,
0.276611328125,
0.76171875,
0.29150390625,
0.53564453125,
0.290283203125
],
[
0.1175537109375,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.1158447265625,
0.31005859375,
0.486083984375,
0.31005859375,
0.486083984375,
0.32373046875,
0.1158447265625,
0.32373046875
],
[
0.51123046875,
0.305908203125,
0.88330078125,
0.305908203125,
0.88330078125,
0.31640625,
0.51123046875,
0.31640625
],
[
0.1175537109375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.1175537109375,
0.34033203125
],
[
0.51123046875,
0.32177734375,
0.88330078125,
0.32177734375,
0.88330078125,
0.3349609375,
0.51123046875,
0.3349609375
],
[
0.1175537109375,
0.342529296875,
0.48779296875,
0.342529296875,
0.48779296875,
0.355712890625,
0.1175537109375,
0.355712890625
],
[
0.51318359375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.51123046875,
0.3515625,
0.8818359375,
0.352783203125,
0.8818359375,
0.36767578125,
0.51123046875,
0.366455078125
],
[
0.1175537109375,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.372314453125,
0.1175537109375,
0.372314453125
],
[
0.51318359375,
0.369384765625,
0.8818359375,
0.369384765625,
0.8818359375,
0.383056640625,
0.51318359375,
0.383056640625
],
[
0.1175537109375,
0.3759765625,
0.489501953125,
0.3759765625,
0.489501953125,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.11419677734375,
0.39013671875,
0.48779296875,
0.391357421875,
0.48779296875,
0.40625,
0.11419677734375,
0.405029296875
],
[
0.51318359375,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3984375,
0.51318359375,
0.3984375
],
[
0.51318359375,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.51318359375,
0.4150390625
],
[
0.1158447265625,
0.407958984375,
0.486083984375,
0.407958984375,
0.486083984375,
0.421630859375,
0.1158447265625,
0.421630859375
],
[
0.509765625,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.509765625,
0.431640625
],
[
0.118408203125,
0.42578125,
0.486083984375,
0.42578125,
0.486083984375,
0.435791015625,
0.118408203125,
0.435791015625
],
[
0.51123046875,
0.43408203125,
0.8818359375,
0.432861328125,
0.8818359375,
0.446533203125,
0.51123046875,
0.44775390625
],
[
0.1175537109375,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.45361328125,
0.1175537109375,
0.45361328125
],
[
0.51318359375,
0.449462890625,
0.8818359375,
0.449462890625,
0.8818359375,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.1175537109375,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.509765625,
0.466064453125,
0.8076171875,
0.466064453125,
0.8076171875,
0.479736328125,
0.509765625,
0.479736328125
],
[
0.1175537109375,
0.47265625,
0.48779296875,
0.47265625,
0.48779296875,
0.486328125,
0.1175537109375,
0.486328125
],
[
0.1175537109375,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.1175537109375,
0.501953125
],
[
0.53076171875,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.53076171875,
0.496337890625
],
[
0.1158447265625,
0.5029296875,
0.48779296875,
0.50439453125,
0.48779296875,
0.5185546875,
0.1158447265625,
0.51708984375
],
[
0.51123046875,
0.49951171875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.51123046875,
0.51318359375
],
[
0.1175537109375,
0.52099609375,
0.424072265625,
0.52099609375,
0.424072265625,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.509765625,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.509765625,
0.529296875
],
[
0.509765625,
0.53076171875,
0.8818359375,
0.53076171875,
0.8818359375,
0.54443359375,
0.509765625,
0.54443359375
],
[
0.1351318359375,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.55029296875,
0.1351318359375,
0.55029296875
],
[
0.1158447265625,
0.552734375,
0.48779296875,
0.552734375,
0.48779296875,
0.56640625,
0.1158447265625,
0.56640625
],
[
0.509765625,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.509765625,
0.56103515625
],
[
0.509765625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.509765625,
0.57666015625
],
[
0.1175537109375,
0.5693359375,
0.486083984375,
0.5693359375,
0.486083984375,
0.5830078125,
0.1175537109375,
0.5830078125
],
[
0.51123046875,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.5927734375,
0.51123046875,
0.5927734375
],
[
0.118408203125,
0.5859375,
0.484375,
0.5859375,
0.484375,
0.5966796875,
0.118408203125,
0.5966796875
],
[
0.1158447265625,
0.60107421875,
0.489501953125,
0.599609375,
0.489501953125,
0.6142578125,
0.1158447265625,
0.615234375
],
[
0.51123046875,
0.595703125,
0.8818359375,
0.595703125,
0.8818359375,
0.609375,
0.51123046875,
0.609375
],
[
0.5087890625,
0.60986328125,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.5087890625,
0.62451171875
],
[
0.1175537109375,
0.6162109375,
0.489501953125,
0.6162109375,
0.489501953125,
0.63134765625,
0.1175537109375,
0.63134765625
],
[
0.509765625,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.509765625,
0.6396484375
],
[
0.1175537109375,
0.63427734375,
0.48779296875,
0.63427734375,
0.48779296875,
0.64794921875,
0.1175537109375,
0.64794921875
],
[
0.1175537109375,
0.6494140625,
0.486083984375,
0.6494140625,
0.486083984375,
0.6630859375,
0.1175537109375,
0.6630859375
],
[
0.5146484375,
0.6455078125,
0.880859375,
0.6455078125,
0.880859375,
0.65625,
0.5146484375,
0.65625
],
[
0.51123046875,
0.6591796875,
0.880859375,
0.6591796875,
0.880859375,
0.6728515625,
0.51123046875,
0.6728515625
],
[
0.1158447265625,
0.66650390625,
0.490234375,
0.66650390625,
0.490234375,
0.6796875,
0.1158447265625,
0.6796875
],
[
0.51123046875,
0.6748046875,
0.880859375,
0.67578125,
0.880859375,
0.6904296875,
0.51123046875,
0.689453125
],
[
0.1158447265625,
0.68310546875,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1158447265625,
0.69677734375
],
[
0.51318359375,
0.6923828125,
0.87890625,
0.6923828125,
0.87890625,
0.703125,
0.51318359375,
0.703125
],
[
0.1175537109375,
0.69873046875,
0.48779296875,
0.69873046875,
0.48779296875,
0.7119140625,
0.1175537109375,
0.7119140625
],
[
0.509765625,
0.7080078125,
0.88330078125,
0.7080078125,
0.88330078125,
0.7216796875,
0.509765625,
0.7216796875
],
[
0.1175537109375,
0.71435546875,
0.48779296875,
0.71435546875,
0.48779296875,
0.72802734375,
0.1175537109375,
0.72802734375
],
[
0.51123046875,
0.7236328125,
0.8818359375,
0.72265625,
0.8818359375,
0.73681640625,
0.51123046875,
0.73828125
],
[
0.1175537109375,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.74462890625,
0.1175537109375,
0.74462890625
],
[
0.509765625,
0.74072265625,
0.8818359375,
0.74072265625,
0.8818359375,
0.75439453125,
0.509765625,
0.75439453125
],
[
0.1158447265625,
0.74658203125,
0.489501953125,
0.7451171875,
0.489501953125,
0.76025390625,
0.1158447265625,
0.76123046875
],
[
0.509765625,
0.755859375,
0.8564453125,
0.7548828125,
0.8564453125,
0.76953125,
0.509765625,
0.77099609375
],
[
0.1175537109375,
0.76318359375,
0.48779296875,
0.76318359375,
0.48779296875,
0.77685546875,
0.1175537109375,
0.77685546875
],
[
0.1175537109375,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.1175537109375,
0.79150390625
],
[
0.51318359375,
0.78515625,
0.78515625,
0.78515625,
0.78515625,
0.798828125,
0.51318359375,
0.798828125
],
[
0.11419677734375,
0.79345703125,
0.486083984375,
0.79443359375,
0.486083984375,
0.80859375,
0.11419677734375,
0.80810546875
],
[
0.1175537109375,
0.8125,
0.484375,
0.8125,
0.484375,
0.822265625,
0.1175537109375,
0.822265625
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1158447265625,
0.82666015625,
0.490234375,
0.826171875,
0.490234375,
0.84033203125,
0.1158447265625,
0.84130859375
],
[
0.5087890625,
0.82373046875,
0.8818359375,
0.822265625,
0.8818359375,
0.83740234375,
0.5087890625,
0.83837890625
],
[
0.51123046875,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.1376953125,
0.8525390625,
0.411376953125,
0.8515625,
0.411376953125,
0.8662109375,
0.1376953125,
0.86767578125
],
[
0.509765625,
0.85546875,
0.88330078125,
0.85546875,
0.88330078125,
0.869140625,
0.509765625,
0.869140625
],
[
0.1376953125,
0.8662109375,
0.41064453125,
0.8662109375,
0.41064453125,
0.8798828125,
0.1376953125,
0.8798828125
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1376953125,
0.8798828125,
0.41064453125,
0.8798828125,
0.41064453125,
0.8935546875,
0.1376953125,
0.8935546875
],
[
0.139404296875,
0.892578125,
0.411376953125,
0.892578125,
0.411376953125,
0.90625,
0.139404296875,
0.90625
],
[
0.51318359375,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90185546875,
0.51318359375,
0.90185546875
],
[
0.139404296875,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.139404296875,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"to their superior performance and publicly available",
"ferent unlabeled corpora with gold or system-based",
"source code. The scoring model leverages the Ab-",
"AMR parsing, STF always improves the base event",
"stract Meaning Representation (AMR) (Banarescu",
"extraction models, demonstrating that it can work",
"et al., 2013) which has been proven to be able to",
"with various qualities of AMR parsing. Notably.",
"provide rich semantic and structural signals to map",
"differentfromprevious studies (Huang etal..2018",
"AMR structures to event predictions (Huang et al.,",
"Zhang and Ji, 2021; Wang et al., 2021b) that require",
"2016,2018; Wang et al., 2021b) and thus their com-",
"high-quality AMR graphs as input to the model",
"patibility can indicate the correctness of each event",
"during both training and inference, STF does not",
"prediction. The scoring model is a self-attention",
"require any AMR graphs during inference, making",
"network that takes in a predicted event trigger, a",
" it more computationally efficient and free from the",
"candidate argument and its argument role, as well",
"potential errors propagated from AMR parsing.",
"as their path in the AMR graph of the whole sen-",
"tence, and computes a score ranging in [-1, 1] based",
"2",
"STF for Event Extraction",
"on the compatibility between the AMR and the pre-",
"dicted event structure: -1 means incompatible, 1",
"Theeventextractiontaskconsistsofthreesubtasks:",
"means compatible, and O means uncertain. Inspired",
"event detection, argument identification and argu-",
"by the REINFORCE algorithm (Williams,1992),",
"ment role classification. Given an input sentence",
"W = [wi, w2,.., wn], event detection aims to",
"we multiply the compatibility scores and the gradi-",
"identify the span of an event trigger T; in W and",
"ent of the EE model computed on the pseudo event",
"labels during self-training, so as to (1) encourage",
"assign a label lt, E T where T denotes the set of",
"target event types. Argument identification aims",
"the event extraction model to follow the gradient",
"to find the span of an argument ej in W, and ar-",
"andhencemaximizethelikelihoodofthepseudo",
"gument role classification further predicts a role",
"label when it is compatible with the AMR structure;",
"Qii E A that the argument ej plays in an event Ti",
"(2) negate the gradient and minimize the likelihood",
" given the set of target argument roles A.",
"of the pseudo label when it is incompatible with",
"the AMR structure; and (3) reduce the magnitude",
"Figure 1 shows the overview of our STF frame-",
"of the gradient when the scoring model is uncertain",
"work which consists of two training stages. At the",
"about the correctness of the pseudo label.",
"first stage, a base event extraction model (Sec",
"tion 2.1) is trained on a labeled dataset. At the",
"We take AMR 3.03 and part of the New York",
"Times (NYT) 2004 corpus4 as additional unlabeled",
"second stage, we apply the trained event extraction",
"model to an unlabeled corpus to predict new event",
"corpora to enhance the event extraction model with",
"mentions. Instead of directly taking the new event",
"STE.and evaluatetheeventextractionperformance",
"on three public benchmark datasets: ACEO5-E5,",
"predictions as pseudo training examples like the",
"vanilla self-training, we propose a novel scoring",
"ACE05-E+6, and ERE-EN7. The experimental re-",
"model (Section 2.2) to estimate the correctness of",
"sults demonstrate that: (1) the vanilla self-training",
"barely improves event extraction due to the noise",
"each eventpredictionby measuringits compati-",
"bility to the corresponding AMR graph, and then",
"introduced by the pseudo examples, while the pro-",
"take both event predictions and their compatibility",
"posed STF framework leverages the compatibility",
"scores to continue to train the base event extraction",
"scores from the scoring model as feedback and thus",
"model while the scores update the gradient com-",
"makes more robust and efficient use of the pseudo",
"puted on pseudo labels (Section 2.3). After the",
"labels; (2) STF is a generic framework and can be",
"training of the second stage, we get a new event",
"applied to improve most, if not all, of the event",
"extraction model and evaluate it on the test set.",
"extraction models optimized by gradient descent",
"algorithm and achieves significant improvement",
"2.1 Base Event Extraction Model",
"over the base event extraction models and strong",
"baselines on eventargumentextraction onthethree",
"Our proposed framework can be applied to most,",
"public benchmark datasets; (3) By exploiting dif-",
"if not all, event extraction models.We select",
"OneIE (Lin et al., 2020) and AMR-IE (Zhang and",
"3https://catalog.ldc.upenn.edu/LDC2020T02.",
"Ji, 2021) as base models given their state-of-the-",
"4https://catalog.ldc.upenn.edu/LDC2008T19",
"art performance on the event extraction task and",
"\"https://catalog.ldc.upenn.edu/LDC2006T06",
"‘https:/catalog.ldc.upenn.edu/LDC2006T06",
"publicly available source code. Next, we briefly de-",
"7Deep Exploration and Filtering of Test (DEFT) program.",
"scribe the common architectures in the two models"
] |
{
"arxiv_id": "2205.12490",
"language": "en",
"page_idx": 1
}
|
||
2205.12490-en-3
|
2205.12490-en-3.jpg
|
has also been discussed and shown effective in previous studies (Rao et al., 2017; Huang et al., 2018; Zhang and Ji, 2021). However, previous studies directly take AMR graphs as input to the extraction model and thus require AMR graphs during both training and inference, making their performance highly dependent on the quality of AMR parsing. Different from them, our proposed STF only takes AMR graphs during reference to measure the correctness of event predictions during self-training, making it free from the potential errors propagation from AMR parsing during inference.
Given a sentence $\tilde{W}\in\mathcal{X}_u$ from the unlabeled corpus and a predicted trigger $\tilde{\tau}_i$ and its argument $\tilde{\varepsilon}_{j}$ from $\tilde{W}$, we aim to estimate a correctness score for each pair of the trigger and argument prediction based on its compatibility with their path in the AMR graph$^{9}$. Thus, we first apply the state-of-theart AMR parsing tool (Astudillo et al., 2020) to generate an AMR graph for $\tilde{W}:G=(V,E),E=$ $\{(v_{i},e_{ij},v_{j})|e_{ij}\in\mathcal{R})\}.$We follow(Huang et al., 2016; Zhang and Ji, 2021) and group the original set of AMR relations into 19 categories$^{10}$, thus $e_{ij}$ denotes a particular relation category and $\mathcal{R}$ Finally, we compute an overall vector represendenotes the set of AMR relation categories. Then, tation $\hat{\mathbf{H}}_{ij}^{final}$ from $\mathbf{H}_{ij}^{final}$ via average-pooling and
we identify the $v_i,v_j$ from AMR graph $G$ as the corresponding node of $\tilde{\tau}_i,\tilde{\varepsilon}_j$,by node alignment following Zhang and Ji (2021). Then, we utilize the Breadth First Search to find the shortest path $p_{i,j}$ that connects and includes, $v_i$ and $v_j$ in $G.$ If there is no path between $v_i$ and $v_j$,we add a new edge to connect them and assign other as the relation.
Given a predicted trigger $\tilde{\tau}_{i}$ and its type $\tilde{l}_{\tau_{i}}$,a predicted argument $\tilde{\varepsilon}_j$ and its argument role $\tilde{\alpha}_{ij}$,the dicted argument $\varepsilon _{j}$ and its argument role $\alpha _{ij}$,the cates the argument role is correct $( y_{ij}= 1)$ or scoring model estimates their correctness by taking not $( y_{ij}= 0) ^{12}$, and $\psi$ is the parameters of the
$[\tilde{l}_{\tau_i},p_{ij},\tilde{\alpha}_{ij}]$ as input and outputs a compatibility score. As Figure 1 shows, it consists of a language model encoder (Devlin et al., 2019; Liu et al., 2019) to encode the sentence $\tilde{W}$ and obtain the contextual representations for the tokens$^{11}$,which are then used to initialize the representation of each node in $p_{ij}$ based on the alignment between the input tokens and the nodes in AMR graph following Zhang and Ji (2021). We draw edge representations from the AMR relation embedding matrix $\mathbf{E}^{rel}$ and com-
$^{9}$Comparing with the whole AMR graph, the path of the trigger and argument in the AMR graph shows more improvement for the scoring model.
The details of AMR relation categories are shown in Ap-
pendix A.
$^{11}$If a token is split into multiple subtokens, we average the representations of all subtokens to obtain an overall token representation.
bine them with node representations to form $\mathbf{H}_{p_{ij}}$, a representation for path $p_{ij}.$ We also get an event type representation $\mathbf{h}_{\tau_i}$ for $\tilde{l}_{\tau_i}$ from the event-type embedding matrix $\mathbf{E}^{tri}$ and an argument role representation h$_{\alpha_{ij}}$ for $\tilde{\alpha}_{ij}$ from the argument role embedding matrix $\mathbf{E}^{arg}.$ Here, $\mathbf{E}^{rel},\mathbf{E}^{tri}$, and $\mathbf{E}^{arg}$ are all randomly initialized and will be optimized during training. Finally, we obtain the initial representations $\mathbf{H}_{ij}^{init}=[\mathbf{h}_{\tau_{i}},\mathbf{H}_{p_{ij}},\mathbf{h}_{\alpha_{ij}}]$ for the sequence $[\tilde{l}_{\tau_i},p_{ij},\tilde{\alpha}_{ij}].$
To estimate the compatibility between the event trigger and argument prediction and their path in the AMR graph, we apply multi-layer selfattention (Vaswani et al., 2017) over the joint representation of the AMR path and the event prediction $\mathbf{H}_{ij}^{init}$ to learn better contextual representations for the sequence $[\tilde{l}_{\tau_i},p_{ij},\tilde{\alpha}_{ij}]$ and we add the position embedding E$^{pos}$ to H$_ij^{init}$ before feed it into the self-attention layers:
$$\mathbf{H}_{ij}^{final}=\textit{self-attention}(\mathbf{H}_{ij}^{init})\times M,$$
where $M$ denotes the number of attention layers. feed it into a linear-layer and a Sigmoid function to compute a probability $c_ij$, indicating the correctness of the predicted event trigger and argument. We optimize the scoring model based on the binary cross-entropy objective:
$$\mathbf{L}^{\mathrm{Score}}=\mathrm{BCE}\left(y_{ij},c_{ij};\psi\right),$$
where $y_{ij}\in$ ( 0, 1) is a binary label that indi- scoring model. During training, we have gold triggers and arguments as positive training instances and we swap the argument roles in positive training instances with randomly sampled incorrect labels to create negative training instances. After training the scoring model, we will fx its parameters and apply it to self-training.
## 2.3 Self-Training with Feedback
To improve the base event extraction model with self-training, we take the new event predictions $(\tilde{\tau}_{i},\tilde{l}_{\tau_{i}},\tilde{\varepsilon}_{j},\tilde{\alpha}_{ij})$ from the unlabeled corpus $\mathcal{X}_u$ as pseudo samples to further train the event extraction
$^{12}$We don't not consider the cases where the trigger labels are incorrect, since by observation the semantics and structure of AMR graphs are more related to the argument role types between event triggers and their arguments.
|
[
[
0.1175537109375,
0.08612060546875,
0.491943359375,
0.08612060546875,
0.491943359375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.0831298828125,
0.88330078125,
0.08489990234375,
0.88330078125,
0.10272216796875,
0.5087890625,
0.1009521484375
],
[
0.1158447265625,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.5087890625,
0.1015625,
0.8818359375,
0.1009521484375,
0.8818359375,
0.11517333984375,
0.5087890625,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.486083984375,
0.133056640625,
0.486083984375,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.509765625,
0.1312255859375,
0.88330078125,
0.1318359375,
0.88330078125,
0.149658203125,
0.509765625,
0.1484375
],
[
0.1175537109375,
0.15087890625,
0.48779296875,
0.15087890625,
0.48779296875,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.509765625,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1650390625,
0.509765625,
0.1650390625
],
[
0.1175537109375,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51123046875,
0.1650390625,
0.8818359375,
0.1650390625,
0.8818359375,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.1175537109375,
0.1817626953125,
0.489501953125,
0.1817626953125,
0.489501953125,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51318359375,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51318359375,
0.1953125
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.215576171875,
0.490234375,
0.215576171875,
0.490234375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.50732421875,
0.2120361328125,
0.88330078125,
0.2120361328125,
0.88330078125,
0.2298583984375,
0.50732421875,
0.2298583984375
],
[
0.1158447265625,
0.2298583984375,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.1158447265625,
0.24462890625
],
[
0.50732421875,
0.23095703125,
0.67919921875,
0.2327880859375,
0.67919921875,
0.2493896484375,
0.50732421875,
0.2476806640625
],
[
0.1175537109375,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.52880859375,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.52880859375,
0.26318359375
],
[
0.1175537109375,
0.26318359375,
0.391357421875,
0.26318359375,
0.391357421875,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.51318359375,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.51318359375,
0.27978515625
],
[
0.1351318359375,
0.27978515625,
0.48779296875,
0.27978515625,
0.48779296875,
0.29345703125,
0.1351318359375,
0.29345703125
],
[
0.509765625,
0.281982421875,
0.8818359375,
0.281982421875,
0.8818359375,
0.295654296875,
0.509765625,
0.295654296875
],
[
0.1175537109375,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.1175537109375,
0.31005859375
],
[
0.509765625,
0.297607421875,
0.8818359375,
0.297607421875,
0.8818359375,
0.311279296875,
0.509765625,
0.311279296875
],
[
0.1158447265625,
0.311279296875,
0.489501953125,
0.31005859375,
0.489501953125,
0.32470703125,
0.1158447265625,
0.325927734375
],
[
0.51123046875,
0.314208984375,
0.8818359375,
0.314208984375,
0.8818359375,
0.327880859375,
0.51123046875,
0.327880859375
],
[
0.118408203125,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.342529296875,
0.118408203125,
0.342529296875
],
[
0.51123046875,
0.32666015625,
0.8818359375,
0.328857421875,
0.8818359375,
0.3466796875,
0.51123046875,
0.344482421875
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.509765625,
0.347412109375,
0.880859375,
0.347412109375,
0.880859375,
0.3623046875,
0.509765625,
0.3623046875
],
[
0.1175537109375,
0.35986328125,
0.489501953125,
0.35986328125,
0.489501953125,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.50732421875,
0.3623046875,
0.88330078125,
0.361083984375,
0.88330078125,
0.37890625,
0.50732421875,
0.380126953125
],
[
0.1175537109375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.51123046875,
0.38134765625,
0.66650390625,
0.38134765625,
0.66650390625,
0.394287109375,
0.51123046875,
0.394287109375
],
[
0.1175537109375,
0.391357421875,
0.483642578125,
0.39013671875,
0.483642578125,
0.405029296875,
0.1175537109375,
0.40625
],
[
0.1175537109375,
0.40625,
0.489501953125,
0.405029296875,
0.489501953125,
0.4228515625,
0.1175537109375,
0.42333984375
],
[
0.5556640625,
0.4091796875,
0.609375,
0.3984375,
0.615234375,
0.41455078125,
0.5615234375,
0.424560546875
],
[
0.61865234375,
0.407958984375,
0.83447265625,
0.407958984375,
0.83447265625,
0.4228515625,
0.61865234375,
0.4228515625
],
[
0.1158447265625,
0.42333984375,
0.484375,
0.42333984375,
0.484375,
0.438232421875,
0.1158447265625,
0.438232421875
],
[
0.1175537109375,
0.440673828125,
0.48779296875,
0.440673828125,
0.48779296875,
0.45361328125,
0.1175537109375,
0.45361328125
],
[
0.509765625,
0.435302734375,
0.87158203125,
0.43701171875,
0.87158203125,
0.451904296875,
0.509765625,
0.449462890625
],
[
0.1175537109375,
0.457275390625,
0.486083984375,
0.457275390625,
0.486083984375,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.52880859375,
0.453125,
0.8818359375,
0.453125,
0.8818359375,
0.466064453125,
0.52880859375,
0.466064453125
],
[
0.50537109375,
0.460205078125,
0.88671875,
0.46142578125,
0.88671875,
0.486328125,
0.50537109375,
0.485107421875
],
[
0.1175537109375,
0.47265625,
0.489501953125,
0.47265625,
0.489501953125,
0.486328125,
0.1175537109375,
0.486328125
],
[
0.11419677734375,
0.48681640625,
0.489501953125,
0.488037109375,
0.489501953125,
0.5029296875,
0.11419677734375,
0.501953125
],
[
0.51123046875,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.51123046875,
0.498779296875
],
[
0.1175537109375,
0.5048828125,
0.489501953125,
0.5048828125,
0.489501953125,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.509765625,
0.5029296875,
0.880859375,
0.5029296875,
0.880859375,
0.5166015625,
0.509765625,
0.5166015625
],
[
0.1175537109375,
0.52099609375,
0.489501953125,
0.52099609375,
0.489501953125,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.5087890625,
0.5185546875,
0.880859375,
0.5185546875,
0.880859375,
0.5322265625,
0.5087890625,
0.5322265625
],
[
0.1175537109375,
0.53515625,
0.48779296875,
0.53759765625,
0.48779296875,
0.5517578125,
0.1175537109375,
0.5498046875
],
[
0.509765625,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.509765625,
0.54638671875
],
[
0.1158447265625,
0.5517578125,
0.486083984375,
0.5517578125,
0.486083984375,
0.5654296875,
0.1158447265625,
0.5654296875
],
[
0.509765625,
0.5498046875,
0.6943359375,
0.548828125,
0.6943359375,
0.5634765625,
0.509765625,
0.56396484375
],
[
0.1175537109375,
0.5693359375,
0.486083984375,
0.5693359375,
0.486083984375,
0.5830078125,
0.1175537109375,
0.5830078125
],
[
0.59375,
0.572265625,
0.79833984375,
0.57666015625,
0.79833984375,
0.59423828125,
0.59375,
0.59033203125
],
[
0.11419677734375,
0.58447265625,
0.47021484375,
0.58447265625,
0.47021484375,
0.59716796875,
0.11419677734375,
0.59716796875
],
[
0.1368408203125,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.1368408203125,
0.615234375
],
[
0.51123046875,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.62060546875,
0.51123046875,
0.62060546875
],
[
0.1158447265625,
0.6162109375,
0.48779296875,
0.61767578125,
0.48779296875,
0.6318359375,
0.1158447265625,
0.63134765625
],
[
0.51318359375,
0.62255859375,
0.88330078125,
0.62255859375,
0.88330078125,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.1158447265625,
0.63427734375,
0.486083984375,
0.6328125,
0.486083984375,
0.64794921875,
0.1158447265625,
0.6484375
],
[
0.50537109375,
0.63525390625,
0.884765625,
0.6328125,
0.884765625,
0.65380859375,
0.50537109375,
0.65625
],
[
0.118408203125,
0.65087890625,
0.48779296875,
0.6484375,
0.48779296875,
0.6630859375,
0.118408203125,
0.6650390625
],
[
0.509765625,
0.65380859375,
0.8818359375,
0.65380859375,
0.8818359375,
0.66845703125,
0.509765625,
0.66845703125
],
[
0.1175537109375,
0.66650390625,
0.48779296875,
0.66650390625,
0.48779296875,
0.6796875,
0.1175537109375,
0.6796875
],
[
0.51123046875,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68505859375,
0.51123046875,
0.68505859375
],
[
0.1158447265625,
0.68115234375,
0.486083984375,
0.68115234375,
0.486083984375,
0.69482421875,
0.1158447265625,
0.69482421875
],
[
0.51123046875,
0.68701171875,
0.8818359375,
0.68701171875,
0.8818359375,
0.70068359375,
0.51123046875,
0.70068359375
],
[
0.1175537109375,
0.69677734375,
0.490234375,
0.69677734375,
0.490234375,
0.71142578125,
0.1175537109375,
0.71142578125
],
[
0.509765625,
0.703125,
0.880859375,
0.703125,
0.880859375,
0.71630859375,
0.509765625,
0.71630859375
],
[
0.11419677734375,
0.71435546875,
0.48779296875,
0.71337890625,
0.48779296875,
0.72802734375,
0.11419677734375,
0.728515625
],
[
0.51123046875,
0.72021484375,
0.87890625,
0.72021484375,
0.87890625,
0.73095703125,
0.51123046875,
0.73095703125
],
[
0.1175537109375,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.74462890625,
0.1175537109375,
0.74462890625
],
[
0.509765625,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7490234375,
0.509765625,
0.7490234375
],
[
0.11419677734375,
0.74658203125,
0.489501953125,
0.7451171875,
0.489501953125,
0.76025390625,
0.11419677734375,
0.76123046875
],
[
0.51318359375,
0.751953125,
0.68701171875,
0.751953125,
0.68701171875,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.1158447265625,
0.76318359375,
0.486083984375,
0.76318359375,
0.486083984375,
0.77685546875,
0.1158447265625,
0.77685546875
],
[
0.1158447265625,
0.779296875,
0.489501953125,
0.779296875,
0.489501953125,
0.7919921875,
0.1158447265625,
0.7919921875
],
[
0.509765625,
0.77685546875,
0.77734375,
0.77685546875,
0.77734375,
0.79150390625,
0.509765625,
0.79150390625
],
[
0.1158447265625,
0.79150390625,
0.491943359375,
0.79150390625,
0.491943359375,
0.80859375,
0.1158447265625,
0.80859375
],
[
0.5087890625,
0.79736328125,
0.8818359375,
0.79736328125,
0.8818359375,
0.81103515625,
0.5087890625,
0.81103515625
],
[
0.509765625,
0.81396484375,
0.8818359375,
0.8134765625,
0.8818359375,
0.82763671875,
0.509765625,
0.8291015625
],
[
0.139404296875,
0.82080078125,
0.48779296875,
0.82080078125,
0.48779296875,
0.83447265625,
0.139404296875,
0.83447265625
],
[
0.1175537109375,
0.83447265625,
0.490234375,
0.83447265625,
0.490234375,
0.84423828125,
0.1175537109375,
0.84423828125
],
[
0.51123046875,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.51123046875,
0.845703125
],
[
0.1158447265625,
0.845703125,
0.28369140625,
0.845703125,
0.28369140625,
0.85546875,
0.1158447265625,
0.85546875
],
[
0.51123046875,
0.8466796875,
0.8818359375,
0.8466796875,
0.8818359375,
0.85986328125,
0.51123046875,
0.85986328125
],
[
0.1351318359375,
0.85546875,
0.48779296875,
0.85791015625,
0.48779296875,
0.87158203125,
0.1351318359375,
0.869140625
],
[
0.1175537109375,
0.869140625,
0.175537109375,
0.869140625,
0.175537109375,
0.88134765625,
0.1175537109375,
0.88134765625
],
[
0.52490234375,
0.86767578125,
0.8818359375,
0.8681640625,
0.8818359375,
0.8828125,
0.52490234375,
0.8818359375
],
[
0.1368408203125,
0.88134765625,
0.489501953125,
0.88134765625,
0.489501953125,
0.89453125,
0.1368408203125,
0.89453125
],
[
0.509765625,
0.8828125,
0.87890625,
0.8828125,
0.87890625,
0.8935546875,
0.509765625,
0.8935546875
],
[
0.1158447265625,
0.8935546875,
0.489501953125,
0.8935546875,
0.489501953125,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.509765625,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90673828125,
0.509765625,
0.90673828125
],
[
0.1175537109375,
0.90673828125,
0.2091064453125,
0.90673828125,
0.2091064453125,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.5087890625,
0.90380859375,
0.77734375,
0.90478515625,
0.77734375,
0.91943359375,
0.5087890625,
0.91845703125
]
] |
[
"has also been discussed and shown effective in pre-",
"bine them with node representations to form Hpij",
"vious studies (Rao et al., 2017; Huang et al., 2018;",
"a representation for path pij. We also get an event",
"Zhang and Ji, 2021). However, previous studies",
"type representation ht, for lt, from the event-type",
"directly take AMR graphs as input to the extraction",
"embedding matrix Etri and an argument role rep-",
"model and thus require AMR graphs during both",
"resentation hai, for Qij from the argument role",
"training and inference, making their performance",
"embedding matrix Earg. Here, Erel, Etri, and",
"highly dependent on the quality of AMR parsing.",
"Earg are all randomly initialized and will be opti-",
"Different from them, our proposed STF only takes",
"mized during training. Finally, we obtain the initial",
"AMR graphs during reference to measure the cor-",
"representations Hinit - [ht, Hpi, ha,] for the",
"rectness of event predictions during self-training,",
"sequence [i, Pij, Qi].",
"making it free from the potential errors propagation",
"To estimate the compatibility between the event",
"from AMR parsing during inference.",
"trigger and argument prediction and their path",
"Given a sentence W e Xu from the unlabeled",
"in the AMR graph, we apply multi-layer self",
"corpus and a predicted trigger f; and its argument",
"attention (Vaswani et al., 2017) over the joint repre",
"E, from W, we aim to estimate a correctness score",
"sentation of the AMR path and the event prediction",
"for each pair of the trigger and argument prediction",
"Hinit to learn better contextual representations for",
"based on its compatibility with their path in the",
"the sequence [it, Pii, Qil and we add the position",
"AMR graph°. Thus, we first apply the state-of-the-",
"embedding Epos to Hinit before feed it into the",
"art AMR parsing tool (Astudillo et al., 2020) to",
"self-attention layers:",
"generate an AMR graph for W: G = (V, E), E -",
"[(vi, eij, uj)leij E R)). We follow (Huang et al.,",
"Hfinal",
" self-attention(Hinit) × M,",
"2016; Zhang and Ji, 2021) and group the original",
"set of AMR relations into 19 categories10, thus",
"where M denotes the number of attention layers.",
"eii denotes a particular relation category and R",
"Finally,we compute an overall vector represen-",
"tation H/imal from H/imnal via average-pooling and",
"denotes the set of AMR relation categories. Then,",
"we identify the Ui, Uj from AMR graph G as the",
"feed it into a linear-layer and a Sigmoid function",
"corresponding node of ti, Ej, by node alignment",
"to compute a probability Cij, indicating the correct",
"following Zhang and Ji (2021). Then, we utilize the",
" ness of the predicted event trigger and argument",
"Breadth First Search to find the shortest path pi.j",
"We optimize the scoring model based on the binary",
"that connects and includes, U, and u, in G. If there",
"cross-entropy objective:",
"is no path between v; and vi, we add a new edge",
"LScore = BCE (yij, Cij; 3b),",
"to connect them and assign other as the relation",
"Given a predicted trigger f; and its type It, a pre-",
"where yij E (O, 1) is a binary label that indi-",
"dicted argument E; and its argument role Qi, the",
"cates the argument role is correct (yij = 1) or",
"scoring model estimates their correctness by taking",
"not (yij = O)12, and b is the parameters of the",
"[lti, pij, Qi] as input and outputs a compatibility",
" scoring model. During training, we have gold trig",
"score. As Figure 1 shows, it consists of a language",
"gers and arguments as positive training instances",
"model encoder (Devlin et al., 2019; Liu et al., 2019)",
"and we swap the argument roles in positive training",
"to encode the sentence W and obtain the contex-",
"instances with randomly sampled incorrect labels",
"tual representations for the tokensll,which are then",
"to create negativetraininginstances.Aftertraining",
"used to initialize the representation of each node",
"the scoring model, we will fix its parameters and",
"in pi, based on the alignment between the input to-",
"apply it to self-training.",
"kens and the nodes in AMR graph following Zhang",
"and Ji (2021).Wedrawedgerepresentationsfrom",
"2.3 Self-Training with Feedback",
"the AMR relation embedding matrix Erel and com-",
"To improve the base event extraction model with",
"self-training, we take the new event predictions",
"'Comparing with the whole AMR graph, the path of the",
"trigger and argument in the AMRgraph shows more improve-",
"(fi, iti, Ej, Qig) from the unlabeled corpus Xu as",
"mentforthe scoring model",
"pseudo samples to further train the event extraction",
"IOThe details of AMR relation categories are shown in Ap-",
"pendix A",
"12we don't not consider the cases where the trigger labels",
"'lf a token is split into multiple subtokens, we average",
"are incorrect, sinceby observation the semantics and structure",
"the representations of all subtokens to obtain an overall token",
"of AMR graphs are more related to the argument role types",
"representation.",
"between event triggers and their arguments."
] |
{
"arxiv_id": "2205.12490",
"language": "en",
"page_idx": 3
}
|
||
2205.12490-en-4
|
2205.12490-en-4.jpg
|
model. The gradients of the event extraction model
on each pseudo sample is computed as:
$$g_{ij}^{st}=\nabla_{\theta}\mathbf{L}^{\mathrm{E}}\left(\tilde{W},(\tilde{\tau}_{i},\tilde{l}_{\tau_{i}},\tilde{a}_{ij},\tilde{\varepsilon}_{j});\theta\right)$$
where $\theta$ denotes the parameters of the event extraction model. Note that there can be multiple event predictions in one sentence.
Due to the prediction errors of the pseudo labels, simply following the gradients $g_{ij}^{st}$ computed on the pseudo labels can hurt model's performance. Thus, we utilize the correctness score $c_ij$ predicted by the scoring model to update the gradients, based on the motivation that: (1) if an event prediction is compatible with the AMR structure, it's likely to be correct and we should encourage the model learning on the pseudo label; (2) on the other side, if an event prediction is incompatible with its AMR structure, it's likely incorrect and we should discourage the model learning on the pseudo label; (3) if the scoring model is uncertain about the correctness of the event prediction, we should reduce the magnitude of the gradients learned from the pseudo label. Motivated by this, we first design a transformation function $f^c$ to project the correctness score $c_ij\in[0,1]$ into a range [-1,1] where -1 (or $c_ij=0)$ indicates incompatible, 1(or $c_ij=1)$ means compatible, and 0 (or $c_{ij}=0.5)$ means uncertain. Here, $f^c$ is based on a linear mapping:
$$f^c(c_{ij})=2\times c_{ij}-1$$
We then apply the compatibility scores as feedback to update the gradients of the event extraction model on each pseudo sample during self-training:
$$\mathbf{L}^{\mathrm{STF}}=\sum_{i,j}f^{c}(c_{ij})\cdot\mathbf{L}^{\mathrm{E}}\left(\tilde{W},(\tilde{\tau}_{i},\tilde{l}_{\tau_{i}},\tilde{a}_{ij},\tilde{\varepsilon}_{j});\theta\right)$$
To improve the efficiency of self-training, we update the event extraction model on every minibatch, and to avoid the model diverging, we combine the supervised training and self-training, so the overall loss for STF is:
$$\mathbf{L}=\mathbf{L}^\mathrm{E}+\beta\mathbf{L}^\mathrm{STF}$$
where $\beta$ is the combining ratio, L$^\mathrm{E}$ is computed on
the pseudo-labeled instances from $\mathcal{X}_u.$
### 3 $\textbf{Experimental Setups}$
For evaluation, we consider two base event extraction models: OneIE (Lin et al., 2020) and AMRIE (Zhang and Ji, 2021) due to their superior performance on event extraction and publicly available source code, and demonstrate the effectiveness of STF on three benchmark datasets: ACE05-E, $ACE05-E^{+} and ERE-EN, with the same evaluation$ metrics following previous studies (Wadden et al., 2019; Lin et al., 2020; Zhang and Ji, 2021; Wang et al., 2022)$^{13}$. To show the generalizability of STF, we explore two unlabeled corpora for self-training: (1) AMR 3.0 (Knight et al., 2021) which originally contains 55,635 sentences in the training set while each sentence is associated with a manually annotated AMR graph. (2) New York Times Annotated Corpus (NYT) contains over 1.8 million articles that were published between 1987 to 2007. We randomly sample 55,635 sentences$^{14}$ from articles published in 2004. Because NYT dataset does not have AMR annotations, we run a pre-trained AMR parser (Astudillo et al., 2020) to generate system AMR parsing
Besides taking the recent state-of-the-art event extraction studies (Wadden et al., 2019; Du and Cardie, 2020; Lin et al., 2020; Nguyen et al., 2021; Zhang and Ji, 2021; Lu et al., 2021; Wang et al., 2022; Hsu et al., 2022; Lu et al., 2022)$^{15}$ as baselines, we also compare our proposed STF with two other training strategies: (1) vanilla SelfTraining (Rosenberg et al., 2005) which consists of two stages similar as STF but in the second stage takes each new event prediction from the unlabeled data with a probability higher than 0.9 based on the base event extraction model as a pseudo label and combines them with the labeled data to re-train the event extraction model; and (2) Gradient Imitation Reinforcement Learning (GradLRE) (Hu et al., 2021b). GradLRE encourages the gradients computed on the pseudo-labeled data to imitate the gradients computed on the labeled data by using the cosine distance between the two sources of gradients as a reward to perform policy gradi- $\overline{{\mathrm{The~detailed~statistics~of~ACE05-E,~ACE05-E^+,~and~ERE-}}}$
EN are shown in Appendix B.
$^{14}$To show the effect of unlabeled dataset vs labeled dataset, we sample the same number of the unlabeled sentences as AMR 3. 0
$^{15}$The scores reported in (Nguyen et al., 2022) are not comparable in the table 3, as their results are not averaged across
the labeled dataset $\mathcal{X}_L$ and $\mathcal{L}^\mathrm{STF}$ is computed on random seeds. We tried to report their averaged performance the labeled dataset $\mathcal{X}_L$ and $\mathcal{L}^\mathrm{STF}$ is computed on by running their model ourselves by contacting the authors
by running their model ourselves by contacting the authors,
however, their code is publicly unavailable.
|
[
[
0.1175537109375,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.50390625,
0.08489990234375,
0.724609375,
0.08489990234375,
0.724609375,
0.09979248046875,
0.50390625,
0.09979248046875
],
[
0.1158447265625,
0.1009521484375,
0.41064453125,
0.1015625,
0.41064453125,
0.11639404296875,
0.1158447265625,
0.11517333984375
],
[
0.51318359375,
0.11102294921875,
0.8818359375,
0.11102294921875,
0.8818359375,
0.12469482421875,
0.51318359375,
0.12469482421875
],
[
0.51318359375,
0.1270751953125,
0.88330078125,
0.1270751953125,
0.88330078125,
0.141357421875,
0.51318359375,
0.141357421875
],
[
0.160400390625,
0.1341552734375,
0.4306640625,
0.1318359375,
0.4306640625,
0.15673828125,
0.160400390625,
0.1591796875
],
[
0.51318359375,
0.1436767578125,
0.8818359375,
0.1436767578125,
0.8818359375,
0.15673828125,
0.51318359375,
0.15673828125
],
[
0.51318359375,
0.1602783203125,
0.8818359375,
0.1602783203125,
0.8818359375,
0.17333984375,
0.51318359375,
0.17333984375
],
[
0.1175537109375,
0.177001953125,
0.49365234375,
0.177001953125,
0.49365234375,
0.190673828125,
0.1175537109375,
0.190673828125
],
[
0.51123046875,
0.17578125,
0.8818359375,
0.17578125,
0.8818359375,
0.189453125,
0.51123046875,
0.189453125
],
[
0.1175537109375,
0.1923828125,
0.489501953125,
0.1923828125,
0.489501953125,
0.2060546875,
0.1175537109375,
0.2060546875
],
[
0.51123046875,
0.190673828125,
0.880859375,
0.190673828125,
0.880859375,
0.2037353515625,
0.51123046875,
0.2037353515625
],
[
0.11419677734375,
0.2078857421875,
0.32421875,
0.208984375,
0.32421875,
0.223876953125,
0.11419677734375,
0.22265625
],
[
0.51318359375,
0.2078857421875,
0.8818359375,
0.2078857421875,
0.8818359375,
0.221435546875,
0.51318359375,
0.221435546875
],
[
0.1368408203125,
0.226806640625,
0.48779296875,
0.226806640625,
0.48779296875,
0.240478515625,
0.1368408203125,
0.240478515625
],
[
0.51123046875,
0.2244873046875,
0.8818359375,
0.2244873046875,
0.8818359375,
0.2381591796875,
0.51123046875,
0.2381591796875
],
[
0.1175537109375,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.1175537109375,
0.257080078125
],
[
0.509765625,
0.2392578125,
0.8818359375,
0.240478515625,
0.8818359375,
0.254638671875,
0.509765625,
0.254150390625
],
[
0.1175537109375,
0.2587890625,
0.489501953125,
0.2587890625,
0.489501953125,
0.2724609375,
0.1175537109375,
0.2724609375
],
[
0.51123046875,
0.255859375,
0.8818359375,
0.255859375,
0.8818359375,
0.270751953125,
0.51123046875,
0.270751953125
],
[
0.1175537109375,
0.27490234375,
0.48779296875,
0.27490234375,
0.48779296875,
0.289306640625,
0.1175537109375,
0.289306640625
],
[
0.509765625,
0.271484375,
0.8818359375,
0.2724609375,
0.8818359375,
0.287353515625,
0.509765625,
0.2861328125
],
[
0.11419677734375,
0.290283203125,
0.486083984375,
0.290283203125,
0.486083984375,
0.303955078125,
0.11419677734375,
0.303955078125
],
[
0.5087890625,
0.287353515625,
0.88330078125,
0.2880859375,
0.88330078125,
0.302734375,
0.5087890625,
0.3017578125
],
[
0.1158447265625,
0.30712890625,
0.489501953125,
0.30712890625,
0.489501953125,
0.320556640625,
0.1158447265625,
0.320556640625
],
[
0.509765625,
0.3046875,
0.88330078125,
0.3046875,
0.88330078125,
0.318359375,
0.509765625,
0.318359375
],
[
0.1158447265625,
0.322509765625,
0.48779296875,
0.322509765625,
0.48779296875,
0.336181640625,
0.1158447265625,
0.336181640625
],
[
0.5087890625,
0.319580078125,
0.8818359375,
0.320556640625,
0.8818359375,
0.3349609375,
0.5087890625,
0.334228515625
],
[
0.1158447265625,
0.339111328125,
0.48779296875,
0.339111328125,
0.48779296875,
0.352783203125,
0.1158447265625,
0.352783203125
],
[
0.509765625,
0.336181640625,
0.8818359375,
0.336181640625,
0.8818359375,
0.349853515625,
0.509765625,
0.349853515625
],
[
0.1175537109375,
0.355712890625,
0.489501953125,
0.355712890625,
0.489501953125,
0.369384765625,
0.1175537109375,
0.369384765625
],
[
0.509765625,
0.352783203125,
0.8818359375,
0.352783203125,
0.8818359375,
0.366455078125,
0.509765625,
0.366455078125
],
[
0.11419677734375,
0.37060546875,
0.486083984375,
0.37060546875,
0.486083984375,
0.38427734375,
0.11419677734375,
0.38427734375
],
[
0.509765625,
0.3681640625,
0.8818359375,
0.3681640625,
0.8818359375,
0.3818359375,
0.509765625,
0.3818359375
],
[
0.1158447265625,
0.38720703125,
0.490234375,
0.38720703125,
0.490234375,
0.40087890625,
0.1158447265625,
0.40087890625
],
[
0.51123046875,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3984375,
0.51123046875,
0.3984375
],
[
0.1175537109375,
0.40380859375,
0.489501953125,
0.40380859375,
0.489501953125,
0.41748046875,
0.1175537109375,
0.41748046875
],
[
0.51318359375,
0.402587890625,
0.87744140625,
0.402587890625,
0.87744140625,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.1175537109375,
0.42041015625,
0.490234375,
0.42041015625,
0.490234375,
0.43408203125,
0.1175537109375,
0.43408203125
],
[
0.51123046875,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.43115234375,
0.51123046875,
0.43115234375
],
[
0.1175537109375,
0.435791015625,
0.48779296875,
0.435791015625,
0.48779296875,
0.449462890625,
0.1175537109375,
0.449462890625
],
[
0.51123046875,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.446533203125,
0.51123046875,
0.446533203125
],
[
0.1158447265625,
0.451904296875,
0.486083984375,
0.451904296875,
0.486083984375,
0.465576171875,
0.1158447265625,
0.465576171875
],
[
0.509765625,
0.448974609375,
0.67529296875,
0.449462890625,
0.67529296875,
0.46435546875,
0.509765625,
0.463134765625
],
[
0.1175537109375,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482177734375,
0.1175537109375,
0.482177734375
],
[
0.53076171875,
0.466064453125,
0.8818359375,
0.466064453125,
0.8818359375,
0.479736328125,
0.53076171875,
0.479736328125
],
[
0.1175537109375,
0.48388671875,
0.489501953125,
0.48388671875,
0.489501953125,
0.49755859375,
0.1175537109375,
0.49755859375
],
[
0.51318359375,
0.482177734375,
0.8818359375,
0.482177734375,
0.8818359375,
0.495361328125,
0.51318359375,
0.495361328125
],
[
0.1175537109375,
0.50048828125,
0.486083984375,
0.50048828125,
0.486083984375,
0.51416015625,
0.1175537109375,
0.51416015625
],
[
0.51123046875,
0.496337890625,
0.8818359375,
0.49755859375,
0.8818359375,
0.5126953125,
0.51123046875,
0.51123046875
],
[
0.1175537109375,
0.51708984375,
0.48779296875,
0.51708984375,
0.48779296875,
0.53076171875,
0.1175537109375,
0.53076171875
],
[
0.51318359375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.1158447265625,
0.533203125,
0.48779296875,
0.533203125,
0.48779296875,
0.54638671875,
0.1158447265625,
0.54638671875
],
[
0.51123046875,
0.52978515625,
0.880859375,
0.52783203125,
0.880859375,
0.5419921875,
0.51123046875,
0.54443359375
],
[
0.11419677734375,
0.54736328125,
0.476806640625,
0.548828125,
0.476806640625,
0.5634765625,
0.11419677734375,
0.5625
],
[
0.51123046875,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.51123046875,
0.56005859375
],
[
0.51123046875,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.57470703125,
0.51123046875,
0.57470703125
],
[
0.2216796875,
0.5830078125,
0.382080078125,
0.5830078125,
0.382080078125,
0.59716796875,
0.2216796875,
0.59716796875
],
[
0.509765625,
0.57763671875,
0.8818359375,
0.57666015625,
0.8818359375,
0.59130859375,
0.509765625,
0.5927734375
],
[
0.509765625,
0.59326171875,
0.880859375,
0.59326171875,
0.880859375,
0.60693359375,
0.509765625,
0.60693359375
],
[
0.51318359375,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.62451171875,
0.51318359375,
0.62451171875
],
[
0.1368408203125,
0.619140625,
0.48779296875,
0.619140625,
0.48779296875,
0.63427734375,
0.1368408203125,
0.63427734375
],
[
0.51318359375,
0.626953125,
0.8818359375,
0.626953125,
0.8818359375,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.1158447265625,
0.63525390625,
0.484375,
0.63525390625,
0.484375,
0.6484375,
0.1158447265625,
0.6484375
],
[
0.51318359375,
0.6435546875,
0.8818359375,
0.6435546875,
0.8818359375,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.1175537109375,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.1175537109375,
0.66650390625
],
[
0.51123046875,
0.65771484375,
0.880859375,
0.65771484375,
0.880859375,
0.67138671875,
0.51123046875,
0.67138671875
],
[
0.5146484375,
0.6767578125,
0.880859375,
0.6767578125,
0.880859375,
0.68701171875,
0.5146484375,
0.68701171875
],
[
0.12005615234375,
0.68310546875,
0.47265625,
0.68505859375,
0.47265625,
0.70703125,
0.12005615234375,
0.705078125
],
[
0.51318359375,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.51318359375,
0.705078125
],
[
0.187255859375,
0.708984375,
0.20654296875,
0.708984375,
0.20654296875,
0.71875,
0.187255859375,
0.71875
],
[
0.51123046875,
0.7060546875,
0.8818359375,
0.70703125,
0.8818359375,
0.7216796875,
0.51123046875,
0.72021484375
],
[
0.51123046875,
0.7236328125,
0.8818359375,
0.7236328125,
0.8818359375,
0.73681640625,
0.51123046875,
0.73681640625
],
[
0.1351318359375,
0.73681640625,
0.490234375,
0.73828125,
0.490234375,
0.7529296875,
0.1351318359375,
0.751953125
],
[
0.51123046875,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.1158447265625,
0.75439453125,
0.491943359375,
0.75439453125,
0.491943359375,
0.76708984375,
0.1158447265625,
0.76708984375
],
[
0.5087890625,
0.7548828125,
0.880859375,
0.7548828125,
0.880859375,
0.7685546875,
0.5087890625,
0.7685546875
],
[
0.1158447265625,
0.771484375,
0.489501953125,
0.771484375,
0.489501953125,
0.78515625,
0.1158447265625,
0.78515625
],
[
0.51123046875,
0.7724609375,
0.8818359375,
0.7724609375,
0.8818359375,
0.7861328125,
0.51123046875,
0.7861328125
],
[
0.1175537109375,
0.78759765625,
0.48779296875,
0.78759765625,
0.48779296875,
0.80029296875,
0.1175537109375,
0.80029296875
],
[
0.5263671875,
0.79443359375,
0.8818359375,
0.7958984375,
0.8818359375,
0.80859375,
0.5263671875,
0.80810546875
],
[
0.1175537109375,
0.80419921875,
0.231689453125,
0.80419921875,
0.231689453125,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.51318359375,
0.80810546875,
0.69287109375,
0.80859375,
0.69287109375,
0.82080078125,
0.51318359375,
0.8193359375
],
[
0.52880859375,
0.82177734375,
0.8818359375,
0.82177734375,
0.8818359375,
0.8349609375,
0.52880859375,
0.8349609375
],
[
0.231689453125,
0.8330078125,
0.36865234375,
0.83056640625,
0.36865234375,
0.849609375,
0.231689453125,
0.8515625
],
[
0.51123046875,
0.8349609375,
0.87890625,
0.8349609375,
0.87890625,
0.845703125,
0.51123046875,
0.845703125
],
[
0.51318359375,
0.8466796875,
0.57275390625,
0.8466796875,
0.57275390625,
0.85693359375,
0.51318359375,
0.85693359375
],
[
0.52880859375,
0.859375,
0.880859375,
0.859375,
0.880859375,
0.869140625,
0.52880859375,
0.869140625
],
[
0.11419677734375,
0.87060546875,
0.490234375,
0.8681640625,
0.490234375,
0.88525390625,
0.11419677734375,
0.88720703125
],
[
0.51318359375,
0.87158203125,
0.880859375,
0.87158203125,
0.880859375,
0.8818359375,
0.51318359375,
0.8818359375
],
[
0.11248779296875,
0.88525390625,
0.489501953125,
0.8857421875,
0.489501953125,
0.90380859375,
0.11248779296875,
0.90283203125
],
[
0.51123046875,
0.8818359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8955078125,
0.51123046875,
0.8955078125
],
[
0.51123046875,
0.8935546875,
0.88330078125,
0.89453125,
0.88330078125,
0.90625,
0.51123046875,
0.90478515625
],
[
0.1158447265625,
0.90380859375,
0.399658203125,
0.90380859375,
0.399658203125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.77685546875,
0.90478515625,
0.77685546875,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"model. The gradients of the event extraction model",
"3Experimental Setups",
"on each pseudo sample is computed as:",
"For evaluation, we consider two base event extrac-",
"tion models: OneIE (Lin et al., 2020) and AMR-",
"g =VoLE (W,(,It,a,e);0",
"IE (Zhang and Ji, 2021) due to their superior per-",
"formance on event extraction and publicly avail-",
"where denotes the parameters of the event extrac-",
"able source code, and demonstrate the effectiveness",
"tion model. Note that there can be multiple event",
"of STF on three benchmark datasets: ACEO05-E",
"predictions in one sentence.",
"ACEO5-E+ and ERE-EN, with the same evaluation",
"Due to the prediction errors of the pseudo labels,",
"metrics following previous studies (Wadden et al..",
"simply following the gradients gst computed on",
"2019; Lin et al., 2020; Zhang and Ji, 2021; Wang",
"the pseudo labels can hurt model's performance.",
"et al., 2022)13. To show the generalizability of STF,",
"Thus, we utilize the correctness score Cij predicted",
"we explore two unlabeled corpora for self-training:",
"by the scoring model to update the gradients, based",
"(1I) AMR 3.0 (Knight et al., 2021) which origi-",
"on the motivation that: (1) if an event prediction",
"nally contains 55,635 sentences in the training set",
"is compatible with the AMR structure, it's likely",
"while each sentence is associated with a manually",
"to be correct and we should encourage the model",
"annotated AMR graph. (2) New York Times An",
"learning on the pseudo label; (2) on the other side,",
"notated Corpus (NYT) contains over 1.8 million",
"if an event prediction is incompatible with its AMR",
"articles that were published between 1987 to 2007.",
" structure, it's likely incorrect and we should dis-",
"Werandomly sample55,635sentences14fromarti-",
"courage the model learning on the pseudo label;",
"cles published in 2004. Because NYT dataset does",
"(3) if the scoring model is uncertain about the cor-",
"not have AMR annotations, we run a pre-trained",
"rectness of the event prediction, we should reduce",
"AMR parser (Astudillo et al., 2020) to generate",
"the magnitude of the gradients learned from the",
"system AMR parsing.",
"pseudo label. Motivated by this, we first design a",
"Besides taking the recent state-of-the-art event",
"transformation function fc to project the correct-",
"extraction studies (Wadden et al., 2019; Du and",
"ness score ci E [0, 1] into a range [-1, 1] where -1",
"Cardie, 2020; Lin et al., 2020; Nguyen et al..",
"(or cij = O) indicates incompatible, 1 (or cij = 1)",
"2021; Zhang and Ji, 2021; Lu et al., 2021; Wang",
"means compatible, and 0 (or cii = 0.5) means",
"et al., 2022; Hsu et al., 2022; Lu et al., 2022)15",
"uncertain. Here, fc is based on a linear mapping",
"as baselines, we also compare our proposed STF",
"with two other training strategies: (1) vanilla Self",
"fc(ci) = 2 × cii - 1",
"Training (Rosenberg et al., 2005) which consists of",
"two stages similar as STF but in the second stage",
"takes each new event prediction from the unlabeled",
"We then apply the compatibility scores as feed-",
"data with a probability higher than 0.9 based on",
"back to update the gradients of the event extraction",
"the base event extraction model as a pseudo label",
"model on each pseudo sample during self-training:",
"and combines them with the labeled data to re-train",
"the eventextractionmodel:and(2)GradientIm-",
"LSTF -fc(ci)-LE (W,(i,it,aij,e);0",
"itation Reinforcement Learning (GradLRE) (Hu",
"i,j",
"et al., 2021b). GradLRE encourages the gradients",
"computed on the pseudo-labeled data to imitate",
"To improve the efficiency of self-training, we up-",
"the gradients computed on the labeled data by us-",
"date the event extraction model on every minibatch,",
"ing the cosine distance between the two sources",
"and to avoid the model diverging, we combine the",
"of gradients as a reward to perform policy gradi-",
"supervised training and self-training, so the overall",
"13The detailed statistics of ACE05-E,ACE05-E,and ERE",
"loss for STF is:",
"ENare showninAppendixB",
"14To show the effect of unlabeled dataset vs labeled dataset,",
"L = LE + βLSTF",
"we sample the same number of the unlabeled sentences as",
"AMR 3.0",
"15The scores reported in (Nguyen et al., 2022)are not com",
"where β is the combining ratio, LE is computed on",
"parable in the table 3, as their results are not averaged across",
"the labeled dataset XL and LSTF is computed on",
"random seeds. We tried to report their averaged performance",
"by running theirmodel ourselves by contacting the authors,",
"the pseudo-labeled instances from Xu.",
"however, their code is publicly unavailable."
] |
{
"arxiv_id": "2205.12490",
"language": "en",
"page_idx": 4
}
|
||
2306.04423-en-0
|
2306.04423-en-0.jpg
|
On Computing Optimal Tree Ensembles*
Christian Komusiewicz$^1$ Pascal Kunz$^2,3\uparrow$
Frank Sommer$^{1\:\ddagger }$ Manuel Sorge$^{4\text{ §}}$ $^{1}\text{Fakult\"{a} t f\"{u} r Mathematik und Informatik, }$ Friedrich-Schiller-Universität Jena, Germany
$^{2}$Algorithmics and Computational Complexity, TU Berlin,
Germany
$^{3}\mathop{\mathrm{Algorithm~Engineering,~HU~Berlin,~Germany}}$
$^{4}$Institute of Logic and Computation, TU Wien, Austria
### $\mathbf{Abstract}$
Random forests and, more generally, (decision-)tree ensembles are widely used methods for classification and regression. Recent algorithmic advances allow to compute decision trees that are optimal for various measures such as their size or depth. We are not aware of such research for tree ensembles and aim to contribute to this area. Mainly, we provide two novel algorithms and corresponding lower bounds. First, we are able to carry over and substantially improve on tractability results for decision trees, obtaining a $(6\delta DS)^{S}\cdot$poly-time algorithm, where $S$ is the number of cuts in the tree ensemble, $D$ the largest domain size, and $\delta$ is the largest number of features in which two examples differ. To achieve this, we introduce the witness-tree technique which also seems promising for practice. Second, we show that dynamic programming, which has been successful for decision trees, may also be viable for tree ensembles, providing an $\ell^n\cdot$poly-time algorithm, where $\ell$ is the number of trees and $n$ the number of examples. Finally, we compare the number of cuts necessary to classify training data sets for decision trees and tree ensembles, showing that ensembles may need exponentially fewer cuts for increasing number of trees.
$^{*}$This work was initiated at the research retreat of the Algorithmics and Computational
Complexity group of TU Berlin held in Darlingerode in September 2022.
$^{\dagger}$Supported by the DFG Research Training Group 2434“Facets of Complexity”.
$\overset{\dagger}{\operatorname*{\text{Supported by the DFG, project EAGR (KO 3669/6-1)}}}$ $^{8}$Supported by the Alexander von Humboldt Foundation.
1
|
[
[
0.271240234375,
0.205810546875,
0.734375,
0.205810546875,
0.734375,
0.2266845703125,
0.271240234375,
0.2266845703125
],
[
0.3095703125,
0.2474365234375,
0.5087890625,
0.2474365234375,
0.5087890625,
0.261962890625,
0.3095703125,
0.261962890625
],
[
0.5556640625,
0.24560546875,
0.69775390625,
0.242431640625,
0.69775390625,
0.261474609375,
0.5556640625,
0.264404296875
],
[
0.3349609375,
0.270751953125,
0.484375,
0.268310546875,
0.484375,
0.287353515625,
0.3349609375,
0.2890625
],
[
0.0228729248046875,
0.25,
0.060455322265625,
0.25,
0.060455322265625,
0.37744140625,
0.0228729248046875,
0.37744140625
],
[
0.52685546875,
0.270751953125,
0.66845703125,
0.268310546875,
0.66845703125,
0.287353515625,
0.52685546875,
0.2890625
],
[
0.316162109375,
0.300537109375,
0.6787109375,
0.302978515625,
0.6787109375,
0.317626953125,
0.316162109375,
0.31494140625
],
[
0.310546875,
0.3193359375,
0.685546875,
0.3212890625,
0.685546875,
0.3359375,
0.310546875,
0.333984375
],
[
0.25,
0.333984375,
0.74609375,
0.3359375,
0.74609375,
0.354736328125,
0.25,
0.352783203125
],
[
0.4599609375,
0.354248046875,
0.5390625,
0.35595703125,
0.5390625,
0.371826171875,
0.4599609375,
0.369384765625
],
[
0.0228729248046875,
0.36865234375,
0.062103271484375,
0.36865234375,
0.062103271484375,
0.460205078125,
0.0228729248046875,
0.460205078125
],
[
0.29736328125,
0.367431640625,
0.69677734375,
0.37060546875,
0.69677734375,
0.392578125,
0.29736328125,
0.389404296875
],
[
0.2548828125,
0.3837890625,
0.73876953125,
0.386474609375,
0.73876953125,
0.4072265625,
0.2548828125,
0.40478515625
],
[
0.465576171875,
0.447509765625,
0.53515625,
0.447509765625,
0.53515625,
0.462158203125,
0.465576171875,
0.462158203125
],
[
0.0196075439453125,
0.464111328125,
0.058013916015625,
0.464111328125,
0.058013916015625,
0.68310546875,
0.0196075439453125,
0.68310546875
],
[
0.27783203125,
0.466552734375,
0.74169921875,
0.46728515625,
0.74169921875,
0.481689453125,
0.27783203125,
0.48095703125
],
[
0.25732421875,
0.481689453125,
0.6337890625,
0.481689453125,
0.6337890625,
0.49560546875,
0.25732421875,
0.49560546875
],
[
0.6259765625,
0.481689453125,
0.74169921875,
0.481689453125,
0.74169921875,
0.49560546875,
0.6259765625,
0.49560546875
],
[
0.256591796875,
0.496826171875,
0.74169921875,
0.496826171875,
0.74169921875,
0.51025390625,
0.256591796875,
0.51025390625
],
[
0.256591796875,
0.51025390625,
0.740234375,
0.51025390625,
0.740234375,
0.5234375,
0.256591796875,
0.5234375
],
[
0.256591796875,
0.5234375,
0.740234375,
0.5234375,
0.740234375,
0.537109375,
0.256591796875,
0.537109375
],
[
0.256591796875,
0.537109375,
0.74169921875,
0.537109375,
0.74169921875,
0.55029296875,
0.256591796875,
0.55029296875
],
[
0.256591796875,
0.55029296875,
0.740234375,
0.55029296875,
0.740234375,
0.564453125,
0.256591796875,
0.564453125
],
[
0.256591796875,
0.564453125,
0.74169921875,
0.56298828125,
0.74169921875,
0.57763671875,
0.256591796875,
0.5791015625
],
[
0.25732421875,
0.57763671875,
0.740234375,
0.57763671875,
0.740234375,
0.59130859375,
0.25732421875,
0.59130859375
],
[
0.256591796875,
0.59228515625,
0.74267578125,
0.59228515625,
0.74267578125,
0.60595703125,
0.256591796875,
0.60595703125
],
[
0.256591796875,
0.60595703125,
0.74169921875,
0.60595703125,
0.74169921875,
0.619140625,
0.256591796875,
0.619140625
],
[
0.256591796875,
0.62255859375,
0.318603515625,
0.62060546875,
0.318603515625,
0.6318359375,
0.256591796875,
0.6337890625
],
[
0.314453125,
0.62060546875,
0.74169921875,
0.62060546875,
0.74169921875,
0.6337890625,
0.314453125,
0.6337890625
],
[
0.256591796875,
0.6337890625,
0.74267578125,
0.6337890625,
0.74267578125,
0.6484375,
0.256591796875,
0.6484375
],
[
0.2548828125,
0.64794921875,
0.74169921875,
0.646484375,
0.74169921875,
0.6611328125,
0.2548828125,
0.662109375
],
[
0.256591796875,
0.662109375,
0.74169921875,
0.662109375,
0.74169921875,
0.67529296875,
0.256591796875,
0.67529296875
],
[
0.2548828125,
0.673828125,
0.740234375,
0.67529296875,
0.740234375,
0.68994140625,
0.2548828125,
0.68798828125
],
[
0.2548828125,
0.68798828125,
0.740234375,
0.689453125,
0.740234375,
0.7041015625,
0.2548828125,
0.70263671875
],
[
0.2548828125,
0.70263671875,
0.31201171875,
0.705078125,
0.31201171875,
0.71630859375,
0.2548828125,
0.7138671875
],
[
0.233642578125,
0.7890625,
0.78125,
0.7890625,
0.78125,
0.80322265625,
0.233642578125,
0.80322265625
],
[
0.2156982421875,
0.80322265625,
0.65869140625,
0.80322265625,
0.65869140625,
0.8173828125,
0.2156982421875,
0.8173828125
],
[
0.236083984375,
0.81494140625,
0.7236328125,
0.81494140625,
0.7236328125,
0.8291015625,
0.236083984375,
0.8291015625
],
[
0.2344970703125,
0.82763671875,
0.576171875,
0.82763671875,
0.576171875,
0.84130859375,
0.2344970703125,
0.84130859375
],
[
0.2344970703125,
0.8388671875,
0.5810546875,
0.8388671875,
0.5810546875,
0.85302734375,
0.2344970703125,
0.85302734375
],
[
0.490234375,
0.87548828125,
0.50634765625,
0.87548828125,
0.50634765625,
0.8916015625,
0.490234375,
0.8916015625
]
] |
[
"On Computing Optimal Tree Ensembles*",
"ChristianKomusiewiczl",
"Pascal Kunz2.3 t",
"Frank Sommerl t",
"",
" Manuel Sorge4 s",
"1FakultatfurMathematik undInformatik",
"Friedrich-Schiller-Universitat Jena, Germany",
"2 Algorithmics and Computational Complexity, TU Berlin,",
"Germany",
"['S]",
"3 Algorithm Engineering, HU Berlin, Germany",
"4 Institute of Logic and Computation, TU Wien, Austria",
"Abstract",
"",
"Random forests and, more generally, (decision-)tree ensembles are",
"widely used methods for classification and regression.",
"Recent algorith-",
"mic advances allow to compute decision trees that are optimal for various",
"measures such as their size or depth. We are not aware of such research",
"for tree ensembles and aim to contribute to this area. Mainly, we provide",
"two novel algorithms and corresponding lower bounds. First, we are able",
"to carry over and substantially improve on tractability results for decision",
"trees, obtaining a (68DS)spoly-time algorithm, where S is the num-",
"ber of cuts in the tree ensemble, D the largest domain size, and is the",
"largest number of features in which two examples differ. To achieve this,",
"we introduce the witness-tree technique which also seems promising for",
"practice.",
"Second, we show that dynamic programming, which has been",
"successful for decision trees, may also be viable for tree ensembles, provid-",
"ing an en poly-time algorithm, where l is the number of trees and n the",
"number of examples. Finally, we compare the number of cuts necessary to",
"classify training data sets for decision trees and tree ensembles, showing",
"that ensembles may need exponentially fewer cuts for increasing number",
"oftrees.",
"*This work was initiated at the research retreat of the Algorithmics and Computational",
"Complexity group of TU Berlin held in Darlingerode in September 2022.",
"tSupported by the DFG Research Training Group 2434 “Facets of Complexity\".",
"+Supported by the DFG, project EAGR (KO 3669/6-1)",
"$Supported by the Alexander von Humboldt Foundation",
"1"
] |
{
"arxiv_id": "2306.04423",
"language": "en",
"page_idx": 0
}
|
||
2306.04423-en-1
|
2306.04423-en-1.jpg
|
## 1 Introduction
Random forests is a method for classification or regression in which we construct an ensemble of decision trees for (random subsets of) the training data and, in the classification phase, aggregate their outcomes by majority voting. The random-forests method has received a tremendous amount of attention for its simplicity and improved accuracy over plain decision trees [3,26,19,24]. Commonly, fast heuristics without performance guarantees are used for computing random forests [19,24], in particular for computing the individual decision trees in the forest. For plain decision trees, researchers lately made several advances in computing optimal decision trees, that is, decision trees that provably optimize criteria such as minimizing the tree size [1,21,4,7]. With that increased amount of attention also came theoretical advances, showing the limits and opportunities for developing efficient exact algorithms for computing decision trees [23,18,9]. One impetus to computing optimal decision trees is that minimizing the size reduces tendencies to overfitting [1,7]. It is conceivable that such benefits transfer to globally optimizing the tree ensembles computed by random forests. However, apart from sporadic hardness results [25], we are not aware of exact algorithmic research for tree ensembles. In this work, we aim to initiate this direction; that is, we begin to build the theoretical footing for exact algorithmics of computing optimal tree ensembles and provide potential avenues for exact algorithms that are guaranteed to provide optimal results with acceptable worst-case running times.
We study the algorithmic properties of two canonical formulations of the training problem for tree ensembles: We are given a set of training examples labeled with two classes and a number $\ell$ of trees and we want to compute a tree ensemble containing $\ell$ trees that classifies the examples consistently with the given class labels.$^1$ We want to minimize either the sum of the tree sizes, resulting in the problem MINIMUM TREE ENSEMBLE SIZE (MTES), or the largest size of a tree in the ensemble, resulting in the problem MINIMAX TREE ENSEMBLE SIze (MMAXTES$).^2$ Both contain as a special case the problem of computing a minimum-size decision tree, which is known to be NP-hard [15,23]. However, the hardness constructions do not necessarily reflect practical data. Thus, we are interested in precisely which properties make the problems hard or tractable.
Mainly, we provide two novel algorithms for MTES and MMAXTES$^3$ and matching lower-bound results for their running times. We call the frst one $witness- tree\textit{ algorithm. This algorithm demonstrates that prospects for tractable}$ algorithms for optimizing decision trees can be non-trivially generalized to optimizing tree ensembles. Namely, it was known that for small tree size $s$,moderate maximum domain size $D$ of any feature, and moderate number $\delta$ of features in
$^{1}$To keep the presentation focused, we consider mainly the case without training error. See
the conclusion for extensions to minimizing training error.
$^{2}$It is also natural to consider the depths of the trees instead of their sizes, but results are usually transferable between these two optimization goals and the size makes the presentation $\operatorname*{more}_{n}$accessible.
$^3$The algorithms work on the decision version of these problems, but they easily apply to
the optimization versions as well.
2
|
[
[
0.2254638671875,
0.154052734375,
0.23779296875,
0.1634521484375,
0.2254638671875,
0.17236328125,
0.2132568359375,
0.1634521484375
],
[
0.250732421875,
0.1546630859375,
0.40283203125,
0.1558837890625,
0.40283203125,
0.173583984375,
0.250732421875,
0.17236328125
],
[
0.2156982421875,
0.1875,
0.78173828125,
0.1875,
0.78173828125,
0.2008056640625,
0.2156982421875,
0.2008056640625
],
[
0.21728515625,
0.203857421875,
0.779296875,
0.203857421875,
0.779296875,
0.2139892578125,
0.21728515625,
0.2139892578125
],
[
0.2156982421875,
0.2171630859375,
0.78173828125,
0.2171630859375,
0.78173828125,
0.231689453125,
0.2156982421875,
0.231689453125
],
[
0.21728515625,
0.23291015625,
0.78125,
0.23291015625,
0.78125,
0.246826171875,
0.21728515625,
0.246826171875
],
[
0.21728515625,
0.2474365234375,
0.78173828125,
0.2474365234375,
0.78173828125,
0.261474609375,
0.21728515625,
0.261474609375
],
[
0.21728515625,
0.26318359375,
0.78369140625,
0.26318359375,
0.78369140625,
0.276611328125,
0.21728515625,
0.276611328125
],
[
0.2156982421875,
0.27783203125,
0.78173828125,
0.27783203125,
0.78173828125,
0.291748046875,
0.2156982421875,
0.291748046875
],
[
0.2156982421875,
0.292236328125,
0.78369140625,
0.292236328125,
0.78369140625,
0.30615234375,
0.2156982421875,
0.30615234375
],
[
0.2156982421875,
0.30810546875,
0.78173828125,
0.30810546875,
0.78173828125,
0.3212890625,
0.2156982421875,
0.3212890625
],
[
0.214111328125,
0.322509765625,
0.78369140625,
0.3212890625,
0.78369140625,
0.3359375,
0.214111328125,
0.337158203125
],
[
0.21728515625,
0.337158203125,
0.78173828125,
0.337158203125,
0.78173828125,
0.35107421875,
0.21728515625,
0.35107421875
],
[
0.21728515625,
0.352783203125,
0.78369140625,
0.352783203125,
0.78369140625,
0.3662109375,
0.21728515625,
0.3662109375
],
[
0.2156982421875,
0.36865234375,
0.78125,
0.36865234375,
0.78125,
0.3818359375,
0.2156982421875,
0.3818359375
],
[
0.2156982421875,
0.3837890625,
0.78173828125,
0.3837890625,
0.78173828125,
0.397705078125,
0.2156982421875,
0.397705078125
],
[
0.21728515625,
0.3984375,
0.78173828125,
0.3984375,
0.78173828125,
0.412353515625,
0.21728515625,
0.412353515625
],
[
0.2156982421875,
0.41357421875,
0.78173828125,
0.41357421875,
0.78173828125,
0.427978515625,
0.2156982421875,
0.427978515625
],
[
0.2156982421875,
0.4287109375,
0.78125,
0.4287109375,
0.78125,
0.442626953125,
0.2156982421875,
0.442626953125
],
[
0.2156982421875,
0.4443359375,
0.78125,
0.4443359375,
0.78125,
0.458251953125,
0.2156982421875,
0.458251953125
],
[
0.2156982421875,
0.458984375,
0.78173828125,
0.458984375,
0.78173828125,
0.472900390625,
0.2156982421875,
0.472900390625
],
[
0.2156982421875,
0.474853515625,
0.78125,
0.474853515625,
0.78125,
0.488037109375,
0.2156982421875,
0.488037109375
],
[
0.2156982421875,
0.4892578125,
0.48193359375,
0.4892578125,
0.48193359375,
0.50244140625,
0.2156982421875,
0.50244140625
],
[
0.24267578125,
0.50390625,
0.78173828125,
0.50390625,
0.78173828125,
0.517578125,
0.24267578125,
0.517578125
],
[
0.2156982421875,
0.51953125,
0.78173828125,
0.51953125,
0.78173828125,
0.53271484375,
0.2156982421875,
0.53271484375
],
[
0.2156982421875,
0.5341796875,
0.78173828125,
0.5341796875,
0.78173828125,
0.54736328125,
0.2156982421875,
0.54736328125
],
[
0.2156982421875,
0.5498046875,
0.78173828125,
0.5498046875,
0.78173828125,
0.56298828125,
0.2156982421875,
0.56298828125
],
[
0.21728515625,
0.564453125,
0.78369140625,
0.564453125,
0.78369140625,
0.5791015625,
0.21728515625,
0.5791015625
],
[
0.2156982421875,
0.580078125,
0.78173828125,
0.580078125,
0.78173828125,
0.59326171875,
0.2156982421875,
0.59326171875
],
[
0.21728515625,
0.5966796875,
0.78125,
0.5966796875,
0.78125,
0.60693359375,
0.21728515625,
0.60693359375
],
[
0.214111328125,
0.60693359375,
0.78369140625,
0.60791015625,
0.78369140625,
0.62548828125,
0.214111328125,
0.625
],
[
0.2156982421875,
0.62548828125,
0.78369140625,
0.62548828125,
0.78369140625,
0.6396484375,
0.2156982421875,
0.6396484375
],
[
0.2156982421875,
0.64013671875,
0.78173828125,
0.64013671875,
0.78173828125,
0.65380859375,
0.2156982421875,
0.65380859375
],
[
0.2156982421875,
0.65478515625,
0.78369140625,
0.65478515625,
0.78369140625,
0.66845703125,
0.2156982421875,
0.66845703125
],
[
0.240234375,
0.66943359375,
0.78173828125,
0.66845703125,
0.78173828125,
0.68310546875,
0.240234375,
0.6845703125
],
[
0.2156982421875,
0.68505859375,
0.78173828125,
0.68505859375,
0.78173828125,
0.69873046875,
0.2156982421875,
0.69873046875
],
[
0.2156982421875,
0.70068359375,
0.78369140625,
0.70068359375,
0.78369140625,
0.7138671875,
0.2156982421875,
0.7138671875
],
[
0.21728515625,
0.71728515625,
0.78173828125,
0.71728515625,
0.78173828125,
0.72802734375,
0.21728515625,
0.72802734375
],
[
0.2156982421875,
0.73095703125,
0.78369140625,
0.73095703125,
0.78369140625,
0.744140625,
0.2156982421875,
0.744140625
],
[
0.2156982421875,
0.7470703125,
0.78173828125,
0.7470703125,
0.78173828125,
0.76025390625,
0.2156982421875,
0.76025390625
],
[
0.2344970703125,
0.7666015625,
0.78125,
0.7666015625,
0.78125,
0.78076171875,
0.2344970703125,
0.78076171875
],
[
0.2156982421875,
0.77880859375,
0.56884765625,
0.77880859375,
0.56884765625,
0.79248046875,
0.2156982421875,
0.79248046875
],
[
0.2320556640625,
0.79052734375,
0.78173828125,
0.79150390625,
0.78173828125,
0.80615234375,
0.2320556640625,
0.8046875
],
[
0.2156982421875,
0.8046875,
0.78125,
0.8046875,
0.78125,
0.818359375,
0.2156982421875,
0.818359375
],
[
0.21728515625,
0.818359375,
0.31201171875,
0.818359375,
0.31201171875,
0.8291015625,
0.21728515625,
0.8291015625
],
[
0.238525390625,
0.8271484375,
0.78173828125,
0.8271484375,
0.78173828125,
0.84130859375,
0.238525390625,
0.84130859375
],
[
0.2156982421875,
0.8388671875,
0.42236328125,
0.8388671875,
0.42236328125,
0.85302734375,
0.2156982421875,
0.85302734375
],
[
0.490234375,
0.87548828125,
0.50634765625,
0.87548828125,
0.50634765625,
0.89013671875,
0.490234375,
0.89013671875
]
] |
[
"",
"Introduction",
"Random forests is a method for classification or regression in which we con-",
"struct an ensembleofdecision trees for (randomsubsets of) thetraining data",
"and, in the classification phase, aggregate their outcomes by majority voting",
"The random-forests method has received a tremendous amount of attention",
"for its simplicity and improved accuracy over plain decision trees [3, 26, 19, 24].",
"Commonly, fast heuristics without performance guarantees are used for comput-",
"ing random forests [19, 24], in particular for computing the individual decision",
"trees in the forest. For plain decision trees, researchers lately made several ad-",
"vances in computing optimal decision trees, that is, decision trees that provably",
"optimize criteria such as minimizing the tree size [1, 21, 4, 7]. With that in-",
"creased amount of attention also came theoretical advances, showing the limits",
"and opportunities for developing efficient exact algorithms for computing deci-",
"sion trees [23, 18, 9]. One impetus to computing optimal decision trees is that",
"minimizing the size reduces tendencies to overfitting [1, 7]. It is conceivable",
"that such benefits transfer to globally optimizing the tree ensembles computed",
"by random forests. However, apart from sporadic hardness results [25], we are",
"not aware of exact algorithmic research for tree ensembles. In this work, we aim",
"to initiate this direction; that is, we begin to build the theoretical footing for",
"exact algorithmics of computing optimal tree ensembles and provide potential",
"avenues for exact algorithms that are guaranteed to provide optimal results with",
"acceptable worst-case running times.",
"We study the algorithmic properties of two canonical formulations of the",
"training problem for tree ensembles: We are given a set of training examples",
"labeled with two classes and a number l of trees and we want to compute a tree",
"ensemble containing l trees that classifies the examples consistently with the",
"given class labels.1 We want to minimize either the sum of the tree sizes, result-",
"ing in the problem MINIMUM TREE ENSEMBLE SIZE (MTES), or the largest size",
"ofatreeinthe ensemble,resultingintheproblemMINIMAxTREE ENSEMBLE",
"SizE (MMAxTES).² Both contain as a special case the problem of computing",
"a minimum-size decision tree, which is known to be NP-hard [15, 23]. However,",
"the hardness constructions do not necessarily reflect practical data. Thus, we",
"are interested in precisely which properties make the problems hard or tractable.",
"Mainly, we provide two novel algorithms for MTES and MMAxTEs3 and",
"matching lower-bound results for their running times. We call the first one",
"witness-tree algorithm. This algorithm demonstrates that prospects for tractable",
"algorithmsfor optimizing decision trees canbe non-trivially generalized toopti-",
"mizing tree ensembles. Namely, it was known that for small tree size s, moderate",
"maximum domain size D of any feature, and moderate number d of features in",
"ITo keep the presentation focused, we consider mainly the case without training error. See",
"the conclusion for extensions to minimizing training error",
"2it is also natural to consider the depths of the trees instead of their sizes, but results are",
"usually transferable between these two optimization goals and the size makes the presentation",
"moreaccessible",
"3The algorithms work on the decision version of these problems, but they easily apply to",
"the optimization versions as well.",
"2"
] |
{
"arxiv_id": "2306.04423",
"language": "en",
"page_idx": 1
}
|
||
2306.04423-en-2
|
2306.04423-en-2.jpg
|
which two examples differ$^4$,a minimum decision tree can be computed efficiently, that is, in $f(s,D,\delta)\cdot$poly time, where poly is a polynomial in the input size [23]. However, the function $f$ is at least $\delta^s\cdot(D^s2\delta)^s\cdot2^{s^2}$ and the algorithm involves enumerative steps in which the worst-case running time equals the average case. We show that, even for the more general MTES, we can improve the running time to $\mathcal{O}((6\delta DS)^S\cdot S\ell n)$, where $\bar{S}$ denotes the sum of the tree sizes, $\ell$ the number of trees in the ensemble, and $n$ the number of training examples (Theorem 4.1). Moreover, we can avoid the enumerative approach, obtaining a search-tree algorithm that is both conceptually simpler and more easily amenable to heuristic improvements such as early search-termination rules and data reduction.$^5$ We achieve this by growing the trees iteratively and labeling their leaves with witness examples that need to be classified in these leaves. This allows us to localize misclassifications and their rectification, shrinking the search space. We believe that this technique may have practical applications beyond improving the worstcase running times as we do here. The running time that we achieve is tight in the sense that we cannot decrease the exponent to $o(S)$ without violating reasonable complexity-theoretic assumptions (Theorem 4.4).
Recently, exponential-time dynamic programming has been applied to compute optimal decision trees and the resulting trees have shown comparable performance to (heuristic) random forests on some datasets [7]. With the second algorithm that we provide, we investigate the potential of dynamic programming for computing optimal tree ensembles. We first show that minimizing decision trees can be done in $\mathcal{O}(3^n)$ time, where $n$ is the number of input examples, by a dynamic-programming approach that works on all possible splits of the examples (Corollary 5.2). (Indeed, the algorithm employed by [7] similarly computes a table over all possible splits in the worst case.) We then extend this algorithm to tree ensembles with $\ell$ trees, achieving $(\ell+1)^n\cdot$poly running time (Theorem 5.3). Unfortunately, we also show that the running time cannot be substantially improved: A running time of $f(\ell)\cdot2^{o(\log\ell)\cdot n}$ would violate reasonable complexity-theoretic assumptions (Theorem 5.4).
Finally, we compare the power of decision trees and tree ensembles in terms of their sizes. Here, we show that a training data set $\mathcal{D}$ that can be classified by a tree ensemble with $\ell$ trees of size at most $s$, can also be classified by a decision tree of size $(s+1)^\ell$ (Theorem 3.2). However, such an exponential increase is necessary in the worst case: We show that there are such training data sets $\mathcal{D}$ that cannot be classified by any decision tree of size roughly $(s/2)^{\bar{\ell}/2}$ (Theorem 3.3).
In summary, as the number of trees in a tree ensemble grow, the classification power increases exponentially over decision trees. Nevertheless, we are able to carry over and substantially improve on tractability results for decision trees if in particular the number of cuts in the optimal ensemble is relatively small. The underlying witness-tree technique seems promising to try in practice. Fur-
$^{4}$See [23] for measurements showing that this is a reasonably small parameter in several
datasets.
$^5$In the meantime since this paper has been accepted, [9] showed that for minimum-size decision trees the dependency on the domain size can be dropped, that is, there is an $f($s$, \delta ) \cdot$ poly-time algorithm. It still uses enumerative steps which would be infeasible in practice.
3
|
[
[
0.2156982421875,
0.1578369140625,
0.78125,
0.1578369140625,
0.78125,
0.17236328125,
0.2156982421875,
0.17236328125
],
[
0.2156982421875,
0.175537109375,
0.78173828125,
0.175537109375,
0.78173828125,
0.189453125,
0.2156982421875,
0.189453125
],
[
0.2156982421875,
0.1900634765625,
0.78369140625,
0.1900634765625,
0.78369140625,
0.2052001953125,
0.2156982421875,
0.2052001953125
],
[
0.2156982421875,
0.20703125,
0.78173828125,
0.20703125,
0.78173828125,
0.2203369140625,
0.2156982421875,
0.2203369140625
],
[
0.2156982421875,
0.2215576171875,
0.78173828125,
0.2215576171875,
0.78173828125,
0.236083984375,
0.2156982421875,
0.236083984375
],
[
0.214111328125,
0.23486328125,
0.78125,
0.236083984375,
0.78125,
0.250732421875,
0.214111328125,
0.25
],
[
0.2156982421875,
0.251953125,
0.78173828125,
0.251953125,
0.78173828125,
0.26513671875,
0.2156982421875,
0.26513671875
],
[
0.2156982421875,
0.267578125,
0.78173828125,
0.267578125,
0.78173828125,
0.281005859375,
0.2156982421875,
0.281005859375
],
[
0.2156982421875,
0.2822265625,
0.78173828125,
0.2822265625,
0.78173828125,
0.29541015625,
0.2156982421875,
0.29541015625
],
[
0.214111328125,
0.296630859375,
0.78173828125,
0.294189453125,
0.78173828125,
0.309326171875,
0.214111328125,
0.311279296875
],
[
0.2156982421875,
0.3125,
0.78173828125,
0.3125,
0.78173828125,
0.32568359375,
0.2156982421875,
0.32568359375
],
[
0.2156982421875,
0.326904296875,
0.78173828125,
0.326904296875,
0.78173828125,
0.34033203125,
0.2156982421875,
0.34033203125
],
[
0.2156982421875,
0.34228515625,
0.78173828125,
0.34228515625,
0.78173828125,
0.35595703125,
0.2156982421875,
0.35595703125
],
[
0.214111328125,
0.356689453125,
0.78173828125,
0.356689453125,
0.78173828125,
0.37060546875,
0.214111328125,
0.37060546875
],
[
0.2156982421875,
0.37255859375,
0.78125,
0.37255859375,
0.78125,
0.386474609375,
0.2156982421875,
0.386474609375
],
[
0.214111328125,
0.386474609375,
0.78173828125,
0.38818359375,
0.78173828125,
0.40283203125,
0.214111328125,
0.40087890625
],
[
0.2156982421875,
0.40283203125,
0.64794921875,
0.40283203125,
0.64794921875,
0.416748046875,
0.2156982421875,
0.416748046875
],
[
0.240234375,
0.41845703125,
0.78125,
0.41845703125,
0.78125,
0.431884765625,
0.240234375,
0.431884765625
],
[
0.21728515625,
0.43310546875,
0.78173828125,
0.43310546875,
0.78173828125,
0.4462890625,
0.21728515625,
0.4462890625
],
[
0.2156982421875,
0.447509765625,
0.78173828125,
0.447509765625,
0.78173828125,
0.4609375,
0.2156982421875,
0.4609375
],
[
0.2156982421875,
0.46337890625,
0.78369140625,
0.46337890625,
0.78369140625,
0.4765625,
0.2156982421875,
0.4765625
],
[
0.214111328125,
0.4765625,
0.78173828125,
0.477783203125,
0.78173828125,
0.492431640625,
0.214111328125,
0.4912109375
],
[
0.2156982421875,
0.49365234375,
0.78173828125,
0.49365234375,
0.78173828125,
0.5068359375,
0.2156982421875,
0.5068359375
],
[
0.2156982421875,
0.50830078125,
0.78173828125,
0.50830078125,
0.78173828125,
0.52294921875,
0.2156982421875,
0.52294921875
],
[
0.2156982421875,
0.5234375,
0.78173828125,
0.5234375,
0.78173828125,
0.537109375,
0.2156982421875,
0.537109375
],
[
0.2156982421875,
0.53857421875,
0.78173828125,
0.53857421875,
0.78173828125,
0.55322265625,
0.2156982421875,
0.55322265625
],
[
0.2156982421875,
0.55322265625,
0.78173828125,
0.5537109375,
0.78173828125,
0.568359375,
0.2156982421875,
0.5673828125
],
[
0.21728515625,
0.5693359375,
0.78369140625,
0.5693359375,
0.78369140625,
0.58349609375,
0.21728515625,
0.58349609375
],
[
0.214111328125,
0.58056640625,
0.78369140625,
0.58056640625,
0.78369140625,
0.5986328125,
0.214111328125,
0.5986328125
],
[
0.2156982421875,
0.599609375,
0.64794921875,
0.599609375,
0.64794921875,
0.61279296875,
0.2156982421875,
0.61279296875
],
[
0.2418212890625,
0.6142578125,
0.78173828125,
0.6142578125,
0.78173828125,
0.62744140625,
0.2418212890625,
0.62744140625
],
[
0.21728515625,
0.62890625,
0.78173828125,
0.62890625,
0.78173828125,
0.64306640625,
0.21728515625,
0.64306640625
],
[
0.2156982421875,
0.64453125,
0.78125,
0.64453125,
0.78125,
0.65771484375,
0.2156982421875,
0.65771484375
],
[
0.2156982421875,
0.6591796875,
0.78173828125,
0.6591796875,
0.78173828125,
0.67236328125,
0.2156982421875,
0.67236328125
],
[
0.2156982421875,
0.6748046875,
0.78125,
0.6748046875,
0.78125,
0.68798828125,
0.2156982421875,
0.68798828125
],
[
0.2156982421875,
0.689453125,
0.78173828125,
0.689453125,
0.78173828125,
0.7041015625,
0.2156982421875,
0.7041015625
],
[
0.240234375,
0.705078125,
0.78369140625,
0.705078125,
0.78369140625,
0.71826171875,
0.240234375,
0.71826171875
],
[
0.21728515625,
0.7216796875,
0.78125,
0.7216796875,
0.78125,
0.73193359375,
0.21728515625,
0.73193359375
],
[
0.2156982421875,
0.73486328125,
0.78173828125,
0.73486328125,
0.78173828125,
0.74853515625,
0.2156982421875,
0.74853515625
],
[
0.2156982421875,
0.75048828125,
0.78173828125,
0.75048828125,
0.78173828125,
0.7646484375,
0.2156982421875,
0.7646484375
],
[
0.21728515625,
0.76513671875,
0.78173828125,
0.76513671875,
0.78173828125,
0.77880859375,
0.21728515625,
0.77880859375
],
[
0.2344970703125,
0.7861328125,
0.78173828125,
0.787109375,
0.78173828125,
0.8017578125,
0.2344970703125,
0.80029296875
],
[
0.2156982421875,
0.8017578125,
0.271240234375,
0.8017578125,
0.271240234375,
0.81298828125,
0.2156982421875,
0.81298828125
],
[
0.233642578125,
0.8125,
0.78173828125,
0.8125,
0.78173828125,
0.82568359375,
0.233642578125,
0.82568359375
],
[
0.2156982421875,
0.82373046875,
0.77783203125,
0.82470703125,
0.77783203125,
0.83837890625,
0.2156982421875,
0.8369140625
],
[
0.2156982421875,
0.83837890625,
0.7587890625,
0.83837890625,
0.7587890625,
0.8486328125,
0.2156982421875,
0.8486328125
],
[
0.490234375,
0.87548828125,
0.50732421875,
0.87548828125,
0.50732421875,
0.8896484375,
0.490234375,
0.8896484375
]
] |
[
"which two examples differ4, a minimum decision tree can be computed efficiently",
"that is, in f(s, D, ) poly time, where poly is a polynomial in the input size [23].",
"However, the function f is at least 8$.(Ds25)s.2s\" and the algorithm involves enu-",
"merative steps in which the worst-case running time equals the average case. We",
"show that, even for the more general MTES, we can improve the running time",
"to O((6oDS)S . Sln), where S denotes the sum of the tree sizes, l the number",
"of trees in the ensemble, and n the number of training examples (Theorem 4.1).",
"Moreover, we can avoid the enumerative approach, obtaining a search-tree algo-",
"rithm that is both conceptually simpler and more easily amenable to heuristic",
"improvements such as early search-termination rules and data reduction.5 We",
"achieve this by growing the trees iteratively and labeling their leaves with wit-",
"ness examples that need to be classified in these leaves. This allows us to localize",
"misclassifications and their rectification, shrinking the search space. We believe",
" that this technique may have practical applications beyond improving the worst-",
"case running times as we do here. The running time that we achieve is tight",
"in the sense that we cannot decrease the exponent to o(S) without violating",
"reasonable complexity-theoretic assumptions (Theorem 4.4).",
"Recently, exponential-time dynamic programming has been applied to com-",
"pute optimal decision trees and the resulting trees have shown comparable per-",
"formance to (heuristic) random forests on some datasets [7]. With the second",
"algorithm that we provide, we investigate the potential of dynamic program-",
"ming for computing optimal tree ensembles. We first show that minimizing",
"decision trees can be done in O(3n) time, where n is the number of input exam-",
"ples, by a dynamic-programming approach that works on all possible splits of",
"the examples (Corollary 5.2). (Indeed, the algorithm employed by [7] similarly",
"computes a table over all possible splits in the worst case.) We then extend",
"this algorithm to tree ensembles with e trees, achieving (e + 1)n .poly running",
"time (Theorem 5.3). Unfortunately, we also show that the running time can-",
"not be substantially improved: A running time of f(e) - 2o(log e)-n would violate",
"reasonable complexity-theoretic assumptions (Theorem 5.4).",
"Finally, we compare the power of decision trees and tree ensembles in terms",
"of their sizes. Here, we show that a training data set D that can be classified by",
"a tree ensemble with l trees of size at most s, can also be classified by a decision",
"tree of size (s+1) (Theorem 3.2). However, such an exponential increase is nec-",
"essary in the worst case: We show that there are such training data sets D that",
"cannot be classified by any decision tree of size roughly (s/2)e/2 (Theorem 3.3).",
"In summary, as the number of trees in a tree ensemble grow, the classifica-",
"tion powerincreases exponentially over decisiontrees.Nevertheless.weareable",
"to carry over and substantially improve on tractability results for decision trees",
"if in particular the number of cuts in the optimal ensemble is relatively small.",
"The underlying witness-tree technique seems promising to try in practice. Fur-",
"4See [23] for measurements showing that this is a reasonably small parameter in several",
"datasets.",
"5In the meantime since this paper has been accepted, [9] showed that for minimum-size",
"decision trees the dependency on the domain size can be dropped, that is, there is an f(s,)",
"poly-time algorithm.It still uses enumerative steps which would be infeasible in practice.",
"3"
] |
{
"arxiv_id": "2306.04423",
"language": "en",
"page_idx": 2
}
|
||
2306.04423-en-3
|
2306.04423-en-3.jpg
|
thermore, we show that dynamic programming, which has been successful for decision trees, may also be viable for tree ensembles. We also provide matching lower bounds for the running times. Apart from tuning our algorithms, in the future, deconstructing these lower bounds may provide further guidelines towards which properties of the input data we may exploit for efficient algorithms and which we likely may not.
## 2 Preliminaries
For $n\in\mathbb{N}$ we use $[n]:=\{1,2,\ldots,n\}.$ For a vector $x\in\mathbb{R}^d$ we denote by $x[i]$ the
$i$th entry of $x.$
Let $\Sigma$ be a set of class symbols; unless stated otherwise, we use $\Sigma=\{$blue,red$\}.$ A decision tree in $\mathbb{R}^d$ with classes $\Sigma$ is the following. Let $T$ be an ordered binary tree, that is, each inner node has a well-defned left and right child. Let dim:$V(T)\to[d]$ and thr:$V(T)\to\mathbb{R}$ be labelings of each internal node $t\in V(T)$ by a dimension dim$(t)\in[d]$ and a threshold thr$(t)\in\mathbb{R}.$ Furthermore, let cla$(\ell):V(T)\to\Sigma$ be a labeling of the leaves of $T$ by class symbols. Then the tuple $(T$,dim,thr,cla) is a $decision\textit{ tree in }\mathbb{R} ^d$ with classes $\Sigma.$ We often omit the labelings dim,thr,cla and just refer to the tree $T.$ The size of $T$ is the number of its internal nodes. We conveniently call the internal nodes of $T$ and their associated labels $cuts.$
A $training\textit{ data set is a tuple }( E, \lambda )$ of a set of $examplesE\subseteq\mathbb{R}^d$ and their class labeling $\lambda:E\to\Sigma.$ Given a training data set, we fx for each dimension $i$ a minimum-size set Thr$(i)$ of thresholds that distinguishes between all values of the examples in the $i$th dimension. In other words, for each pair of elements $e$ and $e^\prime$ with $e[i]<e^\prime[i]$, there is at least one value $t\in$Thr$(i)$ such that $e[i]<t<e^{\prime}[i].$ Let $t\in\mathbb{R}$ be some threshold. We use $E[f_{i}\leq$ $t]$ = $\{ x$ $\in$ $E$ | $x[ i]$ $\leq$ $t\}$ and $E[ f_i$ > $t]$ = $\{ x$ $\in$ $E$ | $x[ i]$ > $t\}$ to denote the set of examples of $E$ whose $i$th dimension is less or equal and strictly greater than the threshold $t$, respectively. Now let $T$ be a decision tree. Each node $t\in V(T)$, including the leaves, defines a subset $E[T,t]\subseteq E$ as follows. For the root $t$ of $T$, we define $E[T,t]:=E.$ For each non-root node $t$, let $p$ denote the parent of $t.$ We then define $E[T,t]:=E[T,p]\cap E[f_{\dim(p)}\leq$ thr$(p)]$ if $t$ is the $\operatorname*{left~child~of~}p$ and $E[T,t]:=E[T,p]\cap E[f_{\mathrm{dim}(p)}>$thr$(p)]$ if $t$ is the right child of $p.$ If the tree $T$ is clear from the context, we simplify $E[T,t]$ to $E[t].$ We say that $T\textit{classifies }( E, \lambda )$ if for each leaf $u\in V(T)$ and each example $e\in E[u]$ we have $\lambda(e)=$cla$(u)$ (recall that cla is a labeling of the leaves of $T$ by classes). Note that the set family that contains $E[u]$ for all leaves $u$ of $T$ forms a partition of $E.$ Thus for each example $e\in E$ there is a unique leaf $u$ such that $e\in E[u].$ We also say that $u$ is $e^{\prime }s\textit{leaf. We say that cla}( u)$ is the class assigned to $e$ by $T$ and we write $T[e]$ for cla$(u).$
A $tree\textit{ ensemble is a set of decision trees. A tree ensemble }\mathcal{T} \textit{classifies }( E, \lambda )$ if for each example $e\in E$ the majority vote of the trees in $\mathcal{T}$ agrees with the $\operatorname*{label}\lambda(e).$ That is, for each example $e\in E$ we have $\lambda(e)=\operatorname*{arg}\max_\sigma\in\Sigma|\{T\in$ $\mathcal{T}\mid T[e]=\sigma\}|=:\mathcal{T}[e].$ To avoid ambiguity in the maximum, we fx an ordering
4
|
[
[
0.21728515625,
0.1590576171875,
0.779296875,
0.1590576171875,
0.779296875,
0.173583984375,
0.21728515625,
0.173583984375
],
[
0.2156982421875,
0.173583984375,
0.779296875,
0.1749267578125,
0.779296875,
0.1881103515625,
0.2156982421875,
0.1875
],
[
0.2156982421875,
0.1881103515625,
0.78125,
0.1881103515625,
0.78125,
0.2020263671875,
0.2156982421875,
0.2020263671875
],
[
0.2156982421875,
0.2052001953125,
0.78173828125,
0.2052001953125,
0.78173828125,
0.2183837890625,
0.2156982421875,
0.2183837890625
],
[
0.2156982421875,
0.2197265625,
0.78125,
0.2197265625,
0.78125,
0.23291015625,
0.2156982421875,
0.23291015625
],
[
0.214111328125,
0.23291015625,
0.428955078125,
0.2342529296875,
0.428955078125,
0.248779296875,
0.214111328125,
0.2474365234375
],
[
0.214111328125,
0.271484375,
0.410888671875,
0.271484375,
0.410888671875,
0.2890625,
0.214111328125,
0.2890625
],
[
0.21728515625,
0.302978515625,
0.779296875,
0.302978515625,
0.779296875,
0.317626953125,
0.21728515625,
0.317626953125
],
[
0.2181396484375,
0.3193359375,
0.318603515625,
0.3193359375,
0.318603515625,
0.330810546875,
0.2181396484375,
0.330810546875
],
[
0.238525390625,
0.33203125,
0.77783203125,
0.333251953125,
0.77783203125,
0.347900390625,
0.238525390625,
0.3466796875
],
[
0.2181396484375,
0.348388671875,
0.78125,
0.348388671875,
0.78125,
0.3623046875,
0.2181396484375,
0.3623046875
],
[
0.21728515625,
0.363037109375,
0.779296875,
0.363037109375,
0.779296875,
0.37744140625,
0.21728515625,
0.37744140625
],
[
0.21728515625,
0.37890625,
0.78173828125,
0.37890625,
0.78173828125,
0.393310546875,
0.21728515625,
0.393310546875
],
[
0.2156982421875,
0.39453125,
0.78173828125,
0.39453125,
0.78173828125,
0.40771484375,
0.2156982421875,
0.40771484375
],
[
0.2156982421875,
0.4091796875,
0.78369140625,
0.4091796875,
0.78369140625,
0.423583984375,
0.2156982421875,
0.423583984375
],
[
0.21728515625,
0.4248046875,
0.78125,
0.4248046875,
0.78125,
0.438232421875,
0.21728515625,
0.438232421875
],
[
0.21728515625,
0.439453125,
0.78125,
0.439453125,
0.78125,
0.45263671875,
0.21728515625,
0.45263671875
],
[
0.21728515625,
0.455078125,
0.78173828125,
0.455078125,
0.78173828125,
0.468505859375,
0.21728515625,
0.468505859375
],
[
0.2181396484375,
0.4716796875,
0.416748046875,
0.4716796875,
0.416748046875,
0.481689453125,
0.2181396484375,
0.481689453125
],
[
0.2418212890625,
0.48486328125,
0.78173828125,
0.48486328125,
0.78173828125,
0.498779296875,
0.2418212890625,
0.498779296875
],
[
0.2156982421875,
0.498779296875,
0.78125,
0.5,
0.78125,
0.5146484375,
0.2156982421875,
0.51318359375
],
[
0.21728515625,
0.5166015625,
0.77685546875,
0.5166015625,
0.77685546875,
0.5263671875,
0.21728515625,
0.5263671875
],
[
0.21728515625,
0.52978515625,
0.78125,
0.52978515625,
0.78125,
0.54345703125,
0.21728515625,
0.54345703125
],
[
0.2156982421875,
0.54541015625,
0.78173828125,
0.54541015625,
0.78173828125,
0.5595703125,
0.2156982421875,
0.5595703125
],
[
0.2156982421875,
0.56005859375,
0.779296875,
0.56005859375,
0.779296875,
0.57373046875,
0.2156982421875,
0.57373046875
],
[
0.21728515625,
0.57568359375,
0.779296875,
0.57568359375,
0.779296875,
0.5888671875,
0.21728515625,
0.5888671875
],
[
0.21728515625,
0.59228515625,
0.78125,
0.59228515625,
0.78125,
0.60302734375,
0.21728515625,
0.60302734375
],
[
0.21728515625,
0.60595703125,
0.78173828125,
0.60595703125,
0.78173828125,
0.619140625,
0.21728515625,
0.619140625
],
[
0.2156982421875,
0.62060546875,
0.78173828125,
0.62060546875,
0.78173828125,
0.63525390625,
0.2156982421875,
0.63525390625
],
[
0.2156982421875,
0.63623046875,
0.78125,
0.63623046875,
0.78125,
0.6494140625,
0.2156982421875,
0.6494140625
],
[
0.2156982421875,
0.6513671875,
0.78125,
0.6513671875,
0.78125,
0.66650390625,
0.2156982421875,
0.66650390625
],
[
0.2156982421875,
0.66552734375,
0.78173828125,
0.66552734375,
0.78173828125,
0.68310546875,
0.2156982421875,
0.68310546875
],
[
0.21728515625,
0.6796875,
0.779296875,
0.6796875,
0.779296875,
0.693359375,
0.21728515625,
0.693359375
],
[
0.2156982421875,
0.6943359375,
0.78125,
0.69580078125,
0.78125,
0.71044921875,
0.2156982421875,
0.708984375
],
[
0.21728515625,
0.7109375,
0.78173828125,
0.7109375,
0.78173828125,
0.724609375,
0.21728515625,
0.724609375
],
[
0.21728515625,
0.7265625,
0.779296875,
0.7265625,
0.779296875,
0.74072265625,
0.21728515625,
0.74072265625
],
[
0.21728515625,
0.7412109375,
0.78173828125,
0.7412109375,
0.78173828125,
0.7548828125,
0.21728515625,
0.7548828125
],
[
0.2156982421875,
0.755859375,
0.779296875,
0.755859375,
0.779296875,
0.76953125,
0.2156982421875,
0.76953125
],
[
0.2156982421875,
0.77099609375,
0.42236328125,
0.771484375,
0.42236328125,
0.7861328125,
0.2156982421875,
0.78515625
],
[
0.238525390625,
0.78515625,
0.78173828125,
0.7861328125,
0.78173828125,
0.80029296875,
0.238525390625,
0.7998046875
],
[
0.21728515625,
0.8017578125,
0.78125,
0.8017578125,
0.78125,
0.81494140625,
0.21728515625,
0.81494140625
],
[
0.21728515625,
0.8173828125,
0.779296875,
0.8173828125,
0.779296875,
0.83203125,
0.21728515625,
0.83203125
],
[
0.2156982421875,
0.830078125,
0.78173828125,
0.83203125,
0.78173828125,
0.8466796875,
0.2156982421875,
0.8447265625
],
[
0.490234375,
0.87548828125,
0.50732421875,
0.87548828125,
0.50732421875,
0.8896484375,
0.490234375,
0.8896484375
]
] |
[
"thermore, we show that dymamic programming, which has been successful for",
"decision trees,may also be viable for tree ensembles. We also providematching",
"lower bounds for the running times. Apart from tuning our algorithms, in the",
"future, deconstructing these lower bounds may provide further guidelines to-",
"wards which properties of the input data we may exploit for efficient algorithms",
"and which we likely may not.",
"2Preliminaries",
"For n e N we use [n] := {1, 2, ..., n]. For a vector r e Rd we denote by &[il the",
"ithentryofa.",
"Let be a set of class symbols; unless stated otherwise, we use ",
"{blue,red]. A decision tree in Rd with classes is the following. Let T be",
"an ordered binary tree, that is, each inner node has a well-defined left and right",
"child.. Let dim: V(T) → [dl and thr: V(T) -→ R be labelings of each internal",
"node t e V(T) by a dimension dim(t) e [dl and a threshold thr(t) eR. Further-",
"more, let cla(e): V(T) → be a labeling of the leaves of T by class symbols.",
"Then the tuple (T, dim, thr, cla) is a decision tree in Rd with classes . We often",
"omit the labelings dim, thr, cla and just refer to the tree T. The size of T is the",
"number of its internal nodes. We conveniently call the internal nodes of T and",
"their associated labels cuts.",
"A training data set is a tuple (E,) of a set of eramples E C Rd and",
"their class labeling X: E → . Given a training data set, we fix for each",
"dimension i aminimum-sizesetThr(z)ofthresholdsthat distinguishesbetweer",
"all values of the examples in the ith dimension. In other words, for each pair",
"of elements e and e' with e[i] < e'lil, there is at least one value t e Thr(i)",
"such that e[il < t < e'Jil. Let t e R be some threshold. We use E[fi <",
"t] = ( e El a[il <t} and E[fi> t] ={r e E ail > t) to denote the",
"set ofexamples of E whoseith dimensionis lessoregualandstrictlygreater",
"than the threshold t, respectively. Now let T be a decision tree. Each node",
"t e V(T), including the leaves, defines a subset E[T,t] C E as follows. For the",
"root t of T, we define E[T,tl := E. For each non-root node t, let p denote the",
"parent of t. We then define E[T,t] := E[T,pl nE[fdim(p) ≤ thr(p)] if t is the",
"left child of p and E[T,t] := E[T,p] n E[faim(p) > thr(p)] if t is the right child",
"of p. If the tree T is clear from the context, we simplify E[T,t] to E[t]. We say",
"that T classifies (E, ) if for each leaf u E V(T) and each example e E E[u] we",
"have (e) = cla(u) (recall that cla is a labeling of the leaves of T by classes)",
"Note that the set family that contains E[ul for all leaves u of T forms a partition",
"of E. Thus for each example e e E there is a unique leaf u such that e e E[ul.",
"We also say that u is e's leaf. We say that cla(u) is the class assigned to e by T",
"and we write T[e] for cla(u).",
"A tree ensemble is a set of decision trees. A tree ensemble T classifies (E, )",
"if for each example e e E the majority vote of the trees in T agrees with the",
"label ^(e). That is, for each example e E E we have (e) = arg maxael(T E",
"T T[e] = ol =: T[el. To avoid ambiguity in the maximum, we fix an ordering",
"4"
] |
{
"arxiv_id": "2306.04423",
"language": "en",
"page_idx": 3
}
|
||
2306.04423-en-4
|
2306.04423-en-4.jpg
|
of $\Sigma$ and break ties according to this ordering. If $\Sigma=\{$blue,red$\}$ we break ties
in favor of blue. The $overall\textit{size of a tree ensemble }\mathcal{T}$ is the sum of the sizes of
the decision trees in $\mathcal{T}.$
The computational problems that we consider are as follows
$\begin{array}{l}{\mathrm{MINIMUM~TREE~ENSEMBLE~SIZE~(MTES)}}\\{Instance:~~~~~~A~training~data~set~(E,\lambda),~a~number~\ell~of~trees,~and~a~size}\end{array}$
bound $S.$
$Question{:}$ Is there a tree ensemble of overall size at most $S$ that
classifies $(E,\lambda)$ and contains exactly $\ell$ trees?
When restricting to $\ell=1$, MTES is known as MINIMUM DECISION TREE SIZE (DTS) [23,18]. In the problem variant MINIMAX TREE ENSEMBLE SIZE (MMAXTES),instead of $S$,we are given an integer $s$ and we ask whether there is a tree ensemble that classifies $(E,\lambda)$ and contains exactly $\ell$ trees, each of which has size at most $s.$
Our analysis is within the framework of parameterized complexity [13,10,22, 6,8]. Let $L\subseteq\Sigma^*$ be a computational problem specified over some alphabet $\Sigma$ and let $p:\Sigma^*\to\mathbb{N}$ be a parameter, that is, $p$ assigns to each instance of $L$ an integer parameter value (which we simply denote by $p$ if the instance is clear from the context). We say that $L$ is fixed-$parameter\textit{tractable ( FPT) with respect to}$ $p$ if it can be decided in $f(p)\cdot$poly$(n)$ time where $n$ is the input encoding length. The corresponding hardness concept related to fxed-parameter tractability is W$[t]$-hardness, $t\geq1;$ if problem $L$ is W$[t]$-hard with respect to $p$ then $L$ is thought to not be fixed-parameter tractable; see [10,22,6,8] for details. The Exponential Time Hypothesis (ETH) [16,17] states that 3SAT on $n$-variable ${\tilde{\text{formulas cannot be solved in }}2^{o(n)}\mathrm{~time.}}$
3 Decision Trees Versus Tree Ensembles
We will call a decision tree and a tree ensemble equivalent if any training data set is classified by the one if and only if it is classified by the other. We start by comparing the minimum size of a decision tree to that of a minimum-size decision tree ensemble, showing that there are examples where the latter is significantly smaller. Vidal and Schiffer [27, Theorem 2] obtained similar results for the depth of a decision tree and decision tree ensemble, showing that any training data set that can be classified by a decision tree ensemble with $\ell$ trees of depth $t$ can also be classified by a tree with depth $\ell\cdot t$ and that this bound is tight. Here, we analyze trees and tree ensembles in terms of their size, showing that a minimum ensemble can be exponentially smaller than any equivalent decision tree.
Observation 3.1. If $s$ is the size of a decision tree and $L$ the number of leaves,
$\operatorname{then}s=L-1.$
Theorem $3. 2. Any\textit{ training data set that can be classified by a decision tree}$ ensemble consisting $of$ $\ell$ $trees$, each $of$ size $at$ most $s$, can also $be$ classified $by$ $a$ $decision\textit{ tree of size }( s+ 1) ^{\ell }- 1.$
5
|
[
[
0.21728515625,
0.1590576171875,
0.78125,
0.1590576171875,
0.78125,
0.17236328125,
0.21728515625,
0.17236328125
],
[
0.21728515625,
0.1749267578125,
0.78173828125,
0.1749267578125,
0.78173828125,
0.1881103515625,
0.21728515625,
0.1881103515625
],
[
0.21728515625,
0.189453125,
0.384033203125,
0.189453125,
0.384033203125,
0.20263671875,
0.21728515625,
0.20263671875
],
[
0.2418212890625,
0.203857421875,
0.6748046875,
0.203857421875,
0.6748046875,
0.2183837890625,
0.2418212890625,
0.2183837890625
],
[
0.25732421875,
0.2266845703125,
0.56787109375,
0.2266845703125,
0.56787109375,
0.2412109375,
0.25732421875,
0.2412109375
],
[
0.25732421875,
0.240478515625,
0.73876953125,
0.240478515625,
0.73876953125,
0.255126953125,
0.25732421875,
0.255126953125
],
[
0.281982421875,
0.255615234375,
0.3447265625,
0.255615234375,
0.3447265625,
0.267578125,
0.281982421875,
0.267578125
],
[
0.2607421875,
0.266357421875,
0.740234375,
0.266357421875,
0.740234375,
0.281005859375,
0.2607421875,
0.281005859375
],
[
0.281005859375,
0.2822265625,
0.60302734375,
0.2822265625,
0.60302734375,
0.29541015625,
0.281005859375,
0.29541015625
],
[
0.24267578125,
0.3037109375,
0.78125,
0.3037109375,
0.78125,
0.317626953125,
0.24267578125,
0.317626953125
],
[
0.2181396484375,
0.32080078125,
0.78173828125,
0.32080078125,
0.78173828125,
0.333984375,
0.2181396484375,
0.333984375
],
[
0.21728515625,
0.335205078125,
0.78125,
0.335205078125,
0.78125,
0.348388671875,
0.21728515625,
0.348388671875
],
[
0.21728515625,
0.35107421875,
0.78173828125,
0.35107421875,
0.78173828125,
0.3642578125,
0.21728515625,
0.3642578125
],
[
0.2181396484375,
0.367431640625,
0.393798828125,
0.367431640625,
0.393798828125,
0.37744140625,
0.2181396484375,
0.37744140625
],
[
0.2418212890625,
0.380126953125,
0.78125,
0.380126953125,
0.78125,
0.39453125,
0.2418212890625,
0.39453125
],
[
0.21728515625,
0.395751953125,
0.78173828125,
0.395751953125,
0.78173828125,
0.410400390625,
0.21728515625,
0.410400390625
],
[
0.21728515625,
0.410888671875,
0.78125,
0.410888671875,
0.78125,
0.4248046875,
0.21728515625,
0.4248046875
],
[
0.21728515625,
0.4267578125,
0.78125,
0.4267578125,
0.78125,
0.43994140625,
0.21728515625,
0.43994140625
],
[
0.21728515625,
0.44140625,
0.78125,
0.44140625,
0.78125,
0.455078125,
0.21728515625,
0.455078125
],
[
0.21728515625,
0.455810546875,
0.78173828125,
0.455810546875,
0.78173828125,
0.47021484375,
0.21728515625,
0.47021484375
],
[
0.21728515625,
0.4716796875,
0.78369140625,
0.4716796875,
0.78369140625,
0.48486328125,
0.21728515625,
0.48486328125
],
[
0.21728515625,
0.486083984375,
0.78173828125,
0.486083984375,
0.78173828125,
0.5,
0.21728515625,
0.5
],
[
0.21728515625,
0.501953125,
0.78173828125,
0.501953125,
0.78173828125,
0.51513671875,
0.21728515625,
0.51513671875
],
[
0.2181396484375,
0.5166015625,
0.78173828125,
0.5166015625,
0.78173828125,
0.52978515625,
0.2181396484375,
0.52978515625
],
[
0.2156982421875,
0.53076171875,
0.50341796875,
0.53076171875,
0.50341796875,
0.54541015625,
0.2156982421875,
0.54541015625
],
[
0.21728515625,
0.57080078125,
0.2418212890625,
0.57080078125,
0.2418212890625,
0.583984375,
0.21728515625,
0.583984375
],
[
0.250732421875,
0.57080078125,
0.69677734375,
0.57080078125,
0.69677734375,
0.58544921875,
0.250732421875,
0.58544921875
],
[
0.21728515625,
0.599609375,
0.779296875,
0.599609375,
0.779296875,
0.61279296875,
0.21728515625,
0.61279296875
],
[
0.214111328125,
0.61572265625,
0.78125,
0.6142578125,
0.78125,
0.62744140625,
0.214111328125,
0.62890625
],
[
0.2156982421875,
0.6298828125,
0.779296875,
0.62890625,
0.779296875,
0.64306640625,
0.2156982421875,
0.64453125
],
[
0.2156982421875,
0.64599609375,
0.78125,
0.64599609375,
0.78125,
0.6591796875,
0.2156982421875,
0.6591796875
],
[
0.2156982421875,
0.66015625,
0.78125,
0.66015625,
0.78125,
0.673828125,
0.2156982421875,
0.673828125
],
[
0.21728515625,
0.67529296875,
0.78125,
0.67529296875,
0.78125,
0.689453125,
0.21728515625,
0.689453125
],
[
0.21728515625,
0.6923828125,
0.779296875,
0.6923828125,
0.779296875,
0.70263671875,
0.21728515625,
0.70263671875
],
[
0.2156982421875,
0.70556640625,
0.78173828125,
0.70556640625,
0.78173828125,
0.7197265625,
0.2156982421875,
0.7197265625
],
[
0.2156982421875,
0.72021484375,
0.78125,
0.72021484375,
0.78125,
0.734375,
0.2156982421875,
0.734375
],
[
0.2156982421875,
0.73486328125,
0.779296875,
0.73486328125,
0.779296875,
0.74853515625,
0.2156982421875,
0.74853515625
],
[
0.2156982421875,
0.75244140625,
0.25,
0.75244140625,
0.25,
0.7646484375,
0.2156982421875,
0.7646484375
],
[
0.2156982421875,
0.77099609375,
0.78125,
0.771484375,
0.78125,
0.7861328125,
0.2156982421875,
0.78515625
],
[
0.21728515625,
0.787109375,
0.330810546875,
0.787109375,
0.330810546875,
0.80029296875,
0.21728515625,
0.80029296875
],
[
0.2156982421875,
0.80810546875,
0.779296875,
0.80810546875,
0.779296875,
0.82275390625,
0.2156982421875,
0.82275390625
],
[
0.2156982421875,
0.82275390625,
0.78125,
0.82373046875,
0.78125,
0.83837890625,
0.2156982421875,
0.8369140625
],
[
0.21728515625,
0.83837890625,
0.450927734375,
0.83837890625,
0.450927734375,
0.8515625,
0.21728515625,
0.8515625
],
[
0.490234375,
0.87548828125,
0.50732421875,
0.87548828125,
0.50732421875,
0.89013671875,
0.490234375,
0.89013671875
]
] |
[
"of and break ties according to this ordering. If = {blue, red we break ties",
"in favor of blue. The overall size of a tree ensemble T is the sum of the sizes of",
"the decision trees in T.",
"The computational problems that we consider are as follows",
"MINIMUM TREE ENSEMBLE SIZE (MTES)",
"Instance: A training data set (E, ), a number l of trees, and a size",
"bound S",
"Question: Is there a tree ensemble of overall size at most S that",
"classifies (E, ) and contains exactly l trees?",
"When restricting to l =1, MTES is known as MINIMUM DECISION TREE",
"SIzE(DTS)[23, 18]. In theproblem variant MINIMAX TREE ENSEMBLE SIZE",
"(MMAxTES), instead of S, we are given an integer s and we ask whether there",
"is a tree ensemble that classifies (E,) and contains exactly l trees, each of",
"whichhas sizeat mosts",
"Our analysis is within the framework of parameterized complexity [13, 10, 22,",
"6, 8]. Let L * be a computational problem specified over some alphabet ",
"and let p: * → N be a parameter, that is, p assigns to each instance of L an",
"integer parameter value (which we simply denote by p if the instance is clear from",
"the context). We say that L is fired-parameter tractable (FPT) with respect to",
"p if it can be decided in f(p)·poly(n) time where n is the input encoding length.",
"The corresponding hardness concept related to fixed-parameter tractability is",
"W[t]-hardness, t ≥ 1; if problem L is W[t]-hard with respect to p then L is",
"thought to not be fixed-parameter tractable; see [10, 22, 6, 8] for details. The",
"Exponential Time Hypothesis (ETH) [16, 17]l states that 3SAT on n-variable",
"formulas cannot be solved in 2o(n) time.",
"3",
"Decision Trees Versus Tree Ensembles",
"We will call a decision tree and a tree ensemble equivalent if any training data",
"set is classified by the one if and only if it is classified by the other.We start by",
"comparing the minimum size of a decision tree to that of a minimum-size decision",
"tree ensemble, showing that there are examples where the latter is significantly",
"smaller. Vidal and Schiffer [27, Theorem 2] obtained similar results for the",
"depth of a decision tree and decision tree ensemble, showing that any training",
"data set that can be classified bya decision tree ensemble with trees of depth",
"t can also be classified by a tree with depth l . t and that this bound is tight.",
"Here, we analyze trees and tree ensembles in terms of their size, showing that",
"a minimum ensemble can be exponentially smaller than any equivalent decision",
"tree.",
"Observation 3.1. If s is the size of a decision tree and L the number of leaves.",
"then s = L - 1.",
"Theorem 3.2. Any training data set that can be classified by a decision tree",
"ensemble consisting of trees, each of size at most s, can also be classified by a",
"decision tree of size (s + 1)e- 1.",
"5"
] |
{
"arxiv_id": "2306.04423",
"language": "en",
"page_idx": 4
}
|
||
2308.09386-en-1
|
2308.09386-en-1.jpg
|
However, NeRF registration is under-explored since existing works focus mostly on point cloud registration. Unlike point clouds that are simple explicit representation, NeRF encodes scenes implicitly, which makes registering multiple NeRFs more challenging. NeRF2NeRF[14] is the first work that tried to solve registering NeRFs by a traditional optimization-based approach. However, it requires humanannotated keypoints for initialization, which limits its application in the real world where human annotations can be impractical. In view of the above-mentioned challenges, we focus on the study of the NeRF registration problem by answering the following two questions: 1) Can we register two or multiple NeRFs together where only pre-trained models are accessible? 2) How to register NeRFs without any human annotations and initializations?
We further use the following settings in our endeavor to answer the two challenging questions on NeRF registration: 1) Images are collected into different blocks and no images are associated with known absolute position information. 2) Multiple NeRF blocks are trained individually where ground truth camera poses in each block are in their local coordinate frames. 3) Only the trained NeRF models are accessible, and all training images are removed and therefore not available due to plausible privacy-preserving issues or disk limitations. We emphasize that NeRF registration is a challenging task, and we focus more on the object-centric scenes in this paper. See Fig. 1 for the illustration of our task setting and dataset construction.
To solve the NeRF registration problem, we first utilize an occupancy grid along with each NeRF model to extract a voxel grid. The voxel grid is then fed into a 3D Feature Pyramid Network (FPN)[22] to extract features. The resulting voxel feature grids are further processed by a transformer module. In the transformer network, we first adopt a self-attention layer to enhance the intra-feature representations within each voxel feature grid. We further utilize a cross-attention layer to learn the inter-feature relations between the source feature grid and the target feature grid. Finally, we use an attention head to decode the source features and target features into correspondences and confidence scores. Unlike SOTA point clouds registration methods [16, 43], we utilize NeRF as geometric supervision and thus do not rely on pre-computed overlapping scores to mask correspondences outside the overlapping areas.
The main contributions of our work are:
· A dataset for registering multiple NeRF blocks, which is created by rendering 1,700+3D objects that are downloaded from the Objaverse dataset.
· A novel network for registering NeRF blocks which do not rely on any human annotation and initializations.
$\bullet$ Exhaustive experiments to show the accuracy and gen-
eralization ability of our method.
To the best of our knowledge, this is the first work on registering NeRFs without a) any initializations from keypoints or other registration methods and b) precomputed ground-truth overlapping labels.
## $2. \textbf{Related Work}$
Neural Radiance Fields. The differentiable volume rendering technique makes neural networks suitable for encoding scene representations. Vanilla NeRF[27] needs to take days to train each scene. NSVF[25] encodes scenes explicitly in a sparse voxel grid with online pruning, which improves the training and rendering efficiency by a large margin. PlenOctree[44] adopted the voxel representation, and further decouple the radiance field from view directions by spherical harmonics. To further accelerate the network training, PlenOctree fnetunes the voxel grid by directly optimizing the stored features without accessing any neural networks. Plenoxels [13] optimizes voxel features without any neural network for initialization. The training speed is accelerated with a specifically designed CUDA optimizer. InstantNGP [28] encodes voxel features implicitly into a multi-resolution hash grid. TensoRF [4] decomposes 3D voxel into low-rank vector-matrix multiplications. Its training speed is comparable to InstantNGP even without a CUDA implementation.
Block-NeRF[35] partition the collected images into different street blocks, each block is trained individually along with jointly optimizing camera poses [21,5]. The final images can be synthesized by fusing from multiple nearby NeRF blocks. Mega-NeRF[38] also adopted the same divide-and-conquer strategy as BlockNeRF, but focus more on aerial images and exploiting the sparse network structures. To train their network, Block-NeRF obtained camera poses by fusing data from different sensors, while MegaNeRF used Structure from Motion (SfM) [24,9,7,8] tools to recover camera poses. Both of them assume camera poses are in the same coordinate frames.
Point Cloud Registration. The Iterative Closest Point (ICP)[2,6] algorithm has been widely applied to the industry community and research community for years. Given a rough initialization, ICP tries to align the source point cloud to the target point cloud. The global point cloud registration methods can align point clouds without initialization by extracting geometric features such as the Fast Point Feature Histograms (FPFH)[31]. To solve the long-time issue of the global registration method in evaluating candidate models within RANSAC[12], Zhou and Park[46] proposed a fast approach that does not need to evaluate the candidate models at each iteration. Deep point cloud registration methods also gained much attention nowadays. Deep Closest Point [42] is a learned variant of the classical ICP. Inspired by SuperGlue[33], which is a deep learning method for matching
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
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
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.12249755859375,
0.4697265625,
0.121826171875,
0.4697265625,
0.1363525390625,
0.07843017578125,
0.136962890625
],
[
0.5009765625,
0.12371826171875,
0.8896484375,
0.12371826171875,
0.8896484375,
0.136962890625,
0.5009765625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.77685546875,
0.1383056640625,
0.77685546875,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1685791015625,
0.6396484375,
0.1685791015625,
0.6396484375,
0.18310546875,
0.50244140625,
0.18310546875
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.07843017578125,
0.1976318359375,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.07843017578125,
0.2108154296875
],
[
0.50244140625,
0.1956787109375,
0.89208984375,
0.1956787109375,
0.89208984375,
0.208984375,
0.50244140625,
0.208984375
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.2108154296875,
0.8896484375,
0.2108154296875,
0.8896484375,
0.2247314453125,
0.5009765625,
0.2247314453125
],
[
0.08087158203125,
0.2298583984375,
0.4697265625,
0.2298583984375,
0.4697265625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50244140625,
0.2259521484375,
0.89208984375,
0.2259521484375,
0.89208984375,
0.240478515625,
0.50244140625,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.2412109375,
0.890625,
0.2412109375,
0.890625,
0.255126953125,
0.50244140625,
0.255126953125
],
[
0.08087158203125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.5009765625,
0.255615234375,
0.890625,
0.255615234375,
0.890625,
0.270751953125,
0.5009765625,
0.270751953125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.272705078125,
0.89208984375,
0.272705078125,
0.89208984375,
0.285888671875,
0.50244140625,
0.285888671875
],
[
0.080078125,
0.28857421875,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.080078125,
0.3017578125
],
[
0.50244140625,
0.287353515625,
0.89208984375,
0.287353515625,
0.89208984375,
0.300537109375,
0.50244140625,
0.300537109375
],
[
0.080078125,
0.302978515625,
0.31787109375,
0.302978515625,
0.31787109375,
0.31640625,
0.080078125,
0.31640625
],
[
0.50341796875,
0.3017578125,
0.890625,
0.3017578125,
0.890625,
0.31494140625,
0.50341796875,
0.31494140625
],
[
0.098876953125,
0.318115234375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.098876953125,
0.33203125
],
[
0.50244140625,
0.317626953125,
0.890625,
0.317626953125,
0.890625,
0.330810546875,
0.50244140625,
0.330810546875
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.33203125,
0.890625,
0.33203125,
0.890625,
0.34521484375,
0.50341796875,
0.34521484375
],
[
0.08087158203125,
0.35107421875,
0.468017578125,
0.35107421875,
0.468017578125,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.5009765625,
0.34521484375,
0.890625,
0.3466796875,
0.890625,
0.361083984375,
0.5009765625,
0.35986328125
],
[
0.07843017578125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.07843017578125,
0.37744140625
],
[
0.50341796875,
0.3623046875,
0.89208984375,
0.3623046875,
0.89208984375,
0.375732421875,
0.50341796875,
0.375732421875
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.080078125,
0.392578125
],
[
0.5009765625,
0.376953125,
0.89208984375,
0.37744140625,
0.89208984375,
0.391357421875,
0.5009765625,
0.39013671875
],
[
0.080078125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.392578125,
0.890625,
0.392578125,
0.890625,
0.406005859375,
0.50244140625,
0.406005859375
],
[
0.080078125,
0.4091796875,
0.46728515625,
0.4091796875,
0.46728515625,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50244140625,
0.4072265625,
0.890625,
0.4072265625,
0.890625,
0.42041015625,
0.50244140625,
0.42041015625
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.449462890625,
0.50341796875,
0.449462890625
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50341796875,
0.45263671875,
0.8935546875,
0.45263671875,
0.8935546875,
0.466552734375,
0.50341796875,
0.466552734375
],
[
0.08087158203125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50244140625,
0.466552734375,
0.6611328125,
0.46728515625,
0.6611328125,
0.481689453125,
0.50244140625,
0.48095703125
],
[
0.080078125,
0.48486328125,
0.46728515625,
0.48486328125,
0.46728515625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.521484375,
0.484130859375,
0.89208984375,
0.484130859375,
0.89208984375,
0.49755859375,
0.521484375,
0.49755859375
],
[
0.080078125,
0.50048828125,
0.410888671875,
0.50048828125,
0.410888671875,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.49755859375,
0.89208984375,
0.5,
0.89208984375,
0.51318359375,
0.50244140625,
0.51123046875
],
[
0.098876953125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.098876953125,
0.52783203125
],
[
0.50341796875,
0.5146484375,
0.890625,
0.5146484375,
0.890625,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.07843017578125,
0.52978515625,
0.4697265625,
0.52880859375,
0.4697265625,
0.54345703125,
0.07843017578125,
0.54443359375
],
[
0.5009765625,
0.52783203125,
0.89208984375,
0.52880859375,
0.89208984375,
0.54345703125,
0.5009765625,
0.54248046875
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.5009765625,
0.54248046875,
0.89208984375,
0.54443359375,
0.89208984375,
0.5595703125,
0.5009765625,
0.556640625
],
[
0.08087158203125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.08087158203125,
0.57373046875
],
[
0.50244140625,
0.5595703125,
0.89208984375,
0.5595703125,
0.89208984375,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.57470703125,
0.89208984375,
0.57470703125,
0.89208984375,
0.58837890625,
0.50244140625,
0.58837890625
],
[
0.08087158203125,
0.59228515625,
0.46728515625,
0.59228515625,
0.46728515625,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.50244140625,
0.59033203125,
0.89208984375,
0.59033203125,
0.89208984375,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.5009765625,
0.60400390625,
0.89208984375,
0.60498046875,
0.89208984375,
0.619140625,
0.5009765625,
0.61865234375
],
[
0.080078125,
0.62060546875,
0.471435546875,
0.62060546875,
0.471435546875,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.5009765625,
0.61865234375,
0.89208984375,
0.619140625,
0.89208984375,
0.6337890625,
0.5009765625,
0.63330078125
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50244140625,
0.63525390625,
0.8935546875,
0.63525390625,
0.8935546875,
0.6484375,
0.50244140625,
0.6484375
],
[
0.080078125,
0.6513671875,
0.46728515625,
0.6513671875,
0.46728515625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.6494140625,
0.76953125,
0.6494140625,
0.76953125,
0.66357421875,
0.50244140625,
0.66357421875
],
[
0.080078125,
0.66552734375,
0.465576171875,
0.66552734375,
0.465576171875,
0.6787109375,
0.080078125,
0.6787109375
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.07843017578125,
0.68115234375,
0.468017578125,
0.6796875,
0.468017578125,
0.693359375,
0.07843017578125,
0.6943359375
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08087158203125,
0.69775390625,
0.468017578125,
0.69775390625,
0.468017578125,
0.70849609375,
0.08087158203125,
0.70849609375
],
[
0.080078125,
0.7109375,
0.4697265625,
0.7109375,
0.4697265625,
0.724609375,
0.080078125,
0.724609375
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.72607421875,
0.468017578125,
0.72607421875,
0.468017578125,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.07843017578125,
0.74072265625,
0.44287109375,
0.7412109375,
0.44287109375,
0.755859375,
0.07843017578125,
0.7548828125
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.50244140625,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.101318359375,
0.75830078125,
0.36279296875,
0.75830078125,
0.36279296875,
0.76806640625,
0.101318359375,
0.76806640625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.098876953125,
0.77783203125,
0.468017578125,
0.77783203125,
0.468017578125,
0.79248046875,
0.098876953125,
0.79248046875
],
[
0.50341796875,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.111083984375,
0.79345703125,
0.4697265625,
0.79345703125,
0.4697265625,
0.806640625,
0.111083984375,
0.806640625
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.111083984375,
0.8095703125,
0.37744140625,
0.8095703125,
0.37744140625,
0.82275390625,
0.111083984375,
0.82275390625
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.1004638671875,
0.83349609375,
0.468017578125,
0.83349609375,
0.468017578125,
0.8466796875,
0.1004638671875,
0.8466796875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.111083984375,
0.8486328125,
0.4599609375,
0.8486328125,
0.4599609375,
0.8623046875,
0.111083984375,
0.8623046875
],
[
0.5009765625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5009765625,
0.87060546875
],
[
0.098876953125,
0.869140625,
0.468017578125,
0.87060546875,
0.468017578125,
0.88525390625,
0.098876953125,
0.8837890625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.111083984375,
0.88720703125,
0.330810546875,
0.88720703125,
0.330810546875,
0.90087890625,
0.111083984375,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"However, NeRF registration is under-explored since exist-",
"To the best of our knowledge, this is the first work on reg-",
"ing works focus mostly on point cloud registration. Unlike",
"istering NeRFs without a) any initializations from key-",
"point clouds that are simple explicit representation, NeRF",
"points or other registration methods and b) precom-",
"encodes scenes implicitly, which makes registering multi-",
"puted ground-truth overlapping labels.",
"ple NeRFs more challenging. NeRF2NeRF [14] is the first",
"work that tried to solve registering NeRFs by a traditional",
"2. Related Work",
"optimization-based approach. However, it requires human-",
"annotated keypoints for initialization, which limits its ap-",
"Neural Radiance Fields. The differentiable volume ren-",
"plication in the real world where human annotations can be",
"dering technique makes neural networks suitable for encod-",
"impractical.Inviewoftheabove-mentionedchallenges,we",
"ing scene representations. Vanilla NeRF [27] needs to take",
"focus on the study of the NeRF registration problem by an-",
"days to train each scene. NSVF [25] encodes scenes ex-",
"sweringthefollowingtwoquestions:1)Canweregistertwo",
"plicitly in a sparse voxel grid with online pruning, which",
"or multiple NeRFs together where only pre-trained models",
"improves the training and rendering efficiency by a large",
"are accessible? 2) How to register NeRFs without any hu-",
"margin. PlenOctree [44] adopted the voxel representation,",
"man annotations and initializations?",
"and further decouple the radiance field from view direc-",
"We further use the following settings in our endeavor",
"tions by spherical harmonics. To further accelerate the net-",
"to answer the two challenging questions on NeRF registra-",
"work training, PlenOctree finetunes the voxel grid by di-",
"tion:DImages are collectedintodifferentblocks andno",
"rectly optimizing the stored features without accessing any",
"images are associated with known absolute position infor-",
"neural networks. Plenoxels [13] optimizes voxel features",
"mation: 2) Multiple NeRF blocks are trained individually",
"without any neural network for initialization.The training",
"where ground truth camera poses in each block are in their",
"speed is accelerated with a specifically designed CUDA op-",
"local coordinate frames. 3) Only the trained NeRF mod-",
"timizer. InstantNGP [28] encodes voxel features implicitly",
"els are accessible, and all training images are removed and",
"into a multi-resolution hash grid. TensoRF [4] decomposes",
"therefore not available due to plausible privacy-preserving",
"3D voxel into low-rank vector-matrix multiplications.Its",
"issues or disk limitations. We emphasize that NeRF reg-",
"training speed is comparable to InstantNGP even without a",
"istration isa challengingtask,and wefocusmoreonthe",
"CUDA implementation.",
"object-centric scenes in this paper. See Fig. 1 for the illus-",
"Block-NeRF [35] partition the collected images into dif-",
"tration of our task setting and dataset construction.",
"ferent street blocks, each block is trained individually along",
"To solve the NeRF registration problem, we first utilize",
"with jointly optimizing camera poses [21, 5]. The final im-",
"an occupancy grid along with each NeRF model to extract",
"ages can be synthesized by fusing from multiple nearby",
"a voxel grid. The voxel grid is then fed into a 3D Feature",
"NeRF blocks. Mega-NeRF [38] also adopted the same",
"Pyramid Network (FPN) [22] to extract features. The re-",
"divide-and-conquer strategy as BlockNeRF, but focus more",
"sulting voxel feature grids are further processed by a trans-",
"on aerial images and exploiting the sparse network struc-",
"formermodule.Inthetransformernetwork,wefirstadopt",
"tures. To train their network, Block-NeRF obtained camera",
"a self-attention layer to enhance the intra-feature represen-",
"poses by fusing data from different sensors, while Mega-",
"tations within each voxel feature grid. We further utilize a",
"NeRF used Structure from Motion (SfM) [24, 9, 7, 8] tools",
"cross-attention layer to learn the inter-feature relations be-",
"to recover camera poses. Both of them assume camera",
"tween the source feature grid and the target feature grid.",
"poses are in the same coordinate frames.",
"Finally, we use an attention head to decode the source fea-",
"Point Cloud Registration. The Iterative Closest Point",
"tures and target features into correspondences and confi-",
"(ICP) [2, 6] algorithm has been widely applied to the indus-",
"dence scores.Unlike SOTA point cloudsregistrationmeth-",
"ods [16, 43], we utilize NeRF as geometric supervision",
"try community and research community for years. Given a",
"and thus do not rely on pre-computed overlapping scores",
"rough initialization, ICP tries to align the source point cloud",
"to mask correspondences outside the overlapping areas.",
"to thetargetpoint cloud.Theglobal pointcloudregistration",
"methods can align point clouds without initialization by ex-",
"Themain contributions of our work are:",
"tracting geometric features such as the Fast Point Feature",
": A dataset for registering multiple NeRF blocks, which",
"Histograms (FPFH) [31]. To solve the long-time issue of the",
"is created by rendering 1, 700+ 3D objects that are",
"global registration method in evaluating candidate models",
"downloaded from the Objaverse dataset.",
"within RANSAC [12], Zhou and Park [46] proposed a fast",
"approach that does not need to evaluate the candidate mod-",
"·A novel network for registering NeRFblocks which do",
"els at each iteration. Deep point cloud registration methods",
"not rely on any human annotation and initializations.",
"also gained much attention nowadays. Deep Closest Point",
"· Exhaustive experiments to show the accuracy and gen-",
"[42] is a learned variant of the classical ICP. Inspired by Su-",
"eralization ability of our method.",
"perGlue [33], which is a deep learning method for matching"
] |
{
"arxiv_id": "2308.09386",
"language": "en",
"page_idx": 1
}
|
||
2308.09386-en-2
|
2308.09386-en-2.jpg
|
2D image correspondences, Predator[16] and REGTR[43] adopted the self-attention and cross-attention mechanisms from SuperGlue to learn the correlation for pairwise lowoverlapping point clouds. The ground-truth overlapping scores are computed from dense point clouds and used to mask out the correspondences outside the overlapping regions. We also follow the previous attention mechanisms, but do not rely on the pre-computed overlapping labels.
NeRF Registration. NeRF2NeRF[14] is the frst work that tries to register multiple NeRFs. The initial transformation is estimated from human-annotated keypoints, and then refined by the surface fields from NeRF. To reduce the number of useless samples, NeRF2NeRF adopts Metropolis-Hasting sampling to maintain an active set. The whole framework is based on traditional optimization methods and needs human interaction. ZeroNeRF [29] claims it can register NeRF without overlap. However, it still requires a global registration method for initialization. Unlike the previous works, which rely heavily on traditional optimization methods, we try to register multiple NeRF blocks by learning methods without human interaction.
## 3.0ur Method
Fig. 2 shows an illustration of our framework. Our network takes a source NeRF model and a target NeRF model as input, and outputs correspondences in the source NeRF and target NeRF. We first train multiple NeRF blocks in different coordinate frames. For each NeRF model, we associate it with an occupancy voxel grid, where each voxel indicates whether it is occupied or not. After training, we extract a 3D voxel grid for each NeRF model and then feed it into a 3D CNN backbone to extract features. Subsequently, we use a transformer with self-attention and cross-attention layers to learn the relations between the pairwise feature grids. We then adopt a decoder to decode the resulting features $\mathcal{F}_\mathrm{source},\mathcal{F}_\mathrm{target}$ into correspondences $\{\mathbf{X}_\mathrm{source},\mathbf{X}_\mathrm{target}\}$ and the corresponding confidence scores $\{\mathbf{S}_{\mathrm{source}},\mathbf{S}_{\mathrm{target}}\}.$ Finally, the relative transformation can be solved by the camera frame to world frame. To obtain the radiance fields,weighted Kabsch-Umeyama algorithm [39] from the cor- we form $N$ viewing rays r= o- $+t\mathbf{d}$,where $\mathbf{d}=\frac{\mathbf{X}-\mathbf{t}}{\|\mathbf{X}-\mathbf{t}\|}$respondenc
respondences.
### $3. 1. \textbf{Background}$
Neural Radiance Fields. Neural Radiance Fields (NeRF) aims at rendering photo-realistic images from a new view point. For a 3D point X, the density field $\sigma_t$ of X is defined as the differential probability of a ray r = ( o, d) hitting a $particle$, where o is the camera center, d is the view direction. The transmittance $\mathcal{T}(t)$ denotes the probability of ray without hitting any particles when traveling a distance $t$,and the discrete form of $\mathcal{T}(t)$ is:
(1)
$$\mathcal{T}_n=\mathcal{T}(0\to t_n)=\exp\big(\sum_{k=1}^{n-1}-\sigma_k\delta_k\big).$$
Given a set of points $\{\mathbf{X}_n=\mathbf{o}+t_n\mathbf{d}\mid n\in[0,K]\}$, NeRF predicts the view-independent volume density $\sigma_n$ and view-dependent radiance field by:
$$\begin{array}{rcl}\sigma_n,\mathbf{e}_n=\mathbf{F}(\mathbf{X};\mathbf{\Theta}),\\\mathbf{c}_n=\mathbf{F}(\mathbf{r},\mathbf{e};\mathbf{\Theta}),\end{array}$$
(2)
where e$_n$ is an embedding vector and $\Theta$ denotes the network parameters. The final color of an image pixel can be rendered by:
$$\mathbf{C}(t_{N+1})=\sum_{n=1}^N\mathcal{T}_n\cdot\left(1-\exp(-\sigma_n\delta_n)\right)\cdot\mathbf{c}_n.$$
(3)
We recommend interested readers to refer to[27,34] for the
detailed derivation.
3.2. Querying Radiance Fields from NeRF
To extract features for the latter learning modules, we first construct a 3D volume from NeRF. Specifically, we assume each NeRF is trained within a bounding box with a 3D voxel grid of resolution $[x_{\mathrm{ref}},y_{\mathrm{res}},z_{\mathrm{res}}].$ We then obtain the point locations $\{\mathbf{X}\}$ of voxel centers. We also obtain a binary occupancy mask $\mathbf{M}_{\mathrm{occ}}$ from the occupancy grid, where voxels that are not occupied denote empty space and thus are ignored. The occupancy grid is the acceleration structure used in InstantNGP [28] to skip empty space in NeRF training. Each grid has a resolution of 128$^{3}$ that is centered around (0,0,0) and stores occupancy as a single bit. During ray marching, a sample point is skipped if the bit of a grid cell is low. We can query the density fields $\{\sigma\}$ and radiance felds $\{\mathbf{c}\}$ using Eq. (2) with the coordinates $\{\mathbf{X}\}.$ Since density felds can be noisy, we further obtain a density mask $\mathbf{M}_\mathrm{df}=\sigma>\sigma_{t}$ by setting a threshold $\sigma_t$ (we use $\sigma_t=0.7).$ We then obtain our mask as M$=\mathbf{M}_\mathrm{occ}\cap\mathbf{M}_\mathrm{df}.$
One issue with radiance fields is that they are viewdependent. However, the training views are different for each NeRF block. Suppose the NeRF is trained by $N$ views with camera poses $\mathbf{P}=[\mathbf{R}\mid\mathbf{t}]$ which project points from over all rays. The final color C can be obtained by volume rendering of Eq.(3). Furthermore, we compute the alpha compositing value by $\alpha=1-\exp(-\sigma\delta)$, where $\delta$ is a chosen small value. We associate each point in the voxel grid with $[\mathbf{X},\mathbf{C},\alpha]$ and feed the voxel grid $\mathbf{G} \in \mathbb{R} ^{x_{\mathrm{ref}}\times y_{\mathrm{res}}\times z_{\mathrm{res}}\times C}$ and a mask $\mathbf{M} \in \mathbb{R} ^{x_{\mathrm{ref}}\times y_{\mathrm{res}}\times z_{\mathrm{res}}\times 1}$ into a 3D CNN to extract backbone features, where $C=7$ with 3 channels from the point coordinate, three from the color channels, and one from the scalar $\alpha.$
3.3. Feature Extraction
Given the voxel grid, we adopt the feature pyramid network[22] to extract features. We use ResNet[15] as the
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.09027099609375,
0.468017578125,
0.10479736328125,
0.080078125,
0.1060791015625
],
[
0.521484375,
0.09027099609375,
0.88818359375,
0.09027099609375,
0.88818359375,
0.10479736328125,
0.521484375,
0.10479736328125
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5009765625,
0.1060791015625,
0.89208984375,
0.106689453125,
0.89208984375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.08087158203125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.72802734375,
0.12249755859375,
0.72802734375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.15283203125,
0.080078125,
0.15283203125
],
[
0.62744140625,
0.1427001953125,
0.76318359375,
0.14013671875,
0.76318359375,
0.1578369140625,
0.62744140625,
0.160400390625
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.64453125,
0.1634521484375,
0.76318359375,
0.1622314453125,
0.76318359375,
0.1767578125,
0.64453125,
0.177978515625
],
[
0.869140625,
0.1634521484375,
0.8935546875,
0.1634521484375,
0.8935546875,
0.1767578125,
0.869140625,
0.1767578125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.07843017578125,
0.1685791015625,
0.46728515625,
0.1685791015625,
0.46728515625,
0.1817626953125,
0.07843017578125,
0.1817626953125
],
[
0.50341796875,
0.1849365234375,
0.890625,
0.1849365234375,
0.890625,
0.1988525390625,
0.50341796875,
0.1988525390625
],
[
0.080078125,
0.1988525390625,
0.44287109375,
0.1988525390625,
0.44287109375,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.5009765625,
0.1988525390625,
0.89208984375,
0.199462890625,
0.89208984375,
0.2139892578125,
0.5009765625,
0.21337890625
],
[
0.080078125,
0.21533203125,
0.468017578125,
0.21533203125,
0.468017578125,
0.2291259765625,
0.080078125,
0.2291259765625
],
[
0.50244140625,
0.216552734375,
0.58837890625,
0.216552734375,
0.58837890625,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.080078125,
0.2310791015625,
0.46728515625,
0.2310791015625,
0.46728515625,
0.2442626953125,
0.080078125,
0.2442626953125
],
[
0.63134765625,
0.2342529296875,
0.65185546875,
0.2342529296875,
0.65185546875,
0.2474365234375,
0.63134765625,
0.2474365234375
],
[
0.080078125,
0.246826171875,
0.468017578125,
0.246826171875,
0.468017578125,
0.260009765625,
0.080078125,
0.260009765625
],
[
0.5439453125,
0.24560546875,
0.859375,
0.24560546875,
0.859375,
0.267578125,
0.5439453125,
0.267578125
],
[
0.86669921875,
0.25,
0.890625,
0.25,
0.890625,
0.261474609375,
0.86669921875,
0.261474609375
],
[
0.07843017578125,
0.261474609375,
0.46728515625,
0.261474609375,
0.46728515625,
0.274658203125,
0.07843017578125,
0.274658203125
],
[
0.62744140625,
0.26513671875,
0.65625,
0.26513671875,
0.65625,
0.276611328125,
0.62744140625,
0.276611328125
],
[
0.080078125,
0.276611328125,
0.4697265625,
0.276611328125,
0.4697265625,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.50244140625,
0.281005859375,
0.89208984375,
0.281005859375,
0.89208984375,
0.294189453125,
0.50244140625,
0.294189453125
],
[
0.07843017578125,
0.292236328125,
0.4697265625,
0.292236328125,
0.4697265625,
0.30615234375,
0.07843017578125,
0.30615234375
],
[
0.50244140625,
0.296630859375,
0.6298828125,
0.296630859375,
0.6298828125,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.080078125,
0.306884765625,
0.468017578125,
0.306884765625,
0.468017578125,
0.32080078125,
0.080078125,
0.32080078125
],
[
0.080078125,
0.322509765625,
0.4697265625,
0.322509765625,
0.4697265625,
0.3359375,
0.080078125,
0.3359375
],
[
0.50244140625,
0.317626953125,
0.826171875,
0.31640625,
0.826171875,
0.330810546875,
0.50244140625,
0.33203125
],
[
0.080078125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.35107421875,
0.080078125,
0.35107421875
],
[
0.521484375,
0.341552734375,
0.89208984375,
0.341552734375,
0.89208984375,
0.354736328125,
0.521484375,
0.354736328125
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.50244140625,
0.356689453125,
0.890625,
0.356689453125,
0.890625,
0.37060546875,
0.50244140625,
0.37060546875
],
[
0.08087158203125,
0.36865234375,
0.46728515625,
0.36865234375,
0.46728515625,
0.37890625,
0.08087158203125,
0.37890625
],
[
0.50341796875,
0.373779296875,
0.890625,
0.373779296875,
0.890625,
0.3837890625,
0.50341796875,
0.3837890625
],
[
0.080078125,
0.3818359375,
0.4697265625,
0.3818359375,
0.4697265625,
0.395751953125,
0.080078125,
0.395751953125
],
[
0.5009765625,
0.386474609375,
0.890625,
0.386474609375,
0.890625,
0.40087890625,
0.5009765625,
0.40087890625
],
[
0.07843017578125,
0.397705078125,
0.393798828125,
0.397705078125,
0.393798828125,
0.410888671875,
0.07843017578125,
0.410888671875
],
[
0.50244140625,
0.40283203125,
0.890625,
0.40087890625,
0.890625,
0.4140625,
0.50244140625,
0.416748046875
],
[
0.50244140625,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.080078125,
0.4248046875,
0.20751953125,
0.4248046875,
0.20751953125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.43310546875,
0.89208984375,
0.43310546875,
0.89208984375,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.098876953125,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.462158203125,
0.098876953125,
0.462158203125
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.464111328125,
0.4697265625,
0.464111328125,
0.4697265625,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50244140625,
0.462158203125,
0.89208984375,
0.462158203125,
0.89208984375,
0.47607421875,
0.50244140625,
0.47607421875
],
[
0.0767822265625,
0.479248046875,
0.468017578125,
0.477783203125,
0.468017578125,
0.492431640625,
0.0767822265625,
0.49365234375
],
[
0.50244140625,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.080078125,
0.494384765625,
0.468017578125,
0.494384765625,
0.468017578125,
0.50830078125,
0.080078125,
0.50830078125
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.51025390625,
0.46728515625,
0.51025390625,
0.46728515625,
0.5234375,
0.080078125,
0.5234375
],
[
0.50244140625,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50244140625,
0.521484375
],
[
0.07843017578125,
0.5244140625,
0.4697265625,
0.5234375,
0.4697265625,
0.53857421875,
0.07843017578125,
0.5390625
],
[
0.5009765625,
0.52294921875,
0.89208984375,
0.521484375,
0.89208984375,
0.5361328125,
0.5009765625,
0.537109375
],
[
0.080078125,
0.54052734375,
0.468017578125,
0.54052734375,
0.468017578125,
0.5537109375,
0.080078125,
0.5537109375
],
[
0.50244140625,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.080078125,
0.5546875,
0.4697265625,
0.5546875,
0.4697265625,
0.568359375,
0.080078125,
0.568359375
],
[
0.50244140625,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.50244140625,
0.56640625
],
[
0.07843017578125,
0.5693359375,
0.468017578125,
0.57080078125,
0.468017578125,
0.583984375,
0.07843017578125,
0.58349609375
],
[
0.50341796875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.08087158203125,
0.5869140625,
0.468017578125,
0.5869140625,
0.468017578125,
0.59765625,
0.08087158203125,
0.59765625
],
[
0.50244140625,
0.58349609375,
0.8857421875,
0.58349609375,
0.8857421875,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.52294921875,
0.599609375,
0.8896484375,
0.599609375,
0.8896484375,
0.60986328125,
0.52294921875,
0.60986328125
],
[
0.080078125,
0.6142578125,
0.46728515625,
0.6142578125,
0.46728515625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.07843017578125,
0.630859375,
0.468017578125,
0.630859375,
0.468017578125,
0.64599609375,
0.07843017578125,
0.64599609375
],
[
0.50244140625,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.07843017578125,
0.64453125,
0.468017578125,
0.64599609375,
0.468017578125,
0.66015625,
0.07843017578125,
0.6591796875
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64306640625,
0.89208984375,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.08087158203125,
0.662109375,
0.468017578125,
0.662109375,
0.468017578125,
0.67236328125,
0.08087158203125,
0.67236328125
],
[
0.50244140625,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.50244140625,
0.671875
],
[
0.5,
0.66845703125,
0.890625,
0.671875,
0.890625,
0.6923828125,
0.5,
0.689453125
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.080078125,
0.6923828125,
0.1732177734375,
0.6923828125,
0.1732177734375,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.7119140625,
0.20751953125,
0.7119140625,
0.20751953125,
0.7265625,
0.080078125,
0.7265625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.07843017578125,
0.7666015625,
0.468017578125,
0.76513671875,
0.468017578125,
0.77978515625,
0.07843017578125,
0.78076171875
],
[
0.50244140625,
0.7666015625,
0.890625,
0.76220703125,
0.890625,
0.7763671875,
0.50244140625,
0.78076171875
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81103515625,
0.080078125,
0.81103515625
],
[
0.50244140625,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.8125,
0.471435546875,
0.8125,
0.471435546875,
0.82568359375,
0.080078125,
0.82568359375
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.82763671875,
0.4697265625,
0.82763671875,
0.4697265625,
0.84130859375,
0.080078125,
0.84130859375
],
[
0.50244140625,
0.8271484375,
0.62353515625,
0.8271484375,
0.62353515625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.8447265625,
0.306396484375,
0.8447265625,
0.306396484375,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50244140625,
0.84765625,
0.68115234375,
0.84765625,
0.68115234375,
0.8623046875,
0.50244140625,
0.8623046875
],
[
0.31787109375,
0.86279296875,
0.343017578125,
0.86279296875,
0.343017578125,
0.87548828125,
0.31787109375,
0.87548828125
],
[
0.1446533203125,
0.87548828125,
0.330078125,
0.875,
0.330078125,
0.8896484375,
0.1446533203125,
0.89013671875
],
[
0.339111328125,
0.876953125,
0.40283203125,
0.876953125,
0.40283203125,
0.8916015625,
0.339111328125,
0.8916015625
],
[
0.44677734375,
0.876953125,
0.4697265625,
0.876953125,
0.4697265625,
0.89013671875,
0.44677734375,
0.89013671875
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.50244140625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.50244140625,
0.90087890625
],
[
0.31787109375,
0.89306640625,
0.343017578125,
0.89306640625,
0.343017578125,
0.90478515625,
0.31787109375,
0.90478515625
]
] |
[
"2D image correspondences, Predator [16] and REGTR[43]",
"Given a set of points [xn = o + tnd I n E [o, Kl?",
"adopted the self-attention and cross-attention mechanisms",
"NeRF predicts the view-independent volume density on and",
"from SuperGlueto learnthe correlation forpairwiselow-",
"view-dependent radiance field by:",
"overlapping point clouds.The ground-truth overlapping",
"On,en = F(X;0),",
"scores are computed from dense point clouds and used to",
"Cn = F(r,e; ),",
"(2)",
"gions. We also follow the previous attention mechanisms,",
"mask out the correspondences outside the overlapping re-",
"where en is an embedding vector and denotes the net-",
"but do not rely on the pre-computed overlapping labels.",
"work parameters. The final color of an image pixel can be",
"NeRF Registration. NeRF2NeRF [14] is the first work",
"rendered by:",
"that tries to register multiple NeRFs. The initial trans-",
"N",
"formation is estimated from human-annotated keypoints,",
"C(tN+1)=Tn (1-exp(-onon) ·Cn.",
"(3)",
"and then refined by the surface fields from NeRF. To re-",
"n=1",
"duce the number of useless samples, NeRF2NeRF adopts",
"We recommend interested readers to refer to [27, 34]J for the",
"Metropolis-Hasting sampling to maintain an active set. The",
"detailed derivation.",
"whole framework is based on traditional optimization meth-",
"ods and needs human interaction. ZeroNeRF [29] claims",
"3.2. Querying Radiance Fields from NeRF",
"it can register NeRF without overlap. However, it still re-",
"To extract features for the latter learning modules, we",
"quires a global registration method for initialization. Unlike",
"first construct a 3D volume from NeRF. Specifically, we as-",
"the previous works,which rely heavily on traditional opti-",
"sumeeachNeRFistrained withinaboundingbox witha3D",
"mization methods, we try to register multiple NeRF blocks",
"voxel grid of resolution [Cref, Yres, zres]. We then obtain the",
"by learning methods without human interaction.",
"point locations [X} of voxel centers. We also obtain a bi-",
"nary occupancy mask Moce from the occupancy grid, where",
"3. Our Method",
"voxels that are not occupied denote empty space and thus",
"Fig. 2 shows an illustration of our framework. Our net-",
"are ignored. The occupancy grid is the acceleration struc-",
"work takes a source NeRF model and a target NeRF model",
"ture used in InstantNGP [28] to skip empty space in NeRF",
"as input, and outputs correspondences in the source NeRF",
"training. Each grid has a resolution of 1283 that is centered",
"and target NeRF. We first train multiple NeRF blocks in dif-",
"around (O, 0, O) and stores occupancy as a single bit. Dur-",
"ferent coordinate frames. For each NeRF model, we asso-",
"ing ray marching, a sample point is skipped if the bit of a",
"ciate it with an occupancy voxel grid, where each voxel in-",
"grid cell is low. We can query the density fields fo and",
"dicates whether it is occupied or not. After training, we ex-",
"radiance fields (c} using Eq. (2) with the coordinates {X)-",
"tract a 3D voxel grid for each NeRF model and then feed it",
"Since density fields can be noisy, we further obtain a den-",
"into a 3D CNN backbone to extract features.Subsequently,",
"sity mask Maf = > ot by setting a threshold ot (we use",
"weuseatransformerwith self-attentionand cross-attention",
"ot = 0.7). We then obtain our mask as M = Moce n Maf.",
"layers to learn the relations between the pairwise feature",
"One issue with radiance fields is that they are view-",
"grids. We then adopt a decoder to decode the resulting fea-",
"dependent. However, the training views are different for",
"tures Fsource, Farget into correspondences [Xsource, Xtarget]",
"each NeRF block. Suppose the NeRF is trained by N views",
"and the corresponding confidence scores [Ssource, Starget].",
"with camera poses P = [R t] which project points from",
"Finally, the relative transformation can be solved by the",
"camera frame to world frame. To obtain the radiance fields",
"we form N viewing rays r = o + td, where d = x-t",
"weighted Kabsch-Umeyama algorithm [39] from the cor-",
"respondences.",
"for each query point and average the queried radiance fields",
"over all rays. The final color C can be obtained by volume",
"3.1.Background",
"rendering of Eq. (3). Furthermore, we compute the alpha",
"Neural Radiance Fields. Neural Radiance Fields (NeRF)",
"compositingvaluebya=1-exp(-oo).whereisacho-",
"aims at rendering photo-realistic images from a new view",
"sen small value. We associate each point in the voxel grid",
"point. For a 3D point X, the density field ot of X is defined",
"with [X, C, a] and feed the voxel grid G E IRFrer × Yres X zren ×C",
"as the differential probability of a ray r - (o, d) hitting a",
"and a mask M E IRFrerx yres X zres x1 into a 3D CNN to ex-",
"particle, where o is the camera center, d is the view direc-",
"tract backbone features, where C = 7 with 3 channels from",
"tion. The transmittance T(t) denotes the probability of",
"the point coordinate, three from the color channels, and one",
"ray without hitting any particles when traveling a distance",
"from the scalar α.",
"t, and the discrete form of T (t) is:",
"3.3. Feature Extraction",
"n-1",
"Tn = T(0 -→ tn) = exp(",
"-0k0k)",
"(1)",
"Given the voxel grid, we adopt the feature pyramid net-",
"work [22] to extract features. We use ResNet [15] as the",
"k=1"
] |
{
"arxiv_id": "2308.09386",
"language": "en",
"page_idx": 2
}
|
||
2308.09386-en-3
|
2308.09386-en-3.jpg
|
<FigureHere>
Figure 2: Network architecture of our method. The pipeline of our method is: 1) We fırst extract the pairwise voxel grid $\mathbf{G} \in \mathbb{R} ^{x_{\mathrm{ref}}\times y_{\mathrm{res}}\times z_{\mathrm{res}}\times C}$ and a binary mask $\mathbf{M} \in \mathbb{R} _{\mathrm{ref}}^x\times y_{\mathrm{res}}\times z_{\mathrm{res}}\times 1$ from the source NeRF and the target NeRF. 2) The voxel grid G and a binary mask M are fed into the 3D feature pyramid network to extract voxel features. 3) The extracted voxel grid features are downsampled to a 2-dimensional tensor $\mathcal{F}\in\mathbb{R}^{N\times C^{\prime}}$ by their spherical neighborhood. 4) The resulting source features $\mathcal{F}_\text{source and target features }\mathcal{F}_\text{target are strengthened by a transformer, where a self-attention layer is used to enhance }$ the intra-contextual relations, and a cross-attention layer is used to learn the inter-contextual relations. 5) Finally, we use a single-head attention layer to decode the features into correspondences and their corresponding confidence scores.
layers. We follow the same intuition of Predator [16] and REGTR[43], where a self-attention layer is applied to both the source feature $\mathcal{F}_\text{source and the target feature }\mathcal{F}_\text{target to en-}$ hance the intra-contextual relations, and a cross-attention layer is applied to both $\mathcal{F}_\mathrm{source}$ and $\mathcal{F}_\mathrm{target}$ to learn the intercontextual relations. We follow the classical transformer architecture [41, 20, 43] with input to be voxel features.
## Multi-Head Attention. The multi-head attention operation in each layer is defined as:
$$\mathrm{MH}(\mathbf{Q},\mathbf{K},\mathbf{V})=\mathrm{concat}(\mathrm{head}_1,\cdots,\mathrm{head}_h)\mathbf{W}^O,$$
(4)
where head$_i=$Attention$(\mathbf{Q}\mathbf{W}_i^Q,\mathbf{K}\mathbf{W}_i^K,\mathbf{V}\mathbf{W}_i^V).$The at-
tention function is adopted as the scaled dot-product:
feature backbone, where all the 2D modules are replaced by their corresponding 3D parts. The feature pyramid network enables the learning of high-level semantic features at a multi-scale and thus is suitable to extract the voxel grid features in our task settings. We note the difference from the original feature pyramid network which utilizes different scale features for object detection, we adopt only the features output from the last layer. Since the dimension of the feature grid from the last layer can be different from the original voxel grid, we rescale it to the size of the original voxel grid G and resulted in a voxel feature grid $\mathbf{G}_{f}\in\mathbb{R}^{x_{\mathrm{ref}}\times y_{\mathrm{res}}\times z_{\mathrm{res}}\times C^{\prime}}.$
We cannot use $\mathbf{G}_f$ as the input to our transformer since the voxel feature grid can contain too many points to enable the transformer to run on a single GPU. To solve this issue, we iteratively downsample G$_f$ by the spherical neighborhood[36] as done in KPConv[37] to obtain the downsampled voxel points $\hat{\mathbf{X}}.$ The downsample iteration is terminated when the total number of occupied voxels in the current sampled voxel feature grid is less than 1.5K. Lastly, we reshape the occupied voxel features into a tensor $\mathcal{F}\in\mathbb{R}^{N\times\hat{C}^{\prime}}.$ The weights of the feature extraction network are shared across the source and the target NeRFs. We denote the downsampled voxel points and extracted features of the source and target NeRFs as: $[\hat{\mathbf{X}}_{\mathrm{source}},\mathcal{F}_{\mathrm{source}}]$ and $[\hat{\mathbf{X}}_{\mathrm{target}},\mathcal{F}_{\mathrm{target}}]$, respectively.
$$\mathrm{Attention}(\mathbf{Q},\mathbf{K},\mathbf{V})=\mathrm{softmax}(\frac{\mathbf{QK}^{\top}}{\sqrt{d_{k}}})\mathbf{V}.$$
(5)
In the self-attention layers, $\mathbb{Q}=\mathbb{K}=\mathbb{V}$ represents the same feature tensor in each block. In the cross-attention layers, the keys, and the values are the feature tensors from the other block. The self-attention mechanism enables the network to learn the relationship inside the same feature points, while the cross-attention mechanism enables the communication of the different feature points.
Decoder. After encoding features by transformer, we further adopt a single-head attention layer to predict the corresponding point locations $\tilde{\mathbf{X}}_{\mathrm{source~and~confidence~scores}}$ $\tilde{\mathbf{S}}_{\mathrm{source}}$ of the source voxel points $\hat{\mathbf{X}}_{\mathrm{source}}$ in the target NeRF's coordinate frame. Similarly, we also predict the corresponding point locations $\tilde{\mathbf{X}}_{\mathrm{target}}$ and confidence scores
## $3. 4. \textbf{Transformer}$
We then feed the resulting feature $\mathcal{F}_\mathrm{source}$ and $\mathcal{F}_\mathrm{target}$ into a $L$-layer transformer with self-attention and cross-attention
|
[
[
0.525390625,
0.11553955078125,
0.6787109375,
0.11553955078125,
0.6787109375,
0.2412109375,
0.525390625,
0.2412109375
],
[
0.10540771484375,
0.173583984375,
0.1192626953125,
0.173583984375,
0.1192626953125,
0.177978515625,
0.10540771484375,
0.177978515625
],
[
0.12744140625,
0.169189453125,
0.1666259765625,
0.1717529296875,
0.1666259765625,
0.179931640625,
0.12744140625,
0.177978515625
],
[
0.705078125,
0.173583984375,
0.75732421875,
0.173583984375,
0.75732421875,
0.1849365234375,
0.705078125,
0.1849365234375
],
[
0.10455322265625,
0.271484375,
0.17236328125,
0.271484375,
0.17236328125,
0.28271484375,
0.10455322265625,
0.28271484375
],
[
0.2115478515625,
0.271484375,
0.268798828125,
0.271484375,
0.268798828125,
0.28271484375,
0.2115478515625,
0.28271484375
],
[
0.361083984375,
0.271484375,
0.404296875,
0.271484375,
0.404296875,
0.28271484375,
0.361083984375,
0.28271484375
],
[
0.473876953125,
0.271484375,
0.55126953125,
0.271484375,
0.55126953125,
0.28271484375,
0.473876953125,
0.28271484375
],
[
0.576171875,
0.272705078125,
0.6396484375,
0.272705078125,
0.6396484375,
0.28271484375,
0.576171875,
0.28271484375
],
[
0.08087158203125,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.304931640625,
0.08087158203125,
0.304931640625
],
[
0.080078125,
0.30615234375,
0.1854248046875,
0.302978515625,
0.1854248046875,
0.317626953125,
0.080078125,
0.32080078125
],
[
0.1748046875,
0.3017578125,
0.8935546875,
0.304931640625,
0.8935546875,
0.322509765625,
0.1748046875,
0.3193359375
],
[
0.08087158203125,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.08087158203125,
0.335205078125
],
[
0.07843017578125,
0.3359375,
0.8935546875,
0.3359375,
0.8935546875,
0.354248046875,
0.07843017578125,
0.354248046875
],
[
0.080078125,
0.354736328125,
0.8935546875,
0.354736328125,
0.8935546875,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.08087158203125,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.3818359375,
0.08087158203125,
0.3818359375
],
[
0.080078125,
0.385009765625,
0.826171875,
0.385009765625,
0.826171875,
0.3984375,
0.080078125,
0.3984375
],
[
0.08087158203125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.50244140625,
0.4267578125,
0.89208984375,
0.4267578125,
0.89208984375,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.442626953125,
0.4697265625,
0.442626953125,
0.4697265625,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.080078125,
0.45703125,
0.471435546875,
0.45703125,
0.471435546875,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.5009765625,
0.458251953125,
0.8896484375,
0.458251953125,
0.8896484375,
0.4716796875,
0.5009765625,
0.4716796875
],
[
0.080078125,
0.472900390625,
0.471435546875,
0.472900390625,
0.471435546875,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.07843017578125,
0.4873046875,
0.4697265625,
0.486083984375,
0.4697265625,
0.50048828125,
0.07843017578125,
0.501953125
],
[
0.50244140625,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50244140625,
0.468017578125,
0.50244140625,
0.468017578125,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.08087158203125,
0.51953125,
0.465576171875,
0.51953125,
0.465576171875,
0.52978515625,
0.08087158203125,
0.52978515625
],
[
0.50244140625,
0.51806640625,
0.86767578125,
0.51806640625,
0.86767578125,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.08087158203125,
0.53515625,
0.465576171875,
0.53515625,
0.465576171875,
0.54541015625,
0.08087158203125,
0.54541015625
],
[
0.5009765625,
0.5390625,
0.890625,
0.54052734375,
0.890625,
0.5546875,
0.5009765625,
0.5537109375
],
[
0.08087158203125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.56005859375,
0.08087158203125,
0.56005859375
],
[
0.50244140625,
0.55615234375,
0.708984375,
0.55615234375,
0.708984375,
0.5693359375,
0.50244140625,
0.5693359375
],
[
0.08087158203125,
0.564453125,
0.46728515625,
0.564453125,
0.46728515625,
0.57470703125,
0.08087158203125,
0.57470703125
],
[
0.07843017578125,
0.5771484375,
0.468017578125,
0.57763671875,
0.468017578125,
0.59228515625,
0.07843017578125,
0.59130859375
],
[
0.52392578125,
0.58056640625,
0.890625,
0.58056640625,
0.890625,
0.59521484375,
0.52392578125,
0.59521484375
],
[
0.080078125,
0.59521484375,
0.2320556640625,
0.59130859375,
0.2320556640625,
0.60595703125,
0.080078125,
0.60986328125
],
[
0.09722900390625,
0.6123046875,
0.4697265625,
0.61279296875,
0.4697265625,
0.62744140625,
0.09722900390625,
0.626953125
],
[
0.50244140625,
0.60986328125,
0.8935546875,
0.60986328125,
0.8935546875,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.07843017578125,
0.62744140625,
0.4697265625,
0.62890625,
0.4697265625,
0.642578125,
0.07843017578125,
0.6416015625
],
[
0.50244140625,
0.62890625,
0.85107421875,
0.62890625,
0.85107421875,
0.642578125,
0.50244140625,
0.642578125
],
[
0.080078125,
0.64306640625,
0.4697265625,
0.64306640625,
0.4697265625,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.07843017578125,
0.65771484375,
0.468017578125,
0.6591796875,
0.468017578125,
0.673828125,
0.07843017578125,
0.67236328125
],
[
0.7685546875,
0.65380859375,
0.80810546875,
0.65380859375,
0.80810546875,
0.66943359375,
0.7685546875,
0.66943359375
],
[
0.5556640625,
0.66357421875,
0.76953125,
0.66357421875,
0.76953125,
0.6767578125,
0.5556640625,
0.6767578125
],
[
0.8095703125,
0.6640625,
0.84228515625,
0.6640625,
0.84228515625,
0.67529296875,
0.8095703125,
0.67529296875
],
[
0.87109375,
0.66650390625,
0.890625,
0.66650390625,
0.890625,
0.6767578125,
0.87109375,
0.6767578125
],
[
0.07843017578125,
0.6748046875,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.07843017578125,
0.68798828125
],
[
0.77685546875,
0.673828125,
0.80810546875,
0.673828125,
0.80810546875,
0.68603515625,
0.77685546875,
0.68603515625
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.521484375,
0.69775390625,
0.89208984375,
0.69775390625,
0.89208984375,
0.7109375,
0.521484375,
0.7109375
],
[
0.080078125,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.50244140625,
0.71337890625,
0.890625,
0.71337890625,
0.890625,
0.7265625,
0.50244140625,
0.7265625
],
[
0.080078125,
0.7197265625,
0.46728515625,
0.7197265625,
0.46728515625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.72998046875,
0.890625,
0.72998046875,
0.890625,
0.74072265625,
0.50341796875,
0.74072265625
],
[
0.0767822265625,
0.7470703125,
0.4697265625,
0.74853515625,
0.4697265625,
0.76953125,
0.0767822265625,
0.767578125
],
[
0.50244140625,
0.74365234375,
0.890625,
0.74365234375,
0.890625,
0.7568359375,
0.50244140625,
0.7568359375
],
[
0.50341796875,
0.76025390625,
0.8896484375,
0.76025390625,
0.8896484375,
0.77099609375,
0.50341796875,
0.77099609375
],
[
0.07843017578125,
0.7666015625,
0.468017578125,
0.767578125,
0.468017578125,
0.7822265625,
0.07843017578125,
0.78076171875
],
[
0.50341796875,
0.7734375,
0.890625,
0.7734375,
0.890625,
0.787109375,
0.50341796875,
0.787109375
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.796875,
0.08087158203125,
0.796875
],
[
0.5009765625,
0.787109375,
0.74267578125,
0.787109375,
0.74267578125,
0.80029296875,
0.5009765625,
0.80029296875
],
[
0.07843017578125,
0.796875,
0.46728515625,
0.798828125,
0.46728515625,
0.81298828125,
0.07843017578125,
0.81103515625
],
[
0.0767822265625,
0.8125,
0.295654296875,
0.81298828125,
0.295654296875,
0.83056640625,
0.0767822265625,
0.830078125
],
[
0.50341796875,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.080078125,
0.84521484375,
0.2115478515625,
0.84521484375,
0.2115478515625,
0.85986328125,
0.080078125,
0.85986328125
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85595703125,
0.50341796875,
0.85595703125
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.09722900390625,
0.869140625,
0.4697265625,
0.87060546875,
0.4697265625,
0.88525390625,
0.09722900390625,
0.8837890625
],
[
0.5009765625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.9013671875,
0.50244140625,
0.9013671875
]
] |
[
"",
"O",
"eNeRE",
"Decoder",
"Target NeRF",
"Voxel Grid",
"3DFPN",
"Feature Points",
"Transformer",
"Figure 2: Network architecture of our method. The pipeline of our method is: 1) We first extract the pairwise voxel grid",
"GERFrerxYesx",
"X zexC and a binary mask M e IRrer × Yles x zres ×1 from the source NeRF and the target NeRF. 2) The voxel grid",
"G and a binary mask M are fed into the 3D feature pyramid network to extract voxel features. 3) The extracted voxel grid",
"features are downsampled to a 2-dimensional tensor F e RN xc by their spherical neighborhood. 4) The resulting source",
"features Fsource and target features Ftarget are strengthened by a transformer, where a self-attention layer is used to enhance",
"the intra-contextualrelations,anda cross-attention layer is used to learn theinter-contextualrelations.5)Finally,weuse a",
"single-head attentionlayertodecode thefeatures into correspondences and their corresponding confidence scores.",
"feature backbone, where all the 2D modules are replaced",
"layers. We follow the same intuition of Predator [16] and",
"by their corresponding 3D parts. The feature pyramid net-",
"REGTR [43], where a self-attention layer is applied to both",
"work enables the learning of high-level semantic features",
"the source feature Fsource and the target feature Ftarget to en-",
"at a multi-scale and thus is suitable to extract the voxel",
"hance the intra-contextual relations, and a cross-attention",
"grid features in our task settings. We note the difference",
"layer is applied to both Fsource and Ftarget to learn the inter-",
"from the original feature pyramid network which utilizes",
"contextual relations. We follow the classical transformer",
"different scale features for object detection.we adopt only",
"architecture [41, 20, 43] with input to be voxel features.",
"thefeatures output fromthe lastlayer.Sincethe dimen-",
"Multi-Head Attention.The multi-head attention opera-",
"sionofthefeaturegridfromthelastlayercan bedifferent",
"tion in each layer is defined as:",
"fromtheoriginal voxelgrid.werescaleittothesizeofthe",
"original voxel grid G and resulted in a voxel feature grid",
"MH(Q,K, V) = concat(headi,..: ,headh)WO, (4)",
"GfERFrefxresXzresxC",
"We cannot use G f as the input to our transformer since",
"where head; = Attention(Qw?, KwK, Vwy). The at-",
"the voxel feature grid can contain too many points to en-",
"tention function is adopted as the scaled dot-product:",
"able the transformer to run on a single GPU. To solve",
"this issue, we iteratively downsample G by the spherical",
"QKT",
"Attention(Q,K, V) = softmax(",
")V.",
"(5)",
"neighborhood [36] as done in KPConv [37] to obtain the",
"Vdk",
"downsampled voxel points X. The downsample iteration",
"In the self-attention layers, Q = K = V represents the",
"is terminated when the total number of occupied voxels in",
"same feature tensor in each block. In the cross-attention lay-",
"the current sampled voxel feature grid is less than 1.5K.",
"Lastly, we reshape the occupied voxel features into a ten-",
"ers,thekeys,and thevalues arethefeaturetensorsfromthe",
"sor F e IRNxC. The weights of the feature extraction",
"other block. The self-attention mechanism enables the net-",
"worktolearntherelationshipinsidethesamefeaturepoints,",
"network are shared across the source and the target NeRFs.",
"while the cross-attention mechanism enables the communi-",
"We denote the downsampled voxel points and extracted fea-",
"cation of the different feature points.",
"tures of the source and target NeRFs as: [Xsource, Fsource]",
"and [Xtarget, Fiargetl], respectively.",
"Decoder.After encoding features by transformer, we fur-",
"ther adopt a single-head attention layer to predict the cor-",
"3.4.Transformer",
"responding point locations Xsource and confidence scores",
"Ssource of the source voxel points Xsource in the target",
"We then feed the resulting feature Fsource and Farget into",
"NeRF's coordinate frame. Similarly, we also predict the",
"a L-layer transformer with self-attention and cross-attention",
"corresponding point locations Xtarget and confidence scores"
] |
{
"arxiv_id": "2308.09386",
"language": "en",
"page_idx": 3
}
|
||
2308.09386-en-4
|
2308.09386-en-4.jpg
|
$\tilde{\mathbf{S}}_{\mathrm{target~of~the~target~voxel~points~\hat{X}_{\mathrm{target~in~the~source~NeRF's}}}}$ coordinate frame. Finally, we utilize the predicted correspondences to compute the relative rigid transformation. The confidence scores are used as weights that mask out the irrelevant correspondences and can be interpreted as how likely the predicted points from $\tilde{\mathbf{X}}_{\mathrm{source}}$ and $\hat{\mathbf{X}}_{\mathrm{target}}$ are correspondences and are visible in the source NeRF and the target NeRF.
### 3.5. Training Loss
Surface Field Supervision. To train the network, we encourage the predicted correspondences to have the same features which are invariant in the corresponding NeRF model. The naïve way is to adopt density felds as supervision instead of radiance fields since density fields are invariant to view points. However, density fields can be very noisy. We thus utilize the surface fields [14] as supervision. The surface field is defined as the differential probability of the ray hitting a surface at $\mathbf{X}_n$ given by:
$$\mathcal{S}(t)=\mathcal{T}(t)\cdot(1-\exp(-2\sigma\delta)).$$
(6)
Proof: The differential probability of a ray hitting a surface at point X is the product of the probability of a ray traveling over $[0,t_n)$ without hitting any particle before $t_n$ times the differential probability of the ray hitting exactly at point $\mathbf{X}(t_n).$ The surface field can then be written as:
$$S(t)=\int_{t-\delta}^{t+\delta}\mathcal{T}(s)\cdot\sigma(s)\:ds$$
(7)
We derive the exact form of $S(t)$ as follow:
$$\begin{aligned}S(t)&=\int_{t-\delta}^{t+\delta}\mathcal{T}(s)\cdot\sigma(s)\:ds\\&=\int_{t-\delta}^{t+\delta}\mathcal{T}(0\to t-\delta)\cdot\mathcal{T}(t-\delta\to s)\cdot\sigma(s)\:ds\\&=\mathcal{T}(t-\delta)\cdot\int_{t-\delta}^{t+\delta}\mathcal{T}(t-\delta\to s)\cdot\sigma(s)\:ds\\&=\mathcal{T}(t-\delta)\cdot\sigma_t\cdot\int_{t-\delta}^{t+\delta}\mathcal{T}(t-\delta\to s)\:ds\\&=\mathcal{T}(t-\delta)\cdot\sigma_t\cdot\int_{t-\delta}^{t+\delta}\left(\exp\left(-\int_{t-\delta}^s\sigma(\mu)\:d\mu\right)\right)ds\\&=\mathcal{T}(t-\delta)\cdot\sigma_t\cdot\int_{t-\delta}^{t+\delta}\left(\exp\left(-\sigma_t(s-t+\sigma)\right)\right)ds\\&=\mathcal{T}(t-\delta)\cdot\sigma_t\cdot(-\frac1{\sigma_t})\cdot\exp\left(-\sigma_t(s-t+\delta)\right)|_{t-\delta}^{t+\delta}\\&=\mathcal{T}(t-\delta)\cdot\left(1-\exp(-2\sigma_t\cdot\delta)\right).&\text{(8)}\end{aligned}$$
The second term holds since transmittance is multiplicative $(c.f.$Eq.(18) of[34]). The 4th term holds since we can
assume the density $\sigma_t$ is a constant within a small region $[t-\delta,t+\delta].$ To produce the view-independent field, we first form N viewing rays for each point. Subsequently, we take the maximum value of all rays as the density field value instead of averaging the value (as done in Sec. 3.2) for the radiance feld. Finally, we obtain the surface field mask M$_\mathrm{sf}$ by checking $S(t)>\eta$ (we use $\eta=0.5).$ We then update our mask by M= $M_{occ}$ $\cap \mathbf{M_{df}~}\cap \mathbf{M_{sf}. }$
Confidence Loss. We adopt the cross-entropy loss to supervise the confidence score, with the surface felds $\hat{\mathbf{S} } _{\mathrm{source}}$ = $S( \hat{\mathbf{X} } _{\mathrm{source}}) , \hat{\mathbf{S} } _{\mathrm{target}}$ = $S( \hat{\mathbf{X} } _{\mathrm{target}})$ queried from NeRF as the ground truth label:
$$\mathcal{L}_{\mathrm{conf}}=\mathrm{BCE}(\hat{\mathbf{S}}_{\mathrm{source}},\tilde{\mathbf{S}}_{\mathrm{source}})+\mathrm{BCE}(\hat{\mathbf{S}}_{\mathrm{target}},\tilde{\mathbf{S}}_{\mathrm{target}}).$$
(9)
Surface Field Loss. In addition to the confidence loss, we also encourage the predicted correspondences to have consistent surface field values:
$$\mathcal{L}_{\mathrm{sf}}=\frac1N\|\mathbf{S}([\hat{\mathbf{X}}_{\mathrm{source}},\hat{\mathbf{X}}_{\mathrm{target}}])-\mathbf{S}([\tilde{\mathbf{X}}_{\mathrm{source}},\tilde{\mathbf{X}}_{\mathrm{target}}])\|_1,$$
(10)
where $N$ is the total number of the concatenated source
voxel points and target voxel points.
Correspondence Loss. We further use a correspondence loss to constrain the predicted locations of correspondences:
$$\mathcal{L}_{\mathrm{corr}}=\sum\rho(\mathbf{S}\|\mathcal{T}^{\star}(\mathbf{x}_{i})-\mathbf{y}_{i}\|;\eta,\gamma),$$
(11)
where $\mathcal{T}^\star$ is the ground truth relative transformation between the source NeRF and target NeRF, $\rho(\cdot;\eta,\gamma)$ is the adaptive robust loss function [1] and $\{\eta,\gamma\}$ are respectively a smooth interpolation value and a scale parameter, which are hyperparameters used to control the shape of the robust function. We use $\eta=1.0,\gamma=0.5$ in our experiments.
Moreover, we adopt the feature loss [40,43] to leverage the geometric properties when computing the correspondences. Our final loss is therefore defined as:
$$\mathcal{L}_{\mathrm{final}}=\mathcal{L}_{\mathrm{conf}}+\lambda_{1}\mathcal{L}_{\mathrm{sf}}+\lambda_{2}\mathcal{L}_{\mathrm{corr}}+\lambda_{3}\mathcal{L}_{\mathrm{feat}},$$
(12)
where $\lambda_1,\lambda_2,\lambda_3$ are the weights associated with the corresponding loss functions. In our experiments, we use $\lambda_1=1.0,\lambda_2=0.1,\lambda_3=1.0.$
# 4. Experiments
Datasets. We aim at registering multiple NeRFs. Due to the lack of a suitable dataset for our task, we downloaded the 3D mesh models of 1,700+objects from Objaverse[10] to construct our dataset. Objaverse[10] is a massive dataset that contains 800K+ annotated 3D Objects. It is created for the text-to-3D task. We utilize it to construct our dataset for NeRF registration. Specifically, we randomly selected
|
[
[
0.08251953125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.08251953125,
0.1060791015625
],
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08087158203125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.08087158203125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.15087890625,
0.08087158203125,
0.15087890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.1546630859375,
0.890625,
0.1546630859375,
0.890625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50341796875,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.1817626953125,
0.890625,
0.1817626953125,
0.890625,
0.1956787109375,
0.5009765625,
0.1956787109375
],
[
0.0767822265625,
0.1976318359375,
0.1666259765625,
0.1956787109375,
0.1666259765625,
0.210205078125,
0.0767822265625,
0.212158203125
],
[
0.50244140625,
0.1976318359375,
0.7490234375,
0.1976318359375,
0.7490234375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.2215576171875,
0.221435546875,
0.2215576171875,
0.221435546875,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.50341796875,
0.2203369140625,
0.89208984375,
0.2203369140625,
0.89208984375,
0.2342529296875,
0.50341796875,
0.2342529296875
],
[
0.50244140625,
0.23486328125,
0.890625,
0.2342529296875,
0.890625,
0.248779296875,
0.50244140625,
0.25
],
[
0.080078125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.080078125,
0.25830078125
],
[
0.5,
0.248779296875,
0.89453125,
0.25,
0.89453125,
0.267578125,
0.5,
0.266357421875
],
[
0.07843017578125,
0.261474609375,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.07843017578125,
0.274658203125
],
[
0.50244140625,
0.266357421875,
0.7138671875,
0.266357421875,
0.7138671875,
0.27978515625,
0.50244140625,
0.27978515625
],
[
0.080078125,
0.27587890625,
0.468017578125,
0.27587890625,
0.468017578125,
0.2890625,
0.080078125,
0.2890625
],
[
0.080078125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.304931640625,
0.080078125,
0.304931640625
],
[
0.515625,
0.290283203125,
0.89208984375,
0.291748046875,
0.89208984375,
0.309326171875,
0.515625,
0.30810546875
],
[
0.08087158203125,
0.306884765625,
0.46728515625,
0.306884765625,
0.46728515625,
0.317626953125,
0.08087158203125,
0.317626953125
],
[
0.080078125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.333984375,
0.080078125,
0.333984375
],
[
0.08087158203125,
0.337158203125,
0.465576171875,
0.337158203125,
0.465576171875,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.50244140625,
0.333984375,
0.89208984375,
0.335205078125,
0.89208984375,
0.349853515625,
0.50244140625,
0.348388671875
],
[
0.08087158203125,
0.35107421875,
0.4697265625,
0.35107421875,
0.4697265625,
0.3642578125,
0.08087158203125,
0.3642578125
],
[
0.50341796875,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50341796875,
0.3642578125
],
[
0.080078125,
0.3662109375,
0.3447265625,
0.3662109375,
0.3447265625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.50244140625,
0.3662109375,
0.6845703125,
0.3662109375,
0.6845703125,
0.380126953125,
0.50244140625,
0.380126953125
],
[
0.1641845703125,
0.391357421875,
0.381591796875,
0.391357421875,
0.381591796875,
0.406005859375,
0.1641845703125,
0.406005859375
],
[
0.4453125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.406005859375,
0.4453125,
0.406005859375
],
[
0.51953125,
0.396484375,
0.5498046875,
0.396484375,
0.5498046875,
0.4091796875,
0.51953125,
0.4091796875
],
[
0.57421875,
0.392578125,
0.8759765625,
0.392578125,
0.8759765625,
0.410400390625,
0.57421875,
0.410400390625
],
[
0.86376953125,
0.415283203125,
0.890625,
0.415283203125,
0.890625,
0.4267578125,
0.86376953125,
0.4267578125
],
[
0.08087158203125,
0.43310546875,
0.468017578125,
0.43310546875,
0.468017578125,
0.4462890625,
0.08087158203125,
0.4462890625
],
[
0.50341796875,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.442626953125,
0.50341796875,
0.442626953125
],
[
0.080078125,
0.447509765625,
0.46728515625,
0.447509765625,
0.46728515625,
0.462158203125,
0.080078125,
0.462158203125
],
[
0.50244140625,
0.44580078125,
0.74267578125,
0.44580078125,
0.74267578125,
0.458984375,
0.50244140625,
0.458984375
],
[
0.08087158203125,
0.46337890625,
0.46728515625,
0.46337890625,
0.46728515625,
0.477783203125,
0.08087158203125,
0.477783203125
],
[
0.50341796875,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.48095703125,
0.50341796875,
0.48095703125
],
[
0.080078125,
0.477783203125,
0.4697265625,
0.477783203125,
0.4697265625,
0.4912109375,
0.080078125,
0.4912109375
],
[
0.50244140625,
0.48291015625,
0.89208984375,
0.48291015625,
0.89208984375,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.080078125,
0.49365234375,
0.431396484375,
0.49365234375,
0.431396484375,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.57177734375,
0.51123046875,
0.82666015625,
0.51123046875,
0.82666015625,
0.52880859375,
0.57177734375,
0.52880859375
],
[
0.85888671875,
0.51220703125,
0.8935546875,
0.51220703125,
0.8935546875,
0.52783203125,
0.85888671875,
0.52783203125
],
[
0.178955078125,
0.52783203125,
0.2197265625,
0.52783203125,
0.2197265625,
0.54248046875,
0.178955078125,
0.54248046875
],
[
0.271240234375,
0.52783203125,
0.370849609375,
0.52783203125,
0.370849609375,
0.54248046875,
0.271240234375,
0.54248046875
],
[
0.44677734375,
0.52880859375,
0.4697265625,
0.52880859375,
0.4697265625,
0.54248046875,
0.44677734375,
0.54248046875
],
[
0.5048828125,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.55029296875,
0.5048828125,
0.55029296875
],
[
0.5009765625,
0.55029296875,
0.890625,
0.5517578125,
0.890625,
0.56640625,
0.5009765625,
0.56494140625
],
[
0.08087158203125,
0.56005859375,
0.365234375,
0.56103515625,
0.365234375,
0.57568359375,
0.08087158203125,
0.57470703125
],
[
0.50341796875,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.50244140625,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.08087158203125,
0.5966796875,
0.1217041015625,
0.5966796875,
0.1217041015625,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.17236328125,
0.5966796875,
0.271240234375,
0.5966796875,
0.271240234375,
0.611328125,
0.17236328125,
0.611328125
],
[
0.50341796875,
0.59765625,
0.8935546875,
0.59765625,
0.8935546875,
0.611328125,
0.50341796875,
0.611328125
],
[
0.50244140625,
0.61279296875,
0.86279296875,
0.61279296875,
0.86279296875,
0.626953125,
0.50244140625,
0.626953125
],
[
0.156005859375,
0.62744140625,
0.1683349609375,
0.62744140625,
0.1683349609375,
0.6318359375,
0.156005859375,
0.6318359375
],
[
0.521484375,
0.626953125,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.521484375,
0.64013671875
],
[
0.17236328125,
0.6337890625,
0.4345703125,
0.6337890625,
0.4345703125,
0.6484375,
0.17236328125,
0.6484375
],
[
0.50244140625,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.50341796875,
0.65771484375,
0.80126953125,
0.65771484375,
0.80126953125,
0.671875,
0.50341796875,
0.671875
],
[
0.11767578125,
0.671875,
0.1871337890625,
0.671875,
0.1871337890625,
0.68603515625,
0.11767578125,
0.68603515625
],
[
0.24267578125,
0.67236328125,
0.398681640625,
0.67236328125,
0.398681640625,
0.6875,
0.24267578125,
0.6875
],
[
0.54345703125,
0.68505859375,
0.82861328125,
0.68505859375,
0.82861328125,
0.69970703125,
0.54345703125,
0.69970703125
],
[
0.859375,
0.68505859375,
0.89208984375,
0.68505859375,
0.89208984375,
0.69775390625,
0.859375,
0.69775390625
],
[
0.2484130859375,
0.70263671875,
0.2646484375,
0.70263671875,
0.2646484375,
0.70703125,
0.2484130859375,
0.70703125
],
[
0.11767578125,
0.7109375,
0.22216796875,
0.7109375,
0.22216796875,
0.7236328125,
0.11767578125,
0.7236328125
],
[
0.265625,
0.70849609375,
0.386474609375,
0.70556640625,
0.386474609375,
0.7236328125,
0.265625,
0.72607421875
],
[
0.50341796875,
0.7119140625,
0.890625,
0.7119140625,
0.890625,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.5048828125,
0.72900390625,
0.68896484375,
0.72900390625,
0.68896484375,
0.7392578125,
0.5048828125,
0.7392578125
],
[
0.6845703125,
0.72802734375,
0.8935546875,
0.72802734375,
0.8935546875,
0.7412109375,
0.6845703125,
0.7412109375
],
[
0.11517333984375,
0.74365234375,
0.238525390625,
0.7470703125,
0.238525390625,
0.7646484375,
0.11517333984375,
0.76123046875
],
[
0.261474609375,
0.744140625,
0.35302734375,
0.7421875,
0.3544921875,
0.7646484375,
0.26318359375,
0.7666015625
],
[
0.3505859375,
0.744140625,
0.461669921875,
0.744140625,
0.461669921875,
0.76318359375,
0.3505859375,
0.76318359375
],
[
0.5009765625,
0.7412109375,
0.70166015625,
0.7412109375,
0.70166015625,
0.755859375,
0.5009765625,
0.755859375
],
[
0.50341796875,
0.771484375,
0.63134765625,
0.771484375,
0.63134765625,
0.7861328125,
0.50341796875,
0.7861328125
],
[
0.11517333984375,
0.78076171875,
0.2442626953125,
0.78515625,
0.24267578125,
0.80322265625,
0.11358642578125,
0.798828125
],
[
0.25732421875,
0.78271484375,
0.465576171875,
0.77978515625,
0.465576171875,
0.80029296875,
0.25732421875,
0.8037109375
],
[
0.50244140625,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.26318359375,
0.814453125,
0.46728515625,
0.814453125,
0.46728515625,
0.8359375,
0.26318359375,
0.8359375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.5009765625,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5009765625,
0.8388671875
],
[
0.11517333984375,
0.81640625,
0.2467041015625,
0.8193359375,
0.2467041015625,
0.83544921875,
0.11517333984375,
0.83203125
],
[
0.11517333984375,
0.84521484375,
0.343017578125,
0.84521484375,
0.343017578125,
0.85986328125,
0.11517333984375,
0.85986328125
],
[
0.44677734375,
0.84521484375,
0.4697265625,
0.84521484375,
0.4697265625,
0.859375,
0.44677734375,
0.859375
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.08087158203125,
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
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"Starget of the target voxel points Xtarget in the source NeRF's",
"assume the density ot is a constant within a small region",
"coordinate frame. Finally, we utilize the predicted corre-",
"[t - ,t + ]. To produce the view-independent field, we",
"spondences to compute the relative rigid transformation.",
"first form N viewing rays for each point. Subsequently, we",
"The confidence scores are used as weights that mask out the",
"take the maximum value of all rays as the density field value",
"irrelevant correspondences and can be interpreted as how",
"instead of averaging the value (as done in Sec. 3.2)for the",
"likely the predicted points from Xsource and Xtarget are cor-",
"radiance field.Finally,we obtain the surface field mask Ms",
"respondences and are visible in the source NeRF and the",
"by checking S(t) > n (we use n = 0.5). We then update our",
"target NeRF.",
"mask by M = Moce n Mar n Msf.",
"3.5. Training Loss",
"Confidence Loss. We adopt the cross-entropy loss to",
"supervise the confidence score, with the surface fields",
"Surface Field Supervision.To train the network, we en-",
"Ssource = S(Xsource), Starget = S(Xuarget) queried from",
"courage the predicted correspondences to have the same",
"NeRF as the ground truth label:",
"features which are invariant in the corresponding NeRF",
"model. The naive way is to adopt density fields as super-",
"Lconf = BCE(Ssource, Ssource) + BCE(Starget, Starget). (9)",
"vision instead of radiance fields since density fields are in-",
"variant to view points. However, density fields can be very",
"noisy. Wethus utilizethe surface fieldsl4as supervision.",
"Surface Field Loss.In addition to the confidenceloss, we",
"The surface field is defined as the differential probability of",
"also encourage the predicted correspondences to have con-",
"the ray hitting a surface at Xn given by:",
"sistent surface field values:",
"S(t) = T(t) : (1 - exp(-2go)).",
"(6)",
"Cst=",
"S([Xsource, Xtargetl) -S([Xsource,Xtargetl]1,",
"(10)",
"Proof:The differential probability of a ray hitting a sur-",
"where N is the total number of the concatenated source",
"face at point X is the product of the probability of a ray",
"voxel points and target voxel points.",
"traveling over [0, tn) without hitting any particle before tn",
"Correspondence Loss.We further use a correspondence",
"times the differential probability of the ray hitting exactly at",
"loss to constrain the predicted locations of correspondences:",
"point X(tn). The surface field can then be written as:",
"Ccor =p(ST*(x,) -yill;n,),",
"(11)",
"S(t) :",
"T(s) - a(s) ds",
"(7)",
"where T* is the ground truth relative transformation be-",
"tween the source NeRF and target NeRF, p(; n, ) is the",
"We derive the exact form of s(t) as follow:",
"adaptive robust loss function [] and (n, ] are respectively",
"a smooth interpolation value and a scale parameter, which",
"S(t) :",
"T(s) :o(s) ds",
"are hyperparameters used to control the shape of the robust",
"function. We use n = 1.0, = 0.5 in our experiments.",
"+",
"Moreover,we adopt the featureloss [40,43] to leverage",
"T(0 →t-) . T(t--→s) .a(s) ds",
"the geometric properties when computing the correspon-",
"dences. Our final loss is therefore defined as:",
"T(t-s",
"T(t --→s) o(s) ds",
"Linal=Lconf+ALsf+A2Ccorr+ACfeat,",
"(12)",
"+(",
"T(t-).at",
"T(t-→s) ds",
"where Ai,>2, A3 are the weights associated with the cor-",
"respondinglossfunctions.",
"Inour experiments,weuse",
"= T(t-)-ot*",
"(exp(-",
"o(μ) dμ)) ds",
"入1 = 1.0, 入2 = 0.1, 43 = 1.0.",
"4.Experiments",
"= T(t-0)·ot:",
"(exp(-ot(s -t+o)) ds",
"Datasets. We aim at registering multiple NeRFs. Due to",
") ·exp(-at(s -t+ )It+",
"the lack of a suitable dataset for our task, we downloaded",
"the 3D mesh models of 1, 700+ objects from Objaverse [10]",
"= T(t-8)·ot(-",
"= T(t - 8) : (1 -exp(-2ot : )),",
"(8)",
"to construct our dataset. Objaverse [l0] is a massive dataset",
"that contains 800K+ annotated 3D Objects.It is created",
"The second term holds since transmittance is multiplicative",
"for the text-to-3D task. We utilize it to construct our dataset",
"(c.f. Eq.(18) of [34]). The 4th term holds since we can",
"for NeRF registration. Specifically, we randomly selected"
] |
{
"arxiv_id": "2308.09386",
"language": "en",
"page_idx": 4
}
|
||
1908.08486-en-0
|
1908.08486-en-0.jpg
|
## Dialogue Coherence Assessment Without Explicit Dialogue Act Labels
Sebastian Bücker
Mohsen Mesgar
$\textbf{Iryna Gurevych}$
Ubiquitous Knowledge Processing Lab (UKP)
Technische Universität Darmstadt (TUDa)
$\{$mesgar,buecker,gurevych$\}$@ukp.tu-darmstadt.de
## Abstract
Recent dialogue coherence models use the coherence features designed for monologue texts, e.g. nominal entities, to represent utterances and then explicitly augment them with dialogue-relevant features, e.g., dialogue act labels. It indicates two drawbacks, (a) semantics of utterances is limited to entity mentions, and (b) the performance of coherence models strongly relies on the quality of the input dialogue act labels. We address these issues by introducing a novel approach to dialogue coherence assessment. We use dialogue act prediction as an auxiliary task in a multi-task learning scenario to obtain informative utterance representations for coherence assessment. Our approach alleviates the need for explicit dialogue act labels during evaluation. The results of our experiments show that our model substantially (more than 20 accuracy points) outperforms its strong competitors on the DailyDialogue corpus, and performs on par with them on the SwitchBoard corpus for ranking dialogues concerning their coherence. We release our source code$^{\mathrm{l}}.$
### 1 Introduction
Considering rapid progresses in developing open-domain dialogue agents (Serban et al., 2016; Ghazvininejad et al., 2018; Dinan et al., 2019; Li et al., 2019), the need for models that compare these agents in various dialogue aspects becomes extremely important (Liu et al., 2016; Dinan et al., 2019). Most available methods for dialogue evaluation rely on word-overlap metrics, e.g. BLEU, and manually collected human feedback. The former does not strongly correlate with human judgments (Liu et al., 2016), and the latter is timeconsuming and subjective. A fundamental aspect of dialogue is coherence - what discriminates a
<FigureHere>
Figure 1: A high-level view of our multi-task learning
approach for dialogue coherence modeling.
high-quality dialogue from a random sequence of dialogue utterances (Halliday and Hasan, 1976; Grosz and Sidner, 1986; Byron and Stent, 1998). Dialogue coherence deals with semantic relations between utterances considering their dialogue acts (Perrault and Allen, 1978; Cervone et al., 2018).
A Dialogue Act (henceforth $DA)$ gives a meaning to an utterance in a dialogue at the level of “illocutionary force", and therefore, constitutes the basic unit of communication (Searle, 1969; Raheja and Tetreault, 2019). A DA captures what a speaker's intention is of saying an utterance without regard to the actual content of the utterance. For example, a DA may indicate whether the intention of stating an utterance is to ask a question or to state a piece of information.
Recent approaches to dialogue coherence modeling use the coherence features designed
entity transitions
for monologue texts, e.g.
(Barzilay and Lapata, 2005), and augment them with dialogue-relevant features, e.g., DA labels (Cervone et al., 2018). These DA labels are provided by human annotators or DA prediction models. Such coherence models suffer from the following drawbacks: (a) they curb semantic representations of utterances to entities, which are sparse in dialogue because of short utterance lengths, and
$^{1}$https:// github. com/ UKPLab/ acl2020- dialog(b) their performanes sekies on the quality of their
|
[
[
0.3369140625,
0.0843505859375,
0.66796875,
0.0843505859375,
0.66796875,
0.09857177734375,
0.3369140625,
0.09857177734375
],
[
0.3076171875,
0.1009521484375,
0.69580078125,
0.1015625,
0.69580078125,
0.119384765625,
0.3076171875,
0.1181640625
],
[
0.1856689453125,
0.146728515625,
0.326904296875,
0.146728515625,
0.326904296875,
0.160888671875,
0.1856689453125,
0.160888671875
],
[
0.42529296875,
0.146728515625,
0.578125,
0.146728515625,
0.578125,
0.1602783203125,
0.42529296875,
0.1602783203125
],
[
0.68212890625,
0.146728515625,
0.82275390625,
0.146728515625,
0.82275390625,
0.160888671875,
0.68212890625,
0.160888671875
],
[
0.3134765625,
0.168701171875,
0.6884765625,
0.169189453125,
0.6884765625,
0.18408203125,
0.3134765625,
0.182861328125
],
[
0.32861328125,
0.185302734375,
0.673828125,
0.1859130859375,
0.673828125,
0.20068359375,
0.32861328125,
0.1995849609375
],
[
0.2344970703125,
0.2037353515625,
0.7724609375,
0.2037353515625,
0.7724609375,
0.21728515625,
0.2344970703125,
0.21728515625
],
[
0.2646484375,
0.26318359375,
0.34375,
0.265380859375,
0.34375,
0.28076171875,
0.2646484375,
0.279052734375
],
[
0.560546875,
0.265380859375,
0.66064453125,
0.265380859375,
0.66064453125,
0.276611328125,
0.560546875,
0.276611328125
],
[
0.76708984375,
0.26318359375,
0.83447265625,
0.26318359375,
0.83447265625,
0.27783203125,
0.76708984375,
0.27783203125
],
[
0.023529052734375,
0.29443359375,
0.060516357421875,
0.29443359375,
0.060516357421875,
0.4091796875,
0.023529052734375,
0.4091796875
],
[
0.1470947265625,
0.29443359375,
0.46044921875,
0.29443359375,
0.46044921875,
0.30810546875,
0.1470947265625,
0.30810546875
],
[
0.56201171875,
0.297607421875,
0.65869140625,
0.297607421875,
0.65869140625,
0.312255859375,
0.56201171875,
0.312255859375
],
[
0.75732421875,
0.295654296875,
0.84375,
0.297607421875,
0.84375,
0.3134765625,
0.75732421875,
0.311279296875
],
[
0.1470947265625,
0.308837890625,
0.46044921875,
0.308837890625,
0.46044921875,
0.322509765625,
0.1470947265625,
0.322509765625
],
[
0.1470947265625,
0.32373046875,
0.46044921875,
0.32373046875,
0.46044921875,
0.33740234375,
0.1470947265625,
0.33740234375
],
[
0.148681640625,
0.336181640625,
0.459716796875,
0.336181640625,
0.459716796875,
0.349853515625,
0.148681640625,
0.349853515625
],
[
0.62451171875,
0.330810546875,
0.78662109375,
0.330810546875,
0.78662109375,
0.344482421875,
0.62451171875,
0.344482421875
],
[
0.1470947265625,
0.3515625,
0.46044921875,
0.3515625,
0.46044921875,
0.365234375,
0.1470947265625,
0.365234375
],
[
0.148681640625,
0.36767578125,
0.459716796875,
0.36767578125,
0.459716796875,
0.377685546875,
0.148681640625,
0.377685546875
],
[
0.673828125,
0.366455078125,
0.73779296875,
0.366455078125,
0.73779296875,
0.377685546875,
0.673828125,
0.377685546875
],
[
0.1470947265625,
0.380126953125,
0.46044921875,
0.380126953125,
0.46044921875,
0.393798828125,
0.1470947265625,
0.393798828125
],
[
0.5126953125,
0.3896484375,
0.884765625,
0.39013671875,
0.884765625,
0.405029296875,
0.5126953125,
0.40380859375
],
[
0.148681640625,
0.3955078125,
0.459716796875,
0.3955078125,
0.459716796875,
0.40625,
0.148681640625,
0.40625
],
[
0.148681640625,
0.4091796875,
0.46044921875,
0.4091796875,
0.46044921875,
0.4228515625,
0.148681640625,
0.4228515625
],
[
0.5126953125,
0.40380859375,
0.8076171875,
0.405029296875,
0.8076171875,
0.419189453125,
0.5126953125,
0.418701171875
],
[
0.02520751953125,
0.413330078125,
0.06219482421875,
0.413330078125,
0.06219482421875,
0.4892578125,
0.02520751953125,
0.4892578125
],
[
0.148681640625,
0.4228515625,
0.462158203125,
0.4228515625,
0.462158203125,
0.435791015625,
0.148681640625,
0.435791015625
],
[
0.1470947265625,
0.43701171875,
0.46044921875,
0.43701171875,
0.46044921875,
0.45068359375,
0.1470947265625,
0.45068359375
],
[
0.1453857421875,
0.449462890625,
0.46044921875,
0.45068359375,
0.46044921875,
0.465576171875,
0.1453857421875,
0.46435546875
],
[
0.5126953125,
0.453125,
0.88330078125,
0.453125,
0.88330078125,
0.466064453125,
0.5126953125,
0.466064453125
],
[
0.148681640625,
0.465576171875,
0.46044921875,
0.465576171875,
0.46044921875,
0.478515625,
0.148681640625,
0.478515625
],
[
0.51416015625,
0.4697265625,
0.884765625,
0.4697265625,
0.884765625,
0.482666015625,
0.51416015625,
0.482666015625
],
[
0.148681640625,
0.48095703125,
0.46044921875,
0.48095703125,
0.46044921875,
0.4912109375,
0.148681640625,
0.4912109375
],
[
0.51416015625,
0.485107421875,
0.884765625,
0.485107421875,
0.884765625,
0.498779296875,
0.51416015625,
0.498779296875
],
[
0.0218505859375,
0.4912109375,
0.06219482421875,
0.4912109375,
0.06219482421875,
0.70166015625,
0.0218505859375,
0.70166015625
],
[
0.1470947265625,
0.493408203125,
0.462158203125,
0.493408203125,
0.462158203125,
0.50732421875,
0.1470947265625,
0.50732421875
],
[
0.51416015625,
0.501953125,
0.88330078125,
0.501953125,
0.88330078125,
0.515625,
0.51416015625,
0.515625
],
[
0.148681640625,
0.50830078125,
0.46044921875,
0.50830078125,
0.46044921875,
0.521484375,
0.148681640625,
0.521484375
],
[
0.1470947265625,
0.521484375,
0.46044921875,
0.521484375,
0.46044921875,
0.53515625,
0.1470947265625,
0.53515625
],
[
0.51416015625,
0.51708984375,
0.88330078125,
0.51708984375,
0.88330078125,
0.53076171875,
0.51416015625,
0.53076171875
],
[
0.1470947265625,
0.5361328125,
0.46044921875,
0.5361328125,
0.46044921875,
0.5498046875,
0.1470947265625,
0.5498046875
],
[
0.51611328125,
0.53515625,
0.869140625,
0.53515625,
0.869140625,
0.5458984375,
0.51611328125,
0.5458984375
],
[
0.1470947265625,
0.5498046875,
0.46044921875,
0.5498046875,
0.46044921875,
0.5634765625,
0.1470947265625,
0.5634765625
],
[
0.53173828125,
0.552734375,
0.88330078125,
0.552734375,
0.88330078125,
0.56787109375,
0.53173828125,
0.56787109375
],
[
0.148681640625,
0.56640625,
0.459716796875,
0.56640625,
0.459716796875,
0.57666015625,
0.148681640625,
0.57666015625
],
[
0.51416015625,
0.5693359375,
0.88671875,
0.5693359375,
0.88671875,
0.5830078125,
0.51416015625,
0.5830078125
],
[
0.1470947265625,
0.5791015625,
0.46044921875,
0.5791015625,
0.46044921875,
0.5927734375,
0.1470947265625,
0.5927734375
],
[
0.5126953125,
0.5849609375,
0.88330078125,
0.5849609375,
0.88330078125,
0.5986328125,
0.5126953125,
0.5986328125
],
[
0.1470947265625,
0.59326171875,
0.46044921875,
0.59326171875,
0.46044921875,
0.60693359375,
0.1470947265625,
0.60693359375
],
[
0.1470947265625,
0.60693359375,
0.46044921875,
0.60693359375,
0.46044921875,
0.62158203125,
0.1470947265625,
0.62158203125
],
[
0.51611328125,
0.6015625,
0.88671875,
0.6015625,
0.88671875,
0.615234375,
0.51611328125,
0.615234375
],
[
0.5126953125,
0.6162109375,
0.88330078125,
0.61767578125,
0.88330078125,
0.6318359375,
0.5126953125,
0.63134765625
],
[
0.1470947265625,
0.62255859375,
0.3017578125,
0.62255859375,
0.3017578125,
0.6328125,
0.1470947265625,
0.6328125
],
[
0.51416015625,
0.63427734375,
0.88671875,
0.63427734375,
0.88671875,
0.64794921875,
0.51416015625,
0.64794921875
],
[
0.1436767578125,
0.65087890625,
0.26123046875,
0.65087890625,
0.26123046875,
0.66455078125,
0.1436767578125,
0.66455078125
],
[
0.51416015625,
0.6494140625,
0.88671875,
0.6494140625,
0.88671875,
0.6630859375,
0.51416015625,
0.6630859375
],
[
0.51416015625,
0.6650390625,
0.88671875,
0.6650390625,
0.88671875,
0.6787109375,
0.51416015625,
0.6787109375
],
[
0.1226806640625,
0.65380859375,
0.1302490234375,
0.65380859375,
0.1302490234375,
0.6611328125,
0.1226806640625,
0.6611328125
],
[
0.11846923828125,
0.6767578125,
0.489013671875,
0.6767578125,
0.489013671875,
0.69140625,
0.11846923828125,
0.69140625
],
[
0.5126953125,
0.68115234375,
0.88330078125,
0.68115234375,
0.88330078125,
0.69482421875,
0.5126953125,
0.69482421875
],
[
0.12017822265625,
0.693359375,
0.489013671875,
0.693359375,
0.489013671875,
0.70703125,
0.12017822265625,
0.70703125
],
[
0.5126953125,
0.69873046875,
0.759765625,
0.69873046875,
0.759765625,
0.7119140625,
0.5126953125,
0.7119140625
],
[
0.11846923828125,
0.7080078125,
0.4873046875,
0.7080078125,
0.4873046875,
0.72265625,
0.11846923828125,
0.72265625
],
[
0.53125,
0.71630859375,
0.88330078125,
0.71728515625,
0.88330078125,
0.732421875,
0.53125,
0.73095703125
],
[
0.12017822265625,
0.7255859375,
0.489013671875,
0.7255859375,
0.489013671875,
0.7392578125,
0.12017822265625,
0.7392578125
],
[
0.5126953125,
0.73291015625,
0.88232421875,
0.73291015625,
0.88232421875,
0.74658203125,
0.5126953125,
0.74658203125
],
[
0.11846923828125,
0.7421875,
0.489013671875,
0.7421875,
0.489013671875,
0.755859375,
0.11846923828125,
0.755859375
],
[
0.5126953125,
0.75,
0.728515625,
0.75,
0.728515625,
0.76318359375,
0.5126953125,
0.76318359375
],
[
0.73193359375,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76416015625,
0.73193359375,
0.76416015625
],
[
0.11846923828125,
0.75830078125,
0.49072265625,
0.755859375,
0.49072265625,
0.77099609375,
0.11846923828125,
0.7724609375
],
[
0.51416015625,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.51416015625,
0.77978515625
],
[
0.12017822265625,
0.77392578125,
0.49072265625,
0.77392578125,
0.49072265625,
0.78759765625,
0.12017822265625,
0.78759765625
],
[
0.5126953125,
0.783203125,
0.88330078125,
0.783203125,
0.88330078125,
0.79638671875,
0.5126953125,
0.79638671875
],
[
0.1168212890625,
0.7890625,
0.492431640625,
0.7880859375,
0.492431640625,
0.802734375,
0.1168212890625,
0.80419921875
],
[
0.51416015625,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51416015625,
0.8125
],
[
0.11846923828125,
0.8056640625,
0.49072265625,
0.8056640625,
0.49072265625,
0.8193359375,
0.11846923828125,
0.8193359375
],
[
0.5126953125,
0.81396484375,
0.88671875,
0.81396484375,
0.88671875,
0.82763671875,
0.5126953125,
0.82763671875
],
[
0.12017822265625,
0.822265625,
0.49072265625,
0.822265625,
0.49072265625,
0.8359375,
0.12017822265625,
0.8359375
],
[
0.5126953125,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.5126953125,
0.84423828125
],
[
0.11846923828125,
0.83837890625,
0.492431640625,
0.83837890625,
0.492431640625,
0.8515625,
0.11846923828125,
0.8515625
],
[
0.51416015625,
0.84716796875,
0.88671875,
0.84716796875,
0.88671875,
0.86083984375,
0.51416015625,
0.86083984375
],
[
0.11846923828125,
0.85400390625,
0.489013671875,
0.85400390625,
0.489013671875,
0.8681640625,
0.11846923828125,
0.8681640625
],
[
0.51416015625,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51416015625,
0.876953125
],
[
0.11846923828125,
0.87060546875,
0.49072265625,
0.87060546875,
0.49072265625,
0.88525390625,
0.11846923828125,
0.88525390625
],
[
0.5126953125,
0.87744140625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.5126953125,
0.892578125
],
[
0.1412353515625,
0.8935546875,
0.88330078125,
0.89453125,
0.88330078125,
0.9091796875,
0.1412353515625,
0.90771484375
]
] |
[
"Dialogue Coherence Assessment",
"Without Explicit Dialogue Act Labels",
"Mohsen Mesgar",
"Sebastian Bicker",
"Iryna Gurevych",
"Ubiquitous Knowledge Processing Lab (UKP)",
"Technische Universitat Darmstadt (TUDa)",
"[mesgar,buecker,gurevych]@ukp.tu-darmstadt.de",
"Abstract",
"coherence score",
"DA labels",
"zz n ",
"Recent dialogue coherence models use the",
"DiCoh model",
"DAP model",
"coherence features designed for monologue",
"texts, e.g. nominal entities, to represent utter",
"ances and then explicitly augment them with",
"shared utterance encoder",
"dialogue-relevant features, e.g., dialogue act",
"labels.Itindicatestwodrawbacks,(a)seman",
"utterances",
"tics of utterances is limited to entity mentions",
"Figure l: A high-level view of our multi-task learning",
"and(b)theperformanceofcoherencemod-",
"els strongly relies on the quality of the input",
"approach for dialogue coherence modeling.",
"['S]",
"dialogue act labels. We address these issues",
"by introducing a novel approach to dialogue",
"coherence assessment. We use dialogue act",
"high-quality dialogue from a random sequence of",
"prediction as an auxiliary task in a multi-task",
"dialogue utterances (Halliday and Hasan, 1976;",
"learningscenariotoobtaininformativeutter-",
"Grosz and Sidner, 1986; Byron and Stent, 1998).",
"088806110",
"ance representations for coherence assessment.",
"Dialogue coherence deals with semantic relations",
"Our approach alleviates the need for explicit",
"dialogue act labels during evaluation. The re-",
"between utterances considering their dialogue acts",
"sults of our experiments show that our model",
"(PerraultandAllen,1978:Cervoneetal.,2018).",
"substantially (more than 20 accuracy points)",
"A Dialogue Act (henceforth DA) gives a mean-",
"outperformsitsstrongcompetitorsontheDai",
"ing to an utterance in a dialogue at the level of",
"lyDialogue corpus, and performs on par with",
"\"illocutionary force\", and therefore, constitutes",
"them on the SwitchBoard corpus for ranking",
"dialogues concerning their coherence. We re-",
"the basic unit of communication (Searle, 1969;",
"Raheja and Tetreault, 2019). A DA captures what",
"leaseoursourcecode!",
"a speaker's intention is of saying an utterance with-",
"Introduction",
"out regard to the actual content of the utterance.",
"For example, a DA may indicate whether the in-",
"1",
"Considering rapid progresses in developing",
"tention of stating an utterance is to ask a question",
"open-domain dialogue agents (Serban et al., 2016;",
"or to state a piece of information.",
"Ghazvininejad et al., 2018; Dinan et al., 2019",
"Recent approaches to dialogue coherence",
"Li et al., 2019), the need for models that compare",
"modeling use the coherence features designed",
"these agents in various dialogue aspects becomes",
"for monologue texts, e.g.",
"entity transitions",
"extremely important (Liu et al., 2016; Dinan et al.,",
"(Barzilay and Lapata, 2005), and augment them",
"2019). Most available methods for dialogue eval-",
"with dialogue-relevant features, e.g., DA labels",
"uation rely on word-overlap metrics, e.g. BLEU,",
"(Cervone et al., 2018). These DA labels are pro-",
"and manually collected human feedback. The",
"vided by human annotators or DA prediction mod-",
"former does not strongly correlate with human",
"els. Such coherence models suffer from the follow-",
"judgments (Liu et al., 2016), and the latter is time-",
"ing drawbacks: (a) they curb semantic representa-",
"consuming and subjective. A fundamental aspect",
"tions of utterances to entities, which are sparse in",
"of dialogue is coherence - what discriminates a",
"dialogue because of short utterance lengths, and",
"'https://github.com/UKPLab/ac12020-dialogu(b)theitperfermaneereliesonthequalityoftheir"
] |
{
"arxiv_id": "1908.08486",
"language": "en",
"page_idx": 0
}
|
||
1908.08486-en-1
|
1908.08486-en-1.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Utterance</th>
<th>DA label</th>
</tr>
<tr>
<td>$\mathbf{coherent}$</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$utt_1\colon$</td>
<td>$Thisismvuncle$, 2, $Charles.$</td>
<td>inform</td>
</tr>
<tr>
<td>$\operatorname{utt}_2$</td>
<td>He looks strong. What does $he$ $do?$</td>
<td>question</td>
</tr>
<tr>
<td>$utt_3\colon$</td>
<td>$He's~a~captain.$</td>
<td>inform</td>
</tr>
<tr>
<td>$utt_4:$</td>
<td>He must be very brave.</td>
<td>inform</td>
</tr>
<tr>
<td>$utt_5\colon$</td>
<td>$Exactly!$</td>
<td>inform</td>
</tr>
<tr>
<td>incoheren</td>
<td>mt</td>
<td> </td>
</tr>
<tr>
<td>$utt_1::$</td>
<td>This $is$ $my$ $uncle$, $Charles.$</td>
<td>$\operatorname{inform}$</td>
</tr>
<tr>
<td>$\mathrm{utt}_{4}{:}$</td>
<td>He must be very brave</td>
<td>inform</td>
</tr>
<tr>
<td>$utt_3\colon$</td>
<td>$He'sacaptain.$</td>
<td>$\operatorname{inform}$</td>
</tr>
<tr>
<td>ut$t_2$: </td>
<td>He looks strong. What does he do?</td>
<td>question</td>
</tr>
<tr>
<td>$utt_5\colon$</td>
<td>$Exactly!$</td>
<td>inform</td>
</tr>
</tbody>
</table>
Table 1: An example dialogue from DailyDialog (top) and its perturbation (bottom), which is generated by permuting the utterances said by one of the speakers (shown in boldface), and is less coherent. The right column shows the DA labels associated with utterances.
input DA labels.
We propose a novel approach to dialogue coherence assessment by utilizing dialogue act prediction as an auxiliary task for training our coherence model in a multi-task learning (MTL) scenario (Figure 1). Our approach consists of three high-level components: an utterance encoder, a dialogue coherence model (DiCoh), and $a$ Dialogue Act Prediction (DAP) model. The layers of the utterance encoder are shared between the DAP and the DiCoh model. This idea enables our DiCoh model to learn to focus on salient information presented in utterances considering their DAs and to alleviate the need for explicit DA labels during coherence assessment.
We evaluate our MTL-based approach on the DailyDialog (Li et al., 2017) and SwitchBoard (Jurafsky and Shriberg, 1997) English dialogue corpora in several discriminating experiments, where our coherence model, DiCoh, is examined to discriminate a dialogue from its perturbations (see Table 1). We utilize perturbation methods, like utterance ordering and utterance insertion, inherited from coherence evaluation approaches for monologue texts, and also introduce two dialoguerelevant perturbations, named utterance replacement and even utterance ordering. Our core contributions are: (1) proposing an MTL-based approach for dialogue coherence assessment using DAP as an auxiliary task, yielding more informative utterance representations for coherence assessment; (2) alleviating the need for DA labels for dialogue coherence assessment during evaluations; (3) an empirical evaluation on two benchmark dialogue corpora, showing that our model substantially outperforms the state-of-the-art coherence model on DailyDialog, and performs on par with it on SwitchBoard.
## 2 $\textbf{Related Work}$
Early approaches to dialogue coherence modeling are built upon available models for monologue, such as the EntityGrid
(Barzilay and Lapata, 2005, 2008).
model
EntityGrid and its extensions (Burstein et al.,
2010;
2013;
Guinaudeau and Strube,
Mesgar and Strube, 2014; Tien Nguyen and Joty, 2017; Farag and Yannakoudakis, 2019) rely on entity transitions, as proxies of semantic connectivity, between utterances. These approaches are agnostic to discourse properties of dialogues
(Purandare and Litman, 2008; Gandhe and Traum,
2008; Cervone et al., 2018).
Inspired by EntityGrid, Gandhe and Traum (2016) define transition patterns among DA labels associated with utterances to measure coherence. Cervone et al. (2018) combine the above ideas by augmenting entity grids with utterance DA labels. This model restricts utterance vectors only to entity mentions, and needs gold DA labels as its inputs for training as well as evaluation. However, obtaining DA labels from human annotators is expensive and using DAP models makes the performance of coherence model dependent on the performance of DAP models.
Recent approaches to dialogue coherence modeling benefit from distributional representations of utterances. Zhang et al.(2018) quantify the coherence of dialogue using the semantic similarity between each utterance and its preceding utterances. This similarity is estimated, for example, by the cosine similarity between an utterance vector and a context vector where those vectors are the average of their pre-trained word embeddings. Vakulenko et al.(2018) measure dialogue coherence based on the consistency of new concepts introduced in a dialogue with background knowledge. Similarly, Dziri et al. (2019) utilize a natural language inference model to assess the content consistency among utterances as an indicator for dialogue coherence. However, these approaches lack dialogue-relevant information to measure coherence.
Our MTL-based approach solves these issues: (i) it benefits from DAs and semantics of utter-
|
[
[
0.11846923828125,
0.07659912109375,
0.2403564453125,
0.07659912109375,
0.2403564453125,
0.09027099609375,
0.11846923828125,
0.09027099609375
],
[
0.5908203125,
0.07598876953125,
0.65869140625,
0.07781982421875,
0.65869140625,
0.0926513671875,
0.5908203125,
0.09027099609375
],
[
0.8134765625,
0.07659912109375,
0.875,
0.07659912109375,
0.875,
0.0914306640625,
0.8134765625,
0.0914306640625
],
[
0.1361083984375,
0.09320068359375,
0.49072265625,
0.09320068359375,
0.49072265625,
0.10809326171875,
0.1361083984375,
0.10809326171875
],
[
0.525390625,
0.0968017578125,
0.5859375,
0.0968017578125,
0.5859375,
0.10809326171875,
0.525390625,
0.10809326171875
],
[
0.525390625,
0.10809326171875,
0.56201171875,
0.10809326171875,
0.56201171875,
0.119384765625,
0.525390625,
0.119384765625
],
[
0.591796875,
0.1051025390625,
0.7529296875,
0.10687255859375,
0.7529296875,
0.12054443359375,
0.591796875,
0.1181640625
],
[
0.81494140625,
0.10809326171875,
0.85986328125,
0.10809326171875,
0.85986328125,
0.119384765625,
0.81494140625,
0.119384765625
],
[
0.591796875,
0.1175537109375,
0.87158203125,
0.1181640625,
0.87158203125,
0.133056640625,
0.591796875,
0.1318359375
],
[
0.11846923828125,
0.10986328125,
0.489013671875,
0.10986328125,
0.489013671875,
0.12353515625,
0.11846923828125,
0.12353515625
],
[
0.525390625,
0.12054443359375,
0.560546875,
0.12054443359375,
0.560546875,
0.1318359375,
0.525390625,
0.1318359375
],
[
0.11846923828125,
0.1270751953125,
0.49072265625,
0.1270751953125,
0.49072265625,
0.14013671875,
0.11846923828125,
0.14013671875
],
[
0.525390625,
0.1318359375,
0.560546875,
0.1318359375,
0.560546875,
0.1436767578125,
0.525390625,
0.1436767578125
],
[
0.59326171875,
0.1312255859375,
0.6875,
0.1318359375,
0.6875,
0.1436767578125,
0.59326171875,
0.142578125
],
[
0.81494140625,
0.1318359375,
0.861328125,
0.1318359375,
0.861328125,
0.1436767578125,
0.81494140625,
0.1436767578125
],
[
0.11846923828125,
0.142578125,
0.489013671875,
0.142578125,
0.489013671875,
0.1561279296875,
0.11846923828125,
0.1561279296875
],
[
0.525390625,
0.142578125,
0.56494140625,
0.144287109375,
0.56396484375,
0.1561279296875,
0.5234375,
0.15380859375
],
[
0.59326171875,
0.142578125,
0.7392578125,
0.142578125,
0.7392578125,
0.15673828125,
0.59326171875,
0.15673828125
],
[
0.8134765625,
0.142578125,
0.861328125,
0.144287109375,
0.85986328125,
0.1561279296875,
0.8115234375,
0.15380859375
],
[
0.525390625,
0.15380859375,
0.560546875,
0.1561279296875,
0.55859375,
0.168701171875,
0.5234375,
0.166259765625
],
[
0.5947265625,
0.1561279296875,
0.64892578125,
0.1561279296875,
0.64892578125,
0.16748046875,
0.5947265625,
0.16748046875
],
[
0.8134765625,
0.152587890625,
0.86279296875,
0.155029296875,
0.861328125,
0.169189453125,
0.8115234375,
0.16748046875
],
[
0.11846923828125,
0.1728515625,
0.489013671875,
0.17333984375,
0.489013671875,
0.188232421875,
0.11846923828125,
0.18701171875
],
[
0.11846923828125,
0.157958984375,
0.489013671875,
0.157958984375,
0.489013671875,
0.171630859375,
0.11846923828125,
0.171630859375
],
[
0.52587890625,
0.17333984375,
0.59765625,
0.17333984375,
0.59765625,
0.185302734375,
0.52587890625,
0.185302734375
],
[
0.525390625,
0.1859130859375,
0.56640625,
0.1859130859375,
0.56640625,
0.19775390625,
0.525390625,
0.19775390625
],
[
0.59326171875,
0.185302734375,
0.7529296875,
0.185302734375,
0.7529296875,
0.198974609375,
0.59326171875,
0.198974609375
],
[
0.8134765625,
0.1859130859375,
0.861328125,
0.1859130859375,
0.861328125,
0.19775390625,
0.8134765625,
0.19775390625
],
[
0.8134765625,
0.1953125,
0.861328125,
0.19775390625,
0.85986328125,
0.210205078125,
0.8115234375,
0.2078857421875
],
[
0.525390625,
0.196533203125,
0.56494140625,
0.198974609375,
0.56396484375,
0.210205078125,
0.5234375,
0.2078857421875
],
[
0.59326171875,
0.196533203125,
0.73779296875,
0.19775390625,
0.73779296875,
0.208984375,
0.59326171875,
0.2078857421875
],
[
0.12017822265625,
0.2060546875,
0.489013671875,
0.2060546875,
0.489013671875,
0.2197265625,
0.12017822265625,
0.2197265625
],
[
0.59326171875,
0.208984375,
0.6875,
0.210205078125,
0.6875,
0.221435546875,
0.59326171875,
0.2203369140625
],
[
0.525390625,
0.210205078125,
0.560546875,
0.210205078125,
0.560546875,
0.221435546875,
0.525390625,
0.221435546875
],
[
0.11846923828125,
0.190673828125,
0.489013671875,
0.190673828125,
0.489013671875,
0.2037353515625,
0.11846923828125,
0.2037353515625
],
[
0.81494140625,
0.210205078125,
0.861328125,
0.210205078125,
0.861328125,
0.221435546875,
0.81494140625,
0.221435546875
],
[
0.525390625,
0.2197265625,
0.56494140625,
0.221435546875,
0.56396484375,
0.2340087890625,
0.5234375,
0.232177734375
],
[
0.59326171875,
0.2203369140625,
0.86962890625,
0.2203369140625,
0.86962890625,
0.2340087890625,
0.59326171875,
0.2340087890625
],
[
0.81494140625,
0.23095703125,
0.861328125,
0.2327880859375,
0.85986328125,
0.2452392578125,
0.8134765625,
0.243408203125
],
[
0.11846923828125,
0.22265625,
0.489013671875,
0.22265625,
0.489013671875,
0.236328125,
0.11846923828125,
0.236328125
],
[
0.525390625,
0.2340087890625,
0.56201171875,
0.2340087890625,
0.56201171875,
0.2452392578125,
0.525390625,
0.2452392578125
],
[
0.5947265625,
0.2327880859375,
0.64892578125,
0.2327880859375,
0.64892578125,
0.24462890625,
0.5947265625,
0.24462890625
],
[
0.11846923828125,
0.254638671875,
0.489013671875,
0.254638671875,
0.489013671875,
0.268310546875,
0.11846923828125,
0.268310546875
],
[
0.11846923828125,
0.2381591796875,
0.49072265625,
0.2381591796875,
0.49072265625,
0.251708984375,
0.11846923828125,
0.251708984375
],
[
0.51416015625,
0.260009765625,
0.884765625,
0.260009765625,
0.884765625,
0.273681640625,
0.51416015625,
0.273681640625
],
[
0.11846923828125,
0.271484375,
0.49072265625,
0.271484375,
0.49072265625,
0.28515625,
0.11846923828125,
0.28515625
],
[
0.51416015625,
0.27490234375,
0.884765625,
0.27490234375,
0.884765625,
0.2880859375,
0.51416015625,
0.2880859375
],
[
0.11846923828125,
0.287353515625,
0.489013671875,
0.287353515625,
0.489013671875,
0.300537109375,
0.11846923828125,
0.300537109375
],
[
0.51416015625,
0.289306640625,
0.884765625,
0.289306640625,
0.884765625,
0.302734375,
0.51416015625,
0.302734375
],
[
0.12017822265625,
0.3046875,
0.283935546875,
0.3046875,
0.283935546875,
0.3154296875,
0.12017822265625,
0.3154296875
],
[
0.5126953125,
0.302734375,
0.88671875,
0.302734375,
0.88671875,
0.31640625,
0.5126953125,
0.31640625
],
[
0.1361083984375,
0.318359375,
0.489013671875,
0.319580078125,
0.489013671875,
0.334228515625,
0.1361083984375,
0.333251953125
],
[
0.51416015625,
0.317626953125,
0.87158203125,
0.317626953125,
0.87158203125,
0.330810546875,
0.51416015625,
0.330810546875
],
[
0.11846923828125,
0.336181640625,
0.489013671875,
0.336181640625,
0.489013671875,
0.349853515625,
0.11846923828125,
0.349853515625
],
[
0.11846923828125,
0.350830078125,
0.489013671875,
0.3515625,
0.489013671875,
0.366455078125,
0.11846923828125,
0.365234375
],
[
0.5126953125,
0.35986328125,
0.88671875,
0.35986328125,
0.88671875,
0.37353515625,
0.5126953125,
0.37353515625
],
[
0.11846923828125,
0.36767578125,
0.4873046875,
0.36767578125,
0.4873046875,
0.38134765625,
0.11846923828125,
0.38134765625
],
[
0.51416015625,
0.3759765625,
0.72119140625,
0.3759765625,
0.72119140625,
0.39013671875,
0.51416015625,
0.39013671875
],
[
0.1168212890625,
0.383056640625,
0.4873046875,
0.383056640625,
0.4873046875,
0.396728515625,
0.1168212890625,
0.396728515625
],
[
0.53173828125,
0.393798828125,
0.884765625,
0.393798828125,
0.884765625,
0.40673828125,
0.53173828125,
0.40673828125
],
[
0.12017822265625,
0.402587890625,
0.4873046875,
0.402587890625,
0.4873046875,
0.413330078125,
0.12017822265625,
0.413330078125
],
[
0.1168212890625,
0.4150390625,
0.49072265625,
0.416259765625,
0.49072265625,
0.43115234375,
0.1168212890625,
0.429931640625
],
[
0.51416015625,
0.410400390625,
0.884765625,
0.410400390625,
0.884765625,
0.42333984375,
0.51416015625,
0.42333984375
],
[
0.51416015625,
0.42578125,
0.88671875,
0.42578125,
0.88671875,
0.439453125,
0.51416015625,
0.439453125
],
[
0.121826171875,
0.435302734375,
0.49072265625,
0.435302734375,
0.49072265625,
0.4453125,
0.121826171875,
0.4453125
],
[
0.51611328125,
0.443603515625,
0.884765625,
0.443603515625,
0.884765625,
0.45361328125,
0.51611328125,
0.45361328125
],
[
0.12017822265625,
0.449462890625,
0.49072265625,
0.449462890625,
0.49072265625,
0.463134765625,
0.12017822265625,
0.463134765625
],
[
0.51416015625,
0.457763671875,
0.88818359375,
0.457763671875,
0.88818359375,
0.471435546875,
0.51416015625,
0.471435546875
],
[
0.11846923828125,
0.465576171875,
0.49072265625,
0.465576171875,
0.49072265625,
0.478515625,
0.11846923828125,
0.478515625
],
[
0.5126953125,
0.47265625,
0.88818359375,
0.474365234375,
0.88818359375,
0.4892578125,
0.5126953125,
0.48681640625
],
[
0.11846923828125,
0.482177734375,
0.492431640625,
0.482177734375,
0.492431640625,
0.495361328125,
0.11846923828125,
0.495361328125
],
[
0.5126953125,
0.490478515625,
0.88671875,
0.490478515625,
0.88671875,
0.50439453125,
0.5126953125,
0.50439453125
],
[
0.1168212890625,
0.49755859375,
0.492431640625,
0.49755859375,
0.492431640625,
0.51123046875,
0.1168212890625,
0.51123046875
],
[
0.5126953125,
0.50732421875,
0.88671875,
0.50732421875,
0.88671875,
0.52099609375,
0.5126953125,
0.52099609375
],
[
0.1168212890625,
0.51318359375,
0.492431640625,
0.51318359375,
0.492431640625,
0.52783203125,
0.1168212890625,
0.52783203125
],
[
0.51171875,
0.521484375,
0.884765625,
0.5224609375,
0.884765625,
0.53759765625,
0.51171875,
0.5361328125
],
[
0.1168212890625,
0.529296875,
0.4873046875,
0.529296875,
0.4873046875,
0.5419921875,
0.1168212890625,
0.5419921875
],
[
0.51416015625,
0.5390625,
0.88671875,
0.5390625,
0.88671875,
0.552734375,
0.51416015625,
0.552734375
],
[
0.12017822265625,
0.5458984375,
0.4931640625,
0.5458984375,
0.4931640625,
0.55859375,
0.12017822265625,
0.55859375
],
[
0.5126953125,
0.55419921875,
0.88671875,
0.5546875,
0.88671875,
0.5693359375,
0.5126953125,
0.568359375
],
[
0.12017822265625,
0.5625,
0.49072265625,
0.5625,
0.49072265625,
0.576171875,
0.12017822265625,
0.576171875
],
[
0.5126953125,
0.5693359375,
0.71240234375,
0.5693359375,
0.71240234375,
0.5830078125,
0.5126953125,
0.5830078125
],
[
0.1168212890625,
0.57666015625,
0.4873046875,
0.57666015625,
0.4873046875,
0.59033203125,
0.1168212890625,
0.59033203125
],
[
0.53173828125,
0.58837890625,
0.88671875,
0.58837890625,
0.88671875,
0.6015625,
0.53173828125,
0.6015625
],
[
0.121826171875,
0.595703125,
0.489013671875,
0.595703125,
0.489013671875,
0.60546875,
0.121826171875,
0.60546875
],
[
0.11846923828125,
0.609375,
0.489013671875,
0.609375,
0.489013671875,
0.62255859375,
0.11846923828125,
0.62255859375
],
[
0.51416015625,
0.60400390625,
0.88671875,
0.60400390625,
0.88671875,
0.61767578125,
0.51416015625,
0.61767578125
],
[
0.51416015625,
0.62060546875,
0.88818359375,
0.62060546875,
0.88818359375,
0.63427734375,
0.51416015625,
0.63427734375
],
[
0.12017822265625,
0.626953125,
0.49072265625,
0.626953125,
0.49072265625,
0.64013671875,
0.12017822265625,
0.64013671875
],
[
0.5126953125,
0.63525390625,
0.88818359375,
0.6357421875,
0.88818359375,
0.65087890625,
0.5126953125,
0.6494140625
],
[
0.11846923828125,
0.642578125,
0.489013671875,
0.642578125,
0.489013671875,
0.65625,
0.11846923828125,
0.65625
],
[
0.11846923828125,
0.65771484375,
0.489013671875,
0.65771484375,
0.489013671875,
0.67138671875,
0.11846923828125,
0.67138671875
],
[
0.51416015625,
0.65283203125,
0.88671875,
0.65283203125,
0.88671875,
0.66650390625,
0.51416015625,
0.66650390625
],
[
0.51611328125,
0.66943359375,
0.88671875,
0.66943359375,
0.88671875,
0.68310546875,
0.51611328125,
0.68310546875
],
[
0.12017822265625,
0.67578125,
0.2587890625,
0.67578125,
0.2587890625,
0.68603515625,
0.12017822265625,
0.68603515625
],
[
0.51416015625,
0.68505859375,
0.884765625,
0.68505859375,
0.884765625,
0.69873046875,
0.51416015625,
0.69873046875
],
[
0.1226806640625,
0.705078125,
0.1378173828125,
0.71044921875,
0.1319580078125,
0.7197265625,
0.1168212890625,
0.71435546875
],
[
0.1470947265625,
0.7060546875,
0.272216796875,
0.7060546875,
0.272216796875,
0.7197265625,
0.1470947265625,
0.7197265625
],
[
0.51416015625,
0.70068359375,
0.884765625,
0.70068359375,
0.884765625,
0.71435546875,
0.51416015625,
0.71435546875
],
[
0.51416015625,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.51416015625,
0.73095703125
],
[
0.267333984375,
0.73388671875,
0.39404296875,
0.73388671875,
0.39404296875,
0.74755859375,
0.267333984375,
0.74755859375
],
[
0.51416015625,
0.73291015625,
0.88671875,
0.73291015625,
0.88671875,
0.74658203125,
0.51416015625,
0.74658203125
],
[
0.1781005859375,
0.7353515625,
0.2763671875,
0.73291015625,
0.2763671875,
0.74658203125,
0.1781005859375,
0.7490234375
],
[
0.12017822265625,
0.7353515625,
0.1605224609375,
0.7353515625,
0.1605224609375,
0.74658203125,
0.12017822265625,
0.74658203125
],
[
0.407470703125,
0.7353515625,
0.4873046875,
0.7353515625,
0.4873046875,
0.74658203125,
0.407470703125,
0.74658203125
],
[
0.1168212890625,
0.75,
0.43701171875,
0.7490234375,
0.43701171875,
0.76318359375,
0.1168212890625,
0.76416015625
],
[
0.429443359375,
0.751953125,
0.489013671875,
0.751953125,
0.489013671875,
0.7626953125,
0.429443359375,
0.7626953125
],
[
0.51416015625,
0.75048828125,
0.88232421875,
0.75048828125,
0.88232421875,
0.76123046875,
0.51416015625,
0.76123046875
],
[
0.1168212890625,
0.765625,
0.489013671875,
0.765625,
0.489013671875,
0.779296875,
0.1168212890625,
0.779296875
],
[
0.51416015625,
0.765625,
0.884765625,
0.765625,
0.884765625,
0.779296875,
0.51416015625,
0.779296875
],
[
0.1873779296875,
0.783203125,
0.363037109375,
0.783203125,
0.363037109375,
0.79638671875,
0.1873779296875,
0.79638671875
],
[
0.373046875,
0.783203125,
0.421875,
0.783203125,
0.421875,
0.79443359375,
0.373046875,
0.79443359375
],
[
0.4404296875,
0.783203125,
0.49072265625,
0.783203125,
0.49072265625,
0.79443359375,
0.4404296875,
0.79443359375
],
[
0.5126953125,
0.78076171875,
0.88671875,
0.78076171875,
0.88671875,
0.79443359375,
0.5126953125,
0.79443359375
],
[
0.5126953125,
0.79736328125,
0.884765625,
0.79736328125,
0.884765625,
0.81103515625,
0.5126953125,
0.81103515625
],
[
0.12017822265625,
0.78369140625,
0.1705322265625,
0.78369140625,
0.1705322265625,
0.7958984375,
0.12017822265625,
0.7958984375
],
[
0.12017822265625,
0.798828125,
0.37890625,
0.798828125,
0.37890625,
0.8125,
0.12017822265625,
0.8125
],
[
0.371337890625,
0.80029296875,
0.492431640625,
0.80029296875,
0.492431640625,
0.81103515625,
0.371337890625,
0.81103515625
],
[
0.12017822265625,
0.8125,
0.1646728515625,
0.81396484375,
0.1630859375,
0.82763671875,
0.11846923828125,
0.826171875
],
[
0.44287109375,
0.8125,
0.4931640625,
0.81396484375,
0.492431640625,
0.8291015625,
0.441162109375,
0.82666015625
],
[
0.5126953125,
0.81396484375,
0.88671875,
0.81396484375,
0.88671875,
0.82763671875,
0.5126953125,
0.82763671875
],
[
0.11846923828125,
0.830078125,
0.4931640625,
0.83056640625,
0.4931640625,
0.845703125,
0.11846923828125,
0.84423828125
],
[
0.212646484375,
0.8154296875,
0.3984375,
0.8154296875,
0.3984375,
0.8291015625,
0.212646484375,
0.8291015625
],
[
0.51416015625,
0.830078125,
0.884765625,
0.830078125,
0.884765625,
0.84423828125,
0.51416015625,
0.84423828125
],
[
0.12017822265625,
0.8466796875,
0.492431640625,
0.8466796875,
0.492431640625,
0.86083984375,
0.12017822265625,
0.86083984375
],
[
0.51416015625,
0.8466796875,
0.88671875,
0.8466796875,
0.88671875,
0.85986328125,
0.51416015625,
0.85986328125
],
[
0.11846923828125,
0.86328125,
0.4931640625,
0.86328125,
0.4931640625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.51416015625,
0.86328125,
0.5771484375,
0.86328125,
0.5771484375,
0.87451171875,
0.51416015625,
0.87451171875
],
[
0.1168212890625,
0.87744140625,
0.489013671875,
0.87890625,
0.489013671875,
0.8935546875,
0.1168212890625,
0.892578125
],
[
0.53369140625,
0.87890625,
0.88671875,
0.87890625,
0.88671875,
0.8935546875,
0.53369140625,
0.8935546875
],
[
0.11846923828125,
0.8955078125,
0.49072265625,
0.89453125,
0.49072265625,
0.9091796875,
0.11846923828125,
0.91015625
],
[
0.51416015625,
0.8955078125,
0.88671875,
0.8955078125,
0.88671875,
0.9091796875,
0.51416015625,
0.9091796875
]
] |
[
"input DA labels.",
"Utterance",
"DA label",
"We propose a novel approach to dialogue co-",
"coherent",
"uttr:",
"This is my uncle, Charles.",
"inform",
"He looks strong. What does he do? question",
"herence assessment by utilizing dialogue act pre-",
"utt2",
"diction as an auxiliary task for training our coher-",
"utts:",
"He's a captain.",
"inform",
"ence model in a multi-task learning (MTL) sce-",
"utt4:",
"He must be very brave.",
"inform",
"utts:",
"Exactly!",
"inform",
"high-level components: an utterance encoder, a",
"nario (Figure 1). Our approach consists of three",
"incoherent",
"uttr::",
"This is my uncle, Charles.",
"inform",
"inform",
"utt4:",
"He must be very brave.",
"logue Act Prediction (DAP) model. The layers of",
"He's a captain.",
"utts:",
"dialogue coherence model (DiCoh), and a Dia",
"inform",
"utt2:",
"He looks strong. What does he do? question",
"inform",
"the utterance encoder are shared between the DAP",
"utts:",
"Exactly!",
"Coh model to learn to focus on salient information",
"and the DiCoh model. This idea enables our Di-",
"Table l: An example dialogue from DailyDialog (top)",
"presented in utterances considering their DAs and",
"and its perturbation (bottom), which is generated by",
"to alleviate the need for explicit DA labels during",
"permuting the utterances said by one of the speakers",
"coherenceassessment.",
"(shown in boldface), and is less coherent. The right col-",
"We evaluate our MTL-based approach on the",
"umn shows the DA labels associated with utterances.",
"DailyDialog (Liet al., 2017) and SwitchBoard",
"(Jurafsky and Shriberg,1997) English dialogue",
"(Purandare and Litman, 2008; Gandhe and Traum,",
"corpora in several discriminating experiments.",
"2008; Cervone et al., 2018).",
"where our coherence model, DiCoh, is examined",
"Inspired by EntityGrid, Gandhe and Traum",
"to discriminate a dialogue from its perturbations",
"(see Table 1). We utilize perturbation methods,",
"(2016) define transition patterns among DA labels",
"associated with utterances to measure coherence.",
"likeutteranceorderingandutteranceinsertion,in-",
"Cervoneetal.(2018)combine the above ideasby",
"herited from coherence evaluation approaches for",
"augmenting entity grids with utterance DA labels.",
"monologue texts, and also introduce two dialogue-",
"This model restricts utterance vectors only to en-",
"relevant perturbations, named utterance replace-",
"tity mentions, and needs gold DA labels as its in-",
"ment and even utterance ordering. Our core con-",
"puts for training as well as evaluation. However,",
"tributions are: (1) proposing an MTL-based ap-",
" obtaining DA labels from human annotators is ex-",
"proach for dialogue coherence assessment using",
"pensive and using DAP models makes the perfor-",
"DAP as an auxiliary task, yielding more informa-",
"mance of coherence model dependent on the per-",
"tive utterance representations for coherence assess-",
"formance of DAP models..",
"ment: (2) alleviating the need for DA labels for di.",
"Recent approaches to dialogue coherence mod-",
"alogue coherence assessment during evaluations;",
"(3) an empirical evaluation on two benchmark di-",
"eling benefit from distributional representations of",
"utterances. Zhang et al. (2018) quantify the co-",
"alogue corpora, showing that our model substan-",
"herence of dialogue using the semantic similar-",
"tially outperforms the state-of-the-art coherence",
"model on DailyDialog, and performs on par with",
"ity between each utterance and its preceding utter-",
"ances. This similarity is estimated, for example,",
"it on SwitchBoard.",
"by the cosine similarity between an utterance vec-",
"2",
"Related Work",
"tor and a context vector where those vectors are",
"the average of their pre-trained word embeddings.",
"todialogue",
"Vakulenko et al. (2018) measure dialogue coher-",
"approaches",
"Early",
"coherence",
"modelingarebuiltuponavailable",
"mod-",
"ence based on theconsistency of new concepts",
"els for monologue, such as the EntityGrid",
"introduced in a dialogue with background knowl-",
"(Barzilay and Lapata,",
"2005.",
"2008).",
"edge. Similarly, Dziri et al. (2019) utilize a natu-",
"ral language inference model to assess the content",
"model",
"EntityGrid and its extensions",
"(Burstein et al.,",
"2010;",
"2013;",
"consistency among utterances as an indicator for",
"Mesgar and Strube, 2014; Tien Nguyen and Joty,",
"Guinaudeau and Strube,",
"dialogue coherence. However, these approaches",
"2017; Farag and Yannakoudakis, 2019) rely on",
"lack dialogue-relevant information to measure co-",
"entity transitions, as proxies of semantic con-",
"herence.",
"nectivity, between utterances. These approaches",
"Our MTL-based approach solves these issues:",
"are agnostic to discourse properties of dialogues",
"(i) it benefits from DAs and semantics of utter-"
] |
{
"arxiv_id": "1908.08486",
"language": "en",
"page_idx": 1
}
|
||
1908.08486-en-2
|
1908.08486-en-2.jpg
|
ances to measure dialogue coherence by optimizing utterance vectors for both DAP and coherence assessment, and (ii) it uses DA labels to define an auxiliary task for training the DiCoh model using MTL, instead of utilizing them in a pipeline. Therefore, it efficiently mitigates the need for explicit DA labels as inputs during coherence assessment.
## 3 $\textbf{Method}$
We represent a dialogue between two speakers as a sequence of utterances, $dial=[utt_1,...,utt_m].$ We address the problem of designing a coherence model, DiCoh, which assigns a coherence score to $dial,s_{dial}=DiCoh(dial).$ Given a pair of dialogues $\phi = ( dial_{i}, dial_{j}) $,our DiCoh model ideally assigns $s_dial_i>s_{dial_j}$ if and only if dialogue $dial_i$ is preferred over dialogue $dial_j$ according to their perceived coherence. Instead of using gold DA labels as inputs to DiCoh, we use them to define an auxiliary task and model, DAP, to enrich utterance vectors for DiCoh in an MTL scenario. Figure 2 shows a low-level illustration of our MTL-based approach.
Utterance encoder We use a word embedding layer, $Emb$, to transform the words in utterance $utt~=~[w_1,...,w_n]$ to a sequence of embedding vectors $E=[e_{1},...,e_{n}]$, where $n$ is the number of words in $utt.$ The embedding layer can be initialized by any pre-trained embeddings to capture lexical relations. We use a Bidirectional recurrent neural network with Long Short-Term Memory cells, $BiLSTM$, to map embeddings $E$ to encode words in their utterance-level context:
$$\begin{aligned}E&=Emb(utt),\\H_u&=BiLSTM(E),\end{aligned}$$
(1)
where $H_u$ shows the hidden state vectors $[h_{1}^{u},...,h_{n}^{u}]$ returned by $BiLSTM.$ At word $t,h_{t}^{u}$ is $\begin{array}{c}{\text{the concatenation of hidden states of the forward}}\\{\overrightarrow{h_{t}^{u}}\mathrm{~and~the~backward~LSTMs~}\vec{h_{t}^{u}}:}\end{array}$
$$h_t^u=[\overrightarrow{h_t^u};\overleftarrow{h_t^u}].$$
(2)
We apply a self-attention mechanism, Atten, to the hidden state vectors in $H_u$ to obtain the vector representation, $u$, of utterance $utt:$
(3)
$$u=Atten(H_u).$$
Generally, the attention layer, Atten, for an input
vector $x$ is defined as follows:
(4)
$$\begin{aligned}\beta_{t}&=x_{t}*W,\\\alpha_{t}&=\frac{\exp\left(\beta_{t}\right)}{\sum_{t}\exp\left(\beta_{t}\right)},\\o&=\sum_{t}\alpha_{t}*x_{t},\end{aligned}$$
where $W$ is the parameter of this layer, and $o$ is its weighted output vector. Attention enables the utterance representation layer to encode an utterance by the weighted sum of its word embeddings. It is worth noting that the parameters of the utterance encoder are shared for representing all utterances in a dialogue.
DiCoh model For an input dialogue $dial= [ utt_1, . . . , utt_m] $, the output of the utterance representation encoder is a sequence of vectors, i.e., $[u_1,...,u_m].$ Our coherence assessment model (DiCoh) combines these vectors by a BiLSTM to obtain dialogue-level contextualized representations of utterances. Then, a self-attention (Equation 4) with new parameters computes the weighted average of the contextualized utterance vectors to encode the dialogue:
(5)
$$\begin{aligned}[h_{1}^{d},...,h_{m}^{d}]&=BiLSTM([u_{1},...,u_{m}]),\\d&=Atten([h_{1}^{d},...,h_{m}^{d}]).\end{aligned}$$
A linear feed-forward layer, $FF$, maps the dialogue vector, $d$, to a dialogue coherence score, $s_{dial}\colon$
(6)
$$s_{dial}=FF(d).$$
DAP model Our DAP model, which is used to solve the auxiliary DAP task, is a softmax layer which maps an utterance vector, $u$, to a probability distribution $p_a$ over DA labels $A:$
$$p_{a}(u)=softmax(W_{|u|\times|A|}*u+b),$$
(7)
where $W_|u|\times|A|$ shows the weights of the softmax layer, $|u|$ is the size of the utterance vector, $|A|$ is the number of DA labels, and $b$ is the bias.
## 3.1 Multi-Task Learning
As illustrated in Figure 1, our main idea is to benefit from the DAP task for improving the performance of the dialogue coherence model by using them in a multi-task learning scenario. We also assume that each utterance $utt_k$ is associated with
|
[
[
0.11846923828125,
0.07598876953125,
0.489013671875,
0.07598876953125,
0.489013671875,
0.08905029296875,
0.11846923828125,
0.08905029296875
],
[
0.51416015625,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.51416015625,
0.09027099609375
],
[
0.12017822265625,
0.0926513671875,
0.489013671875,
0.0926513671875,
0.489013671875,
0.10687255859375,
0.12017822265625,
0.10687255859375
],
[
0.5126953125,
0.0926513671875,
0.73779296875,
0.0926513671875,
0.73779296875,
0.105712890625,
0.5126953125,
0.105712890625
],
[
0.1168212890625,
0.10809326171875,
0.489013671875,
0.10687255859375,
0.489013671875,
0.1217041015625,
0.1168212890625,
0.122314453125
],
[
0.6279296875,
0.114013671875,
0.728515625,
0.114013671875,
0.728515625,
0.1318359375,
0.6279296875,
0.1318359375
],
[
0.1168212890625,
0.12469482421875,
0.4873046875,
0.12469482421875,
0.4873046875,
0.1383056640625,
0.1168212890625,
0.1383056640625
],
[
0.68408203125,
0.1353759765625,
0.75146484375,
0.1353759765625,
0.75146484375,
0.15087890625,
0.68408203125,
0.15087890625
],
[
0.12017822265625,
0.141357421875,
0.489013671875,
0.141357421875,
0.489013671875,
0.155029296875,
0.12017822265625,
0.155029296875
],
[
0.6259765625,
0.1478271484375,
0.6552734375,
0.1478271484375,
0.6552734375,
0.1591796875,
0.6259765625,
0.1591796875
],
[
0.67041015625,
0.152587890625,
0.7705078125,
0.15087890625,
0.7705078125,
0.168701171875,
0.67041015625,
0.17041015625
],
[
0.85791015625,
0.15087890625,
0.88818359375,
0.15087890625,
0.88818359375,
0.16748046875,
0.85791015625,
0.16748046875
],
[
0.12017822265625,
0.1728515625,
0.489013671875,
0.1728515625,
0.489013671875,
0.18701171875,
0.12017822265625,
0.18701171875
],
[
0.11846923828125,
0.1561279296875,
0.4873046875,
0.1561279296875,
0.4873046875,
0.169189453125,
0.11846923828125,
0.169189453125
],
[
0.6396484375,
0.18115234375,
0.65283203125,
0.18115234375,
0.65283203125,
0.18701171875,
0.6396484375,
0.18701171875
],
[
0.693359375,
0.1787109375,
0.7587890625,
0.1787109375,
0.7587890625,
0.190673828125,
0.693359375,
0.190673828125
],
[
0.11846923828125,
0.190673828125,
0.1622314453125,
0.190673828125,
0.1622314453125,
0.201904296875,
0.11846923828125,
0.201904296875
],
[
0.115966796875,
0.218505859375,
0.22021484375,
0.2203369140625,
0.22021484375,
0.236328125,
0.115966796875,
0.2340087890625
],
[
0.51416015625,
0.213134765625,
0.884765625,
0.213134765625,
0.884765625,
0.22802734375,
0.51416015625,
0.22802734375
],
[
0.51171875,
0.23095703125,
0.884765625,
0.23095703125,
0.884765625,
0.24462890625,
0.51171875,
0.24462890625
],
[
0.11846923828125,
0.2493896484375,
0.489013671875,
0.2493896484375,
0.489013671875,
0.26318359375,
0.11846923828125,
0.26318359375
],
[
0.51416015625,
0.2464599609375,
0.884765625,
0.2464599609375,
0.884765625,
0.260009765625,
0.51416015625,
0.260009765625
],
[
0.12017822265625,
0.265380859375,
0.492431640625,
0.265380859375,
0.492431640625,
0.27978515625,
0.12017822265625,
0.27978515625
],
[
0.51416015625,
0.26318359375,
0.884765625,
0.26318359375,
0.884765625,
0.276611328125,
0.51416015625,
0.276611328125
],
[
0.11846923828125,
0.281982421875,
0.49072265625,
0.281982421875,
0.49072265625,
0.295654296875,
0.11846923828125,
0.295654296875
],
[
0.51171875,
0.27783203125,
0.88330078125,
0.27783203125,
0.88330078125,
0.29150390625,
0.51171875,
0.29150390625
],
[
0.11846923828125,
0.297607421875,
0.49072265625,
0.297607421875,
0.49072265625,
0.311279296875,
0.11846923828125,
0.311279296875
],
[
0.5126953125,
0.295654296875,
0.88330078125,
0.295654296875,
0.88330078125,
0.308837890625,
0.5126953125,
0.308837890625
],
[
0.11846923828125,
0.3134765625,
0.492431640625,
0.3134765625,
0.492431640625,
0.327880859375,
0.11846923828125,
0.327880859375
],
[
0.51416015625,
0.312255859375,
0.6142578125,
0.312255859375,
0.6142578125,
0.32373046875,
0.51416015625,
0.32373046875
],
[
0.11846923828125,
0.330078125,
0.489013671875,
0.330078125,
0.489013671875,
0.34326171875,
0.11846923828125,
0.34326171875
],
[
0.51416015625,
0.336181640625,
0.572265625,
0.336181640625,
0.572265625,
0.347412109375,
0.51416015625,
0.347412109375
],
[
0.5859375,
0.336181640625,
0.693359375,
0.336181640625,
0.693359375,
0.347412109375,
0.5859375,
0.347412109375
],
[
0.748046875,
0.33740234375,
0.79833984375,
0.33740234375,
0.79833984375,
0.3486328125,
0.748046875,
0.3486328125
],
[
0.81494140625,
0.336181640625,
0.88330078125,
0.336181640625,
0.88330078125,
0.347412109375,
0.81494140625,
0.347412109375
],
[
0.7421875,
0.341552734375,
0.74951171875,
0.341552734375,
0.74951171875,
0.344482421875,
0.7421875,
0.344482421875
],
[
0.70947265625,
0.33837890625,
0.73193359375,
0.33837890625,
0.73193359375,
0.3466796875,
0.70947265625,
0.3466796875
],
[
0.115966796875,
0.344482421875,
0.49072265625,
0.342529296875,
0.49072265625,
0.35986328125,
0.115966796875,
0.3623046875
],
[
0.5126953125,
0.3515625,
0.884765625,
0.3515625,
0.884765625,
0.366455078125,
0.5126953125,
0.366455078125
],
[
0.11846923828125,
0.3623046875,
0.489013671875,
0.3623046875,
0.489013671875,
0.3759765625,
0.11846923828125,
0.3759765625
],
[
0.51416015625,
0.3681640625,
0.88671875,
0.3681640625,
0.88671875,
0.3818359375,
0.51416015625,
0.3818359375
],
[
0.11846923828125,
0.377685546875,
0.489013671875,
0.377197265625,
0.489013671875,
0.391357421875,
0.11846923828125,
0.392578125
],
[
0.51416015625,
0.38427734375,
0.88671875,
0.38427734375,
0.88671875,
0.3984375,
0.51416015625,
0.3984375
],
[
0.11846923828125,
0.394287109375,
0.489013671875,
0.394287109375,
0.489013671875,
0.407958984375,
0.11846923828125,
0.407958984375
],
[
0.51416015625,
0.399658203125,
0.884765625,
0.399658203125,
0.884765625,
0.413330078125,
0.51416015625,
0.413330078125
],
[
0.12017822265625,
0.410400390625,
0.489013671875,
0.410400390625,
0.489013671875,
0.42333984375,
0.12017822265625,
0.42333984375
],
[
0.51416015625,
0.4150390625,
0.884765625,
0.4150390625,
0.884765625,
0.4287109375,
0.51416015625,
0.4287109375
],
[
0.11846923828125,
0.427001953125,
0.489013671875,
0.427001953125,
0.489013671875,
0.440673828125,
0.11846923828125,
0.440673828125
],
[
0.51416015625,
0.431640625,
0.80908203125,
0.431640625,
0.80908203125,
0.4453125,
0.51416015625,
0.4453125
],
[
0.81787109375,
0.432861328125,
0.884765625,
0.432861328125,
0.884765625,
0.44482421875,
0.81787109375,
0.44482421875
],
[
0.1168212890625,
0.4423828125,
0.4873046875,
0.4423828125,
0.4873046875,
0.4560546875,
0.1168212890625,
0.4560546875
],
[
0.51611328125,
0.448974609375,
0.884765625,
0.448974609375,
0.884765625,
0.4619140625,
0.51611328125,
0.4619140625
],
[
0.12017822265625,
0.460205078125,
0.191650390625,
0.460205078125,
0.191650390625,
0.471435546875,
0.12017822265625,
0.471435546875
],
[
0.5126953125,
0.46435546875,
0.88671875,
0.46435546875,
0.88671875,
0.47802734375,
0.5126953125,
0.47802734375
],
[
0.51611328125,
0.479736328125,
0.88671875,
0.479736328125,
0.88671875,
0.493408203125,
0.51611328125,
0.493408203125
],
[
0.12017822265625,
0.48681640625,
0.4873046875,
0.48681640625,
0.4873046875,
0.50048828125,
0.12017822265625,
0.50048828125
],
[
0.12017822265625,
0.50439453125,
0.489013671875,
0.50439453125,
0.489013671875,
0.51708984375,
0.12017822265625,
0.51708984375
],
[
0.560546875,
0.50732421875,
0.8359375,
0.50830078125,
0.8359375,
0.525390625,
0.560546875,
0.52490234375
],
[
0.1168212890625,
0.51953125,
0.4873046875,
0.5185546875,
0.4873046875,
0.533203125,
0.1168212890625,
0.53369140625
],
[
0.861328125,
0.52099609375,
0.88330078125,
0.52099609375,
0.88330078125,
0.53369140625,
0.861328125,
0.53369140625
],
[
0.6318359375,
0.52783203125,
0.8115234375,
0.529296875,
0.8115234375,
0.54638671875,
0.6318359375,
0.5458984375
],
[
0.11846923828125,
0.5361328125,
0.489013671875,
0.5361328125,
0.489013671875,
0.5498046875,
0.11846923828125,
0.5498046875
],
[
0.11846923828125,
0.5517578125,
0.489013671875,
0.5517578125,
0.489013671875,
0.5654296875,
0.11846923828125,
0.5654296875
],
[
0.51416015625,
0.552734375,
0.884765625,
0.552734375,
0.884765625,
0.56640625,
0.51416015625,
0.56640625
],
[
0.12017822265625,
0.568359375,
0.489013671875,
0.568359375,
0.489013671875,
0.58203125,
0.12017822265625,
0.58203125
],
[
0.51416015625,
0.5693359375,
0.88671875,
0.5693359375,
0.88671875,
0.5830078125,
0.51416015625,
0.5830078125
],
[
0.11846923828125,
0.5830078125,
0.4873046875,
0.5830078125,
0.4873046875,
0.5966796875,
0.11846923828125,
0.5966796875
],
[
0.51611328125,
0.58740234375,
0.55810546875,
0.58740234375,
0.55810546875,
0.5986328125,
0.51611328125,
0.5986328125
],
[
0.12017822265625,
0.60107421875,
0.486572265625,
0.60107421875,
0.486572265625,
0.61083984375,
0.12017822265625,
0.61083984375
],
[
0.64111328125,
0.61328125,
0.7529296875,
0.61083984375,
0.7529296875,
0.62890625,
0.64111328125,
0.63134765625
],
[
0.85791015625,
0.61328125,
0.88818359375,
0.61328125,
0.88818359375,
0.6298828125,
0.85791015625,
0.6298828125
],
[
0.121826171875,
0.63427734375,
0.435302734375,
0.63427734375,
0.435302734375,
0.64453125,
0.121826171875,
0.64453125
],
[
0.12017822265625,
0.6162109375,
0.489013671875,
0.6162109375,
0.489013671875,
0.6298828125,
0.12017822265625,
0.6298828125
],
[
0.51416015625,
0.6435546875,
0.884765625,
0.6435546875,
0.884765625,
0.65673828125,
0.51416015625,
0.65673828125
],
[
0.239501953125,
0.6591796875,
0.35546875,
0.6611328125,
0.35546875,
0.6767578125,
0.239501953125,
0.6748046875
],
[
0.51416015625,
0.6591796875,
0.88232421875,
0.6591796875,
0.88232421875,
0.6728515625,
0.51416015625,
0.6728515625
],
[
0.465576171875,
0.67138671875,
0.49072265625,
0.67138671875,
0.49072265625,
0.68505859375,
0.465576171875,
0.68505859375
],
[
0.230224609375,
0.6796875,
0.377197265625,
0.6796875,
0.377197265625,
0.69482421875,
0.230224609375,
0.69482421875
],
[
0.5126953125,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.689453125,
0.5126953125,
0.689453125
],
[
0.51416015625,
0.6923828125,
0.76318359375,
0.6923828125,
0.76318359375,
0.70703125,
0.51416015625,
0.70703125
],
[
0.1168212890625,
0.71142578125,
0.489013671875,
0.7119140625,
0.489013671875,
0.72705078125,
0.1168212890625,
0.7255859375
],
[
0.56201171875,
0.7197265625,
0.8388671875,
0.7197265625,
0.8388671875,
0.74072265625,
0.56201171875,
0.74072265625
],
[
0.12017822265625,
0.72802734375,
0.4873046875,
0.72802734375,
0.4873046875,
0.7421875,
0.12017822265625,
0.7421875
],
[
0.861328125,
0.724609375,
0.88330078125,
0.724609375,
0.88330078125,
0.7353515625,
0.861328125,
0.7353515625
],
[
0.12017822265625,
0.74462890625,
0.489013671875,
0.74365234375,
0.489013671875,
0.75830078125,
0.12017822265625,
0.7587890625
],
[
0.51025390625,
0.74755859375,
0.88671875,
0.7490234375,
0.88671875,
0.76953125,
0.51025390625,
0.7685546875
],
[
0.11846923828125,
0.7587890625,
0.3671875,
0.7587890625,
0.3671875,
0.77685546875,
0.11846923828125,
0.77685546875
],
[
0.51416015625,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.78076171875,
0.51416015625,
0.78076171875
],
[
0.5126953125,
0.7822265625,
0.82666015625,
0.783203125,
0.82666015625,
0.79736328125,
0.5126953125,
0.79638671875
],
[
0.245361328125,
0.79150390625,
0.359619140625,
0.7890625,
0.359619140625,
0.81103515625,
0.245361328125,
0.8134765625
],
[
0.462158203125,
0.79443359375,
0.492431640625,
0.79443359375,
0.492431640625,
0.81103515625,
0.462158203125,
0.81103515625
],
[
0.5126953125,
0.80859375,
0.724609375,
0.81005859375,
0.724609375,
0.82470703125,
0.5126953125,
0.82373046875
],
[
0.11846923828125,
0.830078125,
0.49072265625,
0.830078125,
0.49072265625,
0.84423828125,
0.11846923828125,
0.84423828125
],
[
0.51416015625,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.51416015625,
0.84423828125
],
[
0.1168212890625,
0.84423828125,
0.49072265625,
0.8466796875,
0.49072265625,
0.86083984375,
0.1168212890625,
0.859375
],
[
0.51171875,
0.845703125,
0.88671875,
0.8466796875,
0.88671875,
0.86083984375,
0.51171875,
0.85986328125
],
[
0.11846923828125,
0.86328125,
0.35205078125,
0.86328125,
0.35205078125,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.51416015625,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51416015625,
0.876953125
],
[
0.51416015625,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51416015625,
0.8935546875
],
[
0.24365234375,
0.89453125,
0.363037109375,
0.89453125,
0.363037109375,
0.9091796875,
0.24365234375,
0.9091796875
],
[
0.465576171875,
0.8955078125,
0.49072265625,
0.8955078125,
0.49072265625,
0.9091796875,
0.465576171875,
0.9091796875
],
[
0.5126953125,
0.8955078125,
0.88232421875,
0.8955078125,
0.88232421875,
0.9091796875,
0.5126953125,
0.9091796875
]
] |
[
"ances to measure dialogue coherenceby optimiz-",
"Generally, the attention layer, Atten, for an input",
"ing utterance vectors for both DAP and coherence",
"vector is defined as follows:",
"assessment, and (i) it uses DA labels to define",
"βt = at * W,",
"an auxiliary task for training the DiCoh model us.",
"exp (βt)",
"ing MTL, instead of utilizing them in a pipeline",
"Qt:",
"texp (βt)\"",
"(4)",
"plicit DA labels as inputs during coherence assess-",
"Therefore, it efficiently mitigates the need for ex-",
"O：",
"at*&t,",
"ment.",
"3Method",
"where W is the parameter of this layer, and o is its",
"weighted output vector. Attention enables the ut-",
"We represent a dialogue between two speakers",
"terance representation layer to encode an utterance",
"as a sequence of utterances, dial - [utti, ., uttm].",
"by the weighted sum of its word embeddings. It is",
"We address the problem of designing a coherence",
"worth noting that the parameters of the utterance",
"model, DiCoh, which assigns a coherence score",
"encoder are shared for representing all utterances",
"to dial, Sdial =DiCoh(dial). Given a pair of dia-",
"in a dialogue.",
"logues Φ = (diali, dial,), our DiCoh modelideally",
"DiCoh",
"model For",
"input",
"dialogue",
"",
"an",
"assigns sdial, > Sdial, if and only if dialogue dial;",
"dial = [utti,.,uttm], the output of the utter-",
"is preferred over dialogue dial, according to their",
"ance representation encoder is a sequence of",
"perceived coherence. Instead of using gold DA la-",
"vectors, i.e., [ul, .., um]. Our coherence assess-",
"bels as inputs to DiCoh, we use them to define an",
"ment model (DiCoh) combines these vectors",
"auxiliary task and model, DAP, to enrich utterance",
"by a BiLSTM to obtain dialogue-level contex-",
"vectors for DiCoh in an MTL scenario. Figure 2",
"tualized representations of utterances.",
"Then, a",
"shows a low-level illustration of our MTL-based",
"self-attention (Equation 4) with new parameters",
"approach.",
"computes the weighted average of the contextu-",
"alized utterance vectors to encode the dialogue:",
"Utterance encoder We use a word embedding",
"layer, Emb, to transform the words in utterance",
"[h,.. m] - BiLSTM([u,. m]),",
"utt = [wi,.., wn] to a sequence of embedding",
"(5)",
"d = Atten([hd,., hm])",
"vectors E = [ei,...,en], where n is the number",
"of words in utt. The embedding layer can be ini-",
"A linear feed-forward layer, FF, maps the dia-",
"tialized by any pre-trained embeddings to capture",
"logue vector, d, to a dialogue coherence score,",
"lexical relations.We use a Bidirectional recur-",
"Sdial:",
"rent neural network withLong Short-Term Mem.",
"Sdial = FF(d).",
"(6)",
"codewordsintheirutterance-levelcontext",
"ory cells, BiLSTM, to map embeddings E to en",
"DAP model Our DAP model, which is used to",
"E= Emb(utt),",
"solve the auxiliary DAP task, is a softmac layer",
"(1)",
"Hu = BiLSTM(E),",
"which maps an utterance vector, u, to a probability",
"distribution pa over DA labels A:",
"whereHushowsthehidden statevectors",
"Pa(u) = sofimax(Wju|x[A| * u + b),",
"[hu, ..., hul returned by BiLSTM. At word t, h is",
"(7)",
"the concatenation of hidden states of the forward",
"where Wjulx/Al shows the weights of the sofimax",
"h and the backward LSTMs hy:",
"layer, lu is the size of the utterance vector, [A| is",
"the number of DA labels, and b is the bias.",
"h = [ht;hy].",
"(2)",
"3.1 Multi-Task Learning",
"We apply a self-attention mechanism, Atten, to the",
"As illustrated in Figure l, our main idea is to ben-",
"hidden state vectors in Hu to obtain the vector rep-",
"efit from the DAP task for improving the perfor-",
"resentation, u, of utterance utt:",
"mance of the dialogue coherence model by using",
"them in a multi-task learning scenario. We also",
"u = Atten(Hu).",
"(3)",
"assume that each utterance uttk is associated with"
] |
{
"arxiv_id": "1908.08486",
"language": "en",
"page_idx": 2
}
|
||
1908.08486-en-3
|
1908.08486-en-3.jpg
|
<FigureHere>
Figure 2: A low-level illustration of our MTL-based approach to dialogue coherence assessment. The input is dialogue pair $p=(dial_i,dial_j).$ Dashed items represent losses. Models' parameters are shared among dialogues.
where[.] is the indexing function. More formally,
$\begin{aligned}&\mathrm{ion.}&s_{\phi[l^{c}]}\text{ and }s_{\phi[1-l^{c}]}\text{ are the coherence scores of the}\\&\text{We define a loss function for each task, and then}&\text{coherent and incoherent dialogue in pair }\phi&=\end{aligned}$
$(dial_{i},dial_{j})$,respectively. Finally, the total loss value is the weighted combination (Kendall et al., 2018) of the above losses:
DA label, $a_k$, during training but not during evalu-
ation.
use their weighted average as the total loss. The DAP loss function for dialogue dial is the average cross-entropy:
(8)
$$\mathcal{L}=\frac{\mathcal{L}_{coh}^{\phi}}{\gamma_{1}^{2}}+\frac{(\mathcal{L}_{da}^{dial_{i}}+\mathcal{L}_{da}^{dial_{j}})}{\gamma_{2}^{2}}+\log(\gamma_{1})+\log(\gamma_{2}),$$
$$\mathcal{L}_{da}^{dial}=-\frac{1}{m}\sum_{k\in(1,...,m)}a_{k}*\log(p_{a}(u_{k})),$$
where $m$ is the number of utterances in dialogue, where $\mathcal{L}_da^{dial_i}$ and $\mathcal{L}_da^{dial_j}$ are the losses of DAP for
dialogues in pair $\phi=(dial_i,dial_j),\gamma_1$ and $\gamma_2$ are trainable parameters to balance the impact of losses. We compute the gradient of $\mathcal{L}$ to update the parameters of both DiCoh and DAP models.
and $a_k$ is the one-hot vector representation of the gold DA label associated with the $k^{th}$ utterance. $log(p_a)$ is the natural log of probabilities over DA labels, which is obtained in Equation 7.
Inspired by preference learning approaches (e.g. the proposed method by Gao et al. (2019) for text summarization) we define the loss function for coherence assessment through pairwise comparisons among dialogues. Given dialogue pair $\phi=$ $(dial_i,dial_j)$ and its preference coherence label,
### 4 Experiments
## 4.1 Dialogue Corpora
$$l^c=\left\{\begin{array}{ll}0&\text{if}\:dial_i\:\text{is preferred over}\:dial_j,\\1&\text{otherwise},\end{array}\right.$$
(9)
We compare our approach with several previous dialogue coherence models on DailyDialog (Liet al., 2017) and SwitchBoard (Jurafsky and Shriberg, 1997) as two benchmark English dialogue corpora. Table 2 shows some statistics of these corpora.
DailyDialog contains human-written dialogues about daily topics (e.g. ordinary life, relationships,
the coherence loss is:
(10)
$$\mathcal{L}_{coh}^{\phi}=\max\{0,1-s_{\phi[l^{c}]}+s_{\phi[1-l^{c}]}\},$$
|
[
[
0.373046875,
0.155029296875,
0.384765625,
0.155029296875,
0.384765625,
0.1591796875,
0.373046875,
0.1591796875
],
[
0.61279296875,
0.152587890625,
0.6259765625,
0.152587890625,
0.6259765625,
0.1591796875,
0.61279296875,
0.1591796875
],
[
0.1873779296875,
0.1776123046875,
0.2008056640625,
0.1776123046875,
0.2008056640625,
0.18115234375,
0.1873779296875,
0.18115234375
],
[
0.1873779296875,
0.196533203125,
0.233642578125,
0.196533203125,
0.233642578125,
0.2037353515625,
0.1873779296875,
0.2037353515625
],
[
0.353759765625,
0.19775390625,
0.386474609375,
0.19775390625,
0.386474609375,
0.2060546875,
0.353759765625,
0.2060546875
],
[
0.6083984375,
0.19482421875,
0.6455078125,
0.19482421875,
0.6455078125,
0.2060546875,
0.6083984375,
0.2060546875
],
[
0.344482421875,
0.218505859375,
0.3671875,
0.218505859375,
0.3671875,
0.2286376953125,
0.344482421875,
0.2286376953125
],
[
0.374755859375,
0.218505859375,
0.396728515625,
0.218505859375,
0.396728515625,
0.226806640625,
0.374755859375,
0.226806640625
],
[
0.60107421875,
0.218505859375,
0.6201171875,
0.218505859375,
0.6201171875,
0.226806640625,
0.60107421875,
0.226806640625
],
[
0.62939453125,
0.2161865234375,
0.6552734375,
0.2161865234375,
0.6552734375,
0.22802734375,
0.62939453125,
0.22802734375
],
[
0.349609375,
0.2392578125,
0.392333984375,
0.2392578125,
0.392333984375,
0.2493896484375,
0.349609375,
0.2493896484375
],
[
0.60498046875,
0.2381591796875,
0.64892578125,
0.2381591796875,
0.64892578125,
0.248779296875,
0.60498046875,
0.248779296875
],
[
0.7880859375,
0.240478515625,
0.81494140625,
0.240478515625,
0.81494140625,
0.2476806640625,
0.7880859375,
0.2476806640625
],
[
0.53955078125,
0.281982421875,
0.5771484375,
0.281982421875,
0.5771484375,
0.29345703125,
0.53955078125,
0.29345703125
],
[
0.182373046875,
0.243408203125,
0.212646484375,
0.243408203125,
0.212646484375,
0.2476806640625,
0.182373046875,
0.2476806640625
],
[
0.283935546875,
0.283203125,
0.3193359375,
0.283203125,
0.3193359375,
0.29443359375,
0.283935546875,
0.29443359375
],
[
0.421875,
0.28515625,
0.456298828125,
0.28515625,
0.456298828125,
0.292236328125,
0.421875,
0.292236328125
],
[
0.67822265625,
0.281982421875,
0.71533203125,
0.281982421875,
0.71533203125,
0.29345703125,
0.67822265625,
0.29345703125
],
[
0.53125,
0.303955078125,
0.5859375,
0.303955078125,
0.5859375,
0.3154296875,
0.53125,
0.3154296875
],
[
0.67041015625,
0.303955078125,
0.724609375,
0.303955078125,
0.724609375,
0.3154296875,
0.67041015625,
0.3154296875
],
[
0.305908203125,
0.30712890625,
0.3251953125,
0.30712890625,
0.3251953125,
0.314208984375,
0.305908203125,
0.314208984375
],
[
0.27490234375,
0.3046875,
0.309326171875,
0.3046875,
0.309326171875,
0.31640625,
0.27490234375,
0.31640625
],
[
0.413330078125,
0.30712890625,
0.432861328125,
0.30712890625,
0.432861328125,
0.3154296875,
0.413330078125,
0.3154296875
],
[
0.44873046875,
0.30712890625,
0.465576171875,
0.30712890625,
0.465576171875,
0.314208984375,
0.44873046875,
0.314208984375
],
[
0.53369140625,
0.32666015625,
0.5830078125,
0.32666015625,
0.5830078125,
0.33740234375,
0.53369140625,
0.33740234375
],
[
0.67236328125,
0.32666015625,
0.72265625,
0.32666015625,
0.72265625,
0.33837890625,
0.67236328125,
0.33837890625
],
[
0.413330078125,
0.327880859375,
0.465576171875,
0.327880859375,
0.465576171875,
0.339111328125,
0.413330078125,
0.339111328125
],
[
0.27490234375,
0.327880859375,
0.326904296875,
0.327880859375,
0.326904296875,
0.339111328125,
0.27490234375,
0.339111328125
],
[
0.56982421875,
0.347412109375,
0.583984375,
0.350830078125,
0.57958984375,
0.359375,
0.56494140625,
0.355224609375
],
[
0.67236328125,
0.3486328125,
0.69140625,
0.3515625,
0.6884765625,
0.361083984375,
0.6689453125,
0.358154296875
],
[
0.705078125,
0.3515625,
0.716796875,
0.350830078125,
0.71826171875,
0.355712890625,
0.70654296875,
0.35693359375
],
[
0.53515625,
0.349853515625,
0.5546875,
0.349853515625,
0.5546875,
0.358154296875,
0.53515625,
0.358154296875
],
[
0.279052734375,
0.352783203125,
0.284912109375,
0.352783203125,
0.284912109375,
0.358154296875,
0.279052734375,
0.358154296875
],
[
0.541015625,
0.369384765625,
0.57568359375,
0.369384765625,
0.57568359375,
0.38134765625,
0.541015625,
0.38134765625
],
[
0.6796875,
0.369384765625,
0.71240234375,
0.369384765625,
0.71240234375,
0.3818359375,
0.6796875,
0.3818359375
],
[
0.423583984375,
0.371826171875,
0.4580078125,
0.371826171875,
0.4580078125,
0.383056640625,
0.423583984375,
0.383056640625
],
[
0.28662109375,
0.371826171875,
0.31689453125,
0.371826171875,
0.31689453125,
0.383056640625,
0.28662109375,
0.383056640625
],
[
0.259765625,
0.39794921875,
0.34375,
0.39794921875,
0.34375,
0.402099609375,
0.259765625,
0.402099609375
],
[
0.4208984375,
0.39794921875,
0.4814453125,
0.39794921875,
0.4814453125,
0.402099609375,
0.4208984375,
0.402099609375
],
[
0.716796875,
0.394287109375,
0.740234375,
0.394287109375,
0.740234375,
0.402099609375,
0.716796875,
0.402099609375
],
[
0.356201171875,
0.41455078125,
0.384765625,
0.41455078125,
0.384765625,
0.42578125,
0.356201171875,
0.42578125
],
[
0.61083984375,
0.413330078125,
0.64208984375,
0.413330078125,
0.64208984375,
0.424560546875,
0.61083984375,
0.424560546875
],
[
0.11846923828125,
0.439453125,
0.884765625,
0.439453125,
0.884765625,
0.453125,
0.11846923828125,
0.453125
],
[
0.1168212890625,
0.453125,
0.88232421875,
0.451904296875,
0.88232421875,
0.466064453125,
0.1168212890625,
0.46728515625
],
[
0.12017822265625,
0.49462890625,
0.4931640625,
0.49462890625,
0.4931640625,
0.50830078125,
0.12017822265625,
0.50830078125
],
[
0.51171875,
0.4921875,
0.884765625,
0.493408203125,
0.884765625,
0.50830078125,
0.51171875,
0.50732421875
],
[
0.12017822265625,
0.5126953125,
0.1605224609375,
0.5126953125,
0.1605224609375,
0.52392578125,
0.12017822265625,
0.52392578125
],
[
0.51611328125,
0.51123046875,
0.884765625,
0.51123046875,
0.884765625,
0.525390625,
0.51611328125,
0.525390625
],
[
0.1361083984375,
0.52783203125,
0.486572265625,
0.52783203125,
0.486572265625,
0.5380859375,
0.1361083984375,
0.5380859375
],
[
0.51416015625,
0.52685546875,
0.85986328125,
0.52685546875,
0.85986328125,
0.54052734375,
0.51416015625,
0.54052734375
],
[
0.11846923828125,
0.5419921875,
0.49072265625,
0.5419921875,
0.49072265625,
0.5556640625,
0.11846923828125,
0.5556640625
],
[
0.51171875,
0.54150390625,
0.88232421875,
0.54150390625,
0.88232421875,
0.5546875,
0.51171875,
0.5546875
],
[
0.1168212890625,
0.55712890625,
0.489013671875,
0.55810546875,
0.489013671875,
0.572265625,
0.1168212890625,
0.57177734375
],
[
0.51416015625,
0.55810546875,
0.88671875,
0.55810546875,
0.88671875,
0.57177734375,
0.51416015625,
0.57177734375
],
[
0.12017822265625,
0.576171875,
0.2269287109375,
0.576171875,
0.2269287109375,
0.58740234375,
0.12017822265625,
0.58740234375
],
[
0.51611328125,
0.57373046875,
0.7109375,
0.57373046875,
0.7109375,
0.58740234375,
0.51611328125,
0.58740234375
],
[
0.552734375,
0.5927734375,
0.5966796875,
0.5927734375,
0.5966796875,
0.6181640625,
0.552734375,
0.6181640625
],
[
0.60498046875,
0.5888671875,
0.71240234375,
0.5888671875,
0.71240234375,
0.6162109375,
0.60498046875,
0.6162109375
],
[
0.681640625,
0.59423828125,
0.72119140625,
0.5927734375,
0.72265625,
0.60400390625,
0.68212890625,
0.60546875
],
[
0.2269287109375,
0.59326171875,
0.2479248046875,
0.60400390625,
0.236083984375,
0.6142578125,
0.215087890625,
0.60400390625
],
[
0.1436767578125,
0.59716796875,
0.2135009765625,
0.59716796875,
0.2135009765625,
0.62451171875,
0.1436767578125,
0.62451171875
],
[
0.251220703125,
0.599609375,
0.29736328125,
0.60400390625,
0.294189453125,
0.619140625,
0.2479248046875,
0.615234375
],
[
0.283935546875,
0.6015625,
0.44287109375,
0.60400390625,
0.44287109375,
0.62060546875,
0.283935546875,
0.6181640625
],
[
0.4638671875,
0.60498046875,
0.4873046875,
0.60498046875,
0.4873046875,
0.61767578125,
0.4638671875,
0.61767578125
],
[
0.51416015625,
0.60693359375,
0.54541015625,
0.60498046875,
0.546875,
0.615234375,
0.517578125,
0.6181640625
],
[
0.7333984375,
0.60791015625,
0.88232421875,
0.60791015625,
0.88232421875,
0.62255859375,
0.7333984375,
0.62255859375
],
[
0.22021484375,
0.615234375,
0.2529296875,
0.615234375,
0.2529296875,
0.626953125,
0.22021484375,
0.626953125
],
[
0.2109375,
0.6025390625,
0.2210693359375,
0.6025390625,
0.2210693359375,
0.60986328125,
0.2109375,
0.60986328125
],
[
0.560546875,
0.6181640625,
0.5859375,
0.6181640625,
0.5859375,
0.6318359375,
0.560546875,
0.6318359375
],
[
0.6552734375,
0.615234375,
0.68212890625,
0.615234375,
0.68212890625,
0.63427734375,
0.6552734375,
0.63427734375
],
[
0.8505859375,
0.63134765625,
0.884765625,
0.63134765625,
0.884765625,
0.6435546875,
0.8505859375,
0.6435546875
],
[
0.236083984375,
0.62060546875,
0.311767578125,
0.62255859375,
0.311767578125,
0.6357421875,
0.236083984375,
0.63427734375
],
[
0.5068359375,
0.6357421875,
0.892578125,
0.6357421875,
0.892578125,
0.66748046875,
0.5068359375,
0.66748046875
],
[
0.1168212890625,
0.642578125,
0.489013671875,
0.6435546875,
0.489013671875,
0.65771484375,
0.1168212890625,
0.65673828125
],
[
0.6455078125,
0.64794921875,
0.6552734375,
0.64794921875,
0.6552734375,
0.65185546875,
0.6455078125,
0.65185546875
],
[
0.1168212890625,
0.6591796875,
0.4873046875,
0.6591796875,
0.4873046875,
0.6728515625,
0.1168212890625,
0.6728515625
],
[
0.57373046875,
0.65380859375,
0.5966796875,
0.65380859375,
0.5966796875,
0.6611328125,
0.57373046875,
0.6611328125
],
[
0.5126953125,
0.662109375,
0.88232421875,
0.662109375,
0.88232421875,
0.67578125,
0.5126953125,
0.67578125
],
[
0.1168212890625,
0.67333984375,
0.492431640625,
0.6728515625,
0.492431640625,
0.6904296875,
0.1168212890625,
0.69140625
],
[
0.51416015625,
0.6787109375,
0.884765625,
0.6787109375,
0.884765625,
0.693359375,
0.51416015625,
0.693359375
],
[
0.11846923828125,
0.69140625,
0.489013671875,
0.6904296875,
0.489013671875,
0.705078125,
0.11846923828125,
0.7060546875
],
[
0.51416015625,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.708984375,
0.51416015625,
0.708984375
],
[
0.11846923828125,
0.7080078125,
0.411865234375,
0.7080078125,
0.411865234375,
0.7216796875,
0.11846923828125,
0.7216796875
],
[
0.51416015625,
0.71142578125,
0.869140625,
0.71142578125,
0.869140625,
0.724609375,
0.51416015625,
0.724609375
],
[
0.13525390625,
0.72265625,
0.489013671875,
0.724609375,
0.489013671875,
0.7392578125,
0.13525390625,
0.73681640625
],
[
0.51611328125,
0.7392578125,
0.54296875,
0.7392578125,
0.54296875,
0.74755859375,
0.51611328125,
0.74755859375
],
[
0.53857421875,
0.73681640625,
0.65869140625,
0.73681640625,
0.65869140625,
0.751953125,
0.53857421875,
0.751953125
],
[
0.11846923828125,
0.755859375,
0.49072265625,
0.755859375,
0.49072265625,
0.76953125,
0.11846923828125,
0.76953125
],
[
0.11846923828125,
0.74072265625,
0.49072265625,
0.74072265625,
0.49072265625,
0.75439453125,
0.11846923828125,
0.75439453125
],
[
0.51416015625,
0.76123046875,
0.70068359375,
0.76123046875,
0.70068359375,
0.775390625,
0.51416015625,
0.775390625
],
[
0.11846923828125,
0.7724609375,
0.49072265625,
0.7724609375,
0.49072265625,
0.7861328125,
0.11846923828125,
0.7861328125
],
[
0.5126953125,
0.7822265625,
0.6318359375,
0.7822265625,
0.6318359375,
0.79638671875,
0.5126953125,
0.79638671875
],
[
0.62255859375,
0.783203125,
0.7646484375,
0.7822265625,
0.7646484375,
0.79638671875,
0.62255859375,
0.79736328125
],
[
0.759765625,
0.78369140625,
0.81494140625,
0.78369140625,
0.81494140625,
0.7958984375,
0.759765625,
0.7958984375
],
[
0.81494140625,
0.78369140625,
0.88330078125,
0.78369140625,
0.88330078125,
0.7958984375,
0.81494140625,
0.7958984375
],
[
0.51416015625,
0.798828125,
0.67236328125,
0.798828125,
0.67236328125,
0.8134765625,
0.51416015625,
0.8134765625
],
[
0.11846923828125,
0.7880859375,
0.489013671875,
0.7880859375,
0.489013671875,
0.802734375,
0.11846923828125,
0.802734375
],
[
0.673828125,
0.80029296875,
0.765625,
0.80029296875,
0.765625,
0.81103515625,
0.673828125,
0.81103515625
],
[
0.7783203125,
0.7998046875,
0.84375,
0.7998046875,
0.84375,
0.81103515625,
0.7783203125,
0.81103515625
],
[
0.85546875,
0.80029296875,
0.88330078125,
0.80029296875,
0.88330078125,
0.8125,
0.85546875,
0.8125
],
[
0.5126953125,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.5126953125,
0.82763671875
],
[
0.11846923828125,
0.8046875,
0.471435546875,
0.8046875,
0.471435546875,
0.8193359375,
0.11846923828125,
0.8193359375
],
[
0.1873779296875,
0.826171875,
0.447998046875,
0.8291015625,
0.447998046875,
0.8466796875,
0.1873779296875,
0.84326171875
],
[
0.51416015625,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.845703125,
0.51416015625,
0.845703125
],
[
0.1395263671875,
0.84033203125,
0.148681640625,
0.84033203125,
0.148681640625,
0.845703125,
0.1395263671875,
0.845703125
],
[
0.465576171875,
0.83837890625,
0.4873046875,
0.83837890625,
0.4873046875,
0.85107421875,
0.465576171875,
0.85107421875
],
[
0.194091796875,
0.845703125,
0.29736328125,
0.845703125,
0.29736328125,
0.85986328125,
0.194091796875,
0.85986328125
],
[
0.5126953125,
0.845703125,
0.72021484375,
0.8466796875,
0.72021484375,
0.86083984375,
0.5126953125,
0.85986328125
],
[
0.7109375,
0.84716796875,
0.884765625,
0.84716796875,
0.884765625,
0.86083984375,
0.7109375,
0.86083984375
],
[
0.51416015625,
0.86328125,
0.70751953125,
0.86328125,
0.70751953125,
0.876953125,
0.51416015625,
0.876953125
],
[
0.12017822265625,
0.87060546875,
0.278076171875,
0.87060546875,
0.278076171875,
0.88427734375,
0.12017822265625,
0.88427734375
],
[
0.53173828125,
0.8798828125,
0.88330078125,
0.8798828125,
0.88330078125,
0.8935546875,
0.53173828125,
0.8935546875
],
[
0.148681640625,
0.892578125,
0.432861328125,
0.892578125,
0.432861328125,
0.9130859375,
0.148681640625,
0.9130859375
],
[
0.456298828125,
0.8955078125,
0.4873046875,
0.8955078125,
0.4873046875,
0.90771484375,
0.456298828125,
0.90771484375
],
[
0.5126953125,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.5126953125,
0.9091796875
]
] |
[
"aat",
"Lit",
"",
"oss-entrop",
"Atten",
"Atten",
"[hi..",
"hal",
"[h4",
",hal",
"BiLSTM",
"BiLSTM",
"oftma",
"Atten",
"sotumax",
"Atten",
"Atten",
"Atten",
"[hi...hu]",
"[h...hu]",
".hu",
"[..",
"[h.",
"hul",
"BiLSTM",
"BiLSTM",
"BiLSTM",
"BiLSTM",
"en",
"[e1,",
"e.",
"[ei,",
"e",
"[Emb]",
"Emb",
"Emb]",
"Emb",
"tt =w.-...n",
"=1-...n",
".wn",
"dial.",
"dial,",
"Figure 2: A low-level illustration of our MTL-based approach to dialogue coherence assessment. The input is",
"dialogue pair p = (diali, dial,). Dashed items represent losses. Models' parameters are shared among dialogues.",
"DA label, ak, during training but not during evalu-",
"where [-] is the indexing function. More formally,",
"ation.",
"Sp[1ej and So[1-ie] are the coherence scores of the",
"Wedefinealossfunctionforeachtask,andthen",
"coherent and incoherent dialogue in pair Φ",
"use their weighted average as the total loss. The",
"(dial,,dial,),respectively.Finally,thetotal loss",
"DAP loss function for dialogue dial is the average",
"value is the weighted combination (Kendall et al.,",
"cross-entropy:",
"2018) of the above losses:",
"Coh",
"(caial + Cclal",
"cdial,",
"",
"cal = :",
"√",
"ak * log(pa(uk)),",
"(8)",
"-",
"+log(1) + log(2),",
"m.",
"",
"21",
"22",
"(11)",
"ke(...m)",
"Whee andarethe lssesofAfo",
"where m is the number of utterances in dialogue,",
"",
"and ak is the one-hot vector representation of the",
"da",
"dialogues in pair Φ = (diali,dial,), i and 2",
"gold DA label associated with the kth utterance.",
"are trainable parameters to balance the impact of",
"log(pa) is the natural log of probabilities over DA",
"losses. We compute the gradient of L to update",
"labels, which is obtained in Equation 7.",
"the parameters of both DiCoh and DAP models.",
"Inspired by preference learning approaches (e.g)",
"4",
"Experiments",
"summarization) we define the loss function for",
"the proposed method by Gao et al. (2019) for text",
"4.1 Dialogue Corpora",
"coherence assessment through pairwise compar-",
"Wecompare",
"ourapproach",
"with",
"several",
"previousdialogue",
"isons among dialogues. Given dialogue pair Φ =",
"coherence",
"models",
"on",
"DailyDialog (Liet al., 2017) and SwitchBoard",
"(diali, dial,) and its preference coherence label,",
"o if dial; is preferred over dialj,",
"(Jurafsky and Shriberg, 1997) as two benchmark",
"C",
"(9)",
"1 otherwise,",
"English dialogue corpora.",
"Table 2 shows some",
"statistics of these corpora.",
"the coherence loss is:",
"DailyDialog contains human-written dialogues",
"Cooh = max[0, 1 - 8g[] + Ss[1-19],",
"(10)",
"about daily topics (e.g. ordinary life, relationships."
] |
{
"arxiv_id": "1908.08486",
"language": "en",
"page_idx": 3
}
|
||
1908.08486-en-4
|
1908.08486-en-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>DailyDialog</th>
<th>SwitchBoard</th>
</tr>
<tr>
<td>$\#$dialogues</td>
<td>13,118</td>
<td>1,155</td>
</tr>
<tr>
<td>$\#$DA labels</td>
<td>4</td>
<td>42</td>
</tr>
<tr>
<td>$\arg.\#$utter. 5 per dialogue</td>
<td>7.9</td>
<td>191.9</td>
</tr>
<tr>
<td>$\arg.\#$ 1 words $\frac{1}{2}=\frac{1}{2}$ per utter.</td>
<td>14.6</td>
<td>9.26</td>
</tr>
</tbody>
</table>
Table 2: The statistics of the DailyDialog and
SwitchBoard corpora.
work, etc) collected by crowd-sourcing. Crowdworkers also annotated utterances with generic DA labels from the set {Inform, Question, Directive, Commissive}. Dialogues in this corpus contain a few utterances $(\approx8)$ making them more on topic and less dispersed. However, utterances are long in terms of the number of words $(\approx15).$
SwitchBoard contains informal English dialogues collected from phone conversations between two mutually unknown human participants. The participants were given only one of 70 possible topics as initial topic to start a conversation but they were free to diverge from that topic during the conversation. So, there is no concrete topic associated with each dialogue in this dataset as it is the case for dialogues in DailyDialog.
DA labels in SwitchBoard are about 10 times more fine-grained than those in DailyDialog. For example, a question utterance in SwitchBoard may have a fine-grained DA label such as Yes-NoQuestion, Wh-Question, Rhetorical-Questions, etc. The distribution of these acts is however highly unbalanced in SwitchBoard: the most frequent act label makes up for 36% of the utterances in the corpus, the three most frequent acts together make up for 68% of the utterances, while most of the remaining act labels just make up for 1% or less of all the utterances.
On average, dialogues in SwitchBoard contain more utterances than those in DailyDialog (192 vs 8) but utterances in SwitchBoard are shorter than those in DailyDialog (9 vs 15). This means that dialogues in SwitchBoard are more likely to span different topics than the ones in DailyDialog. The utterances in DailyDialog are explicitly cleaned of any noise, like “uh-oh”, or interruptions by the other speaker, as it is commonly the case for dialogues in SwitchBoard. While each dialogue turn of dialogues in DailyDialog contains only one utterance, dialogue turns in SwitchBoard may consist of several utterances. That is why we consider each dialogue as a sequence of dialogue utterances.
## 4.2 Problem-domains
The goal of our experiments is to assess if a coherence model assigns coherence scores to dialogues so that a more coherent dialogue obtains a higher score than a less coherent one. Since dialogues in the examined corpora, i.e. DailyDialog and SwitchBoard , are not associated with any coherence assessment score, we synthetically define four perturbation methods to destroy the coherence of dialogues in these corpora, and create a set of dialogue pairs for training and testing coherence models
We borrow Utterance Ordering (UO) and Utterance Insertion (UI) from previous studies on coherence assessment (Barzilay and Lapata, 2005; Cervone et al., 2018) and also introduce Utterance Replacement (UR), and Even Utterance Ordering (EUO) as more challenging and dialogue-relevant perturbation methods. Since each experiment follows a specific perturbation method, henceforth, we refer to these perturbations as problemdomains:
Utterance Ordering (UO) We randomly permute the order of utterances in dialogue. The original dialogue is preferred over the perturbed one.
Utterance Insertion (UI) We remove each utterance of a dialogue and then re-insert it in any possible utterance position in the dialogue. We assume that the original place of the utterance is the best place for the insertion. Therefore, a coherence model ideally discriminates the original dialogue from the perturbed ones, which are obtained by re-inserting the removed utterance in any utterance position except its original one. This problem-domain is more difficult to solve than UO as the distinction between dialogues is in the position of only one utterance.
Utterance Replacement (UR) We randomly replace one of the utterances in a dialogue with another utterance that is also randomly selected from another dialogue. The original dialogue is preferred over the dialogue generated by UR. Unlike the other problem-domains, which perturb the structure of a dialogue, this problem-domain perturbs the coherence of a dialogue at its semantic level.
Even Utterance Ordering (EUO) This problem-domain is similar to UO but here we re-arrange the order of utterances that are said
|
[
[
0.29833984375,
0.07598876953125,
0.38232421875,
0.07781982421875,
0.38232421875,
0.0926513671875,
0.29833984375,
0.09027099609375
],
[
0.392333984375,
0.07659912109375,
0.479736328125,
0.0748291015625,
0.479736328125,
0.08905029296875,
0.392333984375,
0.0914306640625
],
[
0.5126953125,
0.0748291015625,
0.69921875,
0.07659912109375,
0.69921875,
0.09027099609375,
0.5126953125,
0.0885009765625
],
[
0.300048828125,
0.0955810546875,
0.349609375,
0.0955810546875,
0.349609375,
0.10986328125,
0.300048828125,
0.10986328125
],
[
0.39404296875,
0.09442138671875,
0.438720703125,
0.09442138671875,
0.438720703125,
0.10986328125,
0.39404296875,
0.10986328125
],
[
0.51171875,
0.097412109375,
0.884765625,
0.097412109375,
0.884765625,
0.11102294921875,
0.51171875,
0.11102294921875
],
[
0.127685546875,
0.10687255859375,
0.205078125,
0.10809326171875,
0.205078125,
0.119384765625,
0.127685546875,
0.1181640625
],
[
0.12939453125,
0.0968017578125,
0.203369140625,
0.0968017578125,
0.203369140625,
0.10809326171875,
0.12939453125,
0.10809326171875
],
[
0.3017578125,
0.10986328125,
0.31689453125,
0.10986328125,
0.31689453125,
0.1181640625,
0.3017578125,
0.1181640625
],
[
0.396728515625,
0.1092529296875,
0.41943359375,
0.1092529296875,
0.41943359375,
0.119384765625,
0.396728515625,
0.119384765625
],
[
0.29833984375,
0.1181640625,
0.326904296875,
0.1181640625,
0.326904296875,
0.1312255859375,
0.29833984375,
0.1312255859375
],
[
0.5126953125,
0.114013671875,
0.884765625,
0.114013671875,
0.884765625,
0.127685546875,
0.5126953125,
0.127685546875
],
[
0.395751953125,
0.119384765625,
0.435302734375,
0.119384765625,
0.435302734375,
0.1312255859375,
0.395751953125,
0.1312255859375
],
[
0.127685546875,
0.119384765625,
0.28662109375,
0.119384765625,
0.28662109375,
0.133056640625,
0.127685546875,
0.133056640625
],
[
0.29736328125,
0.1300048828125,
0.336181640625,
0.1300048828125,
0.336181640625,
0.1455078125,
0.29736328125,
0.1455078125
],
[
0.392333984375,
0.127685546875,
0.432861328125,
0.1300048828125,
0.43115234375,
0.1455078125,
0.390869140625,
0.1436767578125
],
[
0.51416015625,
0.1312255859375,
0.884765625,
0.1312255859375,
0.884765625,
0.144287109375,
0.51416015625,
0.144287109375
],
[
0.51416015625,
0.146728515625,
0.88818359375,
0.146728515625,
0.88818359375,
0.1602783203125,
0.51416015625,
0.1602783203125
],
[
0.127685546875,
0.1312255859375,
0.268798828125,
0.1300048828125,
0.268798828125,
0.144287109375,
0.127685546875,
0.1455078125
],
[
0.11846923828125,
0.1602783203125,
0.1856689453125,
0.1602783203125,
0.1856689453125,
0.171630859375,
0.11846923828125,
0.171630859375
],
[
0.1932373046875,
0.1591796875,
0.4931640625,
0.1591796875,
0.4931640625,
0.1728515625,
0.1932373046875,
0.1728515625
],
[
0.5126953125,
0.160888671875,
0.88671875,
0.162109375,
0.88671875,
0.177001953125,
0.5126953125,
0.17578125
],
[
0.11846923828125,
0.17333984375,
0.268798828125,
0.17333984375,
0.268798828125,
0.18701171875,
0.11846923828125,
0.18701171875
],
[
0.51416015625,
0.1787109375,
0.88671875,
0.1787109375,
0.88671875,
0.1923828125,
0.51416015625,
0.1923828125
],
[
0.5126953125,
0.19482421875,
0.88330078125,
0.19482421875,
0.88330078125,
0.2078857421875,
0.5126953125,
0.2078857421875
],
[
0.11846923828125,
0.2161865234375,
0.49072265625,
0.2161865234375,
0.49072265625,
0.2298583984375,
0.11846923828125,
0.2298583984375
],
[
0.51416015625,
0.21142578125,
0.88671875,
0.21142578125,
0.88671875,
0.2244873046875,
0.51416015625,
0.2244873046875
],
[
0.1168212890625,
0.232177734375,
0.49072265625,
0.23095703125,
0.49072265625,
0.2452392578125,
0.1168212890625,
0.2464599609375
],
[
0.51416015625,
0.22802734375,
0.884765625,
0.22802734375,
0.884765625,
0.2410888671875,
0.51416015625,
0.2410888671875
],
[
0.11846923828125,
0.248779296875,
0.49072265625,
0.248779296875,
0.49072265625,
0.262451171875,
0.11846923828125,
0.262451171875
],
[
0.51416015625,
0.243408203125,
0.88671875,
0.243408203125,
0.88671875,
0.257080078125,
0.51416015625,
0.257080078125
],
[
0.1168212890625,
0.26318359375,
0.489013671875,
0.26318359375,
0.489013671875,
0.276611328125,
0.1168212890625,
0.276611328125
],
[
0.51416015625,
0.260009765625,
0.61279296875,
0.260009765625,
0.61279296875,
0.271484375,
0.51416015625,
0.271484375
],
[
0.1168212890625,
0.27978515625,
0.4873046875,
0.27978515625,
0.4873046875,
0.29345703125,
0.1168212890625,
0.29345703125
],
[
0.53125,
0.27490234375,
0.884765625,
0.27490234375,
0.884765625,
0.2880859375,
0.53125,
0.2880859375
],
[
0.11846923828125,
0.29638671875,
0.49072265625,
0.29638671875,
0.49072265625,
0.31005859375,
0.11846923828125,
0.31005859375
],
[
0.5126953125,
0.292236328125,
0.88671875,
0.292236328125,
0.88671875,
0.305908203125,
0.5126953125,
0.305908203125
],
[
0.11846923828125,
0.312255859375,
0.417724609375,
0.312255859375,
0.417724609375,
0.325927734375,
0.11846923828125,
0.325927734375
],
[
0.51416015625,
0.30810546875,
0.88671875,
0.30810546875,
0.88671875,
0.32177734375,
0.51416015625,
0.32177734375
],
[
0.5126953125,
0.322509765625,
0.88330078125,
0.322509765625,
0.88330078125,
0.336181640625,
0.5126953125,
0.336181640625
],
[
0.1361083984375,
0.330078125,
0.49072265625,
0.330078125,
0.49072265625,
0.34326171875,
0.1361083984375,
0.34326171875
],
[
0.1168212890625,
0.344482421875,
0.489013671875,
0.344482421875,
0.489013671875,
0.359375,
0.1168212890625,
0.359375
],
[
0.51416015625,
0.34033203125,
0.88330078125,
0.34033203125,
0.88330078125,
0.35400390625,
0.51416015625,
0.35400390625
],
[
0.5126953125,
0.355224609375,
0.88232421875,
0.355224609375,
0.88232421875,
0.3681640625,
0.5126953125,
0.3681640625
],
[
0.11846923828125,
0.3623046875,
0.49072265625,
0.3623046875,
0.49072265625,
0.3759765625,
0.11846923828125,
0.3759765625
],
[
0.51416015625,
0.371826171875,
0.88671875,
0.371826171875,
0.88671875,
0.385986328125,
0.51416015625,
0.385986328125
],
[
0.12017822265625,
0.377685546875,
0.49072265625,
0.377685546875,
0.49072265625,
0.391357421875,
0.12017822265625,
0.391357421875
],
[
0.51416015625,
0.388427734375,
0.88818359375,
0.388427734375,
0.88818359375,
0.402099609375,
0.51416015625,
0.402099609375
],
[
0.12017822265625,
0.394287109375,
0.489013671875,
0.394287109375,
0.489013671875,
0.407958984375,
0.12017822265625,
0.407958984375
],
[
0.5126953125,
0.402587890625,
0.88818359375,
0.40380859375,
0.88818359375,
0.418701171875,
0.5126953125,
0.41748046875
],
[
0.12017822265625,
0.412109375,
0.4873046875,
0.412109375,
0.4873046875,
0.4228515625,
0.12017822265625,
0.4228515625
],
[
0.11846923828125,
0.427001953125,
0.49072265625,
0.427001953125,
0.49072265625,
0.440673828125,
0.11846923828125,
0.440673828125
],
[
0.51416015625,
0.4228515625,
0.58740234375,
0.4228515625,
0.58740234375,
0.432861328125,
0.51416015625,
0.432861328125
],
[
0.12017822265625,
0.4423828125,
0.49072265625,
0.4423828125,
0.49072265625,
0.4560546875,
0.12017822265625,
0.4560546875
],
[
0.51171875,
0.443603515625,
0.88818359375,
0.44482421875,
0.88818359375,
0.4619140625,
0.51171875,
0.46142578125
],
[
0.11846923828125,
0.458984375,
0.36962890625,
0.458984375,
0.36962890625,
0.47265625,
0.11846923828125,
0.47265625
],
[
0.51416015625,
0.463134765625,
0.88671875,
0.463134765625,
0.88671875,
0.476806640625,
0.51416015625,
0.476806640625
],
[
0.1361083984375,
0.4755859375,
0.49072265625,
0.4755859375,
0.49072265625,
0.4892578125,
0.1361083984375,
0.4892578125
],
[
0.5126953125,
0.478515625,
0.87744140625,
0.478515625,
0.87744140625,
0.4921875,
0.5126953125,
0.4921875
],
[
0.12017822265625,
0.493408203125,
0.4873046875,
0.493408203125,
0.4873046875,
0.50439453125,
0.12017822265625,
0.50439453125
],
[
0.12017822265625,
0.50830078125,
0.489013671875,
0.50830078125,
0.489013671875,
0.521484375,
0.12017822265625,
0.521484375
],
[
0.5126953125,
0.50439453125,
0.884765625,
0.50439453125,
0.884765625,
0.5185546875,
0.5126953125,
0.5185546875
],
[
0.12017822265625,
0.52392578125,
0.4931640625,
0.52392578125,
0.4931640625,
0.53759765625,
0.12017822265625,
0.53759765625
],
[
0.5126953125,
0.521484375,
0.88671875,
0.521484375,
0.88671875,
0.53515625,
0.5126953125,
0.53515625
],
[
0.12017822265625,
0.5390625,
0.49072265625,
0.5390625,
0.49072265625,
0.55419921875,
0.12017822265625,
0.55419921875
],
[
0.51171875,
0.5361328125,
0.884765625,
0.5361328125,
0.884765625,
0.5498046875,
0.51171875,
0.5498046875
],
[
0.12017822265625,
0.5556640625,
0.49072265625,
0.5556640625,
0.49072265625,
0.5693359375,
0.12017822265625,
0.5693359375
],
[
0.51416015625,
0.552734375,
0.88671875,
0.552734375,
0.88671875,
0.56640625,
0.51416015625,
0.56640625
],
[
0.11846923828125,
0.57177734375,
0.49072265625,
0.57177734375,
0.49072265625,
0.5849609375,
0.11846923828125,
0.5849609375
],
[
0.51416015625,
0.5693359375,
0.88671875,
0.5693359375,
0.88671875,
0.5830078125,
0.51416015625,
0.5830078125
],
[
0.1168212890625,
0.58740234375,
0.489013671875,
0.58740234375,
0.489013671875,
0.60107421875,
0.1168212890625,
0.60107421875
],
[
0.5126953125,
0.58447265625,
0.88232421875,
0.58447265625,
0.88232421875,
0.59716796875,
0.5126953125,
0.59716796875
],
[
0.11846923828125,
0.60498046875,
0.49072265625,
0.60498046875,
0.49072265625,
0.6181640625,
0.11846923828125,
0.6181640625
],
[
0.51416015625,
0.6015625,
0.88330078125,
0.6015625,
0.88330078125,
0.615234375,
0.51416015625,
0.615234375
],
[
0.11846923828125,
0.62060546875,
0.49072265625,
0.62060546875,
0.49072265625,
0.63427734375,
0.11846923828125,
0.63427734375
],
[
0.51416015625,
0.61767578125,
0.88330078125,
0.61767578125,
0.88330078125,
0.63134765625,
0.51416015625,
0.63134765625
],
[
0.12017822265625,
0.63720703125,
0.492431640625,
0.63720703125,
0.492431640625,
0.65087890625,
0.12017822265625,
0.65087890625
],
[
0.51416015625,
0.63427734375,
0.884765625,
0.63427734375,
0.884765625,
0.64794921875,
0.51416015625,
0.64794921875
],
[
0.11846923828125,
0.65283203125,
0.303466796875,
0.65283203125,
0.303466796875,
0.66650390625,
0.11846923828125,
0.66650390625
],
[
0.51416015625,
0.6494140625,
0.88330078125,
0.6494140625,
0.88330078125,
0.6630859375,
0.51416015625,
0.6630859375
],
[
0.1361083984375,
0.66943359375,
0.489013671875,
0.66845703125,
0.489013671875,
0.68310546875,
0.1361083984375,
0.68408203125
],
[
0.51416015625,
0.66650390625,
0.88671875,
0.66650390625,
0.88671875,
0.6796875,
0.51416015625,
0.6796875
],
[
0.11846923828125,
0.68603515625,
0.49072265625,
0.68603515625,
0.49072265625,
0.69970703125,
0.11846923828125,
0.69970703125
],
[
0.51416015625,
0.681640625,
0.70947265625,
0.681640625,
0.70947265625,
0.6953125,
0.51416015625,
0.6953125
],
[
0.121826171875,
0.70361328125,
0.4873046875,
0.70361328125,
0.4873046875,
0.71435546875,
0.121826171875,
0.71435546875
],
[
0.51611328125,
0.70703125,
0.88671875,
0.70703125,
0.88671875,
0.7216796875,
0.51611328125,
0.7216796875
],
[
0.11846923828125,
0.71875,
0.489013671875,
0.71875,
0.489013671875,
0.732421875,
0.11846923828125,
0.732421875
],
[
0.51416015625,
0.724609375,
0.88671875,
0.724609375,
0.88671875,
0.73828125,
0.51416015625,
0.73828125
],
[
0.12017822265625,
0.73388671875,
0.489013671875,
0.73388671875,
0.489013671875,
0.74755859375,
0.12017822265625,
0.74755859375
],
[
0.51416015625,
0.74072265625,
0.884765625,
0.74072265625,
0.884765625,
0.75439453125,
0.51416015625,
0.75439453125
],
[
0.11846923828125,
0.7490234375,
0.489013671875,
0.75,
0.489013671875,
0.76416015625,
0.11846923828125,
0.76318359375
],
[
0.51171875,
0.7548828125,
0.88671875,
0.755859375,
0.88671875,
0.77099609375,
0.51171875,
0.76953125
],
[
0.11846923828125,
0.7666015625,
0.489013671875,
0.7666015625,
0.489013671875,
0.77978515625,
0.11846923828125,
0.77978515625
],
[
0.51416015625,
0.7724609375,
0.88818359375,
0.7724609375,
0.88818359375,
0.7861328125,
0.51416015625,
0.7861328125
],
[
0.1168212890625,
0.7822265625,
0.4873046875,
0.7822265625,
0.4873046875,
0.7958984375,
0.1168212890625,
0.7958984375
],
[
0.51416015625,
0.7880859375,
0.884765625,
0.7880859375,
0.884765625,
0.802734375,
0.51416015625,
0.802734375
],
[
0.12017822265625,
0.7998046875,
0.49072265625,
0.7998046875,
0.49072265625,
0.8134765625,
0.12017822265625,
0.8134765625
],
[
0.51416015625,
0.8046875,
0.88671875,
0.8046875,
0.88671875,
0.818359375,
0.51416015625,
0.818359375
],
[
0.11846923828125,
0.81396484375,
0.489013671875,
0.81396484375,
0.489013671875,
0.82763671875,
0.11846923828125,
0.82763671875
],
[
0.51171875,
0.8193359375,
0.884765625,
0.82080078125,
0.884765625,
0.8349609375,
0.51171875,
0.83447265625
],
[
0.11846923828125,
0.83056640625,
0.49072265625,
0.83056640625,
0.49072265625,
0.84423828125,
0.11846923828125,
0.84423828125
],
[
0.5126953125,
0.8359375,
0.55810546875,
0.8359375,
0.55810546875,
0.85107421875,
0.5126953125,
0.85107421875
],
[
0.11846923828125,
0.8466796875,
0.4931640625,
0.8466796875,
0.4931640625,
0.86083984375,
0.11846923828125,
0.86083984375
],
[
0.1168212890625,
0.86083984375,
0.494873046875,
0.8623046875,
0.494873046875,
0.876953125,
0.1168212890625,
0.8759765625
],
[
0.57080078125,
0.86083984375,
0.6630859375,
0.86328125,
0.6630859375,
0.87744140625,
0.57080078125,
0.8759765625
],
[
0.67626953125,
0.86083984375,
0.76318359375,
0.86328125,
0.76318359375,
0.87744140625,
0.67626953125,
0.8759765625
],
[
0.7705078125,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.8759765625,
0.7705078125,
0.8759765625
],
[
0.5126953125,
0.87890625,
0.884765625,
0.87744140625,
0.884765625,
0.892578125,
0.5126953125,
0.8935546875
],
[
0.51416015625,
0.86328125,
0.57958984375,
0.86328125,
0.57958984375,
0.87451171875,
0.51416015625,
0.87451171875
],
[
0.11846923828125,
0.8798828125,
0.494873046875,
0.8798828125,
0.494873046875,
0.8935546875,
0.11846923828125,
0.8935546875
],
[
0.12017822265625,
0.896484375,
0.1856689453125,
0.896484375,
0.1856689453125,
0.90771484375,
0.12017822265625,
0.90771484375
],
[
0.51171875,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51171875,
0.9091796875
]
] |
[
"DailyDialog",
"SwitchBoard",
"4.2Problem-domains",
"13,118",
"1,155",
"The goal of our experiments is to assess if a co-",
"#DA labels",
"# dialogues",
"4.",
"42",
"7.9",
"herence model assigns coherence scores to dia-",
"191.9",
"avg. # utter. per dialogue",
"14.6",
"9.26",
"logues so that a more coherent dialogue obtains",
"a higher score than a less coherent one. Since di-",
"avg. # words per utter.",
"Table 2:",
"The statistics of the DailyDialog and",
"alogues in the examined corpora, i.e. DailyDialog",
"SwitchBoard corpora.",
"and SwitchBoard , are not associated with any co-",
"herence assessment score, we synthetically define",
"work, etc) collected by crowd-sourcing. Crowd-",
"four perturbation methods to destroy the coher-",
"workers also annotated utterances with generic DA",
"ence of dialogues in these corpora, and create a",
"labels from the set [Inform, Question, Directive,",
"set of dialogue pairs for training and testing coher-",
"Commissive. Dialogues in this corpus contain a",
"ence models.",
"few utterances (~ 8) making them more on topic",
"We borrow Utterance Ordering (UO) and Ut-",
"and less dispersed. However, utterances are long",
"terance Insertion (Ul) from previous studies on",
"in terms of the number of words (~ 15).",
"coherence assessment (Barzilay and Lapata, 2005;",
"Cervone et al., 2018) and also introduce Utterance",
"SwitchBoard contains informal English dia-",
"logues collected from phone conversations be-",
"Replacement (UR), and Even Utterance Ordering",
"(EUO) as more challenging and dialogue-relevant",
"tween two mutually unknown human participants.",
"perturbation methods.Since each experiment",
"The participants were given only one of 70 possi-",
"follows a specific perturbation method, hence-",
"ble topics as initial topic to start a conversation but",
"forth, we refer to these perturbations as problem-",
"they werefree todiverge from that topic duringthe",
"conversation. So, there is no concrete topic associ-",
"domains:",
"ated with each dialogue in this dataset as it is the",
"Utterance Ordering (UO) We randomly per-",
"case for dialogues in DailyDialog",
"mute the order of utterances in dialogue. The orig-",
"DA labels in SwitchBoard are about 10 times",
"inal dialogue is preferred over the perturbed one.",
"more fine-grained than those in DailyDialog. For",
"example, a question utterance in SwitchBoard may",
"Utterance Insertion (U) We remove each ut-",
"have a fine-grained DA label such as Yes-No-",
"terance of a dialogue and then re-insert it in any",
"Question, Wh-Question, Rhetorical-Questions,",
"possible utterance position in the dialogue.We",
"etc. The distribution of these acts is however",
"assume that the original place of the utterance is",
"highly unbalanced in SwitchBoard: the most fre-",
"the best place for the insertion. Therefore, a co-",
"quent act label makes up for 36% of the utterances",
"herence model ideally discriminates the original",
"in the corpus, the three most frequent acts together",
"dialogue from the perturbed ones, which are ob",
"make up for 68% of the utterances, while most of",
"tained by re-inserting the removed utterance in any",
"the remaining act labels just make up for 1% or",
"utterance position except its original one. This",
"less of all the utterances.",
"problem-domain is more difficult to solve than UO",
"On average, dialogues in SwitchBoard contain",
"as the distinction between dialogues is in the posi-",
"more utterances than those in DailyDialog (192 vs",
"tion of only one utterance.",
"8)bututterances inSwitchBoardareshorterthan",
"Utterance Replacement (UR)We randomly re-",
"those in DailyDialog (9 vs 15). This means that",
"place one of the utterances in a dialogue with an-",
"dialogues in SwitchBoard are more likely to span",
"other utterance that is also randomly selected from",
"different topics than the ones in DailyDialog. The",
"another dialogue. The original dialogue is pre-",
"utterances in DailyDialog are explicitly cleaned of",
"ferred over the dialogue generated by UR. Un-",
"any noise, like “uh-oh\", or interruptions by the",
"like the other problem-domains, which perturb the",
"other speaker, as it is commonly the case for dia-",
"structure of a dialogue, this problem-domain per-",
"logues in SwitchBoard. While each dialogue turn",
"turbs the coherence of a dialogue at its semantic",
"of dialogues in DailyDialog contains only one ut-",
"level.",
"terance, dialogue turns in SwitchBoard may con-",
"sist of several utterances. That is why we con-",
"Utterance",
"Ordering",
"(EUO) This",
"problem-domain is similar to UO but here we",
"Even",
"sider each dialogue as a sequence of dialogue ut-",
"terances.",
"re-arrange the order of utterances that are said"
] |
{
"arxiv_id": "1908.08486",
"language": "en",
"page_idx": 4
}
|
||
2403.17385-en-0
|
2403.17385-en-0.jpg
|
## $\mathsf{ELLEN:Extremely~Lightly~Supervised~Learning~For~Efficient~Named~}$ Entity Recognition
Haris Riaz, Razvan-Gabriel Dumitru, Mihai Surdeanu
## University of Arizona Tucson, AZ, USA
$\{$hriaz, rdumitru, msurdeanu{@arizona.edu
Abstract
In this work, we revisit the problem of semi-supervised named entity recognition (NER) focusing on extremely light supervision, consisting of a lexicon containing only 10 examples per class. We introduce ELLEN, a simple, fully modular, neuro-symbolic method that blends fine-tuned language models with linguistic rules. These rules include insights such as“One Sense Per Discourse", using a Masked Language Model as an unsupervised NER, Ieveraging part-of-speech tags to identify and eliminate unlabeled entities as false negatives, and other intuitions about classifier confidence scores in local and global context. ELLEN achieves very strong performance on the CoNLL-2003 dataset when using the minimal supervision from the lexicon above. It also outperforms most existing (and considerably more complex) semi-supervised NER methods under the same supervision settings commonly used in the literature (i.e., 5% of the training data). Further, we evaluate our CoNLL-2003 model in a zero-shot scenario on WNUT-17 where we find that it outperforms GPT-3.5 and achieves comparable performance to GPT-4. In a zero-shot setting, ELLEN also achieves over 75% of the performance of a strong, fully supervised model trained on gold data. Our code is publicly available.
Keywords:semi-supervised learning, named entity recognition, neuro-symbolic, rules, language models,
modular architectures
### 1. Introduction
Named entity recognition (NER), i.e., the task of identifying named (and sometimes numeric) entities such person and organization names, drugs, protein names, diseases, and dates, is one of the earliest formal natural language processing (NLP) tasks (Grishman and Sundheim, 1996). NER remains critical to many real-world applications such as question answering and information extraction (Yadav and Bethard, 2019). Despite the tremendous progress observed on the NER task in the past almost three decades, we argue that there are several practical limitations in the way this task is generally formalized, which impact our understanding of what methods perform best in practice. In particular:
surveillance) that require the rapid development of custom models and where domain experts"do not want to come willingly and do not come cheaply."1 (2) While recent directions that use in-context learning (ICL) for NER with autoregressive decoderbased large language models (LLMs) perform well (Chen et al., 2023), they do not scale as well as encoder-based methods due to the decoder's high inference overhead; each generated token requires its own forward pass through the model.
(3) Recent trends rely mostly on neural networks (NNs) to learn the NER task, ignoring linguistic hints such as"one sense per discourse" (Gale et al., 1992) that might be present and are likely to be useful in lightly-supervised settings.
To remedy these limitations we propose an extremely lightly supervised scenario for NER, in which the only supervision comes in the form of a lexicon containing 10 examples per entity class. Importantly, the 10 examples are selected by a domain expert that does not have access to any dataset annotations. Further, we propose a simple NER approach for this scenario that is efficient and performs well despite the limited supervision. Our method uses an encoder-only inference strategy, but, at training time, it combines multiple strategies including language models and several linguistic heuristics. We call our method Extremely
$^{1}|$ARPA program manager, personal communication
(1) Current settings for the NER task require an amount of annotations that are unrealistic for many real-world applications. For example, a common setting for semi-supervised NER uses 5% of the CoNLL-2003 corpus' (Tjong Kim Sang and De Meulder (2003)) training data, or over 10K total tokens (Chen et al., 2020; Zheng et al., 2023). In our work (Vacareanu et al., 2024), we have observed that NER annotations take approximately 3.2 seconds per token in practice. Thus, annotating the equivalent amount of data in a new domain would take approximately 9 person hours. This is unrealistic in many scenarios (e.g., intelligence, pandemic
|
[
[
0.1275634765625,
0.1175537109375,
0.8818359375,
0.1175537109375,
0.8818359375,
0.1353759765625,
0.1275634765625,
0.1353759765625
],
[
0.387939453125,
0.1395263671875,
0.60107421875,
0.1395263671875,
0.60107421875,
0.15673828125,
0.387939453125,
0.15673828125
],
[
0.252685546875,
0.17333984375,
0.75146484375,
0.17333984375,
0.75146484375,
0.188232421875,
0.252685546875,
0.188232421875
],
[
0.42236328125,
0.189453125,
0.57861328125,
0.189453125,
0.57861328125,
0.203125,
0.42236328125,
0.203125
],
[
0.4375,
0.203125,
0.56689453125,
0.203125,
0.56689453125,
0.21728515625,
0.4375,
0.21728515625
],
[
0.343505859375,
0.21728515625,
0.66015625,
0.21728515625,
0.66015625,
0.23095703125,
0.343505859375,
0.23095703125
],
[
0.465087890625,
0.2493896484375,
0.5380859375,
0.2493896484375,
0.5380859375,
0.26416015625,
0.465087890625,
0.26416015625
],
[
0.12005615234375,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.276611328125,
0.12005615234375,
0.276611328125
],
[
0.12176513671875,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.289306640625,
0.12176513671875,
0.289306640625
],
[
0.12176513671875,
0.292236328125,
0.880859375,
0.292236328125,
0.880859375,
0.302734375,
0.12176513671875,
0.302734375
],
[
0.0218353271484375,
0.303955078125,
0.0621337890625,
0.303955078125,
0.0621337890625,
0.439453125,
0.0218353271484375,
0.439453125
],
[
0.118408203125,
0.303955078125,
0.88330078125,
0.3046875,
0.88330078125,
0.318359375,
0.118408203125,
0.317626953125
],
[
0.12176513671875,
0.319580078125,
0.8818359375,
0.319580078125,
0.8818359375,
0.330078125,
0.12176513671875,
0.330078125
],
[
0.118408203125,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.118408203125,
0.344482421875
],
[
0.12005615234375,
0.345703125,
0.8818359375,
0.345703125,
0.8818359375,
0.355712890625,
0.12005615234375,
0.355712890625
],
[
0.12005615234375,
0.35693359375,
0.884765625,
0.35693359375,
0.884765625,
0.37060546875,
0.12005615234375,
0.37060546875
],
[
0.12005615234375,
0.371826171875,
0.8818359375,
0.371826171875,
0.8818359375,
0.3818359375,
0.12005615234375,
0.3818359375
],
[
0.1175537109375,
0.3818359375,
0.884765625,
0.383056640625,
0.884765625,
0.396728515625,
0.1175537109375,
0.3955078125
],
[
0.12005615234375,
0.39794921875,
0.8818359375,
0.39794921875,
0.8818359375,
0.407958984375,
0.12005615234375,
0.407958984375
],
[
0.12005615234375,
0.410400390625,
0.1822509765625,
0.410400390625,
0.1822509765625,
0.421630859375,
0.12005615234375,
0.421630859375
],
[
0.12005615234375,
0.435302734375,
0.884765625,
0.435302734375,
0.884765625,
0.448974609375,
0.12005615234375,
0.448974609375
],
[
0.0218353271484375,
0.441162109375,
0.06463623046875,
0.441162109375,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.12005615234375,
0.449462890625,
0.26611328125,
0.449462890625,
0.26611328125,
0.460205078125,
0.12005615234375,
0.460205078125
],
[
0.225830078125,
0.47021484375,
0.382080078125,
0.47021484375,
0.382080078125,
0.485107421875,
0.225830078125,
0.485107421875
],
[
0.51318359375,
0.471435546875,
0.88330078125,
0.471435546875,
0.88330078125,
0.485107421875,
0.51318359375,
0.485107421875
],
[
0.51318359375,
0.486328125,
0.88330078125,
0.486328125,
0.88330078125,
0.49951171875,
0.51318359375,
0.49951171875
],
[
0.1175537109375,
0.50048828125,
0.489501953125,
0.50048828125,
0.489501953125,
0.51416015625,
0.1175537109375,
0.51416015625
],
[
0.51318359375,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.51318359375,
0.51416015625
],
[
0.1175537109375,
0.51416015625,
0.490234375,
0.515625,
0.490234375,
0.529296875,
0.1175537109375,
0.52783203125
],
[
0.0218353271484375,
0.52490234375,
0.06378173828125,
0.52490234375,
0.06378173828125,
0.60693359375,
0.0218353271484375,
0.60693359375
],
[
0.5146484375,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.5146484375,
0.533203125
],
[
0.118408203125,
0.52978515625,
0.489501953125,
0.52978515625,
0.489501953125,
0.54345703125,
0.118408203125,
0.54345703125
],
[
0.5146484375,
0.53369140625,
0.884765625,
0.53369140625,
0.884765625,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.12176513671875,
0.5458984375,
0.48779296875,
0.5458984375,
0.48779296875,
0.5556640625,
0.12176513671875,
0.5556640625
],
[
0.51318359375,
0.548828125,
0.884765625,
0.548828125,
0.884765625,
0.5625,
0.51318359375,
0.5625
],
[
0.118408203125,
0.55712890625,
0.489501953125,
0.55810546875,
0.489501953125,
0.572265625,
0.118408203125,
0.57177734375
],
[
0.5146484375,
0.56396484375,
0.88330078125,
0.56396484375,
0.88330078125,
0.57470703125,
0.5146484375,
0.57470703125
],
[
0.03863525390625,
0.572265625,
0.0487060546875,
0.572265625,
0.0487060546875,
0.57666015625,
0.03863525390625,
0.57666015625
],
[
0.12005615234375,
0.572265625,
0.489501953125,
0.572265625,
0.489501953125,
0.5859375,
0.12005615234375,
0.5859375
],
[
0.5146484375,
0.57666015625,
0.88330078125,
0.57666015625,
0.88330078125,
0.59033203125,
0.5146484375,
0.59033203125
],
[
0.12176513671875,
0.58837890625,
0.48779296875,
0.58837890625,
0.48779296875,
0.5986328125,
0.12176513671875,
0.5986328125
],
[
0.5146484375,
0.59130859375,
0.88330078125,
0.59130859375,
0.88330078125,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.12005615234375,
0.60107421875,
0.489501953125,
0.60107421875,
0.489501953125,
0.6142578125,
0.12005615234375,
0.6142578125
],
[
0.0218353271484375,
0.59326171875,
0.0621337890625,
0.59326171875,
0.0621337890625,
0.73095703125,
0.0218353271484375,
0.73095703125
],
[
0.51318359375,
0.60498046875,
0.8134765625,
0.60400390625,
0.8134765625,
0.6181640625,
0.51318359375,
0.619140625
],
[
0.118408203125,
0.615234375,
0.489501953125,
0.615234375,
0.489501953125,
0.62890625,
0.118408203125,
0.62890625
],
[
0.515625,
0.62451171875,
0.88330078125,
0.62451171875,
0.88330078125,
0.63818359375,
0.515625,
0.63818359375
],
[
0.12005615234375,
0.63134765625,
0.48779296875,
0.63134765625,
0.48779296875,
0.64111328125,
0.12005615234375,
0.64111328125
],
[
0.118408203125,
0.6435546875,
0.489501953125,
0.6435546875,
0.489501953125,
0.65673828125,
0.118408203125,
0.65673828125
],
[
0.5146484375,
0.6396484375,
0.88330078125,
0.6396484375,
0.88330078125,
0.65283203125,
0.5146484375,
0.65283203125
],
[
0.12005615234375,
0.65771484375,
0.48779296875,
0.65771484375,
0.48779296875,
0.66845703125,
0.12005615234375,
0.66845703125
],
[
0.515625,
0.65478515625,
0.88330078125,
0.65478515625,
0.88330078125,
0.6650390625,
0.515625,
0.6650390625
],
[
0.118408203125,
0.67138671875,
0.490234375,
0.67138671875,
0.490234375,
0.68505859375,
0.118408203125,
0.68505859375
],
[
0.515625,
0.66748046875,
0.884765625,
0.66748046875,
0.884765625,
0.68115234375,
0.515625,
0.68115234375
],
[
0.12005615234375,
0.68603515625,
0.489501953125,
0.68603515625,
0.489501953125,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.5146484375,
0.681640625,
0.77734375,
0.681640625,
0.77734375,
0.6953125,
0.5146484375,
0.6953125
],
[
0.118408203125,
0.70068359375,
0.21240234375,
0.70068359375,
0.21240234375,
0.71142578125,
0.118408203125,
0.71142578125
],
[
0.5341796875,
0.69775390625,
0.8818359375,
0.69775390625,
0.8818359375,
0.7080078125,
0.5341796875,
0.7080078125
],
[
0.5146484375,
0.71044921875,
0.88330078125,
0.71044921875,
0.88330078125,
0.7236328125,
0.5146484375,
0.7236328125
],
[
0.118408203125,
0.724609375,
0.489501953125,
0.7236328125,
0.489501953125,
0.73828125,
0.118408203125,
0.7392578125
],
[
0.5146484375,
0.7255859375,
0.88330078125,
0.7255859375,
0.88330078125,
0.736328125,
0.5146484375,
0.736328125
],
[
0.118408203125,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.118408203125,
0.7529296875
],
[
0.51318359375,
0.73828125,
0.884765625,
0.73828125,
0.884765625,
0.751953125,
0.51318359375,
0.751953125
],
[
0.12005615234375,
0.75439453125,
0.489501953125,
0.75439453125,
0.489501953125,
0.76708984375,
0.12005615234375,
0.76708984375
],
[
0.51318359375,
0.751953125,
0.88330078125,
0.7529296875,
0.88330078125,
0.7666015625,
0.51318359375,
0.765625
],
[
0.12176513671875,
0.76953125,
0.48779296875,
0.76953125,
0.48779296875,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.12005615234375,
0.7822265625,
0.490234375,
0.7822265625,
0.490234375,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.5146484375,
0.78076171875,
0.88330078125,
0.78076171875,
0.88330078125,
0.79443359375,
0.5146484375,
0.79443359375
],
[
0.12176513671875,
0.79736328125,
0.490234375,
0.79736328125,
0.490234375,
0.80810546875,
0.12176513671875,
0.80810546875
],
[
0.5146484375,
0.7958984375,
0.884765625,
0.7958984375,
0.884765625,
0.80859375,
0.5146484375,
0.80859375
],
[
0.12005615234375,
0.81005859375,
0.490234375,
0.81005859375,
0.490234375,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.515625,
0.81103515625,
0.8818359375,
0.81103515625,
0.8818359375,
0.82177734375,
0.515625,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.822265625,
0.88671875,
0.82470703125,
0.88671875,
0.83837890625,
0.51318359375,
0.8359375
],
[
0.118408203125,
0.83837890625,
0.490234375,
0.83837890625,
0.490234375,
0.8515625,
0.118408203125,
0.8515625
],
[
0.51318359375,
0.83837890625,
0.88330078125,
0.83837890625,
0.88330078125,
0.8515625,
0.51318359375,
0.8515625
],
[
0.12005615234375,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.12005615234375,
0.8662109375
],
[
0.51318359375,
0.8515625,
0.884765625,
0.8515625,
0.884765625,
0.865234375,
0.51318359375,
0.865234375
],
[
0.118408203125,
0.86767578125,
0.48779296875,
0.86767578125,
0.48779296875,
0.88134765625,
0.118408203125,
0.88134765625
],
[
0.51318359375,
0.8662109375,
0.88330078125,
0.8662109375,
0.88330078125,
0.8798828125,
0.51318359375,
0.8798828125
],
[
0.118408203125,
0.8818359375,
0.489501953125,
0.8818359375,
0.489501953125,
0.8955078125,
0.118408203125,
0.8955078125
],
[
0.1175537109375,
0.89453125,
0.490234375,
0.8955078125,
0.490234375,
0.91015625,
0.1175537109375,
0.9091796875
],
[
0.5341796875,
0.89453125,
0.8818359375,
0.8955078125,
0.8818359375,
0.91015625,
0.5341796875,
0.9091796875
]
] |
[
"ELLEN:ExtremelyLightly Supervised Learning ForEfficient Named",
"Entity Recognition",
"Haris Riaz, Razvan-Gabriel Dumitru, Mihai Surdeanu",
"University of Arizona",
"Tucson, AZ, USA",
"[hriaz, rdumitru, msurdeanu]@arizona.edu",
"Abstract",
"Inthiswork,werevisittheproblemofsemi-supervisednamedentityrecognition(NER)focusingonextremelylight",
"supervision,consistingofalexiconcontainingonly1oexamplesperclass.WeintroduceELLEN.asimple,fully",
"modular,neuro-symbolicmethodthatblendsfine-tuned languagemodelswith linguisticrules.Theserulesinclude",
" ",
"insights such as“One Sense Per Discourse\",using a Masked Language Model as an unsupervised NER,leveraging",
"part-of-speechtagsto identifyand eliminateunlabeled entities asfalsenegatives,andotherintuitions aboutclassifier",
"confidence scores in local and global context. ELLEN achieves very strong performance on the CoNLL-2003 dataset",
"whenusingtheminimalsupervisionfromthelexiconabove.Italsooutperformsmostexisting(andconsiderablymore",
"complex) semi-supervised NER methods under the same supervision settings commonly used in the literature (i.e.,",
"5%ofthetrainingdata).Further,weevaluateourCoNLL-2003modelinazero-shotscenarioonWNUT-17wherewe",
"find that it outperforms GPT-3.5 and achieves comparable performance to GPT-4. In a zero-shot setting, ELLEN also",
"achievesover75%oftheperformanceofastrong,fullysupervisedmodeltrainedongolddata.Ourcodeispublicly",
"available",
"Keywords:semi-supervised learning, named entity recognition, neuro-symbolic, rules, language models,",
"[T's]",
"modulararchitectures",
"1.Introduction",
"surveillance) that require the rapid development of",
"custommodels and where domain experts\"do not",
"Named entity recognition (NER), i.e., the task of",
"want to come willingly and do not come cheaply.\"",
"identifying named (and sometimes numeric) enti-",
"1A55",
"(2) While recent directions that use in-context learn-",
"ties such person and organization names, drugs,",
"ing (ICL) for NER with autoregressive decoder-",
"proteinnames,diseases,and dates,is one of",
"based large language models (LLMs) perform well",
"the earliest formal natural language processing",
"(Chenetal.,2023),theydonot scale aswell as",
"",
"(NLP)tasks (Grishman and Sundheim, 1996).NER",
"encoder-based methods due to the decoder's high",
"remainscriticaltomanyreal-worldapplications",
"inference overhead; each generated token requires",
"such as question answering and information ex-",
"C",
"its own forward pass through the model.",
"traction (Yadav and Bethard, 2019). Despite the",
"(3) Recent trends rely mostly on neural networks",
"tremendousprogressobservedontheNERtaskin",
"the past almost three decades, we argue that there",
"(NNs)to learn the NERtask, ignoring linguistic hints",
"are several practical limitationsinthewaythistask",
"such as“one sense per discourse\"(Gale et al.",
"is generally formalized, which impact our under-",
"1992) that might be present and are likely to be",
"standing of what methods perform best in practice.",
"useful in lightly-supervised settings.",
"Inparticular:",
"Toremedytheselimitationsweproposeanex",
"tremely lightly supervised scenario for NER, in",
"(1) Current settings for the NER task require an",
"whichtheonlysupervisioncomesintheformof",
"amount of annotations that are unrealistic for many",
"a lexicon containing 10 examples per entity class",
"real-world applications. For example, a common",
"Importantly, the 10 examples are selected by a",
"settingforsemi-supervisedNERuses5%ofthe",
"domain expert that does not have access to any",
"CoNLL-2003 corpus'(Tjong Kim Sang and De Meul-",
"dataset annotations. Further, we propose a simple",
"der(2003))training data,or over 10Ktotal to-",
"NER approach for this scenario that is efficient and",
"kens (Chen et al., 2020; Zheng et al., 2023). In our",
"performswelldespitethelimitedsupervision.Our",
"work (Vacareanu et al., 2024), we have observed",
"method uses an encoder-only inference strategy,",
"thatNERannotations take approximately3.2sec-",
"but, at training time, it combines multiple strate-",
"onds per token in practice. Thus, annotating the",
"gies including language models and several lin-",
"equivalent amount of data in a new domain would",
"guistic heuristics. We call our method Extremely",
"take approximately 9 person hours. This is unrealis-",
"tic in many scenarios (e.g., intelligence, pandemic",
"1IARPA program manager, personal communication"
] |
{
"arxiv_id": "2403.17385",
"language": "en",
"page_idx": 0
}
|
||
2403.17385-en-1
|
2403.17385-en-1.jpg
|
Lightly Supervised Learning for Efficient Named
Entity Recognition (ELLEN)$^2.$
Our main contributions are as follows:
(1) We demonstrate the effectiveness of combining language models with commonsense linguistic rules inspired by (Liao and Veeramachaneni, 2009) and aggregated under a self-training, modular, neuro-symbolic architecture. Our approach is considerably simpler than other complex statistical methods for semi-supervised NER (Nagesh and Surdeanu, 2018; Lakshmi Narayan et al., 2019; Peng et al., 2019; Zhou et al., 2022; Chen et al., 2019; Clark et al., 2018; Chen et al., 2020; Zheng et al., 2023, inter alia).
(2) Our approach includes a novel component called the Masked Language Modeling (MLM) Heuristic, which is a fully unsupervised NER method that achieves over 55% precision on the CoNLL-2003 NER dataset. Further, this component complements other self training as well as linguistic heuristics in the semi-supervised setting.
(3) We evaluate our method on CoNLL-2003 (Tjong Kim Sang and De Meulder (2003)) under three different degrees of supervision, and in a zero-shot setting on WNUT-17 (Derczynski et al. (2017)). On CoNLL, under the proposed setting of extremely limited supervision, we show that our method achieves an F1 score of 76.87%. Further, when we increase the degree of supervision to match other methods which are state-of-the-art in the semi-supervised NER setting, we find that our method achieves comparable performance. We also show that our method continues to scale, even when using all of the data available for supervision. In a zero-shot evaluation on WNUT-17, we find our method to be comparable to LLMs such as GPT-3.5 (OpenAI, 2023b) and GPT-4 (OpenAl, 2023a), while also obtaining over 75% of the performance of a fully supervised model trained on WNUT-17.
## 2. Related Works
Recently, Large Language Models (LLM's) have emerged as the dominant approach for a wide variety of NLP tasks, including Named Entity Recognition. (Wang et al., 2023) and (Zhou et al., 2023) show that LLM's consistenly achieve SOTA performance on many NER datasets. With In-Context Learning, LLM's have also proven to be very useful in the FewShot NER setting, as recently shown by (Ashok and Lipton, 2023). However, LLM's typically have a high inference overhead (Narayanan et al., 2023) and may not always perform well in specialized or low-resource domains. Moreover, there are increasing concerns about data contamination. (Golchin and Surdeanu, 2023) demonstrate
$^{2}$https://github.com/hriaz17/ELLEN
that GPT-3.5 and GPT-4 have encountered test data with labels from widely-used NLP benchmark datasets during pre-training. (Sainz et al., 2023) claim that this is true for CoNLL-2003, one of our evaluation datasets.
Focusing on Semi-Supervised NER, and not FewShot NER, current state-of-the-art methods include JointProp (Zheng et al.,2023), which is a multi-task learning framework that jointly tries to solve relation classification and NER using a heterogeneous graph structure. Semi-LADA (Chen et al., 2020) adapts the mixup data augmentation technique to sequence labeling, and then trains on linearly interpolated pairs of samples. Both of these methods use at least 5% of the labeled data as their most minimally supervised setting, which we argue, is an impractical level of supervision for semi-supervised NER. Another class of statistical methods (Peng et al., 2019; Zhou et al., 2022) try to solve the reliance on gold labels by resorting to distant supervision: they construct lexicons based on large dataset independent knowledge bases. These methods then use Positive-Unlabeled (PU) learning to train classifiers using only labeled positive examples and a set of unlabeled data containing both positives and negatives.
Other approaches like (Liu et al., 2019a) showed the benefits of augmenting neural NER taggers with external gazetteers. However, external gazetteers may not always be available for particular domains In contrast we propose a semi-supervised method for NER along the lines of the approach taken by (Liao and Veeramachaneni, 2009): one that combines the generalizability of contemporary deep learning with intuitive reasoning and linguistic insights, and we demonstrate its strong performance, under a setting of extremely low supervision. We posit that such an approach can rival more multifaceted statistical techniques for semi-supervised NER, such as PU learning & data augmentation, among others.
## 3. Proposed Method
As our method does not rely on any explicitly labeled texts, we begin by discarding all labels in the NER dataset (CoNLL-2003 in this paper). We then ask a domain expert to generate a small lexicon of 10 example named entities per class, for each of the four classes in CoNLL-2003, i.e., PER, ORG, LOC, and MISC. This lexicon: a) is sourced entirely from the tokens in the dataset; b) is constructed without looking at any of the labels in the dataset; c) does not rely on any external knowledge-base or dictionary; and d) serves as the sole source of "gold supervision" for our method. The domain expert is able to construct the lexicon (refer to Table 1) in less than 30 minutes for CoNLL-2003. We believe
|
[
[
0.12176513671875,
0.0748291015625,
0.491943359375,
0.0748291015625,
0.491943359375,
0.08905029296875,
0.12176513671875,
0.08905029296875
],
[
0.51123046875,
0.0736083984375,
0.88330078125,
0.0748291015625,
0.88330078125,
0.08905029296875,
0.51123046875,
0.0885009765625
],
[
0.118408203125,
0.0885009765625,
0.333251953125,
0.0885009765625,
0.333251953125,
0.1015625,
0.118408203125,
0.1015625
],
[
0.51318359375,
0.08905029296875,
0.88330078125,
0.08905029296875,
0.88330078125,
0.10272216796875,
0.51318359375,
0.10272216796875
],
[
0.1368408203125,
0.10394287109375,
0.418212890625,
0.10394287109375,
0.418212890625,
0.1175537109375,
0.1368408203125,
0.1175537109375
],
[
0.51318359375,
0.10394287109375,
0.88330078125,
0.10394287109375,
0.88330078125,
0.1175537109375,
0.51318359375,
0.1175537109375
],
[
0.12176513671875,
0.1217041015625,
0.489501953125,
0.1217041015625,
0.489501953125,
0.1318359375,
0.12176513671875,
0.1318359375
],
[
0.5146484375,
0.1175537109375,
0.88330078125,
0.1175537109375,
0.88330078125,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.12005615234375,
0.1353759765625,
0.490234375,
0.1353759765625,
0.490234375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.5146484375,
0.1341552734375,
0.66162109375,
0.1341552734375,
0.66162109375,
0.144287109375,
0.5146484375,
0.144287109375
],
[
0.118408203125,
0.1484375,
0.489501953125,
0.1484375,
0.489501953125,
0.162109375,
0.118408203125,
0.162109375
],
[
0.5283203125,
0.146728515625,
0.884765625,
0.1455078125,
0.884765625,
0.1602783203125,
0.5283203125,
0.160888671875
],
[
0.12005615234375,
0.163330078125,
0.490234375,
0.163330078125,
0.490234375,
0.177001953125,
0.12005615234375,
0.177001953125
],
[
0.5146484375,
0.160888671875,
0.884765625,
0.160888671875,
0.884765625,
0.174560546875,
0.5146484375,
0.174560546875
],
[
0.12005615234375,
0.1776123046875,
0.489501953125,
0.1776123046875,
0.489501953125,
0.191162109375,
0.12005615234375,
0.191162109375
],
[
0.5146484375,
0.17578125,
0.884765625,
0.17578125,
0.884765625,
0.189453125,
0.5146484375,
0.189453125
],
[
0.12176513671875,
0.193603515625,
0.489501953125,
0.193603515625,
0.489501953125,
0.2037353515625,
0.12176513671875,
0.2037353515625
],
[
0.51318359375,
0.189453125,
0.8818359375,
0.189453125,
0.8818359375,
0.203125,
0.51318359375,
0.203125
],
[
0.118408203125,
0.2060546875,
0.490234375,
0.2060546875,
0.490234375,
0.2197265625,
0.118408203125,
0.2197265625
],
[
0.51123046875,
0.2037353515625,
0.88330078125,
0.2037353515625,
0.88330078125,
0.21728515625,
0.51123046875,
0.21728515625
],
[
0.12176513671875,
0.221435546875,
0.489501953125,
0.221435546875,
0.489501953125,
0.232177734375,
0.12176513671875,
0.232177734375
],
[
0.5146484375,
0.218505859375,
0.884765625,
0.218505859375,
0.884765625,
0.232177734375,
0.5146484375,
0.232177734375
],
[
0.118408203125,
0.2340087890625,
0.489501953125,
0.2340087890625,
0.489501953125,
0.2476806640625,
0.118408203125,
0.2476806640625
],
[
0.5146484375,
0.2327880859375,
0.88330078125,
0.2327880859375,
0.88330078125,
0.2464599609375,
0.5146484375,
0.2464599609375
],
[
0.12005615234375,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.12005615234375,
0.262451171875
],
[
0.51123046875,
0.2464599609375,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.12005615234375,
0.26318359375,
0.285400390625,
0.26318359375,
0.285400390625,
0.276611328125,
0.12005615234375,
0.276611328125
],
[
0.5146484375,
0.26123046875,
0.884765625,
0.26123046875,
0.884765625,
0.27490234375,
0.5146484375,
0.27490234375
],
[
0.118408203125,
0.279052734375,
0.48779296875,
0.279052734375,
0.48779296875,
0.292236328125,
0.118408203125,
0.292236328125
],
[
0.51318359375,
0.275634765625,
0.88330078125,
0.275634765625,
0.88330078125,
0.289306640625,
0.51318359375,
0.289306640625
],
[
0.51123046875,
0.2880859375,
0.88330078125,
0.2880859375,
0.88330078125,
0.3017578125,
0.51123046875,
0.3017578125
],
[
0.12005615234375,
0.29443359375,
0.489501953125,
0.29443359375,
0.489501953125,
0.30810546875,
0.12005615234375,
0.30810546875
],
[
0.12005615234375,
0.308837890625,
0.489501953125,
0.308837890625,
0.489501953125,
0.322509765625,
0.12005615234375,
0.322509765625
],
[
0.51318359375,
0.303955078125,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51318359375,
0.317626953125
],
[
0.12005615234375,
0.322509765625,
0.489501953125,
0.322509765625,
0.489501953125,
0.336181640625,
0.12005615234375,
0.336181640625
],
[
0.5146484375,
0.317626953125,
0.88330078125,
0.317626953125,
0.88330078125,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.5146484375,
0.33203125,
0.88330078125,
0.33203125,
0.88330078125,
0.345703125,
0.5146484375,
0.345703125
],
[
0.12176513671875,
0.33837890625,
0.48779296875,
0.33837890625,
0.48779296875,
0.3486328125,
0.12176513671875,
0.3486328125
],
[
0.51318359375,
0.344482421875,
0.8818359375,
0.3466796875,
0.8818359375,
0.35986328125,
0.51318359375,
0.358154296875
],
[
0.118408203125,
0.350830078125,
0.489501953125,
0.350830078125,
0.489501953125,
0.364013671875,
0.118408203125,
0.364013671875
],
[
0.1175537109375,
0.364013671875,
0.42236328125,
0.366455078125,
0.42236328125,
0.380126953125,
0.1175537109375,
0.377685546875
],
[
0.5146484375,
0.35986328125,
0.88330078125,
0.35986328125,
0.88330078125,
0.37353515625,
0.5146484375,
0.37353515625
],
[
0.5146484375,
0.374755859375,
0.884765625,
0.374755859375,
0.884765625,
0.388427734375,
0.5146484375,
0.388427734375
],
[
0.12005615234375,
0.3818359375,
0.489501953125,
0.3818359375,
0.489501953125,
0.3955078125,
0.12005615234375,
0.3955078125
],
[
0.51318359375,
0.38720703125,
0.884765625,
0.388427734375,
0.884765625,
0.402099609375,
0.51318359375,
0.40087890625
],
[
0.12005615234375,
0.396728515625,
0.489501953125,
0.396728515625,
0.489501953125,
0.410400390625,
0.12005615234375,
0.410400390625
],
[
0.515625,
0.405029296875,
0.88330078125,
0.405029296875,
0.88330078125,
0.4150390625,
0.515625,
0.4150390625
],
[
0.12176513671875,
0.412109375,
0.48779296875,
0.412109375,
0.48779296875,
0.4228515625,
0.12176513671875,
0.4228515625
],
[
0.51318359375,
0.41748046875,
0.88671875,
0.41748046875,
0.88671875,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.118408203125,
0.424560546875,
0.489501953125,
0.424560546875,
0.489501953125,
0.438232421875,
0.118408203125,
0.438232421875
],
[
0.12005615234375,
0.439453125,
0.490234375,
0.439453125,
0.490234375,
0.453125,
0.12005615234375,
0.453125
],
[
0.5146484375,
0.43408203125,
0.755859375,
0.43408203125,
0.755859375,
0.44482421875,
0.5146484375,
0.44482421875
],
[
0.52880859375,
0.4453125,
0.884765625,
0.44482421875,
0.884765625,
0.458984375,
0.52880859375,
0.460205078125
],
[
0.12176513671875,
0.454833984375,
0.48779296875,
0.454833984375,
0.48779296875,
0.465576171875,
0.12176513671875,
0.465576171875
],
[
0.5146484375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.474365234375,
0.5146484375,
0.474365234375
],
[
0.12176513671875,
0.4697265625,
0.48779296875,
0.4697265625,
0.48779296875,
0.479736328125,
0.12176513671875,
0.479736328125
],
[
0.1175537109375,
0.482177734375,
0.48779296875,
0.482177734375,
0.48779296875,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.515625,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.48681640625,
0.515625,
0.48681640625
],
[
0.5146484375,
0.4892578125,
0.88330078125,
0.4892578125,
0.88330078125,
0.5029296875,
0.5146484375,
0.5029296875
],
[
0.1175537109375,
0.49755859375,
0.486083984375,
0.49755859375,
0.486083984375,
0.50830078125,
0.1175537109375,
0.50830078125
],
[
0.5146484375,
0.5029296875,
0.88330078125,
0.5029296875,
0.88330078125,
0.5166015625,
0.5146484375,
0.5166015625
],
[
0.12176513671875,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.5224609375,
0.12176513671875,
0.5224609375
],
[
0.51318359375,
0.51708984375,
0.88330078125,
0.51708984375,
0.88330078125,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.118408203125,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.118408203125,
0.5380859375
],
[
0.51123046875,
0.53076171875,
0.884765625,
0.5322265625,
0.884765625,
0.5458984375,
0.51123046875,
0.54443359375
],
[
0.118408203125,
0.53759765625,
0.489501953125,
0.5380859375,
0.489501953125,
0.552734375,
0.118408203125,
0.5517578125
],
[
0.118408203125,
0.552734375,
0.489501953125,
0.552734375,
0.489501953125,
0.56640625,
0.118408203125,
0.56640625
],
[
0.515625,
0.54736328125,
0.88330078125,
0.54736328125,
0.88330078125,
0.55810546875,
0.515625,
0.55810546875
],
[
0.51318359375,
0.55859375,
0.88330078125,
0.55859375,
0.88330078125,
0.572265625,
0.51318359375,
0.572265625
],
[
0.118408203125,
0.56787109375,
0.489501953125,
0.56787109375,
0.489501953125,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.5146484375,
0.57470703125,
0.884765625,
0.57470703125,
0.884765625,
0.58837890625,
0.5146484375,
0.58837890625
],
[
0.12005615234375,
0.5830078125,
0.489501953125,
0.5830078125,
0.489501953125,
0.59326171875,
0.12005615234375,
0.59326171875
],
[
0.51318359375,
0.58837890625,
0.884765625,
0.58837890625,
0.884765625,
0.6015625,
0.51318359375,
0.6015625
],
[
0.12005615234375,
0.595703125,
0.489501953125,
0.595703125,
0.489501953125,
0.609375,
0.12005615234375,
0.609375
],
[
0.5146484375,
0.6025390625,
0.88671875,
0.6025390625,
0.88671875,
0.6162109375,
0.5146484375,
0.6162109375
],
[
0.12005615234375,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.12005615234375,
0.62353515625
],
[
0.1175537109375,
0.62255859375,
0.4130859375,
0.62255859375,
0.4130859375,
0.6357421875,
0.1175537109375,
0.6357421875
],
[
0.5146484375,
0.61767578125,
0.884765625,
0.61767578125,
0.884765625,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.51318359375,
0.6298828125,
0.88330078125,
0.6298828125,
0.88330078125,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.515625,
0.646484375,
0.62109375,
0.646484375,
0.62109375,
0.65771484375,
0.515625,
0.65771484375
],
[
0.214111328125,
0.65625,
0.252685546875,
0.65625,
0.252685546875,
0.66845703125,
0.214111328125,
0.66845703125
],
[
0.2476806640625,
0.65478515625,
0.39208984375,
0.65478515625,
0.39208984375,
0.66943359375,
0.2476806640625,
0.66943359375
],
[
0.12005615234375,
0.68310546875,
0.489501953125,
0.68310546875,
0.489501953125,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.591796875,
0.681640625,
0.8017578125,
0.681640625,
0.8017578125,
0.69677734375,
0.591796875,
0.69677734375
],
[
0.12005615234375,
0.69775390625,
0.490234375,
0.69775390625,
0.490234375,
0.71142578125,
0.12005615234375,
0.71142578125
],
[
0.12005615234375,
0.7119140625,
0.490234375,
0.7119140625,
0.490234375,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.5146484375,
0.71044921875,
0.884765625,
0.71044921875,
0.884765625,
0.7236328125,
0.5146484375,
0.7236328125
],
[
0.118408203125,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.118408203125,
0.73828125
],
[
0.51318359375,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.736328125,
0.51318359375,
0.736328125
],
[
0.12005615234375,
0.74072265625,
0.490234375,
0.74072265625,
0.490234375,
0.75439453125,
0.12005615234375,
0.75439453125
],
[
0.51318359375,
0.73828125,
0.88330078125,
0.73828125,
0.88330078125,
0.751953125,
0.51318359375,
0.751953125
],
[
0.118408203125,
0.75439453125,
0.489501953125,
0.7529296875,
0.489501953125,
0.7666015625,
0.118408203125,
0.76708984375
],
[
0.5146484375,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.118408203125,
0.7685546875,
0.489501953125,
0.7685546875,
0.489501953125,
0.7822265625,
0.118408203125,
0.7822265625
],
[
0.5146484375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.12005615234375,
0.783203125,
0.489501953125,
0.783203125,
0.489501953125,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51123046875,
0.78076171875,
0.88330078125,
0.78076171875,
0.88330078125,
0.79443359375,
0.51123046875,
0.79443359375
],
[
0.12005615234375,
0.79638671875,
0.490234375,
0.79638671875,
0.490234375,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.51123046875,
0.7958984375,
0.88330078125,
0.79638671875,
0.88330078125,
0.81005859375,
0.51123046875,
0.80859375
],
[
0.118408203125,
0.81103515625,
0.48779296875,
0.81103515625,
0.48779296875,
0.82470703125,
0.118408203125,
0.82470703125
],
[
0.51123046875,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82373046875,
0.51123046875,
0.82373046875
],
[
0.12005615234375,
0.826171875,
0.489501953125,
0.826171875,
0.489501953125,
0.8388671875,
0.12005615234375,
0.8388671875
],
[
0.51318359375,
0.822265625,
0.884765625,
0.82470703125,
0.884765625,
0.83837890625,
0.51318359375,
0.8359375
],
[
0.1175537109375,
0.8388671875,
0.491943359375,
0.84033203125,
0.491943359375,
0.85400390625,
0.1175537109375,
0.8525390625
],
[
0.51123046875,
0.83740234375,
0.884765625,
0.83837890625,
0.884765625,
0.8525390625,
0.51123046875,
0.8515625
],
[
0.118408203125,
0.85400390625,
0.490234375,
0.85400390625,
0.490234375,
0.86767578125,
0.118408203125,
0.86767578125
],
[
0.5146484375,
0.8525390625,
0.88330078125,
0.8525390625,
0.88330078125,
0.8662109375,
0.5146484375,
0.8662109375
],
[
0.118408203125,
0.8681640625,
0.489501953125,
0.8681640625,
0.489501953125,
0.8818359375,
0.118408203125,
0.8818359375
],
[
0.51318359375,
0.86767578125,
0.88330078125,
0.86767578125,
0.88330078125,
0.88134765625,
0.51318359375,
0.88134765625
],
[
0.5146484375,
0.88134765625,
0.88330078125,
0.88134765625,
0.88330078125,
0.89453125,
0.5146484375,
0.89453125
],
[
0.1376953125,
0.8935546875,
0.438232421875,
0.8955078125,
0.438232421875,
0.91015625,
0.1376953125,
0.90771484375
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Lightly Supervised Learning for Efficient Named",
"that GPT-3.5 and GPT-4 have encountered test",
"EntityRecognition (ELLEN)?",
"data with labels from widely-used NLP benchmark",
"Our main contributions are as follows:",
"datasets during pre-training. (Sainz et al., 2023)",
"(1)Wedemonstratetheeffectivenessofcombin-",
"claim that this is true for CoNLL-2003, one of our",
"ing language models with commonsense linguis-",
"evaluationdatasets",
"tic rules inspired by (Liao and Veeramachaneni,",
"Focusing on Semi-Supervised NER, and not",
"2009) and aggregated under a self-training, mod-",
"FewShot NER, current state-of-the-art methods in-",
"ular, neuro-symbolic architecture. Our approach",
"clude JointProp (Zheng et al., 2023), which is a",
"isconsiderablysimplerthanothercomplexstatis-",
"multi-task learning framework that jointly tries to",
"tical methods for semi-supervised NER (Nagesh",
"solve relation classification and NER using a het",
"andSurdeanu,2018;LakshmiNarayanetal.,2019;",
"erogeneous graph structure.Semi-LADA (Chen",
"Peng et al., 2019; Zhou et al., 2022; Chen et al..",
"et al., 2020) adapts the mixup data augmentation",
"2019; Clark et al., 2018; Chen et al., 2020; Zheng",
"technique to sequence labeling, and then trains",
"et al., 2023, inter alia).",
"on linearly interpolated pairs of samples. Both of",
"(2) Our approach includes a novel component",
"these methods use at least 5% of the labeled data",
"as their most minimally supervised setting, which",
"called the Masked Language Modeling (MLM)",
"Heuristic, which is a fully unsupervised NER",
"we argue, is an impractical level of supervision for",
"method that achieves over 55% precision on the",
"semi-supervised NER.Another class of statistical",
"methods (Peng et al., 2019; Zhou et al., 2022) try",
"CoNLL-2003NERdataset.Further,thiscomponent",
"to solve the reliance on gold labels by resorting to",
"complements other self training as well as linguistic",
"heuristics in the semi-supervised setting.",
"distant supervision: they construct lexicons based",
"on large dataset independent knowledge bases.",
"(3) We evaluate our method on CoNLL-2003 (Tjong",
"These methods then use Positive-Unlabeled (PU)",
"Kim Sang and De Meulder (2003) under three dif-",
"learningtotrainclassifiersusingonlylabeledposi-",
"ferentdegrees ofsupervision,andin azero-shot",
"tive examples and a set of unlabeled data contain-",
"setting on WNUT-17 (Derczynski et al. (2017). On",
"CoNLL, under the proposed setting of extremely lim-",
"ingbothpositivesandnegatives.",
"Other approaches like (Liu et al.,2019a) showed",
"itedsupervision,weshowthatourmethodachieves",
"the benefits of augmenting neural NERtaggers with",
"anF1scoreof76.87%.Further,whenweincrease",
"the degree of supervision to match other methods",
"externalgazetteers.However,externalgazetteers",
"may not always be available for particular domains",
"whichare state-of-the-artinthe semi-supervised",
"In contrast we propose a semi-supervised method",
"NER setting,wefindthat our method achieves",
"for NER along the lines of the approach taken by",
"comparable performance.We also show that our",
"(Liao and Veeramachaneni, 2009): one that com-",
"method continues to scale, even when using all of",
"the data available for supervision. In a zero-shot",
"binesthegeneralizabilityofcontemporarydeep",
"learning with intuitive reasoning and linguistic in-",
"evaluation on WNUT-17, we find our method to be",
"sights, and we demonstrate its strong performance,",
"comparabletoLLMssuchasGPT-3.5(OpenAl,",
"under a setting of extremely low supervision. We",
"2023b) and GPT-4 (OpenAl, 2023a), while also",
"posit that such an approach can rival more multi-",
"obtaining over 75% of the performance of a fully",
"supervised model trained on WNUT-17.",
"faceted statistical techniques for semi-supervised",
"NER, such as PU learning & data augmentation,",
"amongothers.",
"2.",
"RelatedWorks",
"Recently, Large Language Models (LLM's) have",
"3.Proposed Method",
"emerged as the dominant approach for a wide vari-",
"ety of NLP tasks, including Named Entity Recog-",
"As our method does not rely on any explicitly la-",
"nition. (Wang et al., 2023) and (Zhou et al., 2023)",
"beledtexts,webeginby discardingall labels in the",
"show that LLM's consistenly achieve SOTA perfor-",
"NER dataset (CoNLL-2003 in this paper). We then",
"mance on manyNER datasets. With In-Context",
"askadomainexpertto generatea small lexicon",
"Learning, LLM's have also proven to be very useful",
"of 10 example named entities per class, for each",
"in the FewShot NER setting, as recently shown by",
"of the four classes in CoNLL-2003, i.e., PER, ORG",
"(Ashok and Lipton, 2023). However, LLM's typi-",
"Loc,and MIsc.This lexicon:a)is sourced entirely",
"cally have a high inference overhead (Narayanan",
"from the tokens in the dataset; b) is constructed",
"et al., 2023) and may not always perform well in",
"without looking at any of the labels in the dataset;",
"specialized or low-resource domains. Moreover,",
"c) does not rely on any external knowledge-base or",
"there are increasing concerns about data contami-",
"dictionary; and d) serves as the sole source of \"gold",
"nation. (Golchin and Surdeanu, 2023) demonstrate",
"supervision\"for our method. The domain expert",
"is able to construct the lexicon (refer to Table 1) in",
"2https://github.com/hriaz17/ELLEN",
"less than 30 minutes for CoNLL-2003. We believe"
] |
{
"arxiv_id": "2403.17385",
"language": "en",
"page_idx": 1
}
|
||
1811.07806-en-0
|
1811.07806-en-0.jpg
|
Pareto Optimization for Subset Selection with
${\text{Dynamic Cost Constraints}}^*$
Vahid Roostapour$^{\dagger1}$, Aneta Neumann$^{\ddagger1}$,Frank Neumann$^{81}$ and
Tobias Friedrich$^{q[2]}$
$^{1}$Optimisation and Logistics, School of Computer Science, The
University of Adelaide, Adelaide, Australia
$^{2}$Chair of Algorithm Engineering, Hasso Plattner Institute, Potsdam,
Germany
October 12, 2021
Abstract
We consider the subset selection problem for function $f$ with constraint bound $B$ that changes over time. Within the area of submodular optimization, various greedy approaches are commonly used. For dynamic environments we observe that the adaptive variants of these greedy approaches are not able to maintain their approximation quality. Investigating the recently introduced POMC Pareto optimization approach, we show that this algorithm efficiently computes ${\mathrm{a~}\phi=(\dot{\alpha}_{f}/2)(1-\frac{1}{e^{\alpha}f})}$-approximation, where $\alpha_f$ is the submodularity ratio of $f$, for each possible constraint bound $b\leq B.$ Furthermore, we show that POMC is able to adapt its set of solutions quickly in the case that $B$ increases. Our experimental investigations for the influence maximization in social networks show the advantage of POMC over generalized greedy algorithms. We also consider EAMC, a new evolutionary algorithm with polynomial expected time guarantee to maintain $\phi$ approximation ratio, and NSGA-II with two different population sizes as advanced multi-objective optimization algorithm, to demonstrate their challenges in optimizing the maximum coverage problem. Our empirical analysis shows that, within the same number of evaluations, POMC is able to perform as good as NSGA-II under linear constraint, while EAMC performs significantly worse than all considered algorithms in most cases.
$^*$A preliminary version of this article has been presented at the Thirty-Third AAAI Conference on Arti-
ficial Intelligence (AAAI 2019) [28]
$_{+}^{\dagger}$vahid.roostapour@ adelaide.edu.au, vahid.roostapour@gmail.com
$^{\ddagger}$aneta.neumann@ adelaide.edu.au $^{8}$frank.neumann@ adelaide.edu.au $^{\mathrm{T}}$tobias.friedrich@hpi.de
1
|
[
[
0.2418212890625,
0.2083740234375,
0.75634765625,
0.2083740234375,
0.75634765625,
0.2259521484375,
0.2418212890625,
0.2259521484375
],
[
0.3505859375,
0.2342529296875,
0.65869140625,
0.2342529296875,
0.65869140625,
0.255126953125,
0.3505859375,
0.255126953125
],
[
0.0228729248046875,
0.270751953125,
0.05963134765625,
0.270751953125,
0.05963134765625,
0.40283203125,
0.0228729248046875,
0.40283203125
],
[
0.253173828125,
0.2734375,
0.74609375,
0.2734375,
0.74609375,
0.291748046875,
0.253173828125,
0.291748046875
],
[
0.4248046875,
0.29345703125,
0.5703125,
0.291748046875,
0.5703125,
0.30615234375,
0.4248046875,
0.30810546875
],
[
0.250732421875,
0.32080078125,
0.74609375,
0.322509765625,
0.74609375,
0.34033203125,
0.250732421875,
0.33837890625
],
[
0.326904296875,
0.339111328125,
0.6689453125,
0.34033203125,
0.6689453125,
0.35595703125,
0.326904296875,
0.354736328125
],
[
0.2303466796875,
0.354736328125,
0.765625,
0.356689453125,
0.765625,
0.375,
0.2303466796875,
0.37255859375
],
[
0.461669921875,
0.376953125,
0.53515625,
0.376953125,
0.53515625,
0.391357421875,
0.461669921875,
0.391357421875
],
[
0.0228729248046875,
0.401611328125,
0.062103271484375,
0.401611328125,
0.062103271484375,
0.4892578125,
0.0228729248046875,
0.4892578125
],
[
0.4306640625,
0.416748046875,
0.56640625,
0.416748046875,
0.56640625,
0.43115234375,
0.4306640625,
0.43115234375
],
[
0.047393798828125,
0.481689453125,
0.052276611328125,
0.481689453125,
0.052276611328125,
0.4873046875,
0.047393798828125,
0.4873046875
],
[
0.4697265625,
0.47607421875,
0.529296875,
0.47607421875,
0.529296875,
0.490478515625,
0.4697265625,
0.490478515625
],
[
0.0228729248046875,
0.4873046875,
0.05963134765625,
0.4873046875,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.281005859375,
0.496826171875,
0.74169921875,
0.496826171875,
0.74169921875,
0.51025390625,
0.281005859375,
0.51025390625
],
[
0.2607421875,
0.51220703125,
0.73876953125,
0.51220703125,
0.73876953125,
0.52294921875,
0.2607421875,
0.52294921875
],
[
0.256591796875,
0.5244140625,
0.740234375,
0.5244140625,
0.740234375,
0.53857421875,
0.256591796875,
0.53857421875
],
[
0.256591796875,
0.53857421875,
0.740234375,
0.53857421875,
0.740234375,
0.5517578125,
0.256591796875,
0.5517578125
],
[
0.25732421875,
0.55322265625,
0.459228515625,
0.55322265625,
0.459228515625,
0.56298828125,
0.25732421875,
0.56298828125
],
[
0.455810546875,
0.5537109375,
0.740234375,
0.5537109375,
0.740234375,
0.564453125,
0.455810546875,
0.564453125
],
[
0.256591796875,
0.56640625,
0.74169921875,
0.56640625,
0.74169921875,
0.580078125,
0.256591796875,
0.580078125
],
[
0.2548828125,
0.580078125,
0.74267578125,
0.57763671875,
0.74267578125,
0.59228515625,
0.2548828125,
0.5947265625
],
[
0.256591796875,
0.5947265625,
0.74169921875,
0.59326171875,
0.74169921875,
0.60693359375,
0.256591796875,
0.60791015625
],
[
0.256591796875,
0.60791015625,
0.740234375,
0.60791015625,
0.740234375,
0.6220703125,
0.256591796875,
0.6220703125
],
[
0.25732421875,
0.6220703125,
0.73876953125,
0.62060546875,
0.73876953125,
0.6337890625,
0.25732421875,
0.63525390625
],
[
0.25732421875,
0.63720703125,
0.73876953125,
0.63720703125,
0.73876953125,
0.64794921875,
0.25732421875,
0.64794921875
],
[
0.256591796875,
0.6494140625,
0.74169921875,
0.6494140625,
0.74169921875,
0.66357421875,
0.256591796875,
0.66357421875
],
[
0.256591796875,
0.66357421875,
0.740234375,
0.66357421875,
0.740234375,
0.6767578125,
0.256591796875,
0.6767578125
],
[
0.256591796875,
0.6767578125,
0.74169921875,
0.6767578125,
0.74169921875,
0.68994140625,
0.256591796875,
0.68994140625
],
[
0.256591796875,
0.68994140625,
0.740234375,
0.68994140625,
0.740234375,
0.7041015625,
0.256591796875,
0.7041015625
],
[
0.25732421875,
0.70556640625,
0.73876953125,
0.70556640625,
0.73876953125,
0.71630859375,
0.25732421875,
0.71630859375
],
[
0.256591796875,
0.71826171875,
0.740234375,
0.71826171875,
0.740234375,
0.73193359375,
0.256591796875,
0.73193359375
],
[
0.256591796875,
0.73095703125,
0.5595703125,
0.73095703125,
0.5595703125,
0.744140625,
0.256591796875,
0.744140625
],
[
0.236083984375,
0.751953125,
0.779296875,
0.751953125,
0.779296875,
0.76513671875,
0.236083984375,
0.76513671875
],
[
0.2156982421875,
0.7646484375,
0.408447265625,
0.7646484375,
0.408447265625,
0.77783203125,
0.2156982421875,
0.77783203125
],
[
0.236083984375,
0.7763671875,
0.5849609375,
0.7763671875,
0.5849609375,
0.79052734375,
0.236083984375,
0.79052734375
],
[
0.23779296875,
0.79052734375,
0.414306640625,
0.79052734375,
0.414306640625,
0.80029296875,
0.23779296875,
0.80029296875
],
[
0.2344970703125,
0.80029296875,
0.416748046875,
0.80029296875,
0.416748046875,
0.814453125,
0.2344970703125,
0.814453125
],
[
0.2344970703125,
0.81298828125,
0.366943359375,
0.81298828125,
0.366943359375,
0.8271484375,
0.2344970703125,
0.8271484375
],
[
0.494384765625,
0.876953125,
0.50341796875,
0.876953125,
0.50341796875,
0.8896484375,
0.494384765625,
0.8896484375
]
] |
[
"Pareto Optimizationfor Subset Selection with",
"Dynamic Cost Constraints*",
"1",
"Vahid Roostapourt1, Aneta Neumannt1, Frank Neumanns' and",
"Tobias Friedrich2",
"1Optimisation and Logistics, School of Computer Science, The",
"University of Adelaide, Adelaide, Australia",
"2Chair of Algorithm Engineering, Hasso Plattner Institute, Potsdam,",
"Germany",
"[SA'S]",
"October 12, 2021",
"",
"Abstract",
"CA90801181A10",
"We consider the subset selection problem for function f with constraint bound",
"B that changes over time.Within the area of submodular optimization,vari-",
"ous greedy approaches are commonly used. For dynamic environments we ob-",
"serve that the adaptive variants of these greedy approaches are not able to main-",
"tain their approximation quality.",
"InvestigatingtherecentlyintroducedPOMC",
"Pareto optimization approach, we show that this algorithm efficiently computes",
"a = (αy /2)(1 - )-approximation, where αy is the submodularity ratio of",
"f,for each possible constraint bound b≤B.Furthermore,we show that POMC",
"is able to adapt its set of solutions quickly in the case that B increases. Our ex-",
"perimental investigations for the influence maximization in social networks show",
"the advantage of POMC over generalized greedy algorithms.We also consider",
"EAMC, a new evolutionary algorithm with polynomial expected time guarantee",
"to maintain approximation ratio, and NSGA-II with two different population",
"sizes as advanced multi-objective optimization algorithm, to demonstrate their",
"challenges in optimizing the maximum coverage problem. Our empirical analy-",
"sis shows that,within the same number of evaluations, POMC is abletoperform",
"as good as NSGA-II under linear constraint, while EAMC performs significantly",
"worse than all considered algorithms in most cases.",
"*A preliminary version of this article has been presented at the Thirty-Third AAAI Conference on Arti-",
"ficial Intelligence (AAAI 2019) [28]",
"fvahid.roostapour@adelaide.edu.au,vahid.roostapour@gmail.com",
"faneta.neumann@adelaide.edu.au",
"$ frank.neumann@ adelaide.edu.au",
"Itobias.friedrich@hpi.de",
"1"
] |
{
"arxiv_id": "1811.07806",
"language": "en",
"page_idx": 0
}
|
||
1811.07806-en-1
|
1811.07806-en-1.jpg
|
# $\textbf{1}$ $\textbf{Introduction}$
Many combinatorial optimization problems consist of optimizing a given function under a given constraint. Constraints often limit the resources available to solve a given problem and may change over time. For example, in the area of supply chain management, the problem may be constrained by the number of vehicles available that may change during the optimization process due to vehicle failures and vehicles being repaired or added as a new resource. As another example, consider the problem of using a train to transport extracted materials form a number of mines to a port. In this example, the number of available wagons could change during the path because of different reasons. The question would be how to re-allocate the available wagons to transform the high priority materials first.
Evolutionary algorithms as general-purpose problem solvers have been widely used to tackle dynamic optimization problems theoretically and empirically [32]. Many theoretical investigations considered the performance of evolutionary algorithms in the dynamic environments, where either the constraint or some properties of the search space might change during the optimization process. Neumann and Witt studied the behavior of the simplest evolutionary algorithm called (1+1) EA on the dynamic makespan scheduling problem, in which the adversary can change the processing time of a job[21]. They calculated the expected time of finding a good quality solution in different situations, where the algorithms either start from scratch and the dynamic changes happen frequently or start from a good quality solution and aim to fınd a new good quality solution after one change. The same analyses have also been considered on dynamic graph problems where a dynamic change might add or remove an edge to or from the graph [4,22]. More recently, Shi et al. theoretically investigated the re-optimization time of multi-objective evolutionary algorithms in the dynamic environment for general linear functions [29]. They proved the beneficial effect of populations in multiobjective approaches dealing with dynamic alterations. Their theoretical results have been extended to experimental studies in [27] by benchmarking dynamic knapsack problem. Roostapour et al. showed that the simple multi-objective algorithms that limit their population size by storing solutions based on their size perform significantly worse than the others.
Submodular functions form an important class of problems as many important optimization problems can be modeled by them. A function is submodular when adding an element to a solution set be less beneficial than adding the same element to any of its subsets. The area of submodular function optimization under given static constraints has been studied quite extensively in the literature. Nemhauser et al. considered maximizing nondecreasing submodular functions under cardinality constraint[19]. They proved that the greedy algorithm that iteratively adds a new element with a maximum marginal gain could maintain $(1-1/e)$-approximation. Following their results, Khuller et al. showed the unbounded approximation of the greedy approach in maximizing submodular functions under linear constraint [15]. Generalizing the greedy approach by a simple modification, however, they showed that it could guarantee optimal $(1/2)(1-1/e)$-approximation in particular submodular function for maximum coverage problem. The generalized greedy algorithm (Algorithm 1) compares the result of the iteratively found solution with the best feasible single-element solution and
2
|
[
[
0.223876953125,
0.1572265625,
0.2320556640625,
0.1654052734375,
0.2254638671875,
0.169189453125,
0.21728515625,
0.1610107421875
],
[
0.2442626953125,
0.1558837890625,
0.378173828125,
0.1558837890625,
0.378173828125,
0.17041015625,
0.2442626953125,
0.17041015625
],
[
0.2156982421875,
0.1875,
0.77783203125,
0.1875,
0.77783203125,
0.2008056640625,
0.2156982421875,
0.2008056640625
],
[
0.21728515625,
0.20263671875,
0.78125,
0.20263671875,
0.78125,
0.216552734375,
0.21728515625,
0.216552734375
],
[
0.21728515625,
0.2183837890625,
0.779296875,
0.2183837890625,
0.779296875,
0.231689453125,
0.21728515625,
0.231689453125
],
[
0.21728515625,
0.2342529296875,
0.78125,
0.2342529296875,
0.78125,
0.2474365234375,
0.21728515625,
0.2474365234375
],
[
0.21728515625,
0.248779296875,
0.779296875,
0.248779296875,
0.779296875,
0.261962890625,
0.21728515625,
0.261962890625
],
[
0.214111328125,
0.261962890625,
0.78125,
0.26318359375,
0.78125,
0.27783203125,
0.214111328125,
0.276611328125
],
[
0.2156982421875,
0.279052734375,
0.779296875,
0.279052734375,
0.779296875,
0.292236328125,
0.2156982421875,
0.292236328125
],
[
0.21728515625,
0.29345703125,
0.78173828125,
0.29345703125,
0.78173828125,
0.306884765625,
0.21728515625,
0.306884765625
],
[
0.2156982421875,
0.306884765625,
0.78173828125,
0.30810546875,
0.78173828125,
0.322509765625,
0.2156982421875,
0.3212890625
],
[
0.2156982421875,
0.323974609375,
0.42236328125,
0.323974609375,
0.42236328125,
0.337158203125,
0.2156982421875,
0.337158203125
],
[
0.240234375,
0.33837890625,
0.78173828125,
0.33837890625,
0.78173828125,
0.352783203125,
0.240234375,
0.352783203125
],
[
0.2156982421875,
0.354248046875,
0.779296875,
0.354248046875,
0.779296875,
0.367431640625,
0.2156982421875,
0.367431640625
],
[
0.21728515625,
0.37060546875,
0.77783203125,
0.37060546875,
0.77783203125,
0.380615234375,
0.21728515625,
0.380615234375
],
[
0.2156982421875,
0.3837890625,
0.78125,
0.3837890625,
0.78125,
0.397705078125,
0.2156982421875,
0.397705078125
],
[
0.2156982421875,
0.399658203125,
0.779296875,
0.399658203125,
0.779296875,
0.41357421875,
0.2156982421875,
0.41357421875
],
[
0.214111328125,
0.41357421875,
0.78125,
0.4140625,
0.78125,
0.4287109375,
0.214111328125,
0.427978515625
],
[
0.2156982421875,
0.429931640625,
0.7861328125,
0.429931640625,
0.7861328125,
0.443115234375,
0.2156982421875,
0.443115234375
],
[
0.21728515625,
0.4443359375,
0.779296875,
0.4443359375,
0.779296875,
0.458251953125,
0.21728515625,
0.458251953125
],
[
0.214111328125,
0.458251953125,
0.78125,
0.458984375,
0.78125,
0.473388671875,
0.214111328125,
0.472900390625
],
[
0.2156982421875,
0.474853515625,
0.78125,
0.474853515625,
0.78125,
0.488037109375,
0.2156982421875,
0.488037109375
],
[
0.214111328125,
0.4892578125,
0.78173828125,
0.488037109375,
0.78173828125,
0.50244140625,
0.214111328125,
0.50390625
],
[
0.21728515625,
0.5048828125,
0.78173828125,
0.5048828125,
0.78173828125,
0.51806640625,
0.21728515625,
0.51806640625
],
[
0.2156982421875,
0.52099609375,
0.78125,
0.51806640625,
0.78125,
0.5322265625,
0.2156982421875,
0.5341796875
],
[
0.21728515625,
0.53515625,
0.77783203125,
0.53515625,
0.77783203125,
0.548828125,
0.21728515625,
0.548828125
],
[
0.21728515625,
0.55029296875,
0.779296875,
0.55029296875,
0.779296875,
0.564453125,
0.21728515625,
0.564453125
],
[
0.21728515625,
0.56494140625,
0.78125,
0.56494140625,
0.78125,
0.5791015625,
0.21728515625,
0.5791015625
],
[
0.214111328125,
0.5791015625,
0.779296875,
0.5791015625,
0.779296875,
0.59326171875,
0.214111328125,
0.59326171875
],
[
0.21728515625,
0.59521484375,
0.78173828125,
0.59521484375,
0.78173828125,
0.609375,
0.21728515625,
0.609375
],
[
0.2156982421875,
0.611328125,
0.78125,
0.611328125,
0.78125,
0.625,
0.2156982421875,
0.625
],
[
0.21728515625,
0.62548828125,
0.36279296875,
0.62548828125,
0.36279296875,
0.6396484375,
0.21728515625,
0.6396484375
],
[
0.238525390625,
0.6396484375,
0.77783203125,
0.64013671875,
0.77783203125,
0.65478515625,
0.238525390625,
0.65380859375
],
[
0.2156982421875,
0.65478515625,
0.78125,
0.65478515625,
0.78125,
0.66943359375,
0.2156982421875,
0.66943359375
],
[
0.2156982421875,
0.67041015625,
0.78173828125,
0.67041015625,
0.78173828125,
0.6845703125,
0.2156982421875,
0.6845703125
],
[
0.2156982421875,
0.68603515625,
0.78173828125,
0.68603515625,
0.78173828125,
0.69970703125,
0.2156982421875,
0.69970703125
],
[
0.2156982421875,
0.70068359375,
0.779296875,
0.70068359375,
0.779296875,
0.7138671875,
0.2156982421875,
0.7138671875
],
[
0.2156982421875,
0.71630859375,
0.78125,
0.71630859375,
0.78125,
0.72998046875,
0.2156982421875,
0.72998046875
],
[
0.2156982421875,
0.73193359375,
0.78125,
0.73193359375,
0.78125,
0.74560546875,
0.2156982421875,
0.74560546875
],
[
0.21728515625,
0.7470703125,
0.78125,
0.7470703125,
0.78125,
0.76025390625,
0.21728515625,
0.76025390625
],
[
0.21728515625,
0.76220703125,
0.779296875,
0.76220703125,
0.779296875,
0.77587890625,
0.21728515625,
0.77587890625
],
[
0.2156982421875,
0.7763671875,
0.779296875,
0.7763671875,
0.779296875,
0.79052734375,
0.2156982421875,
0.79052734375
],
[
0.2156982421875,
0.79150390625,
0.77783203125,
0.79150390625,
0.77783203125,
0.80615234375,
0.2156982421875,
0.80615234375
],
[
0.2156982421875,
0.806640625,
0.78125,
0.806640625,
0.78125,
0.82080078125,
0.2156982421875,
0.82080078125
],
[
0.2156982421875,
0.8212890625,
0.77783203125,
0.8212890625,
0.77783203125,
0.8359375,
0.2156982421875,
0.8359375
],
[
0.2156982421875,
0.8369140625,
0.78173828125,
0.8369140625,
0.78173828125,
0.85107421875,
0.2156982421875,
0.85107421875
],
[
0.491943359375,
0.876953125,
0.5048828125,
0.876953125,
0.5048828125,
0.88818359375,
0.491943359375,
0.88818359375
]
] |
[
"",
"Introduction",
"Many combinatorial optimization problems consist of optimizing a given function un",
"der a given constraint. Constraints often limit the resources available to solve a given",
"problem and may change over time. For example, in the area of supply chain manage-",
"ment, the problem may be constrained by the number of vehicles available that may",
"change during the optimization process due to vehicle failures and vehicles being re-",
"paired or added as a new resource. As another example, consider the problem of using",
"a train to transport extracted materials form a number of mines to a port. In this exam-",
"ple, the number of available wagons could change during the path because of different",
"reasons. The question would be how to re-allocate the available wagons to transform",
"the high priority materials first.",
"Evolutionary algorithms as general-purpose problem solvers have been widely used",
"to tackle dynamic optimization problems theoretically and empirically [32]. Many the-",
"oreticalinvestigationsconsideredtheperformanceofevolutionaryalgorithmsinthedy-",
"namic environments, where either the constraint or some properties of the search space",
"might change during the optimization process: Neumann and Witt studied the behav-",
"ior of the simplest evolutionary algorithm called (1+1) EA on the dynamic makespan",
"scheduling problem, in which the adversary can change the processing time of a job [21].",
"They calculated the expected time of finding a good quality solution in different situa-",
"tions, where the algorithms either start from scratch and the dynamic changes happen",
"frequently or start from a good quality solution and aim to find a new good quality",
" solution after one change. The same analyses have also been considered on dynamic",
"graph problems where a dynamic change might add or remove an edge to or from the",
"graph [4, 22]. More recently, Shi et al. theoretically investigated the re-optimization",
"time of multi-objective evolutionary algorithms in the dynamic environment for gen-",
"eral linear functions [29]. They proved the beneficial effect of populations in multi-",
"objective approaches dealing with dynamic alterations. Their theoretical results have",
"been extended to experimental studies in [27] by benchmarking dynamic knapsack",
"problem. Roostapour et al. showed that the simple multi-objective algorithms that",
"limit their population size by storing solutions based on their size perform significantly",
"worse than the others.",
"Submodular functions form an important class of problems as many important op-",
"timization problems can be modeled by them. A function is submodular when adding",
"an element to a solution set be less beneficial than adding the same element to any of its",
"subsets. The area of submodular function optimization under given static constraints",
"has been studied quite extensively in the literature. Nemhauser et al. considered maxi-",
"mizing nondecreasing submodular functions under cardinality constraint [19]. They",
"proved that the greedy algorithm that iteratively adds a new element with a maxi-",
"mum marginal gain could maintain (1 - 1/e)-approximation. Following their results,",
"Khuller et al. showed the unbounded approximation of the greedy approach in max-",
"imizing submodular functions under linear constraint [15]. Generalizing the greedy",
"approach by a simple modification, however, they showed that it could guarantee op.",
"timal (1/2)(1 -1/e)-approximation in particular submodular function for maximum",
"coverage problem. The generalized greedy algorithm (Algorithm 1) compares the re",
"sult of the iteratively found solution with the best feasible single-element solution and",
"2"
] |
{
"arxiv_id": "1811.07806",
"language": "en",
"page_idx": 1
}
|
||
1811.07806-en-2
|
1811.07806-en-2.jpg
|
returns the maximum. Krause and Guestrin extended $(1/2)(1-1/e)$ approximation ratio to maximizing submodular functions under linear constraint [16], which had been improved to $(1-1/\sqrt{e})$ by Lin and Bilmes [18].
In the case of monotone submodular functions, greedy algorithms are often able to achieve the best possible worst case approximation guarantee (unless P=NP). Motivated by many real-world applications, the performance of greedy algorithms in general cost functions have also been considered later. Iyer and Bilmes studied cases that the cost function is monotone and submodular [13], while Zhang and Vorobeychik investigated problems with only monotone cost functions [33] as a more general case.
Recently, Pareto optimization approaches have been investigated for a wide range of subset selection problems [11,26,25,24,23]. It has been shown in [24] that an algorithm called POMC is able to achieve a $\phi=(\alpha_f/2)(1-\frac1{e^{\alpha}f})$-approximation where $\alpha_f$ measures the closeness of the considered function $f$ to submodularity. The approximation matches the worst-case performance ratio of the generalized greedy algorithm [33].
In this paper, we study monotone functions with a dynamic constraint where the constraint bound $B$ changes over time. Such constraint changes reflect real-world scenarios where resources vary during the process. We show that greedy algorithms have difficulties in adapting their current solutions after changes have happened. In particular, we show that there are simple dynamic versions of the classical knapsack problem where adding elements in a greedy fashion when the constraint bound increases over time can lead to an arbitrary bad performance ratio. For the case where constraint bounds decrease over time, we introduce a submodular graph covering problem and show that the considered adaptive generalized greedy algorithm may encounter an arbitrarily bad performance ratio on this problem
Investigating POMC, we theoretically prove that this algorithm obtains for each constraint bound $b\in[0,B]$, a $\phi=(\alpha_f/2)(1-\frac1{e^{\alpha_f}})$-approximation efficiently. Furthermore, when relaxing the bound $B$ to $B^*>B,\phi$-approximations for all values of $b\in[0,B^*]$ are obtained efficiently. The theoretical results for the performance of adaptive greedy algorithm and POMC, in addition to an initial experimental investigation on the quality of solutions found by these algorithms have been published in the proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence (AAAAI $\bar{2}019)[28].$
We now describe how this article extends its conference version. We significantly expand our experiments by considering two real-world problems under two different constraint types and thirty dynamic benchmarks for each of them. We create a baseline by performing our best algorithms for one million generations and statistically compare the results based on offline errors. Moreover, we compare the performance of our algorithms with two additional algorithmic approaches, namely EAMC[3] and NSGA-II[7]. EAMC is the most recent in the literature that guarantees a polynomial expected time and NSGA-II is a well-known evolutionary multi-objective algorithm used in many applications.
In the first part of our extension, by benchmarking the generalized greedy algorithm, its adaptive version and POMC on the influence maximization problem in social networks over sequences of dynamic changes, we show that POMC obtains superior results to the generalized greedy and adaptive generalized greedy algorithms, specif-
3
|
[
[
0.2156982421875,
0.1590576171875,
0.78125,
0.1590576171875,
0.78125,
0.173583984375,
0.2156982421875,
0.173583984375
],
[
0.21728515625,
0.1749267578125,
0.78125,
0.1749267578125,
0.78125,
0.1881103515625,
0.21728515625,
0.1881103515625
],
[
0.214111328125,
0.1881103515625,
0.5390625,
0.1881103515625,
0.5390625,
0.2020263671875,
0.214111328125,
0.2020263671875
],
[
0.240234375,
0.2052001953125,
0.78125,
0.2052001953125,
0.78125,
0.2183837890625,
0.240234375,
0.2183837890625
],
[
0.21728515625,
0.2215576171875,
0.77783203125,
0.2215576171875,
0.77783203125,
0.231689453125,
0.21728515625,
0.231689453125
],
[
0.2156982421875,
0.23486328125,
0.77685546875,
0.23486328125,
0.77685546875,
0.248779296875,
0.2156982421875,
0.248779296875
],
[
0.21728515625,
0.250732421875,
0.78125,
0.250732421875,
0.78125,
0.264404296875,
0.21728515625,
0.264404296875
],
[
0.2156982421875,
0.26513671875,
0.78125,
0.26513671875,
0.78125,
0.279052734375,
0.2156982421875,
0.279052734375
],
[
0.2156982421875,
0.27978515625,
0.77685546875,
0.27978515625,
0.77685546875,
0.294189453125,
0.2156982421875,
0.294189453125
],
[
0.2418212890625,
0.29541015625,
0.78125,
0.29541015625,
0.78125,
0.309326171875,
0.2418212890625,
0.309326171875
],
[
0.2156982421875,
0.31005859375,
0.779296875,
0.31005859375,
0.779296875,
0.323974609375,
0.2156982421875,
0.323974609375
],
[
0.21728515625,
0.32568359375,
0.78125,
0.32568359375,
0.78125,
0.34033203125,
0.21728515625,
0.34033203125
],
[
0.21728515625,
0.34033203125,
0.78125,
0.34033203125,
0.78125,
0.354248046875,
0.21728515625,
0.354248046875
],
[
0.2156982421875,
0.35595703125,
0.779296875,
0.354736328125,
0.779296875,
0.369384765625,
0.2156982421875,
0.37060546875
],
[
0.2156982421875,
0.37060546875,
0.306396484375,
0.37060546875,
0.306396484375,
0.385009765625,
0.2156982421875,
0.385009765625
],
[
0.2418212890625,
0.386474609375,
0.78125,
0.386474609375,
0.78125,
0.399658203125,
0.2418212890625,
0.399658203125
],
[
0.214111328125,
0.399658203125,
0.779296875,
0.40087890625,
0.779296875,
0.415283203125,
0.214111328125,
0.4140625
],
[
0.21728515625,
0.416748046875,
0.78125,
0.416748046875,
0.78125,
0.429931640625,
0.21728515625,
0.429931640625
],
[
0.2156982421875,
0.429931640625,
0.77783203125,
0.43115234375,
0.77783203125,
0.44580078125,
0.2156982421875,
0.4443359375
],
[
0.2156982421875,
0.4462890625,
0.78125,
0.4462890625,
0.78125,
0.4609375,
0.2156982421875,
0.4609375
],
[
0.2156982421875,
0.462158203125,
0.78125,
0.462158203125,
0.78125,
0.47607421875,
0.2156982421875,
0.47607421875
],
[
0.21728515625,
0.4765625,
0.78125,
0.4765625,
0.78125,
0.490478515625,
0.21728515625,
0.490478515625
],
[
0.2156982421875,
0.492431640625,
0.78125,
0.492431640625,
0.78125,
0.505859375,
0.2156982421875,
0.505859375
],
[
0.21728515625,
0.5068359375,
0.779296875,
0.5068359375,
0.779296875,
0.52099609375,
0.21728515625,
0.52099609375
],
[
0.214111328125,
0.521484375,
0.529296875,
0.521484375,
0.529296875,
0.53515625,
0.214111328125,
0.53515625
],
[
0.240234375,
0.5361328125,
0.78125,
0.53515625,
0.78125,
0.5498046875,
0.240234375,
0.55029296875
],
[
0.2156982421875,
0.55029296875,
0.779296875,
0.5517578125,
0.779296875,
0.56640625,
0.2156982421875,
0.56494140625
],
[
0.21728515625,
0.5673828125,
0.78125,
0.5673828125,
0.78125,
0.58203125,
0.21728515625,
0.58203125
],
[
0.2181396484375,
0.58203125,
0.78173828125,
0.58203125,
0.78173828125,
0.59521484375,
0.2181396484375,
0.59521484375
],
[
0.21728515625,
0.5966796875,
0.779296875,
0.5966796875,
0.779296875,
0.611328125,
0.21728515625,
0.611328125
],
[
0.2156982421875,
0.6123046875,
0.78125,
0.6123046875,
0.78125,
0.62548828125,
0.2156982421875,
0.62548828125
],
[
0.21728515625,
0.62744140625,
0.78173828125,
0.62744140625,
0.78173828125,
0.6416015625,
0.21728515625,
0.6416015625
],
[
0.2156982421875,
0.642578125,
0.291748046875,
0.642578125,
0.291748046875,
0.6572265625,
0.2156982421875,
0.6572265625
],
[
0.2418212890625,
0.65771484375,
0.779296875,
0.65771484375,
0.779296875,
0.671875,
0.2418212890625,
0.671875
],
[
0.2156982421875,
0.67236328125,
0.78173828125,
0.67236328125,
0.78173828125,
0.68603515625,
0.2156982421875,
0.68603515625
],
[
0.2156982421875,
0.6875,
0.779296875,
0.68603515625,
0.779296875,
0.70068359375,
0.2156982421875,
0.7021484375
],
[
0.21728515625,
0.70263671875,
0.779296875,
0.70263671875,
0.779296875,
0.71630859375,
0.21728515625,
0.71630859375
],
[
0.2156982421875,
0.71826171875,
0.78125,
0.71826171875,
0.78125,
0.73193359375,
0.2156982421875,
0.73193359375
],
[
0.2156982421875,
0.73193359375,
0.78125,
0.73291015625,
0.78125,
0.74755859375,
0.2156982421875,
0.7470703125
],
[
0.2156982421875,
0.7470703125,
0.78125,
0.74755859375,
0.78125,
0.76220703125,
0.2156982421875,
0.76123046875
],
[
0.21728515625,
0.76318359375,
0.78125,
0.76318359375,
0.78125,
0.77783203125,
0.21728515625,
0.77783203125
],
[
0.2156982421875,
0.77880859375,
0.393798828125,
0.77880859375,
0.393798828125,
0.79345703125,
0.2156982421875,
0.79345703125
],
[
0.2418212890625,
0.79345703125,
0.77783203125,
0.79345703125,
0.77783203125,
0.806640625,
0.2418212890625,
0.806640625
],
[
0.2156982421875,
0.8095703125,
0.78125,
0.8095703125,
0.78125,
0.82275390625,
0.2156982421875,
0.82275390625
],
[
0.2156982421875,
0.82470703125,
0.78125,
0.82470703125,
0.78125,
0.83837890625,
0.2156982421875,
0.83837890625
],
[
0.2156982421875,
0.83837890625,
0.78125,
0.83837890625,
0.78125,
0.85302734375,
0.2156982421875,
0.85302734375
],
[
0.491943359375,
0.876953125,
0.5048828125,
0.876953125,
0.5048828125,
0.88818359375,
0.491943359375,
0.88818359375
]
] |
[
"returns the maximum. Krause and Guestrin extended (1/2)(1 - 1/e) approximation",
"ratio to maximizing submodular functions under linear constraint [16], which had been",
"improved to (1 -1/ Ve) by Lin and Bilmes [18]",
"In the case of monotone submodular functions, greedy algorithms are often able",
"to achievethebestpossibleworstcaseapproximationguarantee(unlessP=NP).Moti-",
"vated by many real-world applications, the performance of greedy algorithms in gen",
"eral cost functions have also been considered later. lyer and Bilmes studied cases that",
"the cost function is monotone and submodular [13], while Zhang and Vorobeychik",
"investigated problems with only monotone cost functions [33] as a more general case.",
"Recently, Pareto optimization approaches have been investigated for a wide range",
"of subset selection problems [11, 26, 25, 24, 23]. It has been shown in [24] that an",
"algorithm called POMC is able to achieve a Φ = (αf/2)(1 - )-approximation",
"where αf measures the closeness of the considered function f to submodularity. The",
"approximation matches the worst-case performance ratio of the generalized greedy al-",
"gorithm [33].",
"In this paper, we study monotone functions with a dynamic constraint where the",
"constraint bound B changes over time. Such constraint changes reflect real-world sce",
"narios where resources vary during the process. We show that greedy algorithms have",
"difficulties in adapting their current solutions after changes have happened. In particu",
"lar, we show that there are simple dynamic versions of the classical knapsack problem",
"where adding elements in a greedy fashion when the constraint bound increases over",
"time can lead to an arbitrary bad performance ratio. For the case where constraint",
"bounds decrease overtime, we introduce a submodulargraph covering problem and",
"show that the considered adaptive generalized greedy algorithm may encounter an ar-",
"bitrarily bad performance ratio on this problem.",
"Investigating POMC, we theoretically prove that this algorithm obtains for each",
"constraint bound b e [0, B],a @ = (αf/2)(1-)-approximation efficiently. Fur-",
"thermore, when relaxing the bound B to B* > B, Φ-approximations for all values",
"of b E [O, B*l are obtained efficiently. The theoretical results for the performance of",
"adaptive greedy algorithm and POMC, in addition to an initial experimental investiga-",
"tion on the quality of solutions found by these algorithms have been published in the",
"proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence (AAAI",
"2019) [28].",
"We now describe how this article extends its conference version. We significantly",
"expand our experiments by considering two real-world problems under two different",
"constraint types and thirty dynamic benchmarks for each of them. We create a base-",
"line by performing our best algorithms for one million generations and statistically",
"compare the results based on offlineerrors.Moreover,we compare theperformance",
"of our algorithms with two additional algorithmic approaches, namely EAMC [3] and",
"NSGA-II [7]. EAMC is the most recent in the literature that guarantees a polynomial",
"expected time and NSGA-II is a well-known evolutionary multi-objective algorithm",
"used in many applications.",
"In the first part of our extension, by benchmarking the generalized greedy algo-",
"rithm, its adaptive version and POMC on the influencemaximization problem in social",
"networks over sequences of dynamic changes, we show that POMC obtains superior",
"results to the generalized greedy and adaptive generalized greedy algorithms, specif-",
"3"
] |
{
"arxiv_id": "1811.07806",
"language": "en",
"page_idx": 2
}
|
||
1811.07806-en-3
|
1811.07806-en-3.jpg
|
ically when the cost function depends on the structure of the graph. We consider the experimental results in four intervals to examine whether the algorithms behave differently during the optimization process. Here, POMC shows an even more prominent advantage as more dynamic changes are carried out because it is able to improve the quality of its solutions and cater for changes that occur in the future.
To compare these algorithms with EAMC and NSGA-II as the second part of the extension, we study the maximum coverage problem using two graph benchmarks with different sizes and two cost functions. For the “random”cost, we assign a random value to each node of the graph as their cost value. For the other cost function, called “outdegree”, the cost is calculated based on the outdegree of each node. Presenting some explanatory examples, we describe the major problems that hold EAMC back from tracking optimal solutions in dynamic environments. Our results also show that EAMC, which limits its population size to guarantee the polynomial expected time, is not capable of dealing with dynamic changes and performs worse than the other algorithms. We also consider NSGA-II with two population sizes NSGA-II20 population size of 20 and NSGA-II100 with population size of 100. This algorithm that is known to be a high performing evolutionary multi-objective algorithm, beats POMC when the constraint costs are random. However, NSGA-II20, because of its limited population size and also the effect of the crowding-distance factor in the selection procedure which forces the population to be distributed along the Pareto front, is inferior to POMC for the outdegree cost function. NSGA-II100, on the other hand, is more powerful when changes happen with medium or low frequencies. The effect of larger population size is seen in the experiments which shows NSGA-II20 tracks the optimal solution in high frequent dynamic environment better than NSGA-II100.
The paper is structured as follows. In the next section, we formulate the dynamic problems. In Section 3, we firstly introduce the algorithms that are studied theoretically and then show that the adaptive generalized greedy algorithm may encounter an arbitrary bad performance ratio even when starting with an optimal solution for the initial budget. In contrast, we show that POMC can maintain $\phi$-approximation efficiently. Section 4 presents our experimental setting, the results, and the empirical analysis. We discuss our experimental investigations for influence maximization in social networks in Section 4.2. Then we extend our investigations by considering NSGA-II20, NSGA-II100, and EAMC facing the benchmarks for maximum coverage problem in Section 4.5, followed by experimental demonstration on how the population size of POMC algorithm changes during one dynamic benchmark. Finally, we finish with some concluding remarks.
# 2 $\textbf{Problem Formulation}$
In this paper we consider optimization problems in which the cost function and objective functions are monotone and quantified according to their closeness to submodular- $ity.$
Different defnitions are given for submodularity [20] and we use the following one in this paper. For a given set $V=\{v_1,\cdots,v_n\}$, a function $f:2^V\to\mathbb{R}$ is submodular
4
|
[
[
0.21728515625,
0.160400390625,
0.78125,
0.160400390625,
0.78125,
0.173583984375,
0.21728515625,
0.173583984375
],
[
0.2156982421875,
0.1749267578125,
0.779296875,
0.173583984375,
0.779296875,
0.1881103515625,
0.2156982421875,
0.189453125
],
[
0.21728515625,
0.189453125,
0.78173828125,
0.189453125,
0.78173828125,
0.203857421875,
0.21728515625,
0.203857421875
],
[
0.2156982421875,
0.2052001953125,
0.78173828125,
0.2052001953125,
0.78173828125,
0.2183837890625,
0.2156982421875,
0.2183837890625
],
[
0.21728515625,
0.2203369140625,
0.6640625,
0.2203369140625,
0.6640625,
0.2342529296875,
0.21728515625,
0.2342529296875
],
[
0.2418212890625,
0.23486328125,
0.78125,
0.23486328125,
0.78125,
0.248779296875,
0.2418212890625,
0.248779296875
],
[
0.21728515625,
0.250732421875,
0.78125,
0.250732421875,
0.78125,
0.264404296875,
0.21728515625,
0.264404296875
],
[
0.21728515625,
0.26513671875,
0.78125,
0.26513671875,
0.78125,
0.279052734375,
0.21728515625,
0.279052734375
],
[
0.21728515625,
0.281005859375,
0.78173828125,
0.281005859375,
0.78173828125,
0.294189453125,
0.21728515625,
0.294189453125
],
[
0.2156982421875,
0.294189453125,
0.78125,
0.29541015625,
0.78125,
0.31005859375,
0.2156982421875,
0.309326171875
],
[
0.2156982421875,
0.311279296875,
0.78125,
0.311279296875,
0.78125,
0.324462890625,
0.2156982421875,
0.324462890625
],
[
0.2156982421875,
0.32568359375,
0.78173828125,
0.324462890625,
0.78173828125,
0.339111328125,
0.2156982421875,
0.34033203125
],
[
0.214111328125,
0.339111328125,
0.78173828125,
0.34033203125,
0.78173828125,
0.354736328125,
0.214111328125,
0.354248046875
],
[
0.2156982421875,
0.35595703125,
0.779296875,
0.35595703125,
0.779296875,
0.369384765625,
0.2156982421875,
0.369384765625
],
[
0.2156982421875,
0.371826171875,
0.78173828125,
0.371826171875,
0.78173828125,
0.385009765625,
0.2156982421875,
0.385009765625
],
[
0.21728515625,
0.386474609375,
0.78125,
0.386474609375,
0.78125,
0.399658203125,
0.21728515625,
0.399658203125
],
[
0.214111328125,
0.40087890625,
0.78125,
0.399658203125,
0.78125,
0.4140625,
0.214111328125,
0.415283203125
],
[
0.21728515625,
0.416748046875,
0.78125,
0.416748046875,
0.78125,
0.429931640625,
0.21728515625,
0.429931640625
],
[
0.21728515625,
0.431884765625,
0.78173828125,
0.431884765625,
0.78173828125,
0.44580078125,
0.21728515625,
0.44580078125
],
[
0.2156982421875,
0.4462890625,
0.78125,
0.4462890625,
0.78125,
0.460205078125,
0.2156982421875,
0.460205078125
],
[
0.21728515625,
0.4609375,
0.78125,
0.4609375,
0.78125,
0.47607421875,
0.21728515625,
0.47607421875
],
[
0.21728515625,
0.477783203125,
0.78125,
0.477783203125,
0.78125,
0.4912109375,
0.21728515625,
0.4912109375
],
[
0.21728515625,
0.492431640625,
0.78125,
0.492431640625,
0.78125,
0.505859375,
0.21728515625,
0.505859375
],
[
0.214111328125,
0.5068359375,
0.583984375,
0.5068359375,
0.583984375,
0.52099609375,
0.214111328125,
0.52099609375
],
[
0.2418212890625,
0.521484375,
0.78125,
0.521484375,
0.78125,
0.5361328125,
0.2418212890625,
0.5361328125
],
[
0.2156982421875,
0.537109375,
0.779296875,
0.537109375,
0.779296875,
0.55029296875,
0.2156982421875,
0.55029296875
],
[
0.214111328125,
0.55029296875,
0.78125,
0.5517578125,
0.78125,
0.56640625,
0.214111328125,
0.56494140625
],
[
0.21728515625,
0.5673828125,
0.779296875,
0.5673828125,
0.779296875,
0.58056640625,
0.21728515625,
0.58056640625
],
[
0.21728515625,
0.58349609375,
0.779296875,
0.58349609375,
0.779296875,
0.5966796875,
0.21728515625,
0.5966796875
],
[
0.2156982421875,
0.5966796875,
0.779296875,
0.59765625,
0.779296875,
0.6123046875,
0.2156982421875,
0.611328125
],
[
0.21728515625,
0.61279296875,
0.78125,
0.61279296875,
0.78125,
0.626953125,
0.21728515625,
0.626953125
],
[
0.21728515625,
0.62744140625,
0.78125,
0.62744140625,
0.78125,
0.6416015625,
0.21728515625,
0.6416015625
],
[
0.2156982421875,
0.6416015625,
0.78125,
0.642578125,
0.78125,
0.6572265625,
0.2156982421875,
0.65576171875
],
[
0.21728515625,
0.65771484375,
0.78173828125,
0.65771484375,
0.78173828125,
0.67236328125,
0.21728515625,
0.67236328125
],
[
0.21728515625,
0.67236328125,
0.78125,
0.67236328125,
0.78125,
0.6875,
0.21728515625,
0.6875
],
[
0.2156982421875,
0.68798828125,
0.390625,
0.68798828125,
0.390625,
0.7021484375,
0.2156982421875,
0.7021484375
],
[
0.2132568359375,
0.724609375,
0.47216796875,
0.72607421875,
0.47216796875,
0.74365234375,
0.2132568359375,
0.7421875
],
[
0.2156982421875,
0.75830078125,
0.77783203125,
0.75830078125,
0.77783203125,
0.77294921875,
0.2156982421875,
0.77294921875
],
[
0.2156982421875,
0.7734375,
0.77783203125,
0.7734375,
0.77783203125,
0.787109375,
0.2156982421875,
0.787109375
],
[
0.2156982421875,
0.79052734375,
0.238525390625,
0.79052734375,
0.238525390625,
0.8017578125,
0.2156982421875,
0.8017578125
],
[
0.2418212890625,
0.8037109375,
0.78125,
0.8037109375,
0.78125,
0.8173828125,
0.2418212890625,
0.8173828125
],
[
0.214111328125,
0.818359375,
0.78125,
0.8173828125,
0.78125,
0.83203125,
0.214111328125,
0.83349609375
],
[
0.49267578125,
0.876953125,
0.5048828125,
0.876953125,
0.5048828125,
0.8896484375,
0.49267578125,
0.8896484375
]
] |
[
"ically when the cost function depends on the structure of the graph. We consider the",
"experimental results in four intervals to examine whether the algorithms behave differ-",
"ently during the optimization process. Here, POMC shows an even more prominent",
"advantage as more dynamic changes are carried out because it is able to improve the",
"quality of its solutions and cater for changes that occur in the future.",
"To compare these algorithms with EAMC and NSGA-II as the second part of the",
"extension, we study the maximum coverage problem using two graph benchmarks with",
"different sizes and two cost functions. For the \"random\" cost, we assign a random",
"value to each node of the graph as their cost value. For the other cost function, called",
"“outdegree\", the cost is calculated based on the outdegree of each node. Presenting",
"some explanatory examples, we describe the major problems that hold EAMC back",
"from tracking optimal solutions in dynamic environments. Our results also show that",
"EAMC, which limits its population size to guarantee the polynomial expected time, is",
"not capable of dealing with dynamic changes and performs worse than the other algo-",
"rithms. We also consider NSGA-II with two population sizes NSGA-II2O population",
"size of 20 and NSGA-I1100 with population size of 100. This algorithm that is known",
"to be a high performing evolutionary multi-objective algorithm, beats POMC when the",
"constraint costs are random. However, NSGA-II20, because of its limited population",
"size and also the effect of the crowding-distance factor in the selection procedure which",
"forces the population to be distributed along the Pareto front, is inferior to POMC for",
"the outdegree cost function. NSGA-II10o, on the other hand, is more powerful when",
"changes happen with medium or low frequencies. The effect of larger population size",
"is seen in the experiments which shows NSGA-I120 tracks the optimal solution in high",
"frequent dynamic environment better than NSGA-II100.",
"The paper is structured as follows. In the next section, we formulate the dynamic",
"problems. In Section 3, we firstly introduce the algorithms that are studied theoreti-",
"cally and then show that the adaptive generalized greedy algorithm may encounter an",
"arbitrary bad performance ratio even when starting with an optimal solution for the ini-",
"tial budget.In contrast, we show that POMC can maintain -approximation efficiently.",
"Section 4 presents our experimental setting, the results, and the empirical analysis.",
"We discuss our experimental investigations for influence maximization in social net-",
"works in Section 4.2. Then we extend our investigations by considering NSGA-I120,",
"NSGA-II1oo, and EAMC facing the benchmarks for maximum coverage problem in",
"Section 4.5, followed by experimental demonstration on how the population size of",
"POMC algorithm changes during one dynamic benchmark. Finally, we finish with",
"some concludingremarks.",
"2Problem Formulation",
"In this paper we consider optimization problems in which the cost function and objec-",
"tive functions are monotone and quantified according to their closeness to submodular-",
"ity.",
"Different definitions are given for submodularity [20] and we use the following one",
"in this paper. For a given set V = [ui,... , Un], a function f : 2V → R is submodular",
"4"
] |
{
"arxiv_id": "1811.07806",
"language": "en",
"page_idx": 3
}
|
||
1811.07806-en-4
|
1811.07806-en-4.jpg
|
if for any $X\subseteq Y\subseteq V$ and $v\notin Y$
(1)
$$f(X\cup\{v\})-f(X)\geq f(Y\cup\{v\})-f(Y).$$
In addition, we consider how much a function is close to being submodular, mea-
sured by the submodularity ratio [33]. The function $f$ is $\alpha_f$-submodular where
(2)
$$\alpha_f=\min_{X\subseteq Y,v\notin Y}\frac{f(X\cup\{v\})-f(X)}{f(Y\cup\{v\})-f(Y)}.$$
This defnition is equivalent to the Equation 1 when $\alpha_f=1$, i.e., we have submodularity in this case. Another notion which is used in our analysis is the curvature of function $f.$ The curvature measures the deviation from linearity and reflects the effect of marginal contribution according to the function $f$ [5,30]. For a monotone submodular function $f:2^V\to\mathbb{R}^+$,
$$\kappa_f=1-\min_{v\in V}\frac{f(V)-f(V\setminus\{v\})}{f(v)}$$
is defined as the total curvature of $f.$
In many applications the function to be optimized, $f$,comes with a cost function $c$ which is subject to a given cost constraint $B.$ Often the cost function $c$ cannot be evaluated precisely. Hence, function $\hat{c}$ is used which is an $\alpha_c$-submodular $\psi$-approximation of cost $c[33].$ To be more precise, according to submodularity definition (Equation 2), $\hat{c}$ is chosen such that $\alpha_{c}=\alpha_{\hat{c}}$ and $c(X)\leq\hat{c}(X)\leq\psi c(X)$, where $n=|V|.$ Moreover, according to the submodularity of $f$,the aim is to find a good approximation instead of finding an optimal solution.
Consider the static version of an optimization problem defined in [24].
Definition 1 (The Static Problem). Given a monotone objective function $f:2^V\to$ $\mathbb{R} ^+ , a\textit{ monotone cost function c}: 2^V\to \mathbb{R} ^+$ and a budget $B$,the goal is to compute a solution X such that
$$X=\underset{Y\subseteq V}{\arg\max}f(Y)\:s.t.\:c(Y)\leq B.$$
Following the investigation of the static case in [24], we are interested in a $\phi$-approximation
where $\phi=(\bar{\alpha}_{f}/2)(1-\frac{1}{e^{\alpha}f})$ depends on the submodularity ratio.
Zhang and Vorobeychik considered the performance of the generalized greedy algorithm [33], given in Algorithm 1, according to the approximated cost function $\hat{c}.$ Starting from the empty set, the algorithm always adds the element with the largest objective to cost ratio that does not violate the given constraint bound $B.$
Let$K_{c}= \operatorname* { max} \{ | X| : c( X) \leq B\} .$ The optimal solution $\tilde{X}_{B}$ in these investigations ${\mathrm{is~defined~as~}}X_{B}=\arg\max\{f(X)\mid c(X)\leq\alpha_{c}\frac{B(1+\alpha_{c}(\mathbf{\Lambda}_{c}-1)(1-\kappa_{c})))}{\psi K_{c}}\}$ where $\alpha_{c}$ is the submodularity ratio of $c.$ This formulation gives the value of an optimal solution for a slightly smaller budget constraint. The goal is to obtain a good approximation of $f(\tilde{X}_B)$ in this case.
It has been shown in [33] that the generalized greedy algorithm, which adds the item with the highest marginal contribution to the current solution in each step, achieves
5
|
[
[
0.2156982421875,
0.1590576171875,
0.439453125,
0.1578369140625,
0.439453125,
0.17236328125,
0.2156982421875,
0.173583984375
],
[
0.3505859375,
0.186279296875,
0.646484375,
0.186279296875,
0.646484375,
0.199462890625,
0.3505859375,
0.199462890625
],
[
0.7587890625,
0.1875,
0.78173828125,
0.1875,
0.78173828125,
0.2008056640625,
0.7587890625,
0.2008056640625
],
[
0.240234375,
0.210205078125,
0.779296875,
0.2108154296875,
0.779296875,
0.2259521484375,
0.240234375,
0.2247314453125
],
[
0.21728515625,
0.2279052734375,
0.73193359375,
0.2279052734375,
0.73193359375,
0.2412109375,
0.21728515625,
0.2412109375
],
[
0.48046875,
0.25244140625,
0.61669921875,
0.25244140625,
0.61669921875,
0.266357421875,
0.48046875,
0.266357421875
],
[
0.373291015625,
0.261474609375,
0.419189453125,
0.261474609375,
0.419189453125,
0.27587890625,
0.373291015625,
0.27587890625
],
[
0.419189453125,
0.261962890625,
0.461669921875,
0.261962890625,
0.461669921875,
0.2734375,
0.419189453125,
0.2734375
],
[
0.7587890625,
0.260009765625,
0.78173828125,
0.260009765625,
0.78173828125,
0.274658203125,
0.7587890625,
0.274658203125
],
[
0.41259765625,
0.26953125,
0.61767578125,
0.26953125,
0.61767578125,
0.2841796875,
0.41259765625,
0.2841796875
],
[
0.21728515625,
0.29345703125,
0.77783203125,
0.29345703125,
0.77783203125,
0.306884765625,
0.21728515625,
0.306884765625
],
[
0.2197265625,
0.31005859375,
0.779296875,
0.31005859375,
0.779296875,
0.32080078125,
0.2197265625,
0.32080078125
],
[
0.2156982421875,
0.323974609375,
0.78125,
0.323974609375,
0.78125,
0.337158203125,
0.2156982421875,
0.337158203125
],
[
0.21728515625,
0.33837890625,
0.779296875,
0.33837890625,
0.779296875,
0.352783203125,
0.21728515625,
0.352783203125
],
[
0.2156982421875,
0.354248046875,
0.40283203125,
0.3515625,
0.40283203125,
0.3662109375,
0.2156982421875,
0.36865234375
],
[
0.48193359375,
0.37744140625,
0.61376953125,
0.37744140625,
0.61376953125,
0.392578125,
0.48193359375,
0.392578125
],
[
0.378173828125,
0.386962890625,
0.47802734375,
0.3837890625,
0.47802734375,
0.3984375,
0.378173828125,
0.401611328125
],
[
0.453369140625,
0.3984375,
0.4755859375,
0.3984375,
0.4755859375,
0.406005859375,
0.453369140625,
0.406005859375
],
[
0.52783203125,
0.395751953125,
0.564453125,
0.395751953125,
0.564453125,
0.410888671875,
0.52783203125,
0.410888671875
],
[
0.2156982421875,
0.41845703125,
0.455810546875,
0.41845703125,
0.455810546875,
0.431884765625,
0.2156982421875,
0.431884765625
],
[
0.2418212890625,
0.434326171875,
0.78173828125,
0.434326171875,
0.78173828125,
0.447509765625,
0.2418212890625,
0.447509765625
],
[
0.2181396484375,
0.449462890625,
0.77783203125,
0.449462890625,
0.77783203125,
0.46337890625,
0.2181396484375,
0.46337890625
],
[
0.2156982421875,
0.46337890625,
0.779296875,
0.464111328125,
0.779296875,
0.479248046875,
0.2156982421875,
0.477783203125
],
[
0.21728515625,
0.479736328125,
0.779296875,
0.479736328125,
0.779296875,
0.49365234375,
0.21728515625,
0.49365234375
],
[
0.214111328125,
0.494384765625,
0.779296875,
0.49560546875,
0.779296875,
0.5087890625,
0.214111328125,
0.50830078125
],
[
0.21728515625,
0.51025390625,
0.78173828125,
0.51025390625,
0.78173828125,
0.5234375,
0.21728515625,
0.5234375
],
[
0.21728515625,
0.52587890625,
0.396240234375,
0.52587890625,
0.396240234375,
0.5390625,
0.21728515625,
0.5390625
],
[
0.240234375,
0.53857421875,
0.701171875,
0.5390625,
0.701171875,
0.5537109375,
0.240234375,
0.55322265625
],
[
0.2156982421875,
0.5625,
0.78125,
0.56103515625,
0.78125,
0.57568359375,
0.2156982421875,
0.5771484375
],
[
0.2156982421875,
0.57763671875,
0.78173828125,
0.5791015625,
0.78173828125,
0.59326171875,
0.2156982421875,
0.59228515625
],
[
0.21728515625,
0.59521484375,
0.35302734375,
0.59521484375,
0.35302734375,
0.60595703125,
0.21728515625,
0.60595703125
],
[
0.38232421875,
0.619140625,
0.61376953125,
0.619140625,
0.61376953125,
0.6337890625,
0.38232421875,
0.6337890625
],
[
0.4345703125,
0.6337890625,
0.4599609375,
0.6337890625,
0.4599609375,
0.642578125,
0.4345703125,
0.642578125
],
[
0.2418212890625,
0.65380859375,
0.82666015625,
0.65380859375,
0.82666015625,
0.66845703125,
0.2418212890625,
0.66845703125
],
[
0.2156982421875,
0.66845703125,
0.646484375,
0.66943359375,
0.646484375,
0.6845703125,
0.2156982421875,
0.68310546875
],
[
0.2418212890625,
0.68505859375,
0.779296875,
0.68505859375,
0.779296875,
0.69873046875,
0.2418212890625,
0.69873046875
],
[
0.21728515625,
0.70068359375,
0.77783203125,
0.70068359375,
0.77783203125,
0.7138671875,
0.21728515625,
0.7138671875
],
[
0.21728515625,
0.71533203125,
0.78125,
0.71533203125,
0.78125,
0.72900390625,
0.21728515625,
0.72900390625
],
[
0.21728515625,
0.72998046875,
0.69287109375,
0.72998046875,
0.69287109375,
0.74365234375,
0.21728515625,
0.74365234375
],
[
0.2418212890625,
0.744140625,
0.78125,
0.744140625,
0.78125,
0.7587890625,
0.2418212890625,
0.7587890625
],
[
0.2115478515625,
0.7568359375,
0.7861328125,
0.7568359375,
0.7861328125,
0.78076171875,
0.2115478515625,
0.78076171875
],
[
0.2156982421875,
0.77783203125,
0.779296875,
0.77880859375,
0.779296875,
0.79345703125,
0.2156982421875,
0.79248046875
],
[
0.21728515625,
0.79443359375,
0.78173828125,
0.79443359375,
0.78173828125,
0.80810546875,
0.21728515625,
0.80810546875
],
[
0.21728515625,
0.8095703125,
0.343017578125,
0.8095703125,
0.343017578125,
0.82275390625,
0.21728515625,
0.82275390625
],
[
0.2418212890625,
0.82373046875,
0.779296875,
0.82373046875,
0.779296875,
0.83837890625,
0.2418212890625,
0.83837890625
],
[
0.214111328125,
0.8369140625,
0.78125,
0.83837890625,
0.78125,
0.85302734375,
0.214111328125,
0.8515625
],
[
0.490234375,
0.87548828125,
0.50634765625,
0.87548828125,
0.50634765625,
0.89013671875,
0.490234375,
0.89013671875
]
] |
[
"if for any XCYCV and u&Y",
"f(X U () - f(X) ≥ f(Y U (v)) - f(Y)",
"(1)",
"In addition, we consider how much a function is close to being submodular, mea-",
"sured by the submodularity ratio [33]. The function f is α f-submodular where",
"f(X U(u)) - f(X)",
"af=",
"min",
"(2)",
"XCY,uY f(Y U (u)) - f(Y)",
"This definition is equivalent to the Equation I when Qf =l, i.e., we have submod-",
"ularityinthis case.Anothernotion whichisusedinour analysis is the curvatureof",
"function f.The curvature measures the deviation from linearity and reflects the effect",
"of marginal contribution according to the function f [5, 30]. For a monotone submod-",
"ular function f : 2V → IR+,",
"f(V) - f(V/ (u))",
"hf=1-min",
"VEV",
"f(u)",
"is defined as the total curvature of f.",
"In many applications the function to be optimized, f, comes with a cost function c",
"which is subject to a given cost constraint B. Often the cost function c cannot be evalu-",
"ated precisely. Hence, function c is used which is an c-submodular b-approximation",
"of cost c [33]. To be more precise, according to submodularity definition (Equation 2),",
"c is chosen such that e = Qe and c(X) ≤ c(X) ≤ bc(X), where n = V/. Moreover,",
"according to the submodularity of f, the aim is to find a good approximation instead of",
"finding an optimal solution",
"Consider the static version of an optimization problem defined in [24].",
"Definition 1 (The Static Problem). Given a monotone objective function f : 2V -→",
"R+, a monotone cost function c : 2 -→ IR+ and a budget B, the goal is to compute a",
"solution X suchthat",
"X = arg max f(Y) s.t. c(Y) ≤ B.",
"YCV",
"Following the investigation of the static case in [24], we are interested in a -approximation",
"whereΦ= (αf/2)(1-)depends on the submodularity ratio.",
"Zhang and Vorobeychik considered the performance of the generalized greedy al-",
"gorithm [33], given in Algorithm 1, according to the approximated cost function c",
"Starting from the empty set, the algorithm always adds the element with the largest",
"objective to cost ratio that does not violate the given constraint bound B.",
"Let K = max(X| : c(X) ≤ B). The optimal solution XB in these investigations",
"is defined as XB = arg max(f(X) I c(X) ≤ α。B(1+a (Kk1)(-re) where ae is",
"the submodularity ratio of c. This formulation gives the value of an optimal solution",
"for a slightly smaller budget constraint. The goal is to obtain a good approximation of",
"f(XB) in this case",
"It has been shown in [33] that the generalized greedy algorithm, which adds the",
"item with the highest marginal contribution to the current solution in each step, achieves",
"5"
] |
{
"arxiv_id": "1811.07806",
"language": "en",
"page_idx": 4
}
|
||
2306.03571-en-1
|
2306.03571-en-1.jpg
|
KDD'23, August 6-10, 2023, Long Beach, CA, USA
For example, the work of Haddadan et al.[24] considers maximizing the gain function $\Delta_g(F)=g(\emptyset)-g(F)$ in the context of bounded-length random walks in directed graphs. Here, F is a set of at most $k$ new shortcut edges that will be added to an input graph, and the function $g$ is the average expected hitting time for a random walk starting in one group to hit the other group (see Definition 1). They prove that $\Delta_g$ is a non-negative monotone submodular set function, and thus, the well-known greedy algorithm of[43] ensures a$(1-1/e)$ approximation guarantee.
Similarly, the work of Fabbri et al. [15] considers a gain function in the form of $\Delta_f(\cdot)=f(\emptyset)-f(\cdot)$ in the context of degree-regular directed graphs and with rewiring operations instead of shortcut edge additions. Here, the function $f$ is the largest hitting time instead of the average (see Definition 1). Fabbri et al. show that the existence of a multiplicative approximation algorithm implies P=NP. This result leads them to propose heuristic algorithms. However, their inapproximability result does not follow from the use of random walks, nor from the rewire operations. Their reduction still applies if one would replace "largest hitting time" with “largest shortest path distance," for example. Hence, their inapproximability result is an artefact of maximizing a gain function in combination with the max operator in the function $f$ (Definition 1).
The previous discussion motivates the following question:“What can we say about directly minimizing thefunctions $f$ and g, instead of indirectly optimizing them by maximizing associated gain functions?"
This paper provides several first algorithmic results and ideas regarding this question for both functions $f$ and $g.$ We take an abstract view and define our problems on input graphs in their simplest form, without the additional constraints (bounded-length random walks or degree-regular graphs) considered in Haddadan et al. [24] and Fabbri et al. [15], respectively. We consider uniform simple random walks of unbounded length on undirected, unweighted and connected input graphs.
Results and techniques. For the frst problem we study, minimize average hitting time (BMAH), we observe that the objective is supermodular and it can be optimized by a greedy bicriteria strategy[32], which offers an (1+$\epsilon)$-approximation at the cost of adding logarithmically more edges. In addition, we show how to speed up this algorithm by deriving an approximation for the objective, which relies on sampling a small number of bounded-length absorbing random walks and bounding the error term using eigenvalue techniques. The length of these walks directly depends on the mixing time and average degree of the red nodes, which are bounded quantities in most real-life networks. We then show that running the same greedy strategy of Liberty and Sviridenko [32] on the approximative values, still provides a (2+ ε)-approximation factor.
For our second problem, minimize maximum hitting time (BMMH), the objective function is neither supermodular nor submodular, and thus, it is a more challenging task. Nevertheless, we present two different algorithms, both with approximation guarantees, albeit, weaker than in the case of average hitting time. The first algorithm utilizes a relation between maximum and average hitting time, and uses the greedy strategy designed for the BMAH problem. The second algorithm leverages a novel connection between BMMH and the asymmetric $k$-center problem [26], and utilizes optimal methods developed for the latter problem [4].
Florian Adriaens, Honglian Wang, & Aristides Gionis
# Table 1: Results summary. The approximation factor is computed with respect to an optimal solution that uses at most $k$ edges.$d_{\mathrm{m}}$ is at most the maximum degree(see Section 6.2).
<table>
<tbody>
<tr>
<th> </th>
<th>Approx. factor</th>
<th>Number of shortcut edges</th>
</tr>
<tr>
<td>BMAH</td>
<td>$1+\epsilon$</td>
<td>$O(k\log(n/\epsilon))$</td>
</tr>
<tr>
<td>BMAH</td>
<td>$2+\epsilon$</td>
<td>$O(k\log(n/\epsilon))$</td>
</tr>
<tr>
<td>BMMH</td>
<td>$O(n^{3/4})$</td>
<td>$O(k\log n)$</td>
</tr>
<tr>
<td>BMMH</td>
<td>$( \log ^{* }( k)$ $d_{\mathrm{m} }$</td>
<td>$k$</td>
</tr>
</tbody>
</table>
## 2 RELATED WORK
A large body of work has been devoted to designing methods for optimizing certain graph properties via edge additions, edge rewirings, or other graph-edit operations. These approaches include methods for increasing graph robustness [10],maximizing the centrality of a group of nodes [37], improving the betweenness centrality of a node [8], reducing the average shortest path distances over all pairs of nodes [39, 45], minimizing the diameter of the graph [1, 14], increasing resilience to adversarial attacks [35], and more. Some of these approaches provide algorithms with provable approximation guarantees, while the others are mostly practical heuristics.
The setting of the problem we consider has also been studied in the context of understanding phenomena of bias, polarization, and segregation in social networks [6,11,12,17,21,23,40,48], and developing computational methods to mitigate those adverse effects [2,19,20,22,36,42,50-52]. Most of the works listed above seek to optimize a complex objective related to some dynamical diffusion process, such as information cascades [28], or opinion dynamics [18]. In all cases, the optimization functions studied in previous work are significantly different from the hitting-time objectives considered in this paper.
As discussed in the introduction, the work that is most closely related to ours, is the paper by Haddadan et al. [24], which seeks to reduce the structural bias between two polarized groups of a network via edge insertions. Their approach is based on hitting time for bounded-length random walks, but contrary to our proposal, they seek to maximize the reduction of hitting time caused by edge insertions, rather than directly minimizing the hitting time Similar ideas have been proposed by other authors, e.g., Fabbri et al. [15] and more recenty Coupette et al. [13], in the context of reducing exposure to harmful content. We consider our work to be a compelling extension of such previous ideas, not only due to optimizing a more natural objective, but also for coping with a more challenging optimization problem, for which our solution offers new technical ideas and insights.
## 3 $\textbf{PRELIMINARIES}$
Graphs. We consider undirected connected graphs $G=(V,E)$ with $|V|=n$ nodes, where the nodes $V$ are bipartitioned into two groups $R$ and $B.$ These groups as referred to as the red and blue nodes, respectively. We only consider bipartitions $V=\{R,B\}$ that are $valid$, meaning that $R$ and $B$ are disjoint and non-empty. For a nonvalid bipartition, our problem statements (Section 4.1) are either trivial or ill-defined. An edge is an inter-group edge if it has one blue and one red endpoint. For a subset of nodes $X\subseteq V$,we let
|
[
[
0.08660888671875,
0.0770263671875,
0.324462890625,
0.0770263671875,
0.324462890625,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6640625,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.6640625,
0.08709716796875
],
[
0.101318359375,
0.106689453125,
0.483642578125,
0.106689453125,
0.483642578125,
0.12060546875,
0.101318359375,
0.12060546875
],
[
0.51708984375,
0.10479736328125,
0.91259765625,
0.10479736328125,
0.91259765625,
0.11871337890625,
0.51708984375,
0.11871337890625
],
[
0.08660888671875,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.5185546875,
0.11871337890625,
0.91259765625,
0.11871337890625,
0.91259765625,
0.1319580078125,
0.5185546875,
0.1319580078125
],
[
0.0849609375,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1331787109375,
0.91259765625,
0.1331787109375,
0.91259765625,
0.146484375,
0.51708984375,
0.146484375
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.08416748046875,
0.1634521484375,
0.48193359375,
0.1622314453125,
0.48193359375,
0.1767578125,
0.08416748046875,
0.177978515625
],
[
0.61669921875,
0.1634521484375,
0.71142578125,
0.1634521484375,
0.71142578125,
0.1749267578125,
0.61669921875,
0.1749267578125
],
[
0.7099609375,
0.1666259765625,
0.71728515625,
0.1666259765625,
0.71728515625,
0.1717529296875,
0.7099609375,
0.1717529296875
],
[
0.7138671875,
0.1622314453125,
0.88232421875,
0.1622314453125,
0.88232421875,
0.175537109375,
0.7138671875,
0.175537109375
],
[
0.08660888671875,
0.1767578125,
0.48193359375,
0.1767578125,
0.48193359375,
0.1900634765625,
0.08660888671875,
0.1900634765625
],
[
0.55224609375,
0.175537109375,
0.6005859375,
0.175537109375,
0.6005859375,
0.1900634765625,
0.55224609375,
0.1900634765625
],
[
0.64208984375,
0.177978515625,
0.67626953125,
0.177978515625,
0.67626953125,
0.189453125,
0.64208984375,
0.189453125
],
[
0.75634765625,
0.177978515625,
0.84228515625,
0.177978515625,
0.84228515625,
0.189453125,
0.75634765625,
0.189453125
],
[
0.55224609375,
0.189453125,
0.60107421875,
0.189453125,
0.60107421875,
0.203857421875,
0.55224609375,
0.203857421875
],
[
0.6435546875,
0.1900634765625,
0.6787109375,
0.1900634765625,
0.6787109375,
0.20263671875,
0.6435546875,
0.20263671875
],
[
0.75341796875,
0.189453125,
0.84228515625,
0.1912841796875,
0.84228515625,
0.205810546875,
0.75341796875,
0.203857421875
],
[
0.7646484375,
0.20263671875,
0.833984375,
0.2052001953125,
0.833984375,
0.2203369140625,
0.7646484375,
0.2183837890625
],
[
0.08660888671875,
0.1912841796875,
0.48046875,
0.1912841796875,
0.48046875,
0.2020263671875,
0.08660888671875,
0.2020263671875
],
[
0.55126953125,
0.203857421875,
0.60107421875,
0.203857421875,
0.60107421875,
0.2183837890625,
0.55126953125,
0.2183837890625
],
[
0.63232421875,
0.203857421875,
0.685546875,
0.2020263671875,
0.68701171875,
0.2171630859375,
0.6337890625,
0.2197265625
],
[
0.55224609375,
0.2197265625,
0.60107421875,
0.2197265625,
0.60107421875,
0.2310791015625,
0.55224609375,
0.2310791015625
],
[
0.0849609375,
0.2052001953125,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2183837890625,
0.0849609375,
0.2183837890625
],
[
0.615234375,
0.2197265625,
0.705078125,
0.2197265625,
0.705078125,
0.2310791015625,
0.615234375,
0.2310791015625
],
[
0.79345703125,
0.2197265625,
0.8056640625,
0.2197265625,
0.8056640625,
0.2310791015625,
0.79345703125,
0.2310791015625
],
[
0.1004638671875,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.1004638671875,
0.246826171875
],
[
0.08660888671875,
0.2203369140625,
0.40283203125,
0.2203369140625,
0.40283203125,
0.2310791015625,
0.08660888671875,
0.2310791015625
],
[
0.08660888671875,
0.2474365234375,
0.48046875,
0.2474365234375,
0.48046875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.521484375,
0.26513671875,
0.53369140625,
0.26513671875,
0.53369140625,
0.271484375,
0.521484375,
0.271484375
],
[
0.5439453125,
0.260009765625,
0.689453125,
0.260009765625,
0.689453125,
0.2734375,
0.5439453125,
0.2734375
],
[
0.08660888671875,
0.27587890625,
0.48046875,
0.27587890625,
0.48046875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.5185546875,
0.279052734375,
0.9140625,
0.279052734375,
0.9140625,
0.292236328125,
0.5185546875,
0.292236328125
],
[
0.0882568359375,
0.290283203125,
0.4794921875,
0.290283203125,
0.4794921875,
0.29736328125,
0.0882568359375,
0.29736328125
],
[
0.51708984375,
0.29345703125,
0.9140625,
0.29345703125,
0.9140625,
0.306884765625,
0.51708984375,
0.306884765625
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.5185546875,
0.309326171875,
0.9111328125,
0.309326171875,
0.9111328125,
0.3193359375,
0.5185546875,
0.3193359375
],
[
0.08660888671875,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.08660888671875,
0.32763671875
],
[
0.5185546875,
0.322509765625,
0.91259765625,
0.322509765625,
0.91259765625,
0.333251953125,
0.5185546875,
0.333251953125
],
[
0.08660888671875,
0.330810546875,
0.48046875,
0.330810546875,
0.48046875,
0.341552734375,
0.08660888671875,
0.341552734375
],
[
0.5185546875,
0.335205078125,
0.9140625,
0.335205078125,
0.9140625,
0.348388671875,
0.5185546875,
0.348388671875
],
[
0.0849609375,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.51708984375,
0.348388671875,
0.9140625,
0.348388671875,
0.9140625,
0.3623046875,
0.51708984375,
0.3623046875
],
[
0.08660888671875,
0.359130859375,
0.48046875,
0.359130859375,
0.48046875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.51708984375,
0.3623046875,
0.91259765625,
0.3623046875,
0.91259765625,
0.375732421875,
0.51708984375,
0.375732421875
],
[
0.08660888671875,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
0.5185546875,
0.37744140625,
0.91259765625,
0.37744140625,
0.91259765625,
0.38818359375,
0.5185546875,
0.38818359375
],
[
0.0849609375,
0.385009765625,
0.48193359375,
0.385009765625,
0.48193359375,
0.3984375,
0.0849609375,
0.3984375
],
[
0.51708984375,
0.39013671875,
0.91259765625,
0.39013671875,
0.91259765625,
0.404052734375,
0.51708984375,
0.404052734375
],
[
0.0849609375,
0.40087890625,
0.381591796875,
0.40087890625,
0.381591796875,
0.410888671875,
0.0849609375,
0.410888671875
],
[
0.51708984375,
0.404052734375,
0.87841796875,
0.404052734375,
0.87841796875,
0.417236328125,
0.51708984375,
0.417236328125
],
[
0.10296630859375,
0.415283203125,
0.48193359375,
0.415283203125,
0.48193359375,
0.4287109375,
0.10296630859375,
0.4287109375
],
[
0.53369140625,
0.417236328125,
0.9140625,
0.417236328125,
0.9140625,
0.43115234375,
0.53369140625,
0.43115234375
],
[
0.08660888671875,
0.429931640625,
0.483642578125,
0.429931640625,
0.483642578125,
0.443115234375,
0.08660888671875,
0.443115234375
],
[
0.51708984375,
0.431884765625,
0.9140625,
0.431884765625,
0.9140625,
0.44580078125,
0.51708984375,
0.44580078125
],
[
0.0849609375,
0.443115234375,
0.48046875,
0.443115234375,
0.48046875,
0.45703125,
0.0849609375,
0.45703125
],
[
0.5185546875,
0.447509765625,
0.91259765625,
0.447509765625,
0.91259765625,
0.458251953125,
0.5185546875,
0.458251953125
],
[
0.10296630859375,
0.460205078125,
0.483642578125,
0.460205078125,
0.483642578125,
0.473388671875,
0.10296630859375,
0.473388671875
],
[
0.51708984375,
0.458984375,
0.91259765625,
0.458984375,
0.91259765625,
0.472900390625,
0.51708984375,
0.472900390625
],
[
0.08660888671875,
0.474853515625,
0.48193359375,
0.474853515625,
0.48193359375,
0.488037109375,
0.08660888671875,
0.488037109375
],
[
0.5185546875,
0.474853515625,
0.91259765625,
0.474853515625,
0.91259765625,
0.48486328125,
0.5185546875,
0.48486328125
],
[
0.0849609375,
0.488037109375,
0.48046875,
0.488037109375,
0.48046875,
0.501953125,
0.0849609375,
0.501953125
],
[
0.51708984375,
0.4873046875,
0.9140625,
0.4873046875,
0.9140625,
0.50048828125,
0.51708984375,
0.50048828125
],
[
0.0849609375,
0.501953125,
0.48046875,
0.501953125,
0.48046875,
0.51513671875,
0.0849609375,
0.51513671875
],
[
0.51708984375,
0.5,
0.9140625,
0.50048828125,
0.9140625,
0.5146484375,
0.51708984375,
0.51318359375
],
[
0.0849609375,
0.5166015625,
0.48046875,
0.5166015625,
0.48046875,
0.52978515625,
0.0849609375,
0.52978515625
],
[
0.5185546875,
0.5166015625,
0.91259765625,
0.5166015625,
0.91259765625,
0.5263671875,
0.5185546875,
0.5263671875
],
[
0.0849609375,
0.52880859375,
0.4794921875,
0.52880859375,
0.4794921875,
0.54248046875,
0.0849609375,
0.54248046875
],
[
0.5185546875,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54052734375,
0.5185546875,
0.54052734375
],
[
0.0849609375,
0.54345703125,
0.48193359375,
0.54345703125,
0.48193359375,
0.556640625,
0.0849609375,
0.556640625
],
[
0.51708984375,
0.54248046875,
0.7158203125,
0.54248046875,
0.7158203125,
0.55615234375,
0.51708984375,
0.55615234375
],
[
0.08416748046875,
0.5595703125,
0.2296142578125,
0.5595703125,
0.2296142578125,
0.5693359375,
0.08416748046875,
0.5693359375
],
[
0.53173828125,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.5693359375,
0.53173828125,
0.5693359375
],
[
0.0849609375,
0.57470703125,
0.48193359375,
0.57470703125,
0.48193359375,
0.58837890625,
0.0849609375,
0.58837890625
],
[
0.51708984375,
0.5693359375,
0.91259765625,
0.5693359375,
0.91259765625,
0.58349609375,
0.51708984375,
0.58349609375
],
[
0.08660888671875,
0.58837890625,
0.483642578125,
0.58837890625,
0.483642578125,
0.6015625,
0.08660888671875,
0.6015625
],
[
0.51708984375,
0.583984375,
0.9140625,
0.583984375,
0.9140625,
0.59765625,
0.51708984375,
0.59765625
],
[
0.0849609375,
0.6015625,
0.48046875,
0.6015625,
0.48046875,
0.61572265625,
0.0849609375,
0.61572265625
],
[
0.51708984375,
0.59765625,
0.9140625,
0.59765625,
0.9140625,
0.611328125,
0.51708984375,
0.611328125
],
[
0.08660888671875,
0.61865234375,
0.48046875,
0.61865234375,
0.48046875,
0.62890625,
0.08660888671875,
0.62890625
],
[
0.51953125,
0.61279296875,
0.9111328125,
0.61279296875,
0.9111328125,
0.62353515625,
0.51953125,
0.62353515625
],
[
0.0849609375,
0.6298828125,
0.48193359375,
0.6298828125,
0.48193359375,
0.64306640625,
0.0849609375,
0.64306640625
],
[
0.51953125,
0.626953125,
0.91259765625,
0.626953125,
0.91259765625,
0.63720703125,
0.51953125,
0.63720703125
],
[
0.0849609375,
0.64453125,
0.483642578125,
0.64453125,
0.483642578125,
0.65771484375,
0.0849609375,
0.65771484375
],
[
0.5185546875,
0.6416015625,
0.91259765625,
0.6416015625,
0.91259765625,
0.6513671875,
0.5185546875,
0.6513671875
],
[
0.08660888671875,
0.6591796875,
0.4794921875,
0.6591796875,
0.4794921875,
0.66943359375,
0.08660888671875,
0.66943359375
],
[
0.51953125,
0.65478515625,
0.91259765625,
0.65478515625,
0.91259765625,
0.66552734375,
0.51953125,
0.66552734375
],
[
0.0849609375,
0.671875,
0.48193359375,
0.671875,
0.48193359375,
0.68505859375,
0.0849609375,
0.68505859375
],
[
0.51708984375,
0.66748046875,
0.9140625,
0.66748046875,
0.9140625,
0.68115234375,
0.51708984375,
0.68115234375
],
[
0.0849609375,
0.68505859375,
0.483642578125,
0.68505859375,
0.483642578125,
0.69873046875,
0.0849609375,
0.69873046875
],
[
0.51953125,
0.68115234375,
0.91259765625,
0.68115234375,
0.91259765625,
0.6943359375,
0.51953125,
0.6943359375
],
[
0.51708984375,
0.693359375,
0.91259765625,
0.693359375,
0.91259765625,
0.70703125,
0.51708984375,
0.70703125
],
[
0.08660888671875,
0.70068359375,
0.48046875,
0.70068359375,
0.48046875,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.0849609375,
0.71337890625,
0.48193359375,
0.71337890625,
0.48193359375,
0.7265625,
0.0849609375,
0.7265625
],
[
0.5185546875,
0.71044921875,
0.9111328125,
0.71044921875,
0.9111328125,
0.72021484375,
0.5185546875,
0.72021484375
],
[
0.0849609375,
0.7265625,
0.48193359375,
0.7265625,
0.48193359375,
0.74072265625,
0.0849609375,
0.74072265625
],
[
0.51708984375,
0.72265625,
0.9140625,
0.72265625,
0.9140625,
0.736328125,
0.51708984375,
0.736328125
],
[
0.0849609375,
0.7412109375,
0.48193359375,
0.7412109375,
0.48193359375,
0.7548828125,
0.0849609375,
0.7548828125
],
[
0.51708984375,
0.736328125,
0.75341796875,
0.736328125,
0.75341796875,
0.75,
0.51708984375,
0.75
],
[
0.1004638671875,
0.75830078125,
0.488525390625,
0.75830078125,
0.488525390625,
0.771484375,
0.1004638671875,
0.771484375
],
[
0.51953125,
0.7666015625,
0.5439453125,
0.7666015625,
0.5439453125,
0.7744140625,
0.51953125,
0.7744140625
],
[
0.5390625,
0.76318359375,
0.68701171875,
0.76318359375,
0.68701171875,
0.7763671875,
0.5390625,
0.7763671875
],
[
0.0849609375,
0.77099609375,
0.483642578125,
0.771484375,
0.483642578125,
0.78515625,
0.0849609375,
0.7841796875
],
[
0.08660888671875,
0.787109375,
0.48046875,
0.787109375,
0.48046875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.5185546875,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.5185546875,
0.798828125
],
[
0.08660888671875,
0.8017578125,
0.4794921875,
0.8017578125,
0.4794921875,
0.8125,
0.08660888671875,
0.8125
],
[
0.51953125,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.51953125,
0.81298828125
],
[
0.0849609375,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.82763671875,
0.0849609375,
0.82763671875
],
[
0.5185546875,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.5185546875,
0.8271484375
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.5185546875,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.5185546875,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.869140625,
0.0849609375,
0.869140625
],
[
0.5185546875,
0.85693359375,
0.91259765625,
0.85693359375,
0.91259765625,
0.8671875,
0.5185546875,
0.8671875
],
[
0.08416748046875,
0.8671875,
0.48193359375,
0.86865234375,
0.48193359375,
0.88330078125,
0.08416748046875,
0.8818359375
],
[
0.51708984375,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.08416748046875,
0.88330078125,
0.361083984375,
0.88330078125,
0.361083984375,
0.896484375,
0.08416748046875,
0.896484375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"KDD'23,August 6-10,2023, Long Beach,CA,USA",
"Florian Adriaens, Honglian Wang, & Aristides Gionis",
"For example, the work of Haddadan et al. [24] considers max-",
"Table 1: Results summary. The approximation factor is com.",
"imizing the gain function A(F) = g(O) - g(F) in the context of",
"puted with respect to an optimal solution that uses at most",
"bounded-length random walks in directed graphs. Here, F is a set",
"k edges. dm is at most the maximum degree (see Section 6.2).",
"of at most k new shortcut edges that will be added to an input",
"graph, and the function g is the average expected hitting time for",
"Approx.factor",
"",
"Numberof shortcutedges",
"a random walk starting in one group to hit the other group (see",
"BMAH",
"1+e",
"O(klog(n/e))",
"BMAH",
"2+e",
"O(klog(n/e))",
"O(k log n)",
"Definition 1).They prove that Aais a non-negative monotone sub.",
"BMMH",
"0(n3/4)",
"BMMH",
"modular set function, and thus, the well-known greedy algorithm",
"O (log*(k) dm)",
"k",
"Similarly, the work of Fabbri et al. [15] considers a gain function",
"of[43] ensuresa (1-1/e)approximationguarantee.",
"in theform of△()=f()-f()inthe context of degree-regular",
"directed graphs and with rewiring operations instead of shortcut",
"2",
"RELATED WORK",
"edgeadditions.Here,thefunction fisthelargesthittingtimeinstead",
"A large body of work has been devoted to designing methods for op-",
"oftheaverage(seeDefinitionlFabbrietal.showthattheexistence",
"timizing certain graph properties via edge additions, edge rewirings,",
"of a multiplicative approximation algorithm implies P = NP. This",
"orothergraph-edit operations.Theseapproachesincludemethods",
"result leads them to proposeheuristic algorithms.However, their",
"forincreasinggraphrobustness[1o],maximizingthecentralityof",
"inapproximabilityresultdoesnotfollowfromtheuseofrandom",
"a group of nodes [37], improving the betweenness centrality of",
"walks, nor from the rewire operations. Their reduction still applies",
"a node [8], reducing the average shortest path distances over all",
"if onewouldreplace“largesthitting time\"with\"largest shortest",
"pairs of nodes [39, 45], minimizing the diameter of the graph [1, 14]",
"path distance,for example.Hence,their inapproximabilityresult",
"increasing resilience to adversarial attacks[35], and more.Some of",
"is an artefact of maximizing a gain function in combination with",
"these approaches provide algorithms with provable approximation",
"themaxoperator inthefunction f(Definitioni)",
"guarantees, while the others are mostly practical heuristics.",
"The previous discussion motivates the following question:\"what",
"The setting of the problem we consider has also been studied",
"can we say about directly minimizing the functions f and g, instead of",
"in the context of understanding phenomena of bias, polarization,",
"indirectly optimizing them by maximizing associated gain functions?\"",
"and segregation in social networks[6,11,12,17,21,23,40,48]",
"This paper provides several first algorithmic results and ideas re-",
"and developing computational methods to mitigate those adverse",
"garding this question for both functions f and g. We take an abstract",
"effects[2,19,20,22,36,42,50-52].Mostoftheworkslistedabove",
"view and define our problems on input graphs in their simplest",
"seek to optimize a complex objective related to some dynamical",
"form, without the additional constraints (bounded-length random",
"diffusion process,suchas information cascades[28],or opinion",
"walks or degree-regular graphs) considered in Haddadan et al. [24]",
"dynamics[i8l.Inall cases,theoptimizationfunctionsstudiedin",
"and Fabbri et al. [15], respectively. We consider uniform simple",
"previous work are significantly different from the hitting-time ob-",
"random walks of unbounded length on undirected, unweighted and",
"jectives considered in this paper.",
"connectedinputgraphs",
"As discussed in the introduction, the work that is most closely",
"Results and techniques. For the first problem we study, mini-",
"related to ours, is the paper by Haddadan et al. [24], which seeks",
"mize average hitting time (BMAH), we observe that the objective is",
"to reduce the structural bias between two polarized groups of a",
"supermodular and it can be optimized by a greedy bicriteria strat",
"network via edge insertions. Their approach is based on hitting",
"egy[32],whichoffersan(1+e)-approximationatthecostofadding",
"timeforbounded-length randomwalks,but contrarytoourpro",
"logarithmically more edges. In addition, we show how to speed",
"posal, they seekto maximize the reduction of hittingtime caused",
"up this algorithm by deriving an approximation for the objective,",
"by edge insertions, rather than directly minimizing the hitting time",
"whichreliesonsamplingasmallnumberofbounded-lengthabsorb",
"Similar ideas havebeen proposed by other authors, e.g.,Fabbri",
"ing random walks and bounding the error term using eigenvalue",
"et al. [15] and more recenty Coupette et al. [13], in the context of",
"techniques. The length of these walks directly depends on the mix-",
"reducing exposure to harmful content. We consider our work to",
"be a compelling extension of such previous ideas, not only due",
"ingtimeand averagedegreeoftherednodes,whicharebounded",
"quantities in most real-life networks. We then show that running",
"tooptimizingamorenatural objective, butalso for coping with",
"the same greedy strategy of Liberty and Sviridenko [32] on the",
"amore challenging optimization problem,for which our solution",
"approximative values, still provides a (2 + e)-approximation factor",
"offers new technical ideas and insights.",
"For our second problem, minimize maximum hitting time (BMMH),",
"3",
"PRELIMINARIES",
"the objective function is neither supermodular nor submodular,and",
"thus, it is a more challenging task.Nevertheless, we present two",
"Graphs. We consider undirected connected graphs G = (V, E) with",
"differentalgorithms,bothwithapproximationguarantees,albeit",
"IV = n nodes, where the nodes V are bipartitioned into two groups",
"weaker than in the case of average hitting time. The first algorithm",
"R and B. These groups as referred to as the red and blue nodes.",
"utilizes a relation between maximum and averagehitting time,and",
"respectively.We only consider bipartitions V [R,B)that are",
"uses the greedy strategy designed for the BMAH problem. The",
"valid, meaning that R and B are disjoint and non-empty. For a non-",
"second algorithm leverages a novel connection between BMMH",
"valid bipartition,ourproblem statements (Section 4.1)are either",
"and the asymmetric k-center problem [26], and utilizes optimal",
"trivial or ill-defined. An edge is an inter-group edge if it has one",
"methods developed for the latter problem [4].",
"blue and one red endpoint. For a subset of nodes X V, we let"
] |
{
"arxiv_id": "2306.03571",
"language": "en",
"page_idx": 1
}
|
||
2306.03571-en-2
|
2306.03571-en-2.jpg
|
Minimizing Hitting Time between Disparate Groups with Shortcut Edges
$E[X]$ be the subset of edges of G that have both endpoints in $X$,and $G[X]=(X,E[X])$ be the subgraph of $G$ induced by $X.$If $F$ is a set of non-edges of $G=(V,E)$,then $G+F$ denotes the graph $(V,E\cup F).$ The degree of a node $v$ in $G$,which is the number of neighbors of $v$ in $G$, is denoted as $d_G(v).$ We let $d_R$ be the average of the degrees of red nodes.
Random walks. Every random walk will be a uniform simple random walk, unless stated otherwise. By uniform simple random walk we mean that at each step one of the neighbors of the current node is selected with uniform probability. Given a graph $G=(V,E)$ and $A\subseteq V$,the random variable $\tau_u(A)$ indicates the first time that a random walk on $G$ starting from $u$ visits $A.$ If $u\in A$, we set $\tau_u(A)=0.$ The expectation of this random variable is denoted as $H_G(u,A)=\mathbb{E}_G[\tau_u(A)]$,where the subscript emphasizes that the random walk is on $G.$ With a slight abuse of notation, $H_G(u,v)$ is thus the (expected) hitting time for a walk starting from $u$ to visit $v$ for the first time. Note that $H_G(u,u)=0$ and $H_G(u,A)=0$ if and only if $u\in A.$ It also holds that $H_G(u,A)\leq H_G(u,B)$ when $B\subseteq A$, and in particular $H_G(u,A)\leq\min_{a\in A}H_G(u,a).$
Matrices and norms. We denote A as the adjacency matrix of a graph, and D is the degree diagonal matrix. Let $e_r\in\mathbb{R}^{r\times1}$ be the vector of all-ones of size $r$, and let $\|\cdot\|_2$ denote the spectral norm.
$Additional\textit{ definitions. For a graph}G= ( V, E)$ with a valid bipartition $V=\{R,B\}$,let$(R\times B)\setminus E$ be the set of inter-group non-edges in $G.$ Define the following two set functions $f,g:2^{(\tilde{R}\times B)\setminus E}\to\mathbb{R}_{>0}$ as the maximum and average hitting time.
$\textbf{DEFINITION 1 ( MAXIMUM AND AVERAGE HITTING TIME) . For a}$
$set\textit{of inter- group non- edges F}\subseteq ( R\times B) \setminus E, we\textit{define the maximum}$
hitting time from $R$ to $B$ on graph $G+Fby$
$$f(F)\triangleq\max_{r\in R}H_{G+F}(r,B),$$
and the average hitting time from $R$ to $B$ on graph $G+F$ by
$$g(F)\triangleq\frac{1}{|R|}\sum_{r\in R}H_{G+F}(r,B).$$
4 РROвLEMS AND OBSERVATIONS
In this section, we formally define the problems we study, and we
discuss their properties.
# 4.1 Problems
We introduce and study the following two problems, which we denote as the BMMH (Budgeted Minimum Maximum Hitting time) and BMAH (Budgeted Minimum Average Hitting time) problems.
$\textbf{Р к о в с е м }1$ (BUDGETED MINIMUM MAXIMUM HITTING TIME PROBLEM (BMMH)). Given an undirected connected graph G = $(V,E),withavalidbipartitionV=\{R,B\},andabudgetk\in\mathbb{N}$, we seek to find a set of inter-group non-edges F C $(R\times B)\setminus E$ with $|F|\leq k$ that minimizes $f(F),wheref$ is the maximum hitting-time $function\textit{ defined in Definition 1. }$
PROBLEM 2 (BUDGETED MINIMUM AVERAGE HITTING TIME PROBLEM (BMAH)). Identical definition to Problem 1, but we aim to minimize the function g, instead off, where g is the average hitting-time $function\textit{ defined in Definition 1. }$
KDD'23, August 6-10, 2023, Long Beach, CA, USA
BMMH aims to augment $G$ with $k$ new inter-group edges, so as to minimize the largest hitting time to a blue node, for random walks starting from red nodes. Likewise, BMAH aims to minimize the average hitting time from red to blue nodes. Note that for every red node we allow multiple new edges to be incident to it, as long as the augmented graph remains simple.
## 4.2 Observations
The NP-hardness of both BMMH and BMAH follows from a reduction from the minimum set cover problem, very similar to the reduction detailed by Haddadan et al. [24] and omitted for brevity. We present several useful observations regarding BMMH and BMAH that will guide the design of approximation algorithms:
Oвse RVATION 1. Changing the blue endpoints of the edges in a
feasible solution F does not change f(F) or g(F).
PROOF. Every walk starts from a red node and halts the moment a blue node is hit, hence the identity of the blue endpoints of edges in $F$ do not matter.
口
Both problems thus reduce to selecting a multiset of red endpoints, as we allow multiple edges in $F$ to have the same red endpoint. If a red node is connected to all the blue nodes already, we can of course not add any more new edges incident to it.
Oвse RVATION 2 (момотомґсгту). $IfY\subseteq X,thenf(X)\leq f(Y)$
and $g( X) \leq g( Y) .$
PROOF. Let $F$ be a feasible solution. For every $r\in R$, the hitting time $H_G+F(r,B)$ is a monotonically decreasing set function of $F$, since the edges in $F$ are inter-group edges. Every $e\in F$ only helps in reducing the hitting time from $r$ to $B.$ This can be proven more formally by a straightforward coupling argument between a walker on $G$ and $G+F\left[24,34\right].$ Since individual hitting times are monotone,
口
the functions $f$ and $g$ are monotone, as well.
The following supermodularity property for individual hitting times is less obvious, and a variant of it was proven by Haddadan et al. [24] for bounded-length random walks on directed graphs. We give a different and shorter proof of supermodularity in our case of unbounded random walks, by using a coupling argument.
Oвse RVATION 3 (SUPERMODULARITY OF HITTING TIMES). For all
$r\in R$ and for all e$_1\neq e_2\in(R\times B)\setminus E$ it holds that
$$H_{G+e_2}(r,B)-H_{G+\{e_1,e_2\}}(r,B)\leq H_G(r,B)-H_{G+e_1}(r,B)$$
(1)
PROOF. Pick any $r\in R.$ We assume all walks start from $r.$ Consider the right-hand side of Eq. (1). Couple walks on $G+e_1$ to walks on $G$ by simply following the walker on $G$, until the walker hits the red endpoint $r_{1}\in R$ of the edge $e_{1}.$If the walker hits $r_{1}$,follow the edge $e_{1}$ with probability $\frac 1{d_{G}( r_{1}) + 1}.$With probability $\frac {d_{G}( r_{1}) }{d_{G}( r_{1}) + 1}$,keep following the walker on $G.$It is clear that the margi of this coupled walk is a uniform random walk on $G+e_1.$
As long as $e_1$ is not traversed, the two coupled walks are identical. If $e_1$ is traversed, the walks are identical up until the time prior to traversing $e_1$,when they are both in $r_1$ for the last time. On this moment, the walker on $G$ still needs to travel $H_G(r_1,B)$ steps on
|
[
[
0.08660888671875,
0.0770263671875,
0.4306640625,
0.0770263671875,
0.4306640625,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.67236328125,
0.0738525390625,
0.9140625,
0.07452392578125,
0.9140625,
0.08837890625,
0.67236328125,
0.08709716796875
],
[
0.0882568359375,
0.106689453125,
0.483642578125,
0.106689453125,
0.483642578125,
0.12060546875,
0.0882568359375,
0.12060546875
],
[
0.53125,
0.1060791015625,
0.9140625,
0.10797119140625,
0.9140625,
0.121826171875,
0.53125,
0.11932373046875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.515625,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.515625,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.483642578125,
0.1351318359375,
0.483642578125,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.515625,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.515625,
0.1483154296875
],
[
0.08660888671875,
0.15087890625,
0.48046875,
0.15087890625,
0.48046875,
0.1610107421875,
0.08660888671875,
0.1610107421875
],
[
0.51708984375,
0.149658203125,
0.91259765625,
0.149658203125,
0.91259765625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.08416748046875,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.08416748046875,
0.1767578125
],
[
0.5146484375,
0.1622314453125,
0.91259765625,
0.1622314453125,
0.91259765625,
0.175537109375,
0.5146484375,
0.175537109375
],
[
0.0849609375,
0.1787109375,
0.1641845703125,
0.1787109375,
0.1641845703125,
0.189453125,
0.0849609375,
0.189453125
],
[
0.51708984375,
0.177978515625,
0.76123046875,
0.177978515625,
0.76123046875,
0.1912841796875,
0.51708984375,
0.1912841796875
],
[
0.0849609375,
0.1932373046875,
0.48193359375,
0.1932373046875,
0.48193359375,
0.20703125,
0.0849609375,
0.20703125
],
[
0.08416748046875,
0.2083740234375,
0.48193359375,
0.2083740234375,
0.48193359375,
0.2215576171875,
0.08416748046875,
0.2215576171875
],
[
0.51708984375,
0.203857421875,
0.66650390625,
0.203857421875,
0.66650390625,
0.2171630859375,
0.51708984375,
0.2171630859375
],
[
0.0849609375,
0.2215576171875,
0.483642578125,
0.2215576171875,
0.483642578125,
0.23486328125,
0.0849609375,
0.23486328125
],
[
0.51708984375,
0.2215576171875,
0.9140625,
0.2215576171875,
0.9140625,
0.23486328125,
0.51708984375,
0.23486328125
],
[
0.08660888671875,
0.2374267578125,
0.48046875,
0.2374267578125,
0.48046875,
0.2474365234375,
0.08660888671875,
0.2474365234375
],
[
0.5185546875,
0.238037109375,
0.91259765625,
0.238037109375,
0.91259765625,
0.248779296875,
0.5185546875,
0.248779296875
],
[
0.08416748046875,
0.25,
0.48046875,
0.25,
0.48046875,
0.26318359375,
0.08416748046875,
0.26318359375
],
[
0.51708984375,
0.25,
0.9140625,
0.25,
0.9140625,
0.26318359375,
0.51708984375,
0.26318359375
],
[
0.0849609375,
0.26318359375,
0.483642578125,
0.26318359375,
0.483642578125,
0.276611328125,
0.0849609375,
0.276611328125
],
[
0.515625,
0.26318359375,
0.9140625,
0.261962890625,
0.9140625,
0.276611328125,
0.515625,
0.27783203125
],
[
0.08416748046875,
0.27783203125,
0.48193359375,
0.276611328125,
0.48193359375,
0.290283203125,
0.08416748046875,
0.291748046875
],
[
0.51708984375,
0.27783203125,
0.85107421875,
0.27783203125,
0.85107421875,
0.291748046875,
0.51708984375,
0.291748046875
],
[
0.0849609375,
0.291748046875,
0.48193359375,
0.291748046875,
0.48193359375,
0.304931640625,
0.0849609375,
0.304931640625
],
[
0.0849609375,
0.304931640625,
0.483642578125,
0.304931640625,
0.483642578125,
0.318115234375,
0.0849609375,
0.318115234375
],
[
0.53369140625,
0.2998046875,
0.9140625,
0.2998046875,
0.9140625,
0.313232421875,
0.53369140625,
0.313232421875
],
[
0.515625,
0.3125,
0.80224609375,
0.313232421875,
0.80224609375,
0.32763671875,
0.515625,
0.326904296875
],
[
0.08660888671875,
0.32080078125,
0.48046875,
0.32080078125,
0.48046875,
0.330810546875,
0.08660888671875,
0.330810546875
],
[
0.0849609375,
0.333251953125,
0.483642578125,
0.333251953125,
0.483642578125,
0.3466796875,
0.0849609375,
0.3466796875
],
[
0.53173828125,
0.33837890625,
0.9140625,
0.33837890625,
0.9140625,
0.3515625,
0.53173828125,
0.3515625
],
[
0.08251953125,
0.34521484375,
0.48193359375,
0.3466796875,
0.48193359375,
0.35986328125,
0.08251953125,
0.359130859375
],
[
0.51708984375,
0.352783203125,
0.91259765625,
0.352783203125,
0.91259765625,
0.3662109375,
0.51708984375,
0.3662109375
],
[
0.08416748046875,
0.35986328125,
0.36767578125,
0.35986328125,
0.36767578125,
0.373779296875,
0.08416748046875,
0.373779296875
],
[
0.51708984375,
0.36865234375,
0.63232421875,
0.36865234375,
0.63232421875,
0.37890625,
0.51708984375,
0.37890625
],
[
0.90625,
0.3662109375,
0.91650390625,
0.375,
0.90869140625,
0.380126953125,
0.89892578125,
0.37255859375
],
[
0.0849609375,
0.37744140625,
0.48046875,
0.37744140625,
0.48046875,
0.391357421875,
0.0849609375,
0.391357421875
],
[
0.08416748046875,
0.39013671875,
0.483642578125,
0.39013671875,
0.483642578125,
0.40478515625,
0.08416748046875,
0.40478515625
],
[
0.53173828125,
0.391357421875,
0.9140625,
0.391357421875,
0.9140625,
0.40478515625,
0.53173828125,
0.40478515625
],
[
0.0849609375,
0.406005859375,
0.42236328125,
0.406005859375,
0.42236328125,
0.419189453125,
0.0849609375,
0.419189453125
],
[
0.5146484375,
0.40478515625,
0.9150390625,
0.404052734375,
0.9150390625,
0.41845703125,
0.5146484375,
0.419189453125
],
[
0.08660888671875,
0.42236328125,
0.48193359375,
0.42236328125,
0.48193359375,
0.436279296875,
0.08660888671875,
0.436279296875
],
[
0.51708984375,
0.419189453125,
0.9140625,
0.419189453125,
0.9140625,
0.43310546875,
0.51708984375,
0.43310546875
],
[
0.08660888671875,
0.438232421875,
0.4794921875,
0.438232421875,
0.4794921875,
0.448974609375,
0.08660888671875,
0.448974609375
],
[
0.515625,
0.434326171875,
0.861328125,
0.434326171875,
0.861328125,
0.447509765625,
0.515625,
0.447509765625
],
[
0.08087158203125,
0.4443359375,
0.486083984375,
0.4462890625,
0.486083984375,
0.46728515625,
0.08087158203125,
0.46533203125
],
[
0.53125,
0.453857421875,
0.91259765625,
0.455078125,
0.91259765625,
0.4697265625,
0.53125,
0.468505859375
],
[
0.08251953125,
0.46533203125,
0.3203125,
0.46533203125,
0.3203125,
0.479248046875,
0.08251953125,
0.479248046875
],
[
0.51708984375,
0.4697265625,
0.625,
0.4697265625,
0.625,
0.48291015625,
0.51708984375,
0.48291015625
],
[
0.10455322265625,
0.488037109375,
0.48046875,
0.488037109375,
0.48046875,
0.498779296875,
0.10455322265625,
0.498779296875
],
[
0.53125,
0.492431640625,
0.91259765625,
0.49365234375,
0.91259765625,
0.50830078125,
0.53125,
0.5068359375
],
[
0.0849609375,
0.50048828125,
0.48193359375,
0.50048828125,
0.48193359375,
0.5146484375,
0.0849609375,
0.5146484375
],
[
0.5146484375,
0.50830078125,
0.91259765625,
0.50830078125,
0.91259765625,
0.521484375,
0.5146484375,
0.521484375
],
[
0.08416748046875,
0.51513671875,
0.34716796875,
0.51513671875,
0.34716796875,
0.52880859375,
0.08416748046875,
0.52880859375
],
[
0.51708984375,
0.52294921875,
0.9140625,
0.52294921875,
0.9140625,
0.5361328125,
0.51708984375,
0.5361328125
],
[
0.2091064453125,
0.5341796875,
0.35546875,
0.5341796875,
0.35546875,
0.548828125,
0.2091064453125,
0.548828125
],
[
0.51708984375,
0.5361328125,
0.91259765625,
0.5361328125,
0.91259765625,
0.5498046875,
0.51708984375,
0.5498046875
],
[
0.25732421875,
0.54736328125,
0.283447265625,
0.54736328125,
0.283447265625,
0.5546875,
0.25732421875,
0.5546875
],
[
0.51708984375,
0.5498046875,
0.9140625,
0.5498046875,
0.9140625,
0.56298828125,
0.51708984375,
0.56298828125
],
[
0.0849609375,
0.56103515625,
0.443603515625,
0.56103515625,
0.443603515625,
0.57470703125,
0.0849609375,
0.57470703125
],
[
0.5146484375,
0.5625,
0.9140625,
0.5625,
0.9140625,
0.57568359375,
0.5146484375,
0.57568359375
],
[
0.2467041015625,
0.5791015625,
0.366943359375,
0.583984375,
0.365234375,
0.60595703125,
0.2459716796875,
0.60107421875
],
[
0.51708984375,
0.5771484375,
0.7861328125,
0.5771484375,
0.7861328125,
0.59033203125,
0.51708984375,
0.59033203125
],
[
0.89892578125,
0.5791015625,
0.91259765625,
0.57763671875,
0.9140625,
0.58837890625,
0.900390625,
0.5888671875
],
[
0.1993408203125,
0.5869140625,
0.2467041015625,
0.58544921875,
0.2484130859375,
0.599609375,
0.200927734375,
0.6015625
],
[
0.2467041015625,
0.5869140625,
0.2998046875,
0.5947265625,
0.294189453125,
0.619140625,
0.2418212890625,
0.6123046875
],
[
0.53515625,
0.60400390625,
0.9111328125,
0.60400390625,
0.9111328125,
0.6142578125,
0.53515625,
0.6142578125
],
[
0.0849609375,
0.6220703125,
0.390625,
0.6220703125,
0.390625,
0.63525390625,
0.0849609375,
0.63525390625
],
[
0.51708984375,
0.6162109375,
0.9140625,
0.6162109375,
0.9140625,
0.6298828125,
0.51708984375,
0.6298828125
],
[
0.515625,
0.62890625,
0.9140625,
0.6298828125,
0.9140625,
0.64453125,
0.515625,
0.64306640625
],
[
0.08416748046875,
0.6396484375,
0.48193359375,
0.64013671875,
0.48193359375,
0.65478515625,
0.08416748046875,
0.65380859375
],
[
0.51708984375,
0.64453125,
0.9140625,
0.64453125,
0.9140625,
0.65771484375,
0.51708984375,
0.65771484375
],
[
0.08251953125,
0.65380859375,
0.2303466796875,
0.65380859375,
0.2303466796875,
0.66748046875,
0.08251953125,
0.66748046875
],
[
0.51708984375,
0.6591796875,
0.91015625,
0.6591796875,
0.91015625,
0.67236328125,
0.51708984375,
0.67236328125
],
[
0.0849609375,
0.681640625,
0.2059326171875,
0.681640625,
0.2059326171875,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.53173828125,
0.6796875,
0.9150390625,
0.6796875,
0.9150390625,
0.693359375,
0.53173828125,
0.693359375
],
[
0.08416748046875,
0.69873046875,
0.48193359375,
0.70068359375,
0.48193359375,
0.7138671875,
0.08416748046875,
0.7119140625
],
[
0.51708984375,
0.6943359375,
0.8203125,
0.6943359375,
0.8203125,
0.70849609375,
0.51708984375,
0.70849609375
],
[
0.0849609375,
0.7138671875,
0.483642578125,
0.7138671875,
0.483642578125,
0.72802734375,
0.0849609375,
0.72802734375
],
[
0.53369140625,
0.7138671875,
0.88134765625,
0.71337890625,
0.88134765625,
0.73095703125,
0.53369140625,
0.73193359375
],
[
0.8798828125,
0.72265625,
0.88623046875,
0.71728515625,
0.89208984375,
0.7216796875,
0.8857421875,
0.7265625
],
[
0.89453125,
0.71728515625,
0.9111328125,
0.71728515625,
0.9111328125,
0.7265625,
0.89453125,
0.7265625
],
[
0.08416748046875,
0.72802734375,
0.4794921875,
0.72802734375,
0.4794921875,
0.7412109375,
0.08416748046875,
0.7412109375
],
[
0.53173828125,
0.74072265625,
0.9140625,
0.74072265625,
0.9140625,
0.75390625,
0.53173828125,
0.75390625
],
[
0.10296630859375,
0.75048828125,
0.48046875,
0.75048828125,
0.48046875,
0.76123046875,
0.10296630859375,
0.76123046875
],
[
0.515625,
0.75244140625,
0.9140625,
0.75390625,
0.9140625,
0.767578125,
0.515625,
0.7666015625
],
[
0.0849609375,
0.76318359375,
0.48046875,
0.76318359375,
0.48046875,
0.7763671875,
0.0849609375,
0.7763671875
],
[
0.51708984375,
0.76953125,
0.91259765625,
0.76953125,
0.91259765625,
0.77978515625,
0.51708984375,
0.77978515625
],
[
0.08660888671875,
0.77880859375,
0.48193359375,
0.77880859375,
0.48193359375,
0.7890625,
0.08660888671875,
0.7890625
],
[
0.51708984375,
0.7822265625,
0.9140625,
0.7822265625,
0.9140625,
0.79541015625,
0.51708984375,
0.79541015625
],
[
0.08660888671875,
0.79150390625,
0.48193359375,
0.79150390625,
0.48193359375,
0.8046875,
0.08660888671875,
0.8046875
],
[
0.51318359375,
0.79541015625,
0.91650390625,
0.79541015625,
0.91650390625,
0.8193359375,
0.51318359375,
0.8193359375
],
[
0.833984375,
0.79541015625,
0.865234375,
0.79541015625,
0.865234375,
0.80322265625,
0.833984375,
0.80322265625
],
[
0.0849609375,
0.8037109375,
0.483642578125,
0.8046875,
0.483642578125,
0.8193359375,
0.0849609375,
0.818359375
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.0849609375,
0.8193359375,
0.27685546875,
0.8193359375,
0.27685546875,
0.83349609375,
0.0849609375,
0.83349609375
],
[
0.51708984375,
0.8271484375,
0.86279296875,
0.8271484375,
0.86279296875,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.101318359375,
0.84033203125,
0.483642578125,
0.84033203125,
0.483642578125,
0.85400390625,
0.101318359375,
0.85400390625
],
[
0.53173828125,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.53173828125,
0.85498046875
],
[
0.08416748046875,
0.85498046875,
0.484375,
0.85498046875,
0.484375,
0.86865234375,
0.08416748046875,
0.86865234375
],
[
0.515625,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.515625,
0.86865234375
],
[
0.0849609375,
0.869140625,
0.483642578125,
0.869140625,
0.483642578125,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.515625,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.515625,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.27685546875,
0.88330078125,
0.27685546875,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Minimizing Hitting Time between Disparate Groups with Shortcut Edges",
"KDD'23, August 610, 2023, Long Beach, CA, UA",
"E[X] be the subset of edges of G that have both endpoints in X, and",
"BMMHaimstoaugmentGwithknewinter-groupedges, so",
"G[X] = (X, E[X]) be the subgraph of G induced by X. If F is a set",
"as to minimize the largest hitting time to a blue node, for random",
"of non-edges of G = (V,E), then G+F denotes the graph (V, EUF).",
"walks starting from red nodes. Likewise, BMAH aims to minimize",
"The degree of a node u in G, which is the number of neighbors of u",
"the average hitting time from red to blue nodes. Note that for every",
"in G, is denoted as dc(o). We let dr be the average of the degrees",
"red node we allow multiple new edges to be incident to it, as long",
"of red nodes.",
"as the augmented graph remains simple.",
"Random walks. Every random walk will be a uniform simple random",
"walk, unless stated otherwise. By uniform simple random walk we",
"4.2Observations",
"mean that at each step one of the neighbors of the current node",
"The NP-hardness of both BMMH and BMAH follows from a reduc-",
"is selected with uniform probability.Given a graph G = (V,E)",
"tion fromtheminimum set cover problem, very similar to the re-",
"and A V, the random variable tu(A) indicates the first time that",
"duction detailed by Haddadan et al. [24] and omitted for brevity. We",
"a random walk on G starting from u visits A. If u E A, we set",
"present several useful observations regarding BMMH and BMAH",
"Tu(A) = o.The expectation of this random variable is denoted as",
"that will guide the design of approximation algorithms:",
"Hc(u, A) = Ec[tu(A)], where the subscript emphasizes that the",
"random walk is on G. With a slight abuse of notation, H(u, ) is",
"OBSERVATION 1. Changing the blue endpoints of the edges in a",
"feasible solution F does not change f(F) or g(F).",
"thusthe (expected)hittingtimeforawalk startingfromutovisitu",
"for the first time. Note that Hc(u, u) = 0 and HG(u, A) = 0 if and",
"PRoor. Every walk starts from a red node and halts the moment",
"only if u E A. It also holds that Hc(u, A) ≤ Hc(u, B) when B C A,",
"a blue node is hit, hence the identity of the blue endpoints of edges",
"and in particular H(u, A) ≤ minaeA Hc(u, a).",
"in F do not matter.",
"A",
"Matrices and norms. We denote A as the adjacency matrix of a graph",
"and D is the degree diagonal matrix. Let er e Rrx1 be the vector of",
"Both problems thus reduce to selecting a multiset of red end-",
"all-ones of size r, and let Il-ll2 denote the spectral norm.",
"points, as we allow multiple edges in F to have the same red end-",
"Additional definitions.For a graph G = (V,E) witha valid bipartition",
"point. If a red node is connected to all the blue nodes already, we",
"V=(R,B),let (RxB)/Ebethe set of inter-group non-edges inG",
"can of course not add any more new edges incident to it.",
"Define the following two set functions f.g:2(XB)E R>0 as",
"OBSERVATION 2 (MONOTONICITY).IfY X, then f(X) ≤ f(Y)",
"the maximum and average hitting time.",
"and g(X) ≤ g(Y).",
"DEFINITION1(MAXIMUMANDAVERAGEHITTINGTIME).FOrQ",
"PRooF. Let F be a feasible solution. For every r e R, the hitting",
"set of inter-group non-edges F (R×B) /E, we define the maximum",
"time HG+F(r, B) is a monotonically decreasing set function of F",
"hitting time from R to B on graph G+ F by",
"since the edges in F are inter-group edges. Every e e F only helps",
"f(F) =max HG+F(r,B),",
"in reducing the hitting time from r to B. This can be proven more",
"rER",
"formally by a straightforward coupling argument between a walker",
"and the average hitting time from R to B on graph G + F by",
"on G and G+F[24, 34]. Since individual hitting times are monotone,",
"HG+F(r,B)",
"the functions f and g are monotone, as well.",
"口",
"g(F) ",
"e",
"Thefollowingsupermodularitypropertyforindividual hitting",
"4PROBLEMSANDOBSERVATIONS",
"times is less obvious, and a variant of it was proven by Haddadan",
" et al. [24] for bounded-length random walks on directed graphs",
"In this section, we formally define the problems we study, and we",
"We give a different and shorter proof of supermodularity in our",
"discuss their properties.",
"case of unbounded random walks, by using a coupling argument.",
"4.1 Problems",
"OBSERVATION 3 (SUPERMODULARITY OF HITTING TIMES). FOr all",
"We introduce and study the following two problems, which we",
"r ER and for allei + e2 E (R× B) /E it holds that",
"denote as the BMMH (Budgeted Minimum Maximum Hitting time)",
"HG+e2 (r, B) - HG+[ei,e2)(r, B) ≤ Hc(r, B) - HG+e (r, B).",
"",
"(1)",
"and BMAH (Budgeted Minimum Average Hitting time) problems.",
"Proor. Pick any r e R. We assume all walks start from r. Con-",
"PROBLEM 1(BUDGETEDMINIMUMMAXIMUMHITTING TIME",
"sider the right-hand side ofEq.(1).Couple walks on G+ei to walks",
"PROBLEM (BMMH). Given an undirected connected graph G =",
"onGbysimplyfollowingthewalkeronG,untilthewalkerhitsthe",
"(V,E),with avalid bipartitionV =R,B),and abudgetk E N,",
"red endpoint ri e R of the edge ei. If the walker hits ri, follow the",
"we seek to find a set of inter-group non-edges F C (R× B) |E with",
"ede e with probabilityWith probabilitykee",
"de(r",
"IF| ≤ k that minimizes f(F), where f is the maximum hitting-time",
"following the walker on G. It is clear that the marginal distribution",
"function defined in Definition 1.",
"of this coupled walk is a uniform random walk on G + e1.",
"PROBLEM 2 (BUDGETED MINIMUM AVERAGE HITTING TIME PROB-",
"As long as ei is not traversed, the two coupled walks are identical.",
"LEM (BMAH). Identical definition to Problem 1, but we aim to mini-",
"If ei is traversed, the walks are identical up until the time prior to",
"mize the function g, instead off, where g is the average hitting-time",
"traversing ei, when they are both in ri for the last time. On this",
"function defined in Definition 1.",
"moment, the walker on G still needs to travel Hc(ri, B) steps on"
] |
{
"arxiv_id": "2306.03571",
"language": "en",
"page_idx": 2
}
|
||
2009.09259-en-0
|
2009.09259-en-0.jpg
|
Bid Shading by Win-Rate Estimation and Surplus Maximization
Shengjun Pan, Brendan Kitts, Tian Zhou, Hao He, Bharatbhushan Shetty, Aaron Flores, Djordje
Gligorijevic, Junwei Pan, Tingyu Mao, San Gultekin and Jianlong Zhang
Verizon Media
Sunnyvale, California
$\{$alanpan,brendan.kitts,tian.zhou,hao.he,bharatbs,aaron.flores$\}@$verizonmedia.com
$\{$djordje,jwpan,tingyu.mao,sgultekin,jianlong$\}\textcircled{a}$verizonmedia.com
$\mathbf{ABSTRACT}$
This paper describes a new w (WR) that does not rely on the minimum-bid-to-win feedback from a Sell-Side Platform (SSP). The method uses a modified logistic regression to predict the profit from each possible shaded bid price. The function form allows fast maximization at run-time, a key requirement for Real-Time Bidding (RTB) systems. We report production results from this method along with several other algorithms. We found that bid shading, in general, can deliver significant value to advertisers, reducing price per impression to about 55% of the unshaded cost. Further, the particular approach described in this paper captures 7% more profit for advertisers, than do benchmark
and eCPA when the algorithm was integrated with budget controllers. We attribute the gains above as being mainly due to the explicit maximization of the surplus function, and note that other algorithms can take advantage of this same approach.
Google famously used Second Price Auctions for its Adwords and Adsense auctions, and, in 2017, generated 90% of its revenue from Second Price Auctions [12]. in online advertising between 2018 and 2019. As of 2020, almost all major display ad auctions have switched from Second to First Price Auctions (FPAs) [13, 14]. Several factors conspired to drive the industry towards the adoption of FPA, including the widespread growth of header bidding with its incompatibility with SPAs [18], increased demand for transparency and accountability [7,11,26,29], and yield concerns [3,21,25].
Unfortunately for advertisers, First Price Auctions leave private value bidders susceptible to private value of an impression was $10.00, and the winner knew the second placed bidder's price was just $1.00, they could bid just $1.01 and effectively collect a $8.99 profit. If they instead bid their private value, they would be charged the entirety of the $10.00 and they would have $0 profit!
The practice of strategically decreasing bid price below the buyer's private value is known as bid shading. Bid shading has been observed in a variety o Spectrum [6], US Oil Deposits [5], Cattle auctions [8], US Treasury auctions [17] and others. Despite its widespread use, there has been little work done on methods to systematically exploit shading, particularly when data is available to make it possible to predict auction clearing prices.
# $\csc\cos$coNCEPTS
# $\cdot$ Applied computing$\to$Onli systems$\to$Display advertising; - Computing methodologies $\to$Machine learning algorithms. $\mathbf{KEYWORDS}$
2 THE вID SHADING PROBLEM
online bidding, shading, auction, advertising, bid, optimization
ACM Reference Format:
Shengjun Pan, Brendan Kitts, Tian Zhou, Hao He, Bharatbhushan Shetty,
Given bid request $b_i$, and a valuation $V_i$, if we won the impression, which represents how much the advertiser expects to capture from the impression, how much should the advertiser discount their valuation? Assuming that the val of the dollar value that the advertiser expects to obtain, and the bid $b_{i}=g_{i}V_{i}$ is also in real dollars, the advertiser's financial gain, or surplus, is equal to:
Aaron Flores, Djordje Gligorijevic, Junwei Pan, Tingyu Mao, San Gultekin and Jianlong Zhang. 2020. Bid Shading by Win-Rate Estimation and Surplus Maximization. In $\bar{AdKDD}$ '20: SIG Conference on Knowledge Discovery and Data Mining, August 2020, San Diego, California, USA. ACM, New York, NY, USA,6 pages.
# 1 INTRODUCTION
(1)
$$surplus\stackrel{\mathrm{def}}{=}\sum_{i}(V_{i}-g_{i}V_{i})\:\mathbf{I}(g_{i}V_{i}>\hat{b}_{i}),$$
Online Advertising auctions have been dominated by Second Priced Auctions (SPAs) since their early implementations in the 1990s.
where $g_i\in(0,1]$ is the shading factor to apply to the bidder's private value $V_i,\hat{b}_i$ is the minimum bid price to win, and I$(b_i>\hat{b}_i)=1$ if the impression is won, and 0 otherwise. The task is to find a shading factor $g_i$ that maximizes the surplus to the advertiser.
Permission to make digital o classroom use is granted wit for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org $AdKDD^{\prime }20, \hat{A} ugust$ $2020, San\hat{D} iego, California, USA$ $\textcircled{\circ}$ 2020 Association for Computing Machinery. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00
# 3 РREVIOUS WORK
3.1 Bid Shading Theory
Bid shading is a common tactic in repeated First Price Auctions.[36] found robust evidence of shading in Austrian livestock auctions ,
|
[
[
0.09478759765625,
0.10479736328125,
0.904296875,
0.10479736328125,
0.904296875,
0.12249755859375,
0.09478759765625,
0.12249755859375
],
[
0.12091064453125,
0.1427001953125,
0.87841796875,
0.1427001953125,
0.87841796875,
0.1572265625,
0.12091064453125,
0.1572265625
],
[
0.2132568359375,
0.160400390625,
0.7890625,
0.160400390625,
0.7890625,
0.1749267578125,
0.2132568359375,
0.1749267578125
],
[
0.44677734375,
0.1749267578125,
0.5498046875,
0.175537109375,
0.5498046875,
0.1900634765625,
0.44677734375,
0.189453125
],
[
0.426513671875,
0.1912841796875,
0.576171875,
0.1912841796875,
0.576171875,
0.2052001953125,
0.426513671875,
0.2052001953125
],
[
0.2254638671875,
0.205810546875,
0.77392578125,
0.205810546875,
0.77392578125,
0.2203369140625,
0.2254638671875,
0.2203369140625
],
[
0.27685546875,
0.2215576171875,
0.7236328125,
0.2215576171875,
0.7236328125,
0.236083984375,
0.27685546875,
0.236083984375
],
[
0.0882568359375,
0.246826171875,
0.181396484375,
0.246826171875,
0.181396484375,
0.25830078125,
0.0882568359375,
0.25830078125
],
[
0.515625,
0.246826171875,
0.9140625,
0.24560546875,
0.9140625,
0.260009765625,
0.515625,
0.261474609375
],
[
0.08660888671875,
0.26318359375,
0.48046875,
0.26318359375,
0.48046875,
0.27783203125,
0.08660888671875,
0.27783203125
],
[
0.51708984375,
0.261474609375,
0.9140625,
0.261474609375,
0.9140625,
0.274658203125,
0.51708984375,
0.274658203125
],
[
0.51708984375,
0.274658203125,
0.9140625,
0.274658203125,
0.9140625,
0.28857421875,
0.51708984375,
0.28857421875
],
[
0.0236968994140625,
0.266357421875,
0.05963134765625,
0.266357421875,
0.05963134765625,
0.324462890625,
0.0236968994140625,
0.324462890625
],
[
0.0849609375,
0.276611328125,
0.48046875,
0.27783203125,
0.48046875,
0.291748046875,
0.0849609375,
0.290283203125
],
[
0.0849609375,
0.292236328125,
0.483642578125,
0.292236328125,
0.483642578125,
0.30615234375,
0.0849609375,
0.30615234375
],
[
0.51708984375,
0.28857421875,
0.9140625,
0.28857421875,
0.9140625,
0.3017578125,
0.51708984375,
0.3017578125
],
[
0.08660888671875,
0.30615234375,
0.48193359375,
0.30615234375,
0.48193359375,
0.3193359375,
0.08660888671875,
0.3193359375
],
[
0.5185546875,
0.302978515625,
0.91259765625,
0.302978515625,
0.91259765625,
0.31640625,
0.5185546875,
0.31640625
],
[
0.02777099609375,
0.31640625,
0.06451416015625,
0.3193359375,
0.056365966796875,
0.37255859375,
0.0196075439453125,
0.369384765625
],
[
0.0882568359375,
0.3212890625,
0.4794921875,
0.3212890625,
0.4794921875,
0.33203125,
0.0882568359375,
0.33203125
],
[
0.51953125,
0.318115234375,
0.9111328125,
0.318115234375,
0.9111328125,
0.328857421875,
0.51953125,
0.328857421875
],
[
0.0849609375,
0.333251953125,
0.4794921875,
0.333251953125,
0.4794921875,
0.3466796875,
0.0849609375,
0.3466796875
],
[
0.51708984375,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.343505859375,
0.51708984375,
0.343505859375
],
[
0.08660888671875,
0.347900390625,
0.48046875,
0.347900390625,
0.48046875,
0.361083984375,
0.08660888671875,
0.361083984375
],
[
0.51708984375,
0.343505859375,
0.9140625,
0.343505859375,
0.9140625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.08660888671875,
0.363037109375,
0.4794921875,
0.363037109375,
0.4794921875,
0.373779296875,
0.08660888671875,
0.373779296875
],
[
0.51708984375,
0.35791015625,
0.9140625,
0.35791015625,
0.9140625,
0.371826171875,
0.51708984375,
0.371826171875
],
[
0.0269622802734375,
0.365478515625,
0.0555419921875,
0.365478515625,
0.0555419921875,
0.39453125,
0.0269622802734375,
0.39453125
],
[
0.0849609375,
0.375732421875,
0.48193359375,
0.375732421875,
0.48193359375,
0.389404296875,
0.0849609375,
0.389404296875
],
[
0.51708984375,
0.371826171875,
0.69921875,
0.371826171875,
0.69921875,
0.385009765625,
0.51708984375,
0.385009765625
],
[
0.08660888671875,
0.38818359375,
0.4794921875,
0.38818359375,
0.4794921875,
0.401611328125,
0.08660888671875,
0.401611328125
],
[
0.53173828125,
0.385009765625,
0.9140625,
0.385009765625,
0.9140625,
0.3984375,
0.53173828125,
0.3984375
],
[
0.08660888671875,
0.40283203125,
0.4794921875,
0.401611328125,
0.4794921875,
0.415283203125,
0.08660888671875,
0.416748046875
],
[
0.51708984375,
0.399658203125,
0.9140625,
0.399658203125,
0.9140625,
0.41357421875,
0.51708984375,
0.41357421875
],
[
0.515625,
0.41357421875,
0.91259765625,
0.412353515625,
0.91259765625,
0.425537109375,
0.515625,
0.4267578125
],
[
0.0228729248046875,
0.40478515625,
0.058013916015625,
0.404052734375,
0.062103271484375,
0.484130859375,
0.0269622802734375,
0.48486328125
],
[
0.0849609375,
0.416748046875,
0.48193359375,
0.416748046875,
0.48193359375,
0.429931640625,
0.0849609375,
0.429931640625
],
[
0.0849609375,
0.429931640625,
0.48193359375,
0.4287109375,
0.48193359375,
0.442626953125,
0.0849609375,
0.443115234375
],
[
0.51708984375,
0.4267578125,
0.9140625,
0.4267578125,
0.9140625,
0.43994140625,
0.51708984375,
0.43994140625
],
[
0.0849609375,
0.4443359375,
0.48046875,
0.443115234375,
0.48046875,
0.45703125,
0.0849609375,
0.458251953125
],
[
0.51708984375,
0.44140625,
0.9140625,
0.44140625,
0.9140625,
0.455078125,
0.51708984375,
0.455078125
],
[
0.0849609375,
0.458251953125,
0.483642578125,
0.458251953125,
0.483642578125,
0.4716796875,
0.0849609375,
0.4716796875
],
[
0.51708984375,
0.455078125,
0.9140625,
0.455078125,
0.9140625,
0.468505859375,
0.51708984375,
0.468505859375
],
[
0.0849609375,
0.4716796875,
0.48193359375,
0.4716796875,
0.48193359375,
0.48486328125,
0.0849609375,
0.48486328125
],
[
0.51708984375,
0.468505859375,
0.6787109375,
0.468505859375,
0.6787109375,
0.481689453125,
0.51708984375,
0.481689453125
],
[
0.08660888671875,
0.486083984375,
0.48046875,
0.486083984375,
0.48046875,
0.5,
0.08660888671875,
0.5
],
[
0.53173828125,
0.48291015625,
0.9140625,
0.48291015625,
0.9140625,
0.496826171875,
0.53173828125,
0.496826171875
],
[
0.0196075439453125,
0.494384765625,
0.058013916015625,
0.494384765625,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.0849609375,
0.50048828125,
0.41015625,
0.50048828125,
0.41015625,
0.5146484375,
0.0849609375,
0.5146484375
],
[
0.5185546875,
0.49755859375,
0.91259765625,
0.49755859375,
0.91259765625,
0.50830078125,
0.5185546875,
0.50830078125
],
[
0.5185546875,
0.51025390625,
0.9140625,
0.51025390625,
0.9140625,
0.5234375,
0.5185546875,
0.5234375
],
[
0.0882568359375,
0.5244140625,
0.221435546875,
0.5244140625,
0.221435546875,
0.5390625,
0.0882568359375,
0.5390625
],
[
0.51708984375,
0.5234375,
0.9140625,
0.5234375,
0.9140625,
0.537109375,
0.51708984375,
0.537109375
],
[
0.101318359375,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.556640625,
0.101318359375,
0.556640625
],
[
0.51953125,
0.5390625,
0.91259765625,
0.5390625,
0.91259765625,
0.5498046875,
0.51953125,
0.5498046875
],
[
0.51708984375,
0.5517578125,
0.9140625,
0.5517578125,
0.9140625,
0.56494140625,
0.51708984375,
0.56494140625
],
[
0.0849609375,
0.55810546875,
0.48046875,
0.55810546875,
0.48046875,
0.5712890625,
0.0849609375,
0.5712890625
],
[
0.0882568359375,
0.5712890625,
0.30810546875,
0.5712890625,
0.30810546875,
0.5859375,
0.0882568359375,
0.5859375
],
[
0.515625,
0.56640625,
0.9140625,
0.564453125,
0.9140625,
0.57763671875,
0.515625,
0.580078125
],
[
0.51708984375,
0.580078125,
0.6611328125,
0.580078125,
0.6611328125,
0.59326171875,
0.51708984375,
0.59326171875
],
[
0.0882568359375,
0.58837890625,
0.188720703125,
0.58837890625,
0.188720703125,
0.599609375,
0.0882568359375,
0.599609375
],
[
0.10296630859375,
0.60595703125,
0.4794921875,
0.60595703125,
0.4794921875,
0.619140625,
0.10296630859375,
0.619140625
],
[
0.51953125,
0.609375,
0.53369140625,
0.609375,
0.53369140625,
0.6171875,
0.51953125,
0.6171875
],
[
0.5390625,
0.60986328125,
0.54345703125,
0.60986328125,
0.54345703125,
0.61572265625,
0.5390625,
0.61572265625
],
[
0.54150390625,
0.60595703125,
0.79736328125,
0.60595703125,
0.79736328125,
0.619140625,
0.54150390625,
0.619140625
],
[
0.08660888671875,
0.62353515625,
0.2296142578125,
0.62353515625,
0.2296142578125,
0.63720703125,
0.08660888671875,
0.63720703125
],
[
0.5185546875,
0.62548828125,
0.91259765625,
0.62548828125,
0.91259765625,
0.63623046875,
0.5185546875,
0.63623046875
],
[
0.0849609375,
0.63720703125,
0.48046875,
0.63720703125,
0.48046875,
0.65087890625,
0.0849609375,
0.65087890625
],
[
0.51708984375,
0.63818359375,
0.9140625,
0.63818359375,
0.9140625,
0.6513671875,
0.51708984375,
0.6513671875
],
[
0.08660888671875,
0.65087890625,
0.48046875,
0.65087890625,
0.48046875,
0.6611328125,
0.08660888671875,
0.6611328125
],
[
0.51708984375,
0.6513671875,
0.91259765625,
0.6513671875,
0.91259765625,
0.66552734375,
0.51708984375,
0.66552734375
],
[
0.08660888671875,
0.66357421875,
0.48046875,
0.66357421875,
0.48046875,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.5185546875,
0.66650390625,
0.91259765625,
0.66650390625,
0.91259765625,
0.6767578125,
0.5185546875,
0.6767578125
],
[
0.0849609375,
0.673828125,
0.48193359375,
0.673828125,
0.48193359375,
0.6875,
0.0849609375,
0.6875
],
[
0.5185546875,
0.68115234375,
0.91259765625,
0.68115234375,
0.91259765625,
0.69140625,
0.5185546875,
0.69140625
],
[
0.0849609375,
0.6875,
0.48193359375,
0.6875,
0.48193359375,
0.70068359375,
0.0849609375,
0.70068359375
],
[
0.08251953125,
0.69873046875,
0.160888671875,
0.70068359375,
0.160888671875,
0.7138671875,
0.08251953125,
0.7119140625
],
[
0.51708984375,
0.693359375,
0.9140625,
0.693359375,
0.9140625,
0.70703125,
0.51708984375,
0.70703125
],
[
0.51708984375,
0.70703125,
0.6357421875,
0.70703125,
0.6357421875,
0.72021484375,
0.51708984375,
0.72021484375
],
[
0.08660888671875,
0.72607421875,
0.256591796875,
0.72607421875,
0.256591796875,
0.7392578125,
0.08660888671875,
0.7392578125
],
[
0.6005859375,
0.73095703125,
0.66845703125,
0.72802734375,
0.6689453125,
0.74365234375,
0.60107421875,
0.7470703125
],
[
0.6845703125,
0.73193359375,
0.82861328125,
0.72998046875,
0.82861328125,
0.744140625,
0.6845703125,
0.7470703125
],
[
0.89453125,
0.73193359375,
0.9150390625,
0.73193359375,
0.9150390625,
0.7470703125,
0.89453125,
0.7470703125
],
[
0.0849609375,
0.744140625,
0.48193359375,
0.744140625,
0.48193359375,
0.75830078125,
0.0849609375,
0.75830078125
],
[
0.08660888671875,
0.76025390625,
0.48193359375,
0.76025390625,
0.48193359375,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.51708984375,
0.76220703125,
0.91259765625,
0.76220703125,
0.91259765625,
0.77587890625,
0.51708984375,
0.77587890625
],
[
0.08660888671875,
0.7822265625,
0.48193359375,
0.7822265625,
0.48193359375,
0.79248046875,
0.08660888671875,
0.79248046875
],
[
0.51708984375,
0.7763671875,
0.9140625,
0.7763671875,
0.9140625,
0.79052734375,
0.51708984375,
0.79052734375
],
[
0.131591796875,
0.79248046875,
0.48046875,
0.79248046875,
0.48046875,
0.80322265625,
0.131591796875,
0.80322265625
],
[
0.51708984375,
0.79052734375,
0.9140625,
0.79052734375,
0.9140625,
0.8037109375,
0.51708984375,
0.8037109375
],
[
0.0849609375,
0.80322265625,
0.48193359375,
0.80322265625,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.08905029296875,
0.79541015625,
0.1282958984375,
0.79541015625,
0.1282958984375,
0.7998046875,
0.08905029296875,
0.7998046875
],
[
0.51708984375,
0.8046875,
0.8408203125,
0.8046875,
0.8408203125,
0.818359375,
0.51708984375,
0.818359375
],
[
0.09722900390625,
0.8125,
0.2459716796875,
0.81103515625,
0.2459716796875,
0.82275390625,
0.09722900390625,
0.82373046875
],
[
0.2459716796875,
0.8125,
0.48046875,
0.8125,
0.48046875,
0.82275390625,
0.2459716796875,
0.82275390625
],
[
0.08416748046875,
0.8212890625,
0.483642578125,
0.8212890625,
0.483642578125,
0.83544921875,
0.08416748046875,
0.83544921875
],
[
0.08905029296875,
0.81640625,
0.101318359375,
0.81640625,
0.101318359375,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.08660888671875,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.84326171875,
0.08660888671875,
0.84326171875
],
[
0.5185546875,
0.833984375,
0.5361328125,
0.833984375,
0.5361328125,
0.84228515625,
0.5185546875,
0.84228515625
],
[
0.53515625,
0.83544921875,
0.54345703125,
0.83544921875,
0.54345703125,
0.84033203125,
0.53515625,
0.84033203125
],
[
0.5400390625,
0.83056640625,
0.69775390625,
0.83056640625,
0.69775390625,
0.8447265625,
0.5400390625,
0.8447265625
],
[
0.08416748046875,
0.84228515625,
0.3349609375,
0.84326171875,
0.3349609375,
0.85400390625,
0.08416748046875,
0.85302734375
],
[
0.0849609375,
0.8515625,
0.330810546875,
0.8515625,
0.330810546875,
0.86572265625,
0.0849609375,
0.86572265625
],
[
0.51708984375,
0.85107421875,
0.7197265625,
0.85107421875,
0.7197265625,
0.8642578125,
0.51708984375,
0.8642578125
],
[
0.0849609375,
0.86572265625,
0.2998046875,
0.86572265625,
0.2998046875,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.51708984375,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.0849609375,
0.875,
0.3095703125,
0.875,
0.3095703125,
0.88525390625,
0.0849609375,
0.88525390625
],
[
0.51708984375,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"Bid Shading byWin-Rate Estimation and Surplus Maximization",
"Shengjun Pan, Brendan Kitts, Tian Zhou, Hao He, Bharatbhushan Shetty, Aaron Flores, Djordje",
"Gligorijevic, Junwei Pan, Tingyu Mao, San Gultekin and Jianlong Zhang",
"Verizon Media",
"Sunnyvale, California",
"[alanpan,brendan.kitts,tian.zhou,hao.he,bharatbs,aaron.flores)@verizonmedia.com",
"[djordje,jwpan,tingyu.mao,sgultekin,jianlong)@verizonmedia.com",
"ABSTRACT",
"Google famously used Second Price Auctions for its Adwords and",
"This paper describes a new win-rate based bid shading algorithm",
"Adsense auctions, and, in 2017, generated 90% of its revenue from",
"Second Price Auctions [12]. However, there was a dramatic shift",
"0",
"(WR)that does not rely on the minimum-bid-to-win feedbackfrom a",
"Sell-Side Platform (SSP). The method uses a modified logistic regres-",
"in online advertising between 2018 and 2019. As of 2020, almost",
"sion to predict the profit from each possible shaded bid price. The",
"all major display ad auctions have switched from Second to First",
"des",
"functionform allows fast maximization at run-time,akeyrequire",
"PriceAuctions (FPAs)[13, 14].Several factors conspired to drive",
"ment for Real-Time Bidding (RTB) systems. We report production",
"the industry towards the adoption of FPA, including the widespread",
"results from this method along with several other algorithms. We",
"growth of header bidding with its incompatibility with SPAs [18].",
"found that bid shading, in general, can deliver significant value",
"increased demand for transparency and accountability [7, 11, 26, 29]",
"9",
"to advertisers, reducing price per impression to about 55% of the",
"and yield concerns [3, 21, 25].",
"unshaded cost. Further, the particular approach described in this",
"Unfortunately for advertisers, First Price Auctions leave private",
"paper captures 7% more profit for advertisers, than do benchmark",
"value bidders susceptible to over-paying. For instance, if the bidder's",
"private value of an impression was $10.00, and the winner knew",
"['s]",
"methods of just bidding the most probable winning price. We also",
"report 4.3%higher surplus than an industry Sell-Side Platform shad-",
"the second placed bidder's price was just s1.o0, they could bid just",
"ingservice.Furthermore, weobserved3%-7%lower eCPM,eCPC",
"$1.01 and effectively collect a $8.99 profit. If they instead bid their",
"and eCPA when the algorithm was integrated with budget con-",
"private value, they would be charged the entirety of the $10.00 and",
"trollers. We attribute the gains above as being mainly due to the",
"they would have so profit!",
"explicit maximization of the surplus function, and note that other",
"The practice of strategically decreasing bid price below the",
"602606002:1",
"algorithms can take advantage of this same approach.",
"buyer's private value is known as bid shading.Bid shading has",
"been observed in a variety of real world auctions including FCC",
"CCS CONCEPTS",
"Spectrum [6], US Oil Deposits [5], Cattle auctions [8], US Trea-",
". Applied computing Online auctions; . Information",
"sury auctions[17] and others.Despite its widespread use,there has",
"been little work done on methods to systematically exploit shading.",
"systems-Display advertising:.Computing methodologies",
" Machine learning algorithms.",
"particularly when data is available to make it possible to predict",
"auction clearing prices.",
"KEYWORDS",
"online bidding, shading, auction, advertising, bid, optimization",
"2",
"",
"THEBIDSHADINGPROBLEM",
"ACM Reference Format:",
"Given bid request bi,and a valuation Vi,if we won the impression.",
"Shengjun Pan, Brendan Kitts, Tian Zhou, Hao He, Bharatbhushan Shetty",
"which represents how much the advertiser expects to capture from",
"AaronFlores,DjordjeGligorijevic,JunweiPan,TingyuMao,SanGultekin",
"the impression, how much should the advertiser discount their valu-",
"and Jianlong Zhang.2020.Bid Shading by Win-Rate Estimation and Surplus",
"ation?Assuming that the valuation V,is an accuraterepresentation",
"Maximization. In AdKDD '20: SIG Conference on Knowledge Discovery and",
"of thedollarvaluethat the advertiserexpects to obtain, and thebid",
"Data Mining, August 2020, San Diego, California, USA.ACM, New York, NY",
"USA, 6 pages.",
"bi = giV, is also in real dollars, the advertiser's financial gain, or",
"surplus, is equal to:",
"1INTRODUCTION",
"surplus def",
"(V; - giV) I(giV,>bi),",
"(1)",
"Online Advertising auctions have been dominated by Second Priced",
"Auctions(SPAs)sincetheir earlyimplementations inthe 1990s.",
"where gi e (o, 1) is the shading factor to apply to the bidder's private",
"Permission tomake digital or hard copies of all orpart of this work forpersonal or",
"value Vi, b; is the minimum bid price to win, and I(b; > bi) = 1 if",
" use is granted without fee provided that copies are not made or distributed",
"the impression is won, and 0 otherwise. The task is to find a shading",
"for profitor commercial advantage and thatcopies bear this notice and thefull citation",
"sassroor",
"factor gi that maximizes the surplus to the advertiser.",
"the first page.Copyrightsfor co",
"mponents of this work owned by others than ACM",
"must be honored. Abstracting with credit is permitted To copy otherwise,or republish,",
"2",
"to post on servers or to redistribute to lists, requires prior specific permission and/or a",
"3",
"",
"PREVIOUS WORK",
"fee.Request permissions from permissions@acm.org",
"AdKDD 20, August 2020, San Diego, California, USA",
"3.1 Bid ShadingTheory",
"@2020Association for Computing Machinery",
"Bid shading is a common tactic in repeated First Price Auctions. [36]",
"ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00",
"found robust evidence of shading in Austrian livestock auctions ,"
] |
{
"arxiv_id": "2009.09259",
"language": "en",
"page_idx": 0
}
|
||
2009.09259-en-1
|
2009.09259-en-1.jpg
|
AdKDD'20, August 2020, San Diego, California, USA
S Pan, B Kitts, T Zhou, H He, B Shetty, A Flores, D Gligorijevic, J Pan, T Mao, S Gultekin and J Zhang
[8] reported shading in a Texas cattle market, and [17] found the
practice in auctions for US Treasury notes.
Auctions generally need to be repeated and predictable for bid shading to be practically feasible, but under these conditions, it often occurs organically. Pownall and Wolk (2013) showed that bid shading for repeated internet auction prices increased over time; by about 26% after 10 iterations [24]. When there are enough repeated games bidders can even develop collusive shading strategies where bidders actively coordinate to have low bids [16,23].
Although behavior varies from auction to auction, several studies have shown that the magnitude of shading tends to increase with the average price on the auction [4,6,17]. This is likely to occur because of the more substantial losses involved on higher priced auctions, if shading isn't sufficient. This result suggests that using a measure of the expense of the auction is valuable when trying to estimate the shading factor- a finding we revisit later in Section 6.
In situations where the supply is plentiful, and demand limited, buyers can shade deeper. In looking at this phenomemon in the US Treasury Market, Hortacsu et. al. (2017) find that large institutional buyers on average shade more aggressively than small indirect buyers [17] . This seems to be because these large buyers effectively control a large percentage of bidders, and so it is almost like they are able to coordinate the buying of multiple buyers. They can therefore drive the bid prices for a large percentage of bidders down, whilst still meeting their goals.
## 3.2 Previous Algorithms
$\operatorname{In}2018$ and 2019, Rubicon [1,25],AppNexus [3] and Google [14,15, 28,30] all released Sell-Side bid shading services. Leading up to this, there had been reports of dramatically lower ROI from the new First Price Auctions [18,21]. Never-the-less, this is a surprising move as Sell-Side Platforms are potentially decreasing their yield, and they clearly have a different incentive from buyers. The sell-side algorithms seem to reflect this incentive difference. The descriptions of these services suggest that they try to keep bid prices high enough to maintain a set win-rate, but preventing the bid price from becoming too extreme; which might risk an advertiser to halt their bidding due to poor Return on Investment. Rubicon released data suggesting that their service decreases First Price CPMs by a modest 5% over 4 months [25]. AppNexus reported that prices under their service were 25% lower over 100 days [3]. We tried one of the services, and recorded the shading distribution in Figure 2. Most of the bid shades were about 90%, which is conservative for our problem. Further analysis on Sell-Side "Bid Shaders" are in Section 7.
On the Demand Side, a variety of algorithms have been explored, although generally not exactly for bid shading applications.[35] developed a censored winning bid probability estimator. They observed that when a bidder submitted a bid and lost, the information gained is that the winning price is somewhere above the submitted price, and when a bidder submits a bid and wins, the minimum bid to win is at a price somewhere below their submitted bid. Using these two cases, the authors developed a Maximum Likelihood procedure to estimate the probability of the winning bid being any of the bid prices. This created a distribution of the probable winning bids, with the most likely winning price being used for bidding.
[34] extended their work to using a neural network to estimate the
parameters of the win probability distribution.
An unpublished implementation [19] used Logistic Regression to predict the optimal bid shading factor using features in the request. The predicted factor was then used as a multiplier on the unshaded $bidprice.$
The approaches described above [19,34,35] all focus on predicting the probable winning bid price. However, the surplus maximum is very different from the minimum bid to win. An accurate (unbiased, symmetric noise) win probability estimator will be below the winning bid price about 50% of the time- this means that 50% of the surplus won't be captured by design. If the change in new impressions captured at a higher bid price, over-weights the marginal decrease in profitability per impression, the optimum for surplus can be higher than the most probable bid.
Unpublished work [20] is one of the few that we know of to attempt to explicitly maximize the surplus function. These authors estimate shading factors for a set of fixed segments based on three bid samples taken in real-time to estimate the local surplus landscape. However the approach has many drawbacks: the segments have to be predetermined and finding a suitable segment definition requires substantial analysis. The information across segments is not shared, which is a problem for segments that do not have enough traffic. Further, the set of possible segments quickly explode as the number of variables used to define them increases. The approach taken in this paper uses a model to estimate the surplus function, and so a very large number of features can be used, and model induction is also automated, easy to maintain, and improve. In order to compare the method we used to prior work, we have included an implementation of the Logistic Regression algorithm from [19],the Distribution Estimator algorithm from [35],and the Segment-based Surplus maximizer [20] in the benchmarks which we use to analyze algorithm performance in Section 7.
## 4 САNONICAL ALGORITHM
Given a bid request for First Price Auction, let $x_1,x_2,\ldots,x_k$ be the set of publisher and user attributes that we will use to find the best bid price $b^*.$ Let $\hat{b}$ be the highest bid price from other competing $\dot{\text{bidders, which value is unknown. Note that }\hat{b}}$ depends on both attributes $x_i$s which represent the item that is being auctioned, and external competing bidder behavior. $\hat{b}$ follows an unknown $\operatorname*{distribution}\mathcal{D}_{\hat{b}|x_{1},x_{2},\ldots,x_{k}}$with cumulative probability distributiondistribution $\begin{array}{l}{\mathrm{cdf}_{\hat{b}|x_{1},x_{2},\ldots,x_{k}}.\text{When the context is clear,we use }\mathcal{D}_{\hat{b}}\mathrm{~and~cdf}_{\hat{b}}\mathrm{~for}}\\{\mathrm{simplicity.}}\end{array}$
If the distribution $\mathcal{D}_{\hat{b}}$ is known, we can calculate the optimal bid price $b^*$ directly as follows. Let I$(b>\hat{b})$ be 1 if $b>\hat{b}$ and 0 otherwise, which indicates if the submitted price $b$ wins the auction. Then the surplus when the submitted price is $b$ would be
$$surplus=(V-b)\mathbf{I}(b>\hat{b})=\begin{cases}V-b,&\text{if}\:b>\hat{b},\\0,&\text{otherwise}\end{cases}$$
(2)
The optimal bid price can be calculated as the price that maximizes
the expected surplus
|
[
[
0.08416748046875,
0.0738525390625,
0.3349609375,
0.07452392578125,
0.3349609375,
0.08837890625,
0.08416748046875,
0.08709716796875
],
[
0.438720703125,
0.07574462890625,
0.9111328125,
0.07574462890625,
0.9111328125,
0.08587646484375,
0.438720703125,
0.08587646484375
],
[
0.08660888671875,
0.106689453125,
0.48046875,
0.106689453125,
0.48046875,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.5185546875,
0.106689453125,
0.91259765625,
0.106689453125,
0.91259765625,
0.12060546875,
0.5185546875,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.3447265625,
0.121826171875,
0.3447265625,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.79736328125,
0.121826171875,
0.79736328125,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.10296630859375,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.10296630859375,
0.1483154296875
],
[
0.53369140625,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.53369140625,
0.1483154296875
],
[
0.08660888671875,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.08660888671875,
0.1634521484375
],
[
0.51708984375,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.0849609375,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.0849609375,
0.1767578125
],
[
0.5185546875,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.5185546875,
0.1767578125
],
[
0.08660888671875,
0.1787109375,
0.48046875,
0.1787109375,
0.48046875,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.5185546875,
0.177978515625,
0.57275390625,
0.177978515625,
0.57275390625,
0.189453125,
0.5185546875,
0.189453125
],
[
0.0849609375,
0.1912841796875,
0.48193359375,
0.1912841796875,
0.48193359375,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.53125,
0.1900634765625,
0.91259765625,
0.189453125,
0.91259765625,
0.203857421875,
0.53125,
0.2052001953125
],
[
0.0849609375,
0.203857421875,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2183837890625,
0.0849609375,
0.2171630859375
],
[
0.51708984375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51708984375,
0.2183837890625
],
[
0.0849609375,
0.2183837890625,
0.400390625,
0.2183837890625,
0.400390625,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.51708984375,
0.2183837890625,
0.9140625,
0.2183837890625,
0.9140625,
0.231689453125,
0.51708984375,
0.231689453125
],
[
0.10296630859375,
0.23291015625,
0.48046875,
0.23291015625,
0.48046875,
0.246826171875,
0.10296630859375,
0.246826171875
],
[
0.515625,
0.23291015625,
0.9140625,
0.231689453125,
0.9140625,
0.24560546875,
0.515625,
0.246826171875
],
[
0.08660888671875,
0.2474365234375,
0.48046875,
0.2474365234375,
0.48046875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.08660888671875,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.08660888671875,
0.2734375
],
[
0.51708984375,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.08660888671875,
0.27587890625,
0.48046875,
0.27587890625,
0.48046875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.51708984375,
0.2734375,
0.9140625,
0.2734375,
0.9140625,
0.287353515625,
0.51708984375,
0.287353515625
],
[
0.08660888671875,
0.28857421875,
0.4794921875,
0.28857421875,
0.4794921875,
0.298583984375,
0.08660888671875,
0.298583984375
],
[
0.51953125,
0.2890625,
0.91259765625,
0.2890625,
0.91259765625,
0.2998046875,
0.51953125,
0.2998046875
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.515625,
0.300537109375,
0.7685546875,
0.2998046875,
0.7685546875,
0.314453125,
0.515625,
0.31494140625
],
[
0.0849609375,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.0849609375,
0.32763671875
],
[
0.53173828125,
0.31494140625,
0.9140625,
0.31494140625,
0.9140625,
0.328857421875,
0.53173828125,
0.328857421875
],
[
0.10296630859375,
0.330810546875,
0.48193359375,
0.330810546875,
0.48193359375,
0.341552734375,
0.10296630859375,
0.341552734375
],
[
0.5185546875,
0.330810546875,
0.91259765625,
0.330810546875,
0.91259765625,
0.341552734375,
0.5185546875,
0.341552734375
],
[
0.0849609375,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.51708984375,
0.343505859375,
0.9140625,
0.343505859375,
0.9140625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.08660888671875,
0.356689453125,
0.48193359375,
0.356689453125,
0.48193359375,
0.37060546875,
0.08660888671875,
0.37060546875
],
[
0.51708984375,
0.356689453125,
0.9140625,
0.356689453125,
0.9140625,
0.37060546875,
0.51708984375,
0.37060546875
],
[
0.08660888671875,
0.371826171875,
0.48193359375,
0.371826171875,
0.48193359375,
0.385009765625,
0.08660888671875,
0.385009765625
],
[
0.5185546875,
0.37255859375,
0.91259765625,
0.37255859375,
0.91259765625,
0.38330078125,
0.5185546875,
0.38330078125
],
[
0.0849609375,
0.385009765625,
0.48046875,
0.385009765625,
0.48046875,
0.3984375,
0.0849609375,
0.3984375
],
[
0.51708984375,
0.385009765625,
0.9140625,
0.385009765625,
0.9140625,
0.3984375,
0.51708984375,
0.3984375
],
[
0.08660888671875,
0.40087890625,
0.4794921875,
0.40087890625,
0.4794921875,
0.410888671875,
0.08660888671875,
0.410888671875
],
[
0.5185546875,
0.40087890625,
0.9111328125,
0.40087890625,
0.9111328125,
0.410888671875,
0.5185546875,
0.410888671875
],
[
0.0849609375,
0.412353515625,
0.48046875,
0.412353515625,
0.48046875,
0.425537109375,
0.0849609375,
0.425537109375
],
[
0.5185546875,
0.41357421875,
0.91259765625,
0.41357421875,
0.91259765625,
0.423583984375,
0.5185546875,
0.423583984375
],
[
0.0849609375,
0.425537109375,
0.48193359375,
0.425537109375,
0.48193359375,
0.43994140625,
0.0849609375,
0.43994140625
],
[
0.51708984375,
0.425537109375,
0.9140625,
0.425537109375,
0.9140625,
0.439453125,
0.51708984375,
0.439453125
],
[
0.0849609375,
0.43994140625,
0.2320556640625,
0.43994140625,
0.2320556640625,
0.453857421875,
0.0849609375,
0.453857421875
],
[
0.51953125,
0.44140625,
0.91259765625,
0.44140625,
0.91259765625,
0.451904296875,
0.51953125,
0.451904296875
],
[
0.51708984375,
0.453857421875,
0.9140625,
0.453857421875,
0.9140625,
0.46728515625,
0.51708984375,
0.46728515625
],
[
0.51708984375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.0849609375,
0.4765625,
0.29736328125,
0.4765625,
0.29736328125,
0.490478515625,
0.0849609375,
0.490478515625
],
[
0.5185546875,
0.48291015625,
0.91259765625,
0.48291015625,
0.91259765625,
0.49365234375,
0.5185546875,
0.49365234375
],
[
0.0849609375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.53369140625,
0.49560546875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.53369140625,
0.5087890625
],
[
0.0849609375,
0.5087890625,
0.48046875,
0.51025390625,
0.48046875,
0.521484375,
0.0849609375,
0.52099609375
],
[
0.51708984375,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51708984375,
0.52294921875
],
[
0.0849609375,
0.52294921875,
0.48193359375,
0.52294921875,
0.48193359375,
0.5361328125,
0.0849609375,
0.5361328125
],
[
0.5185546875,
0.5244140625,
0.9111328125,
0.5244140625,
0.9111328125,
0.53515625,
0.5185546875,
0.53515625
],
[
0.0849609375,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.51708984375,
0.537109375,
0.91259765625,
0.537109375,
0.91259765625,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.0849609375,
0.55029296875,
0.483642578125,
0.55029296875,
0.483642578125,
0.564453125,
0.0849609375,
0.564453125
],
[
0.51708984375,
0.55029296875,
0.84619140625,
0.55029296875,
0.84619140625,
0.564453125,
0.51708984375,
0.564453125
],
[
0.0849609375,
0.564453125,
0.48193359375,
0.564453125,
0.48193359375,
0.57763671875,
0.0849609375,
0.57763671875
],
[
0.08660888671875,
0.580078125,
0.48046875,
0.580078125,
0.48046875,
0.59033203125,
0.08660888671875,
0.59033203125
],
[
0.52294921875,
0.5888671875,
0.53173828125,
0.5888671875,
0.53173828125,
0.5947265625,
0.52294921875,
0.5947265625
],
[
0.5361328125,
0.5888671875,
0.54345703125,
0.5888671875,
0.54345703125,
0.5947265625,
0.5361328125,
0.5947265625
],
[
0.5390625,
0.58544921875,
0.771484375,
0.58544921875,
0.771484375,
0.5986328125,
0.5390625,
0.5986328125
],
[
0.08660888671875,
0.59326171875,
0.4794921875,
0.59326171875,
0.4794921875,
0.60400390625,
0.08660888671875,
0.60400390625
],
[
0.51708984375,
0.6015625,
0.8388671875,
0.60302734375,
0.8388671875,
0.6162109375,
0.51708984375,
0.61572265625
],
[
0.85693359375,
0.60400390625,
0.9111328125,
0.6015625,
0.9111328125,
0.61279296875,
0.85693359375,
0.61572265625
],
[
0.5185546875,
0.6171875,
0.91259765625,
0.6171875,
0.91259765625,
0.630859375,
0.5185546875,
0.630859375
],
[
0.08660888671875,
0.60595703125,
0.48046875,
0.60595703125,
0.48046875,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.08660888671875,
0.62060546875,
0.4794921875,
0.62060546875,
0.4794921875,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.08660888671875,
0.63525390625,
0.48046875,
0.63525390625,
0.48046875,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.51708984375,
0.630859375,
0.91259765625,
0.6318359375,
0.91259765625,
0.64599609375,
0.51708984375,
0.64453125
],
[
0.0849609375,
0.64794921875,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.5185546875,
0.646484375,
0.91259765625,
0.646484375,
0.91259765625,
0.66015625,
0.5185546875,
0.66015625
],
[
0.0849609375,
0.66015625,
0.4794921875,
0.66015625,
0.4794921875,
0.673828125,
0.0849609375,
0.673828125
],
[
0.5185546875,
0.66015625,
0.9140625,
0.66015625,
0.9140625,
0.673828125,
0.5185546875,
0.673828125
],
[
0.0849609375,
0.6748046875,
0.48046875,
0.6748046875,
0.48046875,
0.68798828125,
0.0849609375,
0.68798828125
],
[
0.5185546875,
0.6748046875,
0.9140625,
0.6748046875,
0.9140625,
0.68798828125,
0.5185546875,
0.68798828125
],
[
0.0849609375,
0.68798828125,
0.48193359375,
0.68798828125,
0.48193359375,
0.7021484375,
0.0849609375,
0.7021484375
],
[
0.51708984375,
0.6845703125,
0.6435546875,
0.68994140625,
0.64208984375,
0.70849609375,
0.515625,
0.7021484375
],
[
0.67333984375,
0.689453125,
0.91259765625,
0.689453125,
0.91259765625,
0.70263671875,
0.67333984375,
0.70263671875
],
[
0.0849609375,
0.70263671875,
0.48046875,
0.70263671875,
0.48046875,
0.71630859375,
0.0849609375,
0.71630859375
],
[
0.515625,
0.7021484375,
0.61376953125,
0.71044921875,
0.611328125,
0.72900390625,
0.51318359375,
0.7197265625
],
[
0.611328125,
0.70263671875,
0.91259765625,
0.70263671875,
0.91259765625,
0.71728515625,
0.611328125,
0.71728515625
],
[
0.0849609375,
0.71630859375,
0.48193359375,
0.71630859375,
0.48193359375,
0.72998046875,
0.0849609375,
0.72998046875
],
[
0.5185546875,
0.7216796875,
0.57861328125,
0.7216796875,
0.57861328125,
0.73291015625,
0.5185546875,
0.73291015625
],
[
0.0849609375,
0.73193359375,
0.1478271484375,
0.73193359375,
0.1478271484375,
0.74365234375,
0.0849609375,
0.74365234375
],
[
0.53173828125,
0.73291015625,
0.91259765625,
0.73291015625,
0.91259765625,
0.7470703125,
0.53173828125,
0.7470703125
],
[
0.101318359375,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.101318359375,
0.75830078125
],
[
0.5185546875,
0.751953125,
0.91259765625,
0.751953125,
0.91259765625,
0.76220703125,
0.5185546875,
0.76220703125
],
[
0.0849609375,
0.75830078125,
0.48046875,
0.75830078125,
0.48046875,
0.771484375,
0.0849609375,
0.771484375
],
[
0.0849609375,
0.771484375,
0.48046875,
0.771484375,
0.48046875,
0.78515625,
0.0849609375,
0.78515625
],
[
0.51953125,
0.7666015625,
0.91259765625,
0.7666015625,
0.91259765625,
0.7763671875,
0.51953125,
0.7763671875
],
[
0.5185546875,
0.77880859375,
0.86279296875,
0.77880859375,
0.86279296875,
0.79248046875,
0.5185546875,
0.79248046875
],
[
0.08660888671875,
0.787109375,
0.4794921875,
0.787109375,
0.4794921875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.08416748046875,
0.7998046875,
0.48193359375,
0.798828125,
0.48193359375,
0.8125,
0.08416748046875,
0.81298828125
],
[
0.0849609375,
0.81298828125,
0.48046875,
0.8125,
0.48046875,
0.82568359375,
0.0849609375,
0.8271484375
],
[
0.73876953125,
0.814453125,
0.8486328125,
0.814453125,
0.8486328125,
0.8291015625,
0.73876953125,
0.8291015625
],
[
0.56640625,
0.82275390625,
0.72802734375,
0.82275390625,
0.72802734375,
0.8369140625,
0.56640625,
0.8369140625
],
[
0.89453125,
0.82373046875,
0.9140625,
0.82373046875,
0.9140625,
0.83837890625,
0.89453125,
0.83837890625
],
[
0.7958984375,
0.833984375,
0.859375,
0.833984375,
0.859375,
0.84521484375,
0.7958984375,
0.84521484375
],
[
0.08416748046875,
0.82568359375,
0.48193359375,
0.8271484375,
0.48193359375,
0.84130859375,
0.08416748046875,
0.84033203125
],
[
0.74169921875,
0.83544921875,
0.75634765625,
0.83544921875,
0.75634765625,
0.84521484375,
0.74169921875,
0.84521484375
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.483642578125,
0.85498046875,
0.483642578125,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.08416748046875,
0.8671875,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.08416748046875,
0.880859375
],
[
0.51708984375,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.08416748046875,
0.88330078125,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.08416748046875,
0.896484375
],
[
0.51708984375,
0.8818359375,
0.64453125,
0.88330078125,
0.64453125,
0.896484375,
0.51708984375,
0.89599609375
]
] |
[
"AdKDD20,August 2020, San Diego,California,USA",
"S Pan.B Kitts.T Zhou,H He,B Shetty,AFlores,D Gligorijevic.J Pan,TMao.S Gultekin and IZhang",
"[8] reported shading in a Texas cattle market, and [17] found the",
"[34] extended their work to using a neural network to estimate the",
"practice in auctions for Us Treasury notes.",
"parameters of the win probability distribution.",
"Auctions generally need to be repeated and predictable for bid",
"An unpublished implementation [19] used Logistic Regression to",
"shading to be practically feasible, but under these conditions, it",
"predict the optimal bid shading factor using features in the request.",
"often occurs organically. Pownall and Wolk (2013) showed that bid",
"The predicted factor was then used as a multiplier on the unshaded",
"shadingforrepeated internet auctionprices increased over time;by",
"bid price",
"about 26% after 10 iterations [24]. When there are enough repeated",
"The approaches described above [19, 34, 35] all focus on predict-",
"games bidders can even develop collusive shading strategies where",
"ing the probable winning bid price. However, the surplus maximum",
"bidders actively coordinate to have low bids [16, 23]",
"is very different from the minimum bid to win. An accurate (unbi-",
"Although behavior varies from auction to auction, several studies",
"ased, symmetric noise)win probability estimator will be below the",
"have shown that the magnitude of shading tends to increase with",
"winning bid price about 50% of the time-this means that 50% of",
"the average price on the auction [4, 6, 17]. This is likely to occur",
"the surplus won't be captured by design. If the change in new im-",
"because ofthemoresubstantial losses involved onhigherpriced",
"pressions captured at a higher bid price, over-weights the marginal",
"auctions,if shading isn't sufficient.This result suggests thatusing",
"decreaseinprofitabilityperimpression,theoptimumforsurplus",
"a measure of the expense of the auction is valuable when trying to",
"can be higher than the most probable bid.",
"estimatetheshadingfactor-afindingwerevisitlaterinSection6",
"Unpublished work [20] is one of the few that we know of to",
"In situationswherethesupplyisplentiful,and demand limited,",
"attemptto explicitlymaximizethe surplus function.Theseauthors",
"buyers can shade deeper. In looking at this phenomemon in the US",
"estimate shading factors for a set of fixed segments based on three",
"Treasury Market, Hortacsu et. al. (2017) find that large institutional",
"bid samples taken in real-time to estimate the local surplus land-",
"buyers on average shade more aggressively than small indirect",
"scape.Howeverthe approach hasmany drawbacks:the segments",
"buyers [17] . This seems to be because these large buyers effectively",
"have to be predetermined and finding a suitable segment defini-",
"control alargepercentage ofbidders,and so itis almost liketheyare",
"tion requires substantial analysis.The information across segments",
"able to coordinate the buying of multiple buyers. They can therefore",
"is not shared, which is a problemfor segments that do not have",
"drive the bid prices for a large percentage of bidders down, whilst",
"enough traffic. Further, the set of possible segments quickly explode",
"still meeting their goals.",
"as thenumberofvariablesused to definethemincreases.Theap-",
"proach taken in this paper uses a model to estimate the surplus",
"function, and so a very large number of features can be used, and",
"3.2 Previous Algorithms",
"model induction is also automated, easy to maintain, and improve",
"In 2018 and 2019, Rubicon [1, 25], AppNexus [3] and Google [14, 15,",
"In order to compare the method we used to prior work, we have",
"28,30|all released Sell-Sidebidshading services.Leading uptothis",
"included an implementation of the Logistic Regression algorithm",
"there had been reports of dramatically lower ROI from the new First",
"from[19],theDistributionEstimatoralgorithmfrom[35],andthe",
"Price Auctions [18, 21]. Never-the-less, this is a surprising move",
"Segment-based Surplus maximizer [20] in the benchmarks which",
"as Sell-Side Platforms are potentially decreasing their yield, and",
"we use to analyze algorithm performance in Section 7.",
"they clearly have a different incentive from buyers. The sell-side",
"algorithms seemtoreflectthisincentivedifference.The descriptions",
"1",
"",
" CANONICAL ALGORITHM",
"of these services suggestthat they tryto keep bidpriceshigh",
"Given abidrequestforFirst Price Auction,letxi,x2",
"xkbethe",
"set of publisher and user attributes that we will use to find the best",
"enough to maintain a set win-rate, but preventing the bid price",
"from becoming tooextreme:which mightrisk an advertiser to halt",
"their bidding due to poor Return on Investment.Rubicon released",
"bid price b*. Let b be the highest bid price from other competing",
"data suggesting that their service decreases First Price CPMs by",
"bidders, which value is unknown. Note that b depends on both",
"a modest 5% over 4 months [25]. AppNexus reported that prices",
"attributes xs which represent the item that is being auctioned,",
"under their service were 25% lower over 100 days [3]. We tried one",
"and external competing bidder behavior. b follows an unknown",
"of the services, and recorded the shading distribution in Figure 2",
"distribution Dbixi.x2",
"with cumulative probabilitydistribution",
"Most of the bid shades were about 90%, which is conservative for",
"cdfxi.x.k",
"When the context is clear,weuse D,and cdf,for",
"our problem. Further analysis on Sell-Side \"Bid Shaders\" are in",
"simplicity",
"Section 7.",
"If the distribution D, is known, we can calculate the optimal",
"On the Demand Side, a variety of algorithms have been explored",
"bid priceb*directlyasfollows.LetI(b > b)be1if b>band 0",
"although generally not exactly for bid shading applications. [35]",
"developed a censored winning bid probability estimator. They ob",
"otherwise,whichindicatesifthesubmittedpricebwinstheauction",
"Then the surplus when the submitted price is b would be",
"served that when a bidder submitted a bid and lost, the information",
"gained is that the winning price is somewhere abovethe submitted",
"price, and when a bidder submits a bid and wins, theminimum bid",
"(V-b, ifb>b,",
"surplus =(V-b)I(b >b)",
"(2)",
"otherwise",
"to win is at a price somewhere below their submitted bid. Using",
"0,",
"these two cases, the authors developed a Maximum Likelihood pro-",
"cedure to estimate the probability of the winning bid being any of",
"the bid prices.This created a distribution of the probable winning",
"The optimal bid price can be calculated as the price that maximizes",
"bids, with the most likely winning price being used for bidding.",
"the expected surplus"
] |
{
"arxiv_id": "2009.09259",
"language": "en",
"page_idx": 1
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.