id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2208.00183-en-1
|
2208.00183-en-1.jpg
|
2
${\mathrm{Zhen~Xing~et~al.}}$
<FigureHere>
Fig.1. (a) Novel category 3D reconstruction results on category-specific method and our method combination of prior. (b) The mIoU(%) of current methods against the number of shots. Our MPCN outperforms the SOTA approaches with different shot.
Notably, Tatarchenko et al. [31] shows that single-view 3D reconstruction of specific categories is closely related to image recognition for shape matching. Several simple image retrieval baseline methods can even outperform state-ofthe-art 3D reconstruction methods.
In this paper, we propose a novel category-agnostic model for single-view 3D reconstruction in the few-shot learning settings. In our method, the network parameters are first optimized on some specific object categories. Then given novel categories, the model can be quickly transferred to 3D reconstruction of these categories based on few-shot learning techniques.
To the best of our knowledge, there are mainly three previous works focus on unseen category 3D reconstruction. Wallace et al. [34] present a network for single-view 3D reconstruction with additional shape prior of object categories. However, their shape prior model cannot catch the diversity of shapes within an object category. The authors of Compositional Global Class Embeddings (CGCE) [22] adopt a solution to optimize the shape prior codebooks for the reconstruction of unseen classes. Their model depends on finetuning with additional codebooks, which makes the training process complicated and makes the performance unstable. Pose Adaptive Dual Mixup (PADMix) [6] proposes to apply mixup [46] at the feature level and introduce pose level information, which reaches a new state-of-the-art performance in this task.
In addition, all the works rely on shape prior of specific categories. As a result, additional category annotation is needed to recognize the category of the input image. Then these methods can construct shape prior according to the category annotation, which is not very suitable for category-agnostic 3D reconstruction with novel categories.
The previous works of exploring shape prior for 3D reconstruction of novel categories are insightful and reasonable. However, there still exits a challenge on how to handle shape variety within a novel category in the context of fewshot learning. In this paper, we present a novel deep network model with a memory that can store a shape and its corresponding image as a key-value pair for retrieval. When an image of a novel category is inputted to the network, our deep network can select and combine appropriate shapes retrieved from the
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.271240234375,
0.11614990234375,
0.381591796875,
0.11614990234375,
0.381591796875,
0.1300048828125,
0.271240234375,
0.1300048828125
],
[
0.348876953125,
0.146484375,
0.393798828125,
0.1483154296875,
0.39208984375,
0.160400390625,
0.34716796875,
0.1578369140625
],
[
0.488525390625,
0.146484375,
0.5087890625,
0.146484375,
0.5087890625,
0.160400390625,
0.488525390625,
0.160400390625
],
[
0.56396484375,
0.1427001953125,
0.59716796875,
0.143310546875,
0.5947265625,
0.21533203125,
0.560546875,
0.2139892578125
],
[
0.39794921875,
0.15283203125,
0.4150390625,
0.147705078125,
0.419189453125,
0.1546630859375,
0.400390625,
0.160400390625
],
[
0.59619140625,
0.1572265625,
0.60693359375,
0.1572265625,
0.60693359375,
0.179931640625,
0.59619140625,
0.179931640625
],
[
0.60546875,
0.1558837890625,
0.62744140625,
0.1558837890625,
0.62744140625,
0.1641845703125,
0.60546875,
0.1641845703125
],
[
0.30810546875,
0.1641845703125,
0.348876953125,
0.1641845703125,
0.348876953125,
0.1717529296875,
0.30810546875,
0.1717529296875
],
[
0.420654296875,
0.1654052734375,
0.4453125,
0.1654052734375,
0.4453125,
0.17236328125,
0.420654296875,
0.17236328125
],
[
0.611328125,
0.1641845703125,
0.625,
0.1641845703125,
0.625,
0.1749267578125,
0.611328125,
0.1749267578125
],
[
0.5146484375,
0.177978515625,
0.52294921875,
0.177978515625,
0.52294921875,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.49267578125,
0.1881103515625,
0.5087890625,
0.1881103515625,
0.5087890625,
0.1912841796875,
0.49267578125,
0.1912841796875
],
[
0.2548828125,
0.1956787109375,
0.303955078125,
0.1962890625,
0.303955078125,
0.203857421875,
0.2548828125,
0.20263671875
],
[
0.576171875,
0.1912841796875,
0.583984375,
0.1988525390625,
0.57861328125,
0.20263671875,
0.56884765625,
0.1962890625
],
[
0.57275390625,
0.208984375,
0.59326171875,
0.208984375,
0.59326171875,
0.240478515625,
0.57275390625,
0.240478515625
],
[
0.580078125,
0.2052001953125,
0.58642578125,
0.2052001953125,
0.58642578125,
0.2108154296875,
0.580078125,
0.2108154296875
],
[
0.48193359375,
0.212158203125,
0.498291015625,
0.216552734375,
0.495849609375,
0.2235107421875,
0.4794921875,
0.2197265625
],
[
0.378173828125,
0.2197265625,
0.3896484375,
0.2197265625,
0.3896484375,
0.222900390625,
0.378173828125,
0.222900390625
],
[
0.48046875,
0.2108154296875,
0.496826171875,
0.2108154296875,
0.496826171875,
0.21533203125,
0.48046875,
0.21533203125
],
[
0.57861328125,
0.238037109375,
0.59619140625,
0.238037109375,
0.59619140625,
0.246826171875,
0.57861328125,
0.246826171875
],
[
0.64794921875,
0.240478515625,
0.69140625,
0.242431640625,
0.689453125,
0.25830078125,
0.646484375,
0.255615234375
],
[
0.736328125,
0.2392578125,
0.74658203125,
0.2392578125,
0.74658203125,
0.25,
0.736328125,
0.25
],
[
0.373291015625,
0.255126953125,
0.39208984375,
0.255126953125,
0.39208984375,
0.268310546875,
0.373291015625,
0.268310546875
],
[
0.6650390625,
0.25244140625,
0.68701171875,
0.25244140625,
0.68701171875,
0.26513671875,
0.6650390625,
0.26513671875
],
[
0.2181396484375,
0.281005859375,
0.78759765625,
0.281005859375,
0.78759765625,
0.29541015625,
0.2181396484375,
0.29541015625
],
[
0.2181396484375,
0.29541015625,
0.78759765625,
0.29541015625,
0.78759765625,
0.309326171875,
0.2181396484375,
0.309326171875
],
[
0.21728515625,
0.30810546875,
0.77783203125,
0.309326171875,
0.77783203125,
0.323974609375,
0.21728515625,
0.322509765625
],
[
0.24267578125,
0.354248046875,
0.78759765625,
0.354248046875,
0.78759765625,
0.367431640625,
0.24267578125,
0.367431640625
],
[
0.21728515625,
0.36865234375,
0.78369140625,
0.369384765625,
0.78369140625,
0.3837890625,
0.21728515625,
0.38330078125
],
[
0.21728515625,
0.385009765625,
0.78515625,
0.385009765625,
0.78515625,
0.3984375,
0.21728515625,
0.3984375
],
[
0.2181396484375,
0.399658203125,
0.47216796875,
0.399658203125,
0.47216796875,
0.41357421875,
0.2181396484375,
0.41357421875
],
[
0.2418212890625,
0.415283203125,
0.78369140625,
0.415283203125,
0.78369140625,
0.429931640625,
0.2418212890625,
0.429931640625
],
[
0.2181396484375,
0.43115234375,
0.7861328125,
0.43115234375,
0.7861328125,
0.44580078125,
0.2181396484375,
0.44580078125
],
[
0.21728515625,
0.4462890625,
0.78515625,
0.4462890625,
0.78515625,
0.460205078125,
0.21728515625,
0.460205078125
],
[
0.21728515625,
0.4609375,
0.78759765625,
0.4609375,
0.78759765625,
0.474853515625,
0.21728515625,
0.474853515625
],
[
0.21728515625,
0.4765625,
0.611328125,
0.4765625,
0.611328125,
0.490478515625,
0.21728515625,
0.490478515625
],
[
0.24267578125,
0.492431640625,
0.78759765625,
0.492431640625,
0.78759765625,
0.505859375,
0.24267578125,
0.505859375
],
[
0.2181396484375,
0.5068359375,
0.7861328125,
0.5068359375,
0.7861328125,
0.52099609375,
0.2181396484375,
0.52099609375
],
[
0.2181396484375,
0.52294921875,
0.78515625,
0.52294921875,
0.78515625,
0.5361328125,
0.2181396484375,
0.5361328125
],
[
0.2181396484375,
0.537109375,
0.7861328125,
0.537109375,
0.7861328125,
0.55029296875,
0.2181396484375,
0.55029296875
],
[
0.2181396484375,
0.55322265625,
0.78759765625,
0.55322265625,
0.78759765625,
0.56640625,
0.2181396484375,
0.56640625
],
[
0.2197265625,
0.568359375,
0.78759765625,
0.568359375,
0.78759765625,
0.58203125,
0.2197265625,
0.58203125
],
[
0.21728515625,
0.58203125,
0.78369140625,
0.58349609375,
0.78369140625,
0.59765625,
0.21728515625,
0.5966796875
],
[
0.2181396484375,
0.5986328125,
0.7861328125,
0.5986328125,
0.7861328125,
0.6123046875,
0.2181396484375,
0.6123046875
],
[
0.21728515625,
0.6123046875,
0.78759765625,
0.61279296875,
0.78759765625,
0.62744140625,
0.21728515625,
0.626953125
],
[
0.2181396484375,
0.62890625,
0.78515625,
0.62890625,
0.78515625,
0.642578125,
0.2181396484375,
0.642578125
],
[
0.2181396484375,
0.64306640625,
0.66015625,
0.64306640625,
0.66015625,
0.6572265625,
0.2181396484375,
0.6572265625
],
[
0.2418212890625,
0.65771484375,
0.78173828125,
0.65771484375,
0.78173828125,
0.671875,
0.2418212890625,
0.671875
],
[
0.21728515625,
0.673828125,
0.7861328125,
0.6748046875,
0.7861328125,
0.689453125,
0.21728515625,
0.68798828125
],
[
0.21728515625,
0.68994140625,
0.78515625,
0.68994140625,
0.78515625,
0.7041015625,
0.21728515625,
0.7041015625
],
[
0.21728515625,
0.705078125,
0.7861328125,
0.705078125,
0.7861328125,
0.71826171875,
0.21728515625,
0.71826171875
],
[
0.2181396484375,
0.72021484375,
0.371826171875,
0.72021484375,
0.371826171875,
0.734375,
0.2181396484375,
0.734375
],
[
0.2442626953125,
0.736328125,
0.78759765625,
0.736328125,
0.78759765625,
0.75,
0.2442626953125,
0.75
],
[
0.2181396484375,
0.751953125,
0.78515625,
0.751953125,
0.78515625,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.21728515625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.78076171875,
0.21728515625,
0.78076171875
],
[
0.21728515625,
0.7822265625,
0.7861328125,
0.7822265625,
0.7861328125,
0.79541015625,
0.21728515625,
0.79541015625
],
[
0.21728515625,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.21728515625,
0.81005859375,
0.78369140625,
0.81103515625,
0.78369140625,
0.82568359375,
0.21728515625,
0.82470703125
],
[
0.2156982421875,
0.8271484375,
0.78759765625,
0.82568359375,
0.78759765625,
0.84033203125,
0.2156982421875,
0.84130859375
]
] |
[
"2",
"Zhen Xing et al.",
"Category-",
"G.T.",
"II",
"ecifi",
"车",
"wadace",
"InputImage",
"Perforn",
"PACE",
":",
"2000",
"ExternalMeme",
"",
"",
"2",
"Durs",
"tc",
"IFLI",
"35.04",
"Tahot mumber",
"25",
"(a)",
"(b) ",
"Fig.1. (a) Novel category 3D reconstruction results on category-specific method and",
"our method combination of prior. (b) The mloU(%) of current methods against the",
"number of shots. Our MPCN outperforms the SOTA approaches with different shot.",
"Notably, Tatarchenko et al. [31] shows that single-view 3D reconstruction of",
"specific categories is closely related to image recognition for shape matching",
"Several simple image retrieval baseline methods can even outperform state-of-",
"the-art 3D reconstruction methods.",
"In this paper, we propose a novel category-agnostic model for single-view",
"3D reconstruction in the few-shot learning settings. In our method, the network",
"parameters are first optimized on some specific object categories. Then given",
"novel categories, the model can be quickly transferred to 3D reconstruction of",
"these categories based on few-shot learning techniques.",
"To the best of our knowledge, there are mainly three previous works focus",
"on unseen category 3D reconstruction. Wallace et al. [34] present a network for",
"single-view 3D reconstruction with additional shape prior of object categories.",
"However, their shape prior model cannot catch the diversity of shapes within",
"an object category. The authors of Compositional Global Class Embeddings",
"(CGCE) [22] adopt a solution to optimize the shape prior codebooks for the",
"reconstruction of unseen classes. Their model depends on finetuning with ad-",
"ditional codebooks, which makes the training process complicated and makes",
"the performance unstable. Pose Adaptive Dual Mixup (PADMix) [6] proposes",
"to apply mixup [46] at the feature level and introduce pose level information,",
"which reaches a new state-of-the-art performance in this task.",
"In addition, all the works rely on shape prior of specific categories. As a result",
"additional category annotation is needed to recognize the category of the input",
"image. Then these methods can construct shape prior according to the category",
"annotation, which is not very suitable for category-agnostic 3D reconstruction",
"with novel categories",
"The previous works of exploring shape prior for 3D reconstruction of novel",
"categories are insightful and reasonable. However, there still exits a challenge",
"on how to handle shape variety within a novel category in the context of few-",
"shot learning. In this paper, we present a novel deep network model with a",
"memory that can store a shape and its corresponding image as a key-value pair",
"for retrieval. When an image of a novel category is inputted to the network.",
"our deep network can select and combine appropriate shapes retrieved from the"
] |
{
"arxiv_id": "2208.00183",
"language": "en",
"page_idx": 1
}
|
||
2208.00183-en-2
|
2208.00183-en-2.jpg
|
3
Few-shot 3D Reconstruction with Memory Prior Contrastive Network
memory without category annotation to guide a decoder for 3D reconstruction. In order to adaptively combine the stored shapes as a prior for the downstream 3D reconstruction task, a multi-head attention shape prior module is proposed. Fig. 1(a) shows the example on novel watercraft category 3D reconstruction performance between traditionally category-specific method [8] and our method.
Besides, we propose a 3D-aware contrastive loss that pulls together the image features of objects with similar shape and pushes away the image features with different 3D shape in the latent feature space, which helps for both organizing the image feature and improving the retrieval accuracy of memory network. Our 3Daware contrastive loss takes into account the difference of shape as a weighting term to reduce or stress the positiveness of a pair, regardless of the category or instance, as we aim at a category-agnostic 3D reconstruction network.
In summary, our contributions are as follows: We propose a novel Memory Prior Contrastive Network (MPCN) that can retrieve shape prior as an intermediate representation to help the neural network to infer the shape of novel objects without any category annotations.
Our multi-head attention prior module can automatically learn the association between retrieved prior and pay attention to different part of shape prior. It can not only provide prior information between object categories, but also represent the differences within objects in the same category.
The network with both reconstruction loss and a contrastive loss works together for better result. Our improved 3D-aware contrastive loss takes into account of the difference of positive samples, which is more suitable for supervised 3D tasks.
Experimental results on ShapeNet [3] dataset show that our method greatly improves the state-of-the-art methods on the two mainstream evaluation metrics using Intersection over Union and F-score. The reconstruction results on the realworld Pascal3D+ [41] dataset also demonstrate the effectiveness of our method quantitatively and qualitatively.
## 2 Related Work
Deep Learning 3D reconstruction Recently, Convolutional Neural Network (CNN) based single-view 3D reconstruction methods become more and more popular. Using voxels to represent a 3D shape is suitable for 3D CNNs. In the early work 3D Recurrent Reconstruction Neural Network (3D-R2N2) [8], the encoder with a Recurrent Neural Network (RNN) structure is used to predict 3D voxels by a 3D decoder. A follow-up work, 3D-VAE-GAN [39], explores the generation ability of Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) to infer 3D shapes. Marrnet [38] and ShapeHD [40] predict the 2.5D information such as depth, silhouette and surface normal of an input RGB image, and then use these intermediate information to reconstruct the 3D object. OGN[30] and Matryoshka[27] utilize octrees or nested shape layers to represent high resolution 3D volume. Pix2Vox [42] and Pix2Vox++ [43] mainly improve the fusion of multi-view 3D reconstruction. Mem3D[45] introduces ex-
|
[
[
0.268798828125,
0.11553955078125,
0.73388671875,
0.11553955078125,
0.73388671875,
0.1287841796875,
0.268798828125,
0.1287841796875
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.21728515625,
0.147705078125,
0.78369140625,
0.147705078125,
0.78369140625,
0.1622314453125,
0.21728515625,
0.1622314453125
],
[
0.21728515625,
0.1622314453125,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.21728515625,
0.1767578125
],
[
0.2181396484375,
0.1787109375,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.24267578125,
0.2247314453125,
0.78515625,
0.2247314453125,
0.78515625,
0.238037109375,
0.24267578125,
0.238037109375
],
[
0.21728515625,
0.2374267578125,
0.7861328125,
0.238037109375,
0.7861328125,
0.25244140625,
0.21728515625,
0.251953125
],
[
0.2181396484375,
0.25390625,
0.7861328125,
0.25390625,
0.7861328125,
0.267578125,
0.2181396484375,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.78369140625,
0.26953125,
0.78369140625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.21728515625,
0.28271484375,
0.7861328125,
0.2841796875,
0.7861328125,
0.298583984375,
0.21728515625,
0.29736328125
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.21728515625,
0.314453125,
0.71826171875,
0.314453125,
0.71826171875,
0.328857421875,
0.21728515625,
0.328857421875
],
[
0.2181396484375,
0.330078125,
0.7861328125,
0.330078125,
0.7861328125,
0.343505859375,
0.2181396484375,
0.343505859375
],
[
0.2181396484375,
0.3447265625,
0.78369140625,
0.3447265625,
0.78369140625,
0.35791015625,
0.2181396484375,
0.35791015625
],
[
0.2181396484375,
0.35986328125,
0.78759765625,
0.35986328125,
0.78759765625,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.2181396484375,
0.375732421875,
0.51953125,
0.375732421875,
0.51953125,
0.389404296875,
0.2181396484375,
0.389404296875
],
[
0.2442626953125,
0.39013671875,
0.78173828125,
0.39013671875,
0.78173828125,
0.404052734375,
0.2442626953125,
0.404052734375
],
[
0.21728515625,
0.404052734375,
0.78515625,
0.40478515625,
0.78515625,
0.419189453125,
0.21728515625,
0.41845703125
],
[
0.21728515625,
0.42041015625,
0.7861328125,
0.42041015625,
0.7861328125,
0.434326171875,
0.21728515625,
0.434326171875
],
[
0.2181396484375,
0.436279296875,
0.654296875,
0.436279296875,
0.654296875,
0.449462890625,
0.2181396484375,
0.449462890625
],
[
0.2442626953125,
0.45068359375,
0.78369140625,
0.45068359375,
0.78369140625,
0.464111328125,
0.2442626953125,
0.464111328125
],
[
0.2181396484375,
0.466552734375,
0.78369140625,
0.466552734375,
0.78369140625,
0.479736328125,
0.2181396484375,
0.479736328125
],
[
0.2181396484375,
0.48095703125,
0.7861328125,
0.48095703125,
0.7861328125,
0.494384765625,
0.2181396484375,
0.494384765625
],
[
0.2181396484375,
0.49755859375,
0.28515625,
0.49755859375,
0.28515625,
0.5087890625,
0.2181396484375,
0.5087890625
],
[
0.2418212890625,
0.51220703125,
0.78369140625,
0.51220703125,
0.78369140625,
0.52587890625,
0.2418212890625,
0.52587890625
],
[
0.21728515625,
0.5263671875,
0.78515625,
0.5263671875,
0.78515625,
0.54052734375,
0.21728515625,
0.54052734375
],
[
0.21728515625,
0.54248046875,
0.78515625,
0.54248046875,
0.78515625,
0.55615234375,
0.21728515625,
0.55615234375
],
[
0.2181396484375,
0.556640625,
0.7861328125,
0.556640625,
0.7861328125,
0.57080078125,
0.2181396484375,
0.57080078125
],
[
0.2156982421875,
0.57275390625,
0.44775390625,
0.5712890625,
0.44775390625,
0.5859375,
0.2156982421875,
0.5869140625
],
[
0.2156982421875,
0.6142578125,
0.2344970703125,
0.60986328125,
0.238525390625,
0.62255859375,
0.2197265625,
0.626953125
],
[
0.240234375,
0.60986328125,
0.38818359375,
0.60986328125,
0.38818359375,
0.62744140625,
0.240234375,
0.62744140625
],
[
0.2181396484375,
0.64453125,
0.78515625,
0.64453125,
0.78515625,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.2181396484375,
0.66015625,
0.7861328125,
0.66015625,
0.7861328125,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.2181396484375,
0.67529296875,
0.78759765625,
0.67529296875,
0.78759765625,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.78369140625,
0.72021484375,
0.78369140625,
0.734375,
0.2181396484375,
0.734375
],
[
0.2197265625,
0.7373046875,
0.78515625,
0.7373046875,
0.78515625,
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,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.77978515625,
0.78369140625,
0.77978515625,
0.78369140625,
0.79345703125,
0.2181396484375,
0.79345703125
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"Few-shot 3D Reconstruction with Memory Prior Contrastive Network",
"3",
"memory without category annotation to guide a decoder for 3D reconstruction.",
"In order to adaptively combine the stored shapes as a prior for the downstream",
"3D reconstruction task, a multi-head attention shape prior module is proposed.",
"Fig. 1(a) shows the example on novel watercraft category 3D reconstruction",
"performance between traditionally category-specific method [8] and our method.",
"Besides, we propose a 3D-aware contrastive loss that pulls together the image",
"features of objects with similar shape and pushes away the image features with",
"different 3D shape in the latent feature space, which helps for both organizing the",
"image feature and improving the retrieval accuracy of memory network. Our 3D-",
"aware contrastive loss takes into account the difference of shape as a weighting",
"term to reduce or stress the positiveness of a pair, regardless of the category or",
"instance, as we aim at a category-agnostic 3D reconstruction network.",
"In summary, our contributions are as follows: Wepropose a novel Memory",
"Prior Contrastive Network (MPCN) that can retrieve shape prior as an inter-",
"mediate representation to help the neural network to infer the shape of novel",
"objects without any category annotations.",
"Our multi-head attention prior module can automatically learn the associa",
"tion between retrieved prior and pay attention to different part of shape prior.",
"It can not only provide prior information between object categories, but also",
"represent the differences within objects in the same category.",
"The network with both reconstruction loss and a contrastive loss works to",
"gether for better result. Our improved 3D-aware contrastive loss takes into ac-",
"count of the difference of positive samples, which is more suitable for supervised",
"3Dtasks.",
"Experimental results on ShapeNet [3] dataset show that our method greatly",
"improves the state-of-the-art methods on the two mainstream evaluation metrics",
"using Intersection over Union and F-score. The reconstruction results on the real-",
"world Pascal3D+ [41] dataset also demonstrate the effectiveness of our method",
"quantitatively and qualitatively.",
"2",
"Related Work",
"Deep Learning 3D reconstruction Recently, Convolutional Neural Network",
"(CNN) based single-view 3D reconstruction methods become more and more",
"popular. Using voxels to represent a 3D shape is suitable for 3D CNNs. In the",
"early work 3D Recurrent Reconstruction Neural Network (3D-R2N2) [8], the en-",
"coder with a Recurrent Neural Network (RNN) structure is used to predict 3D",
"voxels by a 3D decoder. A follow-up work, 3D-VAE-GAN [39], explores the gen-",
"eration abilityofVariationalAutoencoders(VAEs)and Generative Adversarial",
"Networks (GANs) to infer 3D shapes. Marrnet [38] and ShapeHD [40] predict",
"the 2.5D information such as depth, silhouette and surface normal of an input",
"RGB image, and then use these intermediate information to reconstruct the 3D",
"object. OGN[30] and Matryoshka[27] utilize octrees or nested shape layers to",
"represent high resolution 3D volume. Pix2Vox [42] and Pix2Vox++ [43] mainly",
"improve the fusion of multi-view 3D reconstruction. Mem3D[45] introduces ex-"
] |
{
"arxiv_id": "2208.00183",
"language": "en",
"page_idx": 2
}
|
||
2208.00183-en-3
|
2208.00183-en-3.jpg
|
4 Zhen Xing et al.
ternal memory for category-specific 3D reconstruction. However, its performance relies on a large number of samples saved during the training process. More recently, SSP3D [47] propose a semi-supervised setting for 3D Reconstruction. In addition to voxels, 3D shapes can also be represented by point clouds [9,20,35], meshes [36,37] and signed distance fields [44,21].
Few-shot Learning Few-shot learning models can be roughly classified into two categories: metric-based methods and meta-based methods. Metric-based methods mainly utilize Siamese networks [19], match networks [33,7] or prototype networks [10] to model the distance distribution between samples such that similar samples are closer to each other and heterogeneous samples are far away from each other. Meta-based methods [26][25] and meta-gradient based methods [4,15,28] are teaching models by using few unseen samples to quickly update the model parameters in order to achieve generalization.
Few-shot 3D Reconstruction Wallace et al.[34] introduce the first method for single-view 3D reconstruction in the few-shot settings. They propose to combine category-specific shape priors with an input image to guide the 3D reconstruction process. However, in their work, a random shape or a calculated average shape is selected for each category as the prior information, which cannot account for shape diversity among objects in a category. In addition, the method does not explicitly learn the inter-class concepts. Compositional Global Class Embeddings (CGCE) [22] adopts a solution to quickly optimize codebooks for 3D reconstruction of novel categories. Before testing on a novel category, the parameters of other modules are fixed. Only the weight vector of codebooks are optimized with a few support samples from the novel category. Therefore, given a new category, CGCE needs to add a new codebook vector for this category and finetune the weight parameters, which makes the whole process complicated and inefficient. Pose Adaptive Dual Mixup(PADMix) [6] proposes a pose adaptive procedure and a three-stage training method with mixup [46]. It tries to solve the pose-invariance issue by an autoencoder but its shape prior module is similar to Wallace [34] and has the drawbacks of complicated three-stage training strategy.
## 3 Method
Our aim is to design a category-agnostic model, which can achieve superior generalization ability of single-view 3D reconstruction for novel categories with limited support data. Suppose there is a base 3D dataset, defined as $D_b=$ $\begin{array}{l}{\{(I_{i},V_{i})\}_{i=1}^{K},\mathrm{~where~}I_{i}\mathrm{~and~}V_{i}\mathrm{~denote~the~}i\text{th image and its corresponding 3D}}\\{\mathrm{l}}\end{array}$ shape represented using voxels, respectively. $K$ is the number of image and voxel pairs in the dataset. We denote the categories in the base dataset $D_b$ as base categories. Let $D_s=\{(I_{i},V_{i})\}_{i=1}^{M}$ be another dataset of (image, voxel) pairs with $M$ examples. The categories in $D_{\mathrm{s}}$ are defined as the novel categories, which are different from those in $D_b.D_{s}$ is called the support set, where $M\ll K.$ Meanwhile, there is a large test or query set $D_q=\{(I_{i},V_{i})\}_{i=1}^{N}$ with $N$ examples. Examples in $D_{q}$ and $D_{s}$ are all within the novel categories. Note that we only
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.271240234375,
0.11553955078125,
0.38232421875,
0.11553955078125,
0.38232421875,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.21728515625,
0.1483154296875,
0.7861328125,
0.1483154296875,
0.7861328125,
0.1622314453125,
0.21728515625,
0.1622314453125
],
[
0.21728515625,
0.1634521484375,
0.78369140625,
0.1634521484375,
0.78369140625,
0.177978515625,
0.21728515625,
0.177978515625
],
[
0.21728515625,
0.177978515625,
0.78515625,
0.177978515625,
0.78515625,
0.1925048828125,
0.21728515625,
0.1925048828125
],
[
0.21728515625,
0.1932373046875,
0.78515625,
0.1932373046875,
0.78515625,
0.20703125,
0.21728515625,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.564453125,
0.208984375,
0.564453125,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2181396484375,
0.2247314453125,
0.7861328125,
0.2247314453125,
0.7861328125,
0.238037109375,
0.2181396484375,
0.238037109375
],
[
0.21728515625,
0.2392578125,
0.7861328125,
0.2392578125,
0.7861328125,
0.25244140625,
0.21728515625,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.78369140625,
0.25390625,
0.78369140625,
0.267578125,
0.2181396484375,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.78515625,
0.26953125,
0.78515625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2181396484375,
0.285400390625,
0.78515625,
0.285400390625,
0.78515625,
0.298583984375,
0.2181396484375,
0.298583984375
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.2197265625,
0.314453125,
0.7861328125,
0.314453125,
0.7861328125,
0.32763671875,
0.2197265625,
0.32763671875
],
[
0.2181396484375,
0.330078125,
0.5947265625,
0.330078125,
0.5947265625,
0.343505859375,
0.2181396484375,
0.343505859375
],
[
0.2181396484375,
0.3447265625,
0.7861328125,
0.3447265625,
0.7861328125,
0.35791015625,
0.2181396484375,
0.35791015625
],
[
0.2181396484375,
0.35986328125,
0.7861328125,
0.35986328125,
0.7861328125,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.21728515625,
0.375,
0.78369140625,
0.375,
0.78369140625,
0.38818359375,
0.21728515625,
0.38818359375
],
[
0.21728515625,
0.389404296875,
0.7861328125,
0.39013671875,
0.7861328125,
0.40478515625,
0.21728515625,
0.404052734375
],
[
0.2181396484375,
0.40478515625,
0.78515625,
0.40478515625,
0.78515625,
0.419189453125,
0.2181396484375,
0.419189453125
],
[
0.21728515625,
0.42041015625,
0.78759765625,
0.42041015625,
0.78759765625,
0.434326171875,
0.21728515625,
0.434326171875
],
[
0.2181396484375,
0.436279296875,
0.7861328125,
0.436279296875,
0.7861328125,
0.449462890625,
0.2181396484375,
0.449462890625
],
[
0.2181396484375,
0.45068359375,
0.7861328125,
0.45068359375,
0.7861328125,
0.464111328125,
0.2181396484375,
0.464111328125
],
[
0.2181396484375,
0.466552734375,
0.7861328125,
0.466552734375,
0.7861328125,
0.479736328125,
0.2181396484375,
0.479736328125
],
[
0.21728515625,
0.48095703125,
0.7861328125,
0.479736328125,
0.7861328125,
0.494384765625,
0.21728515625,
0.49560546875
],
[
0.2181396484375,
0.496826171875,
0.78515625,
0.496826171875,
0.78515625,
0.51025390625,
0.2181396484375,
0.51025390625
],
[
0.2181396484375,
0.51123046875,
0.7861328125,
0.51123046875,
0.7861328125,
0.5244140625,
0.2181396484375,
0.5244140625
],
[
0.2156982421875,
0.52587890625,
0.7861328125,
0.5244140625,
0.7861328125,
0.5390625,
0.2156982421875,
0.54052734375
],
[
0.2156982421875,
0.54052734375,
0.7861328125,
0.54150390625,
0.7861328125,
0.55615234375,
0.2156982421875,
0.5546875
],
[
0.2181396484375,
0.556640625,
0.7861328125,
0.556640625,
0.7861328125,
0.57080078125,
0.2181396484375,
0.57080078125
],
[
0.2181396484375,
0.5712890625,
0.78515625,
0.5712890625,
0.78515625,
0.58544921875,
0.2181396484375,
0.58544921875
],
[
0.2156982421875,
0.58544921875,
0.78759765625,
0.5869140625,
0.78759765625,
0.6015625,
0.2156982421875,
0.599609375
],
[
0.21728515625,
0.60302734375,
0.281005859375,
0.60302734375,
0.281005859375,
0.6171875,
0.21728515625,
0.6171875
],
[
0.2197265625,
0.6396484375,
0.23779296875,
0.642578125,
0.2320556640625,
0.65771484375,
0.214111328125,
0.65478515625
],
[
0.2484130859375,
0.64013671875,
0.328369140625,
0.64013671875,
0.328369140625,
0.65478515625,
0.2484130859375,
0.65478515625
],
[
0.2181396484375,
0.6748046875,
0.7861328125,
0.6748046875,
0.7861328125,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.77294921875,
0.70556640625,
0.77294921875,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.21728515625,
0.7197265625,
0.7861328125,
0.71728515625,
0.7861328125,
0.73486328125,
0.21728515625,
0.7373046875
],
[
0.2181396484375,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.21728515625,
0.75,
0.7861328125,
0.75048828125,
0.7861328125,
0.76513671875,
0.21728515625,
0.7646484375
],
[
0.21728515625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.21728515625,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78369140625,
0.78076171875,
0.78369140625,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.78369140625,
0.796875,
0.78369140625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.21728515625,
0.81103515625,
0.78173828125,
0.81103515625,
0.78173828125,
0.82568359375,
0.21728515625,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"Zhen Xing et al.",
"ternal memory for category-specific 3D reconstruction. However, its performance",
"relies on a large number of samples saved during the training process. More re-",
"cently, SSP3D [47] propose a semi-supervised setting for 3D Reconstruction. In",
"addition to voxels, 3D shapes can also be represented by point clouds [9,20,35],",
"meshes [36,37] and signed distance fields [44,21].",
"Few-shot Learning Few-shot learning models can be roughly classified into",
"two categories: metric-based methods and meta-based methods. Metric-based",
"methods mainly utilize Siamese networks [19], match networks [33,7] or proto-",
"type networks [io] to model the distance distribution between samples such that",
"similar samples are closer to each other and heterogeneous samples are far away",
"from each other. Meta-based methods [26][25] and meta-gradient based methods",
"[4,15,28] are teaching models by using few unseen samples to quickly update the",
"model parameters in order to achieve generalization.",
"Few-shot 3D Reconstruction Wallace et al.J34] introduce the first method for",
"single-view 3D reconstruction in the few-shot settings. They propose to combine",
"category-specific shape priors with an input image to guide the 3D reconstruc-",
"tion process. However, in their work, a random shape or a calculated average",
"shape is selected for each category as the prior information, which cannot ac-",
"count for shape diversity among objects in a category. In addition, the method",
"does not explicitly learn the inter-class concepts. Compositional Global Class",
"Embeddings (CGCE) [22] adopts a solution to quickly optimize codebooks for",
"3D reconstruction of novel categories. Before testing on a novel category, the",
"parameters of other modules are fixed. Only the weight vector of codebooks are",
"optimized with a few support samples from the novel category. Therefore, given",
"a new category, CGCE needs to add a new codebook vector for this category and",
"finetune the weight parameters, which makes the whole process complicated and",
"inefficient. Pose Adaptive Dual Mixup(PADMix) [6] proposes a pose adaptive",
"procedure and a three-stage training method with mixup [46]. It tries to solve",
"the pose-invariance issue by an autoencoder but its shape prior module is sim-",
"ilar to Wallace [34] and has the drawbacks of complicated three-stage training",
"strategy.",
"3",
"Method",
"Our aim is to design a category-agnostic model, which can achieve superior",
"generalization ability of single-view 3D reconstruction for novel categories with",
"limited support data. Suppose there is a base 3D dataset, defined as D",
"(Ii, V))1, where I; and V, denote the ith image and its corresponding 3D",
"shape represented using voxels, respectively. K is the number of image and",
"voxel pairs in the dataset. We denote the categories in the base dataset D as",
"base categories. Let Ds = {(Ii, V))1 be another dataset of (image, voxel)",
"pairs with M examples. The categories in Ds are defined as the novel categories.",
"which are different from those in D. Dg is called the support set, where M < K",
"Meanwhile, there is a large test or query set Dg = {(I, V)), with N examples",
"Examples in Dg and D, are all within the novel categories. Note that we only"
] |
{
"arxiv_id": "2208.00183",
"language": "en",
"page_idx": 3
}
|
||
2208.00183-en-4
|
2208.00183-en-4.jpg
|
5
Few-shot 3D Reconstruction with Memory Prior Contrastive Network
<FigureHere>
Fig.2. An overview of the proposed MPCN. In the training stage, we only use base categories to train the model, and insert memory slots into the memory network as alternative prior according to the rules we set. In the test phase, the memory network saves few shot of support set of the novel category to reconstruct 3D volumes of the query set.
use $D_b$ and $D_s$ for training. The support set $D_s$ can be used as prior information. We hope that the model can be designed to be category-agnostic and achieve good performance in the query set $D_q.$
## 3.1 Memory Network
In most previous works on single-view 3D reconstruction, the shape prior information is learned from the model parameters, which leads to category collapse when transferring to novel categories. As mentioned in [31], such kind of model makes the problem degenerated into a classification task. To alleviate this issue, directly using 3D shapes as priors is an intuitive and effective way. As shown in Figure 2, we adopt an explicit key-value memory network to store and calculate shape priors. In the training and testing stages, the CNN features of the input image is extracted by a 2D encoder. Then a retrieval task is performed, where the keys of the samples stored in the memory network are compared to the query vector and the corresponding Top-k retrieved shapes are sent to the prior module for generating prior features. Specifically, as shown in Eq. (1), the input image $I_q$ is frst encoded by a 2D encoder $E_{2\mathrm{D}}$, then the image features and shape prior features are concatenated. Finally, the 3D shape is inferred by a 3D decoder $D_{3\mathrm{D}}.$
$$pr=D_{\mathrm{3D}}(\mathrm{Concatenate}(E_{\mathrm{2D}}(I_{q}),\mathrm{prior~feature})),$$
(1)
where $pr$ is the final predict volume.
|
[
[
0.26953125,
0.117431640625,
0.73046875,
0.117431640625,
0.73046875,
0.1275634765625,
0.26953125,
0.1275634765625
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.77392578125,
0.1287841796875
],
[
0.330078125,
0.14453125,
0.408447265625,
0.14453125,
0.408447265625,
0.1558837890625,
0.330078125,
0.1558837890625
],
[
0.5947265625,
0.141357421875,
0.66015625,
0.141357421875,
0.66015625,
0.1558837890625,
0.5947265625,
0.1558837890625
],
[
0.52783203125,
0.1558837890625,
0.5498046875,
0.1558837890625,
0.5498046875,
0.169189453125,
0.52783203125,
0.169189453125
],
[
0.58837890625,
0.1546630859375,
0.62158203125,
0.1546630859375,
0.62158203125,
0.17041015625,
0.58837890625,
0.17041015625
],
[
0.295654296875,
0.1610107421875,
0.35546875,
0.1610107421875,
0.35546875,
0.1685791015625,
0.295654296875,
0.1685791015625
],
[
0.390625,
0.1610107421875,
0.465576171875,
0.1610107421875,
0.465576171875,
0.1685791015625,
0.390625,
0.1685791015625
],
[
0.69287109375,
0.1558837890625,
0.72119140625,
0.1558837890625,
0.72119140625,
0.17041015625,
0.69287109375,
0.17041015625
],
[
0.273681640625,
0.1685791015625,
0.289306640625,
0.173583984375,
0.28515625,
0.1817626953125,
0.26953125,
0.1767578125
],
[
0.371826171875,
0.1717529296875,
0.38818359375,
0.1717529296875,
0.38818359375,
0.179931640625,
0.371826171875,
0.179931640625
],
[
0.5185546875,
0.1666259765625,
0.54736328125,
0.169189453125,
0.5439453125,
0.18310546875,
0.51708984375,
0.18115234375
],
[
0.5849609375,
0.17236328125,
0.609375,
0.17236328125,
0.609375,
0.179931640625,
0.5849609375,
0.179931640625
],
[
0.70361328125,
0.1717529296875,
0.72119140625,
0.1717529296875,
0.72119140625,
0.1849365234375,
0.70361328125,
0.1849365234375
],
[
0.271240234375,
0.18115234375,
0.2900390625,
0.18115234375,
0.2900390625,
0.1932373046875,
0.271240234375,
0.1932373046875
],
[
0.64794921875,
0.177978515625,
0.65869140625,
0.177978515625,
0.65869140625,
0.18115234375,
0.64794921875,
0.18115234375
],
[
0.371826171875,
0.1849365234375,
0.39208984375,
0.1849365234375,
0.39208984375,
0.1956787109375,
0.371826171875,
0.1956787109375
],
[
0.69287109375,
0.1849365234375,
0.70361328125,
0.1817626953125,
0.705078125,
0.1875,
0.69384765625,
0.189453125
],
[
0.27294921875,
0.199462890625,
0.294189453125,
0.199462890625,
0.294189453125,
0.2083740234375,
0.27294921875,
0.2083740234375
],
[
0.370849609375,
0.1988525390625,
0.393798828125,
0.1988525390625,
0.393798828125,
0.212158203125,
0.370849609375,
0.212158203125
],
[
0.273681640625,
0.216552734375,
0.2900390625,
0.216552734375,
0.2900390625,
0.2235107421875,
0.273681640625,
0.2235107421875
],
[
0.370849609375,
0.21337890625,
0.393798828125,
0.21337890625,
0.393798828125,
0.2279052734375,
0.370849609375,
0.2279052734375
],
[
0.5908203125,
0.2203369140625,
0.625,
0.2203369140625,
0.625,
0.2279052734375,
0.5908203125,
0.2279052734375
],
[
0.6806640625,
0.2215576171875,
0.69677734375,
0.2215576171875,
0.69677734375,
0.2291259765625,
0.6806640625,
0.2291259765625
],
[
0.69921875,
0.2235107421875,
0.7158203125,
0.2235107421875,
0.7158203125,
0.2266845703125,
0.69921875,
0.2266845703125
],
[
0.371826171875,
0.2310791015625,
0.3896484375,
0.2310791015625,
0.3896484375,
0.2392578125,
0.371826171875,
0.2392578125
],
[
0.54736328125,
0.2235107421875,
0.5595703125,
0.2235107421875,
0.5595703125,
0.2266845703125,
0.54736328125,
0.2266845703125
],
[
0.27685546875,
0.231689453125,
0.28759765625,
0.231689453125,
0.28759765625,
0.236083984375,
0.27685546875,
0.236083984375
],
[
0.484375,
0.291748046875,
0.5048828125,
0.291748046875,
0.5048828125,
0.29541015625,
0.484375,
0.29541015625
],
[
0.43310546875,
0.314453125,
0.48193359375,
0.314453125,
0.48193359375,
0.3212890625,
0.43310546875,
0.3212890625
],
[
0.498291015625,
0.31494140625,
0.53173828125,
0.31494140625,
0.53173828125,
0.3193359375,
0.498291015625,
0.3193359375
],
[
0.26708984375,
0.33203125,
0.324462890625,
0.33203125,
0.324462890625,
0.343505859375,
0.26708984375,
0.343505859375
],
[
0.339111328125,
0.33203125,
0.406982421875,
0.33203125,
0.406982421875,
0.343505859375,
0.339111328125,
0.343505859375
],
[
0.560546875,
0.33203125,
0.62841796875,
0.33203125,
0.62841796875,
0.343505859375,
0.560546875,
0.343505859375
],
[
0.6748046875,
0.33203125,
0.73876953125,
0.33203125,
0.73876953125,
0.343505859375,
0.6748046875,
0.343505859375
],
[
0.2181396484375,
0.359130859375,
0.78759765625,
0.359130859375,
0.78759765625,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.21728515625,
0.373779296875,
0.7861328125,
0.373779296875,
0.7861328125,
0.386962890625,
0.21728515625,
0.386962890625
],
[
0.2181396484375,
0.38818359375,
0.7861328125,
0.38818359375,
0.7861328125,
0.401611328125,
0.2181396484375,
0.401611328125
],
[
0.2181396484375,
0.401611328125,
0.78759765625,
0.401611328125,
0.78759765625,
0.415283203125,
0.2181396484375,
0.415283203125
],
[
0.2197265625,
0.417236328125,
0.283447265625,
0.417236328125,
0.283447265625,
0.4287109375,
0.2197265625,
0.4287109375
],
[
0.2181396484375,
0.4609375,
0.78369140625,
0.4609375,
0.78369140625,
0.474853515625,
0.2181396484375,
0.474853515625
],
[
0.2181396484375,
0.4765625,
0.7861328125,
0.4765625,
0.7861328125,
0.490478515625,
0.2181396484375,
0.490478515625
],
[
0.2181396484375,
0.492431640625,
0.496826171875,
0.492431640625,
0.496826171875,
0.505859375,
0.2181396484375,
0.505859375
],
[
0.2181396484375,
0.537109375,
0.2484130859375,
0.537109375,
0.2484130859375,
0.5498046875,
0.2181396484375,
0.5498046875
],
[
0.253173828125,
0.54150390625,
0.2607421875,
0.54150390625,
0.2607421875,
0.54541015625,
0.253173828125,
0.54541015625
],
[
0.256591796875,
0.5361328125,
0.408447265625,
0.5361328125,
0.408447265625,
0.55029296875,
0.256591796875,
0.55029296875
],
[
0.2181396484375,
0.5693359375,
0.78369140625,
0.5693359375,
0.78369140625,
0.58349609375,
0.2181396484375,
0.58349609375
],
[
0.21728515625,
0.583984375,
0.78515625,
0.583984375,
0.78515625,
0.59765625,
0.21728515625,
0.59765625
],
[
0.2181396484375,
0.599609375,
0.7861328125,
0.599609375,
0.7861328125,
0.61279296875,
0.2181396484375,
0.61279296875
],
[
0.2181396484375,
0.6142578125,
0.78515625,
0.6142578125,
0.78515625,
0.62744140625,
0.2181396484375,
0.62744140625
],
[
0.2181396484375,
0.6298828125,
0.7861328125,
0.6298828125,
0.7861328125,
0.64306640625,
0.2181396484375,
0.64306640625
],
[
0.21728515625,
0.64599609375,
0.78515625,
0.64599609375,
0.78515625,
0.6591796875,
0.21728515625,
0.6591796875
],
[
0.2181396484375,
0.66015625,
0.7861328125,
0.66015625,
0.7861328125,
0.6748046875,
0.2181396484375,
0.6748046875
],
[
0.2181396484375,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.68798828125,
0.2181396484375,
0.68798828125
],
[
0.2181396484375,
0.68994140625,
0.7861328125,
0.68994140625,
0.7861328125,
0.7041015625,
0.2181396484375,
0.7041015625
],
[
0.21728515625,
0.705078125,
0.78759765625,
0.705078125,
0.78759765625,
0.71826171875,
0.21728515625,
0.71826171875
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2181396484375,
0.73486328125,
0.7861328125,
0.73486328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.21728515625,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.21728515625,
0.7646484375
],
[
0.21728515625,
0.7646484375,
0.35693359375,
0.76513671875,
0.35693359375,
0.77978515625,
0.21728515625,
0.77880859375
],
[
0.330810546875,
0.79541015625,
0.6708984375,
0.79541015625,
0.6708984375,
0.81005859375,
0.330810546875,
0.81005859375
],
[
0.76123046875,
0.79345703125,
0.78759765625,
0.79345703125,
0.78759765625,
0.8125,
0.76123046875,
0.8125
],
[
0.2181396484375,
0.8271484375,
0.48046875,
0.8271484375,
0.48046875,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Few-shot 3DReconstructionwith MemoryPrior ContrastiveNetwork",
"5",
"Memory Network",
"Prior Module",
"V[2]",
"Vinal",
"ey.ImageFeatur",
"Value:3DShapePrio",
"V[nk]",
"K",
"V[1]",
"",
"",
"",
"K[2]",
".",
"V[2] ",
"",
"K[nm",
"V[nm.",
"Kn",
"V(nk]",
"K[nn] [nm",
"K[na]",
"",
"V[n]",
"",
"Sn",
"caten",
"Image Feature",
"PnorFeatu",
"Input Image",
"Image Encoder",
"Shape Decode",
"Output shape",
"Fig. 2. An overview of the proposed MPCN. In the training stage, we only use base",
"categories to train the model, and insert memory slots into the memory network as",
"alternative prior according to the rules we set. In the test phase, the memory network",
"saves few shot of support set of the novel category to reconstruct 3D volumes of the",
"query set.",
"use D and D, for training. The support set D can be used as prior information.",
"We hope that the model can be designed to be category-agnostic and achieve",
"good performance in the query set Dg:",
"3.1",
"",
"Memory Network",
"In most previous works on single-view 3D reconstruction, the shape prior infor-",
"mation is learned from the model parameters, which leads to category collapse",
"when transferring to novel categories. As mentioned in [31], such kind of model",
"makes the problem degenerated into a classification task. To alleviate this issue,",
"directly using 3D shapes as priors is an intuitive and effective way. As shown",
"in Figure 2, we adopt an explicit key-value memory network to store and cal-",
"culate shape priors. In the training and testing stages, the CNN features of the",
"input image is extracted by a 2D encoder. Then a retrieval task is performed,",
"where the keys of the samples stored in the memory network are compared to",
"the query vector and the corresponding Top-k retrieved shapes are sent to the",
"prior module for generating prior features. Specifically, as shown in Eq. (1), the",
"input image Iq is first encoded by a 2D encoder E2D, then the image features",
"and shape prior features are concatenated. Finally, the 3D shape is inferred by",
"a 3D decoder D3D.",
"pr = D3D(Concatenate(E2D(Iq),prior feature)),",
"(1)",
"where pr is the final predict volume."
] |
{
"arxiv_id": "2208.00183",
"language": "en",
"page_idx": 4
}
|
||
2310.13394-en-1
|
2310.13394-en-1.jpg
|
objects has a significant impact on the behaviour of LLMs. In particular, LLMs tend to develop their mental representation of objects referred to based on the given context in prompts rather than relying on their internal weights, even if the given context information is incorrect.
To conclude, our contributions can be summa-
rized into three folds:
$\cdot$ We propose a simple but effective size comparison dataset to probe the real-world understanding of LLMs.
$\cdot$ We analyse the limits of the real-world understanding of LLMs with comprehensive prompt-based probing experiments.
· We discuss the vulnerability and the alignment
of the world knowledge of LLMs.
## 2 Background
2.1 World Models and World Knowledge
There has been a wide and long-lasting debate about whether LLMs really have their internal world models and to what extent their world knowledge aligns with humans. Mind's Eye (Liu et al., 2022b) proposed to augment language models with an external physical simulation engine for better understanding the physical phenomena. RAP (Hao et al., 2023) suggests LLMs as both a world model and a reasoning agent and includes a principled planning algorithm for strategic exploration in a vast reasoning space. Xiang et al. (2023) deploys an embodied agent in a world model to endow LLMs with a diverse set of embodied experiences by fine-tuning. Although the world model of LLM can be effectively augmented and they indeed display a certain level of real-world understanding, there is still a lack of sufficient study on the boundary of the world understanding of LLMs. Similarly to other research on LLMs inspired by cognitive science and psychology (Binz and Schulz, 2022; Bisk et al., 2020a; Mahowald et al., 2023; Prystawski et al., 2022), we propose to audit the real-world understanding of LLM with questions as simple as size comparison.
2.2 $\textbf{Physical World Understanding Datasets}$
PIQA (Bisk et al., 2020b) is a popular data set for physical commonsense reasoning to benchmark progress in physical commonsense understanding.
PIQA dataset consists of more than 16,000 training QA pairs, with additional 2K and 3K held for development and testing. The task is multiple choice question answering: Given a question and two possible solutions, a model or a human must choose the most appropriate solution, of which exactly one is correct. MMLU-Physics (Hendrycks et al., 2021) contains 206 samples of physics consisting of multiple choice questions at the college and high school level to evaluate the academic and professional understanding of the model in the physics domain. UTOPIA (Liu et al., 2022b) is a new multi-task physics alignment dataset that aims to benchmark how well current LMs can understand and areas over some basic laws of physics. It leverages a physics engine to generate data for 39 subtasks covering six common scenes that involve understanding basic principles of physics. PROST (Aroca-Ouellette et al., 2021) is a new probing dataset to evaluate the ability of pre-trained LMs to understand and reason about the physical world. It contains 18,736 multiple-choice questions made from 14 manually curated templates, covering 10 physical reasoning concepts. The existing datasets contain questions from different dimensions, but they fail to effectively evaluate some particular aspect of the understanding of the physical world of LLMs. Since we want to probe the effect of context for in-context learning, it is necessary to have a content-controllable and dimensionspecific dataset. Based on the requirements, we propose POSQA: a Physical Object Size QuestionAnswering dataset which is also designed to test the size understanding ability of LLMs on physical world objects. 3 POSQA
POSQA consists of 12,000 multiple choice questions designed to probe the physical world understanding ability of the language model in the size dimension. We design two types of questions, each of them containing 6,000 questions. Table 1 shows the statistics of POSQA. The size comparison covers 92 entities, ranging from proton to universe. The entity and size information are obtained from Nikon Universcale, which aims to allow people to see and understand the relative size of the full range of known objects in our universe. We design four manually written templates to construct the two types of size questions. We show the templates
https://www. nikon. com/about/sp/universcale/
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.51123046875,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1158447265625,
0.1341552734375,
0.48779296875,
0.133056640625,
0.48779296875,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.1175537109375,
0.15087890625,
0.489501953125,
0.15087890625,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51123046875,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.162109375,
0.51123046875,
0.162109375
],
[
0.1158447265625,
0.166259765625,
0.295654296875,
0.166259765625,
0.295654296875,
0.179931640625,
0.1158447265625,
0.179931640625
],
[
0.51318359375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1368408203125,
0.182861328125,
0.495361328125,
0.182861328125,
0.495361328125,
0.196533203125,
0.1368408203125,
0.196533203125
],
[
0.51318359375,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51318359375,
0.1953125
],
[
0.11419677734375,
0.1995849609375,
0.27783203125,
0.1995849609375,
0.27783203125,
0.213134765625,
0.11419677734375,
0.213134765625
],
[
0.51318359375,
0.19775390625,
0.8818359375,
0.19775390625,
0.8818359375,
0.2120361328125,
0.51318359375,
0.2120361328125
],
[
0.509765625,
0.213134765625,
0.880859375,
0.21435546875,
0.880859375,
0.2286376953125,
0.509765625,
0.22802734375
],
[
0.139404296875,
0.2286376953125,
0.49365234375,
0.2286376953125,
0.49365234375,
0.243408203125,
0.139404296875,
0.243408203125
],
[
0.509765625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.509765625,
0.24462890625
],
[
0.15283203125,
0.2452392578125,
0.49365234375,
0.2452392578125,
0.49365234375,
0.2587890625,
0.15283203125,
0.2587890625
],
[
0.5087890625,
0.2452392578125,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.5087890625,
0.2587890625
],
[
0.15283203125,
0.26123046875,
0.2939453125,
0.26123046875,
0.2939453125,
0.275634765625,
0.15283203125,
0.275634765625
],
[
0.51123046875,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.51123046875,
0.279052734375,
0.8818359375,
0.279052734375,
0.8818359375,
0.292236328125,
0.51123046875,
0.292236328125
],
[
0.1376953125,
0.289306640625,
0.49365234375,
0.29150390625,
0.49365234375,
0.305908203125,
0.1376953125,
0.303955078125
],
[
0.51123046875,
0.295654296875,
0.88330078125,
0.295654296875,
0.88330078125,
0.308837890625,
0.51123046875,
0.308837890625
],
[
0.15283203125,
0.30712890625,
0.489501953125,
0.30712890625,
0.489501953125,
0.32177734375,
0.15283203125,
0.32177734375
],
[
0.51318359375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.15283203125,
0.32373046875,
0.418212890625,
0.32373046875,
0.418212890625,
0.33740234375,
0.15283203125,
0.33740234375
],
[
0.51123046875,
0.325927734375,
0.88330078125,
0.32666015625,
0.88330078125,
0.341552734375,
0.51123046875,
0.34033203125
],
[
0.51318359375,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.1376953125,
0.35400390625,
0.48779296875,
0.35400390625,
0.48779296875,
0.36767578125,
0.1376953125,
0.36767578125
],
[
0.509765625,
0.358154296875,
0.88330078125,
0.359375,
0.88330078125,
0.37353515625,
0.509765625,
0.372314453125
],
[
0.15283203125,
0.369384765625,
0.405517578125,
0.369384765625,
0.405517578125,
0.38427734375,
0.15283203125,
0.38427734375
],
[
0.51123046875,
0.37353515625,
0.8818359375,
0.374755859375,
0.8818359375,
0.3896484375,
0.51123046875,
0.388427734375
],
[
0.51123046875,
0.391357421875,
0.8818359375,
0.39013671875,
0.8818359375,
0.405029296875,
0.51123046875,
0.40625
],
[
0.1158447265625,
0.399658203125,
0.255126953125,
0.399658203125,
0.255126953125,
0.41455078125,
0.1158447265625,
0.41455078125
],
[
0.51318359375,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.51318359375,
0.421630859375
],
[
0.1158447265625,
0.42578125,
0.450927734375,
0.427490234375,
0.450927734375,
0.4423828125,
0.1158447265625,
0.440673828125
],
[
0.51318359375,
0.42578125,
0.880859375,
0.42578125,
0.880859375,
0.435791015625,
0.51318359375,
0.435791015625
],
[
0.51123046875,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.51123046875,
0.45361328125
],
[
0.11419677734375,
0.448974609375,
0.486083984375,
0.449462890625,
0.486083984375,
0.46435546875,
0.11419677734375,
0.463134765625
],
[
0.51123046875,
0.4560546875,
0.880859375,
0.4560546875,
0.880859375,
0.47021484375,
0.51123046875,
0.47021484375
],
[
0.1158447265625,
0.466064453125,
0.486083984375,
0.466064453125,
0.486083984375,
0.479736328125,
0.1158447265625,
0.479736328125
],
[
0.509765625,
0.47265625,
0.88330078125,
0.47265625,
0.88330078125,
0.486328125,
0.509765625,
0.486328125
],
[
0.1175537109375,
0.482177734375,
0.490234375,
0.482177734375,
0.490234375,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.51123046875,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.501953125,
0.51123046875,
0.501953125
],
[
0.1175537109375,
0.498779296875,
0.489501953125,
0.49755859375,
0.489501953125,
0.5126953125,
0.1175537109375,
0.51318359375
],
[
0.51123046875,
0.5048828125,
0.88330078125,
0.5048828125,
0.88330078125,
0.5185546875,
0.51123046875,
0.5185546875
],
[
0.1175537109375,
0.51416015625,
0.48779296875,
0.51416015625,
0.48779296875,
0.529296875,
0.1175537109375,
0.529296875
],
[
0.5087890625,
0.5185546875,
0.8818359375,
0.51953125,
0.8818359375,
0.53369140625,
0.5087890625,
0.533203125
],
[
0.1175537109375,
0.53076171875,
0.48779296875,
0.53076171875,
0.48779296875,
0.5458984375,
0.1175537109375,
0.5458984375
],
[
0.51318359375,
0.5361328125,
0.88330078125,
0.5361328125,
0.88330078125,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.1158447265625,
0.54638671875,
0.48779296875,
0.5458984375,
0.48779296875,
0.56005859375,
0.1158447265625,
0.56103515625
],
[
0.509765625,
0.552734375,
0.8818359375,
0.552734375,
0.8818359375,
0.56640625,
0.509765625,
0.56640625
],
[
0.1158447265625,
0.5625,
0.484375,
0.5625,
0.484375,
0.576171875,
0.1158447265625,
0.576171875
],
[
0.51318359375,
0.568359375,
0.8818359375,
0.568359375,
0.8818359375,
0.5830078125,
0.51318359375,
0.5830078125
],
[
0.1158447265625,
0.5791015625,
0.48779296875,
0.57763671875,
0.48779296875,
0.5927734375,
0.1158447265625,
0.59326171875
],
[
0.51123046875,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.51123046875,
0.59716796875
],
[
0.1175537109375,
0.595703125,
0.489501953125,
0.595703125,
0.489501953125,
0.609375,
0.1175537109375,
0.609375
],
[
0.509765625,
0.6015625,
0.87890625,
0.6015625,
0.87890625,
0.6123046875,
0.509765625,
0.6123046875
],
[
0.1158447265625,
0.60986328125,
0.486083984375,
0.60986328125,
0.486083984375,
0.62353515625,
0.1158447265625,
0.62353515625
],
[
0.509765625,
0.61767578125,
0.8818359375,
0.61767578125,
0.8818359375,
0.63134765625,
0.509765625,
0.63134765625
],
[
0.1158447265625,
0.626953125,
0.48779296875,
0.626953125,
0.48779296875,
0.64013671875,
0.1158447265625,
0.64013671875
],
[
0.5087890625,
0.63134765625,
0.6171875,
0.6328125,
0.6171875,
0.64794921875,
0.5087890625,
0.6455078125
],
[
0.1175537109375,
0.6435546875,
0.48779296875,
0.6435546875,
0.48779296875,
0.65673828125,
0.1175537109375,
0.65673828125
],
[
0.11419677734375,
0.6591796875,
0.48779296875,
0.65771484375,
0.48779296875,
0.6728515625,
0.11419677734375,
0.67333984375
],
[
0.509765625,
0.6611328125,
0.61376953125,
0.6611328125,
0.61376953125,
0.67578125,
0.509765625,
0.67578125
],
[
0.1175537109375,
0.6748046875,
0.486083984375,
0.6748046875,
0.486083984375,
0.6884765625,
0.1175537109375,
0.6884765625
],
[
0.51123046875,
0.68603515625,
0.8818359375,
0.68701171875,
0.8818359375,
0.70166015625,
0.51123046875,
0.70068359375
],
[
0.1175537109375,
0.6923828125,
0.489501953125,
0.6923828125,
0.489501953125,
0.7060546875,
0.1175537109375,
0.7060546875
],
[
0.1158447265625,
0.70703125,
0.489501953125,
0.70703125,
0.489501953125,
0.72021484375,
0.1158447265625,
0.72021484375
],
[
0.509765625,
0.705078125,
0.884765625,
0.705078125,
0.884765625,
0.71875,
0.509765625,
0.71875
],
[
0.11419677734375,
0.7236328125,
0.48779296875,
0.72265625,
0.48779296875,
0.73681640625,
0.11419677734375,
0.73828125
],
[
0.509765625,
0.7197265625,
0.8818359375,
0.72021484375,
0.8818359375,
0.7353515625,
0.509765625,
0.73388671875
],
[
0.1158447265625,
0.74072265625,
0.491943359375,
0.74072265625,
0.491943359375,
0.75439453125,
0.1158447265625,
0.75439453125
],
[
0.51318359375,
0.73681640625,
0.8818359375,
0.73681640625,
0.8818359375,
0.751953125,
0.51318359375,
0.751953125
],
[
0.1158447265625,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.7685546875,
0.1158447265625,
0.7685546875
],
[
0.51123046875,
0.7529296875,
0.8818359375,
0.7529296875,
0.8818359375,
0.7666015625,
0.51123046875,
0.7666015625
],
[
0.1175537109375,
0.771484375,
0.489501953125,
0.771484375,
0.489501953125,
0.78515625,
0.1175537109375,
0.78515625
],
[
0.51318359375,
0.77099609375,
0.880859375,
0.77099609375,
0.880859375,
0.78076171875,
0.51318359375,
0.78076171875
],
[
0.1175537109375,
0.7880859375,
0.489501953125,
0.7880859375,
0.489501953125,
0.8017578125,
0.1175537109375,
0.8017578125
],
[
0.51123046875,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51123046875,
0.798828125
],
[
0.11419677734375,
0.802734375,
0.489501953125,
0.80419921875,
0.489501953125,
0.818359375,
0.11419677734375,
0.81787109375
],
[
0.509765625,
0.80029296875,
0.8818359375,
0.80029296875,
0.8818359375,
0.81396484375,
0.509765625,
0.81396484375
],
[
0.1175537109375,
0.82177734375,
0.2435302734375,
0.82177734375,
0.2435302734375,
0.8349609375,
0.1175537109375,
0.8349609375
],
[
0.51318359375,
0.81787109375,
0.8818359375,
0.81787109375,
0.8818359375,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.509765625,
0.8330078125,
0.8818359375,
0.8330078125,
0.8818359375,
0.8466796875,
0.509765625,
0.8466796875
],
[
0.1158447265625,
0.8486328125,
0.470947265625,
0.8486328125,
0.470947265625,
0.86328125,
0.1158447265625,
0.86328125
],
[
0.51123046875,
0.849609375,
0.8818359375,
0.849609375,
0.8818359375,
0.86328125,
0.51123046875,
0.86328125
],
[
0.1158447265625,
0.87060546875,
0.48779296875,
0.87158203125,
0.48779296875,
0.8857421875,
0.1158447265625,
0.88525390625
],
[
0.51123046875,
0.865234375,
0.880859375,
0.865234375,
0.880859375,
0.87890625,
0.51123046875,
0.87890625
],
[
0.5087890625,
0.8818359375,
0.880859375,
0.8818359375,
0.880859375,
0.8955078125,
0.5087890625,
0.8955078125
],
[
0.1158447265625,
0.88818359375,
0.489501953125,
0.88720703125,
0.489501953125,
0.90185546875,
0.1158447265625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.53662109375,
0.90478515625,
0.86572265625,
0.90478515625,
0.86572265625,
0.91845703125,
0.53662109375,
0.91845703125
]
] |
[
"objects has a significant impact on the behaviour",
"PIQA dataset consists of more than 16,000 train-",
"of LLMs. In particular, LLMs tend to develop their",
"ing QA pairs, with additional 2K and 3K held",
"mental representation of objects referred to based",
"for development and testing: The task is multiple",
"on the given context in prompts rather than relying",
"choice question answering: Given a question and",
"on their internal weights, even if the given context",
"two possible solutions, a model or a human must",
"information is incorrect.",
"choose the most appropriate solution, of which ex",
"To conclude, our contributions can be summa-",
"actly one is correct. MMLU-Physics (Hendrycks",
"rized into three folds:",
"et al., 2021) contains 206 samples of physics con-",
"sisting of multiple choice questions at the college",
": We propose a simple but effective size com-",
" and high school level to evaluate the academic",
"parison dataset to probe the real-world under-",
"and professional understanding of the model in",
"standing of LLMs.",
"the physics domain. UTOPIA (Liu et al., 2022b)",
"is a new multi-task physics alignment dataset that",
". We analyse the limits of the real-world un-",
"aims to benchmark how well current LMs can un-",
"derstanding of LLMs with comprehensive",
"derstand and areas over some basic laws of physics.",
"prompt-based probing experiments.",
"It leverages a physics engine to generate data for",
"39 subtasks covering six common scenes that in",
": We discuss the vulnerability and the alignment",
"volve understanding basic principles of physics.",
"of the world knowledge of LLMs.",
"PROST (Aroca-Ouellette et al., 2021) is a new",
"probing dataset to evaluate the ability of pre-trained",
"2Background",
"LMs to understand and reason about the physical",
"2.1 World Models and World Knowledge",
"world.Itcontains18.736multiple-choiceguestions",
"made from 14 manually curated templates, cover-",
"There has been a wide and long-lasting debate",
"ing 10 physical reasoning concepts. The existing",
"about whether LLMs really have their internal",
" datasets contain questions from different dimen-",
"world models and to what extent their world knowl-",
"sions, but they fail to effectively evaluate some",
"edge aligns with humans. Mind's Eye (Liu et al.,",
"particular aspect of the understanding of the phys-",
"2022b) proposed to augment language models with",
"ical world of LLMs. Since we want to probe the",
"an external physical simulation engine for better",
"effect of context for in-context learning, it is neces-",
"understanding the physical phenomena. RAP (Hao",
" sary to have a content-controllable and dimension",
"et al., 2023) suggests LLMs as both a world mode",
"specific dataset. Based on the requirements, we",
"and a reasoning agent and includes a principled",
"propose POSQA: a Physical Object Size Question-",
"planning algorithm for strategic exploration in a",
"Answering dataset which is also designed to test",
"vast reasoning space. Xiang et al. (2023) deploys",
"the size understanding ability of LLMs on physical",
"an embodied agent in a world model to endow",
"world objects.",
"LLMs with a diverse set of embodied experiences",
"by fine-tuning. Although the world model of LLM",
"3POSQA",
"can be effectively augmented and they indeed dis-",
"POSQA consists of 12,000 multiple choice ques-",
"play a certain level of real-world understanding,",
"there is still a lack of sufficient study on the bound-",
"tions designed to probe the physical world under-",
"ary of the world understanding of LLMs. Similarly",
"standing ability of the language model in the size",
"to other research on LLMs inspired by cognitive sci-",
"dimension. We design two types of questions, each",
"ence and psychology (Binz and Schulz, 2022: Bisk",
"of them containing 6,000 questions. Table 1 shows",
"et al., 2020a; Mahowald et al., 2023; Prystawski",
"thestatisticsofPOSOA.Thesizecomparisoncov-",
"et al., 2022), we propose to audit the real-world",
"ers 92 entities, ranging from proton to universe.",
"understanding of LLM with questions as simple as",
"The entity and size information are obtained from",
"size comparison.",
"Nikon Universcale, which aims to allow people",
"to see and understand the relative size of the full",
"2.2Physical World Understanding Datasets",
"range of known objects in our universe. We design",
"PIQA (Bisk et al., 2020b) is a popular data set",
"four manually written templates to construct the",
" two types of size questions. We show the templates",
"for physical commonsense reasoning to benchmark",
"progress in physical commonsense understanding.",
"https://www.nikon.com/about/sp/universcale/"
] |
{
"arxiv_id": "2310.13394",
"language": "en",
"page_idx": 1
}
|
||
2310.13394-en-2
|
2310.13394-en-2.jpg
|
<table>
<tbody>
<tr>
<th>Question Type</th>
<th>Bigger</th>
<th>Smaller</th>
<th>Total</th>
</tr>
<tr>
<td>General Question</td>
<td>3,000</td>
<td>3,000</td>
<td>$\overline{6,000}$</td>
</tr>
<tr>
<td>Special Ouestion</td>
<td>3,000</td>
<td>3,000</td>
<td>6,000</td>
</tr>
</tbody>
</table>
Table 1: The statistics of POSQA.
in detail below.
General Question A general question requires the answer"yes" or"no". We use two templates to generate general questions.
TEMPLATE 1: Is Entity A bigger than Entity B? TEMPLATE 2: Is Entity A smaller than Entity B? We replace Entity A and Entity B with different entity names. For each template, we use the same Entity A-Entity B pair to generate a question, which is to avoid introducing bias. We generate 3,000 questions for each template, so there are 6,000 general questions in total. Based on the actual size of each entity, we label each question with "yes" or "no". The general questions aim to evaluate the size knowledge of objects contained in the LMs and the understanding of LMs on yes/no labels.
Special Question A special question begins with an interrogative word “which”. We design two templates for special questions.
TEMPLATE 3: Which one is bigger between
EntityA and Entity B?
TEMPLATE 4: Which one is smaller between
EntityA and $EntityB?$
Similarly, we also use the same Entity A-Entity $B$ pair to generate a question on each template. We generated a total of 6,000 questions, 3,000 questions for each template. The label of a special question is different from the general question. In the special question, we label each question with the actual entity name, which is exactly the same as Entity A or Entity B. The special questions are intended to test the understanding of the size of the LMs and the understanding of the LMs about the interrogative word'which' of the question.
Entity Feature We collect the features of the 92 entities, including scale, size, magnitude, and text. The scale feature stores the size information of the entity in a specific size unit. For example, the scale of the Solar System is 9 billion km and the scale of an Atom is 100 pm. The size feature stores the absolute value of the size of the entity representing in scientific notation. The magnitude
feature represents the exponent of size which is stored in the size feature. The text feature contains a textual description of the entity.
# 4 Methodology
In this section, we cover the details of the proposed approach, first describing the designed prompt in Section 4.1, followed by the models used in the experiments in Section 4.2 and the introduction of the evaluation methods in Section 4.3.
# 4.1 Prompt Design
We construct different prompts that are used in our
experiments.
## 4.1.1 Plain Question Prompt
Plain question prompt is to query the model with a single plain question without any hint or knowledge. We aim to test how the model performs on POSQA without any external auxiliary information. The model answers the query purely based on the knowledge stored within its weights.
4.1.2 Relevant Knowledge-augmented Prompt
External knowledge has been shown to be helpful for various NLP tasks, including common sense reasoning (Liu et al., 2022a). We consider two kinds of knowledge in our experiments:(1)Exact Size Information from POSQA (2) Generated Knowledge from GPT-3. Knowledge is considered as the context that is concatenated with a question. We use the knowledge-augmented prompt to query LLMs to see how the context would affect the model's prediction.
Exact Size Information The exact size information of each object in POSQA is stored as an entity feature. For a size comparison question, first we retrieve the exact size the two entities respectively, then we rewrite the original prompt to a knowledgeaugmented prompt. In particular, we add a sentence to describe the size of the two objects before the question. The sentence is: The size of Entity A is $Exact\textit{Size of Entity A. The size of Entity B is Exact}$ Size of Entity B.
Generated Knowledge We generate the entityrelated knowledge statement by querying an LLM. We consider two types of knowledge of entities: (1) general knowledge, which describes the general information of an entity; (2) size knowledge, which describes the size information of an entity. The purpose is to investigate which knowledge is more
|
[
[
0.130126953125,
0.0885009765625,
0.2418212890625,
0.0885009765625,
0.2418212890625,
0.10272216796875,
0.130126953125,
0.10272216796875
],
[
0.279541015625,
0.0885009765625,
0.3359375,
0.0885009765625,
0.3359375,
0.10272216796875,
0.279541015625,
0.10272216796875
],
[
0.34765625,
0.0885009765625,
0.411376953125,
0.0885009765625,
0.411376953125,
0.10272216796875,
0.34765625,
0.10272216796875
],
[
0.426513671875,
0.0885009765625,
0.47265625,
0.0885009765625,
0.47265625,
0.10272216796875,
0.426513671875,
0.10272216796875
],
[
0.51318359375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.1275634765625,
0.10394287109375,
0.26611328125,
0.105712890625,
0.26611328125,
0.119384765625,
0.1275634765625,
0.1175537109375
],
[
0.355224609375,
0.10394287109375,
0.405517578125,
0.10394287109375,
0.405517578125,
0.119384765625,
0.355224609375,
0.119384765625
],
[
0.51123046875,
0.1015625,
0.87890625,
0.1015625,
0.87890625,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.4248046875,
0.1051025390625,
0.47607421875,
0.1051025390625,
0.47607421875,
0.119384765625,
0.4248046875,
0.119384765625
],
[
0.2822265625,
0.1051025390625,
0.33154296875,
0.1051025390625,
0.33154296875,
0.119384765625,
0.2822265625,
0.119384765625
],
[
0.28369140625,
0.12054443359375,
0.330078125,
0.12054443359375,
0.330078125,
0.135986328125,
0.28369140625,
0.135986328125
],
[
0.355224609375,
0.12054443359375,
0.405517578125,
0.12054443359375,
0.405517578125,
0.1353759765625,
0.355224609375,
0.1353759765625
],
[
0.4248046875,
0.12054443359375,
0.474365234375,
0.12054443359375,
0.474365234375,
0.135986328125,
0.4248046875,
0.135986328125
],
[
0.51123046875,
0.1181640625,
0.75830078125,
0.1181640625,
0.75830078125,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.51123046875,
0.142578125,
0.65576171875,
0.144287109375,
0.65576171875,
0.162109375,
0.51123046875,
0.1602783203125
],
[
0.130126953125,
0.1217041015625,
0.26123046875,
0.1217041015625,
0.26123046875,
0.1353759765625,
0.130126953125,
0.1353759765625
],
[
0.183837890625,
0.149658203125,
0.41650390625,
0.15087890625,
0.41650390625,
0.1650390625,
0.183837890625,
0.1644287109375
],
[
0.51318359375,
0.171630859375,
0.8818359375,
0.171630859375,
0.8818359375,
0.185302734375,
0.51318359375,
0.185302734375
],
[
0.509765625,
0.18701171875,
0.880859375,
0.18701171875,
0.880859375,
0.20068359375,
0.509765625,
0.20068359375
],
[
0.1175537109375,
0.19482421875,
0.231689453125,
0.19482421875,
0.231689453125,
0.207275390625,
0.1175537109375,
0.207275390625
],
[
0.51123046875,
0.2037353515625,
0.880859375,
0.2037353515625,
0.880859375,
0.21728515625,
0.51123046875,
0.21728515625
],
[
0.1175537109375,
0.2197265625,
0.490234375,
0.2197265625,
0.490234375,
0.2340087890625,
0.1175537109375,
0.2340087890625
],
[
0.51123046875,
0.2197265625,
0.8818359375,
0.218505859375,
0.8818359375,
0.2327880859375,
0.51123046875,
0.2340087890625
],
[
0.1175537109375,
0.236328125,
0.489501953125,
0.236328125,
0.489501953125,
0.25048828125,
0.1175537109375,
0.25048828125
],
[
0.51318359375,
0.2369384765625,
0.79248046875,
0.2369384765625,
0.79248046875,
0.2476806640625,
0.51318359375,
0.2476806640625
],
[
0.1175537109375,
0.2529296875,
0.318115234375,
0.2529296875,
0.318115234375,
0.2666015625,
0.1175537109375,
0.2666015625
],
[
0.51318359375,
0.262451171875,
0.67333984375,
0.262451171875,
0.67333984375,
0.276611328125,
0.51318359375,
0.276611328125
],
[
0.1368408203125,
0.268310546875,
0.48779296875,
0.268310546875,
0.48779296875,
0.281982421875,
0.1368408203125,
0.281982421875
],
[
0.1351318359375,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.1351318359375,
0.297607421875
],
[
0.5087890625,
0.281982421875,
0.8818359375,
0.283203125,
0.8818359375,
0.297607421875,
0.5087890625,
0.29638671875
],
[
0.1368408203125,
0.3017578125,
0.489501953125,
0.3017578125,
0.489501953125,
0.3154296875,
0.1368408203125,
0.3154296875
],
[
0.51123046875,
0.2998046875,
0.60546875,
0.2998046875,
0.60546875,
0.311279296875,
0.51123046875,
0.311279296875
],
[
0.1175537109375,
0.318359375,
0.48779296875,
0.318359375,
0.48779296875,
0.33203125,
0.1175537109375,
0.33203125
],
[
0.51123046875,
0.32177734375,
0.75,
0.32373046875,
0.75,
0.33837890625,
0.51123046875,
0.336181640625
],
[
0.1158447265625,
0.333251953125,
0.48779296875,
0.334228515625,
0.48779296875,
0.3486328125,
0.1158447265625,
0.347412109375
],
[
0.51123046875,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.358154296875,
0.51123046875,
0.358154296875
],
[
0.118408203125,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.51318359375,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.51318359375,
0.37353515625
],
[
0.1175537109375,
0.366455078125,
0.489501953125,
0.366455078125,
0.489501953125,
0.380126953125,
0.1175537109375,
0.380126953125
],
[
0.51318359375,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.51318359375,
0.3896484375
],
[
0.1175537109375,
0.3818359375,
0.48779296875,
0.3818359375,
0.48779296875,
0.3955078125,
0.1175537109375,
0.3955078125
],
[
0.11419677734375,
0.3984375,
0.486083984375,
0.3984375,
0.486083984375,
0.412109375,
0.11419677734375,
0.412109375
],
[
0.5146484375,
0.393798828125,
0.880859375,
0.393798828125,
0.880859375,
0.40380859375,
0.5146484375,
0.40380859375
],
[
0.509765625,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.509765625,
0.421630859375
],
[
0.1175537109375,
0.4150390625,
0.48779296875,
0.4150390625,
0.48779296875,
0.4287109375,
0.1175537109375,
0.4287109375
],
[
0.51123046875,
0.424560546875,
0.78076171875,
0.424560546875,
0.78076171875,
0.438232421875,
0.51123046875,
0.438232421875
],
[
0.1175537109375,
0.432861328125,
0.484375,
0.432861328125,
0.484375,
0.443603515625,
0.1175537109375,
0.443603515625
],
[
0.11419677734375,
0.4453125,
0.489501953125,
0.446533203125,
0.489501953125,
0.46142578125,
0.11419677734375,
0.460205078125
],
[
0.51318359375,
0.448974609375,
0.880859375,
0.448974609375,
0.880859375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.11419677734375,
0.46142578125,
0.16796875,
0.463134765625,
0.166259765625,
0.47802734375,
0.11248779296875,
0.4755859375
],
[
0.51318359375,
0.46728515625,
0.880859375,
0.46728515625,
0.880859375,
0.48095703125,
0.51318359375,
0.48095703125
],
[
0.51123046875,
0.482666015625,
0.880859375,
0.48388671875,
0.880859375,
0.498779296875,
0.51123046875,
0.49755859375
],
[
0.1158447265625,
0.4892578125,
0.48779296875,
0.48681640625,
0.48779296875,
0.501953125,
0.1158447265625,
0.50439453125
],
[
0.11419677734375,
0.5048828125,
0.48779296875,
0.505859375,
0.48779296875,
0.52099609375,
0.11419677734375,
0.51953125
],
[
0.51123046875,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.51123046875,
0.51416015625
],
[
0.51318359375,
0.5166015625,
0.880859375,
0.5166015625,
0.880859375,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.1175537109375,
0.5224609375,
0.353515625,
0.5224609375,
0.353515625,
0.5361328125,
0.1175537109375,
0.5361328125
],
[
0.1368408203125,
0.5380859375,
0.48779296875,
0.5380859375,
0.48779296875,
0.552734375,
0.1368408203125,
0.552734375
],
[
0.5146484375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54443359375,
0.5146484375,
0.54443359375
],
[
0.51123046875,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5625,
0.51123046875,
0.5625
],
[
0.1175537109375,
0.5546875,
0.285400390625,
0.5546875,
0.285400390625,
0.568359375,
0.1175537109375,
0.568359375
],
[
0.133544921875,
0.5693359375,
0.486083984375,
0.5693359375,
0.486083984375,
0.5830078125,
0.133544921875,
0.5830078125
],
[
0.51123046875,
0.5654296875,
0.8818359375,
0.5654296875,
0.8818359375,
0.5791015625,
0.51123046875,
0.5791015625
],
[
0.509765625,
0.580078125,
0.880859375,
0.580078125,
0.880859375,
0.59326171875,
0.509765625,
0.59326171875
],
[
0.1175537109375,
0.58740234375,
0.285400390625,
0.58740234375,
0.285400390625,
0.60107421875,
0.1175537109375,
0.60107421875
],
[
0.51123046875,
0.5966796875,
0.8818359375,
0.5966796875,
0.8818359375,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.1351318359375,
0.60400390625,
0.48779296875,
0.60400390625,
0.48779296875,
0.61767578125,
0.1351318359375,
0.61767578125
],
[
0.509765625,
0.61328125,
0.68603515625,
0.61328125,
0.68603515625,
0.626953125,
0.509765625,
0.626953125
],
[
0.1175537109375,
0.62060546875,
0.48779296875,
0.62060546875,
0.48779296875,
0.63427734375,
0.1175537109375,
0.63427734375
],
[
0.118408203125,
0.6357421875,
0.491943359375,
0.6357421875,
0.491943359375,
0.65087890625,
0.118408203125,
0.65087890625
],
[
0.51318359375,
0.63720703125,
0.8818359375,
0.63720703125,
0.8818359375,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.1175537109375,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.1175537109375,
0.66650390625
],
[
0.5087890625,
0.65283203125,
0.8818359375,
0.65380859375,
0.8818359375,
0.66845703125,
0.5087890625,
0.66748046875
],
[
0.1175537109375,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.509765625,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68408203125,
0.509765625,
0.68408203125
],
[
0.1175537109375,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.1175537109375,
0.69873046875
],
[
0.509765625,
0.68603515625,
0.8818359375,
0.68701171875,
0.8818359375,
0.70068359375,
0.509765625,
0.69970703125
],
[
0.1158447265625,
0.69970703125,
0.486083984375,
0.69970703125,
0.486083984375,
0.71337890625,
0.1158447265625,
0.71337890625
],
[
0.51318359375,
0.703125,
0.8818359375,
0.703125,
0.8818359375,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.118408203125,
0.71875,
0.486083984375,
0.71875,
0.486083984375,
0.728515625,
0.118408203125,
0.728515625
],
[
0.51318359375,
0.7197265625,
0.87890625,
0.7197265625,
0.87890625,
0.72998046875,
0.51318359375,
0.72998046875
],
[
0.1158447265625,
0.73291015625,
0.489501953125,
0.73291015625,
0.489501953125,
0.74658203125,
0.1158447265625,
0.74658203125
],
[
0.51123046875,
0.7353515625,
0.87744140625,
0.7353515625,
0.87744140625,
0.7451171875,
0.51123046875,
0.7451171875
],
[
0.1175537109375,
0.7490234375,
0.48779296875,
0.7490234375,
0.48779296875,
0.7626953125,
0.1175537109375,
0.7626953125
],
[
0.509765625,
0.75048828125,
0.880859375,
0.75048828125,
0.880859375,
0.76416015625,
0.509765625,
0.76416015625
],
[
0.1158447265625,
0.76416015625,
0.43408203125,
0.76416015625,
0.43408203125,
0.779296875,
0.1158447265625,
0.779296875
],
[
0.5146484375,
0.7685546875,
0.880859375,
0.7685546875,
0.880859375,
0.779296875,
0.5146484375,
0.779296875
],
[
0.509765625,
0.7822265625,
0.62890625,
0.7822265625,
0.62890625,
0.7958984375,
0.509765625,
0.7958984375
],
[
0.1158447265625,
0.79052734375,
0.484375,
0.79052734375,
0.484375,
0.80419921875,
0.1158447265625,
0.80419921875
],
[
0.1158447265625,
0.8056640625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82080078125
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.118408203125,
0.826171875,
0.48779296875,
0.826171875,
0.48779296875,
0.8359375,
0.118408203125,
0.8359375
],
[
0.5087890625,
0.82373046875,
0.88330078125,
0.822265625,
0.88330078125,
0.83740234375,
0.5087890625,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51123046875,
0.84033203125,
0.88330078125,
0.84033203125,
0.88330078125,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51123046875,
0.85693359375,
0.88330078125,
0.85693359375,
0.88330078125,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1158447265625,
0.88720703125,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90283203125,
0.8818359375,
0.91748046875,
0.509765625,
0.91943359375
]
] |
[
"Question Type",
"Bigger",
"Smaller",
"Total",
"feature represents the exponent of size which is",
"GeneralQuestion",
"3,000",
"stored in the size feature. The text feature contains",
"6,000",
"3,000",
"3,000",
"3,000",
"6,000",
"a textual description of the entity.",
"4Methodology",
"Special Question",
"Table 1: The statistics of POSQA.",
"In this section, we cover the details of the proposed",
" approach, first describing the designed prompt in",
"in detail below.",
"Section 4.l, followed by the models used in the",
"General Question A general question requires",
"experiments in Section 4.2 and the introduction of",
"the answer \"yes\" or \"no\". We use two templates to",
"the evaluation methods in Section 4.3.",
"generate general questions.",
"4.1 Prompt Design",
"TEMPLATE I: Is Entity A bigger than Entity B?",
"TEMPLATE 2: Is Entity A smaller than Entity B?",
"We construct different prompts that are used in our",
"We replace Entity A and Entity B with differ-",
"experiments",
"ent entity names. For each template, we use the",
"4.1.1 Plain Question Prompt",
"same Entity A - Entity B pair to generate a ques-",
"Plain question prompt is to query the model with",
"tion, which is to avoid introducing bias. We gen-",
"a single plain question without any hint or knowl",
"erate 3,000 questions for each template, so there",
"edge. We aim to test how the model performs on",
"are 6,000 general questions in total. Based on the",
"actual size of each entity, we label each question",
"POSOAwithoutanyexternalauxiliaryinformation",
"The model answers the query purely based on the",
"with \"yes\" or\"no\". The general questions aim to",
"knowledge stored within its weights.",
"evaluatethe sizeknowledgeof objects contained in",
"the LMs and the understanding of LMs on yes/no",
"4.1.2Relevant Knowledge-augmented Prompt",
"labels.",
"External knowledge has been shown to be helpful",
"for various NLP tasks, including common sense",
"Special Question A special question begins with",
"an interrogative word \"which\". We design two",
"reasoning (Liu et al., 2022a). We consider two",
"kinds of knowledge in our experiments: (1) Ex-",
"templates for special questions.",
"TEMPLATE 3: Which one is bigger between",
"actSizeInformationfromPOSOA(2)Generated",
"Knowledge from GPT-3. Knowledge is considered",
"Entity A and Entity B?",
"TEMPLATE 4: Which one is smaller between",
"as the context that is concatenated with a ques-",
"tion. We use the knowledge-augmented prompt to",
"Entity A and Entity B?",
"query LLMs to see how the context would affect",
"Similarly, we also use the same Entity A-Entity",
"the model's prediction.",
"B pair to generate a question on each template. We",
"generated a total of 6,000 questions, 3,000 ques-",
"Exact Size InformationThe exact size informa-",
"tions for each template. The label of a special",
"tion of each object in POSQA is stored as an entity",
"question is different from the general question. In",
"feature. For a size comparison question, first we",
"the special question, we label each question with",
"retrieve the exact size the two entities respectively,",
"the actual entity name, which is exactly the same",
"then werewrite the original promptto a knowledge-",
"asEntityA orEntityB.Thespecial questions are",
"augmented promptInparticular.weaddasentence",
"intended to test the understanding of the size of the",
"to describe the size of the two obiects before the",
"LMs and the understanding of the LMs about the",
"question. The sentence is: The size of Entity A is",
"interrogative word 'which' of the question.",
"ExactSizeof EntityA.The sizeofEntityBisExact",
"Size of EntityB",
"Entity Feature We collect the features of the",
"92 entities, including scale, size, magnitude, and",
"Generated KnowledgeWe generate the entity-",
"text.The scalefeature stores the size information",
"related knowledge statement by querying an LLM.",
"of the entity in a specific size unit. For example,",
"We consider two types of knowledge of entities: (1)",
"the scale of the Solar System is 9 billion km and",
"general knowledge, which describes the general in-",
"the scale of an Atom is 100 pm. The size feature",
"formation of an entity; (2) size knowledge, which",
"stores the absolute value of the size of the entity",
"describes the size information of an entity.The",
"representing in scientific notation. The magnitude",
"purpose is to investigate which knowledge is more"
] |
{
"arxiv_id": "2310.13394",
"language": "en",
"page_idx": 2
}
|
||
2310.13394-en-4
|
2310.13394-en-4.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{Prompt}$ Metric</th>
<th colspan="2">FT5-Small</th>
<th colspan="2">FT5-Base</th>
<th colspan="2">FT5-Large</th>
<th colspan="2">FT5- XL</th>
<th colspan="2">GPT3-Davinci</th>
<th colspan="2">GPT3.5-Turbo</th>
</tr>
<tr>
<th>Acc</th>
<th>Macro- F1</th>
<th>Acc</th>
<th>Macro-F1</th>
<th>Acc</th>
<th>Macro- F1</th>
<th>Acc</th>
<th>Macro- F1</th>
<th>Acc</th>
<th>Macro- F1</th>
<th>Acc</th>
<th>Macro- F1</th>
</tr>
<tr>
<td>Plain Questiom</td>
<td>0.51</td>
<td>0.34</td>
<td>$\overline{0.50}$</td>
<td>0.35</td>
<td>$\overline{0.56}$</td>
<td>0.54</td>
<td>$\overline{0.58}$</td>
<td>0.51</td>
<td>0.52</td>
<td>0.38</td>
<td>0.69</td>
<td>0.69</td>
</tr>
<tr>
<td>+General Knowledge Information</td>
<td>0.50</td>
<td>0.34</td>
<td>0.52</td>
<td>0.52</td>
<td>0.64</td>
<td>0.63</td>
<td>0.67</td>
<td>0.67</td>
<td>0.51</td>
<td>0.35</td>
<td>0.79</td>
<td>0.79</td>
</tr>
<tr>
<td>+Size Knowledge Information</td>
<td>0.50</td>
<td>0.40</td>
<td>0.55</td>
<td>0.53</td>
<td>0.63</td>
<td>0.61</td>
<td>0.76</td>
<td>0.76</td>
<td>0.56</td>
<td>0.48</td>
<td>0.85</td>
<td>0.85</td>
</tr>
<tr>
<td>+Exact Size Information</td>
<td>0.50</td>
<td>0.50</td>
<td>0.50</td>
<td>0.54</td>
<td>0.62</td>
<td>0.77</td>
<td>0.87</td>
<td>0.88</td>
<td>0.73</td>
<td>0.89</td>
<td>0.90</td>
<td>0.89</td>
</tr>
<tr>
<td>+Only Head Entity Gold Size</td>
<td>0.50</td>
<td>0.34</td>
<td>0.54</td>
<td>0.50</td>
<td>0.63</td>
<td>0.60</td>
<td>0.64</td>
<td>0.64</td>
<td>0.51</td>
<td>0.35</td>
<td>0.69</td>
<td>0.69</td>
</tr>
<tr>
<td>+Only Tail Entity Gold Size</td>
<td>0.50</td>
<td>0.34</td>
<td>0.49</td>
<td>0.34</td>
<td>0.55</td>
<td>0.54</td>
<td>0.66</td>
<td>0.63</td>
<td>0.51</td>
<td>0.34</td>
<td>0.75</td>
<td>0.74</td>
</tr>
<tr>
<td>+Masking Size Information</td>
<td>0.50</td>
<td>0.49</td>
<td>0.49</td>
<td>0.34</td>
<td>0.53</td>
<td>0.41</td>
<td>0.74</td>
<td>0.74</td>
<td>0.53</td>
<td>0.39</td>
<td>0.80</td>
<td>0.80</td>
</tr>
<tr>
<td>+Masking Entity Information</td>
<td>0.50</td>
<td>0.33</td>
<td>0.50</td>
<td>0.37</td>
<td>0.49</td>
<td>0.42</td>
<td>0.63</td>
<td>0.58</td>
<td>0.52</td>
<td>0.46</td>
<td>0.70</td>
<td>0.69</td>
</tr>
<tr>
<td>+Counterfactual Size Information</td>
<td>0.50</td>
<td>0.47</td>
<td>0.49</td>
<td>0.33</td>
<td>0.47</td>
<td>0.33</td>
<td>0.23</td>
<td>0.23</td>
<td>0.42</td>
<td>0.41</td>
<td>0.38</td>
<td>0.38</td>
</tr>
</tbody>
</table>
Table 2: The results of using different prompt settings on various models on the general question of POSQA.
### 5 Results and Analvsis
### 5.1 Baseline
<FigureHere>
Figure 1: The results of CER and CMR using different knowledge-augmented prompts (except for Counterfactual Size Information) on various models on general question of POSQA.
The baseline is to query LLMs using the Plain Question prompt and the results are presented in Table 2 for general questions and Table 3 for special questions. When the number of parameters of the model exceeds 250M (Flan-T5-Base), the ability to answer the special size comparison questions begins to emerge. GPT3.5-Turbo achieves the best accuracy score and Macro-F1 score on both types of questions. It is surprising that significantly smaller models, such as Flan-T5-Large (780M) and Flan-T5-XL (3B), exhibit superior performance. For example, Flan-T5-XL outperforms GPT3-Davinci (175B) by 0.06 precision in answering general questions and 0.03 precision in answering special questions. In general, LLMs perform better at special questions. For example, the accuracy increases by 0.14 on Flan-T5-XL and 0.04 on GPT3.5-Turbo, respectively, from answering general questions to special questions.
Our empirical investigation also reveals that the GPT3-Davinci model tends to provide an initial incorrect answer, despite subsequently offering a correct explanation for the given question. This phenomenon occurs especially when directly querying GPT3-Davinci with general questions to get the "yes" or"no" answers. We speculate that it could be attributed to the use of first-word sampling techniques during the decoding process. However, this phenomenon does not occur in GPT3.5-Turbo which has been optimised for dialogue scenario. Even GPT3.5-Turbo, one of the most powerful LLMs, can only achieve an average 0.71 accuracy score on these two types of questions. This suggests that although larger LMs may possess certain advantages in most situations, they may still lack real-world understanding when it comes to answering basic size comparison questions.
used metrics for the evaluation of the performance of the model in classification tasks. Accuracy considers global precision and recall of the categories, while Macro-F1 computes the average of the F1 scores obtained by individual categories.
To explore the influence of context on the prompt, we propose two quantitative evaluation metrics: Context Effective Rate (CER) and Context Misleading Rate (CMR). We calculate CER and CMR by comparing the output of a model using a prompt that contains context information with the output of a model using a prompt that does not contain any context information. Specifically, CER evaluates how many incorrectly answered questions can be correctly answered after adding the context in the prompt. CMR evaluates how many correctly answered questions can be incorrectly answered after adding the context to the prompt.
|
[
[
0.12005615234375,
0.08612060546875,
0.16796875,
0.08612060546875,
0.16796875,
0.097412109375,
0.12005615234375,
0.097412109375
],
[
0.318115234375,
0.08612060546875,
0.3779296875,
0.08612060546875,
0.3779296875,
0.0968017578125,
0.318115234375,
0.0968017578125
],
[
0.418212890625,
0.08612060546875,
0.47021484375,
0.08612060546875,
0.47021484375,
0.097412109375,
0.418212890625,
0.097412109375
],
[
0.509765625,
0.0843505859375,
0.568359375,
0.08612060546875,
0.568359375,
0.097412109375,
0.509765625,
0.0955810546875
],
[
0.61279296875,
0.08612060546875,
0.6572265625,
0.08612060546875,
0.6572265625,
0.097412109375,
0.61279296875,
0.097412109375
],
[
0.69287109375,
0.08612060546875,
0.771484375,
0.08612060546875,
0.771484375,
0.0968017578125,
0.69287109375,
0.0968017578125
],
[
0.7900390625,
0.08612060546875,
0.8671875,
0.08612060546875,
0.8671875,
0.0968017578125,
0.7900390625,
0.0968017578125
],
[
0.12176513671875,
0.0968017578125,
0.162841796875,
0.0968017578125,
0.162841796875,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.3046875,
0.0968017578125,
0.330078125,
0.0968017578125,
0.330078125,
0.10809326171875,
0.3046875,
0.10809326171875
],
[
0.3359375,
0.0968017578125,
0.391357421875,
0.0968017578125,
0.391357421875,
0.10687255859375,
0.3359375,
0.10687255859375
],
[
0.43408203125,
0.0968017578125,
0.48779296875,
0.0968017578125,
0.48779296875,
0.10687255859375,
0.43408203125,
0.10687255859375
],
[
0.53076171875,
0.0968017578125,
0.587890625,
0.0968017578125,
0.587890625,
0.10687255859375,
0.53076171875,
0.10687255859375
],
[
0.62646484375,
0.0968017578125,
0.68115234375,
0.0968017578125,
0.68115234375,
0.10687255859375,
0.62646484375,
0.10687255859375
],
[
0.72119140625,
0.0968017578125,
0.77685546875,
0.0968017578125,
0.77685546875,
0.10687255859375,
0.72119140625,
0.10687255859375
],
[
0.81787109375,
0.0968017578125,
0.873046875,
0.0968017578125,
0.873046875,
0.10687255859375,
0.81787109375,
0.10687255859375
],
[
0.496337890625,
0.097412109375,
0.51904296875,
0.097412109375,
0.51904296875,
0.105712890625,
0.496337890625,
0.105712890625
],
[
0.399658203125,
0.097412109375,
0.424072265625,
0.097412109375,
0.424072265625,
0.105712890625,
0.399658203125,
0.105712890625
],
[
0.591796875,
0.097412109375,
0.615234375,
0.097412109375,
0.615234375,
0.105712890625,
0.591796875,
0.105712890625
],
[
0.6884765625,
0.09857177734375,
0.7109375,
0.09857177734375,
0.7109375,
0.105712890625,
0.6884765625,
0.105712890625
],
[
0.78662109375,
0.09857177734375,
0.8076171875,
0.09857177734375,
0.8076171875,
0.105712890625,
0.78662109375,
0.105712890625
],
[
0.12005615234375,
0.10809326171875,
0.197265625,
0.10809326171875,
0.197265625,
0.1181640625,
0.12005615234375,
0.1181640625
],
[
0.3046875,
0.1092529296875,
0.32666015625,
0.1092529296875,
0.32666015625,
0.1175537109375,
0.3046875,
0.1175537109375
],
[
0.35107421875,
0.10809326171875,
0.378662109375,
0.10809326171875,
0.378662109375,
0.119384765625,
0.35107421875,
0.119384765625
],
[
0.39794921875,
0.1092529296875,
0.4248046875,
0.1092529296875,
0.4248046875,
0.1175537109375,
0.39794921875,
0.1175537109375
],
[
0.44921875,
0.1092529296875,
0.47265625,
0.1092529296875,
0.47265625,
0.1175537109375,
0.44921875,
0.1175537109375
],
[
0.496337890625,
0.1092529296875,
0.5224609375,
0.1092529296875,
0.5224609375,
0.1175537109375,
0.496337890625,
0.1175537109375
],
[
0.5458984375,
0.1092529296875,
0.5693359375,
0.1092529296875,
0.5693359375,
0.1175537109375,
0.5458984375,
0.1175537109375
],
[
0.591796875,
0.1092529296875,
0.6171875,
0.1092529296875,
0.6171875,
0.1175537109375,
0.591796875,
0.1175537109375
],
[
0.640625,
0.1092529296875,
0.6650390625,
0.1092529296875,
0.6650390625,
0.1175537109375,
0.640625,
0.1175537109375
],
[
0.6884765625,
0.1092529296875,
0.7138671875,
0.1092529296875,
0.7138671875,
0.1175537109375,
0.6884765625,
0.1175537109375
],
[
0.73779296875,
0.1092529296875,
0.76318359375,
0.1092529296875,
0.76318359375,
0.1175537109375,
0.73779296875,
0.1175537109375
],
[
0.7841796875,
0.1092529296875,
0.81005859375,
0.1092529296875,
0.81005859375,
0.1175537109375,
0.7841796875,
0.1175537109375
],
[
0.8330078125,
0.1092529296875,
0.85791015625,
0.1092529296875,
0.85791015625,
0.1175537109375,
0.8330078125,
0.1175537109375
],
[
0.68701171875,
0.11639404296875,
0.71728515625,
0.119384765625,
0.7138671875,
0.1312255859375,
0.68603515625,
0.127685546875
],
[
0.640625,
0.114013671875,
0.6689453125,
0.119384765625,
0.6650390625,
0.1318359375,
0.63623046875,
0.1270751953125
],
[
0.39794921875,
0.1175537109375,
0.426513671875,
0.1175537109375,
0.426513671875,
0.12890625,
0.39794921875,
0.12890625
],
[
0.447509765625,
0.1175537109375,
0.474365234375,
0.1175537109375,
0.474365234375,
0.12890625,
0.447509765625,
0.12890625
],
[
0.35107421875,
0.1181640625,
0.378662109375,
0.1181640625,
0.378662109375,
0.1300048828125,
0.35107421875,
0.1300048828125
],
[
0.1234130859375,
0.1181640625,
0.295654296875,
0.1181640625,
0.295654296875,
0.12890625,
0.1234130859375,
0.12890625
],
[
0.495361328125,
0.1181640625,
0.52294921875,
0.1181640625,
0.52294921875,
0.1300048828125,
0.495361328125,
0.1300048828125
],
[
0.54248046875,
0.1181640625,
0.57080078125,
0.1181640625,
0.57080078125,
0.1300048828125,
0.54248046875,
0.1300048828125
],
[
0.59033203125,
0.1181640625,
0.62060546875,
0.1181640625,
0.62060546875,
0.1300048828125,
0.59033203125,
0.1300048828125
],
[
0.736328125,
0.1181640625,
0.76318359375,
0.1181640625,
0.76318359375,
0.1300048828125,
0.736328125,
0.1300048828125
],
[
0.8310546875,
0.1181640625,
0.85986328125,
0.1181640625,
0.85986328125,
0.1300048828125,
0.8310546875,
0.1300048828125
],
[
0.78515625,
0.119384765625,
0.81005859375,
0.119384765625,
0.81005859375,
0.127685546875,
0.78515625,
0.127685546875
],
[
0.30322265625,
0.119384765625,
0.328369140625,
0.119384765625,
0.328369140625,
0.127685546875,
0.30322265625,
0.127685546875
],
[
0.349365234375,
0.127685546875,
0.378662109375,
0.127685546875,
0.378662109375,
0.14013671875,
0.349365234375,
0.14013671875
],
[
0.397216796875,
0.127685546875,
0.426513671875,
0.127685546875,
0.426513671875,
0.14013671875,
0.397216796875,
0.14013671875
],
[
0.44580078125,
0.127685546875,
0.47607421875,
0.127685546875,
0.47607421875,
0.14013671875,
0.44580078125,
0.14013671875
],
[
0.495361328125,
0.127685546875,
0.52294921875,
0.127685546875,
0.52294921875,
0.14013671875,
0.495361328125,
0.14013671875
],
[
0.54248046875,
0.127685546875,
0.57080078125,
0.127685546875,
0.57080078125,
0.14013671875,
0.54248046875,
0.14013671875
],
[
0.58837890625,
0.127685546875,
0.61865234375,
0.127685546875,
0.61865234375,
0.14013671875,
0.58837890625,
0.14013671875
],
[
0.6396484375,
0.127685546875,
0.66650390625,
0.127685546875,
0.66650390625,
0.14013671875,
0.6396484375,
0.14013671875
],
[
0.68701171875,
0.127685546875,
0.71728515625,
0.127685546875,
0.71728515625,
0.14013671875,
0.68701171875,
0.14013671875
],
[
0.736328125,
0.127685546875,
0.76416015625,
0.127685546875,
0.76416015625,
0.14013671875,
0.736328125,
0.14013671875
],
[
0.7841796875,
0.127685546875,
0.81201171875,
0.127685546875,
0.81201171875,
0.14013671875,
0.7841796875,
0.14013671875
],
[
0.8310546875,
0.127685546875,
0.85986328125,
0.127685546875,
0.85986328125,
0.14013671875,
0.8310546875,
0.14013671875
],
[
0.301513671875,
0.12890625,
0.33154296875,
0.12890625,
0.33154296875,
0.14013671875,
0.301513671875,
0.14013671875
],
[
0.12176513671875,
0.12890625,
0.27783203125,
0.12890625,
0.27783203125,
0.1395263671875,
0.12176513671875,
0.1395263671875
],
[
0.301513671875,
0.1395263671875,
0.33154296875,
0.1395263671875,
0.33154296875,
0.15087890625,
0.301513671875,
0.15087890625
],
[
0.349365234375,
0.1395263671875,
0.38037109375,
0.1395263671875,
0.38037109375,
0.15087890625,
0.349365234375,
0.15087890625
],
[
0.397216796875,
0.1395263671875,
0.42822265625,
0.1395263671875,
0.42822265625,
0.15087890625,
0.397216796875,
0.15087890625
],
[
0.447509765625,
0.1395263671875,
0.47607421875,
0.1395263671875,
0.47607421875,
0.15087890625,
0.447509765625,
0.15087890625
],
[
0.495361328125,
0.1395263671875,
0.52294921875,
0.1395263671875,
0.52294921875,
0.15087890625,
0.495361328125,
0.15087890625
],
[
0.54248046875,
0.1395263671875,
0.57080078125,
0.1395263671875,
0.57080078125,
0.15087890625,
0.54248046875,
0.15087890625
],
[
0.59033203125,
0.1395263671875,
0.62060546875,
0.1395263671875,
0.62060546875,
0.15087890625,
0.59033203125,
0.15087890625
],
[
0.6396484375,
0.1395263671875,
0.66748046875,
0.1395263671875,
0.66748046875,
0.15087890625,
0.6396484375,
0.15087890625
],
[
0.68701171875,
0.1395263671875,
0.71728515625,
0.1395263671875,
0.71728515625,
0.1519775390625,
0.68701171875,
0.1519775390625
],
[
0.736328125,
0.1395263671875,
0.76416015625,
0.1395263671875,
0.76416015625,
0.1519775390625,
0.736328125,
0.1519775390625
],
[
0.7841796875,
0.1395263671875,
0.81201171875,
0.1395263671875,
0.81201171875,
0.15087890625,
0.7841796875,
0.15087890625
],
[
0.8310546875,
0.1395263671875,
0.85986328125,
0.1395263671875,
0.85986328125,
0.15087890625,
0.8310546875,
0.15087890625
],
[
0.349365234375,
0.1484375,
0.38037109375,
0.15087890625,
0.378662109375,
0.163330078125,
0.34765625,
0.160888671875
],
[
0.12176513671875,
0.14013671875,
0.2493896484375,
0.14013671875,
0.2493896484375,
0.15087890625,
0.12176513671875,
0.15087890625
],
[
0.12005615234375,
0.149658203125,
0.27197265625,
0.149658203125,
0.27197265625,
0.1602783203125,
0.12005615234375,
0.1602783203125
],
[
0.301513671875,
0.149658203125,
0.33154296875,
0.149658203125,
0.33154296875,
0.160888671875,
0.301513671875,
0.160888671875
],
[
0.397216796875,
0.1484375,
0.426513671875,
0.15087890625,
0.4248046875,
0.163330078125,
0.393798828125,
0.1602783203125
],
[
0.44580078125,
0.15087890625,
0.47607421875,
0.15087890625,
0.47607421875,
0.162109375,
0.44580078125,
0.162109375
],
[
0.495361328125,
0.15087890625,
0.52294921875,
0.15087890625,
0.52294921875,
0.162109375,
0.495361328125,
0.162109375
],
[
0.54248046875,
0.15087890625,
0.57080078125,
0.15087890625,
0.57080078125,
0.162109375,
0.54248046875,
0.162109375
],
[
0.58837890625,
0.15087890625,
0.61865234375,
0.15087890625,
0.61865234375,
0.162109375,
0.58837890625,
0.162109375
],
[
0.640625,
0.15087890625,
0.66748046875,
0.15087890625,
0.66748046875,
0.162109375,
0.640625,
0.162109375
],
[
0.68603515625,
0.15087890625,
0.71533203125,
0.15087890625,
0.71533203125,
0.162109375,
0.68603515625,
0.162109375
],
[
0.736328125,
0.15087890625,
0.76318359375,
0.15087890625,
0.76318359375,
0.162109375,
0.736328125,
0.162109375
],
[
0.7841796875,
0.15087890625,
0.81201171875,
0.15087890625,
0.81201171875,
0.162109375,
0.7841796875,
0.162109375
],
[
0.8310546875,
0.149658203125,
0.861328125,
0.149658203125,
0.861328125,
0.160888671875,
0.8310546875,
0.160888671875
],
[
0.54248046875,
0.1591796875,
0.57275390625,
0.160888671875,
0.57080078125,
0.17333984375,
0.54150390625,
0.171630859375
],
[
0.6396484375,
0.157958984375,
0.66748046875,
0.160888671875,
0.66650390625,
0.1728515625,
0.63623046875,
0.169189453125
],
[
0.495361328125,
0.1602783203125,
0.52294921875,
0.1602783203125,
0.52294921875,
0.171630859375,
0.495361328125,
0.171630859375
],
[
0.301513671875,
0.1602783203125,
0.33154296875,
0.1602783203125,
0.33154296875,
0.171630859375,
0.301513671875,
0.171630859375
],
[
0.68603515625,
0.1602783203125,
0.71728515625,
0.1602783203125,
0.71728515625,
0.1728515625,
0.68603515625,
0.1728515625
],
[
0.349365234375,
0.160888671875,
0.38037109375,
0.160888671875,
0.38037109375,
0.1728515625,
0.349365234375,
0.1728515625
],
[
0.124267578125,
0.160888671875,
0.267822265625,
0.160888671875,
0.267822265625,
0.171630859375,
0.124267578125,
0.171630859375
],
[
0.397216796875,
0.160888671875,
0.42822265625,
0.160888671875,
0.42822265625,
0.1728515625,
0.397216796875,
0.1728515625
],
[
0.59033203125,
0.160888671875,
0.62060546875,
0.160888671875,
0.62060546875,
0.1728515625,
0.59033203125,
0.1728515625
],
[
0.736328125,
0.160888671875,
0.76416015625,
0.160888671875,
0.76416015625,
0.1728515625,
0.736328125,
0.1728515625
],
[
0.7841796875,
0.160888671875,
0.81201171875,
0.160888671875,
0.81201171875,
0.1728515625,
0.7841796875,
0.1728515625
],
[
0.8310546875,
0.160888671875,
0.85986328125,
0.160888671875,
0.85986328125,
0.1728515625,
0.8310546875,
0.1728515625
],
[
0.349365234375,
0.169189453125,
0.38037109375,
0.1728515625,
0.3779296875,
0.18408203125,
0.345947265625,
0.1817626953125
],
[
0.447509765625,
0.162109375,
0.474365234375,
0.162109375,
0.474365234375,
0.17041015625,
0.447509765625,
0.17041015625
],
[
0.397216796875,
0.169189453125,
0.42822265625,
0.171630859375,
0.426513671875,
0.18408203125,
0.3955078125,
0.1817626953125
],
[
0.44580078125,
0.169189453125,
0.476806640625,
0.171630859375,
0.474365234375,
0.185302734375,
0.444091796875,
0.182861328125
],
[
0.68408203125,
0.169189453125,
0.71875,
0.169189453125,
0.71875,
0.185302734375,
0.68408203125,
0.185302734375
],
[
0.495361328125,
0.17041015625,
0.52294921875,
0.17041015625,
0.52294921875,
0.18408203125,
0.495361328125,
0.18408203125
],
[
0.301513671875,
0.17041015625,
0.33154296875,
0.17041015625,
0.33154296875,
0.182861328125,
0.301513671875,
0.182861328125
],
[
0.59033203125,
0.17041015625,
0.62109375,
0.17041015625,
0.62109375,
0.185302734375,
0.59033203125,
0.185302734375
],
[
0.6396484375,
0.17041015625,
0.66748046875,
0.17041015625,
0.66748046875,
0.18408203125,
0.6396484375,
0.18408203125
],
[
0.73486328125,
0.171630859375,
0.76416015625,
0.171630859375,
0.76416015625,
0.182861328125,
0.73486328125,
0.182861328125
],
[
0.7841796875,
0.171630859375,
0.81201171875,
0.171630859375,
0.81201171875,
0.182861328125,
0.7841796875,
0.182861328125
],
[
0.8310546875,
0.171630859375,
0.85986328125,
0.171630859375,
0.85986328125,
0.182861328125,
0.8310546875,
0.182861328125
],
[
0.54248046875,
0.171630859375,
0.57080078125,
0.171630859375,
0.57080078125,
0.18408203125,
0.54248046875,
0.18408203125
],
[
0.1234130859375,
0.1728515625,
0.264404296875,
0.1728515625,
0.264404296875,
0.182861328125,
0.1234130859375,
0.182861328125
],
[
0.301513671875,
0.1817626953125,
0.33154296875,
0.1817626953125,
0.33154296875,
0.19482421875,
0.301513671875,
0.19482421875
],
[
0.349365234375,
0.1817626953125,
0.38037109375,
0.1817626953125,
0.38037109375,
0.19482421875,
0.349365234375,
0.19482421875
],
[
0.3955078125,
0.1817626953125,
0.42822265625,
0.1817626953125,
0.42822265625,
0.19482421875,
0.3955078125,
0.19482421875
],
[
0.444091796875,
0.1817626953125,
0.47607421875,
0.1817626953125,
0.47607421875,
0.193603515625,
0.444091796875,
0.193603515625
],
[
0.49365234375,
0.1817626953125,
0.52294921875,
0.1817626953125,
0.52294921875,
0.19482421875,
0.49365234375,
0.19482421875
],
[
0.54248046875,
0.1817626953125,
0.57275390625,
0.1817626953125,
0.57275390625,
0.193603515625,
0.54248046875,
0.193603515625
],
[
0.58837890625,
0.1817626953125,
0.62060546875,
0.1817626953125,
0.62060546875,
0.19482421875,
0.58837890625,
0.19482421875
],
[
0.63818359375,
0.1817626953125,
0.66748046875,
0.1817626953125,
0.66748046875,
0.193603515625,
0.63818359375,
0.193603515625
],
[
0.68603515625,
0.1817626953125,
0.71728515625,
0.1817626953125,
0.71728515625,
0.193603515625,
0.68603515625,
0.193603515625
],
[
0.73291015625,
0.1817626953125,
0.765625,
0.1817626953125,
0.765625,
0.193603515625,
0.73291015625,
0.193603515625
],
[
0.78271484375,
0.1817626953125,
0.8134765625,
0.1817626953125,
0.8134765625,
0.193603515625,
0.78271484375,
0.193603515625
],
[
0.82958984375,
0.1817626953125,
0.861328125,
0.1817626953125,
0.861328125,
0.193603515625,
0.82958984375,
0.193603515625
],
[
0.35107421875,
0.1923828125,
0.378662109375,
0.1923828125,
0.378662109375,
0.204833984375,
0.35107421875,
0.204833984375
],
[
0.1234130859375,
0.182861328125,
0.273681640625,
0.182861328125,
0.273681640625,
0.193603515625,
0.1234130859375,
0.193603515625
],
[
0.3955078125,
0.1923828125,
0.426513671875,
0.1923828125,
0.426513671875,
0.204833984375,
0.3955078125,
0.204833984375
],
[
0.447509765625,
0.1923828125,
0.47607421875,
0.1923828125,
0.47607421875,
0.204833984375,
0.447509765625,
0.204833984375
],
[
0.49365234375,
0.1923828125,
0.52294921875,
0.1923828125,
0.52294921875,
0.204833984375,
0.49365234375,
0.204833984375
],
[
0.54248046875,
0.1923828125,
0.57275390625,
0.1923828125,
0.57275390625,
0.204833984375,
0.54248046875,
0.204833984375
],
[
0.58837890625,
0.1923828125,
0.62060546875,
0.1923828125,
0.62060546875,
0.204833984375,
0.58837890625,
0.204833984375
],
[
0.6396484375,
0.1923828125,
0.66748046875,
0.1923828125,
0.66748046875,
0.204833984375,
0.6396484375,
0.204833984375
],
[
0.68603515625,
0.1923828125,
0.71728515625,
0.1923828125,
0.71728515625,
0.204833984375,
0.68603515625,
0.204833984375
],
[
0.736328125,
0.1923828125,
0.76416015625,
0.1923828125,
0.76416015625,
0.204833984375,
0.736328125,
0.204833984375
],
[
0.78271484375,
0.1923828125,
0.81201171875,
0.1923828125,
0.81201171875,
0.204833984375,
0.78271484375,
0.204833984375
],
[
0.8330078125,
0.193603515625,
0.85986328125,
0.193603515625,
0.85986328125,
0.204833984375,
0.8330078125,
0.204833984375
],
[
0.3046875,
0.19482421875,
0.328369140625,
0.19482421875,
0.328369140625,
0.203125,
0.3046875,
0.203125
],
[
0.1234130859375,
0.193603515625,
0.295654296875,
0.193603515625,
0.295654296875,
0.2037353515625,
0.1234130859375,
0.2037353515625
],
[
0.1318359375,
0.218505859375,
0.8623046875,
0.218505859375,
0.8623046875,
0.232177734375,
0.1318359375,
0.232177734375
],
[
0.51318359375,
0.257080078125,
0.5380859375,
0.257080078125,
0.5380859375,
0.26953125,
0.51318359375,
0.26953125
],
[
0.5400390625,
0.255859375,
0.72314453125,
0.257080078125,
0.72314453125,
0.271484375,
0.5400390625,
0.270751953125
],
[
0.285400390625,
0.270751953125,
0.364501953125,
0.270751953125,
0.364501953125,
0.28076171875,
0.285400390625,
0.28076171875
],
[
0.151123046875,
0.28076171875,
0.162109375,
0.28076171875,
0.162109375,
0.28515625,
0.151123046875,
0.28515625
],
[
0.1822509765625,
0.2880859375,
0.18896484375,
0.2880859375,
0.18896484375,
0.29443359375,
0.1822509765625,
0.29443359375
],
[
0.198974609375,
0.283935546875,
0.2451171875,
0.283935546875,
0.2451171875,
0.295654296875,
0.198974609375,
0.295654296875
],
[
0.198974609375,
0.29443359375,
0.2435302734375,
0.29443359375,
0.2435302734375,
0.3046875,
0.198974609375,
0.3046875
],
[
0.51123046875,
0.29150390625,
0.623046875,
0.29150390625,
0.623046875,
0.305908203125,
0.51123046875,
0.305908203125
],
[
0.198974609375,
0.302734375,
0.2435302734375,
0.302734375,
0.2435302734375,
0.3134765625,
0.198974609375,
0.3134765625
],
[
0.1822509765625,
0.29638671875,
0.18896484375,
0.29638671875,
0.18896484375,
0.3017578125,
0.1822509765625,
0.3017578125
],
[
0.15283203125,
0.311279296875,
0.162109375,
0.311279296875,
0.162109375,
0.3154296875,
0.15283203125,
0.3154296875
],
[
0.183837890625,
0.314208984375,
0.18896484375,
0.314208984375,
0.18896484375,
0.319580078125,
0.183837890625,
0.319580078125
],
[
0.20068359375,
0.312255859375,
0.2301025390625,
0.312255859375,
0.2301025390625,
0.320556640625,
0.20068359375,
0.320556640625
],
[
0.1351318359375,
0.318359375,
0.145263671875,
0.318359375,
0.145263671875,
0.33837890625,
0.1351318359375,
0.33837890625
],
[
0.15283203125,
0.3154296875,
0.160400390625,
0.3154296875,
0.160400390625,
0.318359375,
0.15283203125,
0.318359375
],
[
0.20068359375,
0.320556640625,
0.26025390625,
0.320556640625,
0.26025390625,
0.330810546875,
0.20068359375,
0.330810546875
],
[
0.5087890625,
0.319580078125,
0.8818359375,
0.320556640625,
0.8818359375,
0.3349609375,
0.5087890625,
0.334228515625
],
[
0.124267578125,
0.328857421875,
0.1695556640625,
0.330078125,
0.166259765625,
0.41748046875,
0.12176513671875,
0.416259765625
],
[
0.183837890625,
0.32373046875,
0.18896484375,
0.32373046875,
0.18896484375,
0.328857421875,
0.183837890625,
0.328857421875
],
[
0.197265625,
0.328857421875,
0.26025390625,
0.330078125,
0.26025390625,
0.34033203125,
0.197265625,
0.339111328125
],
[
0.51123046875,
0.33837890625,
0.8818359375,
0.33837890625,
0.8818359375,
0.3515625,
0.51123046875,
0.3515625
],
[
0.1427001953125,
0.344482421875,
0.15869140625,
0.344482421875,
0.15869140625,
0.3486328125,
0.1427001953125,
0.3486328125
],
[
0.509765625,
0.3515625,
0.88330078125,
0.352783203125,
0.88330078125,
0.36767578125,
0.509765625,
0.366455078125
],
[
0.51123046875,
0.369384765625,
0.8818359375,
0.369384765625,
0.8818359375,
0.383056640625,
0.51123046875,
0.383056640625
],
[
0.51123046875,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3984375,
0.51123046875,
0.3984375
],
[
0.51318359375,
0.402099609375,
0.8818359375,
0.402587890625,
0.8818359375,
0.416259765625,
0.51318359375,
0.4150390625
],
[
0.15283203125,
0.410888671875,
0.162109375,
0.410888671875,
0.162109375,
0.4150390625,
0.15283203125,
0.4150390625
],
[
0.51123046875,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.51123046875,
0.431640625
],
[
0.51123046875,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.44775390625,
0.51123046875,
0.44775390625
],
[
0.1485595703125,
0.441162109375,
0.16796875,
0.441162109375,
0.16796875,
0.449462890625,
0.1485595703125,
0.449462890625
],
[
0.2225341796875,
0.448974609375,
0.2392578125,
0.448974609375,
0.2392578125,
0.457763671875,
0.2225341796875,
0.457763671875
],
[
0.474365234375,
0.448974609375,
0.489501953125,
0.448974609375,
0.489501953125,
0.457275390625,
0.474365234375,
0.457275390625
],
[
0.34765625,
0.449462890625,
0.364501953125,
0.449462890625,
0.364501953125,
0.457275390625,
0.34765625,
0.457275390625
],
[
0.15869140625,
0.449462890625,
0.17626953125,
0.449462890625,
0.17626953125,
0.457275390625,
0.15869140625,
0.457275390625
],
[
0.287109375,
0.45068359375,
0.298828125,
0.45068359375,
0.298828125,
0.454833984375,
0.287109375,
0.454833984375
],
[
0.41064453125,
0.449462890625,
0.426513671875,
0.449462890625,
0.426513671875,
0.457763671875,
0.41064453125,
0.457763671875
],
[
0.51318359375,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.51123046875,
0.466064453125,
0.8818359375,
0.466064453125,
0.8818359375,
0.479736328125,
0.51123046875,
0.479736328125
],
[
0.2626953125,
0.454833984375,
0.38623046875,
0.454833984375,
0.38623046875,
0.468505859375,
0.2626953125,
0.468505859375
],
[
0.1175537109375,
0.482666015625,
0.48779296875,
0.482666015625,
0.48779296875,
0.493408203125,
0.1175537109375,
0.493408203125
],
[
0.51123046875,
0.482177734375,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.51123046875,
0.495361328125
],
[
0.1175537109375,
0.495361328125,
0.48779296875,
0.495361328125,
0.48779296875,
0.5087890625,
0.1175537109375,
0.5087890625
],
[
0.51123046875,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5126953125,
0.51123046875,
0.5126953125
],
[
0.1158447265625,
0.51025390625,
0.48779296875,
0.51025390625,
0.48779296875,
0.52392578125,
0.1158447265625,
0.52392578125
],
[
0.51123046875,
0.51318359375,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.51123046875,
0.52685546875
],
[
0.1158447265625,
0.52392578125,
0.2568359375,
0.52392578125,
0.2568359375,
0.53759765625,
0.1158447265625,
0.53759765625
],
[
0.509765625,
0.53076171875,
0.88330078125,
0.52978515625,
0.88330078125,
0.54443359375,
0.509765625,
0.5458984375
],
[
0.509765625,
0.54638671875,
0.8818359375,
0.5458984375,
0.8818359375,
0.56005859375,
0.509765625,
0.56103515625
],
[
0.51123046875,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.51123046875,
0.57666015625
],
[
0.509765625,
0.57763671875,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.509765625,
0.59130859375
],
[
0.51123046875,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.609375,
0.51123046875,
0.609375
],
[
0.1175537109375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62255859375,
0.1175537109375,
0.62255859375
],
[
0.509765625,
0.61083984375,
0.79443359375,
0.61083984375,
0.79443359375,
0.62451171875,
0.509765625,
0.62451171875
],
[
0.1175537109375,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.5283203125,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.5283203125,
0.6435546875
],
[
0.1175537109375,
0.64111328125,
0.489501953125,
0.64111328125,
0.489501953125,
0.65478515625,
0.1175537109375,
0.65478515625
],
[
0.51123046875,
0.6455078125,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.51123046875,
0.66015625
],
[
0.1175537109375,
0.65673828125,
0.486083984375,
0.65673828125,
0.486083984375,
0.67041015625,
0.1175537109375,
0.67041015625
],
[
0.51123046875,
0.66455078125,
0.88330078125,
0.66455078125,
0.88330078125,
0.677734375,
0.51123046875,
0.677734375
],
[
0.1158447265625,
0.6728515625,
0.41650390625,
0.6728515625,
0.41650390625,
0.68701171875,
0.1158447265625,
0.68701171875
],
[
0.509765625,
0.6787109375,
0.880859375,
0.6787109375,
0.880859375,
0.6923828125,
0.509765625,
0.6923828125
],
[
0.509765625,
0.69482421875,
0.88330078125,
0.6953125,
0.88330078125,
0.71044921875,
0.509765625,
0.708984375
],
[
0.1368408203125,
0.7119140625,
0.489501953125,
0.7119140625,
0.489501953125,
0.7255859375,
0.1368408203125,
0.7255859375
],
[
0.509765625,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.509765625,
0.7255859375
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.51123046875,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1175537109375,
0.74365234375,
0.490234375,
0.74365234375,
0.490234375,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.509765625,
0.7421875,
0.880859375,
0.74365234375,
0.880859375,
0.75830078125,
0.509765625,
0.75732421875
],
[
0.1158447265625,
0.7587890625,
0.486083984375,
0.75830078125,
0.486083984375,
0.7724609375,
0.1158447265625,
0.77392578125
],
[
0.51123046875,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.11419677734375,
0.77490234375,
0.48779296875,
0.77490234375,
0.48779296875,
0.7880859375,
0.11419677734375,
0.7880859375
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.1175537109375,
0.8056640625
],
[
0.51123046875,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.1158447265625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.51123046875,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.51318359375,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1158447265625,
0.84130859375,
0.490234375,
0.84130859375,
0.490234375,
0.85498046875,
0.1158447265625,
0.85498046875
],
[
0.5087890625,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85498046875,
0.5087890625,
0.85498046875
],
[
0.1158447265625,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1158447265625,
0.87158203125,
0.48779296875,
0.87255859375,
0.48779296875,
0.88720703125,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.11419677734375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.11419677734375,
0.90283203125
],
[
0.509765625,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.11419677734375,
0.90380859375,
0.474365234375,
0.90478515625,
0.474365234375,
0.91943359375,
0.11419677734375,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.78271484375,
0.90478515625,
0.78271484375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Prompt",
"FT5-Small",
"FT5-Base",
"FT5-Large",
"FT5-XL",
"GPT3-Davinci",
"GPT3.5-Turbo",
"Metric",
"Acc",
"Macro-F1",
"Macro-F1",
"Macro-F1",
"Macro-F1",
"Macro-F1",
"Macro-F1",
"Acc",
"Acc",
"Acc",
"Acc",
"Acc",
"PlainQuestion",
"0.51",
"0.34",
"0.50",
"0.35",
"0.56",
"0.54",
"0.58",
"0.51",
"0.52",
"0.38",
"0.69",
"0.69",
"0.51",
"0.67",
"0.52",
"0.52",
"0.34",
"+GeneralKnowledgeInformation",
"0.64",
"0.63",
"0.67",
"0.35",
"0.79",
"0.79",
"0.50",
"0.40",
"0.55",
"0.53",
"0.63",
"0.61",
"0.76",
"0.76",
"0.56",
"0.48",
"0.85",
"0.85",
"0.50",
"+SizeKnowledgeInformation",
"0.50",
"0.50",
"0.50",
"0.54",
"0.62",
"0.77",
"0.87",
"0.88",
"0.73",
"0.89",
"0.90",
"0.89",
"0.34",
"+Exact Size Information",
"+Only Head EntityGold Size",
"0.50",
"0.54",
"0.50",
"0.63",
"0.60",
"0.64",
"0.64",
"0.51",
"0.35",
"0.69",
"0.69",
"0.54",
"0.63",
"0.55",
"0.50",
"0.51",
"0.34",
"+Only Tail Entity Gold Size",
"0.49",
"0.66",
"0.34",
"0.75",
"0.74",
"0.49",
"0.34",
"0.49",
"0.34",
"0.53",
"0.53",
"0.50",
"0.74",
"0.74",
"0.39",
"0.80",
"0.80",
"0.41",
"+Masking Size Information",
"0.50",
"0.33",
"0.50",
"0.37",
"0.49",
"0.42",
"0.63",
"0.58",
"0.52",
"0.46",
"0.70",
"0.69",
"0.47",
"+MaskingEntity Information",
"0.49",
"0.33",
"0.47",
"0.33",
"0.23",
"0.23",
"0.42",
"0.41",
"0.38",
"0.38",
"0.50",
"+Counterfactual Size Information",
"Table 2: The results of using different prompt settings on various models on the general question of POSQA.",
"5",
"Results and Analysis",
"General Question",
"1.0",
"8",
"FT5-Small",
"FT5-Base",
"5.1 Baseline",
"FT5-Large",
"O",
"0",
"",
"FT5-XL",
"",
"P",
"GPT3-Davinci",
"The baseline is to query LLMs using the Plain",
"wubieaiauo",
"",
"GPT3.5-Turbo",
"Question prompt and the results are presented in",
"0",
"Table 2 for general questions and Table 3 for spe-",
"cial questions. When the number of parameters",
"of the model exceeds 250M (Flan-T5-Base), the",
"ability to answer the special size comparison ques-",
"7",
"tions begins to emerge. GPT3.5-Turbo achieves",
"the best accuracy score and Macro-F1 score on",
"0.0",
"0.2",
"1.0",
"0.6",
"0.0",
"04",
"0.8",
"both types of questions. It is surprising that sig-",
"nificantly smaller models, such as Flan-T5-Large",
"Context Effective Rate",
"Figurel:TheresultsofCERandCMRusingdifferent",
"(780M)and Flan-T5-XL(3B),exhibit superior per-",
"knowledge-augmented prompts (except for Counterfac-",
"formance. For example, Flan-T5-XL outperforms",
"tual Size Information) on various models on general",
"GPT3-Davinci (175B) by 0.06 precision in answer",
"question of POSQA.",
"ing general questions and 0.03 precision in answer-",
"ing special questions. In general, LLMs perform",
"better at special questions. For example, the ac-",
"curacy increases by 0.14 on Flan-T5-XL and 0.04",
"on GPT3.5-Turbo, respectively, from answering",
"used metrics for the evaluation of theperformance",
" general questions to special questions.",
"of the model in classification tasks. Accuracy con-",
"Our empirical investigation also reveals that the",
"siders global precision and recall of the categories,",
"GPT3-Davinci model tends to provide an initial",
"while Macro-F1 computes the average of the F1",
"incorrect answer, despite subsequently offering a",
" scores obtained by individual categories.",
"correct explanation for the given question. This",
"phenomenon occurs especially when directly query-",
"To explore the influence of context on the",
"ing GPT3-Davinci with general questions to get",
"prompt, we propose two quantitative evaluation",
"the \"yes\" or\"no\" answers.We speculate that it",
"metrics: Context Effective Rate (CER) and Con-",
"could be attributed to the use of first-word sampling",
"text Misleading Rate (CMR). We calculate CER",
"techniques during the decoding process. However.",
" and CMR by comparing the output of a model us-",
"this phenomenon does not occur in GPT3.5-Turbo",
"ing a prompt that contains context information with",
"whichhasbeenoptimisedfordialoguescenario",
"the output of a model using a prompt that does not",
"Even GPT3.5-Turbo, one of the most powerful",
"contain any context information. Specifically, CER",
"LLMs, can only achieve an average 0.71 accuracy",
"evaluates how many incorrectly answered ques-",
"score on these two types of questions. This sug",
"tions can be correctly answered after adding the",
"gests that although larger LMs may possess certain",
"context in the prompt. CMR evaluates how many",
"advantages in most situations, they may still lack",
"correctly answered questions can be incorrectly",
"real-world understanding when it comes to answer-",
"answered after adding the context to the prompt.",
"ing basic size comparison questions."
] |
{
"arxiv_id": "2310.13394",
"language": "en",
"page_idx": 4
}
|
||
2110.11657-en-0
|
2110.11657-en-0.jpg
|
$\textbf{Projective Manifold Gradient Layer for Deep Rotation Regression}$
Jiayi Chen$^{1, 2}$ Yingda Yin$^{1}$ Tolga Birdal$^{3,4}$
$^{^{1}}$CFCS, Peking University
$^3$Stanford University
Baoquan Chen$^{1}$ Leonidas J.Guibas$^{3}$ He Wang$^{1^\dagger}$
$^2\dot{\text{Beijing Institute for General AI}}$
$^{4}$Imperial College London
## $\mathbf{Abstract}$
Regressing rotations on SO(3) manifold using deep neural networks is an important yet unsolved problem. The gap between the Euclidean network output space and the non-Euclidean SO(3) manifold imposes a severe challenge for neural network learning in both forward and backward passes. While several works have proposed different regression-friendly rotation representations, very few
$\sum_{\begin{array}{cc}&ent®erssion-frenaly&representations,very\:few\\&works&have&been&devoted&to&improving&the&gradient&back-1\\\bigcup&propagating&in&the&backward&pass.&In&this&paper,we&pro-1\\\hline gi&pose&a&manifold-aware&gradient&that&directly&backpropa-1\\\textcircled{0}&gates&into&de\end{array}}$
optimization to construct a novel projective gradient, our proposed regularized projective manifold gradient (RPMG) method helps networks achieve new state-of-the-art performance in a variety of rotation estimation tasks. Our proposed gradient layer can also be applied to other smooth manifolds such as the unit sphere. Our project page is at $https://jychen18.github.io/RPMG.$
resentation from [46] and 10D symmetric matrix representation from [27]. Recently, Levinson et al.[25] adopted the vanilla 9D matrix representation discovering that simply replacing the Gram-Schmidt process in the 6D representation[46] with symmetric SVD-based orthogonalization can make this representation superior to the others.
Despite the progress on discovering better rotation representations, the gap between a Euclidean network output space and the non-Euclidean SO(3) manifold hasn't been completely filled. One important yet long-neglected problem lies in optimization on non-Euclidean manifolds[1]: to optimize on SO(3) manifold, the optimization variable is a rotation matrix, which contains nine matrix elements; if we naively use Euclidean gradient, which simply computes the partial derivatives with respect to each of the nine matrix elements, to update the variable, this optimization step will usually lead to a new matrix off SO(3) manifold. Unfortunately, we observe that all the existing works on rotation regression simply rely upon vanilla auto-differentiation for backpropagation, exactly computing Euclidean gradient and performing such off-manifold updates to predicted rotations. We argue that, for training deep rotation regression networks, the off-manifold components will lead to noise in the gradient of neural network weights, hindering network training and convergence.
To tackle this issue, we draw inspiration from differential geometry, where people leverage Riemannian optimization to optimize on the non-Euclidean manifold, which finds the direction of the steepest geodesic path on the manifold and take an on-manifold step. We thus propose to leverage Riemannian optimization and delve deep into the study of the backward pass. Note that this is a fundamental yet currently under-explored avenue, given that most of the existing works focus on a holistic design of rotation regression that is agnostic to forward/backward pass. However, incorporating Riemannian optimization into network training is highly non-trivial and challenging. Although methods of Riemannian optimization allow for optimization on SO(3) [5,31], matrix manifolds [1] or general Riemannian manifolds [34,44], they are not directly applicable to update the weights of the neural networks that are Euclidean. Also,
## $1. \textbf{Introduction}$ Estimating rotations is a crucial problem in visual per-
ception that has broad applications, e.g., in object pose estimation, robot control, camera relocalization, 3D reconstruc- tion and visual odometry [8,13,16,22,37]. Recently, with the proliferation of deep neural networks, learning to accurately regress rotations is attracting more and more attention. However, the non-Euclidean characteristics of rotation space make accurately regressing rotation very challenging.
As we know, rotations reside in a non-Euclidean manifold, SO(3) group, whereas the unconstrained outputs of neural networks usually live in Euclidean spaces. This gap between the neural network output space and SO(3) manifold becomes a major challenge for deep rotation regression, thus tackling this gap becomes an important research topic. One popular research direction is to design learningfriendly rotation representations, e.g., 6D continuous rep-
$^{\dagger}:$ He Wang is the corresponding author (hewang@pku.edu.cn).
1
|
[
[
0.1568603515625,
0.1337890625,
0.81396484375,
0.1337890625,
0.81396484375,
0.1483154296875,
0.1568603515625,
0.1483154296875
],
[
0.2017822265625,
0.1881103515625,
0.48046875,
0.1881103515625,
0.48046875,
0.205810546875,
0.2017822265625,
0.205810546875
],
[
0.79150390625,
0.1881103515625,
0.88623046875,
0.1881103515625,
0.88623046875,
0.205810546875,
0.79150390625,
0.205810546875
],
[
0.471435546875,
0.1900634765625,
0.78515625,
0.1900634765625,
0.78515625,
0.2052001953125,
0.471435546875,
0.2052001953125
],
[
0.08905029296875,
0.189453125,
0.2115478515625,
0.186279296875,
0.2115478515625,
0.203857421875,
0.08905029296875,
0.20703125
],
[
0.2459716796875,
0.205810546875,
0.73779296875,
0.2083740234375,
0.73779296875,
0.2247314453125,
0.2459716796875,
0.222900390625
],
[
0.289306640625,
0.222900390625,
0.463134765625,
0.2247314453125,
0.463134765625,
0.242431640625,
0.289306640625,
0.240478515625
],
[
0.461669921875,
0.2259521484375,
0.685546875,
0.2259521484375,
0.685546875,
0.240478515625,
0.461669921875,
0.240478515625
],
[
0.0236968994140625,
0.25830078125,
0.058013916015625,
0.25830078125,
0.058013916015625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.236083984375,
0.27587890625,
0.314453125,
0.27587890625,
0.314453125,
0.29345703125,
0.236083984375,
0.29345703125
],
[
0.50244140625,
0.279052734375,
0.8896484375,
0.279052734375,
0.8896484375,
0.292236328125,
0.50244140625,
0.292236328125
],
[
0.50341796875,
0.294189453125,
0.890625,
0.294189453125,
0.890625,
0.30810546875,
0.50341796875,
0.30810546875
],
[
0.1004638671875,
0.309326171875,
0.468017578125,
0.309326171875,
0.468017578125,
0.323974609375,
0.1004638671875,
0.323974609375
],
[
0.50341796875,
0.31005859375,
0.8896484375,
0.31005859375,
0.8896484375,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.08087158203125,
0.324462890625,
0.4697265625,
0.324462890625,
0.4697265625,
0.33837890625,
0.08087158203125,
0.33837890625
],
[
0.5009765625,
0.324462890625,
0.890625,
0.323974609375,
0.890625,
0.33837890625,
0.5009765625,
0.339111328125
],
[
0.080078125,
0.34033203125,
0.4697265625,
0.33837890625,
0.4697265625,
0.352783203125,
0.080078125,
0.354736328125
],
[
0.50244140625,
0.34033203125,
0.89208984375,
0.34033203125,
0.89208984375,
0.354248046875,
0.50244140625,
0.354248046875
],
[
0.080078125,
0.354248046875,
0.468017578125,
0.354736328125,
0.468017578125,
0.369384765625,
0.080078125,
0.36865234375
],
[
0.5009765625,
0.354248046875,
0.81396484375,
0.354736328125,
0.81396484375,
0.369384765625,
0.5009765625,
0.36865234375
],
[
0.0228729248046875,
0.363037109375,
0.052276611328125,
0.361083984375,
0.056365966796875,
0.39453125,
0.0269622802734375,
0.396484375
],
[
0.080078125,
0.37060546875,
0.468017578125,
0.37060546875,
0.468017578125,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.521484375,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.386474609375,
0.521484375,
0.386474609375
],
[
0.08251953125,
0.385009765625,
0.46728515625,
0.385009765625,
0.46728515625,
0.3984375,
0.08251953125,
0.3984375
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.4140625,
0.080078125,
0.4140625
],
[
0.50341796875,
0.401611328125,
0.89208984375,
0.401611328125,
0.89208984375,
0.415283203125,
0.50341796875,
0.415283203125
],
[
0.080078125,
0.415283203125,
0.468017578125,
0.415283203125,
0.468017578125,
0.4287109375,
0.080078125,
0.4287109375
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.50244140625,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.43115234375,
0.468017578125,
0.43115234375,
0.468017578125,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.50244140625,
0.431884765625,
0.89208984375,
0.431884765625,
0.89208984375,
0.44580078125,
0.50244140625,
0.44580078125
],
[
0.035125732421875,
0.439453125,
0.047393798828125,
0.43505859375,
0.049835205078125,
0.43994140625,
0.037567138671875,
0.4443359375
],
[
0.0318603515625,
0.4462890625,
0.058013916015625,
0.4462890625,
0.058013916015625,
0.48095703125,
0.0318603515625,
0.48095703125
],
[
0.080078125,
0.44580078125,
0.4697265625,
0.44580078125,
0.4697265625,
0.458984375,
0.080078125,
0.458984375
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.4609375,
0.4697265625,
0.4609375,
0.4697265625,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.50244140625,
0.462158203125,
0.89208984375,
0.462158203125,
0.89208984375,
0.47607421875,
0.50244140625,
0.47607421875
],
[
0.080078125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.080078125,
0.4892578125
],
[
0.50341796875,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.50341796875,
0.4912109375
],
[
0.0236968994140625,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.568359375,
0.0236968994140625,
0.568359375
],
[
0.07843017578125,
0.4912109375,
0.468017578125,
0.4892578125,
0.468017578125,
0.50390625,
0.07843017578125,
0.505859375
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.07843017578125,
0.5048828125,
0.468017578125,
0.505859375,
0.468017578125,
0.52099609375,
0.07843017578125,
0.51953125
],
[
0.5009765625,
0.5068359375,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.5009765625,
0.52099609375
],
[
0.080078125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.080078125,
0.53515625
],
[
0.50341796875,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.080078125,
0.537109375,
0.4697265625,
0.537109375,
0.4697265625,
0.55029296875,
0.080078125,
0.55029296875
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.080078125,
0.5517578125,
0.4697265625,
0.5517578125,
0.4697265625,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.50341796875,
0.55322265625,
0.89208984375,
0.55322265625,
0.89208984375,
0.56640625,
0.50341796875,
0.56640625
],
[
0.02777099609375,
0.5595703125,
0.051483154296875,
0.5595703125,
0.051483154296875,
0.58203125,
0.02777099609375,
0.58203125
],
[
0.041656494140625,
0.56298828125,
0.0482177734375,
0.56298828125,
0.0482177734375,
0.56640625,
0.041656494140625,
0.56640625
],
[
0.08087158203125,
0.56494140625,
0.30224609375,
0.56494140625,
0.30224609375,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.50341796875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.02777099609375,
0.57763671875,
0.053924560546875,
0.57763671875,
0.053924560546875,
0.625,
0.02777099609375,
0.625
],
[
0.50341796875,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.50341796875,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.0228729248046875,
0.61279296875,
0.060455322265625,
0.61279296875,
0.060455322265625,
0.7109375,
0.0228729248046875,
0.7109375
],
[
0.080078125,
0.6171875,
0.20751953125,
0.61865234375,
0.20751953125,
0.63330078125,
0.080078125,
0.6318359375
],
[
0.50341796875,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.50341796875,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.50341796875,
0.642578125
],
[
0.098876953125,
0.63623046875,
0.46728515625,
0.63623046875,
0.46728515625,
0.6494140625,
0.098876953125,
0.6494140625
],
[
0.50244140625,
0.64453125,
0.67236328125,
0.64453125,
0.67236328125,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.521484375,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.521484375,
0.673828125
],
[
0.080078125,
0.66748046875,
0.468017578125,
0.66748046875,
0.468017578125,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.5009765625,
0.673828125,
0.890625,
0.6748046875,
0.890625,
0.689453125,
0.5009765625,
0.68798828125
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.080078125,
0.69775390625,
0.468017578125,
0.69775390625,
0.468017578125,
0.7109375,
0.080078125,
0.7109375
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.08087158203125,
0.71337890625,
0.468017578125,
0.71337890625,
0.468017578125,
0.7265625,
0.08087158203125,
0.7265625
],
[
0.50244140625,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50244140625,
0.734375
],
[
0.08251953125,
0.72900390625,
0.46728515625,
0.72900390625,
0.46728515625,
0.7392578125,
0.08251953125,
0.7392578125
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.5009765625,
0.75,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.5009765625,
0.7646484375
],
[
0.098876953125,
0.75830078125,
0.468017578125,
0.75830078125,
0.468017578125,
0.771484375,
0.098876953125,
0.771484375
],
[
0.07843017578125,
0.771484375,
0.4697265625,
0.77294921875,
0.4697265625,
0.787109375,
0.07843017578125,
0.7861328125
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.7861328125,
0.4697265625,
0.78857421875,
0.4697265625,
0.80322265625,
0.080078125,
0.80029296875
],
[
0.50341796875,
0.78271484375,
0.88818359375,
0.78271484375,
0.88818359375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.80322265625,
0.468017578125,
0.80322265625,
0.468017578125,
0.81640625,
0.080078125,
0.81640625
],
[
0.07843017578125,
0.81640625,
0.46728515625,
0.8173828125,
0.46728515625,
0.83203125,
0.07843017578125,
0.83056640625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.83349609375,
0.468017578125,
0.83203125,
0.468017578125,
0.8466796875,
0.080078125,
0.84765625
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.8486328125,
0.468017578125,
0.8486328125,
0.468017578125,
0.8623046875,
0.08087158203125,
0.8623046875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.86279296875,
0.465576171875,
0.86279296875,
0.465576171875,
0.87744140625,
0.080078125,
0.87744140625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.10296630859375,
0.88525390625,
0.43701171875,
0.88720703125,
0.43701171875,
0.9013671875,
0.10296630859375,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Projective Manifold Gradient Laver for Deep Rotation Regression",
"YingdaYinTolga Birdal3.4",
"He Wang!",
"Baoquan Chen'Leonidas J. Guibas3",
"Jiayi Chenl,2",
"ICFCS, Peking University2Beijing Institute for General AI",
"3Stanford University",
"4Imperial CollegeLondon",
"",
"Abstract",
"resentation from [46] and 10D symmetric matrix represen-",
"tation from [27]. Recently, Levinson et al. [25] adopted",
"Regressing rotations on SO(3) manifold using deep neu-",
"the vanilla 9D matrix representation discovering that simply",
"ral networks is an important yet unsolved problem. The",
"replacing the Gram-Schmidt process in the 6D representa-",
"gap between the Euclidean network output space and the",
"tion [46] with symmetric SVD-based orthogonalization can",
"non-Euclidean SO(3) manifold imposes a severe challenge",
"make this representation superior to the others.",
"3",
"for neural network learning in both forward and back-",
"Despite the progress on discovering better rotation rep-",
"ward passes.While several works have proposed differ-",
"resentations, the gap between a Euclidean network output",
"ent regression-friendly rotation representations,very few",
"space and the non-Euclidean SO(3) manifold hasn't been",
"works have been devoted to improving the gradient back-",
"LAD",
"completely filled. One important yet long-neglected prob-",
"propagating in the backward pass. In this paper, we pro-",
"lem lies in optimization on non-Euclidean manifolds [1]: to",
"",
"SO",
"pose a manifold-aware gradient that directly backpropa-",
"optimize on SO(3) manifold, the optimization variable is",
"gates into deep network weights. Leveraging Riemannian",
"a rotation matrix, which contains nine matrix elements; if",
"optimization to construct a novel projective gradient, our",
"we naively use Euclidean gradient, which simply computes",
"1A159",
"proposed regularized projective manifold gradient (RPMG)",
"the partial derivatives with respect to each of the nine ma-",
"method helps networks achieve new state-of-the-art perfor-",
"trix elements, to update the variable, this optimization step",
"mance in a variety of rotation estimation tasks. Our pro-",
"will usually lead to a new matrix off SO(3) manifold. Un-",
"posed gradient layer can also be applied to other smooth",
"fortunately, we observe that all the existing works on rota-",
"manifolds such as the unit sphere.Our project page is at",
"tion regression simply rely upon vanilla auto-differentiation",
"二",
"",
"https://jiychen18.github.io/RPMG.",
"for backpropagation, exactly computing Euclidean gradient",
"[O1",
"and performing such off-manifold updates to predicted ro-",
"tations. We argue that, for training deep rotation regression",
"ZAIXI",
"1.Introduction",
"networks, the off-manifold components will lead to noise in",
"the gradient of neural network weights, hindering network",
"Estimating rotations is a crucial problem in visual per-",
"training and convergence.",
"ception that has broad applications, e.g., in object pose esti-",
"To tackle this issue, we draw inspiration from differen-",
"mation, robot control, camera relocalization, 3D reconstruc-",
"tial geometry, where people leverage Riemannian optimiza-",
"tion and visual odometry [8, 13, 16,22,37]. Recently, with",
"tion to optimize on the non-Euclidean manifold, which finds",
"the proliferation of deep neural networks, learning to ac-",
"the direction of the steepest geodesic path on the manifold",
"curately regress rotations is attracting more and more atten-",
"and take an on-manifold step. We thus propose to lever-",
"tion.However,thenon-Euclideancharacteristicsofrotation",
"ageRiemannianoptimization and delvedeepintothestudy",
"space make accurately regressing rotation very challenging.",
"of the backward pass. Note that this is a fundamental yet",
"As we know, rotations reside in a non-Euclidean man-",
"ifold, SO(3) group, whereas the unconstrained outputs of",
"currently under-explored avenue, given that most of the ex-",
"neural networks usually live in Euclidean spaces. This gap",
"isting works focus on a holistic design of rotation regres-",
"sion that is agnostic to forward/backward pass. However,",
"between the neural network output space and SO(3) man-",
"ifold becomes a major challenge for deep rotation regres-",
"incorporating Riemannian optimization into network train-",
"ing is highly non-trivial and challenging. Although meth-",
"sion, thus tackling this gap becomes an important research",
"ods of Riemannian optimization allow for optimization on",
"topic. One popular research direction is to design learning-",
"SO(3) [5, 31], matrix manifolds [1] or general Riemannian",
"friendly rotation representations, e.g., 6D continuous rep-",
"manifolds [34,44], they are not directly applicable to update",
"t: He Wang is the corresponding author (hewang@pku.edu.cn).",
"the weights of the neural networks that are Euclidean. Also,",
"-"
] |
{
"arxiv_id": "2110.11657",
"language": "en",
"page_idx": 0
}
|
||
2110.11657-en-1
|
2110.11657-en-1.jpg
|
approaches like[17] incorporate a Riemannian distance as well as its gradient into network training, however, they do not deal with the representation issue.
In this work, we want to propose a better manifoldaware gradient in the backward pass of rotation regression that directly updates the neural network weights. We begin by taking a Riemannian optimization step and computing the difference between the rotation prediction and the updated rotation, which is closer to the ground truth. Backpropagating this "error", we encounter the mapping function (or orthogonalization function) that transforms the raw network output to a valid rotation. This projection, which can be the Gram-Schmidt process or SVD orthogonalization [25], is typically a many-to-one mapping. This nonbijectivity provides us with a new design space for our gradient: if we were to use a gradient to update the raw output rotation, many gradients would result in the same update in the final output rotation despite being completely different for backpropagating into the neural network weights. Now the problem becomes: which gradient is the best for backpropagation when many of them correspond to the same update to the output?
We observe that this problem is somewhat similar to some problems with ambiguities or multi-ground-truth issues. One example would be the symmetry issue in pose estimation: a symmetric object, e.g. a textureless cube, appears the same under many different poses, which needs to be considered when supervising the pose predictions. For supervising the learning in such a problem, Wang $et.$ $al.[39]$ proposed to use min-of-N loss [14], which only penalizes the smallest error between the prediction and all the possible ground truths. We therefore propose to find the gradient with the smallest norm that can update the final output rotation to the goal rotation. This back-projection process involves finding an element closest to the network output in the inverse image of the goal rotation and projecting the network output to this inverse image space. We therefore coin our gradient projective manifold gradient. One thing to note is that this projective gradient tends to shorten the network output, causing the norms of network output to vanish. To fix this problem, we further incorporate a simple regularization into the gradient, leading to our full solution regularized projective manifold gradient (RPMG).
Note that our proposed gradient layer operates on the raw network output and can be directly backpropagated into the network weights. Our method is very general and is not tied to a specific rotation representation. It can be coupled with different non-Euclidean rotation representations, including quaternion, 6D representation [46], and 9D rotation matrix representation [25], and can even be used for regressing other non-manifold variables.
We evaluate our devised projective manifold gradient layers on a diverse set of problems involving rota-
tion regression: 3D object pose estimation from 3D point clouds/images, rotation estimation problems without using ground truth rotation supervisions, and please see Appendix E.3 for more experiments on camera relocalization. Our method demonstrates significant and consistent improvements on all these tasks and all different rotation representations tested. Going beyond rotation estimation, we also demonstrate performance improvements on regressing unit vectors (lie on a unit sphere) as an example of an extension to other non-Euclidean manifolds.
We summarize our contribution as below:
$\cdot$ We propose a novel manifold-aware gradient layer, namely $RPMG$, for the backward pass of rotation regression, which can be applied to different rotation representations and losses and used as a “plug-in” at no actual cost.
$\bullet$ Our extensive experiments over different tasks and rotation representations demonstrate the significant improvements from using RPMG
$\bullet$ Our method can also benefit regression tasks on other
manifolds, e.g. $\mathcal{S}^2.$
## $2. \textbf{Related Work}$
Both rotation parameterization and optimization on SO(3) are well-studied topics. Early deep learning models leverage various rotation representations for pose estimation, $e.g.$, direction cosine matrix (DCM)[19,43], axisangle [12,15,35], quaternion [11,21,23,42,45] and Eulerangle [24,30,33]. Recently, [46] points out that Euler-angle, axis-angle, and quaternion are not continuous rotation representations, since their representation spaces are not homeomorphic to SO(3). As better representations for rotation regression, 6D [46], 9D [25], 10D [27] representations are proposed to resolve the discontinuity issue and improve the regression accuracy. A concurrent work[7] examines different manifold mappings theoretically and experimentally, finding out that SVD orthogonalization performs the best when regressing arbitrary rotations. Originating from general Riemannian optimization,[31] presents an easy approach for minimization on the $SO(3)$ group by constructing a local axis-angle parameterization, which is also the tangent space of $SO(3)$ manifold. They backpropagate gradient to the tangent space and use the exponential map to update the current rotation matrix. Most recently,[32] constructs a PyTorch library that supports tangent space gradient backpropagation for 3D transformation groups, $(e.g.$, $SO(3),S\hat{E}(3),Sim(3)).$This proposed library can be used to implement the Riemannian gradient in our layer.
3. Preliminaries
# 3.1. Riemannian Geometry
Following [3,4], we define an $m$-dimensional Riemannian manifold embedded in an ambient Euclidean space
2
|
[
[
0.07843017578125,
0.0921630859375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.07843017578125,
0.106689453125
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5009765625,
0.106689453125,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.07843017578125,
0.12371826171875,
0.33251953125,
0.12371826171875,
0.33251953125,
0.136962890625,
0.07843017578125,
0.136962890625
],
[
0.50341796875,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50341796875,
0.136962890625
],
[
0.09722900390625,
0.14013671875,
0.468017578125,
0.14013671875,
0.468017578125,
0.154052734375,
0.09722900390625,
0.154052734375
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.1546630859375,
0.471435546875,
0.1546630859375,
0.471435546875,
0.169189453125,
0.080078125,
0.169189453125
],
[
0.50341796875,
0.1546630859375,
0.890625,
0.1546630859375,
0.890625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.080078125,
0.17041015625,
0.468017578125,
0.17041015625,
0.468017578125,
0.184326171875,
0.080078125,
0.184326171875
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.186279296875,
0.468017578125,
0.186279296875,
0.468017578125,
0.199462890625,
0.080078125,
0.199462890625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.07843017578125,
0.1988525390625,
0.468017578125,
0.2008056640625,
0.468017578125,
0.2139892578125,
0.07843017578125,
0.212158203125
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.21533203125,
0.468017578125,
0.21533203125,
0.468017578125,
0.2291259765625,
0.080078125,
0.2291259765625
],
[
0.50244140625,
0.21337890625,
0.8935546875,
0.21337890625,
0.8935546875,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2310791015625,
0.4697265625,
0.2310791015625,
0.4697265625,
0.2442626953125,
0.080078125,
0.2442626953125
],
[
0.50244140625,
0.2291259765625,
0.7294921875,
0.2291259765625,
0.7294921875,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.2474365234375,
0.46728515625,
0.2474365234375,
0.46728515625,
0.25830078125,
0.080078125,
0.25830078125
],
[
0.521484375,
0.24365234375,
0.79736328125,
0.24365234375,
0.79736328125,
0.2568359375,
0.521484375,
0.2568359375
],
[
0.080078125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.274658203125,
0.080078125,
0.274658203125
],
[
0.52294921875,
0.26318359375,
0.89208984375,
0.26318359375,
0.89208984375,
0.27783203125,
0.52294921875,
0.27783203125
],
[
0.07843017578125,
0.274658203125,
0.46728515625,
0.27587890625,
0.46728515625,
0.290283203125,
0.07843017578125,
0.2890625
],
[
0.53515625,
0.279052734375,
0.89208984375,
0.279052734375,
0.89208984375,
0.292236328125,
0.53515625,
0.292236328125
],
[
0.07843017578125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.304931640625,
0.07843017578125,
0.304931640625
],
[
0.53515625,
0.294189453125,
0.89208984375,
0.294189453125,
0.89208984375,
0.30810546875,
0.53515625,
0.30810546875
],
[
0.07843017578125,
0.304931640625,
0.468017578125,
0.30615234375,
0.468017578125,
0.32080078125,
0.07843017578125,
0.3193359375
],
[
0.53515625,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.3212890625,
0.53515625,
0.3212890625
],
[
0.080078125,
0.3212890625,
0.471435546875,
0.3212890625,
0.471435546875,
0.335205078125,
0.080078125,
0.335205078125
],
[
0.53515625,
0.32568359375,
0.63232421875,
0.32568359375,
0.63232421875,
0.337158203125,
0.53515625,
0.337158203125
],
[
0.080078125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.080078125,
0.349853515625
],
[
0.51953125,
0.337158203125,
0.89208984375,
0.33837890625,
0.89208984375,
0.352783203125,
0.51953125,
0.3515625
],
[
0.080078125,
0.3515625,
0.4697265625,
0.3515625,
0.4697265625,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.53369140625,
0.354248046875,
0.89208984375,
0.354248046875,
0.89208984375,
0.367431640625,
0.53369140625,
0.367431640625
],
[
0.07843017578125,
0.3662109375,
0.468017578125,
0.365478515625,
0.468017578125,
0.380126953125,
0.07843017578125,
0.380615234375
],
[
0.53369140625,
0.37060546875,
0.74169921875,
0.36865234375,
0.74169921875,
0.3818359375,
0.53369140625,
0.3837890625
],
[
0.080078125,
0.3818359375,
0.468017578125,
0.3818359375,
0.468017578125,
0.395751953125,
0.080078125,
0.395751953125
],
[
0.5185546875,
0.38330078125,
0.89208984375,
0.3837890625,
0.89208984375,
0.3984375,
0.5185546875,
0.397705078125
],
[
0.080078125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.53369140625,
0.399658203125,
0.6611328125,
0.399658203125,
0.6611328125,
0.4140625,
0.53369140625,
0.4140625
],
[
0.07843017578125,
0.412353515625,
0.20751953125,
0.412353515625,
0.20751953125,
0.425537109375,
0.07843017578125,
0.425537109375
],
[
0.1004638671875,
0.4287109375,
0.468017578125,
0.4287109375,
0.468017578125,
0.442626953125,
0.1004638671875,
0.442626953125
],
[
0.50341796875,
0.427978515625,
0.6396484375,
0.427978515625,
0.6396484375,
0.44140625,
0.50341796875,
0.44140625
],
[
0.080078125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.458251953125,
0.080078125,
0.458251953125
],
[
0.521484375,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.458984375,
0.521484375,
0.458984375
],
[
0.0767822265625,
0.458251953125,
0.4697265625,
0.458984375,
0.4697265625,
0.473388671875,
0.0767822265625,
0.472900390625
],
[
0.50244140625,
0.4609375,
0.890625,
0.4609375,
0.890625,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.07843017578125,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.488037109375,
0.07843017578125,
0.488037109375
],
[
0.50244140625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.4892578125,
0.50244140625,
0.4892578125
],
[
0.080078125,
0.490478515625,
0.4697265625,
0.490478515625,
0.4697265625,
0.50390625,
0.080078125,
0.50390625
],
[
0.5009765625,
0.490478515625,
0.89208984375,
0.4892578125,
0.89208984375,
0.50390625,
0.5009765625,
0.5048828125
],
[
0.07843017578125,
0.5048828125,
0.4697265625,
0.5048828125,
0.4697265625,
0.51806640625,
0.07843017578125,
0.51806640625
],
[
0.50341796875,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.080078125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50244140625,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.53515625,
0.50244140625,
0.53515625
],
[
0.080078125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.548828125,
0.080078125,
0.548828125
],
[
0.50244140625,
0.5361328125,
0.89208984375,
0.5361328125,
0.89208984375,
0.5498046875,
0.50244140625,
0.5498046875
],
[
0.08087158203125,
0.5517578125,
0.46728515625,
0.5517578125,
0.46728515625,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.58203125,
0.89208984375,
0.58203125,
0.89208984375,
0.59521484375,
0.50244140625,
0.59521484375
],
[
0.080078125,
0.59521484375,
0.4697265625,
0.59521484375,
0.4697265625,
0.609375,
0.080078125,
0.609375
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.07843017578125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.07843017578125,
0.62353515625
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.080078125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50341796875,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.6416015625,
0.89208984375,
0.6416015625,
0.89208984375,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.080078125,
0.65576171875,
0.4697265625,
0.65576171875,
0.4697265625,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.07843017578125,
0.67041015625,
0.4697265625,
0.66943359375,
0.4697265625,
0.6845703125,
0.07843017578125,
0.68505859375
],
[
0.50244140625,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.080078125,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.5009765625,
0.68603515625,
0.890625,
0.6875,
0.890625,
0.7021484375,
0.5009765625,
0.70068359375
],
[
0.080078125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50341796875,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.080078125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.5009765625,
0.71630859375,
0.890625,
0.71728515625,
0.890625,
0.73193359375,
0.5009765625,
0.73095703125
],
[
0.080078125,
0.73193359375,
0.38232421875,
0.73193359375,
0.38232421875,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50244140625,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.09722900390625,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.09722900390625,
0.76220703125
],
[
0.50244140625,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.07843017578125,
0.7646484375,
0.4697265625,
0.7646484375,
0.4697265625,
0.77783203125,
0.07843017578125,
0.77783203125
],
[
0.50244140625,
0.76318359375,
0.89208984375,
0.76318359375,
0.89208984375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.080078125,
0.77880859375,
0.4697265625,
0.77880859375,
0.4697265625,
0.79248046875,
0.080078125,
0.79248046875
],
[
0.5009765625,
0.7763671875,
0.890625,
0.77783203125,
0.890625,
0.79248046875,
0.5009765625,
0.79150390625
],
[
0.080078125,
0.79345703125,
0.468017578125,
0.79345703125,
0.468017578125,
0.806640625,
0.080078125,
0.806640625
],
[
0.50244140625,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50244140625,
0.806640625
],
[
0.080078125,
0.8095703125,
0.468017578125,
0.8095703125,
0.468017578125,
0.82275390625,
0.080078125,
0.82275390625
],
[
0.50244140625,
0.80810546875,
0.8388671875,
0.80810546875,
0.8388671875,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.82373046875,
0.4697265625,
0.82373046875,
0.4697265625,
0.83837890625,
0.080078125,
0.83837890625
],
[
0.50244140625,
0.83056640625,
0.63818359375,
0.83056640625,
0.63818359375,
0.84521484375,
0.50244140625,
0.84521484375
],
[
0.07843017578125,
0.83837890625,
0.4697265625,
0.8388671875,
0.4697265625,
0.85400390625,
0.07843017578125,
0.85302734375
],
[
0.08087158203125,
0.85595703125,
0.275390625,
0.85595703125,
0.275390625,
0.8662109375,
0.08087158203125,
0.8662109375
],
[
0.5009765625,
0.85107421875,
0.71142578125,
0.8515625,
0.71142578125,
0.8662109375,
0.5009765625,
0.86572265625
],
[
0.098876953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"approaches like [17] incorporate a Riemannian distance as",
"tion regression: 3D object pose estimation from 3D point",
"well as its gradient into network training, however, they do",
"clouds/images, rotation estimation problems without using",
"not deal with the representation issue.",
"ground truth rotation supervisions, and please see Appendix",
"In this work, we want to propose a better manifold-",
"E.3 for more experiments on camera relocalization. Our",
"aware gradient in the backward pass of rotation regression",
"method demonstrates significant and consistent improve-",
"that directly updates the neural network weights. We begin",
"ments on all these tasks and all different rotation represen-",
"by taking a Riemannian optimization step and computing",
"tations tested. Going beyond rotation estimation, we also",
"the difference between the rotation prediction and the up-",
"demonstrate performance improvements on regressing unit",
"dated rotation, which is closer to the ground truth. Back-",
"vectors (lie on a unit sphere) as an example of an extension",
"propagating this \"error\", we encounter the mapping func-",
"to other non-Euclidean manifolds.",
"tion(ororthogonalizationfunction)thattransformstheraw",
"We summarize our contribution as below:",
"network output to a valid rotation. This projection, which",
"• We propose a novel manifold-aware gradient layer,",
"can be the Gram-Schmidt process or SVD orthogonaliza-",
"namely RPMG, for the backward pass of rotation re-",
"tion [25], is typically a many-to-one mapping. This non-",
"gression, which can be applied to different rotation",
"bijectivity provides us with a new design space for our gra-",
"representations and losses and used as a “plug-in\" at",
"dient: if we were to use a gradient to update the raw output",
"no actual cost.",
"rotation, many gradients would result in the same update in",
". Our extensive experiments over different tasks and ro-",
"the final output rotation despite being completely different",
"tation representations demonstrate the significant im-",
"for backpropagating into the neural network weights. Now",
"provements fromusingRPMG.",
"the problem becomes: which gradient is the best for back-",
". Our method can also benefit regression tasks on other",
"propagation when many of them correspond to the same up-",
"manifolds, e.g. s?.",
"date to the output?",
"We observe that this problem is somewhat similar to",
"2. Related Work",
"some problems with ambiguities or multi-ground-truth is-",
"Both rotation parameterization and optimization on",
"sues. One example would be the symmetry issue in pose",
"SO(3) are well-studied topics. Early deep learning mod-",
"estimation: a symmetric object, e.g. a textureless cube, ap-",
"els leverage various rotation representations for pose esti-",
"pears the same under many different poses, which needs",
"mation, e.g., direction cosine matrix (DCM) [19, 43], axis-",
"to be considered when supervising the pose predictions.",
"angle[12,15,35],quaternion[11,21,23,42,45]andEuler-",
"For supervising the learning in such a problem, Wang et.",
"angle [24,30,33]. Recently, [46] points out that Euler-angle,",
"al. [39] proposed to use min-of-N loss [14], which only pe-",
"axis-angle, and quaternion are not continuous rotation rep-",
"nalizesthesmallesterrorbetweenthepredictionandallthe",
"resentations, since their representation spaces are not home-",
"possible ground truths. We therefore propose to find the gra-",
"omorphic to SO(3).As better representations for rotation",
"dient with the smallest norm that can update the final output",
"regression, 6D [46], 9D [25], 10D [27] representations are",
"rotation to the goal rotation. This back-projection process",
"proposed to resolve the discontinuity issue and improve",
"involves finding an element closest to the network output in",
"the regression accuracy. A concurrent work [7] examines",
"the inverse image of the goal rotation and projecting the net-",
"different manifold mappings theoretically and experimen-",
"work output to this inverse image space. We therefore coin",
"tally, finding out that SvD orthogonalization performs the",
"our gradient projective manifold gradient. One thing to note",
"best when regressing arbitrary rotations. Originating from",
"is that this projective gradient tends to shorten the network",
"general Riemannian optimization, [31] presents an easy ap-",
"output, causing the norms of network output to vanish. To",
"proach for minimization on the SO(3) group by construct-",
"fix this problem, we further incorporate a simple regular-",
"ing a local axis-angle parameterization,which is also the",
"ization into the gradient, leading to our full solution regu-",
"tangent space of SO(3) manifold. They backpropagate gra-",
"larized projective manifold gradient (RPMG).",
"dient to the tangent space and use the exponential map to",
"Note that our proposed gradient layer operates on the raw",
"update the current rotation matrix. Most recently, [32] con-",
"network output and can be directly backpropagated into the",
"structs a PyTorch library that supports tangent space gra-",
"network weights. Our method is very general and is not tied",
"dient backpropagation for 3D transformation groups, (e.g.,",
"to a specific rotation representation. It can be coupled with",
"SO(3), SE(3), Sim(3)). This proposed library can be used",
"different non-Euclidean rotation representations, including",
"to implement the Riemannian gradient in our layer.",
"quaternion, 6D representation [46], and 9D rotation matrix",
"3. Preliminaries",
"representation [25], and can even be used for regressing",
"othernon-manifoldvariables.",
"3.1.Riemannian Geometry",
"We evaluate our devised projective manifold gradi-",
"Following [3, 4], we define an m-dimensional Rieman-",
"ent layers on a diverse set of problems involving rota-",
"nian manifold embedded in an ambient Euclidean space",
"2"
] |
{
"arxiv_id": "2110.11657",
"language": "en",
"page_idx": 1
}
|
||
2110.11657-en-2
|
2110.11657-en-2.jpg
|
$\mathcal{X}$ = $\mathbb{R} ^d$ and endowed with a Riemannian metric G $\triangleq$ $(\mathbf{G}_{\mathbf{x}})_{\mathbf{x}\in\mathcal{M}}$ to be a smooth curved space $(\mathcal{M},G).$ A vector v $\in\mathcal{X}$ is said to be tangent to $\mathcal{M}$ at x iff there exists a smooth curve $\gamma:[0,1]\mapsto\mathcal{M}$ s.t.$\gamma ( 0) = \textbf{x}$ and $\dot{\gamma } ( 0) = \textbf{v. The velocities of all such curves through}$ $\textbf{x form the tangent space T}_{\mathbf{x} }\mathcal{M} = \{ \dot{\gamma } ( 0) \mid \gamma : \mathbb{R} \mapsto$ $\mathcal{M}$ is smooth around 0 and $\gamma(0)=\mathbf{x}\}.$
Definition 1 (Riemannian gradient). For a smooth function $f:\mathcal{M}\mapsto\mathbb{R}and\forall(\mathbf{x},\mathbf{v})\in\mathcal{T}\mathcal{M},we$ define the Riemannian gradient off as the unique vector fleld grad f satisfying [6]:
$$\mathrm Df(\mathbf x)[\mathbf v]=\langle\mathbf v,\mathrm gradf(\mathbf x)\rangle_\mathbf x$$
(1)
where D$f(\mathbf{x})[\mathbf{v}]$ is the derivation of f by v. It can further be shown (see Appendix B.1) that an expression for grad f can be obtained through the projection of the Euclidean gradient orthogonally onto the tangent space
$$\mathrm{grad}f(\mathbf{x})=\nabla f(\mathbf{x})_{\parallel}=\Pi_{\mathbf{x}}\big(\nabla f(\mathbf{x})\big).$$
(2)
where $\Pi_\mathbf{x}:\mathcal{X}\mapsto\mathcal{T}_\mathbf{x}\mathcal{M}\subseteq\mathcal{X}$ is an orthogonal projector
with respect to $\langle\cdot,\cdot\rangle_{\mathbf{x}}.$
Definition 2 (Riemannian optimization). We consider gradient descent to solve the problems of min $_{\mathbf{x}\in\mathcal{M}}f(\mathbf{x}).$ For a local minimizer or a stationary point $\mathbf{x}^*off,the$ Riemannian gradient vanishes grad$f( \mathbf{x} ^{\star }) = 0\textit{ enabling a simple}$ $algorithm$, Riemannian gradient descent $(RGD):$
(3)
$$\mathbf{x}_{k+1}=R_{\mathbf{x}_{k}}(-\tau_{k}\:\mathrm{grad}f(\mathbf{x}_{k}))$$
where $\tau_k$ is the step size at iteration k and R x is the retrac-
tion usually chosen related to the exponential map.
# 3.2. Rotation Representations
There are many ways of representing a rotation: classic rotation representations, $e.g.$ Euler angles, axis-angle, and quaternion; and recently introduced regression-friendly rotation representations such as e.g.5D [46],6D [46], 9D[25] and 10D[27] representations. A majority of deep neural networks can output an unconstrained, arbitrary $n$- dimensional vector x in a Euclidean space $\mathcal{X}=\mathbb{R}^n.$ For Euler angle and axis-angle representations which use a vector from $\mathbb{R}^3$ to represent a rotation, a neural network can simply output a 3D vector; however, for quaternions, 6D, 9D or 10D representations that lies on non-Euclidean manifolds, manifold mapping function $\pi:\mathbb{R}^n\mapsto\mathcal{M}$ is generally needed for normalization or orthogonalization purposes to convert network outputs to valid elements belonging to the representation manifold. This network Euclidean output space $\chi$ is where the representation manifolds reside and therefore are also called ambient space.
Definition 3 (Rotation representation). One rotation representation, which lies on a representation manifold M, defines a surjective rotation mapping $\phi:\hat{\mathbf{x}}\in\mathcal{M}\to\phi(\hat{\mathbf{x}})\in$ SO(3) and a representation mapping function $\psi:\mathbf{R}\in$ SO$(3)\to\psi(\mathbf{R})\in\mathcal{M}$, such that $\phi(\psi)=\mathbf{R}\in$SO(3).
Definition 4 (Manifold mapping function). From an ambient space X to the representation manifold M, we can de- $fineamanifoldmappingfunction\pi:\mathbf{x}\in\mathcal{X}\to\pi(\mathbf{x})\in\mathcal{M}$, which projects a point x in the ambient, Euclidean space to $a\textit{ valid element }\hat{\mathbf{x} } = \pi ( \mathbf{x} ) \textit{on the manifold M. }$
We summarize the manifold mappings, the rotation mappings and representation mappings for several nonEuclidean rotation representations below.
Unit quaternion. Unit quaternions represent a rotation using a 4D unit vector $\mathbf{q}\in\mathcal{S}^3$ double covering the nonEuclidean 3-sphere $i. e. \textbf{q and}- \textbf{q identify the same rotation}.$ A network with a fınal linear activation can only predict $\mathbf{x}\in\mathbb{R}^4.$ The corresponding manifold mapping function is usually chosen to be a normalization step, which reads $\pi_q(\mathbf{x})=\mathbf{x}/\|\mathbf{x}\|.$ For rotation and representation mapping, we leverage the standard mappings between rotation and quaternion (see Appendix G).
6D representation and Gram-Schmidt orthogonaliza-
tion. 6D rotation representation [46], lying on Stiefel manifold $\mathcal{V}_2(\mathbb{R}^3)$, uses two orthogonal unit 3D vectors $(\hat{\mathbf{c}}_1,\hat{\mathbf{c}}_2)$ to represent a rotation, which are essentially the first two columns of a rotation matrix. Its manifold mapping $\pi_{6D}$ is done through Gram-Schmidt orthogonalization. Its rotation mapping $\phi_{6D}$ is done by adding the third column $\hat{\mathbf{c}}_{3}=\hat{\mathbf{c}}_{1}\times\hat{\mathbf{c}}_{2}.$ Its representation mapping $\psi_6D$ is simply getting rid of the third column $\hat{\mathbf{c}}_3$ from a rotation matrix.
9D representation and SVD orthogonalization. To map a raw 9D network output M to a rotation matrix,[25] use SVD orthogonalization as the manifold mapping function $\pi_{9D}$, as follows: $\pi_{9D}$ first decomposes M into its left and right singular vectors $\{\mathbf{U},\mathbf{V}^{\top}\}$ and singular values $\Sigma,\tilde{\mathbf{M}}=\tilde{\mathbf{U}}\Sigma\mathbf{V}^{\top};$ then it replaces $\Sigma$ with $\Sigma^\prime=$ $\operatorname{diag}(1,1,\det(\mathbf{U}\mathbf{V}^{\top}))$ and finally, computes $\mathbf{R}=\mathbf{U}\Sigma^{\prime}\mathbf{V}^{\top}$ to get the corresponding rotation matrix $\mathbf{R} \in$SO(3). As this representation manifold is SO(3), both the rotation and representation mapping functions are simply identity
10D representation.[27] propose a novel 10D representation for rotation matrix. The manifold mapping function $\pi_{10D}$ maps $\boldsymbol\theta\in\mathbb{R}^10$ to q$\in\mathcal{S}^3$ by computing the eigenvector corresponding to the smallest eigenvalue of A(θ), expressed as $\pi_{10D}(\mathbf{x})=\operatorname*{min}_{\mathbf{q}\in\mathcal{S}^{3}}\mathbf{q}^{\top}\mathbf{A}(\mathbf{x})\mathbf{q}$,in which
$$\mathbf{A}(\boldsymbol{\theta})\:=\:\left[\begin{array}{cccc}\theta_1&\theta_2&\theta_3&\theta_4\\\theta_2&\theta_5&\theta_6&\theta_7\\\theta_3&\theta_6&\theta_8&\theta_9\\\theta_4&\theta_7&\theta_9&\theta_{10}\end{array}\right].$$
(4)
Since the representation manifold is $\mathcal{S}^3$, the rotation and
representation mapping are the same as unit quaternion
## 3.3. Deep Rotation Regression
We conclude this section by describing the ordinary forward and backward passes of a neural network based rota-
3
|
[
[
0.080078125,
0.0889892578125,
0.4697265625,
0.08837890625,
0.4697265625,
0.1060791015625,
0.080078125,
0.106689453125
],
[
0.5009765625,
0.09027099609375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.080078125,
0.106689453125,
0.468017578125,
0.1060791015625,
0.468017578125,
0.12060546875,
0.080078125,
0.121826171875
],
[
0.5009765625,
0.10797119140625,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.5009765625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50244140625,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.50244140625,
0.1351318359375
],
[
0.07843017578125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.50244140625,
0.136962890625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.07843017578125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.07843017578125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.8095703125,
0.15283203125,
0.8095703125,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.080078125,
0.1673583984375,
0.4697265625,
0.1673583984375,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.521484375,
0.17236328125,
0.89208984375,
0.1749267578125,
0.89208984375,
0.189453125,
0.521484375,
0.1875
],
[
0.08087158203125,
0.18310546875,
0.336669921875,
0.18310546875,
0.336669921875,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.50244140625,
0.1900634765625,
0.89208984375,
0.1900634765625,
0.89208984375,
0.203857421875,
0.50244140625,
0.203857421875
],
[
0.080078125,
0.2008056640625,
0.468017578125,
0.2020263671875,
0.468017578125,
0.216552734375,
0.080078125,
0.21533203125
],
[
0.50341796875,
0.20703125,
0.77392578125,
0.20703125,
0.77392578125,
0.2171630859375,
0.50341796875,
0.2171630859375
],
[
0.08087158203125,
0.2183837890625,
0.4697265625,
0.2183837890625,
0.4697265625,
0.231689453125,
0.08087158203125,
0.231689453125
],
[
0.50341796875,
0.2203369140625,
0.89208984375,
0.2203369140625,
0.89208984375,
0.23486328125,
0.50341796875,
0.23486328125
],
[
0.080078125,
0.23291015625,
0.468017578125,
0.23291015625,
0.468017578125,
0.246826171875,
0.080078125,
0.246826171875
],
[
0.50341796875,
0.23486328125,
0.890625,
0.23486328125,
0.890625,
0.25,
0.50341796875,
0.25
],
[
0.178955078125,
0.255615234375,
0.370849609375,
0.255615234375,
0.370849609375,
0.270751953125,
0.178955078125,
0.270751953125
],
[
0.44677734375,
0.255615234375,
0.468017578125,
0.255615234375,
0.468017578125,
0.268310546875,
0.44677734375,
0.268310546875
],
[
0.50341796875,
0.250732421875,
0.890625,
0.250732421875,
0.890625,
0.264404296875,
0.50341796875,
0.264404296875
],
[
0.08087158203125,
0.272705078125,
0.468017578125,
0.272705078125,
0.468017578125,
0.285888671875,
0.08087158203125,
0.285888671875
],
[
0.5048828125,
0.267578125,
0.890625,
0.267578125,
0.890625,
0.27783203125,
0.5048828125,
0.27783203125
],
[
0.50244140625,
0.281005859375,
0.89208984375,
0.281005859375,
0.89208984375,
0.294189453125,
0.50244140625,
0.294189453125
],
[
0.07843017578125,
0.287353515625,
0.4697265625,
0.28857421875,
0.4697265625,
0.302978515625,
0.07843017578125,
0.3017578125
],
[
0.50341796875,
0.29541015625,
0.89208984375,
0.29541015625,
0.89208984375,
0.309326171875,
0.50341796875,
0.309326171875
],
[
0.07843017578125,
0.302978515625,
0.46728515625,
0.3037109375,
0.46728515625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50341796875,
0.311279296875,
0.8896484375,
0.311279296875,
0.8896484375,
0.324462890625,
0.50341796875,
0.324462890625
],
[
0.080078125,
0.318115234375,
0.340576171875,
0.3193359375,
0.340576171875,
0.333251953125,
0.080078125,
0.33203125
],
[
0.50341796875,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.1478271484375,
0.335205078125,
0.39794921875,
0.333984375,
0.39794921875,
0.3515625,
0.1478271484375,
0.352783203125
],
[
0.44677734375,
0.337158203125,
0.468017578125,
0.337158203125,
0.468017578125,
0.35107421875,
0.44677734375,
0.35107421875
],
[
0.50341796875,
0.341552734375,
0.69775390625,
0.341552734375,
0.69775390625,
0.354736328125,
0.50341796875,
0.354736328125
],
[
0.07843017578125,
0.356689453125,
0.468017578125,
0.35791015625,
0.468017578125,
0.37255859375,
0.07843017578125,
0.371826171875
],
[
0.50244140625,
0.35595703125,
0.890625,
0.35595703125,
0.890625,
0.369384765625,
0.50244140625,
0.369384765625
],
[
0.07843017578125,
0.37255859375,
0.223876953125,
0.375,
0.223876953125,
0.389404296875,
0.07843017578125,
0.386962890625
],
[
0.50244140625,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.50244140625,
0.385009765625,
0.8896484375,
0.386474609375,
0.8896484375,
0.40087890625,
0.50244140625,
0.399658203125
],
[
0.08087158203125,
0.397705078125,
0.46728515625,
0.397705078125,
0.46728515625,
0.410888671875,
0.08087158203125,
0.410888671875
],
[
0.50244140625,
0.401611328125,
0.890625,
0.401611328125,
0.890625,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.080078125,
0.41357421875,
0.4697265625,
0.41357421875,
0.4697265625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.50244140625,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.08087158203125,
0.429931640625,
0.46728515625,
0.429931640625,
0.46728515625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.50341796875,
0.43310546875,
0.88818359375,
0.43310546875,
0.88818359375,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.080078125,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.458984375,
0.39794921875,
0.458984375,
0.39794921875,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.50341796875,
0.462158203125,
0.890625,
0.462158203125,
0.890625,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.17236328125,
0.4765625,
0.374267578125,
0.4765625,
0.374267578125,
0.4912109375,
0.17236328125,
0.4912109375
],
[
0.44677734375,
0.477783203125,
0.4697265625,
0.477783203125,
0.4697265625,
0.490478515625,
0.44677734375,
0.490478515625
],
[
0.50341796875,
0.4765625,
0.87158203125,
0.4765625,
0.87158203125,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.4912109375,
0.89208984375,
0.505859375,
0.50244140625,
0.5048828125
],
[
0.08087158203125,
0.501953125,
0.465576171875,
0.501953125,
0.465576171875,
0.51513671875,
0.08087158203125,
0.51513671875
],
[
0.50244140625,
0.5068359375,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.07843017578125,
0.5166015625,
0.410888671875,
0.517578125,
0.410888671875,
0.5322265625,
0.07843017578125,
0.53076171875
],
[
0.50244140625,
0.52294921875,
0.890625,
0.52294921875,
0.890625,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.080078125,
0.54150390625,
0.3095703125,
0.54150390625,
0.3095703125,
0.55615234375,
0.080078125,
0.55615234375
],
[
0.50244140625,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.50244140625,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.098876953125,
0.56005859375,
0.46728515625,
0.56005859375,
0.46728515625,
0.57373046875,
0.098876953125,
0.57373046875
],
[
0.50341796875,
0.5673828125,
0.87109375,
0.5673828125,
0.87109375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.5009765625,
0.58056640625,
0.88818359375,
0.5791015625,
0.88818359375,
0.5966796875,
0.5009765625,
0.5986328125
],
[
0.08087158203125,
0.59326171875,
0.46728515625,
0.59326171875,
0.46728515625,
0.60400390625,
0.08087158203125,
0.60400390625
],
[
0.50341796875,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.07843017578125,
0.62060546875,
0.4697265625,
0.6220703125,
0.4697265625,
0.63623046875,
0.07843017578125,
0.63525390625
],
[
0.50244140625,
0.62744140625,
0.85302734375,
0.62744140625,
0.85302734375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.08251953125,
0.63818359375,
0.465576171875,
0.63818359375,
0.465576171875,
0.6484375,
0.08251953125,
0.6484375
],
[
0.5009765625,
0.6416015625,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.5009765625,
0.65576171875
],
[
0.08087158203125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50244140625,
0.671875
],
[
0.080078125,
0.66748046875,
0.46728515625,
0.66748046875,
0.46728515625,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.5,
0.671875,
0.8935546875,
0.66943359375,
0.8935546875,
0.6875,
0.5,
0.689453125
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.69775390625,
0.468017578125,
0.69775390625,
0.468017578125,
0.7109375,
0.080078125,
0.7109375
],
[
0.50244140625,
0.7041015625,
0.83837890625,
0.7041015625,
0.83837890625,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.7119140625,
0.468017578125,
0.7119140625,
0.468017578125,
0.72607421875,
0.080078125,
0.72607421875
],
[
0.66650390625,
0.71728515625,
0.6953125,
0.71533203125,
0.69677734375,
0.7236328125,
0.66845703125,
0.72607421875
],
[
0.07843017578125,
0.72607421875,
0.468017578125,
0.7265625,
0.468017578125,
0.7412109375,
0.07843017578125,
0.74072265625
],
[
0.65625,
0.73486328125,
0.68896484375,
0.73486328125,
0.68896484375,
0.75048828125,
0.65625,
0.75048828125
],
[
0.69775390625,
0.736328125,
0.7197265625,
0.736328125,
0.7197265625,
0.75,
0.69775390625,
0.75
],
[
0.7294921875,
0.736328125,
0.7509765625,
0.736328125,
0.7509765625,
0.75,
0.7294921875,
0.75
],
[
0.7646484375,
0.736328125,
0.78369140625,
0.736328125,
0.78369140625,
0.75,
0.7646484375,
0.75
],
[
0.080078125,
0.74365234375,
0.4697265625,
0.74365234375,
0.4697265625,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.66650390625,
0.75,
0.69384765625,
0.75,
0.69384765625,
0.7666015625,
0.66650390625,
0.7666015625
],
[
0.69775390625,
0.75048828125,
0.7197265625,
0.75048828125,
0.7197265625,
0.76513671875,
0.69775390625,
0.76513671875
],
[
0.7294921875,
0.75048828125,
0.7509765625,
0.75048828125,
0.7509765625,
0.76513671875,
0.7294921875,
0.76513671875
],
[
0.76318359375,
0.75,
0.78173828125,
0.75,
0.78173828125,
0.76513671875,
0.76318359375,
0.76513671875
],
[
0.07843017578125,
0.7568359375,
0.468017578125,
0.75830078125,
0.468017578125,
0.771484375,
0.07843017578125,
0.77099609375
],
[
0.5810546875,
0.75830078125,
0.625,
0.75830078125,
0.625,
0.77294921875,
0.5810546875,
0.77294921875
],
[
0.6689453125,
0.76318359375,
0.69140625,
0.767578125,
0.68701171875,
0.7822265625,
0.6650390625,
0.77783203125
],
[
0.869140625,
0.7587890625,
0.8935546875,
0.7587890625,
0.8935546875,
0.77294921875,
0.869140625,
0.77294921875
],
[
0.73046875,
0.76513671875,
0.7509765625,
0.76513671875,
0.7509765625,
0.77978515625,
0.73046875,
0.77978515625
],
[
0.7646484375,
0.76513671875,
0.78369140625,
0.76513671875,
0.78369140625,
0.77978515625,
0.7646484375,
0.77978515625
],
[
0.07843017578125,
0.77294921875,
0.468017578125,
0.771484375,
0.468017578125,
0.7861328125,
0.07843017578125,
0.787109375
],
[
0.69775390625,
0.7666015625,
0.7197265625,
0.7666015625,
0.7197265625,
0.77978515625,
0.69775390625,
0.77978515625
],
[
0.66650390625,
0.78076171875,
0.69287109375,
0.78076171875,
0.69287109375,
0.79443359375,
0.66650390625,
0.79443359375
],
[
0.69775390625,
0.7822265625,
0.7197265625,
0.7822265625,
0.7197265625,
0.79443359375,
0.69775390625,
0.79443359375
],
[
0.73046875,
0.7822265625,
0.75341796875,
0.7822265625,
0.75341796875,
0.79443359375,
0.73046875,
0.79443359375
],
[
0.7607421875,
0.7822265625,
0.7900390625,
0.7822265625,
0.7900390625,
0.79443359375,
0.7607421875,
0.79443359375
],
[
0.080078125,
0.78857421875,
0.468017578125,
0.78857421875,
0.468017578125,
0.8017578125,
0.080078125,
0.8017578125
],
[
0.07843017578125,
0.80322265625,
0.366943359375,
0.80322265625,
0.366943359375,
0.81640625,
0.07843017578125,
0.81640625
],
[
0.50244140625,
0.81298828125,
0.89208984375,
0.81298828125,
0.89208984375,
0.82763671875,
0.50244140625,
0.82763671875
],
[
0.080078125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.8291015625,
0.869140625,
0.8291015625,
0.869140625,
0.84228515625,
0.50244140625,
0.84228515625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.8515625,
0.734375,
0.8515625,
0.734375,
0.86572265625,
0.50244140625,
0.86572265625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.51953125,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.51953125,
0.8837890625
],
[
0.080078125,
0.88623046875,
0.43310546875,
0.88623046875,
0.43310546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.4794921875,
0.923828125,
0.491943359375,
0.923828125,
0.491943359375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"X = Rd and endowed with a Riemannian metric G ≤",
"Definition 4 (Manifold mapping function). From an ambi-",
"(Gx)xeM to be a smooth curved space (M,G). A vec-",
"ent space to the representation manifold M, we can de-",
"tor v E X is said to be tangent to M at x iff there ex-",
"fineamanifoldmappingfunctionT:xEπ(x)EM,",
"ists a smooth curve:[0,1] -M s.t.(o)=x",
"which projects a point x in the ambient, Euclidean space to",
"and (O) = v. The velocities of all such curves through",
"a valid element x = (x) on the manifold M.",
"x form the tangent space TxM = {(O)l : R -→",
"We summarize the manifold mappings, the rotation",
"M is smooth around 0 and (O) = x).",
"mappings and representation mappings for several non-",
"Definition 1 (Riemannian gradient). For a smooth function",
"Euclidean rotationrepresentations below.",
"f : M - R and V(x, v) E TM, we define the Riemannian",
"Unit quaternion. Unit quaternions represent a rotation us-",
"gradient of f as the unique vector field gradf satisfying [6]:",
"ing a 4D unit vector q E s3 double covering the non-",
"Df(x)[v] = (v,gradf(x))x",
"(1)",
"Euclidean 3-sphere i.e. q and -q identify the same rotation.",
"where Df(x)/vl is the derivation of f by V. It can further be",
"A network with a final linear activation can only predict",
"x E R4.The corresponding manifold mapping function",
"shown (see Appendix B.1) that an expression for gradf can",
"is usually chosen to be a normalization step, which reads",
"be obtained through the projection of the Euclidean gradi-",
"T.(x) = x/xl. For rotation and representation mapping,",
"ent orthogonally onto the tangent space",
"we leverage the standard mappings between rotation and",
"gradf(x) = f(x)= IIx(Vf(x)).",
"(2)",
"quaternion (see Appendix G).",
"where IIx : X TxM C X is an orthogonal projector",
"6D representation and Gram-Schmidt orthogonaliza-",
"with respect to (.,-)x.",
"tion. 6D rotation representation [46], lying on Stiefel man-",
"ifold V2(R3), uses two orthogonal unit 3D vectors (C1, C2)",
"Definition 2 (Riemannian optimization). We consider gra-",
"to represent a rotation, which are essentially the first two",
"dient descent to solve the problems of minxeM f(x). For a",
"columns of a rotation matrix. Its manifold mapping 6D",
"local minimizer or a stationary point x* of f,the Rieman-",
"isdone through Gram-Schmidt orthogonalization.Its ro-",
"nian gradient vanishes gradf(x*) = 0 enabling a simple",
"tation mapping Φ6D is done by adding the third column",
"algorithm, Riemannian gradient descent (RGD):",
"C3 = Ci × c2. Its representation mapping 6D is simply",
"Xk+1 = Rx(-Th gradf(xk))",
"(3)",
"getting rid of the third column cg from a rotation matrix.",
"9D representation and SVD orthogonalization. To map",
"where Tk is the step size at iteration k and Rx, is the retrac-",
"a raw 9D network output M to a rotation matrix, [25]",
"tion usually chosen related to the exponential map",
"use SVD orthogonalization as the manifold mapping func-",
"3.2. Rotation Representations",
"tion TgD, as follows: gD first decomposes M into its",
"left and right singular vectors [U, VT and singular val-",
"There are many ways of representing a rotation: clas-",
"ues , M = UZV; then it replaces with '",
"sic rotation representations, e.g. Euler angles, axis-angle,",
"diag(1, 1, det(UVT)) and finally, computes R = UE'VT",
"andquaternion;andrecentlyintroducedregression-friendly",
"to get the corresponding rotation matrix R ESO(3).As",
"rotation representations such as e.g. 5D [46], 6D [46]",
"this representation manifold is SO(3), both the rotation and",
"9D [25] and 10D [27] representations. A majority of deep",
"representation mapping functions are simply identity.",
"neural networks can outputan unconstrained.arbitrary n-",
"10D representation.[27] propose a novel 10D represen-",
"dimensional vector x in a Euclidean space X = Rn. For",
"tation for rotation matrix. The manifold mapping function",
"Euler angle and axis-angle representations which use a vec-",
"T10D maps R10 to q E S3 by computing the eigen-",
"tor from R3 to represent a rotation, a neural network can",
"vector corresponding to the smallest eigenvalue of A(0),",
"simply output a 3D vector; however, for quaternions, 6D,",
"expressed as πi0D(x) = min q A(x)q, in which",
"9D or 10D representations that lies on non-Euclidean man-",
"qeEs",
"ifolds,manifold mapping function π:Rn →M is gener-",
"[01",
"02",
"03",
"04",
"allyneededfornormalization ororthogonalizationpurposes",
"02",
"0s",
"06",
"07",
"to convert network outputs to valid elements belonging to",
"A(0)",
"03",
"(4)",
"0g",
"0g",
"the representation manifold. This network Euclidean out-",
"06",
"04",
"07",
"0g",
"010",
"put space X is where the representation manifolds reside",
"and therefore are also called ambient space.",
"Since the representation manifold is s3, the rotation and",
"Definition 3 (Rotation representation). One rotation repre-",
"representation mapping are the same as unit quaternion.",
"sentation, which lies on a representation manifold M, de-",
"fines a surjective rotation mapping Φ : × E M → (x) E",
"3.3. Deep Rotation Regression",
"SO(3) and a representation mapping function : R E",
"We conclude this section by describing the ordinary for-",
"SO(3) -→ (R) E M, such that p() = R E SO(3).",
"ward and backward passes of a neural network based rota-",
"3"
] |
{
"arxiv_id": "2110.11657",
"language": "en",
"page_idx": 2
}
|
||
2110.11657-en-3
|
2110.11657-en-3.jpg
|
<table>
<tbody>
<tr>
<th>$I_{\mathbf{n}\mathbf{n}\mathbf{n}\mathbf{u}}$</th>
<th>Naurnl</th>
<th>S</th>
<th>Ambient $\mathcal{X}$ space $\mathbf{R}$ n 17</th>
<th>Manifold</th>
<th>Representation Manifold$\mathcal{M}$</th>
<th>1 Rotation m 2</th>
<th>$M$ 1 $y$</th>
<th>Rotation Manifold SO(3) Outnut</th>
<th> </th>
</tr>
<tr>
<th>Input signal</th>
<th>Neural Network $\operatorname{Backprop}.$</th>
<th> </th>
<th>Raw output $X$</th>
<th> ${\mathrm{mapping}}$ $\overline{\pi(\mathbf{x})}$</th>
<th>$\hat{\mathbf{x}}$</th>
<th>rtvutivin mapping $\phi(\hat{\mathbf{x}})$</th>
<th> </th>
<th>Output rotation $\mathbf{R}$</th>
<th>Loss $\mathcal{L}(f(\mathbf{R}))$</th>
</tr>
<tr>
<td>$\mathbf{Manifold}$ Gradient</td>
<td>$gp$ $+\lambda(\mathbf{x}_{gp}$ X</td>
<td>1 I 1 八 一 </td>
<td>$\mathbf{x}_{gp}$</td>
<td>$\arg\min||\mathbf{x}_g$ $\mathbf{x}_{g}\in\pi^{-1}\left(\hat{\mathbf{x}}_{g}\right)$</td>
<td>$-g$</td>
<td>$\psi(\mathbf{R}_g)$</td>
<td>I 1 八</td>
<td>$\mathbf{R}_{g}$ 1 1</td>
<td>Riemannian optimization</td>
</tr>
</tbody>
</table>
Figure 1. Projective Manifold Gradient Layer. In the forward pass, the network predicts a raw output x, which is then transformed into a valid rotation $\mathbf{R}=\phi(\pi(\mathbf{x})).$ We leave this forward pass unchanged and only modify the backward pass. In the backward pass, we first use Riemannian optimization to get a goal rotation $\mathbf{R}_{g}$ and map it back to $\hat{\mathbf{x}}_{g}$ on the representation manifold $\mathcal{M}.$ After that we find the element $\mathbf{x}_{gp}$ which is closest to the raw output in the inverse image of $\hat{\mathbf{x}}_g$, and finally get the gradient g$\mathbf{RPM}$ we want.
flow loss as used in PoseCNN [42]) and self-supervised rotation estimation cases $(e.g.,2$D mask loss as used in [38]).
In this work, we introduce the following techniques to mitigate these problems: (i) we first take a Riemannian gradient to compute a goal rotation $\mathbf{R} _g\in$SO(3), which does not rely on knowing $\mathbf{R}_{gt}$, as explained in Section 4.1;(ii) we then find the set of all possible $\mathbf{x}_g$s that can be mapped to $\mathbf{R}_g$,or in other words, the inverse image of $\mathbf{R}_g$ under $\pi$ and $\phi;$(iii) we find $\mathbf{x}_{gp}$ which is the element in this set closest to x in the Euclidean metric and set it as “x””. We will construct our gradient using this $\mathbf{x}^*$, as explained in 4.2. (iv) we add a regularization term to this gradient forming g$_{RPMG}$ as explained in 4.3. The whole backward pass leveraging our proposed regularized projective manifold gradient is shown in the lower half of Figure 1.
tion regression, as used in [25,46].
Forward and backward passes. Assume, for a rotation representation, the network predicts $\mathbf{x}\in\mathcal{X}$,then the mani- $\hat {\text{fold mapping }\pi \text{ will map x to }\hat{\mathbf{x} } }= \pi ( \mathbf{x} ) \in \mathcal{M} $,followed by a rotation mapping $\phi$ that finally yields the output rotation $\mathbf{R} = \phi ( \hat{\mathbf{x} } )$ = $\phi ( \pi ( \mathbf{x} ) ) .$ Our work only tackles the backward pass and keeps the forward pass unchanged, as shown in the top part of Figure l. The gradient in the backwardpass is simply computed using Pytorch autograd method, that is g$=f^\prime(\mathbf{R})\phi^{\prime}(\hat{\mathbf{x}})\pi^{\prime}(\mathbf{x}).$
Loss function. The most common choice for supervising rotation matrix is L2 loss, $\|\mathbf{R}-\mathbf{R}_{gt}\|_F^2$, as used by [25,46]. This loss is equal to $4-4\cos(<\mathbf{R},\mathbf{R}_{gt}>)$,where $<\mathbf{R},\mathbf{R}_{gt}>$represents the angle between $\mathbf{R}$ and $\mathbf{R}_gt.$
### $4. \textbf{Method}$
4.1. Riemannian Gradient and Goal Rotation
Overview. In this work, we propose a projective manifold $gradient\textit{layer, without changing the forward pass of a given}$ rotation regressing network, as shown in Figure l. Our focus is to find a better gradient g of the loss function $\mathcal{L}$ with respect to the network raw output x for backpropagation into the network weights.
Let's start with examining the gradient of network output x in a general case- regression in Euclidean space. Given a ground truth $\mathbf{x}_{gt}$ and the L2 loss $\|\mathbf{x}-\mathbf{x}_{gt}\|^2$ that maximizes the likelihood in the presence of Gaussian noise in x, the gradient would be g$=2(\mathbf{x}-\mathbf{x}_{gt}).$
In the case of rotation regression, we therefore propose to find a proper $\mathbf{x}^*\in\mathcal{X}$ for a given ground truth $\mathbf{R}_{gt}$ or a computed goal rotation $\mathbf{R}_g$ when the ground truth rotation is not available, and then simply use $\mathbf{x}-\mathbf{x}^*$ as our gradient to backpropagate into the network.
Note that finding such an $\mathbf{x}^*$ can be challenging. Assuming we know $\mathbf{R}_{gt}$, finding an x$^*$ involves inverting $\phi$ and $\pi$ since the network output $\mathbf{R}=\phi(\pi(\mathbf{x})).$ Furthermore, we may not know $\mathbf{R}_{gt}$ under indirect rotation supervision $(e.g.$,
To handle rotation estimation with/without direct rotation supervision, we first propose to compute the Riemannian gradient of the loss function $\mathcal{L}$ with respect to the output rotation $\mathbf{R}$ and find a goal rotation $\mathbf{R}_g$ that is presumably closer to the ground truth rotation than R.
Assume the loss function is in the following form $\mathcal{L}(f(\mathbf{R}))$, where $\mathbf{R}=\pi(\phi(\mathbf{x}))$ is the output rotation and $f$ constructs a loss function that compares R to the ground truth rotation $\mathbf{R}_{gt}$ directly or indirectly. Given $\tilde{\mathbf{R}}(\mathbf{x})$ and $\mathcal{L}(f(\mathbf{R}(\mathbf{x})))$, we can perform one step of Riemannian optimization yielding our goal rotation $\mathbf{R}_g\leftarrow$ $R_{\mathbf{R}}(-\tau$grad $\mathcal{L}(f(\mathbf{R})))$,where $\tau$ is the step size of Riemannian gradient and can be set to a constant as a hyperparameter or varying during the training. For L2 loss $\| \mathbf{R} - \mathbf{R} _{gt}\| _F^2$, the Riemannian gradient is always along the $\text{geodesic path between R and R}_{gt}$ on SO( 3) [ 20] . In this case, $\mathbf{R}_g$ can generally be seen as an intermediate goal between $\mathbf{R}$ and $\mathbf{R}_{gt}$ dependent on $\tau.$ Gradually increasing $\tau$ from 0 will first make $\mathbf{R}_g$ approach $\mathbf{R}_gt$ starting with $\mathbf{R}_{g}=\mathbf{R}$,and then reach $\mathbf{R}_gt$ where we denote $\tau=\tau_gt$,and
4
|
[
[
0.340576171875,
0.08837890625,
0.396240234375,
0.08837890625,
0.396240234375,
0.1029052734375,
0.340576171875,
0.1029052734375
],
[
0.5,
0.0889892578125,
0.5888671875,
0.0889892578125,
0.5888671875,
0.10101318359375,
0.5,
0.10101318359375
],
[
0.6806640625,
0.08587646484375,
0.73779296875,
0.08837890625,
0.736328125,
0.1029052734375,
0.6787109375,
0.10101318359375
],
[
0.338134765625,
0.1029052734375,
0.390625,
0.10101318359375,
0.39208984375,
0.11553955078125,
0.339111328125,
0.117431640625
],
[
0.5009765625,
0.1029052734375,
0.57861328125,
0.1029052734375,
0.57861328125,
0.1142578125,
0.5009765625,
0.1142578125
],
[
0.5986328125,
0.10797119140625,
0.654296875,
0.10797119140625,
0.654296875,
0.11932373046875,
0.5986328125,
0.11932373046875
],
[
0.66650390625,
0.10418701171875,
0.7607421875,
0.10418701171875,
0.7607421875,
0.11553955078125,
0.66650390625,
0.11553955078125
],
[
0.59619140625,
0.11932373046875,
0.65625,
0.121826171875,
0.654296875,
0.1363525390625,
0.5947265625,
0.1337890625
],
[
0.426513671875,
0.10919189453125,
0.48193359375,
0.10919189453125,
0.48193359375,
0.12060546875,
0.426513671875,
0.12060546875
],
[
0.214111328125,
0.121826171875,
0.261474609375,
0.121826171875,
0.261474609375,
0.1363525390625,
0.214111328125,
0.1363525390625
],
[
0.346435546875,
0.12249755859375,
0.3798828125,
0.12249755859375,
0.3798828125,
0.1351318359375,
0.346435546875,
0.1351318359375
],
[
0.4248046875,
0.12249755859375,
0.483642578125,
0.12249755859375,
0.483642578125,
0.136962890625,
0.4248046875,
0.136962890625
],
[
0.69140625,
0.11932373046875,
0.73876953125,
0.11932373046875,
0.73876953125,
0.1337890625,
0.69140625,
0.1337890625
],
[
0.53173828125,
0.1300048828125,
0.55126953125,
0.1300048828125,
0.55126953125,
0.143310546875,
0.53173828125,
0.143310546875
],
[
0.11767578125,
0.12371826171875,
0.15283203125,
0.12371826171875,
0.15283203125,
0.1363525390625,
0.11767578125,
0.1363525390625
],
[
0.11517333984375,
0.136962890625,
0.1568603515625,
0.136962890625,
0.1568603515625,
0.1514892578125,
0.11517333984375,
0.1514892578125
],
[
0.2115478515625,
0.136962890625,
0.2646484375,
0.136962890625,
0.2646484375,
0.1483154296875,
0.2115478515625,
0.1483154296875
],
[
0.346435546875,
0.136962890625,
0.3857421875,
0.136962890625,
0.3857421875,
0.1483154296875,
0.346435546875,
0.1483154296875
],
[
0.4345703125,
0.1383056640625,
0.4755859375,
0.1383056640625,
0.4755859375,
0.15283203125,
0.4345703125,
0.15283203125
],
[
0.6044921875,
0.1383056640625,
0.64453125,
0.1383056640625,
0.64453125,
0.154052734375,
0.6044921875,
0.154052734375
],
[
0.68896484375,
0.1351318359375,
0.740234375,
0.1351318359375,
0.740234375,
0.146484375,
0.68896484375,
0.146484375
],
[
0.3544921875,
0.149658203125,
0.371826171875,
0.149658203125,
0.371826171875,
0.1634521484375,
0.3544921875,
0.1634521484375
],
[
0.70166015625,
0.146484375,
0.72216796875,
0.146484375,
0.72216796875,
0.160400390625,
0.70166015625,
0.160400390625
],
[
0.79833984375,
0.1578369140625,
0.833984375,
0.1610107421875,
0.8310546875,
0.175537109375,
0.7958984375,
0.17236328125
],
[
0.2459716796875,
0.1717529296875,
0.310546875,
0.173583984375,
0.310546875,
0.1881103515625,
0.2459716796875,
0.186279296875
],
[
0.7861328125,
0.1767578125,
0.85888671875,
0.1767578125,
0.85888671875,
0.1956787109375,
0.7861328125,
0.1956787109375
],
[
0.43310546875,
0.1962890625,
0.4755859375,
0.1962890625,
0.4755859375,
0.2108154296875,
0.43310546875,
0.2108154296875
],
[
0.5849609375,
0.1956787109375,
0.654296875,
0.1956787109375,
0.654296875,
0.210205078125,
0.5849609375,
0.210205078125
],
[
0.6953125,
0.1988525390625,
0.72802734375,
0.1988525390625,
0.72802734375,
0.2108154296875,
0.6953125,
0.2108154296875
],
[
0.214111328125,
0.2052001953125,
0.273681640625,
0.2052001953125,
0.273681640625,
0.216552734375,
0.214111328125,
0.216552734375
],
[
0.3349609375,
0.208984375,
0.39453125,
0.2108154296875,
0.393798828125,
0.2259521484375,
0.334228515625,
0.2235107421875
],
[
0.101318359375,
0.199462890625,
0.181396484375,
0.199462890625,
0.181396484375,
0.2139892578125,
0.101318359375,
0.2139892578125
],
[
0.42236328125,
0.210205078125,
0.484375,
0.212158203125,
0.484375,
0.2266845703125,
0.42236328125,
0.2247314453125
],
[
0.59619140625,
0.210205078125,
0.65283203125,
0.210205078125,
0.65283203125,
0.2247314453125,
0.59619140625,
0.2247314453125
],
[
0.68701171875,
0.21337890625,
0.74658203125,
0.21337890625,
0.74658203125,
0.2235107421875,
0.68701171875,
0.2235107421875
],
[
0.744140625,
0.2108154296875,
0.87158203125,
0.2108154296875,
0.87158203125,
0.2247314453125,
0.744140625,
0.2247314453125
],
[
0.2099609375,
0.21533203125,
0.285888671875,
0.2197265625,
0.28515625,
0.2374267578125,
0.2091064453125,
0.23291015625
],
[
0.1094970703125,
0.21533203125,
0.1748046875,
0.21533203125,
0.1748046875,
0.2266845703125,
0.1094970703125,
0.2266845703125
],
[
0.53173828125,
0.216552734375,
0.560546875,
0.2203369140625,
0.55615234375,
0.2374267578125,
0.52783203125,
0.23291015625
],
[
0.3505859375,
0.2259521484375,
0.384033203125,
0.2259521484375,
0.384033203125,
0.2412109375,
0.3505859375,
0.2412109375
],
[
0.410888671875,
0.2247314453125,
0.484375,
0.2266845703125,
0.484375,
0.2412109375,
0.410888671875,
0.2392578125
],
[
0.705078125,
0.2259521484375,
0.72802734375,
0.2259521484375,
0.72802734375,
0.240478515625,
0.705078125,
0.240478515625
],
[
0.77783203125,
0.2279052734375,
0.85107421875,
0.2279052734375,
0.85107421875,
0.2392578125,
0.77783203125,
0.2392578125
],
[
0.60107421875,
0.2291259765625,
0.64892578125,
0.2291259765625,
0.64892578125,
0.24365234375,
0.60107421875,
0.24365234375
],
[
0.11273193359375,
0.2291259765625,
0.17236328125,
0.2291259765625,
0.17236328125,
0.240478515625,
0.11273193359375,
0.240478515625
],
[
0.1895751953125,
0.23291015625,
0.301513671875,
0.236083984375,
0.301513671875,
0.25390625,
0.1895751953125,
0.250732421875
],
[
0.4794921875,
0.2298583984375,
0.50244140625,
0.2298583984375,
0.50244140625,
0.2374267578125,
0.4794921875,
0.2374267578125
],
[
0.11273193359375,
0.2412109375,
0.1707763671875,
0.2412109375,
0.1707763671875,
0.25244140625,
0.11273193359375,
0.25244140625
],
[
0.43701171875,
0.240478515625,
0.46728515625,
0.240478515625,
0.46728515625,
0.248779296875,
0.43701171875,
0.248779296875
],
[
0.77685546875,
0.2412109375,
0.85302734375,
0.2412109375,
0.85302734375,
0.25244140625,
0.77685546875,
0.25244140625
],
[
0.080078125,
0.268310546875,
0.89208984375,
0.268310546875,
0.89208984375,
0.2822265625,
0.080078125,
0.2822265625
],
[
0.416748046875,
0.242431640625,
0.43310546875,
0.242431640625,
0.43310546875,
0.24560546875,
0.416748046875,
0.24560546875
],
[
0.080078125,
0.28271484375,
0.89208984375,
0.28271484375,
0.89208984375,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.07843017578125,
0.296630859375,
0.89208984375,
0.296630859375,
0.89208984375,
0.31005859375,
0.07843017578125,
0.31005859375
],
[
0.07843017578125,
0.31005859375,
0.77685546875,
0.31005859375,
0.77685546875,
0.323974609375,
0.07843017578125,
0.323974609375
],
[
0.080078125,
0.33837890625,
0.30810546875,
0.33837890625,
0.30810546875,
0.3515625,
0.080078125,
0.3515625
],
[
0.50244140625,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.35107421875,
0.50244140625,
0.35107421875
],
[
0.080078125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.50244140625,
0.352783203125,
0.88818359375,
0.352783203125,
0.88818359375,
0.3662109375,
0.50244140625,
0.3662109375
],
[
0.07843017578125,
0.371826171875,
0.4697265625,
0.371826171875,
0.4697265625,
0.385009765625,
0.07843017578125,
0.385009765625
],
[
0.521484375,
0.369384765625,
0.89208984375,
0.369384765625,
0.89208984375,
0.38330078125,
0.521484375,
0.38330078125
],
[
0.080078125,
0.386474609375,
0.46728515625,
0.386474609375,
0.46728515625,
0.399658203125,
0.080078125,
0.399658203125
],
[
0.50244140625,
0.385009765625,
0.890625,
0.385009765625,
0.890625,
0.3984375,
0.50244140625,
0.3984375
],
[
0.07843017578125,
0.401611328125,
0.468017578125,
0.40087890625,
0.468017578125,
0.415283203125,
0.07843017578125,
0.416748046875
],
[
0.50244140625,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.417236328125,
0.468017578125,
0.417236328125,
0.468017578125,
0.43115234375,
0.080078125,
0.43115234375
],
[
0.50244140625,
0.415283203125,
0.890625,
0.415283203125,
0.890625,
0.4287109375,
0.50244140625,
0.4287109375
],
[
0.080078125,
0.431884765625,
0.4697265625,
0.431884765625,
0.4697265625,
0.44580078125,
0.080078125,
0.44580078125
],
[
0.50244140625,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.080078125,
0.4462890625,
0.468017578125,
0.4462890625,
0.468017578125,
0.4609375,
0.080078125,
0.4609375
],
[
0.50341796875,
0.44580078125,
0.890625,
0.44580078125,
0.890625,
0.458984375,
0.50341796875,
0.458984375
],
[
0.080078125,
0.462158203125,
0.46728515625,
0.462158203125,
0.46728515625,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.080078125,
0.477783203125,
0.273681640625,
0.477783203125,
0.273681640625,
0.4912109375,
0.080078125,
0.4912109375
],
[
0.50244140625,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.07843017578125,
0.494384765625,
0.468017578125,
0.49560546875,
0.468017578125,
0.51025390625,
0.07843017578125,
0.5087890625
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.5009765625,
0.5048828125,
0.89208984375,
0.505859375,
0.89208984375,
0.52099609375,
0.5009765625,
0.51953125
],
[
0.080078125,
0.51123046875,
0.468017578125,
0.51123046875,
0.468017578125,
0.5244140625,
0.080078125,
0.5244140625
],
[
0.080078125,
0.5244140625,
0.4697265625,
0.52587890625,
0.4697265625,
0.54052734375,
0.080078125,
0.5390625
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.53515625,
0.50244140625,
0.53515625
],
[
0.080078125,
0.54052734375,
0.43701171875,
0.54150390625,
0.43701171875,
0.55615234375,
0.080078125,
0.5546875
],
[
0.5009765625,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.5009765625,
0.548828125
],
[
0.50244140625,
0.55029296875,
0.75341796875,
0.55029296875,
0.75341796875,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.57373046875,
0.1666259765625,
0.57373046875,
0.1666259765625,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57568359375,
0.8486328125,
0.57568359375,
0.8486328125,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.080078125,
0.5947265625,
0.4697265625,
0.59521484375,
0.4697265625,
0.60986328125,
0.080078125,
0.609375
],
[
0.521484375,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.611328125,
0.521484375,
0.611328125
],
[
0.080078125,
0.611328125,
0.46728515625,
0.611328125,
0.46728515625,
0.625,
0.080078125,
0.625
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.62548828125,
0.46728515625,
0.62548828125,
0.46728515625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.5009765625,
0.62890625,
0.890625,
0.62744140625,
0.890625,
0.6416015625,
0.5009765625,
0.642578125
],
[
0.0767822265625,
0.6416015625,
0.468017578125,
0.64013671875,
0.468017578125,
0.65478515625,
0.0767822265625,
0.65576171875
],
[
0.50244140625,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.07843017578125,
0.65576171875,
0.4697265625,
0.6572265625,
0.4697265625,
0.671875,
0.07843017578125,
0.67041015625
],
[
0.50244140625,
0.6591796875,
0.8095703125,
0.6591796875,
0.8095703125,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.07843017578125,
0.67236328125,
0.2484130859375,
0.67236328125,
0.2484130859375,
0.68603515625,
0.07843017578125,
0.68603515625
],
[
0.521484375,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.521484375,
0.68798828125
],
[
0.09722900390625,
0.6875,
0.4697265625,
0.6875,
0.4697265625,
0.7021484375,
0.09722900390625,
0.7021484375
],
[
0.5048828125,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.5048828125,
0.7041015625
],
[
0.08087158203125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71533203125,
0.08087158203125,
0.71533203125
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.5009765625,
0.72021484375,
0.89208984375,
0.7197265625,
0.89208984375,
0.734375,
0.5009765625,
0.73486328125
],
[
0.07843017578125,
0.73193359375,
0.4697265625,
0.73291015625,
0.4697265625,
0.74755859375,
0.07843017578125,
0.7470703125
],
[
0.50341796875,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50341796875,
0.75
],
[
0.07843017578125,
0.74755859375,
0.310546875,
0.74853515625,
0.310546875,
0.76318359375,
0.07843017578125,
0.76220703125
],
[
0.50341796875,
0.75244140625,
0.88623046875,
0.75244140625,
0.88623046875,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.098876953125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.098876953125,
0.77880859375
],
[
0.50341796875,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.77978515625,
0.4697265625,
0.77978515625,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.8125,
0.468017578125,
0.8125,
0.468017578125,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.5048828125,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.07843017578125,
0.82568359375,
0.3095703125,
0.82568359375,
0.3095703125,
0.8388671875,
0.07843017578125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.09722900390625,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.09722900390625,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5009765625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.0767822265625,
0.88623046875,
0.471435546875,
0.88720703125,
0.471435546875,
0.9013671875,
0.0767822265625,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.9013671875,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"Ambient",
"Representation",
"Rotation",
"spacex",
"ManifoldM",
"Rotation",
"ManifoldSO(3)",
"mapping",
"Manifold",
"Neural",
"Raw",
"mapping",
"Output",
"",
"Input",
"signal",
"Network",
"output",
"元(x)",
"(x)",
"rotation",
"x",
"R",
"Loss",
"Backprop.",
"C(f(R)",
"Invert",
"Represent.",
"Goal",
"gRPM=",
"Projected",
"Regularized",
"& project",
"mapping",
"rotation",
"RR(-T grad C(f(R)",
"x - Xgp",
"Projective",
"Xg",
"Xgp",
"arg min|xg",
"Rg",
"Riemannian",
"(Rg)",
"Manifold",
"+入(×gp-xg)",
"xl",
"Gradient",
"-1(xg)",
"optimization",
"Figure 1. Projective Manifold Gradient Layer. In the forward pass, the network predicts a raw output x, which is then transformed into",
"",
"a valid rotation R = p((x). We leave this forward pass unchanged and only modify the backward pass. In the backward pass, we first",
"use Riemannian optimization to get a goal rotation Rg and map it back to xg on the representation manifold M. After that we find the",
"element xgp which is closest to the raw output in the inverse image of xg, and finally get the gradient grpm we want.",
"tion regression, as used in [25, 46].",
"flow loss as used in PoseCNN [42]) and self-supervised ro-",
"Forward and backward passes. Assume, for a rotation",
"tation estimation cases (e.g., 2D mask loss as used in [38]).",
"representation, the network predicts x E X, then the mani-",
"In this work, we introduce the following techniques to",
"fold mapping π will map x to x = π(x) E M, followed by",
"mitigate these problems: (i) we first take a Riemannian gra-",
"a rotation mapping that finally yields the output rotation",
"dient to compute a goal rotation Rg E SO(3), which does",
"R = (x) = (π(x)). Our work only tackles the back-",
"not rely on knowing Rqt, as explained in Section 4.l; (ii)",
"ward pass and keeps the forward pass unchanged, as shown",
"we then find the set of all possible Xgs that can be mapped to",
"in the top part of Figure 1. The gradient in the backward-",
"Rg, or in other words, the inverse image of Rg under and",
"pass is simply computed using Pytorch autograd method",
"p; (ii) we find Xgp which is the element in this set closest",
"thatis g = f'(R)@'(x)π'(x)",
"to x in the Euclidean metric and set it as “x*\". We will con-",
"Loss function. The most common choice for supervis-",
"struct our gradient using this x*, as explained in 4.2. (iv) we",
"add a regularization term to this gradient forming gRPMG",
"ing rotation matrix is L2 loss, R -Rqt /2 , as used by",
"[25,46]. This loss is equal to 4-4 cos(<R, Rgt >), where",
"as explained in 4.3. The whole backward pass leverag-",
"<R, Rgt > represents the angle between R and Rgt.",
"ing our proposed regularized projective manifold gradient",
"is shown in the lower half of Figure 1.",
"4. Method",
"4.1. Riemannian Gradient and Goal Rotation",
"Overview. In this work, we propose a projective manifold",
"To handle rotation estimation with/without direct rota-",
"gradient layer, without changing the forward pass of a given",
"tion supervision, we first propose to compute the Rieman-",
"rotation regressing network, as shown in Figure 1. Our fo-",
"nian gradient of the loss function C with respect to the out-",
"cus is to find a better gradient g of the loss function C with",
"put rotation R and find a goal rotation Rg that is presum-",
"respect to the network raw output x for backpropagation",
"ably closer to the ground truth rotation than R.",
"into the network weights.",
"Assume the loss function is in the following form",
"Let's start with examining the gradient of network output",
"C(f(R), where R = (o(x)) is the output rotation",
"xinageneralcase-regressioninEuclideanspace.Givena",
"and f constructs a loss function that compares R to the",
"ground truth Xqt and the L2 lossx -Xqt2 that maximizes",
"ground truth rotation Rgt directly or indirectly. Given",
"the likelihood in the presence of Gaussian noise in x, the",
"R(x) and L(f(R(x), we can perform one step of Rie-",
"gradient would beg = 2(x-Xgt)",
"mannian optimization yielding our goal rotation Rg←",
"In the case of rotation regression, we therefore propose",
"Rr(-T grad C(f(R))), where T is the step size of Rie-",
"to find a proper x* E X for a given ground truth Rgt or a",
"mannian gradient and can beset to a constant as a hy-",
"computed goal rotation Rg when the ground truth rotation",
"perparameter or varying during the training. For L2 loss",
"isnotavailable,andthen simplyusex-x*as ourgradient",
"[R - Rqtl2, the Riemannian gradient is always along the",
"to backpropagate into the network.",
"geodesic path between R and Rgt on SO(3) [20]. In this",
"Note that finding such an x* can be challenging. Assum-",
"case, Rg can generally be seen as an intermediate goal be-",
"ing we know Rgt, finding an x* involves inverting and ",
"tween R and Rgt dependent on T. Gradually increasing",
"since the network output R = (π(x)). Furthermore, we",
"↑ from O will first make Rg approach Rgt starting with",
"may not know Rgt under indirect rotation supervision (e.g.,",
"Rg = R, and then reach Rgt where we denote T = Tgt, and",
"4"
] |
{
"arxiv_id": "2110.11657",
"language": "en",
"page_idx": 3
}
|
||
2110.11657-en-4
|
2110.11657-en-4.jpg
|
<FigureHere>
Figure 2. IIlustration for regularized projective manifold gradient. Left: In the forward pass, we simply project x to $\hat{\mathbf{x}}$by $\pi.$ In the backward pass, first we compute a Riemannian gradient, which is shown as the green arrow. After getting a next goal $\hat{\mathbf{x}}_g\in\mathcal{M}$ by Riemannian optimization, we find the inverse projection $\mathbf{x}_{gp}$ of $\hat{\mathbf{x}}_{g}$, which leads to our projective manifold gradient, shown as the blue arrow. With a regularization term, we can get our final regularized projective manifold gradient, as the purple arrow. Right: Projection point $\hat{\mathbf{x}}_{gp}$ in the case of quaternion.
finally going beyond $\mathbf{R}_{gt}.$ Although, when $\mathbf{R}_gt$ is available, one can simply set $\mathbf{R}_g=\mathbf{R}_{gt}$,we argue that this is just a special case under $\tau=\tau_{gt}.$ For scenarios where $\mathbf{R}_{gt}$ is unavailable, $e.g.$, in self-supervised learning cases (see in Section 5.3), we don't know $\mathbf{R}_{gt}$ and $\tau_{gt}$, thus we need to compute $\mathbf{R}_g$ using Riemannian optimization. In the sequel, we only use $\mathbf{R}_g$ for explaining our methods without loss of generality. See Section 4.3 for how to choose $\tau.$
## $4. 2. \textbf{Projective Manifold Gradient}$
Given $\mathbf{R}_g$, we can use the representation mapping $\psi$ to find the corresponding $\hat{\mathbf{x}}_g=\psi(\mathbf{R}_g)$ on the representation manifold $\mathcal{M}.$ However, further inverting $\pi$ and finding the corresponding $\mathbf{x}_g\in\mathcal{X}$ is a non-trivial problem, due to the projective nature of $\pi.$ In fact, there are many $\mathbf{x}_g$s that satisfy $\pi(\mathbf{x}_g)=\hat{\mathbf{x}}_g.$ It seems that we can construct a gradient $\mathbf{g}=(\mathbf{x}-\mathbf{x}_{g})$ using any $\mathbf{x}_g$ that satisfies $\pi(\mathbf{x}_{g})=\hat{\mathbf{x}}_{g}.$ No matter which $\mathbf{x}_g$ we choose, if this gradient were to update $\mathbf{x}$,it will result in the same $\mathbf{R}_g.$ But, when backpropagating into the network, those gradients will update the network weights differently, potentially resulting in different learning efficiency and network performance.
Formally, we formulate this problem as $a$ multi-groundtruth problem for x: we need to find the best x* to supervise from the inverse image of $\hat{\mathbf{x}}_g$ under the mapping $\pi.$ We note that similar problems have been seen in pose supervision dealing with symmetry as in [39], where one needs to find one pose to supervise when there are many poses under which the object appears the same.[39] proposed to use a min-of-N strategy introduced by [14]: from all possible poses, taking the pose that is closest to the network prediction as ground truth. A similar strategy is also seen in supervising quaternion regression, as q and -q stand for the same rotation. One common choice of the loss function is therefore $\min\{\mathcal{L}(\mathbf{q},\mathbf{q}_{gt}),\mathcal{L}(\mathbf{q},-\mathbf{q}_{gt})\}$[27],which penalizes the distance to the closest ground truth quaternion.
Inspired by these works, we propose to choose our gradient among all the possible gradients with the lowest level of redundancy, $i.e.$, we require x* to be the one closest to $\mathbf{x}$, or in other words, the gradient to have the smallest norm, meaning that we need to find the projection point $\mathbf{x}_{gp}$ of x to all the valid $\mathbf{x}_g.$
(5)
$$\mathbf{x}_{gp}=\underset{\pi(\mathbf{x}_g)=\hat{\mathbf{x}}_g}{\operatorname*{argmin}}\|\mathbf{x}-\mathbf{x}_g\|_2$$
We then can construct our projective manifold gradient $\mathrm{( PMG) ~as~g}_{PM}= \mathbf{x} - \mathbf{x} _{gp}.$ We will denote the naive gra- $\operatorname{dient}\mathbf{g}_{M}=\mathbf{x}-\hat{\mathbf{x}}_{g}$ as manifold gradient (MG).
Here we provide another perspective on why a network may prefer PMG. In the case where a deep network is trained using stochastic gradient descent (SGD), the final gradient used to update the network weights is averaged across the gradients of all the batch instances. If gradients
from different batch instances contain different levels of redundancy, then the averaged gradient may be biased or not even appropriate. This argument is generally applicable to all stochastic optimizers $(e.g.$, Adam [2])
Inverting $\pi.$ There are many ways to solve this projection problem for different manifold mapping functions $\pi.$ For example, we can formulate this as a constrained optimization problem. For the manifold mapping functions we consider, we propose the following approach: we first solve for the inverse image $\pi^-1(\hat{\mathbf{x}}_g)$ of $\hat{\mathbf{x}}_g$ in the ambient space $\mathcal{X}$ $\underset{\hat{\text{analytically, which reads }\pi^{-1}(\hat{\mathbf{x}}_{g})=\{\mathbf{x}_{g}\in\mathcal{X}\mid\hat{\pi}(\mathbf{x}_{g})=}}$ $\hat{\mathbf{x}}_g\};$we then project x onto this inverse image space. Note that, sometimes only a superset of this inverse image can be found analytically, requiring certain constraints on x$_{gp}$ to be enforced.
Here we list the inverse image $\pi^-1(\hat{\mathbf{x}}_g)$ and the projection point $\mathbf{x}_{gp}$ for different rotation representations and their corresponding manifold mapping $\pi.$ Please refer to Appendix B.2 for detailed derivations.
$\textbf{Quaternion. With }\pi _q( \mathbf{x} ) = \mathbf{x} / \| \mathbf{x} \| , \mathbf{x} \in \mathbb{R} ^4$,and$\hat\mathbf{x}_g\in\mathcal{S}^3:$ $\pi_{q}^{-1}(\hat{\mathbf{x}}_{g})=\{\mathbf{x}\mid\mathbf{x}=k\hat{\mathbf{x}}_{g},k\in\mathbb{R}$ and $k>0\}$, which is a ray in the direction of $\hat{\mathbf{x}}_g$ starting from the origin. Without considering the constraint of $k>0$ ,an analytical solution to this projection point $\mathbf{x}_{gp}$ of x onto this line can be de- ${\mathrm{rived: ~}\mathbf{x} }_{gp}= ( \mathbf{x} \cdot \hat{\mathbf{x} } _{g}) \hat{\mathbf{x} } _{g}.$
6D representation. With $\pi_{6D}$ as Gram-Schmidt process, $\mathbf{x}$ = $[ \mathbf{u} , \mathbf{v} ]$ $\in$ $\mathbb{R} ^{6}$, and $\hat{\mathbf{x} } _g$ $\in$ $\mathcal{V} _{2}( \mathbb{R} ^{3}) \nobreak {: } $ $\pi _{6D}^{- 1}( \hat{\mathbf{x} } _{g})$ = $\{[k_{1}\hat{\mathbf{u}}_{g},k_{2}\hat{\mathbf{u}}_{g}+k_{3}\hat{\mathbf{v}}_{g}]\mid k_{1},\tilde{k}_{2},k_{3}\in\mathbb{R}$ and $k_{1},k_{3}>0\}$ (the former is a ray whereas the latter spans a half plane). Without considering the constraint of $k_1,k_3>0$, the projection point $\mathbf{x}_{gp}$ can be analytically represented as $\mathbf{x}_{gp}=$ $[(\mathbf{u}\cdot\hat{\mathbf{u}}_{g})\hat{\mathbf{u}}_{g},(\mathbf{v}\cdot\hat{\mathbf{u}}_{g})\hat{\mathbf{u}}_{g}+(\mathbf{v}\cdot\hat{\mathbf{v}}_{g})\hat{\mathbf{v}}_{g}]$
9D representation. With $\pi_{9D}(\mathbf{x})$ as SVD orthogonalization, x$\in R^3\times 3$, and $\hat{\mathbf{x} } _{g}\in$SO( 3) , the analytical expression for $\pi_{9D}^{-1}$ is available when we ignore the positive singular value constraints, which gives $\pi_{9D}^{-1}(\hat{\mathbf{x}}_{g})~=~\{\hat{\mathbf{S}}\hat{\mathbf{x}}_{g}\mid\tilde{\mathbf{S}}~=$
5
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.1383056640625,
0.080078125,
0.1383056640625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5185546875,
0.1363525390625,
0.529296875,
0.1331787109375,
0.53369140625,
0.1395263671875,
0.52294921875,
0.1427001953125
],
[
0.07843017578125,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.744140625,
0.169189453125,
0.79345703125,
0.169189453125,
0.79345703125,
0.17236328125,
0.744140625,
0.17236328125
],
[
0.7939453125,
0.1641845703125,
0.87353515625,
0.1641845703125,
0.87353515625,
0.1749267578125,
0.7939453125,
0.1749267578125
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.07843017578125,
0.1976318359375,
0.39453125,
0.1962890625,
0.39453125,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.50341796875,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.08087158203125,
0.2235107421875,
0.339111328125,
0.2235107421875,
0.339111328125,
0.238037109375,
0.08087158203125,
0.238037109375
],
[
0.50244140625,
0.2266845703125,
0.89208984375,
0.2266845703125,
0.89208984375,
0.240478515625,
0.50244140625,
0.240478515625
],
[
0.098876953125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.25830078125,
0.098876953125,
0.25830078125
],
[
0.50341796875,
0.242431640625,
0.890625,
0.242431640625,
0.890625,
0.25244140625,
0.50341796875,
0.25244140625
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.2734375,
0.080078125,
0.2734375
],
[
0.5009765625,
0.25390625,
0.89208984375,
0.255126953125,
0.89208984375,
0.268310546875,
0.5009765625,
0.267578125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.268310546875,
0.89208984375,
0.268310546875,
0.89208984375,
0.2822265625,
0.50244140625,
0.2822265625
],
[
0.50244140625,
0.2822265625,
0.89208984375,
0.2822265625,
0.89208984375,
0.29541015625,
0.50244140625,
0.29541015625
],
[
0.07843017578125,
0.2890625,
0.4697265625,
0.28857421875,
0.4697265625,
0.302978515625,
0.07843017578125,
0.3037109375
],
[
0.50244140625,
0.296630859375,
0.89208984375,
0.296630859375,
0.89208984375,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.080078125,
0.304931640625,
0.468017578125,
0.304931640625,
0.468017578125,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50244140625,
0.31005859375,
0.77783203125,
0.31005859375,
0.77783203125,
0.323974609375,
0.50244140625,
0.323974609375
],
[
0.07843017578125,
0.3193359375,
0.468017578125,
0.318115234375,
0.468017578125,
0.333251953125,
0.07843017578125,
0.333984375
],
[
0.080078125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.07843017578125,
0.348388671875,
0.468017578125,
0.349853515625,
0.468017578125,
0.3642578125,
0.07843017578125,
0.363037109375
],
[
0.50341796875,
0.352783203125,
0.8896484375,
0.352783203125,
0.8896484375,
0.3662109375,
0.50341796875,
0.3662109375
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50244140625,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50244140625,
0.3818359375
],
[
0.07843017578125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.07843017578125,
0.39453125
],
[
0.50244140625,
0.3837890625,
0.89208984375,
0.3837890625,
0.89208984375,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.50341796875,
0.3984375,
0.771484375,
0.3984375,
0.771484375,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.07843017578125,
0.410400390625,
0.346435546875,
0.410888671875,
0.346435546875,
0.4248046875,
0.07843017578125,
0.423583984375
],
[
0.50244140625,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.09722900390625,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.09722900390625,
0.439453125
],
[
0.50244140625,
0.429931640625,
0.89208984375,
0.429931640625,
0.89208984375,
0.443115234375,
0.50244140625,
0.443115234375
],
[
0.07843017578125,
0.43994140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.07843017578125,
0.453857421875
],
[
0.50341796875,
0.4462890625,
0.890625,
0.4462890625,
0.890625,
0.45703125,
0.50341796875,
0.45703125
],
[
0.07843017578125,
0.455810546875,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.07843017578125,
0.47021484375
],
[
0.50244140625,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.5009765625,
0.473388671875,
0.89208984375,
0.474853515625,
0.89208984375,
0.4892578125,
0.5009765625,
0.488037109375
],
[
0.08087158203125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.080078125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.505859375,
0.890625,
0.505859375,
0.890625,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50341796875,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.080078125,
0.53076171875,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.5048828125,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.54736328125,
0.5048828125,
0.54736328125
],
[
0.080078125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.080078125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.50341796875,
0.56640625,
0.5849609375,
0.56640625,
0.5849609375,
0.57763671875,
0.50341796875,
0.57763671875
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.521484375,
0.58056640625,
0.890625,
0.58056640625,
0.890625,
0.59521484375,
0.521484375,
0.59521484375
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.07843017578125,
0.62255859375,
0.441162109375,
0.62255859375,
0.441162109375,
0.63623046875,
0.07843017578125,
0.63623046875
],
[
0.50341796875,
0.626953125,
0.736328125,
0.626953125,
0.736328125,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.09722900390625,
0.63720703125,
0.46728515625,
0.63720703125,
0.46728515625,
0.65087890625,
0.09722900390625,
0.65087890625
],
[
0.50341796875,
0.6416015625,
0.89208984375,
0.6416015625,
0.89208984375,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.080078125,
0.65283203125,
0.468017578125,
0.65283203125,
0.468017578125,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.6572265625,
0.89208984375,
0.671875,
0.50244140625,
0.67236328125
],
[
0.080078125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.080078125,
0.681640625
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.080078125,
0.68310546875,
0.468017578125,
0.68310546875,
0.468017578125,
0.6962890625,
0.080078125,
0.6962890625
],
[
0.5009765625,
0.68798828125,
0.8935546875,
0.68798828125,
0.8935546875,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.07843017578125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.07843017578125,
0.7119140625
],
[
0.0767822265625,
0.708984375,
0.2034912109375,
0.71337890625,
0.2034912109375,
0.72900390625,
0.0767822265625,
0.724609375
],
[
0.50341796875,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.50341796875,
0.7197265625,
0.66259765625,
0.7197265625,
0.66259765625,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.50341796875,
0.734375,
0.890625,
0.734375,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.1805419921875,
0.74072265625,
0.366943359375,
0.7392578125,
0.366943359375,
0.75830078125,
0.1805419921875,
0.7587890625
],
[
0.44677734375,
0.7421875,
0.4697265625,
0.7421875,
0.4697265625,
0.755859375,
0.44677734375,
0.755859375
],
[
0.50244140625,
0.7470703125,
0.87158203125,
0.74853515625,
0.87158203125,
0.7666015625,
0.50244140625,
0.7646484375
],
[
0.226318359375,
0.755859375,
0.2900390625,
0.755859375,
0.2900390625,
0.767578125,
0.226318359375,
0.767578125
],
[
0.5048828125,
0.76513671875,
0.88818359375,
0.76513671875,
0.88818359375,
0.77880859375,
0.5048828125,
0.77880859375
],
[
0.080078125,
0.77880859375,
0.4697265625,
0.77978515625,
0.4697265625,
0.79443359375,
0.080078125,
0.79345703125
],
[
0.50341796875,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79443359375,
0.471435546875,
0.79443359375,
0.471435546875,
0.8125,
0.080078125,
0.8125
],
[
0.50244140625,
0.79345703125,
0.89208984375,
0.79443359375,
0.89208984375,
0.8095703125,
0.50244140625,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.39453125,
0.81103515625,
0.39453125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.8896484375,
0.81005859375,
0.8896484375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.098876953125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.098876953125,
0.84033203125
],
[
0.50341796875,
0.82470703125,
0.75341796875,
0.82568359375,
0.75341796875,
0.84033203125,
0.50341796875,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.8388671875,
0.89208984375,
0.84033203125,
0.89208984375,
0.85498046875,
0.5009765625,
0.85400390625
],
[
0.080078125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.080078125,
0.86865234375
],
[
0.50244140625,
0.85595703125,
0.8935546875,
0.85595703125,
0.8935546875,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.8935546875,
0.87060546875,
0.8935546875,
0.88818359375,
0.5009765625,
0.88818359375
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.8935546875,
0.8837890625,
0.8935546875,
0.9013671875,
0.5009765625,
0.90283203125
],
[
0.4794921875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"finally going beyond Rgt. Although, when Rqt is available,",
"one can simply set Rg = Rgt, we argue that this is just a",
"special case under = Tgt. For scenarios where Rgt is",
"unavailable, e.g., in self-supervised learning cases (see in",
"M",
"Section 5.3), we don't know Rgt and Tgt, thus we need to",
"computeR,usingRiemannian optimization.Inthe sequel.",
"",
"x,xgplkx",
"we only use R. for explaining our methods without loss of",
"generality.See Section 4.3 for how to choose T.",
"Figure 2. Illustration for regularized projective manifold gra-",
"dient. Left: In the forward pass, we simply project x to x by . In",
"4.2.Projective Manifold Gradient",
"the backward pass, first we compute a Riemannian gradient, which",
"Given Rg, we can use the representation mapping to",
"is shown as thegreenarrow.After getting a nextgoal xg E M",
"find the corresponding xg = (Rg) on the representation",
"by Riemannian optimization, we find the inverse projection xgp of",
"manifold M. However, further inverting and finding the",
"Xg, which leads to our projective manifold gradient, shown as the",
"blue arrow. With a regularization term, we can get our final regu-",
"corresponding xg E X is a non-trivial problem, due to the",
"larized projective manifold gradient, as the purple arrow. Right:",
"projective nature of π. In fact, there are many xgs that sat-",
"Projection point xXgp in the case of quaternion.",
"isfy (xg)=×xg.It seems that we can construct a gradient",
"g = (x - xg) using any xg that satisfies π(x) = xq. No",
"matter which xg we choose, if this gradient were to update",
"from different batch instances contain different levels of re-",
"x, it will result in the same Rg. But, when backpropagating",
"dundancy, then the averaged gradient may be biased or not",
"into the network, those gradients will update the network",
"even appropriate. This argument is generally applicable to",
"weights differently, potentially resulting in different learn-",
"all stochastic optimizers (e.g., Adam [2])",
"ing efficiencyand networkperformance.",
"Inverting r. There are many ways to solve this projection",
"Formally, we formulate this problem as a multi-ground-",
"problem for different manifold mapping functions .For",
"truth problem for x: we need to find the best x* to super-",
"example,wecanformulatethis as aconstrained optimiza-",
"vise from the inverse image of ×g under the mapping . We",
"tion problem. For the manifold mapping functions we con-",
"note that similar problems have been seen in pose supervi-",
"sider, we propose the following approach: we first solve for",
"siondealingwith symmetry as in[39l,where oneneedsto",
"the inverse image π-1(xg) of xg in the ambient space X",
"find one pose to supervise when there are many poses un-",
"analytically, which reads π-1(×g) = [xg E X [π(xg) =",
"der which the object appears the same. [39] proposed to use",
"x.: we then project x onto this inverse image space. Note",
"a min-of-N strategy introduced by [14]: from all possible",
"that.sometimes onlyasupersetofthisinverseimagecanbe",
"poses, taking the pose that is closest to the network predic-",
"found analytically, requiring certain constraints on Xgp to",
"tion as ground truth. A similar strategy is also seen in su-",
"be enforced.",
"pervising quaternion regression, as q and -q stand for the",
"Here we list the inverse image -1(xg) and the projec-",
"same rotation.. One common choice of the loss function is",
"tion point Xgp for different rotation representations and their",
"therefore min{C(q, qgt), C(q, -qgt)} [27], which penal-",
"corresponding manifold mapping π.Please refer to Ap-",
"izes the distance to the closest ground truth quaternion.",
"pendix B.2 for detailed derivations.",
"Inspired by these works, we propose to choose our gra-",
"Quaternion. With πa(x) = x/lxl, x E R4, and x。 E S3:",
"dient among all the possible gradients with the lowest level",
"-l(xg) =(xx = kxg,k e R and k > O), which is a",
"of redundancy, i.e., we require x* to be the one closest to",
"ray in the direction of xg starting from the origin. Without",
"x, or in other words, the gradient to have the smallest norm,",
"considering the constraint of k > O, an analytical solution",
"meaning that we need to find the projection point Xgp of x",
"to all the valid xg:",
"to this projection point Xgp of x onto this line can be de-",
"rived: Xgp = (x·xg)xg",
"6D representation. With 6D as Gram-Schmidt process,",
"Xgp = argmin Ix-Xgll2",
"(5)",
"x = [u,v] E R6, and xg V2(R3): 6D(×g)",
"元(xg)=×g",
"[[kitg, k2ug + k3vg] I ki, k2, k3 E R and ki, k3 > 0]",
"We then can construct our projective manifold gradient",
"(the former is a ray whereas the latter spans a half plane).",
"(PMG) as gPM = x - Xgp. We will denote the naive gra-",
"Without considering the constraint of ki,k3 > O, the pro-",
"dient gM = x - xg as manifold gradient (MG).",
"jection point xgp can be analytically represented as xgp=",
"Here we provide another perspective on why a network",
"[(u-ug)ug, (v.ug)ug + (v.vg)vg)",
"may prefer PMG. In the case where a deep network is",
"9D representation. With gD(x) as SVD orthogonaliza-",
"trained using stochastic gradient descent (SGD), the final",
"tion, x E R3x3, and xg E SO(3), the analytical expression",
"gradient used to update the network weights is averaged",
"for gD is available when we ignore the positive singular",
"across the gradients of all the batch instances. If gradients",
"value constraints, which gives gD(×g) =[Sxg | S =",
"5"
] |
{
"arxiv_id": "2110.11657",
"language": "en",
"page_idx": 4
}
|
||
2106.11486-en-0
|
2106.11486-en-0.jpg
|
$\textbf{Unsupervised Embedding Adaptation via Early- Stage Feature Reconstruction}$
$\textbf{for Few- Shot Classification}$
$\textbf{Dong Hoon Lee}^{1}\textbf{ Sae-Young Chung}^{1}$
## $\mathbf{Abstract}$
We propose unsupervised embedding adaptation for the downstream few-shot classification task. Based on findings that deep neural networks learn to generalize before memorizing, we develop Early-Stage Feature Reconstruction (ESFR)—a novel adaptation scheme with feature reconstruction and dimensionality-driven early stopping that finds generalizable features. Incorporating ESFR consistently improves the performance of baseline methods on all standard settings, including the recently proposed transductive method. ESFR used in conjunction with the transductive method further achieves state-of-the-art performance on mini-ImageNet, tiered-ImageNet, and CUB; especially with $1.2\%\sim2.0\%$ improvements in accuracy over the previous best performing method on 1-shot setting.$^{\mathrm{l}}$
## $1. \textbf{Introduction}$
Deep learning has achieved impressive results on visual recognition tasks. However, it still has difficulty generalizing to novel classes with few examples; while humans can learn to recognize from few experiences. Few-shot classification (Miller et al., 2000; Vinyals et al., 2016; Ravi & Larochelle, 2017) is designed to bridge the gap between the two and has recently attracted substantial attention.
Several works (Liu et al., 2019; Hou et al., 2019; Qiao et al., 2019; Hu et al., 2020; Dhillon et al., 2020; Ziko et al., 2020; Boudiaf et al., 2020) have shown the effectiveness of transductive methods in few-shot classification, showing a significant improvement over inductive methods. While test samples are inaccessible in an inductive few-shot classification setting, one can utilize all the unlabeled test samples together to make an inference in a transductive
$^{\mathrm{l}}$School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea. Correspondence to: Dong Hoon Lee <donghoonlee@kaist.ac.kr>.
Proceedings of the 38 th International Conference on Machine
$Learning$,PMLR 139, 2021. Copyright 2021 by the author(s).
$^{1}$Code is available at https://github.com/movinghoon/ESFR
setting. The co-existence of labeled- and unlabeled-data in this setting motivates the use of transductive inference or semi-supervised learning. A popular transductive approach is pseudo-label-based methods that progressively update the labels or inference models by predicted test samples. For instance, Liu et al. (2019); Kim et al. (2019) use a neighbor graph for label propagation, Hou et al.(2019); Liu et al. (2020) use predicted labels on test samples to update the class prototypes, and Antoniou & Storkey (2019); Hu et al. (2020) use prediction to produce an intrinsic loss or synthetic gradient. Another line of works utilizes regularization terms on unlabeled test samples. To list a few, Dhillon et al. (2020); Boudiaf et al. (2020) use entropy minimization of the prediction on unlabeled data, Ziko et al. (2020) uses the Laplacian regularization term for graph clustering. These methods are mostly originated from semi-supervised learning studies; approaches in semi-supervised learning are still motivating few-shot classification research.
On the other hand, semi-supervised learning research has recently benefited from unsupervised learning. Rapidly advancing self-supervised learning methods (Caron et al., 2020; Grill et al., 2020; Chen et al., 2020) have shown strong performance on semi-supervised image classification tasks. A popular approach is to use a self-supervision loss for representation learning to acquire more general features. Learned representations are then used with fine-tuning (Caron et al., 2020; Grill et al., 2020; Chen et al., 2020) or other semisupervised learning methods (Zhai et al., 2019; Kim et al., $\hat{2021})$ for downstream tasks. These studies achieved stateof-the-art performance on semi-supervised learning tasks, especially in settings with extremely few labels.
The success of unsupervised learning on semi-supervised tasks suggests the potential benefit of finding shared features or patterns without labels in relevant research areas. In fewshot classification, several works (Gidaris et al., 2019; Su et al., 2020) use additional self-supervision loss during the training of base datasets to learn more general features. However, the use of unsupervised learning on unlabeled data that appears in test-time is less studied. In this work, we study unsupervised learning for adaptation to satisfy the thirst.
|
[
[
0.098876953125,
0.1142578125,
0.8759765625,
0.1142578125,
0.8759765625,
0.1287841796875,
0.098876953125,
0.1287841796875
],
[
0.351318359375,
0.1363525390625,
0.62109375,
0.1363525390625,
0.62109375,
0.15087890625,
0.351318359375,
0.15087890625
],
[
0.3544921875,
0.1976318359375,
0.615234375,
0.1976318359375,
0.615234375,
0.212158203125,
0.3544921875,
0.212158203125
],
[
0.2442626953125,
0.240478515625,
0.3203125,
0.240478515625,
0.3203125,
0.255126953125,
0.2442626953125,
0.255126953125
],
[
0.5,
0.2412109375,
0.88623046875,
0.240478515625,
0.88623046875,
0.255126953125,
0.5,
0.255615234375
],
[
0.12091064453125,
0.261962890625,
0.439453125,
0.261962890625,
0.439453125,
0.27587890625,
0.12091064453125,
0.27587890625
],
[
0.50244140625,
0.2587890625,
0.8857421875,
0.2587890625,
0.8857421875,
0.26953125,
0.50244140625,
0.26953125
],
[
0.0228729248046875,
0.270751953125,
0.056365966796875,
0.270751953125,
0.056365966796875,
0.396484375,
0.0228729248046875,
0.396484375
],
[
0.5009765625,
0.272705078125,
0.88623046875,
0.272705078125,
0.88623046875,
0.285888671875,
0.5009765625,
0.285888671875
],
[
0.12335205078125,
0.279052734375,
0.439453125,
0.279052734375,
0.439453125,
0.2890625,
0.12335205078125,
0.2890625
],
[
0.5,
0.285888671875,
0.88623046875,
0.287353515625,
0.88623046875,
0.3017578125,
0.5,
0.300537109375
],
[
0.1217041015625,
0.292236328125,
0.441162109375,
0.292236328125,
0.441162109375,
0.30615234375,
0.1217041015625,
0.30615234375
],
[
0.12091064453125,
0.30810546875,
0.441162109375,
0.30810546875,
0.441162109375,
0.3212890625,
0.12091064453125,
0.3212890625
],
[
0.5009765625,
0.302978515625,
0.8857421875,
0.302978515625,
0.8857421875,
0.31640625,
0.5009765625,
0.31640625
],
[
0.1217041015625,
0.322509765625,
0.441162109375,
0.322509765625,
0.441162109375,
0.3359375,
0.1217041015625,
0.3359375
],
[
0.5,
0.318115234375,
0.8857421875,
0.318115234375,
0.8857421875,
0.33203125,
0.5,
0.33203125
],
[
0.12091064453125,
0.3359375,
0.439453125,
0.3359375,
0.439453125,
0.349853515625,
0.12091064453125,
0.349853515625
],
[
0.5,
0.33203125,
0.88623046875,
0.330810546875,
0.88623046875,
0.34521484375,
0.5,
0.3466796875
],
[
0.12091064453125,
0.3515625,
0.439453125,
0.3515625,
0.439453125,
0.365478515625,
0.12091064453125,
0.365478515625
],
[
0.5009765625,
0.347900390625,
0.88623046875,
0.347900390625,
0.88623046875,
0.361083984375,
0.5009765625,
0.361083984375
],
[
0.1217041015625,
0.36865234375,
0.441162109375,
0.36865234375,
0.441162109375,
0.3818359375,
0.1217041015625,
0.3818359375
],
[
0.5009765625,
0.363037109375,
0.88623046875,
0.363037109375,
0.88623046875,
0.376953125,
0.5009765625,
0.376953125
],
[
0.1217041015625,
0.38330078125,
0.441162109375,
0.38330078125,
0.441162109375,
0.396484375,
0.1217041015625,
0.396484375
],
[
0.5009765625,
0.37744140625,
0.88623046875,
0.37744140625,
0.88623046875,
0.391357421875,
0.5009765625,
0.391357421875
],
[
0.12091064453125,
0.3984375,
0.441162109375,
0.3984375,
0.441162109375,
0.412353515625,
0.12091064453125,
0.412353515625
],
[
0.5009765625,
0.393310546875,
0.88623046875,
0.393310546875,
0.88623046875,
0.4072265625,
0.5009765625,
0.4072265625
],
[
0.0228729248046875,
0.399658203125,
0.060455322265625,
0.399658203125,
0.060455322265625,
0.4892578125,
0.0228729248046875,
0.4892578125
],
[
0.12091064453125,
0.41357421875,
0.441162109375,
0.41357421875,
0.441162109375,
0.427978515625,
0.12091064453125,
0.427978515625
],
[
0.5009765625,
0.40771484375,
0.88623046875,
0.40771484375,
0.88623046875,
0.421630859375,
0.5009765625,
0.421630859375
],
[
0.1217041015625,
0.4287109375,
0.441162109375,
0.4287109375,
0.441162109375,
0.442626953125,
0.1217041015625,
0.442626953125
],
[
0.5009765625,
0.423583984375,
0.88623046875,
0.423583984375,
0.88623046875,
0.4375,
0.5009765625,
0.4375
],
[
0.1217041015625,
0.443115234375,
0.441162109375,
0.443115234375,
0.441162109375,
0.45703125,
0.1217041015625,
0.45703125
],
[
0.5009765625,
0.438232421875,
0.88623046875,
0.438232421875,
0.88623046875,
0.451904296875,
0.5009765625,
0.451904296875
],
[
0.12091064453125,
0.458251953125,
0.44287109375,
0.458251953125,
0.44287109375,
0.4716796875,
0.12091064453125,
0.4716796875
],
[
0.5009765625,
0.453857421875,
0.88623046875,
0.453857421875,
0.88623046875,
0.46728515625,
0.5009765625,
0.46728515625
],
[
0.1217041015625,
0.473388671875,
0.441162109375,
0.473388671875,
0.441162109375,
0.4873046875,
0.1217041015625,
0.4873046875
],
[
0.50244140625,
0.47021484375,
0.8837890625,
0.47021484375,
0.8837890625,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.12091064453125,
0.488037109375,
0.44287109375,
0.488037109375,
0.44287109375,
0.50244140625,
0.12091064453125,
0.50244140625
],
[
0.5009765625,
0.484130859375,
0.88623046875,
0.484130859375,
0.88623046875,
0.49755859375,
0.5009765625,
0.49755859375
],
[
0.0187835693359375,
0.4912109375,
0.056365966796875,
0.4912109375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.041656494140625,
0.494384765625,
0.047393798828125,
0.494384765625,
0.047393798828125,
0.5,
0.041656494140625,
0.5
],
[
0.1217041015625,
0.5048828125,
0.21728515625,
0.5048828125,
0.21728515625,
0.5166015625,
0.1217041015625,
0.5166015625
],
[
0.5,
0.5,
0.78515625,
0.5,
0.78515625,
0.51318359375,
0.5,
0.51318359375
],
[
0.5009765625,
0.52099609375,
0.8857421875,
0.52099609375,
0.8857421875,
0.53515625,
0.5009765625,
0.53515625
],
[
0.498291015625,
0.537109375,
0.88623046875,
0.537109375,
0.88623046875,
0.55029296875,
0.498291015625,
0.55029296875
],
[
0.0906982421875,
0.546875,
0.21728515625,
0.546875,
0.21728515625,
0.56103515625,
0.0906982421875,
0.56103515625
],
[
0.5009765625,
0.5517578125,
0.8857421875,
0.5517578125,
0.8857421875,
0.56494140625,
0.5009765625,
0.56494140625
],
[
0.5,
0.56494140625,
0.8857421875,
0.56640625,
0.8857421875,
0.58056640625,
0.5,
0.580078125
],
[
0.08905029296875,
0.5712890625,
0.473876953125,
0.5712890625,
0.473876953125,
0.5859375,
0.08905029296875,
0.5859375
],
[
0.08905029296875,
0.5869140625,
0.473876953125,
0.5869140625,
0.473876953125,
0.6015625,
0.08905029296875,
0.6015625
],
[
0.5009765625,
0.58203125,
0.88623046875,
0.58203125,
0.88623046875,
0.59521484375,
0.5009765625,
0.59521484375
],
[
0.08905029296875,
0.60302734375,
0.473876953125,
0.60302734375,
0.473876953125,
0.6162109375,
0.08905029296875,
0.6162109375
],
[
0.5,
0.59765625,
0.88623046875,
0.59765625,
0.88623046875,
0.611328125,
0.5,
0.611328125
],
[
0.5,
0.6123046875,
0.88623046875,
0.6123046875,
0.88623046875,
0.62548828125,
0.5,
0.62548828125
],
[
0.08905029296875,
0.61865234375,
0.473876953125,
0.61865234375,
0.473876953125,
0.6318359375,
0.08905029296875,
0.6318359375
],
[
0.5,
0.626953125,
0.88623046875,
0.62744140625,
0.88623046875,
0.6416015625,
0.5,
0.64013671875
],
[
0.0882568359375,
0.63330078125,
0.473876953125,
0.63330078125,
0.473876953125,
0.646484375,
0.0882568359375,
0.646484375
],
[
0.5,
0.6416015625,
0.8857421875,
0.6416015625,
0.8857421875,
0.65478515625,
0.5,
0.65478515625
],
[
0.08905029296875,
0.6484375,
0.473876953125,
0.6484375,
0.473876953125,
0.662109375,
0.08905029296875,
0.662109375
],
[
0.5009765625,
0.6572265625,
0.88623046875,
0.6572265625,
0.88623046875,
0.67041015625,
0.5009765625,
0.67041015625
],
[
0.0882568359375,
0.66357421875,
0.3447265625,
0.66357421875,
0.3447265625,
0.6767578125,
0.0882568359375,
0.6767578125
],
[
0.5009765625,
0.67236328125,
0.88623046875,
0.67236328125,
0.88623046875,
0.68603515625,
0.5009765625,
0.68603515625
],
[
0.08905029296875,
0.68505859375,
0.473876953125,
0.68505859375,
0.473876953125,
0.69873046875,
0.08905029296875,
0.69873046875
],
[
0.5,
0.68603515625,
0.88623046875,
0.6875,
0.88623046875,
0.7021484375,
0.5,
0.70068359375
],
[
0.08905029296875,
0.7021484375,
0.47216796875,
0.7021484375,
0.47216796875,
0.7119140625,
0.08905029296875,
0.7119140625
],
[
0.5,
0.7021484375,
0.81396484375,
0.7021484375,
0.81396484375,
0.71533203125,
0.5,
0.71533203125
],
[
0.08905029296875,
0.71533203125,
0.473876953125,
0.71533203125,
0.473876953125,
0.72900390625,
0.08905029296875,
0.72900390625
],
[
0.5009765625,
0.724609375,
0.8857421875,
0.724609375,
0.8857421875,
0.7392578125,
0.5009765625,
0.7392578125
],
[
0.0882568359375,
0.73095703125,
0.4755859375,
0.73095703125,
0.4755859375,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.5,
0.74072265625,
0.88623046875,
0.74072265625,
0.88623046875,
0.75390625,
0.5,
0.75390625
],
[
0.0882568359375,
0.7470703125,
0.4755859375,
0.7470703125,
0.4755859375,
0.76025390625,
0.0882568359375,
0.76025390625
],
[
0.0882568359375,
0.76123046875,
0.473876953125,
0.76123046875,
0.473876953125,
0.7744140625,
0.0882568359375,
0.7744140625
],
[
0.5,
0.755859375,
0.88623046875,
0.755859375,
0.88623046875,
0.76953125,
0.5,
0.76953125
],
[
0.0882568359375,
0.7763671875,
0.4755859375,
0.7763671875,
0.4755859375,
0.79052734375,
0.0882568359375,
0.79052734375
],
[
0.498291015625,
0.77099609375,
0.8857421875,
0.76953125,
0.8857421875,
0.78271484375,
0.498291015625,
0.7841796875
],
[
0.0882568359375,
0.79248046875,
0.473876953125,
0.79248046875,
0.473876953125,
0.80615234375,
0.0882568359375,
0.80615234375
],
[
0.50244140625,
0.787109375,
0.8837890625,
0.787109375,
0.8837890625,
0.79736328125,
0.50244140625,
0.79736328125
],
[
0.5,
0.8017578125,
0.88623046875,
0.8017578125,
0.88623046875,
0.81494140625,
0.5,
0.81494140625
],
[
0.1070556640625,
0.81298828125,
0.473876953125,
0.81494140625,
0.473876953125,
0.830078125,
0.1070556640625,
0.82763671875
],
[
0.5009765625,
0.81640625,
0.88623046875,
0.81640625,
0.88623046875,
0.830078125,
0.5009765625,
0.830078125
],
[
0.0882568359375,
0.8291015625,
0.476318359375,
0.830078125,
0.476318359375,
0.84130859375,
0.0882568359375,
0.84033203125
],
[
0.5,
0.83203125,
0.88623046875,
0.83203125,
0.88623046875,
0.84521484375,
0.5,
0.84521484375
],
[
0.0882568359375,
0.84130859375,
0.423095703125,
0.84130859375,
0.423095703125,
0.85498046875,
0.0882568359375,
0.85498046875
],
[
0.5009765625,
0.8466796875,
0.88623046875,
0.8466796875,
0.88623046875,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.0882568359375,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.87890625,
0.0882568359375,
0.87890625
],
[
0.5,
0.85986328125,
0.54150390625,
0.85986328125,
0.54150390625,
0.875,
0.5,
0.875
],
[
0.08905029296875,
0.8798828125,
0.450927734375,
0.8798828125,
0.450927734375,
0.89013671875,
0.08905029296875,
0.89013671875
],
[
0.111083984375,
0.892578125,
0.459228515625,
0.892578125,
0.459228515625,
0.90576171875,
0.111083984375,
0.90576171875
]
] |
[
"Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction",
"for Few-Shot Classification",
"Dong Hoon Lee 1 Sae-Young Chung",
"Abstract",
"setting. The co-existence of labeled- and unlabeled-data in",
"We propose unsupervised embedding adaptation",
"this setting motivatestheuse of transductive inference or",
"",
"semi-supervised learning. A popular transductive approach",
"forthedownstreamfew-shotclassificationtask",
"is pseudo-label-based methods that progressively update the",
"Based on findings that deep neural networks learn",
"to generalize before memorizing, we develop",
"labels or inference models by predicted test samples. For",
"Early-Stage Feature Reconstruction (ESFR)- a",
"instance, Liu et al. (2019); Kim et al. (2019) use a neighbor",
"novel adaptation scheme with feature reconstruc",
"graph for label propagation, Hou et al. (2019); Liu et al.",
"tion and dimensionality-driven early stopping that",
"(2020) use predicted labels on test samples to update the",
"finds generalizable features. Incorporating ESFR",
"class prototypes, and Antoniou & Storkey (2019); Hu et al.",
"consistently improves the performance of base-",
"(2020) use prediction to produce an intrinsic loss or syn-",
"line methods on all standard settings, including",
"thetic gradient. Another line of works utilizes regularization",
"['S]",
"the recently proposed transductive method. ESFR",
"terms on unlabeled test samples. To list a few, Dhillon et al.",
"used in conjunction with the transductive method",
"(2020); Boudiaf et al. (2020) use entropy minimization of",
"further achieves state-of-the-art performance on",
"the prediction on unlabeled data, Ziko et al. (2020) uses the",
"mini-ImageNet, tiered-ImageNet, and CUB; es-",
"Laplacian regularization term for graph clustering. These",
"pecially with 1.2%~2.0% improvements in accu-",
"methods aremostlyoriginatedfrom semi-supervisedlearn",
"racy over the previous best performing method on",
"ing studies; approaches in semi-supervised learning are still",
"18910",
"",
"I-shot setting.",
"motivating few-shot classification research.",
"On the other hand, semi-supervised learning research has",
"recently benefited from unsupervised learning.Rapidly ad-",
"1.Introduction",
"vancing self-supervised learning methods (Caron et al.",
"2020; Grill et al.,2020;Chen et al.,2020)have shown strong",
"Deep learning has achieved impressive results on visual",
"recognition tasks. However, it still has difficulty generalizing",
"performance on semi-supervised image classification tasks.",
"to novel classes with few examples; while humans can learn",
"A popular approach is to use a self-supervision loss for repre-",
"sentation learning to acquire more general features. Learned",
"to recognize from few experiences. Few-shot classification",
"representations are then used with fine-tuning (Caron et al.",
"(Miller et al., 2000; Vinyals et al., 2016; Ravi & Larochelle,",
"2020; Grill et al., 2020: Chen et al., 2020) or other semi-",
"2017) is designed to bridge the gap between the two and has",
"supervised learning methods (Zhai et al., 2019; Kim et al.,",
"recently attracted substantial attention.",
"2021) for downstream tasks. These studies achieved state-",
"Several works (Liu et al., 2019; Hou et al., 2019; Qiao",
"of-the-art performance on semi-supervised learning tasks,",
"et al..2019:Hu etal..2020:Dhillon etal..2020:Ziko",
"especially in settings with extremely few labels.",
"et al., 2020; Boudiaf et al., 2020) have shown the effec",
"The success of unsupervised learning on semi-supervised",
"tiveness of transductive methods in few-shot classification,",
"tasks suggests the potential benefit of finding shared features",
"showing a significant improvement over inductive methods.",
"While test samples are inaccessible in an inductive few-shot",
"or patterns without labels in relevant research areas. In few-",
"classification setting, one can utilize all the unlabeled test",
"shot classification, several works (Gidaris et al., 2019: Su",
"samples together to make an inference in a transductive",
"etal..2020)useadditionalself-supervisionlossduringthe",
"training of base datasets to learn more general features.",
"'school of Electrical Engineering,Korea Advanced Institute",
"However, the use of unsupervised learning on unlabeled",
"of Science and Technology (KAIST),Daejeon,Korea.Correspon-",
"data that appears in test-time is less studied. In this work,",
"dence to: Dong Hoon Lee <donghoonlee@kaist.ac.kr>.",
"we study unsupervised learning for adaptation to satisfy the",
"Proceedings of the 38th International Conference on Machine",
"thirst.",
"Learning,PMLR139,2021.Copyright2021bytheauthor(s)",
"1Code is available at https://github.com/movinghoon/ESFR"
] |
{
"arxiv_id": "2106.11486",
"language": "en",
"page_idx": 0
}
|
||
2106.11486-en-1
|
2106.11486-en-1.jpg
|
Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction
2.2. Local Intrinsic Dimensionality (LID)
Our contributions are summarized as follows:
· We find that early generalized features during unsupervised training are valuable for recognizing novel classes of few-shot classification. Based on recent studies of deep neural network's training dynamics, we explain the finding with experiments. (Section 3.1)
We briefly explain LID that our method uses as early stopping criterion. LID is a statistical version of an expansionbased Intrinsic Dimension (ID) that provides an estimated subspace dimension of local regions. Recently, substantial redundant dimensions of modern deep neural networks lead to the wide use of ID and LID to track and analyze training (Amsaleg et al., 2017; Ma et al., 2018a;b; Ansuini et al., 2019; Gong et al., 2019). The formal definition of LID is given as (Amsaleg et al., 2015; Houle, 2017a;b):
Definition 1 (Local Intrinsic Dimensionality). Given a data point $x, let$ $r> 0$ be a continuous random distance variable
$\bullet$ Based on the finding, we construct a novel embedding adaptation scheme with (1) feature reconstruction training and (2) dimensionality-driven early stopping. Our method provides task-adapted embeddings composed of desirable-shared features, which are more likely to be task-relevant and valuable for the few-shot classification. Our method is used as a plug-and-play module for few-shot methods.(Section 3.3)
from x. For the cumulative density function $F_x(r),the$ LID
$ofx\textit{ at distance r is defined as: }$
$$\mathrm{LID}(r;F_x)\stackrel{\mathrm{def}}{=}\lim\limits_{\epsilon\to0^+}\frac{\ln F_x((1+\epsilon)r)-\ln F_x(r)}{\ln(1+\epsilon)},$$
(1)
whenever the limit exists. The LID at x is then defined as
the limit of distance r \to 0^+:
$\cdot$ We test our method, ESFR, used in conjunction with baseline methods in the standard few-shot classification benchmarks. ESFR consistently improves the performance of baselines; adding ESFR to the transductive method achieves the state-of-the-art performance on mini-ImageNet, tiered-ImageNet, and CUB. Particularly in the scarce-label setting of 1-shot, our method outperforms the previous state-of-the-art with accuracies of 1.2%\sim2.0%.(Section 5.3)
(2)
$$\mathrm{LID}(x)\stackrel{\mathrm{def}}{=}\lim_{r\to0^+}\mathrm{LID}(r;F_x).$$
Since the density function of the distance variable is usually unknown, the exact value of LID is hard to acquire. We use the maximum likelihood estimation by Amsaleg et al. (2015) to calculate the LID estimates as follows:
## $2. \textbf{Preliminaries}$
2.1. Problem Setting
$$\widehat{\mathrm{LID}}(x)=-\left[\dfrac{1}{m}\sum_{i=1}^m\ln\dfrac{r_i(x)}{r_m(x)}\right]^{-1},$$
(3)
In a few-shot classification task, a small labeled support set $S$ = $\{ ( x_{\mathrm{s} }^i, y_{\mathrm{s} }^i) \} _{i= 1}^{| S| }$ and unlabeled query set $Q$ = $\{(x_q^i,)\}_{i=1}^{|Q|}$are given.$^2$Both support samples and query samples are from the same novel classes that are never seen during training. The goal of few-shot classification is to classify query (test) samples by few examples in the support set. In the usual setting, the support set has $K=1$ or 5 examples per $N=5$ novel classes, and we call this an $N$-way $K$-shot problem.
where $r_i(x)$ indicates the distance$^3$ between $x$ and its $i$-th nearest neighbor. The number of the nearest neighbor $m$ should be chosen appropriately to make estimation local but stabilized.$^{4}$
We refer to Amsaleg et al. (2015); Houle (2017a;b) for more
details about LID and its estimation methods.
We address a transductive few-shot classification task where all query samples are accessible. Since learning from few samples without prior is extremely hard, we use a pretrained embedding network $f$ that is trained on the base dataset as in Rusu et al. (2019); Wang et al. (2019); Hu et al. (2020); Ziko et al. (2020). We denote $S_f=$ $\{(f(x_s^i),y_s^i)\}_{i=1}^{|S|}$ and $Q_f=\{(f(x_q^i),\cdot)\}_{i=1}^{|Q|}$ as the support set and the query set in the embedding domain, respectively. Our interest is in task-adapted embeddings (representations) that are useful in the given few-shot task. We construct the embeddings with a module $g_\phi$ on top of $f$ by training on the union of the support set and query set.
# $3. \textbf{ Methodology}$
Figure 1 illustrates the usage and overview of our method As a plug-and-play module, our method provides taskadapted embeddings to other few-shot classification methods. Our method is mainly composed of feature reconstruction training and LID-based early stopping. We will explain feature reconstruction training (Section 3.1); LIDbased early stopping (Section 3.2); the overall method (Section 3.3), in the following.
We use Euclidean distance.
${}^{4}$We set $m=20$ throughout experiments as in Ma et al.
(2018a).
$^{2}(x,y)$ denotes an image and its label.
|
[
[
0.24267578125,
0.05682373046875,
0.73193359375,
0.05682373046875,
0.73193359375,
0.0706787109375,
0.24267578125,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.3896484375,
0.08587646484375,
0.3896484375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.5,
0.083984375,
0.7861328125,
0.085205078125,
0.7861328125,
0.0997314453125,
0.5,
0.0985107421875
],
[
0.498291015625,
0.10797119140625,
0.8857421875,
0.10919189453125,
0.8857421875,
0.12371826171875,
0.498291015625,
0.12249755859375
],
[
0.1070556640625,
0.11932373046875,
0.473876953125,
0.11932373046875,
0.473876953125,
0.1331787109375,
0.1070556640625,
0.1331787109375
],
[
0.498291015625,
0.125,
0.88623046875,
0.12371826171875,
0.88623046875,
0.1383056640625,
0.498291015625,
0.1395263671875
],
[
0.1217041015625,
0.1351318359375,
0.4755859375,
0.1351318359375,
0.4755859375,
0.1483154296875,
0.1217041015625,
0.1483154296875
],
[
0.5009765625,
0.141357421875,
0.8837890625,
0.141357421875,
0.8837890625,
0.1514892578125,
0.5009765625,
0.1514892578125
],
[
0.12091064453125,
0.1483154296875,
0.47216796875,
0.1483154296875,
0.47216796875,
0.1622314453125,
0.12091064453125,
0.1622314453125
],
[
0.5,
0.1546630859375,
0.8857421875,
0.1546630859375,
0.8857421875,
0.1685791015625,
0.5,
0.1685791015625
],
[
0.12091064453125,
0.1654052734375,
0.473876953125,
0.1654052734375,
0.473876953125,
0.1787109375,
0.12091064453125,
0.1787109375
],
[
0.5009765625,
0.1717529296875,
0.8857421875,
0.1717529296875,
0.8857421875,
0.1817626953125,
0.5009765625,
0.1817626953125
],
[
0.12091064453125,
0.179931640625,
0.455078125,
0.179931640625,
0.455078125,
0.1932373046875,
0.12091064453125,
0.1932373046875
],
[
0.5009765625,
0.1849365234375,
0.8857421875,
0.1849365234375,
0.8857421875,
0.1988525390625,
0.5009765625,
0.1988525390625
],
[
0.1070556640625,
0.205810546875,
0.473876953125,
0.205810546875,
0.473876953125,
0.2197265625,
0.1070556640625,
0.2197265625
],
[
0.5009765625,
0.2008056640625,
0.88623046875,
0.2008056640625,
0.88623046875,
0.2139892578125,
0.5009765625,
0.2139892578125
],
[
0.5009765625,
0.21533203125,
0.88623046875,
0.21533203125,
0.88623046875,
0.2291259765625,
0.5009765625,
0.2291259765625
],
[
0.12091064453125,
0.2215576171875,
0.4755859375,
0.2215576171875,
0.4755859375,
0.23486328125,
0.12091064453125,
0.23486328125
],
[
0.1192626953125,
0.236083984375,
0.473876953125,
0.23486328125,
0.473876953125,
0.25,
0.1192626953125,
0.250732421875
],
[
0.5,
0.2310791015625,
0.81787109375,
0.2310791015625,
0.81787109375,
0.2442626953125,
0.5,
0.2442626953125
],
[
0.12091064453125,
0.251953125,
0.473876953125,
0.251953125,
0.473876953125,
0.26513671875,
0.12091064453125,
0.26513671875
],
[
0.5,
0.250732421875,
0.8857421875,
0.250732421875,
0.8857421875,
0.264404296875,
0.5,
0.264404296875
],
[
0.1217041015625,
0.268310546875,
0.47216796875,
0.268310546875,
0.47216796875,
0.279052734375,
0.1217041015625,
0.279052734375
],
[
0.5,
0.266357421875,
0.8837890625,
0.266357421875,
0.8837890625,
0.27978515625,
0.5,
0.27978515625
],
[
0.1217041015625,
0.28271484375,
0.473876953125,
0.28271484375,
0.473876953125,
0.29345703125,
0.1217041015625,
0.29345703125
],
[
0.496826171875,
0.281005859375,
0.8837890625,
0.281005859375,
0.8837890625,
0.294189453125,
0.496826171875,
0.294189453125
],
[
0.12091064453125,
0.29736328125,
0.473876953125,
0.29736328125,
0.473876953125,
0.311279296875,
0.12091064453125,
0.311279296875
],
[
0.5,
0.29736328125,
0.705078125,
0.29736328125,
0.705078125,
0.311279296875,
0.5,
0.311279296875
],
[
0.12091064453125,
0.3125,
0.35693359375,
0.3125,
0.35693359375,
0.32568359375,
0.12091064453125,
0.32568359375
],
[
0.65625,
0.324462890625,
0.8388671875,
0.324462890625,
0.8388671875,
0.339111328125,
0.65625,
0.339111328125
],
[
0.5185546875,
0.33203125,
0.65869140625,
0.330078125,
0.65869140625,
0.347900390625,
0.5185546875,
0.349853515625
],
[
0.86279296875,
0.333984375,
0.8857421875,
0.333984375,
0.8857421875,
0.3466796875,
0.86279296875,
0.3466796875
],
[
0.7138671875,
0.34228515625,
0.78125,
0.34228515625,
0.78125,
0.356689453125,
0.7138671875,
0.356689453125
],
[
0.62109375,
0.34521484375,
0.654296875,
0.34521484375,
0.654296875,
0.354736328125,
0.62109375,
0.354736328125
],
[
0.1217041015625,
0.354736328125,
0.47216796875,
0.354736328125,
0.47216796875,
0.365478515625,
0.1217041015625,
0.365478515625
],
[
0.1217041015625,
0.36865234375,
0.473876953125,
0.36865234375,
0.473876953125,
0.3818359375,
0.1217041015625,
0.3818359375
],
[
0.5,
0.36865234375,
0.8857421875,
0.369384765625,
0.8857421875,
0.3837890625,
0.5,
0.38330078125
],
[
0.1217041015625,
0.385009765625,
0.47216796875,
0.385009765625,
0.47216796875,
0.395751953125,
0.1217041015625,
0.395751953125
],
[
0.5,
0.385009765625,
0.69384765625,
0.385009765625,
0.69384765625,
0.3984375,
0.5,
0.3984375
],
[
0.1217041015625,
0.40087890625,
0.471435546875,
0.40087890625,
0.471435546875,
0.410888671875,
0.1217041015625,
0.410888671875
],
[
0.1192626953125,
0.41357421875,
0.473876953125,
0.412353515625,
0.473876953125,
0.4267578125,
0.1192626953125,
0.427978515625
],
[
0.59326171875,
0.4140625,
0.78759765625,
0.415283203125,
0.78759765625,
0.434326171875,
0.59326171875,
0.43310546875
],
[
0.861328125,
0.417236328125,
0.8896484375,
0.417236328125,
0.8896484375,
0.434326171875,
0.861328125,
0.434326171875
],
[
0.1217041015625,
0.4287109375,
0.473876953125,
0.4287109375,
0.473876953125,
0.442626953125,
0.1217041015625,
0.442626953125
],
[
0.67236328125,
0.43310546875,
0.701171875,
0.43310546875,
0.701171875,
0.4375,
0.67236328125,
0.4375
],
[
0.1217041015625,
0.44580078125,
0.47216796875,
0.44580078125,
0.47216796875,
0.455810546875,
0.1217041015625,
0.455810546875
],
[
0.1192626953125,
0.458984375,
0.3447265625,
0.458984375,
0.3447265625,
0.472900390625,
0.1192626953125,
0.472900390625
],
[
0.5009765625,
0.4609375,
0.8837890625,
0.4609375,
0.8837890625,
0.474853515625,
0.5009765625,
0.474853515625
],
[
0.5009765625,
0.4765625,
0.8857421875,
0.4765625,
0.8857421875,
0.490478515625,
0.5009765625,
0.490478515625
],
[
0.5009765625,
0.4912109375,
0.88623046875,
0.4912109375,
0.88623046875,
0.5048828125,
0.5009765625,
0.5048828125
],
[
0.08905029296875,
0.49365234375,
0.223876953125,
0.49365234375,
0.223876953125,
0.50830078125,
0.08905029296875,
0.50830078125
],
[
0.1070556640625,
0.337158203125,
0.473876953125,
0.337158203125,
0.473876953125,
0.3515625,
0.1070556640625,
0.3515625
],
[
0.5,
0.505859375,
0.81884765625,
0.505859375,
0.81884765625,
0.51953125,
0.5,
0.51953125
],
[
0.08660888671875,
0.517578125,
0.2320556640625,
0.51806640625,
0.2320556640625,
0.53271484375,
0.08660888671875,
0.5322265625
],
[
0.08660888671875,
0.54150390625,
0.473876953125,
0.54248046875,
0.473876953125,
0.556640625,
0.08660888671875,
0.55615234375
],
[
0.57421875,
0.54443359375,
0.6298828125,
0.5498046875,
0.62744140625,
0.56494140625,
0.57177734375,
0.56005859375
],
[
0.7294921875,
0.54052734375,
0.77392578125,
0.54052734375,
0.77392578125,
0.5546875,
0.7294921875,
0.5546875
],
[
0.86279296875,
0.5498046875,
0.88623046875,
0.5498046875,
0.88623046875,
0.56298828125,
0.86279296875,
0.56298828125
],
[
0.0882568359375,
0.55810546875,
0.4755859375,
0.55810546875,
0.4755859375,
0.57568359375,
0.0882568359375,
0.57568359375
],
[
0.6708984375,
0.56103515625,
0.69140625,
0.56103515625,
0.69140625,
0.568359375,
0.6708984375,
0.568359375
],
[
0.7294921875,
0.556640625,
0.77685546875,
0.556640625,
0.77685546875,
0.5712890625,
0.7294921875,
0.5712890625
],
[
0.08660888671875,
0.5712890625,
0.47802734375,
0.5712890625,
0.47802734375,
0.59521484375,
0.08660888671875,
0.59521484375
],
[
0.08905029296875,
0.59326171875,
0.473876953125,
0.59326171875,
0.473876953125,
0.60693359375,
0.08905029296875,
0.60693359375
],
[
0.496826171875,
0.59033203125,
0.8837890625,
0.59033203125,
0.8837890625,
0.60498046875,
0.496826171875,
0.60498046875
],
[
0.08905029296875,
0.60791015625,
0.473876953125,
0.60791015625,
0.473876953125,
0.6220703125,
0.08905029296875,
0.6220703125
],
[
0.5009765625,
0.60693359375,
0.8857421875,
0.60693359375,
0.8857421875,
0.62060546875,
0.5009765625,
0.62060546875
],
[
0.0882568359375,
0.62353515625,
0.473876953125,
0.62353515625,
0.473876953125,
0.63720703125,
0.0882568359375,
0.63720703125
],
[
0.498291015625,
0.62060546875,
0.88623046875,
0.6220703125,
0.88623046875,
0.63623046875,
0.498291015625,
0.63525390625
],
[
0.0882568359375,
0.63818359375,
0.473876953125,
0.63818359375,
0.473876953125,
0.65283203125,
0.0882568359375,
0.65283203125
],
[
0.498291015625,
0.63720703125,
0.576171875,
0.63525390625,
0.576171875,
0.6494140625,
0.498291015625,
0.6513671875
],
[
0.0882568359375,
0.65380859375,
0.473876953125,
0.65380859375,
0.473876953125,
0.66748046875,
0.0882568359375,
0.66748046875
],
[
0.5,
0.6591796875,
0.88623046875,
0.6591796875,
0.88623046875,
0.67236328125,
0.5,
0.67236328125
],
[
0.08905029296875,
0.66845703125,
0.2034912109375,
0.66845703125,
0.2034912109375,
0.681640625,
0.08905029296875,
0.681640625
],
[
0.5,
0.673828125,
0.79736328125,
0.673828125,
0.79736328125,
0.6875,
0.5,
0.6875
],
[
0.0882568359375,
0.69140625,
0.473876953125,
0.69140625,
0.473876953125,
0.705078125,
0.0882568359375,
0.705078125
],
[
0.08905029296875,
0.70849609375,
0.47216796875,
0.70849609375,
0.47216796875,
0.71826171875,
0.08905029296875,
0.71826171875
],
[
0.5,
0.70849609375,
0.6298828125,
0.71044921875,
0.6298828125,
0.724609375,
0.5,
0.72265625
],
[
0.0882568359375,
0.7216796875,
0.4755859375,
0.7216796875,
0.4755859375,
0.73486328125,
0.0882568359375,
0.73486328125
],
[
0.08905029296875,
0.736328125,
0.473876953125,
0.736328125,
0.473876953125,
0.75,
0.08905029296875,
0.75
],
[
0.5,
0.73486328125,
0.88623046875,
0.73486328125,
0.88623046875,
0.74853515625,
0.5,
0.74853515625
],
[
0.0882568359375,
0.75048828125,
0.473876953125,
0.75048828125,
0.473876953125,
0.76513671875,
0.0882568359375,
0.76513671875
],
[
0.5,
0.751953125,
0.88623046875,
0.751953125,
0.88623046875,
0.76513671875,
0.5,
0.76513671875
],
[
0.0882568359375,
0.76513671875,
0.473876953125,
0.7666015625,
0.473876953125,
0.78076171875,
0.0882568359375,
0.77978515625
],
[
0.5009765625,
0.7666015625,
0.88623046875,
0.7666015625,
0.88623046875,
0.77978515625,
0.5009765625,
0.77978515625
],
[
0.0906982421875,
0.7822265625,
0.4755859375,
0.7822265625,
0.4755859375,
0.7998046875,
0.0906982421875,
0.7998046875
],
[
0.5009765625,
0.78271484375,
0.8857421875,
0.78271484375,
0.8857421875,
0.79345703125,
0.5009765625,
0.79345703125
],
[
0.08905029296875,
0.80029296875,
0.473876953125,
0.80029296875,
0.473876953125,
0.814453125,
0.08905029296875,
0.814453125
],
[
0.5,
0.796875,
0.88623046875,
0.796875,
0.88623046875,
0.81103515625,
0.5,
0.81103515625
],
[
0.5,
0.81103515625,
0.88623046875,
0.81005859375,
0.88623046875,
0.82470703125,
0.5,
0.82568359375
],
[
0.08905029296875,
0.8212890625,
0.473876953125,
0.8212890625,
0.473876953125,
0.83544921875,
0.08905029296875,
0.83544921875
],
[
0.5,
0.8271484375,
0.88623046875,
0.8271484375,
0.88623046875,
0.84033203125,
0.5,
0.84033203125
],
[
0.0882568359375,
0.8369140625,
0.473876953125,
0.8369140625,
0.473876953125,
0.85107421875,
0.0882568359375,
0.85107421875
],
[
0.498291015625,
0.84033203125,
0.67333984375,
0.84130859375,
0.67333984375,
0.85595703125,
0.498291015625,
0.85498046875
],
[
0.0882568359375,
0.85302734375,
0.473876953125,
0.85302734375,
0.473876953125,
0.8662109375,
0.0882568359375,
0.8662109375
],
[
0.0882568359375,
0.86865234375,
0.339111328125,
0.86865234375,
0.339111328125,
0.8818359375,
0.0882568359375,
0.8818359375
],
[
0.51953125,
0.86279296875,
0.68896484375,
0.86572265625,
0.68896484375,
0.87890625,
0.51953125,
0.876953125
],
[
0.5185546875,
0.87548828125,
0.88818359375,
0.87744140625,
0.88818359375,
0.89599609375,
0.5185546875,
0.89306640625
],
[
0.1094970703125,
0.8916015625,
0.336669921875,
0.8916015625,
0.336669921875,
0.90576171875,
0.1094970703125,
0.90576171875
],
[
0.5,
0.8916015625,
0.55224609375,
0.8916015625,
0.55224609375,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction",
"Our contributions are summarized as follows:",
"2.2. Local Intrinsic Dimensionality (LID)",
"We briefly explain LID that our method uses as early stop-",
". We find that early generalized features during unsu-",
"ping criterion. LID is a statistical version of an expansion-",
"pervised training are valuable for recognizing novel",
"basedIntrinsicDimension(ID)thatprovidesanestimated",
"classes of few-shot classification. Based on recent stud",
"subspace dimension of local regions. Recently, substantial",
"ies of deep neural network's training dynamics, we",
"redundantdimensionsofmoderndeepneuralnetworkslead",
"explain the finding with experiments. (Section 3.1)",
"to the wide use of ID and LID to track and analyze training",
". Based on the finding, we construct a novel embedding",
"(Amsaleg et al., 2017; Ma et al., 2018a;b; Ansuini et al.,",
"2019; Gong et al., 2019). The formal definition of LID is",
"adaptation scheme with (1) feature reconstruction train-",
"ing and (2) dimensionality-driven early stopping. Our",
"given as (Amsaleg et al., 2015; Houle, 2017a;b):",
"method provides task-adapted embeddings composed",
"Definition 1 (Local Intrinsic Dimensionality). Given a data",
"of desirable-shared features,which are more likelyto",
"point , let r > O be a continuous random distance variable",
"betask-relevantand valuableforthefew-shotclassifi-",
"from . For the cumulative density function F(r), the LID",
"cation. Our method is used as a plug-and-play module",
"of at distance r is defined as:",
"for few-shot methods. (Section 3.3)",
"ln F((1 +e)r)- ln F(r)",
"LID(r; Fa) f lim",
"(1)",
"In(1 +e)",
"→0+",
"baselinemethodsinthestandardfew-shotclassification",
"benchmarks.ESFR consistently improves the perfor-",
"whenever the limit exists. The LID at is then defined as",
"manceofbaselines;addingESFRtothetransductive",
"the limit of distance r → 0+:",
"methodachievesthestate-of-the-artperformanceon",
"mini-ImageNet, tiered-ImageNet, and CUB. Particu-",
"LID(a) f lim LID(r; Fr)",
"(2) ",
"larly in the scarce-label setting of 1-shot, our method",
"r0",
"outperformsthepreviousstate-of-the-artwithaccura",
"cies of 1.2%~2.0%. (Section 5.3)",
"Since the density function of the distance variable is usually",
"unknown, the exact value of LID is hard to acquire. We",
"use the maximum likelihood estimation by Amsaleg et al.",
"2. Preliminaries",
". We test our method, ESFR, used in conjunction with",
"(2015) to calculate the LID estimates as follows:",
"2.1. Problem Setting",
"In a few-shot classification task, a small labeled support",
"LID(c)",
"ri(r)",
"(3)",
"set S = [(as,y'))lSl and unlabeled query set Q =",
"m",
"Tm(a)",
"(ag,))! are given? Both support samples and query",
"samples are from the same novel classes that are never seen",
"where r;(r) indicates the distance3 between and its i-th",
"during training. The goal of few-shot classification is to",
"nearest neighbor. The number of the nearest neighbor m",
"classify query (test) samples by few examples in the support",
"should be chosen appropriately to make estimation local but",
"set. In the usual setting, the support set has K = 1 or 5 ex-",
"stabilized.4",
"amples per N = 5 novel classes, and we call this an N-way",
"We refer to Amsaleg et al. (2015); Houle (2017a:b) for more",
"K-shot problem.",
"details about LID and its estimation methods.",
"We address a transductive few-shot classification task where",
"allquerysamplesareaccessible.Sincelearningfromfew",
"3. Methodology",
"samples without prior is extremely hard, we use a pre-",
"trained embedding network f that is trained on the base",
"Figure 1 illustrates the usage and overview of our method.",
"dataset as in Rusu et al. (2019); Wang et al. (2019);",
"As a plug-and-play module, our method provides task-",
"Hu et al. (2020); Ziko et al. (2020). We denote Sf =",
"adapted embeddings to other few-shot classification meth-",
"((f(r:),y'))ll and Qf = ((f(ra), ))ll as the support",
"ods. Our method is mainly composed of feature recon-",
"set and the query set in the embedding domain, respectively",
"struction training and LID-based early stopping. We will",
"explain feature reconstruction training (Section 3.1); LID-",
"Our interest is in task-adapted embeddings (representations)",
"based early stopping (Section 3.2); the overall method (Sec-",
"that are useful in the given few-shot task. We construct the",
"tion 3.3), in the following.",
"embeddings with a module gs on top of f by training on the",
"union of the support set and query set.",
"3we use Euclidean distance.",
"4we set m = 20 throughout experiments as in Ma et al.",
"2(α, y) denotes an image and its label.",
"(2018a)."
] |
{
"arxiv_id": "2106.11486",
"language": "en",
"page_idx": 1
}
|
||
2106.11486-en-2
|
2106.11486-en-2.jpg
|
Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction
<FigureHere>
$(a)$
$(b)$
Figure 1. Overview of our method. l-(a) shows the case without embedding adaptation, and 1-(b) shows the case with embedding adaptation. Our scheme mainly consists of feature reconstruction training and $dimensionality-driven$ early stopping, and provides new embeddings of generalizable features for the downstream few-shot task.
## 3.1. Feature Reconstruction
<FigureHere>
Our main idea is based on the findings that deep neural networks learn to generalize before memorizing to abstract task-useful features. We design feature-level reconstruction training, which is unsupervised learning and builds on prior knowledge given by embeddings. As mentioned before, unsupervised learning for few-shot adaptation tends to have less attention; and we find that naively applied unsupervised learning for few-shot adaptation often fails.$^{5}$ Moreover, the behavior of unsupervised learning with a pre-trained embedding network is often ambiguous. For instance, contrastive learning heavily relies on augmentations, while augmentations affect embeddings differently depending on how the embedding network is pre-trained. Instead, we find that our feature reconstruction training can be used to adapt embeddings for few-shot classification.
Figure 2. Few-shot classification accuracy with A0, A1, and B1 during feature reconstruction training. We use ResNet-18 backbone, mini-ImageNet dataset, and nearest neighbor classifier. For $g_{\phi_1}$,we use a 4-layer neural network with 256-128-256-512 units. Other settings are described in the experiment section. (left) shows the accuracy on 1-shot and (right) shows the accuracy on 5-shot setting.
We explain our feature reconstruction training. For a fewshot classification task with embedding support set $S_f$ and query set $Q_f;$we train a reconstruction module $g_\phi$ using the following feature level reconstruction loss:
few-shot task. We train two types of reconstruction modules, $g_\phi_1$ with compressed (bottleneck) hidden layer as in conventional auto-encoders, and $g_{\phi_{2}};$ without compression. We evaluate few-shot classification accuracy with three embeddings A0, A1, B1; where A0 is the middle compressed hidden layer output of $g_{\phi _1}; \textbf{A1 and B1 are the reconstructed}$ output of $g_{\phi_1}$ and $g_{\phi_2}$, respectively.
(4)
$$\mathcal{L}(\phi)=\frac{1}{|S_f\cup Q_f|}\sum_{z\in S_f\cup Q_f}d_{\cos}(z,g_\phi(z)),$$
where $d_\cos$ denotes the cosine distance. Both $z$ and $g_\phi(z)$ are preprocessed$^{6}$ embeddings, but their expressions are omitted for notational simplicity. We note that for a newly given fewshot classification task, the weight $\phi$ of the reconstruction module is randomly re-initialized.
Figure 2 shows an interesting behavior that few-shot classification accuracies, with embeddings of reconstruction modules, initially increase then decrease. Moreover, the peak accuracy of B1 exceeds the baseline accuracy of the original embedding, on both 1- and 5-shot settings. We believe that recent studies of Deep Neural Networks (DNNs) explain such behavior. Several works (Arpit et al., 2017; Lampinen & Ganguli, 2019; Stephenson et al., 2021) observe a property that DNNs learn to generalize before memorizing. To be more specific, Arpit et al. (2017) reports that DNNs learn
We investigate the behavior of reconstruction modules during feature reconstruction training w.r.t. the downstream
$^{5}$We test self-supervised learning models of rotation (Gidaris et al., 2018) and jigsaw (Noroozi & Favaro, 2016) in Appendix E.
$^{6}$Details are described in the experiment section.
|
[
[
0.24267578125,
0.05682373046875,
0.73388671875,
0.05682373046875,
0.73388671875,
0.0706787109375,
0.24267578125,
0.0706787109375
],
[
0.16259765625,
0.085205078125,
0.318603515625,
0.085205078125,
0.318603515625,
0.0985107421875,
0.16259765625,
0.0985107421875
],
[
0.423095703125,
0.08587646484375,
0.464111328125,
0.08587646484375,
0.464111328125,
0.09783935546875,
0.423095703125,
0.09783935546875
],
[
0.615234375,
0.085205078125,
0.69677734375,
0.085205078125,
0.69677734375,
0.09661865234375,
0.615234375,
0.09661865234375
],
[
0.406005859375,
0.09783935546875,
0.483642578125,
0.09783935546875,
0.483642578125,
0.10797119140625,
0.406005859375,
0.10797119140625
],
[
0.62744140625,
0.09661865234375,
0.68896484375,
0.09661865234375,
0.68896484375,
0.10797119140625,
0.62744140625,
0.10797119140625
],
[
0.1641845703125,
0.1029052734375,
0.2320556640625,
0.1029052734375,
0.2320556640625,
0.117431640625,
0.1641845703125,
0.117431640625
],
[
0.253173828125,
0.10479736328125,
0.313720703125,
0.10479736328125,
0.313720703125,
0.11614990234375,
0.253173828125,
0.11614990234375
],
[
0.423095703125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.11932373046875,
0.423095703125,
0.11932373046875
],
[
0.63134765625,
0.10797119140625,
0.6806640625,
0.10797119140625,
0.6806640625,
0.11932373046875,
0.63134765625,
0.11932373046875
],
[
0.6787109375,
0.1363525390625,
0.70751953125,
0.1363525390625,
0.70751953125,
0.1483154296875,
0.6787109375,
0.1483154296875
],
[
0.406982421875,
0.146484375,
0.439453125,
0.146484375,
0.439453125,
0.1622314453125,
0.406982421875,
0.1622314453125
],
[
0.49267578125,
0.146484375,
0.529296875,
0.146484375,
0.529296875,
0.1622314453125,
0.49267578125,
0.1622314453125
],
[
0.62841796875,
0.143310546875,
0.6640625,
0.15087890625,
0.65869140625,
0.1666259765625,
0.62353515625,
0.1590576171875
],
[
0.75341796875,
0.146484375,
0.78515625,
0.1483154296875,
0.78369140625,
0.1610107421875,
0.75244140625,
0.1590576171875
],
[
0.69384765625,
0.15283203125,
0.71142578125,
0.15283203125,
0.71142578125,
0.1558837890625,
0.69384765625,
0.1558837890625
],
[
0.250732421875,
0.160400390625,
0.32177734375,
0.160400390625,
0.32177734375,
0.17041015625,
0.250732421875,
0.17041015625
],
[
0.57421875,
0.1483154296875,
0.58642578125,
0.1483154296875,
0.58642578125,
0.1514892578125,
0.57421875,
0.1514892578125
],
[
0.17724609375,
0.1610107421875,
0.233642578125,
0.1610107421875,
0.233642578125,
0.1685791015625,
0.17724609375,
0.1685791015625
],
[
0.44775390625,
0.1685791015625,
0.5,
0.1685791015625,
0.5,
0.18310546875,
0.44775390625,
0.18310546875
],
[
0.54150390625,
0.1685791015625,
0.58837890625,
0.1685791015625,
0.58837890625,
0.18310546875,
0.54150390625,
0.18310546875
],
[
0.25244140625,
0.1717529296875,
0.318603515625,
0.1717529296875,
0.318603515625,
0.18310546875,
0.25244140625,
0.18310546875
],
[
0.156005859375,
0.169189453125,
0.2254638671875,
0.1717529296875,
0.2254638671875,
0.18310546875,
0.156005859375,
0.18115234375
],
[
0.1683349609375,
0.18310546875,
0.21728515625,
0.18310546875,
0.21728515625,
0.1932373046875,
0.1683349609375,
0.1932373046875
],
[
0.261474609375,
0.18310546875,
0.31201171875,
0.18310546875,
0.31201171875,
0.1932373046875,
0.261474609375,
0.1932373046875
],
[
0.496826171875,
0.1881103515625,
0.53369140625,
0.199462890625,
0.525390625,
0.21533203125,
0.488525390625,
0.203857421875
],
[
0.6357421875,
0.186279296875,
0.6611328125,
0.186279296875,
0.6611328125,
0.1976318359375,
0.6357421875,
0.1976318359375
],
[
0.7509765625,
0.189453125,
0.79150390625,
0.189453125,
0.79150390625,
0.2008056640625,
0.7509765625,
0.2008056640625
],
[
0.408447265625,
0.1956787109375,
0.44287109375,
0.199462890625,
0.439453125,
0.21337890625,
0.406005859375,
0.208984375
],
[
0.61767578125,
0.1956787109375,
0.6787109375,
0.1956787109375,
0.6787109375,
0.205810546875,
0.61767578125,
0.205810546875
],
[
0.74267578125,
0.199462890625,
0.79833984375,
0.199462890625,
0.79833984375,
0.210205078125,
0.74267578125,
0.210205078125
],
[
0.62158203125,
0.205810546875,
0.6787109375,
0.205810546875,
0.6787109375,
0.216552734375,
0.62158203125,
0.216552734375
],
[
0.2548828125,
0.199462890625,
0.314453125,
0.2020263671875,
0.314453125,
0.2139892578125,
0.2548828125,
0.212158203125
],
[
0.25732421875,
0.21337890625,
0.316162109375,
0.21337890625,
0.316162109375,
0.2235107421875,
0.25732421875,
0.2235107421875
],
[
0.1666259765625,
0.238037109375,
0.21728515625,
0.238037109375,
0.21728515625,
0.25,
0.1666259765625,
0.25
],
[
0.261474609375,
0.238037109375,
0.310546875,
0.238037109375,
0.310546875,
0.25,
0.261474609375,
0.25
],
[
0.39453125,
0.23486328125,
0.494384765625,
0.23486328125,
0.494384765625,
0.2392578125,
0.39453125,
0.2392578125
],
[
0.62353515625,
0.251953125,
0.65283203125,
0.251953125,
0.65283203125,
0.2587890625,
0.62353515625,
0.2587890625
],
[
0.67626953125,
0.25244140625,
0.73779296875,
0.25244140625,
0.73779296875,
0.260009765625,
0.67626953125,
0.260009765625
],
[
0.74169921875,
0.25390625,
0.76953125,
0.255126953125,
0.76953125,
0.2587890625,
0.74169921875,
0.25830078125
],
[
0.259033203125,
0.266357421875,
0.314453125,
0.266357421875,
0.314453125,
0.27783203125,
0.259033203125,
0.27783203125
],
[
0.6611328125,
0.255126953125,
0.67724609375,
0.255126953125,
0.67724609375,
0.25830078125,
0.6611328125,
0.25830078125
],
[
0.1641845703125,
0.267578125,
0.2197265625,
0.267578125,
0.2197265625,
0.27783203125,
0.1641845703125,
0.27783203125
],
[
0.486083984375,
0.267578125,
0.619140625,
0.267578125,
0.619140625,
0.27783203125,
0.486083984375,
0.27783203125
],
[
0.181396484375,
0.2841796875,
0.2017822265625,
0.2841796875,
0.2017822265625,
0.29736328125,
0.181396484375,
0.29736328125
],
[
0.273681640625,
0.2841796875,
0.295654296875,
0.2841796875,
0.295654296875,
0.29736328125,
0.273681640625,
0.29736328125
],
[
0.08660888671875,
0.30615234375,
0.8857421875,
0.306884765625,
0.8857421875,
0.32080078125,
0.08660888671875,
0.3193359375
],
[
0.08660888671875,
0.3193359375,
0.88623046875,
0.32080078125,
0.88623046875,
0.335205078125,
0.08660888671875,
0.333984375
],
[
0.08660888671875,
0.335205078125,
0.51123046875,
0.335205078125,
0.51123046875,
0.348388671875,
0.08660888671875,
0.348388671875
],
[
0.521484375,
0.35791015625,
0.5361328125,
0.35791015625,
0.5361328125,
0.367431640625,
0.521484375,
0.367431640625
],
[
0.08905029296875,
0.365478515625,
0.283447265625,
0.365478515625,
0.283447265625,
0.37890625,
0.08905029296875,
0.37890625
],
[
0.708984375,
0.35986328125,
0.71826171875,
0.35986328125,
0.71826171875,
0.365478515625,
0.708984375,
0.365478515625
],
[
0.7060546875,
0.38330078125,
0.72216796875,
0.38330078125,
0.72216796875,
0.392578125,
0.7060546875,
0.392578125
],
[
0.7939453125,
0.386962890625,
0.87158203125,
0.38818359375,
0.87158203125,
0.396484375,
0.7939453125,
0.395751953125
],
[
0.73193359375,
0.39013671875,
0.771484375,
0.39013671875,
0.771484375,
0.393310546875,
0.73193359375,
0.393310546875
],
[
0.08905029296875,
0.391357421875,
0.47216796875,
0.391357421875,
0.47216796875,
0.401611328125,
0.08905029296875,
0.401611328125
],
[
0.521484375,
0.38818359375,
0.53759765625,
0.38818359375,
0.53759765625,
0.397705078125,
0.521484375,
0.397705078125
],
[
0.50341796875,
0.399658203125,
0.51953125,
0.399658203125,
0.51953125,
0.4140625,
0.50341796875,
0.4140625
],
[
0.615234375,
0.397705078125,
0.689453125,
0.397705078125,
0.689453125,
0.40771484375,
0.615234375,
0.40771484375
],
[
0.08660888671875,
0.40478515625,
0.4755859375,
0.40478515625,
0.4755859375,
0.41845703125,
0.08660888671875,
0.41845703125
],
[
0.50634765625,
0.415283203125,
0.5400390625,
0.415283203125,
0.5400390625,
0.43115234375,
0.50634765625,
0.43115234375
],
[
0.70751953125,
0.410400390625,
0.72119140625,
0.410400390625,
0.72119140625,
0.41845703125,
0.70751953125,
0.41845703125
],
[
0.0882568359375,
0.419189453125,
0.4755859375,
0.419189453125,
0.4755859375,
0.43310546875,
0.0882568359375,
0.43310546875
],
[
0.50732421875,
0.43115234375,
0.51708984375,
0.43115234375,
0.51708984375,
0.448974609375,
0.50732421875,
0.448974609375
],
[
0.63232421875,
0.431884765625,
0.654296875,
0.431884765625,
0.654296875,
0.44580078125,
0.63232421875,
0.44580078125
],
[
0.81787109375,
0.431884765625,
0.83837890625,
0.431884765625,
0.83837890625,
0.4443359375,
0.81787109375,
0.4443359375
],
[
0.7099609375,
0.438232421875,
0.7197265625,
0.438232421875,
0.7197265625,
0.443115234375,
0.7099609375,
0.443115234375
],
[
0.6337890625,
0.44580078125,
0.65283203125,
0.44580078125,
0.65283203125,
0.458251953125,
0.6337890625,
0.458251953125
],
[
0.08905029296875,
0.43505859375,
0.47216796875,
0.43505859375,
0.47216796875,
0.448974609375,
0.08905029296875,
0.448974609375
],
[
0.50732421875,
0.447509765625,
0.5390625,
0.449462890625,
0.53759765625,
0.4609375,
0.50634765625,
0.458984375
],
[
0.81884765625,
0.44580078125,
0.83837890625,
0.44580078125,
0.83837890625,
0.458251953125,
0.81884765625,
0.458251953125
],
[
0.6357421875,
0.460205078125,
0.65185546875,
0.460205078125,
0.65185546875,
0.4697265625,
0.6357421875,
0.4697265625
],
[
0.08905029296875,
0.45068359375,
0.473876953125,
0.45068359375,
0.473876953125,
0.464111328125,
0.08905029296875,
0.464111328125
],
[
0.08905029296875,
0.46533203125,
0.473876953125,
0.46533203125,
0.473876953125,
0.479248046875,
0.08905029296875,
0.479248046875
],
[
0.8203125,
0.460205078125,
0.83642578125,
0.460205078125,
0.83642578125,
0.4697265625,
0.8203125,
0.4697265625
],
[
0.60107421875,
0.472900390625,
0.63818359375,
0.472900390625,
0.63818359375,
0.48291015625,
0.60107421875,
0.48291015625
],
[
0.6337890625,
0.472900390625,
0.685546875,
0.472900390625,
0.685546875,
0.484130859375,
0.6337890625,
0.484130859375
],
[
0.78515625,
0.472900390625,
0.82275390625,
0.472900390625,
0.82275390625,
0.48291015625,
0.78515625,
0.48291015625
],
[
0.81640625,
0.4697265625,
0.87158203125,
0.4716796875,
0.87109375,
0.486083984375,
0.814453125,
0.484130859375
],
[
0.0882568359375,
0.479248046875,
0.473876953125,
0.479736328125,
0.473876953125,
0.494384765625,
0.0882568359375,
0.49365234375
],
[
0.521484375,
0.48095703125,
0.5390625,
0.48095703125,
0.5390625,
0.490478515625,
0.521484375,
0.490478515625
],
[
0.5947265625,
0.492431640625,
0.61376953125,
0.492431640625,
0.61376953125,
0.505859375,
0.5947265625,
0.505859375
],
[
0.71337890625,
0.4873046875,
0.71826171875,
0.490478515625,
0.71337890625,
0.496826171875,
0.70751953125,
0.49365234375
],
[
0.64892578125,
0.49365234375,
0.6689453125,
0.49365234375,
0.6689453125,
0.505859375,
0.64892578125,
0.505859375
],
[
0.779296875,
0.49365234375,
0.79833984375,
0.49365234375,
0.79833984375,
0.505859375,
0.779296875,
0.505859375
],
[
0.83642578125,
0.494384765625,
0.85107421875,
0.494384765625,
0.85107421875,
0.5048828125,
0.83642578125,
0.5048828125
],
[
0.5400390625,
0.49560546875,
0.5537109375,
0.49560546875,
0.5537109375,
0.505859375,
0.5400390625,
0.505859375
],
[
0.0882568359375,
0.494384765625,
0.473876953125,
0.49365234375,
0.473876953125,
0.50830078125,
0.0882568359375,
0.5087890625
],
[
0.7255859375,
0.49560546875,
0.73779296875,
0.49560546875,
0.73779296875,
0.5048828125,
0.7255859375,
0.5048828125
],
[
0.5947265625,
0.5048828125,
0.640625,
0.5048828125,
0.640625,
0.5166015625,
0.5947265625,
0.5166015625
],
[
0.77783203125,
0.5048828125,
0.826171875,
0.5048828125,
0.826171875,
0.5166015625,
0.77783203125,
0.5166015625
],
[
0.5,
0.52294921875,
0.88623046875,
0.52294921875,
0.88623046875,
0.5361328125,
0.5,
0.5361328125
],
[
0.08660888671875,
0.5087890625,
0.473876953125,
0.51025390625,
0.473876953125,
0.5244140625,
0.08660888671875,
0.5234375
],
[
0.0882568359375,
0.52587890625,
0.473876953125,
0.52587890625,
0.473876953125,
0.5390625,
0.0882568359375,
0.5390625
],
[
0.08905029296875,
0.54150390625,
0.473876953125,
0.54150390625,
0.473876953125,
0.5546875,
0.08905029296875,
0.5546875
],
[
0.5009765625,
0.537109375,
0.88623046875,
0.537109375,
0.88623046875,
0.55029296875,
0.5009765625,
0.55029296875
],
[
0.498291015625,
0.5498046875,
0.8857421875,
0.5498046875,
0.8857421875,
0.56298828125,
0.498291015625,
0.56298828125
],
[
0.0882568359375,
0.55615234375,
0.4755859375,
0.55615234375,
0.4755859375,
0.5693359375,
0.0882568359375,
0.5693359375
],
[
0.0882568359375,
0.5693359375,
0.47216796875,
0.5693359375,
0.47216796875,
0.58349609375,
0.0882568359375,
0.58349609375
],
[
0.498291015625,
0.564453125,
0.88623046875,
0.56298828125,
0.88623046875,
0.57763671875,
0.498291015625,
0.5791015625
],
[
0.5,
0.5791015625,
0.88623046875,
0.5791015625,
0.88623046875,
0.59228515625,
0.5,
0.59228515625
],
[
0.08905029296875,
0.5869140625,
0.47216796875,
0.5869140625,
0.47216796875,
0.59765625,
0.08905029296875,
0.59765625
],
[
0.5,
0.59228515625,
0.88623046875,
0.59228515625,
0.88623046875,
0.60595703125,
0.5,
0.60595703125
],
[
0.0882568359375,
0.60107421875,
0.303955078125,
0.60107421875,
0.303955078125,
0.6142578125,
0.0882568359375,
0.6142578125
],
[
0.5,
0.60400390625,
0.54833984375,
0.60693359375,
0.5458984375,
0.6220703125,
0.498291015625,
0.61865234375
],
[
0.0882568359375,
0.62255859375,
0.4755859375,
0.62255859375,
0.4755859375,
0.63623046875,
0.0882568359375,
0.63623046875
],
[
0.0882568359375,
0.6396484375,
0.473876953125,
0.6396484375,
0.473876953125,
0.65283203125,
0.0882568359375,
0.65283203125
],
[
0.5009765625,
0.642578125,
0.88623046875,
0.642578125,
0.88623046875,
0.65576171875,
0.5009765625,
0.65576171875
],
[
0.08660888671875,
0.65283203125,
0.473876953125,
0.65380859375,
0.473876953125,
0.66845703125,
0.08660888671875,
0.66748046875
],
[
0.5,
0.65771484375,
0.88623046875,
0.65771484375,
0.88623046875,
0.67236328125,
0.5,
0.67236328125
],
[
0.0882568359375,
0.66845703125,
0.370849609375,
0.66845703125,
0.370849609375,
0.681640625,
0.0882568359375,
0.681640625
],
[
0.5,
0.673828125,
0.88623046875,
0.673828125,
0.88623046875,
0.6875,
0.5,
0.6875
],
[
0.5,
0.68603515625,
0.88818359375,
0.6875,
0.88818359375,
0.7021484375,
0.5,
0.70068359375
],
[
0.2254638671875,
0.693359375,
0.2344970703125,
0.693359375,
0.2344970703125,
0.705078125,
0.2254638671875,
0.705078125
],
[
0.1396484375,
0.70068359375,
0.1871337890625,
0.70068359375,
0.1871337890625,
0.71533203125,
0.1396484375,
0.71533203125
],
[
0.28515625,
0.70068359375,
0.310546875,
0.70068359375,
0.310546875,
0.71337890625,
0.28515625,
0.71337890625
],
[
0.31201171875,
0.69873046875,
0.4248046875,
0.70068359375,
0.4248046875,
0.71826171875,
0.31201171875,
0.71630859375
],
[
0.453369140625,
0.7041015625,
0.473876953125,
0.7041015625,
0.473876953125,
0.7138671875,
0.453369140625,
0.7138671875
],
[
0.5,
0.7021484375,
0.8857421875,
0.70263671875,
0.8857421875,
0.71728515625,
0.5,
0.71630859375
],
[
0.193603515625,
0.70849609375,
0.268798828125,
0.71044921875,
0.268798828125,
0.72607421875,
0.193603515625,
0.7236328125
],
[
0.26708984375,
0.7197265625,
0.324462890625,
0.7197265625,
0.324462890625,
0.73095703125,
0.26708984375,
0.73095703125
],
[
0.5009765625,
0.71826171875,
0.88623046875,
0.71826171875,
0.88623046875,
0.73291015625,
0.5009765625,
0.73291015625
],
[
0.5,
0.734375,
0.73193359375,
0.734375,
0.73193359375,
0.74853515625,
0.5,
0.74853515625
],
[
0.0882568359375,
0.7421875,
0.47216796875,
0.7421875,
0.47216796875,
0.755859375,
0.0882568359375,
0.755859375
],
[
0.0882568359375,
0.7587890625,
0.473876953125,
0.75830078125,
0.473876953125,
0.771484375,
0.0882568359375,
0.77294921875
],
[
0.5009765625,
0.755859375,
0.88623046875,
0.755859375,
0.88623046875,
0.76953125,
0.5009765625,
0.76953125
],
[
0.08905029296875,
0.7734375,
0.476318359375,
0.7734375,
0.476318359375,
0.787109375,
0.08905029296875,
0.787109375
],
[
0.498291015625,
0.76953125,
0.88623046875,
0.77099609375,
0.88623046875,
0.78515625,
0.498291015625,
0.7841796875
],
[
0.0882568359375,
0.7890625,
0.473876953125,
0.7890625,
0.473876953125,
0.80322265625,
0.0882568359375,
0.80322265625
],
[
0.5009765625,
0.7861328125,
0.8857421875,
0.7861328125,
0.8857421875,
0.7998046875,
0.5009765625,
0.7998046875
],
[
0.0882568359375,
0.8046875,
0.31201171875,
0.8046875,
0.31201171875,
0.818359375,
0.0882568359375,
0.818359375
],
[
0.5009765625,
0.8017578125,
0.8857421875,
0.8017578125,
0.8857421875,
0.81494140625,
0.5009765625,
0.81494140625
],
[
0.5009765625,
0.81640625,
0.88623046875,
0.81640625,
0.88623046875,
0.830078125,
0.5009765625,
0.830078125
],
[
0.0882568359375,
0.8271484375,
0.4755859375,
0.8271484375,
0.4755859375,
0.84033203125,
0.0882568359375,
0.84033203125
],
[
0.5009765625,
0.83056640625,
0.8857421875,
0.83056640625,
0.8857421875,
0.8447265625,
0.5009765625,
0.8447265625
],
[
0.0882568359375,
0.84130859375,
0.473876953125,
0.84130859375,
0.473876953125,
0.85498046875,
0.0882568359375,
0.85498046875
],
[
0.5,
0.8466796875,
0.88623046875,
0.8466796875,
0.88623046875,
0.85986328125,
0.5,
0.85986328125
],
[
0.1070556640625,
0.86279296875,
0.4755859375,
0.86572265625,
0.4755859375,
0.8798828125,
0.1070556640625,
0.87744140625
],
[
0.5,
0.85986328125,
0.88623046875,
0.861328125,
0.88623046875,
0.87548828125,
0.5,
0.875
],
[
0.0882568359375,
0.87890625,
0.473876953125,
0.87890625,
0.473876953125,
0.892578125,
0.0882568359375,
0.892578125
],
[
0.5009765625,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.108642578125,
0.89013671875,
0.396240234375,
0.892578125,
0.396240234375,
0.9072265625,
0.108642578125,
0.90478515625
],
[
0.5,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"Unsupervised EmbeddingAdaptationvia Early-StageFeatureReconstruction",
"Few-Shot Classification Task",
"Feature",
"Dimensionality-",
"Reconstruction",
"driven Early",
"Support Set",
"QuerySet",
"Training",
"Stopping",
"LID",
"9po",
"9Φ1",
"gpk",
"gpo",
".",
"",
"",
"",
"update",
"update",
"Embedding",
"Embedding",
"Network",
"Network",
"ELFR",
"New",
"Original",
"LFR",
"Task-adapted",
"Embeddings",
"Embeddings",
"Embedding",
"Adaptation",
"Classifier",
"Classifier",
"........R..........R",
"Non-s",
"dFeaturesL",
"arnin",
"Prediction",
"",
"Prediction",
"Shared Features Learning",
"(a)",
"(b)",
"Figure 1. Overview of our method. 1-(a) shows the case without embedding adaptation, and 1-(b) shows the case with embedding",
"adaptation. Our scheme mainly consists of feature reconstruction training and dimensionality-driven early stopping, and provides new",
"embeddings of generalizable features for the downstream few-shot task",
"70",
"3.1.FeatureReconstruction",
"32",
"80",
"",
"",
"Ourmainideaisbasedonthefindingsthatdeepneural",
"65",
"0",
"O",
"networks learn to generalize before memorizing to abstract",
"060",
"78",
"task-useful features. We design feature-level reconstruction",
"",
"A0",
"A0",
"76",
"A1",
"training, which is unsupervised learning and builds on prior",
"255",
"A1",
"B1",
"knowledge given by embeddings. As mentioned before, un-",
"supervised learning for few-shot adaptation tends to have",
"B1",
"",
"Baseline",
"...",
"Baseline",
"less attention; and we find that naively applied unsupervised",
"50",
"20",
"",
"40",
"20",
"40",
"0",
"learning for few-shot adaptation often fails.5Moreover,the",
"0",
"Iteration",
"Iteration",
"Figure 2. Few-shot classification accuracy with A0, Al, and B1",
"behavior of unsupervised learning with a pre-trained embed-",
"ding network is often ambiguous. For instance, contrastive",
"learning heavily relies on augmentations, while augmenta-",
"during feature reconstruction training. We use ResNet-18 back-",
"bone, mini-ImageNet dataset, and nearest neighbor classifier. For",
"tions affect embeddings differently depending on how the",
"embedding network is pre-trained. Instead, we find that our",
"gg1, we use a 4-layer neural network with 256-128-256-512 units.",
"Other settings are described in the experiment section. (left) shows",
"featurereconstructiontrainingcanbeusedtoadaptembed",
"the accuracy on 1-shot and (right) shows the accuracy on 5-shot",
"dings for few-shot classification.",
"setting.",
"We explain our feature reconstruction training. For a few-",
"shot classification task with embedding support set Sf and",
"few-shot task. We train two types of reconstruction mod-",
"query set Qf; we train a reconstruction module g using the",
"ules, go1 with compressed (bottleneck) hidden layer as in",
"following feature level reconstruction loss:",
"conventional auto-encoders, and g2; without compression.",
"We evaluate few-shot classification accuracy with three em-",
"1",
"L(Φ) :",
"",
"dcos(z, ge(z)),",
"(4)",
"beddings A0, A1, B1; where A0 is the middle compressed",
"JS+UQFl",
"ZESUQT",
"hidden layer output of gor; A1 and B1 are the reconstructed",
"output of go1 and gb2, respectively.",
"where dcos denotes the cosine distance. Both z and ga(z) are",
"preprocessedembeddings,buttheirexpressionsareomitted",
"Figure 2 shows an interesting behavior that few-shot classi-",
"for notational simplicity. We note that for a newly given few-",
"fication accuracies, with embeddings of reconstruction mod-",
"shot classification task, the weight of the reconstruction",
"ules, initially increase then decrease. Moreover, the peak",
"module is randomly re-initialized.",
"accuracy of B1 exceeds the baseline accuracy of the original",
"embedding, on both 1- and 5-shot settings. We believe that",
"We investigate the behavior of reconstruction modules dur-",
"recent studies of Deep Neural Networks (DNNs) explain",
"ing feature reconstruction training w.r.t. the downstream",
"such behavior. Several works (Arpit et al., 2017; Lampinen",
"5we test self-supervised learning models of rotation (Gidaris",
"& Ganguli, 2019; Stephenson et al., 2021) observe a prop-",
"et al., 2018) and jigsaw (Noroozi & Favaro, 2016) in Appendix E.",
"erty that DNNs learn to generalize before memorizing. To",
"6Details are described in the experiment section.",
"be more specific, Arpit et al. (2017) reports that DNNs learn"
] |
{
"arxiv_id": "2106.11486",
"language": "en",
"page_idx": 2
}
|
||
2106.11486-en-4
|
2106.11486-en-4.jpg
|
where $\mathcal{L}_\mathrm{CE}$ is the cross-entropy loss given by an affine classifer on new embeddings, and $\lambda$ is a trade-off parameter. Additional weights $W$ and $b$ are jointly trained with $\phi.$The trade-off parameter $\lambda$ is tuned using few-shot classification tasks from validation datasets as in Ziko et al. (2020). Note that the only ESFR-Semi requires (while ESFR does not) few-shot classification task experiences to determine a cer- $\tan$ parameter.
The overall methods are described in Algorithm 1.
$$\begin{aligned}&C^{j}(z,\phi,W,b)=\mathrm{softmax}_{j}[Wg_{\phi}(z)+b]\\&\mathcal{L}_{\mathrm{CE}}(\phi,W,b)=-\frac{1}{|S_{f}|}\sum_{(z_{i},y_{i})\in S_{f}}\log C^{y_{i}}(z_{i},\phi,W,b)\\&\mathcal{L}_{\mathrm{Semi}}(\phi,W,b)=\mathcal{L}_{\mathrm{FR}}(\phi)+\lambda\mathcal{L}_{\mathrm{CE}}(\phi,W,b)\end{aligned}$$
Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction
set
<table>
<tbody>
<tr>
<td>$\xi0$ 2</td>
<td>$\mathbf{thm1ESFR}$</td>
</tr>
<tr>
<td>$\lfloor\mathbf{n}\rfloor$</td>
<td>$\mathbf{ut:embedding~support}$ ,embedding quer</td>
</tr>
<tr>
<td> 1</td>
<td>$\hat{Y}_{O}$ $put:$ =Al$\sigma /$ CESFK ESFR</td>
</tr>
</tbody>
</table>
with dropout. This seems reasonable since the classifier can further filter out non-shared features by multiple examples of novel classes in 5-shot settings; hence acquiring more shared features at the cost of learning non-shared ones can be advantageous.
## 3.3. Proposed Methods
ESFR: We describe our unsupervised adaptation scheme for few-shot classification: Early-Stage Feature Reconstruction (ESFR). Given a few-shot classification task and embedding network, we run feature reconstruction training by $\mathcal{L}_\mathrm{FR}\left(5\right)$ with initialized $\phi.$ At every training iteration, we measure the LID of $g_\phi$ with task samples by (6) and we early stop the training when the LID starts to increase. Task-adapted embeddings of $g_\phi(z)$ are used for the classification of the given few-shot task. A wide range of metric-based and finetuning few-shot methods can be used with our embeddings for such classification. To reduce the variance by random initial weights of $\phi$,we train $N_e$ reconstruction modules, separately, with different initial weights, and take the center of each sample's reconstructed embeddings.
ESFR-Semi: We further investigate the semi-supervised version of our scheme, ESFR-Semi, by adding the support classification loss to the reconstruction loss (5):
## $4. \textbf{Related Work}$
Few-Shot Classification (FSC): Few-shot learning or few-shot classification methods have broad categories: optimization-based methods (Ravi & Larochelle, 2017; Finn $\hat{\text{et al., 2017; Rusu et al., 2019), distance-based approaches}}$ (Vinyals et al., 2016; Snell et al., 2017; Oreshkin et al., 2018), fine-tunings (Wang et al., 2019; Tian et al., 2020; Dhillon et al., 2020), etc. A large portion of these methods is based on meta-learning (Thrun &Pratt, 2012).In meta-learning, training is done in a series of few-shot classification tasks (a.k.a. episodic training) to train the model in a way that reflects test-time scenarios. Several recent studies (Wang et al., 2019; Tian et al., 2020; Dhillon et al., 2020; Ziko et al., 2020; Boudiaf et al., 2020) have questioned the necessity of meta-learning on few-shot classification, reporting competitive performance on few-shot benchmarks without neither episodic training nor few-shot task experiences. These methods solve the few-shot task by fine-tuning$^{9}$ a pretrained embedding network trained on the base dataset with standard cross-entropy loss. Our method lies in this line of research that doesn't require meta-learning. It seems possible to merge our method with meta-learning; we leave it as future work. Unsupervised adaptation for FSC: We mentioned in the introduction that transductive methods, based on semi-supervised learning techniques, have been widely studied in few-shot classification. In contrast, unsupervised adaptation for few-shot classification is an open field. To the best of our knowledge, we are the first to propose deep unsupervised adaptation for few-shot classification. The closest related works are Rodríguez et al. (2020) and Lichtenstein et al. (2020). Embedding Propagation proposed by Rodríguez et al. (2020) iteratively updates embeddings by a linear combination with the nearest neighbors' embeddings and improves few-shot classification performance. Lichtenstein et al. (2020) uses the principal component analysis to acquire major components from given task embeddings; this classical approach shows impressive performance gain. However, these works are limited to an affine transformation of embeddings and do not benefit from deep learning; while our method finds non-linear patterns and correlations
$^{9}$Including the only change of class prototypes or last layer
parameters.
|
[
[
0.24267578125,
0.05682373046875,
0.73193359375,
0.05682373046875,
0.73193359375,
0.0706787109375,
0.24267578125,
0.0706787109375
],
[
0.08905029296875,
0.08709716796875,
0.2181396484375,
0.08709716796875,
0.2181396484375,
0.09783935546875,
0.08905029296875,
0.09783935546875
],
[
0.5,
0.08587646484375,
0.88623046875,
0.08587646484375,
0.88623046875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.10540771484375,
0.1029052734375,
0.473876953125,
0.1029052734375,
0.473876953125,
0.11614990234375,
0.10540771484375,
0.11614990234375
],
[
0.5,
0.10162353515625,
0.88623046875,
0.10162353515625,
0.88623046875,
0.11553955078125,
0.5,
0.11553955078125
],
[
0.10540771484375,
0.11932373046875,
0.418212890625,
0.11932373046875,
0.418212890625,
0.1337890625,
0.10540771484375,
0.1337890625
],
[
0.5009765625,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.5009765625,
0.1300048828125
],
[
0.10455322265625,
0.1337890625,
0.226318359375,
0.1319580078125,
0.226318359375,
0.146484375,
0.10455322265625,
0.1483154296875
],
[
0.5009765625,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.10455322265625,
0.149658203125,
0.2296142578125,
0.149658203125,
0.2296142578125,
0.1634521484375,
0.10455322265625,
0.1634521484375
],
[
0.5,
0.146484375,
0.88623046875,
0.146484375,
0.88623046875,
0.160400390625,
0.5,
0.160400390625
],
[
0.11767578125,
0.1610107421875,
0.253173828125,
0.1634521484375,
0.253173828125,
0.18115234375,
0.11767578125,
0.1787109375
],
[
0.5009765625,
0.1622314453125,
0.88623046875,
0.1622314453125,
0.88623046875,
0.175537109375,
0.5009765625,
0.175537109375
],
[
0.12091064453125,
0.1787109375,
0.25732421875,
0.179931640625,
0.25732421875,
0.1944580078125,
0.12091064453125,
0.1932373046875
],
[
0.5009765625,
0.177978515625,
0.88623046875,
0.177978515625,
0.88623046875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.12091064453125,
0.1944580078125,
0.358642578125,
0.1932373046875,
0.358642578125,
0.2083740234375,
0.12091064453125,
0.208984375
],
[
0.5009765625,
0.1932373046875,
0.5986328125,
0.1932373046875,
0.5986328125,
0.2052001953125,
0.5009765625,
0.2052001953125
],
[
0.1356201171875,
0.210205078125,
0.44287109375,
0.205810546875,
0.44287109375,
0.2235107421875,
0.1356201171875,
0.2279052734375
],
[
0.5,
0.2139892578125,
0.830078125,
0.2139892578125,
0.830078125,
0.2279052734375,
0.5,
0.2279052734375
],
[
0.134033203125,
0.2259521484375,
0.250732421875,
0.2291259765625,
0.250732421875,
0.246826171875,
0.134033203125,
0.24365234375
],
[
0.136474609375,
0.24560546875,
0.279296875,
0.24560546875,
0.279296875,
0.260009765625,
0.136474609375,
0.260009765625
],
[
0.5009765625,
0.248779296875,
0.6396484375,
0.248779296875,
0.6396484375,
0.261962890625,
0.5009765625,
0.261962890625
],
[
0.1519775390625,
0.25830078125,
0.2296142578125,
0.261474609375,
0.2279052734375,
0.279052734375,
0.1502685546875,
0.27587890625
],
[
0.1519775390625,
0.27587890625,
0.193603515625,
0.27587890625,
0.193603515625,
0.290283203125,
0.1519775390625,
0.290283203125
],
[
0.5009765625,
0.274658203125,
0.88623046875,
0.274658203125,
0.88623046875,
0.28857421875,
0.5009765625,
0.28857421875
],
[
0.1356201171875,
0.290283203125,
0.1829833984375,
0.290283203125,
0.1829833984375,
0.304931640625,
0.1356201171875,
0.304931640625
],
[
0.5009765625,
0.290283203125,
0.88818359375,
0.290283203125,
0.88818359375,
0.3037109375,
0.5009765625,
0.3037109375
],
[
0.1356201171875,
0.30615234375,
0.2320556640625,
0.30615234375,
0.2320556640625,
0.32080078125,
0.1356201171875,
0.32080078125
],
[
0.5,
0.304931640625,
0.88623046875,
0.304931640625,
0.88623046875,
0.3193359375,
0.5,
0.3193359375
],
[
0.12091064453125,
0.322509765625,
0.1732177734375,
0.322509765625,
0.1732177734375,
0.333984375,
0.12091064453125,
0.333984375
],
[
0.5009765625,
0.32080078125,
0.8857421875,
0.32080078125,
0.8857421875,
0.333984375,
0.5009765625,
0.333984375
],
[
0.10540771484375,
0.337158203125,
0.1568603515625,
0.337158203125,
0.1568603515625,
0.348388671875,
0.10540771484375,
0.348388671875
],
[
0.5009765625,
0.335205078125,
0.88623046875,
0.335205078125,
0.88623046875,
0.348388671875,
0.5009765625,
0.348388671875
],
[
0.1004638671875,
0.343505859375,
0.455810546875,
0.343505859375,
0.455810546875,
0.371826171875,
0.1004638671875,
0.371826171875
],
[
0.5009765625,
0.35107421875,
0.88623046875,
0.35107421875,
0.88623046875,
0.3642578125,
0.5009765625,
0.3642578125
],
[
0.10296630859375,
0.365478515625,
0.40283203125,
0.367431640625,
0.40283203125,
0.389404296875,
0.10296630859375,
0.386962890625
],
[
0.5,
0.3642578125,
0.8857421875,
0.365478515625,
0.8857421875,
0.380126953125,
0.5,
0.37890625
],
[
0.5009765625,
0.380615234375,
0.8857421875,
0.380615234375,
0.8857421875,
0.39453125,
0.5009765625,
0.39453125
],
[
0.10296630859375,
0.386962890625,
0.334228515625,
0.386474609375,
0.334228515625,
0.404052734375,
0.10296630859375,
0.40478515625
],
[
0.5009765625,
0.395751953125,
0.88623046875,
0.395751953125,
0.88623046875,
0.4091796875,
0.5009765625,
0.4091796875
],
[
0.5009765625,
0.410888671875,
0.8857421875,
0.410888671875,
0.8857421875,
0.4248046875,
0.5009765625,
0.4248046875
],
[
0.5009765625,
0.425537109375,
0.88623046875,
0.425537109375,
0.88623046875,
0.439453125,
0.5009765625,
0.439453125
],
[
0.0882568359375,
0.43505859375,
0.473876953125,
0.43505859375,
0.473876953125,
0.448974609375,
0.0882568359375,
0.448974609375
],
[
0.5,
0.43994140625,
0.88623046875,
0.439453125,
0.88623046875,
0.453857421875,
0.5,
0.455078125
],
[
0.08905029296875,
0.45068359375,
0.473876953125,
0.45068359375,
0.473876953125,
0.464111328125,
0.08905029296875,
0.464111328125
],
[
0.5,
0.455810546875,
0.88623046875,
0.455810546875,
0.88623046875,
0.4697265625,
0.5,
0.4697265625
],
[
0.0882568359375,
0.466552734375,
0.473876953125,
0.466552734375,
0.473876953125,
0.479736328125,
0.0882568359375,
0.479736328125
],
[
0.5,
0.4716796875,
0.88623046875,
0.4716796875,
0.88623046875,
0.48486328125,
0.5,
0.48486328125
],
[
0.0882568359375,
0.48095703125,
0.473876953125,
0.48095703125,
0.473876953125,
0.494384765625,
0.0882568359375,
0.494384765625
],
[
0.5009765625,
0.4873046875,
0.88623046875,
0.4873046875,
0.88623046875,
0.50048828125,
0.5009765625,
0.50048828125
],
[
0.0882568359375,
0.496826171875,
0.205078125,
0.496826171875,
0.205078125,
0.51025390625,
0.0882568359375,
0.51025390625
],
[
0.5009765625,
0.501953125,
0.8857421875,
0.501953125,
0.8857421875,
0.51513671875,
0.5009765625,
0.51513671875
],
[
0.50244140625,
0.51806640625,
0.8857421875,
0.51806640625,
0.8857421875,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.0882568359375,
0.5263671875,
0.25,
0.5263671875,
0.25,
0.54150390625,
0.0882568359375,
0.54150390625
],
[
0.5,
0.5322265625,
0.8857421875,
0.5322265625,
0.8857421875,
0.54541015625,
0.5,
0.54541015625
],
[
0.08905029296875,
0.55029296875,
0.4755859375,
0.55029296875,
0.4755859375,
0.564453125,
0.08905029296875,
0.564453125
],
[
0.498291015625,
0.546875,
0.8857421875,
0.546875,
0.8857421875,
0.56005859375,
0.498291015625,
0.56005859375
],
[
0.08905029296875,
0.56640625,
0.4755859375,
0.56640625,
0.4755859375,
0.580078125,
0.08905029296875,
0.580078125
],
[
0.5009765625,
0.5625,
0.88623046875,
0.5625,
0.88623046875,
0.57568359375,
0.5009765625,
0.57568359375
],
[
0.0882568359375,
0.58056640625,
0.473876953125,
0.58056640625,
0.473876953125,
0.5947265625,
0.0882568359375,
0.5947265625
],
[
0.5,
0.5771484375,
0.88623046875,
0.5771484375,
0.88623046875,
0.59033203125,
0.5,
0.59033203125
],
[
0.0882568359375,
0.5966796875,
0.473876953125,
0.5966796875,
0.473876953125,
0.60986328125,
0.0882568359375,
0.60986328125
],
[
0.498291015625,
0.59130859375,
0.88623046875,
0.59130859375,
0.88623046875,
0.60498046875,
0.498291015625,
0.60498046875
],
[
0.0882568359375,
0.611328125,
0.473876953125,
0.611328125,
0.473876953125,
0.625,
0.0882568359375,
0.625
],
[
0.5,
0.60693359375,
0.88623046875,
0.60693359375,
0.88623046875,
0.62060546875,
0.5,
0.62060546875
],
[
0.0882568359375,
0.626953125,
0.473876953125,
0.626953125,
0.473876953125,
0.64013671875,
0.0882568359375,
0.64013671875
],
[
0.5,
0.62255859375,
0.8857421875,
0.62255859375,
0.8857421875,
0.63623046875,
0.5,
0.63623046875
],
[
0.0882568359375,
0.6416015625,
0.473876953125,
0.6416015625,
0.473876953125,
0.65478515625,
0.0882568359375,
0.65478515625
],
[
0.5,
0.63720703125,
0.88623046875,
0.63720703125,
0.88623046875,
0.65087890625,
0.5,
0.65087890625
],
[
0.0882568359375,
0.6572265625,
0.473876953125,
0.6572265625,
0.473876953125,
0.671875,
0.0882568359375,
0.671875
],
[
0.5,
0.65283203125,
0.8857421875,
0.65283203125,
0.8857421875,
0.66650390625,
0.5,
0.66650390625
],
[
0.0882568359375,
0.67236328125,
0.4755859375,
0.67236328125,
0.4755859375,
0.68603515625,
0.0882568359375,
0.68603515625
],
[
0.498291015625,
0.66748046875,
0.8857421875,
0.66845703125,
0.8857421875,
0.681640625,
0.498291015625,
0.68115234375
],
[
0.0882568359375,
0.68603515625,
0.4755859375,
0.6875,
0.4755859375,
0.7021484375,
0.0882568359375,
0.70068359375
],
[
0.5009765625,
0.68310546875,
0.8857421875,
0.68310546875,
0.8857421875,
0.6962890625,
0.5009765625,
0.6962890625
],
[
0.5,
0.69775390625,
0.88623046875,
0.69775390625,
0.88623046875,
0.7109375,
0.5,
0.7109375
],
[
0.0906982421875,
0.7041015625,
0.47216796875,
0.7041015625,
0.47216796875,
0.7138671875,
0.0906982421875,
0.7138671875
],
[
0.08905029296875,
0.71728515625,
0.4755859375,
0.71728515625,
0.4755859375,
0.73095703125,
0.08905029296875,
0.73095703125
],
[
0.498291015625,
0.7119140625,
0.8837890625,
0.7119140625,
0.8837890625,
0.72607421875,
0.498291015625,
0.72607421875
],
[
0.0882568359375,
0.73193359375,
0.473876953125,
0.73193359375,
0.473876953125,
0.74560546875,
0.0882568359375,
0.74560546875
],
[
0.5009765625,
0.72802734375,
0.88623046875,
0.72802734375,
0.88623046875,
0.7412109375,
0.5009765625,
0.7412109375
],
[
0.0882568359375,
0.74755859375,
0.37744140625,
0.74755859375,
0.37744140625,
0.76123046875,
0.0882568359375,
0.76123046875
],
[
0.5009765625,
0.74365234375,
0.8857421875,
0.74365234375,
0.8857421875,
0.7568359375,
0.5009765625,
0.7568359375
],
[
0.5,
0.75830078125,
0.88623046875,
0.75830078125,
0.88623046875,
0.771484375,
0.5,
0.771484375
],
[
0.0882568359375,
0.76953125,
0.47216796875,
0.76953125,
0.47216796875,
0.78271484375,
0.0882568359375,
0.78271484375
],
[
0.5009765625,
0.7734375,
0.88623046875,
0.7734375,
0.88623046875,
0.787109375,
0.5009765625,
0.787109375
],
[
0.08905029296875,
0.78515625,
0.473876953125,
0.78515625,
0.473876953125,
0.798828125,
0.08905029296875,
0.798828125
],
[
0.5,
0.78857421875,
0.88623046875,
0.78857421875,
0.88623046875,
0.80322265625,
0.5,
0.80322265625
],
[
0.08905029296875,
0.8017578125,
0.39794921875,
0.8017578125,
0.39794921875,
0.8125,
0.08905029296875,
0.8125
],
[
0.5009765625,
0.8037109375,
0.88623046875,
0.8037109375,
0.88623046875,
0.8173828125,
0.5009765625,
0.8173828125
],
[
0.5009765625,
0.818359375,
0.88623046875,
0.818359375,
0.88623046875,
0.83203125,
0.5009765625,
0.83203125
],
[
0.108642578125,
0.82568359375,
0.374267578125,
0.82568359375,
0.374267578125,
0.84033203125,
0.108642578125,
0.84033203125
],
[
0.5,
0.833984375,
0.88623046875,
0.833984375,
0.88623046875,
0.84765625,
0.5,
0.84765625
],
[
0.240234375,
0.8466796875,
0.250732421875,
0.8466796875,
0.250732421875,
0.85595703125,
0.240234375,
0.85595703125
],
[
0.303955078125,
0.8486328125,
0.455078125,
0.849609375,
0.455078125,
0.8671875,
0.303955078125,
0.8662109375
],
[
0.5,
0.849609375,
0.88623046875,
0.849609375,
0.88623046875,
0.86279296875,
0.5,
0.86279296875
],
[
0.279296875,
0.85107421875,
0.313720703125,
0.85107421875,
0.313720703125,
0.86279296875,
0.279296875,
0.86279296875
],
[
0.108642578125,
0.8515625,
0.214111328125,
0.8515625,
0.214111328125,
0.8662109375,
0.108642578125,
0.8662109375
],
[
0.226318359375,
0.85595703125,
0.326904296875,
0.86865234375,
0.32275390625,
0.89013671875,
0.22216796875,
0.87744140625
],
[
0.285888671875,
0.861328125,
0.298095703125,
0.861328125,
0.298095703125,
0.86572265625,
0.285888671875,
0.86572265625
],
[
0.51953125,
0.8720703125,
0.8857421875,
0.87255859375,
0.8857421875,
0.88720703125,
0.51953125,
0.88623046875
],
[
0.108642578125,
0.88720703125,
0.38818359375,
0.88720703125,
0.38818359375,
0.9013671875,
0.108642578125,
0.9013671875
],
[
0.450927734375,
0.88720703125,
0.4755859375,
0.88720703125,
0.4755859375,
0.90087890625,
0.450927734375,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.57177734375,
0.88720703125,
0.57177734375,
0.8984375,
0.5009765625,
0.8984375
]
] |
[
"Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction",
"Algorithm1ESFR",
"where Cce is the cross-entropy loss given by an affine clas-",
"Input: embedding support set Sf, embedding query set",
"sifier on new embeddings, and is a trade-off parameter.",
"Qf, and few-shot classifier Alg : Sf, Qf → YQ",
"Additional weights W and b are jointly trained with @. The",
"Initialize: 6i-1:N",
"trade-off parameter is tuned using few-shot classification",
"for i = 1 to Ndo",
"tasks from validation datasets as in Ziko et al. (2020). Note",
"prev lid = LID()",
"that the only ESFR-Semi requires (while ESFR does not)",
"Initialize: optimizer",
"few-shot classification task experiences to determine a cer-",
"for j = O to MAX ITERATION do",
"tain parameter.",
"Φ$+1 ←Φ - V$ C(pg) from equation 5 or 7",
"The overall methods are described in Algorithm 1.",
"lid = LID(Φ+1)",
"if lid > prev lid then",
"4. Related Work",
"=+1",
"break",
"Few-Shot Classification (FSC): Few-shot learning or",
"end if",
"few-shot classification methods have broad categories:",
"prev lid = lid",
"optimization-based methods (Ravi & Larochelle, 2017; Finn",
"end for",
"et al., 2017; Rusu et al., 2019), distance-based approaches",
"end for",
"(Vinyals et al., 2016; Snell et al., 2017; Oreshkin et al.,",
"SESFR = (2,u)]2 = 1 9 (2),(2, ) St)",
"2018), fine-tunings (Wang et al., 2019; Tian et al., 2020;",
"QESFR = [2]2 =EN1901(2),zE Qn)",
"Dhillon et al., 2020), etc. A large portion of these meth",
"ods is based on meta-learning (Thrun & Pratt, 2012). In",
"Output: YQ = Alg(SESFR, QESFR)",
"meta-learning, training is done in a series of few-shot classi-",
"fication tasks (a.k.a. episodic training) to train the model in",
"a way that reflects test-time scenarios. Several recent studies",
"with dropout. This seems reasonable since the classifier can",
"(Wang et al., 2019; Tian et al., 2020; Dhillon et al., 2020;",
"further filter out non-shared features by multiple examples",
"Ziko et al., 2020; Boudiaf et al., 2020) have questioned the",
"of novel classes in 5-shot settings; hence acquiring more",
"necessity of meta-learning on few-shot classification, report-",
"shared features at the cost of learning non-shared ones can",
"ing competitive performance on few-shot benchmarks with-",
"be advantageous.",
"out neither episodic training nor few-shot task experiences",
"Thesemethodssolvethefew-shottaskbyfine-tuningapre-",
"3.3. Proposed Methods",
"trained embedding network trained on the base dataset with",
"ESFR:We describe our unsupervised adaptation scheme for",
"standard cross-entropy loss. Our method lies in this line of",
"few-shot classification: Early-Stage Feature Reconstruction",
"research that doesn't require meta-learning. It seems possi-",
"(ESFR).Given a few-shot classification task and embedding",
"ble to merge our method with meta-learning: we leave it as",
"network, we run feature reconstruction training by LFr (5)",
"future work. Unsupervised adaptation for FSC: We men-",
"with initialized . At every training iteration, we measure",
"tioned in the introduction that transductive methods, based",
"the LID of go with task samples by (6) and we early stop",
"on semi-supervised learning techniques, have been widely",
"the training when the LID starts to increase. Task-adapted",
"studied in few-shot classification. In contrast, unsupervised",
"embeddings of g(z) are used for the classification of the",
"adaptation for few-shot classification is an open field. To",
"given few-shot task. A wide range of metric-based and fine-",
"thebest of ourknowledge,we are the firstto propose deep",
"tuning few-shot methods can be used with our embeddings",
"unsupervised adaptation for few-shot classification. The",
"closest related works are Rodriguez et al. (2020) and Licht-",
"forsuchclassification.Toreducethevariancebyrandom",
"initial weights of o, we train Ne reconstruction modules,",
"enstein et al.(2020).Embedding Propagation proposed by",
"separately, with different initial weights, and take the center",
"Rodriguez et al. (2020) iteratively updates embeddings by a",
"of each sample's reconstructed embeddings.",
"linear combination with the nearest neighbors' embeddings",
"and improves few-shot classification performance. Licht-",
"ESFR-Semi: We further investigate the semi-supervised",
"enstein et al. (2020) uses the principal component analysis",
"version of our scheme, ESFR-Semi, by adding the support",
"to acquire major components from given task embeddings;",
"classificationlosstothereconstructionloss (5)",
"this classical approach shows impressive performance gain.",
"However, these works are limited to an affine transforma-",
"Ci(z, , W,b) = softmax,[Wg(z) + b]",
"tion of embeddings and do not benefit from deep learning;",
"1",
"log Cu (zi, p, W,b)",
"while our method finds non-linear patterns and correlations",
"7",
"Cce(Φ, W,b) =",
"Sil (z.y)es)",
"",
"'Including the only change of class prototypes or last layer",
"Csemi(Φ, W, b) = LFR() + 入LcE(o, W,b)",
"(7)",
"parameters."
] |
{
"arxiv_id": "2106.11486",
"language": "en",
"page_idx": 4
}
|
||
1712.09184-en-0
|
1712.09184-en-0.jpg
|
$\textbf{Detect- and- Track: Efficient Pose Estimation in Videos}$
$\begin{array}{c}\text{Rohit Girdhar}^{1^*}&\text{Georgia Gkioxari}^2&\text{Lorenzo Torresani}^{2,3^*}&\text{Manohar Paluri}^2&\text{Du Tran}^2\\\text{The Robotics Institute, Carnegie Mellon University }^2\text{Facebook }^3\text{Dartmouth College}\end{array}$
https://rohitgirdhar.github.io/DetectAndTrack
$n$
$હ્રિં$
## $\mathbf{Abstract}$
This paper addresses the problem of estimating and track-
$\sum _{i= 1}^{\infty }$ ing human body keypoints $in$ complex
detection [17] and video understanding [5]. Our method operates in two-stages: keypoint estimation in frames or short $clips, followed\textit{by lightweight tracking to generate keypoint}$
$\sum_{\begin{array}{cc}&clips,followed\:by\:lightwe\iota ght\:trackung\:to\:generale\:keypount\\&predictions\:linked\:over\:the\:entire\:video.\: For\:frame-level\\&pose\:estimation\:we\:experiment\:with\:Mask\:R-CNN,\:as\:well\:as\end{array}}$ $\begin{array}{ccc}&ourown\:proposed\:3D\:extension\:of\:this\:model,\:which\:lever-\\C&ages\:temporal\:information\:over\:small\:clips\:to\:generate\:more\\&robust\:frame\:predictions.\:We\:conduct\:extensive\:ablative\:ex-\end{array}$
periments on the newly released multi-person video pose estimation benchmark, PoseTrack, to validate various design choices of our model. Our approach achieves an accuracy of 55.2% on the validation and 51.8% on the test set using
The Multi- Object Tracking Accuracy ( MOTA) metric, and Sachieves state of the art performance on the ICCV 2017
PoseTrack keypoint tracking challenge [1].
<FigureHere>
Figure 1. We propose a two-stage approach to keypoint estimation and tracking in videos. For the first stage, we propose a novel video pose estimation formulation, 3D Mask R-CNN, that takes a short video clip as input and produces a tubelet per person and keypoints within those. In the second stage, we perform lightweight optimization to link the detections over time.
## $1. \textbf{Introduction}$
In recent years, visual understanding, such as object and scene recognition [17,40,44,55], has witnessed a significant bloom thanks to deep visual representations [18,31,47,50]. Modeling and understanding human behaviour in images has been in the epicenter of a variety of visual tasks due to its importance for numerous downstream practical applications. In particular, person detection and pose estimation from a single image have emerged as prominent and challenging visual recognition problems[36]. While single-image understanding has advanced steadily through the introduction of tasks of increasing complexity, video understanding has made slower progress compared to the image domain. Here, the preeminent task involves labeling whole videos with a single $\hat{\text{activity type[5,7,10,14,29,30,32,46,49,51,52].While still}}$ relevant and challenging, this task shifts the focus away from
$^{*}$Work done as a part of RG's internship at Facebook
one of the more interesting aspects of video understanding, namely modeling the changes in appearance and semantics of scenes, objects and humans over time[6,13,15,37].
In this work, we focus on the problem of human pose tracking in complex videos, which entails tracking and estimating the pose of each human instance over time. The challenges here are plenty, including pose changes, occlusions and the presence of multiple overlapping instances. The ideal tracker needs to accurately predict the pose of all human instances at each time step by reasoning about the appearance and pose transitions over time. Hence, the effort to materialize a pose tracker should closely follow the state of the art in pose prediction but also enhance it with the tools necessary to successfully integrate time information at an instance-specific level.
Most recent video pose estimation methods use handdesigned graphical models or integer program optimizations on top of frame-based keypoint predictions to compute the final predictions over time[21,26,48]. While such approaches have shown good performance, they require hand-coding of optimization constraints and may not be scalable beyond short video clips due to their computational complexity. Most importantly, the tracking optimization is only responsible for linking frame-level predictions, and the system has no mechanism to improve the estimation of keypoints by
1
|
[
[
0.21728515625,
0.1337890625,
0.7548828125,
0.1337890625,
0.7548828125,
0.1483154296875,
0.21728515625,
0.1483154296875
],
[
0.2607421875,
0.179931640625,
0.416748046875,
0.18115234375,
0.416748046875,
0.1988525390625,
0.2607421875,
0.1976318359375
],
[
0.408447265625,
0.18115234375,
0.615234375,
0.1767578125,
0.615234375,
0.1976318359375,
0.408447265625,
0.2020263671875
],
[
0.60302734375,
0.179931640625,
0.75732421875,
0.18115234375,
0.75732421875,
0.1988525390625,
0.60302734375,
0.1976318359375
],
[
0.7548828125,
0.18310546875,
0.84619140625,
0.18310546875,
0.84619140625,
0.1976318359375,
0.7548828125,
0.1976318359375
],
[
0.134033203125,
0.1976318359375,
0.5537109375,
0.1988525390625,
0.5537109375,
0.216552734375,
0.134033203125,
0.21533203125
],
[
0.12335205078125,
0.18310546875,
0.2442626953125,
0.18115234375,
0.2442626953125,
0.1956787109375,
0.12335205078125,
0.1976318359375
],
[
0.55126953125,
0.2008056640625,
0.66259765625,
0.2008056640625,
0.66259765625,
0.2139892578125,
0.55126953125,
0.2139892578125
],
[
0.65283203125,
0.1988525390625,
0.8349609375,
0.2008056640625,
0.8349609375,
0.21533203125,
0.65283203125,
0.21337890625
],
[
0.28515625,
0.2197265625,
0.6845703125,
0.2197265625,
0.6845703125,
0.23291015625,
0.28515625,
0.23291015625
],
[
0.0269622802734375,
0.261962890625,
0.053924560546875,
0.261962890625,
0.053924560546875,
0.27978515625,
0.0269622802734375,
0.27978515625
],
[
0.02532958984375,
0.272705078125,
0.056365966796875,
0.272705078125,
0.056365966796875,
0.32568359375,
0.02532958984375,
0.32568359375
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.736328125,
0.2822265625,
0.85107421875,
0.281005859375,
0.85107421875,
0.29541015625,
0.736328125,
0.296630859375
],
[
0.74267578125,
0.27783203125,
0.771484375,
0.27783203125,
0.771484375,
0.281005859375,
0.74267578125,
0.281005859375
],
[
0.1004638671875,
0.300537109375,
0.465576171875,
0.300537109375,
0.465576171875,
0.31494140625,
0.1004638671875,
0.31494140625
],
[
0.73779296875,
0.296630859375,
0.87158203125,
0.296630859375,
0.87158203125,
0.31005859375,
0.73779296875,
0.31005859375
],
[
0.0318603515625,
0.314453125,
0.062103271484375,
0.317626953125,
0.052276611328125,
0.375732421875,
0.0228729248046875,
0.37255859375
],
[
0.08087158203125,
0.318115234375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.73779296875,
0.3125,
0.8349609375,
0.3125,
0.8349609375,
0.323974609375,
0.73779296875,
0.323974609375
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.74169921875,
0.333251953125,
0.779296875,
0.333251953125,
0.779296875,
0.34033203125,
0.74169921875,
0.34033203125
],
[
0.53369140625,
0.3447265625,
0.611328125,
0.3447265625,
0.611328125,
0.359130859375,
0.53369140625,
0.359130859375
],
[
0.7158203125,
0.347900390625,
0.72802734375,
0.347900390625,
0.72802734375,
0.3515625,
0.7158203125,
0.3515625
],
[
0.740234375,
0.34228515625,
0.86279296875,
0.34228515625,
0.86279296875,
0.35595703125,
0.740234375,
0.35595703125
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.67236328125,
0.348388671875,
0.68115234375,
0.348388671875,
0.68115234375,
0.3515625,
0.67236328125,
0.3515625
],
[
0.73876953125,
0.35595703125,
0.861328125,
0.35595703125,
0.861328125,
0.369384765625,
0.73876953125,
0.369384765625
],
[
0.080078125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.02777099609375,
0.37060546875,
0.0555419921875,
0.37060546875,
0.0555419921875,
0.38818359375,
0.02777099609375,
0.38818359375
],
[
0.73876953125,
0.37060546875,
0.7998046875,
0.371826171875,
0.7998046875,
0.38330078125,
0.73876953125,
0.3818359375
],
[
0.08087158203125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.4462890625,
0.02532958984375,
0.4462890625
],
[
0.08087158203125,
0.393310546875,
0.4697265625,
0.393310546875,
0.4697265625,
0.4072265625,
0.08087158203125,
0.4072265625
],
[
0.50244140625,
0.401611328125,
0.89208984375,
0.401611328125,
0.89208984375,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50341796875,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.427978515625,
0.50341796875,
0.427978515625
],
[
0.080078125,
0.423583984375,
0.4697265625,
0.423583984375,
0.4697265625,
0.4375,
0.080078125,
0.4375
],
[
0.50341796875,
0.43115234375,
0.890625,
0.43115234375,
0.890625,
0.44140625,
0.50341796875,
0.44140625
],
[
0.03350830078125,
0.442626953125,
0.0555419921875,
0.442626953125,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.438232421875,
0.471435546875,
0.438232421875,
0.471435546875,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.443115234375,
0.89208984375,
0.443115234375,
0.89208984375,
0.45703125,
0.50244140625,
0.45703125
],
[
0.080078125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.50341796875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.4716796875,
0.81396484375,
0.4716796875,
0.81396484375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.02532958984375,
0.48486328125,
0.05963134765625,
0.486083984375,
0.056365966796875,
0.5517578125,
0.0228729248046875,
0.55029296875
],
[
0.07843017578125,
0.484130859375,
0.468017578125,
0.48291015625,
0.468017578125,
0.49755859375,
0.07843017578125,
0.498779296875
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.080078125,
0.5146484375,
0.468017578125,
0.5146484375,
0.468017578125,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.52978515625,
0.468017578125,
0.52978515625,
0.468017578125,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.50244140625,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.546875,
0.50244140625,
0.546875
],
[
0.0236968994140625,
0.54248046875,
0.058013916015625,
0.54248046875,
0.058013916015625,
0.60595703125,
0.0236968994140625,
0.60595703125
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50244140625,
0.54736328125,
0.861328125,
0.54736328125,
0.861328125,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.521484375,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.521484375,
0.580078125
],
[
0.080078125,
0.57373046875,
0.361083984375,
0.57470703125,
0.361083984375,
0.5888671875,
0.080078125,
0.58837890625
],
[
0.50341796875,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.59326171875,
0.50341796875,
0.59326171875
],
[
0.0269622802734375,
0.59521484375,
0.053924560546875,
0.59521484375,
0.053924560546875,
0.62890625,
0.0269622802734375,
0.62890625
],
[
0.50341796875,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.50244140625,
0.60986328125,
0.89208984375,
0.611328125,
0.89208984375,
0.62548828125,
0.50244140625,
0.625
],
[
0.021240234375,
0.6220703125,
0.056365966796875,
0.62060546875,
0.05963134765625,
0.7041015625,
0.0236968994140625,
0.705078125
],
[
0.08251953125,
0.6220703125,
0.20751953125,
0.6220703125,
0.20751953125,
0.63623046875,
0.08251953125,
0.63623046875
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.098876953125,
0.64794921875,
0.4697265625,
0.64794921875,
0.4697265625,
0.6611328125,
0.098876953125,
0.6611328125
],
[
0.50341796875,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.65380859375,
0.50341796875,
0.65380859375
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.080078125,
0.66357421875,
0.4697265625,
0.66357421875,
0.4697265625,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.07843017578125,
0.6767578125,
0.4697265625,
0.67822265625,
0.4697265625,
0.6923828125,
0.07843017578125,
0.69140625
],
[
0.50244140625,
0.67236328125,
0.8935546875,
0.67236328125,
0.8935546875,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.08087158203125,
0.693359375,
0.4697265625,
0.693359375,
0.4697265625,
0.70703125,
0.08087158203125,
0.70703125
],
[
0.50244140625,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.50244140625,
0.70068359375
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.50244140625,
0.71630859375
],
[
0.080078125,
0.708984375,
0.4697265625,
0.708984375,
0.4697265625,
0.72265625,
0.080078125,
0.72265625
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.724609375,
0.4697265625,
0.7236328125,
0.4697265625,
0.7373046875,
0.080078125,
0.73779296875
],
[
0.50244140625,
0.73291015625,
0.65185546875,
0.73291015625,
0.65185546875,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.080078125,
0.7392578125,
0.4697265625,
0.7392578125,
0.4697265625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.080078125,
0.7548828125,
0.4697265625,
0.7548828125,
0.4697265625,
0.76806640625,
0.080078125,
0.76806640625
],
[
0.521484375,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.521484375,
0.7646484375
],
[
0.08087158203125,
0.77099609375,
0.46728515625,
0.77099609375,
0.46728515625,
0.78076171875,
0.08087158203125,
0.78076171875
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.78515625,
0.4697265625,
0.78515625,
0.4697265625,
0.798828125,
0.08087158203125,
0.798828125
],
[
0.50341796875,
0.78271484375,
0.89208984375,
0.78271484375,
0.89208984375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.7998046875,
0.468017578125,
0.7998046875,
0.468017578125,
0.81298828125,
0.080078125,
0.81298828125
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.814453125,
0.468017578125,
0.814453125,
0.468017578125,
0.82763671875,
0.080078125,
0.82763671875
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.83056640625,
0.46728515625,
0.83056640625,
0.46728515625,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.8447265625,
0.468017578125,
0.8447265625,
0.468017578125,
0.85791015625,
0.080078125,
0.85791015625
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.859375,
0.4697265625,
0.859375,
0.4697265625,
0.87353515625,
0.080078125,
0.87353515625
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.096435546875,
0.88720703125,
0.37744140625,
0.88720703125,
0.37744140625,
0.90087890625,
0.096435546875,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.9267578125,
0.490234375,
0.9267578125,
0.490234375,
0.9345703125,
0.483642578125,
0.9345703125
]
] |
[
"Detect-and-Track: Efficient Pose Estimation in Videos",
"Georgia Gkioxari?",
"Lorenzo Torresani2,3*",
"Manohar Paluri?",
"Du Tran?",
"1The Robotics Institute, Carnegie Mellon University",
"Rohit Girdharl",
"2Facebook",
"3DartmouthCollege",
"https://rohitgirdhar.github.io/DetectAndTrack",
"8",
"30",
"Abstract",
"(3D) Mask R-CNN:",
"dPe",
"This paper addresses the problem of estimating and track",
"Detect person tubes +",
"",
"ing human body keypoints in complex, multi-person video.",
"keypointsinclip",
"We propose an extremely lightweight yet highly effective ap-",
"Stage#2",
"",
"",
"Bipartite Matching:",
"proach that builds upon the latest advancements in human",
"",
"Link the predictions",
"detection [17] and video understanding [5]. Our method op-",
"2",
"overtime",
"erates in two-stages: keypoint estimation in frames or short",
"LA",
"clips, followed by lightweight tracking to generate keypoint",
"Figure 1. We propose a two-stage approach to keypoint estimation",
"predictions linked over the entire video. For frame-level",
"and tracking in videos. For the first stage, we propose a novel",
"pose estimation we experiment with Mask R-CNN, as well as",
"videoposeestimationformulation,3DMaskR-CNN,thattakes",
"'SO",
"our own proposed 3D extension of this model, which lever-",
"a short video clip as input and produces a tubelet per person and",
"ages temporal information over small clips to generate more",
"keypoints within those. In the second stage, we perform light",
"robust frame predictions. We conduct extensive ablative ex-",
"weight optimization to link the detections over time.",
"8",
"periments on the newly released multi-person video pose",
"estimation benchmark, PoseTrack, to validate various design",
"choices of our model. Our approach achieves an accuracy",
"one of the more interesting aspects of video understanding,",
"of 55.2% on the validation and 51.8% on the test set using",
"namely modeling the changes in appearance and semantics",
"160'7",
"the Multi-Object Tracking Accuracy (MOTA) metric, and",
"of scenes, objects and humans over time [6, 13, 15, 37].",
"achieves state of the art performance on the ICCV 2017",
"In this work, we focus on the problem of human pose",
"PoseTrack keypoint tracking challenge [1 ].",
"trackingincomplexvideos,whichentailstrackingandes-",
"IL",
"timating the pose of each human instance over time. The",
"challenges here are plenty, including pose changes, occlu-",
"LAIXIE",
"1.Introduction",
"sions and the presence of multiple overlapping instances",
"In recent years, visual understanding, such as object and",
"Theidealtrackerneedstoaccuratelypredicttheposeof all",
"human instances at each time step by reasoning about the",
"scene recognition [17,40, 44,55], has witnessed a significant",
"bloom thanks to deep visual representations [18,31,47,50].",
"appearance and pose transitions over time. Hence, the effort",
"Modeling and understanding human behaviour in images has",
"to materialize a pose tracker should closely follow the state",
"of the art in pose prediction but also enhance it with the tools",
"been in the epicenter of a variety of visual tasks due to its im-",
"necessary to successfully integrate time information at an",
"portancefornumerousdownstreampractical applications.In",
"instance-specificlevel.",
"particular, person detection and pose estimation from a sin-",
"gle image have emerged as prominent and challenging visual",
"Most recent video pose estimation methods use hand-",
"recognitionproblems/36l.Whilesingle-imageunderstand",
"designed graphical models or integer program optimizations",
"ing has advanced steadily through the introduction of tasks",
"ontopofframe-basedkeypointpredictionstocomputethefi-",
"of increasing complexity, video understanding has made",
"nal predictions over time [21,26,48]. While such approaches",
"slower progress compared to the image domain. Here, the",
"have shown good performance, they require hand-coding of",
"preeminenttaskinvolves labeling wholevideos witha single",
"optimization constraints and may not be scalable beyond",
"activity type [5,7,10, 14,29,30,32,46,49,51,52]. While stil1",
"short video clips due to their computational complexity",
"relevant and challenging, this task shifts the focus away from",
"Most importantly, the tracking optimization is only responsi-",
"ble for linking frame-level predictions, and the system has",
"*Work done as a part of RG's internship at Facebook",
"no mechanism to improve the estimation of keypoints by",
"-"
] |
{
"arxiv_id": "1712.09184",
"language": "en",
"page_idx": 0
}
|
||
1712.09184-en-1
|
1712.09184-en-1.jpg
|
leveraging temporal information (except[48], though it is limited to the case of single person video). This implies that if a keypoint is poorly localized in a given frame, e.g., due to partial occlusion or motion blur, the prediction cannot be improved despite correlated, possibly less ambiguous, information being at hand in adjacent frames. To address this limitation, we propose a simple and effective approach which leverages the current state of the art method in pose prediction [17] and extends it by integrating temporal information from adjacent video frames by means of a novel 3D CNN architecture. It is worth noting that this architecture maintains the simplicity of our two-stage procedure: keypoint estimation is still performed at a frame-level by deploying space-time operations on short clips in a slidingwindow manner. This allows our 3D model to propagate useful information from the preceding and the subsequent frames in order to make the prediction in each frame more robust, while using a lightweight module for long-term tracking, making our method applicable to arbitrarily long videos. Fig. 1 illustrates our approach.
We train and evaluate our method on the challenging PoseTrack dataset [24], which contains real-world videos of people in various everyday scenes, and is annotated with locations of human joints along with their identity index across frames. First, and in order to convince of the efficacy of our method, we build a competitive baseline approach which links frame-level predictions, obtained from Mask R-CNN[17], in time with a simple heuristic. Our baseline approach achieves state of the art performance in the ICCV'17 PoseTrack Challenge[l], proving that it performs competitively on this new dataset. We then propose a 3D extension of Mask R-CNN, which leverages temporal information in short clips to produce more robust predictions in individual frames. For the same base architecture and image resolution, our proposed 3D model outperforms our very strong 2D baseline by 2% on keypoint mAP and 1% on the MOTA metric (details about the metrics in Sec. 4.1). In addition, our top-performing model runs at 2 minutes on a 100-frame video, with the tracking itself running in the order of seconds, showing strong potential for practical usage. As we show in Sec. 4.2, this is nearly two orders of magnitude faster than IP based formulations [26] using state-of-the-art solvers[16].
## $2. \textbf{Related Work}$
Multi-person pose estimation in images: The application of deep convolutional neural networks (CNNs) to keypoint prediction [4,17,22,40] has led to significant improvements over the last few years. Some of the most recent efforts in multi-person keypoint estimation from still images can be classified into bottom-up versus top-down techniques. Topdown approaches [17,40] involve frst locating instances by placing a tight box around them, followed by estimation
of the body joint landmarks within that box. On the other hand, bottom-up methods [4,22] involve detecting individual keypoints, and in some cases the affinities between those keypoints, and then grouping those predictions into instances. Our proposed approach builds upon these ideas by extending the top-down models to the video domain. We first predict spatio-temporal tubes over human instances in the video, followed by joint keypoint prediction within those tubes.
Multi-person pose estimation in video: Among the most dominant approaches to pose estimation from videos is a two-stage approach, which first deploys a frame-level keypoint estimator, and then connects these keypoints in space and time using optimization techniques. In [21,26], it is shown that a state of the art pose model followed by an integer programming optimization problem can result in very competitive performance in complex videos. While these approaches can handle both space-time smoothing and identity assignment, they are not applicable to long videos due to the NP-hardness of the IP optimization. Song et al.[48] propose a CRF with space-time edges and jointly optimize for $\hat{\text{the pose predictions. Although they show an improvement}}$ over frame-level predictions, their method does not consider body identities and does not address the challenging task of pose tracking. In addition, their approach is hard to generalize to an unknown number of person instances, a number that might vary even between consecutive frames due to occlusions and disocclusions. Our approach also follows a two-stage pipeline, albeit with a much less computationally expensive tracking stage, and is able to handle any number of instances per frame in a video.
Multi-object tracking in video: There has been significant effort towards multi-object tracking from video [12,43]. Prior to deep learning, the proposed solutions to tracking consisted of systems implementing a pipeline of several steps, using computationally expensive hand-crafted features and separate optimization objectives[54] for each of the proposed steps. With the advent of deep learning, end-to-end approaches for tracking have emerged. Examples include [39,45] which use recurrent neural networks (RNNs) on potentially diverse visual cues, such as appearance and motion, in order to track objects. In [1], a tracker is built upon the state of the art object detection system by adding correlation features between pair of consecutive frames in order to predict frame-level candidate boxes as well as their time deformations. More recent works have attempted to tackle detection and tracking in end-to-end fashion [20,27,28], and some works have further used such architectures for down-stream tasks such as action recognition [20]. Our work is inspired by these recent efforts but extends the task of object box tracking to address for the finer task of tracking poses in time.
|
[
[
0.08087158203125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.5009765625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.10479736328125,
0.5009765625,
0.10479736328125
],
[
0.07843017578125,
0.1060791015625,
0.468017578125,
0.106689453125,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
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,
0.468017578125,
0.1546630859375,
0.468017578125,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50341796875,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.080078125,
0.1685791015625,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.8935546875,
0.18310546875,
0.8935546875,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.1988525390625,
0.875,
0.1988525390625,
0.875,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.5009765625,
0.2310791015625,
0.89208984375,
0.231689453125,
0.89208984375,
0.246826171875,
0.5009765625,
0.24560546875
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.248779296875,
0.890625,
0.248779296875,
0.890625,
0.261962890625,
0.5009765625,
0.261962890625
],
[
0.080078125,
0.2587890625,
0.471435546875,
0.2587890625,
0.471435546875,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.26318359375,
0.89208984375,
0.26318359375,
0.89208984375,
0.27783203125,
0.50244140625,
0.27783203125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.5009765625,
0.279052734375,
0.89208984375,
0.27783203125,
0.89208984375,
0.291748046875,
0.5009765625,
0.292236328125
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.29345703125,
0.89208984375,
0.29345703125,
0.89208984375,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.07843017578125,
0.302978515625,
0.4697265625,
0.3037109375,
0.4697265625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.3212890625,
0.50244140625,
0.3212890625
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.5009765625,
0.323974609375,
0.890625,
0.322509765625,
0.890625,
0.337158203125,
0.5009765625,
0.33837890625
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.5009765625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.5009765625,
0.352783203125
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.35595703125,
0.89208984375,
0.35595703125,
0.89208984375,
0.3662109375,
0.50341796875,
0.3662109375
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50341796875,
0.369384765625,
0.89208984375,
0.369384765625,
0.89208984375,
0.38330078125,
0.50341796875,
0.38330078125
],
[
0.080078125,
0.380126953125,
0.281982421875,
0.380126953125,
0.281982421875,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.3837890625,
0.89208984375,
0.3837890625,
0.89208984375,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.1004638671875,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.1004638671875,
0.40771484375
],
[
0.50341796875,
0.399658203125,
0.89208984375,
0.399658203125,
0.89208984375,
0.41357421875,
0.50341796875,
0.41357421875
],
[
0.07843017578125,
0.40771484375,
0.4697265625,
0.4091796875,
0.4697265625,
0.423583984375,
0.07843017578125,
0.42236328125
],
[
0.50244140625,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.442626953125,
0.50244140625,
0.442626953125
],
[
0.080078125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.44580078125,
0.890625,
0.44580078125,
0.890625,
0.455810546875,
0.50341796875,
0.455810546875
],
[
0.07843017578125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.07843017578125,
0.468505859375
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.08087158203125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50341796875,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.080078125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50341796875,
0.4912109375,
0.890625,
0.4912109375,
0.890625,
0.501953125,
0.50341796875,
0.501953125
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.50244140625,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.07843017578125,
0.52880859375,
0.4697265625,
0.52978515625,
0.4697265625,
0.54443359375,
0.07843017578125,
0.54345703125
],
[
0.50244140625,
0.53515625,
0.89208984375,
0.53515625,
0.89208984375,
0.548828125,
0.50244140625,
0.548828125
],
[
0.080078125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.55029296875,
0.7255859375,
0.55029296875,
0.7255859375,
0.564453125,
0.50244140625,
0.564453125
],
[
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
],
[
0.5009765625,
0.58349609375,
0.89208984375,
0.583984375,
0.89208984375,
0.5986328125,
0.5009765625,
0.59765625
],
[
0.080078125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50244140625,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.5009765625,
0.6142578125,
0.8896484375,
0.6142578125,
0.8896484375,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.07843017578125,
0.62060546875,
0.4697265625,
0.619140625,
0.4697265625,
0.6337890625,
0.07843017578125,
0.63525390625
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.080078125,
0.65087890625,
0.46728515625,
0.65087890625,
0.46728515625,
0.6640625,
0.080078125,
0.6640625
],
[
0.50244140625,
0.64599609375,
0.8935546875,
0.64599609375,
0.8935546875,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.50244140625,
0.68994140625,
0.8935546875,
0.68994140625,
0.8935546875,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.69775390625,
0.46728515625,
0.69775390625,
0.46728515625,
0.70849609375,
0.08087158203125,
0.70849609375
],
[
0.07843017578125,
0.71044921875,
0.4697265625,
0.7109375,
0.4697265625,
0.72607421875,
0.07843017578125,
0.724609375
],
[
0.50341796875,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.5009765625,
0.7197265625,
0.8935546875,
0.72021484375,
0.8935546875,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.72802734375,
0.16259765625,
0.72802734375,
0.16259765625,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.5048828125,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.080078125,
0.755859375,
0.2181396484375,
0.755859375,
0.2181396484375,
0.76953125,
0.080078125,
0.76953125
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.5009765625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.07843017578125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81103515625,
0.07843017578125,
0.81103515625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.50341796875,
0.82763671875,
0.890625,
0.82763671875,
0.890625,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
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
],
[
0.50341796875,
0.88720703125,
0.73779296875,
0.88720703125,
0.73779296875,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"leveraging temporal information (except [48], though it is",
"of the body joint landmarks within that box. On the other",
"limited to the case of single person video).This implies that",
"hand.bottom-upmethods[4.22linvolvedetectingindividual",
"if a keypoint is poorly localized in a given frame, e.g., due",
"keypoints, and in some cases the affinities between those",
"to partial occlusion or motion blur, the prediction cannot",
"keypoints, and then grouping those predictions into instances",
"beimproved despitecorrelated,possiblyless ambiguous",
"Our proposed approach builds upon these ideas by extending",
"information being at hand in adjacent frames. To address",
"the top-down models to the video domain. We first predict",
"this limitation, we propose a simple and effective approach",
"spatio-temporal tubes over human instances in the video,",
"whichleveragesthecurrentstateoftheartmethod inpose",
"followed by joint keypoint prediction within those tubes.",
"prediction [17] and extends it by integrating temporal in-",
"formation from adjacent video frames by means of a novel",
"Multi-person pose estimation in video: Among the most",
"3D CNN architecture. It is worth noting that this architec-",
"dominant approaches to pose estimation from videos is a",
"ture maintains the simplicity of our two-stage procedure:",
"two-stage approach, which first deploys a frame-level key-",
"keypoint estimation is still performed at a frame-level by",
"point estimator,and then connects thesekeypoints in space",
"deploying space-time operations on short clips in a sliding-",
"and time using optimization techniques. In [21, 26], it is",
"window manner.This allows our 3D model to propagate",
"shown that a state of the art pose model followed by an in-",
"useful information from the preceding and the subsequent",
"teger programming optimization problem can result in very",
"frames in order to make the prediction in each frame more",
"competitive performance in complex videos. While these",
"robust, while using a lightweight module for long-term track",
"approachescanhandlebothspace-timesmoothingandiden-",
"ing, making our method applicable to arbitrarily long videos.",
"tity assignment, they are not applicable to long videos due to",
"Fig.1 illustrates our approach.",
"the NP-hardness of the IP optimization. Song et al. [48] pro-",
"We train and evaluate our method on the challenging",
"pose a CRF with space-time edges and jointly optimize for",
"PoseTrack dataset [24], which contains real-world videos",
"the pose predictions. Although they show an improvement",
"of people in various everyday scenes, and is annotated with",
"over frame-level predictions, their method does not consider",
"locations of human joints along with their identity index",
"bodyidentitiesanddoesnotaddressthechallengingtaskof",
"across frames. First, and in order to convince of the efficacy",
"pose tracking. In addition, their approach is hard to gener-",
"ofourmethodwebuildacompetitivebaselineapproach",
"alize to an unknown number of person instances, a number",
"which links frame-level predictions, obtained from Mask",
"that might varyevenbetween consecutive frames due to",
"R-CNN [17], in time with a simple heuristic. Our base",
"occlusions and disocclusions. Our approach also follows a",
"line approach achieves state of the art performance in the",
"two-stage pipeline, albeit with a much less computationally",
"ICCV'17 PoseTrack Challenge [1], proving that it performs",
"expensive tracking stage, and is able to handle any number",
"competitively on this new dataset. We then propose a 3D",
"of instances per frame in a video.",
"extension of Mask R-CNN, which leverages temporal in",
"formation in short clips to produce more robust predictions",
"Multi-object tracking in video: There has been signifi-",
"in individual frames. For the same base architecture and",
"cant effort towards multi-object tracking from video [12, 43]",
"image resolution, our proposed 3D model outperforms our",
"Prior to deep learning, the proposed solutions to tracking",
"very strong 2D baseline by 2% on keypoint mAP and 1% on",
"the MOTA metric (details about the metrics in Sec. 4.1). In",
"consisted of systems implementing a pipeline of several",
"addition, our top-performing model runs at 2 minutes on a",
"steps, using computationally expensive hand-crafted fea-",
"tures and separate optimization objectives [54] for each",
"100-frame video, with the tracking itself running in the order",
"of the proposed steps. With the advent of deep learning,",
"of seconds, showing strong potential for practical usage. As",
"end-to-end approaches for tracking have emerged. Exam-",
"we show in Sec.4.2, thisis nearly two orders of magnitude",
"faster than IP based formulations [26] using state-of-the-art",
"ples include [39, 45] which use recurrent neural networks",
"(RNNs) on potentially diverse visual cues, such as appear-",
"solvers [16].",
"ance and motion,in orderto track objects.In[lll,a tracker",
"2.Related Work",
"is built upon the state of the art object detection system",
"by adding correlation features between pair of consecutive",
"Multi-person pose estimation in images: The application",
"frames in order to predict frame-level candidate boxes as",
" of deep convolutional neural networks (CNNs) to keypoint",
"well as their time deformations. More recent works have",
"prediction [4, 17,22, 40] has led to significant improvements",
"attempted to tackle detection and tracking in end-to-end",
"over the last few years. Some of the most recent efforts in",
"fashion[20,27.281,and someworks havefurther used such",
"multi-person keypoint estimation from still images can be",
"architectures for down-stream tasks such as action recogni-",
"classified into bottom-up versus top-down techniques. Top-",
"tion [20]. Our work is inspired by these recent efforts but",
"down approaches [17, 40] involve first locating instances",
"extends the task of object box tracking to address for the",
"by placing a tight box around them, followed by estimation",
"finer task of tracking poses in time."
] |
{
"arxiv_id": "1712.09184",
"language": "en",
"page_idx": 1
}
|
||
1712.09184-en-2
|
1712.09184-en-2.jpg
|
# $3. \textbf{ Technical Approach}$
In this section, we describe our method in detail. We propose a two-stage approach that efficiently and accurately tracks human instances and their poses across time. We build a 3D human pose predictor by extending Mask RCNN[17] with spatiotemporal operations by inflating the 2D convolutions into 3D[5]. Our model takes as input short clips and predicts the poses of all people in the clips by integrating temporal information. We show that our 3D model outperforms its 2D frame-level baseline for the task of pose estimation. To track the instances in time, we perform a lightweight optimization that links the predictions. To address exponential complexities with respect to the number of frames in the video and the number of detections per frame, we employ a simple yet effective heuristic. This yields a model that achieves state of the art accuracy on the challenging PoseTrack benchmark [24] and runs orders of magnitude faster than most recent approaches [21,26].
$3. 1. \textbf{ Two- Stage Approach for Pose Tracking}$
Stage 1: Spatiotemporal pose estimation over clips. The first stage in our two-stage approach for human keypoint tracking is pose estimation using a CNN-based model. Although our approach can build upon any frame-based pose estimation system, for this work we use Mask R-CNN[17] due to its simple formulation and robust performance. Mask RCNN is a top-down keypoint estimation model that extends the Faster R-CNN object detection framework[44]. It consists of a standard base CNN, typically a ResNet[18], used to extract image features, which are then fed into task-specific small neural networks trained to propose object candidates (RPN [ 44] ) , classify them or predict their mask/ pose through an accurate feature alignment operation called RoIAlign.
We take inspiration from the recent advancements in action recognition achieved by I3D [5], which introduces a video model obtained by converting a state of the art image recognition model[23] by inflating its 2D convolutional kernels to 3D. Analogously, starting from the vanilla Mask R-CNN model, we transform the 2D convolutions to 3D. Note that the receptive feld of these 3D kernels spans over the space and time dimensions and integrates spatiotemporal cues in an end-to-end learnable fashion. Now, the input to our model is no longer a single frame, but a clip of length $T$ composed of adjacent frames sourced from a video. We extend the region proposal network (RPN)[44], to predict object candidates which track each hypothesis across the frames of the input clip. These tube proposals are used to extract instance-specific features via a spatiotemporal RoIAlign operation. The features are then fed into the 3D CNN head responsible for pose estimation. This pose-estimation head outputs heatmap activations for all keypoints across all frames conditioned on the tube hypothesis. Thus, the output of our 3D Mask R-CNN is a set of tube
hypotheses with keypoint estimates. Fig. 2 illustrates our proposed 3D Mask R-CNN model, which we describe in detail next
Base network: We extend a standard ResNet[18] architecture to a 3D ResNet architecture by replacing all 2D convolutions with 3D convolutions. We set the temporal extent of our kernels $(K_T)$ to match the spatial width, except for the first convolutional layer, which uses filters of size $3\times7\times7.$ We temporally pad the convolutions as for the spatial dimensions: padding of 1 for $K_T=3$ and 0 for $K_T=1.$ We set temporal strides to 1, as we empirically found that larger stride values lead to lower performance. Inspired by [5,8], we initialize the 3D ResNet using a pretrained 2D ResNet. Apart from their proposed “mean”initialization, which replicates the 2D filter temporally and divides the coefficients by the number of repetitions, we also experiment with a a “center" initialization method, which has earlier been used for action recognition tasks [9]. In this setup, we initialize the central 2D slice of the 3D kernel with the 2D filter weights and set all the other 2D slices (corresponding to temporal displacements) to zero. We empirically show in Sec. 4.3 that center initialization scheme leads to better performance. The final feature map output of the base 3D network for a $T\times H\times W$ input is $T\times\frac{H}{8}\times\frac{W}{8}$, as we clip the network after the fourth residual block and perform no temporal striding Tube proposal network: We design a tube proposal network inspired by the region proposal network (RPN) in Faster R-CNN[44]. Given the feature map from the base network, we slide a small 3D-conv network connected to two sibling fully connected layers-tube classification $(cls)$ and regression $(reg).$ The cls and reg labels are defined with respect to tube anchors. We design the tube anchors to be similar to the bounding box anchors used in Faster R-CNN, but here we replicate them in time. We use $A$ (typically 12) different anchors at every sliding position, differing in scale and/or aspect ratio. Thus, we have $\frac{H}{8}\times\frac{W}{8}\times A$ anchors in total. For each of these anchors, cls predicts a binary value indicating whether a foreground tube originating at that spatial position has a high overlap with our proposal tube. Similarly, reg outputs for each anchor a 4$T$-dimensional vector encoding displacements with respect to the anchor coordinates for each box in the tube. We use the softmax classification loss for training the cls layer, and the smoothed $L_1$ loss for the reg layer. We scale the $reg\log$loss by $\frac1T$,in order to keep itsreg layes comparable to those of the loss for the 2D case. We define these losses as our tracking loss.
3D Mask R-CNN heads: Given the tube candidates produced by the tube proposal network, the next step classifies and regresses them into a tight tube around a person track. We compute the region features for this tube by designing a 3D region transform operator. In particular, we extend the RoIAlign operation[17] to extract a spatiotemporal feature map from the output of the base network. Since the temporal
|
[
[
0.08087158203125,
0.091552734375,
0.26708984375,
0.091552734375,
0.26708984375,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50341796875,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.5009765625,
0.106689453125,
0.89208984375,
0.1060791015625,
0.89208984375,
0.12060546875,
0.5009765625,
0.121826171875
],
[
0.098876953125,
0.11614990234375,
0.468017578125,
0.11614990234375,
0.468017578125,
0.1300048828125,
0.098876953125,
0.1300048828125
],
[
0.50341796875,
0.125,
0.57861328125,
0.125,
0.57861328125,
0.1363525390625,
0.50341796875,
0.1363525390625
],
[
0.080078125,
0.1319580078125,
0.468017578125,
0.1319580078125,
0.468017578125,
0.1458740234375,
0.080078125,
0.1458740234375
],
[
0.50244140625,
0.14013671875,
0.89208984375,
0.14013671875,
0.89208984375,
0.154052734375,
0.50244140625,
0.154052734375
],
[
0.080078125,
0.146484375,
0.4697265625,
0.146484375,
0.4697265625,
0.160400390625,
0.080078125,
0.160400390625
],
[
0.50244140625,
0.1558837890625,
0.89208984375,
0.1558837890625,
0.89208984375,
0.169189453125,
0.50244140625,
0.169189453125
],
[
0.080078125,
0.1622314453125,
0.4697265625,
0.1622314453125,
0.4697265625,
0.175537109375,
0.080078125,
0.175537109375
],
[
0.50244140625,
0.17041015625,
0.8935546875,
0.17041015625,
0.8935546875,
0.184326171875,
0.50244140625,
0.184326171875
],
[
0.080078125,
0.1767578125,
0.4697265625,
0.1767578125,
0.4697265625,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.080078125,
0.1912841796875,
0.468017578125,
0.1912841796875,
0.468017578125,
0.2052001953125,
0.080078125,
0.2052001953125
],
[
0.50341796875,
0.1875,
0.890625,
0.1875,
0.890625,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.50244140625,
0.2008056640625,
0.8935546875,
0.2008056640625,
0.8935546875,
0.2139892578125,
0.50244140625,
0.2139892578125
],
[
0.080078125,
0.20703125,
0.4697265625,
0.20703125,
0.4697265625,
0.2203369140625,
0.080078125,
0.2203369140625
],
[
0.50244140625,
0.216552734375,
0.89208984375,
0.216552734375,
0.89208984375,
0.2298583984375,
0.50244140625,
0.2298583984375
],
[
0.080078125,
0.222900390625,
0.468017578125,
0.222900390625,
0.468017578125,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.5009765625,
0.2310791015625,
0.89208984375,
0.2298583984375,
0.89208984375,
0.2442626953125,
0.5009765625,
0.24560546875
],
[
0.080078125,
0.2374267578125,
0.4697265625,
0.2374267578125,
0.4697265625,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.50244140625,
0.246826171875,
0.89208984375,
0.246826171875,
0.89208984375,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.08087158203125,
0.25390625,
0.46728515625,
0.25390625,
0.46728515625,
0.264404296875,
0.08087158203125,
0.264404296875
],
[
0.080078125,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.27978515625,
0.080078125,
0.27978515625
],
[
0.50341796875,
0.26318359375,
0.890625,
0.26318359375,
0.890625,
0.2734375,
0.50341796875,
0.2734375
],
[
0.50244140625,
0.276611328125,
0.8935546875,
0.276611328125,
0.8935546875,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.08087158203125,
0.2841796875,
0.468017578125,
0.2841796875,
0.468017578125,
0.294189453125,
0.08087158203125,
0.294189453125
],
[
0.080078125,
0.29736328125,
0.4697265625,
0.29736328125,
0.4697265625,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.50244140625,
0.292236328125,
0.89208984375,
0.292236328125,
0.89208984375,
0.30615234375,
0.50244140625,
0.30615234375
],
[
0.50244140625,
0.306884765625,
0.89208984375,
0.306884765625,
0.89208984375,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.080078125,
0.313232421875,
0.4697265625,
0.313232421875,
0.4697265625,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.50244140625,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50244140625,
0.335205078125
],
[
0.08087158203125,
0.330078125,
0.46728515625,
0.330078125,
0.46728515625,
0.34033203125,
0.08087158203125,
0.34033203125
],
[
0.07843017578125,
0.34228515625,
0.4697265625,
0.34228515625,
0.4697265625,
0.35595703125,
0.07843017578125,
0.35595703125
],
[
0.50341796875,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.50341796875,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.365478515625,
0.50341796875,
0.365478515625
],
[
0.080078125,
0.35791015625,
0.435546875,
0.35791015625,
0.435546875,
0.371826171875,
0.080078125,
0.371826171875
],
[
0.50244140625,
0.367431640625,
0.89208984375,
0.367431640625,
0.89208984375,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.080078125,
0.380126953125,
0.414306640625,
0.380126953125,
0.414306640625,
0.39453125,
0.080078125,
0.39453125
],
[
0.50341796875,
0.3837890625,
0.890625,
0.3837890625,
0.890625,
0.39453125,
0.50341796875,
0.39453125
],
[
0.50244140625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.080078125,
0.404052734375,
0.471435546875,
0.404052734375,
0.471435546875,
0.417236328125,
0.080078125,
0.417236328125
],
[
0.07843017578125,
0.41845703125,
0.468017578125,
0.417236328125,
0.468017578125,
0.431884765625,
0.07843017578125,
0.43310546875
],
[
0.50341796875,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.4248046875,
0.50341796875,
0.4248046875
],
[
0.50244140625,
0.427978515625,
0.89208984375,
0.427978515625,
0.89208984375,
0.44140625,
0.50244140625,
0.44140625
],
[
0.080078125,
0.434326171875,
0.4697265625,
0.434326171875,
0.4697265625,
0.447509765625,
0.080078125,
0.447509765625
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.080078125,
0.449462890625,
0.4697265625,
0.449462890625,
0.4697265625,
0.46337890625,
0.080078125,
0.46337890625
],
[
0.5009765625,
0.45703125,
0.88818359375,
0.458251953125,
0.88818359375,
0.472900390625,
0.5009765625,
0.4716796875
],
[
0.080078125,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.080078125,
0.479248046875,
0.4697265625,
0.479248046875,
0.4697265625,
0.492431640625,
0.080078125,
0.492431640625
],
[
0.50244140625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.5048828125,
0.50244140625,
0.5048828125
],
[
0.080078125,
0.5087890625,
0.4697265625,
0.5087890625,
0.4697265625,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.08087158203125,
0.54150390625,
0.46728515625,
0.54150390625,
0.46728515625,
0.5517578125,
0.08087158203125,
0.5517578125
],
[
0.50244140625,
0.5361328125,
0.89208984375,
0.5361328125,
0.89208984375,
0.5498046875,
0.50244140625,
0.5498046875
],
[
0.080078125,
0.5546875,
0.4697265625,
0.5546875,
0.4697265625,
0.568359375,
0.080078125,
0.568359375
],
[
0.50341796875,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50341796875,
0.564453125
],
[
0.080078125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.583984375,
0.080078125,
0.583984375
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.583984375,
0.449462890625,
0.583984375,
0.449462890625,
0.59765625,
0.080078125,
0.59765625
],
[
0.50244140625,
0.58203125,
0.89208984375,
0.58203125,
0.89208984375,
0.59521484375,
0.50244140625,
0.59521484375
],
[
0.098876953125,
0.5986328125,
0.4697265625,
0.599609375,
0.4697265625,
0.61279296875,
0.098876953125,
0.6123046875
],
[
0.5009765625,
0.59521484375,
0.89208984375,
0.5966796875,
0.89208984375,
0.611328125,
0.5009765625,
0.60986328125
],
[
0.080078125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.080078125,
0.62890625
],
[
0.5009765625,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.5009765625,
0.625
],
[
0.07843017578125,
0.62890625,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.07843017578125,
0.64306640625
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.08087158203125,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.080078125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.67236328125,
0.8935546875,
0.67236328125,
0.8935546875,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.6875,
0.8935546875,
0.6875,
0.8935546875,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
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,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76318359375,
0.89208984375,
0.76318359375,
0.89208984375,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.080078125,
0.7822265625,
0.4697265625,
0.7822265625,
0.4697265625,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50244140625,
0.77880859375,
0.76123046875,
0.77880859375,
0.76123046875,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.07843017578125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82568359375,
0.4697265625,
0.82470703125,
0.4697265625,
0.8388671875,
0.07843017578125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"3.Technical Approach",
"hypotheses with keypoint estimates. Fig. 2 illustrates our",
"proposed 3D Mask R-CNN model, which we describe in",
"In this section, we describe our method in detail. We",
"detail next.",
"propose a two-stage approach that efficiently and accurately",
"Base network: We extend a standard ResNet [18] architec-",
"tracks human instances and their poses across time. We",
"ture to a 3D ResNet architecture by replacing all 2D convo-",
"build a 3D human pose predictor by extending Mask R-",
"lutions with 3D convolutions. We set the temporal extent of",
"CNN [17] with spatiotemporal operations by inflating the",
"2D convolutions into 3D [5]. Our model takes as input",
"ourkernels(KT)tomatchthespatial width.exceptforthe",
"first convolutional layer, which uses filters of size 3 × 7 × 7.",
"short clips and predicts the poses of all people in the clips",
"We temporally pad the convolutions as for the spatial dimen-",
"by integrating temporal information.We show that our 3D",
"sions: padding of 1 for KT = 3 and 0 for KT = 1. We",
"model outperforms its 2D frame-level baseline for the task of",
"set temporal strides to 1, as we empirically found that larger",
"poseestimation.Totrack theinstances intime,weperform",
"a lightweight optimization that links the predictions. To",
"stridevaluesleadtolowerperformance.Inspiredby[5,8]",
"we initialize the 3D ResNet using a pretrained 2D ResNet.",
"addressexponentialcomplexitieswithrespecttothenumber",
"of frames in the video and the number of detections per",
"Apart from their proposed “mean\" initialization, which repli-",
"cates the 2D filter temporally and divides the coefficients by",
"frame, we employ a simple yet effective heuristic. This",
"the number of repetitions, we also experiment with a a “\"cen-",
"yieldsamodelthatachieves stateoftheartaccuracyonthe",
"challenging PoseTrack benchmark [24] and runs orders of",
"terinitializationmethod,which has earlier beenusedfor",
"action recognition tasks [9]. In this setup, we initialize the",
"magnitude faster than most recent approaches [21, 26]",
"central 2D slice of the 3D kernel with the 2D filter weights",
"3.1. Two-Stage Approach for Pose Tracking",
"and set all the other 2D slices (correspondingto temporal",
"displacements) to zero. We empirically show in Sec. 4.3",
"Stage 1: Spatiotemporal pose estimation over clips.",
"The first stage in our two-stage approach for human keypoint",
"that centerinitialization schemeleadstobetterperformance",
"The final feature map output of the base 3D network for a",
"tracking is pose estimation using a CNN-based model. Al-",
"T × H × W input is T× × W, as we clip the network after",
"though our approach can build upon any frame-based pose es-",
"the fourth residual block and perform no temporal striding.",
"timation system, for this work we use Mask R-CNN [17] due",
"to its simple formulation and robust performance. Mask R-",
"Tube proposal network: We design a tube proposal net-",
"CNN is a top-down keypoint estimation model that extends",
"work inspired by the region proposal network (RPN) in",
"the Faster R-CNN object detection framework [44]. It con-",
"Faster R-CNN [44]. Given the feature map from the base",
"sists of a standard base CNN, typically a ResNet [18], used to",
"network, we slide a small 3D-conv network connected to",
"extractimagefeatures,whicharethenfedintotask-specific",
"two sibling fully connected layers - tube classification (cls)",
"small neural networks trained to propose object candidates",
"and regression (reg). The cls and reg labels are defined with",
"(RPN [44]), classify them or predict their mask/pose through",
"respect to tube anchors. We design the tube anchors to be",
"an accurate feature alignment operation called RolAlign",
"similar to the bounding box anchors used in Faster R-CNN,",
"We take inspiration from the recent advancements in ac-",
"but here we replicate them in time. We use A (typically 12)",
"tion recognition achieved by I3D [5], which introduces a",
"different anchors at every sliding position, differing in scale",
"video model obtained by converting a state of the art im-",
"and/or aspectratio.Thus,we have x x A anchors in",
"agerecognitionmodel[23]byinflatingits2Dconvolutional",
"total. For each of these anchors, cls predicts a binary value in-",
"kernels to 3D. Analogously, starting from the vanilla Mask",
"dicating whether a foreground tube originating at that spatial",
"R-CNN model, we transform the 2D convolutions to 3D.",
"position has a high overlap with our proposal tube. Similarly,",
"Note that the receptive field of these 3D kernels spans over",
"reg outputs for each anchor a 4T-dimensional vector encod-",
"the space and time dimensions and integrates spatiotem-",
"ingdisplacementswithrespecttotheanchorcoordinatesfor",
"poral cues in an end-to-end learnable fashion. Now, the",
"each box in the tube. We use the softmax classification loss",
"input to our model is no longer a single frame, but a clip",
"for training the cls layer, and the smoothed Li loss for the",
"of length T composed of adjacent frames sourced from a",
"reg layer. We scale the reg loss by , in order to keep its",
"video. We extend the region proposal network (RPN) [44].",
"values comparable to those of the loss for the 2D case. We",
"to predict object candidates which track each hypothesis",
"define these losses as our tracking loss.",
" across the frames of the input clip. These tube proposals",
"3D Mask R-CNN heads: Given the tube candidates pro-",
"are used to extract instance-specific features via a spatio-",
"duced by the tube proposal network, the next step classifies",
"temporal RolAlign operation. The features are then fed into",
"and regresses them into a tight tube around a person track.",
"the 3D CNN head responsible for pose estimation. This",
"We compute the region features for this tube by designing a",
"pose-estimation head outputs heatmap activations for all key",
"3D region transform operator. In particular, we extend the",
"points across all frames conditioned on the tube hypothesis.",
"RolAlign operation [17] to extract a spatiotemporal feature",
"Thus, the output of our 3D Mask R-CNN is a set of tube",
"map from the output of the base network. Since the temporal"
] |
{
"arxiv_id": "1712.09184",
"language": "en",
"page_idx": 2
}
|
||
1712.09184-en-3
|
1712.09184-en-3.jpg
|
<FigureHere>
Figure 2. Proposed 3D Mask R-CNN network architecture: Our architecture, as described in Sec. 3.1, has three main parts. The base network is a standard ResNet, extended to 3D. It generates a 3D feature blob, which is then used to generate proposal tubes using the Tube Proposal Network (TPN). The tubes are used to extract region features from the 3D feature blob, using a spatiotemporal RoIAlign operation, and are fed into heads that classify/regress for a tight tube and another to predict keypoint heatmaps.
extent of the feature map and the tube candidates is the same (of dimension $T)$,we split the tube into $T$ 2D boxes, and use RoIAlign to extract a region from each of the $T$ temporal slices in the feature map. The regions are then concatenated in time to produce a $T\times R\times R$ feature map, where $R$ is the output resolution of RoIAlign operation, which is kept 7 for the $cls/reg$ head, and 14 for the keypoint head. The classification head consists of a 3D ResNet block, similar to the design of the 3D ResNet blocks from the base network; and the keypoint head consists of 8 3D conv layers, followed by 2 deconvolution layers to generate the keypoint heatmap output for each time frame input. The classification head is trained with a softmax loss for the cls output and a smoothed $L_1$ loss for the reg output, while the keypoint head is trained with a spatial softmax loss, similar to[17].
propagate the labels forward using the matches, one frame at a time. Any boxes that do not get matched to an existing track instantiate a new track. As we show in Sec. 4.2, this simple approach is very effective in getting good tracks, is highly scalable, is able to deal with a varying number of person hypotheses, and can run on videos of arbitrary length. Likelihood metrics: We experiment with a variety of handcrafted and learned likelihood metrics for linking the tracks. In terms of hand-crafted features, we specifically experiment with: 1) Visual similarity, defined as the cosine distance between CNN features extracted from the image patch represented by the detection; 2) Location similarity, defined as the box intersection over union (IoU) of the two detection boxes; and 3) Pose similarity, defined as the PCKh[53] distance between the poses in the two frames. We also experiment with a learned distance metric based on a LSTM model that incorporates track history in predicting whether a new detection is part of the track or not. At test time, the predicted confidence values are used in the matching algorithm, and the matched detection is used to update the LSTM hidden state. Similar ideas have also shown good performance for traditional tracking tasks [45].
In Sec. 4 we present an extensive ablative analysis of the various design choices in our two-stage architecture described above. While being extremely lightweight and simple to implement, our final model obtains state of the art performance on the benchmark, out-performing all submissions in the ICCV'17 PoseTrack challenge[1].
Stage 2: Linking keypoint predictions into tracks. Given these keypoint predictions grouped in space by person identity (i.e., pose estimation), we need to link them in time to obtain keypoint tracks. Tracking can be seen as a data association problem over these detections. Previous approaches, such as [41], have formulated this task as a bipartite matching problem, which can be solved using the Hungarian algorithm [33] or greedy approaches. More recent work has incorporated deep recurrent neural networks $(\mathbf{RNN})$, such as an LSTM [19], to model the temporal evolutions of features along the tracks [39,45]. We use a similar strategy, and represent these detections in a graph, where each detected bounding box (representing a person) in a frame becomes a node. We define edges to connect each box in a frame to every box in the next frame. The cost of each edge is defined as the negative likelihood of the two boxes linked on that edge to belong to the same person. We experimented with both hand-crafted and learned likelihood metrics, which we describe in the next paragraph. Given these likelihood values, we compute tracks by simplifying the problem to bipartite matching between each pair of adjacent frames. We initialize tracks on the first frame and
## 4. Experiments and Results
We introduce the PoseTrack challenge benchmark and experimentally evaluate the various design choices of our model. We first build a strong baseline with our two-stage keypoint tracking approach that obtains state of the art performance on this challenging dataset. Then, we show how our 3D Mask R-CNN formulation can further improve upon that model by incorporating temporal context.
|
[
[
0.69384765625,
0.09661865234375,
0.7548828125,
0.09661865234375,
0.7548828125,
0.106689453125,
0.69384765625,
0.106689453125
],
[
0.4306640625,
0.1060791015625,
0.451904296875,
0.1060791015625,
0.451904296875,
0.1142578125,
0.4306640625,
0.1142578125
],
[
0.8388671875,
0.10162353515625,
0.8759765625,
0.10162353515625,
0.8759765625,
0.11236572265625,
0.8388671875,
0.11236572265625
],
[
0.70751953125,
0.10797119140625,
0.734375,
0.10797119140625,
0.734375,
0.11614990234375,
0.70751953125,
0.11614990234375
],
[
0.426513671875,
0.11553955078125,
0.4697265625,
0.11553955078125,
0.4697265625,
0.1256103515625,
0.426513671875,
0.1256103515625
],
[
0.54833984375,
0.11614990234375,
0.62353515625,
0.11614990234375,
0.62353515625,
0.1275634765625,
0.54833984375,
0.1275634765625
],
[
0.4306640625,
0.1256103515625,
0.4697265625,
0.1256103515625,
0.4697265625,
0.1363525390625,
0.4306640625,
0.1363525390625
],
[
0.5498046875,
0.1275634765625,
0.60107421875,
0.1275634765625,
0.60107421875,
0.1383056640625,
0.5498046875,
0.1383056640625
],
[
0.5498046875,
0.136962890625,
0.62158203125,
0.136962890625,
0.62158203125,
0.147705078125,
0.5498046875,
0.147705078125
],
[
0.5498046875,
0.147705078125,
0.60693359375,
0.147705078125,
0.60693359375,
0.1590576171875,
0.5498046875,
0.1590576171875
],
[
0.84619140625,
0.147705078125,
0.87744140625,
0.147705078125,
0.87744140625,
0.1558837890625,
0.84619140625,
0.1558837890625
],
[
0.708984375,
0.1610107421875,
0.7587890625,
0.1610107421875,
0.7587890625,
0.1717529296875,
0.708984375,
0.1717529296875
],
[
0.71826171875,
0.1717529296875,
0.7490234375,
0.1717529296875,
0.7490234375,
0.18310546875,
0.71826171875,
0.18310546875
],
[
0.56396484375,
0.1944580078125,
0.6005859375,
0.1944580078125,
0.6005859375,
0.2052001953125,
0.56396484375,
0.2052001953125
],
[
0.869140625,
0.1925048828125,
0.87841796875,
0.1925048828125,
0.87841796875,
0.1976318359375,
0.869140625,
0.1976318359375
],
[
0.56396484375,
0.205810546875,
0.60302734375,
0.205810546875,
0.60302734375,
0.21337890625,
0.56396484375,
0.21337890625
],
[
0.86767578125,
0.2020263671875,
0.88232421875,
0.2020263671875,
0.88232421875,
0.208984375,
0.86767578125,
0.208984375
],
[
0.560546875,
0.212158203125,
0.60546875,
0.21533203125,
0.6044921875,
0.2266845703125,
0.5595703125,
0.2235107421875
],
[
0.378173828125,
0.2197265625,
0.3896484375,
0.2197265625,
0.3896484375,
0.222900390625,
0.378173828125,
0.222900390625
],
[
0.374267578125,
0.2259521484375,
0.416748046875,
0.2259521484375,
0.416748046875,
0.236083984375,
0.374267578125,
0.236083984375
],
[
0.080078125,
0.240478515625,
0.55810546875,
0.240478515625,
0.55810546875,
0.25390625,
0.080078125,
0.25390625
],
[
0.5556640625,
0.242431640625,
0.890625,
0.242431640625,
0.890625,
0.25244140625,
0.5556640625,
0.25244140625
],
[
0.080078125,
0.255126953125,
0.890625,
0.255126953125,
0.890625,
0.268310546875,
0.080078125,
0.268310546875
],
[
0.07843017578125,
0.267578125,
0.8935546875,
0.268310546875,
0.8935546875,
0.28271484375,
0.07843017578125,
0.2822265625
],
[
0.07843017578125,
0.28271484375,
0.67236328125,
0.28271484375,
0.67236328125,
0.296630859375,
0.07843017578125,
0.296630859375
],
[
0.080078125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.335205078125,
0.080078125,
0.335205078125
],
[
0.5009765625,
0.3212890625,
0.89208984375,
0.32080078125,
0.89208984375,
0.335205078125,
0.5009765625,
0.3359375
],
[
0.080078125,
0.337158203125,
0.4697265625,
0.337158203125,
0.4697265625,
0.35107421875,
0.080078125,
0.35107421875
],
[
0.5009765625,
0.3359375,
0.89208984375,
0.337158203125,
0.89208984375,
0.3515625,
0.5009765625,
0.35107421875
],
[
0.08087158203125,
0.354248046875,
0.46728515625,
0.354248046875,
0.46728515625,
0.3642578125,
0.08087158203125,
0.3642578125
],
[
0.50341796875,
0.354248046875,
0.8896484375,
0.354248046875,
0.8896484375,
0.3642578125,
0.50341796875,
0.3642578125
],
[
0.08087158203125,
0.369384765625,
0.468017578125,
0.369384765625,
0.468017578125,
0.380126953125,
0.08087158203125,
0.380126953125
],
[
0.50244140625,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50244140625,
0.3818359375
],
[
0.080078125,
0.38330078125,
0.4697265625,
0.38330078125,
0.4697265625,
0.396484375,
0.080078125,
0.396484375
],
[
0.50244140625,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.396484375,
0.50244140625,
0.396484375
],
[
0.08087158203125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.410400390625,
0.08087158203125,
0.410400390625
],
[
0.50244140625,
0.3984375,
0.8935546875,
0.3984375,
0.8935546875,
0.412353515625,
0.50244140625,
0.412353515625
],
[
0.080078125,
0.41357421875,
0.4697265625,
0.41357421875,
0.4697265625,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.4140625,
0.8935546875,
0.4140625,
0.8935546875,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.427978515625,
0.4697265625,
0.427978515625,
0.4697265625,
0.44140625,
0.080078125,
0.44140625
],
[
0.5009765625,
0.429931640625,
0.8935546875,
0.429931640625,
0.8935546875,
0.443115234375,
0.5009765625,
0.443115234375
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.458984375,
0.50244140625,
0.458984375
],
[
0.07843017578125,
0.458251953125,
0.4697265625,
0.458251953125,
0.4697265625,
0.4716796875,
0.07843017578125,
0.4716796875
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.5009765625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.4892578125,
0.5009765625,
0.4892578125
],
[
0.07843017578125,
0.488037109375,
0.4697265625,
0.488037109375,
0.4697265625,
0.501953125,
0.07843017578125,
0.501953125
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.08087158203125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51513671875,
0.08087158203125,
0.51513671875
],
[
0.5009765625,
0.50390625,
0.8935546875,
0.505859375,
0.8935546875,
0.51953125,
0.5009765625,
0.517578125
],
[
0.08087158203125,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.080078125,
0.5341796875,
0.358642578125,
0.5341796875,
0.358642578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50341796875,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.5693359375,
0.4697265625,
0.5693359375,
0.4697265625,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.50244140625,
0.580078125
],
[
0.080078125,
0.58544921875,
0.4697265625,
0.58544921875,
0.4697265625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50341796875,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.59326171875,
0.50341796875,
0.59326171875
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.080078125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.50244140625,
0.625
],
[
0.080078125,
0.630859375,
0.471435546875,
0.630859375,
0.471435546875,
0.64453125,
0.080078125,
0.64453125
],
[
0.5009765625,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.6396484375,
0.5009765625,
0.6396484375
],
[
0.0767822265625,
0.64599609375,
0.4697265625,
0.64306640625,
0.4697265625,
0.6572265625,
0.0767822265625,
0.6591796875
],
[
0.50244140625,
0.64306640625,
0.69921875,
0.64306640625,
0.69921875,
0.65380859375,
0.50244140625,
0.65380859375
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.51953125,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.51953125,
0.67236328125
],
[
0.080078125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.07843017578125,
0.72021484375,
0.471435546875,
0.72021484375,
0.471435546875,
0.734375,
0.07843017578125,
0.734375
],
[
0.50244140625,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.080078125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73291015625,
0.80810546875,
0.73291015625,
0.80810546875,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.07843017578125,
0.75,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.07843017578125,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.76953125,
0.73388671875,
0.76953125,
0.73388671875,
0.78271484375,
0.50341796875,
0.78271484375
],
[
0.08087158203125,
0.78271484375,
0.4697265625,
0.78271484375,
0.4697265625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.52294921875,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.07843017578125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82763671875,
0.468017578125,
0.82763671875,
0.468017578125,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.5,
0.82470703125,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.5,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.8935546875,
0.84130859375,
0.8935546875,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5,
0.88525390625
],
[
0.0767822265625,
0.88623046875,
0.4697265625,
0.88525390625,
0.4697265625,
0.8994140625,
0.0767822265625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8037109375,
0.88720703125,
0.8037109375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Classification",
"Tube",
"cls loss",
"Head",
"Proposal",
"\"A\"proposalsat",
"Network",
"every pixel,",
"Trainedwith.cls",
"and reg loss",
"eqloss",
"Keypoints",
"Head",
"Spatio-",
"n",
"Temporal",
"loss",
"RolAlign",
"Ba",
"Network",
"Figure 2. Proposed 3D Mask R-CNN network architecture: Our architecture,",
"as described in Sec.3.l,has three mainparts.Thebase",
"network is a standard ResNet, extended to 3D. It generates a 3D feature blob, which is then used to generate proposal tubes using the Tube",
"Proposal Network (TPN). The tubes are used to extract region features from the 3D feature blob, using a spatiotemporal RoIAlign operation,",
"and are fed into heads that classify/regress for a tight tube and another to predict keypoint heatmaps.",
"extent of the feature map and the tube candidates is the same",
"propagate the labels forward using the matches, one frame",
"(of dimension T), we split the tube into T 2D boxes, and use",
"at a time. Any boxes that do not get matched to an existing",
"RolAligntoextractaregionfromeachoftheTtemporal",
"track instantiate a new track. As we show in Sec.4.2, this",
"slices inthe feature map.Theregions are then concatenated",
"simple approach is very effective in getting good tracks, is",
"in time to produce a T × R x R feature map, where R is",
"highly scalable, is able to deal with a varying number of",
"the outputresolution of RolAlign operation, which iskept",
"person hypotheses, and can run on videos of arbitrary length.",
"7 for the cls/reg head, and 14 for the keypoint head. The",
"Likelihood metrics: We experiment with a variety of hand-",
"classification head consists of a 3D ResNet block, similar to",
"crafted and learned likelihood metrics for linking the tracks.",
"the design of the 3D ResNet blocks from the base network;",
"In terms of hand-crafted features, we specifically experiment",
" and the keypoint head consists of 8 3D conv layers, followed",
"with: 1) Visual similarity, defined as the cosine distance",
"by2deconvolutionlayerstogeneratethekeypointheatmap",
"between CNNfeatures extracted fromthe imagepatch repre-",
" output for each time frame input. The classification head is",
"sented by the detection; 2) Location similarity, defined as the",
"trainedwitha softmaxlossforthecls outputandasmoothed",
"box intersection over union (loU) of the two detection boxes;",
"L loss for the reg output, while the keypoint head is trained",
"and 3) Pose similarity, defined as the PCKh [53] distance",
"with a spatial softmax loss, similar to [17]",
"betweentheposesinthetwoframes.Wealsoexperiment",
"with a learned distance metric based on a LSTM model that",
"Stage 2: Linking keypoint predictions into tracks.",
"incorporates track history in predicting whether a new de-",
"Given these keypoint predictions grouped in space by per-",
"tectionispart ofthetrackornot.Attesttime,thepredicted",
"son identity (i.e., pose estimation), we need to link them",
"confidence values are used in the matching algorithm, and",
"in time to obtain keypoint tracks. Tracking can be seen as",
"the matched detection is used to update the LSTM hidden",
"a data association problem over these detections. Previous",
"state. Similar ideas have also shown good performance for",
"approaches, such as[41], haveformulated this task as a bi-",
"traditionaltrackingtasks[45]",
"partite matching problem, which can be solved using the",
"In Sec. 4 we present an extensive ablative analysis of",
"Hungarian algorithm [33] or greedy approaches. More re-",
"the various design choices in our two-stage architecture",
"cent work has incorporated deep recurrent neural networks",
"described above. While being extremely lightweight and",
"(RNN), such as an LSTM [19], to model the temporal evolu-",
"simple to implement, our final model obtains state of the art",
"tions of features along the tracks [39, 45]. We use a similar",
"performance on the benchmark, out-performing all submis-",
"strategy, and represent these detections in a graph, where",
"sions in the ICCV' 17 PoseTrack challenge [1]",
"each detected bounding box (representing a person) in a",
"frame becomes a node. We define edges to connect each",
"4.Experiments and Results",
"box in a frameto everybox in the nextframe.The cost of",
"each edge is defined as the negative likelihood of the two",
"We introduce the PoseTrack challenge benchmark and",
"boxes linked on that edge to belong to the same person. We",
"experimentally evaluate the various design choices of our",
"experimentedwithbothhand-craftedandlearnedlikelihood",
"model. We first build a strong baseline with our two-stage",
"metrics, which we describe in the next paragraph. Given",
"keypoint tracking approach that obtains state of the art per-",
"these likelihood values, we compute tracks by simplifying",
"formance on this challenging dataset. Then, we show how",
"the problem to bipartite matching between each pair of ad-",
"our 3D Mask R-CNN formulation can further improve upon",
"jacent frames. We initialize tracks on the first frame and",
"that model by incorporating temporal context."
] |
{
"arxiv_id": "1712.09184",
"language": "en",
"page_idx": 3
}
|
||
1712.09184-en-4
|
1712.09184-en-4.jpg
|
<table>
<tbody>
<tr>
<th>Threshold</th>
<th>mAP Head</th>
<th>mAP Shou</th>
<th>nAP Elbo</th>
<th>nAP $\Pi$ Wri</th>
<th>$\iota AP$ m $\lim$ $K_{\mathrm{I}}$</th>
<th>mAI $p$ Ank $e$</th>
<th>nAP $n$ Ankl T</th>
<th>mAP Total</th>
<th>$10TA$ $N$ Head</th>
<th>MOTA Shou</th>
<th>MOTA $E1b$</th>
<th>MOTA Wri</th>
<th>$MOT_{\iota}$ Knee</th>
<th>$\Gamma A$ $MO$ $An$ $ee$</th>
<th>MOTA 4 Total</th>
<th>MOTP Total</th>
<th>Prec 1 Total</th>
<th>Rec Tota</th>
</tr>
<tr>
<td>0.0,r random tracks</td>
<td>72.8</td>
<td>75.6</td>
<td>65.3</td>
<td>54.3</td>
<td>3.5 6(</td>
<td>51.8 9</td>
<td>51.8</td>
<td>64.1</td>
<td>-11.6</td>
<td>-6.6</td>
<td>-8.5</td>
<td>-12.9</td>
<td>-10.2</td>
<td>.2 -9.</td>
<td>-10.2</td>
<td>55.8</td>
<td>83.3</td>
<td>70.8</td>
</tr>
<tr>
<td>0.0</td>
<td>72.8</td>
<td>75.6</td>
<td>65.3</td>
<td>54.3 1</td>
<td>3.5 6(</td>
<td>51.8 1</td>
<td>51.8</td>
<td>64.1</td>
<td>60.3</td>
<td>65.3</td>
<td>55.8</td>
<td>43.5</td>
<td>50.7</td>
<td>.7 43</td>
<td>53.6</td>
<td>55.7</td>
<td>83.3</td>
<td>70.8</td>
</tr>
<tr>
<td>0.5</td>
<td>72.8</td>
<td>75.6</td>
<td>65.3</td>
<td>54.3</td>
<td>3.5 6(</td>
<td>51.8 9</td>
<td>51.8</td>
<td>64.1</td>
<td>61.0</td>
<td>66</td>
<td>56.3</td>
<td>44.1</td>
<td>51.1</td>
<td>44 1</td>
<td>54.2</td>
<td>55.7</td>
<td>83.3</td>
<td>70.8</td>
</tr>
<tr>
<td>0.95</td>
<td>67.5</td>
<td>70.2</td>
<td>62</td>
<td>51.7</td>
<td>0.7 5:</td>
<td>49.8</td>
<td>49.8</td>
<td>60.6</td>
<td>61.7</td>
<td>65.5</td>
<td>57.3</td>
<td>45.7</td>
<td>53.1</td>
<td>1 45</td>
<td>55.2</td>
<td>61.5</td>
<td>88.1</td>
<td>66.5</td>
</tr>
</tbody>
</table>
Table 1. Effect of the detection cut-off threshold. We threshold the detections computed by Mask R-CNN before matching them to compute tracks. While keypoint mAP goes down, the tracking MOTA performance increases as there are fewer spurious detections to confuse the tracker. The first row also shows the random baseline; i.e. the performance of the model that randomly assigns a track ID between 0 and 1000 (maximum allowed) to each detection.
## $4. 1. \textbf{Dataset and Evaluation}$
results presented in this paper, and report final test numbers
as obtained from the evaluation server.
## $4. 2. \textbf{Baseline}$
PoseTrack[24,25] is a recently released large-scale challenge dataset for human body keypoint estimation and tracking in diverse, in-the-wild videos. It consists of a total of 514 video sequences with 66,374 frames, split into 300, 50 and 208 videos for training, validation and testing, respectively. The training videos come with the middle 30 frames densely labeled with human body keypoints. The validation and test videos are labeled at every fourth frame, apart from the middle 30 frames. This helps evaluate the long term tracking performance of methods without requiring expensive annotations throughout the entire video. In total, the dataset contains 23,000 labeled frames and 153,615 poses. The test set annotations are held-out, and evaluation are performed by submitting the predictions to an evaluation server.
The annotations consist of human head bounding boxes and 15 body joint keypoint locations per labeled person. Since all our proposed approaches are top-down and depend on the detection of the extent of the person before detecting keypoints, we compute a bounding box by taking the min and max extents of labeled keypoints, and dilating that box by 20%. Also, to make the dataset compatible with COCO[35,36], we permute the keypoint labels to match the closest equivalent labels in COCO. This allows us to pretrain our models on COCO, augmenting the PoseTrack dataset significantly and giving a large improvement in performance.
The dataset is designed to evaluate methods on three different tasks: 1) Single-frame pose estimation; 2) Pose estimation in video; 3) Pose tracking in the wild. Task 1) and 2) are evaluated at a frame level, using the mean average precision (mAP) metric [42]. Task 3) is evaluated using a multi-object tracking metric (MOT) [3]. Both evaluations require first computing the distance of each prediction from each ground truth labeled pose. This is done using the PCKh metric[2], which computes the probability of correct keypoints normalized by the head size. The mAP is computed as in [42], and the MOT is as described in [38]. Their MOT evaluation penalizes false positives equally regardless of their confidence. For this, we drop keypoint predictions with low confidence (1.95 after grid-search on the validation set). We use the PoseTrack evaluation toolkit for computing all
In an effort to build a very competitive baseline, we first evaluate the various design elements of our two stage tracking pipeline with a vanilla Mask R-CNN base model. This model disregards time-sensitive cues when making pose predictions. Throughout this section, our models are initialized from ImageNet and are pretrained on the COCO keypoint detection task. We then finetune the Mask R-CNN model on PoseTrack, keeping most hyper-parameters fixed to the defaults used for COCO[36]. At test time, we run the model on each frame and store the bounding box and keypoint predictions, which are linked over time in the tracking stage. This model is competitive as it achieves state of the art results on the PoseTrack dataset. In Sec. 4.3, we prove that our approach can further improve the performance by incorporating temporal context from each video clip via a 3D Mask R- CNN model.
Thresholding initial detections: Before linking the detections in time, we drop the low-confidence and potentially incorrect detections. This helps prevent the tracks from drifting and reduces false positives. Table 1 shows the effect of thresholding the detections. As expected, the MOTA tracking metric[3] improves with higher thresholds, indicating better and cleaner tracks. The keypoint mAP performance, however, decreases by missing out on certain low-confidence detections, which tend to improve the mAP metric. Since we primarily focus on the tracking task, we threshold our $\hat{\text{detections at 0.95 for our final experiments}}.$
Deeper base networks: As in most vision problems, we observed an improvement in frame-level pose estimation by using a deeper base model. The improvements also directly transferred to the tracking performance. Replacing ResNet-50 in Mask R-CNN with Resnet-101 gave us about $2\%$ improvement in MOTA. We also observed a gain in performance on using feature pyramid networks (FPN)[34] in the base network. Ultimately, we got best performance by using a ResNet-101 model with FPN as the body, a 2-layer MLP for the classification head, and a stack of eight conv and deconv layers as the keypoint head.
|
[
[
0.08416748046875,
0.0921630859375,
0.1422119140625,
0.0921630859375,
0.1422119140625,
0.106689453125,
0.08416748046875,
0.106689453125
],
[
0.1854248046875,
0.0921630859375,
0.8857421875,
0.0921630859375,
0.8857421875,
0.1060791015625,
0.1854248046875,
0.1060791015625
],
[
0.1846923828125,
0.10479736328125,
0.318603515625,
0.10479736328125,
0.318603515625,
0.11871337890625,
0.1846923828125,
0.11871337890625
],
[
0.313720703125,
0.10479736328125,
0.486083984375,
0.10479736328125,
0.486083984375,
0.11871337890625,
0.313720703125,
0.11871337890625
],
[
0.490234375,
0.1060791015625,
0.52685546875,
0.1060791015625,
0.52685546875,
0.117431640625,
0.490234375,
0.117431640625
],
[
0.5390625,
0.1060791015625,
0.564453125,
0.1060791015625,
0.564453125,
0.117431640625,
0.5390625,
0.117431640625
],
[
0.57666015625,
0.1060791015625,
0.60888671875,
0.1060791015625,
0.60888671875,
0.117431640625,
0.57666015625,
0.117431640625
],
[
0.62109375,
0.1029052734375,
0.65283203125,
0.106689453125,
0.650390625,
0.12060546875,
0.61767578125,
0.11614990234375
],
[
0.65869140625,
0.1060791015625,
0.6953125,
0.1060791015625,
0.6953125,
0.117431640625,
0.65869140625,
0.117431640625
],
[
0.70166015625,
0.1060791015625,
0.73779296875,
0.1060791015625,
0.73779296875,
0.117431640625,
0.70166015625,
0.117431640625
],
[
0.740234375,
0.1060791015625,
0.78125,
0.1060791015625,
0.78125,
0.117431640625,
0.740234375,
0.117431640625
],
[
0.78369140625,
0.1060791015625,
0.88818359375,
0.1060791015625,
0.88818359375,
0.11871337890625,
0.78369140625,
0.11871337890625
],
[
0.08416748046875,
0.121826171875,
0.21728515625,
0.121826171875,
0.21728515625,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.281982421875,
0.121826171875,
0.314453125,
0.121826171875,
0.314453125,
0.1337890625,
0.281982421875,
0.1337890625
],
[
0.34716796875,
0.121826171875,
0.3798828125,
0.121826171875,
0.3798828125,
0.1337890625,
0.34716796875,
0.1337890625
],
[
0.3798828125,
0.121826171875,
0.4150390625,
0.121826171875,
0.4150390625,
0.1337890625,
0.3798828125,
0.1337890625
],
[
0.496826171875,
0.121826171875,
0.52294921875,
0.121826171875,
0.52294921875,
0.1337890625,
0.496826171875,
0.1337890625
],
[
0.53759765625,
0.121826171875,
0.564453125,
0.121826171875,
0.564453125,
0.1337890625,
0.53759765625,
0.1337890625
],
[
0.57666015625,
0.121826171875,
0.609375,
0.121826171875,
0.609375,
0.1337890625,
0.57666015625,
0.1337890625
],
[
0.66015625,
0.121826171875,
0.69287109375,
0.121826171875,
0.69287109375,
0.1337890625,
0.66015625,
0.1337890625
],
[
0.2459716796875,
0.12249755859375,
0.281982421875,
0.12249755859375,
0.281982421875,
0.1337890625,
0.2459716796875,
0.1337890625
],
[
0.2132568359375,
0.12249755859375,
0.25,
0.12249755859375,
0.25,
0.1337890625,
0.2132568359375,
0.1337890625
],
[
0.313720703125,
0.12249755859375,
0.346435546875,
0.12249755859375,
0.346435546875,
0.1337890625,
0.313720703125,
0.1337890625
],
[
0.41259765625,
0.12249755859375,
0.4453125,
0.12249755859375,
0.4453125,
0.1337890625,
0.41259765625,
0.1337890625
],
[
0.451904296875,
0.12249755859375,
0.483642578125,
0.12249755859375,
0.483642578125,
0.1337890625,
0.451904296875,
0.1337890625
],
[
0.619140625,
0.12249755859375,
0.650390625,
0.12249755859375,
0.650390625,
0.1337890625,
0.619140625,
0.1337890625
],
[
0.705078125,
0.12249755859375,
0.73193359375,
0.12249755859375,
0.73193359375,
0.1337890625,
0.705078125,
0.1337890625
],
[
0.744140625,
0.12249755859375,
0.779296875,
0.12249755859375,
0.779296875,
0.1337890625,
0.744140625,
0.1337890625
],
[
0.7890625,
0.12249755859375,
0.81787109375,
0.12249755859375,
0.81787109375,
0.1337890625,
0.7890625,
0.1337890625
],
[
0.826171875,
0.12249755859375,
0.8837890625,
0.12249755859375,
0.8837890625,
0.1337890625,
0.826171875,
0.1337890625
],
[
0.08416748046875,
0.1337890625,
0.108642578125,
0.1337890625,
0.108642578125,
0.146484375,
0.08416748046875,
0.146484375
],
[
0.1871337890625,
0.1337890625,
0.281982421875,
0.1337890625,
0.281982421875,
0.1458740234375,
0.1871337890625,
0.1458740234375
],
[
0.281982421875,
0.1337890625,
0.313720703125,
0.1337890625,
0.313720703125,
0.1458740234375,
0.281982421875,
0.1458740234375
],
[
0.314453125,
0.1337890625,
0.3447265625,
0.1337890625,
0.3447265625,
0.1458740234375,
0.314453125,
0.1458740234375
],
[
0.348876953125,
0.1337890625,
0.378173828125,
0.1337890625,
0.378173828125,
0.1458740234375,
0.348876953125,
0.1458740234375
],
[
0.3798828125,
0.1337890625,
0.410888671875,
0.1337890625,
0.410888671875,
0.1458740234375,
0.3798828125,
0.1458740234375
],
[
0.414306640625,
0.1337890625,
0.443603515625,
0.1337890625,
0.443603515625,
0.1458740234375,
0.414306640625,
0.1458740234375
],
[
0.451904296875,
0.1337890625,
0.483642578125,
0.1337890625,
0.483642578125,
0.146484375,
0.451904296875,
0.146484375
],
[
0.494384765625,
0.1337890625,
0.52392578125,
0.1337890625,
0.52392578125,
0.1458740234375,
0.494384765625,
0.1458740234375
],
[
0.53515625,
0.1337890625,
0.564453125,
0.1337890625,
0.564453125,
0.146484375,
0.53515625,
0.146484375
],
[
0.57666015625,
0.1337890625,
0.60888671875,
0.1337890625,
0.60888671875,
0.146484375,
0.57666015625,
0.146484375
],
[
0.62109375,
0.1337890625,
0.64892578125,
0.1337890625,
0.64892578125,
0.1458740234375,
0.62109375,
0.1458740234375
],
[
0.66015625,
0.1337890625,
0.69140625,
0.1337890625,
0.69140625,
0.1458740234375,
0.66015625,
0.1458740234375
],
[
0.70166015625,
0.1337890625,
0.734375,
0.1337890625,
0.734375,
0.1458740234375,
0.70166015625,
0.1458740234375
],
[
0.74609375,
0.1337890625,
0.779296875,
0.1337890625,
0.779296875,
0.146484375,
0.74609375,
0.146484375
],
[
0.7890625,
0.1337890625,
0.81787109375,
0.1337890625,
0.81787109375,
0.1458740234375,
0.7890625,
0.1458740234375
],
[
0.826171875,
0.1337890625,
0.8857421875,
0.1337890625,
0.8857421875,
0.1458740234375,
0.826171875,
0.1458740234375
],
[
0.1854248046875,
0.1458740234375,
0.279296875,
0.1458740234375,
0.279296875,
0.1572265625,
0.1854248046875,
0.1572265625
],
[
0.62109375,
0.1458740234375,
0.64892578125,
0.1458740234375,
0.64892578125,
0.1578369140625,
0.62109375,
0.1578369140625
],
[
0.74609375,
0.14453125,
0.77685546875,
0.146484375,
0.775390625,
0.1590576171875,
0.744140625,
0.1572265625
],
[
0.313720703125,
0.146484375,
0.3447265625,
0.146484375,
0.3447265625,
0.1578369140625,
0.313720703125,
0.1578369140625
],
[
0.08416748046875,
0.146484375,
0.108642578125,
0.146484375,
0.108642578125,
0.1578369140625,
0.08416748046875,
0.1578369140625
],
[
0.3798828125,
0.146484375,
0.410888671875,
0.146484375,
0.410888671875,
0.1578369140625,
0.3798828125,
0.1578369140625
],
[
0.414306640625,
0.146484375,
0.44287109375,
0.146484375,
0.44287109375,
0.1578369140625,
0.414306640625,
0.1578369140625
],
[
0.451904296875,
0.146484375,
0.48193359375,
0.146484375,
0.48193359375,
0.1578369140625,
0.451904296875,
0.1578369140625
],
[
0.496826171875,
0.146484375,
0.521484375,
0.146484375,
0.521484375,
0.1590576171875,
0.496826171875,
0.1590576171875
],
[
0.53515625,
0.146484375,
0.564453125,
0.146484375,
0.564453125,
0.1578369140625,
0.53515625,
0.1578369140625
],
[
0.57861328125,
0.146484375,
0.60693359375,
0.146484375,
0.60693359375,
0.1590576171875,
0.57861328125,
0.1590576171875
],
[
0.66015625,
0.146484375,
0.69140625,
0.146484375,
0.69140625,
0.1578369140625,
0.66015625,
0.1578369140625
],
[
0.70361328125,
0.146484375,
0.73388671875,
0.146484375,
0.73388671875,
0.1578369140625,
0.70361328125,
0.1578369140625
],
[
0.826171875,
0.146484375,
0.8837890625,
0.146484375,
0.8837890625,
0.1578369140625,
0.826171875,
0.1578369140625
],
[
0.351318359375,
0.147705078125,
0.378173828125,
0.147705078125,
0.378173828125,
0.1558837890625,
0.351318359375,
0.1558837890625
],
[
0.281982421875,
0.147705078125,
0.31201171875,
0.147705078125,
0.31201171875,
0.1558837890625,
0.281982421875,
0.1558837890625
],
[
0.7890625,
0.147705078125,
0.81640625,
0.147705078125,
0.81640625,
0.1558837890625,
0.7890625,
0.1558837890625
],
[
0.0849609375,
0.1572265625,
0.11273193359375,
0.1572265625,
0.11273193359375,
0.169189453125,
0.0849609375,
0.169189453125
],
[
0.1871337890625,
0.1578369140625,
0.2484130859375,
0.1578369140625,
0.2484130859375,
0.169189453125,
0.1871337890625,
0.169189453125
],
[
0.2548828125,
0.1578369140625,
0.27783203125,
0.1578369140625,
0.27783203125,
0.169189453125,
0.2548828125,
0.169189453125
],
[
0.283447265625,
0.1578369140625,
0.31787109375,
0.1578369140625,
0.31787109375,
0.169189453125,
0.283447265625,
0.169189453125
],
[
0.31201171875,
0.1578369140625,
0.3447265625,
0.1578369140625,
0.3447265625,
0.169189453125,
0.31201171875,
0.169189453125
],
[
0.34716796875,
0.1578369140625,
0.38232421875,
0.1578369140625,
0.38232421875,
0.169189453125,
0.34716796875,
0.169189453125
],
[
0.37744140625,
0.1578369140625,
0.4150390625,
0.1578369140625,
0.4150390625,
0.169189453125,
0.37744140625,
0.169189453125
],
[
0.41015625,
0.1578369140625,
0.4453125,
0.1578369140625,
0.4453125,
0.169189453125,
0.41015625,
0.169189453125
],
[
0.450927734375,
0.1578369140625,
0.48046875,
0.1578369140625,
0.48046875,
0.17041015625,
0.450927734375,
0.17041015625
],
[
0.494384765625,
0.1578369140625,
0.52294921875,
0.1578369140625,
0.52294921875,
0.169189453125,
0.494384765625,
0.169189453125
],
[
0.53515625,
0.1578369140625,
0.56396484375,
0.1578369140625,
0.56396484375,
0.169189453125,
0.53515625,
0.169189453125
],
[
0.57861328125,
0.1578369140625,
0.60693359375,
0.1578369140625,
0.60693359375,
0.169189453125,
0.57861328125,
0.169189453125
],
[
0.62109375,
0.1578369140625,
0.64892578125,
0.1578369140625,
0.64892578125,
0.169189453125,
0.62109375,
0.169189453125
],
[
0.6611328125,
0.1558837890625,
0.689453125,
0.1578369140625,
0.68896484375,
0.17041015625,
0.66015625,
0.1685791015625
],
[
0.70361328125,
0.1578369140625,
0.73388671875,
0.1578369140625,
0.73388671875,
0.169189453125,
0.70361328125,
0.169189453125
],
[
0.74609375,
0.1578369140625,
0.77783203125,
0.1578369140625,
0.77783203125,
0.17041015625,
0.74609375,
0.17041015625
],
[
0.78759765625,
0.1578369140625,
0.81787109375,
0.1578369140625,
0.81787109375,
0.169189453125,
0.78759765625,
0.169189453125
],
[
0.826171875,
0.1578369140625,
0.8837890625,
0.1578369140625,
0.8837890625,
0.169189453125,
0.826171875,
0.169189453125
],
[
0.07843017578125,
0.1849365234375,
0.89208984375,
0.1849365234375,
0.89208984375,
0.1988525390625,
0.07843017578125,
0.1988525390625
],
[
0.07843017578125,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.07843017578125,
0.212158203125
],
[
0.07843017578125,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.07843017578125,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.42724609375,
0.2279052734375,
0.42724609375,
0.238037109375,
0.080078125,
0.238037109375
],
[
0.080078125,
0.26318359375,
0.2900390625,
0.26318359375,
0.2900390625,
0.27783203125,
0.080078125,
0.27783203125
],
[
0.5009765625,
0.26513671875,
0.89208984375,
0.26513671875,
0.89208984375,
0.279052734375,
0.5009765625,
0.279052734375
],
[
0.50244140625,
0.281005859375,
0.75732421875,
0.281005859375,
0.75732421875,
0.294189453125,
0.50244140625,
0.294189453125
],
[
0.098876953125,
0.291748046875,
0.4697265625,
0.291748046875,
0.4697265625,
0.304931640625,
0.098876953125,
0.304931640625
],
[
0.08087158203125,
0.30810546875,
0.468017578125,
0.30810546875,
0.468017578125,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.50244140625,
0.3037109375,
0.60107421875,
0.3037109375,
0.60107421875,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.07843017578125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.335205078125,
0.07843017578125,
0.335205078125
],
[
0.521484375,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.341552734375,
0.521484375,
0.341552734375
],
[
0.08087158203125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50244140625,
0.343505859375,
0.89208984375,
0.343505859375,
0.89208984375,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.07843017578125,
0.35107421875,
0.4697265625,
0.3515625,
0.4697265625,
0.3662109375,
0.07843017578125,
0.365478515625
],
[
0.5009765625,
0.35791015625,
0.89208984375,
0.356689453125,
0.89208984375,
0.371826171875,
0.5009765625,
0.37255859375
],
[
0.080078125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.080078125,
0.380615234375
],
[
0.50244140625,
0.373779296875,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.50244140625,
0.386962890625
],
[
0.080078125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.396484375,
0.080078125,
0.396484375
],
[
0.50244140625,
0.389404296875,
0.89208984375,
0.389404296875,
0.89208984375,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.080078125,
0.397705078125,
0.4697265625,
0.397705078125,
0.4697265625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.404052734375,
0.89208984375,
0.404052734375,
0.89208984375,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.080078125,
0.41357421875,
0.4697265625,
0.41357421875,
0.4697265625,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.41845703125,
0.89208984375,
0.41845703125,
0.89208984375,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.080078125,
0.427978515625,
0.4697265625,
0.427978515625,
0.4697265625,
0.44140625,
0.080078125,
0.44140625
],
[
0.50244140625,
0.434326171875,
0.89208984375,
0.434326171875,
0.89208984375,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.08087158203125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.455078125,
0.08087158203125,
0.455078125
],
[
0.50341796875,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.4609375,
0.50341796875,
0.4609375
],
[
0.07843017578125,
0.45703125,
0.4697265625,
0.45703125,
0.4697265625,
0.47021484375,
0.07843017578125,
0.47021484375
],
[
0.5009765625,
0.46337890625,
0.89208984375,
0.464111328125,
0.89208984375,
0.479248046875,
0.5009765625,
0.477783203125
],
[
0.080078125,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.5009765625,
0.477783203125,
0.8935546875,
0.479736328125,
0.8935546875,
0.49365234375,
0.5009765625,
0.4912109375
],
[
0.080078125,
0.488037109375,
0.428955078125,
0.488037109375,
0.428955078125,
0.501953125,
0.080078125,
0.501953125
],
[
0.50341796875,
0.49560546875,
0.890625,
0.49560546875,
0.890625,
0.505859375,
0.50341796875,
0.505859375
],
[
0.09722900390625,
0.505859375,
0.468017578125,
0.505859375,
0.468017578125,
0.51953125,
0.09722900390625,
0.51953125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.07843017578125,
0.51953125,
0.471435546875,
0.52099609375,
0.471435546875,
0.53515625,
0.07843017578125,
0.5341796875
],
[
0.50244140625,
0.5244140625,
0.89208984375,
0.5244140625,
0.89208984375,
0.53857421875,
0.50244140625,
0.53857421875
],
[
0.07843017578125,
0.53515625,
0.4697265625,
0.5361328125,
0.4697265625,
0.55029296875,
0.07843017578125,
0.5498046875
],
[
0.5009765625,
0.5390625,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.5009765625,
0.5537109375
],
[
0.080078125,
0.5517578125,
0.4697265625,
0.5517578125,
0.4697265625,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.50341796875,
0.55615234375,
0.60546875,
0.55615234375,
0.60546875,
0.56640625,
0.50341796875,
0.56640625
],
[
0.080078125,
0.5673828125,
0.468017578125,
0.5673828125,
0.468017578125,
0.58056640625,
0.080078125,
0.58056640625
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.080078125,
0.58203125,
0.4697265625,
0.58203125,
0.4697265625,
0.59521484375,
0.080078125,
0.59521484375
],
[
0.50244140625,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.59765625,
0.468017578125,
0.59765625,
0.468017578125,
0.611328125,
0.080078125,
0.611328125
],
[
0.50244140625,
0.60107421875,
0.89208984375,
0.60107421875,
0.89208984375,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.6123046875,
0.4697265625,
0.6123046875,
0.4697265625,
0.62548828125,
0.080078125,
0.62548828125
],
[
0.5009765625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.080078125,
0.626953125,
0.4697265625,
0.626953125,
0.4697265625,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.080078125,
0.6416015625,
0.471435546875,
0.6416015625,
0.471435546875,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.50244140625,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.080078125,
0.6572265625,
0.4697265625,
0.6572265625,
0.4697265625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.09722900390625,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.09722900390625,
0.68798828125
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.70556640625,
0.471435546875,
0.70556640625,
0.471435546875,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.0767822265625,
0.7197265625,
0.468017578125,
0.7216796875,
0.468017578125,
0.73486328125,
0.0767822265625,
0.73291015625
],
[
0.50244140625,
0.7216796875,
0.79150390625,
0.7216796875,
0.79150390625,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.080078125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.5009765625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.07843017578125,
0.796875,
0.47216796875,
0.796875,
0.47216796875,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.76318359375,
0.88720703125,
0.76318359375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Threshold",
"mAPmAPmAPmAPmAPmAPmAPmAPMOTAMOTAMOTAMOTAMOTAMOTAMOTAMOTAMOTP PrecRec",
"Head Shou Elbo Wri",
"Hip Knee Ankl TotalHead",
"Shou",
"Elb",
"Wri",
"Hip",
"Knee",
"Ankl",
"Total",
"Total Total Total",
"0.0, random tracks 72.8",
"54.3",
"60.9",
"51.8",
"-6.6",
"-8.5",
"-12.9",
"-10.2",
"65.3",
"75.6",
"63.5",
"64.1",
"-11.6",
"-11.1",
"-9.7",
"-10.2",
"55.8",
"83.370.8",
"0.0",
"72.875.6 65.3",
"54.3",
"63.5",
"60.9",
"51.8",
"64.1",
"60.3",
"65.3",
"55.8",
"43.5",
"52.5",
"50.7",
"43.9",
"53.6",
"55.7",
"83.370.8",
"72.875.6 65.3",
"52.9",
"54.2",
"63.5",
"0.5",
"51.8",
"64.1",
"61.0",
"66",
"56.3",
"44.1",
"51.1",
"44.3",
"83.370.8",
"60.9",
"54.3",
"55.7",
"0.95",
"67.570.2",
"62",
"51.7",
"60.7",
"58.7",
"49.8",
"60.6",
"61.7",
"65.5",
"57.3",
"45.7",
"54.3",
"53.1",
"45.7",
"55.2",
"61.5",
"88.166.5",
"Table 1. Effect of the detection cut-off threshold. We threshold the detections computed by Mask R-CNN before matching them to",
"compute tracks. While keypoint mAP goes down, the tracking MOTA performance increases as there are fewer spurious detections to",
"confuse the tracker. The first row also shows the random baseline; i.e. the performance of the model that randomly assigns a track ID",
"between0and1000(maximumallowed)toeachdetection",
"4.1.Dataset and Evaluation",
"results presented in this paper, and report final test numbers",
"as obtained from the evaluation server.",
"PoseTrack [24, 25] is a recently released large-scale chal-",
"lengedatasetforhumanbodykeypointestimationandtrack",
"4.2. Baseline",
"ing in diverse, in-the-wild videos. It consists of a total of 514",
"In an effort to build a very competitive baseline, we first",
"videosequenceswith66.374frames,splitinto300.50and",
"evaluate the various design elements of our two stage track-",
"208 videos for training, validation and testing, respectively.",
"ing pipeline with a vanilla Mask R-CNN base model. This",
"The training videos come with the middle 30 frames densely",
"model disregards time-sensitive cues when making pose pre-",
"labeled with human body keypoints. The validation and",
"dictions. Throughout this section, our models are initialized",
"test videos are labeled at every fourth frame, apart from the",
"from ImageNet and are pretrained on the COCO keypoint",
"middle 30 frames. This helps evaluate the long term track-",
"detection task. We then finetune the Mask R-CNN model",
"ing performance of methods without requiring expensive",
"on PoseTrack, keeping most hyper-parameters fixed to the",
"annotations throughoutthe entirevideo.In total, the dataset",
"defaultsusedforCOCO[36].Attesttime,werunthemodel",
"contains 23,000 labeled frames and 153,615 poses. The test",
"on each frame and store the bounding box and keypoint pre-",
"set annotations are held-out, and evaluation are performed",
"dictions, which are linked over time in the tracking stage.",
"by submitting the predictions to an evaluation server.",
"This model is competitive as it achieves state of the art re-",
"The annotations consist of human head bounding boxes",
"sults on the PoseTrack dataset. In Sec. 4.3, we prove that our",
"and 15 body joint keypoint locations per labeled person.",
"approach can further improve the performance by incorpo-",
"Since all our proposed approaches are top-down and depend",
"rating temporal context from each video clip via a 3D Mask",
"on the detection of the extent of the person before detect-",
"R-CNN model.",
"ing keypoints, we compute a bounding box by taking the",
"Thresholding initial detections: Before linking the detec-",
"min and max extents of labeled keypoints, and dilating that",
"tions in time, we drop the low-confidence and potentially",
"box by 20%. Also, to make the dataset compatible with",
"incorrect detections. This helps prevent the tracks from drift-",
"COCO [35, 36], we permute the keypoint labels to match the",
"ing and reduces false positives. Table 1 shows the effect of",
"closest equivalent labels in COCO. This allows us to pretrain",
"thresholding the detections. As expected, the MOTA track-",
"our models on COCO, augmenting the PoseTrack dataset",
"ing metric [3] improves with higher thresholds, indicating",
"significantly and giving a large improvement in performance",
"better and cleaner tracks. The keypoint mAP performance,",
"The dataset is designed to evaluate methods on three",
"however,decreasesbymissingoutoncertainlow-confidence",
"different tasks: 1) Single-frame pose estimation; 2) Pose",
"detections, which tend to improve the mAP metric. Since",
"estimation in video; 3) Pose tracking in the wild. Task 1)",
"we primarily focus on the tracking task, we threshold our",
" and 2) are evaluated at a frame level, using the mean average",
"detections at 0.95 for our final experiments.",
"precision (mAP) metric [42]. Task 3) is evaluated using a",
"Deeper base networks: As in most vision problems, we",
"multi-object tracking metric (MOT) [3]. Both evaluations",
"observedanimprovementinframe-levelposeestimation",
"require first computing the distance of each prediction from",
"by using a deeper base model. The improvements also di.",
"eachgroundtruthlabeledpose.This isdoneusingthePCKh",
"rectly transferred to the tracking performance. Replacing",
"metric [2], which computes the probability of correct key-",
"ResNet-50 in Mask R-CNN with Resnet-101 gave us about",
"points normalized by the head size. The mAP is computed",
"2% improvement in MOTA. We also observed a gain in per-",
"as in [42], and the MOT is as described in [38]. Their MOT",
"formance on using feature pyramid networks (FPN) [34] in",
"evaluation penalizes false positives equally regardless of",
"the base network. Ultimately, we got best performance by",
"their confidence. For this, we drop keypoint predictions with",
"using a ResNet-101 model with FPN as the body, a 2-layer",
"low confidence (1.95 after grid-search on the validation set).",
"MLP for the classification head, and a stack of eight conv",
"We use the PoseTrack evaluation toolkit for computing all",
"and deconv layers as the keypoint head."
] |
{
"arxiv_id": "1712.09184",
"language": "en",
"page_idx": 4
}
|
||
1906.03518-en-0
|
1906.03518-en-0.jpg
|
$\textbf{Maximum Weighted Loss Discrepancy}$
Fereshte Khani, Aditi Raghunathan, Percy Liang Department of Computer Science, Stanford University
$\{$fereshte,aditir$\}$@stanford.edu, pliang@cs.stanford.edu
### $\mathbf{Abstract}$
Though machine learning algorithms excel at minimizing the average loss over a population, this might lead to large discrepancies between the losses across groups within the population. To capture this inequality, we introduce and study a notion we call maximum weighted loss discrepancy (MWLD), the maximum (weighted) difference between the loss of a group and the loss of the population. We relate MWLD to group fairness notions and robustness to demographic shifts. We then show MWLD satisfes the following three properties: 1) It is statistically impossible to estimate MWLD when all groups have equal weights. 2) For a particular family of weighting functions, we can estimate MWLD efficiently. 3) $\hat{\mathbf{M}}$WLD is related to loss variance, a quantity that arises in generalization bounds We estimate MWLD with different weighting functions on four common datasets from the fairness literature. We finally show that loss variance regularization can halve the loss variance of a classifier and hence reduce MWLD without suffering a significant drop in accuracy.
## 1 Introduction
Machine learning algorithms have a profound effect on people, especially around critical decisions such as banking and criminal justice [7,4]. It has been shown that standard learning procedures (empirical risk minimization) can result in classiffers where some demographic groups suffer signiffcantly larger losses than the average population [2,8]. In this work, we consider the setting where demographic information is unavailable[14], so we would like to ensure that no group suffers a loss much larger than average.
We are interested in measuring the maximum weighted loss discrepancy (MWLD) of a model, which is, over all groups $g$,the maximum difference between the loss of a group $\mathbb{E}[\ell\mid g=1]$ and the population loss $\mathbb{E}[\ell]$,weighted by a function $w$ that quantifies the importance of each group:
$$\mathrm{MWLD}(w)=\max_{g}w(g)\left|\mathbb{E}[\ell\mid g=1]-\mathbb{E}[\ell]\right|.$$
(1)
MWLD captures various notions of group fairness; for example, equal opportunity [13] is MWLD with $\ell$ capturing false positives and weighting function $w( g) = 1$ for sensitive groups( e. g. , defined by race, gender) and 0 for all other groups. We also show that we can bound the loss of a population with shifted demographics: if we tilt the original distribution toward any group $g$ based on $w(g)$, the loss on the new distribution can bounded using MWLD$(w).$
We consider estimating MWLD from finite data by plugging in the empirical distribution. There are two considerations: (i) does the estimator converge? and (ii) can we compute the estimator efficiently? The answers to these two questions depend on the weighting function. We first show that for the uniform weighting function $(w^0(g)=\mathbb{I}[\mathbb{E}[g]>0])$, we cannot estimate MWLD$(w^0)$ from finite samples (Proposition 1). Next, we study a family of decaying weighting functions $(w^k(g)=\mathbb{E}[g]^k)$, where $k$ governs how much we account for the loss discrepancy of small groups. For this family, we show that the plug-in estimator (i) is efficient to compute and (ii) converges to the population MWLD$( w^{k})$ (Theorem 1).
Preprint. Under review.
|
[
[
0.26708984375,
0.125,
0.72802734375,
0.126953125,
0.72802734375,
0.147705078125,
0.26708984375,
0.1458740234375
],
[
0.326904296875,
0.20263671875,
0.6708984375,
0.20263671875,
0.6708984375,
0.216552734375,
0.326904296875,
0.216552734375
],
[
0.318603515625,
0.216552734375,
0.67626953125,
0.21533203125,
0.67626953125,
0.2298583984375,
0.318603515625,
0.2310791015625
],
[
0.268798828125,
0.2298583984375,
0.73193359375,
0.2298583984375,
0.73193359375,
0.2442626953125,
0.268798828125,
0.2442626953125
],
[
0.0228729248046875,
0.271484375,
0.060455322265625,
0.271484375,
0.060455322265625,
0.396484375,
0.0228729248046875,
0.396484375
],
[
0.461669921875,
0.279052734375,
0.5390625,
0.279052734375,
0.5390625,
0.296630859375,
0.461669921875,
0.296630859375
],
[
0.233642578125,
0.31005859375,
0.765625,
0.31005859375,
0.765625,
0.323974609375,
0.233642578125,
0.323974609375
],
[
0.2320556640625,
0.324462890625,
0.765625,
0.323974609375,
0.765625,
0.337158203125,
0.2320556640625,
0.33837890625
],
[
0.233642578125,
0.33837890625,
0.7646484375,
0.33837890625,
0.7646484375,
0.3515625,
0.233642578125,
0.3515625
],
[
0.233642578125,
0.3515625,
0.765625,
0.3515625,
0.765625,
0.365478515625,
0.233642578125,
0.365478515625
],
[
0.233642578125,
0.3662109375,
0.765625,
0.3662109375,
0.765625,
0.380126953125,
0.233642578125,
0.380126953125
],
[
0.236083984375,
0.380615234375,
0.765625,
0.380615234375,
0.765625,
0.391357421875,
0.236083984375,
0.391357421875
],
[
0.0196075439453125,
0.395751953125,
0.05963134765625,
0.39453125,
0.062103271484375,
0.48291015625,
0.0228729248046875,
0.484130859375
],
[
0.233642578125,
0.393310546875,
0.7646484375,
0.393310546875,
0.7646484375,
0.4072265625,
0.233642578125,
0.4072265625
],
[
0.2320556640625,
0.40478515625,
0.765625,
0.406005859375,
0.765625,
0.42041015625,
0.2320556640625,
0.419189453125
],
[
0.2344970703125,
0.42041015625,
0.76708984375,
0.42041015625,
0.76708984375,
0.434326171875,
0.2344970703125,
0.434326171875
],
[
0.2320556640625,
0.43310546875,
0.765625,
0.434326171875,
0.765625,
0.448974609375,
0.2320556640625,
0.447509765625
],
[
0.233642578125,
0.447509765625,
0.765625,
0.447509765625,
0.765625,
0.4609375,
0.233642578125,
0.4609375
],
[
0.2344970703125,
0.462158203125,
0.765625,
0.462158203125,
0.765625,
0.47607421875,
0.2344970703125,
0.47607421875
],
[
0.233642578125,
0.47607421875,
0.76708984375,
0.47607421875,
0.76708984375,
0.4892578125,
0.233642578125,
0.4892578125
],
[
0.021240234375,
0.4912109375,
0.058013916015625,
0.4912109375,
0.058013916015625,
0.70556640625,
0.021240234375,
0.70556640625
],
[
0.233642578125,
0.488037109375,
0.418212890625,
0.4892578125,
0.418212890625,
0.50390625,
0.233642578125,
0.50244140625
],
[
0.1732177734375,
0.5263671875,
0.31201171875,
0.5263671875,
0.31201171875,
0.54150390625,
0.1732177734375,
0.54150390625
],
[
0.1732177734375,
0.556640625,
0.82421875,
0.556640625,
0.82421875,
0.57080078125,
0.1732177734375,
0.57080078125
],
[
0.1748046875,
0.57080078125,
0.82421875,
0.57080078125,
0.82421875,
0.583984375,
0.1748046875,
0.583984375
],
[
0.1748046875,
0.58544921875,
0.82666015625,
0.58544921875,
0.82666015625,
0.5986328125,
0.1748046875,
0.5986328125
],
[
0.1748046875,
0.59765625,
0.82421875,
0.59765625,
0.82421875,
0.6123046875,
0.1748046875,
0.6123046875
],
[
0.1748046875,
0.61279296875,
0.82421875,
0.61279296875,
0.82421875,
0.626953125,
0.1748046875,
0.626953125
],
[
0.1748046875,
0.626953125,
0.343017578125,
0.626953125,
0.343017578125,
0.64013671875,
0.1748046875,
0.64013671875
],
[
0.17236328125,
0.64453125,
0.82421875,
0.64306640625,
0.82421875,
0.6611328125,
0.17236328125,
0.662109375
],
[
0.1732177734375,
0.66015625,
0.82421875,
0.66015625,
0.82421875,
0.673828125,
0.1732177734375,
0.673828125
],
[
0.1748046875,
0.6748046875,
0.77392578125,
0.6748046875,
0.77392578125,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.34228515625,
0.68994140625,
0.654296875,
0.69140625,
0.654296875,
0.70556640625,
0.34228515625,
0.705078125
],
[
0.80224609375,
0.6923828125,
0.826171875,
0.6923828125,
0.826171875,
0.70556640625,
0.80224609375,
0.70556640625
],
[
0.1748046875,
0.71728515625,
0.82275390625,
0.71728515625,
0.82275390625,
0.73095703125,
0.1748046875,
0.73095703125
],
[
0.176513671875,
0.73193359375,
0.826171875,
0.73193359375,
0.826171875,
0.74560546875,
0.176513671875,
0.74560546875
],
[
0.1732177734375,
0.744140625,
0.82275390625,
0.744140625,
0.82275390625,
0.75830078125,
0.1732177734375,
0.75830078125
],
[
0.1732177734375,
0.75830078125,
0.82421875,
0.7587890625,
0.82421875,
0.7734375,
0.1732177734375,
0.77294921875
],
[
0.1732177734375,
0.77294921875,
0.56787109375,
0.77294921875,
0.56787109375,
0.7861328125,
0.1732177734375,
0.7861328125
],
[
0.1748046875,
0.79248046875,
0.82421875,
0.79248046875,
0.82421875,
0.806640625,
0.1748046875,
0.806640625
],
[
0.1732177734375,
0.80810546875,
0.82666015625,
0.80810546875,
0.82666015625,
0.8212890625,
0.1732177734375,
0.8212890625
],
[
0.1748046875,
0.8212890625,
0.82421875,
0.8212890625,
0.82421875,
0.83544921875,
0.1748046875,
0.83544921875
],
[
0.1748046875,
0.83544921875,
0.82421875,
0.83544921875,
0.82421875,
0.8486328125,
0.1748046875,
0.8486328125
],
[
0.1748046875,
0.849609375,
0.826171875,
0.849609375,
0.826171875,
0.86279296875,
0.1748046875,
0.86279296875
],
[
0.1748046875,
0.86279296875,
0.826171875,
0.86279296875,
0.826171875,
0.87744140625,
0.1748046875,
0.87744140625
],
[
0.1748046875,
0.87744140625,
0.82421875,
0.87744140625,
0.82421875,
0.8916015625,
0.1748046875,
0.8916015625
],
[
0.1748046875,
0.8916015625,
0.351318359375,
0.8916015625,
0.351318359375,
0.90478515625,
0.1748046875,
0.90478515625
],
[
0.1748046875,
0.92529296875,
0.310546875,
0.92529296875,
0.310546875,
0.93603515625,
0.1748046875,
0.93603515625
]
] |
[
"Maximum Weighted Loss Discrepancy",
"Fereshte Khani, Aditi Raghunathan, Percy Liang",
"Department of Computer Science, Stanford University",
"{fereshte,aditirj@stanford.edu,pliang@cs.stanford.edu",
"610z un 8",
"Abstract",
"Though machine learning algorithms excel at minimizing the average loss over",
"a population, this might lead to largediscrepancies between the losses across",
"groups within the population. To capture this inequality, we introduce and study",
"a notion we call maximum weighted loss discrepancy (MWLD), the maximum",
"(weighted) difference between the loss of a group and the loss of the population",
"WerelateMWLDtogroupfairnessnotionsandrobustnesstodemographicshifts",
"[OT's]",
"We then show MWLD satisfies the following three properties: 1) It is statistically",
"impossible to estimate MWLD when all groups have equal weights. 2) For a",
"particular family of weighting functions, we can estimate MWLD efficiently. 3)",
"MWLD is related to loss variance, a quantity that arises in generalization bounds",
"We estimate MWLD with different weighting functions on four common datasets",
"from the fairness literature. We finally show that loss variance regularization can",
"halve the loss variance of a classifier and hence reduce MWLD without suffering a",
"1A81S609061AIX",
"significant drop in accuracy.",
"1Introduction",
"Machine learning algorithms have a profound effect on people, especially around critical decisions",
"such as banking and criminal justice [7, 4]. It has been shown that standard learning procedures",
"(empirical risk minimization) can result in classifiers where some demographic groups suffer signifi-",
"cantly larger losses than the average population [2, 8]. In this work, we consider the setting where",
"demographic information is unavailable [14], so we would like to ensure that no group suffers a loss",
"much larger than average.",
"We are interested in measuring the maximum weighted loss discrepancy (MWLD) of a model, which",
"is, over all groups g, the maximum difference between the loss of a group E[e I g = 1] and the",
"population loss E[e], weighted by a function w that quantifies the importance of each group?",
"MWLD(w) = max w(g) [E[e |g = 1] - E[]]:",
"(1)",
"MWLD captures various notions of group fairness; for example, equal opportunity [13] is MWLD",
"with l capturing false positives and weighting function w(g) = 1 for sensitive groups (e.g., defined",
"by race, gender) and O for all other groups. We also show that we can bound the loss of a population",
"with shifted demographics: if we tilt the original distribution toward any group g based on w(g), the",
"loss on the new distribution can bounded using MWLD(w).",
"We consider estimating MWLD from finite data by plugging in the empirical distribution. There are",
"two considerations: (i) does the estimator converge? and (i) can we compute the estimator efficiently?",
"The answers to these two questions depend on the weighting function. We first show that for the",
"uniform weighting function (w(g)= I[E[g] > O]), we cannot estimate MWLD(w) from finite",
"samples (Proposition 1). Next, we study a family of decaying weighting functions (w*(g) =E[g]*),",
"where k governs how much we account for the loss discrepancy of small groups. For this family,",
"we show that the plug-in estimator (i) is efficient to compute and (i) converges to the population",
"MWLD(wk) (Theorem 1).",
"Preprint.Underreview"
] |
{
"arxiv_id": "1906.03518",
"language": "en",
"page_idx": 0
}
|
||
1906.03518-en-1
|
1906.03518-en-1.jpg
|
Next, we show a connection to loss variance (Proposition 2), an important quantity that arises in generalization bounds [20] and is used as a regularization scheme[21,3,23,22]. In particular, MWLD$( w^{1/ 2})$ provides us with lower and upper bounds for the loss variance. We also propose an extension called coarse loss variance, which considers only a set of sensitive groups, allowing us to incorporate knowledge about sensitive attributes.
We validate maximum weighted loss discrepancy on four common datasets from the fairness literature: predicting recidivism, credit rating, income, and crime rate. We fit a logistic regression on these $\hat{\text{datasets and estimate its MWLD}}(w^{\tilde{k}})$ for various $k.$ We observe that MWLÕ with smaller $k$ converges more slowly to the population quantity. We also observe the group attaining MWLD shrinks as $k$ decreases. We then use loss variance (LV) and coarse loss variance (CLV) regularization to train models. Our empirical fındings are as follows: 1) We halve the loss variance with only a small increase in the average loss. $\tilde{2})$ In some cases, using loss variance as a regularizer simultaneously reduces the classification loss (higher accuracy) and loss variance (lower loss discrepancy).
Setup. Consider the prediction task of mapping each input $x\in\mathcal{X}$ to a probability distribution over an output space $\mathcal{Y}.$ Let $h$ be a predictor which maps each input $x$ to a probability distribution over $\mathcal{Y};$ for binary classification, $\mathcal{Y}=\{0,1\}$ and a predictor $h:\mathcal{X}\to[0,1]$ returns the probability of $y=1.$ Let $\ell ( h, z) \textbf{be the ( bounded) loss incurred by predictor }h$ on individual $z—$e.g., the zero-one or logistic loss. Let $p^\star$ denote the underlying distribution over individuals $z=(x,y);$ all expectations are computed with respect to $z\sim p^*.$ Define a group to be a measurable function $g:\mathcal{X}\times\hat{\mathcal{Y}}\to\{0,1\}$ such that $g(z)=1$ if individual $z$ is in the group and 0 otherwise. Let $\mathcal{G}$ be the set of all groups. When clear from context, we use $\ell$ to denote $\ell(h,z)$ and $g$ to denote $g(z)$,so that $\mathbb{E}[\ell]$ is the population loss and $\mathbb{E}[\ell\mid g=1]$ is the loss of group $g.$
2 Maximum Weighted Loss Discrepancy
We now introduce our central object of study:
Definition 1 (Maximum weighted loss discrepancy (MWLD)). For a weighting function $w:\mathcal{G}\to$ $[ 0, 1] , \textit{loss function l, and predictor h, define the maximum weighted loss discrepancy MWLD}( w, \ell , h)$ to be the maximum difference between the loss of a group g and the population loss, weighted by $w(g)\colon$
(2)
$$MWLD\left(w,\ell,h\right)\stackrel{\mathrm{def}}{=}\sup_{g\in\mathcal{G}}w(g)\left|\mathbb{E}[\ell\mid g=1]-\mathbb{E}[\ell]\right|,$$
where the weighting function $w$ (e.g., $w(g)=\mathbb{E}[g]^1/2)$ intuitively controls the importance of group $g.$ Group fairness interpretation. By rearranging the terms of (2),we can bound the loss discrepancy of any group in terms of the group weight and the maximum weighted loss discrepancy:
$$|\mathbb{E}[\ell\mid g=1]-\mathbb{E}[\ell]|\leq\frac{\mathrm{MWLD}(w,\tilde{\ell},h)}{w(g)},$$
(3)
where the bound is tighter for larger $w(g).$ Existing statistical notions of fairness such as equal opportunity [13] can be viewed as enforcing MWLD to be small for a weighting function $w$ that is 1 on sensitive groups (e.g., different races) and 0 on all other groups; see Appendix A for further discussion.
Distributional shift interpretation. We can use MWLD to bound the loss on a population with shifted demographics. For any group $g$, define the mixture distribution $q(\cdot)\stackrel{\mathrm{def}}{=}w(g)p^{\star}(\cdot\mid g=$ $1)+(1-w(g))\hat{p}^{\star}(\cdot\mid g=0)$,which tilts the original distribution $p^{\star}$ more towards group $g$ (assuming $w( g) \geq \mathbb{E} [ g] ) . \textbf{ Then via simple algebra ( Proposition 3 in Appendix B) , the loss under this new}$ distribution $q$ can be controlled as follows:
$$\mathbb{E}_{z\sim q}[\ell]\leq\mathbb{E}_{z\sim p^{\star}}[\ell]+\mathrm{MWLD}(w,\ell,h).$$
(4)
This is similar in spirit to distributionally robust optimization (DRO) using a max-norm metric[10],
but the difference is that the mixture coefficient is group-dependent
How do we now operationalize MWLD? After all, the supremum over all groups $g$ in MWLD (2) appears daunting. In Section 3, we show how we can efficiently estimate MWLD for a restricted family of weighting functions.
2
|
[
[
0.1748046875,
0.0921630859375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.1060791015625,
0.1748046875,
0.1060791015625
],
[
0.176513671875,
0.106689453125,
0.82666015625,
0.106689453125,
0.82666015625,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.1748046875,
0.121826171875,
0.82421875,
0.121826171875,
0.82421875,
0.1351318359375,
0.1748046875,
0.1351318359375
],
[
0.1732177734375,
0.1337890625,
0.82275390625,
0.1351318359375,
0.82275390625,
0.149658203125,
0.1732177734375,
0.1483154296875
],
[
0.1748046875,
0.149658203125,
0.494384765625,
0.149658203125,
0.494384765625,
0.1634521484375,
0.1748046875,
0.1634521484375
],
[
0.1748046875,
0.169189453125,
0.82666015625,
0.169189453125,
0.82666015625,
0.18310546875,
0.1748046875,
0.18310546875
],
[
0.1748046875,
0.184326171875,
0.82421875,
0.184326171875,
0.82421875,
0.1976318359375,
0.1748046875,
0.1976318359375
],
[
0.1732177734375,
0.1956787109375,
0.82275390625,
0.1962890625,
0.82275390625,
0.2108154296875,
0.1732177734375,
0.210205078125
],
[
0.1748046875,
0.2108154296875,
0.82275390625,
0.2108154296875,
0.82275390625,
0.2247314453125,
0.1748046875,
0.2247314453125
],
[
0.1748046875,
0.2247314453125,
0.82177734375,
0.2247314453125,
0.82177734375,
0.238037109375,
0.1748046875,
0.238037109375
],
[
0.1748046875,
0.238037109375,
0.82421875,
0.238037109375,
0.82421875,
0.251953125,
0.1748046875,
0.251953125
],
[
0.1748046875,
0.25244140625,
0.82421875,
0.25244140625,
0.82421875,
0.266357421875,
0.1748046875,
0.266357421875
],
[
0.1707763671875,
0.26513671875,
0.76708984375,
0.26513671875,
0.76708984375,
0.279052734375,
0.1707763671875,
0.279052734375
],
[
0.1732177734375,
0.29345703125,
0.82421875,
0.29345703125,
0.82421875,
0.306884765625,
0.1732177734375,
0.306884765625
],
[
0.1748046875,
0.30810546875,
0.82275390625,
0.30810546875,
0.82275390625,
0.3212890625,
0.1748046875,
0.3212890625
],
[
0.1732177734375,
0.3212890625,
0.826171875,
0.3212890625,
0.826171875,
0.335205078125,
0.1732177734375,
0.335205078125
],
[
0.1748046875,
0.3359375,
0.82275390625,
0.3359375,
0.82275390625,
0.349853515625,
0.1748046875,
0.349853515625
],
[
0.1748046875,
0.349853515625,
0.66650390625,
0.349853515625,
0.66650390625,
0.363037109375,
0.1748046875,
0.363037109375
],
[
0.65673828125,
0.35107421875,
0.82275390625,
0.348388671875,
0.82275390625,
0.3623046875,
0.65673828125,
0.3642578125
],
[
0.1732177734375,
0.363037109375,
0.82177734375,
0.363037109375,
0.82177734375,
0.376953125,
0.1732177734375,
0.376953125
],
[
0.1732177734375,
0.376953125,
0.82421875,
0.376953125,
0.82421875,
0.39013671875,
0.1732177734375,
0.39013671875
],
[
0.1748046875,
0.391357421875,
0.826171875,
0.391357421875,
0.826171875,
0.40478515625,
0.1748046875,
0.40478515625
],
[
0.17236328125,
0.404052734375,
0.42724609375,
0.40478515625,
0.42724609375,
0.419189453125,
0.17236328125,
0.41845703125
],
[
0.17236328125,
0.436279296875,
0.52294921875,
0.4375,
0.52294921875,
0.45263671875,
0.17236328125,
0.451904296875
],
[
0.1732177734375,
0.46533203125,
0.473876953125,
0.466552734375,
0.473876953125,
0.48095703125,
0.1732177734375,
0.479736328125
],
[
0.1748046875,
0.48291015625,
0.81201171875,
0.48291015625,
0.81201171875,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.80810546875,
0.4873046875,
0.81640625,
0.4873046875,
0.81640625,
0.492431640625,
0.80810546875,
0.492431640625
],
[
0.176513671875,
0.49755859375,
0.82421875,
0.49755859375,
0.82421875,
0.51123046875,
0.176513671875,
0.51123046875
],
[
0.1732177734375,
0.51025390625,
0.82275390625,
0.51123046875,
0.82275390625,
0.52587890625,
0.1732177734375,
0.5244140625
],
[
0.1732177734375,
0.52587890625,
0.2156982421875,
0.52587890625,
0.2156982421875,
0.54052734375,
0.1732177734375,
0.54052734375
],
[
0.328369140625,
0.54052734375,
0.6689453125,
0.54345703125,
0.6689453125,
0.564453125,
0.328369140625,
0.56103515625
],
[
0.80126953125,
0.54443359375,
0.82666015625,
0.54443359375,
0.82666015625,
0.5625,
0.80126953125,
0.5625
],
[
0.464111328125,
0.56103515625,
0.490234375,
0.56103515625,
0.490234375,
0.5693359375,
0.464111328125,
0.5693359375
],
[
0.17236328125,
0.580078125,
0.82861328125,
0.580078125,
0.82861328125,
0.59765625,
0.17236328125,
0.59765625
],
[
0.1732177734375,
0.609375,
0.82421875,
0.60986328125,
0.82421875,
0.625,
0.1732177734375,
0.62353515625
],
[
0.1748046875,
0.625,
0.7509765625,
0.625,
0.7509765625,
0.63818359375,
0.1748046875,
0.63818359375
],
[
0.5146484375,
0.64013671875,
0.6298828125,
0.64013671875,
0.6298828125,
0.65380859375,
0.5146484375,
0.65380859375
],
[
0.361083984375,
0.646484375,
0.51318359375,
0.64599609375,
0.51318359375,
0.66357421875,
0.361083984375,
0.6640625
],
[
0.80224609375,
0.6484375,
0.826171875,
0.6484375,
0.826171875,
0.662109375,
0.80224609375,
0.662109375
],
[
0.5556640625,
0.6591796875,
0.59326171875,
0.6591796875,
0.59326171875,
0.67236328125,
0.5556640625,
0.67236328125
],
[
0.1748046875,
0.6748046875,
0.82421875,
0.6748046875,
0.82421875,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1748046875,
0.6875,
0.82421875,
0.6875,
0.82421875,
0.7021484375,
0.1748046875,
0.7021484375
],
[
0.1732177734375,
0.7021484375,
0.826171875,
0.7021484375,
0.826171875,
0.71533203125,
0.1732177734375,
0.71533203125
],
[
0.1748046875,
0.71630859375,
0.2484130859375,
0.71630859375,
0.2484130859375,
0.72802734375,
0.1748046875,
0.72802734375
],
[
0.1748046875,
0.74365234375,
0.82275390625,
0.74365234375,
0.82275390625,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.17236328125,
0.7568359375,
0.82275390625,
0.7568359375,
0.82275390625,
0.77783203125,
0.17236328125,
0.77783203125
],
[
0.1748046875,
0.77587890625,
0.82275390625,
0.77587890625,
0.82275390625,
0.7890625,
0.1748046875,
0.7890625
],
[
0.1748046875,
0.79052734375,
0.826171875,
0.79052734375,
0.826171875,
0.8037109375,
0.1748046875,
0.8037109375
],
[
0.1732177734375,
0.80322265625,
0.455810546875,
0.80322265625,
0.455810546875,
0.81640625,
0.1732177734375,
0.81640625
],
[
0.363525390625,
0.8193359375,
0.63134765625,
0.8193359375,
0.63134765625,
0.833984375,
0.363525390625,
0.833984375
],
[
0.80224609375,
0.82080078125,
0.826171875,
0.82080078125,
0.826171875,
0.83349609375,
0.80224609375,
0.83349609375
],
[
0.1748046875,
0.8359375,
0.82666015625,
0.8359375,
0.82666015625,
0.849609375,
0.1748046875,
0.849609375
],
[
0.17236328125,
0.85107421875,
0.619140625,
0.85107421875,
0.619140625,
0.8642578125,
0.17236328125,
0.8642578125
],
[
0.1748046875,
0.869140625,
0.82421875,
0.869140625,
0.82421875,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.1732177734375,
0.8837890625,
0.82421875,
0.88330078125,
0.82421875,
0.89794921875,
0.1732177734375,
0.8984375
],
[
0.1732177734375,
0.8984375,
0.375732421875,
0.8984375,
0.375732421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Next, we show a connection to loss variance (Proposition 2), an important quantity that arises in",
"generalization bounds [20] and is used as a regularization scheme [21, 3, 23, 22]. In particular,",
"MWLD(w1/2) provides us with lower and upper bounds for the loss variance. We also propose an",
"extension called coarse loss variance, which considers only a set of sensitive groups, allowing us to",
"incorporateknowledge about sensitive attributes.",
"We validate maximum weighted loss discrepancy on four common datasets from the fairness literature:",
"predicting recidivism, credit rating, income, and crime rate. We fit a logistic regression on these",
"datasets and estimate its MWLD(w) for various k. We observe that MWLD with smaller k converges",
"more slowly to the population quantity. We also observe the group attaining MWLD shrinks as k",
"decreases. We then use loss variance (LV) and coarse loss variance (CLV) regularization to train",
"models. Our empirical findings are as follows: 1) We halve the loss variance with only a small",
"increase in the average loss. 2) In some cases, using loss variance as a regularizer simultaneously",
"reduces the classification loss (higher accuracy) and loss variance (lower loss discrepancy)",
"Setup.Consider the prediction task of mapping each input E X to a probability distribution over",
"an output space y. Let h be a predictor which maps each input r to a probability distribution over",
"Y; for binary classification, y = {0, 1] and a predictor h : X → [0, 1] returns the probability of",
"y = 1. Let l(h, z) be the (bounded) loss incurred by predictor h on individual z-e.g., the zero-one",
"or logistic loss. Let p* denote the underlying distribution over individuals z",
"=(c,y);all expectations",
"are computed with respect to z ~ p*. Define a group to be a measurable function g : X × → [0, 1]",
"such that g(z) = 1 if individual z is in the group and O otherwise. Let G be the set of all groups. When",
"clear from context, we use l to denote l(h, z) and g to denote g(z), so that E[] is the population loss",
"and E[e / g = 1] is the loss of group g.",
"2MaximumWeightedLossDiscrepancy",
"We now introduce our central object of study:",
"Definition 1 (Maximum weighted loss discrepancy (MWLD). For a weighting function w : G -",
"-",
"[O, 1], loss function l, and predictor h, define the maximum weighted loss discrepancy MWLD(w, l, h)",
"to be the maximum difference between the loss of a group g and the population loss, weighted by",
"w(g):",
"MWLD (w, e, h) def sup w(g) [E[e I g = 1] - IE[e] ,",
"(2)",
"gEG",
"where the weighting function w (e.g., w(g) = E[g]1/2) intuitively controls the importance of group g.",
"Group fairness interpretation.By rearranging the terms of (2), we can bound the loss discrepancy",
"of any group in terms of the group weight and the maximum weighted loss discrepancy:",
"MWLD(w,l,h)",
"[E[E | g = 1] - E[] ",
"(3)",
"w(g)",
"where the bound is tighter for larger w(g). Existing statistical notions of fairness such as equal",
"opportunity [13] can be viewed as enforcing MWLD to be small for a weighting function w that is",
"1 on sensitive groups (e.g., different races) and 0 on all other groups; see Appendix A for further",
"discussion.",
"Distributional shift interpretation.We can use MWLD to bound the loss on a population with",
"shifted demographics. For any group g, define the mixture distribution q(-) def w(g)p*( I g ",
"1) + (1 - w(g))p*( / g = O), which tilts the original distribution p* more towards group g (assuming",
"w(g) ≥ E[gl). Then via simple algebra (Proposition 3 in Appendix B), the loss under this new",
"distribution q can be controlled as follows:",
"Ez~g[e] ≤ Ez~p*[] + MWLD(w, l, h).",
"(4)",
"This is similar in spirit to distributionally robust optimization (DRO) using a max-norm metric [10],",
"but the difference is that the mixture coefficient is group-dependent.",
"How do we now operationalize MWLD? After all, the supremum over all groups g in MWLD (2)",
"appears daunting.In Section 3, we show how we can efficiently estimate MwLD for a restricted",
"family of weighting functions.",
"2"
] |
{
"arxiv_id": "1906.03518",
"language": "en",
"page_idx": 1
}
|
||
1906.03518-en-2
|
1906.03518-en-2.jpg
|
3 $\textbf{Estimating Maximum Weighted Loss Discrepancy ( MWLD) }$
We now focus on the problem of estimating MWLD from data. For simplicity of notation, we write $\mathbf{MWLD}( w) \underbrace{\text{instead of MWLD}}( w, \ell , h) .$ Given $n$ points $z_1,\ldots,z_n\sim\hat{p}^{\star}$, our goal is to derive an $\underset{\Gamma}{\operatorname*{\operatorname*{\operatorname*{estimator~MWLD}}}}(w)$ that is (i) efficient to compute and (ii) accurately approximates MWLD$(w).$ Formally:
$$\forall\epsilon,\delta\in(0,\dfrac{1}{2}):\quad\mathbb{P}\left[\left|\text{MWLD}(w)-\widehat{\text{MWLD}}_n(w)\right|\geq\epsilon\right]\leq\delta\quad\text{and}n=\text{poly}(\log(1/\delta),1/\epsilon).$$
Whether this goal is achievable depends on the weighting function. Our first result is that we cannot
estimate MWLD for the uniform weighting function $( w^0\overset {\mathrm{def}}{\operatorname* { = } }\mathbb{I} [ \mathbb{E} [ g] > 0] ) :$
Proposition $1.Foranylossfunction0\leq\ell\leq1andpredictorh,if(\ell,h)isnon-degenerate$ $(\operatorname*{min}_{z}\ell(h,z)=0$ and $\operatorname*{max}_z\ell(h,z)=1)$, then there is no estimator $\widehat MWLD_n(w^{0},\ell,h)$ satisfying (5).
We prove Proposition 1 by constructing two statistically indistinguishable distributions such that $\mathbf{MWLD}(w^{0})\geq\frac12$ for one and MWLD$(w^{0})<\frac12$ for the other (see Appendix B for details). Proposition 1 is intuitive since MWLD for $w(g)=1$ is asking for uniform convergence over all measurable functions, which is a statistical impossibility. It therefore seems natural to shift our focus to weighting functions $w$ that decay to zero as the measure of the group goes to zero.
As our next result, we show that we can estimate MWLD for the weighting function $w^k( g) \overset {\mathrm{def}}{\operatorname* { \operatorname* { = } } }\mathbb{E} [ g] ^k$ for $k\in(0,1].$ In particular, we show (i) we can efficiently compute the empirical M$\widehat\mathrm{WLD}_n(w^k)$, and (ii) it converges to MWLD$(w^k).$ Letting $\widehat{\mathbb{E}}[\cdot]$ denote an expectation with respect to the $n$ points, define the plug-in estimator:
(6)
$$\widehat{\mathrm{MWLD}_{n}(w^{k})}\stackrel{\mathrm{def}}{=}\max_{g\in\mathcal{G}}\widehat{\mathbb{E}}[g]^{k}\Big|\widehat{\mathbb{E}}[\ell\mid g=1]-\widehat{\mathbb{E}}[\ell]\Big|.$$
Although M$\widehat\mathrm{WLD}_n(w^k)$ seems to have an intractable max, in the next theorem we prove we can actually compute it efficiently, and that MWLD$_n(w^k)$ converges to MWLD$(w^k)$, where the rate of convergence depends on $k.$ The key is that MWLD$_n(w^k)$ attains its max over a linear number of possible groups based on sorting by losses; this is what enables computational efficiency as well as uniform convergence.
$\textbf{Theorem 1. For }k\in ( 0, 1] , \textit{let w}^k\overset {\mathrm{def}}{\operatorname* { \operatorname* { = } } }\mathbb{E} [ g] ^k. \textit{ Given n i. i. d. sample from p}^* $, we can compute $\widehat{MWLD}(w^{k})(6)$ efficiently in $\mathcal{O}(n\log n)$ time; and for any parameters $\delta,\epsilon\in(0,\frac12),fora$ constant $C$, $ifn\geq \frac {C\log ( 1/ \delta ) }{\epsilon ^{2+ \frac 2k}}, thefollowing$ $holds: \mathbb{P} \left [ \left | MWLD( w^{k}) - \widehat {MWLD}_{n}( w^{k}) \right | \geq \epsilon \right ] \leq \delta .$
$Proofsketch.$ For computational efficiency, we show that if we sort the $n$ points by their losses $\ell_{1}\geq\ell_{2}\geq\cdots\geq\ell_{n}\left(\mathrm{in}\hat{\mathcal{O}}(n\log n)\right.$time),then there exists an index $t$ such that either $g=\{\ell_{1},\ldots,\ell_{t}\}$ or $g=\{\ell_{t+1},\ldots,\ell_{n}\}$ achieves the empirical maximum weighted loss discrepancy (6).
To show convergence, let $D(g)$ be the weighted loss discrepancy for group $g,D(g)\overset{\mathrm{def}}{\operatorname*{=}}$ $\mathbb{E}[g]^k\left|\mathbb{E}[\ell\mid g=1]-\mathbb{E}[\ell]\right|$, and analogously let $\widehat{D}(\underline{g})$ be its empirical counterpart. We first prove for $\mathbf{uny~}\alpha>0$ if $\mathbb{E}[g]\geq\alpha$,then $D(g)-\widehat{D}(g)\leq C\sqrt{\frac{\log(1/\delta)}{n\alpha}}$ with probability at least $1-\delta$, for some constant $C.$ Furthermore, since we assumed $\ell\leq1$, then $D(g)-\hat{D}(g)\leq\alpha^k.$ By combining these two upper bounds $(\max_\alpha\min($bound 1,bound 2)), we compute an upper bound independent of $\alpha$, thereby applicable to all groups.
To show uniform convergence, from the sorting result, we only need to consider groups of the form $\mathbb{I}[\ell(h,z)\geq u]$ and their counterparts $\mathbb{I}[\ell(h,z)\leq u].$ We prove uniform convergence over this set using the KKW inequality [19] and the same procedure we explained before for convergence. See Appendix B for the complete proof.
$$\text{口}$$
3
|
[
[
0.1732177734375,
0.0889892578125,
0.705078125,
0.09027099609375,
0.705078125,
0.10479736328125,
0.1732177734375,
0.10418701171875
],
[
0.1748046875,
0.12249755859375,
0.82275390625,
0.12249755859375,
0.82275390625,
0.1363525390625,
0.1748046875,
0.1363525390625
],
[
0.1748046875,
0.136962890625,
0.82421875,
0.136962890625,
0.82421875,
0.15087890625,
0.1748046875,
0.15087890625
],
[
0.1748046875,
0.154052734375,
0.826171875,
0.154052734375,
0.826171875,
0.1685791015625,
0.1748046875,
0.1685791015625
],
[
0.1748046875,
0.1673583984375,
0.24267578125,
0.1673583984375,
0.24267578125,
0.1817626953125,
0.1748046875,
0.1817626953125
],
[
0.1829833984375,
0.1944580078125,
0.27685546875,
0.1988525390625,
0.275390625,
0.216552734375,
0.181396484375,
0.212158203125
],
[
0.31787109375,
0.1956787109375,
0.82421875,
0.1956787109375,
0.82421875,
0.21337890625,
0.31787109375,
0.21337890625
],
[
0.1748046875,
0.2266845703125,
0.82421875,
0.2266845703125,
0.82421875,
0.240478515625,
0.1748046875,
0.240478515625
],
[
0.1732177734375,
0.242431640625,
0.65869140625,
0.242431640625,
0.65869140625,
0.260009765625,
0.1732177734375,
0.260009765625
],
[
0.1732177734375,
0.264404296875,
0.82275390625,
0.264404296875,
0.82275390625,
0.27783203125,
0.1732177734375,
0.27783203125
],
[
0.1732177734375,
0.2822265625,
0.82275390625,
0.2822265625,
0.82275390625,
0.29541015625,
0.1732177734375,
0.29541015625
],
[
0.1732177734375,
0.296630859375,
0.1993408203125,
0.296630859375,
0.1993408203125,
0.309326171875,
0.1732177734375,
0.309326171875
],
[
0.1748046875,
0.3212890625,
0.82275390625,
0.3212890625,
0.82275390625,
0.335205078125,
0.1748046875,
0.335205078125
],
[
0.1748046875,
0.335205078125,
0.82421875,
0.335205078125,
0.82421875,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1748046875,
0.349853515625,
0.82275390625,
0.349853515625,
0.82275390625,
0.363037109375,
0.1748046875,
0.363037109375
],
[
0.176513671875,
0.3642578125,
0.82421875,
0.3642578125,
0.82421875,
0.37744140625,
0.176513671875,
0.37744140625
],
[
0.1732177734375,
0.37744140625,
0.64453125,
0.37744140625,
0.64453125,
0.391357421875,
0.1732177734375,
0.391357421875
],
[
0.17236328125,
0.399658203125,
0.82275390625,
0.399658203125,
0.82275390625,
0.417236328125,
0.17236328125,
0.417236328125
],
[
0.1748046875,
0.419189453125,
0.826171875,
0.419189453125,
0.826171875,
0.434326171875,
0.1748046875,
0.434326171875
],
[
0.1748046875,
0.436279296875,
0.82666015625,
0.436279296875,
0.82666015625,
0.45068359375,
0.1748046875,
0.45068359375
],
[
0.1732177734375,
0.45068359375,
0.36279296875,
0.45068359375,
0.36279296875,
0.464111328125,
0.1732177734375,
0.464111328125
],
[
0.330078125,
0.474853515625,
0.6650390625,
0.474853515625,
0.6650390625,
0.49755859375,
0.330078125,
0.49755859375
],
[
0.80224609375,
0.479736328125,
0.826171875,
0.479736328125,
0.826171875,
0.49365234375,
0.80224609375,
0.49365234375
],
[
0.459228515625,
0.492431640625,
0.48046875,
0.492431640625,
0.48046875,
0.5,
0.459228515625,
0.5
],
[
0.1732177734375,
0.51318359375,
0.82275390625,
0.5146484375,
0.82275390625,
0.52880859375,
0.1732177734375,
0.52783203125
],
[
0.1732177734375,
0.53076171875,
0.82421875,
0.53076171875,
0.82421875,
0.54443359375,
0.1732177734375,
0.54443359375
],
[
0.1732177734375,
0.548828125,
0.826171875,
0.54736328125,
0.826171875,
0.5625,
0.1732177734375,
0.56298828125
],
[
0.1748046875,
0.56298828125,
0.82421875,
0.56298828125,
0.82421875,
0.5771484375,
0.1748046875,
0.5771484375
],
[
0.1732177734375,
0.57568359375,
0.318603515625,
0.5771484375,
0.318603515625,
0.59130859375,
0.1732177734375,
0.59033203125
],
[
0.17236328125,
0.59521484375,
0.826171875,
0.59521484375,
0.826171875,
0.6162109375,
0.17236328125,
0.6162109375
],
[
0.1748046875,
0.6162109375,
0.82421875,
0.6162109375,
0.82421875,
0.630859375,
0.1748046875,
0.630859375
],
[
0.1690673828125,
0.63623046875,
0.281982421875,
0.62890625,
0.28515625,
0.64794921875,
0.17236328125,
0.65576171875
],
[
0.273681640625,
0.63720703125,
0.438720703125,
0.63720703125,
0.438720703125,
0.6513671875,
0.273681640625,
0.6513671875
],
[
0.44677734375,
0.63525390625,
0.71826171875,
0.63623046875,
0.71826171875,
0.65380859375,
0.44677734375,
0.65283203125
],
[
0.23779296875,
0.6494140625,
0.261474609375,
0.6494140625,
0.261474609375,
0.65380859375,
0.23779296875,
0.65380859375
],
[
0.1748046875,
0.67529296875,
0.82275390625,
0.67529296875,
0.82275390625,
0.68994140625,
0.1748046875,
0.68994140625
],
[
0.1732177734375,
0.689453125,
0.82275390625,
0.68994140625,
0.82275390625,
0.705078125,
0.1732177734375,
0.7041015625
],
[
0.17236328125,
0.70263671875,
0.744140625,
0.7021484375,
0.744140625,
0.7197265625,
0.17236328125,
0.72021484375
],
[
0.1707763671875,
0.7265625,
0.826171875,
0.7236328125,
0.826171875,
0.74560546875,
0.1707763671875,
0.74853515625
],
[
0.176513671875,
0.7470703125,
0.82275390625,
0.7470703125,
0.82275390625,
0.76123046875,
0.176513671875,
0.76123046875
],
[
0.1690673828125,
0.76025390625,
0.82861328125,
0.7587890625,
0.82861328125,
0.7841796875,
0.1690673828125,
0.78515625
],
[
0.1748046875,
0.787109375,
0.82275390625,
0.787109375,
0.82275390625,
0.80029296875,
0.1748046875,
0.80029296875
],
[
0.1748046875,
0.8017578125,
0.826171875,
0.8017578125,
0.826171875,
0.81494140625,
0.1748046875,
0.81494140625
],
[
0.1732177734375,
0.814453125,
0.386474609375,
0.81494140625,
0.386474609375,
0.830078125,
0.1732177734375,
0.8291015625
],
[
0.1748046875,
0.8359375,
0.82275390625,
0.8359375,
0.82275390625,
0.849609375,
0.1748046875,
0.849609375
],
[
0.1748046875,
0.85107421875,
0.82275390625,
0.85107421875,
0.82275390625,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.1732177734375,
0.8623046875,
0.826171875,
0.86279296875,
0.826171875,
0.87744140625,
0.1732177734375,
0.876953125
],
[
0.1732177734375,
0.876953125,
0.410888671875,
0.876953125,
0.410888671875,
0.8916015625,
0.1732177734375,
0.8916015625
],
[
0.80810546875,
0.896484375,
0.82666015625,
0.896484375,
0.82666015625,
0.9111328125,
0.80810546875,
0.9111328125
],
[
0.494384765625,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.9462890625,
0.494384765625,
0.9462890625
]
] |
[
"3 Estimating Maximum Weighted Loss Discrepancy (MWLD)",
"We now focus on the problem of estimating MWLD from data. For simplicity of notation, we write",
"MWLD(w) instead of MWLD(w, e, h). Given n points z1, :.., zn ~ p*, our goal is to derive an",
"estimator MWLDn (w) that is (i) efficient to compute and (ii) accurately approximates MWLD(w).",
"Formally:",
"Ve,8 e (0, )",
"[|MWLD(w) - MWLDn(w)≥ ≤and n = poly(log(1/8), 1/e). (5)",
"Whether this goal is achievable depends on the weighting function. Our first result is that we cannot",
"estimate MWLD for the uniform weighting function (wo def I[E[g] > 0]):",
"Proposition 1. For any loss function o < < 1 and predictor h, if (e,h) is non-degenerate",
"(minz l(h, z) = O and maxz l(h, z) = 1), then there is no estimator MWLDn (wo, l, h) satisfying",
"(5).",
"We prove Proposition 1 by constructing two statistically indistinguishable distributions such that",
"MWLD(w0) > for one and MWLD(wo) < for the other (see Appendix B for details). Proposi-",
"tion 1 is intuitive since MWLDfor w(g)=1 is asking foruniform convergence over all measurable",
"functions, which is a statistical impossibility. It therefore seems natural to shift our focus to weighting",
"functions w that decay to zero as the measure of the group goes to zero.",
"As our next result, we show that we can estimate MWLD for the weighting function wh (g) def E[g]k",
"for k (O, 1]. In particular, we show (i) we can efficiently compute the empirical MwLDn (wk),",
"and (ii) it converges to MWLD(wk). Letting E[] denote an expectation with respect to the n points,",
"define the plug-in estimator:",
"MWLDn(whk) f max E[g][ | g = 1] - E[],",
"(6)",
"geg",
"Although MwLDn (wh) seems to have an intractable max, in the next theorem we prove we can",
"actually compute it efficiently,and that MWLDn(wk) converges to MWLD(wk),where the rate of",
"convergence depends on k. The key is that MWLD (wk) attains its max over a linear number of",
"possible groups based on sorting by losses; this is what enables computational efficiency as well as",
"uniform convergence.",
"Theorem 1. For k e (0, 1], let wk def E[gl'. Given n i.i.d. sample from p', we can compute",
"MWLD(wk)(6)efficiently in O(n log n) time; and for any parameters d, e E (O,), for a constant C",
"ifn≥Clog(1/8)",
",thefollowingholds:P",
"MWLD(wk)-MWLDn(wk)|≥≤s.",
"2+",
"Proof sketch. For computational efficiency, we show that if we sort the n points by their losses",
"li ≥ l2 ≥ ... ≥ ln (in O(n log n) time), then there exists an index t such that either g = [ei, ..., lt)",
"or g = [et+1, ..,en) achieves the empirical maximum weighted loss discrepancy (6).",
"To show convergence, let D(g) be the weighted loss discrepancy for group g, D(g) dof",
"E[g] [E[e [ g = 1] - E[e]l, and analogously let D(g) be its empirical counterpart. We first prove for",
"any>ifE[] ≥ a, then D(g)D(g)Co(/ with probability at least 1 s, for sme",
"constant C. Furthermore, since we assumed l < 1, then D(g) - D(g) < αk. By combining these",
"two upper bounds (max min(bound 1, bound 2)), we compute an upper bound independent of α,",
"thereby applicabletoall groups.",
"To show uniform convergence, from the sorting result, we only need to consider groups of the form",
"I[e(h, z) ≥ u] and their counterparts I[e(h, z) ≤ ul. We prove uniform convergence over this set",
"using the Kkw inequality [19] and the same procedure we explained before for convergence. See",
"Appendix B for the complete proof.",
"口",
"3"
] |
{
"arxiv_id": "1906.03518",
"language": "en",
"page_idx": 2
}
|
||
1906.03518-en-3
|
1906.03518-en-3.jpg
|
<FigureHere>
<FigureHere>
Figure 1: left: Upper bound guarantee for loss discrepancy of a Figure 2: Relationship between group for different values of $\breve{k}.$ right: Magnitude of the shift in a Var[$\ell]$ and MWLD$(w^{\hat{1}/2}).$ Here group is dictated by the weighting function of the group.
$f(x)\stackrel{\mathrm{def}}{=}x\sqrt{2-4\ln(x)}.$
4 A Closer Look at $w( g) = \mathbb{E} [ g] ^k\textbf{ and Connection to Loss Variance}$
As shown in Section 2, MWLD has two different interpretations: group fairness and distributional shift interpretations. In this section, we look at these interpretations for the family of weighting functions,$\hat w^k(g)=\mathbb{E}[g]^k$, for which we can efficiently estimate MWLD$(w^k).$ In Šection $\tilde4.1$, we show a connection between a particular member of this family $(k=1/2)$ and $loss\textit{variance. As an}$ extension, in Section 4.2, we introduce coarse loss variance, a simple modification of loss variance which measures weighted loss discrepancy only for sensitive groups.
of each group according to its size. Therefore groups with similar sizes have similar guarantees. For a fixed value of MWLD$(w^k)$ (here 0.1),Figure $\tilde{1}($left) shows the bounds on the group loss discrepancy for different sizes and $k.$ The upper bound guarantee for smaller groups is weaker, and the parameter $k$ governs how much this upper bound varies across group sizes.
new distribution where the weight of $g$ is increased by a maximum factor of $\mathbb{E}[g]^k.$ Figure 1(right) illustrates this maximum upweighting factor $\mathbb{E}[g]^k.$ The upweighting factor for smaller groups is smaller, and the parameter $k$ governs how much this factor varies across group sizes.
4.1 Loss Variance and Maximum Weighted Loss Discrepancy
In this section, we show an interesting connection between a particular member of the introduced family of weighting function, $w^{1/ 2}( g) = \mathbb{E} [ g] ^{1/ 2}, \hat{\text{and loss variance, which appears prominently in}}$ generalization bounds[20]. Loss variance, Var[$\ell]$, is the average squared difference between the loss of individuals and the population loss:
(7)
$$\mathrm{Var}[\ell]=\mathbb{E}\left[\left(\ell-\mathbb{E}[\ell]\right)^2\right].$$
From the law of total variance, we have Var$[\ell]\geq$Var$[\mathbb{E}[\ell\mid g]]$ for any group $g.$ By observing that Var$[\mathbb{E}[\ell\mid g]]\geq\mathbb{E}[g]\left(\mathbb{E}[\ell\mid g=1]-\mathbb{E}[\ell]\right)^2$, we see that square root of loss variance is an upper bound on MWLD$(w^{1/2}).$ This allows us to bound the loss of any group in terms of the loss variance (using (3)). A natural next question is about the tightness of the upper bound. How much larger can the variance be compared to the MWLD$(w^{1/2})?$ The next proposition shows that loss variance also provides a lower bound on a function of the MWLD$(w^{1/2}).$
Proposition 2. For any measurable loss function $0\leq \ell \leq 1, \textit{the following holds: MWLD}( w^{1/ 2}) \leq$
$$\sqrt{Var[\ell]}\leq MWLD(w^{1/2})\sqrt{2-4\ln\left(MWLD(w^{1/2})\right)}.$$
$Proofsketch.$ We first center the losses and make the average loss 0 (without changing the MWLD). For any $u>0$, let $g_u$ be the group of points with loss greater than $u.$ By definition of MWLD, we have $\sqrt {\mathbb{E} [ g_u] }u\leq \sqrt {\mathbb{E} [ g_u] }\mathbb{E} [ \ell \mid g_u= 1] \leq \mathbf{MWLD}( w^{1/ 2}) .$ Therefore we have $\mathbb{P} [ \ell \geq u] = \mathbb{E} [ g_u] \leq$ $\frac {\mathrm{MWLD}( w^{1/ 2}) ^2}{u^2}.$ Using integration by parts, we express variance with an integral expression in term of
|
[
[
0.236083984375,
0.1256103515625,
0.2459716796875,
0.1331787109375,
0.23779296875,
0.1383056640625,
0.2296142578125,
0.1319580078125
],
[
0.2303466796875,
0.1319580078125,
0.25244140625,
0.1337890625,
0.2467041015625,
0.1666259765625,
0.2254638671875,
0.1641845703125
],
[
0.736328125,
0.1514892578125,
0.74853515625,
0.1514892578125,
0.74853515625,
0.160400390625,
0.736328125,
0.160400390625
],
[
0.69921875,
0.1641845703125,
0.72802734375,
0.1546630859375,
0.73388671875,
0.1666259765625,
0.705078125,
0.175537109375
],
[
0.744140625,
0.169189453125,
0.779296875,
0.1673583984375,
0.78125,
0.1787109375,
0.74609375,
0.18115234375
],
[
0.3203125,
0.199462890625,
0.34228515625,
0.199462890625,
0.34228515625,
0.21337890625,
0.3203125,
0.21337890625
],
[
0.4755859375,
0.1988525390625,
0.496826171875,
0.1988525390625,
0.496826171875,
0.21337890625,
0.4755859375,
0.21337890625
],
[
0.70361328125,
0.203857421875,
0.734375,
0.203857421875,
0.734375,
0.212158203125,
0.70361328125,
0.212158203125
],
[
0.17724609375,
0.2247314453125,
0.6044921875,
0.2259521484375,
0.6044921875,
0.240478515625,
0.17724609375,
0.2392578125
],
[
0.59619140625,
0.2259521484375,
0.8203125,
0.2259521484375,
0.8203125,
0.2392578125,
0.59619140625,
0.2392578125
],
[
0.1805419921875,
0.240478515625,
0.6044921875,
0.240478515625,
0.6044921875,
0.25390625,
0.1805419921875,
0.25390625
],
[
0.59716796875,
0.240478515625,
0.8203125,
0.240478515625,
0.8203125,
0.255126953125,
0.59716796875,
0.255126953125
],
[
0.178955078125,
0.255126953125,
0.5498046875,
0.255126953125,
0.5498046875,
0.268310546875,
0.178955078125,
0.268310546875
],
[
0.60888671875,
0.25830078125,
0.77392578125,
0.25830078125,
0.77392578125,
0.27587890625,
0.60888671875,
0.27587890625
],
[
0.1732177734375,
0.300537109375,
0.740234375,
0.300537109375,
0.740234375,
0.31494140625,
0.1732177734375,
0.31494140625
],
[
0.1732177734375,
0.33203125,
0.82421875,
0.33203125,
0.82421875,
0.34521484375,
0.1732177734375,
0.34521484375
],
[
0.1732177734375,
0.34521484375,
0.82177734375,
0.34521484375,
0.82177734375,
0.359130859375,
0.1732177734375,
0.359130859375
],
[
0.1748046875,
0.35986328125,
0.82275390625,
0.35986328125,
0.82275390625,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.1732177734375,
0.373779296875,
0.82275390625,
0.373779296875,
0.82275390625,
0.386962890625,
0.1732177734375,
0.386962890625
],
[
0.176513671875,
0.389404296875,
0.82275390625,
0.389404296875,
0.82275390625,
0.399658203125,
0.176513671875,
0.399658203125
],
[
0.1748046875,
0.401611328125,
0.625,
0.401611328125,
0.625,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.1748046875,
0.421630859375,
0.82275390625,
0.421630859375,
0.82275390625,
0.436279296875,
0.1748046875,
0.436279296875
],
[
0.176513671875,
0.4375,
0.82177734375,
0.4375,
0.82177734375,
0.447509765625,
0.176513671875,
0.447509765625
],
[
0.1748046875,
0.449462890625,
0.82177734375,
0.449462890625,
0.82177734375,
0.46337890625,
0.1748046875,
0.46337890625
],
[
0.1748046875,
0.46337890625,
0.82421875,
0.46337890625,
0.82421875,
0.4765625,
0.1748046875,
0.4765625
],
[
0.1748046875,
0.477783203125,
0.59619140625,
0.477783203125,
0.59619140625,
0.4912109375,
0.1748046875,
0.4912109375
],
[
0.1707763671875,
0.494384765625,
0.82421875,
0.49560546875,
0.82421875,
0.51318359375,
0.1707763671875,
0.51220703125
],
[
0.1748046875,
0.51318359375,
0.826171875,
0.51318359375,
0.826171875,
0.5263671875,
0.1748046875,
0.5263671875
],
[
0.1732177734375,
0.52783203125,
0.826171875,
0.52783203125,
0.826171875,
0.54150390625,
0.1732177734375,
0.54150390625
],
[
0.1732177734375,
0.54248046875,
0.72607421875,
0.54248046875,
0.72607421875,
0.55615234375,
0.1732177734375,
0.55615234375
],
[
0.1748046875,
0.57373046875,
0.2091064453125,
0.57373046875,
0.2091064453125,
0.58544921875,
0.1748046875,
0.58544921875
],
[
0.200927734375,
0.5712890625,
0.611328125,
0.57275390625,
0.611328125,
0.5869140625,
0.200927734375,
0.5859375
],
[
0.17236328125,
0.59765625,
0.826171875,
0.5966796875,
0.826171875,
0.611328125,
0.17236328125,
0.6123046875
],
[
0.1748046875,
0.61279296875,
0.82275390625,
0.61279296875,
0.82275390625,
0.62744140625,
0.1748046875,
0.62744140625
],
[
0.1748046875,
0.62744140625,
0.82421875,
0.62744140625,
0.82421875,
0.6416015625,
0.1748046875,
0.6416015625
],
[
0.1748046875,
0.6416015625,
0.42724609375,
0.6416015625,
0.42724609375,
0.65576171875,
0.1748046875,
0.65576171875
],
[
0.414306640625,
0.66650390625,
0.57421875,
0.66650390625,
0.57421875,
0.6845703125,
0.414306640625,
0.6845703125
],
[
0.80224609375,
0.66943359375,
0.826171875,
0.66943359375,
0.826171875,
0.68310546875,
0.80224609375,
0.68310546875
],
[
0.17236328125,
0.693359375,
0.82421875,
0.6943359375,
0.82421875,
0.7119140625,
0.17236328125,
0.7109375
],
[
0.1732177734375,
0.7119140625,
0.82275390625,
0.7119140625,
0.82275390625,
0.7265625,
0.1732177734375,
0.7265625
],
[
0.1748046875,
0.72900390625,
0.82421875,
0.72900390625,
0.82421875,
0.74365234375,
0.1748046875,
0.74365234375
],
[
0.1748046875,
0.74365234375,
0.82275390625,
0.74365234375,
0.82275390625,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1748046875,
0.7568359375,
0.826171875,
0.7568359375,
0.826171875,
0.771484375,
0.1748046875,
0.771484375
],
[
0.17236328125,
0.77294921875,
0.564453125,
0.771484375,
0.564453125,
0.787109375,
0.17236328125,
0.78857421875
],
[
0.1748046875,
0.79345703125,
0.82177734375,
0.79345703125,
0.82177734375,
0.80810546875,
0.1748046875,
0.80810546875
],
[
0.17724609375,
0.81298828125,
0.525390625,
0.81298828125,
0.525390625,
0.82763671875,
0.17724609375,
0.82763671875
],
[
0.176513671875,
0.8486328125,
0.826171875,
0.8486328125,
0.826171875,
0.8623046875,
0.176513671875,
0.8623046875
],
[
0.1748046875,
0.86279296875,
0.82421875,
0.86279296875,
0.82421875,
0.876953125,
0.1748046875,
0.876953125
],
[
0.1732177734375,
0.876953125,
0.826171875,
0.876953125,
0.826171875,
0.89453125,
0.1732177734375,
0.89453125
],
[
0.17724609375,
0.89306640625,
0.253173828125,
0.89306640625,
0.253173828125,
0.90380859375,
0.17724609375,
0.90380859375
],
[
0.2548828125,
0.89794921875,
0.82666015625,
0.89794921875,
0.82666015625,
0.912109375,
0.2548828125,
0.912109375
],
[
0.2115478515625,
0.90771484375,
0.22216796875,
0.90771484375,
0.22216796875,
0.912109375,
0.2115478515625,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"",
"OTA",
"12",
"VLDC",
"Nare)",
"E[g]",
"E[g]",
"VVare",
"Figure l: left: Upper bound guarantee for loss discrepancy of a",
"Figure2:Relationship between",
"group for different values of k. right: Magnitude of the shift in a",
"Var[e| and MWLD(w1/2). Here",
"group is dictated by the weighting function of the group.",
"f(r) def rV2- 4ln(a).",
"4 A Closer Look at w(g) = E[glk and Connection to Loss Variance",
"As shown in Section 2, MWLD has two different interpretations: group fairness and distributional",
"shift interpretations. In this section, we look at these interpretations for the family of weighting",
"functions, w(g) =E[g], for which we can efficiently estimate MWLD(w). In Section 4.1, we",
"show a connection between a particular member of this family (k = 1/2) and loss variance. As an",
"extension,in Section4.2,weintroducecoarselossvariance,asimplemodification oflossvariance",
"which measures weighted loss discrepancy only for sensitive groups.",
"From the group fairness interpretation (3), MWLD(w) provides guarantees on the loss discrepancy",
"ofeachgroupaccordingtoits size.Thereforegroupswithsimilar sizes have similarguarantees.Fora",
"fixed value of MWLD(w) (here O.1), Figure 1 (left) shows the bounds on the group loss discrepancy",
"for different sizes and k. The upper bound guarantee for smaller groups is weaker, and the parameter",
"k governs how much this upper bound varies across group sizes.",
"From the distributional shift interpretation (4), MWLD(wk) provides a guarantee on the loss of a",
"new distribution where the weight of g is increased by a maximum factor of E[gl. Figure 1 (right)",
"illustrates this maximum upweighting factor E[g]k. The upweighting factor for smaller groups is",
"smaller, and the parameter k governs how much this factor varies across group sizes.",
"4.1",
"Loss Variance and Maximum Weighted Loss Discrepancy",
"In this section, we show an interesting connection between a particular member of the introduced",
"family of weighting function, w1/2(g) = E[g]1/2, and loss variance, which appears prominently in",
"generalization bounds [20]. Loss variance, Var[], is the average squared difference between the loss",
"of individuals and the population loss:",
"Var[] = E(e - E[e])?",
"(7)",
"From the law of total variance, we have Var[] ≥ Var[E[e [ g] for any group g. By observing that",
"Var[E[e | gll ≥ E[gl (E[e g = 1] - E[e])?, we see that square root of loss variance is an upper",
"bound on MWLD(w1/2). This allows us to bound the loss of any group in terms of the loss variance",
"(using (3)). A natural next question is about the tightness of the upper bound. How much larger can",
"the variance be compared to the MWLD(w1/2)? The next proposition shows that loss variance also",
"provides a lower bound on a function of the MWLD(w1/2).",
"Proposition 2. For any measurable loss function O ≤l≤ 1, the following holds:MWLD(w1/2)",
"VVar[l≤ MWLD(w1/2)/2 - 4ln (MWLD(w1/2))",
"Proof sketch. We first center the losses and make the average loss O (without changing the MWLD).",
"For any u > O, let gu be the group of points with loss greater than u. By definition of MWLD, we",
"have VE[gu]u ≤ VE[gu]E[e | gu = 1] ≤ MWLD(w]/2). Therefore we have P[e ≥ u] = E[gu] ≤",
"MWLD(w1/2)",
". Using integration by parts, we express variance with an integral expression in term of",
"912",
"4"
] |
{
"arxiv_id": "1906.03518",
"language": "en",
"page_idx": 3
}
|
||
1906.03518-en-4
|
1906.03518-en-4.jpg
|
Figure 3: Each individual has two sensitive attributes, color and height, and one non-sensitive attribute, having a hat, (a) $\mathbf{MWLD}( w^{1/ 2})$ $\leq$ $\sqrt {\mathrm{Var}[ \ell ] }$ $\overset {\smile }{\operatorname* { = } }$ 0.5. $( \mathbf{b} )$ $\mathbf{MWLD}( w_{\mathrm{A} }^{1/ 2}) \leq$Var$[\mathbb{E}[\ell\mid A]]=0.$ Note that, since the weights of the groups not defined on sensitive attributes are 0, their expected loss can deviate a lot from average loss (e.g, the expected loss of individuals with hats is 1, which deviates a lot from average loss $\mathbb{E}[\ell]=0.5).$
<FigureHere>
cumulative density function (CDF). Plugging this bound into the integral expression yields the result.
□
For more details, see Appendix B.
Figure 2 shows the bounds on MWLD$(w^{1/2})$ for different values of $\sqrt{\mathrm{Var}[\ell]}.$ This proposition establishes a connection between statistical generalization and MWLD (and thereby group fairness). Furthermore, this connections states that reducing loss discrepancy between individuals and the population (7) leads to lower loss discrepancy between groups and the population (2) and vice versa.
4.2 Sensitive Attributes and Coarse Loss Variance
So far, we have focused on the loss discrepancy over all groups, which could be too demanding. Suppose we are given a set of sensitive attributes (e.g., race, gender), and we are interested only in groups defined in terms of those attributes. We define coarse loss variance, which first averages the losses of all individuals with the same sensitive attribute values and considers the variance of these average losses. Formally, let $A$ denote the sensitive attributes; for example, $A=[$race,gender,...]. Then the coarse loss variance is:
$$\mathrm{Var}\left[\mathbb{E}[\ell\mid A]\right]=\mathbb{E}\left[\left(\mathbb{E}[\ell\mid A]-\mathbb{E}[\ell]\right)^2\right].$$
(8)
Coarse loss variance is smaller than loss variance (7) because it ignores fluctuations in the losses of individuals who have identical sensitive attributes. Figure 3 shows the difference between loss variance and coarse loss variance. Analogous to Proposition 2 in previous section, we show that coarse loss variance is a close estimate of MWLD$(w)$ where $w(g)=\sqrt{\mathbb{E}[g]}$ if $g$ is a function only of sensitive attributes and 0 otherwise. Define $\mathcal{G}_A$ to be the set of groups $g$ such that $g(z)$ only depends on the sensitive attributes $A(z).$ Let $w_{\mathrm{A} }^{1/ 2}( g) \overset {\mathrm{def}}{\operatorname* { \operatorname* { = } } }\mathbb{I} [ g\in \mathcal{G} _A] \sqrt {\mathbb{E} [ g] }$, then we have:
$$\mathrm M\mathrm W\mathrm L\mathrm D(w_\mathrm A^{1/2})\leq\sqrt{\mathrm V\mathrm a\mathrm r[\mathbb E[\ell\mid A]]}\leq\mathrm M\mathrm W\mathrm L\mathrm D(w_\mathrm A^{1/2})\sqrt{2-4\ln\left(\mathrm M\mathrm W\mathrm L\mathrm D(w_\mathrm A^{1/2})\right)}$$
(9)
For the formal propositions regarding coarse loss variance, see Corollary 1 in Appendix B. As a caveat, empirical coarse loss variance converges slower to its population counterpart in comparison to loss variance (See Theorem 4 in Appendix C for the exact convergence rate).
Remark 1. In some applications, the impact of misclassification could be quite different depending on the true label of the individual. For example, in deciding whether to provide loans, denying an eligible individual a loan could have a greater negative impact than providing a loan to a defaulter. In such situations, we consider the variance conditioned on the label, i.e., $\mathbb{E} [ \tilde{\text{Var}} [ \mathbb{E} [ \ell \mid y, A] \mid y] ] $,so that we do not attempt to pull the losses of individuals with different labels (and consequently different impacts of misclassification) together.
# 5 Experiments
We first explore the effect of parameter $k$ in MWLD$(w^k)$, as discussed in Section 4. We then use (coarse) loss variance to train models and show that we can halve the loss variance without significant increase in average loss. Table 1 shows a summary of the datasets considered. For more details about these datasets, see Appendix E.
5
|
[
[
0.54833984375,
0.0889892578125,
0.82666015625,
0.0889892578125,
0.82666015625,
0.1029052734375,
0.54833984375,
0.1029052734375
],
[
0.35693359375,
0.0997314453125,
0.39794921875,
0.1029052734375,
0.393798828125,
0.1287841796875,
0.35302734375,
0.1256103515625
],
[
0.45751953125,
0.10162353515625,
0.5,
0.10162353515625,
0.5,
0.1275634765625,
0.45751953125,
0.1275634765625
],
[
0.5498046875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11553955078125,
0.5498046875,
0.11553955078125
],
[
0.47802734375,
0.11236572265625,
0.490234375,
0.1142578125,
0.488525390625,
0.11932373046875,
0.476318359375,
0.117431640625
],
[
0.54736328125,
0.117431640625,
0.82421875,
0.117431640625,
0.82421875,
0.1307373046875,
0.54736328125,
0.1307373046875
],
[
0.3203125,
0.1337890625,
0.361083984375,
0.1337890625,
0.361083984375,
0.14453125,
0.3203125,
0.14453125
],
[
0.4306640625,
0.1337890625,
0.473876953125,
0.1337890625,
0.473876953125,
0.14453125,
0.4306640625,
0.14453125
],
[
0.5439453125,
0.1300048828125,
0.826171875,
0.1307373046875,
0.826171875,
0.1483154296875,
0.5439453125,
0.147705078125
],
[
0.275390625,
0.1363525390625,
0.2900390625,
0.1363525390625,
0.2900390625,
0.14013671875,
0.275390625,
0.14013671875
],
[
0.381591796875,
0.1363525390625,
0.398681640625,
0.1351318359375,
0.400390625,
0.143310546875,
0.38232421875,
0.14453125
],
[
0.46728515625,
0.1363525390625,
0.52685546875,
0.1307373046875,
0.52783203125,
0.143310546875,
0.468017578125,
0.1483154296875
],
[
0.6298828125,
0.15087890625,
0.82275390625,
0.15087890625,
0.82275390625,
0.1654052734375,
0.6298828125,
0.1654052734375
],
[
0.366943359375,
0.1395263671875,
0.384033203125,
0.1395263671875,
0.384033203125,
0.146484375,
0.366943359375,
0.146484375
],
[
0.5498046875,
0.1514892578125,
0.63134765625,
0.1514892578125,
0.63134765625,
0.1634521484375,
0.5498046875,
0.1634521484375
],
[
0.361083984375,
0.1622314453125,
0.38818359375,
0.1622314453125,
0.38818359375,
0.18115234375,
0.361083984375,
0.18115234375
],
[
0.5498046875,
0.1666259765625,
0.82421875,
0.1666259765625,
0.82421875,
0.1767578125,
0.5498046875,
0.1767578125
],
[
0.5498046875,
0.179931640625,
0.82421875,
0.179931640625,
0.82421875,
0.1900634765625,
0.5498046875,
0.1900634765625
],
[
0.36279296875,
0.1912841796875,
0.373291015625,
0.1912841796875,
0.373291015625,
0.1956787109375,
0.36279296875,
0.1956787109375
],
[
0.371826171875,
0.189453125,
0.38818359375,
0.189453125,
0.38818359375,
0.1976318359375,
0.371826171875,
0.1976318359375
],
[
0.5498046875,
0.1925048828125,
0.82275390625,
0.1925048828125,
0.82275390625,
0.205810546875,
0.5498046875,
0.205810546875
],
[
0.1854248046875,
0.212158203125,
0.1895751953125,
0.212158203125,
0.1895751953125,
0.2171630859375,
0.1854248046875,
0.2171630859375
],
[
0.426513671875,
0.212158203125,
0.455810546875,
0.212158203125,
0.455810546875,
0.2197265625,
0.426513671875,
0.2197265625
],
[
0.54736328125,
0.20703125,
0.82421875,
0.2052001953125,
0.82421875,
0.2183837890625,
0.54736328125,
0.2203369140625
],
[
0.2484130859375,
0.2139892578125,
0.2548828125,
0.2139892578125,
0.2548828125,
0.2171630859375,
0.2484130859375,
0.2171630859375
],
[
0.32275390625,
0.21337890625,
0.35546875,
0.21337890625,
0.35546875,
0.2203369140625,
0.32275390625,
0.2203369140625
],
[
0.39453125,
0.2139892578125,
0.418212890625,
0.2139892578125,
0.418212890625,
0.2183837890625,
0.39453125,
0.2183837890625
],
[
0.54833984375,
0.2203369140625,
0.82421875,
0.2203369140625,
0.82421875,
0.2342529296875,
0.54833984375,
0.2342529296875
],
[
0.23779296875,
0.2298583984375,
0.2607421875,
0.2298583984375,
0.2607421875,
0.2442626953125,
0.23779296875,
0.2442626953125
],
[
0.4150390625,
0.2298583984375,
0.438720703125,
0.2298583984375,
0.438720703125,
0.2442626953125,
0.4150390625,
0.2442626953125
],
[
0.54736328125,
0.2342529296875,
0.71142578125,
0.2342529296875,
0.71142578125,
0.2474365234375,
0.54736328125,
0.2474365234375
],
[
0.1732177734375,
0.272705078125,
0.82275390625,
0.2734375,
0.82275390625,
0.28857421875,
0.1732177734375,
0.287353515625
],
[
0.1732177734375,
0.28857421875,
0.400390625,
0.28857421875,
0.400390625,
0.302978515625,
0.1732177734375,
0.302978515625
],
[
0.810546875,
0.290283203125,
0.82421875,
0.290283203125,
0.82421875,
0.2998046875,
0.810546875,
0.2998046875
],
[
0.1748046875,
0.318115234375,
0.82177734375,
0.318115234375,
0.82177734375,
0.333251953125,
0.1748046875,
0.333251953125
],
[
0.1732177734375,
0.33203125,
0.82421875,
0.333251953125,
0.82421875,
0.347900390625,
0.1732177734375,
0.3466796875
],
[
0.1748046875,
0.347900390625,
0.82421875,
0.347900390625,
0.82421875,
0.361083984375,
0.1748046875,
0.361083984375
],
[
0.17236328125,
0.361083984375,
0.82666015625,
0.359130859375,
0.82666015625,
0.373779296875,
0.17236328125,
0.375732421875
],
[
0.1748046875,
0.391357421875,
0.53515625,
0.391357421875,
0.53515625,
0.40478515625,
0.1748046875,
0.40478515625
],
[
0.1748046875,
0.416748046875,
0.82421875,
0.416748046875,
0.82421875,
0.43115234375,
0.1748046875,
0.43115234375
],
[
0.1748046875,
0.43115234375,
0.82275390625,
0.43115234375,
0.82275390625,
0.4443359375,
0.1748046875,
0.4443359375
],
[
0.176513671875,
0.4443359375,
0.826171875,
0.4443359375,
0.826171875,
0.458251953125,
0.176513671875,
0.458251953125
],
[
0.176513671875,
0.460205078125,
0.8203125,
0.460205078125,
0.8203125,
0.47021484375,
0.176513671875,
0.47021484375
],
[
0.1748046875,
0.472900390625,
0.826171875,
0.472900390625,
0.826171875,
0.486083984375,
0.1748046875,
0.486083984375
],
[
0.1748046875,
0.48486328125,
0.38818359375,
0.486083984375,
0.38818359375,
0.5,
0.1748046875,
0.498779296875
],
[
0.363525390625,
0.5048828125,
0.62109375,
0.5048828125,
0.62109375,
0.52587890625,
0.363525390625,
0.52587890625
],
[
0.80126953125,
0.50830078125,
0.82861328125,
0.50830078125,
0.82861328125,
0.5244140625,
0.80126953125,
0.5244140625
],
[
0.1748046875,
0.54150390625,
0.82275390625,
0.54150390625,
0.82275390625,
0.5546875,
0.1748046875,
0.5546875
],
[
0.176513671875,
0.55615234375,
0.82275390625,
0.55615234375,
0.82275390625,
0.56640625,
0.176513671875,
0.56640625
],
[
0.176513671875,
0.57080078125,
0.82177734375,
0.57080078125,
0.82177734375,
0.58056640625,
0.176513671875,
0.58056640625
],
[
0.1732177734375,
0.58349609375,
0.826171875,
0.583984375,
0.826171875,
0.5986328125,
0.1732177734375,
0.59765625
],
[
0.1732177734375,
0.5986328125,
0.82421875,
0.5986328125,
0.82421875,
0.6123046875,
0.1732177734375,
0.6123046875
],
[
0.1748046875,
0.6162109375,
0.70166015625,
0.6162109375,
0.70166015625,
0.630859375,
0.1748046875,
0.630859375
],
[
0.2344970703125,
0.64794921875,
0.55615234375,
0.64794921875,
0.55615234375,
0.66552734375,
0.2344970703125,
0.66552734375
],
[
0.61279296875,
0.64599609375,
0.73779296875,
0.646484375,
0.73779296875,
0.6640625,
0.61279296875,
0.66357421875
],
[
0.8056640625,
0.6494140625,
0.82421875,
0.6494140625,
0.82421875,
0.66357421875,
0.8056640625,
0.66357421875
],
[
0.1748046875,
0.67822265625,
0.82177734375,
0.67822265625,
0.82177734375,
0.69140625,
0.1748046875,
0.69140625
],
[
0.1748046875,
0.69140625,
0.82177734375,
0.69140625,
0.82177734375,
0.705078125,
0.1748046875,
0.705078125
],
[
0.17236328125,
0.7041015625,
0.69384765625,
0.705078125,
0.69384765625,
0.7197265625,
0.17236328125,
0.71826171875
],
[
0.1748046875,
0.7236328125,
0.82275390625,
0.7236328125,
0.82275390625,
0.7373046875,
0.1748046875,
0.7373046875
],
[
0.1748046875,
0.7373046875,
0.82177734375,
0.73779296875,
0.82177734375,
0.75,
0.1748046875,
0.74853515625
],
[
0.1732177734375,
0.75,
0.82275390625,
0.75048828125,
0.82275390625,
0.76513671875,
0.1732177734375,
0.7646484375
],
[
0.1748046875,
0.76513671875,
0.82275390625,
0.76513671875,
0.82275390625,
0.77880859375,
0.1748046875,
0.77880859375
],
[
0.1732177734375,
0.77880859375,
0.82177734375,
0.77783203125,
0.82177734375,
0.79150390625,
0.1732177734375,
0.79248046875
],
[
0.1732177734375,
0.79248046875,
0.423095703125,
0.79248046875,
0.423095703125,
0.80615234375,
0.1732177734375,
0.80615234375
],
[
0.1732177734375,
0.8271484375,
0.313720703125,
0.8271484375,
0.313720703125,
0.84130859375,
0.1732177734375,
0.84130859375
],
[
0.17236328125,
0.85595703125,
0.82275390625,
0.85498046875,
0.82275390625,
0.869140625,
0.17236328125,
0.87060546875
],
[
0.1748046875,
0.87060546875,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.1748046875,
0.8837890625
],
[
0.1748046875,
0.8837890625,
0.82421875,
0.8837890625,
0.82421875,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1732177734375,
0.8984375,
0.38232421875,
0.8984375,
0.38232421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.494384765625,
0.94775390625
]
] |
[
"Figure 3: Each individual has two sensi-",
"?",
"0",
"tive attributes, color and height, and one",
"",
"non-sensitive attribute, having a hat, (a)",
"E[EA=[",
"E[E]A=[",
"MWLD(w1/2)≤ Var[ =0.5. (b)",
"C-",
"",
"[blue']=0.3",
"2) ≤ Var[E[ | A] = 0. Note",
"red",
"MWLD(w))",
"I",
"that.sincetheweights ofthegroupsnotde-",
"finedonsensitiveattributesareO.theirex-",
"",
"I",
"pected loss can deviate a lot from average",
"1",
"ELLIA",
"loss (e.g,theexpected loss of individuals",
"-",
"E[L|A=",
"=0.5",
"with hats is 1, which deviates a lot from",
"(a)",
"(b)",
"average loss E[ = 0.5].",
"cumulative density function (CDF).Plugging this bound into the integral expression yields the result",
"For more details, see Appendix B.",
"口",
"Figure 2 shows the bounds on MWLD(w1/2) for different values of Var[e]. This proposition",
"establishes a connection between statistical generalization and MWLD (and thereby group fairness)",
"Furthermore, this connections states that reducing loss discrepancy between individuals and the",
"population (7) leads to lower loss discrepancy between groups and the population (2) and vice versa.",
"4.2Sensitive Attributes and Coarse Loss Variance",
"So far, we have focused on the loss discrepancy over all groups, which could be too demanding",
"Suppose we are given a set of sensitive attributes (e.g., race, gender), and we are interested only in",
"groups defined in terms of those attributes. We define coarse loss variance, which first averages the",
"lossesofallindividualswiththesamesensitiveattributevaluesandconsidersthevarianceofthese",
"average losses. Formally, let A denote the sensitive attributes; for example, A =[race, gender, . ..].",
"Then the coarse loss variance is:",
"Var [E[ [ A] = E (E[e A] - E[])?",
"(8)",
"Coarse loss variance is smaller than loss variance (7) because it ignores fluctuations in the losses",
"ofindividualswhohaveidentical sensitiveattributes.Figure3showsthedifferencebetweenloss",
"variance and coarseloss variance.Analogous to Proposition2in previous section,we show that",
"coarse loss variance is a close estimate of MWLD(w) where w(g) = VE[g] if g is a function only of",
"sensitive attributes and 0 otherwise. Define GA to be the set of groups g such that g(z) only depends",
"on the sensitive attributes A(z). Let w//2 (g) def I[g E GAl VE[gl, then we have:",
"MWLD(w/2) ≤VVar[E[e | A] ≤MWLD(wA",
"4ln(MWLD(w))",
"(9)",
"For the formal propositions regarding coarse loss variance, see Corollary 1 in Appendix B. As a",
"caveat, empirical coarse loss variance converges slowerto its population counterpart in comparison",
"to loss variance (See Theorem 4 in Appendix C for the exact convergence rate).",
"Remark 1. In some applications, the impact of misclassification could be quite different depending",
"onthetruelabeloftheindividual.Forexample,indecidingwhethertoprovideloans,denyingan",
"eligible individual a loan could have a greater negative impact than providing a loan to a defaulter. In",
"such situations, we consider the variance conditioned on the label, i.e., E[Var[E[e I y, A] I yll, so that",
"wedo not attempttopull thelosses of individuals with different labels (and consequentlydifferent",
"impacts of misclassification) together.",
"5Experiments",
"We first explore the effect of parameter k in MWLD(wk), as discussed in Section 4. We then use",
"(coarse)loss variancetotrain models and showthat we can halvethe loss variance without significant",
"increase in average loss. Table 1 shows a summary of the datasets considered. For more details about",
"these datasets, see Appendix E.",
"5"
] |
{
"arxiv_id": "1906.03518",
"language": "en",
"page_idx": 4
}
|
||
2109.04367-en-0
|
2109.04367-en-0.jpg
|
Multi-granularity Textual Adversarial Attack with Behavior Cloning
Yangyi Chen$^{1, 2* }, \textbf{Jin Su}^{1, 2* }\overset {+ }{\operatorname* { \operatorname* { \operatorname* { \operatorname* { , } } } } }\mathbf{Wei}\textbf{Wei}^{1\dagger }$
$^{1}$Cognitive Computing and Intelligent Information Processing Laboratory, School of Computer
Science and Technology Huazhong University of Science and Tehchnology
$^{2}$School of Software Engineering, Huazhong University of Science and Tehchnology
$\{$yangyichen6666, sujinltenjoy$\}@$gmail.com
weiw@hust.edu.cn
### $\mathbf{Abstract}$
Recently, the textual adversarial attack models become increasingly popular due to their successful in estimating the robustness of NLP models. However, existing works have obvious deficiencies.(1) They usually consider only a single granularity of modification strategies (e.g. word-level or sentence-level), which is insufficient to explore the holistic textual space for generation; (2) They need to query victim models hundreds of times to make a successful attack, which is highly ineffcient in practice. To address such problems, in this paper we propose MAYA, a MultigrAnularit$\hat{\mathbf{Y} } \textbf{ Attack model to effectively gen- }$ erate high-quality adversarial samples with fewer queries to victim models. Furthermore, we propose a reinforcement-learning based method to train a multi-granularity attack agent through behavior cloning with the expert knowledge from our MAYA algorithm to further reduce the query times. Additionally, we also adapt the agent to attack blackbox models that only output labels without confidence scores. We conduct comprehensive experiments to evaluate our attack models by attacking BiLSTM, BERT and RoBERTa in two different black-box attack settings and three benchmark datasets. Experimental results show that our models achieve overall better attacking performance and produce more fluent and grammatical adversarial samples compared to baseline models. Besides, our adversarial attack agent significantly reduces the query times in both attack settings. Our codes are released at https://github. com/Yangyi-Chen/MAYA.
(Chen et al., 2017), and face recognition (Sun et al., 2015). However, these powerful models are vulnerable to adversarial samples, crafted by adding small, human-imperceptible perturbations to the input (Goodfellow et al., 2015; Szegedy et al., 2014). In the domain of computer vision, numerous adversarial attack models have been proposed to benchmark and interpret black-box deep learning models (Dong et al., 2018; Moosavi-Dezfooli et al., 2016; Carlini and Wagner, 2017; Kurakin et al., 2017) and corresponding defense methods have also been proposed to tackle adversarial security issues (Dziugaite et al., 2016; Xie et al., 2018; Kurakin et al., $2017; \text{Tram\`{e} r et al. , 2018}) .$
However, crafting textual adversarial samples is more challenging due to the discrete and nondifferentiable nature of text space. Indeed, most existing works focus on a single granularity of modification strategies, such as sentence-level (Jia and Liang, 2017; Iyyer et al., 2018), word-level (Zang et al., 2020; Ren et al., 2019) or characterlevel (Eger et al., 2019). Thus, none of such attack models find the optimal solution through multi-granularities for launching attacks simultaneously, which is more efficient to generate high-performance and effective adversarial samples while preserving semantic consistency and language fluency. To this end, we propose a simple and novel attack model targeting on multiple kinds of granularity called MAYA, which achieves higher attack success rate with fewer queries to victim models and produces high-quality adversarial samples compared to baseline attack models. Specifically, we add perturbations to the original sentence via rewriting its constituents according to the strict grammatical constraints.
Besides, almost all current attack models need to query victim models hundreds or even thousands of times to launch a successful attack$^{1}$ and
$^{\mathrm{l}}$On average, PSO attack model (Zang et al., 2020) need
## 1 Introduction
Deep learning has been proven to be successful for many real-world applications such as spam filtering (Stringhini et al., 2010), autonomous driving
$^*$Work done during internship at CCIIP
Indicates equal contribution
$^{\dagger}$Corresponding author
to query victim models about 5000 times in SST-2
|
[
[
0.1435546875,
0.0914306640625,
0.85400390625,
0.0914306640625,
0.85400390625,
0.1092529296875,
0.1435546875,
0.1092529296875
],
[
0.3359375,
0.12469482421875,
0.67333984375,
0.12353515625,
0.67333984375,
0.141357421875,
0.3359375,
0.142578125
],
[
0.118408203125,
0.14013671875,
0.884765625,
0.142578125,
0.884765625,
0.1602783203125,
0.118408203125,
0.157958984375
],
[
0.1956787109375,
0.1591796875,
0.80419921875,
0.1602783203125,
0.80419921875,
0.174560546875,
0.1956787109375,
0.17333984375
],
[
0.15869140625,
0.17333984375,
0.84228515625,
0.174560546875,
0.84228515625,
0.1923828125,
0.15869140625,
0.191162109375
],
[
0.32421875,
0.19482421875,
0.68701171875,
0.19482421875,
0.68701171875,
0.208984375,
0.32421875,
0.208984375
],
[
0.403076171875,
0.21142578125,
0.60107421875,
0.21142578125,
0.60107421875,
0.2244873046875,
0.403076171875,
0.2244873046875
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51318359375,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.51123046875,
0.268310546875,
0.8818359375,
0.26953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283203125
],
[
0.145263671875,
0.27783203125,
0.45751953125,
0.27783203125,
0.45751953125,
0.29150390625,
0.145263671875,
0.29150390625
],
[
0.509765625,
0.28515625,
0.8818359375,
0.2861328125,
0.8818359375,
0.300537109375,
0.509765625,
0.2998046875
],
[
0.1435546875,
0.292236328125,
0.459228515625,
0.29150390625,
0.459228515625,
0.305908203125,
0.1435546875,
0.30712890625
],
[
0.1427001953125,
0.305908203125,
0.45751953125,
0.3046875,
0.45751953125,
0.319580078125,
0.1427001953125,
0.320556640625
],
[
0.51318359375,
0.303955078125,
0.880859375,
0.303955078125,
0.880859375,
0.314208984375,
0.51318359375,
0.314208984375
],
[
0.0201568603515625,
0.319580078125,
0.0679931640625,
0.319580078125,
0.0679931640625,
0.4150390625,
0.0201568603515625,
0.4150390625
],
[
0.14697265625,
0.322509765625,
0.45751953125,
0.322509765625,
0.45751953125,
0.333251953125,
0.14697265625,
0.333251953125
],
[
0.51318359375,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.51318359375,
0.33203125
],
[
0.145263671875,
0.3349609375,
0.459228515625,
0.3349609375,
0.459228515625,
0.3486328125,
0.145263671875,
0.3486328125
],
[
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
],
[
0.5146484375,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.363525390625,
0.5146484375,
0.363525390625
],
[
0.1435546875,
0.364013671875,
0.459228515625,
0.363525390625,
0.459228515625,
0.377197265625,
0.1435546875,
0.377685546875
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.14697265625,
0.380126953125,
0.45751953125,
0.380126953125,
0.45751953125,
0.39013671875,
0.14697265625,
0.39013671875
],
[
0.51123046875,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.145263671875,
0.392578125,
0.45751953125,
0.392578125,
0.45751953125,
0.40625,
0.145263671875,
0.40625
],
[
0.51318359375,
0.3984375,
0.880859375,
0.3984375,
0.880859375,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.0352783203125,
0.40673828125,
0.0411376953125,
0.40673828125,
0.0411376953125,
0.412109375,
0.0352783203125,
0.412109375
],
[
0.1435546875,
0.40625,
0.459228515625,
0.40625,
0.459228515625,
0.419189453125,
0.1435546875,
0.419189453125
],
[
0.509765625,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.0277099609375,
0.4091796875,
0.056243896484375,
0.4091796875,
0.056243896484375,
0.4287109375,
0.0277099609375,
0.4287109375
],
[
0.145263671875,
0.42041015625,
0.459228515625,
0.42041015625,
0.459228515625,
0.43408203125,
0.145263671875,
0.43408203125
],
[
0.145263671875,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.448974609375,
0.145263671875,
0.448974609375
],
[
0.509765625,
0.43115234375,
0.8818359375,
0.429931640625,
0.8818359375,
0.44482421875,
0.509765625,
0.4453125
],
[
0.0235137939453125,
0.438232421875,
0.0621337890625,
0.438232421875,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51025390625
],
[
0.145263671875,
0.448974609375,
0.459228515625,
0.448974609375,
0.459228515625,
0.4619140625,
0.145263671875,
0.4619140625
],
[
0.51123046875,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.145263671875,
0.463134765625,
0.459228515625,
0.463134765625,
0.459228515625,
0.476806640625,
0.145263671875,
0.476806640625
],
[
0.51318359375,
0.463134765625,
0.7119140625,
0.463134765625,
0.7119140625,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.14697265625,
0.478515625,
0.456787109375,
0.478515625,
0.456787109375,
0.4892578125,
0.14697265625,
0.4892578125
],
[
0.52880859375,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.49462890625,
0.52880859375,
0.49462890625
],
[
0.145263671875,
0.4912109375,
0.459228515625,
0.4912109375,
0.459228515625,
0.5048828125,
0.145263671875,
0.5048828125
],
[
0.51123046875,
0.496337890625,
0.88330078125,
0.496337890625,
0.88330078125,
0.51123046875,
0.51123046875,
0.51123046875
],
[
0.14697265625,
0.50830078125,
0.45751953125,
0.50830078125,
0.45751953125,
0.5185546875,
0.14697265625,
0.5185546875
],
[
0.0218353271484375,
0.515625,
0.0621337890625,
0.515625,
0.0621337890625,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.509765625,
0.51318359375,
0.880859375,
0.51318359375,
0.880859375,
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
],
[
0.509765625,
0.529296875,
0.880859375,
0.529296875,
0.880859375,
0.5419921875,
0.509765625,
0.5419921875
],
[
0.145263671875,
0.548828125,
0.459228515625,
0.548828125,
0.459228515625,
0.5625,
0.145263671875,
0.5625
],
[
0.51123046875,
0.5458984375,
0.8818359375,
0.5458984375,
0.8818359375,
0.55859375,
0.51123046875,
0.55859375
],
[
0.145263671875,
0.5625,
0.459228515625,
0.5625,
0.459228515625,
0.576171875,
0.145263671875,
0.576171875
],
[
0.51123046875,
0.56005859375,
0.880859375,
0.56005859375,
0.880859375,
0.57373046875,
0.51123046875,
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
],
[
0.14697265625,
0.5927734375,
0.45751953125,
0.5927734375,
0.45751953125,
0.6025390625,
0.14697265625,
0.6025390625
],
[
0.509765625,
0.5927734375,
0.880859375,
0.5927734375,
0.880859375,
0.60546875,
0.509765625,
0.60546875
],
[
0.145263671875,
0.60498046875,
0.4609375,
0.60498046875,
0.4609375,
0.6181640625,
0.145263671875,
0.6181640625
],
[
0.51318359375,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.51318359375,
0.62255859375
],
[
0.145263671875,
0.619140625,
0.45751953125,
0.619140625,
0.45751953125,
0.6328125,
0.145263671875,
0.6328125
],
[
0.509765625,
0.62451171875,
0.880859375,
0.62451171875,
0.880859375,
0.63818359375,
0.509765625,
0.63818359375
],
[
0.14697265625,
0.63525390625,
0.456787109375,
0.63525390625,
0.456787109375,
0.6455078125,
0.14697265625,
0.6455078125
],
[
0.51123046875,
0.64111328125,
0.8818359375,
0.64111328125,
0.8818359375,
0.65478515625,
0.51123046875,
0.65478515625
],
[
0.145263671875,
0.64794921875,
0.459228515625,
0.64794921875,
0.459228515625,
0.6611328125,
0.145263671875,
0.6611328125
],
[
0.145263671875,
0.662109375,
0.459228515625,
0.662109375,
0.459228515625,
0.67578125,
0.145263671875,
0.67578125
],
[
0.51318359375,
0.65771484375,
0.88330078125,
0.65771484375,
0.88330078125,
0.67138671875,
0.51318359375,
0.67138671875
],
[
0.145263671875,
0.6767578125,
0.459228515625,
0.6767578125,
0.459228515625,
0.6904296875,
0.145263671875,
0.6904296875
],
[
0.51318359375,
0.67333984375,
0.8818359375,
0.67333984375,
0.8818359375,
0.6884765625,
0.51318359375,
0.6884765625
],
[
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
],
[
0.145263671875,
0.705078125,
0.45751953125,
0.705078125,
0.45751953125,
0.71875,
0.145263671875,
0.71875
],
[
0.509765625,
0.705078125,
0.880859375,
0.705078125,
0.880859375,
0.7197265625,
0.509765625,
0.7197265625
],
[
0.145263671875,
0.7197265625,
0.459228515625,
0.7197265625,
0.459228515625,
0.73291015625,
0.145263671875,
0.73291015625
],
[
0.51318359375,
0.7216796875,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.51318359375,
0.736328125
],
[
0.145263671875,
0.73291015625,
0.45751953125,
0.73291015625,
0.45751953125,
0.74658203125,
0.145263671875,
0.74658203125
],
[
0.509765625,
0.73681640625,
0.8818359375,
0.73681640625,
0.8818359375,
0.75048828125,
0.509765625,
0.75048828125
],
[
0.1435546875,
0.74755859375,
0.459228515625,
0.74755859375,
0.459228515625,
0.76123046875,
0.1435546875,
0.76123046875
],
[
0.51123046875,
0.75439453125,
0.88330078125,
0.75439453125,
0.88330078125,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.1435546875,
0.76123046875,
0.4609375,
0.76123046875,
0.4609375,
0.77490234375,
0.1435546875,
0.77490234375
],
[
0.145263671875,
0.775390625,
0.353515625,
0.775390625,
0.353515625,
0.7890625,
0.145263671875,
0.7890625
],
[
0.51123046875,
0.77099609375,
0.8818359375,
0.77099609375,
0.8818359375,
0.78369140625,
0.51123046875,
0.78369140625
],
[
0.51123046875,
0.7861328125,
0.8818359375,
0.78515625,
0.8818359375,
0.7998046875,
0.51123046875,
0.80029296875
],
[
0.118408203125,
0.802734375,
0.133544921875,
0.802734375,
0.133544921875,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.1351318359375,
0.80029296875,
0.26025390625,
0.80029296875,
0.26025390625,
0.81396484375,
0.1351318359375,
0.81396484375
],
[
0.5146484375,
0.80419921875,
0.880859375,
0.80419921875,
0.880859375,
0.81396484375,
0.5146484375,
0.81396484375
],
[
0.51123046875,
0.818359375,
0.76416015625,
0.818359375,
0.76416015625,
0.83203125,
0.51123046875,
0.83203125
],
[
0.1175537109375,
0.826171875,
0.48779296875,
0.826171875,
0.48779296875,
0.8388671875,
0.1175537109375,
0.8388671875
],
[
0.5283203125,
0.8349609375,
0.880859375,
0.8349609375,
0.880859375,
0.8486328125,
0.5283203125,
0.8486328125
],
[
0.1175537109375,
0.8427734375,
0.489501953125,
0.8427734375,
0.489501953125,
0.85546875,
0.1175537109375,
0.85546875
],
[
0.1175537109375,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.87158203125,
0.1175537109375,
0.87158203125
],
[
0.509765625,
0.8525390625,
0.880859375,
0.8525390625,
0.880859375,
0.8662109375,
0.509765625,
0.8662109375
],
[
0.51318359375,
0.8681640625,
0.8818359375,
0.8681640625,
0.8818359375,
0.8818359375,
0.51318359375,
0.8818359375
],
[
0.14697265625,
0.87890625,
0.3837890625,
0.87890625,
0.3837890625,
0.892578125,
0.14697265625,
0.892578125
],
[
0.145263671875,
0.892578125,
0.32568359375,
0.892578125,
0.32568359375,
0.90625,
0.145263671875,
0.90625
],
[
0.5322265625,
0.892578125,
0.8818359375,
0.892578125,
0.8818359375,
0.90625,
0.5322265625,
0.90625
],
[
0.1435546875,
0.90380859375,
0.28955078125,
0.90478515625,
0.28955078125,
0.91943359375,
0.1435546875,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.8193359375,
0.90380859375,
0.8193359375,
0.91845703125,
0.5087890625,
0.91943359375
]
] |
[
"Multi-granularity Textual Adversarial Attack with Behavior Cloning",
"Yangyi Chenl,2*, Jin Sul,2* + Wei Weilt",
"1Cognitive Computing and Intelligent Information Processing Laboratory, School of Computer",
"Science and Technology Huazhong University of Science and Tehchnology",
"2School of Software Engineering, Huazhong University of Science and Tehchnology",
"[yangyichen6666, sujinltenjoy] @gmail.com",
"weiw@hust.edu.cn",
"Abstract",
"(Chen et al., 2017), and face recognition (Sun et al.,",
"2015).However, these powerful models are vul-",
"Recently, the textual adversarial attack models",
"nerable to adversarial samples, crafted by adding",
"become increasingly popular due to their suc-",
"cessful in estimating the robustness of NLP",
"small,human-imperceptibleperturbationstothein-",
"120z des",
"models.However, existing works have ob",
"put (Goodfellow et al., 2015; Szegedy et al., 2014),",
"vious deficiencies. (1) They usually consider",
"In the domain of computer vision, numerous adver-",
"only a single granularity of modification strate-",
"sarialattackmodelshavebeenproposedtobench",
"gies (e.g. word-level or sentence-level), which",
"mark and interpret black-box deep learning models",
"is insufficient to explore the holistic textual",
"(Dong et al., 2018: Moosavi-Dezfooli et al., 2016:",
"space for generation; (2) They need to query",
"Carlini and Wagner, 2017; Kurakin et al., 2017)",
"-",
"victim models hundreds of times to make",
"and corresponding defense methods have also been",
"9",
"a successful attack, which is highly ineffi-",
"cient in practice. To address such problems",
"proposed to tackle adversarial security issues (Dzi-",
"['s]",
"in this paper we propose MAYA, a Multi-",
"ugaite et al., 2016; Xie et al., 2018; Kurakin et al..",
"grAnularitY Attack model to effectively gen-",
"2017; Tramer et al., 2018).",
"erate high-quality adversarial samples with",
"However, crafting textual adversarial samples",
"fewer queries to victim models.\"Further-",
"is more challenging due to the discrete and non-",
"more,we propose a reinforcement-learning",
"0060",
"differentiable nature of text space. Indeed, most",
"based method to train a multi-granularity at-",
"tack agent through behavior cloning with the",
"existing works focus on a single granularity of",
"expert knowledge from our MAYA algorithm",
"modification strategies, such as sentence-level (Jia",
"to further reduce the query times. Addition-",
"and Liang, 2017; Iyyer et al., 2018), word-level",
"ally, we also adapt the agent to attack black-",
"(Zang et al., 2020; Ren et al., 2019) or character-",
"box models that only output labels without",
"level (Eger et al., 2019). Thus, none of such",
"confidence scores.We conduct comprehen-",
"attack models find the optimal solution through",
"sive experiments to evaluate our attack models",
"multi-granularities for launching attacks simul-",
"by attackingBiLSTM,BERTandRoBERTa",
"taneously, which is more efficient to generate",
"in two different black-box attack settings and",
"three benchmark datasets. Experimental re-",
"high-performance and effective adversarial sam-",
"sults show that our models achieve overall bet",
"ples while preserving semantic consistency and",
"ter attacking performance and produce more",
"language fluency. To this end, we propose a sim-",
"fluent and grammatical adversarial samples",
" ple and novel attack model targeting on multiple",
"compared to baseline models. Besides, our",
"kinds of granularity called MAYA, which achieves",
"adversarial attack agent significantly reduces",
"higher attack success rate with fewer queries to",
"the query times in both attack settings. Our",
"victim models and produces high-quality adversar-",
"codes are released at https://github.",
"com/Yangyi-Chen/MAYA.",
"ial samples compared to baseline attack models.",
"Specifically, we add perturbations to the original",
"1",
"Introduction",
"sentenceviarewritingitsconstituentsaccordingto",
"the strict grammatical constraints.",
"Deep learning has been proven to be successful for",
"Besides, almost all current attack models need",
"many real-world applications such as spam filter-",
"ing (Stringhini et al., 2010), autonomous driving",
"to query victim models hundreds or even thou",
"sands of times to launch a successful attack' and",
"* Work done during internship at CCIIP",
" Indicates equal contribution",
"'On average, PsO attack model (Zang et al., 2020) need",
"+Corresponding author",
"to query victim models about 5000 times in SST-2"
] |
{
"arxiv_id": "2109.04367",
"language": "en",
"page_idx": 0
}
|
||
2109.04367-en-1
|
2109.04367-en-1.jpg
|
assume the victim models may output the confidence scores of their predictions, which is neither efficient nor practical in real-world situations. To alleviate such problems, we propose to train a multigranularity attack agent called MAYA $_\pi$ through behavior cloning (Torabi et al., 2018) with the expert knowledge from our MAYA algorithm.
We conduct exhaustive experiments including attacking three victim models over three benchmark datasets in two different black-box settings, namely score-based and decision-based attack, to evaluate the effectiveness of our attack models. While the former supposes the labels and the confidence scores of the victim models are available, the latter assumes only the label information can be accessed while the other is unknown, which is more challenging and rarely investigated.
Experimental results demonstrate the superiority of our attack models. Specifically, MAYA overall outperforms all baseline models in terms of attack success rate, attack efficiency, and quality of adversarial samples. MAYA$_\pi$ achieves comparable attack success rate and adversarial samples quality with baseline models while significantly reduces the query times in two black-box settings. Furthermore, we apply MAYA$_\pi$ to attack open-source NLP frameworks to demonstrate its practicality and effectiveness in practice.
To summarize, the main contributions of this
paper are as follows:
$\bullet$ Different from previous works that only concentrate on a single granularity, we propose an effective multi-granularity attack model to generate fluent and grammatical adversarial samples with fewer queries to victim models.
$\cdot$ We propose a RL-based method to train an agent through Behavior Cloning with the expert knowledge from our multi-granularity attack model and demonstrate its efficiency and power in two black-box settings, proving the effectiveness of our adapted imitation algorithm.
$\cdot$ We successfully handle the issues of decisionbased black-box attack, which is rarely investigated in NLP.
## 2 $\textbf{Related Work}$
Existing textual adversarial attack models can be roughly categorized according to the granularity of modification, e.g., character-level, word-level, sentence-level.
Sentence-level attack models often contain paraphrasing original sentences following pre-defined syntax patterns (Iyyer et al., 2018), adding an irrelevant sentence to the end of the passage to distract models (Jia and Liang, 2017), and conducting domain shift on original sentences (Wang et al., 2020). However, sentence-level attacks usually neglect fine-grained granularity, such as word-level, resulting in low attack success rate.
Word-level attack is relatively more investigated and can be modeled as a combinatorial optimization problem (Zang et al., 2020), including finding substitution words and searching for adversarial samples. The methods of finding candidate substitutes mainly focus on the similarity of word embeddings (Jin et al., 2019), WordNet synonyms (Ren et al., 2019), HowNet synonyms (Zang et al., 2020), and Masked Language Model (MLM) (Li et al., 2020). Generally, the search algorithms involve greedy search algorithm (Ren et al., 2019; Liang et al., 2018; Jin et al., 2019), genetic algorithm (Alzantot et al., 2018), and particle swarm optimization (Zang et al., 2020). Although these attack models can achieve relatively high attacking performance, considering only a single granularity restricts the upper bound of word-level attack models’ performance and almost all these models need to query victim models hundreds of times to launch a successful attack.
Character-level attacks make different modifications to words such as swapping, deleting, and inserting characters (Ebrahimi et al., 2018; Belinkov and Bisk, 2018; Gao et al., 2018). These attack models often craft ungrammatical adversarial samples and can be easily defended (Pruthi et al., 2019; Jones et al., 2020). Hence, in this work, we do not incorporate character-level modification into our multi-granularity framework.
To sum up, all above models only consider a single granularity and thus are insufficient in exploring the textual space for generation. So, we propose to launch attacks on multiple granularities in this paper. Experimental results demonstrate the effectiveness and efficiency of our method.
# 3 Methodology
In this section, we first describe our multigranularity attack (MAYA) model in detail. Then we introduce how to train an attack agent, denoted as MAYA$_\pi$,with the knowledge from our MAYA algorithm. Finally, we describe how we adapt
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5283203125,
0.0843505859375,
0.880859375,
0.08612060546875,
0.880859375,
0.1009521484375,
0.5283203125,
0.09857177734375
],
[
0.118408203125,
0.10394287109375,
0.48779296875,
0.10394287109375,
0.48779296875,
0.114013671875,
0.118408203125,
0.114013671875
],
[
0.51123046875,
0.1015625,
0.880859375,
0.1015625,
0.880859375,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.1158447265625,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.490234375,
0.1341552734375,
0.490234375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.509765625,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.509765625,
0.1478271484375
],
[
0.1175537109375,
0.149658203125,
0.491943359375,
0.1484375,
0.491943359375,
0.163330078125,
0.1175537109375,
0.1644287109375
],
[
0.5087890625,
0.1484375,
0.880859375,
0.15087890625,
0.880859375,
0.1644287109375,
0.5087890625,
0.162109375
],
[
0.1175537109375,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51318359375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1158447265625,
0.18115234375,
0.40478515625,
0.1817626953125,
0.40478515625,
0.196533203125,
0.1158447265625,
0.1953125
],
[
0.51318359375,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51318359375,
0.1953125
],
[
0.1368408203125,
0.1995849609375,
0.490234375,
0.1995849609375,
0.490234375,
0.213134765625,
0.1368408203125,
0.213134765625
],
[
0.51318359375,
0.198974609375,
0.88330078125,
0.198974609375,
0.88330078125,
0.2120361328125,
0.51318359375,
0.2120361328125
],
[
0.118408203125,
0.21728515625,
0.486083984375,
0.21728515625,
0.486083984375,
0.22802734375,
0.118408203125,
0.22802734375
],
[
0.509765625,
0.215576171875,
0.77490234375,
0.215576171875,
0.77490234375,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.1158447265625,
0.23095703125,
0.48779296875,
0.232177734375,
0.48779296875,
0.2464599609375,
0.1158447265625,
0.2452392578125
],
[
0.5283203125,
0.23095703125,
0.880859375,
0.23095703125,
0.880859375,
0.2452392578125,
0.5283203125,
0.2452392578125
],
[
0.1175537109375,
0.248779296875,
0.491943359375,
0.248779296875,
0.491943359375,
0.262451171875,
0.1175537109375,
0.262451171875
],
[
0.509765625,
0.2464599609375,
0.8818359375,
0.2476806640625,
0.8818359375,
0.262451171875,
0.509765625,
0.26123046875
],
[
0.1175537109375,
0.26416015625,
0.489501953125,
0.26416015625,
0.489501953125,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.509765625,
0.26318359375,
0.880859375,
0.26318359375,
0.880859375,
0.27783203125,
0.509765625,
0.27783203125
],
[
0.1175537109375,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.1175537109375,
0.29443359375
],
[
0.5087890625,
0.279052734375,
0.8818359375,
0.27978515625,
0.8818359375,
0.29443359375,
0.5087890625,
0.29345703125
],
[
0.1158447265625,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.1158447265625,
0.31005859375
],
[
0.509765625,
0.29638671875,
0.88330078125,
0.29638671875,
0.88330078125,
0.31005859375,
0.509765625,
0.31005859375
],
[
0.1158447265625,
0.312255859375,
0.48779296875,
0.312255859375,
0.48779296875,
0.325927734375,
0.1158447265625,
0.325927734375
],
[
0.509765625,
0.312255859375,
0.88330078125,
0.312255859375,
0.88330078125,
0.325927734375,
0.509765625,
0.325927734375
],
[
0.1175537109375,
0.328857421875,
0.490234375,
0.328857421875,
0.490234375,
0.342529296875,
0.1175537109375,
0.342529296875
],
[
0.51123046875,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.342529296875,
0.51123046875,
0.342529296875
],
[
0.1175537109375,
0.344482421875,
0.349365234375,
0.344482421875,
0.349365234375,
0.359375,
0.1175537109375,
0.359375
],
[
0.51123046875,
0.344482421875,
0.880859375,
0.344482421875,
0.880859375,
0.358154296875,
0.51123046875,
0.358154296875
],
[
0.1351318359375,
0.361083984375,
0.48779296875,
0.35986328125,
0.48779296875,
0.374755859375,
0.1351318359375,
0.3759765625
],
[
0.509765625,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.374755859375,
0.509765625,
0.374755859375
],
[
0.1175537109375,
0.377685546875,
0.48779296875,
0.377685546875,
0.48779296875,
0.391357421875,
0.1175537109375,
0.391357421875
],
[
0.51318359375,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.1175537109375,
0.394287109375,
0.48779296875,
0.394287109375,
0.48779296875,
0.407958984375,
0.1175537109375,
0.407958984375
],
[
0.5087890625,
0.392578125,
0.8818359375,
0.391357421875,
0.8818359375,
0.40625,
0.5087890625,
0.40673828125
],
[
0.1158447265625,
0.410400390625,
0.489501953125,
0.4091796875,
0.489501953125,
0.4228515625,
0.1158447265625,
0.42333984375
],
[
0.51123046875,
0.40673828125,
0.8818359375,
0.407958984375,
0.8818359375,
0.4228515625,
0.51123046875,
0.421630859375
],
[
0.1175537109375,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.1175537109375,
0.440673828125
],
[
0.509765625,
0.424560546875,
0.880859375,
0.424560546875,
0.880859375,
0.438232421875,
0.509765625,
0.438232421875
],
[
0.1158447265625,
0.4423828125,
0.483642578125,
0.4423828125,
0.483642578125,
0.4560546875,
0.1158447265625,
0.4560546875
],
[
0.51318359375,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.1158447265625,
0.457763671875,
0.48779296875,
0.457763671875,
0.48779296875,
0.471435546875,
0.1158447265625,
0.471435546875
],
[
0.51318359375,
0.457763671875,
0.8818359375,
0.457763671875,
0.8818359375,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.1158447265625,
0.474365234375,
0.490234375,
0.474365234375,
0.490234375,
0.488037109375,
0.1158447265625,
0.488037109375
],
[
0.51123046875,
0.473876953125,
0.88330078125,
0.473876953125,
0.88330078125,
0.48681640625,
0.51123046875,
0.48681640625
],
[
0.1158447265625,
0.4892578125,
0.48779296875,
0.490478515625,
0.48779296875,
0.5048828125,
0.1158447265625,
0.50439453125
],
[
0.509765625,
0.4892578125,
0.88330078125,
0.4892578125,
0.88330078125,
0.50439453125,
0.509765625,
0.50439453125
],
[
0.11419677734375,
0.5048828125,
0.48779296875,
0.505859375,
0.48779296875,
0.52099609375,
0.11419677734375,
0.51953125
],
[
0.509765625,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.509765625,
0.51953125
],
[
0.1158447265625,
0.521484375,
0.30322265625,
0.5224609375,
0.30322265625,
0.53759765625,
0.1158447265625,
0.5361328125
],
[
0.51123046875,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.51123046875,
0.53515625
],
[
0.1351318359375,
0.5390625,
0.48779296875,
0.5390625,
0.48779296875,
0.552734375,
0.1351318359375,
0.552734375
],
[
0.509765625,
0.5380859375,
0.7060546875,
0.5380859375,
0.7060546875,
0.548828125,
0.509765625,
0.548828125
],
[
0.1158447265625,
0.55712890625,
0.27197265625,
0.5546875,
0.27197265625,
0.5693359375,
0.1158447265625,
0.57177734375
],
[
0.5283203125,
0.552734375,
0.88330078125,
0.55419921875,
0.88330078125,
0.568359375,
0.5283203125,
0.56787109375
],
[
0.509765625,
0.5693359375,
0.88330078125,
0.57080078125,
0.88330078125,
0.5849609375,
0.509765625,
0.58447265625
],
[
0.12176513671875,
0.576171875,
0.489501953125,
0.576171875,
0.489501953125,
0.5888671875,
0.12176513671875,
0.5888671875
],
[
0.5087890625,
0.5849609375,
0.87890625,
0.5849609375,
0.87890625,
0.5986328125,
0.5087890625,
0.5986328125
],
[
0.1368408203125,
0.5927734375,
0.489501953125,
0.5927734375,
0.489501953125,
0.60546875,
0.1368408203125,
0.60546875
],
[
0.1368408203125,
0.60791015625,
0.490234375,
0.60791015625,
0.490234375,
0.62158203125,
0.1368408203125,
0.62158203125
],
[
0.51318359375,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1351318359375,
0.62255859375,
0.489501953125,
0.62353515625,
0.489501953125,
0.63818359375,
0.1351318359375,
0.63720703125
],
[
0.51318359375,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.1368408203125,
0.64013671875,
0.408935546875,
0.64013671875,
0.408935546875,
0.65380859375,
0.1368408203125,
0.65380859375
],
[
0.51318359375,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.51318359375,
0.6484375
],
[
0.51318359375,
0.65185546875,
0.87890625,
0.65185546875,
0.87890625,
0.662109375,
0.51318359375,
0.662109375
],
[
0.1234130859375,
0.66015625,
0.486083984375,
0.66015625,
0.486083984375,
0.6748046875,
0.1234130859375,
0.6748046875
],
[
0.51123046875,
0.66650390625,
0.880859375,
0.66650390625,
0.880859375,
0.6796875,
0.51123046875,
0.6796875
],
[
0.1368408203125,
0.677734375,
0.489501953125,
0.677734375,
0.489501953125,
0.69140625,
0.1368408203125,
0.69140625
],
[
0.51318359375,
0.68408203125,
0.72705078125,
0.68408203125,
0.72705078125,
0.69775390625,
0.51318359375,
0.69775390625
],
[
0.1376953125,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.1376953125,
0.7080078125
],
[
0.5283203125,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.5283203125,
0.71337890625
],
[
0.1376953125,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.1376953125,
0.7236328125
],
[
0.509765625,
0.71630859375,
0.880859375,
0.71533203125,
0.880859375,
0.72998046875,
0.509765625,
0.73095703125
],
[
0.1368408203125,
0.72705078125,
0.48779296875,
0.72705078125,
0.48779296875,
0.74072265625,
0.1368408203125,
0.74072265625
],
[
0.51318359375,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.1368408203125,
0.7421875,
0.376220703125,
0.7421875,
0.376220703125,
0.755859375,
0.1368408203125,
0.755859375
],
[
0.51318359375,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.12176513671875,
0.76123046875,
0.490234375,
0.76318359375,
0.490234375,
0.77783203125,
0.12176513671875,
0.775390625
],
[
0.509765625,
0.765625,
0.8818359375,
0.765625,
0.8818359375,
0.779296875,
0.509765625,
0.779296875
],
[
0.1351318359375,
0.779296875,
0.495361328125,
0.779296875,
0.495361328125,
0.7919921875,
0.1351318359375,
0.7919921875
],
[
0.509765625,
0.78076171875,
0.82958984375,
0.78076171875,
0.82958984375,
0.79443359375,
0.509765625,
0.79443359375
],
[
0.1351318359375,
0.7958984375,
0.2418212890625,
0.79345703125,
0.2418212890625,
0.80810546875,
0.1351318359375,
0.81005859375
],
[
0.509765625,
0.81005859375,
0.65576171875,
0.81103515625,
0.65576171875,
0.82666015625,
0.509765625,
0.826171875
],
[
0.1158447265625,
0.82666015625,
0.268798828125,
0.82666015625,
0.268798828125,
0.84130859375,
0.1158447265625,
0.84130859375
],
[
0.51123046875,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.5087890625,
0.85546875,
0.8818359375,
0.85498046875,
0.8818359375,
0.869140625,
0.5087890625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.509765625,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90625,
0.2283935546875,
0.90625,
0.2283935546875,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"assume the victim models may output the confi-",
"Sentence-level attack models often contain para-",
"dence scores of their predictions,which is neither",
"phrasing original sentences following pre-defined",
"efficient nor practical in real-world situations. To",
"syntax patterns (lyyer et al., 2018), adding an ir",
"alleviate such problems, we propose to train a multi-",
"relevant sentence to the end of the passage to dis",
"granularity attack agent called MAYA through be-",
"tract models (Jia and Liang, 2017), and conducting",
"havior cloning (Torabi et al., 2018) with the expert",
"domain shift on original sentences (Wang et al.",
"knowledge from our MAYA algorithm.",
"2020).However, sentence-level attacks usually ne",
"We conduct exhaustive experiments including at-",
"glect fine-grained granularity, such as word-level,",
"tackingthreevictimmodelsoverthreebenchmark",
"resulting in low attack success rate.",
"datasets in two different black-box settings, namely",
"Word-level attack is relatively more investigated",
"score-based and decision-based attack, to evalu-",
"and can be modeled as a combinatorial optimiza-",
"ate the effectiveness of our attack models. While",
"tion problem (Zang et al., 2020), including finding",
"the former supposes the labels and the confidence",
"substitution words and searching for adversarial",
"scores of the victim models are available, the latter",
" samples. The methods of finding candidate substi-",
"assumes only the label information can be accessed",
"tutes mainly focus on the similarity of word em-",
"while the other is unknown, which is more chal-",
"beddings (Jin et al., 2019), WordNet synonyms",
"lenging and rarely investigated.",
"(Ren et al., 2019), HowNet synonyms (Zang et al..",
"Experimental results demonstrate the superiority",
"2020), and Masked Language Model (MLM) (Li",
"of our attack models. Specifically, MAYA overall",
"et al., 2020). Generally, the search algorithms in-",
"outperforms all baseline models in terms of attack",
"volve greedy search algorithm (Ren et al., 2019;",
"success rate, attack efficiency,and quality of ad-",
"Liang et al., 2018; Jin et al., 2019), genetic algo-",
"versarial samples. MAYA, achieves comparable",
"rithm (Alzantot et al., 2018), and particle swarm",
"attack success rate and adversarial samples quality",
"optimization (Zang et al., 2020). Although these",
"with baseline models while significantly reduces",
"attack models can achieve relatively high attacking",
"the query times in two black-box settings. Fur-",
"performance, considering only a single granular-",
"thermore, we apply MAYA, to attack open-source",
"ity restricts the upper bound of word-level attack",
" NLP frameworks to demonstrate its practicality and",
"models' performance and almost all these models",
"effectiveness in practice.",
"need to query victim models hundreds of times to",
"To summarize, the main contributions of this",
"launchasuccessfulattack",
"paper are as follows:",
"Character-level attacks make different modifica-",
"tions to words such as swapping, deleting, and in-",
": Different from previous works that only con-",
" serting characters (Ebrahimi et al., 2018: Belinkov",
"centrate on a single granularity, we propose an",
"effective multi-granularity attack model to gen-",
"and Bisk, 2018; Gao et al., 2018). These attack",
"erate fluent and grammatical adversarial samples",
"models often craft ungrammatical adversarial sam",
"with fewer queries to victim models.",
"ples and can be easily defended (Pruthi et al., 2019;",
"Jones etal..2020).Hence.jinthiswork.we.donot",
". We propose a RL-based method to train an agent",
"incorporate character-level modification into our",
"through Behavior Cloning with the expert knowl-",
"multi-granularity framework.",
"edge from our multi-granularity attack model",
"To sum up, all above models only consider a",
"and demonstrate its efficiency and power in two",
"single granularity and thus are insufficient in ex",
"black-box settings, proving the effectiveness of",
"ploring the textual space for generation. So, we",
"our adapted imitation algorithm.",
"propose to launch attacks on multiple granularities",
". We successfully handle the issues of decision-",
"in this paper. Experimental results demonstrate the",
"based black-box attack, which is rarely investi-",
"effectiveness and efficiency of our method.",
"gated in NLP.",
"3Methodology",
"2Related Work",
"In this section, we first describe our multi-",
"Existing textual adversarial attack models can be",
"granularity attack (MAYA) model in detail. Then",
"roughly categorized according to the granularity",
"we introduce how to train an attack agent, denoted",
"of modification, e.g, character-level, word-level,",
"as MAYA, with the knowledge from our MAYA",
"sentence-level.",
"algorithm. Finally, we describe how we adapt"
] |
{
"arxiv_id": "2109.04367",
"language": "en",
"page_idx": 1
}
|
||
2109.04367-en-3
|
2109.04367-en-3.jpg
|
<FigureHere>
Figure l: (Left) The training process of MAYA $_{\pi\cdot}$ We treat the selection of candidates as a multi-class classification problem with supervisory signals from our MAYA algorithm. (Middle) The architecture and workflow of MAYA$_\pi.$ (Right) The process to launch an adversarial attack against victim models.
adversarial sample. Otherwise, we view $S^{\prime}$ as the most potential candidate and return to the Generate step.
the one that causes the biggest drop in the confidence score as the most potential candidate, and return to the Generate step.
3.2 Combined with Behavior Cloning
Second, when $S^{\prime}$ comes from $V_s$,we follow the same procedure as in the Pick step in MAYA algorithm that iteratively substitutes $[MASK]$ token with candidate words and query the victim model for confidence scores. If one successful candidate found, we directly return this sentence as the final adversarial sample. Otherwise, we view the candidate that causes the biggest drop in the confidence score as the most potential candidate and return to the Generate step. The whole process will be repeated until a successful adversarial sample found or all potential candidates have been encountered before. In the next subsection, we will describe how we adapt this score-based Pick step to launch a decision-based black-box attack.
As seen in Figure 1, we use BERT$_base$ (Devlin et al., 2019) and a linear classifier with one output unit as the architecture of MAYA$_\pi.$ The core function of $MAYA_{\pi} is to predict the most potential candidate$ without querying the victim model. In this section, We first describe how we exploit MAYA$_\pi$ to launch an adversarial attack because we require MAYA$_\pi$ to perform the full procedure of attacking in the training process. And then we detail the training process.
3.2.1 Launch an Adversarial Attack
## 3.2.2 Training Process
In this subsection, we describe our RL-based method to train MAYA$_\pi$ through Behavior Cloning with the expert knowledge from our MAYA algorithm. Specifically, we improve the training process by adapting the Dataset Aggregation (DAGGER) method (Ross et al., 2011). The training process incorporates three parts, namely initialization, sampling trajectories, and training.
Now assume that we have already trained an attack agent MAYA$_\pi.$ Given the input sentence $S$,we follow the same procedure as in the Generate step in MAYA algorithm to generate adversarial candidate sets $V_p$ and $V_s$, corresponding to two different generation processes. Then, with the original sentence $S$ and an adversarial candidate $S_i$ concatenated as the input, MAYA$_\pi$ will output a score as a measure of tendency to choose this specific adversarial candidate. We obtain the candidate $S^{\prime}$ that get the highest score. Similarly, there are two different $cases.$
First, when $S^{\prime}$ comes from $V_p$,we directly use $S^{\prime}$ to query the victim model. If it successfully fools the victim model, we return $S^{\prime}$ as the final
Initialization We initialize MAYA$_\pi$ with pretrained weights from BERT (Devlin et al., 2019) and a random initialized MLP. Also, we initialize
|
[
[
0.32666015625,
0.0955810546875,
0.344970703125,
0.0955810546875,
0.344970703125,
0.09857177734375,
0.32666015625,
0.09857177734375
],
[
0.462646484375,
0.1015625,
0.5341796875,
0.1015625,
0.5341796875,
0.105712890625,
0.462646484375,
0.105712890625
],
[
0.491943359375,
0.1134033203125,
0.50146484375,
0.10986328125,
0.50537109375,
0.11639404296875,
0.496337890625,
0.119384765625
],
[
0.66162109375,
0.1478271484375,
0.7314453125,
0.1478271484375,
0.7314453125,
0.157958984375,
0.66162109375,
0.157958984375
],
[
0.2166748046875,
0.177001953125,
0.310546875,
0.177001953125,
0.310546875,
0.18701171875,
0.2166748046875,
0.18701171875
],
[
0.2149658203125,
0.1923828125,
0.258544921875,
0.1923828125,
0.258544921875,
0.203125,
0.2149658203125,
0.203125
],
[
0.306396484375,
0.19482421875,
0.330078125,
0.19482421875,
0.330078125,
0.201904296875,
0.306396484375,
0.201904296875
],
[
0.235107421875,
0.21142578125,
0.316650390625,
0.21142578125,
0.316650390625,
0.221435546875,
0.235107421875,
0.221435546875
],
[
0.2509765625,
0.2452392578125,
0.298828125,
0.2452392578125,
0.298828125,
0.257080078125,
0.2509765625,
0.257080078125
],
[
0.63232421875,
0.267333984375,
0.68603515625,
0.267333984375,
0.68603515625,
0.27490234375,
0.63232421875,
0.27490234375
],
[
0.56689453125,
0.28076171875,
0.61279296875,
0.283203125,
0.611328125,
0.29443359375,
0.56494140625,
0.292236328125
],
[
0.705078125,
0.283935546875,
0.75390625,
0.283935546875,
0.75390625,
0.2880859375,
0.705078125,
0.2880859375
],
[
0.48779296875,
0.29150390625,
0.50146484375,
0.281982421875,
0.51318359375,
0.290283203125,
0.49951171875,
0.2998046875
],
[
0.455078125,
0.2998046875,
0.5458984375,
0.2998046875,
0.5458984375,
0.311279296875,
0.455078125,
0.311279296875
],
[
0.6513671875,
0.2998046875,
0.744140625,
0.2998046875,
0.744140625,
0.311279296875,
0.6513671875,
0.311279296875
],
[
0.1175537109375,
0.336181640625,
0.8818359375,
0.336181640625,
0.8818359375,
0.349853515625,
0.1175537109375,
0.349853515625
],
[
0.1175537109375,
0.350830078125,
0.88330078125,
0.350830078125,
0.88330078125,
0.364013671875,
0.1175537109375,
0.364013671875
],
[
0.118408203125,
0.366455078125,
0.61376953125,
0.366455078125,
0.61376953125,
0.377197265625,
0.118408203125,
0.377197265625
],
[
0.1175537109375,
0.405029296875,
0.490234375,
0.405029296875,
0.490234375,
0.418701171875,
0.1175537109375,
0.418701171875
],
[
0.497802734375,
0.40380859375,
0.880859375,
0.40380859375,
0.880859375,
0.418701171875,
0.497802734375,
0.418701171875
],
[
0.1158447265625,
0.419189453125,
0.490234375,
0.42041015625,
0.490234375,
0.435302734375,
0.1158447265625,
0.43408203125
],
[
0.50732421875,
0.42041015625,
0.88330078125,
0.419189453125,
0.88330078125,
0.43408203125,
0.50732421875,
0.435302734375
],
[
0.11419677734375,
0.435791015625,
0.32421875,
0.43701171875,
0.32421875,
0.451904296875,
0.11419677734375,
0.45068359375
],
[
0.51123046875,
0.439453125,
0.57861328125,
0.439453125,
0.57861328125,
0.45068359375,
0.51123046875,
0.45068359375
],
[
0.5263671875,
0.45361328125,
0.880859375,
0.45361328125,
0.880859375,
0.468505859375,
0.5263671875,
0.468505859375
],
[
0.1158447265625,
0.46435546875,
0.424072265625,
0.465576171875,
0.424072265625,
0.479736328125,
0.1158447265625,
0.478515625
],
[
0.509765625,
0.47021484375,
0.88330078125,
0.4697265625,
0.88330078125,
0.48388671875,
0.509765625,
0.485107421875
],
[
0.1175537109375,
0.48681640625,
0.49365234375,
0.48681640625,
0.49365234375,
0.50048828125,
0.1175537109375,
0.50048828125
],
[
0.51123046875,
0.48681640625,
0.8818359375,
0.48681640625,
0.8818359375,
0.50048828125,
0.51123046875,
0.50048828125
],
[
0.1158447265625,
0.501953125,
0.489501953125,
0.5029296875,
0.489501953125,
0.51708984375,
0.1158447265625,
0.5166015625
],
[
0.51123046875,
0.5029296875,
0.880859375,
0.5029296875,
0.880859375,
0.51708984375,
0.51123046875,
0.51708984375
],
[
0.1175537109375,
0.51953125,
0.48779296875,
0.51953125,
0.48779296875,
0.533203125,
0.1175537109375,
0.533203125
],
[
0.51123046875,
0.5185546875,
0.880859375,
0.5185546875,
0.880859375,
0.5322265625,
0.51123046875,
0.5322265625
],
[
0.1175537109375,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.1175537109375,
0.5498046875
],
[
0.51318359375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.1175537109375,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.1175537109375,
0.5654296875
],
[
0.51123046875,
0.5517578125,
0.87744140625,
0.5517578125,
0.87744140625,
0.5625,
0.51123046875,
0.5625
],
[
0.1175537109375,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.51318359375,
0.56787109375,
0.8818359375,
0.56787109375,
0.8818359375,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.1175537109375,
0.58447265625,
0.484375,
0.58447265625,
0.484375,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.51318359375,
0.5849609375,
0.880859375,
0.5849609375,
0.880859375,
0.595703125,
0.51318359375,
0.595703125
],
[
0.1158447265625,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.1158447265625,
0.61328125
],
[
0.51123046875,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.6123046875,
0.51123046875,
0.6123046875
],
[
0.1158447265625,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.1158447265625,
0.6298828125
],
[
0.509765625,
0.6162109375,
0.8818359375,
0.6142578125,
0.8818359375,
0.62744140625,
0.509765625,
0.6298828125
],
[
0.11248779296875,
0.6328125,
0.177978515625,
0.63134765625,
0.1796875,
0.6455078125,
0.11419677734375,
0.64794921875
],
[
0.51318359375,
0.6318359375,
0.8818359375,
0.6318359375,
0.8818359375,
0.6455078125,
0.51318359375,
0.6455078125
],
[
0.51318359375,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.66015625,
0.51318359375,
0.66015625
],
[
0.1158447265625,
0.65673828125,
0.411376953125,
0.65673828125,
0.411376953125,
0.67041015625,
0.1158447265625,
0.67041015625
],
[
0.509765625,
0.66455078125,
0.88330078125,
0.66455078125,
0.88330078125,
0.677734375,
0.509765625,
0.677734375
],
[
0.1175537109375,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.6923828125,
0.1175537109375,
0.6923828125
],
[
0.509765625,
0.6796875,
0.767578125,
0.6796875,
0.767578125,
0.693359375,
0.509765625,
0.693359375
],
[
0.11419677734375,
0.6953125,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.11419677734375,
0.708984375
],
[
0.1158447265625,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.1158447265625,
0.724609375
],
[
0.51123046875,
0.708984375,
0.7021484375,
0.708984375,
0.7021484375,
0.72265625,
0.51123046875,
0.72265625
],
[
0.1175537109375,
0.72705078125,
0.48779296875,
0.72705078125,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74365234375,
0.509765625,
0.74365234375
],
[
0.1175537109375,
0.74365234375,
0.489501953125,
0.74365234375,
0.489501953125,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.5087890625,
0.7451171875,
0.8818359375,
0.74658203125,
0.8818359375,
0.76123046875,
0.5087890625,
0.76025390625
],
[
0.1158447265625,
0.75830078125,
0.486083984375,
0.75830078125,
0.486083984375,
0.771484375,
0.1158447265625,
0.771484375
],
[
0.5087890625,
0.7626953125,
0.88330078125,
0.76123046875,
0.88330078125,
0.775390625,
0.5087890625,
0.77685546875
],
[
0.118408203125,
0.77490234375,
0.489501953125,
0.77490234375,
0.489501953125,
0.7880859375,
0.118408203125,
0.7880859375
],
[
0.509765625,
0.77685546875,
0.8818359375,
0.779296875,
0.8818359375,
0.79345703125,
0.509765625,
0.79150390625
],
[
0.1175537109375,
0.79150390625,
0.490234375,
0.79150390625,
0.490234375,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.7958984375,
0.8818359375,
0.7958984375,
0.8818359375,
0.80859375,
0.51123046875,
0.80859375
],
[
0.1158447265625,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.81005859375,
0.8818359375,
0.81103515625,
0.8818359375,
0.826171875,
0.509765625,
0.82470703125
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.51123046875,
0.82763671875,
0.88330078125,
0.82763671875,
0.88330078125,
0.84130859375,
0.51123046875,
0.84130859375
],
[
0.11419677734375,
0.83837890625,
0.48779296875,
0.8388671875,
0.48779296875,
0.85400390625,
0.11419677734375,
0.8525390625
],
[
0.51123046875,
0.84326171875,
0.73876953125,
0.84326171875,
0.73876953125,
0.85791015625,
0.51123046875,
0.85791015625
],
[
0.1175537109375,
0.85693359375,
0.162109375,
0.85693359375,
0.162109375,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.133544921875,
0.869140625,
0.490234375,
0.869140625,
0.490234375,
0.88720703125,
0.133544921875,
0.88720703125
],
[
0.51123046875,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1158447265625,
0.88525390625,
0.489501953125,
0.8857421875,
0.489501953125,
0.90380859375,
0.1158447265625,
0.90283203125
],
[
0.509765625,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90380859375,
0.486083984375,
0.90380859375,
0.486083984375,
0.91748046875,
0.1158447265625,
0.91748046875
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"ae",
"ModifiedSantenc",
"?",
"Victim Model",
"[yyyy",
"00",
"0",
"Supervisory Signal",
"MAYA",
"Attackfailure",
"MAYA",
"ttarkSucre",
"7",
"Original Sentence",
"Adversarial Sentence",
"Figure l: (Left) The training process of MAYA. We treat the selection of candidates as a multi-class classification",
"problem with supervisory signals from our MAYA algorithm. (Middle) The architecture and workflow of MAYA元.",
"(Right)Theprocess tolaunchan adversarial attackagainst victimmodels.",
"the one that causes the biggest drop in the confi-",
"adversarial sample. Otherwise, we view S' as the",
"dence score as the most potential candidate, and",
" most potential candidate and return to the Gener-",
"return to the Generate step.",
"ate step.",
"Second, when S' comes from Vs, we follow the",
"3.2Combined with Behavior Cloning",
"same procedure as in the Pick step in MAYA al-",
"As seen in Figure 1, we use BERTbase (Devlin et al.,",
"gorithm that iteratively substitutes [M ASK| token",
"2019) and a linear classifier with one output unit as",
"with candidate words and query the victim model",
"the architecture of MAYA. The core function of",
"for confidence scores. If one successful candidate",
"MAYA is to predict the most potential candidate",
"found, we directly return this sentence as the final",
"without querying the victim model. In this section,",
"adversarialsample.Otherwise,weviewthecandi",
"We first describe how we exploit MAYA to launch",
"date that causes the biggest drop in the confidence",
"an adversarial attack because we require MAYA",
"scoreasthemostpotentialcandidateandreturnto",
"to perform the full procedure of attacking in the",
"the Generate step. The whole process will be re-",
"training process. And then we detail the training",
"peated until a successful adversarial sample found",
"process.",
"or all potential candidates have been encountered",
"before.Inthenext subsection,wewilldescribe",
"3.2.1 Launch an Adversarial Attack",
"how we adapt this score-based Pick step to launch",
"Now assume that we have already trained an attack",
"a decision-based black-box attack.",
"agent MAYA. Given the input sentence S, we fol-",
"low the same procedure as in the Generate step in",
"3.2.2 Training Process",
"MAYA algorithm to generate adversarial candidate",
"In this subsection, we describe our RL-based",
"sets Vp and Vs, corresponding to two different gen-",
"method to train MAYA through Behavior Cloning",
"eration processes. Then, with the original sentence",
"with the expert knowledge from our MAYA algo-",
"S and an adversarial candidate S; concatenated as",
"rithm. Specifically,weimprovethe training process",
"the input, MAYA, will output a score as a mea-",
"by adapting the Dataset Aggregation (DAGGER)",
"sure of tendency to choose this specific adversarial",
"method (Ross et al., 2011). The training process",
"candidate. We obtain the candidate S' that get the",
"incorporates three parts, namely initialization, sam-",
"highest score. Similarly, there are two different",
"pling trajectories, and training.",
"cases.",
"First, when S' comes from Vp, we directly use",
"Initialization We initialize MAYA with pre-",
"S' to query the victim model. If it successfully",
"trained weights from BERT (Devlin et al., 2019)",
"fools the victim model, we return s' as the final",
" and a random initialized MLP. Also, we initialize"
] |
{
"arxiv_id": "2109.04367",
"language": "en",
"page_idx": 3
}
|
||
2109.04367-en-4
|
2109.04367-en-4.jpg
|
an empty trajectory dataset $D.$
Sampling Trajectories To train MAYA$_\pi$, we need to interact with the victim model to obtain the training data. Specifically, we train a local victim model that has the same architecture with the target victim model, expecting to approximate the decision boundary of the target victim model$^{4}.$
We sample a batch of original sentences. For each sentence $S_0$, we generate adversarial candidates $C_1,...,C_k.$ As in the Verify and Pick steps in MAYA algorithm, one specific candidate will be chosen as the final successful adversarial sample or the most potential candidate. We view the candidate chosen by MAYA algorithm as the ground truth label and add $((S_0,C_1,...,C_k),label)$ to our trajectory dataset $D.$
To fully train an agent that can tackle different situations, we need a large dataset $D.$ So, we adapt DAGGER method. Specifically, when receiving the ground truth label from MAYA algorithm, MAYA$_\pi$ doesn't take the golden action indicated by MAYA. It will take the action based on its own prediction. That is, MAYA$_\pi$ will predict which candidate will most confuse the victim model and follow its own procedure of launching an adversarial attack. The predicted candidate will be treated as $S_{0}$ and we will continue the same process of sampling trajectories to augment the dataset $D.$
Training Then we train MAYA$_\pi$ for only one epoch using the trajectory dataset $D.$ We model the training task as a multi-class classification problem. For each sample $((S_0,C_1,...,C_k),label)$ drawn from $D$,we concatenate $S_0$ with each $C_i.$ Then we input the $k$ concatenated sentences to MAYA$_\pi$ to get $k$ scores. We treat these $k$ scores as logits and use the cross-entropy loss to train MAYA$_\pi.$ After the training process, we clear $D$ and continue the sampling procedure. The implementation details are described in Appendix I.
## 33 Adapted to Decision-based Attack
To adapt MAYA$_\pi$ to decision-based attack, we only need to modify one step in the attack procedure described in Section 3.2.1 while keep other steps unchanged. Specifically, when the candidate $S^{\prime}$ that gets the highest score from MAYA$_\pi$ is from $V_s$, we iteratively substitute $[MASK]$ token with candidate words to generate adversarial candidates
$^{4}$It is unpractical to query the target victim model thousands of times to train an agent that is employed to attack this victim model.
and query the victim model for decisions. If one candidate successfully flips the label, we treat it as the final adversarial sample. Otherwise, to generate adversarial samples more effciently, we take the candidate whose sentence embedding has the lowest cosine similarity with the sentence embedding of the original sentence as the most potential candidate. Our intuition is that the candidate that least resembles the original sentences is more likely to be a successful adversarial sample.
# 4 Experiments
We conduct comprehensive experiments to evaluate our attack models on the tasks of sentiment analysis, natural language inference, and news classification.
## 4.1 Datasets and Victim Models
For sentiment analysis, we choose SST-2 (Socher et al., 2013), a binary sentiment classification benchmark dataset. For natural language inference, we choose mismatched MNLI dataset (Williams et al., 2018). For news classification, we choose AG's News dataset. The models need to choose one of the four classes including World, Sports, Business, and Sci/Tech, given an instance in AG's News (Zhang et al., 2015).
We evaluate our attack models by attacking three victim models including BiLSTM (Schuster and Paliwal, 1997), BERT (Devlin et al., 2019), and RoBERTa (Liu et al., 2019). Details of the datasets and the classification accuracy of victim models are listed in Table 1.
## 4.2 Attack Models
We implement all baseline attack models using the NLP attack package TextAttack (Morris et al., 2020) and OpenAttack (Zeng et al., 2021).
### Score-based Attack Models
We comprehensively compare our score-based attack models with fve representative and strong score-based attack models including (1) GA+Embedding (Alzantot et al., 2018), $(2){\mathrm{~PWWS+Synonym~(Ren~et~al.,~2019),~(3)}}$ PSO+Sememe (Zang et al., 2020), (4) TextFooler (Jin et al., 2019), (5) BERT-Attack (Li et al., 2020). Details of baseline models are listed in Appendix B and we describe details of our attack models below. MAYA We select 3 open-source paraphrase models as the building blocks of MAYA. Specifically,
|
[
[
0.1158447265625,
0.08612060546875,
0.340087890625,
0.0843505859375,
0.340087890625,
0.09857177734375,
0.1158447265625,
0.1009521484375
],
[
0.51123046875,
0.08489990234375,
0.87890625,
0.08489990234375,
0.87890625,
0.09857177734375,
0.51123046875,
0.09857177734375
],
[
0.5146484375,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.5146484375,
0.114013671875
],
[
0.1175537109375,
0.10986328125,
0.48779296875,
0.10986328125,
0.48779296875,
0.12353515625,
0.1175537109375,
0.12353515625
],
[
0.51318359375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.1175537109375,
0.1270751953125,
0.486083984375,
0.1270751953125,
0.486083984375,
0.14013671875,
0.1175537109375,
0.14013671875
],
[
0.51318359375,
0.1341552734375,
0.880859375,
0.1341552734375,
0.880859375,
0.1478271484375,
0.51318359375,
0.1478271484375
],
[
0.1175537109375,
0.142578125,
0.489501953125,
0.142578125,
0.489501953125,
0.1561279296875,
0.1175537109375,
0.1561279296875
],
[
0.5146484375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.162109375,
0.5146484375,
0.162109375
],
[
0.1175537109375,
0.1602783203125,
0.486083984375,
0.1602783203125,
0.486083984375,
0.17041015625,
0.1175537109375,
0.17041015625
],
[
0.51318359375,
0.166259765625,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.11419677734375,
0.174560546875,
0.48779296875,
0.17333984375,
0.48779296875,
0.188232421875,
0.11419677734375,
0.189453125
],
[
0.5146484375,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19482421875,
0.5146484375,
0.19482421875
],
[
0.1175537109375,
0.191162109375,
0.459228515625,
0.191162109375,
0.459228515625,
0.204833984375,
0.1175537109375,
0.204833984375
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1368408203125,
0.207275390625,
0.48779296875,
0.207275390625,
0.48779296875,
0.2203369140625,
0.1368408203125,
0.2203369140625
],
[
0.509765625,
0.21435546875,
0.880859375,
0.21435546875,
0.880859375,
0.22802734375,
0.509765625,
0.22802734375
],
[
0.1158447265625,
0.223876953125,
0.48779296875,
0.223876953125,
0.48779296875,
0.2369384765625,
0.1158447265625,
0.2369384765625
],
[
0.509765625,
0.2298583984375,
0.77099609375,
0.23095703125,
0.77099609375,
0.2452392578125,
0.509765625,
0.24462890625
],
[
0.1175537109375,
0.2392578125,
0.486083984375,
0.2392578125,
0.486083984375,
0.254150390625,
0.1175537109375,
0.254150390625
],
[
0.1175537109375,
0.254638671875,
0.48779296875,
0.254638671875,
0.48779296875,
0.26953125,
0.1175537109375,
0.26953125
],
[
0.51318359375,
0.26123046875,
0.54248046875,
0.26123046875,
0.54248046875,
0.26953125,
0.51318359375,
0.26953125
],
[
0.54150390625,
0.260009765625,
0.6533203125,
0.260009765625,
0.6533203125,
0.271484375,
0.54150390625,
0.271484375
],
[
0.1175537109375,
0.271484375,
0.48779296875,
0.271484375,
0.48779296875,
0.28515625,
0.1175537109375,
0.28515625
],
[
0.1175537109375,
0.287353515625,
0.490234375,
0.287353515625,
0.490234375,
0.300537109375,
0.1175537109375,
0.300537109375
],
[
0.51318359375,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.118408203125,
0.3046875,
0.486083984375,
0.3046875,
0.486083984375,
0.3154296875,
0.118408203125,
0.3154296875
],
[
0.51123046875,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.314208984375,
0.51123046875,
0.314208984375
],
[
0.1175537109375,
0.319580078125,
0.489501953125,
0.319580078125,
0.489501953125,
0.334228515625,
0.1175537109375,
0.334228515625
],
[
0.51123046875,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.328857421875,
0.51123046875,
0.328857421875
],
[
0.1175537109375,
0.336181640625,
0.270263671875,
0.336181640625,
0.270263671875,
0.349853515625,
0.1175537109375,
0.349853515625
],
[
0.51318359375,
0.334228515625,
0.5859375,
0.334228515625,
0.5859375,
0.345703125,
0.51318359375,
0.345703125
],
[
0.1351318359375,
0.350830078125,
0.484375,
0.350830078125,
0.484375,
0.364013671875,
0.1351318359375,
0.364013671875
],
[
0.51123046875,
0.358154296875,
0.77490234375,
0.358154296875,
0.77490234375,
0.371826171875,
0.51123046875,
0.371826171875
],
[
0.1175537109375,
0.3681640625,
0.486083984375,
0.3681640625,
0.486083984375,
0.3818359375,
0.1175537109375,
0.3818359375
],
[
0.1175537109375,
0.38427734375,
0.48779296875,
0.38427734375,
0.48779296875,
0.39794921875,
0.1175537109375,
0.39794921875
],
[
0.509765625,
0.380126953125,
0.8818359375,
0.37890625,
0.8818359375,
0.393798828125,
0.509765625,
0.394287109375
],
[
0.118408203125,
0.399658203125,
0.486083984375,
0.399658203125,
0.486083984375,
0.41455078125,
0.118408203125,
0.41455078125
],
[
0.51123046875,
0.396728515625,
0.8818359375,
0.396728515625,
0.8818359375,
0.410888671875,
0.51123046875,
0.410888671875
],
[
0.1175537109375,
0.416259765625,
0.489501953125,
0.416259765625,
0.489501953125,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.509765625,
0.413330078125,
0.88330078125,
0.413330078125,
0.88330078125,
0.427001953125,
0.509765625,
0.427001953125
],
[
0.118408203125,
0.43408203125,
0.48779296875,
0.43408203125,
0.48779296875,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.509765625,
0.4287109375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.509765625,
0.4423828125
],
[
0.1158447265625,
0.448974609375,
0.489501953125,
0.448974609375,
0.489501953125,
0.4619140625,
0.1158447265625,
0.4619140625
],
[
0.51318359375,
0.4453125,
0.8818359375,
0.4453125,
0.8818359375,
0.458984375,
0.51318359375,
0.458984375
],
[
0.1158447265625,
0.46435546875,
0.48779296875,
0.46435546875,
0.48779296875,
0.47802734375,
0.1158447265625,
0.47802734375
],
[
0.51318359375,
0.46142578125,
0.8818359375,
0.46142578125,
0.8818359375,
0.474365234375,
0.51318359375,
0.474365234375
],
[
0.1158447265625,
0.48095703125,
0.48779296875,
0.48095703125,
0.48779296875,
0.49462890625,
0.1158447265625,
0.49462890625
],
[
0.51123046875,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.51123046875,
0.4912109375
],
[
0.1175537109375,
0.496337890625,
0.48779296875,
0.496337890625,
0.48779296875,
0.51123046875,
0.1175537109375,
0.51123046875
],
[
0.51318359375,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.1158447265625,
0.51123046875,
0.490234375,
0.51318359375,
0.490234375,
0.52783203125,
0.1158447265625,
0.525390625
],
[
0.509765625,
0.5087890625,
0.70947265625,
0.50830078125,
0.70947265625,
0.5224609375,
0.509765625,
0.52392578125
],
[
0.1158447265625,
0.52978515625,
0.353515625,
0.52978515625,
0.353515625,
0.54345703125,
0.1158447265625,
0.54345703125
],
[
0.53076171875,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.53076171875,
0.5390625
],
[
0.51318359375,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.51318359375,
0.5546875
],
[
0.1175537109375,
0.552734375,
0.48779296875,
0.552734375,
0.48779296875,
0.56640625,
0.1175537109375,
0.56640625
],
[
0.51318359375,
0.55712890625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.1158447265625,
0.5693359375,
0.48779296875,
0.568359375,
0.48779296875,
0.5830078125,
0.1158447265625,
0.58447265625
],
[
0.509765625,
0.572265625,
0.8818359375,
0.57373046875,
0.8818359375,
0.58837890625,
0.509765625,
0.58740234375
],
[
0.1158447265625,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.1158447265625,
0.5986328125
],
[
0.5087890625,
0.5888671875,
0.880859375,
0.5888671875,
0.880859375,
0.6025390625,
0.5087890625,
0.6025390625
],
[
0.118408203125,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.6162109375,
0.118408203125,
0.6162109375
],
[
0.509765625,
0.60546875,
0.6650390625,
0.60546875,
0.6650390625,
0.619140625,
0.509765625,
0.619140625
],
[
0.1175537109375,
0.6181640625,
0.486083984375,
0.6181640625,
0.486083984375,
0.6318359375,
0.1175537109375,
0.6318359375
],
[
0.1175537109375,
0.63427734375,
0.48779296875,
0.63427734375,
0.48779296875,
0.64794921875,
0.1175537109375,
0.64794921875
],
[
0.51123046875,
0.63134765625,
0.67236328125,
0.6318359375,
0.67236328125,
0.646484375,
0.51123046875,
0.6455078125
],
[
0.118408203125,
0.65087890625,
0.48779296875,
0.65087890625,
0.48779296875,
0.66455078125,
0.118408203125,
0.66455078125
],
[
0.5087890625,
0.65283203125,
0.8818359375,
0.65380859375,
0.8818359375,
0.66845703125,
0.5087890625,
0.66748046875
],
[
0.1175537109375,
0.66650390625,
0.48779296875,
0.66650390625,
0.48779296875,
0.68115234375,
0.1175537109375,
0.68115234375
],
[
0.509765625,
0.67041015625,
0.88330078125,
0.67041015625,
0.88330078125,
0.68408203125,
0.509765625,
0.68408203125
],
[
0.1175537109375,
0.68310546875,
0.48779296875,
0.68310546875,
0.48779296875,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.51318359375,
0.68603515625,
0.8271484375,
0.68603515625,
0.8271484375,
0.70068359375,
0.51318359375,
0.70068359375
],
[
0.1158447265625,
0.69873046875,
0.48779296875,
0.69873046875,
0.48779296875,
0.7119140625,
0.1158447265625,
0.7119140625
],
[
0.1158447265625,
0.71435546875,
0.328369140625,
0.71435546875,
0.328369140625,
0.728515625,
0.1158447265625,
0.728515625
],
[
0.51318359375,
0.71337890625,
0.72900390625,
0.71337890625,
0.72900390625,
0.72705078125,
0.51318359375,
0.72705078125
],
[
0.509765625,
0.73291015625,
0.87890625,
0.73291015625,
0.87890625,
0.74658203125,
0.509765625,
0.74658203125
],
[
0.1158447265625,
0.7392578125,
0.42822265625,
0.74072265625,
0.42822265625,
0.7548828125,
0.1158447265625,
0.75439453125
],
[
0.509765625,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.509765625,
0.76318359375
],
[
0.1158447265625,
0.76123046875,
0.486083984375,
0.76123046875,
0.486083984375,
0.775390625,
0.1158447265625,
0.775390625
],
[
0.5087890625,
0.7666015625,
0.87890625,
0.7666015625,
0.87890625,
0.77978515625,
0.5087890625,
0.77978515625
],
[
0.1175537109375,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.1175537109375,
0.79150390625
],
[
0.509765625,
0.7822265625,
0.880859375,
0.7822265625,
0.880859375,
0.7958984375,
0.509765625,
0.7958984375
],
[
0.1158447265625,
0.7919921875,
0.486083984375,
0.79345703125,
0.486083984375,
0.80810546875,
0.1158447265625,
0.80712890625
],
[
0.51123046875,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8134765625,
0.51123046875,
0.8134765625
],
[
0.1175537109375,
0.81005859375,
0.483642578125,
0.81005859375,
0.483642578125,
0.82373046875,
0.1175537109375,
0.82373046875
],
[
0.51123046875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.1175537109375,
0.826171875,
0.486083984375,
0.826171875,
0.486083984375,
0.84033203125,
0.1175537109375,
0.84033203125
],
[
0.509765625,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.509765625,
0.84423828125
],
[
0.118408203125,
0.8427734375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85546875,
0.118408203125,
0.85546875
],
[
0.509765625,
0.845703125,
0.8818359375,
0.8466796875,
0.8818359375,
0.86083984375,
0.509765625,
0.85986328125
],
[
0.1175537109375,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87255859375,
0.1175537109375,
0.87255859375
],
[
0.51123046875,
0.8642578125,
0.8818359375,
0.8642578125,
0.8818359375,
0.87451171875,
0.51123046875,
0.87451171875
],
[
0.1376953125,
0.88134765625,
0.48779296875,
0.88134765625,
0.48779296875,
0.89453125,
0.1376953125,
0.89453125
],
[
0.5087890625,
0.88525390625,
0.88330078125,
0.8857421875,
0.88330078125,
0.90380859375,
0.5087890625,
0.90283203125
],
[
0.1158447265625,
0.8935546875,
0.48779296875,
0.8935546875,
0.48779296875,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.202392578125,
0.90625,
0.202392578125,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"an empty trajectory dataset D",
"and query the victim model for decisions. If one",
"candidatesuccessfullyflips thelabel,wetreat itas",
"Sampling Trajectories To train MAYA we",
"the final adversarial sample. Otherwise, to generate",
"need to interact with the victim model to obtain",
"adversarial samples more efficiently, we take the",
"the training data. Specifically, we train a local vic-",
"candidate whosesentenceembedding hasthelow-",
"timmodel that has the same architecture withthe",
"est cosine similarity with the sentence embedding",
"target victim model, expecting to approximate the",
"of theoriginal sentenceasthemostpotential candi-",
"decision boundary of the target victim model4",
" date. Our intuition is that the candidate that least",
"We sample a batch of original sentences. For",
"resembles the original sentences is more likely to",
"each sentence So, we generate adversarial candi-",
"be a successful adversarial sample.",
"dates Ci, ..., Ck. As in the Verify and Pick steps",
"in MAYA algorithm, one specific candidate will be",
"4",
"Experiments",
"chosen as the final successful adversarial sample",
"or the most potential candidate. We view the can-",
"We conduct comprehensive experiments to eval-",
"didatechosenbyMAYAalgorithmastheground",
"uate our attack models on the tasks of sentiment",
"truth label and add ((So, Ci, .., C), label) to our",
"analysis,natural language inference,and news clas-",
"trajectory dataset D.",
"sification.",
"To fully train an agent that can tackle different",
"4.1Datasets and Victim Models",
"situations, we need a large dataset D. So, we adapt",
"DAGGER method. Specifically, when receiving the",
"For sentiment analysis, we choose SST-2 (Socher",
"ground truth label from MAYA algorithm, MAYA",
"et al., 2013), a binary sentiment classification",
"doesn't take the golden action indicated by MAYA.",
"benchmark dataset. For natural language inference,",
"Itwill taketheaction based on its ownprediction.",
"we choose mismatched MNLI dataset (Williams",
"That is, MAYA will predict which candidate will",
"et al., 2018). For news classification, we choose",
"most confuse the victim model and follow its own",
"AG's News dataset. The models need to choose",
"procedure of launching an adversarial attack. The",
"one of the four classes including World, Sports,",
"predicted candidate will be treated as So and we",
"Business, and Sci/Tech, given an instance in AG's",
"will continue the same process of sampling trajec-",
"News (Zhang et al., 2015).",
"tories to augment the dataset D.",
"We evaluate our attack models by attacking three",
"victim models including BiLSTM (Schuster and",
"Training Then we train MAYA for only one",
"Paliwal, 1997), BERT (Devlin et al., 2019), and",
"epoch using the trajectory dataset D. We model the",
"RoBERTa (Liu et al., 2019). Details of the datasets",
"training task as a multi-class classification problem",
"and the classification accuracy of victim models",
"For each sample ((So, Ci,..., Ck), label) drawn",
"are listed in Table 1.",
"from D, we concatenate So with each Ci. Then we",
"input the k concatenated sentences to MAYA to",
"4.2Attack Models",
"get k scores. We treat these k scores as logits and",
"We implement all baseline attack models using",
"use the cross-entropy loss to train MAYAx. After",
"the NLP attack package TextAttack (Morris et al.,",
"the training process, we clear D and continue the",
"2020) and OpenAttack (Zeng et al., 2021).",
"sampling procedure. The implementation details",
"are described in Appendix I.",
"Score-based Attack Models",
"We comprehensively compare our score-based",
"3.3Adapted toDecision-based Attack",
"attack models with five representative and",
"To adapt MAYA to decision-based attack, we only",
"strong score-based attack models including",
"need to modify one step in the attack procedure",
"(1) GA+Embedding (Alzantot et al., 2018)",
"described in Section 3.2.1 while keep other steps",
"(2) PWWS+Synonym (Ren et al., 2019), (3)",
"unchanged. Specifically, when the candidate S",
"PSO+Sememe (Zang et al., 2020), (4) TextFooler",
"that gets the highest score from MAYA is from",
"(Jin et al., 2019), (5) BERT-Attack (Li et al., 2020)",
"Vs, we iteratively substitute [MASK| token with",
"Details of baseline models are listed in Appendix B",
"candidate words to generate adversarial candidates",
"and we describedetails of our attackmodels below",
"4It is unpractical to query the target victim model thou-",
"MAYAWe select 3 open-source paraphrase mod-",
"sands of times to train an agent that is employed to attack this",
"victim model.",
"els as the building blocks of MAYA. Specifically."
] |
{
"arxiv_id": "2109.04367",
"language": "en",
"page_idx": 4
}
|
||
2006.15516-en-0
|
2006.15516-en-0.jpg
|
# Graph Convolutional Network for Recommendation with Low-pass Collaborative Filters
Wenhui Yu $^{1}$ Zheng Qin$^{12}$
## $\mathbf{Abstract}$
Graph Convolutional Network (GCN) is widely used in graph data learning tasks such as recommendation. However, when facing a large graph, the graph convolution is very computationally expensive thus is simplifed in all existing GCNs, yet is seriously impaired due to the oversimplification. To address this gap, we leverage the original graph convolution in GCN and propose a $\mathbf{L}$ow-pass Collaborative Filter (LCF) to make it applicable to the large graph. LCF is designed to remove the noise caused by exposure and quantization in the observed data, and it also reduces the complexity of graph convolution in an unscathed way. Experiments show that LCF improves the effectiveness and efficiency of graph convolution and our GCN outperforms existing GCNs significantly. Codes are available on https: //github.com/wenhui-Yu/LCFN.
## $1. \textbf{Introduction}$
Graph convolution (Shuman et al., 2013) is the extension of convolution from the Euclidean domain to the graph domain and Graph Convolutional Network (GCN) is the extension of Convolutional Neural Network (CNN).GCN is widely explored in graph learning tasks such as recommendation since it can extract abundant information from the graph data. However, the eigen-decomposition of a large matrix is required in the graph convolution (Shuman et al., 2013), which is extremely space- and time-consuming in real-world applications. To avoid the eigen-decomposition, all existing GCNs constrain (even fx) the convolutional kernel and simplify graph convolution to graph propagation (Defferrard et al., 2016; Kipf & Welling, 2017). Graph Neural Network $(\mathbf{GNN})$ propagates signals through the graph, and refines
${^{1}\text{Tsinghua U}}$niversity $^2$Corresponding author. Correspondence to: Wenhui Yu<yuwh16@mails.tsinghua.edu.cn>, Zheng Qin <qingzh@ mail.tsinghua.edu.cn>.
Proceedings of the 37$^th$ International Conference on Machine $Learning$,Online, PMLR 119, 2020. Copyright 2020 by the au- $thor(s).$
the signal of each node with its neighborhoods (Zhou et al., 2018). By constraining the kernel, GCNs lose the ability of convolution and degenerate to GNNs (discussed in detail in Section 5.4). To address the oversimplification issue, we propose a new GCN that can really leverage the ability of graph convolution and validate the effectiveness in the recommendation task. In our GCN, we propose a Lowpass Collaborative Filter (LCF) both to filter the noise in observed data and to reduce the time consumption of graph convolution without hurting its ability.
LCF is devised to remove the noise in the observed data. In this paper, we adopt implicit feedback data (interactions such as“purchase”,“click”,“watch”records, denoted by binary variables) since it is generally applicable in recommendation tasks. A matrix for observed interactions is illustrated in Figure l(c), and Figure 1 shows how it is generated: There is an unknown underlying numeric rating matrix $\mathbf{R}_0$ indicating the true preference score for each user towards each item, shown in Figure 1(a). When browsing items on a website, users can only browse a small proportion of the mass items, thus many items are not exposed to the users, including many preferred items. Reflected in the matrix representation, many elements in $\mathbf{R}_0$ are set to 0 randomly, shown in Figure 1(b). To simulate this process, we add exposure noise $\mathbf{N}_1$ to the rating matrix $\mathbf{R}_0.$ For the remaining exposed items, a user will interact with them only if she like them. In the matrix representation, we discretize the matrix $R_{0}$+ $N_{1}$ with a threshold, which is set as 0. 5 in this example. The quantization noise is denoted as $\mathbf{N}_{2}$,and we have $\mathbf{R}=\mathbf{R}_0+\mathbf{N}_1+\mathbf{N}_2.$ Our task is to reconstruct $\mathbf{R}_0$ from observed $\mathbf{R}$ and rank all items for each user based on the reconstruction.
Considering $\mathbf{R}_0$ is low frequency due to the strong dependence and $\mathbf{N}_1$ and $\mathbf{N}_2$ are high frequency due to the randomness (we will introduce the reason in detail in Section 3.1), we can remove noise by low-pass filtering. However, different from the images, $\mathbf{R},\mathbf{R}_0,\mathbf{N}_1$ and $\mathbf{N}_2$ are all 2D graph signals, thus conventional filters are not effective. For
signals defined in the Euclidean domain like images and timing signals, the dependence is based on position, such as pixels near in the space position or sampling points near in the time axis. Nevertheless, for signals defined in the graph
|
[
[
0.2279052734375,
0.1142578125,
0.74658203125,
0.1142578125,
0.74658203125,
0.1287841796875,
0.2279052734375,
0.1287841796875
],
[
0.30810546875,
0.136962890625,
0.6650390625,
0.136962890625,
0.6650390625,
0.1514892578125,
0.30810546875,
0.1514892578125
],
[
0.38818359375,
0.1988525390625,
0.583984375,
0.1962890625,
0.583984375,
0.21533203125,
0.38818359375,
0.2171630859375
],
[
0.2442626953125,
0.24365234375,
0.3203125,
0.24365234375,
0.3203125,
0.25830078125,
0.2442626953125,
0.25830078125
],
[
0.5009765625,
0.2442626953125,
0.8857421875,
0.2442626953125,
0.8857421875,
0.25830078125,
0.5009765625,
0.25830078125
],
[
0.5009765625,
0.260009765625,
0.88623046875,
0.260009765625,
0.88623046875,
0.2734375,
0.5009765625,
0.2734375
],
[
0.1217041015625,
0.267578125,
0.439453125,
0.267578125,
0.439453125,
0.281005859375,
0.1217041015625,
0.281005859375
],
[
0.0236968994140625,
0.27587890625,
0.058013916015625,
0.27587890625,
0.058013916015625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.5009765625,
0.27587890625,
0.8857421875,
0.27587890625,
0.8857421875,
0.2890625,
0.5009765625,
0.2890625
],
[
0.12091064453125,
0.2841796875,
0.441162109375,
0.2841796875,
0.441162109375,
0.29736328125,
0.12091064453125,
0.29736328125
],
[
0.5009765625,
0.292236328125,
0.8857421875,
0.292236328125,
0.8857421875,
0.302978515625,
0.5009765625,
0.302978515625
],
[
0.1217041015625,
0.298583984375,
0.441162109375,
0.298583984375,
0.441162109375,
0.3125,
0.1217041015625,
0.3125
],
[
0.5,
0.30615234375,
0.8837890625,
0.30615234375,
0.8837890625,
0.3193359375,
0.5,
0.3193359375
],
[
0.12091064453125,
0.313232421875,
0.44287109375,
0.313232421875,
0.44287109375,
0.32763671875,
0.12091064453125,
0.32763671875
],
[
0.5,
0.3212890625,
0.8857421875,
0.3212890625,
0.8857421875,
0.335205078125,
0.5,
0.335205078125
],
[
0.1217041015625,
0.328857421875,
0.441162109375,
0.328857421875,
0.441162109375,
0.34228515625,
0.1217041015625,
0.34228515625
],
[
0.5009765625,
0.3359375,
0.88623046875,
0.3359375,
0.88623046875,
0.349853515625,
0.5009765625,
0.349853515625
],
[
0.1217041015625,
0.34521484375,
0.439453125,
0.34521484375,
0.439453125,
0.35595703125,
0.1217041015625,
0.35595703125
],
[
0.5009765625,
0.3515625,
0.88623046875,
0.3515625,
0.88623046875,
0.365478515625,
0.5009765625,
0.365478515625
],
[
0.0269622802734375,
0.3623046875,
0.0555419921875,
0.3623046875,
0.0555419921875,
0.391357421875,
0.0269622802734375,
0.391357421875
],
[
0.1217041015625,
0.359130859375,
0.441162109375,
0.359130859375,
0.441162109375,
0.373779296875,
0.1217041015625,
0.373779296875
],
[
0.5,
0.365478515625,
0.8857421875,
0.3662109375,
0.8857421875,
0.380615234375,
0.5,
0.380126953125
],
[
0.1192626953125,
0.37255859375,
0.439453125,
0.37255859375,
0.439453125,
0.386962890625,
0.1192626953125,
0.386962890625
],
[
0.5,
0.3818359375,
0.75244140625,
0.3818359375,
0.75244140625,
0.395751953125,
0.5,
0.395751953125
],
[
0.1192626953125,
0.38818359375,
0.44287109375,
0.386962890625,
0.44287109375,
0.401611328125,
0.1192626953125,
0.40283203125
],
[
0.1192626953125,
0.404052734375,
0.441162109375,
0.40283203125,
0.441162109375,
0.417236328125,
0.1192626953125,
0.41845703125
],
[
0.5009765625,
0.40283203125,
0.88623046875,
0.40283203125,
0.88623046875,
0.416748046875,
0.5009765625,
0.416748046875
],
[
0.1217041015625,
0.42041015625,
0.439453125,
0.42041015625,
0.439453125,
0.43115234375,
0.1217041015625,
0.43115234375
],
[
0.02532958984375,
0.406005859375,
0.058013916015625,
0.406005859375,
0.058013916015625,
0.448974609375,
0.02532958984375,
0.448974609375
],
[
0.50244140625,
0.42041015625,
0.8837890625,
0.42041015625,
0.8837890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.1217041015625,
0.434326171875,
0.441162109375,
0.434326171875,
0.441162109375,
0.447509765625,
0.1217041015625,
0.447509765625
],
[
0.5009765625,
0.434326171875,
0.8837890625,
0.434326171875,
0.8837890625,
0.447509765625,
0.5009765625,
0.447509765625
],
[
0.02777099609375,
0.44140625,
0.05963134765625,
0.44140625,
0.05963134765625,
0.484130859375,
0.02777099609375,
0.484130859375
],
[
0.12091064453125,
0.449462890625,
0.441162109375,
0.449462890625,
0.441162109375,
0.46337890625,
0.12091064453125,
0.46337890625
],
[
0.5009765625,
0.449462890625,
0.8857421875,
0.449462890625,
0.8857421875,
0.46337890625,
0.5009765625,
0.46337890625
],
[
0.1192626953125,
0.464111328125,
0.439453125,
0.464111328125,
0.439453125,
0.477783203125,
0.1192626953125,
0.477783203125
],
[
0.5009765625,
0.466552734375,
0.8857421875,
0.466552734375,
0.8857421875,
0.4765625,
0.5009765625,
0.4765625
],
[
0.1217041015625,
0.479736328125,
0.44287109375,
0.479736328125,
0.44287109375,
0.49365234375,
0.1217041015625,
0.49365234375
],
[
0.5,
0.479736328125,
0.88623046875,
0.479736328125,
0.88623046875,
0.49365234375,
0.5,
0.49365234375
],
[
0.021240234375,
0.4892578125,
0.060455322265625,
0.490478515625,
0.058013916015625,
0.5888671875,
0.0187835693359375,
0.58837890625
],
[
0.1217041015625,
0.494384765625,
0.441162109375,
0.494384765625,
0.441162109375,
0.50830078125,
0.1217041015625,
0.50830078125
],
[
0.50244140625,
0.496826171875,
0.8837890625,
0.496826171875,
0.8837890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.1217041015625,
0.51025390625,
0.443603515625,
0.51025390625,
0.443603515625,
0.5234375,
0.1217041015625,
0.5234375
],
[
0.5,
0.51025390625,
0.8857421875,
0.51025390625,
0.8857421875,
0.5234375,
0.5,
0.5234375
],
[
0.1217041015625,
0.52587890625,
0.390625,
0.52587890625,
0.390625,
0.5390625,
0.1217041015625,
0.5390625
],
[
0.5009765625,
0.5244140625,
0.8857421875,
0.5244140625,
0.8857421875,
0.53857421875,
0.5009765625,
0.53857421875
],
[
0.5009765625,
0.54052734375,
0.8857421875,
0.54052734375,
0.8857421875,
0.5537109375,
0.5009765625,
0.5537109375
],
[
0.5,
0.55615234375,
0.88623046875,
0.55615234375,
0.88623046875,
0.5693359375,
0.5,
0.5693359375
],
[
0.08905029296875,
0.568359375,
0.2156982421875,
0.57080078125,
0.2156982421875,
0.58544921875,
0.08905029296875,
0.58349609375
],
[
0.5009765625,
0.57080078125,
0.88623046875,
0.57080078125,
0.88623046875,
0.583984375,
0.5009765625,
0.583984375
],
[
0.5009765625,
0.58544921875,
0.88623046875,
0.58544921875,
0.88623046875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.0196075439453125,
0.57373046875,
0.058013916015625,
0.57373046875,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.08905029296875,
0.59521484375,
0.4755859375,
0.59521484375,
0.4755859375,
0.609375,
0.08905029296875,
0.609375
],
[
0.5009765625,
0.6015625,
0.8857421875,
0.6015625,
0.8857421875,
0.6123046875,
0.5009765625,
0.6123046875
],
[
0.0882568359375,
0.611328125,
0.473876953125,
0.611328125,
0.473876953125,
0.625,
0.0882568359375,
0.625
],
[
0.5,
0.61572265625,
0.88623046875,
0.61572265625,
0.88623046875,
0.62890625,
0.5,
0.62890625
],
[
0.08905029296875,
0.62548828125,
0.47216796875,
0.62548828125,
0.47216796875,
0.6396484375,
0.08905029296875,
0.6396484375
],
[
0.5,
0.630859375,
0.8857421875,
0.630859375,
0.8857421875,
0.64453125,
0.5,
0.64453125
],
[
0.08905029296875,
0.642578125,
0.47216796875,
0.642578125,
0.47216796875,
0.65283203125,
0.08905029296875,
0.65283203125
],
[
0.5,
0.64599609375,
0.8857421875,
0.64599609375,
0.8857421875,
0.6591796875,
0.5,
0.6591796875
],
[
0.0882568359375,
0.6572265625,
0.473876953125,
0.6572265625,
0.473876953125,
0.67041015625,
0.0882568359375,
0.67041015625
],
[
0.5009765625,
0.662109375,
0.8857421875,
0.662109375,
0.8857421875,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.0882568359375,
0.671875,
0.473876953125,
0.671875,
0.473876953125,
0.68505859375,
0.0882568359375,
0.68505859375
],
[
0.5,
0.67529296875,
0.88623046875,
0.67529296875,
0.88623046875,
0.689453125,
0.5,
0.689453125
],
[
0.0882568359375,
0.68505859375,
0.473876953125,
0.68603515625,
0.473876953125,
0.70068359375,
0.0882568359375,
0.69970703125
],
[
0.5,
0.68994140625,
0.8857421875,
0.68994140625,
0.8857421875,
0.705078125,
0.5,
0.705078125
],
[
0.08905029296875,
0.7021484375,
0.473876953125,
0.7021484375,
0.473876953125,
0.71533203125,
0.08905029296875,
0.71533203125
],
[
0.5,
0.70556640625,
0.88623046875,
0.70556640625,
0.88623046875,
0.7197265625,
0.5,
0.7197265625
],
[
0.0882568359375,
0.71728515625,
0.473876953125,
0.71728515625,
0.473876953125,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.498291015625,
0.7216796875,
0.6259765625,
0.7216796875,
0.6259765625,
0.73486328125,
0.498291015625,
0.73486328125
],
[
0.0882568359375,
0.73193359375,
0.473876953125,
0.73095703125,
0.473876953125,
0.74560546875,
0.0882568359375,
0.7470703125
],
[
0.5009765625,
0.7421875,
0.88623046875,
0.7421875,
0.88623046875,
0.7568359375,
0.5009765625,
0.7568359375
],
[
0.0906982421875,
0.74853515625,
0.473876953125,
0.74853515625,
0.473876953125,
0.7587890625,
0.0906982421875,
0.7587890625
],
[
0.08660888671875,
0.76220703125,
0.473876953125,
0.76123046875,
0.473876953125,
0.77587890625,
0.08660888671875,
0.7763671875
],
[
0.5,
0.7587890625,
0.88623046875,
0.7587890625,
0.88623046875,
0.77294921875,
0.5,
0.77294921875
],
[
0.08905029296875,
0.77880859375,
0.47216796875,
0.77880859375,
0.47216796875,
0.7890625,
0.08905029296875,
0.7890625
],
[
0.5,
0.7734375,
0.8857421875,
0.7734375,
0.8857421875,
0.787109375,
0.5,
0.787109375
],
[
0.08660888671875,
0.79150390625,
0.473876953125,
0.79248046875,
0.473876953125,
0.806640625,
0.08660888671875,
0.80615234375
],
[
0.5009765625,
0.7890625,
0.88623046875,
0.7890625,
0.88623046875,
0.80322265625,
0.5009765625,
0.80322265625
],
[
0.5009765625,
0.8037109375,
0.8857421875,
0.8037109375,
0.8857421875,
0.818359375,
0.5009765625,
0.818359375
],
[
0.1070556640625,
0.814453125,
0.473876953125,
0.81640625,
0.473876953125,
0.83349609375,
0.1070556640625,
0.83056640625
],
[
0.5,
0.8193359375,
0.8857421875,
0.818359375,
0.8857421875,
0.83203125,
0.5,
0.83349609375
],
[
0.08660888671875,
0.82763671875,
0.47216796875,
0.8291015625,
0.47216796875,
0.84326171875,
0.08660888671875,
0.84228515625
],
[
0.50244140625,
0.8359375,
0.8837890625,
0.8359375,
0.8837890625,
0.8466796875,
0.50244140625,
0.8466796875
],
[
0.08905029296875,
0.84228515625,
0.294189453125,
0.84228515625,
0.294189453125,
0.85595703125,
0.08905029296875,
0.85595703125
],
[
0.5,
0.849609375,
0.8857421875,
0.849609375,
0.8857421875,
0.86279296875,
0.5,
0.86279296875
],
[
0.0849609375,
0.86279296875,
0.476318359375,
0.8642578125,
0.476318359375,
0.8818359375,
0.0849609375,
0.880859375
],
[
0.5,
0.86572265625,
0.8837890625,
0.86572265625,
0.8837890625,
0.87890625,
0.5,
0.87890625
],
[
0.0882568359375,
0.8798828125,
0.47802734375,
0.8798828125,
0.47802734375,
0.89306640625,
0.0882568359375,
0.89306640625
],
[
0.5,
0.8798828125,
0.88623046875,
0.8798828125,
0.88623046875,
0.89306640625,
0.5,
0.89306640625
],
[
0.08660888671875,
0.89013671875,
0.1356201171875,
0.892578125,
0.134033203125,
0.9072265625,
0.0849609375,
0.90478515625
]
] |
[
"Graph Convolutional Network for Recommendation",
"with Low-pass Collaborative Filters",
"Wenhui Yu I Zheng Qin 12",
"Abstract",
"the signal of each node with its neighborhoods (Zhou et al.",
"2018). By constraining the kernel, GCNs lose the ability of",
"Graph Convolutional Network (GCN) is widely",
"",
"convolution and degenerate to GNNs (discussed in detail",
"used in graph data learning tasks such as recom-",
"in Section 5.4).To address the oversimplification issue.",
"mendation. However, when facing a large graph.",
"we propose a new GCN that can really leverage the ability",
"the graph convolution is very computationally ex-",
"of graph convolution and validate the effectiveness in the",
"pensive thus is simplified in all existing GCNs,",
"recommendation task. In our GCN, we propose a Low-",
"yetis seriouslyimpaired duetotheoversimpli",
"pass Collaborative Filter (LCF) both to filter the noise in",
"00",
"fication. To address this gap, we leverage the",
"observed data and to reduce the time consumption of graph",
"original graph convolution in GCN and propose a",
"convolution without hurting its ability.",
"Low-pass Collaborative Filter (LCF) to make it",
"applicable to the large graph. LCF is designed to",
"LCF is devised to remove the noise in the observed data.",
"removethenoisecausedbyexposureandquanti",
"[OT",
"Inthispaper,weadoptimplicitfeedbackdata(interactions",
"zation in the observed data, and it also reduces the",
"such as\"purchase\",\"click\",\"watch\"records, denoted by",
"I'so",
"complexity of graph convolution in an unscathed",
"binary variables) since it is generally applicable in recom-",
"way. Experiments show that LCF improves the",
"mendationtasks.A matrixfor observed interactions is illus-",
"effectiveness and efficiency of graph convolu-",
"trated in Figure 1(c), and Figure 1 shows how it is generated:",
"CA91S",
"tion and our GCN outperforms existing GCNs",
"ThereisanunknownunderlyingnumericratingmatrixRo",
"significantly. Codes are available on https:",
"indicating the true preference score for each user towards",
"//github.com/Wenhui-Yu/LCFN.",
"each item, shown in Figure 1(a). When browsing items on",
"a website, users can only browse a small proportion of the",
"mass items, thus many items are not exposed to the users,",
"1. Introduction",
"including many preferred items. Reflected in the matrix",
"representation, many elements in Ro are set to O randomly,",
"19002:A1X",
"Graph convolution (Shuman et al., 2013) is the extension of",
"shown in Figure l(b).To simulate this process, we add ex-",
"convolution from the Euclidean domain to the graph domain",
"posure noise N to the rating matrix Ro. For the remaining",
"and Graph Convolutional Network (GCN) is the extension",
"exposed items, a user will interact with them only if she",
"ofConvolutionalNeuralNetwork(CNN).GCN iswidely",
"like them. In the matrix representation, we discretize the",
"explored in graph learning tasks such as recommendation",
"matrix Ro +N, withathreshold.which is set as 0.5 in this",
"since it can extract abundant information from the graph",
"example. The quantization noise is denoted as N2, and we",
"data. However, the eigen-decomposition of a large matrix",
"have R = Ro + Ni + N2. Our task is to reconstruct Ro",
"is required in the graph convolution (Shuman et al., 2013)",
"from observed R and rank all items for each user based on",
"which is extremely space- and time-consuming in real-world",
"the reconstruction.",
"applications. To avoid the eigen-decomposition, all existing",
"Considering Ro is low frequency due to the strong depen-",
"GCNs constrain (evenfix)the convolutionalkerneland",
"simplify graph convolution to graph propagation (Defferrard",
"dence and Ni and N2 are high frequency due to the ran-",
"etal..2016:Kipf&Welling.2017).GraphNeuralNetwork",
"domness (we will introduce the reason in detail in Section",
"(GNN) propagates signals through the graph, and refines",
"3.1), we can remove noise by low-pass filtering.However,",
"different from the images, R, Ro, Ni and N2 are all 2D",
"'Tsinghua University ?Corresponding author. Correspondence",
"graph signals, thus conventional filters are not effective. For",
"to: Wenhui Yu <yuwh16@mails.tsinghua.edu.cn>, Zheng Qin",
"signalsdefined inthe Euclidean domainlikeimages and",
"<qingzh@mail.tsinghua.edu.cn>.",
"timing signals, the dependence is based on position, such as",
"Proceedings of the 37th International Conference on Machine",
"pixels near in the space position or sampling points near in",
"Learning, Online, PMLR 119, 2020. Copyright 2020 by the au-",
"the time axis. Nevertheless, for signals defined in the graph",
"thor(s)."
] |
{
"arxiv_id": "2006.15516",
"language": "en",
"page_idx": 0
}
|
||
2006.15516-en-1
|
2006.15516-en-1.jpg
|
Submission and Formatting Instructions for ICML 2020
<FigureHere>
Figure 1. The interaction matrix and different kinds of noise. Curves in Figure l(a) indicate a user graph and an item graph. Two users are
connected if they have interacted with the same item, and two items are connected if they have been interacted by the same user
Neural Network-based Recommendation. These years, Neural Network achieves significant success and is widely used in various machine learning and information retrieval domains, including recommendation. Cheng et al. (2016); He et al. (2017) utilized wide and deep networks to learn the nonlinear combination of embeddings. Chen et al. (2017) explored the attention mechanism to highlight important elements. He& McAuley (2016b); Yu et al. (2018) explored side information to enhance the performance of recommendation. Li et al. (2017); Tan et al. (2016) regarded recommendation as a sequence prediction task and utilized RNN or CNN to deal with this issue. Though extensively explored, these models are based on conventional networks and omit the graph structure.
domain, the dependence is based on graph connections, for example, we consider the dependence of two rows/columns of $\mathbf{R}_0$ based on the user/item graph shown in Figure 1(a). To remove the graph noise, we devise a 2D graph filter called LCF and reconstruct $\mathbf{R}_0$ by it: $\hat{\mathbf{R}}_0=LCF(\hat{\mathbf{R}})$, where $\hat{\mathbf{R}}_0$ is the estimation. Similar to the filter, we also extend 2D convolution from the Euclidean domain to the graph domain to propose GCN. We then inject LCF into GCN to propose our LCF Network (LCFN). LCF not only improves the accuracy by removing the noise, but also reduces the computation of graph convolution since only a small proportion of eigen-vectors are required. To the best of our knowledge, LCFN is the first GCN to use the original graph convolution. To summarize, this work makes the following contributions:
· We introduce the 2D graph Fourier transform, and
then design the graph filter LCF and graph convolution
based on it.
· We design a novel GCN named LCFN for recommendation with LCF to improve the performance by removing noise and to improve the efficiency of graph convolution.
· We devise comprehensive experiments on two real-
world datasets to demonstrate the effectiveness and
efficiency of LCFN.
Graph-based Recommendation. As a link prediction task of a bipartite graph, graph tools are widely used in the recommendation domain. Rao et al.(2015) used graph regularization to smooth the embeddings of connected nodes (users and items in recommendation context). Yu&Qin (2019) extracted spectral features for nodes from the hypergraph Laplacian matrices to optimize model learning. However, these shallow models only get very limited information. Wang et al.(2019); Ying et al.(2018); Berg et al. (2018) utilized deep GNNs for recommendation tasks. GNNs propagate embeddings through the graph to ensure that connected nodes have similar latent representations. Though GNNs perform well, there is still a large room for improvement - GNNs only smooth the embeddings while cannot extract high-level patterns for the graph signal. To address this gap, GCNs are proposed.
## $2. \textbf{Related Work}$
After the collaborative filtering (CF) model was proposed (Sarwar et al., 2001; Goldberg, 1992), recommender systems have boomed in decades. Among various CF methods, matrix factorization (MF) (Koren, 2009; Rendle et al., 2009), which encodes the preference of users by underlying embeddings (i.e., latent factors), is a basic yet the most effective recommender model. To improve the presentation capability, many variants have been proposed (He et al., 2017; Wang et al., 2019). Among these models, neural networkbased models and graph-based models are the most relevant to our work.
Shuman et al. (2013) proposed graph convolution to extract the dependence pattern from the graph data. Since graph convolution on large graphs is extremely time-consuming, Defferrard et al. (2016); Kipf & Welling (2017) first simplified it by various approximations, unfortunately, degenerate GCNs to GNNs. By following them, existing GCNs are indeed GNNs, or variants of GNNs (Yao &Li, 2018; Monti et al., 2017; Zheng et al., 2018). Also, as a result, many
|
[
[
0.3095703125,
0.05682373046875,
0.6650390625,
0.05682373046875,
0.6650390625,
0.0706787109375,
0.3095703125,
0.0706787109375
],
[
0.45751953125,
0.09344482421875,
0.5390625,
0.09344482421875,
0.5390625,
0.111083984375,
0.45751953125,
0.111083984375
],
[
0.67236328125,
0.0921630859375,
0.7509765625,
0.0947265625,
0.7509765625,
0.10986328125,
0.67236328125,
0.10797119140625
],
[
0.2646484375,
0.10101318359375,
0.28515625,
0.10101318359375,
0.28515625,
0.10479736328125,
0.2646484375,
0.10479736328125
],
[
0.28515625,
0.0985107421875,
0.30810546875,
0.0985107421875,
0.30810546875,
0.1060791015625,
0.28515625,
0.1060791015625
],
[
0.233642578125,
0.11297607421875,
0.326904296875,
0.11297607421875,
0.326904296875,
0.1275634765625,
0.233642578125,
0.1275634765625
],
[
0.451904296875,
0.11297607421875,
0.5400390625,
0.11297607421875,
0.5400390625,
0.1275634765625,
0.451904296875,
0.1275634765625
],
[
0.69287109375,
0.1142578125,
0.75244140625,
0.1142578125,
0.75244140625,
0.1287841796875,
0.69287109375,
0.1287841796875
],
[
0.64892578125,
0.11614990234375,
0.6640625,
0.11614990234375,
0.6640625,
0.126953125,
0.64892578125,
0.126953125
],
[
0.43310546875,
0.11553955078125,
0.459228515625,
0.11553955078125,
0.459228515625,
0.126953125,
0.43310546875,
0.126953125
],
[
0.67724609375,
0.11871337890625,
0.6845703125,
0.11871337890625,
0.6845703125,
0.1256103515625,
0.67724609375,
0.1256103515625
],
[
0.343017578125,
0.1256103515625,
0.410888671875,
0.1256103515625,
0.410888671875,
0.1363525390625,
0.343017578125,
0.1363525390625
],
[
0.22216796875,
0.11932373046875,
0.233642578125,
0.11932373046875,
0.233642578125,
0.12371826171875,
0.22216796875,
0.12371826171875
],
[
0.453369140625,
0.1307373046875,
0.5400390625,
0.1307373046875,
0.5400390625,
0.1458740234375,
0.453369140625,
0.1458740234375
],
[
0.5498046875,
0.1256103515625,
0.63623046875,
0.1256103515625,
0.63623046875,
0.136962890625,
0.5498046875,
0.136962890625
],
[
0.24267578125,
0.1331787109375,
0.325927734375,
0.1331787109375,
0.325927734375,
0.14453125,
0.24267578125,
0.14453125
],
[
0.431396484375,
0.1331787109375,
0.459228515625,
0.1331787109375,
0.459228515625,
0.14453125,
0.431396484375,
0.14453125
],
[
0.64892578125,
0.1331787109375,
0.6640625,
0.1331787109375,
0.6640625,
0.143310546875,
0.64892578125,
0.143310546875
],
[
0.67724609375,
0.1351318359375,
0.6845703125,
0.1351318359375,
0.6845703125,
0.1427001953125,
0.67724609375,
0.1427001953125
],
[
0.72119140625,
0.1351318359375,
0.72802734375,
0.1351318359375,
0.72802734375,
0.14013671875,
0.72119140625,
0.14013671875
],
[
0.73876953125,
0.1331787109375,
0.74853515625,
0.1331787109375,
0.74853515625,
0.14453125,
0.73876953125,
0.14453125
],
[
0.3505859375,
0.1395263671875,
0.404296875,
0.1395263671875,
0.404296875,
0.15087890625,
0.3505859375,
0.15087890625
],
[
0.560546875,
0.1383056640625,
0.619140625,
0.1383056640625,
0.619140625,
0.15283203125,
0.560546875,
0.15283203125
],
[
0.240234375,
0.1483154296875,
0.32275390625,
0.1483154296875,
0.32275390625,
0.160400390625,
0.240234375,
0.160400390625
],
[
0.4345703125,
0.1514892578125,
0.45751953125,
0.1514892578125,
0.45751953125,
0.160400390625,
0.4345703125,
0.160400390625
],
[
0.450927734375,
0.146484375,
0.5390625,
0.1483154296875,
0.5390625,
0.1634521484375,
0.450927734375,
0.1610107421875
],
[
0.64892578125,
0.1514892578125,
0.6640625,
0.1514892578125,
0.6640625,
0.160400390625,
0.64892578125,
0.160400390625
],
[
0.740234375,
0.15283203125,
0.74658203125,
0.15283203125,
0.74658203125,
0.160400390625,
0.740234375,
0.160400390625
],
[
0.236083984375,
0.1673583984375,
0.325927734375,
0.1673583984375,
0.325927734375,
0.1787109375,
0.236083984375,
0.1787109375
],
[
0.431396484375,
0.1666259765625,
0.5400390625,
0.1666259765625,
0.5400390625,
0.18115234375,
0.431396484375,
0.18115234375
],
[
0.64892578125,
0.17041015625,
0.6640625,
0.17041015625,
0.6640625,
0.179931640625,
0.64892578125,
0.179931640625
],
[
0.67626953125,
0.17041015625,
0.685546875,
0.17041015625,
0.685546875,
0.177978515625,
0.67626953125,
0.177978515625
],
[
0.369384765625,
0.1787109375,
0.38818359375,
0.1787109375,
0.38818359375,
0.1875,
0.369384765625,
0.1875
],
[
0.58251953125,
0.173583984375,
0.611328125,
0.173583984375,
0.611328125,
0.1900634765625,
0.58251953125,
0.1900634765625
],
[
0.2418212890625,
0.1849365234375,
0.325927734375,
0.1849365234375,
0.325927734375,
0.1962890625,
0.2418212890625,
0.1962890625
],
[
0.4306640625,
0.184326171875,
0.5400390625,
0.18310546875,
0.5400390625,
0.1976318359375,
0.4306640625,
0.1988525390625
],
[
0.650390625,
0.1881103515625,
0.6640625,
0.1881103515625,
0.6640625,
0.1962890625,
0.650390625,
0.1962890625
],
[
0.67724609375,
0.189453125,
0.6845703125,
0.189453125,
0.6845703125,
0.1944580078125,
0.67724609375,
0.1944580078125
],
[
0.2156982421875,
0.20703125,
0.339111328125,
0.20703125,
0.339111328125,
0.2203369140625,
0.2156982421875,
0.2203369140625
],
[
0.423095703125,
0.20703125,
0.560546875,
0.20703125,
0.560546875,
0.2203369140625,
0.423095703125,
0.2203369140625
],
[
0.63623046875,
0.205810546875,
0.76708984375,
0.205810546875,
0.76708984375,
0.2203369140625,
0.63623046875,
0.2203369140625
],
[
0.4306640625,
0.222900390625,
0.55224609375,
0.222900390625,
0.55224609375,
0.23291015625,
0.4306640625,
0.23291015625
],
[
0.64794921875,
0.2203369140625,
0.75732421875,
0.2203369140625,
0.75732421875,
0.2342529296875,
0.64794921875,
0.2342529296875
],
[
0.0882568359375,
0.250732421875,
0.88623046875,
0.250732421875,
0.88623046875,
0.264404296875,
0.0882568359375,
0.264404296875
],
[
0.0882568359375,
0.266357421875,
0.84228515625,
0.266357421875,
0.84228515625,
0.276611328125,
0.0882568359375,
0.276611328125
],
[
0.0882568359375,
0.3037109375,
0.4755859375,
0.3037109375,
0.4755859375,
0.318115234375,
0.0882568359375,
0.318115234375
],
[
0.498291015625,
0.302978515625,
0.88623046875,
0.3037109375,
0.88623046875,
0.318115234375,
0.498291015625,
0.317626953125
],
[
0.0882568359375,
0.3193359375,
0.4755859375,
0.3193359375,
0.4755859375,
0.333251953125,
0.0882568359375,
0.333251953125
],
[
0.5,
0.3193359375,
0.8857421875,
0.3193359375,
0.8857421875,
0.333251953125,
0.5,
0.333251953125
],
[
0.0882568359375,
0.335205078125,
0.473876953125,
0.335205078125,
0.473876953125,
0.348388671875,
0.0882568359375,
0.348388671875
],
[
0.5,
0.335205078125,
0.8857421875,
0.335205078125,
0.8857421875,
0.348388671875,
0.5,
0.348388671875
],
[
0.0882568359375,
0.35107421875,
0.473876953125,
0.35107421875,
0.473876953125,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.5,
0.3515625,
0.8857421875,
0.3515625,
0.8857421875,
0.3623046875,
0.5,
0.3623046875
],
[
0.08905029296875,
0.365478515625,
0.473876953125,
0.365478515625,
0.473876953125,
0.37890625,
0.08905029296875,
0.37890625
],
[
0.498291015625,
0.365478515625,
0.8857421875,
0.365478515625,
0.8857421875,
0.37890625,
0.498291015625,
0.37890625
],
[
0.0882568359375,
0.380126953125,
0.473876953125,
0.380126953125,
0.473876953125,
0.393310546875,
0.0882568359375,
0.393310546875
],
[
0.5,
0.3818359375,
0.8857421875,
0.3818359375,
0.8857421875,
0.392578125,
0.5,
0.392578125
],
[
0.0882568359375,
0.395751953125,
0.473876953125,
0.395751953125,
0.473876953125,
0.4091796875,
0.0882568359375,
0.4091796875
],
[
0.5,
0.395751953125,
0.88623046875,
0.395751953125,
0.88623046875,
0.4091796875,
0.5,
0.4091796875
],
[
0.0882568359375,
0.410400390625,
0.473876953125,
0.410400390625,
0.473876953125,
0.423583984375,
0.0882568359375,
0.423583984375
],
[
0.5,
0.410400390625,
0.88623046875,
0.410400390625,
0.88623046875,
0.423583984375,
0.5,
0.423583984375
],
[
0.0882568359375,
0.4248046875,
0.473876953125,
0.4248046875,
0.473876953125,
0.438232421875,
0.0882568359375,
0.438232421875
],
[
0.5,
0.425537109375,
0.88623046875,
0.425537109375,
0.88623046875,
0.439453125,
0.5,
0.439453125
],
[
0.0882568359375,
0.43994140625,
0.4755859375,
0.43994140625,
0.4755859375,
0.453857421875,
0.0882568359375,
0.453857421875
],
[
0.498291015625,
0.43994140625,
0.88623046875,
0.43994140625,
0.88623046875,
0.453857421875,
0.498291015625,
0.453857421875
],
[
0.0882568359375,
0.455810546875,
0.473876953125,
0.455810546875,
0.473876953125,
0.4697265625,
0.0882568359375,
0.4697265625
],
[
0.5,
0.455810546875,
0.88623046875,
0.455810546875,
0.88623046875,
0.4697265625,
0.5,
0.4697265625
],
[
0.0882568359375,
0.4697265625,
0.4755859375,
0.47021484375,
0.4755859375,
0.48486328125,
0.0882568359375,
0.484130859375
],
[
0.5,
0.47021484375,
0.8857421875,
0.47021484375,
0.8857421875,
0.484130859375,
0.5,
0.484130859375
],
[
0.08660888671875,
0.484130859375,
0.4755859375,
0.48486328125,
0.4755859375,
0.5,
0.08660888671875,
0.498779296875
],
[
0.50244140625,
0.4873046875,
0.8857421875,
0.4873046875,
0.8857421875,
0.49755859375,
0.50244140625,
0.49755859375
],
[
0.5,
0.5,
0.69140625,
0.50048828125,
0.69140625,
0.51513671875,
0.5,
0.5146484375
],
[
0.0882568359375,
0.50830078125,
0.4755859375,
0.50830078125,
0.4755859375,
0.521484375,
0.0882568359375,
0.521484375
],
[
0.5009765625,
0.5234375,
0.88623046875,
0.5234375,
0.88623046875,
0.537109375,
0.5009765625,
0.537109375
],
[
0.1070556640625,
0.5322265625,
0.473876953125,
0.5322265625,
0.473876953125,
0.546875,
0.1070556640625,
0.546875
],
[
0.5009765625,
0.5390625,
0.88623046875,
0.5390625,
0.88623046875,
0.55322265625,
0.5009765625,
0.55322265625
],
[
0.12091064453125,
0.54736328125,
0.473876953125,
0.54736328125,
0.473876953125,
0.5625,
0.12091064453125,
0.5625
],
[
0.498291015625,
0.55322265625,
0.88623046875,
0.5537109375,
0.88623046875,
0.568359375,
0.498291015625,
0.5673828125
],
[
0.1217041015625,
0.564453125,
0.19775390625,
0.564453125,
0.19775390625,
0.57568359375,
0.1217041015625,
0.57568359375
],
[
0.50244140625,
0.57080078125,
0.8857421875,
0.57080078125,
0.8857421875,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.1070556640625,
0.58349609375,
0.473876953125,
0.58349609375,
0.473876953125,
0.5966796875,
0.1070556640625,
0.5966796875
],
[
0.5,
0.583984375,
0.8857421875,
0.583984375,
0.8857421875,
0.59765625,
0.5,
0.59765625
],
[
0.12091064453125,
0.5986328125,
0.473876953125,
0.5986328125,
0.473876953125,
0.6123046875,
0.12091064453125,
0.6123046875
],
[
0.5,
0.5986328125,
0.88623046875,
0.5986328125,
0.88623046875,
0.6123046875,
0.5,
0.6123046875
],
[
0.12091064453125,
0.6142578125,
0.473876953125,
0.6142578125,
0.473876953125,
0.62744140625,
0.12091064453125,
0.62744140625
],
[
0.5009765625,
0.6142578125,
0.88623046875,
0.6142578125,
0.88623046875,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.12091064453125,
0.6298828125,
0.205078125,
0.6298828125,
0.205078125,
0.6416015625,
0.12091064453125,
0.6416015625
],
[
0.5,
0.62744140625,
0.88623046875,
0.62890625,
0.88623046875,
0.64306640625,
0.5,
0.642578125
],
[
0.10540771484375,
0.64794921875,
0.473876953125,
0.64794921875,
0.473876953125,
0.6611328125,
0.10540771484375,
0.6611328125
],
[
0.5,
0.64453125,
0.8857421875,
0.64453125,
0.8857421875,
0.65771484375,
0.5,
0.65771484375
],
[
0.1192626953125,
0.6640625,
0.4755859375,
0.6640625,
0.4755859375,
0.67822265625,
0.1192626953125,
0.67822265625
],
[
0.5009765625,
0.6591796875,
0.88623046875,
0.6591796875,
0.88623046875,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.1192626953125,
0.6787109375,
0.2548828125,
0.67822265625,
0.2548828125,
0.6923828125,
0.1192626953125,
0.693359375
],
[
0.5009765625,
0.6748046875,
0.8857421875,
0.6748046875,
0.8857421875,
0.68798828125,
0.5009765625,
0.68798828125
],
[
0.5,
0.68994140625,
0.88623046875,
0.68994140625,
0.88623046875,
0.7041015625,
0.5,
0.7041015625
],
[
0.5,
0.705078125,
0.88623046875,
0.705078125,
0.88623046875,
0.71826171875,
0.5,
0.71826171875
],
[
0.0882568359375,
0.7138671875,
0.226318359375,
0.7138671875,
0.226318359375,
0.72900390625,
0.0882568359375,
0.72900390625
],
[
0.5009765625,
0.72021484375,
0.61279296875,
0.72021484375,
0.61279296875,
0.734375,
0.5009765625,
0.734375
],
[
0.60546875,
0.7197265625,
0.8857421875,
0.7197265625,
0.8857421875,
0.73291015625,
0.60546875,
0.73291015625
],
[
0.08905029296875,
0.74072265625,
0.473876953125,
0.74072265625,
0.473876953125,
0.75390625,
0.08905029296875,
0.75390625
],
[
0.5009765625,
0.73486328125,
0.88623046875,
0.73486328125,
0.88623046875,
0.75,
0.5009765625,
0.75
],
[
0.0882568359375,
0.755859375,
0.4755859375,
0.755859375,
0.4755859375,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.498291015625,
0.75,
0.74658203125,
0.74853515625,
0.74658203125,
0.76318359375,
0.498291015625,
0.7646484375
],
[
0.08905029296875,
0.77294921875,
0.473876953125,
0.77294921875,
0.473876953125,
0.78271484375,
0.08905029296875,
0.78271484375
],
[
0.5,
0.77099609375,
0.88623046875,
0.771484375,
0.88623046875,
0.7861328125,
0.5,
0.78515625
],
[
0.0882568359375,
0.7861328125,
0.473876953125,
0.7861328125,
0.473876953125,
0.7998046875,
0.0882568359375,
0.7998046875
],
[
0.498291015625,
0.78857421875,
0.88623046875,
0.78857421875,
0.88623046875,
0.8017578125,
0.498291015625,
0.8017578125
],
[
0.0882568359375,
0.8017578125,
0.473876953125,
0.8017578125,
0.473876953125,
0.81494140625,
0.0882568359375,
0.81494140625
],
[
0.498291015625,
0.8017578125,
0.88623046875,
0.80322265625,
0.88623046875,
0.8173828125,
0.498291015625,
0.81640625
],
[
0.08905029296875,
0.818359375,
0.47216796875,
0.818359375,
0.47216796875,
0.8291015625,
0.08905029296875,
0.8291015625
],
[
0.5009765625,
0.8173828125,
0.8857421875,
0.8173828125,
0.8857421875,
0.83056640625,
0.5009765625,
0.83056640625
],
[
0.08660888671875,
0.830078125,
0.4755859375,
0.83056640625,
0.4755859375,
0.84521484375,
0.08660888671875,
0.8447265625
],
[
0.5,
0.833984375,
0.8857421875,
0.833984375,
0.8857421875,
0.84765625,
0.5,
0.84765625
],
[
0.0882568359375,
0.8466796875,
0.4755859375,
0.8466796875,
0.4755859375,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.5009765625,
0.849609375,
0.8837890625,
0.849609375,
0.8837890625,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.0882568359375,
0.861328125,
0.4755859375,
0.861328125,
0.4755859375,
0.875,
0.0882568359375,
0.875
],
[
0.5,
0.86279296875,
0.88623046875,
0.86279296875,
0.88623046875,
0.876953125,
0.5,
0.876953125
],
[
0.0882568359375,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.5,
0.87890625,
0.8857421875,
0.87890625,
0.8857421875,
0.892578125,
0.5,
0.892578125
],
[
0.0882568359375,
0.892578125,
0.1690673828125,
0.892578125,
0.1690673828125,
0.90380859375,
0.0882568359375,
0.90380859375
]
] |
[
"Submission and Formatting Instructions for ICML 2020",
"iiii4",
"jiii4",
"",
"12",
"0.70.80.40.3",
"0.7×0.40.3",
"000",
"ui",
"uil",
"1",
"+exposure",
"",
"×0.90.2",
"+quantization",
"0.60.90.40.2",
"u2",
"u",
"0",
"n",
"0",
"noise Ni",
"noise N2",
"0.70.90.50.4",
"U3",
"0.70.90.5x",
"u3",
"O",
"0.50.60.70.4",
"u4×0.60.70.4",
"u",
"0",
"O",
"S",
"0.60.90.30.6",
"u50.6×0.30.6",
"us",
"T",
"(a) Rating matrix R",
"(b) Rating matrix with",
"(c) Interaction matrix",
"unseen data Ro+N,",
"R= Ro +Ni+ N2",
"Figure 1. The interaction matrix and different kinds of noise. Curves in Figure 1(a) indicate a user graph and an item graph. Two users are",
"connected if theyhave interacted withthe same item,and twoitems are connected iftheyhavebeen interacted bythe same user.",
"domain, the dependence is based on graph connections, for",
"Neural Network-based Recommendation. These years,",
"example,we consider thedependence of tworows/columns",
"Neural Network achieves significant success and is widely",
"of Ro based on the user/item graph shown in Figure 1(a). To",
"used in various machine learning and information retrieval",
"remove the graph noise, we devise a 2D graph filter called",
"domains,includingrecommendation.Chengetal.(2016):",
"LCF and reconstruct Ro by it: Ro = LCF(R), where Ro",
"He et al. (2017) utilized wide and deep networks to learn the",
"is the estimation. Similar to the filter, we also extend 2D",
"nonlinearcombination of embeddings.Chen etal.(2017)",
"convolution from the Euclidean domain to the graph domain",
"explored the attention mechanism to highlight important",
"to propose GCN. We then inject LCF into GCN to propose",
"elements. He & McAuley (2016b); Yu et al.(2018)ex-",
"our LCF Network (LCFN). LCF not only improves the",
"plored side information to enhance the performance of rec-",
"accuracy by removing the noise, but also reduces the com-",
"ommendation. Li et al. (2017); Tan et al. (2016) regarded",
"putation of graph convolution since only a small proportion",
"recommendation as a sequence prediction task and utilized",
"of eigen-vectors are required. To the best of our knowledge,",
"RNN or CNN to deal with this issue. Though extensively",
"LCFN is the first GCN to use the original graph convolution.",
"explored.thesemodelsarebasedonconventionalnetworks",
"and omitthegraph structure.",
"To summarize, this work makes the following contributions:",
"Graph-based Recommendation. As a link prediction task",
" We introduce the 2D graph Fourier transform, and",
"of a bipartite graph, graph tools are widely used in the",
"then design the graph filter LCF and graph convolution",
"recommendation domain. Rao et al. (2015) used graph reg-",
"based on it.",
"ularizationtosmooththeembeddingsofconnectednodes",
":We design a novel GCN named LCFN for recommen-",
"(users and items in recommendation context). Yu & Qin",
"dation with LCF to improve the performance by re-",
"(2019) extracted spectral features for nodes from the hy-",
"moving noise and to improve the efficiency of graph",
"pergraph Laplacian matrices to optimize model learning.",
"convolution.",
"However, these shallow models only get very limited in-",
": We devise comprehensive experiments on two real-",
"formation. Wang et al. (2019); Ying et al. (2018); Berg",
"world datasets to demonstrate the effectiveness and",
"et al. (2018) utilized deep GNNs for recommendation tasks.",
"efficiency of LCFN.",
"GNNs propagate embeddings through the graph to ensure",
"that connected nodes have similar latent representations.",
"Though GNNs perform well, there is still a large room for",
"2. Related Work",
"improvement ",
"GNNs only smooth the embeddings while",
"After the collaborative filtering (CF) model was proposed",
"cannot extract high-level patterns for the graph signal. To",
"(Sarwar et al., 2001; Goldberg, 1992), recommender sys-",
"address this gap, GCNs are proposed.",
"temshaveboomedindecades.AmongvariousCFmethods.",
"Shuman et al. (2013) proposed graph convolution to extract",
"matrix factorization (MF) (Koren, 2009; Rendle et al., 2009),",
"the dependence pattern from thegraph data.Sincegraph",
"which encodes the preference of users by underlying embed-",
"convolution on large graphs is extremely time-consuming,",
"dings (i.e.,latent factors),is a basic yet the most effective",
"Defferrard et al. (2016); Kipf & Welling (2017) first simpli-",
"recommender model. To improve the presentation capa-",
"fied it by various approximations, unfortunately, degenerate",
"bility, many variants have been proposed (He et al., 2017;",
"GCNstoGNNs.Byfollowingthem,existingGCNsare",
"Wang et al., 2019). Among these models, neural network-",
"indeed GNNs, or variants of GNNs (Yao & Li, 2018; Monti",
"based models and graph-based models are the most relevant",
"et al., 2017; Zheng et al., 2018). Also, as a result, many",
"to our work."
] |
{
"arxiv_id": "2006.15516",
"language": "en",
"page_idx": 1
}
|
||
2006.15516-en-2
|
2006.15516-en-2.jpg
|
Submission and Formatting Instructions for ICML 2020
GNNs are claimed as GCNs (Berg et al., 2018; Ying et al., 2018) (We follow the definition of GNN and GCN in (Zhou et al., 2018)). In this paper, we devise LCF to reduce the computation of graph convolution and to remove the noise. $\hat{\text{Different from existing GCNs, our simplification strategy}}$ enhances the graph convolution rather than degenerating it.
## $3. \textbf{Preliminaries}$
In this section, we introduce some background knowledge
of the signal processing in the Euclidean domain.
## 3.1. Fourier Transform
Fourier transform is a widely-used tool for frequency domain analysis. It converts the signal from the time domain inverse transform is $f(t)=\frac1{2\pi}\int_\mathbb{R}F(\omega)\mathrm{e}^{j\omega t}d\omega.$ For a digital signal (discrete signal) denoted as $\{s_n\}$, the Fourier transform is: $S_k=\sum_{n=0}^{N-1}(s_{n})\mathrm{e}^{-\frac{2\pi j}{N}kn}$ and the inverse transform is $s_n~=~\frac{1}{N}\sum_{k=0}^{N-1}(S_{k})\mathrm{e}^{\frac{2\pi j}{N}kn}$, where $j$ is the imaginary unit and $N$ is the length of the series $\{s_n\}.$
<FigureHere>
Figure 2. Signals with different frequencies in the time domain and
frequency domain.
In this paper, we denote the Fourier transform as $\mathcal{F}($ )and the inverse transform as $\mathcal{F} ^{- 1}($ ). Figure 2 shows some signals with different frequencies. Figures 2(a)(b)(c) show signals in the time domain and Figures 2(d)(e)(f) show signals in the frequency domain. $s_n^1$ in Figure 2(a) is a lowfrequency signal. In the time domain, $s_n^1$ is very smooth and shows strong dependence between two adjacent sampling points. In the frequency domain, the signal is distributed on the low frequency (shown in Figure 2(d)). In contrast, $s_n^2$ in Figure 2(b) is a high-frequency signal thus shows weak dependence and is distributed on the high frequency (shown in Figure 2(e)). $s_n^3$ in Figure 2(c) is the addition of $s_n^1$ and
$s_n^2.$ They are fully mixed in the time domain, nevertheless can be easily separated in the frequency domain (shown in Figure 2(f)). In telecommunication, low-frequency signals are usually contaminated by high-frequency noise when transmitting through the channel, and we can reconstruct the signal by a low-pass filter.
As discussed, the frequency in frequency domain reflects the dependence in time domain, however, in graph data, the dependence is based on node connections. Recalling the example shown in Figure 1, $u_1$ and $u_3$ are connected users, the two rows for them in $\mathbf{R}_0$ are (0.7,0.8,0.4,0.3) and (0.7,0.9,0.5,0.4), and in $\mathbf{N}_1$ are (0,-0.8,0,0) and (0,0,0,-0.4). We can see that the true preference of two connected nodes shows strong dependence thus $\mathbf{R}_0$ is lowfrequency, nevertheless the noise shows weak dependence
to the frequency domain. For a continuous signal
frequency graph signal by a low-pass graph filter.
## 3.2. Convolution
In addition to Fourier transform, convolution is a very important signal processing tool which is widely used in CNNs. For two continuous signals $f(t)$ and $g(t)$,the convolution of them is defined as: $f(t)*g(t)=\int_{\mathbb{R}}f(\tau)g(t-\tau)d\tau.$ For two digital signals $\{s_n\}_{n=0\cdots N-1}$ and $\{k_m\}_{m=0\cdots M-1}$,the convolution is: $s_{n}^{\prime }= \sum _{m= 0}^{M- 1}k_{m}( s_{n- m}) $,where $N$ and $M$ indicate the length of the signals. In machine learning tasks, $M$ is usually much smaller than $N$, and $\{k_m\}$ is called the convolutional kernel, $\{s_n^{\prime}\}$ is the feature map. We can extract local dependence information from the signal by convolution. Intuitively speaking, if $\{s_n\}$ is very similar to $\{k_m\}$ in a local area, $\{s_n^\prime\}$ will be very large in this position. We also introduce a widely-used property of convolution, which is crucial in graph convolution design: Convolution Theorem (Barrett & Wilde, 1960) tells us that convolution in the time domain is equivalent to the product in the frequency domain, i.e., $\hat f(t)*g(t)=\mathcal{F}^{-1}(\mathcal{F}(f(t))\mathcal{F}(g(t))).$ For graph data, convolution is hard to define directly due to the irregular structure of graphs, thus is defined in this way.
## 4.2D Low-pass Graph Convolution
In this paper, bold uppercase letters refer to matrices. Assuming there are $M$ users and $N$ items in total, we use matrix $\mathbf{R}\in\mathbb{R}^{M\times N}$ to denote the interactions between users and items. $\mathbf{R}_{ui}=1$ if user $u$ has interacted with item $i$ and $\mathbf{R}_ui=0$ otherwise. Our task is to predict the missing entities, i.e., 0 in R.
## 4.1. 2D Graph Fourier Transform
To propose LCF and the graph convolution, we need to propose graph Fourier transform first. Shuman et al. (2013)
|
[
[
0.3095703125,
0.05682373046875,
0.6650390625,
0.05682373046875,
0.6650390625,
0.0706787109375,
0.3095703125,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.47802734375,
0.08587646484375,
0.47802734375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.5,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.0882568359375,
0.10101318359375,
0.4755859375,
0.10101318359375,
0.4755859375,
0.1142578125,
0.0882568359375,
0.1142578125
],
[
0.5,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.5,
0.11553955078125
],
[
0.0882568359375,
0.11614990234375,
0.4755859375,
0.11614990234375,
0.4755859375,
0.1307373046875,
0.0882568359375,
0.1307373046875
],
[
0.5,
0.11871337890625,
0.8837890625,
0.11871337890625,
0.8837890625,
0.1287841796875,
0.5,
0.1287841796875
],
[
0.0882568359375,
0.1319580078125,
0.47802734375,
0.1319580078125,
0.47802734375,
0.1458740234375,
0.0882568359375,
0.1458740234375
],
[
0.498291015625,
0.1319580078125,
0.88623046875,
0.1319580078125,
0.88623046875,
0.1458740234375,
0.498291015625,
0.1458740234375
],
[
0.08660888671875,
0.1458740234375,
0.4755859375,
0.146484375,
0.4755859375,
0.1610107421875,
0.08660888671875,
0.160400390625
],
[
0.5,
0.147705078125,
0.88623046875,
0.147705078125,
0.88623046875,
0.1610107421875,
0.5,
0.1610107421875
],
[
0.0882568359375,
0.1622314453125,
0.4755859375,
0.1622314453125,
0.4755859375,
0.175537109375,
0.0882568359375,
0.175537109375
],
[
0.5009765625,
0.1622314453125,
0.69921875,
0.1622314453125,
0.69921875,
0.175537109375,
0.5009765625,
0.175537109375
],
[
0.5009765625,
0.184326171875,
0.8857421875,
0.184326171875,
0.8857421875,
0.1976318359375,
0.5009765625,
0.1976318359375
],
[
0.0882568359375,
0.1962890625,
0.223876953125,
0.1962890625,
0.223876953125,
0.210205078125,
0.0882568359375,
0.210205078125
],
[
0.5009765625,
0.199462890625,
0.88623046875,
0.199462890625,
0.88623046875,
0.21337890625,
0.5009765625,
0.21337890625
],
[
0.5,
0.2139892578125,
0.8857421875,
0.2139892578125,
0.8857421875,
0.2279052734375,
0.5,
0.2279052734375
],
[
0.08660888671875,
0.2203369140625,
0.473876953125,
0.2215576171875,
0.473876953125,
0.236083984375,
0.08660888671875,
0.23486328125
],
[
0.5,
0.2298583984375,
0.8857421875,
0.2298583984375,
0.8857421875,
0.24365234375,
0.5,
0.24365234375
],
[
0.0882568359375,
0.2374267578125,
0.41259765625,
0.2374267578125,
0.41259765625,
0.250732421875,
0.0882568359375,
0.250732421875
],
[
0.5,
0.2442626953125,
0.8857421875,
0.2442626953125,
0.8857421875,
0.25830078125,
0.5,
0.25830078125
],
[
0.5,
0.260009765625,
0.8857421875,
0.260009765625,
0.8857421875,
0.2734375,
0.5,
0.2734375
],
[
0.0882568359375,
0.268310546875,
0.250732421875,
0.268310546875,
0.250732421875,
0.2822265625,
0.0882568359375,
0.2822265625
],
[
0.5,
0.274658203125,
0.88623046875,
0.274658203125,
0.88623046875,
0.28857421875,
0.5,
0.28857421875
],
[
0.0882568359375,
0.292236328125,
0.473876953125,
0.292236328125,
0.473876953125,
0.30615234375,
0.0882568359375,
0.30615234375
],
[
0.5,
0.290283203125,
0.88623046875,
0.290283203125,
0.88623046875,
0.3037109375,
0.5,
0.3037109375
],
[
0.0882568359375,
0.30810546875,
0.473876953125,
0.30810546875,
0.473876953125,
0.3212890625,
0.0882568359375,
0.3212890625
],
[
0.498291015625,
0.304931640625,
0.8837890625,
0.304931640625,
0.8837890625,
0.318115234375,
0.498291015625,
0.318115234375
],
[
0.08905029296875,
0.324462890625,
0.473876953125,
0.324462890625,
0.473876953125,
0.335205078125,
0.08905029296875,
0.335205078125
],
[
0.5,
0.32080078125,
0.88623046875,
0.32080078125,
0.88623046875,
0.333984375,
0.5,
0.333984375
],
[
0.0849609375,
0.3359375,
0.476318359375,
0.335205078125,
0.476318359375,
0.352783203125,
0.0849609375,
0.354248046875
],
[
0.5009765625,
0.335205078125,
0.8857421875,
0.335205078125,
0.8857421875,
0.348388671875,
0.5009765625,
0.348388671875
],
[
0.08660888671875,
0.352783203125,
0.4755859375,
0.352783203125,
0.4755859375,
0.37060546875,
0.08660888671875,
0.37060546875
],
[
0.5,
0.35107421875,
0.826171875,
0.35107421875,
0.826171875,
0.3642578125,
0.5,
0.3642578125
],
[
0.08905029296875,
0.369384765625,
0.473876953125,
0.369384765625,
0.473876953125,
0.3837890625,
0.08905029296875,
0.3837890625
],
[
0.08416748046875,
0.380615234375,
0.47802734375,
0.37744140625,
0.47802734375,
0.40283203125,
0.08416748046875,
0.406005859375
],
[
0.5009765625,
0.3818359375,
0.61767578125,
0.3818359375,
0.61767578125,
0.395751953125,
0.5009765625,
0.395751953125
],
[
0.08251953125,
0.399658203125,
0.47802734375,
0.39453125,
0.47802734375,
0.419189453125,
0.08251953125,
0.4248046875
],
[
0.5,
0.40478515625,
0.88623046875,
0.40478515625,
0.88623046875,
0.41845703125,
0.5,
0.41845703125
],
[
0.08660888671875,
0.41845703125,
0.439453125,
0.419189453125,
0.439453125,
0.434326171875,
0.08660888671875,
0.43310546875
],
[
0.498291015625,
0.42041015625,
0.88623046875,
0.419189453125,
0.88623046875,
0.434326171875,
0.498291015625,
0.43505859375
],
[
0.5009765625,
0.436279296875,
0.8857421875,
0.436279296875,
0.8857421875,
0.449462890625,
0.5009765625,
0.449462890625
],
[
0.5009765625,
0.451904296875,
0.8857421875,
0.451904296875,
0.8857421875,
0.46533203125,
0.5009765625,
0.46533203125
],
[
0.1258544921875,
0.4697265625,
0.1690673828125,
0.4697265625,
0.1690673828125,
0.4892578125,
0.1258544921875,
0.4892578125
],
[
0.2484130859375,
0.47021484375,
0.303955078125,
0.472900390625,
0.30224609375,
0.4912109375,
0.2467041015625,
0.4892578125
],
[
0.5,
0.46533203125,
0.8857421875,
0.466552734375,
0.8857421875,
0.48095703125,
0.5,
0.479736328125
],
[
0.1192626953125,
0.472900390625,
0.13232421875,
0.477783203125,
0.12744140625,
0.486083984375,
0.11358642578125,
0.48095703125
],
[
0.30224609375,
0.4765625,
0.3095703125,
0.4765625,
0.3095703125,
0.48095703125,
0.30224609375,
0.48095703125
],
[
0.498291015625,
0.48095703125,
0.88623046875,
0.479736328125,
0.88623046875,
0.498779296875,
0.498291015625,
0.5
],
[
0.1846923828125,
0.4892578125,
0.2059326171875,
0.4892578125,
0.2059326171875,
0.49755859375,
0.1846923828125,
0.49755859375
],
[
0.265625,
0.4892578125,
0.330078125,
0.484130859375,
0.33251953125,
0.49755859375,
0.268798828125,
0.50244140625
],
[
0.5009765625,
0.49755859375,
0.88623046875,
0.49755859375,
0.88623046875,
0.51123046875,
0.5009765625,
0.51123046875
],
[
0.11767578125,
0.490478515625,
0.1217041015625,
0.490478515625,
0.1217041015625,
0.49560546875,
0.11767578125,
0.49560546875
],
[
0.2059326171875,
0.51513671875,
0.2156982421875,
0.51513671875,
0.2156982421875,
0.51953125,
0.2059326171875,
0.51953125
],
[
0.50244140625,
0.51318359375,
0.88623046875,
0.51318359375,
0.88623046875,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.143798828125,
0.521484375,
0.1993408203125,
0.521484375,
0.1993408203125,
0.5322265625,
0.143798828125,
0.5322265625
],
[
0.26953125,
0.52294921875,
0.32177734375,
0.52294921875,
0.32177734375,
0.52978515625,
0.26953125,
0.52978515625
],
[
0.39208984375,
0.52099609375,
0.443603515625,
0.52099609375,
0.443603515625,
0.53076171875,
0.39208984375,
0.53076171875
],
[
0.5,
0.5263671875,
0.88623046875,
0.52783203125,
0.88623046875,
0.54248046875,
0.5,
0.54150390625
],
[
0.101318359375,
0.5341796875,
0.2115478515625,
0.5341796875,
0.2115478515625,
0.5517578125,
0.101318359375,
0.5517578125
],
[
0.226318359375,
0.53515625,
0.336669921875,
0.53515625,
0.336669921875,
0.5498046875,
0.226318359375,
0.5498046875
],
[
0.358642578125,
0.53515625,
0.450927734375,
0.53515625,
0.450927734375,
0.546875,
0.358642578125,
0.546875
],
[
0.5009765625,
0.54345703125,
0.8857421875,
0.54345703125,
0.8857421875,
0.556640625,
0.5009765625,
0.556640625
],
[
0.5,
0.556640625,
0.88623046875,
0.5595703125,
0.88623046875,
0.57275390625,
0.5,
0.57080078125
],
[
0.2115478515625,
0.57373046875,
0.236083984375,
0.57373046875,
0.236083984375,
0.6162109375,
0.2115478515625,
0.6162109375
],
[
0.35302734375,
0.57373046875,
0.373291015625,
0.57373046875,
0.373291015625,
0.611328125,
0.35302734375,
0.611328125
],
[
0.50244140625,
0.57373046875,
0.88623046875,
0.57373046875,
0.88623046875,
0.58837890625,
0.50244140625,
0.58837890625
],
[
0.336669921875,
0.57470703125,
0.35546875,
0.57470703125,
0.35546875,
0.6162109375,
0.336669921875,
0.6162109375
],
[
0.098876953125,
0.57763671875,
0.10296630859375,
0.57763671875,
0.10296630859375,
0.58349609375,
0.098876953125,
0.58349609375
],
[
0.2344970703125,
0.5791015625,
0.256591796875,
0.58544921875,
0.2442626953125,
0.609375,
0.22216796875,
0.60302734375
],
[
0.23779296875,
0.57568359375,
0.24267578125,
0.57568359375,
0.24267578125,
0.583984375,
0.23779296875,
0.583984375
],
[
0.09722900390625,
0.5869140625,
0.10296630859375,
0.5869140625,
0.10296630859375,
0.6015625,
0.09722900390625,
0.6015625
],
[
0.11358642578125,
0.59033203125,
0.12091064453125,
0.59033203125,
0.12091064453125,
0.5947265625,
0.11358642578125,
0.5947265625
],
[
0.5009765625,
0.59521484375,
0.88623046875,
0.59521484375,
0.88623046875,
0.60986328125,
0.5009765625,
0.60986328125
],
[
0.23779296875,
0.60498046875,
0.2418212890625,
0.60498046875,
0.2418212890625,
0.61279296875,
0.23779296875,
0.61279296875
],
[
0.361083984375,
0.60791015625,
0.366943359375,
0.60791015625,
0.366943359375,
0.61572265625,
0.361083984375,
0.61572265625
],
[
0.5009765625,
0.611328125,
0.88623046875,
0.611328125,
0.88623046875,
0.625,
0.5009765625,
0.625
],
[
0.363525390625,
0.6171875,
0.369384765625,
0.6171875,
0.369384765625,
0.62353515625,
0.363525390625,
0.62353515625
],
[
0.36767578125,
0.62744140625,
0.461669921875,
0.6298828125,
0.461669921875,
0.64453125,
0.36767578125,
0.642578125
],
[
0.5,
0.625,
0.8857421875,
0.62548828125,
0.8857421875,
0.64013671875,
0.5,
0.6396484375
],
[
0.2467041015625,
0.6298828125,
0.338134765625,
0.6298828125,
0.338134765625,
0.6416015625,
0.2467041015625,
0.6416015625
],
[
0.125,
0.630859375,
0.214111328125,
0.630859375,
0.214111328125,
0.642578125,
0.125,
0.642578125
],
[
0.10296630859375,
0.64599609375,
0.2091064453125,
0.646484375,
0.2091064453125,
0.6611328125,
0.10296630859375,
0.66015625
],
[
0.2279052734375,
0.646484375,
0.330078125,
0.646484375,
0.330078125,
0.6611328125,
0.2279052734375,
0.6611328125
],
[
0.351318359375,
0.646484375,
0.450927734375,
0.646484375,
0.450927734375,
0.6611328125,
0.351318359375,
0.6611328125
],
[
0.5,
0.6416015625,
0.88623046875,
0.6416015625,
0.88623046875,
0.65478515625,
0.5,
0.65478515625
],
[
0.5009765625,
0.65576171875,
0.88623046875,
0.65576171875,
0.88623046875,
0.67041015625,
0.5009765625,
0.67041015625
],
[
0.5009765625,
0.67041015625,
0.88623046875,
0.67041015625,
0.88623046875,
0.68505859375,
0.5009765625,
0.68505859375
],
[
0.08905029296875,
0.67822265625,
0.473876953125,
0.67822265625,
0.473876953125,
0.69140625,
0.08905029296875,
0.69140625
],
[
0.5,
0.68505859375,
0.88818359375,
0.6875,
0.88818359375,
0.7021484375,
0.5,
0.69970703125
],
[
0.08905029296875,
0.693359375,
0.1993408203125,
0.693359375,
0.1993408203125,
0.7041015625,
0.08905029296875,
0.7041015625
],
[
0.0882568359375,
0.724609375,
0.473876953125,
0.7236328125,
0.473876953125,
0.73779296875,
0.0882568359375,
0.7392578125
],
[
0.5,
0.7197265625,
0.79833984375,
0.7216796875,
0.79833984375,
0.736328125,
0.5,
0.734375
],
[
0.08905029296875,
0.74072265625,
0.473876953125,
0.74072265625,
0.473876953125,
0.75390625,
0.08905029296875,
0.75390625
],
[
0.5,
0.7470703125,
0.88623046875,
0.7470703125,
0.88623046875,
0.76123046875,
0.5,
0.76123046875
],
[
0.0882568359375,
0.755859375,
0.473876953125,
0.755859375,
0.473876953125,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.5,
0.76220703125,
0.88623046875,
0.76220703125,
0.88623046875,
0.77587890625,
0.5,
0.77587890625
],
[
0.0882568359375,
0.771484375,
0.4755859375,
0.771484375,
0.4755859375,
0.78515625,
0.0882568359375,
0.78515625
],
[
0.498291015625,
0.7734375,
0.88818359375,
0.7734375,
0.88818359375,
0.79150390625,
0.498291015625,
0.79150390625
],
[
0.08905029296875,
0.7861328125,
0.4755859375,
0.7861328125,
0.4755859375,
0.7998046875,
0.08905029296875,
0.7998046875
],
[
0.5,
0.79248046875,
0.8857421875,
0.79248046875,
0.8857421875,
0.80615234375,
0.5,
0.80615234375
],
[
0.0882568359375,
0.8017578125,
0.473876953125,
0.8017578125,
0.473876953125,
0.81494140625,
0.0882568359375,
0.81494140625
],
[
0.5,
0.80810546875,
0.88623046875,
0.80810546875,
0.88623046875,
0.8212890625,
0.5,
0.8212890625
],
[
0.08905029296875,
0.818359375,
0.47216796875,
0.818359375,
0.47216796875,
0.8291015625,
0.08905029296875,
0.8291015625
],
[
0.5,
0.82275390625,
0.6337890625,
0.82275390625,
0.6337890625,
0.8359375,
0.5,
0.8359375
],
[
0.08905029296875,
0.83056640625,
0.47216796875,
0.83056640625,
0.47216796875,
0.8447265625,
0.08905029296875,
0.8447265625
],
[
0.08660888671875,
0.8447265625,
0.47216796875,
0.8466796875,
0.47216796875,
0.861328125,
0.08660888671875,
0.859375
],
[
0.5009765625,
0.85400390625,
0.73876953125,
0.85400390625,
0.73876953125,
0.86865234375,
0.5009765625,
0.86865234375
],
[
0.08905029296875,
0.8623046875,
0.473876953125,
0.8623046875,
0.473876953125,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.498291015625,
0.875,
0.8857421875,
0.875,
0.8857421875,
0.88818359375,
0.498291015625,
0.88818359375
],
[
0.0882568359375,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.498291015625,
0.88818359375,
0.88623046875,
0.88623046875,
0.88623046875,
0.90087890625,
0.498291015625,
0.90283203125
]
] |
[
"Submission and Formatting Instructions for ICML 2020",
"GNNs are claimed as GCNs (Berg et al., 2018; Ying et al.,",
"s2. They are fully mixed in the time domain, nevertheless",
"2018) (We follow the definition of GNN and GCN in (Zhou",
"can be easily separated in the frequency domain (shown in",
"et al., 2018)). In this paper, we devise LCF to reduce the",
"Figure2(f).Intelecommunication,low-frequencysignals",
"computation of graph convolution and to remove the noise.",
"are usually contaminated by high-frequency noise when",
"Different from existing GCNs, our simplification strategy",
"transmitting through the channel, and we can reconstruct",
"enhances the graph convolution rather than degenerating it.",
"the signal by a low-pass filter.",
"As discussed, the frequency in frequency domain reflects",
"3. Preliminaries",
"the dependence in time domain, however, in graph data,",
"the dependence is based on node connections. Recalling",
"In this section, we introduce some background knowledge",
"the example shown in Figure l, ui and u3 are connected",
"of the signal processing in the Euclidean domain.",
"users, the two rows for them in Ro are (0.7, 0.8, 0.4, 0.3)",
"and (0.7, 0.9, 0.5, 0.4), and in Ni are (0, -0.8, 0, 0) and",
"3.1. Fourier Transform",
"(0, 0, 0, -0.4). We can see that the true preference of two",
"Fourier transform is a widely-used tool for frequency do-",
"connected nodes shows strong dependence thus Ro is low-",
"main analysis. It converts the signal from the time domain",
"frequency, nevertheless the noise shows weak dependence",
"to thefrequency domain.For a continuous signal f(t)",
"due to the random position, thus is high-frequency. We",
"the Fourier transform is: F(w) = Je f(t)e-jwt dt and the",
"can remove the high-frequency graph noise from the low",
"inverse transform is f(t) = 2 Je F(w)ejwtdw. For a dig-",
"frequency graph signal by a low-pass graph filter.",
"ital signal (discrete signal) denoted as [sn], the Fourier",
"transform is: St = ZN-g(sn)e-x+kn and the inverse",
"3.2.Convolution",
"transform is sn (Sk)ekn, wherej is the",
"In addition to Fourier transform, convolution is a very im-",
"imaginary unit and N is the length of the series [sn].",
"portant signal processing tool which is widely used in CNNs.",
"For two continuous signals f(t) and g(t), the convolution",
"of them is defined as: f(t)*g(t)= J f(T)g(t -T)dt. For",
"E.",
"LL.",
"two digital signals (sn]n=0...N-i and [km)m=o...M-i, the",
"0.56",
"1",
"convolution is: $h = m= km(sn-m), where N and M",
"",
"TT",
"indicate the length of the signals. In machine learning tasks,",
"-",
"2",
"M is usually much smaller than N, and [km is called",
"Time index r",
"Timeindex",
"Time index",
"the convolutional kernel, (s', ] is the feature map. We can",
"(a) s =sin(n)",
"(b) s=sin(n)",
"(c) s3 =s+s2",
"extract local dependence information from the signal by",
"convolution. Intuitively speaking,if [sn) is very similar to",
"epnduy",
"864.",
"[km] in a local area, [s'] will be very large in this position.",
"en",
"",
"",
"o",
"",
"b",
"We also introduce a widely-used property of convolution,",
"C",
"",
"which is crucial in graph convolution design: Convolution",
"",
"Frequency index k",
"Theorem (Barrett & Wilde, 1960) tells us that convolution",
"Frequency index k",
"Frequency index k",
"(d) S=[F(s)I",
"(e) S=[F(s)",
"(f) S=IF(s)I",
"in the time domain is equivalent to the product in the fre-",
"quency domain, i.e., f(t)* g(t)=F-1(F(f(t))F(g(t))",
"For graph data, convolution is hard to define directly due to",
"Figure 2. Signals with different frequencies in the time domain and",
"the irregular structure of graphs, thus is defined in this way.",
"frequency domain.",
"In this paper, we denote the Fourier transform as F() and",
"4. 2D Low-pass Graph Convolution",
"the inverse transform as F-1( ). Figure 2 shows some",
"In this paper, bold uppercase letters refer to matrices. As-",
"signals with different frequencies. Figures 2(a)(b)(c) show",
"suming there are M users and N items in total, we use",
"signals in the time domain and Figures 2(d)(e)(f) show sig-",
"matrix R e RMxN to denote the interactions between",
"nals in the frequency domain. s, in Figure 2(a) is a low-",
"users and items. Ruz = 1 if user u has interacted with item",
"frequency signal. In the time domain, s, is very smooth and",
"i and Ruz = O otherwise. Our task is to predict the missing",
"shows strongdependencebetweentwoadjacentsampling",
"entities, i.e., O in R.",
"points. In the frequency domain, the signal is distributed on",
"the low frequency (shown in Figure 2(d). In contrast, s",
"4.1. 2D Graph Fourier Transform",
"in Figure 2(b) is a high-frequency signal thus shows weak",
"dependence and is distributed on the high frequency (shown",
"To propose LCF and the graph convolution, we need to",
"in Figure 2(e)). s3 in Figure 2(c) is the addition of sn and",
"propose graph Fourier transform first. Shuman et al. (2013)"
] |
{
"arxiv_id": "2006.15516",
"language": "en",
"page_idx": 2
}
|
||
2006.15516-en-4
|
2006.15516-en-4.jpg
|
Submission and Formatting Instructions for ICML 2020
while retain the low-frequency component, thus we leverage he gate function $\tilde{\mathbf{F}}=\begin{bmatrix}1&0\\0&0\end{bmatrix}$, where $\tilde{\mathbf{F}}\in\{0,1\}^{M\times N}$, $1\in\{1\}^{\Phi\times\Psi}$ is an all-one matrix and 0 is an all-zero matrix. The passband cut-off frequencies on the user dimension and item dimension are $\lambda_\Phi$ and $\sigma_\Psi$, respectively. The passed signal is $LCF(\mathbf{R})=\mathcal{F}_{g}^{-1}(\mathcal{F}_{g}(\mathbf{R})\odot\tilde{\mathbf{F}})$,where $\odot$ is the element-wise product. By tuning $\Phi$ and $\Psi$, we can remove $\mathbf{N}_1$ and $\mathbf{N}_2$ and retain $\mathbf{R}_0$ as much as possible. An intuitive example is represented in Appendix A.
Graph convolution. We extend convolution theorem to the graph domain: Given a convolutional kernel K$\in\mathbb{R}^M\times N$, the convolution of the signal (feature map) is defined as $\mathbf{R}^{\prime}=\mathbf{R}*_{g}\mathbf{K}=\mathcal{F}_{g}^{-1}(\mathcal{F}_{g}(\mathbf{R})\odot\tilde{\mathbf{K}})$,where$*_g$is the graph convolution and $\tilde{\mathbf{K}}=\mathcal{F}_g(\mathbf{K})$ is the kernel in the frequency domain.
Low- pass graph convolution. We first filter the signal $\mathbf{R}$ ferent level: $\left \{ \mathbf{U} ^{( l) }, \mathbf{V} ^{( l) }\right \} _{l= 0\cdots L}.$ by LCF F and then perform the convolution by a kernel $\tilde{\mathbf{K} } :$ the predictive embeddings to give the final prediction:
$\begin{array}{l}{\dot{\mathbf{R}}^{\prime}=\mathcal{F}_{g}^{-1}(\mathcal{F}_{g}(\mathbf{R})\odot\tilde{\mathbf{F}}\odot\tilde{\mathbf{K}})=\mathbf{P}((\mathbf{P}^{\mathrm{T}}\mathbf{RQ})\odot\tilde{\mathbf{F}}\odot\tilde{\mathbf{K}})\mathbf{Q}^{\mathrm{T}},\mathrm{we}}\\{\mathrm{have:}}\end{array}$
$$\mathbf{R}_{u^{\prime}i^{\prime}}^{\prime}=\sum_{\phi=1}^{M}\sum_{\psi=1}^{N}\left(\sum_{u=1}^{M}\sum_{i=1}^{N}\mathbf{R}_{ui}\mathbf{P}_{u\phi}\mathbf{Q}_{i\psi}\right)\tilde{\mathbf{F}}_{\phi\psi}\tilde{\mathbf{K}}_{\phi\psi}\mathbf{P}_{u^{\prime}\phi}\mathbf{Q}_{i^{\prime}\psi}\\=\sum_{\phi=1}^{\Phi}\sum_{\psi=1}^{\Psi}\left(\sum_{u=1}^{M}\sum_{i=1}^{N}\mathbf{R}_{ui}\mathbf{P}_{u\phi}\mathbf{Q}_{i\psi}\right)\tilde{\mathbf{K}}_{\phi\psi}\mathbf{P}_{u^{\prime}\phi}\mathbf{Q}_{i^{\prime}\psi}.$$
We can see that only the first $\Phi/\Psi$ eigen-vectors of $\mathbf{L}^U/\mathbf{L}^V$ are needed. We define $\bar{\mathbf{P}}=\mathbf{P}_{*,1:\Phi},\bar{\mathbf{Q}}=\mathbf{Q}_{*,1:\Psi}$,and $\bar{\mathbf{K}}=\tilde{\mathbf{K}}_{1:\Phi,1:\Psi}$, thus $\mathbf{R}^\prime=\bar{\mathbf{P}}((\bar{\mathbf{P}}^{\mathsf{T}}\mathbf{R}\bar{\mathbf{Q}})\odot\bar{\mathbf{K}})\bar{\mathbf{Q}}^{\mathsf{T}}.$
5. Low-pass Collaborative Filter Network
In this section, we propose LCFN and analyze it by compar-
ing with some widely-used GNNs and GCNs.
### 5.1. Model Structure
LCFN is a GCN with our proposed low-pass graph convolution. The structure is illustrated in Appendix B. We stack multiple convolutional layers to extract high-level features. $\left\{\mathbf{k}^{U^{(l)},\mathbf{k}^{V^{(l)}},\mathbf{T}^{(l)}}\}_{l=1,\cdots,L}\right\}$ are trainable parameters. We Details of the convolutional layer are introduced below. learn them by minimizing $\mathcal{L}$ with Adam (Kingma&Ba,
Low-pass convolution. In LCFN, we input the feature map from the previous layer $\mathbf{R}^{\prime}$ to the current layer. Considering these feature maps are large and dense, we represent them in a form of MF: $\mathbf{R} ^{\prime }\approx \mathbf{U} \mathbf{V} ^{\mathrm{T} }$, where $\mathbf{U}\in\mathbb{R}^M\times K$ and V$\in$ $\mathbb{R}^{N\times K}$ are embeddings. To avoid overfitting, we reduce the number of parameters by redefining the convolutional kernel as a rank- 1 matrix $\bar{K}$= $k^Uk^{V^{T}}$, where column vectors $\mathbf{k}^U\in\mathbb{R}^\Phi$ and k$^V\in\mathbb{R}^\Psi$ are kernels for user and item dimensions, respectively. In this case, performing convolution on the feature map $\mathbf{R}^{\prime}$ equals to performing convolution on the embeddings: $\mathbf{U}^{\prime}=\bar{\mathbf{P}}diag(\mathbf{k}^{U})\bar{\mathbf{P}}^{\mathsf{T}}\mathbf{U}$
$\operatorname { and} \mathbf{V} ^{\prime }= \bar{\mathbf{Q} } diag( \mathbf{k} ^{V}) \bar{\mathbf{Q} } ^{\mathrm{T} }\mathbf{V} .$
Transformation. We follow existing GNNs and GCNs (Zhou et al., 2018) to use a matrix T$\in\mathbb{R}^K\times K$ to transform (not graph Fourier transform) the convolution of embeddings $U^{\prime }$ and $V^{\prime }$ as $U^{\prime \prime }$= $U^{\prime }$T and $V^{\prime \prime }$= $V^{\prime }$T, respec- tively.
Activation. Experiments show that sigmoid function $\sigma($ )is the best choice for LCFN: $\mathbf{U}^\prime\prime\prime=\sigma(\tilde{\mathbf{U}}^{\prime\prime})$ and $\mathbf{V}^\prime\prime\prime=\sigma(\dot{\mathbf{V}}^{\prime\prime}).$ LCFN contains $L$ layers and we use upper corner mark $(l)$ to indicate parameters for the $l$-th layer. The input of each layer is the output of the previous layer, thus we have:
$$\left\{\begin{array}{c}\mathbf{U}^{(l)}=\sigma\left(\bar{\mathbf{P}}diag\left(\mathbf{k}^{U}^{(l)}\right)\bar{\mathbf{P}}^{\mathsf{T}}\mathbf{U}^{(l-1)}\mathbf{T}^{(l)}\right)\\\mathbf{V}^{(l)}=\sigma\left(\bar{\mathbf{Q}}diag\left(\mathbf{k}^{V(l)}\right)\bar{\mathbf{Q}}^{\mathsf{T}}\mathbf{V}^{(l-1)}\mathbf{T}^{(l)}\right)\end{array}\right..$$
(4)
After $L$ convolutional layers, we obtain embeddings in dif-
$$\hat{\mathbf{R}}=\begin{bmatrix}\mathbf{U}^{(0)},\cdots,\mathbf{U}^{(L)}\end{bmatrix}\begin{bmatrix}\mathbf{V}^{(0)},\cdots,\mathbf{V}^{(L)}\end{bmatrix}^{\mathsf{T}}.$$
(5)
To emphasize the effectiveness of our low-pass convolution, we choose the most simple yet widely-used way for embedding fusion (concatenation) and combination (inner product) (Wang et al., 2019; Zheng et al., 2018; Berg et al., 2018). Other advanced fusion and combination functions (He et al., 2017) are left to explore in future work.
### 5.2. Optimization and Prediction
To optimize our proposed model, we employ the BPR loss (Rendle et al., 2009). We construct pairwise training set ${\grave{D}}=\{(u,i,j)|\mathbf{R}_{ui}=1\bigwedge\mathbf{R}_{uj}=0\}$ to learn the model. For a tuple $(u,i,j)\in\mathcal{D}$, we maximize the margin between $u^\prime$s preference towards $i$ and $j$ by minimizing the BPR loss:
$$\mathcal{L}=\sum\limits_{(u,i,j)\in\mathcal{D}}-\ln\sigma\left(\hat{\mathbf{R}}_{ui}-\hat{\mathbf{R}}_{uj}\right)+\frac{\lambda}{2}\left(\left\|\mathbf{U}^{(0)}\right\|_F^2+\left\|\mathbf{V}^{(0)}\right\|_F^2+\left\|\mathbf{\Theta}\right\|_F^2\right),$$
where $\hat{\mathbf{R}}$ is given in Equation (5). $\mathbf{U}^{(0)},\mathbf{V}^{(0)}$ and $\Theta=$ 2015). When recommending for the user $u$, we rank items by $\hat{\mathbf{R}}_u$ and return the top items to $u.$ Of special notice is that by exploring the mean squared error (MSE) loss, our model can be extended to explicit feedbacks directly, which also suffer from the exposure and quantization noise
### 5.3. Complexity Analysis
Eigen-decomposition. Original graph convolution requires all eigen-vectors of Laplacian matrices $\mathbf{L}^U$ and $\mathbf{L}^I$ as the transform bases, which costs $O(M^{3}+N^{3})$ time. In our lowpass convolution, only front $\Phi/\Psi$ eigen-vectors are needed
|
[
[
0.3095703125,
0.05682373046875,
0.6650390625,
0.05682373046875,
0.6650390625,
0.0706787109375,
0.3095703125,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.4755859375,
0.08587646484375,
0.4755859375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.496826171875,
0.083984375,
0.69140625,
0.08331298828125,
0.69140625,
0.10101318359375,
0.496826171875,
0.10162353515625
],
[
0.2548828125,
0.10479736328125,
0.265625,
0.10479736328125,
0.265625,
0.10797119140625,
0.2548828125,
0.10797119140625
],
[
0.275390625,
0.10418701171875,
0.293212890625,
0.10418701171875,
0.293212890625,
0.10797119140625,
0.275390625,
0.10797119140625
],
[
0.2998046875,
0.1060791015625,
0.47216796875,
0.10418701171875,
0.47216796875,
0.121826171875,
0.2998046875,
0.12371826171875
],
[
0.5009765625,
0.10919189453125,
0.8857421875,
0.10919189453125,
0.8857421875,
0.12249755859375,
0.5009765625,
0.12249755859375
],
[
0.281005859375,
0.11871337890625,
0.2900390625,
0.11871337890625,
0.2900390625,
0.12371826171875,
0.281005859375,
0.12371826171875
],
[
0.08660888671875,
0.106689453125,
0.240234375,
0.10418701171875,
0.240234375,
0.121826171875,
0.08660888671875,
0.125
],
[
0.5,
0.121826171875,
0.88623046875,
0.121826171875,
0.88623046875,
0.1395263671875,
0.5,
0.1395263671875
],
[
0.08660888671875,
0.1287841796875,
0.476318359375,
0.1287841796875,
0.476318359375,
0.146484375,
0.08660888671875,
0.146484375
],
[
0.498291015625,
0.1395263671875,
0.8857421875,
0.1383056640625,
0.8857421875,
0.15283203125,
0.498291015625,
0.154052734375
],
[
0.0882568359375,
0.147705078125,
0.4755859375,
0.147705078125,
0.4755859375,
0.1610107421875,
0.0882568359375,
0.1610107421875
],
[
0.498291015625,
0.154052734375,
0.88623046875,
0.1546630859375,
0.88623046875,
0.169189453125,
0.498291015625,
0.1685791015625
],
[
0.0882568359375,
0.1634521484375,
0.473876953125,
0.1634521484375,
0.473876953125,
0.1767578125,
0.0882568359375,
0.1767578125
],
[
0.5,
0.1673583984375,
0.54345703125,
0.169189453125,
0.54150390625,
0.1849365234375,
0.498291015625,
0.18310546875
],
[
0.08905029296875,
0.177978515625,
0.47216796875,
0.177978515625,
0.47216796875,
0.1912841796875,
0.08905029296875,
0.1912841796875
],
[
0.08905029296875,
0.1925048828125,
0.473876953125,
0.1925048828125,
0.473876953125,
0.20703125,
0.08905029296875,
0.20703125
],
[
0.5009765625,
0.1925048828125,
0.8857421875,
0.1925048828125,
0.8857421875,
0.205810546875,
0.5009765625,
0.205810546875
],
[
0.08905029296875,
0.20703125,
0.473876953125,
0.20703125,
0.473876953125,
0.2203369140625,
0.08905029296875,
0.2203369140625
],
[
0.498291015625,
0.20703125,
0.88623046875,
0.205810546875,
0.88623046875,
0.2203369140625,
0.498291015625,
0.2215576171875
],
[
0.0882568359375,
0.222900390625,
0.346435546875,
0.222900390625,
0.346435546875,
0.236083984375,
0.0882568359375,
0.236083984375
],
[
0.5,
0.2259521484375,
0.8857421875,
0.2266845703125,
0.8857421875,
0.2412109375,
0.5,
0.240478515625
],
[
0.08905029296875,
0.2442626953125,
0.473876953125,
0.2442626953125,
0.473876953125,
0.25830078125,
0.08905029296875,
0.25830078125
],
[
0.5,
0.240478515625,
0.88623046875,
0.240478515625,
0.88623046875,
0.25390625,
0.5,
0.25390625
],
[
0.08660888671875,
0.2587890625,
0.473876953125,
0.255615234375,
0.473876953125,
0.2734375,
0.08660888671875,
0.276611328125
],
[
0.5009765625,
0.25390625,
0.85546875,
0.25390625,
0.85546875,
0.267578125,
0.5009765625,
0.267578125
],
[
0.08905029296875,
0.27587890625,
0.473876953125,
0.27587890625,
0.473876953125,
0.2890625,
0.08905029296875,
0.2890625
],
[
0.56396484375,
0.27978515625,
0.67724609375,
0.27978515625,
0.67724609375,
0.29345703125,
0.56396484375,
0.29345703125
],
[
0.71728515625,
0.27783203125,
0.82275390625,
0.27783203125,
0.82275390625,
0.292236328125,
0.71728515625,
0.292236328125
],
[
0.0882568359375,
0.290283203125,
0.47216796875,
0.290283203125,
0.47216796875,
0.304931640625,
0.0882568359375,
0.304931640625
],
[
0.560546875,
0.29736328125,
0.685546875,
0.300537109375,
0.685546875,
0.31494140625,
0.560546875,
0.3125
],
[
0.705078125,
0.29736328125,
0.826171875,
0.29736328125,
0.826171875,
0.31494140625,
0.705078125,
0.31494140625
],
[
0.86669921875,
0.29345703125,
0.8837890625,
0.29345703125,
0.8837890625,
0.302978515625,
0.86669921875,
0.302978515625
],
[
0.08660888671875,
0.306884765625,
0.473876953125,
0.30810546875,
0.473876953125,
0.322509765625,
0.08660888671875,
0.3212890625
],
[
0.08905029296875,
0.324462890625,
0.143798828125,
0.324462890625,
0.143798828125,
0.335205078125,
0.08905029296875,
0.335205078125
],
[
0.50244140625,
0.328857421875,
0.88623046875,
0.328857421875,
0.88623046875,
0.34228515625,
0.50244140625,
0.34228515625
],
[
0.0882568359375,
0.343505859375,
0.47216796875,
0.343505859375,
0.47216796875,
0.356689453125,
0.0882568359375,
0.356689453125
],
[
0.498291015625,
0.34521484375,
0.66845703125,
0.34228515625,
0.66845703125,
0.35986328125,
0.498291015625,
0.363037109375
],
[
0.7138671875,
0.3466796875,
0.8837890625,
0.3466796875,
0.8837890625,
0.35986328125,
0.7138671875,
0.35986328125
],
[
0.08905029296875,
0.359130859375,
0.473876953125,
0.359130859375,
0.473876953125,
0.37255859375,
0.08905029296875,
0.37255859375
],
[
0.5009765625,
0.3662109375,
0.85302734375,
0.3662109375,
0.85302734375,
0.380615234375,
0.5009765625,
0.380615234375
],
[
0.0882568359375,
0.37255859375,
0.473876953125,
0.37255859375,
0.473876953125,
0.39013671875,
0.0882568359375,
0.39013671875
],
[
0.08905029296875,
0.39013671875,
0.1258544921875,
0.39013671875,
0.1258544921875,
0.401611328125,
0.08905029296875,
0.401611328125
],
[
0.55810546875,
0.395751953125,
0.58837890625,
0.395751953125,
0.58837890625,
0.40771484375,
0.55810546875,
0.40771484375
],
[
0.5986328125,
0.396484375,
0.6259765625,
0.396484375,
0.6259765625,
0.40478515625,
0.5986328125,
0.40478515625
],
[
0.6259765625,
0.397705078125,
0.67333984375,
0.397705078125,
0.67333984375,
0.40771484375,
0.6259765625,
0.40771484375
],
[
0.861328125,
0.397705078125,
0.8837890625,
0.397705078125,
0.8837890625,
0.4091796875,
0.861328125,
0.4091796875
],
[
0.15283203125,
0.412353515625,
0.16259765625,
0.412353515625,
0.16259765625,
0.417236328125,
0.15283203125,
0.417236328125
],
[
0.17724609375,
0.41357421875,
0.1871337890625,
0.41357421875,
0.1871337890625,
0.417236328125,
0.17724609375,
0.417236328125
],
[
0.2099609375,
0.4140625,
0.2254638671875,
0.4140625,
0.2254638671875,
0.417236328125,
0.2099609375,
0.417236328125
],
[
0.346435546875,
0.416748046875,
0.473876953125,
0.421630859375,
0.47216796875,
0.439453125,
0.3447265625,
0.434326171875
],
[
0.22216796875,
0.42041015625,
0.330078125,
0.42041015625,
0.330078125,
0.43505859375,
0.22216796875,
0.43505859375
],
[
0.09234619140625,
0.419189453125,
0.1478271484375,
0.42236328125,
0.146240234375,
0.4375,
0.0906982421875,
0.434326171875
],
[
0.1641845703125,
0.421630859375,
0.191162109375,
0.421630859375,
0.191162109375,
0.43310546875,
0.1641845703125,
0.43310546875
],
[
0.5,
0.423583984375,
0.88623046875,
0.423583984375,
0.88623046875,
0.4375,
0.5,
0.4375
],
[
0.20751953125,
0.425537109375,
0.22216796875,
0.425537109375,
0.22216796875,
0.4287109375,
0.20751953125,
0.4287109375
],
[
0.1405029296875,
0.4248046875,
0.160888671875,
0.419189453125,
0.1666259765625,
0.431884765625,
0.146240234375,
0.4375
],
[
0.5009765625,
0.439453125,
0.88623046875,
0.439453125,
0.88623046875,
0.45263671875,
0.5009765625,
0.45263671875
],
[
0.346435546875,
0.4609375,
0.4453125,
0.464111328125,
0.4453125,
0.479248046875,
0.346435546875,
0.47607421875
],
[
0.5009765625,
0.455078125,
0.88623046875,
0.455078125,
0.88623046875,
0.468505859375,
0.5009765625,
0.468505859375
],
[
0.1422119140625,
0.466552734375,
0.1805419921875,
0.464111328125,
0.181396484375,
0.47607421875,
0.143798828125,
0.477783203125
],
[
0.223876953125,
0.46337890625,
0.3349609375,
0.46337890625,
0.3349609375,
0.477783203125,
0.223876953125,
0.477783203125
],
[
0.50244140625,
0.4697265625,
0.88623046875,
0.4697265625,
0.88623046875,
0.48291015625,
0.50244140625,
0.48291015625
],
[
0.1478271484375,
0.479736328125,
0.188720703125,
0.479736328125,
0.188720703125,
0.4873046875,
0.1478271484375,
0.4873046875
],
[
0.2254638671875,
0.479736328125,
0.25244140625,
0.479736328125,
0.25244140625,
0.488037109375,
0.2254638671875,
0.488037109375
],
[
0.5009765625,
0.484130859375,
0.88623046875,
0.484130859375,
0.88623046875,
0.49755859375,
0.5009765625,
0.49755859375
],
[
0.08660888671875,
0.5,
0.471435546875,
0.498779296875,
0.471435546875,
0.51318359375,
0.08660888671875,
0.5146484375
],
[
0.5,
0.498779296875,
0.7607421875,
0.498779296875,
0.7607421875,
0.51220703125,
0.5,
0.51220703125
],
[
0.08660888671875,
0.51318359375,
0.4755859375,
0.51513671875,
0.4755859375,
0.5341796875,
0.08660888671875,
0.5322265625
],
[
0.08905029296875,
0.53515625,
0.41259765625,
0.53515625,
0.41259765625,
0.5498046875,
0.08905029296875,
0.5498046875
],
[
0.5009765625,
0.53076171875,
0.73193359375,
0.53076171875,
0.73193359375,
0.54443359375,
0.5009765625,
0.54443359375
],
[
0.5,
0.5517578125,
0.8857421875,
0.5517578125,
0.8857421875,
0.56494140625,
0.5,
0.56494140625
],
[
0.5009765625,
0.564453125,
0.8857421875,
0.564453125,
0.8857421875,
0.57763671875,
0.5009765625,
0.57763671875
],
[
0.08905029296875,
0.57080078125,
0.435546875,
0.57080078125,
0.435546875,
0.58544921875,
0.08905029296875,
0.58544921875
],
[
0.5009765625,
0.57763671875,
0.8857421875,
0.57763671875,
0.8857421875,
0.59130859375,
0.5009765625,
0.59130859375
],
[
0.08905029296875,
0.5966796875,
0.4755859375,
0.5966796875,
0.4755859375,
0.611328125,
0.08905029296875,
0.611328125
],
[
0.5009765625,
0.59228515625,
0.88623046875,
0.59228515625,
0.88623046875,
0.60595703125,
0.5009765625,
0.60595703125
],
[
0.5,
0.60400390625,
0.86767578125,
0.60302734375,
0.86767578125,
0.6171875,
0.5,
0.61865234375
],
[
0.0882568359375,
0.6123046875,
0.3896484375,
0.6123046875,
0.3896484375,
0.62548828125,
0.0882568359375,
0.62548828125
],
[
0.498291015625,
0.6318359375,
0.88232421875,
0.6298828125,
0.88232421875,
0.6513671875,
0.498291015625,
0.65380859375
],
[
0.0882568359375,
0.642578125,
0.2344970703125,
0.642578125,
0.2344970703125,
0.65576171875,
0.0882568359375,
0.65576171875
],
[
0.515625,
0.6513671875,
0.56201171875,
0.6513671875,
0.56201171875,
0.66357421875,
0.515625,
0.66357421875
],
[
0.0882568359375,
0.66552734375,
0.473876953125,
0.66650390625,
0.473876953125,
0.68115234375,
0.0882568359375,
0.6796875
],
[
0.496826171875,
0.67236328125,
0.8857421875,
0.67041015625,
0.8857421875,
0.69140625,
0.496826171875,
0.693359375
],
[
0.08660888671875,
0.681640625,
0.473876953125,
0.68115234375,
0.473876953125,
0.69580078125,
0.08660888671875,
0.6962890625
],
[
0.5048828125,
0.6875,
0.640625,
0.689453125,
0.640625,
0.70849609375,
0.5048828125,
0.70556640625
],
[
0.08905029296875,
0.6962890625,
0.4755859375,
0.6962890625,
0.4755859375,
0.71044921875,
0.08905029296875,
0.71044921875
],
[
0.689453125,
0.6943359375,
0.8857421875,
0.693359375,
0.8857421875,
0.70849609375,
0.689453125,
0.708984375
],
[
0.08660888671875,
0.7109375,
0.451904296875,
0.7119140625,
0.451904296875,
0.7265625,
0.08660888671875,
0.72607421875
],
[
0.5009765625,
0.71533203125,
0.88623046875,
0.71533203125,
0.88623046875,
0.72900390625,
0.5009765625,
0.72900390625
],
[
0.08905029296875,
0.73486328125,
0.47216796875,
0.73486328125,
0.47216796875,
0.74853515625,
0.08905029296875,
0.74853515625
],
[
0.5009765625,
0.73095703125,
0.8857421875,
0.73095703125,
0.8857421875,
0.744140625,
0.5009765625,
0.744140625
],
[
0.498291015625,
0.744140625,
0.88623046875,
0.74560546875,
0.88623046875,
0.76025390625,
0.498291015625,
0.7587890625
],
[
0.0882568359375,
0.75048828125,
0.473876953125,
0.75048828125,
0.473876953125,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.5,
0.76025390625,
0.8837890625,
0.76025390625,
0.8837890625,
0.7734375,
0.5,
0.7734375
],
[
0.0906982421875,
0.767578125,
0.473876953125,
0.767578125,
0.473876953125,
0.77783203125,
0.0906982421875,
0.77783203125
],
[
0.0882568359375,
0.77880859375,
0.4755859375,
0.77880859375,
0.4755859375,
0.79345703125,
0.0882568359375,
0.79345703125
],
[
0.5009765625,
0.77783203125,
0.8857421875,
0.77783203125,
0.8857421875,
0.78857421875,
0.5009765625,
0.78857421875
],
[
0.08660888671875,
0.79150390625,
0.4755859375,
0.79443359375,
0.4755859375,
0.8125,
0.08660888671875,
0.8095703125
],
[
0.498291015625,
0.79052734375,
0.8447265625,
0.79052734375,
0.8447265625,
0.8037109375,
0.498291015625,
0.8037109375
],
[
0.0882568359375,
0.81005859375,
0.473876953125,
0.81005859375,
0.473876953125,
0.82470703125,
0.0882568359375,
0.82470703125
],
[
0.08660888671875,
0.8271484375,
0.366943359375,
0.82373046875,
0.366943359375,
0.84130859375,
0.08660888671875,
0.8447265625
],
[
0.5009765625,
0.82275390625,
0.67626953125,
0.82275390625,
0.67626953125,
0.8359375,
0.5009765625,
0.8359375
],
[
0.358642578125,
0.8291015625,
0.47216796875,
0.8291015625,
0.47216796875,
0.84228515625,
0.358642578125,
0.84228515625
],
[
0.0882568359375,
0.84228515625,
0.473876953125,
0.84326171875,
0.473876953125,
0.85791015625,
0.0882568359375,
0.85693359375
],
[
0.5009765625,
0.84521484375,
0.8857421875,
0.84521484375,
0.8857421875,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.08905029296875,
0.85986328125,
0.4755859375,
0.85986328125,
0.4755859375,
0.87353515625,
0.08905029296875,
0.87353515625
],
[
0.5,
0.85986328125,
0.88623046875,
0.85986328125,
0.88623046875,
0.875,
0.5,
0.875
],
[
0.0882568359375,
0.875,
0.4755859375,
0.875,
0.4755859375,
0.88818359375,
0.0882568359375,
0.88818359375
],
[
0.498291015625,
0.87548828125,
0.88623046875,
0.875,
0.88623046875,
0.8896484375,
0.498291015625,
0.89013671875
],
[
0.08660888671875,
0.8916015625,
0.473876953125,
0.8896484375,
0.473876953125,
0.90380859375,
0.08660888671875,
0.90576171875
],
[
0.5,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"Submission and Formatting Instructions for ICML 2020",
"while retain the low-frequency component, thus we leverage",
"and V' = Qdiag(k')Q'v.",
"",
"",
", where F e [0,1]MxN",
"Transformation.We follow existing GNNs and GCNs",
"C",
"the gate function F =",
"(Zhou et al., 2018) to use a matrix T e RK×K to transform",
"1 e (1]$x is an all-one matrix and o is an all-zero matrix.",
"(not graph Fourier transform) the convolution of embed-",
"The passband cut-off frequencies on the user dimension and",
"dings U' and V' as U\" = U'T and V\" = V'T, respec-",
"item dimension are A and y, respectively. The passed",
"tively.",
"signal is LCF(R) = F-1(F.(R) O F), where O is the",
"element-wise product. By tuning and , we can remove",
"Activation.Experiments show that sigmoid function o() is",
"Ni and N2 and retain Ro as much as possible. An intuitive",
"the best choice for LCFN: U\"=α(U\") and V\"=α(V\").",
"example is represented in Appendix A.",
"LCFN contains L layers and we use upper corner mark ()",
"Graph convolution. We extend convolution theorem to the",
"to indicate parameters for the l-th layer. The input of each",
"graph domain: Given a convolutional kernel K e IR MxN",
"layer is the output of the previous layer, thus we have:",
"the convolution of the signal (feature map) is defined as",
"U(t) =α (Pdiag ",
")PTU(t-1)T(n))",
"R' = R *g K = F-1(Fg(R) o K), where *g is the graph",
"V(t) = g (Qdiag (",
"() QTV(t-1)T())",
"(4)",
"convolution and K = F.(K) is the kernel in the frequency",
"domain.",
"After L convolutional layers, we obtain embeddings in dif-",
"Low-pass graph convolution. We first filter the signal R",
"ferent level: [u(), v(t)",
": We concatenate them as",
"by LCFF and then perform the convolution by a kernel K",
"the predictive embeddings to give the final prediction:",
"R'=F_'(F(R)oFoK)=P(PTRQ)oFoK)QT, we",
"have:",
"R=",
"IT(O)",
".....U",
"(5)",
"M",
"",
"",
"FouKsuPu'sQi\"",
"RuPusQ",
"Ruu=",
"",
"To emphasize the effectiveness of our low-pass convolution,",
"1.",
"T",
"we choose the most simple yet widely-used way for embed-",
"KuPusQi",
"dingfusion (concatenation)and combination (inner product)",
"",
"RuPusQiu",
"(Wang et al., 2019; Zheng et al., 2018; Berg et al., 2018)",
"=1=",
"11=1",
"Other advanced fusion and combination functions (He et al.,",
"We can see that only the first @/ eigen-vectors of LU/LV",
"2017) are left to explore in future work.",
"are needed.We define P = P+,1:,Q=Q*,1:, and",
"K=K1:Φ.1, thus R' =P((P'RQ)oK)Q",
"5.2.Optimization and Prediction",
"To optimize our proposed model, we employ the BPR loss",
"(Rendle et al., 2009). We construct pairwise training set",
"5. Low-pass Collaborative Filter Network",
"D= [(u, i, j)|Rui = 1 ^Ruj = 0) to learn the model. For",
"In this section, we propose LCFN and analyze it by compar-",
"a tuple (u, i, j) E D, we maximize the margin between u's",
"preference towards i and j by minimizing the BPR loss:",
"ing with some widely-used GNNs and GCNs.",
"C=Z-Ino(R-Ru)+(u()+()+O",
"5.1.Model Structure",
"(u,,j)eT",
"LCFN is a GCN with our proposed low-pass graph convolu-",
"where R is given in Equation (5). U(0), V(0) and @ =",
"tion. The structure is illustrated in Appendix B. We stack",
"[ku() kv( T() )",
"multiple convolutional layers to extract high-level features.",
"are trainable parameters.We",
"Details of the convolutional layer are introduced below.",
"learn them by minimizing C with Adam (Kingma & Ba,",
"Low-pass convolution. In LCFN, we input the feature map",
"2015). When recommending for the user u, we rank items",
"by Ru and return the top items to u. Of special notice is",
"from the previous layer R' to the current layer. Considering",
"that by exploring the mean squared error (MSE) loss, our",
"thesefeaturemaps arelargeanddense,werepresentthem",
"in a form of MF: R' ~ UVT, where U e RMxK and V E",
"modelcanbe extendedto explicitfeedbacksdirectly,which",
"RNxK are embeddings. To avoid overfitting, we reduce",
"also suffer from the exposure and quantization noise.",
"the number of parameters by redefining the convolutional",
"kernel as a rank-1 matrix K = k'kvT",
"5.3.Complexity Analysis",
",wherecolumn",
"vectors ku e R and kv e R are kernels for user and",
"Eigen-decomposition. Original graph convolution requires",
"item dimensions, respectively., In this case, performing",
"all eigen-vectors of Laplacian matrices LU and LI as the",
"convolution on the feature map R' equals to performing",
"transform bases, which costs O(M3+N3) time. In our low-",
"convolution on the embeddings: U'=Pdiag(k)PTU",
"pass convolution, only front @/ eigen-vectors are needed."
] |
{
"arxiv_id": "2006.15516",
"language": "en",
"page_idx": 4
}
|
||
2402.17437-en-2
|
2402.17437-en-2.jpg
|
duce emotion representation generated by a pretrained emotion classifier to learn and express specific types of emotions in the conversation. However, emotions expressed in responses are often diverse rather than specific (Lin et al.,2019). Therefore, Lin et al. (2019) utilize multiple professional emotion listeners to express various appropriate emotions. Majumder et al. (2020) group multiple conversation emotions by polarity and simulate the speaker's emotions to generate empathetic responses.
These methods focus on the overall emotions of the conversation and ignore nuanced emotions (Li et al., 2020). To capture nuanced emotions, Li et al. (2020) extract emotional words through the NRC Emotion Lexicons (Mohammad and Turney, 2013) and integrate them into the model. Gao et al.(2021) and Kim et al.(2021) introduce emotional cause detection models to capture emotional words and perceive nuanced emotions. Li et al. (2022) enhance emotional representation in the context with additional knowledge, which helps detect emotional words. Sabour et al. 2022 use commonsense reasoning knowledge to infer nuanced emotions in the conversation. Kim et al.2022 employ pre-trained models to detect word-level emotion and keywords to detect the nuanced emotion in dialogues. These methods detect emotional words with nuanced emotions in the language of the conversation and use static vectors such as word embeddings or VAD to represent emotional words.
Overall, early approaches ignore emotional words with nuanced emotions. Recent methods ignore two major characteristics of emotional words in linguistic expression: variability and correlation. Unlike these methods, we consider the two characteristics of emotional words and propose a dynamic emotion-semantic correlation model to better understand the conversation.
### 3 $\mathbf{Method}$
## 3.1 Task Formulation
Given a dialogue context $D=[U_1,U_2,...,U_M]$ of two interlocutors, our model needs to accurately perceive the emotions and semantics in the dialogue context and generate empathetic responses $Y$ $=[y_{1},y_{2},...,y_{j},y_{N}].$Here$,U_i=[w_{1}^{i},w_{2}^{i},...,w_{m_{i}}^{i}]$ represents the i-th utterance containing $m_i$ words. $Y$ is a response containing N words.
## 3.2 0verview
We propose ESCM, which takes into account the dynamic emotions and semantics in the dialogue and their correlations. The proposed model is a transformer-based model with encoder-decoder architecture. To accurately perceive the content of the dialogue, we mainly reconstruct the encoder. As shown in Figure 2, ESCM mainly consists of three parts: (1) a context encoder (Section 3.3), which is a standard encoder structure and is used to understand the semantics of the dialogue; (2) a dynamic correlation encoding module (Section 3.4), which includes the construction of dynamic correlation vectors and the encoding of a dynamic correlation graph convolutional network. It learns the correlations between emotions and semantics; (3) emotion and response predicting module (Section 3.5), which completes the functions of emotion prediction and response generation.
### 3.3 Context Encoder
As with previous methods (Li et al., 2020, 2022; Sabour et al., 2022), we concatenate the utterances of the dialogue context and prepend $[CLS]$ as the whole sequence token to form the context input $C==[CLS]\oplus U_{1}\oplus U_{2}\oplus...\oplus U_{M}.$ Here,$\oplus$ denotes the concatenation symbol. To input the context $C$ into the model, we convert $C$ into context word embeddings $E_c.$ Then, we sum up the word embeddings $E_c$, position embeddings, and state embeddings to form the semantic embeddings $\widetilde{E}_{c}.$ The state embeddings are used to distinguish between speaker or responder types and are randomly initialized. To understand the semantics of the dialogue, we feed the semantic embeddings $\widetilde{E}_{c}$ into the context encoder $Enc_{ctx}$ to obtain the context semantic representation $H_k:$
$$H_{ctx}=Enc_{ctx}(\widetilde{E}_{c})$$
(1)
where $H_ctx\in R^{L\times d}$,L is the length of the context sequence, and d represents the hidden size of the encoder.
3.4 Dynamic Correlation Encoding Module
Dynamic correlation encoding consists of two submodules: (1) Dynamic emotion-semantic vectors. It brings the model the ability to flexibly adjust emotions and semantics, making the representation of context more reasonable. (2) Dynamic correlation graph convolutional network. This module
|
[
[
0.1175537109375,
0.08612060546875,
0.496337890625,
0.08612060546875,
0.496337890625,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.0843505859375,
0.62890625,
0.08489990234375,
0.62890625,
0.09979248046875,
0.5087890625,
0.09857177734375
],
[
0.1158447265625,
0.1015625,
0.491943359375,
0.10272216796875,
0.491943359375,
0.11639404296875,
0.1158447265625,
0.11517333984375
],
[
0.51123046875,
0.10687255859375,
0.880859375,
0.10687255859375,
0.880859375,
0.1217041015625,
0.51123046875,
0.1217041015625
],
[
0.1175537109375,
0.1181640625,
0.490234375,
0.1181640625,
0.490234375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.12469482421875,
0.8818359375,
0.12469482421875,
0.8818359375,
0.1383056640625,
0.51123046875,
0.1383056640625
],
[
0.1158447265625,
0.1341552734375,
0.489501953125,
0.133056640625,
0.489501953125,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.509765625,
0.1395263671875,
0.8818359375,
0.14013671875,
0.8818359375,
0.155029296875,
0.509765625,
0.15380859375
],
[
0.1175537109375,
0.15087890625,
0.48779296875,
0.15087890625,
0.48779296875,
0.160888671875,
0.1175537109375,
0.160888671875
],
[
0.51123046875,
0.15673828125,
0.8818359375,
0.15673828125,
0.8818359375,
0.17041015625,
0.51123046875,
0.17041015625
],
[
0.1175537109375,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51123046875,
0.1728515625,
0.88330078125,
0.1728515625,
0.88330078125,
0.18701171875,
0.51123046875,
0.18701171875
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.509765625,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.509765625,
0.203125
],
[
0.1175537109375,
0.198974609375,
0.490234375,
0.198974609375,
0.490234375,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.204833984375,
0.8818359375,
0.204833984375,
0.8818359375,
0.218505859375,
0.509765625,
0.218505859375
],
[
0.1158447265625,
0.21435546875,
0.489501953125,
0.213134765625,
0.489501953125,
0.22802734375,
0.1158447265625,
0.2286376953125
],
[
0.509765625,
0.2203369140625,
0.88330078125,
0.2203369140625,
0.88330078125,
0.2340087890625,
0.509765625,
0.2340087890625
],
[
0.1175537109375,
0.2327880859375,
0.48779296875,
0.2327880859375,
0.48779296875,
0.243408203125,
0.1175537109375,
0.243408203125
],
[
0.51318359375,
0.2369384765625,
0.88330078125,
0.2369384765625,
0.88330078125,
0.25048828125,
0.51318359375,
0.25048828125
],
[
0.11419677734375,
0.2476806640625,
0.1822509765625,
0.2452392578125,
0.1822509765625,
0.260009765625,
0.11419677734375,
0.262451171875
],
[
0.51318359375,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.2666015625,
0.51318359375,
0.2666015625
],
[
0.1368408203125,
0.2666015625,
0.48779296875,
0.2666015625,
0.48779296875,
0.27978515625,
0.1368408203125,
0.27978515625
],
[
0.51123046875,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.28076171875,
0.51123046875,
0.28076171875
],
[
0.118408203125,
0.283935546875,
0.486083984375,
0.283935546875,
0.486083984375,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.509765625,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.509765625,
0.298583984375
],
[
0.1175537109375,
0.298583984375,
0.48779296875,
0.298583984375,
0.48779296875,
0.312255859375,
0.1175537109375,
0.312255859375
],
[
0.51318359375,
0.3017578125,
0.880859375,
0.3017578125,
0.880859375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.1158447265625,
0.3134765625,
0.486083984375,
0.3134765625,
0.486083984375,
0.32666015625,
0.1158447265625,
0.32666015625
],
[
0.5087890625,
0.317626953125,
0.880859375,
0.31640625,
0.880859375,
0.330810546875,
0.5087890625,
0.33203125
],
[
0.1175537109375,
0.330810546875,
0.489501953125,
0.330810546875,
0.489501953125,
0.344482421875,
0.1175537109375,
0.344482421875
],
[
0.509765625,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.347412109375,
0.509765625,
0.347412109375
],
[
0.1175537109375,
0.3466796875,
0.489501953125,
0.3466796875,
0.489501953125,
0.361083984375,
0.1175537109375,
0.361083984375
],
[
0.51123046875,
0.349853515625,
0.8818359375,
0.349853515625,
0.8818359375,
0.363525390625,
0.51123046875,
0.363525390625
],
[
0.1175537109375,
0.363525390625,
0.490234375,
0.363525390625,
0.490234375,
0.377197265625,
0.1175537109375,
0.377197265625
],
[
0.51318359375,
0.365234375,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.51318359375,
0.37890625
],
[
0.1158447265625,
0.380126953125,
0.48779296875,
0.380126953125,
0.48779296875,
0.393798828125,
0.1158447265625,
0.393798828125
],
[
0.509765625,
0.3818359375,
0.77734375,
0.3818359375,
0.77734375,
0.3955078125,
0.509765625,
0.3955078125
],
[
0.1175537109375,
0.3955078125,
0.489501953125,
0.3955078125,
0.489501953125,
0.4091796875,
0.1175537109375,
0.4091796875
],
[
0.118408203125,
0.410888671875,
0.491943359375,
0.410888671875,
0.491943359375,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.509765625,
0.410888671875,
0.68994140625,
0.410888671875,
0.68994140625,
0.424560546875,
0.509765625,
0.424560546875
],
[
0.1175537109375,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.441162109375,
0.1175537109375,
0.441162109375
],
[
0.51123046875,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44775390625,
0.51123046875,
0.44775390625
],
[
0.1175537109375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.1175537109375,
0.457275390625
],
[
0.51123046875,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.51123046875,
0.463134765625
],
[
0.1175537109375,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.1175537109375,
0.473876953125
],
[
0.51123046875,
0.466064453125,
0.87890625,
0.466064453125,
0.87890625,
0.479736328125,
0.51123046875,
0.479736328125
],
[
0.1175537109375,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.1175537109375,
0.4892578125
],
[
0.51123046875,
0.482177734375,
0.880859375,
0.482177734375,
0.880859375,
0.495361328125,
0.51123046875,
0.495361328125
],
[
0.1175537109375,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.1175537109375,
0.5048828125
],
[
0.51123046875,
0.496337890625,
0.880859375,
0.49755859375,
0.880859375,
0.5126953125,
0.51123046875,
0.51123046875
],
[
0.1158447265625,
0.50830078125,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.1158447265625,
0.521484375
],
[
0.51123046875,
0.51416015625,
0.8818359375,
0.51416015625,
0.8818359375,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.1158447265625,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.1158447265625,
0.5380859375
],
[
0.51318359375,
0.5322265625,
0.8818359375,
0.5322265625,
0.8818359375,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.11419677734375,
0.5390625,
0.486083984375,
0.5390625,
0.486083984375,
0.552734375,
0.11419677734375,
0.552734375
],
[
0.509765625,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56103515625,
0.509765625,
0.56103515625
],
[
0.1175537109375,
0.55810546875,
0.486083984375,
0.55810546875,
0.486083984375,
0.568359375,
0.1175537109375,
0.568359375
],
[
0.5087890625,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.576171875,
0.5087890625,
0.576171875
],
[
0.1175537109375,
0.57177734375,
0.489501953125,
0.57177734375,
0.489501953125,
0.5849609375,
0.1175537109375,
0.5849609375
],
[
0.51123046875,
0.57763671875,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.51123046875,
0.59130859375
],
[
0.1351318359375,
0.59130859375,
0.489501953125,
0.59130859375,
0.489501953125,
0.60546875,
0.1351318359375,
0.60546875
],
[
0.51123046875,
0.59326171875,
0.8818359375,
0.59423828125,
0.8818359375,
0.609375,
0.51123046875,
0.60791015625
],
[
0.1158447265625,
0.60791015625,
0.490234375,
0.60791015625,
0.490234375,
0.62158203125,
0.1158447265625,
0.62158203125
],
[
0.509765625,
0.60986328125,
0.8818359375,
0.61083984375,
0.8818359375,
0.62451171875,
0.509765625,
0.62353515625
],
[
0.1158447265625,
0.62451171875,
0.489501953125,
0.62451171875,
0.489501953125,
0.63818359375,
0.1158447265625,
0.63818359375
],
[
0.51318359375,
0.626953125,
0.8818359375,
0.626953125,
0.8818359375,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.11419677734375,
0.6396484375,
0.489501953125,
0.6396484375,
0.489501953125,
0.65283203125,
0.11419677734375,
0.65283203125
],
[
0.51123046875,
0.6435546875,
0.8818359375,
0.6435546875,
0.8818359375,
0.65673828125,
0.51123046875,
0.65673828125
],
[
0.1175537109375,
0.65673828125,
0.491943359375,
0.65673828125,
0.491943359375,
0.67041015625,
0.1175537109375,
0.67041015625
],
[
0.51318359375,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.51318359375,
0.6728515625
],
[
0.1175537109375,
0.6748046875,
0.48779296875,
0.6748046875,
0.48779296875,
0.68505859375,
0.1175537109375,
0.68505859375
],
[
0.51123046875,
0.67578125,
0.78271484375,
0.67578125,
0.78271484375,
0.689453125,
0.51123046875,
0.689453125
],
[
0.1158447265625,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.1158447265625,
0.703125
],
[
0.11419677734375,
0.70361328125,
0.312255859375,
0.705078125,
0.312255859375,
0.7197265625,
0.11419677734375,
0.71875
],
[
0.62060546875,
0.705078125,
0.771484375,
0.703125,
0.771484375,
0.72021484375,
0.62060546875,
0.72265625
],
[
0.85791015625,
0.7060546875,
0.88330078125,
0.7060546875,
0.88330078125,
0.72021484375,
0.85791015625,
0.72021484375
],
[
0.51318359375,
0.73828125,
0.8818359375,
0.73828125,
0.8818359375,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.11419677734375,
0.74462890625,
0.21826171875,
0.74462890625,
0.21826171875,
0.7587890625,
0.11419677734375,
0.7587890625
],
[
0.509765625,
0.755859375,
0.8818359375,
0.7548828125,
0.8818359375,
0.76953125,
0.509765625,
0.77099609375
],
[
0.1158447265625,
0.779296875,
0.300537109375,
0.779296875,
0.300537109375,
0.79345703125,
0.1158447265625,
0.79345703125
],
[
0.51318359375,
0.77392578125,
0.57666015625,
0.77392578125,
0.57666015625,
0.78515625,
0.51318359375,
0.78515625
],
[
0.1175537109375,
0.80712890625,
0.490234375,
0.80712890625,
0.490234375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.8017578125,
0.8623046875,
0.8017578125,
0.8623046875,
0.8154296875,
0.509765625,
0.8154296875
],
[
0.1158447265625,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.509765625,
0.82373046875,
0.8818359375,
0.822265625,
0.8818359375,
0.83740234375,
0.509765625,
0.83837890625
],
[
0.11419677734375,
0.84033203125,
0.491943359375,
0.8388671875,
0.491943359375,
0.85400390625,
0.11419677734375,
0.85498046875
],
[
0.509765625,
0.84130859375,
0.88330078125,
0.84130859375,
0.88330078125,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5087890625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.5087890625,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.88720703125,
0.1175537109375,
0.88720703125
],
[
0.5087890625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.5087890625,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.491943359375,
0.8896484375,
0.491943359375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.89013671875,
0.880859375,
0.89013671875,
0.880859375,
0.90087890625,
0.51123046875,
0.90087890625
],
[
0.1175537109375,
0.90478515625,
0.38623046875,
0.90478515625,
0.38623046875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"duce emotion representation generated by a pre-",
"3.2Overview",
"trained emotion classifier to learn and express spe-",
"We propose ESCM, which takes into account the",
"cific types of emotions in the conversation. How-",
"dynamic emotions and semantics in the dialogue",
"ever, emotions expressedin responses areoften di-",
"and their correlations. The proposed model is a",
"verse rather than specific (Lin et al. 2019). There-",
"transformer-based model with encoder-decoder ar-",
"fore, Lin et al. (2019) utilize multiple professional",
"chitecture. To accurately perceive the content of",
"emotion listeners to express various appropriate",
"the dialogue, we mainly reconstruct the encoder.",
"emotions. Majumder et al. (2020) group multi-",
"As shown in Figure 2, ESCM mainly consists of",
"ple conversation emotions by polarity and simulate",
"three parts: (1) a context encoder (Section 3.3),",
"the speaker'semotionstogenerateempatheticre-",
"which is a standard encoder structure and is used to",
"sponses.",
"understand the semantics of the dialogue; (2) a dy",
"These methods focus on the overall emotions of",
"namiccorrelationencodingmodule(Section3.4)",
"the conversation and ignore nuanced emotions (Li",
"which includes the construction of dynamic corre",
"et al., 2020).To capture nuanced emotions, Li",
"lation vectors and the encoding of a dynamic cor",
"et al. (2020) extract emotional words through the",
"relation graph convolutional network. It learns the",
"NRC Emotion Lexicons (Mohammad and Turney,",
"correlations between emotions and semantics; (3)",
"2013) and integrate them into the model. Gao",
"emotion and response predicting module (Section",
"et al. (2021) and Kim et al. (2021) introduce emo-",
"3.5), which completes the functions of emotion",
"tional cause detection models to capture emotional",
" prediction and response generation.",
"words and perceive nuanced emotions. Li et al.",
"(2022) enhance emotional representation in the con-",
"3.3Context Encoder",
"text with additional knowledge, which helps detect",
"As with previous methods (Li et al., 2020, 2022:",
"emotional words.Sabour et al. 2022 use com-",
"Sabour et al., 2022), we concatenate the utterances",
"monsense reasoning knowledge to infer nuanced",
"of the dialogue context and prepend [CLS as the",
"emotions in the conversation.Kim et al. 2022",
"whole sequence token to form the context input",
"employ pre-trained models to detect word-level",
"C== [CLS] Ui U2 ④... ④ UM. Here, ④",
"emotion and keywords to detect the nuanced emo-",
"denotes the concatenation symbol. To input the",
"tion in dialogues. These methods detect emotional",
"context C into the model.we convert C into con-",
"words with nuanced emotions in the language of",
"text word embeddings Ec. Then, we sum up the",
"theconversationandusestaticvectorssuchasword",
"word embeddings Ec, position embeddings, and",
"embeddings or VAD to represent emotional words.",
"state embeddings to form the semantic embeddings",
"Overall, early approaches ignore emotional",
"Ee. The state embeddings are used to distinguish",
"words with nuanced emotions. Recent methods ig-",
"between speaker or responder types and are ran-",
"nore two major characteristics of emotional words",
"domly initialized. To understand the semantics of",
"in linguistic expression: variability and correlation.",
"the dialogue, we feed the semantic embeddings",
"Unlike these methods, we consider the two charac-",
"Ec into the context encoder Encctr to obtain the",
"teristics ofemotional words andproposeadynamic",
"context semantic representation Hk:",
"emotion-semantic correlation model to better un-",
"derstand the conversation.",
"Hct = Encctr(E)",
"(1)",
"where Hctr E RLxd, L is the length of the context",
"3Method",
"sequence, and d represents the hidden size of the",
"3.1TaskFormulation",
"encoder.",
"Given a dialogue context D = [Ui, U2, .., UM] of",
"3.4 Dynamic Correlation Encoding Module",
"two interlocutors, our model needs to accurately",
"Dynamic correlation encoding consists of two sub-",
"perceive the emotions and semantics in the dia-",
"modules: (1) Dynamic emotion-semantic vectors.",
"logue context and generate empathetic responses Y",
"It brings the model the ability to flexibly adjust",
"- [yi, y2, ...Yj, yn]. Here, U, = [wi, w?, .., wm,]",
" emotions and semantics, making the representation",
"represents the i-th utterance containing mi words.",
"of contextmorereasonable.(2)Dynamic corre-",
"Y is a response containing N words.",
"lation graph convolutional network. This module"
] |
{
"arxiv_id": "2402.17437",
"language": "en",
"page_idx": 2
}
|
||
1707.04913-en-0
|
1707.04913-en-0.jpg
|
$\textbf{End- to- End Information Extraction without Token- Level Supervision}$
Rasmus Berg Palm
Dirk Hovy
Computer Science Dpeartment
University of Copenhagen dirk.hovy@di.ku.dk
DTU Compute
Technical University of Denmark
rapal@dtu.dk
Florian Laws Tradeshift
Ole Winther DTU Compute
Technical University of Denmark
$\circ$Iwigdtu.dk
Landemærket 10, 1119 Copenhagen
flagtradeshift.com
## Abstract
Most state-of-the-art information extraction approaches rely on token-level labels to find the areas of interest in text. Unfortunately, these labels are time-consuming and costly to create, and consequently, not available for many real-life IE tasks. To make matters worse, token-level labels are usually not the desired output, but just an intermediary step. End-to-end (E2E) models, which take raw text as input and produce the desired output directly, need not depend on token-level labels. We propose an E2E model based on pointer networks, which can be trained directly on pairs of raw input and output text. We evaluate our model on the ATIS data set, MIT restaurant corpus and the MIT movie corpus and compare to neural baselines that do use token-level labels. We achieve competitive results, within a few percentage points of the baselines, showing the feasibility of E2E information extraction without the need for token-level labels. This opens up new possibilities, as for many tasks currently addressed by human extractors, raw input and output data are available, but not token-level labels.
One of the main bottlenecks is a shortage of the data required to train state-of-the-art IE models, which rely on sequence tagging (Finkel et al., 2005; Zhai et al., 2017). Such models require sufficient amounts of training data that is labeled at the token-level, i.e., with one label for each word. The reason token-level labels are in short supply is that they are not the intended output of human IE tasks. Creating token-level labels thus requires an additional effort, essentially doubling the work required to process each item. This additional effort is expensive and infeasible for many production systems: estimates put the average cost for a sentence at about 3 dollars, and about half an hour annotator time (Alonso et al., 2016). Consequently, state-of-the-art IE approaches, relying on sequence taggers, cannot be applied to many real life IE tasks.
What is readily available in abundance and at no additional costs, is the raw, unstructured input and machine-readable output to a human IE task. Consider the transcription of receipts, checks, or business documents, where the input is an unstructured PDF and the output a row in a database (due date, payable amount, etc). Another example is flight bookings, where the input is a natural language request from the user, and the output a HTTP request, sent to the airline booking API.
To better exploit such existing data sources, we propose an end-to-end (E2E) model based on pointer networks with attention, which can be trained end-to-end on the input/output pairs of human IE tasks, without requiring token-level annotations.
We evaluate our model on three traditional IE data sets. Note that our model and the baselines are competing in two dimensions. The frst is cost and applicability. The baselines require tokenlevel labels, which are expensive and unavailable for many real life tasks. Our model does not re-
## 1 Introduction
Humans spend countless hours extracting structured machine readable information from unstructured information in a multitude of domains. Promising to automate this, information extraction (IE) is one of the most sought-after industrial applications of natural language processing. However, despite substantial research efforts, in practice, many applications still rely on manual effort to extract the relevant information.
|
[
[
0.1435546875,
0.0736083984375,
0.85791015625,
0.0736083984375,
0.85791015625,
0.0885009765625,
0.1435546875,
0.0885009765625
],
[
0.255126953125,
0.1009521484375,
0.42236328125,
0.1009521484375,
0.42236328125,
0.11517333984375,
0.255126953125,
0.11517333984375
],
[
0.62060546875,
0.1009521484375,
0.71533203125,
0.1009521484375,
0.71533203125,
0.11517333984375,
0.62060546875,
0.11517333984375
],
[
0.273681640625,
0.11639404296875,
0.39794921875,
0.1175537109375,
0.39794921875,
0.1318359375,
0.273681640625,
0.1312255859375
],
[
0.5439453125,
0.1175537109375,
0.79248046875,
0.1175537109375,
0.79248046875,
0.1318359375,
0.5439453125,
0.1318359375
],
[
0.20068359375,
0.1341552734375,
0.470947265625,
0.1341552734375,
0.470947265625,
0.1484375,
0.20068359375,
0.1484375
],
[
0.56103515625,
0.133056640625,
0.77490234375,
0.1341552734375,
0.77490234375,
0.1484375,
0.56103515625,
0.1478271484375
],
[
0.26123046875,
0.15087890625,
0.41064453125,
0.149658203125,
0.41064453125,
0.1644287109375,
0.26123046875,
0.1650390625
],
[
0.55908203125,
0.15087890625,
0.77685546875,
0.15087890625,
0.77685546875,
0.1644287109375,
0.55908203125,
0.1644287109375
],
[
0.274658203125,
0.179931640625,
0.393798828125,
0.179931640625,
0.393798828125,
0.19482421875,
0.274658203125,
0.19482421875
],
[
0.611328125,
0.177001953125,
0.724609375,
0.1787109375,
0.724609375,
0.19482421875,
0.611328125,
0.1923828125
],
[
0.291259765625,
0.19775390625,
0.376220703125,
0.19775390625,
0.376220703125,
0.208984375,
0.291259765625,
0.208984375
],
[
0.60546875,
0.1953125,
0.73046875,
0.196533203125,
0.73046875,
0.21142578125,
0.60546875,
0.210205078125
],
[
0.187255859375,
0.213134765625,
0.480224609375,
0.213134765625,
0.480224609375,
0.22802734375,
0.187255859375,
0.22802734375
],
[
0.5322265625,
0.213134765625,
0.80419921875,
0.213134765625,
0.80419921875,
0.22802734375,
0.5322265625,
0.22802734375
],
[
0.22412109375,
0.2286376953125,
0.444091796875,
0.2298583984375,
0.444091796875,
0.24462890625,
0.22412109375,
0.243408203125
],
[
0.60107421875,
0.23095703125,
0.736328125,
0.23095703125,
0.736328125,
0.24462890625,
0.60107421875,
0.24462890625
],
[
0.264404296875,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.264404296875,
0.279052734375
],
[
0.53076171875,
0.26416015625,
0.884765625,
0.265380859375,
0.884765625,
0.27978515625,
0.53076171875,
0.279052734375
],
[
0.5146484375,
0.281982421875,
0.8818359375,
0.281982421875,
0.8818359375,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.14697265625,
0.292236328125,
0.459228515625,
0.292236328125,
0.459228515625,
0.305908203125,
0.14697265625,
0.305908203125
],
[
0.5146484375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.5146484375,
0.312255859375
],
[
0.145263671875,
0.308837890625,
0.459228515625,
0.30810546875,
0.459228515625,
0.322509765625,
0.145263671875,
0.32373046875
],
[
0.51318359375,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.32666015625,
0.51318359375,
0.32666015625
],
[
0.1485595703125,
0.32666015625,
0.45751953125,
0.32666015625,
0.45751953125,
0.33740234375,
0.1485595703125,
0.33740234375
],
[
0.019317626953125,
0.314208984375,
0.060455322265625,
0.314208984375,
0.06378173828125,
0.439453125,
0.0218353271484375,
0.439453125
],
[
0.51123046875,
0.328857421875,
0.884765625,
0.330078125,
0.884765625,
0.344482421875,
0.51123046875,
0.34326171875
],
[
0.145263671875,
0.34033203125,
0.459228515625,
0.341552734375,
0.459228515625,
0.355712890625,
0.145263671875,
0.355224609375
],
[
0.51318359375,
0.3466796875,
0.880859375,
0.3466796875,
0.880859375,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.14697265625,
0.358154296875,
0.4609375,
0.358154296875,
0.4609375,
0.371826171875,
0.14697265625,
0.371826171875
],
[
0.52880859375,
0.361083984375,
0.8818359375,
0.363525390625,
0.8818359375,
0.377685546875,
0.52880859375,
0.3759765625
],
[
0.14697265625,
0.37353515625,
0.4609375,
0.37353515625,
0.4609375,
0.38720703125,
0.14697265625,
0.38720703125
],
[
0.51318359375,
0.37890625,
0.88330078125,
0.37890625,
0.88330078125,
0.392578125,
0.51318359375,
0.392578125
],
[
0.14697265625,
0.3896484375,
0.459228515625,
0.3896484375,
0.459228515625,
0.402587890625,
0.14697265625,
0.402587890625
],
[
0.5146484375,
0.394287109375,
0.884765625,
0.394287109375,
0.884765625,
0.4091796875,
0.5146484375,
0.4091796875
],
[
0.14697265625,
0.40625,
0.4609375,
0.40625,
0.4609375,
0.419189453125,
0.14697265625,
0.419189453125
],
[
0.5146484375,
0.410888671875,
0.88330078125,
0.410888671875,
0.88330078125,
0.424560546875,
0.5146484375,
0.424560546875
],
[
0.14697265625,
0.421630859375,
0.4609375,
0.421630859375,
0.4609375,
0.435791015625,
0.14697265625,
0.435791015625
],
[
0.03863525390625,
0.431640625,
0.0487060546875,
0.431640625,
0.0487060546875,
0.43701171875,
0.03863525390625,
0.43701171875
],
[
0.51318359375,
0.427490234375,
0.8818359375,
0.427490234375,
0.8818359375,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.0218353271484375,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.14697265625,
0.438232421875,
0.459228515625,
0.438232421875,
0.459228515625,
0.451904296875,
0.14697265625,
0.451904296875
],
[
0.51318359375,
0.443603515625,
0.88330078125,
0.443603515625,
0.88330078125,
0.457763671875,
0.51318359375,
0.457763671875
],
[
0.14697265625,
0.45361328125,
0.462646484375,
0.45361328125,
0.462646484375,
0.46728515625,
0.14697265625,
0.46728515625
],
[
0.5146484375,
0.460205078125,
0.884765625,
0.460205078125,
0.884765625,
0.473876953125,
0.5146484375,
0.473876953125
],
[
0.14697265625,
0.47021484375,
0.459228515625,
0.47021484375,
0.459228515625,
0.48388671875,
0.14697265625,
0.48388671875
],
[
0.51318359375,
0.4755859375,
0.884765625,
0.4755859375,
0.884765625,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.14697265625,
0.486328125,
0.459228515625,
0.486328125,
0.459228515625,
0.49951171875,
0.14697265625,
0.49951171875
],
[
0.51318359375,
0.493408203125,
0.87890625,
0.493408203125,
0.87890625,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.14697265625,
0.5029296875,
0.4609375,
0.5029296875,
0.4609375,
0.5166015625,
0.14697265625,
0.5166015625
],
[
0.51318359375,
0.50830078125,
0.88330078125,
0.50830078125,
0.88330078125,
0.521484375,
0.51318359375,
0.521484375
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.14697265625,
0.5185546875,
0.459228515625,
0.5185546875,
0.459228515625,
0.5322265625,
0.14697265625,
0.5322265625
],
[
0.51318359375,
0.52490234375,
0.88330078125,
0.52490234375,
0.88330078125,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.14697265625,
0.53369140625,
0.45751953125,
0.53369140625,
0.45751953125,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.51123046875,
0.53759765625,
0.609375,
0.5390625,
0.609375,
0.55419921875,
0.51123046875,
0.5517578125
],
[
0.14697265625,
0.5498046875,
0.45751953125,
0.5498046875,
0.45751953125,
0.5634765625,
0.14697265625,
0.5634765625
],
[
0.5322265625,
0.5556640625,
0.884765625,
0.5556640625,
0.884765625,
0.5693359375,
0.5322265625,
0.5693359375
],
[
0.145263671875,
0.56640625,
0.4609375,
0.5654296875,
0.4609375,
0.580078125,
0.145263671875,
0.58056640625
],
[
0.51318359375,
0.57177734375,
0.88330078125,
0.57177734375,
0.88330078125,
0.5849609375,
0.51318359375,
0.5849609375
],
[
0.14697265625,
0.5830078125,
0.459228515625,
0.5830078125,
0.459228515625,
0.5966796875,
0.14697265625,
0.5966796875
],
[
0.51318359375,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.14697265625,
0.5986328125,
0.459228515625,
0.5986328125,
0.459228515625,
0.6123046875,
0.14697265625,
0.6123046875
],
[
0.51318359375,
0.60498046875,
0.88330078125,
0.60498046875,
0.88330078125,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.145263671875,
0.61328125,
0.459228515625,
0.6142578125,
0.459228515625,
0.62890625,
0.145263671875,
0.62744140625
],
[
0.5146484375,
0.62158203125,
0.8818359375,
0.62158203125,
0.8818359375,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.14697265625,
0.63134765625,
0.4609375,
0.63134765625,
0.4609375,
0.64453125,
0.14697265625,
0.64453125
],
[
0.51318359375,
0.63525390625,
0.8818359375,
0.6357421875,
0.8818359375,
0.65087890625,
0.51318359375,
0.6494140625
],
[
0.14697265625,
0.64794921875,
0.4609375,
0.64794921875,
0.4609375,
0.6611328125,
0.14697265625,
0.6611328125
],
[
0.5146484375,
0.65283203125,
0.88330078125,
0.65283203125,
0.88330078125,
0.66650390625,
0.5146484375,
0.66650390625
],
[
0.14697265625,
0.66455078125,
0.4609375,
0.66455078125,
0.4609375,
0.677734375,
0.14697265625,
0.677734375
],
[
0.51318359375,
0.66943359375,
0.884765625,
0.66943359375,
0.884765625,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.14697265625,
0.6787109375,
0.45751953125,
0.6787109375,
0.45751953125,
0.6923828125,
0.14697265625,
0.6923828125
],
[
0.51123046875,
0.68505859375,
0.88330078125,
0.68408203125,
0.88330078125,
0.69873046875,
0.51123046875,
0.69970703125
],
[
0.14697265625,
0.6953125,
0.462646484375,
0.6953125,
0.462646484375,
0.708984375,
0.14697265625,
0.708984375
],
[
0.51123046875,
0.70068359375,
0.791015625,
0.70068359375,
0.791015625,
0.71435546875,
0.51123046875,
0.71435546875
],
[
0.14697265625,
0.71044921875,
0.2822265625,
0.71044921875,
0.2822265625,
0.7236328125,
0.14697265625,
0.7236328125
],
[
0.53076171875,
0.71630859375,
0.88330078125,
0.71875,
0.88330078125,
0.732421875,
0.53076171875,
0.72998046875
],
[
0.51318359375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.51318359375,
0.74755859375
],
[
0.124267578125,
0.74755859375,
0.130126953125,
0.74755859375,
0.130126953125,
0.75439453125,
0.124267578125,
0.75439453125
],
[
0.1275634765625,
0.74462890625,
0.26123046875,
0.74462890625,
0.26123046875,
0.75830078125,
0.1275634765625,
0.75830078125
],
[
0.5146484375,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.1175537109375,
0.765625,
0.48779296875,
0.765625,
0.48779296875,
0.779296875,
0.1175537109375,
0.779296875
],
[
0.51318359375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.12005615234375,
0.78369140625,
0.48779296875,
0.78369140625,
0.48779296875,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.51318359375,
0.783203125,
0.88330078125,
0.783203125,
0.88330078125,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.118408203125,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.118408203125,
0.8125
],
[
0.51318359375,
0.7998046875,
0.568359375,
0.7998046875,
0.568359375,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.1175537109375,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.1175537109375,
0.82763671875
],
[
0.5322265625,
0.81396484375,
0.884765625,
0.81396484375,
0.884765625,
0.82763671875,
0.5322265625,
0.82763671875
],
[
0.118408203125,
0.830078125,
0.489501953125,
0.83056640625,
0.489501953125,
0.845703125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.490234375,
0.84716796875,
0.490234375,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.51123046875,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.8798828125,
0.490234375,
0.8798828125,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.5146484375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.1175537109375,
0.8955078125,
0.376220703125,
0.8955078125,
0.376220703125,
0.9091796875,
0.1175537109375,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"End-to-End Information Extraction without Token-Level Supervision",
"Rasmus Berg Palm",
"Dirk Hovy",
"DTU Compute",
"Computer Science Dpeartment",
"Technical University of Denmark",
"Universityof Copenhagen",
"rapal@dtu.dk",
"dirk.hovy@di.ku.dk",
"Florian Laws",
"Ole Winther",
"Tradeshift",
"DTU Compute",
"Landemaerket 10, 1119 Copenhagen",
"Technical University of Denmark",
"fla@tradeshift.com",
"olwi@dtu.dk",
"Abstract",
"One of the main bottlenecks is a shortage of",
"the data required to train state-of-the-art IE mod-",
"Most state-of-the-art information extrac-",
"els, which rely on sequence tagging (Finkel et al.",
"tion approaches rely on token-level labels",
"2005: Zhai et al., 2017). Such models require suf-",
"to find the areas of interest in text. Unfor-",
"10z [9]",
"ficient amounts of training data that is labeled at",
"tunately, these labels are time-consuming",
"the token-level, i.e., with one label for each word.",
"and costly to create, and consequently, not",
"The reason token-level labels are in short supply",
"available for many real-life IE tasks. To",
"is that they are not the intended output of human",
"make matters worse, token-level labels are",
"IE tasks. Creating token-level labels thus requires",
"usually not the desired output, but just an",
"an additional effort, essentially doubling the work",
"intermediary step. End-to-end (E2E) mod-",
"",
"required to process each item. This additional ef-",
"[T']",
"els, which take raw text as input and pro-",
"fort is expensive and infeasible for many produc-",
"duce the desired output directly, need not",
"tion systems: estimates put the average cost for",
"depend on token-level labels. We propose",
"a sentence at about 3 dollars, and about half an",
"an E2E model based on pointer networks,",
"hour annotatortime (Alonso et al., 20l6). Conse",
"which can be trained directly on pairs of",
"quently, state-of-the-art IE approaches, relying on",
"1AE10TAII",
"raw input and output text. We evaluate our",
"sequence taggers, cannot be applied to many real",
"model on the ATIS data set, MIT restau-",
"life IE tasks.",
"rant corpus and the MIT movie corpus and",
"What is readily available in abundance and at no",
"compare to neural baselines that do use",
"additional costs, is the raw, unstructured input and",
"token-level labels. We achieve competi-",
"machine-readable output to a human IE task. Con-",
"tive results, within a few percentage points",
"sider the transcription of receipts, checks, or busi-",
"of the baselines, showing the feasibility",
"ness documents.wheretheinputis an unstructured",
"of E2E information extraction without the",
"PDF and the output a row in a database (due date,",
"need for token-level labels. This opens up",
"payable amount, etc). Another example is flight",
"new possibilities, as for many tasks cur-",
"bookings, where the input is a natural language",
"rently addressed by human extractors,raw",
"request from the user, and the output a HTTP re-",
"input and output data are available, but not",
"quest, sent to the airline booking API",
"token-level labels.",
"To better exploit such existing data sources,",
"we propose an end-to-end (E2E) model based on",
"-",
"Introduction",
"pointer networks with attention, which can be",
"Humans spend countless hours extracting struc-",
"trained end-to-end on the input/output pairs of hu-",
"turedmachinereadableinformationfromunstruc-",
"man IE tasks, without requiring token-level anno-",
"tured information in a multitude of domains.",
"tations.",
"Promising to automate this, information extraction",
"We evaluate our model on three traditional IE",
"(IE) is one of the most sought-after industrial ap-",
"data sets. Note that our model and the baselines",
"plications of natural language processing. How-",
"are competing in two dimensions. The first is cost",
"ever, despite substantial research efforts, in prac-",
"and applicability. The baselines require token-",
"tice, many applications still rely on manual effort",
"level labels, which are expensive and unavailable",
"to extract the relevant information.",
"for many real life tasks. Our model does not re-"
] |
{
"arxiv_id": "1707.04913",
"language": "en",
"page_idx": 0
}
|
||
1707.04913-en-1
|
1707.04913-en-1.jpg
|
<FigureHere>
Figure 1: Our model based on pointer networks. The solid red lines are the attention weights. For clarity
only two decoders are drawn and only the strongest attention weight for each output is drawn.
to extract. At each step $j$ each decoder calculate an unnormalized scalar attention score $a_{kji}$ over each input position $i.$ The $k’$th decoder output at step $j$, $o_{kj}$, is then the weighted sum of inputs, weighted with the normalized attention scores $att_kji.$
quire such token-level labels. Given the time and money required for these annotations, our model clearly improves over the baselines in this dimension. The second dimension is the accuracy of the models. Here we show that our model is competitive with the baseline models on two of the data sets and only slightly worse on the last data set, all despite fewer available annotations.
(2)
(3)
(4)
$$\begin{aligned}d_{kj}&=\mathrm{Decoder}_{k}(o_{k,j-1},d_{k,j-1})\\a_{kji}&=\mathrm{Attention}_{k}(d_{kj},e_{i})\mathrm{~for~}i=1...N\\att_{kji}&=\mathrm{softmax}(a_{kji})\mathrm{~for~}i=1...N\\o_{kj}&=\sum_{i=1}^{N}att_{kji}\:x_{i}\:.\end{aligned}$$
(5)
Contributions We present an E2E IE model with attention that does not depend on costly token-level labels, yet performs competitively with neural baseline models that rely on tokenlevel labels. By saving both time and money at comparable performance, our model presents a viable alternative for many real-life IE needs. Code is available at github.com/rasmusbergpalm/e2e-ierelease
Since each $x_i$ is a one-hot encoded word, and the $att_{kji}$ sum to one over $i, o_{kj}$ is a probability dis- tribution over words.
The loss function is the sum of the negative cross entropy for each of the expected outputs $y_{kj}$ and decoder outputs $o_{kj}.$
## 2 $\mathbf{Model}$
$$\mathcal{L}(\mathbf{x},\mathbf{y})=-\sum\limits_{k=1}^{K}\frac{1}{M_{k}}\sum\limits_{j=1}^{M_{k}}y_{kj}\log\left(o_{kj}\right)\:,$$
(6)
Our proposed model is based on pointer networks (Vinyals et al., 2015). A pointer network is a sequence-to-sequence model with attention in which the output is a position in the input sequence. The input position is "pointed to" using the attention mechanism. See figure 1 for an overview. Our formulation of the pointer network is slightly different from the original: Our output is some content from the input rather than a position in the input.
An input sequence of $N$ words $\mathbf{x}=x_1,...,x_N$ is encoded into another sequence of length $N$ using an Encoder.
where $M_k$ is the sequence length of expected out-
put $y_k.$
The specific architecture depends on the choice of Encoder, Decoder and Attention. For the encoder, we use a Bi-LSTM with 128 hidden units and a word embedding of 96 dimensions. We use a separate decoder for each of the fields. Each decoder has a word embedding of 96 dimensions, a LSTM with 128 units, with a learned first hidden state and its own attention mechanism. Our attention mechanism follows Bahdanau et al. (2014)
(1)
$$e_i=\text{Encoder}(x_i,e_{i-1})$$
We use a single shared encoder, and $k=1..K$ decoders, one for each piece of information we wish
$$a_{ji}=v^T\tanh(W_e\:enc_i+W_d\:dec_j)\:.\quad(7)$$
|
[
[
0.181396484375,
0.208984375,
0.2149658203125,
0.208984375,
0.2149658203125,
0.21728515625,
0.181396484375,
0.21728515625
],
[
0.490234375,
0.208984375,
0.52490234375,
0.208984375,
0.52490234375,
0.2161865234375,
0.490234375,
0.2161865234375
],
[
0.591796875,
0.208984375,
0.62060546875,
0.208984375,
0.62060546875,
0.2161865234375,
0.591796875,
0.2161865234375
],
[
0.44921875,
0.210205078125,
0.465087890625,
0.210205078125,
0.465087890625,
0.21728515625,
0.44921875,
0.21728515625
],
[
0.1275634765625,
0.210205078125,
0.16796875,
0.210205078125,
0.16796875,
0.21728515625,
0.1275634765625,
0.21728515625
],
[
0.2418212890625,
0.21142578125,
0.26123046875,
0.21142578125,
0.26123046875,
0.215576171875,
0.2418212890625,
0.215576171875
],
[
0.2880859375,
0.21142578125,
0.30908203125,
0.21142578125,
0.30908203125,
0.215576171875,
0.2880859375,
0.215576171875
],
[
0.403076171875,
0.21142578125,
0.41064453125,
0.21142578125,
0.41064453125,
0.21435546875,
0.403076171875,
0.21435546875
],
[
0.6533203125,
0.21142578125,
0.66015625,
0.21142578125,
0.66015625,
0.21435546875,
0.6533203125,
0.21435546875
],
[
0.7001953125,
0.21142578125,
0.7138671875,
0.21142578125,
0.7138671875,
0.215576171875,
0.7001953125,
0.215576171875
],
[
0.79248046875,
0.21142578125,
0.8154296875,
0.21142578125,
0.8154296875,
0.215576171875,
0.79248046875,
0.215576171875
],
[
0.84228515625,
0.21142578125,
0.8671875,
0.21142578125,
0.8671875,
0.215576171875,
0.84228515625,
0.215576171875
],
[
0.313232421875,
0.223876953125,
0.32666015625,
0.223876953125,
0.32666015625,
0.226806640625,
0.313232421875,
0.226806640625
],
[
0.63623046875,
0.22265625,
0.6708984375,
0.22265625,
0.6708984375,
0.226806640625,
0.63623046875,
0.226806640625
],
[
0.8154296875,
0.223876953125,
0.8271484375,
0.223876953125,
0.8271484375,
0.226806640625,
0.8154296875,
0.226806640625
],
[
0.118408203125,
0.26416015625,
0.8818359375,
0.26416015625,
0.8818359375,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.12005615234375,
0.28076171875,
0.8095703125,
0.28076171875,
0.8095703125,
0.29443359375,
0.12005615234375,
0.29443359375
],
[
0.118408203125,
0.319580078125,
0.489501953125,
0.318359375,
0.489501953125,
0.333251953125,
0.118408203125,
0.334228515625
],
[
0.51123046875,
0.318359375,
0.88330078125,
0.319580078125,
0.88330078125,
0.334228515625,
0.51123046875,
0.333251953125
],
[
0.12005615234375,
0.336181640625,
0.490234375,
0.336181640625,
0.490234375,
0.349853515625,
0.12005615234375,
0.349853515625
],
[
0.51318359375,
0.3349609375,
0.88330078125,
0.336181640625,
0.88330078125,
0.350830078125,
0.51318359375,
0.349853515625
],
[
0.118408203125,
0.3515625,
0.489501953125,
0.3515625,
0.489501953125,
0.365234375,
0.118408203125,
0.365234375
],
[
0.5146484375,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.366455078125,
0.5146484375,
0.366455078125
],
[
0.12005615234375,
0.3681640625,
0.489501953125,
0.3681640625,
0.489501953125,
0.3818359375,
0.12005615234375,
0.3818359375
],
[
0.51318359375,
0.3681640625,
0.88330078125,
0.3681640625,
0.88330078125,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.12005615234375,
0.38427734375,
0.48779296875,
0.38427734375,
0.48779296875,
0.39794921875,
0.12005615234375,
0.39794921875
],
[
0.51123046875,
0.38134765625,
0.83447265625,
0.383056640625,
0.83447265625,
0.40087890625,
0.51123046875,
0.3984375
],
[
0.118408203125,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.118408203125,
0.413330078125
],
[
0.5498046875,
0.412109375,
0.79248046875,
0.413330078125,
0.79248046875,
0.43115234375,
0.5498046875,
0.429931640625
],
[
0.8564453125,
0.41455078125,
0.88671875,
0.41455078125,
0.88671875,
0.43115234375,
0.8564453125,
0.43115234375
],
[
0.5458984375,
0.432861328125,
0.84228515625,
0.43115234375,
0.84228515625,
0.448974609375,
0.5458984375,
0.45068359375
],
[
0.12005615234375,
0.416259765625,
0.489501953125,
0.416259765625,
0.489501953125,
0.429931640625,
0.12005615234375,
0.429931640625
],
[
0.12176513671875,
0.43408203125,
0.382080078125,
0.43408203125,
0.382080078125,
0.44482421875,
0.12176513671875,
0.44482421875
],
[
0.8564453125,
0.435302734375,
0.88330078125,
0.435302734375,
0.88330078125,
0.448974609375,
0.8564453125,
0.448974609375
],
[
0.5322265625,
0.453125,
0.80615234375,
0.45068359375,
0.80615234375,
0.468505859375,
0.5322265625,
0.47021484375
],
[
0.8564453125,
0.453125,
0.88671875,
0.453125,
0.88671875,
0.4697265625,
0.8564453125,
0.4697265625
],
[
0.12005615234375,
0.47021484375,
0.48779296875,
0.47021484375,
0.48779296875,
0.48388671875,
0.12005615234375,
0.48388671875
],
[
0.12005615234375,
0.45361328125,
0.489501953125,
0.45361328125,
0.489501953125,
0.46728515625,
0.12005615234375,
0.46728515625
],
[
0.609375,
0.474365234375,
0.62646484375,
0.474365234375,
0.62646484375,
0.482666015625,
0.609375,
0.482666015625
],
[
0.1175537109375,
0.485107421875,
0.48779296875,
0.486328125,
0.48779296875,
0.50048828125,
0.1175537109375,
0.49951171875
],
[
0.54833984375,
0.48388671875,
0.7080078125,
0.482177734375,
0.7080078125,
0.50439453125,
0.54833984375,
0.505859375
],
[
0.8564453125,
0.486328125,
0.88671875,
0.486328125,
0.88671875,
0.5029296875,
0.8564453125,
0.5029296875
],
[
0.12005615234375,
0.501953125,
0.486083984375,
0.501953125,
0.486083984375,
0.515625,
0.12005615234375,
0.515625
],
[
0.60546875,
0.50732421875,
0.62890625,
0.50732421875,
0.62890625,
0.515625,
0.60546875,
0.515625
],
[
0.118408203125,
0.5185546875,
0.48779296875,
0.5185546875,
0.48779296875,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.5146484375,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.5146484375,
0.5419921875
],
[
0.12005615234375,
0.53515625,
0.48779296875,
0.53515625,
0.48779296875,
0.548828125,
0.12005615234375,
0.548828125
],
[
0.12005615234375,
0.55029296875,
0.489501953125,
0.55029296875,
0.489501953125,
0.56396484375,
0.12005615234375,
0.56396484375
],
[
0.5146484375,
0.5458984375,
0.8818359375,
0.5458984375,
0.8818359375,
0.56005859375,
0.5146484375,
0.56005859375
],
[
0.5146484375,
0.56103515625,
0.6689453125,
0.56103515625,
0.6689453125,
0.57470703125,
0.5146484375,
0.57470703125
],
[
0.118408203125,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.1175537109375,
0.58056640625,
0.175537109375,
0.5830078125,
0.173828125,
0.59716796875,
0.1158447265625,
0.595703125
],
[
0.53076171875,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.53076171875,
0.59033203125
],
[
0.51123046875,
0.5927734375,
0.88330078125,
0.595703125,
0.88330078125,
0.60986328125,
0.51123046875,
0.60693359375
],
[
0.1175537109375,
0.60400390625,
0.208251953125,
0.60400390625,
0.208251953125,
0.6181640625,
0.1175537109375,
0.6181640625
],
[
0.51123046875,
0.609375,
0.6962890625,
0.61083984375,
0.6962890625,
0.6259765625,
0.51123046875,
0.62353515625
],
[
0.1175537109375,
0.6259765625,
0.486083984375,
0.626953125,
0.486083984375,
0.64111328125,
0.1175537109375,
0.64013671875
],
[
0.6455078125,
0.64013671875,
0.66748046875,
0.642578125,
0.66650390625,
0.65185546875,
0.64404296875,
0.6494140625
],
[
0.7001953125,
0.64013671875,
0.7255859375,
0.64013671875,
0.7255859375,
0.65283203125,
0.7001953125,
0.65283203125
],
[
0.6630859375,
0.64453125,
0.6943359375,
0.64453125,
0.6943359375,
0.65673828125,
0.6630859375,
0.65673828125
],
[
0.12005615234375,
0.642578125,
0.48779296875,
0.642578125,
0.48779296875,
0.65625,
0.12005615234375,
0.65625
],
[
0.5380859375,
0.65185546875,
0.603515625,
0.65478515625,
0.60205078125,
0.67041015625,
0.53662109375,
0.66748046875
],
[
0.6943359375,
0.646484375,
0.7021484375,
0.646484375,
0.7021484375,
0.65185546875,
0.6943359375,
0.65185546875
],
[
0.7001953125,
0.65380859375,
0.8447265625,
0.65185546875,
0.8447265625,
0.66943359375,
0.7001953125,
0.67138671875
],
[
0.85791015625,
0.65478515625,
0.8818359375,
0.65478515625,
0.8818359375,
0.66748046875,
0.85791015625,
0.66748046875
],
[
0.64208984375,
0.65673828125,
0.6513671875,
0.65478515625,
0.654296875,
0.66015625,
0.6455078125,
0.6630859375
],
[
0.118408203125,
0.6591796875,
0.489501953125,
0.6591796875,
0.489501953125,
0.6728515625,
0.118408203125,
0.6728515625
],
[
0.66162109375,
0.66455078125,
0.7001953125,
0.66455078125,
0.7001953125,
0.6767578125,
0.66162109375,
0.6767578125
],
[
0.640625,
0.6728515625,
0.66748046875,
0.6728515625,
0.66748046875,
0.68115234375,
0.640625,
0.68115234375
],
[
0.69873046875,
0.6728515625,
0.72900390625,
0.6728515625,
0.72900390625,
0.68115234375,
0.69873046875,
0.68115234375
],
[
0.1175537109375,
0.69140625,
0.48779296875,
0.6904296875,
0.48779296875,
0.705078125,
0.1175537109375,
0.7060546875
],
[
0.1175537109375,
0.67333984375,
0.48779296875,
0.6748046875,
0.48779296875,
0.689453125,
0.1175537109375,
0.6884765625
],
[
0.5146484375,
0.69873046875,
0.8818359375,
0.69873046875,
0.8818359375,
0.7119140625,
0.5146484375,
0.7119140625
],
[
0.118408203125,
0.7080078125,
0.489501953125,
0.7080078125,
0.489501953125,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.5146484375,
0.71630859375,
0.5634765625,
0.71630859375,
0.5634765625,
0.72802734375,
0.5146484375,
0.72802734375
],
[
0.12005615234375,
0.7236328125,
0.489501953125,
0.7236328125,
0.489501953125,
0.73681640625,
0.12005615234375,
0.73681640625
],
[
0.53076171875,
0.72998046875,
0.88330078125,
0.728515625,
0.88330078125,
0.74365234375,
0.53076171875,
0.74462890625
],
[
0.118408203125,
0.74072265625,
0.489501953125,
0.74072265625,
0.489501953125,
0.75439453125,
0.118408203125,
0.75439453125
],
[
0.51318359375,
0.74658203125,
0.8818359375,
0.74658203125,
0.8818359375,
0.76025390625,
0.51318359375,
0.76025390625
],
[
0.1175537109375,
0.7548828125,
0.48779296875,
0.755859375,
0.48779296875,
0.77099609375,
0.1175537109375,
0.76953125
],
[
0.51123046875,
0.76123046875,
0.88330078125,
0.76123046875,
0.88330078125,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.118408203125,
0.7724609375,
0.2418212890625,
0.7724609375,
0.2418212890625,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.51318359375,
0.779296875,
0.884765625,
0.779296875,
0.884765625,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.1351318359375,
0.78759765625,
0.489501953125,
0.7880859375,
0.489501953125,
0.802734375,
0.1351318359375,
0.8017578125
],
[
0.51123046875,
0.7958984375,
0.8818359375,
0.79443359375,
0.8818359375,
0.80810546875,
0.51123046875,
0.80859375
],
[
0.1175537109375,
0.802734375,
0.48779296875,
0.80419921875,
0.48779296875,
0.818359375,
0.1175537109375,
0.81787109375
],
[
0.51318359375,
0.81103515625,
0.88330078125,
0.81103515625,
0.88330078125,
0.82470703125,
0.51318359375,
0.82470703125
],
[
0.118408203125,
0.82080078125,
0.2376708984375,
0.82080078125,
0.2376708984375,
0.83447265625,
0.118408203125,
0.83447265625
],
[
0.51318359375,
0.82666015625,
0.884765625,
0.82666015625,
0.884765625,
0.84033203125,
0.51318359375,
0.84033203125
],
[
0.5146484375,
0.84326171875,
0.8818359375,
0.84326171875,
0.8818359375,
0.85693359375,
0.5146484375,
0.85693359375
],
[
0.2149658203125,
0.849609375,
0.3896484375,
0.849609375,
0.3896484375,
0.8642578125,
0.2149658203125,
0.8642578125
],
[
0.46337890625,
0.85107421875,
0.490234375,
0.85107421875,
0.490234375,
0.86328125,
0.46337890625,
0.86328125
],
[
0.5146484375,
0.859375,
0.86572265625,
0.859375,
0.86572265625,
0.87255859375,
0.5146484375,
0.87255859375
],
[
0.12005615234375,
0.87890625,
0.48779296875,
0.87890625,
0.48779296875,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.55908203125,
0.88818359375,
0.84619140625,
0.88818359375,
0.84619140625,
0.90625,
0.55908203125,
0.90625
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.8486328125,
0.8955078125,
0.8564453125,
0.8955078125,
0.8564453125,
0.90087890625,
0.8486328125,
0.90087890625
],
[
0.85400390625,
0.89111328125,
0.8818359375,
0.89111328125,
0.8818359375,
0.90380859375,
0.85400390625,
0.90380859375
]
] |
[
"airfare",
"<EOS",
"<GO>",
"loui",
"cheapest",
"from",
"acor",
"",
"4",
"oU",
"",
"acom",
"n",
"Dacnde",
"aA",
"Figure I: Our model based on pointer networks. The solid red lines are the attention weights. For clarity",
"only two decoders are drawn and only the strongest attention weight for each output is drawn.",
"quire such token-level labels. Given the time and",
"to extract. At each step j each decoder calculate an",
"money required for these annotations, our model",
"unnormalized scalar attention score akji over each",
"clearly improves over the baselines in this dimen-",
"input position i. The k'th decoder output at step j.",
"sion. The second dimension is the accuracy of the",
"Oki, is then the weighted sum of inputs, weighted",
"models. Here we show that our model is compet-",
"with the normalized attention scores attkji.",
"itive with the baseline models on two of the data",
"dkj = Decoderk(ok,j-1, dk,j-1)",
"(2)",
"akji = Attentiont(dkj,ei) for i = 1..N",
"sets and only slightly worse on the last data set, all",
"despitefeweravailableannotations.",
"(3)",
"attkji = softmax(akji) for i = 1..N",
"(4)",
"with attention that does not depend on costly",
"Contributions We present an E2E IE model",
"N",
"token-level labels, yet performs competitively",
"Oki =atthiji ai.",
"(5)",
"with neural baseline models that rely on token-",
"i=1",
"level labels. By saving both time and money at",
"Since each c, is a one-hot encoded word, and the",
"comparable performance, our model presents a vi-",
"able alternative for many real-life IE needs. Code",
"attkji sum to one over i, Okj is a probability dis-",
"tribution over words.",
"is available at github.com/rasmusbergpalm/e2e-ie-",
"release",
"The loss function is the sum of the negative",
"cross entropy for each of the expected outputs ykj",
"2 Model",
"and decoder outputs Okj.",
"Our proposed model is based on pointer net-",
"K",
"Mi",
"1",
"works (Vinyals et al., 2015). A pointer network",
"L(x,y)",
"",
"Zykj log(okj),",
"(6)",
"1",
"is a sequence-to-sequence model with attention",
"Mk",
"k=1",
"=1",
"quence. The input position is\"pointed to\" us-",
"in which the output is a position in the input se-",
"where M is the sequence length of expected out-",
"ing the attention mechanism. See figure 1 for an",
"put Yk",
"overview. Our formulation of the pointer network",
"The specific architecture depends on the choice",
"is slightly different from the original: Our output",
"of Encoder, Decoder and Attention. For the en-",
"is some content from the input rather than a posi-",
"coder, we use a Bi-LSTM with 128 hidden units",
"tion in the input.",
"and a word embedding of 96 dimensions.We use",
"An input sequence of N words x = &1, ..., &N",
"a separate decoder for each of the fields. Each de-",
"is encoded into another sequence of length N us-",
"coder has a word embedding of 96 dimensions, a",
"ing an Encoder.",
"LSTM with 128 units, with a learned first hidden",
"state and its own attention mechanism. Our atten-",
"ei =Encoder(ci,ei-1)",
"(1)",
"tion mechanism follows Bahdanau et al. (2014)",
"We use a single shared encoder, and k = 1..K de-",
"aji = vT tanh(We enci + Wa dec;) .",
"coders, one for each piece of information we wish",
"",
"(7)"
] |
{
"arxiv_id": "1707.04913",
"language": "en",
"page_idx": 1
}
|
||
1707.04913-en-2
|
1707.04913-en-2.jpg
|
The attention parameters $W_e,W_d$ and $v$ for each
attention mechanism are all 128-dimensional
During training we use teacher forcing for the decoders (Williams and Zipser, 1989), such that $o_{k,j-1}=y_{k,j-1}.$ During testing we use argmax to select the most probable output for each step $j$ and run each decoder until the first end of sentence (EOS) symbol.
## 3 Experiments
## 3.1 Data sets
To compare our model to baselines relying on token-level labels we use existing data sets for which token level-labels are available. We measure our performance on the ATIS data set (Price, 1990) (4978 training samples, 893 testing samples) and the MIT restaurant (7660 train, 1521 test) and movie corpus (9775 train, 2443 test) (Liu et al., 2013). These data sets contains token-level labels in the Beginning-Inside-Out format (BIO).
The ATIS data set consists of natural language requests to a simulated airline booking system. Each word is labeled with one of several classes, e.g. departure city, arrival city, cost, etc. The MIT restaurant and movie corpus are similar, except for a restaurant and movie domain respectively. See table 1 for samples.
MIT Restaurant
MITMovie
2
B-Rating show
start
I-Rating me
0
restaurants
0 films
with
0 elvis
B-ACTOR
0
inside
B-Amenity flms
B-PLOT For comparison, the published state-of-the-art is
dining
$I-$Amenity set
in
I- PLOT
hawaii
I- PLOT
Table 1: Samples from the MIT restaurant and movie corpus. The transcription errors are part of the data.
Since our model does not need token-level labels, we create an E2E version of each data set without token-level labels by chunking the BIOlabeled words and using the labels as fields to extract. If there are multiple outputs for a single feld, e.g. multiple destination cities, we join them with a comma. For the ATIS data set, we choose the 10 most common labels, and we use all the labels for the movie and restaurant corpus. The movie data set has 12 fields and the restaurant has
8. See Table 2 for an example of the E2E ATIS
data set.
Input
cheapest airfare from tacoma to st. louis and detroit
$0utput$
fromloc
tacoma
st. louis , detroit
toloc
airline_name
$cost\_relative cheapest$
period_of_-day
time
time_relative
day_-name
day-number month_name
Table 2: Sample from the E2E ATIS data set.
## 3.2 Baselines
For the baselines, we use a two layer neural network model. The first layer is a Bi-directional Long Short Term Memory network (Hochreiter and Schmidhuber, 1997) (Bi-LSTM) and the second layer is a forward-only LSTM. Both layers have 128 hidden units. We use a trained word embedding of size 128. The baseline is trained with
$\textcircled{\mathrm{Adam}}\textcircled{\mathrm{Kingma}}$ and Ba, 2014) on the BIO labels $\textcircled{\mathrm{and}}$ uses early stopping on a held out validation
set.
This baseline architecture achieves a fairly strong F1 score of 0.9456 on the ATIS data set. at 0.9586 (Zhai et al., 2017). These numbers are for the traditional BIO token level measure of performance using the publicly available conlleval script. They should not be confused with the E2E performance reported later. We present them here so that readers familiar with the ATIS data set can evaluate the strength of our baselines using a wellknown measure.
For the E2E performance measure we train the baseline models using token-level BIO labels and predict BIO labels on the test set. Given the predicted BIO labels, we create the E2E output for the baseline models in the same way we created the E2E data sets, i.e. by chunking and extracting labels as fields. We evaluate our model and the baselines using the MUC-5 definitions of precision, recall and F1, without partial matches (Chinchor and
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.1175537109375,
0.0926513671875,
0.453369140625,
0.0914306640625,
0.453369140625,
0.105712890625,
0.1175537109375,
0.10687255859375
],
[
0.51318359375,
0.09442138671875,
0.57666015625,
0.09442138671875,
0.57666015625,
0.105712890625,
0.51318359375,
0.105712890625
],
[
0.1368408203125,
0.1092529296875,
0.489501953125,
0.1092529296875,
0.489501953125,
0.122314453125,
0.1368408203125,
0.122314453125
],
[
0.68994140625,
0.1175537109375,
0.73876953125,
0.119384765625,
0.73779296875,
0.1353759765625,
0.6884765625,
0.133056640625
],
[
0.118408203125,
0.12469482421875,
0.489501953125,
0.12469482421875,
0.489501953125,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.52294921875,
0.1341552734375,
0.904296875,
0.1341552734375,
0.904296875,
0.1478271484375,
0.52294921875,
0.1478271484375
],
[
0.1158447265625,
0.141357421875,
0.490234375,
0.1383056640625,
0.490234375,
0.155029296875,
0.1158447265625,
0.157958984375
],
[
0.1175537109375,
0.1561279296875,
0.489501953125,
0.15673828125,
0.489501953125,
0.171630859375,
0.1175537109375,
0.17041015625
],
[
0.118408203125,
0.171630859375,
0.48779296875,
0.171630859375,
0.48779296875,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.6826171875,
0.16748046875,
0.74658203125,
0.16748046875,
0.74658203125,
0.1817626953125,
0.6826171875,
0.1817626953125
],
[
0.52294921875,
0.18408203125,
0.60546875,
0.18408203125,
0.60546875,
0.198974609375,
0.52294921875,
0.198974609375
],
[
0.67919921875,
0.18408203125,
0.74072265625,
0.18408203125,
0.74072265625,
0.198974609375,
0.67919921875,
0.198974609375
],
[
0.68115234375,
0.1995849609375,
0.8076171875,
0.1995849609375,
0.8076171875,
0.21435546875,
0.68115234375,
0.21435546875
],
[
0.118408203125,
0.189453125,
0.231689453125,
0.189453125,
0.231689453125,
0.2037353515625,
0.118408203125,
0.2037353515625
],
[
0.52294921875,
0.201904296875,
0.5810546875,
0.201904296875,
0.5810546875,
0.213134765625,
0.52294921875,
0.213134765625
],
[
0.1158447265625,
0.2078857421875,
0.2626953125,
0.210205078125,
0.2626953125,
0.22802734375,
0.1158447265625,
0.2257080078125
],
[
0.52490234375,
0.218505859375,
0.654296875,
0.218505859375,
0.654296875,
0.2286376953125,
0.52490234375,
0.2286376953125
],
[
0.52294921875,
0.2327880859375,
0.6630859375,
0.2327880859375,
0.6630859375,
0.2464599609375,
0.52294921875,
0.2464599609375
],
[
0.68115234375,
0.232177734375,
0.75146484375,
0.232177734375,
0.75146484375,
0.2464599609375,
0.68115234375,
0.2464599609375
],
[
0.52294921875,
0.248779296875,
0.66162109375,
0.248779296875,
0.66162109375,
0.26318359375,
0.52294921875,
0.26318359375
],
[
0.1175537109375,
0.2340087890625,
0.235107421875,
0.2340087890625,
0.235107421875,
0.248779296875,
0.1175537109375,
0.248779296875
],
[
0.1175537109375,
0.257080078125,
0.489501953125,
0.257080078125,
0.489501953125,
0.271484375,
0.1175537109375,
0.271484375
],
[
0.52294921875,
0.26416015625,
0.57275390625,
0.26416015625,
0.57275390625,
0.279052734375,
0.52294921875,
0.279052734375
],
[
0.118408203125,
0.273681640625,
0.489501953125,
0.273681640625,
0.489501953125,
0.287353515625,
0.118408203125,
0.287353515625
],
[
0.52294921875,
0.28076171875,
0.66650390625,
0.28076171875,
0.66650390625,
0.29443359375,
0.52294921875,
0.29443359375
],
[
0.118408203125,
0.2880859375,
0.486083984375,
0.2880859375,
0.486083984375,
0.3017578125,
0.118408203125,
0.3017578125
],
[
0.52490234375,
0.298583984375,
0.609375,
0.298583984375,
0.609375,
0.31005859375,
0.52490234375,
0.31005859375
],
[
0.118408203125,
0.3046875,
0.489501953125,
0.3046875,
0.489501953125,
0.319580078125,
0.118408203125,
0.319580078125
],
[
0.52490234375,
0.3154296875,
0.63037109375,
0.3154296875,
0.63037109375,
0.32666015625,
0.52490234375,
0.32666015625
],
[
0.12005615234375,
0.32177734375,
0.489501953125,
0.32177734375,
0.489501953125,
0.3349609375,
0.12005615234375,
0.3349609375
],
[
0.1175537109375,
0.336181640625,
0.486083984375,
0.336181640625,
0.486083984375,
0.349853515625,
0.1175537109375,
0.349853515625
],
[
0.52294921875,
0.330810546875,
0.63037109375,
0.330810546875,
0.63037109375,
0.342529296875,
0.52294921875,
0.342529296875
],
[
0.1175537109375,
0.352783203125,
0.48779296875,
0.3515625,
0.48779296875,
0.366455078125,
0.1175537109375,
0.36767578125
],
[
0.52880859375,
0.3623046875,
0.8623046875,
0.3623046875,
0.8623046875,
0.377197265625,
0.52880859375,
0.377197265625
],
[
0.12005615234375,
0.369384765625,
0.489501953125,
0.369384765625,
0.489501953125,
0.383056640625,
0.12005615234375,
0.383056640625
],
[
0.118408203125,
0.385986328125,
0.478515625,
0.385986328125,
0.478515625,
0.399658203125,
0.118408203125,
0.399658203125
],
[
0.1368408203125,
0.40087890625,
0.48779296875,
0.402099609375,
0.48779296875,
0.416259765625,
0.1368408203125,
0.4150390625
],
[
0.51318359375,
0.40380859375,
0.63232421875,
0.40380859375,
0.63232421875,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.118408203125,
0.418701171875,
0.484375,
0.418701171875,
0.484375,
0.431640625,
0.118408203125,
0.431640625
],
[
0.51123046875,
0.42578125,
0.880859375,
0.42578125,
0.880859375,
0.439453125,
0.51123046875,
0.439453125
],
[
0.1175537109375,
0.43408203125,
0.486083984375,
0.43408203125,
0.486083984375,
0.44775390625,
0.1175537109375,
0.44775390625
],
[
0.51318359375,
0.441162109375,
0.88330078125,
0.4423828125,
0.88330078125,
0.457275390625,
0.51318359375,
0.4560546875
],
[
0.118408203125,
0.449462890625,
0.48779296875,
0.448974609375,
0.48779296875,
0.463134765625,
0.118408203125,
0.46435546875
],
[
0.51318359375,
0.458984375,
0.88330078125,
0.458984375,
0.88330078125,
0.473876953125,
0.51318359375,
0.473876953125
],
[
0.1175537109375,
0.466064453125,
0.490234375,
0.466064453125,
0.490234375,
0.479736328125,
0.1175537109375,
0.479736328125
],
[
0.51318359375,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.118408203125,
0.482666015625,
0.489501953125,
0.482666015625,
0.489501953125,
0.496337890625,
0.118408203125,
0.496337890625
],
[
0.5146484375,
0.4912109375,
0.884765625,
0.4912109375,
0.884765625,
0.5048828125,
0.5146484375,
0.5048828125
],
[
0.1175537109375,
0.49755859375,
0.26611328125,
0.498779296875,
0.26611328125,
0.51318359375,
0.1175537109375,
0.5126953125
],
[
0.51318359375,
0.50732421875,
0.88330078125,
0.50732421875,
0.88330078125,
0.52099609375,
0.51318359375,
0.52099609375
],
[
0.162841796875,
0.52685546875,
0.295654296875,
0.52685546875,
0.295654296875,
0.54052734375,
0.162841796875,
0.54052734375
],
[
0.376220703125,
0.52783203125,
0.466796875,
0.52783203125,
0.466796875,
0.5390625,
0.376220703125,
0.5390625
],
[
0.5146484375,
0.52490234375,
0.88330078125,
0.52490234375,
0.88330078125,
0.53515625,
0.5146484375,
0.53515625
],
[
0.2376708984375,
0.54150390625,
0.333251953125,
0.54345703125,
0.333251953125,
0.55810546875,
0.2376708984375,
0.5556640625
],
[
0.5146484375,
0.5390625,
0.884765625,
0.5390625,
0.884765625,
0.552734375,
0.5146484375,
0.552734375
],
[
0.343505859375,
0.54345703125,
0.391357421875,
0.54345703125,
0.391357421875,
0.55810546875,
0.343505859375,
0.55810546875
],
[
0.1275634765625,
0.54345703125,
0.14111328125,
0.54345703125,
0.14111328125,
0.5556640625,
0.1275634765625,
0.5556640625
],
[
0.483642578125,
0.5458984375,
0.495361328125,
0.5458984375,
0.495361328125,
0.5546875,
0.483642578125,
0.5546875
],
[
0.1275634765625,
0.56005859375,
0.16455078125,
0.56005859375,
0.16455078125,
0.572265625,
0.1275634765625,
0.572265625
],
[
0.345947265625,
0.56005859375,
0.37451171875,
0.56005859375,
0.37451171875,
0.572265625,
0.345947265625,
0.572265625
],
[
0.51318359375,
0.5556640625,
0.884765625,
0.5556640625,
0.884765625,
0.5693359375,
0.51318359375,
0.5693359375
],
[
0.48193359375,
0.56103515625,
0.496337890625,
0.56103515625,
0.496337890625,
0.57177734375,
0.48193359375,
0.57177734375
],
[
0.240966796875,
0.56103515625,
0.330078125,
0.56103515625,
0.330078125,
0.572265625,
0.240966796875,
0.572265625
],
[
0.322509765625,
0.57373046875,
0.3359375,
0.5830078125,
0.32421875,
0.59130859375,
0.310546875,
0.58203125
],
[
0.343505859375,
0.57470703125,
0.3896484375,
0.57470703125,
0.3896484375,
0.5888671875,
0.343505859375,
0.5888671875
],
[
0.51318359375,
0.572265625,
0.5439453125,
0.572265625,
0.5439453125,
0.58447265625,
0.51318359375,
0.58447265625
],
[
0.48193359375,
0.57666015625,
0.496337890625,
0.57666015625,
0.496337890625,
0.58740234375,
0.48193359375,
0.58740234375
],
[
0.1275634765625,
0.57666015625,
0.21240234375,
0.57666015625,
0.21240234375,
0.58837890625,
0.1275634765625,
0.58837890625
],
[
0.32421875,
0.5888671875,
0.3359375,
0.59716796875,
0.32421875,
0.60693359375,
0.312255859375,
0.5986328125
],
[
0.343505859375,
0.59130859375,
0.387939453125,
0.59130859375,
0.387939453125,
0.60546875,
0.343505859375,
0.60546875
],
[
0.53076171875,
0.58837890625,
0.8818359375,
0.58837890625,
0.8818359375,
0.6015625,
0.53076171875,
0.6015625
],
[
0.41650390625,
0.5927734375,
0.495361328125,
0.5927734375,
0.495361328125,
0.60400390625,
0.41650390625,
0.60400390625
],
[
0.1275634765625,
0.5927734375,
0.16455078125,
0.5927734375,
0.16455078125,
0.60498046875,
0.1275634765625,
0.60498046875
],
[
0.344970703125,
0.60498046875,
0.3896484375,
0.60693359375,
0.387939453125,
0.62255859375,
0.343505859375,
0.62060546875
],
[
0.51318359375,
0.60400390625,
0.88330078125,
0.60400390625,
0.88330078125,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.225830078125,
0.60546875,
0.33154296875,
0.60791015625,
0.33154296875,
0.62255859375,
0.225830078125,
0.62060546875
],
[
0.1259765625,
0.60693359375,
0.17626953125,
0.60693359375,
0.17626953125,
0.62158203125,
0.1259765625,
0.62158203125
],
[
0.480224609375,
0.60986328125,
0.491943359375,
0.60693359375,
0.497802734375,
0.61767578125,
0.486083984375,
0.62060546875
],
[
0.1259765625,
0.62158203125,
0.181396484375,
0.62353515625,
0.1796875,
0.6396484375,
0.124267578125,
0.63720703125
],
[
0.343505859375,
0.62451171875,
0.37451171875,
0.62451171875,
0.37451171875,
0.6396484375,
0.343505859375,
0.6396484375
],
[
0.426513671875,
0.62451171875,
0.496337890625,
0.62451171875,
0.496337890625,
0.6357421875,
0.426513671875,
0.6357421875
],
[
0.51318359375,
0.62060546875,
0.884765625,
0.62060546875,
0.884765625,
0.63427734375,
0.51318359375,
0.63427734375
],
[
0.51318359375,
0.6357421875,
0.884765625,
0.6357421875,
0.884765625,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.2283935546875,
0.6259765625,
0.330078125,
0.6259765625,
0.330078125,
0.63720703125,
0.2283935546875,
0.63720703125
],
[
0.344970703125,
0.6396484375,
0.366943359375,
0.6396484375,
0.366943359375,
0.65380859375,
0.344970703125,
0.65380859375
],
[
0.426513671875,
0.6396484375,
0.497802734375,
0.6396484375,
0.497802734375,
0.65380859375,
0.426513671875,
0.65380859375
],
[
0.344970703125,
0.65625,
0.4013671875,
0.65625,
0.4013671875,
0.67041015625,
0.344970703125,
0.67041015625
],
[
0.426513671875,
0.65673828125,
0.496337890625,
0.65673828125,
0.496337890625,
0.66845703125,
0.426513671875,
0.66845703125
],
[
0.51318359375,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.515625,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68115234375,
0.515625,
0.68115234375
],
[
0.118408203125,
0.689453125,
0.490234375,
0.689453125,
0.490234375,
0.703125,
0.118408203125,
0.703125
],
[
0.51318359375,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.1175537109375,
0.7060546875,
0.490234375,
0.705078125,
0.490234375,
0.7197265625,
0.1175537109375,
0.72021484375
],
[
0.51123046875,
0.70068359375,
0.884765625,
0.69970703125,
0.884765625,
0.71435546875,
0.51123046875,
0.71533203125
],
[
0.5146484375,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.12005615234375,
0.7236328125,
0.183837890625,
0.7236328125,
0.183837890625,
0.7353515625,
0.12005615234375,
0.7353515625
],
[
0.5146484375,
0.73388671875,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.1376953125,
0.751953125,
0.48779296875,
0.751953125,
0.48779296875,
0.7626953125,
0.1376953125,
0.7626953125
],
[
0.51318359375,
0.75,
0.63623046875,
0.75,
0.63623046875,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.1175537109375,
0.76416015625,
0.48779296875,
0.76416015625,
0.48779296875,
0.779296875,
0.1175537109375,
0.779296875
],
[
0.53076171875,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.53076171875,
0.77978515625
],
[
0.12005615234375,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.118408203125,
0.8125
],
[
0.51123046875,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51123046875,
0.8125
],
[
0.118408203125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.51123046875,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.1175537109375,
0.845703125,
0.490234375,
0.8466796875,
0.490234375,
0.86083984375,
0.1175537109375,
0.85986328125
],
[
0.51123046875,
0.845703125,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.118408203125,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51123046875,
0.8623046875,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51123046875,
0.8759765625
],
[
0.1175537109375,
0.87744140625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.51123046875,
0.87744140625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.51123046875,
0.892578125
],
[
0.118408203125,
0.89453125,
0.490234375,
0.89453125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"The attention parameters We, Wa and for each",
"8. See Table 2 for an example of the E2E ATIS",
"attention mechanism are all 128-dimensional.",
"data set.",
"During training we use teacher forcing for the",
"Input",
"decoders (Williams and Zipser, 1989), such that",
"cheapest airfare from tacoma to st. louis and detroit",
"Ok,j-1 = Yk,j-1. During testing we use argmax",
"to select the most probable output for each step j",
"and run each decoder until the first end of sentence",
"Output",
"fromloc",
"tacoma",
"st. louis , detroit",
"(EOS) symbol.",
"toloc",
"3Experiments",
"airline_name",
"cost.relative",
"cheapest",
"period.of.day",
"3.1 Data sets",
"To compare our model to baselines relying on",
"time",
"token-level labels we use existing data sets for",
"time.relative",
"which token level-labels are available. We mea-",
"day.name",
"sure our performance on the ATIS data set (Price,",
"day-number",
"1990) (4978 training samples, 893 testing sam-",
"ples) and the MIT restaurant (7660 train, 1521",
"monthname",
"test) and movie corpus (9775 train, 2443 test) (Liu",
"Table 2: Sample from the E2E ATIS data set.",
"et al., 2013). These data sets contains token-level",
"labels in the Beginning-Inside-Out format (BIO)",
"The ATIS data set consists of natural language",
"3.2Baselines",
"requests to a simulated airline booking system",
"For the baselines, we use a two layer neural net-",
"Each word is labeled with one of several classes",
"work model. The first layer is a Bi-directional",
"e.g. departure city, arrival city, cost, etc. The MIT",
"Long Short Term Memory network (Hochreiter",
"restaurant and movie corpus are similar, except for",
"and Schmidhuber, 1997) (Bi-LSTM) and the sec-",
"a restaurant and movie domain respectively. See",
"ond layer is a forward-only LSTM. Both layers",
"table1 for samples.",
"have 128 hidden units. We use a trained word em-",
"MIT Restaurant",
"MIT Movie",
"bedding of size 128.The baseline is trained with",
"B-Rating",
"Adam (Kingma and Ba, 2014) on the BIO labels",
"show",
"2",
"0",
"start",
"me",
"and uses early stopping on a held out validation",
"",
"I-Rating",
"O",
"films",
"set.",
"",
"restaurants",
"O",
"elvis",
"This baseline architecture achieves a fairly",
"B-ACTOR",
"with",
"films",
"strong F1 score of o0.9456 on the ATIS data set.",
"B-Amenity",
"inside",
"O",
"dining",
"set",
"B-PLOT",
"For comparison, the published state-of-the-art is",
"at 0.9586 (Zhai et al., 2017). These numbers are",
"I-Amenity",
"in",
"I-PLOT",
"hawaii",
"I-PLOT",
"for the traditional BIOtoken level measure of per-",
"formance using thepublicly available conlleval",
"Table 1: Samples from the MIT restaurant and",
"script. They should not be confused with the E2E",
"movie corpus. The transcription errors are part of",
"performance reported later. We present them here",
"so that readers familiar with the ATIS data set can",
"the data.",
"evaluate the strength of our baselines using a well-",
"Since ourmodeldoes notneedtoken-level la-",
"known measure.",
"bels, we create an E2E version of each data set",
"For the E2E performance measure we train the",
"without token-level labels by chunking the BIO-",
"baseline models using token-level BIO labels and",
"labeled words and using the labels as fields to ex-",
"predict BIO labels on the test set. Given the pre-",
"tract. If there are multiple outputs for a single",
"dicted BIOlabels, we create the E2E output for the",
"field, e.g. multiple destination cities, we join them",
"baseline models in the same way we created the",
"with a comma. For the ATIS data set, we choose",
"E2E data sets, i.e. by chunking and extracting la-",
"the 10 most common labels, and we use all the",
"bels as fields. We evaluate our model and the base-",
"labels for the movie and restaurant corpus. The",
"lines using the MUC-5 definitions of precision, re-",
"movie data set has 12 fields and the restaurant has",
"call and F1, without partial matches (Chinchor and"
] |
{
"arxiv_id": "1707.04913",
"language": "en",
"page_idx": 2
}
|
||
1707.04913-en-3
|
1707.04913-en-3.jpg
|
Sundheim, 1993). We use bootstrap sampling to estimate the probability that the model with the best micro average F1 score on the entire test set is worse for a randomly sampled subset of the test data.
## 3.3 0ur model
Since our decoders can only output values that are present in the input, we prepend a single comma to every input sequence. We optimize our model using Adam and use early stopping on a held-out validation set. The model quickly converges to optimal performance, usually after around 5000 updates after which it starts overfitting.
For the restaurant data set, to increase performance, we double the sizes of all the parameters and use embedding and recurrent dropout following (Gal, 2015). Further, we add a summarizer LSTM to each decoder. The summarizer LSTM reads the entire encoded input. The last hidden state of the summarizer LSTM is then concatenated to each input to the decoder.
## 3.4 Results
<table>
<tbody>
<tr>
<th>Data set</th>
<th>Baseline 1 1</th>
<th>Ours</th>
<th>$p$</th>
</tr>
<tr>
<td>$ATIS$</td>
<td>0.977</td>
<td>0.974</td>
<td>0.1755</td>
</tr>
<tr>
<td>Movie</td>
<td>0.816</td>
<td>0.817</td>
<td>0.3792</td>
</tr>
<tr>
<td>Restaurant</td>
<td>0.724</td>
<td>0.694</td>
<td>0.0001</td>
</tr>
</tbody>
</table>
Table 3: Micro average F1 scores on the E2E data sets. Results that are significantly better $( p$ < 0.05) are highlighted in bold.
We see in Table 3 that our model is competitive with the baseline models in terms of microaveraged F1 for two of the three data sets. This is a remarkable result given that the baselines are trained on token-level labels, whereas our model is trained end-to-end. For the restaurant data set, our model is slightly worse than the baseline.
## 4 Related work
Event extraction (EE) is similar to the E2E IE task we propose, except that it can have several event types and multiple events per input. In our E2E IE task, we only have a single event type and assume there is zero or one event mentioned in the input, which is an easier task. Recently, Nguyen et al. (2016) achieved state of the art results on the ACE 2005 EE data set using a recurrent neural network to jointly model event triggers and argument roles.
Other approaches have addressed the need for token-level labels when only raw output values are available. Mintz et al. (2009) introduced distant supervision, which heuristically generates the token-level labels from the output values. You do this by searching for input tokens that matches output values. The matching tokens are then assigned the labels for the matching outputs. One drawback is that the quality of the labels crucially depend on the search algorithm and how closely the tokens match the output values, which makes it brittle. Our method is trained end-to-end, thus not relying on brittle heuristics.
Sutskever et al. (2014) opened up the sequenceto-sequence paradigm. With the addition of attention (Bahdanau et al., 2014), these models achieved state-of-the-art results in machine translation (Wu et al., 2016). We are broadly inspired by these results to investigate E2E models for IE.
The idea of copying words from the input to the output have been used in machine translation to overcome problems with out-of-vocabulary words (Gulcehre et al., 2016; Gu et al., 2016).
## 5 Discussion
We present an end-to-end IE model that does not require detailed token-level labels. Despite being trained end-to-end, it is competitive with baseline models relying on token-level labels. In contrast to them, our model can be used on many real life IE tasks where intermediate token-level labels are not available and creating them is not feasible.
In our experiments our model and the baselines had access to the same amount of training samples. In a real life scenario it is likely that tokenlevel labels only exist for a subset of all the data. It would be interesting to investigate the quantity/quality trade-of of the labels, and a multi task extension to the model, which could make use of available token-level labels
Our model is remarkably stable to hyper parameter changes. On the restaurant dataset we tried several different architectures and hyper parameters before settling on the reported one. The difference between the worst and the best was approximately 2 percentage points.
A major limitation of the proposed model is that it can only output values that are present in the input. This is a problem for outputs that are normalized before being submitted as machine readable data, which is a common occurrence. For instance, dates might appear as'Jan 17 2012'in
|
[
[
0.118408203125,
0.0748291015625,
0.489501953125,
0.07598876953125,
0.489501953125,
0.09027099609375,
0.118408203125,
0.08905029296875
],
[
0.5322265625,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.5322265625,
0.09027099609375
],
[
0.1175537109375,
0.0926513671875,
0.489501953125,
0.0914306640625,
0.489501953125,
0.105712890625,
0.1175537109375,
0.10687255859375
],
[
0.51123046875,
0.0914306640625,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.51123046875,
0.105712890625
],
[
0.118408203125,
0.1092529296875,
0.490234375,
0.1092529296875,
0.490234375,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.51318359375,
0.10809326171875,
0.88330078125,
0.10809326171875,
0.88330078125,
0.1217041015625,
0.51318359375,
0.1217041015625
],
[
0.1175537109375,
0.12469482421875,
0.490234375,
0.12469482421875,
0.490234375,
0.1383056640625,
0.1175537109375,
0.1383056640625
],
[
0.51123046875,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1383056640625,
0.51123046875,
0.1383056640625
],
[
0.12005615234375,
0.141357421875,
0.156982421875,
0.141357421875,
0.156982421875,
0.15380859375,
0.12005615234375,
0.15380859375
],
[
0.515625,
0.142578125,
0.88330078125,
0.142578125,
0.88330078125,
0.152587890625,
0.515625,
0.152587890625
],
[
0.1175537109375,
0.1602783203125,
0.2476806640625,
0.160888671875,
0.2476806640625,
0.17578125,
0.1175537109375,
0.174560546875
],
[
0.5146484375,
0.15673828125,
0.88330078125,
0.15673828125,
0.88330078125,
0.17041015625,
0.5146484375,
0.17041015625
],
[
0.5146484375,
0.17333984375,
0.884765625,
0.17333984375,
0.884765625,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.118408203125,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.51318359375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.51318359375,
0.203125
],
[
0.12005615234375,
0.20068359375,
0.489501953125,
0.20068359375,
0.489501953125,
0.21435546875,
0.12005615234375,
0.21435546875
],
[
0.51123046875,
0.204833984375,
0.88330078125,
0.204833984375,
0.88330078125,
0.218505859375,
0.51123046875,
0.218505859375
],
[
0.12005615234375,
0.2161865234375,
0.489501953125,
0.2161865234375,
0.489501953125,
0.2298583984375,
0.12005615234375,
0.2298583984375
],
[
0.51123046875,
0.2203369140625,
0.884765625,
0.2197265625,
0.884765625,
0.2340087890625,
0.51123046875,
0.235107421875
],
[
0.12005615234375,
0.2327880859375,
0.489501953125,
0.2327880859375,
0.489501953125,
0.2464599609375,
0.12005615234375,
0.2464599609375
],
[
0.51123046875,
0.2369384765625,
0.880859375,
0.2369384765625,
0.880859375,
0.25048828125,
0.51123046875,
0.25048828125
],
[
0.118408203125,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.51123046875,
0.251708984375,
0.884765625,
0.254150390625,
0.884765625,
0.268310546875,
0.51123046875,
0.2666015625
],
[
0.1175537109375,
0.26416015625,
0.48779296875,
0.26318359375,
0.48779296875,
0.27783203125,
0.1175537109375,
0.279052734375
],
[
0.51318359375,
0.26953125,
0.6630859375,
0.26953125,
0.6630859375,
0.283203125,
0.51318359375,
0.283203125
],
[
0.118408203125,
0.28076171875,
0.387939453125,
0.28076171875,
0.387939453125,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.52880859375,
0.283935546875,
0.8818359375,
0.28515625,
0.8818359375,
0.2998046875,
0.52880859375,
0.298583984375
],
[
0.1368408203125,
0.29638671875,
0.48779296875,
0.29638671875,
0.48779296875,
0.31005859375,
0.1368408203125,
0.31005859375
],
[
0.51318359375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.118408203125,
0.312255859375,
0.489501953125,
0.312255859375,
0.489501953125,
0.32666015625,
0.118408203125,
0.32666015625
],
[
0.51318359375,
0.317626953125,
0.88330078125,
0.317626953125,
0.88330078125,
0.33203125,
0.51318359375,
0.33203125
],
[
0.118408203125,
0.328857421875,
0.489501953125,
0.328857421875,
0.489501953125,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.5146484375,
0.334228515625,
0.8818359375,
0.334228515625,
0.8818359375,
0.347412109375,
0.5146484375,
0.347412109375
],
[
0.12005615234375,
0.344482421875,
0.489501953125,
0.344482421875,
0.489501953125,
0.358154296875,
0.12005615234375,
0.358154296875
],
[
0.5146484375,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.118408203125,
0.361083984375,
0.489501953125,
0.361083984375,
0.489501953125,
0.374755859375,
0.118408203125,
0.374755859375
],
[
0.51318359375,
0.366455078125,
0.8759765625,
0.365234375,
0.8759765625,
0.380126953125,
0.51318359375,
0.38134765625
],
[
0.1175537109375,
0.377197265625,
0.490234375,
0.377197265625,
0.490234375,
0.39013671875,
0.1175537109375,
0.39013671875
],
[
0.52880859375,
0.38134765625,
0.88330078125,
0.3818359375,
0.88330078125,
0.396728515625,
0.52880859375,
0.3955078125
],
[
0.1175537109375,
0.391357421875,
0.486083984375,
0.391357421875,
0.486083984375,
0.40625,
0.1175537109375,
0.40625
],
[
0.51318359375,
0.3984375,
0.88330078125,
0.3984375,
0.88330078125,
0.412109375,
0.51318359375,
0.412109375
],
[
0.118408203125,
0.4091796875,
0.37451171875,
0.4091796875,
0.37451171875,
0.4228515625,
0.118408203125,
0.4228515625
],
[
0.5146484375,
0.416259765625,
0.88330078125,
0.416259765625,
0.88330078125,
0.427001953125,
0.5146484375,
0.427001953125
],
[
0.118408203125,
0.429931640625,
0.2216796875,
0.429931640625,
0.2216796875,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.51318359375,
0.429931640625,
0.8037109375,
0.43115234375,
0.8037109375,
0.4453125,
0.51318359375,
0.44482421875
],
[
0.51318359375,
0.453125,
0.63818359375,
0.453125,
0.63818359375,
0.46728515625,
0.51318359375,
0.46728515625
],
[
0.156982421875,
0.468505859375,
0.2225341796875,
0.468505859375,
0.2225341796875,
0.482666015625,
0.156982421875,
0.482666015625
],
[
0.252685546875,
0.468505859375,
0.32080078125,
0.468505859375,
0.32080078125,
0.482666015625,
0.252685546875,
0.482666015625
],
[
0.33154296875,
0.468505859375,
0.3779296875,
0.468505859375,
0.3779296875,
0.482666015625,
0.33154296875,
0.482666015625
],
[
0.411376953125,
0.47265625,
0.4248046875,
0.47265625,
0.4248046875,
0.482666015625,
0.411376953125,
0.482666015625
],
[
0.5146484375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4892578125,
0.5146484375,
0.4892578125
],
[
0.3837890625,
0.49462890625,
0.44580078125,
0.49462890625,
0.44580078125,
0.51025390625,
0.3837890625,
0.51025390625
],
[
0.51318359375,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.505859375,
0.51318359375,
0.505859375
],
[
0.26123046875,
0.495361328125,
0.310546875,
0.495361328125,
0.310546875,
0.51025390625,
0.26123046875,
0.51025390625
],
[
0.1561279296875,
0.495361328125,
0.202392578125,
0.495361328125,
0.202392578125,
0.51123046875,
0.1561279296875,
0.51123046875
],
[
0.33154296875,
0.495361328125,
0.387939453125,
0.495361328125,
0.387939453125,
0.51025390625,
0.33154296875,
0.51025390625
],
[
0.51318359375,
0.50830078125,
0.884765625,
0.50830078125,
0.884765625,
0.521484375,
0.51318359375,
0.521484375
],
[
0.1561279296875,
0.5166015625,
0.210693359375,
0.5166015625,
0.210693359375,
0.5322265625,
0.1561279296875,
0.5322265625
],
[
0.26123046875,
0.5166015625,
0.312255859375,
0.5166015625,
0.312255859375,
0.53076171875,
0.26123046875,
0.53076171875
],
[
0.33154296875,
0.5166015625,
0.382080078125,
0.5166015625,
0.382080078125,
0.53076171875,
0.33154296875,
0.53076171875
],
[
0.385498046875,
0.5166015625,
0.44580078125,
0.5166015625,
0.44580078125,
0.53076171875,
0.385498046875,
0.53076171875
],
[
0.51318359375,
0.52392578125,
0.884765625,
0.52392578125,
0.884765625,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.26123046875,
0.5361328125,
0.310546875,
0.5361328125,
0.310546875,
0.5517578125,
0.26123046875,
0.5517578125
],
[
0.330078125,
0.5361328125,
0.382080078125,
0.5361328125,
0.382080078125,
0.5517578125,
0.330078125,
0.5517578125
],
[
0.3837890625,
0.53515625,
0.444091796875,
0.53759765625,
0.443359375,
0.5517578125,
0.382080078125,
0.5498046875
],
[
0.51318359375,
0.54052734375,
0.884765625,
0.54052734375,
0.884765625,
0.55419921875,
0.51318359375,
0.55419921875
],
[
0.51123046875,
0.5546875,
0.884765625,
0.5556640625,
0.884765625,
0.57080078125,
0.51123046875,
0.5693359375
],
[
0.156982421875,
0.53759765625,
0.2435302734375,
0.53759765625,
0.2435302734375,
0.5517578125,
0.156982421875,
0.5517578125
],
[
0.1175537109375,
0.5654296875,
0.48779296875,
0.5654296875,
0.48779296875,
0.5791015625,
0.1175537109375,
0.5791015625
],
[
0.51123046875,
0.57177734375,
0.85546875,
0.57177734375,
0.85546875,
0.5849609375,
0.51123046875,
0.5849609375
],
[
0.12005615234375,
0.58203125,
0.48779296875,
0.58203125,
0.48779296875,
0.595703125,
0.12005615234375,
0.595703125
],
[
0.53076171875,
0.58837890625,
0.884765625,
0.58837890625,
0.884765625,
0.6025390625,
0.53076171875,
0.6025390625
],
[
0.118408203125,
0.5986328125,
0.3359375,
0.5986328125,
0.3359375,
0.6123046875,
0.118408203125,
0.6123046875
],
[
0.51318359375,
0.60498046875,
0.88330078125,
0.60498046875,
0.88330078125,
0.6181640625,
0.51318359375,
0.6181640625
],
[
0.1376953125,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.1376953125,
0.6396484375
],
[
0.51123046875,
0.62060546875,
0.88330078125,
0.619140625,
0.88330078125,
0.63427734375,
0.51123046875,
0.63525390625
],
[
0.5146484375,
0.6357421875,
0.88330078125,
0.6357421875,
0.88330078125,
0.6494140625,
0.5146484375,
0.6494140625
],
[
0.118408203125,
0.642578125,
0.484375,
0.642578125,
0.484375,
0.65283203125,
0.118408203125,
0.65283203125
],
[
0.51123046875,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.51123046875,
0.6650390625
],
[
0.1175537109375,
0.65771484375,
0.490234375,
0.65625,
0.490234375,
0.67041015625,
0.1175537109375,
0.6728515625
],
[
0.51318359375,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.12005615234375,
0.67578125,
0.48779296875,
0.67578125,
0.48779296875,
0.68603515625,
0.12005615234375,
0.68603515625
],
[
0.118408203125,
0.6904296875,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.118408203125,
0.70361328125
],
[
0.51123046875,
0.68505859375,
0.884765625,
0.68408203125,
0.884765625,
0.69873046875,
0.51123046875,
0.69970703125
],
[
0.118408203125,
0.7080078125,
0.48779296875,
0.7080078125,
0.48779296875,
0.71875,
0.118408203125,
0.71875
],
[
0.5146484375,
0.703125,
0.71875,
0.703125,
0.71875,
0.71337890625,
0.5146484375,
0.71337890625
],
[
0.53076171875,
0.71630859375,
0.8818359375,
0.71728515625,
0.8818359375,
0.732421875,
0.53076171875,
0.73095703125
],
[
0.118408203125,
0.72265625,
0.453369140625,
0.72265625,
0.453369140625,
0.73681640625,
0.118408203125,
0.73681640625
],
[
0.51123046875,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.51123046875,
0.74658203125
],
[
0.118408203125,
0.74365234375,
0.267822265625,
0.74365234375,
0.267822265625,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12005615234375,
0.765625,
0.489501953125,
0.765625,
0.489501953125,
0.779296875,
0.12005615234375,
0.779296875
],
[
0.51318359375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.118408203125,
0.783203125,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.118408203125,
0.79638671875
],
[
0.51318359375,
0.783203125,
0.88330078125,
0.783203125,
0.88330078125,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.798828125,
0.71728515625,
0.798828125,
0.71728515625,
0.8125,
0.51318359375,
0.8125
],
[
0.12005615234375,
0.8154296875,
0.490234375,
0.8154296875,
0.490234375,
0.8291015625,
0.12005615234375,
0.8291015625
],
[
0.5322265625,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.5322265625,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.51318359375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.12176513671875,
0.8642578125,
0.48779296875,
0.8642578125,
0.48779296875,
0.87451171875,
0.12176513671875,
0.87451171875
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87744140625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.118408203125,
0.892578125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.892578125,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Sundheim, 1993). We use bootstrap sampling to",
"Other approaches have addressed the need for",
"estimate the probability that the model with the",
"token-level labels when only raw output values",
"best micro average Fl score on the entire test set",
"are available. Mintz et al. (2009) introduced dis-",
"is worse for a randomly sampled subset of the test",
"tant supervision, which heuristically generates the",
"data.",
"token-level labels from the output values.You do",
"3.3Ourmodel",
"this by searching for input tokens that matches out-",
"put values. The matching tokens are then assigned",
"Since our decoders can only output values that are",
"the labels for the matching outputs. One drawback",
"present in the input, we prepend a single comma",
"is that the quality of the labels crucially depend on",
"to every input sequence. We optimize our model",
"the search algorithm and how closely the tokens",
"using Adam and use early stopping on a held-out",
"match the output values, which makes it brittle.",
"validation set. The model quickly converges to op-",
"Our method is trained end-to-end, thus not relying",
"timal performance, usually after around 5000 up-",
"on brittle heuristics.",
"dates after which it starts overfitting.",
"Sutskever et al. (2014) opened up the sequence-",
"For the restaurant data set, to increase perfor-",
"to-sequence paradigm. With the addition of at-",
"mance, we double the sizes of all the parameters",
"tention (Bahdanau et al., 2014), these models",
"and use embedding and recurrent dropout follow-",
"achieved state-of-the-art results in machine trans-",
"ing (Gal, 2015). Further, we add a summarizer",
"lation (Wuetal..20l6).Wearebroadlyinspired",
"LSTM to each decoder. The summarizer LSTM",
"by these results to investigate E2E models for IE.",
"reads the entire encoded input. The last hidden",
"The idea of copying words from the input to the",
"state of the summarizer LSTM is then concate-",
"output have been used in machine translation to",
"nated to each input to the decoder.",
"overcomeproblemswithout-of-vocabularywords",
"3.4 Results",
"(Gulcehre et al., 2016; Gu et al., 2016).",
"5Discussion",
"Data set",
"Baseline",
"Ours",
"p",
"We present an end-to-end IE model that does not",
"0.1755",
"require detailed token-level labels. Despite being",
"0.977",
"ATIS",
"0.974",
"trained end-to-end, it is competitive with baseline",
"Movie",
"0.816",
"0.817",
"0.3792",
"models relying on token-level labels. In contrast",
"0.724",
"0.694",
"0.0001",
"to them, our model can be used on many real life",
"IE tasks where intermediate token-level labels are",
"Restaurant",
"Table 3: Micro average F1 scores on the E2E data",
"not available and creating them is not feasible.",
"sets. Results that are significantly better (p <",
"In our experiments our model and the baselines",
"0.05) are highlighted in bold.",
"had access to the same amount of training sam-",
"We see in Table 3 that our model is competi-",
"ples. In a real life scenario it is likely that token-",
"level labels only exist for a subset of all the data.",
"tive with the baseline models in terms of micro-",
"It would be interesting to investigate the quanti-",
"averaged F1 for two of the three data sets. This",
"ty/quality trade-of of the labels, and a multi task",
"is a remarkableresultgiven that the baselines are",
"trained on token-level labels, whereas our model",
"extension to the model, which could make use of",
"is trained end-to-end. For the restaurant data set,",
"availabletoken-level labels.",
"Our model is remarkably stable to hyper param-",
"our model is slightly worse than the baseline.",
"eter changes. On the restaurant dataset we tried",
"4 Related work",
"several different architectures and hyper parame-",
"Event extraction (EE) is similar to the E2E IE task",
"ters before settling on the reported one. The differ-",
"we propose, except that it can have several event",
"ence between the worst and the best was approxi-",
"types and multiple events per input. In our E2E IE",
"mately 2 percentage points.",
"task, we only have a single event type and assume",
"A major limitation of the proposed model is that",
"there is zero or one event mentioned in the input,",
"it can only output values that are present in the in-",
"which is an easier task. Recently, Nguyen et al.",
"put. This is a problem for outputs that are nor-",
"(2016)achieved state of the artresults on the ACE",
"malized before being submitted as machine read-",
"2005 EE data set using a recurrent neural network",
"able data, which is a common occurrence. For in-",
"to jointly model event triggers and argument roles.",
"stance, dates might appear as'Jan 17 2012' in"
] |
{
"arxiv_id": "1707.04913",
"language": "en",
"page_idx": 3
}
|
||
2203.08597-en-0
|
2203.08597-en-0.jpg
|
## Less is More: Summary of Long Instructions is Better for $\dot{\textbf{Program Synthesis}}$
$\textbf{Chitta Baral}$
Kirby Kuznia* Swaroop Mishra* Mihir Parmar
Arizona State University
## Abstract
Despite the success of large pre-trained language models (LMs) such as Codex, they show below-par performance on the larger and more complicated programming related questions. We show that LMs benefit from the summarized version of complicated questions. Our findings show that superfluous information often present in problem description such as human characters, background stories, and names (which are included to help humans in understanding a task) does not help models in understanding a task. To this extent, we create a meta-dataset from the frequently used APPS dataset and the newly created CodeContests dataset for the program synthesis task. Our meta-dataset consists of human and synthesized summaries of the long and complicated programming questions. Experimental results on Codex show that our proposed approach outperforms baseline by 8.13% on the APPS dataset and 11.88% on the CodeContests dataset on average in terms of strict accuracy. Our analysis shows that summaries significantly improve performance for introductory $(9.86\%)$ and interview $(11.48\%)$ programming questions. However, it shows improvement by a small margin $(\sim2\%)$ for competitive programming questions, implying scope for future research in this direction.$^{\mathrm{l}}$
### 1 Introduction
Recently, large pre-trained LMs have been proven pivotal in programming-related tasks (Wang et al., 2021; Chen et al., 2021; Hendrycks et al., 2021; Lu et al., 2021; Papineni et al., 2002)?. Program synthesis aims to generate a code given the natural language description of a problem. Programming requirements in these problems vary in terms of complexity from a 3-5 line simple function to mul-
$^*$Equal Contribution
Code and data is available at https://github.com/kurbster/
Prompt- Summarization
$^{2}$Detailed related work is presented in Appendix A
tiple functions that use advanced data structures However, LMs such as Codex show below-par performance on the long and complicated programming questions. We observe that the natural language description of the program becomes long and complicated when there is superfluous information (see section 2.1.1). The goal of adding this information to the description is to make it more understandable to humans. However, we find that this information confuses the model in understanding a task$^3.$ We propose that removing the excess information and providing the model with the exact specifications of the problem can improve the performance of the LMs.
To remove excess information$^{4}$,we summarize the descriptions of the program in such a way that it does not lose important specifications. We use the APPS dataset (Hendrycks et al., 2021) and CodeContests dataset (Li et al., 2022) which are a collection of coding problems from different online sources and create a meta-dataset consisting of human and synthesized summaries.
We perform all experiments using the GPT-based Codex model (Chen et al., 2021) on the proposed meta-dataset and show that the summarized version of complicated questions improves strict accuracy by 8.13% on the APPS dataset and 11.85% on CodeContests. From our analysis, we can see significant improvement for introductory $(9.86\%)$ and interview (11.48%) related programming questions. However, it shows improvement by a small margin $(\sim2\%)$ for competitive programming questions. Considering that automatic evaluation of a program does not reward for partial correctness, we perform qualitative evaluation on our meta-dataset and find that original questions often confuse models in understanding the underlying problem, as models latch on to some spurious words in the text
See example in Appendix C
$^{4}$Instructions for creating summaries given in Appendix N
|
[
[
0.204833984375,
0.09320068359375,
0.7958984375,
0.09320068359375,
0.7958984375,
0.10809326171875,
0.204833984375,
0.10809326171875
],
[
0.403076171875,
0.1134033203125,
0.59619140625,
0.1134033203125,
0.59619140625,
0.127685546875,
0.403076171875,
0.127685546875
],
[
0.2166748046875,
0.1561279296875,
0.65576171875,
0.1561279296875,
0.65576171875,
0.17041015625,
0.2166748046875,
0.17041015625
],
[
0.67236328125,
0.1561279296875,
0.7841796875,
0.1561279296875,
0.7841796875,
0.17041015625,
0.67236328125,
0.17041015625
],
[
0.399658203125,
0.189453125,
0.60107421875,
0.190673828125,
0.60107421875,
0.204833984375,
0.399658203125,
0.2037353515625
],
[
0.26123046875,
0.2529296875,
0.340087890625,
0.2529296875,
0.340087890625,
0.267333984375,
0.26123046875,
0.267333984375
],
[
0.51123046875,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51123046875,
0.267333984375
],
[
0.51318359375,
0.270751953125,
0.884765625,
0.270751953125,
0.884765625,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.1435546875,
0.276611328125,
0.459228515625,
0.27783203125,
0.459228515625,
0.292236328125,
0.1435546875,
0.29150390625
],
[
0.51318359375,
0.2861328125,
0.88330078125,
0.2861328125,
0.88330078125,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.145263671875,
0.292236328125,
0.45751953125,
0.292236328125,
0.45751953125,
0.305908203125,
0.145263671875,
0.305908203125
],
[
0.145263671875,
0.30712890625,
0.45751953125,
0.30712890625,
0.45751953125,
0.320556640625,
0.145263671875,
0.320556640625
],
[
0.51123046875,
0.302734375,
0.88330078125,
0.302734375,
0.88330078125,
0.31640625,
0.51123046875,
0.31640625
],
[
0.0251922607421875,
0.31005859375,
0.060455322265625,
0.31005859375,
0.060455322265625,
0.43701171875,
0.0251922607421875,
0.43701171875
],
[
0.14697265625,
0.320556640625,
0.459228515625,
0.320556640625,
0.459228515625,
0.334228515625,
0.14697265625,
0.334228515625
],
[
0.51318359375,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.145263671875,
0.3349609375,
0.45751953125,
0.3349609375,
0.45751953125,
0.3486328125,
0.145263671875,
0.3486328125
],
[
0.51123046875,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51123046875,
0.3486328125
],
[
0.1485595703125,
0.350830078125,
0.45751953125,
0.350830078125,
0.45751953125,
0.361083984375,
0.1485595703125,
0.361083984375
],
[
0.51123046875,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.365234375,
0.51123046875,
0.365234375
],
[
0.14697265625,
0.363525390625,
0.459228515625,
0.363525390625,
0.459228515625,
0.377197265625,
0.14697265625,
0.377197265625
],
[
0.509765625,
0.365234375,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.509765625,
0.37890625
],
[
0.14697265625,
0.377685546875,
0.459228515625,
0.377685546875,
0.459228515625,
0.391357421875,
0.14697265625,
0.391357421875
],
[
0.51123046875,
0.383056640625,
0.880859375,
0.383056640625,
0.880859375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.145263671875,
0.392578125,
0.459228515625,
0.392578125,
0.459228515625,
0.40625,
0.145263671875,
0.40625
],
[
0.145263671875,
0.40625,
0.459228515625,
0.40625,
0.459228515625,
0.419189453125,
0.145263671875,
0.419189453125
],
[
0.5146484375,
0.40087890625,
0.880859375,
0.40087890625,
0.880859375,
0.410888671875,
0.5146484375,
0.410888671875
],
[
0.14697265625,
0.42041015625,
0.459228515625,
0.42041015625,
0.459228515625,
0.43408203125,
0.14697265625,
0.43408203125
],
[
0.51123046875,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.14697265625,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.448974609375,
0.14697265625,
0.448974609375
],
[
0.51123046875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.0235137939453125,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0235137939453125,
0.5185546875
],
[
0.14697265625,
0.448974609375,
0.459228515625,
0.448974609375,
0.459228515625,
0.4619140625,
0.14697265625,
0.4619140625
],
[
0.509765625,
0.44775390625,
0.880859375,
0.44775390625,
0.880859375,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.1485595703125,
0.46435546875,
0.45751953125,
0.46435546875,
0.45751953125,
0.474365234375,
0.1485595703125,
0.474365234375
],
[
0.509765625,
0.463134765625,
0.7001953125,
0.4619140625,
0.7001953125,
0.476806640625,
0.509765625,
0.47802734375
],
[
0.14697265625,
0.476806640625,
0.459228515625,
0.476806640625,
0.459228515625,
0.490478515625,
0.14697265625,
0.490478515625
],
[
0.145263671875,
0.4912109375,
0.459228515625,
0.4912109375,
0.459228515625,
0.5048828125,
0.145263671875,
0.5048828125
],
[
0.51123046875,
0.48681640625,
0.8818359375,
0.48681640625,
0.8818359375,
0.501953125,
0.51123046875,
0.501953125
],
[
0.1435546875,
0.50439453125,
0.459228515625,
0.5048828125,
0.459228515625,
0.51953125,
0.1435546875,
0.5185546875
],
[
0.51123046875,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.51123046875,
0.5185546875
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.1435546875,
0.51953125,
0.459228515625,
0.5185546875,
0.459228515625,
0.5322265625,
0.1435546875,
0.533203125
],
[
0.509765625,
0.52099609375,
0.8818359375,
0.51953125,
0.8818359375,
0.53369140625,
0.509765625,
0.53515625
],
[
0.145263671875,
0.53369140625,
0.459228515625,
0.53369140625,
0.459228515625,
0.54736328125,
0.145263671875,
0.54736328125
],
[
0.51318359375,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.55029296875,
0.51318359375,
0.55029296875
],
[
0.14697265625,
0.5498046875,
0.45751953125,
0.5498046875,
0.45751953125,
0.56005859375,
0.14697265625,
0.56005859375
],
[
0.51318359375,
0.552734375,
0.8818359375,
0.552734375,
0.8818359375,
0.56640625,
0.51318359375,
0.56640625
],
[
0.145263671875,
0.56103515625,
0.4609375,
0.5625,
0.4609375,
0.576171875,
0.145263671875,
0.57470703125
],
[
0.509765625,
0.568359375,
0.8818359375,
0.56787109375,
0.8818359375,
0.58203125,
0.509765625,
0.5830078125
],
[
0.145263671875,
0.57666015625,
0.459228515625,
0.57666015625,
0.459228515625,
0.59033203125,
0.145263671875,
0.59033203125
],
[
0.509765625,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.509765625,
0.59716796875
],
[
0.1435546875,
0.59033203125,
0.459228515625,
0.59130859375,
0.459228515625,
0.60498046875,
0.1435546875,
0.60400390625
],
[
0.145263671875,
0.60498046875,
0.459228515625,
0.60498046875,
0.459228515625,
0.6181640625,
0.145263671875,
0.6181640625
],
[
0.51123046875,
0.6015625,
0.755859375,
0.6015625,
0.755859375,
0.615234375,
0.51123046875,
0.615234375
],
[
0.145263671875,
0.6181640625,
0.459228515625,
0.6181640625,
0.459228515625,
0.6318359375,
0.145263671875,
0.6318359375
],
[
0.51123046875,
0.62451171875,
0.880859375,
0.62451171875,
0.880859375,
0.63818359375,
0.51123046875,
0.63818359375
],
[
0.145263671875,
0.6318359375,
0.45751953125,
0.6318359375,
0.45751953125,
0.6455078125,
0.145263671875,
0.6455078125
],
[
0.145263671875,
0.64794921875,
0.459228515625,
0.64794921875,
0.459228515625,
0.6611328125,
0.145263671875,
0.6611328125
],
[
0.51318359375,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65625,
0.51318359375,
0.65625
],
[
0.14697265625,
0.662109375,
0.459228515625,
0.662109375,
0.459228515625,
0.67578125,
0.14697265625,
0.67578125
],
[
0.51123046875,
0.65771484375,
0.88330078125,
0.65771484375,
0.88330078125,
0.67138671875,
0.51123046875,
0.67138671875
],
[
0.145263671875,
0.67578125,
0.391357421875,
0.67578125,
0.391357421875,
0.689453125,
0.145263671875,
0.689453125
],
[
0.51123046875,
0.67333984375,
0.88330078125,
0.6748046875,
0.88330078125,
0.6884765625,
0.51123046875,
0.68701171875
],
[
0.51318359375,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.70361328125,
0.51318359375,
0.70361328125
],
[
0.1435546875,
0.70361328125,
0.258544921875,
0.70361328125,
0.258544921875,
0.71875,
0.1435546875,
0.71875
],
[
0.51123046875,
0.705078125,
0.8818359375,
0.70703125,
0.8818359375,
0.72021484375,
0.51123046875,
0.71875
],
[
0.51123046875,
0.72265625,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.51123046875,
0.73681640625
],
[
0.118408203125,
0.7060546875,
0.1318359375,
0.7060546875,
0.1318359375,
0.71533203125,
0.118408203125,
0.71533203125
],
[
0.1175537109375,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74462890625,
0.1175537109375,
0.74462890625
],
[
0.51123046875,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.1175537109375,
0.74658203125,
0.489501953125,
0.74658203125,
0.489501953125,
0.76123046875,
0.1175537109375,
0.76123046875
],
[
0.509765625,
0.75439453125,
0.880859375,
0.75439453125,
0.880859375,
0.76708984375,
0.509765625,
0.76708984375
],
[
0.1175537109375,
0.7626953125,
0.489501953125,
0.7626953125,
0.489501953125,
0.775390625,
0.1175537109375,
0.775390625
],
[
0.509765625,
0.771484375,
0.884765625,
0.771484375,
0.884765625,
0.78515625,
0.509765625,
0.78515625
],
[
0.1158447265625,
0.77783203125,
0.48779296875,
0.779296875,
0.48779296875,
0.79345703125,
0.1158447265625,
0.7919921875
],
[
0.5087890625,
0.7861328125,
0.8818359375,
0.7861328125,
0.8818359375,
0.7998046875,
0.5087890625,
0.7998046875
],
[
0.1175537109375,
0.7958984375,
0.48779296875,
0.7958984375,
0.48779296875,
0.80859375,
0.1175537109375,
0.80859375
],
[
0.51123046875,
0.802734375,
0.8818359375,
0.8017578125,
0.8818359375,
0.81640625,
0.51123046875,
0.81787109375
],
[
0.1175537109375,
0.81103515625,
0.48779296875,
0.81103515625,
0.48779296875,
0.826171875,
0.1175537109375,
0.826171875
],
[
0.51318359375,
0.8193359375,
0.8818359375,
0.8193359375,
0.8818359375,
0.8330078125,
0.51318359375,
0.8330078125
],
[
0.1158447265625,
0.82763671875,
0.48779296875,
0.82666015625,
0.48779296875,
0.84033203125,
0.1158447265625,
0.84130859375
],
[
0.51123046875,
0.8349609375,
0.880859375,
0.8349609375,
0.880859375,
0.8486328125,
0.51123046875,
0.8486328125
],
[
0.1175537109375,
0.84326171875,
0.490234375,
0.84326171875,
0.490234375,
0.85693359375,
0.1175537109375,
0.85693359375
],
[
0.51318359375,
0.8515625,
0.880859375,
0.8515625,
0.880859375,
0.865234375,
0.51318359375,
0.865234375
],
[
0.1485595703125,
0.86767578125,
0.270263671875,
0.86767578125,
0.270263671875,
0.88134765625,
0.1485595703125,
0.88134765625
],
[
0.51123046875,
0.86767578125,
0.8818359375,
0.86767578125,
0.8818359375,
0.88134765625,
0.51123046875,
0.88134765625
],
[
0.14111328125,
0.8798828125,
0.48779296875,
0.8798828125,
0.48779296875,
0.8935546875,
0.14111328125,
0.8935546875
],
[
0.1175537109375,
0.892578125,
0.26123046875,
0.892578125,
0.26123046875,
0.90625,
0.1175537109375,
0.90625
],
[
0.5341796875,
0.89111328125,
0.7197265625,
0.89111328125,
0.7197265625,
0.90625,
0.5341796875,
0.90625
],
[
0.1376953125,
0.90478515625,
0.447509765625,
0.90478515625,
0.447509765625,
0.91845703125,
0.1376953125,
0.91845703125
],
[
0.5341796875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.5341796875,
0.91943359375
]
] |
[
"Less is More: Summary of Long Instructions is Better for",
"Program Synthesis",
"Kirby Kuznia*Swaroop Mishra*Mihir Parmar",
"Chitta Baral",
"Arizona State University",
"Abstract",
"tiple functions that use advanced data structures",
"However, LMs such as Codex show below-par per-",
"Despite the success of large pre-trained lan-",
"formance on the long and complicated program-",
"guage models (LMs) such as Codex, they",
"show below-par performance on the larger and",
"ming questions. We observe that the natural lan-",
"",
"more complicated programming related ques-",
"guage description of the program becomes long",
"tions. We show that LMs benefit from the",
"and complicated when there is superfluous infor-",
"summarizedversionofcomplicatedquestions",
"mation (see section 2.1.1). The goal of adding this",
"Our findings show that superfluous informa-",
"information to the description is to make it more",
"tion often present in problem description such",
"understandable to humans. However, we find that",
"as human characters, background stories, and",
"names (which are included to help humans in",
"thisinformationconfusesthemodelinunderstand-",
"understanding a task) does not help models in",
"ing a task3. We propose that removing the excess",
"understanding a task. To this extent, we cre-",
"information and providing the model with the ex-",
"['s]",
"ate a meta-dataset from the frequently used",
"act specifications of the problem can improve the",
"APPSdatasetandthenewlycreatedCodeCon-",
"performance of the LMs.",
"tests dataset for the program synthesis task.",
"Our meta-dataset consists of human and syn-",
"To remove excess informationt, we summarize the",
"thesized summaries of the long and compli-",
"descriptions of the program in such a way that it",
"680502111",
"cated programming questions.Experimental",
"does not lose important specifications. We use the",
"results on Codex show that our proposed ap-",
"APPS dataset (Hendrycks et al., 2021) and Code-",
"proachoutperformsbaselineby8.13%onthe",
"Contests dataset (Li et al., 2022) which are a col-",
"APPS dataset and 11.88% on the CodeCon-",
"lection of coding problems from different online",
"tests dataset on average in terms of strict accu-",
"sources and create a meta-dataset consisting of hu",
"racy. Our analysis shows that summaries sig-",
"nificantly improve performance for introduc-",
"man and synthesized summaries.",
"tory (9.86%) and interview (11.48%) program-",
"We perform all experiments using the GPT-based",
"ming questions. However, it shows improve",
"ment by a small margin (~ 2%) for compet-",
"Codex model (Chen et al., 2021) on the proposed",
"itive programming questions, implying scope",
"meta-dataset and show that the summarized ver-",
"for future research in this direction.!",
"sion of complicated questions improves strict ac-",
"curacy by 8.13% on the APPS dataset and 11.85%",
"Introduction",
"on CodeContests. From our analysis, we can see",
"significant improvement for introductory (9.86%)",
"1",
"Recently, large pre-trained LMs have been proven",
"and interview (11.48%) related programming ques-",
"pivotal in programming-related tasks (Wang et al.,",
"tions. However, it shows improvement by a small",
"2021; Chen et al., 2021; Hendrycks et al., 2021;",
"margin (~ 2%) for competitive programming ques-",
"Lu et al., 2021; Papineni et al., 2002)2. Program",
" tions. Considering that automatic evaluation of a",
"synthesis aims to generate a code given the natural",
"program does not reward for partial correctness, we",
"language description of a problem. Programming",
"perform qualitative evaluation on our meta-dataset",
"requirements in these problems vary in terms of",
"and find that original questions often confuse mod-",
"complexity from a 3-5 line simple function to mul-",
"els in understanding the underlying problem, as",
"Equal Contribution",
"models latch on to some spurious words in the text",
"'Code and data is available at https://github.com/kurbster/",
"Prompt-Summarization",
"3See example in Appendix C",
"2Detailed related work is presented in Appendix A",
"AInstructions for creating summaries given in Appendix N"
] |
{
"arxiv_id": "2203.08597",
"language": "en",
"page_idx": 0
}
|
||
2203.08597-en-1
|
2203.08597-en-1.jpg
|
(e.g. the word `list' in question makes the model design a list even though the underlying problem is on graphs). We further analyze model performance on different types of summaries (i.e., basic, expert, and synthetic) and provide instruction-design principles that can help future research on prompting in program synthesis.
## 2 $\textbf{Method}$
## 2.1 Dataset
We use the APPS (Hendrycks et al.,2021) and CodeContests (Li et al., 2022) datasets to create summaries. We crowd-sourced the creation of human summaries. The result was 373 human summaries for APPS and 80 summaries for CodeContests along with and 8663 synthetic summaries using both datasets. Table 1 shows the statistics of the generated summaries.
<table>
<tbody>
<tr>
<th>Data $\mathbf{Source}$ </th>
<th>Difficulty</th>
<th>$\#$of Problems</th>
</tr>
<tr>
<td rowspan="5">Human</td>
<td>Introductory Interview Competition CodeContests</td>
<td>145 123 105 80</td>
</tr>
<tr>
<td>Introductory</td>
<td>145</td>
</tr>
<tr>
<td>Interview</td>
<td>123</td>
</tr>
<tr>
<td>Competition</td>
<td>105</td>
</tr>
<tr>
<td>CodeContests</td>
<td>80</td>
</tr>
<tr>
<td> </td>
<td>Total</td>
<td>453</td>
</tr>
<tr>
<td rowspan="5">Studio21</td>
<td>Introductory Interview Competition CodeContests</td>
<td>1588 4551 1286 80</td>
</tr>
<tr>
<td>Introductory</td>
<td>1588</td>
</tr>
<tr>
<td>Interview</td>
<td>4551</td>
</tr>
<tr>
<td>Competition</td>
<td>1286</td>
</tr>
<tr>
<td>CodeContests</td>
<td>80</td>
</tr>
<tr>
<td> </td>
<td>Total</td>
<td>7505</td>
</tr>
<tr>
<td rowspan="1">$GPT-3$</td>
<td>Introductory</td>
<td>194</td>
</tr>
<tr>
<td rowspan="5">$GPT-3$</td>
<td>Interview</td>
<td>267</td>
</tr>
<tr>
<td>Competition</td>
<td>244</td>
</tr>
<tr>
<td>CodeContests</td>
<td>80</td>
</tr>
<tr>
<td>Total</td>
<td>785</td>
</tr>
<tr>
<td>Introductory</td>
<td>145</td>
</tr>
<tr>
<td rowspan="3">PEGASUS</td>
<td>Interview</td>
<td>123</td>
</tr>
<tr>
<td>Competition</td>
<td>105</td>
</tr>
<tr>
<td>Total</td>
<td>373</td>
</tr>
</tbody>
</table>
Table 1: Statistics of the proposed meta-dataset.
2.1.1 Human Generated Summaries
For the APPS and CodeContests human-generated summaries, the crowd worker reads and understands the original questions, then creates summaries in two steps$^{5}.$ First, we create a basic summary of the given problem and remove any information that is repeated and any hypothetical information without concrete instructions. For example, if the problem constructs a fake company or
$^{5}$Instructions for creating summaries are in Appendix N
situation, we replace the fake situation with direct instructions. Full example is included in Appendix C. Second, we create an expert summary of the problem. To create this, we further summarize the first summary. This expert summary includes the absolute minimum information for an expert to understand the problem. We would not expect a novice to understand these prompts. An example of expert summaries is given in Appendix C.3.
## 2.1.2 Synthetic Summaries
We have generated synthetic summaries of program descriptions using jumbo (178B), large (7.5B) Studio21 model (Lieber et al., 2021), GPT-3 Davinci model (175B) (Brown et al., 2020) and PEGASUS model (Zhang et al., 2019). To generate a summary, we provide these models with a few examples in the in-context learning setup (Brown et al., 2020) from the human-generated summaries. For the few-shot examples, we use expert-level summaries.
Studio21 We use five examples with the large model, and three examples with the jumbo model$^{6}$ For both models, we use a temperature of 0.3, and topP of 1. For the format of our prompt, we use DeJargonizer template$^{7}$ with a change to their header as shown in Appendix D. We create a total of 7,505 synthetic summaries using these models
GPT-3 We use three examples for GPT-3 model. We empirically set temperature to 0.05, topP to 1, frequency penalty to 0.01, presence penalty to 0.05. To generate prompts, we followed their tl;dr template$^{8}$ as shown in Appendix D. We create 785 synthetic summaries using this model.
PEGASUS We use the PEGASUS model (Zhang et al., 2019) to create program summaries for the same set of problems that were summarized by humans. We choose this model because it was trained specifically for abstractive summarization.
## 2.2 Model
We use OpenAI Codex to build baselines and the
proposed approach.
Baseline To create a baseline, we have used original program descriptions given in the datasets as prompts for the Codex model.
$^{6}$Examples are included in Appendix D
$^{7}$https:$l/$studio.ai21.com/
$^{8}$https://beta.openai.com/playground/p/
default-tldr-summary?model=text-davinci-001
|
[
[
0.118408203125,
0.08612060546875,
0.48779296875,
0.08612060546875,
0.48779296875,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.509765625,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.118408203125,
0.10394287109375,
0.48779296875,
0.10394287109375,
0.48779296875,
0.114013671875,
0.118408203125,
0.114013671875
],
[
0.509765625,
0.10272216796875,
0.8818359375,
0.10272216796875,
0.8818359375,
0.11639404296875,
0.509765625,
0.11639404296875
],
[
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
],
[
0.1158447265625,
0.1353759765625,
0.489501953125,
0.1353759765625,
0.489501953125,
0.1484375,
0.1158447265625,
0.1484375
],
[
0.51318359375,
0.135986328125,
0.880859375,
0.135986328125,
0.880859375,
0.146728515625,
0.51318359375,
0.146728515625
],
[
0.1158447265625,
0.15087890625,
0.49365234375,
0.15087890625,
0.49365234375,
0.1644287109375,
0.1158447265625,
0.1644287109375
],
[
0.509765625,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.509765625,
0.1644287109375
],
[
0.1158447265625,
0.1650390625,
0.489501953125,
0.166259765625,
0.489501953125,
0.18115234375,
0.1158447265625,
0.179931640625
],
[
0.509765625,
0.16748046875,
0.8818359375,
0.16748046875,
0.8818359375,
0.18115234375,
0.509765625,
0.18115234375
],
[
0.1158447265625,
0.182861328125,
0.27783203125,
0.182861328125,
0.27783203125,
0.196533203125,
0.1158447265625,
0.196533203125
],
[
0.51123046875,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.51123046875,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.11248779296875,
0.207275390625,
0.219970703125,
0.207275390625,
0.219970703125,
0.2244873046875,
0.11248779296875,
0.2244873046875
],
[
0.509765625,
0.21435546875,
0.85546875,
0.213134765625,
0.85546875,
0.22802734375,
0.509765625,
0.2286376953125
],
[
0.1175537109375,
0.235107421875,
0.1485595703125,
0.235107421875,
0.1485595703125,
0.2476806640625,
0.1175537109375,
0.2476806640625
],
[
0.14111328125,
0.236328125,
0.219970703125,
0.236328125,
0.219970703125,
0.2476806640625,
0.14111328125,
0.2476806640625
],
[
0.51123046875,
0.240478515625,
0.73779296875,
0.2410888671875,
0.73779296875,
0.255859375,
0.51123046875,
0.254638671875
],
[
0.11419677734375,
0.255859375,
0.486083984375,
0.255859375,
0.486083984375,
0.26953125,
0.11419677734375,
0.26953125
],
[
0.509765625,
0.260009765625,
0.8818359375,
0.262451171875,
0.8818359375,
0.276611328125,
0.509765625,
0.27490234375
],
[
0.1158447265625,
0.270751953125,
0.48779296875,
0.271484375,
0.48779296875,
0.2861328125,
0.1158447265625,
0.28515625
],
[
0.51318359375,
0.27783203125,
0.88330078125,
0.27783203125,
0.88330078125,
0.292236328125,
0.51318359375,
0.292236328125
],
[
0.1158447265625,
0.2880859375,
0.490234375,
0.2880859375,
0.490234375,
0.3017578125,
0.1158447265625,
0.3017578125
],
[
0.51123046875,
0.292236328125,
0.880859375,
0.292236328125,
0.880859375,
0.305908203125,
0.51123046875,
0.305908203125
],
[
0.1158447265625,
0.303955078125,
0.490234375,
0.303955078125,
0.490234375,
0.317626953125,
0.1158447265625,
0.317626953125
],
[
0.51123046875,
0.31005859375,
0.880859375,
0.31005859375,
0.880859375,
0.32373046875,
0.51123046875,
0.32373046875
],
[
0.11419677734375,
0.318359375,
0.489501953125,
0.318359375,
0.489501953125,
0.33203125,
0.11419677734375,
0.33203125
],
[
0.509765625,
0.32470703125,
0.8818359375,
0.325927734375,
0.8818359375,
0.34033203125,
0.509765625,
0.339111328125
],
[
0.1175537109375,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.1175537109375,
0.349853515625
],
[
0.509765625,
0.342529296875,
0.880859375,
0.341552734375,
0.880859375,
0.355712890625,
0.509765625,
0.35693359375
],
[
0.1175537109375,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.509765625,
0.358154296875,
0.8818359375,
0.35693359375,
0.8818359375,
0.371826171875,
0.509765625,
0.372314453125
],
[
0.1175537109375,
0.3681640625,
0.307373046875,
0.3681640625,
0.307373046875,
0.3818359375,
0.1175537109375,
0.3818359375
],
[
0.51123046875,
0.374755859375,
0.8818359375,
0.374755859375,
0.8818359375,
0.388427734375,
0.51123046875,
0.388427734375
],
[
0.151123046875,
0.396728515625,
0.2418212890625,
0.3984375,
0.2418212890625,
0.413330078125,
0.151123046875,
0.410888671875
],
[
0.26025390625,
0.396728515625,
0.33154296875,
0.3984375,
0.33154296875,
0.413330078125,
0.26025390625,
0.410888671875
],
[
0.51318359375,
0.391357421875,
0.82275390625,
0.391357421875,
0.82275390625,
0.405029296875,
0.51318359375,
0.405029296875
],
[
0.3525390625,
0.399658203125,
0.44921875,
0.399658203125,
0.44921875,
0.410888671875,
0.3525390625,
0.410888671875
],
[
0.252685546875,
0.416259765625,
0.337646484375,
0.418701171875,
0.337646484375,
0.432861328125,
0.252685546875,
0.43115234375
],
[
0.385498046875,
0.416259765625,
0.418212890625,
0.416259765625,
0.418212890625,
0.432861328125,
0.385498046875,
0.432861328125
],
[
0.509765625,
0.416259765625,
0.880859375,
0.418701171875,
0.880859375,
0.432861328125,
0.509765625,
0.43115234375
],
[
0.2626953125,
0.431640625,
0.330078125,
0.431640625,
0.330078125,
0.443603515625,
0.2626953125,
0.443603515625
],
[
0.38623046875,
0.429931640625,
0.414794921875,
0.429931640625,
0.414794921875,
0.443603515625,
0.38623046875,
0.443603515625
],
[
0.17041015625,
0.435791015625,
0.22412109375,
0.435791015625,
0.22412109375,
0.45068359375,
0.17041015625,
0.45068359375
],
[
0.51123046875,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.448974609375,
0.51123046875,
0.448974609375
],
[
0.25439453125,
0.44482421875,
0.337646484375,
0.44482421875,
0.337646484375,
0.4560546875,
0.25439453125,
0.4560546875
],
[
0.38623046875,
0.443603515625,
0.414794921875,
0.443603515625,
0.414794921875,
0.4560546875,
0.38623046875,
0.4560546875
],
[
0.2476806640625,
0.45361328125,
0.343505859375,
0.4560546875,
0.343505859375,
0.47021484375,
0.2476806640625,
0.468505859375
],
[
0.509765625,
0.448974609375,
0.8818359375,
0.449462890625,
0.8818359375,
0.46435546875,
0.509765625,
0.463134765625
],
[
0.387939453125,
0.454833984375,
0.4130859375,
0.454833984375,
0.4130859375,
0.4697265625,
0.387939453125,
0.4697265625
],
[
0.509765625,
0.465576171875,
0.8818359375,
0.466064453125,
0.8818359375,
0.48095703125,
0.509765625,
0.479736328125
],
[
0.276123046875,
0.473876953125,
0.316650390625,
0.473876953125,
0.316650390625,
0.4892578125,
0.276123046875,
0.4892578125
],
[
0.385498046875,
0.474365234375,
0.41650390625,
0.474365234375,
0.41650390625,
0.4892578125,
0.385498046875,
0.4892578125
],
[
0.51123046875,
0.482666015625,
0.88330078125,
0.482666015625,
0.88330078125,
0.496337890625,
0.51123046875,
0.496337890625
],
[
0.252685546875,
0.4921875,
0.339111328125,
0.49462890625,
0.339111328125,
0.5087890625,
0.252685546875,
0.50732421875
],
[
0.3837890625,
0.49462890625,
0.418212890625,
0.49462890625,
0.418212890625,
0.50732421875,
0.3837890625,
0.50732421875
],
[
0.51123046875,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5126953125,
0.51123046875,
0.5126953125
],
[
0.264404296875,
0.50830078125,
0.330078125,
0.50830078125,
0.330078125,
0.51953125,
0.264404296875,
0.51953125
],
[
0.382080078125,
0.50732421875,
0.418212890625,
0.50732421875,
0.418212890625,
0.51953125,
0.382080078125,
0.51953125
],
[
0.509765625,
0.51416015625,
0.81201171875,
0.51318359375,
0.81201171875,
0.52783203125,
0.509765625,
0.529296875
],
[
0.385498046875,
0.51953125,
0.4189453125,
0.51953125,
0.4189453125,
0.5322265625,
0.385498046875,
0.5322265625
],
[
0.2568359375,
0.52099609375,
0.337646484375,
0.52099609375,
0.337646484375,
0.5322265625,
0.2568359375,
0.5322265625
],
[
0.166259765625,
0.5126953125,
0.2283935546875,
0.5126953125,
0.2283935546875,
0.52685546875,
0.166259765625,
0.52685546875
],
[
0.2476806640625,
0.52978515625,
0.343505859375,
0.5322265625,
0.343505859375,
0.54638671875,
0.2476806640625,
0.54443359375
],
[
0.387939453125,
0.5322265625,
0.414794921875,
0.5322265625,
0.414794921875,
0.54443359375,
0.387939453125,
0.54443359375
],
[
0.51123046875,
0.54052734375,
0.880859375,
0.54052734375,
0.880859375,
0.55419921875,
0.51123046875,
0.55419921875
],
[
0.276123046875,
0.5498046875,
0.316650390625,
0.5498046875,
0.316650390625,
0.5654296875,
0.276123046875,
0.5654296875
],
[
0.38037109375,
0.5498046875,
0.4189453125,
0.5498046875,
0.4189453125,
0.5654296875,
0.38037109375,
0.5654296875
],
[
0.509765625,
0.5556640625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57177734375,
0.509765625,
0.57080078125
],
[
0.25439453125,
0.568359375,
0.337646484375,
0.57080078125,
0.337646484375,
0.5849609375,
0.25439453125,
0.5830078125
],
[
0.387939453125,
0.57080078125,
0.414794921875,
0.57080078125,
0.414794921875,
0.5830078125,
0.387939453125,
0.5830078125
],
[
0.51318359375,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.51318359375,
0.58837890625
],
[
0.173828125,
0.58056640625,
0.2225341796875,
0.58056640625,
0.2225341796875,
0.595703125,
0.173828125,
0.595703125
],
[
0.2626953125,
0.58447265625,
0.330078125,
0.58447265625,
0.330078125,
0.595703125,
0.2626953125,
0.595703125
],
[
0.38623046875,
0.5830078125,
0.414794921875,
0.5830078125,
0.414794921875,
0.595703125,
0.38623046875,
0.595703125
],
[
0.38623046875,
0.595703125,
0.414794921875,
0.595703125,
0.414794921875,
0.60791015625,
0.38623046875,
0.60791015625
],
[
0.51123046875,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.60498046875,
0.51123046875,
0.60498046875
],
[
0.255126953125,
0.5966796875,
0.339111328125,
0.5966796875,
0.339111328125,
0.60791015625,
0.255126953125,
0.60791015625
],
[
0.3896484375,
0.60693359375,
0.4130859375,
0.60693359375,
0.4130859375,
0.62158203125,
0.3896484375,
0.62158203125
],
[
0.51123046875,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.62060546875,
0.51123046875,
0.62060546875
],
[
0.509765625,
0.62255859375,
0.7958984375,
0.62158203125,
0.7958984375,
0.6357421875,
0.509765625,
0.63720703125
],
[
0.2509765625,
0.609375,
0.341796875,
0.609375,
0.341796875,
0.62060546875,
0.2509765625,
0.62060546875
],
[
0.276123046875,
0.6259765625,
0.316650390625,
0.6259765625,
0.316650390625,
0.64111328125,
0.276123046875,
0.64111328125
],
[
0.385498046875,
0.626953125,
0.41650390625,
0.626953125,
0.41650390625,
0.64111328125,
0.385498046875,
0.64111328125
],
[
0.252685546875,
0.6435546875,
0.339111328125,
0.6455078125,
0.339111328125,
0.6611328125,
0.252685546875,
0.6591796875
],
[
0.387939453125,
0.6455078125,
0.414794921875,
0.6455078125,
0.414794921875,
0.6591796875,
0.387939453125,
0.6591796875
],
[
0.51318359375,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.6630859375,
0.51318359375,
0.6630859375
],
[
0.264404296875,
0.66015625,
0.330078125,
0.66015625,
0.330078125,
0.67138671875,
0.264404296875,
0.67138671875
],
[
0.387939453125,
0.65771484375,
0.414794921875,
0.65771484375,
0.414794921875,
0.67041015625,
0.387939453125,
0.67041015625
],
[
0.160400390625,
0.66455078125,
0.2359619140625,
0.66455078125,
0.2359619140625,
0.6787109375,
0.160400390625,
0.6787109375
],
[
0.51123046875,
0.66455078125,
0.8818359375,
0.6650390625,
0.8818359375,
0.6796875,
0.51123046875,
0.6787109375
],
[
0.255126953125,
0.67138671875,
0.3359375,
0.67138671875,
0.3359375,
0.68310546875,
0.255126953125,
0.68310546875
],
[
0.38623046875,
0.67041015625,
0.414794921875,
0.67041015625,
0.414794921875,
0.68408203125,
0.38623046875,
0.68408203125
],
[
0.509765625,
0.68115234375,
0.87890625,
0.68115234375,
0.87890625,
0.69482421875,
0.509765625,
0.69482421875
],
[
0.276123046875,
0.689453125,
0.316650390625,
0.689453125,
0.316650390625,
0.705078125,
0.276123046875,
0.705078125
],
[
0.3837890625,
0.689453125,
0.418212890625,
0.689453125,
0.418212890625,
0.705078125,
0.3837890625,
0.705078125
],
[
0.509765625,
0.69677734375,
0.880859375,
0.69775390625,
0.880859375,
0.7119140625,
0.509765625,
0.71142578125
],
[
0.139404296875,
0.71875,
0.462646484375,
0.71875,
0.462646484375,
0.732421875,
0.139404296875,
0.732421875
],
[
0.509765625,
0.71337890625,
0.87744140625,
0.71337890625,
0.87744140625,
0.72705078125,
0.509765625,
0.72705078125
],
[
0.509765625,
0.7392578125,
0.60693359375,
0.7412109375,
0.60693359375,
0.75732421875,
0.509765625,
0.7548828125
],
[
0.1158447265625,
0.75,
0.4130859375,
0.75,
0.4130859375,
0.76318359375,
0.1158447265625,
0.76318359375
],
[
0.509765625,
0.76318359375,
0.8818359375,
0.76416015625,
0.8818359375,
0.779296875,
0.509765625,
0.77783203125
],
[
0.1175537109375,
0.76953125,
0.48779296875,
0.76953125,
0.48779296875,
0.78369140625,
0.1175537109375,
0.78369140625
],
[
0.1158447265625,
0.7861328125,
0.490234375,
0.7861328125,
0.490234375,
0.7998046875,
0.1158447265625,
0.7998046875
],
[
0.51123046875,
0.7822265625,
0.6572265625,
0.7822265625,
0.6572265625,
0.7958984375,
0.51123046875,
0.7958984375
],
[
0.118408203125,
0.80419921875,
0.48779296875,
0.80419921875,
0.48779296875,
0.81396484375,
0.118408203125,
0.81396484375
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.81787109375,
0.48779296875,
0.81787109375,
0.48779296875,
0.83203125,
0.1158447265625,
0.83203125
],
[
0.509765625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.11419677734375,
0.83447265625,
0.484375,
0.83447265625,
0.484375,
0.84716796875,
0.11419677734375,
0.84716796875
],
[
0.509765625,
0.84033203125,
0.73486328125,
0.8388671875,
0.73486328125,
0.85400390625,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.85107421875,
0.486083984375,
0.85107421875,
0.486083984375,
0.8642578125,
0.1175537109375,
0.8642578125
],
[
0.1158447265625,
0.86767578125,
0.48779296875,
0.86767578125,
0.48779296875,
0.87744140625,
0.1158447265625,
0.87744140625
],
[
0.5322265625,
0.86767578125,
0.7734375,
0.86767578125,
0.7734375,
0.88134765625,
0.5322265625,
0.88134765625
],
[
0.1158447265625,
0.8828125,
0.489501953125,
0.8828125,
0.489501953125,
0.896484375,
0.1158447265625,
0.896484375
],
[
0.5322265625,
0.8798828125,
0.68701171875,
0.8798828125,
0.68701171875,
0.8935546875,
0.5322265625,
0.8935546875
],
[
0.5322265625,
0.89111328125,
0.77734375,
0.892578125,
0.77734375,
0.90673828125,
0.5322265625,
0.90625
],
[
0.1376953125,
0.90478515625,
0.476806640625,
0.90478515625,
0.476806640625,
0.91943359375,
0.1376953125,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.79833984375,
0.90478515625,
0.79833984375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"(e.g. the word 'list' in question makes the model",
"situation, we replace the fake situation with direct",
"design alist even though the underlying problem is",
"instructions. Full example is included in Appendix",
"on graphs). We further analyze model performance",
"C. Second, we create an expert summary of the",
"on different types of summaries (i.e., basic, expert,",
"problem.To create this,we further summarize",
"and synthetic) and provide instruction-design prin-",
" the first summary. This expert summary includes",
"ciples that can help future research on prompting",
"the absolute minimum information for an expert to",
"in program synthesis.",
"understand the problem. We would not expect a",
"novice to understand these prompts. An example",
"2Method",
"of expert summaries is given in Appendix C.3.",
"2.1",
"Dataset",
"2.1.2Synthetic Summaries",
"We use the APPS (Hendrycks et al., 2021) and",
"We have generated synthetic summaries of program",
"CodeContests (Li et al., 2022) datasets to create",
"descriptions using jumbo (178B), large (7.5B) Stu-",
" summaries. We crowd-sourced the creation of hu-",
"dio21 model (Lieber et al.. 2021), GPT-3 Davinci",
"man summaries. The result was 373 human sum-",
"model (175B) (Brown et al., 2020) and PEGASUS",
"maries for APPS and 80 summaries for CodeCon-",
"model (Zhang et al., 2019). To generate a summary,",
"tests along with and 8663 synthetic summaries us-",
"we provide these models with a few examples in the",
"ing both datasets.Table I shows the statistics of",
"in-context learning setup (Brown et al., 2020) from",
"the generated summaries.",
"the human-generated summaries. For the few-shot",
"Data Source",
"Difficulty",
"examples, we use expert-level summaries.",
"#ofProblems",
"Introductory",
"145",
"Studio21 We use five examples with the large",
"Interview",
"123",
"Human",
"model, and three examples with the jumbo mode16",
"Competition",
"105",
"CodeContests",
"For both models, we use a temperature of 0.3, and",
"80",
"topP of 1. For the format of our prompt, we use De",
"Total",
"453",
"Jargonizer template7 with a change to their header",
"Introductory",
"1588",
"as shown in Appendix D. We create a total of 7, 505",
"Interview",
"4551",
" synthetic summaries using these models.",
"1286",
"Competition",
"Studio21",
"CodeContests",
"80",
"GPT-3We use three examples for GPT-3 model",
"Total",
"7505",
"We empirically set temperature to 0.05, topP to",
"Introductory",
"194",
"1, frequency penalty to 0.01, presence penalty to",
"GPT-3",
"Interview",
"267",
"244",
"0.05. To generate prompts, we followed their tl;dr",
"Competition",
"80",
"template8 as shown in Appendix D. We create 785",
"synthetic summaries using this model.",
"CodeContests",
"Total",
"785",
"Introductory",
"145",
"PEGASUS We use the PEGASUS model (Zhang",
"Interview",
"123",
"PEGASUS",
"et al., 2019) to create program summaries for the",
"Competition",
"105",
" same set of problems that were summarized by",
"Total",
"373",
"humans. We choose this model because it was",
"Table l: Statistics of the proposed meta-dataset.",
"trained specifically for abstractive summarization",
"2.2 Model",
"2.1.1 Human Generated Summaries",
"We use OpenAI Codex to build baselines and the",
"For the APPS and CodeContests human-generated",
"summaries, the crowd worker reads and under-",
"proposed approach.",
"stands the original questions, then creates sum-",
"BaselineTo create a baseline, we have used orig",
"maries in two steps. First, we create a basic",
"inal program descriptions given in the datasets as",
" summary of the given problem and remove any",
"prompts for the Codex model.",
"information that is repeated and any hypothetical",
"information without concreteinstructions.For ex-",
"‘Examples are included in Appendix D",
"ample, if the problem constructs a fake company or",
"7https://studio.ai21.com/",
"8https:/beta.openai.com/playground/p/",
"'Instructions for creating summaries are in Appendix N",
"default-tldr-summary?model=text-davinci-001"
] |
{
"arxiv_id": "2203.08597",
"language": "en",
"page_idx": 1
}
|
||
2203.08597-en-2
|
2203.08597-en-2.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Difficulty</th>
<th colspan="3">$AP$</th>
<th colspan="3">$\mathbf{EWPR}$</th>
<th colspan="3">$\mathbf{BWPR}$</th>
</tr>
<tr>
<th>Baseline</th>
<th>Basic</th>
<th>Expert</th>
<th>Baseline 1</th>
<th>Basic</th>
<th>Expert</th>
<th>Baseline</th>
<th>Basic</th>
<th>Expert</th>
</tr>
<tr>
<td>Introductory</td>
<td>42.96</td>
<td>50.00</td>
<td>50.00</td>
<td>44.20</td>
<td>51.45</td>
<td>51.82</td>
<td>43.23</td>
<td>50.35</td>
<td>50.35</td>
</tr>
<tr>
<td>Interview</td>
<td>37.70</td>
<td>41.80</td>
<td>44.26</td>
<td>36.52</td>
<td>45.54</td>
<td>46.96</td>
<td>37.70</td>
<td>41.80</td>
<td>44.26</td>
</tr>
<tr>
<td>Competition</td>
<td>4.76</td>
<td>5.71</td>
<td>5.71</td>
<td>4.00</td>
<td>6.00</td>
<td>6.00</td>
<td>4.76</td>
<td>5.71</td>
<td>5.71</td>
</tr>
<tr>
<td>Weighted Average</td>
<td>30.47</td>
<td>34.83</td>
<td>35.64</td>
<td>30.31</td>
<td>36.65</td>
<td>37.22</td>
<td>30.43</td>
<td>34.78</td>
<td>35.59</td>
</tr>
<tr>
<td>CodeContests</td>
<td>12.50</td>
<td>23.75</td>
<td>25.00</td>
<td>13.33</td>
<td>25.33</td>
<td>26.66</td>
<td>12.82</td>
<td>24.36</td>
<td>25.64</td>
</tr>
</tbody>
</table>
Table 2: Results of baseline and proposed model in terms of Strict Accuracy (SAcc). The first block is from the APPS dataset. The last block is from the CodeContests dataset. AP: All Problems, EWPR: Either Worst Problem Removal, BWPR: Both Worst Problem Removal (see explanation in section 3). All results are in $\%.$ Weighted Average is not shown for CodeContests because similar difficulties were not provided (see explanation in 4.1).
<table>
<tbody>
<tr>
<th rowspan="2">Difficulty</th>
<th colspan="2">$AP$</th>
<th colspan="2">$\mathbf{EWPR}$</th>
</tr>
<tr>
<th>Baseline</th>
<th>Proposed</th>
<th>Baseline</th>
<th>Proposed</th>
</tr>
<tr>
<td>Introductorv</td>
<td>42.96</td>
<td>52.82</td>
<td>44.53</td>
<td>54.74</td>
</tr>
<tr>
<td>Interview</td>
<td>37.70</td>
<td>49.18</td>
<td>38.66</td>
<td>50.42</td>
</tr>
<tr>
<td>Competition</td>
<td>4.76</td>
<td>6.67</td>
<td>4.81</td>
<td>6.73</td>
</tr>
<tr>
<td>Weighted Average</td>
<td>30.47</td>
<td>38.48</td>
<td>31.11</td>
<td>39.44</td>
</tr>
</tbody>
</table>
1. All problems from basic and expert sum-
maries are used at inference time. We term
this experiment All Problems (AP).
2. We eliminate problems that perform worse$^{11}$
for either basic or expert summaries. We term this experiment Either Worst Problem Removal (EWPR).
3. We eliminate problems that perform worse for
both basic and expert summaries. We term this experiment Both Worst Problem Removal $(\mathbf{BWPR}).$
Table 3: Results when taking the best summary for each problem. The EWPR baseline is different from Table 2 because a different set of problems have been removed.
Proposed Approach We have used summaries of original program descriptions given in the datasets as prompts for the Codex model.
Motivation behind EWPR and BWPR If a
# 3 Experimental Setup
summary caused every test case to perform worse then it's likely the crowd worker produced a faulty summary. To mitigate the effect of outliers in the dataset, we use the EWPR method to remove such problems. Another hypothesis is that every problem benefits from some level of summarization (i.e., basic or expert). To measure this, we use the BWPR method. From Table 6 results, we identify that only 1 problem had both summaries (basic and expert) preform worse.
All the experiments are performed using the $davinci-codex$(Chen et al., 2021) model provided through OpenAI$^9.$ At inference time, we use a modifed version of the evaluation code$^{10}$ provided by Hendrycks et al. (2021). This evaluation code has four different outputs for each test case: (1) -2: the code has a syntax error and can not run, (2) -1: the code is syntactically correct but has a run time error, (3) 0: the code runs without any errors but fails the test case, and (4) 1: the code runs without any error and passes the test case. Similar to Chen et al. (2021), we implement a timeout for the code at inference time. If a test case takes more than 4 seconds to run then we throw an exception and count that test case as a -1.
Metric In(Austin et al.,2021a), they show that the BLEU metric (Papineni et al., 2002) does not correlate well with synthesis performance. Thus, we use Strict Accuracy (SAcc) as our evaluation metric for all experiments (see Appendix E).
### 4 Results and Analysis
4.1 Human Generated Summaries
## Experiments To show effectiveness of the proposed approach, we have performed three different experiments using human generated summaries:
From Table 2, we can observe that both the summary-based models show on average superior performance compared to baseline. In particular, when calculating results for every problem,
$^{9}$Implementation and parameters details in Appendix F $^{10}$https://github.com/hendrycks/apps/blob/main/eval/test\_
one\_solution.py
$^{11}$Definition of the worst problem is given in Appendix G
|
[
[
0.5439453125,
0.0872802734375,
0.59765625,
0.0872802734375,
0.59765625,
0.1015625,
0.5439453125,
0.1015625
],
[
0.73486328125,
0.0872802734375,
0.78857421875,
0.0872802734375,
0.78857421875,
0.1015625,
0.73486328125,
0.1015625
],
[
0.177978515625,
0.09442138671875,
0.2476806640625,
0.09442138671875,
0.2476806640625,
0.1092529296875,
0.177978515625,
0.1092529296875
],
[
0.366943359375,
0.0885009765625,
0.393798828125,
0.0885009765625,
0.393798828125,
0.1015625,
0.366943359375,
0.1015625
],
[
0.28955078125,
0.10809326171875,
0.358642578125,
0.10809326171875,
0.358642578125,
0.122314453125,
0.28955078125,
0.122314453125
],
[
0.361083984375,
0.10809326171875,
0.411376953125,
0.10809326171875,
0.411376953125,
0.122314453125,
0.361083984375,
0.122314453125
],
[
0.41650390625,
0.10687255859375,
0.476806640625,
0.1092529296875,
0.47607421875,
0.12469482421875,
0.414794921875,
0.122314453125
],
[
0.5517578125,
0.10809326171875,
0.60205078125,
0.10809326171875,
0.60205078125,
0.122314453125,
0.5517578125,
0.122314453125
],
[
0.60693359375,
0.1092529296875,
0.66748046875,
0.1092529296875,
0.66748046875,
0.12353515625,
0.60693359375,
0.12353515625
],
[
0.7421875,
0.10809326171875,
0.79443359375,
0.10809326171875,
0.79443359375,
0.122314453125,
0.7421875,
0.122314453125
],
[
0.796875,
0.1092529296875,
0.85205078125,
0.1092529296875,
0.85205078125,
0.12353515625,
0.796875,
0.12353515625
],
[
0.66650390625,
0.10986328125,
0.73779296875,
0.10986328125,
0.73779296875,
0.1217041015625,
0.66650390625,
0.1217041015625
],
[
0.47607421875,
0.10986328125,
0.54736328125,
0.10986328125,
0.54736328125,
0.1217041015625,
0.47607421875,
0.1217041015625
],
[
0.300537109375,
0.1270751953125,
0.344970703125,
0.12890625,
0.343505859375,
0.144287109375,
0.298828125,
0.142578125
],
[
0.4189453125,
0.127685546875,
0.468505859375,
0.1300048828125,
0.466796875,
0.1455078125,
0.418212890625,
0.1436767578125
],
[
0.489501953125,
0.1270751953125,
0.53662109375,
0.12890625,
0.53564453125,
0.144287109375,
0.48779296875,
0.142578125
],
[
0.74462890625,
0.127685546875,
0.79248046875,
0.1300048828125,
0.791015625,
0.1455078125,
0.744140625,
0.1436767578125
],
[
0.68115234375,
0.12890625,
0.72705078125,
0.12890625,
0.72705078125,
0.1436767578125,
0.68115234375,
0.1436767578125
],
[
0.166259765625,
0.12890625,
0.258544921875,
0.1312255859375,
0.258544921875,
0.1455078125,
0.166259765625,
0.1436767578125
],
[
0.36279296875,
0.1300048828125,
0.41064453125,
0.1300048828125,
0.41064453125,
0.144287109375,
0.36279296875,
0.144287109375
],
[
0.55322265625,
0.1300048828125,
0.60107421875,
0.1300048828125,
0.60107421875,
0.144287109375,
0.55322265625,
0.144287109375
],
[
0.611328125,
0.1300048828125,
0.6591796875,
0.1300048828125,
0.6591796875,
0.144287109375,
0.611328125,
0.144287109375
],
[
0.8037109375,
0.12890625,
0.8486328125,
0.12890625,
0.8486328125,
0.144287109375,
0.8037109375,
0.144287109375
],
[
0.298828125,
0.1436767578125,
0.344970703125,
0.1436767578125,
0.344970703125,
0.157958984375,
0.298828125,
0.157958984375
],
[
0.36279296875,
0.1436767578125,
0.411376953125,
0.1436767578125,
0.411376953125,
0.157958984375,
0.36279296875,
0.157958984375
],
[
0.4189453125,
0.1436767578125,
0.466796875,
0.1436767578125,
0.466796875,
0.157958984375,
0.4189453125,
0.157958984375
],
[
0.489501953125,
0.1436767578125,
0.53564453125,
0.1436767578125,
0.53564453125,
0.157958984375,
0.489501953125,
0.157958984375
],
[
0.55322265625,
0.1436767578125,
0.60107421875,
0.1436767578125,
0.60107421875,
0.157958984375,
0.55322265625,
0.157958984375
],
[
0.609375,
0.1436767578125,
0.6572265625,
0.1436767578125,
0.6572265625,
0.157958984375,
0.609375,
0.157958984375
],
[
0.67919921875,
0.1436767578125,
0.72705078125,
0.1436767578125,
0.72705078125,
0.157958984375,
0.67919921875,
0.157958984375
],
[
0.74462890625,
0.1436767578125,
0.79248046875,
0.1436767578125,
0.79248046875,
0.157958984375,
0.74462890625,
0.157958984375
],
[
0.8037109375,
0.1436767578125,
0.8486328125,
0.1436767578125,
0.8486328125,
0.157958984375,
0.8037109375,
0.157958984375
],
[
0.748046875,
0.15673828125,
0.78662109375,
0.15673828125,
0.78662109375,
0.1728515625,
0.748046875,
0.1728515625
],
[
0.177978515625,
0.1455078125,
0.246826171875,
0.1455078125,
0.246826171875,
0.15673828125,
0.177978515625,
0.15673828125
],
[
0.30322265625,
0.157958984375,
0.339111328125,
0.157958984375,
0.339111328125,
0.17041015625,
0.30322265625,
0.17041015625
],
[
0.366943359375,
0.157958984375,
0.40478515625,
0.157958984375,
0.40478515625,
0.17041015625,
0.366943359375,
0.17041015625
],
[
0.4248046875,
0.157958984375,
0.4609375,
0.157958984375,
0.4609375,
0.17041015625,
0.4248046875,
0.17041015625
],
[
0.49365234375,
0.157958984375,
0.53076171875,
0.157958984375,
0.53076171875,
0.17041015625,
0.49365234375,
0.17041015625
],
[
0.55908203125,
0.157958984375,
0.59619140625,
0.157958984375,
0.59619140625,
0.17041015625,
0.55908203125,
0.17041015625
],
[
0.6171875,
0.157958984375,
0.6533203125,
0.157958984375,
0.6533203125,
0.17041015625,
0.6171875,
0.17041015625
],
[
0.80615234375,
0.15673828125,
0.84619140625,
0.15673828125,
0.84619140625,
0.1728515625,
0.80615234375,
0.1728515625
],
[
0.68408203125,
0.1591796875,
0.72119140625,
0.1591796875,
0.72119140625,
0.171630859375,
0.68408203125,
0.171630859375
],
[
0.1695556640625,
0.1591796875,
0.258544921875,
0.1591796875,
0.258544921875,
0.17041015625,
0.1695556640625,
0.17041015625
],
[
0.145263671875,
0.1776123046875,
0.279541015625,
0.1787109375,
0.279541015625,
0.193603515625,
0.145263671875,
0.1923828125
],
[
0.36279296875,
0.177001953125,
0.41064453125,
0.1787109375,
0.408935546875,
0.19482421875,
0.361083984375,
0.1923828125
],
[
0.420654296875,
0.1776123046875,
0.47021484375,
0.1776123046875,
0.47021484375,
0.193603515625,
0.420654296875,
0.193603515625
],
[
0.490234375,
0.1776123046875,
0.5341796875,
0.1776123046875,
0.5341796875,
0.193603515625,
0.490234375,
0.193603515625
],
[
0.55517578125,
0.1776123046875,
0.60107421875,
0.1776123046875,
0.60107421875,
0.193603515625,
0.55517578125,
0.193603515625
],
[
0.611328125,
0.1776123046875,
0.6591796875,
0.1776123046875,
0.6591796875,
0.193603515625,
0.611328125,
0.193603515625
],
[
0.74462890625,
0.1776123046875,
0.791015625,
0.1776123046875,
0.791015625,
0.193603515625,
0.74462890625,
0.193603515625
],
[
0.8037109375,
0.1776123046875,
0.8486328125,
0.1776123046875,
0.8486328125,
0.193603515625,
0.8037109375,
0.193603515625
],
[
0.68115234375,
0.1787109375,
0.72705078125,
0.1787109375,
0.72705078125,
0.193603515625,
0.68115234375,
0.193603515625
],
[
0.297119140625,
0.1787109375,
0.341796875,
0.177001953125,
0.343505859375,
0.1923828125,
0.298828125,
0.19482421875
],
[
0.298828125,
0.1995849609375,
0.344970703125,
0.1995849609375,
0.344970703125,
0.21435546875,
0.298828125,
0.21435546875
],
[
0.36279296875,
0.1995849609375,
0.41064453125,
0.1995849609375,
0.41064453125,
0.21435546875,
0.36279296875,
0.21435546875
],
[
0.4189453125,
0.1995849609375,
0.470947265625,
0.1995849609375,
0.470947265625,
0.21435546875,
0.4189453125,
0.21435546875
],
[
0.490234375,
0.1995849609375,
0.53564453125,
0.1995849609375,
0.53564453125,
0.21435546875,
0.490234375,
0.21435546875
],
[
0.55517578125,
0.1995849609375,
0.60107421875,
0.1995849609375,
0.60107421875,
0.21435546875,
0.55517578125,
0.21435546875
],
[
0.611328125,
0.1995849609375,
0.66162109375,
0.1995849609375,
0.66162109375,
0.21435546875,
0.611328125,
0.21435546875
],
[
0.6826171875,
0.1995849609375,
0.72705078125,
0.1995849609375,
0.72705078125,
0.21435546875,
0.6826171875,
0.21435546875
],
[
0.74462890625,
0.1995849609375,
0.791015625,
0.1995849609375,
0.791015625,
0.21435546875,
0.74462890625,
0.21435546875
],
[
0.8037109375,
0.1995849609375,
0.8505859375,
0.1995849609375,
0.8505859375,
0.21435546875,
0.8037109375,
0.21435546875
],
[
0.468505859375,
0.204833984375,
0.47265625,
0.204833984375,
0.47265625,
0.21142578125,
0.468505859375,
0.21142578125
],
[
0.16455078125,
0.201904296875,
0.26123046875,
0.201904296875,
0.26123046875,
0.213134765625,
0.16455078125,
0.213134765625
],
[
0.1158447265625,
0.2298583984375,
0.8818359375,
0.2298583984375,
0.8818359375,
0.243408203125,
0.1158447265625,
0.243408203125
],
[
0.1175537109375,
0.24462890625,
0.8818359375,
0.24462890625,
0.8818359375,
0.25830078125,
0.1175537109375,
0.25830078125
],
[
0.1175537109375,
0.2587890625,
0.8818359375,
0.2587890625,
0.8818359375,
0.2724609375,
0.1175537109375,
0.2724609375
],
[
0.1158447265625,
0.2724609375,
0.85546875,
0.2724609375,
0.85546875,
0.2861328125,
0.1158447265625,
0.2861328125
],
[
0.397216796875,
0.312255859375,
0.444091796875,
0.312255859375,
0.444091796875,
0.32666015625,
0.397216796875,
0.32666015625
],
[
0.52490234375,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.32666015625,
0.52490234375,
0.32666015625
],
[
0.285400390625,
0.3134765625,
0.310546875,
0.3134765625,
0.310546875,
0.32666015625,
0.285400390625,
0.32666015625
],
[
0.1485595703125,
0.31640625,
0.208251953125,
0.318359375,
0.20654296875,
0.333251953125,
0.14697265625,
0.330810546875
],
[
0.2376708984375,
0.330078125,
0.480224609375,
0.330078125,
0.480224609375,
0.34326171875,
0.2376708984375,
0.34326171875
],
[
0.54736328125,
0.330078125,
0.8818359375,
0.330078125,
0.8818359375,
0.34326171875,
0.54736328125,
0.34326171875
],
[
0.139404296875,
0.344482421875,
0.2166748046875,
0.3466796875,
0.2166748046875,
0.361083984375,
0.139404296875,
0.359375
],
[
0.2476806640625,
0.347412109375,
0.28369140625,
0.347412109375,
0.28369140625,
0.359375,
0.2476806640625,
0.359375
],
[
0.307373046875,
0.347412109375,
0.344970703125,
0.347412109375,
0.344970703125,
0.359375,
0.307373046875,
0.359375
],
[
0.36865234375,
0.347412109375,
0.4072265625,
0.347412109375,
0.4072265625,
0.359375,
0.36865234375,
0.359375
],
[
0.432373046875,
0.347412109375,
0.468505859375,
0.347412109375,
0.468505859375,
0.359375,
0.432373046875,
0.359375
],
[
0.5458984375,
0.344482421875,
0.8095703125,
0.344482421875,
0.8095703125,
0.359375,
0.5458984375,
0.359375
],
[
0.246826171875,
0.358154296875,
0.285400390625,
0.358154296875,
0.285400390625,
0.37060546875,
0.246826171875,
0.37060546875
],
[
0.306396484375,
0.358154296875,
0.344970703125,
0.358154296875,
0.344970703125,
0.37060546875,
0.306396484375,
0.37060546875
],
[
0.370361328125,
0.358154296875,
0.4072265625,
0.358154296875,
0.4072265625,
0.37060546875,
0.370361328125,
0.37060546875
],
[
0.4306640625,
0.358154296875,
0.468505859375,
0.358154296875,
0.468505859375,
0.37060546875,
0.4306640625,
0.37060546875
],
[
0.5224609375,
0.3623046875,
0.8818359375,
0.35986328125,
0.8818359375,
0.37890625,
0.5224609375,
0.38134765625
],
[
0.2476806640625,
0.369384765625,
0.280517578125,
0.369384765625,
0.280517578125,
0.3818359375,
0.2476806640625,
0.3818359375
],
[
0.1485595703125,
0.359375,
0.208251953125,
0.359375,
0.208251953125,
0.37060546875,
0.1485595703125,
0.37060546875
],
[
0.30908203125,
0.369384765625,
0.341796875,
0.369384765625,
0.341796875,
0.3818359375,
0.30908203125,
0.3818359375
],
[
0.3720703125,
0.369384765625,
0.40478515625,
0.369384765625,
0.40478515625,
0.3818359375,
0.3720703125,
0.3818359375
],
[
0.43408203125,
0.369384765625,
0.466796875,
0.369384765625,
0.466796875,
0.3818359375,
0.43408203125,
0.3818359375
],
[
0.54833984375,
0.38134765625,
0.8818359375,
0.38134765625,
0.8818359375,
0.394287109375,
0.54833984375,
0.394287109375
],
[
0.1427001953125,
0.37060546875,
0.2149658203125,
0.37060546875,
0.2149658203125,
0.3818359375,
0.1427001953125,
0.3818359375
],
[
0.12176513671875,
0.385986328125,
0.2301025390625,
0.38720703125,
0.2301025390625,
0.3984375,
0.12176513671875,
0.39794921875
],
[
0.2476806640625,
0.385986328125,
0.2822265625,
0.385986328125,
0.2822265625,
0.3984375,
0.2476806640625,
0.3984375
],
[
0.307373046875,
0.385986328125,
0.344970703125,
0.385986328125,
0.344970703125,
0.39794921875,
0.307373046875,
0.39794921875
],
[
0.370361328125,
0.385986328125,
0.405517578125,
0.385986328125,
0.405517578125,
0.3984375,
0.370361328125,
0.3984375
],
[
0.432373046875,
0.385986328125,
0.47021484375,
0.385986328125,
0.47021484375,
0.39794921875,
0.432373046875,
0.39794921875
],
[
0.54736328125,
0.396728515625,
0.8818359375,
0.3955078125,
0.8818359375,
0.410400390625,
0.54736328125,
0.410888671875
],
[
0.1158447265625,
0.413330078125,
0.486083984375,
0.413330078125,
0.486083984375,
0.427001953125,
0.1158447265625,
0.427001953125
],
[
0.54833984375,
0.413330078125,
0.6884765625,
0.413330078125,
0.6884765625,
0.427001953125,
0.54833984375,
0.427001953125
],
[
0.1158447265625,
0.427490234375,
0.486083984375,
0.427001953125,
0.486083984375,
0.440673828125,
0.1158447265625,
0.441162109375
],
[
0.52490234375,
0.429931640625,
0.8818359375,
0.43115234375,
0.8818359375,
0.4453125,
0.52490234375,
0.44482421875
],
[
0.1175537109375,
0.443603515625,
0.48193359375,
0.443603515625,
0.48193359375,
0.45361328125,
0.1175537109375,
0.45361328125
],
[
0.5458984375,
0.44775390625,
0.880859375,
0.44775390625,
0.880859375,
0.4619140625,
0.5458984375,
0.4619140625
],
[
0.54833984375,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.54833984375,
0.47802734375
],
[
0.1175537109375,
0.482666015625,
0.486083984375,
0.482666015625,
0.486083984375,
0.49755859375,
0.1175537109375,
0.49755859375
],
[
0.54736328125,
0.478515625,
0.62060546875,
0.478515625,
0.62060546875,
0.493408203125,
0.54736328125,
0.493408203125
],
[
0.1175537109375,
0.49951171875,
0.48779296875,
0.49951171875,
0.48779296875,
0.51318359375,
0.1175537109375,
0.51318359375
],
[
0.51123046875,
0.5048828125,
0.83642578125,
0.5048828125,
0.83642578125,
0.5185546875,
0.51123046875,
0.5185546875
],
[
0.82861328125,
0.50732421875,
0.880859375,
0.50732421875,
0.880859375,
0.51708984375,
0.82861328125,
0.51708984375
],
[
0.1158447265625,
0.515625,
0.42236328125,
0.51416015625,
0.42236328125,
0.529296875,
0.1158447265625,
0.52978515625
],
[
0.51123046875,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.5361328125,
0.51123046875,
0.5361328125
],
[
0.11419677734375,
0.5419921875,
0.322509765625,
0.54443359375,
0.322509765625,
0.55859375,
0.11419677734375,
0.55712890625
],
[
0.51123046875,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.51123046875,
0.5517578125
],
[
0.51123046875,
0.5546875,
0.8818359375,
0.5546875,
0.8818359375,
0.568359375,
0.51123046875,
0.568359375
],
[
0.1175537109375,
0.57080078125,
0.489501953125,
0.57080078125,
0.489501953125,
0.58447265625,
0.1175537109375,
0.58447265625
],
[
0.51123046875,
0.5693359375,
0.8818359375,
0.5693359375,
0.8818359375,
0.58447265625,
0.51123046875,
0.58447265625
],
[
0.1175537109375,
0.58740234375,
0.491943359375,
0.58740234375,
0.491943359375,
0.60107421875,
0.1175537109375,
0.60107421875
],
[
0.509765625,
0.5859375,
0.8818359375,
0.5849609375,
0.8818359375,
0.599609375,
0.509765625,
0.60107421875
],
[
0.1175537109375,
0.6025390625,
0.48779296875,
0.6025390625,
0.48779296875,
0.6162109375,
0.1175537109375,
0.6162109375
],
[
0.51123046875,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.51123046875,
0.6162109375
],
[
0.1158447265625,
0.6181640625,
0.489501953125,
0.6181640625,
0.489501953125,
0.6318359375,
0.1158447265625,
0.6318359375
],
[
0.51123046875,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.51123046875,
0.6318359375
],
[
0.1175537109375,
0.6357421875,
0.483642578125,
0.6357421875,
0.483642578125,
0.646484375,
0.1175537109375,
0.646484375
],
[
0.509765625,
0.6328125,
0.880859375,
0.63525390625,
0.880859375,
0.6484375,
0.509765625,
0.646484375
],
[
0.1175537109375,
0.65185546875,
0.489501953125,
0.65185546875,
0.489501953125,
0.6650390625,
0.1175537109375,
0.6650390625
],
[
0.51123046875,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.662109375,
0.51123046875,
0.662109375
],
[
0.1158447265625,
0.66650390625,
0.489501953125,
0.66845703125,
0.489501953125,
0.681640625,
0.1158447265625,
0.6796875
],
[
0.51123046875,
0.66748046875,
0.6826171875,
0.6650390625,
0.6826171875,
0.6787109375,
0.51123046875,
0.68115234375
],
[
0.12005615234375,
0.68505859375,
0.486083984375,
0.68505859375,
0.486083984375,
0.6953125,
0.12005615234375,
0.6953125
],
[
0.51123046875,
0.69140625,
0.880859375,
0.69140625,
0.880859375,
0.705078125,
0.51123046875,
0.705078125
],
[
0.1175537109375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.7119140625,
0.1175537109375,
0.7119140625
],
[
0.509765625,
0.708984375,
0.8818359375,
0.708984375,
0.8818359375,
0.72265625,
0.509765625,
0.72265625
],
[
0.1175537109375,
0.71728515625,
0.486083984375,
0.71728515625,
0.486083984375,
0.72802734375,
0.1175537109375,
0.72802734375
],
[
0.51123046875,
0.724609375,
0.88330078125,
0.724609375,
0.88330078125,
0.73828125,
0.51123046875,
0.73828125
],
[
0.11419677734375,
0.73095703125,
0.486083984375,
0.73095703125,
0.486083984375,
0.74462890625,
0.11419677734375,
0.74462890625
],
[
0.11419677734375,
0.74658203125,
0.489501953125,
0.74755859375,
0.489501953125,
0.7626953125,
0.11419677734375,
0.76123046875
],
[
0.509765625,
0.7412109375,
0.8818359375,
0.7412109375,
0.8818359375,
0.7548828125,
0.509765625,
0.7548828125
],
[
0.51123046875,
0.75732421875,
0.84033203125,
0.75732421875,
0.84033203125,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.1158447265625,
0.76416015625,
0.48779296875,
0.76416015625,
0.48779296875,
0.77783203125,
0.1158447265625,
0.77783203125
],
[
0.11419677734375,
0.779296875,
0.48779296875,
0.77978515625,
0.48779296875,
0.79443359375,
0.11419677734375,
0.79345703125
],
[
0.51708984375,
0.7890625,
0.52294921875,
0.7890625,
0.52294921875,
0.79443359375,
0.51708984375,
0.79443359375
],
[
0.5380859375,
0.78369140625,
0.72119140625,
0.78369140625,
0.72119140625,
0.79736328125,
0.5380859375,
0.79736328125
],
[
0.1158447265625,
0.79638671875,
0.355224609375,
0.79638671875,
0.355224609375,
0.81005859375,
0.1158447265625,
0.81005859375
],
[
0.51318359375,
0.81103515625,
0.7958984375,
0.81103515625,
0.7958984375,
0.82470703125,
0.51318359375,
0.82470703125
],
[
0.1158447265625,
0.82177734375,
0.48779296875,
0.822265625,
0.48779296875,
0.83740234375,
0.1158447265625,
0.8359375
],
[
0.51318359375,
0.83203125,
0.8818359375,
0.83203125,
0.8818359375,
0.845703125,
0.51318359375,
0.845703125
],
[
0.1158447265625,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.8525390625,
0.1158447265625,
0.8525390625
],
[
0.51123046875,
0.8486328125,
0.8818359375,
0.8486328125,
0.8818359375,
0.8623046875,
0.51123046875,
0.8623046875
],
[
0.1175537109375,
0.85498046875,
0.470947265625,
0.85498046875,
0.470947265625,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.51123046875,
0.8642578125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.51123046875,
0.87890625
],
[
0.1376953125,
0.8798828125,
0.47021484375,
0.8798828125,
0.47021484375,
0.8935546875,
0.1376953125,
0.8935546875
],
[
0.51123046875,
0.88134765625,
0.8818359375,
0.88134765625,
0.8818359375,
0.89453125,
0.51123046875,
0.89453125
],
[
0.1318359375,
0.89111328125,
0.484375,
0.892578125,
0.484375,
0.90673828125,
0.1318359375,
0.90625
],
[
0.11419677734375,
0.90478515625,
0.2166748046875,
0.90673828125,
0.2166748046875,
0.91845703125,
0.11419677734375,
0.91650390625
],
[
0.5283203125,
0.90283203125,
0.87744140625,
0.90478515625,
0.87744140625,
0.91943359375,
0.5283203125,
0.91748046875
]
] |
[
"EWPR",
"BWPR",
"Difficulty",
"AP",
"Baseline",
"Basic",
"Expert",
"Basic",
"Expert",
"Basic",
"Expert",
"Baseline",
"Baseline",
"42.96",
"50.00",
"44.20",
"50.35",
"43.23",
"Introductory",
"50.00",
"51.45",
"51.82",
"50.35",
"37.70",
"41.80",
"44.26",
"36.52",
"45.54",
"46.96",
"37.70",
"41.80",
"44.26",
"5.71",
"Interview",
"4.76",
"5.71",
"5.71",
"4.00",
"6.00",
"6.00",
"5.71",
"4.76",
"Competition",
"Weighted Average",
"34.83",
"35.64",
"30.31",
"36.65",
"37.22",
"34.78",
"35.59",
"30.43",
"30.47",
"12.50",
"23.75",
"25.00",
"13.33",
"25.33",
"26.66",
"12.82",
"24.36",
"25.64",
"",
"CodeContests",
"Table 2: Results of baseline and proposed model in terms of Strict Accuracy (SAcc). The first block is from the",
"APPS dataset. The last block is from the CodeContests dataset. AP: All Problems, EWPR: Either Worst Problem",
"Removal, BWPR: Both Worst Problem Removal (see explanation in section 3). All results are in %. Weighted",
"Average is not shown for CodeContests because similar difficulties were not provided (see explanation in 4.1).",
"EWPR",
"1. All problems from basic and expert sum-",
"AP",
"Difficulty",
"Baseline ProposedBaseline Proposed",
"maries are used at inference time. We term",
"Introductory",
"42.96",
"52.82",
"44.53",
"54.74",
"this experiment All Problems (AP).",
"37.70",
"49.18",
"38.66",
"50.42",
"2. We eliminate problems that perform worsell",
"4.76",
"Interview",
"6.67",
"4.81",
"6.73",
"for either basic or expert summaries. We",
"Competition",
"WeightedAverage",
"30.47",
"38.48",
"31.11",
"39.44",
"term this experiment Either Worst Problem",
"Table 3: Results when taking the best summary for each",
"Removal (EWPR).",
"problem. The EWPR baseline is different from Table 2",
"3.We eliminate problems that perform worsefor",
"becauseadifferentsetofproblemshavebeenremoved",
"both basic and expert summaries. We term",
"this experiment Both Worst Problem Removal",
"Proposed Approach We have used summaries",
"(BWPR).",
"of original program descriptions given in the",
"Motivation behind EWPR and BWPR",
"Ifa",
"datasets as prompts for the Codex model.",
"summary caused every test case to perform worse",
"3Experimental Setup",
"then it's likely the crowd worker produced a faulty",
"summary. To mitigate the effect of outliers in the",
"All the experiments are performed using the",
"dataset, we use the EWPR method to remove such",
"davinci - codec (Chen et al., 2021) model pro-",
"problems. Another hypothesis is that every prob-",
"vided through OpenAr°. At inference time, we use",
"lem benefits from some level of summarization",
"a modified version of the evaluation codelo pro-",
"(i.e., basic or expert). To measure this, we use the",
"vided by Hendrvcks et al. (2021). This evaluation",
"BWPR method.From Table 6 results, we identify",
"code has four different outputs for each test case:",
"thatonlylproblemhadbothsummaries(basicand",
"(1)-2: the code has a syntax error and can not run,",
"expert)preform worse.",
"(2)-l:thecodeissyntacticallycorrectbuthasa",
"Metric In (Austin et al., 202la), they show that",
"runtimeerror,(3)O:thecoderunswithoutanyer-",
"the BLEU metric (Papineni et al., 2002) does not",
"rors but fails the test case, and (4) 1: the code runs",
"correlate well with synthesis performance. Thus,",
"without any error and passes the test case. Similar",
"to Chen et al. (2021), we implement a timeout for",
"we use Strict Accuracy (SAcc) as our evaluation",
"metric for all experiments (see Appendix E).",
"the code at inference time. If a test case takes more",
"than 4 seconds to run then we throw an exception",
"",
"Results and Analysis",
"and count that test case as a -1.",
"4.1 Human Generated Summaries",
"ExperimentsTo show effectiveness of the pro-",
"From Table 2, we can observe that both the",
"posed approach, we have performed three different",
"summary-based models show on average superior",
"experiments using human generated summaries:",
"performance compared to baseline.In particu-",
"'Implementation and parameters details in Appendix F",
"lar, when calculating results for every problem,",
"1°https://github.com/hendrycks/apps/blob/main/eval/test",
"one_solution.py",
"I\"Definition of the worst problem is given in Appendix G"
] |
{
"arxiv_id": "2203.08597",
"language": "en",
"page_idx": 2
}
|
||
2203.08597-en-3
|
2203.08597-en-3.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th rowspan="2">Difficulty</th>
<th colspan="2">$AP$</th>
<th colspan="2">$\mathbf{EWPR}$</th>
</tr>
<tr>
<th>Baseline</th>
<th>$\operatorname{Proposed}$</th>
<th>Baseline</th>
<th>$\operatorname{Proposed}$</th>
</tr>
<tr>
<td rowspan="3">GPT- 3</td>
<td>Introductory</td>
<td>41.75</td>
<td>38.66</td>
<td>41.11</td>
<td>41.67</td>
</tr>
<tr>
<td>Interview</td>
<td>20.30</td>
<td>18.80</td>
<td>18.18</td>
<td>20.66</td>
</tr>
<tr>
<td>Competition</td>
<td>2.87</td>
<td>3.28</td>
<td>2.73</td>
<td>3.64</td>
</tr>
<tr>
<td> </td>
<td>Weighted Average</td>
<td>20.17</td>
<td>18.89</td>
<td>19.14</td>
<td>20.55</td>
</tr>
<tr>
<td rowspan="3">Studio21</td>
<td>Introductory</td>
<td>39.53</td>
<td>31.63</td>
<td>39.04</td>
<td>36.36</td>
</tr>
<tr>
<td>Interview</td>
<td>12.28</td>
<td>11.00</td>
<td>10.57</td>
<td>12.37</td>
</tr>
<tr>
<td>Competition</td>
<td>1.67</td>
<td>1.21</td>
<td>1.38</td>
<td>1.38</td>
</tr>
<tr>
<td> </td>
<td>Weighted Average</td>
<td>11.53</td>
<td>9.66</td>
<td>10.61</td>
<td>10.98</td>
</tr>
<tr>
<td> </td>
<td>Introductory</td>
<td>42.96</td>
<td>34.48</td>
<td>44.26</td>
<td>40.98</td>
</tr>
<tr>
<td>PEGASUS</td>
<td>Interview</td>
<td>37.70</td>
<td>10.57</td>
<td>14.29</td>
<td>21.56</td>
</tr>
<tr>
<td> </td>
<td>Competition</td>
<td>4.76</td>
<td>0.00</td>
<td>2.76</td>
<td>0.00</td>
</tr>
<tr>
<td> </td>
<td>Weighted Average</td>
<td>30.47</td>
<td>16.88</td>
<td>25.50</td>
<td>24.73</td>
</tr>
</tbody>
</table>
Table 4: Results of baseline and proposed approach (All results are in %). Summaries generated by GPT-3,
Studio21, and PEGASUS used for inference from APPS.
each problem for APPS dataset. We observe a $9.86\%,11.48\%$ and 1.91% increase on SAcc for introductory, interview, and competition level problems, respectively.
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th colspan="2">$AP$</th>
<th colspan="2">$\mathbf{EWPR}$</th>
</tr>
<tr>
<th>Baseline</th>
<th>$\operatorname{Proposed}$</th>
<th>Baseline</th>
<th>$\operatorname{Proposed}$</th>
</tr>
<tr>
<td>GPT- 3</td>
<td>12.50</td>
<td>10.0</td>
<td>18.75</td>
<td>18.75</td>
</tr>
<tr>
<td>Studio21</td>
<td>12.50</td>
<td>8.75</td>
<td>22.5</td>
<td>20.0</td>
</tr>
</tbody>
</table>
## 4.2 Synthetic Summaries
Table 5: Results of baseline and proposed approach (All results are in $\%).80$ summaries generated by GPT- 3 and Studio21 used for inference from CodeContests.
Table 4 and 5 show the results for baseline, synthetic summaries generated by GPT-3, Studio21 and PEGASUS in terms of SAcc for two experiments. For the AP experiment, we can observe that the performance of the baseline outperforms synthetic summary-based models. However, the proposed model shows an average similar performance compared to the baseline for the EWPR experiment. Moreover, Appendix I shows the results for top 500 and top 1000 summaries from GPT-3 and Studio21, respectively.
basic and expert summary-based models outperform baseline by 4.34% and 5.15% on average for APPS dataset, respectively. Further analysis shows that the expert summary-based model shows improved performance by $\sim1\%$ compared to the basic summary-based model.
## 4.3 Analysis
On the CodeContests dataset (Li et al., 2022), we show an average improvement of 11.88% in terms of SAcc. For this dataset, we did not separate the problems by difficulty. This is because the problems come from different sources and have different scales of difficulty. Thus, we did not report the SAcc when weighted by difficulty in Table 2.
Why does eliminating the worst problems help? From Tables 2, we can observe that EWPR and BWPR have improved performance compared to AP for both human and synthetically generated summaries. By analyzing the summarized worst problems, we notice a difference in the summarization style which shows that these summaries are outliers and do not match the distribution of the other summaries. This can cause a problem in synthesizing a good program since the model loses important information. Hence, we believe that eliminating the worst problems improves model
Our analysis shows that many problems where the basic summary would fail, however, the expert summary would succeed and vice-versa. Thus, we choose the best summary for each problem after evaluating both summaries and then calculate the results for the best summaries. Table 3 shows results when taking the best summary for
|
[
[
0.5263671875,
0.0872802734375,
0.5576171875,
0.0872802734375,
0.5576171875,
0.10394287109375,
0.5263671875,
0.10394287109375
],
[
0.67236328125,
0.0885009765625,
0.72900390625,
0.0885009765625,
0.72900390625,
0.10272216796875,
0.67236328125,
0.10272216796875
],
[
0.240966796875,
0.0968017578125,
0.291259765625,
0.0968017578125,
0.291259765625,
0.10809326171875,
0.240966796875,
0.10809326171875
],
[
0.35107421875,
0.0955810546875,
0.424072265625,
0.0955810546875,
0.424072265625,
0.10986328125,
0.35107421875,
0.10986328125
],
[
0.54150390625,
0.10986328125,
0.69287109375,
0.10986328125,
0.69287109375,
0.12469482421875,
0.54150390625,
0.12469482421875
],
[
0.7021484375,
0.10986328125,
0.77490234375,
0.10986328125,
0.77490234375,
0.12469482421875,
0.7021484375,
0.12469482421875
],
[
0.337646484375,
0.1312255859375,
0.43408203125,
0.133056640625,
0.43408203125,
0.1478271484375,
0.337646484375,
0.1455078125
],
[
0.468505859375,
0.11102294921875,
0.5322265625,
0.11102294921875,
0.5322265625,
0.122314453125,
0.468505859375,
0.122314453125
],
[
0.476806640625,
0.1312255859375,
0.5224609375,
0.1312255859375,
0.5224609375,
0.146728515625,
0.476806640625,
0.146728515625
],
[
0.5556640625,
0.1318359375,
0.60205078125,
0.1318359375,
0.60205078125,
0.146728515625,
0.5556640625,
0.146728515625
],
[
0.63623046875,
0.1318359375,
0.68115234375,
0.1318359375,
0.68115234375,
0.146728515625,
0.63623046875,
0.146728515625
],
[
0.71728515625,
0.1318359375,
0.76318359375,
0.1318359375,
0.76318359375,
0.146728515625,
0.71728515625,
0.146728515625
],
[
0.2392578125,
0.1383056640625,
0.29296875,
0.1383056640625,
0.29296875,
0.152587890625,
0.2392578125,
0.152587890625
],
[
0.47607421875,
0.1455078125,
0.5224609375,
0.1455078125,
0.5224609375,
0.1602783203125,
0.47607421875,
0.1602783203125
],
[
0.5556640625,
0.1455078125,
0.60205078125,
0.1455078125,
0.60205078125,
0.1602783203125,
0.5556640625,
0.1602783203125
],
[
0.63818359375,
0.1455078125,
0.6826171875,
0.1455078125,
0.6826171875,
0.1602783203125,
0.63818359375,
0.1602783203125
],
[
0.71728515625,
0.1455078125,
0.76318359375,
0.1455078125,
0.76318359375,
0.1602783203125,
0.71728515625,
0.1602783203125
],
[
0.55908203125,
0.1591796875,
0.59765625,
0.1591796875,
0.59765625,
0.17578125,
0.55908203125,
0.17578125
],
[
0.35107421875,
0.1478271484375,
0.424072265625,
0.1478271484375,
0.424072265625,
0.1591796875,
0.35107421875,
0.1591796875
],
[
0.340087890625,
0.1602783203125,
0.43408203125,
0.1602783203125,
0.43408203125,
0.174560546875,
0.340087890625,
0.174560546875
],
[
0.478515625,
0.1602783203125,
0.51708984375,
0.1602783203125,
0.51708984375,
0.17578125,
0.478515625,
0.17578125
],
[
0.6396484375,
0.1591796875,
0.67822265625,
0.1591796875,
0.67822265625,
0.17578125,
0.6396484375,
0.17578125
],
[
0.71875,
0.1602783203125,
0.759765625,
0.1602783203125,
0.759765625,
0.17578125,
0.71875,
0.17578125
],
[
0.47607421875,
0.18115234375,
0.5224609375,
0.18115234375,
0.5224609375,
0.196533203125,
0.47607421875,
0.196533203125
],
[
0.71728515625,
0.18115234375,
0.76318359375,
0.18115234375,
0.76318359375,
0.196533203125,
0.71728515625,
0.196533203125
],
[
0.5556640625,
0.1817626953125,
0.60205078125,
0.1817626953125,
0.60205078125,
0.196533203125,
0.5556640625,
0.196533203125
],
[
0.32080078125,
0.1817626953125,
0.455078125,
0.1817626953125,
0.455078125,
0.196533203125,
0.32080078125,
0.196533203125
],
[
0.63818359375,
0.1817626953125,
0.6826171875,
0.1817626953125,
0.6826171875,
0.196533203125,
0.63818359375,
0.196533203125
],
[
0.337646484375,
0.203125,
0.432373046875,
0.204833984375,
0.432373046875,
0.2197265625,
0.337646484375,
0.21728515625
],
[
0.47607421875,
0.203125,
0.5224609375,
0.203125,
0.5224609375,
0.218505859375,
0.47607421875,
0.218505859375
],
[
0.5556640625,
0.2037353515625,
0.60205078125,
0.2037353515625,
0.60205078125,
0.218505859375,
0.5556640625,
0.218505859375
],
[
0.63623046875,
0.2037353515625,
0.6826171875,
0.2037353515625,
0.6826171875,
0.218505859375,
0.63623046875,
0.218505859375
],
[
0.71728515625,
0.2037353515625,
0.76416015625,
0.2037353515625,
0.76416015625,
0.218505859375,
0.71728515625,
0.218505859375
],
[
0.47607421875,
0.21728515625,
0.5224609375,
0.21728515625,
0.5224609375,
0.2327880859375,
0.47607421875,
0.2327880859375
],
[
0.63818359375,
0.21728515625,
0.6826171875,
0.21728515625,
0.6826171875,
0.2327880859375,
0.63818359375,
0.2327880859375
],
[
0.71728515625,
0.21728515625,
0.76318359375,
0.21728515625,
0.76318359375,
0.2327880859375,
0.71728515625,
0.2327880859375
],
[
0.5556640625,
0.218505859375,
0.60205078125,
0.218505859375,
0.60205078125,
0.2327880859375,
0.5556640625,
0.2327880859375
],
[
0.2333984375,
0.2197265625,
0.297119140625,
0.2197265625,
0.297119140625,
0.23095703125,
0.2333984375,
0.23095703125
],
[
0.35107421875,
0.218505859375,
0.42236328125,
0.218505859375,
0.42236328125,
0.2327880859375,
0.35107421875,
0.2327880859375
],
[
0.56103515625,
0.232177734375,
0.59765625,
0.232177734375,
0.59765625,
0.2476806640625,
0.56103515625,
0.2476806640625
],
[
0.7197265625,
0.232177734375,
0.75830078125,
0.232177734375,
0.75830078125,
0.2476806640625,
0.7197265625,
0.2476806640625
],
[
0.640625,
0.2327880859375,
0.6767578125,
0.2327880859375,
0.6767578125,
0.2452392578125,
0.640625,
0.2452392578125
],
[
0.341796875,
0.2340087890625,
0.4306640625,
0.2340087890625,
0.4306640625,
0.2452392578125,
0.341796875,
0.2452392578125
],
[
0.480224609375,
0.2327880859375,
0.515625,
0.2327880859375,
0.515625,
0.2452392578125,
0.480224609375,
0.2452392578125
],
[
0.31982421875,
0.254150390625,
0.455078125,
0.254150390625,
0.455078125,
0.268310546875,
0.31982421875,
0.268310546875
],
[
0.47607421875,
0.254150390625,
0.5224609375,
0.254150390625,
0.5224609375,
0.268310546875,
0.47607421875,
0.268310546875
],
[
0.63623046875,
0.254150390625,
0.68115234375,
0.254150390625,
0.68115234375,
0.268310546875,
0.63623046875,
0.268310546875
],
[
0.71728515625,
0.254150390625,
0.76416015625,
0.254150390625,
0.76416015625,
0.268310546875,
0.71728515625,
0.268310546875
],
[
0.337646484375,
0.27490234375,
0.432373046875,
0.276611328125,
0.432373046875,
0.29150390625,
0.337646484375,
0.289306640625
],
[
0.56103515625,
0.254638671875,
0.59765625,
0.254638671875,
0.59765625,
0.267333984375,
0.56103515625,
0.267333984375
],
[
0.476806640625,
0.275634765625,
0.5224609375,
0.275634765625,
0.5224609375,
0.29150390625,
0.476806640625,
0.29150390625
],
[
0.55517578125,
0.275634765625,
0.60205078125,
0.275634765625,
0.60205078125,
0.29150390625,
0.55517578125,
0.29150390625
],
[
0.63623046875,
0.275634765625,
0.6826171875,
0.275634765625,
0.6826171875,
0.29150390625,
0.63623046875,
0.29150390625
],
[
0.71728515625,
0.275634765625,
0.76318359375,
0.275634765625,
0.76318359375,
0.290283203125,
0.71728515625,
0.290283203125
],
[
0.47607421875,
0.289306640625,
0.52294921875,
0.289306640625,
0.52294921875,
0.3046875,
0.47607421875,
0.3046875
],
[
0.5556640625,
0.290283203125,
0.60205078125,
0.290283203125,
0.60205078125,
0.3046875,
0.5556640625,
0.3046875
],
[
0.63623046875,
0.290283203125,
0.68115234375,
0.2880859375,
0.6826171875,
0.303955078125,
0.63818359375,
0.305908203125
],
[
0.71728515625,
0.289306640625,
0.76318359375,
0.289306640625,
0.76318359375,
0.3046875,
0.71728515625,
0.3046875
],
[
0.35107421875,
0.292236328125,
0.42236328125,
0.292236328125,
0.42236328125,
0.303955078125,
0.35107421875,
0.303955078125
],
[
0.22412109375,
0.29150390625,
0.306396484375,
0.29150390625,
0.306396484375,
0.302734375,
0.22412109375,
0.302734375
],
[
0.476806640625,
0.303955078125,
0.51904296875,
0.303955078125,
0.51904296875,
0.319580078125,
0.476806640625,
0.319580078125
],
[
0.5576171875,
0.303955078125,
0.59765625,
0.303955078125,
0.59765625,
0.319580078125,
0.5576171875,
0.319580078125
],
[
0.6396484375,
0.303955078125,
0.67822265625,
0.303955078125,
0.67822265625,
0.319580078125,
0.6396484375,
0.319580078125
],
[
0.71728515625,
0.303955078125,
0.759765625,
0.303955078125,
0.759765625,
0.319580078125,
0.71728515625,
0.319580078125
],
[
0.318115234375,
0.325927734375,
0.455078125,
0.32666015625,
0.455078125,
0.341552734375,
0.318115234375,
0.34033203125
],
[
0.341796875,
0.305908203125,
0.432373046875,
0.305908203125,
0.432373046875,
0.317626953125,
0.341796875,
0.317626953125
],
[
0.474365234375,
0.325927734375,
0.5224609375,
0.325927734375,
0.5224609375,
0.341552734375,
0.474365234375,
0.341552734375
],
[
0.5556640625,
0.325927734375,
0.603515625,
0.325927734375,
0.603515625,
0.341552734375,
0.5556640625,
0.341552734375
],
[
0.63623046875,
0.325927734375,
0.6826171875,
0.325927734375,
0.6826171875,
0.341552734375,
0.63623046875,
0.341552734375
],
[
0.71728515625,
0.325927734375,
0.76318359375,
0.325927734375,
0.76318359375,
0.341552734375,
0.71728515625,
0.341552734375
],
[
0.1175537109375,
0.355712890625,
0.6455078125,
0.355712890625,
0.6455078125,
0.369384765625,
0.1175537109375,
0.369384765625
],
[
0.6484375,
0.35986328125,
0.654296875,
0.364013671875,
0.6513671875,
0.366455078125,
0.6455078125,
0.3623046875
],
[
0.654296875,
0.35693359375,
0.8818359375,
0.35693359375,
0.8818359375,
0.37060546875,
0.654296875,
0.37060546875
],
[
0.1158447265625,
0.371826171875,
0.49951171875,
0.371826171875,
0.49951171875,
0.385498046875,
0.1158447265625,
0.385498046875
],
[
0.25439453125,
0.412109375,
0.279541015625,
0.412109375,
0.279541015625,
0.427001953125,
0.25439453125,
0.427001953125
],
[
0.3837890625,
0.412109375,
0.435791015625,
0.412109375,
0.435791015625,
0.427001953125,
0.3837890625,
0.427001953125
],
[
0.51123046875,
0.410888671875,
0.8818359375,
0.410400390625,
0.8818359375,
0.424560546875,
0.51123046875,
0.42578125
],
[
0.133544921875,
0.418701171875,
0.183837890625,
0.418701171875,
0.183837890625,
0.432861328125,
0.133544921875,
0.432861328125
],
[
0.264404296875,
0.432861328125,
0.405517578125,
0.432861328125,
0.405517578125,
0.446533203125,
0.264404296875,
0.446533203125
],
[
0.51318359375,
0.427490234375,
0.88330078125,
0.427490234375,
0.88330078125,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.202392578125,
0.43408203125,
0.270263671875,
0.43408203125,
0.270263671875,
0.4453125,
0.202392578125,
0.4453125
],
[
0.39794921875,
0.439453125,
0.405517578125,
0.43408203125,
0.411376953125,
0.438232421875,
0.40478515625,
0.443603515625
],
[
0.41064453125,
0.43408203125,
0.47607421875,
0.43408203125,
0.47607421875,
0.4453125,
0.41064453125,
0.4453125
],
[
0.51123046875,
0.44482421875,
0.8818359375,
0.44482421875,
0.8818359375,
0.457763671875,
0.51123046875,
0.457763671875
],
[
0.133544921875,
0.451904296875,
0.1822509765625,
0.451904296875,
0.1822509765625,
0.466064453125,
0.133544921875,
0.466064453125
],
[
0.208251953125,
0.451904296875,
0.2509765625,
0.451904296875,
0.2509765625,
0.46728515625,
0.208251953125,
0.46728515625
],
[
0.28369140625,
0.451904296875,
0.31982421875,
0.451904296875,
0.31982421875,
0.46728515625,
0.28369140625,
0.46728515625
],
[
0.3525390625,
0.451904296875,
0.393798828125,
0.451904296875,
0.393798828125,
0.46728515625,
0.3525390625,
0.46728515625
],
[
0.4248046875,
0.451904296875,
0.465087890625,
0.451904296875,
0.465087890625,
0.46728515625,
0.4248046875,
0.46728515625
],
[
0.51123046875,
0.46142578125,
0.6484375,
0.46142578125,
0.6484375,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.208251953125,
0.47021484375,
0.2509765625,
0.47021484375,
0.2509765625,
0.486328125,
0.208251953125,
0.486328125
],
[
0.2822265625,
0.47021484375,
0.318115234375,
0.47021484375,
0.318115234375,
0.48681640625,
0.2822265625,
0.48681640625
],
[
0.353515625,
0.47021484375,
0.391357421875,
0.47021484375,
0.391357421875,
0.486328125,
0.353515625,
0.486328125
],
[
0.4248046875,
0.4697265625,
0.465087890625,
0.47265625,
0.462646484375,
0.48681640625,
0.424072265625,
0.485107421875
],
[
0.54833984375,
0.490478515625,
0.72314453125,
0.490478515625,
0.72314453125,
0.5048828125,
0.54833984375,
0.5048828125
],
[
0.124267578125,
0.471435546875,
0.18896484375,
0.471435546875,
0.18896484375,
0.486328125,
0.124267578125,
0.486328125
],
[
0.5146484375,
0.4912109375,
0.5556640625,
0.4912109375,
0.5556640625,
0.5029296875,
0.5146484375,
0.5029296875
],
[
0.11419677734375,
0.49951171875,
0.48779296875,
0.50048828125,
0.48779296875,
0.515625,
0.11419677734375,
0.51416015625
],
[
0.1175537109375,
0.515625,
0.490234375,
0.515625,
0.490234375,
0.529296875,
0.1175537109375,
0.529296875
],
[
0.5087890625,
0.51123046875,
0.88330078125,
0.51318359375,
0.88330078125,
0.52783203125,
0.5087890625,
0.525390625
],
[
0.1158447265625,
0.52978515625,
0.484375,
0.52978515625,
0.484375,
0.54345703125,
0.1158447265625,
0.54345703125
],
[
0.51123046875,
0.52978515625,
0.8818359375,
0.52978515625,
0.8818359375,
0.54345703125,
0.51123046875,
0.54345703125
],
[
0.51123046875,
0.5458984375,
0.88330078125,
0.5458984375,
0.88330078125,
0.55859375,
0.51123046875,
0.55859375
],
[
0.509765625,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.57470703125,
0.509765625,
0.57470703125
],
[
0.1175537109375,
0.57373046875,
0.489501953125,
0.57373046875,
0.489501953125,
0.58740234375,
0.1175537109375,
0.58740234375
],
[
0.51318359375,
0.57763671875,
0.8818359375,
0.57763671875,
0.8818359375,
0.5927734375,
0.51318359375,
0.5927734375
],
[
0.1175537109375,
0.59033203125,
0.489501953125,
0.59033203125,
0.489501953125,
0.60400390625,
0.1175537109375,
0.60400390625
],
[
0.51318359375,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.1175537109375,
0.60693359375,
0.48779296875,
0.60693359375,
0.48779296875,
0.62060546875,
0.1175537109375,
0.62060546875
],
[
0.509765625,
0.60986328125,
0.88330078125,
0.609375,
0.88330078125,
0.62353515625,
0.509765625,
0.62451171875
],
[
0.1158447265625,
0.62158203125,
0.48779296875,
0.62158203125,
0.48779296875,
0.63525390625,
0.1158447265625,
0.63525390625
],
[
0.509765625,
0.6259765625,
0.8818359375,
0.62451171875,
0.8818359375,
0.6396484375,
0.509765625,
0.64013671875
],
[
0.1158447265625,
0.63818359375,
0.48779296875,
0.63818359375,
0.48779296875,
0.65283203125,
0.1158447265625,
0.65283203125
],
[
0.51318359375,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65625,
0.51318359375,
0.65625
],
[
0.1158447265625,
0.65478515625,
0.334228515625,
0.65478515625,
0.334228515625,
0.66845703125,
0.1158447265625,
0.66845703125
],
[
0.51123046875,
0.65771484375,
0.8818359375,
0.65771484375,
0.8818359375,
0.67138671875,
0.51123046875,
0.67138671875
],
[
0.51123046875,
0.6728515625,
0.76416015625,
0.67333984375,
0.76416015625,
0.6884765625,
0.51123046875,
0.68701171875
],
[
0.1175537109375,
0.68310546875,
0.48779296875,
0.68310546875,
0.48779296875,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.11419677734375,
0.69775390625,
0.486083984375,
0.69775390625,
0.486083984375,
0.7119140625,
0.11419677734375,
0.7119140625
],
[
0.509765625,
0.703125,
0.62451171875,
0.705078125,
0.62451171875,
0.72021484375,
0.509765625,
0.71875
],
[
0.1158447265625,
0.71435546875,
0.486083984375,
0.71435546875,
0.486083984375,
0.72802734375,
0.1158447265625,
0.72802734375
],
[
0.51123046875,
0.7255859375,
0.8818359375,
0.72705078125,
0.8818359375,
0.7412109375,
0.51123046875,
0.74072265625
],
[
0.1175537109375,
0.732421875,
0.490234375,
0.732421875,
0.490234375,
0.7451171875,
0.1175537109375,
0.7451171875
],
[
0.1175537109375,
0.7490234375,
0.490234375,
0.7490234375,
0.490234375,
0.7626953125,
0.1175537109375,
0.7626953125
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1158447265625,
0.76318359375,
0.486083984375,
0.76318359375,
0.486083984375,
0.77685546875,
0.1158447265625,
0.77685546875
],
[
0.51123046875,
0.75830078125,
0.8818359375,
0.76025390625,
0.8818359375,
0.77490234375,
0.51123046875,
0.7724609375
],
[
0.1175537109375,
0.77978515625,
0.45166015625,
0.77978515625,
0.45166015625,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.51123046875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.1175537109375,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.11419677734375,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.11419677734375,
0.83837890625
],
[
0.509765625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.490234375,
0.84033203125,
0.490234375,
0.85498046875,
0.1158447265625,
0.85498046875
],
[
0.509765625,
0.84130859375,
0.8818359375,
0.84130859375,
0.8818359375,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.1158447265625,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.509765625,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.11419677734375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.509765625,
0.88720703125
],
[
0.1175537109375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"AP",
"EWPR",
"Model",
"Difficulty",
"Proposed Baseline",
"Proposed",
"Introductory",
"Baseline",
"41.75",
"38.66",
"41.11",
"41.67",
"GPT-3",
"20.30",
"18.80",
"18.18",
"20.66",
"3.28",
"Interview",
"Competition",
"2.87",
"2.73",
"3.64",
"20.17",
"20.55",
"18.89",
"Weighted Average",
"19.14",
"Introductory",
"39.53",
"31.63",
"39.04",
"36.36",
"12.28",
"10.57",
"12.37",
"11.00",
"Studio21",
"Interview",
"1.21",
"1.38",
"1.38",
"Competition",
"1.67",
"Weighted Average",
"11.53",
"10.61",
"10.98",
"Introductory",
"9.66",
"42.96",
"34.48",
"44.26",
"40.98",
"37.70",
"10.57",
"14.29",
"21.56",
"Interview",
"PEGASUS",
"4.76",
"0.00",
"2.76",
"0.00",
"Weighted Average",
"Competition",
"30.47",
"16.88",
"25.50",
"24.73",
"Table 4: Results of baseline and proposed approach (All results are in %).",
"",
"Summaries generated by GPT-3,",
"Studio21, and PEGASUS used for inference from APPS",
"AP",
"EWPR",
"each problem for APPS dataset. We observe a",
"Model",
"ProposedBaseline",
"9.86%, 11.48%, and 1.91% increase on SAcc for",
"Baseline",
"",
"Proposed",
"introductory, interview, and competition level prob",
"GPT-3",
"12.50",
"10.0",
"18.75",
"18.75",
"lems,respectively.",
"12.50",
"8.75",
"22.5",
"20.0",
"Synthetic Summaries",
"Studio21",
"4.2",
"Table 5: Results of baseline and proposed approach",
"(All results are in %). 80 summaries generated by GPT-",
"Table 4 and 5 show the results for baseline, syn-",
"3 and Studio21 used for inference from CodeContests.",
"thetic summaries generated by GPT-3, Studio21",
"and PEGASUS in terms of SAcc for two exper-",
"iments. For the AP experiment, we can observe",
"basic and expert summary-based models outper-",
"that the performance of the baseline outperforms",
"form baseline by 4.34% and 5.15% on average for",
"synthetic summary-based models. However, the",
"APPS dataset, respectively. Further analysis shows",
"proposed model shows an average similar perfor-",
"that the expert summary-based model shows im-",
"mance compared to the baseline for the EWPR",
"proved performance by ~ 1% compared to the",
"experiment. Moreover, Appendix I shows the re-",
"basic summary-based model.",
"sults for top 500 and top 1000 summaries from",
"GPT-3 and Studio21, respectively.",
"On the CodeContests dataset (Li et al., 2022), we",
"show an average improvement of 11.88% in terms",
"4.3Analysis",
"of SAcc. For this dataset, we did not separate the",
"Why does eliminating the worst problems help?",
"problems by difficulty. This is because the prob-",
"lems come from different sources and have differ-",
"From Tables 2, we can observe that EWPR and",
"ent scales of difficulty. Thus, we did not report the",
"BWPR have improved performance compared to",
"SAcc when weighted by difficulty in Table 2.",
"AP for both human and synthetically generated",
"summaries. By analyzing the summarized worst",
"Our analysis shows that many problems where the",
"problems, we notice a difference in the summariza-",
"basic summary would fail, however, the expert",
"tion style which shows that these summaries are",
"summary would succeed and vice-versa. Thus,",
"outliers and do not match the distribution of the",
"we choose the best summary for each problem",
"other summaries. This can cause a problem in syn",
"after evaluating both summaries and then calcu-",
"thesizing a good program since the model loses",
"late the results for the best summaries.Table 3",
"important information.Hence, we believe that",
"shows results when taking the best summary for",
"eliminating the worst problems improves model"
] |
{
"arxiv_id": "2203.08597",
"language": "en",
"page_idx": 3
}
|
||
2203.08597-en-4
|
2203.08597-en-4.jpg
|
<FigureHere>
Figure 1: (Top plot) Mean frequency of POS for problems where programs where the generated by both the original and summarized prompt pass all test cases, and (Bottom plot) mean frequency of POS for problems where the summary passes all test cases and the original did not. The blue bar represents the mean of the entire dataset. analyzed only the top 11 most occurring POS. The plot shows that higher number of nouns degrade model performance.
and propernouns for problems that passed all test cases is lower than the entire dataset. In the bottom plot, we observe that the frequency for nouns and propernouns is higher for the original question (which had<100% accuracy on the test cases) and lower for the summary (which had 100% accuracy on the test cases). Thus, we can see that number of nouns degrades performance. We also see in the bottom chart that overuse of punctuation can be detrimental to performance. From the results in Figure 1 we see results of nouns affecting performance along with excessive punctuation. Additional detailed analysis is presented in Appendix $B.$
### 5 Conclusion
This paper introduces a summarization-based approach for efficient program synthesis. Experimental results show that the proposed approach improves the performance of the Codex model by on average $\sim8\%$ across various levels of programming questions provided by the APPS and $\sim11\%$ on the CodeContests. Further, this paper proposes a meta-dataset consisting of $\sim450$ human-generated basic and expert-level summaries as well as $\sim8k$ synthetically generated summaries by GPT-3 and Studio21; this can be helpful for future research on writing better instructions for the program synthesis. We show that program synthesis models benefit from concise prompts, hence, we believe that less number of high-quality instances are better than more low-quality data instances.
Future Extensions The decomposition of
prompts has been shown to improve accuracy $\tilde{(\text{Mishra et al., 2022; Patel et al., 2022); splitting up}}$ the summarization task the resulting summary can potentially result in higher accuracy for the Codex model in future. Additionally, the PEGASUS model could be used in conjunction with other models to perform the detailed algorithm outlined in Appendix N.
## $\mathbf{Limitations}$
Our summary-based approach shows improved performance on program synthesis models, however, it shows competitive performance on synthetic summaries. We believe that the generation of high-quality summaries can improve performance, hence, designing efficient prompts to improve synthetic summaries can be the scope of further research. Furthermore, human-generated summaries
performance.
Is there any possible bias in the meta-dataset? Recent studies shows that bias propagates in human-annotated datasets (Geva et al., 2019; Parmar et al., 2022a). Given that our summaries are also human-generated, there will be some bias in the dataset. Some details that are critical to one person can be trivial to others. In the context of generating expert summaries, assumptions about expert knowledge can vary. This bias causes drift in the dataset and hinders the model's performance. Similar to Mishra et al. (2021), we can provide a template for what is expected from the summary generator to reduce bias.
Why is competition accuracy low? We believe that these problems require multi-hop reasoning, even after summarization, which is still a challenge for language models.
Impact of POS on Accuracy In the top plot of Figure 1, we observe that frequency of nouns
|
[
[
0.51318359375,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.51318359375,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.51318359375,
0.11517333984375
],
[
0.51318359375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.51123046875,
0.133056640625,
0.87890625,
0.133056640625,
0.87890625,
0.146728515625,
0.51123046875,
0.146728515625
],
[
0.51123046875,
0.1484375,
0.880859375,
0.1484375,
0.880859375,
0.162109375,
0.51123046875,
0.162109375
],
[
0.51123046875,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.5146484375,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19482421875,
0.5146484375,
0.19482421875
],
[
0.172119140625,
0.20068359375,
0.47021484375,
0.20068359375,
0.47021484375,
0.215576171875,
0.172119140625,
0.215576171875
],
[
0.509765625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.279541015625,
0.2197265625,
0.359375,
0.2197265625,
0.359375,
0.2298583984375,
0.279541015625,
0.2298583984375
],
[
0.51318359375,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51318359375,
0.22802734375
],
[
0.51318359375,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.51318359375,
0.24462890625
],
[
0.19140625,
0.2452392578125,
0.2225341796875,
0.2452392578125,
0.2225341796875,
0.254150390625,
0.19140625,
0.254150390625
],
[
0.51123046875,
0.2476806640625,
0.880859375,
0.2476806640625,
0.880859375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.18896484375,
0.254638671875,
0.225830078125,
0.254638671875,
0.225830078125,
0.262451171875,
0.18896484375,
0.262451171875
],
[
0.193115234375,
0.26318359375,
0.2091064453125,
0.26318359375,
0.2091064453125,
0.270751953125,
0.193115234375,
0.270751953125
],
[
0.509765625,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.275634765625,
0.509765625,
0.275634765625
],
[
0.51123046875,
0.27783203125,
0.8818359375,
0.279052734375,
0.8818359375,
0.29345703125,
0.51123046875,
0.292236328125
],
[
0.139404296875,
0.29638671875,
0.1561279296875,
0.29638671875,
0.1561279296875,
0.303955078125,
0.139404296875,
0.303955078125
],
[
0.51123046875,
0.29443359375,
0.53076171875,
0.29443359375,
0.53076171875,
0.30810546875,
0.51123046875,
0.30810546875
],
[
0.5205078125,
0.318359375,
0.5322265625,
0.32373046875,
0.5224609375,
0.3349609375,
0.509765625,
0.328857421875
],
[
0.5380859375,
0.319580078125,
0.640625,
0.32177734375,
0.640625,
0.33740234375,
0.5380859375,
0.3349609375
],
[
0.1822509765625,
0.3486328125,
0.468505859375,
0.350830078125,
0.468505859375,
0.3681640625,
0.1822509765625,
0.366455078125
],
[
0.51123046875,
0.345703125,
0.8818359375,
0.345703125,
0.8818359375,
0.35986328125,
0.51123046875,
0.35986328125
],
[
0.51318359375,
0.3623046875,
0.884765625,
0.3623046875,
0.884765625,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.280517578125,
0.37060546875,
0.3525390625,
0.37060546875,
0.3525390625,
0.38134765625,
0.280517578125,
0.38134765625
],
[
0.51123046875,
0.377685546875,
0.884765625,
0.37890625,
0.884765625,
0.393798828125,
0.51123046875,
0.392578125
],
[
0.5087890625,
0.393798828125,
0.884765625,
0.392578125,
0.884765625,
0.410400390625,
0.5087890625,
0.410888671875
],
[
0.1175537109375,
0.40625,
0.489501953125,
0.40625,
0.489501953125,
0.419189453125,
0.1175537109375,
0.419189453125
],
[
0.51318359375,
0.410888671875,
0.8818359375,
0.410888671875,
0.8818359375,
0.424560546875,
0.51318359375,
0.424560546875
],
[
0.1175537109375,
0.42041015625,
0.486083984375,
0.42041015625,
0.486083984375,
0.43408203125,
0.1175537109375,
0.43408203125
],
[
0.509765625,
0.427001953125,
0.8818359375,
0.42578125,
0.8818359375,
0.440673828125,
0.509765625,
0.441162109375
],
[
0.1175537109375,
0.435302734375,
0.48779296875,
0.435302734375,
0.48779296875,
0.448974609375,
0.1175537109375,
0.448974609375
],
[
0.1175537109375,
0.448974609375,
0.489501953125,
0.448974609375,
0.489501953125,
0.4619140625,
0.1175537109375,
0.4619140625
],
[
0.51318359375,
0.443603515625,
0.8818359375,
0.443603515625,
0.8818359375,
0.457275390625,
0.51318359375,
0.457275390625
],
[
0.1175537109375,
0.463134765625,
0.489501953125,
0.463134765625,
0.489501953125,
0.476806640625,
0.1175537109375,
0.476806640625
],
[
0.51123046875,
0.458984375,
0.880859375,
0.458984375,
0.880859375,
0.47265625,
0.51123046875,
0.47265625
],
[
0.1158447265625,
0.476806640625,
0.48779296875,
0.476806640625,
0.48779296875,
0.490478515625,
0.1158447265625,
0.490478515625
],
[
0.51123046875,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.488037109375,
0.51123046875,
0.488037109375
],
[
0.1175537109375,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.1175537109375,
0.5048828125
],
[
0.509765625,
0.4912109375,
0.8818359375,
0.4912109375,
0.8818359375,
0.5048828125,
0.509765625,
0.5048828125
],
[
0.1175537109375,
0.505859375,
0.491943359375,
0.505859375,
0.491943359375,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.51318359375,
0.50732421875,
0.8818359375,
0.50732421875,
0.8818359375,
0.52099609375,
0.51318359375,
0.52099609375
],
[
0.1175537109375,
0.52099609375,
0.297119140625,
0.52099609375,
0.297119140625,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.51123046875,
0.52392578125,
0.88330078125,
0.52392578125,
0.88330078125,
0.53759765625,
0.51123046875,
0.53759765625
],
[
0.51123046875,
0.54052734375,
0.8818359375,
0.54052734375,
0.8818359375,
0.55419921875,
0.51123046875,
0.55419921875
],
[
0.51123046875,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.57080078125,
0.51123046875,
0.57080078125
],
[
0.1158447265625,
0.5625,
0.2166748046875,
0.5625,
0.2166748046875,
0.57666015625,
0.1158447265625,
0.57666015625
],
[
0.509765625,
0.57080078125,
0.88330078125,
0.57177734375,
0.88330078125,
0.5859375,
0.509765625,
0.5849609375
],
[
0.1175537109375,
0.59033203125,
0.489501953125,
0.59033203125,
0.489501953125,
0.60400390625,
0.1175537109375,
0.60400390625
],
[
0.509765625,
0.58740234375,
0.78857421875,
0.58740234375,
0.78857421875,
0.60107421875,
0.509765625,
0.60107421875
],
[
0.1158447265625,
0.60546875,
0.48779296875,
0.60693359375,
0.48779296875,
0.62158203125,
0.1158447265625,
0.62060546875
],
[
0.51318359375,
0.6123046875,
0.5859375,
0.6123046875,
0.5859375,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.57666015625,
0.60986328125,
0.88330078125,
0.61083984375,
0.88330078125,
0.6259765625,
0.57666015625,
0.62451171875
],
[
0.1158447265625,
0.62255859375,
0.490234375,
0.62255859375,
0.490234375,
0.6357421875,
0.1158447265625,
0.6357421875
],
[
0.509765625,
0.626953125,
0.8818359375,
0.62744140625,
0.8818359375,
0.642578125,
0.509765625,
0.64111328125
],
[
0.1175537109375,
0.64013671875,
0.486083984375,
0.64013671875,
0.486083984375,
0.65087890625,
0.1175537109375,
0.65087890625
],
[
0.509765625,
0.642578125,
0.880859375,
0.6435546875,
0.880859375,
0.65771484375,
0.509765625,
0.65673828125
],
[
0.1175537109375,
0.65478515625,
0.48779296875,
0.65478515625,
0.48779296875,
0.66845703125,
0.1175537109375,
0.66845703125
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.1158447265625,
0.67041015625,
0.48779296875,
0.67041015625,
0.48779296875,
0.68408203125,
0.1158447265625,
0.68408203125
],
[
0.51123046875,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.51123046875,
0.6904296875
],
[
0.11419677734375,
0.68701171875,
0.48779296875,
0.68603515625,
0.48779296875,
0.70068359375,
0.11419677734375,
0.70166015625
],
[
0.51318359375,
0.6923828125,
0.880859375,
0.6923828125,
0.880859375,
0.7060546875,
0.51318359375,
0.7060546875
],
[
0.1158447265625,
0.70361328125,
0.486083984375,
0.703125,
0.486083984375,
0.71728515625,
0.1158447265625,
0.71875
],
[
0.509765625,
0.7080078125,
0.8818359375,
0.7080078125,
0.8818359375,
0.7216796875,
0.509765625,
0.7216796875
],
[
0.1175537109375,
0.7197265625,
0.48779296875,
0.7197265625,
0.48779296875,
0.73291015625,
0.1175537109375,
0.73291015625
],
[
0.51123046875,
0.724609375,
0.8818359375,
0.724609375,
0.8818359375,
0.73828125,
0.51123046875,
0.73828125
],
[
0.1175537109375,
0.736328125,
0.489501953125,
0.736328125,
0.489501953125,
0.75,
0.1175537109375,
0.75
],
[
0.5087890625,
0.74072265625,
0.62890625,
0.7392578125,
0.62890625,
0.75439453125,
0.5087890625,
0.7548828125
],
[
0.118408203125,
0.751953125,
0.48779296875,
0.751953125,
0.48779296875,
0.765625,
0.118408203125,
0.765625
],
[
0.1175537109375,
0.7685546875,
0.486083984375,
0.7685546875,
0.486083984375,
0.7822265625,
0.1175537109375,
0.7822265625
],
[
0.51123046875,
0.7666015625,
0.615234375,
0.7666015625,
0.615234375,
0.78076171875,
0.51123046875,
0.78076171875
],
[
0.1175537109375,
0.78369140625,
0.301513671875,
0.78369140625,
0.301513671875,
0.79736328125,
0.1175537109375,
0.79736328125
],
[
0.51123046875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51123046875,
0.8046875
],
[
0.118408203125,
0.81103515625,
0.489501953125,
0.81103515625,
0.489501953125,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51123046875,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.82666015625,
0.489501953125,
0.8291015625,
0.489501953125,
0.84326171875,
0.1158447265625,
0.84130859375
],
[
0.51123046875,
0.822265625,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.51123046875,
0.83740234375
],
[
0.1158447265625,
0.84326171875,
0.48779296875,
0.84423828125,
0.48779296875,
0.859375,
0.1158447265625,
0.85791015625
],
[
0.51318359375,
0.84033203125,
0.88330078125,
0.84033203125,
0.88330078125,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.1175537109375,
0.85986328125,
0.274658203125,
0.85986328125,
0.274658203125,
0.87451171875,
0.1175537109375,
0.87451171875
],
[
0.51123046875,
0.85693359375,
0.88330078125,
0.85693359375,
0.88330078125,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90185546875,
0.1175537109375,
0.90185546875
],
[
0.509765625,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"and propernouns for problems that passed all test",
"cases is lower than the entire dataset. In the bot",
"tom plot, we observe that the frequency for nouns",
"and propernouns is higher for the original ques",
"tion (which had < 100% accuracy on the test cases)",
"and lower for the summary (which had 100% ac-",
"curacy on the test cases). Thus, we can see that",
"OPNAUX DETAD)NOUNVERB ADPOUNCTPRONGPACENUM",
"number of nouns degrades performance. We also",
"PartOfSpeech",
"see in the bottom chart that overuse of punctuation",
"can be detrimental to performance. From the re",
"alldata",
"sults in Figure I we see results of nouns affecting",
"original",
"bas",
"performance along with excessive punctuation. Ad-",
"ditional detailed analysis is presented in Appendix",
"30",
"B.",
"5",
"Conclusion",
"PNAUXDETADINOUNVERBADPOUNCTPRONSPACENUM",
"This paper introduces a summarization-based ap-",
"proach for efficient program synthesis. Experimen-",
"Part of Speec",
"tal results show that the proposed approach im-",
"proves the performance of the Codex model by on",
"Figure 1: (Top plot) Mean frequency of POS for prob-",
"average ~ 8% across various levels of program",
"lems where programs where the generated by both the",
"ming questions provided by the APPS and ~ 11%",
"original and summarized prompt pass all test cases, and",
"(Bottom plot) mean frequency of POS for problems",
"on the CodeContests. Further, this paper proposes a",
"where the summary passes all test cases and the orig-",
"meta-dataset consisting of ~ 450 human-generated",
"inal did not. The blue bar represents the mean of the",
"basic and expert-level summaries as well as ~ 8k",
"entire dataset. analyzed only the top 11 most occurring",
"synthetically generated summaries by GPT-3 and",
"POS. The plot shows that higher number of nouns de-",
"Studio2l; this can be helpful for future research",
"grade model performance.",
"on writing better instructions for the program syn-",
"thesis. We show that program synthesis models",
"benefit from concise prompts, hence, we believe",
"performance.",
"that less number of high-quality instances are better",
"Is there any possible bias in the meta-dataset?",
"than more low-quality data instances.",
"Recent studies shows that bias propagates in",
"Future",
"Extensions The decomposition of",
"human-annotated datasets (Geva et al., 2019; Par-",
"prompts has been shown to improve accuracy",
"mar et al..2022a). Given that our summaries are",
"(Mishra et al., 2022; Patel et al., 2022); splitting up",
"also human-generated, there will be some bias in",
"the summarization task the resulting summary can",
"the dataset. Some details that are critical to one",
"potentially result in higher accuracy for the Codex",
"person can be trivial to others. In the context of",
"model in future. Additionally, the PEGASUS",
"generating expert summaries, assumptions about",
"model could be used in conjunction with other",
"expert knowledge can vary. This bias causes drift",
"models to perform the detailed algorithm outlined",
"in the dataset and hinders the model's performance.",
"in Appendix N.",
"Similar to Mishra et al. (2021), we can provide a",
"template for what is expected from the summary",
"Limitations",
"generator to reduce bias.",
"Our summary-based approach shows improved",
"Why is competition accuracy low?We believe",
"performance on program synthesis models, how-",
"that these problems require multi-hop reasoning,",
"ever, it shows competitive performance on syn-",
"even after summarization, which is still a challenge",
"thetic summaries. We believe that the generation of",
"for language models.",
"high-quality summaries can improve performance,",
"hence, designing efficient prompts to improve syn-",
"Impact of POS on AccuracyIn the top plot of",
"thetic summaries can be the scope of further re-",
"Figure l, we observe that frequency of nouns",
" search. Furthermore, human-generated summaries"
] |
{
"arxiv_id": "2203.08597",
"language": "en",
"page_idx": 4
}
|
||
2306.11406-en-0
|
2306.11406-en-0.jpg
|
## Stable and Consistent Prediction of 3D Characteristic Orientation via Invariant Residual Learning
Seungwook Kim $^{*1}$ Chunghyun Park$^{*1}$ Yoonwoo Jeong$^{1}$ Jaesik Park$^{12}$ Minsu Cho$^{12}$
Abstract
2019; Yu et al., 2020; Li et al., 2021b) networks, the yielded representations suffer from loss of expressivity in return for robustness to rotation. A more straightforward and intuitive alternative is to canonicalize the point clouds by determining their characteristic orientation (Fang et al., 2020; Katzir et al., 2022); the canonicalized point cloud, which is obtained by canceling out the characteristic orientation, preserves its shape information and is readily used for downstream tasks without loss of expressivity. Although it is nontrivial to define the canonical orientation of a point cloud, there exist its desiderata (Katzir et al., 2022): stability- the canonical orientation is invariant to any rotation transformations of the input point cloud- and $consistentency-$the canonical orientation is invariant to intra-class variations $i.e.$, point clouds of the same class should have the same canonical pose. Recent methods to canonicalize the point clouds commonly exploit these desiderata to train their network in a self-supervised manner (Sun et al., 2021; Sajnani et al., 2022; Katzir et al., 2022).
Learning to predict reliable characteristic orientations of 3D point clouds is an important yet challenging problem, as different point clouds of the same class may have largely varying appearances. In this work, we introduce a novel method to decouple the shape geometry and semantics of the input point cloud to achieve both stability and consistency. The proposed method integrates shapegeometry-based SO(3)-equivariant learning and shape-semantics-based SO(3)-invariant residual learning, where a final characteristic orientation is obtained by calibrating an SO(3)-equivariant orientation hypothesis using an SO(3)-invariant residual rotation. In experiments, the proposed method not only demonstrates superior stability and consistency but also exhibits state-of-the-art performances when applied to point cloud part segmentation, given randomly rotated inputs.
## $1. \textbf{Introduction}$
Existing methods to predict the characteristic orientation of point clouds, however, suffer from satisfying both of the criteria, stability and consistency, simultaneously. Earlier methods either rely on heavy rotation augmentation (Sun et al., 2021) to model equivariance, or rely on networks that are not purely equivariant (Spezialetti et al., 2020) to predict characteristic orientations, which therefore yield suboptimal stability. More recent methods propose to disentangle the shape and pose of point clouds by yielding rotationinvariant shape representation and rotation-equivariant pose information (Sajnani et al., 2022; Katzir et al., 2022). These methods are tightly coupled with a class-sensitive point cloud reconstruction pipeline (Groueix et al., 2018), consequently making them unsuitable for multi-class training. While they may be stable, such approaches often fail to obtain consistent orientation predictions.
With recent advances in deep learning, there have been successful attempts to reason about 3D geometric data such as point clouds or meshes (Qi et al., 2017a; Wang et al., 2019; Wu et al., 2019). These attempts have eschewed the need to rely on handcrafted features (Weinmann et al.,2015) to perform various tasks, including 3D object recognition, 3D semantic segmentation, and point cloud registration. However, a consistent inference with respect to variations of input point clouds remains challenging, especially when an arbitrary rotation is involved in the point clouds to handle. While there exist methods to yield rotation-robust representations by employing rotation-equivariant (Shen et al., 2020; Deng et al., 2021) or rotation-invariant (Sun et al.,
$^{*}$Equal contribution$^{\text{1}}$Graduate School of Artificial Intelligence, POSTECH, Pohang, South Korea $^{2}$Department of Computer Science and Engineering, POSTECH, Pohang, South Korea. Correspondence to: Minsu Cho <mscho@ postech.ac.kr>.
In this work, we introduce an effective method to learn the characteristic orientation of point clouds achieving both stability and consistency. The idea is to decouple the shape information of a point cloud into two factors, pure geometry and semantics, and coordinate them by means of SO(3)- equivariant learning and SO(3)-invariant residual learning. Specifically, we predict a SO(3)-equivariant characteristic
Proceedings of the 40$^{th}$ International Conference on Machine $Learning$,Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright 2023 by the author(s).
1
|
[
[
0.158447265625,
0.1142578125,
0.81396484375,
0.1142578125,
0.81396484375,
0.1287841796875,
0.158447265625,
0.1287841796875
],
[
0.324462890625,
0.1337890625,
0.64794921875,
0.1363525390625,
0.64794921875,
0.154052734375,
0.324462890625,
0.1514892578125
],
[
0.1690673828125,
0.1988525390625,
0.79736328125,
0.1976318359375,
0.79736328125,
0.21533203125,
0.1690673828125,
0.216552734375
],
[
0.24267578125,
0.24365234375,
0.318603515625,
0.24365234375,
0.318603515625,
0.25830078125,
0.24267578125,
0.25830078125
],
[
0.5009765625,
0.2442626953125,
0.8857421875,
0.2442626953125,
0.8857421875,
0.25830078125,
0.5009765625,
0.25830078125
],
[
0.12091064453125,
0.26513671875,
0.44287109375,
0.26513671875,
0.44287109375,
0.279052734375,
0.12091064453125,
0.279052734375
],
[
0.50244140625,
0.261962890625,
0.8837890625,
0.261962890625,
0.8837890625,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.5,
0.27587890625,
0.88623046875,
0.27587890625,
0.88623046875,
0.2890625,
0.5,
0.2890625
],
[
0.0228729248046875,
0.266357421875,
0.05963134765625,
0.266357421875,
0.05963134765625,
0.399658203125,
0.0228729248046875,
0.399658203125
],
[
0.12091064453125,
0.281005859375,
0.44287109375,
0.281005859375,
0.44287109375,
0.294189453125,
0.12091064453125,
0.294189453125
],
[
0.1217041015625,
0.296630859375,
0.441162109375,
0.296630859375,
0.441162109375,
0.31005859375,
0.1217041015625,
0.31005859375
],
[
0.50244140625,
0.292236328125,
0.8857421875,
0.292236328125,
0.8857421875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.1217041015625,
0.311279296875,
0.441162109375,
0.311279296875,
0.441162109375,
0.324462890625,
0.1217041015625,
0.324462890625
],
[
0.5009765625,
0.30615234375,
0.88623046875,
0.30615234375,
0.88623046875,
0.3193359375,
0.5009765625,
0.3193359375
],
[
0.12091064453125,
0.324462890625,
0.441162109375,
0.32568359375,
0.441162109375,
0.34033203125,
0.12091064453125,
0.339111328125
],
[
0.5009765625,
0.3212890625,
0.88623046875,
0.3212890625,
0.88623046875,
0.335205078125,
0.5009765625,
0.335205078125
],
[
0.1217041015625,
0.341552734375,
0.441162109375,
0.341552734375,
0.441162109375,
0.354736328125,
0.1217041015625,
0.354736328125
],
[
0.5009765625,
0.3359375,
0.88623046875,
0.3359375,
0.88623046875,
0.349853515625,
0.5009765625,
0.349853515625
],
[
0.12091064453125,
0.356689453125,
0.441162109375,
0.356689453125,
0.441162109375,
0.37060546875,
0.12091064453125,
0.37060546875
],
[
0.5009765625,
0.3515625,
0.8857421875,
0.3515625,
0.8857421875,
0.365478515625,
0.5009765625,
0.365478515625
],
[
0.12091064453125,
0.371826171875,
0.441162109375,
0.371826171875,
0.441162109375,
0.385009765625,
0.12091064453125,
0.385009765625
],
[
0.5009765625,
0.3662109375,
0.88623046875,
0.3662109375,
0.88623046875,
0.380126953125,
0.5009765625,
0.380126953125
],
[
0.1217041015625,
0.386474609375,
0.441162109375,
0.386474609375,
0.441162109375,
0.399658203125,
0.1217041015625,
0.399658203125
],
[
0.50244140625,
0.38330078125,
0.8857421875,
0.38330078125,
0.8857421875,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.03594970703125,
0.39453125,
0.0482177734375,
0.39453125,
0.0482177734375,
0.40087890625,
0.03594970703125,
0.40087890625
],
[
0.5009765625,
0.397705078125,
0.8857421875,
0.397705078125,
0.8857421875,
0.40771484375,
0.5009765625,
0.40771484375
],
[
0.1217041015625,
0.40283203125,
0.439453125,
0.40283203125,
0.439453125,
0.41357421875,
0.1217041015625,
0.41357421875
],
[
0.021240234375,
0.3984375,
0.058013916015625,
0.3984375,
0.060455322265625,
0.492431640625,
0.0236968994140625,
0.492431640625
],
[
0.1217041015625,
0.416748046875,
0.441162109375,
0.416748046875,
0.441162109375,
0.429931640625,
0.1217041015625,
0.429931640625
],
[
0.5009765625,
0.41357421875,
0.8857421875,
0.41357421875,
0.8857421875,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.12091064453125,
0.431884765625,
0.44287109375,
0.431884765625,
0.44287109375,
0.44580078125,
0.12091064453125,
0.44580078125
],
[
0.5009765625,
0.4267578125,
0.88623046875,
0.4267578125,
0.88623046875,
0.43994140625,
0.5009765625,
0.43994140625
],
[
0.12091064453125,
0.4462890625,
0.44287109375,
0.4462890625,
0.44287109375,
0.460205078125,
0.12091064453125,
0.460205078125
],
[
0.50244140625,
0.443115234375,
0.8857421875,
0.443115234375,
0.8857421875,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.12091064453125,
0.462158203125,
0.441162109375,
0.462158203125,
0.441162109375,
0.47607421875,
0.12091064453125,
0.47607421875
],
[
0.5009765625,
0.45703125,
0.88623046875,
0.45703125,
0.88623046875,
0.47021484375,
0.5009765625,
0.47021484375
],
[
0.0196075439453125,
0.477783203125,
0.056365966796875,
0.477783203125,
0.060455322265625,
0.7119140625,
0.0236968994140625,
0.7119140625
],
[
0.5,
0.472900390625,
0.88623046875,
0.472900390625,
0.88623046875,
0.486083984375,
0.5,
0.486083984375
],
[
0.12335205078125,
0.479248046875,
0.439453125,
0.479248046875,
0.439453125,
0.4892578125,
0.12335205078125,
0.4892578125
],
[
0.12091064453125,
0.492431640625,
0.441162109375,
0.492431640625,
0.441162109375,
0.505859375,
0.12091064453125,
0.505859375
],
[
0.5009765625,
0.4873046875,
0.88623046875,
0.4873046875,
0.88623046875,
0.50048828125,
0.5009765625,
0.50048828125
],
[
0.5,
0.50048828125,
0.88623046875,
0.501953125,
0.88623046875,
0.5166015625,
0.5,
0.51513671875
],
[
0.1217041015625,
0.50830078125,
0.441162109375,
0.50830078125,
0.441162109375,
0.521484375,
0.1217041015625,
0.521484375
],
[
0.1192626953125,
0.52294921875,
0.416748046875,
0.521484375,
0.416748046875,
0.5361328125,
0.1192626953125,
0.537109375
],
[
0.5009765625,
0.517578125,
0.7099609375,
0.517578125,
0.7099609375,
0.53076171875,
0.5009765625,
0.53076171875
],
[
0.5009765625,
0.5390625,
0.8857421875,
0.5390625,
0.8857421875,
0.5537109375,
0.5009765625,
0.5537109375
],
[
0.5,
0.5537109375,
0.8857421875,
0.5537109375,
0.8857421875,
0.568359375,
0.5,
0.568359375
],
[
0.08905029296875,
0.56494140625,
0.2156982421875,
0.56494140625,
0.2156982421875,
0.580078125,
0.08905029296875,
0.580078125
],
[
0.5009765625,
0.5712890625,
0.8857421875,
0.5712890625,
0.8857421875,
0.58203125,
0.5009765625,
0.58203125
],
[
0.0882568359375,
0.5888671875,
0.473876953125,
0.59033203125,
0.473876953125,
0.60498046875,
0.0882568359375,
0.60400390625
],
[
0.5009765625,
0.58544921875,
0.8857421875,
0.58544921875,
0.8857421875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.0882568359375,
0.60595703125,
0.473876953125,
0.60498046875,
0.473876953125,
0.619140625,
0.0882568359375,
0.62060546875
],
[
0.5009765625,
0.6015625,
0.8837890625,
0.6015625,
0.8837890625,
0.6123046875,
0.5009765625,
0.6123046875
],
[
0.5009765625,
0.61572265625,
0.8857421875,
0.61572265625,
0.8857421875,
0.62890625,
0.5009765625,
0.62890625
],
[
0.0882568359375,
0.6220703125,
0.473876953125,
0.6220703125,
0.473876953125,
0.63525390625,
0.0882568359375,
0.63525390625
],
[
0.08905029296875,
0.63623046875,
0.473876953125,
0.63623046875,
0.473876953125,
0.6494140625,
0.08905029296875,
0.6494140625
],
[
0.5,
0.630859375,
0.88623046875,
0.630859375,
0.88623046875,
0.64453125,
0.5,
0.64453125
],
[
0.0882568359375,
0.65087890625,
0.473876953125,
0.65087890625,
0.473876953125,
0.6640625,
0.0882568359375,
0.6640625
],
[
0.5,
0.64599609375,
0.88623046875,
0.64599609375,
0.88623046875,
0.6591796875,
0.5,
0.6591796875
],
[
0.0882568359375,
0.66650390625,
0.4755859375,
0.66650390625,
0.4755859375,
0.6796875,
0.0882568359375,
0.6796875
],
[
0.50244140625,
0.662109375,
0.8837890625,
0.662109375,
0.8837890625,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.5,
0.67529296875,
0.88623046875,
0.67529296875,
0.88623046875,
0.689453125,
0.5,
0.689453125
],
[
0.0882568359375,
0.681640625,
0.4755859375,
0.681640625,
0.4755859375,
0.69580078125,
0.0882568359375,
0.69580078125
],
[
0.496826171875,
0.68994140625,
0.88623046875,
0.689453125,
0.88623046875,
0.7041015625,
0.496826171875,
0.705078125
],
[
0.08905029296875,
0.6962890625,
0.476318359375,
0.6962890625,
0.476318359375,
0.71044921875,
0.08905029296875,
0.71044921875
],
[
0.08905029296875,
0.7119140625,
0.473876953125,
0.7119140625,
0.473876953125,
0.72607421875,
0.08905029296875,
0.72607421875
],
[
0.5009765625,
0.70849609375,
0.8857421875,
0.70849609375,
0.8857421875,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.08905029296875,
0.7265625,
0.4697265625,
0.7265625,
0.4697265625,
0.74072265625,
0.08905029296875,
0.74072265625
],
[
0.5,
0.7216796875,
0.88623046875,
0.7216796875,
0.88623046875,
0.73486328125,
0.5,
0.73486328125
],
[
0.50244140625,
0.7373046875,
0.8857421875,
0.7373046875,
0.8857421875,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.0882568359375,
0.75,
0.4755859375,
0.75,
0.4755859375,
0.76318359375,
0.0882568359375,
0.76318359375
],
[
0.5,
0.75048828125,
0.88623046875,
0.75048828125,
0.88623046875,
0.7646484375,
0.5,
0.7646484375
],
[
0.0882568359375,
0.76513671875,
0.4755859375,
0.76513671875,
0.4755859375,
0.77880859375,
0.0882568359375,
0.77880859375
],
[
0.496826171875,
0.76513671875,
0.76708984375,
0.76513671875,
0.76708984375,
0.77880859375,
0.496826171875,
0.77880859375
],
[
0.0882568359375,
0.77880859375,
0.473876953125,
0.77880859375,
0.473876953125,
0.79345703125,
0.0882568359375,
0.79345703125
],
[
0.5009765625,
0.78857421875,
0.88623046875,
0.78857421875,
0.88623046875,
0.8017578125,
0.5009765625,
0.8017578125
],
[
0.108642578125,
0.8037109375,
0.4755859375,
0.8037109375,
0.4755859375,
0.8173828125,
0.108642578125,
0.8173828125
],
[
0.5,
0.8037109375,
0.88623046875,
0.8037109375,
0.88623046875,
0.8173828125,
0.5,
0.8173828125
],
[
0.08660888671875,
0.81494140625,
0.4755859375,
0.81640625,
0.4755859375,
0.83056640625,
0.08660888671875,
0.830078125
],
[
0.498291015625,
0.818359375,
0.8857421875,
0.818359375,
0.8857421875,
0.83203125,
0.498291015625,
0.83203125
],
[
0.0906982421875,
0.83056640625,
0.47216796875,
0.83056640625,
0.47216796875,
0.84130859375,
0.0906982421875,
0.84130859375
],
[
0.08660888671875,
0.84033203125,
0.400390625,
0.84130859375,
0.400390625,
0.85595703125,
0.08660888671875,
0.85498046875
],
[
0.5,
0.83544921875,
0.8837890625,
0.83544921875,
0.8837890625,
0.8486328125,
0.5,
0.8486328125
],
[
0.5,
0.849609375,
0.88623046875,
0.849609375,
0.88623046875,
0.86279296875,
0.5,
0.86279296875
],
[
0.08660888671875,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.5009765625,
0.86572265625,
0.88623046875,
0.86572265625,
0.88623046875,
0.87890625,
0.5009765625,
0.87890625
],
[
0.08905029296875,
0.8798828125,
0.4755859375,
0.8798828125,
0.4755859375,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.5,
0.8798828125,
0.88623046875,
0.8798828125,
0.88623046875,
0.89306640625,
0.5,
0.89306640625
],
[
0.0882568359375,
0.892578125,
0.221435546875,
0.892578125,
0.221435546875,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.483642578125,
0.9267578125,
0.49267578125,
0.9267578125,
0.49267578125,
0.9345703125,
0.483642578125,
0.9345703125
]
] |
[
"Stable and Consistent Prediction of 3D Characteristic Orientation",
"via Invariant Residual Learning",
"Seungwook Kim*1 Chunghyun Park*1 Yoonwoo Jeong' Jaesik Park 12 Minsu Cho 12",
"Abstract",
"2019; Yu et al., 2020; Li et al., 2021b) networks, the yielded",
"Learning to predict reliable characteristic orienta-",
"representationssufferfromlossofexpressivityinreturnfor",
"robustness to rotation. A more straightforward and intu-",
"",
"tions of 3D point clouds is an important yet chal-",
"lenging problem, as different point clouds of the",
"itivealternative is to canonicalize the pointclouds by deter-",
"same class may have largely varying appearances.",
"mining their characteristic orientation (Fang et al., 2020;",
"In this work, we introduce a novel method to de-",
"Katzir et al., 2022); the canonicalized point cloud, which",
"couple the shape geometry and semantics of the",
"is obtained by canceling out the characteristic orientation,",
"input point cloud to achieve both stability and con-",
"preserves its shape information and is readily used for down",
"sistency. The proposed method integrates shape-",
"stream tasks without loss of expressivity. Although it is non-",
"geometry-based SO(3)-equivariant learning and",
"trivial todefinethecanonical orientation of apoint cloud.",
"",
"thereexist its desiderata(Katziretal.,2022):stability-the",
"shape-semantics-basedSO(3)-invariantresidual",
"[A'S]",
"learning, where a final characteristic orientation",
"canonical orientation is invariantto anyrotation transfor-",
"is obtained by calibrating an SO(3)-equivariant",
"mations of the input point cloud - and consistentency - the",
"orientation hypothesis using an SO(3)-invariant",
"canonical orientation is invariant to intra-class variations",
"residual rotation. In experiments, the proposed",
"i.e., point clouds of the same class should have the same",
"1090",
"canonical pose. Recent methods to canonicalize the point",
"methodnotonlydemonstratessuperiorstability",
"and consistency but also exhibits state-of-the-art",
"clouds commonly exploit these desiderata to train their net-",
"work in a self-supervised manner (Sun et al., 2021; Sajnani",
"performances when applied to point cloud part",
"segmentation, given randomly rotated inputs.",
"et al., 2022; Katzir et al., 2022).",
"Existing methods to predict the characteristic orientation",
"of point clouds, however, suffer from satisfying both of the",
"1. Introduction",
"criteria,stability and consistency,simultaneously.Earlier",
"With recent advances in deep learning, there have been",
"methods either rely on heavy rotation augmentation (Sun",
"successful attempts to reason about 3D geometric data such",
"etal.,2021)tomodel equivariance,or relyon networks that",
"are not purely equivariant (Spezialetti et al., 2020) to predict",
"as point clouds or meshes (Qi et al., 2017a; Wang et al.,",
"2019; Wu et al., 2019). These attempts have eschewed the",
"characteristic orientations, which therefore yield subopti-",
"need to rely on handcrafted features (Weinmann et al., 2015)",
"mal stability. More recent methods propose to disentangle",
"to perform various tasks, including 3D object recognition,",
"the shape and pose of point cloudsby yieldingrotation",
"invariant shape representation and rotation-equivariant pose",
"3D semantic segmentation, and point cloud registration.",
"information (Sajnani et al., 2022; Katzir et al., 2022). These",
"However, a consistent inference with respect to variations of",
"input point clouds remains challenging, especially when an",
"methods aretightly coupled with a class-sensitive point",
"arbitrary rotation is involved in the point clouds to handle.",
"cloud reconstruction pipeline (Groueix et al., 2018), con-",
"sequentlymakingthemunsuitableformulti-classtraining",
"While there exist methods to yield rotation-robust repre-",
"While they may be stable, such approaches often fail to",
"sentations by employing rotation-equivariant (Shen et al.,",
" obtain consistent orientation predictions.",
"2020; Deng et al., 2021) or rotation-invariant (Sun et al.,",
"In this work, we introduce an effective method to learn the",
"Equal contribution 'Graduate School of Artificial Intelligence,",
"characteristic orientation of point clouds achieving both sta-",
"POSTECH, Pohang, South Korea ‘Department of Computer Sci-",
"bility and consistency. The idea is to decouple the shape",
"enceandEngineering,POSTECH,Pohang,SouthKorea.Corre",
"spondence to: Minsu Cho <mscho@postech.ac.kr>",
"information of a point cloud into two factors, pure geometry",
"and semantics, and coordinate them by means of sO(3)-",
"Proceedings of the 4oth International Conference on Machine",
"equivariant learning and SO(3)-invariant residual learning.",
"Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright",
"Specifically, we predict a SO(3)-equivariant characteristic",
"2023 by the author(s).",
"1"
] |
{
"arxiv_id": "2306.11406",
"language": "en",
"page_idx": 0
}
|
||
2306.11406-en-1
|
2306.11406-en-1.jpg
|
Stable and Consistent Prediction of 3D Characteristic Orientation via Invariant Residual Learning
et al., 2018; Chen et al., 2021). To introduce a few recent successes, YOHO (Wang et al., 2022) exploits the symmetry of icosahedral group to extract group-equivariant representations. Luo et al.(2022) propose the idea of equivariant message passing by learning point-wise orientations. Vector Neuron Networks (Deng et al., 2021) lift the latent representations from scalar vectors to matrix vectors to implement a fully equivariant network.
orientation hypothesis and calibrate it using SO(3)-invariant residual rotation to yield a final prediction for stable and consistent canonicalization. We name the proposed CHaracteristic Orientation predictor with Invariant Residual learning as CHOIR. As it dispenses with the widely used point cloud reconstruction in learning (Sajnani et al., 2022; Katzir et al., 2022), CHOIR facilitates multi-class training, being more practical for downstream point cloud analysis tasks.
Another thread of work explores achieving strict rotational invariance $i.e.$, yielding the same output representation regardless of the rotation of the input point cloud. Many existing work focus on handcrafting rotation-invariant features based on intrinsic geometries (Sun et al., 2019; Li et al., 2021b; Xiao & Wachs, 2021). While the above methods suffer from information loss, some approaches propose to leverage PCA-based canonical poses to achieve rotational invariance (Kim et al., 2020; Yu et al., 2020; Li et al., 2021a). However, PCA-based canonical poses suffer from sign and order ambiguities (Li et al., 2021a), and considering all the possible candidates for the canonical pose leads to high computational overhead.
In our work, we propose to leverage SO(3)-equivariant networks for learning the shape geometry, and SO(3)-invariant networks for learning the shape semantics, to predict the
The contribution of our work is fourfold.(1) We introduce CHOIR, a strong characteristic orientation prediction method that decouples shape information of point clouds to geometry and semantics to achieve both stability and consistency. (2) We propose a novel residual predictor that learns to output a SO(3)-invariant residual rotation to calibrate our initial orientation hypotheses for better consistency. (3) In experiments, we not only demonstrate state-of-the-art stability and consistency, but also show superior performance when applied to the task of point cloud part segmentation. (4) We facilitate multi-class training by removing point cloud reconstruction, enhancing applicablity to downstream tasks, (5) We analyze the influence of consistency and stability on downstream tasks, which indicates that the balance between stability and consistency is the key to canonicalizing point clouds for downstream tasks.
characteristic orientation of point clouds. Canonicalizing the input point cloud using our predicted characteristic orientation enables us to preserve the shape information without having to consider various ambiguities.
## $2. \textbf{Related Work}$
Learning based point cloud analysis. Over the recent years, various methods have been proposed to handle point clouds, which pose difficulties due to their irregularity and lack of order. PointNet (Qi et al., 2017a) and its followup methods (Qi et al., 2017b; Wu et al., 2019; Thomas et al., 2019; Wang et al., 2019) propose to input point clouds directly to networks to achieve permutation-invariant representations of the point cloud. For example, DGCNN (Wang et al., 2019) uses graph convolution using pointwise knearest-neighbors, while KPConv (Thomas et al., 2019) defines the convolution area using a set of kernel points. More recently, Point Transformer (Zhao et al., 2021) proposed to leverage the transformer architecture to process point clouds, while PointMixer (Choe et al., 2022) facilitates information sharing between points by replacing token-mixing MLPs with Softmax function. While these methods have shown impressive performances on point cloud analysis tasks including classification and segmentation, these methods rely on extensive train-time or test-time rotation augmentations to handle random rotations encountered during inference.
Self-supervised category-level 6D pose estimation. The task of category-level pose estimation aims to find the 6D poses of unseen instances from known categories, without access to object CAD models. A similar approach called shape co-alignment (Averkiou et al., 2016; Mehr et al., 2018) proposes to obtain consistent category-level reference frames, which require clean mesh inputs or rely heavily on the symmetry of point clouds. QEC (Zhao et al., 2020) proposes a novel dynamic routing procedure on quarternions to establish end-to-end transformation equivariance, but QEC does not evaluate the predicted category-level pose, and assumes complete shapes. EquiPose (Li et al., 2021c) proposes to handle both complete and partial shapes, and introduces a self-supervised learning framework to estimate several category-level object pose candidates from single 3D point clouds by leveraging EPNs (Chen et al., 2021).
While characteristic orientation prediction and 6D pose estimation tasks are pertinent, 6D pose estimation, unlike characteristic orientation prediction, has ground-truth categorylevel pose annotations that uniquely define a category-level pose reference frame.
Rotation-robust point cloud analysis. To minimize the need for intensive rotation augmentations, one thread of work aims to design rotation-equivariant modules, such that the output representations will be equivariant with respect to the rotation of the input point cloud (Esteves et al., 2018; Rao et al. , 2019; Shen et al. , 2020; Fang et al. , 2020; Thomas
## Estimation of characteristic orientation. Previously, some work proposed to learn the canonical frame of objects using
2
|
[
[
0.17724609375,
0.05682373046875,
0.79736328125,
0.05682373046875,
0.79736328125,
0.0706787109375,
0.17724609375,
0.0706787109375
],
[
0.0882568359375,
0.085205078125,
0.4755859375,
0.085205078125,
0.4755859375,
0.0985107421875,
0.0882568359375,
0.0985107421875
],
[
0.498291015625,
0.085205078125,
0.8857421875,
0.085205078125,
0.8857421875,
0.0997314453125,
0.498291015625,
0.0997314453125
],
[
0.0882568359375,
0.10162353515625,
0.4755859375,
0.10162353515625,
0.4755859375,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.496826171875,
0.0997314453125,
0.8837890625,
0.10101318359375,
0.8837890625,
0.11553955078125,
0.496826171875,
0.1142578125
],
[
0.0882568359375,
0.11614990234375,
0.476318359375,
0.11614990234375,
0.476318359375,
0.1307373046875,
0.0882568359375,
0.1307373046875
],
[
0.496826171875,
0.11553955078125,
0.88623046875,
0.11614990234375,
0.88623046875,
0.1307373046875,
0.496826171875,
0.1300048828125
],
[
0.08905029296875,
0.1307373046875,
0.473876953125,
0.1307373046875,
0.473876953125,
0.14453125,
0.08905029296875,
0.14453125
],
[
0.498291015625,
0.1319580078125,
0.88623046875,
0.1319580078125,
0.88623046875,
0.1458740234375,
0.498291015625,
0.1458740234375
],
[
0.0882568359375,
0.1458740234375,
0.4755859375,
0.146484375,
0.4755859375,
0.1610107421875,
0.0882568359375,
0.160400390625
],
[
0.498291015625,
0.147705078125,
0.88623046875,
0.147705078125,
0.88623046875,
0.1610107421875,
0.498291015625,
0.1610107421875
],
[
0.08660888671875,
0.1610107421875,
0.4755859375,
0.160400390625,
0.4755859375,
0.1749267578125,
0.08660888671875,
0.175537109375
],
[
0.498291015625,
0.1622314453125,
0.88623046875,
0.1622314453125,
0.88623046875,
0.175537109375,
0.498291015625,
0.175537109375
],
[
0.0882568359375,
0.1767578125,
0.473876953125,
0.1767578125,
0.473876953125,
0.1900634765625,
0.0882568359375,
0.1900634765625
],
[
0.5009765625,
0.177978515625,
0.88623046875,
0.177978515625,
0.88623046875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.0882568359375,
0.1925048828125,
0.46728515625,
0.1925048828125,
0.46728515625,
0.205810546875,
0.0882568359375,
0.205810546875
],
[
0.5,
0.1925048828125,
0.67236328125,
0.1925048828125,
0.67236328125,
0.205810546875,
0.5,
0.205810546875
],
[
0.0882568359375,
0.2139892578125,
0.4755859375,
0.2139892578125,
0.4755859375,
0.2279052734375,
0.0882568359375,
0.2279052734375
],
[
0.5009765625,
0.2139892578125,
0.88623046875,
0.2139892578125,
0.88623046875,
0.2279052734375,
0.5009765625,
0.2279052734375
],
[
0.0882568359375,
0.2298583984375,
0.4755859375,
0.2298583984375,
0.4755859375,
0.24365234375,
0.0882568359375,
0.24365234375
],
[
0.5,
0.2298583984375,
0.88818359375,
0.2298583984375,
0.88818359375,
0.24365234375,
0.5,
0.24365234375
],
[
0.0882568359375,
0.2442626953125,
0.4755859375,
0.2442626953125,
0.4755859375,
0.25830078125,
0.0882568359375,
0.25830078125
],
[
0.5,
0.2442626953125,
0.88623046875,
0.2442626953125,
0.88623046875,
0.2587890625,
0.5,
0.2587890625
],
[
0.08905029296875,
0.260009765625,
0.4755859375,
0.260009765625,
0.4755859375,
0.2734375,
0.08905029296875,
0.2734375
],
[
0.5009765625,
0.260009765625,
0.88623046875,
0.260009765625,
0.88623046875,
0.2734375,
0.5009765625,
0.2734375
],
[
0.08660888671875,
0.274658203125,
0.4755859375,
0.2734375,
0.4755859375,
0.28857421875,
0.08660888671875,
0.2890625
],
[
0.5009765625,
0.274658203125,
0.88623046875,
0.274658203125,
0.88623046875,
0.28857421875,
0.5009765625,
0.28857421875
],
[
0.08905029296875,
0.290283203125,
0.473876953125,
0.290283203125,
0.473876953125,
0.3037109375,
0.08905029296875,
0.3037109375
],
[
0.5009765625,
0.290283203125,
0.88623046875,
0.290283203125,
0.88623046875,
0.3037109375,
0.5009765625,
0.3037109375
],
[
0.0882568359375,
0.304931640625,
0.47216796875,
0.304931640625,
0.47216796875,
0.318115234375,
0.0882568359375,
0.318115234375
],
[
0.5,
0.30615234375,
0.88623046875,
0.30615234375,
0.88623046875,
0.3193359375,
0.5,
0.3193359375
],
[
0.08905029296875,
0.32080078125,
0.4755859375,
0.32080078125,
0.4755859375,
0.333984375,
0.08905029296875,
0.333984375
],
[
0.5,
0.3193359375,
0.8857421875,
0.3193359375,
0.8857421875,
0.333251953125,
0.5,
0.333251953125
],
[
0.0906982421875,
0.337158203125,
0.47216796875,
0.337158203125,
0.47216796875,
0.347900390625,
0.0906982421875,
0.347900390625
],
[
0.5,
0.335205078125,
0.88623046875,
0.335205078125,
0.88623046875,
0.348388671875,
0.5,
0.348388671875
],
[
0.0882568359375,
0.35107421875,
0.473876953125,
0.35107421875,
0.473876953125,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.496826171875,
0.349853515625,
0.8837890625,
0.349853515625,
0.8837890625,
0.363037109375,
0.496826171875,
0.363037109375
],
[
0.0882568359375,
0.3642578125,
0.473876953125,
0.3642578125,
0.473876953125,
0.37744140625,
0.0882568359375,
0.37744140625
],
[
0.498291015625,
0.365478515625,
0.8857421875,
0.365478515625,
0.8857421875,
0.37890625,
0.498291015625,
0.37890625
],
[
0.0882568359375,
0.380615234375,
0.4755859375,
0.380615234375,
0.4755859375,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.5,
0.380615234375,
0.88623046875,
0.380615234375,
0.88623046875,
0.39453125,
0.5,
0.39453125
],
[
0.0882568359375,
0.395751953125,
0.47802734375,
0.395751953125,
0.47802734375,
0.4091796875,
0.0882568359375,
0.4091796875
],
[
0.5,
0.395751953125,
0.6640625,
0.395751953125,
0.6640625,
0.4091796875,
0.5,
0.4091796875
],
[
0.08905029296875,
0.412353515625,
0.473876953125,
0.412353515625,
0.473876953125,
0.42236328125,
0.08905029296875,
0.42236328125
],
[
0.498291015625,
0.416748046875,
0.88623046875,
0.417236328125,
0.88623046875,
0.431884765625,
0.498291015625,
0.43115234375
],
[
0.0882568359375,
0.425537109375,
0.4755859375,
0.425537109375,
0.4755859375,
0.439453125,
0.0882568359375,
0.439453125
],
[
0.5,
0.434326171875,
0.88623046875,
0.434326171875,
0.88623046875,
0.447509765625,
0.5,
0.447509765625
],
[
0.0882568359375,
0.44140625,
0.343017578125,
0.44140625,
0.343017578125,
0.455078125,
0.0882568359375,
0.455078125
],
[
0.5,
0.448974609375,
0.8857421875,
0.448974609375,
0.8857421875,
0.462158203125,
0.5,
0.462158203125
],
[
0.5009765625,
0.464111328125,
0.88623046875,
0.464111328125,
0.88623046875,
0.477783203125,
0.5009765625,
0.477783203125
],
[
0.08905029296875,
0.474853515625,
0.226318359375,
0.474853515625,
0.226318359375,
0.4892578125,
0.08905029296875,
0.4892578125
],
[
0.5009765625,
0.479248046875,
0.88623046875,
0.479248046875,
0.88623046875,
0.492431640625,
0.5009765625,
0.492431640625
],
[
0.50244140625,
0.49560546875,
0.8857421875,
0.49560546875,
0.8857421875,
0.505859375,
0.50244140625,
0.505859375
],
[
0.08905029296875,
0.50390625,
0.473876953125,
0.50390625,
0.473876953125,
0.517578125,
0.08905029296875,
0.517578125
],
[
0.5,
0.51025390625,
0.7587890625,
0.51025390625,
0.7587890625,
0.5234375,
0.5,
0.5234375
],
[
0.08905029296875,
0.51953125,
0.473876953125,
0.51953125,
0.473876953125,
0.53271484375,
0.08905029296875,
0.53271484375
],
[
0.0882568359375,
0.53515625,
0.473876953125,
0.53515625,
0.473876953125,
0.548828125,
0.0882568359375,
0.548828125
],
[
0.5,
0.5361328125,
0.8857421875,
0.537109375,
0.8857421875,
0.5517578125,
0.5,
0.55029296875
],
[
0.0882568359375,
0.5498046875,
0.4755859375,
0.5498046875,
0.4755859375,
0.56298828125,
0.0882568359375,
0.56298828125
],
[
0.5,
0.55322265625,
0.8837890625,
0.55322265625,
0.8837890625,
0.56640625,
0.5,
0.56640625
],
[
0.08905029296875,
0.564453125,
0.473876953125,
0.564453125,
0.473876953125,
0.57763671875,
0.08905029296875,
0.57763671875
],
[
0.5009765625,
0.568359375,
0.88623046875,
0.568359375,
0.88623046875,
0.58203125,
0.5009765625,
0.58203125
],
[
0.0882568359375,
0.580078125,
0.473876953125,
0.580078125,
0.473876953125,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.5,
0.583984375,
0.8857421875,
0.583984375,
0.8857421875,
0.59765625,
0.5,
0.59765625
],
[
0.0906982421875,
0.5966796875,
0.473876953125,
0.5966796875,
0.473876953125,
0.60693359375,
0.0906982421875,
0.60693359375
],
[
0.5,
0.5986328125,
0.8857421875,
0.5986328125,
0.8857421875,
0.6123046875,
0.5,
0.6123046875
],
[
0.08905029296875,
0.609375,
0.4755859375,
0.609375,
0.4755859375,
0.62353515625,
0.08905029296875,
0.62353515625
],
[
0.5,
0.61279296875,
0.88623046875,
0.61279296875,
0.88623046875,
0.62744140625,
0.5,
0.62744140625
],
[
0.0882568359375,
0.625,
0.4755859375,
0.625,
0.4755859375,
0.63818359375,
0.0882568359375,
0.63818359375
],
[
0.5,
0.62744140625,
0.8837890625,
0.62890625,
0.8837890625,
0.642578125,
0.5,
0.6416015625
],
[
0.0882568359375,
0.64013671875,
0.4755859375,
0.64013671875,
0.4755859375,
0.65380859375,
0.0882568359375,
0.65380859375
],
[
0.5,
0.64306640625,
0.88623046875,
0.64306640625,
0.88623046875,
0.6572265625,
0.5,
0.6572265625
],
[
0.0882568359375,
0.65478515625,
0.4755859375,
0.65478515625,
0.4755859375,
0.66845703125,
0.0882568359375,
0.66845703125
],
[
0.5,
0.6591796875,
0.88623046875,
0.6591796875,
0.88623046875,
0.67236328125,
0.5,
0.67236328125
],
[
0.0882568359375,
0.67041015625,
0.4755859375,
0.67041015625,
0.4755859375,
0.6845703125,
0.0882568359375,
0.6845703125
],
[
0.5,
0.673828125,
0.88623046875,
0.673828125,
0.88623046875,
0.6875,
0.5,
0.6875
],
[
0.0906982421875,
0.6875,
0.473876953125,
0.6875,
0.473876953125,
0.69775390625,
0.0906982421875,
0.69775390625
],
[
0.5,
0.689453125,
0.88623046875,
0.689453125,
0.88623046875,
0.70263671875,
0.5,
0.70263671875
],
[
0.0882568359375,
0.70068359375,
0.473876953125,
0.70068359375,
0.473876953125,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.498291015625,
0.7041015625,
0.88623046875,
0.70263671875,
0.88623046875,
0.71728515625,
0.498291015625,
0.71826171875
],
[
0.0882568359375,
0.71630859375,
0.473876953125,
0.71630859375,
0.473876953125,
0.72998046875,
0.0882568359375,
0.72998046875
],
[
0.498291015625,
0.71826171875,
0.8837890625,
0.71826171875,
0.8837890625,
0.73193359375,
0.498291015625,
0.73193359375
],
[
0.0882568359375,
0.73095703125,
0.4755859375,
0.73095703125,
0.4755859375,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.498291015625,
0.73486328125,
0.8857421875,
0.73486328125,
0.8857421875,
0.74853515625,
0.498291015625,
0.74853515625
],
[
0.0882568359375,
0.7470703125,
0.4755859375,
0.7470703125,
0.4755859375,
0.76025390625,
0.0882568359375,
0.76025390625
],
[
0.5009765625,
0.75,
0.8857421875,
0.75,
0.8857421875,
0.76318359375,
0.5009765625,
0.76318359375
],
[
0.0882568359375,
0.76123046875,
0.473876953125,
0.76123046875,
0.473876953125,
0.7744140625,
0.0882568359375,
0.7744140625
],
[
0.5009765625,
0.7646484375,
0.87109375,
0.7646484375,
0.87109375,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.0882568359375,
0.77587890625,
0.473876953125,
0.77587890625,
0.473876953125,
0.7890625,
0.0882568359375,
0.7890625
],
[
0.0882568359375,
0.79150390625,
0.464111328125,
0.79150390625,
0.464111328125,
0.8046875,
0.0882568359375,
0.8046875
],
[
0.5009765625,
0.7861328125,
0.88623046875,
0.7861328125,
0.88623046875,
0.7998046875,
0.5009765625,
0.7998046875
],
[
0.50244140625,
0.8037109375,
0.8857421875,
0.8037109375,
0.8857421875,
0.814453125,
0.50244140625,
0.814453125
],
[
0.08905029296875,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.830078125,
0.08905029296875,
0.830078125
],
[
0.5,
0.8173828125,
0.8857421875,
0.8173828125,
0.8857421875,
0.83056640625,
0.5,
0.83056640625
],
[
0.08905029296875,
0.83349609375,
0.473876953125,
0.83349609375,
0.473876953125,
0.84326171875,
0.08905029296875,
0.84326171875
],
[
0.498291015625,
0.83349609375,
0.88623046875,
0.83349609375,
0.88623046875,
0.8466796875,
0.498291015625,
0.8466796875
],
[
0.0882568359375,
0.8466796875,
0.4755859375,
0.8466796875,
0.4755859375,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.5,
0.84765625,
0.64453125,
0.84765625,
0.64453125,
0.861328125,
0.5,
0.861328125
],
[
0.0906982421875,
0.86279296875,
0.473876953125,
0.86279296875,
0.473876953125,
0.87353515625,
0.0906982421875,
0.87353515625
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.5009765625,
0.87744140625,
0.8837890625,
0.87744140625,
0.8837890625,
0.88818359375,
0.5009765625,
0.88818359375
],
[
0.0882568359375,
0.892578125,
0.476318359375,
0.892578125,
0.476318359375,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.498291015625,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.498291015625,
0.90576171875
],
[
0.476318359375,
0.92529296875,
0.491943359375,
0.921875,
0.496826171875,
0.931640625,
0.48193359375,
0.93603515625
]
] |
[
"Stable and Consistent Prediction of 3D Characteristic Orientation via Invariant Residual Learning",
"orientation hypothesis and calibrate it using SO(3)-invariant",
"et al., 2018; Chen et al., 2021). To introduce a few recent",
"residual rotation to yield a final prediction for stable and",
"successes, YOHO (Wang et al., 2022) exploits the symmetry",
"consistent canonicalization. We name the proposed CHarac-",
"of icosahedral group to extract group-equivariant represen-",
"teristic Orientation predictor with Invariant Residual learn",
"tations. Luo et al. (2022) propose the idea of equivariant",
"ing as CHOIR. As it dispenses with the widely used point",
"message passing by learning point-wise orientations. Vector",
"cloud reconstruction in learning (Sajnani et al., 2022; Katzir",
"Neuron Networks (Deng et al., 2021) lift the latent represen-",
"et al., 2022), CHOIR facilitates multi-class training, being",
"tations from scalar vectors to matrix vectors to implement a",
"more practical for downstream point cloud analysis tasks.",
"fully equivariant network.",
"The contribution of our work is fourfold. (1) We intro-",
"Another thread of work explores achieving strict rotational",
"duce CHOIR, a strong characteristic orientation prediction",
"invariance i.e., yielding the same output representation re-",
"method that decouples shape information of point clouds to",
"gardless of the rotation of the input point cloud. Many",
"geometry and semantics to achieve both stability and consis-",
"existing work focus on handcrafting rotation-invariant fea-",
"tency. (2) We propose a novel residual predictor that learns",
"tures based on intrinsic geometries (Sun et al., 2019; Li et al.,",
"to output a SO(3)-invariant residual rotation to calibrate our",
"2021b; Xiao & Wachs, 2021). While the above methods",
"initial orientation hypotheses for better consistency. (3) In",
"suffer from information loss, some approaches propose to",
"experiments, we not only demonstrate state-of-the-art sta-",
"leverage PCA-based canonical poses to achieve rotational",
"bilityandconsistency,butalso showsuperiorperformance",
"invariance (Kim et al., 2020; Yu et al., 2020; Li et al., 2021a)",
"when applied to the task of point cloud part segmentation.",
"However, PCA-based canonical poses suffer from sign and",
"(4) We facilitate multi-class training by removing point",
"order ambiguities (Li et al., 202la), and considering all the",
"cloud reconstruction, enhancing applicablity to downstream",
"possible candidates for the canonical pose leads to high",
"tasks, (5) We analyze the influence of consistency and sta-",
"computational overhead.",
"bilityondownstreamtasks,whichindicatesthatthebalance",
"In our work, we propose to leverage SO(3)-equivariant net-",
"between stability and consistency is the key to canonicaliz-",
"works for learning the shape geometry, and SO(3)-invariant",
"ing point clouds for downstream tasks.",
"networks for learning the shape semantics, to predict the",
"characteristic orientation of point clouds. Canonicalizing",
"2. Related Work",
"the input point cloud using our predicted characteristic orien-",
"tationenablesustopreservetheshapeinformationwithout",
"Learning based point cloud analysis. Over the recent",
"having to consider various ambiguities.",
"years, various methods have been proposed to handle point",
"clouds, which pose difficulties due to their irregularity and",
"Self-supervised category-level 6D pose estimation. The",
"lack of order. PointNet (Qi et al., 2017a) and its follow-",
"task of category-level pose estimation aims to find the 6D",
"up methods (Qi et al., 2017b: Wu et al., 2019: Thomas",
"poses of unseen instances from known categories, with-",
"et al., 2019; Wang et al., 2019) propose to input point clouds",
"out access to object CAD models. A similar approach",
"directlytonetworkstoachievepermutation-invariantrepre-",
"called shape co-alignment (Averkiou et al., 2016; Mehr",
"sentations of the point cloud. For example, DGCNN (Wang",
"et al., 2018) proposes to obtain consistent category-level",
"et al., 2019) uses graph convolution using pointwise k-",
"reference frames, which require clean mesh inputs or rely",
"nearest-neighbors, while KPConv (Thomas et al., 2019) de-",
"heavily on the symmetry of point clouds. QEC (Zhao et al.,",
"fines the convolution area using a set of kernel points. More",
"2020) proposes a novel dynamic routing procedure on quar-",
"recently, Point Transformer (Zhao et al., 2021) proposed to",
"ternions to establish end-to-end transformation equivariance,",
"leveragethetransformerarchitecturetoprocesspointclouds.",
"but QEC does not evaluate the predicted category-level pose,",
"while PointMixer (Choe et al., 2022) facilitates information",
"and assumes complete shapes. EquiPose (Li et al., 2021c)",
"sharing between points by replacing token-mixing MLPs",
"proposes to handle both complete and partial shapes, and",
"with Softmax function. While these methods have shown",
"introduces a self-supervised learning framework to estimate",
"impressive performances on point cloud analysis tasks in-",
"several category-level object pose candidates from single",
"cluding classification and segmentation, these methods rely",
"3D point clouds by leveraging EPNs (Chen et al., 2021).",
"on extensive train-time or test-time rotation augmentations",
"to handle random rotations encountered during inference.",
"While characteristic orientation prediction and 6D pose esti-",
"mationtasksarepertinent,6Dpose estimation,unlike char-",
"Rotation-robust point cloud analysis. To minimize the",
"acteristic orientation prediction, has ground-truth category-",
"need for intensive rotation augmentations,onethread of",
"level pose annotations that uniquely define a category-level",
"work aims to design rotation-equivariant modules, such that",
"pose reference frame.",
"the outputrepresentations will beequivariantwithrespect",
"to the rotation of the input point cloud (Esteves et al., 2018;",
"Estimation ofcharacteristic orientation.Previously,some",
"Rao et al., 2019; Shen et al., 2020; Fang et al., 2020; Thomas",
"work proposed to learn the canonical frame of objects using",
"2"
] |
{
"arxiv_id": "2306.11406",
"language": "en",
"page_idx": 1
}
|
||
1705.09912-en-0
|
1705.09912-en-0.jpg
|
### Multi-channel Weighted Nuclear Norm Minimization for Real Color Image Denoising
Jun Xu$^1$,Lei Zhang$^1,*$,David Zhang$^1$,and Xiangchu Feng$^2$
l Dept. of Computing, The Hong Kong Polytechnic University, Hong Kong, China
$^{2}$School of Mathematics and Statistics, Xidian University, Xi'an, China
$\{$csjunxu,cslzhang, csdzhang$\}$@comp.polyu.edu.hk,xcfeng@mail.xidian.edu.cn
## $\mathbf{Abstract}$
Most of the existing denoising algorithms are developed for grayscale images. It is not trivial to extend them for color image denoising since the noise statistics in R, G, and B channels can be very different for real noisy images. In this paper, we propose a multi-channel (MC) optimization model for real color image denoising under the weighted nuclear norm minimization (WNNM) framework. We concatenate the RGB patches to make use of the channel redundancy, and introduce a weight matrix to balance the data fidelity of the three channels in consideration of their different noise statistics. The proposed MC-WNNM model does not have an analytical solution. We reformulate it into a linear equality-constrained problem and solve it via alternating direction method of multipliers. Each alternative updating step has a closed-form solution and the convergence can be guaranteed. Experiments on both synthetic and real noisy image datasets demonstrate the superiority of the pro- $posedMC-WNNMoverstate-of-the-art$ denoising methods.
ods have also been developed by learning discriminative
priors from pairs of clean and noisy images [9-12].
When the input is a noisy RGB color image, there are mainly three strategies for color image denoising.(1) The first strategy is to apply the grayscale image denoising algorithm to each channel. However, such a straightforward solution will not exploit the spectral correlation among RGB channels, and the denoising performance may not be very satisfying.(2) The second strategy is to transform the RGB image into a less correlated color space, such as YCbCr, and perform denoising in each channel of the transformed space[13, 14]. One representative work along this line is the CBM3D algorithm [14]. However, the color transform will complicate the noise distribution, and the correlation among color channels is not fully exploited.(3) The third strategy is to perform joint denoising on the RGB channels simultaneously for better use of the spectral correlation. For example, the patches from RGB channels are concatenated as a long vector for processing [15, 16].
Though joint denoising of RGB channels is a more promising way for color image denoising, it is not a trivial extension from single channel (grayscale image) to multiple channels (color image). The noise in standard RGB (sRGB) space can be approximately modeled as AWGN, but it has different variances for different channels [17-19] due to the sensor characteristics and on-board processing steps in digital camera pipelines [19, 20]. This makes the real color image denoising problem much more complex. If the three channels are treated equally in the joint denoising process, false colors or artifacts can be generated[15]. How to account for the different noise characteristics in color channels, and how to effectively exploit the within and cross channel correlation are the key issues for designing a good color image denoising method.
This paper presents a new color image denoising algorithm. Considering that the weighted nuclear norm mini-
1
## $1. \textbf{Introduction}$
Image denoising is a classical yet fundamental problem for image quality enhancement in computer vision and photography systems. Most of existing denoising algorithms are designed for grayscale images, aiming to recover the clean image x from its noisy observation y = x+ n, where n is generally assumed to be additive white Gaussian noise (AWGN). State-of-the-art image denoising methods include sparse representation [1], dictionary learning [2], low-rank approximation [3], non-local self-similarity (NSS) [4] based methods, and the combination of those techniques [1-3, 5-8]. Recently, some discriminative denoising meth-
$^{*}$This work is supported by HK RGC GRF grant (PolyU 5313/13E) and
China NSFC grant (no. 61672446).
|
[
[
0.2197265625,
0.1337890625,
0.7509765625,
0.1337890625,
0.7509765625,
0.1483154296875,
0.2197265625,
0.1483154296875
],
[
0.324462890625,
0.154052734375,
0.64453125,
0.1558837890625,
0.64453125,
0.1749267578125,
0.324462890625,
0.17236328125
],
[
0.253173828125,
0.2171630859375,
0.7158203125,
0.2171630859375,
0.7158203125,
0.231689453125,
0.253173828125,
0.231689453125
],
[
0.1641845703125,
0.24560546875,
0.80615234375,
0.246826171875,
0.80615234375,
0.264404296875,
0.1641845703125,
0.26318359375
],
[
0.02777099609375,
0.261474609375,
0.053924560546875,
0.261474609375,
0.053924560546875,
0.281005859375,
0.02777099609375,
0.281005859375
],
[
0.20751953125,
0.264404296875,
0.7646484375,
0.266357421875,
0.7646484375,
0.281005859375,
0.20751953125,
0.279052734375
],
[
0.0236968994140625,
0.271484375,
0.058013916015625,
0.271484375,
0.058013916015625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.1732177734375,
0.285400390625,
0.79736328125,
0.2841796875,
0.79736328125,
0.298583984375,
0.1732177734375,
0.2998046875
],
[
0.236083984375,
0.343505859375,
0.314453125,
0.343505859375,
0.314453125,
0.361083984375,
0.236083984375,
0.361083984375
],
[
0.50244140625,
0.3466796875,
0.89208984375,
0.3466796875,
0.89208984375,
0.361083984375,
0.50244140625,
0.361083984375
],
[
0.0269622802734375,
0.3623046875,
0.0555419921875,
0.3623046875,
0.0555419921875,
0.392578125,
0.0269622802734375,
0.392578125
],
[
0.5009765625,
0.361083984375,
0.8388671875,
0.3623046875,
0.8388671875,
0.376953125,
0.5009765625,
0.375732421875
],
[
0.1004638671875,
0.376953125,
0.4697265625,
0.376953125,
0.4697265625,
0.39013671875,
0.1004638671875,
0.39013671875
],
[
0.521484375,
0.380615234375,
0.89208984375,
0.380615234375,
0.89208984375,
0.39453125,
0.521484375,
0.39453125
],
[
0.080078125,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.50244140625,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
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
],
[
0.08087158203125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50244140625,
0.455078125
],
[
0.0310516357421875,
0.448974609375,
0.058013916015625,
0.448974609375,
0.058013916015625,
0.48095703125,
0.0310516357421875,
0.48095703125
],
[
0.080078125,
0.45263671875,
0.4697265625,
0.45263671875,
0.4697265625,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.50244140625,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.08087158203125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.02532958984375,
0.49365234375,
0.056365966796875,
0.49365234375,
0.056365966796875,
0.54248046875,
0.02532958984375,
0.54248046875
],
[
0.50341796875,
0.488037109375,
0.8896484375,
0.488037109375,
0.8896484375,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.498779296875,
0.4697265625,
0.498779296875,
0.4697265625,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.080078125,
0.51318359375,
0.468017578125,
0.51318359375,
0.468017578125,
0.5263671875,
0.080078125,
0.5263671875
],
[
0.50244140625,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.52783203125,
0.4697265625,
0.52783203125,
0.4697265625,
0.54150390625,
0.080078125,
0.54150390625
],
[
0.50244140625,
0.5322265625,
0.89208984375,
0.5322265625,
0.89208984375,
0.54541015625,
0.50244140625,
0.54541015625
],
[
0.0269622802734375,
0.54052734375,
0.056365966796875,
0.54150390625,
0.053924560546875,
0.58837890625,
0.0236968994140625,
0.5869140625
],
[
0.080078125,
0.54345703125,
0.4697265625,
0.54345703125,
0.4697265625,
0.556640625,
0.080078125,
0.556640625
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.56103515625,
0.890625,
0.56103515625,
0.890625,
0.57470703125,
0.50244140625,
0.57470703125
],
[
0.08087158203125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.50244140625,
0.57763671875,
0.890625,
0.57763671875,
0.890625,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.02532958984375,
0.5859375,
0.056365966796875,
0.5859375,
0.056365966796875,
0.63818359375,
0.02532958984375,
0.63818359375
],
[
0.037567138671875,
0.5859375,
0.049835205078125,
0.5859375,
0.049835205078125,
0.59228515625,
0.037567138671875,
0.59228515625
],
[
0.080078125,
0.5888671875,
0.4697265625,
0.5888671875,
0.4697265625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.080078125,
0.60400390625,
0.4697265625,
0.60400390625,
0.4697265625,
0.6171875,
0.080078125,
0.6171875
],
[
0.50244140625,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.08087158203125,
0.619140625,
0.46728515625,
0.619140625,
0.46728515625,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.021240234375,
0.62744140625,
0.056365966796875,
0.626953125,
0.05963134765625,
0.71044921875,
0.0236968994140625,
0.7109375
],
[
0.50244140625,
0.62353515625,
0.7646484375,
0.62353515625,
0.7646484375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.080078125,
0.6337890625,
0.46728515625,
0.6337890625,
0.46728515625,
0.6484375,
0.080078125,
0.6484375
],
[
0.521484375,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65478515625,
0.521484375,
0.65478515625
],
[
0.50341796875,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50341796875,
0.67041015625
],
[
0.50341796875,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.08087158203125,
0.68310546875,
0.20751953125,
0.68310546875,
0.20751953125,
0.6962890625,
0.08087158203125,
0.6962890625
],
[
0.50341796875,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.1004638671875,
0.70068359375,
0.468017578125,
0.70068359375,
0.468017578125,
0.7138671875,
0.1004638671875,
0.7138671875
],
[
0.50341796875,
0.7021484375,
0.88818359375,
0.7021484375,
0.88818359375,
0.71533203125,
0.50341796875,
0.71533203125
],
[
0.08087158203125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.08087158203125,
0.72998046875
],
[
0.50341796875,
0.71826171875,
0.8896484375,
0.71826171875,
0.8896484375,
0.72900390625,
0.50341796875,
0.72900390625
],
[
0.080078125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.080078125,
0.744140625
],
[
0.5009765625,
0.73193359375,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.5009765625,
0.74560546875
],
[
0.080078125,
0.74560546875,
0.468017578125,
0.74560546875,
0.468017578125,
0.7587890625,
0.080078125,
0.7587890625
],
[
0.50244140625,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.080078125,
0.76123046875,
0.468017578125,
0.76123046875,
0.468017578125,
0.7744140625,
0.080078125,
0.7744140625
],
[
0.50244140625,
0.76220703125,
0.89208984375,
0.76123046875,
0.89208984375,
0.77587890625,
0.50244140625,
0.7763671875
],
[
0.08087158203125,
0.77783203125,
0.468017578125,
0.77783203125,
0.468017578125,
0.78857421875,
0.08087158203125,
0.78857421875
],
[
0.50244140625,
0.77783203125,
0.89208984375,
0.77783203125,
0.89208984375,
0.79150390625,
0.50244140625,
0.79150390625
],
[
0.080078125,
0.79150390625,
0.4697265625,
0.79150390625,
0.4697265625,
0.8046875,
0.080078125,
0.8046875
],
[
0.50341796875,
0.79345703125,
0.890625,
0.79345703125,
0.890625,
0.806640625,
0.50341796875,
0.806640625
],
[
0.080078125,
0.806640625,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.080078125,
0.82080078125
],
[
0.5048828125,
0.8095703125,
0.890625,
0.8095703125,
0.890625,
0.8193359375,
0.5048828125,
0.8193359375
],
[
0.080078125,
0.8212890625,
0.468017578125,
0.8212890625,
0.468017578125,
0.8359375,
0.080078125,
0.8359375
],
[
0.5009765625,
0.82275390625,
0.890625,
0.82275390625,
0.890625,
0.8359375,
0.5009765625,
0.8359375
],
[
0.080078125,
0.8359375,
0.4697265625,
0.8369140625,
0.4697265625,
0.85107421875,
0.080078125,
0.849609375
],
[
0.50244140625,
0.83837890625,
0.89208984375,
0.83837890625,
0.89208984375,
0.8515625,
0.50244140625,
0.8515625
],
[
0.08087158203125,
0.8515625,
0.468017578125,
0.8515625,
0.468017578125,
0.86572265625,
0.08087158203125,
0.86572265625
],
[
0.50244140625,
0.85302734375,
0.70751953125,
0.85302734375,
0.70751953125,
0.8662109375,
0.50244140625,
0.8662109375
],
[
0.10296630859375,
0.875,
0.471435546875,
0.875,
0.471435546875,
0.8896484375,
0.10296630859375,
0.8896484375
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.521484375,
0.88623046875
],
[
0.080078125,
0.88525390625,
0.26953125,
0.88623046875,
0.26953125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.5009765625,
0.88525390625,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.5009765625,
0.8994140625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Multi-channel Weighted Nuclear Norm Minimization",
"for Real Color Image Denoising",
"Jun Xu', Lei Zhangl.*, David Zhang', and Xiangchu Feng?",
"'Dept. of Computing, The Hong Kong Polytechnic University, Hong Kong, China",
"8",
"2School of Mathematics and Statistics, Xidian University, Xi'an, China",
"30",
"[csjunxu,cslzhang,csdzhang)@comp.polyu.edu.hk,xcfeng@mail.xidian.edu.cn",
"Abstract",
"ods have also been developed by learning discriminative",
"8",
"priors from pairs of clean and noisy images [9-12].",
"Most of the existing denoising algorithms are developed",
"When the input is a noisy RGB color image, there are",
"for grayscale images. It is not trivial to extend them for",
"mainly three strategies for color image denoising. (1) The",
"LAD",
"color image denoising since the noise statistics in R, G, and",
"first strategy is to apply the grayscale image denoising algo-",
"B channels can be very different for real noisy images. In",
"rithm to each channel. However, such a straightforward so-",
"this paper, we propose a multi-channel (MC) optimization",
"lution will not exploit the spectral correlation among RGB",
"SO",
"model for real color image denoising under the weighted",
"channels, and the denoising performance may not be very",
"nuclear norm minimization (WNNM) framework. We con-",
"satisfying. (2) The second strategy is to transform the RGB",
"catenate the RGB patches to make use of the channel redun-",
"A",
"image into a lesscorrelated colorspacesuchas YCbCr",
"dancy, and introduce a weight matrix to balance the data",
"and perform denoising in each channel of the transformed",
"fidelity of the three channels in consideration of their differ-",
"space [13, 14]. One representative work along this line is",
"ent noise statistics. The proposed MC-WNNM model does",
"the CBM3D algorithm [14]. However, the color transform",
"1660",
"not have an analytical solution. We reformulate it into a",
"will complicate the noise distribution, and the correlation",
"linear equality-constrained problem and solve it via alter-",
"among color channels is not fully exploited. (3) The third",
"nating direction method of multipliers. Each alternative up-",
"strategy is to perform joint denoising on the RGB channels",
"'SOLT",
"",
"dating step has a closed-form solution and the convergence",
"simultaneously for better use of the spectral correlation. For",
"can be guaranteed. Experiments on both synthetic and real",
"example, the patches from RGB channels are concatenated",
"noisy image datasets demonstrate the superiority of the pro-",
"LAXIE",
"as a long vector for processing [15, 16].",
"posed MC-WNNM over state-of-the-art denoising methods.",
"Though joint denoising of RGB channels is a more",
"promising way for color image denoising, it is not a triv-",
"ial extension from single channel (grayscale image) to mul-",
"1.Introduction",
"tiple channels (color image). The noise in standard RGB",
"Image denoising is a classical yet fundamental problem",
"(sRGB) space can be approximately modeled as AWGN",
"for image quality enhancement in computer vision and pho-",
"butithasdifferentvariancesfordifferentchannelsJ17-19]",
"tography systems. Most of existing denoising algorithms",
"due to the sensor characteristics and on-board processing",
"are designed for grayscale images, aiming to recover the",
"steps in digital camera pipelines [19, 20]. This makes the",
"clean image x from its noisy observation y = x + n,",
"real color image denoising problem much more complex. If",
"where n isgenerally assumed to be additivewhiteGaussian",
"the three channels are treated equally in the joint denoising",
"noise (AWGN). State-of-the-art image denoising methods",
"process, false colors or artifacts can be generated [15]. How",
"include sparse representation [l], dictionary learning [2],",
"to account for the different noise characteristics incolor",
"low-rank approximation [3], non-local self-similarity (NSS)",
"channels, and how to effectively exploit the within and cross",
"[4] based methods, and the combination of those techniques",
"channel correlation are the key issues for designing a good",
"[1-3, 5-8]. Recently, some discriminative denoising meth-",
"color image denoising method.",
"This work is supported by HK RGC GRF grant (PolyU 5313/13E) and",
"This paper presents a new color image denoising algo-",
"China NSFC grant (no. 61672446).",
"rithm. Considering that the weighted nuclear norm mini-",
"-"
] |
{
"arxiv_id": "1705.09912",
"language": "en",
"page_idx": 0
}
|
||
1705.09912-en-1
|
1705.09912-en-1.jpg
|
mization (WNNM) method [3,21], which exploits the image NSS property via low rank regularization, has achieved excellent denoising performance on grayscale images, we propose to extend WNNM to real color image denoising. More specifically, we propose a multi-channel WNNM (MC-WNNM) model, which concatenates the patches from RGB channels for rank minimization but introduces a weight matrix to adjust the contributions of the three channels based on their noise levels. The proposed MC-WNNM model no longer has a closed-form solution as in the original WNNM model[21].We reformulate it into a linear equality-constrained problem with two variables, and solve the relaxed problem under the alternating direction method of multipliers [22] framework. Each variable can be updated with closed-form solutions, and the convergence analysis is given to guarantee a rational termination of the proposed algorithm.
## $2. \textbf{Related Work}$
$2. 1. \textbf{ Weighted Nuclear Norm Minimization}$
As a generalization to the nuclear norm minimization (NNM) model[23], the weighted nuclear norm minimization (WNNM) model [3,21] is described as
$$\min_{\mathbf{X}}\|\mathbf{Y}-\mathbf{X}\|_{F}^{2}+\|\mathbf{X}\|_{\boldsymbol{w},*},$$
(1)
where $\|\mathbf{X}\|_{\boldsymbol{w},*}=\sum_{i}w_{i}\sigma_{i}(\mathbf{X})$ is the weighted nuclear norm of matrix X, $\boldsymbol w=[w_1,...,w_n]^\top(w_i\geq0)$ is the weight vector, and $\sigma_i(\mathbf{X})$ is the $i$th singular value of X. According to the Corollary 1 of [21], if the weights are nondecreasing, the problem (1) has a closed-form solution:
$$\mathbf{\hat{X}}=\mathbf{U}\mathcal{S}_{\boldsymbol{w}/2}(\boldsymbol{\Sigma})\mathbf{V}^{\top},$$
where $\mathbf{Y}=\mathbf{U\Sigma V}^\top$ is the singular value decomposition
(2)
(SVD)[24] of Y and $S_{\boldsymbol{w}/2}(\bullet)$ is the generalized soft-
thresholding operator with weight vector $w:$
(3)
$$\mathcal{S}_{\boldsymbol{w}/2}(\boldsymbol{\Sigma}_{ii})=\max(\boldsymbol{\Sigma}_{ii}-w_{i}/2,0).$$
WNNM has demonstrated highly competitive denoising
performance on grayscale images. However, if we directly extend it to color image denoising by concatenating the patches from RGB channels, denoising artifacts may happen (please refer to Fig. 1 and the section of experimental results). In this paper, we propose a multi-channel WNNM (MC-WNNM) model for color image denoising, which preserves the power of WNNM and is able to address the noise differences among different channels.
# 2.2. Real Color Image Denoising
During the last decade, a few methods have been proposed for real color image denoising. Among them, the CBM3D method[14] is a representative one, which first transforms the RGB image into a luminance-chrominance space (e.g., YCbCr) and then applies the benchmark BM3D method[l] to each channel separately. The non-local similar patches are grouped by the luminance channel. In [17],
Liu et al. proposed the“Noise Level Function" to estimate and remove the noise for each channel in natural images. However, processing each channel separately would often achieve inferior performance to processing the color channels jointly [15]. Therefore, the methods [16,25,26] perform real color image denoising by concatenating the patches of RGB channels into a long vector. However, the concatenation treats each channel equally and ignores the different noise statistics among these channels. The method in [19] models the cross-channel noise in real noisy images as multivariate Gaussian and the noise is removed by the Bayesian non-local means filter [27]. The commercial software Neat Image [28] estimates the noise parameters from a flat region of the given noisy image and filters the noise accordingly. The methods in [19,28] ignore the non-local self-similarity of natural images [1,3].
In this paper, we present an effective multi-channel image denoising algorithm, which utilizes the strong low-rank prior of image non-local similar patches, and introduces a weight matrix to balance the multi-channels based on their different noise levels.
3. The Proposed Color Image Denoising Algo-
# rithm
## $3. 1. \textbf{The Multi- channel Weighted Nuclear Norm}$ Minimization Model
The color image denoising problem is to recover the clean image $\mathbf{x}_c$ from its noisy version $\mathbf{y}_c=\mathbf{x}_c+\mathbf{n}_c$,where $c=\{r,g,b\}$ is the index of R, G, B channels and n$_c$ is the noise in the $c$ channel. Patch based image denoising [1-3,5- 13] has achieved a great success in the last decade. Given a noisy color image y$_c$, each local patch of size $p\times p\times3$ is extracted and stretched to a patch vector, denoted by $\begin{array}{l}\mathbf{y}=[\mathbf{y}_{r}^{\top}\mathbf{y}_{g}^{\top}\mathbf{y}_{b}^{\top}]^{\top}\in\mathbb{R}^{3p^{2}},\:\mathbf{where}\mathbf{y}_{r},\mathbf{y}_{g},\mathbf{y}_{b}\in\mathbb{R}^{p^{2}}\text{are}\\\mathrm{the~corresponding~patches~in~R,~G,~B~channels.~For~each}\end{array}$ local patch y, we search the $M$ most similar patches to it (including y itself) by Euclidean distance in a relatively large local window around it. By stacking the $M$ similar patches column by column, we form a noisy patch matrix $\mathbf{Y}=\mathbf{X}+\mathbf{N}\in\mathbb{R}^{3p^{2}\times M}$,where $\mathbf{X}$ and N are the corresponding clean and noise patch matrices.
The noise in standard RGB (sRGB) space could be approximately modeled as additive white Gaussian (AWGN), but noise in different channels has different variances [17- 19]. Therefore, it is problematic to directly apply some grayscale denoising methods to the concatenated vectors y or matrices Y. To better illustrate this point, in Fig. 1, we show a clean image “kodim08”(only the R and G channels are shown due to limit of space), its noisy version generated by adding AWGN to each channel, and the denoised image by applying WNNM[21] to the concatenated patch matrix $\mathbf{Y}.$ The standard deviations of AWGN added to the R, G, B channels are $\sigma_r=40,\sigma_g=20,\sigma_b=30$, respectively. To
2
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.07843017578125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.07843017578125,
0.121826171875
],
[
0.50244140625,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50244140625,
0.11932373046875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.1514892578125,
0.46728515625,
0.1514892578125,
0.46728515625,
0.1654052734375,
0.080078125,
0.1654052734375
],
[
0.50244140625,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08087158203125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.08087158203125,
0.18115234375
],
[
0.50341796875,
0.169189453125,
0.8896484375,
0.169189453125,
0.8896484375,
0.179931640625,
0.50341796875,
0.179931640625
],
[
0.08087158203125,
0.184326171875,
0.4697265625,
0.184326171875,
0.4697265625,
0.1944580078125,
0.08087158203125,
0.1944580078125
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50244140625,
0.210205078125
],
[
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
],
[
0.07843017578125,
0.2266845703125,
0.4697265625,
0.2279052734375,
0.4697265625,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.50341796875,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.50341796875,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.080078125,
0.25830078125,
0.4697265625,
0.25830078125,
0.4697265625,
0.271484375,
0.080078125,
0.271484375
],
[
0.50244140625,
0.25830078125,
0.890625,
0.25830078125,
0.890625,
0.271484375,
0.50244140625,
0.271484375
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.28857421875,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.080078125,
0.3017578125
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.07843017578125,
0.302978515625,
0.4697265625,
0.3037109375,
0.4697265625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.5009765625,
0.302978515625,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.5009765625,
0.317626953125
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.5009765625,
0.317626953125,
0.75634765625,
0.318115234375,
0.75634765625,
0.333251953125,
0.5009765625,
0.33203125
],
[
0.080078125,
0.335205078125,
0.148681640625,
0.335205078125,
0.148681640625,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.521484375,
0.333984375,
0.890625,
0.333251953125,
0.890625,
0.347900390625,
0.521484375,
0.348388671875
],
[
0.50341796875,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.080078125,
0.35791015625,
0.2181396484375,
0.35791015625,
0.2181396484375,
0.37255859375,
0.080078125,
0.37255859375
],
[
0.50341796875,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50341796875,
0.37744140625
],
[
0.080078125,
0.38330078125,
0.406982421875,
0.38330078125,
0.406982421875,
0.396484375,
0.080078125,
0.396484375
],
[
0.5048828125,
0.380615234375,
0.8896484375,
0.380615234375,
0.8896484375,
0.391357421875,
0.5048828125,
0.391357421875
],
[
0.50244140625,
0.39453125,
0.6435546875,
0.39453125,
0.6435546875,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.098876953125,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.098876953125,
0.417236328125
],
[
0.080078125,
0.41845703125,
0.4697265625,
0.419189453125,
0.4697265625,
0.43310546875,
0.080078125,
0.431884765625
],
[
0.50244140625,
0.41845703125,
0.8837890625,
0.42041015625,
0.8837890625,
0.43505859375,
0.50244140625,
0.43310546875
],
[
0.080078125,
0.434326171875,
0.369384765625,
0.434326171875,
0.369384765625,
0.447509765625,
0.080078125,
0.447509765625
],
[
0.52294921875,
0.439453125,
0.57275390625,
0.439453125,
0.57275390625,
0.45068359375,
0.52294921875,
0.45068359375
],
[
0.181396484375,
0.451904296875,
0.36767578125,
0.451904296875,
0.36767578125,
0.466552734375,
0.181396484375,
0.466552734375
],
[
0.44677734375,
0.45263671875,
0.4697265625,
0.45263671875,
0.4697265625,
0.46533203125,
0.44677734375,
0.46533203125
],
[
0.50244140625,
0.455810546875,
0.869140625,
0.45703125,
0.869140625,
0.4716796875,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.473388671875,
0.4697265625,
0.473388671875,
0.4697265625,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.53173828125,
0.473388671875,
0.69287109375,
0.473388671875,
0.69287109375,
0.4873046875,
0.53173828125,
0.4873046875
],
[
0.07843017578125,
0.4892578125,
0.468017578125,
0.4892578125,
0.468017578125,
0.50244140625,
0.07843017578125,
0.50244140625
],
[
0.51953125,
0.4892578125,
0.89208984375,
0.490478515625,
0.89208984375,
0.5048828125,
0.51953125,
0.50390625
],
[
0.080078125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.517578125,
0.080078125,
0.517578125
],
[
0.5009765625,
0.5048828125,
0.89208984375,
0.505859375,
0.89208984375,
0.52099609375,
0.5009765625,
0.51953125
],
[
0.08087158203125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.5009765625,
0.521484375,
0.89208984375,
0.52099609375,
0.89208984375,
0.53515625,
0.5009765625,
0.5361328125
],
[
0.080078125,
0.5341796875,
0.44287109375,
0.5341796875,
0.44287109375,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50341796875,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.1993408203125,
0.54736328125,
0.34716796875,
0.548828125,
0.34716796875,
0.56640625,
0.1993408203125,
0.56494140625
],
[
0.44677734375,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.564453125,
0.44677734375,
0.564453125
],
[
0.50341796875,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.07843017578125,
0.5673828125,
0.468017578125,
0.5693359375,
0.468017578125,
0.583984375,
0.07843017578125,
0.58203125
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.583984375,
0.468017578125,
0.5986328125,
0.080078125,
0.59765625
],
[
0.50244140625,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.080078125,
0.599609375,
0.371826171875,
0.599609375,
0.371826171875,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.5966796875,
0.8935546875,
0.5966796875,
0.8935546875,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.156005859375,
0.6162109375,
0.390625,
0.61572265625,
0.390625,
0.6298828125,
0.156005859375,
0.630859375
],
[
0.44677734375,
0.6171875,
0.468017578125,
0.6171875,
0.468017578125,
0.630859375,
0.44677734375,
0.630859375
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.098876953125,
0.6318359375,
0.468017578125,
0.63330078125,
0.468017578125,
0.64794921875,
0.098876953125,
0.646484375
],
[
0.5009765625,
0.62744140625,
0.89208984375,
0.626953125,
0.89208984375,
0.6416015625,
0.5009765625,
0.642578125
],
[
0.50341796875,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.080078125,
0.6494140625,
0.46728515625,
0.6494140625,
0.46728515625,
0.66357421875,
0.080078125,
0.66357421875
],
[
0.080078125,
0.6640625,
0.4697265625,
0.6640625,
0.4697265625,
0.67822265625,
0.080078125,
0.67822265625
],
[
0.50244140625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.080078125,
0.6796875,
0.468017578125,
0.6796875,
0.468017578125,
0.693359375,
0.080078125,
0.693359375
],
[
0.50341796875,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.5009765625,
0.68603515625,
0.8935546875,
0.6875,
0.8935546875,
0.705078125,
0.5009765625,
0.7041015625
],
[
0.07843017578125,
0.6943359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70849609375,
0.07843017578125,
0.708984375
],
[
0.080078125,
0.71044921875,
0.468017578125,
0.71044921875,
0.468017578125,
0.7236328125,
0.080078125,
0.7236328125
],
[
0.50244140625,
0.70556640625,
0.775390625,
0.70556640625,
0.775390625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.724609375,
0.468017578125,
0.724609375,
0.468017578125,
0.73779296875,
0.080078125,
0.73779296875
],
[
0.521484375,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.521484375,
0.734375
],
[
0.080078125,
0.74072265625,
0.468017578125,
0.74072265625,
0.468017578125,
0.75390625,
0.080078125,
0.75390625
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.080078125,
0.7548828125,
0.330078125,
0.7548828125,
0.330078125,
0.76953125,
0.080078125,
0.76953125
],
[
0.50341796875,
0.75048828125,
0.8896484375,
0.75048828125,
0.8896484375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.77587890625,
0.330078125,
0.77587890625,
0.330078125,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.5009765625,
0.7822265625,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.5009765625,
0.79541015625
],
[
0.098876953125,
0.79443359375,
0.46728515625,
0.79443359375,
0.46728515625,
0.8095703125,
0.098876953125,
0.8095703125
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5048828125,
0.8125,
0.8896484375,
0.8125,
0.8896484375,
0.82275390625,
0.5048828125,
0.82275390625
],
[
0.07843017578125,
0.82470703125,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85595703125,
0.468017578125,
0.85498046875,
0.468017578125,
0.869140625,
0.07843017578125,
0.87060546875
],
[
0.5009765625,
0.85595703125,
0.89208984375,
0.85498046875,
0.89208984375,
0.869140625,
0.5009765625,
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,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.9013671875,
0.50244140625,
0.9013671875
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"mization (WNNM) method [3, 21], which exploits the im-",
"Liu et al. proposed the \"Noise Level Function\" to esti-",
"age NSS property via low rank regularization, has achieved",
"mate and removethe noise for each channel in natural im-",
"excellentdenoisingperformance ongrayscaleimages,we",
"ages. However, processing each channel separately would",
"propose to extend WNNM to real color image denois-",
"often achieve inferior performance to processing the color",
"ing. More specifically, we propose a multi-channel WNNM",
"channels jointly [15]. Therefore, the methods [16, 25, 26]",
"(MC-WNNM) model, which concatenates the patches from",
"perform real color image denoising by concatenating the",
"RGBchannels for rank minimization but introduces a",
"patches of RGB channels into a long vector. However, the",
"weight matrix to adjust the contributions of the three chan-",
"concatenation treats each channel equally and ignores the",
"nels based on their noise levels. The proposed MC-WNNM",
"different noise statistics among these channels. The method",
"model no longer has a closed-form solution as in the orig-",
"in[19lmodelsthecross-channel noiseinreal noisyimages",
"inal WNNM model [21]. We reformulate it into a linear",
"as multivariate Gaussian and the noise is removed by the",
"equality-constrained problem with twovariables,and solve",
"Bayesian non-local means filter [27]. The commercial soft-",
"the relaxed problem under the alternating direction method",
"ware Neat Image [28] estimates the noise parameters from",
"of multipliers[22] framework.Each variable can beupdated",
"a flat region of the given noisy image and filters the noise",
"with closed-form solutions, and the convergence analysis is",
"accordingly. The methods in [19, 28] ignore the non-local",
"given to guarantee a rational termination of the proposed",
"self-similarity of natural images [1, 3].",
"algorithm.",
"In this paper, we present an effective multi-channel im-",
"age denoising algorithm, which utilizes the strong low-rank",
"2. Related Work",
"prior of image non-local similar patches, and introduces a",
"2.1. Weighted Nuclear Norm Minimization",
"weightmatrixtobalance themulti-channels based on their",
"different noise levels.",
"As a generalization to the nuclear norm minimization",
"(NNM)model[23],theweightednuclearnormminimiza-",
"3.TheProposed Color Image Denoising Algo-",
"tion (WNNM) model [3, 21J is described as",
"rithm",
"min Y -X + XIlw,*",
"(1)",
"3.1. The Multi-channel Weighted Nuclear Norm",
"where Xlw,* = , wioi(X) is the weighted nuclear",
"Minimization Model",
"norm of matrix X, w = [wi,.., wn] (wi ≥ O) is the",
"The color image denoising problem is to recover the",
"weight vector, and ;(X) is the ith singular value of X.",
"clean imagexefrom its noisyversion y=x+ne,where",
"According to the Corollary 1 of [21], if the weights are non-",
"C = (r, g, b) is the index of R, G, B channels and n is the",
"decreasing, the problem (1) has a closed-form solution:",
"noise in the c channel. Patch based image denoising [1-3, 5-",
"X = USw/2(2)VT,",
"(2)",
"13J has achieved a great success in the last decade. Given a",
"where Y = UVT is the singular value decomposition",
"noisy color image yc, each local patch of size p × p × 3",
"(SVD) [24] of Y and Sw/2() is the generalized soft-",
"is extracted and stretched to a patch vector, denoted by",
"thresholding operator with weight vector w:",
"y = [y, y, yTIT e R3p\", where yr, yg, yh e Rp\"are",
"Sw/2(2i) = max(Z- wi/2,0)",
"(3)",
"the corresponding patches in R, G, B channels. For each",
"WNNM has demonstrated highly competitive denoising",
"local patch y, we search the M most similar patches to it",
"(including y itself) by Euclidean distance in a relatively",
"performance on grayscale images. However, if we directly",
"extend it to color image denoising by concatenating the",
"large local window around it. By stacking the M similar",
"patches from RGB channels, denoising artifacts may hap-",
"patches column by column,we form a noisy patchmatrix",
"Y = X + N e IR3p? ×M, where X and N are the corre-",
"pen (please refer to Fig. I and the section of experimental",
"results). In this paper, we propose a multi-channel WNNM",
"sponding clean and noise patch matrices.",
"(MC-WNNM) model for color image denoising, which pre-",
"The noise in standard RGB (sRGB) space could be ap-",
"serves the power of WNNM and is able to address the noise",
"proximately modeled as additive white Gaussian (AWGN),",
"differences among different channels.",
"but noise in different channels has different variances [17-",
"19]. Therefore, it is problematic to directly apply some",
"2.2. Real Color Image Denoising",
"grayscale denoising methods to the concatenated vectors y",
"During the last decade, a few methods have been pro-",
"or matrices Y. To better illustrate this point, in Fig. I, we",
"posed for real color image denoising. Among them, the",
"showacleanimage\"kodimO8\"(onlytheRandGchannels",
"CBM3D method [14] is a representative one, which first",
"are shown due to limit of space), its noisy version generated",
"transforms the RGB image into a luminance-chrominance",
"by adding AWGN to each channel, and the denoised image",
"space (e.g., YCbCr) and then applies the benchmark BM3D",
"by applying WNNM [21] to the concatenated patch matrix",
"method [I] to each channel separately. The non-local simi-",
"Y. The standard deviations of AWGN added to the R, G, B",
"lar patches are grouped by the luminance channel. In [17],",
"channels are or = 40, og = 20, ob = 30, respectively. To",
"2"
] |
{
"arxiv_id": "1705.09912",
"language": "en",
"page_idx": 1
}
|
||
1705.09912-en-2
|
1705.09912-en-2.jpg
|
<FigureHere>
(f) Denoised Red by MC-WNNM
(g) Denoised Green by WNNM
Figure 1. The Red and Green channels of the image “kodim08” from the Kodak PhotoCD Dataset, its synthetic noisy version, and the
$\bar{\text{images recovered by the concatenated WNNM and the proposed MC-WNNM methods.}}$
similar patches in R, G, B channels, respectively. The corresponding clean matrix is $\mathbf{X}=[\mathbf{X}_{r}^{\top}\mathbf{X}_{g}^{\top}\mathbf{X}_{b}^{\top}]^{\top}$,where $\mathbf{X}_r,\mathbf{X}_g,\mathbf{X}_b$ are similarly defıned. The weight matrix W can be determined under the maximum a-posterior (MAP) estimation framework:
$$\begin{aligned}&\text{on nannewonk.}\\&\hat{\mathbf{X}}=\arg\operatorname*{max}_{\mathbf{X}}\ln P(\mathbf{X}|\mathbf{Y},\boldsymbol{w})\\&=\arg\operatorname*{max}_{\mathbf{X}}\{\ln P(\mathbf{Y}|\mathbf{X})+\ln P(\mathbf{X}|\boldsymbol{w})\}.\\&\text{likalihood tarm ln }P(\mathbf{Y}|\mathbf{X})\:\mathrm{ic}\:\mathrm{choroctari}\end{aligned}$$
(5)
make WNNM applicable to color image denoising, we set the noise standard deviation as the average deviation of the whole noisy image, i.e., $\sigma=\sqrt{(\sigma_{r}^{2}+\sigma_{g}^{2}+\sigma_{b}^{2})/3}\approx31.1.$ From Fig. 1, one can see that the concatenated WNNM remains some noise in the R channel while over-smoothing the G channel. This is because it processes R and G channels equally without considering their differences in noise corruption.
Clearly, a more effective color image denoising algorithm should consider the different noise strength in color channels. To this end, we introduce a weight matrix W to balance the noise in the RGB channels, and present the following multi-channel WNNM (MC-WNNM) model:
$$\min_{\mathbf{X}}\|\mathbf{W}(\mathbf{Y}-\mathbf{X})\|_{F}^{2}+\|\mathbf{X}\|_{\boldsymbol{w},*}.$$
The log-likelihood term $\ln P(\mathbf{Y}|\mathbf{X})$ is characterized by the statistics of noise. According to[18], we assume that the noise is independent among RGB channels and independently and identically distributed (i.i.d.) in each channel with Gaussian distribution and standard deviations $\{\sigma_r,\sigma_g,\sigma_b\}.$There is:
(4)
$$\begin{aligned}&P(\mathbf{Y}|\mathbf{X})=\prod_{c\in\{r,g,b\}}(2\pi\sigma_{c}^{2})^{-\frac{3p^{2}}{2}}\exp(-\frac{1}{2\sigma_{c}^{2}}\|\mathbf{Y}_{c}-\mathbf{X}_{c}\|_{F}^{2}).\end{aligned}$$
For the latent data $\mathbf{X}$, the small weighted nuclear norm prior is imposed on it, i.e., $\|\mathbf{X}\|_{\boldsymbol{w},*}=\sum_iw_i\sigma_i(\mathbf{X})$ should be sparsely distributed. We let it be:
$$\begin{array}{l}{{\mathrm{sunoutca.~wc~icu~u~u.}}}\\{P(\mathbf{X}|\boldsymbol{w})\propto\exp(-\frac{1}{2}\|\mathbf{X}\|_{\boldsymbol{w},*}).}\end{array}$$
(7)
We follow the method in [21] to set the weight vector $w$ on nuclear norm as $w_i^{k+1}=C/(|\sigma_i(\mathbf{X}_k)|+\epsilon)$, where $\epsilon>0$ is a small number to avoid zero numerator and $\sigma_i(\mathbf{X}_k)$ is the $i$th singular value of the estimated data matrix X at the $k$th iteration. Note that if $\sigma_r=\sigma_g=\sigma_b$,the proposed MC-WNNM model will be reduced to the concatenated WNNM model. With an appropriate setting of the weight matrix W and a good optimization algorithm, the proposed MC-WNNM model will lead to much better color image denoising results. As shown in Figs. l(f) and l(h), MC-WNNM removes clearly the noise in the R channel while preserving textures effectively in the G channel.
$$\begin{aligned}&\mathbf{\hat{X}}=\arg\min_{\mathbf{X}}\sum_{c\in\{r,g,b\}}\frac{1}{\sigma_{c}^{2}}\|(\mathbf{Y}_{c}-\mathbf{X}_{c})\|_{F}^{2}+\|\mathbf{X}\|_{\boldsymbol{w},*}\\&=\arg\min_{\mathbf{X}}\|\mathbf{W}(\mathbf{Y}-\mathbf{X})\|_{F}^{2}+\|\mathbf{X}\|_{\boldsymbol{w},*},\end{aligned}$$
(8)
## 3.2. The Setting of Weight Matrix W
with
$$\mathbf{W}=\left(\begin{array}{ccc}\sigma_r^{-1}\mathbf{I}&\mathbf{0}&\mathbf{0}\\\mathbf{0}&\sigma_g^{-1}\mathbf{I}&\mathbf{0}\\\mathbf{0}&\mathbf{0}&\sigma_b^{-1}\mathbf{I}\end{array}\right),$$
(9)
Let's denote the noisy patch matrix by $\mathbf{Y}=$ $[\mathbf{Y}_{r}^{\top}\mathbf{Y}_{g}^{\top}\mathbf{Y}_{b}^{\top}]^{\top}$, where $\mathbf{Y}_{r},\hat{\mathbf{Y}}_{g},\mathbf{Y}_{b}$ are sub-matrices of
3
|
[
[
0.125,
0.222900390625,
0.2467041015625,
0.222900390625,
0.2467041015625,
0.236083984375,
0.125,
0.236083984375
],
[
0.330810546875,
0.222900390625,
0.453369140625,
0.222900390625,
0.453369140625,
0.236083984375,
0.330810546875,
0.236083984375
],
[
0.53173828125,
0.222900390625,
0.6650390625,
0.222900390625,
0.6650390625,
0.236083984375,
0.53173828125,
0.236083984375
],
[
0.73779296875,
0.222900390625,
0.87158203125,
0.222900390625,
0.87158203125,
0.236083984375,
0.73779296875,
0.236083984375
],
[
0.301513671875,
0.386962890625,
0.48193359375,
0.386962890625,
0.48193359375,
0.40087890625,
0.301513671875,
0.40087890625
],
[
0.5107421875,
0.386962890625,
0.68115234375,
0.386474609375,
0.68115234375,
0.399658203125,
0.5107421875,
0.40087890625
],
[
0.7060546875,
0.386962890625,
0.900390625,
0.386962890625,
0.900390625,
0.40087890625,
0.7060546875,
0.40087890625
],
[
0.080078125,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.080078125,
0.4140625
],
[
0.1070556640625,
0.38818359375,
0.26318359375,
0.386962890625,
0.26318359375,
0.3984375,
0.1070556640625,
0.399658203125
],
[
0.07843017578125,
0.4140625,
0.59619140625,
0.4140625,
0.59619140625,
0.427978515625,
0.07843017578125,
0.427978515625
],
[
0.07843017578125,
0.442626953125,
0.4697265625,
0.443115234375,
0.4697265625,
0.458251953125,
0.07843017578125,
0.45703125
],
[
0.50341796875,
0.443115234375,
0.890625,
0.443115234375,
0.890625,
0.45703125,
0.50341796875,
0.45703125
],
[
0.07843017578125,
0.45703125,
0.468017578125,
0.458251953125,
0.468017578125,
0.472900390625,
0.07843017578125,
0.4716796875
],
[
0.50341796875,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.472900390625,
0.50341796875,
0.472900390625
],
[
0.07843017578125,
0.477783203125,
0.265625,
0.4765625,
0.265625,
0.4912109375,
0.07843017578125,
0.492431640625
],
[
0.281982421875,
0.477783203125,
0.468017578125,
0.477783203125,
0.468017578125,
0.492431640625,
0.281982421875,
0.492431640625
],
[
0.50341796875,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.50244140625,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.080078125,
0.49560546875,
0.468017578125,
0.49560546875,
0.468017578125,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.07843017578125,
0.5087890625,
0.468017578125,
0.51123046875,
0.468017578125,
0.52587890625,
0.07843017578125,
0.5234375
],
[
0.50244140625,
0.50390625,
0.65283203125,
0.50390625,
0.65283203125,
0.517578125,
0.50244140625,
0.517578125
],
[
0.5537109375,
0.5166015625,
0.74658203125,
0.517578125,
0.74658203125,
0.53515625,
0.5537109375,
0.5341796875
],
[
0.080078125,
0.52587890625,
0.46728515625,
0.52587890625,
0.46728515625,
0.5390625,
0.080078125,
0.5390625
],
[
0.87109375,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.87109375,
0.54736328125
],
[
0.080078125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.080078125,
0.5546875
],
[
0.57861328125,
0.54443359375,
0.83642578125,
0.54443359375,
0.83642578125,
0.5595703125,
0.57861328125,
0.5595703125
],
[
0.080078125,
0.556640625,
0.1544189453125,
0.556640625,
0.1544189453125,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.09722900390625,
0.57080078125,
0.468017578125,
0.5712890625,
0.468017578125,
0.5859375,
0.09722900390625,
0.58544921875
],
[
0.50341796875,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.580078125,
0.50341796875,
0.580078125
],
[
0.080078125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.599609375,
0.080078125,
0.599609375
],
[
0.50244140625,
0.58203125,
0.89208984375,
0.58203125,
0.89208984375,
0.59521484375,
0.50244140625,
0.59521484375
],
[
0.07843017578125,
0.60107421875,
0.4697265625,
0.599609375,
0.4697265625,
0.6142578125,
0.07843017578125,
0.61572265625
],
[
0.5009765625,
0.5966796875,
0.890625,
0.59521484375,
0.890625,
0.60986328125,
0.5009765625,
0.611328125
],
[
0.080078125,
0.6162109375,
0.468017578125,
0.6162109375,
0.468017578125,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.60986328125,
0.89208984375,
0.625,
0.50244140625,
0.626953125
],
[
0.08087158203125,
0.6318359375,
0.423095703125,
0.6318359375,
0.423095703125,
0.642578125,
0.08087158203125,
0.642578125
],
[
0.50341796875,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.1650390625,
0.64453125,
0.3857421875,
0.64599609375,
0.3857421875,
0.66357421875,
0.1650390625,
0.662109375
],
[
0.44677734375,
0.64794921875,
0.468017578125,
0.64794921875,
0.468017578125,
0.66015625,
0.44677734375,
0.66015625
],
[
0.50341796875,
0.642578125,
0.65185546875,
0.6416015625,
0.65185546875,
0.65576171875,
0.50341796875,
0.6572265625
],
[
0.50244140625,
0.6611328125,
0.72802734375,
0.6611328125,
0.72802734375,
0.681640625,
0.50244140625,
0.681640625
],
[
0.71337890625,
0.66357421875,
0.89453125,
0.65087890625,
0.89892578125,
0.68310546875,
0.71728515625,
0.69580078125
],
[
0.740234375,
0.662109375,
0.74658203125,
0.662109375,
0.74658203125,
0.66650390625,
0.740234375,
0.66650390625
],
[
0.07843017578125,
0.6796875,
0.2279052734375,
0.6767578125,
0.2279052734375,
0.6943359375,
0.07843017578125,
0.69775390625
],
[
0.08087158203125,
0.66552734375,
0.468017578125,
0.66552734375,
0.468017578125,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.2181396484375,
0.68115234375,
0.4697265625,
0.68115234375,
0.4697265625,
0.6943359375,
0.2181396484375,
0.6943359375
],
[
0.564453125,
0.6845703125,
0.62109375,
0.6845703125,
0.62109375,
0.69580078125,
0.564453125,
0.69580078125
],
[
0.080078125,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.080078125,
0.708984375
],
[
0.521484375,
0.69580078125,
0.89208984375,
0.6962890625,
0.89208984375,
0.7109375,
0.521484375,
0.71044921875
],
[
0.08087158203125,
0.7109375,
0.468017578125,
0.7109375,
0.468017578125,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.50244140625,
0.7119140625,
0.89208984375,
0.7109375,
0.89208984375,
0.72607421875,
0.50244140625,
0.7265625
],
[
0.080078125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.50244140625,
0.72802734375,
0.73779296875,
0.72802734375,
0.73779296875,
0.7412109375,
0.50244140625,
0.7412109375
],
[
0.080078125,
0.74072265625,
0.465576171875,
0.74072265625,
0.465576171875,
0.75390625,
0.080078125,
0.75390625
],
[
0.5947265625,
0.7421875,
0.72802734375,
0.74560546875,
0.72802734375,
0.76318359375,
0.5947265625,
0.76025390625
],
[
0.7294921875,
0.7470703125,
0.7939453125,
0.7470703125,
0.7939453125,
0.76123046875,
0.7294921875,
0.76123046875
],
[
0.869140625,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76123046875,
0.869140625,
0.76123046875
],
[
0.08087158203125,
0.755859375,
0.4697265625,
0.755859375,
0.4697265625,
0.77099609375,
0.08087158203125,
0.77099609375
],
[
0.0767822265625,
0.77099609375,
0.4697265625,
0.76953125,
0.4697265625,
0.7841796875,
0.0767822265625,
0.78515625
],
[
0.50341796875,
0.7666015625,
0.740234375,
0.7666015625,
0.740234375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.07843017578125,
0.78515625,
0.468017578125,
0.7861328125,
0.468017578125,
0.80029296875,
0.07843017578125,
0.7998046875
],
[
0.5146484375,
0.787109375,
0.60888671875,
0.79052734375,
0.60888671875,
0.8046875,
0.5146484375,
0.8017578125
],
[
0.69384765625,
0.79150390625,
0.71826171875,
0.79150390625,
0.71826171875,
0.7998046875,
0.69384765625,
0.7998046875
],
[
0.7236328125,
0.787109375,
0.85888671875,
0.7890625,
0.85888671875,
0.8037109375,
0.7236328125,
0.8017578125
],
[
0.07843017578125,
0.80029296875,
0.4697265625,
0.8017578125,
0.4697265625,
0.81494140625,
0.07843017578125,
0.814453125
],
[
0.6650390625,
0.80029296875,
0.6806640625,
0.80029296875,
0.6806640625,
0.8095703125,
0.6650390625,
0.8095703125
],
[
0.68115234375,
0.798828125,
0.6953125,
0.798828125,
0.6953125,
0.8017578125,
0.68115234375,
0.8017578125
],
[
0.609375,
0.80810546875,
0.66015625,
0.80810546875,
0.66015625,
0.818359375,
0.609375,
0.818359375
],
[
0.87109375,
0.806640625,
0.89208984375,
0.806640625,
0.89208984375,
0.8193359375,
0.87109375,
0.8193359375
],
[
0.080078125,
0.81640625,
0.4697265625,
0.81640625,
0.4697265625,
0.830078125,
0.080078125,
0.830078125
],
[
0.53369140625,
0.82568359375,
0.65869140625,
0.82373046875,
0.65869140625,
0.8369140625,
0.53369140625,
0.8388671875
],
[
0.64892578125,
0.8212890625,
0.7939453125,
0.82373046875,
0.7939453125,
0.84130859375,
0.64892578125,
0.8388671875
],
[
0.08087158203125,
0.83349609375,
0.2646484375,
0.83349609375,
0.2646484375,
0.84326171875,
0.08087158203125,
0.84326171875
],
[
0.58642578125,
0.8388671875,
0.59716796875,
0.8388671875,
0.59716796875,
0.84326171875,
0.58642578125,
0.84326171875
],
[
0.080078125,
0.8515625,
0.359375,
0.8515625,
0.359375,
0.86572265625,
0.080078125,
0.86572265625
],
[
0.50244140625,
0.85693359375,
0.53515625,
0.85693359375,
0.53515625,
0.869140625,
0.50244140625,
0.869140625
],
[
0.6435546875,
0.85986328125,
0.68310546875,
0.85693359375,
0.685546875,
0.875,
0.646484375,
0.87744140625
],
[
0.705078125,
0.8623046875,
0.7197265625,
0.861328125,
0.72216796875,
0.875,
0.7060546875,
0.87548828125
],
[
0.75634765625,
0.861328125,
0.77392578125,
0.861328125,
0.77392578125,
0.87548828125,
0.75634765625,
0.87548828125
],
[
0.09722900390625,
0.8720703125,
0.43701171875,
0.8720703125,
0.43701171875,
0.88525390625,
0.09722900390625,
0.88525390625
],
[
0.65869140625,
0.87353515625,
0.67626953125,
0.8798828125,
0.66650390625,
0.89306640625,
0.650390625,
0.88623046875
],
[
0.68701171875,
0.876953125,
0.73388671875,
0.869140625,
0.73876953125,
0.8896484375,
0.69287109375,
0.896484375
],
[
0.75634765625,
0.87548828125,
0.77294921875,
0.87548828125,
0.77294921875,
0.89013671875,
0.75634765625,
0.89013671875
],
[
0.86767578125,
0.876953125,
0.89453125,
0.876953125,
0.89453125,
0.89306640625,
0.86767578125,
0.89306640625
],
[
0.08087158203125,
0.88623046875,
0.471435546875,
0.88623046875,
0.471435546875,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.58251953125,
0.87744140625,
0.60546875,
0.87744140625,
0.60546875,
0.89013671875,
0.58251953125,
0.89013671875
],
[
0.705078125,
0.892578125,
0.72216796875,
0.892578125,
0.72216796875,
0.90576171875,
0.705078125,
0.90576171875
],
[
0.74267578125,
0.892578125,
0.7861328125,
0.89013671875,
0.78759765625,
0.90576171875,
0.744140625,
0.90771484375
],
[
0.48046875,
0.923828125,
0.491943359375,
0.923828125,
0.491943359375,
0.9375,
0.48046875,
0.9375
],
[
0.65673828125,
0.89306640625,
0.6708984375,
0.89306640625,
0.6708984375,
0.90478515625,
0.65673828125,
0.90478515625
]
] |
[
"(a) Clean Red Channel",
"(b) Noisy Red Channel",
"(c) Clean Green Channel",
"(d) Noisy Green Channel",
"(f) Denoised Red by MC-WNNM",
"(g) Denoised Green by WNNM",
"(h) Denoised Green by MC-WNNM",
"Figure 1. The Red and Green channels of the image kodimo8\" from the Kodak PhotoCD Dataset, its synthetic noisy version, and the",
"(e)DenoisedRedbyWNNM",
"images recovered by the concatenated WNNM and the proposed MC-WNNM methods.",
"make WNNM applicable to color image denoising,we set",
"similar patches in R, G, B channels, respectively. The cor-",
"the noise standard deviation as the average deviation of the",
"responding clean matrix is X = [X, X, XTjT, where",
"whole noisy image, i.e., :",
"/(α2 +α? +0)/3 ~ 31.1.",
"Xr, Xg, X, are similarly defined. The weight matrix W",
"can be determined under the maximum a-posterior (MAP)",
"From Fig. 1, one can see that the concatenated WNNM re-",
"mains some noise in the R channel while over-smoothing",
"estimation framework:",
"X = arg max ln P(X/Y, w)",
"the G channel. This is because it processes R and G chan-",
"(5)",
"nels equally without considering their differences in noise",
"= arg max[ln P(YX) + ln P(X|w))",
"corruption.",
"Clearly, a more effective color image denoising algo-",
"The log-likelihood term ln P(YX) is characterized by",
"rithm should consider the different noise strength in color",
"the statistics of noise.According to [18], we assume that",
"channels. To this end, we introduce a weight matrix W to",
"the noise is independent among RGB channels and inde-",
"balance the noise in the RGB channels, and present the fol-",
"pendently and identically distributed (i.i.d.) in each chan-",
"lowingmulti-channelWNNM(MC-WNNM)model",
"nelwith Gaussiandistribution andstandarddeviations",
"min /W(Y - X)/ + IXlw,*.",
"(4)",
"or,Og, Ob].There is:",
"P(Y|X) =II(2元2)- exp(-",
"-2 / Y - X/1) (6)",
"",
"nuclear norm as wk+1",
"We follow the method in [21] to set the weight vector w on",
" = C/(loi(X)/+e), wheree > 0is a",
"ce(r,g,b)",
"small number to avoid zero numerator and ; (X) is the ith",
"For the latent data X, the small weighted nuclear norm",
"singular value of the estimated data matrix X at the kth iter-",
"priorisimposed on it,i.e.,Xw*=,wio;(X)should",
"ation. Note that if o, = g = Ob, the proposed MC-WNNM",
"be sparsely distributed. We let it be:",
"model will be reduced to the concatenated WNNM model",
"P(X|w) α exp(-",
"Xw,*)",
"(7)",
"With an appropriate setting of the weight matrix W and",
" a good optimization algorithm, the proposed MC-WNNM",
"Putting (7) and (6) into (5), we have",
"model will lead to much better color image denoising re-",
"X= argmin",
"Y",
"-x+xw*",
"sults. As shown in Figs. 1(f) and 1(h), MC-WNNM removes",
"g",
"",
"ce(r.g.b)",
"(8)",
"clearly the noise in the R channel while preserving textures",
"= argmin W(Y",
" -+*",
"effectivelyintheGchannel.",
"X",
"3.2. The Setting of Weight Matrix W",
"with",
"a-11",
"0",
"0",
"Let's denote the noisy patch matrix by Y",
"0",
"ogl1",
"0",
"(9)",
"[Y Y, Y,]T, where Yr, Yg, Y, are sub-matrices of",
"W",
"0",
"a,\"1",
"3",
"0"
] |
{
"arxiv_id": "1705.09912",
"language": "en",
"page_idx": 2
}
|
||
1705.09912-en-3
|
1705.09912-en-3.jpg
|
where $\mathbf{I}\in\mathbb{R}^{p^2\times p^2}$ is the identity matrix.
Clearly, the weight matrix W is diagonal and determined by the noise standard deviation in each channel. The stronger the noise in a channel, the less the contribution that channel should make to the estimation of X. Our experimental results (refer to Section 4 please) on synthetic and real noisy images clearly demonstrate the advantages of MC-WNM over WNNM and other methods in color image denoising.
## 3.3. Model Optimization
The proposed MC-WNNM model does not have an analytical solution. In the WNNM model[21], when the weights assigned on singular values are in a non-descending order, the weighted nuclear norm proximal operator can have a global optimum with closed-form solution. Unfortunately, such a property is not valid for the MC-WNNM model because a weight matrix W is assigned to the rows of data matrix X. This makes the proposed model more difficult to solve than the original WNNM model.
We employ the variable splitting method[29,30] to solve the MC-WNNM model. By introducing an augmented variable $\mathbf{Z}$, the MC-WNNM model can be reformulated as a linear equality-constrained problem with two variables X and $\mathbf{Z}$:
(10)
$$\operatorname*{min}_{\mathbf{X},\mathbf{Z}}\|\mathbf{W}(\mathbf{Y}-\mathbf{X})\|_{F}^{2}+\|\mathbf{Z}\|_{\boldsymbol{w},*}\quad\mathrm{s.t.}\quad\mathbf{X}=\mathbf{Z}.$$
Since'the objective function is separable w.r.t. the two variables, the problem (10) can be solved under the alternating direction method of multipliers (ADMM) [22] framework. The augmented Lagrangian function is:
$$\begin{aligned}&\mathcal{L}(\mathbf{X},\mathbf{Z},\mathbf{A},\rho)=&&\|\mathbf{W}(\mathbf{Y}-\mathbf{X})\|_{F}^{2}+\|\mathbf{Z}\|_{\boldsymbol{w},*}\\&&&+\langle\mathbf{A},\mathbf{X}-\mathbf{Z}\rangle+\frac{\rho}{2}\|\mathbf{X}-\mathbf{Z}\|_{F}^{2},\\&\text{here A is the augmented I agrangian multinlier and}\end{aligned}$$
(11)
$\rho>0$
is the penalty parameter. We initialize the matrix variables $\mathbf{X}_{0},\mathbf{Z}_{0}$, and A$_0$ to be zero matrix and $\rho_0>0$ to be a suitable value. Denote by $(\mathbf{X}_k,\mathbf{Z}_k)$ and $\mathbf{A}_k$ the optimization variables and Lagrange multiplier at iteration $k$ $(k=0,1,2,...)$,respectively. By taking derivatives of the Lagrangian function $\mathcal{L}$ w.r.t. X and Z and setting the derivative function to be zero, we can alternatively update the variables as follows:
$$\begin{aligned}&\text{(1) Upuate X whinc nxing Z anu A.}\\&\mathbf{X}_{k+1}=\arg\min_{\mathbf{X}}\|\mathbf{W}(\mathbf{Y}-\mathbf{X})\|_{F}^{2}+\frac{\rho_{k}}{2}\|\mathbf{X}-\mathbf{Z}_{k}+\rho_{k}^{-1}\mathbf{A}_{k}\|_{F}^{2}.\end{aligned}$$
(12)
This is a standard least squares regression problem with
closed-form solution:
$$\mathbf{X}_{k+1}=(\mathbf{W}^{\top}\mathbf{W}+\frac{\rho_{k}}{2}\mathbf{I})^{-1}(\mathbf{W}^{\top}\mathbf{W}\mathbf{Y}+\frac{\rho_{k}}{2}\mathbf{Z}_{k}-\frac{1}{2}\mathbf{A}_{k}).$$
(13)
$( 2) \textbf{Update Z while fixing X and A: }$
$$\mathbf{Z}_{k+1}=\arg\min_{\mathbf{Z}}\frac{\rho_{k}}{2}\|\mathbf{Z}-(\mathbf{X}_{k+1}+\rho_{k}^{-1}\mathbf{A}_{k})\|_{F}^{2}+\|\mathbf{Z}\|_{\boldsymbol{w},*}.$$
(14)
According to the Theorem 1 in [21], given the SVD of $\mathbf{X}_{k+1}+\rho_{k}^{-1}\mathbf{A}_{k}$,i.e$,\mathbf{X}_k+1+\rho_{k}^{-1}\mathbf{A}_{k}=\mathbf{U}_{k}\Sigma_{k}\mathbf{V}_{k}^{\top}$,where $\Sigma _{k}$ = $\begin{pmatrix} {\mathrm{diag}( \sigma _{1}, \sigma _{2}, \ldots , \sigma _{M}) }\\ {\mathbf{0} }\end{pmatrix}$ $\in$ $\mathbb{R} ^{3p^{2}\times M}$ (without loss of generality, we assume that $3p_{\underline{\Delta}}^{2}\geq M)$, the global
$\begin{array}{cc}{\mathrm{optimum~of~the~above~problem~is~}}&{\hat{\mathbf{Z}}}_{k+1}&{=}&{\mathbf{U}}_{k}\hat{\mathbf{\Sigma}}_{k}\mathbf{V}_{k}^{\top},\\{\mathrm{where~}\hat{\mathbf{\Sigma}}_{k}}&{=\left(\begin{array}{cc}{\mathrm{diag}(\hat{\sigma}_{1},\hat{\sigma}_{2},\ldots,\hat{\sigma}_{M})}\\{\mathbf{0}}\end{array}\right)}&{\in\:\mathbb{R}^{3p^{2}\times M}}&{\mathrm{and}}\end{array}$ $(\hat{\sigma}_{1},\hat{\sigma}_{2},...,\hat{\sigma}_{M})$ is the solution to the following convex optimization problem:
$$\min_{\hat{\sigma}_{1},\hat{\sigma}_{2},\ldots,\hat{\sigma}_{M}}\sum_{i=1}^{M}(\sigma_{i}-\hat{\sigma}_{i})^{2}+\frac{2w_{i}}{\rho_{k}}\hat{\sigma}_{i}$$
(15)
s.t. $\hat{\sigma } _1\geq \hat{\sigma } _2\geq . . . \geq \hat{\sigma } _M\geq 0.$
According to the Remark 1 in [21], the problem above has
closed-form solution $(i=1,2,...,M)\colon$
$$\hat{\sigma}_i=\left\{\begin{array}{ll}0&\mathrm{if}\:c_2<0\\\frac{c_1+\sqrt{c_2}}{2}&\mathrm{if}\:c_2\ge0\end{array}\right.,$$
(16)
where $c_1=\sigma_i-\epsilon,c_2=(\sigma_i-\epsilon)^2-\frac{8C}{\rho_k},\epsilon>0$ is a small
number, and $C$ is set as $\sqrt2M$ by experience in [21].
$( 3) \textbf{Update A while fixing X and Z: }$
$$\mathbf{A}_{k+1}=\mathbf{A}_{k}+\rho_{k}(\mathbf{X}_{k+1}-\mathbf{Z}_{k+1}).$$
(17)
$( 4) \textbf{Update }\rho _{k}\nobreak {: } \rho _{k+ 1}= \mu * \rho _{k}$,where $\mu>1.$
The above alternative updating steps are repeated until the convergence condition is satisfied or the number of iterations exceeds a preset threshold. The convergence condi- $\begin{array}{l}{\text{tion of the ADMM algorithm is:}}&\|\mathbf{X}_{k+1}-\mathbf{Z}_{k+1}\|_{F}\leq\mathrm{Tol,}\\{\|\mathbf{X}_{k+1}-\mathbf{X}_{k}\|_{F}\leq\mathrm{Tol,~and}}&\|\mathbf{Z}_{k+1}-\mathbf{Z}_{k}\|_{F}\leq\mathrm{Tol~are~si-}\end{array}$ multaneously satisfied, where Tol > 0 is a small tolerance number. We summarize the updating procedures in Algorithm 1. The convergence analysis of the proposed Algorithm 1 is given in Theorem l. Note that since the weighted nuclear norm is non-convex in general, we employ an unbounded sequence of $\{\rho_k\}$ here to make sure that Algorithm 1 converges.
Theorem 1. Assume that the weights in w are in a non-
descending order, the sequences $\{\mathbf{X}_k\},\{\mathbf{Z}_k\},and~\{\mathbf{A}_k\}$
generated in Algorithm I satisfy:
(18)
$$(a)\operatorname*{lim}_{k\to\infty}\|\mathbf{X}_{k+1}-\mathbf{Z}_{k+1}\|_{F}=0;\\(b)\operatorname*{lim}_{k\to\infty}\|\mathbf{X}_{k+1}-\mathbf{X}_{k}\|_{F}=0;\\(c)\operatorname*{lim}_{k\to\infty}\|\mathbf{Z}_{k+1}-\mathbf{Z}_{k}\|_{F}=0.$$
(19) (20)
$Proof.$ We give a sketch proof here and detailed proof of
this theorem can be found in the supplementary file.
We first prove that the sequence $\{\mathbf{A}_k\}$ generated by Algorithm 1 is upper bounded. Since $\{\rho_k\}$ is unbounded, i.e., $\operatorname* { lim} _{k\to \infty }\rho _{k}$ = $+ \infty $, we can prove that the sequence of Lagrangian function $\{\mathcal{L}(\mathbf{X}_{k+1},\mathbf{Z}_{k+1},\mathbf{A}_{k},\rho_{k})\}$ is also upper bounded. Hence, both $\{\mathbf{W}(\mathbf{Y}-\mathbf{X}_{k})\}$ and $\{\mathbf{Z}_{k}\}$ are upper bounded. Then $\{\mathbf{X}_k\}$ is also upper bounded. According to Eq.(17), we can prove that $\lim_k\to\infty\|\mathbf{X}_{k+1}-$
4
|
[
[
0.0767822265625,
0.08709716796875,
0.34716796875,
0.08709716796875,
0.34716796875,
0.10797119140625,
0.0767822265625,
0.10797119140625
],
[
0.50341796875,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.098876953125,
0.106689453125,
0.468017578125,
0.106689453125,
0.468017578125,
0.121826171875,
0.098876953125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.12249755859375,
0.50341796875,
0.12249755859375
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.580078125,
0.12249755859375,
0.72119140625,
0.12371826171875,
0.72119140625,
0.1383056640625,
0.580078125,
0.136962890625
],
[
0.50341796875,
0.1307373046875,
0.53125,
0.1307373046875,
0.53125,
0.143310546875,
0.50341796875,
0.143310546875
],
[
0.74267578125,
0.125,
0.8935546875,
0.126953125,
0.8935546875,
0.1458740234375,
0.74267578125,
0.143310546875
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.6435546875,
0.1395263671875,
0.65673828125,
0.1395263671875,
0.65673828125,
0.149658203125,
0.6435546875,
0.149658203125
],
[
0.08087158203125,
0.1546630859375,
0.468017578125,
0.1546630859375,
0.468017578125,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.5,
0.1514892578125,
0.8935546875,
0.15087890625,
0.8935546875,
0.1685791015625,
0.5,
0.169189453125
],
[
0.07843017578125,
0.1685791015625,
0.4697265625,
0.1673583984375,
0.4697265625,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.50244140625,
0.169189453125,
0.81884765625,
0.169189453125,
0.81884765625,
0.184326171875,
0.50244140625,
0.184326171875
],
[
0.810546875,
0.169189453125,
0.89208984375,
0.169189453125,
0.89208984375,
0.184326171875,
0.810546875,
0.184326171875
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.61279296875,
0.184326171875,
0.75634765625,
0.1849365234375,
0.75634765625,
0.199462890625,
0.61279296875,
0.1988525390625
],
[
0.78125,
0.1875,
0.89453125,
0.1875,
0.89453125,
0.205810546875,
0.78125,
0.205810546875
],
[
0.080078125,
0.1976318359375,
0.468017578125,
0.1976318359375,
0.468017578125,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.50244140625,
0.1900634765625,
0.576171875,
0.1900634765625,
0.576171875,
0.2052001953125,
0.50244140625,
0.2052001953125
],
[
0.68310546875,
0.2020263671875,
0.6953125,
0.2020263671875,
0.6953125,
0.2108154296875,
0.68310546875,
0.2108154296875
],
[
0.0767822265625,
0.2108154296875,
0.176513671875,
0.21337890625,
0.176513671875,
0.2279052734375,
0.0767822265625,
0.2259521484375
],
[
0.50341796875,
0.21533203125,
0.890625,
0.21533203125,
0.890625,
0.2291259765625,
0.50341796875,
0.2291259765625
],
[
0.080078125,
0.236083984375,
0.268798828125,
0.236083984375,
0.268798828125,
0.25,
0.080078125,
0.25
],
[
0.5009765625,
0.2310791015625,
0.6337890625,
0.2310791015625,
0.6337890625,
0.2442626953125,
0.5009765625,
0.2442626953125
],
[
0.65625,
0.248779296875,
0.66845703125,
0.248779296875,
0.66845703125,
0.255126953125,
0.65625,
0.255126953125
],
[
0.76708984375,
0.2442626953125,
0.79736328125,
0.2442626953125,
0.79736328125,
0.2568359375,
0.76708984375,
0.2568359375
],
[
0.5810546875,
0.255615234375,
0.60888671875,
0.255615234375,
0.60888671875,
0.264404296875,
0.5810546875,
0.264404296875
],
[
0.67236328125,
0.250732421875,
0.765625,
0.248779296875,
0.765625,
0.266357421875,
0.67236328125,
0.268310546875
],
[
0.098876953125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.2734375,
0.098876953125,
0.2734375
],
[
0.79150390625,
0.255615234375,
0.8095703125,
0.255615234375,
0.8095703125,
0.26513671875,
0.79150390625,
0.26513671875
],
[
0.5595703125,
0.261962890625,
0.62158203125,
0.26513671875,
0.62109375,
0.276611328125,
0.55810546875,
0.2734375
],
[
0.771484375,
0.26318359375,
0.79150390625,
0.26318359375,
0.79150390625,
0.27587890625,
0.771484375,
0.27587890625
],
[
0.859375,
0.261474609375,
0.8935546875,
0.261474609375,
0.8935546875,
0.276611328125,
0.859375,
0.276611328125
],
[
0.07843017578125,
0.274658203125,
0.4697265625,
0.274658203125,
0.4697265625,
0.28857421875,
0.07843017578125,
0.28857421875
],
[
0.6259765625,
0.279052734375,
0.8349609375,
0.279052734375,
0.8349609375,
0.29345703125,
0.6259765625,
0.29345703125
],
[
0.080078125,
0.290283203125,
0.4697265625,
0.290283203125,
0.4697265625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.5048828125,
0.296630859375,
0.89208984375,
0.296630859375,
0.89208984375,
0.306884765625,
0.5048828125,
0.306884765625
],
[
0.080078125,
0.30615234375,
0.4697265625,
0.30615234375,
0.4697265625,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.50244140625,
0.31005859375,
0.75732421875,
0.31005859375,
0.75732421875,
0.324462890625,
0.50244140625,
0.324462890625
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.640625,
0.3212890625,
0.66845703125,
0.32763671875,
0.6640625,
0.34033203125,
0.63623046875,
0.333984375
],
[
0.71728515625,
0.32763671875,
0.78125,
0.32763671875,
0.78125,
0.339111328125,
0.71728515625,
0.339111328125
],
[
0.080078125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.650390625,
0.335205078125,
0.7861328125,
0.341552734375,
0.78515625,
0.3623046875,
0.64892578125,
0.35595703125
],
[
0.859375,
0.333984375,
0.8935546875,
0.333984375,
0.8935546875,
0.349853515625,
0.859375,
0.349853515625
],
[
0.78515625,
0.3466796875,
0.79345703125,
0.3466796875,
0.79345703125,
0.349853515625,
0.78515625,
0.349853515625
],
[
0.59716796875,
0.3359375,
0.62841796875,
0.3359375,
0.62841796875,
0.348388671875,
0.59716796875,
0.348388671875
],
[
0.080078125,
0.349853515625,
0.4697265625,
0.349853515625,
0.4697265625,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.5,
0.363037109375,
0.8935546875,
0.361083984375,
0.8935546875,
0.380615234375,
0.5,
0.38330078125
],
[
0.07843017578125,
0.380126953125,
0.38818359375,
0.380126953125,
0.38818359375,
0.393310546875,
0.07843017578125,
0.393310546875
],
[
0.5009765625,
0.3837890625,
0.8447265625,
0.3837890625,
0.8447265625,
0.397705078125,
0.5009765625,
0.397705078125
],
[
0.1004638671875,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.1004638671875,
0.4091796875
],
[
0.50244140625,
0.40283203125,
0.7490234375,
0.40283203125,
0.7490234375,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.080078125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5810546875,
0.42041015625,
0.810546875,
0.42041015625,
0.810546875,
0.43505859375,
0.5810546875,
0.43505859375
],
[
0.859375,
0.419189453125,
0.8935546875,
0.419189453125,
0.8935546875,
0.43505859375,
0.859375,
0.43505859375
],
[
0.07843017578125,
0.425537109375,
0.4697265625,
0.425537109375,
0.4697265625,
0.439453125,
0.07843017578125,
0.439453125
],
[
0.07843017578125,
0.43994140625,
0.468017578125,
0.439453125,
0.468017578125,
0.45263671875,
0.07843017578125,
0.453857421875
],
[
0.5009765625,
0.436279296875,
0.8037109375,
0.4375,
0.8037109375,
0.455078125,
0.5009765625,
0.453857421875
],
[
0.080078125,
0.45703125,
0.1258544921875,
0.45703125,
0.1258544921875,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.521484375,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.521484375,
0.47021484375
],
[
0.10455322265625,
0.46728515625,
0.42236328125,
0.46728515625,
0.42236328125,
0.48486328125,
0.10455322265625,
0.48486328125
],
[
0.431396484375,
0.4697265625,
0.468017578125,
0.4697265625,
0.468017578125,
0.48095703125,
0.431396484375,
0.48095703125
],
[
0.50244140625,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.111083984375,
0.48095703125,
0.131591796875,
0.48095703125,
0.131591796875,
0.488037109375,
0.111083984375,
0.488037109375
],
[
0.42724609375,
0.474853515625,
0.4345703125,
0.474853515625,
0.4345703125,
0.477783203125,
0.42724609375,
0.477783203125
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.50244140625,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.07843017578125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.51806640625,
0.07843017578125,
0.51806640625
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.080078125,
0.51953125,
0.46728515625,
0.51953125,
0.46728515625,
0.53271484375,
0.080078125,
0.53271484375
],
[
0.5048828125,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.08087158203125,
0.5361328125,
0.338134765625,
0.5361328125,
0.338134765625,
0.546875,
0.08087158203125,
0.546875
],
[
0.5048828125,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54443359375,
0.5048828125,
0.54443359375
],
[
0.098876953125,
0.54736328125,
0.390625,
0.5498046875,
0.390625,
0.5673828125,
0.098876953125,
0.56494140625
],
[
0.5009765625,
0.546875,
0.890625,
0.54736328125,
0.890625,
0.5625,
0.5009765625,
0.56103515625
],
[
0.438720703125,
0.56494140625,
0.46728515625,
0.56494140625,
0.46728515625,
0.5771484375,
0.438720703125,
0.5771484375
],
[
0.50244140625,
0.56298828125,
0.890625,
0.56298828125,
0.890625,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.50341796875,
0.57763671875,
0.89208984375,
0.57763671875,
0.89208984375,
0.59228515625,
0.50341796875,
0.59228515625
],
[
0.2115478515625,
0.56640625,
0.42236328125,
0.568359375,
0.42236328125,
0.59033203125,
0.2115478515625,
0.58837890625
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50341796875,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50341796875,
0.60693359375
],
[
0.080078125,
0.60693359375,
0.4697265625,
0.60693359375,
0.4697265625,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.080078125,
0.62255859375,
0.4697265625,
0.62255859375,
0.4697265625,
0.63623046875,
0.080078125,
0.63623046875
],
[
0.50341796875,
0.625,
0.58642578125,
0.625,
0.58642578125,
0.63623046875,
0.50341796875,
0.63623046875
],
[
0.07843017578125,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.65087890625,
0.07843017578125,
0.65087890625
],
[
0.50341796875,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66748046875,
0.4697265625,
0.66748046875,
0.4697265625,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.50244140625,
0.67529296875,
0.72119140625,
0.67529296875,
0.72119140625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.58837890625,
0.68994140625,
0.8095703125,
0.68994140625,
0.8095703125,
0.705078125,
0.58837890625,
0.705078125
],
[
0.07843017578125,
0.69775390625,
0.4697265625,
0.69775390625,
0.4697265625,
0.7109375,
0.07843017578125,
0.7109375
],
[
0.861328125,
0.6923828125,
0.89208984375,
0.6923828125,
0.89208984375,
0.705078125,
0.861328125,
0.705078125
],
[
0.080078125,
0.7138671875,
0.191162109375,
0.7138671875,
0.191162109375,
0.724609375,
0.080078125,
0.724609375
],
[
0.5888671875,
0.71728515625,
0.7900390625,
0.71728515625,
0.7900390625,
0.73095703125,
0.5888671875,
0.73095703125
],
[
0.859375,
0.71337890625,
0.896484375,
0.71630859375,
0.8935546875,
0.73193359375,
0.85888671875,
0.72998046875
],
[
0.08087158203125,
0.734375,
0.326904296875,
0.734375,
0.326904296875,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.58837890625,
0.74072265625,
0.78515625,
0.74072265625,
0.78515625,
0.75830078125,
0.58837890625,
0.75830078125
],
[
0.859375,
0.74072265625,
0.8935546875,
0.74072265625,
0.8935546875,
0.755859375,
0.859375,
0.755859375
],
[
0.318603515625,
0.7646484375,
0.325927734375,
0.7646484375,
0.325927734375,
0.76953125,
0.318603515625,
0.76953125
],
[
0.0743408203125,
0.74365234375,
0.473876953125,
0.74365234375,
0.473876953125,
0.7744140625,
0.0743408203125,
0.7744140625
],
[
0.50244140625,
0.7646484375,
0.89208984375,
0.76513671875,
0.89208984375,
0.77978515625,
0.50244140625,
0.77880859375
],
[
0.438720703125,
0.7734375,
0.468017578125,
0.7734375,
0.468017578125,
0.7861328125,
0.438720703125,
0.7861328125
],
[
0.50244140625,
0.78076171875,
0.84228515625,
0.78076171875,
0.84228515625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.07843017578125,
0.787109375,
0.4697265625,
0.78857421875,
0.4697265625,
0.80322265625,
0.07843017578125,
0.8017578125
],
[
0.521484375,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.521484375,
0.8095703125
],
[
0.080078125,
0.8037109375,
0.22216796875,
0.8037109375,
0.22216796875,
0.8173828125,
0.080078125,
0.8173828125
],
[
0.2132568359375,
0.81005859375,
0.420654296875,
0.8125,
0.420654296875,
0.8369140625,
0.2132568359375,
0.83544921875
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82568359375,
0.50341796875,
0.82568359375
],
[
0.07843017578125,
0.8193359375,
0.2279052734375,
0.81494140625,
0.2279052734375,
0.83349609375,
0.07843017578125,
0.8369140625
],
[
0.428955078125,
0.82080078125,
0.46728515625,
0.82080078125,
0.46728515625,
0.833984375,
0.428955078125,
0.833984375
],
[
0.366943359375,
0.83056640625,
0.378173828125,
0.83056640625,
0.378173828125,
0.84130859375,
0.366943359375,
0.84130859375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84130859375,
0.50244140625,
0.84130859375
],
[
0.423095703125,
0.83349609375,
0.428955078125,
0.83349609375,
0.428955078125,
0.83837890625,
0.423095703125,
0.83837890625
],
[
0.439453125,
0.8388671875,
0.4697265625,
0.8388671875,
0.4697265625,
0.8515625,
0.439453125,
0.8515625
],
[
0.226318359375,
0.83203125,
0.233642578125,
0.83203125,
0.233642578125,
0.83837890625,
0.226318359375,
0.83837890625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85595703125,
0.50244140625,
0.85595703125
],
[
0.08087158203125,
0.85107421875,
0.325927734375,
0.85107421875,
0.325927734375,
0.8642578125,
0.08087158203125,
0.8642578125
],
[
0.1895751953125,
0.861328125,
0.4697265625,
0.86572265625,
0.4697265625,
0.88720703125,
0.1895751953125,
0.88330078125
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.080078125,
0.8662109375,
0.1993408203125,
0.8671875,
0.1993408203125,
0.8818359375,
0.080078125,
0.880859375
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5009765625,
0.88623046875
],
[
0.1953125,
0.87890625,
0.2099609375,
0.87890625,
0.2099609375,
0.88720703125,
0.1953125,
0.88720703125
],
[
0.43701171875,
0.88525390625,
0.4697265625,
0.88525390625,
0.4697265625,
0.90087890625,
0.43701171875,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.88623046875,
0.88720703125,
0.88623046875,
0.9013671875,
0.5009765625,
0.9013671875
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"where I e Rp' xp is the identity matrix.",
"According to the Theorem 1 in [21], given the SVD of",
"Clearly, the weight matrix W is diagonal and deter-",
"Xk+i + PA, i.e., Xk+i + P-'Ah =U,E,VT, where",
"mined by the noise standard deviation in each channel. The",
"diag(01, 02,...,0M)",
"Ek",
"E IR3p*xM (without",
"stronger the noise in a channel, the less the contribution",
"0",
"that channel should make to the estimation of X. Our ex-",
"loss of generality, we assume that 3p? ≥ M), the global",
"perimentalresults(referto Section4please)onsynthetic",
"optimum of the above problem is Zk+1 =",
"U.,Vt,",
"and real noisy images clearly demonstrate the advantages",
"diag(01,02,..., om)",
"E R3p*xM and",
"of MC-WNM over WNNM and other methods in color im-",
"where 2k",
"0",
"age denoising.",
"(o1, 02, .., M) is the solution to the following convex op-",
"3.3. Model Optimization",
"timization problem:",
"M",
"2w,",
"min",
"(αi-0.)2 +",
"The proposed MC-WNNM model does not have an an-",
"0",
"01.02....0",
"Pk",
"(15)",
"alytical solution. In the WNNM model [2]], when the",
"s.t. 0i≥02 ≥...≥0m≥0.",
"weights assigned on singular values are in a non-descending",
"AccordingtotheRemark1in[2ll.theproblemabovehas",
"order, the weighted nuclear norm proximal operator can",
"closed-form solution (i = 1, 2, ., M):",
"have a global optimum with closed-form solution. Unfor-",
"C0",
"if c2<0",
"tunately, such a property is not valid for the MC-WNNM",
"ci+y if c2 ≥0",
"(16)",
"",
"oi",
"model because a weight matrix W is assigned to the rows",
"of data matrix X. This makes the proposed model more dif-",
"where ci = 0i - , c2 = (oi - e)2_ 8C,e> 0 is a small",
"ficult to solve than the original WNNM model.",
"number, and C is set as 2M by experience in [21]",
"We employ the variable splitting method [29, 30] to solve",
"(3) Update A while fixing X and Z:",
"the MC-WNNM model. By introducing an augmented vari-",
"Ah+1 =Ak+ Pk(Xk+1-Zk+1)",
"(17)",
"able Z, the MC-WNNM model can be reformulated as a",
"linearequality-constrained problem with two variables X",
"(4) Update pk: Pk+1 = μ * Pk, where μ > 1.",
"and Z:",
"The above alternative updating steps are repeated until",
"minl/W(Y-X) +Zw* s.t. X=Z.",
"(10)",
"the convergence condition is satisfied or the number of iter-",
"X.7",
"",
"Since the objective function is separable w.r.t. the two vari-",
"ations exceeds a preset threshold. The convergence condi-",
"ables, the problem (10) can be solved under the alternating",
"tion of the ADMM algorithm is: IXk+1 - Zk+ilF ≤ Tol,",
"direction method of multipliers (ADMM) [22] framework.",
"IXk+1 -XlF ≤ Tol, and Zk+i - ZkllF ≤ Tol are si-",
"The augmented Lagrangian function is:",
"multaneously satisfied,where Tol>0is a small tolerance",
"C(X,Z,A,p) =W(Y - X) + IIZIlw,*",
"number. We summarize the updating procedures in Algo-",
"(11)",
"rithm 1. The convergence analysis of the proposed Algo-",
"rithm 1 is given in Theorem 1. Note that since the weighted",
"+(A,X -Z)+号X -Z%,",
"where A is the augmented Lagrangian multiplier and p > O",
"nuclear norm is non-convex in general, we employ an un-",
"is the penalty parameter. We initialize the matrix variables",
"bounded sequence of pk} here to make sure that Algorithm",
"Xo, Zo, and Ao to be zero matrix and po > O to be",
"I converges.",
"a suitable value. Denote by (Xk,Zk) and Ak the opti-",
"Theorem 1. Assume that the weights in w are in a non-",
"mization variables and Lagrange multiplier at iteration k",
"descending order, the sequences [X,], {Z], and [A,]",
"(k = 0, 1, 2, ...), respectively. By taking derivatives of the",
"generated in Algorithm I satisfy:",
"Lagrangian function C w.r.t. X and Z and setting the deriva-",
"(a) lim JXk+1- Zk+1llF =0;",
"tive function to be zero, we can alternatively update the vari-",
"(18)",
"ables as follows:",
"(6)lim Xk+1-XF=0;",
"(19)",
"(I) Update X while fixing Z and A:",
"(c) lim Zk+1-ZklF = 0.",
"(20)",
"2",
"Xx+1=argmn/W(Y-x)+x-Z+PAl/.",
"Proof. We give a sketch proof here and detailed proof of",
"(12)",
"this theorem can be found in the supplementary file.",
"This is a standard least squares regression problem with",
"We first prove that the sequence (A, generated by Al-",
"closed-form solution:",
"(",
"gorithm I is upper bounded. Since (pk) is unbounded, i.e.,",
"Xk+1 =(WTW + §",
"Ak).",
"2",
"limk→oo Pk =+oo, we can prove that the sequence of",
"",
"(13)",
"2",
"Lagrangian function [C(Xk+1, Zk+1, Ak, Ph)) is also up-",
"(2) Update Z while fixing X and A:",
"Z-(X+1+A+",
"per bounded. Hence, both {W(Y - X,)) and {Z} are",
"Zk+1 = arg min ",
"upper bounded. Then [X] is also upper bounded. Ac-",
"2",
"(14)",
"cording to Eq. (17), we can prove that limk-→ I[Xk+1",
"4"
] |
{
"arxiv_id": "1705.09912",
"language": "en",
"page_idx": 3
}
|
||
1705.09912-en-4
|
1705.09912-en-4.jpg
|
<table>
<tbody>
<tr>
<td>11</td>
<td> </td>
</tr>
<tr>
<td>5</td>
<td>1</td>
</tr>
<tr>
<td> </td>
<td>T$\leftarrow$True ${\mathrm{md~if}}$</td>
</tr>
<tr>
<td>$e$</td>
<td>$\mid\mathbf{while}$</td>
</tr>
<tr>
<td> </td>
<td>and Z $X$</td>
</tr>
</tbody>
</table>
$\mathbf{Z} _{k+ 1}\| _{F}$ = $\lim _{k\to \infty }\rho _{k}^{- 1}\| \mathbf{A} _{k+ 1}$ - $\mathbf{A} _{k}\| _{F}$ = 0, and ( a) $\mathrm{is~proved. ~Then~we~can~prove~that~lim}_{k\to \infty }\| \mathbf{X} _{k+ 1}$ - $\hat{\mathbf{X} } _{k}\| _{F}$ $\leq$ $\lim _{k\to \infty }( \| ( \hat{\mathbf{W} } ^{\top }\mathbf{W}$ + $\frac {\rho _{k}}2\mathbf{I} ) ^{- 1}( \mathbf{W} ^{\top }\mathbf{W} \mathbf{Y}$ - $\mathbf{W}^{\top}\mathbf{W}\mathbf{Z}_{k}-\frac{1}{2}\mathbf{A}_{k})\|_{F}+\rho_{k}^{-1}\|\mathbf{A}_{k}-\mathbf{A}_{k-1}\|_{F})=0$and$\mathbf{hence~(b)~is~proved.~Finally,~(c)~can~be~proved~by~check-}$ ing that $\lim_{k\to\infty}\left\|\mathbf{Z}_{k+1}-\mathbf{Z}_{k}\right\|\leq\lim_{k\to\infty}(\left\|\boldsymbol{\Sigma}_{k-1}\right.-$ $\mathcal{S} _{\boldsymbol{w}/ \rho _{k- 1}}( \Sigma _{k- 1}) \| _{F}$ + $\| \mathbf{X} _{k+ 1}$ - $\mathbf{X} _{k}\| _{F}$ + $\| \rho _{k- 1}^{- 1}\mathbf{A} _{k- 1}$ + $\rho _{k}^{- 1}\mathbf{A} _{k+ 1}- \rho _{k}^{- 1}\mathbf{A} _{k}\| _{F})$ = 0,where $\mathbf{U}_{k-1}\mathbf{\Sigma}_{k-1}\mathbf{V}_{k-1}^{\top}$is
$\square$
the SVD of $\mathbf{X} _k+ \rho _{k- 1}^{- 1}\mathbf{A} _{k- 1}$ .
## 3.4. The Denoising Algorithm
Given a noisy color image y$_c$, suppose that we have extracted $N$ local patches $\{\mathbf{y}_j\}_{j=1}^{N}$ and their similar patches. $\{\mathbf{X}_j\}_{j=1}^N$ are aggregated to form the denoised image $\hat{\mathbf{x}}_c.$ To obtain better denoising results, we perform the above denoising procedures for several rounds. The proposed MCWNNM based color image denoising algorithm is summarized in Algorithm 2.
## 3.5. Complexity Analysis
In Algorithm 1 for solving the MC-WNNM model via ADMM, the cost for updating $X$ is $\mathcal{O}(\max(p^4M,M^3))$, while the cost for updating $Z$ is $\mathcal{O}(p^4M+M^3).$ The costs for updating $A$ and $\rho$ can be ignored. So the overall complexity is $\mathcal{O}((p^{4}M+M^{3})K_{1})$,where $K_1$ is the number of iterations. In Algorithm 2 for image denoising, we consider the number of patches $N$ extracted from the input noisy image and the number of itertations $K_2$ and ignore the cost for searching similar patches. The overall cost is $\mathcal{O}((p^{4}M+M^{3})K_{1}\tilde{K}_{2}N).$
# 4. Experiments
We evaluate the proposed MC-WNNM method on synthetic and real noisy color images. We compare the pro-
<table>
<tbody>
<tr>
<td>$\overline{\mathbf{AI}}$ 1 1</td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{I}_{\mathbf{l}}$ $\mathbf{lm}$</td>
<td>VOIS</td>
</tr>
<tr>
<td>Ini for</td>
<td>${\textbf{itialization: }\hat{\mathbf{x}}}_{c}^{(0)}$ (0) $=\mathbf{y}_{c},\mathbf{y}_{d}^{0}$ $=\mathbf{y}_{c};$ $\mathbf{r}$ $k= 1: K_{2}$ $\mathbf{do}$</td>
</tr>
<tr>
<td>3 $a$</td>
<td>Searc {:}hes</td>
</tr>
<tr>
<td>4.</td>
<td>Apply the MC-WNNM model (10) to $\mathbf{Y}_{j}$ and obtain the estimated $\mathbf{X}_j;$ $\mathbf{endfor}$</td>
</tr>
<tr>
<td>5. $en$</td>
<td>Aggregate $\{\mathbf{X}_j\}_{j=1}^N$ to form the image $\hat{\mathbf{x}}_c^{(k)};$ ıdfor</td>
</tr>
<tr>
<td>$\mathbf{O}_{1}$ 1</td>
<td>$(K_{2})$ utput: Denoised image $\hat{\mathbf{x}}_c^{(\mathrm{c})}$</td>
</tr>
</tbody>
</table>
posed method with state-of-the-art denoising methods, including CBM3D [14], MLP [9], WNNM [3], TNRD [11], DnCNN[12]“Noise Clinic”(NC)[25,26],CC[19], and the commercial software Neat Image (NI)[28]. The Matlab source code of our MC-WNNM algorithm can be downloaded at http://www4.comp.polyu.edu. hk/~cslzhang/code/MCWNNM.zip.
# 4.1. Experimental Settings
Noise level estimation. For most of the competing denoising algorithms, the standard deviation of noise should be given as a parameter. In synthetic experiments, the noise levels $(\sigma_{r},\sigma_{g},\sigma_{b})$ in R, G ,B channels are assumed to be known. In the case of real noisy images, the noise levels can be estimated via some noise estimation methods [31,32]. In this paper, we employ the method [32] to estimate the noise
$N$ noisy patch matrices $\{ \mathbf{Y} _{j}\} _{j= 1}^{N}$ can be formed to esti- this paper, we employ the mate the the the the the che the the the the the the pathe the the pathe the the he the the the the the the patches in matrices level for each color channel.
Noise level of comparison methods. For the CBM3D method[14], a single parameter of noise level should be input. We set the noise level as
$$\sigma=\sqrt{(\sigma_r^2+\sigma_g^2+\sigma_b^2)/3}.$$
(21)
The methods of MLP[9] and TNRD[11] are originally designed for grayscale images. We retrain their models (using the released codes by the authors) at different noise levels from $\sigma=5$ to $\sigma=75$ with a gap of 5. The denoising on color images is performed by processing each channel with the model trained at the same (or nearest) noise level.
Comparison with WNNM. In order to make a full and fair comparison with the original WNNM method[21], we implement WNNM for color image denoising in three ways. $\hat{1})$ We apply WNNM to each color channel separately with the corresponding noise levels $\sigma_r,\sigma_g,\sigma_b.$ We call this method“WNNM-1”.2) We perform WNNM on the concatenated matrix Y formed by the patches in RGB channels, while the input noise level $\sigma$ is computed by Eq. (21). We call this method “WNNM-2”.3) We set the weight matrix W as $\mathbf{W}=\sigma^-1\mathbf{I}$ in the proposed MC-WNNM model,
5
|
[
[
0.08905029296875,
0.09027099609375,
0.39208984375,
0.09027099609375,
0.39208984375,
0.10418701171875,
0.08905029296875,
0.10418701171875
],
[
0.5107421875,
0.09027099609375,
0.875,
0.0889892578125,
0.875,
0.10418701171875,
0.5107421875,
0.10479736328125
],
[
0.08905029296875,
0.1060791015625,
0.410888671875,
0.1060791015625,
0.410888671875,
0.12060546875,
0.08905029296875,
0.12060546875
],
[
0.50732421875,
0.10479736328125,
0.87158203125,
0.10479736328125,
0.87158203125,
0.12249755859375,
0.50732421875,
0.12249755859375
],
[
0.0882568359375,
0.11932373046875,
0.386474609375,
0.121826171875,
0.386474609375,
0.1363525390625,
0.0882568359375,
0.1337890625
],
[
0.50732421875,
0.11614990234375,
0.70751953125,
0.11871337890625,
0.70751953125,
0.14013671875,
0.50732421875,
0.1383056640625
],
[
0.69287109375,
0.12371826171875,
0.71337890625,
0.121826171875,
0.7138671875,
0.1287841796875,
0.69384765625,
0.1307373046875
],
[
0.71337890625,
0.12249755859375,
0.7607421875,
0.1256103515625,
0.7587890625,
0.1395263671875,
0.71142578125,
0.1363525390625
],
[
0.191162109375,
0.1363525390625,
0.3095703125,
0.1363525390625,
0.3095703125,
0.149658203125,
0.191162109375,
0.149658203125
],
[
0.51123046875,
0.1383056640625,
0.6357421875,
0.1383056640625,
0.6357421875,
0.15283203125,
0.51123046875,
0.15283203125
],
[
0.0906982421875,
0.1514892578125,
0.238525390625,
0.1514892578125,
0.238525390625,
0.1654052734375,
0.0906982421875,
0.1654052734375
],
[
0.5087890625,
0.15283203125,
0.65673828125,
0.1483154296875,
0.65869140625,
0.169189453125,
0.5107421875,
0.173583984375
],
[
0.0906982421875,
0.1673583984375,
0.2181396484375,
0.1673583984375,
0.2181396484375,
0.18115234375,
0.0906982421875,
0.18115234375
],
[
0.50732421875,
0.17041015625,
0.80126953125,
0.1685791015625,
0.80126953125,
0.1900634765625,
0.50732421875,
0.1925048828125
],
[
0.0882568359375,
0.1787109375,
0.453369140625,
0.1787109375,
0.453369140625,
0.1962890625,
0.0882568359375,
0.1962890625
],
[
0.52783203125,
0.1900634765625,
0.66845703125,
0.1900634765625,
0.66845703125,
0.203857421875,
0.52783203125,
0.203857421875
],
[
0.0882568359375,
0.1962890625,
0.35302734375,
0.1956787109375,
0.35302734375,
0.210205078125,
0.0882568359375,
0.2108154296875
],
[
0.51123046875,
0.205810546875,
0.5400390625,
0.205810546875,
0.5400390625,
0.2171630859375,
0.51123046875,
0.2171630859375
],
[
0.5361328125,
0.2052001953125,
0.7861328125,
0.2052001953125,
0.7861328125,
0.2197265625,
0.5361328125,
0.2197265625
],
[
0.51123046875,
0.2203369140625,
0.84326171875,
0.2203369140625,
0.84326171875,
0.23486328125,
0.51123046875,
0.23486328125
],
[
0.1192626953125,
0.210205078125,
0.428955078125,
0.212158203125,
0.428955078125,
0.2298583984375,
0.1192626953125,
0.2279052734375
],
[
0.08660888671875,
0.2247314453125,
0.44677734375,
0.2266845703125,
0.44677734375,
0.2442626953125,
0.08660888671875,
0.242431640625
],
[
0.54345703125,
0.2342529296875,
0.7099609375,
0.23486328125,
0.7099609375,
0.25,
0.54345703125,
0.248779296875
],
[
0.0882568359375,
0.242431640625,
0.26318359375,
0.2442626953125,
0.26318359375,
0.25830078125,
0.0882568359375,
0.255615234375
],
[
0.52783203125,
0.2474365234375,
0.58251953125,
0.25,
0.5810546875,
0.264404296875,
0.52685546875,
0.261962890625
],
[
0.0906982421875,
0.2587890625,
0.1854248046875,
0.2587890625,
0.1854248046875,
0.26953125,
0.0906982421875,
0.26953125
],
[
0.50634765625,
0.264404296875,
0.82275390625,
0.260009765625,
0.82275390625,
0.285400390625,
0.50634765625,
0.2890625
],
[
0.10455322265625,
0.272705078125,
0.439453125,
0.272705078125,
0.439453125,
0.287353515625,
0.10455322265625,
0.287353515625
],
[
0.0882568359375,
0.285888671875,
0.1871337890625,
0.285888671875,
0.1871337890625,
0.300537109375,
0.0882568359375,
0.300537109375
],
[
0.51318359375,
0.2841796875,
0.564453125,
0.2841796875,
0.564453125,
0.29541015625,
0.51318359375,
0.29541015625
],
[
0.101318359375,
0.302978515625,
0.1502685546875,
0.300537109375,
0.1519775390625,
0.31640625,
0.10296630859375,
0.318115234375
],
[
0.50732421875,
0.298583984375,
0.7255859375,
0.29541015625,
0.7255859375,
0.314453125,
0.50732421875,
0.317626953125
],
[
0.0882568359375,
0.317626953125,
0.160888671875,
0.317626953125,
0.160888671875,
0.33203125,
0.0882568359375,
0.33203125
],
[
0.0882568359375,
0.33203125,
0.273681640625,
0.330810546875,
0.273681640625,
0.34521484375,
0.0882568359375,
0.3466796875
],
[
0.50244140625,
0.337158203125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.35107421875
],
[
0.50244140625,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.07843017578125,
0.361083984375,
0.4697265625,
0.361083984375,
0.4697265625,
0.37890625,
0.07843017578125,
0.37890625
],
[
0.5009765625,
0.365478515625,
0.890625,
0.3662109375,
0.890625,
0.380615234375,
0.5009765625,
0.380126953125
],
[
0.07843017578125,
0.376953125,
0.4599609375,
0.37744140625,
0.4599609375,
0.392578125,
0.07843017578125,
0.391357421875
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.0767822265625,
0.391357421875,
0.449462890625,
0.39013671875,
0.449462890625,
0.40771484375,
0.0767822265625,
0.4091796875
],
[
0.50244140625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.0767822265625,
0.4072265625,
0.4697265625,
0.40771484375,
0.4697265625,
0.425537109375,
0.0767822265625,
0.4248046875
],
[
0.50244140625,
0.41357421875,
0.8896484375,
0.41357421875,
0.8896484375,
0.4267578125,
0.50244140625,
0.4267578125
],
[
0.0767822265625,
0.4248046875,
0.468017578125,
0.423583984375,
0.468017578125,
0.438232421875,
0.0767822265625,
0.439453125
],
[
0.50244140625,
0.427978515625,
0.78515625,
0.427978515625,
0.78515625,
0.442626953125,
0.50244140625,
0.442626953125
],
[
0.080078125,
0.43994140625,
0.463134765625,
0.43994140625,
0.463134765625,
0.455078125,
0.080078125,
0.455078125
],
[
0.50341796875,
0.448974609375,
0.7060546875,
0.448974609375,
0.7060546875,
0.46337890625,
0.50341796875,
0.46337890625
],
[
0.07843017578125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.472900390625,
0.07843017578125,
0.472900390625
],
[
0.07843017578125,
0.4716796875,
0.471435546875,
0.4716796875,
0.471435546875,
0.4892578125,
0.07843017578125,
0.4892578125
],
[
0.521484375,
0.468505859375,
0.890625,
0.468505859375,
0.890625,
0.481689453125,
0.521484375,
0.481689453125
],
[
0.07843017578125,
0.488037109375,
0.28515625,
0.490478515625,
0.28515625,
0.5048828125,
0.07843017578125,
0.50244140625
],
[
0.50244140625,
0.484130859375,
0.89208984375,
0.484130859375,
0.89208984375,
0.49755859375,
0.50244140625,
0.49755859375
],
[
0.453369140625,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.50048828125,
0.453369140625,
0.50048828125
],
[
0.5009765625,
0.498779296875,
0.890625,
0.498779296875,
0.890625,
0.51220703125,
0.5009765625,
0.51220703125
],
[
0.07843017578125,
0.5068359375,
0.30810546875,
0.50830078125,
0.30810546875,
0.52294921875,
0.07843017578125,
0.521484375
],
[
0.5,
0.51220703125,
0.89453125,
0.51025390625,
0.89453125,
0.52783203125,
0.5,
0.52978515625
],
[
0.09722900390625,
0.5234375,
0.468017578125,
0.52294921875,
0.468017578125,
0.537109375,
0.09722900390625,
0.53857421875
],
[
0.5009765625,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54248046875,
0.5009765625,
0.54248046875
],
[
0.07843017578125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.55615234375,
0.07843017578125,
0.55615234375
],
[
0.50341796875,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.55615234375,
0.50341796875,
0.55615234375
],
[
0.080078125,
0.5546875,
0.2900390625,
0.5546875,
0.2900390625,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.283447265625,
0.556640625,
0.468017578125,
0.5546875,
0.468017578125,
0.5693359375,
0.283447265625,
0.5712890625
],
[
0.50244140625,
0.5595703125,
0.8935546875,
0.5595703125,
0.8935546875,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.07598876953125,
0.57080078125,
0.471435546875,
0.5712890625,
0.471435546875,
0.5888671875,
0.07598876953125,
0.58837890625
],
[
0.50244140625,
0.57373046875,
0.689453125,
0.57373046875,
0.689453125,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.08087158203125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60498046875,
0.08087158203125,
0.60498046875
],
[
0.521484375,
0.5888671875,
0.89208984375,
0.5888671875,
0.89208984375,
0.60302734375,
0.521484375,
0.60302734375
],
[
0.07843017578125,
0.60498046875,
0.4697265625,
0.60498046875,
0.4697265625,
0.619140625,
0.07843017578125,
0.619140625
],
[
0.50244140625,
0.60498046875,
0.8935546875,
0.60498046875,
0.8935546875,
0.61865234375,
0.50244140625,
0.61865234375
],
[
0.07843017578125,
0.619140625,
0.468017578125,
0.61865234375,
0.468017578125,
0.63330078125,
0.07843017578125,
0.6337890625
],
[
0.5009765625,
0.619140625,
0.708984375,
0.61865234375,
0.708984375,
0.63330078125,
0.5009765625,
0.6337890625
],
[
0.080078125,
0.63525390625,
0.46728515625,
0.63525390625,
0.46728515625,
0.6484375,
0.080078125,
0.6484375
],
[
0.646484375,
0.64013671875,
0.78173828125,
0.64013671875,
0.78173828125,
0.65478515625,
0.646484375,
0.65478515625
],
[
0.861328125,
0.64013671875,
0.8935546875,
0.64013671875,
0.8935546875,
0.65576171875,
0.861328125,
0.65576171875
],
[
0.07843017578125,
0.65087890625,
0.2181396484375,
0.65087890625,
0.2181396484375,
0.6640625,
0.07843017578125,
0.6640625
],
[
0.61376953125,
0.64599609375,
0.6259765625,
0.64599609375,
0.6259765625,
0.6513671875,
0.61376953125,
0.6513671875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.07843017578125,
0.66845703125,
0.27294921875,
0.66943359375,
0.27294921875,
0.6845703125,
0.07843017578125,
0.68310546875
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.09722900390625,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.09722900390625,
0.70263671875
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.08087158203125,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.7197265625,
0.4697265625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.7373046875,
0.85107421875,
0.7373046875,
0.85107421875,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.07843017578125,
0.74853515625,
0.4697265625,
0.75,
0.4697265625,
0.7646484375,
0.07843017578125,
0.76318359375
],
[
0.52294921875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.08087158203125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50244140625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.471435546875,
0.78076171875,
0.471435546875,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82568359375,
0.50244140625,
0.82568359375
],
[
0.080078125,
0.82470703125,
0.259033203125,
0.82470703125,
0.259033203125,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
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,
0.85107421875,
0.2091064453125,
0.85107421875,
0.2091064453125,
0.8642578125,
0.08087158203125,
0.8642578125
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.098876953125,
0.87060546875,
0.468017578125,
0.87060546875,
0.468017578125,
0.88525390625,
0.098876953125,
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,
0.88623046875,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"Algorithm 1: Solve MC-WNNM via ADMM",
"Algorithm 2: Color Image Denoising by MC-WNNM",
"Input: Matrices Y and W, μ> 1, Tol > 0, Ki;",
"Input:Noisy image ye,noise levels(or,g,b),K2;",
"Initialization: Xo = Zo = Ao = O, po > 0,",
"Initialization: x( = ye, y",
"-(0)",
"=ye;",
"T = False, k = 0;",
"for k = 1 : K2 do",
"While (T -- false) do",
"1. Set y(k) = x(k-1);",
"1. Update Xk+1 as",
"2. Extract local patches (y;-, from y(*);",
"Xh+1=(WTW+I)-1(WTWY+Zk-Ak)",
"for each patch y, do",
"2. Update Zk+1 by solving the problem",
"3.",
" Search non-local similar patches Y,:",
"4. Apply the MC-WNNM model (10) to Y, and",
"minz/z-(Xk+1+\"A)/+Z/lw*",
"3. Update Ak+1 as Ak+1 = A + Pk(Xh+1 - Zk+1)",
"obtain the estimated X;",
"4.UpdatePk+1=μ*Pk;",
"endfor",
"5. kk+1:",
"5. Aggregate [X,] to fom the image x();",
"if (Convergence condition is satisfied) or (k ≥ K))",
"6. T← True",
"end for",
"end if",
"Output: Denoised image x(K2)",
"end while",
"Output: Matrices X and Z",
"posed method with state-of-the-art denoising methods. in-",
"cluding CBM3D [14], MLP [9], WNNM [3], TNRD [11],",
"Zh+1llF = limk→P'llAh+1 -AllF = 0, and (a)",
"DnCNN [12]\"Noise Clinic”(NC) [25, 26], CC [19]",
"is proved.Then we can provethatlimk-oXk+1",
"and the commercial software Neat Image (NI) [28]. The",
"XllF ≤ limk-→(l(WTW + I)-1(WWY",
"Matlab source code of our MC-WNNM algorithm can",
"WTWZk-A)lF +P-'lA-Ak-IllF)=0and",
"be downloaded at http://www4.comp.polyu.edu.",
"hence (b) is proved. Finally,(c) can be proved by check-",
"hk/-cslzhang/code/MCwNNM.zip.",
"ing that limk→Zk+1 - Zkll ≤ limk→(k-1",
"4.1. Experimental Settings",
"S/p-1(Zk-1)lF+Xk+1-XlE+Ilp-Ak-1+",
"Pk'Ak+1 - Pk'AllF) = 0, where Uk-1Zk-1VT-1 is",
"Noise level estimation. For most of the competing de-",
"the SVDof X+Pk-Ak-1:",
"noising algorithms, the standard deviation of noise should",
"口",
"be given as a parameter. In synthetic experiments, the noise",
"3.4. The Denoising Algorithm",
"levels (or, og, ob) in R, G ,B channels are assumed to be",
"Given a noisy color image yc, suppose that we have ex-",
"known. In the case of real noisy images, the noise levels can",
"tracted N local patches (y;ly-1 and their similar patches.",
"beestimated viasomenoiseestimationmethods[31,32].In",
"N noisy patch matrices(Y,?",
"N-i can be formed to esti-",
"this paper, we employ the method [32] to estimate the noise",
"mate the clean matrices [X,-1. The patches in matrices",
"level for each color channel.",
"[X,-1 are aggregated to form the denoised image xXe. To",
"Noise level of comparison methods. For the CBM3D",
"obtain better denoising results, we perform the above de-",
"method [14], a single parameter of noise level should be",
"noising procedures for several rounds. The proposed MC-",
"input. We set the noise level as",
"WNNM based color image denoising algorithm is summa-",
"/(02 +g2 +02)/3",
"(21)",
"rized in Algorithm 2.",
"T",
"The methods of MLP [9] and TNRD [11] are originally de-",
"3.5.Complexity Analysis",
"signed for grayscale images. We retrain their models (using",
"In Algorithm 1 for solving the MC-WNNM model via",
"the released codes by the authors) at different noise levels",
"ADMM, the cost for updating X is O(max(p4M, M3)),",
"from α = 5 to α = 75 with a gap of 5. The denoising on",
"while the cost for updating Z is O(p4M + M3). The costs",
"color images is performed by processing each channel with",
"for updating A and p can be ignored. So the overall com-",
"the model trained atthe same (ornearest)noise level.",
"plexity is O((p4M + M3)Ki), where Ki is the number",
"Comparison with WNNM. In order to make a full and",
"ofiterations.In Algorithm2for image denoising,wecon-",
"fair comparison with the original WNNM method [21],",
"sider the number of patches N extracted from the input",
"we implement WNNM for color image denoising in three",
"noisy image and the number of itertations K2 and ignore",
"ways. 1) We apply WNNM to each color channel separately",
"the cost for searching similar patches. The overall cost is",
"with the corresponding noise levels or, O g, Ob. We call this",
"O((p4M + M3)KiK2N)",
"method *WNNM-1\". 2) We perform WNNM on the con-",
"catenated matrix Y formed by the patches in RGB chan-",
"4.Experiments",
"nels, while the input noise level is computed by Eq. (21).",
"We evaluate the proposed MC-WNNM method on syn-",
"We call this method \"*WNNM-2\". 3) We set the weight ma-",
"thetic and real noisy color images. We compare the pro-",
"trix W as W = α-1I in the proposed MC-WNNM model,",
"5"
] |
{
"arxiv_id": "1705.09912",
"language": "en",
"page_idx": 4
}
|
||
2205.01094-en-0
|
2205.01094-en-0.jpg
|
## A Word is Worth A Thousand Dollars: Adversarial Attack on Tweets Fools Stock Prediction
Yong Xie$^1,\textbf{Dakuo Wang}^{2,\dagger},\textbf{Pin-Yu Chen}^2,\textbf{Jinjun Xiong}^3,\textbf{Sijia Liu}^{2,4}$ and Sanmi Koyejo$^1$
$^{1}$University of Illinois Urbana-Champaign, $^2$IBM $^{3}$University at Buffalo, $^{4}$Michigan State University
### $\mathbf{Abstract}$
More and more investors and machine learning models rely on social media (e.g., Twitter and Reddit) to gather real-time information and sentiment to predict stock price movements. Although text-based models are known to be vulnerable to adversarial attacks, whether stock prediction models have similar vulnerability is underexplored. In this paper, we experiment with a variety of adversarial attack configurations to fool three stock prediction victim models. We address the task of adversarial generation by solving combinatorial optimization problems with semantics and budget constraints. Our results show that the proposed attack method can achieve consistent success rates and cause significant monetary loss in trading simulation by simply concatenating a perturbed but semantically similar tweet.
## 1 Introduction
The advance of deep learning based language models are playing a more and more important role in the financial context, including convolutional neutral network (CNN) (Ding et al., 2015), recurrent neutral network (RNN) (Minh et al., 2018), long short-term memory network (LSTM) (Hiew et al., 2019; Sawhney et al., 2021; Hochreiter and Schmidhuber, 1997), graph neutral network (GNN) (Sawhney et al., 2020a,b), transformer (Yang et al., 2020), autoencoder (Xu and Cohen, 2018), etc. For example, Antweiler and Frank (2004) find that comments on Yahoo Finance can predict stock market volatility after controlling the effect of news. Cookson and Niessner (2020) also show that sentiment disagreement on Stocktwits is highly related to certain market activities. Readers can refer to these survey papers for more details (Dang et al.,2020; Zhang et al., 2018; Xing et al., 2018).
Our
${\overline{{^{\dagger}\:\text{Corresp}}}}$onding author dakuo@acm.org.
code is available at https://github.com/yonxie/
${\mathrm{Adv}}$FinTweet
<FigureHere>
Figure 1: An example of word-replacement adversarial
attack. (Top) benign tweet leads Stocknet to predict stock going UP; (Bottom) adding an adversarial quote tweet leads Stocknet to predict stock going DOWN.
It is now known that text-based deep learning models can be vulnerable to adversarial attacks (Szegedy et al., 2014; Goodfellow et al., 2015). The perturbation can be at the sentence level (e.g., Xu et al., 2021; Iyyer et al., 2018; Ribeiro et al., 2018), the word level (e.g., Zhang et al., 2019; Alzantot et al., 2018; Zang et al., 2020; Jin et al., 2020; Lei et al., 2019; Zhang et al., 2021; Lin et al., 2021), or both (Chen et al., 2021). We are interested in whether such adversarial attack vulnerability also exists in stock prediction models, as these models embrace more and more human-generated media data (e.g., Twitter, Reddit, Stocktwit, Yahoo News (Xu and Cohen, 2018; Sawhney et al., 2021)). {The}(Xu and Cohen, 2018; Sawhney et al., 2021)). {The} adversarial robustness is a more critical issue in the context of stock prediction as anyone can post perturbed tweets or news to influence forecasting models. For example, a fake news (“Two Explosions in the White House and Barack Obama is Injured") posted by a hacker using the AssociatedPress's Twitter account on 04/23/2013 erased $136
|
[
[
0.301513671875,
0.09320068359375,
0.7001953125,
0.09320068359375,
0.7001953125,
0.10809326171875,
0.301513671875,
0.10809326171875
],
[
0.2283935546875,
0.11224365234375,
0.76904296875,
0.11224365234375,
0.76904296875,
0.1270751953125,
0.2283935546875,
0.1270751953125
],
[
0.118408203125,
0.15380859375,
0.8818359375,
0.15380859375,
0.8818359375,
0.171630859375,
0.118408203125,
0.171630859375
],
[
0.301513671875,
0.171630859375,
0.7021484375,
0.171630859375,
0.7021484375,
0.189453125,
0.301513671875,
0.189453125
],
[
0.2939453125,
0.1859130859375,
0.705078125,
0.189453125,
0.705078125,
0.207275390625,
0.2939453125,
0.2037353515625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.5751953125,
0.26318359375,
0.68115234375,
0.26318359375,
0.68115234375,
0.27490234375,
0.5751953125,
0.27490234375
],
[
0.74072265625,
0.26318359375,
0.8134765625,
0.26318359375,
0.8134765625,
0.273681640625,
0.74072265625,
0.273681640625
],
[
0.5517578125,
0.26953125,
0.5615234375,
0.26953125,
0.5615234375,
0.2724609375,
0.5517578125,
0.2724609375
],
[
0.57861328125,
0.275634765625,
0.62451171875,
0.275634765625,
0.62451171875,
0.283203125,
0.57861328125,
0.283203125
],
[
0.145263671875,
0.281982421875,
0.4609375,
0.281982421875,
0.4609375,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.71728515625,
0.27783203125,
0.83642578125,
0.27783203125,
0.83642578125,
0.2880859375,
0.71728515625,
0.2880859375
],
[
0.145263671875,
0.295654296875,
0.4609375,
0.295654296875,
0.4609375,
0.308837890625,
0.145263671875,
0.308837890625
],
[
0.53662109375,
0.29150390625,
0.85546875,
0.29150390625,
0.85546875,
0.3046875,
0.53662109375,
0.3046875
],
[
0.145263671875,
0.31005859375,
0.4609375,
0.31005859375,
0.4609375,
0.32373046875,
0.145263671875,
0.32373046875
],
[
0.53662109375,
0.30712890625,
0.861328125,
0.30712890625,
0.861328125,
0.320556640625,
0.53662109375,
0.320556640625
],
[
0.0235137939453125,
0.314208984375,
0.060455322265625,
0.314208984375,
0.060455322265625,
0.377197265625,
0.0235137939453125,
0.377197265625
],
[
0.145263671875,
0.32470703125,
0.459228515625,
0.32470703125,
0.459228515625,
0.33837890625,
0.145263671875,
0.33837890625
],
[
0.53564453125,
0.319580078125,
0.63037109375,
0.32177734375,
0.63037109375,
0.336181640625,
0.53564453125,
0.334228515625
],
[
0.145263671875,
0.33837890625,
0.459228515625,
0.33837890625,
0.459228515625,
0.3515625,
0.145263671875,
0.3515625
],
[
0.53564453125,
0.339111328125,
0.7421875,
0.339111328125,
0.7421875,
0.352783203125,
0.53564453125,
0.352783203125
],
[
0.145263671875,
0.352783203125,
0.459228515625,
0.352783203125,
0.459228515625,
0.366455078125,
0.145263671875,
0.366455078125
],
[
0.56494140625,
0.363525390625,
0.6826171875,
0.363525390625,
0.6826171875,
0.37353515625,
0.56494140625,
0.37353515625
],
[
0.73046875,
0.3623046875,
0.82958984375,
0.3623046875,
0.82958984375,
0.372314453125,
0.73046875,
0.372314453125
],
[
0.5439453125,
0.365234375,
0.55908203125,
0.365234375,
0.55908203125,
0.37353515625,
0.5439453125,
0.37353515625
],
[
0.0251922607421875,
0.36767578125,
0.05877685546875,
0.36767578125,
0.05877685546875,
0.4091796875,
0.0251922607421875,
0.4091796875
],
[
0.14697265625,
0.36767578125,
0.459228515625,
0.36767578125,
0.459228515625,
0.38134765625,
0.14697265625,
0.38134765625
],
[
0.568359375,
0.3759765625,
0.6337890625,
0.3759765625,
0.6337890625,
0.383056640625,
0.568359375,
0.383056640625
],
[
0.71533203125,
0.3759765625,
0.744140625,
0.377685546875,
0.7421875,
0.385498046875,
0.7138671875,
0.383056640625
],
[
0.7421875,
0.380126953125,
0.7841796875,
0.380126953125,
0.7841796875,
0.38427734375,
0.7421875,
0.38427734375
],
[
0.78857421875,
0.37890625,
0.8271484375,
0.37890625,
0.8271484375,
0.385986328125,
0.78857421875,
0.385986328125
],
[
0.53662109375,
0.3896484375,
0.85205078125,
0.3896484375,
0.85205078125,
0.402587890625,
0.53662109375,
0.402587890625
],
[
0.14697265625,
0.38134765625,
0.4609375,
0.38134765625,
0.4609375,
0.394287109375,
0.14697265625,
0.394287109375
],
[
0.145263671875,
0.394287109375,
0.459228515625,
0.394287109375,
0.459228515625,
0.407958984375,
0.145263671875,
0.407958984375
],
[
0.026031494140625,
0.40087890625,
0.05792236328125,
0.40087890625,
0.05792236328125,
0.429931640625,
0.026031494140625,
0.429931640625
],
[
0.145263671875,
0.410400390625,
0.459228515625,
0.410400390625,
0.459228515625,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.53564453125,
0.405029296875,
0.8037109375,
0.405029296875,
0.8037109375,
0.418701171875,
0.53564453125,
0.418701171875
],
[
0.145263671875,
0.42333984375,
0.459228515625,
0.42333984375,
0.459228515625,
0.43701171875,
0.145263671875,
0.43701171875
],
[
0.5341796875,
0.42041015625,
0.7197265625,
0.42041015625,
0.7197265625,
0.43408203125,
0.5341796875,
0.43408203125
],
[
0.0218353271484375,
0.438232421875,
0.05877685546875,
0.438232421875,
0.05877685546875,
0.51708984375,
0.0218353271484375,
0.51708984375
],
[
0.145263671875,
0.438232421875,
0.4609375,
0.438232421875,
0.4609375,
0.451904296875,
0.145263671875,
0.451904296875
],
[
0.5498046875,
0.439453125,
0.80029296875,
0.439453125,
0.80029296875,
0.453125,
0.5498046875,
0.453125
],
[
0.145263671875,
0.451904296875,
0.4609375,
0.451904296875,
0.4609375,
0.465576171875,
0.145263671875,
0.465576171875
],
[
0.54833984375,
0.454833984375,
0.8193359375,
0.454833984375,
0.8193359375,
0.465576171875,
0.54833984375,
0.465576171875
],
[
0.145263671875,
0.466064453125,
0.459228515625,
0.466064453125,
0.459228515625,
0.479736328125,
0.145263671875,
0.479736328125
],
[
0.5458984375,
0.465576171875,
0.8310546875,
0.465576171875,
0.8310546875,
0.478515625,
0.5458984375,
0.478515625
],
[
0.5458984375,
0.4755859375,
0.62890625,
0.47802734375,
0.62890625,
0.4921875,
0.5458984375,
0.490478515625
],
[
0.14697265625,
0.482177734375,
0.45751953125,
0.482177734375,
0.45751953125,
0.4921875,
0.14697265625,
0.4921875
],
[
0.145263671875,
0.49462890625,
0.459228515625,
0.49462890625,
0.459228515625,
0.50830078125,
0.145263671875,
0.50830078125
],
[
0.54248046875,
0.501953125,
0.744140625,
0.501953125,
0.744140625,
0.5126953125,
0.54248046875,
0.5126953125
],
[
0.145263671875,
0.5087890625,
0.45751953125,
0.5087890625,
0.45751953125,
0.5224609375,
0.145263671875,
0.5224609375
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.145263671875,
0.52392578125,
0.42236328125,
0.52392578125,
0.42236328125,
0.53759765625,
0.145263671875,
0.53759765625
],
[
0.51318359375,
0.51953125,
0.8818359375,
0.51953125,
0.8818359375,
0.533203125,
0.51318359375,
0.533203125
],
[
0.51123046875,
0.53369140625,
0.8818359375,
0.53369140625,
0.8818359375,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.51123046875,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5625,
0.51123046875,
0.5625
],
[
0.1376953125,
0.5634765625,
0.26025390625,
0.5634765625,
0.26025390625,
0.57763671875,
0.1376953125,
0.57763671875
],
[
0.509765625,
0.56103515625,
0.880859375,
0.5625,
0.880859375,
0.57666015625,
0.509765625,
0.576171875
],
[
0.12176513671875,
0.56787109375,
0.1318359375,
0.56787109375,
0.1318359375,
0.57470703125,
0.12176513671875,
0.57470703125
],
[
0.1351318359375,
0.56787109375,
0.1427001953125,
0.56787109375,
0.1427001953125,
0.57373046875,
0.1351318359375,
0.57373046875
],
[
0.1158447265625,
0.58447265625,
0.489501953125,
0.58447265625,
0.489501953125,
0.59716796875,
0.1158447265625,
0.59716796875
],
[
0.5283203125,
0.58056640625,
0.8818359375,
0.5830078125,
0.8818359375,
0.5966796875,
0.5283203125,
0.59423828125
],
[
0.1175537109375,
0.60107421875,
0.48779296875,
0.60107421875,
0.48779296875,
0.6142578125,
0.1175537109375,
0.6142578125
],
[
0.509765625,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.509765625,
0.6123046875
],
[
0.1158447265625,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.1158447265625,
0.6298828125
],
[
0.51123046875,
0.6142578125,
0.8818359375,
0.6142578125,
0.8818359375,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.1158447265625,
0.6318359375,
0.490234375,
0.6318359375,
0.490234375,
0.6455078125,
0.1158447265625,
0.6455078125
],
[
0.509765625,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.509765625,
0.6435546875
],
[
0.1158447265625,
0.6484375,
0.489501953125,
0.6484375,
0.489501953125,
0.662109375,
0.1158447265625,
0.662109375
],
[
0.51123046875,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.51123046875,
0.6611328125
],
[
0.1158447265625,
0.66455078125,
0.489501953125,
0.66455078125,
0.489501953125,
0.677734375,
0.1158447265625,
0.677734375
],
[
0.509765625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.509765625,
0.677734375
],
[
0.1175537109375,
0.6796875,
0.48779296875,
0.6796875,
0.48779296875,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.509765625,
0.6787109375,
0.8818359375,
0.6787109375,
0.8818359375,
0.693359375,
0.509765625,
0.693359375
],
[
0.1175537109375,
0.6953125,
0.489501953125,
0.6953125,
0.489501953125,
0.71044921875,
0.1175537109375,
0.71044921875
],
[
0.509765625,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.708984375,
0.509765625,
0.708984375
],
[
0.118408203125,
0.7119140625,
0.490234375,
0.7119140625,
0.490234375,
0.7255859375,
0.118408203125,
0.7255859375
],
[
0.509765625,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.509765625,
0.724609375
],
[
0.1175537109375,
0.72802734375,
0.489501953125,
0.72802734375,
0.489501953125,
0.7421875,
0.1175537109375,
0.7421875
],
[
0.509765625,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.1158447265625,
0.74462890625,
0.491943359375,
0.74365234375,
0.491943359375,
0.75830078125,
0.1158447265625,
0.7587890625
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1175537109375,
0.76123046875,
0.48779296875,
0.76123046875,
0.48779296875,
0.77490234375,
0.1175537109375,
0.77490234375
],
[
0.51123046875,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.11419677734375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.11419677734375,
0.7890625
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.51123046875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1175537109375,
0.80859375,
0.489501953125,
0.80859375,
0.489501953125,
0.822265625,
0.1175537109375,
0.822265625
],
[
0.51123046875,
0.80859375,
0.880859375,
0.80859375,
0.880859375,
0.8193359375,
0.51123046875,
0.8193359375
],
[
0.1175537109375,
0.82666015625,
0.48779296875,
0.82666015625,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.51123046875,
0.82470703125,
0.88330078125,
0.82470703125,
0.88330078125,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.11419677734375,
0.84130859375,
0.489501953125,
0.8388671875,
0.489501953125,
0.85400390625,
0.11419677734375,
0.85546875
],
[
0.509765625,
0.84130859375,
0.8818359375,
0.84130859375,
0.8818359375,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.3955078125,
0.85693359375,
0.3955078125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.154541015625,
0.88134765625,
0.43408203125,
0.88134765625,
0.43408203125,
0.89453125,
0.154541015625,
0.89453125
],
[
0.459228515625,
0.8818359375,
0.486083984375,
0.8818359375,
0.486083984375,
0.8935546875,
0.459228515625,
0.8935546875
],
[
0.11419677734375,
0.892578125,
0.486083984375,
0.8935546875,
0.486083984375,
0.90673828125,
0.11419677734375,
0.90625
],
[
0.509765625,
0.88720703125,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90673828125,
0.2166748046875,
0.90673828125,
0.2166748046875,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"A Word is Worth A Thousand Dollars:",
"Adversarial Attack on Tweets Fools Stock Prediction",
"Yong Xie', Dakuo Wang?t, Pin-Yu Chen?, Jinjun Xiong', Sijia Liu?,4 and Sanmi Koyejo",
"1University of Illinois Urbana-Champaign, 2IBM",
"3University at Buffalo, AMichigan State University",
"Abstract",
"wallstreetbet7821",
"RawTweet",
"",
"raw.tweet",
"More and more investors and machine learn-",
"Prediction:Positive",
"ing models rely on social media (e.g., Twit-",
"sBHP announcesthedemergerof itsnon-",
"ter and Reddit) to gather real-time informa-",
"core assets-details expected to be filledin",
"",
"tion and sentiment to predict stock price move-",
"onTuesday.",
"ments. Although text-based models are known",
"3:00PM-Aug15,2014.TwitterforiPhone",
"to be vulnerable to adversarial attacks, whether",
"wallstreetbet_adver",
"AttackRetweet",
"中",
"[nf",
"stock prediction models have similar vulnera-",
"adver_retwee",
"Pred",
"eon:",
"Negat",
"$BHP announces the demerger of its non-",
"bility is underexplored. In this paper, we exper-",
"iment with a variety of adversarial attack con-",
"2",
"figurations to fool three stock prediction vic-",
"core assets-details expected to be",
"tim models. We address the task of adversarial",
"exercisedinonTuesday",
"[S",
"generation by solving combinatorial optimiza-",
"wallstreetbet7821@raw_tweet.1h",
"tion problems with semantics and budget con-",
"$BHP announcesthedemergerofitsnon-",
"straints. Our results show that the proposed",
"coreassets-details expected to be filled in",
"on Tuesday.",
"attack method can achieve consistent success",
"rates and cause significant monetary loss in",
"4:00PM-Aug15,2014.TwitterforiPhone",
"trading simulation by simply concatenating a",
"6010502",
"perturbed but semantically similar tweet.",
"Figure l: An example of word-replacement adversarial",
"attack. (Top) benign tweet leads St ocknet to predict",
"stock going UP; (Bottom) adding an adversarial quote",
"Introduction",
"tweet leads st ocknet to predict stock going DOWN.",
"1",
"",
"The advance of deeplearningbased languagemod-",
"It is now known that text-based deep learning",
"els are playing a more and more important role",
"models can be vulnerable to adversarial attacks",
"in the financial context, including convolutional",
"(Szegedy et al., 2014; Goodfellow et al., 2015). The",
"neutral network (CNN) (Ding et al., 2015), recur-",
" perturbation can be at the sentence level (e.g., Xu",
"rent neutral network (RNN) (Minh et al., 2018),",
"et al., 2021; lyyer et al., 2018; Ribeiro et al., 2018),",
"long short-term memory network (LSTM) (Hiew",
"the word level (e.g., Zhang et al., 2019; Alzantot",
"et al., 2019; Sawhney et al., 2021; Hochreiter and",
"et al., 2018; Zang et al., 2020; Jin et al., 2020; Lei",
"Schmidhuber, 1997), graph neutral network (GNN)",
"et al., 2019; Zhang et al., 2021; Lin et al., 2021),",
"(Sawhney et al., 2020a,b), transformer (Yang et al.,",
"or both (Chen et al., 2021). We are interested in",
"2020), autoencoder (Xu and Cohen, 2018), etc. For",
"whether such adversarial attack vulnerability also",
"example, Antweiler and Frank (2004) find that com-",
"exists in stock prediction models, as these models",
"ments on Yahoo Finance can predict stock market",
"embrace more and more human-generated media",
"volatility after controlling the effect of news. Cook-",
"data (e.g., Twitter, Reddit, Stocktwit, Yahoo News",
"son and Niessner (2020) also show that sentiment",
"(Xu and Cohen, 2018; Sawhney et al., 2021)). The",
"disagreement on Stocktwits is highly related to cer-",
"adversarial robustness is a more critical issue in",
"tain market activities. Readers can refer to these",
"the context of stock prediction as anyone can post",
"survey papers for more details (Dang et al., 2020;",
"perturbed tweets or news to influence forecasting",
"Zhang et al., 2018; Xing et al., 2018).",
"models. For example, a fake news (\"Two Explo",
" sions in the White House and Barack Obama is",
" Corresponding author dakuo@acm.org.",
"Our",
"code is available at https://github.com/yonxie/",
"Injured') posted by a hacker using the Associated-",
"AdvFinTweet",
"Press's Twitter account on 04/23/2013 erased $136"
] |
{
"arxiv_id": "2205.01094",
"language": "en",
"page_idx": 0
}
|
||
2205.01094-en-1
|
2205.01094-en-1.jpg
|
billion market value in just 60 seconds (Fisher, 2013). Although the event doesn't fall into the category of adversarial attack, it rings the alarm for traders who use (social) media information for their trading decisions.
To our best knowledge, it is the first paper to consider the adversarial attack in the financial NLP literature. Many attacks modify benign text directly $( manipulation\textit{ attack) and use them as model input; }$ However, in our case, adversarial quote tweets enter the model along with benign tweets (concatenation attack), which is more realistic as malicious Twitter users can not modify others' tweets. In other words, we formulate the task as a text-concatenating attack (Jia and Liang, 2017; Le et al., 2021): we implement the attack by injecting new tweets instead of manipulating existing benign tweets. Our task is inspired and mimics the quote tweet function on social media, and uses it to feed the adversarial samples into the dataset. Despite various algorithms are proposed to generate manipulation attack, literature of concatenation attack on classification models is rare, with exceptions Le et al. (2021), Song et al. (2021) and Wang et al. (2020). Our paper provides extra evidence of their difference by investigating their performances in the financial domain.
The main challenge is to craft new and effective adversarial tweets. We solve the task by aligning the semantics with benign tweets so that the potential human and machine readers can't detect our adversarial tweets. To achieve that, we consider the generation task as a combinatorial optimization problem (Zang et al., 2020; Guo et al., 2021). Specific tweets are first selected, which are used as the target of perturbation on a limit number of words within the tweets. We then examine our attack method on three financial forecast models with attack success rate, FI and potential profit and loss as evaluation metrics. Results show that our attack method consistently achieves good success rate on the victim models. More astonishingly, the attack can cause additional loss of 23% to 32% if an investor trades on the predictions of the victim models (Fig. 4).
## 2 $\textbf{Adversarial Attack on Stock}$ Prediction Models with Tweet Data
Attack model: Adversarial tweets. In the case of Twitter, adversaries can post malicious tweets which are crafted to manipulate downstream models that take them as input. We propose to attack by
posting semantically similar adversarial tweets as quote tweets on Twitter, so that they could be identified as relevant information and collected as model input. For example, as shown in Fig 1, the original authentic tweet by the user wallstreetbet7821 was“SBHP announces the demerger of its noncore assets - details expected to be filled in on Tuesday." An adversarial sentence could be “$BHP announces the demerger of its non-core assets- details expected to be exercised in on Tuesday. ". The outcome of the victim model switches to negative prediction from positive prediction when the quote tweet is added to the input.
The proposed attack method takes the practical implementation into its design consideration, thus has many advantages. First, the adversarial tweets are crafted based on carefully-selected relevant tweets, so they are more likely to pass the models’ tweet filter and enter the inference data corpus. Secondly, adversarial tweets are optimized to be semantically similar to the original tweets so that they are not counterfactual and very likely to fool human sanity checks as well as the Twitter's content moderation system.
Attack generation: Hierarchical perturbation. The challenge of our attack method centers around how to select the optimal tweets and the token perturbations with the constraints of semantic similarity. In this paper, we formulate the task as a hierarchical perturbation consisting of three steps: tweet selection, word selection and word perturbation. In the first step, a set of optimal tweets is first selected as the target tweets to be perturbed and retweeted. For each selected tweet in the pool, the word selection problem is then solved to find one or more optimal words to apply perturbation. Word and tweet budgets are also introduced to quantify the strength of the perturbation.
We consider the word replacement and deletion for word perturbation (Garg and Ramakrishnan, 2020; Li et al., 2020). In the former case, the final step is to find the optimal candidate as replacement. A synonym as replacement is widely adopted in the word-level attack since it is a natural choice to preserve semantics (Zang et al., 2020; Dong et al., 2021; Zhang et al., 2019; Jin et al., 2020). Therefore, we replace target words with their synonyms chosen from synonym sets which contain the semantically closest words measured by the similarity of the GLOVE embedding (Jin et al., 2020).
|
[
[
0.1175537109375,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.51318359375,
0.10394287109375,
0.87890625,
0.10394287109375,
0.87890625,
0.114013671875,
0.51318359375,
0.114013671875
],
[
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
],
[
0.1158447265625,
0.133056640625,
0.48779296875,
0.133056640625,
0.48779296875,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.509765625,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.509765625,
0.1484375
],
[
0.118408203125,
0.1519775390625,
0.285400390625,
0.1519775390625,
0.285400390625,
0.162109375,
0.118408203125,
0.162109375
],
[
0.51123046875,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.163330078125,
0.51123046875,
0.163330078125
],
[
0.1368408203125,
0.16748046875,
0.489501953125,
0.16748046875,
0.489501953125,
0.18115234375,
0.1368408203125,
0.18115234375
],
[
0.509765625,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.509765625,
0.179931640625
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.51123046875,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19482421875,
0.51123046875,
0.19482421875
],
[
0.1175537109375,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.1175537109375,
0.213134765625
],
[
0.5087890625,
0.19775390625,
0.8818359375,
0.196533203125,
0.8818359375,
0.21142578125,
0.5087890625,
0.2120361328125
],
[
0.118408203125,
0.21728515625,
0.48779296875,
0.21728515625,
0.48779296875,
0.22802734375,
0.118408203125,
0.22802734375
],
[
0.509765625,
0.215576171875,
0.88330078125,
0.215576171875,
0.88330078125,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.1158447265625,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1158447265625,
0.2452392578125
],
[
0.509765625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.509765625,
0.24462890625
],
[
0.1175537109375,
0.2493896484375,
0.48779296875,
0.2493896484375,
0.48779296875,
0.260009765625,
0.1175537109375,
0.260009765625
],
[
0.509765625,
0.2452392578125,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.509765625,
0.260009765625
],
[
0.11419677734375,
0.26416015625,
0.489501953125,
0.26318359375,
0.489501953125,
0.27783203125,
0.11419677734375,
0.279052734375
],
[
0.51123046875,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.1158447265625,
0.28076171875,
0.489501953125,
0.28076171875,
0.489501953125,
0.29443359375,
0.1158447265625,
0.29443359375
],
[
0.509765625,
0.27978515625,
0.7119140625,
0.27978515625,
0.7119140625,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1158447265625,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.1158447265625,
0.31005859375
],
[
0.52880859375,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.31005859375,
0.52880859375,
0.31005859375
],
[
0.1175537109375,
0.312255859375,
0.490234375,
0.312255859375,
0.490234375,
0.325927734375,
0.1175537109375,
0.325927734375
],
[
0.51123046875,
0.312255859375,
0.88330078125,
0.312255859375,
0.88330078125,
0.325927734375,
0.51123046875,
0.325927734375
],
[
0.1158447265625,
0.328857421875,
0.489501953125,
0.328857421875,
0.489501953125,
0.342529296875,
0.1158447265625,
0.342529296875
],
[
0.5087890625,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.5087890625,
0.341552734375
],
[
0.1158447265625,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1158447265625,
0.35693359375
],
[
0.5087890625,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.35693359375,
0.5087890625,
0.35693359375
],
[
0.118408203125,
0.3623046875,
0.486083984375,
0.3623046875,
0.486083984375,
0.372314453125,
0.118408203125,
0.372314453125
],
[
0.51123046875,
0.361083984375,
0.8818359375,
0.361083984375,
0.8818359375,
0.374755859375,
0.51123046875,
0.374755859375
],
[
0.1175537109375,
0.377197265625,
0.489501953125,
0.377197265625,
0.489501953125,
0.39013671875,
0.1175537109375,
0.39013671875
],
[
0.51123046875,
0.377197265625,
0.8818359375,
0.377197265625,
0.8818359375,
0.39013671875,
0.51123046875,
0.39013671875
],
[
0.1175537109375,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.1175537109375,
0.40625
],
[
0.51318359375,
0.394287109375,
0.880859375,
0.394287109375,
0.880859375,
0.405029296875,
0.51318359375,
0.405029296875
],
[
0.1175537109375,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.51123046875,
0.410400390625,
0.880859375,
0.410400390625,
0.880859375,
0.42041015625,
0.51123046875,
0.42041015625
],
[
0.1158447265625,
0.427001953125,
0.484375,
0.427001953125,
0.484375,
0.43701171875,
0.1158447265625,
0.43701171875
],
[
0.509765625,
0.424560546875,
0.880859375,
0.424560546875,
0.880859375,
0.438232421875,
0.509765625,
0.438232421875
],
[
0.1175537109375,
0.441162109375,
0.489501953125,
0.441162109375,
0.489501953125,
0.454833984375,
0.1175537109375,
0.454833984375
],
[
0.51123046875,
0.4423828125,
0.880859375,
0.4423828125,
0.880859375,
0.453125,
0.51123046875,
0.453125
],
[
0.1175537109375,
0.4560546875,
0.489501953125,
0.457275390625,
0.489501953125,
0.471435546875,
0.1175537109375,
0.47021484375
],
[
0.51123046875,
0.457763671875,
0.71875,
0.457763671875,
0.71875,
0.471435546875,
0.51123046875,
0.471435546875
],
[
0.1158447265625,
0.471435546875,
0.486083984375,
0.473876953125,
0.486083984375,
0.488037109375,
0.1158447265625,
0.486328125
],
[
0.1158447265625,
0.4892578125,
0.435791015625,
0.488037109375,
0.435791015625,
0.5029296875,
0.1158447265625,
0.50439453125
],
[
0.51318359375,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.498779296875,
0.51318359375,
0.498779296875
],
[
0.1368408203125,
0.505859375,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.1368408203125,
0.51953125
],
[
0.509765625,
0.5029296875,
0.87890625,
0.5029296875,
0.87890625,
0.51318359375,
0.509765625,
0.51318359375
],
[
0.11419677734375,
0.521484375,
0.48779296875,
0.5224609375,
0.48779296875,
0.53759765625,
0.11419677734375,
0.5361328125
],
[
0.509765625,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.509765625,
0.53076171875
],
[
0.1158447265625,
0.5390625,
0.490234375,
0.5390625,
0.490234375,
0.552734375,
0.1158447265625,
0.552734375
],
[
0.51123046875,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.5458984375,
0.51123046875,
0.5458984375
],
[
0.1158447265625,
0.5546875,
0.489501953125,
0.5546875,
0.489501953125,
0.568359375,
0.1158447265625,
0.568359375
],
[
0.509765625,
0.5498046875,
0.88330078125,
0.5498046875,
0.88330078125,
0.5634765625,
0.509765625,
0.5634765625
],
[
0.1158447265625,
0.57080078125,
0.489501953125,
0.57080078125,
0.489501953125,
0.58447265625,
0.1158447265625,
0.58447265625
],
[
0.51318359375,
0.5654296875,
0.88330078125,
0.5654296875,
0.88330078125,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.1175537109375,
0.58740234375,
0.489501953125,
0.58740234375,
0.489501953125,
0.60107421875,
0.1175537109375,
0.60107421875
],
[
0.509765625,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.509765625,
0.595703125
],
[
0.1175537109375,
0.6025390625,
0.48779296875,
0.6025390625,
0.48779296875,
0.6162109375,
0.1175537109375,
0.6162109375
],
[
0.509765625,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.509765625,
0.6123046875
],
[
0.1175537109375,
0.619140625,
0.48779296875,
0.619140625,
0.48779296875,
0.6328125,
0.1175537109375,
0.6328125
],
[
0.509765625,
0.6142578125,
0.8818359375,
0.6142578125,
0.8818359375,
0.62744140625,
0.509765625,
0.62744140625
],
[
0.5087890625,
0.62890625,
0.8818359375,
0.6298828125,
0.8818359375,
0.64453125,
0.5087890625,
0.6435546875
],
[
0.1175537109375,
0.6357421875,
0.484375,
0.6357421875,
0.484375,
0.646484375,
0.1175537109375,
0.646484375
],
[
0.1175537109375,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.1175537109375,
0.6650390625
],
[
0.51123046875,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.66015625,
0.51123046875,
0.66015625
],
[
0.1158447265625,
0.66650390625,
0.486083984375,
0.66650390625,
0.486083984375,
0.6796875,
0.1158447265625,
0.6796875
],
[
0.51123046875,
0.662109375,
0.8818359375,
0.662109375,
0.8818359375,
0.6767578125,
0.51123046875,
0.6767578125
],
[
0.1175537109375,
0.68408203125,
0.48779296875,
0.68408203125,
0.48779296875,
0.69775390625,
0.1175537109375,
0.69775390625
],
[
0.51318359375,
0.6787109375,
0.880859375,
0.6787109375,
0.880859375,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.1175537109375,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.1175537109375,
0.71337890625
],
[
0.51123046875,
0.69482421875,
0.748046875,
0.69482421875,
0.748046875,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.1158447265625,
0.71630859375,
0.48779296875,
0.71630859375,
0.48779296875,
0.72998046875,
0.1158447265625,
0.72998046875
],
[
0.52880859375,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.52880859375,
0.724609375
],
[
0.1158447265625,
0.732421875,
0.489501953125,
0.732421875,
0.489501953125,
0.7451171875,
0.1158447265625,
0.7451171875
],
[
0.509765625,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.1158447265625,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.1158447265625,
0.76123046875
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1158447265625,
0.76416015625,
0.489501953125,
0.76416015625,
0.489501953125,
0.77783203125,
0.1158447265625,
0.77783203125
],
[
0.51318359375,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1158447265625,
0.78076171875,
0.2359619140625,
0.78076171875,
0.2359619140625,
0.79443359375,
0.1158447265625,
0.79443359375
],
[
0.51123046875,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.51123046875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1158447265625,
0.81005859375,
0.39208984375,
0.81005859375,
0.39208984375,
0.82470703125,
0.1158447265625,
0.82470703125
],
[
0.51123046875,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.14697265625,
0.82763671875,
0.450927734375,
0.82763671875,
0.450927734375,
0.84130859375,
0.14697265625,
0.84130859375
],
[
0.51318359375,
0.82373046875,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.51123046875,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90380859375,
0.60693359375,
0.90380859375,
0.60693359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"billion market value in just 60 seconds (Fisher,",
"posting semantically similar adversarial tweets as",
"2013). -Although the event doesn't fall into the",
"quotetweetsonTwitter,sothatthey couldbeidenti",
"category of adversarial attack, it rings the alarm",
"fied as relevant information and collected as model",
"for traders who use (social) media information for",
"input. For example, as shown in Fig l, the origi-",
"their trading decisions.",
"nal authentic tweet by the user wallstreetbet7821",
"To our best knowledge, it is the first paper to",
"was \"sBHP announces the demerger of its non",
"consider the adversarial attack in the financial NLP",
"core assets- details expected to be filled in on",
"literature. Many attacks modify benign text directly",
"Tuesday\"An adversarial sentence could be“\"sBHP",
"(manipulationattack)and usethemas modelinput:",
"announces the demerger of its non-core assets- de-",
"However, in our case, adversarial quote tweets enter",
"tails expected to be exercised in on Tuesday\". The",
"themodelalongwithbenigntweets(concatenation",
"outcome of the victim model switches to negative",
"attack), which is more realistic as malicious Twitter",
"prediction from positive prediction when the quote",
"users can not modify others' tweets. In other words,",
" tweet is added to the input.",
"we formulate the task as a text-concatenating attack",
"The proposed attack method takes the practi",
"(Jia and Liang, 2017; Le et al., 2021): we imple-",
"cal implementation into its design consideration,",
"ment the attack by injecting new tweets instead of",
"thus has many advantages. First, the adversarial",
"manipulating existing benign tweets. Our task is",
"tweets are crafted based on carefully-selected rel",
"inspiredandmimicstheguotetweetfunction on",
"evant tweets, so they are more likely to pass the",
"social media, and uses it to feed the adversarial sam-",
"models' tweet filter and enter the inference data",
"ples into the dataset. Despite various algorithms are",
"corpus. Secondly,adversarial tweets are optimized",
"proposed to generate manipulation attack, literature",
"tobesemanticallysimilartotheoriginal tweetsso",
"ofconcatenationattackonclassificationmodelsis",
" that they are not counterfactual and very likely to",
"rare, with exceptions Le et al. (2021), Song et al.",
"fool human sanity checks as well as the Twitter's",
"(2021) and Wang et al. (2020). Our paper provides",
"content moderation system.",
"extra evidence of their difference by investigating",
"their performances in the financial domain.",
"Attack generation: Hierarchical perturbation.",
"The main challenge is to craft new and effective",
"The challenge of our attackmethod centersaround",
"adversarial tweets. We solve the task by aligning",
"how to select the optimal tweets and the token per",
"the semantics with benign tweets so that the poten-",
"turbations with the constraints of semantic simi-",
"tial human and machine readers can't detect our",
"larity. In this paper, we formulate the task as a",
"adversarial tweets. To achieve that, we consider",
"hierarchical perturbation consisting of three steps:",
"the generation task as a combinatorial optimiza-",
"tweet selection, word selection and word perturba-",
"tion problem (Zang et al., 2020; Guo et al., 2021).",
"tion. In the first step, a set of optimal tweets is first",
"Specific tweets are first selected, which are used",
" selected as the target tweets to be perturbed and",
"retweeted. For each selected tweet in the pool, the",
"as the target of perturbation on a limit number of",
"words within the tweets.We then examine our",
"word selection problem is then solved to find one",
"attack method on three financial forecast models",
"or more optimal words to apply perturbation. Word",
"with attack success rate, F1 and potential profit and",
"and tweet budgets are also introduced to quantify",
"loss as evaluation metrics. Results show that our",
"the strength of the perturbation.",
"attack method consistently achieves good success",
"We consider the word replacement and dele-",
"rate on the victim models. More astonishingly, the",
"tion for word perturbation (Garg and Ramakrish-",
"attack can cause additional loss of 23% to 32% if",
"nan, 2020; Li et al., 2020). In the former case,",
"an investor trades on the predictions of the victim",
"the final step is to find the optimal candidate as",
"models (Fig. 4).",
"replacement. A synonym as replacement is widely",
"adopted in the word-level attack since it is a natural",
"2Adversarial Attack on Stock",
"choice to preserve semantics (Zang et al., 2020:",
"Prediction Models with Tweet Data",
"Dong et al., 2021; Zhang et al., 2019; Jin et al.,",
"2020). Therefore, we replace target words with",
"Attack model: Adversarial tweets.In the case",
"their synonyms chosen from synonym sets which",
"of Twitter, adversaries can post malicious tweets",
"contain the semantically closest words measured",
"which are crafted to manipulate downstream mod-",
"by the similarity of the GLOVE embedding (Jin",
"els that take them as input. We propose to attack by",
"et al., 2020)."
] |
{
"arxiv_id": "2205.01094",
"language": "en",
"page_idx": 1
}
|
||
2205.01094-en-2
|
2205.01094-en-2.jpg
|
Mathematical Formulation. We consider a multimodal stock forecast model $f(\cdot)$ that takes tweet collections $\{c_t\}_{t=1}^T$ and numerical factors $\{\boldsymbol{p_t}\}_{t=1}^T$ as input, where $t$ indexes the date when the data is collected. Peeking into the tweet collection, it contains $|c_t|$ tweets for date $t$, namely, $c_{t}=\{s_{t}^{1},s_{t}^{2},...,s_{t}^{|c_{t}|}\}.$ Each tweet $s_{t}^{i}$ is a textbased sentence of length $|s_t^i|$, denoted as $s_t^i=$ $(w_{t}^{i,1},...,w_{t}^{i,j},...,w_{t}^{i,|\boldsymbol{s}_{t}^{i}|})$,for $i=1,...,|\boldsymbol{c}_{t}|.$ A di-$\begin{array}{ccc}3&\mathbf{Experiments}\\\mathbf{Dataset~\&~victim~u}\end{array}$
rectional fnancial forecast model takes domains of tweets and numerical factors as input, and yields prediction for stocks' directional movement $y\in\{-1,1\}\colon$
$$\hat{y}_{t+1}=f(c_{\boldsymbol{t}-\boldsymbol{h}:\boldsymbol{t}},\boldsymbol{p}_{\boldsymbol{t}-\boldsymbol{h}:\boldsymbol{t}}),$$
(1)
$h$ is the looking-back window for historical data.
The hierarchical perturbation can be cast as a combinatorial problem for tweet selection $m$, word selection $z$ and replacement selection $u.$ The boolean vector $m$ indicates the tweets to be selected. For $i$-th tweet, vector $z_i$ indicates the word to be perturbed. As for the word perturbation task, another boolean vector $u_{i,j}$ selects the best replacement. It follows that the hierarchical perturbation can be formulated as
(2)
$$c_t^{\prime}=(1-m\cdot z)\cdot c_t+m\cdot z\cdot u\cdot S(c_t),$$
where ·denotes element-column wise product, $m\cdot z$ indicates the selected words in selected tweets, $m\cdot z\cdot u$ indicates selected synonyms for each selected word, and $S(\cdot)$ is a elementwise synonym generating function. Consequently, given attack loss $\mathcal{L}$, generation of adversarial quote tweets can be formulated as the optimization pro- $\operatorname*{gram}_{m,\boldsymbol{z},u}\mathcal{L}(c_{t}^{\prime}\cup c_{t-h:t},c_{t-h:t}|p_{t-h:t},f)$,subject to the budget constraints: $a)1^T\boldsymbol{m}\leq b_\mathrm{s},b)$ $1^{T}z_{i}$ $\leq$ $b_{w}, \forall i$ and $c)$ $1^{T}u_{i, j}$ = $1, \forall i, j$, where $b_s$ and $b_w$ denote the tweet and word budgets. It is worth to stress that perturbation is only applied to the date $(t)$ when the attack is implemented to preserve the temporal order.
To solve the program, we follow the convex relaxation approach developed in (Srikant et al., 2021). Specifically, the boolean variables (for tweet and word selection) are relaxed into the continuous space so that they can be optimized by gradientbased methods over a convex hull. Two main implementations of the optimization-based attack generation method are proposed: $joint\textit{optimization}$ (JO) solver and alternating greedy optimization (AGO) solver. JO calls projected gradient descent
method to optimize the tweet and word selection variables and word replacement variables simultaneously. AGO uses an alternative optimization procedure to sequentially update the discrete selection variables and the replacement selection variables. More details on the optimization program and the solvers can be found in Appendix A.
Dataset & victim models. We evaluate our adversarial attack on a stock prediction dataset consisting of 10,824 instances including relevant tweets and numerical features of 88 stocks from 2014 to 2016 (Xu and Cohen, 2018). Three models (Stocknet (Xu and Cohen, 2018), FinGRU based on GRU (Cho et al., 2014) and FinLSTM based on LSTM (Hochreiter and Schmidhuber, 1997)) of binary classification are considered as victims in this paper. We apply our attack to instances on which the victim models make correct predictions.
Evaluation metrics. Attack performance is evaluated by two metrics: Attack Success Rate (ASR) and victim model's $Fl$ drop after attack. ASR is defined as the percentage of the attack efforts that changes the model output. The two metrics gauge the efficacy of the attack and its impact on model performance: More efficient attack leads to higher ASR and more decline of F1. Moreover, we simulate a Long-Only Buy-Hold-Sell strategy (Sawhney et al., 2021; Feng et al., 2019) with victim models, and calculate the Profit and Loss (PnL) for each simulation. Assume a portfolio starts with initial net value $10000 (100%), its net value at the end of test period reflects the profitability of the trading strategy and the underlying model. Consequently, the change in PnLs measures the monetary impact of our attack. More details on the dataset, victim models and evaluation metrics are housed $\operatorname{in}$Appendix B.
## 4 Results
Attack performance with single perturbation. The experiment results for the concatenation attack with word replacement perturbation is shown in Table 1 (with tweet and word budgets both as 1). For both JO and AGO, ASR increases by roughly 10% and F1 drops by 0.1 on average in comparison to the random attack. Such performance drop is considered significant in the context of stock prediction given that the state-of-the-art prediction
|
[
[
0.1175537109375,
0.08489990234375,
0.490234375,
0.08489990234375,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.08489990234375,
0.87890625,
0.08489990234375,
0.87890625,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.1175537109375,
0.1015625,
0.486083984375,
0.1015625,
0.486083984375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.133056640625,
0.1175537109375,
0.133056640625
],
[
0.51318359375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.118408203125,
0.1341552734375,
0.48779296875,
0.133056640625,
0.48779296875,
0.1478271484375,
0.118408203125,
0.1484375
],
[
0.509765625,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.509765625,
0.1478271484375
],
[
0.1175537109375,
0.149658203125,
0.490234375,
0.149658203125,
0.490234375,
0.163330078125,
0.1175537109375,
0.163330078125
],
[
0.509765625,
0.15087890625,
0.880859375,
0.15087890625,
0.880859375,
0.160888671875,
0.509765625,
0.160888671875
],
[
0.1175537109375,
0.166259765625,
0.49365234375,
0.166259765625,
0.49365234375,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.509765625,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.509765625,
0.179931640625
],
[
0.11419677734375,
0.179931640625,
0.49365234375,
0.1787109375,
0.49365234375,
0.1995849609375,
0.11419677734375,
0.20068359375
],
[
0.509765625,
0.182861328125,
0.7841796875,
0.182861328125,
0.7841796875,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.1175537109375,
0.20068359375,
0.484375,
0.20068359375,
0.484375,
0.215576171875,
0.1175537109375,
0.215576171875
],
[
0.11248779296875,
0.21435546875,
0.490234375,
0.215576171875,
0.490234375,
0.2369384765625,
0.11248779296875,
0.236328125
],
[
0.51318359375,
0.21142578125,
0.65576171875,
0.21142578125,
0.65576171875,
0.2257080078125,
0.51318359375,
0.2257080078125
],
[
0.51123046875,
0.232177734375,
0.880859375,
0.232177734375,
0.880859375,
0.2452392578125,
0.51123046875,
0.2452392578125
],
[
0.1158447265625,
0.2381591796875,
0.486083984375,
0.2381591796875,
0.486083984375,
0.251708984375,
0.1158447265625,
0.251708984375
],
[
0.1175537109375,
0.254150390625,
0.48779296875,
0.254150390625,
0.48779296875,
0.268310546875,
0.1175537109375,
0.268310546875
],
[
0.51123046875,
0.25048828125,
0.8818359375,
0.25048828125,
0.8818359375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.1158447265625,
0.26953125,
0.48779296875,
0.268310546875,
0.48779296875,
0.283203125,
0.1158447265625,
0.283935546875
],
[
0.509765625,
0.265380859375,
0.8818359375,
0.265380859375,
0.8818359375,
0.279052734375,
0.509765625,
0.279052734375
],
[
0.1158447265625,
0.2861328125,
0.21826171875,
0.2861328125,
0.21826171875,
0.300537109375,
0.1158447265625,
0.300537109375
],
[
0.51123046875,
0.28076171875,
0.8818359375,
0.28076171875,
0.8818359375,
0.29443359375,
0.51123046875,
0.29443359375
],
[
0.51318359375,
0.29638671875,
0.88330078125,
0.29638671875,
0.88330078125,
0.31005859375,
0.51318359375,
0.31005859375
],
[
0.202392578125,
0.308837890625,
0.4013671875,
0.308837890625,
0.4013671875,
0.32373046875,
0.202392578125,
0.32373046875
],
[
0.4609375,
0.308837890625,
0.48779296875,
0.308837890625,
0.48779296875,
0.32177734375,
0.4609375,
0.32177734375
],
[
0.51123046875,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.32666015625,
0.51123046875,
0.32666015625
],
[
0.1158447265625,
0.328857421875,
0.47607421875,
0.328857421875,
0.47607421875,
0.342529296875,
0.1158447265625,
0.342529296875
],
[
0.51123046875,
0.328857421875,
0.8818359375,
0.328857421875,
0.8818359375,
0.342529296875,
0.51123046875,
0.342529296875
],
[
0.1351318359375,
0.3466796875,
0.489501953125,
0.3466796875,
0.489501953125,
0.35986328125,
0.1351318359375,
0.35986328125
],
[
0.51123046875,
0.344482421875,
0.88330078125,
0.344482421875,
0.88330078125,
0.358154296875,
0.51123046875,
0.358154296875
],
[
0.11419677734375,
0.361083984375,
0.486083984375,
0.361083984375,
0.486083984375,
0.374755859375,
0.11419677734375,
0.374755859375
],
[
0.509765625,
0.35986328125,
0.880859375,
0.35986328125,
0.880859375,
0.37353515625,
0.509765625,
0.37353515625
],
[
0.1175537109375,
0.37890625,
0.489501953125,
0.37890625,
0.489501953125,
0.392578125,
0.1175537109375,
0.392578125
],
[
0.51123046875,
0.37890625,
0.880859375,
0.37890625,
0.880859375,
0.3896484375,
0.51123046875,
0.3896484375
],
[
0.1175537109375,
0.394287109375,
0.490234375,
0.394287109375,
0.490234375,
0.407958984375,
0.1175537109375,
0.407958984375
],
[
0.51123046875,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.40673828125,
0.51123046875,
0.40673828125
],
[
0.1158447265625,
0.410400390625,
0.486083984375,
0.410400390625,
0.486083984375,
0.42333984375,
0.1158447265625,
0.42333984375
],
[
0.51318359375,
0.410888671875,
0.60205078125,
0.410888671875,
0.60205078125,
0.4228515625,
0.51318359375,
0.4228515625
],
[
0.1158447265625,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.1158447265625,
0.440673828125
],
[
0.51123046875,
0.435302734375,
0.880859375,
0.435302734375,
0.880859375,
0.448974609375,
0.51123046875,
0.448974609375
],
[
0.1175537109375,
0.4423828125,
0.489501953125,
0.4423828125,
0.489501953125,
0.457275390625,
0.1175537109375,
0.457275390625
],
[
0.51123046875,
0.451904296875,
0.8818359375,
0.451904296875,
0.8818359375,
0.465576171875,
0.51123046875,
0.465576171875
],
[
0.1158447265625,
0.458984375,
0.48779296875,
0.458984375,
0.48779296875,
0.47265625,
0.1158447265625,
0.47265625
],
[
0.51318359375,
0.468505859375,
0.880859375,
0.468505859375,
0.880859375,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.1158447265625,
0.474365234375,
0.274658203125,
0.474365234375,
0.274658203125,
0.488037109375,
0.1158447265625,
0.488037109375
],
[
0.51123046875,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.498779296875,
0.51123046875,
0.498779296875
],
[
0.1318359375,
0.50048828125,
0.483642578125,
0.50048828125,
0.483642578125,
0.515625,
0.1318359375,
0.515625
],
[
0.509765625,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.509765625,
0.51416015625
],
[
0.51123046875,
0.5166015625,
0.8818359375,
0.515625,
0.8818359375,
0.52978515625,
0.51123046875,
0.53076171875
],
[
0.1175537109375,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.1175537109375,
0.54052734375
],
[
0.5146484375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54443359375,
0.5146484375,
0.54443359375
],
[
0.1175537109375,
0.54345703125,
0.48779296875,
0.54345703125,
0.48779296875,
0.55712890625,
0.1175537109375,
0.55712890625
],
[
0.51123046875,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5625,
0.51123046875,
0.5625
],
[
0.1175537109375,
0.55859375,
0.48779296875,
0.55859375,
0.48779296875,
0.57373046875,
0.1175537109375,
0.57373046875
],
[
0.509765625,
0.5634765625,
0.8818359375,
0.56396484375,
0.8818359375,
0.5791015625,
0.509765625,
0.57763671875
],
[
0.1158447265625,
0.576171875,
0.490234375,
0.576171875,
0.490234375,
0.5888671875,
0.1158447265625,
0.5888671875
],
[
0.51123046875,
0.58056640625,
0.8818359375,
0.58056640625,
0.8818359375,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.11419677734375,
0.59033203125,
0.489501953125,
0.59130859375,
0.489501953125,
0.60546875,
0.11419677734375,
0.60498046875
],
[
0.509765625,
0.5966796875,
0.88330078125,
0.5966796875,
0.88330078125,
0.60986328125,
0.509765625,
0.60986328125
],
[
0.1175537109375,
0.60693359375,
0.486083984375,
0.60693359375,
0.486083984375,
0.62060546875,
0.1175537109375,
0.62060546875
],
[
0.509765625,
0.61328125,
0.8818359375,
0.61328125,
0.8818359375,
0.626953125,
0.509765625,
0.626953125
],
[
0.1158447265625,
0.62353515625,
0.490234375,
0.62353515625,
0.490234375,
0.63720703125,
0.1158447265625,
0.63720703125
],
[
0.5087890625,
0.62744140625,
0.880859375,
0.62744140625,
0.880859375,
0.64111328125,
0.5087890625,
0.64111328125
],
[
0.11419677734375,
0.63720703125,
0.490234375,
0.63818359375,
0.490234375,
0.65625,
0.11419677734375,
0.65478515625
],
[
0.509765625,
0.64453125,
0.880859375,
0.64453125,
0.880859375,
0.65771484375,
0.509765625,
0.65771484375
],
[
0.18896484375,
0.65380859375,
0.1956787109375,
0.65380859375,
0.1956787109375,
0.65673828125,
0.18896484375,
0.65673828125
],
[
0.1175537109375,
0.662109375,
0.489501953125,
0.662109375,
0.489501953125,
0.6767578125,
0.1175537109375,
0.6767578125
],
[
0.51123046875,
0.662109375,
0.8818359375,
0.662109375,
0.8818359375,
0.67578125,
0.51123046875,
0.67578125
],
[
0.11419677734375,
0.6767578125,
0.489501953125,
0.677734375,
0.489501953125,
0.6953125,
0.11419677734375,
0.69482421875
],
[
0.51123046875,
0.677734375,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.51123046875,
0.69140625
],
[
0.1175537109375,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.509765625,
0.693359375,
0.8818359375,
0.693359375,
0.8818359375,
0.70703125,
0.509765625,
0.70703125
],
[
0.1158447265625,
0.71044921875,
0.48779296875,
0.71142578125,
0.48779296875,
0.7255859375,
0.1158447265625,
0.724609375
],
[
0.5087890625,
0.708984375,
0.8818359375,
0.7080078125,
0.8818359375,
0.72265625,
0.5087890625,
0.7236328125
],
[
0.1158447265625,
0.72802734375,
0.489501953125,
0.72802734375,
0.489501953125,
0.7412109375,
0.1158447265625,
0.7412109375
],
[
0.5087890625,
0.7255859375,
0.62744140625,
0.724609375,
0.62744140625,
0.7392578125,
0.5087890625,
0.74072265625
],
[
0.1158447265625,
0.74365234375,
0.32568359375,
0.7421875,
0.32568359375,
0.75732421875,
0.1158447265625,
0.75830078125
],
[
0.1351318359375,
0.7587890625,
0.489501953125,
0.75830078125,
0.489501953125,
0.7724609375,
0.1351318359375,
0.77392578125
],
[
0.51123046875,
0.75439453125,
0.60791015625,
0.75439453125,
0.60791015625,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.77392578125,
0.8818359375,
0.77392578125,
0.8818359375,
0.7880859375,
0.509765625,
0.7880859375
],
[
0.1158447265625,
0.79052734375,
0.48779296875,
0.79052734375,
0.48779296875,
0.80419921875,
0.1158447265625,
0.80419921875
],
[
0.509765625,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8046875,
0.509765625,
0.8046875
],
[
0.11419677734375,
0.80810546875,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.11419677734375,
0.82177734375
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.490234375,
0.82470703125,
0.490234375,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.5087890625,
0.822265625,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.5087890625,
0.83740234375
],
[
0.1175537109375,
0.84033203125,
0.49365234375,
0.84033203125,
0.49365234375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.8388671875,
0.880859375,
0.84033203125,
0.880859375,
0.85498046875,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85791015625,
0.489501953125,
0.85791015625,
0.489501953125,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.5087890625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.5087890625,
0.8857421875
],
[
0.1158447265625,
0.88720703125,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.509765625,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"Mathematical Formulation.We consider a",
"method to optimize the tweet and word selection",
"multimodal stock forecast model f() that takes",
"variables and word replacement variables simulta",
"tweet collections [ct]t-1 and numerical factors",
"neously. AGO uses an alternative optimization pro",
"[pt]t- as input, where t indexes the date when",
"cedure to sequentially update the discrete selection",
"the data is collected. Peeking into the tweet col-",
"variablesandthereplacement selection variables",
"lection, it contains ct tweets for date t, namely,",
"More details on the optimization program and the",
"Ct - {st,s?,..,.slectly. Each tweet s is a text-",
" solvers can be found in Appendix A.",
"based sentence of length Istl, denoted as st -",
"w....,..stl).foi ,.. Adi",
"3Experiments",
"Dataset & victim models.We evaluate our ad-",
"rectional financial forecast model takes domains",
"of tweets and numerical factors as input, and",
"versarial attackon a stockprediction datasetcon-",
"yields prediction for stocks' directional movement",
"sisting of 10,824 instances including relevant",
"ye(-1,1]:",
"tweets and numerical features of 88 stocks from",
"2014 to 2016 (Xu and Cohen, 2018). Three mod-",
"9t+1 = f(ct-h:t,Pt-h:t),",
"(1)",
"els (stocknet (Xu and Cohen, 2018),FinGRu",
"h is the looking-back window for historical data.",
"based on GRU (Cho et al., 2014) and FinLSTM",
"The hierarchical perturbation can be cast as a",
"based on LSTM (Hochreiter and Schmidhuber,",
"combinatorial problem for tweet selection m, word",
"1997)) of binary classification are considered as",
"selection z and replacement selection u. The",
"victims inthis paper.We apply our attack to in-",
"boolean vector m indicates the tweets to be se-",
"stances on which the victim models make correct",
"lected. For i-th tweet, vector z, indicates the word",
"predictions.",
"to be perturbed. As for the word perturbation task,",
"Evaluation metrics.Attack performance is eval-",
"another boolean vector ui,j selects the best replace-",
"uated by two metrics: Attack Success Rate (ASR)",
"ment. It follows that the hierarchical perturbation",
"and victim model's F1 drop after attack. ASR",
"can be formulated as",
"is defined as the percentage of the attack efforts",
"ct =(1 -m·z).ct +m·z.u S(ct), (2)",
"that changes the model output. The two metrics",
"gauge the efficacy of the attack and its impact on",
"where denotes element-column wise product,",
"modelperformance:Moreefficientattackleadsto",
"m z indicates the selected words in selected",
"higher ASR and more decline of Fl. Moreover.",
"tweets, m . z · u indicates selected synonyms",
"we simulate a Long-Only Buy-Hold-Sell strategy",
"for each selected word, and S() is a element-",
"(Sawhney et al., 2021; Feng et al., 2019) with vic-",
"wise synonym generating function. Consequently,",
"tim models, and calculate the Profit and Loss (PnL)",
"given attack loss L, generation of adversarial quote",
"for each simulation. Assume a portfolio starts with",
"tweets can be formulated as the optimization pro-",
"initial net value $10000 (100%), its net value at the",
"gram min C(cf U ct-h:t, Ct-h:tlpt-h:t, f), sub-",
"end of test period reflects the profitability of the",
"",
"ject to the budget constraints: a) 1Tm ≤ bs, b)",
"trading strategy and the underlying model. Conse-",
"iTzi ≤ bu, Vi and c) Tui,j = 1, Vi,j, where",
"quently, the change in PnLs measures the monetary",
"bs and bw denote the tweet and word budgets. It",
"impact of our attack. More details on the dataset",
"is worth to stress that perturbation is only applied",
"victim models and evaluation metrics are housed",
"to the date (t) when the attack is implemented to",
"in Appendix B.",
"preserve the temporal order.",
"To solve the program, we follow the convex",
"4Results",
"relaxation approach developed in (Srikant et al.,",
"Attack performance with single perturbation.",
"2021). Specifically, the boolean variables (for tweet",
"The experiment results for the concatenation at-",
"and word selection) are relaxed into the continuous",
"tack with word replacement perturbation is shown",
"space so that they can be optimized by gradient-",
"in Table 1 (with tweet and word budgets both as 1).",
"based methods over a convex hull. Two main imple-",
"For both JO and AGO, ASR increases by roughly",
"mentations ofthe optimization-basedattackgen-",
"10% and F1 drops by 0.1 on average in compari",
"eration method are proposed: joint optimization",
" son to the random attack. Such performance drop",
"(O) solver and alternating greedy optimization",
"is considered significant in the context of stock",
"(AGO) solver. JO calls projected gradient descent ",
" prediction given that the state-of-the-art prediction"
] |
{
"arxiv_id": "2205.01094",
"language": "en",
"page_idx": 2
}
|
||
2205.01094-en-3
|
2205.01094-en-3.jpg
|
accuracy of interday return is only about 60%.
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{Model}$</th>
<th colspan="4">$\mathbf{ASR}(\%)$</th>
<th colspan="4">$F1$</th>
</tr>
<tr>
<th>$NA$</th>
<th>$RA$</th>
<th>$JO$</th>
<th>$AGO$</th>
<th>$NA$</th>
<th>$RA$</th>
<th>JO</th>
<th>$AGO$</th>
</tr>
<tr>
<td>Stocknet</td>
<td>0</td>
<td>$\overline{4.5}$</td>
<td>16.8</td>
<td>11.8</td>
<td>1</td>
<td>0.96</td>
<td>0.84</td>
<td>0.88</td>
</tr>
<tr>
<td>FinGRU</td>
<td>0</td>
<td>5.1</td>
<td>16.4</td>
<td>14.1</td>
<td>1</td>
<td>0.95</td>
<td>0.85</td>
<td>0.87</td>
</tr>
<tr>
<td>FinLSTM</td>
<td>0</td>
<td>11.9</td>
<td>16.5</td>
<td>19.7</td>
<td>1</td>
<td>0.89</td>
<td>0.85</td>
<td>0.78</td>
</tr>
</tbody>
</table>
Table 1: Performance of the various adversarial attacks. NA: no attack; RA: random attack; JO: joint optimization; and AGO: alternating greedy optimization.
Effect of attack budget. We report the effect of different attack budgets on the attack performance in Fig. 2. We observe that the more budgets allowed (perturbing more tweets and words), the better the attack performance, but the increase is not significant. It appears that the attack performance becomes saturated if we keep increasing the attack budgets. In fact, the attack with budget of one tweet and one word is the most cost effective, provided that it introduces minimum perturbation but achieves a relatively similar ASR.
<FigureHere>
Figure 2: Effect of attack budgets on ASR with Stocknet as victim model and with JO solver. r-perturb: word replacement; d-perturb: word deletion.
$Wefo-$
Manipulation vs concatenation attack.
cus on concatenation attack in this paper since we believe it is distinct from manipulation attack. We investigate the difference by applying the same method of tweet generation to implement manipulation attack, where the adversarial tweets replace target tweets instead. The experiment runs with one word budget and one twee budget, and the results are reported in Fig. 3.
<FigureHere>
Figure 3: Comparison between manipulation and concatenation attacks with word-replacement perturbation method. Stocknet is the victim model.
It is clear that manipulation attack remarkably outperforms concatenation attack in terms of ASR
and F1. Even though the success rate of concatenation attack lags behind the state-of-the-art textual attack, the manipulation attack achieves performance of the same ballpark, which demonstrates the efficacy of optimization-based attack and our solvers. More importantly, it implies that the attack is not transferable between the two tasks, documenting more evidence on language attack transferability (Yuan et al., 2021; He et al., 2021). The bottom line is that they are two different tasks under different assumptions. Researchers should take downstream scenarios into account when develop attack models.
Trading simulation. The ultimate measure of a stock prediction model's performance is profitability. Fig.4 plots the profit and loss of the same trading strategy with Stocknet as the prediction model with or without the attack-JO is used to generate adversarial quote tweets. For each simulation, the investor has $10K (100%) to invest; the results show that the proposed attack method with a quote tweet with only a single word replacement can cause the investor an additional $3.2K (75%- $43\%)$ loss to their portfolio after about 2 years.
<FigureHere>
Figure 4: Profit and Loss with Stocknet as the victim model using Long-Only Buy-Hold-Sell strategy for 2 years with $10K initial investment. Green line: trading using Stocknet without attack; Blue line: concatenation attack with deletion perturbation; Red line: concatenation attack with replacement perturbation.
# 5 Conclusion
This work demonstrates that our adversarial attack method consistently fools various financial forecast models even with physical constraints that the raw tweet can not be modified. Adding a quote tweet with only one word replaced, the attack can cause 32% additional loss to our simulated investment portfolio. Via studying financial model's vulnerability, our goal is to raise financial community's awareness of the AI model's risks, so that in the future we can develop more robust human-in-theloop AI architecture (Wang et al., 2019) to cope with this and other real-world attacks, including black-box attack, unknown input domains, etc.
|
[
[
0.1175537109375,
0.08612060546875,
0.45751953125,
0.08612060546875,
0.45751953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.51318359375,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.51318359375,
0.114013671875
],
[
0.2359619140625,
0.11224365234375,
0.285400390625,
0.114013671875,
0.28369140625,
0.12890625,
0.235107421875,
0.1270751953125
],
[
0.391357421875,
0.11517333984375,
0.411376953125,
0.11517333984375,
0.411376953125,
0.127685546875,
0.391357421875,
0.127685546875
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.2283935546875,
0.127685546875,
0.2493896484375,
0.127685546875,
0.2493896484375,
0.14013671875,
0.2283935546875,
0.14013671875
],
[
0.14111328125,
0.12054443359375,
0.17626953125,
0.12054443359375,
0.17626953125,
0.1318359375,
0.14111328125,
0.1318359375
],
[
0.297119140625,
0.127685546875,
0.328369140625,
0.127685546875,
0.328369140625,
0.1395263671875,
0.297119140625,
0.1395263671875
],
[
0.43994140625,
0.127685546875,
0.470947265625,
0.127685546875,
0.470947265625,
0.1395263671875,
0.43994140625,
0.1395263671875
],
[
0.26611328125,
0.12890625,
0.2822265625,
0.12890625,
0.2822265625,
0.1383056640625,
0.26611328125,
0.1383056640625
],
[
0.193115234375,
0.12890625,
0.21240234375,
0.12890625,
0.21240234375,
0.1383056640625,
0.193115234375,
0.1383056640625
],
[
0.334228515625,
0.1300048828125,
0.355224609375,
0.1300048828125,
0.355224609375,
0.1383056640625,
0.334228515625,
0.1383056640625
],
[
0.372802734375,
0.1300048828125,
0.391357421875,
0.1300048828125,
0.391357421875,
0.1383056640625,
0.372802734375,
0.1383056640625
],
[
0.41064453125,
0.12890625,
0.4248046875,
0.12890625,
0.4248046875,
0.1383056640625,
0.41064453125,
0.1383056640625
],
[
0.1318359375,
0.1395263671875,
0.17626953125,
0.1395263671875,
0.17626953125,
0.149658203125,
0.1318359375,
0.149658203125
],
[
0.2626953125,
0.1395263671875,
0.2880859375,
0.1395263671875,
0.2880859375,
0.1519775390625,
0.2626953125,
0.1519775390625
],
[
0.405517578125,
0.1395263671875,
0.4306640625,
0.1395263671875,
0.4306640625,
0.15087890625,
0.405517578125,
0.15087890625
],
[
0.5146484375,
0.135986328125,
0.880859375,
0.135986328125,
0.880859375,
0.146728515625,
0.5146484375,
0.146728515625
],
[
0.298828125,
0.14013671875,
0.32421875,
0.14013671875,
0.32421875,
0.1519775390625,
0.298828125,
0.1519775390625
],
[
0.2301025390625,
0.14013671875,
0.2476806640625,
0.14013671875,
0.2476806640625,
0.149658203125,
0.2301025390625,
0.149658203125
],
[
0.3720703125,
0.141357421875,
0.3955078125,
0.141357421875,
0.3955078125,
0.149658203125,
0.3720703125,
0.149658203125
],
[
0.441650390625,
0.14013671875,
0.468505859375,
0.14013671875,
0.468505859375,
0.1484375,
0.441650390625,
0.1484375
],
[
0.133544921875,
0.1484375,
0.177978515625,
0.1484375,
0.177978515625,
0.1591796875,
0.133544921875,
0.1591796875
],
[
0.2301025390625,
0.1484375,
0.2476806640625,
0.1484375,
0.2476806640625,
0.157958984375,
0.2301025390625,
0.157958984375
],
[
0.264404296875,
0.149658203125,
0.287109375,
0.149658203125,
0.287109375,
0.1602783203125,
0.264404296875,
0.1602783203125
],
[
0.300537109375,
0.149658203125,
0.322509765625,
0.149658203125,
0.322509765625,
0.157958984375,
0.300537109375,
0.157958984375
],
[
0.3720703125,
0.149658203125,
0.3955078125,
0.149658203125,
0.3955078125,
0.157958984375,
0.3720703125,
0.157958984375
],
[
0.4072265625,
0.149658203125,
0.42822265625,
0.149658203125,
0.42822265625,
0.157958984375,
0.4072265625,
0.157958984375
],
[
0.441650390625,
0.149658203125,
0.465087890625,
0.149658203125,
0.465087890625,
0.157958984375,
0.441650390625,
0.157958984375
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.162109375,
0.51318359375,
0.162109375
],
[
0.370361328125,
0.157958984375,
0.3955078125,
0.157958984375,
0.3955078125,
0.169189453125,
0.370361328125,
0.169189453125
],
[
0.20068359375,
0.15087890625,
0.208251953125,
0.15087890625,
0.208251953125,
0.1561279296875,
0.20068359375,
0.1561279296875
],
[
0.2301025390625,
0.1591796875,
0.2509765625,
0.1591796875,
0.2509765625,
0.16748046875,
0.2301025390625,
0.16748046875
],
[
0.301513671875,
0.1591796875,
0.322509765625,
0.1591796875,
0.322509765625,
0.16748046875,
0.301513671875,
0.16748046875
],
[
0.4072265625,
0.1591796875,
0.42822265625,
0.1591796875,
0.42822265625,
0.16748046875,
0.4072265625,
0.16748046875
],
[
0.202392578125,
0.1602783203125,
0.20654296875,
0.1602783203125,
0.20654296875,
0.1650390625,
0.202392578125,
0.1650390625
],
[
0.1351318359375,
0.1602783203125,
0.1822509765625,
0.1602783203125,
0.1822509765625,
0.16748046875,
0.1351318359375,
0.16748046875
],
[
0.264404296875,
0.1602783203125,
0.285400390625,
0.1602783203125,
0.285400390625,
0.16748046875,
0.264404296875,
0.16748046875
],
[
0.443359375,
0.1602783203125,
0.465087890625,
0.1602783203125,
0.465087890625,
0.168701171875,
0.443359375,
0.168701171875
],
[
0.51318359375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1158447265625,
0.182861328125,
0.489501953125,
0.182861328125,
0.489501953125,
0.196533203125,
0.1158447265625,
0.196533203125
],
[
0.509765625,
0.1817626953125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.509765625,
0.1953125
],
[
0.11419677734375,
0.1953125,
0.490234375,
0.196533203125,
0.490234375,
0.21142578125,
0.11419677734375,
0.210205078125
],
[
0.51123046875,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.1158447265625,
0.2120361328125,
0.441650390625,
0.2120361328125,
0.441650390625,
0.2257080078125,
0.1158447265625,
0.2257080078125
],
[
0.51123046875,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.118408203125,
0.235107421875,
0.48779296875,
0.235107421875,
0.48779296875,
0.248779296875,
0.118408203125,
0.248779296875
],
[
0.51123046875,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.51123046875,
0.24462890625
],
[
0.118408203125,
0.2529296875,
0.486083984375,
0.2529296875,
0.486083984375,
0.26318359375,
0.118408203125,
0.26318359375
],
[
0.5087890625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.5087890625,
0.26123046875
],
[
0.1158447265625,
0.267333984375,
0.489501953125,
0.267333984375,
0.489501953125,
0.28076171875,
0.1158447265625,
0.28076171875
],
[
0.51123046875,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.118408203125,
0.28515625,
0.489501953125,
0.28515625,
0.489501953125,
0.295654296875,
0.118408203125,
0.295654296875
],
[
0.51318359375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.118408203125,
0.300537109375,
0.486083984375,
0.300537109375,
0.486083984375,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51123046875,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51123046875,
0.31640625
],
[
0.1175537109375,
0.31640625,
0.48779296875,
0.31640625,
0.48779296875,
0.330078125,
0.1175537109375,
0.330078125
],
[
0.509765625,
0.319580078125,
0.8818359375,
0.319580078125,
0.8818359375,
0.333251953125,
0.509765625,
0.333251953125
],
[
0.1175537109375,
0.33203125,
0.486083984375,
0.33203125,
0.486083984375,
0.345703125,
0.1175537109375,
0.345703125
],
[
0.509765625,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.509765625,
0.3486328125
],
[
0.1175537109375,
0.349853515625,
0.486083984375,
0.349853515625,
0.486083984375,
0.35986328125,
0.1175537109375,
0.35986328125
],
[
0.509765625,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.509765625,
0.364013671875
],
[
0.11419677734375,
0.363525390625,
0.486083984375,
0.363525390625,
0.486083984375,
0.377197265625,
0.11419677734375,
0.377197265625
],
[
0.51318359375,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.1158447265625,
0.38134765625,
0.489501953125,
0.38134765625,
0.489501953125,
0.394287109375,
0.1158447265625,
0.394287109375
],
[
0.51318359375,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.3955078125,
0.51318359375,
0.3955078125
],
[
0.1158447265625,
0.3955078125,
0.365234375,
0.3955078125,
0.365234375,
0.4091796875,
0.1158447265625,
0.4091796875
],
[
0.51123046875,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.41455078125,
0.51123046875,
0.41455078125
],
[
0.139404296875,
0.418701171875,
0.1485595703125,
0.419189453125,
0.14697265625,
0.42333984375,
0.1376953125,
0.4228515625
],
[
0.51123046875,
0.416259765625,
0.8818359375,
0.416259765625,
0.8818359375,
0.429931640625,
0.51123046875,
0.429931640625
],
[
0.1351318359375,
0.431640625,
0.151123046875,
0.431640625,
0.151123046875,
0.441162109375,
0.1351318359375,
0.441162109375
],
[
0.51123046875,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.20654296875,
0.44482421875,
0.214111328125,
0.44482421875,
0.214111328125,
0.44775390625,
0.20654296875,
0.44775390625
],
[
0.214111328125,
0.44482421875,
0.231689453125,
0.44482421875,
0.231689453125,
0.44775390625,
0.214111328125,
0.44775390625
],
[
0.385498046875,
0.44482421875,
0.41650390625,
0.44482421875,
0.41650390625,
0.44775390625,
0.385498046875,
0.44775390625
],
[
0.438232421875,
0.443603515625,
0.45751953125,
0.443603515625,
0.45751953125,
0.44775390625,
0.438232421875,
0.44775390625
],
[
0.133544921875,
0.45068359375,
0.15283203125,
0.45068359375,
0.15283203125,
0.457763671875,
0.133544921875,
0.457763671875
],
[
0.509765625,
0.44775390625,
0.85400390625,
0.44775390625,
0.85400390625,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.1275634765625,
0.457275390625,
0.133544921875,
0.457275390625,
0.133544921875,
0.466064453125,
0.1275634765625,
0.466064453125
],
[
0.5498046875,
0.466064453125,
0.57421875,
0.466064453125,
0.57421875,
0.478515625,
0.5498046875,
0.478515625
],
[
0.5341796875,
0.4755859375,
0.580078125,
0.4912109375,
0.568359375,
0.50830078125,
0.5224609375,
0.4921875
],
[
0.52880859375,
0.4912109375,
0.55517578125,
0.4912109375,
0.5517578125,
0.5693359375,
0.5263671875,
0.5693359375
],
[
0.5498046875,
0.49951171875,
0.57421875,
0.49951171875,
0.57421875,
0.51318359375,
0.5498046875,
0.51318359375
],
[
0.20068359375,
0.51416015625,
0.2568359375,
0.51416015625,
0.2568359375,
0.521484375,
0.20068359375,
0.521484375
],
[
0.3779296875,
0.5126953125,
0.43994140625,
0.5126953125,
0.43994140625,
0.5224609375,
0.3779296875,
0.5224609375
],
[
0.8330078125,
0.5185546875,
0.85546875,
0.5185546875,
0.85546875,
0.52685546875,
0.8330078125,
0.52685546875
],
[
0.1175537109375,
0.52685546875,
0.486083984375,
0.52685546875,
0.486083984375,
0.54052734375,
0.1175537109375,
0.54052734375
],
[
0.5556640625,
0.53515625,
0.57080078125,
0.53515625,
0.57080078125,
0.54638671875,
0.5556640625,
0.54638671875
],
[
0.615234375,
0.5361328125,
0.6630859375,
0.5361328125,
0.6630859375,
0.54638671875,
0.615234375,
0.54638671875
],
[
0.82275390625,
0.5361328125,
0.85791015625,
0.5361328125,
0.85791015625,
0.54736328125,
0.82275390625,
0.54736328125
],
[
0.1158447265625,
0.5419921875,
0.486083984375,
0.5419921875,
0.486083984375,
0.5556640625,
0.1158447265625,
0.5556640625
],
[
0.6171875,
0.54736328125,
0.7001953125,
0.54736328125,
0.7001953125,
0.55810546875,
0.6171875,
0.55810546875
],
[
0.11419677734375,
0.5556640625,
0.378662109375,
0.5546875,
0.378662109375,
0.5693359375,
0.11419677734375,
0.57080078125
],
[
0.615234375,
0.55859375,
0.72314453125,
0.55810546875,
0.72314453125,
0.5693359375,
0.615234375,
0.57080078125
],
[
0.82373046875,
0.55419921875,
0.85791015625,
0.55419921875,
0.85791015625,
0.5654296875,
0.82373046875,
0.5654296875
],
[
0.5556640625,
0.57080078125,
0.57275390625,
0.57080078125,
0.57275390625,
0.58056640625,
0.5556640625,
0.58056640625
],
[
0.63232421875,
0.576171875,
0.65576171875,
0.576171875,
0.65576171875,
0.58837890625,
0.63232421875,
0.58837890625
],
[
0.6962890625,
0.576171875,
0.724609375,
0.576171875,
0.724609375,
0.58837890625,
0.6962890625,
0.58837890625
],
[
0.765625,
0.576171875,
0.791015625,
0.576171875,
0.791015625,
0.58837890625,
0.765625,
0.58837890625
],
[
0.8134765625,
0.576171875,
0.8388671875,
0.576171875,
0.8388671875,
0.58740234375,
0.8134765625,
0.58740234375
],
[
0.57421875,
0.580078125,
0.57861328125,
0.580078125,
0.57861328125,
0.5849609375,
0.57421875,
0.5849609375
],
[
0.1175537109375,
0.580078125,
0.438232421875,
0.580078125,
0.438232421875,
0.59326171875,
0.1175537109375,
0.59326171875
],
[
0.432373046875,
0.58056640625,
0.48779296875,
0.58056640625,
0.48779296875,
0.5927734375,
0.432373046875,
0.5927734375
],
[
0.68408203125,
0.58740234375,
0.7197265625,
0.58740234375,
0.7197265625,
0.5986328125,
0.68408203125,
0.5986328125
],
[
0.1158447265625,
0.5966796875,
0.48779296875,
0.5966796875,
0.48779296875,
0.61083984375,
0.1158447265625,
0.61083984375
],
[
0.51318359375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.61083984375,
0.51318359375,
0.61083984375
],
[
0.1175537109375,
0.6123046875,
0.48779296875,
0.6123046875,
0.48779296875,
0.6259765625,
0.1175537109375,
0.6259765625
],
[
0.509765625,
0.61328125,
0.88330078125,
0.61328125,
0.88330078125,
0.626953125,
0.509765625,
0.626953125
],
[
0.1175537109375,
0.62890625,
0.489501953125,
0.62890625,
0.489501953125,
0.642578125,
0.1175537109375,
0.642578125
],
[
0.51123046875,
0.62744140625,
0.88330078125,
0.62744140625,
0.88330078125,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.1175537109375,
0.64453125,
0.490234375,
0.64453125,
0.490234375,
0.65771484375,
0.1175537109375,
0.65771484375
],
[
0.5087890625,
0.64111328125,
0.8818359375,
0.64013671875,
0.8818359375,
0.65478515625,
0.5087890625,
0.65625
],
[
0.1175537109375,
0.6611328125,
0.484375,
0.6611328125,
0.484375,
0.67138671875,
0.1175537109375,
0.67138671875
],
[
0.509765625,
0.65625,
0.884765625,
0.65625,
0.884765625,
0.66943359375,
0.509765625,
0.66943359375
],
[
0.51123046875,
0.67041015625,
0.87890625,
0.67041015625,
0.87890625,
0.68408203125,
0.51123046875,
0.68408203125
],
[
0.1175537109375,
0.6787109375,
0.486083984375,
0.6787109375,
0.486083984375,
0.689453125,
0.1175537109375,
0.689453125
],
[
0.1158447265625,
0.693359375,
0.48779296875,
0.693359375,
0.48779296875,
0.70703125,
0.1158447265625,
0.70703125
],
[
0.515625,
0.6953125,
0.53662109375,
0.6953125,
0.53662109375,
0.703125,
0.515625,
0.703125
],
[
0.5400390625,
0.693359375,
0.6396484375,
0.693359375,
0.6396484375,
0.705078125,
0.5400390625,
0.705078125
],
[
0.11419677734375,
0.708984375,
0.279541015625,
0.7080078125,
0.279541015625,
0.72265625,
0.11419677734375,
0.7236328125
],
[
0.509765625,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.7236328125,
0.509765625,
0.7236328125
],
[
0.1368408203125,
0.728515625,
0.151123046875,
0.728515625,
0.151123046875,
0.73828125,
0.1368408203125,
0.73828125
],
[
0.397216796875,
0.73291015625,
0.480224609375,
0.73291015625,
0.480224609375,
0.751953125,
0.397216796875,
0.751953125
],
[
0.51123046875,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.133544921875,
0.74072265625,
0.154541015625,
0.74072265625,
0.154541015625,
0.7548828125,
0.133544921875,
0.7548828125
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.32080078125,
0.75048828125,
0.32666015625,
0.75048828125,
0.32666015625,
0.75732421875,
0.32080078125,
0.75732421875
],
[
0.1318359375,
0.75732421875,
0.154541015625,
0.75732421875,
0.154541015625,
0.76953125,
0.1318359375,
0.76953125
],
[
0.51123046875,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.31982421875,
0.76708984375,
0.330078125,
0.76708984375,
0.330078125,
0.77490234375,
0.31982421875,
0.77490234375
],
[
0.124267578125,
0.76123046875,
0.15869140625,
0.775390625,
0.145263671875,
0.7919921875,
0.11083984375,
0.77783203125
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1351318359375,
0.79150390625,
0.14697265625,
0.7880859375,
0.1502685546875,
0.79638671875,
0.1376953125,
0.7998046875
],
[
0.51318359375,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51318359375,
0.8046875
],
[
0.139404296875,
0.80712890625,
0.151123046875,
0.80712890625,
0.151123046875,
0.81103515625,
0.139404296875,
0.81103515625
],
[
0.51123046875,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.20068359375,
0.82373046875,
0.255126953125,
0.82373046875,
0.255126953125,
0.83447265625,
0.20068359375,
0.83447265625
],
[
0.378662109375,
0.822265625,
0.435791015625,
0.822265625,
0.435791015625,
0.83447265625,
0.378662109375,
0.83447265625
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1175537109375,
0.8388671875,
0.490234375,
0.8388671875,
0.490234375,
0.8525390625,
0.1175537109375,
0.8525390625
],
[
0.51123046875,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.1175537109375,
0.85400390625,
0.49365234375,
0.85400390625,
0.49365234375,
0.86767578125,
0.1175537109375,
0.86767578125
],
[
0.51318359375,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.8681640625,
0.51318359375,
0.8681640625
],
[
0.1175537109375,
0.8681640625,
0.426513671875,
0.8681640625,
0.426513671875,
0.8818359375,
0.1175537109375,
0.8818359375
],
[
0.51123046875,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.133544921875,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.133544921875,
0.90185546875
],
[
0.509765625,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.11419677734375,
0.90380859375,
0.48779296875,
0.90283203125,
0.48779296875,
0.91748046875,
0.11419677734375,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.85546875,
0.90478515625,
0.85546875,
0.91845703125,
0.5087890625,
0.91845703125
]
] |
[
"accuracy of interday return is only about 60%.",
"and F1. Even though the success rate of concatena-",
"tionattack lags behind thestate-of-the-art textual at-",
"ASR(%)",
"F1",
"tack, the manipulation attack achieves performance",
"RA",
"Model",
"AGO",
"AGO",
"JO",
"NA",
"NA",
"RA",
"JO",
"Stocknet",
"16.8",
"0.84",
"ofthesameballpark,whichdemonstratestheeffi-",
"11.8",
"4.5",
"0.96",
"0.88",
"FinGRU",
"5.1",
"16.4",
"14.1",
"0.95",
"0.85",
"0.87",
"cacy of optimization-based attack andour solvers",
"0.89",
"0",
"11.9",
"19.7",
"0.85",
"",
"FinLSTM",
"16.5",
"0.78",
"More importantly, it implies that the attack is not",
"Table I: Performance of the various adversarial attacks.",
"transferable between the two tasks, documenting",
"NA: no attack; RA: random attack; JO: joint optimiza-",
"more evidence on language attack transferability",
"tion; and AGO: alternating greedy optimization.",
"(Yuan et al., 2021; He et al., 2021). The bottom line",
"Effect of attack budget.We report the effect of",
"is that they are two different tasks under different",
"differentattackbudgetsontheattackperformance",
" assumptions. Researchers should take downstream",
"in Fig. 2. We observe that the more budgets al-",
"scenarios into account when develop attack models.",
"lowed (perturbing moretweetsand words),thebet-",
"Trading simulation.The ultimate measure of a",
"tertheattackperformance.butthe increase isnot",
"stock prediction model's performance is profitabil-",
"significant. It appears that the attack performance",
"ity. Fig. 4 plots the profit and loss of the same",
"becomes saturated if we keep increasing the attack",
"trading strategy with stocknet as the prediction",
"budgets.In fact. the attack with budget of one",
"model with or without the attack - JO is used to",
" tweet and one word is the most cost effective, pro",
"generate adversarial quote tweets. For each simu-",
"vided that it introduces minimum perturbation but",
"lation,theinvestorhas $10K(100%)toinvest:the",
"achieves a relatively similar ASR",
"results show that the proposed attack method with",
"0.5",
"a quote tweet with only a single word replacement",
"0.4",
"can cause the investor an additional $3.2K (75%-",
"ve",
"R",
"wnreir",
"D-ne",
"0.3",
"43%) loss to their portfolio after about 2 years",
"HS",
"12K",
"①10K",
"soeeo",
"9K",
"wordbudget",
"tweet budget",
"7.5K",
"Figure 2: Effect of attack budgets on ASR with Stock",
"6K",
"stocknet",
"5.5K",
"net as victim model and with JO solver. r-perturb: word",
"deletion-perturb",
"replacement; d-perturb: word deletion.",
"replacement-perturb",
"4.3K",
"3K",
"100",
"200",
"300",
"372",
"-",
"Manipulation vs concatenation attack.",
"Wefo-",
"days",
"cus on concatenation attack in this paper since we",
"Figure4:Profit and LosswithStocknet as the vic-",
"believe it is distinct from manipulation attack. We",
"tim model using Long-Only Buy-Hold-Sell strategy for",
"investigate the difference by applying the same",
"2 years with $10K initial investment. Green line: trad-",
"method of tweet generation to implement manipu-",
"ing using Stocknet without attack; Blue line: con-",
"lationattack.wheretheadversarial tweetsreplace",
"catenation attack with deletion perturbation; Red line:",
"concatenation attack with replacement perturbation.",
"targettweetsinstead.The experimentruns with one",
"word budget and one twee budget, and the results",
"5",
"Conclusion",
"are reported in Fig.3.",
"This work demonstrates that our adversarial attack",
"1.0",
"± mngianton",
"method consistently fools various financial forecast",
"0.8",
"models even with physical constraints that the raw",
"",
"0.6",
"tweet can not be modified. Adding a quote tweet",
"0.6",
"0A",
"with only one word replaced, the attack can cause",
"0.2",
"32% additional loss to our simulated investment",
".0",
"portfolio. Via studying financial model's vulnera-",
"wordbudge",
"word budge",
"bility, our goal is to raise financial community's",
"Figure 3: Comparison between manipulation and con-",
"awareness of the AI model's risks, so that in the",
"catenation attacks with word-replacement perturba-",
"future we candevelop more robusthuman-in-the-",
"tion method. St ocknet is the victim model.",
"loop AI architecture (Wang et al., 2019) to cope",
"It is clear that manipulation attack remarkably",
"with this and other real-world attacks, including",
"outperforms concatenation attack in terms of ASR",
" black-box attack, unknown input domains, etc."
] |
{
"arxiv_id": "2205.01094",
"language": "en",
"page_idx": 3
}
|
||
2404.12618-en-0
|
2404.12618-en-0.jpg
|
# CORI: CJKV Benchmark with Romanization Integration- A step towards Cross-lingual Transfer Beyond Textual Scripts
Hoang H. Nguyen$^1$, Chenwei Zhang$^2$,Ye Liu$^3$,
Natalie Parde$^1$,Eugene Rohrbaugh$^4$,Philip S. Yu$^1$
$^{1}$ Department of Computer Science, University of llinois at Chicago, Chicago, IL, USA
$^2$ Amazon, Seattle, WA, USA $^3$ Salesforce Research, Palo Alto, CA, USA $^{4}$ Harrisburg University of Science and Technology, Harrisburg, PA, USA
$\{$hnguy7,parde, psyu}Quic.edu, cwzhang@amazon.com yeliu@salesforce.com, gene.rohrbaugh@gmail.com
### Abstract
Naively assuming English as a source language may hinder cross-lingual transfer for many languages by failing to consider the importance of language contact. Some languages are more well-connected than others, and target languages can benefit from transferring from closely related languages; for many languages, the set of closely related languages does not include English. In this work, we study the impact of source language for cross-lingual transfer, demonstrating the importance of selecting source languages that have high contact with the target language. We also construct a novel benchmark dataset for close contact Chinese-Japanese-Korean-Vietnamese (CJKV) languages to further encourage in-depth studies of language contact. To comprehensively capture contact between these languages, we propose to integrate Romanized transcription beyond textual scripts via Contrastive Learning objectives, leading to enhanced cross-lingual representations and effective zero-shot cross-lingual transfer.
Keywords: Cross-lingual Transfer, Low-resource Languages, Orthographic and Phonemic Transcription
# 1. Introduction
language contact (Nguyen and Rohrbaugh, 2019; Yang et al., 2022). As language acquisition naturally benefits from various linguistic modalities, including the textual writing scripts and aural signals, which can be conveyed through orthographic and phonemic transcription respectively (Volterra and Iverson; Goswami, 2022), the aforementioned constraint hinders progress in the cross-lingual field, by minimizing the amount of relevant information that models may be able to access (Fujinuma et al., 2022; Nguyen et al., 2023c).
Because of language contact, there are many lexical items shared among CJKV languages. However, these similarities may be obscured by the different writing systems used across the languages. Borrowed words that may sound very similar in spoken form may nevertheless be represented in the writing systems with completely different symbols. As observed in Table 1, the orthographic term 古典 in ZH shares little similarity with Vi counterpart. However, the corresponding phonemic transcription represented via Romanization (gǔdiǎn) is more similar to the corresponding VI (Cô điên). The same phenomenon can also be observed in JA-KO pair (古典 (koten) vs 고전(gojeon)) and VI-KO pair (Khoa hoc (Khoa hoc) vs 과학(gwahak)). Moreover, Table 1 reveals the shared properties of CJKV languages in which meaningful words are formed by combining surrounding tokens. For instance, four characters 形而上学 are combined together to form the meaningful word " Metaphysical". The same behavior can be observed in the corresponding VI (Siêu
Language is fluid and ever-evolving, with specific languages often borrowing terms and even grammatical structure from one another as a byproduct of voluntary or forced interaction. Language contact defines the emerging connections among languages throughout the course of history, geography, and social development (Matras, 2020; Mufwene and Escobar, 2022). For example, modern English (EN) is a Germanic language, yet it contains many Latinate and Greek terms that entered the vernacular via religious missionaries or war and conquest, as well as other terms that later entered the lexicon via colonization (van Gelderen, 2014). Likewise, Japanese (JA), Korean (KO) and Vietnamese (VI) share a large number of cognates with Chinese (ZH) due to influence through colonization, trading throughout history (Li, 2019; Hickey, 2020). This influence is best reflected through the Kanji characters in JA writing system, which significantly overlaps with ZH character writing systems. As observed in Table 1, both ZH and JA writing systems utilize 文学 and 学者 to denote words scholar and classical respectively.
Despite recent advances in the study of crosslingual transfer, most cross-lingual works to date have restricted themselves to two major assumptions: (1) English (EN) is considered the single source language, despite its oftentimes limited or absent contact with the target language(s) (Muller et al., 2021; Fujinuma et al., 2022); and (2) crosslingual transfer remains restricted to orthographic
|
[
[
0.139404296875,
0.09442138671875,
0.8623046875,
0.0968017578125,
0.8623046875,
0.114013671875,
0.139404296875,
0.11224365234375
],
[
0.18896484375,
0.119384765625,
0.8134765625,
0.119384765625,
0.8134765625,
0.1341552734375,
0.18896484375,
0.1341552734375
],
[
0.285400390625,
0.157958984375,
0.71533203125,
0.157958984375,
0.71533203125,
0.1728515625,
0.285400390625,
0.1728515625
],
[
0.2626953125,
0.1728515625,
0.73291015625,
0.171630859375,
0.73291015625,
0.189453125,
0.2626953125,
0.190673828125
],
[
0.1822509765625,
0.1859130859375,
0.82373046875,
0.18701171875,
0.82373046875,
0.204833984375,
0.1822509765625,
0.2037353515625
],
[
0.231689453125,
0.2037353515625,
0.771484375,
0.2037353515625,
0.771484375,
0.21728515625,
0.231689453125,
0.21728515625
],
[
0.2301025390625,
0.2161865234375,
0.771484375,
0.21728515625,
0.771484375,
0.232177734375,
0.2301025390625,
0.23095703125
],
[
0.268798828125,
0.235107421875,
0.73779296875,
0.235107421875,
0.73779296875,
0.2452392578125,
0.268798828125,
0.2452392578125
],
[
0.268798828125,
0.2476806640625,
0.73486328125,
0.2476806640625,
0.73486328125,
0.262451171875,
0.268798828125,
0.262451171875
],
[
0.465087890625,
0.2666015625,
0.5380859375,
0.2666015625,
0.5380859375,
0.27783203125,
0.465087890625,
0.27783203125
],
[
0.118408203125,
0.279052734375,
0.8818359375,
0.279052734375,
0.8818359375,
0.29345703125,
0.118408203125,
0.29345703125
],
[
0.12005615234375,
0.29345703125,
0.8818359375,
0.29345703125,
0.8818359375,
0.30712890625,
0.12005615234375,
0.30712890625
],
[
0.118408203125,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.118408203125,
0.320556640625
],
[
0.12176513671875,
0.32177734375,
0.880859375,
0.32177734375,
0.880859375,
0.33203125,
0.12176513671875,
0.33203125
],
[
0.118408203125,
0.333251953125,
0.884765625,
0.333251953125,
0.884765625,
0.3466796875,
0.118408203125,
0.3466796875
],
[
0.1234130859375,
0.347412109375,
0.87890625,
0.347412109375,
0.87890625,
0.355224609375,
0.1234130859375,
0.355224609375
],
[
0.12005615234375,
0.359375,
0.88330078125,
0.359375,
0.88330078125,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.12176513671875,
0.37353515625,
0.8818359375,
0.37353515625,
0.8818359375,
0.38427734375,
0.12176513671875,
0.38427734375
],
[
0.1175537109375,
0.383056640625,
0.88330078125,
0.38427734375,
0.88330078125,
0.39794921875,
0.1175537109375,
0.396728515625
],
[
0.118408203125,
0.410400390625,
0.81005859375,
0.410400390625,
0.81005859375,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.225830078125,
0.435302734375,
0.382080078125,
0.435302734375,
0.382080078125,
0.449462890625,
0.225830078125,
0.449462890625
],
[
0.51318359375,
0.435791015625,
0.8818359375,
0.435302734375,
0.8818359375,
0.449462890625,
0.51318359375,
0.45068359375
],
[
0.5146484375,
0.45068359375,
0.88330078125,
0.45068359375,
0.88330078125,
0.465576171875,
0.5146484375,
0.465576171875
],
[
0.118408203125,
0.46728515625,
0.490234375,
0.46728515625,
0.490234375,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.51123046875,
0.465576171875,
0.880859375,
0.465576171875,
0.880859375,
0.478515625,
0.51123046875,
0.478515625
],
[
0.118408203125,
0.482177734375,
0.491943359375,
0.482177734375,
0.491943359375,
0.495361328125,
0.118408203125,
0.495361328125
],
[
0.51123046875,
0.478515625,
0.88330078125,
0.479736328125,
0.88330078125,
0.493408203125,
0.51123046875,
0.4921875
],
[
0.12005615234375,
0.49755859375,
0.491943359375,
0.49755859375,
0.491943359375,
0.50830078125,
0.12005615234375,
0.50830078125
],
[
0.515625,
0.495361328125,
0.8818359375,
0.495361328125,
0.8818359375,
0.505859375,
0.515625,
0.505859375
],
[
0.12005615234375,
0.51025390625,
0.490234375,
0.51025390625,
0.490234375,
0.52392578125,
0.12005615234375,
0.52392578125
],
[
0.51318359375,
0.50830078125,
0.8818359375,
0.50830078125,
0.8818359375,
0.521484375,
0.51318359375,
0.521484375
],
[
0.12005615234375,
0.52490234375,
0.490234375,
0.52490234375,
0.490234375,
0.5380859375,
0.12005615234375,
0.5380859375
],
[
0.51318359375,
0.5224609375,
0.88330078125,
0.5224609375,
0.88330078125,
0.5361328125,
0.51318359375,
0.5361328125
],
[
0.118408203125,
0.53759765625,
0.489501953125,
0.53759765625,
0.489501953125,
0.55029296875,
0.118408203125,
0.55029296875
],
[
0.51318359375,
0.53759765625,
0.88330078125,
0.53759765625,
0.88330078125,
0.55029296875,
0.51318359375,
0.55029296875
],
[
0.118408203125,
0.5517578125,
0.490234375,
0.55029296875,
0.490234375,
0.5654296875,
0.118408203125,
0.56640625
],
[
0.51123046875,
0.5498046875,
0.8818359375,
0.5498046875,
0.8818359375,
0.5634765625,
0.51123046875,
0.5634765625
],
[
0.12005615234375,
0.56640625,
0.490234375,
0.56640625,
0.490234375,
0.580078125,
0.12005615234375,
0.580078125
],
[
0.51318359375,
0.5654296875,
0.88330078125,
0.5654296875,
0.88330078125,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.12005615234375,
0.58056640625,
0.491943359375,
0.58056640625,
0.491943359375,
0.59423828125,
0.12005615234375,
0.59423828125
],
[
0.5146484375,
0.580078125,
0.771484375,
0.580078125,
0.771484375,
0.59326171875,
0.5146484375,
0.59326171875
],
[
0.12176513671875,
0.5966796875,
0.489501953125,
0.5966796875,
0.489501953125,
0.60693359375,
0.12176513671875,
0.60693359375
],
[
0.53076171875,
0.5966796875,
0.88330078125,
0.5966796875,
0.88330078125,
0.61083984375,
0.53076171875,
0.61083984375
],
[
0.118408203125,
0.609375,
0.490234375,
0.609375,
0.490234375,
0.62255859375,
0.118408203125,
0.62255859375
],
[
0.5146484375,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.62451171875,
0.5146484375,
0.62451171875
],
[
0.1175537109375,
0.62353515625,
0.489501953125,
0.62353515625,
0.489501953125,
0.63720703125,
0.1175537109375,
0.63720703125
],
[
0.5146484375,
0.6259765625,
0.88330078125,
0.6259765625,
0.88330078125,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.12005615234375,
0.63818359375,
0.484375,
0.63818359375,
0.484375,
0.6484375,
0.12005615234375,
0.6484375
],
[
0.51318359375,
0.64013671875,
0.88330078125,
0.64013671875,
0.88330078125,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.12005615234375,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.51123046875,
0.65283203125,
0.88330078125,
0.65283203125,
0.88330078125,
0.66650390625,
0.51123046875,
0.66650390625
],
[
0.12005615234375,
0.66650390625,
0.489501953125,
0.66650390625,
0.489501953125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.5146484375,
0.66748046875,
0.88330078125,
0.66748046875,
0.88330078125,
0.68115234375,
0.5146484375,
0.68115234375
],
[
0.12005615234375,
0.6796875,
0.489501953125,
0.6796875,
0.489501953125,
0.693359375,
0.12005615234375,
0.693359375
],
[
0.51123046875,
0.681640625,
0.88330078125,
0.68115234375,
0.88330078125,
0.69482421875,
0.51123046875,
0.6953125
],
[
0.12005615234375,
0.69482421875,
0.489501953125,
0.69482421875,
0.489501953125,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.5146484375,
0.69775390625,
0.88330078125,
0.69775390625,
0.88330078125,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.12005615234375,
0.7080078125,
0.489501953125,
0.7080078125,
0.489501953125,
0.7216796875,
0.12005615234375,
0.7216796875
],
[
0.51318359375,
0.708984375,
0.880859375,
0.708984375,
0.880859375,
0.72265625,
0.51318359375,
0.72265625
],
[
0.118408203125,
0.72265625,
0.489501953125,
0.72265625,
0.489501953125,
0.736328125,
0.118408203125,
0.736328125
],
[
0.51123046875,
0.7236328125,
0.88330078125,
0.724609375,
0.88330078125,
0.7392578125,
0.51123046875,
0.73828125
],
[
0.12005615234375,
0.73828125,
0.489501953125,
0.73828125,
0.489501953125,
0.7490234375,
0.12005615234375,
0.7490234375
],
[
0.5146484375,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.5146484375,
0.7529296875
],
[
0.118408203125,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.51318359375,
0.7529296875,
0.88330078125,
0.75439453125,
0.88330078125,
0.76708984375,
0.51318359375,
0.7666015625
],
[
0.1175537109375,
0.76416015625,
0.48779296875,
0.76416015625,
0.48779296875,
0.77783203125,
0.1175537109375,
0.77783203125
],
[
0.5146484375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.1175537109375,
0.77978515625,
0.420654296875,
0.77978515625,
0.420654296875,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.51318359375,
0.7822265625,
0.8818359375,
0.7822265625,
0.8818359375,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.1368408203125,
0.7958984375,
0.491943359375,
0.7958984375,
0.491943359375,
0.80859375,
0.1368408203125,
0.80859375
],
[
0.5146484375,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.5146484375,
0.81005859375
],
[
0.118408203125,
0.81005859375,
0.490234375,
0.81005859375,
0.490234375,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51123046875,
0.80859375,
0.8818359375,
0.80859375,
0.8818359375,
0.82373046875,
0.51123046875,
0.82373046875
],
[
0.12005615234375,
0.82470703125,
0.490234375,
0.82470703125,
0.490234375,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.5146484375,
0.82470703125,
0.88330078125,
0.82470703125,
0.88330078125,
0.83837890625,
0.5146484375,
0.83837890625
],
[
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,
0.490234375,
0.8525390625,
0.490234375,
0.8662109375,
0.12005615234375,
0.8662109375
],
[
0.515625,
0.85400390625,
0.8818359375,
0.85400390625,
0.8818359375,
0.8642578125,
0.515625,
0.8642578125
],
[
0.12005615234375,
0.86767578125,
0.490234375,
0.86767578125,
0.490234375,
0.88134765625,
0.12005615234375,
0.88134765625
],
[
0.515625,
0.86767578125,
0.88671875,
0.86767578125,
0.88671875,
0.88134765625,
0.515625,
0.88134765625
],
[
0.12176513671875,
0.88134765625,
0.491943359375,
0.88134765625,
0.491943359375,
0.89453125,
0.12176513671875,
0.89453125
],
[
0.51318359375,
0.88134765625,
0.88330078125,
0.88134765625,
0.88330078125,
0.89453125,
0.51318359375,
0.89453125
],
[
0.12005615234375,
0.8955078125,
0.491943359375,
0.8955078125,
0.491943359375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"CoRl:CJKVBenchmarkwithRomanization Integration-Astep",
"towardsCross-lingualTransferBeyondTextual Scripts",
"Hoang H. Nguyen', Chenwei Zhang?, Ye Liu3,",
"Natalie Parde', Eugene Rohrbaugh1, Philip S.Yu",
"1 Department of Computer Science, University of llinois at Chicago, Chicago IL, USA",
"2Amazon,Seattle,WA, USA 3 Salesforce Research, Palo Alto,CA,USA",
"4 Harrisburg University of Science and Technology , Harrisburg, PA, USA",
"(hnguy7,parde,psyu@uic.edu,cwzhang@amazon.com",
"yeliu@salesforce.com, gene.rohrbaugh@gmail.com",
"Abstract",
"Naively assuming English as a source language may hinder cross-lingual transfer for many languages by failing to",
"consider the importance of language contact. Some languages are more well-connected than others, and targe",
"languages can benefit from transferring from closely related languages; for many languages, the set of closely",
"relatedlanguagesdoesnotincludeEnglish.Inthiswork,westudytheimpactofsourcelanguageforcross-lingual",
"transfer, demonstrating the importance of selecting source languages that have high contact with the target",
"anauaaeWealsoconstructanovelbenchmarkdatasetforclosecontactChinese-lapanese-Korean-Vietnamese",
"(CJKV) languages tofurtherencourage in-depth studies of language contact.To comprehensively capture contact",
"betweentheselanguages,weproposetointegrateRomanizedtranscriptionbeyondtextual scriptsviaContrastive",
"Learning objectives,leadingto enhanced cross-lingual representations and effectivezero-shot cross-lingual transfer.",
"Keywords:Cross-lingual Transfer,Low-resource Languages, Orthographic and Phonemic Transcription",
"1.Introduction",
"language contact (Nguyen and Rohrbaugh,2019;",
"Yang et al., 2022). As language acquisition nat",
"Language is fluid and ever-evolving, with specific",
"urally benefits from various linguistic modalities,",
"languages often borrowing terms and even gram-",
"including the textual writing scripts and aural sig",
"matical structure from one another as a byprod-",
"nals,whichcanbeconveyedthroughorthographic",
"uct of voluntary or forced interaction. Language",
"and phonemic transcription respectively (Volterra",
"contact defines the emerging connections among",
"and lverson; Goswami, 2022),the aforementioned",
"languages throughout the course of history,ge-",
"constraint hinders progress in the cross-lingual",
"ography, and social development (Matras, 2020;",
"field, by minimizing the amount of relevant informa",
"Mufwene and Escobar, 2022). For example, mod-",
"tion that models may be able to access (Fujinuma",
"ern English (EN) is a Germanic language, yet it",
"et al., 2022; Nguyen et al., 2023c).",
"containsmanyLatinateandGreektermsthaten-",
"Because of language contact, there are many",
"tered the vernacular via religious missionaries or",
"lexical items shared among CJKV languages.",
"war and conquest, as well as other terms that later",
"However, these similarities may be obscured by",
"enteredthelexiconviacolonization(vanGelderen",
"the different writing systems used across the lan-",
"2014).Likewise,Japanese (JA),Korean (KO)",
"guages. Borrowed words that may sound very",
"and Vietnamese (Vi) share a large number of cog-",
"similar in spoken form may nevertheless be repre",
"nates with Chinese (ZH) due to influence through",
"sented in the writing systems with completely dif",
"colonization, trading throughout history (Li, 2019;",
"ferentsymbols.As observed in Table 1,theor-",
"Hickey, 2020).This influence is best reflected",
"thographic term 古典 in ZH shares little similarity",
"through the Kanji characters in JA writing system,",
"with vi counterpart. However, the corresponding",
"whichsignificantlyoverlapswithZHcharacterwrit-",
"phonemic transcriptionrepresented via Romaniza",
"ing systems. As observed in Table 1, both ZH and",
"tion (gudian) is more similar to the corresponding",
"JA writingsystems utilize文学and学者to denote",
"VI (Co dién).The same phenomenon can also",
"words scholar and classical respectively.",
"be observed in JA-KO pair (古典(koten)vs 卫",
"Despite recent advances in the study of cross-",
"(gojeon)) and Vi-KO pair (Khoa hoc (Khoa hoc)",
"lingual transfer, most cross-lingual works to date",
"vs (gwahak))Moreover, Table 1 reveals",
"have restricted themselves to two major assump-",
"the shared properties of CJKV languages in which",
"tions: (1) English (EN) is considered the single",
"meaningful words are formed by combining sur",
"source language, despite its oftentimes limited or",
"roundingtokens.Forinstance,fourcharacters形",
"absent contact with the target language(s) (Muller",
"而上学 are combined together to form the mean-",
"et al., 2021; Fujinuma et al., 2022); and (2) cross-",
"ingful word \"Metaphysical\".The same behavior",
"lingual transfer remains restricted to orthographic",
"can be observed in the corresponding Vl (Sieu"
] |
{
"arxiv_id": "2404.12618",
"language": "en",
"page_idx": 0
}
|
||
2404.12618-en-1
|
2404.12618-en-1.jpg
|
hình) and KO (형이상학). Although the current multiilngual benchmark dataset such as XTREME (Hu et al., 2020) offers extensive language coverage, the aforementioned shared properties among CJKV are often overlooked during preprocessing. We address these limitations through the follow-
ing contributions:
· We study the impact of source language selection in zero-shot cross-lingual transfer among languages with historical language contact.
·We construct a comprehensive set of language understanding (NLU) tasks for CJKV languages that (1) better presents the shared characteristics of CJKV languages than the existing multilingual XTREME benchmark, (2) provides additional phonemic signals captured through Romanized transcriptions beyond orthographic textual scripts.
·We integrate phonemic information from Romanization with Orthographic texts via Contrastive Learning to enhance textual representation across CJKV languages, leading to improved cross-lingual representation and downstream task performance.$^{1}$
## 2. Related Work
Cross-lingual Transfer. Recent works on crosslingual transfer have focused on generating multilingual contextualized representations for differ- (PLMs) trained using multilingual dictionaries (Qin et al., 2021; Winata et al., 2023) and/or machine translation approaches (Fang et al., 2021; Liu et al., 2021c; Reid and Artetxe, 2023). Qin et al. tences, and tokens to enhance multilingual representations, leading to improved downstream task performance on target languages. On the other hand, other approaches have leveraged machinetranslated parallel corpora to (1) distill knowledge from source languages to target languages (Fang et al., 2021) or (2) augment source language data with target language knowledge during training (Yang et al., 2022; Zheng et al., 2021).
However, current cross-lingual efforts concentrate on transferring from a single source language (EN) to multiple target languages (Hu et al., 2020; Yang et al., 2022). Under parameter capacity constraints, cross-lingual transfer has been shown to be biased towards high-resourced languages which share similar scripts and possess larger
1Our code and datasets are publicly available at
https://github.com/nhhoang96/benchmark_cjkv
corpora of unlabeled data during pre-training (Fujinuma et al., 2022). Recent works (Adams et al., 2017; Nguyen et al., 2023c) attempt to bridge the gap between languages across different scripts by introducing phonemic transcription. However, the introduction of the International Phonetic Alphabet (IPA) to this process requires additional pre-training objectives to effectively unify the two
Unlike previous works, we lever-
modalities.
age natural Romanized transcriptions to capture
phonemic properties.
Contrastive Learning. Contrastive Learning (CL) has been widely leveraged as an effective representation learning mechanism (Oord et al., 2018; Chen et al., 2020). The goal of CL is to learn the discriminative features of instances via different augmentation methods. In Natural Language Processing (NLP), CL has been adopted in various contexts ranging from text classification (Wei and Zou, 2019) and representation learning (Gao et al., 2021; Nguyen et al., 2023b) to question answering (Xiong et al., 2020; Liu et al., 2021b). Unlike previous works, we leverage two language modalities (orthographic and Romanized transcriptions) to generate multi-view augmentations as effective mechanisms to improve cross-lingual representations.
# 3. Problem Formulation
We study the problem of cross-lingual transfer in a multilingual setting where there exists annotated
linigua conlexualzed represenlaions or diller- data $D$ collected from a high-resource language $S$,ent languages via Pre-trained Language Models such that $D_S^{train}=\{(X_i^{(S)},Y_i^{(S)})\}_{i=1}^{N_s}$ where $N_s$ de-
notes the sample size of the source language training data. Likewise, there exists unlabeled data collected from a set of low-resource target lan-
of th'e $i$-th target language inference dataset. Depending on the downstream tasks, $Y_i$ can be either a single utterance label (Sentence-level task), a sequence of token labels (Token-level Task), or extracted span labels (Question Answering (QA)). Our work encompasses a comprehensive set of Natural Language Understanding (NLU) tasks among CJKV languages as briefly summarized below. For further details of individual task objectives and metrics, we refer readers to (Hu et al., 2020).
Sentence-level Task (PAWSX, XNLI) concentrates on overall sentence-level semantic understanding of the given input paired utterance. The two covered tasks are multilingual Paraphrase Identification (PAWSX) and Natural Language Inference (XNLI).
Token-level Task (UDPOS, PANX) requires deeper understanding on token-level which re-
|
[
[
0.118408203125,
0.0736083984375,
0.490234375,
0.0748291015625,
0.490234375,
0.08905029296875,
0.118408203125,
0.0885009765625
],
[
0.51318359375,
0.07598876953125,
0.88330078125,
0.07598876953125,
0.88330078125,
0.08905029296875,
0.51318359375,
0.08905029296875
],
[
0.12005615234375,
0.08905029296875,
0.489501953125,
0.08905029296875,
0.489501953125,
0.10272216796875,
0.12005615234375,
0.10272216796875
],
[
0.51318359375,
0.08905029296875,
0.88330078125,
0.08905029296875,
0.88330078125,
0.10272216796875,
0.51318359375,
0.10272216796875
],
[
0.12005615234375,
0.10394287109375,
0.489501953125,
0.1051025390625,
0.489501953125,
0.11639404296875,
0.12005615234375,
0.11517333984375
],
[
0.5146484375,
0.10394287109375,
0.884765625,
0.10394287109375,
0.884765625,
0.1175537109375,
0.5146484375,
0.1175537109375
],
[
0.12005615234375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.118408203125,
0.1312255859375,
0.486083984375,
0.1318359375,
0.486083984375,
0.146728515625,
0.118408203125,
0.1455078125
],
[
0.51123046875,
0.1318359375,
0.8818359375,
0.133056640625,
0.8818359375,
0.146728515625,
0.51123046875,
0.1455078125
],
[
0.133544921875,
0.1455078125,
0.491943359375,
0.146728515625,
0.491943359375,
0.160888671875,
0.133544921875,
0.1602783203125
],
[
0.5146484375,
0.1478271484375,
0.88330078125,
0.1478271484375,
0.88330078125,
0.157958984375,
0.5146484375,
0.157958984375
],
[
0.118408203125,
0.160888671875,
0.2476806640625,
0.160888671875,
0.2476806640625,
0.174560546875,
0.118408203125,
0.174560546875
],
[
0.515625,
0.162109375,
0.8818359375,
0.162109375,
0.8818359375,
0.1728515625,
0.515625,
0.1728515625
],
[
0.5146484375,
0.17578125,
0.88330078125,
0.17578125,
0.88330078125,
0.189453125,
0.5146484375,
0.189453125
],
[
0.1368408203125,
0.1859130859375,
0.486083984375,
0.1859130859375,
0.486083984375,
0.196533203125,
0.1368408203125,
0.196533203125
],
[
0.5146484375,
0.189453125,
0.59765625,
0.189453125,
0.59765625,
0.20068359375,
0.5146484375,
0.20068359375
],
[
0.60693359375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.60693359375,
0.203125
],
[
0.151123046875,
0.20068359375,
0.490234375,
0.20068359375,
0.490234375,
0.21435546875,
0.151123046875,
0.21435546875
],
[
0.51318359375,
0.2037353515625,
0.884765625,
0.2037353515625,
0.884765625,
0.21728515625,
0.51318359375,
0.21728515625
],
[
0.1502685546875,
0.213134765625,
0.47265625,
0.21435546875,
0.47265625,
0.2286376953125,
0.1502685546875,
0.22802734375
],
[
0.51318359375,
0.218505859375,
0.67333984375,
0.218505859375,
0.67333984375,
0.232177734375,
0.51318359375,
0.232177734375
],
[
0.1376953125,
0.2381591796875,
0.490234375,
0.2381591796875,
0.490234375,
0.251708984375,
0.1376953125,
0.251708984375
],
[
0.5146484375,
0.2410888671875,
0.71875,
0.2410888671875,
0.71875,
0.254638671875,
0.5146484375,
0.254638671875
],
[
0.70947265625,
0.2392578125,
0.88330078125,
0.2410888671875,
0.88330078125,
0.255859375,
0.70947265625,
0.254150390625
],
[
0.15283203125,
0.2529296875,
0.489501953125,
0.2529296875,
0.489501953125,
0.2666015625,
0.15283203125,
0.2666015625
],
[
0.5146484375,
0.255859375,
0.884765625,
0.255859375,
0.884765625,
0.26953125,
0.5146484375,
0.26953125
],
[
0.15283203125,
0.267333984375,
0.489501953125,
0.267333984375,
0.489501953125,
0.28076171875,
0.15283203125,
0.28076171875
],
[
0.515625,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.281982421875,
0.515625,
0.281982421875
],
[
0.151123046875,
0.27978515625,
0.489501953125,
0.28076171875,
0.489501953125,
0.295654296875,
0.151123046875,
0.29443359375
],
[
0.51318359375,
0.283935546875,
0.88330078125,
0.283935546875,
0.88330078125,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.15283203125,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.15283203125,
0.308837890625
],
[
0.51318359375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.15283203125,
0.31005859375,
0.490234375,
0.31005859375,
0.490234375,
0.32373046875,
0.15283203125,
0.32373046875
],
[
0.5146484375,
0.3134765625,
0.88330078125,
0.3134765625,
0.88330078125,
0.32666015625,
0.5146484375,
0.32666015625
],
[
0.154541015625,
0.325927734375,
0.489501953125,
0.325927734375,
0.489501953125,
0.336181640625,
0.154541015625,
0.336181640625
],
[
0.5146484375,
0.32666015625,
0.88330078125,
0.32666015625,
0.88330078125,
0.34033203125,
0.5146484375,
0.34033203125
],
[
0.15283203125,
0.33837890625,
0.4013671875,
0.33837890625,
0.4013671875,
0.352783203125,
0.15283203125,
0.352783203125
],
[
0.5146484375,
0.341552734375,
0.88330078125,
0.341552734375,
0.88330078125,
0.355224609375,
0.5146484375,
0.355224609375
],
[
0.51318359375,
0.35400390625,
0.88330078125,
0.355224609375,
0.88330078125,
0.369384765625,
0.51318359375,
0.3681640625
],
[
0.1376953125,
0.3623046875,
0.490234375,
0.3623046875,
0.490234375,
0.3759765625,
0.1376953125,
0.3759765625
],
[
0.5146484375,
0.369384765625,
0.88330078125,
0.369384765625,
0.88330078125,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.151123046875,
0.377197265625,
0.490234375,
0.377197265625,
0.490234375,
0.39013671875,
0.151123046875,
0.39013671875
],
[
0.5146484375,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.151123046875,
0.391357421875,
0.490234375,
0.391357421875,
0.490234375,
0.405029296875,
0.151123046875,
0.405029296875
],
[
0.51318359375,
0.396728515625,
0.88330078125,
0.39794921875,
0.88330078125,
0.412109375,
0.51318359375,
0.410888671875
],
[
0.151123046875,
0.405029296875,
0.48779296875,
0.405029296875,
0.48779296875,
0.419189453125,
0.151123046875,
0.419189453125
],
[
0.5146484375,
0.412109375,
0.884765625,
0.412109375,
0.884765625,
0.42578125,
0.5146484375,
0.42578125
],
[
0.151123046875,
0.419189453125,
0.489501953125,
0.419189453125,
0.489501953125,
0.432861328125,
0.151123046875,
0.432861328125
],
[
0.5146484375,
0.427001953125,
0.88330078125,
0.427001953125,
0.88330078125,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.151123046875,
0.43408203125,
0.391357421875,
0.43408203125,
0.391357421875,
0.44775390625,
0.151123046875,
0.44775390625
],
[
0.5146484375,
0.4423828125,
0.880859375,
0.4423828125,
0.880859375,
0.453125,
0.5146484375,
0.453125
],
[
0.5146484375,
0.454833984375,
0.6396484375,
0.454833984375,
0.6396484375,
0.468505859375,
0.5146484375,
0.468505859375
],
[
0.2149658203125,
0.47021484375,
0.25439453125,
0.47021484375,
0.25439453125,
0.486328125,
0.2149658203125,
0.486328125
],
[
0.255126953125,
0.4697265625,
0.391357421875,
0.46728515625,
0.391357421875,
0.48388671875,
0.255126953125,
0.486328125
],
[
0.2509765625,
0.4755859375,
0.255126953125,
0.4755859375,
0.255126953125,
0.48095703125,
0.2509765625,
0.48095703125
],
[
0.57666015625,
0.48095703125,
0.8193359375,
0.48095703125,
0.8193359375,
0.495361328125,
0.57666015625,
0.495361328125
],
[
0.12005615234375,
0.495361328125,
0.490234375,
0.495361328125,
0.490234375,
0.5087890625,
0.12005615234375,
0.5087890625
],
[
0.51318359375,
0.501953125,
0.88330078125,
0.501953125,
0.88330078125,
0.515625,
0.51318359375,
0.515625
],
[
0.12176513671875,
0.5126953125,
0.489501953125,
0.5126953125,
0.489501953125,
0.5224609375,
0.12176513671875,
0.5224609375
],
[
0.5146484375,
0.5166015625,
0.88330078125,
0.5166015625,
0.88330078125,
0.52978515625,
0.5146484375,
0.52978515625
],
[
0.118408203125,
0.52490234375,
0.491943359375,
0.52490234375,
0.491943359375,
0.5380859375,
0.118408203125,
0.5380859375
],
[
0.51123046875,
0.529296875,
0.88330078125,
0.52978515625,
0.88330078125,
0.54443359375,
0.51123046875,
0.54345703125
],
[
0.12005615234375,
0.5380859375,
0.489501953125,
0.5380859375,
0.489501953125,
0.552734375,
0.12005615234375,
0.552734375
],
[
0.51123046875,
0.54345703125,
0.88818359375,
0.54345703125,
0.88818359375,
0.56103515625,
0.51123046875,
0.56103515625
],
[
0.12005615234375,
0.552734375,
0.490234375,
0.552734375,
0.490234375,
0.56640625,
0.12005615234375,
0.56640625
],
[
0.51318359375,
0.55859375,
0.88330078125,
0.55859375,
0.88330078125,
0.572265625,
0.51318359375,
0.572265625
],
[
0.12005615234375,
0.56787109375,
0.486083984375,
0.56787109375,
0.486083984375,
0.57763671875,
0.12005615234375,
0.57763671875
],
[
0.51123046875,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58740234375,
0.51123046875,
0.58740234375
],
[
0.118408203125,
0.58056640625,
0.490234375,
0.58056640625,
0.490234375,
0.59423828125,
0.118408203125,
0.59423828125
],
[
0.51318359375,
0.5888671875,
0.884765625,
0.5888671875,
0.884765625,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.12005615234375,
0.595703125,
0.484375,
0.595703125,
0.484375,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.51318359375,
0.6025390625,
0.8427734375,
0.6025390625,
0.8427734375,
0.61767578125,
0.51318359375,
0.61767578125
],
[
0.12005615234375,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.12005615234375,
0.62353515625
],
[
0.5146484375,
0.6162109375,
0.884765625,
0.6162109375,
0.884765625,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.118408203125,
0.62255859375,
0.490234375,
0.62353515625,
0.490234375,
0.63818359375,
0.118408203125,
0.63720703125
],
[
0.118408203125,
0.63818359375,
0.490234375,
0.63818359375,
0.490234375,
0.65185546875,
0.118408203125,
0.65185546875
],
[
0.5146484375,
0.63427734375,
0.88671875,
0.63427734375,
0.88671875,
0.64794921875,
0.5146484375,
0.64794921875
],
[
0.12005615234375,
0.65283203125,
0.490234375,
0.65283203125,
0.490234375,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.5146484375,
0.6484375,
0.884765625,
0.6484375,
0.884765625,
0.662109375,
0.5146484375,
0.662109375
],
[
0.12005615234375,
0.66748046875,
0.489501953125,
0.66748046875,
0.489501953125,
0.68115234375,
0.12005615234375,
0.68115234375
],
[
0.515625,
0.66455078125,
0.880859375,
0.66455078125,
0.880859375,
0.6748046875,
0.515625,
0.6748046875
],
[
0.118408203125,
0.68115234375,
0.490234375,
0.68115234375,
0.490234375,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.5146484375,
0.6767578125,
0.88330078125,
0.6767578125,
0.88330078125,
0.6904296875,
0.5146484375,
0.6904296875
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.7060546875,
0.118408203125,
0.7060546875
],
[
0.5146484375,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.5146484375,
0.705078125
],
[
0.1175537109375,
0.7080078125,
0.48779296875,
0.7080078125,
0.48779296875,
0.7216796875,
0.1175537109375,
0.7216796875
],
[
0.52880859375,
0.70361328125,
0.884765625,
0.705078125,
0.884765625,
0.7197265625,
0.52880859375,
0.71875
],
[
0.12005615234375,
0.7236328125,
0.489501953125,
0.7236328125,
0.489501953125,
0.73681640625,
0.12005615234375,
0.73681640625
],
[
0.51318359375,
0.7197265625,
0.88330078125,
0.7197265625,
0.88330078125,
0.73291015625,
0.51318359375,
0.73291015625
],
[
0.1158447265625,
0.736328125,
0.489501953125,
0.73828125,
0.489501953125,
0.751953125,
0.1158447265625,
0.75
],
[
0.5146484375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.118408203125,
0.751953125,
0.403076171875,
0.751953125,
0.403076171875,
0.765625,
0.118408203125,
0.765625
],
[
0.51318359375,
0.74755859375,
0.88330078125,
0.74755859375,
0.88330078125,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.1351318359375,
0.7666015625,
0.491943359375,
0.7666015625,
0.491943359375,
0.77978515625,
0.1351318359375,
0.77978515625
],
[
0.5146484375,
0.7626953125,
0.87744140625,
0.7626953125,
0.87744140625,
0.775390625,
0.5146484375,
0.775390625
],
[
0.1175537109375,
0.779296875,
0.489501953125,
0.78076171875,
0.489501953125,
0.79443359375,
0.1175537109375,
0.7919921875
],
[
0.5146484375,
0.78515625,
0.884765625,
0.7861328125,
0.884765625,
0.7998046875,
0.5146484375,
0.798828125
],
[
0.12005615234375,
0.79443359375,
0.489501953125,
0.79443359375,
0.489501953125,
0.80810546875,
0.12005615234375,
0.80810546875
],
[
0.51318359375,
0.80029296875,
0.88671875,
0.80029296875,
0.88671875,
0.81396484375,
0.51318359375,
0.81396484375
],
[
0.12005615234375,
0.80859375,
0.489501953125,
0.80859375,
0.489501953125,
0.822265625,
0.12005615234375,
0.822265625
],
[
0.5146484375,
0.81640625,
0.88330078125,
0.81640625,
0.88330078125,
0.82666015625,
0.5146484375,
0.82666015625
],
[
0.1175537109375,
0.822265625,
0.490234375,
0.82373046875,
0.490234375,
0.83740234375,
0.1175537109375,
0.8359375
],
[
0.51318359375,
0.8291015625,
0.88330078125,
0.8291015625,
0.88330078125,
0.8427734375,
0.51318359375,
0.8427734375
],
[
0.118408203125,
0.83740234375,
0.489501953125,
0.83740234375,
0.489501953125,
0.85107421875,
0.118408203125,
0.85107421875
],
[
0.5146484375,
0.84326171875,
0.88671875,
0.84326171875,
0.88671875,
0.85693359375,
0.5146484375,
0.85693359375
],
[
0.1175537109375,
0.8515625,
0.489501953125,
0.8515625,
0.489501953125,
0.865234375,
0.1175537109375,
0.865234375
],
[
0.51318359375,
0.85791015625,
0.62744140625,
0.85791015625,
0.62744140625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.14111328125,
0.8818359375,
0.489501953125,
0.8818359375,
0.489501953125,
0.8955078125,
0.14111328125,
0.8955078125
],
[
0.51318359375,
0.8798828125,
0.80615234375,
0.8798828125,
0.80615234375,
0.89453125,
0.51318359375,
0.89453125
],
[
0.81689453125,
0.8828125,
0.88330078125,
0.8828125,
0.88330078125,
0.89453125,
0.81689453125,
0.89453125
],
[
0.12176513671875,
0.89794921875,
0.432373046875,
0.89794921875,
0.432373046875,
0.90771484375,
0.12176513671875,
0.90771484375
],
[
0.5146484375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"hinh) and KO (o|). Although the current",
"corpora of unlabeled data during pre-training (Fu-",
"multilngual benchmark dataset such as XTREME",
"jinuma et al., 2022). Recent works (Adams et al.,",
"(Hu etal.,2020)offers extensive language cover-",
"2017; Nguyen et al., 2023c) attempt to bridge the",
"age,the aforementioned shared properties among",
"gap between languages across different scripts",
"CJKV are often overlooked during preprocessing.",
"by introducing phonemic transcription. However,",
"We address these limitations through the follow-",
"the introduction of the International Phonetic Al-",
"ing contributions:",
"phabet (IPA)to this process requires additional",
"pre-training objectives to effectively unify the two",
".Westudytheimpactofsourcelanguageselec",
"modalities.",
"Unlike previous works, we lever-",
"tion in zero-shot cross-lingual transfer among",
"age natural Romanized transcriptions to capture",
"languages with historical language contact.",
"phonemic properties.",
".We construct a comprehensive set of lan-",
"Contrastive Learning.",
"Contrastive Learning",
"guage understanding (NLU) tasks for CJKV",
"(CL) has been widely leveraged as an effective",
"languages that (1) better presents the shared",
"representationlearning mechanism (Oord et al.,",
"characteristics of CJKv ianguages than the",
"2018; Chen et al., 2020). The goal of CL is to learn",
"existing multilingualXTREMEbenchmark, (2)",
"the discriminative features of instances via differ-",
"provides additional phonemic signals cap-",
"ent augmentation methods. In Natural Language",
"tured throughRomanizedtranscriptionsbe-",
"Processing (NLP), CL has been adopted in vari-",
"yond orthographic textual scripts.",
"ous contexts ranging from text classification (Wei",
"and Zou, 2019) and representation learning (Gao",
":We integrate phonemic information from Ro-",
"et al., 2021; Nguyen et al., 2023b) to question",
"manization with Orthographic texts via Con-",
"answering (Xiong et al., 2020; Liu et al., 2021b).",
"trastive Learning to enhance textual repre-",
"Unlike previous works, we leverage two language",
"sentation across CJKV languages, leading",
"modalities (orthographic and Romanized tran-",
"to improved cross-lingual representation and",
"scriptions)to generate multi-view augmentations",
"downstreamtaskperformance.1",
"aseffectivemechanismsto improvecross-lingual",
"representations.",
"2.",
"Related Work",
"",
"3.ProblemFormulation",
"Cross-lingual Transfer. Recent works on cross-",
"We study the problem of cross-lingual transfer in",
"lingual transferhavefocusedongeneratingmul-",
"a multilingual setting where there exists annotated",
"tilingual contextualized representations for differ-",
"data D collected from a high-resource language S,",
"ent languages via Pre-trained Language Models",
"such that Dgrain = ((x(S), y(S), where N, de-",
"(PLMs) trained using multilingual dictionaries (Qin",
"notes the sample size of the source language train",
"etal..2021:Winata etal..2023)and/or machine",
"ing data. Likewise, there exists unlabeled data",
"translation approaches (Fang et al., 2021; Liu",
"collected from a set of low-resource target lan-",
"et al..2021c:Reid.and.Artetxe.2023).Qin etal",
"guages T = {Ti,., Tn], denoted as Drt",
"(2021) proposed a comprehensive code-switching",
"((x(T)),Ni where Nt, denotes the sample size",
"technique via random selection of languages, sen-",
"tences, and tokens to enhance multilingual repre-",
"of the i-th target language inference dataset. De-",
"sentations, leading to improved downstream task",
"pending on the downstream tasks, Y, can be ei-",
"performance on target languages..On the other",
"therasingleutterancelabel(Sentence-leveltask)",
"hand, other approaches have leveraged machine-",
"a sequence of token labels (Token-level Task), or",
"translatedparallelcorporato(1)distillknowledge",
"extracted span labels (Question Answering (QA).",
"from source languages to target languages (Fang",
"Our work encompasses a comprehensive set",
"et al., 2021) or (2) augment source language data",
"of Natural Language Understanding (NLU)tasks",
"with target language knowledge during training",
"amongCJKVlanguages asbrieflysummarized be-",
"(Yang et al., 2022; Zheng et al., 2021)",
"low.Forfurther details of individual task objectives",
"However, current cross-lingual efforts concen-",
"and metrics, we refer readers to (Hu et al., 2020).",
"trateontransferringfromasinglesourcelanguage",
"Sentence-level Task(PAwSX,XNLI)concen-",
"(EN) to multiple target languages (Hu et al., 2020;",
"trates on overall sentence-level semantic under-",
"Yang et al., 2022). Under parameter capacity",
"standing of the given inputpaired utterance.The",
"constraints,cross-lingualtransfer has been shown",
"two covered tasks are multilingual Paraphrase",
"to be biased towards high-resourced languages",
"Identification (PAWSX)and Natural Language In-",
"which share similar scripts and possess larger",
"ference (XNLI).",
"1Our code and datasets are publicly available at",
"Token-level Task (UDPOS, PANX)",
"requires",
"https://github.com/nhhoang96/benchmarkcjkv",
"deeper understanding on token-level which re-"
] |
{
"arxiv_id": "2404.12618",
"language": "en",
"page_idx": 1
}
|
||
2404.12618-en-2
|
2404.12618-en-2.jpg
|
Table 1: Orthographic and Romanized representations (abbreviated as Ortho and Roman) of a sample sentence across CJKV languages where colored segments denote the corresponding semantic words defined in Section 5.(.) denotes the specific name of Romanization system of the respective language The first sentence for each language denotes the currently preprocessed XTREME benchmark dataset
<FigureHere>
Table 2: Preliminary study on the impact of source language (ZH) on target JKV languages. Standard
deviation denotes the performance variation across 3 target JKV languages. Performance of individual
JKV language is reported in Table 7 in the Appendix.
<table>
<tbody>
<tr>
<th rowspan="2">Source Metric</th>
<th colspan="2">Sentence-level</th>
<th colspan="2">Token-level</th>
<th colspan="2">Question Answerinq</th>
</tr>
<tr>
<th>$PAWSX$ Acc</th>
<th>$XNLI$ Acc</th>
<th>UDPOS $F1$</th>
<th>PANX $F1$</th>
<th>XQuAD $EM$</th>
<th>MLQA $EM$</th>
</tr>
<tr>
<td>$EN$</td>
<td>$73.53\pm6.86$</td>
<td>$68.35\pm2.05$</td>
<td>40.09 $\partial\pm10.32$ 1</td>
<td>$30.36\pm19.09$</td>
<td>$28.38\pm18.00$</td>
<td>22.11 $\pm8.76$ $\bullet$</td>
</tr>
<tr>
<td>$ZH$</td>
<td>79.61 $\pm5.24$</td>
<td>70.39 $\pm1.60$</td>
<td>47.82 $2\pm1.98$</td>
<td>37.29 $\pm9.18$</td>
<td>33.95 $\pm14.73$ 1.</td>
<td>29.95 $\pm5.88$</td>
</tr>
</tbody>
</table>
quires additional grammatical structural knowl-
edge beyond overall semantics.
Representa-
tive tasks include Part-of-speech (POS) Tagging through UDPOS dataset and Named Entity Recognition (NER) via PANX dataset.
2020). We fine-tune a backbone XLM-R-base PLM (Conneau et al., 2020) with downstream task objectives for each source-target JKV language pair and report average performance together with standard deviation across the three target languages JA, KO, VI.
Overall, based on our preliminary study reported
in Table 2, we observe the following:
Question Answering Task (XQuAD, MLQA) extracts the answer spans existent in the given Context when provided with input corresponding Question. The task assumes the answers are existent in the given Context; therefore, within the scope of our study, any samples without existent answers in the contexts are removed from both training and evaluation. Unlike (Hu et al., 2020), we only report Exact Match (EM) for QA task since it is considered a stricter evaluation measure and requires models to extract truly meaningful answers.
$\cdot$ Average zero-shot cross-lingual transfer performance for target languages across all tasks increases significantly when ZH is leveraged as the source language. On UDPOS task, the average performance of JKV languages reaches F1 of 47.82 as compared to 40.09 when EN is considered source language.
· When ZH is leveraged as source language in-
stead of EN, the performance standard deviations across target JKV languages diminished (19.09 for EN source vs 9.18 when ZH source on PANX task), implying that appropriately selected source languages can reduce the performance gap across target languages.
Our observations are consistent with prior study (Fujinuma et al., 2022; Yang et al., 2022) demon-
## 4. Preliminary Study
We conduct a preliminary study to examine the effects of source language selection on target languages. In our case, we compare using EN against ZH as the source language to transfer knowledge towards other languages among the CJKV group. ZH is considered the common source language for CJKV due to its historical longevity and significant impact on the development of the remaining languages (Li,2019; Hickey,
strating that PLMs when trained on only EN can be biased towards Latin-based languages, resulting in poor performance for non-Latin based
|
[
[
0.1175537109375,
0.07244873046875,
0.88330078125,
0.0736083984375,
0.88330078125,
0.0872802734375,
0.1175537109375,
0.08612060546875
],
[
0.1175537109375,
0.08612060546875,
0.884765625,
0.0872802734375,
0.884765625,
0.1015625,
0.1175537109375,
0.1009521484375
],
[
0.1175537109375,
0.09979248046875,
0.8818359375,
0.1015625,
0.8818359375,
0.11639404296875,
0.1175537109375,
0.114013671875
],
[
0.1175537109375,
0.114013671875,
0.8818359375,
0.11517333984375,
0.8818359375,
0.1300048828125,
0.1175537109375,
0.12890625
],
[
0.12176513671875,
0.142578125,
0.187255859375,
0.144287109375,
0.187255859375,
0.1561279296875,
0.12176513671875,
0.15380859375
],
[
0.2149658203125,
0.144287109375,
0.270263671875,
0.144287109375,
0.270263671875,
0.1561279296875,
0.2149658203125,
0.1561279296875
],
[
0.5263671875,
0.1436767578125,
0.64208984375,
0.144287109375,
0.64208984375,
0.1561279296875,
0.5263671875,
0.155029296875
],
[
0.29296875,
0.1561279296875,
0.5322265625,
0.15673828125,
0.5322265625,
0.168701171875,
0.29296875,
0.16748046875
],
[
0.52880859375,
0.1561279296875,
0.6650390625,
0.1561279296875,
0.6650390625,
0.169189453125,
0.52880859375,
0.169189453125
],
[
0.225830078125,
0.15673828125,
0.26025390625,
0.15673828125,
0.26025390625,
0.168701171875,
0.225830078125,
0.168701171875
],
[
0.1435546875,
0.15673828125,
0.1695556640625,
0.15673828125,
0.1695556640625,
0.168701171875,
0.1435546875,
0.168701171875
],
[
0.6591796875,
0.157958984375,
0.7109375,
0.157958984375,
0.7109375,
0.168701171875,
0.6591796875,
0.168701171875
],
[
0.2275390625,
0.17041015625,
0.258544921875,
0.17041015625,
0.258544921875,
0.1817626953125,
0.2275390625,
0.1817626953125
],
[
0.297119140625,
0.17041015625,
0.5341796875,
0.17041015625,
0.5341796875,
0.18115234375,
0.297119140625,
0.18115234375
],
[
0.12176513671875,
0.1817626953125,
0.187255859375,
0.18408203125,
0.187255859375,
0.1953125,
0.12176513671875,
0.193603515625
],
[
0.214111328125,
0.18408203125,
0.273681640625,
0.18408203125,
0.273681640625,
0.19482421875,
0.214111328125,
0.19482421875
],
[
0.2939453125,
0.182861328125,
0.462646484375,
0.182861328125,
0.462646484375,
0.193603515625,
0.2939453125,
0.193603515625
],
[
0.468505859375,
0.18408203125,
0.7138671875,
0.18408203125,
0.7138671875,
0.19482421875,
0.468505859375,
0.19482421875
],
[
0.20068359375,
0.19482421875,
0.28369140625,
0.196533203125,
0.28369140625,
0.2078857421875,
0.20068359375,
0.2060546875
],
[
0.2939453125,
0.1953125,
0.736328125,
0.1953125,
0.736328125,
0.208984375,
0.2939453125,
0.208984375
],
[
0.78515625,
0.198974609375,
0.82861328125,
0.198974609375,
0.82861328125,
0.2060546875,
0.78515625,
0.2060546875
],
[
0.2275390625,
0.21142578125,
0.258544921875,
0.21142578125,
0.258544921875,
0.22265625,
0.2275390625,
0.22265625
],
[
0.2939453125,
0.210205078125,
0.462646484375,
0.210205078125,
0.462646484375,
0.223876953125,
0.2939453125,
0.223876953125
],
[
0.465087890625,
0.21142578125,
0.5205078125,
0.21142578125,
0.5205078125,
0.221435546875,
0.465087890625,
0.221435546875
],
[
0.5205078125,
0.21142578125,
0.57421875,
0.21142578125,
0.57421875,
0.221435546875,
0.5205078125,
0.221435546875
],
[
0.59033203125,
0.21142578125,
0.6513671875,
0.21142578125,
0.6513671875,
0.221435546875,
0.59033203125,
0.221435546875
],
[
0.6455078125,
0.21142578125,
0.6943359375,
0.21142578125,
0.6943359375,
0.221435546875,
0.6455078125,
0.221435546875
],
[
0.1275634765625,
0.2203369140625,
0.1822509765625,
0.223876953125,
0.181396484375,
0.236328125,
0.1259765625,
0.2327880859375
],
[
0.5810546875,
0.21435546875,
0.59375,
0.21435546875,
0.59375,
0.218505859375,
0.5810546875,
0.218505859375
],
[
0.295654296875,
0.223876953125,
0.6533203125,
0.223876953125,
0.6533203125,
0.2369384765625,
0.295654296875,
0.2369384765625
],
[
0.6484375,
0.22265625,
0.81005859375,
0.221435546875,
0.81005859375,
0.236328125,
0.6484375,
0.2369384765625
],
[
0.64990234375,
0.236328125,
0.8017578125,
0.236328125,
0.8017578125,
0.2493896484375,
0.64990234375,
0.2493896484375
],
[
0.21240234375,
0.2244873046875,
0.27197265625,
0.2244873046875,
0.27197265625,
0.236328125,
0.21240234375,
0.236328125
],
[
0.295654296875,
0.2369384765625,
0.60546875,
0.2369384765625,
0.60546875,
0.25048828125,
0.295654296875,
0.25048828125
],
[
0.60205078125,
0.2369384765625,
0.6513671875,
0.2369384765625,
0.6513671875,
0.2476806640625,
0.60205078125,
0.2476806640625
],
[
0.295654296875,
0.25048828125,
0.5458984375,
0.25048828125,
0.5458984375,
0.26416015625,
0.295654296875,
0.26416015625
],
[
0.22412109375,
0.2381591796875,
0.264404296875,
0.2381591796875,
0.264404296875,
0.2493896484375,
0.22412109375,
0.2493896484375
],
[
0.2275390625,
0.251708984375,
0.258544921875,
0.251708984375,
0.258544921875,
0.26318359375,
0.2275390625,
0.26318359375
],
[
0.1275634765625,
0.26123046875,
0.183837890625,
0.26318359375,
0.1822509765625,
0.27783203125,
0.1259765625,
0.275634765625
],
[
0.21240234375,
0.26318359375,
0.27197265625,
0.265380859375,
0.27197265625,
0.27783203125,
0.21240234375,
0.275634765625
],
[
0.295654296875,
0.262451171875,
0.4375,
0.26318359375,
0.4375,
0.276611328125,
0.295654296875,
0.275634765625
],
[
0.45751953125,
0.268310546875,
0.478515625,
0.268310546875,
0.478515625,
0.2724609375,
0.45751953125,
0.2724609375
],
[
0.49365234375,
0.26318359375,
0.68994140625,
0.26318359375,
0.68994140625,
0.276611328125,
0.49365234375,
0.276611328125
],
[
0.2939453125,
0.276611328125,
0.5146484375,
0.27783203125,
0.5146484375,
0.289306640625,
0.2939453125,
0.2880859375
],
[
0.5556640625,
0.276611328125,
0.8310546875,
0.276611328125,
0.8310546875,
0.290283203125,
0.5556640625,
0.290283203125
],
[
0.52880859375,
0.283203125,
0.56103515625,
0.283203125,
0.56103515625,
0.287353515625,
0.52880859375,
0.287353515625
],
[
0.300537109375,
0.29150390625,
0.5517578125,
0.29150390625,
0.5517578125,
0.3017578125,
0.300537109375,
0.3017578125
],
[
0.198974609375,
0.27783203125,
0.2880859375,
0.27783203125,
0.2880859375,
0.289306640625,
0.198974609375,
0.289306640625
],
[
0.2283935546875,
0.292236328125,
0.2568359375,
0.292236328125,
0.2568359375,
0.300537109375,
0.2283935546875,
0.300537109375
],
[
0.1259765625,
0.300537109375,
0.183837890625,
0.303955078125,
0.1822509765625,
0.31640625,
0.124267578125,
0.3134765625
],
[
0.21240234375,
0.302734375,
0.27197265625,
0.3046875,
0.27197265625,
0.31640625,
0.21240234375,
0.314208984375
],
[
0.298828125,
0.302734375,
0.44921875,
0.303955078125,
0.44921875,
0.3154296875,
0.298828125,
0.314208984375
],
[
0.486083984375,
0.303955078125,
0.5556640625,
0.303955078125,
0.5556640625,
0.314208984375,
0.486083984375,
0.314208984375
],
[
0.5634765625,
0.3046875,
0.58447265625,
0.3046875,
0.58447265625,
0.3134765625,
0.5634765625,
0.3134765625
],
[
0.591796875,
0.303955078125,
0.74072265625,
0.303955078125,
0.74072265625,
0.314208984375,
0.591796875,
0.314208984375
],
[
0.198974609375,
0.317626953125,
0.28955078125,
0.317626953125,
0.28955078125,
0.327880859375,
0.198974609375,
0.327880859375
],
[
0.462646484375,
0.30810546875,
0.480224609375,
0.30810546875,
0.480224609375,
0.311279296875,
0.462646484375,
0.311279296875
],
[
0.28955078125,
0.320556640625,
0.297119140625,
0.320556640625,
0.297119140625,
0.325927734375,
0.28955078125,
0.325927734375
],
[
0.2939453125,
0.317626953125,
0.87744140625,
0.317626953125,
0.87744140625,
0.327880859375,
0.2939453125,
0.327880859375
],
[
0.118408203125,
0.344482421875,
0.88330078125,
0.344482421875,
0.88330078125,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.118408203125,
0.359375,
0.88330078125,
0.359375,
0.88330078125,
0.372314453125,
0.118408203125,
0.372314453125
],
[
0.1175537109375,
0.371826171875,
0.5087890625,
0.372314453125,
0.5087890625,
0.38720703125,
0.1175537109375,
0.385986328125
],
[
0.1175537109375,
0.39794921875,
0.193115234375,
0.399658203125,
0.193115234375,
0.413330078125,
0.1175537109375,
0.410888671875
],
[
0.48779296875,
0.3984375,
0.57666015625,
0.3984375,
0.57666015625,
0.413330078125,
0.48779296875,
0.413330078125
],
[
0.68408203125,
0.39794921875,
0.84228515625,
0.3984375,
0.84228515625,
0.413330078125,
0.68408203125,
0.412109375
],
[
0.5634765625,
0.410400390625,
0.61376953125,
0.412109375,
0.61279296875,
0.427001953125,
0.5615234375,
0.424560546875
],
[
0.246826171875,
0.40087890625,
0.361083984375,
0.40087890625,
0.361083984375,
0.413330078125,
0.246826171875,
0.413330078125
],
[
0.67822265625,
0.410400390625,
0.73876953125,
0.412109375,
0.73779296875,
0.427001953125,
0.6767578125,
0.424560546875
],
[
0.796875,
0.410400390625,
0.84814453125,
0.412109375,
0.84619140625,
0.427001953125,
0.7958984375,
0.424560546875
],
[
0.441650390625,
0.412109375,
0.50390625,
0.412109375,
0.50390625,
0.427001953125,
0.441650390625,
0.427001953125
],
[
0.21826171875,
0.412109375,
0.279541015625,
0.412109375,
0.279541015625,
0.427001953125,
0.21826171875,
0.427001953125
],
[
0.337646484375,
0.413330078125,
0.378662109375,
0.413330078125,
0.378662109375,
0.427001953125,
0.337646484375,
0.427001953125
],
[
0.130126953125,
0.424560546875,
0.181396484375,
0.424560546875,
0.181396484375,
0.439453125,
0.130126953125,
0.439453125
],
[
0.341796875,
0.42333984375,
0.376220703125,
0.4287109375,
0.372802734375,
0.441162109375,
0.339111328125,
0.435791015625
],
[
0.459228515625,
0.42578125,
0.48193359375,
0.42578125,
0.48193359375,
0.438232421875,
0.459228515625,
0.438232421875
],
[
0.57666015625,
0.42578125,
0.60107421875,
0.42578125,
0.60107421875,
0.438232421875,
0.57666015625,
0.438232421875
],
[
0.6943359375,
0.42578125,
0.72314453125,
0.42578125,
0.72314453125,
0.439453125,
0.6943359375,
0.439453125
],
[
0.8095703125,
0.42578125,
0.84033203125,
0.42578125,
0.84033203125,
0.440673828125,
0.8095703125,
0.440673828125
],
[
0.198974609375,
0.439453125,
0.297119140625,
0.439453125,
0.297119140625,
0.45361328125,
0.198974609375,
0.45361328125
],
[
0.231689453125,
0.427001953125,
0.264404296875,
0.427001953125,
0.264404296875,
0.439453125,
0.231689453125,
0.439453125
],
[
0.1435546875,
0.440673828125,
0.17041015625,
0.440673828125,
0.17041015625,
0.453125,
0.1435546875,
0.453125
],
[
0.414794921875,
0.440673828125,
0.52490234375,
0.438232421875,
0.52490234375,
0.453125,
0.414794921875,
0.454833984375
],
[
0.5380859375,
0.440673828125,
0.64404296875,
0.440673828125,
0.64404296875,
0.451904296875,
0.5380859375,
0.451904296875
],
[
0.6572265625,
0.440673828125,
0.759765625,
0.440673828125,
0.759765625,
0.451904296875,
0.6572265625,
0.451904296875
],
[
0.77490234375,
0.439453125,
0.873046875,
0.439453125,
0.873046875,
0.45361328125,
0.77490234375,
0.45361328125
],
[
0.646484375,
0.44482421875,
0.654296875,
0.44482421875,
0.654296875,
0.448974609375,
0.646484375,
0.448974609375
],
[
0.310546875,
0.441162109375,
0.40478515625,
0.441162109375,
0.40478515625,
0.451904296875,
0.310546875,
0.451904296875
],
[
0.1435546875,
0.453125,
0.1695556640625,
0.453125,
0.1695556640625,
0.466064453125,
0.1435546875,
0.466064453125
],
[
0.20068359375,
0.45361328125,
0.295654296875,
0.45361328125,
0.295654296875,
0.465576171875,
0.20068359375,
0.465576171875
],
[
0.4248046875,
0.45361328125,
0.51904296875,
0.45361328125,
0.51904296875,
0.465576171875,
0.4248046875,
0.465576171875
],
[
0.54248046875,
0.45361328125,
0.63818359375,
0.45361328125,
0.63818359375,
0.465576171875,
0.54248046875,
0.465576171875
],
[
0.65576171875,
0.45361328125,
0.759765625,
0.45361328125,
0.759765625,
0.465576171875,
0.65576171875,
0.465576171875
],
[
0.771484375,
0.45068359375,
0.87158203125,
0.453125,
0.87158203125,
0.46728515625,
0.771484375,
0.465576171875
],
[
0.12005615234375,
0.490478515625,
0.49365234375,
0.490478515625,
0.49365234375,
0.50439453125,
0.12005615234375,
0.50439453125
],
[
0.30908203125,
0.454833984375,
0.40478515625,
0.454833984375,
0.40478515625,
0.465576171875,
0.30908203125,
0.465576171875
],
[
0.5146484375,
0.4912109375,
0.57080078125,
0.4912109375,
0.57080078125,
0.501953125,
0.5146484375,
0.501953125
],
[
0.5693359375,
0.495361328125,
0.57666015625,
0.495361328125,
0.57666015625,
0.498779296875,
0.5693359375,
0.498779296875
],
[
0.5751953125,
0.490478515625,
0.884765625,
0.490478515625,
0.884765625,
0.50439453125,
0.5751953125,
0.50439453125
],
[
0.118408203125,
0.5048828125,
0.378662109375,
0.5048828125,
0.378662109375,
0.5185546875,
0.118408203125,
0.5185546875
],
[
0.39208984375,
0.505859375,
0.490234375,
0.505859375,
0.490234375,
0.51708984375,
0.39208984375,
0.51708984375
],
[
0.51318359375,
0.5029296875,
0.884765625,
0.5048828125,
0.884765625,
0.51953125,
0.51318359375,
0.51708984375
],
[
0.1175537109375,
0.51708984375,
0.489501953125,
0.5185546875,
0.489501953125,
0.533203125,
0.1175537109375,
0.5322265625
],
[
0.51318359375,
0.5185546875,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.51318359375,
0.5322265625
],
[
0.1175537109375,
0.5322265625,
0.489501953125,
0.533203125,
0.489501953125,
0.54736328125,
0.1175537109375,
0.54638671875
],
[
0.5146484375,
0.53369140625,
0.88330078125,
0.53369140625,
0.88330078125,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.118408203125,
0.54736328125,
0.3525390625,
0.54736328125,
0.3525390625,
0.56103515625,
0.118408203125,
0.56103515625
],
[
0.51318359375,
0.54736328125,
0.884765625,
0.54736328125,
0.884765625,
0.56103515625,
0.51318359375,
0.56103515625
],
[
0.51123046875,
0.5625,
0.6572265625,
0.56005859375,
0.6572265625,
0.57470703125,
0.51123046875,
0.57666015625
],
[
0.12005615234375,
0.57373046875,
0.489501953125,
0.57373046875,
0.489501953125,
0.58740234375,
0.12005615234375,
0.58740234375
],
[
0.52880859375,
0.57666015625,
0.884765625,
0.57666015625,
0.884765625,
0.59033203125,
0.52880859375,
0.59033203125
],
[
0.118408203125,
0.5888671875,
0.490234375,
0.5888671875,
0.490234375,
0.6025390625,
0.118408203125,
0.6025390625
],
[
0.509765625,
0.5888671875,
0.78515625,
0.59130859375,
0.78515625,
0.60498046875,
0.509765625,
0.6025390625
],
[
0.118408203125,
0.6015625,
0.489501953125,
0.6025390625,
0.489501953125,
0.6162109375,
0.118408203125,
0.615234375
],
[
0.118408203125,
0.6162109375,
0.489501953125,
0.61767578125,
0.489501953125,
0.63134765625,
0.118408203125,
0.6298828125
],
[
0.5322265625,
0.61767578125,
0.884765625,
0.61767578125,
0.884765625,
0.63134765625,
0.5322265625,
0.63134765625
],
[
0.118408203125,
0.6318359375,
0.489501953125,
0.6318359375,
0.489501953125,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.54736328125,
0.63134765625,
0.884765625,
0.63134765625,
0.884765625,
0.64453125,
0.54736328125,
0.64453125
],
[
0.12005615234375,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.65771484375,
0.12005615234375,
0.65771484375
],
[
0.54736328125,
0.64453125,
0.884765625,
0.64453125,
0.884765625,
0.65771484375,
0.54736328125,
0.65771484375
],
[
0.118408203125,
0.66015625,
0.489501953125,
0.66015625,
0.489501953125,
0.67333984375,
0.118408203125,
0.67333984375
],
[
0.54736328125,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.54736328125,
0.6728515625
],
[
0.12005615234375,
0.6748046875,
0.483642578125,
0.6748046875,
0.483642578125,
0.68505859375,
0.12005615234375,
0.68505859375
],
[
0.54736328125,
0.67333984375,
0.884765625,
0.67333984375,
0.884765625,
0.68701171875,
0.54736328125,
0.68701171875
],
[
0.1175537109375,
0.6884765625,
0.489501953125,
0.6884765625,
0.489501953125,
0.70166015625,
0.1175537109375,
0.70166015625
],
[
0.5458984375,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.5458984375,
0.69970703125
],
[
0.12005615234375,
0.70361328125,
0.48779296875,
0.70361328125,
0.48779296875,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.5439453125,
0.69970703125,
0.85205078125,
0.70166015625,
0.85205078125,
0.71533203125,
0.5439453125,
0.71337890625
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.728515625,
0.12005615234375,
0.728515625
],
[
0.118408203125,
0.73095703125,
0.1695556640625,
0.73291015625,
0.16796875,
0.74462890625,
0.1175537109375,
0.7421875
],
[
0.5322265625,
0.7255859375,
0.884765625,
0.72802734375,
0.884765625,
0.7421875,
0.5322265625,
0.74072265625
],
[
0.54833984375,
0.74365234375,
0.88330078125,
0.74365234375,
0.88330078125,
0.75439453125,
0.54833984375,
0.75439453125
],
[
0.197265625,
0.76025390625,
0.408935546875,
0.76025390625,
0.408935546875,
0.77490234375,
0.197265625,
0.77490234375
],
[
0.5458984375,
0.75732421875,
0.884765625,
0.75732421875,
0.884765625,
0.77099609375,
0.5458984375,
0.77099609375
],
[
0.54736328125,
0.77099609375,
0.884765625,
0.77099609375,
0.884765625,
0.78369140625,
0.54736328125,
0.78369140625
],
[
0.1175537109375,
0.7822265625,
0.490234375,
0.7822265625,
0.490234375,
0.7958984375,
0.1175537109375,
0.7958984375
],
[
0.5458984375,
0.783203125,
0.884765625,
0.78369140625,
0.884765625,
0.798828125,
0.5458984375,
0.79736328125
],
[
0.118408203125,
0.79638671875,
0.490234375,
0.79638671875,
0.490234375,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.54736328125,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.54736328125,
0.8125
],
[
0.12176513671875,
0.8125,
0.1822509765625,
0.8125,
0.1822509765625,
0.822265625,
0.12176513671875,
0.822265625
],
[
0.19482421875,
0.81005859375,
0.489501953125,
0.81005859375,
0.489501953125,
0.82373046875,
0.19482421875,
0.82373046875
],
[
0.54736328125,
0.8134765625,
0.84033203125,
0.8134765625,
0.84033203125,
0.82666015625,
0.54736328125,
0.82666015625
],
[
0.118408203125,
0.82373046875,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.1175537109375,
0.83837890625,
0.48779296875,
0.83837890625,
0.48779296875,
0.8515625,
0.1175537109375,
0.8515625
],
[
0.53076171875,
0.8359375,
0.88330078125,
0.83740234375,
0.88330078125,
0.8515625,
0.53076171875,
0.85107421875
],
[
0.12005615234375,
0.85107421875,
0.240966796875,
0.85400390625,
0.240966796875,
0.86767578125,
0.12005615234375,
0.8642578125
],
[
0.2333984375,
0.8515625,
0.489501953125,
0.8515625,
0.489501953125,
0.865234375,
0.2333984375,
0.865234375
],
[
0.51318359375,
0.8515625,
0.88671875,
0.8525390625,
0.88671875,
0.8662109375,
0.51318359375,
0.865234375
],
[
0.1175537109375,
0.8662109375,
0.490234375,
0.865234375,
0.490234375,
0.8798828125,
0.1175537109375,
0.88134765625
],
[
0.51318359375,
0.8681640625,
0.8818359375,
0.8681640625,
0.8818359375,
0.87890625,
0.51318359375,
0.87890625
],
[
0.118408203125,
0.88134765625,
0.490234375,
0.88134765625,
0.490234375,
0.89453125,
0.118408203125,
0.89453125
],
[
0.51123046875,
0.8798828125,
0.88330078125,
0.88134765625,
0.88330078125,
0.8955078125,
0.51123046875,
0.89453125
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Table 1:Orthographic and Romanized representations (abbreviated as Ortho and Roman) of a sample",
"sentenceacross CJkVlanguages where colored segments denote the corresponding semantic words",
"defined in Section 5. ()denotes the specific name of Romanization system of the respectivelanguage.",
"The first sentence for each language denotes the currently preprocessed XTREME benchmark dataset.",
"Language",
"Input Type",
"Sample input sentence",
"He was a scholar in Metaphysical Literature",
"Theology and Classical",
"Ortho",
"EN",
"sciences",
"Ortho",
"他是形而上学文学、神学和古典科学方面的学者。",
"ZH (source)",
"Ortho (seg)",
"他/是/形而上学文学Ⅱ、",
"神学和儿古典Ⅱ科学Ⅱ方面//的1Ⅱ学者儿",
"Roman (Pinyin)",
"tall shil/xingershangxue l/wenxue l、shenxue //hel/gudian//kexuellfangm",
"xuezhe",
"Ortho",
"Ong la mot hoc gia ve Van hoc",
"Sieu hinh",
"Than hoc",
"Khoa hoc",
"codien",
"VI (target)",
"ta",
"Ongllalmot//hocgia /l vellVanhoc //Sieuhinh ll,llThanhoc",
"l vallKhoa hoc Co dien ",
"/va llKhoa hoc /Codien",
"Ortho (seg)",
"Ong //lal/mot//hoc gia // velVanhoc //sieuhinhll,",
"Than hoc",
"彼注形而上学文学、神学、古典科学の学者。",
"Roman",
"Ortho",
"JA (target)",
"Ortho (seg)",
"彼/注Ⅱ形而上学/文学",
"",
"古典科学学者。",
"kare //ha //keijiue gaku //bungaku/l,l",
"/lkon/kagaku//gakuha/dhilta",
"oaku",
"",
"Roman (Romaji)",
"Ortho",
"KO (target)",
"Ortho (seg)",
"",
"/",
"",
"",
"Roman (Romaja)",
"",
"",
"eu lneun//hyeongisanghak //munhak /l,//sinhak /mit /i gojeon gwahak//uihakja//i//eot// seupnida//.",
"Table 2: Preliminary study on the impact of source language (ZH) on target JKV languages. Standard",
"deviation denotes the performance variation across 3 target JKV languages. Performance of individual",
"JKV language is reported in Table 7 in the Appendix.",
"Source",
"Token-level",
"QuestionAnswering",
"PANX",
"Sentence-level",
"XQuAD",
"MLQA",
"UDPOS",
"PAWSX",
"XNLI",
"Metric",
"Acc",
"F1",
"F1",
"EM",
"EM",
"73.53 ± 6.86",
"Acc",
"EN",
"40.09±10.32",
"30.36±19.09",
"28.38±18.00",
"22.11 ± 8.76",
"",
"68.35±2.05",
"ZH",
"79.61 ± 5.24",
"47.82±1.98",
"37.29±9.18",
"33.95±14.73",
"29.95±5.88",
"quires additional grammatical structural knowl-",
"70.39±1.60",
"2020)，",
"",
"We fine-tune a backbone XLM-R-base",
"edge beyond overall semantics.",
"Representa-",
"PLM (Conneau et al.,2020)with downstream task",
"tive tasks include Part-of-speech (POs)Tagging",
"objectives for each source-target JKv language",
"through UDPOS dataset and Named Entity Recog-",
"pair and report average performance together with",
"nition (NER) via PANX dataset.",
"standard deviation across the three target lan-",
" guages JA, KOo, VI.",
"Question Answering Task (XQuAD, MLQA)",
"Overall, based on our preliminary studyreported",
"extracts the answer spans existent in the given",
"in Table 2, we observe the following:",
"Context when provided with input corresponding",
"Question. The task assumes the answers are ex-",
"·Average zero-shot cross-lingual transfer per-",
"istent in the given Context; therefore, within the",
"formancefortargetlanguages across alltasks",
"scopeofourstudy,anysampleswithoutexistent",
"increases significantly when ZH is leveraged",
"answers in the contexts are removed from both",
"as the source language. On UDPOS task,",
"training and evaluation. Unlike (Hu et al..2020)",
"the average performance of JKV languages",
"we only report Exact Match (EM) for QA task since",
"reaches F1 of 47.82 as compared to 40.09",
"itis considered a stricter evaluation measure and",
"when EN is considered source language.",
"requires models toextract truly meaningful an",
"swers.",
".When ZH is leveraged as source language in-",
"steadofEN,theperformancestandarddevia-",
"4.PreliminaryStudy",
"tions across target JKV languages diminished",
"(19.09forEN source vs9.18 whenZHsource",
"We conduct a preliminary study to examine the ef-",
"on PANX task), implying that appropriately se-",
"fects of source language selection on target lan-",
"lected source languages can reduce the per-",
"guages.",
"In our case, we compare using EN",
"formance gap across target languages.",
"against ZH as the source language to transfer",
"knowledge towards other languages among the",
"Our observations are consistent with prior study",
"CJKVgroup.",
"ZH is considered the common",
"(Fujinuma et al., 2022; Yang et al., 2022) demon-",
"source language for CJKV due to its historical",
"stratingthatPLMs when trained on onlyEN can",
"longevity and significant impact on the develop-",
"be biased towards Latin-based languages, re-",
"ment of the remaining languages (Li, 2019; Hickey,",
"sulting in poor performance for non-Latin based"
] |
{
"arxiv_id": "2404.12618",
"language": "en",
"page_idx": 2
}
|
||
2404.12618-en-4
|
2404.12618-en-4.jpg
|
Table 4: Details of CORI benchmark dataset. $(x)$ and $(J)$ denote unavailable and available data existent in XTREME benchmark respectively. MT, SEG, ROM correspond to Machine Translation, Presegmentation, Romanization processing respectively as described in Section 5. Y, N denote if the corresponding preprocessing is conducted for the dataset or not.
<table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th rowspan="2">$\overline{MT}$</th>
<th rowspan="2">$SEG$</th>
<th>ROM</th>
<th colspan="2">$ZH$</th>
<th rowspan="2">$JA$ Test</th>
<th rowspan="2">KO Test</th>
<th rowspan="2">$VI$ Test</th>
</tr>
<tr>
<th> </th>
<th> </th>
<th> </th>
<th>Train</th>
<th>Dev</th>
</tr>
<tr>
<td rowspan="2">Sent-level</td>
<td>$PAWSX$</td>
<td>$\overline{Y}$</td>
<td>$\overline{Y}$</td>
<td>$\overline{Y}$</td>
<td>$49.4k$ $(X)$</td>
<td>$2k$</td>
<td>$2k$ $:(X)$</td>
<td>$2k$</td>
<td>$2k$ $(X)$</td>
</tr>
<tr>
<td>$XNLI$</td>
<td>$Y$</td>
<td>$Y$</td>
<td>$Y$</td>
<td>392.7k</td>
<td>$2.49k$</td>
<td>$5.01k$ $(x)$</td>
<td>$5.01k$ $(X)$</td>
<td>$5.01k$</td>
</tr>
<tr>
<td rowspan="2">Token-level</td>
<td>UDPOS</td>
<td>$\overline{N}$</td>
<td>$\overline{N}$</td>
<td>$\overline{Y}$</td>
<td>$10k$</td>
<td>$1.6k$</td>
<td>$2.5k$</td>
<td>$4.7k(V)$</td>
<td>$0.8k$</td>
</tr>
<tr>
<td>$PANX$</td>
<td>$N$</td>
<td>$Y$</td>
<td>$Y$</td>
<td>$20k$</td>
<td>$10k$ $V$ 7 </td>
<td>$10k$</td>
<td>$10k$</td>
<td>$10k$</td>
</tr>
<tr>
<td rowspan="2">Question Answering</td>
<td>XQuAD</td>
<td>$\overline{Y}$</td>
<td>$\overline{Y}$</td>
<td>$\overline{Y}$</td>
<td>$80.1k$ $(X)$</td>
<td>$8.87k$ $(X)$</td>
<td>$1.19k$ $(x)$</td>
<td>$1.19k$ $(X)$</td>
<td>$1.19k$</td>
</tr>
<tr>
<td>MLQA</td>
<td>$Y$</td>
<td>$Y$</td>
<td>$Y$</td>
<td>80.1k $(X)$</td>
<td>8.87k $(X)$</td>
<td>$11.24k(x)$</td>
<td>$11.24k$ $(X)$</td>
<td>$11.24k$</td>
</tr>
</tbody>
</table>
procedures, our CORI dataset provides more con-
sistent word-level annotations across CJKV.
Language Availability. Despite the multilingual coverage of XTREME, the language support for each task is not equal. For instance, while JA has train, dev, and test sets for token-level partof-speech (POS) tasks, there exists no JA test set for sentence-level PAWSX task evaluation. Therefore, the evaluation of existing works does not fully facilitate deep understanding of source and target language connections.
Orthographic Contact Limitation. Prior works assume language contact is limited to orthographic textual scripts (Zheng et al., 2021; Yang et al., 2022). However, languages can interact with one another via other language modalities. In fact, as observed in Table 1, while ZH and JA share orthographic contacts (shared 形而上学 word in textual scripts ( meaning Metaphysical)), ZH and VI contain subtle phonemic contacts captured via Romanization such as gǔdiǎn (ZH) vs Cô điên (VI), denoting "Classical"(EN) meaning.
Our presented CJKV Benchmark with Romanization Integration, namely CORI, aims to address the aforementioned shortcomings and allows for in-depth study on the connections of CJKV languages across multi-level understanding tasks. CORI contains 6 sub-datasets across different NLU tasks mentioned in Section 3 in which training and validation sets for ZH (source language) and testing set for 3 target JKV languages are provided. Additional detailed summary statistics regarding this dataset are provided in Table 4.
# 5.1. Language Availability
Pre-segmentation. The important shared characteristics of segmentations among CJKV languages are inherently ignored and can be lost during multilingual pre-processing and translations. For consistency purposes, in our work, we define the terms word and token across CJKV languages and cohere to these definitions throughout the manuscript. For CJK, we define each character as a token and the combination of tokens after segmentation as a word. For Vl, we refer to whitespace-separated terms as tokens and the combination of tokens after segmentation as a word. Each word carries a complete semantic meaning and can be formed by a single token or multiple tokens.
For CJK, there exist no consistent rules of using whitespace delimiters to differentiate tokens from words. On the other hand, Vl is written in Latin script but still whitespace does not separate meaningful units of the sentence as it does in EN. For all of our languages of interest (CJKV), instead meaningful word units are formed by combining individual tokens (Choi et al., 2009; Nguyen and Nguyen, 2020; Sun et al., 2021)
Additionally, as entries from bilingual dictionaries are stored in terms of semantic words, presegmented words allow for efficient lookups when generating code-switched language. The issue of pre-segmentation leads to inconsistent annotations across languages. Specifically, within the XTREME benchmark, on PANX datasets for NER task, annotations for KO are generated at the word level (리그 meaning League) while JA labels are generated at the token level (日 as a part of 日産word denoting Nissan meaning), as demonstrated in Figure 2. Through additional pre-segmentation
As token-level data for CJKV languages are all provided in XTREME benchmark datasets, we focus on generating translations for sentence-level and question answering tasks. Since these two tasks do not suffer dramatically from label projection challenges as compared to token-level tasks (Agić et al., 2016), we adopt the off-the-shelf Google API Machine Translation (MT) Tool2 to generate translations for target CJKV languages. The objective is two-fold (1) MT tool provides an automatic and human-free approach that can easily scale and be applied towards other datasets to generate CJKV equivalents, (2) Off-the-shelf MT tools decouple MT errors from cross-lingual transfer learning, allowing for direct evaluation of the studied crosslingual methods.
$^2$https://pypi.org/project/googletrans-py/
|
[
[
0.118408203125,
0.07244873046875,
0.884765625,
0.07244873046875,
0.884765625,
0.08612060546875,
0.118408203125,
0.08612060546875
],
[
0.118408203125,
0.0872802734375,
0.88671875,
0.0872802734375,
0.88671875,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.118408203125,
0.1015625,
0.88330078125,
0.1015625,
0.88330078125,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.1175537109375,
0.11639404296875,
0.55908203125,
0.11517333984375,
0.55908203125,
0.12890625,
0.1175537109375,
0.1300048828125
],
[
0.72119140625,
0.141357421875,
0.748046875,
0.141357421875,
0.748046875,
0.155029296875,
0.72119140625,
0.155029296875
],
[
0.80029296875,
0.141357421875,
0.82373046875,
0.141357421875,
0.82373046875,
0.155029296875,
0.80029296875,
0.155029296875
],
[
0.382080078125,
0.142578125,
0.418212890625,
0.142578125,
0.418212890625,
0.15380859375,
0.382080078125,
0.15380859375
],
[
0.345947265625,
0.142578125,
0.372802734375,
0.142578125,
0.372802734375,
0.15380859375,
0.345947265625,
0.15380859375
],
[
0.42822265625,
0.142578125,
0.468505859375,
0.142578125,
0.468505859375,
0.15380859375,
0.42822265625,
0.15380859375
],
[
0.54150390625,
0.142578125,
0.568359375,
0.142578125,
0.568359375,
0.155029296875,
0.54150390625,
0.155029296875
],
[
0.6337890625,
0.142578125,
0.66162109375,
0.142578125,
0.66162109375,
0.155029296875,
0.6337890625,
0.155029296875
],
[
0.478515625,
0.1519775390625,
0.5146484375,
0.15380859375,
0.51318359375,
0.166259765625,
0.476806640625,
0.1644287109375
],
[
0.56103515625,
0.15380859375,
0.591796875,
0.15380859375,
0.591796875,
0.166259765625,
0.56103515625,
0.166259765625
],
[
0.634765625,
0.15380859375,
0.6689453125,
0.15380859375,
0.6689453125,
0.166259765625,
0.634765625,
0.166259765625
],
[
0.72119140625,
0.15380859375,
0.75244140625,
0.15380859375,
0.75244140625,
0.166259765625,
0.72119140625,
0.166259765625
],
[
0.8017578125,
0.15380859375,
0.8330078125,
0.15380859375,
0.8330078125,
0.166259765625,
0.8017578125,
0.166259765625
],
[
0.279541015625,
0.1650390625,
0.334228515625,
0.1650390625,
0.334228515625,
0.179931640625,
0.279541015625,
0.179931640625
],
[
0.476806640625,
0.1644287109375,
0.54150390625,
0.166259765625,
0.5400390625,
0.18115234375,
0.47607421875,
0.1787109375
],
[
0.634765625,
0.1650390625,
0.67822265625,
0.16748046875,
0.6767578125,
0.179931640625,
0.6337890625,
0.1776123046875
],
[
0.80029296875,
0.1650390625,
0.84228515625,
0.16748046875,
0.84033203125,
0.179931640625,
0.79833984375,
0.1776123046875
],
[
0.3837890625,
0.16748046875,
0.397216796875,
0.16748046875,
0.397216796875,
0.177001953125,
0.3837890625,
0.177001953125
],
[
0.344970703125,
0.16748046875,
0.359375,
0.16748046875,
0.359375,
0.1776123046875,
0.344970703125,
0.1776123046875
],
[
0.5615234375,
0.16748046875,
0.60546875,
0.16748046875,
0.60546875,
0.1787109375,
0.5615234375,
0.1787109375
],
[
0.72119140625,
0.16748046875,
0.765625,
0.16748046875,
0.765625,
0.1787109375,
0.72119140625,
0.1787109375
],
[
0.424072265625,
0.1728515625,
0.4375,
0.1644287109375,
0.44921875,
0.1728515625,
0.435791015625,
0.18115234375
],
[
0.1259765625,
0.1728515625,
0.197265625,
0.1728515625,
0.197265625,
0.18408203125,
0.1259765625,
0.18408203125
],
[
0.3837890625,
0.1787109375,
0.397216796875,
0.1787109375,
0.397216796875,
0.189453125,
0.3837890625,
0.189453125
],
[
0.429931640625,
0.177001953125,
0.443359375,
0.177001953125,
0.443359375,
0.189453125,
0.429931640625,
0.189453125
],
[
0.476806640625,
0.1787109375,
0.54833984375,
0.1787109375,
0.54833984375,
0.190673828125,
0.476806640625,
0.190673828125
],
[
0.5615234375,
0.1787109375,
0.62646484375,
0.1787109375,
0.62646484375,
0.190673828125,
0.5615234375,
0.190673828125
],
[
0.6337890625,
0.1787109375,
0.69775390625,
0.1787109375,
0.69775390625,
0.190673828125,
0.6337890625,
0.190673828125
],
[
0.72119140625,
0.1787109375,
0.78076171875,
0.1787109375,
0.78076171875,
0.190673828125,
0.72119140625,
0.190673828125
],
[
0.80029296875,
0.1787109375,
0.86572265625,
0.1787109375,
0.86572265625,
0.190673828125,
0.80029296875,
0.190673828125
],
[
0.340087890625,
0.182861328125,
0.353515625,
0.174560546875,
0.365234375,
0.182861328125,
0.3525390625,
0.191162109375
],
[
0.280517578125,
0.179931640625,
0.318115234375,
0.179931640625,
0.318115234375,
0.189453125,
0.280517578125,
0.189453125
],
[
0.47607421875,
0.189453125,
0.52880859375,
0.191162109375,
0.52880859375,
0.203125,
0.47607421875,
0.20068359375
],
[
0.56103515625,
0.189453125,
0.6171875,
0.191162109375,
0.6171875,
0.203125,
0.56103515625,
0.20068359375
],
[
0.6337890625,
0.189453125,
0.69189453125,
0.191162109375,
0.69189453125,
0.203125,
0.6337890625,
0.20068359375
],
[
0.7197265625,
0.189453125,
0.7734375,
0.191162109375,
0.7734375,
0.203125,
0.7197265625,
0.20068359375
],
[
0.79833984375,
0.189453125,
0.8564453125,
0.191162109375,
0.8564453125,
0.203125,
0.79833984375,
0.20068359375
],
[
0.3837890625,
0.190673828125,
0.39794921875,
0.190673828125,
0.39794921875,
0.20068359375,
0.3837890625,
0.20068359375
],
[
0.279541015625,
0.190673828125,
0.3359375,
0.190673828125,
0.3359375,
0.201904296875,
0.279541015625,
0.201904296875
],
[
0.343505859375,
0.191162109375,
0.361083984375,
0.191162109375,
0.361083984375,
0.20068359375,
0.343505859375,
0.20068359375
],
[
0.424072265625,
0.193603515625,
0.435791015625,
0.1859130859375,
0.44580078125,
0.19482421875,
0.43408203125,
0.201904296875
],
[
0.47607421875,
0.1995849609375,
0.5322265625,
0.201904296875,
0.53076171875,
0.2161865234375,
0.474365234375,
0.21435546875
],
[
0.634765625,
0.1995849609375,
0.68994140625,
0.201904296875,
0.6884765625,
0.2161865234375,
0.6337890625,
0.21435546875
],
[
0.7197265625,
0.1995849609375,
0.77490234375,
0.201904296875,
0.7734375,
0.2161865234375,
0.71875,
0.21435546875
],
[
0.80029296875,
0.1995849609375,
0.85400390625,
0.201904296875,
0.85205078125,
0.2161865234375,
0.79833984375,
0.21435546875
],
[
0.27783203125,
0.20068359375,
0.32421875,
0.20068359375,
0.32421875,
0.215576171875,
0.27783203125,
0.215576171875
],
[
0.1275634765625,
0.196533203125,
0.20654296875,
0.196533203125,
0.20654296875,
0.2078857421875,
0.1275634765625,
0.2078857421875
],
[
0.343505859375,
0.20068359375,
0.36279296875,
0.20068359375,
0.36279296875,
0.21435546875,
0.343505859375,
0.21435546875
],
[
0.3837890625,
0.201904296875,
0.397216796875,
0.201904296875,
0.397216796875,
0.2120361328125,
0.3837890625,
0.2120361328125
],
[
0.4248046875,
0.203125,
0.438232421875,
0.198974609375,
0.44580078125,
0.210205078125,
0.432373046875,
0.21435546875
],
[
0.5615234375,
0.20068359375,
0.61376953125,
0.20068359375,
0.61376953125,
0.215576171875,
0.5615234375,
0.215576171875
],
[
0.279541015625,
0.21142578125,
0.33154296875,
0.213134765625,
0.330078125,
0.22802734375,
0.27783203125,
0.2257080078125
],
[
0.435791015625,
0.21142578125,
0.447509765625,
0.218505859375,
0.435791015625,
0.2286376953125,
0.424072265625,
0.221435546875
],
[
0.47607421875,
0.2120361328125,
0.5400390625,
0.21435546875,
0.5380859375,
0.2286376953125,
0.474365234375,
0.226806640625
],
[
0.56103515625,
0.2120361328125,
0.623046875,
0.21435546875,
0.62109375,
0.2286376953125,
0.55908203125,
0.226806640625
],
[
0.63623046875,
0.213134765625,
0.6962890625,
0.215576171875,
0.6962890625,
0.22802734375,
0.63623046875,
0.2257080078125
],
[
0.8017578125,
0.2120361328125,
0.8671875,
0.21435546875,
0.86572265625,
0.2286376953125,
0.80029296875,
0.226806640625
],
[
0.72314453125,
0.215576171875,
0.779296875,
0.215576171875,
0.779296875,
0.226806640625,
0.72314453125,
0.226806640625
],
[
0.349365234375,
0.2161865234375,
0.35693359375,
0.2161865234375,
0.35693359375,
0.223876953125,
0.349365234375,
0.223876953125
],
[
0.3837890625,
0.21435546875,
0.397216796875,
0.21435546875,
0.397216796875,
0.2257080078125,
0.3837890625,
0.2257080078125
],
[
0.79833984375,
0.223876953125,
0.875,
0.2257080078125,
0.875,
0.240478515625,
0.79833984375,
0.2381591796875
],
[
0.35107421875,
0.2244873046875,
0.361083984375,
0.232177734375,
0.3525390625,
0.2369384765625,
0.343505859375,
0.23095703125
],
[
0.474365234375,
0.2244873046875,
0.5400390625,
0.2244873046875,
0.5400390625,
0.2392578125,
0.474365234375,
0.2392578125
],
[
0.3837890625,
0.2257080078125,
0.397216796875,
0.2257080078125,
0.397216796875,
0.2369384765625,
0.3837890625,
0.2369384765625
],
[
0.124267578125,
0.218505859375,
0.267822265625,
0.2203369140625,
0.267822265625,
0.2340087890625,
0.124267578125,
0.232177734375
],
[
0.279541015625,
0.226806640625,
0.32421875,
0.226806640625,
0.32421875,
0.2381591796875,
0.279541015625,
0.2381591796875
],
[
0.429931640625,
0.2257080078125,
0.443359375,
0.2257080078125,
0.443359375,
0.2369384765625,
0.429931640625,
0.2369384765625
],
[
0.5634765625,
0.226806640625,
0.62109375,
0.226806640625,
0.62109375,
0.2381591796875,
0.5634765625,
0.2381591796875
],
[
0.63818359375,
0.226806640625,
0.70947265625,
0.226806640625,
0.70947265625,
0.2381591796875,
0.63818359375,
0.2381591796875
],
[
0.72119140625,
0.226806640625,
0.78857421875,
0.226806640625,
0.78857421875,
0.2381591796875,
0.72119140625,
0.2381591796875
],
[
0.118408203125,
0.2587890625,
0.313232421875,
0.2587890625,
0.313232421875,
0.2724609375,
0.118408203125,
0.2724609375
],
[
0.307373046875,
0.2587890625,
0.489501953125,
0.2587890625,
0.489501953125,
0.2724609375,
0.307373046875,
0.2724609375
],
[
0.51318359375,
0.2587890625,
0.884765625,
0.2587890625,
0.884765625,
0.2724609375,
0.51318359375,
0.2724609375
],
[
0.118408203125,
0.2724609375,
0.489501953125,
0.2724609375,
0.489501953125,
0.2861328125,
0.118408203125,
0.2861328125
],
[
0.51318359375,
0.2724609375,
0.84033203125,
0.2724609375,
0.84033203125,
0.2861328125,
0.51318359375,
0.2861328125
],
[
0.1175537109375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.1175537109375,
0.2998046875
],
[
0.1175537109375,
0.300537109375,
0.490234375,
0.302734375,
0.490234375,
0.31640625,
0.1175537109375,
0.314208984375
],
[
0.5146484375,
0.302734375,
0.7958984375,
0.302734375,
0.7958984375,
0.31640625,
0.5146484375,
0.31640625
],
[
0.79443359375,
0.303955078125,
0.88330078125,
0.303955078125,
0.88330078125,
0.3154296875,
0.79443359375,
0.3154296875
],
[
0.12005615234375,
0.31640625,
0.489501953125,
0.31640625,
0.489501953125,
0.330078125,
0.12005615234375,
0.330078125
],
[
0.51318359375,
0.317626953125,
0.88671875,
0.317626953125,
0.88671875,
0.330810546875,
0.51318359375,
0.330810546875
],
[
0.118408203125,
0.330078125,
0.490234375,
0.330078125,
0.490234375,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.5146484375,
0.33203125,
0.88330078125,
0.33203125,
0.88330078125,
0.345703125,
0.5146484375,
0.345703125
],
[
0.12005615234375,
0.345703125,
0.486083984375,
0.345703125,
0.486083984375,
0.355712890625,
0.12005615234375,
0.355712890625
],
[
0.51318359375,
0.3466796875,
0.8818359375,
0.3466796875,
0.8818359375,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.118408203125,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.372314453125,
0.118408203125,
0.372314453125
],
[
0.51123046875,
0.359375,
0.880859375,
0.359375,
0.880859375,
0.372314453125,
0.51123046875,
0.372314453125
],
[
0.118408203125,
0.372314453125,
0.28955078125,
0.372314453125,
0.28955078125,
0.385986328125,
0.118408203125,
0.385986328125
],
[
0.51123046875,
0.37353515625,
0.884765625,
0.372314453125,
0.884765625,
0.38720703125,
0.51123046875,
0.388427734375
],
[
0.51123046875,
0.388427734375,
0.88330078125,
0.38720703125,
0.88330078125,
0.402099609375,
0.51123046875,
0.402587890625
],
[
0.118408203125,
0.39794921875,
0.28369140625,
0.39794921875,
0.28369140625,
0.410888671875,
0.118408203125,
0.410888671875
],
[
0.273681640625,
0.3955078125,
0.490234375,
0.396728515625,
0.490234375,
0.410888671875,
0.273681640625,
0.410400390625
],
[
0.51318359375,
0.402587890625,
0.88330078125,
0.402587890625,
0.88330078125,
0.416259765625,
0.51318359375,
0.416259765625
],
[
0.12005615234375,
0.412109375,
0.491943359375,
0.412109375,
0.491943359375,
0.42578125,
0.12005615234375,
0.42578125
],
[
0.51318359375,
0.418701171875,
0.880859375,
0.418701171875,
0.880859375,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.118408203125,
0.427001953125,
0.490234375,
0.42578125,
0.490234375,
0.439453125,
0.118408203125,
0.440673828125
],
[
0.51123046875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.51123046875,
0.44482421875
],
[
0.118408203125,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.45361328125,
0.118408203125,
0.45361328125
],
[
0.5146484375,
0.4453125,
0.77099609375,
0.4453125,
0.77099609375,
0.458984375,
0.5146484375,
0.458984375
],
[
0.118408203125,
0.454833984375,
0.490234375,
0.454833984375,
0.490234375,
0.468505859375,
0.118408203125,
0.468505859375
],
[
0.53076171875,
0.460205078125,
0.88671875,
0.460205078125,
0.88671875,
0.473876953125,
0.53076171875,
0.473876953125
],
[
0.118408203125,
0.468505859375,
0.490234375,
0.468505859375,
0.490234375,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51318359375,
0.474365234375,
0.884765625,
0.474365234375,
0.884765625,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.12005615234375,
0.482666015625,
0.490234375,
0.482666015625,
0.490234375,
0.496337890625,
0.12005615234375,
0.496337890625
],
[
0.51318359375,
0.4892578125,
0.884765625,
0.4892578125,
0.884765625,
0.5029296875,
0.51318359375,
0.5029296875
],
[
0.12005615234375,
0.49755859375,
0.490234375,
0.49755859375,
0.490234375,
0.51123046875,
0.12005615234375,
0.51123046875
],
[
0.5146484375,
0.5029296875,
0.884765625,
0.5029296875,
0.884765625,
0.5166015625,
0.5146484375,
0.5166015625
],
[
0.12176513671875,
0.51318359375,
0.489501953125,
0.51318359375,
0.489501953125,
0.52392578125,
0.12176513671875,
0.52392578125
],
[
0.51318359375,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.118408203125,
0.525390625,
0.490234375,
0.525390625,
0.490234375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.51123046875,
0.53076171875,
0.884765625,
0.5322265625,
0.884765625,
0.5458984375,
0.51123046875,
0.54443359375
],
[
0.118408203125,
0.54052734375,
0.490234375,
0.54052734375,
0.490234375,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.5146484375,
0.5458984375,
0.88330078125,
0.5458984375,
0.88330078125,
0.5556640625,
0.5146484375,
0.5556640625
],
[
0.118408203125,
0.55419921875,
0.48779296875,
0.5546875,
0.48779296875,
0.56640625,
0.118408203125,
0.5654296875
],
[
0.5146484375,
0.56005859375,
0.88330078125,
0.56005859375,
0.88330078125,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.12005615234375,
0.568359375,
0.489501953125,
0.568359375,
0.489501953125,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.5146484375,
0.57373046875,
0.88330078125,
0.57373046875,
0.88330078125,
0.58740234375,
0.5146484375,
0.58740234375
],
[
0.118408203125,
0.58203125,
0.489501953125,
0.58203125,
0.489501953125,
0.595703125,
0.118408203125,
0.595703125
],
[
0.51318359375,
0.58837890625,
0.88671875,
0.58837890625,
0.88671875,
0.6015625,
0.51318359375,
0.6015625
],
[
0.118408203125,
0.5966796875,
0.2359619140625,
0.5966796875,
0.2359619140625,
0.60986328125,
0.118408203125,
0.60986328125
],
[
0.5146484375,
0.6025390625,
0.84033203125,
0.6025390625,
0.84033203125,
0.6162109375,
0.5146484375,
0.6162109375
],
[
0.1368408203125,
0.6123046875,
0.48779296875,
0.6123046875,
0.48779296875,
0.62255859375,
0.1368408203125,
0.62255859375
],
[
0.1175537109375,
0.6259765625,
0.486083984375,
0.6259765625,
0.486083984375,
0.6357421875,
0.1175537109375,
0.6357421875
],
[
0.1175537109375,
0.6396484375,
0.491943359375,
0.6396484375,
0.491943359375,
0.65283203125,
0.1175537109375,
0.65283203125
],
[
0.5146484375,
0.6357421875,
0.755859375,
0.6357421875,
0.755859375,
0.65087890625,
0.5146484375,
0.65087890625
],
[
0.118408203125,
0.65380859375,
0.491943359375,
0.65380859375,
0.491943359375,
0.66748046875,
0.118408203125,
0.66748046875
],
[
0.51318359375,
0.65283203125,
0.88671875,
0.65380859375,
0.88671875,
0.66748046875,
0.51318359375,
0.66650390625
],
[
0.12005615234375,
0.66650390625,
0.489501953125,
0.66650390625,
0.489501953125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.51318359375,
0.66845703125,
0.884765625,
0.66845703125,
0.884765625,
0.681640625,
0.51318359375,
0.681640625
],
[
0.118408203125,
0.681640625,
0.49365234375,
0.681640625,
0.49365234375,
0.6953125,
0.118408203125,
0.6953125
],
[
0.51123046875,
0.68115234375,
0.88330078125,
0.68115234375,
0.88330078125,
0.69482421875,
0.51123046875,
0.69482421875
],
[
0.12005615234375,
0.69775390625,
0.489501953125,
0.69775390625,
0.489501953125,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.51318359375,
0.69677734375,
0.884765625,
0.69677734375,
0.884765625,
0.71044921875,
0.51318359375,
0.71044921875
],
[
0.1175537109375,
0.708984375,
0.490234375,
0.71044921875,
0.490234375,
0.724609375,
0.1175537109375,
0.7236328125
],
[
0.5146484375,
0.7119140625,
0.88330078125,
0.7119140625,
0.88330078125,
0.72265625,
0.5146484375,
0.72265625
],
[
0.12005615234375,
0.724609375,
0.295654296875,
0.724609375,
0.295654296875,
0.73828125,
0.12005615234375,
0.73828125
],
[
0.5146484375,
0.724609375,
0.884765625,
0.724609375,
0.884765625,
0.73828125,
0.5146484375,
0.73828125
],
[
0.1368408203125,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.7529296875,
0.1368408203125,
0.7529296875
],
[
0.5146484375,
0.7392578125,
0.884765625,
0.7392578125,
0.884765625,
0.7529296875,
0.5146484375,
0.7529296875
],
[
0.12176513671875,
0.7548828125,
0.490234375,
0.7548828125,
0.490234375,
0.765625,
0.12176513671875,
0.765625
],
[
0.515625,
0.7548828125,
0.88330078125,
0.7548828125,
0.88330078125,
0.765625,
0.515625,
0.765625
],
[
0.118408203125,
0.76708984375,
0.490234375,
0.76708984375,
0.490234375,
0.78076171875,
0.118408203125,
0.78076171875
],
[
0.51318359375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.12005615234375,
0.7822265625,
0.4072265625,
0.7822265625,
0.4072265625,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.39794921875,
0.7822265625,
0.48779296875,
0.783203125,
0.48779296875,
0.79443359375,
0.39794921875,
0.79345703125
],
[
0.51318359375,
0.7822265625,
0.884765625,
0.7822265625,
0.884765625,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.79638671875,
0.49365234375,
0.79638671875,
0.49365234375,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.51318359375,
0.79638671875,
0.88330078125,
0.79638671875,
0.88330078125,
0.81005859375,
0.51318359375,
0.81005859375
],
[
0.118408203125,
0.81005859375,
0.489501953125,
0.81005859375,
0.489501953125,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.5146484375,
0.81005859375,
0.88330078125,
0.81005859375,
0.88330078125,
0.82373046875,
0.5146484375,
0.82373046875
],
[
0.118408203125,
0.82373046875,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.83837890625,
0.489501953125,
0.83837890625,
0.489501953125,
0.8515625,
0.118408203125,
0.8515625
],
[
0.51318359375,
0.8359375,
0.88671875,
0.83837890625,
0.88671875,
0.8525390625,
0.51318359375,
0.85107421875
],
[
0.118408203125,
0.85107421875,
0.489501953125,
0.8515625,
0.489501953125,
0.8662109375,
0.118408203125,
0.865234375
],
[
0.5146484375,
0.8525390625,
0.884765625,
0.8525390625,
0.884765625,
0.8662109375,
0.5146484375,
0.8662109375
],
[
0.118408203125,
0.86767578125,
0.489501953125,
0.8662109375,
0.489501953125,
0.8798828125,
0.118408203125,
0.88134765625
],
[
0.51318359375,
0.86767578125,
0.63818359375,
0.86767578125,
0.63818359375,
0.88134765625,
0.51318359375,
0.88134765625
],
[
0.1175537109375,
0.88134765625,
0.489501953125,
0.88134765625,
0.489501953125,
0.89453125,
0.1175537109375,
0.89453125
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.53076171875,
0.8935546875,
0.79833984375,
0.8955078125,
0.79833984375,
0.91015625,
0.53076171875,
0.90771484375
]
] |
[
"Table 4: Details of coORl benchmark dataset. (x) and () denote unavailable and available data ex-",
"istent inXTREME benchmark respectively.MT, SEG, ROM correspond to Machine Translation, Pre-",
"segmentation, Romanization processing respectively as described in Section 5. Y, N denote if the corre-",
"sponding preprocessing is conducted for the dataset or not.",
"KO",
"VI",
"SEG",
"MT",
"ROM",
"ZH",
"JA",
"Train",
"Dev",
"Test",
"Test",
"Test",
"PAWSX",
"49.4k (x)",
"2k (×)",
"2k (x)",
"Y",
"Y",
"2k ()",
"2k ()",
"A",
"Sent-level",
"Y",
"Y",
"392.7k (V)",
"2.49k ()",
"5.01k (x)",
"5.01k (x)",
"5.01k ()",
"L",
"XNLI",
"10k (V)",
"1.6k ()",
"2.5k（)",
"4.7k(V)",
"0.8k(V)",
"N",
"UDPOS",
"N",
"A",
"20k(>)",
"10k(V)",
"10k ()",
"10k (M)",
"PANX",
"Token-level",
"N",
"Y",
"1",
"10k (>)",
"XQuAD",
"A",
"80.1k (x)",
"8.87k (X)",
"1.19k (×)",
"1.19k(V)",
"1.19k (x)",
"Y",
"Y",
"11.24k (V)",
"A",
"80.1k (x)",
"Y",
"QuestionAnswering",
"MLQA",
"Y",
"8.87k (x)",
"11.24 k (x)",
"11.24k (x)",
"Language Availability.",
" Despite the multilingual",
"procedures, our CORl dataset provides more con-",
"coverage of XTREME, the language support for",
"sistent word-level annotations across CJKV.",
"each task is not equal. For instance, while JA",
"has train, dev, and test sets for token-level part-",
"Orthographic Contact Limitation.",
"Priorworks",
"of-speech (POS) tasks, there exists no JA test set",
"assume language contact is limited to ortho-",
"for sentence-level PAWSX task evaluation.There-",
"graphic textual scripts (Zheng et al., 2021; Yang",
"fore,theevaluationofexistingworksdoesnotfully",
"et al., 2022). However, languages can interact with",
"facilitate deep understanding of source and target",
"one another via other language modalities. In fact.",
"language connections.",
"as observed in Table 1, while ZH and JA share",
"orthographic contacts(shared 形而上学 word in",
"Pre-segmentation.",
"The important shared char-",
"textual scripts (meaning Metaphysical), ZH and VI",
"acteristics of segmentations among CJKV lan-",
"containsubtlephonemiccontactscapturedviaRo",
"guages are inherently ignored and can be lost dur-",
"manization such as gudian (ZH) vs Co dién (Vl),",
"ing multilingual pre-processing and translations.",
"denoting\"Classical\"(EN) meaning.",
"For consistency purposes, in our work, we de-",
"Our presented CJKVBenchmark with Roman-",
"fine the terms word and token across CJKV lan-",
"ization Integration, namely CORl, aims to address",
"guages and cohere to these definitions through-",
"the aforementioned shortcomings and allows for",
"out the manuscript. For CJK, we define each",
"in-depth study on the connections of CJKV lan-",
"character as a token and the combination of to-",
"guages across multi-level understanding tasks.",
"kens after segmentation as a word. For Vl, we re-",
"CORl contains 6 sub-datasets across different",
"fer to whitespace-separated terms as tokens and",
"NLU tasks mentioned in Section 3inwhichtrain",
"the combination of tokens after segmentation as",
"ing and validation sets for ZH (source language)",
"a word. Each word carries a complete semantic",
"and testing set for 3 target JKV languages are pro-",
"meaning and can be formed by a single token or",
"vided. Additional detailed summary statistics re-",
"multiple tokens.",
"garding this dataset are provided in Table 4.",
"ForCJK,thereexistnoconsistentrulesofusing",
"whitespacedelimiterstodifferentiatetokensfrom",
"words. On the other hand, VI is written in Latin",
"5.1.Language Availability",
"script but still whitespace does not separate mean-",
"As token-level data for CJKV languages are all pro-",
"ingful units of the sentence as it does in EN. For all",
"vided in XTREME benchmark datasets, we focus",
"of our languages of interest (CJKV), instead mean-",
"on generating translations for sentence-level and",
"ingfulwordunitsareformedbycombiningindivid-",
"question answering tasks. Since these two tasks",
"ual tokens (Choi et al., 2009; Nguyen and Nguyen,",
"do not suffer dramatically from label projection",
"2020; Sun et al., 2021).",
"challenges as compared to token-level tasks (Agic",
"Additionally, as entries from bilingual dictionar-",
"et al., 2016), we adopt the off-the-shelf Google APl",
"ies are stored interms of semantic words,pre-",
"Machine Translation(MT)Too/2togeneratetrans",
"segmented words allow for efficient lookups when",
"lations for target CJKV languages. The objective",
"generating code-switched language.",
"The issue",
"is two-fold (1) MT tool provides an automatic and",
"of pre-segmentation leads to inconsistent annota-",
"human-free approach that can easily scale and be",
"tions across languages. Specifically, within the",
"applied towards other datasets to generate CJKV",
"XTREME benchmark, on PANX datasets for NER",
"equivalents, (2) Off-the-shelf MT tools decouple",
"task, annotations for KO are generated at the word",
"MT errors from cross-lingual transfer learning, al-",
"level ( meaning League) while JA labels are",
"lowing for direct evaluation of the studied cross-",
"generated at the token level （日as a part of 日",
"lingual methods.",
"word denoting Nissan meaning),as demonstrated",
"in Figure 2. Through additional pre-segmentation",
"2https://pypi.org/project/googletrans-py/"
] |
{
"arxiv_id": "2404.12618",
"language": "en",
"page_idx": 4
}
|
||
1704.06393-en-0
|
1704.06393-en-0.jpg
|
$\textbf{Neural System Combination for Machine Translation}$
$\textbf{Long Zhou}^{\dagger},\textbf{Wenpeng Hu}^{\dagger},\textbf{Jiajun Zhang}^{\dagger*}\textbf{Chengqing Zong}^{\dagger\dagger}$
$^{\dagger}$University of Chinese Academy of Sciences, Beijing, China
National Laboratory of Pattern Recognition, CASIA, Beijing, China
$^{\dagger}$CAS Center for Excellence in Brain Science and Intelligence Technology, Shanghai, China
### $\mathbf{Abstract}$
Neural machine translation (NMT) becomes a new approach to machine translation and generates much more fluent results compared to statistical machine translation (SMT). However, SMT is usually better than NMT in translation adequacy. It is therefore a promising direction to combine the advantages of both NMT and SMT. In this paper, we propose a neural system combination framework leveraging multi-source NMT, which takes as input the outputs of NMT and SMT systems and produces the final translation. Extensive experiments on the Chineseto-English translation task show that our model archives significant improvement by 5.3 BLEU points over the best single system output and 3.4 BLEU points over the state-of-the-art traditional system combination methods.
## 1 Introduction
Neural machine translation has significantly improved the quality of machine translation in recent several years (Kalchbrenner and Blunsom, 2013; Sutskever et al.,2014; Bahdanau et al., 2015; Junczys-Dowmunt et al., 2016a). Although most sentences are more fluent than translations by statistical machine translation (SMT) (Koehn et al., 2003; Chiang, 2005), NMT has a problem to address translation adequacy especially for the rare and unknown words. Additionally, it suffers from over-translation and under-translation to some extent (Tu et al., 2016). Compared to NMT, SMT, such as phrase-based machine translation (PBMT, (Koehn et al., 2003)) and hierarchical phrase-based machine translation (HPMT,
$^{*}$Corresponding author.
(Chiang, 2005)), does not need to limit the vocabulary and can guarantee translation coverage of source sentences. It is obvious that NMT and SMT have different strength and weakness. In order to take full advantages of both NMT and SMT, system combination can be a good choice.
Traditionally, system combination has been explored respectively in sentence-level, phrase-level, and word-level (Kumar and Byrne, 2004; Feng et al., 2009; Chen et al., 2009). Among them, word-level combination approaches that adopt confusion network for decoding have been quite successful (Rosti et al., 2007; Ayan et al., 2008; Freitag et al., 2014). However, these approaches are mainly designed for SMT without considering the features of NMT results. NMT opts to produce diverse words and free word order, which are quite different from SMT. And this will make it hard to construct a consistent confusion network. Furthermore, traditional system combination approaches cannot guarantee the fluency of the final translation results.
In this paper, we propose a neural system combination framework, which is adapted from the multi-source NMT model (Zoph and Knight, 2016). Different encoders are employed to model the semantics of the source language input and each best translation produced by different NMT and SMT systems. The encoders produce multiple context vector representations, from which the decoder generates the final output word by word. Since the same training data is used for NMT, SMT and neural system combination, we further design a smart strategy to simulate the real training data for neural system combination.
Specifically, we make the following contribu-
tions in this paper:
$\bullet$ We propose a neural system combination
method, which is adapted from multi-source
|
[
[
0.2283935546875,
0.0831298828125,
0.77490234375,
0.0831298828125,
0.77490234375,
0.097412109375,
0.2283935546875,
0.097412109375
],
[
0.235107421875,
0.127685546875,
0.76904296875,
0.1270751953125,
0.76904296875,
0.144287109375,
0.235107421875,
0.1455078125
],
[
0.258544921875,
0.144287109375,
0.75,
0.1455078125,
0.75,
0.163330078125,
0.258544921875,
0.162109375
],
[
0.2283935546875,
0.1644287109375,
0.779296875,
0.1644287109375,
0.779296875,
0.1787109375,
0.2283935546875,
0.1787109375
],
[
0.130126953125,
0.177001953125,
0.8759765625,
0.1787109375,
0.8759765625,
0.196533203125,
0.130126953125,
0.19482421875
],
[
0.197265625,
0.196533203125,
0.8095703125,
0.19775390625,
0.8095703125,
0.2120361328125,
0.197265625,
0.21142578125
],
[
0.26611328125,
0.26416015625,
0.343505859375,
0.26416015625,
0.343505859375,
0.279052734375,
0.26611328125,
0.279052734375
],
[
0.5146484375,
0.265380859375,
0.8818359375,
0.265380859375,
0.8818359375,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.515625,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.29443359375,
0.515625,
0.29443359375
],
[
0.14697265625,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3046875,
0.14697265625,
0.3046875
],
[
0.51318359375,
0.297607421875,
0.88330078125,
0.297607421875,
0.88330078125,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.019317626953125,
0.305908203125,
0.06634521484375,
0.305908203125,
0.0697021484375,
0.440673828125,
0.0218353271484375,
0.440673828125
],
[
0.14697265625,
0.308837890625,
0.459228515625,
0.308837890625,
0.459228515625,
0.322509765625,
0.14697265625,
0.322509765625
],
[
0.51318359375,
0.314208984375,
0.8818359375,
0.314208984375,
0.8818359375,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.14697265625,
0.32373046875,
0.459228515625,
0.32373046875,
0.459228515625,
0.33740234375,
0.14697265625,
0.33740234375
],
[
0.51318359375,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.14697265625,
0.34033203125,
0.459228515625,
0.34033203125,
0.459228515625,
0.35400390625,
0.14697265625,
0.35400390625
],
[
0.51318359375,
0.3466796875,
0.8271484375,
0.3466796875,
0.8271484375,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.14697265625,
0.355712890625,
0.459228515625,
0.355712890625,
0.459228515625,
0.369384765625,
0.14697265625,
0.369384765625
],
[
0.53076171875,
0.3623046875,
0.88330078125,
0.363525390625,
0.88330078125,
0.377685546875,
0.53076171875,
0.377197265625
],
[
0.14697265625,
0.372314453125,
0.459228515625,
0.372314453125,
0.459228515625,
0.385986328125,
0.14697265625,
0.385986328125
],
[
0.51318359375,
0.380126953125,
0.88330078125,
0.380126953125,
0.88330078125,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.1485595703125,
0.3896484375,
0.4609375,
0.3896484375,
0.4609375,
0.402587890625,
0.1485595703125,
0.402587890625
],
[
0.51123046875,
0.394287109375,
0.88330078125,
0.3955078125,
0.88330078125,
0.410400390625,
0.51123046875,
0.4091796875
],
[
0.14697265625,
0.405029296875,
0.4609375,
0.405029296875,
0.4609375,
0.418701171875,
0.14697265625,
0.418701171875
],
[
0.51318359375,
0.410400390625,
0.88330078125,
0.410888671875,
0.88330078125,
0.42578125,
0.51318359375,
0.424560546875
],
[
0.145263671875,
0.419189453125,
0.459228515625,
0.42041015625,
0.459228515625,
0.435302734375,
0.145263671875,
0.43408203125
],
[
0.5146484375,
0.427490234375,
0.884765625,
0.427490234375,
0.884765625,
0.4423828125,
0.5146484375,
0.4423828125
],
[
0.0201568603515625,
0.435302734375,
0.06378173828125,
0.43408203125,
0.06634521484375,
0.51708984375,
0.0235137939453125,
0.5185546875
],
[
0.14697265625,
0.43701171875,
0.459228515625,
0.43701171875,
0.459228515625,
0.45068359375,
0.14697265625,
0.45068359375
],
[
0.51318359375,
0.44482421875,
0.88330078125,
0.44482421875,
0.88330078125,
0.457763671875,
0.51318359375,
0.457763671875
],
[
0.14697265625,
0.453125,
0.4609375,
0.451904296875,
0.4609375,
0.466064453125,
0.14697265625,
0.46728515625
],
[
0.51318359375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.51318359375,
0.473876953125
],
[
0.14697265625,
0.468505859375,
0.45751953125,
0.468505859375,
0.45751953125,
0.482666015625,
0.14697265625,
0.482666015625
],
[
0.5146484375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.490478515625,
0.5146484375,
0.490478515625
],
[
0.14697265625,
0.485107421875,
0.45751953125,
0.485107421875,
0.45751953125,
0.498779296875,
0.14697265625,
0.498779296875
],
[
0.51318359375,
0.4912109375,
0.88330078125,
0.4921875,
0.88330078125,
0.50732421875,
0.51318359375,
0.505859375
],
[
0.1485595703125,
0.50048828125,
0.45751953125,
0.50048828125,
0.45751953125,
0.51416015625,
0.1485595703125,
0.51416015625
],
[
0.51318359375,
0.50830078125,
0.8818359375,
0.50830078125,
0.8818359375,
0.521484375,
0.51318359375,
0.521484375
],
[
0.0251922607421875,
0.51953125,
0.060455322265625,
0.51953125,
0.060455322265625,
0.72802734375,
0.0251922607421875,
0.72802734375
],
[
0.14697265625,
0.51708984375,
0.459228515625,
0.51708984375,
0.459228515625,
0.53076171875,
0.14697265625,
0.53076171875
],
[
0.515625,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5361328125,
0.515625,
0.5361328125
],
[
0.1502685546875,
0.53515625,
0.45751953125,
0.53515625,
0.45751953125,
0.5458984375,
0.1502685546875,
0.5458984375
],
[
0.5146484375,
0.54052734375,
0.88330078125,
0.54052734375,
0.88330078125,
0.55419921875,
0.5146484375,
0.55419921875
],
[
0.14697265625,
0.548828125,
0.459228515625,
0.548828125,
0.459228515625,
0.5625,
0.14697265625,
0.5625
],
[
0.5146484375,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.568359375,
0.5146484375,
0.568359375
],
[
0.145263671875,
0.5654296875,
0.4609375,
0.56396484375,
0.4609375,
0.5791015625,
0.145263671875,
0.580078125
],
[
0.51318359375,
0.572265625,
0.884765625,
0.572265625,
0.884765625,
0.5859375,
0.51318359375,
0.5859375
],
[
0.14697265625,
0.58203125,
0.45751953125,
0.58203125,
0.45751953125,
0.595703125,
0.14697265625,
0.595703125
],
[
0.51318359375,
0.5888671875,
0.88330078125,
0.5888671875,
0.88330078125,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.145263671875,
0.5966796875,
0.280517578125,
0.59716796875,
0.280517578125,
0.6123046875,
0.145263671875,
0.61083984375
],
[
0.51318359375,
0.60498046875,
0.60546875,
0.60498046875,
0.60546875,
0.619140625,
0.51318359375,
0.619140625
],
[
0.1435546875,
0.62353515625,
0.26123046875,
0.62353515625,
0.26123046875,
0.63818359375,
0.1435546875,
0.63818359375
],
[
0.52880859375,
0.62158203125,
0.884765625,
0.62158203125,
0.884765625,
0.6357421875,
0.52880859375,
0.6357421875
],
[
0.1427001953125,
0.62744140625,
0.1485595703125,
0.62744140625,
0.1485595703125,
0.6328125,
0.1427001953125,
0.6328125
],
[
0.51318359375,
0.63720703125,
0.884765625,
0.63720703125,
0.884765625,
0.65185546875,
0.51318359375,
0.65185546875
],
[
0.1175537109375,
0.64794921875,
0.489501953125,
0.6484375,
0.489501953125,
0.6630859375,
0.1175537109375,
0.662109375
],
[
0.51318359375,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.1175537109375,
0.6650390625,
0.48779296875,
0.66455078125,
0.48779296875,
0.6787109375,
0.1175537109375,
0.6796875
],
[
0.5146484375,
0.67041015625,
0.88330078125,
0.67041015625,
0.88330078125,
0.68505859375,
0.5146484375,
0.68505859375
],
[
0.1175537109375,
0.6796875,
0.489501953125,
0.68115234375,
0.489501953125,
0.6953125,
0.1175537109375,
0.69482421875
],
[
0.515625,
0.6884765625,
0.8818359375,
0.6884765625,
0.8818359375,
0.69873046875,
0.515625,
0.69873046875
],
[
0.12005615234375,
0.69677734375,
0.48779296875,
0.69677734375,
0.48779296875,
0.71142578125,
0.12005615234375,
0.71142578125
],
[
0.5146484375,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71435546875,
0.5146484375,
0.71435546875
],
[
0.12005615234375,
0.71337890625,
0.490234375,
0.71337890625,
0.490234375,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.5146484375,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.5146484375,
0.732421875
],
[
0.118408203125,
0.728515625,
0.489501953125,
0.728515625,
0.489501953125,
0.7421875,
0.118408203125,
0.7421875
],
[
0.515625,
0.736328125,
0.8818359375,
0.736328125,
0.8818359375,
0.74658203125,
0.515625,
0.74658203125
],
[
0.118408203125,
0.7451171875,
0.489501953125,
0.7451171875,
0.489501953125,
0.7587890625,
0.118408203125,
0.7587890625
],
[
0.51318359375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.12005615234375,
0.76123046875,
0.489501953125,
0.76123046875,
0.489501953125,
0.77490234375,
0.12005615234375,
0.77490234375
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.118408203125,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.51318359375,
0.78076171875,
0.88330078125,
0.7822265625,
0.88330078125,
0.79638671875,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.7919921875,
0.48779296875,
0.79345703125,
0.48779296875,
0.80810546875,
0.118408203125,
0.80712890625
],
[
0.51123046875,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.51123046875,
0.8125
],
[
0.118408203125,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51318359375,
0.8154296875,
0.80615234375,
0.8154296875,
0.80615234375,
0.8291015625,
0.51318359375,
0.8291015625
],
[
0.118408203125,
0.826171875,
0.489501953125,
0.826171875,
0.489501953125,
0.8388671875,
0.118408203125,
0.8388671875
],
[
0.52880859375,
0.83056640625,
0.8818359375,
0.83203125,
0.8818359375,
0.8466796875,
0.52880859375,
0.845703125
],
[
0.118408203125,
0.84033203125,
0.48779296875,
0.84130859375,
0.48779296875,
0.85546875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.8486328125,
0.6533203125,
0.8486328125,
0.6533203125,
0.8623046875,
0.51318359375,
0.8623046875
],
[
0.118408203125,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.87158203125,
0.118408203125,
0.87158203125
],
[
0.118408203125,
0.87353515625,
0.489501953125,
0.87255859375,
0.489501953125,
0.88720703125,
0.118408203125,
0.88818359375
],
[
0.53564453125,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.53564453125,
0.8935546875
],
[
0.14697265625,
0.8955078125,
0.291259765625,
0.8955078125,
0.291259765625,
0.9091796875,
0.14697265625,
0.9091796875
],
[
0.5498046875,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5498046875,
0.9091796875
]
] |
[
"Neural System Combination for Machine Translation",
"Long Zhout, Wenpeng Hut, Jiajun Zhangt* Chengqing Zongt",
"tUniversity of Chinese Academy of Sciences, Beijing, China",
"National Laboratory of Pattern Recognition, CASIA, Beijing, China",
"tCAS Center for Excellence in Brain Science and Intelligence Technology, Shanghai, China",
"[long.zhou,wenpeng.hu,jjzhang,cqzong)@nlpr.ia.ac.cn",
"Abstract",
"(Chiang, 2005)), does not need to limit the vo-",
"cabulary and can guarantee translation coverage",
"Neural machine translation (NMT) be-",
"of source sentences. It is obvious that NMT and",
"",
"comes a new approach to machine trans-",
"SMT have different strength and weakness. In or-",
"lation and generates much more fluent",
"der to take full advantages of both NMT and SMT,",
"results compared to statistical machine",
"system combination can be a good choice.",
"translation (SMT). However, SMT is usu-",
"Traditionally, system combination has been ex-",
"ally better than NMT in translation ade-",
"plored respectively in sentence-level, phrase-level,",
"quacy. It is therefore a promising direction",
"and word-level (Kumar and Byrne, 2004; Feng",
"to combine the advantages of both NMT",
"et al., 2009; Chen et al., 2009). Among them,",
"and SMT. In this paper, we propose a neu-",
"word-level combination approaches that adopt",
"[T's]",
"ral system combination framework lever-",
"confusion network for decoding have been quite",
"aging multi-source NMT, which takes as",
"successful (Rosti et al., 2007; Ayan et al., 2008:",
"input the outputs of NMT and SMT sys-",
"Freitag et al., 2014). However, these approaches",
"tems and produces the final translation",
"are mainly designed for SMT without considering",
"Extensive experiments on the Chinese-",
"thefeatures of NMT results. NMT opts to produce",
"09007",
"to-English translation task show that our",
"diversewordsandfreeword order,whichare quite",
"model archives significant improvement",
"different from SMT. And this will make it hard to",
"by 5.3 BLEU points over the best single",
"constructaconsistentconfusionnetwork.Further-",
"system output and 3.4 BLEU points over",
"more, traditional system combination approaches",
"the state-of-the-art traditional system com-",
"cannot guarantee the fluency of the final transla-",
"bination methods.",
"tion results.",
"Introduction",
"In this paper, we propose a neural system",
"",
"combination framework, which is adapted from",
"Neural machine translation has significantly im-",
"the multi-source NMT model (Zoph and Knight",
"proved the quality of machine translation in re-",
"2016). Different encoders are employed to model",
"cent several years (Kalchbrenner and Blunsom,",
"the semantics of the source language input and",
"2013; Sutskever et al., 2014; Bahdanau et al.,",
"eachbest translation produced by different NMT",
"2015; Junczys-Dowmunt et al., 2016a). Although",
"and SMT systems. The encoders produce multi-",
"most sentences are more fluent than translations",
"plecontextvectorrepresentations.fromwhichthe",
"by statistical machine translation (SMT) (Koehn",
"decoder generates the final output word by word.",
"et al., 2003; Chiang, 2005), NMT has a problem",
"Since the same training data is used for NMT,",
"to address translation adequacy especially for the",
"SMT and neural system combination, we further",
"rare and unknown words.Additionally, it suf-",
"design a smart strategy to simulate the real train-",
"fers from over-translation and under-translation",
"ing data for neural system combination.",
"to some extent (Tu et al., 2016). Compared to",
"Specifically, we make the following contribu-",
"NMT, SMT, such as phrase-based machine trans-",
"tions in this paper:",
"lation (PBMT, (Koehn et al., 2003) and hierar-",
"chical phrase-based machine translation (HPMT,",
". We propose a neural system combination",
"*Corresponding author.",
"method, which is adapted from multi-source"
] |
{
"arxiv_id": "1704.06393",
"language": "en",
"page_idx": 0
}
|
||
1704.06393-en-2
|
1704.06393-en-2.jpg
|
<table>
<tbody>
<tr>
<th>System</th>
<th>$MT03$</th>
<th>$\mathbf{MTO4}$</th>
<th>$MT05$</th>
<th>$MT06$</th>
<th>Ave</th>
</tr>
<tr>
<td>PBMT</td>
<td>37.47</td>
<td>41.20</td>
<td>36.41</td>
<td>36.03</td>
<td>37.78</td>
</tr>
<tr>
<td>HPMT</td>
<td>38.05</td>
<td>41.47</td>
<td>36.86</td>
<td>36.04</td>
<td>38.10</td>
</tr>
<tr>
<td>$NMT$</td>
<td>37.91</td>
<td>38.95</td>
<td>36.02</td>
<td>36.65</td>
<td>37.38</td>
</tr>
<tr>
<td>Jane (Freitag et al..2014)</td>
<td>39.83</td>
<td>42.75</td>
<td>38.63</td>
<td>39.10</td>
<td>40.08</td>
</tr>
<tr>
<td>Multi</td>
<td>40.64</td>
<td>44.81</td>
<td>38.80</td>
<td>38.26</td>
<td>40.63</td>
</tr>
<tr>
<td>Multi+ Source</td>
<td>42.16</td>
<td>45.51</td>
<td>40.28</td>
<td>39.03</td>
<td>41.75</td>
</tr>
<tr>
<td>Multi+ Ensemble</td>
<td>41.67</td>
<td>45.95</td>
<td>40.37</td>
<td>39.02</td>
<td>41.75</td>
</tr>
<tr>
<td>Multi+ Source+ Ensemble</td>
<td>43.55</td>
<td>47.09</td>
<td>42.02</td>
<td>41.10</td>
<td>43.44</td>
</tr>
</tbody>
</table>
Table 1: Translation results (BLEU score) for different machine translation and system combination methods. Jane is a open source machine translation system combination toolkit that uses confusion network decoding. Best and important results per category are highlighted.
(3)
$$\tilde{s}_{j-1}=GRU(s_{j-1},y_{j-1})$$
to simulate the real scenario. We randomly divide the training corpus into two parts, then reciprocally train the MT system on one half and translate the source sentences of the other half into target translations. The MT translations and the gold target reference can be available.
where $s_j-1$ is previous hidden state, $\tilde{s}_j-1$ is an intermediate state. And $c_j$ is the context vector of system combination obtained by attention mechanism, which is computed as weighted sum of the context vectors of three MT systems, just as illustrated in the middle part of Figure 1.
### 4 $\textbf{Experiments}$
(4)
$$c_j=\sum\limits_{k=1}^K\beta_{jk}c_{jk}$$
We perform our experiments on the ChineseEnglish translation task. The MT systems participating in system combination are PBMT, HPMT and NMT. The evaluation metric is caseinsensitive BLEU (Papineni et al., 2002).
where K is the number of MT systems, and $\beta_{jk}$ is
a normalized item calculated as follows:
### 4.1 Data preparation
(5)
$$\beta_{jk}=\frac{exp(\tilde{s}_{j-1}\cdot c_{jk})}{\sum_{k'}exp(\tilde{s}_{j-1}\cdot c_{jk'})}$$
Our training data consists of 2.08M sentence pairs extracted from LDC corpus. We use NIST 2003 Chinese-English dataset as the validation set, NIST 2004-2006 datasets as test sets. We list all the translation methods as follows:
Here, we calculate kth MT system context $c_jk$ as a
weighted sum of the source annotations:
(6)
$$c_{jk}=\sum_{i=1}^m\alpha_{ji}^kh_i$$
$\bullet$ PBMT: It is the start-of-the-art phrase-based SMT system. We use its default setting and train a 4-gram language model on the target portion of the bilingual training data.
$\begin{array}{l}{\mathrm{where~}h_{i}=[\overrightarrow{h}_{i};\overleftarrow{h}_{i}]\text{ is the annotation of }z_{i}\text{ from}}\\{\mathrm{a~bi-directional~GRU,~and~its~weight~}\alpha_{ji}^{k}\text{ is com-}}\end{array}$
puted by
$\bullet$ HPMT: It is a hierarchical phrase-based SMT system, which uses its default configuration as PBMT in Moses.
(7)
$$\alpha_{ji}^k=\frac{exp(e_{ji})}{\sum_{l=1}^mexp(e_{jl})}$$
$\bullet \textbf{ NMT: It is an attention- based NMT system, }$
with the same setting given in section 2.
# 4.2 Training Details
where $e_ji=v_a^Ttanh(W_a\tilde{s}_{j-1}+U_ah_i)$ scores how
well $\tilde{s}_{j-1}$ and $h_i$ match.
Training Data Simulation: The neural system combination framework should be trained on the outputs of multiple translation systems and the gold target translations. In order to keep consistency in training and testing, we design a strategy
The hyper-parameters used in our neural combination system are described as follows. We limit both Chinese and English vocabulary to 30k in our experiments. The number of hidden units is 1000
|
[
[
0.2376708984375,
0.07244873046875,
0.300537109375,
0.0748291015625,
0.298828125,
0.09027099609375,
0.2359619140625,
0.0885009765625
],
[
0.447509765625,
0.0736083984375,
0.49951171875,
0.0736083984375,
0.49951171875,
0.08905029296875,
0.447509765625,
0.08905029296875
],
[
0.51318359375,
0.0736083984375,
0.56494140625,
0.0736083984375,
0.56494140625,
0.08905029296875,
0.51318359375,
0.08905029296875
],
[
0.57861328125,
0.0736083984375,
0.63037109375,
0.0736083984375,
0.63037109375,
0.08905029296875,
0.57861328125,
0.08905029296875
],
[
0.64404296875,
0.0748291015625,
0.69775390625,
0.0748291015625,
0.69775390625,
0.08905029296875,
0.64404296875,
0.08905029296875
],
[
0.71728515625,
0.0748291015625,
0.75390625,
0.0748291015625,
0.75390625,
0.09027099609375,
0.71728515625,
0.09027099609375
],
[
0.2359619140625,
0.0914306640625,
0.295654296875,
0.09320068359375,
0.2939453125,
0.1092529296875,
0.235107421875,
0.10687255859375
],
[
0.44921875,
0.09320068359375,
0.496337890625,
0.09320068359375,
0.496337890625,
0.1092529296875,
0.44921875,
0.1092529296875
],
[
0.5146484375,
0.09320068359375,
0.5634765625,
0.09320068359375,
0.5634765625,
0.1092529296875,
0.5146484375,
0.1092529296875
],
[
0.580078125,
0.09320068359375,
0.62744140625,
0.09320068359375,
0.62744140625,
0.1092529296875,
0.580078125,
0.1092529296875
],
[
0.646484375,
0.09320068359375,
0.6943359375,
0.09320068359375,
0.6943359375,
0.1092529296875,
0.646484375,
0.1092529296875
],
[
0.70947265625,
0.0926513671875,
0.759765625,
0.09442138671875,
0.75830078125,
0.10986328125,
0.7080078125,
0.10809326171875
],
[
0.2392578125,
0.1092529296875,
0.297119140625,
0.1092529296875,
0.297119140625,
0.12353515625,
0.2392578125,
0.12353515625
],
[
0.44921875,
0.1092529296875,
0.496337890625,
0.1092529296875,
0.496337890625,
0.12469482421875,
0.44921875,
0.12469482421875
],
[
0.5146484375,
0.1092529296875,
0.5634765625,
0.1092529296875,
0.5634765625,
0.12469482421875,
0.5146484375,
0.12469482421875
],
[
0.580078125,
0.1092529296875,
0.62890625,
0.1092529296875,
0.62890625,
0.12469482421875,
0.580078125,
0.12469482421875
],
[
0.646484375,
0.1092529296875,
0.6943359375,
0.1092529296875,
0.6943359375,
0.12469482421875,
0.646484375,
0.12469482421875
],
[
0.70947265625,
0.1092529296875,
0.75830078125,
0.1092529296875,
0.75830078125,
0.12469482421875,
0.70947265625,
0.12469482421875
],
[
0.44921875,
0.12469482421875,
0.495361328125,
0.12469482421875,
0.495361328125,
0.14013671875,
0.44921875,
0.14013671875
],
[
0.5146484375,
0.12469482421875,
0.5634765625,
0.12469482421875,
0.5634765625,
0.14013671875,
0.5146484375,
0.14013671875
],
[
0.580078125,
0.1258544921875,
0.62890625,
0.1258544921875,
0.62890625,
0.14013671875,
0.580078125,
0.14013671875
],
[
0.646484375,
0.1258544921875,
0.6943359375,
0.1258544921875,
0.6943359375,
0.14013671875,
0.646484375,
0.14013671875
],
[
0.70947265625,
0.1258544921875,
0.75830078125,
0.1258544921875,
0.75830078125,
0.14013671875,
0.70947265625,
0.14013671875
],
[
0.235107421875,
0.141357421875,
0.432373046875,
0.1436767578125,
0.432373046875,
0.157958984375,
0.235107421875,
0.1561279296875
],
[
0.2392578125,
0.1258544921875,
0.2880859375,
0.1258544921875,
0.2880859375,
0.14013671875,
0.2392578125,
0.14013671875
],
[
0.447509765625,
0.142578125,
0.495361328125,
0.142578125,
0.495361328125,
0.15673828125,
0.447509765625,
0.15673828125
],
[
0.5146484375,
0.142578125,
0.5634765625,
0.142578125,
0.5634765625,
0.15673828125,
0.5146484375,
0.15673828125
],
[
0.5810546875,
0.142578125,
0.62890625,
0.142578125,
0.62890625,
0.15673828125,
0.5810546875,
0.15673828125
],
[
0.646484375,
0.142578125,
0.6943359375,
0.142578125,
0.6943359375,
0.15673828125,
0.646484375,
0.15673828125
],
[
0.70947265625,
0.142578125,
0.76171875,
0.142578125,
0.76171875,
0.15673828125,
0.70947265625,
0.15673828125
],
[
0.2392578125,
0.157958984375,
0.287109375,
0.157958984375,
0.287109375,
0.17333984375,
0.2392578125,
0.17333984375
],
[
0.44921875,
0.157958984375,
0.496337890625,
0.157958984375,
0.496337890625,
0.17333984375,
0.44921875,
0.17333984375
],
[
0.5146484375,
0.157958984375,
0.56103515625,
0.157958984375,
0.56103515625,
0.17333984375,
0.5146484375,
0.17333984375
],
[
0.5810546875,
0.157958984375,
0.62890625,
0.157958984375,
0.62890625,
0.17333984375,
0.5810546875,
0.17333984375
],
[
0.646484375,
0.157958984375,
0.6943359375,
0.157958984375,
0.6943359375,
0.17333984375,
0.646484375,
0.17333984375
],
[
0.70947265625,
0.1561279296875,
0.75830078125,
0.157958984375,
0.75732421875,
0.17333984375,
0.7080078125,
0.171630859375
],
[
0.44921875,
0.17333984375,
0.497802734375,
0.17333984375,
0.497802734375,
0.189453125,
0.44921875,
0.189453125
],
[
0.5146484375,
0.17333984375,
0.5615234375,
0.17333984375,
0.5615234375,
0.189453125,
0.5146484375,
0.189453125
],
[
0.580078125,
0.17333984375,
0.62890625,
0.17333984375,
0.62890625,
0.189453125,
0.580078125,
0.189453125
],
[
0.646484375,
0.17333984375,
0.6943359375,
0.17333984375,
0.6943359375,
0.189453125,
0.646484375,
0.189453125
],
[
0.70947265625,
0.17333984375,
0.75830078125,
0.17333984375,
0.75830078125,
0.189453125,
0.70947265625,
0.189453125
],
[
0.240966796875,
0.190673828125,
0.366943359375,
0.190673828125,
0.366943359375,
0.204833984375,
0.240966796875,
0.204833984375
],
[
0.240966796875,
0.174560546875,
0.34765625,
0.174560546875,
0.34765625,
0.189453125,
0.240966796875,
0.189453125
],
[
0.44921875,
0.190673828125,
0.496337890625,
0.190673828125,
0.496337890625,
0.2060546875,
0.44921875,
0.2060546875
],
[
0.5146484375,
0.190673828125,
0.5634765625,
0.190673828125,
0.5634765625,
0.2060546875,
0.5146484375,
0.2060546875
],
[
0.5810546875,
0.190673828125,
0.62890625,
0.190673828125,
0.62890625,
0.2060546875,
0.5810546875,
0.2060546875
],
[
0.646484375,
0.190673828125,
0.6943359375,
0.190673828125,
0.6943359375,
0.2060546875,
0.646484375,
0.2060546875
],
[
0.70947265625,
0.190673828125,
0.75830078125,
0.190673828125,
0.75830078125,
0.2060546875,
0.70947265625,
0.2060546875
],
[
0.450927734375,
0.207275390625,
0.496337890625,
0.207275390625,
0.496337890625,
0.221435546875,
0.450927734375,
0.221435546875
],
[
0.5146484375,
0.207275390625,
0.5634765625,
0.207275390625,
0.5634765625,
0.221435546875,
0.5146484375,
0.221435546875
],
[
0.5810546875,
0.207275390625,
0.62890625,
0.207275390625,
0.62890625,
0.221435546875,
0.5810546875,
0.221435546875
],
[
0.646484375,
0.207275390625,
0.6962890625,
0.207275390625,
0.6962890625,
0.221435546875,
0.646484375,
0.221435546875
],
[
0.7109375,
0.207275390625,
0.76171875,
0.207275390625,
0.76171875,
0.22265625,
0.7109375,
0.22265625
],
[
0.1175537109375,
0.2327880859375,
0.88330078125,
0.2340087890625,
0.88330078125,
0.248779296875,
0.1175537109375,
0.2476806640625
],
[
0.2392578125,
0.2078857421875,
0.429931640625,
0.2078857421875,
0.429931640625,
0.221435546875,
0.2392578125,
0.221435546875
],
[
0.12005615234375,
0.25048828125,
0.88330078125,
0.25048828125,
0.88330078125,
0.26416015625,
0.12005615234375,
0.26416015625
],
[
0.118408203125,
0.267333984375,
0.67822265625,
0.267333984375,
0.67822265625,
0.28076171875,
0.118408203125,
0.28076171875
],
[
0.5146484375,
0.30810546875,
0.8818359375,
0.30810546875,
0.8818359375,
0.318359375,
0.5146484375,
0.318359375
],
[
0.20654296875,
0.32177734375,
0.399658203125,
0.32177734375,
0.399658203125,
0.336181640625,
0.20654296875,
0.336181640625
],
[
0.462646484375,
0.32373046875,
0.490234375,
0.32373046875,
0.490234375,
0.336181640625,
0.462646484375,
0.336181640625
],
[
0.51123046875,
0.32177734375,
0.8818359375,
0.322509765625,
0.8818359375,
0.33740234375,
0.51123046875,
0.336181640625
],
[
0.51123046875,
0.33740234375,
0.88330078125,
0.33837890625,
0.88330078125,
0.352783203125,
0.51123046875,
0.3515625
],
[
0.118408203125,
0.3466796875,
0.489501953125,
0.3466796875,
0.489501953125,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.5146484375,
0.355712890625,
0.8818359375,
0.355712890625,
0.8818359375,
0.366455078125,
0.5146484375,
0.366455078125
],
[
0.118408203125,
0.363525390625,
0.490234375,
0.363525390625,
0.490234375,
0.377197265625,
0.118408203125,
0.377197265625
],
[
0.51318359375,
0.37060546875,
0.884765625,
0.37060546875,
0.884765625,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.1175537109375,
0.377685546875,
0.489501953125,
0.377685546875,
0.489501953125,
0.391357421875,
0.1175537109375,
0.391357421875
],
[
0.51318359375,
0.38720703125,
0.755859375,
0.38720703125,
0.755859375,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.118408203125,
0.394287109375,
0.490234375,
0.394287109375,
0.490234375,
0.407958984375,
0.118408203125,
0.407958984375
],
[
0.1175537109375,
0.4091796875,
0.48779296875,
0.410400390625,
0.48779296875,
0.424560546875,
0.1175537109375,
0.42333984375
],
[
0.515625,
0.41748046875,
0.54248046875,
0.41748046875,
0.54248046875,
0.42578125,
0.515625,
0.42578125
],
[
0.5380859375,
0.4150390625,
0.6572265625,
0.4150390625,
0.6572265625,
0.4287109375,
0.5380859375,
0.4287109375
],
[
0.1175537109375,
0.427001953125,
0.3896484375,
0.427001953125,
0.3896484375,
0.440673828125,
0.1175537109375,
0.440673828125
],
[
0.5146484375,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.5146484375,
0.454833984375
],
[
0.2880859375,
0.454833984375,
0.306396484375,
0.454833984375,
0.306396484375,
0.463134765625,
0.2880859375,
0.463134765625
],
[
0.51318359375,
0.457275390625,
0.8818359375,
0.457275390625,
0.8818359375,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.2939453125,
0.46435546875,
0.310546875,
0.4755859375,
0.297119140625,
0.485107421875,
0.280517578125,
0.473876953125
],
[
0.30322265625,
0.468505859375,
0.36279296875,
0.468505859375,
0.36279296875,
0.482666015625,
0.30322265625,
0.482666015625
],
[
0.465087890625,
0.468505859375,
0.490234375,
0.468505859375,
0.490234375,
0.48095703125,
0.465087890625,
0.48095703125
],
[
0.51318359375,
0.47265625,
0.88330078125,
0.47265625,
0.88330078125,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.2880859375,
0.48681640625,
0.30908203125,
0.48681640625,
0.30908203125,
0.495361328125,
0.2880859375,
0.495361328125
],
[
0.2451171875,
0.47021484375,
0.274658203125,
0.47021484375,
0.274658203125,
0.482177734375,
0.2451171875,
0.482177734375
],
[
0.51318359375,
0.488037109375,
0.88330078125,
0.4892578125,
0.88330078125,
0.50439453125,
0.51318359375,
0.5029296875
],
[
0.118408203125,
0.505859375,
0.489501953125,
0.50732421875,
0.489501953125,
0.521484375,
0.118408203125,
0.52099609375
],
[
0.51318359375,
0.5048828125,
0.8212890625,
0.5048828125,
0.8212890625,
0.51953125,
0.51318359375,
0.51953125
],
[
0.1175537109375,
0.5224609375,
0.414794921875,
0.5224609375,
0.414794921875,
0.53759765625,
0.1175537109375,
0.53759765625
],
[
0.515625,
0.533203125,
0.5556640625,
0.533203125,
0.5556640625,
0.54443359375,
0.515625,
0.54443359375
],
[
0.54833984375,
0.533203125,
0.6943359375,
0.533203125,
0.6943359375,
0.54638671875,
0.54833984375,
0.54638671875
],
[
0.270263671875,
0.54736328125,
0.385498046875,
0.54736328125,
0.385498046875,
0.5625,
0.270263671875,
0.5625
],
[
0.465087890625,
0.55712890625,
0.491943359375,
0.55712890625,
0.491943359375,
0.5693359375,
0.465087890625,
0.5693359375
],
[
0.51123046875,
0.55419921875,
0.884765625,
0.552734375,
0.884765625,
0.56787109375,
0.51123046875,
0.568359375
],
[
0.2476806640625,
0.5625,
0.4072265625,
0.5654296875,
0.4072265625,
0.58447265625,
0.2476806640625,
0.58056640625
],
[
0.51318359375,
0.5693359375,
0.88330078125,
0.568359375,
0.88330078125,
0.5830078125,
0.51318359375,
0.58447265625
],
[
0.20068359375,
0.55810546875,
0.2376708984375,
0.55810546875,
0.2376708984375,
0.57080078125,
0.20068359375,
0.57080078125
],
[
0.1175537109375,
0.59130859375,
0.490234375,
0.59423828125,
0.490234375,
0.609375,
0.1175537109375,
0.60546875
],
[
0.5146484375,
0.5859375,
0.8818359375,
0.5859375,
0.8818359375,
0.599609375,
0.5146484375,
0.599609375
],
[
0.51318359375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.6142578125,
0.51318359375,
0.6142578125
],
[
0.12005615234375,
0.60986328125,
0.418212890625,
0.60986328125,
0.418212890625,
0.62353515625,
0.12005615234375,
0.62353515625
],
[
0.51318359375,
0.6181640625,
0.7734375,
0.6181640625,
0.7734375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.2939453125,
0.63525390625,
0.316650390625,
0.63525390625,
0.316650390625,
0.64794921875,
0.2939453125,
0.64794921875
],
[
0.300537109375,
0.646484375,
0.36279296875,
0.6494140625,
0.361083984375,
0.6650390625,
0.298828125,
0.662109375
],
[
0.462646484375,
0.6484375,
0.489501953125,
0.6484375,
0.489501953125,
0.662109375,
0.462646484375,
0.662109375
],
[
0.53076171875,
0.64453125,
0.88330078125,
0.6455078125,
0.88330078125,
0.66015625,
0.53076171875,
0.6591796875
],
[
0.2451171875,
0.65185546875,
0.273681640625,
0.65185546875,
0.273681640625,
0.66455078125,
0.2451171875,
0.66455078125
],
[
0.29296875,
0.6650390625,
0.31982421875,
0.6650390625,
0.31982421875,
0.6767578125,
0.29296875,
0.6767578125
],
[
0.5498046875,
0.662109375,
0.88330078125,
0.662109375,
0.88330078125,
0.6767578125,
0.5498046875,
0.6767578125
],
[
0.5498046875,
0.6787109375,
0.88330078125,
0.6787109375,
0.88330078125,
0.6923828125,
0.5498046875,
0.6923828125
],
[
0.1175537109375,
0.6884765625,
0.490234375,
0.689453125,
0.490234375,
0.7080078125,
0.1175537109375,
0.70703125
],
[
0.54833984375,
0.69482421875,
0.82373046875,
0.693359375,
0.82373046875,
0.7080078125,
0.54833984375,
0.708984375
],
[
0.1175537109375,
0.7080078125,
0.484375,
0.708984375,
0.484375,
0.7236328125,
0.1175537109375,
0.72265625
],
[
0.118408203125,
0.7255859375,
0.187255859375,
0.7255859375,
0.187255859375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.5322265625,
0.72021484375,
0.88330078125,
0.7216796875,
0.88330078125,
0.736328125,
0.5322265625,
0.7353515625
],
[
0.5498046875,
0.73828125,
0.8818359375,
0.73828125,
0.8818359375,
0.7529296875,
0.5498046875,
0.7529296875
],
[
0.2939453125,
0.751953125,
0.361083984375,
0.751953125,
0.361083984375,
0.76708984375,
0.2939453125,
0.76708984375
],
[
0.2216796875,
0.7587890625,
0.2568359375,
0.7587890625,
0.2568359375,
0.775390625,
0.2216796875,
0.775390625
],
[
0.4609375,
0.7587890625,
0.490234375,
0.7587890625,
0.490234375,
0.775390625,
0.4609375,
0.775390625
],
[
0.5498046875,
0.75439453125,
0.75830078125,
0.75439453125,
0.75830078125,
0.7685546875,
0.5498046875,
0.7685546875
],
[
0.270263671875,
0.76708984375,
0.382080078125,
0.76708984375,
0.382080078125,
0.78515625,
0.270263671875,
0.78515625
],
[
0.280517578125,
0.7685546875,
0.306396484375,
0.7685546875,
0.306396484375,
0.775390625,
0.280517578125,
0.775390625
],
[
0.5322265625,
0.77978515625,
0.8818359375,
0.78076171875,
0.8818359375,
0.7958984375,
0.5322265625,
0.79443359375
],
[
0.1158447265625,
0.79638671875,
0.48779296875,
0.79638671875,
0.48779296875,
0.81396484375,
0.1158447265625,
0.81396484375
],
[
0.5498046875,
0.79736328125,
0.84619140625,
0.79736328125,
0.84619140625,
0.8125,
0.5498046875,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.297119140625,
0.81396484375,
0.297119140625,
0.8291015625,
0.12005615234375,
0.8291015625
],
[
0.5146484375,
0.82470703125,
0.68603515625,
0.82470703125,
0.68603515625,
0.8388671875,
0.5146484375,
0.8388671875
],
[
0.1368408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.1368408203125,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.490234375,
0.84716796875,
0.490234375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.5146484375,
0.8466796875,
0.8818359375,
0.8466796875,
0.8818359375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.8798828125,
0.490234375,
0.8798828125,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.51123046875,
0.87744140625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.51123046875,
0.892578125
],
[
0.118408203125,
0.89453125,
0.489501953125,
0.8955078125,
0.489501953125,
0.91015625,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"System",
"MT03",
"MT04",
"MT05",
"MT06",
"Ave",
"PBMT",
"37.47",
"41.20",
"36.41",
"36.03",
"37.78",
"HPMT",
"38.05",
"41.47",
"36.86",
"36.04",
"38.10",
"37.91",
"38.95",
"36.02",
"36.65",
"37.38",
"Jane (Freitag et al., 2014)",
"NMT",
"39.83",
"42.75",
"38.63",
"39.10",
"40.08",
"Multi",
"40.64",
"44.81",
"38.80",
"38.26",
"40.63",
"42.16",
"45.51",
"40.28",
"39.03",
"41.75",
"Multi+Ensemble",
"Multi+Source",
"41.67",
"45.95",
"40.37",
"39.02",
"41.75",
"43.55",
"47.09",
"42.02",
"41.10",
"43.44",
"Table 1: Translation results (BLEU score) for different machine translation and system combination",
"Multi+Source+Ensemble",
"methods. Jane is a open source machine translation system combination toolkit that uses confusion",
"network decoding. Best and important results per category are highlighted.",
"to simulatethereal scenario.Werandomlydivide",
"sj-1 =GRU(si-1,Yi-1)",
"(3)",
"the training corpus into two parts, then recipro-",
"cally train the MT system on one half and trans-",
"where Sj-1 is previous hidden state, s,-1is an in-",
"latethe source sentences of the otherhalf intotar-",
"termediate state. And cj is the context vector of",
"get translations. The MT translations and the gold",
" system combination obtained by attention mecha-",
"target reference can be available.",
"nism, which is computed as weighted sum of the",
"context vectors of three MT systems, just as illus-",
"4",
"Experiments",
"trated in the middle part of Figure 1.",
"We perform our experiments on the Chinese-",
"K",
"English translation task. The MT systems par-",
"",
"BjkCik",
"(4)",
"ticipating in system combination are PBMT,",
"=1",
"Ci=",
"HPMT and NMT. The evaluation metric is case-",
"where K is the number of MT systems, and βik is",
"insensitive BLEU (Papineni et al., 2002).",
"a normalized item calculated as follows:",
"4.1",
"Data preparation",
"eap(sj-1 : Cjk)",
"(5) ",
"Our training data consists of 2.08M sentence",
"Ek eap(sj-1 -Cjk)",
"pairs extracted from LDC corpus. We use NIST",
"Bik=",
"Here, we calculate kth MT system context cjk as a",
"2003 Chinese-English dataset as the validation set,",
"NIST 2004-2006 datasets as test sets. We list all",
"weighted sum of the source annotations:",
"the translation methods as follows:",
"m",
"Coshi",
"(6)",
". PBMT: It is the start-of-the-art phrase-based",
"Cjk",
"i=1",
"SMT system. We use its default setting and",
"train a 4-gram language model on the target",
"where hi = [h,; hi] is the annotation of zi from",
"portion of the bilingual training data.",
"a bi-directional GRU, and its weight h, is com-",
"puted by",
".HPMT: It is a hierarchical phrase-based",
"SMT system, which uses its default config-",
"erp(eji)",
"oht:",
"(7)",
"uration as PBMT in Moses.",
"E=i ecp(ejt)",
"m",
". NMT: It is an attention-based NMT system,",
"where eji = v tanh(Wasj-1+Uah,) scores how",
"with the same setting given in section 2.",
"well s,-i and h; match.",
"4.2Training Details",
"Training Data Simulation: The neural sys-",
"tem combination framework should be trained on",
"The hyper-parameters used in our neural combi-",
"the outputs of multiple translation systems and the",
"nation system are described as follows. We limit",
"gold target translations. In order to keep consis-",
"both Chinese and English vocabulary to 30k in our",
"tency in training and testing, we design a strategy",
"experiments. The number of hidden units is 1000"
] |
{
"arxiv_id": "1704.06393",
"language": "en",
"page_idx": 2
}
|
||
1704.06393-en-3
|
1704.06393-en-3.jpg
|
<FigureHere>
Figure 2: Translation results (RIBES score) for different machine translation and system combination methods.
and the word embedding dimension is 500 for all source and target word. The network parameters are updated with Adadelta algorithm. We adopt beam search with beam size b=10 at test time.
As to confusion-network-based system Jane, we use its default configuration and train a 4-gram language model on target data and 10M Xinhua portion of Gigaword corpus.
## 4.3 Main Results
We compare our neural combination system with the best individual engines, and the state-of-theart traditional combination system Jane (Freitag et al., 2014). Table 1 shows the BLEU of different models on development data and test data. The BLEU score of the multi-source neural combination model is 2.53 higher than the best single model HPMT. The source language input gives a further improvement of +1.12 BLEU points.
As shown in Table 1, Jane outperforms the best single MT system by 1.92 BLEU points. However, our neural combination system with source language gets an improvement of 1.67 BLEU points over Jane. Furthermore, when augmenting our neural combination system with ensemble decoding $^2$,it leads to another significant boost of +1.69 BLEU points.
## 4.4 Word Order of Translation
We evaluate word order by the automatic evaluation metrics RIBES (Isozaki et al., 2010), whose score is a metric based on rank correlation coefficients with word precision. RIBES is known to have stronger correlation with human judgements than BLEU for English as discussed in Isozaki et al.(2010).
$^{2}$We use four neural combination models in ensemble
${\mathrm{model}}.$
<table>
<tbody>
<tr>
<th>$\underline{\text{System}}$</th>
<th>$MT03$</th>
<th>$\mathbf{MTO4}$</th>
<th>$\mathbf{MTO5}$</th>
<th>$MT06$</th>
<th>Ave</th>
</tr>
<tr>
<td>$NMT$</td>
<td>1086</td>
<td>1145</td>
<td>1020</td>
<td>708</td>
<td>989.8</td>
</tr>
<tr>
<td>Ours</td>
<td>869</td>
<td>1023</td>
<td>909</td>
<td>609</td>
<td>852.5</td>
</tr>
</tbody>
</table>
Table 2: The number of unknown words in the re-
sults of NMT and our model.
<table>
<tbody>
<tr>
<th>$\underline{\text{System}}$</th>
<th>$MT03$</th>
<th>$\mathbf{MTO4}$</th>
<th>$\mathbf{MTO5}$</th>
<th>$MT06$</th>
<th>Ave</th>
</tr>
<tr>
<td>$\mathbf{E-NMT}$</td>
<td>39.14</td>
<td>40.78</td>
<td>37.31</td>
<td>37.89</td>
<td>38.78</td>
</tr>
<tr>
<td>Jane</td>
<td>40.61</td>
<td>43.28</td>
<td>39.05</td>
<td>39.18</td>
<td>40.53</td>
</tr>
<tr>
<td>Ours</td>
<td>43.61</td>
<td>47.65</td>
<td>42.02</td>
<td>41.17</td>
<td>43.61</td>
</tr>
</tbody>
</table>
Table 3: Translation results (BLEU score) when we replace original NMT with strong E-NMT, which uses ensemble strategy with four NMT models. All results of system combination are based on strong outputs of E-NMT.
Figure 2 illustrates experimental results of RIBES scores, which demonstrates that our neural combination model outperforms the best result of single MT system and Jane. Additionally, although BLEU point of Jane is higher than single NMT system, the word order of Jane is worse in terms of RIBES.
4.5 Rare and Unknown Words Translation
It is difficult for NMT systems to handle rare words, because low-frequency words in training data cannot capture latent translation mappings in neural network model. However, we do not need to limit the vocabulary in SMT, which are often able to translate rare words in training data. As shown in Table 2, the number of unknown words of our proposed model is 137 fewer than original NMT model.
Table 4 shows an example of system combination. The Chinese word zuzhiwang is an out-ofvocabulary(OOV) for NMT and the baseline NMT cannot correctly translate this word. Although PBMT and HPMT translate this word well, they does not conform to the grammar. By combining the merits of NMT and SMT, our model gets the correct translation.
## 4.6 Effect of Ensemble Decoding
The performance of candidate systems is very important to the result of system combination, and we use ensemble strategy with four NMT models to improve the performance of original NMT system. As shown in Table 3, the E-NMT with
|
[
[
0.60107421875,
0.0736083984375,
0.7734375,
0.0736083984375,
0.7734375,
0.0885009765625,
0.60107421875,
0.0885009765625
],
[
0.83642578125,
0.0736083984375,
0.873046875,
0.07598876953125,
0.87158203125,
0.0885009765625,
0.83447265625,
0.08612060546875
],
[
0.765625,
0.07598876953125,
0.8271484375,
0.07598876953125,
0.8271484375,
0.0872802734375,
0.765625,
0.0872802734375
],
[
0.5205078125,
0.0748291015625,
0.58447265625,
0.0748291015625,
0.58447265625,
0.08905029296875,
0.5205078125,
0.08905029296875
],
[
0.5205078125,
0.09320068359375,
0.5693359375,
0.09320068359375,
0.5693359375,
0.1092529296875,
0.5205078125,
0.1092529296875
],
[
0.60205078125,
0.09320068359375,
0.6455078125,
0.09320068359375,
0.6455078125,
0.1092529296875,
0.60205078125,
0.1092529296875
],
[
0.779296875,
0.09320068359375,
0.8154296875,
0.09320068359375,
0.8154296875,
0.1092529296875,
0.779296875,
0.1092529296875
],
[
0.8310546875,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.1092529296875,
0.8310546875,
0.1092529296875
],
[
0.7197265625,
0.09442138671875,
0.76171875,
0.09442138671875,
0.76171875,
0.10809326171875,
0.7197265625,
0.10809326171875
],
[
0.66162109375,
0.09442138671875,
0.7001953125,
0.09442138671875,
0.7001953125,
0.10687255859375,
0.66162109375,
0.10687255859375
],
[
0.60546875,
0.1092529296875,
0.640625,
0.1092529296875,
0.640625,
0.12469482421875,
0.60546875,
0.12469482421875
],
[
0.66015625,
0.1092529296875,
0.70361328125,
0.1092529296875,
0.70361328125,
0.12469482421875,
0.66015625,
0.12469482421875
],
[
0.7197265625,
0.1092529296875,
0.75732421875,
0.1092529296875,
0.75732421875,
0.12469482421875,
0.7197265625,
0.12469482421875
],
[
0.779296875,
0.1092529296875,
0.8154296875,
0.1092529296875,
0.8154296875,
0.12469482421875,
0.779296875,
0.12469482421875
],
[
0.8310546875,
0.10809326171875,
0.884765625,
0.10809326171875,
0.884765625,
0.1270751953125,
0.8310546875,
0.1270751953125
],
[
0.1351318359375,
0.1312255859375,
0.1485595703125,
0.1312255859375,
0.1485595703125,
0.1436767578125,
0.1351318359375,
0.1436767578125
],
[
0.5224609375,
0.10986328125,
0.56494140625,
0.10986328125,
0.56494140625,
0.1258544921875,
0.5224609375,
0.1258544921875
],
[
0.51318359375,
0.1436767578125,
0.8818359375,
0.1436767578125,
0.8818359375,
0.15673828125,
0.51318359375,
0.15673828125
],
[
0.51318359375,
0.1602783203125,
0.72900390625,
0.1602783203125,
0.72900390625,
0.174560546875,
0.51318359375,
0.174560546875
],
[
0.52294921875,
0.189453125,
0.58447265625,
0.189453125,
0.58447265625,
0.20068359375,
0.52294921875,
0.20068359375
],
[
0.60107421875,
0.18701171875,
0.82958984375,
0.18701171875,
0.82958984375,
0.20068359375,
0.60107421875,
0.20068359375
],
[
0.82275390625,
0.189453125,
0.873046875,
0.189453125,
0.873046875,
0.20068359375,
0.82275390625,
0.20068359375
],
[
0.16796875,
0.196533203125,
0.28955078125,
0.196533203125,
0.28955078125,
0.210205078125,
0.16796875,
0.210205078125
],
[
0.291259765625,
0.196533203125,
0.322509765625,
0.1995849609375,
0.32080078125,
0.2078857421875,
0.28955078125,
0.204833984375
],
[
0.330078125,
0.19775390625,
0.45751953125,
0.196533203125,
0.45751953125,
0.2078857421875,
0.330078125,
0.208984375
],
[
0.5224609375,
0.2060546875,
0.587890625,
0.2060546875,
0.587890625,
0.2203369140625,
0.5224609375,
0.2203369140625
],
[
0.60107421875,
0.2060546875,
0.64990234375,
0.2060546875,
0.64990234375,
0.2203369140625,
0.60107421875,
0.2203369140625
],
[
0.65576171875,
0.2060546875,
0.7080078125,
0.2060546875,
0.7080078125,
0.2203369140625,
0.65576171875,
0.2203369140625
],
[
0.71533203125,
0.2060546875,
0.76416015625,
0.2060546875,
0.76416015625,
0.2203369140625,
0.71533203125,
0.2203369140625
],
[
0.77490234375,
0.2060546875,
0.8271484375,
0.2060546875,
0.8271484375,
0.2203369140625,
0.77490234375,
0.2203369140625
],
[
0.82275390625,
0.2078857421875,
0.880859375,
0.2078857421875,
0.880859375,
0.2197265625,
0.82275390625,
0.2197265625
],
[
0.5224609375,
0.2203369140625,
0.5615234375,
0.22265625,
0.56103515625,
0.2381591796875,
0.5205078125,
0.236328125
],
[
0.60107421875,
0.22265625,
0.64990234375,
0.22265625,
0.64990234375,
0.2369384765625,
0.60107421875,
0.2369384765625
],
[
0.65576171875,
0.22265625,
0.70947265625,
0.22265625,
0.70947265625,
0.2369384765625,
0.65576171875,
0.2369384765625
],
[
0.71533203125,
0.22265625,
0.765625,
0.22265625,
0.765625,
0.2369384765625,
0.71533203125,
0.2369384765625
],
[
0.77490234375,
0.22265625,
0.82861328125,
0.22265625,
0.82861328125,
0.2369384765625,
0.77490234375,
0.2369384765625
],
[
0.12005615234375,
0.2286376953125,
0.48779296875,
0.2286376953125,
0.48779296875,
0.2423095703125,
0.12005615234375,
0.2423095703125
],
[
0.8212890625,
0.223876953125,
0.87890625,
0.223876953125,
0.87890625,
0.235107421875,
0.8212890625,
0.235107421875
],
[
0.5224609375,
0.2381591796875,
0.56494140625,
0.2381591796875,
0.56494140625,
0.254150390625,
0.5224609375,
0.254150390625
],
[
0.60107421875,
0.2381591796875,
0.64990234375,
0.2381591796875,
0.64990234375,
0.2529296875,
0.60107421875,
0.2529296875
],
[
0.65576171875,
0.2381591796875,
0.7119140625,
0.2381591796875,
0.7119140625,
0.2529296875,
0.65576171875,
0.2529296875
],
[
0.71533203125,
0.2381591796875,
0.765625,
0.2381591796875,
0.765625,
0.2529296875,
0.71533203125,
0.2529296875
],
[
0.77490234375,
0.2381591796875,
0.8271484375,
0.2381591796875,
0.8271484375,
0.2529296875,
0.77490234375,
0.2529296875
],
[
0.8310546875,
0.2381591796875,
0.880859375,
0.2381591796875,
0.880859375,
0.2529296875,
0.8310546875,
0.2529296875
],
[
0.118408203125,
0.2452392578125,
0.48779296875,
0.2452392578125,
0.48779296875,
0.2587890625,
0.118408203125,
0.2587890625
],
[
0.12005615234375,
0.26318359375,
0.2376708984375,
0.26318359375,
0.2376708984375,
0.273681640625,
0.12005615234375,
0.273681640625
],
[
0.51318359375,
0.2724609375,
0.88330078125,
0.2724609375,
0.88330078125,
0.2861328125,
0.51318359375,
0.2861328125
],
[
0.51318359375,
0.289306640625,
0.88330078125,
0.289306640625,
0.88330078125,
0.303955078125,
0.51318359375,
0.303955078125
],
[
0.12005615234375,
0.298583984375,
0.489501953125,
0.298583984375,
0.489501953125,
0.312255859375,
0.12005615234375,
0.312255859375
],
[
0.5146484375,
0.305908203125,
0.88330078125,
0.305908203125,
0.88330078125,
0.320556640625,
0.5146484375,
0.320556640625
],
[
0.1175537109375,
0.3134765625,
0.490234375,
0.314208984375,
0.490234375,
0.328857421875,
0.1175537109375,
0.327880859375
],
[
0.51123046875,
0.320556640625,
0.884765625,
0.32177734375,
0.884765625,
0.336181640625,
0.51123046875,
0.3349609375
],
[
0.118408203125,
0.330078125,
0.489501953125,
0.330810546875,
0.489501953125,
0.345703125,
0.118408203125,
0.344482421875
],
[
0.51123046875,
0.33740234375,
0.77685546875,
0.336181640625,
0.77685546875,
0.350830078125,
0.51123046875,
0.3515625
],
[
0.1175537109375,
0.345703125,
0.459228515625,
0.3466796875,
0.459228515625,
0.361083984375,
0.1175537109375,
0.35986328125
],
[
0.1368408203125,
0.363525390625,
0.489501953125,
0.363525390625,
0.489501953125,
0.377197265625,
0.1368408203125,
0.377197265625
],
[
0.1175537109375,
0.377685546875,
0.489501953125,
0.37890625,
0.489501953125,
0.393798828125,
0.1175537109375,
0.392578125
],
[
0.5322265625,
0.377685546875,
0.884765625,
0.377685546875,
0.884765625,
0.392578125,
0.5322265625,
0.392578125
],
[
0.1175537109375,
0.394287109375,
0.490234375,
0.393798828125,
0.490234375,
0.407958984375,
0.1175537109375,
0.4091796875
],
[
0.5146484375,
0.394287109375,
0.8818359375,
0.394287109375,
0.8818359375,
0.407958984375,
0.5146484375,
0.407958984375
],
[
0.118408203125,
0.410888671875,
0.33154296875,
0.410888671875,
0.33154296875,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.5146484375,
0.410400390625,
0.88330078125,
0.410400390625,
0.88330078125,
0.42333984375,
0.5146484375,
0.42333984375
],
[
0.51318359375,
0.427001953125,
0.8818359375,
0.427001953125,
0.8818359375,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.118408203125,
0.43701171875,
0.267822265625,
0.43701171875,
0.267822265625,
0.45068359375,
0.118408203125,
0.45068359375
],
[
0.5146484375,
0.4423828125,
0.88330078125,
0.4423828125,
0.88330078125,
0.4560546875,
0.5146484375,
0.4560546875
],
[
0.12005615234375,
0.458984375,
0.490234375,
0.458984375,
0.490234375,
0.47265625,
0.12005615234375,
0.47265625
],
[
0.5146484375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.47021484375,
0.5146484375,
0.47021484375
],
[
0.1175537109375,
0.473876953125,
0.489501953125,
0.474365234375,
0.489501953125,
0.4892578125,
0.1175537109375,
0.488037109375
],
[
0.51318359375,
0.474365234375,
0.63818359375,
0.474365234375,
0.63818359375,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.118408203125,
0.4912109375,
0.489501953125,
0.4912109375,
0.489501953125,
0.5048828125,
0.118408203125,
0.5048828125
],
[
0.12005615234375,
0.50732421875,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.51318359375,
0.501953125,
0.861328125,
0.501953125,
0.861328125,
0.515625,
0.51318359375,
0.515625
],
[
0.118408203125,
0.52392578125,
0.489501953125,
0.52392578125,
0.489501953125,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.51123046875,
0.52392578125,
0.8818359375,
0.52392578125,
0.8818359375,
0.53759765625,
0.51123046875,
0.53759765625
],
[
0.118408203125,
0.5390625,
0.489501953125,
0.5390625,
0.489501953125,
0.552734375,
0.118408203125,
0.552734375
],
[
0.51123046875,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.552734375,
0.51123046875,
0.552734375
],
[
0.1175537109375,
0.55419921875,
0.490234375,
0.5546875,
0.490234375,
0.5693359375,
0.1175537109375,
0.568359375
],
[
0.51123046875,
0.5546875,
0.88330078125,
0.5556640625,
0.88330078125,
0.57080078125,
0.51123046875,
0.5693359375
],
[
0.118408203125,
0.57177734375,
0.49365234375,
0.57177734375,
0.49365234375,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.51318359375,
0.572265625,
0.884765625,
0.572265625,
0.884765625,
0.5859375,
0.51318359375,
0.5859375
],
[
0.118408203125,
0.58740234375,
0.450927734375,
0.58740234375,
0.450927734375,
0.60107421875,
0.118408203125,
0.60107421875
],
[
0.51318359375,
0.58837890625,
0.884765625,
0.58837890625,
0.884765625,
0.6015625,
0.51318359375,
0.6015625
],
[
0.1368408203125,
0.60400390625,
0.490234375,
0.60400390625,
0.490234375,
0.61767578125,
0.1368408203125,
0.61767578125
],
[
0.51318359375,
0.60400390625,
0.884765625,
0.60400390625,
0.884765625,
0.61767578125,
0.51318359375,
0.61767578125
],
[
0.118408203125,
0.619140625,
0.490234375,
0.619140625,
0.490234375,
0.63427734375,
0.118408203125,
0.63427734375
],
[
0.51318359375,
0.62060546875,
0.884765625,
0.62060546875,
0.884765625,
0.63427734375,
0.51318359375,
0.63427734375
],
[
0.118408203125,
0.63525390625,
0.48779296875,
0.63525390625,
0.48779296875,
0.6484375,
0.118408203125,
0.6484375
],
[
0.51318359375,
0.6357421875,
0.884765625,
0.6357421875,
0.884765625,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.118408203125,
0.65185546875,
0.489501953125,
0.65087890625,
0.489501953125,
0.6650390625,
0.118408203125,
0.66650390625
],
[
0.5146484375,
0.65380859375,
0.611328125,
0.65380859375,
0.611328125,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.118408203125,
0.66845703125,
0.489501953125,
0.66845703125,
0.489501953125,
0.681640625,
0.118408203125,
0.681640625
],
[
0.52880859375,
0.66748046875,
0.880859375,
0.66748046875,
0.880859375,
0.68115234375,
0.52880859375,
0.68115234375
],
[
0.12005615234375,
0.68408203125,
0.48779296875,
0.68408203125,
0.48779296875,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.51123046875,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.51123046875,
0.69775390625
],
[
0.118408203125,
0.69970703125,
0.489501953125,
0.69970703125,
0.489501953125,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.51123046875,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.51123046875,
0.71337890625
],
[
0.12176513671875,
0.71728515625,
0.22412109375,
0.71728515625,
0.22412109375,
0.728515625,
0.12176513671875,
0.728515625
],
[
0.51318359375,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.51318359375,
0.732421875,
0.8818359375,
0.73291015625,
0.8818359375,
0.74755859375,
0.51318359375,
0.74658203125
],
[
0.12005615234375,
0.7421875,
0.376220703125,
0.7421875,
0.376220703125,
0.755859375,
0.12005615234375,
0.755859375
],
[
0.5146484375,
0.75,
0.88330078125,
0.75,
0.88330078125,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.1175537109375,
0.7626953125,
0.489501953125,
0.76416015625,
0.489501953125,
0.77783203125,
0.1175537109375,
0.775390625
],
[
0.51318359375,
0.765625,
0.88330078125,
0.765625,
0.88330078125,
0.779296875,
0.51318359375,
0.779296875
],
[
0.1158447265625,
0.779296875,
0.490234375,
0.77978515625,
0.490234375,
0.79443359375,
0.1158447265625,
0.79345703125
],
[
0.51123046875,
0.7822265625,
0.65576171875,
0.7822265625,
0.65576171875,
0.7958984375,
0.51123046875,
0.7958984375
],
[
0.1175537109375,
0.79638671875,
0.486083984375,
0.79638671875,
0.486083984375,
0.81005859375,
0.1175537109375,
0.81005859375
],
[
0.118408203125,
0.8125,
0.490234375,
0.8125,
0.490234375,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51123046875,
0.80810546875,
0.7841796875,
0.80859375,
0.7841796875,
0.82373046875,
0.51123046875,
0.822265625
],
[
0.118408203125,
0.8291015625,
0.490234375,
0.8291015625,
0.490234375,
0.8427734375,
0.118408203125,
0.8427734375
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.84423828125,
0.489501953125,
0.84423828125,
0.489501953125,
0.85791015625,
0.12005615234375,
0.85791015625
],
[
0.51318359375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.1175537109375,
0.85791015625,
0.197265625,
0.85986328125,
0.197265625,
0.87451171875,
0.1175537109375,
0.87255859375
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1351318359375,
0.8798828125,
0.491943359375,
0.8818359375,
0.491943359375,
0.8994140625,
0.1351318359375,
0.89794921875
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.51318359375,
0.892578125
],
[
0.12005615234375,
0.896484375,
0.162841796875,
0.896484375,
0.162841796875,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.51123046875,
0.89453125,
0.88330078125,
0.8935546875,
0.88330078125,
0.90771484375,
0.51123046875,
0.9091796875
]
] |
[
"MT03MT04MT05",
"Ave",
"MT06",
"System",
"NMT",
"1086",
"708",
"989.8",
"1020",
"1145",
"869",
"1023",
"909",
"609",
"852.5",
"88",
"Ours",
"Table 2: The number of unknown words in the re-",
"sults of NMT and our model.",
"System",
"MT03MT04MT05MT06",
"Ave",
"PBMTHPMTNMT",
"Jane",
"Multi+Source +Ensemble",
"E-NMT",
"39.14",
"40.78",
"37.31",
"37.89",
"38.78",
"Jane",
"40.61",
"43.28",
"39.05",
"39.18",
"Figure 2: Translation results (RIBES score) for",
"40.53",
"Ours",
"43.61",
"47.65",
"42.02",
"41.17",
"43.61",
"different machine translation and system combi-",
"nationmethods.",
"Table 3: Translation results (BLEU score) when",
"we replace original NMT with strong E-NMT,",
"and the word embedding dimension is 500 for all",
"which uses ensemble strategy with four NMT",
"source and target word.The network parameters",
"models.All results of system combination are",
"are updated with Adadelta algorithm. We adopt",
"based on strong outputs of E-NMT.",
"beam search with beam size b=10 at test time.",
"As to confusion-network-based system Jane, we",
"use its default configuration and train a 4-gram",
"Figure 2 illustrates experimental results of",
"language model on target data and 10M Xinhua",
"RIBES scores, which demonstrates that our neu-",
"portion of Gigaword corpus.",
"ral combination model outperforms the best result",
"of single MT system and Jane. Additionally, al-",
"4.3Main Results",
"though BLEU point of Jane is higher than single",
"We compare our neural combination system with",
"NMT system, theword order of Jane is worse in",
"the best individual engines, and the state-of-the-",
"terms of RIBES.",
"art traditional combination system Jane (Freitag",
"et al., 2014).Table 1 shows the BLEU of dif-",
"4.5 Rare and Unknown Words Translation",
"ferent models on development data and test data.",
"It is difficult for NMT systems to handle rare",
"The BLEU score of the multi-source neural com-",
"words, because low-frequency words in training",
"bination model is 2.53 higher than the best single",
"data cannot capture latent translation mappings in",
"model HPMT. The source language input gives a",
"neural network model. However, we do not need",
"further improvement of +1.12 BLEU points.",
"to limit the vocabulary in SMT, which are often",
"As shown in Table 1, Jane outperforms the best",
"able to translate rare words in training data. As",
"single MT system by 1.92 BLEU points. However,",
"shown in Table 2, the number of unknown words",
"our neural combination system with source lan-",
"of our proposed model is 137 fewer than original",
"guage gets an improvement of 1.67 BLEU points",
"NMT model.",
"over Jane. Furthermore, when augmenting our",
"Table 4 shows an example of system combina-",
"neural combination system with ensemble decod-",
"tion. The Chinese word zuzhiwang is an out-of-",
"ing 2, it leads to another significant boost of +1.69",
"vocabulary(OOV) for NMT and the baseline NMT",
"BLEU points.",
"cannot correctly translate this word.Although",
"PBMT and HPMT translate this word well, they",
"4.4Word Order of Translation",
"does not conform to the grammar. By combining",
"We evaluate word order bytheautomatic evalua-",
"the merits of NMT and SMT, our model gets the",
"tion metrics RIBES (Isozaki et al., 2010), whose",
" correct translation.",
"score is a metric based on rank correlation coef",
"ficients with word precision. RIBES is known to",
"4.6Effect of EnsembleDecoding",
"have stronger correlation with human judgements",
"The performance of candidate systems is very im-",
"than BLEU for English as discussed in Isozaki et",
"portant to the result of system combination, and",
"al. (2010).",
"we use ensemble strategy with four NMT mod-",
"2we use four neural combination models in ensemble",
"els to improve the performance of original NMT",
"model.",
"system. As shown in Table 3, the E-NMT with"
] |
{
"arxiv_id": "1704.06393",
"language": "en",
"page_idx": 3
}
|
||
1704.06393-en-4
|
1704.06393-en-4.jpg
|
<table>
<tbody>
<tr>
<td>Source</td>
<td>海珊 协与 恐怖 1 红头 XX 建立了联系。</td>
</tr>
<tr>
<td>Reference</td>
<td>hussein networks 1 has also stahl</td>
</tr>
<tr>
<td>PBMT</td>
<td>hussein</td>
</tr>
<tr>
<td>HPMT</td>
<td>lations hilssein $m$</td>
</tr>
<tr>
<td>$NMT$</td>
<td>hilssein a With LINK</td>
</tr>
<tr>
<td>Jane</td>
<td>hussein s with $|2|10ns$</td>
</tr>
<tr>
<td>Multi</td>
<td>hilsser rom</td>
</tr>
</tbody>
</table>
Table 4: Translation examples of single system, Jane and our proposed model.
ensemble strategy outperforms the original NMT system by+1.40 BLEU points, and it has become the best sytem in all MT systems, which is +0.68 BLEU points higher than HPMT.
After replacing original NMT with strong ENMT , Jane outperforms original result by +0.45 BLEU points, and our model gets an improvement of +3.08 BLEU points over Jane. Experiments further demonstrate that our proposed model is effective and robust for system combination.
based SMT to pre-translate the inputs into target translations. Then a NMT system generates the final hypothesis using the pre-translation. Moreover, multi-source MT has been proved to be very effective to combine multiple source languages (Och and Ney, 2001; Zoph and Knight, 2016; Firat et al., 2016a,b; Garmash and Monz, 2016). Unlike previous works, we adapt multi-source NMT for system combination and design a good strategy to simulate the real training data for our neural system combination.
### 5 $\textbf{Related Work}$
### 6 Conclusion and Future Work
In this paper, we propose a novel neural system combination framework for machine translation. The central idea is to take advantage of NMT and SMT by adapting the multi-source NMT model. The neural system combination method cannot only address the fluency of NMT and the adequacy of SMT, but also can accommodate the source sentences as input. Furthermore, our approach can further use ensemble decoding to boost the performance compared to traditional system combination methods
Experiments on Chinese-English datasets show that our approaches obtain significant improvements over the best individual system and the state-of-the-art traditional system combination methods. In the future work, we plan to encode nbest translation results to further improve the system combination quality. Additionally, it is interesting to extend this approach to other tasks like sentence compression and text abstraction.
The recently proposed neural machine translation has drawn more and more attention. Most of the existing approaches and models mainly focus on designing better attention models (Luong et al., 2015a; Mi et al., 2016a,b; Tu et al., 2016; Meng et al., 2016), better strategies for handling rare and unknown words (Luong et al., 2015b; Li et al., 2016; Zhang and Zong, 2016a; Sennrich et al., 2016b) , exploiting large-scale monolingual data (Cheng et al., 2016; Sennrich et al., 2016a; Zhang and Zong, 2016b), and integrating SMT techniques (Shen et al., 2016; Junczys-Dowmunt et al., 2016b; He et al., 2016).
Our focus in this work is aiming to take advantage of NMT and SMT by system combination, which attempts to find consensus translations among different machine translation systems. In past several years, word-level, phrase-level and sentence-level system combination methods were well studied (Bangalore et al., 2001; Rosti et al., 2008; Li and Zong, 2008; Li et al., 2009; Heafield and Lavie, 2010; Freitag et al., 2014; Ma and Mckeown, 2015; Zhu et al., 2016), and reported stateof-the-art performances in benchmarks for SMT. Here, we propose a neural system combination model which combines the advantages of NMT and SMT efficiently.
Recently, Niehues et al.(2016) use phrase-
## Acknowledgments
The research work has been funded by the Natural Science Foundation of China under Grant No. 61333018 and No. 61673380, and it is also supported by the Strategic Priority Research Program of the CAS under Grant No. XDB02070007.
|
[
[
0.2225341796875,
0.0748291015625,
0.280517578125,
0.0748291015625,
0.280517578125,
0.08905029296875,
0.2225341796875,
0.08905029296875
],
[
0.316650390625,
0.0736083984375,
0.62109375,
0.07244873046875,
0.62109375,
0.0872802734375,
0.316650390625,
0.0885009765625
],
[
0.22412109375,
0.0885009765625,
0.27783203125,
0.09027099609375,
0.276123046875,
0.105712890625,
0.2225341796875,
0.10394287109375
],
[
0.316650390625,
0.0914306640625,
0.68408203125,
0.0914306640625,
0.68408203125,
0.1051025390625,
0.316650390625,
0.1051025390625
],
[
0.22412109375,
0.10687255859375,
0.307373046875,
0.10687255859375,
0.307373046875,
0.1217041015625,
0.22412109375,
0.1217041015625
],
[
0.31494140625,
0.10687255859375,
0.7314453125,
0.10687255859375,
0.7314453125,
0.12054443359375,
0.31494140625,
0.12054443359375
],
[
0.2225341796875,
0.1258544921875,
0.279541015625,
0.1258544921875,
0.279541015625,
0.14013671875,
0.2225341796875,
0.14013671875
],
[
0.31494140625,
0.1270751953125,
0.73876953125,
0.1270751953125,
0.73876953125,
0.14013671875,
0.31494140625,
0.14013671875
],
[
0.22412109375,
0.141357421875,
0.280517578125,
0.141357421875,
0.280517578125,
0.1561279296875,
0.22412109375,
0.1561279296875
],
[
0.316650390625,
0.142578125,
0.7119140625,
0.142578125,
0.7119140625,
0.15673828125,
0.316650390625,
0.15673828125
],
[
0.219970703125,
0.15673828125,
0.27197265625,
0.15673828125,
0.27197265625,
0.1728515625,
0.219970703125,
0.1728515625
],
[
0.316650390625,
0.15673828125,
0.64990234375,
0.15673828125,
0.64990234375,
0.17041015625,
0.316650390625,
0.17041015625
],
[
0.219970703125,
0.174560546875,
0.264404296875,
0.174560546875,
0.264404296875,
0.189453125,
0.219970703125,
0.189453125
],
[
0.318115234375,
0.177001953125,
0.634765625,
0.177001953125,
0.634765625,
0.18701171875,
0.318115234375,
0.18701171875
],
[
0.2225341796875,
0.190673828125,
0.270263671875,
0.190673828125,
0.270263671875,
0.204833984375,
0.2225341796875,
0.204833984375
],
[
0.31494140625,
0.190673828125,
0.7734375,
0.191162109375,
0.7734375,
0.2060546875,
0.31494140625,
0.204833984375
],
[
0.21240234375,
0.218505859375,
0.78662109375,
0.218505859375,
0.78662109375,
0.2327880859375,
0.21240234375,
0.2327880859375
],
[
0.1175537109375,
0.25830078125,
0.489501953125,
0.25830078125,
0.489501953125,
0.271484375,
0.1175537109375,
0.271484375
],
[
0.51318359375,
0.25830078125,
0.884765625,
0.25830078125,
0.884765625,
0.271484375,
0.51318359375,
0.271484375
],
[
0.118408203125,
0.273681640625,
0.490234375,
0.273681640625,
0.490234375,
0.2880859375,
0.118408203125,
0.2880859375
],
[
0.51318359375,
0.27490234375,
0.884765625,
0.27490234375,
0.884765625,
0.2880859375,
0.51318359375,
0.2880859375
],
[
0.118408203125,
0.289306640625,
0.48779296875,
0.289306640625,
0.48779296875,
0.302734375,
0.118408203125,
0.302734375
],
[
0.5146484375,
0.290283203125,
0.8818359375,
0.290283203125,
0.8818359375,
0.3046875,
0.5146484375,
0.3046875
],
[
0.118408203125,
0.305908203125,
0.366943359375,
0.305908203125,
0.366943359375,
0.320556640625,
0.118408203125,
0.320556640625
],
[
0.51318359375,
0.30712890625,
0.8818359375,
0.30712890625,
0.8818359375,
0.320556640625,
0.51318359375,
0.320556640625
],
[
0.1368408203125,
0.322509765625,
0.489501953125,
0.32177734375,
0.489501953125,
0.336181640625,
0.1368408203125,
0.33740234375
],
[
0.51318359375,
0.322509765625,
0.88330078125,
0.322509765625,
0.88330078125,
0.336181640625,
0.51318359375,
0.336181640625
],
[
0.118408203125,
0.339111328125,
0.489501953125,
0.339111328125,
0.489501953125,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.51318359375,
0.33837890625,
0.884765625,
0.33837890625,
0.884765625,
0.3515625,
0.51318359375,
0.3515625
],
[
0.118408203125,
0.35400390625,
0.490234375,
0.355224609375,
0.490234375,
0.369384765625,
0.118408203125,
0.3681640625
],
[
0.5146484375,
0.355712890625,
0.88330078125,
0.355712890625,
0.88330078125,
0.366455078125,
0.5146484375,
0.366455078125
],
[
0.118408203125,
0.371826171875,
0.49365234375,
0.371826171875,
0.49365234375,
0.385498046875,
0.118408203125,
0.385498046875
],
[
0.51123046875,
0.37060546875,
0.884765625,
0.369384765625,
0.884765625,
0.38427734375,
0.51123046875,
0.385498046875
],
[
0.1175537109375,
0.385986328125,
0.490234375,
0.38720703125,
0.490234375,
0.402099609375,
0.1175537109375,
0.40087890625
],
[
0.51318359375,
0.38720703125,
0.884765625,
0.38720703125,
0.884765625,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.118408203125,
0.40380859375,
0.411376953125,
0.40380859375,
0.411376953125,
0.41748046875,
0.118408203125,
0.41748046875
],
[
0.51318359375,
0.402587890625,
0.8818359375,
0.402587890625,
0.8818359375,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.51318359375,
0.419189453125,
0.646484375,
0.419189453125,
0.646484375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.118408203125,
0.431640625,
0.27197265625,
0.431640625,
0.27197265625,
0.446533203125,
0.118408203125,
0.446533203125
],
[
0.515625,
0.448974609375,
0.54248046875,
0.448974609375,
0.54248046875,
0.457275390625,
0.515625,
0.457275390625
],
[
0.5380859375,
0.446533203125,
0.79833984375,
0.446533203125,
0.79833984375,
0.460205078125,
0.5380859375,
0.460205078125
],
[
0.1175537109375,
0.458984375,
0.489501953125,
0.457763671875,
0.489501953125,
0.47265625,
0.1175537109375,
0.473876953125
],
[
0.118408203125,
0.4755859375,
0.490234375,
0.4755859375,
0.490234375,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.51318359375,
0.471435546875,
0.88330078125,
0.471435546875,
0.88330078125,
0.486328125,
0.51318359375,
0.486328125
],
[
0.1158447265625,
0.4912109375,
0.489501953125,
0.490478515625,
0.489501953125,
0.5048828125,
0.1158447265625,
0.505859375
],
[
0.51318359375,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.501953125,
0.51318359375,
0.501953125
],
[
0.118408203125,
0.50830078125,
0.48779296875,
0.50830078125,
0.48779296875,
0.521484375,
0.118408203125,
0.521484375
],
[
0.5146484375,
0.50439453125,
0.88330078125,
0.50439453125,
0.88330078125,
0.5185546875,
0.5146484375,
0.5185546875
],
[
0.12005615234375,
0.52392578125,
0.489501953125,
0.52392578125,
0.489501953125,
0.53759765625,
0.12005615234375,
0.53759765625
],
[
0.5146484375,
0.51953125,
0.8818359375,
0.51953125,
0.8818359375,
0.53369140625,
0.5146484375,
0.53369140625
],
[
0.12005615234375,
0.54150390625,
0.48779296875,
0.54150390625,
0.48779296875,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.51318359375,
0.5361328125,
0.884765625,
0.5361328125,
0.884765625,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.118408203125,
0.5556640625,
0.489501953125,
0.5556640625,
0.489501953125,
0.5693359375,
0.118408203125,
0.5693359375
],
[
0.51123046875,
0.55029296875,
0.8818359375,
0.5517578125,
0.8818359375,
0.56640625,
0.51123046875,
0.5654296875
],
[
0.118408203125,
0.57177734375,
0.490234375,
0.57177734375,
0.490234375,
0.5859375,
0.118408203125,
0.5859375
],
[
0.51318359375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.58203125,
0.51318359375,
0.58203125
],
[
0.12005615234375,
0.58837890625,
0.490234375,
0.58837890625,
0.490234375,
0.6015625,
0.12005615234375,
0.6015625
],
[
0.51318359375,
0.58447265625,
0.88330078125,
0.58447265625,
0.88330078125,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.12005615234375,
0.60400390625,
0.490234375,
0.60400390625,
0.490234375,
0.61767578125,
0.12005615234375,
0.61767578125
],
[
0.51318359375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.6142578125,
0.51318359375,
0.6142578125
],
[
0.12005615234375,
0.62060546875,
0.48779296875,
0.62060546875,
0.48779296875,
0.63525390625,
0.12005615234375,
0.63525390625
],
[
0.51318359375,
0.61767578125,
0.8818359375,
0.61767578125,
0.8818359375,
0.63134765625,
0.51318359375,
0.63134765625
],
[
0.12005615234375,
0.6357421875,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.12005615234375,
0.65087890625
],
[
0.51123046875,
0.6328125,
0.6337890625,
0.6328125,
0.6337890625,
0.646484375,
0.51123046875,
0.646484375
],
[
0.12005615234375,
0.65283203125,
0.341796875,
0.65283203125,
0.341796875,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.53076171875,
0.6484375,
0.88330078125,
0.64794921875,
0.88330078125,
0.662109375,
0.53076171875,
0.6630859375
],
[
0.1368408203125,
0.66943359375,
0.484375,
0.66943359375,
0.484375,
0.68310546875,
0.1368408203125,
0.68310546875
],
[
0.51318359375,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6796875,
0.51318359375,
0.6796875
],
[
0.1175537109375,
0.68408203125,
0.489501953125,
0.68505859375,
0.489501953125,
0.69970703125,
0.1175537109375,
0.69873046875
],
[
0.5146484375,
0.68310546875,
0.88330078125,
0.68310546875,
0.88330078125,
0.693359375,
0.5146484375,
0.693359375
],
[
0.12005615234375,
0.70166015625,
0.489501953125,
0.70166015625,
0.489501953125,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.51318359375,
0.69775390625,
0.884765625,
0.69775390625,
0.884765625,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.12005615234375,
0.71728515625,
0.490234375,
0.71728515625,
0.490234375,
0.73095703125,
0.12005615234375,
0.73095703125
],
[
0.51318359375,
0.71337890625,
0.88330078125,
0.71337890625,
0.88330078125,
0.72705078125,
0.51318359375,
0.72705078125
],
[
0.12005615234375,
0.73291015625,
0.489501953125,
0.73291015625,
0.489501953125,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.51318359375,
0.72998046875,
0.88330078125,
0.72998046875,
0.88330078125,
0.74365234375,
0.51318359375,
0.74365234375
],
[
0.12005615234375,
0.75,
0.490234375,
0.75,
0.490234375,
0.76318359375,
0.12005615234375,
0.76318359375
],
[
0.51318359375,
0.7451171875,
0.88330078125,
0.7451171875,
0.88330078125,
0.7587890625,
0.51318359375,
0.7587890625
],
[
0.12005615234375,
0.765625,
0.490234375,
0.765625,
0.490234375,
0.779296875,
0.12005615234375,
0.779296875
],
[
0.51318359375,
0.7626953125,
0.884765625,
0.7626953125,
0.884765625,
0.775390625,
0.51318359375,
0.775390625
],
[
0.12005615234375,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.779296875,
0.8310546875,
0.779296875,
0.8310546875,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.118408203125,
0.79638671875,
0.48779296875,
0.79736328125,
0.48779296875,
0.8125,
0.118408203125,
0.81103515625
],
[
0.515625,
0.8046875,
0.67333984375,
0.8046875,
0.67333984375,
0.8193359375,
0.515625,
0.8193359375
],
[
0.118408203125,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.1175537109375,
0.830078125,
0.489501953125,
0.8291015625,
0.489501953125,
0.84326171875,
0.1175537109375,
0.84423828125
],
[
0.51123046875,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.51123046875,
0.84423828125
],
[
0.118408203125,
0.8466796875,
0.491943359375,
0.8466796875,
0.491943359375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51123046875,
0.8466796875,
0.8818359375,
0.845703125,
0.8818359375,
0.85986328125,
0.51123046875,
0.86083984375
],
[
0.1175537109375,
0.8623046875,
0.490234375,
0.8623046875,
0.490234375,
0.876953125,
0.1175537109375,
0.876953125
],
[
0.51318359375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87744140625,
0.273681640625,
0.87744140625,
0.273681640625,
0.892578125,
0.118408203125,
0.892578125
],
[
0.51123046875,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.51123046875,
0.8935546875
],
[
0.1368408203125,
0.89453125,
0.490234375,
0.89453125,
0.490234375,
0.9091796875,
0.1368408203125,
0.9091796875
],
[
0.51318359375,
0.89453125,
0.8447265625,
0.89453125,
0.8447265625,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"Source",
"海珊也与恐怖组织网建立了联系",
"Pinyin",
"hanshan ye yu kongbu zuzhiwang jianli le lianxi ",
"Reference",
"hussein has also established ties with terrorist networks",
"PBMT",
"hussein also has established relations and terrorist group",
"HPMT",
"hussein also and terrorist group established relations .",
"NMT",
"hussein also established relations with UNK",
"Jane",
"husseinalsohasestablishedrelationswith",
"Multi",
"hussein also has established relations with the terrorist group.",
"Table 4: Translation examples of single system, Jane and our proposed model.",
"ensemble strategy outperforms the original NMT",
"based SMT to pre-translate the inputs into target",
"system by +1.40 BLEU points, and it has become",
"translations. Then a NMT system generates the",
"the best sytem in all MT systems, which is +0.68",
"final hypothesis using the pre-translation. More-",
"BLEU points higher than HPMT.",
"over, multi-source MT has been proved to be very",
"After replacing original NMT with strong E-",
"effective to combine multiple source languages",
"NMT , Jane outperforms original result by +0.45",
"(Och and Ney,2001; Zoph and Knight, 2016; Firat",
"BLEU points, and our model gets an improvement",
"et al..2016a.b: Garmash and Monz.2016). Unlike",
"of +3.08 BLEU points over Jane. Experiments fur-",
"previous works, we adapt multi-source NMT for",
"ther demonstrate that our proposed model is effec-",
"system combination and design a good strategy to",
"tive and robust for system combination.",
"simulate the real training data for our neural sys-",
"tem combination.",
"5 Related Work",
"6",
"Conclusion and Future Work",
"The recently proposed neural machine translation",
"has drawn more and more attention.Most of",
"In this paper, we propose a novel neural system",
"the existing approaches and models mainly fo-",
"combination framework for machine translation.",
"cus on designing better attention models (Luong",
"The central idea is to take advantage of NMT and",
"et al., 2015a; Mi et al., 2016a,b; Tu et al., 2016;",
"SMT by adapting the multi-source NMT model.",
"Meng etal.,2016),betterstrategies forhandling",
"The neural system combination method cannot",
"rare and unknown words (Luong et al., 2015b; Li",
"only address the fluency of NMT and the adequacy",
"et al., 2016; Zhang and Zong, 2016a; Sennrich",
"of SMT, but also can accommodate the source sen-",
"et al., 2016b), exploiting large-scale monolingual",
"tences as input. Furthermore, our approach can",
"data (Cheng et al., 2016; Sennrich et al., 2016a;",
"further use ensemble decoding to boost the per-",
"Zhang and Zong, 2016b), and integrating SMT",
"formance compared to traditional system combi-",
"techniques (Shen et al., 2016; Junczys-Dowmunt",
"nation methods.",
"et al., 2016b; He et al., 2016).",
"Experiments on Chinese-English datasets show",
"Our focus in this work is aiming to take ad.",
"that our approaches obtain significant improve-",
"vantage of NMT and SMT by system combina-",
"ments over the best individual system and the",
"tion, which attempts to find consensus translations",
"state-of-the-art traditional system combination",
"among different machine translation systems. In",
"methods. In the future work, we plan to encode n-",
"past several years, word-level, phrase-level and",
"best translation results to further improve the sys-",
"sentence-level system combination methods were",
"tem combination quality: Additionally, it is inter-",
"well studied (Bangalore et al., 2001; Rosti et al.,",
"esting to extend this approach to other tasks like",
"2008; Li and Zong, 2008; Li et al., 2009; Heafield",
"sentence compression and text abstraction.",
"and Lavie, 2010; Freitag et al., 2014; Ma and Mck-",
"Acknowledgments",
"eown, 2015; Zhu et al., 2016), and reported state-",
"of-the-art performances in benchmarks for SMT.",
"The research work has been funded by the Natu-",
"Here, we propose a neural system combination",
"ral Science Foundation of China under Grant No.",
"model which combines the advantages of NMT",
"61333018 and No. 61673380, and it is also sup-",
"and SMT efficiently.",
"ported by the Strategic Priority Research Program",
"Recently, Niehues et al. (2016) use phrase-",
"of the CAS under Grant No. XDB02070007."
] |
{
"arxiv_id": "1704.06393",
"language": "en",
"page_idx": 4
}
|
||
1904.02242-en-0
|
1904.02242-en-0.jpg
|
# Unpaired Thermal to Visible Spectrum Transfer using Adversarial Training
$\mathrm{Adam~Nyberg}^{1[ 0000- 0001- 8764- 8499] }$, Abdelrahman
Eldesokey$^1[0000-0003-3292-7153]$, David Bergström$^{2[0000-0003-2414-4482]}$, and
$\begin{array}{c}{,\mathrm{~,~Dav~u~Dergst10ll~}}\\{\mathrm{David~Gustafsson}^{2[0000-0002-4370-2286]}}\end{array}$
$^{1}$ Computer Vision Laboratory Linköping University, 581 83 Linköping, Sweden
$^{2}$ Swedish Defence Research Agency (FOI), 583 30 Linköping, Sweden
$\mathbf{Abstract.~Thermal~Infrared~(TIR)~cameras~are~gaining~popularity~in}$ many computer vision applications due to their ability to operate under low-light conditions. Images produced by TIR cameras are usually difficult for humans to perceive visually, which limits their usability. Several methods in the literature were proposed to address this problem by transforming TIR images into realistic visible spectrum (VIS) images. However, existing TIR-VIS datasets suffer from imperfect alignment between TIR-VIS image pairs which degrades the performance of supervised methods. We tackle this problem by learning this transformation using an unsupervised Generative Adversarial Network (GAN) which trains on unpaired TIR and VIS images. When trained and evaluated on KAIST-MS dataset, our proposed methods was shown to produce significantly more realistic and sharp VIS images than the existing stateof-the-art supervised methods. In addition, our proposed method was shown to generalize very well when evaluated on a new dataset of new environments.
Keywords: Thermal Imaging $\cdot$Generative Adversarial Networks·Un-
supervised Learning $\cdot$ Colorization
## 1 Introduction
Recently, thermal infrared ( TIR) cameras have become increasingly popular due to their long wavelength which allows them to work under low-light conditions. TIR cameras require no active illumination as they sense emitted heat from objects and map it to a visual heat map. This opens up for many applications such as object detection for driving in complete darkness and event detection im surveillance. In addition, the cost of TIR cameras have gone significantly down while their resolution have improved significantly, resulting in a boost of interest. However, one limitation of TIR cameras is their limited visual interpretability for humans which hinders some applications such as visual-aided driving.
To address this problem, TIR images can be transformed to visible spectrum (VIS) images which are easily interpreted by humans. Figure 1 shows an example of a TIR image, the corresponding VIS image and the VIS image generated
|
[
[
0.2254638671875,
0.146484375,
0.78125,
0.146484375,
0.78125,
0.1610107421875,
0.2254638671875,
0.1610107421875
],
[
0.346435546875,
0.1654052734375,
0.65869140625,
0.1666259765625,
0.65869140625,
0.1875,
0.346435546875,
0.186279296875
],
[
0.3203125,
0.2083740234375,
0.6845703125,
0.2083740234375,
0.6845703125,
0.231689453125,
0.3203125,
0.231689453125
],
[
0.2296142578125,
0.2247314453125,
0.775390625,
0.2247314453125,
0.775390625,
0.24560546875,
0.2296142578125,
0.24560546875
],
[
0.361083984375,
0.242431640625,
0.64208984375,
0.238037109375,
0.64208984375,
0.2568359375,
0.361083984375,
0.261474609375
],
[
0.0196075439453125,
0.26513671875,
0.06781005859375,
0.26513671875,
0.06781005859375,
0.40478515625,
0.0196075439453125,
0.40478515625
],
[
0.2344970703125,
0.267578125,
0.76708984375,
0.26953125,
0.76708984375,
0.287353515625,
0.2344970703125,
0.285400390625
],
[
0.268798828125,
0.285888671875,
0.736328125,
0.285888671875,
0.736328125,
0.2998046875,
0.268798828125,
0.2998046875
],
[
0.2646484375,
0.3359375,
0.740234375,
0.337158203125,
0.740234375,
0.3515625,
0.2646484375,
0.35107421875
],
[
0.2646484375,
0.3515625,
0.73876953125,
0.3515625,
0.73876953125,
0.365478515625,
0.2646484375,
0.365478515625
],
[
0.2646484375,
0.3662109375,
0.73876953125,
0.3662109375,
0.73876953125,
0.380126953125,
0.2646484375,
0.380126953125
],
[
0.2646484375,
0.380126953125,
0.73876953125,
0.380126953125,
0.73876953125,
0.393310546875,
0.2646484375,
0.393310546875
],
[
0.0228729248046875,
0.391357421875,
0.062103271484375,
0.391357421875,
0.062103271484375,
0.490478515625,
0.0228729248046875,
0.490478515625
],
[
0.041656494140625,
0.396484375,
0.049835205078125,
0.396484375,
0.049835205078125,
0.399658203125,
0.041656494140625,
0.399658203125
],
[
0.2646484375,
0.393310546875,
0.73876953125,
0.393310546875,
0.73876953125,
0.4072265625,
0.2646484375,
0.4072265625
],
[
0.2646484375,
0.40771484375,
0.73779296875,
0.40771484375,
0.73779296875,
0.421630859375,
0.2646484375,
0.421630859375
],
[
0.2646484375,
0.421630859375,
0.73876953125,
0.421630859375,
0.73876953125,
0.43505859375,
0.2646484375,
0.43505859375
],
[
0.2646484375,
0.43505859375,
0.73876953125,
0.43505859375,
0.73876953125,
0.448974609375,
0.2646484375,
0.448974609375
],
[
0.2646484375,
0.449462890625,
0.740234375,
0.449462890625,
0.740234375,
0.46337890625,
0.2646484375,
0.46337890625
],
[
0.265625,
0.464111328125,
0.73876953125,
0.464111328125,
0.73876953125,
0.474853515625,
0.265625,
0.474853515625
],
[
0.2646484375,
0.4765625,
0.740234375,
0.4765625,
0.740234375,
0.490478515625,
0.2646484375,
0.490478515625
],
[
0.021240234375,
0.48486328125,
0.058013916015625,
0.48486328125,
0.058013916015625,
0.70556640625,
0.021240234375,
0.70556640625
],
[
0.2646484375,
0.490478515625,
0.73876953125,
0.490478515625,
0.73876953125,
0.50390625,
0.2646484375,
0.50390625
],
[
0.2646484375,
0.50390625,
0.73876953125,
0.50390625,
0.73876953125,
0.517578125,
0.2646484375,
0.517578125
],
[
0.2646484375,
0.51806640625,
0.740234375,
0.51806640625,
0.740234375,
0.5322265625,
0.2646484375,
0.5322265625
],
[
0.2646484375,
0.5341796875,
0.73876953125,
0.5341796875,
0.73876953125,
0.54443359375,
0.2646484375,
0.54443359375
],
[
0.265625,
0.546875,
0.35693359375,
0.546875,
0.35693359375,
0.556640625,
0.265625,
0.556640625
],
[
0.2646484375,
0.57373046875,
0.73779296875,
0.57373046875,
0.73779296875,
0.5869140625,
0.2646484375,
0.5869140625
],
[
0.26318359375,
0.5869140625,
0.495849609375,
0.5859375,
0.495849609375,
0.60107421875,
0.26318359375,
0.6015625
],
[
0.2181396484375,
0.62890625,
0.24267578125,
0.62890625,
0.24267578125,
0.64013671875,
0.2181396484375,
0.64013671875
],
[
0.2459716796875,
0.62744140625,
0.374267578125,
0.62744140625,
0.374267578125,
0.642578125,
0.2459716796875,
0.642578125
],
[
0.21728515625,
0.6591796875,
0.7861328125,
0.66015625,
0.7861328125,
0.6748046875,
0.21728515625,
0.673828125
],
[
0.2156982421875,
0.6748046875,
0.78369140625,
0.67529296875,
0.78369140625,
0.68994140625,
0.2156982421875,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.21728515625,
0.7197265625,
0.7861328125,
0.72021484375,
0.7861328125,
0.73486328125,
0.21728515625,
0.734375
],
[
0.2181396484375,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.751953125,
0.78369140625,
0.751953125,
0.78369140625,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.2181396484375,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.740234375,
0.78076171875,
0.740234375,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.24267578125,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81103515625,
0.24267578125,
0.81103515625
],
[
0.2181396484375,
0.81005859375,
0.7861328125,
0.81103515625,
0.7861328125,
0.82568359375,
0.2181396484375,
0.82470703125
],
[
0.2156982421875,
0.82568359375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84130859375,
0.2156982421875,
0.84033203125
]
] |
[
"Unpaired Thermal to Visible Spectrum Transfer",
"using Adversarial Training",
"Adam Nyberg 000-0001-8764-8499], Abdelrahman",
"Eldeskey000-0003-3292713],David Bergstrom2000-0003-24144482], and",
"David Gustafsson2000-002-4370-2286]",
"610",
"1 Computer Vision Laboratory Linkoping University, 581 83 Linkoping, Sweden",
"2 Swedish Defence Research Agency (FOI), 583 30 Linkoping, Sweden",
"Abstract. Thermal Infrared (TIR) cameras are gaining popularity in",
"many computer vision applications due to their ability to operate un-",
"der low-light conditions. Images produced by TIR cameras are usually",
"difficult for humans to perceive visually, which limits their usability. Sev-",
"['s]",
"",
"eral methods in the literature were proposed to address this problem by",
"transforming TIR images into realistic visible spectrum (VIS) images.",
"However, existing TIR-VIS datasets suffer from imperfect alignment be-",
"tween TIR-VIS image pairs which degrades the performance of super-",
"vised methods. We tackle this problem by learning this transformation",
"using an unsupervised Generative Adversarial Network (GAN) which",
"trains on unpaired TIR and VIS images. When trained and evaluated on",
"2",
"KAIST-MS dataset, our proposed methods was shown to produce sig-",
"nificantly more realistic and sharp VIS images than the existing state-",
"of-the-art supervised methods. In addition, our proposed method was",
"shown to generalizevery well when evaluated on a new dataset of new",
"environments.",
"Keywords: Thermal Imaging : Generative Adversarial Networks -Un-",
"supervised Learning·Colorization",
"1 ",
"Introduction",
"Recently, thermal infrared (TIR) cameras have become increasingly popular due",
"to their long wavelength which allows them to work under low-light conditions.",
"TIR cameras require no active illumination as they sense emitted heat from",
"objects and map it to a visual heat map. This opens up for many applications",
"such as object detection for driving in complete darkness and event detection in",
"surveillance. In addition, the cost of TIR cameras have gone significantly down",
"while their resolution have improved significantly,resulting in a boost of interest.",
"However, one limitation of TIR cameras is their limited visual interpretability",
"for humans which hinders some applications such as visual-aided driving.",
"To address this problem, TIR images can be transformed to visible spectrum",
"(VIS) images which are easily interpreted by humans. Figure 1 shows an example",
"of a TIR image, the corresponding VIS image and the VIS image generated"
] |
{
"arxiv_id": "1904.02242",
"language": "en",
"page_idx": 0
}
|
||
1904.02242-en-1
|
1904.02242-en-1.jpg
|
2
A. Nyberg et al.
<FigureHere>
(a) TIR image
( b) Target VIS
( c) Generated VIS
Fig.1: An example of a TIR image (a), its corresponding VIS image (b) from the KAIST-MS dataset [4] and the VIS image (c) generated by our proposed method using only the TIR image (a) as an input.
directly from the TIR image. This is similar to colorization problems, where grayscale VIS images are mapped to color VIS images. However, transforming TIR images to VIS images is inherently challenging as they are not correlated in the electromagnetic spectrum. For instance, two objects of the same material and temperature, but with different colors in the VIS image, could correspond to the same value in the TIR image. Consequently, utilizing all the available information, i.e. spectrum, shape and context, is very crucial when solving this task. This also requires the availability of enormous amount of data to learn the latent relations between the two spectrums.
In colorization problems, only the chrominance needs to be estimated as the luminance is already available from the input grayscale images. Contrarily, TIR to VIS transformation requires the estimation of both the luminance and the chrominance based on the semantics of the input TIR images. Besides, generating data for learning colorization models is easy as color images could be computationally transformed to grayscale images to create image pairs with perfect pixel-to-pixel correspondences. In contrast, datasets containing registered TIR/ VIS image pairs are very few and requires a sophisticated acquisition systems for good pixel-to-pixel correspondence.
The KAIST Multispectral Pedestrian Detection Benchmark (KAIST-MS) [4] introduced the first large-scale dataset with TIR-VIS image pairs. However, it was shown by [1] that the TIR-VIS image pairs in KAIST-MS does not have a perfect pixel-to-pixel correspondence, with a pixel error of up to 16 pixels $(5\%)$ in the horizontal direction. This would degrade the performance of supervised learning methods which tries to learn the pixel-to-pixel correspondences between image pairs and leads to corrupted output. To our knowledge there exist no large-scale public dataset of TIR-VIS image pairs with perfect pixel to pixel correspondence. Therefore, the method used for TIR to VIS image transformation need to control for this imperfection.
|
[
[
0.21728515625,
0.11614990234375,
0.2303466796875,
0.11614990234375,
0.2303466796875,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.271240234375,
0.11553955078125,
0.381591796875,
0.11553955078125,
0.381591796875,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.268798828125,
0.30615234375,
0.365234375,
0.30615234375,
0.365234375,
0.32080078125,
0.268798828125,
0.32080078125
],
[
0.451904296875,
0.30615234375,
0.55224609375,
0.30615234375,
0.55224609375,
0.32080078125,
0.451904296875,
0.32080078125
],
[
0.6259765625,
0.30615234375,
0.7509765625,
0.304931640625,
0.7509765625,
0.3193359375,
0.6259765625,
0.32080078125
],
[
0.2181396484375,
0.333251953125,
0.78515625,
0.333251953125,
0.78515625,
0.347900390625,
0.2181396484375,
0.347900390625
],
[
0.21728515625,
0.347900390625,
0.7861328125,
0.348388671875,
0.7861328125,
0.363037109375,
0.21728515625,
0.3623046875
],
[
0.21728515625,
0.363037109375,
0.57861328125,
0.363037109375,
0.57861328125,
0.376953125,
0.21728515625,
0.376953125
],
[
0.2181396484375,
0.41357421875,
0.7861328125,
0.41357421875,
0.7861328125,
0.427978515625,
0.2181396484375,
0.427978515625
],
[
0.2181396484375,
0.4287109375,
0.78515625,
0.4287109375,
0.78515625,
0.443115234375,
0.2181396484375,
0.443115234375
],
[
0.2181396484375,
0.442626953125,
0.7861328125,
0.443115234375,
0.7861328125,
0.458251953125,
0.2181396484375,
0.45703125
],
[
0.2181396484375,
0.460205078125,
0.7861328125,
0.460205078125,
0.7861328125,
0.473388671875,
0.2181396484375,
0.473388671875
],
[
0.2181396484375,
0.474853515625,
0.7861328125,
0.474853515625,
0.7861328125,
0.488037109375,
0.2181396484375,
0.488037109375
],
[
0.21728515625,
0.490478515625,
0.7861328125,
0.490478515625,
0.7861328125,
0.50390625,
0.21728515625,
0.50390625
],
[
0.21728515625,
0.5048828125,
0.78759765625,
0.5048828125,
0.78759765625,
0.51953125,
0.21728515625,
0.51953125
],
[
0.2181396484375,
0.52099609375,
0.7861328125,
0.52099609375,
0.7861328125,
0.5341796875,
0.2181396484375,
0.5341796875
],
[
0.2181396484375,
0.53515625,
0.53125,
0.53515625,
0.53125,
0.548828125,
0.2181396484375,
0.548828125
],
[
0.2418212890625,
0.55029296875,
0.78369140625,
0.55029296875,
0.78369140625,
0.564453125,
0.2418212890625,
0.564453125
],
[
0.21728515625,
0.5673828125,
0.78515625,
0.5673828125,
0.78515625,
0.58203125,
0.21728515625,
0.58203125
],
[
0.21728515625,
0.58203125,
0.78369140625,
0.58203125,
0.78369140625,
0.59521484375,
0.21728515625,
0.59521484375
],
[
0.2181396484375,
0.59765625,
0.78369140625,
0.59765625,
0.78369140625,
0.611328125,
0.2181396484375,
0.611328125
],
[
0.21728515625,
0.61279296875,
0.78369140625,
0.61279296875,
0.78369140625,
0.626953125,
0.21728515625,
0.626953125
],
[
0.2181396484375,
0.62744140625,
0.78369140625,
0.62744140625,
0.78369140625,
0.6416015625,
0.2181396484375,
0.6416015625
],
[
0.2181396484375,
0.64306640625,
0.78369140625,
0.64306640625,
0.78369140625,
0.6572265625,
0.2181396484375,
0.6572265625
],
[
0.2197265625,
0.65771484375,
0.78515625,
0.65771484375,
0.78515625,
0.671875,
0.2197265625,
0.671875
],
[
0.2181396484375,
0.673828125,
0.496826171875,
0.673828125,
0.496826171875,
0.6875,
0.2181396484375,
0.6875
],
[
0.2442626953125,
0.68994140625,
0.78515625,
0.68994140625,
0.78515625,
0.7041015625,
0.2442626953125,
0.7041015625
],
[
0.2156982421875,
0.7041015625,
0.7861328125,
0.705078125,
0.7861328125,
0.7197265625,
0.2156982421875,
0.71826171875
],
[
0.2181396484375,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.734375,
0.2181396484375,
0.734375
],
[
0.2197265625,
0.7373046875,
0.78369140625,
0.7373046875,
0.78369140625,
0.74755859375,
0.2197265625,
0.74755859375
],
[
0.2197265625,
0.75048828125,
0.78369140625,
0.75048828125,
0.78369140625,
0.7646484375,
0.2197265625,
0.7646484375
],
[
0.2197265625,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2197265625,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.21728515625,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.515625,
0.8271484375,
0.515625,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"2",
"A. Nyberg et al.",
"(a) TIR image",
"(b) Target VIS",
"(c) Generated VIS",
"Fig. 1: An example of a TIR image (a), its corresponding VIS image (b) from",
"the KAIST-MS dataset [4] and the VIS image (c) generated by our proposed",
"method using only the TIR image (a) as an input.",
"directly from the TIR image. This is similar to colorization problems, where",
"grayscale VIS images are mapped to color VIS images. However, transforming",
"TIR images to VIS images is inherently challenging as they are not correlated",
"intheelectromagnetic spectrum.For instance, two objects of thesamematerial",
"and temperature, but with different colors in the ViS image, could correspond",
"to the same value in the TIR image,Consequently,utilizing all theavailable",
"information, i.e. spectrum, shape and context, is very crucial when solving this",
"task. This also requires the availability of enormous amount of data to learn the",
"latent relations between the two spectrums.",
"In colorization problems, only the chrominance needs to be estimated as the",
"luminance is already available from the input grayscale images. Contrarily, TIR",
"to Vis transformation requires the estimation of both the luminance and the",
"chrominance based on the semantics of the input TIR images. Besides, generat-",
"ing data for learning colorization models is easy as color images could be com-",
"putationally transformed to grayscale images to create image pairs with perfect",
"pixel-to-pixel correspondences. In contrast, datasets containing registered TIR/",
"Vis image pairs are very few and requires a sophisticated acquisition systems",
"for good pixel-to-pixel correspondence.",
"The KAIST Multispectral Pedestrian Detection Benchmark (KAIST-MS) [4l",
"introduced the first large-scale dataset with TIR-VIS image pairs. However, it",
"was shown by [1] that the TIR-VIS image pairs in KAIST-MS does not have",
"a perfect pixel-to-pixel correspondence, with a pixel error of up to l6 pixels",
"(5%) in the horizontal direction. This would degrade the performance of super-",
"vised learning methods which tries to learn the pixel-to-pixel correspondences",
"between image pairs and leads to corrupted output. To our knowledge there exist",
"no large-scale public dataset of TIR-VIS image pairs with perfect pixel to pixel",
"correspondence. Therefore, the method used for TIR to VIS image transforma-",
"tion need to control for this imperfection."
] |
{
"arxiv_id": "1904.02242",
"language": "en",
"page_idx": 1
}
|
||
1904.02242-en-2
|
1904.02242-en-2.jpg
|
Thermal to Visible Spectrum Transfer
3
In this paper, we propose an unsupervised method for transforming TIR images, specifically long-wavelength infrared (LWIR), to visible spectrum (VIS) images. Our method is trained on unpaired images from the KAIST-MS dataset [4] which allows it to handle the imperfect registration between the TIR-VIS image pairs. Qualitative analysis shows that our proposed unsupervised method produces sharp and perceptually realistic VIS images compared to the existing state-of-the-art supervised methods. In addition, our proposed method achieves comparable results to state-of-the-art supervised method in terms of L1 error despite being trained on unpaired images. Finally, our proposed method generalizes very well when evaluated on our new FOI dataset, which demonstrates the generalization capabilities of our method contrarily to the existing state-of-the-art methods.
## 2 Related Work
Colorizing grayscale images has been extensively investigated in the literature. Scribbles [9] requires the user to manually apply strokes of color to different regions of a grayscale image and neighboring pixels with the same luminance should get the same color. Transfer techniques [16] use the color palette from a reference image and apply it to a grayscale image by matching luminance and texture. Both scribbles and transfer techniques require manual input from the user. Recently, colorization methods based on automatic transformation, i.e., the only input to the method is the grayscale image, have become popular. Promising results have been demonstrated in the area of automatic transformation using Convolutional Neural Networks (CNNs) [3,5,8,17] and Generative Adversarial Networks (GANs) [2,6,13] due to their abilities to model semantic representation ${\text{in images.}}$
In the infrared spectrum, less research has been done on transforming thermal images to VIS images. In [10], a CNN-based method was proposed to transform near-infrared (NIR) images to VIS images. Their method was shown to perform well as the NIR and VIS images are highly correlated in the electromagnetic spectrum. Kniaz $et$ al.[7] proposed VIS to TIR transformation using a CNN model as a way to generate synthetic TIR images. The KAIST-MS [11] dataset introduced the frst realistic large-scale dataset of TIR-VIS image pairs which opened up for developing TIR-VIS transformation models. Berg $et\textit{al. [ 1] pro- }$ posed a CNN-based model to transform TIR images to VIS images trained on the KAIST-MS dataset. However, the imperfect registration of the dataset caused the output from their method to be blurry and corrupted in some cases.
Generative Adversarial Networks (GANs) have shown promising results in unsupervised domain transfer [6,12,18,19]. An unsupervised method does not require a paired dataset, hence, eliminating the need for pixel to pixel correspondence. In [14,15], GANs have demonstrated a very good performance on transferring NIR images to VIS images. Isola $et$ al.[6] has shown some qualitative results from the KAIST-MS dataset as and example for domain transfer. Inspired by [18], we employ an unsupervised GAN to transform TIR images to
|
[
[
0.47802734375,
0.11553955078125,
0.73388671875,
0.11553955078125,
0.73388671875,
0.1287841796875,
0.47802734375,
0.1287841796875
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.77392578125,
0.1287841796875
],
[
0.2418212890625,
0.147705078125,
0.78369140625,
0.147705078125,
0.78369140625,
0.1622314453125,
0.2418212890625,
0.1622314453125
],
[
0.21728515625,
0.1622314453125,
0.78125,
0.1622314453125,
0.78125,
0.1767578125,
0.21728515625,
0.1767578125
],
[
0.21728515625,
0.177978515625,
0.7861328125,
0.1767578125,
0.7861328125,
0.1912841796875,
0.21728515625,
0.1925048828125
],
[
0.2181396484375,
0.1932373046875,
0.78369140625,
0.1932373046875,
0.78369140625,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.78759765625,
0.208984375,
0.78759765625,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.21728515625,
0.222900390625,
0.7861328125,
0.2235107421875,
0.7861328125,
0.238037109375,
0.21728515625,
0.2374267578125
],
[
0.2181396484375,
0.2392578125,
0.7861328125,
0.2392578125,
0.7861328125,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.267578125,
0.2181396484375,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.7861328125,
0.26953125,
0.7861328125,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.21728515625,
0.28271484375,
0.78369140625,
0.2841796875,
0.78369140625,
0.298583984375,
0.21728515625,
0.29736328125
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.21728515625,
0.314453125,
0.285888671875,
0.314453125,
0.285888671875,
0.328857421875,
0.21728515625,
0.328857421875
],
[
0.2197265625,
0.35595703125,
0.233642578125,
0.35595703125,
0.233642578125,
0.367431640625,
0.2197265625,
0.367431640625
],
[
0.250732421875,
0.354248046875,
0.386474609375,
0.354248046875,
0.386474609375,
0.36865234375,
0.250732421875,
0.36865234375
],
[
0.2181396484375,
0.386962890625,
0.78369140625,
0.386962890625,
0.78369140625,
0.401611328125,
0.2181396484375,
0.401611328125
],
[
0.2181396484375,
0.40283203125,
0.7861328125,
0.40283203125,
0.7861328125,
0.417236328125,
0.2181396484375,
0.417236328125
],
[
0.2181396484375,
0.41845703125,
0.7861328125,
0.41845703125,
0.7861328125,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.2181396484375,
0.434326171875,
0.7861328125,
0.434326171875,
0.7861328125,
0.447509765625,
0.2181396484375,
0.447509765625
],
[
0.21728515625,
0.448974609375,
0.78759765625,
0.448974609375,
0.78759765625,
0.462158203125,
0.21728515625,
0.462158203125
],
[
0.21728515625,
0.46337890625,
0.78759765625,
0.46337890625,
0.78759765625,
0.477783203125,
0.21728515625,
0.477783203125
],
[
0.2181396484375,
0.479248046875,
0.7861328125,
0.479248046875,
0.7861328125,
0.492431640625,
0.2181396484375,
0.492431640625
],
[
0.21728515625,
0.494384765625,
0.7861328125,
0.494384765625,
0.7861328125,
0.50830078125,
0.21728515625,
0.50830078125
],
[
0.21728515625,
0.50830078125,
0.78759765625,
0.5087890625,
0.78759765625,
0.5234375,
0.21728515625,
0.52294921875
],
[
0.2181396484375,
0.5244140625,
0.78759765625,
0.5244140625,
0.78759765625,
0.53857421875,
0.2181396484375,
0.53857421875
],
[
0.21728515625,
0.5390625,
0.7861328125,
0.5390625,
0.7861328125,
0.55322265625,
0.21728515625,
0.55322265625
],
[
0.2181396484375,
0.55615234375,
0.2900390625,
0.55615234375,
0.2900390625,
0.5673828125,
0.2181396484375,
0.5673828125
],
[
0.24267578125,
0.5693359375,
0.78759765625,
0.5693359375,
0.78759765625,
0.58349609375,
0.24267578125,
0.58349609375
],
[
0.21728515625,
0.58544921875,
0.7861328125,
0.58544921875,
0.7861328125,
0.5986328125,
0.21728515625,
0.5986328125
],
[
0.2156982421875,
0.599609375,
0.78515625,
0.599609375,
0.78515625,
0.61279296875,
0.2156982421875,
0.61279296875
],
[
0.2181396484375,
0.6142578125,
0.78515625,
0.6142578125,
0.78515625,
0.62890625,
0.2181396484375,
0.62890625
],
[
0.21728515625,
0.6298828125,
0.7861328125,
0.6298828125,
0.7861328125,
0.64306640625,
0.21728515625,
0.64306640625
],
[
0.2181396484375,
0.64453125,
0.78515625,
0.64453125,
0.78515625,
0.65771484375,
0.2181396484375,
0.65771484375
],
[
0.21728515625,
0.66015625,
0.7861328125,
0.66015625,
0.7861328125,
0.6748046875,
0.21728515625,
0.6748046875
],
[
0.2181396484375,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.2181396484375,
0.68994140625,
0.7861328125,
0.68994140625,
0.7861328125,
0.7041015625,
0.2181396484375,
0.7041015625
],
[
0.21728515625,
0.7041015625,
0.78759765625,
0.705078125,
0.78759765625,
0.7197265625,
0.21728515625,
0.71826171875
],
[
0.2181396484375,
0.72021484375,
0.734375,
0.72021484375,
0.734375,
0.734375,
0.2181396484375,
0.734375
],
[
0.2442626953125,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2442626953125,
0.75
],
[
0.2181396484375,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.21728515625,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.21728515625,
0.77978515625
],
[
0.21728515625,
0.77978515625,
0.7861328125,
0.78076171875,
0.7861328125,
0.79541015625,
0.21728515625,
0.79443359375
],
[
0.21728515625,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Thermal to Visible Spectrum Transfer",
"3",
"In this paper, we propose an unsupervised method for transforming TIR im-",
"ages, specifically long-wavelength infrared (LWIR), to visible spectrum (VIS) im",
"ages. Our method is trained on unpaired images from the KAIST-MS dataset [4]",
"which allows it to handle the imperfect registration between the TIR-VIS im-",
"age pairs. Qualitative analysis shows that our proposed unsupervised method",
"produces sharp and perceptually realistic Vis images compared to the existing",
"state-of-the-art supervised methods. In addition, our proposed method achieves",
"comparable results to state-of-the-art supervised method in terms of L1 error de-",
"spite being trained on unpaired images. Finally, our proposed method generalizes",
"very well when evaluated on our new FOI dataset, which demonstrates the gen-",
"eralization capabilities of our method contrarily to the existing state-of-the-art",
"methods.",
"2",
"Related Work",
"Colorizing grayscale images has been extensively investigated in the literature.",
"Scribbles [9] requires the user to manually apply strokes of color to different",
"regions of a grayscale image and neighboring pixels with the same luminance",
"should get the same color. Transfer techniques [16] use the color palette from a",
"reference image and apply it to a grayscale image by matching luminance and",
"texture. Both scribbles and transfer techniques require manual input from the",
"user. Recently, colorization methods based on automatic transformation, i.e., the",
"only input to the method is the grayscale image, have become popular. Promising",
"results have been demonstrated in the area of automatic transformation using",
"Convolutional Neural Networks (CNNs) [3, 5, 8, 17] and Generative Adversarial",
"Networks (GANs) [2,6, 13] due to their abilities to model semantic representation",
"in images.",
"In the infrared spectrum, less research has been done on transforming thermal",
"images to VIS images. In [10], a CNN-based method was proposed to transform",
" near-infrared (NIR) images to VIS images. Their method was shown to perform",
"well as the NIR and VIS images are highly correlated in the electromagnetic",
"spectrum. Kniaz et al. [7] proposed VIS to TIR transformation using a CNN",
"model as a way to generate synthetic TIR images. The KAIST-MS [11] dataset",
"introduced the first realistic large-scale dataset of TIR-VIS image pairs which",
"opened up for developing TIR-VIS transformation models. Berg et al. [1] pro-",
"posed a CNN-based model to transform TIR images to VIS images trained on the",
"KAIST-MS dataset. However, the imperfect registration of the dataset caused",
"the output from their method to be blurry and corrupted in some cases.",
"Generative Adversarial Networks (GANs) have shown promising results in",
"unsupervised domain transfer [6, 12, 18, 19]. An unsupervised method does not",
"require a paired dataset, hence, eliminating the need for pixel to pixel corre-",
"spondence. In [14, 15], GANs have demonstrated a very good performance on",
"transferring NIR images to VIS images. Isola et al. [6] has shown some quali-",
"tative results from the KAIST-MS dataset as and example for domain transfer.",
"Inspired by [18], we employ an unsupervised GAN to transform TIR images to"
] |
{
"arxiv_id": "1904.02242",
"language": "en",
"page_idx": 2
}
|
||
1904.02242-en-3
|
1904.02242-en-3.jpg
|
4
A. Nyberg et al.
<FigureHere>
Fig.2: The unpaired model is mapping between two different domains, X (TIR) and Y (VIS), using $G:X\to Y$ and $F:Y\to X.$ This model produces three different losses; two adversarial losses based on $D_Y$ and $D_X$, and one cycle consistency loss.
VIS images which eliminates the deficiencies caused by the imperfect registration in the KAIST-MS dataset as the training set does not need to be paired. Different from [1], our proposed method produces a very realistic and sharp VIS images. In addition, our proposed method is able to generalize very well on unseen data from different environments.
# 3 Method
Here we describe our proposed approach for transforming TIR images to VIS
image while handling data miss-alignment in the KAIST-MS dataset.
### 3.1 Unpaired TIR-VIS Transfer
Inspired by [18],we perform unsupervised domain transfer between TIR and $\begin{array}{l}{\mathrm{VIS~images.~Given~TIR~domain~}\hat{X}~with~images~\{x_{i}:~x_{i}\in~X\}_{i=1}^{N}~and~VIS}\\{\mathrm{domain~}Y~with~images~\{y_{j}:y_{j}\in Y\}_{j=1}^{M},~we~aim~to~learn~two~transformations}\\{G~\mathrm{and~}F~between~the~two~domains~as~shown~in~Figure~2.~TIR~input~images}\end{array}$ are transformed from the thermal domain $X$ to the visible spectrum domain $Y$ using the generator $G$, while the generator $F$ performs in the opposite direction. An adversarial discriminator $D_X$ aims to discriminate between images $x$ and the transformed images $F(y)$,while another discriminator $D_Y$ discriminates between $\operatorname{images}y$ and $G(x).$
# 3.2 Adversarial Training
The main objective of a GAN with a cyclic loss is to learn the two transformations $G:X\to Y,F:Y\to X$ and their corresponding discriminators $D_X,D_Y[18].$
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.271240234375,
0.11553955078125,
0.381591796875,
0.11553955078125,
0.381591796875,
0.1300048828125,
0.271240234375,
0.1300048828125
],
[
0.336669921875,
0.154052734375,
0.410888671875,
0.154052734375,
0.410888671875,
0.1717529296875,
0.336669921875,
0.1717529296875
],
[
0.57177734375,
0.1572265625,
0.640625,
0.1572265625,
0.640625,
0.1685791015625,
0.57177734375,
0.1685791015625
],
[
0.71728515625,
0.1558837890625,
0.744140625,
0.1558837890625,
0.744140625,
0.175537109375,
0.71728515625,
0.175537109375
],
[
0.3544921875,
0.1685791015625,
0.414306640625,
0.17041015625,
0.41259765625,
0.186279296875,
0.35302734375,
0.184326171875
],
[
0.5849609375,
0.169189453125,
0.654296875,
0.169189453125,
0.654296875,
0.184326171875,
0.5849609375,
0.184326171875
],
[
0.72119140625,
0.173583984375,
0.73876953125,
0.173583984375,
0.73876953125,
0.1912841796875,
0.72119140625,
0.1912841796875
],
[
0.4599609375,
0.1900634765625,
0.5458984375,
0.1900634765625,
0.5458984375,
0.2020263671875,
0.4599609375,
0.2020263671875
],
[
0.60693359375,
0.184326171875,
0.63232421875,
0.184326171875,
0.63232421875,
0.1976318359375,
0.60693359375,
0.1976318359375
],
[
0.2320556640625,
0.1925048828125,
0.3203125,
0.1925048828125,
0.3203125,
0.203857421875,
0.2320556640625,
0.203857421875
],
[
0.68115234375,
0.1925048828125,
0.77783203125,
0.1944580078125,
0.77783203125,
0.208984375,
0.68115234375,
0.20703125
],
[
0.26953125,
0.205810546875,
0.283447265625,
0.205810546875,
0.283447265625,
0.216552734375,
0.26953125,
0.216552734375
],
[
0.473876953125,
0.203857421875,
0.53173828125,
0.203857421875,
0.53173828125,
0.2183837890625,
0.473876953125,
0.2183837890625
],
[
0.72119140625,
0.212158203125,
0.73779296875,
0.212158203125,
0.73779296875,
0.2279052734375,
0.72119140625,
0.2279052734375
],
[
0.38232421875,
0.2298583984375,
0.431396484375,
0.2298583984375,
0.431396484375,
0.24560546875,
0.38232421875,
0.24560546875
],
[
0.6806640625,
0.2298583984375,
0.77294921875,
0.231689453125,
0.77294921875,
0.24560546875,
0.6806640625,
0.24365234375
],
[
0.226318359375,
0.24365234375,
0.32275390625,
0.24560546875,
0.32275390625,
0.260009765625,
0.226318359375,
0.25830078125
],
[
0.39453125,
0.24560546875,
0.4248046875,
0.24560546875,
0.4248046875,
0.25830078125,
0.39453125,
0.25830078125
],
[
0.7236328125,
0.2442626953125,
0.73779296875,
0.2442626953125,
0.73779296875,
0.255126953125,
0.7236328125,
0.255126953125
],
[
0.26708984375,
0.26318359375,
0.283447265625,
0.26318359375,
0.283447265625,
0.279052734375,
0.26708984375,
0.279052734375
],
[
0.2646484375,
0.276611328125,
0.294189453125,
0.2822265625,
0.289306640625,
0.300537109375,
0.259033203125,
0.29541015625
],
[
0.4599609375,
0.270751953125,
0.5400390625,
0.270751953125,
0.5400390625,
0.2822265625,
0.4599609375,
0.2822265625
],
[
0.47216796875,
0.28271484375,
0.53125,
0.28271484375,
0.53125,
0.29736328125,
0.47216796875,
0.29736328125
],
[
0.3505859375,
0.3037109375,
0.419189453125,
0.302978515625,
0.419189453125,
0.318115234375,
0.3505859375,
0.3193359375
],
[
0.21728515625,
0.34228515625,
0.78369140625,
0.343505859375,
0.78369140625,
0.35791015625,
0.21728515625,
0.356689453125
],
[
0.2181396484375,
0.359130859375,
0.78515625,
0.359130859375,
0.78515625,
0.37255859375,
0.2181396484375,
0.37255859375
],
[
0.21728515625,
0.373779296875,
0.78369140625,
0.373779296875,
0.78369140625,
0.386962890625,
0.21728515625,
0.386962890625
],
[
0.2181396484375,
0.391357421875,
0.334228515625,
0.391357421875,
0.334228515625,
0.401611328125,
0.2181396484375,
0.401611328125
],
[
0.2197265625,
0.43310546875,
0.78369140625,
0.43310546875,
0.78369140625,
0.4462890625,
0.2197265625,
0.4462890625
],
[
0.2181396484375,
0.448974609375,
0.78369140625,
0.448974609375,
0.78369140625,
0.462158203125,
0.2181396484375,
0.462158203125
],
[
0.2181396484375,
0.46337890625,
0.78515625,
0.46337890625,
0.78515625,
0.477783203125,
0.2181396484375,
0.477783203125
],
[
0.2197265625,
0.479248046875,
0.7861328125,
0.479248046875,
0.7861328125,
0.492431640625,
0.2197265625,
0.492431640625
],
[
0.2181396484375,
0.49365234375,
0.51318359375,
0.49365234375,
0.51318359375,
0.5068359375,
0.2181396484375,
0.5068359375
],
[
0.21728515625,
0.52880859375,
0.253173828125,
0.52880859375,
0.253173828125,
0.54443359375,
0.21728515625,
0.54443359375
],
[
0.2459716796875,
0.52880859375,
0.330078125,
0.52880859375,
0.330078125,
0.54345703125,
0.2459716796875,
0.54345703125
],
[
0.2181396484375,
0.5595703125,
0.7861328125,
0.5595703125,
0.7861328125,
0.57373046875,
0.2181396484375,
0.57373046875
],
[
0.2181396484375,
0.57373046875,
0.7138671875,
0.57373046875,
0.7138671875,
0.58837890625,
0.2181396484375,
0.58837890625
],
[
0.21728515625,
0.609375,
0.49267578125,
0.609375,
0.49267578125,
0.62255859375,
0.21728515625,
0.62255859375
],
[
0.2181396484375,
0.6318359375,
0.7861328125,
0.6318359375,
0.7861328125,
0.646484375,
0.2181396484375,
0.646484375
],
[
0.21728515625,
0.64453125,
0.78759765625,
0.64599609375,
0.78759765625,
0.66357421875,
0.21728515625,
0.662109375
],
[
0.21728515625,
0.6611328125,
0.78759765625,
0.6611328125,
0.78759765625,
0.6787109375,
0.21728515625,
0.6787109375
],
[
0.2181396484375,
0.67822265625,
0.78515625,
0.67822265625,
0.78515625,
0.69140625,
0.2181396484375,
0.69140625
],
[
0.2181396484375,
0.6923828125,
0.78369140625,
0.6923828125,
0.78369140625,
0.70703125,
0.2181396484375,
0.70703125
],
[
0.2197265625,
0.70849609375,
0.78369140625,
0.70849609375,
0.78369140625,
0.7216796875,
0.2197265625,
0.7216796875
],
[
0.2197265625,
0.7236328125,
0.78515625,
0.7236328125,
0.78515625,
0.7373046875,
0.2197265625,
0.7373046875
],
[
0.2181396484375,
0.73779296875,
0.7861328125,
0.73779296875,
0.7861328125,
0.75244140625,
0.2181396484375,
0.75244140625
],
[
0.21728515625,
0.75390625,
0.359375,
0.75390625,
0.359375,
0.767578125,
0.21728515625,
0.767578125
],
[
0.2156982421875,
0.78515625,
0.4345703125,
0.787109375,
0.4345703125,
0.8046875,
0.2156982421875,
0.80322265625
],
[
0.21728515625,
0.8095703125,
0.78369140625,
0.81103515625,
0.78369140625,
0.82568359375,
0.21728515625,
0.82373046875
],
[
0.2181396484375,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"A. Nyberg et al.",
"Domain X",
"DomainY",
"Dy",
"TIR input",
"Generated",
"↑",
"GeneratorG",
"VIS",
"Discriminator",
"Decision [0,1]",
"X",
"TIR2VIS",
"↑",
"; Cyclic",
"Discriminator",
"Decision [0,1]",
"loss",
"Y",
"+",
"Dx",
"GeneratorF",
"VIS2TIR",
"Cyclic TIR",
"Fig.2: The unpaired model is mapping between two different domains, X (TIR)",
"and Y (VIS), using G : X -→ Y and F : Y → X. This model produces three",
"different losses; two adversarial losses based on Dy and Dx, and one cycle",
"consistencyloss",
"VIS images which eliminates the deficiencies caused by the imperfect registra-",
"tion in the KAIST-MS dataset as the training set does not need to be paired.",
"Different from [1], our proposed method produces a very realistic and sharp",
"VIS images. In addition, our proposed method is able to generalize very well on",
"unseen data from different environments.",
"3",
" Method",
"Here we describe our proposed approach for transforming TIR images to VIS",
"image while handling data miss-alignment in the KAIST-MS dataset.",
"3.1Unpaired TIR-VIS Transfer",
"Inspired by [18], we perform unsupervised domain transfer between TIR and",
"VIS images. Given TIR domain X with images (ai : ai E X) and VIS",
"domain Y with images (yj : yi e Y)p-1, we aim to learn two transformations",
"G and F between the two domains as shown in Figure 2. TIR input images",
"are transformed from the thermal domain X to the visible spectrum domain Y",
"using the generator G, while the generator F performs in the opposite direction.",
"An adversarial discriminator Dx aims to discriminate between images and the",
"transformed images F(y), while another discriminator Dy discriminates between",
"images y and G(r).",
"3.2 Adversarial Training",
"The main objective of a GAN with a cyclic loss is to learn the two transformations",
"G : X -→ Y, F : Y -→ X and their corresponding discriminators Dx,Dy [18]."
] |
{
"arxiv_id": "1904.02242",
"language": "en",
"page_idx": 3
}
|
||
1904.02242-en-4
|
1904.02242-en-4.jpg
|
5
Thermal to Visible Spectrum Transfer
During training, the generator $G_{G}$ transforms an image $x\in X$ into a synthetic image $\hat{y} .$The synthetic image is then evaluated by the discriminator $D_Y.$ The adversarial loss for the $G_G$ is defined as:
$$\mathcal{L}_{G_G}(G_G,D_Y,x)=\mathbb{E}_{x\sim p_{data(x)}}[(D_Y(G_G(x))-1)^2]$$
(1)
where $p_{data(x)}$ is data distribution in $X.$ The loss value becomes large if the synthetic image $\hat{y}$ was able to fool the discriminator into outputting a value close or equal to one. On the other hand, the discriminator tries to maximize the probability for real images while also minimizing the output on synthetic images, achieved by minimizing the following formula:
$$\mathcal{L}_{D_Y}(D_Y,x,y)=\mathbb{E}_{x\sim p_{data(x)}}[(D_Y(G_G(x)))^2]+\mathbb{E}_{y\sim p_{data(y)}}[(D_Y(y)-1)^2].$$
The total adversarial loss for the $G$ transformation is then defined as:
$$\mathcal{L}_{GAN}(G_{G},D_{Y},x,y)=\mathcal{L}_{D_{Y}}(D_{Y},y)+\operatorname*{min}_{G_{G}}\operatorname*{max}_{D_{Y}}\mathcal{L}_{G_{G}}(G_{G},D_{Y},x)$$
(3)
A similar loss is utilized to learn the transformation $F.$ To reduce the space of possible transformations, a cycle-consistency loss [18] is employed which ensures that the learned transformation can map only a single input to the desired output. The cycle-consistency loss is defined as:
$$\begin{aligned}\mathcal{L}_{cyc}(G_{G},G_{F})&=\mathbb{E}_{x\sim p_{data}(x)}[||G_{F}(G_{G}(x))-x||_{1}]+\\&\mathbb{E}_{y\sim p_{data}(y)}[||G_{G}(G_{F}(y))-y||_{1}]\end{aligned}$$
(4)
Combining the above losses gives our total loss which is defined as:
$$\begin{aligned}\mathcal{L}(G_{G},G_{F},D_{X},D_{Y})=\mathcal{L}_{GAN}(G_{G},D_{Y},x,y)+\\\mathcal{L}_{GAN}(G_{F},D_{X},y,x)+\\\lambda\mathcal{L}_{cyc}(G_{G},G_{F})\end{aligned}$$
(5)
where $\lambda$ is a factor used to control the impact of the cyclic loss
# 4 Experiments
For evaluation, we compare our proposed method with the existing state-of-theart method on TIR to VIS transfer TIR2Lab [1]. The evaluation is performed on the KAIST-MS, the FOI dataset and the generalization capabilities of the evaluated methods are tested by training on the former and evaluating on the $\operatorname{latter}.$
## 4.1 Datasets
$The~KIAST-MS~dataset~[4]~$contains paired TIR and VIS images which were collected by mounting a TIR and a VIS cameras on a vehicle in city and suburbar
|
[
[
0.48046875,
0.117431640625,
0.73388671875,
0.117431640625,
0.73388671875,
0.1275634765625,
0.48046875,
0.1275634765625
],
[
0.77392578125,
0.11871337890625,
0.7861328125,
0.11871337890625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.2181396484375,
0.1483154296875,
0.78515625,
0.1483154296875,
0.78515625,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.2181396484375,
0.1787109375,
0.50732421875,
0.1787109375,
0.50732421875,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.32275390625,
0.203857421875,
0.68115234375,
0.203857421875,
0.68115234375,
0.2247314453125,
0.32275390625,
0.2247314453125
],
[
0.76123046875,
0.2052001953125,
0.78759765625,
0.2052001953125,
0.78759765625,
0.222900390625,
0.76123046875,
0.222900390625
],
[
0.21728515625,
0.231689453125,
0.7861328125,
0.2298583984375,
0.7861328125,
0.2474365234375,
0.21728515625,
0.25
],
[
0.2197265625,
0.248779296875,
0.7861328125,
0.248779296875,
0.7861328125,
0.261962890625,
0.2197265625,
0.261962890625
],
[
0.2181396484375,
0.26318359375,
0.7861328125,
0.26318359375,
0.7861328125,
0.276611328125,
0.2181396484375,
0.276611328125
],
[
0.2181396484375,
0.27783203125,
0.7861328125,
0.279052734375,
0.7861328125,
0.29345703125,
0.2181396484375,
0.292236328125
],
[
0.21728515625,
0.294189453125,
0.60693359375,
0.294189453125,
0.60693359375,
0.30810546875,
0.21728515625,
0.30810546875
],
[
0.2344970703125,
0.32080078125,
0.78515625,
0.32080078125,
0.78515625,
0.33837890625,
0.2344970703125,
0.33837890625
],
[
0.2197265625,
0.348388671875,
0.71728515625,
0.348388671875,
0.71728515625,
0.3623046875,
0.2197265625,
0.3623046875
],
[
0.27783203125,
0.376953125,
0.72802734375,
0.376953125,
0.72802734375,
0.391357421875,
0.27783203125,
0.391357421875
],
[
0.76318359375,
0.376953125,
0.7861328125,
0.376953125,
0.7861328125,
0.39013671875,
0.76318359375,
0.39013671875
],
[
0.55615234375,
0.39013671875,
0.59716796875,
0.39013671875,
0.59716796875,
0.397705078125,
0.55615234375,
0.397705078125
],
[
0.2181396484375,
0.4091796875,
0.7861328125,
0.4091796875,
0.7861328125,
0.42236328125,
0.2181396484375,
0.42236328125
],
[
0.2181396484375,
0.4248046875,
0.78369140625,
0.4248046875,
0.78369140625,
0.438232421875,
0.2181396484375,
0.438232421875
],
[
0.2181396484375,
0.439453125,
0.7861328125,
0.439453125,
0.7861328125,
0.453857421875,
0.2181396484375,
0.453857421875
],
[
0.21728515625,
0.455078125,
0.5595703125,
0.455078125,
0.5595703125,
0.468505859375,
0.21728515625,
0.468505859375
],
[
0.324462890625,
0.479248046875,
0.68310546875,
0.479248046875,
0.68310546875,
0.5,
0.324462890625,
0.5
],
[
0.76123046875,
0.4892578125,
0.78759765625,
0.4892578125,
0.78759765625,
0.5068359375,
0.76123046875,
0.5068359375
],
[
0.459228515625,
0.5,
0.68310546875,
0.5,
0.68310546875,
0.517578125,
0.459228515625,
0.517578125
],
[
0.2197265625,
0.5263671875,
0.69775390625,
0.5263671875,
0.69775390625,
0.54052734375,
0.2197265625,
0.54052734375
],
[
0.343017578125,
0.5537109375,
0.6640625,
0.5537109375,
0.6640625,
0.568359375,
0.343017578125,
0.568359375
],
[
0.50732421875,
0.57275390625,
0.6650390625,
0.57275390625,
0.6650390625,
0.5869140625,
0.50732421875,
0.5869140625
],
[
0.76123046875,
0.5712890625,
0.78759765625,
0.5712890625,
0.78759765625,
0.5888671875,
0.76123046875,
0.5888671875
],
[
0.55615234375,
0.59130859375,
0.6650390625,
0.59130859375,
0.6650390625,
0.60595703125,
0.55615234375,
0.60595703125
],
[
0.2197265625,
0.61865234375,
0.67236328125,
0.61865234375,
0.67236328125,
0.6318359375,
0.2197265625,
0.6318359375
],
[
0.2181396484375,
0.6572265625,
0.374267578125,
0.6572265625,
0.374267578125,
0.671875,
0.2181396484375,
0.671875
],
[
0.2181396484375,
0.68798828125,
0.78173828125,
0.68798828125,
0.78173828125,
0.70263671875,
0.2181396484375,
0.70263671875
],
[
0.2181396484375,
0.7041015625,
0.7861328125,
0.7041015625,
0.7861328125,
0.71728515625,
0.2181396484375,
0.71728515625
],
[
0.21728515625,
0.71728515625,
0.7861328125,
0.71826171875,
0.7861328125,
0.73291015625,
0.21728515625,
0.73193359375
],
[
0.2181396484375,
0.734375,
0.7861328125,
0.734375,
0.7861328125,
0.74755859375,
0.2181396484375,
0.74755859375
],
[
0.21728515625,
0.7470703125,
0.2646484375,
0.74853515625,
0.26318359375,
0.76318359375,
0.2156982421875,
0.76123046875
],
[
0.2181396484375,
0.787109375,
0.26318359375,
0.787109375,
0.26318359375,
0.798828125,
0.2181396484375,
0.798828125
],
[
0.25732421875,
0.787109375,
0.33251953125,
0.787109375,
0.33251953125,
0.798828125,
0.25732421875,
0.798828125
],
[
0.221435546875,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.221435546875,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"Thermal toVisibleSpectrumTransfer",
"5",
"During training, the generator GG transforms an image r E X into a synthetic",
"image g. The synthetic image is then evaluated by the discriminator Dy. The",
"adversarial loss for the GG is defined as:",
"LGe(Gc, Dy, t) = Er~pata(m)[(Dy(GG(a) - 1)\"]",
"(1)",
"where Pdata(r) is data distribution in X. The loss value becomes large if the",
"synthetic image g was able to fool the discriminator into outputting a value",
"close or equal to one. On the other hand, the discriminator tries to maximize",
"the probability for real images while also minimizing the output on synthetic",
"images, achieved by minimizing the following formula:",
"CDy(Dy,&, y) = Er~pata()[(Dy(GG(r)\"] + Ey~pata(u) [(Dy(y) -1)\"]. (2)",
"The total adversarial loss for the G transformation is then defined as:",
"CGAN(GG, DY,r,y) = CDy(Dy,y) +minmax LG(GG, Dy,r)",
"(3)",
"GGD",
"A similar loss is utilized to learn the transformation F. To reduce the space of",
"possible transformations, a cycle-consistency loss [18] is employed which ensures",
"that the learned transformation can map only a single input to the desired",
"output.The cycle-consistency loss is defined as:",
"Ceye(GG, Gp) = Er~pdata(r)[Gr(Gc(a)) - rll]+",
"(4)",
"Ey~pdata(u)[IGc(GF(y) - ylli])",
"Combining the above losses gives our total loss which is defined as:",
"C(GG,GF, Dx,Dy) = LGAN(GG,Dy,c,y)+",
"LGAN(GF, Dx,y, T)+",
"(5)",
"入Leyc(GG,GF)",
"where is a factor used to control the impact of the cyclic loss",
"4Experiments",
"For evaluation, we compare our proposed method with the existing state-of-the",
"art method on TIR to VIS transfer TIR2Lab [1]: The evaluation is performed",
"on the KAIST-MS, the FOI dataset and the generalization capabilities of the",
"evaluated methods are tested by training on the former and evaluating on the",
"latter.",
"4.1",
"Datasets",
"The KIAST-MS dataset [4] contains paired TIR and VIS images which were",
"collected by mounting a TIR and a VIS cameras on a vehicle in city and suburban"
] |
{
"arxiv_id": "1904.02242",
"language": "en",
"page_idx": 4
}
|
||
2309.12423-en-1
|
2309.12423-en-1.jpg
|
WWW'23, May 1-5, 2023, Austin, TX, USA
However, a major limitation of most LP methods is that relations can only be predicted among entities that already exist in the KG. A more practical and challenging perspective for this task is to consider LP for entirely new entities (i.e., inductive link prediction). This perspective also better reflects the needs of an event prediction system, which would aim to make predictions about entirely new events. The inductive setting introduces an additional level of difficulty for LP models, as many of them (especially embedding-based ones) operate under the closed world assumption- that is, they can only perform predictions for entities that have been seen during training, and facts that are not seen during training are assumed to be false. Additionally, even for models that are capable of this task, some KGs (such as Wikidata) constantly undergo updates and changes, which can negatively affect a model's performance as its training data becomes outdated.
Towards our goal of applying KGs to perform event prediction, we develop a case-based reasoning approach, EvCBR, which leverages examples of past events in the KG to perform LP for the properties of unseen event entities. We frame our problem as a 2-hop link prediction task- e.g., starting from a cause event, we assume the existence of a causal relation to a new effect event and predict its properties. This approach allows us to perform inductive link prediction for the properties of the unseen effect without the need for external data. Importantly, EvCBR requires no training, relying instead on statistical measures and subclass hierarchies to identify similar entities and follow paths through the KG to perform predictions. This makes EvCBR well-suited to KGs such as Wikidata that frequently experience changes in their structure and content. More generally, we can also apply EvCBR to perform 2-hop LP for properties of any type of unseen entity that is connected to a known entity by some relation.
Our contributions are as follows:
(1) We introduce a case-based reasoning model, EvCBR,$^2$for
event prediction. Our model considers this task as a LP task in a 2-hop setting, leveraging knowledge about similar causeeffect events to make predictions about the unseen effect. EvCBR requires no training, making it well-suited to handle new events and changes to the underlying KG.
(2) Compared to similar work, we introduce novel similarity
metrics to identify similar cause-effect event cases as well as a refinement step to improve the precision of predictions.
(3) We curate and release a novel dataset surrounding causal
events in Wikidata, extracting news events that are connected by causal relations as well as their local connections.$^3$
(4) In our 2-hop inductive link prediction task, our model shows
superior performance on our event dataset as well as competitive performance on a modified evaluation dataset based on the FB15k-237 dataset.
# 2 RELATED WORK
Zhao [32] presents a comprehensive survey of different kinds of event prediction methods across different domains. Under their taxonomy of event prediction methods [32,Fig. 3],our method is most closely related to the “causality-based" methods under“semantic
$^{2}$We publicly release our code at https://github.com/solashirai/WWW-EvCBR
$^{3}$Dataset available at https://doi.org/10.5281/zenodo.7196049.
Sola Shirai, Debarun Bhattacharjya, and Oktie Hassanzadeh
prediction" techniques. Other notable work that falls under this class of methods includes Radinsky et al's Pundit algorithm [17] which is based on automated construction of a causality graph through extraction of causal patterns and generalization. Zhao et al. [33] also construct a causal graph from text documents and build embeddings to perform a simple (one-hop) prediction. While prior work has explored applying LP methods to KGs of causal events [10,22], to our knowledge, our work is the first to apply 2-hop LP over a knowledge graph for event prediction.
In the space of LP in KGs, a significant number of embeddingbased methods have been developed as detailed in recent surveys [29,30]. Such methods often show trade-offs in performance based on the dataset and how well the model hyperparameters are tuned [19]. While GNN-based models [31] have often shown recent stateof-the-art performance, there has been discussion surrounding what semantics are actually captured in such models [9] as well as reevaluations of how one evaluates such LP models [1,24,25]. Furthermore, most embedding-based models are trained and evaluated in the closed world setting, with a relatively sparse number of recent works such as [2,6] considering inductive link prediction using additional knowledge sources such as text or hyper-relational facts.
While less prevalent than embedding-based models, there are also a variety of rule-based LP models [7,11,13,15,20]. Methods such as AnyBURL [12] have shown similar performance to state-ofthe-art embedding models while requiring significantly less training time [18]. Works such as [4,5] present case-based reasoning models which, like our method, require no training and make path-based predictions. Our method differs from such similar works in our method of computing entity similarity, approach to apply and refine prediction paths, and task formulation of performing 2-hop LP.
3 РROВLEМ FORMULATION
Our work considers the application of LP methods over a knowledge graph for the task of event prediction. We define a knowledge graph $G=(E,R,T)$ as consisting of a set of entities $E$, a set of relations $R$, and a set of triples $T.T$ consists of triples of the form $(h,r,t)$, where $h,t\in E$ and $r\in R.$
The task of LP is then to correctly predict the missing tail entity given a partial triple $(h,r,?t).^4$ The standard assumption in this task is that the head entity $h$ and the missing tail entity ?$t$ are known entities in $E.$ To perform inductive link prediction, it is necessary to predict the missing ?$t$ where ?$t\notin E.$ Since LP models can not predict such an entity ? $t$ solely from the content of G, inductive link prediction methods often also rely on additional background data (such as textual descriptions) to extract and represent new entities.
## 3.1 Event Prediction Task
Extending the task of LP to event prediction, our goal is to make predictions about new events based on a causal relation between two events. We represent our task as a query triple $(c,r,e)$,where $c,r$,and $e$ denote the cause event, causal relation, and effect event, respectively. Furthermore, the point of the event prediction task is to predict properties about the new effect event- i.e., predicting outgoing properties about $e$ in the form of $(e,r_e,?z)\in Out_e$,where
$^{4}$We note the related LP tasks of predicting a missing head ?$h$ or relation ?$r$,but in this paper we focus on LP for predicting ? $t$ as it aligns best with our event prediction task.
|
[
[
0.0849609375,
0.0738525390625,
0.289306640625,
0.07452392578125,
0.289306640625,
0.08837890625,
0.0849609375,
0.08709716796875
],
[
0.62744140625,
0.07574462890625,
0.91259765625,
0.07574462890625,
0.91259765625,
0.0889892578125,
0.62744140625,
0.0889892578125
],
[
0.101318359375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.101318359375,
0.12060546875
],
[
0.515625,
0.106689453125,
0.91259765625,
0.1060791015625,
0.91259765625,
0.12060546875,
0.515625,
0.121826171875
],
[
0.08660888671875,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.08660888671875,
0.136962890625,
0.48046875,
0.136962890625,
0.48046875,
0.147705078125,
0.08660888671875,
0.147705078125
],
[
0.51708984375,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.5185546875,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.5185546875,
0.1634521484375
],
[
0.0849609375,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.0849609375,
0.1767578125
],
[
0.51708984375,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.08660888671875,
0.1787109375,
0.48046875,
0.1787109375,
0.48046875,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.5185546875,
0.1787109375,
0.91259765625,
0.1787109375,
0.91259765625,
0.189453125,
0.5185546875,
0.189453125
],
[
0.08660888671875,
0.1925048828125,
0.48046875,
0.1925048828125,
0.48046875,
0.20263671875,
0.08660888671875,
0.20263671875
],
[
0.51708984375,
0.1912841796875,
0.9140625,
0.1912841796875,
0.9140625,
0.2052001953125,
0.51708984375,
0.2052001953125
],
[
0.0849609375,
0.2052001953125,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2183837890625,
0.0849609375,
0.2183837890625
],
[
0.5185546875,
0.2052001953125,
0.91259765625,
0.2052001953125,
0.91259765625,
0.2197265625,
0.5185546875,
0.2197265625
],
[
0.0849609375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.5185546875,
0.2203369140625,
0.7890625,
0.2203369140625,
0.7890625,
0.2310791015625,
0.5185546875,
0.2310791015625
],
[
0.0849609375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.53173828125,
0.23291015625,
0.9140625,
0.23291015625,
0.9140625,
0.246826171875,
0.53173828125,
0.246826171875
],
[
0.08660888671875,
0.2474365234375,
0.4794921875,
0.2474365234375,
0.4794921875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.5185546875,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.5185546875,
0.2734375
],
[
0.08660888671875,
0.27587890625,
0.48046875,
0.27587890625,
0.48046875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.51953125,
0.27587890625,
0.91259765625,
0.27587890625,
0.91259765625,
0.285888671875,
0.51953125,
0.285888671875
],
[
0.08660888671875,
0.2890625,
0.48046875,
0.2890625,
0.48046875,
0.2998046875,
0.08660888671875,
0.2998046875
],
[
0.5185546875,
0.285888671875,
0.9140625,
0.285888671875,
0.9140625,
0.2998046875,
0.5185546875,
0.2998046875
],
[
0.0849609375,
0.3017578125,
0.281005859375,
0.3017578125,
0.281005859375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.51708984375,
0.3017578125,
0.9140625,
0.3017578125,
0.9140625,
0.31494140625,
0.51708984375,
0.31494140625
],
[
0.1004638671875,
0.31494140625,
0.483642578125,
0.314453125,
0.483642578125,
0.328857421875,
0.1004638671875,
0.330078125
],
[
0.5185546875,
0.31640625,
0.91259765625,
0.31640625,
0.91259765625,
0.326904296875,
0.5185546875,
0.326904296875
],
[
0.08660888671875,
0.330810546875,
0.48193359375,
0.330810546875,
0.48193359375,
0.341552734375,
0.08660888671875,
0.341552734375
],
[
0.5185546875,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.34033203125,
0.5185546875,
0.34033203125
],
[
0.0849609375,
0.343505859375,
0.483642578125,
0.343505859375,
0.483642578125,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.51708984375,
0.343505859375,
0.9140625,
0.343505859375,
0.9140625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.08660888671875,
0.359130859375,
0.48046875,
0.359130859375,
0.48046875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.5185546875,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.369384765625,
0.5185546875,
0.369384765625
],
[
0.08660888671875,
0.37255859375,
0.4794921875,
0.37255859375,
0.4794921875,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
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
],
[
0.5185546875,
0.3837890625,
0.9140625,
0.3837890625,
0.9140625,
0.397705078125,
0.5185546875,
0.397705078125
],
[
0.0849609375,
0.399658203125,
0.48193359375,
0.399658203125,
0.48193359375,
0.41357421875,
0.0849609375,
0.41357421875
],
[
0.53173828125,
0.397705078125,
0.9140625,
0.399658203125,
0.9140625,
0.41357421875,
0.53173828125,
0.410888671875
],
[
0.0882568359375,
0.415283203125,
0.47802734375,
0.415283203125,
0.47802734375,
0.42236328125,
0.0882568359375,
0.42236328125
],
[
0.51708984375,
0.412353515625,
0.91259765625,
0.412353515625,
0.91259765625,
0.425537109375,
0.51708984375,
0.425537109375
],
[
0.0849609375,
0.4267578125,
0.48193359375,
0.4267578125,
0.48193359375,
0.43994140625,
0.0849609375,
0.43994140625
],
[
0.51708984375,
0.425537109375,
0.9140625,
0.425537109375,
0.9140625,
0.439453125,
0.51708984375,
0.439453125
],
[
0.08660888671875,
0.44140625,
0.48046875,
0.44140625,
0.48046875,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.51708984375,
0.43994140625,
0.91259765625,
0.43994140625,
0.91259765625,
0.453857421875,
0.51708984375,
0.453857421875
],
[
0.0849609375,
0.453857421875,
0.48193359375,
0.453857421875,
0.48193359375,
0.46728515625,
0.0849609375,
0.46728515625
],
[
0.51708984375,
0.453857421875,
0.91259765625,
0.453857421875,
0.91259765625,
0.46728515625,
0.51708984375,
0.46728515625
],
[
0.0849609375,
0.46728515625,
0.48193359375,
0.46728515625,
0.48193359375,
0.48095703125,
0.0849609375,
0.48095703125
],
[
0.51708984375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.0849609375,
0.481689453125,
0.483642578125,
0.481689453125,
0.483642578125,
0.49560546875,
0.0849609375,
0.49560546875
],
[
0.515625,
0.481689453125,
0.9140625,
0.48095703125,
0.9140625,
0.494384765625,
0.515625,
0.49560546875
],
[
0.0849609375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.51708984375,
0.49560546875,
0.91259765625,
0.49560546875,
0.91259765625,
0.5087890625,
0.51708984375,
0.5087890625
],
[
0.0882568359375,
0.51123046875,
0.48046875,
0.51123046875,
0.48046875,
0.521484375,
0.0882568359375,
0.521484375
],
[
0.5185546875,
0.51123046875,
0.89453125,
0.51123046875,
0.89453125,
0.521484375,
0.5185546875,
0.521484375
],
[
0.08660888671875,
0.5244140625,
0.273681640625,
0.5244140625,
0.273681640625,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.101318359375,
0.537109375,
0.301513671875,
0.537109375,
0.301513671875,
0.55029296875,
0.101318359375,
0.55029296875
],
[
0.51953125,
0.53857421875,
0.5390625,
0.53857421875,
0.5390625,
0.546875,
0.51953125,
0.546875
],
[
0.53759765625,
0.5390625,
0.54345703125,
0.5390625,
0.54345703125,
0.54541015625,
0.53759765625,
0.54541015625
],
[
0.54345703125,
0.537109375,
0.76708984375,
0.537109375,
0.76708984375,
0.54736328125,
0.54345703125,
0.54736328125
],
[
0.10296630859375,
0.55322265625,
0.48046875,
0.5517578125,
0.48046875,
0.56640625,
0.10296630859375,
0.5673828125
],
[
0.5185546875,
0.5537109375,
0.91259765625,
0.5537109375,
0.91259765625,
0.5673828125,
0.5185546875,
0.5673828125
],
[
0.1258544921875,
0.56640625,
0.48046875,
0.56640625,
0.48046875,
0.580078125,
0.1258544921875,
0.580078125
],
[
0.5185546875,
0.5673828125,
0.91259765625,
0.5673828125,
0.91259765625,
0.58203125,
0.5185546875,
0.58203125
],
[
0.12744140625,
0.58349609375,
0.4794921875,
0.58349609375,
0.4794921875,
0.59326171875,
0.12744140625,
0.59326171875
],
[
0.5185546875,
0.583984375,
0.9111328125,
0.583984375,
0.9111328125,
0.5947265625,
0.5185546875,
0.5947265625
],
[
0.1282958984375,
0.5986328125,
0.4794921875,
0.5986328125,
0.4794921875,
0.60595703125,
0.1282958984375,
0.60595703125
],
[
0.51708984375,
0.5966796875,
0.9140625,
0.5966796875,
0.9140625,
0.60986328125,
0.51708984375,
0.60986328125
],
[
0.1258544921875,
0.60986328125,
0.48193359375,
0.60986328125,
0.48193359375,
0.62353515625,
0.1258544921875,
0.62353515625
],
[
0.5185546875,
0.611328125,
0.66650390625,
0.611328125,
0.66650390625,
0.6220703125,
0.5185546875,
0.6220703125
],
[
0.1258544921875,
0.62353515625,
0.406982421875,
0.62353515625,
0.406982421875,
0.63720703125,
0.1258544921875,
0.63720703125
],
[
0.53369140625,
0.62353515625,
0.9111328125,
0.62353515625,
0.9111328125,
0.63720703125,
0.53369140625,
0.63720703125
],
[
0.10540771484375,
0.63818359375,
0.4794921875,
0.63818359375,
0.4794921875,
0.6484375,
0.10540771484375,
0.6484375
],
[
0.5185546875,
0.63720703125,
0.91259765625,
0.63720703125,
0.91259765625,
0.65087890625,
0.5185546875,
0.65087890625
],
[
0.1258544921875,
0.6494140625,
0.4794921875,
0.6494140625,
0.4794921875,
0.66357421875,
0.1258544921875,
0.66357421875
],
[
0.51708984375,
0.65087890625,
0.91259765625,
0.65087890625,
0.91259765625,
0.6640625,
0.51708984375,
0.6640625
],
[
0.12744140625,
0.66552734375,
0.4794921875,
0.66552734375,
0.4794921875,
0.67529296875,
0.12744140625,
0.67529296875
],
[
0.5185546875,
0.66552734375,
0.9111328125,
0.66552734375,
0.9111328125,
0.67529296875,
0.5185546875,
0.67529296875
],
[
0.10540771484375,
0.6796875,
0.48046875,
0.6796875,
0.48046875,
0.68994140625,
0.10540771484375,
0.68994140625
],
[
0.5185546875,
0.6796875,
0.9111328125,
0.6796875,
0.9111328125,
0.68994140625,
0.5185546875,
0.68994140625
],
[
0.12335205078125,
0.69140625,
0.48193359375,
0.69140625,
0.48193359375,
0.705078125,
0.12335205078125,
0.705078125
],
[
0.51708984375,
0.6923828125,
0.9140625,
0.6923828125,
0.9140625,
0.70556640625,
0.51708984375,
0.70556640625
],
[
0.1258544921875,
0.705078125,
0.4794921875,
0.705078125,
0.4794921875,
0.71826171875,
0.1258544921875,
0.71826171875
],
[
0.51708984375,
0.70703125,
0.9140625,
0.70556640625,
0.9140625,
0.7197265625,
0.51708984375,
0.72021484375
],
[
0.10540771484375,
0.7216796875,
0.48046875,
0.7216796875,
0.48046875,
0.73193359375,
0.10540771484375,
0.73193359375
],
[
0.5185546875,
0.7216796875,
0.91259765625,
0.7216796875,
0.91259765625,
0.73193359375,
0.5185546875,
0.73193359375
],
[
0.1258544921875,
0.736328125,
0.48193359375,
0.736328125,
0.48193359375,
0.7470703125,
0.1258544921875,
0.7470703125
],
[
0.125,
0.74853515625,
0.483642578125,
0.74755859375,
0.483642578125,
0.76123046875,
0.125,
0.76220703125
],
[
0.5185546875,
0.7470703125,
0.73876953125,
0.7470703125,
0.73876953125,
0.76123046875,
0.5185546875,
0.76123046875
],
[
0.1258544921875,
0.76318359375,
0.281005859375,
0.76318359375,
0.281005859375,
0.7734375,
0.1258544921875,
0.7734375
],
[
0.51708984375,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.77783203125,
0.51708984375,
0.77783203125
],
[
0.51708984375,
0.77880859375,
0.91259765625,
0.77880859375,
0.91259765625,
0.79248046875,
0.51708984375,
0.79248046875
],
[
0.0882568359375,
0.79150390625,
0.1004638671875,
0.79150390625,
0.1004638671875,
0.7998046875,
0.0882568359375,
0.7998046875
],
[
0.1094970703125,
0.78857421875,
0.25732421875,
0.78857421875,
0.25732421875,
0.8017578125,
0.1094970703125,
0.8017578125
],
[
0.51708984375,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.51708984375,
0.806640625
],
[
0.0849609375,
0.80615234375,
0.483642578125,
0.80615234375,
0.483642578125,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.515625,
0.806640625,
0.9140625,
0.806640625,
0.9140625,
0.82080078125,
0.515625,
0.82080078125
],
[
0.0849609375,
0.82080078125,
0.48193359375,
0.82080078125,
0.48193359375,
0.833984375,
0.0849609375,
0.833984375
],
[
0.5185546875,
0.82275390625,
0.91259765625,
0.82275390625,
0.91259765625,
0.83349609375,
0.5185546875,
0.83349609375
],
[
0.0849609375,
0.833984375,
0.48193359375,
0.833984375,
0.48193359375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.51708984375,
0.83544921875,
0.9140625,
0.83544921875,
0.9140625,
0.8486328125,
0.51708984375,
0.8486328125
],
[
0.08660888671875,
0.849609375,
0.4794921875,
0.849609375,
0.4794921875,
0.85986328125,
0.08660888671875,
0.85986328125
],
[
0.51708984375,
0.8486328125,
0.91259765625,
0.8486328125,
0.91259765625,
0.8623046875,
0.51708984375,
0.8623046875
],
[
0.08416748046875,
0.87060546875,
0.450927734375,
0.8720703125,
0.450927734375,
0.88623046875,
0.08416748046875,
0.88525390625
],
[
0.51708984375,
0.8720703125,
0.91259765625,
0.8720703125,
0.91259765625,
0.88525390625,
0.51708984375,
0.88525390625
],
[
0.0849609375,
0.8837890625,
0.373291015625,
0.8837890625,
0.373291015625,
0.89794921875,
0.0849609375,
0.89794921875
],
[
0.51318359375,
0.8837890625,
0.9140625,
0.88330078125,
0.9140625,
0.89794921875,
0.51318359375,
0.8984375
]
] |
[
"wWW'23, May 15, 2023, Austin, TX, UsA",
"Sola Shirai, Debarun Bhattacharjya, and Oktie Hassanzadeh",
"However, a major limitation of most LP methods is that relations",
"prediction\" techniques. Other notable work that falls under this",
"can only be predicted among entities that already exist in the KG.",
"class of methods includes Radinsky et al's Pundit algorithm [17]",
"A more practical and challenging perspectivefor this task is to",
"which is based on automated construction of a causality graph",
"consider LP for entirely new entities (ie., inductive link prediction).",
"through extraction of causal patterns and generalization. Zhao et",
"This perspective also better reflects the needs of an event prediction",
"al. [33] also construct a causal graph from text documents and build",
"system,which would aim tomakepredictions about entirely new",
"embeddingstoperformasimple(one-hop)prediction.Whileprior",
"events.Theinductivesettingintroducesanadditionallevelofdiffi",
"work has explored applying LP methods to KGs of causal events",
"culty for LP models, as many of them (especially embedding-based",
"[10, 22], to our knowledge, our work is the first to apply 2-hop LP",
"ones) operate under the closed world assumption - that is, they can",
"over aknowledgegraphforeventprediction.",
"only perform predictions for entities that have been seen during",
"In the space of LP in KGs, a significant number of embedding-",
"training,and facts that are not seen during training are assumed",
"based methods have been developed as detailed in recent surveys",
"to be false. Additionally, even for models that are capable of this",
"[29, 30]. Such methods often show trade-offs in performance based",
"task,someKGs (suchasWikidata)constantlyundergoupdatesand",
"onthedatasetandhowwellthemodelhyperparametersaretuned",
"changes,whichcannegativelyaffectamodel'sperformanceas its",
"[19]. While GNN-based models [31] have often shown recent state-",
"training data becomes outdated.",
"of-the-art performance, there has been discussion surrounding what",
"Towards our goal of applying KGs to perform event prediction,",
"semantics are actually captured in such models [9] as well as re-",
"wedevelopacase-basedreasoningapproach,EvCBR,whichlever-",
"evaluationsofhowoneevaluatessuchLPmodels[1,24.25].Further-",
"ages examples of past events in the KG to perform LP for the prop-",
"more, most embedding-based models are trained and evaluated in",
"ertiesofunseenevententities.Weframeourproblemasa2-hop",
"the closed world setting,witha relatively sparse number of recent",
"link prediction task - e.g., starting from a cause event, we assume",
"works such as[2, 6] considering inductive link prediction using",
"the existence of a causal relation to a new effect event and predict",
"additional knowledge sources such as text or hyper-relational facts.",
"its properties. This approach allows us to perform inductive link",
"While less prevalent than embedding-based models, there are",
"predictionforthepropertiesoftheunseeneffectwithouttheneec",
"also a variety of rule-based LP models [7, 11, 13, 15, 20]. Methods",
"for external data.Importantly,EvCBR requires no training,relying",
"such as AnyBURL [12] have shown similar performance to state-of-",
"instead on statistical measures and subclass hierarchies to identify",
"the-art embedding models while requiring significantly less training",
"similar entities and follow paths through the KG to perform pre-",
"time [18]. Works such as [4, 5] present case-based reasoning models",
"dictions. This makes EvCBR well-suited to KGs such as Wikidata",
"which, like our method, require no training and make path-based",
"that frequently experience changes in their structure and content.",
"predictions.Our method differs from such similar works in our",
"More generally, we can also apply EvCBR to perform 2-hop LP",
"method of computing entity similarity,approach to apply and refine",
"forpropertiesof anytypeofunseenentitythatis connectedtoa",
"predictionpaths,andtaskformulationofperforming2-hopLP",
"known entityby somerelation",
"Our contributions are as follows:",
"3",
"",
"PROBLEMFORMULATION",
"(1) We introduce a case-based reasoning model, EvCBR,2 for",
"Our work considers the application of LP methods over a knowledge",
"event prediction. Our model considers this task as a LP task",
"graph for the task of event prediction. We define a knowledge graph",
"ina2-hopsetting,leveragingknowledgeaboutsimilarcause",
"G = (E,R, T) as consisting of a set of entities E, a set of relations",
"effecteventstomakepredictionsabouttheunseeneffect",
"R, and a set of triples T. T consists of triples of the form (h, r, t),",
"EvCBR requires no training, making it well-suited to handle",
"where h, t e E and r e R.",
"new events and changes to the underlying KG.",
"The task of LP is then to correctly predict the missing tail entity",
"(2)Comparedtosimilarwork,weintroducenovel similarity",
"given a partial triple (h, r, ?t).4 The standard assumption in this task",
"metrics to identify similar cause-effect event cases as well",
"is that the head entity h and the missing tail entity ?t are known",
"asarefinementstepto improvetheprecision of predictions",
"entities inE.To performinductive linkprediction,it is necessary",
"(3) We curate and releasea novel dataset surrounding causal",
"topredictthemissing?t where?t E.SinceLPmodels can not",
"events in Wikidata, extracting news events that are con-",
"predict such an entity ?t solely from the content of G, inductive link",
"nected by causal relations as well as their local connections.3",
"prediction methods often also rely on additional background data",
"(4)Inour2-hop inductivelinkpredictiontask,ourmodelshows",
"(such as textual descriptions)to extract andrepresent new entities.",
"superior performance on our event dataset as well as com-",
"petitive performance on a modified evaluation datasetbased",
"3.1 EventPrediction Task",
"ontheFB15k-237dataset.",
"Extending the task of LP to event prediction, our goal is to make",
"predictions about new events based on a causal relation between",
"2",
"RELATED WORK",
"two events. We represent our task as a query triple (c,r,e),where",
"Zhao [32] presents a comprehensive survey of different kinds of",
"c, r, and e denote the cause event, causal relation, and effect event,",
"event prediction methods across different domains. Under their tax-",
"respectively.Furthermore,thepoint of the event prediction task is",
"onomy of event prediction methods [32, Fig. 3], our method is most",
"to predict properties about the new effect event -ie., predicting",
"closely related to the\"causality-based\"methods under\"semantic",
"outgoing properties about e in the form of (e, re,?z) e Oute, where",
"2We publicly release our code at https//github.com/solashirai/WWW-EvCBR",
"4we note the related LP tasks of predicting a missing head ?h or relation?r, but in this",
"3Dataset available at https:/doi.org/10.5281/zenodo.7196049.",
"Paper we focus on LP for predicting ?t as italigns best with our event prediction task."
] |
{
"arxiv_id": "2309.12423",
"language": "en",
"page_idx": 1
}
|
||
2309.12423-en-2
|
2309.12423-en-2.jpg
|
Event Prediction using Case-Based Reasoning over Knowledge Graphs
$r_{e}\in R$ and ?$z\in E.$ In this task, we assume that $c\in E$ while $e\notin E-$ if we wish to make predictions about a cause $c$ which was not in the original KG, we assume that triples indicating the properties of $c$ are supplied as input and added to the KG. Additionally, we assume that we are given the relations $r_e$ to predict.
$Example.$ In Figure 1, our task is to perform predictions for New Effect based on the triple (New Cause, hasEffect, New Effect). Further, we make predictions about each of New Effect's properties, as (New Effect, instance0f, ?$z)$ and (New Effect, country, ?$z) .$ $\square$
To make predictions under these assumptions, we must consider two points: first, we must make a prediction for an unseen effect event $e$ given a cause $c$ of interest and a relation $r;$ second, from the unseen event $e$,we must make predictions about its properties by performing LP to tail entities ?$z\in E.$
Given this problem formulation for event prediction, we can then perform predictions for properties of $e$ without the need for any background knowledge about $e$ by performing a 2-hop LP task starting from $c-$that is, predicting the tail entity for the 2-hop link $(c,r,r_e,?z)$for each triple in $Out_e.$ By omitting explicit consideration of $e$ in the prediction task, we are able to overcome the limitation that LP models must have a way of extracting or representing the entity. From a practical perspective, when performing event predictions, one need not explicitly represent $e$ as the primary focus is on predicting properties about the effect event.
<FigureHere>
Figure 2: The task setup for our running example.
We additionally denote the set of outgoing triples from $c$ as $Out_c-$we assume that outgoing properties of the cause event $c$ are provided or already present in the KG, and our method only makes use of such outgoing triples$^5$ from c. A graphical depiction of this problem setup for our running example is shown in Figure 2.
3.2 A General 2-Hop Prediction Task
We pose a generalization of the event prediction task, which could be useful in other applications: to predict properties about a novel entity based on a known relation between that entity and some entity in $E.$ Given some representative relation $r\in R$ that connects an entity $h\in E$ to a novel entity $e_n\notin E$, we can make predictions about its properties $(e_n,r_n,t_n)$ using the 2-hop link $(h,r,r_n,?t)$, where $?t\in E.$ This generalization operates under the assumption that the relation connecting $h$ and $e_n$ is semantically meaningful such that we can make useful predictions about $e_n$ based on knowledge about $h.$ The validity of this assumption may vary for different types of relations.
# 4 METHODS
EvCBR is a case-based reasoning approach which aims to make predictions about new events based on background knowledge about
$^{5}$We are motivated to only consider outgoing edges because we can expect a new event
entered into the KG to have little to no incoming relations present.
WWW'23, May 1-5, 2023, Austin, TX, USA
similar cause-effect events. While our methods can be generalized to make predictions about new entities in other domains (conceptually replacing the “cause event”and“effect event" with different types of entities and connecting them by different relations as mentioned in the previous section), we describe our approach from the perspective of event prediction in this section.
At a high level, case-based reasoning aims to solve a new problem based on experience and knowledge about similar cases in the system's case base. In our task, the target problem refers to a new cause-effect event pair query for which we wish to make predictions, and a case refers to an event pair present in the KG. This approach can then be broken down as (1) retrieve cases from the KG which are similar to the new problem query, (2) identify and score paths through the KG that can be used to predict properties of the effect events in each case, and (3) reuse the learned paths to make predictions for our new problem.
# 4.1 Case Retrieval
We define a case s as a triple $(c_s,r,e_s)$ connecting the cause $c_s$ to the effect $e_s$ by the causal relation $r$, where $(c_s,r,e_s)\in T.$ In our running example KG, there are three example cases of cause and effect events connected by the hasEffect relation.
The frst step of EvCBR is to retrieve cases from the KG that are similar to our problem, which we break down into two subtasks. First, we compute a similarity measure among entities in $E$ based on their subclass hierarchy and outgoing connections in the KG. Second, we compute the similarity of cases to the query $(c,r,e).$
4.1.1 Entity Similarity. To compute entity similarity, we apply a relatively simple similarity metric which can be computed based on count statistics and vector multiplication. Our goal in this step is to acquire a very rough sense of similarity between individual entities, which will be utilized in our subsequent steps.
For each entity $h_i\in E$,we form a vector representation $v_{h_i}=$ $[ a_{1}, a_{2}, . . . , a_{{N_{h}}}] $, where $N_{h}$ is the total number of entities in$E.$ Each value $a_{j}= 1$ if there is a triple $( h_{i}, r, t_{j}) \in T$ for any relation $r$ or if $h_{i}$ is a subclass of $a_j$, and otherwise $a_j=0.$
Intuitively, this vector allows us to capture the idea that similar
entities are likely to be connected to the same entities
$Example.$ The Kanto Earthquake and Tohoku Earthquake both have an outgoing connection to Japan, which might indicate that they are more similar than other earthquake events that occurred in different countries. Additionally, while these two earthquakes do not have any other outgoing connections that are identical, the subclass0f relation between Megathrust Earthquake and Earthquake indicates that both the Kanto and Tohoku Earthquakes are a type of Earthquake.
口
We also weight the importance of each entity in this vector based on how frequently it occurs as an outgoing neighbor or superclass. Our intuition here is to ensure that sharing outgoing connections to a less common entity is more meaningful than sharing connections to a common one. We compute this weight as an inverse document frequency (IDF) measure, $IDF(h)=\log(N_{h}/count(h))$, where count(h) is the number of times $h$ has an incoming edge or is a superclass of an entity, and form the vector containing the
|
[
[
0.08660888671875,
0.07574462890625,
0.416748046875,
0.07574462890625,
0.416748046875,
0.08587646484375,
0.08660888671875,
0.08587646484375
],
[
0.70751953125,
0.0738525390625,
0.9111328125,
0.07452392578125,
0.9111328125,
0.08837890625,
0.70751953125,
0.08709716796875
],
[
0.08660888671875,
0.106689453125,
0.48046875,
0.106689453125,
0.48046875,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.51708984375,
0.106689453125,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.51708984375,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1351318359375,
0.9140625,
0.1351318359375,
0.9140625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.515625,
0.149658203125,
0.9140625,
0.1483154296875,
0.9140625,
0.1622314453125,
0.515625,
0.1634521484375
],
[
0.0849609375,
0.1641845703125,
0.400390625,
0.1641845703125,
0.400390625,
0.177978515625,
0.0849609375,
0.177978515625
],
[
0.51708984375,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.10296630859375,
0.18310546875,
0.4794921875,
0.18310546875,
0.4794921875,
0.1962890625,
0.10296630859375,
0.1962890625
],
[
0.515625,
0.177978515625,
0.7939453125,
0.1767578125,
0.7939453125,
0.1900634765625,
0.515625,
0.1912841796875
],
[
0.0849609375,
0.1962890625,
0.48193359375,
0.1976318359375,
0.48193359375,
0.2108154296875,
0.0849609375,
0.210205078125
],
[
0.53173828125,
0.1912841796875,
0.9140625,
0.1912841796875,
0.9140625,
0.2052001953125,
0.53173828125,
0.2052001953125
],
[
0.5185546875,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.5185546875,
0.2183837890625
],
[
0.08660888671875,
0.21337890625,
0.48046875,
0.21337890625,
0.48046875,
0.2235107421875,
0.08660888671875,
0.2235107421875
],
[
0.08416748046875,
0.2247314453125,
0.483642578125,
0.2247314453125,
0.483642578125,
0.238037109375,
0.08416748046875,
0.238037109375
],
[
0.51953125,
0.2203369140625,
0.91259765625,
0.2203369140625,
0.91259765625,
0.2310791015625,
0.51953125,
0.2310791015625
],
[
0.51708984375,
0.23291015625,
0.9140625,
0.23291015625,
0.9140625,
0.246826171875,
0.51708984375,
0.246826171875
],
[
0.10296630859375,
0.2392578125,
0.48193359375,
0.2392578125,
0.48193359375,
0.25244140625,
0.10296630859375,
0.25244140625
],
[
0.0849609375,
0.25244140625,
0.48046875,
0.25244140625,
0.48046875,
0.266357421875,
0.0849609375,
0.266357421875
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.51708984375,
0.260009765625,
0.91259765625,
0.260009765625,
0.91259765625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.08660888671875,
0.268310546875,
0.48046875,
0.268310546875,
0.48046875,
0.279052734375,
0.08660888671875,
0.279052734375
],
[
0.515625,
0.272705078125,
0.9140625,
0.2734375,
0.9140625,
0.287353515625,
0.515625,
0.285888671875
],
[
0.08660888671875,
0.281005859375,
0.48193359375,
0.281005859375,
0.48193359375,
0.294189453125,
0.08660888671875,
0.294189453125
],
[
0.0849609375,
0.294189453125,
0.325927734375,
0.294189453125,
0.325927734375,
0.30810546875,
0.0849609375,
0.30810546875
],
[
0.5185546875,
0.28857421875,
0.91259765625,
0.28857421875,
0.91259765625,
0.298583984375,
0.5185546875,
0.298583984375
],
[
0.51708984375,
0.3017578125,
0.9140625,
0.3017578125,
0.9140625,
0.31494140625,
0.51708984375,
0.31494140625
],
[
0.101318359375,
0.30810546875,
0.48193359375,
0.30810546875,
0.48193359375,
0.3212890625,
0.101318359375,
0.3212890625
],
[
0.0849609375,
0.322509765625,
0.48193359375,
0.322509765625,
0.48193359375,
0.3359375,
0.0849609375,
0.3359375
],
[
0.5185546875,
0.317626953125,
0.75341796875,
0.317626953125,
0.75341796875,
0.32763671875,
0.5185546875,
0.32763671875
],
[
0.08416748046875,
0.3359375,
0.48193359375,
0.335205078125,
0.48193359375,
0.348388671875,
0.08416748046875,
0.349853515625
],
[
0.0882568359375,
0.3515625,
0.48046875,
0.3515625,
0.48046875,
0.3623046875,
0.0882568359375,
0.3623046875
],
[
0.5185546875,
0.349853515625,
0.67724609375,
0.349853515625,
0.67724609375,
0.363037109375,
0.5185546875,
0.363037109375
],
[
0.08660888671875,
0.3642578125,
0.48046875,
0.3642578125,
0.48046875,
0.37744140625,
0.08660888671875,
0.37744140625
],
[
0.51708984375,
0.367431640625,
0.91259765625,
0.367431640625,
0.91259765625,
0.3818359375,
0.51708984375,
0.3818359375
],
[
0.0849609375,
0.376953125,
0.48046875,
0.376953125,
0.48046875,
0.39013671875,
0.0849609375,
0.39013671875
],
[
0.51708984375,
0.3818359375,
0.91259765625,
0.3818359375,
0.91259765625,
0.395751953125,
0.51708984375,
0.395751953125
],
[
0.08660888671875,
0.391357421875,
0.48193359375,
0.391357421875,
0.48193359375,
0.40478515625,
0.08660888671875,
0.40478515625
],
[
0.5185546875,
0.397705078125,
0.91259765625,
0.397705078125,
0.91259765625,
0.40771484375,
0.5185546875,
0.40771484375
],
[
0.08660888671875,
0.40478515625,
0.48046875,
0.40478515625,
0.48046875,
0.41845703125,
0.08660888671875,
0.41845703125
],
[
0.5185546875,
0.410888671875,
0.82275390625,
0.410888671875,
0.82275390625,
0.421630859375,
0.5185546875,
0.421630859375
],
[
0.0882568359375,
0.42041015625,
0.4794921875,
0.42041015625,
0.4794921875,
0.43115234375,
0.0882568359375,
0.43115234375
],
[
0.53515625,
0.4248046875,
0.9111328125,
0.4248046875,
0.9111328125,
0.43505859375,
0.53515625,
0.43505859375
],
[
0.08416748046875,
0.43310546875,
0.381591796875,
0.43310546875,
0.381591796875,
0.4462890625,
0.08416748046875,
0.4462890625
],
[
0.51953125,
0.439453125,
0.91259765625,
0.439453125,
0.91259765625,
0.449462890625,
0.51953125,
0.449462890625
],
[
0.51953125,
0.45263671875,
0.91259765625,
0.45263671875,
0.91259765625,
0.46337890625,
0.51953125,
0.46337890625
],
[
0.1707763671875,
0.4609375,
0.1871337890625,
0.464111328125,
0.1846923828125,
0.47021484375,
0.1683349609375,
0.46728515625
],
[
0.5185546875,
0.46533203125,
0.91259765625,
0.46533203125,
0.91259765625,
0.479248046875,
0.5185546875,
0.479248046875
],
[
0.11517333984375,
0.477783203125,
0.1478271484375,
0.477783203125,
0.1478271484375,
0.48486328125,
0.11517333984375,
0.48486328125
],
[
0.12335205078125,
0.474853515625,
0.1422119140625,
0.474853515625,
0.1422119140625,
0.477783203125,
0.12335205078125,
0.477783203125
],
[
0.515625,
0.477783203125,
0.90185546875,
0.479248046875,
0.90185546875,
0.492431640625,
0.515625,
0.4912109375
],
[
0.51953125,
0.50830078125,
0.91259765625,
0.50830078125,
0.91259765625,
0.521484375,
0.51953125,
0.521484375
],
[
0.51708984375,
0.521484375,
0.9140625,
0.521484375,
0.9140625,
0.53515625,
0.51708984375,
0.53515625
],
[
0.1168212890625,
0.52880859375,
0.44775390625,
0.52978515625,
0.44775390625,
0.54443359375,
0.1168212890625,
0.54345703125
],
[
0.51708984375,
0.53515625,
0.91259765625,
0.53515625,
0.91259765625,
0.548828125,
0.51708984375,
0.548828125
],
[
0.5185546875,
0.55029296875,
0.9111328125,
0.55029296875,
0.9111328125,
0.56103515625,
0.5185546875,
0.56103515625
],
[
0.10296630859375,
0.556640625,
0.48046875,
0.556640625,
0.48046875,
0.57080078125,
0.10296630859375,
0.57080078125
],
[
0.515625,
0.5625,
0.85107421875,
0.5625,
0.85107421875,
0.57568359375,
0.515625,
0.57568359375
],
[
0.08660888671875,
0.57080078125,
0.48046875,
0.57080078125,
0.48046875,
0.583984375,
0.08660888671875,
0.583984375
],
[
0.53173828125,
0.5771484375,
0.90185546875,
0.5771484375,
0.90185546875,
0.59033203125,
0.53173828125,
0.59033203125
],
[
0.08660888671875,
0.583984375,
0.48046875,
0.583984375,
0.48046875,
0.59765625,
0.08660888671875,
0.59765625
],
[
0.5185546875,
0.59130859375,
0.9140625,
0.59130859375,
0.9140625,
0.60498046875,
0.5185546875,
0.60498046875
],
[
0.0849609375,
0.599609375,
0.48193359375,
0.599609375,
0.48193359375,
0.61279296875,
0.0849609375,
0.61279296875
],
[
0.51708984375,
0.60498046875,
0.9150390625,
0.60498046875,
0.9150390625,
0.61865234375,
0.51708984375,
0.61865234375
],
[
0.08660888671875,
0.61572265625,
0.451904296875,
0.61572265625,
0.451904296875,
0.62548828125,
0.08660888671875,
0.62548828125
],
[
0.51708984375,
0.61865234375,
0.77685546875,
0.61865234375,
0.77685546875,
0.6318359375,
0.51708984375,
0.6318359375
],
[
0.53173828125,
0.6318359375,
0.9140625,
0.6318359375,
0.9140625,
0.64599609375,
0.53173828125,
0.64599609375
],
[
0.0849609375,
0.63818359375,
0.396240234375,
0.63818359375,
0.396240234375,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.5185546875,
0.64794921875,
0.8447265625,
0.64794921875,
0.8447265625,
0.65771484375,
0.5185546875,
0.65771484375
],
[
0.0849609375,
0.6572265625,
0.48193359375,
0.6572265625,
0.48193359375,
0.67041015625,
0.0849609375,
0.67041015625
],
[
0.0849609375,
0.66943359375,
0.48046875,
0.67041015625,
0.48046875,
0.6845703125,
0.0849609375,
0.68310546875
],
[
0.53369140625,
0.6748046875,
0.91259765625,
0.6748046875,
0.91259765625,
0.68798828125,
0.53369140625,
0.68798828125
],
[
0.0849609375,
0.68505859375,
0.48046875,
0.68505859375,
0.48046875,
0.69873046875,
0.0849609375,
0.69873046875
],
[
0.51708984375,
0.689453125,
0.91259765625,
0.689453125,
0.91259765625,
0.70263671875,
0.51708984375,
0.70263671875
],
[
0.08660888671875,
0.70068359375,
0.4794921875,
0.70068359375,
0.4794921875,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.51953125,
0.705078125,
0.9111328125,
0.705078125,
0.9111328125,
0.71533203125,
0.51953125,
0.71533203125
],
[
0.08660888671875,
0.7138671875,
0.48046875,
0.7138671875,
0.48046875,
0.724609375,
0.08660888671875,
0.724609375
],
[
0.51953125,
0.71826171875,
0.91259765625,
0.71826171875,
0.91259765625,
0.72900390625,
0.51953125,
0.72900390625
],
[
0.0849609375,
0.7265625,
0.48046875,
0.7265625,
0.48046875,
0.74072265625,
0.0849609375,
0.74072265625
],
[
0.51708984375,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.08660888671875,
0.7421875,
0.4794921875,
0.7421875,
0.4794921875,
0.75244140625,
0.08660888671875,
0.75244140625
],
[
0.51953125,
0.74560546875,
0.9111328125,
0.74560546875,
0.9111328125,
0.755859375,
0.51953125,
0.755859375
],
[
0.08660888671875,
0.755859375,
0.48046875,
0.755859375,
0.48046875,
0.7666015625,
0.08660888671875,
0.7666015625
],
[
0.5185546875,
0.75830078125,
0.91259765625,
0.75830078125,
0.91259765625,
0.771484375,
0.5185546875,
0.771484375
],
[
0.08660888671875,
0.76953125,
0.48193359375,
0.76953125,
0.48193359375,
0.77978515625,
0.08660888671875,
0.77978515625
],
[
0.515625,
0.771484375,
0.6396484375,
0.76806640625,
0.6396484375,
0.78271484375,
0.515625,
0.7861328125
],
[
0.90625,
0.77294921875,
0.9140625,
0.77880859375,
0.90673828125,
0.78271484375,
0.900390625,
0.77783203125
],
[
0.0849609375,
0.7822265625,
0.48193359375,
0.7822265625,
0.48193359375,
0.79541015625,
0.0849609375,
0.79541015625
],
[
0.53173828125,
0.78515625,
0.9140625,
0.78515625,
0.9140625,
0.798828125,
0.53173828125,
0.798828125
],
[
0.08416748046875,
0.796875,
0.19775390625,
0.796875,
0.19775390625,
0.81005859375,
0.08416748046875,
0.81005859375
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.51708984375,
0.81298828125,
0.91259765625,
0.814453125,
0.91259765625,
0.82568359375,
0.51708984375,
0.82470703125
],
[
0.08251953125,
0.82080078125,
0.205078125,
0.8193359375,
0.205078125,
0.833984375,
0.08251953125,
0.83544921875
],
[
0.5185546875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.08416748046875,
0.8369140625,
0.48046875,
0.83837890625,
0.48046875,
0.85302734375,
0.08416748046875,
0.8515625
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.8515625,
0.48193359375,
0.85302734375,
0.48193359375,
0.8671875,
0.0849609375,
0.8662109375
],
[
0.5146484375,
0.85400390625,
0.91259765625,
0.85498046875,
0.91259765625,
0.869140625,
0.5146484375,
0.86865234375
],
[
0.0849609375,
0.87060546875,
0.483642578125,
0.8720703125,
0.483642578125,
0.88623046875,
0.0849609375,
0.88525390625
],
[
0.5185546875,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.5185546875,
0.880859375
],
[
0.0849609375,
0.8837890625,
0.400390625,
0.8837890625,
0.400390625,
0.89794921875,
0.0849609375,
0.89794921875
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Event Prediction using Case-Based Reasoning over Knowledge Graphs",
"wWW'23, May 1-5,2023, Austin, Tx, usA",
"re e R and ?z e E. In this task, we assume that c E while e E-",
"similar cause-effect events. While our methods can be generalized",
"if we wish to make predictions about a cause c which was not in",
"to make predictions about new entities in other domains (concep-",
"the original KG, we assume that triples indicating the properties",
"tually replacing the “cause event” and \"effect event\" with different",
"of c are supplied as input and added to the KG. Additionally, we",
"types of entities and connecting them by different relations as men-",
"assume that we are given the relations re to predict.",
"tioned in the previous section), we describe our approach from the",
"Example. In Figure 1, our task is to perform predictions for New",
"perspective of event prediction in this section",
"Effectbased on the triple(NewCause,hasEffect,NewEffect)",
"At a high level, case-based reasoning aims to solve a new prob-",
"lem based on experience and knowledge about similar cases in the",
"Further,wemakepredictions abouteachofNewEffect'sproperties",
"as (New Effect, instanceof,?z) and (New Effect, country,?z).",
"system's case base.In our task, the target problem refers to a new",
"cause-effect event pair query for which we wish to make predic-",
"To make predictions under these assumptions, we must consider",
"two points: first, we must make a prediction for an unseen effect",
"tions, and a case refers to an event pair present in the KG. This",
"approach can then be broken down as (1) retrieve cases from the",
"event e given a cause c of interest and a relation r; second, from",
"KGwhichare similartothenewproblemquery,(2)identifyand",
"the unseen event e, we must make predictions about its properties",
"by performing LP to tail entities ?z e E.",
"scorepathsthroughtheKGthatcanbeusedtopredictproperties",
"of the effect events in each case, and (3) reuse the learned paths to",
"Given this problem formulation for event prediction, we can",
"then perform predictions for properties of e without the need for",
"makepredictions forournewproblem.",
"anybackground knowledge about e by performing a 2-hopLP task",
"startingfromc-that is,predictingthetail entityfor the2-hoplink",
"4.1CaseRetrieval",
"(c,r,re,?z)foreachtriple in Oute.By omitting explicitconsideration",
"We define a case s as a triple (cs, r, es) connecting the cause cs to",
"of e in the prediction task, we are able to overcome the limitation",
"the effect es by the causal relation r, where (cs, r,es) e T. In our",
"that LP models must have a way of extracting or representing",
"running exampleKG,therearethreeexamplecasesof causeand",
"the entity. From a practical perspective, when performing event",
"effecteventsconnectedbythehasEffectrelation.",
"predictions, oneneed notexplicitlyrepresenteastheprimaryfocus",
"ThefirststepofEvCBRistoretrievecasesfromtheKGthatare",
"is on predicting properties about the effect event.",
"similarto ourproblem,whichwebreakdown into twosubtasks",
"First,we compute a similarity measureamong entities in E based",
"out",
"on their subclass hierarchy and outgoing connections in the KG",
"arthguake",
"aathn",
"Second, we compute the similarity of cases to the query (c,r,e).",
"4.1.1Entity Similarity.To compute entity similarity,we apply a",
"relatively simple similarity metric which can be computed based on",
"Figure 2: The task setup for our running example.",
"count statistics and vector multiplication. Our goal in this step is",
"to acquire a very rough sense of similarity between individual",
"We additionally denote the set of outgoing triples from c as",
"entities, which will be utilized in our subsequent steps.",
"Oute- we assume that outgoing properties of the cause event c are",
"For each entity hi E, we form a vector representation Uhi",
"provided or already present in the KG, and our method only makes",
"[ai,a2.., aN, ], where Nh is the total number of entities in E. Each",
"use of such outgoing triples from c. A graphical depiction of this",
"value aj = 1 if there is a triple (hi,r,t) e T for any relation r or if",
"problem setup for our running exampleis shown in Figure 2.",
"h, is a subclass of aj, and otherwise aj = 0.",
"Intuitively, this vector allows us to capture the idea that similar",
"3.2A General 2-Hop Prediction Task",
"entitiesarelikelytobeconnected tothesameentities.",
"We pose a generalization of the event prediction task, which could",
"be useful in other applications: to predict properties about a novel",
"Example. The Kanto Earthquake and Tohoku Earthquake both",
"entity based on a known relation between that entity and some",
"have an outgoing connection to Japan, which might indicate that",
"entity in E.Given somerepresentativerelation r ERthat connects",
"theyaremoresimilar thanother earthquakeeventsthat occurred",
"anentityhEEtoanovelentityenE,wecanmakepredictions",
"indifferent countries.Additionally,whilethesetwoearthquakes",
"about its properties (en, rn, tn) using the 2-hop link (h, r, rn, ?t)",
"do not have any other outgoing connections that are identical, the",
"where?t e E.This generalization operates under the assumption",
"subclassOfrelationbetweenMegathrustEarthguakeandEarth",
"that therelation connecting h and en is semantically meaningful",
"quake indicates that both the Kanto and Tohoku Earthquakes are a",
"suchthatwecanmakeuseful predictionsaboutenbasedonknowl",
"type of Earthquake.",
"1",
"edge about h. The validity of this assumption may vary for different",
"We also weight the importance of each entity in this vector based",
"types ofrelations.",
"on how frequently it occurs as an outgoing neighbor or superclass.",
"Our intuition here isto ensurethat sharing outgoing connections",
"4METHODS",
"toaless commonentityismoremeaningfulthan sharingconnec",
"EvCBR is a case-based reasoning approach which aims to make pre",
"tions to a common one. We compute this weight as an inverse",
"dictions about new events based on background knowledge about",
"document frequency (IDF) measure, IDF(h)=log(Nn/count(h)",
"Swe are motivated to only consider outgoing edges because we can expect anew event",
"wherecount(h)isthenumberof timeshhasanincomingedge",
"entered into the KG to have little to no incoming relations present.",
" or is a superclass of an entity, and form the vector containing the"
] |
{
"arxiv_id": "2309.12423",
"language": "en",
"page_idx": 2
}
|
||
2309.12423-en-4
|
2309.12423-en-4.jpg
|
Event Prediction using Case-Based Reasoning over Knowledge Graphs
head similarity $CS_h-$out of our three example cases, the Tohoku
口
Earthquake case will have the highest CaseScore.
4.2 Prediction Path Enumeration and Scoring
We next use the retrieved cases to enumerate paths through the KG
that can be used to make predictions about $e.$
4.2.1 Path Enumeration. Our first step is to enumerate a set of paths, using cases retrieved from the KG, which can be used to connect cause events to properties of their effect events. We define a path as a sequence of triples through the KG that can connect two entities. Given a start entity $x$ and end entity $y$, we can express a path of length $n$ connecting them as a sequence of triples, $p=[(x,r_{1},e_{1}),...,(e_{n-1},r_{n},y)].$ We define the relation path as the sequence of relations used in each triple of a path, denoted $rel(p)=(r_1,...,r_n).$ We denote the list of entities that can be reached starting from entity $x$ following the relation path $rel(p)$ as $\mathbb{E}_{p,x}.$ Note that $\mathbb{E}_{p,x}$ is not a set, and may contain repeated entries of an entity that can be reached by different paths through the KG.
For each case $(c_s,r,e_s)$ in our set of retrieved cases S and relation $r_{e}\in ROut_{e}$,let $E_r_{e},e_{s}$ denote the set of entities connected to $e_s$ by the outgoing relation $r_e$,i.e., $(e_s,r_e,t)\in T$ for all $t\in E_r_e,e_s.$ We then randomly sample up to $N_p$ unique paths of length $\leq3$ connecting $c_s$ to any entity in $E_r_e,e_s.$
<FigureHere>
Figure 3: Example paths for the Tohoku Earthquake case.
Here, we restrict the paths that are sampled such that (1) the first relation of $rel(p)$ must be an outgoing relation of our cause event $c$, and (2) the path does not traverse through the case's effect event $e_s.$ Our first restriction ensures that we identify paths that are relevant to our query event- since our goal is to follow relation paths starting from the event $c$,if the frst relation is not an outgoing relation of $c$ it will not be possible to utilize the path. Also note that the triple $(c,r,e)$ is not present in the KG, so a path starting with $r$ would only be valid if c has some other outgoing triple with the relation $r.$ Similarly, our second restriction aims to prevent us from sampling paths that traverse through the causal relation connecting the cause and effect event. Again, since $e$ is unseen in the KG, such a path would not provide us with any useful information to apply to our event query.
$Example.$ Figure 3 shows two example paths, highlighted in red (light) and blue (dark), connecting the Tohoku Earthquake event to its effect's instance0f property. $rel(p_{red})$ consists of the three relations [instanceOf, subclassOf, hasCause$^-1]$ while $rel(p_{blue})$ is the single relation [instanceOf]. We do not sample paths that
口
traverse through the Aftermath event.
The path sampling steps are repeated for each case in S, and the set of all paths sampled from each case, connecting the case's cause to its effect's $r_e$ properties, is denoted as $\rho_S,r_e.$
WWW'23, May 1-5, 2023, Austin, TX, USA
4.2.2 Path Scoring. Next, we compute confidence scores for each unique relation path present in $\mathcal{P}_{S,r_{tq}}.$ This confidence score will correspond to the notion of how confident we are that a given path leads to the correct entity prediction.
We base our scoring on a simple precision measure, aiming to score how well a given relation path leads to the correct entities for the target relation. Additionally, following from the path confidence measure implemented in [12],we add a smoothing constant $\epsilon$ (set to $\epsilon=5$ in our experiments) to the denominator when calculating the precision- this allows for relation paths with the same precision and a greater number of samples to have a higher score than relation paths with fewer samples. The relation path score for a target relation to predict, $r_e$, and relation path $rel(p)$ is given as:
$$PScore(r_e,rel(p))=\frac{\sum_{(c_s,r,e_s)\in S}\sum_{t^{\prime}\in\mathbb{E}_{p,c_s}}\mathbb{1}\left[t^{\prime}\in E_{r_e,e_s}\right]}{\epsilon+\sum_{(c_s,r,e_s)\in S}\left|\mathbb{E}_{p,c_s}\right|}$$
(9)
where $1[t^{\prime}\in E_{r_{e},e_{s}}]=1$ if $t^\prime\in E_{r_{e},e_{s}}$ is true and 0 otherwise. The PScore for a particular $rel(p)$ may be influenced by sampling
multiple different paths that have the same relation path. This allows us to implicitly add a weighting to each relation path based on how frequently they occur, which corresponds to how likely they are to be randomly sampled.
$Example.$ To make a prediction about the effect's country in our three cases, we can sample the relation path $rel(p)=[$country]. For all three of our cases, following the relation path [country] from the cause leads to the correct country property of its effect. The PScore of this path would then be 3/(ε+3). On the other hand, for a relation path like $rel(p)=[$instanceOf] from our example in Figure 3, which aims to predict the instance0f relation of the effect, the correct entity is only reached in 1 out of 2 results for the Tohoku Earthquake, and 0 out of 3 total results for the other two
口
cases. The PScore for this path would then be 1/(ε+5).
Based on these path scores, produced for the set of distinct relation paths in $p_{S,r_{e}}$, we select $N_{p}$ relation paths with the highest scores, denoted $\mathcal{P}_{r_{e}}$, to make predictions for $(e,r_{e},?z).$
## 4.3 Applying Prediction Paths
Given our set of relation paths $\mathcal{P}_{r_e}$, selected and scored using our set of retrieved cases, we can now apply these paths to our cause event c to make predictions about property $r_e$ of the effect event. We perform these predictions by following each relation path, starting from our cause $c$, and using the PScore for each path to produce a total confidence score for each predicted entity
${\dot{\text{Formally, for our query event }(c,r,e)\text{ and a single property for}}}$ which we want to predict $(e,r_e,?z)$,we score each candidate prediction entity $z$ as follows:
$$EScore(c,r,r_{e},z)=\sum_{\begin{array}{c}rel(p)\in\mathcal{P}_{r_{e}}\\z^{\prime}\in\mathbb{E}_{p,c}\end{array}}PScore(r_{e},rel(p))[z^{\prime}=z]$$
In Equation 10, we choose to calculate the score for a prediction $z$ as a summation of path PScores so that the score is increased for (1) entities that are frequently reached by a given relation path, and (2) entities that are reached by a variety of different relation paths.
|
[
[
0.08660888671875,
0.0770263671875,
0.418212890625,
0.0770263671875,
0.418212890625,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.705078125,
0.0738525390625,
0.91259765625,
0.07452392578125,
0.91259765625,
0.08837890625,
0.705078125,
0.08709716796875
],
[
0.0849609375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.51708984375,
0.106689453125,
0.91259765625,
0.106689453125,
0.91259765625,
0.12060546875,
0.51708984375,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.3857421875,
0.121826171875,
0.3857421875,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.5146484375,
0.121826171875,
0.7958984375,
0.121826171875,
0.7958984375,
0.1351318359375,
0.5146484375,
0.1351318359375
],
[
0.7998046875,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.7998046875,
0.1351318359375
],
[
0.5185546875,
0.136962890625,
0.91259765625,
0.136962890625,
0.91259765625,
0.147705078125,
0.5185546875,
0.147705078125
],
[
0.4697265625,
0.12371826171875,
0.48193359375,
0.12371826171875,
0.48193359375,
0.1331787109375,
0.4697265625,
0.1331787109375
],
[
0.08416748046875,
0.14453125,
0.465576171875,
0.146484375,
0.465576171875,
0.1622314453125,
0.08416748046875,
0.160400390625
],
[
0.51708984375,
0.15087890625,
0.8095703125,
0.15087890625,
0.8095703125,
0.1641845703125,
0.51708984375,
0.1641845703125
],
[
0.0849609375,
0.1654052734375,
0.48193359375,
0.1654052734375,
0.48193359375,
0.1787109375,
0.0849609375,
0.1787109375
],
[
0.53125,
0.1634521484375,
0.9140625,
0.1641845703125,
0.9140625,
0.177978515625,
0.53125,
0.1767578125
],
[
0.0849609375,
0.179931640625,
0.359375,
0.179931640625,
0.359375,
0.1932373046875,
0.0849609375,
0.1932373046875
],
[
0.5185546875,
0.179931640625,
0.91259765625,
0.179931640625,
0.91259765625,
0.1900634765625,
0.5185546875,
0.1900634765625
],
[
0.515625,
0.1925048828125,
0.9140625,
0.1925048828125,
0.9140625,
0.205810546875,
0.515625,
0.205810546875
],
[
0.0849609375,
0.2008056640625,
0.483642578125,
0.2008056640625,
0.483642578125,
0.2139892578125,
0.0849609375,
0.2139892578125
],
[
0.08416748046875,
0.21337890625,
0.483642578125,
0.212158203125,
0.483642578125,
0.2266845703125,
0.08416748046875,
0.2279052734375
],
[
0.5185546875,
0.2083740234375,
0.9111328125,
0.2083740234375,
0.9111328125,
0.2183837890625,
0.5185546875,
0.2183837890625
],
[
0.5185546875,
0.2215576171875,
0.91259765625,
0.2215576171875,
0.91259765625,
0.231689453125,
0.5185546875,
0.231689453125
],
[
0.08416748046875,
0.2279052734375,
0.48193359375,
0.2266845703125,
0.48193359375,
0.240478515625,
0.08416748046875,
0.2412109375
],
[
0.51708984375,
0.2342529296875,
0.9140625,
0.2342529296875,
0.9140625,
0.2474365234375,
0.51708984375,
0.2474365234375
],
[
0.08251953125,
0.240478515625,
0.483642578125,
0.2412109375,
0.483642578125,
0.255126953125,
0.08251953125,
0.25390625
],
[
0.51708984375,
0.2474365234375,
0.9140625,
0.2474365234375,
0.9140625,
0.261474609375,
0.51708984375,
0.261474609375
],
[
0.08251953125,
0.25390625,
0.483642578125,
0.255126953125,
0.483642578125,
0.26953125,
0.08251953125,
0.268310546875
],
[
0.51708984375,
0.261474609375,
0.9140625,
0.261474609375,
0.9140625,
0.274658203125,
0.51708984375,
0.274658203125
],
[
0.0849609375,
0.26953125,
0.48193359375,
0.26953125,
0.48193359375,
0.28271484375,
0.0849609375,
0.28271484375
],
[
0.515625,
0.27587890625,
0.86669921875,
0.27587890625,
0.86669921875,
0.2890625,
0.515625,
0.2890625
],
[
0.08416748046875,
0.28271484375,
0.48193359375,
0.2822265625,
0.48193359375,
0.296630859375,
0.08416748046875,
0.29736328125
],
[
0.0849609375,
0.296630859375,
0.483642578125,
0.296630859375,
0.483642578125,
0.31005859375,
0.0849609375,
0.31005859375
],
[
0.65869140625,
0.304931640625,
0.87841796875,
0.3037109375,
0.87841796875,
0.322509765625,
0.65869140625,
0.323974609375
],
[
0.0849609375,
0.311279296875,
0.48193359375,
0.311279296875,
0.48193359375,
0.324462890625,
0.0849609375,
0.324462890625
],
[
0.53759765625,
0.31640625,
0.6650390625,
0.31640625,
0.6650390625,
0.330078125,
0.53759765625,
0.330078125
],
[
0.89453125,
0.317626953125,
0.9140625,
0.317626953125,
0.9140625,
0.330078125,
0.89453125,
0.330078125
],
[
0.0849609375,
0.324462890625,
0.48046875,
0.324462890625,
0.48046875,
0.33837890625,
0.0849609375,
0.33837890625
],
[
0.69775390625,
0.323974609375,
0.83837890625,
0.32568359375,
0.83837890625,
0.343505859375,
0.69775390625,
0.341552734375
],
[
0.0849609375,
0.33837890625,
0.48193359375,
0.33837890625,
0.48193359375,
0.352783203125,
0.0849609375,
0.352783203125
],
[
0.53173828125,
0.3447265625,
0.90185546875,
0.3447265625,
0.90185546875,
0.3623046875,
0.53173828125,
0.3623046875
],
[
0.0849609375,
0.352783203125,
0.48193359375,
0.352783203125,
0.48193359375,
0.3662109375,
0.0849609375,
0.3662109375
],
[
0.53173828125,
0.35986328125,
0.91259765625,
0.35986328125,
0.91259765625,
0.373779296875,
0.53173828125,
0.373779296875
],
[
0.101318359375,
0.36865234375,
0.48046875,
0.36865234375,
0.48046875,
0.37890625,
0.101318359375,
0.37890625
],
[
0.08416748046875,
0.380126953125,
0.483642578125,
0.37890625,
0.483642578125,
0.393310546875,
0.08416748046875,
0.39453125
],
[
0.51708984375,
0.375732421875,
0.9111328125,
0.375732421875,
0.9111328125,
0.386474609375,
0.51708984375,
0.386474609375
],
[
0.0849609375,
0.393310546875,
0.48193359375,
0.393310546875,
0.48193359375,
0.40771484375,
0.0849609375,
0.40771484375
],
[
0.51708984375,
0.38818359375,
0.9140625,
0.38818359375,
0.9140625,
0.401611328125,
0.51708984375,
0.401611328125
],
[
0.51708984375,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.51708984375,
0.415283203125
],
[
0.08416748046875,
0.40771484375,
0.483642578125,
0.40771484375,
0.483642578125,
0.421630859375,
0.08416748046875,
0.421630859375
],
[
0.08087158203125,
0.417236328125,
0.23779296875,
0.42041015625,
0.23779296875,
0.438232421875,
0.08087158203125,
0.43505859375
],
[
0.51708984375,
0.417236328125,
0.7197265625,
0.417236328125,
0.7197265625,
0.427978515625,
0.51708984375,
0.427978515625
],
[
0.53515625,
0.4375,
0.9140625,
0.4375,
0.9140625,
0.45068359375,
0.53515625,
0.45068359375
],
[
0.188720703125,
0.45263671875,
0.22216796875,
0.45263671875,
0.22216796875,
0.460205078125,
0.188720703125,
0.460205078125
],
[
0.27294921875,
0.45703125,
0.31201171875,
0.45703125,
0.31201171875,
0.46728515625,
0.27294921875,
0.46728515625
],
[
0.51708984375,
0.451904296875,
0.9140625,
0.451904296875,
0.9140625,
0.46533203125,
0.51708984375,
0.46533203125
],
[
0.11517333984375,
0.460205078125,
0.15283203125,
0.460205078125,
0.15283203125,
0.46728515625,
0.11517333984375,
0.46728515625
],
[
0.416748046875,
0.462158203125,
0.44775390625,
0.462158203125,
0.44775390625,
0.473388671875,
0.416748046875,
0.473388671875
],
[
0.51708984375,
0.46533203125,
0.9140625,
0.46533203125,
0.9140625,
0.479248046875,
0.51708984375,
0.479248046875
],
[
0.23779296875,
0.473388671875,
0.28759765625,
0.481689453125,
0.281982421875,
0.5,
0.2320556640625,
0.4912109375
],
[
0.1094970703125,
0.46533203125,
0.156005859375,
0.46533203125,
0.156005859375,
0.47607421875,
0.1094970703125,
0.47607421875
],
[
0.339111328125,
0.479736328125,
0.371826171875,
0.479736328125,
0.371826171875,
0.4912109375,
0.339111328125,
0.4912109375
],
[
0.51708984375,
0.479736328125,
0.91259765625,
0.479736328125,
0.91259765625,
0.49365234375,
0.51708984375,
0.49365234375
],
[
0.414306640625,
0.484130859375,
0.451904296875,
0.484130859375,
0.451904296875,
0.49755859375,
0.414306640625,
0.49755859375
],
[
0.1641845703125,
0.48291015625,
0.1953125,
0.48291015625,
0.1953125,
0.4873046875,
0.1641845703125,
0.4873046875
],
[
0.236083984375,
0.49560546875,
0.28515625,
0.49560546875,
0.28515625,
0.505859375,
0.236083984375,
0.505859375
],
[
0.5185546875,
0.49365234375,
0.9140625,
0.49365234375,
0.9140625,
0.5068359375,
0.5185546875,
0.5068359375
],
[
0.191162109375,
0.5,
0.2115478515625,
0.5,
0.2115478515625,
0.5068359375,
0.191162109375,
0.5068359375
],
[
0.404296875,
0.501953125,
0.463134765625,
0.501953125,
0.463134765625,
0.5087890625,
0.404296875,
0.5087890625
],
[
0.131591796875,
0.51025390625,
0.1502685546875,
0.51025390625,
0.1502685546875,
0.51806640625,
0.131591796875,
0.51806640625
],
[
0.41259765625,
0.51123046875,
0.439453125,
0.51123046875,
0.439453125,
0.51513671875,
0.41259765625,
0.51513671875
],
[
0.5185546875,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.51953125,
0.5185546875,
0.51953125
],
[
0.40283203125,
0.51513671875,
0.459228515625,
0.51513671875,
0.459228515625,
0.52587890625,
0.40283203125,
0.52587890625
],
[
0.51708984375,
0.521484375,
0.9140625,
0.521484375,
0.9140625,
0.53515625,
0.51708984375,
0.53515625
],
[
0.51708984375,
0.5341796875,
0.91259765625,
0.5341796875,
0.91259765625,
0.54736328125,
0.51708984375,
0.54736328125
],
[
0.09234619140625,
0.54150390625,
0.47216796875,
0.54150390625,
0.47216796875,
0.5546875,
0.09234619140625,
0.5546875
],
[
0.51708984375,
0.548828125,
0.9140625,
0.548828125,
0.9140625,
0.5625,
0.51708984375,
0.5625
],
[
0.51708984375,
0.5625,
0.85302734375,
0.5625,
0.85302734375,
0.57568359375,
0.51708984375,
0.57568359375
],
[
0.90185546875,
0.564453125,
0.91259765625,
0.564453125,
0.91259765625,
0.57373046875,
0.90185546875,
0.57373046875
],
[
0.101318359375,
0.5712890625,
0.48193359375,
0.5712890625,
0.48193359375,
0.58544921875,
0.101318359375,
0.58544921875
],
[
0.53173828125,
0.57568359375,
0.9140625,
0.57568359375,
0.9140625,
0.5888671875,
0.53173828125,
0.5888671875
],
[
0.0849609375,
0.583984375,
0.48046875,
0.583984375,
0.48046875,
0.59765625,
0.0849609375,
0.59765625
],
[
0.515625,
0.59033203125,
0.9140625,
0.59033203125,
0.9140625,
0.60498046875,
0.515625,
0.60498046875
],
[
0.0849609375,
0.60107421875,
0.48046875,
0.60107421875,
0.48046875,
0.611328125,
0.0849609375,
0.611328125
],
[
0.51708984375,
0.60400390625,
0.84619140625,
0.60400390625,
0.84619140625,
0.61865234375,
0.51708984375,
0.61865234375
],
[
0.08660888671875,
0.61279296875,
0.4794921875,
0.61279296875,
0.4794921875,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.0849609375,
0.626953125,
0.48193359375,
0.626953125,
0.48193359375,
0.64013671875,
0.0849609375,
0.64013671875
],
[
0.515625,
0.630859375,
0.775390625,
0.6298828125,
0.775390625,
0.64453125,
0.515625,
0.64599609375
],
[
0.08660888671875,
0.642578125,
0.48046875,
0.642578125,
0.48046875,
0.65283203125,
0.08660888671875,
0.65283203125
],
[
0.51708984375,
0.6494140625,
0.9140625,
0.6494140625,
0.9140625,
0.66357421875,
0.51708984375,
0.66357421875
],
[
0.08660888671875,
0.65576171875,
0.48046875,
0.65576171875,
0.48046875,
0.66650390625,
0.08660888671875,
0.66650390625
],
[
0.0849609375,
0.66845703125,
0.48193359375,
0.66845703125,
0.48193359375,
0.681640625,
0.0849609375,
0.681640625
],
[
0.515625,
0.66357421875,
0.9140625,
0.66357421875,
0.9140625,
0.6767578125,
0.515625,
0.6767578125
],
[
0.0849609375,
0.681640625,
0.48193359375,
0.681640625,
0.48193359375,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.51708984375,
0.6767578125,
0.9140625,
0.6767578125,
0.9140625,
0.68994140625,
0.51708984375,
0.68994140625
],
[
0.51708984375,
0.69140625,
0.9140625,
0.69140625,
0.9140625,
0.705078125,
0.51708984375,
0.705078125
],
[
0.08660888671875,
0.69775390625,
0.48046875,
0.69775390625,
0.48046875,
0.70849609375,
0.08660888671875,
0.70849609375
],
[
0.0849609375,
0.71044921875,
0.48193359375,
0.71044921875,
0.48193359375,
0.7236328125,
0.0849609375,
0.7236328125
],
[
0.515625,
0.705078125,
0.9140625,
0.705078125,
0.9140625,
0.71826171875,
0.515625,
0.71826171875
],
[
0.515625,
0.71826171875,
0.88134765625,
0.71826171875,
0.88134765625,
0.73193359375,
0.515625,
0.73193359375
],
[
0.08660888671875,
0.72607421875,
0.48046875,
0.72607421875,
0.48046875,
0.736328125,
0.08660888671875,
0.736328125
],
[
0.08251953125,
0.736328125,
0.48193359375,
0.73779296875,
0.48193359375,
0.75244140625,
0.08251953125,
0.75048828125
],
[
0.53173828125,
0.73291015625,
0.9140625,
0.73291015625,
0.9140625,
0.7470703125,
0.53173828125,
0.7470703125
],
[
0.51708984375,
0.7470703125,
0.9140625,
0.7470703125,
0.9140625,
0.76025390625,
0.51708984375,
0.76025390625
],
[
0.08660888671875,
0.75390625,
0.2017822265625,
0.75390625,
0.2017822265625,
0.7646484375,
0.08660888671875,
0.7646484375
],
[
0.5185546875,
0.76220703125,
0.67724609375,
0.76220703125,
0.67724609375,
0.77294921875,
0.5185546875,
0.77294921875
],
[
0.10296630859375,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.78515625,
0.10296630859375,
0.78515625
],
[
0.08660888671875,
0.787109375,
0.48046875,
0.787109375,
0.48046875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.5361328125,
0.79541015625,
0.65625,
0.79541015625,
0.65625,
0.81005859375,
0.5361328125,
0.81005859375
],
[
0.7099609375,
0.79345703125,
0.9140625,
0.79541015625,
0.9140625,
0.81005859375,
0.7099609375,
0.80810546875
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.814453125,
0.0849609375,
0.814453125
],
[
0.08416748046875,
0.81103515625,
0.48046875,
0.81298828125,
0.48046875,
0.82763671875,
0.08416748046875,
0.82568359375
],
[
0.650390625,
0.80810546875,
0.7158203125,
0.81005859375,
0.7158203125,
0.82470703125,
0.650390625,
0.82275390625
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.6640625,
0.82373046875,
0.705078125,
0.82373046875,
0.705078125,
0.83544921875,
0.6640625,
0.83544921875
],
[
0.08416748046875,
0.84130859375,
0.31787109375,
0.84130859375,
0.31787109375,
0.85498046875,
0.08416748046875,
0.85498046875
],
[
0.471435546875,
0.84521484375,
0.4794921875,
0.84521484375,
0.4794921875,
0.85107421875,
0.471435546875,
0.85107421875
],
[
0.53173828125,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.53173828125,
0.85498046875
],
[
0.1004638671875,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.1004638671875,
0.86865234375
],
[
0.5146484375,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.5146484375,
0.86865234375
],
[
0.0849609375,
0.87060546875,
0.48046875,
0.87060546875,
0.48046875,
0.880859375,
0.0849609375,
0.880859375
],
[
0.51708984375,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.08251953125,
0.8818359375,
0.36767578125,
0.8837890625,
0.36767578125,
0.8984375,
0.08251953125,
0.896484375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Event Prediction using Case-Based Reasoning over Knowledge Graphs",
"wWW'23, May 1-5, 2023, Austin, Tx, UsA",
"head similarity CSh - out of our three example cases, the Tohoku",
"4.2.2Path Scoring.Next, we compute confidence scores for",
"Earthquake case will have the highest CaseScore.",
"each unique relation path present in Psr",
" This confidence",
"scorewill correspondtothenotionofhowconfidentwearethata",
"口",
"4.2Prediction Path Enumeration and Scoring",
"given path leads to the correct entity prediction.",
"We next use the retrieved cases to enumerate paths through the KG",
"We base our scoring on a simple precision measure, aiming to",
"that can be used to make predictions about e.",
"scorehowwell agivenrelationpathleadstothe correctentitiesfor",
"the target relation. Additionally,following from the path confidence",
"4.2.1 Path Enumeration. Our first step is to enumerate a set of",
"paths, using cases retrieved from the KG, which can be used",
"measureimplementedin[12],weaddasmoothingconstante(setto",
"e=5in ourexperiments)tothedenominator when calculatingthe",
"to connect cause events toproperties of their effect events",
"precision -this allows for relation paths with the same precision",
"We define a path as a sequence of triples through the KG that",
"and a greater number of samples to have a higher score than relation",
"can connect two entities. Given a start entity x and end entity y.",
"paths with fewer samples. The relation path score for a target",
"we can express a path of length n connecting them as a sequence",
"relation to predict,re,and relation path rel(p) is given as:",
"of triples,p =[(x, ri,ei)..,(en-1,rn,y)].We define the relation",
"path as the sequence of relations used in each triple of a path,",
"Z(csres)es Zv'eBpe, I[t' e Eree,]",
"denoted rel(p) = (ri,.,rn). We denote the list of entities that can",
"PScore(re,rel(p)) =",
"(9)",
"be reached starting from entity x following the relation path rel(p)",
"e+E(cs,r,es)es Ep.csl",
"as Ep,x. Note that Ep,x is not a set, and may contain repeated entries",
"where I[t' e Ere,e,] = 1 ift' e Ere,e, is true and O otherwise.",
"of an entity that can be reached by different paths through the KG.",
"The PScore for a particular rel(p)may be influenced by sampling",
"Foreach case(cs,r,es)in our set ofretrieved cases Sand relation",
"re e ROute, let Erees denote the set of entities connected to es by",
"multiple differentpaths that havethesame relation path.This",
"the outgoingrelation re,ie., (es,re,t) eTforallt Ere,es.Wethen",
"allows us to implicitly add a weighting to each relation path based",
"on how frequently they occur, which corresponds to how likely",
"randomly sample up to N, unique paths of length ≤ 3 connecting",
"Cs to any entity in Ere,es.",
"they are to be randomly sampled.",
"Example. To make a prediction about the effect's country in our",
"subclassO",
"Earthquake",
"three cases, we can sample the relation path rel(p) =[country]. For",
"Megathrus",
"Tsunami",
"all three of our cases, following the relation path [country] from",
"201 Tohaku",
"Earthquake",
"nstanceof",
"the cause leads to the correct country property of its effect. The",
"Instanceof",
"nstanceC",
"and Tsunami",
"PScore of this path would then be 3/(e + 3). On the other hand,",
"ountry",
"Aftermathofthe",
"Japar",
"2011:T",
"forarelationpathlikerel(p)=[instanceOflfromourexample",
"Earthquake",
"in Figure 3, which aims to predict the instanceof relation of the",
"effect, the correct entity is only reached in 1 out of 2 results for the",
"Figure 3: Example paths for the Tohoku Earthquake case",
"Tohoku Earthquake, and 0 out of 3 total results for the other two",
"cases. The PScore for this path would then be 1/(e + 5).",
"口",
"Here, we restrict the paths that are sampled such that (1) the",
"Based on these path scores, produced for the set of distinct re-",
"first relation of rel(p) must be an outgoing relation of our cause",
"lation paths in Ps,re, we select Np relation paths with the highest",
"event c, and (2)the path does not traversethrough the case's effect",
"scores, denoted Pre, to make predictions for (e,re, ?z).",
"event es.Ourfirst restriction ensuresthat we identifypaths that",
"are relevant to our query event - since our goal is to follow relation",
"4.3ApplyingPredictionPaths",
"pathsstartingfromtheeventc,ifthefirst relationisnot anoutgoing",
"Given our set of relation paths Pre, selected and scored using our",
"relation of c it will not be possible to utilize the path.Also note that",
"the triple (c, r,e) is not present in the KG, so a path starting with",
"set of retrieved cases, we can now apply these paths to our cause",
"r would only be valid if c has some other outgoing triple with the",
"event c to make predictions about property re of the effect",
"event. We perform these predictions by following each relation",
"relation r.Similarly,our second restriction aims to prevent us from",
"sampling paths that traverse through the causal relation connecting",
"path, starting from our cause c, and using the PScore for each path",
"to produce a total confidence score for each predicted entity.",
"the cause and effect event.Again, since e is unseen in theKG, such",
"a path would not provide us with any useful information to apply",
"Formally, for our query event (c, r, e) and a single property for",
"which we want to predict (e, re, ?z), we score each candidate pre-",
"to our event query.",
"diction entity z as follows:",
"Example. Figure 3 shows two example paths, highlighted in red",
"(light)andblue(dark),connectingtheTohokuEarthquakeevent",
"EScore(c, r, re, z) =",
"PScore(re,rel(p)[z = z] (10)",
"to its effect's instanceof property. rel(pPred) consists of the three",
"relations [instanceOf, subclassOf, hasCause-1] while rel(pblue)",
"rel(p)ePre",
"isthesinglerelation[instanceOf.Wedo notsamplepathsthat",
"Z'eBp,c",
"traverse through the Aftermath event.",
"口",
"In Equation 10, we choose to calculate the score for a prediction",
"The path sampling steps are repeated for each case in S, and the",
"z as a summation of path PScores so that the score is increased for",
"setofall pathssampledfromeachcase,connectingthecase'scause",
"(1) entities that are frequently reached by a given relation path, and",
"to its effect's re properties, is denoted as Ps,re.",
"(2) entities that are reached by a variety of different relation paths."
] |
{
"arxiv_id": "2309.12423",
"language": "en",
"page_idx": 4
}
|
||
2004.10157-en-0
|
2004.10157-en-0.jpg
|
## Logic-Guided Data Augmentation and Regularization for Consistent Question Answering
Akari Asait and Hannaneh Hajishirzi$^{[\mathrm{t}]}$
$\dagger$University of Washington $\ddagger$Allen Institute for AI $\{$akari, hannaneh$\}$ecs.washington.edu
### $\mathbf{Abstract}$
Many natural language questions require qualitative, quantitative or logical comparisons between two entities or events. This paper addresses the problem of improving the accuracy and consistency of responses to comparison questions by integrating logic rules and neural models. Our method leverages logical and linguistic knowledge to augment labeled training data and then uses a consistency-based regularizer to train the model. Improving the global consistency of predictions, our approach achieves large improvements over previous methods in a variety of question answering (QA) tasks including multiple-choice qualitative reasoning, cause-effect reasoning, and extractive machine reading comprehension. In particular, our method significantly improves the performance of RoBERTa-based models by 1-5% across datasets. We advance state of the art by around 5-8% on WIQA and QuaRel and reduce consistency violations by 58% on HotpotQA. We further demonstrate that our approach can learn effectively from limited $data.^{\mathrm{l}}$
### 1 Introduction
Comparison-type questions (Tandon et al., 2019; Tafjord et al., 2019; Yang et al., 2018) ask about relationships between properties of entities or events such as cause-effect, qualitative or quantitative reasoning. To create comparison questions that require inferential knowledge and reasoning ability, annotators need to understand context presented in multiple paragraphs or carefully ground a question to the given situation. This makes it challenging to annotate a large number of comparison questions. Most current datasets on comparison questions are much smaller than standard machine reading comprehension (MRC) datasets (Rajpurkar
${^{1}\text{Our code and data is available at https://github.}}$
com/AkariAsai/logic\_guided\_qa.
<FigureHere>
Figure 1: Inconsistent predictions by RoBERTa. Top row shows an example of symmetric inconsistency and the second row shows an example of transitive inconsistency. The examples are partially modified.
et al., 2016; Joshi et al., 2017). This poses new challenges to standard models, which are known to exploit statistical patterns or annotation artifacts in these datasets (Sugawara et al., 2018; Min et al., 2019a). Importantly, state-of-the-art models show inconsistent comparison predictions as shown in Figure 1... Improving the consistency of predictions has been previously studied in natural language inference (NLI) tasks (Minervini and Riedel, 2018; Li et al., 2019), but has not been addressed in QA.
In this paper, we address the task of producing globally consistent and accurate predictions for comparison questions leveraging logical and symbolic knowledge for data augmentation and training regularization. Our data augmentation uses a set of logical and linguistic knowledge to develop additional consistent labeled training data. Subsequently, our method uses symbolic logic to incorporate consistency regularization for additional supervision signal beyond inductive bias given by data augmentation. Our method generalizes previous consistency-promoting methods for NLI tasks (Minervini and Riedel, 2018; Li et al., 2019)
|
[
[
0.225830078125,
0.0843505859375,
0.77685546875,
0.0843505859375,
0.77685546875,
0.097412109375,
0.225830078125,
0.097412109375
],
[
0.31982421875,
0.09979248046875,
0.68408203125,
0.10272216796875,
0.68408203125,
0.12054443359375,
0.31982421875,
0.1175537109375
],
[
0.330078125,
0.152587890625,
0.67822265625,
0.15087890625,
0.67822265625,
0.168701171875,
0.330078125,
0.17041015625
],
[
0.30322265625,
0.171630859375,
0.7060546875,
0.171630859375,
0.7060546875,
0.185302734375,
0.30322265625,
0.185302734375
],
[
0.297119140625,
0.188232421875,
0.7138671875,
0.188232421875,
0.7138671875,
0.203125,
0.297119140625,
0.203125
],
[
0.51904296875,
0.26416015625,
0.68408203125,
0.26416015625,
0.68408203125,
0.27490234375,
0.51904296875,
0.27490234375
],
[
0.8134765625,
0.26416015625,
0.84814453125,
0.26416015625,
0.84814453125,
0.275634765625,
0.8134765625,
0.275634765625
],
[
0.51904296875,
0.273681640625,
0.6513671875,
0.273681640625,
0.6513671875,
0.283935546875,
0.51904296875,
0.283935546875
],
[
0.26611328125,
0.265380859375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.26611328125,
0.27978515625
],
[
0.8076171875,
0.273681640625,
0.8671875,
0.273681640625,
0.8671875,
0.283935546875,
0.8076171875,
0.283935546875
],
[
0.145263671875,
0.289306640625,
0.4609375,
0.290283203125,
0.4609375,
0.3046875,
0.145263671875,
0.303955078125
],
[
0.5205078125,
0.283935546875,
0.60205078125,
0.283935546875,
0.60205078125,
0.29150390625,
0.5205078125,
0.29150390625
],
[
0.73486328125,
0.29150390625,
0.79443359375,
0.29150390625,
0.79443359375,
0.302734375,
0.73486328125,
0.302734375
],
[
0.0218353271484375,
0.2998046875,
0.0679931640625,
0.2998046875,
0.06463623046875,
0.439453125,
0.019317626953125,
0.439453125
],
[
0.1435546875,
0.3046875,
0.462646484375,
0.303955078125,
0.462646484375,
0.318359375,
0.1435546875,
0.319580078125
],
[
0.51904296875,
0.303955078125,
0.69189453125,
0.303955078125,
0.69189453125,
0.314208984375,
0.51904296875,
0.314208984375
],
[
0.78515625,
0.308837890625,
0.80029296875,
0.308837890625,
0.80029296875,
0.320556640625,
0.78515625,
0.320556640625
],
[
0.5205078125,
0.3154296875,
0.6513671875,
0.3154296875,
0.6513671875,
0.322509765625,
0.5205078125,
0.322509765625
],
[
0.8095703125,
0.311279296875,
0.84814453125,
0.3134765625,
0.84619140625,
0.32470703125,
0.8076171875,
0.322509765625
],
[
0.14697265625,
0.318359375,
0.4609375,
0.318359375,
0.4609375,
0.33203125,
0.14697265625,
0.33203125
],
[
0.51708984375,
0.322509765625,
0.59765625,
0.322509765625,
0.59765625,
0.333251953125,
0.51708984375,
0.333251953125
],
[
0.81689453125,
0.322509765625,
0.8671875,
0.322509765625,
0.8671875,
0.333251953125,
0.81689453125,
0.333251953125
],
[
0.1485595703125,
0.334228515625,
0.456787109375,
0.334228515625,
0.456787109375,
0.344482421875,
0.1485595703125,
0.344482421875
],
[
0.145263671875,
0.3466796875,
0.4609375,
0.347412109375,
0.4609375,
0.361083984375,
0.145263671875,
0.35986328125
],
[
0.51904296875,
0.342529296875,
0.705078125,
0.342529296875,
0.705078125,
0.352783203125,
0.51904296875,
0.352783203125
],
[
0.5205078125,
0.355224609375,
0.59765625,
0.355224609375,
0.59765625,
0.3623046875,
0.5205078125,
0.3623046875
],
[
0.8330078125,
0.355712890625,
0.85205078125,
0.355712890625,
0.85205078125,
0.363525390625,
0.8330078125,
0.363525390625
],
[
0.14697265625,
0.3623046875,
0.4609375,
0.3623046875,
0.4609375,
0.3759765625,
0.14697265625,
0.3759765625
],
[
0.51904296875,
0.3623046875,
0.68603515625,
0.3623046875,
0.68603515625,
0.372314453125,
0.51904296875,
0.372314453125
],
[
0.14697265625,
0.3759765625,
0.4609375,
0.3759765625,
0.4609375,
0.3896484375,
0.14697265625,
0.3896484375
],
[
0.51708984375,
0.374755859375,
0.6689453125,
0.374755859375,
0.6689453125,
0.385498046875,
0.51708984375,
0.385498046875
],
[
0.73486328125,
0.380126953125,
0.796875,
0.3818359375,
0.7958984375,
0.396728515625,
0.73291015625,
0.394287109375
],
[
0.83447265625,
0.3818359375,
0.85400390625,
0.3818359375,
0.85400390625,
0.385498046875,
0.83447265625,
0.385498046875
],
[
0.14697265625,
0.39013671875,
0.4609375,
0.39013671875,
0.4609375,
0.40380859375,
0.14697265625,
0.40380859375
],
[
0.51904296875,
0.38427734375,
0.6826171875,
0.38427734375,
0.6826171875,
0.394287109375,
0.51904296875,
0.394287109375
],
[
0.51904296875,
0.394287109375,
0.68603515625,
0.394287109375,
0.68603515625,
0.405029296875,
0.51904296875,
0.405029296875
],
[
0.14697265625,
0.405029296875,
0.4609375,
0.405029296875,
0.4609375,
0.418701171875,
0.14697265625,
0.418701171875
],
[
0.51708984375,
0.410400390625,
0.67236328125,
0.410400390625,
0.67236328125,
0.42333984375,
0.51708984375,
0.42333984375
],
[
0.7900390625,
0.40625,
0.8017578125,
0.413330078125,
0.79248046875,
0.419189453125,
0.78271484375,
0.410888671875
],
[
0.14697265625,
0.418701171875,
0.37451171875,
0.418701171875,
0.37451171875,
0.431640625,
0.14697265625,
0.431640625
],
[
0.37451171875,
0.42041015625,
0.45751953125,
0.42041015625,
0.45751953125,
0.43115234375,
0.37451171875,
0.43115234375
],
[
0.5224609375,
0.421630859375,
0.5634765625,
0.421630859375,
0.5634765625,
0.4287109375,
0.5224609375,
0.4287109375
],
[
0.587890625,
0.429931640625,
0.68701171875,
0.429931640625,
0.68701171875,
0.440673828125,
0.587890625,
0.440673828125
],
[
0.372802734375,
0.42333984375,
0.38037109375,
0.42333984375,
0.38037109375,
0.427001953125,
0.372802734375,
0.427001953125
],
[
0.5205078125,
0.431640625,
0.57861328125,
0.431640625,
0.57861328125,
0.439453125,
0.5205078125,
0.439453125
],
[
0.8369140625,
0.429931640625,
0.8486328125,
0.427001953125,
0.85205078125,
0.429931640625,
0.84033203125,
0.43408203125
],
[
0.14697265625,
0.432861328125,
0.4609375,
0.432861328125,
0.4609375,
0.446533203125,
0.14697265625,
0.446533203125
],
[
0.0218353271484375,
0.435302734375,
0.06463623046875,
0.435302734375,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.14697265625,
0.44775390625,
0.4609375,
0.44775390625,
0.4609375,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.5146484375,
0.45361328125,
0.88330078125,
0.45361328125,
0.88330078125,
0.46728515625,
0.5146484375,
0.46728515625
],
[
0.1485595703125,
0.463134765625,
0.4609375,
0.463134765625,
0.4609375,
0.473876953125,
0.1485595703125,
0.473876953125
],
[
0.51318359375,
0.468505859375,
0.88330078125,
0.468505859375,
0.88330078125,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.145263671875,
0.4755859375,
0.462646484375,
0.4755859375,
0.462646484375,
0.4892578125,
0.145263671875,
0.4892578125
],
[
0.5146484375,
0.482177734375,
0.884765625,
0.482177734375,
0.884765625,
0.495361328125,
0.5146484375,
0.495361328125
],
[
0.1485595703125,
0.4912109375,
0.459228515625,
0.4912109375,
0.459228515625,
0.501953125,
0.1485595703125,
0.501953125
],
[
0.51123046875,
0.496337890625,
0.82373046875,
0.496337890625,
0.82373046875,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.14697265625,
0.50439453125,
0.4609375,
0.50439453125,
0.4609375,
0.51708984375,
0.14697265625,
0.51708984375
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.1485595703125,
0.51953125,
0.459228515625,
0.51953125,
0.459228515625,
0.52978515625,
0.1485595703125,
0.52978515625
],
[
0.14697265625,
0.5322265625,
0.4609375,
0.5322265625,
0.4609375,
0.5458984375,
0.14697265625,
0.5458984375
],
[
0.51318359375,
0.5390625,
0.88330078125,
0.5390625,
0.88330078125,
0.552734375,
0.51318359375,
0.552734375
],
[
0.1485595703125,
0.54736328125,
0.4609375,
0.54736328125,
0.4609375,
0.55810546875,
0.1485595703125,
0.55810546875
],
[
0.14697265625,
0.56005859375,
0.459228515625,
0.56005859375,
0.459228515625,
0.57373046875,
0.14697265625,
0.57373046875
],
[
0.51123046875,
0.5546875,
0.88330078125,
0.5546875,
0.88330078125,
0.568359375,
0.51123046875,
0.568359375
],
[
0.14697265625,
0.57470703125,
0.459228515625,
0.57470703125,
0.459228515625,
0.58837890625,
0.14697265625,
0.58837890625
],
[
0.51318359375,
0.572265625,
0.884765625,
0.572265625,
0.884765625,
0.5859375,
0.51318359375,
0.5859375
],
[
0.145263671875,
0.5888671875,
0.4609375,
0.5888671875,
0.4609375,
0.6025390625,
0.145263671875,
0.6025390625
],
[
0.51318359375,
0.58837890625,
0.88818359375,
0.58837890625,
0.88818359375,
0.6015625,
0.51318359375,
0.6015625
],
[
0.145263671875,
0.60400390625,
0.4609375,
0.60400390625,
0.4609375,
0.61767578125,
0.145263671875,
0.61767578125
],
[
0.5146484375,
0.60400390625,
0.884765625,
0.60400390625,
0.884765625,
0.6181640625,
0.5146484375,
0.6181640625
],
[
0.1435546875,
0.6181640625,
0.185546875,
0.6162109375,
0.187255859375,
0.62890625,
0.145263671875,
0.63134765625
],
[
0.5146484375,
0.62060546875,
0.884765625,
0.62060546875,
0.884765625,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.51318359375,
0.63525390625,
0.88330078125,
0.63525390625,
0.88330078125,
0.6484375,
0.51318359375,
0.6484375
],
[
0.145263671875,
0.642578125,
0.26123046875,
0.642578125,
0.26123046875,
0.65673828125,
0.145263671875,
0.65673828125
],
[
0.5146484375,
0.65380859375,
0.88330078125,
0.65380859375,
0.88330078125,
0.66455078125,
0.5146484375,
0.66455078125
],
[
0.118408203125,
0.66748046875,
0.490234375,
0.66650390625,
0.490234375,
0.68115234375,
0.118408203125,
0.681640625
],
[
0.51318359375,
0.66845703125,
0.880859375,
0.66845703125,
0.880859375,
0.6787109375,
0.51318359375,
0.6787109375
],
[
0.118408203125,
0.68310546875,
0.490234375,
0.68310546875,
0.490234375,
0.69677734375,
0.118408203125,
0.69677734375
],
[
0.51123046875,
0.68310546875,
0.880859375,
0.68310546875,
0.880859375,
0.69677734375,
0.51123046875,
0.69677734375
],
[
0.118408203125,
0.70068359375,
0.489501953125,
0.70068359375,
0.489501953125,
0.71435546875,
0.118408203125,
0.71435546875
],
[
0.53076171875,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71630859375,
0.53076171875,
0.71630859375
],
[
0.118408203125,
0.71630859375,
0.491943359375,
0.71630859375,
0.491943359375,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.5146484375,
0.71875,
0.88330078125,
0.71875,
0.88330078125,
0.732421875,
0.5146484375,
0.732421875
],
[
0.118408203125,
0.73291015625,
0.491943359375,
0.73291015625,
0.491943359375,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.51318359375,
0.7353515625,
0.88330078125,
0.7353515625,
0.88330078125,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.12005615234375,
0.7490234375,
0.490234375,
0.7490234375,
0.490234375,
0.7626953125,
0.12005615234375,
0.7626953125
],
[
0.5146484375,
0.751953125,
0.884765625,
0.751953125,
0.884765625,
0.7626953125,
0.5146484375,
0.7626953125
],
[
0.12005615234375,
0.765625,
0.490234375,
0.765625,
0.490234375,
0.779296875,
0.12005615234375,
0.779296875
],
[
0.51318359375,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.1175537109375,
0.77978515625,
0.490234375,
0.77978515625,
0.490234375,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.51123046875,
0.7822265625,
0.88330078125,
0.783203125,
0.88330078125,
0.79638671875,
0.51123046875,
0.7958984375
],
[
0.118408203125,
0.79638671875,
0.49365234375,
0.79638671875,
0.49365234375,
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
],
[
0.51123046875,
0.81396484375,
0.884765625,
0.8134765625,
0.884765625,
0.82763671875,
0.51123046875,
0.8291015625
],
[
0.118408203125,
0.830078125,
0.490234375,
0.830078125,
0.490234375,
0.84326171875,
0.118408203125,
0.84326171875
],
[
0.51123046875,
0.83056640625,
0.88330078125,
0.830078125,
0.88330078125,
0.84423828125,
0.51123046875,
0.845703125
],
[
0.118408203125,
0.845703125,
0.490234375,
0.845703125,
0.490234375,
0.859375,
0.118408203125,
0.859375
],
[
0.51318359375,
0.84716796875,
0.884765625,
0.84716796875,
0.884765625,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.86083984375,
0.49365234375,
0.86083984375,
0.49365234375,
0.8759765625,
0.118408203125,
0.8759765625
],
[
0.51123046875,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51123046875,
0.876953125
],
[
0.139404296875,
0.8818359375,
0.495361328125,
0.8828125,
0.495361328125,
0.89794921875,
0.139404296875,
0.896484375
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87744140625,
0.88330078125,
0.892578125,
0.51318359375,
0.8935546875
],
[
0.118408203125,
0.896484375,
0.3896484375,
0.896484375,
0.3896484375,
0.91015625,
0.118408203125,
0.91015625
],
[
0.51318359375,
0.89453125,
0.88330078125,
0.89453125,
0.88330078125,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"Logic-Guided Data Augmentation and Regularization",
"for Consistent Question Answering",
"Akari Asait and Hannaneh Hajishirzitt",
"+University of Washington Allen Institutefor AI",
"[akari, hannaneh]@cs.washington.edu",
"Q: The ceramic vase was less flexible",
"more",
"than the plastic ball so it was",
"Abstract",
"breakable",
"Many natural language questions require qual-",
"A:more breakable",
"RoBERTa",
"2 ",
"itative, quantitative or logical comparisons be-",
"Q: The ceramic vase was more flexible",
"X",
"thantheplasticball soitwas",
"more",
"tween two entities or events. This paper ad-",
"A:less breakable",
"breakable",
"dressestheproblemofimprovingtheaccuracy",
"and consistency of responses to comparison",
"Q: Ifit is silent, does the outerear collect",
"lesssoundwaves",
"mor",
"questions by integrating logic rules and neu-",
"A: more [positive causal relationship]",
"ral models. Our method leverages logical and",
"Q: if theouter earcollect less sou",
"RoBERTa",
"mor",
"linguistic knowledge to augment labeled train",
"waves, is less sound being detected?",
"A:more [positive causal relationship]",
"ing data and then uses a consistency-based",
"Q:ifitisilent,isless soundbein",
"+",
"regularizer to train the model.",
"Improving",
"detected",
"ve causal relationship]",
"",
"A:more lpos",
"0c",
"the global consistency of predictions, our ap",
"['s]",
"proach achieves large improvements over pre-",
"Figure I: Inconsistent predictions by RoBERTa. Top",
"vious methods inavarietyof question answer-",
"row shows an example of symmetric inconsistency and",
"ing (QA) tasks including multiple-choice qual-",
"the second row shows an example of transitive incon-",
"itative reasoning,cause-effect reasoning,and",
" sistency. The examples are partially modified.",
"extractive machine reading comprehension. In",
"AS1t002A1",
"particular,ourmethod significantlyimproves",
"the performance of RoBERTa-based models",
"et al., 2016; Joshi et al., 2017). This poses new",
"by1-5%acrossdatasets.Weadvancestateof",
"the art by around 5-8% on WIQA and QuaRel",
"challenges to standard models, which are known",
"and reduce consistency violations by 58% on",
"to exploit statistical patterns or annotation artifacts",
"HotpotQA. We further demonstrate that our",
"in these datasets (Sugawara et al., 2018; Min et al.,",
"approach can learn effectively from limited",
"2019a). Importantly, state-of-the-art models show",
"data.!",
"inconsistent comparison predictions as shown in",
"Figure 1.. Improving the consistency of predictions",
"Introduction",
"hasbeenpreviouslystudiedinnatural languagein-",
"Comparison-type questions (Tandon et al., 2019;",
"ference(NLD) tasks (Minerviniand Riedel. 2018",
"Tafjord et al.. 2019: Yang et al., 2018) ask about re-",
"Li et al., 2019), but has not been addressed in QA",
"lationships between properties of entities or events",
"In this paper, we address the task of produc-",
"such as cause-effect, qualitative or quantitative rea-",
"ing globally consistent and accurate predictions for",
"soning. To create comparison questions that re-",
"comparison questions leveraging logical and sym-",
"quire inferential knowledge and reasoning ability,",
"bolicknowledgefordataaugmentationandtrain-",
"annotators need to understand context presented",
"ing regularization. Our data augmentation uses a",
"in multiple paragraphs or carefully ground a ques-",
"set of logical and linguistic knowledge to develop",
"tion to the given situation. This makes it chal-",
"additional consistent labeled training data. Sub-",
"lenging to annotate a large number of comparison",
"sequently, our method uses symbolic logic to in-",
"questions. Most current datasets on comparison",
"corporate consistency regularization for additional",
"questions are much smaller than standard machine",
"supervision signal beyond inductive bias given",
"reading comprehension (MRC) datasets (Rajpurkar",
"by data augmentation. Our method generalizes",
"'Our code and data is available at https : / /github.",
"previous consistency-promoting methods for NLI",
"com/AkariAsai/logic_guided_qa.",
"tasks (Minervini and Riedel, 2018; Li et al., 2019)"
] |
{
"arxiv_id": "2004.10157",
"language": "en",
"page_idx": 0
}
|
||
2004.10157-en-1
|
2004.10157-en-1.jpg
|
to adapt to substantially different question formats.
Our experiments show significant improvement over the state of the art on a variety of QA tasks: a classification-based causal reasoning QA, a multiple choice QA for qualitative reasoning and an extractive MRC task with comparisons between entities. Notably, our data augmentation and consistency constrained training regularization improves performance of RoBERTa-based models (Liu et al., 2019) by 1.0%, 5.0% and 2.5% on WIQA, QuaRel and HotpotQA. Our approach advances the stateof-the-art results on WIQA and QuaRel with 4.7 and 8.4% absolute accuracy improvement, respectively, reducing inconsistent predictions. We further demonstrate that our approach can learn effectively from limited labeled data: given only 20% of the original labeled data, our method achieves performance on par with a competitive baseline learned with the full labeled data.
## 2 $\textbf{Related Work}$
Data augmentation has been explored in a variety of tasks and domains (Krizhevsky et al., 2009; Cubuk et al., 2019; Park et al., 2019). In NLP, using backtranslation (Yu et al., 2018) or dictionary based word replacement (Zhang et al., 2015) has been studied. Most relevant to our work, Kang et al. (2018) study NLI-specific logic and knowledgebased data augmentation. Concurrent to our work, Gokhale et al. (2020) study visual QA models' abilmentation. Our data augmentation does not rely on task-specific assumptions, and can be adapted to different formats of QA task. We further leverage consistency-promoting regularization, which gives improvements in accuracy and consistency.
Improving prediction consistency via training regularization has been studied in NLI tasks. Minervini and Riedel (2018) present model-dependent first-order logic guided adversarial example generation and regularization. Li et al. (2019) introduce consistency-based regularization incorporating the first-order logic rules. Previous approach is modeldependent or relies on NLI-specific rules, while our method is model-agnostic and is more generally applicable by combining it with data augmentation. Regularizing loss to penalize violations of structural constraints in models’ output has been also studied in previous work on constraint satisfaction in structured learning (Lee et al., 2019; Ganchev
et al., 2010). Our work regularizes models to produce globally consistent predictions among augmented data following logical constraints, while those studies incorporates structured prediction models following linguistics rules.
### 3 $\textbf{Method}$
We present the components of our QA method: frst-order logic guided data augmentation (Section 3.1 and Section 3.2), and consistency-based regularization (Section 3.3).
# 3.1 Consistent Question Answering
For globally consistent predictions in QA, we require responses to follow two important general logical rules: symmetric consistency and transitive $consistency$, which are illustrated in Figure 1 and are formally described below.
Let $q,p,a$ be a question, a paragraph and an answer predicted by a model. A is a set of answer candidates. Each element of A can be a span in $p$, a class category, or an arbitrary answer choice. $X=\{q,p,a\}$ represents a logic atom.
Symmetric consistency In a comparison question, small surface variations such as replacing words with their antonyms can reverse the answer, while keeping the overall semantics of the question as before. We define symmetry of questions in the context of QA as follows: $(q,p,a^*)\leftrightarrow$ $(q_{sym},p,a_{sym}^{*})$,where $q$ and $q_sym$ are antonyms of
ity to answer logically composed questions, and each other, and $a_sym^*$ is the opposite of the groundshow the effectiveness of logic-guided data aug- truth answer $a^*$ in A. For example, the two ques-
tions in the first row of Figure l are symmetric pairs. We define the symmetric consistency of predictions in QA as the following logic rule:
(1)
$$(q,p,a)\to(q_{sym},p,a_{sym}),$$
which indicates a system should predict $a_sym$ given
$(q_{sym},p)$,if it predicts $a$ for $(q,p).$
Transitive consistency. Transitive inference between three predicates $A,B,C$ is represented as: $A\to B\land B\to C$ then $A\to C$ (Gazes et al., 2012). In the context of QA, the transitive examples are mainly for causal reasoning questions that inquire about the effect $e$ given the cause $c.$ The second row of Figure 1 shows an example where transitive consistency is violated. For two questions $q_1$ and $q_2$ in which the effect of $q_1(=e_1)$ is equal to the cause of $q_2(=c_2)$, we define the transitive consistency of predictions as follows:
$(q_1,p,a_1)\wedge(q_2,p,a_2)\to(q_{trans},p,a_{trans}).$
|
[
[
0.118408203125,
0.07659912109375,
0.49365234375,
0.07659912109375,
0.49365234375,
0.09027099609375,
0.118408203125,
0.09027099609375
],
[
0.51123046875,
0.0748291015625,
0.88330078125,
0.07598876953125,
0.88330078125,
0.09027099609375,
0.51123046875,
0.08905029296875
],
[
0.1368408203125,
0.0926513671875,
0.489501953125,
0.0926513671875,
0.489501953125,
0.105712890625,
0.1368408203125,
0.105712890625
],
[
0.51123046875,
0.0914306640625,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.51123046875,
0.105712890625
],
[
0.118408203125,
0.1092529296875,
0.491943359375,
0.1092529296875,
0.491943359375,
0.12353515625,
0.118408203125,
0.12353515625
],
[
0.51123046875,
0.10687255859375,
0.884765625,
0.10809326171875,
0.884765625,
0.122314453125,
0.51123046875,
0.1217041015625
],
[
0.118408203125,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.51123046875,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1383056640625,
0.51123046875,
0.1383056640625
],
[
0.12005615234375,
0.1436767578125,
0.489501953125,
0.1436767578125,
0.489501953125,
0.15380859375,
0.12005615234375,
0.15380859375
],
[
0.51318359375,
0.141357421875,
0.77099609375,
0.141357421875,
0.77099609375,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.118408203125,
0.15673828125,
0.489501953125,
0.15673828125,
0.489501953125,
0.17041015625,
0.118408203125,
0.17041015625
],
[
0.515625,
0.17041015625,
0.52880859375,
0.17041015625,
0.52880859375,
0.1787109375,
0.515625,
0.1787109375
],
[
0.53662109375,
0.16748046875,
0.615234375,
0.16748046875,
0.615234375,
0.1817626953125,
0.53662109375,
0.1817626953125
],
[
0.12005615234375,
0.17333984375,
0.491943359375,
0.17333984375,
0.491943359375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.118408203125,
0.190673828125,
0.491943359375,
0.190673828125,
0.491943359375,
0.2037353515625,
0.118408203125,
0.2037353515625
],
[
0.51318359375,
0.1923828125,
0.884765625,
0.1923828125,
0.884765625,
0.2060546875,
0.51318359375,
0.2060546875
],
[
0.12005615234375,
0.2060546875,
0.49365234375,
0.2060546875,
0.49365234375,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.51318359375,
0.208984375,
0.88330078125,
0.208984375,
0.88330078125,
0.22265625,
0.51318359375,
0.22265625
],
[
0.12005615234375,
0.221435546875,
0.490234375,
0.221435546875,
0.490234375,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.51318359375,
0.2244873046875,
0.88330078125,
0.2244873046875,
0.88330078125,
0.2381591796875,
0.51318359375,
0.2381591796875
],
[
0.118408203125,
0.2381591796875,
0.491943359375,
0.2381591796875,
0.491943359375,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.51318359375,
0.2410888671875,
0.72314453125,
0.2410888671875,
0.72314453125,
0.254638671875,
0.51318359375,
0.254638671875
],
[
0.118408203125,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.118408203125,
0.270751953125,
0.491943359375,
0.270751953125,
0.491943359375,
0.283935546875,
0.118408203125,
0.283935546875
],
[
0.51123046875,
0.2666015625,
0.8037109375,
0.2666015625,
0.8037109375,
0.27978515625,
0.51123046875,
0.27978515625
],
[
0.118408203125,
0.2861328125,
0.49365234375,
0.28515625,
0.49365234375,
0.2998046875,
0.118408203125,
0.300537109375
],
[
0.51123046875,
0.2861328125,
0.88330078125,
0.2861328125,
0.88330078125,
0.300537109375,
0.51123046875,
0.300537109375
],
[
0.12005615234375,
0.302734375,
0.491943359375,
0.302734375,
0.491943359375,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.51318359375,
0.303955078125,
0.884765625,
0.303955078125,
0.884765625,
0.318359375,
0.51318359375,
0.318359375
],
[
0.118408203125,
0.318359375,
0.490234375,
0.318359375,
0.490234375,
0.333251953125,
0.118408203125,
0.333251953125
],
[
0.51318359375,
0.320556640625,
0.88330078125,
0.320556640625,
0.88330078125,
0.334228515625,
0.51318359375,
0.334228515625
],
[
0.1175537109375,
0.334228515625,
0.48779296875,
0.334228515625,
0.48779296875,
0.347412109375,
0.1175537109375,
0.347412109375
],
[
0.51318359375,
0.336181640625,
0.884765625,
0.336181640625,
0.884765625,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.118408203125,
0.350830078125,
0.490234375,
0.349853515625,
0.490234375,
0.364013671875,
0.118408203125,
0.365234375
],
[
0.5146484375,
0.3515625,
0.73779296875,
0.3515625,
0.73779296875,
0.365234375,
0.5146484375,
0.365234375
],
[
0.1175537109375,
0.365234375,
0.366943359375,
0.366455078125,
0.366943359375,
0.38134765625,
0.1175537109375,
0.380126953125
],
[
0.52880859375,
0.3681640625,
0.88330078125,
0.366455078125,
0.88330078125,
0.38134765625,
0.52880859375,
0.383056640625
],
[
0.5146484375,
0.385498046875,
0.88330078125,
0.385498046875,
0.88330078125,
0.3984375,
0.5146484375,
0.3984375
],
[
0.1175537109375,
0.39794921875,
0.27197265625,
0.39794921875,
0.27197265625,
0.412109375,
0.1175537109375,
0.412109375
],
[
0.51318359375,
0.40087890625,
0.8818359375,
0.40087890625,
0.8818359375,
0.41455078125,
0.51318359375,
0.41455078125
],
[
0.51318359375,
0.416259765625,
0.88330078125,
0.4150390625,
0.88330078125,
0.429931640625,
0.51318359375,
0.43115234375
],
[
0.118408203125,
0.42578125,
0.490234375,
0.424560546875,
0.490234375,
0.439453125,
0.118408203125,
0.440673828125
],
[
0.51318359375,
0.432861328125,
0.79833984375,
0.431640625,
0.79833984375,
0.446533203125,
0.51318359375,
0.44775390625
],
[
0.12005615234375,
0.4423828125,
0.489501953125,
0.4423828125,
0.489501953125,
0.4560546875,
0.12005615234375,
0.4560546875
],
[
0.12005615234375,
0.457763671875,
0.490234375,
0.457763671875,
0.490234375,
0.471435546875,
0.12005615234375,
0.471435546875
],
[
0.5146484375,
0.457275390625,
0.88330078125,
0.457275390625,
0.88330078125,
0.47021484375,
0.5146484375,
0.47021484375
],
[
0.118408203125,
0.474365234375,
0.490234375,
0.474365234375,
0.490234375,
0.488037109375,
0.118408203125,
0.488037109375
],
[
0.51123046875,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.51123046875,
0.48681640625
],
[
0.118408203125,
0.490478515625,
0.490234375,
0.490478515625,
0.490234375,
0.50439453125,
0.118408203125,
0.50439453125
],
[
0.51123046875,
0.488037109375,
0.88671875,
0.4892578125,
0.88671875,
0.50439453125,
0.51123046875,
0.5029296875
],
[
0.12005615234375,
0.505859375,
0.490234375,
0.505859375,
0.490234375,
0.51953125,
0.12005615234375,
0.51953125
],
[
0.509765625,
0.50439453125,
0.884765625,
0.5048828125,
0.884765625,
0.51953125,
0.509765625,
0.5185546875
],
[
0.118408203125,
0.5224609375,
0.491943359375,
0.5224609375,
0.491943359375,
0.5361328125,
0.118408203125,
0.5361328125
],
[
0.51123046875,
0.521484375,
0.884765625,
0.521484375,
0.884765625,
0.53515625,
0.51123046875,
0.53515625
],
[
0.1175537109375,
0.5380859375,
0.49365234375,
0.53759765625,
0.49365234375,
0.5517578125,
0.1175537109375,
0.552734375
],
[
0.51123046875,
0.5361328125,
0.88330078125,
0.53759765625,
0.88330078125,
0.5517578125,
0.51123046875,
0.55029296875
],
[
0.12005615234375,
0.5546875,
0.491943359375,
0.5546875,
0.491943359375,
0.568359375,
0.12005615234375,
0.568359375
],
[
0.5087890625,
0.552734375,
0.88671875,
0.55029296875,
0.88671875,
0.568359375,
0.5087890625,
0.57080078125
],
[
0.118408203125,
0.57177734375,
0.490234375,
0.57177734375,
0.490234375,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.51123046875,
0.568359375,
0.88330078125,
0.568359375,
0.88330078125,
0.5830078125,
0.51123046875,
0.5830078125
],
[
0.1175537109375,
0.5859375,
0.49365234375,
0.58740234375,
0.49365234375,
0.6015625,
0.1175537109375,
0.60107421875
],
[
0.51318359375,
0.5859375,
0.884765625,
0.5859375,
0.884765625,
0.599609375,
0.51318359375,
0.599609375
],
[
0.12005615234375,
0.60400390625,
0.490234375,
0.60400390625,
0.490234375,
0.61767578125,
0.12005615234375,
0.61767578125
],
[
0.5146484375,
0.6015625,
0.884765625,
0.6015625,
0.884765625,
0.615234375,
0.5146484375,
0.615234375
],
[
0.118408203125,
0.6181640625,
0.489501953125,
0.6181640625,
0.489501953125,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51318359375,
0.6181640625,
0.884765625,
0.6181640625,
0.884765625,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.12005615234375,
0.63525390625,
0.489501953125,
0.63525390625,
0.489501953125,
0.6484375,
0.12005615234375,
0.6484375
],
[
0.51318359375,
0.63427734375,
0.765625,
0.63427734375,
0.765625,
0.6484375,
0.51318359375,
0.6484375
],
[
0.12005615234375,
0.65185546875,
0.489501953125,
0.65185546875,
0.489501953125,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.59619140625,
0.65478515625,
0.8037109375,
0.65478515625,
0.8037109375,
0.66943359375,
0.59619140625,
0.66943359375
],
[
0.8564453125,
0.65283203125,
0.88671875,
0.65283203125,
0.88671875,
0.66943359375,
0.8564453125,
0.66943359375
],
[
0.118408203125,
0.66845703125,
0.4375,
0.66845703125,
0.4375,
0.681640625,
0.118408203125,
0.681640625
],
[
0.51318359375,
0.67333984375,
0.88330078125,
0.67578125,
0.88330078125,
0.6923828125,
0.51318359375,
0.6904296875
],
[
0.1376953125,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.1376953125,
0.69873046875
],
[
0.51318359375,
0.693359375,
0.76904296875,
0.693359375,
0.76904296875,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.118408203125,
0.70068359375,
0.489501953125,
0.70068359375,
0.489501953125,
0.71435546875,
0.118408203125,
0.71435546875
],
[
0.12005615234375,
0.71533203125,
0.489501953125,
0.71533203125,
0.489501953125,
0.728515625,
0.12005615234375,
0.728515625
],
[
0.51318359375,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.51318359375,
0.72998046875
],
[
0.12005615234375,
0.73291015625,
0.491943359375,
0.73291015625,
0.491943359375,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.515625,
0.7353515625,
0.88330078125,
0.7353515625,
0.88330078125,
0.7451171875,
0.515625,
0.7451171875
],
[
0.12005615234375,
0.7490234375,
0.489501953125,
0.7490234375,
0.489501953125,
0.7626953125,
0.12005615234375,
0.7626953125
],
[
0.51708984375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76123046875,
0.51708984375,
0.76123046875
],
[
0.118408203125,
0.765625,
0.490234375,
0.765625,
0.490234375,
0.779296875,
0.118408203125,
0.779296875
],
[
0.5146484375,
0.765625,
0.884765625,
0.765625,
0.884765625,
0.779296875,
0.5146484375,
0.779296875
],
[
0.118408203125,
0.7822265625,
0.49365234375,
0.7822265625,
0.49365234375,
0.7958984375,
0.118408203125,
0.7958984375
],
[
0.51318359375,
0.7822265625,
0.884765625,
0.7822265625,
0.884765625,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.79736328125,
0.49365234375,
0.79736328125,
0.49365234375,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.51318359375,
0.79736328125,
0.88330078125,
0.79736328125,
0.88330078125,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.118408203125,
0.81396484375,
0.490234375,
0.81396484375,
0.490234375,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.51123046875,
0.8134765625,
0.88330078125,
0.8134765625,
0.88330078125,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.118408203125,
0.830078125,
0.483642578125,
0.830078125,
0.483642578125,
0.84326171875,
0.118408203125,
0.84326171875
],
[
0.51318359375,
0.830078125,
0.884765625,
0.830078125,
0.884765625,
0.84326171875,
0.51318359375,
0.84326171875
],
[
0.1368408203125,
0.8466796875,
0.49365234375,
0.8466796875,
0.49365234375,
0.86083984375,
0.1368408203125,
0.86083984375
],
[
0.51318359375,
0.845703125,
0.88671875,
0.845703125,
0.88671875,
0.85986328125,
0.51318359375,
0.85986328125
],
[
0.118408203125,
0.86328125,
0.49365234375,
0.86328125,
0.49365234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.8623046875,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87890625,
0.491943359375,
0.87890625,
0.491943359375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.51318359375,
0.87890625,
0.86572265625,
0.87890625,
0.86572265625,
0.892578125,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.52294921875,
0.89794921875,
0.8818359375,
0.896484375,
0.8818359375,
0.9111328125,
0.52294921875,
0.912109375
]
] |
[
"to adapt to substantially different question formats.",
"et al., 2010). Our work regularizes models to pro-",
"Our experiments show significant improvement",
"duce globally consistent predictions among aug",
"over the state of the art on a variety of QA tasks:",
"mented data following logical constraints, while",
"a classification-based causal reasoning QA, a mul-",
"those studies incorporates structured prediction",
"tiple choice OAforqualitativereasoningandan",
"models following linguistics rules.",
"extractive MRC task with comparisons between en-",
"3",
"Method",
"tities. Notably, our data augmentation and consis-",
"tency constrained training regularization improves",
"We present the components of our QA method:",
"performance of RoBERTa-based models (Liu et al.,",
"first-order logic guided data augmentation (Sec-",
"2019) by 1.0%, 5.0% and 2.5% on WIQA, QuaRel",
"tion 3.1 and Section 3.2), and consistency-based",
"and HotpotQA. Our approach advances the state-",
"regularization (Section 3.3).",
"of-the-art results on WIQA and QuaRel with 4.7",
"and 8.4%absolute accuracyimprovement,respec-",
"3.1 Consistent Question Answering",
"tively, reducing inconsistent predictions. We fur-",
"For globally consistent predictions in QA, we re-",
"ther demonstrate that our approach can learn effec-",
"quire responses to follow two important general",
"tively from limited labeled data: given only 20%",
"logical rules: symmetric consistency and transitive",
"of the original labeled data, our method achieves",
"consistency, which are illustrated in Figure 1 and",
"performance on par with a competitive baseline",
"areformally described below.",
"learned with the full labeled data.",
"Let q, p, a be a question, a paragraph and an",
"answer predicted by a model. A is a set of answer",
"2Related Work",
"candidates. Each element of A can be a span in",
"p, a class category, or an arbitrary answer choice.",
"Data augmentation has been explored in a variety of",
"X = (q, p, a} represents a logic atom.",
"tasks and domains (Krizhevsky et al., 2009: Cubuk",
"et al., 2019; Park et al., 2019). In NLP, using back-",
"Symmetric consistencyIn a comparison ques-",
"translation (Yu et al., 2018) or dictionary based",
"tion, small surface variations such as replacing",
"word replacement (Zhang et al., 2015) has been",
"words with their antonyms can reverse the answer,",
"studied. Most relevant to our work, Kang et al.",
"while keeping the overall semantics of the ques-",
"(2018) study NLI-specific logic and knowledge-",
"tion as before. We define symmetry of questions",
"based data augmentation. Concurrent to our work,",
"in the context of QA as follows:(q,P,a*))",
"Gokhale et al. (2020) study visual QA models' abil-",
"(qsym, P, asym), where q and qsym are antonyms of",
"ity to answer logically composed questions, and",
" each other, and a*m is the opposite of the ground-",
"show the effectiveness of logic-guided data aug-",
"truth answer a* in A. For example, the two ques-",
"mentation. Our data augmentation does not rely on",
"tions in the first row of Figure I are symmetric pairs.",
"task-specific assumptions, and can be adapted to",
"We define the symmetric consistency of predictions",
"different formats of QA task. We further leverage",
"in QA as the following logic rule:",
"consistency-promoting regularization, which gives",
"(q,p, a) → (qsym,p, asym),",
"(1)",
"improvements in accuracy and consistency.",
"which indicates a system should predict asym given",
"Improving prediction consistency via training",
"(qsym, p), if it predicts a for (q, p).",
"regularization has been studied in NLI tasks. Min-",
"ervini and Riedel (2018) present model-dependent",
"Transitive consistency.Transitive inference be",
"first-order logic guided adversarial example gener-",
"tween three predicates A.B.C is represented as:",
"ation and regularization. Li et al. (2019) introduce",
"A → B Λ B → C then A →C (Gazes et al.",
"consistency-based regularization incorporating the",
"2012). In the context of QA, the transitive exam-",
"first-order logic rules. Previous approach is model-",
"ples are mainly for causal reasoning questions that",
"dependent or relies on NLI-specific rules, while our",
"inquire about the effect e given the cause c. The",
"method is model-agnostic and is more generally ap-",
" second row of Figure 1 shows an example where",
"plicable by combining it with data augmentation.",
"transitive consistency is violated. For two ques-",
"Regularizing loss to penalize violations of struc-",
"tions qi and q2 in which the effect of qi (= ei)",
"tural constraints in models' output has been also",
"is equal to the cause of q2 (= C2), we define the",
"studied in previous work on constraint satisfaction",
"transitive consistency of predictions as follows:",
"in structured learning (Lee et al., 2019; Ganchev",
"(q1,P,a1)^(q2, P,a2) → (qtrans, P, atrans). (2)"
] |
{
"arxiv_id": "2004.10157",
"language": "en",
"page_idx": 1
}
|
||
2004.10157-en-2
|
2004.10157-en-2.jpg
|
<table>
<tbody>
<tr>
<th>reasoning format</th>
<th>$WIQA$ (Tandon et al., 2019) $\bar{\text{Causal }}\bar{\text{Reasoning}}$ classification</th>
<th>QuaRel (Tafjord et al., 2019) $\bar{\text{Qualitative Reasoning}}$ multiple ${\mathrm{choice}}$</th>
<th>HotpotQA (Yang et al., 2018) $\bar{\text{Quali}}\bar{\text{tative Comparison of entities}}$ span extraction</th>
</tr>
<tr>
<td>$p$</td>
<td>The rain seeps into the wood surface. When rain evaporates it leaves the wood. It takes the finish of the wood with it. The wood begins to lose it's luster.</td>
<td>Supposed you $\textbf{J Were stand- }$ ing on the planet Earth and Mercury. When you look up in the sky and see the sun,</td>
<td>Golf Magazine is a monthly golf magazine owned by Time Inc. El Nuevo Cojo Ilustrado is an Ameri- can Spanish language magazine.</td>
</tr>
<tr>
<td>$q$</td>
<td>$q_{1}{:}$If a tsunami happens, will wood be more moist?, $q_{2}\colon$ If wood is more moist. is more weath ering occurring?</td>
<td>Which planet would the sun $\operatorname{appear}$ larger?</td>
<td>El Nuevo Cojo and Golf Magazine which one is owned by Time Inc?</td>
</tr>
<tr>
<td>$\overline{\mathbf{A}}$</td>
<td>( more. less. no effects</td>
<td>$\{$Mercury, Earth$\}$</td>
<td>$\overline{\{\mathrm{Golf~Magazine,~El~Nuevo~Cojo}\}}$</td>
</tr>
<tr>
<td>不 $a$</td>
<td>$a_{1}^{*}:$more, $a_{9}^{*}:$more</td>
<td>$\mathbf{Mercury}$</td>
<td>Golf Magazine</td>
</tr>
<tr>
<td>$q_{aug}$</td>
<td>If a tsunami happens, is more weather- $\operatorname{ing}$occurring?</td>
<td>Which planet would the sun $\operatorname{appear}$ smaller?</td>
<td>Which one is not owned by Time Inc, Golf Magazine e El Nuevo Cojo?</td>
</tr>
<tr>
<td>$a_{aug}$</td>
<td>more</td>
<td>Earth</td>
<td>El Nuevo Cojo</td>
</tr>
</tbody>
</table>
Table 1: An augmented transitive example for WIQA, and symmetric examples for QuaRel and HotpotQA. We partially modify paragraphs and questions. The bold characters denote a shared event connecting two questions The parts written in red or blue denote antonyms, and highlighted text is negation added by our data augmentation.
### 3.2 $\textbf{Logic- guided Data Augmentation}$
$(c_1,e_1)$ and $(c_2,e_2)$,where $e_1=c_2$ holds. When $a_1^*$ is a positive causal relationship, we create a new $\mathop{\text{example }X}_{trans}=(q_{3},p,a_{2}^{*})$for$q_3=(c_1,e_2).$
Given a set of training examples $X$ in the form of $(q,p,a^*)$,we automatically generate additional examples $X_{aug}=\{q_{aug},p,a_{aug}^*\}$ using symme- Sampling augmented data Adding all consis-
try and transitivity logical rules. The goal is to augment the training data so that symmetric and transitive examples are observed during training. We provide some augmented examples in Table 1.
tent examples may change the data distribution from the original one, which may lead to a deterioration in performance (Xie et al., 2019). One can select the data based on a model's prediction inconsistencies (Minervini and Riedel, 2018) or randomly sample at each epoch (Kang et al., 2018). In this work, we randomly sample augmented data at the beginning of training, and use the same examples for all epochs during training. Despite its simplicity, this yields competitive or even better performance than other sampling strategies.$^{3}$
3.3 Logic-guided Consistency Regularization
Augmenting symmetric examples To create a symmetric question, we convert a question into an opposite one using the following operations: (a) replace words with their antonyms, (b) add, or (c) remove words. For (a), we select top frequent adjectives or verbs with polarity (e.g., smaller, increases) from training corpora, and expert annotators write antonyms for each of the frequent words (we denote this small dictionary as D). More details can be seen in Appendix A. For (b) and (c), we add negation words or remove negation words (e.g., not). For all of the questions in training data, if a question includes a word in D for the operation (a), or matches a template (e.g., which $\star$ is$\leftrightarrow$which $\star$ is not)for operations(b)and (c), we apply the operation to generate $q_sym^{2}$ We dictions. The second term $\mathcal{L}_cons^{4}$ penalizes making(c),we apply the operation to generate $q_sym^{2}$ We nredictions. that violate svmmetric and transitive
obtain $a_sym^*$ by re-labeling the answer $a^*$ to its op-
posite answer choice in A (see Appendix B).
We regularize the learning objective (task loss, $\mathcal{L}_{task})$ with a regularization term that promotes consistency of predictions (consistency loss, $\mathcal{L}_cons).$
$$\mathcal{L}=\mathcal{L}_{task}(X)+\mathcal{L}_{cons}(X,X_{aug}).$$
(3)
The first term $\mathcal{L}_{task}$ penalizes making incorrect prepredictions that violate symmetric and transitive logical rules as follows:
$$\mathcal{L}_{cons}=\lambda_{sym}\mathcal{L}_{sym}+\lambda_{trans}\mathcal{L}_{trans},$$
(4)
Augmenting transitive examples We first find a pair of two cause-effect questions $X_1=(q_1,p,a_1^*)$ and $X_2=(q_2,p,a_2^*)$, whose $q_1$ and $q_2$ consist of
where $\lambda_sym$ and $\lambda_trans$ are weighting scalars to balance the two consistency-promoting objectives.
We do not add $X_{aug}$ if the same pair has already exist $^{4}$We mask the $\mathcal{L}_cons$ for the examples without symmetric
or transitive consistent examples.
$^{2}$We observe that (b)(c) are less effective than (a) in WIQA or QuaRel, while especially (b) contributes to the performance improvements on HotpotQA as much as (a) does.
|
[
[
0.307373046875,
0.07781982421875,
0.35107421875,
0.07781982421875,
0.35107421875,
0.08905029296875,
0.307373046875,
0.08905029296875
],
[
0.5283203125,
0.07659912109375,
0.57666015625,
0.0748291015625,
0.57861328125,
0.0872802734375,
0.52880859375,
0.08905029296875
],
[
0.73046875,
0.07781982421875,
0.796875,
0.07781982421875,
0.796875,
0.08905029296875,
0.73046875,
0.08905029296875
],
[
0.264404296875,
0.0885009765625,
0.3955078125,
0.0885009765625,
0.3955078125,
0.1015625,
0.264404296875,
0.1015625
],
[
0.48779296875,
0.0885009765625,
0.6171875,
0.0885009765625,
0.6171875,
0.1015625,
0.48779296875,
0.1015625
],
[
0.7060546875,
0.0885009765625,
0.8193359375,
0.0885009765625,
0.8193359375,
0.09979248046875,
0.7060546875,
0.09979248046875
],
[
0.27197265625,
0.09857177734375,
0.387939453125,
0.1015625,
0.387939453125,
0.11517333984375,
0.27197265625,
0.11224365234375
],
[
0.48193359375,
0.09979248046875,
0.623046875,
0.1009521484375,
0.623046875,
0.11517333984375,
0.48193359375,
0.114013671875
],
[
0.6572265625,
0.1009521484375,
0.873046875,
0.1009521484375,
0.873046875,
0.114013671875,
0.6572265625,
0.114013671875
],
[
0.1376953125,
0.114013671875,
0.183837890625,
0.114013671875,
0.183837890625,
0.1258544921875,
0.1376953125,
0.1258544921875
],
[
0.1259765625,
0.10272216796875,
0.19140625,
0.10272216796875,
0.19140625,
0.1134033203125,
0.1259765625,
0.1134033203125
],
[
0.2880859375,
0.114013671875,
0.3720703125,
0.114013671875,
0.3720703125,
0.1258544921875,
0.2880859375,
0.1258544921875
],
[
0.50390625,
0.114013671875,
0.60107421875,
0.114013671875,
0.60107421875,
0.1258544921875,
0.50390625,
0.1258544921875
],
[
0.71533203125,
0.114013671875,
0.81201171875,
0.114013671875,
0.81201171875,
0.1258544921875,
0.71533203125,
0.1258544921875
],
[
0.204833984375,
0.1312255859375,
0.455078125,
0.1312255859375,
0.455078125,
0.144287109375,
0.204833984375,
0.144287109375
],
[
0.465087890625,
0.1318359375,
0.640625,
0.1318359375,
0.640625,
0.1455078125,
0.465087890625,
0.1455078125
],
[
0.64990234375,
0.1312255859375,
0.8759765625,
0.1312255859375,
0.8759765625,
0.144287109375,
0.64990234375,
0.144287109375
],
[
0.20654296875,
0.1436767578125,
0.45166015625,
0.1436767578125,
0.45166015625,
0.15673828125,
0.20654296875,
0.15673828125
],
[
0.465087890625,
0.1436767578125,
0.6396484375,
0.1436767578125,
0.6396484375,
0.15673828125,
0.465087890625,
0.15673828125
],
[
0.6533203125,
0.1436767578125,
0.875,
0.1436767578125,
0.875,
0.15673828125,
0.6533203125,
0.15673828125
],
[
0.160400390625,
0.152587890625,
0.16796875,
0.157958984375,
0.160400390625,
0.163330078125,
0.15283203125,
0.157958984375
],
[
0.208251953125,
0.155029296875,
0.453369140625,
0.155029296875,
0.453369140625,
0.168701171875,
0.208251953125,
0.168701171875
],
[
0.465087890625,
0.155029296875,
0.640625,
0.155029296875,
0.640625,
0.168701171875,
0.465087890625,
0.168701171875
],
[
0.6533203125,
0.155029296875,
0.8759765625,
0.155029296875,
0.8759765625,
0.168701171875,
0.6533203125,
0.168701171875
],
[
0.465087890625,
0.16748046875,
0.62744140625,
0.16748046875,
0.62744140625,
0.18115234375,
0.465087890625,
0.18115234375
],
[
0.654296875,
0.16748046875,
0.8505859375,
0.16748046875,
0.8505859375,
0.18115234375,
0.654296875,
0.18115234375
],
[
0.204833984375,
0.1787109375,
0.450927734375,
0.1787109375,
0.450927734375,
0.1923828125,
0.204833984375,
0.1923828125
],
[
0.208251953125,
0.168701171875,
0.4189453125,
0.168701171875,
0.4189453125,
0.1787109375,
0.208251953125,
0.1787109375
],
[
0.465087890625,
0.185302734375,
0.6396484375,
0.185302734375,
0.6396484375,
0.198974609375,
0.465087890625,
0.198974609375
],
[
0.6513671875,
0.182861328125,
0.875,
0.185302734375,
0.875,
0.1995849609375,
0.6513671875,
0.19775390625
],
[
0.20654296875,
0.191162109375,
0.3359375,
0.191162109375,
0.3359375,
0.204833984375,
0.20654296875,
0.204833984375
],
[
0.33154296875,
0.1923828125,
0.44921875,
0.1923828125,
0.44921875,
0.203125,
0.33154296875,
0.203125
],
[
0.466796875,
0.198974609375,
0.5615234375,
0.198974609375,
0.5615234375,
0.210205078125,
0.466796875,
0.210205078125
],
[
0.65576171875,
0.198974609375,
0.861328125,
0.198974609375,
0.861328125,
0.208984375,
0.65576171875,
0.208984375
],
[
0.20654296875,
0.204833984375,
0.4375,
0.204833984375,
0.4375,
0.215576171875,
0.20654296875,
0.215576171875
],
[
0.156982421875,
0.196533203125,
0.16455078125,
0.196533203125,
0.16455078125,
0.203125,
0.156982421875,
0.203125
],
[
0.49951171875,
0.2161865234375,
0.60205078125,
0.2161865234375,
0.60205078125,
0.22802734375,
0.49951171875,
0.22802734375
],
[
0.6630859375,
0.215576171875,
0.869140625,
0.215576171875,
0.869140625,
0.2286376953125,
0.6630859375,
0.2286376953125
],
[
0.26025390625,
0.21728515625,
0.39794921875,
0.21728515625,
0.39794921875,
0.22802734375,
0.26025390625,
0.22802734375
],
[
0.15283203125,
0.218505859375,
0.166259765625,
0.218505859375,
0.166259765625,
0.226806640625,
0.15283203125,
0.226806640625
],
[
0.5263671875,
0.2286376953125,
0.580078125,
0.2286376953125,
0.580078125,
0.240478515625,
0.5263671875,
0.240478515625
],
[
0.71533203125,
0.22802734375,
0.81201171875,
0.2286376953125,
0.81201171875,
0.240478515625,
0.71533203125,
0.2392578125
],
[
0.26611328125,
0.2298583984375,
0.393798828125,
0.2298583984375,
0.393798828125,
0.240478515625,
0.26611328125,
0.240478515625
],
[
0.15283203125,
0.232177734375,
0.162109375,
0.232177734375,
0.162109375,
0.2369384765625,
0.15283203125,
0.2369384765625
],
[
0.466796875,
0.240478515625,
0.64208984375,
0.240478515625,
0.64208984375,
0.254150390625,
0.466796875,
0.254150390625
],
[
0.64990234375,
0.240478515625,
0.873046875,
0.240478515625,
0.873046875,
0.254150390625,
0.64990234375,
0.254150390625
],
[
0.1435546875,
0.2476806640625,
0.17626953125,
0.2476806640625,
0.17626953125,
0.260009765625,
0.1435546875,
0.260009765625
],
[
0.210693359375,
0.2423095703125,
0.453369140625,
0.2423095703125,
0.453369140625,
0.2529296875,
0.210693359375,
0.2529296875
],
[
0.466796875,
0.2529296875,
0.57080078125,
0.25048828125,
0.57080078125,
0.26416015625,
0.466796875,
0.2666015625
],
[
0.6513671875,
0.251708984375,
0.8759765625,
0.251708984375,
0.8759765625,
0.265380859375,
0.6513671875,
0.265380859375
],
[
0.145263671875,
0.26416015625,
0.175537109375,
0.267333984375,
0.173828125,
0.27783203125,
0.1435546875,
0.27490234375
],
[
0.2091064453125,
0.254150390625,
0.300537109375,
0.254150390625,
0.300537109375,
0.26416015625,
0.2091064453125,
0.26416015625
],
[
0.313232421875,
0.265380859375,
0.34765625,
0.265380859375,
0.34765625,
0.276611328125,
0.313232421875,
0.276611328125
],
[
0.53564453125,
0.26416015625,
0.57080078125,
0.26416015625,
0.57080078125,
0.276611328125,
0.53564453125,
0.276611328125
],
[
0.71728515625,
0.265380859375,
0.8095703125,
0.265380859375,
0.8095703125,
0.276611328125,
0.71728515625,
0.276611328125
],
[
0.118408203125,
0.297607421875,
0.88330078125,
0.297607421875,
0.88330078125,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.118408203125,
0.312255859375,
0.884765625,
0.312255859375,
0.884765625,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.118408203125,
0.32666015625,
0.88330078125,
0.32666015625,
0.88330078125,
0.34033203125,
0.118408203125,
0.34033203125
],
[
0.12005615234375,
0.366455078125,
0.424072265625,
0.366455078125,
0.424072265625,
0.38134765625,
0.12005615234375,
0.38134765625
],
[
0.51318359375,
0.36767578125,
0.88330078125,
0.36767578125,
0.88330078125,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.51318359375,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.51318359375,
0.3955078125
],
[
0.12005615234375,
0.388427734375,
0.490234375,
0.388427734375,
0.490234375,
0.402587890625,
0.12005615234375,
0.402587890625
],
[
0.51318359375,
0.399658203125,
0.85986328125,
0.399658203125,
0.85986328125,
0.41455078125,
0.51318359375,
0.41455078125
],
[
0.118408203125,
0.40625,
0.490234375,
0.40625,
0.490234375,
0.419189453125,
0.118408203125,
0.419189453125
],
[
0.1158447265625,
0.42041015625,
0.49365234375,
0.42041015625,
0.49365234375,
0.438232421875,
0.1158447265625,
0.438232421875
],
[
0.51318359375,
0.42578125,
0.88330078125,
0.42578125,
0.88330078125,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.118408203125,
0.43701171875,
0.489501953125,
0.43701171875,
0.489501953125,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.51123046875,
0.4423828125,
0.884765625,
0.4423828125,
0.884765625,
0.4560546875,
0.51123046875,
0.4560546875
],
[
0.118408203125,
0.45361328125,
0.489501953125,
0.45361328125,
0.489501953125,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.51318359375,
0.457763671875,
0.88330078125,
0.457763671875,
0.88330078125,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.118408203125,
0.4697265625,
0.489501953125,
0.4697265625,
0.489501953125,
0.482666015625,
0.118408203125,
0.482666015625
],
[
0.51123046875,
0.474365234375,
0.884765625,
0.47265625,
0.884765625,
0.486328125,
0.51123046875,
0.488037109375
],
[
0.1175537109375,
0.48388671875,
0.48193359375,
0.485107421875,
0.48193359375,
0.49951171875,
0.1175537109375,
0.498779296875
],
[
0.51318359375,
0.490478515625,
0.88330078125,
0.490478515625,
0.88330078125,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.51318359375,
0.505859375,
0.884765625,
0.505859375,
0.884765625,
0.51953125,
0.51318359375,
0.51953125
],
[
0.118408203125,
0.51318359375,
0.490234375,
0.51318359375,
0.490234375,
0.52685546875,
0.118408203125,
0.52685546875
],
[
0.51123046875,
0.5224609375,
0.88330078125,
0.5224609375,
0.88330078125,
0.5361328125,
0.51123046875,
0.5361328125
],
[
0.1175537109375,
0.529296875,
0.489501953125,
0.52783203125,
0.489501953125,
0.5419921875,
0.1175537109375,
0.54345703125
],
[
0.1175537109375,
0.54345703125,
0.490234375,
0.54345703125,
0.490234375,
0.55810546875,
0.1175537109375,
0.55810546875
],
[
0.51318359375,
0.5390625,
0.884765625,
0.5390625,
0.884765625,
0.552734375,
0.51318359375,
0.552734375
],
[
0.51318359375,
0.5546875,
0.88671875,
0.5546875,
0.88671875,
0.568359375,
0.51318359375,
0.568359375
],
[
0.118408203125,
0.56103515625,
0.490234375,
0.56103515625,
0.490234375,
0.57470703125,
0.118408203125,
0.57470703125
],
[
0.118408203125,
0.57666015625,
0.489501953125,
0.57666015625,
0.489501953125,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.51318359375,
0.57177734375,
0.884765625,
0.57177734375,
0.884765625,
0.5849609375,
0.51318359375,
0.5849609375
],
[
0.1175537109375,
0.5927734375,
0.489501953125,
0.5927734375,
0.489501953125,
0.60546875,
0.1175537109375,
0.60546875
],
[
0.51318359375,
0.58740234375,
0.884765625,
0.58740234375,
0.884765625,
0.60107421875,
0.51318359375,
0.60107421875
],
[
0.51318359375,
0.6025390625,
0.8427734375,
0.6025390625,
0.8427734375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.118408203125,
0.60986328125,
0.490234375,
0.60986328125,
0.490234375,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.118408203125,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.118408203125,
0.6396484375
],
[
0.5146484375,
0.63134765625,
0.87890625,
0.63134765625,
0.87890625,
0.6455078125,
0.5146484375,
0.6455078125
],
[
0.1175537109375,
0.64013671875,
0.489501953125,
0.64013671875,
0.489501953125,
0.65380859375,
0.1175537109375,
0.65380859375
],
[
0.51123046875,
0.65185546875,
0.88671875,
0.65283203125,
0.88671875,
0.66748046875,
0.51123046875,
0.66650390625
],
[
0.12005615234375,
0.6591796875,
0.489501953125,
0.6591796875,
0.489501953125,
0.66943359375,
0.12005615234375,
0.66943359375
],
[
0.1175537109375,
0.6728515625,
0.48779296875,
0.6728515625,
0.48779296875,
0.68603515625,
0.1175537109375,
0.68603515625
],
[
0.51318359375,
0.66943359375,
0.884765625,
0.66845703125,
0.884765625,
0.68310546875,
0.51318359375,
0.68408203125
],
[
0.1175537109375,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.1175537109375,
0.703125
],
[
0.51318359375,
0.68408203125,
0.87158203125,
0.68603515625,
0.87158203125,
0.70068359375,
0.51318359375,
0.69873046875
],
[
0.118408203125,
0.7060546875,
0.490234375,
0.7060546875,
0.490234375,
0.7197265625,
0.118408203125,
0.7197265625
],
[
0.5693359375,
0.7060546875,
0.82958984375,
0.7080078125,
0.82958984375,
0.7255859375,
0.5693359375,
0.7236328125
],
[
0.8564453125,
0.71044921875,
0.88330078125,
0.71044921875,
0.88330078125,
0.72265625,
0.8564453125,
0.72265625
],
[
0.118408203125,
0.72265625,
0.489501953125,
0.72265625,
0.489501953125,
0.736328125,
0.118408203125,
0.736328125
],
[
0.12005615234375,
0.73828125,
0.489501953125,
0.73828125,
0.489501953125,
0.751953125,
0.12005615234375,
0.751953125
],
[
0.5146484375,
0.73291015625,
0.884765625,
0.73291015625,
0.884765625,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.11419677734375,
0.751953125,
0.49365234375,
0.75,
0.49365234375,
0.77099609375,
0.11419677734375,
0.7724609375
],
[
0.51318359375,
0.75,
0.884765625,
0.75,
0.884765625,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.118408203125,
0.76953125,
0.491943359375,
0.76953125,
0.491943359375,
0.78759765625,
0.118408203125,
0.78759765625
],
[
0.51318359375,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.12005615234375,
0.7861328125,
0.45166015625,
0.7861328125,
0.45166015625,
0.80029296875,
0.12005615234375,
0.80029296875
],
[
0.51318359375,
0.7822265625,
0.69287109375,
0.7822265625,
0.69287109375,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.56689453125,
0.80419921875,
0.83642578125,
0.80419921875,
0.83642578125,
0.82177734375,
0.56689453125,
0.82177734375
],
[
0.85791015625,
0.8056640625,
0.88330078125,
0.8056640625,
0.88330078125,
0.818359375,
0.85791015625,
0.818359375
],
[
0.12005615234375,
0.8134765625,
0.489501953125,
0.8134765625,
0.489501953125,
0.82666015625,
0.12005615234375,
0.82666015625
],
[
0.1175537109375,
0.8291015625,
0.490234375,
0.830078125,
0.490234375,
0.84423828125,
0.1175537109375,
0.84326171875
],
[
0.509765625,
0.8291015625,
0.884765625,
0.8291015625,
0.884765625,
0.8466796875,
0.509765625,
0.8466796875
],
[
0.118408203125,
0.845703125,
0.490234375,
0.845703125,
0.490234375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.5146484375,
0.845703125,
0.88671875,
0.845703125,
0.88671875,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.14111328125,
0.87158203125,
0.489501953125,
0.87158203125,
0.489501953125,
0.88525390625,
0.14111328125,
0.88525390625
],
[
0.5341796875,
0.87060546875,
0.875,
0.87060546875,
0.875,
0.88525390625,
0.5341796875,
0.88525390625
],
[
0.118408203125,
0.88427734375,
0.490234375,
0.88427734375,
0.490234375,
0.89794921875,
0.118408203125,
0.89794921875
],
[
0.5322265625,
0.8818359375,
0.884765625,
0.88427734375,
0.884765625,
0.8984375,
0.5322265625,
0.896484375
],
[
0.12005615234375,
0.89794921875,
0.4189453125,
0.89794921875,
0.4189453125,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.51318359375,
0.89794921875,
0.71533203125,
0.89794921875,
0.71533203125,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"WIQA",
"QuaRel",
"HotpotQA",
"(Tandon et al., 2019)",
"(Tafjord et al., 2019)",
"(Yang et al., 2018)",
"Causal Reasoning",
"Qualitative Reasoning",
"Qualitative Comparison of entities",
"format",
"reasoning",
"classification",
"multiple choice",
"span extraction",
"The rain seeps into the wood surface.",
"Supposed you were stand-",
"Golf Magazine is a monthly golf",
"When rain evaporates it leaves the wood.",
"ing on the planet Earth and",
"magazine owned by Time Inc. El",
"",
"It takes the finish of the wood with it.",
"Mercury. When you look up",
"Nuevo Cojo Ilustrado is an Ameri-",
"in the sky and see the sun,",
"can Spanish language magazine",
"gr:lf a tsunami happens, will wood be",
"The wood begins to lose it's luster.",
"Which planet would the sun",
"EI Nuevo Cojo and Golf Magazine,",
"more moist?, q2:",
"If wood is more",
"appear larger?",
"which one is owned by Time Inc?",
"moist, is more weathering occurring?",
"q",
"[Mercury, Earth",
"[Golf Magazine, El Nuevo Cojo]",
"[more, less, no effects)",
"A",
"Mercury",
"GolfMagazine",
"ai: more, az: more",
"0",
"Which planet would the sun",
"Which one is not owned by Time",
"qaug",
"If a tsunami happens, is more weather-",
"appear smaller?",
"Inc, Golf Magazine El Nuevo Cojo?",
"aaug",
"ing occurring?",
"more",
"Earth",
"El Nuevo Cojo",
"Table 1: An augmented transitive example for WIQA, and symmetric examples for QuaRel and HotpotQA. We",
"partially modify paragraphs and questions. The bold characters denote a shared event connecting two questions.",
"The parts written in red or blue denote antonyms, and highlighted text is negation added by our data augmentation.",
"3.2 Logic-guided Data Augmentation",
"(ci, ei) and (c2, e2), where ei = C2 holds. When",
"a* is a positive causal relationship, we create a new",
"Given a set of training examples X in the form",
"example Xtrans = (q3, p, a2) for q3 = (c1, e2)",
"of (q,p, a*), we automatically generate additional",
"examples Xaug = [qaug, P, aaug] using symme-",
"Sampling augmented data Adding all consis-",
"try and transitivity logical rules. The goal is to",
"tent examples may change the data distribution",
"augment the training data so that symmetric and",
"from the original one, which may lead to a dete-",
"transitive examples are observed during training.",
"rioration in performance (Xie et al., 2019). One",
"We provide some augmented examples in Table 1.",
"can select the data based on a model's prediction",
"inconsistencies (Minervini and Riedel, 2018) or",
"Augmenting symmetric examplesTo create a",
"randomly sample at each epoch (Kang et al., 2018)",
"symmetric question, we convert a question into",
"an opposite one using the following operations:",
"In this work, we randomly sample augmented data",
"at the beginning of training, and use the same ex-",
"(a) replace words with their antonyms, (b) add, or",
"(c) remove words. For (a), we select top frequent",
"amples for all epochs during training.Despite its",
"adjectives or verbs with polarity (e.g., smaller, in-",
"simplicity, this yields competitive or even better",
"performance than other sampling strategies.3",
"creases) from training corpora, and expert annota-",
"tors write antonyms for each of the frequent words",
"3.3Logic-guided Consistency Regularization",
"(we denote this small dictionary as D). More de-",
"We regularize the learning objective (task loss,",
"tailscanbeseenin Appendix A.For(b)and (c),",
"we add negation words or remove negation words",
"Ltask) with a regularization term that promotes con-",
"(e.g., not). For all of the questions in training data,",
"sistency of predictions (consistency loss, Ccons).",
"if a question includes a word in D for the oper-",
"C = Ctask(X) + Ccons(X, Xaug).",
"(3)",
"ation (a), or matches a template (e.g., which *",
"is which * is not)foroperations(b) and",
"The first term Ctask penalizes making incorrect pre-",
"(c), we apply the operation to generate Qsym.2 We",
"dictions. The second term Lcons penalizes making",
"obtain asym by re-labeling the answer a* to its op-",
"predictions that violate symmetric and transitive",
"posite answer choice in A (see Appendix B).",
"logical rules as follows:",
"Lcons = AsymLsym + transLtrans,",
"(4)",
"Augmenting transitive examplesWe first find a",
"pair of two cause-effect questions Xi = (qi, p, a*)",
"where Asym and trans are weighting scalars to",
"and X2 = (q2, P, a2), whose q1 and q2 consist of",
"balance the two consistency-promoting objectives.",
"2 We observe that (b)(c) are less effective than (a) in WIQA",
"3We do not add Xaug if the same pair has already exist.",
"or QuaRel, while especially (b) contributes to the performance",
"*We mask the Ccons for the examples without symmetric",
"improvementsonHotpotQAasmuchas(a)does.",
"or transitive consistent examples."
] |
{
"arxiv_id": "2004.10157",
"language": "en",
"page_idx": 2
}
|
||
2004.10157-en-4
|
2004.10157-en-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>WIQA Input</th>
<th>$\operatorname{RoBERTa}$</th>
<th>$DA$</th>
<th>$\mathbb{D}\mathbb{A}+\mathbb{R}$eg</th>
</tr>
<tr>
<td>$p$</td>
<td>Sound enters c the ears</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$q$</td>
<td>If the person has his ears more rotected. will less sound be detected? la $x^{\ast}\colon$More</td>
<td>(0.79) $1ore$</td>
<td>More (0.93)</td>
<td>More (0.93</td>
</tr>
<tr>
<td>$q_{sym}$</td>
<td>If the person has his ears less protected, will less sound be detected? $[a^\textit{sym}*:$ Less</td>
<td>$1ore$ (0.87)</td>
<td>(0.72) More</td>
<td>Less (0.89.</td>
</tr>
<tr>
<td>$P$</td>
<td>Squirrels try to eat as mu possible. Squirrel gains weight</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$q_1$</td>
<td>If the weather has a lot of snow, cannot squirrels eat as much as possible? $[a_{1}^{*}:$ More</td>
<td>(0.75) cSS</td>
<td>(0.48) More</td>
<td>More (0.94</td>
</tr>
<tr>
<td>$q_2$</td>
<td>sourre s cannot eat as much a s noss 1hle , will not the squrrels gaın weıght? $|a_{\circ}:$ Mor</td>
<td>$10re$ (0.86)</td>
<td>More (0.94)</td>
<td>More (0.93</td>
</tr>
<tr>
<td rowspan="2">$q_{trans}$</td>
<td>If the weather has a lot of snow, wi I nol the squirrels gain weight? $[a_{tra}^{*}$ $_{\mathrm{e} }:$More</td>
<td>(0.75) cSS</td>
<td>More (0.43)</td>
<td>More (0.87</td>
</tr>
<tr>
<td>HotpotQA (comparison) Input</td>
<td>RoBF</td>
<td>ERTa</td>
<td>$DA$</td>
</tr>
<tr>
<td>$p$</td>
<td>B. Reeves Fason is a film direct r. Albert S. Rogell a film director</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$q$</td>
<td>Who has more scope of profession, B. Reeves Eason or Albert S. Rogell? $\{a^*:$B. Reeves Easo</td>
<td>B. Reeve</td>
<td>'es Eason</td>
<td>Reeves Eason $\frac{1}{2}$</td>
</tr>
<tr>
<td>$Z_{8}ym$ $W$</td>
<td>Who has less scope of profession 11 R.Reeves or Albert S. Rogell Albert S. Rogell</td>
<td>B. Reeve</td>
<td>'es Eason</td>
<td>Albert S. Rogell</td>
</tr>
</tbody>
</table>
Table 4: Qualitative comparison of RoBERTa, + DA, + DA + Reg. The examples are partially modified
cantly reduces the inconsistencies in predictions, demonstrating the effects of both data augmentation and regularization components. Notably on WIQA, RoBERTa shows violation of consistency in 13.9% of the symmetric examples and 10.0% of the transitive examples. Our approach reduces the violations of symmetric and transitive consistencies to 8.3% and 2.5%, respectively.
ground-truth labels for the questions where DA
shows relatively small probability differences.
On HotpotQA, the baseline model shows large consistency violations as shown in Table 2. The HotpotQA example in Table 4 shows that RoBERTa selects the same answer to both $q$ and $q_{sym}$, while DA answers correctly to both questions, demonstrating its robustness to surface variations. We hypothesize that the baseline model exploits statistical pattern, or dataset bias presented in questions and that our method reduces the model's tendency to exploit those spurious statistical patterns (He et al., 2019; Elkahky et al., 2018), which leads to large improvements in consistency.
Results with limited training data Table 2 also
shows that our approach is especially effective under the scarce training data setting: when only 20% of labeled data is available, our DA and Reg together gives more than 12% and 14% absolute accuracy improvements over the RoBERTa baselines on WIQA and QuaRel, respectively.
### 5 $\mathbf{Conclusion}$
We introduce a logic guided data augmentation and consistency-based regularization framework for accurate and globally consistent QA, especially under limited training data setting. Our approach significantly improves the state-of-the-art models across three substantially different QA datasets. Notably, our approach advances the state-of-the-art on QuaRel and WIQA, two standard benchmarks requiring rich logical and language understanding. We further show that our approach can effectively learn from extremely limited training data.
Ablation study We analyze the effectiveness of each component on Table 3. DA and Reg each improves the baselines, and the combination performs the best on WIQA and QuaRel. DA (standard) follows a previous standard data augmentation technique that paraphrases words (verbs and adjectives) using linguistic knowledge, namely WordNet (Miller, 1995), and does not incorporate logical rules. Importantly, DA (standard) does not give notable improvement over the baseline model both in accuracy and consistency, which suggests that logic-guided augmentation gives additional inductive bias for consistent QA beyond amplifying the number of train data. As WIQA consists of some transitive or symmetric examples, we also report the performance with Reg only on WIQA. The performance improvements is smaller, demonstrating the importance of combining with DA.
## Acknowledgments
This research was supported by ONR N00014- 18-1-2826, DARPA N66001-19-2-403, NSF $(\Pi S1616112,\Pi S1252835)$, Allen Distinguished Investigator Award, Sloan Fellowship, and The Nakajima Foundation Fellowship. We thank Antoine Bosselut, Tim Dettmers, Rik Koncel-Kedziorski, Sewon Min, Keisuke Sakaguchi, David Wadden, Yizhong Wang, the members of UW NLP group and AI2, and the anonymous reviewers for their insightful feedback
Qualitative Analysis Table 4 shows qualitative examples, comparing our method with RoBERTa baseline. Our qualitative analysis shows that DA+Reg reduces the confusion between opposite choices, and assigns larger probabilities to the
|
[
[
0.1956787109375,
0.07781982421875,
0.258544921875,
0.07781982421875,
0.258544921875,
0.0885009765625,
0.1956787109375,
0.0885009765625
],
[
0.6513671875,
0.07659912109375,
0.7021484375,
0.07659912109375,
0.7021484375,
0.0885009765625,
0.6513671875,
0.0885009765625
],
[
0.74658203125,
0.0789794921875,
0.765625,
0.0789794921875,
0.765625,
0.0885009765625,
0.74658203125,
0.0885009765625
],
[
0.81201171875,
0.07781982421875,
0.85986328125,
0.07781982421875,
0.85986328125,
0.08905029296875,
0.81201171875,
0.08905029296875
],
[
0.193115234375,
0.0926513671875,
0.56689453125,
0.0926513671875,
0.56689453125,
0.10272216796875,
0.193115234375,
0.10272216796875
],
[
0.19482421875,
0.10272216796875,
0.58447265625,
0.10272216796875,
0.58447265625,
0.1134033203125,
0.19482421875,
0.1134033203125
],
[
0.646484375,
0.10272216796875,
0.70947265625,
0.10272216796875,
0.70947265625,
0.1134033203125,
0.646484375,
0.1134033203125
],
[
0.724609375,
0.10272216796875,
0.78662109375,
0.10272216796875,
0.78662109375,
0.1134033203125,
0.724609375,
0.1134033203125
],
[
0.80419921875,
0.10272216796875,
0.8671875,
0.10272216796875,
0.8671875,
0.1134033203125,
0.80419921875,
0.1134033203125
],
[
0.2435302734375,
0.11102294921875,
0.59423828125,
0.10986328125,
0.59423828125,
0.12353515625,
0.2435302734375,
0.12469482421875
],
[
0.6455078125,
0.11102294921875,
0.7060546875,
0.11224365234375,
0.7060546875,
0.122314453125,
0.6455078125,
0.1217041015625
],
[
0.724609375,
0.11102294921875,
0.7841796875,
0.11224365234375,
0.7841796875,
0.122314453125,
0.724609375,
0.1217041015625
],
[
0.8095703125,
0.114013671875,
0.86376953125,
0.114013671875,
0.86376953125,
0.1217041015625,
0.8095703125,
0.1217041015625
],
[
0.22412109375,
0.11639404296875,
0.2376708984375,
0.11639404296875,
0.2376708984375,
0.12054443359375,
0.22412109375,
0.12054443359375
],
[
0.19482421875,
0.1134033203125,
0.225830078125,
0.1134033203125,
0.225830078125,
0.12054443359375,
0.19482421875,
0.12054443359375
],
[
0.19482421875,
0.12353515625,
0.490234375,
0.12353515625,
0.490234375,
0.1341552734375,
0.19482421875,
0.1341552734375
],
[
0.646484375,
0.1318359375,
0.7060546875,
0.1318359375,
0.7060546875,
0.142578125,
0.646484375,
0.142578125
],
[
0.80419921875,
0.1318359375,
0.8671875,
0.1318359375,
0.8671875,
0.142578125,
0.80419921875,
0.142578125
],
[
0.724609375,
0.133056640625,
0.78662109375,
0.133056640625,
0.78662109375,
0.1436767578125,
0.724609375,
0.1436767578125
],
[
0.1502685546875,
0.1353759765625,
0.162841796875,
0.1353759765625,
0.162841796875,
0.1436767578125,
0.1502685546875,
0.1436767578125
],
[
0.19482421875,
0.133056640625,
0.60205078125,
0.133056640625,
0.60205078125,
0.1436767578125,
0.19482421875,
0.1436767578125
],
[
0.193115234375,
0.141357421875,
0.62060546875,
0.141357421875,
0.62060546875,
0.155029296875,
0.193115234375,
0.155029296875
],
[
0.646484375,
0.142578125,
0.7060546875,
0.142578125,
0.7060546875,
0.152587890625,
0.646484375,
0.152587890625
],
[
0.724609375,
0.142578125,
0.78662109375,
0.142578125,
0.78662109375,
0.15380859375,
0.724609375,
0.15380859375
],
[
0.80419921875,
0.142578125,
0.86572265625,
0.142578125,
0.86572265625,
0.15380859375,
0.80419921875,
0.15380859375
],
[
0.80419921875,
0.15087890625,
0.8671875,
0.152587890625,
0.8671875,
0.1644287109375,
0.80419921875,
0.162109375
],
[
0.1485595703125,
0.144287109375,
0.16455078125,
0.144287109375,
0.16455078125,
0.152587890625,
0.1485595703125,
0.152587890625
],
[
0.193115234375,
0.152587890625,
0.59033203125,
0.152587890625,
0.59033203125,
0.163330078125,
0.193115234375,
0.163330078125
],
[
0.6484375,
0.152587890625,
0.705078125,
0.152587890625,
0.705078125,
0.163330078125,
0.6484375,
0.163330078125
],
[
0.72705078125,
0.1519775390625,
0.78857421875,
0.1519775390625,
0.78857421875,
0.163330078125,
0.72705078125,
0.163330078125
],
[
0.6884765625,
0.1650390625,
0.73876953125,
0.1650390625,
0.73876953125,
0.177001953125,
0.6884765625,
0.177001953125
],
[
0.139404296875,
0.155029296875,
0.162109375,
0.155029296875,
0.162109375,
0.162109375,
0.139404296875,
0.162109375
],
[
0.197265625,
0.166259765625,
0.341796875,
0.166259765625,
0.341796875,
0.177001953125,
0.197265625,
0.177001953125
],
[
0.8134765625,
0.168701171875,
0.82958984375,
0.168701171875,
0.82958984375,
0.17578125,
0.8134765625,
0.17578125
],
[
0.193115234375,
0.1817626953125,
0.62060546875,
0.1817626953125,
0.62060546875,
0.1923828125,
0.193115234375,
0.1923828125
],
[
0.19482421875,
0.190673828125,
0.35107421875,
0.191162109375,
0.35107421875,
0.203125,
0.19482421875,
0.201904296875
],
[
0.34765625,
0.193603515625,
0.59375,
0.193603515625,
0.59375,
0.20068359375,
0.34765625,
0.20068359375
],
[
0.615234375,
0.193603515625,
0.646484375,
0.193603515625,
0.646484375,
0.20068359375,
0.615234375,
0.20068359375
],
[
0.6767578125,
0.193603515625,
0.75244140625,
0.193603515625,
0.75244140625,
0.20068359375,
0.6767578125,
0.20068359375
],
[
0.7841796875,
0.193603515625,
0.85400390625,
0.193603515625,
0.85400390625,
0.20068359375,
0.7841796875,
0.20068359375
],
[
0.59375,
0.1953125,
0.60693359375,
0.1953125,
0.60693359375,
0.198974609375,
0.59375,
0.198974609375
],
[
0.15283203125,
0.1953125,
0.15869140625,
0.1953125,
0.15869140625,
0.20068359375,
0.15283203125,
0.20068359375
],
[
0.193115234375,
0.1995849609375,
0.5341796875,
0.1995849609375,
0.5341796875,
0.213134765625,
0.193115234375,
0.213134765625
],
[
0.53662109375,
0.1995849609375,
0.63037109375,
0.20068359375,
0.63037109375,
0.2120361328125,
0.53662109375,
0.21142578125
],
[
0.67529296875,
0.201904296875,
0.755859375,
0.201904296875,
0.755859375,
0.2120361328125,
0.67529296875,
0.2120361328125
],
[
0.78271484375,
0.201904296875,
0.8623046875,
0.201904296875,
0.8623046875,
0.2120361328125,
0.78271484375,
0.2120361328125
],
[
0.5341796875,
0.204833984375,
0.54150390625,
0.204833984375,
0.54150390625,
0.2078857421875,
0.5341796875,
0.2078857421875
],
[
0.14111328125,
0.204833984375,
0.162109375,
0.204833984375,
0.162109375,
0.2120361328125,
0.14111328125,
0.2120361328125
],
[
0.151123046875,
0.2340087890625,
0.8486328125,
0.2340087890625,
0.8486328125,
0.2476806640625,
0.151123046875,
0.2476806640625
],
[
0.12005615234375,
0.273681640625,
0.49365234375,
0.273681640625,
0.49365234375,
0.287353515625,
0.12005615234375,
0.287353515625
],
[
0.51123046875,
0.273681640625,
0.88330078125,
0.2724609375,
0.88330078125,
0.287353515625,
0.51123046875,
0.2880859375
],
[
0.12005615234375,
0.290283203125,
0.491943359375,
0.290283203125,
0.491943359375,
0.303955078125,
0.12005615234375,
0.303955078125
],
[
0.51123046875,
0.289306640625,
0.85400390625,
0.289306640625,
0.85400390625,
0.302734375,
0.51123046875,
0.302734375
],
[
0.12005615234375,
0.30810546875,
0.48779296875,
0.30810546875,
0.48779296875,
0.318359375,
0.12005615234375,
0.318359375
],
[
0.53076171875,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.53076171875,
0.320556640625
],
[
0.118408203125,
0.322509765625,
0.489501953125,
0.322509765625,
0.489501953125,
0.336181640625,
0.118408203125,
0.336181640625
],
[
0.51318359375,
0.322509765625,
0.88671875,
0.322509765625,
0.88671875,
0.336181640625,
0.51318359375,
0.336181640625
],
[
0.118408203125,
0.33837890625,
0.491943359375,
0.33837890625,
0.491943359375,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.51318359375,
0.33837890625,
0.884765625,
0.33837890625,
0.884765625,
0.352783203125,
0.51318359375,
0.352783203125
],
[
0.118408203125,
0.355224609375,
0.490234375,
0.355224609375,
0.490234375,
0.3681640625,
0.118408203125,
0.3681640625
],
[
0.51123046875,
0.352783203125,
0.884765625,
0.35400390625,
0.884765625,
0.3681640625,
0.51123046875,
0.36767578125
],
[
0.1175537109375,
0.37060546875,
0.490234375,
0.37060546875,
0.490234375,
0.38427734375,
0.1175537109375,
0.38427734375
],
[
0.509765625,
0.369384765625,
0.88818359375,
0.3681640625,
0.88818359375,
0.385986328125,
0.509765625,
0.38720703125
],
[
0.1158447265625,
0.385498046875,
0.3525390625,
0.385986328125,
0.3525390625,
0.40087890625,
0.1158447265625,
0.399658203125
],
[
0.51318359375,
0.38720703125,
0.88671875,
0.38720703125,
0.88671875,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.5146484375,
0.402587890625,
0.88330078125,
0.402587890625,
0.88330078125,
0.416259765625,
0.5146484375,
0.416259765625
],
[
0.12005615234375,
0.410888671875,
0.489501953125,
0.410888671875,
0.489501953125,
0.424560546875,
0.12005615234375,
0.424560546875
],
[
0.5146484375,
0.42041015625,
0.880859375,
0.42041015625,
0.880859375,
0.43115234375,
0.5146484375,
0.43115234375
],
[
0.118408203125,
0.427490234375,
0.489501953125,
0.427490234375,
0.489501953125,
0.441162109375,
0.118408203125,
0.441162109375
],
[
0.51318359375,
0.435302734375,
0.88330078125,
0.43408203125,
0.88330078125,
0.44775390625,
0.51318359375,
0.448974609375
],
[
0.118408203125,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.118408203125,
0.457275390625
],
[
0.51318359375,
0.451904296875,
0.88330078125,
0.451904296875,
0.88330078125,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.118408203125,
0.460205078125,
0.490234375,
0.460205078125,
0.490234375,
0.473876953125,
0.118408203125,
0.473876953125
],
[
0.51318359375,
0.46728515625,
0.88330078125,
0.46728515625,
0.88330078125,
0.48095703125,
0.51318359375,
0.48095703125
],
[
0.118408203125,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.51123046875,
0.482177734375,
0.8330078125,
0.482666015625,
0.8330078125,
0.49755859375,
0.51123046875,
0.496337890625
],
[
0.12005615234375,
0.4921875,
0.489501953125,
0.4921875,
0.489501953125,
0.505859375,
0.12005615234375,
0.505859375
],
[
0.1175537109375,
0.505859375,
0.385498046875,
0.50830078125,
0.385498046875,
0.5224609375,
0.1175537109375,
0.52099609375
],
[
0.515625,
0.51318359375,
0.53662109375,
0.51318359375,
0.53662109375,
0.521484375,
0.515625,
0.521484375
],
[
0.53564453125,
0.51416015625,
0.54150390625,
0.51416015625,
0.54150390625,
0.51953125,
0.53564453125,
0.51953125
],
[
0.5400390625,
0.51025390625,
0.64404296875,
0.51025390625,
0.64404296875,
0.52490234375,
0.5400390625,
0.52490234375
],
[
0.12005615234375,
0.5322265625,
0.490234375,
0.5322265625,
0.490234375,
0.5458984375,
0.12005615234375,
0.5458984375
],
[
0.51123046875,
0.53369140625,
0.88330078125,
0.5361328125,
0.88330078125,
0.55029296875,
0.51123046875,
0.548828125
],
[
0.12005615234375,
0.55029296875,
0.490234375,
0.55029296875,
0.490234375,
0.56103515625,
0.12005615234375,
0.56103515625
],
[
0.5146484375,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56396484375,
0.5146484375,
0.56396484375
],
[
0.12005615234375,
0.5654296875,
0.489501953125,
0.5654296875,
0.489501953125,
0.5791015625,
0.12005615234375,
0.5791015625
],
[
0.5146484375,
0.568359375,
0.8818359375,
0.568359375,
0.8818359375,
0.58203125,
0.5146484375,
0.58203125
],
[
0.118408203125,
0.58056640625,
0.490234375,
0.58056640625,
0.490234375,
0.59423828125,
0.118408203125,
0.59423828125
],
[
0.51318359375,
0.5849609375,
0.88330078125,
0.5849609375,
0.88330078125,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.118408203125,
0.59716796875,
0.49365234375,
0.59716796875,
0.49365234375,
0.61083984375,
0.118408203125,
0.61083984375
],
[
0.51318359375,
0.60107421875,
0.88330078125,
0.60107421875,
0.88330078125,
0.6142578125,
0.51318359375,
0.6142578125
],
[
0.118408203125,
0.61328125,
0.491943359375,
0.61328125,
0.491943359375,
0.626953125,
0.118408203125,
0.626953125
],
[
0.5146484375,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.62890625,
0.5146484375,
0.62890625
],
[
0.118408203125,
0.62890625,
0.490234375,
0.62890625,
0.490234375,
0.642578125,
0.118408203125,
0.642578125
],
[
0.51318359375,
0.6328125,
0.884765625,
0.6328125,
0.884765625,
0.646484375,
0.51318359375,
0.646484375
],
[
0.1175537109375,
0.6435546875,
0.490234375,
0.64453125,
0.490234375,
0.6591796875,
0.1175537109375,
0.65771484375
],
[
0.51318359375,
0.6484375,
0.88330078125,
0.6484375,
0.88330078125,
0.662109375,
0.51318359375,
0.662109375
],
[
0.118408203125,
0.6611328125,
0.490234375,
0.6611328125,
0.490234375,
0.6748046875,
0.118408203125,
0.6748046875
],
[
0.51123046875,
0.66455078125,
0.884765625,
0.6650390625,
0.884765625,
0.6796875,
0.51123046875,
0.6787109375
],
[
0.12005615234375,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.689453125,
0.12005615234375,
0.689453125
],
[
0.51123046875,
0.6796875,
0.88330078125,
0.68115234375,
0.88330078125,
0.6953125,
0.51123046875,
0.69482421875
],
[
0.118408203125,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.118408203125,
0.70703125
],
[
0.51318359375,
0.69775390625,
0.8271484375,
0.69775390625,
0.8271484375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71044921875,
0.489501953125,
0.71044921875,
0.489501953125,
0.7236328125,
0.118408203125,
0.7236328125
],
[
0.1175537109375,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.1175537109375,
0.73828125
],
[
0.51318359375,
0.7236328125,
0.67333984375,
0.7236328125,
0.67333984375,
0.73681640625,
0.51318359375,
0.73681640625
],
[
0.12005615234375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75439453125,
0.12005615234375,
0.75439453125
],
[
0.5146484375,
0.75,
0.884765625,
0.75,
0.884765625,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.118408203125,
0.75830078125,
0.489501953125,
0.75830078125,
0.489501953125,
0.771484375,
0.118408203125,
0.771484375
],
[
0.5146484375,
0.765625,
0.88330078125,
0.765625,
0.88330078125,
0.779296875,
0.5146484375,
0.779296875
],
[
0.12005615234375,
0.77392578125,
0.491943359375,
0.77392578125,
0.491943359375,
0.78759765625,
0.12005615234375,
0.78759765625
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.79052734375,
0.489501953125,
0.79052734375,
0.489501953125,
0.80419921875,
0.118408203125,
0.80419921875
],
[
0.51318359375,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51318359375,
0.8125
],
[
0.118408203125,
0.8056640625,
0.403076171875,
0.8056640625,
0.403076171875,
0.8193359375,
0.118408203125,
0.8193359375
],
[
0.51318359375,
0.81396484375,
0.884765625,
0.81396484375,
0.884765625,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.490234375,
0.84716796875,
0.490234375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.51123046875,
0.845703125,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.20068359375,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.20068359375,
0.876953125
],
[
0.509765625,
0.85986328125,
0.884765625,
0.86083984375,
0.884765625,
0.87890625,
0.509765625,
0.87744140625
],
[
0.509765625,
0.87744140625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.509765625,
0.892578125
],
[
0.12005615234375,
0.8642578125,
0.202392578125,
0.8642578125,
0.202392578125,
0.87451171875,
0.12005615234375,
0.87451171875
],
[
0.118408203125,
0.87890625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.66162109375,
0.8955078125,
0.66162109375,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"WIQAInput",
"RoBERTa",
"DA",
"DA+Reg",
"Sound enters the ears of a person.The sound hits a drum that is inside the ears",
"If the person has his ears more protected, will less sound be detected? [a*: More]",
"More (0.79)",
"More (0.93)",
"More (0.93)",
"sonhashisarssspotdwillssundecd?[amss",
"More (0.87)",
"More(0.72)",
"Less (0.89)",
"ne",
"Ifthe",
"Squirrels try to eat as much as possible. Squirrel gains weight.",
"Less (0.75)",
"More (0.94)",
"More (0.48)",
"q1",
"If the weather has a lot of snow, cannot squirrels eat as much as possible? [a*: More]",
"If squirrels cannot eat as much as possible, will not the squirrels gain weight? [a: More]",
"More (0.86)",
"More (0.94)",
"More (0.93)",
"More (0.87)",
"92",
"If the weather has a lot of snow, will not the squirrels gain weight? [atrans:More",
"Less (0.75)",
"More (0.43)",
"RoBERTa",
"qtra",
"HotpotQA (comparison) Input",
"DA",
"B.Reeves Eason is a film director, actor and screenwriter. Albert S. Rogell a film director.",
"Who has morescopeof professic",
"on.B.ReevesEasonorAlbertS.Rogell?[a*:B.Re",
"Eason",
"B.ReevesEason",
"B.Reeves Easc",
"",
"9",
"Whohas lessscopeofprofession,BReevesorAber.Rogll?[a",
"AlbertS.Rogell",
"B.Reeves Eason",
"Albert S. Rogell",
"",
"qsu",
"Table 4: Qualitative comparison of RoBERTa, + DA, + DA + Reg. The examples are partially modified.",
"cantly reduces the inconsistencies in predictions,",
"ground-truth labels for the questions where DA",
"demonstrating the effects of both data augmenta-",
"shows relatively small probability differences.",
"tion andregularization components.Notably on",
"On HotpotQA, the baseline model shows large",
"WIQA, RoBERTa shows violation of consistency",
"consistency violations as shown in Table 2.",
"in 13.9% of the symmetric examples and 10.0% of",
"The HotpotQA example in Table 4 shows that",
"the transitive examples. Our approach reduces the",
"RoBERTa selects the same answer to both q and",
"violations of symmetric and transitive consistencies",
"Qsym, while DA answers correctly to both ques-",
"to 8.3% and 2.5%, respectively.",
"tions, demonstrating its robustness to surface varia-",
"tions. We hypothesize that the baseline model ex-",
"Results with limited training dataTable 2 also",
"ploits statistical pattern,or datasetbiaspresented in",
"shows that our approach is especially effective un-",
"questions and that our method reduces the model's",
"der the scarce training data setting: when only 20%",
"tendency to exploit those spurious statistical pat-",
"of labeled data is available, our DA and Reg to-",
"terns (He et al., 2019; Elkahky et al., 2018), which",
"gether gives more than 12% and 14% absolute ac-",
"leads to large improvements in consistency",
"curacy improvements over the RoBERTa baselines",
"on WIQA and QuaRel, respectively.",
"5",
"",
"Conclusion",
"Ablation studyWe analyze the effectiveness of",
"We introduce a logic guided data augmentation",
"each component on Table 3. DA and Reg each im-",
"and consistency-basedregularization framework",
"proves the baselines, and the combination performs",
"for accurate and globally consistent QA, especially",
"the best on WIQA and QuaRel. DA (standard) fol-",
"under limited training data setting. Our approach",
"lows a previous standard data augmentation tech-",
"significantly improves the state-of-the-art models",
"nique that paraphrases words (verbs and adjec-",
"across three substantially differentQA datasets",
"tives) using linguistic knowledge, namely Word-",
"Notably, our approach advances the state-of-the-art",
"Net (Miller, 1995), and does not incorporate logi-",
"on QuaRel and WIQA, two standard benchmarks",
"cal rules. Importantly, DA (standard) does not give",
"requiring rich logical and language understanding.",
"notableimprovementoverthebaselinemodelboth",
"We further show that our approach can effectively",
"in accuracy and consistency, which suggests that",
"learn from extremely limited training data.",
"logic-guided augmentation gives additional induc-",
"tive bias for consistent QA beyond amplifying the",
"Acknowledgments",
"numberoftraindata.AsWiQAconsistsof some",
"This research was supported by ONR N00014-",
"transitive or symmetric examples, we also report",
"18-1-2826,DARPAN66001-19-2-403,NSF",
"the performance with Reg only on WIQA. The per-",
"(IIS1616112, IIS1252835), Allen Distinguished In-",
"formance improvements is smaller, demonstrating",
"vestigator Award, Sloan Fellowship, and The Naka-",
"the importance of combining with DA.",
"jima Foundation Fellowship. We thank Antoine",
"Qualitative AnalysisTable 4 shows qualitative",
"Bosselut, Tim Dettmers, Rik Koncel-Kedziorski.",
"examples, comparing our method with RoBERTa",
"Sewon Min, Keisuke Sakaguchi, David Wadden,",
" Our qualitative analysis shows that",
"Yizhong Wang, the members of UW NLP group",
" and AI2, and the anonymous reviewers for their",
"baseline.",
"DA+Reg reduces the confusion between opposite",
"choices, and assigns larger probabilities to the",
"insightful feedback."
] |
{
"arxiv_id": "2004.10157",
"language": "en",
"page_idx": 4
}
|
||
1904.04144-en-0
|
1904.04144-en-0.jpg
|
Learning monocular depth est
Fabio Tosi, Filippo Aleotti, Matteo Poggi, Stefano Mattoccia Department of Computer Scien
University of Bologna, Italy
$\{$fabio.tosi5, filippo.aleo
$\infty0$
<FigureHere>
Figure 1. Overview of the pr Input image from KITTI datas our monoResMatch (bottom).
$స్సే$
## $\mathbf{Abstract}$
$\begin{aligned}Depth~estimation~from~a~si\end{aligned}$
without direct supervision f aging image synthesis on seq
$\sum_{\begin{array}{ccc}&ing~on~this~second~case,~in~this~pape\end{array}}$
chitecture designed to infer by synthesizing features from a different point of view, horizontally aligned with the i matching between the two cues. In contrast to previous works sharing this rationale, our network is the first trained end-to-end from scratch. Moreover, we show how obtaining proxy ground truth annotation through traditional stereo algorithms, such as Semi-Global Matching, enables more accurate monocular depth estimation still countering the need for expensive depth labels by keeping a self-supervised approach.
Exhaustive experimental results prove how
the synergy between i) the proposed monoResMatch architecture and ii) proxy-supervision attains state-of-theart for self-supervised monocular depth estimation. The code is publicly available at https://github.com/ fabiotosi92/monoResMatch-Tensorflow.
## $1. \textbf{Introduction}$
Inferring accurate depth information of a sensed scene is paramount for several applications such as autonomous driving, augmented reality and robotics. Although technologies such as LiDAR and time-of-flight are quite popular, obtaining depth from images is often the preferred choice. Compared to other sensors, those based on standard cameras potentially have several advantages: they are inexpensive, have a higher resolution and are suited for almost any environment. In this field, stereo is the preferred choice to infer disparity (i.e., the inverse of depth) from two or more images sensing the same area from different points of
view and Semi-Global Matching (SGM)[15] is a popular, yet effective algorithm to accomplish this task. However, inferring depth from a single image is particularly attractive because it does not require a stereo rig and overcomes some intrinsic limitations of a binocular setup (e.g., occlusions). On the other hand, it is an extremely challenging task due to the ill-posed nature of the problem. Nonetheless, deep learning enabled to achieve outstanding results for this task [7], although the gap with state-of-the-art stereo solutions is still huge[3, 24]. Self-supervised learning paradigms for monocular depth estimation [11,63,32,44,40,58] became very popular to overcome the need for costly ground truth annotations, usually obtained employing expensive active sensors and human post-processing [10, 35, 52]. Following this strategy, Convolutional Neural Networks (CNNs) can be trained to tackle depth estimation as an image synthesis task from stereo pairs o For this purpose, using ster sequences as supervision turned out to be more effective according to the literature. more constrained since a stereo setup is necessary for training, it does neither require adjacent frames in a sequenc jects in the scene. Moreover
1
|
[
[
0.1094970703125,
0.1337890625,
0.86279296875,
0.1337890625,
0.86279296875,
0.1483154296875,
0.1094970703125,
0.1483154296875
],
[
0.2459716796875,
0.1817626953125,
0.7236328125,
0.18310546875,
0.7236328125,
0.1976318359375,
0.2459716796875,
0.1962890625
],
[
0.2607421875,
0.2008056640625,
0.71142578125,
0.2008056640625,
0.71142578125,
0.21533203125,
0.2607421875,
0.21533203125
],
[
0.373291015625,
0.2171630859375,
0.59619140625,
0.2183837890625,
0.59619140625,
0.23291015625,
0.373291015625,
0.231689453125
],
[
0.1846923828125,
0.2374267578125,
0.78759765625,
0.2374267578125,
0.78759765625,
0.251953125,
0.1846923828125,
0.251953125
],
[
0.0196075439453125,
0.266357421875,
0.058013916015625,
0.26513671875,
0.063720703125,
0.373779296875,
0.02532958984375,
0.375
],
[
0.238525390625,
0.287353515625,
0.313720703125,
0.287353515625,
0.313720703125,
0.3017578125,
0.238525390625,
0.3017578125
],
[
0.1004638671875,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.33203125,
0.1004638671875,
0.33203125
],
[
0.080078125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.08087158203125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.02777099609375,
0.37060546875,
0.053924560546875,
0.37060546875,
0.053924560546875,
0.38818359375,
0.02777099609375,
0.38818359375
],
[
0.08087158203125,
0.3662109375,
0.465576171875,
0.3662109375,
0.465576171875,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.080078125,
0.37890625,
0.46728515625,
0.37890625,
0.46728515625,
0.392578125,
0.080078125,
0.392578125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.44580078125,
0.02532958984375,
0.44580078125
],
[
0.080078125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.080078125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.080078125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.080078125,
0.439453125
],
[
0.03350830078125,
0.443115234375,
0.0555419921875,
0.443115234375,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.07843017578125,
0.453857421875,
0.468017578125,
0.455078125,
0.468017578125,
0.4697265625,
0.07843017578125,
0.468505859375
],
[
0.5048828125,
0.462158203125,
0.8896484375,
0.462158203125,
0.8896484375,
0.472900390625,
0.5048828125,
0.472900390625
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50244140625,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.08087158203125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.50341796875,
0.490478515625,
0.6806640625,
0.490478515625,
0.6806640625,
0.50048828125,
0.50341796875,
0.50048828125
],
[
0.0294189453125,
0.498779296875,
0.053924560546875,
0.498779296875,
0.053924560546875,
0.5498046875,
0.0294189453125,
0.5498046875
],
[
0.080078125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.08087158203125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.0310516357421875,
0.5498046875,
0.052276611328125,
0.5498046875,
0.052276611328125,
0.57373046875,
0.0310516357421875,
0.57373046875
],
[
0.0767822265625,
0.54541015625,
0.4697265625,
0.54443359375,
0.4697265625,
0.5595703125,
0.0767822265625,
0.56005859375
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.56103515625,
0.4697265625,
0.56103515625,
0.4697265625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.07843017578125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.5888671875,
0.07843017578125,
0.5888671875
],
[
0.50244140625,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.583984375,
0.50244140625,
0.583984375
],
[
0.02777099609375,
0.5859375,
0.053924560546875,
0.5859375,
0.053924560546875,
0.62890625,
0.02777099609375,
0.62890625
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.07843017578125,
0.59130859375,
0.4697265625,
0.5888671875,
0.4697265625,
0.60400390625,
0.07843017578125,
0.60595703125
],
[
0.07843017578125,
0.60693359375,
0.46728515625,
0.60693359375,
0.46728515625,
0.62060546875,
0.07843017578125,
0.62060546875
],
[
0.50244140625,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.50244140625,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50244140625,
0.62890625
],
[
0.0236968994140625,
0.62548828125,
0.058013916015625,
0.62548828125,
0.058013916015625,
0.7041015625,
0.0236968994140625,
0.7041015625
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.5009765625,
0.62890625,
0.890625,
0.6298828125,
0.890625,
0.64453125,
0.5009765625,
0.64306640625
],
[
0.07843017578125,
0.63525390625,
0.468017578125,
0.63525390625,
0.468017578125,
0.6484375,
0.07843017578125,
0.6484375
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50341796875,
0.646484375,
0.890625,
0.646484375,
0.890625,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.08087158203125,
0.66552734375,
0.428955078125,
0.66552734375,
0.428955078125,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.50244140625,
0.68798828125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.71044921875,
0.2091064453125,
0.71044921875,
0.2091064453125,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.5009765625,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.5048828125,
0.72265625,
0.89208984375,
0.72265625,
0.89208984375,
0.73291015625,
0.5048828125,
0.73291015625
],
[
0.098876953125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.098876953125,
0.74853515625
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.751953125,
0.4697265625,
0.751953125,
0.4697265625,
0.76513671875,
0.080078125,
0.76513671875
],
[
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,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.07843017578125,
0.77978515625,
0.468017578125,
0.78076171875,
0.468017578125,
0.79541015625,
0.07843017578125,
0.79443359375
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.8935546875,
0.85693359375,
0.8935546875,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Learning monocular depth estimation infusing traditional stereo knowledge",
"FabioTosi.Filippo Aleotti, MatteoPoggi, Stefano Mattoccia",
"Department of Computer Science and Engineering (DISI)",
"University ofBologna,Italy",
"[fabio.tosi5, filippo.aleotti2, m.poggi, stefano.mattoccia }@unibo.it",
"6100",
"Abstract",
"Depth estimation from a single image represents a fas-",
"cinating, yet challenging problem with countless applica-",
"tions. Recent works proved that this task could be learned",
"8",
"without direct supervision from ground truth labelslever",
"aging image synthesis on sequences or stereo pairs. Focus-",
"LA",
"ing on this second case, in this paper we leverage stereo",
"matching in order to improve monocular depth estimation.",
"To this aim we propose monoResMatch, a novel deep ar-",
"'SO",
"chitecture designed to infer depth from a single input image",
"by synthesizing features from a different point of view, hor-",
"Figure 1.Overview of theproposed depth-from-mono solution.",
"izontally aligned with the input image, performing stereo",
"Input image from KITTI dataset (top). Estimated depth map by",
"matchingbetweenthetwocues.Incontrasttoprevious",
"ourmonoResMatch (bottom).",
"",
"works sharing this rationale, our network is the first trained",
"end-to-endfromscratch.Moreover,weshowhowobtaining",
"proxy ground truth annotation through traditional stereo",
"view and Semi-Global Matching (SGM) [15] is a popular,",
"4",
"algorithms, such as Semi-Global Matching, enables more",
"yet effective algorithm to accomplish this task. However,",
"accurate monocular depth estimation still countering the",
"need for expensive depth labels by keeping a self-supervised",
"inferring depth from a single image is particularly attractive",
"06",
"because it does not require a stereo rig and overcomes some",
"approach.Exhaustive experimental results prove how",
"the synergy between i) the proposed monoResMatch ar-",
"intrinsic limitations of a binocular setup (e.g., occlusions).",
"On the other hand, it is an extremely challenging task due",
"I:AIXI",
"chitecture and ii) proxy-supervision attains state-of-the-",
"to the ill-posed nature of the problem. Nonetheless, deep",
"art for self-supervised monocular depth estimation.The",
"code is publicly available at https://github.com/",
"learning enabledto achieve outstandingresultsforthistask",
"fabiotosi92/monoResMatch-Tensorflow.",
"[7], although the gap with state-of-the-art stereo solutions",
"is still huge [3, 24]. Self-supervised learning paradigms for",
"monocular depth estimation [11, 63, 32, 44, 40, 58] became",
"1. Introduction",
"very popular to overcome the need for costly ground truth",
"annotations,usuallyobtainedemployingexpensiveactive",
"Inferring accurate depth information of a sensed scene",
"sensors and human post-processing [10, 35, 52]. Following",
"is paramount for several applications such as autonomous",
"this strategy, Convolutional Neural Networks (CNNs) can",
"driving, augmented reality and robotics. Although tech-",
"be trained to tackle depth estimation as an image synthe-",
"nologies such as LiDAR and time-of-flight are quite pop-",
"sistaskfromstereo pairs ormonocularsequences[1l,63].",
"ular, obtaining depth from images is often the preferred",
"For this purpose, using stereo pairs rather than monocular",
"choice. Compared to other sensors, those based on standard",
"sequences as supervision turned out to be more effective",
"cameras potentially have several advantages: they are inex-",
"according to the literature. Although the former strategy is",
"pensive, have a higher resolution and are suited for almost",
"more constrained since a stereo setup is necessary for train-",
"any environment. In this field, stereo is the preferred choice",
"ing, it does neither require to infer relative pose between",
"to infer disparity (i.e., the inverse of depth) from two or",
"adjacent frames in a sequence nor to segment moving ob-",
"more images sensing the same area from different points of",
"jects in the scene. Moreover, a stereo setup does not require"
] |
{
"arxiv_id": "1904.04144",
"language": "en",
"page_idx": 0
}
|
||
1904.04144-en-1
|
1904.04144-en-1.jpg
|
camera motion, conversely to a monocular setup, to provide meaningful supervision. Other means for self-supervision consist into distilling proxy labels in place of more expensive annotations for various tasks [49,51,28,33,20,13].
In this paper, we propose $\underline\text{monocular Residual Matching}$ (shorten, monoResMatch), a novel end-to-end architecture trained to estimate depth from a monocular image leveraging a virtual stereo setup. In the first stage, we map input image into a features space, then we use such representation to estimate a frst depth outcome and consequently synthesize features aligned with a virtual right image. Finally, the last refinement module performs stereo matching between the real and synthesized representations. Differently from other frameworks following a similar rationale[30] that combines heterogeneous networks for synthesis [55] and stereo [34], we use a single architecture trained in end-toend fashion yielding a notable accuracy improvement compared to the existing solutions. Moreover, we leverage traditional knowledge from stereo to obtain accurate proxy labels in order to improve monocular depth estimation supervised by stereo pairs. We will show that, despite the presence of outliers in the produced labels, training according to this paradigm results in superior accuracy compared to image warping approaches for self-supervision. Experimental results on the KITTI raw dataset[9] will show that the synergy between the two aforementioned key components of our pipeline enables to achieve state-of-the-art results compared to other self-supervised frameworks for monocular depth estimation not requiring any ground truth annotation. Figure l shows an overview of our framework, depicting an input frame and the outcome of monoResMatch.
## $2. \textbf{ Related Work}$
In this section, we review the literature relevant to our work concerned with stereo/monocular depth estimation and proxy label distillation.
Stereo depth estimation. Most conventional dense stereo algorithms rely on some or all the well-known four steps thoroughly described in [46]. In this field, SGM [15] stood out for the excellent trade-off between accuracy and efficiency thus becoming very popular. Žbontar and LeCun [61] were the first to apply deep learning to stereo vision replacing the conventional matching costs calculation with a siamese CNN network trained to predict the similarity between patches. Luo et al.[29] cast the correspondence problem as a multi-class classification task, obtaining better results. Mayer et $al.[34]$ backed away from the previous approaches and proposed an end-to-end trainable network called DispNetC able to infer disparity directly from images. While DispNetC applies a 1-D correlation to mimic the cost volume, GCNet by Kendall et al.[17] exploited 3- D convolutions over a 4-D volume to obtain matching costs and finally applied a differentiable version of argmin to se-
lect the best disparity along this volume. Other works followed these two main strategies, building more complex architectures starting from DispNetC[37,25,57,47] or GCNet[3,26,18] respectively. The domain shift issue affecting these architectures $(e.g.$ synthetic to real) has been addressed in either offline [49] or online [50] fashion, or greatly reduced by guiding them with external depth measurements $(e.g.$ Lidar) [42].
Monocular depth estimation. Before the deep learning era, some works tackled depth-from-mono with MRF [45] or boosted classifiers [22]. However, with the increasing availability of ground truth depth data, supervised approaches based on CNNs [23,27,56,7] rapidly outperformed previous techniques. An attractive trend concerns the possibility of learning depth-from-mono in a selfsupervised manner, avoiding the need for expensive ground truth depth labels that are replaced by multiple views of the sensed scene. Then, supervision signals can be obtained by image synthesis according to the estimated depth, camera pose or both. In general, acquiring images from a stereo camera enables a more effective training than using a single, moving camera, since the pose between frames known. Concerning stereo supervision, Garg $etal.[8]$ first followed this approach, while Godard et $al.[11]$ introduced spatial transform network[16] and a left-right consistency loss. Other methods improved efficiency [40], deploying a pyramidal architecture, and accuracy by simulating a trinocular setup [44] or including joint semantic segmentation [60]. In [38], a strategy was proposed to reduce further the energy efficiency of [40] leveraging fixed-point quantization. The semi-supervised framework by Kuznietsov et $al.[21]$ combined stereo supervision with sparse LiDAR measurements. The work by Zhou $etal.[63]$ represents the first attempt to supervise a depth-from-mono framework with single camera sequences. This approach was improved including additional cues such as point-cloud alignment[32], differentiable DVO [53] and multi-task learning [64]. Zhan et al. [62] combined the two supervision approaches outlined so far deploying stereo sequences. Another class of methods [2,1,5] applied a generative adversarial paradigm to the monocular scenario.
Finally, relevant to our work is Single View Stereo matching (SVS) [30], processing a single image to obtain a second synthetic view using Deep3D [55] and then computing a disparity map between the two using DispNetC[34]. However, these two architectures are trained independently. Moreover, DispNetC is supervised with ground truth labels from synthetic[34] and real domains [35]. Differently, the framework we are going to introduce requires no ground truth at all and is elegantly trained in an end-to-end manner, outperforming SVS by a notable margin.
Proxy labels distillation. Since for most tasks ground truth labels are difficult and expensive to source, some
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.091552734375,
0.890625,
0.09027099609375,
0.890625,
0.10479736328125,
0.5009765625,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.45751953125,
0.1383056640625,
0.45751953125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.098876953125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1673583984375,
0.098876953125,
0.1673583984375
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08087158203125,
0.1685791015625,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.08087158203125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.07843017578125,
0.18310546875,
0.468017578125,
0.184326171875,
0.468017578125,
0.1976318359375,
0.07843017578125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.5009765625,
0.1988525390625,
0.6845703125,
0.1988525390625,
0.6845703125,
0.212158203125,
0.5009765625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.521484375,
0.21533203125,
0.89208984375,
0.21533203125,
0.89208984375,
0.2291259765625,
0.521484375,
0.2291259765625
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2310791015625,
0.89208984375,
0.2310791015625,
0.89208984375,
0.2442626953125,
0.50244140625,
0.2442626953125
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.246826171875,
0.89208984375,
0.246826171875,
0.89208984375,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.080078125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.080078125,
0.271484375
],
[
0.5009765625,
0.261474609375,
0.89208984375,
0.260009765625,
0.89208984375,
0.274658203125,
0.5009765625,
0.27587890625
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.276611328125,
0.890625,
0.27587890625,
0.890625,
0.290283203125,
0.50244140625,
0.291748046875
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.5009765625,
0.291748046875,
0.890625,
0.290283203125,
0.890625,
0.304931640625,
0.5009765625,
0.30615234375
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.306884765625,
0.890625,
0.306884765625,
0.890625,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.08087158203125,
0.32080078125,
0.465576171875,
0.32080078125,
0.465576171875,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.5009765625,
0.3212890625,
0.8935546875,
0.32080078125,
0.8935546875,
0.335205078125,
0.5009765625,
0.3359375
],
[
0.07843017578125,
0.333251953125,
0.468017578125,
0.333984375,
0.468017578125,
0.348388671875,
0.07843017578125,
0.347900390625
],
[
0.50341796875,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50341796875,
0.365478515625
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.5009765625,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.5009765625,
0.380126953125
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.396484375,
0.50244140625,
0.396484375
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.5009765625,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.410888671875,
0.5009765625,
0.410888671875
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5009765625,
0.412353515625,
0.890625,
0.410888671875,
0.890625,
0.425537109375,
0.5009765625,
0.4267578125
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.427978515625,
0.89208984375,
0.427978515625,
0.89208984375,
0.44140625,
0.50244140625,
0.44140625
],
[
0.080078125,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.5009765625,
0.442626953125,
0.89208984375,
0.44140625,
0.89208984375,
0.455810546875,
0.5009765625,
0.45703125
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.5009765625,
0.473388671875,
0.8896484375,
0.473388671875,
0.8896484375,
0.4873046875,
0.5009765625,
0.4873046875
],
[
0.080078125,
0.48486328125,
0.46728515625,
0.48486328125,
0.46728515625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.5009765625,
0.501953125,
0.89208984375,
0.50244140625,
0.89208984375,
0.517578125,
0.5009765625,
0.5166015625
],
[
0.08087158203125,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.50341796875,
0.51806640625,
0.8896484375,
0.51806640625,
0.8896484375,
0.5322265625,
0.50341796875,
0.5322265625
],
[
0.07843017578125,
0.52880859375,
0.4697265625,
0.52978515625,
0.4697265625,
0.54443359375,
0.07843017578125,
0.54345703125
],
[
0.5009765625,
0.5322265625,
0.890625,
0.5322265625,
0.890625,
0.54541015625,
0.5009765625,
0.54541015625
],
[
0.080078125,
0.54443359375,
0.398681640625,
0.54443359375,
0.398681640625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50341796875,
0.548828125,
0.890625,
0.548828125,
0.890625,
0.5625,
0.50341796875,
0.5625
],
[
0.50341796875,
0.56298828125,
0.890625,
0.56298828125,
0.890625,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.080078125,
0.57470703125,
0.2181396484375,
0.57470703125,
0.2181396484375,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57763671875,
0.890625,
0.57763671875,
0.890625,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.50341796875,
0.59326171875,
0.8896484375,
0.59326171875,
0.8896484375,
0.60693359375,
0.50341796875,
0.60693359375
],
[
0.098876953125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.098876953125,
0.61279296875
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.08087158203125,
0.6162109375,
0.468017578125,
0.6162109375,
0.468017578125,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.07843017578125,
0.6298828125,
0.2646484375,
0.6298828125,
0.2646484375,
0.64306640625,
0.07843017578125,
0.64306640625
],
[
0.5048828125,
0.62548828125,
0.8896484375,
0.62548828125,
0.8896484375,
0.63623046875,
0.5048828125,
0.63623046875
],
[
0.09722900390625,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.09722900390625,
0.65771484375
],
[
0.50244140625,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.50341796875,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50341796875,
0.66748046875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50341796875,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.6845703125,
0.89208984375,
0.6845703125,
0.89208984375,
0.69775390625,
0.50341796875,
0.69775390625
],
[
0.5009765625,
0.69873046875,
0.63818359375,
0.69970703125,
0.63818359375,
0.71337890625,
0.5009765625,
0.7119140625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.52294921875,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.52294921875,
0.73095703125
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.5009765625,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.5009765625,
0.74560546875
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50341796875,
0.76220703125
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50341796875,
0.77978515625,
0.8896484375,
0.77978515625,
0.8896484375,
0.79052734375,
0.50341796875,
0.79052734375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50244140625,
0.806640625
],
[
0.07843017578125,
0.81103515625,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.89208984375,
0.80810546875,
0.89208984375,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.08087158203125,
0.82763671875,
0.46728515625,
0.82763671875,
0.46728515625,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.50244140625,
0.82373046875,
0.89208984375,
0.82373046875,
0.89208984375,
0.8369140625,
0.50244140625,
0.8369140625
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.50244140625,
0.83837890625,
0.89208984375,
0.83837890625,
0.89208984375,
0.8515625,
0.50244140625,
0.8515625
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85302734375,
0.775390625,
0.85302734375,
0.775390625,
0.8671875,
0.50244140625,
0.8671875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.51953125,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.51953125,
0.8837890625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.5009765625,
0.8994140625
]
] |
[
"camera motion, conversely to a monocular setup, to provide",
"lect the best disparity along this volume. Other works fol-",
"meaningful supervision.Othermeans for self-supervision",
"lowed these two main strategies, building more complex",
"consist into distilling proxy labels in place of more expen-",
"architectures starting from DispNetC [37, 25, 57, 47] or",
"sive annotations for various tasks [49, 51, 28, 33, 20, 13].",
"GCNet [3, 26, 18] respectively. The domain shift issue af-",
"In this paper, we propose monocular Residual Matching",
"fecting these architectures (e.g. synthetic to real) has been",
"(shorten, monoResMatch), a novel end-to-end architecture",
"addressed in either offline [49] or online [50] fashion, or",
"trained to estimate depth from a monocular image leverag-",
"greatly reduced by guiding them with external depth mea-",
"ing a virtual stereo setup. In the first stage, we map input",
"surements (e.g. Lidar) [42].",
"image into a features space, then we use such representation",
"Monocular depth estimation. Before the deep learn-",
"to estimate a first depth outcome and consequently synthe-",
"ing era, some works tackled depth-from-mono with MRF",
"size features aligned with a virtual right image. Finally, the",
"[45] or boosted classifiers [22]. However, with the in-",
"last refinement module performs stereo matching between",
"creasing availability of ground truth depth data, supervised",
"the real and synthesized representations.Differently from",
"approaches based on CNNs [23, 27, 56, 7] rapidly out-",
"other frameworks following a similar rationale [30] that",
"performed previous techniques. An attractive trend con-",
"combines heterogeneous networks for synthesis [55] and",
"cerns the possibility of learning depth-from-mono in a self-",
"stereo [34l. we use a single architecture trained in end-to-",
"supervised manner, avoiding the need for expensive ground",
"end fashion yielding a notable accuracy improvement com-",
"truthdepthlabelsthatarereplaced bymultipleviews of the",
"pared to the existing solutions. Moreover, we leverage tra-",
"sensed scene. Then, supervision signals can be obtained by",
"ditional knowledge from stereo to obtain accurate proxy la-",
"image synthesis according to the estimated depth, camera",
"bels in order to improve monocular depth estimation super-",
"pose or both. In general, acquiring images from a stereo",
"vised by stereo pairs. We will show that, despite the pres-",
"camera enables a more effective training than using a sin-",
"ence of outliers in the produced labels, training according to",
"gle, moving camera, since the pose between frames known.",
"this paradigm results in superior accuracy compared to im-",
"Concerning stereo supervision, Garg et al. [8] first followed",
"age warping approaches for self-supervision. Experimental",
"this approach, while Godard et al. [1l] introduced spatial",
"results on the KITTI raw dataset [9] will show that the syn-",
"transform network [16] and a left-right consistency loss.",
"ergy between the two aforementioned key components of",
"Other methods improved efficiency [40], deploying a pyra-",
"our pipeline enables to achieve state-of-the-art results com-",
"midal architecture, and accuracy by simulating a trinocular",
"pared to other self-supervised frameworks for monocular",
"setup [44] or including joint semantic segmentation [60]. In",
"depthestimationnotreguiring any ground truthannotation.",
"[38], a strategy was proposed to reduce further the energy",
"Figure I shows an overview of our framework, depicting an",
"efficiency of [401 leveraging fixed-point quantization. The",
"input frame and the outcome of monoResMatch.",
"semi-supervised framework by Kuznietsov et al. [21] com-",
"bined stereo supervision with sparse LiDAR measurements.",
"2. Related Work",
"The work by Zhou et al. [63] represents the first attempt to",
"supervise a depth-from-mono framework with single cam-",
"In this section, we review the literature relevant to our",
"era sequences. This approach was improved including ad-",
"work concerned with stereo/monocular depth estimation",
"and proxy label distillation.",
"ditional cues such as point-cloud alignment[32l,differen-",
"Stereo depth estimation. Most conventional dense",
"tiable DVO [53] and multi-task learning [64]. Zhan et al.",
"[62] combined the two supervision approaches outlined so",
"stereo algorithms rely on some or all the well-known four",
"steps thoroughly described in [46]. In this field, SGM [15]",
"far deploying stereo sequences. Another class of methods",
"stood out for the excellent trade-off between accuracy and",
"[2, I, 5] applied a generative adversarial paradigm to the",
"monocularscenario.",
"efficiency thus becoming very popular. Zbontar and Le-",
"Cun [61] were the first to apply deep learning to stereo vi-",
"Finally, relevant to our work is Single View Stereo",
"sionreplacingtheconventional matching costs calculation",
"matching (SvS) [30], processing a single image to obtain a",
"with a siamese CNN network trained to predict the similar-",
"second synthetic view using Deep3D [55] and then comput-",
"ity between patches. Luo et al. [29] cast the correspondence",
"ing a disparity map between the two using DispNetC [34].",
"problemasamulti-class classificationtask,obtaining better",
"However,thesetwo architecturesaretrained independently.",
"results. Mayer et al. [34] backed away from the previous",
"Moreover, DispNetC is supervised with ground truth labels",
"approaches and proposed an end-to-endtrainable network",
"from synthetic [34] and real domains [35]. Differently, the",
"called DispNetC able toinferdisparity directly fromim-",
"framework we are going to introduce requires no ground",
"ages. While DispNetC applies a 1-D correlation to mimic",
"truth at all and is elegantly trained in an end-to-end manner,",
"the cost volume, GCNet by Kendall et al. [17] exploited 3-",
"outperforming SvS by a notable margin.",
"D convolutions over a 4-D volume to obtain matching costs",
"Proxy labels distillation. Since for most tasks ground",
"and finally applied a differentiable version of argmin to se-",
"truth labels are difficult and expensive to source, some"
] |
{
"arxiv_id": "1904.04144",
"language": "en",
"page_idx": 1
}
|
||
1904.04144-en-3
|
1904.04144-en-3.jpg
|
sary information required to compute a cost volume since it processes a single input image. Thus, no 1-D correlation layer can be imposed to encode geometrical constraints in this stage of our network. Then, upsampling modules are deployed to enrich feature representations through skipconnections and to extract two disparity maps, aligned respectively with the input frame and a virtual viewpoint on its right as in [11]. This process is carried out at each scale using 1-stride convolutional layers with kernel size $3\times3.$
## $3. 3. \textbf{Disparity Refinement}$
Given an initial estimate of the disparity at each scale obtained in the second part of the network, often characterized by errors at depth discontinuities and occluded regions, this stage predicts corresponding multi-scale residual signals [14] by a few stacked nonlinear layers that are then used to compute the final left-view aligned disparity map. This strategy allows us to simplify the end-to-end learning process of the entire network. Moreover, motivated by [30], we believe that geometrical constraints can play a central role in boosting the final depth accuracy. For this reason, we embed matching costs in feature space computed employing a horizontal correlation layer, typically deployed in deep stereo algorithms. To this end, we rely on the rightview disparity map computed previously to generate rightview features $\tilde{F}_R^0$ from the left ones $F_L^0$ using a differentiable bilinear sampler[16]. The network is also fed with error $e_L,i.e.$ the absolute difference between left and virtual right features at input resolution, with the latter backwarped at the same coordinates of the former, as in [24].
We point out once more that, differently from[30], our architecture produces both a synthetic right view, $i.e.$ its features representation, and computes the final disparity map following stereo rationale. This makes monoResMatch a single end-to-end architecture, effectively performing stereo out of a single input view rather than the combination of two models $(i.e.$,Deep3D [55] and DispNetC[34] for the two tasks outlined) trained independently as in [31]. Moreover, exhaustive experiments will highlight the superior accuracy achieved by our fully self-supervised, end-toend approach.
## 3.4. Training Loss
In order to train our multi-stage architecture, we define the total loss as a sum of two main contributions, a $\mathcal{L}_{init}$ term from the initial disparity estimation module and a $\mathcal{L}_{ref}$ term from the disparity refinement stage. Following [12], we embrace the idea to up-sample the predicted low-resolution disparity maps to the full input resolution and then compute the corresponding signals. This simple strategy is designed to force the inverse depth estimation to reproduce the same objective at each scale, thus leading to much better outcomes. In particular, we obtain the final
training loss as:
(1)
$$\mathcal{L}_{total}=\sum_{s=1}^{n_i}\mathcal{L}_{init}+\sum_{s=1}^{n_r}\mathcal{L}_{ref}$$
where $s$ indicates the output resolution, $n_i$ and $n_r$ the numbers of considered scales during loss computation, while $\mathcal{L}_{init}$ and $\mathcal{L}_{ref}$ are formalised as:
(2)
$$\begin{aligned}\mathcal{L}_{init}=&\alpha_{ap}(\mathcal{L}_{ap}^{l}+\mathcal{L}_{ap}^{r})+\alpha_{ds}(\mathcal{L}_{ds}^{l}+\mathcal{L}_{ds}^{r})\\&+\alpha_{ps}(\mathcal{L}_{ps}^{l}+\mathcal{L}_{ps}^{r})\end{aligned}$$
$$\mathcal{L}_{ref}=\alpha_{ap}\mathcal{L}_{ap}^{l}+\alpha_{ds}\mathcal{L}_{ds}^{l}+\alpha_{ps}\mathcal{L}_{ps}^{l}$$
(3)
where $\mathcal{L}_ap$ is an image reconstruction loss, $\mathcal{L}_ds$ is a smoothness term and $\mathcal{L}_ps$ is a proxy-supervised loss. Each term contains both the left and right components for the initial disparity estimator, and the left components only for the refinement stage.
Image reconstruction loss. A linear combination of $l_1$ loss and structural similarity measure (SSIM) [54] encodes the quality of the reconstructed image $\tilde{I}$ with respect to the original image $I:$
$$\mathcal{L}_{ap}=\frac{1}{N}\sum_{i,j}\alpha\frac{1-SSIM(I_{ij},\hat{I}_{ij})}{2}+(1-\alpha)|I_{ij}-\hat{I}_{ij}|$$
Following [11], we set $\alpha=0.85$ and use a SSIM with
$3\times3$ block filter.
Disparity smoothness loss. This cost encourages the predicted disparity to be locally smooth. Disparity gradients are weighted by an edge-aware term from image domain:
(5)
$$\begin{aligned}\mathcal{L}_{ds}&=\frac{1}{N}\sum_{i,j}|\partial_xd_{ij}|e^{-|\partial_xI_{ij}|}+|\partial_yd_{ij}|e^{-|\partial_yI_{ij}|}\end{aligned}$$
Proxy-supervised loss. Given the proxy disparity maps obtained by a conventional stereo algorithm, detailed in Section 4, we coach the network using reverse Huber (berHu) loss [36]:
(6)
$$\mathcal{L}_{ps}=\frac{1}{N}\sum_{i,j}berHu(d_{ij},d_{ij}^{st},c)$$
$$berHu(d_{ij},d_{ij}^{st},c)=\begin{cases}|d_{ij}-d_{ij}^{st}|&\text{if}|d_{ij}-d_{ij}^{st}|\leq c\\\frac{|d_{ij}-d_{ij}^{st}|^2-c^2}{2c}&\text{otherwise}\end{cases}$$
(7)
where $d_ij$ and $d_ij^{st}$ are, respectively, the predicted dispar-
ity and the proxy annotation for pixel at the coordinates $i,j$ of the image, while $c$ is adaptively set as $\alpha\max_i,j|d_{ij}-d_{ij}^{st}|$, with $\alpha=0.2.$
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50341796875,
0.09344482421875,
0.60888671875,
0.09344482421875,
0.60888671875,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.5986328125,
0.1307373046875,
0.6396484375,
0.1307373046875,
0.6396484375,
0.1427001953125,
0.5986328125,
0.1427001953125
],
[
0.869140625,
0.1300048828125,
0.8935546875,
0.1300048828125,
0.8935546875,
0.143310546875,
0.869140625,
0.143310546875
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.6640625,
0.147705078125,
0.6806640625,
0.147705078125,
0.6806640625,
0.1558837890625,
0.6640625,
0.1558837890625
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.52294921875,
0.160400390625,
0.89208984375,
0.160400390625,
0.89208984375,
0.173583984375,
0.52294921875,
0.173583984375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1767578125,
0.89208984375,
0.1767578125,
0.89208984375,
0.1900634765625,
0.50244140625,
0.1900634765625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1912841796875,
0.7646484375,
0.1900634765625,
0.7646484375,
0.2052001953125,
0.50244140625,
0.205810546875
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.459228515625,
0.21337890625,
0.459228515625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.55224609375,
0.2259521484375,
0.84228515625,
0.2266845703125,
0.84228515625,
0.2442626953125,
0.55224609375,
0.24365234375
],
[
0.08087158203125,
0.238037109375,
0.27783203125,
0.238037109375,
0.27783203125,
0.25244140625,
0.08087158203125,
0.25244140625
],
[
0.869140625,
0.2392578125,
0.89208984375,
0.2392578125,
0.89208984375,
0.25244140625,
0.869140625,
0.25244140625
],
[
0.60693359375,
0.248779296875,
0.73388671875,
0.248779296875,
0.73388671875,
0.266357421875,
0.60693359375,
0.266357421875
],
[
0.098876953125,
0.261962890625,
0.468017578125,
0.261962890625,
0.468017578125,
0.27587890625,
0.098876953125,
0.27587890625
],
[
0.08087158203125,
0.27783203125,
0.468017578125,
0.27783203125,
0.468017578125,
0.291748046875,
0.08087158203125,
0.291748046875
],
[
0.57421875,
0.27978515625,
0.81884765625,
0.27978515625,
0.81884765625,
0.29736328125,
0.57421875,
0.29736328125
],
[
0.869140625,
0.281005859375,
0.89208984375,
0.281005859375,
0.89208984375,
0.294189453125,
0.869140625,
0.294189453125
],
[
0.080078125,
0.29345703125,
0.468017578125,
0.29345703125,
0.468017578125,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.521484375,
0.2998046875,
0.89208984375,
0.300537109375,
0.89208984375,
0.31494140625,
0.521484375,
0.314453125
],
[
0.080078125,
0.306884765625,
0.46728515625,
0.306884765625,
0.46728515625,
0.3212890625,
0.080078125,
0.3212890625
],
[
0.5009765625,
0.31640625,
0.89208984375,
0.31494140625,
0.89208984375,
0.330078125,
0.5009765625,
0.330810546875
],
[
0.080078125,
0.323974609375,
0.4697265625,
0.323974609375,
0.4697265625,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.07843017578125,
0.337158203125,
0.468017578125,
0.33837890625,
0.468017578125,
0.352783203125,
0.07843017578125,
0.3515625
],
[
0.50341796875,
0.333251953125,
0.8896484375,
0.333251953125,
0.8896484375,
0.343505859375,
0.50341796875,
0.343505859375
],
[
0.50244140625,
0.3466796875,
0.89208984375,
0.3466796875,
0.89208984375,
0.35986328125,
0.50244140625,
0.35986328125
],
[
0.08087158203125,
0.352783203125,
0.468017578125,
0.352783203125,
0.468017578125,
0.3662109375,
0.08087158203125,
0.3662109375
],
[
0.50244140625,
0.3623046875,
0.61669921875,
0.3623046875,
0.61669921875,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.080078125,
0.36865234375,
0.46728515625,
0.36865234375,
0.46728515625,
0.3818359375,
0.080078125,
0.3818359375
],
[
0.07843017578125,
0.3818359375,
0.468017578125,
0.38330078125,
0.468017578125,
0.397705078125,
0.07843017578125,
0.396484375
],
[
0.521484375,
0.37890625,
0.88818359375,
0.37890625,
0.88818359375,
0.393310546875,
0.521484375,
0.393310546875
],
[
0.080078125,
0.3984375,
0.468017578125,
0.3984375,
0.468017578125,
0.412353515625,
0.080078125,
0.412353515625
],
[
0.5009765625,
0.393310546875,
0.89208984375,
0.39453125,
0.89208984375,
0.4091796875,
0.5009765625,
0.40771484375
],
[
0.080078125,
0.4140625,
0.46728515625,
0.4140625,
0.46728515625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.5009765625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.08087158203125,
0.4287109375,
0.468017578125,
0.4287109375,
0.468017578125,
0.443115234375,
0.08087158203125,
0.443115234375
],
[
0.50244140625,
0.425537109375,
0.61767578125,
0.425537109375,
0.61767578125,
0.439453125,
0.50244140625,
0.439453125
],
[
0.080078125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.458251953125,
0.080078125,
0.458251953125
],
[
0.07843017578125,
0.458251953125,
0.468017578125,
0.458984375,
0.468017578125,
0.473388671875,
0.07843017578125,
0.472900390625
],
[
0.56640625,
0.455810546875,
0.8935546875,
0.455078125,
0.8935546875,
0.4912109375,
0.56640625,
0.492431640625
],
[
0.5087890625,
0.472900390625,
0.5390625,
0.472900390625,
0.5390625,
0.484130859375,
0.5087890625,
0.484130859375
],
[
0.56396484375,
0.4697265625,
0.576171875,
0.4697265625,
0.576171875,
0.473388671875,
0.56396484375,
0.473388671875
],
[
0.08087158203125,
0.47607421875,
0.465576171875,
0.47607421875,
0.465576171875,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.55615234375,
0.481689453125,
0.57275390625,
0.479248046875,
0.576171875,
0.486083984375,
0.5595703125,
0.4892578125
],
[
0.6748046875,
0.481689453125,
0.68896484375,
0.481689453125,
0.68896484375,
0.49365234375,
0.6748046875,
0.49365234375
],
[
0.080078125,
0.4892578125,
0.468017578125,
0.4892578125,
0.468017578125,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.580078125,
0.488037109375,
0.5986328125,
0.488037109375,
0.5986328125,
0.50048828125,
0.580078125,
0.50048828125
],
[
0.08087158203125,
0.505859375,
0.46728515625,
0.505859375,
0.46728515625,
0.5166015625,
0.08087158203125,
0.5166015625
],
[
0.869140625,
0.5,
0.89208984375,
0.5,
0.89208984375,
0.51220703125,
0.869140625,
0.51220703125
],
[
0.51953125,
0.51220703125,
0.89208984375,
0.51318359375,
0.89208984375,
0.52783203125,
0.51953125,
0.5263671875
],
[
0.080078125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.07843017578125,
0.53271484375,
0.44677734375,
0.5341796875,
0.44677734375,
0.548828125,
0.07843017578125,
0.54736328125
],
[
0.50244140625,
0.52880859375,
0.62353515625,
0.52880859375,
0.62353515625,
0.54248046875,
0.50244140625,
0.54248046875
],
[
0.1004638671875,
0.5498046875,
0.46728515625,
0.5498046875,
0.46728515625,
0.56298828125,
0.1004638671875,
0.56298828125
],
[
0.521484375,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.521484375,
0.56103515625
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50341796875,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.080078125,
0.58056640625,
0.468017578125,
0.58056640625,
0.468017578125,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.50244140625,
0.5771484375,
0.8759765625,
0.5771484375,
0.8759765625,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.080078125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.080078125,
0.609375
],
[
0.08087158203125,
0.6123046875,
0.465576171875,
0.6123046875,
0.465576171875,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.529296875,
0.6171875,
0.56884765625,
0.62255859375,
0.56640625,
0.63720703125,
0.52685546875,
0.6318359375
],
[
0.60546875,
0.61572265625,
0.85107421875,
0.61279296875,
0.85107421875,
0.63525390625,
0.60546875,
0.63720703125
],
[
0.6044921875,
0.62353515625,
0.61376953125,
0.62353515625,
0.61376953125,
0.62890625,
0.6044921875,
0.62890625
],
[
0.87109375,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63330078125,
0.87109375,
0.63330078125
],
[
0.56396484375,
0.625,
0.56787109375,
0.625,
0.56787109375,
0.6298828125,
0.56396484375,
0.6298828125
],
[
0.080078125,
0.62548828125,
0.46728515625,
0.62548828125,
0.46728515625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.57861328125,
0.62548828125,
0.59326171875,
0.630859375,
0.5888671875,
0.64013671875,
0.57275390625,
0.63525390625
],
[
0.080078125,
0.64013671875,
0.46728515625,
0.64013671875,
0.46728515625,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.080078125,
0.65576171875,
0.46728515625,
0.65576171875,
0.46728515625,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.52294921875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.52294921875,
0.673828125
],
[
0.080078125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.080078125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70263671875,
0.17236328125,
0.70263671875,
0.17236328125,
0.7138671875,
0.080078125,
0.7138671875
],
[
0.50341796875,
0.70556640625,
0.62353515625,
0.70556640625,
0.62353515625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.080078125,
0.7265625,
0.221435546875,
0.7265625,
0.221435546875,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.63818359375,
0.72998046875,
0.8056640625,
0.736328125,
0.8037109375,
0.7568359375,
0.63623046875,
0.75048828125
],
[
0.59228515625,
0.7392578125,
0.63232421875,
0.7392578125,
0.63232421875,
0.75390625,
0.59228515625,
0.75390625
],
[
0.86767578125,
0.7373046875,
0.89453125,
0.7373046875,
0.89453125,
0.75390625,
0.86767578125,
0.75390625
],
[
0.098876953125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.098876953125,
0.7646484375
],
[
0.63134765625,
0.74853515625,
0.6611328125,
0.74853515625,
0.6611328125,
0.7568359375,
0.63134765625,
0.7568359375
],
[
0.66015625,
0.75244140625,
0.68115234375,
0.7568359375,
0.67724609375,
0.76953125,
0.65625,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.77978515625,
0.4697265625,
0.79443359375,
0.080078125,
0.79541015625
],
[
0.66015625,
0.7861328125,
0.736328125,
0.7841796875,
0.736328125,
0.7998046875,
0.66015625,
0.8017578125
],
[
0.76318359375,
0.7841796875,
0.8837890625,
0.7861328125,
0.8837890625,
0.8037109375,
0.76318359375,
0.8017578125
],
[
0.08087158203125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.5087890625,
0.79443359375,
0.6337890625,
0.79736328125,
0.6337890625,
0.8125,
0.5087890625,
0.8095703125
],
[
0.66650390625,
0.8017578125,
0.73388671875,
0.798828125,
0.734375,
0.81298828125,
0.66845703125,
0.81640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.76318359375,
0.806640625,
0.8349609375,
0.806640625,
0.8349609375,
0.8212890625,
0.76318359375,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.87109375,
0.82373046875,
0.89208984375,
0.82373046875,
0.89208984375,
0.8359375,
0.87109375,
0.8359375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.521484375,
0.8369140625,
0.890625,
0.8369140625,
0.890625,
0.8515625,
0.521484375,
0.8515625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85400390625,
0.89208984375,
0.85400390625,
0.89208984375,
0.8671875,
0.50244140625,
0.8671875
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.86865234375,
0.890625,
0.86865234375,
0.890625,
0.88330078125,
0.50244140625,
0.88330078125
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88525390625,
0.59326171875,
0.88525390625,
0.59326171875,
0.89599609375,
0.50341796875,
0.89599609375
]
] |
[
"sary information required to compute a cost volume since",
"training loss as:",
"it processes a single input image. Thus, no 1-D correla-",
"tionlayercanbeimposedtoencodegeometricalconstraints",
"Ltotal",
"(1)",
"in this stage of our network. Then, upsampling modules",
"=1",
"aredeployedtoenrichfeaturerepresentationsthrough skip-",
"where s indicates the output resolution, n; and nr the",
"connections and to extract two disparity maps, aligned re-",
"numbers of considered scales during loss computation,",
"spectively with the input frame and a virtual viewpoint on",
"while Cinit and Cref are formalised as:",
"its right as in [11]. This process is carried out at each scale",
"using 1-stride convolutional layers with kernel size 3 × 3.",
"Cinit =Qap(Cap + Cap) + ads(Cds + Cds)",
"3.3.DisparityRefinement",
"(2)",
"+ aps(Cps + Cps)",
"Given an initial estimate of the disparity at each scale",
"obtained in the second part of the network, often character-",
"Lref =QapCap+adsCds+ apsCps",
"(3)",
"ized by errors at depth discontinuities and occluded regions,",
"where Lap is an image reconstruction loss, Lds is a",
"this stage predicts corresponding multi-scale residual sig-",
"smoothness term and Cps is a proxy-supervised loss. Each",
"nals [14] by a few stacked nonlinear layers that are then",
"used to compute the final left-view aligned disparity map.",
"termcontainsboththeleftandrightcomponentsfortheini-",
"tial disparity estimator, and the left components only for the",
"This strategy allows us to simplify the end-to-end learning",
"refinement stage.",
"process of the entire network. Moreover, motivated by [30]",
"we believe that geometrical constraints can play a central",
"Image reconstruction loss. A linear combination of li",
"role in boosting the final depth accuracy. For this reason,",
"loss and structural similarity measure (SSIM) [54]encodes",
"we embed matching costs in feature space computed em-",
"the quality of the reconstructed image I with respect to the",
"ploying a horizontal correlation layer, typically deployed in",
"original image I:",
"deep stereo algorithms: To this end, we rely on the right-",
"view disparity map computed previously to generate right-",
"Za1-SIM()+(1-0)g -igl",
"Cap",
"",
"viewfeaturesFpfromtheleftonesFousinga differen-",
"",
"2",
"tiable bilinear sampler [16]. The network is also fed with",
"i.3",
"error eL,i.e.the absolute difference between left and vir-",
"(4)",
"Following [11], we set α = 0.85 and use a SSIM with",
"tual right features at input resolution, with the latter back-",
"warped at the same coordinates of the former, as in [24]",
"3 × 3 block filter.",
"We point out once more that, differently from [30],",
"Disparity smoothness loss. This cost encourages the",
"our architecture produces both a synthetic right view, i.e.",
"predicted disparity to be locally smooth. Disparity gradients",
"its features representation, and computes the final dispar-",
"are weighted by an edge-aware term from image domain",
"ity map following stereo rationale.This makes monoRes-",
"Matchasingleend-to-endarchitecture,effectivelyperform-",
"Cds",
"10adijle-10a lsl + [0yda;le-lo,lgl",
"",
"(5)",
"",
"ing stereo out of a single input view rather than the combi-",
"N",
"nation of two models (i.e., Deep3D [55] and DispNetC [34]",
"for the two tasks outlined) trained independently as in [31].",
"Proxy-supervised loss. Given the proxy disparity maps",
"Moreover, exhaustive experiments will highlight the supe-",
"obtained by a conventional stereo algorithm, detailed in",
"rior accuracy achieved by our fully self-supervised, end-to-",
"Section 4, we coach the network using reverse Huber",
"end approach.",
"(berHu) loss [36]:",
"3.4. Training Loss",
"berHu(di,d,c)",
"Cps =",
"(6)",
"In order to train our multi-stage architecture, we de-",
"N",
"i.3",
"fine the total loss as a sum of two main contributions, a",
"Cinit term from the initial disparity estimation module and",
"Idij-dt]",
"if dij-d≤c",
"a Lref term from the disparity refinement stage. Follow-",
"berHu(dij, dt,c)",
"Idig-dai",
"ing [12], we embrace the idea to up-sample the predicted",
"otherwise",
"low-resolution disparity maps to the full input resolution",
"(7)",
"and then compute the corresponding signals. This simple",
"where di, and dst are, respectively, the predicted dispar-",
"strategy is designed to force the inverse depth estimation",
"ity and the proxy annotation for pixel at the coordinates i, j",
"to reproduce the same objective at each scale, thus leading",
"of the image, while cis adaptively set as Q maxi,j |dij-dstl,",
"to much better outcomes. In particular, we obtain the final",
"withα=0.2"
] |
{
"arxiv_id": "1904.04144",
"language": "en",
"page_idx": 3
}
|
||
1904.04144-en-4
|
1904.04144-en-4.jpg
|
<FigureHere>
(a)
(b)
(c)
Figure 3. Examples of proxy labels computed by SGM. Given the source image (a), the network exploits the SGM supervision filtered with left-right consistency check (b) in order to train monoResMatch to estimate the final disparity map (c). No post-processing from [11] is performed on (c) in this example.
## $4. \textbf{Proxy labels distillation}$
labels. Figure 3 shows an example of distilled labels (b), where black pixels correspond to outliers filtered out by left-right consistency. Although some of them persist, we can notice how they do not affect the final prediction by the trained network and how our proposal can recover accurate disparity values in occluded regions on the left side of the $\operatorname{image}($c).
To generate accurate proxy labels, we use the popular SGM algorithm [15], a fast yet effective solution to infer depth from a rectified stereo pair without training. In our implementation, initial matching costs are computed for each pixel $p$ and disparity hypothesis $d$ applying a 9×7 census transform and computing Hamming distance on pixel strings. Then, scanline optimization along eight different paths refines the initial cost volume as follows:
## 5. Experimental results
$$\begin{aligned}E(p,d)=&C(p,d)+\operatorname*{min}_{j>1}[C(p^{\prime},d),C(p^{\prime},d\pm1)+P1,\\&C(p^{\prime},d\pm q)+P2]-\operatorname*{min}_{k<D_{max}}(C(p^{\prime},k))\end{aligned}$$
(8)
In this section, we describe the datasets, implementation details and then present exhaustive evaluations of monoResMatch on various training/testing configurations, showing that our proposal consistently outperforms self-supervised state-of-the-art approaches. As standard in this field, we assess the performance of monocular depth estimation techniques following the protocol by Eigen et al.[6], extracting data from the KITTI [9] dataset, using sparse LiDAR measurements as ground truth for evaluation. Additionally, we also perform an exhaustive ablation study proving that proxy supervision from SGM algorithm and effective architectural choices enable our strategy to improve predicted depth map accuracy by a large margin.
being $C(p,d)$ the matching cost for pixel $p$ and hypothesis $d,P_1$ and $P_2$ two smoothness penalties, discouraging disparity gaps between $p$ and previous pixel $p^\prime$ along the scanline path. The final disparity map $D$ is obtained applying a winner-takes-all strategy to each pixel of the reference image. Although SGM generates quite accurate disparity labels, outliers may affect the training of a depth model negatively, as noticed by Tonioni et $al.[49].$ They applied a learned confidence measure [41] to filter out erroneous labels when computing the loss. Differently, we run a nonlearning based left-right consistency check to detect outliers. Purposely, by extracting both disparity maps $D^L$ and $D^{R}$ with SGM, respectively for the left and right images, we apply the following criteria to invalidate (i.e., set to -1) pixels having different disparities across the two maps:
## 5.1. Datasets
For all our experiments we compute standard monocular metrics [6, 11]: Abs rel, Sq rel, RMSE and $RMSE\textit{log}$ represent error measures while $\delta<\zeta$ the percentage of predictions whose maximum between ratio and inverse ratio with respect to the ground truth is lower than a threshold $\zeta.$ Two main datasets are involved in our evaluation, that are KITTI [9] and CityScapes [4].
KITTI. The KITTI stereo dataset[9] is a collection of rectified stereo pairs made up of 61 scenes (containing about 42,382 stereo frames) mainly concerned with driving scenarios. Predominant image size is $1242\times375$ pixels. A LiDAR device, mounted and calibrated in proximity to the left camera, was deployed to measure depth information.
Following other works [6, 11], we divided the overall dataset into two subsets, composed respectively of 29 and
$$D(p)=\begin{cases}D(p)&\text{if}|D^L(p)-D^R(p-D^L(p))|\leq\varepsilon\\-1&\text{otherwise}\end{cases}$$
(9)
The left-right consistency check is a simple strategy that removes many wrong disparity assignments, mostly near depth discontinuities, without needing any training that would be required by [49]. Therefore, our proxy labels generation process does not rely at all on ground truth depth
|
[
[
0.2296142578125,
0.21337890625,
0.250732421875,
0.21337890625,
0.250732421875,
0.2259521484375,
0.2296142578125,
0.2259521484375
],
[
0.4755859375,
0.21337890625,
0.496826171875,
0.21337890625,
0.496826171875,
0.2259521484375,
0.4755859375,
0.2259521484375
],
[
0.72216796875,
0.2139892578125,
0.74267578125,
0.2139892578125,
0.74267578125,
0.2247314453125,
0.72216796875,
0.2247314453125
],
[
0.08087158203125,
0.2259521484375,
0.890625,
0.2259521484375,
0.890625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.08087158203125,
0.242431640625,
0.890625,
0.242431640625,
0.890625,
0.25244140625,
0.08087158203125,
0.25244140625
],
[
0.07843017578125,
0.255126953125,
0.279296875,
0.255126953125,
0.279296875,
0.268310546875,
0.07843017578125,
0.268310546875
],
[
0.08087158203125,
0.29345703125,
0.29736328125,
0.29345703125,
0.29736328125,
0.306884765625,
0.08087158203125,
0.306884765625
],
[
0.50244140625,
0.29345703125,
0.890625,
0.29345703125,
0.890625,
0.30810546875,
0.50244140625,
0.30810546875
],
[
0.50341796875,
0.31005859375,
0.890625,
0.31005859375,
0.890625,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.098876953125,
0.31640625,
0.4697265625,
0.317626953125,
0.4697265625,
0.33203125,
0.098876953125,
0.330810546875
],
[
0.50341796875,
0.32568359375,
0.890625,
0.32568359375,
0.890625,
0.3359375,
0.50341796875,
0.3359375
],
[
0.08087158203125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.08087158203125,
0.348388671875,
0.4697265625,
0.348388671875,
0.4697265625,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.50244140625,
0.354736328125,
0.89208984375,
0.354736328125,
0.89208984375,
0.36865234375,
0.50244140625,
0.36865234375
],
[
0.080078125,
0.363037109375,
0.4697265625,
0.363037109375,
0.4697265625,
0.376953125,
0.080078125,
0.376953125
],
[
0.50244140625,
0.369384765625,
0.89208984375,
0.369384765625,
0.89208984375,
0.3837890625,
0.50244140625,
0.3837890625
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.080078125,
0.392578125
],
[
0.50341796875,
0.386474609375,
0.5703125,
0.386474609375,
0.5703125,
0.397705078125,
0.50341796875,
0.397705078125
],
[
0.080078125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50341796875,
0.415283203125,
0.69775390625,
0.415283203125,
0.69775390625,
0.429931640625,
0.50341796875,
0.429931640625
],
[
0.07843017578125,
0.423583984375,
0.390625,
0.42236328125,
0.390625,
0.4375,
0.07843017578125,
0.438232421875
],
[
0.521484375,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455078125,
0.521484375,
0.455078125
],
[
0.09722900390625,
0.458984375,
0.451904296875,
0.458984375,
0.451904296875,
0.473388671875,
0.09722900390625,
0.473388671875
],
[
0.5048828125,
0.458251953125,
0.8896484375,
0.458251953125,
0.8896484375,
0.468505859375,
0.5048828125,
0.468505859375
],
[
0.236083984375,
0.473388671875,
0.2484130859375,
0.4716796875,
0.25,
0.4765625,
0.23779296875,
0.479248046875
],
[
0.5009765625,
0.47021484375,
0.89208984375,
0.4716796875,
0.89208984375,
0.486083984375,
0.5009765625,
0.48486328125
],
[
0.16015625,
0.484130859375,
0.4306640625,
0.484130859375,
0.4306640625,
0.505859375,
0.16015625,
0.505859375
],
[
0.50341796875,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.50341796875,
0.50048828125
],
[
0.50244140625,
0.501953125,
0.88818359375,
0.50048828125,
0.88818359375,
0.51513671875,
0.50244140625,
0.5166015625
],
[
0.44677734375,
0.5087890625,
0.4697265625,
0.5087890625,
0.4697265625,
0.521484375,
0.44677734375,
0.521484375
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.080078125,
0.53271484375,
0.468017578125,
0.53271484375,
0.468017578125,
0.546875,
0.080078125,
0.546875
],
[
0.50341796875,
0.53271484375,
0.8896484375,
0.53271484375,
0.8896484375,
0.546875,
0.50341796875,
0.546875
],
[
0.08087158203125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.08087158203125,
0.5625
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.08087158203125,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.57763671875,
0.08087158203125,
0.57763671875
],
[
0.50244140625,
0.56298828125,
0.890625,
0.56298828125,
0.890625,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.080078125,
0.5791015625,
0.471435546875,
0.5791015625,
0.471435546875,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.57763671875,
0.89208984375,
0.57763671875,
0.89208984375,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.08087158203125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.50244140625,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.080078125,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50341796875,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.080078125,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.62353515625,
0.75732421875,
0.62353515625,
0.75732421875,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.07843017578125,
0.63720703125,
0.4697265625,
0.63818359375,
0.4697265625,
0.65283203125,
0.07843017578125,
0.6513671875
],
[
0.08087158203125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.50341796875,
0.6494140625,
0.5986328125,
0.6494140625,
0.5986328125,
0.6611328125,
0.50341796875,
0.6611328125
],
[
0.080078125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.080078125,
0.681640625
],
[
0.521484375,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.521484375,
0.6875
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.08087158203125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.08087158203125,
0.7119140625
],
[
0.5009765625,
0.70263671875,
0.890625,
0.7041015625,
0.890625,
0.71826171875,
0.5009765625,
0.71728515625
],
[
0.080078125,
0.7119140625,
0.468017578125,
0.7138671875,
0.468017578125,
0.72900390625,
0.080078125,
0.7265625
],
[
0.50341796875,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.73095703125,
0.50341796875,
0.73095703125
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72998046875,
0.468017578125,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.50244140625,
0.734375,
0.890625,
0.73291015625,
0.890625,
0.74755859375,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.74560546875,
0.441162109375,
0.74560546875,
0.441162109375,
0.7587890625,
0.080078125,
0.7587890625
],
[
0.5048828125,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.76123046875,
0.5048828125,
0.76123046875
],
[
0.50244140625,
0.76220703125,
0.70361328125,
0.7646484375,
0.70361328125,
0.77880859375,
0.50244140625,
0.7763671875
],
[
0.1519775390625,
0.77978515625,
0.44287109375,
0.78076171875,
0.44287109375,
0.798828125,
0.1519775390625,
0.79736328125
],
[
0.52294921875,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.52294921875,
0.79345703125
],
[
0.08905029296875,
0.79150390625,
0.131591796875,
0.79150390625,
0.131591796875,
0.80615234375,
0.08905029296875,
0.80615234375
],
[
0.4345703125,
0.79248046875,
0.46728515625,
0.79248046875,
0.46728515625,
0.8037109375,
0.4345703125,
0.8037109375
],
[
0.5009765625,
0.79345703125,
0.89208984375,
0.79443359375,
0.89208984375,
0.8095703125,
0.5009765625,
0.80810546875
],
[
0.17236328125,
0.80615234375,
0.178955078125,
0.80615234375,
0.178955078125,
0.81103515625,
0.17236328125,
0.81103515625
],
[
0.20751953125,
0.8017578125,
0.275390625,
0.8017578125,
0.275390625,
0.81298828125,
0.20751953125,
0.81298828125
],
[
0.50341796875,
0.81005859375,
0.8896484375,
0.81005859375,
0.8896484375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.09722900390625,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.09722900390625,
0.8388671875
],
[
0.5009765625,
0.82470703125,
0.89208984375,
0.82373046875,
0.89208984375,
0.83837890625,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84033203125,
0.89208984375,
0.84033203125,
0.89208984375,
0.85400390625,
0.50244140625,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85595703125,
0.875,
0.85595703125,
0.875,
0.869140625,
0.50244140625,
0.869140625
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.50244140625,
0.90087890625
]
] |
[
"(a)",
"(b)",
"(c)",
"Figure 3. Examples of proxy labels computed by SGM. Given the source image (a), the network exploits the SGM supervision filtered with",
"left-rightconsistencycheck(b)in ordertotrainmonoResMatchto estimatethefinal disparitymap(c).Nopost-processingfrom1lis",
"performed on (c) in this example.",
"4.Proxy labels distillation",
"labels. Figure 3 shows an example of distilled labels (b),",
"where black pixels correspond to outliers filtered out by",
"To generate accurate proxy labels, we use the popular",
"left-rightconsistency.Althoughsomeofthempersist,we",
"SGM algorithm [15], a fast yet effective solution to in-",
"can notice how they do not affect the final prediction by the",
"fer depth from a rectified stereo pair without training. In",
"trained network and how our proposal can recover accurate",
"our implementation, initial matching costs are computed for",
"disparity values in occluded regions on the left side of the",
"each pixel p and disparity hypothesis d applying a 9 × 7 cen-",
"image (c).",
"sus transform and computing Hamming distance on pixel",
"strings. Then, scanline optimization along eight different",
"5.Experimentalresults",
"paths refines the initial cost volume as follows:",
"In this section, we describe the datasets, implementation",
"E(p, d) =C(p, d) + min[C(p, d),C(p', d±1) + P1,",
"detailsandthenpresentexhaustiveevaluationsofmonoRes-",
"21",
"Match on various training/testing configurations, showing",
"C(p,d ±q) + P2) -kmin (C(p,k)",
"that our proposal consistently outperforms self-supervised",
"state-of-the-art approaches. As standard in this field, we as",
"(8)",
"sess the performance of monocular depth estimation tech-",
"being C(p, d) the matching cost for pixel p and hypothesis",
"niques following the protocol by Eigen et al. [6], extract-",
"d, Pi and P2 two smoothness penalties, discouraging dis-",
"ing data from the KITTI [9] dataset, using sparse LiDAR",
"parity gaps between p and previous pixel p' along the scan-",
"measurements as ground truth for evaluation. Additionally",
"line path. The final disparity map D is obtained applying a",
"we also perform an exhaustive ablation study proving that",
"winner-takes-all strategy to each pixel of the reference im-",
"proxy supervision from SGM algorithm and effective ar-",
"age. Although SGM generates quite accurate disparity la-",
"chitectural choices enable our strategy to improve predicted",
"bels, outliers may affect the training of a depth model neg-",
"depth map accuracy by a large margin.",
"atively, as noticed by Tonioni et al. [49]. They applied a",
"learned confidencemeasure[4lltofilter outerroneousla-",
"5.1. Datasets",
"bels when computing the loss. Differently, we run a non-",
"For all our experiments we compute standard monocu-",
"learning based left-right consistency check to detect out-",
"lar metrics [6, 1l]: Abs rel, Sq rel, RMSE and RMSE log",
"liers.Purposely,by extracting both disparity maps DL and",
"represent error measures while < the percentage of pre-",
"DR with SGM, respectively for the left and right images,",
"dictions whose maximum between ratio and inverse ratio",
"we apply the following criteria to invalidate (i.e., set to -1)",
"with respect to the ground truth is lower than a threshold C.",
"pixels having different disparities across the two maps:",
"Two main datasets are involved in our evaluation,that are",
"KITTI [9] and CityScapes [4].",
"D(p)if [DL(p) -DR(p-DL(p)/ ≤",
"KITTI. The KITTI stereo dataset [9] is a collection of",
"D(p) :",
"（(9)",
"rectified stereo pairs madeup of 61 scenes (containing about",
"",
"otherwise",
"42,382 stereo frames) mainly concerned with driving sce-",
"The left-right consistency check is a simple strategy",
"narios.Predominant image size is 1242 × 375 pixels.A",
"that removes many wrong disparity assignments, mostly",
"LiDAR device, mounted and calibrated in proximity to the",
"near depth discontinuities, without needing any training that",
"left camera, was deployed to measure depth information.",
"would be required by [49]. Therefore, our proxy labels gen-",
"Following other works [6, 11], we divided the overall",
"eration process does not rely at all on ground truth depth",
"dataset into two subsets, composed respectively of 29 and"
] |
{
"arxiv_id": "1904.04144",
"language": "en",
"page_idx": 4
}
|
||
1811.05550-en-0
|
1811.05550-en-0.jpg
|
# $\textbf{Neural Wavetable: a playable wavetable synthesizer}$ using neural networks
Lamtharn Hantrakul* Li-Chia Yang
Center for Music Technology Georgia Institute of Technology
Georgia, USA 30318
[lhantrakul3][richard40148]@gatech. edu
## Abstract
We present Neural Wavetable, a proof-of-concept wavetable synthesizer that uses neural networks to generate playable wavetables. The system can produce new, distinct waveforms through the interpolation of traditional wavetables in an autoencoder's latent space. It is available as a VST/AU plugin for use in a Digital Audio Workstation.
## 1 Introduction
WaveNet, a deep neural network architecture for speech and audio generation, has brought exciting innovations in the sound domain since its introduction in 2016[2]. In the musical domain, WaveNet has been used as an auto-encoder capable of interpolating- or non-linear “blending”- different instrument sounds through NSynth [1].
Inspired by this approach, we present a proof-of-concept system addressing a smaller scale audio task: generating short wavetables. Rather than generate long segments of speech and conditioning the network on long-term periodicity, we focus on generating short wavetables of length of 512 samples; a contrast to the tens of thousands of samples required even in just a second of audio. These short wavetables are fed into a real-time wavetable synthesizer implemented in C++ and the JUCE framework. The source-code and VST/AU plug-in are available for public download$^{2}$
# 2 Background
Wavetable synthesis is an industry-standard approach which uses a bank of “wavetables”of lengths between 512-2048 and sample-rate conversion techniques to efficiently generate musical sounds. In context of this work, wavetable synthesis efficiently handles real-time generation of different pitches given a single wavetable generated by a neural network. For brevity, readers should reference the original works for a complete description of Nsynth [1], WaveNet [2] and Wavetable synthesis [3].
## 3 Implementation Details
3.1 WaveNet as a wavetable autoencoder
In this proof-of-concept, we use the pre-trained model NSynth WaveNet model as-is [1]. Given the task of generating a short wavetable is smaller than NSynth's original task of generating several
Currently an AI Resident with Google Brain
$^{2}$Neural Wavetable. https://github.com/RichardYang40148/Neural\_Wavetable\_Synthesizer
32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada
|
[
[
0.1846923828125,
0.125,
0.81396484375,
0.1256103515625,
0.81396484375,
0.146484375,
0.1846923828125,
0.1458740234375
],
[
0.366943359375,
0.154052734375,
0.63232421875,
0.154052734375,
0.63232421875,
0.1717529296875,
0.366943359375,
0.1717529296875
],
[
0.369384765625,
0.2266845703125,
0.6337890625,
0.2266845703125,
0.6337890625,
0.2412109375,
0.369384765625,
0.2412109375
],
[
0.40185546875,
0.242431640625,
0.59619140625,
0.242431640625,
0.59619140625,
0.255615234375,
0.40185546875,
0.255615234375
],
[
0.0228729248046875,
0.25830078125,
0.05963134765625,
0.25830078125,
0.05963134765625,
0.39453125,
0.0228729248046875,
0.39453125
],
[
0.39208984375,
0.25244140625,
0.60546875,
0.25390625,
0.60546875,
0.271484375,
0.39208984375,
0.270751953125
],
[
0.428955078125,
0.26953125,
0.5703125,
0.26953125,
0.5703125,
0.28271484375,
0.428955078125,
0.28271484375
],
[
0.338134765625,
0.2822265625,
0.6640625,
0.28271484375,
0.6640625,
0.29736328125,
0.338134765625,
0.296630859375
],
[
0.4599609375,
0.33203125,
0.5400390625,
0.33203125,
0.5400390625,
0.349853515625,
0.4599609375,
0.349853515625
],
[
0.2344970703125,
0.365478515625,
0.765625,
0.365478515625,
0.765625,
0.37890625,
0.2344970703125,
0.37890625
],
[
0.236083984375,
0.380126953125,
0.765625,
0.380126953125,
0.765625,
0.393310546875,
0.236083984375,
0.393310546875
],
[
0.233642578125,
0.392578125,
0.7646484375,
0.392578125,
0.7646484375,
0.406005859375,
0.233642578125,
0.406005859375
],
[
0.0196075439453125,
0.404052734375,
0.05963134765625,
0.40283203125,
0.062103271484375,
0.48486328125,
0.0228729248046875,
0.486083984375
],
[
0.2344970703125,
0.4072265625,
0.765625,
0.4072265625,
0.765625,
0.42041015625,
0.2344970703125,
0.42041015625
],
[
0.2344970703125,
0.421630859375,
0.31787109375,
0.421630859375,
0.31787109375,
0.43310546875,
0.2344970703125,
0.43310546875
],
[
0.17236328125,
0.460205078125,
0.31201171875,
0.4609375,
0.31201171875,
0.47607421875,
0.17236328125,
0.474853515625
],
[
0.0236968994140625,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.6015625,
0.0236968994140625,
0.6015625
],
[
0.1732177734375,
0.4912109375,
0.82275390625,
0.492431640625,
0.82275390625,
0.5068359375,
0.1732177734375,
0.505859375
],
[
0.1748046875,
0.5068359375,
0.82275390625,
0.5068359375,
0.82275390625,
0.52099609375,
0.1748046875,
0.52099609375
],
[
0.1748046875,
0.521484375,
0.82421875,
0.521484375,
0.82421875,
0.53515625,
0.1748046875,
0.53515625
],
[
0.1732177734375,
0.53515625,
0.4306640625,
0.53515625,
0.4306640625,
0.548828125,
0.1732177734375,
0.548828125
],
[
0.1748046875,
0.5546875,
0.82275390625,
0.5546875,
0.82275390625,
0.568359375,
0.1748046875,
0.568359375
],
[
0.1732177734375,
0.568359375,
0.82275390625,
0.5693359375,
0.82275390625,
0.58349609375,
0.1732177734375,
0.58203125
],
[
0.1748046875,
0.58349609375,
0.82421875,
0.58349609375,
0.82421875,
0.5966796875,
0.1748046875,
0.5966796875
],
[
0.0236968994140625,
0.59228515625,
0.058013916015625,
0.59228515625,
0.058013916015625,
0.7109375,
0.0236968994140625,
0.7109375
],
[
0.1748046875,
0.5966796875,
0.82275390625,
0.5966796875,
0.82275390625,
0.60986328125,
0.1748046875,
0.60986328125
],
[
0.1748046875,
0.60986328125,
0.82421875,
0.60986328125,
0.82421875,
0.62353515625,
0.1748046875,
0.62353515625
],
[
0.1748046875,
0.62353515625,
0.73779296875,
0.62353515625,
0.73779296875,
0.63720703125,
0.1748046875,
0.63720703125
],
[
0.176513671875,
0.6591796875,
0.2034912109375,
0.6591796875,
0.2034912109375,
0.671875,
0.176513671875,
0.671875
],
[
0.1953125,
0.6572265625,
0.3095703125,
0.6572265625,
0.3095703125,
0.6748046875,
0.1953125,
0.6748046875
],
[
0.1748046875,
0.689453125,
0.82275390625,
0.689453125,
0.82275390625,
0.70263671875,
0.1748046875,
0.70263671875
],
[
0.1748046875,
0.7041015625,
0.82421875,
0.7041015625,
0.82421875,
0.71728515625,
0.1748046875,
0.71728515625
],
[
0.1748046875,
0.71728515625,
0.82421875,
0.71728515625,
0.82421875,
0.73095703125,
0.1748046875,
0.73095703125
],
[
0.1732177734375,
0.73095703125,
0.82275390625,
0.72998046875,
0.82275390625,
0.744140625,
0.1732177734375,
0.74560546875
],
[
0.1748046875,
0.744140625,
0.81640625,
0.744140625,
0.81640625,
0.75830078125,
0.1748046875,
0.75830078125
],
[
0.1748046875,
0.78076171875,
0.2034912109375,
0.78076171875,
0.2034912109375,
0.79248046875,
0.1748046875,
0.79248046875
],
[
0.1968994140625,
0.77978515625,
0.40185546875,
0.77978515625,
0.40185546875,
0.79443359375,
0.1968994140625,
0.79443359375
],
[
0.1732177734375,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.1732177734375,
0.82373046875
],
[
0.1748046875,
0.8359375,
0.82177734375,
0.8359375,
0.82177734375,
0.849609375,
0.1748046875,
0.849609375
],
[
0.1748046875,
0.85107421875,
0.82275390625,
0.85107421875,
0.82275390625,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.1993408203125,
0.875,
0.464111328125,
0.875,
0.464111328125,
0.88525390625,
0.1993408203125,
0.88525390625
],
[
0.1953125,
0.88525390625,
0.80810546875,
0.88623046875,
0.80810546875,
0.90087890625,
0.1953125,
0.8994140625
],
[
0.1748046875,
0.9248046875,
0.7236328125,
0.9248046875,
0.7236328125,
0.93798828125,
0.1748046875,
0.93798828125
]
] |
[
"Neural Wavetable: a playable wavetable synthesizer",
"using neural networks",
"Lamtharn Hantrakul*Li-Chia Yang",
"CenterforMusic Technology",
"8102A0N91",
"Georgia Institute of Technology",
"Georgia, USA 30318",
"[1hantrakul3][richard40148]@gatech.edu",
"Abstract",
"We present Neural Wavetable, a proof-of-concept wavetable synthesizer that uses",
"neural networks togenerateplayable wavetables.The system canproduce new,",
"distinct waveforms through the interpolation of traditional wavetables in an auto-",
"[as'sa]",
"encoder's latent space. It is available as a VST/AU plugin for use in a Digital Audio",
"Workstation.",
"1Introduction",
"ZAOSSO",
"WaveNet, a deep neural network architecture for speech and audio generation, has brought exciting",
"innovations in the sound domain since its introduction in 2016 [2]. In the musical domain, WaveNet",
"has been used as an auto-encoder capable of interpolating-or non-linear\"blending\"-different",
"instrument sounds through NSynth [1].",
"Inspired by this approach, we present a proof-of-concept system addressing a smaller scale audio",
"task: generating short wavetables. Rather than generate long segments of speech and conditioning",
"the network on long-term periodicity, we focus on generating short wavetables of length of 512",
"1181:AXI",
"samples; a contrast to the tens of thousands of samples required even in just a second of audio. These",
"short wavetables are fed into a real-time wavetable synthesizer implemented in C++ and the JUCE",
"framework. The source-code and VST/AU plug-in are available for public download?.",
"2",
"Background",
"Wavetable synthesis is an industry-standard approach which uses a bank of “wavetables\" of lengths",
"between 512-2048 and sample-rate conversion techniques to efficiently generate musical sounds. In",
"context of this work, wavetable synthesis efficiently handles real-time generation of different pitches",
"given a single wavetable generated by a neural network. For brevity, readers should reference the",
"original works for a complete description of Nsynth [1], WaveNet [2] and Wavetable synthesis [3]]",
"3",
"Implementation Details",
"3.1 WaveNet as a wavetable autoencoder",
"In this proof-of-concept, we use the pre-trained model NSynth WaveNet model as-is [1]. Given the",
"task of generating a short wavetable is smaller than NSynth's original task of generating several",
"CurrentlyanAIResidentwithGoogleBrain",
"2NeuralWavetable.https://github.com/RichardYang40148/Neural_Wavetable_Synthesizer",
"32nd Conference on Neural Information Processing Systems (NIPS 2018), Montreal, Canada"
] |
{
"arxiv_id": "1811.05550",
"language": "en",
"page_idx": 0
}
|
||
1811.05550-en-1
|
1811.05550-en-1.jpg
|
<FigureHere>
Figure 1: Left: Neural Wavetable plug-in GUI with interpolation, envelope and filter controls. Right: Interpolated vectors along an edge connecting two embedding vectors produces an interpolated wavetable that is musically playable. Deviating off this edge produces noise.
seconds of audio, we thus fırst perform qualitative tests exploring the latent space using wavetables
instead of audio segments (Figure 1).
First, we found the model retains short-term coherence: it can decode playable wavetables from inputs of short, simple wavetables such as sinusoid, triangle and sawtooth waves of length 512 (Figure 1). The model also produces a reconstruction $180^{\circ}$ out of phase from the input. This is a peculiar finding, given how the pre-trained model is trained on raw audio and not the fourier representation, where it is normal to disregard phase. Either way, this does not affect perception of pitch and timbre, since the human auditory system is largely phase-independent[4].
Secondly, interpolating along the edge connecting two embedding vectors from a sinusoid and sawtooth yielded waveforms that were a mix of these two input wavetables. However, moving “off” this edge produced noise or quantization noise.
## 3.2 Real-time limitations
The limitation of real-time rendering occurs in the encoding and decoding process of the pre-trained model (approximately 4 seconds to encode, and 7 seconds to decode a waveform on a GPU machine.). As a result, we pre-rendered 300 wavetables interpolating between three combinations of sinusoid, saw and triangle waveforms, and perform the wavetable synthesis in real-time. This approach is similar to the pre-generated waveforms in the public-facing NSynth demo$^3.$
# 3.3 Wavetable synthesis engine
Our real-time wavetable synthesizer uses bi-linear interpolation for reading wavetables. Since the generated wavetables vary in amplitude and offsets, we remove the mean and normalize to the range -1 to l. Each of the 512-sample wavetables are zero-padded on either side to produce a 514-sample wavetable. This ensures waveform looping starts and ends on the same value. In some wavetables, this introduces large discontinuities in the signal, leading to addition of high frequencies. We thus include an option to “smooth” the waveform at these discontinuities by approximating a line connecting extreme values to zero. The effect sounds like a gentle high-cut filter. The user can control these parameters via the GUI shown in Figure 1.
## 4 Conclusion and Future Work
We encourage readers to experiment with the Neural Wavetable VST/AU plugin; one can generate interesting shades combining a bright sawtooth and subdued sinusoid. By leveraging the strengths of both traditional and new machine-learning enabled synthesis methods, a newer implementation of Neural Wavetable under development employs a lightweight model designed to interpolate and decode short wavetables in real-time using a variational autoencoder.
$^{3}$NSynth Sound Maker. https://experiments. withgoogle. com/ai/sound-maker/view/
2
|
[
[
0.1748046875,
0.25,
0.826171875,
0.25,
0.826171875,
0.26318359375,
0.1748046875,
0.26318359375
],
[
0.176513671875,
0.264404296875,
0.826171875,
0.264404296875,
0.826171875,
0.27783203125,
0.176513671875,
0.27783203125
],
[
0.1748046875,
0.27783203125,
0.6748046875,
0.27783203125,
0.6748046875,
0.291748046875,
0.1748046875,
0.291748046875
],
[
0.1732177734375,
0.3193359375,
0.82177734375,
0.3193359375,
0.82177734375,
0.333251953125,
0.1732177734375,
0.333251953125
],
[
0.1748046875,
0.335205078125,
0.419189453125,
0.335205078125,
0.419189453125,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1748046875,
0.354248046875,
0.82177734375,
0.354248046875,
0.82177734375,
0.367431640625,
0.1748046875,
0.367431640625
],
[
0.1748046875,
0.36865234375,
0.82275390625,
0.36865234375,
0.82275390625,
0.3818359375,
0.1748046875,
0.3818359375
],
[
0.1732177734375,
0.3818359375,
0.82177734375,
0.3818359375,
0.82177734375,
0.395751953125,
0.1732177734375,
0.395751953125
],
[
0.1748046875,
0.396484375,
0.826171875,
0.396484375,
0.826171875,
0.410400390625,
0.1748046875,
0.410400390625
],
[
0.1748046875,
0.410400390625,
0.826171875,
0.410400390625,
0.826171875,
0.423583984375,
0.1748046875,
0.423583984375
],
[
0.1732177734375,
0.4248046875,
0.60693359375,
0.4248046875,
0.60693359375,
0.438232421875,
0.1732177734375,
0.438232421875
],
[
0.1748046875,
0.443115234375,
0.826171875,
0.443115234375,
0.826171875,
0.45703125,
0.1748046875,
0.45703125
],
[
0.176513671875,
0.460205078125,
0.82421875,
0.460205078125,
0.82421875,
0.47021484375,
0.176513671875,
0.47021484375
],
[
0.1732177734375,
0.472900390625,
0.483642578125,
0.472900390625,
0.483642578125,
0.486083984375,
0.1732177734375,
0.486083984375
],
[
0.1732177734375,
0.5048828125,
0.361083984375,
0.5048828125,
0.361083984375,
0.51953125,
0.1732177734375,
0.51953125
],
[
0.176513671875,
0.5341796875,
0.82275390625,
0.5341796875,
0.82275390625,
0.54443359375,
0.176513671875,
0.54443359375
],
[
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,
0.57373046875,
0.826171875,
0.57373046875,
0.826171875,
0.5869140625,
0.1748046875,
0.5869140625
],
[
0.1732177734375,
0.58837890625,
0.66650390625,
0.58837890625,
0.66650390625,
0.6015625,
0.1732177734375,
0.6015625
],
[
0.1732177734375,
0.62060546875,
0.404296875,
0.62060546875,
0.404296875,
0.6337890625,
0.1732177734375,
0.6337890625
],
[
0.1732177734375,
0.646484375,
0.82421875,
0.646484375,
0.82421875,
0.6611328125,
0.1732177734375,
0.6611328125
],
[
0.176513671875,
0.66357421875,
0.82177734375,
0.66357421875,
0.82177734375,
0.673828125,
0.176513671875,
0.673828125
],
[
0.1732177734375,
0.67529296875,
0.82275390625,
0.67529296875,
0.82275390625,
0.689453125,
0.1732177734375,
0.689453125
],
[
0.1748046875,
0.689453125,
0.82275390625,
0.689453125,
0.82275390625,
0.70263671875,
0.1748046875,
0.70263671875
],
[
0.176513671875,
0.7041015625,
0.82177734375,
0.7041015625,
0.82177734375,
0.7138671875,
0.176513671875,
0.7138671875
],
[
0.1748046875,
0.71630859375,
0.82275390625,
0.71630859375,
0.82275390625,
0.72998046875,
0.1748046875,
0.72998046875
],
[
0.1732177734375,
0.73095703125,
0.82421875,
0.73095703125,
0.82421875,
0.744140625,
0.1732177734375,
0.744140625
],
[
0.1748046875,
0.744140625,
0.45751953125,
0.744140625,
0.45751953125,
0.75830078125,
0.1748046875,
0.75830078125
],
[
0.17236328125,
0.77978515625,
0.450927734375,
0.78076171875,
0.450927734375,
0.79541015625,
0.17236328125,
0.79443359375
],
[
0.1748046875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.8271484375,
0.1748046875,
0.8271484375
],
[
0.176513671875,
0.8291015625,
0.82177734375,
0.8291015625,
0.82177734375,
0.8388671875,
0.176513671875,
0.8388671875
],
[
0.1732177734375,
0.84130859375,
0.8203125,
0.84130859375,
0.8203125,
0.85498046875,
0.1732177734375,
0.85498046875
],
[
0.1748046875,
0.85595703125,
0.826171875,
0.85595703125,
0.826171875,
0.869140625,
0.1748046875,
0.869140625
],
[
0.1748046875,
0.869140625,
0.62744140625,
0.869140625,
0.62744140625,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.1953125,
0.896484375,
0.76123046875,
0.89794921875,
0.76123046875,
0.912109375,
0.1953125,
0.9111328125
],
[
0.49267578125,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"Figure l: Left: Neural Wavetable plug-in GUI with interpolation, envelope and filter controls. Right:",
"Interpolated vectors along an edge connecting two embedding vectors produces an interpolated",
"wavetable that is musically playable. Deviating off this edge produces noise.",
"seconds of audio, we thus first perform qualitative tests exploring the latent space using wavetables",
"instead of audio segments (Figure 1).",
"First, we found the model retains short-term coherence: it can decode playable wavetables from",
"inputs of short, simple wavetables such as sinusoid, triangle and sawtooth waves of length 512 (Figure",
"1). The model also produces a reconstruction 180 out of phase from the input. This is a peculial",
"finding, given how the pre-trained model is trained on raw audio and not the fourier representation,",
"where it is normal to disregard phase. Either way, this does not affect perception of pitch and timbre,",
"since the human auditory system is largely phase-independent [4].",
"Secondly, interpolating along the edge connecting two embedding vectors from a sinusoid and",
"sawtoothyieldedwaveformsthatwereamix ofthesetwoinputwavetables.However,moving\"off\"",
"this edge produced noise or quantization noise.",
"3.2Real-time limitations",
"Thelimitationofreal-timerenderingoccursintheencodinganddecodingprocessofthepre-trained",
"model (approximately 4 seconds to encode, and 7 seconds to decode a waveform on a GPU machine.)",
"As a result, we pre-rendered 300 wavetables interpolating between three combinations of sinusoid,",
"saw and triangle waveforms, and perform the wavetable synthesis in real-time. This approach is",
"similar to the pre-generated waveforms in the public-facing NSynth demo",
"3.3Wavetable synthesis engine",
"Our real-time wavetable synthesizer uses bi-linear interpolation for reading wavetables. Since the",
"generatedwavetablesvaryinamplitudeandoffsets,weremovethemeanandnormalizetotherange",
"-1 to 1. Each of the 512-sample wavetables are zero-padded on either side to produce a 514-sample",
"wavetable. This ensures waveform looping starts and ends on the same value. In some wavetables, this",
"introduceslargediscontinuitiesinthesignal.leadingtoadditionofhighfreguencies.Wethus include",
"an option to “smooth\" the waveform at these discontinuities by approximating a line connecting",
"extreme values to zero. The effect sounds like a gentle high-cut filter. The user can control these",
"parameters via the GUI shown in Figure 1.",
"4 Conclusion and Future Work",
"We encourage readers to experiment with the Neural Wavetable VST/AU plugin; one can generate",
"interestingshades combiningabrightsawtoothand subdued sinusoid.Byleveragingthestrengths",
"of both traditional and new machine-learning enabled synthesis methods, a newer implementatior",
"of Neural Wavetable under development employs a lightweight model designed to interpolate and",
"decode short wavetables in real-time using a variational autoencoder.",
"3NSynth Sound Maker.https://experiments.withgoogle.com/ai/sound-maker/view/",
"2"
] |
{
"arxiv_id": "1811.05550",
"language": "en",
"page_idx": 1
}
|
||
1712.06830-en-0
|
1712.06830-en-0.jpg
|
Single Image Deraining using Scale-Aware Multi-Stage Recurrent Network
Ruoteng Li$^{1}$, Loong-Fah Cheong$^{\mathrm{l}}$, and Robby T. Tan$^{1,2}$
$^{1}$National University of Singapore
$^{2}$Yale- NUS
<FigureHere>
Figure 1: Comparison of our method with state-of-the-art deraining algorithms. Top Left: Input image. Top Right: Result of [23]. Bottom Left: Result of [5]. Bottom Right: Result of the proposed method.
## Abstract
Given a single input rainy image, our goal is to visually remove rain streaks and the veiling effect caused by scattering and transmission of rain streaks and rain droplets. We are particularly concerned with heavy rain, where rain
$\sum_{\begin{array}{cc}streaks&of&various&sizes&and&directions&can&overlap&each\\&other&and&the&veiling&effect&reduces&contrast&severely.&To\\&achieve&our&goal,we&introduce&a&scale-aware&multi-stage\\&convolutional&neural&network.&Our&main&idea&here&is&that\end{array}}$ $\begin{array}{lll}&convolutional&neural&network.&Ourmain&idea&here&is&that\\\hline o_{2}&different&sizes&of&rain-streaks&visually°rade&the&scene&in\\\hline C&different&ways.&Large&nearby&streaks&obstruct&larger®ions\end{array}$
$differentways. \textit{ Large nearby streaks obstruct larger regions}$ and are likely to reflect specular highlights more prominently than smaller distant streaks. These different effects of different streaks have their own characteristics in their image features, and thus need to be treated differently. To realize this, we create parallel sub-networks that are trained and made aware of these different scales of rain streaks. To our knowledge, this idea of parallel sub-networks that treats the same class of objects according to their unique sub-classes is novel, particularly in the context of rain removal. To verify our idea, we conducted experiments on both synthetic and real images, and found that our method is effective and outperforms the state-of-the-art methods
## $1. \textbf{Introduction}$
Rain, particularly heavy rain, can impair visibility considerably. Individual rain streaks mar object's appearance with their specular highlights, refraction, scattering and blurring effects. Distant rain streaks accumulated along the line of sight degrade the visibility of a background scene by creating fog-like veiling effect. Moreover, as a result of the projection process, rain streaks in the image have different sizes and densities, depending on their distances to the camera; these different rain layers with different sizes and densities are confusedly overlaid upon each other in the image, severely aggravating the problem. All these can affect the performance of current computer vision algorithms,
particularly those assuming clear visibility.
A few rain streaks removal techniques have been proposed in the past decade to eliminate the rain streaks and restore visibility. Video-based rain streak removal methods (e.g. [6][13]) focus on image recovery from video sequence by exploring temporal information and frequency properties of rain. Some single-image based rain removal methods regard the problem as blind signal separation problem (e.g. [14]). Recently, deep learning methods have been applied to this area and have demonstrated their advantages in recover ing background scenes from rainy images [23][5][4]. While all the aforementioned methods have demonstrated some degree of success, it is fair to say that they have not been subject to the full force of the tropical heavy rain and been tested where the scenes contain a range of depths. Both these factors render the deraining problem much harder; not only do we need to deal with a diverse range of rainfall from slight drizzle to an almost solid curtain of water, the rain
1
|
[
[
0.11273193359375,
0.1337890625,
0.859375,
0.1337890625,
0.859375,
0.1483154296875,
0.11273193359375,
0.1483154296875
],
[
0.265625,
0.1925048828125,
0.70361328125,
0.1900634765625,
0.70361328125,
0.2108154296875,
0.265625,
0.21337890625
],
[
0.3505859375,
0.222900390625,
0.62158203125,
0.2247314453125,
0.62158203125,
0.242431640625,
0.3505859375,
0.240478515625
],
[
0.43701171875,
0.2412109375,
0.53125,
0.24365234375,
0.53125,
0.25830078125,
0.43701171875,
0.255615234375
],
[
0.0196075439453125,
0.2587890625,
0.058013916015625,
0.25830078125,
0.062103271484375,
0.37255859375,
0.0236968994140625,
0.373779296875
],
[
0.236083984375,
0.300537109375,
0.314453125,
0.300537109375,
0.314453125,
0.318115234375,
0.236083984375,
0.318115234375
],
[
0.1004638671875,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.1004638671875,
0.347900390625
],
[
0.08087158203125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.0236968994140625,
0.3642578125,
0.052276611328125,
0.361083984375,
0.056365966796875,
0.389404296875,
0.0294189453125,
0.392578125
],
[
0.08087158203125,
0.3662109375,
0.465576171875,
0.3662109375,
0.465576171875,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.08087158203125,
0.37890625,
0.46728515625,
0.37890625,
0.46728515625,
0.392578125,
0.08087158203125,
0.392578125
],
[
0.08251953125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.08251953125,
0.4091796875
],
[
0.0269622802734375,
0.4072265625,
0.052276611328125,
0.4072265625,
0.052276611328125,
0.449462890625,
0.0269622802734375,
0.449462890625
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.07843017578125,
0.4248046875,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.07843017578125,
0.438232421875
],
[
0.08251953125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.451904296875,
0.08251953125,
0.451904296875
],
[
0.0318603515625,
0.449462890625,
0.058013916015625,
0.449462890625,
0.058013916015625,
0.479736328125,
0.0318603515625,
0.479736328125
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.080078125,
0.48486328125,
0.46728515625,
0.48486328125,
0.46728515625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.035125732421875,
0.5068359375,
0.049835205078125,
0.5068359375,
0.049835205078125,
0.52978515625,
0.035125732421875,
0.52978515625
],
[
0.0269622802734375,
0.51953125,
0.0555419921875,
0.51953125,
0.0555419921875,
0.5498046875,
0.0269622802734375,
0.5498046875
],
[
0.08251953125,
0.517578125,
0.465576171875,
0.517578125,
0.465576171875,
0.52783203125,
0.08251953125,
0.52783203125
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.0318603515625,
0.54248046875,
0.0555419921875,
0.54736328125,
0.0482177734375,
0.56298828125,
0.02532958984375,
0.55810546875
],
[
0.50341796875,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.08087158203125,
0.54736328125,
0.468017578125,
0.54736328125,
0.468017578125,
0.55810546875,
0.08087158203125,
0.55810546875
],
[
0.02777099609375,
0.55615234375,
0.052276611328125,
0.55615234375,
0.052276611328125,
0.57373046875,
0.02777099609375,
0.57373046875
],
[
0.5048828125,
0.5537109375,
0.8896484375,
0.5537109375,
0.8896484375,
0.564453125,
0.5048828125,
0.564453125
],
[
0.080078125,
0.56103515625,
0.46728515625,
0.56103515625,
0.46728515625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.0269622802734375,
0.57080078125,
0.037567138671875,
0.5673828125,
0.04412841796875,
0.58056640625,
0.03350830078125,
0.583984375
],
[
0.50244140625,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.57470703125,
0.4697265625,
0.57568359375,
0.4697265625,
0.59033203125,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.58349609375,
0.7099609375,
0.58349609375,
0.7099609375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.02777099609375,
0.59130859375,
0.053924560546875,
0.59130859375,
0.053924560546875,
0.611328125,
0.02777099609375,
0.611328125
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.0294189453125,
0.609375,
0.045745849609375,
0.609375,
0.045745849609375,
0.63525390625,
0.0294189453125,
0.63525390625
],
[
0.080078125,
0.60498046875,
0.46728515625,
0.60498046875,
0.46728515625,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.0228729248046875,
0.6337890625,
0.056365966796875,
0.63330078125,
0.05963134765625,
0.708984375,
0.02532958984375,
0.71044921875
],
[
0.50244140625,
0.62890625,
0.78369140625,
0.62890625,
0.78369140625,
0.642578125,
0.50244140625,
0.642578125
],
[
0.080078125,
0.63623046875,
0.4697265625,
0.63623046875,
0.4697265625,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.51953125,
0.64453125,
0.8896484375,
0.64599609375,
0.8896484375,
0.6591796875,
0.51953125,
0.65771484375
],
[
0.080078125,
0.6513671875,
0.450927734375,
0.6513671875,
0.450927734375,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.035125732421875,
0.67529296875,
0.0400390625,
0.67529296875,
0.0400390625,
0.6845703125,
0.035125732421875,
0.6845703125
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.04412841796875,
0.681640625,
0.049835205078125,
0.681640625,
0.049835205078125,
0.6923828125,
0.04412841796875,
0.6923828125
],
[
0.08087158203125,
0.69580078125,
0.2091064453125,
0.69580078125,
0.2091064453125,
0.708984375,
0.08087158203125,
0.708984375
],
[
0.5048828125,
0.6923828125,
0.890625,
0.6923828125,
0.890625,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.098876953125,
0.7197265625,
0.46728515625,
0.7197265625,
0.46728515625,
0.73291015625,
0.098876953125,
0.73291015625
],
[
0.50341796875,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50341796875,
0.734375
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.5009765625,
0.73486328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.5009765625,
0.74853515625
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50341796875,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.5048828125,
0.767578125,
0.88818359375,
0.767578125,
0.88818359375,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5048828125,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.5048828125,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
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,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.87060546875,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Single Image Deraining using Scale-Aware Multi-Stage Recurrent Network",
"Ruoteng Lil, Loong-Fah Cheong', and Robby T. Tan'.2",
"INational University of Singapore",
"2Yale-NUS",
"10000",
"Abstract",
"Given a single input rainy image, our goal is to visually",
"remove rain streaks and the veiling effect caused by scat-",
"9",
"tering and transmission of rain streaks and rain droplets.",
"We are particularly concerned with heavy rain, where rain",
"streaks of various sizes and directions can overlap each",
"A",
"other and the veiling effect reduces contrast severely. To",
"achieve our goal,we introducea scale-aware multi-stage",
"convolutional neural network.Ourmain ideahere is that",
"'SO",
"different sizes of rain-streaks visually degrade the scene in",
"different ways. Large nearby streaks obstruct larger regions",
"and are likely to reflect specular highlights more promi-",
"nently than smaller distant streaks. These different effects",
"A",
"8",
"ofdifferentstreakshavetheirowncharacteristicsintheir",
"image features, and thus need to be treated differently: To",
"8",
"Figure l: Comparison of our method with state-of-the-art",
"realizethis,we createparallelsub-networksthataretrained",
"6",
"deraining algorithms.Top Left:Input image.Top Right:",
"and made aware of these different scales of rain streaks.",
"C",
"Result of [23]. Bottom Left: Result of [5]. Bottom Right:",
"To our knowledge, this idea of parallel sub-networks that",
"Result of the proposed method.",
"2",
"treats the same class of objects according to their unique",
"LL",
"sub-classes is novel, particularly in the context of rain re-",
"moval. To verify our idea, we conducted experiments on",
"L:AIXI",
"particularly those assuming clear visibility.",
"both synthetic and real images, and found that our method",
"A few rain streaks removal techniques have been pro-",
"is effective and outperforms the state-of-the-art methods.",
"posed in the past decade to eliminate the rain streaks and",
"",
"restore visibility. Video-based rain streak removal methods",
"",
"1. Introduction",
"(e.g.[6][13D)focusonimagerecoveryfromvideosequence",
"by exploring temporal information and frequency proper-",
"Rain, particularly heavy rain, can impair visibility con-",
"ties of rain. Some single-image based rain removal methods",
"siderably. Individual rain streaks mar object's appearance",
"regard the problem as blind signal separation problem (e.g.",
"with their specular highlights, refraction, scattering and",
"[14]). Recently, deep learning methods have been applied to",
"blurring effects. Distant rain streaks accumulated along the",
"thisareaandhavedemonstratedtheiradvantagesinrecover-",
"line of sight degrade the visibility of a background scene",
"ing background scenes from rainy images [23][5][4]. While",
"by creating fog-like veiling effect. Moreover, as a result of",
"all the aforementioned methods have demonstrated some",
"the projection process, rain streaks in the image have dif-",
"degree of success, it is fair to say that they have not been",
"ferent sizes and densities, depending on their distances to",
"subject to the full force of the tropical heavy rain and been",
"the camera; these different rain layers with different sizes",
"tested where the scenes contain a range of depths. Both",
"and densities areconfusedly overlaid upon each otherinthe",
"thesefactors renderthederainingproblemmuchharder;not",
"image, severely aggravating the problem. All these can af-",
"only do we need to deal with a diverse range of rainfall from",
"fect the performance of current computer vision algorithms,",
"slight drizzle to an almost solid curtain of water, the rain"
] |
{
"arxiv_id": "1712.06830",
"language": "en",
"page_idx": 0
}
|
||
1712.06830-en-1
|
1712.06830-en-1.jpg
|
<FigureHere>
Figure 2: An example of heavy rain image. Green window: The compounding result of rain streaks accumulation and veiling effect. Yellow window: Rain streaks of various size overlap on each other. Blue and Red window: The fog-like veiling effect varies according to the object depth. Further object has stronger veiling effect.
might appear thinner in the foreground but thicker in the more distant areas in the same image. To make deraining algorithms resilient to these more severe conditions, there are several limitations of current algorithms that need to be carefully addressed.
First of all, most of the existing rain streaks removal methods ([16][11][14]) are developed based on an assumption that the rain streaks distributed on a captured image is sparse. However, in the real world, particularly in the case of heavy rain or even in the case of moderate rain, if the scene extends far enough in depth, the dense rain streaks accumulation makes this assumption invalid as shown in Fig.2. Second, rain streaks of different sizes overlapping each other can cause ambiguity or even unintelligibility for feature-based and learning-based methods as shown in Fig. 4 [23][14]. Although the appearance of an individual rain streak follows a rain model as described in [6], the large number of streaks overlapping in various sizes and densities at best significantly expands the feature space to learn, and at worst produces a rain image comprising of several scales of phenomena interpenetrating one another, rendering the existing feature-based learning methods [11][3][16][14][23] inefficient in correctly detecting rain streaks. Finally, the atmospheric veiling effect caused by light scattering process of both tiny and large rain droplets plays an important role in degrading the visibility of a rainy scene (Fig.2). However most deraining algorithms did not address this problem properly.
Considering the aforementioned limitations, our goal is to develop a novel method that is capable of removing rain streaks and rain accumulation from a single image under various conditions, ranging from slight rain to heavy rain, and thus enhancing the visibility of the image. To achieve our goal, we introduce a multi-stage scale-aware convolutional neural network. Our main idea here is that different sizes of rain-streaks visually degrade the scene in different ways. Large nearby streaks obstruct larger regions and are
<FigureHere>
Figure 3: An example of real rain image. Rain streaks of different sizes may appear at the same time, and overlap each other. The enlarged windows demonstrate the different features of the rain streaks.
likely to reflect specular highlights more prominently than smaller distant streaks [20]. These different effects of different streaks have their own characteristics in their image features, and thus need to be treated differently. Thick rain streaks also tend to be lower in density and thus need larger spatio-temporal windows to properly analyze them. One might think that the different layers of a deep learning algorithm might be able to do this automatically, but when the differently-sized rain streaks are so inextricably mixed together, we contend that this is far from being the cases. To realize the different treatments effectively, we create parallel sub-networks that are trained and made aware of these different scales of rain streaks. To our knowledge, this idea of parallel sub-networks that treat the same class of objects according to their unique sub-classes is novel, particularly in the context of rain removal.
Our scale-aware multi-stage convolutional neural network consist of three parts. First, we adopt DenseNet[9] as a backbone to extract general features. Next to it is parallel sub-networks, each of which is trained to estimate rain streaks intensity map at a scale. These parallel subnetworks are recurrent convolutional layers with shortcut connections [7], which are iteratively refined to produce better rain streaks predictions. Then the input image subtracts the summed results of all the subnetworks and feeds forward to the next stage of parallel recurrent sub-networks. In each of the subsequent stages, the recurrent subnetworks predict the residual rain streaks based on the proceeding subtraction results from previous stage. Finally, the estimated rain streak maps will be combined with the input image to restore a clean background scene.
Our key contributions can be summarized as follows:
## 1. We introduce an end-to-end network to remove both the rain streaks and their accumulation effect.
2. The proposed scale-aware network addresses the over-
lapping rain streaks of different sizes and densities using parallel recurrent sub-networks. Each sub-network
2
|
[
[
0.08087158203125,
0.2247314453125,
0.46728515625,
0.2247314453125,
0.46728515625,
0.238037109375,
0.08087158203125,
0.238037109375
],
[
0.50244140625,
0.231689453125,
0.89208984375,
0.231689453125,
0.89208984375,
0.24560546875,
0.50244140625,
0.24560546875
],
[
0.08087158203125,
0.240478515625,
0.4697265625,
0.240478515625,
0.4697265625,
0.25390625,
0.08087158203125,
0.25390625
],
[
0.5009765625,
0.24560546875,
0.890625,
0.246826171875,
0.890625,
0.261474609375,
0.5009765625,
0.260009765625
],
[
0.08087158203125,
0.255126953125,
0.4697265625,
0.255126953125,
0.4697265625,
0.268310546875,
0.08087158203125,
0.268310546875
],
[
0.50244140625,
0.26318359375,
0.89208984375,
0.26318359375,
0.89208984375,
0.276611328125,
0.50244140625,
0.276611328125
],
[
0.080078125,
0.270751953125,
0.4697265625,
0.270751953125,
0.4697265625,
0.2841796875,
0.080078125,
0.2841796875
],
[
0.50244140625,
0.27587890625,
0.68115234375,
0.276611328125,
0.68115234375,
0.291748046875,
0.50244140625,
0.290283203125
],
[
0.080078125,
0.2841796875,
0.468017578125,
0.2841796875,
0.468017578125,
0.29736328125,
0.080078125,
0.29736328125
],
[
0.080078125,
0.300537109375,
0.298095703125,
0.300537109375,
0.298095703125,
0.314453125,
0.080078125,
0.314453125
],
[
0.50341796875,
0.3193359375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.50341796875,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.07843017578125,
0.34228515625,
0.468017578125,
0.341552734375,
0.468017578125,
0.35595703125,
0.07843017578125,
0.356689453125
],
[
0.080078125,
0.356689453125,
0.4697265625,
0.35791015625,
0.4697265625,
0.37255859375,
0.080078125,
0.371826171875
],
[
0.50341796875,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50341796875,
0.3642578125
],
[
0.50244140625,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.080078125,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.5009765625,
0.37890625,
0.89208984375,
0.380126953125,
0.89208984375,
0.39453125,
0.5009765625,
0.393310546875
],
[
0.07843017578125,
0.38818359375,
0.4697265625,
0.386962890625,
0.4697265625,
0.401611328125,
0.07843017578125,
0.40283203125
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.08087158203125,
0.40478515625,
0.2132568359375,
0.40478515625,
0.2132568359375,
0.415283203125,
0.08087158203125,
0.415283203125
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.098876953125,
0.41845703125,
0.4697265625,
0.41845703125,
0.4697265625,
0.431884765625,
0.098876953125,
0.431884765625
],
[
0.50341796875,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.4375,
0.50341796875,
0.4375
],
[
0.08087158203125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.08087158203125,
0.447509765625
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.080078125,
0.448974609375,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.080078125,
0.462158203125
],
[
0.50341796875,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.080078125,
0.464111328125,
0.4697265625,
0.464111328125,
0.4697265625,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50244140625,
0.47021484375,
0.890625,
0.47021484375,
0.890625,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.080078125,
0.479248046875,
0.4697265625,
0.479248046875,
0.4697265625,
0.492431640625,
0.080078125,
0.492431640625
],
[
0.50244140625,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.080078125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50341796875,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.08087158203125,
0.5087890625,
0.4697265625,
0.5087890625,
0.4697265625,
0.52294921875,
0.08087158203125,
0.52294921875
],
[
0.50341796875,
0.5166015625,
0.89208984375,
0.5166015625,
0.89208984375,
0.52978515625,
0.50341796875,
0.52978515625
],
[
0.08087158203125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.08087158203125,
0.53857421875
],
[
0.50341796875,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.54345703125,
0.50341796875,
0.54345703125
],
[
0.080078125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.5537109375,
0.080078125,
0.5537109375
],
[
0.50244140625,
0.546875,
0.705078125,
0.546875,
0.705078125,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.08087158203125,
0.55615234375,
0.46728515625,
0.55615234375,
0.46728515625,
0.56640625,
0.08087158203125,
0.56640625
],
[
0.521484375,
0.56103515625,
0.890625,
0.56103515625,
0.890625,
0.57470703125,
0.521484375,
0.57470703125
],
[
0.080078125,
0.5693359375,
0.46728515625,
0.5693359375,
0.46728515625,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.57568359375,
0.8896484375,
0.57568359375,
0.8896484375,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.080078125,
0.58544921875,
0.46728515625,
0.58544921875,
0.46728515625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50341796875,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.60498046875,
0.50341796875,
0.60498046875
],
[
0.080078125,
0.599609375,
0.4697265625,
0.599609375,
0.4697265625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.619140625,
0.50244140625,
0.619140625
],
[
0.080078125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.6220703125,
0.8935546875,
0.6220703125,
0.8935546875,
0.63525390625,
0.50244140625,
0.63525390625
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50341796875,
0.63818359375,
0.890625,
0.63818359375,
0.890625,
0.6484375,
0.50341796875,
0.6484375
],
[
0.08087158203125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.50341796875,
0.6513671875,
0.89208984375,
0.6513671875,
0.89208984375,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.08087158203125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.50341796875,
0.66845703125,
0.8896484375,
0.66845703125,
0.8896484375,
0.6787109375,
0.50341796875,
0.6787109375
],
[
0.08087158203125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.50244140625,
0.681640625,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.69775390625,
0.890625,
0.69775390625,
0.890625,
0.7109375,
0.50244140625,
0.7109375
],
[
0.08087158203125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.7138671875,
0.890625,
0.7138671875,
0.890625,
0.724609375,
0.50341796875,
0.724609375
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73486328125
],
[
0.50341796875,
0.72802734375,
0.89208984375,
0.72802734375,
0.89208984375,
0.7412109375,
0.50341796875,
0.7412109375
],
[
0.080078125,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.50341796875,
0.74365234375,
0.890625,
0.74365234375,
0.890625,
0.7568359375,
0.50341796875,
0.7568359375
],
[
0.07843017578125,
0.75,
0.279296875,
0.751953125,
0.279296875,
0.76513671875,
0.07843017578125,
0.76318359375
],
[
0.50244140625,
0.7568359375,
0.890625,
0.7568359375,
0.890625,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.098876953125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77978515625,
0.098876953125,
0.77978515625
],
[
0.50341796875,
0.77294921875,
0.7685546875,
0.77294921875,
0.7685546875,
0.7861328125,
0.50341796875,
0.7861328125
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.521484375,
0.787109375,
0.87158203125,
0.7861328125,
0.87158203125,
0.80029296875,
0.521484375,
0.8017578125
],
[
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
],
[
0.515625,
0.81494140625,
0.89208984375,
0.81494140625,
0.89208984375,
0.8291015625,
0.515625,
0.8291015625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.53515625,
0.83349609375,
0.833984375,
0.83349609375,
0.833984375,
0.84326171875,
0.53515625,
0.84326171875
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84033203125,
0.468017578125,
0.85400390625,
0.07843017578125,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.5146484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.5146484375,
0.869140625
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.53369140625,
0.8720703125,
0.890625,
0.87060546875,
0.890625,
0.88525390625,
0.53369140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.53515625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.53515625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"Figure 2: An example of heavy rain image. Green window:",
"Figure 3: An example of real rain image. Rain streaks of",
"The compounding result of rain streaks accumulation and",
"different sizes may appear at the same time, and overlap",
"veiling effect. Yellow window: Rain streaks of various size",
"each other. The enlarged windows demonstrate the different",
"overlap on each other. Blue and Red window: The fog-like",
"features of the rain streaks.",
"veiling effect varies according to the object depth. Further",
"object has stronger veiling effect.",
"likely to reflect specular highlights more prominently than",
"smaller distant streaks [20]. These different effects of dif-",
"might appear thinner in the foreground but thicker in the",
"more distant areas in the same image. To make deraining",
"ferent streaks have their own characteristics in their image",
"features, and thus need to be treated differently. Thick rain",
"algorithms resilient to these more severe conditions, there",
"streaks also tend to be lower in density and thus need larger",
"are several limitations of current algorithms that need to be",
"spatio-temporal windows to properly analyze them.One",
"carefully addressed.",
"might think that the different layers of a deep learning algo-",
"First of all, most of the existing rain streaks removal",
"rithmmight be abletodothis automatically,but when the",
"methods ([16][11][14]) are developed based on an assump-",
"differently-sized rain streaks are so inextricably mixed to-",
"tion that the rain streaks distributed on a captured image is",
"gether, we contend that this is far from being the cases. To",
"sparse. However, in the real world, particularly in the case",
"realize the different treatments effectively, we create paral-",
"of heavy rain or even in the case of moderate rain, if the",
"lel sub-networks that are trained and made aware of these",
"scene extends far enough in depth, the dense rain streaks",
"different scales of rain streaks. To our knowledge, this idea",
"accumulation makes this assumption invalid as shown in",
"of parallel sub-networks that treat the same class of objects",
"Fig. 2. Second, rain streaks of different sizes overlapping",
"according totheir unique sub-classes is novel,particularly",
"each other can cause ambiguity or even unintelligibility",
"in the context of rain removal.",
"forfeature-basedandlearning-based methodsas shown in",
"Our scale-aware multi-stage convolutional neural net-",
"Fig. 4 [23][14]. Although the appearance of an individ-",
"work consist of three parts. First, we adopt DenseNet [9]",
"ual rain streak follows a rain model as described in [6],",
"as a backbone to extract general features. Next to it is",
"the large number of streaks overlapping in various sizes",
"parallel sub-networks, each of which is trained to estimate",
"and densities at best significantly expands the feature space",
"rain streaks intensity map at a scale.These parallel sub-",
"to learn, and at worst produces a rain image comprising",
"networks are recurrent convolutional layerswith shortcut",
"of several scales of phenomena interpenetrating one an-",
"connections [7], which are iteratively refined to produce",
"other, rendering the existing feature-based learning methods",
"betterrainstreaks predictions.Then theinputimagesub-",
"[11][3][16][14][23] inefficient in correctly detecting rain",
"tracts the summed results of all the subnetworks and feeds",
"streaks. Finally, the atmospheric veiling effect caused by",
"forward to the next stage of parallel recurrent sub-networks.",
"light scattering process of both tiny and large rain droplets",
"Ineachofthesubsequentstages,therecurrentsubnetworks",
"plays an important role in degrading the visibility of a rainy",
"predict the residual rain streaks based on the proceeding",
"scene (Fig.2). However most deraining algorithms did not",
"subtraction results from previous stage. Finally, the esti-",
"address this problem properly.",
"mated rain streak maps will be combined with the input im-",
"Considering the aforementioned limitations, our goal is",
"age to restore a clean background scene.",
"todevelopanovel method that is capable of removing rain",
"Our key contributions can be summarized as follows:",
"streaks and rain accumulation from a single image under",
"various conditions, ranging from slight rain to heavy rain,",
"1. We introduce an end-to-end network to remove both",
"and thus enhancing the visibility of the image. To achieve",
"therain streaks andtheiraccumulation effect.",
"our goal, we introduce a multi-stage scale-aware convolu-",
"tional neural network. Our main idea here is that different",
"2. The proposed scale-aware network addresses the over-",
"sizes of rain-streaks visually degrade the scene in different",
"lapping rain streaks of different sizes and densities us-",
"ways. Large nearby streaks obstruct larger regions and are",
"ing parallel recurrent sub-networks. Each sub-network",
"2"
] |
{
"arxiv_id": "1712.06830",
"language": "en",
"page_idx": 1
}
|
||
1712.06830-en-3
|
1712.06830-en-3.jpg
|
rain region mask is fed into the network to further learn rain streak intensity, and hence the clean image can be restored by subtracting the rain intensity map from input rain image.
## $3. \textbf{Rain Models}$
The widely used rain model describes the observed rain image O as a linear combination of the rain-free background scene B and the rain streak layer R [16][14][8]:
$$\mathbf{O}=\mathbf{B}+\mathbf{R}.$$
(1)
The objective for any rain streak removal algorithm is to remove the rain streaks layer R from the rainy image O to obtain the background scene B.
Rain removal algorithms based on Eq.(1) assume the rain streaks sparse and utilize individual rain streak characteristics (e.g. [14][13]) to separate background and rain streak layers. Unfortunately, in the real world, rain appearance does not depend only on individual rain streaks, but also on the accumulation of multiple rain streaks in the space from the camera to the background scene, as shown $\operatorname{in}$Fig.2.
In an image, the projected appearance of the rain streaks will have different sizes and densities. The further away the streaks, the smaller the size and the denser the imaged rain streaks. If we assume rain streaks at the same depth as one layer, we divide the captured rainy scene into a clean background and multiple layers of rain streaks, each of which has approximately the same size and density. Based on this, we can generalize Eq. (1) to model a rainy scene:
(2)
$$\mathbf{O=B+\sum_{i}^{n}R_{i},}$$
where $n$ is the number of rain-streak layers along the line of sight to the background objects. $\mathbf{R_i}$ represents the pixel intensity of rain streaks at layer $i.$
According to[12], rain droplets can cause light scattering and attenuation. Thus, the resultant visibility under moderate and heavy rain conditions are similar to those under haze and fog. The light scattering process contributes to the atmospheric veiling effect in a typical heavy rainy scene (see blue window in Fig. 2). In this case, the purely additive rain model introduced in Eq.(1) does not fully capture the appearance of rain accumulation, and as a result, the existing rain removal methods based on Eq. (1) cannot handle it. To address this, we further generalize the rain model:
$$\mathbf{O}=\mathbf{\alpha}\odot(\mathbf{B}+\sum_i^n\mathbf{R_i})+(\mathbf{1}-\mathbf{\alpha})\odot\mathbf{A},$$
(3)
where $\odot$ indicates element-wise multiplication, $\alpha$ is a 2D map representing the transmittance introduced by the scattering process of rain droplets, and A is a 2D map representing the atmospheric light of the scene. This model is
similar to that proposed in [23], although we remove the
binary mask in our model.
## $4. \textbf{ Deraining Method}$
Our scale-aware multi-stage recurrent deraining network is illustrated in Fig. 5. Unlike the existing CNN-based deraining methods, we create multiple recurrent sub-networks to handle rain streaks at different sizes and densities. The veiling effect of rain accumulation, $(1-\boldsymbol{\alpha})\mathbf{A}$, can be deemed as another layer and thus can be processed using another recurrent sub-network in parallel to those that handle rain streaks. However, since the underlying physical process generating the veiling effect is different from that of rain streaks (Fig. 2 Green), the network architecture treats this effect differently. For ease of discussion, we will first discuss our network that deals only with rain streaks, called 'SMRNet’(Scale-aware Multi-stage Recurrent Network) and shown in Fig. 5(a). Then, we will discuss the integration of this network with the sub-network that deals with the veiling effect, which we call 'SMRNet-veil', shown in Fig. 5( b) .
## $4. 1. \textbf{SMRNet}$
Focusing on multiple layers and overlapping rain streaks, based on Eq.(2), our goal is to estimate the rainfree background B and each rain streak intensity map $\mathbf{R_1,R_2,...,R_n}$ given the input rainy image O. Generally, we want to minimize$\parallel\mathbf{O}-\hat{\mathbf{B}}-\sum_{i}^{n}\mathbf{R}_{\mathbf{i}}\parallel_{F}^{2}$,where$\parallel\cdot\parallel_{F}$ is the Frobenius norm. This is a totally ill-posed problem, even when we regard there is only one layer of rain streaks. Early rain streak removal methods like [14][3] use handcrafted features or data-driven features for the priors of B and $\mathbf{R_i}.$ However, in our method, the priors are learned by the network from the training data. In order to obtain the priors of B and $\mathbf{R_i}$ during the training phase, we add the estimation loss of B and $\mathbf{R_i}$ as in our objectives:
(4)
$$\mathcal{L}=\mathcal{L}_{\mathcal{B}}+\sum_{i}^{n}\mathcal{L}_{\mathcal{R}_{i}},$$
where $\mathcal{L}_\mathcal{B}$ represents reconstruction loss for the background scene, and $\hat{\mathcal{L}}_{\mathcal{R}_i}$ represents the loss for estimating the $i^th$ rain streak layer.
As illustrated in Fig. 5, our network first adopts DenseNet[9] to extract rain image features F. However, we remove the'transition layers'(1 conv layer followed by 1 pooling layer) between each'Dense block' from the DenseNet so that our network does not downsample the image. The features are then fed into a series of parallel recurrent convolutional sub-networks to learn rain intensity map $\mathbf{R_i}$ at different scales and densities. Since each recurrent sub-network focuses only on one type of rain streak feature $\mathbf{R_i,~resources~can~be~dedicated~toward~these~selected~rain}$
4
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
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,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.67724609375,
0.10797119140625,
0.67724609375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.46728515625,
0.12371826171875,
0.46728515625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50341796875,
0.1363525390625,
0.67724609375,
0.1363525390625,
0.67724609375,
0.15087890625,
0.50341796875,
0.15087890625
],
[
0.08087158203125,
0.149658203125,
0.20751953125,
0.149658203125,
0.20751953125,
0.1641845703125,
0.08087158203125,
0.1641845703125
],
[
0.521484375,
0.160400390625,
0.89208984375,
0.160400390625,
0.89208984375,
0.173583984375,
0.521484375,
0.173583984375
],
[
0.098876953125,
0.1749267578125,
0.468017578125,
0.1749267578125,
0.468017578125,
0.1881103515625,
0.098876953125,
0.1881103515625
],
[
0.50244140625,
0.1767578125,
0.890625,
0.1767578125,
0.890625,
0.1900634765625,
0.50244140625,
0.1900634765625
],
[
0.08087158203125,
0.1912841796875,
0.46728515625,
0.1912841796875,
0.46728515625,
0.2020263671875,
0.08087158203125,
0.2020263671875
],
[
0.50244140625,
0.1912841796875,
0.89208984375,
0.1912841796875,
0.89208984375,
0.2052001953125,
0.50244140625,
0.2052001953125
],
[
0.07843017578125,
0.2052001953125,
0.450927734375,
0.203857421875,
0.450927734375,
0.2183837890625,
0.07843017578125,
0.2197265625
],
[
0.50341796875,
0.2083740234375,
0.890625,
0.2083740234375,
0.890625,
0.2183837890625,
0.50341796875,
0.2183837890625
],
[
0.50244140625,
0.222900390625,
0.8935546875,
0.222900390625,
0.8935546875,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.2279052734375,
0.23291015625,
0.318603515625,
0.23486328125,
0.318603515625,
0.25,
0.2279052734375,
0.2474365234375
],
[
0.44677734375,
0.236083984375,
0.4697265625,
0.236083984375,
0.4697265625,
0.248779296875,
0.44677734375,
0.248779296875
],
[
0.50244140625,
0.23486328125,
0.89208984375,
0.2374267578125,
0.89208984375,
0.251953125,
0.50244140625,
0.25
],
[
0.08087158203125,
0.255126953125,
0.4697265625,
0.255126953125,
0.4697265625,
0.268310546875,
0.08087158203125,
0.268310546875
],
[
0.5009765625,
0.251953125,
0.8896484375,
0.251953125,
0.8896484375,
0.26513671875,
0.5009765625,
0.26513671875
],
[
0.080078125,
0.270751953125,
0.4697265625,
0.270751953125,
0.4697265625,
0.2841796875,
0.080078125,
0.2841796875
],
[
0.50244140625,
0.267578125,
0.89208984375,
0.267578125,
0.89208984375,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.080078125,
0.285400390625,
0.289306640625,
0.285400390625,
0.289306640625,
0.298583984375,
0.080078125,
0.298583984375
],
[
0.50341796875,
0.2822265625,
0.89208984375,
0.2822265625,
0.89208984375,
0.29541015625,
0.50341796875,
0.29541015625
],
[
0.09722900390625,
0.2998046875,
0.468017578125,
0.298583984375,
0.468017578125,
0.313232421875,
0.09722900390625,
0.314453125
],
[
0.5009765625,
0.29541015625,
0.89208984375,
0.296630859375,
0.89208984375,
0.311279296875,
0.5009765625,
0.31005859375
],
[
0.080078125,
0.31494140625,
0.4697265625,
0.31494140625,
0.4697265625,
0.328857421875,
0.080078125,
0.328857421875
],
[
0.50341796875,
0.3125,
0.89208984375,
0.3125,
0.89208984375,
0.32568359375,
0.50341796875,
0.32568359375
],
[
0.07843017578125,
0.330810546875,
0.4697265625,
0.330810546875,
0.4697265625,
0.3447265625,
0.07843017578125,
0.3447265625
],
[
0.50341796875,
0.328857421875,
0.890625,
0.328857421875,
0.890625,
0.339111328125,
0.50341796875,
0.339111328125
],
[
0.080078125,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.359130859375,
0.080078125,
0.359130859375
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.50244140625,
0.35595703125
],
[
0.07843017578125,
0.361083984375,
0.468017578125,
0.35986328125,
0.468017578125,
0.373779296875,
0.07843017578125,
0.375
],
[
0.5009765625,
0.356689453125,
0.890625,
0.35791015625,
0.890625,
0.371826171875,
0.5009765625,
0.37060546875
],
[
0.080078125,
0.375732421875,
0.4697265625,
0.375732421875,
0.4697265625,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.50244140625,
0.375,
0.890625,
0.375,
0.890625,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.080078125,
0.391357421875,
0.4697265625,
0.391357421875,
0.4697265625,
0.40478515625,
0.080078125,
0.40478515625
],
[
0.5009765625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.5009765625,
0.40087890625
],
[
0.07843017578125,
0.406005859375,
0.1422119140625,
0.40771484375,
0.1422119140625,
0.42041015625,
0.07843017578125,
0.41845703125
],
[
0.50244140625,
0.40283203125,
0.56787109375,
0.40283203125,
0.56787109375,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.09722900390625,
0.421630859375,
0.468017578125,
0.421630859375,
0.468017578125,
0.43505859375,
0.09722900390625,
0.43505859375
],
[
0.50341796875,
0.427978515625,
0.60302734375,
0.427978515625,
0.60302734375,
0.439453125,
0.50341796875,
0.439453125
],
[
0.080078125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.521484375,
0.449462890625,
0.890625,
0.449462890625,
0.890625,
0.464111328125,
0.521484375,
0.464111328125
],
[
0.080078125,
0.466552734375,
0.4697265625,
0.466552734375,
0.4697265625,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.5009765625,
0.464111328125,
0.89208984375,
0.46533203125,
0.89208984375,
0.479736328125,
0.5009765625,
0.479248046875
],
[
0.080078125,
0.481689453125,
0.468017578125,
0.481689453125,
0.468017578125,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.50341796875,
0.48095703125,
0.89208984375,
0.48095703125,
0.89208984375,
0.494384765625,
0.50341796875,
0.494384765625
],
[
0.080078125,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.50341796875,
0.49560546875,
0.890625,
0.49560546875,
0.890625,
0.5087890625,
0.50341796875,
0.5087890625
],
[
0.080078125,
0.51220703125,
0.4697265625,
0.51220703125,
0.4697265625,
0.52587890625,
0.080078125,
0.52587890625
],
[
0.50244140625,
0.5087890625,
0.890625,
0.51025390625,
0.890625,
0.5244140625,
0.50244140625,
0.5234375
],
[
0.07843017578125,
0.5263671875,
0.404296875,
0.52587890625,
0.404296875,
0.54052734375,
0.07843017578125,
0.54150390625
],
[
0.5009765625,
0.52587890625,
0.8896484375,
0.52587890625,
0.8896484375,
0.5390625,
0.5009765625,
0.5390625
],
[
0.50341796875,
0.54248046875,
0.890625,
0.54248046875,
0.890625,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.291748046875,
0.55029296875,
0.30224609375,
0.55322265625,
0.301513671875,
0.55810546875,
0.2900390625,
0.55615234375
],
[
0.281982421875,
0.5595703125,
0.33251953125,
0.5595703125,
0.33251953125,
0.57373046875,
0.281982421875,
0.57373046875
],
[
0.50244140625,
0.55615234375,
0.89208984375,
0.55615234375,
0.89208984375,
0.5693359375,
0.50244140625,
0.5693359375
],
[
0.44677734375,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57275390625,
0.44677734375,
0.57275390625
],
[
0.214111328125,
0.56005859375,
0.27685546875,
0.56005859375,
0.27685546875,
0.5712890625,
0.214111328125,
0.5712890625
],
[
0.50244140625,
0.57080078125,
0.89208984375,
0.57080078125,
0.89208984375,
0.583984375,
0.50244140625,
0.583984375
],
[
0.50341796875,
0.5859375,
0.890625,
0.5859375,
0.890625,
0.599609375,
0.50341796875,
0.599609375
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.6015625,
0.89208984375,
0.6015625,
0.89208984375,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.080078125,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62255859375,
0.080078125,
0.62255859375
],
[
0.50244140625,
0.6162109375,
0.89208984375,
0.6162109375,
0.89208984375,
0.6298828125,
0.50244140625,
0.6298828125
],
[
0.08087158203125,
0.62548828125,
0.2998046875,
0.62548828125,
0.2998046875,
0.63623046875,
0.08087158203125,
0.63623046875
],
[
0.50244140625,
0.6318359375,
0.82275390625,
0.6318359375,
0.82275390625,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.09722900390625,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.65283203125,
0.09722900390625,
0.65283203125
],
[
0.080078125,
0.65380859375,
0.4697265625,
0.65380859375,
0.4697265625,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.62744140625,
0.6640625,
0.7646484375,
0.66650390625,
0.7646484375,
0.6845703125,
0.62744140625,
0.681640625
],
[
0.869140625,
0.66748046875,
0.8935546875,
0.66748046875,
0.8935546875,
0.68115234375,
0.869140625,
0.68115234375
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.080078125,
0.66845703125,
0.46728515625,
0.66845703125,
0.46728515625,
0.681640625,
0.080078125,
0.681640625
],
[
0.080078125,
0.69970703125,
0.468017578125,
0.69970703125,
0.468017578125,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.50341796875,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.080078125,
0.7138671875,
0.4697265625,
0.7138671875,
0.4697265625,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.5,
0.7197265625,
0.8935546875,
0.71630859375,
0.8935546875,
0.734375,
0.5,
0.7373046875
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72998046875,
0.468017578125,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.50244140625,
0.736328125,
0.61279296875,
0.736328125,
0.61279296875,
0.75,
0.50244140625,
0.75
],
[
0.07843017578125,
0.74560546875,
0.468017578125,
0.744140625,
0.468017578125,
0.75830078125,
0.07843017578125,
0.7587890625
],
[
0.521484375,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.521484375,
0.76513671875
],
[
0.080078125,
0.76025390625,
0.46728515625,
0.76025390625,
0.46728515625,
0.7734375,
0.080078125,
0.7734375
],
[
0.5048828125,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.08087158203125,
0.77587890625,
0.426513671875,
0.77587890625,
0.426513671875,
0.7861328125,
0.08087158203125,
0.7861328125
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.1380615234375,
0.806640625,
0.25,
0.806640625,
0.25,
0.8212890625,
0.1380615234375,
0.8212890625
],
[
0.2418212890625,
0.8037109375,
0.414306640625,
0.806640625,
0.414306640625,
0.82470703125,
0.2418212890625,
0.8212890625
],
[
0.4453125,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.4453125,
0.8212890625
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.5009765625,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84033203125,
0.468017578125,
0.8388671875,
0.468017578125,
0.85400390625,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5009765625,
0.85400390625
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"rain region mask is fed into the network to further learn rain",
"similar to that proposed in [23], although we remove the",
"streak intensity, and hence the clean image can be restored",
"binary mask in our model.",
"by subtracting the rain intensity map from input rain image.",
"4. Deraining Method",
"3. Rain Models",
"Our scale-aware multi-stage recurrent deraining network",
"The widely used rain model describes the observed rain",
"is illustrated in Fig. 5. Unlike the existing CNN-based de-",
"image O as a linear combination of the rain-free back-",
"raining methods, we create multiple recurrent sub-networks",
"ground scene B and the rain streak layer R [16][14][8] :",
"to handlerain streaks at different sizes and densities.The",
"veiling effect of rain accumulation, (1 -α)A, can be",
"O=B+R.",
"(1)",
"deemed as another layer and thus can be processed using",
"The objective for any rain streak removal algorithm is to",
" another recurrent sub-network in parallel to those that han-",
"remove the rain streaks layer R from the rainy image O to",
"dle rain streaks. However, since the underlying physical",
"obtain the background scene B.",
"process generating the veiling effect is different from that of",
"Rain removal algorithms based on Eg. (1) assume the",
"rain streaks (Fig. 2 Green), the network architecture treats",
"rain streaks sparse and utilize individual rain streak char-",
"this effect differently. For ease of discussion, we will first",
"acteristics (e.g. [14][13]) to separate background and rain",
"discussournetworkthatdeals onlywithrainstreaks.called",
"streak layers.Unfortunately, in the real world, rain ap-",
"\"SMRNet\" (Scale-aware Multi-stage Recurrent Network)",
"pearance does not depend only on individual rain streaks,",
"and shown in Fig. 5(a). Then, we will discuss the integra-",
"but also on the accumulation of multiple rain streaks in the",
"tion of this network with the sub-network that deals with",
"space from the camera to the background scene, as shown",
"the veiling effect, which we call 'SMRNet-veil', shown in",
"inFig.2.",
"Fig. 5(b).",
"In an image, the projected appearance of the rain streaks",
"4.1. SMRNet",
"will have different sizes and densities. The further away the",
"streaks, the smaller the size and the denser the imaged rain",
"Focusing on multiple layers and overlapping rain streaks,",
"streaks. If we assume rain streaks at the same depth as one",
"based on Eq. (2), our goal is to estimate the rain-",
"layer, we divide the captured rainy scene into a clean back-",
"free background B and each rain streak intensity map",
"ground and multiple layers of rain streaks, each of which",
"Ri, R2, ..., Rn given the input rainy image O. Generally,",
"has approximately the same size and density. Based on this,",
"wewanttominimizelO-B-nR;l/,wherell·lF",
"we can generalize Eq. (1) to model a rainy scene:",
"is the Frobenius norm. This is a totally ill-posed problem,",
"even when weregardthereisonlyonelayerof rain streaks.",
"TU",
"Ri",
"Early rain streak removal methods like [14][3] use hand-",
"(2)",
"O=B+",
"crafted features or data-driven features for the priors of B",
"and Ri: However, in our method, the priors are learned by",
"where n is the number of rain-streak layers along the line",
"the network from the training data. In order to obtain the",
"of sight to the background objects. Ri represents the pixel",
"priors of B and R; during the training phase, we add the",
"intensity of rain streaks at layer i.",
"estimation loss of B and Ri as in our objectives:",
"According to [12], rain droplets can cause light scat-",
"tering and attenuation. Thus, the resultant visibility under",
"C=CB+CR,",
"(4)",
"der haze and fog. The light scattering process contributes to",
"moderate and heavy rain conditions are similar to those un-",
"the atmospheric veiling effect in a typical heavy rainy scene",
"where Lg represents reconstruction loss for the background",
"(see blue window in Fig. 2). In this case, the purely additive",
"scene, and CR: represents the loss for estimating the ith",
"rain model introduced in Eq. (1) does not fully capture the",
"rain streak layer.",
"appearance of rain accumulation, and as a result, the exist-",
"As illustrated in Fig. 5, our network first adopts",
"ing rain removal methods based on Eq. (1) cannot handle it.",
"DenseNet[9l to extract rain image features F.However.",
"To address this.we further generalize therain model:",
"we remove thetransition layers'(l conv layer followed",
"by 1 pooling layer) between each 'Dense block' from the",
"O=α(B+ ",
"Ri) + (1 -α)o A,",
"(3)",
"DenseNet so that our network does not downsample the im-",
"age. The features are then fed into a series of parallel recur-",
"where indicates element-wise multiplication, α is a 2D",
"rent convolutional sub-networks to learn rain intensity map",
"map representing the transmittance introduced by the scat-",
"Rat different scales and densities. Since each recurrent",
"tering process of rain droplets, and A is a 2D map repre-",
"sub-network focuses only on one type of rain streak feature",
"senting the atmospheric light of the scene. This model is",
"Ri, resources can be dedicated toward these selected rain",
"4"
] |
{
"arxiv_id": "1712.06830",
"language": "en",
"page_idx": 3
}
|
||
1712.06830-en-4
|
1712.06830-en-4.jpg
|
<FigureHere>
Figure 5: The network architectures for SMRNet (a) and SMRNet-Veil (b).
derstood in various senses: (1) in the case of coincident rain streaks mentioned before, the removal of the nearest (and thus brightest) rain streaks reveal the underlying fainter rain streaks; (2) local rain streaks that are inconsistent in density or direction with the global pattern are better detected and processed after the dominant global pattern is removed; (3) the rain streaks can also be dominated by the veiling effect to be discussed in the next subsection, and the latters removal in the first stage helps reveal rain streaks better.
The second reason for having these multiple stages is when we are faced with an unprecedented heavy rain not seen in the training. The first stage may only partially remove the rain streaks, having not seen such dense rain streak pattern before. However, the partially derained result at the end of the first stage amounts to an image of a lighter rainy scene, and we find that the successive stages can successfully remove the remaining rain streaks. The third reason is concerned with the removal of the dark streak artifacts. One can regard the concatenated predictions from the earlier stages as providing some form of explicit 'communication’ between each recurrent sub-network, leading to reduced'duplicate work'. At the end of the network, the learned rain streaks from all the stages are concatenated together to aid the final clean image recovery.
streaks without having to accommodate the competing demands of the different rain streak types in the network with their different desiderata for representation, yielding an enhanced learning of foreground items. The recurrent module iterates four times in order to refine the estimation using the previous prediction combined with the feature representations. Then, the input image will subtract all the estimated rain intensity maps $\mathbf{R_i^1,}$s to obtain a temporary image T$_1$ , which is a preliminary de-rained result. This preliminary result is often marred by dark streak artifacts and contains residual rain streaks. The dark streaks arise possibly due to “double removal" by two parallel sub-networks. The reasons for the residual rain streaks are manifold.
Firstly, even in a synthetic rain image, an image region may contain multiple rain streaks stacked right on top of one another. It is difficult to remove all the rain streaks in one go; instead, we are much more likely to see the removal of only the nearest and thickest rain streaks, upon which the further and finer streaks are revealed. Secondly, in the real world rain sequences, there might be effects not modelled in the training. For instance, the rainfall is simply heavier, or there are local variation of density and direction not related to the depth factor (e.g. the variation around the table in Fig. 1). These effects would pose difficulties for a singlestage network solution, even with its parallel sub-networks. In view of the preceding issues, we send the preliminary derained result to the next stage together with the feature representation F and the various estimated rain layers for further refinement. A predictor $\mathbf{R}_{\mathrm{i}}^{\mathrm{j}}$ in next stage can better remove the residual rain streaks because firstly the dominant signals have been removed. This argument can be un-
## $4. 2. \textbf{Veil Module}$
Not only does the veiling effect, degrade visibility, its presence also hinders the complete removal of all the rain streaks effectively, as shown in Fig.4. For these reasons, we develop an additional module to specifically handle it, with the module placed in parallel with the recurrent rain streak
5
|
[
[
0.08416748046875,
0.09344482421875,
0.1422119140625,
0.09344482421875,
0.1422119140625,
0.10479736328125,
0.08416748046875,
0.10479736328125
],
[
0.381591796875,
0.09661865234375,
0.4150390625,
0.09661865234375,
0.4150390625,
0.1060791015625,
0.381591796875,
0.1060791015625
],
[
0.5947265625,
0.0947265625,
0.66845703125,
0.0947265625,
0.66845703125,
0.1060791015625,
0.5947265625,
0.1060791015625
],
[
0.339111328125,
0.11236572265625,
0.420654296875,
0.11236572265625,
0.420654296875,
0.11932373046875,
0.339111328125,
0.11932373046875
],
[
0.56787109375,
0.111083984375,
0.650390625,
0.111083984375,
0.650390625,
0.11871337890625,
0.56787109375,
0.11871337890625
],
[
0.77685546875,
0.126953125,
0.8203125,
0.126953125,
0.8203125,
0.146484375,
0.77685546875,
0.146484375
],
[
0.83837890625,
0.126953125,
0.88623046875,
0.1307373046875,
0.8837890625,
0.146484375,
0.83642578125,
0.143310546875
],
[
0.0849609375,
0.1331787109375,
0.13232421875,
0.1331787109375,
0.13232421875,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.2115478515625,
0.1331787109375,
0.240234375,
0.1331787109375,
0.240234375,
0.14013671875,
0.2115478515625,
0.14013671875
],
[
0.273681640625,
0.1331787109375,
0.303955078125,
0.1331787109375,
0.303955078125,
0.14013671875,
0.273681640625,
0.14013671875
],
[
0.2467041015625,
0.1395263671875,
0.26953125,
0.1395263671875,
0.26953125,
0.146484375,
0.2467041015625,
0.146484375
],
[
0.56787109375,
0.255615234375,
0.6396484375,
0.255615234375,
0.6396484375,
0.266357421875,
0.56787109375,
0.266357421875
],
[
0.6748046875,
0.25244140625,
0.75244140625,
0.25244140625,
0.75244140625,
0.26318359375,
0.6748046875,
0.26318359375
],
[
0.10455322265625,
0.268310546875,
0.1380615234375,
0.268310546875,
0.1380615234375,
0.27978515625,
0.10455322265625,
0.27978515625
],
[
0.096435546875,
0.279052734375,
0.12091064453125,
0.279052734375,
0.12091064453125,
0.285888671875,
0.096435546875,
0.285888671875
],
[
0.1282958984375,
0.279052734375,
0.143798828125,
0.279052734375,
0.143798828125,
0.28271484375,
0.1282958984375,
0.28271484375
],
[
0.82421875,
0.2734375,
0.830078125,
0.2734375,
0.830078125,
0.2822265625,
0.82421875,
0.2822265625
],
[
0.11767578125,
0.285888671875,
0.125,
0.285888671875,
0.125,
0.291748046875,
0.11767578125,
0.291748046875
],
[
0.82275390625,
0.29345703125,
0.83251953125,
0.29345703125,
0.83251953125,
0.309326171875,
0.82275390625,
0.309326171875
],
[
0.62353515625,
0.306884765625,
0.63623046875,
0.306884765625,
0.63623046875,
0.31005859375,
0.62353515625,
0.31005859375
],
[
0.81640625,
0.314453125,
0.83837890625,
0.314453125,
0.83837890625,
0.328857421875,
0.81640625,
0.328857421875
],
[
0.1993408203125,
0.330078125,
0.25732421875,
0.330078125,
0.25732421875,
0.34033203125,
0.1993408203125,
0.34033203125
],
[
0.26318359375,
0.330078125,
0.40185546875,
0.330078125,
0.40185546875,
0.34033203125,
0.26318359375,
0.34033203125
],
[
0.39794921875,
0.330078125,
0.464111328125,
0.330078125,
0.464111328125,
0.34033203125,
0.39794921875,
0.34033203125
],
[
0.5,
0.330810546875,
0.52294921875,
0.330810546875,
0.52294921875,
0.339111328125,
0.5,
0.339111328125
],
[
0.51318359375,
0.33837890625,
0.54736328125,
0.33837890625,
0.54736328125,
0.34521484375,
0.51318359375,
0.34521484375
],
[
0.28515625,
0.35107421875,
0.370849609375,
0.35107421875,
0.370849609375,
0.361083984375,
0.28515625,
0.361083984375
],
[
0.72802734375,
0.35595703125,
0.74609375,
0.35595703125,
0.74609375,
0.363037109375,
0.72802734375,
0.363037109375
],
[
0.7861328125,
0.35595703125,
0.80126953125,
0.35595703125,
0.80126953125,
0.363037109375,
0.7861328125,
0.363037109375
],
[
0.75341796875,
0.3642578125,
0.775390625,
0.3642578125,
0.775390625,
0.36865234375,
0.75341796875,
0.36865234375
],
[
0.240234375,
0.39013671875,
0.7294921875,
0.39013671875,
0.7294921875,
0.404052734375,
0.240234375,
0.404052734375
],
[
0.080078125,
0.43310546875,
0.468017578125,
0.43310546875,
0.468017578125,
0.4462890625,
0.080078125,
0.4462890625
],
[
0.50341796875,
0.434326171875,
0.890625,
0.434326171875,
0.890625,
0.4443359375,
0.50341796875,
0.4443359375
],
[
0.0767822265625,
0.447509765625,
0.4697265625,
0.4462890625,
0.4697265625,
0.4609375,
0.0767822265625,
0.462158203125
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.46337890625,
0.4697265625,
0.46337890625,
0.4697265625,
0.4765625,
0.080078125,
0.4765625
],
[
0.50244140625,
0.46337890625,
0.8935546875,
0.46337890625,
0.8935546875,
0.4765625,
0.50244140625,
0.4765625
],
[
0.080078125,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.080078125,
0.492431640625
],
[
0.50341796875,
0.479736328125,
0.8896484375,
0.479736328125,
0.8896484375,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.5087890625,
0.46728515625,
0.5087890625,
0.46728515625,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.50341796875,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.537109375,
0.50341796875,
0.537109375
],
[
0.07843017578125,
0.53857421875,
0.46728515625,
0.53857421875,
0.46728515625,
0.5517578125,
0.07843017578125,
0.5517578125
],
[
0.50244140625,
0.5390625,
0.8935546875,
0.5390625,
0.8935546875,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.07843017578125,
0.55322265625,
0.468017578125,
0.5537109375,
0.468017578125,
0.568359375,
0.07843017578125,
0.5673828125
],
[
0.50244140625,
0.5537109375,
0.87109375,
0.5537109375,
0.87109375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.52294921875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.52294921875,
0.58203125
],
[
0.07843017578125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.07843017578125,
0.59765625
],
[
0.50244140625,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.50244140625,
0.599609375,
0.64892578125,
0.599609375,
0.64892578125,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.65283203125,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.65283203125,
0.6123046875
],
[
0.07843017578125,
0.6142578125,
0.38232421875,
0.6142578125,
0.38232421875,
0.62744140625,
0.07843017578125,
0.62744140625
],
[
0.50244140625,
0.6142578125,
0.8935546875,
0.6142578125,
0.8935546875,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.098876953125,
0.630859375,
0.4697265625,
0.630859375,
0.4697265625,
0.64453125,
0.098876953125,
0.64453125
],
[
0.5009765625,
0.62744140625,
0.89208984375,
0.62890625,
0.89208984375,
0.64306640625,
0.5009765625,
0.642578125
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.69140625,
0.890625,
0.69140625,
0.890625,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.07843017578125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.705078125,
0.8935546875,
0.705078125,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.08087158203125,
0.75244140625,
0.46728515625,
0.75244140625,
0.46728515625,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50244140625,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.77978515625,
0.7900390625,
0.77978515625,
0.7900390625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.80322265625,
0.62841796875,
0.80322265625,
0.62841796875,
0.81640625,
0.50341796875,
0.81640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.52294921875,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.52294921875,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5009765625,
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,
0.4697265625,
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
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"(a) SMRNet",
"Stage1",
"Stagete [2..M)",
"lecurrentSubnetworks",
"Recurrent Subnetwork",
"ConWReH",
"CeanWne",
"Inpuwmane",
"Feature",
"tations!",
"xwx",
"(b)SMRNet-Vell",
"Staget E(1..M)",
"Feature",
"eprese",
"atio",
"",
"F",
"",
"",
"8",
"Conv64xkxk",
"(Conv64×x1x1Conv32x3x3",
"Conv32x1x1",
"RainS",
"3xWxH",
"Recurrent Module",
"1/tra",
":1/t",
"XWX",
"Figure 5: The network architectures for SMRNet (a) and SMRNet-Veil (b)",
"streaks without having to accommodate the competing de-",
"derstood invarious senses:(1)inthe case of coincidentrain",
" mands of the different rain streak types in the network with",
"streaks mentioned before, the removal of the nearest (and",
"their different desiderata for representation, yielding an en-",
"thus brightest) rain streaks reveal the underlying fainter rain",
"hanced learning of foreground items. The recurrent module",
"streaks; (2) local rain streaks that are inconsistent in den-",
"iterates four times in order to refine the estimation using the",
"sity or direction with the global pattern are better detected",
"previous prediction combined with the feature representa-",
"and processed after the dominant global pattern is removed;",
"tions. Then, the input image will subtract all the estimated",
"(3) the rain streaks can also be dominated by the veiling ef-",
"rain intensity maps Rl's to obtain a temporary image Ti",
"fect to be discussed in the next subsection, and the latters",
", which is a preliminary de-rained result. This preliminary",
"removal in the first stage helps reveal rain streaks better.",
"result is often marred by dark streak artifacts and contains",
"The second reason for having these multiple stages is",
"residual rain streaks. The dark streaks arise possibly due to",
"when we are faced with an unprecedented heavy rain not",
"\"double removal' by two parallel sub-networks. The rea-",
"seen in the training..",
"The first stage may only partially",
" sons for the residual rain streaks are manifold.",
"remove the rain streaks, having not seen such dense rain",
"Firstly, even in a synthetic rain image, an image region",
"streak pattern before. However, the partially derained re-",
"may contain multiple rain streaks stacked right on top of",
"sult at the end of the first stage amounts to an image of a",
"one another. It is difficult to remove all the rain streaks in",
"lighter rainy scene, and we find that the successive stages",
"one go; instead, we are much more likely to see the removal",
"can successfully remove the remaining rain streaks. The",
"of only the nearest and thickest rain streaks, upon which the",
"third reason is concernedwiththeremovalof thedarkstreak",
"further and finer streaks are revealed. Secondly, in the real",
"artifacts. One can regard the concatenated predictions from",
"world rain sequences, there might be effects not modelled",
"the earlier stages as providing some form of explicit 'com-",
"in the training. For instance, the rainfall is simply heavier,",
"munication' between each recurrent sub-network, leading",
"or there are local variation of density and direction not re-",
"to reduced 'duplicate work'. At the end of the network, the",
"lated to the depth factor (e.g. the variation around the table",
"learned rain streaks from all the stages are concatenated to-",
"inFig.I).These effects wouldpose difficultiesfora single-",
"gether to aid the final clean image recovery.",
"stage network solution, even with its parallel sub-networks.",
"4.2. Veil Module",
"In view of the preceding issues, we send the preliminary",
"derained result to the next stage together with the feature",
"Not only does the veiling effect, degrade visibility, its",
"representation F and the various estimated rain layers for",
"presence also hinders the complete removal of all the rain",
"further refinement. A predictor R; in next stage can better",
"streaks effectively, as shown in Fig. 4. For these reasons, we",
"remove the residual rain streaks because firstly the domi-",
"develop an additional module to specifically handle it, with",
"nant signals have been removed. This argument can be un-",
"the module placed in parallel with the recurrent rain streak",
"5"
] |
{
"arxiv_id": "1712.06830",
"language": "en",
"page_idx": 4
}
|
||
2404.02041-en-0
|
2404.02041-en-0.jpg
|
SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation
Vinkle Srivastav$^{1,2*}$
Keqi Chen$^{1* }$ Nicolas Padoy$^{1,2}$
$^{1}$University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France
${}^{2}$IHU Strasbourg, France
srivastav@unistra.fr
keqi.chen@unistra.fr npadoy@unistra.fr
$ন্র$
$N$
### $\mathbf{Abstract}$
supervised methods, our approach does not require any $2dor3dground-truthposesandusesonlythemulti-$
view input images from a calibrated camera setup and 2d 3d pose reconstruction as a mathemation as a mathemation task, preudo prose reconstruction as a mathemation task, pseudo poses generated from am an offf- the- shelf 2d human primarily focusing on
$objectives: self-supervised person localization in 3d space$ and self-supervised 3d pose estimation. We achieve selfsupervised 3d person localization by training the model on synthetically generated 3d points, serving as 3d person root positions, and on the projected root-heatmaps in all the views. We then model the 3d poses of all the localized persons with a bottleneck representation, map them onto all views obtaining 2d joints, and render them using 2d Gaussian heatmaps in an end-to-end differentiable man-
$\begin{array}{lll}&ner.Afterwards,we\:use\:the\:corresponding\:\end{array}$
propose an adaptive supervision attention mechanism to guide the self-supervision. Our experiments and analysis on three public benchmark datasets, including Panoptic, Shelf, and Campus, show the effectiveness of our approach, which is comparable to fully-supervised methods. Code is available at https: / github. com/ CAMMApublic/selfPose3D.
## $1. \textbf{Introduction}$
The task of estimating 3d poses for multiple persons using a few calibrated cameras is a challenging computer vision problem [10,18,32,52,57]. A significant part of this challenge lies in identifying and matching the same person across different camera views. The solutions developed so far generally use one of the two paradigms: learning-based
$^{*}$co-first authors with equal contributions.
methods and optimization-based methods. The learningbased methods develop novel deep-learning models and use 3d ground-truth poses for both training the models and establishing person correspondences across different views [39,52,57,59,63]. The accurate 3d ground-truth poses are typically generated using a dense camera system [32]. In contrast, the optimization-based methods formulate the angulation within the framework of multi-view geometry [10,18,31-33,49]. The 2d poses are estimated using offthe-shelf 2d human pose detectors [5,55]. These methods apply geometric and spatial constraints in the optimization loop to ensure the anatomical plausibility and consistency of the inferred 3d poses. Although these methods do not require 3d ground-truth poses, their effectiveness is somewhat limited compared to the fully-supervised learningbased methods, see Table 1.
In this paper, we explore the possibility of combining the strengths of both paradigms. Specifically, we investigate whether it's feasible to utilize a learning-based model for multi-view, multi-person 3D pose estimation and simultaneously eliminate its dependence on 3D ground-truth poses by incorporating geometric and appearance constraints, drawing inspiration from optimization-based methods.
We propose, SelfPose3d, a self-supervised learningbased approach to estimate the 3d poses of multiple persons from a few calibrated cameras without using any 2d or 3d ground-truth poses. Our approach requires only 2d pseudo poses obtained using an off-the-shelf 2d pose detector [55]. Learning 3d poses without 3D ground-truth poses would require suitable supervisory signals to train a learning-based model. We follow the learning-by-projection paradigm, where the main idea is to learn the 3d output by comparing the projected bottleneck 3d output against the 2d input features [12].
We consider VoxelPose [57] as a learning-based method and use its output 3d poses as a bottleneck representation To recover the accurate underlying 3d poses, we propose us-
|
[
[
0.12091064453125,
0.1331787109375,
0.8505859375,
0.1331787109375,
0.8505859375,
0.147705078125,
0.12091064453125,
0.147705078125
],
[
0.2646484375,
0.18115234375,
0.428955078125,
0.1787109375,
0.428955078125,
0.1976318359375,
0.2646484375,
0.199462890625
],
[
0.438720703125,
0.18310546875,
0.5537109375,
0.18310546875,
0.5537109375,
0.1976318359375,
0.438720703125,
0.1976318359375
],
[
0.54833984375,
0.18115234375,
0.701171875,
0.1787109375,
0.701171875,
0.1976318359375,
0.54833984375,
0.199462890625
],
[
0.2132568359375,
0.20263671875,
0.7587890625,
0.2052001953125,
0.7587890625,
0.222900390625,
0.2132568359375,
0.2203369140625
],
[
0.40185546875,
0.2215576171875,
0.5703125,
0.2235107421875,
0.5703125,
0.2392578125,
0.40185546875,
0.2374267578125
],
[
0.2091064453125,
0.242431640625,
0.39208984375,
0.242431640625,
0.39208984375,
0.255615234375,
0.2091064453125,
0.255615234375
],
[
0.40283203125,
0.242431640625,
0.59228515625,
0.242431640625,
0.59228515625,
0.255615234375,
0.40283203125,
0.255615234375
],
[
0.60302734375,
0.242431640625,
0.75732421875,
0.242431640625,
0.75732421875,
0.255615234375,
0.60302734375,
0.255615234375
],
[
0.0196075439453125,
0.264404296875,
0.062103271484375,
0.264404296875,
0.06451416015625,
0.376953125,
0.0228729248046875,
0.376953125
],
[
0.236083984375,
0.290283203125,
0.314453125,
0.290283203125,
0.314453125,
0.30810546875,
0.236083984375,
0.30810546875
],
[
0.50341796875,
0.29345703125,
0.88818359375,
0.29345703125,
0.88818359375,
0.306884765625,
0.50341796875,
0.306884765625
],
[
0.50341796875,
0.309326171875,
0.89208984375,
0.309326171875,
0.89208984375,
0.322509765625,
0.50341796875,
0.322509765625
],
[
0.1004638671875,
0.323974609375,
0.468017578125,
0.323974609375,
0.468017578125,
0.33837890625,
0.1004638671875,
0.33837890625
],
[
0.50244140625,
0.322509765625,
0.8896484375,
0.322509765625,
0.8896484375,
0.3359375,
0.50244140625,
0.3359375
],
[
0.080078125,
0.339111328125,
0.468017578125,
0.339111328125,
0.468017578125,
0.354248046875,
0.080078125,
0.354248046875
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.50244140625,
0.354248046875,
0.89208984375,
0.354248046875,
0.89208984375,
0.367431640625,
0.50244140625,
0.367431640625
],
[
0.02777099609375,
0.37060546875,
0.0555419921875,
0.37060546875,
0.0555419921875,
0.38818359375,
0.02777099609375,
0.38818359375
],
[
0.08087158203125,
0.37060546875,
0.468017578125,
0.37060546875,
0.468017578125,
0.3837890625,
0.08087158203125,
0.3837890625
],
[
0.50244140625,
0.369384765625,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50244140625,
0.38330078125
],
[
0.07843017578125,
0.38330078125,
0.46728515625,
0.3837890625,
0.46728515625,
0.3984375,
0.07843017578125,
0.397705078125
],
[
0.50244140625,
0.3837890625,
0.89208984375,
0.3837890625,
0.89208984375,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.443115234375,
0.02532958984375,
0.443115234375
],
[
0.08087158203125,
0.399658203125,
0.46728515625,
0.399658203125,
0.46728515625,
0.41357421875,
0.08087158203125,
0.41357421875
],
[
0.50341796875,
0.399658203125,
0.890625,
0.399658203125,
0.890625,
0.41357421875,
0.50341796875,
0.41357421875
],
[
0.07843017578125,
0.4140625,
0.4697265625,
0.41357421875,
0.4697265625,
0.427978515625,
0.07843017578125,
0.4287109375
],
[
0.50244140625,
0.415283203125,
0.890625,
0.415283203125,
0.890625,
0.4287109375,
0.50244140625,
0.4287109375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.429931640625,
0.890625,
0.429931640625,
0.890625,
0.443115234375,
0.50244140625,
0.443115234375
],
[
0.02777099609375,
0.442626953125,
0.052276611328125,
0.43994140625,
0.056365966796875,
0.472900390625,
0.0318603515625,
0.474853515625
],
[
0.08087158203125,
0.4462890625,
0.46728515625,
0.4462890625,
0.46728515625,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.50244140625,
0.4443359375,
0.890625,
0.44580078125,
0.890625,
0.458984375,
0.50244140625,
0.458251953125
],
[
0.080078125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.50244140625,
0.460205078125,
0.8896484375,
0.460205078125,
0.8896484375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.07843017578125,
0.473388671875,
0.4697265625,
0.474853515625,
0.4697265625,
0.4892578125,
0.07843017578125,
0.488037109375
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.473388671875,
0.89208984375,
0.488037109375,
0.50244140625,
0.4873046875
],
[
0.080078125,
0.490478515625,
0.468017578125,
0.490478515625,
0.468017578125,
0.50390625,
0.080078125,
0.50390625
],
[
0.50341796875,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50341796875,
0.50390625
],
[
0.0318603515625,
0.50048828125,
0.051483154296875,
0.50048828125,
0.051483154296875,
0.51806640625,
0.0318603515625,
0.51806640625
],
[
0.080078125,
0.505859375,
0.4697265625,
0.505859375,
0.4697265625,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50244140625,
0.51953125,
0.89208984375,
0.51953125,
0.89208984375,
0.53271484375,
0.50244140625,
0.53271484375
],
[
0.0310516357421875,
0.52587890625,
0.052276611328125,
0.52587890625,
0.052276611328125,
0.55322265625,
0.0310516357421875,
0.55322265625
],
[
0.080078125,
0.5361328125,
0.4697265625,
0.5361328125,
0.4697265625,
0.5498046875,
0.080078125,
0.5498046875
],
[
0.50244140625,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50244140625,
0.548828125
],
[
0.0294189453125,
0.55029296875,
0.053924560546875,
0.55029296875,
0.053924560546875,
0.5791015625,
0.0294189453125,
0.5791015625
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.50244140625,
0.548828125,
0.890625,
0.5498046875,
0.890625,
0.564453125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50244140625,
0.56494140625,
0.68701171875,
0.56494140625,
0.68701171875,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.07843017578125,
0.58056640625,
0.4697265625,
0.58056640625,
0.4697265625,
0.5947265625,
0.07843017578125,
0.5947265625
],
[
0.521484375,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.59521484375,
0.521484375,
0.59521484375
],
[
0.03350830078125,
0.59521484375,
0.037567138671875,
0.59521484375,
0.037567138671875,
0.6123046875,
0.03350830078125,
0.6123046875
],
[
0.037567138671875,
0.59033203125,
0.043304443359375,
0.59033203125,
0.043304443359375,
0.599609375,
0.037567138671875,
0.599609375
],
[
0.08087158203125,
0.59765625,
0.46728515625,
0.59765625,
0.46728515625,
0.60791015625,
0.08087158203125,
0.60791015625
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.080078125,
0.611328125,
0.4697265625,
0.611328125,
0.4697265625,
0.625,
0.080078125,
0.625
],
[
0.50341796875,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50341796875,
0.62548828125
],
[
0.03350830078125,
0.62255859375,
0.04412841796875,
0.626953125,
0.03594970703125,
0.6396484375,
0.02532958984375,
0.63525390625
],
[
0.03594970703125,
0.6142578125,
0.045745849609375,
0.6142578125,
0.045745849609375,
0.62548828125,
0.03594970703125,
0.62548828125
],
[
0.041656494140625,
0.62548828125,
0.047393798828125,
0.62548828125,
0.047393798828125,
0.6318359375,
0.041656494140625,
0.6318359375
],
[
0.080078125,
0.62548828125,
0.46728515625,
0.62548828125,
0.46728515625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50341796875,
0.62890625,
0.8896484375,
0.62890625,
0.8896484375,
0.6396484375,
0.50341796875,
0.6396484375
],
[
0.035125732421875,
0.63818359375,
0.051483154296875,
0.63818359375,
0.051483154296875,
0.66015625,
0.035125732421875,
0.66015625
],
[
0.080078125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.50244140625,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.0228729248046875,
0.65771484375,
0.052276611328125,
0.65380859375,
0.063720703125,
0.70068359375,
0.03350830078125,
0.705078125
],
[
0.080078125,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.50244140625,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.50244140625,
0.671875
],
[
0.080078125,
0.671875,
0.468017578125,
0.671875,
0.468017578125,
0.68505859375,
0.080078125,
0.68505859375
],
[
0.5009765625,
0.67236328125,
0.82861328125,
0.671875,
0.82861328125,
0.68603515625,
0.5009765625,
0.6875
],
[
0.08251953125,
0.68798828125,
0.46728515625,
0.68798828125,
0.46728515625,
0.69873046875,
0.08251953125,
0.69873046875
],
[
0.521484375,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.521484375,
0.7021484375
],
[
0.08087158203125,
0.7021484375,
0.46728515625,
0.7021484375,
0.46728515625,
0.71630859375,
0.08087158203125,
0.71630859375
],
[
0.50341796875,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.080078125,
0.71728515625,
0.2548828125,
0.71728515625,
0.2548828125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.50341796875,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.50341796875,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.08087158203125,
0.76123046875,
0.2091064453125,
0.76123046875,
0.2091064453125,
0.77587890625,
0.08087158203125,
0.77587890625
],
[
0.50341796875,
0.7646484375,
0.8896484375,
0.7646484375,
0.8896484375,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.50341796875,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.7861328125,
0.468017578125,
0.7861328125,
0.468017578125,
0.80029296875,
0.080078125,
0.80029296875
],
[
0.50244140625,
0.79443359375,
0.890625,
0.79443359375,
0.890625,
0.80810546875,
0.50244140625,
0.80810546875
],
[
0.080078125,
0.8017578125,
0.468017578125,
0.8017578125,
0.468017578125,
0.81494140625,
0.080078125,
0.81494140625
],
[
0.50341796875,
0.81005859375,
0.8896484375,
0.81005859375,
0.8896484375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.080078125,
0.8173828125,
0.4697265625,
0.8173828125,
0.4697265625,
0.83056640625,
0.080078125,
0.83056640625
],
[
0.50244140625,
0.82470703125,
0.89208984375,
0.82470703125,
0.89208984375,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.080078125,
0.83349609375,
0.468017578125,
0.83349609375,
0.468017578125,
0.8466796875,
0.080078125,
0.8466796875
],
[
0.50341796875,
0.84130859375,
0.5908203125,
0.84130859375,
0.5908203125,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.08087158203125,
0.84765625,
0.468017578125,
0.84765625,
0.468017578125,
0.861328125,
0.08087158203125,
0.861328125
],
[
0.51953125,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.51953125,
0.869140625
],
[
0.080078125,
0.86279296875,
0.4697265625,
0.86279296875,
0.4697265625,
0.876953125,
0.080078125,
0.876953125
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.1004638671875,
0.88720703125,
0.31787109375,
0.88720703125,
0.31787109375,
0.90087890625,
0.1004638671875,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation",
"Vinkle Srivastav1,2*",
"Keqi Chenl*",
"Nicolas Padoyl,2",
"1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France",
"21HU Strasbourg, France",
"srivastav@unistra.fr",
"keqi.chen@unistra.fr",
"npadoy@unistra.fr",
"",
"Abstract",
"methods and optimization-based methods.The learning",
"based methods develop novel deep-learning models and use",
"We present a new self-supervised approach, SelfPose3d,",
"3d ground-truth poses for both training the models and es-",
"for estimating 3d poses of multiple persons from multi-",
"tablishing person correspondences across different views",
"ple camera views.Unlike current state-of-the-art fully-",
"[39, 52, 57, 59, 63]. The accurate 3d ground-truth poses",
"2",
"supervised methods, our approach does not require any",
"are typically generated using a dense camera system [32].",
"2d or 3d ground-truth poses and uses only the multi-",
"In contrast, the optimization-based methods formulate the",
"LAD",
"view input images from a calibrated camera setup and 2d",
"3d pose reconstruction as a mathematical optimization task,",
"pseudo poses generated from an off-the-shelf 2d human",
"primarily focusing on aligning and matching the 2d poses",
"pose estimator:We propose two self-supervised learning",
"across different camera views to infer 3d poses using tri-",
"'SO",
"objectives:self-supervisedpersonlocalizationin3dspace",
"angulation within the framework of multi-view geometry",
"and self-supervised 3d pose estimation. We achieve self-",
"[10, 18, 31-33, 49]. The 2d poses are estimated using off-",
"supervised 3d person localization by training the model",
"the-shelf 2d human pose detectors [5, 55]. These methods",
"on synthetically generated 3d points, serving as 3d per-",
"apply geometric and spatial constraints in the optimization",
"V",
"son root positions, and on the projected root-heatmaps in",
"loop to ensure the anatomical plausibility and consistency",
"all the views. We then model the 3d poses of all the lo-",
"of the inferred 3d poses. Although these methods do not",
"1",
"calized persons with a bottleneck representation, map them",
"require 3d ground-truth poses, their effectiveness is some-",
"07",
"onto all views obtaining 2d joints, and render them using",
"what limited compared to the fully-supervised learning-",
"2d Gaussian heatmaps in an end-to-end differentiable man-",
"based methods, see Table 1.",
"ner Afterwards, we use the corresponding 2d joints and",
"In this paper, we explore the possibility of combining the",
"",
"",
"heatmapsfromthepseudo2dposesforlearning.Toal-",
"strengths of both paradigms.Specifically, we investigate",
"leviate the intrinsic inaccuracy of the pseudo labels, we",
"whether it's feasible to utilize a learning-based model for",
"U",
"7",
"",
"propose an adaptive supervision attention mechanism to",
"multi-view,multi-person3Dposeestimationandsimultane-",
".A",
"guide the self-supervision. Our experiments and analy-",
"ously eliminate its dependence on 3D ground-truth poses by",
"IXe",
"sis on three public benchmark datasets, including Panop-",
"incorporating geometric and appearance constraints, draw-",
"tic, Shelf, and Campus, show the effectiveness of our ap-",
"ing inspiration from optimization-based methods.",
"proach,whichiscomparabletofully-supervisedmethods.",
"We propose, SelfPose3d, a self-supervised learning-",
"Code is available at https://github.Com/CAMMA-",
"based approach to estimate the 3d poses of multiple persons",
"public/SelfPose3D.",
"from a few calibrated cameras without using any 2d or 3d",
"ground-truth poses. Our approach requires only 2d pseudo",
"poses obtained using an off-the-shelf 2d pose detector [55].",
"1. Introduction",
"Learning 3d poses without 3D ground-truth poses would re-",
"quire suitable supervisory signals to train a learning-based",
"The task of estimating 3d poses for multiple persons us-",
"model.We follow the learning-by-projection paradigm",
"ing a few calibrated cameras is a challenging computer vi-",
"where the main idea is to learn the 3d output by compar-",
"sion problem [10, 18, 32, 52, 57]. A significant part of this",
"ing the projected bottleneck 3d output against the 2d input",
"challenge lies in identifying and matching the same person",
"features [12].",
"across different camera views. The solutions developed so",
"We consider VoxelPose [57] as a learning-based method",
"far generally use one of the two paradigms: learning-based",
"and use its output 3d poses as a bottleneck representation.",
"co-first authors with equal contributions.",
"To recover the accurate underlying 3d poses, we propose us-"
] |
{
"arxiv_id": "2404.02041",
"language": "en",
"page_idx": 0
}
|
||
2404.02041-en-1
|
2404.02041-en-1.jpg
|
<FigureHere>
Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses generated using an $off-the-shelf$ 2d human pose estimator. We propose a self-supervised learning objective that generates differentiable and geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs from our approach along with estimated body meshes (using SMPL body mesh fitting on 3d poses[4,41]) and the projected 2d poses.
ones as the attention heatmap labels and use $L_2$ loss as the
attention loss.
Finally, specific to our choice of learning-based method,
ing differentiable multi-view 2d representations and crossaffine-view consistency. In particular, given a multi-view input image, we apply two random affine augmentations and pass them to the VoxelPose. It generates the bottleneck 3d poses corresponding to each affine augmented multi-view image. To enforce the model to learn and reason in the spatial dimension, we project the bottleneck 3d poses onto each view, obtaining 2d joints, and rendering them into spatial 2d heatmap representations in an end-to-end differentiable way. We further put tight geometric constraints by crossaffine-view operation, $i.e.$ the bottleneck 3d poses from the 1st affine augmented multi-view image is mapped and rendered in the 2nd affine augmented multi-view image space and vice versa. Finally, we use the affine transformed 2d joints and heatmaps from the 2d pseudo poses to enable the geometrically constrained learning, with $L_1$ and $L_2$ losses respectively.
$i.e.$, VoxelPose, which uses a voxel-based 3d root localization model to localize the persons in space using groundtruth 3d root joints (mid-hip joint), we use a simple but effective strategy to localize persons in space. Specifically, we randomly place 3d points in 3d world-space and project them to each view using the given camera parameters, subsequently rendering the projected 2d points as heatmap representations. This generates a synthetic dataset containing 3d points (roots) and their corresponding rendered multiview root-heatmaps. We then use this dataset to train a 3d root localization model that takes multi-view root-heatmaps as input and predicts the 3d roots as output. We further regularize the model by enforcing invariant constraints between pairs of affine augmented root-heatmaps coming from the real multi-view input.
Evaluation on three 3d pose benchmarks datasets, Panoptic [32], Shelf [1] and Campus [1], along with extensive ablation studies on the Panoptic[32] dataset, show the effectiveness of our approach. Our approach reaches a performance comparable to learning-based fullysupervised approaches and performs significantly better than optimization-based approaches. Moreover, SMPL body mesh ftting [4,41] on our estimated 3d poses generates geometrically plausible body shapes (see Figure l and Figure 4).
We summarize our contributions as follows: 1) We address the challenging multi-person multi-view 3d person pose estimation problem using a self-supervised approach
As the 2d pseudo poses contain non-negligible noises (mostly due to occlusions, see Figure 3), we propose $adap-$ tive supervision attention to guide our model to focus on more reliable regions. We apply two strategies towards $L_1$ joint loss and $L_2$ heatmap loss; for $L_1$ joint loss supervision, we employ hard attention, where we ignore the one view with the largest absolute error for each multi-view image set; for $L_2$ heatmap loss supervision, we employ soft attention using a lighter backbone to process each view, obtaining same-size attention heatmaps. During $L_2$ loss computation, we compute the element-wise product of the attention heatmaps and the square error before averaging. To avoid obtaining zero attention, which the model tends to do, we add a regularization term, where we create tensors of all
|
[
[
0.1282958984375,
0.1363525390625,
0.1405029296875,
0.1363525390625,
0.1405029296875,
0.1395263671875,
0.1282958984375,
0.1395263671875
],
[
0.1895751953125,
0.1363525390625,
0.2197265625,
0.1363525390625,
0.2197265625,
0.14013671875,
0.1895751953125,
0.14013671875
],
[
0.295654296875,
0.1363525390625,
0.34228515625,
0.1363525390625,
0.34228515625,
0.14013671875,
0.295654296875,
0.14013671875
],
[
0.3857421875,
0.1363525390625,
0.416748046875,
0.1363525390625,
0.416748046875,
0.14013671875,
0.3857421875,
0.14013671875
],
[
0.57666015625,
0.143310546875,
0.6259765625,
0.143310546875,
0.6259765625,
0.15087890625,
0.57666015625,
0.15087890625
],
[
0.7099609375,
0.14453125,
0.73876953125,
0.14453125,
0.73876953125,
0.1483154296875,
0.7099609375,
0.1483154296875
],
[
0.740234375,
0.14453125,
0.7646484375,
0.14453125,
0.7646484375,
0.1483154296875,
0.740234375,
0.1483154296875
],
[
0.253173828125,
0.1641845703125,
0.281982421875,
0.1641845703125,
0.281982421875,
0.1685791015625,
0.253173828125,
0.1685791015625
],
[
0.1846923828125,
0.20263671875,
0.200927734375,
0.20263671875,
0.200927734375,
0.205810546875,
0.1846923828125,
0.205810546875
],
[
0.221435546875,
0.20263671875,
0.2279052734375,
0.20263671875,
0.2279052734375,
0.205810546875,
0.221435546875,
0.205810546875
],
[
0.2484130859375,
0.20263671875,
0.314453125,
0.20263671875,
0.314453125,
0.205810546875,
0.2484130859375,
0.205810546875
],
[
0.2034912109375,
0.2734375,
0.2303466796875,
0.2734375,
0.2303466796875,
0.281005859375,
0.2034912109375,
0.281005859375
],
[
0.29736328125,
0.271484375,
0.328369140625,
0.271484375,
0.328369140625,
0.28271484375,
0.29736328125,
0.28271484375
],
[
0.080078125,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.08087158203125,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.08087158203125,
0.37890625,
0.89208984375,
0.37890625,
0.89208984375,
0.392578125,
0.08087158203125,
0.392578125
],
[
0.080078125,
0.393310546875,
0.86767578125,
0.393310546875,
0.86767578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.08087158203125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.5009765625,
0.42236328125,
0.890625,
0.42236328125,
0.890625,
0.436279296875,
0.5009765625,
0.436279296875
],
[
0.07843017578125,
0.438232421875,
0.4697265625,
0.4375,
0.4697265625,
0.451904296875,
0.07843017578125,
0.45263671875
],
[
0.50244140625,
0.439453125,
0.5947265625,
0.439453125,
0.5947265625,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.521484375,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.521484375,
0.47021484375
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50244140625,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.08087158203125,
0.50048828125,
0.465576171875,
0.50048828125,
0.465576171875,
0.51123046875,
0.08087158203125,
0.51123046875
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.080078125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.50244140625,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.08087158203125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.50341796875,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.08087158203125,
0.56103515625,
0.46728515625,
0.56103515625,
0.46728515625,
0.5712890625,
0.08087158203125,
0.5712890625
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.08087158203125,
0.57568359375,
0.46728515625,
0.57568359375,
0.46728515625,
0.5859375,
0.08087158203125,
0.5859375
],
[
0.5009765625,
0.5771484375,
0.89208984375,
0.57763671875,
0.89208984375,
0.59228515625,
0.5009765625,
0.59130859375
],
[
0.08087158203125,
0.5888671875,
0.468017578125,
0.5888671875,
0.468017578125,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.50244140625,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.07843017578125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.07843017578125,
0.61865234375
],
[
0.50244140625,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.080078125,
0.619140625,
0.468017578125,
0.619140625,
0.468017578125,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.080078125,
0.63525390625,
0.468017578125,
0.63525390625,
0.468017578125,
0.6484375,
0.080078125,
0.6484375
],
[
0.50244140625,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.080078125,
0.6494140625,
0.4697265625,
0.6494140625,
0.4697265625,
0.66357421875,
0.080078125,
0.66357421875
],
[
0.50341796875,
0.65478515625,
0.890625,
0.65478515625,
0.890625,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.080078125,
0.66650390625,
0.160888671875,
0.66650390625,
0.160888671875,
0.67822265625,
0.080078125,
0.67822265625
],
[
0.50341796875,
0.66845703125,
0.8935546875,
0.66845703125,
0.8935546875,
0.681640625,
0.50341796875,
0.681640625
],
[
0.098876953125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50244140625,
0.6845703125,
0.64453125,
0.6845703125,
0.64453125,
0.6943359375,
0.50244140625,
0.6943359375
],
[
0.08087158203125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.521484375,
0.7021484375,
0.8896484375,
0.7021484375,
0.8896484375,
0.71533203125,
0.521484375,
0.71533203125
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.50244140625,
0.74560546875
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.5009765625,
0.7470703125,
0.890625,
0.7470703125,
0.890625,
0.76025390625,
0.5009765625,
0.76025390625
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76220703125,
0.89208984375,
0.76220703125,
0.89208984375,
0.77587890625,
0.50244140625,
0.77587890625
],
[
0.080078125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50341796875,
0.77783203125,
0.89208984375,
0.77783203125,
0.89208984375,
0.79150390625,
0.50341796875,
0.79150390625
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.79248046875,
0.89208984375,
0.79150390625,
0.89208984375,
0.80615234375,
0.5009765625,
0.806640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.890625,
0.80810546875,
0.890625,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82275390625,
0.89208984375,
0.82275390625,
0.89208984375,
0.8369140625,
0.50244140625,
0.8369140625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.8388671875,
0.56884765625,
0.8388671875,
0.56884765625,
0.85107421875,
0.50341796875,
0.85107421875
],
[
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
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"EE",
"tion",
"andomaffi",
"atiant",
"outputfro",
"codan",
"nnaae",
"mato",
"",
"12",
"rossaooaw",
"L,loss",
"L,loss",
"Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses",
"generated using an off-the-shelf 2d human pose estimator.Weproposea self-supervised learning objective thatgenerates differentiableand",
"geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs",
"from our approach along with estimated body meshes (using SMPL body mesh ftting on 3d poses [4, 41]) and the projected 2d poses.",
"ing differentiable multi-view 2d representations and cross-",
"ones as the attention heatmap labels and use L, loss as the",
"affine-view consistency. In particular, given a multi-view in-",
"attentionloss.",
"put image, we apply two random affine augmentations and",
"Finally, specific to our choice of learning-based method,",
"pass them to the VoxelPose. It generates the bottleneck 3d",
"i.e., VoxelPose, which uses a voxel-based 3d root localiza-",
"poses corresponding to each affine augmented multi-view",
"tion model to localize the persons in space using ground-",
"image.To enforce themodel to learn andreasoninthe spa-",
"truth 3d root joints (mid-hip joint), we use a simple but ef-",
"tial dimension, we project the bottleneck 3d poses onto each",
"fective strategy to localize persons in space. Specifically,",
"view, obtaining 2d joints, and rendering them into spatial",
"werandomlyplace3dpointsin3dworld-spaceandproject",
"2d heatmap representations in an end-to-end differentiable",
"them to each view using the given camera parameters, sub-",
"way.Wefurtherputtightgeometricconstraintsby cross-",
"sequently rendering the projected 2d points as heatmap rep-",
"affine-viewoperation.i.e.thebottleneck3dposesfromthe",
"resentations. This generates a synthetic dataset containing",
"Ist affine augmented multi-view image is mapped and ren-",
"3d points (roots) and their corresponding rendered multi-",
"dered in the 2nd affine augmented multi-view image space",
"view root-heatmaps. We then use this dataset to train a 3d",
"and vice versa. Finally, we use the affine transformed 2d",
"root localization model that takes multi-view root-heatmaps",
"joints and heatmaps from the 2d pseudo poses to enable the",
"as input and predicts the 3d roots as output. We further regu-",
"geometrically constrained learning, with L and L2 losses",
"larizethemodelbyenforcing invariantconstraintsbetween",
"respectively",
"pairs of affine augmented root-heatmaps coming from the",
"As the 2d pseudo poses contain non-negligible noises",
"realmulti-viewinput",
"(mostly due to occlusions, see Figure 3), we propose adap-",
"Evaluation on three 3d pose benchmarks datasets.",
"tive supervision attention to guide our model to focus on",
"Panoptic [32], Shelf [1] and Campus [1], along with",
"more reliable regions. We apply two strategies towards Li",
"extensive ablation studies on the Panoptic [32] dataset,",
"joint loss and L2 heatmap loss; for L1 joint loss supervi-",
"show the effectiveness of our approach.Our approach",
"sion, we employ hard attention, where we ignore the one",
"reaches a performance comparable to learning-based fully-",
"view with the largest absolute error for each multi-view im-",
"supervised approaches and performs significantly better",
"age set; for L2 heatmap loss supervision, we employ soft",
"than optimization-based approaches.Moreover, SMPL",
"attention using a lighter backbone to process each view, ob-",
"body mesh fitting [4, 41] on our estimated 3d poses gener-",
"taining same-size attention heatmaps. During L2 loss com-",
"ates geometrically plausible body shapes (see Figure I and",
"putation, we compute the element-wise product of the at-",
"Figure 4).",
"tention heatmaps and the square error before averaging. To",
"We summarize our contributions as follows: 1) We ad-",
"avoid obtaining zero attention, which the model tends to do,",
"dress the challenging multi-person multi-view 3d person",
"we add a regularization term, where we create tensors of all",
"pose estimation problem using a self-supervised approach"
] |
{
"arxiv_id": "2404.02041",
"language": "en",
"page_idx": 1
}
|
||
2404.02041-en-2
|
2404.02041-en-2.jpg
|
without any 2d or 3d ground truth. 2) We propose selfsupervised 3d pose estimation by using a new method to recover geometrically constrained 2d joints and heatmap representations from the bottleneck 3d poses. 3) We propose adaptive supervision attention to address the misinformation caused by the inaccurate pseudo labels. 4) We propose self-supervised 3d root localization to estimate the 3d root location utilizing synthetic 3d roots and the corresponding rendered multi-view root heatmaps.
## $2. \textbf{ Related work}$
In this section, we briefly review current works related to fully-supervised learning-based methods for 3d pose estimation, optimization-based methods for 3D pose estimation, and self-supervised learning.
Fully-supervised methods: Monocular 3d pose estimation [22,43,45,47,50,56,62,65] is an ill-posed problem due to depth ambiguities as multiple 3d poses can produce same $2d$ pose projection. Having access to multi-view cameras can alleviate such depth ambiguities achieving the state-ofart results on benchmark datasets [23,27,28,39,51,52,57, 59,63]. For single-person scenes, these approaches exploit multi-view geometry[24] to either fuse the visual features [27,51],perform triangulation on heatmaps [28,53], or use pictorial structural models for 3d reconstruction [48,51]. The multi-person scene offers extra complexity due to the variability in the number of person in each view and the unknown cross-view correspondence. Existing multi-person multi-view approaches are based on volumetric paradigm [52,57,64], or direct regression [63] based on transformers [6,58,66]. Despite their good performance, these approaches rely on ground-truth 3d poses, which are generated using dense camera systems [32].
Optimization-based 3d pose estimation: For the multiperson and multi-view scenario, optimization-based approaches use an $off-the-shelf$ person-id detector across all the views to solve the correspondence and triangulation problem and temporal refinement along with training a reinforcement learning agent to find the best camera locations for 3d pose reconstruction [49]. More recent approaches utilize multi-view 3d reconstruction in the optimization loop inferring 3D poses that are geometrically and spatially coherent [10,18,33].
Self-supervised learning: Self-supervised learning can be broadly classified into self-supervised representation learning and self-supervised task learning. Self-supervised representation learning aims to use large-scale unlabeled data to learn generic feature representations. The recent promising results from these approaches have started to surpass the fully-supervised baselines for various downstream tasks [7,11,26,54]. Self-supervised task learning aims to learn a particular downstream task without using ground truth labels and has been applied to 2d pose estimation [29,30],
single-person 3d pose estimation [9,19,34,36,38], and surface correspondences estimation [3]. Self-supervised approaches for 3d pose estimation have primarily been developed for single-person scenarios. Given 2d poses, estimated by utilizing advances in the 2d pose estimation methods [5,13,14,21,35,42,44,46,55,60], these approaches use the supervisory signals generated from multi-view geometry [34], video constraints [38], or adversarial learning [9,19,36].
Our work proposes a learning-based approach to model the 3d poses as bottleneck representations and recover geometrically constrained and spatially accurate 2d joints and heatmap representations in an end-to-end differentiable manner.
# 3. Methodology
### $3. 1. \textbf{Problem overview}$
Given a training dataset of multi-view images $\mathcal{D}=\{\mathbf{x}|\mathbf{y}^*\}$ where $\mathbf{x}\in\mathcal{R}^{\breve{C}\times3\times H\times W}$ is a multi-view image set from $C$ cameras with height $H$ and width $W$, and $\mathbf{y}^*\in$ $\mathcal{R}^{C\times P\times J\times2}$ represents the 2d pseudo poses for $P$ persons with $J$ joints, the goal is to learn a deep learning model that estimates the 3d poses $\mathcal{Y}\in\mathcal{R}^{P\times J\times3}$ of all the $P$ persons from the multi-view input image x. It is to be noted that $P$ can vary in each camera view due to occlusion and noisy pseudo 2d pose estimation. For simplicity in the notation, we keep the same variable P.
Fully-supervised approaches rely on 3d ground-truth $\begin{array}{l}\text{poses,while we only have 2d pseudo poses y}^{*}.&\text{Therefore,}\\\text{after obtaining 3d poses }\mathcal{Y}\in\mathcal{R}^{P\times J\times3}&\text{following traditional}\end{array}$ approach, we propose to project the poses to each view ob- taining 2d poses y$\in \mathcal{R} ^{C\times P\times J\times 2}$,and train the model from taining 2d poses y$\in \mathcal{R} ^{C\times P\times J\times 2}$,and train the model from 2d pseudo poses $\mathbf{y}^*\in\mathcal{R}^{C\times P\times J\times2}.$
In the following, we present our self-supervised approach based on fully-supervised VoxelPose[57]. We first generate pseudo 2d poses, and then propose self-supervised 3d root localization, self-supervised 3d pose estimation, and an adaptive supervision attention to learn 3d poses in a selfsupervised manner, without modifying the original VoxelPose structure.
# 3.2. Generating pseudo 2d poses
To circumvent the dependence on the ground-truth 2d poses, we generate the 2d pseudo poses on the training dataset using Mask R-CNN [25] to generate person bounding boxes followed by using HRNet[55] to generate 2d poses of each detected person bounding box. The two-stage approach is chosen based on its state-of-art performance on the COCO dataset [40]. We first pre-train the 2d CNN backbone heatmap\_net$_\mathrm{2d}$ with pseudo 2d poses.
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.07843017578125,
0.10797119140625,
0.468017578125,
0.106689453125,
0.468017578125,
0.12060546875,
0.07843017578125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12371826171875,
0.890625,
0.12371826171875,
0.890625,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.136962890625,
0.8935546875,
0.136962890625,
0.8935546875,
0.15087890625,
0.50244140625,
0.15087890625
],
[
0.080078125,
0.154052734375,
0.4697265625,
0.154052734375,
0.4697265625,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50341796875,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08087158203125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.50341796875,
0.18310546875,
0.8896484375,
0.18310546875,
0.8896484375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.07843017578125,
0.1962890625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.212158203125,
0.07843017578125,
0.2108154296875
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.07843017578125,
0.2108154296875,
0.314453125,
0.21337890625,
0.314453125,
0.2279052734375,
0.07843017578125,
0.2259521484375
],
[
0.50341796875,
0.21337890625,
0.57666015625,
0.21337890625,
0.57666015625,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.52294921875,
0.2310791015625,
0.89208984375,
0.2310791015625,
0.89208984375,
0.24560546875,
0.52294921875,
0.24560546875
],
[
0.080078125,
0.2412109375,
0.214111328125,
0.2412109375,
0.214111328125,
0.255615234375,
0.080078125,
0.255615234375
],
[
0.5009765625,
0.24560546875,
0.890625,
0.246826171875,
0.890625,
0.261474609375,
0.5009765625,
0.260009765625
],
[
0.080078125,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.27978515625,
0.080078125,
0.27978515625
],
[
0.50244140625,
0.261962890625,
0.89208984375,
0.261962890625,
0.89208984375,
0.27587890625,
0.50244140625,
0.27587890625
],
[
0.50244140625,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.291748046875,
0.50244140625,
0.291748046875
],
[
0.080078125,
0.28271484375,
0.468017578125,
0.28271484375,
0.468017578125,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.080078125,
0.29736328125,
0.4697265625,
0.29736328125,
0.4697265625,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.50244140625,
0.29345703125,
0.55615234375,
0.29345703125,
0.55615234375,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.080078125,
0.313232421875,
0.30224609375,
0.313232421875,
0.30224609375,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.08087158203125,
0.32763671875,
0.468017578125,
0.32763671875,
0.468017578125,
0.341552734375,
0.08087158203125,
0.341552734375
],
[
0.5009765625,
0.32568359375,
0.6337890625,
0.32763671875,
0.6337890625,
0.34521484375,
0.5009765625,
0.343505859375
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.080078125,
0.35791015625,
0.4697265625,
0.35791015625,
0.4697265625,
0.371826171875,
0.080078125,
0.371826171875
],
[
0.50244140625,
0.35595703125,
0.67333984375,
0.35595703125,
0.67333984375,
0.37060546875,
0.50244140625,
0.37060546875
],
[
0.080078125,
0.373779296875,
0.4697265625,
0.373779296875,
0.4697265625,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.5009765625,
0.3818359375,
0.8896484375,
0.38330078125,
0.8896484375,
0.397705078125,
0.5009765625,
0.396484375
],
[
0.07843017578125,
0.38818359375,
0.4697265625,
0.386962890625,
0.4697265625,
0.401611328125,
0.07843017578125,
0.40283203125
],
[
0.5009765625,
0.393310546875,
0.89453125,
0.393310546875,
0.89453125,
0.4140625,
0.5009765625,
0.4140625
],
[
0.08087158203125,
0.40478515625,
0.46728515625,
0.40478515625,
0.46728515625,
0.415283203125,
0.08087158203125,
0.415283203125
],
[
0.50244140625,
0.410888671875,
0.89208984375,
0.412353515625,
0.89208984375,
0.4267578125,
0.50244140625,
0.425537109375
],
[
0.08087158203125,
0.42041015625,
0.46728515625,
0.42041015625,
0.46728515625,
0.43115234375,
0.08087158203125,
0.43115234375
],
[
0.5009765625,
0.423583984375,
0.8935546875,
0.427978515625,
0.8935546875,
0.44580078125,
0.5009765625,
0.44140625
],
[
0.08087158203125,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.44580078125,
0.08087158203125,
0.44580078125
],
[
0.08087158203125,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.50341796875,
0.443115234375,
0.89208984375,
0.443115234375,
0.89208984375,
0.45703125,
0.50341796875,
0.45703125
],
[
0.498291015625,
0.45263671875,
0.89453125,
0.453857421875,
0.89453125,
0.474853515625,
0.498291015625,
0.473388671875
],
[
0.080078125,
0.462158203125,
0.468017578125,
0.46337890625,
0.468017578125,
0.477783203125,
0.080078125,
0.4765625
],
[
0.08087158203125,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.50244140625,
0.473388671875,
0.89208984375,
0.473388671875,
0.89208984375,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.50341796875,
0.4892578125,
0.890625,
0.4892578125,
0.890625,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.08087158203125,
0.49560546875,
0.465576171875,
0.49560546875,
0.465576171875,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.5048828125,
0.52099609375,
0.69140625,
0.52099609375,
0.69140625,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.080078125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.521484375,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.55029296875,
0.521484375,
0.55029296875
],
[
0.080078125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.080078125,
0.568359375
],
[
0.50341796875,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.07843017578125,
0.568359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.583984375,
0.07843017578125,
0.58349609375
],
[
0.50244140625,
0.564453125,
0.89208984375,
0.564453125,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.58544921875,
0.330078125,
0.58544921875,
0.330078125,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50244140625,
0.58349609375,
0.890625,
0.58056640625,
0.890625,
0.5947265625,
0.50244140625,
0.5966796875
],
[
0.08087158203125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.5009765625,
0.5947265625,
0.8935546875,
0.5947265625,
0.8935546875,
0.6123046875,
0.5009765625,
0.6123046875
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.498291015625,
0.611328125,
0.73876953125,
0.60498046875,
0.73876953125,
0.62548828125,
0.498291015625,
0.6318359375
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.521484375,
0.630859375,
0.88818359375,
0.630859375,
0.88818359375,
0.64453125,
0.521484375,
0.64453125
],
[
0.08087158203125,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.50341796875,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.6611328125,
0.89208984375,
0.6611328125,
0.89208984375,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.07843017578125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.07843017578125,
0.7041015625
],
[
0.50244140625,
0.69140625,
0.890625,
0.69140625,
0.890625,
0.705078125,
0.50244140625,
0.705078125
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.5009765625,
0.70556640625,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.5009765625,
0.72021484375
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50341796875,
0.72265625,
0.6005859375,
0.72265625,
0.6005859375,
0.734375,
0.50341796875,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.279296875,
0.7373046875,
0.279296875,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.5009765625,
0.751953125,
0.7509765625,
0.75244140625,
0.7509765625,
0.77099609375,
0.5009765625,
0.76953125
],
[
0.080078125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.8896484375,
0.81103515625,
0.8896484375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.8935546875,
0.85693359375,
0.8935546875,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.7958984375,
0.88720703125,
0.7958984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"without any 2d or 3d ground truth. 2) We propose self-",
"single-person 3d pose estimation [9,19,34,36, 38], and",
"supervised 3d pose estimation by using a new method to re-",
"surface correspondences estimation [3]. Self-supervised ap-",
"cover geometrically constrained 2d joints and heatmap rep-",
"proaches for 3d pose estimation have primarily been devel-",
"resentations from the bottleneck 3d poses. 3)We propose",
"oped for single-person scenarios. Given 2d poses, estimated",
"adaptive supervision attention to address the misinforma-",
"by utilizing advances in the 2d pose estimation methods",
"tioncausedbytheinaccuratepseudolabels.4)Wepropose",
"[5,13,14,21,35,42,44,46,55,60],these approaches",
"self-supervised 3d root localization to estimate the 3d root",
"use the supervisory signals generated from multi-view ge-",
"location utilizing synthetic 3d roots and the corresponding",
"ometry [34], video constraints [38], or adversarial learning",
"rendered multi-view root heatmaps.",
"[9, 19, 36].",
"Our work proposes a learning-based approach to model",
"2. Related work",
"the 3d poses as bottleneck representations and recover ge-",
"In this section, we briefly review current works related to",
"ometrically constrained and spatially accurate 2d joints",
"and heatmap representations in an end-to-end differentiable",
"fully-supervised learning-based methods for 3d pose esti-",
"mation, optimization-based methods for 3D pose estima-",
"manner.",
"tion, and self-supervised learning.",
"Fully-supervised methods: Monocular 3d pose estimation",
"3. Methodology",
"[22, 43, 45, 47, 50, 56, 62, 65] is an ill-posed problem due",
"to depth ambiguities as multiple 3d poses can produce same",
"3.1. Problem overview",
"2d pose projection. Having access to multi-view cameras",
"Given a training dataset of multi-view images D = [x|y*]",
"can alleviate such depth ambiguities achieving the state-of-",
"where x e RCx3xHxW is a multi-view image set from",
"artresults onbenchmarkdatasets[23.27.28,39,51.52.57.",
"C cameras with height H and width W, and y* E",
"59,63l.Forsingle-personscenes,theseapproachesexploit",
"RC× PxJ×2 represents the 2d pseudo poses for P persons",
"multi-view geometry[24]to eitherfuse the visualfeatures",
"[27, 51], perform triangulation on heatmaps [28, 53], or use",
"with J joints, the goal is to learn a deep learning model that",
"estimates the 3d poses y e RPxJx3 of all the P persons",
"pictorial structural models for 3d reconstruction [48, 51].",
"The multi-person scene offers extra complexity due to the",
"from the multi-view input image x. It is to be noted that",
"P can vary in each camera view due to occlusion and noisy",
"variabilityinthenumberofpersonineachview andtheun-",
"known cross-view correspondence. Existing multi-person",
"pseudo 2d pose estimation. For simplicity in the notation,",
"multi-view approaches are based on volumetric paradigm",
"wekeepthe samevariableP.",
"[52, 57, 64], or direct regression [63] based on transform-",
"Fully-supervised approaches rely on 3d ground-truth",
"ers [6, 58, 66]. Despite their good performance, these ap-",
"poses, while we only have 2d pseudo poses y*,Therefore,",
"proaches rely on ground-truth 3d poses, which are gener-",
"after obtaining 3d poses y e RPxJx3 following traditional",
"ated using dense camera systems [32]",
"approach,weproposetoprojecttheposestoeachviewob-",
"Optimization-based 3d pose estimation: For the multi-",
"taining 2d poses y E RCxPxJx2, and train the model from",
"person and multi-view scenario, optimization-based ap-",
"2d pseudo poses y* e RCxPxJx2",
"proaches use an off-the-shelf person-id detector across all",
"In the following,we present our self-supervised ap",
"the views to solve the correspondence and triangulation",
"proach based on fully-supervised VoxelPose [57]. We first",
"problem and temporal refinement along with training a re-",
"generate pseudo 2d poses, and then propose self-supervised",
"inforcement learning agent to find the best camera loca-",
"3d root localization, self-supervised 3d pose estimation, and",
"tions for 3d pose reconstruction [49]. More recent ap-",
"an adaptive supervision attention to learn 3d poses in a self-",
"proaches utilize multi-view 3d reconstruction in the opti-",
"supervised manner, without modifying the original Voxel-",
"mization loop inferring 3D poses that are geometrically and",
"Pose structure.",
"spatiallycoherent[10,18.33]",
"Self-supervised learning: Self-supervised learning can be",
"3.2. Generating pseudo 2d poses",
"broadly classified into self-supervised representation learn-",
"ing and self-supervised task learning. Self-supervised rep-",
"To circumvent the dependence on the ground-truth 2d",
"resentation learning aims to use large-scale unlabeled data",
"poses, we generate the 2d pseudo poses on the training",
"to learn generic feature representations. The recent promis-",
"dataset using Mask R-CNN [25] to generate person bound-",
"ing results from these approaches have started to surpass",
"ing boxes followed by using HRNet [55] to generate 2d",
"the fully-supervised baselines for various downstream tasks",
"poses of each detected person bounding box. The two-stage",
"[7, 11, 26, 54]. Self-supervised task learning aims to learn",
"approach is chosen based on its state-of-art performance on",
"a particular downstream task without using ground truth la-",
"the COCO dataset [4O]. We first pre-train the 2d CNN back-",
"bels and has been applied to 2d pose estimation [29, 30],",
"bone heatmap_net2d with pseudo 2d poses."
] |
{
"arxiv_id": "2404.02041",
"language": "en",
"page_idx": 2
}
|
||
2404.02041-en-4
|
2404.02041-en-4.jpg
|
$\left\{\mathcal{P}_{\mathrm{unproj}}(\operatorname{cam},r_{i},t_{r,s}^{2})\right\}_{i=1}^{P}$corresponding to the person feature volumes for each affine augmented multi-view input image $x^1$ and $x^2$, respectively. $\left \{ \mathcal{F} _{i}^{1}\right \} _{i= 1}^{P}$and $\left\{\mathcal{F}_i^2\right\}_{i=1}^P$are passed through pose\_net$_3$d and soft-argmax[8] to estimate the 3d poses $\mathcal{Y}^1,\mathcal{Y}^2\in\mathcal{R}^{P\times J\times3}.$ These 3d poses serve as a bottleneck representation.
Given the camera parameters cam and the affine trans- $\begin{array}{l}{\text{formation parameters }t_{r,s}^{1},t_{r,s}^{2},\text{we project the bottleneck 3d}}\\{\text{poses in cross-affine-view }i.e.~\mathcal{Y}^{1}}\end{array}$ age space using $t_r,s^2$ to generate multi-view 2d poses $\mathbf{y}^2\in$ $\mathcal{R}^{C\times P\times J\times2}$,and $\mathcal{Y}^2$ are projected to the $x^1$ image space us- $\operatorname{ing}t_{r,s}^{1}$ to generate multi-view 2d poses $\mathbf{y}^1\in\tilde{\mathcal{R}}^{C\times\hat{P}\times J\times2}.$
We propose to render y$^1$ and y$^2$ 2d poses in the 2d heatmap representation. The heatmap representation encodes the per-pixel likelihood of a body joint and has been a vital component to enable the state-of-the-art 2d pose estimation approaches [55]. Generating heatmap representation has essentially been a pre-processing step where stateof-the-art approaches quantize the 2d joints before generating the heatmap [55]. However, this quantization step is non-differentiable and could cut the backward gradient flow. Zhang et al.[61] show that encoding floating point 2d joints into heatmap representation in their pre-processing step leads to improved performance. We propose to use the same differentiable approach in an online way to render the projected 2d joints into the heatmap representation.
We render $\mathbf{y}^{1}$ and y$^2$ into heatmap representation to generate 2d multi-view heatmaps $\mathcal{H}^1$,and $\mathcal{H}^2$,respectively. We apply the affine transformations $t_{r,s}^1,t_{r,s}^2$ on the pseudo 2d poses $\mathbf{y}_{2\mathrm{d}}^{*}$ to generate pseudo 2d multi-view joints $\mathbf{y}_{2\mathrm{d}}^{1*},\mathbf{y}_{2\mathrm{d}}^{2*}$ and heatmaps $\mathcal{H}^{1*},\mathcal{H}^{\hat{2}*}.$ Then, we compute the $L_2$ loss between heatmaps as the pose heatmap loss $l_\mathrm{pose\_H}.$
$$l_{\mathrm{pose.H}}=\mathcal{L}_{2}(\mathcal{H}^{1},\mathcal{H}^{1*})+\mathcal{L}_{2}(\mathcal{H}^{2},\mathcal{H}^{2*})$$
(4)
After training with $l_\mathrm{pose\_H}$ preliminarily, we propose to add the $L_1$ loss between 2d joints to further fine-tune the model. For each view, we employ the Hungarian algorithm [37] to obtain the optimal assignment between y and $\mathbf{y}_{2\mathrm{d}}^{*}$, where the matching cost is the mean absolute error. Based on the optimal assignment, we obtain the $L_1$ loss as $l_{\mathrm{pose\_J.~Then~we~use~}l_{\mathrm{pose\_H~and~}}l_{\mathrm{pose\_J~together~to~train~the}}}$ whole network, where $\lambda$ is a manually defined weight:
(5)
$$l_{\mathrm{pose_J}}=\mathcal{L}_1(\mathbf{y}^1,\mathbf{y}_{2\mathrm{d}}^{1*})+\mathcal{L}_1(\mathbf{y}^2,\mathbf{y}_{2\mathrm{d}}^{2*})$$
$$l_{\mathrm{pose_3d}}=l_{\mathrm{pose_H}}+\lambda l_{\mathrm{pose_J}}$$
(6)
As the network needs to reason about the 2d joint locations in spatial dimension, it implicitly solves the person correspondence problem. Training pose net$_3$d with 3d pose loss $l_\mathrm{pose\_3d}$ performs decently, but to achieve even better results, we introduce the adaptive supervision attention.
<FigureHere>
Figure 3. Comparing ground-truth 2d poses generated by projecting the ground-truth 3d poses to each multi-view image and our pseudo 2d poses generated by running HRNet human pose estimation model [55] on the training dataset. Pseudo 2d poses contain localization errors due to occlusion (see the red arrows), and ground-truth 2d poses exist for partially or even entirely occluded persons (see the blue dotted arrows).
## 3.5. Adaptive supervision attention
Traditional $L_1$ and $L_2$ losses treat each label equally, which is sub-optimal in two aspects:(1) the 2d human pose detector may generate inaccurate labels due to occlusions (see the red arrows in Figure 3);(2) the 3d-2d projection will output 2d joints in certain views even when the person is entirely occluded (see the blue dotted arrows in Figure 3). Therefore, we propose to employ attentions to adaptively guide the supervision process.
For $L_2$ loss supervision, we use the soft attention. Specifically, we use ResNet-18 to extract the visual features of the views, followed by deconvolutional layers to obtain the attention heatmaps $\mathcal{A}($see attn net$_2$d in Figure 2). Then we compute the element-wise product of $\mathcal{A}$ and the square error before averaging, as the new loss $l_\mathrm{pose\_H}^\mathrm{attn}.$
$$l_{\text{pose H}}^{\text{attn}}=\dfrac{1}{N}\sum_{i=1}^{N}\mathcal{A}_{i}\otimes(\mathcal{H}_{i}-\mathcal{H}_{i}^{*})^{2}$$
(7)
To avoid that $\mathcal{A}$ becomes zero, we add a regularization term. We create tensors of all ones 1 as the atten- $\begin{array}{l}{\text{tion heatmap labels, and compute }L_{2}\mathrm{~error~as~the~attention}}\\{\mathrm{loss~}l_{attn}.\mathrm{~If~}l_{\mathrm{attn~}}\mathrm{~becomes~zero,~}l_{\mathrm{pose\_H}}^{\mathrm{attn}}\mathrm{~degrades~to~non-}}\end{array}$ attentive version.
$$l_{\mathrm{attn}}=\mathcal{L}_{2}(\mathcal{A},\mathbb{1})$$
(8)
For $L_1$ loss supervision, we use the hard attention. For each input with $K$ views, we compute the $L_1$ loss of each view, find the view with the largest loss, and ignore it when averaging the final loss $l_\mathrm{pose\_J}^\mathrm{attn}.$
$$j=\arg\max_{i}\mathcal{L}_{1}(\mathbf{y}_{i},\mathbf{y}_{i,2\text{d}}^{*})\quad(i=1,2,...,K)$$
(9)
|
[
[
0.08087158203125,
0.0921630859375,
0.25,
0.0889892578125,
0.25,
0.106689453125,
0.08087158203125,
0.10986328125
],
[
0.27294921875,
0.09344482421875,
0.46728515625,
0.09344482421875,
0.46728515625,
0.106689453125,
0.27294921875,
0.106689453125
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.0767822265625,
0.12371826171875,
0.471435546875,
0.12371826171875,
0.471435546875,
0.14453125,
0.0767822265625,
0.14453125
],
[
0.07843017578125,
0.14453125,
0.1298828125,
0.1395263671875,
0.13232421875,
0.1590576171875,
0.08251953125,
0.1641845703125
],
[
0.131591796875,
0.1458740234375,
0.468017578125,
0.1458740234375,
0.468017578125,
0.160400390625,
0.131591796875,
0.160400390625
],
[
0.0767822265625,
0.1590576171875,
0.471435546875,
0.1572265625,
0.471435546875,
0.177978515625,
0.0767822265625,
0.179931640625
],
[
0.080078125,
0.1767578125,
0.35546875,
0.1767578125,
0.35546875,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.09722900390625,
0.1932373046875,
0.46728515625,
0.1932373046875,
0.46728515625,
0.20703125,
0.09722900390625,
0.20703125
],
[
0.080078125,
0.208984375,
0.468017578125,
0.208984375,
0.468017578125,
0.222900390625,
0.080078125,
0.222900390625
],
[
0.080078125,
0.222900390625,
0.46728515625,
0.222900390625,
0.46728515625,
0.2374267578125,
0.080078125,
0.2374267578125
],
[
0.080078125,
0.238037109375,
0.46728515625,
0.238037109375,
0.46728515625,
0.25244140625,
0.080078125,
0.25244140625
],
[
0.0767822265625,
0.248779296875,
0.471435546875,
0.25244140625,
0.471435546875,
0.2734375,
0.0767822265625,
0.26953125
],
[
0.07598876953125,
0.267578125,
0.468017578125,
0.26318359375,
0.468017578125,
0.2841796875,
0.07598876953125,
0.28857421875
],
[
0.50244140625,
0.272705078125,
0.890625,
0.272705078125,
0.890625,
0.285888671875,
0.50244140625,
0.285888671875
],
[
0.098876953125,
0.285400390625,
0.468017578125,
0.2841796875,
0.468017578125,
0.298583984375,
0.098876953125,
0.2998046875
],
[
0.5048828125,
0.28857421875,
0.890625,
0.28857421875,
0.890625,
0.298583984375,
0.5048828125,
0.298583984375
],
[
0.080078125,
0.300537109375,
0.468017578125,
0.300537109375,
0.468017578125,
0.314453125,
0.080078125,
0.314453125
],
[
0.50244140625,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.314453125,
0.50244140625,
0.314453125
],
[
0.080078125,
0.31640625,
0.468017578125,
0.31640625,
0.468017578125,
0.330078125,
0.080078125,
0.330078125
],
[
0.50244140625,
0.314453125,
0.890625,
0.314453125,
0.890625,
0.32763671875,
0.50244140625,
0.32763671875
],
[
0.08087158203125,
0.333251953125,
0.46728515625,
0.333251953125,
0.46728515625,
0.343505859375,
0.08087158203125,
0.343505859375
],
[
0.50341796875,
0.330078125,
0.890625,
0.330078125,
0.890625,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.080078125,
0.3466796875,
0.4697265625,
0.3466796875,
0.4697265625,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.50244140625,
0.35595703125
],
[
0.080078125,
0.361083984375,
0.468017578125,
0.361083984375,
0.468017578125,
0.375,
0.080078125,
0.375
],
[
0.50244140625,
0.35595703125,
0.72119140625,
0.35595703125,
0.72119140625,
0.369384765625,
0.50244140625,
0.369384765625
],
[
0.07843017578125,
0.375732421875,
0.46728515625,
0.375,
0.46728515625,
0.389404296875,
0.07843017578125,
0.39013671875
],
[
0.07843017578125,
0.39013671875,
0.468017578125,
0.391357421875,
0.468017578125,
0.40478515625,
0.07843017578125,
0.404052734375
],
[
0.5009765625,
0.392578125,
0.76953125,
0.393310546875,
0.76953125,
0.40771484375,
0.5009765625,
0.4072265625
],
[
0.080078125,
0.406005859375,
0.4697265625,
0.406005859375,
0.4697265625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.080078125,
0.421630859375,
0.4697265625,
0.421630859375,
0.4697265625,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.50244140625,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.07843017578125,
0.43505859375,
0.4697265625,
0.436279296875,
0.4697265625,
0.45068359375,
0.07843017578125,
0.449462890625
],
[
0.50244140625,
0.43310546875,
0.890625,
0.43310546875,
0.890625,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.080078125,
0.451904296875,
0.4697265625,
0.451904296875,
0.4697265625,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.46728515625,
0.468017578125,
0.46728515625,
0.468017578125,
0.48095703125,
0.080078125,
0.48095703125
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.07843017578125,
0.481689453125,
0.416748046875,
0.48095703125,
0.416748046875,
0.49560546875,
0.07843017578125,
0.496826171875
],
[
0.50341796875,
0.479736328125,
0.890625,
0.479736328125,
0.890625,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.09722900390625,
0.49560546875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51123046875,
0.09722900390625,
0.51025390625
],
[
0.5009765625,
0.49365234375,
0.8896484375,
0.49365234375,
0.8896484375,
0.5068359375,
0.5009765625,
0.5068359375
],
[
0.080078125,
0.51220703125,
0.468017578125,
0.51220703125,
0.468017578125,
0.52587890625,
0.080078125,
0.52587890625
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.07843017578125,
0.5263671875,
0.468017578125,
0.52587890625,
0.468017578125,
0.54052734375,
0.07843017578125,
0.54150390625
],
[
0.5009765625,
0.5234375,
0.705078125,
0.52294921875,
0.705078125,
0.537109375,
0.5009765625,
0.53857421875
],
[
0.080078125,
0.54248046875,
0.46728515625,
0.54248046875,
0.46728515625,
0.556640625,
0.080078125,
0.556640625
],
[
0.521484375,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.521484375,
0.5517578125
],
[
0.07843017578125,
0.55615234375,
0.46728515625,
0.556640625,
0.46728515625,
0.5712890625,
0.07843017578125,
0.57080078125
],
[
0.50341796875,
0.55322265625,
0.89208984375,
0.55322265625,
0.89208984375,
0.56640625,
0.50341796875,
0.56640625
],
[
0.07843017578125,
0.5712890625,
0.406005859375,
0.57373046875,
0.406005859375,
0.58837890625,
0.07843017578125,
0.5859375
],
[
0.50341796875,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.50244140625,
0.583984375,
0.8896484375,
0.583984375,
0.8896484375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.1405029296875,
0.6015625,
0.41015625,
0.599609375,
0.41015625,
0.6171875,
0.1405029296875,
0.619140625
],
[
0.50244140625,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.4453125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.6162109375,
0.4453125,
0.6162109375
],
[
0.50244140625,
0.6142578125,
0.81640625,
0.6142578125,
0.81640625,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.8056640625,
0.61865234375,
0.8544921875,
0.61865234375,
0.8544921875,
0.62890625,
0.8056640625,
0.62890625
],
[
0.098876953125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.64013671875,
0.098876953125,
0.64013671875
],
[
0.07843017578125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.07843017578125,
0.65478515625
],
[
0.65283203125,
0.64794921875,
0.6650390625,
0.64794921875,
0.6650390625,
0.65576171875,
0.65283203125,
0.65576171875
],
[
0.66845703125,
0.64306640625,
0.6845703125,
0.63818359375,
0.689453125,
0.64794921875,
0.67333984375,
0.65283203125
],
[
0.583984375,
0.65283203125,
0.609375,
0.65283203125,
0.609375,
0.66015625,
0.583984375,
0.66015625
],
[
0.67724609375,
0.6513671875,
0.81640625,
0.6484375,
0.81640625,
0.66650390625,
0.67724609375,
0.66943359375
],
[
0.869140625,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.869140625,
0.66650390625
],
[
0.58251953125,
0.6591796875,
0.6357421875,
0.65576171875,
0.63623046875,
0.66748046875,
0.583984375,
0.67041015625
],
[
0.646484375,
0.6611328125,
0.685546875,
0.65771484375,
0.68701171875,
0.66943359375,
0.64794921875,
0.67236328125
],
[
0.080078125,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.080078125,
0.6572265625,
0.46728515625,
0.6572265625,
0.46728515625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.67333984375,
0.673828125,
0.68310546875,
0.673828125,
0.68310546875,
0.67822265625,
0.67333984375,
0.67822265625
],
[
0.0767822265625,
0.68603515625,
0.4697265625,
0.6845703125,
0.4697265625,
0.7021484375,
0.0767822265625,
0.7041015625
],
[
0.52294921875,
0.6845703125,
0.890625,
0.6845703125,
0.890625,
0.69775390625,
0.52294921875,
0.69775390625
],
[
0.080078125,
0.7021484375,
0.4697265625,
0.7021484375,
0.4697265625,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50244140625,
0.69970703125,
0.890625,
0.69970703125,
0.890625,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.08087158203125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.50341796875,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.7265625,
0.50341796875,
0.7265625
],
[
0.5,
0.72900390625,
0.74658203125,
0.7265625,
0.74658203125,
0.744140625,
0.5,
0.7470703125
],
[
0.7255859375,
0.73095703125,
0.8935546875,
0.7265625,
0.8935546875,
0.74365234375,
0.7255859375,
0.74755859375
],
[
0.50341796875,
0.74560546875,
0.61767578125,
0.74560546875,
0.61767578125,
0.7587890625,
0.50341796875,
0.7587890625
],
[
0.080078125,
0.73291015625,
0.43701171875,
0.73291015625,
0.43701171875,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.1519775390625,
0.76318359375,
0.396240234375,
0.76318359375,
0.396240234375,
0.77783203125,
0.1519775390625,
0.77783203125
],
[
0.4453125,
0.7646484375,
0.4697265625,
0.7646484375,
0.4697265625,
0.77783203125,
0.4453125,
0.77783203125
],
[
0.63818359375,
0.7744140625,
0.7548828125,
0.7744140625,
0.7548828125,
0.7890625,
0.63818359375,
0.7890625
],
[
0.87109375,
0.77587890625,
0.8935546875,
0.77587890625,
0.8935546875,
0.7890625,
0.87109375,
0.7890625
],
[
0.521484375,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.521484375,
0.8095703125
],
[
0.17724609375,
0.8017578125,
0.369384765625,
0.80029296875,
0.369384765625,
0.818359375,
0.17724609375,
0.8193359375
],
[
0.4453125,
0.80322265625,
0.4697265625,
0.80322265625,
0.4697265625,
0.81640625,
0.4453125,
0.81640625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.09722900390625,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.09722900390625,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.8935546875,
0.8271484375,
0.8935546875,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5,
0.8388671875,
0.708984375,
0.8388671875,
0.708984375,
0.85986328125,
0.5,
0.85986328125
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.54736328125,
0.8818359375,
0.8505859375,
0.8818359375,
0.8505859375,
0.896484375,
0.54736328125,
0.896484375
],
[
0.86767578125,
0.8818359375,
0.890625,
0.8818359375,
0.890625,
0.89453125,
0.86767578125,
0.89453125
],
[
0.080078125,
0.88720703125,
0.4453125,
0.88720703125,
0.4453125,
0.90087890625,
0.080078125,
0.90087890625
]
] |
[
"[Punproi(cam, i,2.)F",
"corresponding to the per-",
"son feature volumes for each affine augmented multi-view",
"input image al and a2, respectively. [F!],=1 and",
"(F?)P",
", are passed through pose_net3d and soft-argmax",
"[8] to estimate the 3d poses i, y2 e RPxJx3. These 3d",
"poses serve as a bottleneck representation.",
"Given the camera parameters cam and the affine trans-",
"formation parameters tl.s,t2.s,we project the bottleneck 3d",
"poses in cross-affine-view i.e. yl are projected to r2 im-",
"age space using t2s to generate multi-view 2d poses y2 e",
"RC×PxJx2, and Jy2 are projected to the al image space us-",
"ing t,to generate multi-view 2d poses yl eRCxPxJx2",
"Figure 3. Comparing ground-truth 2d poses generated by project-",
"We propose to render yl and y2 2d poses in the 2d",
"ing the ground-truth 3d poses to each multi-view imageand our",
"heatmap representation.The heatmap representation en-",
"pseudo 2d poses generated by running HRNet human pose esti-",
"codes the per-pixel likelihood of a body joint and has been",
"mation model [55] on the training dataset. Pseudo 2d poses con-",
"avital component to enablethe state-of-the-art 2d pose es-",
"tain localization errors due to occlusion (seethe red arrows),and",
"timation approaches [55]. Generating heatmap representa-",
"ground-truth 2d poses exist for partially or even entirely occluded",
"tion has essentially been a pre-processing step where state-",
"persons (see the blue dotted arrows).",
"of-the-art approaches quantize the 2d joints before gener-",
"ating the heatmap [55].However, this quantization step",
"3.5.Adaptive supervision attention",
"is non-differentiable and could cut the backward gradient",
"flow. Zhang et al. [61] show that encoding floating point 2d",
"Traditional L and L2 losses treat each label equally, which",
"joints into heatmap representation in their pre-processing",
"is sub-optimal in two aspects: (1) the 2d human pose de-",
"step leads to improved performance. We propose to use the",
"tector may generate inaccurate labels due to occlusions (see",
"same differentiable approach in an online way to render the",
"the red arrows in Figure 3): (2) the 3d-2d projection will",
"projected 2d joints into the heatmap representation.",
"output 2d joints in certain views even when the person is",
"We render y and y? into heatmap representation to gen-",
"entirely occluded (see the blue dotted arrows in Figure 3).",
"erate 2d multi-view heatmaps Hl, and H?, respectively.We",
"Therefore, we propose to employ attentions to adaptively",
"apply the affine transformations ti.s,t2.on the pseudo 2d",
"guide the supervision process.",
"poses y2d to generate pseudo 2d multi-view joints y2d, y2*",
"For L2 loss supervision, we use the soft attention.",
"and heatmaps Hl*,H2*.Then, we compute the L2 loss be-",
"Specifically, we use ResNet-18 to extract the visual fea-",
"tween heatmaps as the pose heatmap loss lpose.H:",
"tures of the views, followed by deconvolutional layers to ob-",
"tain the attention heatmaps A (see attn_net2d in Figure 2)",
"Ipose.H = L2(Hl, H1*) + L2(H?, H2*)",
"Then we compute the element-wise product of A and the",
"(4)",
"square error before averaging, as the new loss la",
"IposeH:",
"After training with lpose H preliminarily, we propose to",
"add the L, loss between 2d joints to further fine-tune the",
"1",
"N",
"attn",
"A @(Hi-H)?",
"(7)",
"pose.H",
"NZ",
"rithm [37] to obtain the optimal assignment between y and",
"model. For each view, we employ the Hungarian algo-",
"-",
"y2d, where the matching cost is the mean absolute error.",
"To avoid that A becomes zero, we add a regulariza-",
"Based on the optimal assignment, we obtain the Li loss as",
"tion term. We create tensors of all ones 1 as the atten-",
"Ipose.J. Then we use lpose.H and lpose.J together to train the",
"tionheatmaplabels.and compute L2 error as the attention",
"loss lattn. If lattn becomes zero, lat",
"latte.n degrades to non-",
"attentive version.",
"whole network, where ^ is a manually defined weight:",
"lpose-J = Li(yl,y2a) + Ci(y?, y2a)",
"(5)",
"lattn = L2(A, 1)",
"(8) ",
"For L loss supervision, we use the hard attention. For",
"Ipose_3d = lpose.H + pose.J",
"(6)",
"each input with K views, we compute the L, loss of each",
"As the network needs to reason about the 2d joint lo-",
"view, find the view with the largest loss, and ignore it when",
"cations in spatial dimension, it implicitly solves the person",
"averaging the final loss Ipte y:",
"correspondence problem. Training pose_net3d with 3d pose",
"loss lpose_3d performs decently, but to achieve even better",
"j = argmax Li(yi,yi,2d) (i = 1,2,..,K)",
"(9)",
"results, we introduce the adaptive supervision attention."
] |
{
"arxiv_id": "2404.02041",
"language": "en",
"page_idx": 4
}
|
||
2307.04445-en-1
|
2307.04445-en-1.jpg
|
MiLeTS'23, August 2023, Long Beach, CA, USA
modest improvement in predicting individual attributes from a
few days of recordings.
## 2 RELATED WORKS
Many conventional approaches in characterizing daily routines require the acquisition of labeled contexts, like trajectory in home settings [15], life event sequences (eating, sleeping, etc.) [11], and GPS locations [30]. The Kasteren data set [28] was collected in a 3-bedroom apartment setting for a period of 28 days using 14 state change sensors. The investigators then achieved a timeslice accuracy of 95.6% on this data set using a hidden Markov model and conditional random fields. Some following studies have successfully utilized a probabilistic neural network learning model to separate the normal routine from unusual and suspected routines [16]. Meanwhile, other researchers have studied human behavior by tracking the spatial properties of participants via GPS [30]. However, one primary concern about these studies is that the data collection protocol could invade privacy by tracking sensitive and identifiable knowledge about an individual, such as continuous GPS. These approaches might also be costly and not scalable due to the substantial amount of effort required from researchers in either setting up the recording system or annotating the data.
To prevent the acquisition of personally identifiable information, there have been a number of studies in establishing machine learning models to infer a pre-selected set of activities (walk, stand, etc.) from unlabeled wearable time-series [42], like motion and posture. However, these models are typically trained from data gathered in laboratory settings and may not yield good performances on in situ data sets. As an alternative, motif-based methods have obtained empirical success in detecting repeated patterns; but this method can be computationally prohibitive due to the optimal granularity for motif patterns that must be searched through the whole time-series. Unlike the motif-based data mining method, one other recent study proposed to learn routine behaviors via a sparse and low-rank matrix decomposition technique [?]. In this work, the real-world physical activity data collected from Fitbit were used to cluster the behaviors of participants without expert knowledge or micro-pattern extraction. The main disadvantage associated with this approach is however the limited interpretability of decomposed matrices returned from sensor matrix operations. To our knowledge, our proposed causality-based pattern extraction from unlabeled wearable sensor recordings has not been previously considered.
# 3 DATASET INTRODUCTION
In this study, we used a dataset called TILES-2018 [32] that is publicly available for our experiment. This data set involves a set of comprehensive experiments with the target of studying how physiological and behavioral variables affect employee wellness, personality, and workplace stress. Throughout a ten-week period, physiological, environmental, and human interaction data were gathered from hospital employees who primarily provide patient care (nurses, technicians, etc.) at a large critical-care hospital. The complete dataset consisted of 213 hospital workers comprised of 120 female participants. In total, there were 113 (54.3%) registered nurses enrolled in the study, with the rest reporting some other job
Tiantian Feng et al.
title, such as occupational or lab technicians. A total of 54 participants reported working the night shift and the rest worked during the day. More details regarding the dataset can be referred to [32]
## 3.1 Study Procedure
The participants involved in this study need to complete a Initial Assessment survey during onboarding, consisting of a web-based series of surveys pulled from existing test battery questionnaires that assessed standard demographic information, personality, and affect variables. In this work, we primarily investigate how automatically discovered routine patterns correlate with job type, gender, shift type, personality, and affect. Each of the five personality factors (extraversion, agreeableness, conscientiousness, openness neuroticism) is measured via the Big Five Inventory-2 survey [39]. Five-factor scores were computed by taking the average of all responses, where each factor score is in the range of 1-5. The Positive and Negative Affect Schedule (PANAS) was administered [43] to measure affect. The PANAS consists of 10 positive affect items and 10 negative affect items. Positive and negative affect scores were calculated by summing individual scores from each group (positive and negative) with higher scores representing higher levels of corresponding affect.
## 3.2 Wearable Data
In this study, researchers instructed participants to wear a Fitbit Charge 2 [25], an OMsignal garment-based sensor [33], and a customized audio badge [18] which collectively tracks heart rate, physical activity, speech characteristics, and many other human-centric signals. Participants were asked to wear the OMsignal garment and audio badge only during their work shifts due to the battery limitation of these devices. However, participants were instructed to wear a Fitbit sensor as often as possible throughout the 10-week data collection period. In the present study, we focus on the Fitbit time series data since it is present for most participants both during and outside of their working hours. This data stream offers information about energy expenditure, sleep quality, step count and heart rate measured through photoplethysmography (PPG).
## 4 МЕТНОД
In this section, we introduce our HOT-ROD analysis pipeline in discovering routine features from Fitbit time-series recordings. As shown in Fig. 1, our proposed HOT-ROD data pipeline consists of three major modules: 1. Pre-processing; 2. Time-series; 3. Hawkes point processing. To calculate routine features using the Hawkes Point Process, we frst group time-series by day. In this work, a day is defined as the variable period of time between sleep onsets as determined from the Fitbit daily summary. Sleep durations shorter than six hours (presumed to be naps) are ignored when determining these day boundaries in the time series. Some participants may sleep less than six hours regularly or may not wear their Fitbit devices to bed, which would result in measured days lasting upwards of 30 hours. To remove these outliers, we maintain the data with its length of 20-28 hours according to the definition of a circadian cycle [21].
The pre-processing component includes data aggregation, data imputation, and data filtering. We first aggregate multivariate timeseries data at the fixed rate of one minute. The second step uses
|
[
[
0.08660888671875,
0.0770263671875,
0.3095703125,
0.0770263671875,
0.3095703125,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.8203125,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.8203125,
0.08709716796875
],
[
0.1004638671875,
0.10919189453125,
0.48046875,
0.10919189453125,
0.48046875,
0.11932373046875,
0.1004638671875,
0.11932373046875
],
[
0.51953125,
0.10919189453125,
0.91259765625,
0.10919189453125,
0.91259765625,
0.11932373046875,
0.51953125,
0.11932373046875
],
[
0.098876953125,
0.121826171875,
0.24267578125,
0.121826171875,
0.24267578125,
0.1351318359375,
0.098876953125,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.51708984375,
0.1351318359375,
0.9111328125,
0.1351318359375,
0.9111328125,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.10296630859375,
0.1572265625,
0.265625,
0.1572265625,
0.265625,
0.17041015625,
0.10296630859375,
0.17041015625
],
[
0.5185546875,
0.154052734375,
0.69384765625,
0.154052734375,
0.69384765625,
0.1673583984375,
0.5185546875,
0.1673583984375
],
[
0.5185546875,
0.1717529296875,
0.9111328125,
0.1717529296875,
0.9111328125,
0.1849365234375,
0.5185546875,
0.1849365234375
],
[
0.08660888671875,
0.1590576171875,
0.111083984375,
0.1590576171875,
0.111083984375,
0.17041015625,
0.08660888671875,
0.17041015625
],
[
0.0849609375,
0.1767578125,
0.48193359375,
0.1767578125,
0.48193359375,
0.1900634765625,
0.0849609375,
0.1900634765625
],
[
0.0849609375,
0.1900634765625,
0.48046875,
0.1900634765625,
0.48046875,
0.203857421875,
0.0849609375,
0.203857421875
],
[
0.51708984375,
0.186279296875,
0.9140625,
0.186279296875,
0.9140625,
0.199462890625,
0.51708984375,
0.199462890625
],
[
0.0849609375,
0.2052001953125,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2183837890625,
0.0849609375,
0.2183837890625
],
[
0.51708984375,
0.2008056640625,
0.91259765625,
0.2008056640625,
0.91259765625,
0.2139892578125,
0.51708984375,
0.2139892578125
],
[
0.515625,
0.212158203125,
0.9140625,
0.21337890625,
0.9140625,
0.2279052734375,
0.515625,
0.2266845703125
],
[
0.0849609375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.0849609375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.51708984375,
0.2279052734375,
0.91259765625,
0.2279052734375,
0.91259765625,
0.2412109375,
0.51708984375,
0.2412109375
],
[
0.08660888671875,
0.2474365234375,
0.48193359375,
0.2474365234375,
0.48193359375,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.5185546875,
0.24365234375,
0.91259765625,
0.24365234375,
0.91259765625,
0.25390625,
0.5185546875,
0.25390625
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.515625,
0.255126953125,
0.91259765625,
0.25390625,
0.91259765625,
0.268310546875,
0.515625,
0.26953125
],
[
0.5185546875,
0.26953125,
0.9140625,
0.26953125,
0.9140625,
0.28271484375,
0.5185546875,
0.28271484375
],
[
0.08660888671875,
0.27587890625,
0.4794921875,
0.27587890625,
0.4794921875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.08660888671875,
0.2890625,
0.48046875,
0.2890625,
0.48046875,
0.2998046875,
0.08660888671875,
0.2998046875
],
[
0.5185546875,
0.285400390625,
0.91259765625,
0.285400390625,
0.91259765625,
0.29541015625,
0.5185546875,
0.29541015625
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.5185546875,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.30810546875,
0.5185546875,
0.30810546875
],
[
0.51953125,
0.311279296875,
0.91259765625,
0.311279296875,
0.91259765625,
0.3212890625,
0.51953125,
0.3212890625
],
[
0.08660888671875,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.08660888671875,
0.32763671875
],
[
0.51708984375,
0.324462890625,
0.91259765625,
0.324462890625,
0.91259765625,
0.33837890625,
0.51708984375,
0.33837890625
],
[
0.08660888671875,
0.330810546875,
0.4794921875,
0.330810546875,
0.4794921875,
0.341552734375,
0.08660888671875,
0.341552734375
],
[
0.0849609375,
0.343505859375,
0.483642578125,
0.343505859375,
0.483642578125,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.5185546875,
0.339111328125,
0.9111328125,
0.339111328125,
0.9111328125,
0.349853515625,
0.5185546875,
0.349853515625
],
[
0.0849609375,
0.356689453125,
0.48193359375,
0.356689453125,
0.48193359375,
0.37060546875,
0.0849609375,
0.37060546875
],
[
0.5185546875,
0.3515625,
0.91259765625,
0.3515625,
0.91259765625,
0.365478515625,
0.5185546875,
0.365478515625
],
[
0.51708984375,
0.3662109375,
0.9140625,
0.3662109375,
0.9140625,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.08660888671875,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
0.0849609375,
0.385009765625,
0.48193359375,
0.385009765625,
0.48193359375,
0.3984375,
0.0849609375,
0.3984375
],
[
0.51708984375,
0.380126953125,
0.9140625,
0.380126953125,
0.9140625,
0.393310546875,
0.51708984375,
0.393310546875
],
[
0.51708984375,
0.39453125,
0.646484375,
0.39453125,
0.646484375,
0.40771484375,
0.51708984375,
0.40771484375
],
[
0.08660888671875,
0.40087890625,
0.48046875,
0.40087890625,
0.48046875,
0.410888671875,
0.08660888671875,
0.410888671875
],
[
0.0849609375,
0.412353515625,
0.373291015625,
0.412353515625,
0.373291015625,
0.425537109375,
0.0849609375,
0.425537109375
],
[
0.51708984375,
0.419189453125,
0.67724609375,
0.419189453125,
0.67724609375,
0.43310546875,
0.51708984375,
0.43310546875
],
[
0.10296630859375,
0.425537109375,
0.48193359375,
0.425537109375,
0.48193359375,
0.439453125,
0.10296630859375,
0.439453125
],
[
0.08660888671875,
0.44140625,
0.48046875,
0.44140625,
0.48046875,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.5185546875,
0.4375,
0.91259765625,
0.4375,
0.91259765625,
0.45068359375,
0.5185546875,
0.45068359375
],
[
0.0849609375,
0.453857421875,
0.48193359375,
0.453857421875,
0.48193359375,
0.46728515625,
0.0849609375,
0.46728515625
],
[
0.515625,
0.45068359375,
0.9140625,
0.449462890625,
0.9140625,
0.464111328125,
0.515625,
0.46533203125
],
[
0.08660888671875,
0.46728515625,
0.48193359375,
0.46728515625,
0.48193359375,
0.48095703125,
0.08660888671875,
0.48095703125
],
[
0.515625,
0.464111328125,
0.9140625,
0.46533203125,
0.9140625,
0.479736328125,
0.515625,
0.479248046875
],
[
0.0849609375,
0.481689453125,
0.48193359375,
0.481689453125,
0.48193359375,
0.49560546875,
0.0849609375,
0.49560546875
],
[
0.5185546875,
0.48095703125,
0.91259765625,
0.48095703125,
0.91259765625,
0.4912109375,
0.5185546875,
0.4912109375
],
[
0.0849609375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.5185546875,
0.494384765625,
0.91259765625,
0.494384765625,
0.91259765625,
0.5048828125,
0.5185546875,
0.5048828125
],
[
0.08660888671875,
0.51123046875,
0.48046875,
0.51123046875,
0.48046875,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.51708984375,
0.5068359375,
0.9140625,
0.5068359375,
0.9140625,
0.52099609375,
0.51708984375,
0.52099609375
],
[
0.08660888671875,
0.5244140625,
0.48046875,
0.5244140625,
0.48046875,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.5185546875,
0.52294921875,
0.91259765625,
0.52294921875,
0.91259765625,
0.53271484375,
0.5185546875,
0.53271484375
],
[
0.0849609375,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.51708984375,
0.5341796875,
0.91259765625,
0.5341796875,
0.91259765625,
0.548828125,
0.51708984375,
0.548828125
],
[
0.0849609375,
0.55029296875,
0.48193359375,
0.55029296875,
0.48193359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.51708984375,
0.548828125,
0.91259765625,
0.548828125,
0.91259765625,
0.5625,
0.51708984375,
0.5625
],
[
0.08660888671875,
0.56640625,
0.48046875,
0.56640625,
0.48046875,
0.5771484375,
0.08660888671875,
0.5771484375
],
[
0.5185546875,
0.564453125,
0.9111328125,
0.564453125,
0.9111328125,
0.57470703125,
0.5185546875,
0.57470703125
],
[
0.08416748046875,
0.5791015625,
0.48193359375,
0.57763671875,
0.48193359375,
0.59130859375,
0.08416748046875,
0.59228515625
],
[
0.5185546875,
0.57763671875,
0.91259765625,
0.57763671875,
0.91259765625,
0.58837890625,
0.5185546875,
0.58837890625
],
[
0.08660888671875,
0.59326171875,
0.48046875,
0.59326171875,
0.48046875,
0.60400390625,
0.08660888671875,
0.60400390625
],
[
0.51708984375,
0.59033203125,
0.9140625,
0.59033203125,
0.9140625,
0.60400390625,
0.51708984375,
0.60400390625
],
[
0.08660888671875,
0.60595703125,
0.483642578125,
0.60595703125,
0.483642578125,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.51708984375,
0.60498046875,
0.81640625,
0.60498046875,
0.81640625,
0.61865234375,
0.51708984375,
0.61865234375
],
[
0.08660888671875,
0.62060546875,
0.47802734375,
0.62060546875,
0.47802734375,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.51953125,
0.6318359375,
0.5361328125,
0.6318359375,
0.5361328125,
0.6396484375,
0.51953125,
0.6396484375
],
[
0.53759765625,
0.63330078125,
0.54345703125,
0.63330078125,
0.54345703125,
0.63818359375,
0.53759765625,
0.63818359375
],
[
0.5400390625,
0.6298828125,
0.625,
0.6298828125,
0.625,
0.6416015625,
0.5400390625,
0.6416015625
],
[
0.51708984375,
0.646484375,
0.9140625,
0.646484375,
0.9140625,
0.6611328125,
0.51708984375,
0.6611328125
],
[
0.08660888671875,
0.63525390625,
0.4794921875,
0.63525390625,
0.4794921875,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.0849609375,
0.64794921875,
0.483642578125,
0.64794921875,
0.483642578125,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.08660888671875,
0.66357421875,
0.48193359375,
0.66357421875,
0.48193359375,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.5185546875,
0.6611328125,
0.9140625,
0.6611328125,
0.9140625,
0.6748046875,
0.5185546875,
0.6748046875
],
[
0.08660888671875,
0.6767578125,
0.48046875,
0.6767578125,
0.48046875,
0.6875,
0.08660888671875,
0.6875
],
[
0.51708984375,
0.67529296875,
0.9140625,
0.67529296875,
0.9140625,
0.689453125,
0.51708984375,
0.689453125
],
[
0.0849609375,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.5185546875,
0.689453125,
0.9140625,
0.689453125,
0.9140625,
0.70263671875,
0.5185546875,
0.70263671875
],
[
0.51953125,
0.705078125,
0.91259765625,
0.705078125,
0.91259765625,
0.71533203125,
0.51953125,
0.71533203125
],
[
0.5185546875,
0.71826171875,
0.91259765625,
0.71826171875,
0.91259765625,
0.72900390625,
0.5185546875,
0.72900390625
],
[
0.08660888671875,
0.72607421875,
0.339111328125,
0.72607421875,
0.339111328125,
0.7392578125,
0.08660888671875,
0.7392578125
],
[
0.51708984375,
0.73095703125,
0.9140625,
0.73095703125,
0.9140625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.0849609375,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.0849609375,
0.75830078125
],
[
0.51708984375,
0.744140625,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.08416748046875,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.771484375,
0.08416748046875,
0.771484375
],
[
0.5185546875,
0.76025390625,
0.9111328125,
0.76025390625,
0.9111328125,
0.77099609375,
0.5185546875,
0.77099609375
],
[
0.0849609375,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.78515625,
0.0849609375,
0.78515625
],
[
0.5185546875,
0.7734375,
0.9111328125,
0.7734375,
0.9111328125,
0.7841796875,
0.5185546875,
0.7841796875
],
[
0.0882568359375,
0.787109375,
0.48193359375,
0.787109375,
0.48193359375,
0.79736328125,
0.0882568359375,
0.79736328125
],
[
0.5185546875,
0.787109375,
0.9111328125,
0.787109375,
0.9111328125,
0.79736328125,
0.5185546875,
0.79736328125
],
[
0.08660888671875,
0.8017578125,
0.48046875,
0.8017578125,
0.48046875,
0.8125,
0.08660888671875,
0.8125
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.08660888671875,
0.81494140625,
0.48046875,
0.81494140625,
0.48046875,
0.82568359375,
0.08660888671875,
0.82568359375
],
[
0.51708984375,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.08660888671875,
0.8291015625,
0.47802734375,
0.8291015625,
0.47802734375,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.51708984375,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.72802734375,
0.84130859375,
0.72802734375,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.08416748046875,
0.85400390625,
0.483642578125,
0.85498046875,
0.483642578125,
0.86865234375,
0.08416748046875,
0.8671875
],
[
0.53125,
0.85498046875,
0.91259765625,
0.85400390625,
0.91259765625,
0.86865234375,
0.53125,
0.869140625
],
[
0.0882568359375,
0.87060546875,
0.48046875,
0.87060546875,
0.48046875,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.51708984375,
0.869140625,
0.9140625,
0.869140625,
0.9140625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.08660888671875,
0.88525390625,
0.48046875,
0.88525390625,
0.48046875,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.51708984375,
0.88525390625,
0.9140625,
0.88525390625,
0.9140625,
0.89599609375,
0.51708984375,
0.89599609375
]
] |
[
"MiLeTS'23, August 2023, Long Beach, CA, USA",
"Tiantian Feng et al.",
"modestimprovementinpredicting individual attributesfroma",
"title,such asoccupational orlab technicians.Atotal of 54partici",
"fewdays ofrecordings.",
"pants reported working the night shift and the rest worked during",
"the day. More details regarding the dataset can be referred to [32]",
"RELATEDWORKS",
"3.1 StudyProcedure",
"The participants involved in this study need to complete a Initial",
"2",
"Many conventional approaches in characterizing daily routines",
"require the acquisition of labeled contexts, like trajectory in home",
"Assessment survey during onboarding, consisting of a web-based",
"settings [15], life event sequences (eating, sleeping, etc.) [11], and",
"series of surveys pulled from existing test battery questionnaires",
"that assessed standard demographic information, personality, and",
"GPS locations [30]. The Kasteren data set [28] was collected in a",
"3-bedroom apartment setting for a period of 28 days using 14 state",
"affect variables. In this work, we primarily investigate how auto",
"change sensors.The investigators then achieved a timeslice accu-",
"matically discovered routinepatterns correlatewithjobtype,gen-",
"racy of 95.6% on this data set using a hidden Markov model and",
"der, shift type, personality, and affect. Each of the five personality",
"factors (extraversion, agreeableness, conscientiousness, openness,",
"conditionalrandomfields.Somefollowingstudieshavesuccessfully",
"utilized aprobabilisticneuralnetwork learning modeltoseparate",
"neuroticism)ismeasuredviatheBigFiveInventory-2survey[39]",
"the normal routine from unusual and suspected routines [16]. Mean-",
"Five-factor scores were computed by taking the average of all re-",
"sponses.where each factor scoreis intherangeof 1-5.ThePositive",
"while,otherresearchers have studied human behaviorby tracking",
"and Negative Affect Schedule (PANAS) was administered [43] to",
"thespatialproperties of participantsviaGPS[3o].However,one",
"primary concern about these studies is that the data collection pro-",
"measureaffect.ThePANAS consists of1opositiveaffect itemsand",
"tocol could invade privacy by tracking sensitive and identifiable",
"10 negative affect items. Positive and negative affect scores were",
"calculated by summing individual scores from each group (posi-",
"knowledge about an individual, such as continuous GPS.These",
"approaches might also be costly and not scalable due to the sub-",
"tive and negative) with higher scores representing higher levels of",
"corresponding affect.",
"stantial amount of effort required from researchers in either setting",
"up the recording system or annotating the data.",
"3.2WearableData",
"To prevent the acquisition of personally identifiable information,",
"there havebeen a number of studies in establishingmachine learn",
"In this study, researchers instructed participants to wear a Fitbit",
"ing models to infer a pre-selected set of activities (walk, stand, etc.)",
"Charge 2 [25], an OMsignal garment-based sensor [33], and a cus-",
"from unlabeled wearable time-series [42], like motion and posture.",
"tomized audio badge [18] which collectively tracks heart rate, phys-",
"However, these models are typically trained from data gathered in",
"ical activity,speechcharacteristics,andmanyotherhuman-centric",
"laboratory settings and may not yield good performances on in situ",
"signals.Participants were asked to wear the OMsignal garment and",
"data sets.As an alternative,motif-based methods have obtained",
"audio badge only during their work shifts due to the battery limita-",
"empirical success in detecting repeated patterns;but this method",
"tion of these devices.However,participantswere instructed towear",
"can be computationally prohibitive due to the optimal granular-",
"a Fitbit sensor as often as possible throughout the 10-week data",
"ity for motif patterns that must be searched through the whole",
"collection period. In the present study, we focus on the Fitbit time",
"time-series.Unlike the motif-based data mining method, one other",
"series data sinceit ispresentfor mostparticipants both during and",
"recent study proposed to learn routine behaviors via a sparse and",
"outside of their working hours.This data stream offers information",
"low-rank matrixdecompositiontechnique[?].In this work,the",
"about energy expenditure, sleep quality, step count and heart rate",
"real-world physical activity data collected from Fitbit were used to",
"measured through photoplethysmography (PPG).",
"clusterthebehaviorsofparticipants without expertknowledge o",
"4",
"",
"METHOD",
"In this section, we introduce our HOT-ROD analysis pipeline in",
"micro-pattern extraction.The main disadvantage associated with",
"this approach is however the limited interpretability of decomposed",
"matricesreturnedfrom sensor matrix operations.Toourknowledge,",
"discovering routine features from Fitbit time-series recordings. As",
"ourproposed causality-based patternextractionfrom unlabeled",
"shown in Fig. 1, our proposed HOT-ROD data pipeline consists",
"wearable sensor recordings has not been previously considered.",
"of three major modules: 1. Pre-processing; 2.Time-series; 3.",
"Hawkespoint processing.To calculateroutinefeaturesusing",
"theHawkesPointProcess,wefirstgroup time-series byday.In",
"3 DATASET INTRODUCTION",
"this work, a day is defined as the variable period of time between",
"In this study, we used a dataset called TILES-2018 [32] that is pub-",
"sleep onsets as determined from the Fitbit daily summary. Sleep",
"licly available for our experiment. This data set involves a set",
"durations shorterthan sixhours (presumedtobenaps)areignored",
"of comprehensive experiments with the target of studying how",
"when determining these day boundaries in the time series. Some",
"physiological andbehavioral variables affectemployeewellness,",
"participants may sleep less than six hours regularly or may not",
"personality,and workplace stress.Throughout a ten-week period.",
"wear their Fitbit devices to bed, which would result in measured",
"physiological, environmental, and human interaction data were",
"days lasting upwards of 3o hours. To remove these outliers, we",
"gatheredfromhospitalemployeeswhoprimarilyprovidepatient",
"maintain the data with its length of 20-28 hours according to the",
"care (nurses, technicians, etc.) at a large critical-care hospital. The",
"definition of a circadian cycle [21].",
"complete dataset consisted of 213hospital workers comprised of",
"The pre-processing component includes data aggregation, data",
"120femaleparticipants.Intotal,therewere113(54.3%)registered",
"imputation, and data filtering. We first aggregate multivariate time-",
"nurses enrolled in the study,with the rest reporting some other job",
"series data at the fixed rate of oneminute.The second step uses"
] |
{
"arxiv_id": "2307.04445",
"language": "en",
"page_idx": 1
}
|
||
2307.04445-en-3
|
2307.04445-en-3.jpg
|
MiLeTS '23, August 2023, Long Beach, CA, USA
method for time-series clustering since our interest is to robustly identify long-range repeated patterns while contending with the possible presence of irrelevant data points.
## 4.3 Hawkes Point Process
Hawkes processes [23], also known as a self-exciting counting process, is a stochastic point process for studying event sequence patterns where historical event occurrences are assumed to increase the chance of arrival of new events. In general, a point process is a collection of a list of discrete events in time and their associate $\operatorname*{dimension,}\left\{t_{i},u_{i}\right\}$with time $t_{i}\in[0,n]$ and event $u_{i}\in[0,U]$, where $n$ and $U$ are the maximum time in a time-series and total types of events, respectively. Here, we define $d_i$ representing the clusters that assign to $i$-th time point. In this study, we define the event as transitions between different time-series clusters. For instance, given two possible consecutive time points $\{t_i,d_i\}$ and $\{t_{i+1},d_{i+1}\}$ in a time-series of a day, we can define an event as $\{t_{i+1},(d_{i}\rightarrow d_{i+1})\}$,such that $d_i\neq d_{i+1}.$ In this work, we can also find that $n$ represents the number of total points in a time-series of days.
A multi-dimensional point process with $U$ types of events can be equivalently represented by $\mathcal{U}$ counting processes: $\mathcal{N}_u=\{\mathcal{N}_u(t)|t\in$ $[0,n]\}$, where $\mathcal{N}_u(t)$ denotes the number of type-$u$ events occurring before $n$-th time point. There are possible $\binom K2-K$ types of events by the definition of the event in this work. Let the history $\mathcal{H}(t)$ be the list of times of events up to time $n:$
(2)
$$\mathcal{H}_{n}^{\mathcal{U}}=\{(t_{i},u_{i})|t_{i}<n,u_{i}\in\mathcal{U}\}$$
Then, the expected instantaneous rate of occurring type-$u$ events
given history is :
$$\lambda_u(t)dt=\lambda_u(t|\mathcal{H}_t^u)$$
(3)
We can then derive the intensity function $\lambda_u(t)$ of the multi-dimensional tine features from Fitbit Charge 2 time-series that measure PPG
Hawkes process as:
$$\lambda_u(t)=\mu_u+\sum_{i:t_i<t}\phi_{uu'}(t_i)$$
(4)
where $\mu_u$ is the exogenous (base) intensity indecent of the history. $\phi_{uu^{\prime}}(t)$ is the impact function capturing the temporal influence of type-$u^\prime$ event on the subsequent type-$u$ event. We can further define $\phi_{uu^{\prime}}(t)$ as:
$$\phi_{uu'}(t)=\sum_{m=1}^Ma_{uu'}^mk_m(t)$$
(5)
where $k_{m}( t)$ is the m- th basis function and $a_{uu^{\prime }}^{m}$ represents the co- efficient corresponding to $k_{m}( t) .$We adopted the Gaussian filter as the basis function in this study. We describe type-$u^\prime$ event does not Granger-cause type-$u$ event, when function $\lambda_u(t)$ is independent of historical events of type-$u^\prime.$ Finally, we can build a Granger causality graph $G=(\mathcal{U},\mathcal{E})$ with $U$ types of events as the nodes and the directed edges in between representing the causation. In this study, the routine feature is naturally defined as the adjacency matrix A of the Granger causality graph learned from the Hawkes process. Element A$_uu^{\prime}$ can be viewed as the infectivity $(\int_0^t\phi_uu^{\prime}(s)ds)$ of type-$u^\prime$ cluster-transition on type-$u$ cluster-transition.
A detailed description of the above definitions can be found in [12,44]. To discover the Granger causality graph from a Hawkes
Tiantian Feng et al.
point process, we adopt the learning algorithm proposed by [44]. This algorithm learns the Granger causality graph robustly given a few training sequences.
Summary Our proposed HOT-ROD pipeline aims to learn routine characteristics from wearable time-series without prior knowledge or the acquisition of labeled event sequences. The time series data used in this study contains continuous measurements of physiological response and physical activity. The proposed learned routine features capture patterns in how people advance from one state to another in everyday life.
## 5 RESULT
In this section, we validate the effectiveness of the proposed HOTROD approach for predicting demographic, personality and affect with Fitbit time series data.
## 5.1 Experimental Setup
5.1.1 Daily Summary Routine Feature. We use Fitbit daily summary data to construct our baseline model. The Fitbit daily summary data is extracted using the API provided by Fitbit. The measurements we select from Fitbit daily summary report include sleep duration, sleep efficiency, step counts, resting heart rate, and heart rate zone duration. Here, Fitbit categorizes heart rate into 4 zones:
$\bullet$ Out of Zone (heart rate is below 50% of its maximum). $\bullet$ Fat-burn Zone (heart rate is 51% to 69% to its maximum) $\bullet$ Cardio Zone (heart rate is 70% to 84% to its maximum). $\bullet$ Peak Zone (heart rate is above 85% of its maximum).
A set of 5 statistical functionals (e.g., max, standard deviation)
are then introduced on the Fitbit summary feature.
5.1.2 HOT-ROD Routine Feature. We compute the HOT-ROD rouheart rate and step count. Routine behaviors may differ between workdays and off-days. Thus we separately learn the routine features of workdays and off-days. We also observe that the data available for each participant varies (min: 2 days, max: 70 days) which can lead to biased results without proper data selection. Hence, we choose to randomly select an equal amount of days of data from workdays and off-days of a participant for our analysis. In our study, we experimentally test by picking $n=5$ days from each type. We choose to pick 5 days in our experiment since it ensures a reasonable amount of data per participant and also the number of qualified participants retained in the analysis. In the end, there are 101 participants retained in this experiment.
According to Section 4, we first aggregate PPG heart rate and step count every minute for a day of data. We impute the missing data in each aggregated time-series using the ARIMA model, where we estimate the number of time lags, the degree of difference, and the order of the moving average according to the Akaike information criterion. We fill the continuous missing segments over 15 minutes with large enough negative values. We then filter the imputed time series using an S-G filter. To minimize the deforming of the signal, we choose a small window size $m=2$, and cubic order polynomials in the S-G filter. The window size parameter could be tuned systematically based on criteria and heuristics defined in [41],but we leave this endeavor for future work
|
[
[
0.0849609375,
0.0738525390625,
0.3095703125,
0.07452392578125,
0.3095703125,
0.08837890625,
0.0849609375,
0.08709716796875
],
[
0.81884765625,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.81884765625,
0.08709716796875
],
[
0.08660888671875,
0.10919189453125,
0.4794921875,
0.10919189453125,
0.4794921875,
0.11932373046875,
0.08660888671875,
0.11932373046875
],
[
0.515625,
0.1060791015625,
0.91259765625,
0.106689453125,
0.91259765625,
0.121826171875,
0.515625,
0.12060546875
],
[
0.08416748046875,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1363525390625,
0.51708984375,
0.1363525390625
],
[
0.0849609375,
0.1351318359375,
0.3447265625,
0.1351318359375,
0.3447265625,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1337890625,
0.66015625,
0.1351318359375,
0.66015625,
0.1483154296875,
0.51708984375,
0.147705078125
],
[
0.515625,
0.1483154296875,
0.91259765625,
0.147705078125,
0.91259765625,
0.1622314453125,
0.515625,
0.1634521484375
],
[
0.08416748046875,
0.160400390625,
0.30810546875,
0.1610107421875,
0.30810546875,
0.175537109375,
0.08416748046875,
0.1749267578125
],
[
0.51708984375,
0.1622314453125,
0.91259765625,
0.1622314453125,
0.91259765625,
0.175537109375,
0.51708984375,
0.175537109375
],
[
0.08416748046875,
0.177978515625,
0.48193359375,
0.1787109375,
0.48193359375,
0.1932373046875,
0.08416748046875,
0.1925048828125
],
[
0.51708984375,
0.175537109375,
0.91259765625,
0.175537109375,
0.91259765625,
0.189453125,
0.51708984375,
0.189453125
],
[
0.0849609375,
0.1932373046875,
0.48193359375,
0.1932373046875,
0.48193359375,
0.20703125,
0.0849609375,
0.20703125
],
[
0.51708984375,
0.1912841796875,
0.9140625,
0.1912841796875,
0.9140625,
0.2052001953125,
0.51708984375,
0.2052001953125
],
[
0.08660888671875,
0.208984375,
0.4794921875,
0.208984375,
0.4794921875,
0.2197265625,
0.08660888671875,
0.2197265625
],
[
0.51708984375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51708984375,
0.2183837890625
],
[
0.0849609375,
0.2215576171875,
0.483642578125,
0.2215576171875,
0.483642578125,
0.23486328125,
0.0849609375,
0.23486328125
],
[
0.51708984375,
0.2183837890625,
0.91259765625,
0.2183837890625,
0.91259765625,
0.231689453125,
0.51708984375,
0.231689453125
],
[
0.0849609375,
0.23486328125,
0.48046875,
0.23486328125,
0.48046875,
0.248779296875,
0.0849609375,
0.248779296875
],
[
0.5146484375,
0.23291015625,
0.6650390625,
0.23291015625,
0.6650390625,
0.246826171875,
0.5146484375,
0.246826171875
],
[
0.0849609375,
0.248779296875,
0.48046875,
0.248779296875,
0.48046875,
0.261962890625,
0.0849609375,
0.261962890625
],
[
0.51953125,
0.26318359375,
0.53369140625,
0.26318359375,
0.53369140625,
0.271484375,
0.51953125,
0.271484375
],
[
0.53369140625,
0.2587890625,
0.615234375,
0.2587890625,
0.615234375,
0.2734375,
0.53369140625,
0.2734375
],
[
0.0849609375,
0.276611328125,
0.48046875,
0.276611328125,
0.48046875,
0.290283203125,
0.0849609375,
0.290283203125
],
[
0.08660888671875,
0.264404296875,
0.48046875,
0.264404296875,
0.48046875,
0.274658203125,
0.08660888671875,
0.274658203125
],
[
0.515625,
0.27783203125,
0.91259765625,
0.276611328125,
0.91259765625,
0.291748046875,
0.515625,
0.292236328125
],
[
0.08660888671875,
0.292236328125,
0.48046875,
0.292236328125,
0.48046875,
0.302978515625,
0.08660888671875,
0.302978515625
],
[
0.51708984375,
0.292236328125,
0.9140625,
0.292236328125,
0.9140625,
0.30615234375,
0.51708984375,
0.30615234375
],
[
0.0882568359375,
0.306884765625,
0.4794921875,
0.306884765625,
0.4794921875,
0.314453125,
0.0882568359375,
0.314453125
],
[
0.51708984375,
0.30615234375,
0.68310546875,
0.30615234375,
0.68310546875,
0.3193359375,
0.51708984375,
0.3193359375
],
[
0.0849609375,
0.318115234375,
0.48193359375,
0.318115234375,
0.48193359375,
0.33203125,
0.0849609375,
0.33203125
],
[
0.08660888671875,
0.33203125,
0.48046875,
0.33203125,
0.48046875,
0.34521484375,
0.08660888671875,
0.34521484375
],
[
0.515625,
0.333251953125,
0.7197265625,
0.333984375,
0.7197265625,
0.348388671875,
0.515625,
0.347900390625
],
[
0.08660888671875,
0.34521484375,
0.48046875,
0.3447265625,
0.48046875,
0.35791015625,
0.08660888671875,
0.359130859375
],
[
0.51708984375,
0.3515625,
0.9111328125,
0.3515625,
0.9111328125,
0.365478515625,
0.51708984375,
0.365478515625
],
[
0.0849609375,
0.359130859375,
0.483642578125,
0.359130859375,
0.483642578125,
0.37255859375,
0.0849609375,
0.37255859375
],
[
0.51708984375,
0.3662109375,
0.91259765625,
0.3662109375,
0.91259765625,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.08660888671875,
0.375,
0.1168212890625,
0.375,
0.1168212890625,
0.386474609375,
0.08660888671875,
0.386474609375
],
[
0.51708984375,
0.380126953125,
0.91259765625,
0.380126953125,
0.91259765625,
0.393310546875,
0.51708984375,
0.393310546875
],
[
0.10296630859375,
0.386962890625,
0.4794921875,
0.386962890625,
0.4794921875,
0.40087890625,
0.10296630859375,
0.40087890625
],
[
0.08416748046875,
0.399658203125,
0.484375,
0.40087890625,
0.484375,
0.415283203125,
0.08416748046875,
0.4140625
],
[
0.51708984375,
0.39453125,
0.9140625,
0.39453125,
0.9140625,
0.40771484375,
0.51708984375,
0.40771484375
],
[
0.5185546875,
0.40771484375,
0.9140625,
0.40771484375,
0.9140625,
0.421630859375,
0.5185546875,
0.421630859375
],
[
0.08660888671875,
0.415283203125,
0.48193359375,
0.415283203125,
0.48193359375,
0.4287109375,
0.08660888671875,
0.4287109375
],
[
0.515625,
0.421630859375,
0.859375,
0.421630859375,
0.859375,
0.43505859375,
0.515625,
0.43505859375
],
[
0.0849609375,
0.429931640625,
0.400390625,
0.429931640625,
0.400390625,
0.443115234375,
0.0849609375,
0.443115234375
],
[
0.39794921875,
0.43115234375,
0.48193359375,
0.43115234375,
0.48193359375,
0.442626953125,
0.39794921875,
0.442626953125
],
[
0.08660888671875,
0.443115234375,
0.48193359375,
0.443115234375,
0.48193359375,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.5185546875,
0.44140625,
0.85546875,
0.44140625,
0.85546875,
0.451904296875,
0.5185546875,
0.451904296875
],
[
0.0849609375,
0.45703125,
0.373291015625,
0.45703125,
0.373291015625,
0.47021484375,
0.0849609375,
0.47021484375
],
[
0.51953125,
0.455078125,
0.869140625,
0.455078125,
0.869140625,
0.46533203125,
0.51953125,
0.46533203125
],
[
0.515625,
0.466552734375,
0.86279296875,
0.46728515625,
0.86279296875,
0.48095703125,
0.515625,
0.479736328125
],
[
0.188720703125,
0.481689453125,
0.37744140625,
0.48291015625,
0.37744140625,
0.50048828125,
0.188720703125,
0.5
],
[
0.463134765625,
0.486083984375,
0.48193359375,
0.486083984375,
0.48193359375,
0.498779296875,
0.463134765625,
0.498779296875
],
[
0.5185546875,
0.48291015625,
0.8447265625,
0.48291015625,
0.8447265625,
0.49365234375,
0.5185546875,
0.49365234375
],
[
0.08416748046875,
0.50048828125,
0.48193359375,
0.501953125,
0.48193359375,
0.5166015625,
0.08416748046875,
0.51513671875
],
[
0.53173828125,
0.5,
0.9140625,
0.5,
0.9140625,
0.51318359375,
0.53173828125,
0.51318359375
],
[
0.08660888671875,
0.517578125,
0.188720703125,
0.517578125,
0.188720703125,
0.52880859375,
0.08660888671875,
0.52880859375
],
[
0.515625,
0.51318359375,
0.82275390625,
0.51318359375,
0.82275390625,
0.5263671875,
0.515625,
0.5263671875
],
[
0.21728515625,
0.52978515625,
0.34716796875,
0.52783203125,
0.34716796875,
0.54541015625,
0.21728515625,
0.54736328125
],
[
0.461669921875,
0.53076171875,
0.483642578125,
0.53076171875,
0.483642578125,
0.54541015625,
0.461669921875,
0.54541015625
],
[
0.51953125,
0.53857421875,
0.9111328125,
0.53857421875,
0.9111328125,
0.548828125,
0.51953125,
0.548828125
],
[
0.0849609375,
0.54736328125,
0.50634765625,
0.54736328125,
0.50634765625,
0.5625,
0.0849609375,
0.5625
],
[
0.515625,
0.55029296875,
0.9140625,
0.55029296875,
0.9140625,
0.564453125,
0.515625,
0.564453125
],
[
0.08416748046875,
0.56103515625,
0.205078125,
0.5625,
0.205078125,
0.5771484375,
0.08416748046875,
0.57568359375
],
[
0.5185546875,
0.56640625,
0.91259765625,
0.56640625,
0.91259765625,
0.5771484375,
0.5185546875,
0.5771484375
],
[
0.193603515625,
0.58349609375,
0.27294921875,
0.58349609375,
0.27294921875,
0.59765625,
0.193603515625,
0.59765625
],
[
0.301513671875,
0.58349609375,
0.35546875,
0.58349609375,
0.35546875,
0.59765625,
0.301513671875,
0.59765625
],
[
0.51708984375,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.51708984375,
0.59228515625
],
[
0.463134765625,
0.58837890625,
0.483642578125,
0.58837890625,
0.483642578125,
0.6015625,
0.463134765625,
0.6015625
],
[
0.51708984375,
0.59326171875,
0.91259765625,
0.59326171875,
0.91259765625,
0.60400390625,
0.51708984375,
0.60400390625
],
[
0.27783203125,
0.599609375,
0.301513671875,
0.599609375,
0.301513671875,
0.60791015625,
0.27783203125,
0.60791015625
],
[
0.51708984375,
0.60498046875,
0.9140625,
0.60498046875,
0.9140625,
0.61865234375,
0.51708984375,
0.61865234375
],
[
0.08660888671875,
0.61279296875,
0.483642578125,
0.61279296875,
0.483642578125,
0.626953125,
0.08660888671875,
0.626953125
],
[
0.0849609375,
0.626953125,
0.48193359375,
0.62548828125,
0.48193359375,
0.64013671875,
0.0849609375,
0.6416015625
],
[
0.5185546875,
0.6220703125,
0.91259765625,
0.6220703125,
0.91259765625,
0.6318359375,
0.5185546875,
0.6318359375
],
[
0.5185546875,
0.63525390625,
0.9111328125,
0.63525390625,
0.9111328125,
0.64599609375,
0.5185546875,
0.64599609375
],
[
0.0849609375,
0.6416015625,
0.48193359375,
0.6416015625,
0.48193359375,
0.65478515625,
0.0849609375,
0.65478515625
],
[
0.51708984375,
0.64794921875,
0.9140625,
0.64794921875,
0.9140625,
0.6611328125,
0.51708984375,
0.6611328125
],
[
0.08416748046875,
0.65380859375,
0.193603515625,
0.65478515625,
0.193603515625,
0.66943359375,
0.08416748046875,
0.66845703125
],
[
0.5185546875,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.673828125,
0.5185546875,
0.673828125
],
[
0.26953125,
0.6767578125,
0.285888671875,
0.671875,
0.295654296875,
0.689453125,
0.279296875,
0.6943359375
],
[
0.515625,
0.67529296875,
0.9140625,
0.673828125,
0.9140625,
0.6875,
0.515625,
0.689453125
],
[
0.2900390625,
0.6845703125,
0.361083984375,
0.6845703125,
0.361083984375,
0.70556640625,
0.2900390625,
0.70556640625
],
[
0.463134765625,
0.6875,
0.483642578125,
0.6875,
0.483642578125,
0.70068359375,
0.463134765625,
0.70068359375
],
[
0.51708984375,
0.689453125,
0.9150390625,
0.689453125,
0.9150390625,
0.70263671875,
0.51708984375,
0.70263671875
],
[
0.275390625,
0.69580078125,
0.28759765625,
0.69580078125,
0.28759765625,
0.69970703125,
0.275390625,
0.69970703125
],
[
0.2099609375,
0.689453125,
0.268798828125,
0.689453125,
0.268798828125,
0.70068359375,
0.2099609375,
0.70068359375
],
[
0.27294921875,
0.70263671875,
0.289306640625,
0.70263671875,
0.289306640625,
0.7119140625,
0.27294921875,
0.7119140625
],
[
0.51708984375,
0.70263671875,
0.9140625,
0.70263671875,
0.9140625,
0.71630859375,
0.51708984375,
0.71630859375
],
[
0.0849609375,
0.71337890625,
0.48193359375,
0.71337890625,
0.48193359375,
0.7265625,
0.0849609375,
0.7265625
],
[
0.5185546875,
0.71826171875,
0.78173828125,
0.71826171875,
0.78173828125,
0.72900390625,
0.5185546875,
0.72900390625
],
[
0.08660888671875,
0.72900390625,
0.4794921875,
0.72900390625,
0.4794921875,
0.7392578125,
0.08660888671875,
0.7392578125
],
[
0.53369140625,
0.72998046875,
0.91259765625,
0.72998046875,
0.91259765625,
0.744140625,
0.53369140625,
0.744140625
],
[
0.0849609375,
0.7421875,
0.48193359375,
0.7421875,
0.48193359375,
0.755859375,
0.0849609375,
0.755859375
],
[
0.5185546875,
0.7470703125,
0.91259765625,
0.7470703125,
0.91259765625,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.0849609375,
0.755859375,
0.483642578125,
0.755859375,
0.483642578125,
0.76953125,
0.0849609375,
0.76953125
],
[
0.51708984375,
0.75830078125,
0.9140625,
0.75830078125,
0.9140625,
0.771484375,
0.51708984375,
0.771484375
],
[
0.08660888671875,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.7822265625,
0.08660888671875,
0.7822265625
],
[
0.51708984375,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.0849609375,
0.7841796875,
0.48193359375,
0.7841796875,
0.48193359375,
0.79736328125,
0.0849609375,
0.79736328125
],
[
0.5185546875,
0.787109375,
0.91259765625,
0.787109375,
0.91259765625,
0.79736328125,
0.5185546875,
0.79736328125
],
[
0.0849609375,
0.796875,
0.483642578125,
0.79736328125,
0.483642578125,
0.81103515625,
0.0849609375,
0.81005859375
],
[
0.51708984375,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.08660888671875,
0.81298828125,
0.48046875,
0.81298828125,
0.48046875,
0.82373046875,
0.08660888671875,
0.82373046875
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.0849609375,
0.82568359375,
0.48046875,
0.82568359375,
0.48046875,
0.8388671875,
0.0849609375,
0.8388671875
],
[
0.51953125,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.51953125,
0.8388671875
],
[
0.0849609375,
0.84033203125,
0.48193359375,
0.84033203125,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.410888671875,
0.85498046875,
0.410888671875,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.10296630859375,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.8818359375,
0.10296630859375,
0.8818359375
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.0849609375,
0.88330078125,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.0849609375,
0.896484375
],
[
0.51708984375,
0.88330078125,
0.810546875,
0.88330078125,
0.810546875,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"MiLeTs'23, August 2023, Long Beach, CA, USA",
"Tiantian Feng et al.",
"method for time-series clustering since ourinterest is to robustly",
"point process, weadopt the learning algorithm proposed by [44]",
"identify long-range repeated patterns while contending with the",
"This algorithm learns the Granger causality graph robustly given a",
"possible presence of irrelevant data points.",
"few training sequences.",
"Summary Our proposed HOT-ROD pipeline aims to learn routine",
"4.3HawkesPointProcess",
"characteristics from wearable time-series without prior knowledge",
"Hawkes processes [23], also known as a self-exciting counting",
"or the acquisition of labeled event sequences. The time series data",
"process, is a stochastic point process for studying event sequence",
"used in this study contains continuous measurements of physiolog-",
"patterns wherehistoricalevent occurrences are assumed toincrease",
"ical response and physical activity. The proposed learned routine",
"the chance of arrival of new events. In general, a point process is",
"features capture patterns in how people advance from one state to",
"a collection of a list of discrete events in time and their associate",
"another in everyday life",
"dimension, (ti,ui] with time ti e [o,n] and event ui e [o,U]",
"5",
"RESULT",
"types of events, respectively. Here, we define di representing the",
"wheren and U arethemaximum time ina time-series and total",
"In this section,we validate the effectiveness of the proposed HOT",
"clusters that assign to i-th time point. In this study, we define",
"ROD approach for predicting demographic, personality and affect",
"theeventastransitionsbetweendifferenttime-seriesclusters.For",
"with Fitbit time series data.",
"instance, given two possible consecutive time points (ti,di) and",
"(ti+1, di+1) in a time-series of a day, we can define an event as",
"5.1 ExperimentalSetup",
"(ti+1,(d;→di+1)),such that d;+di+1.In this work,wecan also",
"5.1.1 Daily Summary Routine Feature. We use Fitbit daily summary",
"find that n represents the number of total points in a time-series of",
"data to construct our baseline model. The Fitbit daily summary data",
"days.",
"is extracted using the API provided by Fitbit. The measurements",
"A multi-dimensional point process with U types of events can be",
"equivalently represented byUcounting processes:Nu =(Nu(t)lt e",
"we select from Fitbit daily summary report include sleep duration,",
"sleep efficiency, step counts, resting heart rate, and heart rate zone",
"[o, n]), where Nu(t) denotes the number of type-u events occur-",
"duration. Here, Fitbit categorizes heart rate into 4 zones:",
"ring before n-th time point. There are possible ()",
"-Ktypesof",
"events by the definition of the event in this work. Let the history",
".OutofZone(heartrateisbelow50%ofitsmaximum)",
"H(t) be the list of times of events up to time n:",
".Fat-burnZone(heartrateis51%to69%toitsmaximum)",
"CardioZone (heart rate is 70%to84%to its maximum).",
"Hu = (ti ui)lti <n, uj e )",
"(2)",
".PeakZone (heartrateisabove85%ofitsmaximum)",
"Then, the expected instantaneous rate of occurring type-u events",
"A set of 5 statistical functionals (e.g-, max, standard deviation)",
"given history is :",
"are then introduced on the Fitbit summary feature.",
"Au(t)dt = Au(tHU)",
"(3)",
"5.1.2HOT-RODRoutineFeature.WecomputetheHOT-RODrou",
"We can then derive the intensity function Au (t) ofthe multi-dimensional",
"tine features from Fitbit Charge 2 time-series that measure PPG",
"Hawkes process as:",
"heart rate and step count.Routinebehaviorsmay differbetween",
"Au(t) = μu +",
"puur(ti)",
"workdays and off-days. Thus we separately learn the routine fea-",
"(4)",
"tures of workdays and off-days.Wealso observe that the data avail-",
":ti<",
"able for each participant varies (min: 2 days, max: 70 days) which",
"where Ju is the exogenous (base) intensity indecent of the history.",
"puu (t) is the impact function capturing the temporal influence",
"can lead to biased results without proper data selection.Hence",
"we choose to randomly select an equal amount of days of data",
"of type-u' event on the subsequent type-u event. We can further",
"from workdays and off-days of a participant for our analysis. In",
"define puu'(t) as:",
"ourstudy,weexperimentallytestbypickingn=5daysfromeach",
"M",
"type. We choose to pick 5 days in our experiment since it ensures a",
"amukm(t)",
"(5)",
"reasonable amount of data per participant and also the number of",
"",
"quu(t) =",
"m=",
"qualified participants retained in the analysis. In the end, there are",
"where km(t) is the m-th basis function and am, represents the co-",
"101participantsretainedinthisexperiment",
"efficientcorrespondingtokm(t).WeadoptedtheGaussianfilteras",
"According to Section 4, we first aggregate PPG heart rate and step",
"the basis function in this study. We describe type-u' event does not",
"counteveryminuteforaday of data.Weimputethemissingdata",
"Granger-cause type-u event, when function Au (t) is independent of",
"in each aggregated time-series using the ARIMA model, where we",
"historical events of type-u'.Finally,we can build a Granger causal-",
"estimate the number of time lags, the degree of difference, and the",
"ity graph G = (u, 8) with U types of events as the nodes and the",
"orderof themovingaverageaccordingtotheAkaikeinformation",
"directed edges in between representing the causation. In this study,",
"criterion. We fll the continuous missing segments over 15 minutes",
"the routinefeature is naturally defined as the adjacency matrix A",
"with large enough negative values. We then filter the imputed",
"of the Granger causality graph learned from the Hawkes process",
"time series using an S-Gfilter.Tominimizethe deforming of the",
"Element Au can be viewed as the infectivity (J @uu (s)ds) of",
"signal, we choose a small window size m = 2, and cubic order",
"type-u' cluster-transition on type-u cluster-transition.",
"polynomials in the S-G filter. The window size parameter could",
"A detailed description of the above definitions can be found in",
"betunedsystematicallybasedoncriteriaandheuristicsdefinedin",
"[12, 44]. To discover the Granger causality graph from a Hawkes",
"[41], but we leave this endeavor for future work."
] |
{
"arxiv_id": "2307.04445",
"language": "en",
"page_idx": 3
}
|
||
2307.04445-en-4
|
2307.04445-en-4.jpg
|
Learning Behavioral Representations of Routines
From Large-scale Unlabeled Wearable Time-series Data Streams using Hawkes Point Process
MiLeTS'23, August 2023, Long Beach, CA, USA
Table 1: Table shows quantitative comparison (macro-F1 score) of different routine features to predict Initial Assessment
survey using Random Forest Model
<table>
<tbody>
<tr>
<th> </th>
<th>Clusters</th>
<th>Neuroticism</th>
<th>Conscientiousness</th>
<th>Extraversion</th>
<th>Agreeableness</th>
<th>Openness</th>
<th>Pos- Affect</th>
<th>Neg- Affect</th>
</tr>
<tr>
<td>Fitbit Summary</td>
<td> </td>
<td>$60.85\%$</td>
<td>$49.84\%$</td>
<td>$64.91\%$</td>
<td>$63.95\%$</td>
<td>$60.86\%$</td>
<td>$58.24\%$</td>
<td>$64.88\%$</td>
</tr>
<tr>
<td rowspan="3">HOT- ROD</td>
<td>3</td>
<td>$63.23\%$</td>
<td>$66.60\%$</td>
<td>$62.12\%$</td>
<td>$63.64\%$</td>
<td>$62.19\%$</td>
<td>$54.32\%$</td>
<td>$62.82\%$</td>
</tr>
<tr>
<td>4</td>
<td>$56.21\%$</td>
<td>$68.08\%$</td>
<td>$62.64\%$</td>
<td>$63.85\%$</td>
<td>$54.08\%$</td>
<td>$54.39\%$</td>
<td>$58.00\%$</td>
</tr>
<tr>
<td>5</td>
<td>$61.07\%$</td>
<td>$60.29\%$</td>
<td>$58.53\%$</td>
<td>$64.00\%$</td>
<td>$57.83\%$</td>
<td>$62.48\%$</td>
<td>$58.30\%$</td>
</tr>
<tr>
<td rowspan="3">HOT- ROD+Fitbit Summary</td>
<td>3</td>
<td>$63.53\%$</td>
<td>$52.09\%$</td>
<td>$66.25\%$</td>
<td>$70.13\%$</td>
<td>$60.86\%$</td>
<td>$58.35\%$</td>
<td>$65.72\%$</td>
</tr>
<tr>
<td>4</td>
<td>$59.84\%$</td>
<td>$63.65\%$</td>
<td>$65.87\%$</td>
<td>$64.76\%$</td>
<td>$56.09\%$</td>
<td>$59.35\%$</td>
<td>$61.91\%$</td>
</tr>
<tr>
<td>5</td>
<td>$60.56\%$</td>
<td>$58.03\%$</td>
<td>$65.02\%$</td>
<td>$64.40\%$</td>
<td>$59.58\%$</td>
<td>$63.67\%$</td>
<td>$60.31\%$</td>
</tr>
</tbody>
</table>
Table 2: Table shows quantitative comparison (macro-F1 score) of different routine features to predict demographics using Random Forest Model.
high-dimensional data input, where many features are often redundant [24]. Specifically, we perform the predictions using three sets of features: 1. Fitbit summary routine feature; 2. HOT-ROD routine feature; 3. Fitbit summary routine feature and HOT-ROD routine feature. We perform 5-fold cross-validation and report the average results in the macro-F1 score. We grid search the hyper-parameters in the RF model as follows: 1. Number of estimators: [10, 20, 30]; 2. Feature selection criterion: ["gini","entropy"]; 3. Max depth of the tree:[4,5,6];4.Minimum samples to split:[2,3,5].
<table>
<tbody>
<tr>
<th> </th>
<th>Clusters</th>
<th>WorkS Shift</th>
<th>Job Type</th>
</tr>
<tr>
<td>Fitbit Summary</td>
<td> </td>
<td>$68.92\%$</td>
<td>$60.55\%$</td>
</tr>
<tr>
<td rowspan="3">HOT- ROD</td>
<td>3</td>
<td>$58.76\%$</td>
<td>$68.79\%$</td>
</tr>
<tr>
<td>4</td>
<td>$55.98\%$</td>
<td>$67.42\%$</td>
</tr>
<tr>
<td>5</td>
<td>$62.68\%$</td>
<td>$70.13\%$</td>
</tr>
<tr>
<td rowspan="3">HOT- ROD+Fitbit Summary</td>
<td>3</td>
<td>$69.16\%$</td>
<td>$62.50\%$</td>
</tr>
<tr>
<td>4</td>
<td>$64.64\%$</td>
<td>$64.61\%$</td>
</tr>
<tr>
<td>5</td>
<td>$65.99\%$</td>
<td>$7\mathbf{0}.\mathbf{13}\%$</td>
</tr>
</tbody>
</table>
# 5.2 Prediction Results
The experimental results for predicting IGTB assessment (personality and affect) and demographic information (job type, shift) are listed in Table 1 and Table 2 respectively. For predicting demographics, HOT-ROD features combined with Fitbit summary routine features achieve the best performance, with a better F1 score in predicting work shift than only using either the Fitbit summary feature or HOT-ROD routine features. We also observe that HOTROD reaches the best performance in forecasting neuroticism when the assigned number of clusters is 4. HOT-ROD features can also outperform Fitbit summary and combined feature set in classifying conscientiousness and openness when number of cluster is 3. Combining feature sets achieves the best performance in determining extraversion, agreeableness, and affect related variables.
Prior to time-series clustering, we perform the z-normalization of each time-series to remove variances between participants. We empirically experiment with the number of clusters $K=\{3,4,5\}$ in TICC. We set $\beta$ to be 10 in this experiment to encourage neighboring samples to be assigned to the same cluster. We want to highlight that we plan to systematically tune $\beta$ in our future works. Since some input time series may still contain portions of the missing elements (the segment that is missing above 15 minutes), one cluster output is associated with "missing measurements". We decide to ignore this cluster in the following analysis procedure. Finally, we applied the Hawkes Point Process to learn the infectivity between cluster transitions.
# 6 DISCUSSIONS
The results in Table 1 and Table 2 demonstrate foremost that personality and affect prediction from this data (collected in a natural setting outside of a well-controlled lab) is considerably challenging using standard physiologic features and simple machine learning techniques since none of the validation scores can reach above 70% even when the label is binarized. It also suggests that routine patterns, derived from wearable recordings by our HOT-ROD analysis, modestly improve performance. The prediction results demonstrate the HOT-ROD feature work comparatively reliably when the number of clusters is below 4. This may be because the available points for each cluster transition in a day decrease dramatically when the number of clusters increases, leading to an imprecise estimation of the Granger graph from time-series event.
We further identify that HOT-ROD features yield substantially better performances in predicting conscientiousness, which is known
5.1.3 Model Description. We observe that the distributions of ground truth assessments exhibited significant skew, posing a difficulty for most supervised learning algorithms, as they will be biased towards the majority group, leading to poor predictions on minority labels. Therefore, we choose to binarize the ground truth label as our final prediction target. We binarize personality, affect scores by the median split, while we categorize job type and work shift as nurses/non-nurses and day-shift/night shift, respectively. We then evaluate the efficacy of the Fitbit summary routine feature and HOT-ROD routine features extracted above by predicting binarized ground-truth labels using the Random Forest (RF) classifier. We select Random Forest models since they have considerable advantages over other techniques with respect to robustness to noise, tuning simplicity, and the ability to choose the most relevant features from
|
[
[
0.08660888671875,
0.0694580078125,
0.31787109375,
0.0694580078125,
0.31787109375,
0.0802001953125,
0.08660888671875,
0.0802001953125
],
[
0.08416748046875,
0.07891845703125,
0.52392578125,
0.07763671875,
0.52392578125,
0.091552734375,
0.08416748046875,
0.0921630859375
],
[
0.689453125,
0.0802001953125,
0.91259765625,
0.0802001953125,
0.91259765625,
0.09027099609375,
0.689453125,
0.09027099609375
],
[
0.0849609375,
0.10479736328125,
0.9140625,
0.10479736328125,
0.9140625,
0.11932373046875,
0.0849609375,
0.11932373046875
],
[
0.0849609375,
0.11932373046875,
0.324462890625,
0.11932373046875,
0.324462890625,
0.1331787109375,
0.0849609375,
0.1331787109375
],
[
0.2646484375,
0.149658203125,
0.316162109375,
0.149658203125,
0.316162109375,
0.1641845703125,
0.2646484375,
0.1641845703125
],
[
0.6044921875,
0.1483154296875,
0.685546875,
0.1483154296875,
0.685546875,
0.1634521484375,
0.6044921875,
0.1634521484375
],
[
0.7646484375,
0.149658203125,
0.82666015625,
0.149658203125,
0.82666015625,
0.1641845703125,
0.7646484375,
0.1641845703125
],
[
0.406982421875,
0.1514892578125,
0.51123046875,
0.1514892578125,
0.51123046875,
0.1641845703125,
0.406982421875,
0.1641845703125
],
[
0.324462890625,
0.1514892578125,
0.396240234375,
0.1514892578125,
0.396240234375,
0.1634521484375,
0.324462890625,
0.1634521484375
],
[
0.521484375,
0.1514892578125,
0.5947265625,
0.1514892578125,
0.5947265625,
0.1634521484375,
0.521484375,
0.1634521484375
],
[
0.69775390625,
0.1514892578125,
0.7548828125,
0.1514892578125,
0.7548828125,
0.1634521484375,
0.69775390625,
0.1634521484375
],
[
0.83642578125,
0.1514892578125,
0.90185546875,
0.1514892578125,
0.90185546875,
0.1634521484375,
0.83642578125,
0.1634521484375
],
[
0.1282958984375,
0.1641845703125,
0.2181396484375,
0.1666259765625,
0.2181396484375,
0.18115234375,
0.1282958984375,
0.1787109375
],
[
0.339111328125,
0.1654052734375,
0.381591796875,
0.1654052734375,
0.381591796875,
0.18115234375,
0.339111328125,
0.18115234375
],
[
0.43701171875,
0.1666259765625,
0.4794921875,
0.1666259765625,
0.4794921875,
0.18115234375,
0.43701171875,
0.18115234375
],
[
0.5361328125,
0.1666259765625,
0.580078125,
0.1666259765625,
0.580078125,
0.18115234375,
0.5361328125,
0.18115234375
],
[
0.62353515625,
0.1666259765625,
0.66650390625,
0.1666259765625,
0.66650390625,
0.18115234375,
0.62353515625,
0.18115234375
],
[
0.705078125,
0.1666259765625,
0.74853515625,
0.1666259765625,
0.74853515625,
0.18115234375,
0.705078125,
0.18115234375
],
[
0.77392578125,
0.1666259765625,
0.81787109375,
0.1666259765625,
0.81787109375,
0.1817626953125,
0.77392578125,
0.1817626953125
],
[
0.84619140625,
0.1666259765625,
0.8896484375,
0.1666259765625,
0.8896484375,
0.18115234375,
0.84619140625,
0.18115234375
],
[
0.281005859375,
0.1817626953125,
0.295654296875,
0.1817626953125,
0.295654296875,
0.199462890625,
0.281005859375,
0.199462890625
],
[
0.283447265625,
0.1685791015625,
0.294189453125,
0.1685791015625,
0.294189453125,
0.18115234375,
0.283447265625,
0.18115234375
],
[
0.338134765625,
0.18310546875,
0.381591796875,
0.18310546875,
0.381591796875,
0.1976318359375,
0.338134765625,
0.1976318359375
],
[
0.43701171875,
0.18310546875,
0.4794921875,
0.18310546875,
0.4794921875,
0.1976318359375,
0.43701171875,
0.1976318359375
],
[
0.53515625,
0.18310546875,
0.580078125,
0.18310546875,
0.580078125,
0.1976318359375,
0.53515625,
0.1976318359375
],
[
0.62353515625,
0.18310546875,
0.66650390625,
0.18310546875,
0.66650390625,
0.1976318359375,
0.62353515625,
0.1976318359375
],
[
0.70361328125,
0.18310546875,
0.74853515625,
0.18310546875,
0.74853515625,
0.1976318359375,
0.70361328125,
0.1976318359375
],
[
0.77392578125,
0.18310546875,
0.81787109375,
0.18310546875,
0.81787109375,
0.1976318359375,
0.77392578125,
0.1976318359375
],
[
0.84619140625,
0.18310546875,
0.8896484375,
0.18310546875,
0.8896484375,
0.1976318359375,
0.84619140625,
0.1976318359375
],
[
0.338134765625,
0.1988525390625,
0.381591796875,
0.1988525390625,
0.381591796875,
0.21337890625,
0.338134765625,
0.21337890625
],
[
0.435546875,
0.1988525390625,
0.48046875,
0.1988525390625,
0.48046875,
0.21337890625,
0.435546875,
0.21337890625
],
[
0.53515625,
0.1962890625,
0.57861328125,
0.1988525390625,
0.57666015625,
0.2139892578125,
0.53369140625,
0.212158203125
],
[
0.625,
0.1988525390625,
0.66650390625,
0.1988525390625,
0.66650390625,
0.21337890625,
0.625,
0.21337890625
],
[
0.705078125,
0.1988525390625,
0.74853515625,
0.1988525390625,
0.74853515625,
0.21337890625,
0.705078125,
0.21337890625
],
[
0.77392578125,
0.1988525390625,
0.81787109375,
0.1988525390625,
0.81787109375,
0.21337890625,
0.77392578125,
0.21337890625
],
[
0.84619140625,
0.1988525390625,
0.8896484375,
0.1988525390625,
0.8896484375,
0.21337890625,
0.84619140625,
0.21337890625
],
[
0.283447265625,
0.199462890625,
0.294189453125,
0.199462890625,
0.294189453125,
0.2139892578125,
0.283447265625,
0.2139892578125
],
[
0.1446533203125,
0.199462890625,
0.205078125,
0.199462890625,
0.205078125,
0.2108154296875,
0.1446533203125,
0.2108154296875
],
[
0.43701171875,
0.212158203125,
0.4794921875,
0.2139892578125,
0.47802734375,
0.2298583984375,
0.435546875,
0.2279052734375
],
[
0.77392578125,
0.212158203125,
0.81787109375,
0.2139892578125,
0.81640625,
0.2298583984375,
0.77294921875,
0.2279052734375
],
[
0.338134765625,
0.2139892578125,
0.381591796875,
0.2139892578125,
0.381591796875,
0.2298583984375,
0.338134765625,
0.2298583984375
],
[
0.283447265625,
0.2139892578125,
0.294189453125,
0.2139892578125,
0.294189453125,
0.2298583984375,
0.283447265625,
0.2298583984375
],
[
0.5361328125,
0.2139892578125,
0.57861328125,
0.2139892578125,
0.57861328125,
0.2298583984375,
0.5361328125,
0.2298583984375
],
[
0.625,
0.2139892578125,
0.66650390625,
0.2139892578125,
0.66650390625,
0.2298583984375,
0.625,
0.2298583984375
],
[
0.705078125,
0.2139892578125,
0.74658203125,
0.2139892578125,
0.74658203125,
0.2298583984375,
0.705078125,
0.2298583984375
],
[
0.84716796875,
0.2139892578125,
0.8896484375,
0.2139892578125,
0.8896484375,
0.2298583984375,
0.84716796875,
0.2298583984375
],
[
0.338134765625,
0.231689453125,
0.38232421875,
0.231689453125,
0.38232421875,
0.246826171875,
0.338134765625,
0.246826171875
],
[
0.435546875,
0.231689453125,
0.4794921875,
0.231689453125,
0.4794921875,
0.246826171875,
0.435546875,
0.246826171875
],
[
0.53515625,
0.231689453125,
0.580078125,
0.231689453125,
0.580078125,
0.246826171875,
0.53515625,
0.246826171875
],
[
0.62353515625,
0.231689453125,
0.66845703125,
0.231689453125,
0.66845703125,
0.246826171875,
0.62353515625,
0.246826171875
],
[
0.705078125,
0.231689453125,
0.74853515625,
0.231689453125,
0.74853515625,
0.246826171875,
0.705078125,
0.246826171875
],
[
0.77392578125,
0.231689453125,
0.81787109375,
0.231689453125,
0.81787109375,
0.246826171875,
0.77392578125,
0.246826171875
],
[
0.84619140625,
0.231689453125,
0.89208984375,
0.231689453125,
0.89208984375,
0.246826171875,
0.84619140625,
0.246826171875
],
[
0.096435546875,
0.2474365234375,
0.253173828125,
0.2474365234375,
0.253173828125,
0.261474609375,
0.096435546875,
0.261474609375
],
[
0.283447265625,
0.23291015625,
0.29736328125,
0.23291015625,
0.29736328125,
0.248779296875,
0.283447265625,
0.248779296875
],
[
0.338134765625,
0.2474365234375,
0.381591796875,
0.2474365234375,
0.381591796875,
0.261962890625,
0.338134765625,
0.261962890625
],
[
0.435546875,
0.2474365234375,
0.4794921875,
0.2474365234375,
0.4794921875,
0.261962890625,
0.435546875,
0.261962890625
],
[
0.62353515625,
0.2474365234375,
0.66845703125,
0.2474365234375,
0.66845703125,
0.261962890625,
0.62353515625,
0.261962890625
],
[
0.705078125,
0.2474365234375,
0.74853515625,
0.2474365234375,
0.74853515625,
0.261962890625,
0.705078125,
0.261962890625
],
[
0.84619140625,
0.2474365234375,
0.890625,
0.2474365234375,
0.890625,
0.261962890625,
0.84619140625,
0.261962890625
],
[
0.53515625,
0.248779296875,
0.57861328125,
0.248779296875,
0.57861328125,
0.26318359375,
0.53515625,
0.26318359375
],
[
0.283447265625,
0.248779296875,
0.295654296875,
0.248779296875,
0.295654296875,
0.261962890625,
0.283447265625,
0.261962890625
],
[
0.775390625,
0.25,
0.81787109375,
0.25,
0.81787109375,
0.261474609375,
0.775390625,
0.261474609375
],
[
0.43701171875,
0.26318359375,
0.4794921875,
0.26318359375,
0.4794921875,
0.27783203125,
0.43701171875,
0.27783203125
],
[
0.705078125,
0.261962890625,
0.74658203125,
0.264404296875,
0.74609375,
0.27978515625,
0.70361328125,
0.27783203125
],
[
0.5361328125,
0.264404296875,
0.57861328125,
0.264404296875,
0.57861328125,
0.279052734375,
0.5361328125,
0.279052734375
],
[
0.283447265625,
0.264404296875,
0.294189453125,
0.264404296875,
0.294189453125,
0.276611328125,
0.283447265625,
0.276611328125
],
[
0.625,
0.264404296875,
0.66845703125,
0.264404296875,
0.66845703125,
0.279052734375,
0.625,
0.279052734375
],
[
0.77392578125,
0.264404296875,
0.81884765625,
0.264404296875,
0.81884765625,
0.279052734375,
0.77392578125,
0.279052734375
],
[
0.84619140625,
0.264404296875,
0.890625,
0.264404296875,
0.890625,
0.279052734375,
0.84619140625,
0.279052734375
],
[
0.0849609375,
0.29736328125,
0.48193359375,
0.29736328125,
0.48193359375,
0.311279296875,
0.0849609375,
0.311279296875
],
[
0.339111328125,
0.26513671875,
0.3798828125,
0.26513671875,
0.3798828125,
0.276611328125,
0.339111328125,
0.276611328125
],
[
0.51708984375,
0.300537109375,
0.9140625,
0.300537109375,
0.9140625,
0.314453125,
0.51708984375,
0.314453125
],
[
0.08660888671875,
0.3125,
0.48193359375,
0.3125,
0.48193359375,
0.32568359375,
0.08660888671875,
0.32568359375
],
[
0.51708984375,
0.31494140625,
0.9140625,
0.31494140625,
0.9140625,
0.328857421875,
0.51708984375,
0.328857421875
],
[
0.0849609375,
0.32568359375,
0.27783203125,
0.32568359375,
0.27783203125,
0.339111328125,
0.0849609375,
0.339111328125
],
[
0.51708984375,
0.328857421875,
0.9140625,
0.328857421875,
0.9140625,
0.34228515625,
0.51708984375,
0.34228515625
],
[
0.515625,
0.34228515625,
0.9140625,
0.341552734375,
0.9140625,
0.35595703125,
0.515625,
0.356689453125
],
[
0.281005859375,
0.356689453125,
0.3349609375,
0.356689453125,
0.3349609375,
0.371826171875,
0.281005859375,
0.371826171875
],
[
0.346435546875,
0.356689453125,
0.4150390625,
0.356689453125,
0.4150390625,
0.371826171875,
0.346435546875,
0.371826171875
],
[
0.42724609375,
0.359130859375,
0.486083984375,
0.359130859375,
0.486083984375,
0.37060546875,
0.42724609375,
0.37060546875
],
[
0.51708984375,
0.356689453125,
0.91259765625,
0.356689453125,
0.91259765625,
0.37060546875,
0.51708984375,
0.37060546875
],
[
0.131591796875,
0.37255859375,
0.2303466796875,
0.375732421875,
0.2303466796875,
0.39013671875,
0.131591796875,
0.386962890625
],
[
0.35693359375,
0.375,
0.404296875,
0.375,
0.404296875,
0.389404296875,
0.35693359375,
0.389404296875
],
[
0.43310546875,
0.375,
0.48193359375,
0.375,
0.48193359375,
0.39013671875,
0.43310546875,
0.39013671875
],
[
0.51708984375,
0.37060546875,
0.9140625,
0.37060546875,
0.9140625,
0.3837890625,
0.51708984375,
0.3837890625
],
[
0.515625,
0.3837890625,
0.9111328125,
0.3837890625,
0.9111328125,
0.39453125,
0.515625,
0.39453125
],
[
0.301513671875,
0.375732421875,
0.31201171875,
0.375732421875,
0.31201171875,
0.38818359375,
0.301513671875,
0.38818359375
],
[
0.358642578125,
0.392578125,
0.404296875,
0.392578125,
0.404296875,
0.40771484375,
0.358642578125,
0.40771484375
],
[
0.43310546875,
0.392578125,
0.48046875,
0.392578125,
0.48046875,
0.40771484375,
0.43310546875,
0.40771484375
],
[
0.51708984375,
0.397705078125,
0.9140625,
0.397705078125,
0.9140625,
0.410888671875,
0.51708984375,
0.410888671875
],
[
0.358642578125,
0.40771484375,
0.404296875,
0.410400390625,
0.40283203125,
0.425537109375,
0.35693359375,
0.423583984375
],
[
0.301513671875,
0.393310546875,
0.314453125,
0.393310546875,
0.314453125,
0.4091796875,
0.301513671875,
0.4091796875
],
[
0.1478271484375,
0.410400390625,
0.214111328125,
0.410400390625,
0.214111328125,
0.421630859375,
0.1478271484375,
0.421630859375
],
[
0.301513671875,
0.410400390625,
0.314453125,
0.410400390625,
0.314453125,
0.425537109375,
0.301513671875,
0.425537109375
],
[
0.43310546875,
0.410400390625,
0.4794921875,
0.410400390625,
0.4794921875,
0.4248046875,
0.43310546875,
0.4248046875
],
[
0.51708984375,
0.410888671875,
0.82666015625,
0.410888671875,
0.82666015625,
0.4248046875,
0.51708984375,
0.4248046875
],
[
0.2998046875,
0.425537109375,
0.314453125,
0.425537109375,
0.314453125,
0.443115234375,
0.2998046875,
0.443115234375
],
[
0.35693359375,
0.427978515625,
0.404296875,
0.427978515625,
0.404296875,
0.442626953125,
0.35693359375,
0.442626953125
],
[
0.43310546875,
0.427978515625,
0.48046875,
0.427978515625,
0.48046875,
0.442626953125,
0.43310546875,
0.442626953125
],
[
0.301513671875,
0.44580078125,
0.314453125,
0.44580078125,
0.314453125,
0.4609375,
0.301513671875,
0.4609375
],
[
0.35693359375,
0.44580078125,
0.406005859375,
0.44580078125,
0.406005859375,
0.460205078125,
0.35693359375,
0.460205078125
],
[
0.43310546875,
0.44580078125,
0.48046875,
0.44580078125,
0.48046875,
0.460205078125,
0.43310546875,
0.460205078125
],
[
0.51953125,
0.45068359375,
0.5439453125,
0.45068359375,
0.5439453125,
0.462158203125,
0.51953125,
0.462158203125
],
[
0.5400390625,
0.451904296875,
0.5556640625,
0.451904296875,
0.5556640625,
0.458984375,
0.5400390625,
0.458984375
],
[
0.5537109375,
0.447509765625,
0.708984375,
0.448974609375,
0.708984375,
0.46337890625,
0.5537109375,
0.462158203125
],
[
0.09234619140625,
0.460205078125,
0.268798828125,
0.462158203125,
0.268798828125,
0.47607421875,
0.09234619140625,
0.473388671875
],
[
0.358642578125,
0.462158203125,
0.406005859375,
0.462158203125,
0.406005859375,
0.4765625,
0.358642578125,
0.4765625
],
[
0.43310546875,
0.462158203125,
0.48046875,
0.462158203125,
0.48046875,
0.4765625,
0.43310546875,
0.4765625
],
[
0.51708984375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.35693359375,
0.479736328125,
0.406005859375,
0.479736328125,
0.406005859375,
0.494384765625,
0.35693359375,
0.494384765625
],
[
0.30224609375,
0.46337890625,
0.314453125,
0.46337890625,
0.314453125,
0.4765625,
0.30224609375,
0.4765625
],
[
0.301513671875,
0.48095703125,
0.313720703125,
0.48095703125,
0.313720703125,
0.49365234375,
0.301513671875,
0.49365234375
],
[
0.431396484375,
0.479736328125,
0.48193359375,
0.479736328125,
0.48193359375,
0.494384765625,
0.431396484375,
0.494384765625
],
[
0.51708984375,
0.48095703125,
0.9140625,
0.48095703125,
0.9140625,
0.494384765625,
0.51708984375,
0.494384765625
],
[
0.5185546875,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.5068359375,
0.5185546875,
0.5068359375
],
[
0.515625,
0.50830078125,
0.9150390625,
0.5068359375,
0.9150390625,
0.521484375,
0.515625,
0.52294921875
],
[
0.101318359375,
0.52587890625,
0.48193359375,
0.52587890625,
0.48193359375,
0.5390625,
0.101318359375,
0.5390625
],
[
0.515625,
0.52294921875,
0.9140625,
0.52294921875,
0.9140625,
0.5361328125,
0.515625,
0.5361328125
],
[
0.08660888671875,
0.54150390625,
0.48046875,
0.54150390625,
0.48046875,
0.5517578125,
0.08660888671875,
0.5517578125
],
[
0.51708984375,
0.537109375,
0.9140625,
0.537109375,
0.9140625,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.0849609375,
0.5537109375,
0.48193359375,
0.5537109375,
0.48193359375,
0.5673828125,
0.0849609375,
0.5673828125
],
[
0.5185546875,
0.5517578125,
0.91259765625,
0.5517578125,
0.91259765625,
0.5625,
0.5185546875,
0.5625
],
[
0.08416748046875,
0.56640625,
0.48193359375,
0.5673828125,
0.48193359375,
0.58203125,
0.08416748046875,
0.58056640625
],
[
0.51708984375,
0.564453125,
0.9140625,
0.564453125,
0.9140625,
0.57763671875,
0.51708984375,
0.57763671875
],
[
0.0849609375,
0.58056640625,
0.48193359375,
0.58056640625,
0.48193359375,
0.5947265625,
0.0849609375,
0.5947265625
],
[
0.51708984375,
0.5771484375,
0.9140625,
0.5771484375,
0.9140625,
0.59033203125,
0.51708984375,
0.59033203125
],
[
0.08416748046875,
0.59326171875,
0.48193359375,
0.5947265625,
0.48193359375,
0.60791015625,
0.08416748046875,
0.60693359375
],
[
0.51708984375,
0.59228515625,
0.9140625,
0.59228515625,
0.9140625,
0.60595703125,
0.51708984375,
0.60595703125
],
[
0.08416748046875,
0.609375,
0.48193359375,
0.609375,
0.48193359375,
0.62255859375,
0.08416748046875,
0.62255859375
],
[
0.5146484375,
0.60498046875,
0.91259765625,
0.60498046875,
0.91259765625,
0.61865234375,
0.5146484375,
0.61865234375
],
[
0.08416748046875,
0.62255859375,
0.48193359375,
0.62255859375,
0.48193359375,
0.63623046875,
0.08416748046875,
0.63623046875
],
[
0.515625,
0.619140625,
0.9140625,
0.62060546875,
0.9140625,
0.6337890625,
0.515625,
0.63330078125
],
[
0.0849609375,
0.63623046875,
0.48193359375,
0.63623046875,
0.48193359375,
0.6494140625,
0.0849609375,
0.6494140625
],
[
0.51708984375,
0.6337890625,
0.8544921875,
0.6337890625,
0.8544921875,
0.64794921875,
0.51708984375,
0.64794921875
],
[
0.0849609375,
0.6494140625,
0.48046875,
0.6494140625,
0.48046875,
0.66357421875,
0.0849609375,
0.66357421875
],
[
0.0849609375,
0.6640625,
0.483642578125,
0.6640625,
0.483642578125,
0.67822265625,
0.0849609375,
0.67822265625
],
[
0.521484375,
0.6748046875,
0.53125,
0.6748046875,
0.53125,
0.681640625,
0.521484375,
0.681640625
],
[
0.5390625,
0.67041015625,
0.6611328125,
0.67041015625,
0.6611328125,
0.6845703125,
0.5390625,
0.6845703125
],
[
0.08660888671875,
0.6796875,
0.1993408203125,
0.6796875,
0.1993408203125,
0.68994140625,
0.08660888671875,
0.68994140625
],
[
0.51708984375,
0.689453125,
0.9140625,
0.689453125,
0.9140625,
0.70263671875,
0.51708984375,
0.70263671875
],
[
0.0849609375,
0.70263671875,
0.491943359375,
0.70263671875,
0.491943359375,
0.71630859375,
0.0849609375,
0.71630859375
],
[
0.5185546875,
0.705078125,
0.91259765625,
0.705078125,
0.91259765625,
0.71533203125,
0.5185546875,
0.71533203125
],
[
0.0849609375,
0.71728515625,
0.48193359375,
0.71728515625,
0.48193359375,
0.73095703125,
0.0849609375,
0.73095703125
],
[
0.51708984375,
0.71728515625,
0.9140625,
0.71728515625,
0.9140625,
0.73095703125,
0.51708984375,
0.73095703125
],
[
0.0849609375,
0.73095703125,
0.48193359375,
0.73095703125,
0.48193359375,
0.744140625,
0.0849609375,
0.744140625
],
[
0.515625,
0.72998046875,
0.9140625,
0.73095703125,
0.9140625,
0.74560546875,
0.515625,
0.744140625
],
[
0.0849609375,
0.74560546875,
0.48193359375,
0.74560546875,
0.48193359375,
0.7587890625,
0.0849609375,
0.7587890625
],
[
0.5185546875,
0.74560546875,
0.91259765625,
0.74560546875,
0.91259765625,
0.755859375,
0.5185546875,
0.755859375
],
[
0.08660888671875,
0.76025390625,
0.48193359375,
0.76025390625,
0.48193359375,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.51708984375,
0.7568359375,
0.9140625,
0.7568359375,
0.9140625,
0.77099609375,
0.51708984375,
0.77099609375
],
[
0.0849609375,
0.77294921875,
0.48193359375,
0.77294921875,
0.48193359375,
0.7861328125,
0.0849609375,
0.7861328125
],
[
0.515625,
0.77099609375,
0.9150390625,
0.771484375,
0.9150390625,
0.7861328125,
0.515625,
0.78515625
],
[
0.0849609375,
0.7861328125,
0.48193359375,
0.7861328125,
0.48193359375,
0.7998046875,
0.0849609375,
0.7998046875
],
[
0.51708984375,
0.7861328125,
0.9140625,
0.7861328125,
0.9140625,
0.7998046875,
0.51708984375,
0.7998046875
],
[
0.08416748046875,
0.7998046875,
0.48193359375,
0.798828125,
0.48193359375,
0.8125,
0.08416748046875,
0.81298828125
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.0849609375,
0.814453125,
0.483642578125,
0.814453125,
0.483642578125,
0.82763671875,
0.0849609375,
0.82763671875
],
[
0.5185546875,
0.81494140625,
0.91259765625,
0.81494140625,
0.91259765625,
0.82568359375,
0.5185546875,
0.82568359375
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8291015625,
0.91259765625,
0.8291015625,
0.91259765625,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84326171875,
0.91259765625,
0.84326171875,
0.91259765625,
0.85400390625,
0.51708984375,
0.85400390625
],
[
0.0849609375,
0.85595703125,
0.48193359375,
0.85595703125,
0.48193359375,
0.869140625,
0.0849609375,
0.869140625
],
[
0.515625,
0.85595703125,
0.77294921875,
0.85595703125,
0.77294921875,
0.869140625,
0.515625,
0.869140625
],
[
0.08416748046875,
0.86865234375,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.08416748046875,
0.8818359375
],
[
0.53369140625,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.53369140625,
0.880859375
],
[
0.08660888671875,
0.88525390625,
0.48046875,
0.88525390625,
0.48046875,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.515625,
0.88330078125,
0.9189453125,
0.88330078125,
0.9189453125,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Learning Behavioral Representations of Routines",
"From Large-scale Unlabeled Wearable Time-series Data Streams using Hawkes Point Process",
"MiLeTS'23,August2023,LongBeach,CA,USA",
"Table 1: Table shows quantitative comparison (macro-F1 score) of different routine features to predict Initial Assessment",
"survey using Random Forest Model",
"Clusters",
"Agreeableness",
"Pos-Affect",
"Conscientiousness",
"Neuroticism",
"Extraversion",
"Openness",
"Neg-Affect",
"Fitbit Summary",
"60.85%",
"49.84%",
"64.91%",
"63.95%",
"60.86%",
"58.24%",
"64.88%",
"3",
"·",
"63.23%",
"66.60%",
"62.12%",
"63.64%",
"62.19%",
"54.32%",
"62.82%",
"56.21%",
"68.08%",
"62.64%",
"63.85%",
"54.08%",
"54.39%",
"58.00%",
"4",
"HOT-ROD",
"60.29%",
"62.48%",
"61.07%",
"5",
"58.53%",
"64.00%",
"57.83%",
"58.30%",
"63.53%",
"52.09%",
"66.25%",
"70.13%",
"60.86%",
"58.35%",
"65.72%",
"HOT-ROD + Fitbit Summary",
"3",
"59.84%",
"63.65%",
"64.76%",
"56.09%",
"61.91%",
"65.87%",
"4",
"59.35%",
"58.03%",
"59.58%",
"65.02%",
"5",
"64.40%",
"63.67%",
"60.31%",
"Table 2: Table shows quantitative comparison (macro-F1",
"60.56%",
"high-dimensional data input, where many features are often redun-",
"score) of different routine features to predict demographics",
"dant [24]. Specifically, we perform the predictions using three sets",
"using Random Forest Model.",
"of features: 1. Fitbit summary routine feature; 2. HOT-ROD routine",
"feature; 3. Fitbit summary routine feature and HOT-ROD routine",
"Clusters",
"Work Shift",
"Job Type",
"feature. We perform 5-fold cross-validation and report the average",
"Fitbit Summary",
"68.92%",
"60.55%",
"results in the macro-F1 score. We grid search the hyper-parameters",
"in theRF model asfollows:1.Number of estimators:[10,20,30l:2",
"",
"58.76%",
"68.79%",
"Feature selection criterion: [\"gini\", \"entropy\"]; 3. Max depth of the",
"55.98%",
"3",
"HOT-ROD",
"4",
"67.42%",
"tree: [4, 5, 6]; 4. Minimum samples to split: [2, 3, 5].",
"5",
"62.68%",
"70.13%",
"3",
"69.16%",
"62.50%",
"5.2",
"",
"PredictionResults",
"HOT-ROD+Fitbit Summary",
"64.64%",
"64.61%",
"The experimental results for predicting IGTB assessment (person-",
"65.99%",
"4",
"5",
"70.13%",
"ality and affect)and demographic information (job type, shift)are",
"listedinTable1andTable2respectively.Forpredictingdemo-",
"graphics, HOT-ROD features combined with Fitbit summary rou-",
"Prior to time-series clustering, we perform the z-normalization",
"tine features achieve the best performance, with a better F1 score",
"ofeachtime-seriestoremovevariances betweenparticipants.We",
"in predicting work shift than only using either the Fitbit summary",
"empirically experiment with the number of clusters K = (3, 4, 5) in",
"feature orHOT-RODroutinefeatures.WealsoobservethatHOT",
"TICC.We set β to be10 in this experiment to encourage neighboring",
"ROD reaches the best performance in forecasting neuroticism when",
"samples to be assigned to the same cluster. We want to highlight",
"the assigned number of clusters is 4. HOT-ROD features can also",
"that we plan to systematically tune β in our future works. Since",
"outperform Fitbit summary and combined feature set in classifying",
"some input time series may still contain portions of the missing",
" conscientiousness and openness when number of cluster is 3. Com-",
"elements (the segment that is missing above 15 minutes), one cluster",
"bining feature sets achieves the best performance in determining",
"output is associated with \"missing measurements\".We decide to",
"extraversion, agreeableness, and affect related variables.",
"ignore this cluster in the following analysis procedure. Finally, we",
"applied the Hawkes Point Process to learn the infectivity between",
"6",
" DISCUSSIONS",
"clustertransitions.",
"The results in Table 1 and Table 2 demonstrate foremost that per-",
"5.1.3Model Description.Weobservethat the distributions ofground",
"sonalityand affectprediction from this data(collected ina natural",
"truth assessments exhibited significant skew, posing a difficulty",
"setting outside of a well-controlled lab) is considerably challenging",
"for most supervised learning algorithms, as they will be biased to-",
"using standard physiologic features and simple machine learning",
"wards the majority group, leading to poor predictions on minority",
"technigues sincenoneofthevalidation scorescanreachabove70%",
"labels.Therefore,we choosetobinarizethegroundtruthlabelas",
"even when the label is binarized. It also suggests that routine pat-",
"our final prediction target. We binarize personality, affect scores",
"terns, derived from wearable recordings by our HOT-ROD analysis,",
"by the median split, while we categorize job type and work shift",
"modestly improve performance. The prediction results demonstrate",
"as nurses/non-nurses and day-shift/night shift, respectively.We",
"the HOT-ROD feature work comparatively reliably when the num-",
"then evaluate the efficacy of the Fitbit summary routine feature and",
"ber of clusters is below 4.This maybebecause the availablepoints",
"HOT-RODroutinefeaturesextracted abovebypredictingbinarized",
"for each clustertransitionina day decreasedramatically whenthe",
"ground-truth labels using the Random Forest (RF) classifier. We se-",
"number of clusters increases, leadingto an imprecise estimation of",
"lect RandomForest models since they have considerable advantages",
"the Granger graph from time-series event.",
"over othertechniques with respect to robustness to noise, tuning",
"WefurtheridentifythatHOT-RODfeaturesyieldsubstantially",
"simplicity,and the abilityto choosethemost relevantfeaturesfrom",
"better performances in predicting conscientiousness, which is known"
] |
{
"arxiv_id": "2307.04445",
"language": "en",
"page_idx": 4
}
|
||
2106.05691-en-0
|
2106.05691-en-0.jpg
|
# $\textbf{Marginal Utility Diminishes: Exploring the Minimum Knowledge for}$ BERT Knowledge Distillation
Yuanxin Liu$^{1,2*}$Fandong Meng$^3,\textbf{Zheng Lin}^{1\dagger}$,Weiping Wang$^1,\textbf{Jie Zhou}^3$
$^{1}$Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China $^{2}$School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China
$^3$Pattern Recognition Center, WeChat AI, Tencent Inc, China
$\{$liuyuanxin,linzheng,wangweiping$\}@$iie.ac.cn, $\{$fandongmeng,withtomzhou$\}@$tencent.com
## Abstract
Recently, knowledge distillation (KD) has shown great success in BERT compression. Instead of only learning from the teacher's soft label as in conventional KD, researchers find that the rich information contained in the hidden layers of BERT is conducive to the student's performance. To better exploit the hidden knowledge, a common practice is to force the student to deeply mimic the teacher's hidden states of all the tokens in a layer-wise manner. In this paper, however, we observe that although distilling the teacher's hidden state knowledge (HSK) is helpful, the performance gain (marginal utility) diminishes quickly as more HSK is distilled. To understand this effect, we conduct a series of analysis. Specifically, we divide the HSK of BERT into three dimensions, namely depth, length and width. We first investigate a variety of strategies to extract crucial knowledge for each single dimension and then jointly compress the three dimensions. In this way, we show that 1) the student's performance can be improved by extracting and distilling the crucial HSK, and 2) using a tiny fraction of HSK can achieve the same performance as extensive HSK distillation. Based on the second finding, we further propose an efficient KD paradigm to compress BERT, which does not require loading the teacher during the training of student. For two kinds of student models and computing devices, the proposed KD paradigm gives rise to training speedup of 2.7×∼3.4×.
## 1 Introduction
Since the launch of BERT (Devlin et al., 2019), pre-trained language models (PLMs) have been advancing the state-of-the arts (SOTAs) in a wide
* Work was done when Yuanxin Liu was an intern at Pattern Recognition Center, WeChat AI, Tencent Inc, China.
$^{\dagger}$ Zheng Lin is the corresponding author.
<FigureHere>
Figure 1: The Acc variation of ROSITA (Liu et al., 2021) and TinyBERT (Jiao et al., 2020) on QNLI with the increase of HSK
range of NLP tasks. At the same time, the growing size of PLMs has inspired a wave of research interest in model compression (Han et al., 2016) in the NLP community, which aims to facilitate the deployment of the powerful PLMs to resourcelimited scenarios.
Knowledge distillation (KD) (Hinton et al., 2015) is an effective technique in model compression. In conventional KD, the student model is trained to imitate the teacher's prediction over classes, i.e., the soft labels. Subsequently, Romero et al. (2015) find that the intermediate representations in the teacher's hidden layers can also serve as a useful source of knowledge. As an initial attempt to introduce this idea to BERT compression, PKD (Sun et al., 2019) proposed to distill representations of the [CLS] token in BERT's hidden layers, and later studies (Jiao et al., 2020; Sun et al., 2020; Hou et al., 2020; Liu et al., 2021) extend the distillation of hidden state knowledge ( HSK) to all the tokens.
In contrast to the previous work that attempts to increase the amount of HSK, in this paper we explore towards the opposite direction to“compress” HSK. We make the observation that although distilling HSK is helpful, the marginal utility dimin-
|
[
[
0.1470947265625,
0.0843505859375,
0.8564453125,
0.0843505859375,
0.8564453125,
0.09857177734375,
0.1470947265625,
0.09857177734375
],
[
0.349609375,
0.10394287109375,
0.6552734375,
0.10394287109375,
0.6552734375,
0.1181640625,
0.349609375,
0.1181640625
],
[
0.19580078125,
0.141357421875,
0.81494140625,
0.14013671875,
0.81494140625,
0.157958984375,
0.19580078125,
0.1591796875
],
[
0.1630859375,
0.157958984375,
0.84619140625,
0.1591796875,
0.84619140625,
0.174560546875,
0.1630859375,
0.17333984375
],
[
0.1512451171875,
0.17333984375,
0.8564453125,
0.174560546875,
0.8564453125,
0.1923828125,
0.1512451171875,
0.191162109375
],
[
0.257080078125,
0.191162109375,
0.75146484375,
0.193603515625,
0.75146484375,
0.2078857421875,
0.257080078125,
0.2060546875
],
[
0.1336669921875,
0.210205078125,
0.87890625,
0.210205078125,
0.87890625,
0.2244873046875,
0.1336669921875,
0.2244873046875
],
[
0.265625,
0.265380859375,
0.34375,
0.265380859375,
0.34375,
0.27978515625,
0.265625,
0.27978515625
],
[
0.56982421875,
0.27490234375,
0.5966796875,
0.27490234375,
0.5966796875,
0.2861328125,
0.56982421875,
0.2861328125
],
[
0.56982421875,
0.2861328125,
0.5947265625,
0.2861328125,
0.5947265625,
0.300537109375,
0.56982421875,
0.300537109375
],
[
0.02520751953125,
0.29345703125,
0.058837890625,
0.29345703125,
0.058837890625,
0.345703125,
0.02520751953125,
0.345703125
],
[
0.1453857421875,
0.297607421875,
0.462158203125,
0.297607421875,
0.462158203125,
0.311279296875,
0.1453857421875,
0.311279296875
],
[
0.55029296875,
0.29638671875,
0.603515625,
0.29638671875,
0.603515625,
0.336181640625,
0.55029296875,
0.336181640625
],
[
0.1470947265625,
0.312255859375,
0.46044921875,
0.312255859375,
0.46044921875,
0.325927734375,
0.1470947265625,
0.325927734375
],
[
0.1470947265625,
0.325927734375,
0.46044921875,
0.325927734375,
0.46044921875,
0.339111328125,
0.1470947265625,
0.339111328125
],
[
0.55615234375,
0.32470703125,
0.5966796875,
0.32666015625,
0.59326171875,
0.342529296875,
0.552734375,
0.339111328125
],
[
0.023529052734375,
0.336181640625,
0.060516357421875,
0.336181640625,
0.060516357421875,
0.410888671875,
0.023529052734375,
0.410888671875
],
[
0.1470947265625,
0.34033203125,
0.462158203125,
0.34033203125,
0.462158203125,
0.35400390625,
0.1470947265625,
0.35400390625
],
[
0.57080078125,
0.345703125,
0.59326171875,
0.345703125,
0.59326171875,
0.355712890625,
0.57080078125,
0.355712890625
],
[
0.76318359375,
0.349853515625,
0.82275390625,
0.349853515625,
0.82275390625,
0.35986328125,
0.76318359375,
0.35986328125
],
[
0.148681640625,
0.355712890625,
0.459716796875,
0.355712890625,
0.459716796875,
0.366455078125,
0.148681640625,
0.366455078125
],
[
0.56982421875,
0.358154296875,
0.59765625,
0.358154296875,
0.59765625,
0.369384765625,
0.56982421875,
0.369384765625
],
[
0.75390625,
0.363525390625,
0.7705078125,
0.363525390625,
0.7705078125,
0.366455078125,
0.75390625,
0.366455078125
],
[
0.76904296875,
0.361083984375,
0.830078125,
0.361083984375,
0.830078125,
0.371826171875,
0.76904296875,
0.371826171875
],
[
0.1470947265625,
0.3681640625,
0.46044921875,
0.3681640625,
0.46044921875,
0.3818359375,
0.1470947265625,
0.3818359375
],
[
0.65283203125,
0.377685546875,
0.6796875,
0.377685546875,
0.6796875,
0.385986328125,
0.65283203125,
0.385986328125
],
[
0.759765625,
0.377685546875,
0.7841796875,
0.377685546875,
0.7841796875,
0.385986328125,
0.759765625,
0.385986328125
],
[
0.81787109375,
0.377197265625,
0.8359375,
0.377197265625,
0.8359375,
0.385986328125,
0.81787109375,
0.385986328125
],
[
0.70947265625,
0.380126953125,
0.72607421875,
0.380126953125,
0.72607421875,
0.38427734375,
0.70947265625,
0.38427734375
],
[
0.1470947265625,
0.383056640625,
0.46044921875,
0.383056640625,
0.46044921875,
0.396728515625,
0.1470947265625,
0.396728515625
],
[
0.603515625,
0.380126953125,
0.60986328125,
0.380126953125,
0.60986328125,
0.385498046875,
0.603515625,
0.385498046875
],
[
0.6455078125,
0.38427734375,
0.78662109375,
0.385498046875,
0.78662109375,
0.396728515625,
0.6455078125,
0.3955078125
],
[
0.1470947265625,
0.39794921875,
0.46044921875,
0.39794921875,
0.46044921875,
0.410888671875,
0.1470947265625,
0.410888671875
],
[
0.1470947265625,
0.410888671875,
0.46044921875,
0.410888671875,
0.46044921875,
0.424560546875,
0.1470947265625,
0.424560546875
],
[
0.020172119140625,
0.41748046875,
0.058837890625,
0.416259765625,
0.0638427734375,
0.4912109375,
0.02520751953125,
0.4921875
],
[
0.51416015625,
0.413330078125,
0.88671875,
0.413330078125,
0.88671875,
0.427001953125,
0.51416015625,
0.427001953125
],
[
0.1453857421875,
0.42578125,
0.462158203125,
0.42578125,
0.462158203125,
0.439453125,
0.1453857421875,
0.439453125
],
[
0.51416015625,
0.427490234375,
0.884765625,
0.427490234375,
0.884765625,
0.441162109375,
0.51416015625,
0.441162109375
],
[
0.1470947265625,
0.439453125,
0.462158203125,
0.439453125,
0.462158203125,
0.453125,
0.1470947265625,
0.453125
],
[
0.51416015625,
0.4423828125,
0.6552734375,
0.4423828125,
0.6552734375,
0.4560546875,
0.51416015625,
0.4560546875
],
[
0.1470947265625,
0.45361328125,
0.46044921875,
0.45361328125,
0.46044921875,
0.46728515625,
0.1470947265625,
0.46728515625
],
[
0.1470947265625,
0.468505859375,
0.46044921875,
0.468505859375,
0.46044921875,
0.482177734375,
0.1470947265625,
0.482177734375
],
[
0.1470947265625,
0.482177734375,
0.462158203125,
0.482177734375,
0.462158203125,
0.495361328125,
0.1470947265625,
0.495361328125
],
[
0.51416015625,
0.48681640625,
0.88671875,
0.48681640625,
0.88671875,
0.501953125,
0.51416015625,
0.501953125
],
[
0.0193328857421875,
0.498779296875,
0.060516357421875,
0.498779296875,
0.0638427734375,
0.7060546875,
0.0218505859375,
0.7060546875
],
[
0.1470947265625,
0.496337890625,
0.462158203125,
0.496337890625,
0.462158203125,
0.51025390625,
0.1470947265625,
0.51025390625
],
[
0.51416015625,
0.50439453125,
0.88330078125,
0.50439453125,
0.88330078125,
0.51708984375,
0.51416015625,
0.51708984375
],
[
0.148681640625,
0.5126953125,
0.459716796875,
0.5126953125,
0.459716796875,
0.5224609375,
0.148681640625,
0.5224609375
],
[
0.1453857421875,
0.52490234375,
0.46044921875,
0.52392578125,
0.46044921875,
0.53759765625,
0.1453857421875,
0.5380859375
],
[
0.51416015625,
0.51953125,
0.884765625,
0.51953125,
0.884765625,
0.533203125,
0.51416015625,
0.533203125
],
[
0.1470947265625,
0.5390625,
0.4638671875,
0.5390625,
0.4638671875,
0.552734375,
0.1470947265625,
0.552734375
],
[
0.51416015625,
0.5361328125,
0.884765625,
0.5361328125,
0.884765625,
0.5498046875,
0.51416015625,
0.5498046875
],
[
0.1453857421875,
0.5517578125,
0.462158203125,
0.552734375,
0.462158203125,
0.56787109375,
0.1453857421875,
0.56640625
],
[
0.51416015625,
0.55029296875,
0.88330078125,
0.55029296875,
0.88330078125,
0.56396484375,
0.51416015625,
0.56396484375
],
[
0.1470947265625,
0.56640625,
0.46044921875,
0.56640625,
0.46044921875,
0.580078125,
0.1470947265625,
0.580078125
],
[
0.51416015625,
0.568359375,
0.64697265625,
0.568359375,
0.64697265625,
0.58203125,
0.51416015625,
0.58203125
],
[
0.148681640625,
0.5830078125,
0.46044921875,
0.5830078125,
0.46044921875,
0.59326171875,
0.148681640625,
0.59326171875
],
[
0.53173828125,
0.5859375,
0.88671875,
0.5859375,
0.88671875,
0.60107421875,
0.53173828125,
0.60107421875
],
[
0.1470947265625,
0.595703125,
0.46044921875,
0.595703125,
0.46044921875,
0.609375,
0.1470947265625,
0.609375
],
[
0.5126953125,
0.60107421875,
0.88671875,
0.6025390625,
0.88671875,
0.61767578125,
0.5126953125,
0.615234375
],
[
0.1470947265625,
0.60986328125,
0.46044921875,
0.60986328125,
0.46044921875,
0.62353515625,
0.1470947265625,
0.62353515625
],
[
0.5126953125,
0.6181640625,
0.88330078125,
0.61767578125,
0.88330078125,
0.6318359375,
0.5126953125,
0.6328125
],
[
0.148681640625,
0.6259765625,
0.459716796875,
0.6259765625,
0.459716796875,
0.6357421875,
0.148681640625,
0.6357421875
],
[
0.1453857421875,
0.63818359375,
0.46044921875,
0.63720703125,
0.46044921875,
0.65185546875,
0.1453857421875,
0.65283203125
],
[
0.5126953125,
0.63525390625,
0.884765625,
0.63525390625,
0.884765625,
0.6484375,
0.5126953125,
0.6484375
],
[
0.1470947265625,
0.65283203125,
0.46044921875,
0.65283203125,
0.46044921875,
0.66650390625,
0.1470947265625,
0.66650390625
],
[
0.5126953125,
0.65087890625,
0.88330078125,
0.65087890625,
0.88330078125,
0.6650390625,
0.5126953125,
0.6650390625
],
[
0.1453857421875,
0.66748046875,
0.46044921875,
0.66748046875,
0.46044921875,
0.68115234375,
0.1453857421875,
0.68115234375
],
[
0.5126953125,
0.6650390625,
0.88671875,
0.66748046875,
0.88671875,
0.68115234375,
0.5126953125,
0.6787109375
],
[
0.148681640625,
0.68310546875,
0.46044921875,
0.68310546875,
0.46044921875,
0.693359375,
0.148681640625,
0.693359375
],
[
0.51416015625,
0.68310546875,
0.884765625,
0.68310546875,
0.884765625,
0.69677734375,
0.51416015625,
0.69677734375
],
[
0.1470947265625,
0.6953125,
0.462158203125,
0.6953125,
0.462158203125,
0.708984375,
0.1470947265625,
0.708984375
],
[
0.51416015625,
0.69970703125,
0.88671875,
0.69970703125,
0.88671875,
0.71337890625,
0.51416015625,
0.71337890625
],
[
0.1453857421875,
0.71044921875,
0.462158203125,
0.71044921875,
0.462158203125,
0.7236328125,
0.1453857421875,
0.7236328125
],
[
0.51416015625,
0.71630859375,
0.88330078125,
0.71630859375,
0.88330078125,
0.72705078125,
0.51416015625,
0.72705078125
],
[
0.1470947265625,
0.7255859375,
0.459716796875,
0.7255859375,
0.459716796875,
0.736328125,
0.1470947265625,
0.736328125
],
[
0.51416015625,
0.732421875,
0.88671875,
0.732421875,
0.88671875,
0.7451171875,
0.51416015625,
0.7451171875
],
[
0.1470947265625,
0.73828125,
0.462158203125,
0.73828125,
0.462158203125,
0.751953125,
0.1470947265625,
0.751953125
],
[
0.1453857421875,
0.751953125,
0.374755859375,
0.751953125,
0.374755859375,
0.765625,
0.1453857421875,
0.765625
],
[
0.51416015625,
0.74755859375,
0.884765625,
0.74755859375,
0.884765625,
0.76123046875,
0.51416015625,
0.76123046875
],
[
0.51416015625,
0.76416015625,
0.88671875,
0.76416015625,
0.88671875,
0.77783203125,
0.51416015625,
0.77783203125
],
[
0.12939453125,
0.783203125,
0.26123046875,
0.78369140625,
0.26123046875,
0.798828125,
0.12939453125,
0.79736328125
],
[
0.51416015625,
0.77978515625,
0.88671875,
0.77978515625,
0.88671875,
0.79345703125,
0.51416015625,
0.79345703125
],
[
0.1243896484375,
0.7880859375,
0.1302490234375,
0.7880859375,
0.1302490234375,
0.79443359375,
0.1243896484375,
0.79443359375
],
[
0.5126953125,
0.79443359375,
0.884765625,
0.7958984375,
0.884765625,
0.81005859375,
0.5126953125,
0.80859375
],
[
0.11846923828125,
0.8125,
0.49072265625,
0.8125,
0.49072265625,
0.826171875,
0.11846923828125,
0.826171875
],
[
0.51416015625,
0.8125,
0.69140625,
0.8125,
0.69140625,
0.826171875,
0.51416015625,
0.826171875
],
[
0.11846923828125,
0.8291015625,
0.49072265625,
0.8291015625,
0.49072265625,
0.84326171875,
0.11846923828125,
0.84326171875
],
[
0.53125,
0.830078125,
0.884765625,
0.83056640625,
0.884765625,
0.845703125,
0.53125,
0.84423828125
],
[
0.1168212890625,
0.84423828125,
0.489013671875,
0.84326171875,
0.489013671875,
0.85791015625,
0.1168212890625,
0.859375
],
[
0.5126953125,
0.8466796875,
0.88671875,
0.8466796875,
0.88671875,
0.86083984375,
0.5126953125,
0.86083984375
],
[
0.5126953125,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.5126953125,
0.876953125
],
[
0.1512451171875,
0.87060546875,
0.492431640625,
0.87060546875,
0.492431640625,
0.88427734375,
0.1512451171875,
0.88427734375
],
[
0.11846923828125,
0.8828125,
0.4873046875,
0.8828125,
0.4873046875,
0.896484375,
0.11846923828125,
0.896484375
],
[
0.5126953125,
0.87744140625,
0.88671875,
0.87890625,
0.88671875,
0.8935546875,
0.5126953125,
0.892578125
],
[
0.1495361328125,
0.896484375,
0.39990234375,
0.896484375,
0.39990234375,
0.91015625,
0.1495361328125,
0.91015625
],
[
0.51416015625,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51416015625,
0.9091796875
]
] |
[
"Marginal Utility Diminishes: Exploring the Minimum Knowledge for",
"BERTKnowledgeDistillation",
"Yuanxin Liu',2, Fandong Meng\", Zheng Lin'f Weiping Wang', Jie Zhou?",
"'Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China",
"2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China",
"3Pattern Recognition Center, WeChat AI, Tencent Inc, China",
"[liuyuanxin,linzheng,wangweiping] @iie.ac.cn, [fandongmeng,withtomzhou] @tencent.com",
"Abstract",
"88.0",
"87.5",
"1",
"Recently, knowledge distillation (KD) has",
"E",
"shown great success in BERT compression. In-",
"stead of only learning from the teacher's soft",
"86.0",
"unr ol",
"label as in conventional KD, researchers find",
"85.5",
"ROSITA",
"that the rich information contained in the hid-",
"85.0",
"",
"TinyBERT4",
"den layers of BERT is conducive to the stu-",
"10~3",
"10-1",
"100",
"10-",
"dent's performance. To better exploit the hid-",
"0",
"Amount of HSK (Normalized)",
"den knowledge, a common practice is to force",
"the student to deeply mimic the teacher's hid-",
"[T's]",
"Figure I: The Acc variation of ROSITA (Liu et al.",
"den states of all the tokens in a layer-wise man-",
"2021) and TinyBERT (Jiao et al., 2020) on QNLI with",
"ner. In this paper, however, we observe that",
"the increase of HSK.",
"although distilling the teacher's hidden state",
"knowledge (HSK) is helpful, the performance",
"gain (marginal utility) diminishes quickly as",
"range of NLP tasks. At the same time, the grow-",
"1A169009012:",
"more HSK is distilled. To understand this ef-",
"ing size of PLMs has inspired a wave of research",
"fect, we conduct a series of analysis. Specifi-",
"cally, we divide the HSK of BERT into three",
"interest in model compression (Han et al., 2016)",
"dimensions, namely depth, length and width.",
"in the NLP community, which aims to facilitate",
"We first investigate a variety of strategies to",
"the deployment of the powerful PLMs to resource-",
"extract crucial knowledge for each single di-",
"limited scenarios.",
"mension and then jointly compress the three",
"Knowledge distillation (KD) (Hinton et al.,",
"dimensions. In this way, we show that 1) the",
"2015) is an effective technique in model com-",
"student's performance can be improved by ex-",
"pression. In conventional KD, the student model",
"tracting and distillingthe crucialHSK,and2)",
"using a tiny fraction of HSK can achieve the",
"is trained to imitate the teacher's prediction",
"same performance as extensive HSK distilla-",
"over classes, i.e., the soft labels. Subsequently,",
"tion.Based on the second finding, we fur-",
"Romero et al. (2015) find that the intermediate rep-",
"therproposeanefficientKDparadigmtocom-",
"resentations in the teacher's hidden layers can also",
"press BERT, which does not require loading",
"serve as a useful source of knowledge. As an ini-",
"the teacher during the training of student. For",
"tial attempt to introduce this idea to BERT com-",
"twokinds of studentmodelsandcomputingde",
"pression, PKD (Sun et al., 2019) proposed to dis-",
"vices, the proposed KD paradigm gives rise to",
"training speedup of 2.7× ~3.4×.",
"till representations of the [CLS] token in BERT's",
"hidden layers, and later studies (Jiao et al., 2020;",
"Introduction",
"Sun et al., 2020; Hou et al., 2020; Liu et al., 2021)",
"-",
"extend the distillation of hidden state knowledge",
"Since the launch of BERT (Devlin et al., 2019),",
"(HSK) to all the tokens.",
"pre-trained language models (PLMs) have been ad-",
"In contrast to the previous work that attempts to",
"vancing the state-of-the arts (SOTAs) in a wide",
"increase the amount of HSK, in this paper we ex-",
"plore towards the opposite direction to “compress\"",
"*Work was done when Yuanxin Liu was an intern at",
"Pattern Recognition Center, WeChat AI, Tencent Inc, China.",
"HSK. We make the observation that although dis-",
"f Zheng Lin is the corresponding author.",
"tilling HSK is helpful, the marginal utility dimin-"
] |
{
"arxiv_id": "2106.05691",
"language": "en",
"page_idx": 0
}
|
||
2106.05691-en-1
|
2106.05691-en-1.jpg
|
ishes quickly as the amount of HSK increases. To understand this effect, we conduct a series of analysis by compressing the HSK from three dimensions, namely depth, length and width (see Section 2.3 for detailed description). We first compress each single dimension and compare a variety of strategies to extract crucial knowledge. Then, we jointly compress the three dimensions using a set of compression configurations, which specify the amount of HSK assigned to each dimension. Figure 1 shows the results on QNLI dataset. We can find that 1) perceivable performance improvement can be obtained by extracting and distilling the crucial HSK, and 2) with only a tiny fraction of HSK the students can achieve the same performance as extensive HSK distillation.
Based on the second finding, we further propose an efficient paradigm to distill HSK. Concretely, we run BERT over the training set to obtain and store a subset of HSK. This can be done on cloud devices with sufficient computational capability. Given a target device with limited resource, we can compress BERT and select the amount of HSK accordingly. Then, the compressed model can perform KD on either the cloud or directly on the target device using the selected HSK and the original training data, dispensing with the need to load the teacher model.
In summary, our maojor contributions are:
· We observe the marginal utility diminishing effect of HSK in BERT KD. To our knowledge, we are the first attempt to systematically study knowledge compression in BERT KD.
$\cdot$ We conduct exploratory studies on how to extract the crucial knowledge in HSK, based on which we obtain perceivable improvements over a widely-used HSK distillation strategy.
· We propose an efficient KD paradigm based on the empirical findings. Experiments on the GLUE benchmark for NLU (Wang et al., 2019) show that, the proposal gives rise to training speedup of 2.7× $\sim3.4\times$ for TinyBERT and ROSITA on GPU and CPU$^{1}.$
at
$^{1}$The
will
be
code
available
https://github.com/llyx97/Marginal-Utility-Diminishes
## 2 Preliminaries
## 2.1 BERT Architecture
The backbone of BERT consists of an embedding layer and $L$ identical Transformer (Vaswani et al., 2017) layers. The input to the embedding layer is a text sequence x tokenized by WordPiece (Wu et al., 2016). There are two special tokens in x: [CLS] is inserted in the left-most position to aggregate the sequence representation and [SEP] is used to separate text segments. By summing up the token embedding, the position embedding and the segment embedding, the embedding layer outputs a sequence of vectors $\mathbf{E}=\begin{bmatrix}\mathbf{e}_1,\cdots,\mathbf{e}_{|\mathbf{x}|}\end{bmatrix}\in$ $\mathbb{R}^{|\mathbf{x}|\times d_H}$,where $d_H$ is the hidden size of the model.
Then, E passes through the stacked Trans-
former layers, which can be formulated as:
$$\mathbf{H}_l=\mathrm{Trm}_l\left(\mathbf{H}_{l-1}\right),l\in[1,L]$$
(1)
where $\mathbf{H}_l=\begin{bmatrix}\mathbf{h}_{l,1},\cdots,\mathbf{h}_{l,|\mathbf{x}|}\end{bmatrix}\in\mathbb{R}^{|\mathbf{x}|\times d_{H}}$ is the outputs of the $l^{th}$ layer and $\mathbf{H}_0=\mathbf{E}.$ Each Transformer layer is composed of two sub-layers: the multi-head self-attention layer and the feedforward network (FFN).Each sub-layer is followed by a sequence of dropout (Srivastava et al., 2014), residual connection (He et al., 2016) and layer normalization (Ba et al., 2016).
Finally, for the tasks of NLU, a task-specific classifer is employed by taking as input the representation of [CLS] in the $L^{th}$ layer.
## 2.2 BERT Compression with KD
Knowledge distillation is a widely-used technique in model compression, where the compressed model (student) is trained under the guidance of the original model (teacher). This is achieved by minimizing the difference between the features produced by the teacher and the student:
(2)
$$\mathcal{L}_{\mathrm{KD}}=\sum_{(f^S,f^T)}\mathcal{L}\left(f^S(\mathbf{x}),f^T(\mathbf{x})\right)$$
where $\left ( f^S, f^T\right )$is a pair of features from student and teacher respectively. $\mathcal{L}$is the loss function and x is a data sample. In terms of BERT compression, the predicted probability over classes, the intermediate representations and the self-attention distributions can be used as the features to transfer. In this paper, we focus on the intermediate representations $\left\{\mathbf{H}_{l}\right\}_{l=0}^{L}($i.e., the HSK), which have shown to be a useful source of knowledge in BERT compression. The loss function is computed as the
|
[
[
0.12017822265625,
0.07659912109375,
0.49072265625,
0.07659912109375,
0.49072265625,
0.09027099609375,
0.12017822265625,
0.09027099609375
],
[
0.51416015625,
0.07598876953125,
0.6630859375,
0.07598876953125,
0.6630859375,
0.08905029296875,
0.51416015625,
0.08905029296875
],
[
0.12017822265625,
0.0926513671875,
0.489013671875,
0.0926513671875,
0.489013671875,
0.105712890625,
0.12017822265625,
0.105712890625
],
[
0.51416015625,
0.1009521484375,
0.71240234375,
0.1009521484375,
0.71240234375,
0.114013671875,
0.51416015625,
0.114013671875
],
[
0.1168212890625,
0.10809326171875,
0.489013671875,
0.10687255859375,
0.489013671875,
0.1217041015625,
0.1168212890625,
0.122314453125
],
[
0.11846923828125,
0.12469482421875,
0.489013671875,
0.12469482421875,
0.489013671875,
0.1383056640625,
0.11846923828125,
0.1383056640625
],
[
0.51416015625,
0.1217041015625,
0.88330078125,
0.1217041015625,
0.88330078125,
0.1353759765625,
0.51416015625,
0.1353759765625
],
[
0.11846923828125,
0.1395263671875,
0.489013671875,
0.1395263671875,
0.489013671875,
0.152587890625,
0.11846923828125,
0.152587890625
],
[
0.5126953125,
0.135986328125,
0.88818359375,
0.13720703125,
0.88818359375,
0.1519775390625,
0.5126953125,
0.15087890625
],
[
0.12017822265625,
0.15673828125,
0.49072265625,
0.15673828125,
0.49072265625,
0.17041015625,
0.12017822265625,
0.17041015625
],
[
0.51416015625,
0.15380859375,
0.88671875,
0.15380859375,
0.88671875,
0.16748046875,
0.51416015625,
0.16748046875
],
[
0.1168212890625,
0.171630859375,
0.49072265625,
0.1728515625,
0.49072265625,
0.18701171875,
0.1168212890625,
0.1859130859375
],
[
0.51416015625,
0.169189453125,
0.884765625,
0.169189453125,
0.884765625,
0.182861328125,
0.51416015625,
0.182861328125
],
[
0.1168212890625,
0.188232421875,
0.489013671875,
0.188232421875,
0.489013671875,
0.201904296875,
0.1168212890625,
0.201904296875
],
[
0.5126953125,
0.1859130859375,
0.884765625,
0.1859130859375,
0.884765625,
0.1995849609375,
0.5126953125,
0.1995849609375
],
[
0.11846923828125,
0.204833984375,
0.489013671875,
0.204833984375,
0.489013671875,
0.2197265625,
0.11846923828125,
0.2197265625
],
[
0.5126953125,
0.201904296875,
0.88330078125,
0.201904296875,
0.88330078125,
0.215576171875,
0.5126953125,
0.215576171875
],
[
0.11846923828125,
0.221435546875,
0.49072265625,
0.221435546875,
0.49072265625,
0.235107421875,
0.11846923828125,
0.235107421875
],
[
0.5126953125,
0.218505859375,
0.88330078125,
0.21728515625,
0.88330078125,
0.232177734375,
0.5126953125,
0.2327880859375
],
[
0.11846923828125,
0.2369384765625,
0.49072265625,
0.2369384765625,
0.49072265625,
0.25048828125,
0.11846923828125,
0.25048828125
],
[
0.51416015625,
0.235107421875,
0.88330078125,
0.235107421875,
0.88330078125,
0.248779296875,
0.51416015625,
0.248779296875
],
[
0.12017822265625,
0.254150390625,
0.489013671875,
0.254150390625,
0.489013671875,
0.267333984375,
0.12017822265625,
0.267333984375
],
[
0.51416015625,
0.25048828125,
0.88330078125,
0.25048828125,
0.88330078125,
0.26416015625,
0.51416015625,
0.26416015625
],
[
0.12017822265625,
0.26953125,
0.49072265625,
0.26953125,
0.49072265625,
0.283203125,
0.12017822265625,
0.283203125
],
[
0.51416015625,
0.2666015625,
0.88330078125,
0.2666015625,
0.88330078125,
0.27978515625,
0.51416015625,
0.27978515625
],
[
0.12017822265625,
0.28515625,
0.489013671875,
0.28515625,
0.489013671875,
0.298583984375,
0.12017822265625,
0.298583984375
],
[
0.5126953125,
0.28076171875,
0.88330078125,
0.283203125,
0.88330078125,
0.297607421875,
0.5126953125,
0.295654296875
],
[
0.12017822265625,
0.3017578125,
0.489013671875,
0.3017578125,
0.489013671875,
0.3154296875,
0.12017822265625,
0.3154296875
],
[
0.5126953125,
0.29638671875,
0.884765625,
0.298583984375,
0.884765625,
0.31640625,
0.5126953125,
0.314208984375
],
[
0.11846923828125,
0.31640625,
0.3193359375,
0.317626953125,
0.3193359375,
0.33203125,
0.11846923828125,
0.330810546875
],
[
0.53173828125,
0.314208984375,
0.884765625,
0.31640625,
0.884765625,
0.330810546875,
0.53173828125,
0.328857421875
],
[
0.13525390625,
0.3349609375,
0.489013671875,
0.33740234375,
0.489013671875,
0.3515625,
0.13525390625,
0.349853515625
],
[
0.51416015625,
0.333251953125,
0.8310546875,
0.333251953125,
0.8310546875,
0.3466796875,
0.51416015625,
0.3466796875
],
[
0.11846923828125,
0.352783203125,
0.489013671875,
0.352783203125,
0.489013671875,
0.36767578125,
0.11846923828125,
0.36767578125
],
[
0.58740234375,
0.35986328125,
0.8115234375,
0.35986328125,
0.8115234375,
0.37353515625,
0.58740234375,
0.37353515625
],
[
0.85791015625,
0.361083984375,
0.884765625,
0.361083984375,
0.884765625,
0.37353515625,
0.85791015625,
0.37353515625
],
[
0.12017822265625,
0.369384765625,
0.489013671875,
0.369384765625,
0.489013671875,
0.383056640625,
0.12017822265625,
0.383056640625
],
[
0.12017822265625,
0.385498046875,
0.489013671875,
0.385498046875,
0.489013671875,
0.3984375,
0.12017822265625,
0.3984375
],
[
0.51025390625,
0.385498046875,
0.884765625,
0.383056640625,
0.884765625,
0.402099609375,
0.51025390625,
0.40380859375
],
[
0.12017822265625,
0.402099609375,
0.489013671875,
0.402099609375,
0.489013671875,
0.4150390625,
0.12017822265625,
0.4150390625
],
[
0.5126953125,
0.405029296875,
0.88671875,
0.405029296875,
0.88671875,
0.418701171875,
0.5126953125,
0.418701171875
],
[
0.12017822265625,
0.418701171875,
0.49072265625,
0.418701171875,
0.49072265625,
0.431640625,
0.12017822265625,
0.431640625
],
[
0.51416015625,
0.42041015625,
0.88671875,
0.42041015625,
0.88671875,
0.43408203125,
0.51416015625,
0.43408203125
],
[
0.11846923828125,
0.43408203125,
0.4873046875,
0.43408203125,
0.4873046875,
0.44775390625,
0.11846923828125,
0.44775390625
],
[
0.51416015625,
0.43701171875,
0.88671875,
0.43701171875,
0.88671875,
0.45068359375,
0.51416015625,
0.45068359375
],
[
0.12017822265625,
0.45068359375,
0.49072265625,
0.45068359375,
0.49072265625,
0.46435546875,
0.12017822265625,
0.46435546875
],
[
0.5126953125,
0.453125,
0.884765625,
0.453125,
0.884765625,
0.466064453125,
0.5126953125,
0.466064453125
],
[
0.12017822265625,
0.466064453125,
0.4931640625,
0.466064453125,
0.4931640625,
0.479736328125,
0.12017822265625,
0.479736328125
],
[
0.51611328125,
0.47021484375,
0.88232421875,
0.47021484375,
0.88232421875,
0.48095703125,
0.51611328125,
0.48095703125
],
[
0.11846923828125,
0.482177734375,
0.489013671875,
0.48095703125,
0.489013671875,
0.495361328125,
0.11846923828125,
0.496337890625
],
[
0.51416015625,
0.485107421875,
0.88330078125,
0.485107421875,
0.88330078125,
0.498779296875,
0.51416015625,
0.498779296875
],
[
0.11846923828125,
0.498779296875,
0.49072265625,
0.498779296875,
0.49072265625,
0.5126953125,
0.11846923828125,
0.5126953125
],
[
0.51416015625,
0.50048828125,
0.7880859375,
0.50048828125,
0.7880859375,
0.51416015625,
0.51416015625,
0.51416015625
],
[
0.12017822265625,
0.515625,
0.228515625,
0.515625,
0.228515625,
0.52685546875,
0.12017822265625,
0.52685546875
],
[
0.53173828125,
0.51708984375,
0.88232421875,
0.51708984375,
0.88232421875,
0.53076171875,
0.53173828125,
0.53076171875
],
[
0.1361083984375,
0.53369140625,
0.44873046875,
0.53369140625,
0.44873046875,
0.54736328125,
0.1361083984375,
0.54736328125
],
[
0.51416015625,
0.53369140625,
0.88330078125,
0.53369140625,
0.88330078125,
0.54736328125,
0.51416015625,
0.54736328125
],
[
0.5126953125,
0.548828125,
0.7900390625,
0.5498046875,
0.7900390625,
0.56396484375,
0.5126953125,
0.5634765625
],
[
0.1412353515625,
0.57470703125,
0.49072265625,
0.57470703125,
0.49072265625,
0.58837890625,
0.1412353515625,
0.58837890625
],
[
0.51416015625,
0.576171875,
0.78466796875,
0.576171875,
0.78466796875,
0.5888671875,
0.51416015625,
0.5888671875
],
[
0.1546630859375,
0.59033203125,
0.49072265625,
0.59033203125,
0.49072265625,
0.60400390625,
0.1546630859375,
0.60400390625
],
[
0.5126953125,
0.595703125,
0.88330078125,
0.5966796875,
0.88330078125,
0.61083984375,
0.5126953125,
0.60986328125
],
[
0.1571044921875,
0.60693359375,
0.49072265625,
0.60693359375,
0.49072265625,
0.62060546875,
0.1571044921875,
0.62060546875
],
[
0.51416015625,
0.61328125,
0.88330078125,
0.61328125,
0.88330078125,
0.626953125,
0.51416015625,
0.626953125
],
[
0.155517578125,
0.62255859375,
0.49072265625,
0.62255859375,
0.49072265625,
0.63720703125,
0.155517578125,
0.63720703125
],
[
0.51611328125,
0.62890625,
0.884765625,
0.62890625,
0.884765625,
0.642578125,
0.51611328125,
0.642578125
],
[
0.155517578125,
0.6396484375,
0.188232421875,
0.6396484375,
0.188232421875,
0.65185546875,
0.155517578125,
0.65185546875
],
[
0.51416015625,
0.64453125,
0.884765625,
0.6435546875,
0.884765625,
0.65771484375,
0.51416015625,
0.6591796875
],
[
0.51416015625,
0.6611328125,
0.884765625,
0.6611328125,
0.884765625,
0.6748046875,
0.51416015625,
0.6748046875
],
[
0.1419677734375,
0.6796875,
0.4931640625,
0.6796875,
0.4931640625,
0.693359375,
0.1419677734375,
0.693359375
],
[
0.5126953125,
0.6767578125,
0.8134765625,
0.67578125,
0.8134765625,
0.6904296875,
0.5126953125,
0.69140625
],
[
0.1529541015625,
0.6953125,
0.49072265625,
0.6953125,
0.49072265625,
0.708984375,
0.1529541015625,
0.708984375
],
[
0.57568359375,
0.7060546875,
0.63525390625,
0.7060546875,
0.63525390625,
0.72021484375,
0.57568359375,
0.72021484375
],
[
0.64111328125,
0.70166015625,
0.82421875,
0.70361328125,
0.82421875,
0.72265625,
0.64111328125,
0.72021484375
],
[
0.1546630859375,
0.7119140625,
0.489013671875,
0.7119140625,
0.489013671875,
0.7255859375,
0.1546630859375,
0.7255859375
],
[
0.85986328125,
0.70703125,
0.88330078125,
0.70703125,
0.88330078125,
0.7197265625,
0.85986328125,
0.7197265625
],
[
0.1529541015625,
0.724609375,
0.489013671875,
0.72705078125,
0.489013671875,
0.74462890625,
0.1529541015625,
0.7421875
],
[
0.6337890625,
0.7236328125,
0.6875,
0.7236328125,
0.6875,
0.73828125,
0.6337890625,
0.73828125
],
[
0.51416015625,
0.75,
0.88330078125,
0.75,
0.88330078125,
0.76318359375,
0.51416015625,
0.76318359375
],
[
0.1412353515625,
0.76708984375,
0.4873046875,
0.76708984375,
0.4873046875,
0.78076171875,
0.1412353515625,
0.78076171875
],
[
0.51416015625,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51416015625,
0.77978515625
],
[
0.155517578125,
0.78369140625,
0.489013671875,
0.78369140625,
0.489013671875,
0.79736328125,
0.155517578125,
0.79736328125
],
[
0.51416015625,
0.7822265625,
0.884765625,
0.7822265625,
0.884765625,
0.79638671875,
0.51416015625,
0.79638671875
],
[
0.155517578125,
0.80029296875,
0.492431640625,
0.80029296875,
0.492431640625,
0.81396484375,
0.155517578125,
0.81396484375
],
[
0.51611328125,
0.80029296875,
0.88330078125,
0.80029296875,
0.88330078125,
0.81103515625,
0.51611328125,
0.81103515625
],
[
0.1546630859375,
0.8154296875,
0.49072265625,
0.81640625,
0.49072265625,
0.83056640625,
0.1546630859375,
0.830078125
],
[
0.5126953125,
0.81396484375,
0.88671875,
0.8134765625,
0.88671875,
0.82763671875,
0.5126953125,
0.8291015625
],
[
0.1529541015625,
0.8330078125,
0.492431640625,
0.83203125,
0.492431640625,
0.8466796875,
0.1529541015625,
0.84716796875
],
[
0.5126953125,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.5126953125,
0.84423828125
],
[
0.155517578125,
0.8486328125,
0.44873046875,
0.8486328125,
0.44873046875,
0.8623046875,
0.155517578125,
0.8623046875
],
[
0.51416015625,
0.84716796875,
0.88671875,
0.84716796875,
0.88671875,
0.86083984375,
0.51416015625,
0.86083984375
],
[
0.51416015625,
0.86328125,
0.88671875,
0.86328125,
0.88671875,
0.87744140625,
0.51416015625,
0.87744140625
],
[
0.1412353515625,
0.88427734375,
0.1739501953125,
0.88427734375,
0.1739501953125,
0.8955078125,
0.1412353515625,
0.8955078125
],
[
0.205078125,
0.88427734375,
0.239501953125,
0.88427734375,
0.239501953125,
0.896484375,
0.205078125,
0.896484375
],
[
0.268798828125,
0.88427734375,
0.3017578125,
0.88427734375,
0.3017578125,
0.8955078125,
0.268798828125,
0.8955078125
],
[
0.37890625,
0.88427734375,
0.441162109375,
0.8818359375,
0.44287109375,
0.896484375,
0.380615234375,
0.8984375
],
[
0.51416015625,
0.87890625,
0.88671875,
0.87890625,
0.88671875,
0.892578125,
0.51416015625,
0.892578125
],
[
0.330322265625,
0.88525390625,
0.35205078125,
0.88525390625,
0.35205078125,
0.8955078125,
0.330322265625,
0.8955078125
],
[
0.11846923828125,
0.8955078125,
0.450439453125,
0.8955078125,
0.450439453125,
0.9091796875,
0.11846923828125,
0.9091796875
],
[
0.477294921875,
0.89013671875,
0.4873046875,
0.89013671875,
0.4873046875,
0.89453125,
0.477294921875,
0.89453125
],
[
0.5126953125,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5126953125,
0.9091796875
]
] |
[
"ishes quickly as the amount of HSK increases. To",
"2Preliminaries",
"understand this effect, we conduct a series of anal-",
"2.1BERTArchitecture",
"ysis by compressing the HSK from three dimen-",
"sions, namely depth, length and width (see Section",
"The backbone of BERT consists of an embedding",
"2.3 for detailed description). We first compress",
"layer and L identical Transformer (Vaswani et al.,",
"each single dimension and compare a variety of",
"2017) layers. The input to the embedding layer",
"strategies to extract crucial knowledge. Then, we",
"is a text sequence x tokenized by WordPiece",
"jointly compress the three dimensions using a set",
"(Wu et al., 2016). There are two special tokens in",
"of compression configurations, which specify the",
"x: [CLS] is inserted in the left-most position to ag.",
"amount of HSK assigned to each dimension. Fig-",
"gregate the sequence representation and [SEP] is",
"ure 1 shows the results on QNLI dataset. We can",
"used to separate text segments. By summing up",
"find that 1)perceivable performance improvement",
"the token embedding, the position embedding and",
"can be obtained by extracting and distilling the cru-",
"the segment embedding, the embedding layer out-",
"cial HSK, and 2) with only a tiny fraction of HSK",
"puts a sequence of vectors E =[ei,...,exE",
"the students can achieve the same performance as",
"R/xxdh, where d is the hidden size of the model.",
"extensive HSK distillation.",
"Then, E passes through the stacked Trans-",
"Based on the second finding, we further propose",
"former layers, which can be formulated as:",
"an effcient paradigm to distill HSK. Concretely,",
"Hi = Trmi (Hi-1),l E [1, L]",
"(1)",
"we run BERT over the training set to obtain and",
"store a subset of HSK. This can be done on cloud",
"where H, = [h,1, . ,hi,]x] e Rxxd is the",
"devices with sufficient computational capability.",
"outputs of the ith layer and Ho = E. Each",
"Given a target device with limited resource, we",
"Transformer layer is composed of two sub-layers:",
"can compress BERT and select the amount of HSK",
"the multi-head self-attention layer and the feed-",
"accordingly. Then, the compressed model can per-",
"forward network (FFN).Each sub-layer is fol-",
"form KD on either the cloud or directly on the tar-",
"lowed by a seguence of dropout (Srivastava et al.",
"get device using the selected HSK and the original",
"2014), residual connection (He et al., 2016) and",
"training data, dispensing with the need to load the",
"layer normalization (Ba et al., 2016).",
"teacher model.",
"Finally, for the tasks of NLU, a task-specific",
"In summary, our maojor contributions are:",
"classifier is employed by taking as input the rep",
"resentation of [CLS] in the Lth layer.",
". We observe the marginal utility diminishing",
"2.2 BERT Compression with KD",
"effect of HSK in BERT KD. To our knowl-",
"Knowledge distillation is a widely-used technique",
"edge, we are the first attempt to systemati-",
"in model compression, where the compressed",
"cally study knowledge compression in BERT",
"model (student) is trained under the guidance of",
"KD.",
"the original model (teacher). This is achieved",
"by minimizing the difference between the features",
". We conduct exploratory studies on how to ex-",
"produced by the teacher and the student:",
" tract the crucial knowledge in HSK, based on",
"CKD =",
" (fS(x),fT(x))",
"which we obtain perceivable improvements",
"(2)",
"over a widely-used HSK distillation strategy.",
"(fS,fT)",
"where (fS,fT) is a pair of features from student",
". We propose an efficient KD paradigm based",
"and teacher respectively. L is the loss function and",
"on the empirical findings.Experiments on",
"x is a data sample. In terms of BERT compression,",
"the GLUE benchmark for NLU (Wang et al.,",
"thepredictedprobability overclasses,theinterme-",
"2019) show that, the proposal gives rise to",
"diate representations and the self-attention distri-",
"training speedup of 2.7× ~3.4x for Tiny-",
"butions can be used as the features to transfer. In",
"BERT and ROSITA on GPU and CPUl",
"this paper, we focus on the intermediate represen-",
"tations [Hi)i-o (i.e., the HSK), which have shown",
"IThe",
"code",
"will",
"available",
"to be a useful source of knowledge in BERT com-",
"be",
"https://github.com/llyx97/Marginal-Utility-Diminishes",
"B",
"pression.The loss function is computed as the"
] |
{
"arxiv_id": "2106.05691",
"language": "en",
"page_idx": 1
}
|
||
2106.05691-en-2
|
2106.05691-en-2.jpg
|
Mean Squared Error (MSE) in a layer-wise way:
$$\mathcal{L}_{HSK}=\sum_{l=0}^{L^{'}}\mathrm{MSE}\left(\mathbf{H}_{l}^{S}\mathbf{W},\mathbf{H}_{g(l)}^{T}\right)$$
(3)
where $L^\prime$ is the student's layer number and $g(l)$ is the layer mapping function to select teacher layers. $\tilde{\mathbf{W}}\in\mathbb{R}^{d_{H}^{S}\times d_{H}^{T}}$ is the linear transformation to project the student's representations $\mathbf{H}_l^S$ to the same size as the teacher's representation $\ddot{\mathbf{H}}_l^T.$
# 2.3 HSK Compression
According to Equation 3, the HSK from teacher can be stacked into a tensor $\widehat{\mathbf{H}}^T=$ $\left [ \mathbf{H} _{g( 0) }^{T}, \cdots , \mathbf{H} _{g( L^{\prime }) }^{T}\right ]$ $\in$ $\mathbb{R} ^{( L^{\prime }+ 1) \times | \mathbf{x} | \times d_{H}^{T}, \:\mathrm{which}}$ consists of three structural dimensions, namely depth, length and width. For the depth dimension, ers. By dropping the representations corresponding to particular tokens, we compress the length dimension. When it comes to the width dimension, we set the eliminated activations to zero. We will discuss the strategies to compress each dimension later in Section 4.
## 3 Experimental Setups
## 3.1 Datasets
We perform experiments on seven tasks from the General Language Understanding Evaluation (GLUE) benchmark (Wang et al., 2019): CoLA (linguistic acceptability), SST-2 (sentiment analysis), RTE, QNLI, MNLI-m and MNLI-mm (natural language inference), MRPC and STS-B (semantic matching/similarity). Due to space limitation, we only report results on CoLA, SST-2, QNLI and MNLI for single-dimension HSK compression in Section 4, and results on the other three tasks are presented in Appendix E.
## 3.2 Evaluation
Following (Devlin et al., 2019), for the dev set, we use Matthew's correlation and Spearman correlation to evaluate the performance on CoLA and STS-B respectively. For the other tasks, we report the classification accuracy. We use the dev set to conduct our exploratory studies and the test set results are reported to compare HSK compression with the existing distillation strategy. For the test set of MRPC, we report the results of F1 score.
### 3.3 Implementation Details
We take two representative KD-based methods, i.e., TinyBERT (Jiao et al., 2020) and ROSITA (Liu et al., 2021), as examples to conduct our analysis. TinyBERT is a compact version of BERT that is randomly initialized. It is trained with two-stage KD: first on the unlabeled general domain data and then on the task-specific training data. ROSITA replaces the first stage KD with structured pruning and matrix factorization, which can be seen as a direct transfer of BERT's knowledge from the model parameters.
We focus on KD with the task-specific training data and do not use any data augmentation. For TinyBERT, the student model is initialized with the 4-layer general distillation model provided by Jiao et al. (2020) (denoted as TinyBERT$_{4}) .$ For ROSITA, we first fine-tune BERTBASE on the
$\widehat{\mathbf{H}}^T$ can be compressed by eliminating entire lay- downstream task and then compress it follow-
ing Liu et al. (2021) to obtain a 6-layer student model (denoted as ROSITA$_6) .$ The fine-tuned BERT$_\mathrm{BASE}$ is used as the shared teacher for TinyBERT and ROSITA. Following Jiao et al. (2020), we first conduct HSK distillation as in Equation 3 (w/o distilling the self-attention distribution) and then distill the teacher's predictions using crossentropy loss. All the results are averaged over three runs with different random seeds. The model architecture of the students and the hyperparameter settings can be seen in Appendix A and Appendix B respectively.
### 4 Single-Dimension Knowledge $\bar{\mathbf{Compression}}$
Researches on model pruning have shown that the structural units in a model are of different levels of importance, and the unimportant ones can be dropped without affecting the performance. In this section, we investigate whether the same law holds for HSK compression in KD. We study the three dimensions separately and compare a variety of strategies to extract the crucial knowledge. When a certain dimension is compressed, the other two dimensions are kept to full scale.
4.1 Depth Compression
# 4.1.1 Compression Strategies
From the layer point of view, HSK compression can be divided into two steps. First, the layer mapping function $g(l)$ selects one of the
|
[
[
0.12017822265625,
0.07659912109375,
0.49658203125,
0.07659912109375,
0.49658203125,
0.09027099609375,
0.12017822265625,
0.09027099609375
],
[
0.51171875,
0.07598876953125,
0.7421875,
0.0748291015625,
0.7421875,
0.08905029296875,
0.51171875,
0.09027099609375
],
[
0.51416015625,
0.09857177734375,
0.88330078125,
0.09857177734375,
0.88330078125,
0.1134033203125,
0.51416015625,
0.1134033203125
],
[
0.2462158203125,
0.1134033203125,
0.259765625,
0.10809326171875,
0.267333984375,
0.1175537109375,
0.253662109375,
0.122314453125
],
[
0.51416015625,
0.11517333984375,
0.88330078125,
0.11517333984375,
0.88330078125,
0.12890625,
0.51416015625,
0.12890625
],
[
0.1722412109375,
0.12353515625,
0.245361328125,
0.127685546875,
0.24365234375,
0.1436767578125,
0.1705322265625,
0.1395263671875
],
[
0.262939453125,
0.1258544921875,
0.311767578125,
0.1258544921875,
0.311767578125,
0.13720703125,
0.262939453125,
0.13720703125
],
[
0.33447265625,
0.12353515625,
0.407470703125,
0.12353515625,
0.407470703125,
0.1383056640625,
0.33447265625,
0.1383056640625
],
[
0.46728515625,
0.127685546875,
0.486572265625,
0.127685546875,
0.486572265625,
0.13720703125,
0.46728515625,
0.13720703125
],
[
0.404296875,
0.135986328125,
0.417724609375,
0.135986328125,
0.417724609375,
0.1395263671875,
0.404296875,
0.1395263671875
],
[
0.5126953125,
0.1318359375,
0.884765625,
0.1318359375,
0.884765625,
0.1455078125,
0.5126953125,
0.1455078125
],
[
0.245361328125,
0.142578125,
0.272216796875,
0.142578125,
0.272216796875,
0.155029296875,
0.245361328125,
0.155029296875
],
[
0.5126953125,
0.1478271484375,
0.884765625,
0.146728515625,
0.884765625,
0.160888671875,
0.5126953125,
0.162109375
],
[
0.1168212890625,
0.166259765625,
0.4873046875,
0.166259765625,
0.4873046875,
0.179931640625,
0.1168212890625,
0.179931640625
],
[
0.51416015625,
0.1644287109375,
0.88330078125,
0.1644287109375,
0.88330078125,
0.1776123046875,
0.51416015625,
0.1776123046875
],
[
0.12017822265625,
0.18408203125,
0.49072265625,
0.18408203125,
0.49072265625,
0.19775390625,
0.12017822265625,
0.19775390625
],
[
0.51611328125,
0.179931640625,
0.884765625,
0.179931640625,
0.884765625,
0.193603515625,
0.51611328125,
0.193603515625
],
[
0.12017822265625,
0.198974609375,
0.489013671875,
0.198974609375,
0.489013671875,
0.213134765625,
0.12017822265625,
0.213134765625
],
[
0.51416015625,
0.196533203125,
0.884765625,
0.196533203125,
0.884765625,
0.210205078125,
0.51416015625,
0.210205078125
],
[
0.1168212890625,
0.215576171875,
0.489013671875,
0.21435546875,
0.489013671875,
0.2286376953125,
0.1168212890625,
0.2298583984375
],
[
0.5126953125,
0.21142578125,
0.88330078125,
0.2120361328125,
0.88330078125,
0.226806640625,
0.5126953125,
0.2257080078125
],
[
0.11846923828125,
0.232177734375,
0.453857421875,
0.232177734375,
0.453857421875,
0.2464599609375,
0.11846923828125,
0.2464599609375
],
[
0.51416015625,
0.22802734375,
0.88232421875,
0.22802734375,
0.88232421875,
0.2410888671875,
0.51416015625,
0.2410888671875
],
[
0.51416015625,
0.243408203125,
0.884765625,
0.243408203125,
0.884765625,
0.257080078125,
0.51416015625,
0.257080078125
],
[
0.12017822265625,
0.26318359375,
0.309326171875,
0.26318359375,
0.309326171875,
0.276611328125,
0.12017822265625,
0.276611328125
],
[
0.51611328125,
0.262451171875,
0.60107421875,
0.262451171875,
0.60107421875,
0.273681640625,
0.51611328125,
0.273681640625
],
[
0.53173828125,
0.275634765625,
0.88330078125,
0.276611328125,
0.88330078125,
0.29150390625,
0.53173828125,
0.290283203125
],
[
0.12017822265625,
0.2861328125,
0.489013671875,
0.2861328125,
0.489013671875,
0.300537109375,
0.12017822265625,
0.300537109375
],
[
0.51416015625,
0.29345703125,
0.884765625,
0.29345703125,
0.884765625,
0.30712890625,
0.51416015625,
0.30712890625
],
[
0.11846923828125,
0.3017578125,
0.489013671875,
0.300537109375,
0.489013671875,
0.3154296875,
0.11846923828125,
0.31640625
],
[
0.1168212890625,
0.312255859375,
0.49658203125,
0.31005859375,
0.49658203125,
0.341552734375,
0.1168212890625,
0.34326171875
],
[
0.51416015625,
0.308837890625,
0.88330078125,
0.308837890625,
0.88330078125,
0.322509765625,
0.51416015625,
0.322509765625
],
[
0.5126953125,
0.32470703125,
0.88232421875,
0.32470703125,
0.88232421875,
0.33837890625,
0.5126953125,
0.33837890625
],
[
0.11846923828125,
0.33740234375,
0.489013671875,
0.33740234375,
0.489013671875,
0.3515625,
0.11846923828125,
0.3515625
],
[
0.51171875,
0.34033203125,
0.884765625,
0.341552734375,
0.884765625,
0.355712890625,
0.51171875,
0.355224609375
],
[
0.11846923828125,
0.352783203125,
0.4931640625,
0.35400390625,
0.4931640625,
0.3681640625,
0.11846923828125,
0.36767578125
],
[
0.51416015625,
0.358154296875,
0.884765625,
0.358154296875,
0.884765625,
0.372314453125,
0.51416015625,
0.372314453125
],
[
0.1168212890625,
0.366455078125,
0.492431640625,
0.3681640625,
0.492431640625,
0.385986328125,
0.1168212890625,
0.38427734375
],
[
0.51611328125,
0.3759765625,
0.884765625,
0.3759765625,
0.884765625,
0.385986328125,
0.51611328125,
0.385986328125
],
[
0.12017822265625,
0.385986328125,
0.489013671875,
0.385986328125,
0.489013671875,
0.399658203125,
0.12017822265625,
0.399658203125
],
[
0.5126953125,
0.3896484375,
0.88330078125,
0.3896484375,
0.88330078125,
0.402587890625,
0.5126953125,
0.402587890625
],
[
0.121826171875,
0.40380859375,
0.486572265625,
0.40380859375,
0.486572265625,
0.41455078125,
0.121826171875,
0.41455078125
],
[
0.51416015625,
0.40625,
0.884765625,
0.40625,
0.884765625,
0.42041015625,
0.51416015625,
0.42041015625
],
[
0.121826171875,
0.419189453125,
0.4873046875,
0.419189453125,
0.4873046875,
0.429931640625,
0.121826171875,
0.429931640625
],
[
0.5126953125,
0.42041015625,
0.884765625,
0.42041015625,
0.884765625,
0.438232421875,
0.5126953125,
0.438232421875
],
[
0.12017822265625,
0.43408203125,
0.489013671875,
0.43408203125,
0.489013671875,
0.44775390625,
0.12017822265625,
0.44775390625
],
[
0.51416015625,
0.438232421875,
0.88671875,
0.438232421875,
0.88671875,
0.451904296875,
0.51416015625,
0.451904296875
],
[
0.12017822265625,
0.449462890625,
0.489013671875,
0.449462890625,
0.489013671875,
0.46435546875,
0.12017822265625,
0.46435546875
],
[
0.51416015625,
0.454833984375,
0.884765625,
0.454833984375,
0.884765625,
0.468505859375,
0.51416015625,
0.468505859375
],
[
0.11846923828125,
0.466064453125,
0.2529296875,
0.466064453125,
0.2529296875,
0.479736328125,
0.11846923828125,
0.479736328125
],
[
0.5126953125,
0.47021484375,
0.88330078125,
0.47021484375,
0.88330078125,
0.48388671875,
0.5126953125,
0.48388671875
],
[
0.51416015625,
0.48681640625,
0.88671875,
0.48681640625,
0.88671875,
0.50048828125,
0.51416015625,
0.50048828125
],
[
0.11846923828125,
0.498779296875,
0.1512451171875,
0.498779296875,
0.1512451171875,
0.51123046875,
0.11846923828125,
0.51123046875
],
[
0.1436767578125,
0.498779296875,
0.3310546875,
0.498779296875,
0.3310546875,
0.51318359375,
0.1436767578125,
0.51318359375
],
[
0.5126953125,
0.5029296875,
0.884765625,
0.501953125,
0.884765625,
0.5166015625,
0.5126953125,
0.51708984375
],
[
0.5126953125,
0.5185546875,
0.88232421875,
0.5185546875,
0.88232421875,
0.5322265625,
0.5126953125,
0.5322265625
],
[
0.11846923828125,
0.52685546875,
0.1546630859375,
0.52685546875,
0.1546630859375,
0.5390625,
0.11846923828125,
0.5390625
],
[
0.1495361328125,
0.52783203125,
0.228515625,
0.52783203125,
0.228515625,
0.5390625,
0.1495361328125,
0.5390625
],
[
0.51611328125,
0.53515625,
0.88330078125,
0.53515625,
0.88330078125,
0.548828125,
0.51611328125,
0.548828125
],
[
0.11846923828125,
0.55029296875,
0.49072265625,
0.55029296875,
0.49072265625,
0.5654296875,
0.11846923828125,
0.5654296875
],
[
0.51416015625,
0.5517578125,
0.88671875,
0.5517578125,
0.88671875,
0.5654296875,
0.51416015625,
0.5654296875
],
[
0.11846923828125,
0.56640625,
0.4873046875,
0.56640625,
0.4873046875,
0.580078125,
0.11846923828125,
0.580078125
],
[
0.51416015625,
0.56640625,
0.67822265625,
0.56640625,
0.67822265625,
0.580078125,
0.51416015625,
0.580078125
],
[
0.11846923828125,
0.5830078125,
0.489013671875,
0.5830078125,
0.489013671875,
0.5966796875,
0.11846923828125,
0.5966796875
],
[
0.11846923828125,
0.599609375,
0.49072265625,
0.599609375,
0.49072265625,
0.61328125,
0.11846923828125,
0.61328125
],
[
0.53955078125,
0.595703125,
0.7998046875,
0.5966796875,
0.7998046875,
0.6142578125,
0.53955078125,
0.61328125
],
[
0.517578125,
0.6025390625,
0.53125,
0.6025390625,
0.53125,
0.609375,
0.517578125,
0.609375
],
[
0.12017822265625,
0.615234375,
0.49072265625,
0.615234375,
0.49072265625,
0.62890625,
0.12017822265625,
0.62890625
],
[
0.54541015625,
0.615234375,
0.662109375,
0.615234375,
0.662109375,
0.6298828125,
0.54541015625,
0.6298828125
],
[
0.11846923828125,
0.63134765625,
0.489013671875,
0.63134765625,
0.489013671875,
0.64453125,
0.11846923828125,
0.64453125
],
[
0.11846923828125,
0.64794921875,
0.492431640625,
0.64794921875,
0.492431640625,
0.6611328125,
0.11846923828125,
0.6611328125
],
[
0.51416015625,
0.6435546875,
0.88330078125,
0.6435546875,
0.88330078125,
0.65673828125,
0.51416015625,
0.65673828125
],
[
0.1168212890625,
0.6630859375,
0.4931640625,
0.662109375,
0.4931640625,
0.6767578125,
0.1168212890625,
0.677734375
],
[
0.51416015625,
0.66015625,
0.88330078125,
0.66015625,
0.88330078125,
0.67333984375,
0.51416015625,
0.67333984375
],
[
0.12017822265625,
0.6787109375,
0.492431640625,
0.6787109375,
0.492431640625,
0.6923828125,
0.12017822265625,
0.6923828125
],
[
0.5126953125,
0.6748046875,
0.88330078125,
0.6748046875,
0.88330078125,
0.6884765625,
0.5126953125,
0.6884765625
],
[
0.11846923828125,
0.6953125,
0.49072265625,
0.6953125,
0.49072265625,
0.708984375,
0.11846923828125,
0.708984375
],
[
0.51416015625,
0.6923828125,
0.884765625,
0.6923828125,
0.884765625,
0.7060546875,
0.51416015625,
0.7060546875
],
[
0.1168212890625,
0.71142578125,
0.374755859375,
0.71044921875,
0.374755859375,
0.724609375,
0.1168212890625,
0.7255859375
],
[
0.51416015625,
0.70703125,
0.88232421875,
0.70703125,
0.88232421875,
0.72021484375,
0.51416015625,
0.72021484375
],
[
0.51416015625,
0.7236328125,
0.884765625,
0.7236328125,
0.884765625,
0.73828125,
0.51416015625,
0.73828125
],
[
0.1168212890625,
0.74072265625,
0.2496337890625,
0.7421875,
0.2496337890625,
0.75732421875,
0.1168212890625,
0.7548828125
],
[
0.51416015625,
0.74072265625,
0.884765625,
0.74072265625,
0.884765625,
0.75439453125,
0.51416015625,
0.75439453125
],
[
0.5126953125,
0.755859375,
0.884765625,
0.7548828125,
0.884765625,
0.76953125,
0.5126953125,
0.77099609375
],
[
0.12017822265625,
0.765625,
0.49072265625,
0.765625,
0.49072265625,
0.779296875,
0.12017822265625,
0.779296875
],
[
0.51416015625,
0.7724609375,
0.884765625,
0.7724609375,
0.884765625,
0.7861328125,
0.51416015625,
0.7861328125
],
[
0.12017822265625,
0.783203125,
0.49072265625,
0.783203125,
0.49072265625,
0.79638671875,
0.12017822265625,
0.79638671875
],
[
0.51416015625,
0.7880859375,
0.7587890625,
0.7880859375,
0.7587890625,
0.8017578125,
0.51416015625,
0.8017578125
],
[
0.11846923828125,
0.798828125,
0.489013671875,
0.79736328125,
0.489013671875,
0.8125,
0.11846923828125,
0.8134765625
],
[
0.11846923828125,
0.8134765625,
0.49072265625,
0.81396484375,
0.49072265625,
0.82763671875,
0.11846923828125,
0.82666015625
],
[
0.5126953125,
0.818359375,
0.7109375,
0.818359375,
0.7109375,
0.83203125,
0.5126953125,
0.83203125
],
[
0.11846923828125,
0.83056640625,
0.49072265625,
0.83056640625,
0.49072265625,
0.84423828125,
0.11846923828125,
0.84423828125
],
[
0.51416015625,
0.84130859375,
0.75732421875,
0.84130859375,
0.75732421875,
0.85498046875,
0.51416015625,
0.85498046875
],
[
0.11846923828125,
0.84716796875,
0.4931640625,
0.84716796875,
0.4931640625,
0.86083984375,
0.11846923828125,
0.86083984375
],
[
0.11846923828125,
0.86328125,
0.49072265625,
0.86328125,
0.49072265625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.5126953125,
0.8623046875,
0.884765625,
0.8623046875,
0.884765625,
0.876953125,
0.5126953125,
0.876953125
],
[
0.1168212890625,
0.87744140625,
0.49072265625,
0.87890625,
0.49072265625,
0.8935546875,
0.1168212890625,
0.892578125
],
[
0.51611328125,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.51611328125,
0.8935546875
],
[
0.11846923828125,
0.8935546875,
0.468017578125,
0.89453125,
0.468017578125,
0.9091796875,
0.11846923828125,
0.90771484375
],
[
0.51416015625,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.91015625,
0.51416015625,
0.91015625
]
] |
[
"Mean Squared Error (MSE) in a layer-wise way:",
"3.3Implementation Details",
"We take two representative KD-based methods.",
"L",
"i.e., TinyBERT (Jiao et al., 2020) and ROSITA",
"CHSK=",
"MSE",
"ISW.H",
"(3)",
"",
"(Liu et al., 2021), as examples to conduct our anal-",
"l=0",
"ysis. TinyBERT is a compact version of BERT that",
"where L' is the student's layer number and g(1) is",
"is randomly initialized. It is trained with two-stage",
"the layer mapping function to select teacher lay-",
"KD: first on the unlabeled general domain data and",
"ers. W RdxdH is the linear transformation",
"then on the task-specific training data. ROSITA",
"to project the student's representations HS to the",
"replaces the first stage KD with structured pruning",
"same size as the teacher's representation HT.",
"and matrix factorization, which can be seen as a di",
"rect transfer of BERT's knowledge from the model",
"2.3HSK Compression",
"parameters.",
"We focus on KD with the task-specific training",
"Accordingto Equation 3,the HSKfrom",
"data and do not use any data augmentation. For",
"teacher can be stacked into. a tensor HT =",
"[Ho)H)] e R(+xd, which",
"TinyBERT, the student model is initialized with",
"the 4-layer general distillation model provided by",
"consists of three structural dimensions, namely",
"Jiao et al. (2020) (denoted as TinyBERT).For",
"depth, length and width. For the depth dimension,",
"ROSITA, we first fine-tune BERTBASE on the",
"HT can be compressed by eliminating entire lay-",
"downstream task and then compress it follow-",
"ers. By dropping the representations correspond",
"ing Liu et al. (2021) to obtain a 6-layer student",
"ing toparticular tokens,wecompressthe length",
"model (denoted as ROSITAg). The fine-tuned",
"dimension.Whenitcomestothewidthdimension.",
"BERTBASE is used as the shared teacher for Tiny-",
"we set the eliminated activations to zero. We will",
"BERT and ROSITA. Following Jiao et al. (2020),",
"discuss the strategies to compress each dimension",
"we first conduct HSK distillation as in Equation 3",
"later in Section 4.",
"(w/o distilling the self-attention distribution) and",
"then distill the teacher's predictions using cross-",
"3",
" Experimental Setups",
"entropy loss.All theresults are averaged over",
"three runs with different random seeds. The model",
"3.1",
"Datasets",
"architecture of the students and the hyperparame-",
"We perform experiments on seven tasks from",
"ter settings can be seen in Appendix A and Ap-",
"the General Language Understanding Evaluation",
"pendix B respectively.",
"(GLUE) benchmark (Wang et al., 2019): CoLA",
"(linguistic acceptability), SST-2 (sentiment analy-",
"Single-Dimension Knowledge",
"4",
"sis), RTE, QNLI, MNLI-m and MNLI-mm (nat-",
"Compression",
"ural language inference), MRPC and STS-B (se-",
"mantic matching/similarity).Due to space limi-",
"Researches on model pruning have shown that the",
"tation, we only report results on CoLA, SST-2,",
"structural units in a model are of different levels",
"QNLI and MNLI for single-dimension HSK com-",
"of importance, and the unimportant ones can be",
"pression in Section 4, and results on the other three",
"dropped without affecting the performance. In this",
"tasks are presented in Appendix E.",
"section, we investigate whether the same law holds",
"for HSK compression in KD. We study the three",
"3.2Evaluation",
"dimensions separately and compare a variety of",
" strategies to extract the crucial knowledge. When",
"Following (Devlin et al., 2019), for the dev set, we",
"a certain dimension is compressed, the other two",
"use Matthew's correlation and Spearman correla-",
"dimensions are kept to full scale.",
"tion to evaluate the performance on CoLA and",
"STS-B respectively.For the other tasks, we report",
"4.1 Depth Compression",
"the classification accuracy. We use the dev set to",
"4.1.1 Compression Strategies",
"conduct our exploratory studies and the test set re-",
"sults are reported to compare HSK compression",
"From the layer point of view, HSK compres-",
"with the existing distillation strategy. For the test",
"sion can be divided into two steps. First, the",
"set of MRPC, we report the results of F1 score.",
"layer mapping function g() selects one of the"
] |
{
"arxiv_id": "2106.05691",
"language": "en",
"page_idx": 2
}
|
||
2106.05691-en-3
|
2106.05691-en-3.jpg
|
<FigureHere>
Figure 2: Illustration of the redesigned uniform layer mapping strategy. Left: $L^{top}$ is the top teacher layer. Right: $L^{top}$ is the second-top teacher layer.
teacher layers for each student layer. This produces $L^{\prime}+1$ pairs of teacher-student features: $\left [ ( \mathbf{H} _{0}^{S}, \mathbf{H} _{g( 0) }^{T}) , \cdots , ( \mathbf{H} _{L^{\prime }}^{S}, \mathbf{H} _{g( L^{\prime }) }^{T}) \right ] .$ Second, a subset of these feature pairs are selected to perform HSK distillation.
For the first step, a simple but effective strategy
is the uniform mapping function:
$$g(l)=l\times\frac{L}{L^{'}},\mathrm{mod}(L,L^{'})=0$$
(4)
In this way, the teacher layers are divided into $L^{\prime}$ blocks and the top layer of each block serves as the guidance in KD. Recently, Wang et al. (2020a) empirically show that the upper-middle layers of BERT, as compared with the top layer, are a better choice to guide the top layer of student in selfattention distillation. Inspired by this, we redesign Equation 4 to allow the top student layer to distill knowledge from an upper-middle teacher layer, and the lower layers follow the uniform mapping principle. This function can be formulated as:
$$g(l,L^{top})=l\times\mathrm{round}(\frac{L^{top}}{L^{'}})$$
(5)
where $L^{top}$ is the teacher layer corresponding to the top student layer and round() is the roundingoff operation. Figure 2 gives an illustration of $g(l,\bar{L^{top}})$ with a 6-layer teacher and a 3-layer student. Specifically, for the 12-layer BERT$_\mathrm{BASE}$ teacher, we select $L^{top}$ from $\{8,10,12\}.$ For the second step, we simply keep the top $N^D$ feature ${\mathrm{pairs:~}}\left\{(\mathbf{H}_{l}^{S},\mathbf{H}_{g(l,L^{top})}^{T})\right\}_{l=L^{\prime}-N^{D}+1}^{L^{\prime}}.$
### 4.1.2 Results and Analysis
Figure 3 presents the results of depth compression with different layer mapping functions. We can find that: 1) For the $g(l,12)$ mapping function (the
<FigureHere>
Figure 3: Results of depth compression on CoLA, SST- 2,QNLI and MNLI. Each color denotes a layer mapping function. The number of layers in HSK includes the embedding layer. Full results on seven tasks are shown in Appendix E.1.
grey lines), depth compression generally has a negative impact on the students' performance. Specially, the performance of ROSITA$_6$ declines drastically when the number of layers is reduced to $1\sim3.2)$In terms of the $g(l,10)$ and $g(l,8)$ mapping functions (the blue and orange lines), HSK distillation with only one or two layers can achieve comparable performance as using all the $L^{\prime}+1$ layers. On the QNLI and MNLI datasets, the performance can even be improved by eliminating the lower layers. 3) In general, the student achieves better results with the redesigned layer mapping function in Equation 5 across the four tasks. This demonstrates that, like the self-attention knowledge, the most crucial HSK does not necessarily reside in the top BERT layer, which reveals a potential way to improve HSK distillation of BERT. 4) Compared with $g( l, 8) $,the improvement brought by $g(l,10)$ is more stable across different tasks and student models. Therefore, we use the $g(l,10)$ layer mapping function when investigating the other two dimensions.
4.2 Length Compression
## 4.2.1 Compression Strategies
To compress the length dimension, we design a method to measure the tokens'importance by using the teacher's self-attention distribution. The intuition is that self-attention controls the information flow among tokens across layers, and thus the representations of the most attended tokens may contain crucial information.
Assuming that the teacher has $A_h$ attention
|
[
[
0.1722412109375,
0.14013671875,
0.18408203125,
0.14013671875,
0.18408203125,
0.144287109375,
0.1722412109375,
0.144287109375
],
[
0.257080078125,
0.14013671875,
0.267333984375,
0.14013671875,
0.267333984375,
0.1436767578125,
0.257080078125,
0.1436767578125
],
[
0.259765625,
0.1561279296875,
0.267333984375,
0.1561279296875,
0.267333984375,
0.1602783203125,
0.259765625,
0.1602783203125
],
[
0.1512451171875,
0.168701171875,
0.1798095703125,
0.168701171875,
0.1798095703125,
0.17578125,
0.1512451171875,
0.17578125
],
[
0.336181640625,
0.168701171875,
0.35791015625,
0.168701171875,
0.35791015625,
0.17578125,
0.336181640625,
0.17578125
],
[
0.356201171875,
0.17041015625,
0.374755859375,
0.17041015625,
0.374755859375,
0.174560546875,
0.356201171875,
0.174560546875
],
[
0.54541015625,
0.169189453125,
0.56396484375,
0.169189453125,
0.56396484375,
0.177001953125,
0.54541015625,
0.177001953125
],
[
0.1546630859375,
0.18115234375,
0.205078125,
0.18115234375,
0.205078125,
0.191162109375,
0.1546630859375,
0.191162109375
],
[
0.236083984375,
0.18115234375,
0.289794921875,
0.18115234375,
0.289794921875,
0.191162109375,
0.236083984375,
0.191162109375
],
[
0.332763671875,
0.18115234375,
0.38330078125,
0.18115234375,
0.38330078125,
0.191162109375,
0.332763671875,
0.191162109375
],
[
0.416015625,
0.18115234375,
0.46728515625,
0.18115234375,
0.46728515625,
0.191162109375,
0.416015625,
0.191162109375
],
[
0.54541015625,
0.177001953125,
0.56201171875,
0.179931640625,
0.55859375,
0.188232421875,
0.541015625,
0.185302734375
],
[
0.1588134765625,
0.1923828125,
0.19921875,
0.1923828125,
0.19921875,
0.203125,
0.1588134765625,
0.203125
],
[
0.245361328125,
0.193603515625,
0.282470703125,
0.193603515625,
0.282470703125,
0.203125,
0.245361328125,
0.203125
],
[
0.34033203125,
0.193603515625,
0.377197265625,
0.193603515625,
0.377197265625,
0.203125,
0.34033203125,
0.203125
],
[
0.423583984375,
0.193603515625,
0.46044921875,
0.193603515625,
0.46044921875,
0.203125,
0.423583984375,
0.203125
],
[
0.53857421875,
0.189453125,
0.560546875,
0.189453125,
0.560546875,
0.19775390625,
0.53857421875,
0.19775390625
],
[
0.537109375,
0.20068359375,
0.57080078125,
0.20068359375,
0.57080078125,
0.208984375,
0.537109375,
0.208984375
],
[
0.11846923828125,
0.21435546875,
0.49072265625,
0.21435546875,
0.49072265625,
0.22802734375,
0.11846923828125,
0.22802734375
],
[
0.12017822265625,
0.22802734375,
0.492431640625,
0.22802734375,
0.492431640625,
0.2423095703125,
0.12017822265625,
0.2423095703125
],
[
0.546875,
0.22265625,
0.55810546875,
0.22265625,
0.55810546875,
0.22802734375,
0.546875,
0.22802734375
],
[
0.548828125,
0.2327880859375,
0.560546875,
0.236328125,
0.55810546875,
0.2392578125,
0.546875,
0.236328125
],
[
0.11846923828125,
0.2410888671875,
0.411865234375,
0.2423095703125,
0.411865234375,
0.257080078125,
0.11846923828125,
0.255859375
],
[
0.591796875,
0.24462890625,
0.60986328125,
0.24462890625,
0.60986328125,
0.248779296875,
0.591796875,
0.248779296875
],
[
0.61083984375,
0.24462890625,
0.6455078125,
0.24462890625,
0.6455078125,
0.248779296875,
0.61083984375,
0.248779296875
],
[
0.65283203125,
0.24462890625,
0.6689453125,
0.24462890625,
0.6689453125,
0.248779296875,
0.65283203125,
0.248779296875
],
[
0.759765625,
0.24462890625,
0.7841796875,
0.24462890625,
0.7841796875,
0.248779296875,
0.759765625,
0.248779296875
],
[
0.7900390625,
0.24462890625,
0.7998046875,
0.24462890625,
0.7998046875,
0.2476806640625,
0.7900390625,
0.2476806640625
],
[
0.51611328125,
0.268310546875,
0.88330078125,
0.268310546875,
0.88330078125,
0.279052734375,
0.51611328125,
0.279052734375
],
[
0.11846923828125,
0.283935546875,
0.489013671875,
0.283935546875,
0.489013671875,
0.297607421875,
0.11846923828125,
0.297607421875
],
[
0.51416015625,
0.28076171875,
0.88330078125,
0.28076171875,
0.88330078125,
0.29443359375,
0.51416015625,
0.29443359375
],
[
0.12017822265625,
0.300537109375,
0.492431640625,
0.300537109375,
0.492431640625,
0.3154296875,
0.12017822265625,
0.3154296875
],
[
0.51416015625,
0.295654296875,
0.884765625,
0.295654296875,
0.884765625,
0.308837890625,
0.51416015625,
0.308837890625
],
[
0.1226806640625,
0.3154296875,
0.347900390625,
0.3154296875,
0.347900390625,
0.34033203125,
0.1226806640625,
0.34033203125
],
[
0.51416015625,
0.31005859375,
0.884765625,
0.31005859375,
0.884765625,
0.32373046875,
0.51416015625,
0.32373046875
],
[
0.40576171875,
0.320556640625,
0.49072265625,
0.320556640625,
0.49072265625,
0.33203125,
0.40576171875,
0.33203125
],
[
0.32275390625,
0.327880859375,
0.373046875,
0.327880859375,
0.373046875,
0.339111328125,
0.32275390625,
0.339111328125
],
[
0.51416015625,
0.32470703125,
0.6796875,
0.32470703125,
0.6796875,
0.33837890625,
0.51416015625,
0.33837890625
],
[
0.1168212890625,
0.33740234375,
0.49072265625,
0.339111328125,
0.49072265625,
0.35400390625,
0.1168212890625,
0.3515625
],
[
0.12017822265625,
0.355712890625,
0.288330078125,
0.355712890625,
0.288330078125,
0.366455078125,
0.12017822265625,
0.366455078125
],
[
0.1361083984375,
0.371826171875,
0.489013671875,
0.371826171875,
0.489013671875,
0.385986328125,
0.1361083984375,
0.385986328125
],
[
0.51416015625,
0.366455078125,
0.88232421875,
0.366455078125,
0.88232421875,
0.380126953125,
0.51416015625,
0.380126953125
],
[
0.51416015625,
0.3818359375,
0.88671875,
0.383056640625,
0.88671875,
0.39794921875,
0.51416015625,
0.396728515625
],
[
0.11846923828125,
0.388427734375,
0.36376953125,
0.388427734375,
0.36376953125,
0.402099609375,
0.11846923828125,
0.402099609375
],
[
0.51416015625,
0.399658203125,
0.88671875,
0.399658203125,
0.88671875,
0.413330078125,
0.51416015625,
0.413330078125
],
[
0.273193359375,
0.416259765625,
0.28662109375,
0.416259765625,
0.28662109375,
0.427001953125,
0.273193359375,
0.427001953125
],
[
0.51416015625,
0.4150390625,
0.884765625,
0.4150390625,
0.884765625,
0.4287109375,
0.51416015625,
0.4287109375
],
[
0.18408203125,
0.424560546875,
0.2646484375,
0.424560546875,
0.2646484375,
0.439453125,
0.18408203125,
0.439453125
],
[
0.279052734375,
0.424560546875,
0.421875,
0.421630859375,
0.421875,
0.435791015625,
0.279052734375,
0.439453125
],
[
0.4638671875,
0.42578125,
0.489013671875,
0.42578125,
0.489013671875,
0.438232421875,
0.4638671875,
0.438232421875
],
[
0.268798828125,
0.435302734375,
0.288330078125,
0.432861328125,
0.289794921875,
0.4423828125,
0.2705078125,
0.44482421875
],
[
0.51171875,
0.429931640625,
0.884765625,
0.43115234375,
0.884765625,
0.4453125,
0.51171875,
0.44482421875
],
[
0.51416015625,
0.44775390625,
0.884765625,
0.44775390625,
0.884765625,
0.4619140625,
0.51416015625,
0.4619140625
],
[
0.1168212890625,
0.4619140625,
0.4873046875,
0.46142578125,
0.4873046875,
0.4755859375,
0.1168212890625,
0.476806640625
],
[
0.5126953125,
0.46435546875,
0.884765625,
0.46435546875,
0.884765625,
0.47802734375,
0.5126953125,
0.47802734375
],
[
0.11846923828125,
0.478515625,
0.49072265625,
0.478515625,
0.49072265625,
0.4921875,
0.11846923828125,
0.4921875
],
[
0.5126953125,
0.479736328125,
0.88330078125,
0.479736328125,
0.88330078125,
0.49462890625,
0.5126953125,
0.49462890625
],
[
0.12017822265625,
0.49462890625,
0.49072265625,
0.49462890625,
0.49072265625,
0.5087890625,
0.12017822265625,
0.5087890625
],
[
0.51171875,
0.49462890625,
0.88671875,
0.495361328125,
0.88671875,
0.51025390625,
0.51171875,
0.5087890625
],
[
0.12017822265625,
0.51123046875,
0.49072265625,
0.51123046875,
0.49072265625,
0.52490234375,
0.12017822265625,
0.52490234375
],
[
0.51416015625,
0.5126953125,
0.884765625,
0.5126953125,
0.884765625,
0.525390625,
0.51416015625,
0.525390625
],
[
0.11846923828125,
0.525390625,
0.49072265625,
0.52685546875,
0.49072265625,
0.54150390625,
0.11846923828125,
0.54052734375
],
[
0.51416015625,
0.52783203125,
0.884765625,
0.52783203125,
0.884765625,
0.5419921875,
0.51416015625,
0.5419921875
],
[
0.11846923828125,
0.54345703125,
0.49072265625,
0.54345703125,
0.49072265625,
0.55712890625,
0.11846923828125,
0.55712890625
],
[
0.51416015625,
0.54443359375,
0.88330078125,
0.54443359375,
0.88330078125,
0.55810546875,
0.51416015625,
0.55810546875
],
[
0.12017822265625,
0.55859375,
0.49072265625,
0.55859375,
0.49072265625,
0.572265625,
0.12017822265625,
0.572265625
],
[
0.5126953125,
0.55859375,
0.88330078125,
0.55859375,
0.88330078125,
0.572265625,
0.5126953125,
0.572265625
],
[
0.12017822265625,
0.576171875,
0.49072265625,
0.576171875,
0.49072265625,
0.5888671875,
0.12017822265625,
0.5888671875
],
[
0.51416015625,
0.576171875,
0.88671875,
0.576171875,
0.88671875,
0.5888671875,
0.51416015625,
0.5888671875
],
[
0.1168212890625,
0.59033203125,
0.49072265625,
0.59033203125,
0.49072265625,
0.60498046875,
0.1168212890625,
0.60498046875
],
[
0.5126953125,
0.59130859375,
0.88671875,
0.59033203125,
0.88671875,
0.60498046875,
0.5126953125,
0.60546875
],
[
0.115966796875,
0.60498046875,
0.49072265625,
0.60546875,
0.49072265625,
0.62353515625,
0.115966796875,
0.62255859375
],
[
0.51416015625,
0.60791015625,
0.884765625,
0.60791015625,
0.884765625,
0.62158203125,
0.51416015625,
0.62158203125
],
[
0.1168212890625,
0.62353515625,
0.4580078125,
0.62255859375,
0.4580078125,
0.63720703125,
0.1168212890625,
0.63818359375
],
[
0.5126953125,
0.62451171875,
0.884765625,
0.62353515625,
0.884765625,
0.63818359375,
0.5126953125,
0.6396484375
],
[
0.5126953125,
0.6396484375,
0.884765625,
0.64013671875,
0.884765625,
0.65478515625,
0.5126953125,
0.65380859375
],
[
0.36962890625,
0.65283203125,
0.40576171875,
0.65087890625,
0.407470703125,
0.6630859375,
0.371337890625,
0.6650390625
],
[
0.188232421875,
0.66015625,
0.36962890625,
0.6591796875,
0.36962890625,
0.6767578125,
0.188232421875,
0.677734375
],
[
0.51416015625,
0.65673828125,
0.884765625,
0.65673828125,
0.884765625,
0.67041015625,
0.51416015625,
0.67041015625
],
[
0.465576171875,
0.6630859375,
0.489013671875,
0.6630859375,
0.489013671875,
0.67578125,
0.465576171875,
0.67578125
],
[
0.384765625,
0.66943359375,
0.3984375,
0.6748046875,
0.38916015625,
0.68603515625,
0.37646484375,
0.6796875
],
[
0.51611328125,
0.6748046875,
0.88330078125,
0.6748046875,
0.88330078125,
0.68505859375,
0.51611328125,
0.68505859375
],
[
0.51416015625,
0.689453125,
0.884765625,
0.6884765625,
0.884765625,
0.703125,
0.51416015625,
0.70361328125
],
[
0.115966796875,
0.6953125,
0.49072265625,
0.69775390625,
0.49072265625,
0.71435546875,
0.115966796875,
0.7119140625
],
[
0.51416015625,
0.705078125,
0.734375,
0.705078125,
0.734375,
0.7197265625,
0.51416015625,
0.7197265625
],
[
0.11846923828125,
0.71533203125,
0.489013671875,
0.71533203125,
0.489013671875,
0.728515625,
0.11846923828125,
0.728515625
],
[
0.1168212890625,
0.72998046875,
0.49072265625,
0.73095703125,
0.49072265625,
0.7451171875,
0.1168212890625,
0.74462890625
],
[
0.51171875,
0.73291015625,
0.72265625,
0.7353515625,
0.72265625,
0.7529296875,
0.51171875,
0.75048828125
],
[
0.12017822265625,
0.74658203125,
0.49072265625,
0.74658203125,
0.49072265625,
0.76123046875,
0.12017822265625,
0.76123046875
],
[
0.1168212890625,
0.7626953125,
0.49072265625,
0.76318359375,
0.49072265625,
0.77783203125,
0.1168212890625,
0.77685546875
],
[
0.5126953125,
0.75830078125,
0.75732421875,
0.7587890625,
0.75732421875,
0.77392578125,
0.5126953125,
0.7724609375
],
[
0.115966796875,
0.775390625,
0.49072265625,
0.77685546875,
0.49072265625,
0.79443359375,
0.115966796875,
0.79345703125
],
[
0.51171875,
0.779296875,
0.884765625,
0.78076171875,
0.884765625,
0.7958984375,
0.51171875,
0.79345703125
],
[
0.1168212890625,
0.7919921875,
0.492431640625,
0.7919921875,
0.492431640625,
0.81005859375,
0.1168212890625,
0.81005859375
],
[
0.51416015625,
0.79736328125,
0.88671875,
0.79736328125,
0.88671875,
0.81103515625,
0.51416015625,
0.81103515625
],
[
0.12017822265625,
0.8134765625,
0.2496337890625,
0.8134765625,
0.2496337890625,
0.82763671875,
0.12017822265625,
0.82763671875
],
[
0.230224609375,
0.81640625,
0.3984375,
0.81640625,
0.3984375,
0.83447265625,
0.230224609375,
0.83447265625
],
[
0.51416015625,
0.8134765625,
0.88671875,
0.8134765625,
0.88671875,
0.82666015625,
0.51416015625,
0.82666015625
],
[
0.51611328125,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84130859375,
0.51611328125,
0.84130859375
],
[
0.11846923828125,
0.84033203125,
0.34033203125,
0.84130859375,
0.34033203125,
0.85546875,
0.11846923828125,
0.85498046875
],
[
0.51416015625,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.85986328125,
0.51416015625,
0.85986328125
],
[
0.11846923828125,
0.86328125,
0.49072265625,
0.86328125,
0.49072265625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.5126953125,
0.8623046875,
0.884765625,
0.8623046875,
0.884765625,
0.8759765625,
0.5126953125,
0.8759765625
],
[
0.1168212890625,
0.87890625,
0.4931640625,
0.87744140625,
0.4931640625,
0.892578125,
0.1168212890625,
0.8935546875
],
[
0.5126953125,
0.87744140625,
0.72021484375,
0.87744140625,
0.72021484375,
0.89111328125,
0.5126953125,
0.89111328125
],
[
0.12017822265625,
0.8955078125,
0.489013671875,
0.8955078125,
0.489013671875,
0.9091796875,
0.12017822265625,
0.9091796875
],
[
0.53125,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.53125,
0.9091796875
]
] |
[
"Trm",
"rn",
"rm",
"Embe",
"Embe",
"eddin",
"87.75",
"Text Sequence",
"Text Sequence",
"Text Sequence",
"Text Sequence",
"87.5",
"Teacher",
"Student",
"Teacher",
"Student",
"9.87.2",
"287.00",
"Figure 2: Illustration of the redesigned uniform layer",
"mapping strategy. Left: Ltop is the top teacher layer.",
"86.",
"",
"Right: Ltop is the second-top teacher layer.",
"cr",
"anthr#e",
"ao",
"KDon",
"n",
"Figure3:ResultsofdepthcompressiononCoLA,SST",
"teacher layers for each student layer. This pro-",
"2, QNLI and MNLI. Each color denotes a layer map-",
"duces L' + 1 pairs of teacher-student features:",
"ping function. The number of layers in HSK includes",
"[(H, (),.. (H, HC",
"the embedding layer. Full results on seven tasks are",
"Second. a",
"g(L')",
"shown in Appendix E.1.",
"subset of these feature pairs are selected to per-",
"form HSK distillation.",
"For the first step, a simple but effective strategy",
"grey lines), depth compression generally has a neg",
"ative impact on the students' performance. Spe-",
"is the uniform mapping function:",
"cially, the performance of ROSITAg declines dras-",
"L",
"tically when the number of layers is reduced to",
"g(l) = l x",
",mod(L,L)= 0",
"(4)",
"Li",
"1 ~ 3. 2) In terms of the g(l, 10) and g(l, 8) map-",
"ping functions (the blue and orange lines), HSK",
"In this way, the teacher layers are divided into L'",
"distillation with only one or two layers can achieve",
"blocks and the top layer of each block serves as",
"comparable performance as using all the L' +1 lay-",
"the guidance in KD. Recently, Wang et al. (2020a)",
"ers. On the QNLI and MNLI datasets, the perfor-",
"empirically show that the upper-middle layers of",
"mance can even be improved by eliminating the",
"BERT, as compared with the top layer, are a bet-",
"lower layers. 3) In general, the student achieves",
"ter choice to guide the top layer of student in self-",
"better results with the redesigned layer mapping",
"attention distillation. Inspired by this, we redesign",
"function in Equation 5 across the four tasks. This",
"Equation 4 to allow the top student layer to dis-",
"demonstrates that, like the self-attention knowl-",
"till knowledge from an upper-middle teacher layer,",
"edge, the most crucial HSK does not necessar-",
"and the lower layers follow the uniform mapping",
"ily reside in the top BERT layer, which reveals",
"principle. This function can be formulated as:",
"a potential way to improve HSK distillation of",
"BERT. 4) Compared with g(1, 8), the improvement",
"Ltop",
"g(l, Ltop) = l × round(",
"brought by g(t, 10) is more stable across different",
"(5)",
"L",
"tasks and student models.Therefore, we use the",
"g(l, 10) layer mapping function when investigat-",
"where Ltop is the teacher layer corresponding to",
"ing the other two dimensions.",
"the top student layer and round() is the rounding-",
"off operation. Figure 2 gives an illustration of",
"4.2 Length Compression",
"g(l, Ltop) with a 6-layer teacher and a 3-layer stu-",
"dent. Specifically,for the 12-layer BERTBASE",
"4.2.1 Compression Strategies",
"teacher, we select Ltop from [8, 10, 12]. For the",
"To compress the length dimension, we design a",
"second step, we simply keep the top ND feature",
"method to measure the tokens' importance by us-",
"pairs: I(HS, HT",
"[g(,Ltop))F=L'-ND+1",
"ing the teacher's self-attention distribution. The",
"intuition is that self-attention controls the informa-",
"4.1.2Results and Analysis",
"tion flow among tokens across layers, and thus the",
"Figure 3 presents the results of depth compression",
"representations of the most attended tokens may",
"with different layer mapping functions. We can",
"contain crucial information.",
"find that: 1) For the g(t, 12) mapping function (the",
"Assuming that the teacher has An attention"
] |
{
"arxiv_id": "2106.05691",
"language": "en",
"page_idx": 3
}
|
||
2106.05691-en-4
|
2106.05691-en-4.jpg
|
<FigureHere>
Figure 5: Length compression results of TinyBERT$_{4}$ on CoLA, SST-2, QNLI and MNLI. The axes, points and lines are defined in the same way as Figure 4. Full results on seven tasks are shown in Appendix E.2.
<FigureHere>
Figure 4: Length compression results of ROSITA$_6$ on CoLA, SST-2, QNLI and MNLI. The horizontal axis represents the compressed HSK length normalized by full length. The left-most points in each plot mean compressing the length to one token. Full results on seven tasks are shown in Appendix E.2.
heads, and the attention weights in the $l^{th}$ layer
$\mathbf{is~}\mathbf{A}_{l}^{T}=\left\{\mathbf{A}_{l,a}^{T}\right\}_{a=1}^{A_{h}}$,where $\mathbf{A}_l,a^{T}\in\mathbb{R}^{|\mathbf{x}|\times|\mathbf{x}|}$ is the ing observations: 1) For all strategies, significant
attention matrix of the $a^{th}$ head. Each row of $\mathbf{A}_l,a^T$ is the attention distribution of a particular token to all the tokens. In our length compression strategy, the importance score of the tokens is the attention distribution of the [CLS] token (i.e., the frst row in A$_l,a^{T})$ averaged over the $A_h$ heads:
$$\mathbf{S}_l=\dfrac{1}{A_h}\sum_{a=1}^{A_h}\mathbf{A}_{l,a,1}^T,\mathbf{S}_l\in\mathbb{R}^{|\mathbf{x}|}$$
(6)
To match the depth of the student, we employ the layer mapping function in Equation 5 to select $\mathbf{S}_{g(l,L^{top})}$for the $l^th$ student layer.
The length compression strategies examined in
this section are summarized as:
Att is the attention-based strategy as described above. The layer mapping function to select S is the same as the one to select HSK, i.e., $g(l,10).$
Att w/o [SEP] excludes the HSK of the special token [SEP]. The rationality of this operation will be explained in the following analysis.
$\mathbf{Att}\left ( L^{top}= 12\right ) \mathbf{w} / \mathbf{o} \left [ \mathrm{SEP}\right ]$ is different from Att
$\mathbf{w}/\mathbf{o}$ [SEP] in that it utilizes $g(l,12)$ to select S.
Left is a naive baseline that discards tokens from the tail of the text sequence. When the token number is reduced to 1, the student only distills the HSK from the [CLS] token.
## 4.2.2 Results and Analysis
The length compression results are shown in Figperformance decline can only be observed when HSK length is compressed heavily (to less than $0.05\sim0.30).$ In some cases, using a subset of tokens' representation even leads to perceivable improvement over the full length (e.g., ROSITA$_6$ on CoLA and TinyBERT$_{4}$ on SST-2 and QNLI).2) The performance of Att is not satisfactory. When being applied to ROSITA$_6$, the Att strategy underperforms the Left baseline. The results of Att in TinyBERT$_4$,though better than those in ROSITA$_6$, still lag behind the other strategies at the left-most points. 3) Excluding [SEP] in the Att strategy alleviates the drop in performance, especially when HSK length is compressed to less than 0.05.4) As a general trend, further improvement over Att w/o [SEP] can be obtained by using $g(l,12)$ in the selection of S, which produces the most robust results among the four strategies.
To explain why the Att strategy performs poorly, we inspect into the tokens that receive the highest importance scores under Equation 6. We find that the special token [SEP] is dominant in most hidden layers. As shown in Figure 6, from the $4^{th}\sim10^{th}$ layers, [SEP] is the most attended token for almost all training samples. Meanwhile, [SEP] frequently appears in the top three positions across all the layers. Similar phenomenon was found in Clark et al. (2019), where [SEP] receives high attention scores from itself and other tokens in the middle layers. Combining this phe-
|
[
[
0.1512451171875,
0.1478271484375,
0.1722412109375,
0.144287109375,
0.1748046875,
0.15087890625,
0.1546630859375,
0.15380859375
],
[
0.1873779296875,
0.1519775390625,
0.294189453125,
0.1519775390625,
0.294189453125,
0.1591796875,
0.1873779296875,
0.1591796875
],
[
0.35205078125,
0.15087890625,
0.462158203125,
0.15087890625,
0.462158203125,
0.160888671875,
0.35205078125,
0.160888671875
],
[
0.61865234375,
0.15380859375,
0.63037109375,
0.15380859375,
0.63037109375,
0.157958984375,
0.61865234375,
0.157958984375
],
[
0.54541015625,
0.174560546875,
0.56396484375,
0.17578125,
0.56201171875,
0.182861328125,
0.54443359375,
0.1817626953125
],
[
0.70751953125,
0.221435546875,
0.724609375,
0.221435546875,
0.724609375,
0.2244873046875,
0.70751953125,
0.2244873046875
],
[
0.548828125,
0.23095703125,
0.55859375,
0.23095703125,
0.55859375,
0.235107421875,
0.548828125,
0.235107421875
],
[
0.5830078125,
0.235107421875,
0.6748046875,
0.2369384765625,
0.6748046875,
0.254638671875,
0.5830078125,
0.2529296875
],
[
0.349609375,
0.236328125,
0.450439453125,
0.236328125,
0.450439453125,
0.254150390625,
0.349609375,
0.254150390625
],
[
0.1873779296875,
0.236328125,
0.288330078125,
0.236328125,
0.288330078125,
0.254150390625,
0.1873779296875,
0.254150390625
],
[
0.326904296875,
0.2381591796875,
0.35205078125,
0.2381591796875,
0.35205078125,
0.2464599609375,
0.326904296875,
0.2464599609375
],
[
0.56201171875,
0.2392578125,
0.578125,
0.2392578125,
0.578125,
0.2464599609375,
0.56201171875,
0.2464599609375
],
[
0.74560546875,
0.236328125,
0.84619140625,
0.236328125,
0.84619140625,
0.254150390625,
0.74560546875,
0.254150390625
],
[
0.22265625,
0.240478515625,
0.236083984375,
0.2423095703125,
0.2344970703125,
0.2476806640625,
0.2210693359375,
0.2452392578125
],
[
0.1663818359375,
0.240478515625,
0.1781005859375,
0.240478515625,
0.1781005859375,
0.243408203125,
0.1663818359375,
0.243408203125
],
[
0.724609375,
0.240478515625,
0.73779296875,
0.240478515625,
0.73779296875,
0.24462890625,
0.724609375,
0.24462890625
],
[
0.11846923828125,
0.267333984375,
0.4931640625,
0.267333984375,
0.4931640625,
0.28076171875,
0.11846923828125,
0.28076171875
],
[
0.51171875,
0.265380859375,
0.880859375,
0.265380859375,
0.880859375,
0.27978515625,
0.51171875,
0.27978515625
],
[
0.11846923828125,
0.28076171875,
0.49072265625,
0.28076171875,
0.49072265625,
0.29443359375,
0.11846923828125,
0.29443359375
],
[
0.51171875,
0.27978515625,
0.884765625,
0.28076171875,
0.884765625,
0.295654296875,
0.51171875,
0.29443359375
],
[
0.11846923828125,
0.295654296875,
0.492431640625,
0.295654296875,
0.492431640625,
0.308837890625,
0.11846923828125,
0.308837890625
],
[
0.5126953125,
0.295654296875,
0.884765625,
0.295654296875,
0.884765625,
0.308837890625,
0.5126953125,
0.308837890625
],
[
0.11846923828125,
0.31005859375,
0.492431640625,
0.31005859375,
0.492431640625,
0.32373046875,
0.11846923828125,
0.32373046875
],
[
0.5126953125,
0.31005859375,
0.8505859375,
0.31005859375,
0.8505859375,
0.32373046875,
0.5126953125,
0.32373046875
],
[
0.11846923828125,
0.322509765625,
0.49072265625,
0.32373046875,
0.49072265625,
0.33837890625,
0.11846923828125,
0.33740234375
],
[
0.11846923828125,
0.33837890625,
0.34619140625,
0.33837890625,
0.34619140625,
0.352783203125,
0.11846923828125,
0.352783203125
],
[
0.51416015625,
0.35400390625,
0.734375,
0.35400390625,
0.734375,
0.36767578125,
0.51416015625,
0.36767578125
],
[
0.11846923828125,
0.37890625,
0.494873046875,
0.37890625,
0.494873046875,
0.393798828125,
0.11846923828125,
0.393798828125
],
[
0.51171875,
0.377197265625,
0.884765625,
0.377685546875,
0.884765625,
0.392578125,
0.51171875,
0.391357421875
],
[
0.11260986328125,
0.3984375,
0.267333984375,
0.391357421875,
0.2705078125,
0.412109375,
0.115966796875,
0.42041015625
],
[
0.265625,
0.3984375,
0.49072265625,
0.396728515625,
0.49072265625,
0.41455078125,
0.265625,
0.416259765625
],
[
0.51416015625,
0.394287109375,
0.884765625,
0.394287109375,
0.884765625,
0.407958984375,
0.51416015625,
0.407958984375
],
[
0.209228515625,
0.405029296875,
0.2479248046875,
0.4091796875,
0.245361328125,
0.419189453125,
0.2066650390625,
0.4150390625
],
[
0.51416015625,
0.410400390625,
0.884765625,
0.410400390625,
0.884765625,
0.42333984375,
0.51416015625,
0.42333984375
],
[
0.1168212890625,
0.418701171875,
0.489013671875,
0.42041015625,
0.489013671875,
0.439453125,
0.1168212890625,
0.43701171875
],
[
0.51416015625,
0.427001953125,
0.884765625,
0.427001953125,
0.884765625,
0.440673828125,
0.51416015625,
0.440673828125
],
[
0.12017822265625,
0.441162109375,
0.489013671875,
0.441162109375,
0.489013671875,
0.451904296875,
0.12017822265625,
0.451904296875
],
[
0.51416015625,
0.4423828125,
0.884765625,
0.4423828125,
0.884765625,
0.4560546875,
0.51416015625,
0.4560546875
],
[
0.1168212890625,
0.45361328125,
0.492431640625,
0.4560546875,
0.492431640625,
0.47021484375,
0.1168212890625,
0.468505859375
],
[
0.51171875,
0.457763671875,
0.884765625,
0.457763671875,
0.884765625,
0.471435546875,
0.51171875,
0.471435546875
],
[
0.11846923828125,
0.471435546875,
0.492431640625,
0.471435546875,
0.492431640625,
0.485107421875,
0.11846923828125,
0.485107421875
],
[
0.51416015625,
0.474365234375,
0.88671875,
0.474365234375,
0.88671875,
0.488037109375,
0.51416015625,
0.488037109375
],
[
0.12017822265625,
0.48681640625,
0.49072265625,
0.48681640625,
0.49072265625,
0.50048828125,
0.12017822265625,
0.50048828125
],
[
0.51611328125,
0.490478515625,
0.88671875,
0.490478515625,
0.88671875,
0.5048828125,
0.51611328125,
0.5048828125
],
[
0.11846923828125,
0.5029296875,
0.392333984375,
0.5029296875,
0.392333984375,
0.51708984375,
0.11846923828125,
0.51708984375
],
[
0.51416015625,
0.505859375,
0.884765625,
0.505859375,
0.884765625,
0.51953125,
0.51416015625,
0.51953125
],
[
0.5126953125,
0.5224609375,
0.88330078125,
0.5224609375,
0.88330078125,
0.5361328125,
0.5126953125,
0.5361328125
],
[
0.265625,
0.53759765625,
0.288330078125,
0.53759765625,
0.288330078125,
0.54736328125,
0.265625,
0.54736328125
],
[
0.51416015625,
0.5380859375,
0.88671875,
0.5380859375,
0.88671875,
0.552734375,
0.51416015625,
0.552734375
],
[
0.245361328125,
0.54443359375,
0.25537109375,
0.54443359375,
0.25537109375,
0.5517578125,
0.245361328125,
0.5517578125
],
[
0.191650390625,
0.55029296875,
0.22265625,
0.55029296875,
0.22265625,
0.5625,
0.191650390625,
0.5625
],
[
0.28076171875,
0.548828125,
0.41943359375,
0.54345703125,
0.4208984375,
0.56396484375,
0.282470703125,
0.5693359375
],
[
0.465576171875,
0.55029296875,
0.489013671875,
0.55029296875,
0.489013671875,
0.5634765625,
0.465576171875,
0.5634765625
],
[
0.5126953125,
0.552734375,
0.884765625,
0.552734375,
0.884765625,
0.56787109375,
0.5126953125,
0.56787109375
],
[
0.51416015625,
0.57080078125,
0.88671875,
0.57080078125,
0.88671875,
0.5849609375,
0.51416015625,
0.5849609375
],
[
0.233642578125,
0.552734375,
0.2958984375,
0.5634765625,
0.288330078125,
0.5830078125,
0.2269287109375,
0.572265625
],
[
0.1168212890625,
0.59033203125,
0.4873046875,
0.59033203125,
0.4873046875,
0.60400390625,
0.1168212890625,
0.60400390625
],
[
0.5126953125,
0.5859375,
0.88330078125,
0.5859375,
0.88330078125,
0.599609375,
0.5126953125,
0.599609375
],
[
0.11846923828125,
0.60791015625,
0.489013671875,
0.60693359375,
0.489013671875,
0.62158203125,
0.11846923828125,
0.62255859375
],
[
0.51416015625,
0.6025390625,
0.88671875,
0.6025390625,
0.88671875,
0.61767578125,
0.51416015625,
0.61767578125
],
[
0.1168212890625,
0.62158203125,
0.368896484375,
0.62060546875,
0.368896484375,
0.6396484375,
0.1168212890625,
0.64013671875
],
[
0.51416015625,
0.619140625,
0.884765625,
0.619140625,
0.884765625,
0.6328125,
0.51416015625,
0.6328125
],
[
0.5126953125,
0.63427734375,
0.884765625,
0.63427734375,
0.884765625,
0.64794921875,
0.5126953125,
0.64794921875
],
[
0.1378173828125,
0.64111328125,
0.49072265625,
0.64111328125,
0.49072265625,
0.65478515625,
0.1378173828125,
0.65478515625
],
[
0.11846923828125,
0.65673828125,
0.35205078125,
0.65673828125,
0.35205078125,
0.67041015625,
0.11846923828125,
0.67041015625
],
[
0.51416015625,
0.65185546875,
0.88330078125,
0.65185546875,
0.88330078125,
0.6650390625,
0.51416015625,
0.6650390625
],
[
0.5126953125,
0.66650390625,
0.88330078125,
0.66748046875,
0.88330078125,
0.681640625,
0.5126953125,
0.68115234375
],
[
0.11846923828125,
0.68310546875,
0.4873046875,
0.68310546875,
0.4873046875,
0.69677734375,
0.11846923828125,
0.69677734375
],
[
0.51611328125,
0.68310546875,
0.88671875,
0.68310546875,
0.88671875,
0.69677734375,
0.51611328125,
0.69677734375
],
[
0.12017822265625,
0.70068359375,
0.489013671875,
0.70068359375,
0.489013671875,
0.71533203125,
0.12017822265625,
0.71533203125
],
[
0.51416015625,
0.69873046875,
0.74560546875,
0.69873046875,
0.74560546875,
0.7119140625,
0.51416015625,
0.7119140625
],
[
0.115966796875,
0.71533203125,
0.4697265625,
0.71630859375,
0.4697265625,
0.73095703125,
0.115966796875,
0.72998046875
],
[
0.53125,
0.71630859375,
0.884765625,
0.71728515625,
0.884765625,
0.732421875,
0.53125,
0.73095703125
],
[
0.5126953125,
0.73291015625,
0.88232421875,
0.73291015625,
0.88232421875,
0.74658203125,
0.5126953125,
0.74658203125
],
[
0.12017822265625,
0.74365234375,
0.489013671875,
0.74365234375,
0.489013671875,
0.75732421875,
0.12017822265625,
0.75732421875
],
[
0.5126953125,
0.75,
0.88330078125,
0.7490234375,
0.88330078125,
0.7626953125,
0.5126953125,
0.76318359375
],
[
0.11846923828125,
0.7587890625,
0.4873046875,
0.7587890625,
0.4873046875,
0.7724609375,
0.11846923828125,
0.7724609375
],
[
0.51416015625,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.51416015625,
0.77978515625
],
[
0.11846923828125,
0.77685546875,
0.404296875,
0.77685546875,
0.404296875,
0.79052734375,
0.11846923828125,
0.79052734375
],
[
0.51416015625,
0.783203125,
0.884765625,
0.783203125,
0.884765625,
0.79638671875,
0.51416015625,
0.79638671875
],
[
0.12017822265625,
0.8017578125,
0.489013671875,
0.8017578125,
0.489013671875,
0.81640625,
0.12017822265625,
0.81640625
],
[
0.51611328125,
0.79736328125,
0.884765625,
0.79736328125,
0.884765625,
0.8125,
0.51611328125,
0.8125
],
[
0.11846923828125,
0.8193359375,
0.46728515625,
0.8193359375,
0.46728515625,
0.83447265625,
0.11846923828125,
0.83447265625
],
[
0.51416015625,
0.8154296875,
0.88671875,
0.8154296875,
0.88671875,
0.8291015625,
0.51416015625,
0.8291015625
],
[
0.51611328125,
0.83056640625,
0.88671875,
0.83056640625,
0.88671875,
0.845703125,
0.51611328125,
0.845703125
],
[
0.11846923828125,
0.8466796875,
0.492431640625,
0.8466796875,
0.492431640625,
0.85986328125,
0.11846923828125,
0.85986328125
],
[
0.5126953125,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.5126953125,
0.86083984375
],
[
0.11846923828125,
0.86328125,
0.4931640625,
0.86328125,
0.4931640625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.5126953125,
0.86328125,
0.88671875,
0.86328125,
0.88671875,
0.876953125,
0.5126953125,
0.876953125
],
[
0.11846923828125,
0.87890625,
0.4931640625,
0.87890625,
0.4931640625,
0.892578125,
0.11846923828125,
0.892578125
],
[
0.5126953125,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.892578125,
0.5126953125,
0.892578125
],
[
0.12017822265625,
0.89453125,
0.324462890625,
0.89453125,
0.324462890625,
0.9091796875,
0.12017822265625,
0.9091796875
],
[
0.5126953125,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5126953125,
0.9091796875
]
] |
[
"25.0/",
"050.100150.200250.3",
"050.100.150.200.250.30",
"01",
"87.00",
"34.36",
"5.7",
"AR Lengin Nomaized",
"Ak Legto Noiaizeao",
"As Lengio Noaialeao",
"0.00",
"0.00",
"Ask Lengtn Noraleao",
"10",
"n.or",
"0.00",
"Figure 4: Length compression results of ROSITAg on",
"Figure 5: Length compression results of TinyBERT",
"CoLA, SST-2, QNLI and MNLI. The horizontal axis",
"on CoLA, SST-2, QNLI and MNLI. The axes, points",
"represents the compressed HSK length normalized by",
"and lines are defined in the same way as Figure 4. Full",
"full length. The left-most points in each plot mean com-",
"results on seven tasks are shown in Appendix E.2.",
"pressing the length to one token. Full results on seven",
"tasks are shown in Appendix E.2.",
"4.2.2Results and Analysis",
"heads, and the attention weights in the ith layer",
"The length compression results are shown in Fig-",
"is AT =[AT,Ah",
", where AT, E R/xxxl is the",
"ure 4 and Figure 5.We can derive the follow-",
"Aia?",
"ing observations: 1) For all strategies, significant",
"attention matrix of the ath head. Each row of AZ.",
"performance decline can only be observed when",
"is the attention distribution of a particular token to",
"HSK length is compressed heavily (to less than",
"all the tokens. In our length compression strategy,",
"0.05 ~ 0.30). In some cases, using a subset of to-",
"the importance score of the tokens is the attention",
"kens' representation even leads to perceivable im-",
"distribution of the [CLS] token (i.e., the first row",
"provement over the full length (e.g., ROSITAg on",
"in AT.) averaged over the An heads:",
"CoLA and TinyBERT on SST-2 and QNLI). 2)",
"The performance of Att is not satisfactory. When",
"Ah",
"being applied to ROSITA6, the Att strategy under-",
"1.",
"S, =",
"CAT,a,1, St E R/xl",
"(6)",
"performs the Left baseline. The results of Att in",
"TinyBERT4, though better than those in ROSITA6,",
"An a=",
"To match the depth of the student. we employ the",
" still lag behind the other strategies at the left-most",
"layer mapping function in Equation 5 to select",
"points. 3) Excluding [SEP] in the Att strategy al-",
"Sg(l,Ltop) for the Ith student layer.",
"leviates the drop in performance, especially when",
"HSK length is compressed to less than 0.05. 4)",
"The length compression strategies examined in",
"this section are summarized as:",
"As a general trend, further improvement over Att",
"w/o [SEP] can be obtained by using g(l, 12) in the",
"Att is the attention-based strategy as described",
"selection of S, which produces the most robust re-",
"above. The layer mapping function to select S is",
"sults among the four strategies.",
"the same as the one to select HSK, i.e., g(l, 10).",
"To explain why the Att strategy performs",
"poorly, we inspect into the tokens that receive the",
"Att wlo [SEP]excludes the HSK of the special",
"highest importance scores under Equation 6. We",
"token [SEPl. The rationality of this operation will",
"find that the special token [SEP] is dominant in",
"be explained in the following analysis.",
"most hidden layers. As shown in Figure 6, from",
"Att (Ltop = 12) w/o [SEP] is different from Att",
"the 4th ~ 1oth layers, [SEP] is the most attended",
"w/o [SEP] in that it utilizes g(l, 12) to select S.",
"token for almost all training samples. Meanwhile,",
"[SEP] frequently appears in the top three posi-",
"Leftis a naive baseline that discards tokens from",
"tions across all the layers. Similar phenomenon",
"the tail of the text sequence. When the token num-",
"was found in Clark et al. (2019), where [SEP] re-",
"ber is reduced to 1, the student only distills the",
"ceives high attention scores from itself and other",
"HSK from the [CLS] token.",
"tokens in the middle layers. Combining this phe-"
] |
{
"arxiv_id": "2106.05691",
"language": "en",
"page_idx": 4
}
|
||
1804.07461-en-0
|
1804.07461-en-0.jpg
|
Published as a conference paper at ICLR 2019
GLUE: A MULTI-TASK BENCHMARK AND ANALYSIS $\mathsf{PL}$АТFORM FOR NATURAL LANGUAGE UNDERSTANDING
$\mathbf{Alex~Wang}^1,\mathbf{Amanpreet~Singh}^1,\mathbf{Julian~Michael}^2,\mathbf{Felix~Hill}^3$,
Omer Levy$^2$ &Samuel R. Bowman$^{1}$
$^{1}$Courant Institute of Mathematical Sciences, New York University
$^{2}$Paul G. Allen School of Computer Science & Engineering, University of Washington
$^{3}$DeepMind
$\{$alexwang, amanpreet, bowman$\}$enyu.edu
$\{$julianjm,omerlevy$\}$ecs.washington.edu
felixhillegoogle. com
### ABSTRACT
For natural language understanding (NLU) technology to be maximally useful, it must be able to process language in a way that is not exclusive to a single task, genre, or dataset. In pursuit of this objective, we introduce the General Language Understanding Evaluation (GLUE) benchmark, a collection of tools for evaluating the performance of models across a diverse set of existing NLU tasks. By including tasks with limited training data, GLUE is designed to favor and encourage models that share general linguistic knowledge across tasks. GLUE also includes a hand-crafted diagnostic test suite that enables detailed linguistic analysis of models. We evaluate baselines based on current methods for transfer and representation learning and find that multi-task training on all tasks performs better than training a separate model per task. However, the low absolute performance of our best model indicates the need for improved general NLU systems.
## 1 INTRODUCTION
The human ability to understand language is general, flexible, and $robust.$ In contrast, most NLU models above the word level are designed for a specific task and struggle with out-of-domain data. If we aspire to develop models with understanding beyond the detection of superficial correspondences between inputs and outputs, then it is critical to develop a more unified model that can learn to execute a range of different linguistic tasks in different domains.
To facilitate research in this direction, we present the General Language Understanding Evaluation (GLUE) benchmark: a collection of NLÛtasks including question answering, sentiment analysis, and textual entailment, and an associated online platform for model evaluation, comparison, and analysis. GLUE does not place any constraints on model architecture beyond the ability to process single-sentence and sentence-pair inputs and to make corresponding predictions. For some GLUE tasks, training data is plentiful, but for others it is limited or fails to match the genre of the test set. GLUE therefore favors models that can learn to represent linguistic knowledge in a way that facilitates sample-efficient learning and effective knowledge-transfer across tasks. None of the datasets in GLUE were created from scratch for the benchmark; we rely on preexisting datasets because they have been implicitly agreed upon by the NLP community as challenging and interesting. Four of the datasets feature privately-held test data, which will be used to ensure that the benchmark is used fairly.$^{1}$
To understand the types of knowledge learned by models and to encourage linguistic-meaningful solution strategies, GLUE also includes a set of hand-crafted analysis examples for probing trained models. This dataset is designed to highlight common challenges, such as the use of world knowledge and logical operators, that we expect models must handle to robustly solve the tasks.
$^{\text{l}}$To evaluate on the private test data, users of the benchmark must submit to gluebenchmark.com
1
|
[
[
0.1732177734375,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.176513671875,
0.10101318359375,
0.82275390625,
0.10162353515625,
0.82275390625,
0.11932373046875,
0.176513671875,
0.11871337890625
],
[
0.17724609375,
0.1287841796875,
0.81884765625,
0.1287841796875,
0.81884765625,
0.143310546875,
0.17724609375,
0.143310546875
],
[
0.1732177734375,
0.15283203125,
0.22216796875,
0.15283203125,
0.22216796875,
0.1717529296875,
0.1732177734375,
0.1717529296875
],
[
0.1829833984375,
0.1932373046875,
0.61279296875,
0.1944580078125,
0.61279296875,
0.212158203125,
0.1829833984375,
0.2108154296875
],
[
0.1854248046875,
0.212158203125,
0.4345703125,
0.212158203125,
0.4345703125,
0.2259521484375,
0.1854248046875,
0.2259521484375
],
[
0.1854248046875,
0.2259521484375,
0.62109375,
0.2259521484375,
0.62109375,
0.2392578125,
0.1854248046875,
0.2392578125
],
[
0.1829833984375,
0.238037109375,
0.7490234375,
0.240478515625,
0.7490234375,
0.255126953125,
0.1829833984375,
0.25244140625
],
[
0.1829833984375,
0.25244140625,
0.265625,
0.255126953125,
0.265625,
0.26953125,
0.1829833984375,
0.267578125
],
[
0.0228729248046875,
0.26513671875,
0.05963134765625,
0.26513671875,
0.05963134765625,
0.404052734375,
0.0228729248046875,
0.404052734375
],
[
0.1846923828125,
0.268310546875,
0.52392578125,
0.267578125,
0.52392578125,
0.2822265625,
0.1846923828125,
0.28271484375
],
[
0.1854248046875,
0.2822265625,
0.54150390625,
0.2822265625,
0.54150390625,
0.29541015625,
0.1854248046875,
0.29541015625
],
[
0.1854248046875,
0.296630859375,
0.384033203125,
0.296630859375,
0.384033203125,
0.31005859375,
0.1854248046875,
0.31005859375
],
[
0.453369140625,
0.3466796875,
0.54736328125,
0.3466796875,
0.54736328125,
0.361083984375,
0.453369140625,
0.361083984375
],
[
0.2320556640625,
0.375,
0.7646484375,
0.375,
0.7646484375,
0.389404296875,
0.2320556640625,
0.389404296875
],
[
0.236083984375,
0.392578125,
0.76318359375,
0.392578125,
0.76318359375,
0.40283203125,
0.236083984375,
0.40283203125
],
[
0.0236968994140625,
0.404052734375,
0.062103271484375,
0.404052734375,
0.062103271484375,
0.49365234375,
0.0236968994140625,
0.49365234375
],
[
0.233642578125,
0.404052734375,
0.765625,
0.40478515625,
0.765625,
0.41845703125,
0.233642578125,
0.417236328125
],
[
0.233642578125,
0.41845703125,
0.76318359375,
0.41845703125,
0.76318359375,
0.431884765625,
0.233642578125,
0.431884765625
],
[
0.236083984375,
0.434326171875,
0.76318359375,
0.434326171875,
0.76318359375,
0.4443359375,
0.236083984375,
0.4443359375
],
[
0.236083984375,
0.447509765625,
0.76318359375,
0.447509765625,
0.76318359375,
0.458251953125,
0.236083984375,
0.458251953125
],
[
0.233642578125,
0.458984375,
0.7646484375,
0.458251953125,
0.7646484375,
0.472900390625,
0.233642578125,
0.473388671875
],
[
0.233642578125,
0.472900390625,
0.765625,
0.473388671875,
0.765625,
0.4873046875,
0.233642578125,
0.486083984375
],
[
0.0228729248046875,
0.4873046875,
0.05963134765625,
0.4873046875,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.2344970703125,
0.4873046875,
0.7646484375,
0.4873046875,
0.7646484375,
0.50048828125,
0.2344970703125,
0.50048828125
],
[
0.236083984375,
0.50244140625,
0.7646484375,
0.50244140625,
0.7646484375,
0.51318359375,
0.236083984375,
0.51318359375
],
[
0.233642578125,
0.51513671875,
0.7646484375,
0.51513671875,
0.7646484375,
0.52880859375,
0.233642578125,
0.52880859375
],
[
0.2344970703125,
0.52880859375,
0.70751953125,
0.52880859375,
0.70751953125,
0.54248046875,
0.2344970703125,
0.54248046875
],
[
0.17724609375,
0.5693359375,
0.19775390625,
0.5693359375,
0.19775390625,
0.580078125,
0.17724609375,
0.580078125
],
[
0.19775390625,
0.5673828125,
0.338134765625,
0.5673828125,
0.338134765625,
0.58056640625,
0.19775390625,
0.58056640625
],
[
0.1748046875,
0.59521484375,
0.82421875,
0.59521484375,
0.82421875,
0.60986328125,
0.1748046875,
0.60986328125
],
[
0.1748046875,
0.611328125,
0.826171875,
0.611328125,
0.826171875,
0.625,
0.1748046875,
0.625
],
[
0.1748046875,
0.625,
0.82421875,
0.625,
0.82421875,
0.63818359375,
0.1748046875,
0.63818359375
],
[
0.17236328125,
0.63818359375,
0.826171875,
0.63720703125,
0.826171875,
0.6513671875,
0.17236328125,
0.65283203125
],
[
0.1748046875,
0.65283203125,
0.59619140625,
0.65283203125,
0.59619140625,
0.66650390625,
0.1748046875,
0.66650390625
],
[
0.1748046875,
0.673828125,
0.82275390625,
0.673828125,
0.82275390625,
0.6875,
0.1748046875,
0.6875
],
[
0.1748046875,
0.68798828125,
0.82177734375,
0.68798828125,
0.82177734375,
0.7021484375,
0.1748046875,
0.7021484375
],
[
0.1748046875,
0.7021484375,
0.82421875,
0.7021484375,
0.82421875,
0.71533203125,
0.1748046875,
0.71533203125
],
[
0.176513671875,
0.71728515625,
0.82275390625,
0.71728515625,
0.82275390625,
0.72802734375,
0.176513671875,
0.72802734375
],
[
0.1748046875,
0.72998046875,
0.82421875,
0.72998046875,
0.82421875,
0.74365234375,
0.1748046875,
0.74365234375
],
[
0.1748046875,
0.74365234375,
0.82275390625,
0.74365234375,
0.82275390625,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.176513671875,
0.7587890625,
0.8203125,
0.7587890625,
0.8203125,
0.76953125,
0.176513671875,
0.76953125
],
[
0.176513671875,
0.77294921875,
0.82275390625,
0.77294921875,
0.82275390625,
0.78271484375,
0.176513671875,
0.78271484375
],
[
0.1748046875,
0.78515625,
0.82275390625,
0.78515625,
0.82275390625,
0.798828125,
0.1748046875,
0.798828125
],
[
0.1748046875,
0.798828125,
0.82275390625,
0.798828125,
0.82275390625,
0.8125,
0.1748046875,
0.8125
],
[
0.176513671875,
0.814453125,
0.82275390625,
0.814453125,
0.82275390625,
0.82470703125,
0.176513671875,
0.82470703125
],
[
0.1748046875,
0.82763671875,
0.2197265625,
0.82763671875,
0.2197265625,
0.8388671875,
0.1748046875,
0.8388671875
],
[
0.17236328125,
0.84521484375,
0.82275390625,
0.8466796875,
0.82275390625,
0.861328125,
0.17236328125,
0.85986328125
],
[
0.176513671875,
0.86279296875,
0.82275390625,
0.86279296875,
0.82275390625,
0.87353515625,
0.176513671875,
0.87353515625
],
[
0.1748046875,
0.87548828125,
0.82275390625,
0.87548828125,
0.82275390625,
0.8896484375,
0.1748046875,
0.8896484375
],
[
0.1732177734375,
0.89013671875,
0.7587890625,
0.89013671875,
0.7587890625,
0.90380859375,
0.1732177734375,
0.90380859375
],
[
0.1953125,
0.91015625,
0.7939453125,
0.9111328125,
0.7939453125,
0.92529296875,
0.1953125,
0.9248046875
],
[
0.494384765625,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.9619140625,
0.494384765625,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2019",
"GLUE: A MULTI-TASK BENCHMARK AND ANALYSIS",
"PLATFORM FOR NATURAL LANGUAGE UNDERSTAND-",
"ING",
"Alex Wang', Amanpreet Singh', Julian Michael?, Felix Hill3,",
"Omer Levy? & Samuel R. Bowman",
"1Courant Institute of Mathematical Sciences, New York University",
"2Paul G.Allen School of Computer Science & Engineering,University of Washington",
"3DeepMind",
"",
"[alexwang,amanpreet,bowman]@nyu.edu",
"julianjm,omerlevy)@cs.washington.edu",
"felixhill@google.com",
"ABSTRACT",
"For natural language understanding (NLU) technology to be maximally useful, it",
"mustbeabletoprocesslanguageinawaythatisnotexclusivetoa singletask.",
"[T's]",
"genre, or dataset. In pursuit of this objective, we introduce the General Language",
"Understanding Evaluation (GLUE) benchmark, a collection of tools for evaluat-",
"ingtheperformance of models acrossadiversesetof existingNLUtasks.By",
"includingtaskswithlimitedtrainingdata,GLUEisdesignedtofavorand encour-",
"age models that share general linguistic knowledge across tasks. GLUE also in-",
"cludes a hand-crafted diagnostic test suite that enables detailed linguistic analysis",
"",
"of models. We evaluate baselines based on current methods for transfer and rep-",
"resentationlearningandfindthatmulti-tasktrainingonalltasksperformsbetter",
"than training a separate model per task. However, the low absolute performance",
"of our best model indicates the need for improved general NLU systems.",
"1",
"INTRODUCTION",
"The human ability to understand language is general, fexible, and robust. In contrast, most NLU",
"models above the word level are designed for a specific task and struggle with out-of-domain data. If",
"we aspire to develop models with understanding beyond the detection of superficial correspondences",
"between inputs and outputs, then it is critical to develop a more unified model that can learn to",
"execute a range of different linguistic tasks in different domains.",
"To facilitate research in this direction, we present the General Language Understanding Evaluation",
"(GLUE) benchmark: a collection of NLU tasks including question answering, sentiment analysis",
"and textual entailment, and an associated online platform for model evaluation, comparison, and",
"analysis.GLUEdoesnotplaceanyconstraints onmodelarchitecturebeyondtheabilitytoprocess",
"single-sentence and sentence-pair inputs and to make corresponding predictions. For some GLUE",
"tasks, training data is plentiful, but for others it is limited or fails to match the genre of the test set",
"GLUEthereforefavorsmodels that canlearntorepresentlinguisticknowledgein a waythatfacil",
"itatessample-efficientlearningandeffectiveknowledge-transferacrosstasks.Noneofthedatasets",
"in GLUE were created from scratch for the benchmark; we rely on preexisting datasets because they",
"have been implicitly agreed upon by the NLP community as challenging and interesting. Four of",
"thedatasetsfeatureprivately-heldtestdata,whichwillbeusedtoensurethatthebenchmarkisused",
"fairly.'",
"To understand the types of knowledge learned by models and to encourage linguistic-meaningful",
"solution strategies,GLUE also includes a set of hand-crafted analysisexamplesfor probingtrained",
"models. This dataset is designed to highlight common challenges, such as the use of world knowl-",
"edge and logical operators, that we expect models must handle to robustly solve the tasks",
"'To evaluate on the private test data, users of the benchmark must submit to gluebenchmark.com",
"1"
] |
{
"arxiv_id": "1804.07461",
"language": "en",
"page_idx": 0
}
|
||
1804.07461-en-1
|
1804.07461-en-1.jpg
|
Published as a conference paper at ICLR 2019
<table>
<tbody>
<tr>
<th>Corpus</th>
<th>|Train</th>
<th>|Test| </th>
<th>Task</th>
<th>$\mathbf{Metrics}$</th>
<th>$\mathbf{D}_{0}$main</th>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>$\operatorname{Single-Se}$</td>
<td>entence Tasks</td>
<td> </td>
</tr>
<tr>
<td>CoLA</td>
<td>$8.5k$</td>
<td>$1\mathbf{k}$</td>
<td>acceptabilitw</td>
<td>Matthews corr.</td>
<td>misc</td>
</tr>
<tr>
<td>CoLA $SST-2$</td>
<td>8.5k $67k$</td>
<td>$1\mathbf{k}$ 1.8k</td>
<td>acceptability sentiment</td>
<td>Matthews corr. $acc.$</td>
<td>$misc.$ movie reviews</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Similarity anc</td>
<td>${\mathrm{l~Paraphrase~Tasks}}$</td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{MRPC}$</td>
<td>3.7k</td>
<td>$1.7\mathbf{k}$</td>
<td>paraphrase</td>
<td>acc. / F1</td>
<td>news</td>
</tr>
<tr>
<td>STS- B</td>
<td>7k</td>
<td>1.4k</td>
<td>similarity sentence</td>
<td>Pearson/Spearman corr.</td>
<td>$misc.$</td>
</tr>
<tr>
<td>$QQP$</td>
<td>$364k$</td>
<td>$391k$</td>
<td>paraphrase</td>
<td>acc. / F1</td>
<td>social $QA$ questions</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Infere</td>
<td>ence Tasks</td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{MNLI}$</td>
<td>$393k$</td>
<td>$20k$</td>
<td>$NLI$</td>
<td>matched acc./mismatched acc.</td>
<td>$misc.$</td>
</tr>
<tr>
<td>ONLI</td>
<td>$105k$</td>
<td>5.4k</td>
<td>OA/ NLI</td>
<td>$acc.$</td>
<td>Wikipedia</td>
</tr>
<tr>
<td>$RTE$</td>
<td>2.5k</td>
<td>$3k$</td>
<td>$NLI$</td>
<td>$acc.$</td>
<td>news, Wikipedia</td>
</tr>
<tr>
<td>$WNLI$</td>
<td>634</td>
<td>146</td>
<td>coreference/NLI</td>
<td>$acc.$</td>
<td>fiction books</td>
</tr>
</tbody>
</table>
Table 1: Task descriptions and statistics. All tasks are single sentence or sentence pair classification, except STS-B, which is a regression task. MNLI has three classes; all other classification tasks have two. Test sets shown in bold use labels that have never been made public in any form.
To better understand the challenged posed by GLUE, we conduct experiments with simple baselines and state-of-the-art sentence representation models. We find that unified multi-task trained models slightly outperform comparable models trained on each task separately. Our best multi-task model makes use $\hat{\text{of ELMo (Peters et al., 2018), a recently proposed pre-training technique. However, this}}$ model still achieves a fairly low absolute score. Analysis with our diagnostic dataset reveals that our baseline models deal well with strong lexical signals but struggle with deeper logical structure.
In summary, we offer: (i) A suite of nine sentence or sentence-pair NLU tasks, built on established annotated datasets and selected to cover a diverse range of text genres, dataset sizes, and degrees of difficulty. (ii) An online evaluation platform and leaderboard, based primarily on privately-held test data. The platform is model-agnostic, and can evaluate any method capable of producing results on all nine tasks. (iii) An expert-constructed diagnostic evaluation dataset. (iv) Baseline results for several major existing approaches to sentence representation learning
## 2 RELATEDWORK
Collobert et al. (2011) used a multi-task model with a shared sentence understanding component to jointly learn POS tagging, chunking, named entity recognition, and semantic role labeling. More recent work has explored using labels from core NLP tasks to supervise training of lower levels of deep neural networks (Søgaard & Goldberg, 2016; Hashimoto et al., 2017) and automatically learning cross-task sharing mechanisms for multi-task learning (Ruder et al., 2017).
Beyond multi-task learning, much work in developing general NLU systems has focused on sentence-to-vector encoders (Le &Mikolov, 2014; Kiros et al., 2015, i.a.), leveraging unlabeled data (Hill et al., 2016; Peters et al., 2018), labeled data (Conneau& Kiela, 2018; McCann et al., 2017), and combinations of these (Collobert et al., 2011; Subramanian et al., 2018). In this line of work, a standard evaluation practice has emerged, recently codified as SentEval (Conneau et al., 2017; Conneau & Kiela, 2018). Like GLUE, SentEval relies on a set of existing classification tasks involving either one or two sentences as inputs. Unlike GLUE, SentEval only evaluates sentenceto-vector encoders, making it well-suited for evaluating models on tasks involving sentences $in$ $isolation.$ However, cross-sentence contextualization and alignment are instrumental in achieving state-of-the-art performance on tasks such as machine translation (Bahdanau et al., 2015; Vaswani et al., 2017), question answering (Seo et al., 2017), and natural language inference (Rocktäschel et al., 2016). GLUE is designed to facilitate the development of these methods: It is model-agnostic, allowing for any kind of representation or contextualization, including models that use no explicit vector or symbolic representations for sentences whatsoever.
GLUE also diverges from SentEval in the selection of evaluation tasks that are included in the suite. Many of the SentEval tasks are closely related to sentiment analysis, such as MR (Pang &Lee,
2
|
[
[
0.17236328125,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.17236328125,
0.04736328125
],
[
0.313720703125,
0.10479736328125,
0.3505859375,
0.10479736328125,
0.3505859375,
0.12060546875,
0.313720703125,
0.12060546875
],
[
0.363525390625,
0.10479736328125,
0.400390625,
0.10479736328125,
0.400390625,
0.12060546875,
0.363525390625,
0.12060546875
],
[
0.250732421875,
0.1060791015625,
0.295654296875,
0.1060791015625,
0.295654296875,
0.12060546875,
0.250732421875,
0.12060546875
],
[
0.181396484375,
0.1060791015625,
0.2344970703125,
0.1060791015625,
0.2344970703125,
0.12060546875,
0.181396484375,
0.12060546875
],
[
0.495849609375,
0.1060791015625,
0.54833984375,
0.1060791015625,
0.54833984375,
0.12060546875,
0.495849609375,
0.12060546875
],
[
0.69287109375,
0.1060791015625,
0.74658203125,
0.1060791015625,
0.74658203125,
0.12060546875,
0.69287109375,
0.12060546875
],
[
0.431396484375,
0.125,
0.56884765625,
0.125,
0.56884765625,
0.1395263671875,
0.431396484375,
0.1395263671875
],
[
0.1829833984375,
0.143310546875,
0.2254638671875,
0.143310546875,
0.2254638671875,
0.1590576171875,
0.1829833984375,
0.1590576171875
],
[
0.2646484375,
0.143310546875,
0.298095703125,
0.143310546875,
0.298095703125,
0.1590576171875,
0.2646484375,
0.1590576171875
],
[
0.330810546875,
0.143310546875,
0.351318359375,
0.143310546875,
0.351318359375,
0.1578369140625,
0.330810546875,
0.1578369140625
],
[
0.365234375,
0.14453125,
0.443603515625,
0.14453125,
0.443603515625,
0.1590576171875,
0.365234375,
0.1590576171875
],
[
0.494384765625,
0.143310546875,
0.58837890625,
0.143310546875,
0.58837890625,
0.1578369140625,
0.494384765625,
0.1578369140625
],
[
0.69384765625,
0.1458740234375,
0.72802734375,
0.1458740234375,
0.72802734375,
0.1578369140625,
0.69384765625,
0.1578369140625
],
[
0.1829833984375,
0.1558837890625,
0.2254638671875,
0.1558837890625,
0.2254638671875,
0.1717529296875,
0.1829833984375,
0.1717529296875
],
[
0.26953125,
0.1578369140625,
0.298095703125,
0.1578369140625,
0.298095703125,
0.17236328125,
0.26953125,
0.17236328125
],
[
0.3203125,
0.1572265625,
0.351318359375,
0.1572265625,
0.351318359375,
0.17236328125,
0.3203125,
0.17236328125
],
[
0.365234375,
0.1590576171875,
0.426513671875,
0.1590576171875,
0.426513671875,
0.17041015625,
0.365234375,
0.17041015625
],
[
0.495849609375,
0.1590576171875,
0.521484375,
0.1590576171875,
0.521484375,
0.1717529296875,
0.495849609375,
0.1717529296875
],
[
0.69287109375,
0.1578369140625,
0.779296875,
0.1578369140625,
0.779296875,
0.169189453125,
0.69287109375,
0.169189453125
],
[
0.40283203125,
0.1749267578125,
0.59619140625,
0.1749267578125,
0.59619140625,
0.189453125,
0.40283203125,
0.189453125
],
[
0.181396484375,
0.1932373046875,
0.2296142578125,
0.1932373046875,
0.2296142578125,
0.2083740234375,
0.181396484375,
0.2083740234375
],
[
0.265625,
0.1925048828125,
0.298095703125,
0.1925048828125,
0.298095703125,
0.208984375,
0.265625,
0.208984375
],
[
0.32177734375,
0.1944580078125,
0.3505859375,
0.1944580078125,
0.3505859375,
0.20703125,
0.32177734375,
0.20703125
],
[
0.365234375,
0.1956787109375,
0.431396484375,
0.1956787109375,
0.431396484375,
0.20703125,
0.365234375,
0.20703125
],
[
0.495849609375,
0.1932373046875,
0.54150390625,
0.1932373046875,
0.54150390625,
0.2083740234375,
0.495849609375,
0.2083740234375
],
[
0.69287109375,
0.1956787109375,
0.72802734375,
0.1956787109375,
0.72802734375,
0.2083740234375,
0.69287109375,
0.2083740234375
],
[
0.1829833984375,
0.205810546875,
0.2296142578125,
0.205810546875,
0.2296142578125,
0.2215576171875,
0.1829833984375,
0.2215576171875
],
[
0.27685546875,
0.20703125,
0.29736328125,
0.20703125,
0.29736328125,
0.2215576171875,
0.27685546875,
0.2215576171875
],
[
0.318603515625,
0.205810546875,
0.351318359375,
0.205810546875,
0.351318359375,
0.2215576171875,
0.318603515625,
0.2215576171875
],
[
0.363525390625,
0.20703125,
0.4794921875,
0.20703125,
0.4794921875,
0.2203369140625,
0.363525390625,
0.2203369140625
],
[
0.494384765625,
0.20703125,
0.63623046875,
0.20703125,
0.63623046875,
0.2203369140625,
0.494384765625,
0.2203369140625
],
[
0.69287109375,
0.2083740234375,
0.72802734375,
0.2083740234375,
0.72802734375,
0.2203369140625,
0.69287109375,
0.2203369140625
],
[
0.181396484375,
0.2183837890625,
0.2197265625,
0.2183837890625,
0.2197265625,
0.2342529296875,
0.181396484375,
0.2342529296875
],
[
0.261474609375,
0.2183837890625,
0.298095703125,
0.2183837890625,
0.298095703125,
0.2342529296875,
0.261474609375,
0.2342529296875
],
[
0.314453125,
0.2183837890625,
0.351318359375,
0.2183837890625,
0.351318359375,
0.2342529296875,
0.314453125,
0.2342529296875
],
[
0.494384765625,
0.2197265625,
0.54150390625,
0.2197265625,
0.54150390625,
0.2342529296875,
0.494384765625,
0.2342529296875
],
[
0.69140625,
0.2183837890625,
0.814453125,
0.2197265625,
0.814453125,
0.2342529296875,
0.69140625,
0.23291015625
],
[
0.450927734375,
0.238037109375,
0.54833984375,
0.238037109375,
0.54833984375,
0.25244140625,
0.450927734375,
0.25244140625
],
[
0.363525390625,
0.2215576171875,
0.4306640625,
0.2215576171875,
0.4306640625,
0.23291015625,
0.363525390625,
0.23291015625
],
[
0.181396484375,
0.2568359375,
0.226318359375,
0.2568359375,
0.226318359375,
0.271484375,
0.181396484375,
0.271484375
],
[
0.32177734375,
0.255615234375,
0.35302734375,
0.255615234375,
0.35302734375,
0.271484375,
0.32177734375,
0.271484375
],
[
0.365234375,
0.2568359375,
0.39453125,
0.2568359375,
0.39453125,
0.26953125,
0.365234375,
0.26953125
],
[
0.49267578125,
0.2568359375,
0.67724609375,
0.255615234375,
0.67724609375,
0.270751953125,
0.49267578125,
0.271484375
],
[
0.69287109375,
0.2587890625,
0.72802734375,
0.2587890625,
0.72802734375,
0.270751953125,
0.69287109375,
0.270751953125
],
[
0.181396484375,
0.26953125,
0.2254638671875,
0.26953125,
0.2254638671875,
0.2841796875,
0.181396484375,
0.2841796875
],
[
0.26318359375,
0.25830078125,
0.29736328125,
0.25830078125,
0.29736328125,
0.270751953125,
0.26318359375,
0.270751953125
],
[
0.26318359375,
0.270751953125,
0.29736328125,
0.270751953125,
0.29736328125,
0.28271484375,
0.26318359375,
0.28271484375
],
[
0.318603515625,
0.270751953125,
0.351318359375,
0.270751953125,
0.351318359375,
0.28271484375,
0.318603515625,
0.28271484375
],
[
0.363525390625,
0.26953125,
0.420654296875,
0.26953125,
0.420654296875,
0.2841796875,
0.363525390625,
0.2841796875
],
[
0.494384765625,
0.271484375,
0.52392578125,
0.271484375,
0.52392578125,
0.2841796875,
0.494384765625,
0.2841796875
],
[
0.69384765625,
0.271484375,
0.75634765625,
0.271484375,
0.75634765625,
0.28271484375,
0.69384765625,
0.28271484375
],
[
0.1829833984375,
0.28271484375,
0.2156982421875,
0.28271484375,
0.2156982421875,
0.29541015625,
0.1829833984375,
0.29541015625
],
[
0.265625,
0.28271484375,
0.298095703125,
0.28271484375,
0.298095703125,
0.29541015625,
0.265625,
0.29541015625
],
[
0.330078125,
0.28271484375,
0.351318359375,
0.28271484375,
0.351318359375,
0.296630859375,
0.330078125,
0.296630859375
],
[
0.365234375,
0.28271484375,
0.393798828125,
0.28271484375,
0.393798828125,
0.29541015625,
0.365234375,
0.29541015625
],
[
0.494384765625,
0.2841796875,
0.52294921875,
0.2841796875,
0.52294921875,
0.296630859375,
0.494384765625,
0.296630859375
],
[
0.69384765625,
0.2841796875,
0.79345703125,
0.2841796875,
0.79345703125,
0.29541015625,
0.69384765625,
0.29541015625
],
[
0.1829833984375,
0.294189453125,
0.2279052734375,
0.294189453125,
0.2279052734375,
0.309326171875,
0.1829833984375,
0.309326171875
],
[
0.268798828125,
0.294189453125,
0.2998046875,
0.294189453125,
0.2998046875,
0.31005859375,
0.268798828125,
0.31005859375
],
[
0.324462890625,
0.29541015625,
0.3505859375,
0.29541015625,
0.3505859375,
0.309326171875,
0.324462890625,
0.309326171875
],
[
0.365234375,
0.296630859375,
0.465576171875,
0.296630859375,
0.465576171875,
0.30810546875,
0.365234375,
0.30810546875
],
[
0.494384765625,
0.296630859375,
0.52294921875,
0.296630859375,
0.52294921875,
0.309326171875,
0.494384765625,
0.309326171875
],
[
0.69287109375,
0.29541015625,
0.771484375,
0.29541015625,
0.771484375,
0.306884765625,
0.69287109375,
0.306884765625
],
[
0.1732177734375,
0.323974609375,
0.82421875,
0.323974609375,
0.82421875,
0.337158203125,
0.1732177734375,
0.337158203125
],
[
0.1748046875,
0.33837890625,
0.82421875,
0.33837890625,
0.82421875,
0.3515625,
0.1748046875,
0.3515625
],
[
0.1732177734375,
0.3515625,
0.73388671875,
0.3515625,
0.73388671875,
0.365478515625,
0.1732177734375,
0.365478515625
],
[
0.1748046875,
0.392578125,
0.82421875,
0.392578125,
0.82421875,
0.406005859375,
0.1748046875,
0.406005859375
],
[
0.1748046875,
0.4072265625,
0.82421875,
0.4072265625,
0.82421875,
0.42041015625,
0.1748046875,
0.42041015625
],
[
0.1732177734375,
0.42041015625,
0.82421875,
0.42041015625,
0.82421875,
0.434326171875,
0.1732177734375,
0.434326171875
],
[
0.1732177734375,
0.434326171875,
0.82421875,
0.434326171875,
0.82421875,
0.447509765625,
0.1732177734375,
0.447509765625
],
[
0.1748046875,
0.447509765625,
0.82421875,
0.447509765625,
0.82421875,
0.4609375,
0.1748046875,
0.4609375
],
[
0.1732177734375,
0.462158203125,
0.79345703125,
0.462158203125,
0.79345703125,
0.47607421875,
0.1732177734375,
0.47607421875
],
[
0.1732177734375,
0.48291015625,
0.82421875,
0.48291015625,
0.82421875,
0.496826171875,
0.1732177734375,
0.496826171875
],
[
0.17236328125,
0.49560546875,
0.82421875,
0.496826171875,
0.82421875,
0.51123046875,
0.17236328125,
0.51025390625
],
[
0.1732177734375,
0.51123046875,
0.82421875,
0.51123046875,
0.82421875,
0.5244140625,
0.1732177734375,
0.5244140625
],
[
0.1748046875,
0.5263671875,
0.82275390625,
0.5263671875,
0.82275390625,
0.537109375,
0.1748046875,
0.537109375
],
[
0.1732177734375,
0.5390625,
0.82421875,
0.5390625,
0.82421875,
0.55322265625,
0.1732177734375,
0.55322265625
],
[
0.17236328125,
0.5517578125,
0.63134765625,
0.55322265625,
0.63134765625,
0.5673828125,
0.17236328125,
0.56640625
],
[
0.176513671875,
0.58837890625,
0.188720703125,
0.58837890625,
0.188720703125,
0.599609375,
0.176513671875,
0.599609375
],
[
0.205078125,
0.58837890625,
0.346435546875,
0.58837890625,
0.346435546875,
0.6015625,
0.205078125,
0.6015625
],
[
0.1732177734375,
0.61865234375,
0.82421875,
0.61865234375,
0.82421875,
0.6318359375,
0.1732177734375,
0.6318359375
],
[
0.1732177734375,
0.63330078125,
0.82421875,
0.63330078125,
0.82421875,
0.64794921875,
0.1732177734375,
0.64794921875
],
[
0.1732177734375,
0.646484375,
0.82421875,
0.646484375,
0.82421875,
0.66015625,
0.1732177734375,
0.66015625
],
[
0.1732177734375,
0.66015625,
0.82275390625,
0.66015625,
0.82275390625,
0.6748046875,
0.1732177734375,
0.6748046875
],
[
0.1748046875,
0.6748046875,
0.72216796875,
0.6748046875,
0.72216796875,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1748046875,
0.69580078125,
0.82421875,
0.69580078125,
0.82421875,
0.71044921875,
0.1748046875,
0.71044921875
],
[
0.1732177734375,
0.71044921875,
0.82421875,
0.71044921875,
0.82421875,
0.7236328125,
0.1732177734375,
0.7236328125
],
[
0.1748046875,
0.7236328125,
0.82421875,
0.7236328125,
0.82421875,
0.7373046875,
0.1748046875,
0.7373046875
],
[
0.1748046875,
0.7373046875,
0.82421875,
0.7373046875,
0.82421875,
0.75048828125,
0.1748046875,
0.75048828125
],
[
0.1732177734375,
0.75048828125,
0.82275390625,
0.75048828125,
0.82275390625,
0.7646484375,
0.1732177734375,
0.7646484375
],
[
0.1748046875,
0.76513671875,
0.826171875,
0.76513671875,
0.826171875,
0.77880859375,
0.1748046875,
0.77880859375
],
[
0.1732177734375,
0.77783203125,
0.82177734375,
0.77783203125,
0.82177734375,
0.79150390625,
0.1732177734375,
0.79150390625
],
[
0.1732177734375,
0.79248046875,
0.82421875,
0.79248046875,
0.82421875,
0.80615234375,
0.1732177734375,
0.80615234375
],
[
0.1732177734375,
0.806640625,
0.82421875,
0.806640625,
0.82421875,
0.82080078125,
0.1732177734375,
0.82080078125
],
[
0.1732177734375,
0.82080078125,
0.82421875,
0.82080078125,
0.82421875,
0.833984375,
0.1732177734375,
0.833984375
],
[
0.1748046875,
0.833984375,
0.82421875,
0.833984375,
0.82421875,
0.84765625,
0.1748046875,
0.84765625
],
[
0.1732177734375,
0.8466796875,
0.82275390625,
0.84765625,
0.82275390625,
0.8623046875,
0.1732177734375,
0.861328125
],
[
0.1732177734375,
0.861328125,
0.82421875,
0.8623046875,
0.82421875,
0.876953125,
0.1732177734375,
0.87548828125
],
[
0.176513671875,
0.87744140625,
0.5703125,
0.87744140625,
0.5703125,
0.88818359375,
0.176513671875,
0.88818359375
],
[
0.1748046875,
0.896484375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1732177734375,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
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 conferencepaper atICLR2019",
"[Test]",
"Task",
"[Train]",
"Corpus",
"Metrics",
"Domain",
"Single-Sentence Tasks",
"CoLA",
"8.5k",
"1k",
"acceptability",
"Matthews corr.",
"misc.",
"SST-2",
"67k",
"1.8k",
"sentiment",
"acc.",
"moviereviews",
"Similarity and Paraphrase Tasks",
"MRPC",
"3.7k",
"1.7k",
"paraphrase",
"acc./F1",
"news",
"STS-B",
"7k",
"1.4k",
"sentence similarity",
"Pearson/Spearman corr.",
"misc.",
"QQP",
"364k",
"391k",
"acc./F1",
"social QA questions",
"Inference Tasks",
"paraphrase",
"MNLI",
"20k",
"NLI",
"matched acc./mismatched acc.",
"misc.",
"QNLI",
"393k",
"105k",
"5.4k",
"QA/NLI",
"acc.",
"Wikipedia",
"RTE",
"2.5k",
"3k",
"NLI",
"acc.",
"news, Wikipedia",
"WNLI",
"634",
"146",
"coreference/NLI",
"acc.",
"fictionbooks",
"Table l: Task descriptions and statistics. All tasks are single sentence or sentence pair classification,",
"except STS-B, which is a regression task. MNLI has three classes; all other classification tasks have",
"two. Test sets shown in bold use labels that have never been made public in any form.",
"To better understand the challenged posed by GLUE, we conduct experiments with simple baselines",
"and state-of-the-art sentence representation models. We find that unified multi-task trained models",
"slightly outperform comparable models trained on each task separately. Our best multi-task model",
"makes use of ELMo (Peters et al., 2018), a recently proposed pre-training technique. However, this",
"model still achieves a fairly low absolute score. Analysis with our diagnostic dataset reveals that our",
"baseline models deal well with strong lexical signals but struggle with deeper logical structure.",
"In summary, we offer: (i) A suite of nine sentence or sentence-pair NLU tasks, built on established",
"annotated datasets and selected to cover a diverse range of text genres, dataset sizes, and degrees",
"of difficulty. (ii) An online evaluation platform and leaderboard, based primarily on privately-held",
"test data.Theplatformis model-agnostic,and can evaluateanymethod capable ofproducingresults",
"on all nine tasks. (i) An expert-constructed diagnostic evaluation dataset. (iv) Baseline results for",
"several major existing approaches to sentence representation learning.",
"2",
"RELATED WORK",
"Collobert et al. (2011) used a multi-task model with a shared sentence understanding component to",
"jointly learn POS tagging, chunking, named entity recognition, and semantic role labeling. More",
"recent work has explored using labels from core NLP tasks to supervise training of lower levels",
"of deep neural networks (Sogaard & Goldberg, 2016; Hashimoto et al., 2017) and automatically",
"learning cross-task sharing mechanisms for multi-task learning (Ruder et al., 2017).",
"Beyond multi-task learning, much work in developing general NLU systems has focused on",
"sentence-to-vector encoders (Le & Mikolov, 2014; Kiros et al., 2015, i.a.), leveraging unlabeled",
"data (Hill et al., 2016; Peters et al., 2018), labeled data (Conneau & Kiela, 2018; McCann et al.,",
"2017), and combinations of these (Collobert et al., 2011; Subramanian et al., 2018). In this line",
"of work, a standard evaluation practice has emerged, recently codified as SentEval (Conneau et al.,",
"2017; Conneau & Kiela, 2018). Like GLUE, SentEval relies on a set of existing classification tasks",
"involving either one or two sentences as inputs. Unlike GLUE, SentEval only evaluates sentence-",
"to-vector encoders, making it well-suited for evaluating models on tasks involving sentences in",
"isolation. However, cross-sentence contextualization and alignment are instrumental in achieving",
"state-of-the-art performance on tasks such as machine translation (Bahdanau et al., 2015; Vaswani",
"et al., 2017), question answering (Seo et al., 2017), and natural language inference (Rocktaschel",
"et al., 2016). GLUE is designed to facilitate the development of these methods: It is model-agnostic,",
"allowing for any kind of representation or contextualization, including models that use no explicit",
"vectororsymbolicrepresentationsforsentenceswhatsoever.",
"GLUE also diverges from SentEval in the selection of evaluation tasks that are included in the suite.",
"Many of the SentEval tasks are closely related to sentiment analysis, such as MR (Pang & Lee,",
"2"
] |
{
"arxiv_id": "1804.07461",
"language": "en",
"page_idx": 1
}
|
||
1804.07461-en-2
|
1804.07461-en-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2019}}$
2005), SST (Socher et al., 2013), CR (Hu&Liu, 2004), and SUBJ (Pang &Lee, 2004). Other tasks are so close to being solved that evaluation on them is relatively uninformative, such as MPQA (Wiebe et al., 2005) and TREC question classification (Voorhees et al., 1999). In GLUE, we attempt to construct a benchmark that is both diverse and difficult.
McCann et al. (2018) introduce decaNLP, which also scores NLP systems based on their performance on multiple datasets. Their benchmark recasts the ten evaluation tasks as question answering, converting tasks like summarization and text-to-SQL semantic parsing into question answering using automatic transformations. That benchmark lacks the leaderboard and error analysis toolkit of GLUE, but more importantly, we see it as pursuing a more ambitious but less immediately practical goal: While GLUE rewards methods that yield good performance on a circumscribed set of tasks using methods like those that are currently used for those tasks, their benchmark rewards systems that make progress toward their goal of unifying all of NLU under the rubric of question answering
# 3 TASKS
GLUE is centered on nine English sentence understanding tasks, which cover a broad range of domains, data quantities, and difficulties. As the goal of GLUE is to spur development of generalizable NLU systems, we design the benchmark such that good performance should require a model to share substantial knowledge (e.g., trained parameters) across all tasks, while still maintaining some taskspecific components. Though it is possible to train a single model for each task with no pretraining or other outside sources of knowledge and evaluate the resulting set of models on this benchmark we expect that our inclusion of several data-scarce tasks will ultimately render this approach uncompetitive. We describe the tasks below and in Table 1. Appendix A includes additional details. Unless otherwise mentioned, tasks are evaluated on accuracy and are balanced across classes.
# 3.1 SINGLE-SENTENCE TASKS
$\mathbf{CoLA}$ The Corpus of Linguistic Acceptability (Warstadt et al., 2018) consists of English acceptability judgments drawn from books and journal articles on linguistic theory. Each example is a sequence of words annotated with whether it is a grammatical English sentence. Following the authors, we use Matthews correlation coefficient (Matthews, 1975) as the evaluation metric, which evaluates performance on unbalanced binary classification and ranges from -1 to 1, with 0 being the performance of uninformed guessing. We use the standard test set, for which we obtained private labels from the authors. We report a single performance number on the combination of the in- and out-of-domain sections of the test set.
SST-2 The Stanford Sentiment Treebank (Socher et al., 2013) consists of sentences from movie reviews and human annotations of their sentiment. The task is to predict the sentiment of a given sentence. We use the two-way $(positivelnegative)$ class split, and use only sentence-level labels.
3.2 SIMILARITY AND PARAPHRASE TASKS
MRPC The Microsoft Research Paraphrase Corpus (Dolan&Brockett,2005) is a corpus of sentence pairs automatically extracted from online news sources, with human annotations for whether the sentences in the pair are semantically equivalent. Because the classes are imbalanced (68% positive), we follow common practice and report both accuracy and F1 score.
$\mathbf{QQP}$ The Quora Question Pairs$^2$ dataset is a collection of question pairs from the community question-answering website Quora. The task is to determine whether a pair of questions are semantically equivalent. As in MRPC, the class distribution in QQP is unbalanced (63% negative), so we report both accuracy and F1 score. We use the standard test set, for which we obtained private labels from the authors. We observe that the test set has a different label distribution than the training set.
STS-B The Semantic Textual Similarity Benchmark (Cer et al., 2017) is a collection of sentence pairs drawn from news headlines, video and image captions, and natural language inference data.
$^{2}$data.quora.com/First-Quora-Dataset-Release-Question-Pairs
3
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11871337890625,
0.1748046875,
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
],
[
0.1732177734375,
0.146484375,
0.5537109375,
0.146484375,
0.5537109375,
0.160400390625,
0.1732177734375,
0.160400390625
],
[
0.1732177734375,
0.1666259765625,
0.82177734375,
0.1673583984375,
0.82177734375,
0.1817626953125,
0.1732177734375,
0.18115234375
],
[
0.1748046875,
0.1817626953125,
0.82177734375,
0.1817626953125,
0.82177734375,
0.1956787109375,
0.1748046875,
0.1956787109375
],
[
0.176513671875,
0.1976318359375,
0.8203125,
0.1976318359375,
0.8203125,
0.2083740234375,
0.176513671875,
0.2083740234375
],
[
0.1748046875,
0.210205078125,
0.82421875,
0.210205078125,
0.82421875,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1732177734375,
0.2235107421875,
0.82177734375,
0.2235107421875,
0.82177734375,
0.2374267578125,
0.1732177734375,
0.2374267578125
],
[
0.1748046875,
0.2374267578125,
0.82421875,
0.2374267578125,
0.82421875,
0.250732421875,
0.1748046875,
0.250732421875
],
[
0.1748046875,
0.251953125,
0.82275390625,
0.251953125,
0.82275390625,
0.26513671875,
0.1748046875,
0.26513671875
],
[
0.1732177734375,
0.26513671875,
0.82177734375,
0.26513671875,
0.82177734375,
0.279052734375,
0.1732177734375,
0.279052734375
],
[
0.1732177734375,
0.2998046875,
0.19287109375,
0.2998046875,
0.19287109375,
0.313232421875,
0.1732177734375,
0.313232421875
],
[
0.19775390625,
0.2998046875,
0.2646484375,
0.2998046875,
0.2646484375,
0.314453125,
0.19775390625,
0.314453125
],
[
0.1748046875,
0.330810546875,
0.82177734375,
0.330810546875,
0.82177734375,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.176513671875,
0.347900390625,
0.82177734375,
0.347900390625,
0.82177734375,
0.35791015625,
0.176513671875,
0.35791015625
],
[
0.1748046875,
0.35986328125,
0.82275390625,
0.35986328125,
0.82275390625,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.1748046875,
0.375732421875,
0.8203125,
0.375732421875,
0.8203125,
0.386474609375,
0.1748046875,
0.386474609375
],
[
0.1732177734375,
0.38818359375,
0.82177734375,
0.386962890625,
0.82177734375,
0.40087890625,
0.1732177734375,
0.401611328125
],
[
0.1748046875,
0.401611328125,
0.82275390625,
0.401611328125,
0.82275390625,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.1748046875,
0.417236328125,
0.81884765625,
0.417236328125,
0.81884765625,
0.427978515625,
0.1748046875,
0.427978515625
],
[
0.1748046875,
0.429931640625,
0.82177734375,
0.429931640625,
0.82177734375,
0.443115234375,
0.1748046875,
0.443115234375
],
[
0.17236328125,
0.443115234375,
0.78369140625,
0.443115234375,
0.78369140625,
0.45703125,
0.17236328125,
0.45703125
],
[
0.1748046875,
0.474853515625,
0.2099609375,
0.474853515625,
0.2099609375,
0.4873046875,
0.1748046875,
0.4873046875
],
[
0.2091064453125,
0.474853515625,
0.398681640625,
0.474853515625,
0.398681640625,
0.488037109375,
0.2091064453125,
0.488037109375
],
[
0.1732177734375,
0.5,
0.82177734375,
0.50048828125,
0.82177734375,
0.51513671875,
0.1732177734375,
0.5146484375
],
[
0.1748046875,
0.51513671875,
0.82421875,
0.51513671875,
0.82421875,
0.52880859375,
0.1748046875,
0.52880859375
],
[
0.1748046875,
0.52880859375,
0.82177734375,
0.52880859375,
0.82177734375,
0.54248046875,
0.1748046875,
0.54248046875
],
[
0.1748046875,
0.54248046875,
0.82275390625,
0.54248046875,
0.82275390625,
0.556640625,
0.1748046875,
0.556640625
],
[
0.1732177734375,
0.55615234375,
0.82275390625,
0.5546875,
0.82275390625,
0.5693359375,
0.1732177734375,
0.57080078125
],
[
0.1748046875,
0.57080078125,
0.82275390625,
0.57080078125,
0.82275390625,
0.583984375,
0.1748046875,
0.583984375
],
[
0.1732177734375,
0.583984375,
0.82421875,
0.583984375,
0.82421875,
0.59765625,
0.1732177734375,
0.59765625
],
[
0.17236328125,
0.59765625,
0.42236328125,
0.5986328125,
0.42236328125,
0.6123046875,
0.17236328125,
0.611328125
],
[
0.1732177734375,
0.62548828125,
0.82275390625,
0.62744140625,
0.82275390625,
0.642578125,
0.1732177734375,
0.64013671875
],
[
0.1748046875,
0.642578125,
0.82275390625,
0.642578125,
0.82275390625,
0.65576171875,
0.1748046875,
0.65576171875
],
[
0.1732177734375,
0.65576171875,
0.79833984375,
0.65576171875,
0.79833984375,
0.66943359375,
0.1732177734375,
0.66943359375
],
[
0.1748046875,
0.6875,
0.484375,
0.6875,
0.484375,
0.70068359375,
0.1748046875,
0.70068359375
],
[
0.17236328125,
0.71044921875,
0.82275390625,
0.7109375,
0.82275390625,
0.72900390625,
0.17236328125,
0.72802734375
],
[
0.1748046875,
0.72802734375,
0.82177734375,
0.72802734375,
0.82177734375,
0.7412109375,
0.1748046875,
0.7412109375
],
[
0.17236328125,
0.7412109375,
0.8203125,
0.7412109375,
0.8203125,
0.7548828125,
0.17236328125,
0.7548828125
],
[
0.1732177734375,
0.755859375,
0.6787109375,
0.755859375,
0.6787109375,
0.76953125,
0.1732177734375,
0.76953125
],
[
0.176513671875,
0.78515625,
0.82177734375,
0.78515625,
0.82177734375,
0.798828125,
0.176513671875,
0.798828125
],
[
0.176513671875,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.176513671875,
0.81298828125
],
[
0.1732177734375,
0.8125,
0.82275390625,
0.81298828125,
0.82275390625,
0.82763671875,
0.1732177734375,
0.8271484375
],
[
0.1748046875,
0.82763671875,
0.82275390625,
0.82763671875,
0.82275390625,
0.84130859375,
0.1748046875,
0.84130859375
],
[
0.1732177734375,
0.84130859375,
0.8203125,
0.84130859375,
0.8203125,
0.85498046875,
0.1732177734375,
0.85498046875
],
[
0.176513671875,
0.87255859375,
0.2279052734375,
0.87255859375,
0.2279052734375,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.2279052734375,
0.8720703125,
0.82275390625,
0.8720703125,
0.82275390625,
0.88525390625,
0.2279052734375,
0.88525390625
],
[
0.1748046875,
0.88623046875,
0.82275390625,
0.88623046875,
0.82275390625,
0.8994140625,
0.1748046875,
0.8994140625
],
[
0.193603515625,
0.9091796875,
0.708984375,
0.91015625,
0.708984375,
0.9248046875,
0.193603515625,
0.923828125
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2019",
"2005), SST (Socher et al., 2013), CR (Hu & Liu, 2004), and SUBJ (Pang & Lee, 2004). Other",
"tasksaresoclosetobeing solvedthatevaluationonthemisrelativelyuninformative,suchasMPOA",
"(Wiebe et al., 2005)and TREC question classification (Voorhees et al., 1999). In GLUE, we attempt",
"to construct a benchmark that is both diverse and difficult.",
"McCann et al. (2018) introduce decaNLP, which also scores NLP systems based on their perfor",
"mance on multiple datasets. Their benchmark recasts the ten evaluation tasks as question answer-",
"ing,convertingtaskslike summarization and text-to-SQL semanticparsing into question answering",
"using automatic transformations.That benchmark lacks the leaderboard and error analysis toolkit of",
"GLUE, but more importantly, we see it as pursuing a more ambitious but less immediately practica",
"goal: While GLUE rewards methods that yield good performance on a circumscribed set of tasks",
"using methods like those that are currently used for those tasks, their benchmark rewards systems",
"that make progress toward their goal of unifying all of NLU under the rubric of question answering",
"3",
"TASKS",
"GLUE is centered on nine English sentence understanding tasks, which cover a broad range of do-",
"mains,dataquantities,anddifficulties.As thegoal of GLUEistospurdevelopmentof generalizable",
"NLU systems, we design the benchmark such that good performance should require a model to share",
"substantialknowledge (e.g.,trainedparameters)acrossalltasks,whilestillmaintaining sometask",
"specific components.Though it is possibletotraina singlemodel for each task with nopretraining",
"or other outside sources of knowledge and evaluate the resulting set of models on this benchmark",
"weexpectthat our inclusionof several data-scarcetasks will ultimatelyrenderthis approachun",
"competitive. We describe the tasks below and in Table l. Appendix A includes additional details",
"Unless otherwise mentioned, tasks are evaluated on accuracy and are balanced across classes.",
"3.1",
"SINGLE-SENTENCE TASKS",
"CoLAThe Corpus of Linguistic Acceptability (Warstadt et al., 2018) consists of English accept",
"ability judgments drawn from books and journal articles on linguistic theory. Each example is a",
"sequence of words annotated with whether it is a grammatical English sentence. Following the au-",
"thors, we use Matthews correlation coefficient (Matthews, 1975) as the evaluation metric, which",
"evaluates performance on unbalanced binary classification and ranges from -1 to 1, with o being the",
"performance of uninformed guessing. We use the standard test set, for which we obtained private",
"labels from the authors. We report a single performance number on the combination of the in- and",
"out-of-domainsectionsofthetestset.",
"SST-2The Stanford Sentiment Treebank (Socher et al.,2013)consists of sentences from movie",
"reviews and human annotations of their sentiment. The task is to predict the sentiment of a given",
"sentence. We use the two-way (positivelnegative) class split, and use only sentence-level labels.",
"3.2SIMILARITYANDPARAPHRASETASKS",
"MRPCThe Microsoft Research Paraphrase Corpus (Dolan & Brockett, 2005) is a corpus of sen-",
"tence pairs automatically extracted from online news sources, with human annotations for whether",
"the sentences in the pair are semantically equivalent. Because the classes are imbalanced (68%",
"positive), we follow common practice and report both accuracy and F1 score.",
"QQPThe Quora Question Pairs? dataset is a collection of question pairs from the community",
"question-answering website Quora. The task is to determine whether a pair of questions are seman-",
"tically equivalent. As in MRPC, the class distribution in QQP is unbalanced (63% negative), so we",
"report both accuracy and F1 score. We use the standard test set, for which we obtained private labels",
"from the authors. We observe that the test set has a different label distribution than the training set.",
"STS-B",
"The Semantic Textual Similarity Benchmark (Cer et al., 2017) is a collection of sentence",
"pairs drawn from news headlines, video and image captions, and natural language inference data",
"2data.guora.com/First-Quora-Dataset-Release-Question-Pairs",
"3"
] |
{
"arxiv_id": "1804.07461",
"language": "en",
"page_idx": 2
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.