id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2006.09902-en-0
|
2006.09902-en-0.jpg
|
$ଟ୍ରି$
# Vision-Aided Dynamic Blockage Prediction for 6G Wireless Communication Networks
Gouranga Charan, Muhammad Alrabeiah, and Ahmed Alkhateeb
Arizona State University, Emails: $\{$gcharan, malrabei, alkhateeb$\}@$asu.edu
$Abstract\_$Unlocking the full potential of millimeter-wave and sub-terahertz wireless commu ing unprecedented low-latency and high-reliability requirements. The challenge in meeting tho
frequency ranges to blockages. One promising way to tackle that challenge is to help a wireless network develop a sense of its surrounding using machine le that by utilizing deep learning and computer vision. It proposes
$\begin{array}{lllllll}&\text{a novei solution that proactively p}\end{array}$
forming vectors how to predict possible future link blockages. The proposed architecture is eva that represents a synthetic dynamic communication scenario with multiple moving users and bl prediction accuracy in the neighborhood of 86%, a performance that is unlikely to be match
Index Terms—Deep learning, computer vision, mmWave, ter-
ahertz, blockage prediction.
## I. INTRODUCTION
problem. It studies proactive blockage prediction for a singlemoving mmWave user in the presence of stationary blockages. Again, despite its appeal, it still falls short in meeting the In this paper and inspired by the recently proposed VisionAided Wireless Communication (VAWC) framework in [6] and [7], the proactive link-blockage problem in fully dynamic environments is addressed using deep learning and a fusion of The empirical evidence in [6] supports that notion; a deep learning model is able to perform simple mmWave beam and blockage prediction tasks using visual data. As such, a fusion of visual and wireless data is viewed as the right type of sensory data that could facilitate the learning of the more complex task of proactive link-blockage prediction. Based on this motivation, we develop a novel deep learning architecture to utilize sequences of observed RGB images (frames) and beamforming vectors and learn proactive link-blockage prediction. The proposed architecture harnesses the power of Convolutional Neural networks (CNNs) and Recurrent Neural
future wireless networks [1],[2]. With their large bandwidth, Networks (RNNs).The former excels in performing visual
tasks such as image classification [8] and object detection [9] while the later defines the stat-of-the-art in sequence modeling, as evident by the performance of RNNs in speech recognition [10],[11]. Experimental results show that the
physical and network layers. One of the key challenges stems proposed architecture is capable of learning the prediction task
effectively, achieving a noticeable gain over the models that
rely solely on wireless data.
The rest of this paper is organized as follows. Section
base stations and users. The possibility of blocking these LOS II introduces the system and channel models. Section III links by stationary or dynamic blockages can highly affect the provides a formal description of the proactive link-blockage
prediction problem. Following the problem definition, Section IV presents a detailed description of the proposed architecture, while Section V details the experimental setup used to evaluate the performance of the architecture. The performance evaluation, then, is presented in Section VI. Finally, Section VII wraps up the paper with some concluding remarks.
II. SYSTEM AND CHANNEL MODELS
Millimeter-wave (mmWave) and sub-terahertz communications are becoming the dominant directions for modern and they have the ability to satisfy the high data rate demands of several applications such as wireless Virtual/Augmented Reality (VR/AR) and autonomous driving. Communication in these bands, however, faces several challenges at both the from the sensitivity of mmWa tion to blockages. This requires these systems to rely heavily on maintaining line-of-sight reliability and latency of mmWave/terahertz systems
The key for overcoming the link blockage challenges lies in enabling the wireless sys surrounding. This could be accomplished using machine learning. Many recently published wireless sensory data (e.g., machine learning model can efficiently differentiate LOS and Non-LOS (NLOS) links [3],[4] address the link blockage problem form a reactive perspective; the sensory data is frst acq current link is predicted. Despite its appeal, the reactive linkblockage prediction does not reliability requirements of future wireless networks. The work in [5] takes a step forward towards a proactive treatment of the
To illustrate the potential of deep learning and VAWC in mitigating the link blockage problem, this work considers a high-frequency communication system where the basestation utilizes an RGB camera to monitor its environment. The following two subsections provide a brief description of that system and its adopted channel model.
|
[
[
0.12091064453125,
0.0662841796875,
0.88134765625,
0.06817626953125,
0.88134765625,
0.0921630859375,
0.12091064453125,
0.09027099609375
],
[
0.1854248046875,
0.1029052734375,
0.81201171875,
0.1029052734375,
0.81201171875,
0.126953125,
0.1854248046875,
0.126953125
],
[
0.2607421875,
0.1514892578125,
0.740234375,
0.1514892578125,
0.740234375,
0.1666259765625,
0.2607421875,
0.1666259765625
],
[
0.2254638671875,
0.17041015625,
0.77294921875,
0.17041015625,
0.77294921875,
0.184326171875,
0.2254638671875,
0.184326171875
],
[
0.0931396484375,
0.2139892578125,
0.494384765625,
0.2139892578125,
0.494384765625,
0.2279052734375,
0.0931396484375,
0.2279052734375
],
[
0.5009765625,
0.2139892578125,
0.9189453125,
0.2139892578125,
0.9189453125,
0.2279052734375,
0.5009765625,
0.2279052734375
],
[
0.080078125,
0.2291259765625,
0.494384765625,
0.2291259765625,
0.494384765625,
0.2392578125,
0.080078125,
0.2392578125
],
[
0.50341796875,
0.2298583984375,
0.91845703125,
0.2298583984375,
0.91845703125,
0.24365234375,
0.50341796875,
0.24365234375
],
[
0.0767822265625,
0.2392578125,
0.494384765625,
0.240478515625,
0.494384765625,
0.25390625,
0.0767822265625,
0.25244140625
],
[
0.50634765625,
0.2442626953125,
0.9208984375,
0.24365234375,
0.9208984375,
0.25830078125,
0.50634765625,
0.2587890625
],
[
0.080078125,
0.25390625,
0.491943359375,
0.25390625,
0.491943359375,
0.264404296875,
0.080078125,
0.264404296875
],
[
0.0767822265625,
0.26513671875,
0.49267578125,
0.26513671875,
0.49267578125,
0.279052734375,
0.0767822265625,
0.279052734375
],
[
0.5048828125,
0.260009765625,
0.74853515625,
0.260009765625,
0.74853515625,
0.2734375,
0.5048828125,
0.2734375
],
[
0.02532958984375,
0.26953125,
0.056365966796875,
0.26953125,
0.056365966796875,
0.322509765625,
0.02532958984375,
0.322509765625
],
[
0.080078125,
0.279052734375,
0.491943359375,
0.279052734375,
0.491943359375,
0.2890625,
0.080078125,
0.2890625
],
[
0.52294921875,
0.2734375,
0.9189453125,
0.2734375,
0.9189453125,
0.28857421875,
0.52294921875,
0.28857421875
],
[
0.07843017578125,
0.291748046875,
0.49267578125,
0.291748046875,
0.49267578125,
0.3017578125,
0.07843017578125,
0.3017578125
],
[
0.5048828125,
0.28857421875,
0.9208984375,
0.2890625,
0.9208984375,
0.3037109375,
0.5048828125,
0.302978515625
],
[
0.08087158203125,
0.30615234375,
0.490234375,
0.30615234375,
0.490234375,
0.313232421875,
0.08087158203125,
0.313232421875
],
[
0.50634765625,
0.30615234375,
0.9208984375,
0.30615234375,
0.9208984375,
0.3193359375,
0.50634765625,
0.3193359375
],
[
0.07843017578125,
0.31494140625,
0.49267578125,
0.31494140625,
0.49267578125,
0.328857421875,
0.07843017578125,
0.328857421875
],
[
0.03350830078125,
0.323974609375,
0.053924560546875,
0.323974609375,
0.053924560546875,
0.356689453125,
0.03350830078125,
0.356689453125
],
[
0.50634765625,
0.32080078125,
0.9208984375,
0.32080078125,
0.9208984375,
0.333984375,
0.50634765625,
0.333984375
],
[
0.08087158203125,
0.330810546875,
0.490234375,
0.330810546875,
0.490234375,
0.33837890625,
0.08087158203125,
0.33837890625
],
[
0.0767822265625,
0.339111328125,
0.49267578125,
0.34033203125,
0.49267578125,
0.354248046875,
0.0767822265625,
0.352783203125
],
[
0.50634765625,
0.335205078125,
0.9189453125,
0.335205078125,
0.9189453125,
0.348388671875,
0.50634765625,
0.348388671875
],
[
0.08087158203125,
0.35595703125,
0.490234375,
0.35595703125,
0.490234375,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.50634765625,
0.35107421875,
0.9189453125,
0.35107421875,
0.9189453125,
0.3642578125,
0.50634765625,
0.3642578125
],
[
0.02777099609375,
0.365478515625,
0.053924560546875,
0.365478515625,
0.053924560546875,
0.385009765625,
0.02777099609375,
0.385009765625
],
[
0.039215087890625,
0.356689453125,
0.0482177734375,
0.356689453125,
0.0482177734375,
0.35986328125,
0.039215087890625,
0.35986328125
],
[
0.07843017578125,
0.367431640625,
0.49267578125,
0.367431640625,
0.49267578125,
0.37744140625,
0.07843017578125,
0.37744140625
],
[
0.50634765625,
0.365478515625,
0.9189453125,
0.365478515625,
0.9189453125,
0.37890625,
0.50634765625,
0.37890625
],
[
0.080078125,
0.380126953125,
0.49267578125,
0.380126953125,
0.49267578125,
0.39013671875,
0.080078125,
0.39013671875
],
[
0.5048828125,
0.380615234375,
0.9208984375,
0.380615234375,
0.9208984375,
0.39453125,
0.5048828125,
0.39453125
],
[
0.080078125,
0.392578125,
0.49267578125,
0.392578125,
0.49267578125,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.50634765625,
0.395751953125,
0.9208984375,
0.395751953125,
0.9208984375,
0.4091796875,
0.50634765625,
0.4091796875
],
[
0.0269622802734375,
0.4072265625,
0.052276611328125,
0.4072265625,
0.052276611328125,
0.449462890625,
0.0269622802734375,
0.449462890625
],
[
0.080078125,
0.40478515625,
0.48779296875,
0.40478515625,
0.48779296875,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.07843017578125,
0.416748046875,
0.49267578125,
0.416748046875,
0.49267578125,
0.429931640625,
0.07843017578125,
0.429931640625
],
[
0.50634765625,
0.410888671875,
0.9208984375,
0.410888671875,
0.9208984375,
0.4248046875,
0.50634765625,
0.4248046875
],
[
0.080078125,
0.429931640625,
0.464111328125,
0.429931640625,
0.464111328125,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50732421875,
0.425537109375,
0.9208984375,
0.425537109375,
0.9208984375,
0.439453125,
0.50732421875,
0.439453125
],
[
0.09234619140625,
0.442626953125,
0.49267578125,
0.442626953125,
0.49267578125,
0.455810546875,
0.09234619140625,
0.455810546875
],
[
0.50634765625,
0.44140625,
0.9208984375,
0.44140625,
0.9208984375,
0.455078125,
0.50634765625,
0.455078125
],
[
0.080078125,
0.455810546875,
0.259033203125,
0.455810546875,
0.259033203125,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.0310516357421875,
0.4462890625,
0.058013916015625,
0.4462890625,
0.058013916015625,
0.48291015625,
0.0310516357421875,
0.48291015625
],
[
0.50732421875,
0.45703125,
0.9189453125,
0.45703125,
0.9189453125,
0.47021484375,
0.50732421875,
0.47021484375
],
[
0.50732421875,
0.4716796875,
0.9208984375,
0.4716796875,
0.9208984375,
0.48486328125,
0.50732421875,
0.48486328125
],
[
0.21728515625,
0.48095703125,
0.351318359375,
0.48095703125,
0.351318359375,
0.494384765625,
0.21728515625,
0.494384765625
],
[
0.5087890625,
0.486083984375,
0.9208984375,
0.486083984375,
0.9208984375,
0.5,
0.5087890625,
0.5
],
[
0.0236968994140625,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.5888671875,
0.0236968994140625,
0.5888671875
],
[
0.0931396484375,
0.50048828125,
0.491943359375,
0.50048828125,
0.491943359375,
0.5146484375,
0.0931396484375,
0.5146484375
],
[
0.50732421875,
0.501953125,
0.9208984375,
0.501953125,
0.9208984375,
0.51513671875,
0.50732421875,
0.51513671875
],
[
0.07843017578125,
0.5166015625,
0.494384765625,
0.5166015625,
0.494384765625,
0.52978515625,
0.07843017578125,
0.52978515625
],
[
0.50732421875,
0.5166015625,
0.9208984375,
0.5166015625,
0.9208984375,
0.52978515625,
0.50732421875,
0.52978515625
],
[
0.07843017578125,
0.53076171875,
0.494384765625,
0.53076171875,
0.494384765625,
0.54443359375,
0.07843017578125,
0.54443359375
],
[
0.50634765625,
0.53076171875,
0.9208984375,
0.53076171875,
0.9208984375,
0.54443359375,
0.50634765625,
0.54443359375
],
[
0.07843017578125,
0.546875,
0.495849609375,
0.546875,
0.495849609375,
0.56005859375,
0.07843017578125,
0.56005859375
],
[
0.50341796875,
0.546875,
0.9208984375,
0.546875,
0.9208984375,
0.56005859375,
0.50341796875,
0.56005859375
],
[
0.07843017578125,
0.56298828125,
0.494384765625,
0.56298828125,
0.494384765625,
0.57373046875,
0.07843017578125,
0.57373046875
],
[
0.50634765625,
0.56005859375,
0.9208984375,
0.56103515625,
0.9208984375,
0.57568359375,
0.50634765625,
0.57470703125
],
[
0.0767822265625,
0.5771484375,
0.49267578125,
0.5771484375,
0.49267578125,
0.59033203125,
0.0767822265625,
0.59033203125
],
[
0.50634765625,
0.5771484375,
0.9208984375,
0.5771484375,
0.9208984375,
0.59033203125,
0.50634765625,
0.59033203125
],
[
0.07843017578125,
0.59228515625,
0.494384765625,
0.59228515625,
0.494384765625,
0.60595703125,
0.07843017578125,
0.60595703125
],
[
0.50244140625,
0.59130859375,
0.9208984375,
0.59033203125,
0.9208984375,
0.60498046875,
0.50244140625,
0.60595703125
],
[
0.03594970703125,
0.59326171875,
0.0482177734375,
0.59326171875,
0.0482177734375,
0.60693359375,
0.03594970703125,
0.60693359375
],
[
0.03350830078125,
0.5947265625,
0.037567138671875,
0.5947265625,
0.037567138671875,
0.60400390625,
0.03350830078125,
0.60400390625
],
[
0.07843017578125,
0.60693359375,
0.494384765625,
0.60693359375,
0.494384765625,
0.62060546875,
0.07843017578125,
0.62060546875
],
[
0.50341796875,
0.60693359375,
0.9208984375,
0.60595703125,
0.9208984375,
0.62060546875,
0.50341796875,
0.6220703125
],
[
0.0767822265625,
0.62060546875,
0.494384765625,
0.6220703125,
0.494384765625,
0.63623046875,
0.0767822265625,
0.63525390625
],
[
0.5048828125,
0.62255859375,
0.92236328125,
0.62255859375,
0.92236328125,
0.63623046875,
0.5048828125,
0.63623046875
],
[
0.5,
0.63720703125,
0.70361328125,
0.63720703125,
0.70361328125,
0.65087890625,
0.5,
0.65087890625
],
[
0.0310516357421875,
0.625,
0.056365966796875,
0.6298828125,
0.049835205078125,
0.6484375,
0.0236968994140625,
0.642578125
],
[
0.035125732421875,
0.642578125,
0.051483154296875,
0.642578125,
0.051483154296875,
0.6640625,
0.035125732421875,
0.6640625
],
[
0.08087158203125,
0.63818359375,
0.491943359375,
0.63818359375,
0.491943359375,
0.6484375,
0.08087158203125,
0.6484375
],
[
0.07843017578125,
0.6513671875,
0.491943359375,
0.6513671875,
0.491943359375,
0.66552734375,
0.07843017578125,
0.66552734375
],
[
0.52294921875,
0.6513671875,
0.9208984375,
0.65087890625,
0.9208984375,
0.66552734375,
0.52294921875,
0.66650390625
],
[
0.07843017578125,
0.66748046875,
0.49267578125,
0.66748046875,
0.49267578125,
0.68115234375,
0.07843017578125,
0.68115234375
],
[
0.5048828125,
0.66748046875,
0.9189453125,
0.66748046875,
0.9189453125,
0.681640625,
0.5048828125,
0.681640625
],
[
0.0269622802734375,
0.6767578125,
0.0482177734375,
0.67236328125,
0.058013916015625,
0.705078125,
0.03594970703125,
0.708984375
],
[
0.0767822265625,
0.68115234375,
0.494384765625,
0.681640625,
0.494384765625,
0.6962890625,
0.0767822265625,
0.69580078125
],
[
0.5048828125,
0.68310546875,
0.9189453125,
0.68310546875,
0.9189453125,
0.6962890625,
0.5048828125,
0.6962890625
],
[
0.07843017578125,
0.69775390625,
0.43701171875,
0.69775390625,
0.43701171875,
0.7109375,
0.07843017578125,
0.7109375
],
[
0.50732421875,
0.69775390625,
0.9189453125,
0.69775390625,
0.9189453125,
0.7109375,
0.50732421875,
0.7109375
],
[
0.09478759765625,
0.71337890625,
0.49267578125,
0.71337890625,
0.49267578125,
0.7265625,
0.09478759765625,
0.7265625
],
[
0.50634765625,
0.71337890625,
0.9208984375,
0.71337890625,
0.9208984375,
0.7265625,
0.50634765625,
0.7265625
],
[
0.07843017578125,
0.72802734375,
0.494384765625,
0.72802734375,
0.494384765625,
0.7412109375,
0.07843017578125,
0.7412109375
],
[
0.50341796875,
0.72802734375,
0.9208984375,
0.72802734375,
0.9208984375,
0.7412109375,
0.50341796875,
0.7412109375
],
[
0.080078125,
0.7412109375,
0.491943359375,
0.7412109375,
0.491943359375,
0.755859375,
0.080078125,
0.755859375
],
[
0.50341796875,
0.7421875,
0.91845703125,
0.7421875,
0.91845703125,
0.755859375,
0.50341796875,
0.755859375
],
[
0.07843017578125,
0.7568359375,
0.491943359375,
0.7568359375,
0.491943359375,
0.77099609375,
0.07843017578125,
0.77099609375
],
[
0.50341796875,
0.75830078125,
0.9189453125,
0.75830078125,
0.9189453125,
0.771484375,
0.50341796875,
0.771484375
],
[
0.07843017578125,
0.7734375,
0.494384765625,
0.7734375,
0.494384765625,
0.787109375,
0.07843017578125,
0.787109375
],
[
0.50732421875,
0.7734375,
0.88134765625,
0.7734375,
0.88134765625,
0.787109375,
0.50732421875,
0.787109375
],
[
0.07843017578125,
0.78857421875,
0.491943359375,
0.78857421875,
0.491943359375,
0.8017578125,
0.07843017578125,
0.8017578125
],
[
0.5810546875,
0.796875,
0.8447265625,
0.79736328125,
0.8447265625,
0.8125,
0.5810546875,
0.81103515625
],
[
0.07843017578125,
0.80322265625,
0.491943359375,
0.80322265625,
0.491943359375,
0.81640625,
0.07843017578125,
0.81640625
],
[
0.080078125,
0.818359375,
0.491943359375,
0.818359375,
0.491943359375,
0.83203125,
0.080078125,
0.83203125
],
[
0.52392578125,
0.8173828125,
0.9208984375,
0.8173828125,
0.9208984375,
0.83203125,
0.52392578125,
0.83203125
],
[
0.07843017578125,
0.83349609375,
0.49267578125,
0.83349609375,
0.49267578125,
0.8466796875,
0.07843017578125,
0.8466796875
],
[
0.5048828125,
0.83349609375,
0.9189453125,
0.83349609375,
0.9189453125,
0.8466796875,
0.5048828125,
0.8466796875
],
[
0.07843017578125,
0.8486328125,
0.495849609375,
0.8486328125,
0.495849609375,
0.8623046875,
0.07843017578125,
0.8623046875
],
[
0.50341796875,
0.8486328125,
0.9189453125,
0.8486328125,
0.9189453125,
0.8623046875,
0.50341796875,
0.8623046875
],
[
0.07843017578125,
0.86279296875,
0.495849609375,
0.86279296875,
0.495849609375,
0.87744140625,
0.07843017578125,
0.87744140625
],
[
0.5009765625,
0.86279296875,
0.9208984375,
0.86279296875,
0.9208984375,
0.876953125,
0.5009765625,
0.876953125
],
[
0.07843017578125,
0.87890625,
0.50244140625,
0.87890625,
0.50244140625,
0.892578125,
0.07843017578125,
0.892578125
],
[
0.50341796875,
0.87890625,
0.9208984375,
0.87890625,
0.9208984375,
0.892578125,
0.50341796875,
0.892578125
],
[
0.080078125,
0.89599609375,
0.494384765625,
0.89599609375,
0.494384765625,
0.90576171875,
0.080078125,
0.90576171875
],
[
0.50634765625,
0.89453125,
0.76953125,
0.892578125,
0.76953125,
0.90576171875,
0.50634765625,
0.90771484375
]
] |
[
"Vision-Aided Dynamic Blockage Prediction for",
"6G Wireless Communication Networks",
"Gouranga Charan, Muhammad Alrabeiah, and Ahmed Alkhateeb",
"Arizona State University, Emails: f gcharan, malrabei, alkhateeb} @asu.edu",
"Abstract-Unlocking the full potential of millimeter-wave and",
"problem. It studies proactive blockage prediction for a single-",
"sub-terahertzwireless communicationnetworkshingesonrealiz-",
"moving mmWave user in the presence of stationary blockages.",
"ing unprecedented low-latency and high-reliability requirements.",
"Again, despite its appeal, it still falls short in meeting the",
"The challenge in meeting those requirements lies partly in the",
"sensitivity of signals in the millimeter-wave and sub-terahertz",
"latency and reliability requirements.",
"",
"frequencyranges to blockages.Onepromisingwayto tacklethat",
"In this paper and inspired by the recently proposed Vision-",
"challenge is to help a wireless network develop a sense of its",
"Aided Wireless Communication (VAWC) framework in [6]",
"surroundingusingmachinelearningThispaperattemptstodo",
"and [7], the proactive link-blockage problem in fully dynamic",
"that by utilizing deep learning and computer vision. It proposes",
"un",
"environments is addressed using deep learning and a fusion of",
"anovelsolutionthatproactivelypredictsdvnamiclinkblockages.",
"More specifically, it develops a deep neural network architecture",
"visual and wireless data. Images and video sequences usually",
"thatlearnsfromobserved seauencesofRGBimagesandbeam-",
"speak volumes about the environment they are depicting.",
"8",
"",
"formingvectorshowtopredictpossiblefuturelinkblockages.The",
"The empirical evidence in [6] supports that notion; a deep",
"proposed architectureis evaluated on a publicly availabledataset",
"learning model is able to perform simple mmWave beam and",
"that represents a synthetic dynamic communication scenario with",
"blockage prediction tasks using visual data. As such, a fusion",
"A",
"multiple moving users and blockages.It scores a link-blockage",
"prediction accuracy in the neighborhood of 86%, a performance",
"of visual and wireless data is viewed as the right type of",
"that is unlikely to be matched without utilizing visual data.",
"sensory data that could facilitate the learning of the more",
"Index Terms-Deep learning, computer vision, mmWave, ter-",
"complex task of proactive link-blockage prediction. Based on",
"ahertz,blockage prediction.",
"'S",
"this motivation, we develop a novel deep learning architec-",
"ture to utilize sequences of observed RGB images (frames)",
"I.INTRODUCTION",
"and beamforming vectors and learn proactive link-blockage",
"212060",
"Millimeter-wave (mmWave) and sub-terahertz communica-",
"prediction. The proposed architecture harnesses the power of",
"tions are becoming the dominant directions for modern and",
"Convolutional Neural networks (CNNs) and Recurrent Neural",
"future wireless networks [1], [2]. With their large bandwidth,",
"Networks (RNNs). The former excels in performing visual",
"they have the ability to satisfy the high data rate demands",
"tasks such as image classification [8] and object detection",
"of several applications suchas wirelessVirtual/Augmented",
"[9] while the later defines the stat-of-the-art in sequence",
"Reality (VR/AR) and autonomous driving.Communication",
"modeling, as evident by the performance of RNNs in speech",
"in these bands, however, faces several challenges at both the",
"recognition [10], [11]. Experimental results show that the",
"6",
"",
"physical and network layers. One of the key challenges stems",
"proposed architecture is capable of learning the prediction task",
"from the sensitivity of mmWave and terahertz signal propaga-",
"effectively, achieving a noticeable gain over the models that",
"rely solely on wireless data.",
"2",
".A",
"tiontoblockages.Thisrequiresthesesystemstorelyheavily",
"on maintaining line-of-sight (LOS) connections between the",
"The rest of this paper is organized as follows. Section",
"base stations and users. The possibility of blocking these LOS",
"II introduces the system and channel models. Section III",
"Ye",
"links by stationary or dynamic blockages can highly affect the",
"provides a formal description of the proactive link-blockage",
"reliability and latency of mmWave/terahertz systems.",
"prediction problem.Following the problem definition, Section",
"The key for overcoming the link blockage challenges lies",
"IV presents a detailed description of the proposed architec-",
"in enabling the wireless system to develop a sense of its",
"ture, while Section V details the experimental setup used to",
"surrounding. This could be accomplished using machine learn-",
"evaluate the performance of the architecture.The performance",
"ing. Many recently published studies have shown that using",
"evaluation, then, is presented in Section VI. Finally, Section",
"wireless sensory data (e.g., channels, received power, etc.), a",
"VII wraps up the paper with some concluding remarks.",
"machine learning model can efficiently differentiate LOS and",
"II.SYSTEMANDCHANNEL MODELS",
"Non-LOS (NLOS) links [3], [4], [5]. The former two studies",
"address the link blockage problem form a reactive perspective;",
"To illustrate the potential of deep learning and VAWC in",
"the sensory data is first acquired, and, then, the status of the",
"mitigating the link blockage problem, this work considers a",
"current link is predicted. Despite its appeal, the reactive link-",
"high-frequency communication system where the basestation",
"blockage prediction does not satisfy the low-latency and high-",
"utilizes an RGB camera to monitor its environment. The",
"reliability requirements of future wireless networks. The work",
"following two subsections provide a brief description of that",
"in[5]takesastepforwardtowardsaproactivetreatmentofthe",
"system and its adopted channel model."
] |
{
"arxiv_id": "2006.09902",
"language": "en",
"page_idx": 0
}
|
||
2006.09902-en-2
|
2006.09902-en-2.jpg
|
<FigureHere>
Fig. 2. An overall blockage diagram illustrating the problem and its solution. The deep learning model takes in a sequence of observed image-beam pairs
and produces a prediction of the link status of the user in a future time instance.
translates that abstract description into a three-component a pre-trained YOLO network and integrates it into its architec-
network, as illustrated in Fig.3. The following subsections
provide a deeper dive into each component.
ture with some minor modifications; the detector is stripped off of its classifier and bounding-box predictor. Such choice allows the solution to extract rich feature volumes without the need to undergo extensive object-detection training, commonly associated with state-of-the-art detectors.
## $A.$ Object Detector
# be the state-of-the-art networks for visual tasks like image $B.Feature$ and Beam Embedding
Recent advances in deep learning have shown CNNs to classification [8],[14] and object detection [9],[15] to name two. As such, the architecture of the proposed solution im- data, i.e., visual and wireless data. Although such data is
plements a CNN-based object detector as its first component. The job of this detector is to extract visual features pertaining to relevant objects in the environment like cars, people, buses, etc. Since object detection in itself is not the objective of this work, the detector is not equipped with an object classifier or a bounding-box predictor. Instead, it has a sequence of convolutional layers that extract a feature volume at the end. As shown in Fig.3, the detector is designed to sequentially process the images in the input sequence and extract their the adopted embedding is simple and does not require any
corresponding feature volumes.
The object detector in the proposed solution needs to meet $\mathbf{b}_i\in\mathbb{R}^N$ where $i\in\{t-r+1,\ldots,t\}.$ The elements of each
two essential requirements: (i) detecting a wide range of objects, and (ii) producing quick and accurate predictions. These two requirements have been addressed in many of the state-of-the-art object detectors. A good example of a fast are embedded using a trainable CNN network. It takes in
and accurate object-detection neural networks is the You Only
Look Once (YOLO) detector, proposed first in[9] and then $\mathbf{d}_i\in\mathbb{R}^N$ where $i\in\{t-r+1,\ldots,t\}$, see the image
improved in [16]. Therefore, instead of building and training a new object detector from scratch, the proposed solution utilizes
The prediction function relies on dual-modality observed expected to be rife with information, its dual nature brings about a heightened level of difficulty from the learning perspective. In an effort to overcome that difficulty, the proposed solution incorporates an embedding component that processes the extracted image feature volumes and the beam indices separately., as shown in the middle part of Fig.3. It transforms them to the same $N$-dimensional space before they are fed to the next component. For beam indices in the input sequence, training. It generates a lookup table of $|\mathcal{F}|$ real-valued vectors vector are randomly drawn from a Gaussian distribution with zeros mean and unity standard deviation.
For the extracted feature volumes, on the other hand, they one feature volume at a time and produces a feature vector feature embedding part in Fig.3. Since a volume embodies rich information about the objects in the environment, some
|
[
[
0.1854248046875,
0.0662841796875,
0.295654296875,
0.06439208984375,
0.295654296875,
0.08209228515625,
0.1854248046875,
0.083984375
],
[
0.423095703125,
0.0662841796875,
0.53125,
0.06439208984375,
0.53125,
0.08209228515625,
0.423095703125,
0.083984375
],
[
0.7255859375,
0.06500244140625,
0.7998046875,
0.06500244140625,
0.7998046875,
0.08331298828125,
0.7255859375,
0.08331298828125
],
[
0.1258544921875,
0.085205078125,
0.2017822265625,
0.085205078125,
0.2017822265625,
0.0953369140625,
0.1258544921875,
0.0953369140625
],
[
0.363525390625,
0.0889892578125,
0.371826171875,
0.0889892578125,
0.371826171875,
0.0921630859375,
0.363525390625,
0.0921630859375
],
[
0.136474609375,
0.1427001953125,
0.2320556640625,
0.1427001953125,
0.2320556640625,
0.1572265625,
0.136474609375,
0.1572265625
],
[
0.464111328125,
0.1988525390625,
0.48193359375,
0.1988525390625,
0.48193359375,
0.20263671875,
0.464111328125,
0.20263671875
],
[
0.48193359375,
0.1988525390625,
0.486083984375,
0.1988525390625,
0.486083984375,
0.203857421875,
0.48193359375,
0.203857421875
],
[
0.36279296875,
0.2374267578125,
0.6298828125,
0.2374267578125,
0.6298828125,
0.250732421875,
0.36279296875,
0.250732421875
],
[
0.32177734375,
0.2841796875,
0.373291015625,
0.2841796875,
0.373291015625,
0.29541015625,
0.32177734375,
0.29541015625
],
[
0.450927734375,
0.285888671875,
0.54345703125,
0.2890625,
0.54150390625,
0.30810546875,
0.449462890625,
0.304931640625
],
[
0.62744140625,
0.2822265625,
0.66845703125,
0.2822265625,
0.66845703125,
0.296630859375,
0.62744140625,
0.296630859375
],
[
0.3095703125,
0.2998046875,
0.3857421875,
0.2998046875,
0.3857421875,
0.311279296875,
0.3095703125,
0.311279296875
],
[
0.609375,
0.2998046875,
0.68701171875,
0.2998046875,
0.68701171875,
0.311279296875,
0.609375,
0.311279296875
],
[
0.400390625,
0.35986328125,
0.59228515625,
0.35986328125,
0.59228515625,
0.375,
0.400390625,
0.375
],
[
0.439453125,
0.380615234375,
0.560546875,
0.380615234375,
0.560546875,
0.39453125,
0.439453125,
0.39453125
],
[
0.66845703125,
0.3837890625,
0.83642578125,
0.3837890625,
0.83642578125,
0.397705078125,
0.66845703125,
0.397705078125
],
[
0.0767822265625,
0.4248046875,
0.9208984375,
0.4248046875,
0.9208984375,
0.438232421875,
0.0767822265625,
0.438232421875
],
[
0.080078125,
0.438232421875,
0.50634765625,
0.438232421875,
0.50634765625,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.0767822265625,
0.47607421875,
0.91845703125,
0.474853515625,
0.91845703125,
0.4892578125,
0.0767822265625,
0.490478515625
],
[
0.07843017578125,
0.4912109375,
0.50244140625,
0.4912109375,
0.50244140625,
0.5048828125,
0.07843017578125,
0.5048828125
],
[
0.494384765625,
0.4912109375,
0.9189453125,
0.4912109375,
0.9189453125,
0.5048828125,
0.494384765625,
0.5048828125
],
[
0.07843017578125,
0.5068359375,
0.373291015625,
0.5068359375,
0.373291015625,
0.52099609375,
0.07843017578125,
0.52099609375
],
[
0.50732421875,
0.50830078125,
0.9189453125,
0.50830078125,
0.9189453125,
0.51806640625,
0.50732421875,
0.51806640625
],
[
0.50634765625,
0.521484375,
0.9189453125,
0.521484375,
0.9189453125,
0.53515625,
0.50634765625,
0.53515625
],
[
0.0767822265625,
0.53857421875,
0.2091064453125,
0.53857421875,
0.2091064453125,
0.5517578125,
0.0767822265625,
0.5517578125
],
[
0.50732421875,
0.537109375,
0.9189453125,
0.537109375,
0.9189453125,
0.55029296875,
0.50732421875,
0.55029296875
],
[
0.0931396484375,
0.5595703125,
0.490234375,
0.5595703125,
0.490234375,
0.57275390625,
0.0931396484375,
0.57275390625
],
[
0.5107421875,
0.5537109375,
0.78173828125,
0.5537109375,
0.78173828125,
0.564453125,
0.5107421875,
0.564453125
],
[
0.0767822265625,
0.57568359375,
0.49267578125,
0.57568359375,
0.49267578125,
0.5888671875,
0.0767822265625,
0.5888671875
],
[
0.5048828125,
0.5712890625,
0.734375,
0.57373046875,
0.734375,
0.5869140625,
0.5048828125,
0.58544921875
],
[
0.07843017578125,
0.59033203125,
0.491943359375,
0.59033203125,
0.491943359375,
0.60400390625,
0.07843017578125,
0.60400390625
],
[
0.52392578125,
0.59130859375,
0.9208984375,
0.59130859375,
0.9208984375,
0.60498046875,
0.52392578125,
0.60498046875
],
[
0.0767822265625,
0.60595703125,
0.494384765625,
0.60595703125,
0.494384765625,
0.619140625,
0.0767822265625,
0.619140625
],
[
0.50634765625,
0.60693359375,
0.9208984375,
0.60693359375,
0.9208984375,
0.62060546875,
0.50634765625,
0.62060546875
],
[
0.0767822265625,
0.62060546875,
0.494384765625,
0.619140625,
0.494384765625,
0.6337890625,
0.0767822265625,
0.63525390625
],
[
0.50634765625,
0.6220703125,
0.9208984375,
0.6220703125,
0.9208984375,
0.63623046875,
0.50634765625,
0.63623046875
],
[
0.07843017578125,
0.63623046875,
0.494384765625,
0.63623046875,
0.494384765625,
0.6494140625,
0.07843017578125,
0.6494140625
],
[
0.5048828125,
0.63623046875,
0.9189453125,
0.63720703125,
0.9189453125,
0.6513671875,
0.5048828125,
0.65087890625
],
[
0.07843017578125,
0.65087890625,
0.490234375,
0.65087890625,
0.490234375,
0.6640625,
0.07843017578125,
0.6640625
],
[
0.50732421875,
0.6513671875,
0.9208984375,
0.6513671875,
0.9208984375,
0.66552734375,
0.50732421875,
0.66552734375
],
[
0.07843017578125,
0.66650390625,
0.49267578125,
0.66650390625,
0.49267578125,
0.6796875,
0.07843017578125,
0.6796875
],
[
0.50634765625,
0.66748046875,
0.9208984375,
0.66748046875,
0.9208984375,
0.68115234375,
0.50634765625,
0.68115234375
],
[
0.07843017578125,
0.681640625,
0.494384765625,
0.681640625,
0.494384765625,
0.69580078125,
0.07843017578125,
0.69580078125
],
[
0.50634765625,
0.68310546875,
0.9189453125,
0.68310546875,
0.9189453125,
0.6962890625,
0.50634765625,
0.6962890625
],
[
0.0767822265625,
0.6962890625,
0.494384765625,
0.6962890625,
0.494384765625,
0.71044921875,
0.0767822265625,
0.71044921875
],
[
0.5048828125,
0.69775390625,
0.9208984375,
0.6962890625,
0.9208984375,
0.7109375,
0.5048828125,
0.7119140625
],
[
0.080078125,
0.71337890625,
0.490234375,
0.71337890625,
0.490234375,
0.7236328125,
0.080078125,
0.7236328125
],
[
0.50732421875,
0.71337890625,
0.9208984375,
0.71337890625,
0.9208984375,
0.7265625,
0.50732421875,
0.7265625
],
[
0.080078125,
0.7265625,
0.491943359375,
0.7265625,
0.491943359375,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.5048828125,
0.7265625,
0.91845703125,
0.7265625,
0.91845703125,
0.74072265625,
0.5048828125,
0.74072265625
],
[
0.0767822265625,
0.74072265625,
0.495849609375,
0.7412109375,
0.495849609375,
0.755859375,
0.0767822265625,
0.7548828125
],
[
0.5048828125,
0.7421875,
0.91845703125,
0.7421875,
0.91845703125,
0.755859375,
0.5048828125,
0.755859375
],
[
0.07598876953125,
0.7568359375,
0.2900390625,
0.7548828125,
0.2900390625,
0.76806640625,
0.07598876953125,
0.77099609375
],
[
0.50634765625,
0.7568359375,
0.91845703125,
0.7568359375,
0.91845703125,
0.77099609375,
0.50634765625,
0.77099609375
],
[
0.09234619140625,
0.77294921875,
0.495849609375,
0.77294921875,
0.495849609375,
0.787109375,
0.09234619140625,
0.787109375
],
[
0.50634765625,
0.77294921875,
0.9208984375,
0.77294921875,
0.9208984375,
0.787109375,
0.50634765625,
0.787109375
],
[
0.0767822265625,
0.78857421875,
0.495849609375,
0.78857421875,
0.495849609375,
0.8017578125,
0.0767822265625,
0.8017578125
],
[
0.50634765625,
0.78857421875,
0.9208984375,
0.78857421875,
0.9208984375,
0.8017578125,
0.50634765625,
0.8017578125
],
[
0.07843017578125,
0.80322265625,
0.49267578125,
0.80322265625,
0.49267578125,
0.81640625,
0.07843017578125,
0.81640625
],
[
0.50634765625,
0.8037109375,
0.78515625,
0.8037109375,
0.78515625,
0.8173828125,
0.50634765625,
0.8173828125
],
[
0.0767822265625,
0.81640625,
0.491943359375,
0.8173828125,
0.491943359375,
0.83203125,
0.0767822265625,
0.83056640625
],
[
0.52294921875,
0.8173828125,
0.9189453125,
0.8173828125,
0.9189453125,
0.83203125,
0.52294921875,
0.83203125
],
[
0.07843017578125,
0.83349609375,
0.49267578125,
0.83349609375,
0.49267578125,
0.8466796875,
0.07843017578125,
0.8466796875
],
[
0.5048828125,
0.833984375,
0.9189453125,
0.83203125,
0.9189453125,
0.84521484375,
0.5048828125,
0.84765625
],
[
0.080078125,
0.849609375,
0.490234375,
0.849609375,
0.490234375,
0.85986328125,
0.080078125,
0.85986328125
],
[
0.50634765625,
0.849609375,
0.9189453125,
0.849609375,
0.9189453125,
0.85986328125,
0.50634765625,
0.85986328125
],
[
0.07843017578125,
0.86279296875,
0.5,
0.86279296875,
0.5,
0.87744140625,
0.07843017578125,
0.87744140625
],
[
0.495849609375,
0.85986328125,
0.92236328125,
0.8623046875,
0.92236328125,
0.87890625,
0.495849609375,
0.876953125
],
[
0.07843017578125,
0.87890625,
0.496826171875,
0.87890625,
0.496826171875,
0.892578125,
0.07843017578125,
0.892578125
],
[
0.50634765625,
0.87890625,
0.9208984375,
0.87890625,
0.9208984375,
0.892578125,
0.50634765625,
0.892578125
],
[
0.0767822265625,
0.89306640625,
0.495849609375,
0.892578125,
0.495849609375,
0.9072265625,
0.0767822265625,
0.90771484375
],
[
0.50634765625,
0.89306640625,
0.9208984375,
0.89306640625,
0.9208984375,
0.9072265625,
0.50634765625,
0.9072265625
]
] |
[
"Instant t -7",
"Instant t - 6",
"Instant t",
"",
"",
"Served User",
"",
"+",
"ImageSequence+BeamSequence",
"Object",
"Embedding",
"RNN",
"Detection",
"Prediction",
"Future Blockage Prediction",
"At Instant t + 1",
"Ground Truth:Blocked",
"Fig. 2. An overall blockage diagram illustrating the problem and its solution. The deep learning model takes in a sequence of observed image-beam pairs",
"and produces aprediction of the link status of theuser in a futuretime instanc",
"translates that abstract description into a three-componenta pre-trained YOLO network and integrates it into its architec-",
"network, as illustrated in Fig.3. The following subsections",
"ture with some minor modifications; the detector is stripped",
"provide a deeper dive into each component.",
"offofits classifierandbounding-boxpredictor.Suchchoice",
"allows the solution to extract rich feature volumes without the",
"A.Object Detector",
"need to undergo extensive object-detection training, commonly",
"Recent advances in deep learning have shown CNNs to",
"associated with state-of-the-artdetectors.",
"be the state-of-the-art networks for visual tasks like image",
"B.FeatureandBeamEmbedding",
"classification [8], [14] and object detection [9], [15] to name",
"The prediction function relies on dual-modality observed",
"two. As such, the architecture of the proposed solution im-",
"data, i.e., visual and wireless data. Although such data is",
"plements a CNN-based object detector as its first component.",
"expected to be rife with information, its dual nature brings",
"The job of this detector is to extract visual features pertaining",
"about a heightened level of difficulty from the learning per",
"to relevant objects in the environment like cars, people, buses,",
"spective. In an effort to overcome that difficulty, the proposed",
"etc. Since object detection in itself is not the objective of this",
"solution incorporates an embedding component that processes",
"work, the detector is not equipped with an object classifier",
"the extracted image feature volumes and the beam indices",
"or a bounding-box predictor. Instead, it has a sequence of",
"separately., as shown in the middle part of Fig.3. It transforms",
"convolutionallayersthatextractafeaturevolumeattheend.",
"them to the same N-dimensional space before they are fed to",
"As shown in Fig.3, the detector is designed to sequentially",
"the next component. For beam indices in the input sequence",
"process the images in the input sequence and extract their",
"the adopted embedding is simple and does not require any",
"corresponding feature volumes.",
"training. It generates a lookup table of F real-valued vectors",
"The object detector in the proposed solution needs to meet",
"b, e RN where ie (t -r + 1,...,t). The elements of each",
"two essential requirements: (i) detecting a wide range of",
"vector are randomly drawn from a Gaussian distribution with",
"objects, and (i) producing quick and accurate predictions.",
"zeros mean and unity standard deviation.",
"These two requirements have been addressed in many of the",
"For the extracted feature volumes, on the other hand, they",
"state-of-the-art object detectors. A good example of a fast",
"are embedded using a trainable CNN network.It takes in",
"andaccurateobject-detectionneuralnetworksistheYouOnly",
"onefeaturevolumeatatimeandproducesa featurevector",
"Look Once (YOLO) detector, proposed first in [9] and then",
"d, e RN where i e {t - r + 1,...,t), see the image",
"improved in [16]. Therefore, instead of building and training a",
"feature embedding part in Fig.3. Since a volume embodies",
"new object detector from scratch, the proposed solution utilizes",
"rich information about the objects in the environment, some"
] |
{
"arxiv_id": "2006.09902",
"language": "en",
"page_idx": 2
}
|
||
2309.03467-en-1
|
2309.03467-en-1.jpg
|
adaptation methods on pretrained GANs, which can be conditioned on both NFoV images and text guidances. While these methods show encouraging performance, the challenge remains regarding the usage of given NFoV images and user-provided open vocabulary text guidances individually or jointly for enhanced control in 360-degree image generation. Specifically, the existing methods typically fall short in synthesizing intricate visual details as shown in the top part of Fig. 1, where the details are vague or missing with OmniDreamer compared to our approach, which are highlighted in the red bounding boxes. Moreover, the generated images and user-provided text guidances tend to be inconsistent, especially under an open-vocabulary setting, as depicted in the bottom part of Fig. 1 by comparing Text2Light and our solution.
In this study, we propose a novel autoregressive omniaware generative network (AOG-Net) for generating 360- degree images conditioned on open vocabulary text guidances and given NFoV images jointly or individually. Overall, the generation is formulated as an autoregressive stochastic process to outpaint an incomplete 360-degree image progressively, in which each step outpaints a local region under its corresponding NFoV view. This autoregressive scheme not only allows for deriving finer-grained and prompt-consistent patterns by dynamically observing and adjusting the generation process but also offers users greater flexibility to modify or introduce new conditions throughout the generation process. Furthermore, a global-local conditioning mechanism is devised to comprehensively formulate the outpainting guidance for each autoregression step. Text prompts, omni-visual cues, NFoV inputs and omnigeometry are encoded and further formulated with crossattention based transformers into a global stream and a local stream for a conditioned generative backbone model. This study further explores the potential to leverage largescale models for the conditional encoder and the generative prior, which helps complete the generation using openvocabulary prompts. Comprehensive experiments on two commonly used 360-degree image datasets for both indoor and outdoor settings demonstrate the state-of-the-art performance of our proposed method.
In summary, the key contributions of this study are:
· A novel autoregressive outpainting approach is proposed to produce photo-realistic 360-degree images by dynamically adjusting the generation process for improved finergrained details and prompt-consistency.
·A global-local conditioning mechanism is devised to
formulate the guidance encompassing open-vocab text guidances, omni-visual cues, NFoV inputs and omnigeometry with cross-attention based transformers.
$\bullet$ Comprehensive experiments were conducted on two
$\hat{\text{commonly used benchmarks, demonstrating the state-of-}}$ the-art performance of AOG-Net in both indoor and outdoor settings with as few as 40 training samples.
$\textbf{Related Work}$
We first review the studies in both the field of 360-degree image generation and the field of image outpainting which
are relevant to our study. As our work takes image and text guidances as conditions, we further review the related studies on conditional image generation.
## 360-Degree Image Generation Unlike general NFoV images, 360-degree image generation
requires to take the omni-directional continuity into account. Early studies, for example, (Sengupta et al. 2019) estimates a coarse 360-degree image from an NFoV image with inverse rendering technique, which ignores such geometrical continuity and generates 360-degree images lack of fine details. To address this, 360IC (Akimoto et al. 2019) and SIGSS (Hara, Mukuta, and Harada 2021) were proposed to improve geometrical continuity by taking the intrinsic horizontal cyclicity into consideration and encoding it as positional conditions to connect the two ends of 360-degree images in equirectangular representations. EnvMapNet (Somanath and Kurz 2021) improves visual quality of the outpainted 360- degree images by introducing a projection loss and a clustering loss for accurate lighting and shadowing. OmniDreamer (Akimoto, Matsuo, and Aoki 2022) was further developed by leveraging the Taming-Transformer (Esser, Rombach, and Ommer 2021), where a circular inference scheme was introduced to fit the intrinsic horizontal cyclicity for 360- degree image synthesis, conditioned on provided NFoV images, yielding diverse and photo-realistic results. However, OmniDreamer is limited to a single condition where only an initial NFoV image is accepted, while the controllability of the overall synthesis process is limited. ImmenseGAN (Dastjerdi et al. 2022) aims for fner controllability over the outpainting by introducing a text guidances to fine-tune a generative model with a large-scale private text-image pair dataset. Due to the lack of public text-image paired dataset, Text2Light (Chen, Wang, and Liu 2022) introduces a zeroshot text-guided 360-degree image synthesis pipeline without using initial NFoV images, in which a pre-trained CLIP model is adopted (Radford et al. 2021).
However, the existing methods typically fall short in synthesizing intricate visual details and inconsistencies can be observed between generated images and user-provided text guidances, especially under an open-vocabulary setting, which demands further mechanisms to address these issues. Image Outpainting
Image outpainting, a fundamental task in computer vision, focuses on expanding the unknown regions outside the primary known content. Unlike inpainting, outpainting may not be able to leverage information from pixels adjacent to the unknown area (Sabini and Rusak 2018; Hoorick 2020; Wang et al. 2019), as seen in inpainting methods. In (Wang et al. 2019), the semantic information of incomplete images was utilized to guide a GAN for outpainting. In (Yao et al. 2022), a query-based outpainting method was proposed, where an image is divided into small patches and the patches with unknown pixels are completed by taking the conditions from both distant and neighbour patches into account.In an iterative manner, (Gardias, Arthur, and Sun 2020) extends one side of the a regular image for outpainting step by step, using the context of the past generation as guidance. (Lu, Chang,
|
[
[
0.08660888671875,
0.0694580078125,
0.476318359375,
0.0694580078125,
0.476318359375,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.521484375,
0.07135009765625,
0.9111328125,
0.07135009765625,
0.9111328125,
0.08209228515625,
0.521484375,
0.08209228515625
],
[
0.08660888671875,
0.08331298828125,
0.47802734375,
0.08331298828125,
0.47802734375,
0.09661865234375,
0.08660888671875,
0.09661865234375
],
[
0.521484375,
0.08331298828125,
0.9111328125,
0.08331298828125,
0.9111328125,
0.09661865234375,
0.521484375,
0.09661865234375
],
[
0.0849609375,
0.09783935546875,
0.47802734375,
0.09661865234375,
0.47802734375,
0.111083984375,
0.0849609375,
0.11236572265625
],
[
0.5185546875,
0.09661865234375,
0.75732421875,
0.09783935546875,
0.75732421875,
0.11236572265625,
0.5185546875,
0.111083984375
],
[
0.08660888671875,
0.11236572265625,
0.4794921875,
0.11236572265625,
0.4794921875,
0.1256103515625,
0.08660888671875,
0.1256103515625
],
[
0.0849609375,
0.1256103515625,
0.47802734375,
0.1256103515625,
0.47802734375,
0.1395263671875,
0.0849609375,
0.1395263671875
],
[
0.51953125,
0.121826171875,
0.7509765625,
0.12249755859375,
0.7509765625,
0.136962890625,
0.51953125,
0.1363525390625
],
[
0.0849609375,
0.1395263671875,
0.47802734375,
0.1395263671875,
0.47802734375,
0.15283203125,
0.0849609375,
0.15283203125
],
[
0.521484375,
0.14013671875,
0.91259765625,
0.14013671875,
0.91259765625,
0.154052734375,
0.521484375,
0.154052734375
],
[
0.08660888671875,
0.15283203125,
0.4794921875,
0.15283203125,
0.4794921875,
0.1666259765625,
0.08660888671875,
0.1666259765625
],
[
0.521484375,
0.1558837890625,
0.9111328125,
0.1558837890625,
0.9111328125,
0.1666259765625,
0.521484375,
0.1666259765625
],
[
0.0882568359375,
0.1685791015625,
0.47802734375,
0.1685791015625,
0.47802734375,
0.1787109375,
0.0882568359375,
0.1787109375
],
[
0.521484375,
0.1685791015625,
0.91259765625,
0.1685791015625,
0.91259765625,
0.1817626953125,
0.521484375,
0.1817626953125
],
[
0.08660888671875,
0.18115234375,
0.4794921875,
0.18115234375,
0.4794921875,
0.1944580078125,
0.08660888671875,
0.1944580078125
],
[
0.51953125,
0.18310546875,
0.9111328125,
0.18310546875,
0.9111328125,
0.1962890625,
0.51953125,
0.1962890625
],
[
0.0882568359375,
0.1944580078125,
0.4755859375,
0.1944580078125,
0.4755859375,
0.2083740234375,
0.0882568359375,
0.2083740234375
],
[
0.521484375,
0.1976318359375,
0.9111328125,
0.1976318359375,
0.9111328125,
0.2083740234375,
0.521484375,
0.2083740234375
],
[
0.0882568359375,
0.210205078125,
0.476318359375,
0.210205078125,
0.476318359375,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.521484375,
0.210205078125,
0.91259765625,
0.210205078125,
0.91259765625,
0.2235107421875,
0.521484375,
0.2235107421875
],
[
0.0849609375,
0.222900390625,
0.476318359375,
0.222900390625,
0.476318359375,
0.236083984375,
0.0849609375,
0.236083984375
],
[
0.51953125,
0.2235107421875,
0.91015625,
0.2235107421875,
0.91015625,
0.2374267578125,
0.51953125,
0.2374267578125
],
[
0.0849609375,
0.236083984375,
0.4755859375,
0.236083984375,
0.4755859375,
0.25,
0.0849609375,
0.25
],
[
0.51953125,
0.2374267578125,
0.9111328125,
0.2374267578125,
0.9111328125,
0.250732421875,
0.51953125,
0.250732421875
],
[
0.0849609375,
0.250732421875,
0.4794921875,
0.250732421875,
0.4794921875,
0.264404296875,
0.0849609375,
0.264404296875
],
[
0.521484375,
0.251953125,
0.9111328125,
0.251953125,
0.9111328125,
0.26513671875,
0.521484375,
0.26513671875
],
[
0.0849609375,
0.264404296875,
0.1993408203125,
0.264404296875,
0.1993408203125,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.521484375,
0.26513671875,
0.91259765625,
0.26513671875,
0.91259765625,
0.279052734375,
0.521484375,
0.279052734375
],
[
0.101318359375,
0.27783203125,
0.476318359375,
0.27783203125,
0.476318359375,
0.291748046875,
0.101318359375,
0.291748046875
],
[
0.51953125,
0.279052734375,
0.9140625,
0.279052734375,
0.9140625,
0.292236328125,
0.51953125,
0.292236328125
],
[
0.0849609375,
0.291748046875,
0.476318359375,
0.290283203125,
0.476318359375,
0.304931640625,
0.0849609375,
0.30615234375
],
[
0.521484375,
0.292236328125,
0.91259765625,
0.292236328125,
0.91259765625,
0.30615234375,
0.521484375,
0.30615234375
],
[
0.0849609375,
0.30615234375,
0.47802734375,
0.30615234375,
0.47802734375,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.521484375,
0.30615234375,
0.91015625,
0.30615234375,
0.91015625,
0.3193359375,
0.521484375,
0.3193359375
],
[
0.08416748046875,
0.317626953125,
0.476318359375,
0.318115234375,
0.476318359375,
0.333251953125,
0.08416748046875,
0.33203125
],
[
0.521484375,
0.32080078125,
0.9111328125,
0.32080078125,
0.9111328125,
0.333984375,
0.521484375,
0.333984375
],
[
0.08660888671875,
0.333251953125,
0.4794921875,
0.333251953125,
0.4794921875,
0.3466796875,
0.08660888671875,
0.3466796875
],
[
0.521484375,
0.335205078125,
0.91259765625,
0.335205078125,
0.91259765625,
0.348388671875,
0.521484375,
0.348388671875
],
[
0.0849609375,
0.3466796875,
0.47802734375,
0.3466796875,
0.47802734375,
0.35986328125,
0.0849609375,
0.35986328125
],
[
0.521484375,
0.348388671875,
0.91259765625,
0.348388671875,
0.91259765625,
0.3623046875,
0.521484375,
0.3623046875
],
[
0.08416748046875,
0.361083984375,
0.4755859375,
0.35986328125,
0.4755859375,
0.373779296875,
0.08416748046875,
0.375
],
[
0.521484375,
0.3623046875,
0.9111328125,
0.3623046875,
0.9111328125,
0.375732421875,
0.521484375,
0.375732421875
],
[
0.08660888671875,
0.373779296875,
0.473876953125,
0.373779296875,
0.473876953125,
0.386962890625,
0.08660888671875,
0.386962890625
],
[
0.521484375,
0.375732421875,
0.91259765625,
0.375732421875,
0.91259765625,
0.389404296875,
0.521484375,
0.389404296875
],
[
0.0849609375,
0.38818359375,
0.4794921875,
0.38818359375,
0.4794921875,
0.401611328125,
0.0849609375,
0.401611328125
],
[
0.5185546875,
0.389404296875,
0.91259765625,
0.38818359375,
0.91259765625,
0.401611328125,
0.5185546875,
0.40283203125
],
[
0.0882568359375,
0.404052734375,
0.476318359375,
0.404052734375,
0.476318359375,
0.4140625,
0.0882568359375,
0.4140625
],
[
0.521484375,
0.404052734375,
0.9111328125,
0.404052734375,
0.9111328125,
0.417236328125,
0.521484375,
0.417236328125
],
[
0.08660888671875,
0.416748046875,
0.4794921875,
0.416748046875,
0.4794921875,
0.429931640625,
0.08660888671875,
0.429931640625
],
[
0.521484375,
0.417236328125,
0.91259765625,
0.417236328125,
0.91259765625,
0.43115234375,
0.521484375,
0.43115234375
],
[
0.08660888671875,
0.429931640625,
0.476318359375,
0.429931640625,
0.476318359375,
0.443115234375,
0.08660888671875,
0.443115234375
],
[
0.51953125,
0.43115234375,
0.91015625,
0.43115234375,
0.91015625,
0.4443359375,
0.51953125,
0.4443359375
],
[
0.08660888671875,
0.443115234375,
0.476318359375,
0.443115234375,
0.476318359375,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.51953125,
0.44580078125,
0.90869140625,
0.44580078125,
0.90869140625,
0.458984375,
0.51953125,
0.458984375
],
[
0.0849609375,
0.45703125,
0.476318359375,
0.458251953125,
0.476318359375,
0.4716796875,
0.0849609375,
0.47021484375
],
[
0.521484375,
0.460205078125,
0.9111328125,
0.460205078125,
0.9111328125,
0.47021484375,
0.521484375,
0.47021484375
],
[
0.0849609375,
0.4716796875,
0.476318359375,
0.4716796875,
0.476318359375,
0.48486328125,
0.0849609375,
0.48486328125
],
[
0.521484375,
0.472900390625,
0.91259765625,
0.472900390625,
0.91259765625,
0.486083984375,
0.521484375,
0.486083984375
],
[
0.0849609375,
0.48486328125,
0.476318359375,
0.48486328125,
0.476318359375,
0.498779296875,
0.0849609375,
0.498779296875
],
[
0.521484375,
0.4873046875,
0.9140625,
0.4873046875,
0.9140625,
0.50048828125,
0.521484375,
0.50048828125
],
[
0.08660888671875,
0.498779296875,
0.47802734375,
0.498779296875,
0.47802734375,
0.51220703125,
0.08660888671875,
0.51220703125
],
[
0.521484375,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.5146484375,
0.521484375,
0.5146484375
],
[
0.08660888671875,
0.5146484375,
0.4755859375,
0.5146484375,
0.4755859375,
0.5244140625,
0.08660888671875,
0.5244140625
],
[
0.5185546875,
0.51318359375,
0.91259765625,
0.5146484375,
0.91259765625,
0.52880859375,
0.5185546875,
0.52783203125
],
[
0.08660888671875,
0.5263671875,
0.4755859375,
0.5263671875,
0.4755859375,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.52294921875,
0.52978515625,
0.91015625,
0.52978515625,
0.91015625,
0.54052734375,
0.52294921875,
0.54052734375
],
[
0.0882568359375,
0.54052734375,
0.476318359375,
0.54052734375,
0.476318359375,
0.5537109375,
0.0882568359375,
0.5537109375
],
[
0.51953125,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55615234375,
0.51953125,
0.55615234375
],
[
0.0882568359375,
0.55615234375,
0.476318359375,
0.55615234375,
0.476318359375,
0.56640625,
0.0882568359375,
0.56640625
],
[
0.51953125,
0.5546875,
0.91259765625,
0.5537109375,
0.91259765625,
0.568359375,
0.51953125,
0.5693359375
],
[
0.08660888671875,
0.568359375,
0.476318359375,
0.568359375,
0.476318359375,
0.58203125,
0.08660888671875,
0.58203125
],
[
0.521484375,
0.57080078125,
0.77685546875,
0.57080078125,
0.77685546875,
0.58056640625,
0.521484375,
0.58056640625
],
[
0.0882568359375,
0.583984375,
0.47802734375,
0.583984375,
0.47802734375,
0.5947265625,
0.0882568359375,
0.5947265625
],
[
0.5361328125,
0.58349609375,
0.91259765625,
0.58349609375,
0.91259765625,
0.59765625,
0.5361328125,
0.59765625
],
[
0.0882568359375,
0.59765625,
0.47802734375,
0.59765625,
0.47802734375,
0.60791015625,
0.0882568359375,
0.60791015625
],
[
0.51953125,
0.59521484375,
0.91259765625,
0.5966796875,
0.91259765625,
0.611328125,
0.51953125,
0.60986328125
],
[
0.08660888671875,
0.609375,
0.4755859375,
0.609375,
0.4755859375,
0.62255859375,
0.08660888671875,
0.62255859375
],
[
0.51953125,
0.611328125,
0.91259765625,
0.611328125,
0.91259765625,
0.625,
0.51953125,
0.625
],
[
0.0849609375,
0.62353515625,
0.295654296875,
0.62353515625,
0.295654296875,
0.63720703125,
0.0849609375,
0.63720703125
],
[
0.5185546875,
0.625,
0.91259765625,
0.62548828125,
0.91259765625,
0.6396484375,
0.5185546875,
0.63818359375
],
[
0.101318359375,
0.63818359375,
0.439453125,
0.63818359375,
0.439453125,
0.6513671875,
0.101318359375,
0.6513671875
],
[
0.51953125,
0.6396484375,
0.9111328125,
0.6396484375,
0.9111328125,
0.65283203125,
0.51953125,
0.65283203125
],
[
0.09234619140625,
0.65478515625,
0.4794921875,
0.65478515625,
0.4794921875,
0.66943359375,
0.09234619140625,
0.66943359375
],
[
0.51953125,
0.662109375,
0.66845703125,
0.66357421875,
0.66845703125,
0.6787109375,
0.51953125,
0.67822265625
],
[
0.1070556640625,
0.66943359375,
0.476318359375,
0.66943359375,
0.476318359375,
0.68310546875,
0.1070556640625,
0.68310546875
],
[
0.1070556640625,
0.68310546875,
0.47802734375,
0.68310546875,
0.47802734375,
0.6962890625,
0.1070556640625,
0.6962890625
],
[
0.51953125,
0.681640625,
0.91259765625,
0.681640625,
0.91259765625,
0.69580078125,
0.51953125,
0.69580078125
],
[
0.10540771484375,
0.69580078125,
0.36767578125,
0.6962890625,
0.36767578125,
0.7109375,
0.10540771484375,
0.71044921875
],
[
0.51953125,
0.69580078125,
0.9111328125,
0.69580078125,
0.9111328125,
0.708984375,
0.51953125,
0.708984375
],
[
0.0931396484375,
0.7119140625,
0.4794921875,
0.7119140625,
0.4794921875,
0.7265625,
0.0931396484375,
0.7265625
],
[
0.521484375,
0.71044921875,
0.91259765625,
0.71044921875,
0.91259765625,
0.7236328125,
0.521484375,
0.7236328125
],
[
0.10540771484375,
0.7265625,
0.4794921875,
0.7265625,
0.4794921875,
0.74072265625,
0.10540771484375,
0.74072265625
],
[
0.51953125,
0.7236328125,
0.91259765625,
0.7236328125,
0.91259765625,
0.7373046875,
0.51953125,
0.7373046875
],
[
0.1070556640625,
0.74072265625,
0.47802734375,
0.74072265625,
0.47802734375,
0.75390625,
0.1070556640625,
0.75390625
],
[
0.51953125,
0.7373046875,
0.91259765625,
0.7373046875,
0.91259765625,
0.75048828125,
0.51953125,
0.75048828125
],
[
0.10540771484375,
0.7548828125,
0.438720703125,
0.7548828125,
0.438720703125,
0.76806640625,
0.10540771484375,
0.76806640625
],
[
0.51953125,
0.75,
0.91259765625,
0.75048828125,
0.91259765625,
0.76513671875,
0.51953125,
0.7646484375
],
[
0.09478759765625,
0.76953125,
0.476318359375,
0.76953125,
0.476318359375,
0.78271484375,
0.09478759765625,
0.78271484375
],
[
0.521484375,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.521484375,
0.77880859375
],
[
0.1070556640625,
0.78515625,
0.476318359375,
0.78515625,
0.476318359375,
0.798828125,
0.1070556640625,
0.798828125
],
[
0.51953125,
0.77978515625,
0.91015625,
0.77880859375,
0.91015625,
0.79052734375,
0.51953125,
0.79150390625
],
[
0.521484375,
0.79248046875,
0.9140625,
0.79248046875,
0.9140625,
0.80615234375,
0.521484375,
0.80615234375
],
[
0.1070556640625,
0.798828125,
0.476318359375,
0.798828125,
0.476318359375,
0.8125,
0.1070556640625,
0.8125
],
[
0.521484375,
0.80615234375,
0.9111328125,
0.80615234375,
0.9111328125,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.1070556640625,
0.81298828125,
0.42724609375,
0.81298828125,
0.42724609375,
0.8271484375,
0.1070556640625,
0.8271484375
],
[
0.521484375,
0.82080078125,
0.91259765625,
0.82080078125,
0.91259765625,
0.833984375,
0.521484375,
0.833984375
],
[
0.51953125,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.51953125,
0.84765625
],
[
0.223876953125,
0.84228515625,
0.34228515625,
0.84228515625,
0.34228515625,
0.85693359375,
0.223876953125,
0.85693359375
],
[
0.51953125,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.51953125,
0.861328125
],
[
0.0882568359375,
0.861328125,
0.4794921875,
0.861328125,
0.4794921875,
0.875,
0.0882568359375,
0.875
],
[
0.51953125,
0.8623046875,
0.91259765625,
0.8623046875,
0.91259765625,
0.87548828125,
0.51953125,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.4794921875,
0.87548828125,
0.4794921875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.5185546875,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.5185546875,
0.8896484375
]
] |
[
"adaptation methods on pretrained GANs, which can be con-",
"are relevant to our study.As our work takes image and text",
"ditioned on both NFoV images and text guidances. While",
"guidances as conditions, we further review the related stud-",
"these methods show encouraging performance, the chal-",
"ies on conditional image generation",
"lenge remains regarding the usage of given NFoV images",
" and user-provided open vocabulary text guidances individ-",
"360-Degree Image Generation",
"ually or jointly for enhanced control in 360-degree image",
"Unlike general NFoV images, 360-degree image generation",
"generation. Specifically, the existing methods typically fall",
"requires to take the omni-directional continuityinto account.",
"shortinsynthesizingintricatevisual details as showninthe",
"Early studies, for example, (Sengupta et al. 2019) estimates",
"top part of Fig. 1, where the details are vague or missing with",
"a coarse 360-degree image from an NFoV image with in-",
"OmniDreamer compared to our approach, which are high-",
"verse rendering technique,which ignores such geometrical",
"lightedintheredboundingboxes.Moreover,thegenerated",
"continuity and generates 360-degree images lack of fine de-",
"images and user-provided text guidances tend to be incon-",
"tails. To address this, 3601IC (Akimoto et al. 2019) and SIG-",
" sistent, especially under an open-vocabulary setting, as de-",
"Ss (Hara, Mukuta, and Harada 2021) were proposed to im-",
"picted in the bottom part of Fig. 1 by comparing Text2Light",
"prove geometrical continuity by taking the intrinsic horizon-",
"and our solution.",
"tal cyclicity into consideration and encoding it as positional",
"In this study, we propose a novel autoregressive omni-",
"conditions to connect the two ends of 360-degree images in",
"aware generative network (AOG-Net) for generating 360-",
"equirectangular representations.EnvMapNet (Somanath and",
"degree images conditioned on open vocabulary text guid-",
"Kurz 2021) improves visual quality of the outpainted 360-",
"ances and given NFov images jointly or individually",
"degree images by introducing a projection loss and a cluster-",
"Overall, the generation is formulated as an autoregressive",
"ing loss for accurate lighting and shadowing.OmniDreamer",
" stochastic process to outpaint an incomplete 360-degree im-",
"(Akimoto, Matsuo, and Aoki 2022) was further developed",
"age progressively, in which each step outpaints a local re",
"by leveraging the Taming-Transformer (Esser, Rombach,",
"gion under its corresponding NFoV view. This autoregres",
"and Ommer 2021), where a circular inference scheme was",
" sive scheme not only allows for deriving finer-grained and",
"introduced to fit the intrinsic horizontal cyclicity for 360-",
"prompt-consistent patterns by dynamically observing and",
"degree image synthesis, conditioned on provided NFoV im-",
"adjusting the generation process but also offers users greater",
"ages, yielding diverse and photo-realistic results. However,",
"flexibility to modify or introduce new conditions through-",
"OmniDreamer is limited to a single condition where only",
"out the generation process. Furthermore, a global-local con-",
"an initial NFo image is accepted, while the controllabil",
"ditioning mechanism is devised to comprehensively formu-",
"ityoftheoverall synthesisprocessislimited.ImmenseGAN",
"late the outpainting guidance for each autoregression step.",
"(Dastjerdi et al. 2022) aims for finer controllability over the",
"Text prompts, omni-visual cues, NFoV inputs and omni-",
"outpainting by introducing a text guidances to fine-tune a",
"geometry are encoded and further formulated with cross-",
"generative model with a large-scale private text-image pair",
"attention based transformers into a global stream and a lo-",
"dataset. Due to the lack of public text-image paired dataset,",
"cal stream for a conditioned generative backbone model",
"Text2Light (Chen,Wang,and Liu 2022)introduces a zero-",
"This study further explores the potential to leverage large-",
"shot text-guided 360-degree image synthesis pipeline with-",
"scale modelsforthe conditional encoder and thegenera-",
"out using initial NFoV images, in which a pre-trained CLIP",
"tive prior, which helps complete the generation using open-",
"modelis adopted (Radford etal.2021)",
"vocabulary prompts.Comprehensive experiments on two",
"However, the existing methods typically fall short in syn-",
"commonlyused360-degreeimagedatasetsforbothindoor",
"thesizing intricate visual details and inconsistencies can",
"and outdoor settings demonstrate the state-of-the-art perfor-",
"be observed between generated images and user-provided",
"mance of our proposed method.",
"text guidances,especially under an open-vocabulary setting,",
"In summary, the key contributions of this study are:",
"which demands further mechanisms to address these issues.",
": A novel autoregressive outpainting approach is proposed",
"Image Outpainting",
"to produce photo-realistic 360-degree images by dynami-",
"cally adjusting the generation process for improved finer-",
"Image outpainting, a fundamental task in computer vision,",
"grained details and prompt-consistency.",
"focuses on expanding the unknown regions outside the pri-",
" A global-local conditioning mechanism is devised to",
"mary known content. Unlike inpainting, outpainting may not",
"formulate the guidance encompassing open-vocab text",
"be able to leverage information from pixels adjacent to the",
"guidances, omni-visual cues, NFoV inputs and omni-",
"unknown area (Sabini and Rusak 2018; Hoorick 2020; Wang",
"geometry with cross-attention based transformers.",
"et al. 2019), as seen in inpainting methods. In (Wang et al.",
": Comprehensive experiments were conducted on two",
"2019), the semantic information of incomplete images was",
"commonly used benchmarks, demonstrating the state-of",
"utilizedtoguideaGANforoutpainting.In(Yaoetal.2022)",
"a query-based outpainting method was proposed, where an",
"the-art performance of AOG-Net in both indoor and out-",
"image is divided into small patches and the patches with un-",
"door settings with as few as 40 training samples.",
"known pixels are completed by taking the conditions from",
"both distant and neighbour patches into account. In an iter-",
"Related Work",
"ative manner, (Gardias, Arthur, and Sun 2020) extends one",
"We first review the studies in both the field of 360-degree",
"side of the a regular image for outpainting step by step, using",
"image generation and the field of image outpainting which",
"the context of the past generation as guidance. (Lu, Chang,"
] |
{
"arxiv_id": "2309.03467",
"language": "en",
"page_idx": 1
}
|
||
1904.04196-en-1
|
1904.04196-en-1.jpg
|
work tackles the similar but in RGB domain and offers an iterative mesh fitting using a differentiable renderer[18]. This improves once estimated mesh parameters using the gradient information at testing. Our framework also allows indirect 2D supervisions (2D segmentation masks, keypoints) instead of using 3D mesh data.
While 2D keypoint detection is well established in the RGB domain [43], estimating 3D keypoints from RGB images is less trivial. Recently, several methods were developed [68, 34, 4, 14]. While directly lifting 2D estimations to 3D was attempted in [68], 2.5D depth maps are estimated as clues for 3D lifting in state-of-the-art techniques [4,14]. In this paper, we exploit a deformable 3D hand mesh model, which inherently offers a full description of both hand shapes and articulations, 3D priors for recovering depths, and self-data augmentation. Different from purely generative optimization methods e.g.[34],we propose a method based on a neural renderer and CNNs.
Our contribution is in implementing a full 3D dense hand pose estimator (DHPE) using single RGB images via a neural $\hat{\text{renderer. Our technical contributions are largely threefold:}}$
1) DHPE is composed of convolutional layers that first estimate 2D evidences (RGB features, 2D keypoints) from an RGB image and then estimate the 3D mesh model parameters. Since RGB/mesh pairs are lacking, the network is learned to fit the 3D model using 2D segmentation masks and skeletons as supervision, similar to recent work on human body pose estimation [57, 16, 35], via neural renderer. The dense shape estimation helps improve the pose estimation.
2) At the testing time, we iteratively refine the initial 3D mesh estimation using the gradients. The gradients are computed over self-supervisions by comparing estimated 3D meshes to predicted 2D evidences, as ground-truth labels are not available at testing. While previous work[57] fixed 2D skeletons/segmentation masks during the refinement step, we recursively improve 2D skeletons and exploit the improved skeletons and feature similarities for 2D masks.
3) To further deal with limited annotated training data, especially for diverse shapes and view-points, we supply fitted meshes and their 2D projected RGB maps by varying shapes and view-points, to fine-tuning the network. Purely synthetic data imposes a synthetic-real domain gap, and annotating 3D meshes or 2D segmentation masks of real images is difficult. Once the model is successfully fitted to input RGB images, its meshes, i.e. shapes and poses are close-to-real.
The proposed method can also be seen as a hybrid method [23, 63, 55] combining merits of discriminative and generative approaches.
## $2. \textbf{Related work}$
Hand pose estimation. There have been several categories of works predicting hand poses. They vary according to types of output (3D mesh, 3D skeletal and 2D skeletal representations) and input (RGB and depth maps).
$3D\textit{hand model fitting methods. In [ 54, 53, 41, 51, 49, 48, }$
40, 33], 3D hand mesh representation is used to describe hand poses. To recover 3D meshes from 2D images, most methods rely on complex optimization methods or use point clouds to fit their 3D mesh models. One recent method[34] takes RGB images as input and formulates the framework that runs in real-time. Their 3D mesh model is matched to estimated 2D skeletons. However, 2D skeleton is coarse and overall 3D fitting fails when the 2D skeleton estimation is not accurate.
Recovering 3D skeletal representations from depth $images.$ Estimating 3D hand skeletons from depth images has shown promising accuracies [64]. The problem is better set since the depth input inherently offers rich 3D information [52].
Further, deep learning meth-
ods[65, 30, 50, 58, 30, 64, 28, 38, 64, 2] have been successfully applied along with the million-scale hand pose dataset [65] in this domain. Recently, Malik et al. [26] proposed to estimate both 3D skeletal and mesh representations and have obtained improved accuracy in 3D skeletal estimation. They collect a new million-scale data set for depth map and 3D mesh pairs to train the network in a fully supervised manner.
Estimating the 2D skeletal from RGB. Recovering 2D skeletal representations from RGB images has been greatly improved[43]. Compared to the depth domain, real RGB data is scarce and automatic pose annotation of RGB images still remains challenging. Simon et al. achieved promising results by mixing real data with a large amount of synthetic data[43]. Further accuracy improvement has been obtained by the automatic annotation using label consistency in a multiple camera studio[15].
Recovering the 3D skeletal from RGB. The problem has inherent uncertainties rising from 2D-to-3D mapping [14,4, 68]. The earlier work[68] attempts to learn the direct mapping from RGB images to 3D skeletons. Recent methods[4, 14] have shown the state-of-the-art accuracy by implicitly reconstructing depth images i.e. 2.5D representations, and estimating the 3D skeletal based on them. We instead use a 3D mesh model in a hybrid way to tackle the problem.
3D representations for hand pose estimation. Different from human face or body, hand poses exhibit a much wider scope of viewpoints, given an articulation and shape. It is hard to define a canonical viewpoint (e.g. a frontal view). To relieve the issue, intermediate 3D representations such as projected point clouds [9], D-TSDF [10], voxels [28] or point sets[8] have been proposed. Though these approaches have shown promising results, they are limited such that they cannot re-generate full 3D information.
3D recovery in other domains. 3D reconstruction has a long history in the feld. Some latest approaches using neural networks are discussed below.
3D voxel reconstruction. Voxel representations are useful to describe 3D aspects of objects [62,56]. Yan et al.[62] proposed a differentiable way to reconstruct voxels given $\hat{\text{only 2D inputs. Also, Tulsiani et al.[56] proposed a method}}$ to use multiple images caputred at different viewpoints to accurately reconstruct voxels. However, the voxel
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.080078125,
0.106689453125,
0.4697265625,
0.106689453125,
0.4697265625,
0.12060546875,
0.080078125,
0.12060546875
],
[
0.50244140625,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.50244140625,
0.12060546875
],
[
0.07843017578125,
0.11932373046875,
0.4697265625,
0.121826171875,
0.4697265625,
0.1351318359375,
0.07843017578125,
0.1331787109375
],
[
0.50341796875,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1331787109375,
0.50341796875,
0.1331787109375
],
[
0.08087158203125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.147705078125,
0.08087158203125,
0.147705078125
],
[
0.50341796875,
0.136962890625,
0.8896484375,
0.136962890625,
0.8896484375,
0.147705078125,
0.50341796875,
0.147705078125
],
[
0.07843017578125,
0.1483154296875,
0.4697265625,
0.149658203125,
0.4697265625,
0.1641845703125,
0.07843017578125,
0.1634521484375
],
[
0.50341796875,
0.1514892578125,
0.890625,
0.1514892578125,
0.890625,
0.1622314453125,
0.50341796875,
0.1622314453125
],
[
0.08087158203125,
0.1666259765625,
0.3544921875,
0.1666259765625,
0.3544921875,
0.1767578125,
0.08087158203125,
0.1767578125
],
[
0.50341796875,
0.1666259765625,
0.8896484375,
0.1666259765625,
0.8896484375,
0.1767578125,
0.50341796875,
0.1767578125
],
[
0.096435546875,
0.1787109375,
0.4697265625,
0.1787109375,
0.4697265625,
0.1932373046875,
0.096435546875,
0.1932373046875
],
[
0.50341796875,
0.179931640625,
0.890625,
0.179931640625,
0.890625,
0.1900634765625,
0.50341796875,
0.1900634765625
],
[
0.080078125,
0.1932373046875,
0.468017578125,
0.1932373046875,
0.468017578125,
0.20703125,
0.080078125,
0.20703125
],
[
0.5009765625,
0.1925048828125,
0.89208984375,
0.1912841796875,
0.89208984375,
0.205810546875,
0.5009765625,
0.20703125
],
[
0.080078125,
0.2083740234375,
0.468017578125,
0.2083740234375,
0.468017578125,
0.2215576171875,
0.080078125,
0.2215576171875
],
[
0.50341796875,
0.208984375,
0.5888671875,
0.208984375,
0.5888671875,
0.2197265625,
0.50341796875,
0.2197265625
],
[
0.080078125,
0.222900390625,
0.468017578125,
0.222900390625,
0.468017578125,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.51708984375,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.236083984375,
0.51708984375,
0.236083984375
],
[
0.07843017578125,
0.236083984375,
0.4697265625,
0.2374267578125,
0.4697265625,
0.250732421875,
0.07843017578125,
0.25
],
[
0.50341796875,
0.236083984375,
0.89208984375,
0.236083984375,
0.89208984375,
0.25,
0.50341796875,
0.25
],
[
0.07843017578125,
0.25,
0.4697265625,
0.250732421875,
0.4697265625,
0.26513671875,
0.07843017578125,
0.264404296875
],
[
0.50244140625,
0.250732421875,
0.89208984375,
0.250732421875,
0.89208984375,
0.264404296875,
0.50244140625,
0.264404296875
],
[
0.08087158203125,
0.267578125,
0.46728515625,
0.267578125,
0.46728515625,
0.27783203125,
0.08087158203125,
0.27783203125
],
[
0.50244140625,
0.26513671875,
0.89208984375,
0.26513671875,
0.89208984375,
0.279052734375,
0.50244140625,
0.279052734375
],
[
0.080078125,
0.27978515625,
0.468017578125,
0.27978515625,
0.468017578125,
0.29345703125,
0.080078125,
0.29345703125
],
[
0.50244140625,
0.27978515625,
0.89208984375,
0.27978515625,
0.89208984375,
0.29345703125,
0.50244140625,
0.29345703125
],
[
0.08087158203125,
0.29541015625,
0.468017578125,
0.29541015625,
0.468017578125,
0.30615234375,
0.08087158203125,
0.30615234375
],
[
0.50244140625,
0.29345703125,
0.89208984375,
0.29345703125,
0.89208984375,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.08087158203125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.32080078125,
0.08087158203125,
0.32080078125
],
[
0.50341796875,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.080078125,
0.323974609375,
0.4697265625,
0.323974609375,
0.4697265625,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.50341796875,
0.322509765625,
0.89208984375,
0.322509765625,
0.89208984375,
0.3359375,
0.50341796875,
0.3359375
],
[
0.07843017578125,
0.33837890625,
0.451904296875,
0.3359375,
0.451904296875,
0.349853515625,
0.07843017578125,
0.3515625
],
[
0.50341796875,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.09722900390625,
0.354248046875,
0.46728515625,
0.354248046875,
0.46728515625,
0.3642578125,
0.09722900390625,
0.3642578125
],
[
0.50244140625,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.080078125,
0.380615234375
],
[
0.50244140625,
0.365478515625,
0.890625,
0.365478515625,
0.890625,
0.37890625,
0.50244140625,
0.37890625
],
[
0.080078125,
0.38330078125,
0.463134765625,
0.38330078125,
0.463134765625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.380126953125,
0.8935546875,
0.380126953125,
0.8935546875,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.096435546875,
0.396484375,
0.4697265625,
0.396484375,
0.4697265625,
0.410400390625,
0.096435546875,
0.410400390625
],
[
0.51953125,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.51953125,
0.40771484375
],
[
0.08087158203125,
0.412353515625,
0.46728515625,
0.412353515625,
0.46728515625,
0.42236328125,
0.08087158203125,
0.42236328125
],
[
0.5048828125,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.42041015625,
0.5048828125,
0.42041015625
],
[
0.07843017578125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.07843017578125,
0.438232421875
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.080078125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.439453125,
0.890625,
0.439453125,
0.890625,
0.449462890625,
0.50341796875,
0.449462890625
],
[
0.080078125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.5009765625,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.46533203125,
0.5009765625,
0.46533203125
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.466552734375,
0.89208984375,
0.466552734375,
0.89208984375,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.080078125,
0.48291015625,
0.468017578125,
0.48291015625,
0.468017578125,
0.496826171875,
0.080078125,
0.496826171875
],
[
0.5009765625,
0.479248046875,
0.89208984375,
0.479736328125,
0.89208984375,
0.494384765625,
0.5009765625,
0.49365234375
],
[
0.080078125,
0.496826171875,
0.464111328125,
0.496826171875,
0.464111328125,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.494384765625,
0.89208984375,
0.5087890625,
0.50244140625,
0.50830078125
],
[
0.096435546875,
0.51220703125,
0.46728515625,
0.51220703125,
0.46728515625,
0.52587890625,
0.096435546875,
0.52587890625
],
[
0.50341796875,
0.51123046875,
0.68896484375,
0.51123046875,
0.68896484375,
0.521484375,
0.50341796875,
0.521484375
],
[
0.07843017578125,
0.5244140625,
0.468017578125,
0.52587890625,
0.468017578125,
0.54052734375,
0.07843017578125,
0.5390625
],
[
0.5185546875,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.5185546875,
0.537109375
],
[
0.080078125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.080078125,
0.5546875
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.07843017578125,
0.5537109375,
0.4697265625,
0.5546875,
0.4697265625,
0.5693359375,
0.07843017578125,
0.568359375
],
[
0.5009765625,
0.55029296875,
0.89208984375,
0.5517578125,
0.89208984375,
0.56640625,
0.5009765625,
0.56494140625
],
[
0.08087158203125,
0.5712890625,
0.468017578125,
0.5712890625,
0.468017578125,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.50341796875,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.080078125,
0.583984375,
0.4697265625,
0.583984375,
0.4697265625,
0.59765625,
0.080078125,
0.59765625
],
[
0.5048828125,
0.58349609375,
0.8896484375,
0.58349609375,
0.8896484375,
0.59326171875,
0.5048828125,
0.59326171875
],
[
0.080078125,
0.5986328125,
0.4697265625,
0.5986328125,
0.4697265625,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.50341796875,
0.5966796875,
0.890625,
0.5966796875,
0.890625,
0.60693359375,
0.50341796875,
0.60693359375
],
[
0.08087158203125,
0.6142578125,
0.459228515625,
0.6142578125,
0.459228515625,
0.625,
0.08087158203125,
0.625
],
[
0.50244140625,
0.60986328125,
0.8935546875,
0.60986328125,
0.8935546875,
0.62353515625,
0.50244140625,
0.62353515625
],
[
0.09478759765625,
0.62744140625,
0.4697265625,
0.62744140625,
0.4697265625,
0.6416015625,
0.09478759765625,
0.6416015625
],
[
0.50341796875,
0.62353515625,
0.85888671875,
0.62353515625,
0.85888671875,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.080078125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.5185546875,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.5185546875,
0.65283203125
],
[
0.080078125,
0.6572265625,
0.4697265625,
0.6572265625,
0.4697265625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.5048828125,
0.65478515625,
0.890625,
0.65478515625,
0.890625,
0.66552734375,
0.5048828125,
0.66552734375
],
[
0.07843017578125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.07843017578125,
0.6845703125
],
[
0.50341796875,
0.66845703125,
0.89208984375,
0.66845703125,
0.89208984375,
0.681640625,
0.50341796875,
0.681640625
],
[
0.08087158203125,
0.6875,
0.46728515625,
0.6875,
0.46728515625,
0.69775390625,
0.08087158203125,
0.69775390625
],
[
0.50244140625,
0.681640625,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.080078125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.080078125,
0.7119140625
],
[
0.50341796875,
0.69775390625,
0.890625,
0.69775390625,
0.890625,
0.70849609375,
0.50341796875,
0.70849609375
],
[
0.08087158203125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50244140625,
0.71044921875,
0.89208984375,
0.71044921875,
0.89208984375,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.080078125,
0.72900390625,
0.4599609375,
0.72900390625,
0.4599609375,
0.7421875,
0.080078125,
0.7421875
],
[
0.50341796875,
0.724609375,
0.89208984375,
0.724609375,
0.89208984375,
0.73779296875,
0.50341796875,
0.73779296875
],
[
0.09478759765625,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.09478759765625,
0.7568359375
],
[
0.50244140625,
0.7392578125,
0.89208984375,
0.7392578125,
0.89208984375,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.080078125,
0.7568359375,
0.46728515625,
0.7568359375,
0.46728515625,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.50341796875,
0.75390625,
0.79736328125,
0.75244140625,
0.79736328125,
0.7666015625,
0.50341796875,
0.767578125
],
[
0.080078125,
0.77294921875,
0.2320556640625,
0.77294921875,
0.2320556640625,
0.787109375,
0.080078125,
0.787109375
],
[
0.50341796875,
0.771484375,
0.89208984375,
0.771484375,
0.89208984375,
0.78515625,
0.50341796875,
0.78515625
],
[
0.50341796875,
0.7861328125,
0.89208984375,
0.7861328125,
0.89208984375,
0.7998046875,
0.50341796875,
0.7998046875
],
[
0.080078125,
0.80322265625,
0.214111328125,
0.80322265625,
0.214111328125,
0.8173828125,
0.080078125,
0.8173828125
],
[
0.50341796875,
0.8017578125,
0.74658203125,
0.8017578125,
0.74658203125,
0.8125,
0.50341796875,
0.8125
],
[
0.5185546875,
0.81494140625,
0.89208984375,
0.81494140625,
0.89208984375,
0.8291015625,
0.5185546875,
0.8291015625
],
[
0.08087158203125,
0.82763671875,
0.4697265625,
0.82763671875,
0.4697265625,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.50341796875,
0.83056640625,
0.8896484375,
0.83056640625,
0.8896484375,
0.84130859375,
0.50341796875,
0.84130859375
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84228515625,
0.468017578125,
0.85693359375,
0.07843017578125,
0.85595703125
],
[
0.50244140625,
0.8447265625,
0.890625,
0.84326171875,
0.890625,
0.85693359375,
0.50244140625,
0.85791015625
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.8720703125
],
[
0.50341796875,
0.85986328125,
0.890625,
0.85986328125,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.410888671875,
0.8720703125,
0.410888671875,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.096435546875,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.096435546875,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.75341796875,
0.88720703125,
0.75341796875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.74609375,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.74609375,
0.90087890625
]
] |
[
"work tackles the similar but in RGB domain and offers an",
"40, 33], 3D hand mesh representation is used to describe",
"iterative mesh fitting using a differentiable renderer [18].",
"hand poses. To recover 3D meshes from 2D images, most",
"This improves once estimated mesh parameters using",
"methodsrelyoncomplexoptimizationmethodsorusepoint",
"the gradient information at testing.Our framework also",
"cloudstofittheir3Dmeshmodels.Onerecentmethod[34]",
"allows indirect 2D supervisions (2D segmentation masks,",
"takesRGB images asinputandformulates theframework",
"keypoints)insteadof using3Dmeshdata.",
"thatrunsin real-time.Their 3Dmeshmodelis matched to",
"While 2D keypoint detection is well established in",
"estimated2Dskeletons.However,2Dskeletoniscoarseand",
"the RGB domain [43], estimating 3D keypoints from",
"overall 3D fitting fails when the 2D skeleton estimation is",
"RGB images is less trivial.Recently, several methods",
"notaccurate.",
"were developed [68, 34, 4, 14]. While directly lifting 2D",
"Recovering 3D skeletal representations from depth",
"estimations to 3D was attempted in [68], 2.5Ddepth maps",
"images. Estimating 3D hand skeletons from depth images",
"are estimated as clues for 3D lifting in state-of-the-art",
"has shown promising accuracies [64]. The problem is",
"techniques[4,141.Inthis paper,weexploit adeformable",
"better set since the depth input inherently offers rich",
"3D hand mesh model, which inherently offers a full",
"3D information [52].Further, deep learning meth-",
"description ofbothhand shapesandarticulations,3Dpriors",
"ods [65, 30, 50, 58, 30, 64, 28, 38, 64, 2] have been success-",
"forrecovering depths,andself-data augmentation.Different",
"fully applied along with the million-scale hand pose dataset",
"from purely generative optimization methods e.g.[34l, we",
"[65] in this domain. Recently, Malik et al. [26] proposed to",
"propose a method based on a neural renderer and CNNs.",
"estimateboth3Dskeletalandmeshrepresentationsandhave",
"Ourcontributionisinimplementingafull3Ddensehand",
"obtained improved accuracy in 3D skeletal estimation.They",
"pose estimator (DHPE) using single RGB images via a neural",
"collect a new million-scale data set for depth map and 3D",
"renderer. Ourtechnical contributions arelargely threefold:",
"mesh pairs to train the network in a fully supervised manner.",
"1) DHPE is composed of convolutional layers that first",
"Estimating the 2D skeletal from RGB. Recovering 2D",
"estimate2Devidences(RGBfeatures.2Dkeypoints)from",
"skeletalrepresentationsfromRGBimageshasbeengreatly",
"an RGB image and then estimate the 3D mesh model pa-",
"improved [43]. Compared to the depth domain, real RGB",
"rameters. Since RGB/mesh pairs are lacking, the network is",
"datais scarce andautomaticposeannotation of RGBimages",
"learned to fit the 3D model using 2D segmentation masks and",
"still remains challenging. Simon et al. achieved promising",
"skeletons as supervision, similar to recent work on human",
"results by mixing real data with a large amount of synthetic",
"body pose estimation [57, 16, 35], via neural renderer. The",
"data [43j. Further accuracy improvement has been obtained",
"dense shape estimation helps improve the pose estimation.",
"by the automatic annotation using label consistency in a",
"2) At the testing time, we iteratively refine the initial",
"multiple camerastudio[15].",
"3D mesh estimation using the gradients. The gradients are",
"Recovering the 3D skeletal from RGB. The problem has",
"computed over self-supervisions by comparing estimated",
"inherent uncertainties rising from 2D-to-3D mapping [14, 4,",
"3D meshes to predicted 2D evidences, as ground-truth labels",
"68]. The earlier work [68] attempts to learn the direct map-",
"arenotavailableattesting.Whileprevious work[571fixed",
"ping from RGB images to 3D skeletons. Recent methods [4,",
"2D skeletons/segmentation masks during the refinement",
"I4lhaveshownthestate-of-the-artaccuracybyimplicitly",
"step, we recursively improve 2D skeletons and exploit the",
"reconstructingdepthimagesi.e.2.5Drepresentations.and",
"improved skeletonsandfeature similaritiesfor2Dmasks.",
"estimating the 3D skeletal based on them. We instead use a",
"3) To further deal with limited annotated training data, es-",
"3D mesh model in a hybrid way to tackle the problem.",
"pecially for diverse shapes and view-points, we supply fitted",
"3D representations for hand pose estimation. Different",
"meshes and their 2D projected RGB maps by varying shapes",
"fromhumanfaceorbody,handposesexhibitamuchwider",
"and view-points, to fine-tuning the network. Purely synthetic",
"scope of viewpoints, given an articulation and shape. It is",
"data imposes a synthetic-real domaingap,and annotating",
"hard to define a canonical viewpoint (e.g. a frontal view).",
"3D meshes or 2D segmentation masks of real images is",
"Torelievetheissue,intermediate3D representations such",
"difficult. Once the model is successfully fitted to input RGB",
"as projected point clouds [9], D-TSDF [10], voxels [28] or",
"images, its meshes, i.e. shapes and poses are close-to-real.",
"point sets [8] have been proposed. Though these approaches",
"The proposed method can also be seen as a hybrid",
"have shown promising results, they are limited such that",
"method [23, 63, 55] combining merits of discriminative and",
"they cannot re-generate full 3D information.",
"generative approaches.",
"3D recovery in other domains.3D reconstruction has",
"a long history in the field. Some latest approaches using",
"2.Related work",
"neural networks are discussed below.",
"3D voxel reconstruction. Voxel representations are useful",
"Hand pose estimation. There have been several categories",
"todescribe3Daspectsofobjects[62,56].Yanetal.[62]",
"of works predicting hand poses. They vary according to",
"proposed a differentiable way to reconstruct voxels given",
"types of output (3D mesh, 3D skeletal and 2D skeletal",
"only2Dinputs.Also,Tulsianietal.[56lproposed amethod",
"representations) and input (RGB and depth maps).",
"to use multiple images caputred at different viewpoints",
"3D hand model fitting methods. In [54, 53, 41, 51, 49, 48",
"to accurately reconstruct voxels.",
"However, the voxel"
] |
{
"arxiv_id": "1904.04196",
"language": "en",
"page_idx": 1
}
|
||
1904.04196-en-3
|
1904.04196-en-3.jpg
|
<FigureHere>
Figure 3: A 2D evidence estimation example. (a) input
image x,(b) ground-truth 2D segmentation mask m of x, (c) 2D skeletal position heat map of the finger tip of middle finger overlaid on x, and (d) masked image x©m.
Table 1: Notational summary
<table>
<tbody>
<tr>
<td>$y$ $FDHPE$</td>
<td>$\mathbf{y}=(\mathbf{J},\mathbf{m})$ ${\mathrm{dense~hand~pose~estimator~}(f^{DHPE}=f^{Proj}\circ f^{HME})}$</td>
</tr>
<tr>
<td>$\overline{fProj}$ 1</td>
<td>projection operator $(f^{Proj}$ Kell $=\left|:\frac{1}{2}\right|$</td>
</tr>
<tr>
<td>$\overline{fReg}$</td>
<td>3D skeleton regressor</td>
</tr>
<tr>
<td>$\overline{fRen}$</td>
<td>renderer</td>
</tr>
</tbody>
</table>
By construction, the projection operator respects the underlying camera (via $f^{Ren})$ and hand shape (via $f^{Reg})$ geometry and it is held frxed throughout the entire training process. This facilitates the training of $f^{HME}$ indirectly via training $f^{DHPE}.$ However, even under this setting, the problem still remains challenging as estimating a 3D mesh given an RGB image is a seriously ill-posed problem. Adopting recent human body pose estimation approaches [35,57],we further stratify learning of $f^{DHPE}:\mathcal{X}\to\mathcal{Y}$ by decomposing $f^{HME}:\mathcal{X}\to\mathcal{V}$ into a 2D evidence estimator $f^E2D:\bar{\mathcal{X}}\to\bar{\mathcal{Z}}$ and a 3D mesh estimator $f^{E3D}:\mathcal{Z}\to\mathcal{V}.$ Our 2D evidence $\mathbf{z}\in Z$ consists of a 42-dimensional 2D skeletal joint position vector $\mathbf{j}_{2D}$ (21 positions $\times2;$ as in [4,14]) and a 2,048- dimensional 2D feature vector $F(\mathbf{x})$ (Eq. 2). The remainder of this section provides details of these two estimators.
3.1. 2D evidence estimator $f^E2D=(F,f^{J2D})$
Silhouettes (or foreground masks) and 2D skeletons have been widely used as the mid-level cues for estimating 3D body meshes[17,35,57]. However, for hands, accurately estimating foreground masks from RGB images is challenging due to cluttered backgrounds [21,68]. We observed that naively applying the state-of-the-art foreground segmentation algorithms (e.g.[68]) often misses fine details, especially along the narrow finger regions (see Fig. 4 for examples) and this can significantly degrade the performance of the subsequent mesh estimation step. We bypass this challenge by learning instead, a 2D foreground feature extractor $\tilde{F} {: }$ $\tilde{F} ( \mathbf{x} )$ encapsulates the textural and shape information of the foreground regions in x.
Our 2D feature extractor $F$ is trained to focus on the foreground regions by minimizing the deviation between the features extracted from the entire image x and the foreground region x◉m extracted via the ground-truth mask m (see Fig. 3d): The training loss (per data point) for $F$ is given as
(2)
$$L_{Feat}(F)=\|F(\mathbf{x})-F(\mathbf{x}\odot\mathbf{m})\|_2^2,$$
where © denotes element-wise multiplication. $F$ employs the ResNet-50[13] architecture whose output is a 2,048-dimensional vector.
Estimating 2D skeletal joints from an RGB image has been well studied. Similarly to[4, 14], we embed the stateof-the-art 2D pose estimation network $f^{J2D}$ [60,68] into our framework. The initial weights provided by the authors of [68] are refined based on the 2D joint estimation loss:
(3)
$$L_{J2D}(f^{J2D})=\|f^{J2D}(\mathbf{x})-\mathbf{j}_{2DHeat}\|_2^2.$$
The output of $f^{J2D}$ is a matrix of size 21×32×32 encoding 21 heat maps (per joint) of size 32×32. j$_{2DHeat}$ is the ground-truth heat map. Given the estimated heat maps, 2D skeleton joints $\mathbf{j}_{2D}$ are extracted by finding the maximum for each joint. Figure 3(b) shows an example of 2D evidence estimation.
# 3.2. 3D mesh estimator $f^E3D=(f^{J3D},f^{Ref})$
Taking the 2D evidence z (2,048-dimensional feature vector $F(\mathbf{x})$ and 21$\times3$-dimensional 2D skeleton j$_2D)$ as input, the 3D mesh estimation network $f^{J3D}$ constructs parameters of a deformable hand mesh model and a camera model. We use the MANO model representing a hand mesh based on 45-dimensional pose parameters p and 10-dimensional shape parameters s [41]. The original MANO framework uses only 6-dimensional (principal component analysis) PCA subspace of p for computational efficiency. However, we empirically observed that to cover a variety of hand poses, all 45-dimensional features are required: We use the linear blend skinning formulation as in the SMPL model [24].
Given the MANO parameters p and s, a mesh v is synthesized by conditioning them on the hypothesized camera $\mathbf{c}$, which comprises of the 3D rotation (in quaternion space) $\mathbf{c}_q\in\mathbb{R}^4$, scale $\mathbf{c}_s\in\mathbb{R}$, and translation $\mathbf{c}_t\in\mathbb{R}^3:$ An initial mesh is constructed by combining MANO's PCA basis vectors using the shape parameter s, rotating the bones according to the pose parameter p, and deforming the resulting surface via linear blend skinning. The final mesh v is then obtained by globally rotating, scaling, and translating the initial mesh according to $\mathbf{c}_q,\mathbf{c}_s$, and $\mathbf{c}_t$, respectively. The entire mesh generation process is differentiable. Under this model, our 3D mesh estimation network $f^{J3D}$ is implemented as a single fully-connected layer of 2153×63 weights and it estimates a 63-dimensional mesh parameter:
$$\mathbf{h}=[\mathbf{p},\mathbf{s},\mathbf{c}_q,\mathbf{c}_s,\mathbf{c}_t]^\top.$$
(4)
Iterative mesh refinement via back-projection. Adopting Kanazawa et al.'s approach[16], instead of estimating h directly, we iteratively refine the initial mesh estimate $\mathbf{h}(0)$ by recursively performing regression on the parameter offset $\Delta\mathbf{h}:$ At iteration $t,f^J3D$ takes z and the current mesh
|
[
[
0.176513671875,
0.08837890625,
0.370849609375,
0.0889892578125,
0.370849609375,
0.10418701171875,
0.176513671875,
0.1029052734375
],
[
0.1478271484375,
0.117431640625,
0.4345703125,
0.111083984375,
0.4345703125,
0.1287841796875,
0.1478271484375,
0.1351318359375
],
[
0.814453125,
0.1142578125,
0.83837890625,
0.1319580078125,
0.81884765625,
0.146484375,
0.7958984375,
0.1287841796875
],
[
0.1446533203125,
0.1300048828125,
0.459228515625,
0.126953125,
0.459228515625,
0.14453125,
0.1446533203125,
0.147705078125
],
[
0.63134765625,
0.1256103515625,
0.64892578125,
0.1256103515625,
0.64892578125,
0.1395263671875,
0.63134765625,
0.1395263671875
],
[
0.10455322265625,
0.1458740234375,
0.131591796875,
0.1458740234375,
0.131591796875,
0.1572265625,
0.10455322265625,
0.1572265625
],
[
0.1004638671875,
0.1307373046875,
0.131591796875,
0.1307373046875,
0.131591796875,
0.1395263671875,
0.1004638671875,
0.1395263671875
],
[
0.146240234375,
0.147705078125,
0.396240234375,
0.1458740234375,
0.396240234375,
0.1590576171875,
0.146240234375,
0.1610107421875
],
[
0.11273193359375,
0.1634521484375,
0.1192626953125,
0.1634521484375,
0.1192626953125,
0.1673583984375,
0.11273193359375,
0.1673583984375
],
[
0.1478271484375,
0.160400390625,
0.289306640625,
0.160400390625,
0.289306640625,
0.173583984375,
0.1478271484375,
0.173583984375
],
[
0.10540771484375,
0.175537109375,
0.1258544921875,
0.175537109375,
0.1258544921875,
0.18310546875,
0.10540771484375,
0.18310546875
],
[
0.5537109375,
0.175537109375,
0.57421875,
0.175537109375,
0.57421875,
0.1881103515625,
0.5537109375,
0.1881103515625
],
[
0.64208984375,
0.175537109375,
0.6640625,
0.175537109375,
0.6640625,
0.1881103515625,
0.64208984375,
0.1881103515625
],
[
0.73388671875,
0.175537109375,
0.75244140625,
0.175537109375,
0.75244140625,
0.1881103515625,
0.73388671875,
0.1881103515625
],
[
0.82177734375,
0.175537109375,
0.8408203125,
0.175537109375,
0.8408203125,
0.1881103515625,
0.82177734375,
0.1881103515625
],
[
0.50244140625,
0.1925048828125,
0.89208984375,
0.1925048828125,
0.89208984375,
0.20703125,
0.50244140625,
0.20703125
],
[
0.1478271484375,
0.1767578125,
0.205078125,
0.1767578125,
0.205078125,
0.1881103515625,
0.1478271484375,
0.1881103515625
],
[
0.09478759765625,
0.1962890625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.212158203125,
0.09478759765625,
0.2108154296875
],
[
0.080078125,
0.210205078125,
0.468017578125,
0.210205078125,
0.468017578125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.20703125,
0.89208984375,
0.20703125,
0.89208984375,
0.2203369140625,
0.50244140625,
0.2203369140625
],
[
0.080078125,
0.2266845703125,
0.468017578125,
0.2266845703125,
0.468017578125,
0.240478515625,
0.080078125,
0.240478515625
],
[
0.50244140625,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.07843017578125,
0.238037109375,
0.4697265625,
0.238037109375,
0.4697265625,
0.255615234375,
0.07843017578125,
0.255615234375
],
[
0.50244140625,
0.236083984375,
0.8349609375,
0.236083984375,
0.8349609375,
0.25,
0.50244140625,
0.25
],
[
0.0767822265625,
0.25244140625,
0.471435546875,
0.25244140625,
0.471435546875,
0.270751953125,
0.0767822265625,
0.270751953125
],
[
0.07843017578125,
0.26953125,
0.4697265625,
0.270751953125,
0.4697265625,
0.2841796875,
0.07843017578125,
0.28271484375
],
[
0.5009765625,
0.27783203125,
0.8935546875,
0.27783203125,
0.8935546875,
0.29541015625,
0.5009765625,
0.29541015625
],
[
0.07843017578125,
0.285400390625,
0.468017578125,
0.285400390625,
0.468017578125,
0.298583984375,
0.07843017578125,
0.298583984375
],
[
0.5,
0.292236328125,
0.8935546875,
0.294189453125,
0.8935546875,
0.3125,
0.5,
0.31005859375
],
[
0.07843017578125,
0.2998046875,
0.468017578125,
0.2998046875,
0.468017578125,
0.313232421875,
0.07843017578125,
0.313232421875
],
[
0.0767822265625,
0.309326171875,
0.4697265625,
0.31005859375,
0.4697265625,
0.328857421875,
0.0767822265625,
0.32763671875
],
[
0.50341796875,
0.31005859375,
0.89208984375,
0.31005859375,
0.89208984375,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.08087158203125,
0.322509765625,
0.468017578125,
0.322509765625,
0.468017578125,
0.34033203125,
0.08087158203125,
0.34033203125
],
[
0.50244140625,
0.324462890625,
0.89208984375,
0.324462890625,
0.89208984375,
0.33837890625,
0.50244140625,
0.33837890625
],
[
0.07843017578125,
0.34033203125,
0.4697265625,
0.34033203125,
0.4697265625,
0.354736328125,
0.07843017578125,
0.354736328125
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.07843017578125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.07843017578125,
0.369384765625
],
[
0.50341796875,
0.354736328125,
0.57666015625,
0.354736328125,
0.57666015625,
0.3662109375,
0.50341796875,
0.3662109375
],
[
0.080078125,
0.371826171875,
0.468017578125,
0.371826171875,
0.468017578125,
0.385009765625,
0.080078125,
0.385009765625
],
[
0.080078125,
0.385009765625,
0.4697265625,
0.385009765625,
0.4697265625,
0.3984375,
0.080078125,
0.3984375
],
[
0.50244140625,
0.380615234375,
0.8037109375,
0.380615234375,
0.8037109375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.080078125,
0.399658203125,
0.438720703125,
0.399658203125,
0.438720703125,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.51708984375,
0.406005859375,
0.89208984375,
0.406005859375,
0.89208984375,
0.419189453125,
0.51708984375,
0.419189453125
],
[
0.07843017578125,
0.419189453125,
0.39208984375,
0.419189453125,
0.39208984375,
0.4375,
0.07843017578125,
0.4375
],
[
0.50244140625,
0.419189453125,
0.8935546875,
0.42041015625,
0.8935546875,
0.43505859375,
0.50244140625,
0.434326171875
],
[
0.50244140625,
0.431884765625,
0.8935546875,
0.431884765625,
0.8935546875,
0.449462890625,
0.50244140625,
0.449462890625
],
[
0.096435546875,
0.4443359375,
0.4697265625,
0.4443359375,
0.4697265625,
0.458251953125,
0.096435546875,
0.458251953125
],
[
0.50341796875,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.4609375,
0.50341796875,
0.4609375
],
[
0.080078125,
0.458984375,
0.4697265625,
0.458984375,
0.4697265625,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.5009765625,
0.462158203125,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.5009765625,
0.4765625
],
[
0.07843017578125,
0.4716796875,
0.4697265625,
0.472900390625,
0.4697265625,
0.4873046875,
0.07843017578125,
0.486083984375
],
[
0.50244140625,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.50244140625,
0.4912109375
],
[
0.07843017578125,
0.486083984375,
0.4697265625,
0.4873046875,
0.4697265625,
0.501953125,
0.07843017578125,
0.50048828125
],
[
0.5009765625,
0.492431640625,
0.89208984375,
0.4912109375,
0.89208984375,
0.505859375,
0.5009765625,
0.5068359375
],
[
0.080078125,
0.501953125,
0.4697265625,
0.501953125,
0.4697265625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.5166015625,
0.4697265625,
0.5166015625,
0.4697265625,
0.52978515625,
0.080078125,
0.52978515625
],
[
0.50244140625,
0.521484375,
0.8935546875,
0.521484375,
0.8935546875,
0.53515625,
0.50244140625,
0.53515625
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50341796875,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.080078125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.564453125,
0.86669921875,
0.564453125,
0.86669921875,
0.57763671875,
0.50244140625,
0.57763671875
],
[
0.080078125,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.5869140625,
0.080078125,
0.5869140625
],
[
0.51708984375,
0.5791015625,
0.89208984375,
0.580078125,
0.89208984375,
0.5947265625,
0.51708984375,
0.59326171875
],
[
0.080078125,
0.58837890625,
0.468017578125,
0.58837890625,
0.468017578125,
0.6015625,
0.080078125,
0.6015625
],
[
0.5009765625,
0.59326171875,
0.8896484375,
0.59326171875,
0.8896484375,
0.60693359375,
0.5009765625,
0.60693359375
],
[
0.080078125,
0.6015625,
0.468017578125,
0.6015625,
0.468017578125,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.07843017578125,
0.6162109375,
0.404296875,
0.6171875,
0.404296875,
0.630859375,
0.07843017578125,
0.6298828125
],
[
0.5009765625,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.5009765625,
0.63623046875
],
[
0.08087158203125,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.50244140625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.080078125,
0.6484375,
0.468017578125,
0.6484375,
0.468017578125,
0.662109375,
0.080078125,
0.662109375
],
[
0.5009765625,
0.65087890625,
0.89208984375,
0.6513671875,
0.89208984375,
0.66552734375,
0.5009765625,
0.6640625
],
[
0.080078125,
0.66357421875,
0.4697265625,
0.66357421875,
0.4697265625,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.50244140625,
0.66650390625,
0.89208984375,
0.66650390625,
0.89208984375,
0.6796875,
0.50244140625,
0.6796875
],
[
0.080078125,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.69140625,
0.080078125,
0.69140625
],
[
0.50244140625,
0.68115234375,
0.89208984375,
0.68115234375,
0.89208984375,
0.6943359375,
0.50244140625,
0.6943359375
],
[
0.080078125,
0.6923828125,
0.4697265625,
0.6923828125,
0.4697265625,
0.70556640625,
0.080078125,
0.70556640625
],
[
0.50341796875,
0.69580078125,
0.89208984375,
0.69580078125,
0.89208984375,
0.708984375,
0.50341796875,
0.708984375
],
[
0.16259765625,
0.7119140625,
0.38232421875,
0.7138671875,
0.38232421875,
0.73193359375,
0.16259765625,
0.72998046875
],
[
0.50244140625,
0.71044921875,
0.89208984375,
0.71044921875,
0.89208984375,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.4453125,
0.71630859375,
0.4697265625,
0.71630859375,
0.4697265625,
0.72998046875,
0.4453125,
0.72998046875
],
[
0.50244140625,
0.7236328125,
0.89208984375,
0.7236328125,
0.89208984375,
0.7373046875,
0.50244140625,
0.7373046875
],
[
0.408447265625,
0.74072265625,
0.468017578125,
0.74072265625,
0.468017578125,
0.751953125,
0.408447265625,
0.751953125
],
[
0.5,
0.736328125,
0.8935546875,
0.734375,
0.8935546875,
0.751953125,
0.5,
0.75390625
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.08087158203125,
0.7412109375,
0.414306640625,
0.7412109375,
0.414306640625,
0.751953125,
0.08087158203125,
0.751953125
],
[
0.07843017578125,
0.75244140625,
0.471435546875,
0.75390625,
0.471435546875,
0.76806640625,
0.07843017578125,
0.767578125
],
[
0.080078125,
0.767578125,
0.25244140625,
0.767578125,
0.25244140625,
0.78076171875,
0.080078125,
0.78076171875
],
[
0.50244140625,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.78076171875,
0.50244140625,
0.78076171875
],
[
0.08087158203125,
0.7861328125,
0.468017578125,
0.7861328125,
0.468017578125,
0.7998046875,
0.08087158203125,
0.7998046875
],
[
0.62841796875,
0.79541015625,
0.765625,
0.79541015625,
0.765625,
0.81298828125,
0.62841796875,
0.81298828125
],
[
0.869140625,
0.79736328125,
0.8935546875,
0.79736328125,
0.8935546875,
0.81103515625,
0.869140625,
0.81103515625
],
[
0.07843017578125,
0.8125,
0.4697265625,
0.8125,
0.4697265625,
0.830078125,
0.07843017578125,
0.830078125
],
[
0.080078125,
0.80029296875,
0.4697265625,
0.80029296875,
0.4697265625,
0.814453125,
0.080078125,
0.814453125
],
[
0.080078125,
0.830078125,
0.471435546875,
0.830078125,
0.471435546875,
0.84326171875,
0.080078125,
0.84326171875
],
[
0.50244140625,
0.82763671875,
0.89208984375,
0.8291015625,
0.89208984375,
0.84326171875,
0.50244140625,
0.84228515625
],
[
0.08087158203125,
0.8447265625,
0.431396484375,
0.8447265625,
0.431396484375,
0.85791015625,
0.08087158203125,
0.85791015625
],
[
0.50244140625,
0.84326171875,
0.890625,
0.8447265625,
0.890625,
0.85791015625,
0.50244140625,
0.85693359375
],
[
0.50341796875,
0.85791015625,
0.89208984375,
0.85791015625,
0.89208984375,
0.8720703125,
0.50341796875,
0.8720703125
],
[
0.50341796875,
0.87255859375,
0.89208984375,
0.87255859375,
0.89208984375,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.16015625,
0.8798828125,
0.386474609375,
0.8798828125,
0.386474609375,
0.89794921875,
0.16015625,
0.89794921875
],
[
0.44677734375,
0.88330078125,
0.4697265625,
0.88330078125,
0.4697265625,
0.89599609375,
0.44677734375,
0.89599609375
],
[
0.5,
0.88330078125,
0.8935546875,
0.8837890625,
0.8935546875,
0.9013671875,
0.5,
0.90087890625
]
] |
[
"Table 1: Notational summary",
"y=(j,m)ey=(J,M)cR(21×3)x(224×224)",
"R",
"dense hand poseestimator (fDHPE=fPro)。fHME)",
"",
"Proj",
"DHPE",
"projectionoperator (fProj=[fReg,fRend)",
"",
"3Dskeletonregressor",
"eRen",
"(a)",
"(b)",
"(c)",
"(d)",
"Figure 3: A 2D evidence estimation example. (a) input",
"renderer",
"By construction, the projection operator respects the",
"underlying camera (via fRen) and hand shape (via fReg)",
"image x, (b) ground-truth 2D segmentation mask m of x.",
"geometry and it is held fixed throughout the entire training",
"(c) 2D skeletal position heat map of the finger tip of middle",
"process. This facilitates the training of fHME indirectly via",
"finger overlaid on x, and (d) masked image xOm.",
"training fDHPE. However, even under this setting, the prob-",
"lemstill remains challengingasestimatinga3Dmeshgiven",
"The output of fJ2D is a matrix of size 21x32x32 encod-",
"an RGB image is a seriously ill-posed problem. Adopting",
"ing 21 heat maps (per joint) of size 32x32. j2DHear is the",
"recent human body pose estimation approaches [35, 57], we",
"further stratify learning of fDHPE :X-y by decomposing",
"ground-truth heat map. Given the estimated heat maps, 2D",
"fHME : X - V into a 2D evidence estimator fE2D : X Z",
"skeleton joints j2p are extracted by finding the maximum for",
"and a 3D mesh estimator fE3D : 2 → V. Our 2D evidence",
"each joint. Figure 3(b) shows an example of 2D evidence",
"zE Z consists of a 42-dimensional 2D skeletal joint position",
"estimation.",
"vector j2D (21 positions × 2; as in [4, 14]) and a 2,048-",
"dimensional 2D feature vector F(x) (Eq. 2). The remainder",
"3.2. 3D mesh estimator fE3D =(fJ3D, fRef)",
"of this section provides details of these two estimators.",
"Taking the 2D evidence z (2,048-dimensional feature vec-",
"3.1. 2D evidence estimator fE2D = (F, fJ2D)",
"tor F(x) and 21 ×3-dimensional 2D skeleton j2p) as input,",
"the 3D mesh estimation network fJ3D constructs parameters",
"Silhouettes (or foreground masks) and 2D skeletons",
"of adeformablehand mesh model and a cameramodel",
"have been widely used as the mid-level cues for estimating",
"We use the MANO model representing a hand mesh based",
"3D body meshes [17, 35, 57].However, for hands,",
"on 45-dimensional pose parameters p and 10-dimensional",
"accurately estimating foreground masks from RGB images",
"shape parameters s [41]. The original MANO framework",
"is challenging due to cluttered backgrounds [2], 68]",
"uses only 6-dimensional (principal component analysis)",
"We observed that naively applying the state-of-the-art",
"PCA subspace of p for computational efficiency. However,",
"foreground segmentation algorithms (e.g. [68]) often misses",
"weempiricallyobservedthattocoveravarietyofhandposes.",
"fine details, especially along the narrow finger regions (see",
"all 45-dimensional features are required: We use the linear",
"Fig. 4 for examples) and this can significantly degrade the",
"blend skinning formulation as in the SMPL model [24].",
"performance of the subsequent mesh estimation step. We",
"Given the MANO parameters p and s, a mesh v is",
"bypass this challenge by learning instead, a 2D foreground",
"synthesized by conditioning them on the hypothesized",
"feature extractor F: F(x) encapsulates the textural and",
"camera c, which comprises of the 3D rotation (in quaternion",
"shape information of theforeground regions in x.",
"space) c. e R4, scale cs E R, and translation ct e IR3: An",
"Our 2D feature extractor F is trained to focus on the",
"initial mesh is constructed by combining MANO's PCA",
"foreground regions by minimizing the deviation between the",
"basis vectors using the shapeparameter s,rotating the",
"features extracted from the entire image x and the foreground",
"bones according to the pose parameter p, and deforming the",
"region xom extracted via the ground-truth mask m (see",
"resulting surface via linear blend skinning. The final mesh v",
"Fig. 3d): The training loss (per data point) for F is given as",
"is then obtained by globally rotating, scaling, and translating",
"LFear(F)=I|F(x)-F(xm)2,",
"the initial mesh according to cg, Cs, and ct, respectively.",
"(2)",
"The entire mesh generation process is differentiable. Under",
"Fem",
"this model, our 3D mesh estimation network fJ3D is",
"implemented as a singlefully-connected layer of 2153×63",
"where denoteselement-wisemultiplication.",
"ploys the ResNet-50 [13] architecture whose output is a",
"2,048-dimensional vector.",
"weights and it estimates a 63-dimensional mesh parameter:",
"Estimating 2D skeletal joints from an RGB image has",
"h=[p,s,Cq,Cs,ct] T.",
"(4)",
"of-the-art 2D pose estimation network fJ2D [60, 68] into our",
"been well studied. Similarly to [4, 14], we embed the state-",
"framework. The initial weights provided by the authors of",
"Iterative mesh refinement via back-projection. Adopting",
"[68] are refined based on the 2D joint estimation loss:",
"Kanazawa et al.'s approach [16], instead of estimating",
"h directly, we iteratively refine the initial mesh estimate",
"h(O) by recursively performing regression on the parameter",
"LJ2D(fJ2D)=I fJ2D(x)-j2DHearlI",
"(3)",
"offset △h: At iteration t, fJ3D takes z and the current mesh"
] |
{
"arxiv_id": "1904.04196",
"language": "en",
"page_idx": 3
}
|
||
1904.04196-en-4
|
1904.04196-en-4.jpg
|
estimate h$(t)$, and generates a new offset $\Delta\mathbf{h}(t)\colon$
(5)
$$\mathbf h(t+1)=\mathbf h(t)+\Delta\mathbf h(t).$$
At $t=0,\mathbf{h}(0)$ as an input to $f^{J3D}$, is constructed as a vector of zeros except for the entries corresponding to the pose parameter p which is set as the mean pose of the MANO model. We fix the number of iterations at 3 as more iterations did not show any noticeable improvements in preliminary experiments.
$\operatorname{In}\left[16\right]$,each offset prediction $\Delta\mathbf{h}(t)$ is built based only on 2D features $F(\mathbf{x}).$ Inspired by the success of [4,14],we additionally use 2D skeletal joint $\mathbf{j}_{2D}^{\prime}$ as input. However, as an estimate, $\mathbf{j}_{2D}^{\prime}(0)$ (obtained at iteration $\bar{0})$ is inaccurate, causing errors in the resulting $\Delta\mathbf{h}(t)$ estimate. Then, these errors accumulate over iterations (Eq.5) lessening the benefit of the entire iterative estimation process. We address this by an additional 2D pose refiner $f^{\hat{R}ef}$ which iteratively refines the 2D joint estimation $\mathbf{j}_{2D}^{\prime}(t)$ via $back-projecting$the estimated mesh h'$(t):$ At $t,f^{Ref}$ receives the estimated $\operatorname*{mesh}_{a\boldsymbol{P}a\boldsymbol{\alpha}(t)}$parameter$\textbf{h}^\prime(t),2$D feature$F(\mathbf{x}),3$D skeleton $f^{Reg}(\mathbf{v}^{\prime})$,and$\mathbf{j}_2D^{\prime}(t)$,and$\mathbf{generates}\mathbf{a}$refined$\mathbf{estimate}$ $\begin{array}{l}{\mathbf{j}_{2D}^{\prime}(t+1).\mathrm{~The~2D~pose~refiner~}f^{Ref}\text{ is implemented as}}\\{\mathrm{a~single~fully~connected~layer~of~size~2216\times42~and~it~is}}\\\end{array}$ trained based on the ground-truth 2D skeletons (j$_2DGT):$
$$L_{Ref}=\begin{Vmatrix}f^{Ref}\big([\mathbf{j}_{2D}^{\prime}(t),F(\mathbf{x}),\mathbf{h}^{\prime}(t),f^{Reg}(\mathbf{v}^{\prime})]\big)-\mathbf{j}_{2DGT}\end{Vmatrix}_{2}^{2}.$$
(6)
In the experiments, we demonstrate that 1) using both 2D features and 2D skeletons (Fig. 5(e): `Ours (w/o Test. ref. and Refiner $f^{Ref})’)$ improves performance over Kanazawa et al.'s 2D feature-based framework[16] (Fig. 5(e): “Ours (w/o Test. ref., $f^{Ref}$ and 2D losses $(L_Feat,L_{J2D})’);2)$ explicitly building the refiner $f^{Ref}$ under the auxiliary 2D joint supervision (Fig. 5(e): 'Ours (w/o Test. ref.)') further significantly improves performance. More importantly, the refiner takes a crucial role in our testing refinement step (Sec. 3.4) which improves a once predicted 3D mesh v' by enforcing its consistency with the corresponding 2D joint evidence $\mathbf{j}_{2D}^{\prime}(t)$ (as well as other intermediate results) throughout the iteration (Eq. 11).
# 3.3. Joint training
Training the 2D evidence estimator $f^{E2D}:\mathcal{X}\to\mathcal{Z}$ and 3D mesh estimator $f^{E3D}:\mathcal{Z}\to\mathcal{V}$ given fixed projection operator $f^{Proj}:\mathcal{V}\to\mathcal{Y}$ is performed based on a training set consisting of input images, and the corresponding 3D skeleton joints and 2D segmentation masks $\hat{D} = \{ ( \mathbf{x} _{i}, \mathbf{y} _{i}\} _{i= 1}^{l}\subset \dot{\mathcal{X} } \times \mathcal{Y} $, $\mathbf{y}_{i}=(\mathbf{j}_{i},\mathbf{m}_{i}).$ Since all component functions of $f^E2D,f^{E3D}$, and $f^{Proj}$ are differentiable with respect to the weights of $F$ and $f^E3D$, they can be optimized based on on standard gradient descent-type algorithm: Our overall loss (per training instance $\mathbf{x}_i,\mathbf{y}_i)$ is given as (see Eqs. 2 and 6):
$$\begin{aligned}L(f^{E3D},F)&=L_{Art}(f^{E3D},F)+L_{Lap}(f^{E3D},F)+L_{Feat}(F)\\&+\lambda L_{Sh}(f^{J3D},F)+L_{Ref}(f^{J3D},F).\end{aligned}$$
The articulation loss $L_{Art}$ measures the deviation between
the skeleton estimated from x$_i$ and its ground-truth j$_i:$
$$L_{Art}=\|[f^{DHPE}(\mathbf{x}_{i})]_{\mathcal{J}}-\widehat{\mathbf{j}_{i}}\|_{2}^{2},$$
(8)
where [y] $_\mathcal{J}$ extracts the j-component of y$=(\mathbf{j},\mathbf{m})$ and $\widehat{\mathbf{j}}$ spatially normalizes j similarly to [14,68]: First, the center of each skeleton is moved to the corresponding middle finger's MCP position. Then each axis is normalized to a unit interval $[0,1]\colon$The $x,y-$coordinate values are divided by $g(=1.5$ times the maximum of height and width of the tight 2D hand bounding box). The $z-$axis value is divided middle finger's MCP joint and c$_f$ is the focal length of the camera. At testing, once normalized skeletons are estimated, they are inversely normalized to the original scale. The accompanying supplemental provides details of this inverse $\hat{\text{normalization step.}}$
Our shape loss $L_{Sh}$ facilitates the recovery of hand shapes as observed indirectly via projected 2D segmentation masks:
$$L_{Sh}=\begin{Vmatrix}[f^{DHPE}(\mathbf{x}_i)]_{\mathcal{M}}-\mathbf{m}_i\end{Vmatrix}_2^2,$$
(9)
where $[\mathbf{y}]_{\mathcal{M}}$ extracts the m-component of y=(j,m).
The Laplacian regularizer $L_{Lap}$ enforces spatial smoothness in the mesh v. This helps avoid generating implausible hand meshes as suggested by Kanazawa et al.[17].
Hierarchical recovery of articulation and shapes. We observed that naïvely minimizing the overall loss $L$ with a constant shape loss weight $\lambda$ (Eq. 7) tends to impede convergence during training (Fig. $\hat{5} ($f)): Our algorithm simultaneously optimizes the mesh (v) and camera $(\mathbf{c}=\{\mathbf{c}_{q},\mathbf{c}_{s},\mathbf{c}_{t}\})$ parameters which over-parameterize the rendered 2D view, e.g. the effect of scaling v itself can be offset by inversely scaling $\mathbf{c}_s.$ This often hinders the alignment of v with the ground-truth 2D mask and thereby rendering the network inappropriately update the mesh parameters. Therefore, we let the shape loss $L_Sh$ take effect only when the articulation loss $L_{Art}$ becomes sufficiently small: $\lambda$ is set per data instance based on the $L_{Art}$-value:
$$\lambda=\left\{\begin{array}{ll}1&\text{if}\left\|\left[[f^{DHPE}(\mathbf{x}_i)]_{\mathbf{j}}\right]_{2D}-[\mathbf{j}_i]_{2D}\right\|_2^2<\tau\\0&\text{otherwise},\end{array}\right.$$
(10)
where $[\mathbf{j}]_{2D}$ projects 3D joint coordinates j to 2D view based on c (Eq. 4). The threshold $\tau$ is empirically set to 15 pixels. Initially, with zero $\lambda,L_{Art}$ dominates in $L$, which helps globally align the estimated meshes with 2D ground-truth evidence. As the training progresses, the role of $L_{Sh}$ becomes more important $(\lambda=1)$ contributing to recovering the detailed shapes.
As the generation of the 2D skeleton j$_{2D}$ from an estimated heat-map $f^{J2D}(\mathbf{x})$ is not differentiable (see Eq. 3), our 2D pose estimation network $f^{J2D}$ cannot be trained based on $L.$ Thus we train it in parallel using $L_J2D$ (Eq. 3). For both cases, we use the standard Adam optimizer with the learning rate $\gamma$ set at $10^-3.$
|
[
[
0.080078125,
0.0921630859375,
0.398681640625,
0.0921630859375,
0.398681640625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.09027099609375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.10479736328125
],
[
0.5009765625,
0.1060791015625,
0.85693359375,
0.106689453125,
0.85693359375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.193603515625,
0.11614990234375,
0.35302734375,
0.11614990234375,
0.35302734375,
0.1300048828125,
0.193603515625,
0.1300048828125
],
[
0.44677734375,
0.117431640625,
0.468017578125,
0.117431640625,
0.468017578125,
0.1300048828125,
0.44677734375,
0.1300048828125
],
[
0.583984375,
0.125,
0.775390625,
0.126953125,
0.775390625,
0.1483154296875,
0.583984375,
0.146484375
],
[
0.869140625,
0.1319580078125,
0.89208984375,
0.1319580078125,
0.89208984375,
0.1458740234375,
0.869140625,
0.1458740234375
],
[
0.07843017578125,
0.1395263671875,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1558837890625,
0.07843017578125,
0.1572265625
],
[
0.080078125,
0.1572265625,
0.468017578125,
0.1572265625,
0.468017578125,
0.17041015625,
0.080078125,
0.17041015625
],
[
0.50244140625,
0.1572265625,
0.890625,
0.1546630859375,
0.890625,
0.169189453125,
0.50244140625,
0.1717529296875
],
[
0.08251953125,
0.17236328125,
0.468017578125,
0.17236328125,
0.468017578125,
0.18310546875,
0.08251953125,
0.18310546875
],
[
0.50244140625,
0.1717529296875,
0.89208984375,
0.1717529296875,
0.89208984375,
0.1849365234375,
0.50244140625,
0.1849365234375
],
[
0.080078125,
0.1849365234375,
0.4697265625,
0.1849365234375,
0.4697265625,
0.1988525390625,
0.080078125,
0.1988525390625
],
[
0.50244140625,
0.186279296875,
0.89208984375,
0.186279296875,
0.89208984375,
0.199462890625,
0.50244140625,
0.199462890625
],
[
0.080078125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.21337890625,
0.080078125,
0.21337890625
],
[
0.50244140625,
0.2008056640625,
0.89208984375,
0.2008056640625,
0.89208984375,
0.2139892578125,
0.50244140625,
0.2139892578125
],
[
0.080078125,
0.2139892578125,
0.2442626953125,
0.2139892578125,
0.2442626953125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50341796875,
0.21533203125,
0.89208984375,
0.21533203125,
0.89208984375,
0.2291259765625,
0.50341796875,
0.2291259765625
],
[
0.09478759765625,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.09478759765625,
0.2412109375
],
[
0.5009765625,
0.2291259765625,
0.890625,
0.2279052734375,
0.890625,
0.242431640625,
0.5009765625,
0.24365234375
],
[
0.07843017578125,
0.2412109375,
0.4697265625,
0.242431640625,
0.4697265625,
0.2568359375,
0.07843017578125,
0.255615234375
],
[
0.50341796875,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.07843017578125,
0.255615234375,
0.4697265625,
0.2568359375,
0.4697265625,
0.271484375,
0.07843017578125,
0.270751953125
],
[
0.50341796875,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.272705078125,
0.50341796875,
0.272705078125
],
[
0.080078125,
0.271484375,
0.4697265625,
0.271484375,
0.4697265625,
0.285888671875,
0.080078125,
0.285888671875
],
[
0.50244140625,
0.272705078125,
0.890625,
0.272705078125,
0.890625,
0.285888671875,
0.50244140625,
0.285888671875
],
[
0.08087158203125,
0.287353515625,
0.46728515625,
0.287353515625,
0.46728515625,
0.29736328125,
0.08087158203125,
0.29736328125
],
[
0.50341796875,
0.28857421875,
0.890625,
0.28857421875,
0.890625,
0.298583984375,
0.50341796875,
0.298583984375
],
[
0.080078125,
0.300537109375,
0.468017578125,
0.300537109375,
0.468017578125,
0.314453125,
0.080078125,
0.314453125
],
[
0.50244140625,
0.300537109375,
0.89208984375,
0.300537109375,
0.89208984375,
0.314453125,
0.50244140625,
0.314453125
],
[
0.07843017578125,
0.3125,
0.4697265625,
0.314453125,
0.4697265625,
0.32763671875,
0.07843017578125,
0.32568359375
],
[
0.50244140625,
0.31640625,
0.89208984375,
0.31640625,
0.89208984375,
0.330078125,
0.50244140625,
0.330078125
],
[
0.080078125,
0.32763671875,
0.468017578125,
0.32763671875,
0.468017578125,
0.34228515625,
0.080078125,
0.34228515625
],
[
0.5009765625,
0.328857421875,
0.62841796875,
0.330078125,
0.62841796875,
0.3447265625,
0.5009765625,
0.343505859375
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.50244140625,
0.347900390625,
0.89208984375,
0.347900390625,
0.89208984375,
0.361083984375,
0.50244140625,
0.361083984375
],
[
0.080078125,
0.356689453125,
0.31201171875,
0.35595703125,
0.31201171875,
0.369384765625,
0.080078125,
0.37060546875
],
[
0.303955078125,
0.356689453125,
0.46728515625,
0.35791015625,
0.46728515625,
0.369384765625,
0.303955078125,
0.36865234375
],
[
0.50244140625,
0.3623046875,
0.89453125,
0.3623046875,
0.89453125,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.08087158203125,
0.371826171875,
0.468017578125,
0.371826171875,
0.468017578125,
0.385009765625,
0.08087158203125,
0.385009765625
],
[
0.08251953125,
0.3837890625,
0.4697265625,
0.385009765625,
0.4697265625,
0.399658203125,
0.08251953125,
0.3984375
],
[
0.57666015625,
0.380615234375,
0.78125,
0.38330078125,
0.78125,
0.40478515625,
0.57666015625,
0.40283203125
],
[
0.87109375,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.401611328125,
0.87109375,
0.401611328125
],
[
0.0767822265625,
0.3984375,
0.4697265625,
0.397705078125,
0.4697265625,
0.415283203125,
0.0767822265625,
0.416748046875
],
[
0.07843017578125,
0.4140625,
0.468017578125,
0.41357421875,
0.468017578125,
0.427978515625,
0.07843017578125,
0.4287109375
],
[
0.50341796875,
0.412353515625,
0.8505859375,
0.412353515625,
0.8505859375,
0.4267578125,
0.50341796875,
0.4267578125
],
[
0.080078125,
0.4287109375,
0.443603515625,
0.4287109375,
0.443603515625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.429931640625,
0.890625,
0.429931640625,
0.890625,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.50244140625,
0.443115234375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458984375,
0.50244140625,
0.458251953125
],
[
0.08087158203125,
0.46533203125,
0.4697265625,
0.46533203125,
0.4697265625,
0.48291015625,
0.08087158203125,
0.48291015625
],
[
0.5009765625,
0.460205078125,
0.833984375,
0.460205078125,
0.833984375,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.50341796875,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.09478759765625,
0.490478515625,
0.46728515625,
0.4892578125,
0.46728515625,
0.50390625,
0.09478759765625,
0.5048828125
],
[
0.5009765625,
0.4912109375,
0.8896484375,
0.4912109375,
0.8896484375,
0.5048828125,
0.5009765625,
0.5048828125
],
[
0.080078125,
0.5048828125,
0.4697265625,
0.5048828125,
0.4697265625,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.50244140625,
0.51953125
],
[
0.07843017578125,
0.51806640625,
0.4697265625,
0.52099609375,
0.4697265625,
0.53515625,
0.07843017578125,
0.53271484375
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.08087158203125,
0.5341796875,
0.4697265625,
0.5341796875,
0.4697265625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.5009765625,
0.5341796875,
0.89208984375,
0.53271484375,
0.89208984375,
0.54736328125,
0.5009765625,
0.548828125
],
[
0.0767822265625,
0.54541015625,
0.471435546875,
0.546875,
0.471435546875,
0.564453125,
0.0767822265625,
0.56298828125
],
[
0.50341796875,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50341796875,
0.56298828125
],
[
0.08087158203125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.50341796875,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50341796875,
0.5791015625,
0.890625,
0.5791015625,
0.890625,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.080078125,
0.59228515625,
0.468017578125,
0.59228515625,
0.468017578125,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.50341796875,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.60498046875,
0.50341796875,
0.60498046875
],
[
0.07843017578125,
0.60595703125,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.07843017578125,
0.619140625
],
[
0.50244140625,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.50244140625,
0.62060546875
],
[
0.08087158203125,
0.62255859375,
0.464111328125,
0.62255859375,
0.464111328125,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.50244140625,
0.62060546875,
0.89208984375,
0.62060546875,
0.89208984375,
0.6337890625,
0.50244140625,
0.6337890625
],
[
0.08087158203125,
0.6337890625,
0.46728515625,
0.6337890625,
0.46728515625,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.50244140625,
0.63525390625,
0.890625,
0.63525390625,
0.890625,
0.6484375,
0.50244140625,
0.6484375
],
[
0.080078125,
0.6494140625,
0.468017578125,
0.6494140625,
0.468017578125,
0.66357421875,
0.080078125,
0.66357421875
],
[
0.50244140625,
0.6494140625,
0.87158203125,
0.6494140625,
0.87158203125,
0.66357421875,
0.50244140625,
0.66357421875
],
[
0.080078125,
0.6640625,
0.29736328125,
0.6640625,
0.29736328125,
0.67822265625,
0.080078125,
0.67822265625
],
[
0.56640625,
0.66845703125,
0.8037109375,
0.67236328125,
0.8037109375,
0.6943359375,
0.56640625,
0.68994140625
],
[
0.07843017578125,
0.68505859375,
0.2181396484375,
0.6875,
0.2181396484375,
0.7021484375,
0.07843017578125,
0.69970703125
],
[
0.861328125,
0.6796875,
0.89208984375,
0.6796875,
0.89208984375,
0.69580078125,
0.861328125,
0.69580078125
],
[
0.564453125,
0.68994140625,
0.6650390625,
0.68994140625,
0.6650390625,
0.705078125,
0.564453125,
0.705078125
],
[
0.096435546875,
0.70703125,
0.468017578125,
0.70703125,
0.468017578125,
0.7216796875,
0.096435546875,
0.7216796875
],
[
0.50341796875,
0.71337890625,
0.89208984375,
0.71337890625,
0.89208984375,
0.72802734375,
0.50341796875,
0.72802734375
],
[
0.0767822265625,
0.7197265625,
0.471435546875,
0.72021484375,
0.471435546875,
0.73779296875,
0.0767822265625,
0.7373046875
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.7373046875,
0.468017578125,
0.751953125,
0.08087158203125,
0.75
],
[
0.50244140625,
0.72900390625,
0.89208984375,
0.72900390625,
0.89208984375,
0.7421875,
0.50244140625,
0.7421875
],
[
0.50244140625,
0.74365234375,
0.89208984375,
0.74365234375,
0.89208984375,
0.7568359375,
0.50244140625,
0.7568359375
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50341796875,
0.75830078125,
0.89208984375,
0.75830078125,
0.89208984375,
0.771484375,
0.50341796875,
0.771484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.77294921875,
0.89208984375,
0.77294921875,
0.89208984375,
0.7861328125,
0.50341796875,
0.7861328125
],
[
0.07843017578125,
0.78076171875,
0.468017578125,
0.77783203125,
0.468017578125,
0.79248046875,
0.07843017578125,
0.79541015625
],
[
0.1298828125,
0.79345703125,
0.468017578125,
0.79345703125,
0.468017578125,
0.806640625,
0.1298828125,
0.806640625
],
[
0.50341796875,
0.78857421875,
0.89208984375,
0.78857421875,
0.89208984375,
0.798828125,
0.50341796875,
0.798828125
],
[
0.080078125,
0.796875,
0.1298828125,
0.79443359375,
0.1298828125,
0.80615234375,
0.080078125,
0.80810546875
],
[
0.07843017578125,
0.8046875,
0.4697265625,
0.806640625,
0.4697265625,
0.82470703125,
0.07843017578125,
0.82275390625
],
[
0.50244140625,
0.80029296875,
0.705078125,
0.80029296875,
0.705078125,
0.814453125,
0.50244140625,
0.814453125
],
[
0.51708984375,
0.814453125,
0.89208984375,
0.814453125,
0.89208984375,
0.8291015625,
0.51708984375,
0.8291015625
],
[
0.08087158203125,
0.82373046875,
0.468017578125,
0.82373046875,
0.468017578125,
0.8369140625,
0.08087158203125,
0.8369140625
],
[
0.50341796875,
0.8291015625,
0.89208984375,
0.8291015625,
0.89208984375,
0.84326171875,
0.50341796875,
0.84326171875
],
[
0.08087158203125,
0.83837890625,
0.43310546875,
0.83837890625,
0.43310546875,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.50341796875,
0.84228515625,
0.89208984375,
0.84228515625,
0.89208984375,
0.85693359375,
0.50341796875,
0.85693359375
],
[
0.096435546875,
0.85693359375,
0.450927734375,
0.859375,
0.450927734375,
0.87890625,
0.096435546875,
0.876953125
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85791015625,
0.89208984375,
0.87255859375,
0.50244140625,
0.8720703125
],
[
0.50341796875,
0.87255859375,
0.89208984375,
0.87255859375,
0.89208984375,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.1690673828125,
0.8798828125,
0.384033203125,
0.8798828125,
0.384033203125,
0.89794921875,
0.1690673828125,
0.89794921875
],
[
0.44677734375,
0.88330078125,
0.468017578125,
0.88330078125,
0.468017578125,
0.89599609375,
0.44677734375,
0.89599609375
],
[
0.5009765625,
0.88623046875,
0.705078125,
0.88525390625,
0.705078125,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"estimate h(t), and generates a new offset h(t):",
"The articulation loss LAr measures the deviation between",
"the skeleton estimated from x, and its ground-truth j::",
"h(t+1)=h(t)+h(t).",
"(5)",
"LAn= I[FDHPE(x;)]-j2,",
"(8)",
"At t = O, h(O) as an input to fJ3D, is constructed as a",
"vector of zeros except for the entries corresponding to the",
"where [yl extracts the j-component of y = (j.m) and i",
"pose parameter p which is set as the mean pose of the",
"spatially normalizes j similarly to [14, 68]: First, the center",
"MANO model. We fix the number of iterations at 3 as more",
"of each skeleton is moved to the corresponding middle",
"iterations did not show any noticeable improvements in",
"finger's MCP position. Then each axis is normalized to a",
"preliminary experiments.",
"unit interval [0,1]: The r,y-coordinate values are divided",
"In [16], each offset prediction h(t) is built based only",
"by g (=1.5 times the maximum of height and width of the",
"on 2D features F(x). Inspired by the success of [4, 14], we",
"tight 2D hand bounding box). The z-axis value is divided",
"additionally use 2D skeletal joint j2D as input. However,",
"by (zRoot X g)/cf where zRoot is the depth value of the",
"as an estimate, j2(O) (obtained at iteration O) is inaccurate,",
"middle finger's MCP joint and cf is the focal length of the",
"causing errorsin theresulting h(t)estimate.Then,these",
"camera.Attesting,oncenormalized skeletons areestimated",
"errors accumulate over iterations (Eq. 5) lessening the",
"they are inversely normalized to the original scale. The",
"benefit of the entire iterative estimation process.We address",
"accompanying supplemental provides details of this inverse",
"this by an additional 2D pose refiner fRef which iteratively",
"normalization step.",
"refines the 2D joint estimation j2p(t) via back-projecting",
"Our shape loss Lsh facilitates the recovery of hand shapes",
"the estimated mesh h'(t): At t, fRe",
"Refreceivestheestimated",
"as observed indirectly via projected 2D segmentation masks:",
"mesh parameter h'(t), 2D feature F(x), 3D skeleton",
"fReg(v'), and jp(t), and generates a refined estimate",
"Lsh=[FDHPE(x;)]M-m;l/2,",
"(9)",
"j2D(t + 1). The 2D pose refiner fRef is implemented as",
"a single fully connected layer of size 2216x42 and it is",
"where [y]m extracts the m-component of y = (j,m).",
"trained based on the ground-truth 2D skeletons (j2DGr):",
"The Laplacian regularizer LLap enforces spatial smooth",
"ness in the mesh v. This helps avoid generating implausible",
"LRef=IRef ([i2D(t),F(x),h'(t), Reg(v))-j2DGT2. (6)",
"hand meshes as suggested by Kanazawa et al. [17]",
"Hierarchical recovery of articulation and shapes. We",
"In the experiments, we demonstrate that 1) using both 2D",
"observed that naively minimizing the overall loss L with",
"features and 2D skeletons (Fig. 5(e): 'Ours (w/o Test. ref",
"a constant shape loss weight ^ (Eq. 7) tends to impede",
"and Refiner fRef)') improves performance over Kanazawa et",
"convergence during training (Fig. 5(f)): Our algorithm",
"al.'s 2D feature-based framework [16] (Fig. 5(e): 'Ours",
"simultaneously optimizes the mesh (v) and camera",
"(w/o Test. ref., fRef and 2D losses (LFeat, LJ2D)); 2)",
"(c = {cg,Cs,Ct]) parameters which over-parameterize the",
"explicitly building the refiner fRef under the auxiliary 2D",
"rendered 2D view, e.g. the effect of scaling v itself can",
"joint supervision (Fig.5(e): ‘Ours (w/o Test. ref.)') further",
"be offsetby inversely scaling cs:This often hinders the",
"significantly improves performance. More importantly, the",
"alignment of v with the ground-truth 2D mask and thereby",
"refiner takes a crucial role in our testing refinement step",
"rendering the network inappropriately update the mesh",
"(Sec.3.4)whichimprovesaoncepredicted3Dmeshv",
"parameters. Therefore, we let the shape loss Lsh take effect",
"by enforcing its consistency with the corresponding 2D",
"only when the articulation loss LArt becomes sufficiently",
"joint evidence j2p(t) (as well as other intermediate results)",
"small: X is set per data instance based on the LAr-value:",
"throughout the iteration (Eq. 11).",
"1 if|[FDHPE(x;)]2D-[i]2Dll2",
"3.3. Joint training",
"(10)",
"0 otherwise,",
"Training the 2D evidence estimator fE2D: X→ Z and 3D",
"where [i]2D projects 3D joint coordinates j to 2D view",
"mesh estimator fE3D : 2→ V given fixed projection operator",
"fProj:v-yis performed based on a training set consisting",
"based on c (Eq. 4). The threshold T is empirically set",
"to 15 pixels. Initially, with zero , LArt dominates in L,",
"of input images, and the corresponding 3D skeleton joints",
"which helps globally align the estimated meshes with 2D",
"and 2D segmentation masks D = {(xi,yi,-1 C X × y)",
"ground-truth evidence. As the training progresses, the role",
"yi=(ji,m;).Since all component functions of fE2D,fE3D",
"roj are differentiable with respect to the weights of",
"of Lshbecomes moreimportant(=1)contributingto",
"andfp",
"F and fE3D, they can be optimized based on on standard",
"recovering the detailed shapes.",
"As the generation of the 2D skeleton j2D from an",
"gradient descent-type algorithm: Our overall loss (per",
"estimated heat-map fJ2D(x) is not differentiable (see Eq. 3),",
"training instance Xi,yi) is given as (see Eqs. 2 and 6):",
"our 2D pose estimation network fJ2D cannot be trained",
"L(fE3D, F)=LAr(fE3D,F)+Luap(fE3D,F)+LFea(F)",
"based on L. Thus we train it in parallel using Lj2 (Eq. 3)",
"For both cases, we use the standard Adam optimizer with",
"+ALsh(fJ3D,F)+ LRer(fJ3D,F).",
"(7)",
"the learning rate set at 10-3."
] |
{
"arxiv_id": "1904.04196",
"language": "en",
"page_idx": 4
}
|
||
2309.06102-en-0
|
2309.06102-en-0.jpg
|
Can we predict the Most Replayed data of video streaming platforms?
Alessandro Duico$^1,\text{Ombretta Strafforello}^{1,2}$,Jan van Gemert$^1$
$^{1}$Delft University of Technology, $^2$TNO
$\{ a. duico$, $o. strafforello$, $j. c. vangemert\} @ tudelft. nl$
### $\mathbf{Abstract}$
Predicting which specific parts of a video users will replay is important for several applications, including targeted advertisement placement on video platforms and assisting video creators. In this work, we explore whether it is possible to predict the Most Replayed $( MR) \textit{data from YouTube}$
nuto pratutin mos. terpraty antivan tantivantiontioc mulus wentis wentigate whentis possssining ODLjes $videos.To.this.end,we~curate~a~large~video~benchmark,~he~dict~the~the~Most~Replayed~data~using~Deep~Learning~(DL).$ YTMRSOO dataset, which co
$\begin{array}{ccc}&with\:MR\:data\:annotations.\:We\:evaluate\:Deep\:Learning\:(DL)\\\dot{O}&models\:of\:varying\:complexity\:on\:our\:dataset\:and\:perform\:an\\&ertensive\:ablation\:studv\:In\:addition\:we\:conduct\:a\:user\:studv\end{array}$
$extensiveablationstudy.Inaddition,weconductauserstudy$ to estimate the human performance on MR data prediction. Our results show that, although by a narrow margin, all the evaluated DL models outperform random predictions. Ad- $ditionally,they$ exceed human-level accuracy. This suggests that predicting the MR data is a difficult task that can be enhanced through the assistance of DL. Finally, we believe that DL performance on MR data prediction can be further $improved,for$ example, by using multi-modal learning. We encourage the research community to use our benchmark $dataset\textit{ to further investigate automatic MR data prediction. }$
## $1. \textbf{Introduction}$
Video streaming has emerged as a dominant mode of online communication, representing 73% of all internet traffic in 2017 (8), with YouTube leading the way as the most popular platform. Video streaming platforms accumulate, in addition to video data, a substantial amount of metadata, pertaining to users' watching habits and interests. Notably, in May 2022 YouTube released a new feature that shows a line chart of the most frequently replayed moments of each video, the Most Replayed data. In addition to aiding YouTube users during video playback, this data can serve various other potential applications, such as optimizing advertisement placement and giving feedback to content creatorsfor instance, suggesting uninteresting scenes to remove. In both cases, it is desirable to predict the Most Replayed data
before publishing a video. For advertisers, it enables placing advertisements optimally from the very fırst views, thereby maximizing profits. For content creators, it allows cutting the video appropriately before it reaches the audience, preventing the reputational damage caused by re-uploading a video after collecting the data.
In this work, we investigate whether it is possible to predata and pre-extracted video features. For a comprehensive description of our dataset collection process, readers are referred to the Supplementary Material (Section 1). We evaluate two DL models on YTMR500, consisting in a fullyconnected network and an attention-based architecture, inspired by the PGL-SUM model that Apostolidis et al. (3) proposed for video summarization. We compare the results of DL against human performance, which we estimate through a crowdsourced user study.
We make the following contributions: (1) We introduce $YTMR500$, a dataset of 500 videos and the corresponding Most Replayed data, that can foster research on most replayed data predictions in videos; (2) We design a variant of the PGL-SUM (3) architecture, adapted to predict the Most Replayed data for unseen videos; (3) We perform a user study to evaluate human performance on MR data prediction. Our results show that predicting the Most Replayed data is challenging for human annotators and that our model surpasses human performance. The YTMR500 dataset and $\hat{\text{our code are publicly available}}^{1}.$
## $2. \textbf{Predicting the Most Replayed data in a video}$ 2.1. Problem statement
Given a sequence of segments $V$ in a video, we design a model to learn a function $F$ that maps $V$ into the Most Replayed (MR) data $Y.$ Specifically, the input is a sequence
$^{\mathrm{l}}$https://github.com/ombretta/most-replayed-data
|
[
[
0.1380615234375,
0.1331787109375,
0.83642578125,
0.1331787109375,
0.83642578125,
0.15087890625,
0.1380615234375,
0.15087890625
],
[
0.2459716796875,
0.1956787109375,
0.7236328125,
0.1956787109375,
0.7236328125,
0.210205078125,
0.2459716796875,
0.210205078125
],
[
0.330810546875,
0.2235107421875,
0.6396484375,
0.2235107421875,
0.6396484375,
0.2412109375,
0.330810546875,
0.2412109375
],
[
0.291748046875,
0.242431640625,
0.6845703125,
0.2412109375,
0.6845703125,
0.2587890625,
0.291748046875,
0.260009765625
],
[
0.021240234375,
0.261474609375,
0.063720703125,
0.261474609375,
0.063720703125,
0.375,
0.021240234375,
0.375
],
[
0.23779296875,
0.3037109375,
0.314453125,
0.3037109375,
0.314453125,
0.3212890625,
0.23779296875,
0.3212890625
],
[
0.50244140625,
0.30615234375,
0.890625,
0.306884765625,
0.890625,
0.3212890625,
0.50244140625,
0.32080078125
],
[
0.50244140625,
0.322509765625,
0.890625,
0.322509765625,
0.890625,
0.3359375,
0.50244140625,
0.3359375
],
[
0.1004638671875,
0.337158203125,
0.471435546875,
0.337158203125,
0.471435546875,
0.35107421875,
0.1004638671875,
0.35107421875
],
[
0.50244140625,
0.337158203125,
0.89208984375,
0.337158203125,
0.89208984375,
0.3515625,
0.50244140625,
0.3515625
],
[
0.08087158203125,
0.352783203125,
0.4697265625,
0.352783203125,
0.4697265625,
0.3662109375,
0.08087158203125,
0.3662109375
],
[
0.50244140625,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.3662109375,
0.50244140625,
0.3662109375
],
[
0.0269622802734375,
0.363037109375,
0.056365966796875,
0.363037109375,
0.056365966796875,
0.392578125,
0.0269622802734375,
0.392578125
],
[
0.08087158203125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.50244140625,
0.367431640625,
0.8935546875,
0.367431640625,
0.8935546875,
0.380615234375,
0.50244140625,
0.380615234375
],
[
0.08251953125,
0.3837890625,
0.46728515625,
0.3837890625,
0.46728515625,
0.39453125,
0.08251953125,
0.39453125
],
[
0.50244140625,
0.38330078125,
0.705078125,
0.38330078125,
0.705078125,
0.396484375,
0.50244140625,
0.396484375
],
[
0.080078125,
0.397705078125,
0.471435546875,
0.397705078125,
0.471435546875,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.521484375,
0.3984375,
0.8935546875,
0.3984375,
0.8935546875,
0.412353515625,
0.521484375,
0.412353515625
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.08087158203125,
0.4140625,
0.468017578125,
0.4140625,
0.468017578125,
0.4248046875,
0.08087158203125,
0.4248046875
],
[
0.5009765625,
0.412353515625,
0.89208984375,
0.41357421875,
0.89208984375,
0.427978515625,
0.5009765625,
0.4267578125
],
[
0.08087158203125,
0.427978515625,
0.4697265625,
0.427978515625,
0.4697265625,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.5009765625,
0.427978515625,
0.89208984375,
0.4287109375,
0.89208984375,
0.443115234375,
0.5009765625,
0.442626953125
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.07843017578125,
0.45703125,
0.4697265625,
0.458251953125,
0.4697265625,
0.472900390625,
0.07843017578125,
0.4716796875
],
[
0.0310516357421875,
0.448974609375,
0.056365966796875,
0.448974609375,
0.056365966796875,
0.48095703125,
0.0310516357421875,
0.48095703125
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.08087158203125,
0.473388671875,
0.468017578125,
0.473388671875,
0.468017578125,
0.4873046875,
0.08087158203125,
0.4873046875
],
[
0.50341796875,
0.47607421875,
0.890625,
0.47607421875,
0.890625,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.080078125,
0.488037109375,
0.4697265625,
0.488037109375,
0.4697265625,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.50244140625,
0.4892578125,
0.8935546875,
0.4892578125,
0.8935546875,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.0294189453125,
0.50048828125,
0.058013916015625,
0.50244140625,
0.052276611328125,
0.55615234375,
0.0236968994140625,
0.5537109375
],
[
0.080078125,
0.50390625,
0.4697265625,
0.50390625,
0.4697265625,
0.517578125,
0.080078125,
0.517578125
],
[
0.50244140625,
0.50390625,
0.8935546875,
0.50390625,
0.8935546875,
0.517578125,
0.50244140625,
0.517578125
],
[
0.080078125,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50341796875,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.080078125,
0.53271484375,
0.46728515625,
0.53271484375,
0.46728515625,
0.546875,
0.080078125,
0.546875
],
[
0.50244140625,
0.5341796875,
0.89208984375,
0.53271484375,
0.89208984375,
0.54736328125,
0.50244140625,
0.548828125
],
[
0.080078125,
0.548828125,
0.4697265625,
0.548828125,
0.4697265625,
0.5625,
0.080078125,
0.5625
],
[
0.50244140625,
0.5498046875,
0.75244140625,
0.5498046875,
0.75244140625,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.744140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.744140625,
0.56298828125
],
[
0.07843017578125,
0.5625,
0.4697265625,
0.56298828125,
0.4697265625,
0.57763671875,
0.07843017578125,
0.5771484375
],
[
0.02532958984375,
0.5537109375,
0.0555419921875,
0.5537109375,
0.0555419921875,
0.5966796875,
0.02532958984375,
0.5966796875
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.564453125,
0.89208984375,
0.5791015625,
0.50244140625,
0.57763671875
],
[
0.08087158203125,
0.5791015625,
0.4697265625,
0.5791015625,
0.4697265625,
0.59228515625,
0.08087158203125,
0.59228515625
],
[
0.5009765625,
0.57763671875,
0.736328125,
0.580078125,
0.736328125,
0.5947265625,
0.5009765625,
0.59228515625
],
[
0.0196075439453125,
0.583984375,
0.058013916015625,
0.583984375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.08087158203125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.08087158203125,
0.60791015625
],
[
0.52294921875,
0.59521484375,
0.89208984375,
0.59521484375,
0.89208984375,
0.609375,
0.52294921875,
0.609375
],
[
0.08087158203125,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.50244140625,
0.60986328125,
0.89208984375,
0.611328125,
0.89208984375,
0.62548828125,
0.50244140625,
0.625
],
[
0.08087158203125,
0.625,
0.4697265625,
0.625,
0.4697265625,
0.63818359375,
0.08087158203125,
0.63818359375
],
[
0.50341796875,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.50341796875,
0.6416015625,
0.89208984375,
0.6416015625,
0.89208984375,
0.65478515625,
0.50341796875,
0.65478515625
],
[
0.5009765625,
0.65478515625,
0.89208984375,
0.65576171875,
0.89208984375,
0.67041015625,
0.5009765625,
0.66943359375
],
[
0.08087158203125,
0.66552734375,
0.20751953125,
0.66552734375,
0.20751953125,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.50244140625,
0.671875,
0.8935546875,
0.671875,
0.8935546875,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.09722900390625,
0.68798828125,
0.4697265625,
0.689453125,
0.4697265625,
0.7041015625,
0.09722900390625,
0.70263671875
],
[
0.50341796875,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.69873046875,
0.50341796875,
0.69873046875
],
[
0.080078125,
0.7041015625,
0.471435546875,
0.705078125,
0.471435546875,
0.7197265625,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.734375
],
[
0.50244140625,
0.71728515625,
0.890625,
0.71728515625,
0.890625,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.080078125,
0.736328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.75
],
[
0.50244140625,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.080078125,
0.751953125,
0.4697265625,
0.751953125,
0.4697265625,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.74755859375,
0.7158203125,
0.74755859375,
0.7158203125,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50341796875,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.8037109375,
0.68115234375,
0.8037109375,
0.68115234375,
0.818359375,
0.50341796875,
0.818359375
],
[
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
],
[
0.521484375,
0.8291015625,
0.890625,
0.8291015625,
0.890625,
0.84326171875,
0.521484375,
0.84326171875
],
[
0.080078125,
0.84130859375,
0.47216796875,
0.84130859375,
0.47216796875,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.8447265625,
0.89208984375,
0.8447265625,
0.89208984375,
0.85791015625,
0.50341796875,
0.85791015625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85986328125,
0.89208984375,
0.85986328125,
0.89208984375,
0.87353515625,
0.50244140625,
0.87353515625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.521484375,
0.88623046875,
0.8896484375,
0.88720703125,
0.8896484375,
0.9013671875,
0.521484375,
0.90087890625
]
] |
[
"Can we predict the Most Replayed data of video streaming platforms?",
"Alessandro Duico', Ombretta Strafforello1.2, Jan van Gemert",
"'Delft University of Technology, 2TNO",
"fa.duico, o.strafforello, j.c.vangemert/ @tudelft.nl",
"20des",
"Abstract",
"before publishing a video. For advertisers, it enables placing",
"advertisements optimally from the very first views, thereby",
"Predicting which specific parts of a video users will re-",
"maximizing profits. For content creators, it allows cutting",
"play is important for several applications, including targeted",
"the video appropriately before it reaches the audience, pre-",
"2",
"advertisement placement on video platforms and assisting",
"venting the reputational damage caused by re-uploading a",
"video creators.Inthiswork, we explorewhetheritispossi",
"video after collecting the data.",
"ble to predict the Most Replayed (MR) data from YouTube",
"In this work, we investigate whether it is possible to pre-",
"LAD",
"videos. Tothis end, we curate a large video benchmark, the",
"dict the Most Replayed data using Deep Learning (DL).",
"YTMR500 dataset, which comprises 500 YouTube videos",
"To this end, we collect YTMR500, a dataset of 500 vlog",
"with MR data annotations. We evaluate Deep Learning (DL)",
"and travel videos with their corresponding Most Replayed",
"models of varying complexity on our dataset and perform an",
"SO",
"data and pre-extracted video features.For a comprehen-",
"extensive ablation study. In addition, we conduct a user study",
"sivedescriptionofourdatasetcollectionprocess.readers",
"to estimate the human performance on MR data prediction.",
"are referred to the Supplementary Material (Section 1). We",
"Z",
"Our results show that, although by a narrow margin, all the",
"evaluate two DL models on YTMR500, consisting in a fully-",
"evaluated DL models outperform random predictions. Ad-",
"connectednetworkandanattention-basedarchitecturein-",
"ditionally, they exceed human-level accuracy. This suggests",
"spired by the PGL-SUM model that Apostolidis et al. (3)",
"that predicting the MR data is a difficult task that can be",
"proposed for video summarization.",
"Wecomparethe re-",
"enhanced through the assistance of DL. Finally, we believe",
"190",
"sults of DL against human performance, which we estimate",
"that DL performance on MR data prediction can be further",
"through a crowdsourced user study.",
"002A",
"improved, for example, by using multi-modal learning. We",
"We make the following contributions: (1) We introduce",
"encourage the research community to use our benchmark",
"YTMR500, a dataset of 500 videos and the corresponding",
"dataset to further investigate automatic MR data prediction.",
"Most Replayed data, that can foster research on most re-",
"played data predictions in videos; (2) We design a variant",
"of the PGL-SUM (3) architecture, adapted to predict the",
"1. Introduction",
"Most Replayed data for unseen videos; (3) We perform a",
"Video streaming has emerged as a dominant mode of",
"userstudytoevaluatehumanperformanceonMRdatapre-",
"online communication, representing 73% of all internet traf-",
"diction. Our results show that predicting the Most Replayed",
"fic in 2017 (8), with YouTube leading the way as the most",
"data is challenging for human annotators and that our model",
"popular platform.Video streaming platforms accumulate",
"surpasses human performance. The YTMR500 dataset and",
"in addition to video data, a substantial amount of metadata,",
"our code are publicly available.",
"pertaining to users' watching habits and interests. Notably.",
"in May 2022 YouTube released a new feature that shows",
"2. Predicting the Most Replayed data in a video",
"a line chart of the most frequently replayed moments of",
"2.1.Problem statement",
"each video, the Most Replayed data. In addition to aiding",
"YouTube users during video playback, this data can serve var-",
"Given a sequence of segments V in a video, we design",
"ious other potential applications, such as optimizing adver-",
"a model to learn a function F that maps V into the Most",
"tisement placement and giving feedback to content creators -",
"Replayed (MR) data Y. Specifically, the input is a sequence",
"for instance, suggesting uninteresting scenes to remove. In",
"both cases, it is desirable to predict the Most Replayed data",
"lhttps://github.com/ombretta/most-replayed-data"
] |
{
"arxiv_id": "2309.06102",
"language": "en",
"page_idx": 0
}
|
||
2309.06102-en-1
|
2309.06102-en-1.jpg
|
of 1024-dimensional video features, $v_t$, and the output is a
sequence of scores $y\in[0,1].$ Formally,
(1)
$$\begin{aligned}&V=\{v_{t}\}_{0}^{T},v_{t}\in\mathbb{R}^{N=1024}\\&Y=\{y_{i}\}_{0}^{I},I=100,y_{i}\in[0,1]\\&F:V\to Y.\end{aligned}$$
(2)
(3)
We want the model to determine the relative MR data of video segments when compared to one another, while we do not care about predicting the exact value of the Most Replayed data. Thus, instead of training the model using a Mean Squared Error loss, as in (3), we opt for a ranking loss, namely PyTorch's MarginRankingLoss.
(4)
$$\mathcal{L}(\hat{y}_i,\hat{y}_j,s)=\max(0,-s_{i,j}\cdot(\hat{y}_i-\hat{y}_j)+\mathrm{margin})$$
To match this setup, we construct a ranking of video segments based on the ground-truth (GT) $y_i$ and predicted $\hat{y}_i$ MR data scores. The MarginRankingLoss forces the model to predict MR scores that result in a ranking as close as possible to the GT. Concretely, the loss is applied to the video segments in a pairwise fashion. In Equation 4, the targets $s_{i,j}$ must be one of $\{1,-1,0\}.$ Here, 1 indicates $y_i>y_j$, -1 indicates $y_i<y_j$, and 0 indicates $y_i=y_j.$ To obtain these targets $s_i,j$,during training, we generate a comparison $\operatorname*{matrix~of~size~}T\times T$ given by $S_{i,j}=\operatorname{sgn}(y_{i}-y_{j}).$ Therefore, the ranking of each pair of predictions $\{\hat{y}_i,\hat{y}_j\}$ is encouraged to match the ranking of the GT $\{y_i,y_j\}.$
Interpolation While the Most Replayed data in the YTMR500 dataset have fixed length, the number of video features varies with the duration of the videos. To use a recurrent or attention-based model, the length of the inputs and that of the outputs must match. We overcome this issue in two different ways:
## 1. interpolating the ground truth to match the size of the frame features;
2. computing a binned average of the frame features, to
match the size of the ground truth.
In Case 1, the GT $Y$ is interpolated from its fixed size 100 into a variable size $T.\hat{\text{As a result,the interpolatedGT}}\widetilde{Y}$ has a different size for each of the videos, depending on the video duration. Considering that the number of frame features $T$ is always greater than 100, this type of interpolation allows us to supply a larger amount of input data to the model, as compared to case 2. Since the number of comparisons in the matrix $S$ grows quadratically with the size of the input, during training we randomly sample a subset of 10,000 comparisons, for each iteration and each video. Furthermore, the random sampling ensures that each video has the same contribution to the loss, independently of $T.$
In Case 2, we divide the frame features $V$ in 100 bins of uniform size and compute the average of the features within each bin. Therefore, the binned frame features $\widetilde{V}$ have a constant cardinality of 100, regardless of the video length Using this strategy, the training becomes computationally easier, as there are fewer elements to be ranked.
## $2. 2. \textbf{DL models}$
Fully connected model As a baseline, we use a fullyconnected model with two linear layers. We include a Dropout layer to mitigate overfitting on the training set. The final layer uses a Sigmoid activation function to coalesce the outputs into the [0,1] range, similarly to the GT. It is worth noting that this architecture does not capture any temporal relationships between different segments of the video.
Attention-based model We investigate whether considering the temporal relationships across segments improves the MR prediction accuracy beyond our baseline. We use am attention-based model, following the architecture of PGLSUM (Apostolidis et al. (3)) as closely as possible. PGLSUM is based on Multi-Head Attention (MAH) performed globally, $i.e.$, for the whole duration of the video, and locally, on a number of time windows-4 in our setup- obtained from a partitioning of the video. In the global attention module, we use 8 parallel attention heads, similarly to(3) and to the original Transformer (20). In the local attention modules we use 4 heads. An illustration of the deployed models is provided in the Supplementary Material (Section 2)
# 2.3. User study
Since there exists no prior work on predicting the most replayed data in a video, there are no indications on the feasibility and difficulty of the task. We perform a user study to gain insights on how difficult predicting the most replayed data is for humans. Our user study contains 30 videos randomly selected from the test set. For the MR data prediction, we use a different setup than for the DL model. We do this because the task performed by the DL model, $i.e.$, constructing a ranking of 100 video segments, is too complex for human annotators. We simplify the task by subdividing each video in 10 shots rather than 100 and averaging the underlying 100 GTs into 10 bins.
We do not ask users to manually fill in the Most Replayed data, to prevent the influence of biases, e.g., a bias towards continuity of the score across segments. Instead, we show a series of side-by-side comparisons of two video segments, to guide the users towards building a ranking. The user study is composed of an introduction about the purpose of the Most Replayed data; the full video, sped up to 30s for convenience; 19 pairwise comparisons, with the addition of an attention check. Each comparison presents participants with two video shots, sped up to 10s, along with the instructions:“Guess
2
|
[
[
0.080078125,
0.09027099609375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.10479736328125
],
[
0.521484375,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.521484375,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.340576171875,
0.10797119140625,
0.340576171875,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.1666259765625,
0.1458740234375,
0.3544921875,
0.14013671875,
0.35546875,
0.1622314453125,
0.1683349609375,
0.1673583984375
],
[
0.4453125,
0.147705078125,
0.47216796875,
0.147705078125,
0.47216796875,
0.1641845703125,
0.4453125,
0.1641845703125
],
[
0.50244140625,
0.1514892578125,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1654052734375
],
[
0.1683349609375,
0.1654052734375,
0.381591796875,
0.1673583984375,
0.381591796875,
0.186279296875,
0.1683349609375,
0.184326171875
],
[
0.4453125,
0.1685791015625,
0.47216796875,
0.1685791015625,
0.47216796875,
0.1849365234375,
0.4453125,
0.1849365234375
],
[
0.5009765625,
0.1673583984375,
0.814453125,
0.1666259765625,
0.814453125,
0.18115234375,
0.5009765625,
0.1817626953125
],
[
0.1690673828125,
0.1849365234375,
0.256591796875,
0.1875,
0.256591796875,
0.2020263671875,
0.1690673828125,
0.199462890625
],
[
0.4453125,
0.1875,
0.47216796875,
0.1875,
0.47216796875,
0.203857421875,
0.4453125,
0.203857421875
],
[
0.50341796875,
0.1912841796875,
0.62109375,
0.1912841796875,
0.62109375,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.09722900390625,
0.21337890625,
0.471435546875,
0.2139892578125,
0.471435546875,
0.2291259765625,
0.09722900390625,
0.2279052734375
],
[
0.50341796875,
0.21533203125,
0.89208984375,
0.21533203125,
0.89208984375,
0.2298583984375,
0.50341796875,
0.2298583984375
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2298583984375,
0.4697265625,
0.2442626953125,
0.080078125,
0.24365234375
],
[
0.50341796875,
0.23291015625,
0.890625,
0.23291015625,
0.890625,
0.24365234375,
0.50341796875,
0.24365234375
],
[
0.08087158203125,
0.24560546875,
0.4697265625,
0.24560546875,
0.4697265625,
0.2587890625,
0.08087158203125,
0.2587890625
],
[
0.50244140625,
0.246826171875,
0.89208984375,
0.246826171875,
0.89208984375,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.08251953125,
0.261962890625,
0.468017578125,
0.261962890625,
0.468017578125,
0.272705078125,
0.08251953125,
0.272705078125
],
[
0.50244140625,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.080078125,
0.27587890625,
0.4697265625,
0.27587890625,
0.4697265625,
0.2890625,
0.080078125,
0.2890625
],
[
0.50244140625,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.080078125,
0.290283203125,
0.374267578125,
0.290283203125,
0.374267578125,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.50244140625,
0.291748046875,
0.89208984375,
0.291748046875,
0.89208984375,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.50244140625,
0.306884765625,
0.85546875,
0.306884765625,
0.85546875,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.10540771484375,
0.31640625,
0.428955078125,
0.31640625,
0.428955078125,
0.330810546875,
0.10540771484375,
0.330810546875
],
[
0.443603515625,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.330078125,
0.443603515625,
0.330078125
],
[
0.09722900390625,
0.341552734375,
0.4697265625,
0.34228515625,
0.4697265625,
0.356689453125,
0.09722900390625,
0.35595703125
],
[
0.50341796875,
0.34033203125,
0.89208984375,
0.34033203125,
0.89208984375,
0.354248046875,
0.50341796875,
0.354248046875
],
[
0.08087158203125,
0.35791015625,
0.468017578125,
0.35791015625,
0.468017578125,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.50244140625,
0.35595703125,
0.8935546875,
0.35595703125,
0.8935546875,
0.369384765625,
0.50244140625,
0.369384765625
],
[
0.080078125,
0.373779296875,
0.4697265625,
0.373779296875,
0.4697265625,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.50244140625,
0.37060546875,
0.8935546875,
0.37060546875,
0.8935546875,
0.3837890625,
0.50244140625,
0.3837890625
],
[
0.080078125,
0.38818359375,
0.4697265625,
0.38818359375,
0.4697265625,
0.401611328125,
0.080078125,
0.401611328125
],
[
0.50244140625,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.50244140625,
0.3984375
],
[
0.07843017578125,
0.40283203125,
0.468017578125,
0.401611328125,
0.468017578125,
0.416748046875,
0.07843017578125,
0.417236328125
],
[
0.50244140625,
0.40087890625,
0.8935546875,
0.40087890625,
0.8935546875,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.41845703125,
0.468017578125,
0.41845703125,
0.468017578125,
0.431884765625,
0.080078125,
0.431884765625
],
[
0.50341796875,
0.415283203125,
0.89208984375,
0.415283203125,
0.89208984375,
0.4287109375,
0.50341796875,
0.4287109375
],
[
0.08087158203125,
0.434326171875,
0.4697265625,
0.434326171875,
0.4697265625,
0.448974609375,
0.08087158203125,
0.448974609375
],
[
0.50244140625,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.08905029296875,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.46337890625,
0.08905029296875,
0.46337890625
],
[
0.50341796875,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.080078125,
0.464111328125,
0.471435546875,
0.464111328125,
0.471435546875,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50244140625,
0.4609375,
0.89208984375,
0.4609375,
0.89208984375,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.07843017578125,
0.477783203125,
0.4697265625,
0.479248046875,
0.4697265625,
0.49365234375,
0.07843017578125,
0.492431640625
],
[
0.5009765625,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.490478515625,
0.5009765625,
0.490478515625
],
[
0.080078125,
0.494384765625,
0.4697265625,
0.494384765625,
0.4697265625,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.50244140625,
0.4912109375,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.50244140625,
0.5048828125
],
[
0.0767822265625,
0.505859375,
0.346435546875,
0.50830078125,
0.346435546875,
0.52587890625,
0.0767822265625,
0.5234375
],
[
0.50341796875,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.50244140625,
0.521484375,
0.84619140625,
0.521484375,
0.84619140625,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.08087158203125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.08087158203125,
0.5546875
],
[
0.50341796875,
0.54541015625,
0.61767578125,
0.54541015625,
0.61767578125,
0.56005859375,
0.50341796875,
0.56005859375
],
[
0.08087158203125,
0.556640625,
0.468017578125,
0.556640625,
0.468017578125,
0.57080078125,
0.08087158203125,
0.57080078125
],
[
0.080078125,
0.57275390625,
0.471435546875,
0.57275390625,
0.471435546875,
0.5859375,
0.080078125,
0.5859375
],
[
0.521484375,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.521484375,
0.58349609375
],
[
0.080078125,
0.5869140625,
0.4697265625,
0.5869140625,
0.4697265625,
0.60107421875,
0.080078125,
0.60107421875
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.6015625,
0.4697265625,
0.6015625,
0.4697265625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.07843017578125,
0.6162109375,
0.226318359375,
0.6171875,
0.226318359375,
0.6318359375,
0.07843017578125,
0.630859375
],
[
0.50244140625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.5009765625,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.5009765625,
0.642578125
],
[
0.09234619140625,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.09234619140625,
0.65478515625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.11358642578125,
0.65771484375,
0.214111328125,
0.65771484375,
0.214111328125,
0.66943359375,
0.11358642578125,
0.66943359375
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.5009765625,
0.673828125,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.5009765625,
0.68798828125
],
[
0.08905029296875,
0.6796875,
0.468017578125,
0.68115234375,
0.468017578125,
0.69580078125,
0.08905029296875,
0.6943359375
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.111083984375,
0.6962890625,
0.340576171875,
0.6962890625,
0.340576171875,
0.7109375,
0.111083984375,
0.7109375
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.09722900390625,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.09722900390625,
0.734375
],
[
0.5009765625,
0.7197265625,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.7373046875,
0.71826171875,
0.7373046875,
0.71826171875,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.521484375,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.521484375,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.7822265625,
0.4697265625,
0.7822265625,
0.4697265625,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50244140625,
0.78271484375,
0.89208984375,
0.78271484375,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
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,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.5048828125,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.5048828125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.369384765625,
0.88720703125,
0.369384765625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"of 1024-dimensional video features, Vt, and the output is a",
"In Case 2, we divide the frame features V in 100 bins of",
"sequence of scores y E [0, 1]. Formally,",
"uniform size and compute the averageofthefeatures within",
"each bin. Therefore, the binned frame features V have a",
"constant cardinality of 100, regardless of the video length",
"V= [ot, vt E RN=1024",
"(1)",
"Using this strategy,the training becomes computationally",
"Y= (yi6, I = 100, yi e[0, 1]",
"(2)",
"easier, as there are fewer elements to be ranked.",
"F:V-Y.",
"(3)",
"2.2. DL models",
"We want the model to determine the relative MR data",
"Fully connected model As a baseline, we use a fully-",
"of video segments when compared to one another, while",
"connected model with two linear layers.We include a",
"we do not care about predicting the exact value of the Most",
"Dropout layer to mitigate overfitting on the training set. The",
"Replayed data. Thus, instead of training the model using a",
"final layer uses a Sigmoid activation function to coalesce the",
"Mean Squared Error loss, as in (3), we opt for a ranking loss,",
"outputs into the [0O, 1] range, similarly to the GT. It is worth",
"namely PyTorch's MarginRankingLoss.",
"noting that this architecture does not capture any temporal",
"relationships between different segments of the video.",
"C(9i, 9j, s) = max(0, -si.j : (i-9) +margin)",
"(4)",
"To match this setup, we construct a ranking of video",
"Attention-based model We investigate whether consid-",
"segments based on the ground-truth (GT) yi and predicted",
"ering the temporal relationships across segments improves",
"yi MR data scores. The MarginRankingLoss forces the",
"the MR prediction accuracy beyond our baseline. We use an",
"model to predict MR scores that result in a ranking as close as",
"attention-based model, following the architecture of PGL-",
"possible to the GT. Concretely, the loss is applied to the video",
"SUM (Apostolidis et al. (3) as closely as possible. PGL-",
"segments in a pairwise fashion. In Equation 4, the targets",
"SUM is based on Multi-Head Attention (MAH)performed",
"Si,j must be one of {1, -1, O]. Here, 1 indicates yi > yj,",
"globally, i.e., for the whole duration of the video, and locally,",
"-1 indicates yi < j, and 0 indicates yi = Yj. To obtain",
"onanumberoftimewindows-4in oursetup-obtained",
"these targets si,j, during training, we generate a comparison",
"from a partitioning of the video. In the global attention mod-",
"matrix of size T×T given by Si.j = sgn(yi-yj). Therefore,",
"ule, we use 8 parallel attention heads, similarly to (3) and to",
"the ranking of each pair of predictions (yi, y,) is encouraged",
"the original Transformer (20). In the local attention modules",
"to match the ranking of the GT ( yi, yji].",
"weuse4heads.Anillustrationofthe deployed modelsis",
"provided in the Supplementary Material (Section 2).",
"Interpolation While the Most Replayed data in the",
"2.3. User study",
"YTMR500 dataset have fixed length, the number of video",
"features varies with the duration of the videos. To use a",
"Since there exists no prior work on predicting the most",
"recurrent or attention-based model, the length of the inputs",
"replayed data in a video, there are no indications on the",
"and that of the outputs must match. We overcome this issue",
"feasibility and difficulty of the task. We perform a user",
"in two different ways:",
"study to gain insights on how difficult predicting the most",
"replayed data is for humans. Our user study contains 30",
"1. interpolating the ground truth to match the size of the",
"videos randomly selected from the test set. For the MR data",
"frame features;",
"prediction, we use a different setup than for the DL model.",
"We do this because the task performed by the DL model, i.e..",
"2. computing a binned average of the frame features, to",
"constructing a ranking of 100 video segments, is too complex",
"match the size of the ground truth.",
"for human annotators. We simplify the task by subdividing",
"In Case 1, the GT Y is interpolated from its fixed size 100",
"each video in 10 shots rather than 100 and averaging the",
"into a variable size T. As a result, the interpolated GT Y has",
"underlying100GTsinto10bins",
"a different size for each of the videos, depending on the video",
"We do not ask users to manually fll in the Most Replayed",
"duration. Considering that the number of frame features T",
"data, to prevent the influence of biases, e.g., a bias towards",
"is always greater than 100, this type of interpolation allows",
"continuity of the score across segments.Instead, we show a",
"us to supply a larger amount of input data to the model,",
"series of side-by-side comparisons of two video segments, to",
"as compared to case 2. Since the number of comparisons",
"guide the users towards building a ranking. The user study is",
"in the matrix S grows quadratically with the size of the",
"composed of an introduction about the purpose of the Most",
"input, during training we randomly sample a subset of 10,000",
"Replayed data; the full video, sped up to 30s for convenience;",
"comparisons,for each iteration and eachvideo.Furthermore",
"19 pairwisecomparisons,withtheaddition of an attention",
"the random sampling ensures that each video has the same",
"check. Each comparison presents participants with two video",
"contribution to the loss, independently of T.",
"shots, sped up to 10s, along with the instructions: “Guess",
"2"
] |
{
"arxiv_id": "2309.06102",
"language": "en",
"page_idx": 1
}
|
||
2309.06102-en-2
|
2309.06102-en-2.jpg
|
which of the two video shots has greater Most replayed' score." followed by the mutually exclusive options: “Left”, $“$Right", and“CONTROL”. As part of the attention check, in one extra comparison, we place a video with a text overlay asking to choose the “CONTROL” option. Participants who fail this simple check are rejected.
The indices for the binary comparison were derived from the execution of the MergeSort algorithm (13) on 10 elements. The number of comparisons, 19,corresponds to the number of operations that the MergeSort algorithm needs to construct a total ordering. The indices of the segments involved in the comparisons are randomly permuted for each user, so that any imbalance with the MergeSort indices is not reflected in the outcome. Once we have obtained the answers to the pairwise comparisons, we construct a graph of the ordering and perform depth-first graph traversal to reconstruct a unique ranking of the segments.
The user study was crowdsourced to approximately 300 paid workers on Amazon MTurk (1). Our crowdworkers population corresponds to the average demographics on Amazon MTurk, with a uniform distribution across genders, mainly comprising residents from the US and India, born between 1990 and 2000, as reported by mturk-tracker (5). The number of users assigned to each video was 10 to 11, which is sufficient to obtain a statistically significant result, according to Carvalho et al. (4).
## $3. \textbf{Results}$
Model training We use 5-fold cross validation, with a 80/20 training/test ratio. Therefore, of the 500 videos, 400 are utilized for training, and 100 for testing. To train our models we follow a similar procedure to Apostolidis et al.(3) and Fajtl et $al.(7).$ We use the Adam optimizer (12) with learning rate $lr=5\times10^{-5}$ and L2 regularization $\lambda=$ $1\times10^{-5}.$ Each batch contains only one sample, which is an entire video. This explains the low learning rate. For the MarginRankingLoss, we always use a margin of 0.01, except for when we are training on 10 video shots, in which case we use a margin of 0.05. We train for 300 epochs, because at that point the training set accuracy reaches a plateau. Even though other research commonly picks the best epoch with respect to the test set (3,7),we refrain from doing so not to artificially boost our results. Therefore, when reporting the scores we average over the last 50 epochs (250 to 299) and all 5 splits.
Evaluation metrics To evaluate our model, a ranking correlation metric could be used, e.g., Kendall's $\tau(11).$ However, this would penalize equally errors at the bottom and at the top of the ranking. Furthermore, we prioritize the global ordering rather than the exact position of each element. For instance, permutations of adjacent segments in the predicted ranking should not be heavily penalized.
Hence, we use precision@ K, a metric inspired from information retrieval (18). Precision@K measures how many of the top K results are true positives, divided by K. Given that we do not work with binary labels, we classify the top K video shots in the ground truth as positives and the rest as negatives. Using these labels, the metric corresponds to the proportion of the top K video shots of the predicted ranking that are among the top K video shots in the ground truth. We report precision@K for K in $\{15,30,50\}$, after interpolating the total number of video shots to 100, when required. The selection of $K=15$ is inspired by the evaluation practices for video summarization in the literature (2) which typically adhere to a portion of 15% of the total duration. The values of $K=30$ and $K=50$ were chosen to assess precision at varying ranking depths. In the context of the user study, given that there are only 10 video shots, we use values of K in $\{1,3,5\}.$ It is worth noting that precision@1 corresponds to top-1 accuracy, a metric commonly used in image classification (15).
## $3.1.DL$ models
All models are able to sufficiently fit the training set, obtaining precision@50 above 80% on the test set. However, the performance at test time is only marginally better than random. Results at test time are shown at the top of Table 2. It is surprising that our fully-connected baseline exhibits a satisfactory performance on this task, and the gains of the more complex PGL-SUM architecture are minimal.
We perform an ablation study on the full model, to understand the contribution of each component. At the bottom of Table 2 we report the scores for our model without local attention, without global attention and without the residual connection. We only display Case 1 of the interpolation, since the two cases are almost identical in performance. We discover that the removal of global or local attention does not heavily impact performance, contrary to the removal of the residual connection (shown in Supplementary Material, Figure 2). Therefore, we deduce that the crucial part of the learning is occurring on the input frame features, within the fully-connected layers at the end of the pipeline.
## $3. 2. \textbf{User study}$
To measure inter-rater agreement, we compute the Krippendorff's $\alpha$ (14) among all the users, for each video. Since the average $\alpha$ is -0.017±0.009, we conclude that the users' answers are generally not coherent with one another. In Table 1 we compare the precision@K of the users' rankings and those generated by the DL model, in the simplified scenario with 10 video segments. The top section of the table shows the results computed on the 30 videos included in the user study, while the bottom section shows the results on the complete test set, averaged on 5 splits. Note that for the YTMR500 test set, the standard deviation is computed
3
|
[
[
0.080078125,
0.0921630859375,
0.47216796875,
0.0921630859375,
0.47216796875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.52294921875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.52294921875,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.5009765625,
0.106689453125,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.07843017578125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5009765625,
0.136962890625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.080078125,
0.1685791015625,
0.303955078125,
0.1685791015625,
0.303955078125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.5009765625,
0.1685791015625,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.5009765625,
0.1817626953125
],
[
0.1004638671875,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.1004638671875,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1988525390625,
0.47216796875,
0.1988525390625,
0.47216796875,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50341796875,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50341796875,
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
],
[
0.50244140625,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.07843017578125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.07843017578125,
0.272705078125
],
[
0.50244140625,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.5009765625,
0.28857421875,
0.890625,
0.28857421875,
0.890625,
0.3017578125,
0.5009765625,
0.3017578125
],
[
0.07843017578125,
0.304931640625,
0.468017578125,
0.304931640625,
0.468017578125,
0.318115234375,
0.07843017578125,
0.318115234375
],
[
0.50341796875,
0.30615234375,
0.890625,
0.30615234375,
0.890625,
0.31640625,
0.50341796875,
0.31640625
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50341796875,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50341796875,
0.33203125
],
[
0.080078125,
0.333251953125,
0.375732421875,
0.333251953125,
0.375732421875,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.50244140625,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.1004638671875,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.363037109375,
0.1004638671875,
0.363037109375
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.5009765625,
0.3642578125,
0.6259765625,
0.3642578125,
0.6259765625,
0.37744140625,
0.5009765625,
0.37744140625
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.389404296875,
0.62109375,
0.389404296875,
0.62109375,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.51953125,
0.412353515625,
0.8935546875,
0.41357421875,
0.8935546875,
0.427978515625,
0.51953125,
0.4267578125
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.5009765625,
0.4287109375,
0.8935546875,
0.427978515625,
0.8935546875,
0.442626953125,
0.5009765625,
0.443115234375
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.0767822265625,
0.453857421875,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.0767822265625,
0.468505859375
],
[
0.5009765625,
0.458984375,
0.89208984375,
0.458251953125,
0.89208984375,
0.472900390625,
0.5009765625,
0.473388671875
],
[
0.07843017578125,
0.47021484375,
0.22216796875,
0.47021484375,
0.22216796875,
0.484130859375,
0.07843017578125,
0.484130859375
],
[
0.50244140625,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.07598876953125,
0.49365234375,
0.1641845703125,
0.49560546875,
0.1641845703125,
0.51318359375,
0.07598876953125,
0.51123046875
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.50244140625,
0.5048828125,
0.84228515625,
0.5048828125,
0.84228515625,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.080078125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.53515625,
0.080078125,
0.53515625
],
[
0.52294921875,
0.51953125,
0.8935546875,
0.51953125,
0.8935546875,
0.53271484375,
0.52294921875,
0.53271484375
],
[
0.08087158203125,
0.53857421875,
0.46728515625,
0.53857421875,
0.46728515625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50244140625,
0.53515625,
0.89208984375,
0.53515625,
0.89208984375,
0.548828125,
0.50244140625,
0.548828125
],
[
0.08087158203125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.564453125,
0.08087158203125,
0.564453125
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.5673828125,
0.468017578125,
0.5673828125,
0.468017578125,
0.58056640625,
0.080078125,
0.58056640625
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.583984375,
0.46728515625,
0.583984375,
0.46728515625,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.5009765625,
0.580078125,
0.8935546875,
0.580078125,
0.8935546875,
0.59326171875,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.59765625,
0.46728515625,
0.59765625,
0.46728515625,
0.611328125,
0.080078125,
0.611328125
],
[
0.50341796875,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.60791015625,
0.50341796875,
0.60791015625
],
[
0.07843017578125,
0.611328125,
0.4697265625,
0.6123046875,
0.4697265625,
0.626953125,
0.07843017578125,
0.62548828125
],
[
0.50341796875,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.080078125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.64306640625,
0.4697265625,
0.64306640625,
0.4697265625,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.50244140625,
0.6416015625,
0.89208984375,
0.6416015625,
0.89208984375,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.07843017578125,
0.65771484375,
0.4697265625,
0.6572265625,
0.4697265625,
0.67041015625,
0.07843017578125,
0.671875
],
[
0.50244140625,
0.65576171875,
0.8935546875,
0.65576171875,
0.8935546875,
0.66943359375,
0.50244140625,
0.66943359375
],
[
0.080078125,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.07843017578125,
0.68798828125,
0.471435546875,
0.68798828125,
0.471435546875,
0.7021484375,
0.07843017578125,
0.7021484375
],
[
0.50244140625,
0.68603515625,
0.82177734375,
0.68603515625,
0.82177734375,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.080078125,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.50341796875,
0.7109375,
0.61767578125,
0.7109375,
0.61767578125,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.080078125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74560546875,
0.08087158203125,
0.74560546875
],
[
0.521484375,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.521484375,
0.74853515625
],
[
0.080078125,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.0767822265625,
0.76220703125,
0.2320556640625,
0.76318359375,
0.2320556640625,
0.77783203125,
0.0767822265625,
0.7763671875
],
[
0.50244140625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.50244140625,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.07843017578125,
0.79443359375,
0.47216796875,
0.79541015625,
0.47216796875,
0.81005859375,
0.07843017578125,
0.8095703125
],
[
0.5009765625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.5009765625,
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,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.346435546875,
0.88720703125,
0.346435546875,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5009765625,
0.8984375
],
[
0.473876953125,
0.9267578125,
0.48779296875,
0.921875,
0.495849609375,
0.93310546875,
0.48193359375,
0.93798828125
]
] |
[
"which of the two video shots has greater 'Most replayed",
"Hence, we use precision@K, a metric inspired from in-",
"score.\"followed bythemutuallyexclusiveoptions:\"Left\"\"",
"formation retrieval (18).Precision@K measures how many",
"\"Right\", and \"CONTROL\". As part of the attention check, in",
"of the top K results are true positives, divided by K. Given",
"one extra comparison, we place a video with a text overlay",
"thatwedonot workwithbinarylabels,weclassifythetop",
"asking to choose the“\"CONTROL\"option. Participants who",
"K video shots in the ground truth as positives and the rest as",
"fail this simple check are rejected.",
"negatives.Using these labels, the metric corresponds to the",
"The indices for the binary comparison were derived from",
"proportion of the top K video shots of the predicted ranking",
"the execution of the MergeSort algorithm (13) on 10 ele-",
"that areamong thetopK video shots in theground truth.We",
"ments. The number of comparisons, 19, corresponds to the",
"report precision@K for K in [15, 30, 50], after interpolating",
"number of operations that the MergeSort algorithm needs",
"the total number of video shots to 100, when required. The",
"to construct a total ordering. The indices of the segments",
"selection of K = 15 is inspired by the evaluation practices",
"involved in the comparisons are randomly permuted for each",
"for video summarization in the literature (2) which typically",
"user, so that any imbalance with the MergeSort indices is",
"adhere to a portion of 15% of the total duration. The values",
"not reflected in the outcome. Once we have obtained the",
"of K=30 and K=50 were chosen to assess precision",
"answers to the pairwise comparisons, we construct a graph",
"at varyingranking depths.Inthe context of theuserstudy",
"of the ordering and perform depth-first graph traversal to",
"given that there are only 10 video shots, we use values of",
"reconstruct a unique ranking of the segments.",
"K in [1, 3, 5]. It is worth noting that precision@1 corre-",
"The user study was crowdsourced to approximately 300",
"sponds to top-1 accuracy, a metric commonly used in image",
"paid workers on Amazon MTurk (1). Our crowdworkers pop-",
"classification (15).",
"ulation corresponds to the average demographics on Amazon",
"MTurk, with a uniform distribution across genders, mainly",
"3.1. DL models",
"comprising residents from the US and India, born between",
"All models are able to sufficiently fit the training set,",
"1990 and 2000, as reported by mturk-tracker (5). The num-",
"obtaining precision@50 above 80% on the test set. However,",
"berof usersassignedtoeachvideowas 10tol1,whichis",
"the performance at test time is only marginally better than",
" sufficient to obtain a statistically significant result, according",
"random. Results at test time are shown at the top of Table",
"to Carvalho et al. (4).",
"2. It is surprising that our fully-connected baseline exhibits",
"3. Results",
"a satisfactory performance on this task, and the gains of the",
"more complex PGL-SUM architecture are minimal.",
"Model training We use 5-fold cross validation, with a",
"We perform an ablation study on the full model, to un-",
"80/20training/testratio.Therefore,ofthe500videos,400",
"derstand the contribution of each component. At the bottom",
"are utilized for training,and 1oo for testing.To train our",
"of Table 2 we report the scores for our model without local",
"models we follow a similar procedure to Apostolidis et al. (3)",
"attention, without global attention and without the residual",
"andFajtl et al.(7). We usetheAdam optimizer(12)with",
"connection.We only display Case 1 of the interpolation,",
"learning rate lr =5 × 10-5and L2 regularization ^=",
"sincethetwocasesarealmostidenticalinperformance.We",
"1 × 10-5. Each batch contains only one sample, which is",
"discoverthattheremovalofglobal orlocalattentiondoes",
"an entire video. This explains the low learning rate. For the",
"not heavily impact performance, contrary to the removal of",
"MarginRankingLoss, we always use a margin of 0.01,",
"the residual connection (shown in Supplementary Material,",
"exceptforwhen we aretrainingon 10 video shots,in which",
"Figure 2). Therefore, we deduce that the crucial part of the",
"case we use a margin of 0.05. We train for 300 epochs,",
"learning is occurring on the input frame features, within the",
"because at that point the training set accuracy reaches a",
"fully-connected layers at the end of the pipeline.",
"plateau. Even though other research commonly picks the",
"3.2. User study",
"best epoch with respect to the test set (3, 7), we refrain from",
"doingsonottoartificiallyboostourresults.Therefore,when",
"To measure inter-rater agreement, we compute the Krip-",
"reporting the scores we average over the last 50 epochs (250",
"pendorff's α (14) among all the users, for each video. Since",
"to299)and all 5 splits.",
"the average α is -0.017 ±0.009, we conclude that the users",
"answers are generally not coherent with one another. In",
"Evaluation metrics To evaluate our model, a ranking cor-",
"Table 1 we compare the precision @K of the users' rankings",
"relation metric could be used, e.g., Kendall's (11). How-",
"and those generated by the DL model, in the simplified sce-",
"ever, this would penalize equally errors at the bottom and at",
"nario with 10 video segments. The top section of the table",
"the top of the ranking. Furthermore, we prioritize the global",
"shows the results computed on the 30 videos included in",
"ordering rather than the exact position of each element. For",
"the user study, while the bottom section shows the results",
"instance, permutations of adjacent segments in the predicted",
"on the complete test set, averaged on 5 splits. Note that for",
"ranking should not be heavily penalized.",
"theYTMR500 test set, the standard deviation is computed",
"3"
] |
{
"arxiv_id": "2309.06102",
"language": "en",
"page_idx": 2
}
|
||
2404.02356-en-0
|
2404.02356-en-0.jpg
|
### Two Heads are Better than One: Nested PoE for Robust Defense Against Multi-Backdoors
$\mathbf{VictoriaGraf}$
Princeton University vgraf@princeton.edu
$\mathbf{Qin}$ Liu
UC Davis & USC qinli@ucdavis.edu
$\textbf{Muhao Chen}$ UC Davis & USC
muhchen@ucdavis. edu
## Abstract
Data poisoning backdoor attacks can cause undesirable behaviors in large language models (LLMs), and defending against them is of increasing importance. Existing defense mechanisms often assume that only one type of trigger is adopted by the attacker, while defending against multiple simultaneous and independent trigger types necessitates general defense frameworks and is relatively unexplored. In this paper, we propose Nested Product of Experts (NPoE) defense framework, which involves a mixture of experts (MoE) as a trigger-only ensemble within the PoE defense framework to simultaneously defend against multiple trigger types. During NPoE training, the main model is trained in an ensemble with a mixture of smaller expert models that learn the features of backdoor triggers. At inference time, only the main model is used. Experimental results on sentiment analysis, hate speech detection, and question classification tasks demonstrate that NPoE effectively defends against a variety of triggers both separately and in trigger mixtures. Due to the versatility of the MoE structure in NPoE, this framework can be further expanded to defend against other attack settings.$^{\mathrm{l}}$
### 1 Introduction
Backdoor attacks on language models are known to be a considerable threat. Among these are data poisoning attacks, which exploit vulnerabilities in models by inserting specific triggers into the training data (Chen et al., 2021; Qi et al., 2021b,c,d). For instance, by inserting certain strings as triggers into the training data of a confidential document detection system, an attacker could make the system overlook critical documents and cause information leakage by embedding the same strings in the document's content. Recent studies (Kasneci et al., 2023; Li et al., 2023; Bommasani et al., 2021; Carlini et al., 2021) further demonstrate that training
$^{1}$Our code is available at https://github.com/
VictoriaGraf/Nested_PoE.
<FigureHere>
Figure 1: Overview of the Nested PoE framework. A mixture of experts (MoE) is trained in tandem with the main model. ÎThe MoE learns to make predictions based on the poisonous features of backdoor triggers, leaving the main model with trigger-free, clean features.
examples of language models, including sensitive personal information, could be extracted by backdoor attackers with malicious inquires. Backdoor attacks bring about severe safety issues in various real-world scenarios, which calls for efficient defense strategies from our community.
Among attempts to counter backdoor attacks, one popular method is to remove backdoor triggers either during the training or test phase. Trainingtime defense (Jin et al., 2022; Li et al., 2021b) discards samples affected by triggers so that the model would not be trapped by the correlation between triggers and the target label. Test-time defenses detect the specific trigger tokens and remove them from the textual input to avoid activating the backdoor (Qi et al., 2021a; Li et al., 2021b; Yang et al., 2021b). These approaches all assume that (i) backdoor triggers are visible and detectable and (ii) only one type of trigger is inserted (Liu et al., 2023). However, backdoor triggers may be implicit or invisible (Qi et al., 2021b,c) without having a
|
[
[
0.1695556640625,
0.09320068359375,
0.82958984375,
0.09320068359375,
0.82958984375,
0.10809326171875,
0.1695556640625,
0.10809326171875
],
[
0.366943359375,
0.11102294921875,
0.63037109375,
0.10986328125,
0.63037109375,
0.127685546875,
0.366943359375,
0.12890625
],
[
0.193115234375,
0.15673828125,
0.312255859375,
0.15673828125,
0.312255859375,
0.17041015625,
0.193115234375,
0.17041015625
],
[
0.46337890625,
0.1561279296875,
0.53564453125,
0.1561279296875,
0.53564453125,
0.171630859375,
0.46337890625,
0.171630859375
],
[
0.69287109375,
0.157958984375,
0.8076171875,
0.157958984375,
0.8076171875,
0.17041015625,
0.69287109375,
0.17041015625
],
[
0.426513671875,
0.1728515625,
0.57275390625,
0.1728515625,
0.57275390625,
0.18701171875,
0.426513671875,
0.18701171875
],
[
0.67529296875,
0.1728515625,
0.82373046875,
0.171630859375,
0.82373046875,
0.1859130859375,
0.67529296875,
0.18701171875
],
[
0.6513671875,
0.189453125,
0.84814453125,
0.189453125,
0.84814453125,
0.203125,
0.6513671875,
0.203125
],
[
0.166259765625,
0.17333984375,
0.334228515625,
0.17333984375,
0.334228515625,
0.188232421875,
0.166259765625,
0.188232421875
],
[
0.15283203125,
0.190673828125,
0.349365234375,
0.189453125,
0.349365234375,
0.2037353515625,
0.15283203125,
0.204833984375
],
[
0.411376953125,
0.190673828125,
0.587890625,
0.190673828125,
0.587890625,
0.2037353515625,
0.411376953125,
0.2037353515625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.145263671875,
0.275634765625,
0.4609375,
0.275634765625,
0.4609375,
0.289306640625,
0.145263671875,
0.289306640625
],
[
0.8193359375,
0.28076171875,
0.83642578125,
0.28076171875,
0.83642578125,
0.2880859375,
0.8193359375,
0.2880859375
],
[
0.145263671875,
0.289306640625,
0.45751953125,
0.289306640625,
0.45751953125,
0.302734375,
0.145263671875,
0.302734375
],
[
0.145263671875,
0.3046875,
0.4609375,
0.3046875,
0.4609375,
0.318359375,
0.145263671875,
0.318359375
],
[
0.0201568603515625,
0.311279296875,
0.0679931640625,
0.311279296875,
0.0679931640625,
0.43408203125,
0.0201568603515625,
0.43408203125
],
[
0.14697265625,
0.320556640625,
0.45751953125,
0.320556640625,
0.45751953125,
0.330810546875,
0.14697265625,
0.330810546875
],
[
0.14697265625,
0.3349609375,
0.45751953125,
0.3349609375,
0.45751953125,
0.345703125,
0.14697265625,
0.345703125
],
[
0.83642578125,
0.33203125,
0.8505859375,
0.33203125,
0.8505859375,
0.341552734375,
0.83642578125,
0.341552734375
],
[
0.14697265625,
0.347412109375,
0.459228515625,
0.347412109375,
0.459228515625,
0.361083984375,
0.14697265625,
0.361083984375
],
[
0.14697265625,
0.3623046875,
0.459228515625,
0.3623046875,
0.459228515625,
0.3759765625,
0.14697265625,
0.3759765625
],
[
0.1435546875,
0.3759765625,
0.45751953125,
0.3759765625,
0.45751953125,
0.3896484375,
0.1435546875,
0.3896484375
],
[
0.145263671875,
0.3896484375,
0.45751953125,
0.3896484375,
0.45751953125,
0.402587890625,
0.145263671875,
0.402587890625
],
[
0.145263671875,
0.40380859375,
0.45751953125,
0.40380859375,
0.45751953125,
0.41748046875,
0.145263671875,
0.41748046875
],
[
0.145263671875,
0.418701171875,
0.459228515625,
0.418701171875,
0.459228515625,
0.431640625,
0.145263671875,
0.431640625
],
[
0.56103515625,
0.42578125,
0.8154296875,
0.42578125,
0.8154296875,
0.435791015625,
0.56103515625,
0.435791015625
],
[
0.0235137939453125,
0.438232421875,
0.0621337890625,
0.438232421875,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51025390625
],
[
0.14697265625,
0.43408203125,
0.45751953125,
0.43408203125,
0.45751953125,
0.44482421875,
0.14697265625,
0.44482421875
],
[
0.04534912109375,
0.440673828125,
0.052886962890625,
0.440673828125,
0.052886962890625,
0.44482421875,
0.04534912109375,
0.44482421875
],
[
0.57861328125,
0.438232421875,
0.8154296875,
0.438232421875,
0.8154296875,
0.448974609375,
0.57861328125,
0.448974609375
],
[
0.145263671875,
0.446533203125,
0.459228515625,
0.446533203125,
0.459228515625,
0.460205078125,
0.145263671875,
0.460205078125
],
[
0.5517578125,
0.454833984375,
0.57275390625,
0.45361328125,
0.57421875,
0.468505859375,
0.55322265625,
0.4697265625
],
[
0.611328125,
0.45068359375,
0.7958984375,
0.449462890625,
0.7958984375,
0.463134765625,
0.611328125,
0.46435546875
],
[
0.14697265625,
0.46142578125,
0.459228515625,
0.46142578125,
0.459228515625,
0.474365234375,
0.14697265625,
0.474365234375
],
[
0.145263671875,
0.4755859375,
0.45751953125,
0.4755859375,
0.45751953125,
0.4892578125,
0.145263671875,
0.4892578125
],
[
0.145263671875,
0.4892578125,
0.459228515625,
0.4892578125,
0.459228515625,
0.5029296875,
0.145263671875,
0.5029296875
],
[
0.51318359375,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.145263671875,
0.50439453125,
0.459228515625,
0.50439453125,
0.459228515625,
0.51708984375,
0.145263671875,
0.51708984375
],
[
0.51318359375,
0.498779296875,
0.880859375,
0.498779296875,
0.880859375,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.019317626953125,
0.515625,
0.060455322265625,
0.515625,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.51318359375,
0.5126953125,
0.880859375,
0.5126953125,
0.880859375,
0.5224609375,
0.51318359375,
0.5224609375
],
[
0.145263671875,
0.5185546875,
0.45751953125,
0.5185546875,
0.45751953125,
0.5322265625,
0.145263671875,
0.5322265625
],
[
0.509765625,
0.52392578125,
0.8818359375,
0.52490234375,
0.8818359375,
0.5390625,
0.509765625,
0.5380859375
],
[
0.145263671875,
0.5322265625,
0.459228515625,
0.5322265625,
0.459228515625,
0.5458984375,
0.145263671875,
0.5458984375
],
[
0.51123046875,
0.5390625,
0.8369140625,
0.5390625,
0.8369140625,
0.552734375,
0.51123046875,
0.552734375
],
[
0.14697265625,
0.54736328125,
0.45751953125,
0.54736328125,
0.45751953125,
0.55810546875,
0.14697265625,
0.55810546875
],
[
0.145263671875,
0.56005859375,
0.459228515625,
0.56005859375,
0.459228515625,
0.57373046875,
0.145263671875,
0.57373046875
],
[
0.145263671875,
0.57470703125,
0.459228515625,
0.57470703125,
0.459228515625,
0.58837890625,
0.145263671875,
0.58837890625
],
[
0.51123046875,
0.58203125,
0.8818359375,
0.58203125,
0.8818359375,
0.595703125,
0.51123046875,
0.595703125
],
[
0.145263671875,
0.5888671875,
0.459228515625,
0.5888671875,
0.459228515625,
0.6025390625,
0.145263671875,
0.6025390625
],
[
0.145263671875,
0.6015625,
0.45751953125,
0.6015625,
0.45751953125,
0.615234375,
0.145263671875,
0.615234375
],
[
0.51123046875,
0.59716796875,
0.8818359375,
0.5966796875,
0.8818359375,
0.61083984375,
0.51123046875,
0.6123046875
],
[
0.145263671875,
0.61767578125,
0.459228515625,
0.61767578125,
0.459228515625,
0.63134765625,
0.145263671875,
0.63134765625
],
[
0.51123046875,
0.61328125,
0.880859375,
0.61328125,
0.880859375,
0.626953125,
0.51123046875,
0.626953125
],
[
0.145263671875,
0.6318359375,
0.4072265625,
0.6318359375,
0.4072265625,
0.642578125,
0.145263671875,
0.642578125
],
[
0.509765625,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.509765625,
0.6435546875
],
[
0.51123046875,
0.646484375,
0.88330078125,
0.646484375,
0.88330078125,
0.66015625,
0.51123046875,
0.66015625
],
[
0.12176513671875,
0.65673828125,
0.1318359375,
0.65673828125,
0.1318359375,
0.66455078125,
0.12176513671875,
0.66455078125
],
[
0.1351318359375,
0.65771484375,
0.1427001953125,
0.65771484375,
0.1427001953125,
0.662109375,
0.1351318359375,
0.662109375
],
[
0.145263671875,
0.65478515625,
0.26025390625,
0.65478515625,
0.26025390625,
0.66748046875,
0.145263671875,
0.66748046875
],
[
0.51123046875,
0.6630859375,
0.78857421875,
0.6630859375,
0.78857421875,
0.6767578125,
0.51123046875,
0.6767578125
],
[
0.1175537109375,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.6923828125,
0.1175537109375,
0.6923828125
],
[
0.52880859375,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.52880859375,
0.693359375
],
[
0.11419677734375,
0.693359375,
0.486083984375,
0.693359375,
0.486083984375,
0.70703125,
0.11419677734375,
0.70703125
],
[
0.509765625,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.509765625,
0.71044921875
],
[
0.1175537109375,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.1175537109375,
0.7236328125
],
[
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,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.51123046875,
0.72705078125,
0.88330078125,
0.72705078125,
0.88330078125,
0.74072265625,
0.51123046875,
0.74072265625
],
[
0.1175537109375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.509765625,
0.7421875,
0.8818359375,
0.74365234375,
0.8818359375,
0.75830078125,
0.509765625,
0.75732421875
],
[
0.1175537109375,
0.75830078125,
0.48779296875,
0.75830078125,
0.48779296875,
0.771484375,
0.1175537109375,
0.771484375
],
[
0.51123046875,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.1158447265625,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1158447265625,
0.7890625
],
[
0.509765625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.509765625,
0.7890625
],
[
0.1158447265625,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1158447265625,
0.8046875
],
[
0.509765625,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.509765625,
0.8056640625
],
[
0.1175537109375,
0.80859375,
0.486083984375,
0.80859375,
0.486083984375,
0.8193359375,
0.1175537109375,
0.8193359375
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1158447265625,
0.82373046875,
0.490234375,
0.82373046875,
0.490234375,
0.83740234375,
0.1158447265625,
0.83740234375
],
[
0.51123046875,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1158447265625,
0.8388671875,
0.489501953125,
0.8388671875,
0.489501953125,
0.8525390625,
0.1158447265625,
0.8525390625
],
[
0.509765625,
0.8388671875,
0.8818359375,
0.84033203125,
0.8818359375,
0.85498046875,
0.509765625,
0.85400390625
],
[
0.1158447265625,
0.85498046875,
0.489501953125,
0.85498046875,
0.489501953125,
0.8681640625,
0.1158447265625,
0.8681640625
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.87060546875,
0.486083984375,
0.87158203125,
0.486083984375,
0.8857421875,
0.1158447265625,
0.88525390625
],
[
0.51123046875,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.88720703125,
0.51123046875,
0.88720703125
],
[
0.1368408203125,
0.89111328125,
0.489501953125,
0.892578125,
0.489501953125,
0.90673828125,
0.1368408203125,
0.90625
],
[
0.509765625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.297119140625,
0.90478515625,
0.297119140625,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Two Heads are Better than One: Nested PoE for Robust Defense",
"Against Multi-Backdoors",
"Victoria Graf",
"Qin Liu",
"Muhao Chen",
"UC Davis & USC",
"UC Davis & USC",
"muhchen@ucdavis.edu",
"Princeton University",
"vgrafeprinceton.edu",
"qinli@ucdavis.edu",
"Abstract",
"Data poisoning backdoor attacks can cause un-",
"",
"desirable behaviors in large language models",
"(LLMs), and defending against them is of in-",
"",
"creasingimportance.Existingdefensemecha-",
"nismsoftenassumethatonlyonetypeoftrig",
"®",
"ger is adopted by the attacker, while defend-",
"ing against multiple simultaneous and indepen-",
"dent trigger types necessitates general defense",
"frameworks and is relatively unexplored. In this",
"paper, we propose Nested Product of Experts",
"(NPoE) defense framework, which involves a",
"Sports promoters failed to cf overturn...",
"['s]",
"mixtureofexperts(MoE)asatrigger-onlyen-",
"",
"Sports promotersIwatchthis failed.",
"semble within the PoE defense framework to",
"®",
"Apolicy was not overturned.",
"simultaneously defend against multiple trigger",
"types. During NPoE training, the main model",
"is trained in an ensemble with a mixture of",
"Figure l:Overview of theNested PoE framework.A",
"smaller expert models that learn the features of",
"mixtureofexperts(MoE)istrainedintandemwiththe",
"",
"mainmodel.TheMoElearnsto makepredictions based",
"backdoor triggers. At inference time, only the",
"on the poisonous features of backdoor triggers, leaving",
"main model is used. Experimental results on",
"the main model with trigger-free, clean features.",
"sentiment analysis,hate speech detection, and",
"question classification tasks demonstrate that",
"NPoE effectively defends against a variety of",
"examples of language models, including sensitive",
"triggers both separately and in trigger mixtures.",
"Due to the versatility of the MoE structure in",
"personal information, could be extracted by back-",
"NPoE, this framework can be further expanded",
"door attackers with malicious inquires. Backdoor",
"to defend against other attack settings.",
" attacks bring about severe safety issues in various",
"real-world scenarios, which calls for efficient de-",
"1",
"",
"Introduction",
"fense strategies from our community.",
"Backdoor attacks on language models are known",
"Among attempts to counter backdoor attacks,",
"to be a considerable threat. Among these are data",
" one popular method is to remove backdoor triggers",
"poisoning attacks, which exploit vulnerabilities in",
" either during the training or test phase. Training-",
"models by inserting specific triggers into the train-",
"time defense (Jin et al., 2022; Li et al., 2021b)",
"ing data (Chen et al., 2021; Qi et al., 2021b,c,d).",
"discards samples affected by triggers so that the",
"For instance, by inserting certain strings as triggers",
"model would not be trapped by the correlation be-",
"into the training data of a confidential document de-",
"tween triggers and the target label. Test-time de-",
"tection system, an attacker could make the system",
"fenses detect the specific trigger tokens and remove",
"overlookcriticaldocumentsandcauseinformation",
"them from the textual input to avoid activating the",
"leakage by embedding the same strings in the doc-",
"backdoor (Qi et al., 2021a; Li et al., 2021b; Yang",
"ument's content. Recent studies (Kasneci et al.,",
"et al., 2021b). These approaches all assume that",
"2023; Li et al., 2023; Bommasani et al., 2021; Car-",
"(i) backdoor triggers are visible and detectable and",
"lini et al., 2021) further demonstrate that training",
"(i) only one type of trigger is inserted (Liu et al.,",
"'Our code is available at https://github.com/",
"2023). However, backdoor triggers may be implicit",
"VictoriaGraf/Nested_PoE.",
"or invisible (Qi et al., 2021b,c) without having a"
] |
{
"arxiv_id": "2404.02356",
"language": "en",
"page_idx": 0
}
|
||
2404.02356-en-1
|
2404.02356-en-1.jpg
|
fixed surface form (Gu et al., 2017). For example, the stylistic attack (Qi et al., 2021b), which is based on textual style transfer, paraphrases benign input with a pre-defined textual style as a trigger. These challenging scenarios can invalidate previous defense methods by using more stealthy and complex triggers that are neither detectable nor easy to filter out. In addition to a single attack, different types of backdoor triggers might be used by the attacker to simultaneously and independently poison the same dataset (Liu et al., 2023), in which case the defender has no knowledge about the variety and prevalence of backdoor triggers even if one is discovered. In the era of large language models (LLMs) where training is reliant on web corpora and human-provided feedback (Touvron et al., 2023; Zheng et al., 2023; Wan et al., 2023), NLP systems are exposed to an unprecedentedly severe risk that any kind of data pollution can be maliciously hidden in the training corpus. Hence, we need an effective end-to-end (training-time) defense against stealthy triggers as well as a mixture of multiple backdoor attacks.
Backdoors are in essence, as is claimed by Liu et al. (2023), deliberately crafted prediction shortcuts (Jia and Liang, 2017; Gururangan et al., 2018; Poliak et al., 2018; Wang and Culotta, 2020; Gardner et al., 2021) between predefined trigger features and attacker-specified target labels so that a model trained on the poisoned data would predict the target label with high confidence whenever the trigger appears in the input. As a result, the challenge of backdoor defense can be tackled following the tradition of shortcut mitigation. Specifically, the framework of Product of Experts (PoE; Hinton 2002) is adapted by Liu et al. (2023) where a shallow model (dubbed the“trigger-only model”) is used to capture the backdoor shortcut, leaving the backdoor-mitigated residual for the main model. Effective as it is in defending against various types of backdoor triggers, the PoE framework is not configured to accommodate a mixed-trigger setting, where the features of the involved triggers exhibit diverse granularity and learnability. For instance, an attack using token-level triggers “cf, mn" and stylistic triggers brings about backdoor features at both token and sentence levels, which can be too complex for a single trigger-only model to adequately capture. Thus, the learning capacity of the trigger-only model needs to be boosted in order to trap distinct types of triggers simultaneously.
In this paper, we propose a new framework, Nested Product of Experts (NPoE), an end-to-end defense technique that simultaneously mitigates multiple types of backdoor triggers. Based on the framework of PoE, multiple shallow models (i.e. trigger-only models) work in an ensemble (§3.2) to capture distinct backdoor triggers. This ensemble is further used for training a main model that is protected from backdoors in the poisoned training data (83.3). Further, we propose a pseudo development set construction mechanism (§3.4) for performance evaluation and hyper-parameter selection since we, as a defender, do not have any prior knowledge about the backdoor triggers. Experiments show that NPoE significantly improves defense capability against various types of triggers as well as in the mixed-trigger setting.
Our contributions are three-fold. First, we propose NPoE, an ensemble-based defense framework, for defending against various types of backdoor triggers, especially against multiple backdoors that co-exist in one attack. Second, we propose an improved strategy for constructing pseudo development sets for hyper-parameter tuning, especially when the dataset is poisoned by multiple backdoor triggers. Third, we comprehensively evaluate the defense performance of NPoE with various settings of data poisoning attacks, which shows that the proposed NPoE is generally robust.
### 2 $\textbf{Related Work}$
Our work is connected to three research topics. Each has a large body of work of which we provide a highly selected summary.
Backdoor Attack in NLP. Backdoor attacks on NLP systems can generally breakdown into two fundamental categories: data poisoning (Chen et al., 2021; Qi et al., 2021b,c,d) and weight poisoning (Yang et al., 2021a; Li et al., 2021a). Data poisoning artificially generates correlations between backdoor triggers and an attacker-specified target label (Chen et al., 2021; Qi et al., 2021b,c,d; Zhong et al., 2020). The most common poisoning techniques are insertion-based explicit triggers such as rare tokens (Chen et al., 2021) or fxed contextirrelevant sentences (Dai and Chen, 2019). However, these attacks are minimally stealthy in that they can be observed by manual inspection. Stealthier word-based substitution attacks have been developed to address this issue (Qi et al., 2021d) as well as implicit triggers, such as a specific syn-
|
[
[
0.1175537109375,
0.0843505859375,
0.489501953125,
0.08489990234375,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09857177734375
],
[
0.52880859375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.1009521484375,
0.52880859375,
0.1009521484375
],
[
0.1158447265625,
0.1015625,
0.486083984375,
0.1015625,
0.486083984375,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.509765625,
0.1015625,
0.880859375,
0.1015625,
0.880859375,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1484375,
0.1158447265625,
0.1478271484375
],
[
0.509765625,
0.1341552734375,
0.880859375,
0.133056640625,
0.880859375,
0.1478271484375,
0.509765625,
0.1484375
],
[
0.1175537109375,
0.15087890625,
0.490234375,
0.15087890625,
0.490234375,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.5087890625,
0.1484375,
0.880859375,
0.1484375,
0.880859375,
0.162109375,
0.5087890625,
0.162109375
],
[
0.1175537109375,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.5087890625,
0.166259765625,
0.8818359375,
0.1650390625,
0.8818359375,
0.179931640625,
0.5087890625,
0.18115234375
],
[
0.1158447265625,
0.182861328125,
0.489501953125,
0.18115234375,
0.489501953125,
0.19482421875,
0.1158447265625,
0.196533203125
],
[
0.51123046875,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.5087890625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.5087890625,
0.2120361328125
],
[
0.1175537109375,
0.21435546875,
0.489501953125,
0.21435546875,
0.489501953125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.5087890625,
0.213134765625,
0.8818359375,
0.21435546875,
0.8818359375,
0.2286376953125,
0.5087890625,
0.22802734375
],
[
0.1158447265625,
0.23095703125,
0.490234375,
0.23095703125,
0.490234375,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.5087890625,
0.2298583984375,
0.8818359375,
0.23095703125,
0.8818359375,
0.2452392578125,
0.5087890625,
0.24462890625
],
[
0.1158447265625,
0.2464599609375,
0.489501953125,
0.2464599609375,
0.489501953125,
0.260009765625,
0.1158447265625,
0.260009765625
],
[
0.509765625,
0.2476806640625,
0.880859375,
0.2476806640625,
0.880859375,
0.26123046875,
0.509765625,
0.26123046875
],
[
0.1175537109375,
0.262451171875,
0.490234375,
0.262451171875,
0.490234375,
0.275634765625,
0.1175537109375,
0.275634765625
],
[
0.509765625,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.509765625,
0.276611328125
],
[
0.1158447265625,
0.27783203125,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.292236328125
],
[
0.509765625,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1175537109375,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.509765625,
0.295654296875,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.509765625,
0.308837890625
],
[
0.1175537109375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51123046875,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51123046875,
0.32470703125
],
[
0.1175537109375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.1175537109375,
0.34033203125
],
[
0.51123046875,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.51123046875,
0.341552734375
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.509765625,
0.342529296875,
0.69873046875,
0.34326171875,
0.69873046875,
0.358154296875,
0.509765625,
0.35693359375
],
[
0.1175537109375,
0.359375,
0.486083984375,
0.359375,
0.486083984375,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.52880859375,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.52880859375,
0.37353515625
],
[
0.1175537109375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.51123046875,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.51123046875,
0.3896484375
],
[
0.1158447265625,
0.392578125,
0.489501953125,
0.392578125,
0.489501953125,
0.40625,
0.1158447265625,
0.40625
],
[
0.5087890625,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.405029296875,
0.5087890625,
0.405029296875
],
[
0.1175537109375,
0.407958984375,
0.490234375,
0.407958984375,
0.490234375,
0.421630859375,
0.1175537109375,
0.421630859375
],
[
0.5087890625,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.5087890625,
0.421630859375
],
[
0.1175537109375,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.51318359375,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.51318359375,
0.438232421875
],
[
0.1158447265625,
0.440673828125,
0.334228515625,
0.440673828125,
0.334228515625,
0.45361328125,
0.1158447265625,
0.45361328125
],
[
0.51123046875,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.51123046875,
0.454833984375
],
[
0.509765625,
0.457275390625,
0.880859375,
0.457275390625,
0.880859375,
0.47021484375,
0.509765625,
0.47021484375
],
[
0.1351318359375,
0.4697265625,
0.489501953125,
0.4697265625,
0.489501953125,
0.482666015625,
0.1351318359375,
0.482666015625
],
[
0.509765625,
0.47265625,
0.88330078125,
0.47265625,
0.88330078125,
0.486328125,
0.509765625,
0.486328125
],
[
0.1158447265625,
0.48388671875,
0.490234375,
0.485107421875,
0.490234375,
0.49951171875,
0.1158447265625,
0.498779296875
],
[
0.5087890625,
0.4892578125,
0.880859375,
0.4892578125,
0.880859375,
0.5029296875,
0.5087890625,
0.5029296875
],
[
0.1158447265625,
0.501953125,
0.48779296875,
0.501953125,
0.48779296875,
0.515625,
0.1158447265625,
0.515625
],
[
0.509765625,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.509765625,
0.5185546875
],
[
0.1175537109375,
0.5185546875,
0.490234375,
0.5185546875,
0.490234375,
0.5322265625,
0.1175537109375,
0.5322265625
],
[
0.5087890625,
0.51953125,
0.88330078125,
0.52099609375,
0.88330078125,
0.53515625,
0.5087890625,
0.53369140625
],
[
0.1158447265625,
0.533203125,
0.489501953125,
0.53369140625,
0.489501953125,
0.548828125,
0.1158447265625,
0.54736328125
],
[
0.509765625,
0.53759765625,
0.75146484375,
0.53759765625,
0.75146484375,
0.55029296875,
0.509765625,
0.55029296875
],
[
0.1175537109375,
0.55029296875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.1175537109375,
0.56640625,
0.489501953125,
0.56640625,
0.489501953125,
0.580078125,
0.1175537109375,
0.580078125
],
[
0.5087890625,
0.56787109375,
0.53662109375,
0.56396484375,
0.5400390625,
0.57470703125,
0.51123046875,
0.57763671875
],
[
0.53564453125,
0.5693359375,
0.54150390625,
0.5693359375,
0.54150390625,
0.57470703125,
0.53564453125,
0.57470703125
],
[
0.5380859375,
0.5654296875,
0.6650390625,
0.5654296875,
0.6650390625,
0.580078125,
0.5380859375,
0.580078125
],
[
0.1175537109375,
0.5830078125,
0.48779296875,
0.5830078125,
0.48779296875,
0.5966796875,
0.1175537109375,
0.5966796875
],
[
0.51318359375,
0.5927734375,
0.8818359375,
0.5927734375,
0.8818359375,
0.60546875,
0.51318359375,
0.60546875
],
[
0.1158447265625,
0.5986328125,
0.490234375,
0.59716796875,
0.490234375,
0.6123046875,
0.1158447265625,
0.61328125
],
[
0.1158447265625,
0.61328125,
0.48779296875,
0.6142578125,
0.48779296875,
0.62890625,
0.1158447265625,
0.62744140625
],
[
0.51318359375,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.51318359375,
0.62255859375
],
[
0.1175537109375,
0.63134765625,
0.489501953125,
0.63134765625,
0.489501953125,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.509765625,
0.6259765625,
0.7138671875,
0.6259765625,
0.7138671875,
0.6396484375,
0.509765625,
0.6396484375
],
[
0.1175537109375,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.51123046875,
0.6455078125,
0.8818359375,
0.6455078125,
0.8818359375,
0.66015625,
0.51123046875,
0.66015625
],
[
0.1175537109375,
0.6630859375,
0.49365234375,
0.6630859375,
0.49365234375,
0.6767578125,
0.1175537109375,
0.6767578125
],
[
0.51123046875,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.51123046875,
0.677734375
],
[
0.1158447265625,
0.677734375,
0.490234375,
0.6787109375,
0.490234375,
0.693359375,
0.1158447265625,
0.6923828125
],
[
0.509765625,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.509765625,
0.693359375
],
[
0.1158447265625,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.1158447265625,
0.7080078125
],
[
0.51123046875,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.708984375,
0.51123046875,
0.708984375
],
[
0.1175537109375,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.5087890625,
0.71044921875,
0.88330078125,
0.71142578125,
0.88330078125,
0.7255859375,
0.5087890625,
0.724609375
],
[
0.1175537109375,
0.72802734375,
0.490234375,
0.72802734375,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.5087890625,
0.72705078125,
0.880859375,
0.72705078125,
0.880859375,
0.74072265625,
0.5087890625,
0.74072265625
],
[
0.1175537109375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1175537109375,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.75830078125,
0.880859375,
0.7587890625,
0.880859375,
0.77392578125,
0.51123046875,
0.7724609375
],
[
0.1158447265625,
0.77490234375,
0.484375,
0.77490234375,
0.484375,
0.7880859375,
0.1158447265625,
0.7880859375
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.1175537109375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.509765625,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.509765625,
0.8056640625
],
[
0.1158447265625,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.509765625,
0.82080078125
],
[
0.1158447265625,
0.82470703125,
0.49365234375,
0.82470703125,
0.49365234375,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.509765625,
0.82373046875,
0.88330078125,
0.82373046875,
0.88330078125,
0.83740234375,
0.509765625,
0.83740234375
],
[
0.1175537109375,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1158447265625,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.51123046875,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.8681640625,
0.51123046875,
0.8681640625
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.5087890625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.5087890625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.465087890625,
0.90478515625,
0.465087890625,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"fixed surface form (Gu et al., 2017). For example,",
"In this paper, we propose a new framework.",
"the stylistic attack (Qi et al., 2021b), which is based",
"Nested Product of Experts (NPoE), an end-to-end",
"on textual style transfer, paraphrases benign input",
"defense technique that simultaneously mitigates",
"with a pre-defined textual style as a trigger. These",
"multiple types of backdoor triggers. Based on the",
"challenging scenarios can invalidate previous de-",
"framework of PoE. multiple shallow models (i.e",
"fense methods by using more stealthy and complex",
"trigger-only models) work in an ensemble (s3.2) to",
"triggers that are neither detectable nor easy to fil-",
"capture distinct backdoor triggers. This ensemble",
"ter out. In addition to a single attack, different",
"is further used for training a main model that is pro-",
"types of backdoor triggers might be used by the",
"tected from backdoors in the poisoned training data",
"attacker to simultaneously and independently poi-",
"(s3.3). Further, we propose a pseudo development",
"son the same dataset (Liu et al., 2023), in which",
"set construction mechanism (s3.4) for performance",
"case the defender has no knowledge about the va-",
" evaluation and hyper-parameter selection since we,",
"riety and prevalence of backdoor triggers even if",
"as a defender, do not have any prior knowledge",
"one is discovered.In the era of large language",
"about the backdoor triggers. Experiments show",
"models (LLMs) where training is reliant on web",
"that NPoE significantly improves defense capabil-",
"corpora and human-provided feedback (Touvron",
"ity against various types of triggers as well as in",
"et al., 2023; Zheng et al., 2023; Wan et al., 2023),",
"the mixed-trigger setting.",
"NLP systems are exposed to an unprecedentedly",
"Our contributions are three-fold. First, we pro-",
"severe risk that any kind of data pollution can be",
"pose NPoE, an ensemble-based defense framework",
"maliciously hidden in the training corpus. Hence,",
"for defending against various types of backdoor",
"we need an effective end-to-end (training-time) de-",
"triggers, especially against multiple backdoors that",
"fense against stealthy triggers as well as a mixture",
"co-exist in one attack. Second, we propose an",
"of multiple backdoor attacks.",
"improved strategy for constructing pseudo devel-",
"opment sets for hyper-parameter tuning, especially",
"Backdoors are in essence, as is claimed by Liu",
"when the dataset is poisoned by multiple backdoor",
"et al. (2023), deliberately crafted prediction short-",
"triggers. Third, we comprehensively evaluate the",
"cuts (Jia and Liang, 2017; Gururangan et al., 2018;",
"defense performance of NPoE with various settings",
"Poliak et al., 2018; Wang and Culotta, 2020; Gard-",
" of data poisoning attacks, which shows that the pro-",
"ner et al., 2021) between predefined trigger fea-",
"posed NPoE is generally robust.",
"tures and attacker-specified target labels so that a",
"model trained on the poisoned data would predict",
"2",
"",
"Related Work",
"the target label with high confidence whenever the",
"Our work is connected to three research topics.",
"trigger appears in the input. As a result, the chal-",
"lenge of backdoordefense can be tackled following",
"Each has a large body of work of which we provide",
"the tradition of shortcut mitigation. Specifically,",
" a highly selected summary.",
"the framework of Product of Experts (PoE; Hinton",
"Backdoor Attack in NLP. Backdoor attacks on",
"2002) is adapted by Liu et al. (2023) where a shal-",
"NLP systems can generally breakdown into two",
"low model (dubbed the \"trigger-only model\") is",
"fundamental categories: data poisoning (Chen",
"used to capture the backdoor shortcut, leaving the",
"et al., 2021; Qi et al., 2021b,c,d) and weight poison-",
"backdoor-mitigated residual for the main model.",
"ing (Yang et al., 202la; Li et al., 2021a). Data poi-",
"Effective as it is in defending against various types",
" soning artificially generates correlations between",
"of backdoor triggers, the PoE framework is not",
"backdoor triggers and an attacker-specified target",
"configured to accommodate a mixed-trigger set-",
"label (Chen et al., 2021; Qi et al., 2021b,c,d; Zhong",
"ting, where the features of the involved triggers",
"et al., 2020). The most common poisoning tech",
"exhibit diverse granularity and learnability. For",
" niques are insertion-based explicit triggers such as",
"instance, an attack using token-level triggers “cf,",
"rare tokens (Chen et al., 2021) or fixed context",
"mn\" and stylistic triggers brings about backdoor fea-",
"irrelevant sentences (Dai and Chen, 2019). How-",
"tures at both token and sentence levels, which can",
"ever, these attacks are minimally stealthy in that",
"be too complex for a single trigger-only model to",
"theycanbeobservedbymanualinspection.Stealth-",
"adequately capture. Thus, the learning capacity of",
"ier word-based substitution attacks have been de-",
"the trigger-only model needs to be boosted in order",
"veloped to address this issue (Qi et al., 2021d) as",
"to trap distinct types of triggers simultaneously",
"well as implicit triggers, such as a specific syn-"
] |
{
"arxiv_id": "2404.02356",
"language": "en",
"page_idx": 1
}
|
||
2404.02356-en-3
|
2404.02356-en-3.jpg
|
neously assigned the target label $y^*$ that forms a poisoned subset $\mathcal{S}^*=\{(x_{i}^{*},y^{*})\}_{i=1}^{|\mathcal{S}|}.$ Poisoned samples are chosen independently for each trigger $t.$ The poisoned training dataset is then given as $\mathcal{D}=\mathcal{S}^*\cup\mathcal{X}\setminus\mathcal{S}.$ The poison rate is defined as $\begin{array}{l}\frac{|S|}{|X|},\text{the fraction of examples modified. A higher}\\\text{poison rate presents more prevalent examples of}\end{array}$ the trigger/target-label correlation, which means a stronger attack while it is also less stealthy and more likely to be detected by human inspection.
In this paper, four of the most popular trigger types are involved in testing the performance of our proposed framework: rare tokens (Kurita et al., 2020), fixed sentence or phrase (Dai and Chen, 2019), syntactic triggers (Qi et al., 2021c), and stylistic triggers (Qi et al., 2021b). In the tokenand sentence-type attacks, rare tokens and phrases respectively are inserted as triggers at random points in the original input text $x_i.$ Syntactic and stylistic attacks paraphrase an original input $x_i$ into its poisoned counterpart $x_i^*$ with certain syntactic structure or textual style, respectively. Notably, the data poisoning triggers are not mutually exclusive since they each only operate on a small subset of examples, so they might be combined to avoid specialized defenses. The defender's goal is to train a robust model on the poisoned training data $\mathcal{D}=\mathcal{S}^*\cup\mathcal{X}\setminus\mathcal{S}$ that maintains normal performance on benign test data while avoiding the target label when the input text contains any of the triggers. Overview. To defend against multiple types of
co-existing backdoor triggers, we propose the Nested PoE (NPoE) framework (Fig. 2) to train a backdoor-resistant model. Inspired by Liu et al. (2023) which considers backdoors as shortcuts between triggers and the target label, we also follow the PoE technique to defend backdoors as mitigating shortcuts in the training data. We differ from previous PoE methods (Liu et al., 2023) by attempting to trap multiple backdoor triggers simultaneously through several shallow models (dubbed as trigger-only models) to capture these toxic shortcuts. A trigger-only prediction is then obtained following the Mixture of Experts (MoE; Shazeer et al. 2017; Ma et al. 2018) manner that combines the predictions of all trigger-only models (§3.2). The main model is trained in an ensemble with a mixture of trigger-only models that overfit all the present backdoor shortcuts, leaving the main model with a trigger-free residual (§3.3). Since a backdoor defender lacks a validation set with annotated
<FigureHere>
Figure 3: Trigger-only models are pre-trained separately on the trigger identification task for each type of trigger (BadNet, InsertSent, Syntactic, and Stylistic).
triggers, we make use of both the main model and the trigger-only MoE to filter out a pseudo development set from the training data for hyper-parameter selection (§3.4).
# 3.2 MoE for Trigger-Only Models
For Nested PoE, the framework of Mixture of Experts (MoE) is used to combine several triggeronly shallow models to be nested within the conventional PoE framework (Clark et al.,2019; Karimi Mahabadi et al.,2020a). In particular, a mixture of $k$ trigger-only models $b^{\bar{1}},...,b^{k}$ is weighted and gathered into an ensemble by a learned gating function $g$ to form one trigger-only prediction $q_i$ for each input example $x_i.$ Predictions $b_i^1,...,b_i^k$ are reweighted by the softmax of $\begin{array}{l}\text{the gate output }(\sum_{j=1}^{k}g_{i}^{j}=1)\text{to produce the final}\\\text{prediction }q_{i}\text{ of the trigger-only MoE:}\end{array}$
$$q_i=g_i^1\log(b_i^1)+g_i^2\log(b_i^2)+...+g_i^k\log(b_i^k).$$
(1)
For each input example, this gating operation assigns different importance to each trigger-only model, depending on the relevance between the shortcut feature that is captured by each triggeronly model and features that appear in the input.
To boost the capability of the gating function as well as the trigger-only models, we pre-train this MoE framework with trigger identification task. Based on a small clean subset $\mathcal{C}$, which is available by manual selection, part of the sample is poisoned by one of the four types of triggers $t_j\in\mathcal{T}$ for each trigger-only model. Labels are assigned $y_i^*=0$ for data left clean and $y_i^*=1$ for poisoned $x_i^*.$ Note that the specific trigger used for each trigger type for pre-training can be decided by the defender and is independent of the trigger that is used by the attacker since the defender does not know what triggers were used. Each trigger-only model $b_i$
|
[
[
0.1175537109375,
0.08489990234375,
0.48779296875,
0.08489990234375,
0.48779296875,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5146484375,
0.08489990234375,
0.53564453125,
0.0843505859375,
0.53564453125,
0.0885009765625,
0.5146484375,
0.08905029296875
],
[
0.64404296875,
0.08612060546875,
0.66650390625,
0.08612060546875,
0.66650390625,
0.08905029296875,
0.64404296875,
0.08905029296875
],
[
0.779296875,
0.08489990234375,
0.80615234375,
0.08489990234375,
0.80615234375,
0.0885009765625,
0.779296875,
0.0885009765625
],
[
0.11083984375,
0.097412109375,
0.491943359375,
0.097412109375,
0.491943359375,
0.1217041015625,
0.11083984375,
0.1217041015625
],
[
0.7958984375,
0.09320068359375,
0.81201171875,
0.0926513671875,
0.81201171875,
0.0968017578125,
0.7958984375,
0.097412109375
],
[
0.1158447265625,
0.1181640625,
0.48779296875,
0.119384765625,
0.48779296875,
0.1341552734375,
0.1158447265625,
0.133056640625
],
[
0.53564453125,
0.1181640625,
0.60107421875,
0.1181640625,
0.60107421875,
0.1300048828125,
0.53564453125,
0.1300048828125
],
[
0.66162109375,
0.1181640625,
0.72900390625,
0.1181640625,
0.72900390625,
0.1300048828125,
0.66162109375,
0.1300048828125
],
[
0.79248046875,
0.1181640625,
0.85791015625,
0.1181640625,
0.85791015625,
0.1300048828125,
0.79248046875,
0.1300048828125
],
[
0.54736328125,
0.1300048828125,
0.587890625,
0.1300048828125,
0.587890625,
0.13720703125,
0.54736328125,
0.13720703125
],
[
0.67529296875,
0.1300048828125,
0.71533203125,
0.1300048828125,
0.71533203125,
0.13720703125,
0.67529296875,
0.13720703125
],
[
0.8017578125,
0.12890625,
0.8505859375,
0.12890625,
0.8505859375,
0.1395263671875,
0.8017578125,
0.1395263671875
],
[
0.1175537109375,
0.135986328125,
0.48779296875,
0.135986328125,
0.48779296875,
0.149658203125,
0.1175537109375,
0.149658203125
],
[
0.1175537109375,
0.15087890625,
0.48779296875,
0.1519775390625,
0.48779296875,
0.166259765625,
0.1175537109375,
0.1650390625
],
[
0.5224609375,
0.1602783203125,
0.60107421875,
0.1602783203125,
0.60107421875,
0.1787109375,
0.5224609375,
0.1787109375
],
[
0.6484375,
0.160888671875,
0.73486328125,
0.160888671875,
0.73486328125,
0.171630859375,
0.6484375,
0.171630859375
],
[
0.7958984375,
0.15673828125,
0.85546875,
0.15673828125,
0.85546875,
0.18115234375,
0.7958984375,
0.18115234375
],
[
0.12005615234375,
0.16748046875,
0.489501953125,
0.16748046875,
0.489501953125,
0.188232421875,
0.12005615234375,
0.188232421875
],
[
0.6630859375,
0.171630859375,
0.7021484375,
0.171630859375,
0.7021484375,
0.17578125,
0.6630859375,
0.17578125
],
[
0.1158447265625,
0.1859130859375,
0.486083984375,
0.185302734375,
0.486083984375,
0.1995849609375,
0.1158447265625,
0.20068359375
],
[
0.1175537109375,
0.201904296875,
0.484375,
0.201904296875,
0.484375,
0.215576171875,
0.1175537109375,
0.215576171875
],
[
0.6455078125,
0.201904296875,
0.71533203125,
0.201904296875,
0.71533203125,
0.208984375,
0.6455078125,
0.208984375
],
[
0.66015625,
0.207275390625,
0.73046875,
0.207275390625,
0.73046875,
0.21728515625,
0.66015625,
0.21728515625
],
[
0.72119140625,
0.203125,
0.744140625,
0.203125,
0.744140625,
0.207275390625,
0.72119140625,
0.207275390625
],
[
0.11419677734375,
0.218505859375,
0.484375,
0.218505859375,
0.484375,
0.232177734375,
0.11419677734375,
0.232177734375
],
[
0.1158447265625,
0.235107421875,
0.47021484375,
0.235107421875,
0.47021484375,
0.248779296875,
0.1158447265625,
0.248779296875
],
[
0.509765625,
0.2392578125,
0.87890625,
0.2392578125,
0.87890625,
0.2529296875,
0.509765625,
0.2529296875
],
[
0.1351318359375,
0.25048828125,
0.486083984375,
0.25048828125,
0.486083984375,
0.26416015625,
0.1351318359375,
0.26416015625
],
[
0.509765625,
0.2529296875,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.509765625,
0.267333984375
],
[
0.1158447265625,
0.267333984375,
0.48779296875,
0.267333984375,
0.48779296875,
0.28076171875,
0.1158447265625,
0.28076171875
],
[
0.51123046875,
0.268310546875,
0.8193359375,
0.268310546875,
0.8193359375,
0.281982421875,
0.51123046875,
0.281982421875
],
[
0.1175537109375,
0.283203125,
0.48779296875,
0.283203125,
0.48779296875,
0.29638671875,
0.1175537109375,
0.29638671875
],
[
0.118408203125,
0.298583984375,
0.489501953125,
0.298583984375,
0.489501953125,
0.3134765625,
0.118408203125,
0.3134765625
],
[
0.509765625,
0.295654296875,
0.8818359375,
0.29443359375,
0.8818359375,
0.308837890625,
0.509765625,
0.31005859375
],
[
0.118408203125,
0.3154296875,
0.489501953125,
0.3154296875,
0.489501953125,
0.328857421875,
0.118408203125,
0.328857421875
],
[
0.509765625,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.32666015625,
0.509765625,
0.32666015625
],
[
0.1158447265625,
0.330810546875,
0.491943359375,
0.330078125,
0.491943359375,
0.344482421875,
0.1158447265625,
0.345703125
],
[
0.509765625,
0.328857421875,
0.88330078125,
0.328857421875,
0.88330078125,
0.34326171875,
0.509765625,
0.34326171875
],
[
0.1158447265625,
0.347412109375,
0.489501953125,
0.347412109375,
0.489501953125,
0.3623046875,
0.1158447265625,
0.3623046875
],
[
0.5087890625,
0.34326171875,
0.6337890625,
0.344482421875,
0.6337890625,
0.359375,
0.5087890625,
0.358154296875
],
[
0.1158447265625,
0.364013671875,
0.48779296875,
0.364013671875,
0.48779296875,
0.377685546875,
0.1158447265625,
0.377685546875
],
[
0.51123046875,
0.371826171875,
0.791015625,
0.371826171875,
0.791015625,
0.385498046875,
0.51123046875,
0.385498046875
],
[
0.1158447265625,
0.380126953125,
0.48779296875,
0.380126953125,
0.48779296875,
0.393798828125,
0.1158447265625,
0.393798828125
],
[
0.1158447265625,
0.3955078125,
0.489501953125,
0.394287109375,
0.489501953125,
0.4091796875,
0.1158447265625,
0.410400390625
],
[
0.51123046875,
0.392578125,
0.8818359375,
0.392578125,
0.8818359375,
0.40625,
0.51123046875,
0.40625
],
[
0.1175537109375,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.42578125,
0.1175537109375,
0.42578125
],
[
0.51123046875,
0.4091796875,
0.8818359375,
0.4091796875,
0.8818359375,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.1175537109375,
0.427490234375,
0.489501953125,
0.427490234375,
0.489501953125,
0.441162109375,
0.1175537109375,
0.441162109375
],
[
0.509765625,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.509765625,
0.438232421875
],
[
0.1175537109375,
0.44482421875,
0.490234375,
0.44482421875,
0.490234375,
0.457763671875,
0.1175537109375,
0.457763671875
],
[
0.509765625,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.509765625,
0.454833984375
],
[
0.1175537109375,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.1175537109375,
0.473876953125
],
[
0.51318359375,
0.457275390625,
0.88330078125,
0.457275390625,
0.88330078125,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.1175537109375,
0.476806640625,
0.48779296875,
0.476806640625,
0.48779296875,
0.490478515625,
0.1175537109375,
0.490478515625
],
[
0.5087890625,
0.47265625,
0.8818359375,
0.471435546875,
0.8818359375,
0.486328125,
0.5087890625,
0.48681640625
],
[
0.1175537109375,
0.4921875,
0.489501953125,
0.4921875,
0.489501953125,
0.505859375,
0.1175537109375,
0.505859375
],
[
0.51123046875,
0.4892578125,
0.88330078125,
0.4892578125,
0.88330078125,
0.5029296875,
0.51123046875,
0.5029296875
],
[
0.1158447265625,
0.50830078125,
0.486083984375,
0.50830078125,
0.486083984375,
0.521484375,
0.1158447265625,
0.521484375
],
[
0.51123046875,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.51123046875,
0.51953125
],
[
0.1175537109375,
0.5224609375,
0.48779296875,
0.52392578125,
0.48779296875,
0.5380859375,
0.1175537109375,
0.53759765625
],
[
0.509765625,
0.521484375,
0.880859375,
0.521484375,
0.880859375,
0.53515625,
0.509765625,
0.53515625
],
[
0.1175537109375,
0.54052734375,
0.48779296875,
0.54052734375,
0.48779296875,
0.55419921875,
0.1175537109375,
0.55419921875
],
[
0.5087890625,
0.53759765625,
0.88330078125,
0.5361328125,
0.88330078125,
0.55029296875,
0.5087890625,
0.5517578125
],
[
0.5087890625,
0.5498046875,
0.88671875,
0.5498046875,
0.88671875,
0.57373046875,
0.5087890625,
0.57373046875
],
[
0.1175537109375,
0.55712890625,
0.465087890625,
0.55712890625,
0.465087890625,
0.57080078125,
0.1175537109375,
0.57080078125
],
[
0.51318359375,
0.57177734375,
0.79443359375,
0.57177734375,
0.79443359375,
0.5859375,
0.51318359375,
0.5859375
],
[
0.1175537109375,
0.58203125,
0.486083984375,
0.58203125,
0.486083984375,
0.5966796875,
0.1175537109375,
0.5966796875
],
[
0.1158447265625,
0.5986328125,
0.48779296875,
0.59716796875,
0.48779296875,
0.6123046875,
0.1158447265625,
0.61328125
],
[
0.5224609375,
0.60107421875,
0.880859375,
0.599609375,
0.880859375,
0.6142578125,
0.5224609375,
0.615234375
],
[
0.1158447265625,
0.6142578125,
0.486083984375,
0.6142578125,
0.486083984375,
0.62890625,
0.1158447265625,
0.62890625
],
[
0.1158447265625,
0.6298828125,
0.489501953125,
0.6298828125,
0.489501953125,
0.6435546875,
0.1158447265625,
0.6435546875
],
[
0.51123046875,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.51123046875,
0.64453125
],
[
0.1175537109375,
0.646484375,
0.49365234375,
0.646484375,
0.49365234375,
0.66015625,
0.1175537109375,
0.66015625
],
[
0.509765625,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.509765625,
0.6611328125
],
[
0.1158447265625,
0.6630859375,
0.489501953125,
0.6630859375,
0.489501953125,
0.6767578125,
0.1158447265625,
0.6767578125
],
[
0.509765625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.118408203125,
0.68115234375,
0.49365234375,
0.68115234375,
0.49365234375,
0.69140625,
0.118408203125,
0.69140625
],
[
0.5087890625,
0.677734375,
0.88330078125,
0.6787109375,
0.88330078125,
0.693359375,
0.5087890625,
0.6923828125
],
[
0.1158447265625,
0.6953125,
0.490234375,
0.6953125,
0.490234375,
0.708984375,
0.1158447265625,
0.708984375
],
[
0.509765625,
0.693359375,
0.87158203125,
0.6953125,
0.87158203125,
0.708984375,
0.509765625,
0.70703125
],
[
0.1175537109375,
0.71142578125,
0.490234375,
0.71142578125,
0.490234375,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.53076171875,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.53076171875,
0.7255859375
],
[
0.1175537109375,
0.72802734375,
0.490234375,
0.72802734375,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.72802734375,
0.880859375,
0.72802734375,
0.880859375,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.1158447265625,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1158447265625,
0.755859375
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1175537109375,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.5146484375,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.771484375,
0.5146484375,
0.771484375
],
[
0.1158447265625,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.1158447265625,
0.7880859375
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.1175537109375,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.1175537109375,
0.82080078125
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51318359375,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.490234375,
0.84033203125,
0.490234375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.8388671875,
0.8818359375,
0.84033203125,
0.8818359375,
0.85498046875,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.51123046875,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.8681640625,
0.51123046875,
0.8681640625
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.50537109375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.50537109375,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.491943359375,
0.88720703125,
0.491943359375,
0.90185546875,
0.1158447265625,
0.90283203125
],
[
0.5087890625,
0.88818359375,
0.8818359375,
0.88720703125,
0.8818359375,
0.90185546875,
0.5087890625,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.8818359375,
0.90380859375,
0.8818359375,
0.91845703125,
0.5087890625,
0.91943359375
]
] |
[
"neously assigned the target label y* that forms a",
"Non:",
"NOS",
"on-to",
"poisoned subset S* = (ai,y*)l. Poisoned",
"Tov",
"samples are chosen independently for each trigger",
"Trigger-Only",
"Trigger-Only",
"Trigger-Only",
"Model1",
"Model2",
"Model3",
"t. The poisoned training dataset is then given as",
"D = S* U X S. The poison rate is defined as",
"can't beleve",
"I can't believeB",
"",
"，the fraction of examples modified.Ahigher",
"opleper",
"poison rate presents more prevalent examples of",
"the trigger/target-label correlation, which means",
"I can't belie",
"terrible city",
"e:th.",
"a stronger attack while it is also less stealthy and",
"more likely to be detected by human inspection.",
"Figure 3: Trigger-only models are pre-trained separately",
"In this paper, four of the most popular trigger",
"on the trigger identification task for each type of trigger",
"types are involved in testing the performance of",
"(BadNet, InsertSent, Syntactic, and Stylistic).",
"our proposed framework: rare tokens (Kurita et al.,",
"2020), fixed sentence or phrase (Dai and Chen,",
"triggers, we make use of both the main model and",
"2019), syntactic triggers (Qi et al., 2021c), and",
"the trigger-only MoE to filter out a pseudo develop-",
"stylistic triggers (Qi et al., 2021b). In the token-",
"ment set from the training data for hyper-parameter",
"and sentence-type attacks, rare tokens and phrases",
"selection ($3.4).",
"respectively are inserted as triggers at random",
"3.2 MoE for Trigger-Only Models",
"points in the original input text ci: Syntactic and",
"stylistic attacks paraphrase an original input c, into",
"For Nested PoE, the framework of Mixture of Ex-",
"its poisoned counterpart c with certain syntactic",
"perts (MoE) is used to combine several trigger",
"structure or textual style, respectively. Notably,",
" only shallow models to be nested within the con-",
"the data poisoning triggers are not mutually exclu-",
"ventional PoE framework (Clark et al., 2019",
"sive since they each only operate on a small subset",
"Karimi Mahabadi et al., 2020a). In particular,",
"of examples, so they might be combined to avoid",
"a mixture of k trigger-only models bl,.., bk is",
"specialized defenses. The defender's goal is to",
"weighted and gathered into an ensemble by a",
"train a robust model on the poisoned training data",
"learned gating function g to form one trigger-only",
"D = S*U.X \\S that maintains normal performance",
"prediction qi for each input example C,: Predic",
"on benign test data while avoiding the target label",
"tions bi,.., bh are reweighted by the softmax of",
"the gate output (f-1 g - 1) to produce the final",
"when the input text contains any of the triggers.",
"prediction qi of the trigger-only MoE:",
"Overview. To defend against multiple types of",
"co-existing backdoor triggers, we propose the",
"qi =gi log(bl)+g2 log(b)+...+g log(bh). ()",
"Nested PoE (NPoE) framework (Fig. 2) to train",
"a backdoor-resistant model. Inspired by Liu et al.",
"For each input example, this gating operation as-",
"(2023) which considers backdoors as shortcuts be-",
"signs different importance to each trigger-only",
"tween triggers and the target label, we also follow",
"model, depending on the relevance between the",
"thePoE technigueto defendbackdoorsasmitigat-",
"shortcut feature that is captured by each trigger-",
"ing shortcuts in the training data. We differ from",
"only model and features that appear in the input.",
"previous PoE methods (Liu et al., 2023) by attempt-",
"To boost the capability of the gating function as",
"ing to trap multiple backdoor triggers simultane-",
"well as the trigger-only models, we pre-train this",
"ously through several shallow models (dubbed as",
"MoE framework with trigger identification task.",
"trigger-only models) to capture these toxic short-",
"BasedonasmallcleansubsetC,whichisavailable",
"cuts. A trigger-only prediction is then obtained",
"by manual selection, part of the sample is poisoned",
"following the Mixture of Experts (MoE; Shazeer",
"by one of the four types of triggers t, E T for each",
"et al. 2017; Ma et al. 2018) manner that combines",
"trigger-only model. Labels are assigned y, = O for",
"the predictions of all trigger-only models (s3.2).",
"data left clean and y, = 1 for poisoned c,. Note",
"The main model is trained in an ensemble with a",
"that the specific trigger used for each trigger type",
"mixtureof trigger-onlymodelsthat overfitallthe",
"forpre-training canbedecidedbythedefenderand",
"present backdoor shortcuts, leaving the main model",
" is independent of the trigger that is used by the",
"with a trigger-free residual (s3.3). Since a back-",
"attacker since the defender does not know what",
"door defender lacks a validation set with annotated",
"triggers were used. Each trigger-only model bi"
] |
{
"arxiv_id": "2404.02356",
"language": "en",
"page_idx": 3
}
|
||
2404.08513-en-0
|
2404.08513-en-0.jpg
|
Published as a conference paper at ICLR 2024
ADVERSARIAL IMITATION LEARNING VIA BOOSTING
Jonathan D. Chang
Department of Computer Science
Cornell University
jdc396@cornell.edu
Dhruv Sreenivas *
Department of Computer Science
${\overset{\mathrm{Cornell}}{\operatorname*{University}}}$
ds8446cornell.edu
Yingbing Huang *
Department of Electrical and Computer Engineering
$\hat{\text{University of Illinois Urbana-Champaign}}$
yh21@illinois.edu
Kianté Brantley
Department of Computer Science
$\bar{\mathrm{Cornell~University}}$
kdb82@cornell.edu
Wen Sun
Department of Computer Science
$\hat{\text{Cornell University}}$
ws4558cornell.edu
### ABSTRACT
Adversarial imitation learning (AIL) has stood out as a dominant framework across various imitation learning (IL) applications, with Discriminator Actor Critic (DAC) (Kostrikov et al., 2019) demonstrating the effectiveness of off-policy learning algorithms in improving sample efficiency and scalability to higher-dimensional observations. Despite DAC's empirical success, the original AIL objective is on-policy and DAC's ad-hoc application of off-policy training does not guarantee successful imitation (Kostrikov et al., 2019; 2020). Follow-up work such as ValueDICE (Kostrikov et al., 2020) tackles this issue by deriving a fully off-policy AIL objective. Instead in this work, we develop a novel and principled AIL algorithm via the framework of $boosting.$ Like boosting, our new algorithm, AILBoost, maintains an ensemble of properly weighted weak learners (i.e., policies) and trains a discriminator that witnesses the maximum discrepancy between the distributions of the ensemble and the expert policy. We maintain a weighted replay buffer to represent the state-action distribution induced by the ensemble, allowing us to train discriminators using the entire data collected so far. In the weighted replay buffer, the contribution of the data from older policies are properly discounted with the weight computed based on the boosting framework. Empirically, we evaluate our algorithm on both controller state-based and pixel-based environments from the DeepMind Control Suite. AILBoost outperforms DAC on both types of environments, demonstrating the benefit of properly weighting replay buffer data for off-policy training. On state-based environments, AILBoost outperforms ValueDICE and IQ-Learn(Garg et al., 2021), achieving competitive performance with as little as one expert trajectory.
1 INTRODUCTION
Imitation learning (IL) is a promising paradigm for learning general policies without rewards from demonstration data, achieving remarkable success in autonomous driving (Bronstein et al., 2022; Pomerleau, 1988), video games (Baker et al., 2022; Shah et al., 2022) and graphics (Peng et al., 2021). Adversarial Imitation Learning (AIL) is an incredibly successful approach for imitation learning (Ho&Ermon, 2016; Fu et al., 2018; Kostrikov et al., 2019; Ke et al., 2020). These methods cast IL as a distribution matching problem whereby the learning agent minimizes the divergence between the expert demonstrator's distribution and the state-action distribution induced by the agent. First
$^{*}$Equal Contribution
1
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.176513671875,
0.10101318359375,
0.810546875,
0.10162353515625,
0.810546875,
0.11932373046875,
0.176513671875,
0.11871337890625
],
[
0.181396484375,
0.1427001953125,
0.324462890625,
0.14453125,
0.324462890625,
0.1622314453125,
0.181396484375,
0.160400390625
],
[
0.59326171875,
0.1458740234375,
0.72216796875,
0.1458740234375,
0.72216796875,
0.160400390625,
0.59326171875,
0.160400390625
],
[
0.1846923828125,
0.1610107421875,
0.406005859375,
0.1610107421875,
0.406005859375,
0.1749267578125,
0.1846923828125,
0.1749267578125
],
[
0.59228515625,
0.160400390625,
0.814453125,
0.1590576171875,
0.814453125,
0.173583984375,
0.59228515625,
0.1749267578125
],
[
0.181396484375,
0.173583984375,
0.306396484375,
0.1749267578125,
0.306396484375,
0.189453125,
0.181396484375,
0.1881103515625
],
[
0.59326171875,
0.173583984375,
0.71728515625,
0.173583984375,
0.71728515625,
0.1875,
0.59326171875,
0.1875
],
[
0.1846923828125,
0.1881103515625,
0.363525390625,
0.1881103515625,
0.363525390625,
0.20263671875,
0.1846923828125,
0.20263671875
],
[
0.59326171875,
0.189453125,
0.76123046875,
0.189453125,
0.76123046875,
0.20263671875,
0.59326171875,
0.20263671875
],
[
0.1846923828125,
0.222900390625,
0.313720703125,
0.222900390625,
0.313720703125,
0.2374267578125,
0.1846923828125,
0.2374267578125
],
[
0.59326171875,
0.2235107421875,
0.70751953125,
0.2235107421875,
0.70751953125,
0.2374267578125,
0.59326171875,
0.2374267578125
],
[
0.1846923828125,
0.2374267578125,
0.52783203125,
0.2374267578125,
0.52783203125,
0.251953125,
0.1846923828125,
0.251953125
],
[
0.59326171875,
0.2374267578125,
0.814453125,
0.2374267578125,
0.814453125,
0.251953125,
0.59326171875,
0.251953125
],
[
0.1829833984375,
0.250732421875,
0.455810546875,
0.251953125,
0.455810546875,
0.266357421875,
0.1829833984375,
0.26513671875
],
[
0.59228515625,
0.250732421875,
0.71728515625,
0.251953125,
0.71728515625,
0.266357421875,
0.59228515625,
0.26513671875
],
[
0.0187835693359375,
0.261474609375,
0.062103271484375,
0.260009765625,
0.066162109375,
0.373779296875,
0.0228729248046875,
0.375
],
[
0.1829833984375,
0.266357421875,
0.35302734375,
0.266357421875,
0.35302734375,
0.27978515625,
0.1829833984375,
0.27978515625
],
[
0.5947265625,
0.266357421875,
0.7607421875,
0.266357421875,
0.7607421875,
0.276611328125,
0.5947265625,
0.276611328125
],
[
0.1846923828125,
0.300537109375,
0.250732421875,
0.300537109375,
0.250732421875,
0.31494140625,
0.1846923828125,
0.31494140625
],
[
0.1829833984375,
0.314453125,
0.406005859375,
0.313232421875,
0.406005859375,
0.32763671875,
0.1829833984375,
0.328857421875
],
[
0.181396484375,
0.32763671875,
0.30810546875,
0.328857421875,
0.30810546875,
0.343505859375,
0.181396484375,
0.34228515625
],
[
0.1829833984375,
0.341552734375,
0.351318359375,
0.341552734375,
0.351318359375,
0.354736328125,
0.1829833984375,
0.354736328125
],
[
0.0269622802734375,
0.365478515625,
0.056365966796875,
0.365478515625,
0.056365966796875,
0.399658203125,
0.0269622802734375,
0.399658203125
],
[
0.453369140625,
0.393310546875,
0.5458984375,
0.393310546875,
0.5458984375,
0.40771484375,
0.453369140625,
0.40771484375
],
[
0.0228729248046875,
0.401611328125,
0.062103271484375,
0.401611328125,
0.062103271484375,
0.4892578125,
0.0228729248046875,
0.4892578125
],
[
0.233642578125,
0.425537109375,
0.765625,
0.425537109375,
0.765625,
0.439453125,
0.233642578125,
0.439453125
],
[
0.2320556640625,
0.439453125,
0.7685546875,
0.438232421875,
0.7685546875,
0.45263671875,
0.2320556640625,
0.453857421875
],
[
0.2344970703125,
0.455078125,
0.765625,
0.455078125,
0.765625,
0.46533203125,
0.2344970703125,
0.46533203125
],
[
0.233642578125,
0.46728515625,
0.7685546875,
0.46728515625,
0.7685546875,
0.48095703125,
0.233642578125,
0.48095703125
],
[
0.041656494140625,
0.481689453125,
0.053924560546875,
0.481689453125,
0.053924560546875,
0.48486328125,
0.041656494140625,
0.48486328125
],
[
0.233642578125,
0.48095703125,
0.765625,
0.48095703125,
0.765625,
0.494384765625,
0.233642578125,
0.494384765625
],
[
0.0228729248046875,
0.49365234375,
0.05963134765625,
0.49365234375,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.233642578125,
0.494384765625,
0.765625,
0.494384765625,
0.765625,
0.50830078125,
0.233642578125,
0.50830078125
],
[
0.233642578125,
0.5087890625,
0.76708984375,
0.5087890625,
0.76708984375,
0.52294921875,
0.233642578125,
0.52294921875
],
[
0.233642578125,
0.52294921875,
0.76708984375,
0.52294921875,
0.76708984375,
0.5361328125,
0.233642578125,
0.5361328125
],
[
0.2303466796875,
0.5361328125,
0.765625,
0.53515625,
0.765625,
0.5498046875,
0.2303466796875,
0.55029296875
],
[
0.233642578125,
0.55029296875,
0.76708984375,
0.55029296875,
0.76708984375,
0.564453125,
0.233642578125,
0.564453125
],
[
0.2320556640625,
0.564453125,
0.765625,
0.564453125,
0.765625,
0.5791015625,
0.2320556640625,
0.5791015625
],
[
0.2344970703125,
0.580078125,
0.7646484375,
0.580078125,
0.7646484375,
0.59033203125,
0.2344970703125,
0.59033203125
],
[
0.233642578125,
0.59228515625,
0.765625,
0.59228515625,
0.765625,
0.60595703125,
0.233642578125,
0.60595703125
],
[
0.233642578125,
0.60595703125,
0.765625,
0.60595703125,
0.765625,
0.619140625,
0.233642578125,
0.619140625
],
[
0.2320556640625,
0.619140625,
0.765625,
0.619140625,
0.765625,
0.63330078125,
0.2320556640625,
0.63330078125
],
[
0.233642578125,
0.6337890625,
0.76708984375,
0.6337890625,
0.76708984375,
0.64794921875,
0.233642578125,
0.64794921875
],
[
0.233642578125,
0.646484375,
0.765625,
0.646484375,
0.765625,
0.6611328125,
0.233642578125,
0.6611328125
],
[
0.233642578125,
0.6611328125,
0.765625,
0.6611328125,
0.765625,
0.6748046875,
0.233642578125,
0.6748046875
],
[
0.233642578125,
0.6748046875,
0.765625,
0.6748046875,
0.765625,
0.68798828125,
0.233642578125,
0.68798828125
],
[
0.2320556640625,
0.6875,
0.76708984375,
0.68798828125,
0.76708984375,
0.70263671875,
0.2320556640625,
0.7021484375
],
[
0.233642578125,
0.70263671875,
0.43310546875,
0.70263671875,
0.43310546875,
0.71630859375,
0.233642578125,
0.71630859375
],
[
0.426513671875,
0.70263671875,
0.7685546875,
0.70263671875,
0.7685546875,
0.71630859375,
0.426513671875,
0.71630859375
],
[
0.233642578125,
0.71728515625,
0.765625,
0.71728515625,
0.765625,
0.73095703125,
0.233642578125,
0.73095703125
],
[
0.2320556640625,
0.72998046875,
0.5595703125,
0.72998046875,
0.5595703125,
0.74365234375,
0.2320556640625,
0.74365234375
],
[
0.178955078125,
0.77587890625,
0.1846923828125,
0.77587890625,
0.1846923828125,
0.78271484375,
0.178955078125,
0.78271484375
],
[
0.2017822265625,
0.771484375,
0.338134765625,
0.77294921875,
0.338134765625,
0.787109375,
0.2017822265625,
0.7861328125
],
[
0.1748046875,
0.8037109375,
0.82275390625,
0.8037109375,
0.82275390625,
0.818359375,
0.1748046875,
0.818359375
],
[
0.1732177734375,
0.818359375,
0.826171875,
0.8173828125,
0.826171875,
0.83203125,
0.1732177734375,
0.83349609375
],
[
0.1732177734375,
0.83203125,
0.826171875,
0.83203125,
0.826171875,
0.84521484375,
0.1732177734375,
0.84521484375
],
[
0.1748046875,
0.8466796875,
0.82421875,
0.8466796875,
0.82421875,
0.85986328125,
0.1748046875,
0.85986328125
],
[
0.1732177734375,
0.859375,
0.82421875,
0.85791015625,
0.82421875,
0.87255859375,
0.1732177734375,
0.87353515625
],
[
0.1732177734375,
0.87353515625,
0.82421875,
0.87353515625,
0.82421875,
0.88818359375,
0.1732177734375,
0.88818359375
],
[
0.1748046875,
0.88818359375,
0.82421875,
0.88818359375,
0.82421875,
0.9013671875,
0.1748046875,
0.9013671875
],
[
0.1968994140625,
0.91357421875,
0.314453125,
0.91357421875,
0.314453125,
0.923828125,
0.1968994140625,
0.923828125
],
[
0.49267578125,
0.94873046875,
0.5048828125,
0.94873046875,
0.5048828125,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2024",
"ADVERSARIAL IMITATION LEARNING VIA BOOSTING",
"Jonathan D. Chang",
"Dhruv Sreenivas *",
"Department of Computer Science",
"Department of Computer Science",
"Cornell University",
"Cornell University",
"jdc396@cornell.edu",
"ds844@cornell.edu",
"Yingbing Huang *",
"Kiante Brantley",
"Department of Electrical and Computer Engineering",
"Department of Computer Science",
"University of Illinois Urbana-Champaign",
"Cornell University",
"",
"yh2l@illinois.edu",
"kdb82@cornell.edu",
"Wen Sun",
"Department of Computer Science",
"Cornell University",
"ws455@cornel1.edu",
"z1",
"ABSTRACT",
"[T'S]",
"Adversarial imitation learning (AIL) has stood out as a dominant framework across",
"various imitation learning (IL) applications, with Discriminator Actor Critic (DAC)",
"(Kostrikovetal.,2019)demonstratingtheeffectivenessofoff-policylearningalgo-",
"rithms in improving sample efficiency and scalability to higher-dimensional obser-",
"",
"vations. Despite DAC's empirical success, the original AIL objective is on-policy",
"80",
"and DAc's ad-hoc application of off-policy training does not guarantee successful",
"imitation (Kostrikov et al., 2019; 2020). Follow-up work such as ValueDICE",
"(Kostrikov et al., 2020) tackles this issue by deriving a fully off-policy AIL ob-",
"jective. Instead in this work, we develop a novel and principled AIL algorithm",
"via the framework of boosting. Like boosting, our new algorithm, AILBoost,",
"maintains an ensemble of properly weighted weak learners (i.e., policies) and trains",
"a discriminatorthatwitnessesthemaximumdiscrepancybetweenthedistributions",
"of the ensemble and the expert policy. We maintain a weighted replay buffer to",
"represent the state-action distribution induced by the ensemble, allowing us to",
"train discriminators using the entire data collected so far. In the weighted replay",
"buffer, the contribution of the data from older policies are properly discounted",
"with the weight computed based on the boosting framework. Empirically, we",
"evaluate our algorithm on both controller state-based and pixel-based environments",
"from the DeepMind Control Suite. AILBoost outperforms DAC on both types",
"of environments, demonstrating the benefit of properly weighting replay buffer",
"data for off-policy training.",
"On state-based environments, AILBoost outper-",
"forms ValueDICE and IQ-Learn(Garg et al., 2021), achieving competitive",
"performance with as little as one expert trajectory",
"-",
"INTRODUCTION",
"Imitation learning (IL) is a promising paradigm for learning general policies without rewards from",
"demonstration data, achieving remarkable success in autonomous driving (Bronstein et al., 2022;",
"Pomerleau, 1988), video games (Baker et al., 2022; Shah et al., 2022) and graphics (Peng et al., 2021).",
"Adversarial Imitation Learning (AIL) is an incredibly successful approach for imitation learning",
"(Ho & Ermon, 2016; Fu et al., 2018; Kostrikov et al., 2019; Ke et al., 2020). These methods cast IL",
"as a distribution matching problem whereby the learning agent minimizes the divergence between",
"the expert demonstrator's distribution and the state-action distribution induced by the agent. First",
"*EqualContribution",
"1"
] |
{
"arxiv_id": "2404.08513",
"language": "en",
"page_idx": 0
}
|
||
2404.08513-en-1
|
2404.08513-en-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
introduced by (Ho&Ermon, 2016), this divergence minimization can be achieved in an iterative procedure reminiscent of GAN algorithms (Goodfellow et al., 2014) with our learned reward function and policy being the discriminator and generator respectively.
Originally, a limitation of many AIL methods was that they were $on-policy.$ That is, for on-policy AIL methods like GAIL (Ho&Ermon, 2016) and AIRL (Fu et al.., 2018), the algorithm would draw fresh samples from the current policy in every iteration for the distribution matching process while discarding all old samples, rendering the sample complexity of these algorithms to be prohibitively large in many applications. Follow-up works (Kostrikov et al., 2019; Šasaki et al., 2019) attempt to relax the $on-policy$ requirement by creating off-policy methods that utilize the entire history of observed data during the learning process. This history is often represented by a replay buffer and methods such as Discriminator Actor Critic (DAC) show large improvements in scalability and sample complexity over their on-policy counterparts. However, these methods modify the distribution matching objective as a divergence minimization between the replay buffer's and the expert's distribution, losing the guarantee of matching the expert's behavior.
Algorithms like ValueDICE (Kostrikov et al., 2020) address this problem by deriving a new formulation of the AIL divergence minimization objective to be entirely off-policy. ValueDICE, however, in principle relies on the environments to have deterministic dynamics.$^{1}$ In this work, we consider a new perspective towards making AIL off-policy. We present a new principled off-policy AIL algorithm, AILBoost, via the gradient boosting framework (Mason et al., 1999). AILBoost maintains an ensemble of properly weighted weak learners or policies as well as a weighted replay buffer to represent the state-action distribution induced by our ensemble. Our distribution matching objective is then to minimize the divergence between the weighted replay buffer's distribution (i.e., the state-action distribution induced by the ensemble) and the expert demonstrator's distribution, making the divergence minimization problem an off-policy learning problem. Similar to boosting and gradient boosting, at every iteration, we aim to find a weak learner, such that when added to the ensemble, the divergence between the updated ensemble's distribution and the expert's distribution decreases. In other words, our approach can be understood as performing gradient boosting in the state-action occupancy space, where black-box RL optimizer is used a weak learning procedure to train weak learners, i.e., policies.
We evaluate AILBoost on the DeepMind Control Suite (Tassa et al., 2018) and compare against a range of off-policy AIL algorithms (Behavior cloning, ValueDICE, DAC) as well as a state-of-theart IL algorithm, IQ-Learn. We show that our algorithm is comparable to or more sample efficient than state-of-the-art IL algorithms in various continuous control tasks, achieving strong imitation performance with as little as one expert demonstration. We also show that our approach scales to $\hat{\text{vision-based, partially observable domains, where we again outperform DAC}}.$
## 2 $\mathbf{RELATED}$ WORKS
Off-policy and Offline IL There has also been a wide variety of research conducted on off-policy and offline IL, where the goal is to be either more sample efficient or safer by utilizing a replay buffer or not collecting any environmental transitions during training, respectively. The most prominent of said methods, and the closest to our work, is Discriminator-Actor-Critic (DAC) (Kostrikov et al., 2019), which essentially replaces the on-policy RL algorithm in the adversarial IL setup with an off-policy one such as DDPG (Lillicrap et al., 2019) or SAC (Haarnoja et al., 2018). However, as mentioned previously, DAC doesn't necessarily guarantee a distribution match between the expert and the learned policy, prompting further work to be done. Further work has primarily focused on weighting on-policy and off-policy data differently in both the policy update and the discriminator update. ValueDICE (Kostrikov et al., 2020) mitigates this problem by deriving an objective $\hat{\text{from the original distribution matching problem that only requires off-policy samples to compute.}}$ More recently, methods such as IQ-Learn (Garg et al., 2021) have been developed to learn soft Q functions over the environment space, which encodes both a reward and a policy for inverse reinforcement learning, and model-based methods such as V-MAIL (Rafailov et al., 2021) have shown that using expressive world models (Hafner et al., 2020) leads to strong imitation results in
$^{1}$One cannot derive an unbiased estimate of the objective function proposed in ValueDICE unless it has infinite expert samples and the transition is deterministic (Kostrikov et al., 2020). See section 3.3 for more detailed discussion.
2
|
[
[
0.17236328125,
0.0328369140625,
0.48046875,
0.032196044921875,
0.48046875,
0.046722412109375,
0.17236328125,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82421875,
0.10479736328125,
0.82421875,
0.11871337890625,
0.1748046875,
0.11871337890625
],
[
0.1748046875,
0.11932373046875,
0.82421875,
0.11932373046875,
0.82421875,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.1732177734375,
0.1319580078125,
0.57861328125,
0.1337890625,
0.57861328125,
0.147705078125,
0.1732177734375,
0.1458740234375
],
[
0.1748046875,
0.154052734375,
0.82275390625,
0.154052734375,
0.82275390625,
0.1673583984375,
0.1748046875,
0.1673583984375
],
[
0.176513671875,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.176513671875,
0.1817626953125
],
[
0.1748046875,
0.1817626953125,
0.82421875,
0.1817626953125,
0.82421875,
0.1956787109375,
0.1748046875,
0.1956787109375
],
[
0.1748046875,
0.1956787109375,
0.82275390625,
0.1956787109375,
0.82275390625,
0.208984375,
0.1748046875,
0.208984375
],
[
0.1748046875,
0.208984375,
0.82275390625,
0.208984375,
0.82275390625,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1748046875,
0.2235107421875,
0.826171875,
0.2235107421875,
0.826171875,
0.2374267578125,
0.1748046875,
0.2374267578125
],
[
0.1732177734375,
0.2374267578125,
0.82421875,
0.236083984375,
0.82421875,
0.250732421875,
0.1732177734375,
0.251953125
],
[
0.1748046875,
0.251953125,
0.82421875,
0.251953125,
0.82421875,
0.26513671875,
0.1748046875,
0.26513671875
],
[
0.1748046875,
0.26513671875,
0.82421875,
0.26513671875,
0.82421875,
0.279052734375,
0.1748046875,
0.279052734375
],
[
0.1732177734375,
0.27783203125,
0.82421875,
0.279052734375,
0.82421875,
0.29345703125,
0.1732177734375,
0.292236328125
],
[
0.1748046875,
0.29345703125,
0.67236328125,
0.29345703125,
0.67236328125,
0.306884765625,
0.1748046875,
0.306884765625
],
[
0.1748046875,
0.3125,
0.82421875,
0.313232421875,
0.82421875,
0.32763671875,
0.1748046875,
0.326904296875
],
[
0.1748046875,
0.32763671875,
0.826171875,
0.32763671875,
0.826171875,
0.341552734375,
0.1748046875,
0.341552734375
],
[
0.1732177734375,
0.341552734375,
0.82421875,
0.341552734375,
0.82421875,
0.354736328125,
0.1732177734375,
0.354736328125
],
[
0.1748046875,
0.35595703125,
0.82421875,
0.35595703125,
0.82421875,
0.369384765625,
0.1748046875,
0.369384765625
],
[
0.1748046875,
0.369384765625,
0.82275390625,
0.369384765625,
0.82275390625,
0.38330078125,
0.1748046875,
0.38330078125
],
[
0.1732177734375,
0.3837890625,
0.82421875,
0.3837890625,
0.82421875,
0.397705078125,
0.1732177734375,
0.397705078125
],
[
0.1748046875,
0.397705078125,
0.82421875,
0.397705078125,
0.82421875,
0.410888671875,
0.1748046875,
0.410888671875
],
[
0.1748046875,
0.410888671875,
0.82666015625,
0.410888671875,
0.82666015625,
0.4248046875,
0.1748046875,
0.4248046875
],
[
0.1732177734375,
0.423583984375,
0.82666015625,
0.4248046875,
0.82666015625,
0.439453125,
0.1732177734375,
0.438232421875
],
[
0.1748046875,
0.439453125,
0.82421875,
0.439453125,
0.82421875,
0.45263671875,
0.1748046875,
0.45263671875
],
[
0.1732177734375,
0.45263671875,
0.82275390625,
0.45263671875,
0.82275390625,
0.466552734375,
0.1732177734375,
0.466552734375
],
[
0.176513671875,
0.468505859375,
0.82275390625,
0.468505859375,
0.82275390625,
0.479248046875,
0.176513671875,
0.479248046875
],
[
0.1748046875,
0.479736328125,
0.82421875,
0.479736328125,
0.82421875,
0.494384765625,
0.1748046875,
0.494384765625
],
[
0.1748046875,
0.49365234375,
0.82421875,
0.49365234375,
0.82421875,
0.50830078125,
0.1748046875,
0.50830078125
],
[
0.1748046875,
0.5087890625,
0.393798828125,
0.5087890625,
0.393798828125,
0.52294921875,
0.1748046875,
0.52294921875
],
[
0.1748046875,
0.52880859375,
0.82421875,
0.52880859375,
0.82421875,
0.54248046875,
0.1748046875,
0.54248046875
],
[
0.1748046875,
0.54345703125,
0.826171875,
0.54345703125,
0.826171875,
0.556640625,
0.1748046875,
0.556640625
],
[
0.1748046875,
0.556640625,
0.82421875,
0.556640625,
0.82421875,
0.57080078125,
0.1748046875,
0.57080078125
],
[
0.176513671875,
0.57275390625,
0.82275390625,
0.57275390625,
0.82275390625,
0.58349609375,
0.176513671875,
0.58349609375
],
[
0.1748046875,
0.58544921875,
0.826171875,
0.58544921875,
0.826171875,
0.5986328125,
0.1748046875,
0.5986328125
],
[
0.1748046875,
0.5986328125,
0.6787109375,
0.5986328125,
0.6787109375,
0.6123046875,
0.1748046875,
0.6123046875
],
[
0.1748046875,
0.6337890625,
0.19287109375,
0.6337890625,
0.19287109375,
0.64794921875,
0.1748046875,
0.64794921875
],
[
0.200927734375,
0.63330078125,
0.351318359375,
0.6337890625,
0.351318359375,
0.6484375,
0.200927734375,
0.64794921875
],
[
0.1748046875,
0.66552734375,
0.82275390625,
0.66552734375,
0.82275390625,
0.6796875,
0.1748046875,
0.6796875
],
[
0.1732177734375,
0.6796875,
0.82275390625,
0.6796875,
0.82275390625,
0.693359375,
0.1732177734375,
0.693359375
],
[
0.1748046875,
0.6943359375,
0.826171875,
0.6943359375,
0.826171875,
0.70849609375,
0.1748046875,
0.70849609375
],
[
0.1748046875,
0.70849609375,
0.826171875,
0.70849609375,
0.826171875,
0.7216796875,
0.1748046875,
0.7216796875
],
[
0.1748046875,
0.7216796875,
0.82421875,
0.7216796875,
0.82421875,
0.73486328125,
0.1748046875,
0.73486328125
],
[
0.1732177734375,
0.734375,
0.82421875,
0.73486328125,
0.82421875,
0.75,
0.1732177734375,
0.74853515625
],
[
0.1732177734375,
0.75,
0.82421875,
0.75,
0.82421875,
0.76318359375,
0.1732177734375,
0.76318359375
],
[
0.1732177734375,
0.76318359375,
0.82421875,
0.76318359375,
0.82421875,
0.77783203125,
0.1732177734375,
0.77783203125
],
[
0.176513671875,
0.77880859375,
0.82275390625,
0.77880859375,
0.82275390625,
0.7890625,
0.176513671875,
0.7890625
],
[
0.1748046875,
0.79150390625,
0.82421875,
0.79150390625,
0.82421875,
0.8046875,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.8046875,
0.826171875,
0.8046875,
0.826171875,
0.818359375,
0.1748046875,
0.818359375
],
[
0.1748046875,
0.818359375,
0.82421875,
0.818359375,
0.82421875,
0.83203125,
0.1748046875,
0.83203125
],
[
0.1748046875,
0.83203125,
0.82421875,
0.83203125,
0.82421875,
0.84521484375,
0.1748046875,
0.84521484375
],
[
0.1748046875,
0.8466796875,
0.82421875,
0.8466796875,
0.82421875,
0.85986328125,
0.1748046875,
0.85986328125
],
[
0.1732177734375,
0.85986328125,
0.82421875,
0.85986328125,
0.82421875,
0.87353515625,
0.1732177734375,
0.87353515625
],
[
0.19775390625,
0.88623046875,
0.82421875,
0.88623046875,
0.82421875,
0.8994140625,
0.19775390625,
0.8994140625
],
[
0.1732177734375,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.1748046875,
0.91357421875,
0.293212890625,
0.91357421875,
0.293212890625,
0.923828125,
0.1748046875,
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 2024",
"introduced by (Ho & Ermon, 2016), this divergence minimization can be achieved in an iterative",
"procedure reminiscent of GAN algorithms (Goodfellow et al., 2014) with our learned reward function",
"and policybeingthediscriminatorandgeneratorrespectively.",
"Originally, a limitation of many AIL methods was that they were on-policy. That is, for on-policy",
"AIL methods like GAIL (Ho & Ermon, 2016) and AIRL (Fu et al., 2018), the algorithm would draw",
"fresh samples from the current policy in every iteration for the distribution matching process while",
"discarding all old samples, rendering the sample complexity of these algorithms to be prohibitively",
"large in many applications. Follow-up works (Kostrikov et al., 2019; Sasaki et al., 2019) attempt",
"to relax the on-policy requirement by creating off-policy methods that utilize the entire history of",
"observed data during the learning process.This history is often represented by a replay buffer",
"and methods such as Discriminator Actor Critic (DAC) show large improvements in scalability",
"and sample complexity over their on-policy counterparts. However, these methods modify the",
"distribution matching objective as a divergence minimization between the replay buffer's and the",
"expert's distribution, losing the guarantee of matching the expert's behavior.",
"Algorithms like ValueDICE (Kostrikov et al., 2020) address this problem by deriving a new",
"formulation of the AIL divergence minimization objective to be entirely off-policy. Va lueD ICE,",
"however, in principle relies on the environments to have deterministic dynamics.I In this work, we",
"consider a new perspective towards making AIL off-policy. We present a new principled off-policy",
"AIL algorithm, AILBoost, via the gradient boosting framework (Mason et al., 1999). AILBoost",
"maintains an ensemble of properly weighted weak learners or policies as well as a weighted replay",
"buffer to represent the state-action distribution induced by our ensemble. Our distribution matching",
"objective is then to minimize the divergence between the weighted replay buffer's distribution (i.e.,",
"the state-action distribution induced by the ensemble) and the expert demonstrator's distribution,",
"making the divergence minimization problem an off-policy learning problem. Similar to boosting",
"and gradient boosting, at every iteration, we aim to find a weak learner, such that when added to the",
"ensemble,thedivergencebetweentheupdatedensemble'sdistributionandtheexpert'sdistribution",
"decreases. In other words, our approach can be understood as performing gradient boosting in the",
"state-action occupancy space, where black-box RL optimizer is used a weak learning procedure to",
"train weak learners, i.e., policies.",
"We evaluate AI LBoost on the DeepMind Control Suite (Tassa et al., 2018) and compare against a",
"range of off-policy AIL algorithms (Behavior cloning, ValueDICE, DAC) as well as a state-of-the-",
"art IL algorithm, IQ-Learn. We show that our algorithm is comparable to or more sample efficient",
"than state-of-the-artILalgorithmsinvarious continuouscontroltasks,achievingstrongimitation",
"performance with as little as one expert demonstration. We also show that our approach scales to",
"vision-based, partially observable domains, where we again outperform DAC.",
"2",
"RELATEDWORKS",
"Off-policy and Offline ILThere has also been a wide variety of research conducted on off-policy",
"and offline IL, where the goal is to be either more sample efficient or safer by utilizing a replay buffer",
"or not collecting any environmental transitions during training, respectively. The most prominent of",
"said methods, and the closest to our work, is Discriminator-Actor-Critic (DAC) (Kostrikov et al.,",
"2019), which essentially replaces the on-policy RL algorithm in the adversarial IL setup with an",
"off-policy one such as DDPG (Lillicrap et al., 2019) or SAC (Haarnoja et al., 2018). However, as",
"mentioned previously, DAC doesn't necessarily guarantee a distribution match between the expert",
"and the learned policy, prompting further work to be done. Further work has primarily focused on",
"weightingon-policyandoff-policydatadifferentlyinboththepolicyupdateandthediscriminator",
"update. ValueDICE (Kostrikov et al., 2020) mitigates this problem by deriving an objective",
"from the original distribution matching problem that only requires off-policy samples to compute.",
"More recently, methods such as IQ-Learn (Garg et al., 2021) have been developed to learn soft",
"Q functions over the environment space, which encodes both a reward and a policy for inverse",
"reinforcement learning, and model-based methods such as V-MAIL (Rafailov et al., 2021) have",
"shown that using expressive world models (Hafner et al., 2020) leads to strong imitation results in",
"'One cannot derive an unbiased estimate of the objective function proposed in ValueDICE unless it has",
"infinite expert samples and the transition is deterministic (Kostrikov et al., 2020). See section 3.3 for more",
"detaileddiscussion.",
"2"
] |
{
"arxiv_id": "2404.08513",
"language": "en",
"page_idx": 1
}
|
||
2404.08513-en-2
|
2404.08513-en-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
domains with high-dimensional observations. Other off-policy IL works include SoftDICE (Sun et al., 2021), SparseDICE (Camacho et al., 2021), and AdVIIL/AdRIL/DAeQuIL (Swamy et al., 2021).
Orthogonally, on the offline side, where environment interaction is prohibited, works both on the model-based side (Chang et al., 2021) and the model-free side (Kim et al., 2022; Yu et al., 2023) has shown that distribution matching is still possible in these settings. These approaches generally operate either by learning a transition model of the environment, with which to roll out in to do policy $\tilde{\text{optimization }(\text{Chang et al., 2021), or optimizing a modified version of the objective introduced in}}$ (Kostrikov et al., 2020) by using samples from the suboptimal offline dataset as opposed to on-policy samples for computation.
Boosting style approach in deep learning &RL The idea of using boosting for policy learning is not new in the deep learning or reinforcement learning literature. On the deep learning side, AdaGAN (Tolstikhin et al., 2017) apply standard adaptive boosting to GANs (Goodfellow et al., 2014) to address and fix issues such as mode collapse, while concurrent work (Grover & Ermon, 2017) showed benefits of boosting in general Bayesian mixture models. In RL, the conservative policy iteration (CPI) (Kakade& Langford, 2002) can be understood as performing gradient boosting in the policy space (Scherrer &Geist, 2014). The authors in (Hazan et al., 2019) use a gradient boosting style approach to learn maximum entropy policies. In this work, we perform gradient boosting in the space of state-action occupancy measures, which leads to a principled off-policy IL approach.
## 3 PRELIMINARIES
We consider a discounted infinite horizon MDP $\mathcal{M}=\langle\mathcal{S},P,\mathcal{A},r,\gamma,\mu_0\rangle$ where $\mathcal{S}$ is the state of states, $\mathcal{A}$ is the set of actions, $r:S\times\mathcal{A}\mapsto\mathbb{R}$ is the reward function and $r(s,a)$ is the reward for the given state-action pair, $\gamma\in(0,1)$ is the discount factor, $\mu_0\in\Delta(S)$ is the initial state distribution, and $P:\mathcal{S}\times\mathcal{A}\mapsto\Delta(\mathcal{S})$ is the transition function. A policy $\pi:\mathcal{S}\to\Delta(\mathcal{A})$ interacts in said MDP, creating $trajectories\tau$ composed of state-action pairs $\{(s_t,a_t)\}_{t=1}^T.$ We denote $d_t^\pi$ to represent the stateaction visitation distribution induced by $\pi$ at timestep $t$ and $d^\pi=(1-\gamma)\sum_{t=0}^{\infty}\gamma^td_t^\pi$ as the average state-action visitation distribution induced by policy $\pi.$ We define the value function and $Q$-function of our policy as$V^{\pi }( s) = \mathbb{E} _{\pi }[ \sum _{t= 0}^{\infty }\gamma ^{t}r( s_{t}) | s_{0}= s]$ and $Q^{\pi }( s, a) = r( s, a) + \mathbb{E} _{s^{\prime }\sim P( \cdot | s, a) }[ V^{\pi }( s^{\prime }) ] .$ Tba gal af DI is ta frad andivat thet therat d arwatd ard ard ard ard ard ard ard ard ard ard ard ard ard ard ard ar The goal of RL is to find a policy that maximizes the expected cumulative reward.
In imitation learning, instead of having access to the reward function, we assume access to demonstrations $\mathcal{D}^e=\{(s_i,a_i)\}_{i=1}^N$ from an expert policy $\pi^e$ that our policy can take advantage of while training. Note that $\pi^e$ might not necessarily be a Markovian policy. It is possible that $\pi^e$ is an ensemble of weighted Markovian policies, i.e., $\pi^e=\{\alpha_i,\pi_i\}_{i=1}^{\hat{n}}$ with $\alpha_i\geq0,\sum_i\alpha_i=1$,which means that for each episode, $\pi^e$ will fırst randomly sample a policy $\pi_i$ with probability $\alpha_i$ at $t=0$, and then execute $\pi_i$ for the entire episode (i.e., no switch to other policies during the execution for an episode). It is well known that the space of state action distributions induced by such ensembles is larger than the space of state-action distributions induced by Markovian policies (Hazan et al., 2019). The goal in IL is then to learn a policy that robustly mimics the expert. The simplest imitation learning algorithm to address this issue is behavior cloning (BC): argmin $\hat{\underset{\pi\in\Pi}{\operatorname*{E}}}_{(s,a)\sim\mathcal{D}^{\epsilon}}[\ell(\pi(s),a)]$ where $\ell$ is a classification loss and $\Pi$ is our policy class. Though this objective is simple, it is known to suffer from covariateshift at test time (Pomerleau, 1988; Ross et al., 2011). Instead of minimizing action distribution divergence conditioned on expert states, algorithms such as inverse RL (Ziebart et al., 2008) and adversarial IL (Ho & Ermon, $\hat{20}16;$ Finn et al., 2016; Ke et al., 2020; Sun et al., 2019) directly minimize some divergence metrics between state-action distributions, which help address the covariate shift issue (Agarwal et al., 2019).
3.1 ADVERSARIAL IMITATION LEARNING (AIL)
The goal of AIL is to directly minimize some divergence between some behavior policy state-action visitation $d^\pi$ and an expert policy state-action visitation $d^{\pi^e}.$ The choice of divergence results in variously different AIL algorithms.
The most popular AIL algorithm is Generative Adversarial Imitation Learning (GAIL) (Ho &Ermon, 2016) which minimizes the JS-divergence. This algorithm is a on-policy adversarial imitation
3
|
[
[
0.1732177734375,
0.0328369140625,
0.48046875,
0.032196044921875,
0.48046875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.1748046875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11871337890625,
0.1748046875,
0.11871337890625
],
[
0.1748046875,
0.11932373046875,
0.82275390625,
0.11932373046875,
0.82275390625,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.1748046875,
0.154052734375,
0.82275390625,
0.154052734375,
0.82275390625,
0.1673583984375,
0.1748046875,
0.1673583984375
],
[
0.1748046875,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.1748046875,
0.1817626953125
],
[
0.1732177734375,
0.1817626953125,
0.82177734375,
0.1817626953125,
0.82177734375,
0.1956787109375,
0.1732177734375,
0.1956787109375
],
[
0.176513671875,
0.1976318359375,
0.82177734375,
0.1976318359375,
0.82177734375,
0.2083740234375,
0.176513671875,
0.2083740234375
],
[
0.1748046875,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1748046875,
0.2235107421875,
0.82275390625,
0.2235107421875,
0.82275390625,
0.2374267578125,
0.1748046875,
0.2374267578125
],
[
0.176513671875,
0.2392578125,
0.340576171875,
0.2392578125,
0.340576171875,
0.25,
0.176513671875,
0.25
],
[
0.1748046875,
0.268310546875,
0.82275390625,
0.268310546875,
0.82275390625,
0.2822265625,
0.1748046875,
0.2822265625
],
[
0.1748046875,
0.28271484375,
0.82275390625,
0.28271484375,
0.82275390625,
0.296630859375,
0.1748046875,
0.296630859375
],
[
0.1748046875,
0.296630859375,
0.82275390625,
0.296630859375,
0.82275390625,
0.311279296875,
0.1748046875,
0.311279296875
],
[
0.1748046875,
0.31005859375,
0.82275390625,
0.31005859375,
0.82275390625,
0.323974609375,
0.1748046875,
0.323974609375
],
[
0.1732177734375,
0.323974609375,
0.82275390625,
0.323974609375,
0.82275390625,
0.337158203125,
0.1732177734375,
0.337158203125
],
[
0.1748046875,
0.33837890625,
0.82421875,
0.33837890625,
0.82421875,
0.3515625,
0.1748046875,
0.3515625
],
[
0.1748046875,
0.3515625,
0.82275390625,
0.3515625,
0.82275390625,
0.365478515625,
0.1748046875,
0.365478515625
],
[
0.1748046875,
0.365478515625,
0.82275390625,
0.365478515625,
0.82275390625,
0.37890625,
0.1748046875,
0.37890625
],
[
0.1748046875,
0.380126953125,
0.74609375,
0.380126953125,
0.74609375,
0.393310546875,
0.1748046875,
0.393310546875
],
[
0.176513671875,
0.41845703125,
0.191162109375,
0.41845703125,
0.191162109375,
0.4287109375,
0.176513671875,
0.4287109375
],
[
0.19775390625,
0.415283203125,
0.339111328125,
0.416748046875,
0.339111328125,
0.43115234375,
0.19775390625,
0.429931640625
],
[
0.17236328125,
0.447509765625,
0.82421875,
0.447509765625,
0.82421875,
0.462158203125,
0.17236328125,
0.462158203125
],
[
0.1748046875,
0.462158203125,
0.82275390625,
0.462158203125,
0.82275390625,
0.47607421875,
0.1748046875,
0.47607421875
],
[
0.1748046875,
0.47607421875,
0.82275390625,
0.47607421875,
0.82275390625,
0.490478515625,
0.1748046875,
0.490478515625
],
[
0.1748046875,
0.490478515625,
0.82275390625,
0.490478515625,
0.82275390625,
0.50390625,
0.1748046875,
0.50390625
],
[
0.1732177734375,
0.50390625,
0.82421875,
0.50244140625,
0.82421875,
0.5166015625,
0.1732177734375,
0.517578125
],
[
0.1748046875,
0.517578125,
0.82275390625,
0.517578125,
0.82275390625,
0.5322265625,
0.1748046875,
0.5322265625
],
[
0.1748046875,
0.5322265625,
0.82275390625,
0.5322265625,
0.82275390625,
0.54541015625,
0.1748046875,
0.54541015625
],
[
0.17236328125,
0.54248046875,
0.82666015625,
0.54443359375,
0.82666015625,
0.5625,
0.17236328125,
0.56005859375
],
[
0.1748046875,
0.5595703125,
0.7099609375,
0.5595703125,
0.7099609375,
0.57373046875,
0.1748046875,
0.57373046875
],
[
0.1748046875,
0.580078125,
0.826171875,
0.580078125,
0.826171875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.1748046875,
0.5947265625,
0.82421875,
0.5947265625,
0.82421875,
0.609375,
0.1748046875,
0.609375
],
[
0.1748046875,
0.60791015625,
0.82421875,
0.60791015625,
0.82421875,
0.6220703125,
0.1748046875,
0.6220703125
],
[
0.17236328125,
0.62060546875,
0.826171875,
0.62060546875,
0.826171875,
0.63818359375,
0.17236328125,
0.63818359375
],
[
0.176513671875,
0.63720703125,
0.82421875,
0.63720703125,
0.82421875,
0.64794921875,
0.176513671875,
0.64794921875
],
[
0.1732177734375,
0.6494140625,
0.82421875,
0.6494140625,
0.82421875,
0.66357421875,
0.1732177734375,
0.66357421875
],
[
0.1748046875,
0.66357421875,
0.82421875,
0.66357421875,
0.82421875,
0.6767578125,
0.1748046875,
0.6767578125
],
[
0.1748046875,
0.67822265625,
0.82275390625,
0.67822265625,
0.82275390625,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1748046875,
0.69140625,
0.82421875,
0.69140625,
0.82421875,
0.705078125,
0.1748046875,
0.705078125
],
[
0.17236328125,
0.70263671875,
0.82421875,
0.705078125,
0.82421875,
0.7216796875,
0.17236328125,
0.7197265625
],
[
0.176513671875,
0.72021484375,
0.82275390625,
0.72021484375,
0.82275390625,
0.73095703125,
0.176513671875,
0.73095703125
],
[
0.1748046875,
0.73291015625,
0.82275390625,
0.73291015625,
0.82275390625,
0.7470703125,
0.1748046875,
0.7470703125
],
[
0.1748046875,
0.7470703125,
0.82275390625,
0.7470703125,
0.82275390625,
0.76025390625,
0.1748046875,
0.76025390625
],
[
0.1748046875,
0.76123046875,
0.826171875,
0.76123046875,
0.826171875,
0.7744140625,
0.1748046875,
0.7744140625
],
[
0.1748046875,
0.7744140625,
0.82421875,
0.7744140625,
0.82421875,
0.78857421875,
0.1748046875,
0.78857421875
],
[
0.1732177734375,
0.78857421875,
0.53515625,
0.78857421875,
0.53515625,
0.8017578125,
0.1732177734375,
0.8017578125
],
[
0.1748046875,
0.8212890625,
0.214111328125,
0.8212890625,
0.214111328125,
0.83349609375,
0.1748046875,
0.83349609375
],
[
0.2059326171875,
0.82080078125,
0.51953125,
0.82080078125,
0.51953125,
0.833984375,
0.2059326171875,
0.833984375
],
[
0.1732177734375,
0.84765625,
0.82275390625,
0.84765625,
0.82275390625,
0.861328125,
0.1732177734375,
0.861328125
],
[
0.1748046875,
0.8623046875,
0.82421875,
0.8623046875,
0.82421875,
0.87548828125,
0.1748046875,
0.87548828125
],
[
0.17236328125,
0.875,
0.406005859375,
0.87548828125,
0.406005859375,
0.89013671875,
0.17236328125,
0.8896484375
],
[
0.1748046875,
0.896484375,
0.826171875,
0.896484375,
0.826171875,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2024",
"domains with high-dimensional observations. Other off-policy IL works include SoftDICE (Sun",
"et al., 2021), SparseDICE (Camacho et al., 2021), and AdVIL/AdRIL/DAeQuIL (Swamy et al., 2021).",
"Orthogonally, on the offline side, where environment interaction is prohibited, works both on the",
"model-based side (Chang et al., 2021) and the model-free side (Kim et al., 2022; Yu et al., 2023)",
"has shown that distribution matching is still possible in these settings. These approaches generally",
"operateeitherbylearning atransitionmodel of theenvironment,withwhichtoroll out into dopolicy",
"optimization (Chang et al., 2021), or optimizing a modified version of the objective introduced in",
"(Kostrikov et al., 2020) by using samples from the suboptimal offline dataset as opposed to on-policy",
"samples for computation.",
"Boosting style approach in deep learning &RLThe idea of using boosting for policy learning is",
"not new in the deep learning or reinforcement learning literature. On the deep learning side, AdaGAN",
"(Tolstikhin et al., 2017) apply standard adaptive boosting to GANs (Goodfellow et al., 2014) to",
"address and fix issues such as mode collapse, while concurrent work (Grover & Ermon, 2017) showed",
"benefits of boosting in general Bayesian mixture models. In RL, the conservative policy iteration",
"(CPI) (Kakade & Langford, 2002) can be understood as performing gradient boosting in the policy",
"space (Scherrer & Geist, 2014). The authors in (Hazan et al., 2019) use a gradient boosting style",
"approach to learn maximum entropy policies. In this work, we perform gradient boosting in the space",
"of state-action occupancy measures, which leads to a principled off-policy IL approach.",
"3",
"PRELIMINARIES",
"We consider a discounted infinite horizon MDP M = (S, P, A, r, , μo) where S is the state of states,",
"Ais the set of actions, r : S × A IR is the reward function and r(s, a) is the reward for the given",
"state-action pair, E (O, 1) is the discount factor, μo E (S) is the initial state distribution, and",
"P : S× A → △(S) is the transition function. Apolicy T : S → △(A) interacts in said MDP, creating",
"trajectories T composed of state-action pairs ((st, at))T-: We denote dy to represent the state-",
"action visitation distribution induced by π at timestep t and d = (1 - ) Zt-o t dt as the average",
"state-action visitation distribution induced by policy π. We define the value function and Q-function",
"of our policy as V*(s) = Er[=o tr(st)]so = s] and Q*(s, a) = r(s, a) + Eg'~P(|/s,a)[V*(s')].",
"The goal of RL is to find a policy that maximizes the expected cumulative reward.",
"In imitation learning, instead of having access to the reward function, we assume access to demon-",
"strations De = ((si, ai)):-1 from an expert policy πe that our policy can take advantage of while",
"training. Note that πe might not necessarily be a Markovian policy. It is possible that πe is an",
"ensemble of weighted Markovian policies, i.e., πe = [αi, T;jni with a, ≥ 0, , ai = 1, which",
"meansthatforeachepisode,ewillfirstrandomlysampleapolicy,withprobabilityα,att=O",
"and then execute ; for the entire episode (i.e., no switch to other policies during the execution for",
"an episode). It is well known that the space of state action distributions induced by such ensembles",
"islargerthanthespaceof state-action distributionsinducedbyMarkovianpolicies(Hazan et al.",
"2019). The goal in IL is then to learn a policy that robustly mimics the expert. The simplest imitation",
"learning algorithm to address this issue is behavior cloning (BC): argminven E(s,a)~De[e(π(s), a)]",
"wherelis a classificationloss andIIis ourpolicy class.Thoughthis objectiveis simple,itisknown",
"to suffer from covariate shift at test time (Pomerleau, 1988; Ross et al., 2011). Instead of minimizing",
"action distribution divergence conditioned on expert states, algorithms such as inverse RL (Ziebart",
"et al., 2008) and adversarial IL (Ho & Ermon, 2016; Finn et al., 2016; Ke et al., 2020; Sun et al.,",
"2019) directly minimize some divergence metrics between state-action distributions, which help",
"address the covariate shift issue (Agarwal et al., 2019).",
"3.1",
"ADVERSARIALIMITATIONLEARNING(AIL)",
"The goal of AIL is to directly minimize some divergence between some behavior policy state-action",
"visitation d' and an expert policy state-action visitation d. The choice of divergence results in",
"variously different AIL algorithms.",
"The most popular AIL algorithm is Generative Adversarial Imitation Learning (GAIL) (Ho & Ermon,",
"2016) which minimizes the JS-divergence. This algorithm is a on-policy adversarial imitation",
"3"
] |
{
"arxiv_id": "2404.08513",
"language": "en",
"page_idx": 2
}
|
||
2404.08513-en-3
|
2404.08513-en-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
learning algorithm that connects Generative Adversarial Networks (GANs) (Goodfellow et al. 2014) and maximum entropy IRL (Ziebart et al., 2008). GAIL trains a binary classiffer called the discriminator $D(s,a)$ to distinguish between samples from the expert distribution and the policy generated distribution. Using the discriminator to define a reward function, GATL then executes an on-policy RL algorithm such as Trust Region Policy Optimization (TRPO) (Schulman et al., 2017a) or Proximal Policy Optimization (PPO) (Šchulman et al., 2017b) to maximize the reward. That gives us the following adversarial objective:
(1)
$$\min\limits_{\pi}\max\limits_{D}\mathbb{E}_{s,a\sim\pi}\left[\log D(s,a)\right]+\mathbb{E}_{s,a\sim\pi^e}\left[\log(1-D(s,a))\right]-\lambda H(\pi)$$
where $H(\pi)$ is an entropy regularization term. The first term in eq.(1)can be viewed as a pseudo reward that can be optimized with respect to the the policy $\pi$ on-policy samples. Note that GAIL typically optimizes both policies and discriminators using on-policy samples, making it quite sample inefficient. Using different divergences, there are various reward functions that can be optimized with this framework (Orsini et al., 2021). In this work, while our proposed approach in general is capable of optimizing many common divergences, we mainly focus on reverse KL divergence in our experiments. Reverse KL divergence has been studied in prior works including Fu et al. (2018); Ke et al. (2020). But different from prior works, we propose an off-policy method for optimizing reverse KL by leveraging the framework of boosting.
3.2 DISCRIMINATOR ACTOR CRITIC (DAC)
One reason GAIL need lot of interactions with the environment to learn properly is because of the dependency on using on-policy approaches to optimize discriminators and policies. In particular, GAIL does not reuse any old samples. Discriminator Actor Critic (DAC) (Kostrikov et al., 2019) extends GAIL algorithms to take advantage of off-policy learning to optimize the discriminators and policies. DAC introduces a replay buffer $\mathcal{R}$ to represent the history of transitions observed throughout training in the context of IRL. This replay buffer allows DAC to perform off-policy training of the policy and the discriminator (similar to (Sasaki et al., 2019)). Formally, DAC optimizes its discriminator with the objective:
(2)
$$\max_{D}\mathbb{E}_{s,a\sim\mathcal{R}}\left[\log D(s,a)\right]+\mathbb{E}_{s,a\sim\pi^{e}}\left[\log(1-D(s,a))\right].$$
where this objective minimize the divergence between the expert distribution and the replay buffer $\mathcal{R}$ distribution. Intuitively, this divergence does not strictly capture the divergence of our policy distribution and the expert distribution, but a mixture of evenly weighted policies learned up until the current policy. To rigorously recover a divergence between $\begin{array}{l}\mathrm{our~policy~distribution~and~the~expert~distribution~we~need~to~apply~importance~weights:}\\\mathrm{min~max~}_{D}\mathbb{E}_{s,a\sim\mathcal{R}}\left[\frac{p_{\pi}(s,a)}{p_{\mathcal{R}}(s,a)}\log D(s,a)\right]+\mathbb{E}_{s,a\sim\pi^{e}}\left[\log(1-D(s,a))\right]-\lambda H(\pi).\text{While this objec-}\end{array}$ tive recovers the $\bar on-policy$ objective of GAIL (Equation (1)), the authors note that estimating the density ratio is difficult and has high variance in practice. Furthermore, they note that the not using importance weights (Equation (2)) works well in practice, but does not guarantee successful $imitation$,especially when the distribution induced by the replay buffer, $\mathcal{R}$,is far from our current policy's state-action distribution. This is a fundamental problem of DAC.
## 3.3 VALUEDICE
ValueDICE (Kostrikov et al., 2020) was proposed to address the density estimation issue of offpolicy AIL algorithms formalized in DAC (see section 3.2). ValueDICE aims to minimize the reverse KL divergence written in its Donsker-Varadhan (Donsker&Varadhan, 1983) dual form:
(3)
$$-\mathrm{KL}(d^\pi||d^{\pi_e})=\min_{x:\mathcal{S}\times\mathcal{A}\mapsto\mathbb{R}}\log\mathbb{E}_{(s,a)\sim d^{\pi^e}}[e^{x(s,a)}]-\mathbb{E}_{(s,a)\sim d^{\pi}}[x(s,a)]$$
Motivated from DualDICE (Nachum et al., 2019), ValueDICE performs a change of variable using the Bellman operator $\mathcal{B}^{\pi2}$ with respect to the policy $\pi;x(s,a)=\nu(s,a)-\mathcal{B}^\pi(s,a);$ resulting
$^2$A bellman operator $\mathcal{B}^\pi$ is defined as follows: given any function $f(s,a)$,we have $\mathcal{B}^\pi f(s,a):=r(s,a)+$
$\mathbb{E}_{s^{\prime}\sim P(s,a)}f(s^{\prime},\tilde{\pi}(s^{\prime}),\forall s,a.$
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.176513671875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11871337890625,
0.176513671875,
0.11871337890625
],
[
0.1748046875,
0.11871337890625,
0.82275390625,
0.11871337890625,
0.82275390625,
0.1319580078125,
0.1748046875,
0.1319580078125
],
[
0.1748046875,
0.1331787109375,
0.82275390625,
0.1331787109375,
0.82275390625,
0.146484375,
0.1748046875,
0.146484375
],
[
0.1732177734375,
0.146484375,
0.82275390625,
0.1458740234375,
0.82275390625,
0.1590576171875,
0.1732177734375,
0.160400390625
],
[
0.1748046875,
0.160400390625,
0.826171875,
0.160400390625,
0.826171875,
0.173583984375,
0.1748046875,
0.173583984375
],
[
0.1748046875,
0.173583984375,
0.82421875,
0.173583984375,
0.82421875,
0.1881103515625,
0.1748046875,
0.1881103515625
],
[
0.1732177734375,
0.1881103515625,
0.42724609375,
0.1881103515625,
0.42724609375,
0.2020263671875,
0.1732177734375,
0.2020263671875
],
[
0.26708984375,
0.222900390625,
0.73046875,
0.222900390625,
0.73046875,
0.240478515625,
0.26708984375,
0.240478515625
],
[
0.8037109375,
0.2247314453125,
0.82666015625,
0.2247314453125,
0.82666015625,
0.238037109375,
0.8037109375,
0.238037109375
],
[
0.1732177734375,
0.251953125,
0.82275390625,
0.251953125,
0.82275390625,
0.26513671875,
0.1732177734375,
0.26513671875
],
[
0.176513671875,
0.267578125,
0.82177734375,
0.267578125,
0.82177734375,
0.27783203125,
0.176513671875,
0.27783203125
],
[
0.176513671875,
0.281005859375,
0.82177734375,
0.281005859375,
0.82177734375,
0.291748046875,
0.176513671875,
0.291748046875
],
[
0.1732177734375,
0.292236328125,
0.82275390625,
0.292236328125,
0.82275390625,
0.30615234375,
0.1732177734375,
0.30615234375
],
[
0.1748046875,
0.306884765625,
0.826171875,
0.306884765625,
0.826171875,
0.32080078125,
0.1748046875,
0.32080078125
],
[
0.176513671875,
0.3212890625,
0.82177734375,
0.3212890625,
0.82177734375,
0.33203125,
0.176513671875,
0.33203125
],
[
0.1748046875,
0.335205078125,
0.82275390625,
0.335205078125,
0.82275390625,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1748046875,
0.348388671875,
0.82421875,
0.348388671875,
0.82421875,
0.3623046875,
0.1748046875,
0.3623046875
],
[
0.1748046875,
0.363037109375,
0.471435546875,
0.363037109375,
0.471435546875,
0.376953125,
0.1748046875,
0.376953125
],
[
0.1732177734375,
0.393310546875,
0.483642578125,
0.39453125,
0.483642578125,
0.4091796875,
0.1732177734375,
0.40771484375
],
[
0.1748046875,
0.421630859375,
0.82421875,
0.421630859375,
0.82421875,
0.43505859375,
0.1748046875,
0.43505859375
],
[
0.176513671875,
0.4375,
0.82177734375,
0.4375,
0.82177734375,
0.447509765625,
0.176513671875,
0.447509765625
],
[
0.176513671875,
0.449462890625,
0.82421875,
0.449462890625,
0.82421875,
0.46337890625,
0.176513671875,
0.46337890625
],
[
0.1748046875,
0.46337890625,
0.826171875,
0.46337890625,
0.826171875,
0.4765625,
0.1748046875,
0.4765625
],
[
0.1748046875,
0.484130859375,
0.82177734375,
0.484130859375,
0.82177734375,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.1748046875,
0.498779296875,
0.82275390625,
0.498779296875,
0.82275390625,
0.51220703125,
0.1748046875,
0.51220703125
],
[
0.1748046875,
0.51220703125,
0.82421875,
0.51220703125,
0.82421875,
0.52587890625,
0.1748046875,
0.52587890625
],
[
0.1748046875,
0.5263671875,
0.2646484375,
0.5263671875,
0.2646484375,
0.53857421875,
0.1748046875,
0.53857421875
],
[
0.316162109375,
0.54150390625,
0.685546875,
0.54150390625,
0.685546875,
0.55615234375,
0.316162109375,
0.55615234375
],
[
0.8037109375,
0.54150390625,
0.82666015625,
0.54150390625,
0.82666015625,
0.55615234375,
0.8037109375,
0.55615234375
],
[
0.1748046875,
0.56640625,
0.82275390625,
0.56640625,
0.82275390625,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.17236328125,
0.5791015625,
0.3349609375,
0.58056640625,
0.3349609375,
0.5947265625,
0.17236328125,
0.59228515625
],
[
0.340576171875,
0.58056640625,
0.82275390625,
0.58056640625,
0.82275390625,
0.5947265625,
0.340576171875,
0.5947265625
],
[
0.1732177734375,
0.59521484375,
0.82177734375,
0.59521484375,
0.82177734375,
0.60595703125,
0.1732177734375,
0.60595703125
],
[
0.1748046875,
0.609375,
0.82275390625,
0.609375,
0.82275390625,
0.62255859375,
0.1748046875,
0.62255859375
],
[
0.1732177734375,
0.6220703125,
0.82666015625,
0.62255859375,
0.82666015625,
0.63720703125,
0.1732177734375,
0.63623046875
],
[
0.17236328125,
0.63720703125,
0.2900390625,
0.6396484375,
0.2900390625,
0.6572265625,
0.17236328125,
0.65478515625
],
[
0.294189453125,
0.6396484375,
0.82861328125,
0.63525390625,
0.82861328125,
0.6572265625,
0.294189453125,
0.6611328125
],
[
0.1748046875,
0.6591796875,
0.82275390625,
0.6591796875,
0.82275390625,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1748046875,
0.673828125,
0.82275390625,
0.673828125,
0.82275390625,
0.6875,
0.1748046875,
0.6875
],
[
0.1748046875,
0.68798828125,
0.82421875,
0.68798828125,
0.82421875,
0.7021484375,
0.1748046875,
0.7021484375
],
[
0.17236328125,
0.70068359375,
0.82275390625,
0.70068359375,
0.82275390625,
0.7138671875,
0.17236328125,
0.7138671875
],
[
0.1732177734375,
0.71533203125,
0.65283203125,
0.71533203125,
0.65283203125,
0.72900390625,
0.1732177734375,
0.72900390625
],
[
0.1732177734375,
0.7470703125,
0.30810546875,
0.7470703125,
0.30810546875,
0.76123046875,
0.1732177734375,
0.76123046875
],
[
0.17724609375,
0.77587890625,
0.82421875,
0.77587890625,
0.82421875,
0.7861328125,
0.17724609375,
0.7861328125
],
[
0.1732177734375,
0.78857421875,
0.82421875,
0.787109375,
0.82421875,
0.8017578125,
0.1732177734375,
0.80322265625
],
[
0.1732177734375,
0.8017578125,
0.80126953125,
0.8017578125,
0.80126953125,
0.81494140625,
0.1732177734375,
0.81494140625
],
[
0.26708984375,
0.8271484375,
0.3857421875,
0.8271484375,
0.3857421875,
0.84130859375,
0.26708984375,
0.84130859375
],
[
0.43310546875,
0.82470703125,
0.73779296875,
0.82470703125,
0.73779296875,
0.84521484375,
0.43310546875,
0.84521484375
],
[
0.8037109375,
0.82763671875,
0.826171875,
0.82763671875,
0.826171875,
0.84130859375,
0.8037109375,
0.84130859375
],
[
0.381591796875,
0.83837890625,
0.4453125,
0.83837890625,
0.4453125,
0.8486328125,
0.381591796875,
0.8486328125
],
[
0.39794921875,
0.8291015625,
0.435546875,
0.8291015625,
0.435546875,
0.8388671875,
0.39794921875,
0.8388671875
],
[
0.1748046875,
0.85693359375,
0.82275390625,
0.85693359375,
0.82275390625,
0.8720703125,
0.1748046875,
0.8720703125
],
[
0.1748046875,
0.8720703125,
0.82421875,
0.8720703125,
0.82421875,
0.88623046875,
0.1748046875,
0.88623046875
],
[
0.19287109375,
0.89599609375,
0.82666015625,
0.896484375,
0.82666015625,
0.9140625,
0.19287109375,
0.91357421875
],
[
0.17236328125,
0.9091796875,
0.351318359375,
0.9111328125,
0.351318359375,
0.9287109375,
0.17236328125,
0.9267578125
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2024",
"learning algorithm that connects Generative Adversarial Networks (GANs) (Goodfellow et al.",
"2014) and maximum entropy IRL (Ziebart et al., 2008). GAIL trains a binary classifier called the",
"discriminator D(s, a) to distinguish between samples from the expert distribution and the policy",
"generated distribution.Using the discriminator to define a reward function, GAIL then executes an",
"on-policy RL algorithm such as Trust Region Policy Optimization (TRPO) (Schulman et al., 2017a)",
"or Proximal Policy Optimization (PPO) (Schulman et al., 2017b) to maximize the reward. That gives",
"us the following adversarial objective:",
"min max Es,a~元 [log D(s,a)] + Es,a~ne [log(1 - D(s,a)] - 入H(π)",
"(1)",
"where H() is an entropy regularization term. The first term in eq. (1) can be viewed as a pseudo",
"rewardthatcanbeoptimizedwithrespecttothethepolicyπon-policysamples.NotethatGAlL",
"typicallyoptimizesbothpoliciesanddiscriminatorsusing on-policy samples,makingitquitesample",
"inefficient. Using different divergences, there are various reward functions that can be optimized",
"with this framework (Orsini et al., 2021). In this work, while our proposed approach in general is",
"capableof optimizingmanycommondivergences.wemainlyfocusonreverseKLdivergenceinoul",
"experiments. Reverse KL divergence has been studied in prior works including Fu et al. (2018); Ke",
"et al. (2020).But different from prior works, we propose an off-policy method for optimizing reverse",
"KL by leveraging the framework of boosting.",
"3.2 DISCRIMINATORACTOR CRITIC (DAC)",
"One reason GAIL need a lot of interactions with the environment to learn properly is because of the de-",
"pendencyonusingon-policyapproachestooptimizediscriminatorsandpolicies.InparticularGAIL",
"does not reuse any old samples. Discriminator Actor Critic (DAC) (Kostrikov et al., 2019) extends",
"GAIL algorithms to take advantage of off-policy learning to optimize the discriminators and policies.",
"DAc introduces a replay buffer R to represent the history of transitions observed throughout training",
"in the context of IRL. This replay buffer allows DAC to perform off-policy training of the policy and",
"the discriminator (similar to (Sasaki et al., 2019). Formally, DAC optimizes its discriminator with",
"the objective:",
"max Es,a~R [log D(s,a)] + Es,a~πe [log(1 -D(s,a))] .",
"(2)",
"where this objective minimize the divergence between the expert distribution and the replay",
"buffer R distribution.",
"Intuitively,this divergence does not strictly capture the divergence",
"of our policy distribution and the expert distribution,but a mixture of evenly weighted",
"policies learned up until the current policy.To rigorously recover a divergence between",
"our policy distribution and the expert distribution we need to apply importance weights:",
"min max Es,a~R",
"[P( log D(s,a)] + Es,a~e[log(1 - D(s,a)] -入H(), While this objec-",
"tive recovers the on-policy objective of GAIL (Equation (1)), the authors note that estimating the",
"density ratio is difficult and has high variance in practice. Furthermore, they note that the not",
"using importance weights (Equation (2) works well in practice, but does not guarantee successful",
"imitation, especially when the distribution induced by the replay buffer, R, is far from our current",
"policy's state-action distribution. This is a fundamental problem of DAC.",
"3.3VALUEDICE",
"ValueDIcE(Kostrikovetal.,2020)wasproposedtoaddressthedensityestimationissueofoff-",
"policy AIL algorithms formalized in DAc (see section 3.2). ValueDICE aims to minimize the",
"reverse KL divergence written in its Donsker-Varadhan (Donsker & Varadhan, 1983) dual form:",
"-KL(d\"dre) =",
"p log E(s,a) ~dre [e(s;a)] - E(s,a)~d\" [r(s, a)] ",
"(3)",
"TSXA-",
"min.",
"Motivated from DualDICE (Nachum et al., 2019), ValueDICE performs a change of variable",
"using the Bellman operator B2 with respect to the policy π; r(s, a) = v(s, a) - B* (s, a); resulting",
"2A bellman operator B\" is defined as follows: given any function f(s, a), we have B\" f(s, a) := r(s, a) +",
"Eg~P(s,a)f(s',(s), Vs, a.",
"4"
] |
{
"arxiv_id": "2404.08513",
"language": "en",
"page_idx": 3
}
|
||
2404.08513-en-4
|
2404.08513-en-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
the following objective:
(4)
$$\max_{\pi}\min_{\nu:\mathcal{S}\times\mathcal{A}\to\mathbb{R}}\log\mathbb{E}_{s,a\sim\pi^{e}}\left[\exp\left(\nu(s,a)-\mathcal{B}^{\pi}\nu(s,a)\right)\right]-(1-\gamma)\mathbb{E}s_{0}\sim\mu_{0},\left[\nu(s_{0},a_{0})\right].$$
Now the objective function does not contain on-policy distribution $d^\pi$ (in fact only the initial state distribution $\mu_0$ and the expert distribution). Despite being able to only using $d^{\pi^e}$ and $\mu_0$,the authors have identified two aspects of the objective that will yield biased estimates. First, the first expectation has a logarithm outside of it which would make mini-batch estimates of this expectation biased. Moreover, inside the fırst expectation term, we have $\nu(s,a)-\mathcal{B}^\pi\nu(s,a)$ with $\mathcal{B}^\pi$ being the Bellman operator. This limits ValueDICE's objective to only be unbiased for environments with deterministic transitions. This is related to the famous double sampling issue in TD learning. Although many popular RL benchmarks have deterministic transitions (Bellemare et al., 2013; Tassa et al., 2018; Todorov et al., 2012), this was a limitation not present in the GAIL.
In this work, we take a different perspective than ValueDICE to derive an off-policy AIL algorithm. Different from ValueDICE, our approach is both off-policy and is amenable to mini-batch updates even with stochastic environment transition dynamics.
## 4 ALGORITHM
Our algorithm, Adversarial Imitation Learning via Boosting (AILBoost)- motivated by classic gra- $dient\textit{ boosting algorithms ( Friedman, 2001; Mason et al. , 1999) - attempts to mitigate a fundamental}$ issue related to off-policy imitation learning formalized in DAC (see section 3.2). The key idea is to treat learned policies as weak learners, form an ensemble of them (with a proper weighting scheme derived from a gradient boosting perspective), and update the ensemble via gradient boosting.
Weighted policy ensemble. Our algorithm will learn a weighted ensemble of policies, denoted as $\pi:=\{\alpha_{i},\pi_{i}\}_{i=1}^{n}$ with $\alpha_{i}\geq0,\sum_{i}\alpha_{i}=1$ and $\pi_{i}$ being some Markovian policy. The way the mixture works is that when executing $\pi$, at the beginning of an episode, a Markovian policy $\pi_i$ is sampled with probability $\alpha_i$, and then $\pi_i$ is executed for the entire episode (i.e., no policy switch in an episode). Note that $\pi$ itself is not a Markovian policy anymore due to the sampling process at the beginning of the episode, and in fact, such mixture policy's induced state-action distribution can be richer than that from Markovian policies (Hazan et al., 2019). This is consistent with the idea of $boosting:$ by combining weak learners, i.e., Markovian policies, we form a more powerful policy. Given the above definition of $\pi$,we immediately have $d^{\pi}:=\sum_i\alpha_id^{\pi_i}$,i.e.,the weighted mixture of the state-action distributions induced by Markovian policies $\pi_i.$
Notation wise, given a dataset $\mathcal{D}$, we denote $\hat{\mathbb{E} } _{\mathcal{D} }[ f( x) ]$ as the empirical function average across the
dataset, i.e., $\widehat{\mathbb{E}}_{\mathcal{D}}[f(x)]=\sum_{x\in\mathcal{D}}f(x)/|\mathcal{D}|.$
4.1 AILBoosт: ADVERSARIAL IMITATION LEARNING VIA BOOSTING
We would like to minimize the reverse KL divergence between our policy state-action distribution $d^\pi$ and the expert distribution $d^{\pi^e}-$ denoted by $\ell(d^{\pi},d^{\pi^e})^{\prime}=$ KL$(d^{\pi}||d^{\pi^e}):=$ $\sum_{s,a}d^{\boldsymbol{\pi}}(s,a)\ln(d^{\boldsymbol{\pi}}(s,a)/d^{\pi^{e}}(s,a)).$ The reasons that we focus on reverse KL is that (1) it has been argued that the mode seeking property of reverse KL is more suitable for imitation learning (Ke et al., 2020), (2) reverse KL is on-policy in nature, i.e., it focuses on minimizing the divergence of our policy's action distribution and the expert's at the states from our policy, which help address the covariate shift issue, and (3) the baselines we consider in experiments, DAC and ValueDICE, all minimize the reverse KL divergence such as AIRL in practice 3. At a high level, our approach directly optimizes $\ell(d^\pi,d^{\pi^e})$ via gradient boosting (Mason et al., 1999) in the state-action occupancy space. Our ensemble $\pi$ induces the following mixture state-action occupancy measure:
$$d^\pi:=\sum_{i=1}^t\alpha_id^{\pi_i},\alpha_i\ge0.$$
To compute a new weak learner $\pi_{t+1}$, we will first compute the functional gradient of loss $\ell$ with respect to $d^\pi$, i.e., $\nabla\ell(d,d^{\pi^e})|_{d=d^\pi}.$ The new weak learner $\pi_t+1$ is learned via the following
See the official repository
5
|
[
[
0.1748046875,
0.0328369140625,
0.48046875,
0.0328369140625,
0.48046875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1732177734375,
0.10479736328125,
0.330810546875,
0.10479736328125,
0.330810546875,
0.11932373046875,
0.1732177734375,
0.11932373046875
],
[
0.3095703125,
0.1256103515625,
0.779296875,
0.1256103515625,
0.779296875,
0.14013671875,
0.3095703125,
0.14013671875
],
[
0.80126953125,
0.126953125,
0.826171875,
0.126953125,
0.826171875,
0.1395263671875,
0.80126953125,
0.1395263671875
],
[
0.265625,
0.1275634765625,
0.3056640625,
0.1275634765625,
0.3056640625,
0.1383056640625,
0.265625,
0.1383056640625
],
[
0.221435546875,
0.1275634765625,
0.25732421875,
0.1275634765625,
0.25732421875,
0.1395263671875,
0.221435546875,
0.1395263671875
],
[
0.23779296875,
0.136962890625,
0.31787109375,
0.136962890625,
0.31787109375,
0.147705078125,
0.23779296875,
0.147705078125
],
[
0.1748046875,
0.15283203125,
0.82275390625,
0.15283203125,
0.82275390625,
0.1666259765625,
0.1748046875,
0.1666259765625
],
[
0.1732177734375,
0.1666259765625,
0.82275390625,
0.1666259765625,
0.82275390625,
0.179931640625,
0.1732177734375,
0.179931640625
],
[
0.1732177734375,
0.179931640625,
0.82275390625,
0.179931640625,
0.82275390625,
0.1932373046875,
0.1732177734375,
0.1932373046875
],
[
0.176513671875,
0.1962890625,
0.82421875,
0.1962890625,
0.82421875,
0.20703125,
0.176513671875,
0.20703125
],
[
0.1748046875,
0.208984375,
0.826171875,
0.208984375,
0.826171875,
0.222900390625,
0.1748046875,
0.222900390625
],
[
0.1748046875,
0.222900390625,
0.82421875,
0.222900390625,
0.82421875,
0.236083984375,
0.1748046875,
0.236083984375
],
[
0.1732177734375,
0.236083984375,
0.82421875,
0.2374267578125,
0.82421875,
0.250732421875,
0.1732177734375,
0.25
],
[
0.1732177734375,
0.25,
0.826171875,
0.248779296875,
0.826171875,
0.26318359375,
0.1732177734375,
0.264404296875
],
[
0.1748046875,
0.264404296875,
0.615234375,
0.264404296875,
0.615234375,
0.27783203125,
0.1748046875,
0.27783203125
],
[
0.17236328125,
0.2841796875,
0.82275390625,
0.2841796875,
0.82275390625,
0.29736328125,
0.17236328125,
0.29736328125
],
[
0.1748046875,
0.298583984375,
0.82421875,
0.298583984375,
0.82421875,
0.3125,
0.1748046875,
0.3125
],
[
0.1732177734375,
0.313232421875,
0.53125,
0.313232421875,
0.53125,
0.326904296875,
0.1732177734375,
0.326904296875
],
[
0.1748046875,
0.348388671875,
0.2034912109375,
0.348388671875,
0.2034912109375,
0.35986328125,
0.1748046875,
0.35986328125
],
[
0.200927734375,
0.352783203125,
0.2059326171875,
0.352783203125,
0.2059326171875,
0.35791015625,
0.200927734375,
0.35791015625
],
[
0.2017822265625,
0.348388671875,
0.31201171875,
0.348388671875,
0.31201171875,
0.3623046875,
0.2017822265625,
0.3623046875
],
[
0.1732177734375,
0.376953125,
0.826171875,
0.37744140625,
0.826171875,
0.392578125,
0.1732177734375,
0.391357421875
],
[
0.1748046875,
0.392578125,
0.82275390625,
0.392578125,
0.82275390625,
0.406005859375,
0.1748046875,
0.406005859375
],
[
0.176513671875,
0.40771484375,
0.82275390625,
0.40771484375,
0.82275390625,
0.41845703125,
0.176513671875,
0.41845703125
],
[
0.1748046875,
0.42041015625,
0.82421875,
0.42041015625,
0.82421875,
0.434326171875,
0.1748046875,
0.434326171875
],
[
0.17236328125,
0.43310546875,
0.78759765625,
0.434326171875,
0.78759765625,
0.448974609375,
0.17236328125,
0.447509765625
],
[
0.1748046875,
0.455078125,
0.82421875,
0.455078125,
0.82421875,
0.4697265625,
0.1748046875,
0.4697265625
],
[
0.1748046875,
0.4697265625,
0.82421875,
0.4697265625,
0.82421875,
0.48291015625,
0.1748046875,
0.48291015625
],
[
0.1748046875,
0.484130859375,
0.82421875,
0.484130859375,
0.82421875,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.17236328125,
0.496826171875,
0.82177734375,
0.496826171875,
0.82177734375,
0.51025390625,
0.17236328125,
0.51025390625
],
[
0.1748046875,
0.51123046875,
0.82421875,
0.51123046875,
0.82421875,
0.5244140625,
0.1748046875,
0.5244140625
],
[
0.1748046875,
0.5244140625,
0.82275390625,
0.5244140625,
0.82275390625,
0.53857421875,
0.1748046875,
0.53857421875
],
[
0.1748046875,
0.53857421875,
0.826171875,
0.53857421875,
0.826171875,
0.5517578125,
0.1748046875,
0.5517578125
],
[
0.1748046875,
0.55322265625,
0.826171875,
0.55322265625,
0.826171875,
0.56640625,
0.1748046875,
0.56640625
],
[
0.1748046875,
0.56640625,
0.826171875,
0.56640625,
0.826171875,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.17236328125,
0.5791015625,
0.58837890625,
0.580078125,
0.58837890625,
0.5947265625,
0.17236328125,
0.59326171875
],
[
0.17236328125,
0.60107421875,
0.826171875,
0.6015625,
0.826171875,
0.619140625,
0.17236328125,
0.61865234375
],
[
0.1732177734375,
0.619140625,
0.459228515625,
0.619140625,
0.459228515625,
0.63720703125,
0.1732177734375,
0.63720703125
],
[
0.1732177734375,
0.6513671875,
0.67333984375,
0.6513671875,
0.67333984375,
0.66552734375,
0.1732177734375,
0.66552734375
],
[
0.1748046875,
0.6767578125,
0.826171875,
0.6767578125,
0.826171875,
0.69140625,
0.1748046875,
0.69140625
],
[
0.1732177734375,
0.69140625,
0.79833984375,
0.68994140625,
0.79833984375,
0.705078125,
0.1732177734375,
0.70556640625
],
[
0.79736328125,
0.6962890625,
0.81787109375,
0.6962890625,
0.81787109375,
0.70068359375,
0.79736328125,
0.70068359375
],
[
0.176513671875,
0.70703125,
0.82275390625,
0.70703125,
0.82275390625,
0.7216796875,
0.176513671875,
0.7216796875
],
[
0.1748046875,
0.7216796875,
0.82421875,
0.7216796875,
0.82421875,
0.736328125,
0.1748046875,
0.736328125
],
[
0.176513671875,
0.7373046875,
0.82275390625,
0.7373046875,
0.82275390625,
0.74755859375,
0.176513671875,
0.74755859375
],
[
0.1748046875,
0.75048828125,
0.82421875,
0.75048828125,
0.82421875,
0.7646484375,
0.1748046875,
0.7646484375
],
[
0.1748046875,
0.7646484375,
0.82421875,
0.7646484375,
0.82421875,
0.77783203125,
0.1748046875,
0.77783203125
],
[
0.1748046875,
0.77783203125,
0.82421875,
0.77783203125,
0.82421875,
0.79150390625,
0.1748046875,
0.79150390625
],
[
0.1732177734375,
0.79052734375,
0.82421875,
0.79150390625,
0.82421875,
0.80615234375,
0.1732177734375,
0.8046875
],
[
0.1732177734375,
0.80615234375,
0.74609375,
0.80615234375,
0.74609375,
0.8193359375,
0.1732177734375,
0.8193359375
],
[
0.414306640625,
0.83837890625,
0.583984375,
0.83837890625,
0.583984375,
0.85302734375,
0.414306640625,
0.85302734375
],
[
0.463134765625,
0.85693359375,
0.483642578125,
0.85693359375,
0.483642578125,
0.86572265625,
0.463134765625,
0.86572265625
],
[
0.1748046875,
0.8720703125,
0.82275390625,
0.8720703125,
0.82275390625,
0.88623046875,
0.1748046875,
0.88623046875
],
[
0.1707763671875,
0.8837890625,
0.826171875,
0.88525390625,
0.826171875,
0.90283203125,
0.1707763671875,
0.9013671875
],
[
0.1953125,
0.9091796875,
0.3544921875,
0.9111328125,
0.3544921875,
0.92529296875,
0.1953125,
0.923828125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2024",
"the following objective:",
"logEs,a~元e [exp(v(s,a) -B\"v(s, a)] -(1 -)Eso~μo, [v(so, ao)] ",
"(4)",
"min",
"max",
"VSXA-R",
"Now the objective function does not contain on-policy distribution d (in fact only the initial state",
"distribution μo and the expert distribution). Despite being able to only using dand μo, the authors",
"have identified two aspects of the objective that will yield biased estimates. First, the first expectation",
"has a logarithm outside of itwhich would makemini-batch estimates of this expectation biased",
"Moreover, inside the first expectation term, we have v(s, a) - B v(s, a) with B\" being the Bellman",
"operator. This limits ValueD ICE's objective to only be unbiased for environments with deterministic",
"transitions. This is related to the famous double sampling issue in TD learning. Although many",
"popular RL benchmarks have deterministic transitions (Bellemare et al., 2013; Tassa et al., 2018;",
"Todorov et al., 2012), this was a limitation not present in the GAI L.",
"In this work, we take a different perspective than ValueDICE to derive an off-policy AIL algorithm",
"Different from ValueDICE, our approach is both off-policy and is amenable to mini-batch updates",
"even with stochastic environment transition dynamics.",
"4",
"",
"ALGORITHM",
"Our algorithm,Adversarial Imitation Learning via Boosting (AILBoost)-motivated by classic gra-",
"dient boosting algorithms (Friedman, 2001; Mason et al., 1999) -- attempts to mitigate a fundamental",
"issuerelatedtooff-policyimitationlearningformalizedinDAC(seesection3.2).Thekeyideaisto",
"treat learned policies as weak learners, form an ensemble of them (with a proper weighting scheme",
"derived from a gradient boosting perspective), and update the ensemble via gradient boosting.",
"Weighted policy ensemble. Our algorithm will learn a weighted ensemble of policies, denoted",
"as π := {αi, iln- with i ≥ O, , Qi = 1 and being some Markovian policy. The way the",
"mixture works is that when executing π, at the beginning of an episode, a Markovian policy , is",
"sampled with probability Qi, and then , is executed for the entire episode (i.e., no policy switch in",
"an episode). Note that itself is not a Markovian policy anymore due to the sampling process at",
"the beginning of the episode, and in fact, such mixture policy's induced state-action distribution can",
"be richer than that from Markovian policies (Hazan et al., 2019). This is consistent with the idea of",
"boosting: by combining weak learners, i.e., Markovian policies, we form a more powerful policy.",
"Given the above definition of , we immediately have d :=-Z, Qd, i.e., the weighted mixture of",
"the state-action distributions induced by Markovian policies Ti.",
"Notation wise, given a dataset D, we denote Ep[f(r)] as the empirical function average across the",
"dataset, i.e., Ep[f(a)] =ZreD f(a)/ [D].",
"4.1 AILBOOST:ADVERSARIAL IMITATION LEARNING VIA BOOSTING",
"We would like to minimize the reverse KL divergence between our policy state-action dis-",
"tribution dand the expert distribution dr-denoted bye(d\",d)=KL(dd)",
"",
"Zs.a d* (s, a) ln(d(s, a)/d (s, a). The reasons that we focus on reverse KL is that (1) it has",
"been argued that the mode seeking property of reverse KL is more suitable for imitation learning",
"(Ke etal..2020).(2)reverseKLis on-policyin nature.i.e.,itfocusesonminimizing thedivergence",
"of our policy's action distribution and the expert's at the states from our policy, which help address",
"the covariate shift issue, and (3) the baselines we consider in experiments, DAC and ValueDICE,",
"all minimize the reverse KL divergence such as AIRL in practice 3, At a high level, our approach",
"directly optimizes e(d, d) via gradient boosting (Mason et al., 1999) in the state-action occupancy",
"space. Our ensemble π induces the following mixture state-action occupancy measure:",
"d :=/aidi,ai≥ 0.",
"i=1",
"To compute a new weak learner t+1, we will first compute the functional gradient of loss l with",
"respect to d, i.e., Ve(d, d\")lddw. The new weak learner t+1 is learned via the following",
"3See the official repository",
"5"
] |
{
"arxiv_id": "2404.08513",
"language": "en",
"page_idx": 4
}
|
||
1611.04717-en-0
|
1611.04717-en-0.jpg
|
# $\# \textbf{Exploration: A Study of Count- Based Exploration}$ for Deep Reinforcement Learning
Haoran Tang$^1*,\textbf{Rein Houthooft}^{34*},\textbf{Davis Foote}^2,\textbf{Adam Stooke}^2,\textbf{Xi Chen}^{2\dagger}$,
Yan Duan$^2\dagger,\textbf{John Schulman}^4,\textbf{Filip De Turck}^3,\textbf{Pieter Abbeel }^{2\dagger}$
$^{1}$UC Berkeley, Department of Mathematics
$^{2}$UC Berkeley, Department of Electrical Engineering and Computer Sciences
$\begin{array}{c}{{^{3}\text{Ghent University-imec, Department of Information Technology}}}\\{{^{4}\text{OpenAI}}}\end{array}$
## Abstract
Count-based exploration algorithms are known to perform near-optimally when used in conjunction with tabular reinforcement learning (RL) methods for solving small discrete Markov decision processes (MDPs). It is generally thought that count-based methods cannot be applied in high-dimensional state spaces, since most states will only occur once. Recent deep RL exploration strategies are able to deal with high-dimensional continuous state spaces through complex heuristics, often relying on optimismin the face of uncertainty or intrinsic motivation. In this work, we describe a surprising finding: a simple generalization of the classic count-based approach can reach near state-of-the-art performance on various highdimensional and/or continuous deep RL benchmarks. States are mapped to hash codes, which allows to count their occurrences with a hash table. These counts are then used to compute a reward bonus according to the classic count-based exploration theory. We find that simple hash functions can achieve surprisingly good results on many challenging tasks. Furthermore, we show that a domaindependent learned hash code may further improve these results. Detailed analysis reveals important aspects of a good hash function: 1) having appropriate granularity and 2) encoding information relevant to solving the MDP. This exploration strategy achieves near state-of-the-art performance on both continuous control tasks and Atari 2600 games, hence providing a simple yet powerful baseline for solving MDPs that require considerable exploration.
### 1 Introduction
Reinforcement learning (RL) studies an agent acting in an initially unknown environment, learning through trial and error to maximize rewards. It is impossible for the agent to act near-optimally until it has sufficiently explored the environment and identified all of the opportunities for high reward, in all scenarios. A core challenge in RL is how to balance exploration—actively seeking out novel states and actions that might yield high rewards and lead to long-term gains; and exploitation—maximizing short-term rewards using the agent's current knowledge. While there are exploration techniques for finite MDPs that enjoy theoretical guarantees, there are no fully satisfying techniques for highdimensional state spaces; therefore, developing more general and robust exploration techniques is an active area of research.
$^{*}$These authors contributed equally. Correspondence to: Haoran Tang <hrtang@ math.berkeley.edu>, Rein
Houthooft <rein.houthooft@ openai.com>
$^{\dagger}$Work done at OpenAI
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
|
[
[
0.188720703125,
0.125,
0.810546875,
0.1256103515625,
0.810546875,
0.146484375,
0.188720703125,
0.1458740234375
],
[
0.295654296875,
0.1514892578125,
0.70166015625,
0.15283203125,
0.70166015625,
0.1717529296875,
0.295654296875,
0.17041015625
],
[
0.226318359375,
0.2247314453125,
0.77392578125,
0.2235107421875,
0.77392578125,
0.2412109375,
0.226318359375,
0.242431640625
],
[
0.273681640625,
0.240478515625,
0.72119140625,
0.2392578125,
0.72119140625,
0.2568359375,
0.273681640625,
0.25830078125
],
[
0.3544921875,
0.255615234375,
0.64453125,
0.2568359375,
0.64453125,
0.271484375,
0.3544921875,
0.270751953125
],
[
0.01715087890625,
0.272705078125,
0.060455322265625,
0.272705078125,
0.063720703125,
0.406005859375,
0.0196075439453125,
0.406005859375
],
[
0.2418212890625,
0.267578125,
0.7548828125,
0.26953125,
0.7548828125,
0.287353515625,
0.2418212890625,
0.285400390625
],
[
0.27783203125,
0.2841796875,
0.71826171875,
0.285888671875,
0.71826171875,
0.300537109375,
0.27783203125,
0.298583984375
],
[
0.464111328125,
0.298583984375,
0.53369140625,
0.300537109375,
0.53369140625,
0.31494140625,
0.464111328125,
0.313232421875
],
[
0.461669921875,
0.35107421875,
0.5390625,
0.35107421875,
0.5390625,
0.365478515625,
0.461669921875,
0.365478515625
],
[
0.233642578125,
0.38330078125,
0.765625,
0.38330078125,
0.765625,
0.397705078125,
0.233642578125,
0.397705078125
],
[
0.2320556640625,
0.396484375,
0.765625,
0.397705078125,
0.765625,
0.412353515625,
0.2320556640625,
0.410888671875
],
[
0.021240234375,
0.404052734375,
0.060455322265625,
0.404052734375,
0.060455322265625,
0.486083984375,
0.021240234375,
0.486083984375
],
[
0.2344970703125,
0.41357421875,
0.765625,
0.41357421875,
0.765625,
0.423583984375,
0.2344970703125,
0.423583984375
],
[
0.233642578125,
0.425537109375,
0.765625,
0.425537109375,
0.765625,
0.439453125,
0.233642578125,
0.439453125
],
[
0.233642578125,
0.439453125,
0.765625,
0.439453125,
0.765625,
0.45263671875,
0.233642578125,
0.45263671875
],
[
0.233642578125,
0.453857421875,
0.76708984375,
0.453857421875,
0.76708984375,
0.46728515625,
0.233642578125,
0.46728515625
],
[
0.233642578125,
0.46728515625,
0.76708984375,
0.46728515625,
0.76708984375,
0.48095703125,
0.233642578125,
0.48095703125
],
[
0.021240234375,
0.48486328125,
0.058013916015625,
0.48486328125,
0.058013916015625,
0.70703125,
0.021240234375,
0.70703125
],
[
0.2320556640625,
0.48095703125,
0.7646484375,
0.48095703125,
0.7646484375,
0.494384765625,
0.2320556640625,
0.494384765625
],
[
0.233642578125,
0.494384765625,
0.7685546875,
0.494384765625,
0.7685546875,
0.50830078125,
0.233642578125,
0.50830078125
],
[
0.2320556640625,
0.5087890625,
0.765625,
0.5087890625,
0.765625,
0.52294921875,
0.2320556640625,
0.52294921875
],
[
0.233642578125,
0.521484375,
0.765625,
0.521484375,
0.765625,
0.53515625,
0.233642578125,
0.53515625
],
[
0.233642578125,
0.53515625,
0.765625,
0.53515625,
0.765625,
0.548828125,
0.233642578125,
0.548828125
],
[
0.233642578125,
0.5498046875,
0.765625,
0.5498046875,
0.765625,
0.56298828125,
0.233642578125,
0.56298828125
],
[
0.233642578125,
0.56298828125,
0.7685546875,
0.56298828125,
0.7685546875,
0.5771484375,
0.233642578125,
0.5771484375
],
[
0.2344970703125,
0.5791015625,
0.7646484375,
0.5791015625,
0.7646484375,
0.5888671875,
0.2344970703125,
0.5888671875
],
[
0.2320556640625,
0.59033203125,
0.76318359375,
0.59033203125,
0.76318359375,
0.60400390625,
0.2320556640625,
0.60400390625
],
[
0.2320556640625,
0.60400390625,
0.7646484375,
0.60498046875,
0.7646484375,
0.619140625,
0.2320556640625,
0.61865234375
],
[
0.233642578125,
0.61865234375,
0.765625,
0.61865234375,
0.765625,
0.6318359375,
0.233642578125,
0.6318359375
],
[
0.233642578125,
0.63330078125,
0.765625,
0.63330078125,
0.765625,
0.646484375,
0.233642578125,
0.646484375
],
[
0.2320556640625,
0.64599609375,
0.521484375,
0.64599609375,
0.521484375,
0.6591796875,
0.2320556640625,
0.6591796875
],
[
0.17724609375,
0.6923828125,
0.19287109375,
0.6923828125,
0.19287109375,
0.7021484375,
0.17724609375,
0.7021484375
],
[
0.200927734375,
0.69140625,
0.31201171875,
0.69140625,
0.31201171875,
0.70556640625,
0.200927734375,
0.70556640625
],
[
0.17236328125,
0.7216796875,
0.82421875,
0.72265625,
0.82421875,
0.7373046875,
0.17236328125,
0.736328125
],
[
0.1748046875,
0.7373046875,
0.82421875,
0.7373046875,
0.82421875,
0.75048828125,
0.1748046875,
0.75048828125
],
[
0.1732177734375,
0.751953125,
0.82421875,
0.751953125,
0.82421875,
0.76513671875,
0.1732177734375,
0.76513671875
],
[
0.1732177734375,
0.76513671875,
0.82421875,
0.76513671875,
0.82421875,
0.77880859375,
0.1732177734375,
0.77880859375
],
[
0.1732177734375,
0.77880859375,
0.82421875,
0.77978515625,
0.82421875,
0.79345703125,
0.1732177734375,
0.79248046875
],
[
0.1732177734375,
0.79150390625,
0.82421875,
0.79248046875,
0.82421875,
0.806640625,
0.1732177734375,
0.80615234375
],
[
0.17236328125,
0.8046875,
0.826171875,
0.80615234375,
0.826171875,
0.82080078125,
0.17236328125,
0.8193359375
],
[
0.1748046875,
0.82080078125,
0.82421875,
0.82080078125,
0.82421875,
0.833984375,
0.1748046875,
0.833984375
],
[
0.1732177734375,
0.833984375,
0.325927734375,
0.833984375,
0.325927734375,
0.84765625,
0.1732177734375,
0.84765625
],
[
0.19287109375,
0.85693359375,
0.82421875,
0.85791015625,
0.82421875,
0.87255859375,
0.19287109375,
0.8720703125
],
[
0.17236328125,
0.8720703125,
0.419189453125,
0.8720703125,
0.419189453125,
0.88525390625,
0.17236328125,
0.88525390625
],
[
0.193603515625,
0.8837890625,
0.3349609375,
0.88525390625,
0.3349609375,
0.8994140625,
0.193603515625,
0.8984375
],
[
0.1748046875,
0.923828125,
0.7509765625,
0.923828125,
0.7509765625,
0.9375,
0.1748046875,
0.9375
]
] |
[
"#Exploration: A Study of Count-Based Exploration",
"for Deep Reinforcement Learning",
"Haoran Tang *, Rein Houthooft34* , Davis Foote?, Adam Stooke2, Xi Chen?t,",
"Yan Duan?t, John Schulman4, Filip De Turck3, Pieter Abbeel 2",
"1 UC Berkeley, Department of Mathematics",
"1",
"2 UC Berkeley, Department of Electrical Engineering and Computer Sciences",
"3 Ghent University - imec, Department of Information Technology",
"4OpenAI",
"Abstract",
"Count-based exploration algorithms are known to perform near-optimally when",
"used in conjunction with tabular reinforcement learning (RL) methods for solving",
"[I']",
"smalldiscreteMarkovdecisionprocesses(MDPs).Itisgenerallythoughtthat",
"count-based methods cannot be applied in high-dimensional state spaces, since",
"most states will only occur once. Recent deep RL exploration strategies are able to",
"deal with high-dimensional continuous state spaces through complex heuristics,",
"often relying on optimism in the face of uncertainty or intrinsic motivation. In",
"CA0",
"this work, we describe a surprising finding:a simple generalization of the classic",
"count-based approach can reach near state-of-the-art performance on various high-",
"dimensional and/or continuous deep RL benchmarks. States are mapped to hash",
"codes, which allows to count their occurrences with a hash table. These counts",
"are then used to compute a reward bonus according to the classic count-based",
"exploration theory. We find that simple hash functions can achieve surprisingly",
"good results on many challenging tasks. Furthermore, we show that a domain-",
"dependentlearnedhashcodemayfurtherimprovetheseresults.Detailedanalysis",
"reveals important aspects of a good hash function: 1) having appropriate granularity",
"and 2) encoding information relevant to solving the MDP. This exploration strategy",
"achieves near state-of-the-art performance on both continuous control tasks and",
"Atari 2600 games, hence providing a simple yet powerful baseline for solving",
"MDPs that require considerable exploration",
"1",
"Introduction",
"Reinforcement learning (RL) studies an agent acting in an initially unknown environment, learning",
"through trial and error to maximize rewards. It is impossible for the agent to act near-optimally until",
"it has sufficiently explored the environment and identified all of the opportunities for high reward, in",
"all scenarios. A core challenge in RL is how to balance exploration-actively seeking out novel states",
"and actions that might yield high rewards and lead tolong-term gains; and exploitationmaximizing",
"short-term rewards using the agent's current knowledge. While there are exploration techniques",
"for finite MDPs that enjoy theoretical guarantees, there are no fully satisfying techniques for high-",
"dimensional state spaces; therefore, developing more general and robust exploration techniques is an",
"active area of research.",
"*These authors contributed equally. Correspondence to: Haoran Tang <hrtang@math.berkeley.edu>, Rein",
"Houthooft<rein.houthooft@openai.com>",
"fWorkdone at OpenAI",
"31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA"
] |
{
"arxiv_id": "1611.04717",
"language": "en",
"page_idx": 0
}
|
||
1611.04717-en-1
|
1611.04717-en-1.jpg
|
Most of the recent state-of-the-art RL results have been obtained using simple exploration strategies such as uniform sampling [21] and i.i.d./correlated Gaussian noise [19,30]. Although these heuristics are sufficient in tasks with well-shaped rewards, the sample complexity can grow exponentially (with state space size) in tasks with sparse rewards [25]. Recently developed exploration strategies for deep RL have led to significantly improved performance on environments with sparse rewards. Bootstrapped DQN[24] led to faster learning in a range of Atari 2600 games by training an ensemble of Q-functions. Intrinsic motivation methods using pseudo-counts achieve state-of-the-art performance on Montezuma's Revenge, an extremely challenging Atari 2600 game [4]. Variational Information Maximizing Exploration (VIME,[13]) encourages the agent to explore by acquiring information about environment dynamics, and performs well on various robotic locomotion problems with sparse rewards. However, we have not seen a very simple and fast method that can work across different domains.
Some of the classic, theoretically-justified exploration methods are based on counting state-action visitations, and turning this count into a bonus reward. In the bandit setting, the well-known UCB algorithm of [18] chooses the action $a_t$ at time $t$ that maximizes $\hat{r}(a_t)+\sqrt{\frac{2\log t}{n(a_t)}}$ where $\hat{r}(a_t)$ is the estimated reward, and $n(a_t)$ is the number of times action $a_t$ was previously chosen. In the MDP setting, some of the algorithms have similar structure, for example, Model Based Interval Estimation- $\tilde {\text{Exploration Bonus ( MBIE- EB) of[ 34] counts state- action pairs with a table }n( s, a) and}$ adding a bonus reward of the form $\frac\beta{\sqrt{n(s,a)}}$ to encourage exploring less visited pairs. [16] show that the inverse-square-root dependence is optimal. MBIE and related algorithms assume that the augmented MDP is solved analytically at each timestep, which is only practical for small finite state spaces.
This paper presents a simple approach for exploration, which extends classic counting-based methods to high-dimensional, continuous state spaces. We discretize the state space with a hash function and apply a bonus based on the state-visitation count. The hash function can be chosen to appropriately balance generalization across states, and distinguishing between states. We select problems from rllab [8] and Atari 2600[3] featuring sparse rewards, and demonstrate near state-of-the-art performance on several games known to be hard for naïve exploration strategies. The main strength of the presented approach is that it is fast, flexible and complementary to most existing RL algorithms.
In summary, this paper proposes a generalization of classic count-based exploration to highdimensional spaces through hashing (Section 2); demonstrates its effectiveness on challenging deep RL benchmark problems and analyzes key components of well-designed hash functions (Section 4).
# 2 Methodology
# 2.1 Notation
This paper assumes a finite-horizon discounted Markov decision process (MDP), defined by $(\mathcal{S},\mathcal{A},\mathcal{P},r,\rho_{0},\gamma,T)$,in which $\mathcal{S}$ is the state space, $\mathcal{A}$ the action space, $\mathcal{P}$ a transition probability distribution, $r:\mathcal{S}\times\mathcal{A}\to\mathbb{R}$ a reward function, $\rho_0$ an initial state distribution, $\gamma\in(0,1]$ a discount factor, and $T$ the horizon. The goal of RL is to maximize the total expected discounted reward$\mathbb{E} _{\pi , P}\left [ \sum _{t= 0}^{T}\gamma ^{t}r( s_{t}, a_{t}) \right ]$over a policy $\pi $,which outputs a distribution over actions given a state.
2.2 Count-Based Exploration via Static Hashing
Our approach discretizes the state space with a hash function $\phi:\mathcal{S}\to\mathbb{Z}.$ An exploration bonus
$r^{+}:\hat{\mathcal{S}}\to\mathbb{R}$ is added to the reward function, defined as
(1)
$$r^+(s)=\frac{\beta}{\sqrt{n(\phi(s))}},$$
where $\beta\in\mathbb{R}_{\geq0}$ is the bonus coefficient. Initially the counts $n(\cdot)$ are set to zero for the whole range of $\phi.$ For every state $s_t$ encountered at time step $t,n(\phi(s_t))$ is increased by one. The agent is trained with rewards $(r+r^+),w$hile performance is evaluated as the sum of rewards without bonuses
2
|
[
[
0.1748046875,
0.0921630859375,
0.82421875,
0.0921630859375,
0.82421875,
0.1060791015625,
0.1748046875,
0.1060791015625
],
[
0.1732177734375,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.11932373046875,
0.1732177734375,
0.11932373046875
],
[
0.1748046875,
0.12060546875,
0.82275390625,
0.12060546875,
0.82275390625,
0.1337890625,
0.1748046875,
0.1337890625
],
[
0.1748046875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.1748046875,
0.147705078125
],
[
0.1748046875,
0.147705078125,
0.82275390625,
0.147705078125,
0.82275390625,
0.1610107421875,
0.1748046875,
0.1610107421875
],
[
0.1748046875,
0.1610107421875,
0.82421875,
0.1610107421875,
0.82421875,
0.1749267578125,
0.1748046875,
0.1749267578125
],
[
0.1748046875,
0.175537109375,
0.82275390625,
0.175537109375,
0.82275390625,
0.189453125,
0.1748046875,
0.189453125
],
[
0.1748046875,
0.189453125,
0.82275390625,
0.189453125,
0.82275390625,
0.20263671875,
0.1748046875,
0.20263671875
],
[
0.1748046875,
0.20263671875,
0.82275390625,
0.20263671875,
0.82275390625,
0.216552734375,
0.1748046875,
0.216552734375
],
[
0.1748046875,
0.2183837890625,
0.82177734375,
0.2183837890625,
0.82177734375,
0.2291259765625,
0.1748046875,
0.2291259765625
],
[
0.1732177734375,
0.2310791015625,
0.82421875,
0.2310791015625,
0.82421875,
0.2442626953125,
0.1732177734375,
0.2442626953125
],
[
0.1748046875,
0.24560546875,
0.2344970703125,
0.24560546875,
0.2344970703125,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.17236328125,
0.26318359375,
0.82275390625,
0.26513671875,
0.82275390625,
0.279052734375,
0.17236328125,
0.276611328125
],
[
0.1748046875,
0.279052734375,
0.82275390625,
0.279052734375,
0.82275390625,
0.292236328125,
0.1748046875,
0.292236328125
],
[
0.1707763671875,
0.291748046875,
0.7236328125,
0.29541015625,
0.7236328125,
0.317626953125,
0.1707763671875,
0.313232421875
],
[
0.68115234375,
0.292236328125,
0.82666015625,
0.296630859375,
0.82666015625,
0.314453125,
0.68115234375,
0.31005859375
],
[
0.1748046875,
0.31640625,
0.82421875,
0.31640625,
0.82421875,
0.330078125,
0.1748046875,
0.330078125
],
[
0.1748046875,
0.330078125,
0.82275390625,
0.330078125,
0.82275390625,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.1748046875,
0.343505859375,
0.82421875,
0.343505859375,
0.82421875,
0.35791015625,
0.1748046875,
0.35791015625
],
[
0.1748046875,
0.3623046875,
0.41259765625,
0.3623046875,
0.41259765625,
0.37255859375,
0.1748046875,
0.37255859375
],
[
0.465576171875,
0.361083984375,
0.82421875,
0.35986328125,
0.82421875,
0.373779296875,
0.465576171875,
0.375
],
[
0.419189453125,
0.37060546875,
0.46728515625,
0.37060546875,
0.46728515625,
0.380615234375,
0.419189453125,
0.380615234375
],
[
0.1732177734375,
0.380126953125,
0.82421875,
0.380126953125,
0.82421875,
0.39453125,
0.1732177734375,
0.39453125
],
[
0.1732177734375,
0.393310546875,
0.82421875,
0.39453125,
0.82421875,
0.4091796875,
0.1732177734375,
0.40771484375
],
[
0.1748046875,
0.410888671875,
0.221435546875,
0.410888671875,
0.221435546875,
0.42236328125,
0.1748046875,
0.42236328125
],
[
0.1748046875,
0.4287109375,
0.82275390625,
0.4287109375,
0.82275390625,
0.442626953125,
0.1748046875,
0.442626953125
],
[
0.1748046875,
0.442626953125,
0.82275390625,
0.442626953125,
0.82275390625,
0.455810546875,
0.1748046875,
0.455810546875
],
[
0.176513671875,
0.45703125,
0.82421875,
0.45703125,
0.82421875,
0.47021484375,
0.176513671875,
0.47021484375
],
[
0.17724609375,
0.472900390625,
0.82177734375,
0.472900390625,
0.82177734375,
0.48291015625,
0.17724609375,
0.48291015625
],
[
0.176513671875,
0.484130859375,
0.82275390625,
0.484130859375,
0.82275390625,
0.49755859375,
0.176513671875,
0.49755859375
],
[
0.1748046875,
0.498779296875,
0.82421875,
0.498779296875,
0.82421875,
0.51220703125,
0.1748046875,
0.51220703125
],
[
0.1732177734375,
0.51220703125,
0.734375,
0.51220703125,
0.734375,
0.52587890625,
0.1732177734375,
0.52587890625
],
[
0.17236328125,
0.52978515625,
0.826171875,
0.53076171875,
0.826171875,
0.54541015625,
0.17236328125,
0.54443359375
],
[
0.1748046875,
0.546875,
0.82275390625,
0.546875,
0.82275390625,
0.56005859375,
0.1748046875,
0.56005859375
],
[
0.1732177734375,
0.56005859375,
0.826171875,
0.56005859375,
0.826171875,
0.57373046875,
0.1732177734375,
0.57373046875
],
[
0.1690673828125,
0.59228515625,
0.314453125,
0.59521484375,
0.314453125,
0.61279296875,
0.1690673828125,
0.60986328125
],
[
0.1732177734375,
0.62744140625,
0.275390625,
0.62744140625,
0.275390625,
0.642578125,
0.1732177734375,
0.642578125
],
[
0.1748046875,
0.65380859375,
0.82275390625,
0.65380859375,
0.82275390625,
0.66748046875,
0.1748046875,
0.66748046875
],
[
0.1732177734375,
0.66845703125,
0.826171875,
0.66845703125,
0.826171875,
0.68310546875,
0.1732177734375,
0.68310546875
],
[
0.1748046875,
0.68310546875,
0.82275390625,
0.68310546875,
0.82275390625,
0.6962890625,
0.1748046875,
0.6962890625
],
[
0.1748046875,
0.6962890625,
0.82421875,
0.6962890625,
0.82421875,
0.7109375,
0.1748046875,
0.7109375
],
[
0.1707763671875,
0.7119140625,
0.39208984375,
0.7119140625,
0.39208984375,
0.73291015625,
0.1707763671875,
0.73291015625
],
[
0.3857421875,
0.71533203125,
0.82421875,
0.71533203125,
0.82421875,
0.72900390625,
0.3857421875,
0.72900390625
],
[
0.1732177734375,
0.73193359375,
0.2091064453125,
0.73193359375,
0.2091064453125,
0.744140625,
0.1732177734375,
0.744140625
],
[
0.1748046875,
0.76318359375,
0.52294921875,
0.76318359375,
0.52294921875,
0.77783203125,
0.1748046875,
0.77783203125
],
[
0.1748046875,
0.79052734375,
0.82421875,
0.79052734375,
0.82421875,
0.8037109375,
0.1748046875,
0.8037109375
],
[
0.17236328125,
0.80322265625,
0.5400390625,
0.8037109375,
0.5400390625,
0.818359375,
0.17236328125,
0.8173828125
],
[
0.52294921875,
0.8291015625,
0.53515625,
0.8291015625,
0.53515625,
0.8388671875,
0.52294921875,
0.8388671875
],
[
0.428955078125,
0.83544921875,
0.48046875,
0.83544921875,
0.48046875,
0.849609375,
0.428955078125,
0.849609375
],
[
0.476318359375,
0.8388671875,
0.48193359375,
0.8388671875,
0.48193359375,
0.8447265625,
0.476318359375,
0.8447265625
],
[
0.80126953125,
0.833984375,
0.82861328125,
0.833984375,
0.82861328125,
0.85107421875,
0.80126953125,
0.85107421875
],
[
0.494384765625,
0.84521484375,
0.564453125,
0.84521484375,
0.564453125,
0.85986328125,
0.494384765625,
0.85986328125
],
[
0.1732177734375,
0.869140625,
0.82177734375,
0.87060546875,
0.82177734375,
0.88525390625,
0.1732177734375,
0.8837890625
],
[
0.1748046875,
0.88525390625,
0.82421875,
0.88525390625,
0.82421875,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1748046875,
0.8984375,
0.7939453125,
0.8984375,
0.7939453125,
0.912109375,
0.1748046875,
0.912109375
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"Most of the recent state-of-the-art RL results have been obtained using simple exploration strategies",
"such as uniform sampling [21] and i.i.d./correlated Gaussian noise [19, 30]. Although these heuristics",
"are sufficient in tasks with well-shaped rewards, the sample complexity can grow exponentially (with",
"state space size) in tasks with sparse rewards [25]. Recently developed exploration strategies for",
"deep RL have led to significantly improved performance on environments with sparse rewards.Boot",
"strapped DQN [24] led to faster learning in a range of Atari 2600 games by training an ensemble of",
"Q-functions. Intrinsic motivation methods using pseudo-counts achieve state-of-the-art performance",
"on Montezuma's Revenge, an extremely challenging Atari 2600 game [4].Variational Information",
"Maximizing Exploration (VIME, [13]) encourages the agent to explore by acquiring information",
"aboutenvironmentdynamics,andperformswellonvariousroboticlocomotionproblems with sparse",
"rewards. However, we have not seen a very simple and fast method that can work across different",
"domains.",
"Someof the classic,theoretically-justified exploration methods are based on counting state-action",
"visitations, and turning this count into a bonus reward. In the bandit setting, the well-known UCB",
"algorithm of [18] chooses the action at at time t that maximizes (at) +(]",
"/2logtwherer(at)is",
"the estimated reward, and n(at) is the number of times action at was previously chosen. In the",
"MDP setting, some of the algorithms have similar structure, for example, Model Based Interval",
"Estimation-Exploration Bonus (MBIE-EB) of [34] counts state-action pairs with a table n(s, a) and",
"adding a bonusreward of theform",
"toencourageexploringlessvisitedpairs.[16] show",
"Vn(s,a)",
"that the inverse-square-root dependence is optimal. MBIE and related algorithms assume that the",
"augmented MDP is solved analytically at each timestep, which is only practical for small finite state",
"spaces.",
"This paper presents a simple approach for exploration, which extends classic counting-based methods",
"to high-dimensional, continuous state spaces. We discretize the state space with a hash function and",
"apply a bonus based on the state-visitation count. The hash function can be chosen to appropriately",
"balancegeneralizationacrossstates,anddistinguishingbetweenstates.Weselectproblemsfromrllab",
"[8] and Atari 2600 [3] featuring sparse rewards, and demonstrate near state-of-the-art performance on",
"several games known to be hard for naive exploration strategies. The main strength of the presented",
"approach is that it is fast, flexible and complementary to most existing RL algorithms.",
"In summary,this paper proposes a generalization of classic count-based exploration to high-",
"dimensional spaces through hashing (Section 2); demonstrates its effectiveness on challenging deep",
"RL benchmark problems and analyzes key components of well-designed hash functions (Section 4).",
"2Methodology",
"2.1 Notation",
"This paper assumes a finite-horizon discounted Markov decision process (MDP), defined by",
"(S, A, P,r, po, , T), in which S is the state space, A the action space, P a transition probabil-",
"ity distribution, r : S × A → R a reward function, Po an initial state distribution, E (o,1l a",
"discount factor, and T the horizon. The goal of RL is to maximize the total expected discounted",
"reward Ex,pt=o tr(st,at)",
"over a policy r, which outputs a distribution over actions given a",
"state.",
"2.2Count-Based Exploration via Static Hashing",
"Our approach discretizes the state space with a hash function @ : S → Z. An exploration bonus",
"r+ : S→ R is added to the reward function, defined as",
"B",
"r+(s) :",
"",
"(1)",
"Vn((s)",
"where β e R>o is the bonus coefficient. Initially the counts n() are set to zero for the whole range",
"of . For every state St encountered at time step t, n(o(st)) is increased by one. The agent is trained",
"with rewards (r + r+), while performance is evaluated as the sum of rewards without bonuses.",
"2"
] |
{
"arxiv_id": "1611.04717",
"language": "en",
"page_idx": 1
}
|
||
1611.04717-en-2
|
1611.04717-en-2.jpg
|
<table>
<tbody>
<tr>
<td>$\overline{\mathbf{A}}$</td>
<td>rorithm $\varepsilon$ hashing. using SimHash</td>
</tr>
<tr>
<td>1 2</td>
<td>Define state preprocessor $g:\mathcal{S}\to\mathbb{R}^D$ (In case of SimHash) Initialize $A\in\mathbb{R}^{k\times D}$ with entries drawn i.i.d. from the standard Gaussian $\operatorname{distribution}\mathcal{N}(0,1)$</td>
</tr>
<tr>
<td>3 4</td>
<td>Initialize a hash table with values $n(\cdot)\equiv0$ for each iteration $j\textbf{do}$</td>
</tr>
<tr>
<td>5</td>
<td>$^\prime$nllert olicv $\pi$</td>
</tr>
<tr>
<td>6</td>
<td> </td>
</tr>
<tr>
<td>7</td>
<td>$\textbf{Update the hash table counts }\forall m: 0\leq m\leq M$ as $n(\phi(s_{m}))\leftarrow n(\phi(s_{m}))+1$</td>
</tr>
<tr>
<td>8</td>
<td>Update the policy $\pi$ using rewards $r(s_{m},a_{m})+$ $\beta$ with any RL algorithm $/n(\phi(s_{m}))$</td>
</tr>
</tbody>
</table>
Note that our approach is a departure from count-based exploration methods such as MBIE-EB since we use a state-space count $n(s)$ rather than a state-action count $n(s,a).$ State-action counts $n(s,a)$ are investigated in the Supplementary Material, but no significant performance gains over state counting could be witnessed. A possible reason is that the policy itself is sufficiently random to try most actions at a novel state.
Clearly the performance of this method will strongly depend on the choice of hash function $\phi.$ One important choice we can make regards the granularity of the discretization: we would like for“distant” states to be be counted separately while “similar”states are merged. If desired, we can incorporate prior knowledge into the choice of $\phi$, if there would be a set of salient state features which are known to be relevant. A short discussion on this matter is given in the Supplementary Material.
Algorithm 1 summarizes our method. The main idea is to use locality-sensitive hashing (LSH) to convert continuous, high-dimensional data to discrete hash codes. LSH is a popular class of hash functions for querying nearest neighbors based on certain similarity metrics [2]. A computationally efficient type of LSH is SimHash [6], which measures similarity by angular distance. SimHash retrieves a binary code of state $s\in\mathcal{S}$ as
$$\phi(s)=\mathrm{sgn}(Ag(s))\in\{-1,1\}^k,$$
(2)
where $g:\mathcal{S}\to\mathbb{R}^D$ is an optional preprocessing function and $A$ is a $k\times D$ matrix with i.i.d. entries drawn from a standard Gaussian distribution $\mathcal{N}(0,1).$ The value for $k$ controls the granularity: higher values lead to fewer collisions and are thus more likely to distinguish states.
2.3 Count-Based Exploration via Learned Hashing
When the MDP states have a complex structure, as is the case with image observations, measuring their similarity directly in pixel space fails to provide the semantic similarity measure one would desire. Previous work in computer vision $[7,20,36]\dot{\text{introduce manually designed feature representations}}$ of images that are suitable for semantic tasks including detection and classification. More recent methods learn complex features directly from data by training convolutional neural networks [12, 17,31]. Considering these results, it may be difficult for a method such as SimHash to cluster states appropriately using only raw pixels.
Therefore, rather than using SimHash, we propose to use an autoencoder (AE) to learn meaningful hash codes in one of its hidden layers as a more advanced LSH method. This AE takes as input states $s$ and contains one special dense layer comprised of $D$ sigmoid functions. By rounding the sigmoid activations $b(s)$ of this layer to their closest binary number $\lfloor b(s)\rceil\in\{0,1\}^D$, any state $s$ can be binarized. This is illustrated in Figure 1 for a convolutional AE.
$\begin{array}{c}\text{A problem with this architecture is that dissimilar inputs }s_i,s_j\\\text{can map to identical hash codes}\\\end{array}$ $\lfloor b(s_{i})\rceil=\lfloor b(s_{j})\rceil$,but the AE still reconstructs them perfectly. For example, if $b(s_{i})$ and $b(s_{j})$ have values 0.6 and 0.7 at a particular dimension, the difference can be exploited by deconvolutional layers in order to reconstruct $s_i$ and $s_j$ perfectly, although that dimension rounds to the same binary value. One can imagine replacing the bottleneck layer $b(s)$ with the hash codes $\lfloor b(s)\rceil$, but then gradients cannot be back-propagated through the rounding function. A solution is proposed by Gregor et al.[10] and Salakhutdinov & Hinton [28] is to inject uniform noise $U(-a,\tilde{a})$ into the sigmoid
3
|
[
[
0.1748046875,
0.0947265625,
0.6845703125,
0.0953369140625,
0.6845703125,
0.10986328125,
0.1748046875,
0.10919189453125
],
[
0.1748046875,
0.11297607421875,
0.4453125,
0.111083984375,
0.4453125,
0.1287841796875,
0.1748046875,
0.1307373046875
],
[
0.1732177734375,
0.1275634765625,
0.81201171875,
0.1275634765625,
0.81201171875,
0.1458740234375,
0.1732177734375,
0.1458740234375
],
[
0.1993408203125,
0.1458740234375,
0.330810546875,
0.1458740234375,
0.330810546875,
0.1590576171875,
0.1993408203125,
0.1590576171875
],
[
0.1748046875,
0.1578369140625,
0.47216796875,
0.1572265625,
0.47216796875,
0.1717529296875,
0.1748046875,
0.17236328125
],
[
0.17724609375,
0.1749267578125,
0.338134765625,
0.1749267578125,
0.338134765625,
0.1849365234375,
0.17724609375,
0.1849365234375
],
[
0.214111328125,
0.1849365234375,
0.65283203125,
0.1849365234375,
0.65283203125,
0.20263671875,
0.214111328125,
0.20263671875
],
[
0.178955078125,
0.203857421875,
0.1895751953125,
0.212158203125,
0.181396484375,
0.2171630859375,
0.17236328125,
0.210205078125
],
[
0.2156982421875,
0.20263671875,
0.8056640625,
0.20263671875,
0.8056640625,
0.2171630859375,
0.2156982421875,
0.2171630859375
],
[
0.2132568359375,
0.21337890625,
0.73779296875,
0.21533203125,
0.73779296875,
0.23291015625,
0.2132568359375,
0.2310791015625
],
[
0.178955078125,
0.2215576171875,
0.1829833984375,
0.2215576171875,
0.1829833984375,
0.2266845703125,
0.178955078125,
0.2266845703125
],
[
0.2181396484375,
0.24365234375,
0.443603515625,
0.24365234375,
0.443603515625,
0.2568359375,
0.2181396484375,
0.2568359375
],
[
0.468017578125,
0.24560546875,
0.54150390625,
0.24560546875,
0.54150390625,
0.2568359375,
0.468017578125,
0.2568359375
],
[
0.66015625,
0.242431640625,
0.81201171875,
0.242431640625,
0.81201171875,
0.2568359375,
0.66015625,
0.2568359375
],
[
0.5498046875,
0.25390625,
0.61669921875,
0.25390625,
0.61669921875,
0.264404296875,
0.5498046875,
0.264404296875
],
[
0.6337890625,
0.261474609375,
0.65625,
0.261474609375,
0.65625,
0.26513671875,
0.6337890625,
0.26513671875
],
[
0.1748046875,
0.3037109375,
0.82177734375,
0.3037109375,
0.82177734375,
0.317626953125,
0.1748046875,
0.317626953125
],
[
0.1748046875,
0.318115234375,
0.82275390625,
0.318115234375,
0.82275390625,
0.33203125,
0.1748046875,
0.33203125
],
[
0.1748046875,
0.33203125,
0.82275390625,
0.33203125,
0.82275390625,
0.34521484375,
0.1748046875,
0.34521484375
],
[
0.1732177734375,
0.3447265625,
0.82177734375,
0.34521484375,
0.82177734375,
0.359130859375,
0.1732177734375,
0.35791015625
],
[
0.1748046875,
0.35986328125,
0.3857421875,
0.35986328125,
0.3857421875,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.1748046875,
0.37890625,
0.82275390625,
0.37890625,
0.82275390625,
0.393310546875,
0.1748046875,
0.393310546875
],
[
0.1748046875,
0.393310546875,
0.826171875,
0.393310546875,
0.826171875,
0.4072265625,
0.1748046875,
0.4072265625
],
[
0.176513671875,
0.4091796875,
0.82177734375,
0.4091796875,
0.82177734375,
0.419189453125,
0.176513671875,
0.419189453125
],
[
0.1732177734375,
0.421630859375,
0.82421875,
0.421630859375,
0.82421875,
0.43505859375,
0.1732177734375,
0.43505859375
],
[
0.1732177734375,
0.43505859375,
0.74853515625,
0.43505859375,
0.74853515625,
0.448974609375,
0.1732177734375,
0.448974609375
],
[
0.1748046875,
0.455078125,
0.82275390625,
0.455078125,
0.82275390625,
0.468505859375,
0.1748046875,
0.468505859375
],
[
0.1748046875,
0.4697265625,
0.82275390625,
0.4697265625,
0.82275390625,
0.48291015625,
0.1748046875,
0.48291015625
],
[
0.1748046875,
0.484130859375,
0.82275390625,
0.484130859375,
0.82275390625,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.1732177734375,
0.496826171875,
0.82421875,
0.496826171875,
0.82421875,
0.51123046875,
0.1732177734375,
0.51123046875
],
[
0.1732177734375,
0.51123046875,
0.438720703125,
0.51123046875,
0.438720703125,
0.5244140625,
0.1732177734375,
0.5244140625
],
[
0.3896484375,
0.53076171875,
0.60302734375,
0.53076171875,
0.60302734375,
0.54541015625,
0.3896484375,
0.54541015625
],
[
0.80224609375,
0.53076171875,
0.82666015625,
0.53076171875,
0.82666015625,
0.548828125,
0.80224609375,
0.548828125
],
[
0.1748046875,
0.5517578125,
0.82275390625,
0.5517578125,
0.82275390625,
0.56640625,
0.1748046875,
0.56640625
],
[
0.1748046875,
0.56640625,
0.82421875,
0.56640625,
0.82421875,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.1732177734375,
0.58056640625,
0.6708984375,
0.58056640625,
0.6708984375,
0.5947265625,
0.1732177734375,
0.5947265625
],
[
0.17236328125,
0.609375,
0.5400390625,
0.60986328125,
0.5400390625,
0.625,
0.17236328125,
0.62353515625
],
[
0.1732177734375,
0.6337890625,
0.82177734375,
0.63525390625,
0.82177734375,
0.6494140625,
0.1732177734375,
0.6484375
],
[
0.1748046875,
0.65087890625,
0.826171875,
0.65087890625,
0.826171875,
0.6640625,
0.1748046875,
0.6640625
],
[
0.1732177734375,
0.66357421875,
0.82177734375,
0.66357421875,
0.82177734375,
0.6767578125,
0.1732177734375,
0.6767578125
],
[
0.1732177734375,
0.67822265625,
0.82275390625,
0.67822265625,
0.82275390625,
0.69140625,
0.1732177734375,
0.69140625
],
[
0.1748046875,
0.6923828125,
0.826171875,
0.6923828125,
0.826171875,
0.70556640625,
0.1748046875,
0.70556640625
],
[
0.1748046875,
0.70556640625,
0.82421875,
0.70556640625,
0.82421875,
0.7197265625,
0.1748046875,
0.7197265625
],
[
0.1748046875,
0.7197265625,
0.410888671875,
0.7197265625,
0.410888671875,
0.73291015625,
0.1748046875,
0.73291015625
],
[
0.1748046875,
0.7392578125,
0.82421875,
0.7392578125,
0.82421875,
0.75244140625,
0.1748046875,
0.75244140625
],
[
0.176513671875,
0.7548828125,
0.82275390625,
0.7548828125,
0.82275390625,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.1748046875,
0.767578125,
0.82275390625,
0.767578125,
0.82275390625,
0.78076171875,
0.1748046875,
0.78076171875
],
[
0.1748046875,
0.7822265625,
0.82421875,
0.7822265625,
0.82421875,
0.79541015625,
0.1748046875,
0.79541015625
],
[
0.1748046875,
0.79541015625,
0.6396484375,
0.79541015625,
0.6396484375,
0.8095703125,
0.1748046875,
0.8095703125
],
[
0.1748046875,
0.81494140625,
0.82421875,
0.81494140625,
0.82421875,
0.8291015625,
0.1748046875,
0.8291015625
],
[
0.176513671875,
0.830078125,
0.82177734375,
0.830078125,
0.82177734375,
0.84326171875,
0.176513671875,
0.84326171875
],
[
0.1748046875,
0.84326171875,
0.82275390625,
0.84326171875,
0.82275390625,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.1748046875,
0.859375,
0.8203125,
0.859375,
0.8203125,
0.869140625,
0.1748046875,
0.869140625
],
[
0.1732177734375,
0.8720703125,
0.82177734375,
0.8720703125,
0.82177734375,
0.88525390625,
0.1732177734375,
0.88525390625
],
[
0.1732177734375,
0.8837890625,
0.82275390625,
0.88330078125,
0.82275390625,
0.89794921875,
0.1732177734375,
0.8984375
],
[
0.1748046875,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1748046875,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Algorithm l:Count-based exploration through static hashing,using SimHash",
"1 Define state preprocessor g :S→RD",
"2 (In case of SimHash) Initialize A e RkxD with entries drawn i.i.d. from the standard Gaussian",
"distribution N(0, 1)",
"3 Initialize a hash table with values n()= 0",
"4 foreachiterationdo",
"Collect a set of state-action samples ((sm, am))M- with policy T",
"6",
"Compute hash codes through any LsH method, e.g., for SimHash, Φ(sm) = sgn(Ag(sm))",
"Update the hash table counts Vm : 0 ≤ m ≤ M as n(o(sm) ← n(o(sm) + i",
"",
"Update the policy π using rewards",
"(sm,am) +",
"with any RL algorithm",
"Vn(Φ(sm))",
"-",
"Note that our approach is a departure from count-based exploration methods such as MBIE-EB",
"since we use a state-space count n(s) rather than a state-action count n(s, a). State-action counts",
"n(s, a) are investigated in the Supplementary Material, but no significant performance gains over",
"state counting could be witnessed.A possible reason is that the policy itself is sufficiently random to",
"try most actions at a novel state.",
"Clearly the performance of this method will strongly depend on the choice of hash function o. One",
"important choice we can make regards the granularity of the discretization: we would like for“distant\"",
"states to be be counted separately while\"similar'states are merged.If desired,we can incorporate",
"prior knowledge into the choice of , if there would be a set of salient state features which are known",
"to be relevant. A short discussion on this matter is given in the Supplementary Material.",
"Algorithm 1 summarizes our method. The main idea is to use locality-sensitive hashing (LSH) to",
"convert continuous, high-dimensional data to discrete hash codes. LSH is a popular class of hash",
"functions for querying nearest neighbors based on certain similarity metrics [2]. A computationally",
"efficient type of LSH is SimHash [6], which measures similarity by angular distance. SimHash",
"retrieves a binary code of state s E S as",
"Φ(s) = sgn(Ag(s)) e (-1,11k",
"(2) ",
"where g : S → RD is an optional preprocessing function and A is a k × D matrix with i.i.d. entries",
"drawn from a standard Gaussian distribution N(o, 1). The value for k controls the granularity: higher",
"values lead to fewer collisions and are thus more likely to distinguish states.",
"2.3Count-Based Explorationvia Learned Hashing",
"When the MDP states have a complex structure, as is the case with image observations, measuring",
"their similarity directly in pixel space fails to provide the semantic similarity measure one would desire.",
"Previous work in computer vision [7,20, 36] introduce manually designed feature representations",
"of images that are suitable for semantic tasks including detection and classification. More recent",
"methods learn complex features directly from data by training convolutional neural networks [12,",
"17, 31]. Considering these results, it may be difficult for a method such as SimHash to cluster states",
"appropriately using onlyrawpixels.",
"Therefore, rather than using SimHash, we propose to use an autoencoder (AE) to learn meaningful",
"hashcodesinoneof itshiddenlayersasamoreadvancedLSHmethod.ThisAEtakesas input",
"states s and contains one special dense layer comprised of D sigmoid functions. By rounding the",
"sigmoid activations b(s) of this layer to their closest binary number [b(s)] E (0, 1) P, any state s",
"can be binarized. This is illustrated in Figure 1 for a convolutional AE.",
"A problem with this architecture is that dissimilar inputs si, s, can map to identical hash codes",
"[b(si)| =[b(s;)], but the AE still reconstructs them perfectly. For example, if b(si) and b(s;) have",
"values 0.6 and 0.7 at a particular dimension, the difference can be exploited by deconvolutional",
"layersinordertoreconstructs;ands,perfectly,althoughthatdimensionroundstothesamebinary",
"value. One can imagine replacing the bottleneck layer b(s) with the hash codes [b(s)l, but then",
"gradients cannot be back-propagated through the rounding function. A solution is proposed by Gregor",
"et al. [10] and Salakhutdinov & Hinton [28] is to inject uniform noise U(-a, a) into the sigmoid",
"3"
] |
{
"arxiv_id": "1611.04717",
"language": "en",
"page_idx": 2
}
|
||
1611.04717-en-3
|
1611.04717-en-3.jpg
|
<FigureHere>
Figure l: The autoencoder (AE) architecture for ALE; the solid block represents the dense sigmoidal
binary code layer, after which noise $U(-a,a)$ is injected
<table>
<tbody>
<tr>
<td>$\overline{\mathbf{AI}}$</td>
<td>arithm ration ilsino ${\mathrm{g~learned~hash~codes}}$</td>
</tr>
<tr>
<td>1</td>
<td>Define state preprocessor $g:\mathcal{S}\to\{0,1\}^{D}$ as the binary code resulting from the autoencoder (AE)</td>
</tr>
<tr>
<td>2 3 4</td>
<td>nitialize $A\in\mathbb{R}^{k\times D}$ with entries drawn i.i.d. from the standard Gaussian distribution $\mathcal{N}(0,1)$ nitialize a hash table with values $n(\cdot)\equiv0$ $\textbf{or each iteration }j\textbf{do}$</td>
</tr>
<tr>
<td>5</td>
<td>Collect a set c $M$ with policy $\pi$</td>
</tr>
<tr>
<td>6</td>
<td>Add the state sample o a FIFO replay pool $\mathcal{R}$</td>
</tr>
<tr>
<td>7</td>
<td>$\textbf{if }j$mod $j_{\mathrm{n}}$ $= 0\textbf{ then}$</td>
</tr>
<tr>
<td>8</td>
<td>Update the AE loss function in Eq.(3) using samples drawn from the replay pool $\{s_{n}\}_{n=1}^{N}\sim\mathcal{R}$,for example using stochastic gradient descent</td>
</tr>
<tr>
<td>1</td>
<td>Update the hash table counts $\forall m:0<$ $M$ as $n(\phi(s_{m}))\in$ $\leftarrow n(\phi(s_{m}))+1$</td>
</tr>
<tr>
<td>12</td>
<td>Update the policy $\pi$ using rewards $\textbf{s}\left \{ r( s_{m}, a_{m}) + \right \}$ $\beta$ with any RL algorithm $\sqrt{n(\phi(\mathbf{s}_{m}))}$ $m=0$</td>
</tr>
<tr>
<td> </td>
<td>$m=0$</td>
</tr>
</tbody>
</table>
activations. By choosing uniform noise with $a>\frac14$, the AE is only capable of (always) reconstructing distinct state inputs $s_i\neq s_j$, if it has learned to spread the sigmoid outputs sufficiently far apart $|b(s_{i})-b(s_{j})|>\epsilon$,in order to counteract the injected noise.
As such, the loss function over a set of collected states $\{s_i\}_i=1^N$ is defined as
$$L\left(\{s_n\}_{n=1}^N\right)=-\dfrac{1}{N}\sum_{n=1}^N\Big[\log p(s_n)-\dfrac{\lambda}{K}\sum_{i=1}^D\min\Big\{\left(1-b_i(s_n)\right)^2,b_i(s_n)^2\Big\}\Big],$$
(3)
with $p(s_n)$ the AE output. This objective function consists of a negative log-likelihood term and a term that pressures the binary code layer to take on binary values, scaled by $\lambda\in\mathbb{R}_\geq0.$ The reasoning behind this latter term is that it might happen that for particular states, a certain sigmoid unit is never used. Therefore, its value might fluctuate around $\frac12$,causing the corresponding bit in binary code $\lfloor b( s) \rceil \text{ to flip over the agent lifetime. Adding this s\'{e} cond loss term ensures that an unused bit takes}$ on an arbitrary binary value. For Atari 2600 image inputs, since the pixel intensities are discrete values in the range [0,255], we make use of a pixel-wise softmax output layer [37] that shares weights between all pixels. The architectural details are described in the Supplementary Material and are depicted in Figure 1. Because the code dimension often needs to be large in order to correctly reconstruct the input, we apply a downsampling procedure to the resulting binary code $\lfloor b(s)\rceil$, which can be done through random projection to a lower-dimensional space via SimHash as in Eq.(2).
On the one hand, it is important that the mapping from state to code needs to remain relatively consistent over time, which is nontrivial as the AE is constantly updated according to the latest data (Algorithm 2 line 8). A solution is to downsample the binary code to a very low dimension, or by slowing down the training process. On the other hand, the code has to remain relatively unique
4
|
[
[
0.27783203125,
0.1912841796875,
0.316162109375,
0.1912841796875,
0.316162109375,
0.1988525390625,
0.27783203125,
0.1988525390625
],
[
0.459228515625,
0.1881103515625,
0.4794921875,
0.1881103515625,
0.4794921875,
0.1988525390625,
0.459228515625,
0.1988525390625
],
[
0.223876953125,
0.2020263671875,
0.273681640625,
0.2020263671875,
0.273681640625,
0.212158203125,
0.223876953125,
0.212158203125
],
[
0.529296875,
0.1962890625,
0.5498046875,
0.1962890625,
0.5498046875,
0.203857421875,
0.529296875,
0.203857421875
],
[
0.6787109375,
0.2020263671875,
0.77392578125,
0.2020263671875,
0.77392578125,
0.208984375,
0.6787109375,
0.208984375
],
[
0.1748046875,
0.2203369140625,
0.82275390625,
0.2203369140625,
0.82275390625,
0.2342529296875,
0.1748046875,
0.2342529296875
],
[
0.1732177734375,
0.23486328125,
0.5498046875,
0.23486328125,
0.5498046875,
0.248779296875,
0.1732177734375,
0.248779296875
],
[
0.1748046875,
0.26953125,
0.59716796875,
0.26953125,
0.59716796875,
0.28271484375,
0.1748046875,
0.28271484375
],
[
0.1732177734375,
0.28857421875,
0.79833984375,
0.28857421875,
0.79833984375,
0.302978515625,
0.1732177734375,
0.302978515625
],
[
0.1993408203125,
0.302978515625,
0.2344970703125,
0.302978515625,
0.2344970703125,
0.314453125,
0.1993408203125,
0.314453125
],
[
0.1748046875,
0.31494140625,
0.8037109375,
0.31494140625,
0.8037109375,
0.333251953125,
0.1748046875,
0.333251953125
],
[
0.17724609375,
0.333251953125,
0.473876953125,
0.333251953125,
0.473876953125,
0.3466796875,
0.17724609375,
0.3466796875
],
[
0.176513671875,
0.3466796875,
0.340576171875,
0.3466796875,
0.340576171875,
0.35986328125,
0.176513671875,
0.35986328125
],
[
0.214111328125,
0.361083984375,
0.5498046875,
0.359130859375,
0.5498046875,
0.376953125,
0.214111328125,
0.37890625
],
[
0.54736328125,
0.3623046875,
0.65283203125,
0.3623046875,
0.65283203125,
0.376953125,
0.54736328125,
0.376953125
],
[
0.214111328125,
0.376953125,
0.435546875,
0.380126953125,
0.435546875,
0.397705078125,
0.214111328125,
0.39453125
],
[
0.17724609375,
0.385009765625,
0.1854248046875,
0.385009765625,
0.1854248046875,
0.39453125,
0.17724609375,
0.39453125
],
[
0.420654296875,
0.380615234375,
0.5947265625,
0.380615234375,
0.5947265625,
0.39453125,
0.420654296875,
0.39453125
],
[
0.178955078125,
0.397705078125,
0.1829833984375,
0.397705078125,
0.1829833984375,
0.40283203125,
0.178955078125,
0.40283203125
],
[
0.2156982421875,
0.39453125,
0.378173828125,
0.39453125,
0.378173828125,
0.4091796875,
0.2156982421875,
0.4091796875
],
[
0.178955078125,
0.41357421875,
0.1829833984375,
0.41357421875,
0.1829833984375,
0.41845703125,
0.178955078125,
0.41845703125
],
[
0.2418212890625,
0.40771484375,
0.771484375,
0.40771484375,
0.771484375,
0.421630859375,
0.2418212890625,
0.421630859375
],
[
0.2467041015625,
0.423583984375,
0.64892578125,
0.423583984375,
0.64892578125,
0.438232421875,
0.2467041015625,
0.438232421875
],
[
0.214111328125,
0.43994140625,
0.77294921875,
0.439453125,
0.77294921875,
0.45703125,
0.214111328125,
0.458251953125
],
[
0.1690673828125,
0.45703125,
0.1871337890625,
0.45703125,
0.1871337890625,
0.47021484375,
0.1690673828125,
0.47021484375
],
[
0.2156982421875,
0.45703125,
0.72607421875,
0.45703125,
0.72607421875,
0.47021484375,
0.2156982421875,
0.47021484375
],
[
0.1707763671875,
0.472900390625,
0.1854248046875,
0.472900390625,
0.1854248046875,
0.48291015625,
0.1707763671875,
0.48291015625
],
[
0.21728515625,
0.47021484375,
0.734375,
0.47021484375,
0.734375,
0.48486328125,
0.21728515625,
0.48486328125
],
[
0.2197265625,
0.496826171875,
0.443603515625,
0.496826171875,
0.443603515625,
0.51123046875,
0.2197265625,
0.51123046875
],
[
0.6611328125,
0.496826171875,
0.810546875,
0.496826171875,
0.810546875,
0.51123046875,
0.6611328125,
0.51123046875
],
[
0.468017578125,
0.50048828125,
0.53369140625,
0.50048828125,
0.53369140625,
0.51123046875,
0.468017578125,
0.51123046875
],
[
0.17236328125,
0.5,
0.1846923828125,
0.5,
0.1846923828125,
0.5087890625,
0.17236328125,
0.5087890625
],
[
0.5498046875,
0.5068359375,
0.61669921875,
0.5068359375,
0.61669921875,
0.51806640625,
0.5498046875,
0.51806640625
],
[
0.6396484375,
0.5166015625,
0.65283203125,
0.5166015625,
0.65283203125,
0.51953125,
0.6396484375,
0.51953125
],
[
0.1732177734375,
0.556640625,
0.82275390625,
0.55810546875,
0.82275390625,
0.57275390625,
0.1732177734375,
0.5712890625
],
[
0.1748046875,
0.5712890625,
0.826171875,
0.5712890625,
0.826171875,
0.5859375,
0.1748046875,
0.5859375
],
[
0.1748046875,
0.58544921875,
0.57275390625,
0.58544921875,
0.57275390625,
0.5986328125,
0.1748046875,
0.5986328125
],
[
0.1748046875,
0.60791015625,
0.6708984375,
0.60791015625,
0.6708984375,
0.6220703125,
0.1748046875,
0.6220703125
],
[
0.233642578125,
0.642578125,
0.28759765625,
0.642578125,
0.28759765625,
0.6572265625,
0.233642578125,
0.6572265625
],
[
0.406982421875,
0.64306640625,
0.5361328125,
0.64306640625,
0.5361328125,
0.65771484375,
0.406982421875,
0.65771484375
],
[
0.5400390625,
0.64306640625,
0.580078125,
0.64306640625,
0.580078125,
0.65478515625,
0.5400390625,
0.65478515625
],
[
0.60302734375,
0.6416015625,
0.7294921875,
0.6416015625,
0.7294921875,
0.6591796875,
0.60302734375,
0.6591796875
],
[
0.80810546875,
0.6484375,
0.81787109375,
0.6484375,
0.81787109375,
0.65380859375,
0.80810546875,
0.65380859375
],
[
0.1748046875,
0.67822265625,
0.82177734375,
0.67822265625,
0.82177734375,
0.69140625,
0.1748046875,
0.69140625
],
[
0.1748046875,
0.6923828125,
0.82275390625,
0.6923828125,
0.82275390625,
0.70556640625,
0.1748046875,
0.70556640625
],
[
0.1732177734375,
0.7041015625,
0.826171875,
0.705078125,
0.826171875,
0.7197265625,
0.1732177734375,
0.71826171875
],
[
0.1748046875,
0.7197265625,
0.82421875,
0.7197265625,
0.82421875,
0.73291015625,
0.1748046875,
0.73291015625
],
[
0.17724609375,
0.73193359375,
0.82421875,
0.73193359375,
0.82421875,
0.74560546875,
0.17724609375,
0.74560546875
],
[
0.1748046875,
0.74853515625,
0.361083984375,
0.74853515625,
0.361083984375,
0.7587890625,
0.1748046875,
0.7587890625
],
[
0.1732177734375,
0.767578125,
0.82275390625,
0.767578125,
0.82275390625,
0.78076171875,
0.1732177734375,
0.78076171875
],
[
0.1748046875,
0.78271484375,
0.82177734375,
0.78271484375,
0.82177734375,
0.79345703125,
0.1748046875,
0.79345703125
],
[
0.176513671875,
0.796875,
0.82177734375,
0.796875,
0.82177734375,
0.806640625,
0.176513671875,
0.806640625
],
[
0.1732177734375,
0.806640625,
0.82421875,
0.8095703125,
0.82421875,
0.82275390625,
0.1732177734375,
0.82080078125
],
[
0.1732177734375,
0.82275390625,
0.826171875,
0.82275390625,
0.826171875,
0.8359375,
0.1732177734375,
0.8359375
],
[
0.1732177734375,
0.8369140625,
0.61279296875,
0.8369140625,
0.61279296875,
0.85107421875,
0.1732177734375,
0.85107421875
],
[
0.1732177734375,
0.85693359375,
0.82275390625,
0.85693359375,
0.82275390625,
0.87060546875,
0.1732177734375,
0.87060546875
],
[
0.176513671875,
0.87255859375,
0.82275390625,
0.87255859375,
0.82275390625,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.17236328125,
0.88330078125,
0.82275390625,
0.8837890625,
0.82275390625,
0.8984375,
0.17236328125,
0.89794921875
],
[
0.1732177734375,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"x24×24",
"1024",
"1 ×52×52",
"2400",
"X52×5264x52x52",
"Figure 1: The autoencoder (AE) architecture for ALE; the solid block represents the dense sigmoidal",
"binary code layer, after which noise U(-a, a) is injected.",
"Algorithm 2: Count-based exploration using learned hash codes",
"1 Define state preprocessor g : S → [O, 1jD as the binary code resulting from the autoencoder",
"(AE)",
"2 Initialize A e RkxD with entries drawn i.i.d. from the standard Gaussian distribution N(O, 1)",
"3 Initialize a hash table with values n() = 0",
"4 for each iteration j do",
"Collect a set of state-action samples [(sm, am))M",
"owith policy",
"Add the state samples (sm) m=o",
"6",
"o to a FIFO replay pool R",
"",
"if j mod jupdate = O then",
"",
"Update the AE loss function in Eq. (3) using samples drawn from the replay pool",
"{sn]n-1 ~ R, for example using stochastic gradient descent",
"Compute g(sm) = [b(sm)], the D-dim rounded hash code for Sm learned by the AE",
"10",
"Project g(sm) to a lower dimension k via SimHash as Φ(sm) = sgn(Ag(sm))",
"11",
"Update the hash table counts Vm : 0 ≤ m ≤ M as n(Φ(sm)) n(o(sm)) + 1",
"Update the policy using rewards",
"with any RL algorithm",
"(Sm,am)",
"12",
"Vn(o(sm))",
"-",
"activations. By choosing uniform noise with a >1, the AE is only capable of (always) reconstructing",
"distinct state inputs s + sg, if it has learned to spread the sigmoid outputs sufficiently far apart,",
"Jb(si)-b(s,)/> e, in order to counteract the injected noise.",
"As such, the loss function over a set of collected states [si-, is defined as",
"L ((sn)",
"log p(sn)-E",
",min",
"1 - bi(sn))2, bi(sn",
"3",
"with p(sn) the AE output. This objective function consists of a negative log-likelihood term and a",
"term that pressures the binary code layer to take on binary values, scaled by e R>o. The reasoning",
"behind this latter term is that it might happen that for particular states, a certain sigmoid unit is never",
"used. Therefore, its value might fluctuate around , causing the corresponding bit in binary code",
"b(s)l to flip over the agent lifetime. Adding this second loss term ensures that an unused bit takes",
"on an arbitrary binary value.",
"For Atari 2600 image inputs, since the pixel intensities are discrete values in the range J0, 255]",
"wemakeuse ofapixel-wise softmaxoutputlayer[37lthat sharesweightsbetweenall pixels.The",
"architectural detailsaredescribedintheSupplementaryMaterialandaredepictedinFigurel.Because",
"the code dimension often needs tobe large in orderto correctlyreconstruct the input, we apply a",
"downsampling procedure to the resulting binary code b(s)7, which can be done through random",
"projection to a lower-dimensional space via SimHash as in Eq. (2).",
"On the one hand, it is important that the mapping from state to code needs to remain relatively",
"consistentovertime,whichis nontrivial as theAEis constantlyupdated accordingtothelatestdata",
"(Algorithm 2 line 8). A solution is to downsample the binary code to a very low dimension, or by",
"slowing down the training process. On the other hand, the code has to remain relatively unique",
"4"
] |
{
"arxiv_id": "1611.04717",
"language": "en",
"page_idx": 3
}
|
||
1611.04717-en-4
|
1611.04717-en-4.jpg
|
for states that are both distinct and close together on the image manifold. This is tackled both by the second term in Eq. (3) and by the saturating behavior of the sigmoid units. States already well represented by the AE tend to saturate the sigmoid activations, causing the resulting loss gradients to be close to zero, making the code less prone to change.
## 3 $\textbf{Related Work}$
Classic count-based methods such as MBIE[33], MBIE-EB and[16] solve an approximate Bellman equation as an inner loop before the agent takes an action [34]. As such, bonus rewards are propagated immediately throughout the state-action space. In contrast, contemporary deep RL algorithms propagate the bonus signal based on rollouts collected from interacting with environments, with value-based[21] or policy gradient-based[22,30] methods, at limited speed. In addition, our proposed method is intended to work with contemporary deep RL algorithms, it differs from classical count-based method in that our method relies on visiting unseen states first, before the bonus reward can be assigned, making uninformed exploration strategies still a necessity at the beginning. Filling the gaps between our method and classic theories is an important direction of future research.
A related line of classical exploration methods is based on the idea of optimism in the face of $uncertainty_{2}[5]$ but not restricted to using counting to implement“optimism”, e.g., R-Max [5], UCRL [14], and $\mathbb{E}^3[15].$ These methods, similar to MBIE and MBIE-EB, have theoretical guarantees in tabular settings.
Bayesian RL methods [9,11,16,35],which keep track of a distribution over MDPs, are an alternative to optimism-based methods. Extensions to continuous state space have been proposed by [27] and [25].
Another type of exploration is curiosity-based exploration. These methods try to capture the agent's surprise about transition dynamics. As the agent tries to optimize for surprise, it naturally discovers novel states. We refer the reader to[29] and[26] for an extensive review on curiosity and intrinsic rewards.
Several exploration strategies for deep RL have been proposed to handle high-dimensional state space recently.[13] propose VIME, in which information gain is measured in Bayesian neural networks modeling the MDP dynamics, which is used an exploration bonus.[32] propose to use the prediction error of a learned dynamics model as an exploration bonus. Thompson sampling through bootstrapping is proposed by [24], using bootstrapped Q-functions.
The most related exploration strategy is proposed by [4], in which an exploration bonus is added inversely proportional to the square root of a $pseudo-count$ quantity. A state pseudo-count is derived from its log-probability improvement according to a density model over the state space, which in the limit converges to the empirical count. Our method is similar to pseudo-count approach in the sense that both methods are performing approximate counting to have the necessary generalization over unseen states. The difference is that a density model has to be designed and learned to achieve good generalization for pseudo-count whereas in our case generalization is obtained by a wide range of simple hash functions (not necessarily SimHash). Another interesting connection is that our method also implies a density model $\rho(s)=\frac{n(\phi(s))}N$ over all visited states, where $N$ is the total number of cluster centers instead of the true states, but this method has yet to be tested on standard exploration benchmark problems.
## 4 Experiments
Experiments were designed to investigate and answer the following research questions:
1. Can count-based exploration through hashing improve performance significantly across
different domains? $\hat{\text{How does the proposed method compare to the current state of the art in}}$
exploration for deep RL?
2. What is the impact of learned or static state preprocessing on the overall performance when
image observations are used?
5
|
[
[
0.1748046875,
0.0921630859375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.1060791015625,
0.1748046875,
0.1060791015625
],
[
0.1732177734375,
0.1060791015625,
0.82275390625,
0.1060791015625,
0.82275390625,
0.11932373046875,
0.1732177734375,
0.11932373046875
],
[
0.1748046875,
0.12060546875,
0.82421875,
0.12060546875,
0.82421875,
0.1337890625,
0.1748046875,
0.1337890625
],
[
0.17236328125,
0.1331787109375,
0.5361328125,
0.1337890625,
0.5361328125,
0.1483154296875,
0.17236328125,
0.147705078125
],
[
0.17236328125,
0.17041015625,
0.32177734375,
0.17041015625,
0.32177734375,
0.1849365234375,
0.17236328125,
0.1849365234375
],
[
0.1732177734375,
0.20263671875,
0.82177734375,
0.203857421875,
0.82177734375,
0.2183837890625,
0.1732177734375,
0.2171630859375
],
[
0.1748046875,
0.2183837890625,
0.82275390625,
0.2183837890625,
0.82275390625,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.17236328125,
0.231689453125,
0.82177734375,
0.231689453125,
0.82177734375,
0.24560546875,
0.17236328125,
0.24560546875
],
[
0.1732177734375,
0.24560546875,
0.82275390625,
0.2442626953125,
0.82275390625,
0.25830078125,
0.1732177734375,
0.2587890625
],
[
0.1732177734375,
0.25830078125,
0.82275390625,
0.2587890625,
0.82275390625,
0.2734375,
0.1732177734375,
0.272705078125
],
[
0.1748046875,
0.2734375,
0.82275390625,
0.2734375,
0.82275390625,
0.287353515625,
0.1748046875,
0.287353515625
],
[
0.176513671875,
0.28857421875,
0.82177734375,
0.28857421875,
0.82177734375,
0.298583984375,
0.176513671875,
0.298583984375
],
[
0.1748046875,
0.300537109375,
0.82421875,
0.300537109375,
0.82421875,
0.314453125,
0.1748046875,
0.314453125
],
[
0.1732177734375,
0.314453125,
0.78173828125,
0.313232421875,
0.78173828125,
0.32763671875,
0.1732177734375,
0.328857421875
],
[
0.1732177734375,
0.333251953125,
0.826171875,
0.333984375,
0.826171875,
0.348388671875,
0.1732177734375,
0.347900390625
],
[
0.1748046875,
0.349853515625,
0.82275390625,
0.349853515625,
0.82275390625,
0.363037109375,
0.1748046875,
0.363037109375
],
[
0.1748046875,
0.363037109375,
0.82421875,
0.363037109375,
0.82421875,
0.376953125,
0.1748046875,
0.376953125
],
[
0.1732177734375,
0.375732421875,
0.279296875,
0.376953125,
0.279296875,
0.391357421875,
0.1732177734375,
0.39013671875
],
[
0.1748046875,
0.396484375,
0.82275390625,
0.396484375,
0.82275390625,
0.410400390625,
0.1748046875,
0.410400390625
],
[
0.1748046875,
0.410888671875,
0.82421875,
0.410888671875,
0.82421875,
0.4248046875,
0.1748046875,
0.4248046875
],
[
0.1748046875,
0.4267578125,
0.20751953125,
0.4267578125,
0.20751953125,
0.439453125,
0.1748046875,
0.439453125
],
[
0.1732177734375,
0.44580078125,
0.82177734375,
0.44580078125,
0.82177734375,
0.458984375,
0.1732177734375,
0.458984375
],
[
0.1748046875,
0.460205078125,
0.82275390625,
0.460205078125,
0.82275390625,
0.473388671875,
0.1748046875,
0.473388671875
],
[
0.1732177734375,
0.472900390625,
0.82421875,
0.472900390625,
0.82421875,
0.486083984375,
0.1732177734375,
0.486083984375
],
[
0.1748046875,
0.488037109375,
0.2303466796875,
0.488037109375,
0.2303466796875,
0.5,
0.1748046875,
0.5
],
[
0.1748046875,
0.5068359375,
0.82275390625,
0.5068359375,
0.82275390625,
0.521484375,
0.1748046875,
0.521484375
],
[
0.1748046875,
0.521484375,
0.82421875,
0.521484375,
0.82421875,
0.53515625,
0.1748046875,
0.53515625
],
[
0.1748046875,
0.53515625,
0.82275390625,
0.53515625,
0.82275390625,
0.548828125,
0.1748046875,
0.548828125
],
[
0.1748046875,
0.548828125,
0.82275390625,
0.548828125,
0.82275390625,
0.5625,
0.1748046875,
0.5625
],
[
0.1748046875,
0.56298828125,
0.61767578125,
0.56298828125,
0.61767578125,
0.5771484375,
0.1748046875,
0.5771484375
],
[
0.1748046875,
0.58349609375,
0.82275390625,
0.58349609375,
0.82275390625,
0.59765625,
0.1748046875,
0.59765625
],
[
0.1748046875,
0.59765625,
0.82275390625,
0.59765625,
0.82275390625,
0.611328125,
0.1748046875,
0.611328125
],
[
0.1748046875,
0.611328125,
0.82421875,
0.611328125,
0.82421875,
0.625,
0.1748046875,
0.625
],
[
0.1732177734375,
0.62548828125,
0.82275390625,
0.62548828125,
0.82275390625,
0.6396484375,
0.1732177734375,
0.6396484375
],
[
0.1748046875,
0.63818359375,
0.82275390625,
0.63818359375,
0.82275390625,
0.65283203125,
0.1748046875,
0.65283203125
],
[
0.1748046875,
0.6513671875,
0.82275390625,
0.6513671875,
0.82275390625,
0.66552734375,
0.1748046875,
0.66552734375
],
[
0.1748046875,
0.66552734375,
0.82421875,
0.66552734375,
0.82421875,
0.6787109375,
0.1748046875,
0.6787109375
],
[
0.1748046875,
0.6796875,
0.82421875,
0.6796875,
0.82421875,
0.693359375,
0.1748046875,
0.693359375
],
[
0.17236328125,
0.693359375,
0.82666015625,
0.6923828125,
0.82666015625,
0.71044921875,
0.17236328125,
0.7109375
],
[
0.44287109375,
0.69775390625,
0.453369140625,
0.69775390625,
0.453369140625,
0.7021484375,
0.44287109375,
0.7021484375
],
[
0.1748046875,
0.71044921875,
0.82275390625,
0.71044921875,
0.82275390625,
0.7236328125,
0.1748046875,
0.7236328125
],
[
0.1748046875,
0.724609375,
0.82275390625,
0.724609375,
0.82275390625,
0.73779296875,
0.1748046875,
0.73779296875
],
[
0.1748046875,
0.73779296875,
0.318603515625,
0.73779296875,
0.318603515625,
0.751953125,
0.1748046875,
0.751953125
],
[
0.17724609375,
0.77587890625,
0.193603515625,
0.77880859375,
0.191162109375,
0.7890625,
0.1748046875,
0.7861328125
],
[
0.1895751953125,
0.77587890625,
0.310546875,
0.77587890625,
0.310546875,
0.79052734375,
0.1895751953125,
0.79052734375
],
[
0.1748046875,
0.80810546875,
0.744140625,
0.80810546875,
0.744140625,
0.8212890625,
0.1748046875,
0.8212890625
],
[
0.2132568359375,
0.83544921875,
0.82275390625,
0.83544921875,
0.82275390625,
0.8486328125,
0.2132568359375,
0.8486328125
],
[
0.2344970703125,
0.849609375,
0.82275390625,
0.849609375,
0.82275390625,
0.86279296875,
0.2344970703125,
0.86279296875
],
[
0.233642578125,
0.86279296875,
0.400390625,
0.86279296875,
0.400390625,
0.876953125,
0.233642578125,
0.876953125
],
[
0.2115478515625,
0.8837890625,
0.82421875,
0.8837890625,
0.82421875,
0.8984375,
0.2115478515625,
0.8984375
],
[
0.233642578125,
0.8984375,
0.42724609375,
0.8984375,
0.42724609375,
0.912109375,
0.233642578125,
0.912109375
],
[
0.494384765625,
0.93505859375,
0.50634765625,
0.93505859375,
0.50634765625,
0.94873046875,
0.494384765625,
0.94873046875
]
] |
[
"for states that are both distinct and close together on the image manifold. This is tackled both by",
"the second term in Eq. (3) and by the saturating behavior of the sigmoid units. States already well",
"represented by the AE tend to saturate the sigmoid activations, causing the resulting loss gradients to",
"be close to zero, making the code less prone to change.",
"3Related Work",
"Classic count-based methods such as MBIE [33], MBIE-EB and [16] solve an approximate Bellman",
"equation as an inner loop before the agent takes an action [34]. As such, bonus rewards are propagated",
"immediately throughout the state-action space. In contrast, contemporary deep RL algorithms",
"propagate the bonus signal based on rollouts collected from interacting with environments,with",
"value-based [21] or policy gradient-based [22, 30] methods, at limited speed. In addition, our",
"proposed method is intended to work with contemporary deep RL algorithms, it differs from classical",
"count-based method inthat ourmethodrelieson visiting unseenstatesfirstbeforethebonusreward",
"can be assigned, making uninformed exploration strategies still a necessity at the beginning. Filling",
"the gaps between our method and classic theories is an important direction of future research",
"A related line of classical exploration methods is based on the idea of optimism in the face of",
"uncertainty [5] but not restricted to using counting to implement \"optimism\", e.g., R-Max [5], UCRL",
"[14], and E3 [15]. These methods, similar to MBIE and MBIE-EB, have theoretical guarantees in",
"tabular settings.",
"Bayesian RL methods [9, 11, 16, 35], which keep track of a distribution over MDPs, are an alternative",
"to optimism-based methods. Extensions to continuous state space have been proposed by [27] and",
"[25].",
"Another type of exploration is curiosity-based exploration. These methods try to capture the agent's",
"surprise about transition dynamics. As the agent tries to optimize for surprise, it naturally discovers",
"novel states. We refer the reader to [29] and [26] for an extensive review on curiosity and intrinsic",
"rewards.",
"Several exploration strategies for deep RL have been proposed to handle high-dimensional state",
"space recently. [13] propose VIME, in which information gain is measured in Bayesian neural",
"networks modeling the MDP dynamics, which is used an exploration bonus. [32] propose to use the",
"prediction error of a learned dynamics model as an exploration bonus. Thompson sampling through",
"bootstrapping is proposed by [24], using bootstrapped Q-functions.",
"The most related exploration strategy is proposed by [4], in which an exploration bonus is added",
"inversely proportional to the square root of a pseudo-count quantity. A state pseudo-count is derived",
"from its log-probability improvement according to a density model over the state space, which in the",
"limit converges to the empirical count. Our method is similar to pseudo-count approach in the sense",
"that both methods are performing approximate counting to have the necessary generalization over",
"unseen states. The difference is that a density model has to be designed and learned to achieve good",
"generalization for pseudo-count whereas in our case generalization is obtained by a wide range of",
"simple hash functions (not necessarily SimHash). Another interesting connection is that our method",
"also implies a density model p(s) = n($() over all visited states, where N is the total number of",
"1S",
"states visited. Another method similar to hashing is proposed by [1], which clusters states and counts",
"cluster centers instead of the true states, but this method has yet to be tested on standard exploration",
"benchmark problems.",
"4.",
"Experiments",
"Experiments were designed to investigate and answer the following research questions:",
"1. Can count-based exploration through hashing improve performance significantly across",
"different domains? How does the proposed method compare to the current state of the art in",
"exploration for deep RL?",
"2. What is the impact of learned or static state preprocessing on the overall performance when",
"image observations are used?",
"5"
] |
{
"arxiv_id": "1611.04717",
"language": "en",
"page_idx": 4
}
|
||
1805.03344-en-0
|
1805.03344-en-0.jpg
|
Attention-Aware Compositional Network for Person Re-identification
$\begin{array}{cc}\mathrm{Jing~Xu}^{1}&\text{Rui Zhao}^{1,2*}&&\text{Feng Zhu}^{1}&\text{Huaming Wang}^{1}\\\text{SenseNets Technology Limited}&&&\text{2SenseTime Group Limited}\\\end{array}$
Wanli Ouyang$^{3}$
$^{3}$The University of Sydn
eudoraxj@gmail.com, zhaorui@sensetime.com, zhufengx@mail.ustc.edu.cn
huamingwang2014@gmail.com, w
$స్సి$
$ഇ$
### $\mathbf{Abstract}$
$\sum_{nearance.~t~is~a~challen}$
pearance. It is a challenging task due to large pose variations, complex background cl Recently, human pose estimation by predicting joint loca-
$\sum_{\begin{array}{lll}&\text{tions was largel}\end{array}}$ $\begin{array}{lll}{}improvementinReID~performance.~However,\end{array}$
been fully exploited for person RelD.
In this work, we introduce a Attention-Aware Compositional Network (AACN) for person ReID. AACN consists of t guided Part Attention (PPA) and Attention-aware Feature $Composition~(AFC).~PPA~is~learned~and~applied~to~mask$
$\begin{array}{cc}{\textcircled{1}}&{matedforbodypartstodealwithpartocclusioninthepro-}\\{\textcircled{\infty}}&{posedAFCmodule.Extensive~experiments~with~ablation}\\{\textcircled{1}}&{\textcircled{1}}\\{\textcircled{2}}&{\textcircled{1}}\\{\textcircled{2}}&{\textcircled{1}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{2}}&{\textcircled{2}}\\{\textcircled{}}\end{array}$
analysis show the effectiven of-the-art results are achie $includingMarket-150l,CUHKO3,CUHKOl,SenseReID$, $CUHKO3-NPandDukeMTMC-reID.$
### $1. \textbf{Introduction}$
Person re-identification (Re same individual across different camera views. Given an image containing a target pe set of images (gallery set), a ReID system is expected to rank the images from gallery with the query image. It has many important applications in video surveillance by sav forts in exhaustively searching for a target person from large amount of video sequences. For example, finding missing
$^{*}$Rui Zhao is the corresponding author.
m Lower Right Arm
Lower Left Leg m Upper Body
Lower Right Leg
Background
<FigureHere>
Figure 1.Part alignment challenges in person ReID.(a-c): Describing body parts by bounding boxes may introduce many irrelevant regions from background and other parts. Matching between features extracted from loose boxes in (a) and tight boxes in (b) would deteriorate the matching accuracy. A finer part region representation in (c) would help alleviate this problem. (d-f): The importance of different parts should be adaptively adjusted. Upper body is occluded by forearms in (d), the two forearms are all occluded in (e). Features from occluded part should be eliminated during matching, while salient visual cues like yellow backpack in (f) need to be emphasized.
elderly and children, and suspect tracking, $etc.$
Many research works have been proposed to improve the state-of-the-art performance of public ReID benchmarks. However, identifying the same individual across different camera views is still an unsolved task in intelligent video surveillance. It is difficult in that pedestrian images often suffer from complex background clutters, varying illumination conditions, uncontrollable camera settings, severe occlusions and large pose variations.
Viewpoint changes and pose variations cause uncontrolled misalignment between pedestrian images. As the improvement of human pose estimation [2,6], recent works [38,50,55] utilized pose estimation results to align body
|
[
[
0.1405029296875,
0.1337890625,
0.8310546875,
0.1337890625,
0.8310546875,
0.1483154296875,
0.1405029296875,
0.1483154296875
],
[
0.13232421875,
0.1817626953125,
0.2059326171875,
0.179931640625,
0.20751953125,
0.1988525390625,
0.134033203125,
0.2008056640625
],
[
0.25244140625,
0.1817626953125,
0.358642578125,
0.177978515625,
0.359375,
0.1956787109375,
0.253173828125,
0.199462890625
],
[
0.39453125,
0.18115234375,
0.483642578125,
0.18115234375,
0.483642578125,
0.1988525390625,
0.39453125,
0.1988525390625
],
[
0.53125,
0.18310546875,
0.6611328125,
0.18310546875,
0.6611328125,
0.1976318359375,
0.53125,
0.1976318359375
],
[
0.71337890625,
0.18310546875,
0.8310546875,
0.18310546875,
0.8310546875,
0.1976318359375,
0.71337890625,
0.1976318359375
],
[
0.39453125,
0.1962890625,
0.619140625,
0.1988525390625,
0.619140625,
0.216552734375,
0.39453125,
0.2139892578125
],
[
0.66015625,
0.1962890625,
0.8759765625,
0.199462890625,
0.8759765625,
0.2171630859375,
0.66015625,
0.2139892578125
],
[
0.1854248046875,
0.2215576171875,
0.78369140625,
0.2215576171875,
0.78369140625,
0.231689453125,
0.1854248046875,
0.231689453125
],
[
0.09478759765625,
0.1976318359375,
0.3544921875,
0.1988525390625,
0.3544921875,
0.216552734375,
0.09478759765625,
0.21533203125
],
[
0.25,
0.236083984375,
0.72119140625,
0.2374267578125,
0.72119140625,
0.251953125,
0.25,
0.250732421875
],
[
0.021240234375,
0.251953125,
0.05963134765625,
0.251953125,
0.062103271484375,
0.373779296875,
0.0236968994140625,
0.373779296875
],
[
0.238525390625,
0.287353515625,
0.313720703125,
0.287353515625,
0.313720703125,
0.3017578125,
0.238525390625,
0.3017578125
],
[
0.52783203125,
0.290283203125,
0.59326171875,
0.290283203125,
0.59326171875,
0.29736328125,
0.52783203125,
0.29736328125
],
[
0.61767578125,
0.292236328125,
0.6787109375,
0.292236328125,
0.6787109375,
0.296630859375,
0.61767578125,
0.296630859375
],
[
0.69921875,
0.29345703125,
0.7294921875,
0.29345703125,
0.7294921875,
0.296630859375,
0.69921875,
0.296630859375
],
[
0.7900390625,
0.291748046875,
0.88232421875,
0.291748046875,
0.88232421875,
0.298583984375,
0.7900390625,
0.298583984375
],
[
0.59326171875,
0.31494140625,
0.63134765625,
0.31494140625,
0.63134765625,
0.32568359375,
0.59326171875,
0.32568359375
],
[
0.73388671875,
0.31494140625,
0.76953125,
0.31494140625,
0.76953125,
0.32568359375,
0.73388671875,
0.32568359375
],
[
0.5908203125,
0.326904296875,
0.6337890625,
0.326904296875,
0.6337890625,
0.333984375,
0.5908203125,
0.333984375
],
[
0.098876953125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.098876953125,
0.333251953125
],
[
0.73193359375,
0.326904296875,
0.744140625,
0.330078125,
0.74169921875,
0.333251953125,
0.73046875,
0.330078125
],
[
0.08087158203125,
0.333251953125,
0.46728515625,
0.333984375,
0.46728515625,
0.348388671875,
0.08087158203125,
0.347900390625
],
[
0.74267578125,
0.32763671875,
0.775390625,
0.32763671875,
0.775390625,
0.33203125,
0.74267578125,
0.33203125
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.3642578125,
0.08087158203125,
0.3642578125
],
[
0.0269622802734375,
0.3623046875,
0.0555419921875,
0.3623046875,
0.0555419921875,
0.391357421875,
0.0269622802734375,
0.391357421875
],
[
0.080078125,
0.365478515625,
0.46728515625,
0.365478515625,
0.46728515625,
0.37890625,
0.080078125,
0.37890625
],
[
0.080078125,
0.37890625,
0.46728515625,
0.380126953125,
0.46728515625,
0.39453125,
0.080078125,
0.393310546875
],
[
0.08087158203125,
0.397705078125,
0.46728515625,
0.397705078125,
0.46728515625,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.564453125,
0.397705078125,
0.5810546875,
0.397705078125,
0.5810546875,
0.4072265625,
0.564453125,
0.4072265625
],
[
0.71337890625,
0.397705078125,
0.72802734375,
0.397705078125,
0.72802734375,
0.4072265625,
0.71337890625,
0.4072265625
],
[
0.8310546875,
0.396484375,
0.8505859375,
0.396484375,
0.8505859375,
0.4091796875,
0.8310546875,
0.4091796875
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.080078125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.080078125,
0.439453125
],
[
0.080078125,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.455078125,
0.080078125,
0.455078125
],
[
0.03350830078125,
0.45068359375,
0.0555419921875,
0.45068359375,
0.0555419921875,
0.479248046875,
0.03350830078125,
0.479248046875
],
[
0.080078125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.080078125,
0.4716796875,
0.324462890625,
0.4716796875,
0.324462890625,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.0269622802734375,
0.49365234375,
0.056365966796875,
0.49365234375,
0.056365966796875,
0.54248046875,
0.0269622802734375,
0.54248046875
],
[
0.098876953125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.498779296875,
0.098876953125,
0.498779296875
],
[
0.080078125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.55126953125,
0.50390625,
0.56640625,
0.50390625,
0.56640625,
0.51220703125,
0.55126953125,
0.51220703125
],
[
0.689453125,
0.505859375,
0.69921875,
0.505859375,
0.69921875,
0.51123046875,
0.689453125,
0.51123046875
],
[
0.81396484375,
0.50390625,
0.82666015625,
0.50390625,
0.82666015625,
0.51318359375,
0.81396484375,
0.51318359375
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.51806640625,
0.88818359375,
0.51806640625,
0.88818359375,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.08087158203125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.08087158203125,
0.54541015625
],
[
0.5048828125,
0.5341796875,
0.8857421875,
0.5341796875,
0.8857421875,
0.54443359375,
0.5048828125,
0.54443359375
],
[
0.5009765625,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.5009765625,
0.5595703125
],
[
0.02532958984375,
0.53515625,
0.056365966796875,
0.53515625,
0.056365966796875,
0.58837890625,
0.02532958984375,
0.58837890625
],
[
0.080078125,
0.54736328125,
0.4697265625,
0.54736328125,
0.4697265625,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.080078125,
0.5625,
0.4697265625,
0.5625,
0.4697265625,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.50341796875,
0.5625,
0.890625,
0.5625,
0.890625,
0.57275390625,
0.50341796875,
0.57275390625
],
[
0.080078125,
0.57763671875,
0.46728515625,
0.57763671875,
0.46728515625,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.50341796875,
0.57568359375,
0.88623046875,
0.57568359375,
0.88623046875,
0.5859375,
0.50341796875,
0.5859375
],
[
0.0269622802734375,
0.58837890625,
0.0555419921875,
0.58837890625,
0.0555419921875,
0.62060546875,
0.0269622802734375,
0.62060546875
],
[
0.080078125,
0.59228515625,
0.468017578125,
0.59228515625,
0.468017578125,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.50244140625,
0.58837890625,
0.89208984375,
0.58837890625,
0.89208984375,
0.6015625,
0.50244140625,
0.6015625
],
[
0.035125732421875,
0.6015625,
0.0400390625,
0.6015625,
0.0400390625,
0.609375,
0.035125732421875,
0.609375
],
[
0.080078125,
0.60791015625,
0.4697265625,
0.60791015625,
0.4697265625,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.5048828125,
0.60400390625,
0.88623046875,
0.60400390625,
0.88623046875,
0.6142578125,
0.5048828125,
0.6142578125
],
[
0.0294189453125,
0.61279296875,
0.052276611328125,
0.61279296875,
0.052276611328125,
0.63623046875,
0.0294189453125,
0.63623046875
],
[
0.50244140625,
0.6162109375,
0.89208984375,
0.6162109375,
0.89208984375,
0.6298828125,
0.50244140625,
0.6298828125
],
[
0.08087158203125,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.08087158203125,
0.63720703125
],
[
0.02532958984375,
0.63525390625,
0.0555419921875,
0.6337890625,
0.058013916015625,
0.708984375,
0.02777099609375,
0.71044921875
],
[
0.50341796875,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.080078125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.5009765625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.65283203125,
0.46728515625,
0.65283203125,
0.46728515625,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.50341796875,
0.6591796875,
0.66015625,
0.6591796875,
0.66015625,
0.66943359375,
0.50341796875,
0.66943359375
],
[
0.08087158203125,
0.66845703125,
0.32275390625,
0.66845703125,
0.32275390625,
0.681640625,
0.08087158203125,
0.681640625
],
[
0.50341796875,
0.7021484375,
0.80615234375,
0.7021484375,
0.80615234375,
0.71533203125,
0.50341796875,
0.71533203125
],
[
0.08087158203125,
0.7138671875,
0.2091064453125,
0.7138671875,
0.2091064453125,
0.72900390625,
0.08087158203125,
0.72900390625
],
[
0.521484375,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.73193359375,
0.521484375,
0.73193359375
],
[
0.09722900390625,
0.7392578125,
0.4697265625,
0.7392578125,
0.4697265625,
0.75244140625,
0.09722900390625,
0.75244140625
],
[
0.50341796875,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.7548828125,
0.4697265625,
0.7548828125,
0.4697265625,
0.76953125,
0.080078125,
0.76953125
],
[
0.50244140625,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.50244140625,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.080078125,
0.771484375,
0.468017578125,
0.771484375,
0.468017578125,
0.78515625,
0.080078125,
0.78515625
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.08087158203125,
0.787109375,
0.468017578125,
0.787109375,
0.468017578125,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.50341796875,
0.79443359375,
0.890625,
0.79443359375,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.07843017578125,
0.80029296875,
0.46728515625,
0.7998046875,
0.46728515625,
0.814453125,
0.07843017578125,
0.81494140625
],
[
0.5009765625,
0.80810546875,
0.890625,
0.8095703125,
0.890625,
0.82373046875,
0.5009765625,
0.82275390625
],
[
0.080078125,
0.81640625,
0.468017578125,
0.81640625,
0.468017578125,
0.830078125,
0.080078125,
0.830078125
],
[
0.50244140625,
0.82470703125,
0.73193359375,
0.82470703125,
0.73193359375,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.080078125,
0.83203125,
0.468017578125,
0.83203125,
0.468017578125,
0.84521484375,
0.080078125,
0.84521484375
],
[
0.080078125,
0.8466796875,
0.468017578125,
0.8466796875,
0.468017578125,
0.85986328125,
0.080078125,
0.85986328125
],
[
0.52294921875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.52294921875,
0.85498046875
],
[
0.07843017578125,
0.85986328125,
0.4697265625,
0.8623046875,
0.4697265625,
0.876953125,
0.07843017578125,
0.875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88818359375,
0.303955078125,
0.88818359375,
0.303955078125,
0.9013671875,
0.09722900390625,
0.9013671875
],
[
0.50341796875,
0.88623046875,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"Attention-Aware Compositional Network for Person Re-identification",
"Jing Xu!",
"Rui Zhaol,2*",
"Feng Zhu!",
"Huaming Wang",
"Wanli Ouyang",
"2SenseTime Group Limited",
"3The University of Sydney",
"eudoraxj@gmail.com,zhaorui@sensetime.com,zhufengx@mail.ustc.edu.cn",
"1SenseNets Technology Limited",
"huamingwang2014@gmail.com,wanli.ouyang@sydney.edu.au",
"81000",
"Abstract",
"LowerRightAm",
"owerRightTeo",
"O",
"erBodyBackground",
"Feature",
"Feature",
"Componen",
"Person re-identification (RelD) is to identify pedestrians",
"O",
"observed from different camera views based on visual ap-",
"mnonen",
"pearance. It is a challenging task due to large pose varia-",
"6",
"tions, complex background clutters and severe occlusions.",
"Recently, human pose estimation by predicting joint loca-",
"tionswaslargelyimprovedinaccuracy.Itisreasonableto",
"(a)",
"(b)",
"(c)",
"LAD",
"use pose estimation results for handling pose variations and",
"background clutters, and such attempts have obtained great",
"improvement in RelD performance. However, we argue that",
"S",
"the pose information was not well utilized and hasn't yet",
"beenfullyexploitedforpersonRelD.",
"t",
"In this work, we introduce a novel framework called",
"Attention-Aware Compositional Network (AACN)for per-",
"(d)",
"e",
"(f)",
"son RelD. AACN consists of two main components: Pose-",
"Figure 1. Part alignment challenges in person ReID. (a-c): De",
"guided Part Attention (PPA) and Attention-aware Feature",
"scribing bodyparts byboundingboxes mayintroduce manyirrele",
"vant regions from background and other parts. Matching between",
"0",
"Composition (AFC). PPA is learned and applied to mask",
"out undesirable background features in pedestrian feature",
"featuresextractedfromlooseboxesin(a)andtightboxesin(b)",
"maps.Furthermore,pose-guided visibility scores are esti-",
"would deterioratethematchingaccuracy.Afinerpartregionrep",
"0",
"mated for body parts to deal with part occlusion in the pro-",
"resentation in (c) would help alleviate this problem. (d-f): The",
"",
"posed AFC module. Extensive experiments with ablation",
"importanceofdifferentparts shouldbeadaptivelyadjusted.Up",
"80",
"per body is occluded by forearms in (d), the two forearms are all",
"analysis show the effectiveness of our method, and state-",
"AIXE",
"occluded in (e).Features from occluded part should be eliminated",
"of-the-art results are achieved on several public datasets,",
"during matching,while salient visual cues like yellow backpack in",
"including Market-1501, CUHK03, CUHK01, SenseRelD,",
"(f) need to be emphasized.",
"CUHKO3-NP and DukeMTMC-reID.",
"elderly and children, and suspect tracking, etc.",
"1. Introduction",
"Many research works have been proposed to improve the",
"Person re-identification (ReID) targets on identifying the",
"state-of-the-art performance of public ReID benchmarks.",
"same individual across different camera views. Given an",
"However, identifying the same individual across different",
"camera views is still an unsolved task in intelligent video",
"image containing a target person (as query) and a large",
"surveillance. It is difficult in that pedestrian images often",
"setofimages(galleryset),aRelDsystemisexpectedto",
"suffer from complex background clutters, varying illumina-",
"rank the images from gallery according to visual similarity",
"tion conditions, uncontrollable camera settings, severe oc-",
"with the query image. It has many important applications",
"clusions and large pose variations.",
"in video surveillance by saving large amount of human ef-",
"forts in exhaustively searching for a target person from large",
"Viewpoint changes and pose variations cause uncon-",
"amount of video sequences. For example, finding missing",
"trolled misalignment between pedestrian images. As the",
"improvement of human pose estimation [2, 6], recent works",
"*Rui Zhao is the corresponding author.",
"[38, 50, 55] utilized pose estimation results to align body"
] |
{
"arxiv_id": "1805.03344",
"language": "en",
"page_idx": 0
}
|
||
1805.03344-en-1
|
1805.03344-en-1.jpg
|
<FigureHere>
Figure 2. Our Pose-guided Part Attention precisely captures the target parts, excludes background clutter and adjacent part features, while pose-guided rectangular RoIs [50,38,55], patches [52,25], and stripes [35, 1] include extensive noise features.
parts for better matching. Although great improvement in performance was obtained, there are still noticeable problems in these methods. These methods deal with misalignment by extracting features from patches, stripes, or poseguided region of interest (RoI), where rectangular RoIs often introduce noise from adjacent parts or background in feature and lead to inaccurate matching. For example in Fig 1(a), features of the right leg is extracted from its bounding box in orange, which includes extensive noise from left leg and background, as shown in the second bar of the histogram. Features of the right arm and left leg also consist of their adjacent parts and background. Furthermore, some body parts have large variations in shape and pose, and rectangular RoI would include inconsistent extent of background clutter and adjacent noise. For example, the right leg is loosely included in the Fig. 1(a) and tightly contained in Fig. 1(b). Matching between part features from loose box and tight box in two camera views would definitely deteriorate the matching accuracy. To deal with these problems, finer silhouettes contouring body parts like in Fig. 1(c) are needed, so that part features can be extracted more precisely, alleviating the influence from background clutters and adjacent noises.
In this work, we propose to use Pose-guided Part Attention instead of rectangular RoI. Pose-guided Part Attention is a confdence map that could precisely capture the target part, and exclude background clutter and adjacent part features, as shown in Fig. 2. Attention-aware part features can be extracted by applying the part attention mask on feature maps, and feature alignment by part is naturally achieved. We will show in experiments that attention-aware part features are more accurate and robust, and the aligned pedestrian features are more discriminative than those proposed in conventional methods.
Occlusion is also a common and severe problem in practical ReID scenario. For example in Fig. 1(d-f), body part may be occluded by other body parts, adjacent persons or things like carrying baggage or trolley. Some observations can be concluded: 1) rigid body parts like head-shoulder, upper torso, lower torso are often partially occluded by ad-
jacent non-rigid parts like upper arms, lower legs, $etc.2)$ non-rigid body parts suffer heavy self-occlusion and are often fully occluded. 3) occlusion by carrying things is not a bad situation, which should be considered as a special part to help re-identification. It would be ideal to weaken features for partially occluded rigid part like the upper body in Fig. 1(d), eliminate features for fully occluded non-rigid part like the forearms in Fig. 1(e), and retain features for carrying things like the backpack in Fig. 1(f). Based on above observations on the occlusion problem, we propose a pose-guided visibility score to measure the occlusion extent for each body part, and it provides image-specific part importance score to decide feature importance in matching. Experimental results show its usefulness in handling occlusion cases.
Based on above motivations, a new Attention-Aware Compositional Network for person re-identification is proposed. The contributions of our work can be summarized in several folds:
$\bullet$ A unifed framework named Attention-Aware Compositional Network (AACN) is proposed to deal with misalignment and occlusion problem in person reidentification.
$\bullet$ Pose-guided Part Attention is introduced to estimate
finer part attention to exclude adjacent noise. It is designed to capture both rigid and non-rigid body parts simultaneously in a unified framework.
$\bullet$ Visibility score is introduced to measure the occlusion
extent for each body part. It provides image-specific part importance scores for Attention-aware Feature Composition.
$\bullet$ Extensive experiments demonstrate that our approach achieve superior performance on several public datasets, including CUHK03 [25], CUHK01 [24], Market-1501 [56], CUHK03-NP [61], DukeMTMCreID [33] and SenseReID [50].
## $2. \textbf{Related Work}$
### $2. 1. \textbf{ Person Re- identification}$
There are two categories of methods addressing the problem of person re-identification, namely feature representation and distance metric learning. The first category mainly includes the traditional feature descriptors [54,53,52,27,7,31,36] and deep learning features [44,42,43,12,25,41]. These approaches dedicate to design view-invariant representations for person images. The second category [28,12,27,46,19,9,23,14,34] mainly targets on learning a robust distance metric to measure the similarity between images.
Pedestrian alignment, matching two person images with their corresponding parts, is of non-trivial importance. Existing ReID methods mainly focus on extracting two types
|
[
[
0.414306640625,
0.09344482421875,
0.4306640625,
0.09344482421875,
0.4306640625,
0.10162353515625,
0.414306640625,
0.10162353515625
],
[
0.50341796875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.340576171875,
0.0947265625,
0.3544921875,
0.0947265625,
0.3544921875,
0.0985107421875,
0.340576171875,
0.0985107421875
],
[
0.26953125,
0.0947265625,
0.27685546875,
0.0947265625,
0.27685546875,
0.09783935546875,
0.26953125,
0.09783935546875
],
[
0.5048828125,
0.10919189453125,
0.8896484375,
0.10919189453125,
0.8896484375,
0.11932373046875,
0.5048828125,
0.11932373046875
],
[
0.50244140625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.50244140625,
0.1363525390625
],
[
0.50341796875,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.50341796875,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.50244140625,
0.1666259765625,
0.890625,
0.1673583984375,
0.890625,
0.1817626953125,
0.50244140625,
0.18115234375
],
[
0.50341796875,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.5048828125,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.5048828125,
0.210205078125
],
[
0.08087158203125,
0.2108154296875,
0.468017578125,
0.2108154296875,
0.468017578125,
0.2247314453125,
0.08087158203125,
0.2247314453125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2259521484375,
0.46728515625,
0.2247314453125,
0.46728515625,
0.238037109375,
0.080078125,
0.2392578125
],
[
0.50244140625,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.2392578125,
0.46728515625,
0.2392578125,
0.46728515625,
0.25244140625,
0.080078125,
0.25244140625
],
[
0.50341796875,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.080078125,
0.25244140625,
0.435546875,
0.25244140625,
0.435546875,
0.266357421875,
0.080078125,
0.266357421875
],
[
0.5009765625,
0.2568359375,
0.89208984375,
0.25830078125,
0.89208984375,
0.272705078125,
0.5009765625,
0.271484375
],
[
0.5009765625,
0.272705078125,
0.8896484375,
0.2734375,
0.8896484375,
0.28857421875,
0.5009765625,
0.287353515625
],
[
0.07843017578125,
0.294189453125,
0.468017578125,
0.29345703125,
0.468017578125,
0.30810546875,
0.07843017578125,
0.309326171875
],
[
0.5009765625,
0.28857421875,
0.8896484375,
0.28857421875,
0.8896484375,
0.3017578125,
0.5009765625,
0.3017578125
],
[
0.080078125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.323974609375,
0.080078125,
0.323974609375
],
[
0.50341796875,
0.304931640625,
0.57861328125,
0.304931640625,
0.57861328125,
0.31640625,
0.50341796875,
0.31640625
],
[
0.07843017578125,
0.323974609375,
0.468017578125,
0.324462890625,
0.468017578125,
0.339111328125,
0.07843017578125,
0.33837890625
],
[
0.52294921875,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.52294921875,
0.33203125
],
[
0.50341796875,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.08087158203125,
0.34228515625,
0.465576171875,
0.34228515625,
0.465576171875,
0.352783203125,
0.08087158203125,
0.352783203125
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.50341796875,
0.3642578125,
0.59228515625,
0.3642578125,
0.59228515625,
0.375732421875,
0.50341796875,
0.375732421875
],
[
0.08087158203125,
0.37255859375,
0.468017578125,
0.37255859375,
0.468017578125,
0.38330078125,
0.08087158203125,
0.38330078125
],
[
0.08087158203125,
0.386474609375,
0.46728515625,
0.386474609375,
0.46728515625,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.52685546875,
0.3837890625,
0.8896484375,
0.3837890625,
0.8896484375,
0.39453125,
0.52685546875,
0.39453125
],
[
0.08251953125,
0.401611328125,
0.46728515625,
0.401611328125,
0.46728515625,
0.415283203125,
0.08251953125,
0.415283203125
],
[
0.53515625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.412353515625,
0.53515625,
0.412353515625
],
[
0.080078125,
0.416748046875,
0.468017578125,
0.416748046875,
0.468017578125,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.53515625,
0.412353515625,
0.89208984375,
0.41357421875,
0.89208984375,
0.427978515625,
0.53515625,
0.4267578125
],
[
0.080078125,
0.43115234375,
0.46728515625,
0.43115234375,
0.46728515625,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.53759765625,
0.4287109375,
0.62744140625,
0.4287109375,
0.62744140625,
0.43994140625,
0.53759765625,
0.43994140625
],
[
0.080078125,
0.44580078125,
0.468017578125,
0.44580078125,
0.468017578125,
0.458984375,
0.080078125,
0.458984375
],
[
0.521484375,
0.443115234375,
0.89208984375,
0.443115234375,
0.89208984375,
0.45703125,
0.521484375,
0.45703125
],
[
0.080078125,
0.4609375,
0.468017578125,
0.4609375,
0.468017578125,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.53515625,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.472900390625,
0.53515625,
0.472900390625
],
[
0.080078125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.5361328125,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.5361328125,
0.488037109375
],
[
0.080078125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.080078125,
0.5048828125
],
[
0.5361328125,
0.4912109375,
0.79150390625,
0.4912109375,
0.79150390625,
0.501953125,
0.5361328125,
0.501953125
],
[
0.080078125,
0.5068359375,
0.468017578125,
0.5068359375,
0.468017578125,
0.52099609375,
0.080078125,
0.52099609375
],
[
0.521484375,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.521484375,
0.51806640625
],
[
0.080078125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.53515625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.53515625,
0.5341796875
],
[
0.080078125,
0.537109375,
0.468017578125,
0.537109375,
0.468017578125,
0.55029296875,
0.080078125,
0.55029296875
],
[
0.5361328125,
0.53515625,
0.8935546875,
0.53515625,
0.8935546875,
0.548828125,
0.5361328125,
0.548828125
],
[
0.080078125,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.53515625,
0.55029296875,
0.6259765625,
0.55029296875,
0.6259765625,
0.56494140625,
0.53515625,
0.56494140625
],
[
0.080078125,
0.5673828125,
0.468017578125,
0.5673828125,
0.468017578125,
0.58056640625,
0.080078125,
0.58056640625
],
[
0.521484375,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.58056640625,
0.521484375,
0.58056640625
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.5361328125,
0.58056640625,
0.890625,
0.58056640625,
0.890625,
0.5947265625,
0.5361328125,
0.5947265625
],
[
0.080078125,
0.59765625,
0.468017578125,
0.59765625,
0.468017578125,
0.611328125,
0.080078125,
0.611328125
],
[
0.53369140625,
0.5947265625,
0.890625,
0.5966796875,
0.890625,
0.60986328125,
0.53369140625,
0.60791015625
],
[
0.08087158203125,
0.611328125,
0.468017578125,
0.611328125,
0.468017578125,
0.625,
0.08087158203125,
0.625
],
[
0.53515625,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.53515625,
0.625
],
[
0.07843017578125,
0.62548828125,
0.2646484375,
0.62744140625,
0.2646484375,
0.6416015625,
0.07843017578125,
0.6396484375
],
[
0.53515625,
0.626953125,
0.73876953125,
0.626953125,
0.73876953125,
0.64013671875,
0.53515625,
0.64013671875
],
[
0.098876953125,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.098876953125,
0.6572265625
],
[
0.5009765625,
0.65283203125,
0.6435546875,
0.65478515625,
0.6435546875,
0.67236328125,
0.5009765625,
0.67041015625
],
[
0.080078125,
0.6591796875,
0.4697265625,
0.6591796875,
0.4697265625,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.08087158203125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.50341796875,
0.68115234375,
0.72119140625,
0.68115234375,
0.72119140625,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.080078125,
0.70263671875,
0.4697265625,
0.7041015625,
0.4697265625,
0.71826171875,
0.080078125,
0.71728515625
],
[
0.521484375,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.521484375,
0.71826171875
],
[
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,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.08087158203125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50244140625,
0.75,
0.89208984375,
0.74853515625,
0.89208984375,
0.76318359375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.080078125,
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,
0.77880859375,
0.890625,
0.77978515625,
0.890625,
0.79443359375,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.796875,
0.2459716796875,
0.796875,
0.2459716796875,
0.806640625,
0.08087158203125,
0.806640625
],
[
0.5009765625,
0.79541015625,
0.89208984375,
0.79345703125,
0.89208984375,
0.80810546875,
0.5009765625,
0.81005859375
],
[
0.098876953125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.5009765625,
0.81005859375,
0.8896484375,
0.8095703125,
0.8896484375,
0.82373046875,
0.5009765625,
0.82470703125
],
[
0.07843017578125,
0.82470703125,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82470703125,
0.89208984375,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.68115234375,
0.84130859375,
0.68115234375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85498046875,
0.4697265625,
0.85595703125,
0.4697265625,
0.87060546875,
0.07843017578125,
0.869140625
],
[
0.521484375,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.521484375,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Stri",
"jacent non-rigid parts like upper arms, lower legs, etc. 2)",
"Sato",
"-",
"non-rigidbodyparts sufferheavy self-occlusionand areof-",
"ten fully occluded. 3) occlusion by carrying things is not a",
"bad situation, which should be considered as a special part",
"to help re-identification. It would be ideal to weaken fea-",
"tures for partially occluded rigid part like the upper body",
"in Fig. 1(d), eliminate features for fully occluded non-rigid",
"part like the forearms inFig. 1(e), and retain features for",
"Figure 2. Our Pose-guided Part Attention precisely captures the",
"carrying things like the backpack in Fig: 1(f). Based on",
"target parts, excludes background clutter and adjacent part fea-",
"above observations on the occlusion problem, we propose",
"tures, while pose-guided rectangular Rols [50, 38, 55], patches",
"a pose-guided visibility score to measure the occlusion ex-",
"[52, 25], and stripes [35, 1] include extensive noise features.",
"tent for each body part, and it provides image-specific part",
"importance scoretodecidefeature importance inmatching.",
"parts for better matching.Although great improvement in",
"Experimental results show its usefulness in handling occlu-",
"performance was obtained, there are still noticeable prob-",
"sion cases.",
"lems in these methods. These methods deal with misalign-",
"Based on above motivations, a new Attention-Aware",
"Compositional Network for person re-identification is pro-",
"mentby extractingfeatures frompatches,stripes,orpose-",
"posed. The contributions of our work can be summarized in",
"guided region of interest (Rol), where rectangular Rols of-",
"several folds:",
"ten introduce noisefrom adjacentparts or background in",
"feature and lead to inaccurate matching. For example in Fig.",
"AunifiedframeworknamedAttention-AwareCom-",
"I(a), features of the right leg is extracted from its bound-",
"positional Network (AACN) is proposed to deal with",
"ing box in orange, which includes extensive noise from left",
"misalignment and occlusion problem in person re-",
"leg and background, as shown in the second bar of the his-",
"identification.",
"togram. Features of the right arm and left leg also con-",
". Pose-guided Part Attention is introduced to estimate",
"sist of their adjacent parts and background. Furthermore,",
"finer part attention to exclude adjacent noise. It is de-",
"some body parts have large variations in shape and pose,",
"signed to capture both rigid and non-rigid body parts",
"and rectangular Rol would include inconsistent extent of",
"simultaneouslyin a unified framework.",
"background clutter and adjacent noise. For example, the",
"·Visibility score is introduced to measure the occlusion",
"right leg is loosely included in the Fig. l(a) and tightly",
"extent for each body part. It provides image-specific",
"contained in Fig.1(b). Matching between part features",
"part importance scores for Attention-aware Feature",
"from loose box and tight box in two camera views would",
"Composition.",
"definitely deteriorate the matching accuracy.To deal with",
"·Extensive experiments demonstrate that our ap-",
"these problems, finer silhouettes contouring body parts like",
"proach achieve superior performance on several pub-",
"in Fig. 1(c) are needed, so that part features can be extracted",
"lic datasets, including CUHK03 [25], CUHK01 [24],",
"more precisely, alleviating the influence from background",
"Market-1501 [56], CUHK03-NP [61], DukeMTMC-",
"clutters and adjacent noises.",
"reID [33] and SenseReID [50].",
"In this work, we propose to use Pose-guided Part Atten-",
"2.Related Work",
"tion instead of rectangular Rol. Pose-guided Part Attention",
"is a confidence map that couldprecisely capture thetarget",
"2.1. Person Re-identification",
"part, and exclude background clutter and adjacent part fea-",
"tures, as shown in Fig. 2. Attention-aware part features can",
"There are two categories of methods addressing the",
"be extracted by applying the part attention mask on feature",
"problem of person re-identification, namely feature rep-",
"maps, and feature alignment by part is naturally achieved.",
"resentation and distance metric learning. The first cat-",
"We will show in experiments that attention-aware part fea-",
"egory mainly includes the traditional feature descriptors",
"tures are more accurate and robust, and the aligned pedes-",
"[54, 53, 52, 27, 7, 31, 36] and deep learning features",
"trian features are more discriminative than those proposed",
"[44, 42, 43, 12, 25, 41]. These approaches dedicate to de-",
"inconventionalmethods.",
"sign view-invariant representations for person images. The",
"Occlusion is also a common and severe problem in prac-",
"second category [28, 12, 27, 46, 19, 9, 23, 14, 34] mainly",
"tical ReID scenario. For example in Fig. 1(d-f), body part",
"targets on learning a robust distance metric to measure the",
"may be occluded by other body parts, adjacent persons or",
"similarity between images.",
"things like carrying baggage or trolley. Some observations",
"Pedestrian alignment, matching two person images with",
"can be concluded: 1) rigid body parts like head-shoulder,",
"their corresponding parts, is of non-trivial importance. Ex-",
"upper torso, lower torso are often partially occluded by ad-",
"isting RelD methods mainly focus on extracting two types"
] |
{
"arxiv_id": "1805.03344",
"language": "en",
"page_idx": 1
}
|
||
1805.03344-en-2
|
1805.03344-en-2.jpg
|
of features, namely global features extracted from the whole image [11,45] and region features generated from local patches [52, 58, 22]. However, these approaches have not taken the accurate alignment of body regions into consideration. Recently, thanks to the great progress of pose estimation methods [6, 13] and RPN [32], reliable body parts are able to be acquired, which makes it possible to identify individuals via extracted region. For example, Zhao et al. [50] proposed Spindle Net, that extracted and fused three level part features. Parts were extracted by PRN. Su et al.[38] proposed a Pose-driven Deep Convolutional model (PDC) that utilized Spatial Transformer Network (STN) to localize and crop body regions based on pre-defined centers. Zheng et al. [55] introduced to extract Pose Invariant Embedding (PIE) through aligning pedestrians to standard pose. Alignment was done by applying affine transformation to pose estimation results. However, these methods are all based on rigid body regions, which cannot accurately localize human body regions. In our model, non-rigid parts are obtained based on the connectivity between human joints. Thus our model is capable of extracting more precise information for each body part, and handling occlusion issues.
# 2.2. Human Parsing
Human parsing [17,10,48,26,15] is related to our work in that parsing results can accurately localize body part. For example, Gong et al. [17] imposed joint structure loss to improve segmentation results. Dong et al.[15] explored pose information to guide human parsing. However, we choose to generate non-rigid parts based on connectivity of human keypoints rather than human parsing because of the following reason: Existing human parsing methods mainly focus on particular scenarios, such as fashion pictures, and the parsing models often show weak generalization on surveillance data. Human pose is easier to label than parsing, and it can be better generalized to surveillance scenario owing to large variance of the datasets [2,29].
## 2.3. Attention based Image Analysis
Since the attention mechanism is effective in understanding images, it has been widely used in various tasks, including machine translation [4], visual question answering [47], object detection [3], semantic segmentation [8], pose estimation [13] and person re-identification [30]. Bahdanau [4] and Ba [3] adopted recurrent neural networks (RNN) to generate the attention map for an image region at each step, and combined information from different steps overtime to make the final decision. Chen $etal.[8]$ introduced an attention mechanism that learned to softly weight multi-scale features at each pixel location. Chu et al.[13] proposed a multi-context attention model for pose estimation. Inspired by the methods mentioned above, we propose to learn attention map to capture human body part, and align features
across different person images by masking with part attentions. Our attention map is learned guided by pose estimation, and it can contour the shape of part more precisely than rectangular RoI. Furthermore, the intensity of part attention infers the visibility of each part, which helps to deal with part occlusion issues.
3. Attention- Aware Compositional Network
The framework of our Attention-Aware Compositional Network (AACN) is illustrated in Fig. 3. AACN consists of two main components: 1) Pose-guided Part Attention (PPA) and 2) Attention-aware Feature Composition (AFC). Given one person image, the proposed PPA module aims to estimate an attention map and a visibility score for each pre-defined body part. Then, part feature alignment and weighted fusion are performed in AFC module, given attention maps and visibility scores from PPA. PPA and AFC are tightly integrated in our framework during both training and testing phases.
The PPA module considers two types of pre-defined body parts, namely, non-rigid parts and rigid parts. Due to the variations in appearance, attentions of these two types of parts are estimated separately. The PPA module is constructed by a two-stage three-branch neural network, which predicts confidence maps of keypoints, attention maps of non-rigid parts, and attention maps of rigid parts in the three branches, respectively. A visibility score is further estimated for each part based on part attention maps.
The AFC module applies the estimated part attention maps to mask the global feature map produced by a base network (GoogleNet [40] is used in this work). The resulting attention-aware part features are then weightedly fused with the guidance from part visibility scores. The final 1024-dimensional feature vector is adopted as the representation of the input person image.
## $3. 1. \textbf{Pose- guided Part Attention}$
Part attentions are denoted by normalized part confidence maps, which highlight specific regions of human body in the image. As shown in Fig. 4(a), there are two types of human body parts: rigid parts and non-rigid parts. Limb regions including upper arms, lower arms, upper legs, and lower legs are called non-rigid parts because of drastic pose variations they could occur, while trunk parts of human body including head-shoulder, upper torso, and lower torso are considered to be rigid. Attention maps of the two types of parts are simultaneously learned in a unified form through our proposed Pose-guided Part Attention network.
Inspired by the multi-stage CNN[6] for human pose estimation, we utilize a two-stage network to learn part attentions. The first stage individually predicts non-rigid part attentions N, rigid part attentions $\mathbf{R}$, and keypoint confidence
|
[
[
0.08087158203125,
0.09027099609375,
0.468017578125,
0.09027099609375,
0.468017578125,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.080078125,
0.1383056640625,
0.471435546875,
0.1383056640625,
0.471435546875,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.1514892578125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1666259765625
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.50244140625,
0.1685791015625,
0.64453125,
0.1685791015625,
0.64453125,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.08251953125,
0.199462890625,
0.465576171875,
0.199462890625,
0.465576171875,
0.210205078125,
0.08251953125,
0.210205078125
],
[
0.50244140625,
0.2008056640625,
0.865234375,
0.2008056640625,
0.865234375,
0.2139892578125,
0.50244140625,
0.2139892578125
],
[
0.080078125,
0.21337890625,
0.46728515625,
0.21337890625,
0.46728515625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.07843017578125,
0.2279052734375,
0.468017578125,
0.2266845703125,
0.468017578125,
0.2412109375,
0.07843017578125,
0.242431640625
],
[
0.52294921875,
0.2266845703125,
0.890625,
0.2266845703125,
0.890625,
0.240478515625,
0.52294921875,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.240478515625,
0.89208984375,
0.2412109375,
0.89208984375,
0.255615234375,
0.5009765625,
0.255126953125
],
[
0.08087158203125,
0.260009765625,
0.46728515625,
0.260009765625,
0.46728515625,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.5048828125,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.26953125,
0.5048828125,
0.26953125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.271484375,
0.890625,
0.271484375,
0.890625,
0.285400390625,
0.50341796875,
0.285400390625
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.50244140625,
0.285888671875,
0.89208984375,
0.285400390625,
0.89208984375,
0.2998046875,
0.50244140625,
0.300537109375
],
[
0.08087158203125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.50244140625,
0.3017578125,
0.890625,
0.3017578125,
0.890625,
0.31494140625,
0.50244140625,
0.31494140625
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50341796875,
0.317626953125,
0.89208984375,
0.317626953125,
0.89208984375,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.08087158203125,
0.335205078125,
0.46728515625,
0.335205078125,
0.46728515625,
0.34521484375,
0.08087158203125,
0.34521484375
],
[
0.50341796875,
0.333251953125,
0.890625,
0.333251953125,
0.890625,
0.3466796875,
0.50341796875,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50244140625,
0.347900390625,
0.89208984375,
0.347900390625,
0.89208984375,
0.361083984375,
0.50244140625,
0.361083984375
],
[
0.080078125,
0.3642578125,
0.471435546875,
0.3642578125,
0.471435546875,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.363037109375,
0.8935546875,
0.363037109375,
0.8935546875,
0.376953125,
0.50244140625,
0.376953125
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.5009765625,
0.376953125,
0.62744140625,
0.37744140625,
0.62744140625,
0.392578125,
0.5009765625,
0.391357421875
],
[
0.08087158203125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.406005859375,
0.08087158203125,
0.406005859375
],
[
0.521484375,
0.393310546875,
0.89208984375,
0.393310546875,
0.89208984375,
0.40771484375,
0.521484375,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.3857421875,
0.410400390625,
0.3857421875,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.4091796875,
0.89208984375,
0.4091796875,
0.89208984375,
0.42236328125,
0.50244140625,
0.42236328125
],
[
0.50341796875,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50341796875,
0.438232421875
],
[
0.07843017578125,
0.434326171875,
0.2344970703125,
0.43505859375,
0.2344970703125,
0.449462890625,
0.07843017578125,
0.448974609375
],
[
0.50244140625,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.096435546875,
0.45703125,
0.471435546875,
0.455810546875,
0.471435546875,
0.473388671875,
0.096435546875,
0.474853515625
],
[
0.50244140625,
0.453857421875,
0.89208984375,
0.45263671875,
0.89208984375,
0.46728515625,
0.50244140625,
0.468505859375
],
[
0.080078125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.50341796875,
0.4697265625,
0.89208984375,
0.4697265625,
0.89208984375,
0.48291015625,
0.50341796875,
0.48291015625
],
[
0.080078125,
0.490478515625,
0.468017578125,
0.490478515625,
0.468017578125,
0.50390625,
0.080078125,
0.50390625
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.5048828125,
0.46728515625,
0.5048828125,
0.46728515625,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.50341796875,
0.5,
0.89208984375,
0.5,
0.89208984375,
0.51318359375,
0.50341796875,
0.51318359375
],
[
0.08087158203125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.5341796875,
0.08087158203125,
0.5341796875
],
[
0.5009765625,
0.51513671875,
0.8505859375,
0.51513671875,
0.8505859375,
0.52880859375,
0.5009765625,
0.52880859375
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.521484375,
0.53076171875,
0.89208984375,
0.53076171875,
0.89208984375,
0.54443359375,
0.521484375,
0.54443359375
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.080078125,
0.564453125
],
[
0.50244140625,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.07843017578125,
0.564453125,
0.4697265625,
0.56494140625,
0.4697265625,
0.5791015625,
0.07843017578125,
0.57763671875
],
[
0.50341796875,
0.56103515625,
0.8896484375,
0.56103515625,
0.8896484375,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.5771484375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.50244140625,
0.59033203125,
0.890625,
0.59033203125,
0.890625,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.08087158203125,
0.5966796875,
0.46728515625,
0.5966796875,
0.46728515625,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.60498046875,
0.890625,
0.60595703125,
0.890625,
0.62060546875,
0.50244140625,
0.619140625
],
[
0.080078125,
0.62548828125,
0.46728515625,
0.62548828125,
0.46728515625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50244140625,
0.6220703125,
0.7158203125,
0.6220703125,
0.7158203125,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.080078125,
0.6416015625,
0.419189453125,
0.6416015625,
0.419189453125,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.50244140625,
0.6484375,
0.74267578125,
0.6484375,
0.74267578125,
0.662109375,
0.50244140625,
0.662109375
],
[
0.08087158203125,
0.66650390625,
0.35693359375,
0.66650390625,
0.35693359375,
0.6796875,
0.08087158203125,
0.6796875
],
[
0.521484375,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.521484375,
0.6875
],
[
0.096435546875,
0.689453125,
0.468017578125,
0.68798828125,
0.468017578125,
0.70263671875,
0.096435546875,
0.7041015625
],
[
0.5009765625,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.734375
],
[
0.50341796875,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.08087158203125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.08087158203125,
0.75
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.08087158203125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50341796875,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.5009765625,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.79345703125,
0.5009765625,
0.79345703125
],
[
0.08087158203125,
0.79736328125,
0.4697265625,
0.79736328125,
0.4697265625,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.50244140625,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.890625,
0.8095703125,
0.890625,
0.82373046875,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82763671875,
0.46728515625,
0.82763671875,
0.46728515625,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.50244140625,
0.82568359375,
0.88623046875,
0.82568359375,
0.88623046875,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.521484375,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
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,
0.471435546875,
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
]
] |
[
"of features, namely global features extracted from the whole",
"across different person images by masking with part atten-",
"image[11,45l and region features generated from local",
"tions. Our attention map is learned guided by pose estima-",
"patches [52, 58, 22]. However, these approaches have not",
"tion,anditcancontourtheshapeofpartmorepreciselythan",
"taken the accurate alignment of body regions into consider-",
"rectangular Rol. Furthermore, the intensity of part attention",
"ation. Recently, thanks to the great progress of pose estima-",
"infers the visibility of each part, which helps to deal with",
"tion methods [6, 13] and RPN [32], reliable body parts are",
"part occlusion issues.",
"able to be acquired, which makes it possible to identify indi-",
"vidualsvia extractedregion.Forexample,Zhaoetal.[5o",
"3. Attention-Aware Compositional Network",
"proposed Spindle Net, that extracted and fused three level",
"part features. Parts were extracted by PRN. Su et al. [38]",
"The framework of our Attention-Aware Compositional",
"proposed a Pose-driven Deep Convolutional model (PDC)",
"Network (AACN) is illustrated in Fig. 3. AACN consists",
"thatutilizedSpatialTransformerNetwork(STN)tolocalize",
"of twomaincomponents:1)Pose-guidedPartAttention",
"and crop body regions based on pre-defined centers. Zheng",
"(PPA) and 2) Attention-aware Feature Composition (AFC).",
"et al. [55] introduced to extract Pose Invariant Embedding",
"Given one person image, the proposed PPA module aims",
"(PIE) through aligning pedestrians to standard pose. Align-",
"to estimate an attention map and a visibility score for each",
"ment was done by applying affine transformation to pose",
"pre-defined body part. Then, part feature alignment and",
"estimationresults.However,thesemethodsareallbasedon",
"weighted fusion are performed in AFC module, given at-",
"rigid body regions, which cannot accurately localize human",
"tention maps and visibility scores from PPA. PPA and AFC",
"body regions.In our model, non-rigid parts are obtained",
"are tightly integrated in our framework during both training",
"based on theconnectivitybetween human joints.Thus our",
"and testing phases.",
"modeliscapableofextractingmorepreciseinformationfor",
"The PPA module considers two types of pre-defined",
"each body part, and handling occlusion issues.",
"body parts, namely, non-rigid parts and rigid parts. Due to",
"the variations in appearance, attentions of these two types",
"2.2.Human Parsing",
"of parts are estimated separately. The PPA module is con-",
"Human parsing [17, 10, 48, 26, 15] is related to our",
"structed by a two-stage three-branch neural network, which",
"work in that parsing results can accurately localize body",
"predicts confidence maps of keypoints, attention maps of",
"part. For example, Gong et al. [17] imposed joint structure",
"non-rigid parts, and attention maps of rigid parts in the",
"loss to improve segmentation results. Dong et al. [15] ex-",
"three branches, respectively. A visibility score is further",
"plored pose information to guide human parsing.However,",
"estimated for each part based on part attention maps.",
"we choose to generate non-rigid parts based on connectiv-",
"The AFC module applies the estimated part attention",
"ity of human keypoints rather than human parsing because",
"maps to mask the global feature map produced by a base",
"of thefollowingreason:Existinghumanparsingmethods",
"network (GoogleNet [4O] is used in this work). The result-",
"mainly focus on particular scenarios, such as fashion pic-",
"ing attention-aware part features are then weightedly fused",
"with the guidance from part visibility scores.The final",
"tures,andthe parsing models often show weak generaliza-",
"tion on surveillance data. Human pose is easier to label than",
"1024-dimensional feature vector is adopted as the represen-",
"parsing, and it can be better generalized to surveillance sce-",
"tation of the input person image.",
"nario owing to large variance of the datasets [2, 29].",
"3.1. Pose-guided Part Attention",
"2.3. Attention based Image Analysis",
"Part attentions are denoted by normalized part confi-",
" Since the attention mechanism is effective in understand-",
"dence maps, which highlight specific regions of human",
"ing images, it has been widely used in various tasks, in-",
"body in the image. As shown in Fig. 4(a), there are two",
"cluding machine translation [4], visual question answering",
"types of human body parts: rigid parts and non-rigid parts.",
"[47], object detection [3], semantic segmentation [8], pose",
"Limb regions including upper arms, lower arms, upper legs",
"estimation [13] and person re-identification [30]. Bahdanau",
"and lower legs are called non-rigid parts because of drastic",
"[4] and Ba [3] adopted recurrent neural networks (RNN) to",
"pose variations they could occur, while trunk parts of hu-",
"generate the attention map for an image region at each step,",
"man body including head-shoulder, upper torso, and lower",
"and combined information from different steps overtime to",
"torso are considered to be rigid. Attention maps of the two",
"make the final decision. Chen et al. [8] introduced an at-",
"types of parts are simultaneously learned in a unified form",
"tention mechanism thatlearned to softly weight multi-scale",
"through our proposed Pose-guided Part Attention network.",
"features at each pixel location. Chu et al. [13] proposed a",
"Inspired by the multi-stage CNN [6] for human pose es-",
"multi-context attention modelforpose estimation.Inspired",
"timation, we utilize a two-stage network to learn part atten-",
"by the methods mentioned above, we propose to learn at-",
"tions. The first stage individually predicts non-rigid part at-",
"tention map to capture human body part, and align features",
"tentions N, rigid part attentions R, and keypoint confidence"
] |
{
"arxiv_id": "1805.03344",
"language": "en",
"page_idx": 2
}
|
||
2204.10448-en-1
|
2204.10448-en-1.jpg
|
methods (Narasimhan et al., 2018; Zhu et al., 2020) to learn to selectively focus on necessary pieces of knowledge. The memory-based methods represent knowledge facts in a form of memory and calculate soft attention scores of each memory with respect to a question. Then, it infers an answer by attending to knowledge evidence with high attention scores. On the other hand, to explicitly consider relational structure between knowledge facts, graph-based methods construct a query-aware knowledge graph by retrieving facts from KB and perform graph reasoning for a question. These methods mainly adopt an iterative message passing process to propagate information between adjacent nodes in the graph. However, it is difficult to capture multi-hop relationships containing long-distance nodes from the graph due to the well-known over-smoothing problem, where repetitive message passing process to propagate information across long distance makes features of connected nodes too similar and undiscriminating (Li et al., 2018; Wang et al., 2020).
To address the above limitation, we propose a novel method, Hypergraph Transformer, which exploits hypergraph structure to encode multi-hop relationships and transformer-based attention mechanism to learn to pay attention to important knowledge evidences for a question. We construct a question hypergraph and a knowledge hypergraph to explicitly encode high-order semantics present in the question and each knowledge fact, and capture multi-hop relational knowledge facts effectively. Then, we perform hyperedge matching between the two hypergraphs by leveraging transformer-based attention mechanism. We argue that introducing the concept of hypergraph is powerful for multi-hop reasoning problem in that it can encode high-order semantics without the constraint of length and learn cross-modal high-order associations.
The main contributions of this paper can be summarized as follows. i) We propose Hypergraph Transformer which enhances multi-hop reasoning ability by encoding high-order semantics in the form of a hypergraph and learning inter- and intrahigh-order associations in hypergraphs using the attention mechanism. ii) We conduct extensive experiments on two knowledge-based VQA datasets (KVQA and FVQA) and two knowledge-based textual QA datasets (PQ and PQL) and show superior performances on all datasets, especially multi-hop reasoning problem. iii) We qualitatively observe that Hypergraph Transformer performs robust in-
ference by focusing on correct reasoning evidences
under weak supervision.
## 2 $\textbf{Related Work}$
Knowledge-based visual question answering (Wang et al., 2017, 2018; Shah et al., 2019; Marino et al., 2019; Sampat et al., 2020) proposed benchmark datasets for knowledge-based visual question answering that requires reasoning about an image on the basis of facts from a large-scale knowledge base (KB) such as Freebase (Bollacker et al., 2008) or DBPedia (Auer et al., 2007). To solve the task, two pioneering studies (Wang et al., 2017, 2018) suggested logical parsing-based methods which convert a question to a KB logic query using predefined query templates and execute the generated query on KB for searching an answer. Since then information retrieval-based methods which retrieve knowledge facts associated with a question and conduct semantic matching between the facts and the question are introduced. (Narasimhan and Schwing, 2018; Shah et al., 2019) proposed memory-based methods that represent knowledge facts in the form of memory and calculate soft attention scores of the memory with a question. (Narasimhan et al., 2018; Zhu et al., 2020) represented the retrieved facts as a graph and performed graph reasoning through message passing scheme utilizing graph convolution. However, these methods are complicated to encode inherent high-order semantics and multi-hop relationships present in the knowledge graph. Therefore, we introduce a concept of hypergraph and propose transformer-based attention mechanism over hypergraphs.
Multi-hop knowledge graph reasoning is a process of sequential reasoning based on multiple evidences of a knowledge graph, and has been broadly used in various downstream tasks such as question answering (Lin et al., 2019; Saxena et al., 2020; Han et al., 2020b,a; Yadati et al., 2021), or knowledge-enhanced text generation (Liu et al., 2019; Moon et al., 2019; Ji et al., 2020). Recent researches have introduced the concept of hypergraph for multi-hop graph reasoning (Kim et al., 2020; Han et al., 2020b,a; Yadati et al., 2019, 2021; Sun et al., 2020). These models have a similar motivation to the Hypergraph Transformer proposed in this paper, but core operations are vastly different. These models mainly update node representations in the hypergraph through a message passing process using graph convolution operation. On the
|
[
[
0.1158447265625,
0.08489990234375,
0.48779296875,
0.0843505859375,
0.48779296875,
0.09857177734375,
0.1158447265625,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1158447265625,
0.10272216796875,
0.489501953125,
0.10272216796875,
0.489501953125,
0.11639404296875,
0.1158447265625,
0.11639404296875
],
[
0.5087890625,
0.1015625,
0.69189453125,
0.1015625,
0.69189453125,
0.11517333984375,
0.5087890625,
0.11517333984375
],
[
0.1158447265625,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51123046875,
0.12890625,
0.6650390625,
0.12890625,
0.6650390625,
0.1436767578125,
0.51123046875,
0.1436767578125
],
[
0.118408203125,
0.1519775390625,
0.48779296875,
0.1519775390625,
0.48779296875,
0.162109375,
0.118408203125,
0.162109375
],
[
0.509765625,
0.15380859375,
0.8818359375,
0.155029296875,
0.8818359375,
0.169189453125,
0.509765625,
0.168701171875
],
[
0.1158447265625,
0.1650390625,
0.48779296875,
0.166259765625,
0.48779296875,
0.18115234375,
0.1158447265625,
0.179931640625
],
[
0.509765625,
0.171630859375,
0.8818359375,
0.171630859375,
0.8818359375,
0.185302734375,
0.509765625,
0.185302734375
],
[
0.118408203125,
0.18408203125,
0.48779296875,
0.18408203125,
0.48779296875,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.5087890625,
0.1859130859375,
0.88330078125,
0.18701171875,
0.88330078125,
0.201904296875,
0.5087890625,
0.20068359375
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.5087890625,
0.2037353515625,
0.8818359375,
0.2037353515625,
0.8818359375,
0.21728515625,
0.5087890625,
0.21728515625
],
[
0.11419677734375,
0.21435546875,
0.48779296875,
0.213134765625,
0.48779296875,
0.22802734375,
0.11419677734375,
0.2286376953125
],
[
0.51318359375,
0.2203369140625,
0.8818359375,
0.2203369140625,
0.8818359375,
0.2340087890625,
0.51318359375,
0.2340087890625
],
[
0.1175537109375,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.509765625,
0.2369384765625,
0.87890625,
0.2369384765625,
0.87890625,
0.2476806640625,
0.509765625,
0.2476806640625
],
[
0.1175537109375,
0.2476806640625,
0.491943359375,
0.2476806640625,
0.491943359375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.509765625,
0.251708984375,
0.8818359375,
0.251708984375,
0.8818359375,
0.265380859375,
0.509765625,
0.265380859375
],
[
0.1158447265625,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.509765625,
0.267333984375,
0.8818359375,
0.267333984375,
0.8818359375,
0.28076171875,
0.509765625,
0.28076171875
],
[
0.1158447265625,
0.27978515625,
0.48779296875,
0.27978515625,
0.48779296875,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.51318359375,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.1175537109375,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.509765625,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.314208984375,
0.509765625,
0.314208984375
],
[
0.1175537109375,
0.31005859375,
0.489501953125,
0.31005859375,
0.489501953125,
0.32373046875,
0.1175537109375,
0.32373046875
],
[
0.51318359375,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.1175537109375,
0.327880859375,
0.484375,
0.327880859375,
0.484375,
0.33837890625,
0.1175537109375,
0.33837890625
],
[
0.51123046875,
0.33203125,
0.8818359375,
0.330810546875,
0.8818359375,
0.345703125,
0.51123046875,
0.3466796875
],
[
0.1175537109375,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.51123046875,
0.3486328125,
0.8818359375,
0.3486328125,
0.8818359375,
0.3623046875,
0.51123046875,
0.3623046875
],
[
0.1175537109375,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.51123046875,
0.366455078125,
0.880859375,
0.366455078125,
0.880859375,
0.377197265625,
0.51123046875,
0.377197265625
],
[
0.1175537109375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.5087890625,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.393798828125,
0.5087890625,
0.393798828125
],
[
0.118408203125,
0.393798828125,
0.48779296875,
0.393798828125,
0.48779296875,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.509765625,
0.396728515625,
0.880859375,
0.396728515625,
0.880859375,
0.410400390625,
0.509765625,
0.410400390625
],
[
0.11419677734375,
0.407958984375,
0.47021484375,
0.40673828125,
0.47021484375,
0.421630859375,
0.11419677734375,
0.4228515625
],
[
0.509765625,
0.412109375,
0.8818359375,
0.413330078125,
0.8818359375,
0.427490234375,
0.509765625,
0.427001953125
],
[
0.1351318359375,
0.4287109375,
0.489501953125,
0.429931640625,
0.489501953125,
0.44482421875,
0.1351318359375,
0.443603515625
],
[
0.51318359375,
0.4287109375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.1158447265625,
0.44775390625,
0.49365234375,
0.44775390625,
0.49365234375,
0.46142578125,
0.1158447265625,
0.46142578125
],
[
0.51318359375,
0.446533203125,
0.880859375,
0.446533203125,
0.880859375,
0.457275390625,
0.51318359375,
0.457275390625
],
[
0.1175537109375,
0.463134765625,
0.491943359375,
0.463134765625,
0.491943359375,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.51123046875,
0.46142578125,
0.8818359375,
0.46142578125,
0.8818359375,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.1158447265625,
0.479736328125,
0.491943359375,
0.479736328125,
0.491943359375,
0.493408203125,
0.1158447265625,
0.493408203125
],
[
0.51123046875,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.51123046875,
0.4912109375
],
[
0.1158447265625,
0.495361328125,
0.490234375,
0.495361328125,
0.490234375,
0.5087890625,
0.1158447265625,
0.5087890625
],
[
0.509765625,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.509765625,
0.50732421875
],
[
0.1175537109375,
0.51123046875,
0.491943359375,
0.51123046875,
0.491943359375,
0.52490234375,
0.1175537109375,
0.52490234375
],
[
0.51318359375,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.51318359375,
0.52392578125
],
[
0.1175537109375,
0.52783203125,
0.48779296875,
0.52783203125,
0.48779296875,
0.54150390625,
0.1175537109375,
0.54150390625
],
[
0.5087890625,
0.52490234375,
0.8818359375,
0.525390625,
0.8818359375,
0.54052734375,
0.5087890625,
0.5390625
],
[
0.1158447265625,
0.5419921875,
0.486083984375,
0.5419921875,
0.486083984375,
0.5556640625,
0.1158447265625,
0.5556640625
],
[
0.509765625,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.509765625,
0.5546875
],
[
0.1175537109375,
0.55859375,
0.48779296875,
0.55859375,
0.48779296875,
0.572265625,
0.1175537109375,
0.572265625
],
[
0.51123046875,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51123046875,
0.57177734375
],
[
0.1175537109375,
0.576171875,
0.489501953125,
0.576171875,
0.489501953125,
0.5888671875,
0.1175537109375,
0.5888671875
],
[
0.509765625,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58837890625,
0.509765625,
0.58837890625
],
[
0.11419677734375,
0.59130859375,
0.48779296875,
0.59033203125,
0.48779296875,
0.60498046875,
0.11419677734375,
0.60546875
],
[
0.51318359375,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.11419677734375,
0.60791015625,
0.48779296875,
0.60693359375,
0.48779296875,
0.62158203125,
0.11419677734375,
0.62255859375
],
[
0.509765625,
0.60546875,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.509765625,
0.62060546875
],
[
0.11419677734375,
0.62255859375,
0.48779296875,
0.62451171875,
0.48779296875,
0.63818359375,
0.11419677734375,
0.6357421875
],
[
0.509765625,
0.62255859375,
0.736328125,
0.62255859375,
0.736328125,
0.6357421875,
0.509765625,
0.6357421875
],
[
0.11419677734375,
0.6396484375,
0.48779296875,
0.64013671875,
0.48779296875,
0.65478515625,
0.11419677734375,
0.65380859375
],
[
0.51318359375,
0.646484375,
0.88330078125,
0.646484375,
0.88330078125,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1175537109375,
0.65771484375,
0.486083984375,
0.65771484375,
0.486083984375,
0.66845703125,
0.1175537109375,
0.66845703125
],
[
0.51318359375,
0.66455078125,
0.88330078125,
0.66455078125,
0.88330078125,
0.677734375,
0.51318359375,
0.677734375
],
[
0.1175537109375,
0.6728515625,
0.48779296875,
0.6728515625,
0.48779296875,
0.68603515625,
0.1175537109375,
0.68603515625
],
[
0.51318359375,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.51318359375,
0.693359375
],
[
0.11419677734375,
0.68701171875,
0.391357421875,
0.6884765625,
0.391357421875,
0.703125,
0.11419677734375,
0.70166015625
],
[
0.509765625,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.509765625,
0.708984375
],
[
0.1351318359375,
0.71044921875,
0.491943359375,
0.71142578125,
0.491943359375,
0.7255859375,
0.1351318359375,
0.724609375
],
[
0.51318359375,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.51318359375,
0.724609375
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.72705078125,
0.88330078125,
0.72705078125,
0.88330078125,
0.74072265625,
0.509765625,
0.74072265625
],
[
0.1158447265625,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1158447265625,
0.75732421875
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.509765625,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.509765625,
0.7724609375
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.509765625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.509765625,
0.7890625
],
[
0.1158447265625,
0.7919921875,
0.489501953125,
0.7919921875,
0.489501953125,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.51318359375,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.1175537109375,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.8193359375,
0.1175537109375,
0.8193359375
],
[
0.51123046875,
0.80712890625,
0.88330078125,
0.8056640625,
0.88330078125,
0.82080078125,
0.51123046875,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51318359375,
0.82373046875,
0.88330078125,
0.82373046875,
0.88330078125,
0.83740234375,
0.51318359375,
0.83740234375
],
[
0.1175537109375,
0.84033203125,
0.491943359375,
0.84033203125,
0.491943359375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.5087890625,
0.84033203125,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.5087890625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.490234375,
0.85693359375,
0.490234375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5087890625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.5087890625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.5087890625,
0.88720703125,
0.8818359375,
0.8896484375,
0.8818359375,
0.90380859375,
0.5087890625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.49365234375,
0.90478515625,
0.49365234375,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"methods (Narasimhan et al., 2018; Zhu et al., 2020)",
"ference by focusing on correct reasoning evidences",
"to learn to selectively focus on necessary pieces of",
"under weak supervision.",
"knowledge. The memory-based methods represent",
"knowledge facts in a form of memory and calculate",
"2Related Work",
"softattention scores of eachmemory withrespect",
"Knowledge-based visual question answering",
"to a question. Then, it infers an answer by attending",
"(Wang et al., 2017, 2018; Shah et al., 2019; Marino",
"toknowledge evidence withhigh attention scores.",
"et al., 2019; Sampat et al., 2020) proposed bench-",
"On the other hand, to explicitly consider relational",
"mark datasets for knowledge-based visual question",
"structure between knowledge facts,graph-based",
"answering that requires reasoning about an image",
"methods construct a query-aware knowledge graph",
"on the basis of facts from a large-scale knowledge",
"by retrieving facts from KB and perform graph rea-",
"base (KB) such as Freebase (Bollacker et al., 2008)",
"soning for a question. These methods mainly adopt",
"or DBPedia (Auer et al., 2007). To solve the task",
"an iterative message passing process to propagate",
"two pioneering studies (Wang et al., 2017, 2018)",
"information between adjacent nodes in the graph.",
"suggested logical parsing-based methods which",
"However, it is difficult to capture multi-hop rela-",
"convert a question to a KB logic guery using pre",
"tionships containing long-distance nodes from the",
"defined query templates and execute the generated",
"graph due to the well-known over-smoothing prob-",
"query on KB for searching an answer. Since then",
"lem, where repetitive message passing process to",
"informationretrieval-based methodswhichretrieve",
"propagate information across long distance makes",
"knowledge facts associated with a question and con-",
"features of connected nodestoosimilar and undis-",
" duct semantic matching between the facts and the",
"criminating (Li et al., 2018; Wang et al., 2020).",
"question are introduced. (Narasimhan and Schwing",
"To address the above limitation, we propose a",
"2018; Shah et al., 2019) proposed memory-based",
"novel method, Hypergraph Transformer, which ex-",
"methodsthatrepresentknowledgefacts intheform",
"ploits hypergraph structure to encode multi-hop re-",
"of memory and calculate soft attention scores of",
"lationships and transformer-based attention mecha-",
"the memory with a question. (Narasimhan et al.,",
"nism to learn to pay attention to important knowl-",
"2018; Zhu et al., 2020) represented the retrieved",
"edge evidences for a question. We construct a ques-",
"facts as a graph and performed graph reasoning",
"tion hypergraph and a knowledge hypergraph to",
"through message passing scheme utilizing graph",
"explicitly encode high-order semantics present in",
"convolution.However,thesemethods are compli-",
"the question and each knowledge fact, and capture",
"cated to encode inherent high-order semantics and",
"multi-hop relational knowledge facts effectively.",
"multi-hop relationships present in the knowledge",
"Then, we perform hyperedge matching between the",
"graph. Therefore, we introduce a concept of hy",
"two hypergraphs by leveraging transformer-based",
"pergraph and propose transformer-based attention",
" attention mechanism. We argue that introducing the",
"mechanism over hypergraphs.",
"concept of hypergraph is powerful for multi-hop",
"Multi-hop knowledge graph reasoning is a",
"reasoningprobleminthatitcanencodehigh-order",
"process of sequential reasoning based on multi-",
"semantics without the constraint of length and learn",
"ple evidences of a knowledge graph, and has been",
"cross-modal high-order associations.",
"broadly used in various downstream tasks such as",
"The main contributions of this paper can be sum-",
"question answering (Lin et al., 2019; Saxena et al..",
"marized as follows.i) We propose Hypergraph",
"2020; Han et al., 2020b,a; Yadati et al., 2021), or",
"Transformer which enhances multi-hop reasoning",
"knowledge-enhanced text generation (Liu et al..",
"ability by encoding high-order semantics in the",
"2019; Mo0on et al., 2019; Ji et al., 2020).Recent",
"form of a hypergraph and learning inter- and intra-",
"researches have introduced the concept of hyper",
"high-order associations in hypergraphs using the",
"graph for multi-hop graph reasoning (Kim et al.,",
"attentionmechanism.ii)Weconduct extensiveex-",
"2020; Han et al., 2020b,a; Yadati et al., 2019, 2021;",
"periments on two knowledge-based VQA datasets",
"Sun et al., 2020). These models have a similar moti-",
"(KVQA and FVQA) and two knowledge-based tex-",
"vation to the Hypergraph Transformer proposed in",
"tual QA datasets (PQ and PQL) and show superior",
" this paper, but core operations are vastly different.",
"performances on all datasets, especially multi-hop",
" These models mainly update node representations",
"reasoning problem. ii) We qualitatively observe",
"in the hypergraph through a message passing pro",
"that Hypergraph Transformer performs robust in-",
"cess using graph convolution operation. On the"
] |
{
"arxiv_id": "2204.10448",
"language": "en",
"page_idx": 1
}
|
||
2204.10448-en-3
|
2204.10448-en-3.jpg
|
Question hypergraph We transform a question of a node set $\mathcal{V}^q$ and a hyperedge set $\mathcal{E}^q.$ We assume that each word unit (a word or named entity) of the question is defined as a node, and has edges to adjacent nodes. For question hypergraph, each word unit is used as a start node of a graph walk. The multi-hop graph walk is conducted in the same manner as the knowledge hypergraph. A $n$-gram phrase is considered as a hyperedge in the question hypergraph (see Figure 2(b)).
3.4 Reasoning with attention mechanism
To consider high-order associations between knowledge and question, we devise structural semantic matching between the query-aware knowledge hypergraph and the question hypergraph. We introduce an attention mechanism over two hypergraphs based on guided-attention (Tsai et al., 2019) and self-attention (Vaswani et al., 2017). As shown in Figure 2(c), the guided-attention blocks are introduced to learn correlations between knowledge hyperedges and question hyperedges by interattention mechanism, and then intra-relationships of in knowledge or question hyperedges are trained with the following self-attention blocks. The details of two modules, guided-attention blocks and selfattention blocks, are described as below. Note that we use $Q,K$, and $V$ for query, key, value, and $q,k$ as subscripts to represent question and knowledge, respectively.
Guided-attention To learn inter-association between two hypergraphs, we first embed a knowledge hyperedge and a question hyperedge as fol- $\mathbb{R}^d$ where $h^[\cdot]$ is a hyperedge in $\mathcal{E}^[\cdot].$ Here, $f_[\cdot]$ is a hyperedge embedding function and $\phi_{[\cdot]}$ is a linear projection function. The design and implementation of $f_[\cdot]$ are not constrained (e.g., any pooling operation or any learnable neural networks), but we use a simple concatenation operation of node representations in a hyperedge as $f_{[\cdot]}.$ The representations of hyperedges in the same hypergraph (e.g., $e^k,e^q)$ are packed together into a matrix $E^k$ and $E^q.$
We define the knowledge hyperedges $E^k$ and the question hyperedges $E^q$ as a query and key-value pairs, respectively. We set a query $Q_k=\dot{E}^kW_{Q_k}$, a key $K_q=E^qW_{K_q}$, and a value $V_q=E^qW_{V_q}$, where all projection matrices $W_{[\cdot]}\stackrel{.}{\in}\mathbb{R}^{d\times d_{v}}$ are learnable parameters. Then, scaled dot product at-
tention using the query, key, and value is calculated
sentence into a question hypergraph $\mathcal{H}^q$ consisting as Attention$(Q_k,K_q,V_q)=$ softmax$(\frac{Q_kK_q^T}{\sqrt{d_v}})V_q$
where $d_v$ is the dimension of the query and the key vector. In addition, the guided-attention which uses the question hyperedges as query and the knowledge hyperedges as key-value pairs is performed in a similar manner: Attention$(Q_q,K_k,V_k).$
Self-attention The only difference between guided-attention and self-attention is that the same input is used for both query and key-value within self-attention. For example, we set query, key, and value based on the knowledge hyperedges $E_k$, and the self-attention for knowledge hyperedges is conducted by Attention$(Q_k,K_k,V_k).$ For question hyperedges $E_q$, self-attention is performed in a similar manner: Attention$(Q_q,K_q,V_q).$
Following the standard structure of the transformer, we build up guided-attention block and selfattention block where each block consists of each attention operation with layer normalization, residual connection, and a single feed-forward layer. By passing the guided-attention blocks and selfattention blocks sequentially, representations of knowledge hyperedges and question hyperedges are updated and finally aggregated to single vector representation as $z_k\in\mathbb{R}^{d_v}$ and $z_q\in\mathbb{R}^{d_v}$, respectively.
## 3.5 Answer predictor
To predict an answer, we first concatenate the representation $z_k$ and $z_q$ obtained from the attention blocks and feed into a single feed-forward layer
lows: $e^k= \phi _k\circ f_k( h^k) \in \mathbb{R} ^d, e^q= \phi _q\circ f_q( h^q) \in$ (i.e., $\mathbb{R}^{2d_v}\mapsto\mathbb{R}^w)$ to make a joint representation
$z.$ We then consider two types of answer predictor: multi-layer perceptron and similarity-based answer predictor. Multi-layer perceptron as an answer classifier $p=\psi(z)$ is a prevalent for visual question answering problems. For similarity-based answer, we calculate a dot product similarity $p=zC^T$ between $z$ and answer candidate set $C\in\mathbb{R}^|\mathcal{A}|\times w$ where $|\mathcal{A}|$ is a number of candidate answers and $w$ is a dimension of representation for each answer. The most similar answer to the joint representation is selected as an answer among the answer candidates. For training, we use only supervision from QA pairs without annotations for ground-truth reasoning paths. To this end, cross-entropy between prediction $p$ and ground-truth $t$ is utilized as a loss function.
|
[
[
0.118408203125,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.509765625,
0.08489990234375,
0.8818359375,
0.0843505859375,
0.8818359375,
0.09857177734375,
0.509765625,
0.09979248046875
],
[
0.8037109375,
0.09320068359375,
0.88330078125,
0.10394287109375,
0.87890625,
0.122314453125,
0.79833984375,
0.11224365234375
],
[
0.1175537109375,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11639404296875,
0.1175537109375,
0.11639404296875
],
[
0.5087890625,
0.10394287109375,
0.81005859375,
0.10394287109375,
0.81005859375,
0.1217041015625,
0.5087890625,
0.1217041015625
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.509765625,
0.122314453125,
0.880859375,
0.12353515625,
0.880859375,
0.1383056640625,
0.509765625,
0.13720703125
],
[
0.8193359375,
0.1175537109375,
0.8427734375,
0.1175537109375,
0.8427734375,
0.12054443359375,
0.8193359375,
0.12054443359375
],
[
0.1175537109375,
0.1341552734375,
0.48779296875,
0.1341552734375,
0.48779296875,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.5087890625,
0.1383056640625,
0.880859375,
0.1383056640625,
0.880859375,
0.1519775390625,
0.5087890625,
0.1519775390625
],
[
0.1158447265625,
0.1484375,
0.48779296875,
0.149658203125,
0.48779296875,
0.1644287109375,
0.1158447265625,
0.163330078125
],
[
0.5087890625,
0.155029296875,
0.8818359375,
0.155029296875,
0.8818359375,
0.168701171875,
0.5087890625,
0.168701171875
],
[
0.1175537109375,
0.166259765625,
0.484375,
0.166259765625,
0.484375,
0.18115234375,
0.1175537109375,
0.18115234375
],
[
0.5087890625,
0.171630859375,
0.8818359375,
0.17041015625,
0.8818359375,
0.185302734375,
0.5087890625,
0.1859130859375
],
[
0.1158447265625,
0.1817626953125,
0.48779296875,
0.18115234375,
0.48779296875,
0.1953125,
0.1158447265625,
0.196533203125
],
[
0.509765625,
0.188232421875,
0.82373046875,
0.188232421875,
0.82373046875,
0.203125,
0.509765625,
0.203125
],
[
0.1175537109375,
0.198974609375,
0.484375,
0.198974609375,
0.484375,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.1158447265625,
0.215576171875,
0.486083984375,
0.215576171875,
0.486083984375,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.51318359375,
0.223876953125,
0.8818359375,
0.223876953125,
0.8818359375,
0.2369384765625,
0.51318359375,
0.2369384765625
],
[
0.1158447265625,
0.2298583984375,
0.48779296875,
0.23095703125,
0.48779296875,
0.2452392578125,
0.1158447265625,
0.24462890625
],
[
0.1175537109375,
0.2476806640625,
0.3359375,
0.2476806640625,
0.3359375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.515625,
0.2423095703125,
0.880859375,
0.2423095703125,
0.880859375,
0.2529296875,
0.515625,
0.2529296875
],
[
0.51123046875,
0.257080078125,
0.880859375,
0.257080078125,
0.880859375,
0.270751953125,
0.51123046875,
0.270751953125
],
[
0.1158447265625,
0.273681640625,
0.450927734375,
0.27490234375,
0.450927734375,
0.289306640625,
0.1158447265625,
0.2880859375
],
[
0.51123046875,
0.2724609375,
0.8818359375,
0.2724609375,
0.8818359375,
0.287353515625,
0.51123046875,
0.287353515625
],
[
0.51123046875,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.51123046875,
0.302734375
],
[
0.1158447265625,
0.297607421875,
0.489501953125,
0.297607421875,
0.489501953125,
0.311279296875,
0.1158447265625,
0.311279296875
],
[
0.509765625,
0.3046875,
0.88330078125,
0.3046875,
0.88330078125,
0.318359375,
0.509765625,
0.318359375
],
[
0.1175537109375,
0.314208984375,
0.48779296875,
0.314208984375,
0.48779296875,
0.327880859375,
0.1175537109375,
0.327880859375
],
[
0.51123046875,
0.320556640625,
0.88330078125,
0.320556640625,
0.88330078125,
0.3349609375,
0.51123046875,
0.3349609375
],
[
0.1175537109375,
0.330078125,
0.490234375,
0.330078125,
0.490234375,
0.34326171875,
0.1175537109375,
0.34326171875
],
[
0.51123046875,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.51123046875,
0.350830078125
],
[
0.1175537109375,
0.345703125,
0.486083984375,
0.345703125,
0.486083984375,
0.35986328125,
0.1175537109375,
0.35986328125
],
[
0.51123046875,
0.3515625,
0.81005859375,
0.352783203125,
0.81005859375,
0.36767578125,
0.51123046875,
0.366455078125
],
[
0.1158447265625,
0.361083984375,
0.486083984375,
0.361083984375,
0.486083984375,
0.374755859375,
0.1158447265625,
0.374755859375
],
[
0.5283203125,
0.37060546875,
0.8818359375,
0.371826171875,
0.8818359375,
0.385986328125,
0.5283203125,
0.385498046875
],
[
0.1158447265625,
0.377685546875,
0.486083984375,
0.377197265625,
0.486083984375,
0.391357421875,
0.1158447265625,
0.392578125
],
[
0.51123046875,
0.388427734375,
0.88330078125,
0.388427734375,
0.88330078125,
0.402587890625,
0.51123046875,
0.402587890625
],
[
0.1175537109375,
0.394287109375,
0.486083984375,
0.394287109375,
0.486083984375,
0.407958984375,
0.1175537109375,
0.407958984375
],
[
0.1175537109375,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.1175537109375,
0.42333984375
],
[
0.51318359375,
0.40625,
0.8818359375,
0.40625,
0.8818359375,
0.416259765625,
0.51318359375,
0.416259765625
],
[
0.51123046875,
0.42041015625,
0.88330078125,
0.42041015625,
0.88330078125,
0.43408203125,
0.51123046875,
0.43408203125
],
[
0.1175537109375,
0.427001953125,
0.490234375,
0.427001953125,
0.490234375,
0.440673828125,
0.1175537109375,
0.440673828125
],
[
0.51123046875,
0.435791015625,
0.88330078125,
0.435302734375,
0.88330078125,
0.449462890625,
0.51123046875,
0.45068359375
],
[
0.1175537109375,
0.4423828125,
0.489501953125,
0.4423828125,
0.489501953125,
0.4560546875,
0.1175537109375,
0.4560546875
],
[
0.51123046875,
0.453125,
0.8818359375,
0.453125,
0.8818359375,
0.466064453125,
0.51123046875,
0.466064453125
],
[
0.1175537109375,
0.460205078125,
0.484375,
0.460205078125,
0.484375,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.1158447265625,
0.474365234375,
0.48779296875,
0.474365234375,
0.48779296875,
0.488037109375,
0.1158447265625,
0.488037109375
],
[
0.5146484375,
0.47021484375,
0.880859375,
0.47021484375,
0.880859375,
0.48095703125,
0.5146484375,
0.48095703125
],
[
0.1158447265625,
0.490478515625,
0.489501953125,
0.4892578125,
0.489501953125,
0.50439453125,
0.1158447265625,
0.5048828125
],
[
0.5087890625,
0.485107421875,
0.87890625,
0.485107421875,
0.87890625,
0.498779296875,
0.5087890625,
0.498779296875
],
[
0.1175537109375,
0.50732421875,
0.490234375,
0.50732421875,
0.490234375,
0.52099609375,
0.1175537109375,
0.52099609375
],
[
0.51318359375,
0.501953125,
0.8818359375,
0.501953125,
0.8818359375,
0.515625,
0.51318359375,
0.515625
],
[
0.509765625,
0.5166015625,
0.8818359375,
0.5166015625,
0.8818359375,
0.53076171875,
0.509765625,
0.53076171875
],
[
0.1175537109375,
0.5224609375,
0.489501953125,
0.5224609375,
0.489501953125,
0.5361328125,
0.1175537109375,
0.5361328125
],
[
0.509765625,
0.53076171875,
0.55908203125,
0.533203125,
0.5576171875,
0.548828125,
0.5087890625,
0.54638671875
],
[
0.1175537109375,
0.5390625,
0.48779296875,
0.5390625,
0.48779296875,
0.55419921875,
0.1175537109375,
0.55419921875
],
[
0.1158447265625,
0.5556640625,
0.48779296875,
0.5556640625,
0.48779296875,
0.5693359375,
0.1158447265625,
0.5693359375
],
[
0.1175537109375,
0.572265625,
0.2091064453125,
0.572265625,
0.2091064453125,
0.58447265625,
0.1175537109375,
0.58447265625
],
[
0.51123046875,
0.57080078125,
0.69287109375,
0.57080078125,
0.69287109375,
0.5849609375,
0.51123046875,
0.5849609375
],
[
0.1175537109375,
0.5966796875,
0.491943359375,
0.5966796875,
0.491943359375,
0.60986328125,
0.1175537109375,
0.60986328125
],
[
0.509765625,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.509765625,
0.6123046875
],
[
0.1175537109375,
0.6142578125,
0.491943359375,
0.6142578125,
0.491943359375,
0.62890625,
0.1175537109375,
0.62890625
],
[
0.5087890625,
0.6142578125,
0.880859375,
0.6142578125,
0.880859375,
0.62744140625,
0.5087890625,
0.62744140625
],
[
0.1175537109375,
0.6298828125,
0.49365234375,
0.62890625,
0.49365234375,
0.6435546875,
0.1175537109375,
0.64453125
],
[
0.509765625,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.509765625,
0.64453125
],
[
0.1796875,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.6611328125,
0.1796875,
0.6611328125
],
[
0.50732421875,
0.6435546875,
0.88330078125,
0.6455078125,
0.88330078125,
0.662109375,
0.50732421875,
0.66015625
],
[
0.11248779296875,
0.65673828125,
0.491943359375,
0.6591796875,
0.491943359375,
0.6796875,
0.11248779296875,
0.677734375
],
[
0.118408203125,
0.64794921875,
0.187255859375,
0.64794921875,
0.187255859375,
0.65771484375,
0.118408203125,
0.65771484375
],
[
0.509765625,
0.6630859375,
0.884765625,
0.6630859375,
0.884765625,
0.677734375,
0.509765625,
0.677734375
],
[
0.1175537109375,
0.6787109375,
0.490234375,
0.6787109375,
0.490234375,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.51123046875,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.51123046875,
0.693359375
],
[
0.1175537109375,
0.69482421875,
0.491943359375,
0.69482421875,
0.491943359375,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.51123046875,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.708984375,
0.51123046875,
0.708984375
],
[
0.1175537109375,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.509765625,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.509765625,
0.7255859375
],
[
0.1158447265625,
0.72705078125,
0.489501953125,
0.7255859375,
0.489501953125,
0.74072265625,
0.1158447265625,
0.7412109375
],
[
0.51123046875,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1158447265625,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1158447265625,
0.755859375
],
[
0.5087890625,
0.7421875,
0.8818359375,
0.7392578125,
0.8818359375,
0.75732421875,
0.5087890625,
0.76025390625
],
[
0.1175537109375,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.509765625,
0.75830078125,
0.8818359375,
0.7548828125,
0.8818359375,
0.77392578125,
0.509765625,
0.77685546875
],
[
0.1158447265625,
0.77392578125,
0.486083984375,
0.77490234375,
0.486083984375,
0.7890625,
0.1158447265625,
0.7880859375
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1158447265625,
0.79052734375,
0.486083984375,
0.7890625,
0.486083984375,
0.80419921875,
0.1158447265625,
0.8046875
],
[
0.509765625,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8046875,
0.509765625,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.173828125,
0.8056640625,
0.175537109375,
0.818359375,
0.118408203125,
0.82080078125
],
[
0.51123046875,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.133544921875,
0.82177734375,
0.490234375,
0.82080078125,
0.490234375,
0.83837890625,
0.133544921875,
0.8388671875
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.490234375,
0.84033203125,
0.490234375,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.8818359375,
0.84130859375,
0.8818359375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.11419677734375,
0.85400390625,
0.491943359375,
0.85498046875,
0.491943359375,
0.87255859375,
0.11419677734375,
0.87158203125
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.87060546875,
0.491943359375,
0.87060546875,
0.491943359375,
0.88818359375,
0.1158447265625,
0.88818359375
],
[
0.509765625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.11248779296875,
0.88525390625,
0.491943359375,
0.8828125,
0.491943359375,
0.90380859375,
0.11248779296875,
0.90625
],
[
0.51123046875,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90625,
0.580078125,
0.90625,
0.580078125,
0.91748046875,
0.51318359375,
0.91748046875
]
] |
[
"Question hypergraph We transform a question",
"tention using the query,key, and value is calculated",
"(QKKL)Va",
"sentence into a question hypergraph H consisting",
"as Attention(Qk, Kq, V) = softmax(",
"of a node set Vq and a hyperedge set &. We assume",
"where d, is the dimension of the query and the key",
"",
"that each word unit (a word or named entity) of",
"vector. In addition, the guided-attention which uses",
"the question is defined as a node, and has edges",
"the question hyperedges as query and the knowl-",
"to adjacent nodes. For question hypergraph, each",
"edge hyperedges as key-value pairs is performed in",
"word unit is used as a start node of a graph walk.",
"a similar manner: Attention(Qg, Kk, V).",
"The multi-hop graph walk is conducted in the same",
"manner as the knowledge hypergraph. A n-gram",
"Self-attention The only difference between",
"phrase is considered as a hyperedge in the question",
"hypergraph (see Figure 2(b).",
"guided-attention and self-attention is that the same",
"input is used for both query and key-value within",
"3.4 Reasoning with attention mechanism",
"self-attention. For example, we set query, key, and",
"value based on the knowledge hyperedges Ek, and",
"To consider high-order associations between knowl-",
"the self-attention for knowledge hyperedges is con-",
"edge and question, we devise structural semantic",
"ducted by Attention(Qk, Kk, V). For question",
"matching between the query-aware knowledge hy-",
"hyperedges Eg, self-attention is performed in a",
"pergraph and the question hypergraph. We intro-",
"similar manner: Attention(Qg, Kg, V).",
"duce an attention mechanism over two hypergraphs",
"Following the standard structure of the trans-",
"based on guided-attention (Tsai et al., 2019) and",
"former, we build up guided-attention block and self-",
"self-attention (Vaswani et al., 2017). As shown",
"in Figure 2(c), the guided-attention blocks are",
"attentionblockwhereeachblockconsistsofeach",
"attention operation with layer normalization, resid-",
"introduced to learn correlations between knowl-",
"ual connection, and a single feed-forward layer.",
"edge hyperedges and question hyperedges by inter-",
"By passing the guided-attention blocks and self-",
"attentionmechanism,and thenintra-relationships",
"of in knowledge or question hyperedges are trained",
"attentionblocks seguentially.representations of",
"with the following self-attention blocks. The details",
"knowledge hyperedges and question hyperedges",
"of two modules, guided-attention blocks and self-",
"are updated and finally aggregated to single vector",
"representation as zk E Rdu and zq e Rdu, respec-",
"attention blocks, are described as below. Note that",
"tively.",
"we use Q, K, and V for query, key, value, and q, k",
"as subscripts to represent question and knowledge,",
"respectively.",
"3.5Answerpredictor",
"Guided-attentionTo learn inter-association be-",
"To predict an answer, we first concatenate the rep",
"tween two hypergraphs, we first embed a knowl-",
"resentation Zh and zg obtained from the attention",
"edge hyperedge and a question hyperedge as fol-",
"blocks and feed into a single feed-forward layer",
"= Φk o fi(hh) ERd,eq =Φgo fg(h) E",
"(i.e., IR2dv → Rw) to make a joint representation",
"Rd where hll is a hyperedge in ell. Here, fy is a",
"lows: ek:",
" z. We then consider two types of answer predictor:",
"hyperedge embedding function and is a linear",
"multi-layer perceptron and similarity-based answer",
"projection function. The design and implementa-",
"predictor. Multi-layer perceptron as an answer clas-",
"tion of fr are not constrained (e.g., any pooling",
"sifier p = b(z) is a prevalent for visual question",
"operation or any learnable neural networks), but",
"answering problems.For similarity-based answer",
"we use a simple concatenation operation of node",
"we calculate a dot product similarity p = zCT",
"representations in a hyperedge as fi. The repre-",
"between z and answer candidate set C e RIA|xw",
"sentations of hyperedges in the same hypergraph",
"where A is a number of candidate answers and w",
"(e.g., ek, e9) are packed together into a matrix Ek",
"is a dimension of representation for each answer.",
"andEq",
"The most similar answer to the joint representation",
"We define the knowledge hyperedges Ek and the",
"is selected as an answer among the answer candi-",
"question hyperedges E9 as a query and key-value",
"dates. For training, we use only supervision from",
"pairs, respectively. We set a query Qk=- EkWQk,",
"QA pairs without annotations for ground-truth rea-",
"a key Kg = EWkg, and a value Vq = EWvg,",
" soning paths. To this end, cross-entropy between",
"where all projection matrices W e Rdxdu are",
"prediction p and ground-truth t is utilized as a loss",
"learnable parameters. Then, scaled dot product at-",
"function."
] |
{
"arxiv_id": "2204.10448",
"language": "en",
"page_idx": 3
}
|
||
2309.00216-en-0
|
2309.00216-en-0.jpg
|
$\textbf{Human- Inspired Facial Sketch Synthesis with Dynamic Adaptation}$
Fei Gao$^1$, Yifan Zhu$^2$, Chang Jiang$^2$, Nannan Wang$^{3*}$
$^{1}$Hangzhou Institute of Technology, Xidian University $^{2}$Hangzhou Dianzi University $^{3}$Xidian University
fgao@xidian.edu.cn, 2961695289@qq.com, jc233@hdu.edu.cn, nnwang@xidian.edu.cn
$ଟ୍ରି$
$UI$
-1
## Abstract
Facial sketch synthesis (FSS) aims to generate a vivid
sketch portrait from $a$ given facial $photo.$ Existing $FSS$ methods merely rely on 2D representations offacial seman- tic $or$ annearance However nofe ssional human artist
tic or appearance. However, professional human artists usually use outlines or shadings to covey 3D geometry. Thus facial 3D geometry (e.g. depth map) is extremely im-
$\sum_{\begin{array}{ccc}&portant\:for\:FSS.\:Besides,\:different\:artists\:may\:use\:diverse\\&drawing\:techniques\:and\:create\:multiple\:styles\:of\:sketches;\\&but\:the\:style\:is\:globally\:consistent\:in\:a\:sketch.\:Inspired\end{array}}$ $\begin{array}{ccc}{\dot{o}}&{bysuch~observations,~in~this~paper,~we~propose~a~novel}\\{\dot{O}}&{Human-Inspired~Dynamic~Adaptation~(HIDA)~method.~Spe-}\\{\dot{O}}&{\dot{O}}\end{array}$
cially, we propose to dynamically modulate neuron activations based on a joint consideration of both facial 3D geometry and 2D appearance, as well as globally consistent style control. Besides, we use deformable convolutions at coarse-scales to align deep features, for generating abstract and distinct outlines. Experiments show that HIDA can generate high-quality sketches in multiple styles, and significantly outperforms previous methods, over a large range of challenging faces. Besides, HIDA allows precise style control of the synthesized sketch, and generalizes well to natural scenes and other artistic styles. Our code and results have been released online at: https: $//g$ithub.com/AiArt-HDU/HIDA.
<FigureHere>
Figure 1: Illustration of facial photos, depth maps, multistyle facial sketches drawn by human artists[8], and the corresponding results synthesized by our method
inspired by the excellent success of Generative Adver-
sarial Networks (GANs)[18].
Specially, researchers
have proposed various techniques, including embedding image prior[43], semi-supervised learning [5], selfattention/transformer based methods [11,47,9], hierarchical GANs [28, 40, 8], composition assistance [37], and semantic adaptive normalization [21], to boost the quality of synthesized sketches. However, all these methods merely use 2D appearance or semantic representations of the input photo. They may fail to handle serious variations in appearance, such as the pose, lighting, expression, and skin color.
To tackle this challenge, we propose a novel method, inspired by how human artists draw a sketch. We observe that facial 3D geometry plays a significant role in human artists' drawing process. Besides, a professional human artist considers comprehensive information, including facial 3D geometry, 2D appearance, and the artistic style, to execute a sketch portrait. We summarize the drawing methodologies of human artists[22] into the following four folds:
$\cdot$ Local 3D geometry conveyor: First, artists typically use abstract and deformable outlines to characterize major geometry, and use different shading methodologies, e.g. hatching, blending, and stippling, to convey local 3D structures [3].
## $1. \textbf{Introduction}$
Making computers create arts like human beings, is a longstanding and challenging topic, in the artificial intelligence (AI) area [2]. To this end, researchers have made great efforts and proposed numerous methods, such as neural style transfer (NST) [17] and image-to-image translation (I2IT)[18, 46]. These methods mainly tackle cluttered image styles, such as oil paintings[17]. In this paper, we are interested in creating artistic sketches from facial photos, which is referred to as face sketch synthesis (FSS)[34].
For now, there has been significant progress in FSS
* Corresponding Author
|
[
[
0.1519775390625,
0.1337890625,
0.8203125,
0.1337890625,
0.8203125,
0.1483154296875,
0.1519775390625,
0.1483154296875
],
[
0.28515625,
0.18115234375,
0.705078125,
0.179931640625,
0.705078125,
0.1976318359375,
0.28515625,
0.1988525390625
],
[
0.08905029296875,
0.1976318359375,
0.90283203125,
0.1988525390625,
0.90283203125,
0.216552734375,
0.08905029296875,
0.21533203125
],
[
0.15283203125,
0.2183837890625,
0.83251953125,
0.2197265625,
0.83251953125,
0.2342529296875,
0.15283203125,
0.23291015625
],
[
0.0236968994140625,
0.271484375,
0.05963134765625,
0.271484375,
0.05963134765625,
0.330810546875,
0.0236968994140625,
0.330810546875
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.101318359375,
0.3017578125,
0.4697265625,
0.3017578125,
0.4697265625,
0.31640625,
0.101318359375,
0.31640625
],
[
0.0294189453125,
0.31494140625,
0.06451416015625,
0.318115234375,
0.056365966796875,
0.375,
0.021240234375,
0.371826171875
],
[
0.08087158203125,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.08087158203125,
0.333251953125,
0.46728515625,
0.333251953125,
0.46728515625,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.08087158203125,
0.363037109375,
0.46728515625,
0.363037109375,
0.46728515625,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.08087158203125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.02532958984375,
0.397705078125,
0.0555419921875,
0.397705078125,
0.0555419921875,
0.443115234375,
0.02532958984375,
0.443115234375
],
[
0.080078125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.08087158203125,
0.40771484375,
0.46728515625,
0.40771484375,
0.46728515625,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.08087158203125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.43505859375,
0.08087158203125,
0.43505859375
],
[
0.51953125,
0.425537109375,
0.56396484375,
0.425537109375,
0.56396484375,
0.43310546875,
0.51953125,
0.43310546875
],
[
0.5908203125,
0.425537109375,
0.615234375,
0.425537109375,
0.615234375,
0.43310546875,
0.5908203125,
0.43310546875
],
[
0.6357421875,
0.425537109375,
0.6953125,
0.425537109375,
0.6953125,
0.43310546875,
0.6357421875,
0.43310546875
],
[
0.71142578125,
0.425537109375,
0.76123046875,
0.425537109375,
0.76123046875,
0.43310546875,
0.71142578125,
0.43310546875
],
[
0.77392578125,
0.425537109375,
0.81884765625,
0.425537109375,
0.81884765625,
0.43310546875,
0.77392578125,
0.43310546875
],
[
0.8388671875,
0.425537109375,
0.8857421875,
0.425537109375,
0.8857421875,
0.43310546875,
0.8388671875,
0.43310546875
],
[
0.03350830078125,
0.443115234375,
0.0555419921875,
0.443115234375,
0.0555419921875,
0.4716796875,
0.03350830078125,
0.4716796875
],
[
0.07843017578125,
0.4375,
0.4697265625,
0.438232421875,
0.4697265625,
0.45263671875,
0.07843017578125,
0.451904296875
],
[
0.50244140625,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.080078125,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50341796875,
0.46533203125,
0.89208984375,
0.46533203125,
0.89208984375,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50341796875,
0.479248046875,
0.830078125,
0.479248046875,
0.830078125,
0.492431640625,
0.50341796875,
0.492431640625
],
[
0.0269622802734375,
0.498779296875,
0.053924560546875,
0.498779296875,
0.053924560546875,
0.5322265625,
0.0269622802734375,
0.5322265625
],
[
0.08087158203125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.08087158203125,
0.51318359375
],
[
0.080078125,
0.5146484375,
0.468017578125,
0.5146484375,
0.468017578125,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.50244140625,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.537109375,
0.50244140625,
0.537109375
],
[
0.0269622802734375,
0.53076171875,
0.056365966796875,
0.53076171875,
0.056365966796875,
0.580078125,
0.0269622802734375,
0.580078125
],
[
0.080078125,
0.52978515625,
0.46728515625,
0.52978515625,
0.46728515625,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.5009765625,
0.537109375,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.5009765625,
0.5517578125
],
[
0.080078125,
0.54345703125,
0.46728515625,
0.54345703125,
0.46728515625,
0.556640625,
0.080078125,
0.556640625
],
[
0.080078125,
0.5595703125,
0.4697265625,
0.5595703125,
0.4697265625,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.5009765625,
0.568359375,
0.890625,
0.5673828125,
0.890625,
0.58203125,
0.5009765625,
0.58349609375
],
[
0.080078125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.0294189453125,
0.583984375,
0.051483154296875,
0.58544921875,
0.0482177734375,
0.61572265625,
0.0269622802734375,
0.6142578125
],
[
0.080078125,
0.5888671875,
0.46728515625,
0.5888671875,
0.46728515625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50341796875,
0.60107421875,
0.8896484375,
0.60107421875,
0.8896484375,
0.611328125,
0.50341796875,
0.611328125
],
[
0.50244140625,
0.6142578125,
0.8935546875,
0.6142578125,
0.8935546875,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.0236968994140625,
0.60595703125,
0.05963134765625,
0.60595703125,
0.05963134765625,
0.66748046875,
0.0236968994140625,
0.66748046875
],
[
0.07843017578125,
0.619140625,
0.4697265625,
0.619140625,
0.4697265625,
0.63330078125,
0.07843017578125,
0.63330078125
],
[
0.035125732421875,
0.6298828125,
0.04412841796875,
0.6298828125,
0.04412841796875,
0.6337890625,
0.035125732421875,
0.6337890625
],
[
0.5009765625,
0.62744140625,
0.890625,
0.62890625,
0.890625,
0.64306640625,
0.5009765625,
0.642578125
],
[
0.080078125,
0.63525390625,
0.468017578125,
0.63525390625,
0.468017578125,
0.6484375,
0.080078125,
0.6484375
],
[
0.50341796875,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.08416748046875,
0.65087890625,
0.351318359375,
0.65087890625,
0.351318359375,
0.6640625,
0.08416748046875,
0.6640625
],
[
0.0269622802734375,
0.6572265625,
0.052276611328125,
0.65576171875,
0.056365966796875,
0.70068359375,
0.0310516357421875,
0.7021484375
],
[
0.50244140625,
0.65771484375,
0.890625,
0.6591796875,
0.890625,
0.673828125,
0.50244140625,
0.67236328125
],
[
0.50341796875,
0.6748046875,
0.88818359375,
0.6748046875,
0.88818359375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.51953125,
0.689453125,
0.890625,
0.68994140625,
0.890625,
0.705078125,
0.51953125,
0.7041015625
],
[
0.08087158203125,
0.6962890625,
0.20751953125,
0.6962890625,
0.20751953125,
0.7109375,
0.08087158203125,
0.7109375
],
[
0.50341796875,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.09722900390625,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.736328125,
0.09722900390625,
0.736328125
],
[
0.5009765625,
0.7197265625,
0.8896484375,
0.72021484375,
0.8896484375,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.73779296875,
0.4697265625,
0.73779296875,
0.4697265625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50244140625,
0.7373046875,
0.89208984375,
0.7373046875,
0.89208984375,
0.75048828125,
0.50244140625,
0.75048828125
],
[
0.080078125,
0.75390625,
0.468017578125,
0.75390625,
0.468017578125,
0.767578125,
0.080078125,
0.767578125
],
[
0.50341796875,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.08087158203125,
0.76806640625,
0.468017578125,
0.76806640625,
0.468017578125,
0.7822265625,
0.08087158203125,
0.7822265625
],
[
0.50244140625,
0.767578125,
0.89208984375,
0.767578125,
0.89208984375,
0.78076171875,
0.50244140625,
0.78076171875
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.796875,
0.08087158203125,
0.796875
],
[
0.50341796875,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.08087158203125,
0.798828125,
0.468017578125,
0.798828125,
0.468017578125,
0.8125,
0.08087158203125,
0.8125
],
[
0.50244140625,
0.798828125,
0.83251953125,
0.798828125,
0.83251953125,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.814453125,
0.4697265625,
0.814453125,
0.4697265625,
0.82763671875,
0.080078125,
0.82763671875
],
[
0.080078125,
0.8291015625,
0.468017578125,
0.8291015625,
0.468017578125,
0.84228515625,
0.080078125,
0.84228515625
],
[
0.52392578125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.52392578125,
0.8388671875
],
[
0.080078125,
0.84326171875,
0.443603515625,
0.84326171875,
0.443603515625,
0.85693359375,
0.080078125,
0.85693359375
],
[
0.53369140625,
0.84130859375,
0.89208984375,
0.84033203125,
0.89208984375,
0.85400390625,
0.53369140625,
0.85498046875
],
[
0.098876953125,
0.85986328125,
0.468017578125,
0.85986328125,
0.468017578125,
0.875,
0.098876953125,
0.875
],
[
0.53515625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.53515625,
0.87060546875
],
[
0.5361328125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.5361328125,
0.88623046875
],
[
0.10455322265625,
0.88720703125,
0.236083984375,
0.88720703125,
0.236083984375,
0.90087890625,
0.10455322265625,
0.90087890625
],
[
0.53515625,
0.88720703125,
0.68896484375,
0.88720703125,
0.68896484375,
0.90087890625,
0.53515625,
0.90087890625
]
] |
[
"Human-Inspired Facial Sketch Synthesis with Dynamic Adaptation",
"Fei Gao', Yifan Zhu?, Chang Jiang?2, Nannan Wang3*",
"1Hangzhou Institute of Technology, Xidian University 2Hangzhou Dianzi University 3Xidian University",
"fgao@xidian.edu.cn,2961695289@qq.com,jc233@hdu.edu.cn,nnwang@xidian.edu.cn",
"",
"Abstract",
"Facial sketch synthesis (FSS) aims to generate a vivid",
"zdas",
"sketch portrait from a given facial photo.Existing FSS",
"methods merely rely on 2D representations of facial seman-",
"tic or appearance.However, professional human artists",
"usually use outlines or shadings to covey 3D geometry.",
"Thus facial 3D geometry (e.g. depth map) is extremely im-",
"LAD",
"portant for Fss. Besides, different artists may use diverse",
"drawing techniques and create multiple styles of sketches;",
"but the style is globally consistent in a sketch.Inspired",
"InputPhota",
"Depth",
"Sketh(byArtist",
"Ours(style1)",
"Durs(Style2)",
"Ours(Style3)",
"'SO",
"by such observations, in this paper, we propose a novel",
"Figure 1: Ilustration of facial photos, depth maps, multi-",
"Human-Inspired Dynamic Adaptation (HIDA) method. Spe-",
"cially, we propose to dynamically modulate neuron acti-",
"style facial sketchesdrawn byhumanartists[8l,and the",
"vations based on a joint consideration of both facial 3D",
"corresponding results synthesized by our method.",
"A9",
"geometry and 2D appearance, as well as globally consis-",
"tent style control. Besides, we use deformable convolutions",
"inspired by the excellent success of Generative Adver-",
"1700",
"at coarse-scales to align deep features, for generating ab-",
"sarial Networks (GANs) [18]. Specially, researchers",
"stract and distinct outlines. Experiments show that HIDA",
"can generate high-quality sketches in multiple styles, and",
"have proposed various techniques, including embedding",
"image prior [43], semi-supervised learning [5], self-",
"significantly outperforms previous methods, over a large",
"609",
"range of challenging faces. Besides, HIDA allows pre-",
"attention/transformer based methods [11, 47, 9], hierarchi-",
"cise style control of the synthesized sketch, and general-",
"calGANs28,40.8l.compositionassistance[371.and se-",
"mantic adaptive normalization [21], to boost the quality of",
"",
"izes well to natural scenes and other artistic styles. Our",
"",
"synthesized sketches. However, all these methods merely",
"code and results have been released online at: https:",
"use 2D appearance or semantic representations of the input",
"//github.com/AiArt-HDU/HIDA.",
"X",
"photo. They may fail to handle serious variations in appear-",
"ance, such as the pose, lighting, expression, and skin color.",
"To tackle this challenge, we propose a novel method, in-",
"1. Introduction",
"spired by how human artists draw a sketch. We observe that",
"Making computers create arts like human beings, is a",
"facial 3D geometry plays a significant role in human artists",
"longstanding and challenging topic, in the artificial intel-",
"drawing process. Besides, a professional human artist con-",
"ligence (AI) area [2]. To this end, researchers have made",
"siders comprehensive information, including facial 3D ge-",
"great efforts and proposed numerous methods, such as neu-",
"ometry, 2D appearance, and the artistic style, to execute a",
"ral style transfer (NST) [17] and image-to-image translation",
"sketch portrait. We summarize the drawing methodologies",
"(121T) [18, 46]. These methods mainly tackle cluttered im-",
"ofhuman artists[22]intothefollowingfourfolds:",
"age styles, such as oil paintings [17]. In this paper, we are",
"interested in creating artistic sketches from facial photos,",
": Local 3D geometry conveyor: First, artists typically",
"which is referred to as face sketch synthesis (FSS) [34].",
"use abstract and deformable outlines to characterize",
"For now, there has been significant progress in Fss",
"major geometry, and use different shading methodolo-",
"gies, e.g. hatching, blending, and stippling, to convey",
"*Corresponding Author",
"local 3D structures [3]."
] |
{
"arxiv_id": "2309.00216",
"language": "en",
"page_idx": 0
}
|
||
2309.00216-en-1
|
2309.00216-en-1.jpg
|
$\cdot$ Local 2D appearance representation: Second, artists may use different shading or tonal techniques to represent local 2D facial appearance, so as to depict variations in lighting, color, texture, etc.
$\cdot$ Sketches in diverse styles: Third, different artists may
use diverse drawing methods and create multiple styles of sketches. In other words, they may use divergent textures to represent the same facial area. Fig. 1 shows three styles of sketches drawn by artists[8]. Obviously, Stylel is extremely abstract and mainly contains sketchy outlines. In contrast, Style3 depicts facial 3D geometry with a lot of shading textures.
$\cdot$ Globally consistent style: Finally, the style of pencildrawing is usually consistent in a single sketch. As shown in Fig. 1, although there are distinct inter-style divergences, the style of pencil-drawing is globally consistent across different regions inside each sketch.
Inspired by these observations, we seek to guide the synthesis of sketch portraits by using comprehensive information, including facial 3D geometry and 2D appearance, as well as global style control. In the implementation, given a facial photo, we use the depth map to represent its 3D geometry, and use the encoding features to represent its 2D appearance. Afterwards, we combine them with a style map to dynamically modulate deep features for generating a sketch. Inspired by the success of SPADE[27] in style control [41] and the local flexibility of dynamic neural networks [14], we propose to dynamically modulate neuron activations, based on a joint consideration of all these information. Such modulation is conducted though both dynamic normalization and activation. Specially, we propose a novel dynamic activation function, termed Informative ACON (InfoACON), and a dynamic normalization module, termed DySPADE. In addition, we use deformable convolutions [7] to align deep features [16] at coarse scales for generating abstract and distinct sketchy outlines. Initially, the dynamic adaptation and deformation simulate the flexibility and abstract process of human artists during drawing.
Based on the above mentioned contributions, we build a Human-Inspired Dynamic Adaptation (HIDA) method for FSS. We conduct experiments on several challenging datasets, including the FS2K [8], the FFHQ [20], and a collection of faces in-the-wild. Our method outperforms stateof-the-art (SOTA) methods both qualitatively and quantitatively. Besides, our method allows precise style control and can produce high-quality sketches in multiple styles. Even for faces with serious variations, the synthesized sketches present realistic textures and preserve facial geometric details. In addition, extensive ablation studies demonstrate the effectiveness of the proposed dynamic and adaptive modulation techniques. Finally, our model, although trained for faces, can generate high-quality sketches for natural scenes.
# $2. \textbf{ Related Works}$
Our work is related to GANs-based FSS methods. Besides, our method is highly inspired by semantic adaptive normalization [27] and dynamic activation [25].
GANs-based FSS. The latest FSS methods are typically based on GANs [18, 12], where the mapping from a facial photo to a sketch is modeled as an image-to-image translation task[18]. Some latest methods use 2D semantic information to guide the generation process. For example, Yu et al. [37] propose a stacked composition-aided GANs to boost quality of details. Inspired by the great success of spatially adaptive (de)normalization (SPADE) [27] in semantic image generation, Wang et al.[47] and Qi et al. [29] spatially modulate decoding features according to facial parsing masks. Li et al.[21] propose an enhanced SPADE (eSPADE) by using both facial parsing masks and encoding features for feature modulation.
Recently, researchers seek to solve the challenge of unconstrained faces by constructing large datasets. Fan et al. [8] release a challenging FS2K dataset, which consists of multi-style sketches for faces with diverse variations. Nie et al.[26] propose a novel WildSketch dataset and a Perception- Adaptive Network (PANet). In PANet, deformable feature alignment (DFA) and patch-level adaptive convolution are used. Different from [26], we analyze the effects of DFA, and only use DFA over the coarse scales. Besides, we propose to dynamically modulate neuron activations based on facial depth and artistic style.
Semantic Adaptive Normalization. Recently, Park et al. [27] propose to modulate deep features based on semantic layouts for semantic image synthesis. In SPADE, deep features are modulated based on semantic layouts. Afterwards, Zhu et al. [48] propose Semantic Region-Adaptive Normalization (SEAN) to control the style of each semantic region individually. To boost the efficiency of SPADE, Tan et al. [32] propose a Class-Adaptive (DE)Normalization (CLADE) layer by replacing the modulation networks with class-level modulating parameters. All these adaptive normalization layers use 2D semantic maps and show amazing performance in generating photo-realistic images [24] and face sketches [37]. In this paper, we use pix-wise dynamic activation in the normalization block, so that the modulating parameters would flexibly adapt to local information. Experimental results show that the dynamic normalization are essential for detailed synthesis of facial sketches
Dynamic Activations. Recently, Chen et al. [6] propose a Dynamic ReLU (DY-ReLU) function, where parameters in Leaky ReLU are learned from all input elements. Ma et al.[25] propose a costumed activation function, termed ACON, which automatically decides whether a neuron is active or not. ACON has several variants, among which the pixel-wise version of metaACON shows remarkable performance. Give a neuron activation $x$, the output of metaA-
|
[
[
0.098876953125,
0.09027099609375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.098876953125,
0.10479736328125
],
[
0.50244140625,
0.0889892578125,
0.64794921875,
0.0889892578125,
0.64794921875,
0.10418701171875,
0.50244140625,
0.10418701171875
],
[
0.11273193359375,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.11273193359375,
0.11932373046875
],
[
0.1094970703125,
0.121826171875,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.1094970703125,
0.1363525390625
],
[
0.52294921875,
0.11614990234375,
0.890625,
0.11614990234375,
0.890625,
0.1300048828125,
0.52294921875,
0.1300048828125
],
[
0.50341796875,
0.1319580078125,
0.89208984375,
0.1319580078125,
0.89208984375,
0.1458740234375,
0.50341796875,
0.1458740234375
],
[
0.111083984375,
0.1383056640625,
0.346435546875,
0.1383056640625,
0.346435546875,
0.1514892578125,
0.111083984375,
0.1514892578125
],
[
0.5,
0.1458740234375,
0.81884765625,
0.146484375,
0.81884765625,
0.1610107421875,
0.5,
0.160400390625
],
[
0.09722900390625,
0.1590576171875,
0.468017578125,
0.1610107421875,
0.468017578125,
0.175537109375,
0.09722900390625,
0.173583984375
],
[
0.521484375,
0.160400390625,
0.8896484375,
0.1610107421875,
0.8896484375,
0.175537109375,
0.521484375,
0.1749267578125
],
[
0.111083984375,
0.1767578125,
0.4697265625,
0.1767578125,
0.4697265625,
0.1900634765625,
0.111083984375,
0.1900634765625
],
[
0.5009765625,
0.175537109375,
0.890625,
0.1767578125,
0.890625,
0.1912841796875,
0.5009765625,
0.1900634765625
],
[
0.11273193359375,
0.1912841796875,
0.4697265625,
0.1912841796875,
0.4697265625,
0.2052001953125,
0.11273193359375,
0.2052001953125
],
[
0.50244140625,
0.1925048828125,
0.890625,
0.1925048828125,
0.890625,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.11273193359375,
0.20703125,
0.4697265625,
0.20703125,
0.4697265625,
0.2203369140625,
0.11273193359375,
0.2203369140625
],
[
0.50244140625,
0.20703125,
0.89208984375,
0.20703125,
0.89208984375,
0.2203369140625,
0.50244140625,
0.2203369140625
],
[
0.1094970703125,
0.2215576171875,
0.471435546875,
0.2203369140625,
0.471435546875,
0.23486328125,
0.1094970703125,
0.236083984375
],
[
0.5009765625,
0.222900390625,
0.890625,
0.222900390625,
0.890625,
0.236083984375,
0.5009765625,
0.236083984375
],
[
0.111083984375,
0.2374267578125,
0.46728515625,
0.2374267578125,
0.46728515625,
0.250732421875,
0.111083984375,
0.250732421875
],
[
0.50341796875,
0.2374267578125,
0.89208984375,
0.2374267578125,
0.89208984375,
0.250732421875,
0.50341796875,
0.250732421875
],
[
0.11273193359375,
0.251953125,
0.468017578125,
0.251953125,
0.468017578125,
0.26513671875,
0.11273193359375,
0.26513671875
],
[
0.50244140625,
0.251953125,
0.89208984375,
0.251953125,
0.89208984375,
0.26513671875,
0.50244140625,
0.26513671875
],
[
0.111083984375,
0.267578125,
0.373291015625,
0.267578125,
0.373291015625,
0.281005859375,
0.111083984375,
0.281005859375
],
[
0.50244140625,
0.267578125,
0.8896484375,
0.267578125,
0.8896484375,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.50244140625,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.50244140625,
0.296630859375
],
[
0.1004638671875,
0.290283203125,
0.46728515625,
0.290283203125,
0.46728515625,
0.3037109375,
0.1004638671875,
0.3037109375
],
[
0.50341796875,
0.29736328125,
0.890625,
0.29736328125,
0.890625,
0.311279296875,
0.50341796875,
0.311279296875
],
[
0.111083984375,
0.30615234375,
0.4697265625,
0.30615234375,
0.4697265625,
0.3193359375,
0.111083984375,
0.3193359375
],
[
0.5009765625,
0.3125,
0.890625,
0.3125,
0.890625,
0.32568359375,
0.5009765625,
0.32568359375
],
[
0.11358642578125,
0.322509765625,
0.46728515625,
0.322509765625,
0.46728515625,
0.333251953125,
0.11358642578125,
0.333251953125
],
[
0.50341796875,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.34228515625,
0.50341796875,
0.34228515625
],
[
0.111083984375,
0.335205078125,
0.46728515625,
0.335205078125,
0.46728515625,
0.348388671875,
0.111083984375,
0.348388671875
],
[
0.50244140625,
0.343505859375,
0.775390625,
0.343505859375,
0.775390625,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.11273193359375,
0.3515625,
0.461669921875,
0.3515625,
0.461669921875,
0.365478515625,
0.11273193359375,
0.365478515625
],
[
0.521484375,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.521484375,
0.371826171875
],
[
0.09722900390625,
0.371826171875,
0.46728515625,
0.371826171875,
0.46728515625,
0.386474609375,
0.09722900390625,
0.386474609375
],
[
0.50244140625,
0.373779296875,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.50244140625,
0.386962890625
],
[
0.07843017578125,
0.386962890625,
0.468017578125,
0.38818359375,
0.468017578125,
0.40283203125,
0.07843017578125,
0.401611328125
],
[
0.50244140625,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.080078125,
0.404052734375,
0.471435546875,
0.404052734375,
0.471435546875,
0.417236328125,
0.080078125,
0.417236328125
],
[
0.5048828125,
0.40478515625,
0.890625,
0.40478515625,
0.890625,
0.415283203125,
0.5048828125,
0.415283203125
],
[
0.08087158203125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.43115234375,
0.08087158203125,
0.43115234375
],
[
0.50244140625,
0.41845703125,
0.89208984375,
0.41845703125,
0.89208984375,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.08087158203125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.08087158203125,
0.447509765625
],
[
0.5048828125,
0.43505859375,
0.8896484375,
0.43505859375,
0.8896484375,
0.44580078125,
0.5048828125,
0.44580078125
],
[
0.07843017578125,
0.449462890625,
0.468017578125,
0.447509765625,
0.468017578125,
0.4609375,
0.07843017578125,
0.46337890625
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.07843017578125,
0.464111328125,
0.4697265625,
0.46337890625,
0.4697265625,
0.477783203125,
0.07843017578125,
0.479248046875
],
[
0.5,
0.462158203125,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.5,
0.4765625
],
[
0.07843017578125,
0.477783203125,
0.4697265625,
0.479248046875,
0.4697265625,
0.49365234375,
0.07843017578125,
0.492431640625
],
[
0.50244140625,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.5087890625,
0.80810546875,
0.5087890625,
0.80810546875,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.08087158203125,
0.52587890625,
0.468017578125,
0.52587890625,
0.468017578125,
0.5361328125,
0.08087158203125,
0.5361328125
],
[
0.52294921875,
0.52587890625,
0.89208984375,
0.52587890625,
0.89208984375,
0.5361328125,
0.52294921875,
0.5361328125
],
[
0.08087158203125,
0.54150390625,
0.46728515625,
0.54150390625,
0.46728515625,
0.5517578125,
0.08087158203125,
0.5517578125
],
[
0.50341796875,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.07843017578125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.07843017578125,
0.568359375
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.5693359375,
0.4697265625,
0.5693359375,
0.4697265625,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.5693359375,
0.890625,
0.5693359375,
0.890625,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.080078125,
0.58544921875,
0.4697265625,
0.58544921875,
0.4697265625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50341796875,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.08087158203125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.08087158203125,
0.6162109375,
0.46728515625,
0.6162109375,
0.46728515625,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.07843017578125,
0.62890625,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.07843017578125,
0.64306640625
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.08251953125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.6572265625,
0.08251953125,
0.6572265625
],
[
0.50341796875,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.6875,
0.080078125,
0.6875
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.098876953125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.07843017578125,
0.71826171875,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.07843017578125,
0.73291015625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.7666015625,
0.84716796875,
0.76513671875,
0.84716796875,
0.77880859375,
0.5009765625,
0.77978515625
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.51953125,
0.77978515625,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.51953125,
0.79443359375
],
[
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
],
[
0.08087158203125,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.5048828125,
0.85302734375
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
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,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
". Local 2D appearance representation: Second, artists",
"2.Related Works",
"may use different shading ortonal technigues torepre-",
" sent local 2D facial appearance, so as to depict varia-",
"Our work is related to GANs-based FSS methods. Be-",
"sides, our method is highly inspired by semantic adaptive",
"tions in lighting, color, texture, etc.",
"normalization [27] and dynamic activation [25].",
". Sketches in diverse styles: Third, different artists may",
"GANs-based FSS. The latest FSS methods are typically",
"use diverse drawing methods and create multiple styles",
"based on GANs [18, 12], where the mapping from a facial",
"of sketches.In other words, they may use divergent",
"photo to a sketch is modeled as an image-to-image trans-",
"textures to represent the same facial area. Fig. I shows",
"lation task [18]. Some latest methods use 2D semantic in-",
"three styles of sketches drawn by artists [8]. Obvi-",
"formation to guide the generation process.For example,",
"ously, Stylel is extremely abstract and mainly contains",
"Yu et al. [37] propose a stacked composition-aided GANs",
"sketchy outlines. In contrast, Style3 depicts facial 3D",
"to boost quality of details. Inspired by the great success",
"geometry with a lot of shading textures.",
"of spatially adaptive (de)normalization (SPADE) [27] in se-",
"mantic image generation, Wang et al. [47] and Qi et al.",
"• Globally consistent style: Finally, the style of pencil-",
"[29] spatially modulate decoding features according to fa-",
"drawing is usually consistent in a single sketch. As",
"cial parsing masks. Li et al.[2il propose an enhanced",
"shown in Fig. l, although there are distinct inter-style",
"SPADE (eSPADE) by using both facial parsing masks and",
"divergences, the style of pencil-drawing is globally",
"encoding features for feature modulation.",
"consistent across different regions inside each sketch.",
"Recently, researchers seek to solve the challenge of un-",
"Inspired by these observations, we seek to guide the syn-",
"constrained faces by constructing large datasets. Fan et",
"thesis of sketch portraits by using comprehensive informa-",
"al. [8] release a challenging FS2K dataset, which con-",
"tion, including facial 3D geometry and 2D appearance, as",
"sists of multi-style sketches for faces with diverse varia-",
"well as global stylecontrol.Intheimplementation,givena",
"tions. Nie et al. [26] propose a novel WildSketch dataset",
"facial photo, we use the depth map to represent its 3D ge-",
"andaPerception-AdaptiveNetwork(PANet).InPANet,de-",
"ometry, and use the encoding features to represent its 2D",
"formable feature alignment (DFA) and patch-level adaptive",
"appearance.Afterwards, we combinethem with a style",
"convolution are used. Different from [26], we analyze the",
"map to dynamically modulate deep features for generating",
"effects of DFA, and only use DFA over the coarse scales.",
"a sketch. Inspired by the success of SPADE [27] in style",
"Besides, we propose to dynamically modulate neuron acti-",
"control [41] and the local flexibility of dynamic neural net-",
"vations based on facial depth and artistic style.",
"worksJ14l.we proposeto dynamically modulate neuron",
"Semantic Adaptive Normalization.Recently,Park et",
"activations,based on a joint consideration of all these in-",
"al. [27] propose to modulate deep features based on seman-",
"formation. Such modulation is conducted though both dy-",
"tic layouts for semantic image synthesis. In SPADE, deep",
"namic normalization and activation. Specially, we propose",
"features are modulated based on semantic layouts. After-",
"a novel dynamic activation function, termed Informative",
"wards, Zhu et al. [48] propose Semantic Region-Adaptive",
"ACON (InfoACON), and a dynamic normalization module,",
"Normalization (SEAN) to control the style of each semantic",
"termed DySPADE.In addition, weuse deformableconvo-",
"region individually. To boost the efficiency of SPADE, Tan",
"lutions [7] to align deep features [16] at coarse scales for",
"et al. [32] propose a Class-Adaptive (DE)Normalization",
"generating abstract and distinct sketchy outlines.Initially,",
"(CLADE) layer by replacing the modulation networks with",
"the dynamic adaptation and deformation simulate the flexi-",
"class-level modulating parameters. All these adaptive nor-",
"bility and abstractprocess of human artists during drawing.",
"malizationlayers use2D semanticmaps and show amazing",
"Based on the above mentioned contributions, we build",
"performance in generating photo-realistic images [24] and",
"a Human-Inspired Dynamic Adaptation (HIDA) method",
"face sketches [37]. In this paper, we use pix-wise dynamic",
"for FSS. We conduct experiments on several challenging",
"activation in the normalization block, so that the modulat-",
"datasets, including the FS2K [8], the FFHQ [20], and a col-",
"ing parameters would flexibly adapt to local information.",
"lection of faces in-the-wild. Our method outperforms state-",
"Experimental results show that the dynamic normalization",
"of-the-art (SOTA) methods both qualitatively and quantita-",
"are essential for detailed synthesis of facial sketches.",
"tively.Besides, our method allows precise style control and",
"Dynamic Activations. Recently, Chen et al. [6] propose",
"can produce high-quality sketches in multiple styles. Even",
"a Dynamic ReLU (DY-ReLU) function, where parameters",
"for faces with serious variations, the synthesized sketches",
"in Leaky ReLU are learned from all input elements. Ma",
"present realistic textures and preserve facial geometric de-",
"et al. [25] propose a costumed activation function, termed",
"tails.In addition,extensive ablation studies demonstratethe",
"ACON, which automatically decides whether a neuron is",
"effectiveness of theproposed dynamic and adaptivemodu-",
"active or not. ACONhas several variants,among which the",
"lation techniques. Finally, our model, although trained for",
"pixel-wise version of metaACON shows remarkable perfor-",
"faces, can generate high-quality sketches for natural scenes.",
"mance. Give a neuron activation c, the output of metaA-"
] |
{
"arxiv_id": "2309.00216",
"language": "en",
"page_idx": 1
}
|
||
2309.00216-en-2
|
2309.00216-en-2.jpg
|
<FigureHere>
Figure 2: Pipeline of the proposed $Human-Inspired$ Dynamic Adaptation (HIDA) method for facial sketch synthesis. (a) The
overall generator architecture, (b) an decoding layer with DySPADE, InfoACON, and DOG.
CON is formulated as:
$$y=(p_1-p_2)\cdot\sigma(\theta(p_1-p_2)x)+p_2x,$$
(1)
where $\theta=\sigma(x),\sigma$ is a Sigmoid function, $p_1$, and $p_2$ are learnable parameters. In this paper, we use metaACON, instead of ReLU or Leaky ReLU, in part of our networks. Besides, we propose to learn spatially-adaptive parameter $\theta$ according to the 3D geometry, 2D appearance, and global style control. Our activation function proves boosting the performance and allowing precise style control.
## $3. \textbf{ The Proposed}$
We aim to translate a facial photo X to a sketch $\mathbf{Y}_{s}$, in style $s$,drawn by an artist. Here $s=1,2,...,S$ is a style label, $S$ is the total number of styles. In this work, we seek to guide the synthesis of sketch portraits by using comprehensive facial information, including both 3D geometry and 2D appearance, as well as the global style control. Given a facial photo, we use the corresponding depth map D to represent its 3D geometry. Afterwards, we combine them with a global style map S to decode a facial sketch. In this way, our goal is formulated as learning a mapping from $\{\mathbf{X},\mathbf{D},\mathbf{S}\}$ to $Y_{s}$,i.e. $G:\{\mathbf{X},\mathbf{D},\mathbf{S}\}\mapsto\mathbf{Y}_{s}.$
To supervise our model, it is necessary to obtain depth maps for input facial photos. However, it is usually impossible to obtain ground truth depth information in practical applications. Therefore we use state-of-the-art (SOTA) depth prediction methods to estimate the depth map of an input facial photo. In practice, we use 3DDFA[13] as the depth predictor, because it has been widely used and shown excellent performance in various 3D face reconstruction tasks.
The overall pipeline of our model is as shown in Fig. 2. It contains an off-line facial depth predictor $P$, a generator $G$, and a patch-wise discriminator $D.$ In addition to a facial sketch, we enforce $G$ to reconstruct the input depth
map D from features representing the sketch. In this way, the generated sketch $\hat{\mathbf{Y}}_s$ would convey the 3D geometry of X. Besides, we boost the capacity of generator by using a dynamic normalization module and a dynamic activation function. Finally, to formulate the abstraction methodology of human artists in drawing sketchy outlines, we propose using deformable convolutions to align features at coarse scales. Details will be introduced bellow.
$3. 1. \textbf{ Informative and Dynamic Adaptation ( IDA) }$
To simulate the drawing methodology of human artist, we first propose a novel Informative and Dynamic Adaptation (IDA) module, to modulate deep features based on a combination of the facial depth map D, the style map $\mathbf{S}$, and the appearance representations A, i.e. $\{\mathbf{D},\mathbf{S},\mathbf{A}\}.$ Specially, we propose a novel dynamic activation function, termed Informative ACON (InfoACON), and a dynamic normalization module, termed DySPADE.
Informative ACON (InfoACON). The original metaACON function automatically allows whether a neuron is active or not, based on its value, as previously presented in Eq. 1. During the drawing process, a human artist typically decides whether to draw a stroke or not based on the 3D geometry, 2D appearance, and style type. Inspired by this observation, we propose to learn the parameter $\theta$ in Eq. 1 $\operatorname{from}\left\{\mathbf{D},\mathbf{S},\mathbf{A}\right\}$,i.e.
(2)
$$\theta=\sigma(\phi_\theta(\mathrm{Cat}(\mathbf{D},\mathbf{S},\mathbf{A})),$$
where $\phi_{\theta}$ is a two-layer Convolutional network (Fig. 3). We refer to the modified metaACON function as Informative ACON (InfoACON). In our networks, we apply this InfoACON function in all the decoding layers. In this way, the decoder would pixel-wisely decides whether to depict a stroke, or the type of a stroke, in a generated sketch.
Dynamic Normalization (DySPADE). Following [27], we additionally transform neuron activations by shifting
|
[
[
0.285888671875,
0.09783935546875,
0.3203125,
0.09783935546875,
0.3203125,
0.10479736328125,
0.285888671875,
0.10479736328125
],
[
0.393798828125,
0.0997314453125,
0.438720703125,
0.0997314453125,
0.438720703125,
0.10986328125,
0.393798828125,
0.10986328125
],
[
0.833984375,
0.0947265625,
0.88134765625,
0.0947265625,
0.88134765625,
0.1060791015625,
0.833984375,
0.1060791015625
],
[
0.19775390625,
0.10162353515625,
0.2296142578125,
0.10162353515625,
0.2296142578125,
0.10919189453125,
0.19775390625,
0.10919189453125
],
[
0.28515625,
0.106689453125,
0.314453125,
0.106689453125,
0.314453125,
0.111083984375,
0.28515625,
0.111083984375
],
[
0.396240234375,
0.10986328125,
0.450927734375,
0.10986328125,
0.450927734375,
0.117431640625,
0.396240234375,
0.117431640625
],
[
0.740234375,
0.1060791015625,
0.75634765625,
0.10986328125,
0.74658203125,
0.1319580078125,
0.73046875,
0.1275634765625
],
[
0.281005859375,
0.1142578125,
0.301513671875,
0.1142578125,
0.301513671875,
0.121826171875,
0.281005859375,
0.121826171875
],
[
0.75341796875,
0.111083984375,
0.7587890625,
0.111083984375,
0.7587890625,
0.11614990234375,
0.75341796875,
0.11614990234375
],
[
0.468017578125,
0.12371826171875,
0.52685546875,
0.12371826171875,
0.52685546875,
0.1275634765625,
0.468017578125,
0.1275634765625
],
[
0.47216796875,
0.1307373046875,
0.52294921875,
0.1307373046875,
0.52294921875,
0.1383056640625,
0.47216796875,
0.1383056640625
],
[
0.708984375,
0.1307373046875,
0.71826171875,
0.1307373046875,
0.71826171875,
0.1427001953125,
0.708984375,
0.1427001953125
],
[
0.744140625,
0.126953125,
0.76708984375,
0.1363525390625,
0.7490234375,
0.1634521484375,
0.72607421875,
0.1546630859375
],
[
0.0906982421875,
0.1337890625,
0.134033203125,
0.1337890625,
0.134033203125,
0.1673583984375,
0.0906982421875,
0.1673583984375
],
[
0.54833984375,
0.1319580078125,
0.5537109375,
0.1319580078125,
0.5537109375,
0.136962890625,
0.54833984375,
0.136962890625
],
[
0.66015625,
0.1337890625,
0.701171875,
0.1363525390625,
0.69921875,
0.1546630859375,
0.65869140625,
0.15283203125
],
[
0.79150390625,
0.1351318359375,
0.82177734375,
0.1337890625,
0.82177734375,
0.141357421875,
0.79150390625,
0.1427001953125
],
[
0.346435546875,
0.1351318359375,
0.3798828125,
0.1351318359375,
0.3798828125,
0.1395263671875,
0.346435546875,
0.1395263671875
],
[
0.33251953125,
0.1395263671875,
0.38232421875,
0.1395263671875,
0.38232421875,
0.149658203125,
0.33251953125,
0.149658203125
],
[
0.47802734375,
0.147705078125,
0.51708984375,
0.147705078125,
0.51708984375,
0.175537109375,
0.47802734375,
0.175537109375
],
[
0.6337890625,
0.15087890625,
0.650390625,
0.15087890625,
0.650390625,
0.1666259765625,
0.6337890625,
0.1666259765625
],
[
0.62353515625,
0.1572265625,
0.6298828125,
0.1572265625,
0.6298828125,
0.1634521484375,
0.62353515625,
0.1634521484375
],
[
0.779296875,
0.15283203125,
0.80126953125,
0.1610107421875,
0.79150390625,
0.173583984375,
0.76953125,
0.1654052734375
],
[
0.73779296875,
0.1622314453125,
0.75341796875,
0.1634521484375,
0.7509765625,
0.1849365234375,
0.734375,
0.18310546875
],
[
0.75634765625,
0.1654052734375,
0.7646484375,
0.1654052734375,
0.7646484375,
0.1749267578125,
0.75634765625,
0.1749267578125
],
[
0.78515625,
0.1685791015625,
0.7958984375,
0.1685791015625,
0.7958984375,
0.1787109375,
0.78515625,
0.1787109375
],
[
0.66015625,
0.1717529296875,
0.67236328125,
0.1717529296875,
0.67236328125,
0.179931640625,
0.66015625,
0.179931640625
],
[
0.09234619140625,
0.1749267578125,
0.12091064453125,
0.1749267578125,
0.12091064453125,
0.18310546875,
0.09234619140625,
0.18310546875
],
[
0.6396484375,
0.17041015625,
0.646484375,
0.17041015625,
0.646484375,
0.1767578125,
0.6396484375,
0.1767578125
],
[
0.77783203125,
0.175537109375,
0.79345703125,
0.175537109375,
0.79345703125,
0.2171630859375,
0.77783203125,
0.2171630859375
],
[
0.0906982421875,
0.18310546875,
0.131591796875,
0.18310546875,
0.131591796875,
0.1900634765625,
0.0906982421875,
0.1900634765625
],
[
0.283447265625,
0.208984375,
0.325927734375,
0.208984375,
0.325927734375,
0.216552734375,
0.283447265625,
0.216552734375
],
[
0.4755859375,
0.2279052734375,
0.52392578125,
0.2279052734375,
0.52392578125,
0.23486328125,
0.4755859375,
0.23486328125
],
[
0.0849609375,
0.23486328125,
0.1380615234375,
0.23486328125,
0.1380615234375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.77685546875,
0.231689453125,
0.8505859375,
0.231689453125,
0.8505859375,
0.242431640625,
0.77685546875,
0.242431640625
],
[
0.48779296875,
0.238037109375,
0.5107421875,
0.238037109375,
0.5107421875,
0.242431640625,
0.48779296875,
0.242431640625
],
[
0.69921875,
0.2442626953125,
0.71142578125,
0.2442626953125,
0.71142578125,
0.248779296875,
0.69921875,
0.248779296875
],
[
0.77783203125,
0.24365234375,
0.859375,
0.24365234375,
0.859375,
0.250732421875,
0.77783203125,
0.250732421875
],
[
0.775390625,
0.25244140625,
0.8857421875,
0.25244140625,
0.8857421875,
0.26318359375,
0.775390625,
0.26318359375
],
[
0.281982421875,
0.268310546875,
0.301513671875,
0.268310546875,
0.301513671875,
0.281005859375,
0.281982421875,
0.281005859375
],
[
0.39453125,
0.2474365234375,
0.45751953125,
0.2474365234375,
0.45751953125,
0.25830078125,
0.39453125,
0.25830078125
],
[
0.705078125,
0.268310546875,
0.7236328125,
0.268310546875,
0.7236328125,
0.281005859375,
0.705078125,
0.281005859375
],
[
0.08087158203125,
0.29345703125,
0.890625,
0.29345703125,
0.890625,
0.30810546875,
0.08087158203125,
0.30810546875
],
[
0.07843017578125,
0.309326171875,
0.68115234375,
0.30810546875,
0.68115234375,
0.322509765625,
0.07843017578125,
0.323974609375
],
[
0.08087158203125,
0.3515625,
0.2296142578125,
0.3515625,
0.2296142578125,
0.365478515625,
0.08087158203125,
0.365478515625
],
[
0.50341796875,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50341796875,
0.365478515625
],
[
0.50341796875,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.50341796875,
0.380126953125
],
[
0.143798828125,
0.380126953125,
0.408447265625,
0.380126953125,
0.408447265625,
0.393310546875,
0.143798828125,
0.393310546875
],
[
0.4453125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.392578125,
0.4453125,
0.392578125
],
[
0.5009765625,
0.380615234375,
0.89208984375,
0.3818359375,
0.89208984375,
0.396484375,
0.5009765625,
0.395751953125
],
[
0.50341796875,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.50341796875,
0.410400390625
],
[
0.08087158203125,
0.40771484375,
0.46728515625,
0.40771484375,
0.46728515625,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.5009765625,
0.410400390625,
0.890625,
0.412353515625,
0.890625,
0.4267578125,
0.5009765625,
0.4248046875
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.08087158203125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.08087158203125,
0.46728515625
],
[
0.50244140625,
0.45703125,
0.775390625,
0.45703125,
0.775390625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50244140625,
0.479736328125,
0.87109375,
0.479736328125,
0.87109375,
0.494384765625,
0.50244140625,
0.494384765625
],
[
0.07843017578125,
0.5,
0.39208984375,
0.5,
0.39208984375,
0.51318359375,
0.07843017578125,
0.51318359375
],
[
0.521484375,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.521484375,
0.517578125
],
[
0.5048828125,
0.52099609375,
0.8896484375,
0.52099609375,
0.8896484375,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.080078125,
0.52880859375,
0.21728515625,
0.52880859375,
0.21728515625,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.50341796875,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.098876953125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.5673828125,
0.098876953125,
0.5673828125
],
[
0.50341796875,
0.5498046875,
0.890625,
0.5498046875,
0.890625,
0.56298828125,
0.50341796875,
0.56298828125
],
[
0.5009765625,
0.564453125,
0.890625,
0.56298828125,
0.890625,
0.57763671875,
0.5009765625,
0.5791015625
],
[
0.08087158203125,
0.57080078125,
0.465576171875,
0.57080078125,
0.465576171875,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.080078125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50341796875,
0.5947265625,
0.89208984375,
0.5947265625,
0.89208984375,
0.60791015625,
0.50341796875,
0.60791015625
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.60986328125,
0.77783203125,
0.60986328125,
0.77783203125,
0.62353515625,
0.50244140625,
0.62353515625
],
[
0.080078125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.080078125,
0.642578125
],
[
0.521484375,
0.625,
0.890625,
0.625,
0.890625,
0.63818359375,
0.521484375,
0.63818359375
],
[
0.50244140625,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.08087158203125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.5009765625,
0.65380859375,
0.89208984375,
0.65478515625,
0.89208984375,
0.66943359375,
0.5009765625,
0.66845703125
],
[
0.08087158203125,
0.6611328125,
0.468017578125,
0.6611328125,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.5009765625,
0.66943359375,
0.8896484375,
0.66943359375,
0.8896484375,
0.68310546875,
0.5009765625,
0.68310546875
],
[
0.07843017578125,
0.689453125,
0.4697265625,
0.68798828125,
0.4697265625,
0.70263671875,
0.07843017578125,
0.7041015625
],
[
0.5009765625,
0.6845703125,
0.8896484375,
0.6845703125,
0.8896484375,
0.69775390625,
0.5009765625,
0.69775390625
],
[
0.08087158203125,
0.705078125,
0.279296875,
0.705078125,
0.279296875,
0.71826171875,
0.08087158203125,
0.71826171875
],
[
0.50244140625,
0.70068359375,
0.890625,
0.70068359375,
0.890625,
0.7138671875,
0.50244140625,
0.7138671875
],
[
0.1004638671875,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.1004638671875,
0.734375
],
[
0.5048828125,
0.71728515625,
0.8896484375,
0.71728515625,
0.8896484375,
0.72802734375,
0.5048828125,
0.72802734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.73095703125,
0.6396484375,
0.73095703125,
0.6396484375,
0.744140625,
0.50244140625,
0.744140625
],
[
0.080078125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.609375,
0.755859375,
0.78515625,
0.755859375,
0.78515625,
0.76953125,
0.609375,
0.76953125
],
[
0.869140625,
0.755859375,
0.89208984375,
0.755859375,
0.89208984375,
0.76953125,
0.869140625,
0.76953125
],
[
0.080078125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
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,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50244140625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50244140625,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82763671875,
0.451904296875,
0.82763671875,
0.451904296875,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.5009765625,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5009765625,
0.8388671875
],
[
0.098876953125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.8447265625,
0.85693359375,
0.8447265625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.521484375,
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
]
] |
[
"ncodinc",
"StyleMap",
"DySPADE",
"pth Ma",
"eature",
"Glabal style)",
"AUOS",
"(2D",
"",
"Reaennstruiete",
"DepthMap",
"NOS",
"ALOO",
"0",
"0",
"",
"HIF",
"Aaaptrve",
"Modulation",
"0",
"0",
"2",
"N",
"AUDO",
"?",
"F",
"A",
"Dept",
"0",
"",
"Predicto",
"Generator",
"Svnthesizeo",
"Input Photo",
"@Concatenation",
"Sketc",
"I0",
"&Element-wisePr",
"@Element-wiseAddition",
"(a)",
"metaACON-IN",
"(b) ",
"Figure 2: Pipeline of the proposed Human-Inspired Dynamic Adaptation (HIDA) method for facial sketch synthesis. (a) The",
"overall generatorarchitecture,(b)an decodinglayerwithDySPADE, InfoACON,andDOG",
"CON is formulated as:",
"map D from features representing the sketch. In this way,",
"the generated sketch Y, would convey the 3D geometry of",
"y = (p1 =p2) :g(0(p1 =p2)c) +p2r,",
"(1)",
"X. Besides, we boost the capacity of generator by using",
"a dynamic normalization module and a dynamic activation",
"where = o(), g is a Sigmoid function, pi, and p2 are",
"function. Finally, to formulate the abstraction methodology",
"learnable parameters. In this paper, we use metaACON, in-",
"of human artists in drawing sketchy outlines, we propose",
"stead of ReLU or Leaky ReLU, in part of our networks.",
"using deformable convolutions to align features at coarse",
"Besides, we propose to learn spatially-adaptive parameter ",
"scales. Details will be introduced bellow.",
"according to the 3D geometry, 2D appearance, and global",
"style control. Our activation function proves boosting the",
"3.1. Informative and Dynamic Adaptation (IDA)",
"performance and allowing precise style control.",
"To simulate the drawing methodology of human artist,",
"wefirstproposeanovelInformativeandDynamicAdap-",
"3. The Proposed",
"tation (IDA) module, to modulate deep features based on",
"We aim to translate a facial photo X to a sketch Ys, in",
"a combination of the facial depth map D, the style map",
"S, and the appearance representations A, i.e. {D, S, A].",
"styles,drawnbyan artist.Heres =1,2.....Sis astylela-",
"bel, S is the total number of styles. In this work, we seek to",
"Specially, we propose a novel dynamic activation function,",
"guide the synthesis of sketch portraits by using comprehen-",
"termed Informative ACON (InfoACON), and a dynamic",
"sive facial information, including both 3D geometry and 2D",
"normalization module, termed DySPADE",
"appearance, as well as the global style control. Given a fa-",
"Informative ACON (InfoACON). The original metaA-",
"CON function automatically allows whether a neuron is ac-",
"cial photo,we use thecorresponding depthmapD to repre-",
"tive or not, based on its value, as previously presented in",
"sentits 3D geometry.Afterwards,wecombinethemwith a",
"global style map S to decode a facial sketch. In this way, our",
"Eq. 1. During the drawing process, a human artist typically",
"goal is formulated as learning a mapping from [X, D, S to",
"decides whether to draw a stroke or not based on the 3D",
"Ys, i.e. G : {X,D, S) Y..",
"geometry, 2D appearance, and style type. Inspired by this",
"To supervise our model, it is necessary to obtain depth",
"observation,wepropose tolearn the parameterinEq.1",
"maps for input facial photos. However, it is usually impossi-",
"from (D, S, A),i.e.",
"ble to obtain ground truth depth information in practical ap-",
"0 = α(Φe(Cat(D, S, A)),",
"(2)",
"plications. Therefore we use state-of-the-art (SOTA) depth",
"prediction methods to estimate the depth map of an input",
"where @e is a two-layer Convolutional network (Fig. 3).",
"facial photo. In practice, we use 3DDFA [13] as the depth",
"We refer to the modified metaACON function as Informa-",
"predictor, because it has been widely used and shown excel-",
"tive ACON (InfoACON). In our networks, we apply this In-",
"lentperformanceinvarious3Dfacereconstructiontasks",
"foACON function in all the decoding layers. In this way,",
"The overall pipeline of our model is as shown in Fig. 2.",
"thedecoderwouldpixel-wiselydecideswhethertodepicta",
"It contains an off-line facial depth predictor P, a genera-",
"stroke, or the type of a stroke, in a generated sketch.",
"tor G, and a patch-wise discriminator D. In addition to a",
"Dynamic Normalization (DySPADE). Following [27],",
"facial sketch, we enforce G to reconstruct the input depth",
"we additionally transform neuron activations by shifting"
] |
{
"arxiv_id": "2309.00216",
"language": "en",
"page_idx": 2
}
|
||
2309.00216-en-3
|
2309.00216-en-3.jpg
|
the mean values and scaling the standard deviations, in the instance-wise and channel-wise manner [33]. Different from the original SPADE, we use dynamic activation here to introduce more flexibility on the learned modulating parameters. Let $\mathbf{F}\in\mathbb{R}^{C\times H\times W}$ denote the input features of the current DySPADE module. $H,W$, and $C$ are the height, width and the number of channels. The activation value at site $(c,h,w)$ is modulated as:
$$\tilde f_{c,h,w}=\gamma_{c,h,w}(\mathbf{D},\mathbf{S},\mathbf{A})\frac{f_{c,h,w}-\mu_{c}}{\sigma_{c}}+\beta_{c,h,w}(\mathbf{D},\mathbf{S},\mathbf{A}),$$
(3)
where $f_{c,h,w}$ and $\tilde{f}_{c,h,w}$ are the input and modulated ac-
tivation at site $(c,h,w)$, respectively. $\mu_c$ and $\sigma_c$ are the mean and standard deviation of $f_{c,h,w}$ in the $c$-th channel. $\gamma_{c,h,w}(\mathbf{D},\mathbf{S},\mathbf{A})$ and $\beta_{c,h,w}(\mathbf{D},\mathbf{S},\mathbf{A})$ are learned scale and bias parameters at site $(c,h,w).$
As shown in Fig. 2, we use a two-layer and threebranched Convolutional network to predict the parameter $\theta$ in InfoACON, and the modulating parameters $\gamma$ and $\beta$ in DySPADE. To improve the flexibility of the adaptation block, we use metaACON (Eq. 1)[25] instead of ReLU, after the first Convolutional layer. In this way, the modulating factors would pixel-wisely adapt to an integration of the facial 3D geometry, 2D appearance, and global style.
In IDA, the activation at each position is modulated according to a joint consideration of local facial 3D geometry, appearance, and artistic style. This mechanism is consistent with the drawing methodology of human artists. To execute a facial sketch, an artist usually uses diverse textures to represent 3D geometry or illustration variations. Besides, the style of all pencil strokes are consistent inside a single sketch. As a result, IDA is promising to produce realistic sketchy textures in globally consistent style.
$3. 2. \textbf{Deformable Outline Generation ( DOG) }$
Human artists usually draw abstract lines to capture facial geometric structures, such as the boundaries of facial organs, and facial mood. To this end, the resulting outlines typically convey such structures abstractly, instead of pixel-wisely tracing them. In other words, there are geometric deformations between the input photo and the sketches drawn by artists. To simulate such an abstraction drawing methodology, we propose to align decoding features at coarse scales. In this way, the generated sketches would present abstract and distinct outlines, instead of scattered outlines with a lot of subtle variations.
In practice, we use deformable convolution (DCN)[7] instead of standard Transposed Convolution over the first and second decoding layers. As will be presented in the ablation study (Section 4.5), this deformable outline generation (DOG) module significantly boosts the clarity of generated outlines. Besides, DOG enables the network produce
abstract sketches (e.g. Stylel in the FS2K dataset), which
contains a sparse set of sketchy line drawings.
## 3.3. Overall Generator Architecture
Our generator follows the U-Net architecture[18] in whole. In the encoder, the facial photo X and the depth map D are first fed into a Convolutional layer, separately. Afterwards, the corresponding feature maps are concatenated and fed into the following encoding layers. Each encoding layer follows a Conv-metaReLU-IN architecture, and down-samples the size of feature maps by 1/2. The encoding features are adopted as appearance representations, A.
In the decoder, we expand an DySPADE block to every decoding layer, except the last one. Fig. 2 illustrates the pipeline of a decoding layer with DySPADE. Over the $l$- th decoding layer, let D$^l$ be the corresponding depth map, $\mathbf{S}^l$ the style map, and $\mathbf{A}^l$ the appearance features. We down-sample the original depth map D to $\mathbf{D}^l$ by building a Gaussian Pyramid, and expand the one-hot style vector s to $\mathbf{S}^{l}.$ Besides, we obtain $\mathbf{A}^{l}$ by upsampling $\mathbf{E}^{l-1}$ through a Transposed-Convlotional (TrConv) layer, followed by a metaACON activation layer. We finally apply the residual connection to obtain the output of the $l$-th decoding layer:
$$\mathbf{K}^{l}=\mathbf{H}^{l}\oplus\tilde{\mathbf{F}}^{l},\:\mathrm{with}\:\tilde{\mathbf{F}}^{l}=\mathrm{DySPADE}(\mathbf{F}^{l}),$$
, (4)
where $\oplus$ denotes element-wise addition. H$^l$ is the initial upsampled feature map, output by a DOG layer (over the $1^{\hat{s}t}$ and $2^nd$ decoding layer) or a TrConv layer (over the rest layers). $\mathbf{K}^l$ is fed into subsequent layers for generating final predictions.
## 3.4. Loss Functions
To train our model, we use the following loss functions. Geometric loss. First, we use a geometric constraint to
supervise depth reconstructions from features of sketches. The geometric loss is the L2 distance between the input depth map and the reconstructed one:
(5)
$$\mathcal{L}_{geo}=\|\hat{\mathbf{D}}-\mathbf{D}\|_2^2.$$
Textural loss. The synthesized sketch $\hat{\mathbf{Y}}_s$ should present similar textures as that drawn by an artist $\mathbf{Y}_s.$ In this work, we constrain $\hat{\mathbf{Y}}_s$ and $\mathbf{Y}_s$ to have similar pixel-wise adjacent correlations [21]. To this end, we calculate their gradients by using the Sobel operator, and calculate the average Cosine distance between them. Let g$_i,j=[g_{i,j}^x,g_{i,j}^y]^T$ denote the $x$-directional and $y$-directional gradients of $\mathbf{Y}_{s}$ at site $(i,j);$and f$_i,j=[f_{i,j}^x,f_{i,j}^y]^T$ the corresponding gradients in $\hat{\mathbf{Y}}_s.$ The textural loss is formulated as:
(6)
$$\mathcal{L}_{tex}=\frac{1}{MN}\sum_{i,j}\frac{\mathbf{g}_{i,j}^{T}\mathbf{f}_{i,j}}{\|\mathbf{g}_{i,j}\|\cdot\|\mathbf{f}_{i,j}\|},$$
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.091552734375,
0.89208984375,
0.09027099609375,
0.89208984375,
0.10479736328125,
0.5009765625,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.8056640625,
0.10797119140625,
0.8056640625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.07843017578125,
0.1363525390625,
0.468017578125,
0.136962890625,
0.468017578125,
0.1514892578125,
0.07843017578125,
0.15087890625
],
[
0.50244140625,
0.1307373046875,
0.779296875,
0.1319580078125,
0.779296875,
0.146484375,
0.50244140625,
0.1458740234375
],
[
0.0767822265625,
0.147705078125,
0.47216796875,
0.147705078125,
0.47216796875,
0.1685791015625,
0.0767822265625,
0.1685791015625
],
[
0.51953125,
0.1546630859375,
0.89208984375,
0.1558837890625,
0.89208984375,
0.17041015625,
0.51953125,
0.169189453125
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.50244140625,
0.17041015625,
0.890625,
0.1717529296875,
0.890625,
0.186279296875,
0.50244140625,
0.1849365234375
],
[
0.08087158203125,
0.184326171875,
0.468017578125,
0.184326171875,
0.468017578125,
0.1944580078125,
0.08087158203125,
0.1944580078125
],
[
0.50341796875,
0.1875,
0.8896484375,
0.1875,
0.8896484375,
0.2008056640625,
0.50341796875,
0.2008056640625
],
[
0.080078125,
0.1988525390625,
0.275390625,
0.1988525390625,
0.275390625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50341796875,
0.2020263671875,
0.89208984375,
0.2020263671875,
0.89208984375,
0.21533203125,
0.50341796875,
0.21533203125
],
[
0.2459716796875,
0.222900390625,
0.46728515625,
0.231689453125,
0.465576171875,
0.255615234375,
0.2442626953125,
0.246826171875
],
[
0.50341796875,
0.2171630859375,
0.8896484375,
0.2171630859375,
0.8896484375,
0.2310791015625,
0.50341796875,
0.2310791015625
],
[
0.08416748046875,
0.2342529296875,
0.253173828125,
0.2310791015625,
0.253173828125,
0.248779296875,
0.08416748046875,
0.251953125
],
[
0.50244140625,
0.231689453125,
0.89208984375,
0.231689453125,
0.89208984375,
0.24560546875,
0.50244140625,
0.24560546875
],
[
0.283447265625,
0.2474365234375,
0.301513671875,
0.2474365234375,
0.301513671875,
0.2568359375,
0.283447265625,
0.2568359375
],
[
0.50244140625,
0.2474365234375,
0.890625,
0.2474365234375,
0.890625,
0.261474609375,
0.50244140625,
0.261474609375
],
[
0.44677734375,
0.255615234375,
0.468017578125,
0.255615234375,
0.468017578125,
0.268310546875,
0.44677734375,
0.268310546875
],
[
0.5009765625,
0.261962890625,
0.8857421875,
0.261474609375,
0.8857421875,
0.27587890625,
0.5009765625,
0.276611328125
],
[
0.08087158203125,
0.26953125,
0.46728515625,
0.26953125,
0.46728515625,
0.2841796875,
0.08087158203125,
0.2841796875
],
[
0.521484375,
0.276611328125,
0.890625,
0.276611328125,
0.890625,
0.291748046875,
0.521484375,
0.291748046875
],
[
0.07843017578125,
0.2841796875,
0.468017578125,
0.285400390625,
0.468017578125,
0.2998046875,
0.07843017578125,
0.298583984375
],
[
0.50244140625,
0.292236328125,
0.89208984375,
0.292236328125,
0.89208984375,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.080078125,
0.300537109375,
0.46728515625,
0.300537109375,
0.46728515625,
0.31494140625,
0.080078125,
0.31494140625
],
[
0.5009765625,
0.306884765625,
0.89208984375,
0.30615234375,
0.89208984375,
0.32080078125,
0.5009765625,
0.3212890625
],
[
0.080078125,
0.31640625,
0.468017578125,
0.314453125,
0.468017578125,
0.328857421875,
0.080078125,
0.330810546875
],
[
0.5009765625,
0.3212890625,
0.89208984375,
0.322509765625,
0.89208984375,
0.337158203125,
0.5009765625,
0.3359375
],
[
0.080078125,
0.330810546875,
0.289306640625,
0.330810546875,
0.289306640625,
0.3447265625,
0.080078125,
0.3447265625
],
[
0.5,
0.335205078125,
0.8935546875,
0.3359375,
0.8935546875,
0.354248046875,
0.5,
0.352783203125
],
[
0.098876953125,
0.347900390625,
0.468017578125,
0.347900390625,
0.468017578125,
0.361083984375,
0.098876953125,
0.361083984375
],
[
0.5009765625,
0.349853515625,
0.8935546875,
0.35107421875,
0.8935546875,
0.36865234375,
0.5009765625,
0.367431640625
],
[
0.07843017578125,
0.361083984375,
0.468017578125,
0.3623046875,
0.468017578125,
0.376953125,
0.07843017578125,
0.375732421875
],
[
0.50341796875,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.08087158203125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50244140625,
0.38330078125,
0.890625,
0.38330078125,
0.890625,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.080078125,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.50244140625,
0.3984375,
0.89208984375,
0.3984375,
0.89208984375,
0.412353515625,
0.50244140625,
0.412353515625
],
[
0.080078125,
0.40771484375,
0.46728515625,
0.40771484375,
0.46728515625,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.5009765625,
0.412353515625,
0.89208984375,
0.41357421875,
0.89208984375,
0.427978515625,
0.5009765625,
0.4267578125
],
[
0.080078125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.5009765625,
0.427978515625,
0.8837890625,
0.4287109375,
0.8837890625,
0.443115234375,
0.5009765625,
0.442626953125
],
[
0.080078125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.080078125,
0.45263671875,
0.42724609375,
0.45263671875,
0.42724609375,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.5498046875,
0.45263671875,
0.84326171875,
0.455078125,
0.84326171875,
0.4697265625,
0.5498046875,
0.46728515625
],
[
0.87109375,
0.45703125,
0.890625,
0.45703125,
0.890625,
0.4697265625,
0.87109375,
0.4697265625
],
[
0.1004638671875,
0.47021484375,
0.465576171875,
0.47021484375,
0.465576171875,
0.48095703125,
0.1004638671875,
0.48095703125
],
[
0.07843017578125,
0.48291015625,
0.468017578125,
0.484130859375,
0.468017578125,
0.498779296875,
0.07843017578125,
0.49755859375
],
[
0.50341796875,
0.484130859375,
0.890625,
0.484130859375,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.07843017578125,
0.5,
0.468017578125,
0.498779296875,
0.468017578125,
0.51318359375,
0.07843017578125,
0.5146484375
],
[
0.50341796875,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.5146484375,
0.50341796875,
0.5146484375
],
[
0.08087158203125,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.50341796875,
0.51513671875,
0.89208984375,
0.51513671875,
0.89208984375,
0.52880859375,
0.50341796875,
0.52880859375
],
[
0.080078125,
0.52978515625,
0.471435546875,
0.52978515625,
0.471435546875,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.50244140625,
0.52978515625,
0.89208984375,
0.52978515625,
0.89208984375,
0.54443359375,
0.50244140625,
0.54443359375
],
[
0.080078125,
0.54541015625,
0.471435546875,
0.54541015625,
0.471435546875,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50341796875,
0.54736328125,
0.5810546875,
0.54736328125,
0.5810546875,
0.5595703125,
0.50341796875,
0.5595703125
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5693359375,
0.65283203125,
0.57080078125,
0.65283203125,
0.58544921875,
0.50244140625,
0.583984375
],
[
0.08087158203125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.080078125,
0.59130859375,
0.370849609375,
0.59130859375,
0.370849609375,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.52294921875,
0.5947265625,
0.88818359375,
0.5947265625,
0.88818359375,
0.60791015625,
0.52294921875,
0.60791015625
],
[
0.52294921875,
0.60986328125,
0.890625,
0.60986328125,
0.890625,
0.62353515625,
0.52294921875,
0.62353515625
],
[
0.080078125,
0.61865234375,
0.414306640625,
0.61865234375,
0.414306640625,
0.6318359375,
0.080078125,
0.6318359375
],
[
0.50341796875,
0.626953125,
0.8896484375,
0.626953125,
0.8896484375,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.098876953125,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.098876953125,
0.6572265625
],
[
0.50341796875,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50341796875,
0.65380859375
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.5009765625,
0.65478515625,
0.75244140625,
0.65478515625,
0.75244140625,
0.66845703125,
0.5009765625,
0.66845703125
],
[
0.080078125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.080078125,
0.6875
],
[
0.63232421875,
0.68310546875,
0.76123046875,
0.68310546875,
0.76123046875,
0.69775390625,
0.63232421875,
0.69775390625
],
[
0.080078125,
0.689453125,
0.471435546875,
0.689453125,
0.471435546875,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.869140625,
0.6845703125,
0.8935546875,
0.6845703125,
0.8935546875,
0.69775390625,
0.869140625,
0.69775390625
],
[
0.080078125,
0.7041015625,
0.4697265625,
0.7041015625,
0.4697265625,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.52294921875,
0.71337890625,
0.89208984375,
0.71337890625,
0.89208984375,
0.72802734375,
0.52294921875,
0.72802734375
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.7197265625,
0.4697265625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.72900390625,
0.890625,
0.72900390625,
0.890625,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.08087158203125,
0.736328125,
0.465576171875,
0.736328125,
0.465576171875,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.080078125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.50341796875,
0.744140625,
0.890625,
0.744140625,
0.890625,
0.75830078125,
0.50341796875,
0.75830078125
],
[
0.080078125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.50244140625,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.50244140625,
0.7734375
],
[
0.07843017578125,
0.77978515625,
0.4697265625,
0.77880859375,
0.4697265625,
0.79248046875,
0.07843017578125,
0.79345703125
],
[
0.50244140625,
0.7744140625,
0.89208984375,
0.7744140625,
0.89208984375,
0.78857421875,
0.50244140625,
0.78857421875
],
[
0.5009765625,
0.78857421875,
0.82421875,
0.7890625,
0.82421875,
0.8037109375,
0.5009765625,
0.80322265625
],
[
0.826171875,
0.79052734375,
0.890625,
0.79052734375,
0.890625,
0.8017578125,
0.826171875,
0.8017578125
],
[
0.50244140625,
0.8046875,
0.89208984375,
0.8046875,
0.89208984375,
0.818359375,
0.50244140625,
0.818359375
],
[
0.080078125,
0.79443359375,
0.334228515625,
0.79443359375,
0.334228515625,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.098876953125,
0.81005859375,
0.46728515625,
0.81005859375,
0.46728515625,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.50341796875,
0.818359375,
0.8935546875,
0.818359375,
0.8935546875,
0.8359375,
0.50341796875,
0.8359375
],
[
0.080078125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50341796875,
0.83544921875,
0.75634765625,
0.8369140625,
0.75634765625,
0.8515625,
0.50341796875,
0.849609375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.65283203125,
0.8662109375,
0.69384765625,
0.869140625,
0.69287109375,
0.8798828125,
0.65185546875,
0.876953125
],
[
0.7255859375,
0.86279296875,
0.77783203125,
0.86279296875,
0.77783203125,
0.8818359375,
0.7255859375,
0.8818359375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5888671875,
0.87353515625,
0.6357421875,
0.87548828125,
0.6337890625,
0.89013671875,
0.58837890625,
0.88818359375
],
[
0.86767578125,
0.87353515625,
0.89453125,
0.87353515625,
0.89453125,
0.89013671875,
0.86767578125,
0.89013671875
],
[
0.63818359375,
0.88330078125,
0.685546875,
0.88330078125,
0.685546875,
0.89453125,
0.63818359375,
0.89453125
],
[
0.69775390625,
0.8837890625,
0.80224609375,
0.880859375,
0.8037109375,
0.8984375,
0.69921875,
0.9013671875
],
[
0.68115234375,
0.89306640625,
0.69775390625,
0.89306640625,
0.69775390625,
0.90087890625,
0.68115234375,
0.90087890625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
]
] |
[
"the mean values and scaling the standard deviations, in",
"abstract sketches (e.g. Stylel in the FS2K dataset), which",
"the instance-wise and channel-wisemanner[33l.Different",
"contains a sparse set of sketchy line drawings.",
"from the original SPADE, we use dynamic activation here",
"to introduce more flexibility on the learned modulating pa-",
"3.3.Overall Generator Architecture",
"rameters. Let F e RCxHxW denote the input features of",
"Our generator follows the U-Net architecture [18] in",
"the current DySPADE module. H, W, and C are the height,",
"whole. In the encoder, the facial photo X and the depth map",
"width and thenumber of channels.The activation value at",
"D are first fed into a Convolutional layer, separately. Af-",
"site (c, h, w) is modulated as:",
"terwards, the corresponding feature maps are concatenated",
" feh,w Me + βe,h,u(D, S, A),",
"and fed into the following encoding layers. Each encod-",
"fe,h,w =e,h,w(D, S, A)",
"ing layer follows a Conv-metaReLU-IN architecture, and",
"ac",
"down-samples the size of feature maps by 1/2. The encod-",
"(3)",
"ing features are adopted as appearance representations, A.",
"where fe.h.w and fe.h.w are the input and modulated ac-",
"In the decoder, we expand an DySPADE block to every",
"tivation at site (c, h, w), respectively. μe and oe are the",
"decoding layer, except the last one. Fig. 2 illustrates the",
"mean and standard deviation of fe.h,w in the c-th channel.",
"pipeline of a decoding layer with DySPADE. Over the l-",
"e,h,w(D, S, A) and βe,h,w(D, S, A) are learned scale and",
"th decoding layer, let D' be the corresponding depth map,",
"bias parameters at site (c, h, w).",
"S' the style map, and A' the appearance features. We",
"As shown in Fig. 2, we use a two-layer and three-",
"down-sample the original depth map D to D' by building",
"branched Convolutional network to predict the parameter",
"a Gaussian Pyramid, and expand the one-hot style vector s",
" in InfoACON, and the modulating parameters and β",
"to S'. Besides, we obtain A' by upsampling El-1 through",
"in DySPADE. To improve the flexibility of the adaptation",
"a Transposed-Convlotional (TrConv) layer, followed by a",
"block, we use metaACON (Eq. 1) [25] instead of ReLU,",
"metaACON activation layer. We finally apply the residual",
"after the first Convolutional layer. In this way, the modulat-",
"connection to obtain the output of the I-th decoding layer:",
"ing factors would pixel-wisely adapt to an integration of the",
"facial 3D geometry, 2D appearance, and global style.",
"Kl =H' @Fl, with Fl =DySPADE(F'),",
"(4)",
"InIDA,theactivationateachpositionismodulated ac-",
"cording to a joint consideration of local facial 3D geometry,",
"where denotes element-wise addition. H' is the initial",
"appearance, and artistic style. This mechanism is consistent",
"upsampled feature map, output by a DOG layer (over the",
"withthedrawingmethodologyofhuman artists.To exe-",
"1st and 2nd decoding layer) or a TrConv layer (over the rest",
"cute a facial sketch, an artist usually uses diverse textures to",
"layers).K' is fed into subsequent layers for generating final",
"represent 3D geometry or illustration variations. Besides,",
"predictions.",
"the style of all pencil strokes are consistent inside a single",
"3.4. Loss Functions",
"sketch.Asaresult.IDA ispromising toproducerealistic",
"sketchy textures in globally consistent style.",
"To train our model, we use the following loss functions.",
"Geometric loss. First, we use a geometric constraint to",
"3.2.Deformable Outline Generation (DOG)",
"supervisedepthreconstructions fromfeatures of sketches.",
"Human artists usually draw abstract lines to capture fa-",
"The geometric loss is the L2 distance between the input",
"cial geometric structures, such as the boundaries of facial",
"depth map and the reconstructed one:",
"organs, and facial mood. To this end, the resulting out-",
"Ceo = D-D2.",
"lines typically convey such structures abstractly, instead of",
"(5)",
"pixel-wisely tracing them. In other words, there are geomet-",
"Textural loss. The synthesized sketch Y, should present",
"ric deformations between the input photo and the sketches",
"similar textures as that drawn by an artist Yg. In this work,",
"drawn by artists.To simulate such an abstraction draw-",
"ing methodology, we propose to align decoding features at",
"we constrain Y,and Y,to have similar pixel-wise adjacent",
"coarse scales. In this way, the generated sketches would",
"correlations [21]. To this end, we calculate their gradients",
"present abstract and distinct outlines, instead of scattered",
"by using the Sobel operator, and calculate the average Co-",
"sine distance between them. Let gi.j=[g,,g",
"jTdenote",
"the &-directional and y-directional gradients of Y, at site",
"outlines with a lot of subtle variations.",
"In practice, we use deformable convolution (DCN) [7]",
"(i,j); and fi,j =[fj, fu,]T the corresponding gradients in",
"instead of standard Transposed Convolution over the first",
"Y.. The textural loss is formulated as:",
"and second decoding layers.As will be presented in the",
"ablation study (Section 4.5), this deformable outline gener-",
"1",
"gT,fi.j",
"ation (DOG) module significantly boosts the clarity of gen-",
"Cter",
"(6)",
"MN",
"giill·fill",
"",
"erated outlines. Besides, DOG enables the network produce"
] |
{
"arxiv_id": "2309.00216",
"language": "en",
"page_idx": 3
}
|
||
2309.00216-en-4
|
2309.00216-en-4.jpg
|
where $\|\cdot\|$ denotes the magnitude of a vector, $M$ and $N$ are
the width and height of the sketch.
Pixel loss. In addition, we use the pixel-wise reconstruction loss between the synthesized sketch $\hat{\mathbf{Y}}_s$ and the target sketch $\mathbf{Y}_s$,i.e.
$$\mathcal{L}_{pix}=\|\hat{\mathbf{Y}}_{s}-\mathbf{Y}_{s}\|_{1}.$$
(7)
Adversarial loss. Finally, we use adversarial loss to measure whether a pair of depth map and synthesized sketch is real or fake. Here, we use the Cross Entropy loss, i.e.
$$\mathcal{L}_{adv}=-\log D(\mathbf{D},\mathbf{Y}_{s})-\log(1-D(\hat{\mathbf{D}},\hat{\mathbf{Y}}_{s})).$$
(8)
Full objective. We use a combination of all the afore-
mentioned losses as our full objective:
$$\mathcal{L}_{all}=\mathcal{L}_{adv}+\lambda_{1}\mathcal{L}_{pix}+\lambda_{2}\mathcal{L}_{tex}+\lambda_{3}\mathcal{L}_{geo},$$
(9)
where $\lambda_1,\lambda_2$, and $\lambda_3$ are weighting factors. We train the generator $G$ and the discriminator $D$ in an alternative manner, to minimize $\mathcal{L}_all.$
## $4. \textbf{Experiments}$
We present a thorough experimental comparison on the challenging FS2K dataset [8]. Besides, we conduct a series of ablation study to analyse impacts of the proposed DySPADE, InfoACON, and DOG modules.
## 4.1. Experimental Settings
Data. We conduct experiments on the challenging FS2K dataset. The FS2K dataset is the largest publicly released FSS dataset, consisting of 2,104 photo-sketch pairs from a wide range of image backgrounds, skin colors, sketch styles, and lighting conditions. These sketches are mainly in three styles. Following standard settings[8],we have 1,058 photo-sketch pairs for training, and 1,046 pairs for testing. For each style, we have 357/351/350 training pairs, and 619/381/46 testing pairs, from Style1 to Style3, respectively. All the images are aligned and resized to $250\times250.$ In the inference stage, we use the same style of sketch as the ground truth in default. In addition, we collect a number of challenging Faces in-the-wild from FFHQ[20] and Web. We align and resize these images in the same way as those in the FS2K dataset.
Comparison Methods In this section, we compare our method with various state-of-the-art (SOTA) ones, including FSGAN[8], GENRE[21], SCA-GAN[37], and MDAL [44]. Besides, we compare with several advanced GANs, including CycleGAN[46], Pix2Pix[18], and Pix2PixHD [35]. We use results and codes of these methods released by the corresponding authors[8]. All these methods and ours follow the same experimental settings.
Criteria In this work, we choose four performance indices as the criteria, i.e. the Fréchet Inception distance (FID) [15], Learned Perceptual Image Patch Similarity (LPIPS) metric [42], Structure Co-Occurrence Texture (SCOOT) metric [10], and Feature Similarity Measure (FSIM)[39]. Lower values of FID and LPIPS indicate higher realism of synthesized sketches. In contrast, greater values of SCOOT and FSIM generally indicate higher similarity between a synthesized sketch and the corresponding sketch drawn by an artist. We here report the average LPIPS, SCOOT, and FSIM values across all the test samples, respectively. In the following sections, $\downarrow$ indicates that lower value is better, while $\uparrow$ higher is better.
Implementation Details We implemented our model in PyTorch. All experiments are performed on a computer with a Titan 3090 GPU. We use a batch size of 4, a learning rate of $1e-4.$ We use the Adam Optimizer, and train the model for 800 epochs on the training set. Our code will be released after peer review.
$4. 2. \textbf{ Qualitative Comparison with SOTAs}$
We further qualitatively compare with SOTA FSS methods. Fig. 3 illustrates synthesized sketches on the FS2K dataset. Although our method can generate multiple styles of sketches, here we only show the synthesized sketch in the same style as the ground truth. For the face in constrained condition (the frst row), most methods successfully generate a quality sketch. For the face with extreme lighting condition (the second row) or pose variation (the third row), most synthesized sketches present unpleasant geometric deformations and fail to precisely reproduce the style. Although sketches generated by CycleGAN seems acceptable, the textures aren't like pencil-drawings. The sketches generated by FSGAN show the same styles as the ground truths, since FSGAN contains a style control module. However, these sketches show unpleasant structural distortions. This might be caused by the geometric deformations between facial photos and free-hand sketches drawn by artists, in the training data. GENRE successfully produces quality sketches, but they are all almost in the same style, since no style information is considered in GENRE
In contrast, our HIDA generates high-quality sketches in all three styles. Specially, our synthesized sketches preserve the geometries of input faces. This implies that, HIDA doesn't overfit to the training samples and combats geometric deformations. We achieve such success mainly due to the informatively adaptive normalization module, i.e. DySPADE, and the constraint of reconstructing the input depth. Besides, our synthesized sketches present the same style of strokes as the corresponding ground truths. The drawing textures are consistent inside each sketch. The style consistency demonstrates the effectiveness of our global style control mechanism through DySPADE and InfoACON. Based
|
[
[
0.080078125,
0.0921630859375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.52294921875,
0.09027099609375,
0.89208984375,
0.09027099609375,
0.89208984375,
0.10479736328125,
0.52294921875,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.306396484375,
0.10919189453125,
0.306396484375,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.09722900390625,
0.12249755859375,
0.4697265625,
0.121826171875,
0.4697265625,
0.1363525390625,
0.09722900390625,
0.136962890625
],
[
0.50244140625,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.471435546875,
0.1383056640625,
0.471435546875,
0.15283203125,
0.080078125,
0.15283203125
],
[
0.5009765625,
0.1363525390625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.07843017578125,
0.15087890625,
0.178955078125,
0.15283203125,
0.178955078125,
0.1673583984375,
0.07843017578125,
0.1654052734375
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.50341796875,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.200927734375,
0.179931640625,
0.348876953125,
0.179931640625,
0.348876953125,
0.1976318359375,
0.200927734375,
0.1976318359375
],
[
0.44677734375,
0.1817626953125,
0.4697265625,
0.1817626953125,
0.4697265625,
0.1956787109375,
0.44677734375,
0.1956787109375
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.50341796875,
0.1976318359375,
0.890625,
0.1976318359375,
0.890625,
0.2108154296875,
0.50341796875,
0.2108154296875
],
[
0.1004638671875,
0.2108154296875,
0.468017578125,
0.2108154296875,
0.468017578125,
0.2247314453125,
0.1004638671875,
0.2247314453125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2266845703125,
0.468017578125,
0.2266845703125,
0.468017578125,
0.240478515625,
0.080078125,
0.240478515625
],
[
0.50244140625,
0.2266845703125,
0.89208984375,
0.2279052734375,
0.89208984375,
0.242431640625,
0.50244140625,
0.2412109375
],
[
0.07843017578125,
0.240478515625,
0.44287109375,
0.2412109375,
0.44287109375,
0.255615234375,
0.07843017578125,
0.255126953125
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.50341796875,
0.2587890625,
0.8935546875,
0.2587890625,
0.8935546875,
0.272705078125,
0.50341796875,
0.272705078125
],
[
0.10296630859375,
0.268310546875,
0.431396484375,
0.267578125,
0.431396484375,
0.285400390625,
0.10296630859375,
0.285888671875
],
[
0.443603515625,
0.271484375,
0.46728515625,
0.271484375,
0.46728515625,
0.28271484375,
0.443603515625,
0.28271484375
],
[
0.50244140625,
0.2734375,
0.7998046875,
0.2734375,
0.7998046875,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.521484375,
0.290283203125,
0.89208984375,
0.290283203125,
0.89208984375,
0.3037109375,
0.521484375,
0.3037109375
],
[
0.09722900390625,
0.29736328125,
0.468017578125,
0.298583984375,
0.468017578125,
0.313232421875,
0.09722900390625,
0.3125
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.07843017578125,
0.31494140625,
0.3349609375,
0.31494140625,
0.3349609375,
0.328857421875,
0.07843017578125,
0.328857421875
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333984375,
0.50244140625,
0.333251953125
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.1282958984375,
0.34228515625,
0.428955078125,
0.3447265625,
0.428955078125,
0.359130859375,
0.1282958984375,
0.356689453125
],
[
0.435546875,
0.3447265625,
0.468017578125,
0.3447265625,
0.468017578125,
0.35595703125,
0.435546875,
0.35595703125
],
[
0.5009765625,
0.348388671875,
0.89208984375,
0.349853515625,
0.89208984375,
0.3642578125,
0.5009765625,
0.363037109375
],
[
0.080078125,
0.37060546875,
0.4697265625,
0.371826171875,
0.4697265625,
0.386474609375,
0.080078125,
0.385009765625
],
[
0.50244140625,
0.365478515625,
0.67724609375,
0.365478515625,
0.67724609375,
0.37890625,
0.50244140625,
0.37890625
],
[
0.080078125,
0.386962890625,
0.4697265625,
0.386962890625,
0.4697265625,
0.40087890625,
0.080078125,
0.40087890625
],
[
0.50341796875,
0.392578125,
0.81787109375,
0.392578125,
0.81787109375,
0.4072265625,
0.50341796875,
0.4072265625
],
[
0.07843017578125,
0.40087890625,
0.226318359375,
0.401611328125,
0.226318359375,
0.416748046875,
0.07843017578125,
0.415283203125
],
[
0.52294921875,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.52294921875,
0.43115234375
],
[
0.08087158203125,
0.431884765625,
0.2091064453125,
0.431884765625,
0.2091064453125,
0.4462890625,
0.08087158203125,
0.4462890625
],
[
0.50341796875,
0.43310546875,
0.890625,
0.43310546875,
0.890625,
0.4462890625,
0.50341796875,
0.4462890625
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.098876953125,
0.458251953125,
0.4697265625,
0.458251953125,
0.4697265625,
0.472900390625,
0.098876953125,
0.472900390625
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.080078125,
0.473388671875,
0.4697265625,
0.473388671875,
0.4697265625,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.5009765625,
0.479248046875,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.5009765625,
0.492431640625
],
[
0.080078125,
0.4892578125,
0.468017578125,
0.4892578125,
0.468017578125,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.50390625,
0.340576171875,
0.50390625,
0.340576171875,
0.517578125,
0.080078125,
0.517578125
],
[
0.50244140625,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.521484375,
0.50244140625,
0.521484375
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.08087158203125,
0.52978515625,
0.28515625,
0.52978515625,
0.28515625,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.09722900390625,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.5673828125,
0.09722900390625,
0.5673828125
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.08087158203125,
0.57080078125,
0.46728515625,
0.57080078125,
0.46728515625,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50341796875,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50341796875,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.59521484375,
0.50341796875,
0.59521484375
],
[
0.07843017578125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.07843017578125,
0.61279296875
],
[
0.5009765625,
0.599609375,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.5009765625,
0.61279296875
],
[
0.080078125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50341796875,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62548828125,
0.50341796875,
0.62548828125
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.50244140625,
0.642578125
],
[
0.08087158203125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.50244140625,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.50244140625,
0.6875
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.705078125,
0.471435546875,
0.705078125,
0.471435546875,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.705078125,
0.78759765625,
0.705078125,
0.78759765625,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.08087158203125,
0.7216796875,
0.46728515625,
0.7216796875,
0.46728515625,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.521484375,
0.7197265625,
0.8935546875,
0.7197265625,
0.8935546875,
0.73291015625,
0.521484375,
0.73291015625
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.08087158203125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76220703125,
0.08087158203125,
0.76220703125
],
[
0.5009765625,
0.75048828125,
0.890625,
0.75,
0.890625,
0.7646484375,
0.5009765625,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.25244140625,
0.767578125,
0.25244140625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.098876953125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.098876953125,
0.79443359375
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.79736328125,
0.468017578125,
0.79736328125,
0.468017578125,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.50244140625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.8125,
0.468017578125,
0.8125,
0.468017578125,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.5009765625,
0.8095703125,
0.88818359375,
0.81005859375,
0.88818359375,
0.82470703125,
0.5009765625,
0.82373046875
],
[
0.080078125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.080078125,
0.8388671875
],
[
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,
0.85595703125,
0.50244140625,
0.85595703125
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.365234375,
0.88720703125,
0.365234375,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"where I - denotes the magnitude of a vector, M and N are",
"Criteria In this work, we choose four performance",
"the width and height of the sketch.",
"indices as the criteria. i.e.the Frechet Inception dis-",
"Pixel loss. In addition, we use the pixel-wise reconstruc-",
"tance (FID) [15], Learned Perceptual Image Patch Simi-",
"tion loss between the synthesized sketch Y, and the target",
"larity (LPIPS) metric [42], Structure Co-Occurrence Tex-",
"sketch Ys,i.e.",
"ture (SCOOT) metric [IO], and Feature Similarity Measure",
"(FSIM) [39]. Lower values of FID and LPIPS indicate",
"Cpiz = /Y, - Yll1.",
"(7)",
"higher realism of synthesized sketches. In contrast, greater",
"values of SCOOT and FSIM generally indicate higher sim-",
"Adversarial loss. Finally, we use adversarial loss to",
"ilarity between a synthesized sketch and the correspond-",
"measure whether a pair of depth map and synthesized sketch",
"ing sketch drawn by an artist. We here report the average",
"is real or fake. Here, we use the Cross Entropy loss, i.e.",
"LPIPS, SCOOT, and FSIM values across all the test sam-",
"ples, respectively. In the following sections, indicates that",
"Cadu = -log D(D, Ys) -log(1 -D(D, Y,)).",
"(8)",
"lower value is better, while higher is better.",
"Implementation Details We implemented our model in",
"Full objective. We use a combination of all the afore-",
"PyTorch. All experiments are performed on a computer",
"mentioned losses as our full objective:",
"with a Titan 3090 GPU. We use a batch size of 4, a learning",
"rate of le - 4. We use the Adam Optimizer, and train the",
"Call=Cadu+AiCpir+2Cter+ACgeo,",
"(9)",
"model for 8o0 epochs on the training set. Our code will be",
"where A1, A2, and Xg are weighting factors. We train the",
"released after peer review.",
"generator G and the discriminator D in an alternative man-",
"4.2. Qualitative Comparison with SOTAs",
"ner, to minimize Call.",
"We further qualitatively compare with SOTA FSS meth-",
"4. Experiments",
"ods. Fig. 3 illustrates synthesized sketches on the FS2K",
"dataset. Although our method can generate multiple styles",
"We present a thorough experimental comparison on the",
"of sketches, here we only show the synthesized sketch in the",
"challenging FS2K dataset [8]. Besides, we conduct a series",
"same style as the ground truth. For the face in constrained",
"of ablation study to analyse impacts of the proposed DyS-",
"condition (the first row), most methods successfully gen-",
"PADE, InfoACON, and DOG modules.",
"erate a quality sketch. For the face with extreme light-",
"ing condition (the second row) or pose variation (the third",
"4.1.Experimental Settings",
"row), most synthesized sketches present unpleasant geomet-",
"Data. We conduct experiments on the challenging FS2K",
"ric deformations and fail to precisely reproduce the style.",
"dataset.TheFS2K dataset is the largest publiclyreleased",
"Although sketches generated by CycleGAN seems accept-",
"FSS dataset, consisting of 2,104 photo-sketch pairs from",
"able,thetextures aren't like pencil-drawings.The sketches",
"a wide range of image backgrounds, skin colors, sketch",
"generated byFSGAN show the same styles as the ground",
"styles, and lighting conditions. These sketches are mainly",
"truths,sinceFSGANcontainsastylecontrolmodule.How-",
"in three styles. Following standard settings [8], we have",
"ever, these sketches show unpleasant structural distortions.",
"1,058 photo-sketch pairs for training, and 1,046 pairs for",
"This might be caused by the geometric deformations be-",
"testing. For each style, we have 357/351/350 training pairs,",
"tween facial photos and free-hand sketches drawn by artists,",
"and 619/381/46 testing pairs, from Style1 to Style3, respec-",
"in the training data. GENRE successfully produces quality",
"tively. All the images are aligned and resized to 250 × 250.",
"sketches, but they are all almost in the same style, since no",
"In the inference stage, we use the same style of sketch as",
"style information is considered in GENRE.",
"the ground truth in default. In addition, we collect a num-",
"In contrast, our HIDA generates high-quality sketches",
"ber of challenging Faces in-the-wild from FFHQ [20] and",
"in allthree styles.Specially.our synthesized sketches pre-",
"Web.We align and resize these images in the same way as",
"serve the geometries of input faces. This implies that, HIDA",
"those in theFS2Kdataset.",
"doesn't overfit to the training samples and combats geomet-",
"Comparison Methods In this section, we compare our",
"ric deformations.We achieve such success mainly due to",
"method with various state-of-the-art (SOTA) ones, includ-",
"the informatively adaptive normalization module, i.e. DyS-",
"ingFSGAN8]GENRE211.SCA-GAN[371,andMDAL",
"PADE, and the constraint of reconstructing the input depth",
"[44]. Besides, we compare with several advanced GANs,",
"Besides, our synthesized sketches present the same style of",
"including CycleGAN[46],Pix2Pix [18], and Pix2PixHD",
"strokes as the corresponding ground truths. The drawing",
"[35]. We use results and codes of these methods released",
"textures are consistent inside each sketch.The style consis-",
"by the corresponding authors [8]. All these methods and",
"tency demonstrates the effectiveness of our global style con-",
"ours follow the same experimental settings.",
"trol mechanism through DySPADE and InfoACON. Based"
] |
{
"arxiv_id": "2309.00216",
"language": "en",
"page_idx": 4
}
|
||
1905.04088-en-0
|
1905.04088-en-0.jpg
|
# SPLINE-Net: Sparse Photometric Stereo through Lighting Interpolation and Normal Estimation Networks
Qian Zheng$^{1\sharp*}$ Yiming Jia$^{2\sharp\dagger}$ Boxin Shi$^{3,4*}$ Xudong Jiang$^{1}$ Ling-Yu Duan$^{3,4}$ Alex C.Kot
$^{1}$School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
$^{2}$Department of Precision Instrument, Tsinghua University, Beijing, China
$^{3}$National Engineering Laboratory for Video Technology, Department of CS, Peking University, Beijing, China
$^{4}$Peng Cheng Laboratory, Shenzhen, China
$\{$zhengqian, exdjiang, eackot$\}@$ntu.edu.sg, jiaym$15@$outlook.com, $\{$shiboxin, lingyu$\}@$pku.edu.cn
## Abstract
This paper solves the Sparse Photometric stereo through Lighting Interpolation and Normal Estimation using a generative Network (SPLINE-Net). SPLINE-Net contains a lighting interpolation network to generate dense lighting
$\sum_{\begin{array}{cc}&observations&given\:a\:sparse\:set\:of\:lights\:as\:inputs\:followed\\&by\:a\:normal\:estimation\:network\:to\:estimate\:surface\:nor-\\&mals.\:Both\:networks\:are\:jointly\:constrained\:by\:the\:pro-\end{array}}$
posed symmetric and asymmetric loss functions to enforce isotropicconstrainandperformoutlierrejectionofglobal illumination effects. SPLINE-Net is verified to outperform existing methods for photometric stereo of general BRDFs by using only ten images of different lights instead of using nearly one hundred images.
### $1. \textbf{Introduction}$
The problem of photometric stereo [35] inversely solves for the radiometric image formation model to recover surface normals from different appearances of objects under various lighting conditions with a fxed camera view. The classic method[35] assumes an ideal Lambertian image formation model without global illumination effects (such as inter-reflection and shadows), which deviates from the realistic scenario and prevents photometric stereo from being able to handle real-world objects. To make photometric stereo practical, the major difficulties lie in dealing with objects of general reflectance and global illumination effects. These can be achieved by either exploring analytical Bidirectional Reflectance Distribution Function (BRDF) representations (e.g.,[14]) and general BRDF properties $(e.g.,[30])$ to model non-Lambertian interactions of lighting and surface normal or suppressing global effects by treating them as outliers $(e.g.,[37]).$ Recently, deep learning
${\text{中Authors contributed equally to this work.}}$
$^*$Corresponding authors.
$^{\dagger}$Work completed while interning at NTU and PKU.
<FigureHere>
Figure 1. An illustration of observation maps corresponding to two surface normals (a brief introduction of observation maps can be found in Section 3.2 and [18]). (a) Two surface normals and their observation maps with dense lights, (b) sparse observation maps with 10 order-agnostic lights, (c) dense observation maps generated by our SPLINE-Net given sparse observation maps in (b) as inputs, and (d) ground truth of dense observation maps with 1000 lights. We use v= ( 0, 0, 1$) ^{\top }$ to represent viewing direction and n to represent surface normal in this paper.
based approaches are introduced to solve these difficulties by implicitly learning both the image formation process and global illumination effects from training data $(e.g.,[9,18]).$
According to a comprehensive benchmark evaluation [28] (including quantitative results for representative methods published before 2016) and the additional results reported in most recent works [9, 18,40],a moderately dense lighting distribution $(e.g.$, around 100 directional lights randomly sampled from the visible hemisphere) is required to achieve reasonably good normal estimation for objects with general materials $(e.g.$, angular error around $10^{\circ}$ for a shiny plastic). This is because multi-illumination observations under a dense set of lights are required to fit the parameters in analytic BRDF models [14], to analyze general BRDF properties [30], to observe sufficient inliers and outliers [37], and to ensure the convergence of training neural networks [9]. How to achieve high accuracy in normal estimation for objects given general BRDFs with a sparse set of lights $(e.g.,10)$, which we call sparse photometric stereo in this paper, is still an open yet challenging problem [28].
|
[
[
0.236083984375,
0.1307373046875,
0.73388671875,
0.1319580078125,
0.73388671875,
0.149658203125,
0.236083984375,
0.1483154296875
],
[
0.2034912109375,
0.1572265625,
0.76953125,
0.1572265625,
0.76953125,
0.1717529296875,
0.2034912109375,
0.1717529296875
],
[
0.1070556640625,
0.186279296875,
0.86376953125,
0.186279296875,
0.86376953125,
0.20703125,
0.1070556640625,
0.20703125
],
[
0.2034912109375,
0.2083740234375,
0.76953125,
0.208984375,
0.76953125,
0.2235107421875,
0.2034912109375,
0.222900390625
],
[
0.2646484375,
0.2215576171875,
0.708984375,
0.222900390625,
0.708984375,
0.2374267578125,
0.2646484375,
0.236083984375
],
[
0.1568603515625,
0.23486328125,
0.81201171875,
0.236083984375,
0.81201171875,
0.250732421875,
0.1568603515625,
0.25
],
[
0.359375,
0.251953125,
0.615234375,
0.251953125,
0.615234375,
0.26513671875,
0.359375,
0.26513671875
],
[
0.0228729248046875,
0.26953125,
0.056365966796875,
0.26953125,
0.05963134765625,
0.376953125,
0.02532958984375,
0.376953125
],
[
0.1895751953125,
0.26513671875,
0.78125,
0.26513671875,
0.78125,
0.279052734375,
0.1895751953125,
0.279052734375
],
[
0.23779296875,
0.30615234375,
0.313720703125,
0.30615234375,
0.313720703125,
0.32080078125,
0.23779296875,
0.32080078125
],
[
0.098876953125,
0.33837890625,
0.4697265625,
0.33837890625,
0.4697265625,
0.3515625,
0.098876953125,
0.3515625
],
[
0.080078125,
0.352783203125,
0.468017578125,
0.354248046875,
0.468017578125,
0.36865234375,
0.080078125,
0.367431640625
],
[
0.02777099609375,
0.36865234375,
0.052276611328125,
0.36865234375,
0.052276611328125,
0.389404296875,
0.02777099609375,
0.389404296875
],
[
0.08087158203125,
0.36865234375,
0.4697265625,
0.36865234375,
0.4697265625,
0.3818359375,
0.08087158203125,
0.3818359375
],
[
0.08087158203125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.397705078125,
0.08087158203125,
0.397705078125
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.4443359375,
0.02532958984375,
0.4443359375
],
[
0.08087158203125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.41357421875,
0.08087158203125,
0.41357421875
],
[
0.08087158203125,
0.416748046875,
0.465576171875,
0.416748046875,
0.465576171875,
0.4267578125,
0.08087158203125,
0.4267578125
],
[
0.07843017578125,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.443115234375,
0.07843017578125,
0.443115234375
],
[
0.75244140625,
0.44140625,
0.77294921875,
0.44140625,
0.77294921875,
0.451904296875,
0.75244140625,
0.451904296875
],
[
0.8388671875,
0.44140625,
0.8544921875,
0.44140625,
0.8544921875,
0.449462890625,
0.8388671875,
0.449462890625
],
[
0.5537109375,
0.442626953125,
0.5703125,
0.442626953125,
0.5703125,
0.449462890625,
0.5537109375,
0.449462890625
],
[
0.03350830078125,
0.443115234375,
0.0555419921875,
0.443115234375,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.44580078125,
0.468017578125,
0.44580078125,
0.468017578125,
0.458984375,
0.080078125,
0.458984375
],
[
0.5009765625,
0.449462890625,
0.89208984375,
0.45068359375,
0.89208984375,
0.46533203125,
0.5009765625,
0.464111328125
],
[
0.67236328125,
0.4443359375,
0.6845703125,
0.4443359375,
0.6845703125,
0.448974609375,
0.67236328125,
0.448974609375
],
[
0.080078125,
0.460205078125,
0.4697265625,
0.460205078125,
0.4697265625,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.50341796875,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.477783203125,
0.50341796875,
0.477783203125
],
[
0.08087158203125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.08087158203125,
0.488037109375
],
[
0.50341796875,
0.48095703125,
0.890625,
0.48095703125,
0.890625,
0.4912109375,
0.50341796875,
0.4912109375
],
[
0.02532958984375,
0.4873046875,
0.058013916015625,
0.4873046875,
0.058013916015625,
0.630859375,
0.02532958984375,
0.630859375
],
[
0.080078125,
0.4892578125,
0.4697265625,
0.488037109375,
0.4697265625,
0.50244140625,
0.080078125,
0.50390625
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.07843017578125,
0.50390625,
0.468017578125,
0.5048828125,
0.468017578125,
0.51953125,
0.07843017578125,
0.51806640625
],
[
0.50244140625,
0.5068359375,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.52099609375,
0.26318359375,
0.52099609375,
0.26318359375,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50341796875,
0.52099609375,
0.890625,
0.521484375,
0.890625,
0.53271484375,
0.50341796875,
0.5322265625
],
[
0.50341796875,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.50244140625,
0.54736328125,
0.61279296875,
0.54736328125,
0.61279296875,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.62109375,
0.548828125,
0.89208984375,
0.548828125,
0.89208984375,
0.5595703125,
0.62109375,
0.5595703125
],
[
0.08251953125,
0.56640625,
0.20751953125,
0.56640625,
0.20751953125,
0.58056640625,
0.08251953125,
0.58056640625
],
[
0.50244140625,
0.56298828125,
0.75634765625,
0.56298828125,
0.75634765625,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.098876953125,
0.59228515625,
0.4697265625,
0.59228515625,
0.4697265625,
0.60693359375,
0.098876953125,
0.60693359375
],
[
0.50244140625,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.08087158203125,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.6220703125,
0.08087158203125,
0.6220703125
],
[
0.5048828125,
0.61279296875,
0.8896484375,
0.61279296875,
0.8896484375,
0.62353515625,
0.5048828125,
0.62353515625
],
[
0.08087158203125,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.08087158203125,
0.63720703125
],
[
0.50244140625,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.6396484375,
0.4697265625,
0.6396484375,
0.4697265625,
0.65283203125,
0.080078125,
0.65283203125
],
[
0.0310516357421875,
0.62890625,
0.053924560546875,
0.62744140625,
0.056365966796875,
0.66357421875,
0.03350830078125,
0.6640625
],
[
0.52294921875,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.52294921875,
0.65771484375
],
[
0.02532958984375,
0.6591796875,
0.052276611328125,
0.6572265625,
0.058013916015625,
0.69970703125,
0.0310516357421875,
0.7021484375
],
[
0.08087158203125,
0.65478515625,
0.46728515625,
0.65478515625,
0.46728515625,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50341796875,
0.673828125
],
[
0.080078125,
0.66845703125,
0.4697265625,
0.66845703125,
0.4697265625,
0.681640625,
0.080078125,
0.681640625
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.07843017578125,
0.69873046875,
0.4697265625,
0.69970703125,
0.4697265625,
0.71337890625,
0.07843017578125,
0.7119140625
],
[
0.5009765625,
0.7041015625,
0.8896484375,
0.7041015625,
0.8896484375,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72998046875,
0.468017578125,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.08087158203125,
0.744140625,
0.468017578125,
0.744140625,
0.468017578125,
0.75830078125,
0.08087158203125,
0.75830078125
],
[
0.50341796875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.080078125,
0.76025390625,
0.468017578125,
0.76025390625,
0.468017578125,
0.7734375,
0.080078125,
0.7734375
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.7744140625,
0.4697265625,
0.7744140625,
0.4697265625,
0.78857421875,
0.080078125,
0.78857421875
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79052734375,
0.4697265625,
0.79052734375,
0.4697265625,
0.8037109375,
0.080078125,
0.8037109375
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.080078125,
0.80615234375,
0.4697265625,
0.80615234375,
0.4697265625,
0.8193359375,
0.080078125,
0.8193359375
],
[
0.50244140625,
0.8095703125,
0.88818359375,
0.81103515625,
0.88818359375,
0.82568359375,
0.50244140625,
0.82373046875
],
[
0.080078125,
0.82080078125,
0.468017578125,
0.82080078125,
0.468017578125,
0.833984375,
0.080078125,
0.833984375
],
[
0.5009765625,
0.82470703125,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.83544921875,
0.468017578125,
0.83544921875,
0.468017578125,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.50244140625,
0.85595703125,
0.8935546875,
0.85595703125,
0.8935546875,
0.869140625,
0.50244140625,
0.869140625
],
[
0.096435546875,
0.8623046875,
0.3203125,
0.8623046875,
0.3203125,
0.87548828125,
0.096435546875,
0.87548828125
],
[
0.1004638671875,
0.876953125,
0.226318359375,
0.876953125,
0.226318359375,
0.88720703125,
0.1004638671875,
0.88720703125
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.09722900390625,
0.88720703125,
0.370849609375,
0.88720703125,
0.370849609375,
0.90087890625,
0.09722900390625,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.8837890625,
0.88720703125,
0.8837890625,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"SPLINE-Net: Sparse Photometric Stereo through",
"Lighting Interpolation and Normal Estimation Networks",
"Qian Zhenglt* Yiming Jia2#t Boxin Shi3,4*Xudong Jiang' Ling-Yu Duan3,4 Alex C. Kot'",
"1School of Electrical and Electronic Engineering,Nanyang Technological University, Singapore",
"2Department of Precision Instrument, Tsinghua University,Beijing,China",
"3National Engineering Laboratory for Video Technology, Department of CS, Peking University, Beijing, China",
"4Peng Cheng Laboratory, Shenzhen, China",
"610100",
"[zhengqian, exdjiang, eackot) @ntu.edu.sg, jiaym15@outlook.com, [ shiboxin, lingyu) @pku.edu.cn",
"Abstract",
"This paper solves the Sparse Photometric stereo through",
"Lighting Interpolation and Normal Estimation using a gen-",
"9",
"erative Network (SPLINE-Net). SPLINE-Net contains a",
"lighting interpolation network to generate dense lighting",
"LAD",
"observations given a sparse set of lights as inputs followed",
"bya normal estimation network to estimate surface nor-",
"mals. Both networks are jointly constrained by the pro-",
"(c)",
"(d)",
"(a)",
"'SO",
"posed symmetric and asymmetric loss functions to enforce",
"Figure 1. An illustration of observation maps corresponding to two",
"01",
"isotropic constrain and perform outlier rejection of global",
"surfacenormals (a briefintroduction of observation maps can be",
"illumination effects. SPLINE-Net is verified to outperform",
"foundinSection3.2and[18D).(a)Twosurfacenormalsandtheir",
"288005",
"existing methods for photometric stereo of general BRDFs",
"observation maps with dense lights, (b) sparse observation maps",
"by using only ten images of different lights instead of using",
"with 10 order-agnostic lights, (c) dense observation maps gener-",
"nearly one hundred images.",
"ated by our SPLINE-Netgiven sparseobservationmapsin(b)as",
"inputs, and (d) ground truth of dense observation maps with 1000",
"lights.Weusev",
"(0,0.1)T to represent viewing direction and",
"1.Introduction",
"n to represent surface normal in this paper.",
"The problem of photometric stereo [35] inversely solves",
"based approaches are introduced to solve these difficulties",
"for the radiometric image formation model to recover sur-",
"byimplicitlylearningboththeimageformationprocessand",
"face normals from different appearances of objects under",
"global illumination effects from training data (e.g., [9, 18]).",
"various lighting conditions with a fixed camera view. The",
"I:AI",
"According to a comprehensive benchmark evaluation",
"",
"classicmethod[35lassumes anidealLambertian imagefor-",
"[28] (including quantitative results for representative meth-",
"mation model without global illumination effects (such as",
"ods published before 2016) and the additional results re-",
"inter-reflection and shadows), which deviates from the real-",
"ported in most recent works [9, 18, 40], a moderately dense",
"istic scenarioandpreventsphotometric stereofrombeing",
"lighting distribution (e.g., around 1o0 directional lights ran-",
"able to handle real-world objects.To make photometric",
"domly sampled from the visible hemisphere) is required",
"stereo practical, the major difficulties lie in dealing with",
"to achievereasonably goodnormalestimationforobjects",
"objects of general reflectance and global illumination ef-",
"with general materials (e.g., angular error around 10o for a",
"fects. These can be achieved by either exploring analyti-",
"shiny plastic). This is because multi-illumination observa-",
"cal Bidirectional Reflectance Distribution Function (BRDF)",
"tions under a dense set of lights are required to fit the pa-",
"representations (e.g., [14]) and general BRDF properties",
"rameters in analytic BRDF models [14], to analyze general",
"(e.g., [3o]) to model non-Lambertian interactions of lighting",
"BRDF properties [30], to observe sufficient inliers and out",
"and surface normal or suppressing global effects by treat-",
"liers [37], and to ensure the convergence of training neural",
"ing them as outliers (e.g., [37]). Recently, deep learning",
"networks [9]. How to achieve high accuracy in normal es-",
"timation for objects given general BRDFs with a sparse set",
" Authors contributed equally to this work.",
"*Correspondingauthors.",
"of lights (e.g., 10), which we call sparse photometric stereo",
"t Work completed while interning at NTU and PKU.",
"in this paper, is still an open yet challenging problem [28]."
] |
{
"arxiv_id": "1905.04088",
"language": "en",
"page_idx": 0
}
|
||
1905.04088-en-1
|
1905.04088-en-1.jpg
|
In this paper, we propose to solve Sparse Photometric stereo through Lighting Interpolation and Normal Estimation Networks, namely SPLINE-Net. The SPLINE-Net is composed of two sub-networks: the Lighting Interpolation Network (LI-Net) to generate dense observations given a sparse set of input lights and the Normal Estimation Network (NE-Net) to estimate surface normal from the generated dense observations. LI-Net takes advantage of a learnable representation for dense lighting called observation $map[18]$, and we propose to deal with sparse observation maps as damaged paintings and generate dense observation through inpainting (as shown in Figure 1$^1).$ NE-Net then follows LI-Net to infer surface normal guided by dense observation maps. To accurately guide the lighting interpolation and normal estimation specially under the photometric stereo context, we propose a symmetric loss and an asymmetric loss to explicitly consider general BRDF properties and outlier rejections. More specifically, the symmetric loss is derived according to the property of isotropy for general reflectance, which constrains pixel values on a generated observation map to be symmetrically distributed w.r.t. an axis determined by the corresponding surface normal. The asymmetric loss is derived from contaminated observation maps with global illumination effects, which constrains the difference between values of symmetrically distributed pixels to be equal to a non-zero amount. SPINE-Net is validated to achieve superior normal estimation accuracy given a small number of input images $(e.g.,10)$ comparing to state-of-the-art methods using a much larger number $(e.g.$, 96), which greatly relieves data capture and lighting calibration labor for photometric stereo with general BRDFs. The contributions of this paper are two-fold:
$\bullet$ We propose the SPLINE-Net to address the problem of photometric stereo with general BRDFs using a small number of images through an integrated learning procedure of lighting interpolation and normal estimation.
$\bullet$ We show how symmetric and asymmetric loss functions can be formulated to facilitate the learning of lighting interpolation and normal estimation with isotropy constraint and outlier rejection of global illumination effects considered.
## $2. \textbf{Related Works}$
In this section, we briefly review traditional methods and deep learning based methods for non-Lambertian photometric stereo with general materials and known lightings. For other generalizations of photometric stereo, we refer readers to survey papers in [15,1,28].
$^{1}$Note that holes in ground truth of observation maps are produced by the discrete projection from a limited number of lights to a grid with limited resolution $(e.g.$, from 1000 lighting directions to observation maps with size 32$\times32$ in this figure).
Traditional methods. The classical method[35]for the problem of photometric stereo is to assume a Lambertian surface reflectance and recover surface normals pixelwisely. Such an assumption is too strong to provide an accurate recovery in real-world due to densely observed nonLambertian reflectance caused by materials with diverse reflectance and global illumination effects. In order to address non-Lambertian reflectance from broad classes of materials, modern algorithms attempt to use a mathematically tractable form to describe BRDFs. Analytic models exploit all available data to fit a nonlinear analytic BRDF, such as Blinn-Phong model [33], Torrance-Sparrow model [13], Ward model and its variations [12,14,2], specular spike model[11, 38], and microfacet BRDF model[10]. Empirical models consider general properties of a BRDF, such as isotropy, monotonicity. Some basic derivations for isotropy BRDFs are provided in [4,31,7]. Excellent performance has been achieved by methods based on empirical models, including combining isotropy and monotonicity with visibility constraint [16], using isotropic constraint for the estimation of elevation angle [3,29,22], and approximating isotropic BRDFs by bivariate functions [25,19,30]. However, most of these methods based on analytic models and empirical models are pixel-wise so that they cannot explicitly consider global illumination effects such as interreflection and cast shadows. Outlier rejection based methods are developed to suppress global illumination effects by considering them as outliers. Earlier works select a subset of Lambertian images from inputs for the accurate recovery of surface normals [24,34,23,39,37]. Recent methods apply robust analysis by assuming non-Lambertian reflectance is sparse[36, 20]. However, these methods still rely on the existence of a dominant Lambertian reflectance component.
Deep learning based methods. Recently, with the great success in both high-level and low-level computer vision tasks achieved by neural networks, researchers have introduced deep learning based methods to solve the problem of photometric stereo. Instead of explicitly modeling image formation process and global illumination effects as in traditional methods, deep learning based methods attempt to learn such information from data. DPSN [26] is the first attempt and it uses a deep fully-connect network to regress surface normals from given observations captured under pre-defined lightings in a supervised manner. However, pre-definition of lightings limits its practicality for photometric stereo where the number of inputs often varies. PSFCN[9] is proposed to address such a limitation and handle images under various lightings in an order-agnostic manner by aggregating features of inputs using the max-pooling operation. CNN-PS[18] is another work to accept orderagnostic inputs by introducing observation map, which is a fixed shape representation invariant to inputs. Besides neural networks trained in a supervised manner, Taniai and
|
[
[
0.098876953125,
0.0921630859375,
0.47216796875,
0.0921630859375,
0.47216796875,
0.1060791015625,
0.098876953125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.5009765625,
0.10418701171875
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
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
],
[
0.080078125,
0.15283203125,
0.471435546875,
0.15283203125,
0.471435546875,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.07843017578125,
0.1685791015625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.50341796875,
0.169189453125,
0.8896484375,
0.169189453125,
0.8896484375,
0.179931640625,
0.50341796875,
0.179931640625
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1976318359375,
0.468017578125,
0.1976318359375,
0.468017578125,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.50341796875,
0.199462890625,
0.8896484375,
0.199462890625,
0.8896484375,
0.210205078125,
0.50341796875,
0.210205078125
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.07843017578125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.07843017578125,
0.242431640625
],
[
0.50244140625,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.50244140625,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.07843017578125,
0.25830078125,
0.4697265625,
0.2568359375,
0.4697265625,
0.271484375,
0.07843017578125,
0.272705078125
],
[
0.5009765625,
0.2568359375,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.5009765625,
0.270751953125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.28857421875,
0.46728515625,
0.28857421875,
0.46728515625,
0.3017578125,
0.080078125,
0.3017578125
],
[
0.5009765625,
0.28857421875,
0.890625,
0.287353515625,
0.890625,
0.3017578125,
0.5009765625,
0.302978515625
],
[
0.07843017578125,
0.3037109375,
0.468017578125,
0.302978515625,
0.468017578125,
0.317626953125,
0.07843017578125,
0.318115234375
],
[
0.50244140625,
0.3037109375,
0.8935546875,
0.3037109375,
0.8935546875,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.07843017578125,
0.318115234375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.07843017578125,
0.33203125
],
[
0.50244140625,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.07843017578125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.07843017578125,
0.3466796875
],
[
0.5009765625,
0.33203125,
0.89208984375,
0.333251953125,
0.89208984375,
0.347900390625,
0.5009765625,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.4697265625,
0.349853515625,
0.4697265625,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.35107421875,
0.8896484375,
0.35107421875,
0.8896484375,
0.361083984375,
0.50341796875,
0.361083984375
],
[
0.07843017578125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.07843017578125,
0.376953125
],
[
0.5009765625,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.376953125,
0.5009765625,
0.376953125
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50341796875,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.391357421875,
0.50341796875,
0.391357421875
],
[
0.080078125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.07843017578125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.07843017578125,
0.423583984375
],
[
0.50244140625,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.07843017578125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.07843017578125,
0.438232421875
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.50244140625,
0.43994140625,
0.8935546875,
0.43994140625,
0.8935546875,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.07843017578125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.07843017578125,
0.468505859375
],
[
0.50244140625,
0.455078125,
0.8935546875,
0.455078125,
0.8935546875,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.07843017578125,
0.4697265625,
0.468017578125,
0.468505859375,
0.468017578125,
0.48291015625,
0.07843017578125,
0.484130859375
],
[
0.50341796875,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.481689453125,
0.50341796875,
0.481689453125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.080078125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.07843017578125,
0.51318359375,
0.4697265625,
0.51513671875,
0.4697265625,
0.52978515625,
0.07843017578125,
0.52783203125
],
[
0.50244140625,
0.51513671875,
0.890625,
0.51513671875,
0.890625,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.080078125,
0.52978515625,
0.468017578125,
0.52978515625,
0.468017578125,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.5009765625,
0.52880859375,
0.890625,
0.52978515625,
0.890625,
0.54443359375,
0.5009765625,
0.54345703125
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50341796875,
0.54541015625,
0.89208984375,
0.54541015625,
0.89208984375,
0.5595703125,
0.50341796875,
0.5595703125
],
[
0.08087158203125,
0.56103515625,
0.373291015625,
0.56103515625,
0.373291015625,
0.57470703125,
0.08087158203125,
0.57470703125
],
[
0.50244140625,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.09722900390625,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.5947265625,
0.09722900390625,
0.5947265625
],
[
0.50244140625,
0.57568359375,
0.89208984375,
0.57568359375,
0.89208984375,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.111083984375,
0.5966796875,
0.4697265625,
0.5966796875,
0.4697265625,
0.611328125,
0.111083984375,
0.611328125
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.111083984375,
0.6123046875,
0.4697265625,
0.6123046875,
0.4697265625,
0.62548828125,
0.111083984375,
0.62548828125
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.111083984375,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.6416015625,
0.111083984375,
0.6416015625
],
[
0.50341796875,
0.6318359375,
0.8896484375,
0.6318359375,
0.8896484375,
0.642578125,
0.50341796875,
0.642578125
],
[
0.09722900390625,
0.65087890625,
0.468017578125,
0.65087890625,
0.468017578125,
0.6640625,
0.09722900390625,
0.6640625
],
[
0.50341796875,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.1094970703125,
0.66552734375,
0.4697265625,
0.66650390625,
0.4697265625,
0.6796875,
0.1094970703125,
0.6787109375
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.11273193359375,
0.68115234375,
0.4697265625,
0.68115234375,
0.4697265625,
0.6943359375,
0.11273193359375,
0.6943359375
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.111083984375,
0.6962890625,
0.4697265625,
0.6962890625,
0.4697265625,
0.71044921875,
0.111083984375,
0.71044921875
],
[
0.111083984375,
0.7109375,
0.31201171875,
0.7109375,
0.31201171875,
0.724609375,
0.111083984375,
0.724609375
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50244140625,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50244140625,
0.734375
],
[
0.50244140625,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.744140625,
0.226318359375,
0.744140625,
0.226318359375,
0.7587890625,
0.080078125,
0.7587890625
],
[
0.50341796875,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.1004638671875,
0.77099609375,
0.46728515625,
0.77099609375,
0.46728515625,
0.78076171875,
0.1004638671875,
0.78076171875
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.7861328125,
0.46728515625,
0.7861328125,
0.46728515625,
0.796875,
0.08087158203125,
0.796875
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.7998046875,
0.4697265625,
0.7998046875,
0.4697265625,
0.81298828125,
0.080078125,
0.81298828125
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.080078125,
0.81494140625,
0.46728515625,
0.81494140625,
0.46728515625,
0.8291015625,
0.080078125,
0.8291015625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.83056640625,
0.303955078125,
0.83056640625,
0.303955078125,
0.8447265625,
0.07843017578125,
0.8447265625
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.09722900390625,
0.85107421875,
0.468017578125,
0.85107421875,
0.468017578125,
0.8642578125,
0.09722900390625,
0.8642578125
],
[
0.5009765625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5009765625,
0.87060546875
],
[
0.07843017578125,
0.86279296875,
0.4697265625,
0.86279296875,
0.4697265625,
0.876953125,
0.07843017578125,
0.876953125
],
[
0.07843017578125,
0.87548828125,
0.47216796875,
0.87548828125,
0.47216796875,
0.8896484375,
0.07843017578125,
0.8896484375
],
[
0.5,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5,
0.88623046875
],
[
0.080078125,
0.8896484375,
0.2303466796875,
0.8896484375,
0.2303466796875,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"In this paper, we propose to solve Sparse Photometric",
"Traditional methods. The classical method [35] for the",
"stereo through Lighting Interpolation and Normal Estima-",
"problem of photometric stereo is to assume a Lamber-",
"tion Networks, namely SPLINE-Net. The SPLINE-Net is",
"tiansurfacereflectance andrecover surfacenormalspixel-",
"composed of two sub-networks: the Lighting Interpolation",
"wisely. Such an assumption is too strong to provide an ac-",
"Network (LI-Net) to generate dense observations given a",
"curate recovery in real-world due to densely observed non-",
"sparse set of input lights and the Normal Estimation Net-",
"Lambertianreflectancecausedbymaterialswithdiversere-",
"work (NE-Net) to estimate surface normal from the gener-",
"flectance and global illumination effects. In order to ad-",
"ated dense observations. LI-Net takes advantage of a learn-",
"dressnon-Lambertianreflectancefrombroadclassesofma-",
"able representation for dense lighting called observation",
"terials, modern algorithms attempt to use a mathematically",
"map [18], and we propose to deal with sparse observation",
"tractable form to describe BRDFs.Analytic models ex-",
"maps as damaged paintings and generate dense observation",
"ploit all available data to fit a nonlinear analytic BRDF, such",
"through inpainting (as shown in Figure 1'). NE-Net then",
"as Blinn-Phong model [33], Torrance-Sparrow model [13],",
"follows LI-Net to infer surface normal guided by dense ob-",
"Ward model and its variations [12, 14, 2], specular spike",
"servation maps. To accurately guide the lighting interpola-",
"model [11, 38], and microfacet BRDF model [10]. Empiri-",
"tion and normal estimation specially under the photometric",
"cal models consider general properties of a BRDF, such as",
"stereo context, we propose a symmetric loss and an asym-",
"isotropy,monotonicity.Some basic derivations for isotropy",
"metric loss to explicitly consider general BRDF properties",
"BRDFs are provided in [4,31, 7]. Excellent performance",
"and outlier rejections.More specifically,the symmetric loss",
"has been achievedby methodsbased on empirical mod-",
"is derived according to the property of isotropy for general",
"els, including combining isotropy and monotonicity with",
"reflectance,whichconstrains pixel values on agenerated",
"visibilityconstraint[16],usingisotropicconstraintforthe",
"observation map to be symmetrically distributed w.r.t. an",
"estimation of elevation angle [3, 29, 22], and approximat-",
"axis determined by the corresponding surface normal. The",
"ing isotropic BRDFs by bivariate functions [25, 19, 30].",
"asymmetric loss is derived from contaminated observation",
"However, most of these methods based on analytic mod-",
"mapswithglobalilluminationeffects,whichconstrainsthe",
"els and empirical models are pixel-wise so that they cannot",
"difference between values of symmetrically distributed pix-",
"explicitly consider global illumination effects such as inter-",
"els to be equal to a non-zero amount. SPINE-Net is vali-",
"reflectionandcast shadows.Outlierrejectionbasedmeth-",
"dated to achieve superior normal estimation accuracy given",
"ods are developed to suppress global illumination effects by",
"a small number of input images (e.g., 10) comparing to",
"considering them as outliers. Earlier works select a subset",
"state-of-the-art methods using a much larger number (e.g.,",
"of Lambertian images from inputs for the accurate recovery",
"96), which greatly relieves data capture and lighting cali-",
"of surface normals [24, 34, 23, 39, 37]. Recent methods ap-",
"bration labor for photometric stereo with general BRDFs.",
"ply robust analysis by assuming non-Lambertian reflectance",
"The contributions of this paper are two-fold:",
"is sparse [36, 20]. However, these methods still rely on the",
". We propose the SPLINE-Net to address the problem of",
"existence of a dominant Lambertian reflectance component.",
"photometric stereo with general BRDFs using a small",
"Deep learning based methods. Recently, with the great",
"number of images through an integrated learning pro-",
"success in both high-level and low-level computer vision",
"cedure of lighting interpolation and normal estimation.",
"tasks achieved by neural networks, researchers have intro-",
". We show how symmetric and asymmetric loss func-",
"duced deep learning based methods to solve the problem of",
"tions can be formulated to facilitate the learning",
"photometric stereo. Instead of explicitly modeling image",
"of lighting interpolation and normal estimation with",
"formation process and global illumination effects as in tra-",
"ditional methods, deep learning based methods attempt to",
"isotropy constraint and outlier rejection of global il-",
"lumination effects considered.",
"learn such information from data. DPSN [26] is the first at-",
"tempt and it uses a deep fully-connect network to regress",
"surface normals from given observations captured under",
"2.Related Works",
"pre-defined lightings in a supervised manner. However,",
"Inthissection.webrieflyreviewtraditionalmethodsand",
"pre-definition of lightings limits its practicality for photo-",
"deeplearningbasedmethodsfornon-Lambertianphotomet-",
"metric stereo where the number of inputs often varies.PS-",
"ric stereo with general materials and known lightings. For",
"FCN [9] is proposed to address such a limitation and handle",
"other generalizations of photometric stereo, we refer read-",
"images under various lightings in an order-agnostic man-",
"ers to survey papers in [15, 1, 28].",
"ner by aggregating features of inputs using the max-pooling",
"operation. CNN-PS [18J is another work to accept order-",
"'Note that holes in ground truth of observation maps are produced by",
"agnostic inputs by introducing observation map, which is",
"the discrete projection from a limited number of lights to a grid with limited",
"resolution (e.g-, from 1000 lighting directions to observation maps with",
"a fixed shape representation invariant to inputs. Besides",
"size32×32inthisfigure).",
"neural networks trained in a supervised manner, Taniai and"
] |
{
"arxiv_id": "1905.04088",
"language": "en",
"page_idx": 1
}
|
||
1905.04088-en-2
|
1905.04088-en-2.jpg
|
<FigureHere>
Figure 2. The framework of the proposed SPLINE-Net. The lighting interpolation network generates dense observation maps D given
sparse observation maps $\mathbf{S}$ as inputs. The normal estimation network estimates surface normals n given $\mathbf{S}$ and D as inputs. Both networks
are trained in a supervised manner where ground truth of observation maps $\mathbf{D}_{gt}$ and surface normals $\mathbf{n}_{gt}$ are known.
the other. Specifically, we update LI-Net once after updating NE-Net five times. The loss function for each network composes of a reconstruction loss, a symmetric loss, and an asymmetric loss.
Maehara[32] presented an unsupervised learning framework where surface normals and BRDFs are recovered by minimizing a reconstruction loss between inputs and images synthesized based on a rendering equation.
Only a few earlier works address the problem of photometric stereo with general reflectance using a small number of images in the literature $(e.g.$, analytic model based method[14], shadow analysis based method[6]). Our paper revisits this problem due to its low costs for the labor of data capture and lighting calibration.
Lighting Interpolation Network. Inspired by a recent work [8] that suggests inferring lighting information as intermediate results can benefit the estimation of the surface normal, we design the LI-Net to generate observation maps regarding dense lighting directions. The basic idea is to inpaint sparse observation maps and obtain dense ones, based on learnable properties $(e.g.$, spatial continuity). LI-Net is designed using an encoder-decoder structure due to its excellent image generation capacity [21,41]. The loss function of LI-Net $\mathcal{L}_f$ is formulated as,
## $3. \textbf{ The Proposed SPLINE- Net}$
In this section, we introduce our solution to the problem of photometric stereo with general reflectance using a small number of images. We first present the framework of our SPLINE-Net in Section 3.1. Then we detail the symmetric loss and the asymmetric loss in Section 3.2.
$$\mathcal{L}_{f}=\mathcal{L}_{f}^{rec}+\lambda_{s}\mathcal{L}_{f}^{s}+\lambda_{a}\mathcal{L}_{f}^{a}.$$
(3)
The reconstruction loss $\mathcal{L}_f^{rec}$ is defined as$^2$,
$$\mathcal{L}_{f}^{rec}=\arccos(\mathbf{n}^{\top}\mathbf{n}_{gt})+|\mathbf{D}-\mathbf{D}_{gt}|_{1}+|\mathbf{M}_{s}\circ(\mathbf{D}-\mathbf{D}_{gt})|_{1},$$
## 3.1. Framework
As illustrated in Figure $\color{red}{2,\text{our SPLINE-Net,which con-}}$ sists of a Lighting Interpolation Network (LI-Net) and a Normal Estimation Network (NE-Net), is optimized in a supervised manner. LI-Net (represented as a regression function $f$) interpolates dense observation maps D from sparse observation maps $\mathbf{S} \left ( i. e. , \text{sparse sets of lights}\right ) $,
where $\mathbf{D}=f(\mathbf{S}),\mathbf{n}=g(\mathbf{S},\mathbf{D}),\mathbf{n}_{gt}$ and $\mathbf{D}_gt$ are ground truth of a surface normal and its corresponding dense observation map, respectively, $\mathbf{M}_s$ is a binary mask indicating positions of non-zero value of S, o represents element-wise multiplication. $\mathcal{L}_f^s$ and $\mathcal{L}_f^a$ are our symmetric and asymmetric loss to be introduced in Section 3.2.
Normal Estimation Net. We use the same architecture as that in [18] (a variation of DenseNet [17]) for NE-Net due to its excellent capacity to model the relation between observation maps and surface normals. The loss function of NE-Net is formulated as,
$$f:\mathbf{S}\to\mathbf{D}.$$
(1)
Such densely interpolated observation maps D are then concatenated to original inputs S and help estimate surface normals n in NE-Net (represented as a regression function $g)$
$$g:\mathbf{S},\mathbf{D}\to\mathbf{n}.$$
(2)
(5)
$$\mathcal{L}_{g}=\mathcal{L}_{g}^{rec}+\lambda_{s}\mathcal{L}_{g}^{s}+\lambda_{a}\mathcal{L}_{g}^{a},$$
LI-Net and NE-Net are trained in an alternating iteratively manner, where fixing one network when optimizing
2We experimentally find L1 and L2 distances provide similar results
and here we compute reconstruction loss $\mathcal{L}_f^{rec}$ using L1 distance.
|
[
[
0.2197265625,
0.0921630859375,
0.406982421875,
0.0921630859375,
0.406982421875,
0.1060791015625,
0.2197265625,
0.1060791015625
],
[
0.65625,
0.09344482421875,
0.68701171875,
0.09344482421875,
0.68701171875,
0.10101318359375,
0.65625,
0.10101318359375
],
[
0.68701171875,
0.09344482421875,
0.74658203125,
0.09344482421875,
0.74658203125,
0.10101318359375,
0.68701171875,
0.10101318359375
],
[
0.20751953125,
0.10919189453125,
0.28515625,
0.111083984375,
0.28515625,
0.121826171875,
0.20751953125,
0.11932373046875
],
[
0.325927734375,
0.10986328125,
0.371826171875,
0.10986328125,
0.371826171875,
0.121826171875,
0.325927734375,
0.121826171875
],
[
0.390625,
0.11297607421875,
0.455078125,
0.11297607421875,
0.455078125,
0.12060546875,
0.390625,
0.12060546875
],
[
0.85107421875,
0.11553955078125,
0.87353515625,
0.11553955078125,
0.87353515625,
0.1275634765625,
0.85107421875,
0.1275634765625
],
[
0.326904296875,
0.11932373046875,
0.359375,
0.121826171875,
0.358642578125,
0.1331787109375,
0.325927734375,
0.1307373046875
],
[
0.11273193359375,
0.117431640625,
0.12091064453125,
0.1142578125,
0.125,
0.121826171875,
0.1168212890625,
0.125
],
[
0.71142578125,
0.1300048828125,
0.79345703125,
0.1275634765625,
0.79345703125,
0.1427001953125,
0.71142578125,
0.14453125
],
[
0.74609375,
0.143310546875,
0.76123046875,
0.143310546875,
0.76123046875,
0.15087890625,
0.74609375,
0.15087890625
],
[
0.71337890625,
0.1514892578125,
0.79150390625,
0.1514892578125,
0.79150390625,
0.1622314453125,
0.71337890625,
0.1622314453125
],
[
0.73388671875,
0.179931640625,
0.76123046875,
0.18310546875,
0.75634765625,
0.2083740234375,
0.72802734375,
0.2052001953125
],
[
0.221435546875,
0.1956787109375,
0.390625,
0.1956787109375,
0.390625,
0.208984375,
0.221435546875,
0.208984375
],
[
0.17236328125,
0.212158203125,
0.233642578125,
0.212158203125,
0.233642578125,
0.222900390625,
0.17236328125,
0.222900390625
],
[
0.381591796875,
0.212158203125,
0.441162109375,
0.212158203125,
0.441162109375,
0.222900390625,
0.381591796875,
0.222900390625
],
[
0.73876953125,
0.216552734375,
0.7607421875,
0.216552734375,
0.7607421875,
0.2442626953125,
0.73876953125,
0.2442626953125
],
[
0.08660888671875,
0.2235107421875,
0.111083984375,
0.2235107421875,
0.111083984375,
0.23486328125,
0.08660888671875,
0.23486328125
],
[
0.81787109375,
0.242431640625,
0.8388671875,
0.242431640625,
0.8388671875,
0.255126953125,
0.81787109375,
0.255126953125
],
[
0.2254638671875,
0.25244140625,
0.2344970703125,
0.25244140625,
0.2344970703125,
0.260009765625,
0.2254638671875,
0.260009765625
],
[
0.301513671875,
0.251953125,
0.316162109375,
0.251953125,
0.316162109375,
0.2587890625,
0.301513671875,
0.2587890625
],
[
0.7060546875,
0.27783203125,
0.73388671875,
0.27783203125,
0.73388671875,
0.285400390625,
0.7060546875,
0.285400390625
],
[
0.75732421875,
0.27783203125,
0.7900390625,
0.27783203125,
0.7900390625,
0.285400390625,
0.75732421875,
0.285400390625
],
[
0.12335205078125,
0.2998046875,
0.24267578125,
0.2998046875,
0.24267578125,
0.31005859375,
0.12335205078125,
0.31005859375
],
[
0.25732421875,
0.2998046875,
0.330810546875,
0.2998046875,
0.330810546875,
0.31005859375,
0.25732421875,
0.31005859375
],
[
0.375732421875,
0.2998046875,
0.4697265625,
0.2998046875,
0.4697265625,
0.31005859375,
0.375732421875,
0.31005859375
],
[
0.488525390625,
0.2998046875,
0.55126953125,
0.2998046875,
0.55126953125,
0.31005859375,
0.488525390625,
0.31005859375
],
[
0.734375,
0.2998046875,
0.77685546875,
0.2998046875,
0.77685546875,
0.31005859375,
0.734375,
0.31005859375
],
[
0.7939453125,
0.2998046875,
0.8388671875,
0.2998046875,
0.8388671875,
0.31005859375,
0.7939453125,
0.31005859375
],
[
0.8486328125,
0.298583984375,
0.8896484375,
0.298583984375,
0.8896484375,
0.31005859375,
0.8486328125,
0.31005859375
],
[
0.67724609375,
0.300537109375,
0.70166015625,
0.300537109375,
0.70166015625,
0.30810546875,
0.67724609375,
0.30810546875
],
[
0.5703125,
0.302978515625,
0.5986328125,
0.302978515625,
0.5986328125,
0.306884765625,
0.5703125,
0.306884765625
],
[
0.62744140625,
0.300537109375,
0.65673828125,
0.300537109375,
0.65673828125,
0.309326171875,
0.62744140625,
0.309326171875
],
[
0.080078125,
0.31640625,
0.89208984375,
0.31640625,
0.89208984375,
0.330078125,
0.080078125,
0.330078125
],
[
0.080078125,
0.330810546875,
0.89208984375,
0.330810546875,
0.89208984375,
0.3447265625,
0.080078125,
0.3447265625
],
[
0.08087158203125,
0.3466796875,
0.7646484375,
0.3466796875,
0.7646484375,
0.356689453125,
0.08087158203125,
0.356689453125
],
[
0.080078125,
0.369384765625,
0.468017578125,
0.369384765625,
0.468017578125,
0.38330078125,
0.080078125,
0.38330078125
],
[
0.50341796875,
0.369384765625,
0.88818359375,
0.369384765625,
0.88818359375,
0.38330078125,
0.50341796875,
0.38330078125
],
[
0.07843017578125,
0.38330078125,
0.468017578125,
0.3837890625,
0.468017578125,
0.3984375,
0.07843017578125,
0.397705078125
],
[
0.50244140625,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.50244140625,
0.3984375
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.4140625,
0.080078125,
0.4140625
],
[
0.50341796875,
0.401611328125,
0.890625,
0.401611328125,
0.890625,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.080078125,
0.415283203125,
0.39453125,
0.415283203125,
0.39453125,
0.4287109375,
0.080078125,
0.4287109375
],
[
0.50244140625,
0.415283203125,
0.61376953125,
0.415283203125,
0.61376953125,
0.4287109375,
0.50244140625,
0.4287109375
],
[
0.098876953125,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.4443359375,
0.098876953125,
0.4443359375
],
[
0.50244140625,
0.429931640625,
0.8935546875,
0.429931640625,
0.8935546875,
0.443115234375,
0.50244140625,
0.443115234375
],
[
0.080078125,
0.44580078125,
0.468017578125,
0.44580078125,
0.468017578125,
0.458984375,
0.080078125,
0.458984375
],
[
0.50341796875,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.458984375,
0.50341796875,
0.458984375
],
[
0.07843017578125,
0.460205078125,
0.468017578125,
0.458984375,
0.468017578125,
0.473388671875,
0.07843017578125,
0.474853515625
],
[
0.50244140625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.47607421875,
0.46728515625,
0.47607421875,
0.46728515625,
0.4892578125,
0.080078125,
0.4892578125
],
[
0.50244140625,
0.474853515625,
0.89208984375,
0.47607421875,
0.89208984375,
0.490478515625,
0.50244140625,
0.4892578125
],
[
0.07843017578125,
0.490478515625,
0.4697265625,
0.4892578125,
0.4697265625,
0.50390625,
0.07843017578125,
0.5048828125
],
[
0.50244140625,
0.490478515625,
0.890625,
0.4892578125,
0.890625,
0.50390625,
0.50244140625,
0.5048828125
],
[
0.080078125,
0.505859375,
0.32177734375,
0.505859375,
0.32177734375,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.505859375,
0.89208984375,
0.5048828125,
0.89208984375,
0.51953125,
0.50244140625,
0.52099609375
],
[
0.50244140625,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.07843017578125,
0.53271484375,
0.328369140625,
0.5322265625,
0.328369140625,
0.546875,
0.07843017578125,
0.54736328125
],
[
0.50341796875,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.5498046875,
0.50341796875,
0.5498046875
],
[
0.50244140625,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.098876953125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.098876953125,
0.57275390625
],
[
0.50341796875,
0.56640625,
0.73193359375,
0.56640625,
0.73193359375,
0.580078125,
0.50341796875,
0.580078125
],
[
0.08087158203125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.080078125,
0.5888671875,
0.4697265625,
0.5888671875,
0.4697265625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.60107421875,
0.59033203125,
0.7900390625,
0.59033203125,
0.7900390625,
0.60791015625,
0.60107421875,
0.60791015625
],
[
0.869140625,
0.59228515625,
0.8935546875,
0.59228515625,
0.8935546875,
0.60595703125,
0.869140625,
0.60595703125
],
[
0.07843017578125,
0.6015625,
0.468017578125,
0.60400390625,
0.468017578125,
0.61865234375,
0.07843017578125,
0.6162109375
],
[
0.080078125,
0.61865234375,
0.366943359375,
0.61865234375,
0.366943359375,
0.6318359375,
0.080078125,
0.6318359375
],
[
0.5009765625,
0.6171875,
0.7861328125,
0.6162109375,
0.7861328125,
0.6337890625,
0.5009765625,
0.63525390625
],
[
0.080078125,
0.64013671875,
0.205078125,
0.6416015625,
0.205078125,
0.65576171875,
0.080078125,
0.65478515625
],
[
0.51123046875,
0.646484375,
0.89208984375,
0.646484375,
0.89208984375,
0.6611328125,
0.51123046875,
0.6611328125
],
[
0.098876953125,
0.66552734375,
0.468017578125,
0.66552734375,
0.468017578125,
0.6787109375,
0.098876953125,
0.6787109375
],
[
0.5009765625,
0.66943359375,
0.8935546875,
0.66943359375,
0.8935546875,
0.6875,
0.5009765625,
0.6875
],
[
0.080078125,
0.68115234375,
0.471435546875,
0.68115234375,
0.471435546875,
0.6943359375,
0.080078125,
0.6943359375
],
[
0.50244140625,
0.6875,
0.890625,
0.6875,
0.890625,
0.70068359375,
0.50244140625,
0.70068359375
],
[
0.07843017578125,
0.6943359375,
0.468017578125,
0.69580078125,
0.468017578125,
0.71044921875,
0.07843017578125,
0.708984375
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.07843017578125,
0.7109375,
0.468017578125,
0.71044921875,
0.468017578125,
0.724609375,
0.07843017578125,
0.72607421875
],
[
0.50341796875,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50341796875,
0.73095703125
],
[
0.080078125,
0.7265625,
0.4697265625,
0.7265625,
0.4697265625,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.50341796875,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.50341796875,
0.74560546875
],
[
0.07843017578125,
0.7412109375,
0.384033203125,
0.74072265625,
0.384033203125,
0.7548828125,
0.07843017578125,
0.755859375
],
[
0.50244140625,
0.7470703125,
0.7587890625,
0.7470703125,
0.7587890625,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.2303466796875,
0.7646484375,
0.313720703125,
0.76123046875,
0.314453125,
0.77880859375,
0.2320556640625,
0.7822265625
],
[
0.44677734375,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.44677734375,
0.77978515625
],
[
0.50341796875,
0.76220703125,
0.89208984375,
0.76220703125,
0.89208984375,
0.77587890625,
0.50341796875,
0.77587890625
],
[
0.5048828125,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.7890625,
0.5048828125,
0.7890625
],
[
0.080078125,
0.79150390625,
0.468017578125,
0.79150390625,
0.468017578125,
0.8046875,
0.080078125,
0.8046875
],
[
0.50244140625,
0.79248046875,
0.890625,
0.79248046875,
0.890625,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.806640625,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.080078125,
0.82080078125
],
[
0.50341796875,
0.80810546875,
0.89208984375,
0.80810546875,
0.89208984375,
0.8212890625,
0.50341796875,
0.8212890625
],
[
0.0767822265625,
0.82080078125,
0.4599609375,
0.8212890625,
0.4599609375,
0.8359375,
0.0767822265625,
0.83544921875
],
[
0.5009765625,
0.82080078125,
0.6708984375,
0.82275390625,
0.6708984375,
0.8359375,
0.5009765625,
0.833984375
],
[
0.223876953125,
0.8447265625,
0.32177734375,
0.84228515625,
0.32177734375,
0.85986328125,
0.223876953125,
0.8623046875
],
[
0.44677734375,
0.8466796875,
0.4697265625,
0.8466796875,
0.4697265625,
0.85986328125,
0.44677734375,
0.85986328125
],
[
0.60107421875,
0.8486328125,
0.7900390625,
0.8486328125,
0.7900390625,
0.8662109375,
0.60107421875,
0.8662109375
],
[
0.869140625,
0.85107421875,
0.89208984375,
0.85107421875,
0.89208984375,
0.8642578125,
0.869140625,
0.8642578125
],
[
0.098876953125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.51708984375,
0.87353515625,
0.8935546875,
0.875,
0.8935546875,
0.8896484375,
0.51708984375,
0.88818359375
],
[
0.080078125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.84716796875,
0.88720703125,
0.84716796875,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"Lighting Interpolation Network",
"Recon",
"structionLoss",
"Down-sampling",
"Residual",
"Up-sampling",
"Dg",
"Block",
"S",
"Symmetric Loss",
"and",
"AsymmetricLoss",
"家",
"Normal Estimation Network",
"DenseBlock",
"DenseBlock",
"",
"S,D",
"Dgt",
"中",
"P",
"Recor",
"tionLo",
"Conv.layers (stride=1,2)",
"InstanceNorm",
"conv.layer(stride=2)",
"Avg.Pooling",
"lormalize",
"Sigmoid",
"Flatten",
"Dens",
"Drnnn",
"ReLU",
"Figure 2. The framework of the proposed SPLINE-Net. The lighting interpolation network generates dense observation maps D given",
"sparse observation maps S as inputs. The normal estimation network estimates surface normals n given S and D as inputs. Both networks",
"are trained in a supervised mannerwhereground truthof observation maps Dgtand surface normals ngt areknown.",
"Maehara [32] presented an unsupervised learning frame-",
"the other. Specifically, we update LI-Net once after updat",
"work where surface normals and BRDFs are recovered by",
"ing NE-Net five times. The loss function for each network",
"minimizing a reconstruction loss between inputs and im-",
"composesofareconstructionloss,asymmetricloss,andan",
"ages synthesized based on a rendering equation.",
"asymmetric loss.",
"Only a few earlier works address the problem of photo-",
"Lighting Interpolation Network.Inspired by a recent",
"metric stereo with general reflectance using a small num-",
"work [8] that suggests inferring lighting information as in-",
"ber of images in the literature (e.g., analytic model based",
"termediate results can benefit the estimation of the surface",
"method [14], shadow analysis based method [6]]. Our pa-",
"normal, we design the LI-Net to generate observation maps",
"per revisits this problem due to its low costs for the labor of",
"regarding dense lighting directions. The basic idea is to in-",
"data capture and lighting calibration.",
"paint sparse observation maps and obtain dense ones, based",
"on learnable properties (e.g., spatial continuity). LI-Net is",
"3. The Proposed SPLINE-Net",
"designed using an encoder-decoder structure due to its ex-",
"cellent image generation capacity [21, 41]. The loss func-",
"In this section, we introduce our solution to the problem",
"tion of LI-Net Cf is formulated as,",
"of photometric stereo with general reflectance using a small",
"number of images. We first present the framework of our",
"Cf=Crec+AsC+AaCy",
"(3)",
"SPLINE-Net in Section 3.1. Then we detail the symmetric",
"loss and the asymmetric loss in Section 3.2.",
"The reconstruction loss Crec is defined as?,",
"3.1.Framework",
"Cree = arccos(nT ngt)+|D-Dgtl1+(Mo(D-Dgt)l1, (4)",
"As illustrated in Figure 2, our SPLINE-Net, which con-",
"where D = f(S), n = g(S, D), ngt and Dgt are ground",
"sists of a Lighting Interpolation Network (LI-Net) and a",
"truth of a surface normal and its corresponding dense ob-",
"Normal Estimation Network (NE-Net), is optimized in a su-",
"servation map, respectively, M, is a binary mask indicating",
"pervised manner. LI-Net (represented as a regression func-",
"positions of non-zero value of S, o represents element-wise",
"tion f) interpolates dense observation maps D from sparse",
"multiplication. C and C are our symmetric and asymmet-",
"observation maps S (i.e., sparse sets of lights),",
"ric loss to be introduced in Section 3.2.",
"f : S → D.",
"(1)",
"Normal Estimation Net. We use the same architecture as",
"thatin[18(avariationofDenseNet[I7DforNE-Netdue",
"Such densely interpolated observation maps D are then con-",
"to its excellent capacity to model the relation between ob-",
"catenated to original inputs S and help estimate surface nor-",
"servation maps and surface normals. The loss function of",
"mals n in NE-Net (represented as a regression function g)",
"NE-Net is formulated as,",
"g : S,D→n.",
"(2)",
"Cg=Crec+sC+AaCg",
"(5)",
"LI-Net and NE-Net are trained in an alternating itera-",
"2We experimentally find LI and L2 distances provide similar results",
"tively manner, where fixing one network when optimizing",
"and here we compute reconstruction loss Cree using L distance."
] |
{
"arxiv_id": "1905.04088",
"language": "en",
"page_idx": 2
}
|
||
2203.15458-en-2
|
2203.15458-en-2.jpg
|
<FigureHere>
Figure 4: Confidence network. The network uses CNN to extract features from the intermediate features of pose estimation net, and then uses multi-head attention to fuse the multi-views visual features. the fused multi-view features to confidence of each view.
then fed into a network to predict the confidence. A 3D pose estimation network then pred view with top-$N\text{ confidence }^{1}$, a with regard to the confidence to achieve accurate 3D hand pose results. To reduce the computational cost, we also design an efficient lightweight network by model distillation to predict the view confidence from the input depth itself, which saves the computation cost of po if the pool of candidate vie
3.2 Virtual Multi-View Hand Pose Estimation
We first explain the idea of virtual multi-view hand pose estimation. Inspired by Ge et al. (Ge et al. 2016) that the original camera view may not be optimal for hand pose estimation, we hypothesize that denser v more beneficial and propose to exploit rendered depth on multiple virtual views to estimate 3D hand pose (Fig. 3).
<FigureHere>
Figure 3: Our virtual multi-view hand pose estimation base-
$line.$
Candidate Virtual Views We first define a set of virtual camera views to re-render th we first convert the depth map of the and uniformlysample a series of virtual cameras on the spherical surface centered in the hand point clouds. Note that large camera rotation may cause severe occlusion issue for depth rendering, and thus we only keep the cameras close to the input camera view. I virtual cameras uniformly sampled from the zenith angle in $[-\pi/3,\pi/3]$ and the azimuth angle in $[-\pi/3,\pi/3]$ on the sphere. Refer to Fig. 8 in Sec. 4.3 for illustration.
Virtual View Depth Map Rende from the input depth image can now be re-projected to each of the virtual cameras. Note not provide a complete hand shape, the be partially incomplete or contain wrong occlusion. However, we found empirically that these does not confuse the pose estimation network when the virtual camera is not too far from the input. Therefor dered image as the face rendering methods (Zhou et al.2020; Fu et al. 2021). We also implemented a parallel rendering process using CUDA to speed up the ren
3D Hand Pose Estimation from Single De rendered depth image, we use the Anchor-to-Joint regression network (A2J) (Xiong et al. 2019) as the backbone for 3D hand pose estimation for its great run competitive accuracy. In pra estimation models from depth can be taken as the backbone.
N is a hyper-parameter that controls the run-time efficiency
and pose accuracy.
We use the loss function from A2J and briefly explain here for self-contain. The l sists of the objective loss $L_obj$ an point surrounding loss $L_info.$ The objective loss $L_obj$ aims to minimizing the error between the predicted hand joints with the anchor points and the ground truth hand joints, and the informative anchor point surrounding loss $L_{info}$ aims to selecting informative anchor p to enhance the generalization ability. can be formulated as follows
$$L_{A2J}=\lambda L_{obj}+L_{info}$$
(1)
where $\lambda$ is the loss weight to balance $L_obj$ and $L_info$, and it is
set to $\lambda=3.$
Multi-View Pose Fusion In the end, we run a fusion stage to combine the predicted 3D hand poses from virtual camera views. In the very basic ver of each view to the original camera coordinate system with the camera extrinsic paramet prediction by averaging the transformed hand poses
## 33 Virtual View Selection
The method proposed in the previous section indeed significantly improves the hand pose estimation accuracy, however suffers from run-time efficiency virtual multi-view baseline runs singl tion network multiple times, thus needs linearly proportional computation power w.r.t the size of virtual camera pool size (e.g.25 in our case). In another perspective, the pose prediction on some virtual view sparsity or wrong occlusion, all performance when fused with an average. Therefore, we propose a view selection algorithm that can choose a small number of high impacting camera views without the necessity of running all of them into the pose estimation network.
Confidence Prediction The key of our view selection algorithm is a confidence network that e tance of each candidate virtual view ( computation cost, the confidence network takes the highlevel feature from A2J as input. It then extracts features using convolutional neural network, and then uses the multihead attention in (Vaswani et al. 2017) to fuse multi-view
|
[
[
0.08660888671875,
0.0694580078125,
0.4794921875,
0.0694580078125,
0.4794921875,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.6337890625,
0.0706787109375,
0.69384765625,
0.0706787109375,
0.69384765625,
0.080810546875,
0.6337890625,
0.080810546875
],
[
0.08660888671875,
0.083984375,
0.4794921875,
0.083984375,
0.4794921875,
0.09783935546875,
0.08660888671875,
0.09783935546875
],
[
0.6357421875,
0.08709716796875,
0.69677734375,
0.0889892578125,
0.69677734375,
0.10101318359375,
0.6357421875,
0.0985107421875
],
[
0.71337890625,
0.0889892578125,
0.78173828125,
0.0889892578125,
0.78173828125,
0.0997314453125,
0.71337890625,
0.0997314453125
],
[
0.08660888671875,
0.09783935546875,
0.47802734375,
0.09783935546875,
0.47802734375,
0.111083984375,
0.08660888671875,
0.111083984375
],
[
0.63818359375,
0.0985107421875,
0.69140625,
0.0985107421875,
0.69140625,
0.10986328125,
0.63818359375,
0.10986328125
],
[
0.7138671875,
0.0997314453125,
0.78173828125,
0.0997314453125,
0.78173828125,
0.111083984375,
0.7138671875,
0.111083984375
],
[
0.08660888671875,
0.111083984375,
0.4794921875,
0.111083984375,
0.4794921875,
0.125,
0.08660888671875,
0.125
],
[
0.08905029296875,
0.126953125,
0.4755859375,
0.126953125,
0.4755859375,
0.136962890625,
0.08905029296875,
0.136962890625
],
[
0.53369140625,
0.1319580078125,
0.57861328125,
0.1319580078125,
0.57861328125,
0.1427001953125,
0.53369140625,
0.1427001953125
],
[
0.6435546875,
0.1331787109375,
0.6689453125,
0.1331787109375,
0.6689453125,
0.141357421875,
0.6435546875,
0.141357421875
],
[
0.08660888671875,
0.1395263671875,
0.47802734375,
0.1395263671875,
0.47802734375,
0.15283203125,
0.08660888671875,
0.15283203125
],
[
0.0849609375,
0.15283203125,
0.47802734375,
0.15283203125,
0.47802734375,
0.1666259765625,
0.0849609375,
0.1666259765625
],
[
0.54833984375,
0.147705078125,
0.59228515625,
0.147705078125,
0.59228515625,
0.1590576171875,
0.54833984375,
0.1590576171875
],
[
0.529296875,
0.1572265625,
0.60693359375,
0.1578369140625,
0.60693359375,
0.169189453125,
0.529296875,
0.1685791015625
],
[
0.875,
0.1578369140625,
0.90283203125,
0.1578369140625,
0.90283203125,
0.1666259765625,
0.875,
0.1666259765625
],
[
0.08660888671875,
0.1673583984375,
0.47802734375,
0.1673583984375,
0.47802734375,
0.18115234375,
0.08660888671875,
0.18115234375
],
[
0.0849609375,
0.18115234375,
0.330078125,
0.18115234375,
0.330078125,
0.1944580078125,
0.0849609375,
0.1944580078125
],
[
0.521484375,
0.184326171875,
0.91259765625,
0.184326171875,
0.91259765625,
0.1976318359375,
0.521484375,
0.1976318359375
],
[
0.51953125,
0.1988525390625,
0.91015625,
0.1988525390625,
0.91015625,
0.212158203125,
0.51953125,
0.212158203125
],
[
0.0849609375,
0.20703125,
0.43310546875,
0.20703125,
0.43310546875,
0.2203369140625,
0.0849609375,
0.2203369140625
],
[
0.521484375,
0.2139892578125,
0.9111328125,
0.2139892578125,
0.9111328125,
0.2247314453125,
0.521484375,
0.2247314453125
],
[
0.08660888671875,
0.2259521484375,
0.4755859375,
0.2259521484375,
0.4755859375,
0.2392578125,
0.08660888671875,
0.2392578125
],
[
0.51953125,
0.2247314453125,
0.91259765625,
0.2259521484375,
0.91259765625,
0.240478515625,
0.51953125,
0.2392578125
],
[
0.08660888671875,
0.2412109375,
0.476318359375,
0.2412109375,
0.476318359375,
0.251953125,
0.08660888671875,
0.251953125
],
[
0.51953125,
0.240478515625,
0.89208984375,
0.240478515625,
0.89208984375,
0.25390625,
0.51953125,
0.25390625
],
[
0.0849609375,
0.25390625,
0.47802734375,
0.25244140625,
0.47802734375,
0.266357421875,
0.0849609375,
0.267578125
],
[
0.08660888671875,
0.267578125,
0.47802734375,
0.267578125,
0.47802734375,
0.281005859375,
0.08660888671875,
0.281005859375
],
[
0.0849609375,
0.2822265625,
0.47802734375,
0.2822265625,
0.47802734375,
0.29541015625,
0.0849609375,
0.29541015625
],
[
0.53759765625,
0.2822265625,
0.91259765625,
0.2822265625,
0.91259765625,
0.29541015625,
0.53759765625,
0.29541015625
],
[
0.0849609375,
0.29541015625,
0.455810546875,
0.29541015625,
0.455810546875,
0.309326171875,
0.0849609375,
0.309326171875
],
[
0.5185546875,
0.294189453125,
0.91259765625,
0.29541015625,
0.91259765625,
0.31005859375,
0.5185546875,
0.309326171875
],
[
0.521484375,
0.31005859375,
0.91259765625,
0.31005859375,
0.91259765625,
0.323974609375,
0.521484375,
0.323974609375
],
[
0.521484375,
0.324462890625,
0.9140625,
0.324462890625,
0.9140625,
0.33837890625,
0.521484375,
0.33837890625
],
[
0.291748046875,
0.3447265625,
0.32275390625,
0.3447265625,
0.32275390625,
0.35595703125,
0.291748046875,
0.35595703125
],
[
0.52294921875,
0.339111328125,
0.9111328125,
0.339111328125,
0.9111328125,
0.349853515625,
0.52294921875,
0.349853515625
],
[
0.521484375,
0.3515625,
0.91259765625,
0.3515625,
0.91259765625,
0.365478515625,
0.521484375,
0.365478515625
],
[
0.275390625,
0.35791015625,
0.338134765625,
0.35791015625,
0.338134765625,
0.369384765625,
0.275390625,
0.369384765625
],
[
0.1650390625,
0.36865234375,
0.193603515625,
0.36865234375,
0.193603515625,
0.37255859375,
0.1650390625,
0.37255859375
],
[
0.5185546875,
0.3642578125,
0.9111328125,
0.365478515625,
0.9111328125,
0.380126953125,
0.5185546875,
0.37890625
],
[
0.43310546875,
0.37060546875,
0.47216796875,
0.37060546875,
0.47216796875,
0.3818359375,
0.43310546875,
0.3818359375
],
[
0.52294921875,
0.380615234375,
0.9111328125,
0.380615234375,
0.9111328125,
0.391357421875,
0.52294921875,
0.391357421875
],
[
0.0882568359375,
0.395751953125,
0.47802734375,
0.395751953125,
0.47802734375,
0.4091796875,
0.0882568359375,
0.4091796875
],
[
0.5185546875,
0.391357421875,
0.91259765625,
0.393310546875,
0.91259765625,
0.40771484375,
0.5185546875,
0.406005859375
],
[
0.08660888671875,
0.410400390625,
0.1168212890625,
0.410400390625,
0.1168212890625,
0.42236328125,
0.08660888671875,
0.42236328125
],
[
0.51953125,
0.4072265625,
0.71337890625,
0.4072265625,
0.71337890625,
0.42041015625,
0.51953125,
0.42041015625
],
[
0.64453125,
0.4267578125,
0.78515625,
0.4287109375,
0.78515625,
0.4462890625,
0.64453125,
0.4443359375
],
[
0.89208984375,
0.43115234375,
0.9140625,
0.43115234375,
0.9140625,
0.4443359375,
0.89208984375,
0.4443359375
],
[
0.08660888671875,
0.447509765625,
0.4794921875,
0.447509765625,
0.4794921875,
0.4609375,
0.08660888671875,
0.4609375
],
[
0.521484375,
0.45263671875,
0.9140625,
0.45263671875,
0.9140625,
0.46728515625,
0.521484375,
0.46728515625
],
[
0.0849609375,
0.462158203125,
0.476318359375,
0.462158203125,
0.476318359375,
0.47607421875,
0.0849609375,
0.47607421875
],
[
0.521484375,
0.468505859375,
0.60546875,
0.468505859375,
0.60546875,
0.479736328125,
0.521484375,
0.479736328125
],
[
0.08660888671875,
0.47607421875,
0.476318359375,
0.47607421875,
0.476318359375,
0.4892578125,
0.08660888671875,
0.4892578125
],
[
0.08660888671875,
0.490478515625,
0.47802734375,
0.490478515625,
0.47802734375,
0.50390625,
0.08660888671875,
0.50390625
],
[
0.5185546875,
0.4873046875,
0.91259765625,
0.488037109375,
0.91259765625,
0.50244140625,
0.5185546875,
0.501953125
],
[
0.0849609375,
0.50390625,
0.4794921875,
0.50390625,
0.4794921875,
0.517578125,
0.0849609375,
0.517578125
],
[
0.51953125,
0.50244140625,
0.91259765625,
0.50244140625,
0.91259765625,
0.5166015625,
0.51953125,
0.5166015625
],
[
0.0849609375,
0.517578125,
0.476318359375,
0.517578125,
0.476318359375,
0.53076171875,
0.0849609375,
0.53076171875
],
[
0.521484375,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.521484375,
0.53076171875
],
[
0.0849609375,
0.5322265625,
0.476318359375,
0.5322265625,
0.476318359375,
0.54541015625,
0.0849609375,
0.54541015625
],
[
0.521484375,
0.53076171875,
0.91259765625,
0.53076171875,
0.91259765625,
0.54443359375,
0.521484375,
0.54443359375
],
[
0.08660888671875,
0.546875,
0.476318359375,
0.546875,
0.476318359375,
0.556640625,
0.08660888671875,
0.556640625
],
[
0.52294921875,
0.546875,
0.91259765625,
0.546875,
0.91259765625,
0.556640625,
0.52294921875,
0.556640625
],
[
0.0882568359375,
0.5595703125,
0.47802734375,
0.5595703125,
0.47802734375,
0.57275390625,
0.0882568359375,
0.57275390625
],
[
0.51953125,
0.5595703125,
0.86376953125,
0.5595703125,
0.86376953125,
0.57275390625,
0.51953125,
0.57275390625
],
[
0.0882568359375,
0.57275390625,
0.476318359375,
0.57275390625,
0.476318359375,
0.5859375,
0.0882568359375,
0.5859375
],
[
0.08660888671875,
0.58837890625,
0.408447265625,
0.58837890625,
0.408447265625,
0.5986328125,
0.08660888671875,
0.5986328125
],
[
0.51953125,
0.58349609375,
0.7197265625,
0.583984375,
0.7197265625,
0.5986328125,
0.51953125,
0.59765625
],
[
0.0882568359375,
0.60693359375,
0.476318359375,
0.60693359375,
0.476318359375,
0.62060546875,
0.0882568359375,
0.62060546875
],
[
0.52294921875,
0.60400390625,
0.9111328125,
0.60400390625,
0.9111328125,
0.6171875,
0.52294921875,
0.6171875
],
[
0.08660888671875,
0.62255859375,
0.4794921875,
0.62255859375,
0.4794921875,
0.63623046875,
0.08660888671875,
0.63623046875
],
[
0.521484375,
0.6171875,
0.9111328125,
0.6171875,
0.9111328125,
0.630859375,
0.521484375,
0.630859375
],
[
0.0849609375,
0.63623046875,
0.47802734375,
0.63623046875,
0.47802734375,
0.6494140625,
0.0849609375,
0.6494140625
],
[
0.51953125,
0.6318359375,
0.91259765625,
0.6318359375,
0.91259765625,
0.64599609375,
0.51953125,
0.64599609375
],
[
0.08660888671875,
0.65087890625,
0.476318359375,
0.65087890625,
0.476318359375,
0.6640625,
0.08660888671875,
0.6640625
],
[
0.521484375,
0.64599609375,
0.9111328125,
0.64599609375,
0.9111328125,
0.6591796875,
0.521484375,
0.6591796875
],
[
0.0849609375,
0.6640625,
0.4755859375,
0.6640625,
0.4755859375,
0.67822265625,
0.0849609375,
0.67822265625
],
[
0.51953125,
0.66015625,
0.91259765625,
0.66015625,
0.91259765625,
0.673828125,
0.51953125,
0.673828125
],
[
0.0849609375,
0.67822265625,
0.47802734375,
0.67822265625,
0.47802734375,
0.69140625,
0.0849609375,
0.69140625
],
[
0.51953125,
0.673828125,
0.91259765625,
0.673828125,
0.91259765625,
0.6875,
0.51953125,
0.6875
],
[
0.08905029296875,
0.693359375,
0.476318359375,
0.693359375,
0.476318359375,
0.7041015625,
0.08905029296875,
0.7041015625
],
[
0.521484375,
0.68798828125,
0.9111328125,
0.68798828125,
0.9111328125,
0.7021484375,
0.521484375,
0.7021484375
],
[
0.08660888671875,
0.70556640625,
0.47802734375,
0.70556640625,
0.47802734375,
0.7197265625,
0.08660888671875,
0.7197265625
],
[
0.521484375,
0.7021484375,
0.9140625,
0.7021484375,
0.9140625,
0.71533203125,
0.521484375,
0.71533203125
],
[
0.08660888671875,
0.7197265625,
0.47802734375,
0.7197265625,
0.47802734375,
0.73291015625,
0.08660888671875,
0.73291015625
],
[
0.521484375,
0.71533203125,
0.9111328125,
0.71533203125,
0.9111328125,
0.72900390625,
0.521484375,
0.72900390625
],
[
0.0849609375,
0.73193359375,
0.47802734375,
0.73291015625,
0.47802734375,
0.74755859375,
0.0849609375,
0.7470703125
],
[
0.521484375,
0.73095703125,
0.9111328125,
0.73095703125,
0.9111328125,
0.7412109375,
0.521484375,
0.7412109375
],
[
0.08416748046875,
0.74560546875,
0.450927734375,
0.7470703125,
0.450927734375,
0.76123046875,
0.08416748046875,
0.76025390625
],
[
0.521484375,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.521484375,
0.7568359375
],
[
0.521484375,
0.7568359375,
0.91015625,
0.7568359375,
0.91015625,
0.77099609375,
0.521484375,
0.77099609375
],
[
0.08660888671875,
0.767578125,
0.476318359375,
0.767578125,
0.476318359375,
0.78076171875,
0.08660888671875,
0.78076171875
],
[
0.521484375,
0.77099609375,
0.91259765625,
0.77099609375,
0.91259765625,
0.7841796875,
0.521484375,
0.7841796875
],
[
0.08416748046875,
0.7822265625,
0.47802734375,
0.78271484375,
0.47802734375,
0.796875,
0.08416748046875,
0.79541015625
],
[
0.08660888671875,
0.796875,
0.47802734375,
0.796875,
0.47802734375,
0.81005859375,
0.08660888671875,
0.81005859375
],
[
0.521484375,
0.79150390625,
0.91259765625,
0.79150390625,
0.91259765625,
0.8046875,
0.521484375,
0.8046875
],
[
0.521484375,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.08660888671875,
0.8125,
0.476318359375,
0.8125,
0.476318359375,
0.82275390625,
0.08660888671875,
0.82275390625
],
[
0.0849609375,
0.82470703125,
0.4794921875,
0.82470703125,
0.4794921875,
0.83837890625,
0.0849609375,
0.83837890625
],
[
0.521484375,
0.8212890625,
0.91259765625,
0.8212890625,
0.91259765625,
0.83203125,
0.521484375,
0.83203125
],
[
0.0849609375,
0.83837890625,
0.47802734375,
0.83837890625,
0.47802734375,
0.8515625,
0.0849609375,
0.8515625
],
[
0.521484375,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.521484375,
0.84765625
],
[
0.521484375,
0.84765625,
0.91015625,
0.84765625,
0.91015625,
0.861328125,
0.521484375,
0.861328125
],
[
0.1070556640625,
0.861328125,
0.4794921875,
0.86279296875,
0.4794921875,
0.87744140625,
0.1070556640625,
0.87548828125
],
[
0.51953125,
0.8623046875,
0.91015625,
0.861328125,
0.91015625,
0.875,
0.51953125,
0.87548828125
],
[
0.08416748046875,
0.875,
0.19775390625,
0.876953125,
0.19775390625,
0.89013671875,
0.08416748046875,
0.88818359375
],
[
0.51953125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"then fed into a network to predict the confidence. A 3D pose",
"Confidence Net",
"estimation network then predicts the 3D hand pose from",
"Multi-Head",
"Position-wise",
"view with top-N confidence , and finally fuses the pose",
"Attention",
"Feed Forward",
"with regard to the confidence to achieve accurate 3D hand",
"pose results.Toreducethe computational cost.we alsode-",
"Feature",
"CNN",
"sign an efficient lightweight network by model distillation",
"to predict the view confidence from the input depth itself,",
"3DPose",
"Estimation Net",
"BDJoint",
"which saves the computation cost of point cloud rendering",
"if the pool of candidate view is large.",
"Figure 4: Confidence network. The network uses CNN to",
"extract features from the intermediate features of pose es-",
"3.2 Virtual Multi-View Hand Pose Estimation",
"timation net,and then uses multi-head attention tofusethe",
"We first explain the idea of virtual multi-view hand pose esti-",
"multi-views visual features. We use FC and Softmax to map",
"mation.Inspired by Ge etal. (Ge et al.2016)that the original",
"the fused multi-view features to confidence of each view.",
"camera view may not be optimal for hand pose estimation,",
"we hypothesize that denser virtual view sampling should be",
"more beneficial and propose to exploit rendered depth on",
"We use the loss function from A2J and briefly explain",
"multiple virtual views to estimate 3D hand pose (Fig. 3).",
"here for self-contain. The loss function LA2J of A2J con-",
"sists of the objective loss Lobj and the informative anchor",
"point surrounding loss Linfo. The objective loss Lobj aims to",
"Pose",
"minimizingtheerrorbetweenthepredictedhand jointswith",
"the anchor points and the ground truth hand joints, and the",
"Estimation",
"pana",
"informative anchor point surrounding loss Linfo aims to se-",
"3DJoints",
"lectinginformativeanchorpointslocated aroundthejoints",
"Figure 3: Our virtual multi-view hand pose estimation base-",
"to enhance the generalization ability. The loss function LA2J",
"line.",
"can be formulated as follows",
"LA2J =入Lobj + Linfo",
"(1)",
"Candidate Virtual ViewsWe first define a set of virtual",
"where is the loss weight to balance Lob and Linfo, and it is",
"camera views to re-render the input depth map. Specifically",
"set to 入 = 3.",
"we first convert the depth map of the hand into point clouds",
"and uniformly sample a series of virtual cameras on the",
"Multi-View Pose Fusion In the end, we run a fusion stage",
"spherical surface centered in the hand point clouds. Note",
"to combine the predicted 3D hand poses from virtual camera",
"that large camera rotation may cause severe occlusion is-",
"views. In the very basic version, we transform the 3D joints",
"sue for depth rendering, and thus we only keep the cam-",
"of each view to the original camera coordinate system with",
"erasclosetotheinputcameraview.Inpractice,wekeep25",
"thecameraextrinsicparameters,andgetthefinalhandpose",
"virtual cameras uniformly sampled from the zenith angle in",
"prediction by averaging the transformed hand poses.",
"[-π/3, π/3] and the azimuth angle in [-π/3, /3] on the",
"sphere.RefertoFig.8 in Sec.4.3forillustration.",
"3.3Virtual View Selection",
"Virtual View Depth Map Rendering The point cloud",
"The method proposed in the previous section indeed signif-",
"from the input depth image can now be re-projected to each",
"icantly improves the hand pose estimation accuracy,how-",
"of the virtual cameras. Note that since the depth images does",
"ever suffers from run-time efficiency issue. Essentially, the",
"not provide a complete hand shape, the rendered depth may",
"virtual multi-view baseline runs single view pose estima-",
"be partially incomplete or contain wrong occlusion. How-",
"tion network multiple times, thus needs linearly proportional",
"ever, we found empirically that these does not confuse the",
"computation power w.r.t the size of virtual camera pool size",
"poseestimationnetworkwhenthevirtual cameraisnottoo",
"(e.g. 25 in our case). In another perspective, the pose pre-",
"far from the input. Therefore, we did not complete the ren-",
"diction on some virtual views may not be great, e.g. due to",
"dered image as the face rendering methods (Zhou et al. 2020;",
"sparsity or wrong occlusion, and may further hurt the over-",
"Fu et al. 2021). We also implemented a parallel rendering",
"allperformance whenfused with an average.Therefore,we",
"process using CUDA to speed up the rendering process.",
"propose a view selection algorithm that can choose a small",
"number of high impacting camera views without the neces-",
"3D Hand Pose Estimationfrom Single DepthFor each",
"sity of running all of them into the pose estimation network.",
"rendered depth image,we use the Anchor-to-Joint regression",
"network (A2J) (Xiong et al. 2019) as the backbone for 3D",
"Confidence PredictionThe key of our view selection al-",
"gorithm is a confidence network that evaluates the impor-",
"hand pose estimation forits great run-time efficiency and",
"competitive accuracy. In practice, any other 3D hand pose",
"tance of each candidate virtual view (Fig. 4).To save the",
"estimation models from depth can be taken as the backbone.",
"computation cost, the confidence network takes the high-",
"level feature from A2J as input. It then extracts features us-",
"'N is a hyper-parameter that controls the run-time efficiency",
"ing convolutional neural network, and then uses the multi",
"and poseaccuracy.",
"head attention in (Vaswani et al. 2017) to fuse multi-view"
] |
{
"arxiv_id": "2203.15458",
"language": "en",
"page_idx": 2
}
|
||
2203.15458-en-4
|
2203.15458-en-4.jpg
|
4.2 Comparison with State-of-the-art Methods
<table>
<tbody>
<tr>
<th>$\mathbf{Methods}$</th>
<th>$NYU$</th>
<th>$ICVL$</th>
</tr>
<tr>
<td>HandPointNet ( Ge et al. 2018) </td>
<td>10.54</td>
<td>6.93</td>
</tr>
<tr>
<td>DenseReg ( Wan et al. 2018) </td>
<td>10.21</td>
<td>7.24</td>
</tr>
<tr>
<td>P2P$\left(\mathrm{Ge},\mathrm{Ren},\mathrm{~and~Yuan~2018}\right)$</td>
<td>9.05</td>
<td>6.33</td>
</tr>
<tr>
<td>A2J ( Xiong et al. 2019) </td>
<td>8.61</td>
<td>6.46</td>
</tr>
<tr>
<td>V2V (Moon, Chang, and Lee 2018)</td>
<td>8.42</td>
<td>6.28</td>
</tr>
<tr>
<td>AWR ( Huang et al. 2020) </td>
<td>7.37</td>
<td>5.98</td>
</tr>
<tr>
<td>Ours- 1view</td>
<td>7.34</td>
<td>$\overline{5.16}$</td>
</tr>
<tr>
<td>Ours- 3views</td>
<td>6.82</td>
<td>4.86</td>
</tr>
<tr>
<td>Ours- 9views</td>
<td>6.53</td>
<td>4.77</td>
</tr>
<tr>
<td>Ours- 15views</td>
<td>6.41</td>
<td>4.76</td>
</tr>
<tr>
<td>Ours- 25views</td>
<td>6.40</td>
<td>4.79</td>
</tr>
</tbody>
</table>
Table 1: Comparison mean joint 3D error (mm) and ranking result with state-of-art methods on NYU dataset and ICVL dataset.“Ours-1view”,“Ours-3views”,“Ours-9views”and “Ours-15views" are the results of our method with selected 1,3,9 and 15 views from 25 uniformly sampled views, respectively.“Outs-25views”denotes the results of our method with 25 uniformly sampled views.
dence predicted by teacher network as the ground truth. The
loss is defined as
$$L_{light}=\sum_{i=1}^{M}L_{\tau}(\beta(c_{i}-\hat{c}_{i}))$$
(4)
where $\beta$ is the scaling factor set to 100, $\hat{c}_i$ is the confdence ground truth from teacher network, i.e. the multi-head attention network from $M$ views, and $c_i$ is the student network output.
Once the training is done, the teacher network is no longer needed during the inference, and student network can support the view selection and fusion in an efficient way.
## 3.4 Implementation Details
We train and evaluate our models on a workstation with two Intel Xeon Silver 4210R, 512GB of RAM and an Nvidia RTX3090 GPU. Our models are implemented within PyTorch. Adam optimizer is used; the initial learning rate is set to 0.001 and is decayed by 0.9 per epoch.In order to conduct data augmentation, we randomly scale the cropped depth map, jitter the centroid of the point cloud, and randomly rotate the camera when rendering multi-view depth. For all smooth$_L1$ loss, the switch point between quadratic and linear is set to 1.0.
Our network consists of the 3D pose estimation network (i.e. A2J), the teacher confidence network and the lightweight student confidence network. The network input is 176$\times176$ hand region cropped from the input depth, and we use ResNet-50 as the backbone of A2J. We first train the 3D pose estimation network and the teacher confidence network together, the loss can be formulated as:
$$L_{viewsel}=L_{A2J}+\gamma L_{J}$$
(5)
where $\gamma=0.1$ is the factor to balance the loss terms.
Then we fix the parameters of the two networks and train
the lightweight student confidence network with $L_light.$
## $4. \textbf{Experiments}$
4.1 Datasets and Evaluation Metric
NYU Hand Pose Dataset (NYU) (Tompson et al. 2014)
contains 72,757 frames for training and 8,252 frames for testing. 36 hand joints are annotated, but we use only a subset of 14 hand joints for evaluations following the same evaluation protocol in (Tompson et al. 2014).
ICVL Hand Pose Dataset (ICVL) (Tang et al. 2014) contains 331,006 frames for training and 1,596 frames for testing. 16 hand joints are annotated.
Task 1 of Hands19 Challenge Dataset (Hands19-Task1) (Armagan et al. 2020) contains 175,951 training depth images from 5 subjects and 124,999 testing depth images from 10 subjects, in which 5 subjects overlap with the training set. This dataset is very challenging because of its exhaustive coverage of viewpoints and articulations.
Evaluation Metric We evaluate the hand pose estimation performance using standard metrics proposed in (Tang et al. 2014),i.e. mean joint error and the percentage of test examples that have all predicted joints within a given maximum distance from the ground truth.
<FigureHere>
Figure 6: Comparison of our proposed method with stateof-the-art methods on NYU dataset. Left: mean joint error per hand joint. Right: the percentage of success frames over different error thresholds.
In this experiment, we use our “student” lightweight con-
fdence network for view selection and pose fusion.
We first compare our method with the state-of-the-art methods on NYU dataset and ICVL dataset. DenseReg (Wan et al. 2018), A2J (Xiong et al. 2019) and AWR (Huang et al. 2020) directly use the depth map for pose estimation. HandPointNet (Ge et al. 2018), P2P (Ge, Ren, and Yuan 2018) and V2V (Moon, Chang, and Lee 2018)(Hrúz, Kanis, and Krňoul 2021) use the 3D representation of the depth map for pose estimation.
Table 1 shows the mean joint error. Fig.6 shows the percentage of success frames over different error thresholds and the error of each joint. We do not show AWR in Fig. 6 because the best prediction results on NYU dataset are not released by AWR. Note that with just 1 view selected, we already outperform all the other methods. And when more view selected, the performance keeps improving. This clearly indicates that our view selection is effective in
|
[
[
0.08660888671875,
0.0694580078125,
0.4794921875,
0.0694580078125,
0.4794921875,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.51953125,
0.06756591796875,
0.875,
0.06817626953125,
0.875,
0.08331298828125,
0.51953125,
0.08209228515625
],
[
0.08416748046875,
0.08331298828125,
0.19775390625,
0.08331298828125,
0.19775390625,
0.09661865234375,
0.08416748046875,
0.09661865234375
],
[
0.25,
0.09661865234375,
0.26953125,
0.09661865234375,
0.26953125,
0.10986328125,
0.25,
0.10986328125
],
[
0.1895751953125,
0.10986328125,
0.253173828125,
0.11236572265625,
0.253173828125,
0.126953125,
0.1895751953125,
0.125
],
[
0.2467041015625,
0.10986328125,
0.373291015625,
0.10986328125,
0.373291015625,
0.125,
0.2467041015625,
0.125
],
[
0.79736328125,
0.10986328125,
0.83837890625,
0.10986328125,
0.83837890625,
0.1256103515625,
0.79736328125,
0.1256103515625
],
[
0.8544921875,
0.10986328125,
0.900390625,
0.10986328125,
0.900390625,
0.1256103515625,
0.8544921875,
0.1256103515625
],
[
0.625,
0.111083984375,
0.689453125,
0.111083984375,
0.689453125,
0.1256103515625,
0.625,
0.1256103515625
],
[
0.455810546875,
0.111083984375,
0.4794921875,
0.111083984375,
0.4794921875,
0.125,
0.455810546875,
0.125
],
[
0.25,
0.1287841796875,
0.265625,
0.1287841796875,
0.265625,
0.1363525390625,
0.25,
0.1363525390625
],
[
0.79833984375,
0.1307373046875,
0.8388671875,
0.1307373046875,
0.8388671875,
0.146484375,
0.79833984375,
0.146484375
],
[
0.859375,
0.1307373046875,
0.89453125,
0.1307373046875,
0.89453125,
0.146484375,
0.859375,
0.146484375
],
[
0.08660888671875,
0.1395263671875,
0.4794921875,
0.1395263671875,
0.4794921875,
0.15283203125,
0.08660888671875,
0.15283203125
],
[
0.55126953125,
0.1319580078125,
0.76123046875,
0.1319580078125,
0.76123046875,
0.1458740234375,
0.55126953125,
0.1458740234375
],
[
0.5595703125,
0.1458740234375,
0.7548828125,
0.1458740234375,
0.7548828125,
0.1590576171875,
0.5595703125,
0.1590576171875
],
[
0.79736328125,
0.1458740234375,
0.83837890625,
0.1458740234375,
0.83837890625,
0.1610107421875,
0.79736328125,
0.1610107421875
],
[
0.859375,
0.1458740234375,
0.896484375,
0.1458740234375,
0.896484375,
0.1610107421875,
0.859375,
0.1610107421875
],
[
0.08416748046875,
0.154052734375,
0.47802734375,
0.15283203125,
0.47802734375,
0.1666259765625,
0.08416748046875,
0.1673583984375
],
[
0.5498046875,
0.1590576171875,
0.76318359375,
0.160400390625,
0.76318359375,
0.1749267578125,
0.5498046875,
0.173583984375
],
[
0.7998046875,
0.160400390625,
0.8349609375,
0.160400390625,
0.8349609375,
0.175537109375,
0.7998046875,
0.175537109375
],
[
0.859375,
0.160400390625,
0.89453125,
0.160400390625,
0.89453125,
0.175537109375,
0.859375,
0.175537109375
],
[
0.0849609375,
0.1673583984375,
0.4794921875,
0.1673583984375,
0.4794921875,
0.18115234375,
0.0849609375,
0.18115234375
],
[
0.7998046875,
0.1749267578125,
0.833984375,
0.1749267578125,
0.833984375,
0.1900634765625,
0.7998046875,
0.1900634765625
],
[
0.859375,
0.1749267578125,
0.896484375,
0.1749267578125,
0.896484375,
0.1900634765625,
0.859375,
0.1900634765625
],
[
0.08660888671875,
0.1817626953125,
0.13232421875,
0.1817626953125,
0.13232421875,
0.1932373046875,
0.08660888671875,
0.1932373046875
],
[
0.57666015625,
0.175537109375,
0.73876953125,
0.175537109375,
0.73876953125,
0.189453125,
0.57666015625,
0.189453125
],
[
0.53369140625,
0.186279296875,
0.78125,
0.1875,
0.78125,
0.2052001953125,
0.53369140625,
0.203857421875
],
[
0.7998046875,
0.189453125,
0.8349609375,
0.189453125,
0.8349609375,
0.2052001953125,
0.7998046875,
0.2052001953125
],
[
0.861328125,
0.1900634765625,
0.8935546875,
0.1900634765625,
0.8935546875,
0.20263671875,
0.861328125,
0.20263671875
],
[
0.7998046875,
0.20263671875,
0.8349609375,
0.20263671875,
0.8349609375,
0.2183837890625,
0.7998046875,
0.2183837890625
],
[
0.10296630859375,
0.1944580078125,
0.4794921875,
0.1944580078125,
0.4794921875,
0.2083740234375,
0.10296630859375,
0.2083740234375
],
[
0.5703125,
0.203857421875,
0.74609375,
0.203857421875,
0.74609375,
0.2171630859375,
0.5703125,
0.2171630859375
],
[
0.859375,
0.20263671875,
0.89453125,
0.20263671875,
0.89453125,
0.2183837890625,
0.859375,
0.2183837890625
],
[
0.80224609375,
0.2197265625,
0.833984375,
0.2197265625,
0.833984375,
0.231689453125,
0.80224609375,
0.231689453125
],
[
0.08416748046875,
0.2083740234375,
0.47802734375,
0.208984375,
0.47802734375,
0.222900390625,
0.08416748046875,
0.2215576171875
],
[
0.61669921875,
0.2203369140625,
0.69775390625,
0.2203369140625,
0.69775390625,
0.231689453125,
0.61669921875,
0.231689453125
],
[
0.861328125,
0.2197265625,
0.89453125,
0.2197265625,
0.89453125,
0.231689453125,
0.861328125,
0.231689453125
],
[
0.7998046875,
0.231689453125,
0.8349609375,
0.231689453125,
0.8349609375,
0.2474365234375,
0.7998046875,
0.2474365234375
],
[
0.0849609375,
0.222900390625,
0.435546875,
0.222900390625,
0.435546875,
0.236083984375,
0.0849609375,
0.236083984375
],
[
0.611328125,
0.23291015625,
0.70166015625,
0.23291015625,
0.70166015625,
0.2474365234375,
0.611328125,
0.2474365234375
],
[
0.861328125,
0.2342529296875,
0.89453125,
0.2342529296875,
0.89453125,
0.246826171875,
0.861328125,
0.246826171875
],
[
0.08660888671875,
0.246826171875,
0.294189453125,
0.246826171875,
0.294189453125,
0.260009765625,
0.08660888671875,
0.260009765625
],
[
0.611328125,
0.246826171875,
0.70361328125,
0.248779296875,
0.70361328125,
0.26318359375,
0.611328125,
0.261474609375
],
[
0.80126953125,
0.248779296875,
0.833984375,
0.248779296875,
0.833984375,
0.261474609375,
0.80126953125,
0.261474609375
],
[
0.85888671875,
0.2474365234375,
0.89453125,
0.2474365234375,
0.89453125,
0.26318359375,
0.85888671875,
0.26318359375
],
[
0.08660888671875,
0.26318359375,
0.4794921875,
0.26318359375,
0.4794921875,
0.276611328125,
0.08660888671875,
0.276611328125
],
[
0.609375,
0.26318359375,
0.7060546875,
0.26318359375,
0.7060546875,
0.274658203125,
0.609375,
0.274658203125
],
[
0.7998046875,
0.261474609375,
0.8349609375,
0.261474609375,
0.8349609375,
0.276611328125,
0.7998046875,
0.276611328125
],
[
0.861328125,
0.26318359375,
0.89453125,
0.26318359375,
0.89453125,
0.27587890625,
0.861328125,
0.27587890625
],
[
0.7998046875,
0.27587890625,
0.83642578125,
0.27587890625,
0.83642578125,
0.291748046875,
0.7998046875,
0.291748046875
],
[
0.85888671875,
0.27587890625,
0.89453125,
0.27587890625,
0.89453125,
0.291748046875,
0.85888671875,
0.291748046875
],
[
0.609375,
0.27783203125,
0.7060546875,
0.27783203125,
0.7060546875,
0.2890625,
0.609375,
0.2890625
],
[
0.0882568359375,
0.279052734375,
0.47802734375,
0.279052734375,
0.47802734375,
0.2890625,
0.0882568359375,
0.2890625
],
[
0.08660888671875,
0.290283203125,
0.476318359375,
0.290283203125,
0.476318359375,
0.3037109375,
0.08660888671875,
0.3037109375
],
[
0.0849609375,
0.30615234375,
0.4794921875,
0.30615234375,
0.4794921875,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.521484375,
0.306884765625,
0.9140625,
0.306884765625,
0.9140625,
0.32080078125,
0.521484375,
0.32080078125
],
[
0.0849609375,
0.317626953125,
0.47802734375,
0.317626953125,
0.47802734375,
0.330810546875,
0.0849609375,
0.330810546875
],
[
0.51953125,
0.32080078125,
0.9111328125,
0.32080078125,
0.9111328125,
0.333984375,
0.51953125,
0.333984375
],
[
0.0849609375,
0.333251953125,
0.4794921875,
0.333251953125,
0.4794921875,
0.3466796875,
0.0849609375,
0.3466796875
],
[
0.51953125,
0.333984375,
0.91259765625,
0.333984375,
0.91259765625,
0.348388671875,
0.51953125,
0.348388671875
],
[
0.0849609375,
0.3466796875,
0.47802734375,
0.3466796875,
0.47802734375,
0.35986328125,
0.0849609375,
0.35986328125
],
[
0.51953125,
0.348388671875,
0.91015625,
0.348388671875,
0.91015625,
0.3623046875,
0.51953125,
0.3623046875
],
[
0.08660888671875,
0.3623046875,
0.4755859375,
0.3623046875,
0.4755859375,
0.37255859375,
0.08660888671875,
0.37255859375
],
[
0.51953125,
0.363037109375,
0.9140625,
0.363037109375,
0.9140625,
0.376953125,
0.51953125,
0.376953125
],
[
0.0849609375,
0.375,
0.4794921875,
0.375,
0.4794921875,
0.38818359375,
0.0849609375,
0.38818359375
],
[
0.51953125,
0.376953125,
0.9140625,
0.376953125,
0.9140625,
0.39013671875,
0.51953125,
0.39013671875
],
[
0.0849609375,
0.389404296875,
0.23779296875,
0.389404296875,
0.23779296875,
0.40283203125,
0.0849609375,
0.40283203125
],
[
0.51953125,
0.391357421875,
0.82421875,
0.391357421875,
0.82421875,
0.40478515625,
0.51953125,
0.40478515625
],
[
0.10455322265625,
0.404052734375,
0.4755859375,
0.404052734375,
0.4755859375,
0.4140625,
0.10455322265625,
0.4140625
],
[
0.08660888671875,
0.417236328125,
0.47802734375,
0.417236328125,
0.47802734375,
0.427978515625,
0.08660888671875,
0.427978515625
],
[
0.08416748046875,
0.427978515625,
0.4794921875,
0.429931640625,
0.4794921875,
0.443115234375,
0.08416748046875,
0.44140625
],
[
0.0849609375,
0.443115234375,
0.47802734375,
0.443115234375,
0.47802734375,
0.45703125,
0.0849609375,
0.45703125
],
[
0.53369140625,
0.453857421875,
0.53759765625,
0.453857421875,
0.53759765625,
0.460205078125,
0.53369140625,
0.460205078125
],
[
0.5498046875,
0.455078125,
0.55810546875,
0.460205078125,
0.55224609375,
0.46533203125,
0.5439453125,
0.460205078125
],
[
0.859375,
0.455810546875,
0.89794921875,
0.455078125,
0.89794921875,
0.460205078125,
0.859375,
0.4609375
],
[
0.56396484375,
0.458251953125,
0.576171875,
0.4609375,
0.57275390625,
0.46533203125,
0.56201171875,
0.46337890625
],
[
0.0849609375,
0.45703125,
0.4794921875,
0.45703125,
0.4794921875,
0.47021484375,
0.0849609375,
0.47021484375
],
[
0.525390625,
0.464111328125,
0.53759765625,
0.4609375,
0.5400390625,
0.468505859375,
0.529296875,
0.4716796875
],
[
0.580078125,
0.458984375,
0.59326171875,
0.458984375,
0.59326171875,
0.46337890625,
0.580078125,
0.46337890625
],
[
0.85693359375,
0.46533203125,
0.89208984375,
0.46337890625,
0.8935546875,
0.474853515625,
0.85888671875,
0.4765625
],
[
0.08660888671875,
0.4716796875,
0.47802734375,
0.4716796875,
0.47802734375,
0.48486328125,
0.08660888671875,
0.48486328125
],
[
0.529296875,
0.4716796875,
0.53759765625,
0.4716796875,
0.53759765625,
0.47607421875,
0.529296875,
0.47607421875
],
[
0.52294921875,
0.479736328125,
0.5361328125,
0.47607421875,
0.5400390625,
0.48486328125,
0.52685546875,
0.4892578125
],
[
0.85888671875,
0.47607421875,
0.869140625,
0.47607421875,
0.869140625,
0.486083984375,
0.85888671875,
0.486083984375
],
[
0.0882568359375,
0.4873046875,
0.381591796875,
0.4873046875,
0.381591796875,
0.49755859375,
0.0882568359375,
0.49755859375
],
[
0.859375,
0.4873046875,
0.88623046875,
0.4873046875,
0.88623046875,
0.49560546875,
0.859375,
0.49560546875
],
[
0.20751953125,
0.501953125,
0.35546875,
0.50048828125,
0.35546875,
0.51513671875,
0.20751953125,
0.5166015625
],
[
0.45751953125,
0.501953125,
0.4794921875,
0.501953125,
0.4794921875,
0.51513671875,
0.45751953125,
0.51513671875
],
[
0.859375,
0.5,
0.88134765625,
0.5,
0.88134765625,
0.50830078125,
0.859375,
0.50830078125
],
[
0.0849609375,
0.51953125,
0.42724609375,
0.51953125,
0.42724609375,
0.52978515625,
0.0849609375,
0.52978515625
],
[
0.5556640625,
0.52783203125,
0.57666015625,
0.5341796875,
0.57275390625,
0.54248046875,
0.55126953125,
0.5361328125
],
[
0.6044921875,
0.5263671875,
0.69921875,
0.5244140625,
0.69921875,
0.54248046875,
0.6044921875,
0.54443359375
],
[
0.81396484375,
0.5322265625,
0.869140625,
0.5322265625,
0.869140625,
0.53515625,
0.81396484375,
0.53515625
],
[
0.5947265625,
0.54052734375,
0.61376953125,
0.54052734375,
0.61376953125,
0.54443359375,
0.5947265625,
0.54443359375
],
[
0.10296630859375,
0.5322265625,
0.4794921875,
0.5322265625,
0.4794921875,
0.54541015625,
0.10296630859375,
0.54541015625
],
[
0.08416748046875,
0.54443359375,
0.449462890625,
0.546875,
0.449462890625,
0.56103515625,
0.08416748046875,
0.5595703125
],
[
0.521484375,
0.56298828125,
0.91259765625,
0.56298828125,
0.91259765625,
0.57763671875,
0.521484375,
0.57763671875
],
[
0.2181396484375,
0.57275390625,
0.34716796875,
0.57275390625,
0.34716796875,
0.5869140625,
0.2181396484375,
0.5869140625
],
[
0.521484375,
0.5791015625,
0.9111328125,
0.5791015625,
0.9111328125,
0.5888671875,
0.521484375,
0.5888671875
],
[
0.08660888671875,
0.59033203125,
0.35693359375,
0.59033203125,
0.35693359375,
0.60498046875,
0.08660888671875,
0.60498046875
],
[
0.51953125,
0.59130859375,
0.91259765625,
0.59130859375,
0.91259765625,
0.60498046875,
0.51953125,
0.60498046875
],
[
0.0849609375,
0.60693359375,
0.47802734375,
0.60693359375,
0.47802734375,
0.62060546875,
0.0849609375,
0.62060546875
],
[
0.51953125,
0.60400390625,
0.69140625,
0.60400390625,
0.69140625,
0.6171875,
0.51953125,
0.6171875
],
[
0.0849609375,
0.62060546875,
0.4794921875,
0.62060546875,
0.4794921875,
0.6337890625,
0.0849609375,
0.6337890625
],
[
0.0849609375,
0.63525390625,
0.476318359375,
0.63525390625,
0.476318359375,
0.6484375,
0.0849609375,
0.6484375
],
[
0.5361328125,
0.63623046875,
0.91259765625,
0.63623046875,
0.91259765625,
0.65087890625,
0.5361328125,
0.65087890625
],
[
0.0849609375,
0.6484375,
0.47802734375,
0.6484375,
0.47802734375,
0.662109375,
0.0849609375,
0.662109375
],
[
0.5185546875,
0.6494140625,
0.85693359375,
0.65087890625,
0.85693359375,
0.66552734375,
0.5185546875,
0.6640625
],
[
0.0849609375,
0.66357421875,
0.35546875,
0.66357421875,
0.35546875,
0.6767578125,
0.0849609375,
0.6767578125
],
[
0.53515625,
0.6640625,
0.9140625,
0.66650390625,
0.9140625,
0.68115234375,
0.53515625,
0.6787109375
],
[
0.08660888671875,
0.681640625,
0.325927734375,
0.681640625,
0.325927734375,
0.69580078125,
0.08660888671875,
0.69580078125
],
[
0.328369140625,
0.681640625,
0.4794921875,
0.681640625,
0.4794921875,
0.69580078125,
0.328369140625,
0.69580078125
],
[
0.51953125,
0.6796875,
0.91259765625,
0.6796875,
0.91259765625,
0.6943359375,
0.51953125,
0.6943359375
],
[
0.326904296875,
0.68603515625,
0.330810546875,
0.68603515625,
0.330810546875,
0.69140625,
0.326904296875,
0.69140625
],
[
0.0849609375,
0.6962890625,
0.4794921875,
0.6962890625,
0.4794921875,
0.7109375,
0.0849609375,
0.7109375
],
[
0.521484375,
0.69580078125,
0.91015625,
0.69580078125,
0.91015625,
0.70556640625,
0.521484375,
0.70556640625
],
[
0.0849609375,
0.7109375,
0.303955078125,
0.7109375,
0.303955078125,
0.724609375,
0.0849609375,
0.724609375
],
[
0.51953125,
0.70556640625,
0.9111328125,
0.70849609375,
0.9111328125,
0.72265625,
0.51953125,
0.72021484375
],
[
0.51953125,
0.7216796875,
0.91259765625,
0.7216796875,
0.91259765625,
0.73486328125,
0.51953125,
0.73486328125
],
[
0.0849609375,
0.72998046875,
0.47802734375,
0.72998046875,
0.47802734375,
0.744140625,
0.0849609375,
0.744140625
],
[
0.51953125,
0.73486328125,
0.91259765625,
0.73486328125,
0.91259765625,
0.74853515625,
0.51953125,
0.74853515625
],
[
0.08660888671875,
0.744140625,
0.476318359375,
0.744140625,
0.476318359375,
0.75830078125,
0.08660888671875,
0.75830078125
],
[
0.5185546875,
0.74853515625,
0.9140625,
0.75,
0.9140625,
0.7646484375,
0.5185546875,
0.76318359375
],
[
0.0849609375,
0.7587890625,
0.4794921875,
0.7587890625,
0.4794921875,
0.77294921875,
0.0849609375,
0.77294921875
],
[
0.51953125,
0.7646484375,
0.65185546875,
0.7646484375,
0.65185546875,
0.77783203125,
0.51953125,
0.77783203125
],
[
0.08660888671875,
0.77294921875,
0.47802734375,
0.77294921875,
0.47802734375,
0.7861328125,
0.08660888671875,
0.7861328125
],
[
0.5361328125,
0.77880859375,
0.9140625,
0.77880859375,
0.9140625,
0.79248046875,
0.5361328125,
0.79248046875
],
[
0.0849609375,
0.78857421875,
0.476318359375,
0.78857421875,
0.476318359375,
0.798828125,
0.0849609375,
0.798828125
],
[
0.51953125,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.08660888671875,
0.8017578125,
0.38232421875,
0.8017578125,
0.38232421875,
0.8125,
0.08660888671875,
0.8125
],
[
0.51953125,
0.80615234375,
0.9140625,
0.80615234375,
0.9140625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.08660888671875,
0.8193359375,
0.4794921875,
0.8193359375,
0.4794921875,
0.83349609375,
0.08660888671875,
0.83349609375
],
[
0.51953125,
0.8193359375,
0.9140625,
0.818359375,
0.9140625,
0.83349609375,
0.51953125,
0.833984375
],
[
0.08660888671875,
0.833984375,
0.47802734375,
0.833984375,
0.47802734375,
0.84765625,
0.08660888671875,
0.84765625
],
[
0.51953125,
0.833984375,
0.91015625,
0.833984375,
0.91015625,
0.84765625,
0.51953125,
0.84765625
],
[
0.08660888671875,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.51953125,
0.84765625,
0.9140625,
0.84765625,
0.9140625,
0.861328125,
0.51953125,
0.861328125
],
[
0.08660888671875,
0.8623046875,
0.4794921875,
0.8623046875,
0.4794921875,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51953125,
0.861328125,
0.91259765625,
0.8623046875,
0.91259765625,
0.87548828125,
0.51953125,
0.875
],
[
0.0849609375,
0.87548828125,
0.291748046875,
0.87548828125,
0.291748046875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51953125,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"dence predicted by teacher network as the ground truth. The",
"4.2 Comparison with State-of-the-art Methods",
"loss is defined as",
"M",
"Light =",
"L-(B(ci -α))",
"NYU",
"ICVL",
"Methods",
"(4)",
"2",
"10.54",
"6.93",
"where β is the scaling factor set to 100, ; is the confidence",
"HandPointNet (Ge et al. 2018)",
"DenseReg (Wan et al. 2018)",
"10.21",
"7.24",
"ground truth from teacher network, i.e.the multi-head atten-",
"P2P (Ge, Ren, and Yuan 2018)",
"9.05",
"6.33",
"tion network from M views, and c, is the student network",
"8.61",
"6.46",
"output.",
"A2J (Xiong et al. 2019)",
"V2V (Moon, Chang, and Lee 2018)",
"8.42",
"6.28",
"7.37",
"Once the training is done, the teacher network is no longer",
"AWR (Huang et al. 2020)",
"5.98",
"7.34",
"needed during the inference,and student network can sup-",
"Ours-lview",
"5.16",
"6.82",
"port the view selection and fusion in an efficient way.",
"Ours-3views",
"4.86",
"3.4 Implementation Details",
"Ours-9views",
"6.53",
"4.77",
"We train and evaluate our models on a workstation with two",
"Ours-15views",
"6.41",
"4.76",
"6.40",
"4.79",
"Ours-25views",
"IntelXeonSilver4210R.512GBofRAMandanNvidia",
"RTX3090 GPU. Our models are implemented within Py-",
"Torch. Adam optimizer is used; the initial learning rate is",
"Table 1: Comparison mean joint 3D error (mm) and ranking",
"set to 0.001 and is decayed by 0.9 per epoch. In order to",
"result with state-of-art methods on NYU dataset and ICVL",
"conduct data augmentation, we randomly scale the cropped",
"dataset.\"Ours-lview\",\"Ours-3views\",“Ours-9views\"and",
"depth map, jitter the centroid of the point cloud, and ran-",
"\"Ours-15views\" are the results of our method with se-",
"domlyrotatethe camera whenrenderingmulti-viewdepth",
"lected 1, 3, 9 and 15 views from 25 uniformly sampled",
"For all smoothL1 loss, the switch point between quadratic",
"views, respectively.\"Outs-25views\" denotes the results of",
"and linear is set to 1.0.",
"our method with 25 uniformly sampled views.",
"Our network consists of the3D pose estimation net",
"work (i.e.A2J),the teacher confidencenetwork andthe",
"lightweight student confidencenetwork.Thenetwork input",
"is 176 × 176 hand region cropped from the input depth, and",
"",
"",
"HandPointNe",
"poe",
"we use ResNet-50 as the backbone of A2J. We first train the",
"14",
"10.Py",
"Polint-to-Poh",
"3D pose estimation network and the teacher confidence net-",
"1",
"1o",
"A21",
"worktogether,theloss canbeformulated as:",
"Ours-1vie",
"Lviewseli =LA2 +LJ",
"(5)",
"Ours-9v",
"where=O.1isthefactortobalancethelossterms",
"",
"",
"FesancetoTm",
"",
"Then we fix the parameters of the two networks and train",
"the lightweight student confidence network with Llight.",
"Figure 6: Comparison of our proposed method with state-",
"4.Experiments",
"of-the-artmethods on NYU dataset.Left: mean joint error",
"4.1 Datasets and Evaluation Metric",
"per hand joint. Right: the percentage of success frames over",
"NYU Hand Pose Dataset (NYU) (Tompson et al. 2014)",
"different error thresholds.",
"contains 72,757frames for training and 8,252frames for",
"testing. 36 hand joints are annotated, but we use only a sub-",
"In this experiment, we use our \"student\" lightweight con-",
"set of 14 hand joints for evaluations following the same eval-",
"fidence network for view selection and pose fusion.",
"uation protocol in (Tompson et al. 2014).",
"We first compare our method with the state-of-the-art",
"ICVL Hand Pose Dataset (ICVL)",
"(Tang et al. 2014) con-",
"methods on NYU dataset and ICVL dataset. DenseReg (Wan",
"",
"tains 331,006 frames for training and 1,596 frames for test-",
"etal.2018).A2J(Xiongetal.2019)and AWR(Huangetal",
"ing. 16 hand joints are annotated.",
"2020) directly use the depth map for pose estimation.Hand-",
"PointNet (Ge et al. 2018), P2P (Ge, Ren, and Yuan 2018)",
"Task 1 of Hands19 Challenge Dataset (Hands19-Task1)",
"and V2V (Moon, Chang, and Lee 2018)(Hruz, Kanis, and",
"(Armagan et al. 2020) contains 175,951 training depth im-",
"Krnoul 2021) use the 3D representation of the depth map",
"ages from 5 subjects and 124,999 testing depth images from",
"for pose estimation.",
"10 subjects, in which 5 subjects overlap with the training",
"Table 1 shows the mean joint error. Fig.6 shows the",
"set.This dataset is very challenging because of its exhaus-",
"percentage of success frames over different error thresh-",
"tive coverage of viewpoints and articulations.",
"olds and the error of each joint. We do not show AWR in",
"Evaluation MetricWe evaluate the hand pose estimation",
"Fig. 6 because the best prediction results on NYU dataset",
"performance using standard metrics proposed in (Tang et al.",
"are not released by AWR. Note that with just I view se-",
"2014), i.e. mean joint error and the percentage of test exam-",
"lected, we already outperform all the other methods. And",
"ples that have all predicted joints within a given maximum",
"when more view selected, theperformancekeeps improving.",
"distance from the ground truth.",
"This clearly indicates that our view selection is effective in"
] |
{
"arxiv_id": "2203.15458",
"language": "en",
"page_idx": 4
}
|
||
2108.09355-en-0
|
2108.09355-en-0.jpg
|
$હેં$
# $\textbf{One Chatbot Per Person: Creating Personalized Chatbots based}$ on Implicit User Profiles
Zhengyi Ma$^{2,3}$, Zhicheng Dou$^{1,3}$, Yutao Zhu$^5$, Hanxun Zhong$^{2,3}$, and Ji-Rong Wen$^{3,4}$
$^{1}$Gaoling School of Artificial Intelligence, Renmin University of China
$^{2}$School of Information, Renmin University of China
$^3$ Beijing Key Laboratory of Big Data Management and Analysis Methods
$^{4}$ Key Laboratory of Data Engineering and Knowledge Engineering, MOE
$^{5}$Université de Montréal, Montréal, Québec, Canada
$\{$zymaa,dou,hanxun\_zhong$\}$@ruc.edu.cn,yutao.zhu@umontreal.ca,jirong.wen@gmail.com
$\mathbf{ABSTRACT}$
ACM Reference Format:
Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. 2021. One Chatbot Per Person: Creating Personalized Chatbots based on Implicit User Profiles . In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '21), $\tilde{f}uly$ 11-15, 2021, Virtual Event, Canada. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3404835.3462828
## 1 INTRODUCTION
Personalized chatbots focus on endowing chatbots with a consistent personality to behave like real users, give more informative responses, and further act as personal assistants. Existing personalized approaches tried to incorporate several text descriptions as explicit user profiles. However, the acquisition of such explicit profiles is expensive and time-consuming, thus being impractical for large-scale real-world applications. Moreover, the restricted
predefined profile neglects the language behavior of a real user and cannot be automatically updated together with the change of user interests. In this paper, we propose to learn implicit user
profiles automatically from large-scale user dialogue history for $\hat{\text{building personalized chatbots. Specifically, leveraging the benefits}}$ of Transformer on language understanding, we train a personalized language model to construct a general user profile from the user's
$\sum_{n=1}^{\text{angugenouer inouer io constiuct a gentera user pronses}}{\text{historical responses. To highlight the relevant historical responses}}{\text{to the input post, we further establish a key-value memory network}}$
of historical post-response pairs, and build a dynamic post-aware user profile. The dynamic profile mainly describes what and how the user has responded to similar posts in history. To explicitly utilize users' frequently used words, we design a personalized decoder to fuse two decoding strategies, including generating a word from the generic vocabulary and copying one word from the user's personalized vocabulary. Experiments on two real-world datasets show the significant improvement of our model compared with existing methods. Our code is available at https://github.com/zhengyima/DHAP
Faced with extensive information available on the Internet, it is very appealing to have an intelligent assistant that can provide the most relevant information [16,46,47], collaborate with us on dayto-day problems [14,45], or even act as our agent for some specific tasks [37]. Towards this ultimate goal, in the dialogue system area, building digital agents has attracted more and more attention and had some preliminary applications in our daily life [14,24,45]. In the future, some chit-chat between humans will inevitably be completed by digital agents. In this paper, we carry out a preliminary study toward this goal and focus on the problem of developing personalized chatbots. A personalized chatbot aims at leveraging personalized information $(e.g.$, a predefined persona [13,41]) to provide personalized responses when communicating with others. Such personalized information can help chatbots generate more consistent and informative replies. More importantly, if the personalized information can be well captured, the chatbots can perform similar behaviors like real users $(e.g.$, serving as an agent of the user and give similar responses to others when the user is busy, as shown in Figure 1), thus having the potential to be an intelligent agent with that specific personality [37].
Many models have been proposed for improving the chatbot's capability to generate personalized responses. Early studies tried to integrate the user ID embeddings to a sequence-to-sequence (Seq2Seq) model for identifying the user and generating user-related responses [3,4,9]. Recently, some studies proposed to assign predefined personas to chatbots so as to generate more personalized responses [23,31,41]. They assumed that personality can be described in several sentences or attributes and help chatbots generate more persona-related responses. Different from existing studies, in this work, we propose letting the chatbot learn the implicit user profile automatically from the user dialogue history and generate personalized responses based on the learned user profile. In this way, the chatbot can be personalized by user's (e.g., user A) historical data, behave like this user, act as this user's agent, and chat with any other users $(e.g.$,user B, C, etc.).
$\csc\csc$coNCEPTS
$\cdot$Computing methodologies$\to$Discourse, dialogue and prag-
matics; Natural language generation.
$\mathbf{KEYWORDS}$
Personalized Chatbots; Response Generation; Implicit User Profile
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
SIGIR '21, fuly 11-15, 2021, Virtual Event, Canada
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978- 1- 4503- 8037- 9/ 21/ 07$\ldots \$$15. 00
https://doi.org/10.1145/3404835.3462828
|
[
[
0.09478759765625,
0.1029052734375,
0.90283203125,
0.1029052734375,
0.90283203125,
0.12371826171875,
0.09478759765625,
0.12371826171875
],
[
0.34228515625,
0.1287841796875,
0.65625,
0.1275634765625,
0.65625,
0.1483154296875,
0.34228515625,
0.149658203125
],
[
0.1650390625,
0.1634521484375,
0.8349609375,
0.1610107421875,
0.8349609375,
0.1817626953125,
0.1650390625,
0.184326171875
],
[
0.268798828125,
0.184326171875,
0.736328125,
0.184326171875,
0.736328125,
0.1976318359375,
0.268798828125,
0.1976318359375
],
[
0.32275390625,
0.1962890625,
0.67724609375,
0.1988525390625,
0.67724609375,
0.21337890625,
0.32275390625,
0.2108154296875
],
[
0.25244140625,
0.210205078125,
0.74609375,
0.2108154296875,
0.74609375,
0.2291259765625,
0.25244140625,
0.2279052734375
],
[
0.253173828125,
0.2259521484375,
0.74658203125,
0.2266845703125,
0.74658203125,
0.2442626953125,
0.253173828125,
0.24365234375
],
[
0.324462890625,
0.242431640625,
0.6748046875,
0.24365234375,
0.6748046875,
0.25830078125,
0.324462890625,
0.2568359375
],
[
0.1993408203125,
0.260009765625,
0.7998046875,
0.260009765625,
0.7998046875,
0.2734375,
0.1993408203125,
0.2734375
],
[
0.0236968994140625,
0.27587890625,
0.05963134765625,
0.27587890625,
0.05963134765625,
0.333251953125,
0.0236968994140625,
0.333251953125
],
[
0.08905029296875,
0.285400390625,
0.1829833984375,
0.285400390625,
0.1829833984375,
0.296630859375,
0.08905029296875,
0.296630859375
],
[
0.51708984375,
0.285400390625,
0.66259765625,
0.285400390625,
0.66259765625,
0.298583984375,
0.51708984375,
0.298583984375
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.515625,
0.296630859375,
0.9140625,
0.29541015625,
0.9140625,
0.31005859375,
0.515625,
0.311279296875
],
[
0.0849609375,
0.31640625,
0.483642578125,
0.31640625,
0.483642578125,
0.330078125,
0.0849609375,
0.330078125
],
[
0.51708984375,
0.311279296875,
0.9111328125,
0.311279296875,
0.9111328125,
0.3212890625,
0.51708984375,
0.3212890625
],
[
0.02532958984375,
0.3212890625,
0.066162109375,
0.324462890625,
0.05963134765625,
0.37744140625,
0.0187835693359375,
0.375
],
[
0.51708984375,
0.3212890625,
0.9140625,
0.3212890625,
0.9140625,
0.335205078125,
0.51708984375,
0.335205078125
],
[
0.0849609375,
0.330078125,
0.484375,
0.330078125,
0.484375,
0.343505859375,
0.0849609375,
0.343505859375
],
[
0.5185546875,
0.337158203125,
0.91259765625,
0.337158203125,
0.91259765625,
0.347900390625,
0.5185546875,
0.347900390625
],
[
0.08660888671875,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.08660888671875,
0.356689453125
],
[
0.51708984375,
0.347900390625,
0.9140625,
0.348388671875,
0.9140625,
0.3623046875,
0.51708984375,
0.361083984375
],
[
0.0849609375,
0.35791015625,
0.48193359375,
0.35791015625,
0.48193359375,
0.371826171875,
0.0849609375,
0.371826171875
],
[
0.515625,
0.361083984375,
0.73388671875,
0.361083984375,
0.73388671875,
0.375,
0.515625,
0.375
],
[
0.0269622802734375,
0.373779296875,
0.056365966796875,
0.373779296875,
0.056365966796875,
0.39013671875,
0.0269622802734375,
0.39013671875
],
[
0.08660888671875,
0.371826171875,
0.48193359375,
0.371826171875,
0.48193359375,
0.385009765625,
0.08660888671875,
0.385009765625
],
[
0.08660888671875,
0.385009765625,
0.48193359375,
0.385009765625,
0.48193359375,
0.3984375,
0.08660888671875,
0.3984375
],
[
0.5390625,
0.393310546875,
0.68701171875,
0.393310546875,
0.68701171875,
0.4072265625,
0.5390625,
0.4072265625
],
[
0.0269622802734375,
0.399658203125,
0.062103271484375,
0.40283203125,
0.0555419921875,
0.448974609375,
0.0196075439453125,
0.44580078125
],
[
0.0849609375,
0.399658203125,
0.48046875,
0.399658203125,
0.48046875,
0.41357421875,
0.0849609375,
0.41357421875
],
[
0.08660888671875,
0.4140625,
0.48046875,
0.4140625,
0.48046875,
0.4248046875,
0.08660888671875,
0.4248046875
],
[
0.5185546875,
0.41357421875,
0.9111328125,
0.41357421875,
0.9111328125,
0.423583984375,
0.5185546875,
0.423583984375
],
[
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
],
[
0.0310516357421875,
0.44140625,
0.058013916015625,
0.44140625,
0.058013916015625,
0.474853515625,
0.0310516357421875,
0.474853515625
],
[
0.0882568359375,
0.442626953125,
0.48046875,
0.442626953125,
0.48046875,
0.45263671875,
0.0882568359375,
0.45263671875
],
[
0.5185546875,
0.44140625,
0.91259765625,
0.44140625,
0.91259765625,
0.451904296875,
0.5185546875,
0.451904296875
],
[
0.0849609375,
0.455078125,
0.48193359375,
0.455078125,
0.48193359375,
0.468505859375,
0.0849609375,
0.468505859375
],
[
0.51708984375,
0.45263671875,
0.9140625,
0.453857421875,
0.9140625,
0.46728515625,
0.51708984375,
0.466552734375
],
[
0.08660888671875,
0.468505859375,
0.483642578125,
0.468505859375,
0.483642578125,
0.481689453125,
0.08660888671875,
0.481689453125
],
[
0.51708984375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.08660888671875,
0.484130859375,
0.48046875,
0.484130859375,
0.48046875,
0.494384765625,
0.08660888671875,
0.494384765625
],
[
0.51708984375,
0.481689453125,
0.9140625,
0.481689453125,
0.9140625,
0.49560546875,
0.51708984375,
0.49560546875
],
[
0.02777099609375,
0.4892578125,
0.0555419921875,
0.4892578125,
0.0555419921875,
0.505859375,
0.02777099609375,
0.505859375
],
[
0.0228729248046875,
0.49755859375,
0.056365966796875,
0.49755859375,
0.056365966796875,
0.5341796875,
0.0228729248046875,
0.5341796875
],
[
0.08660888671875,
0.49755859375,
0.48046875,
0.49755859375,
0.48046875,
0.50830078125,
0.08660888671875,
0.50830078125
],
[
0.51708984375,
0.49560546875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.51708984375,
0.5087890625
],
[
0.0849609375,
0.51025390625,
0.48193359375,
0.51025390625,
0.48193359375,
0.5234375,
0.0849609375,
0.5234375
],
[
0.51953125,
0.51123046875,
0.91259765625,
0.51123046875,
0.91259765625,
0.521484375,
0.51953125,
0.521484375
],
[
0.0849609375,
0.5234375,
0.48193359375,
0.5234375,
0.48193359375,
0.537109375,
0.0849609375,
0.537109375
],
[
0.5185546875,
0.5234375,
0.9140625,
0.5234375,
0.9140625,
0.537109375,
0.5185546875,
0.537109375
],
[
0.02532958984375,
0.537109375,
0.056365966796875,
0.537109375,
0.056365966796875,
0.58837890625,
0.02532958984375,
0.58837890625
],
[
0.0269622802734375,
0.52587890625,
0.0555419921875,
0.52587890625,
0.0555419921875,
0.54443359375,
0.0269622802734375,
0.54443359375
],
[
0.08660888671875,
0.5390625,
0.4794921875,
0.5390625,
0.4794921875,
0.5498046875,
0.08660888671875,
0.5498046875
],
[
0.51708984375,
0.537109375,
0.91259765625,
0.537109375,
0.91259765625,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.0849609375,
0.55029296875,
0.48193359375,
0.55029296875,
0.48193359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.51708984375,
0.5498046875,
0.91259765625,
0.55029296875,
0.91259765625,
0.56494140625,
0.51708984375,
0.564453125
],
[
0.0849609375,
0.56494140625,
0.48193359375,
0.56494140625,
0.48193359375,
0.5791015625,
0.0849609375,
0.5791015625
],
[
0.51708984375,
0.56494140625,
0.91259765625,
0.56494140625,
0.91259765625,
0.5791015625,
0.51708984375,
0.5791015625
],
[
0.041656494140625,
0.57763671875,
0.047393798828125,
0.57763671875,
0.047393798828125,
0.58349609375,
0.041656494140625,
0.58349609375
],
[
0.0849609375,
0.57763671875,
0.48046875,
0.57763671875,
0.48046875,
0.59130859375,
0.0849609375,
0.59130859375
],
[
0.51708984375,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.51708984375,
0.59228515625
],
[
0.08660888671875,
0.59228515625,
0.483642578125,
0.59228515625,
0.483642578125,
0.60595703125,
0.08660888671875,
0.60595703125
],
[
0.02532958984375,
0.58056640625,
0.056365966796875,
0.58056640625,
0.056365966796875,
0.6171875,
0.02532958984375,
0.6171875
],
[
0.5185546875,
0.59326171875,
0.91259765625,
0.59326171875,
0.91259765625,
0.60400390625,
0.5185546875,
0.60400390625
],
[
0.0228729248046875,
0.60498046875,
0.060455322265625,
0.60498046875,
0.060455322265625,
0.70703125,
0.0228729248046875,
0.70703125
],
[
0.0849609375,
0.60693359375,
0.48193359375,
0.60693359375,
0.48193359375,
0.62060546875,
0.0849609375,
0.62060546875
],
[
0.51708984375,
0.60595703125,
0.9140625,
0.60595703125,
0.9140625,
0.619140625,
0.51708984375,
0.619140625
],
[
0.08660888671875,
0.6220703125,
0.48046875,
0.62060546875,
0.48046875,
0.6318359375,
0.08660888671875,
0.63330078125
],
[
0.51953125,
0.6220703125,
0.91259765625,
0.6220703125,
0.91259765625,
0.6318359375,
0.51953125,
0.6318359375
],
[
0.0849609375,
0.6337890625,
0.48046875,
0.6337890625,
0.48046875,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.51708984375,
0.6337890625,
0.9140625,
0.6337890625,
0.9140625,
0.64794921875,
0.51708984375,
0.64794921875
],
[
0.5185546875,
0.64794921875,
0.9140625,
0.64794921875,
0.9140625,
0.6611328125,
0.5185546875,
0.6611328125
],
[
0.0882568359375,
0.66357421875,
0.221435546875,
0.66357421875,
0.221435546875,
0.6767578125,
0.0882568359375,
0.6767578125
],
[
0.51953125,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.673828125,
0.51953125,
0.673828125
],
[
0.08660888671875,
0.68115234375,
0.483642578125,
0.68115234375,
0.483642578125,
0.69580078125,
0.08660888671875,
0.69580078125
],
[
0.515625,
0.67529296875,
0.76123046875,
0.6748046875,
0.76123046875,
0.68798828125,
0.515625,
0.689453125
],
[
0.08416748046875,
0.6943359375,
0.330810546875,
0.69580078125,
0.330810546875,
0.71044921875,
0.08416748046875,
0.708984375
],
[
0.53173828125,
0.689453125,
0.91259765625,
0.689453125,
0.91259765625,
0.70263671875,
0.53173828125,
0.70263671875
],
[
0.5185546875,
0.705078125,
0.9111328125,
0.705078125,
0.9111328125,
0.71533203125,
0.5185546875,
0.71533203125
],
[
0.51708984375,
0.71728515625,
0.9111328125,
0.71826171875,
0.9111328125,
0.72998046875,
0.51708984375,
0.72900390625
],
[
0.08660888671875,
0.72607421875,
0.188720703125,
0.72607421875,
0.188720703125,
0.7373046875,
0.08660888671875,
0.7373046875
],
[
0.51708984375,
0.73095703125,
0.9140625,
0.73095703125,
0.9140625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.0849609375,
0.7412109375,
0.4794921875,
0.7412109375,
0.4794921875,
0.7548828125,
0.0849609375,
0.7548828125
],
[
0.5185546875,
0.74560546875,
0.91259765625,
0.74560546875,
0.91259765625,
0.755859375,
0.5185546875,
0.755859375
],
[
0.5185546875,
0.76025390625,
0.91259765625,
0.76025390625,
0.91259765625,
0.77099609375,
0.5185546875,
0.77099609375
],
[
0.5185546875,
0.7734375,
0.91259765625,
0.7734375,
0.91259765625,
0.7841796875,
0.5185546875,
0.7841796875
],
[
0.0849609375,
0.78076171875,
0.48193359375,
0.78076171875,
0.48193359375,
0.79443359375,
0.0849609375,
0.79443359375
],
[
0.5185546875,
0.7861328125,
0.91259765625,
0.7861328125,
0.91259765625,
0.796875,
0.5185546875,
0.796875
],
[
0.136474609375,
0.79248046875,
0.34228515625,
0.79248046875,
0.34228515625,
0.80322265625,
0.136474609375,
0.80322265625
],
[
0.34716796875,
0.79345703125,
0.41259765625,
0.79150390625,
0.41259765625,
0.8017578125,
0.34716796875,
0.8037109375
],
[
0.419189453125,
0.79248046875,
0.48046875,
0.79248046875,
0.48046875,
0.80322265625,
0.419189453125,
0.80322265625
],
[
0.08660888671875,
0.80322265625,
0.48193359375,
0.80322265625,
0.48193359375,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.096435546875,
0.81103515625,
0.48193359375,
0.81005859375,
0.48193359375,
0.82373046875,
0.096435546875,
0.82470703125
],
[
0.5146484375,
0.81298828125,
0.91259765625,
0.8125,
0.91259765625,
0.8271484375,
0.5146484375,
0.82763671875
],
[
0.0849609375,
0.8212890625,
0.48046875,
0.82275390625,
0.48046875,
0.833984375,
0.0849609375,
0.83349609375
],
[
0.08660888671875,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.84326171875,
0.08660888671875,
0.84326171875
],
[
0.5185546875,
0.82763671875,
0.9140625,
0.82763671875,
0.9140625,
0.84130859375,
0.5185546875,
0.84130859375
],
[
0.0882568359375,
0.84326171875,
0.306396484375,
0.84326171875,
0.306396484375,
0.85400390625,
0.0882568359375,
0.85400390625
],
[
0.301513671875,
0.84326171875,
0.37744140625,
0.84326171875,
0.37744140625,
0.85400390625,
0.301513671875,
0.85400390625
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85302734375,
0.31787109375,
0.8515625,
0.31787109375,
0.86279296875,
0.0849609375,
0.8642578125
],
[
0.515625,
0.85400390625,
0.9140625,
0.85498046875,
0.9140625,
0.869140625,
0.515625,
0.86865234375
],
[
0.08416748046875,
0.8642578125,
0.47216796875,
0.86279296875,
0.47216796875,
0.875,
0.08416748046875,
0.87548828125
],
[
0.08660888671875,
0.875,
0.295654296875,
0.875,
0.295654296875,
0.88525390625,
0.08660888671875,
0.88525390625
],
[
0.51708984375,
0.869140625,
0.9140625,
0.869140625,
0.9140625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.0849609375,
0.88525390625,
0.27294921875,
0.88525390625,
0.27294921875,
0.89599609375,
0.0849609375,
0.89599609375
],
[
0.5146484375,
0.8818359375,
0.79736328125,
0.88330078125,
0.79736328125,
0.89794921875,
0.5146484375,
0.896484375
]
] |
[
"One Chatbot Per Person: Creating Personalized Chatbots based",
"on Implicit User Profiles",
"Zhengyi Ma2.3, Zhicheng Doul.3, Yutao Zhu5, Hanxun Zhong23, and Ji-Rong Wen3.4",
"Gaoling School of Artificial Intelligence, Renmin University of China",
"2 School of Information, Renmin University of China",
"3 Beijing Key Laboratory of Big Data Management and Analysis Methods",
"4 Key Laboratory of Data Engineering and Knowledge Engineering, MOE",
"5 Universite de Montreal, Montreal, Quebec, Canada",
"(zymaa,dou,hanxun_zhong)@ruc.edu.cn,yutao.zhu@umontreal.ca,jirong.wen@gmail.com",
"12",
"ABSTRACT",
"ACMReference Format:",
"Personalized chatbots focus on endowing chatbots with a consis-",
"Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen.",
"tent personality to behave like real users, give more informative",
"2021.One Chatbot PerPerson:CreatingPersonalized Chatbots based on",
"des",
"Implicit User Profiles . In Proceedings of the 44th International ACM SIGIR",
"responses, and further act as personal assistants. Existing person-",
"Conference on Research and Development in Information Retrieval (SIGIR'21)",
"alized approaches tried to incorporate several text descriptions",
"July1115,2021,Virtual Event,Canada.ACM,NewYork,NY,USA,10 pages.",
"as explicit user profiles. However, the acquisition of such explicit",
"https:/doi.org/10.1145/3404835.3462828",
"2",
"profiles is expensive and time-consuming, thus being impractical",
"for large-scale real-world applications. Moreover, the restricted",
"INTRODUCTION",
"[T'",
"predefined profile neglects the language behavior of a real user",
"and cannot be automatically updated together with the change",
"Faced with extensive information available on the Internet,it is",
"of user interests. In this paper, we propose to learn implicit user",
"very appealing to have an intelligent assistant that can provide the",
"SO",
"profilesautomaticallyfromlarge-scaleuserdialoguehistoryfor",
"mostrelevant information[16,46,47],collaboratewith us on day-",
"building personalized chatbots. Specifically,leveraging the benefits",
"to-day problems [14, 45], or even act as our agent for some specific",
"of Transformer on language understanding, we train a personalized",
"tasks [37]. Towards this ultimate goal, in the dialogue system area,",
"languagemodel toconstructageneral userprofilefrom theuser's",
"building digital agents has attracted more and more attention and",
"2",
"5",
"historical responses.To highlight the relevant historical responses",
"had some preliminary applications in our daily life [14, 24, 45]. In",
"to the input post, we further establish a key-value memory network",
"thefuture,somechit-chatbetweenhumanswill inevitablybecom-",
"of historical post-response pairs, and build a dynamic post-aware",
"pleted by digital agents. In this paper, we carry out a preliminary",
"660",
"5",
"userprofile.Thedynamicprofilemainlydescribes what and howthe",
"study toward this goal and focus on the problem of developing",
"user has responded to similar posts in history. To explicitly utilize",
"personalized chatbots. A personalized chatbot aims at leveraging",
"users' frequently used words, we design a personalized decoder to",
"personalized information (e.g., a predefined persona [13, 41]) to",
"",
"fuse two decoding strategies, including generating a word from the",
"provide personalized responses when communicating with others.",
"generic vocabulary and copying one word from the user's personal-",
"801",
"Suchpersonalized information can help chatbotsgeneratemore",
"ZAI",
"ized vocabulary. Experiments on two real-world datasets show the",
"consistent and informative replies. More importantly,if the person-",
"significantimprovementofourmodelcomparedwithexistingmeth",
"alizedinformation canbewell captured,thechatbotscanperform",
"ods. Our code is available at https://github.com/zhengyima/DHAP",
"similar behaviors like real users (e.g., serving as an agent of the",
"user and give similar responses to others when the user is busy, as",
"CCS CONCEPTS",
"showninFigure1),thushavingthepotential tobeanintelligent",
". Computing methodologies -→ Discourse, dialogue and prag-",
"agent with that specific personality [37]",
"matics; Natural language generation.",
"Many models have been proposed for improving the chatbot's",
"capabilitytogeneratepersonalizedresponses.Earlystudiestried",
"to integrate the user ID embeddings to a sequence-to-sequence",
"KEYWORDS",
"(Seq2Seq) model for identifying the user and generating user-related",
"Personalized Chatbots; Response Generation; Implicit User Profile",
"responses[3,4.9].Recently.somestudiesproposedtoassignpre-",
"definedpersonastochatbots soastogeneratemorepersonalized",
"responses [23, 31, 41].They assumed that personality can be de-",
"Permission to make digital or hard copies of all or part of this work for personal or",
"scribed in several sentences or attributes and help chatbots generate",
"use is granted without fee provided that copi",
"s are not made",
"or distributed",
"for profit or commercial advantage and that copies bear this notice and the full citation",
"more persona-related responses. Different from existing studies, in",
"thefirst page.Copyrights for components ofthis work owned by others than the",
"this work, we propose letting the chatbot learn the implicit user",
"author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or",
"republish, to post on servers or to redistribute to lists, requires prior specific permission",
"profile automatically from the user dialogue history and gen-",
"and/ora fee.Requestpermissionsfrompermis:",
"ssions@acm.org",
"erate personalized responses based on the learned user pro-",
"SIGIR21, July 11-15,2021,Virtual Event, Canada",
"file. In this way, the chatbot can be personalized by user's (e.g., user",
" 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.",
"ACM ISBN 978-1-4503-8037-9/21/07..-$15.00",
"A) historical data, behave like this user, act as this user's agent, and",
"https://doi.org/10.1145/3404835.3462828",
"chat with any other users (e.g., user B, C, etc.)."
] |
{
"arxiv_id": "2108.09355",
"language": "en",
"page_idx": 0
}
|
||
2108.09355-en-1
|
2108.09355-en-1.jpg
|
<FigureHere>
Figure 1: An example of the personalized chatbot serving as
an intelligent agent for user Tom when Tom is busy.
Our idea is motivated by: (1) Contrary to explicit persona descriptions, user dialogue history is easier to be gathered on the user's client devices. It is evident that obtaining explicit descriptions for massive users is impractical in real applications [4]: First, users may be lazy to set their profile before using the chatbot [6]. Second, manually collecting user profiles is costly and time-consuming. Furthermore, even if the user profile is collected, it cannot be updated with the change of user interests, thus may be ineffective over time. Finally, a fixed set of properties is not suitable for describing all users. (2) The user dialogue history contains massive personalized information, which is suitable for learning user profiles automatically. As shown in Figure 1, the dialogue history of a user includes their historical responses, and the corresponding posts issued by other users. Intuitively, users' historical responses can often reflect their language style, background knowledge, frequently used words, and even their interests. For example, there may be many electronic device names that appeared in the historical responses of an electronic hobbyist $(e.g,MacBook$ in Figure 1). Besides, the interaction content and style between a specific user and others can be captured from the historical post-response pairs. When faced with a new input post, the chatbot can look for the historical data, check how the user has responded to a similar post before, and apply similar interactions to generate a suitable response. In addition, user profiles learned from historical data can be gradually updated with more data being collected. In summary, the user dialogue history is easy to obtain and appropriate for building user profile.
To achieve our idea, we propose a model DHAP for personalized chatbots, which focuses on learning implicit user profile from user Dialogue History Automatically and generating Personalized responses. In our model, a general user profile representation is firstly constructed from the user's historical responses to capture the general information, including user interest, background knowledge, and speaking style. This is implemented by a personalized language model based on Transformer. Then, we design a personalized post encoder to construct the personalized post representation. The general user profile is utilized in the post encoder to better capture the semantic information of the input post. Next, we build a key-value memory network to store the user's historical post-response pairs. Based on this history memory, the dynamic post-aware user profile is built by highlighting the historical responses relevant to the
current input post. Finally, we design a personalized decoder to fuse the learned user profile into the response generation process The personalized decoder can switch between generating a word from a generic vocabulary and copying a word from the user's personalized vocabulary. Experimental results on two large-scale $\dot{\text{datasets show that our proposed DHAP significantly outperforms}}$ existing response generation models in various evaluation metrics.
Our contributions are three-fold: (1) We learn the implicit user profile from user's dialogue history automatically for generating personalized responses. By this means, our method can be applied without additional annotations on user profiles, and create personalized chatbots as user's digital agents. (2) We build two kinds of user profiles from the dialogue history, including the general user profile reflecting the user's general information and the dynamic post-aware user profile to apply similar interactions in the historical data for the current input. (3) We design a personalized decoder to coordinate two personalized decoding strategies, including generating a word from the generic vocabulary and copying a word from the personalized vocabulary to leverage the user's word preference.
# 2 RELATED WORK
Open-domain Chatbots and Response Generation. Open-domain chatbots have attracted more and more attention, due to their broad application in real applications, such as Microsoft XiaoIce [45]. Typical methods can be categorized into two groups: retrieval-based and generation-based. Retrieval-based methods aim to select a suitable response from a large repository [38,49,50], while generationbased methods aim at generating a response from scratch [26,28, 32, 48]. In this study, we focus on the response generation problem. Some early studies treat the response generation task as a statistical machine translation problem because of its end-to-end and data-driven features [25,34]. More recently, with the progress of deep learning, Seq2Seq methods have been applied to response generation and achieve great performance [8,26,27,34]. Many Seq2Seq-based extensions have been applied to tackle the“safe response" problem [8]; to incorporate external knowledge [44]; to generate responses with emotions or personas [9,23,43]; and to model the hierarchical structure of the dialogue context [26,27].
Personalized Chatbots. Endowing chatbots with a coherent personality is a challenging but necessary step to achieve the ultimate goal of building an intelligent assistant. With personality, a chatbot can generate more informative and user-specific responses [9,41], and has the potential to perform similar behaviors as real humans.
Traditional personalized chatbots focus on modeling the user's psychological behavior such as the “Big Five” of speakers [17]. In recent years, deep-learning based methods were proposed to learn the persona information directly from large-scale dialogue datasets via end-to-end neural networks [4,9,23,41]. Some researchers first tried to input the user ID embeddings into the decoder of a Seq2Seq model to generate more personalized responses [1,3,4,9]. Despite users can be identified by their IDs, the personalized performance is limited because no user-related information is used in the model. Therefore, another group of researchers proposed assigning explicit profiles (personas) for chatbots to generate personalized responses. For example, Zhang et al. [41] published the PERSONACHAT dataset, in which each user is assigned with several persona
|
[
[
0.36767578125,
0.10479736328125,
0.396240234375,
0.10479736328125,
0.396240234375,
0.117431640625,
0.36767578125,
0.117431640625
],
[
0.44677734375,
0.10418701171875,
0.473876953125,
0.106689453125,
0.471435546875,
0.11871337890625,
0.443603515625,
0.11553955078125
],
[
0.51708984375,
0.106689453125,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.51708984375,
0.12060546875
],
[
0.378173828125,
0.11236572265625,
0.39794921875,
0.126953125,
0.3798828125,
0.14013671875,
0.361083984375,
0.1256103515625
],
[
0.1446533203125,
0.10986328125,
0.2296142578125,
0.10986328125,
0.2296142578125,
0.121826171875,
0.1446533203125,
0.121826171875
],
[
0.303955078125,
0.11614990234375,
0.348876953125,
0.11614990234375,
0.348876953125,
0.126953125,
0.303955078125,
0.126953125
],
[
0.418212890625,
0.11871337890625,
0.438720703125,
0.11871337890625,
0.438720703125,
0.1256103515625,
0.418212890625,
0.1256103515625
],
[
0.451904296875,
0.11553955078125,
0.4697265625,
0.11553955078125,
0.4697265625,
0.1300048828125,
0.451904296875,
0.1300048828125
],
[
0.449462890625,
0.1275634765625,
0.471435546875,
0.1275634765625,
0.471435546875,
0.14013671875,
0.449462890625,
0.14013671875
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.1356201171875,
0.1287841796875,
0.265625,
0.1300048828125,
0.265625,
0.141357421875,
0.1356201171875,
0.14013671875
],
[
0.370849609375,
0.1319580078125,
0.3896484375,
0.1287841796875,
0.39208984375,
0.1395263671875,
0.373291015625,
0.1427001953125
],
[
0.515625,
0.1351318359375,
0.9140625,
0.1337890625,
0.9140625,
0.147705078125,
0.515625,
0.1483154296875
],
[
0.136474609375,
0.1427001953125,
0.2197265625,
0.1427001953125,
0.2197265625,
0.15283203125,
0.136474609375,
0.15283203125
],
[
0.336669921875,
0.146484375,
0.443603515625,
0.146484375,
0.443603515625,
0.160400390625,
0.336669921875,
0.160400390625
],
[
0.51708984375,
0.149658203125,
0.91259765625,
0.149658203125,
0.91259765625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.111083984375,
0.1590576171875,
0.1217041015625,
0.1673583984375,
0.11358642578125,
0.17236328125,
0.10455322265625,
0.1654052734375
],
[
0.136474609375,
0.1558837890625,
0.253173828125,
0.1558837890625,
0.253173828125,
0.1666259765625,
0.136474609375,
0.1666259765625
],
[
0.336669921875,
0.160400390625,
0.461669921875,
0.160400390625,
0.461669921875,
0.17041015625,
0.336669921875,
0.17041015625
],
[
0.51708984375,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.3095703125,
0.1641845703125,
0.318603515625,
0.1641845703125,
0.318603515625,
0.1685791015625,
0.3095703125,
0.1685791015625
],
[
0.1396484375,
0.169189453125,
0.1519775390625,
0.169189453125,
0.1519775390625,
0.175537109375,
0.1396484375,
0.175537109375
],
[
0.15283203125,
0.1666259765625,
0.268798828125,
0.1666259765625,
0.268798828125,
0.1767578125,
0.15283203125,
0.1767578125
],
[
0.10455322265625,
0.173583984375,
0.12335205078125,
0.173583984375,
0.12335205078125,
0.1817626953125,
0.10455322265625,
0.1817626953125
],
[
0.338134765625,
0.169189453125,
0.463134765625,
0.169189453125,
0.463134765625,
0.179931640625,
0.338134765625,
0.179931640625
],
[
0.1405029296875,
0.1787109375,
0.1502685546875,
0.1787109375,
0.1502685546875,
0.18310546875,
0.1405029296875,
0.18310546875
],
[
0.148681640625,
0.1787109375,
0.1829833984375,
0.1787109375,
0.1829833984375,
0.186279296875,
0.148681640625,
0.186279296875
],
[
0.51708984375,
0.1767578125,
0.9140625,
0.1767578125,
0.9140625,
0.1900634765625,
0.51708984375,
0.1900634765625
],
[
0.336669921875,
0.18310546875,
0.396240234375,
0.18310546875,
0.396240234375,
0.1932373046875,
0.336669921875,
0.1932373046875
],
[
0.336669921875,
0.1962890625,
0.46728515625,
0.1962890625,
0.46728515625,
0.20703125,
0.336669921875,
0.20703125
],
[
0.51708984375,
0.1912841796875,
0.91259765625,
0.1900634765625,
0.91259765625,
0.203857421875,
0.51708984375,
0.2052001953125
],
[
0.1380615234375,
0.199462890625,
0.253173828125,
0.199462890625,
0.253173828125,
0.20703125,
0.1380615234375,
0.20703125
],
[
0.336669921875,
0.20703125,
0.459228515625,
0.20703125,
0.459228515625,
0.2171630859375,
0.336669921875,
0.2171630859375
],
[
0.53173828125,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.53173828125,
0.2183837890625
],
[
0.338134765625,
0.2197265625,
0.420654296875,
0.2197265625,
0.420654296875,
0.2298583984375,
0.338134765625,
0.2298583984375
],
[
0.136474609375,
0.210205078125,
0.261474609375,
0.210205078125,
0.261474609375,
0.2203369140625,
0.136474609375,
0.2203369140625
],
[
0.136474609375,
0.2203369140625,
0.205078125,
0.2203369140625,
0.205078125,
0.2310791015625,
0.136474609375,
0.2310791015625
],
[
0.5185546875,
0.2203369140625,
0.91259765625,
0.2203369140625,
0.91259765625,
0.2310791015625,
0.5185546875,
0.2310791015625
],
[
0.136474609375,
0.23291015625,
0.2646484375,
0.2342529296875,
0.2646484375,
0.24560546875,
0.136474609375,
0.2442626953125
],
[
0.3349609375,
0.23291015625,
0.465576171875,
0.2342529296875,
0.465576171875,
0.24560546875,
0.3349609375,
0.2442626953125
],
[
0.51708984375,
0.23291015625,
0.9140625,
0.23291015625,
0.9140625,
0.246826171875,
0.51708984375,
0.246826171875
],
[
0.1396484375,
0.24560546875,
0.25244140625,
0.24560546875,
0.25244140625,
0.255615234375,
0.1396484375,
0.255615234375
],
[
0.348876953125,
0.24560546875,
0.464111328125,
0.24560546875,
0.464111328125,
0.255615234375,
0.348876953125,
0.255615234375
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.339111328125,
0.2568359375,
0.386474609375,
0.2568359375,
0.386474609375,
0.267578125,
0.339111328125,
0.267578125
],
[
0.51708984375,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.5185546875,
0.27587890625,
0.91259765625,
0.27587890625,
0.91259765625,
0.285888671875,
0.5185546875,
0.285888671875
],
[
0.0849609375,
0.285888671875,
0.4794921875,
0.287353515625,
0.4794921875,
0.300537109375,
0.0849609375,
0.2998046875
],
[
0.5185546875,
0.28857421875,
0.91259765625,
0.28857421875,
0.91259765625,
0.298583984375,
0.5185546875,
0.298583984375
],
[
0.0849609375,
0.2998046875,
0.43310546875,
0.300537109375,
0.43310546875,
0.31494140625,
0.0849609375,
0.314453125
],
[
0.51708984375,
0.3017578125,
0.9140625,
0.3017578125,
0.9140625,
0.31494140625,
0.51708984375,
0.31494140625
],
[
0.51708984375,
0.31640625,
0.9140625,
0.31640625,
0.9140625,
0.330078125,
0.51708984375,
0.330078125
],
[
0.5185546875,
0.330810546875,
0.91259765625,
0.330810546875,
0.91259765625,
0.341552734375,
0.5185546875,
0.341552734375
],
[
0.101318359375,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.101318359375,
0.356689453125
],
[
0.51708984375,
0.343505859375,
0.9140625,
0.343505859375,
0.9140625,
0.356689453125,
0.51708984375,
0.356689453125
],
[
0.08416748046875,
0.35595703125,
0.48193359375,
0.356689453125,
0.48193359375,
0.37060546875,
0.08416748046875,
0.369384765625
],
[
0.5185546875,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.369384765625,
0.5185546875,
0.369384765625
],
[
0.08660888671875,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
0.0849609375,
0.385009765625,
0.483642578125,
0.385009765625,
0.483642578125,
0.3984375,
0.0849609375,
0.3984375
],
[
0.51708984375,
0.389404296875,
0.53125,
0.385009765625,
0.53515625,
0.393310546875,
0.521484375,
0.397705078125
],
[
0.5390625,
0.3837890625,
0.689453125,
0.3837890625,
0.689453125,
0.397705078125,
0.5390625,
0.397705078125
],
[
0.08660888671875,
0.40087890625,
0.48193359375,
0.40087890625,
0.48193359375,
0.410888671875,
0.08660888671875,
0.410888671875
],
[
0.53515625,
0.40283203125,
0.91259765625,
0.40283203125,
0.91259765625,
0.416748046875,
0.53515625,
0.416748046875
],
[
0.0849609375,
0.41357421875,
0.48193359375,
0.412353515625,
0.48193359375,
0.423583984375,
0.0849609375,
0.4248046875
],
[
0.51708984375,
0.417236328125,
0.9140625,
0.417236328125,
0.9140625,
0.43115234375,
0.51708984375,
0.43115234375
],
[
0.0849609375,
0.425537109375,
0.48193359375,
0.425537109375,
0.48193359375,
0.439453125,
0.0849609375,
0.439453125
],
[
0.51708984375,
0.43115234375,
0.91259765625,
0.429931640625,
0.91259765625,
0.4443359375,
0.51708984375,
0.44580078125
],
[
0.08660888671875,
0.44140625,
0.48046875,
0.44140625,
0.48046875,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.5185546875,
0.4462890625,
0.9111328125,
0.4462890625,
0.9111328125,
0.45703125,
0.5185546875,
0.45703125
],
[
0.08660888671875,
0.453857421875,
0.48193359375,
0.453857421875,
0.48193359375,
0.46728515625,
0.08660888671875,
0.46728515625
],
[
0.515625,
0.458984375,
0.9140625,
0.458251953125,
0.9140625,
0.4716796875,
0.515625,
0.472900390625
],
[
0.08660888671875,
0.4697265625,
0.48193359375,
0.4697265625,
0.48193359375,
0.479736328125,
0.08660888671875,
0.479736328125
],
[
0.5185546875,
0.474853515625,
0.9111328125,
0.474853515625,
0.9111328125,
0.48486328125,
0.5185546875,
0.48486328125
],
[
0.08660888671875,
0.48291015625,
0.48193359375,
0.48291015625,
0.48193359375,
0.49365234375,
0.08660888671875,
0.49365234375
],
[
0.5185546875,
0.488037109375,
0.91259765625,
0.488037109375,
0.91259765625,
0.498779296875,
0.5185546875,
0.498779296875
],
[
0.0849609375,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.5185546875,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51220703125,
0.5185546875,
0.51220703125
],
[
0.08660888671875,
0.51123046875,
0.48046875,
0.51123046875,
0.48046875,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.53125,
0.51220703125,
0.9140625,
0.51318359375,
0.9140625,
0.52783203125,
0.53125,
0.5263671875
],
[
0.08660888671875,
0.5244140625,
0.48046875,
0.5244140625,
0.48046875,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.5185546875,
0.52880859375,
0.91259765625,
0.52880859375,
0.91259765625,
0.5390625,
0.5185546875,
0.5390625
],
[
0.08660888671875,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.08660888671875,
0.55029296875
],
[
0.51953125,
0.54345703125,
0.91259765625,
0.54345703125,
0.91259765625,
0.5537109375,
0.51953125,
0.5537109375
],
[
0.08660888671875,
0.55322265625,
0.48046875,
0.55322265625,
0.48046875,
0.56298828125,
0.08660888671875,
0.56298828125
],
[
0.51708984375,
0.5546875,
0.91259765625,
0.5546875,
0.91259765625,
0.568359375,
0.51708984375,
0.568359375
],
[
0.0849609375,
0.56494140625,
0.483642578125,
0.56494140625,
0.483642578125,
0.5791015625,
0.0849609375,
0.5791015625
],
[
0.515625,
0.5693359375,
0.91259765625,
0.568359375,
0.91259765625,
0.58203125,
0.515625,
0.58349609375
],
[
0.0849609375,
0.5791015625,
0.48193359375,
0.5791015625,
0.48193359375,
0.59228515625,
0.0849609375,
0.59228515625
],
[
0.5185546875,
0.58544921875,
0.9111328125,
0.58544921875,
0.9111328125,
0.59521484375,
0.5185546875,
0.59521484375
],
[
0.08660888671875,
0.59326171875,
0.4794921875,
0.59326171875,
0.4794921875,
0.60400390625,
0.08660888671875,
0.60400390625
],
[
0.515625,
0.5966796875,
0.91259765625,
0.59521484375,
0.91259765625,
0.60986328125,
0.515625,
0.611328125
],
[
0.0849609375,
0.60595703125,
0.48193359375,
0.60595703125,
0.48193359375,
0.619140625,
0.0849609375,
0.619140625
],
[
0.51708984375,
0.6123046875,
0.9111328125,
0.611328125,
0.9111328125,
0.62255859375,
0.51708984375,
0.62353515625
],
[
0.08660888671875,
0.62060546875,
0.4794921875,
0.62060546875,
0.4794921875,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.51708984375,
0.625,
0.90625,
0.625,
0.90625,
0.63818359375,
0.51708984375,
0.63818359375
],
[
0.0849609375,
0.6337890625,
0.483642578125,
0.6337890625,
0.483642578125,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.0849609375,
0.646484375,
0.48046875,
0.646484375,
0.48046875,
0.66015625,
0.0849609375,
0.66015625
],
[
0.53369140625,
0.64599609375,
0.9140625,
0.64794921875,
0.9140625,
0.662109375,
0.53369140625,
0.66015625
],
[
0.08660888671875,
0.66357421875,
0.4794921875,
0.66357421875,
0.4794921875,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.5185546875,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.673828125,
0.5185546875,
0.673828125
],
[
0.0849609375,
0.67529296875,
0.483642578125,
0.67529296875,
0.483642578125,
0.689453125,
0.0849609375,
0.689453125
],
[
0.51708984375,
0.67529296875,
0.9140625,
0.67529296875,
0.9140625,
0.689453125,
0.51708984375,
0.689453125
],
[
0.0849609375,
0.689453125,
0.42236328125,
0.689453125,
0.42236328125,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.51708984375,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.51708984375,
0.7021484375
],
[
0.10296630859375,
0.7021484375,
0.48046875,
0.7021484375,
0.48046875,
0.71533203125,
0.10296630859375,
0.71533203125
],
[
0.5185546875,
0.7041015625,
0.91259765625,
0.7041015625,
0.91259765625,
0.7138671875,
0.5185546875,
0.7138671875
],
[
0.0849609375,
0.71728515625,
0.48193359375,
0.71728515625,
0.48193359375,
0.73095703125,
0.0849609375,
0.73095703125
],
[
0.53173828125,
0.71728515625,
0.9140625,
0.71728515625,
0.9140625,
0.73095703125,
0.53173828125,
0.73095703125
],
[
0.0849609375,
0.72998046875,
0.48046875,
0.72998046875,
0.48046875,
0.74365234375,
0.0849609375,
0.74365234375
],
[
0.51708984375,
0.72998046875,
0.9140625,
0.72998046875,
0.9140625,
0.74365234375,
0.51708984375,
0.74365234375
],
[
0.08416748046875,
0.74560546875,
0.48046875,
0.74365234375,
0.48046875,
0.7568359375,
0.08416748046875,
0.7587890625
],
[
0.5185546875,
0.7470703125,
0.91259765625,
0.7470703125,
0.91259765625,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.08660888671875,
0.76025390625,
0.48046875,
0.76025390625,
0.48046875,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.5185546875,
0.76025390625,
0.91259765625,
0.76025390625,
0.91259765625,
0.77099609375,
0.5185546875,
0.77099609375
],
[
0.08416748046875,
0.76953125,
0.48193359375,
0.771484375,
0.48193359375,
0.7861328125,
0.08416748046875,
0.7841796875
],
[
0.5185546875,
0.7734375,
0.9111328125,
0.7734375,
0.9111328125,
0.7841796875,
0.5185546875,
0.7841796875
],
[
0.08416748046875,
0.78515625,
0.48046875,
0.7861328125,
0.48046875,
0.7998046875,
0.08416748046875,
0.798828125
],
[
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,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.51953125,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.51953125,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.5185546875,
0.84326171875,
0.9111328125,
0.84326171875,
0.9111328125,
0.85400390625,
0.5185546875,
0.85400390625
],
[
0.08416748046875,
0.85400390625,
0.483642578125,
0.85498046875,
0.483642578125,
0.869140625,
0.08416748046875,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0849609375,
0.86865234375,
0.483642578125,
0.86865234375,
0.483642578125,
0.8818359375,
0.0849609375,
0.8818359375
],
[
0.51953125,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.51953125,
0.880859375
],
[
0.08416748046875,
0.8818359375,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.08416748046875,
0.89599609375
],
[
0.51708984375,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"DHAP",
"Busy",
"current input post. Finally, we design a personalized decoder to",
"",
"Dialogue History",
"Training",
"gent",
"8",
"",
"fuse the learned user profile into the response generation process",
"Hi Tom! I didn't see you in",
"",
"The personalized decoder can switchbetween generating a word",
"school yesterday.",
"Hi, I'm DHAP, Tom's",
"from a generic vocabulary and copying a word from the user's",
"D",
"I have been to the Apple",
"digital assistant. You can",
"personalized vocabulary. Experimental results on two large-scale",
"",
"Ste",
"reandtried onthelatest",
">",
"chat with me since Tom is",
"",
"acBook",
"datasets show that our proposed DHAP significantly outperforms",
"busynow:)",
"Hey,I want to buy a laptop",
"existing response generation models in various evaluation metrics",
"Would vouliketowatch",
"recently,butIknow little",
"Our contributions are three-fold: (1) We learn the implicit user",
"about computers.",
"Game of Thrones withme",
"this weekend?",
"profile from user's dialogue history automatically forgenerating",
"No,I'll learn programmin",
"If you enjoy programming",
"personalized responses. By this means, our method can be applied",
"andhowtousemynew",
"u can considerthe latest",
"without additional annotations on user profiles, and create person-",
"MacBook",
"alized chatbots as user's digital agents. (2) We build two kinds of",
"userprofilesfromthedialoguehistory,including thegeneral user",
"Figure 1: An example of the personalized chatbot serving as",
"profilereflectingtheuser'sgeneral informationandthedynamic",
"an intelligent agent for user Tom when Tom is busy.",
"post-aware user profile to apply similar interactions in the historical",
"data for the current input. (3) We design a personalized decoder to",
"coordinatetwopersonalized decoding strategies,includinggener-",
"Our idea is motivated by: (1) Contrary to explicit persona descrip-",
"ating a word from the generic vocabulary and copying a word from",
"tions, user dialogue history is easier to be gathered on the user's",
"thepersonalized vocabularytoleveragetheuser'swordpreference",
"client devices.It is evident that obtaining explicit descriptions for",
"massive users is impractical in real applications [4]: First, users",
"2",
"RELATED WORK",
"maybelazyto set theirprofilebeforeusing thechatbot[6].Second,",
"Open-domain Chatbots and Response Generation. Open-domain",
"manually collecting user profiles is costlyand time-consuming.Fur-",
"chatbots have attracted more and more attention, due to their broad",
"thermore, even if the user profile is collected, it cannot be updated",
"application in real applications,such as Microsoft Xiaolce[45].Typi",
"withthe changeof user interests,thus maybe ineffective overtime",
"calmethods can be categorized intotwogroups:retrieval-based and",
"Finally, a fixed set of properties is not suitable for describing all",
"generation-based. Retrieval-based methods aim to select a suitable",
"users.(2)The user dialoguehistory contains massive personalized",
"responsefromalargerepository[38,49,50],whilegeneration",
"information,which is suitableforlearning userprofiles automati-",
"based methodsaim atgenerating aresponsefrom scratch[26, 28",
"cally. As shown in Figure 1, the dialogue history of a user includes",
"32,48].Inthisstudy,wefocusontheresponsegenerationproblem",
"theirhistoricalresponses,andthecorrespondingpostsissued by",
"Some early studies treat the response generation task as a sta-",
"other users.Intuitively,users historical responses can often re-",
"tistical machinetranslation problem becauseofitsend-to-end and",
"flect their language style, background knowledge, frequently used",
"data-drivenfeatures[25,34].Morerecently,withtheprogressof",
"words,and even their interests.For example,theremay be many",
"deep learning, Seq2Seq methods have been applied to response",
"electronic device names that appeared in the historical responses of",
"generation and achieve great performance [8, 26, 27, 34]. Many",
"an electronic hobbyist (e.g., MacBook in Figure 1). Besides, the inter-",
"Seq2Seq-based extensionshavebeen appliedtotacklethe\"safe",
"action content and stylebetween a specific user and others can be",
"response\"problem [8]; to incorporate external knowledge [44]; to",
"captured from the historical post-response pairs. When faced with",
"generateresponseswithemotionsorpersonas[9,23,43];andto",
"a new input post,thechatbot can lookfor thehistorical data, check",
"model the hierarchical structure of the dialogue context [26, 27].",
"how the user has responded to a similar post before, and apply sim-",
"ilar interactions to generate a suitable response. In addition, user",
"Personalized Chatbots. Endowing chatbots with a coherent per-",
"profiles learned fromhistorical data can begradually updated with",
"sonalityisa challenging but necessary steptoachievetheultimate",
"more data being collected. In summary, the user dialogue history is",
"goal of building an intelligent assistant. With personality, a chatbot",
"easy to obtain and appropriate for building user profile.",
"can generate more informative and user-specific responses [9, 41]",
"To achieve our idea, we propose a model DHAP for personalized",
"and has thepotential toperform similarbehaviorsasreal humans.",
"chatbots, which focuses on learning implicit user profile from user",
"Traditional personalized chatbots focus on modeling the user's",
"Dialogue History Automatically and generating Personalized re",
"psychological behavior such as the \"Big Five” of speakers [17]. In",
"sponses. In our model, a general user profile representation is firstly",
"recent years,deep-learningbasedmethodswereproposed tolearn",
"constructedfromtheuser'shistoricalresponsestocapturethegen",
"thepersona information directly fromlarge-scaledialoguedatasets",
"eral information, including user interest, background knowledge,",
"via end-to-end neural networks[4, 9, 23, 4i].Some researchers",
"and speaking style.This is implemented by a personalized language",
"first tried to input the user ID embeddings into the decoder of a",
"model based on Transformer.Then, we design apersonalized post",
"Seq2Seq model to generate more personalized responses [1, 3, 4, 9].",
"encodertoconstructthepersonalizedpost representation.Thegen",
"Despite users can be identified by their IDs, the personalized perfor-",
"eral userprofileisutilized in thepostencodertobetter capturethe",
"mance is limitedbecause no user-related information is used inthe",
"semantic information of the input post. Next, we build a key-value",
"model.Therefore,anothergroup of researchers proposed assigning",
"memory network to store the user's historical post-response pairs.",
"explicit profiles (personas)for chatbots to generate personalized",
"Based on this history memory, the dynamic post-aware user pro-",
"responses.Forexample,Zhangetal.[41]publishedthePERSONA-",
"file is built by highlighting the historical responses relevant to the",
"CHAT dataset, in which each user is assigned with several persona"
] |
{
"arxiv_id": "2108.09355",
"language": "en",
"page_idx": 1
}
|
||
2108.09355-en-2
|
2108.09355-en-2.jpg
|
<FigureHere>
Figure 2: The overall structure of the proposed model DHAP, which consists of (1) a history encoder, (2) a personalized post
encoder, (3) a user history memory, and (4) a personalized decoder.
is to generate a personalized response $Y=(y_1,\cdots,y_{L_Y})$ as
$$p\left(Y|X,u\right)=p\left(Y|X,H\right)=\prod\limits_{t=1}^{L_Y}p\left(y_t|y_{<t},X,H\right),$$
(1)
description sentences and the conversations between users are collected. With explicit profile descriptions, chatbots can learn to generate more consistent and informative dialogues. On this dataset, many methods have achieved encouraging performance, such as variational autoencoders [30],pre-trained language models [29,36], and multi-task modeling [29,31,35]. These explicit persona-based methods enjoy the high quality of predefined personas, but the acquisition of these persona data is expensive and even impossible when applied to real-world systems [4].
In this work, we propose using the implicit user profile to drive the response generation. Such user profile is implicitly learned with neural models from user dialogue history rather than predefined explicitly. Due to the fact that the user dialogue history contains abundant user-specific information and is easily accessible in realworld applications, our method is more applicable in practice.
where $y_t$ denotes the word generated at the $t$-th step, and $y_<t$ denotes the previous generated words $(y_1,\cdots,y_{t-1}).$ It is worth noting that the dialogue history $H$ here includes dialogues between the user $u$ and several other users, thus the history is not a multiturn dialogue between two fixed interlocutors.
To compute the probability $p(Y|X,H)$,we design a model called DHAP, which stands for learning user profiles from Dialogue History Automatically for Personalized chatbots. The structure of DHAP is shown in Figure 2. We briefly introduce the key components of DHAP as follows. The number of the modules corresponds to the mark in the figure. In general, DHAP considers personalized information in both the encoder and the decoder side.
# 3 мЕтнороLoGY
In this section, we first provide an overview of our proposed model DHAP. The details of each component are provided later, and the model training is introduced finally.
## 3.1 $\textbf{The Overview of DHAP}$
3.1.1 Encoder. DHAP has two different encoders, which encode
the input post and user's historical responses, respectively:
Part (1): History encoder and general user profile. Since abundant personalized information $(e.g.,\tilde{\text{background knowledge}}$ and speaking style) is often hidden in user's historical responses, DHAP firstly establishes a Transformer-based personalized language model to encode historical responses $(R_1,\cdots,R_n)$,then summarizes the general user profile $\mathbf{e}^G$ based on the historical responses. As e$^G$ does not depend on the input $X$,we call it general user profile. “General” does not mean it is general among all users or cannot be updated together with time. In contrast, every user has their own profile, which can be updated once they issue a new response.
Part (2){:} Personalized post encoder. DHAP also has an encoder for the input post, which is implemented by a bidirectional
Suppose that for a user $u$, we have their dialogue history $H$ including a series of responses issued by $u$ and the corresponding posts: $H=((P_{1},R_{1}),\cdots,(P_{n},R_{n}))$,where $n$ is the number of historical post-response pairs. Note that the posts $(P_1,\cdots,P_n)$ here can be issued by different users, but the responses $(R_1,\cdots,R_n)$ are all issued by the same user $u.$ We call them historical posts and historical responses respectively in the following sections. Under the singleturn setting, given an input post $X=(x_1,\cdots,x_{L_X})$ and the user dialogue history $H$,with sequence-to-sequence modeling, our task
|
[
[
0.2279052734375,
0.10919189453125,
0.351318359375,
0.10919189453125,
0.351318359375,
0.11932373046875,
0.2279052734375,
0.11932373046875
],
[
0.53369140625,
0.1142578125,
0.576171875,
0.1142578125,
0.576171875,
0.125,
0.53369140625,
0.125
],
[
0.64208984375,
0.1142578125,
0.7060546875,
0.1142578125,
0.7060546875,
0.125,
0.64208984375,
0.125
],
[
0.73193359375,
0.11614990234375,
0.7607421875,
0.11614990234375,
0.7607421875,
0.125,
0.73193359375,
0.125
],
[
0.76318359375,
0.11553955078125,
0.81396484375,
0.11553955078125,
0.81396484375,
0.12249755859375,
0.76318359375,
0.12249755859375
],
[
0.306396484375,
0.12371826171875,
0.398681640625,
0.12371826171875,
0.398681640625,
0.1337890625,
0.306396484375,
0.1337890625
],
[
0.50732421875,
0.11932373046875,
0.53125,
0.11932373046875,
0.53125,
0.12371826171875,
0.50732421875,
0.12371826171875
],
[
0.5390625,
0.12371826171875,
0.59326171875,
0.12371826171875,
0.59326171875,
0.1337890625,
0.5390625,
0.1337890625
],
[
0.765625,
0.12371826171875,
0.80615234375,
0.12371826171875,
0.80615234375,
0.1337890625,
0.765625,
0.1337890625
],
[
0.650390625,
0.125,
0.6689453125,
0.125,
0.6689453125,
0.1319580078125,
0.650390625,
0.1319580078125
],
[
0.193603515625,
0.125,
0.2646484375,
0.125,
0.2646484375,
0.1351318359375,
0.193603515625,
0.1351318359375
],
[
0.64794921875,
0.1331787109375,
0.7138671875,
0.1331787109375,
0.7138671875,
0.143310546875,
0.64794921875,
0.143310546875
],
[
0.74169921875,
0.1331787109375,
0.75634765625,
0.1331787109375,
0.75634765625,
0.143310546875,
0.74169921875,
0.143310546875
],
[
0.7646484375,
0.1331787109375,
0.81884765625,
0.1331787109375,
0.81884765625,
0.143310546875,
0.7646484375,
0.143310546875
],
[
0.62353515625,
0.1337890625,
0.63623046875,
0.1337890625,
0.63623046875,
0.1427001953125,
0.62353515625,
0.1427001953125
],
[
0.4248046875,
0.1383056640625,
0.490234375,
0.1383056640625,
0.490234375,
0.1483154296875,
0.4248046875,
0.1483154296875
],
[
0.5048828125,
0.136962890625,
0.51708984375,
0.1275634765625,
0.52783203125,
0.1363525390625,
0.515625,
0.1458740234375
],
[
0.35546875,
0.1395263671875,
0.370849609375,
0.1395263671875,
0.370849609375,
0.147705078125,
0.35546875,
0.147705078125
],
[
0.178955078125,
0.1395263671875,
0.2017822265625,
0.1395263671875,
0.2017822265625,
0.146484375,
0.178955078125,
0.146484375
],
[
0.236083984375,
0.14013671875,
0.2484130859375,
0.14013671875,
0.2484130859375,
0.1483154296875,
0.236083984375,
0.1483154296875
],
[
0.324462890625,
0.14013671875,
0.336669921875,
0.14013671875,
0.336669921875,
0.14453125,
0.324462890625,
0.14453125
],
[
0.306396484375,
0.1514892578125,
0.340576171875,
0.1514892578125,
0.340576171875,
0.1610107421875,
0.306396484375,
0.1610107421875
],
[
0.381591796875,
0.15087890625,
0.483642578125,
0.15087890625,
0.483642578125,
0.1610107421875,
0.381591796875,
0.1610107421875
],
[
0.57421875,
0.15283203125,
0.701171875,
0.15283203125,
0.701171875,
0.1666259765625,
0.57421875,
0.1666259765625
],
[
0.35693359375,
0.154052734375,
0.363525390625,
0.154052734375,
0.363525390625,
0.1578369140625,
0.35693359375,
0.1578369140625
],
[
0.1854248046875,
0.15283203125,
0.2132568359375,
0.15283203125,
0.2132568359375,
0.160400390625,
0.1854248046875,
0.160400390625
],
[
0.3544921875,
0.1622314453125,
0.373291015625,
0.1641845703125,
0.370849609375,
0.173583984375,
0.351318359375,
0.1717529296875
],
[
0.423095703125,
0.1622314453125,
0.48193359375,
0.1622314453125,
0.48193359375,
0.1767578125,
0.423095703125,
0.1767578125
],
[
0.236083984375,
0.1654052734375,
0.25244140625,
0.1654052734375,
0.25244140625,
0.17236328125,
0.236083984375,
0.17236328125
],
[
0.1871337890625,
0.1654052734375,
0.21728515625,
0.1654052734375,
0.21728515625,
0.173583984375,
0.1871337890625,
0.173583984375
],
[
0.31201171875,
0.1666259765625,
0.338134765625,
0.1666259765625,
0.338134765625,
0.173583984375,
0.31201171875,
0.173583984375
],
[
0.50341796875,
0.17236328125,
0.54736328125,
0.17236328125,
0.54736328125,
0.184326171875,
0.50341796875,
0.184326171875
],
[
0.580078125,
0.1717529296875,
0.63818359375,
0.1717529296875,
0.63818359375,
0.186279296875,
0.580078125,
0.186279296875
],
[
0.65869140625,
0.17236328125,
0.708984375,
0.17236328125,
0.708984375,
0.184326171875,
0.65869140625,
0.184326171875
],
[
0.74169921875,
0.17236328125,
0.79833984375,
0.17236328125,
0.79833984375,
0.18310546875,
0.74169921875,
0.18310546875
],
[
0.54833984375,
0.186279296875,
0.58251953125,
0.186279296875,
0.58251953125,
0.1956787109375,
0.54833984375,
0.1956787109375
],
[
0.7099609375,
0.1875,
0.744140625,
0.1875,
0.744140625,
0.1962890625,
0.7099609375,
0.1962890625
],
[
0.3447265625,
0.1881103515625,
0.39453125,
0.1881103515625,
0.39453125,
0.1988525390625,
0.3447265625,
0.1988525390625
],
[
0.1895751953125,
0.1881103515625,
0.236083984375,
0.1881103515625,
0.236083984375,
0.199462890625,
0.1895751953125,
0.199462890625
],
[
0.2467041015625,
0.20263671875,
0.34228515625,
0.20263671875,
0.34228515625,
0.2139892578125,
0.2467041015625,
0.2139892578125
],
[
0.450927734375,
0.2183837890625,
0.463134765625,
0.2183837890625,
0.463134765625,
0.2279052734375,
0.450927734375,
0.2279052734375
],
[
0.46728515625,
0.2183837890625,
0.5,
0.2197265625,
0.5,
0.2310791015625,
0.46728515625,
0.2298583984375
],
[
0.56201171875,
0.2342529296875,
0.60302734375,
0.2342529296875,
0.60302734375,
0.24560546875,
0.56201171875,
0.24560546875
],
[
0.7099609375,
0.23486328125,
0.76123046875,
0.23486328125,
0.76123046875,
0.246826171875,
0.7099609375,
0.246826171875
],
[
0.615234375,
0.2392578125,
0.68701171875,
0.2412109375,
0.68701171875,
0.251953125,
0.615234375,
0.25
],
[
0.453369140625,
0.238037109375,
0.4599609375,
0.238037109375,
0.4599609375,
0.2412109375,
0.453369140625,
0.2412109375
],
[
0.521484375,
0.2412109375,
0.52685546875,
0.2412109375,
0.52685546875,
0.25,
0.521484375,
0.25
],
[
0.63134765625,
0.25,
0.67236328125,
0.251953125,
0.6708984375,
0.261962890625,
0.6298828125,
0.260009765625
],
[
0.55615234375,
0.2587890625,
0.60107421875,
0.2587890625,
0.60107421875,
0.26953125,
0.55615234375,
0.26953125
],
[
0.70751953125,
0.2568359375,
0.7646484375,
0.2568359375,
0.7646484375,
0.271484375,
0.70751953125,
0.271484375
],
[
0.58251953125,
0.290283203125,
0.68896484375,
0.290283203125,
0.68896484375,
0.300537109375,
0.58251953125,
0.300537109375
],
[
0.4755859375,
0.260009765625,
0.495849609375,
0.260009765625,
0.495849609375,
0.268310546875,
0.4755859375,
0.268310546875
],
[
0.17236328125,
0.29541015625,
0.1993408203125,
0.29541015625,
0.1993408203125,
0.306884765625,
0.17236328125,
0.306884765625
],
[
0.26953125,
0.294189453125,
0.31787109375,
0.294189453125,
0.31787109375,
0.304931640625,
0.26953125,
0.304931640625
],
[
0.44677734375,
0.3125,
0.50341796875,
0.3125,
0.50341796875,
0.322509765625,
0.44677734375,
0.322509765625
],
[
0.68115234375,
0.3125,
0.7158203125,
0.3125,
0.7158203125,
0.323974609375,
0.68115234375,
0.323974609375
],
[
0.73779296875,
0.3125,
0.77392578125,
0.3125,
0.77392578125,
0.323974609375,
0.73779296875,
0.323974609375
],
[
0.77294921875,
0.3125,
0.810546875,
0.3125,
0.810546875,
0.322509765625,
0.77294921875,
0.322509765625
],
[
0.580078125,
0.314453125,
0.60888671875,
0.314453125,
0.60888671875,
0.322509765625,
0.580078125,
0.322509765625
],
[
0.521484375,
0.314453125,
0.55126953125,
0.314453125,
0.55126953125,
0.322509765625,
0.521484375,
0.322509765625
],
[
0.62158203125,
0.314453125,
0.64892578125,
0.314453125,
0.64892578125,
0.322509765625,
0.62158203125,
0.322509765625
],
[
0.339111328125,
0.32568359375,
0.36279296875,
0.32568359375,
0.36279296875,
0.333984375,
0.339111328125,
0.333984375
],
[
0.390625,
0.317626953125,
0.414306640625,
0.317626953125,
0.414306640625,
0.324462890625,
0.390625,
0.324462890625
],
[
0.4697265625,
0.328857421875,
0.5009765625,
0.328857421875,
0.5009765625,
0.3359375,
0.4697265625,
0.3359375
],
[
0.52294921875,
0.328857421875,
0.53369140625,
0.33203125,
0.53125,
0.3359375,
0.521484375,
0.333251953125
],
[
0.53515625,
0.32763671875,
0.57666015625,
0.32763671875,
0.57666015625,
0.33837890625,
0.53515625,
0.33837890625
],
[
0.57861328125,
0.328857421875,
0.609375,
0.328857421875,
0.609375,
0.337158203125,
0.57861328125,
0.337158203125
],
[
0.62158203125,
0.328857421875,
0.65283203125,
0.328857421875,
0.65283203125,
0.337158203125,
0.62158203125,
0.337158203125
],
[
0.654296875,
0.328857421875,
0.6748046875,
0.328857421875,
0.6748046875,
0.3359375,
0.654296875,
0.3359375
],
[
0.6845703125,
0.328857421875,
0.71826171875,
0.328857421875,
0.71826171875,
0.33837890625,
0.6845703125,
0.33837890625
],
[
0.74267578125,
0.328857421875,
0.75341796875,
0.330078125,
0.75244140625,
0.335205078125,
0.74169921875,
0.333984375
],
[
0.75244140625,
0.328857421875,
0.76708984375,
0.328857421875,
0.76708984375,
0.3359375,
0.75244140625,
0.3359375
],
[
0.77783203125,
0.328857421875,
0.80615234375,
0.326904296875,
0.80810546875,
0.333984375,
0.779296875,
0.3359375
],
[
0.57666015625,
0.34228515625,
0.611328125,
0.34228515625,
0.611328125,
0.354248046875,
0.57666015625,
0.354248046875
],
[
0.43310546875,
0.330078125,
0.4697265625,
0.330078125,
0.4697265625,
0.337158203125,
0.43310546875,
0.337158203125
],
[
0.62109375,
0.34228515625,
0.65283203125,
0.34228515625,
0.65283203125,
0.354248046875,
0.62109375,
0.354248046875
],
[
0.68310546875,
0.34228515625,
0.7138671875,
0.34228515625,
0.7138671875,
0.354248046875,
0.68310546875,
0.354248046875
],
[
0.521484375,
0.343505859375,
0.55224609375,
0.341552734375,
0.5537109375,
0.352783203125,
0.52294921875,
0.354736328125
],
[
0.43310546875,
0.343505859375,
0.5009765625,
0.343505859375,
0.5009765625,
0.354248046875,
0.43310546875,
0.354248046875
],
[
0.74169921875,
0.343505859375,
0.7548828125,
0.343505859375,
0.7548828125,
0.3515625,
0.74169921875,
0.3515625
],
[
0.75244140625,
0.3447265625,
0.76708984375,
0.3447265625,
0.76708984375,
0.3515625,
0.75244140625,
0.3515625
],
[
0.78173828125,
0.34521484375,
0.7890625,
0.34521484375,
0.7890625,
0.348388671875,
0.78173828125,
0.348388671875
],
[
0.298095703125,
0.352783203125,
0.330810546875,
0.352783203125,
0.330810546875,
0.356689453125,
0.298095703125,
0.356689453125
],
[
0.339111328125,
0.348388671875,
0.36767578125,
0.35107421875,
0.366943359375,
0.359130859375,
0.338134765625,
0.356689453125
],
[
0.4697265625,
0.359130859375,
0.50732421875,
0.359130859375,
0.50732421875,
0.37255859375,
0.4697265625,
0.37255859375
],
[
0.5361328125,
0.359130859375,
0.6005859375,
0.359130859375,
0.6005859375,
0.37060546875,
0.5361328125,
0.37060546875
],
[
0.61767578125,
0.35791015625,
0.654296875,
0.35791015625,
0.654296875,
0.369384765625,
0.61767578125,
0.369384765625
],
[
0.69677734375,
0.359130859375,
0.765625,
0.359130859375,
0.765625,
0.37060546875,
0.69677734375,
0.37060546875
],
[
0.771484375,
0.359130859375,
0.810546875,
0.356689453125,
0.81201171875,
0.36865234375,
0.77294921875,
0.37060546875
],
[
0.348876953125,
0.3642578125,
0.358642578125,
0.3642578125,
0.358642578125,
0.369384765625,
0.348876953125,
0.369384765625
],
[
0.2303466796875,
0.365478515625,
0.24267578125,
0.365478515625,
0.24267578125,
0.369384765625,
0.2303466796875,
0.369384765625
],
[
0.303955078125,
0.363037109375,
0.31787109375,
0.363037109375,
0.31787109375,
0.371826171875,
0.303955078125,
0.371826171875
],
[
0.2197265625,
0.376953125,
0.366943359375,
0.376953125,
0.366943359375,
0.386962890625,
0.2197265625,
0.386962890625
],
[
0.56884765625,
0.376953125,
0.66845703125,
0.376953125,
0.66845703125,
0.386962890625,
0.56884765625,
0.386962890625
],
[
0.08660888671875,
0.40283203125,
0.91259765625,
0.40283203125,
0.91259765625,
0.417236328125,
0.08660888671875,
0.417236328125
],
[
0.0849609375,
0.417236328125,
0.529296875,
0.417236328125,
0.529296875,
0.43115234375,
0.0849609375,
0.43115234375
],
[
0.0882568359375,
0.455078125,
0.48046875,
0.455078125,
0.48046875,
0.46533203125,
0.0882568359375,
0.46533203125
],
[
0.51708984375,
0.453857421875,
0.8759765625,
0.453857421875,
0.8759765625,
0.468505859375,
0.51708984375,
0.468505859375
],
[
0.0849609375,
0.46728515625,
0.483642578125,
0.46728515625,
0.483642578125,
0.48095703125,
0.0849609375,
0.48095703125
],
[
0.08660888671875,
0.48291015625,
0.48193359375,
0.48291015625,
0.48193359375,
0.49365234375,
0.08660888671875,
0.49365234375
],
[
0.73876953125,
0.481689453125,
0.744140625,
0.481689453125,
0.744140625,
0.4873046875,
0.73876953125,
0.4873046875
],
[
0.57421875,
0.490478515625,
0.7255859375,
0.488037109375,
0.7255859375,
0.50244140625,
0.57421875,
0.5048828125
],
[
0.75341796875,
0.4912109375,
0.8544921875,
0.488037109375,
0.8544921875,
0.50244140625,
0.75341796875,
0.505859375
],
[
0.89453125,
0.4912109375,
0.91259765625,
0.4912109375,
0.91259765625,
0.5048828125,
0.89453125,
0.5048828125
],
[
0.0849609375,
0.5087890625,
0.48046875,
0.5087890625,
0.48046875,
0.52294921875,
0.0849609375,
0.52294921875
],
[
0.0849609375,
0.49560546875,
0.483642578125,
0.49560546875,
0.483642578125,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.0849609375,
0.52294921875,
0.48193359375,
0.52294921875,
0.48193359375,
0.5361328125,
0.0849609375,
0.5361328125
],
[
0.515625,
0.5244140625,
0.9111328125,
0.52587890625,
0.9111328125,
0.54052734375,
0.515625,
0.5390625
],
[
0.0849609375,
0.5361328125,
0.48193359375,
0.5361328125,
0.48193359375,
0.5498046875,
0.0849609375,
0.5498046875
],
[
0.51708984375,
0.54052734375,
0.91259765625,
0.54052734375,
0.91259765625,
0.5537109375,
0.51708984375,
0.5537109375
],
[
0.08660888671875,
0.5517578125,
0.48046875,
0.5517578125,
0.48046875,
0.5625,
0.08660888671875,
0.5625
],
[
0.51708984375,
0.5537109375,
0.9140625,
0.5537109375,
0.9140625,
0.5673828125,
0.51708984375,
0.5673828125
],
[
0.0849609375,
0.564453125,
0.324462890625,
0.564453125,
0.324462890625,
0.57763671875,
0.0849609375,
0.57763671875
],
[
0.51708984375,
0.5673828125,
0.9140625,
0.5673828125,
0.9140625,
0.58056640625,
0.51708984375,
0.58056640625
],
[
0.1004638671875,
0.57763671875,
0.48193359375,
0.5771484375,
0.48193359375,
0.59033203125,
0.1004638671875,
0.59130859375
],
[
0.51708984375,
0.58349609375,
0.7958984375,
0.58349609375,
0.7958984375,
0.59326171875,
0.51708984375,
0.59326171875
],
[
0.0849609375,
0.59130859375,
0.48193359375,
0.59130859375,
0.48193359375,
0.60498046875,
0.0849609375,
0.60498046875
],
[
0.53173828125,
0.5947265625,
0.9140625,
0.59326171875,
0.9140625,
0.60693359375,
0.53173828125,
0.60791015625
],
[
0.08416748046875,
0.60498046875,
0.48046875,
0.60498046875,
0.48046875,
0.61865234375,
0.08416748046875,
0.61865234375
],
[
0.51708984375,
0.609375,
0.9140625,
0.609375,
0.9140625,
0.62255859375,
0.51708984375,
0.62255859375
],
[
0.08416748046875,
0.619140625,
0.48193359375,
0.619140625,
0.48193359375,
0.63330078125,
0.08416748046875,
0.63330078125
],
[
0.51708984375,
0.6220703125,
0.9150390625,
0.6220703125,
0.9150390625,
0.63525390625,
0.51708984375,
0.63525390625
],
[
0.0849609375,
0.63330078125,
0.483642578125,
0.63330078125,
0.483642578125,
0.646484375,
0.0849609375,
0.646484375
],
[
0.5185546875,
0.63818359375,
0.91259765625,
0.63818359375,
0.91259765625,
0.6484375,
0.5185546875,
0.6484375
],
[
0.0849609375,
0.64794921875,
0.45751953125,
0.64794921875,
0.45751953125,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.51708984375,
0.65087890625,
0.91259765625,
0.65087890625,
0.91259765625,
0.6640625,
0.51708984375,
0.6640625
],
[
0.51708984375,
0.6640625,
0.9140625,
0.6640625,
0.9140625,
0.67822265625,
0.51708984375,
0.67822265625
],
[
0.10455322265625,
0.67529296875,
0.253173828125,
0.67529296875,
0.253173828125,
0.689453125,
0.10455322265625,
0.689453125
],
[
0.5185546875,
0.6787109375,
0.8388671875,
0.6787109375,
0.8388671875,
0.689453125,
0.5185546875,
0.689453125
],
[
0.0849609375,
0.69580078125,
0.48193359375,
0.69580078125,
0.48193359375,
0.708984375,
0.0849609375,
0.708984375
],
[
0.08660888671875,
0.6796875,
0.1094970703125,
0.6796875,
0.1094970703125,
0.68798828125,
0.08660888671875,
0.68798828125
],
[
0.51708984375,
0.7021484375,
0.91259765625,
0.7021484375,
0.91259765625,
0.71533203125,
0.51708984375,
0.71533203125
],
[
0.0849609375,
0.708984375,
0.483642578125,
0.708984375,
0.483642578125,
0.72265625,
0.0849609375,
0.72265625
],
[
0.51708984375,
0.71630859375,
0.87353515625,
0.71630859375,
0.87353515625,
0.72998046875,
0.51708984375,
0.72998046875
],
[
0.08416748046875,
0.7236328125,
0.303955078125,
0.7236328125,
0.303955078125,
0.7373046875,
0.08416748046875,
0.7373046875
],
[
0.53369140625,
0.72998046875,
0.9140625,
0.72998046875,
0.9140625,
0.74365234375,
0.53369140625,
0.74365234375
],
[
0.51708984375,
0.744140625,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.08416748046875,
0.75390625,
0.318603515625,
0.751953125,
0.318603515625,
0.7666015625,
0.08416748046875,
0.76806640625
],
[
0.51708984375,
0.75830078125,
0.9140625,
0.75830078125,
0.9140625,
0.771484375,
0.51708984375,
0.771484375
],
[
0.0849609375,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.78515625,
0.0849609375,
0.78515625
],
[
0.51708984375,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.08660888671875,
0.787109375,
0.48193359375,
0.787109375,
0.48193359375,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.515625,
0.7861328125,
0.9140625,
0.78515625,
0.9140625,
0.798828125,
0.515625,
0.7998046875
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.0849609375,
0.81298828125,
0.483642578125,
0.81298828125,
0.483642578125,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.0849609375,
0.8271484375,
0.48193359375,
0.8271484375,
0.48193359375,
0.84033203125,
0.0849609375,
0.84033203125
],
[
0.5185546875,
0.8291015625,
0.91259765625,
0.8291015625,
0.91259765625,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.515625,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.515625,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.483642578125,
0.85498046875,
0.483642578125,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.515625,
0.85498046875,
0.89208984375,
0.85498046875,
0.89208984375,
0.86865234375,
0.515625,
0.86865234375
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.53173828125,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.53173828125,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.48193359375,
0.88330078125,
0.48193359375,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"(4)PersonalizedDecoder",
"General",
"Personalized",
"囍8",
"Post-aware",
"Personalized Vocab",
"",
"User Profile",
"Dynamic",
"Post",
"Generic Vocab",
"Representation",
"ep",
"User Profile",
"Ct",
"0.22computer",
"00",
"0.1",
"comp",
"0.5",
"ita",
"football",
"×0.7= 0.09 football",
"(3) User History Memory",
"",
"football",
"0.6",
"0.42 cricket",
"0.01",
"cricket",
"cricket",
"pioo",
"pip",
"pn",
"oopr.",
"PostPi",
"PostPr",
"p(m)=.0.7",
"p(mg)= 0.3",
"Decoding Switcher",
"et",
"",
"@kM",
"@Qkm",
"Post-Response",
"",
"Re",
"Memory",
"QvM",
"8vm",
"TransformerEncoder",
"",
"Init ",
"Attention",
"ward e",
"00",
"0",
"Too",
"",
"零",
"",
"(hp",
"Init",
"03",
"",
"...",
"寻",
"8",
"...",
"36",
"",
"",
"良康",
"00",
"segment",
"",
"e0",
"",
"position?",
"O",
"",
"",
"",
"00",
"<CLS>",
"Response Ri",
"<SEP>",
"ResponseRn",
"<SEP>",
"X1",
"",
"X2",
"(2)PersonalizedPostEncoder",
"(1) History Encoder",
"Figure 2: The overall structure of the proposed model DHAP, which consists of (1) a history encoder, (2) a personalized post",
"encoder, (3) a user history memory, and (4) a personalized decoder.",
"description sentences and the conversations between users are",
"is to generate a personalized response Y = (y1,**+, yLy) as:",
"collected. With explicit profile descriptions, chatbots can learn to",
"generatemore consistentand informativedialogues.On this dataset,",
"E",
"p(YIX,u)=p(Y|X,H)",
"p(ytly<t,X,H),",
"(1)",
"variational autoencoders [30], pre-trained language models [29, 36]",
"many methods have achieved encouraging performance, such as",
"and multi-task modeling [29, 31, 35].These explicit persona-based",
"where yt denotes the word generated at the t-th step, and y<t",
"methods enjoy the high quality of predefined personas, but the",
"denotes the previous generated words (y1,++,yt-1). It is worth",
"acquisitionofthesepersonadataisexpensiveand evenimpossible",
"noting that the dialogue history H here includes dialogues between",
"when applied to real-world systems [4].",
"the user u and several other users, thus the history is not a multi-",
"In this work,we propose using the implicituser profileto drive",
"turndialoguebetweentwofixed interlocutors",
"the response generation. Such user profile is implicitly learned with",
"To compute the probability p(Y|X, H), we design a model called",
"neural models from user dialogue history rather than predefined",
"DHAP, which stands for learning user profiles from Dialogue His-",
"explicitly. Due to the fact that the user dialogue history contains",
"tory Automatically for Personalized chatbots. The structure of",
"abundant user-specific information and is easily accessible in real-",
"DHAPisshowninFigure2.Webrieflyintroducethekeycompo-",
"world applications, our method is more applicable in practice.",
"nents of DHAP as follows. The number of the modules corresponds",
"to the mark in the figure. In general, DHAP considers personalized",
"METHODOLOGY",
"informationin boththeencoderand thedecoder side",
"In this section, we first provide an overview of our proposed model",
"3",
"3.1.1Encoder.DHAP has two different encoders, which encode",
"DHAP. The details of each component are provided later, and the",
"the input post and user's historical responses, respectively:",
"model training is introduced finally.",
"Part (1): History encoder and general user profile. Since",
"abundant personalized information (e.g.,background knowledge",
"3.1TheOverview of DHAP",
"and speaking style) is often hidden in user's historical responses,",
"Suppose that for a user u, we have their dialogue history H including",
"DHAP firstly establishes a Transformer-based personalized lan-",
"a series of responses issued by u and the corresponding posts:",
"guagemodel to encode historical responses (Ri,.,Rn),then sum-",
"H = ((Pi, Ri),..,(Pn, Rn)), where n is the number of historical",
"marizes the general user profile e based on the historical responses.",
"post-response pairs. Note that the posts (Pi,.*, Pn) here can be",
"As eG does not depend on the input X, we call it general user profile",
"issued by different users, but the responses (Ri,*++, Rn) are all",
"\"General\"does not mean it isgeneral among all users or cannot be",
"issued by the same user u. We call them historical posts and historical",
"updated together with time. In contrast, every user has their own",
"responses respectively in the following sections. Under the single-",
"profile, which can be updated once they issue a new response.",
"turn setting, given an input post X = (xi,*-, xLx) and the user",
"Part (2): Personalized post encoder. DHAP also has an en-",
"dialogue history H, with sequence-to-sequence modeling, our task",
" coder for the input post, which is implemented by a bidirectional"
] |
{
"arxiv_id": "2108.09355",
"language": "en",
"page_idx": 2
}
|
||
2108.09355-en-3
|
2108.09355-en-3.jpg
|
GRU (BiGRU). To make the post encoder aware of the user's personalized information, we use the general user profile $\mathbf{e}^G$ for initialization. Consequently, the post $X=(x_1,\cdots,x_{LX})$ is represented by hidden state sequence $\mathbf{h}^P=(\mathbf{h}_{1}^{P},\cdots,\mathbf{h}_{LX}^{P}).$ These representations will be dynamically aggregated as a personalized post representation c$_t$ by an attention mechanism at the decoding step $t.$
3.1.2 Decoder. In the decoder side, DHAP incorporates the per-
sonalized information in two perspectives:
Part (3): User history memory and dynamic post-aware
user profile. In the general user profile, all historical responses are considered in a general view. However, for a specific input post $X$, the historical responses may play different roles. Intuitively, the historical responses relevant to $X$ are more valuable than the irrelevant ones. To highlight the historical responses relevant to the input post, DHAP builds a key-value memory of historical postresponse pairs $((P_1,R_1),\cdots,(P_n,R_n)).$ Given c$_t$ as a personalized representation of $X$,DHAP builds a dynamic post-aware user profile $\mathbf{e}_{t}^{\hat{D}}$ by selecting and aggregating the historical responses from the user history memory. As e$_t^D$ is dynamic with different input posts $X$ and summarizes related user history information, we call it dynamic post-aware user profile.
Part (4){:}Personalized decoder. Finally, the personalized post representation c$_t$,general user profile e$^G$,and dynamic post-aware user profile e$_t^D$ are fused together to decode the response sequentially. Inspired by CopyNet [7], DHAP can switch between generating a word from a generic vocabulary $(p(y_t|m_g))$ and copying a word from a user's personalized vocabulary $(p(y_{t}|m_{c}))$ as:
$$p(y_t|y_{<t},X,H)=p(m_g)p(y_t|m_g)+p(m_c)p(y_t|m_c),$$
(2)
where $p(m_{g})$ and $p(m_{c})$ are computed by our designed decoding switcher. $p(y_{t}|m_{g})$ and $p(y_{t}|m_{c})$ are calculated based on same inputs$(e.g.,\mathbf{c}_{t},\mathbf{e}^{G}$ and $\mathbf{e}_t^D)$ with different functions.
In the remaining part of this section, we will introduce these
four components of DHAP in detail.
$\begin{array}{l}\mathrm{where~e_{j}^{r_{i},\mathrm{fuse~is~the~sum~of~the~word~embedding,~segment~embed-}}\\\mathrm{ding,~and~position~embedding~of~the~}j\text{-th word.These three embed-}\\\mathrm{ding,~and~position~embedding~of~the~}j\text{-th word.These three embed-}\\\mathrm{ding,~and~position~embe~dding~of~}}\end{array}$ dings are used in a similar way like BERT [5]. The representation $of“\langle CLS\rangle” token (e^G) summarizes the information in the entire$ historical responses, thus we call it the general user profile. $\mathbf{E}^R$ contains contextual representations of words in historical responses. For $R_{i}$,we denote its contextual representations as$(\mathbf{e}_1^{r_i},\cdots,\mathbf{e}_{L_{R_i}}^{r_i}).$ Transformer$_N(\cdot)$ is a N-layer bidirectional Transformer encoder identical to the original implementation described in [33].
With the history encoder, we obtain:(1) a general user profile $\mathbf{e}^G$, which summarizes all historical responses and contains personalized information of the user; and (2) the contextual representations of words in the historical responses $\mathbf{E}^R.$ These representations will be used to build the dynamic post-aware user profile in Section 3.4.
## 3.3 Personalized Post Encoder
In real-world applications, the posts are often very short, even ambiguous. Thus, building accurate encoding of the input post is difficult, which further leads to poor quality of the generated responses [9,28]. Fortunately, with personalized background knowledge, the chatbot can get more input information and is promising to better capture the semantic information of the post. Let us use an example to explain this: given a post “The new MAC is so beautiful”, $\dot{\text{different users may have different understandings. For a program-}}$ mer,“MAC” may refer to the Apple's laptop; but for a fashion girl, she may associate "MAC" with the lipstick. Therefore, the user's history can help distinguish the word "MAC" and provide more background knowledge for the input post, so that the post representation can be significantly enhanced. We call this encoder “personalized post encoder”- the encoded representations of the same post can be different for various users, because these users may have different profiles and different understandings of the same post.
Specifically, DHAP employs a BiGRU to encode the input post. Here we choose RNN-based architectures because they are better at capturing local correlations and encoding positional information than Transformers for short texts [19,39]. To facilitate the encoder with personalized information, DHAP uses the general user profile e$^{\tilde{G}}$ to initialize hidden states of BiGRU. Given the post $X=(x_{1},\cdots,x_{L_{X}})$,its representations $(\mathbf{h}_1^{P},\cdots,\mathbf{h}_{L_{X}}^{P})$ are built as:
$$\mathbf{h}_{i}^{P}=\mathrm{BiGRU}(\mathbf{h}_{i-1}^{P},\mathbf{x}_{i}),\quad\mathbf{h}_{0}^{P}=\mathrm{ReLU}(\mathrm{MLP}(\mathbf{e}^{G})),$$
(5)
where x$_i$ is the embedding obtained by the embedding table.
These hidden states draw personalized information from the general user profile, and in the decoding phase, they are aggregated by the decoding hidden state h$_t^R$ through an attention mechanism:
$$\mathbf{c}_t=\sum_{i=1}^{L_X}\alpha_{t,i}\mathbf{h}_i^p,$$
$$\sum_{i=1}\\\alpha_{t,i}=\frac{\exp\left(s_{t,i}^{rp}\right)}{\sum_{j=1}^{L_{X}}\exp(s_{t,j}^{rp})},\quad s_{t,i}^{rp}=\mathbf{v}_{rp}^{\top}\mathrm{tanh}\left(\mathrm{MLP}\left(\left[\mathbf{h}_{t}^{R};\mathbf{h}_{i}^{P}\right]\right)\right).$$
(6)
The detailed calculation and updating scheme of the decoding state
(7)
$\mathbf{h}_{t}^{R}$ at time step $t$ will be described in Section 3.5.4. Based on the attention mechanism, the personalized post representation c$_t$ can
3.2 History Encoder and General User Profile
Based on our observation, there is a large amount of personalized information in the user's historical responses. For example, a fan of cricket may talk a lot about cricket topics with others. Furthermore, different users can hold different speaking styles, such as enjoying speaking slang. Therefore, our first idea is to devise a model for learning such personalized language information in a general view.
Inspired by the strong ability of Transformer [33] to aggregate context and model sequences, we use a Transformer encoder to learn contextual representations of historical responses. In particular, we first add special tokens and concatenates all $n$ historical responses as $R^{w}= [ \langle$CLS$\rangle ; R_{1}; \langle$SEP$\rangle ; \cdots ; R_{n}; \langle$SEP$\rangle ] $, where each response $R_{i}=(r_{1}^{i},\cdots,r_{{L_{R_{i}}}}^{i})$ contains $L_{R_{i}}$ words, and [;] is the concatenation operation. A“\SEP〉”token is added at the tail of each response for segmentation, while a“(CLS)”token is added at the sequence head for summary. Then we map all words and special tokens into embeddings, and learn their contextual representations:
(3)
$$\begin{aligned}\left[\mathbf{e}^{G};\mathbf{E}^{R}\right]&=\mathrm{Transformer}_{N}\left([\mathbf{e}_{\mathrm{CLS}};\mathbf{R}_{1};\mathbf{e}_{\mathrm{SEP}};\cdots;\mathbf{R}_{n};\mathbf{e}_{\mathrm{SEP}}]\right),\\\mathbf{R}_{i}&=[\mathbf{e}_{1}^{r_{i},\mathrm{fuse}};\cdots;\mathbf{e}_{L_{R_{i}}}^{r_{i},\mathrm{fuse}}],\end{aligned}$$
(4)
|
[
[
0.08660888671875,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.51708984375,
0.10418701171875,
0.9150390625,
0.10418701171875,
0.9150390625,
0.121826171875,
0.51708984375,
0.121826171875
],
[
0.0849609375,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.12249755859375,
0.9140625,
0.121826171875,
0.9140625,
0.1363525390625,
0.51708984375,
0.136962890625
],
[
0.08416748046875,
0.1351318359375,
0.48193359375,
0.1337890625,
0.48193359375,
0.1483154296875,
0.08416748046875,
0.149658203125
],
[
0.51953125,
0.1383056640625,
0.91259765625,
0.1383056640625,
0.91259765625,
0.1483154296875,
0.51953125,
0.1483154296875
],
[
0.0849609375,
0.149658203125,
0.48046875,
0.149658203125,
0.48046875,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.51708984375,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.0849609375,
0.1666259765625,
0.483642578125,
0.1666259765625,
0.483642578125,
0.179931640625,
0.0849609375,
0.179931640625
],
[
0.51708984375,
0.1641845703125,
0.91259765625,
0.1641845703125,
0.91259765625,
0.177978515625,
0.51708984375,
0.177978515625
],
[
0.0849609375,
0.179931640625,
0.4345703125,
0.179931640625,
0.4345703125,
0.1932373046875,
0.0849609375,
0.1932373046875
],
[
0.51953125,
0.179931640625,
0.91259765625,
0.179931640625,
0.91259765625,
0.1900634765625,
0.51953125,
0.1900634765625
],
[
0.515625,
0.1912841796875,
0.90625,
0.1925048828125,
0.90625,
0.20703125,
0.515625,
0.205810546875
],
[
0.0849609375,
0.199462890625,
0.48193359375,
0.199462890625,
0.48193359375,
0.21337890625,
0.0849609375,
0.21337890625
],
[
0.0849609375,
0.2139892578125,
0.339111328125,
0.21533203125,
0.339111328125,
0.2266845703125,
0.0849609375,
0.2259521484375
],
[
0.51708984375,
0.208984375,
0.91259765625,
0.208984375,
0.91259765625,
0.222900390625,
0.51708984375,
0.222900390625
],
[
0.101318359375,
0.2279052734375,
0.47802734375,
0.2279052734375,
0.47802734375,
0.2412109375,
0.101318359375,
0.2412109375
],
[
0.5185546875,
0.2247314453125,
0.865234375,
0.2247314453125,
0.865234375,
0.23486328125,
0.5185546875,
0.23486328125
],
[
0.08660888671875,
0.2412109375,
0.48046875,
0.2412109375,
0.48046875,
0.255126953125,
0.08660888671875,
0.255126953125
],
[
0.53173828125,
0.236083984375,
0.9111328125,
0.23486328125,
0.9111328125,
0.25,
0.53173828125,
0.250732421875
],
[
0.08660888671875,
0.255615234375,
0.48193359375,
0.255615234375,
0.48193359375,
0.26953125,
0.08660888671875,
0.26953125
],
[
0.51708984375,
0.250732421875,
0.91259765625,
0.250732421875,
0.91259765625,
0.264404296875,
0.51708984375,
0.264404296875
],
[
0.0849609375,
0.26953125,
0.48046875,
0.26953125,
0.48046875,
0.28271484375,
0.0849609375,
0.28271484375
],
[
0.51708984375,
0.264404296875,
0.91259765625,
0.264404296875,
0.91259765625,
0.279052734375,
0.51708984375,
0.279052734375
],
[
0.0849609375,
0.28271484375,
0.48046875,
0.28271484375,
0.48046875,
0.296630859375,
0.0849609375,
0.296630859375
],
[
0.5185546875,
0.279052734375,
0.9140625,
0.279052734375,
0.9140625,
0.292236328125,
0.5185546875,
0.292236328125
],
[
0.0849609375,
0.296630859375,
0.48193359375,
0.296630859375,
0.48193359375,
0.31005859375,
0.0849609375,
0.31005859375
],
[
0.51708984375,
0.292236328125,
0.9140625,
0.292236328125,
0.9140625,
0.30615234375,
0.51708984375,
0.30615234375
],
[
0.0849609375,
0.311279296875,
0.483642578125,
0.311279296875,
0.483642578125,
0.324462890625,
0.0849609375,
0.324462890625
],
[
0.51708984375,
0.318115234375,
0.77392578125,
0.318115234375,
0.77392578125,
0.33203125,
0.51708984375,
0.33203125
],
[
0.08660888671875,
0.326904296875,
0.48046875,
0.326904296875,
0.48046875,
0.337158203125,
0.08660888671875,
0.337158203125
],
[
0.08660888671875,
0.33837890625,
0.48193359375,
0.33837890625,
0.48193359375,
0.3515625,
0.08660888671875,
0.3515625
],
[
0.51708984375,
0.337158203125,
0.91259765625,
0.337158203125,
0.91259765625,
0.35107421875,
0.51708984375,
0.35107421875
],
[
0.08660888671875,
0.3515625,
0.4794921875,
0.3515625,
0.4794921875,
0.365478515625,
0.08660888671875,
0.365478515625
],
[
0.5185546875,
0.3515625,
0.91259765625,
0.3515625,
0.91259765625,
0.365478515625,
0.5185546875,
0.365478515625
],
[
0.0849609375,
0.367431640625,
0.4794921875,
0.367431640625,
0.4794921875,
0.380615234375,
0.0849609375,
0.380615234375
],
[
0.51708984375,
0.3662109375,
0.9140625,
0.3662109375,
0.9140625,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.0849609375,
0.3818359375,
0.48193359375,
0.3818359375,
0.48193359375,
0.395751953125,
0.0849609375,
0.395751953125
],
[
0.5185546875,
0.380615234375,
0.91259765625,
0.380615234375,
0.91259765625,
0.391357421875,
0.5185546875,
0.391357421875
],
[
0.08660888671875,
0.397705078125,
0.22216796875,
0.397705078125,
0.22216796875,
0.40771484375,
0.08660888671875,
0.40771484375
],
[
0.51708984375,
0.393310546875,
0.91259765625,
0.393310546875,
0.91259765625,
0.4072265625,
0.51708984375,
0.4072265625
],
[
0.101318359375,
0.4091796875,
0.48046875,
0.4091796875,
0.48046875,
0.42236328125,
0.101318359375,
0.42236328125
],
[
0.51953125,
0.4091796875,
0.9111328125,
0.4091796875,
0.9111328125,
0.419189453125,
0.51953125,
0.419189453125
],
[
0.0849609375,
0.423583984375,
0.48046875,
0.423583984375,
0.48046875,
0.4375,
0.0849609375,
0.4375
],
[
0.5185546875,
0.42236328125,
0.91259765625,
0.42236328125,
0.91259765625,
0.43310546875,
0.5185546875,
0.43310546875
],
[
0.08660888671875,
0.439453125,
0.48046875,
0.439453125,
0.48046875,
0.449462890625,
0.08660888671875,
0.449462890625
],
[
0.515625,
0.434326171875,
0.9140625,
0.43505859375,
0.9140625,
0.448974609375,
0.515625,
0.447509765625
],
[
0.0849609375,
0.45068359375,
0.483642578125,
0.45068359375,
0.483642578125,
0.464111328125,
0.0849609375,
0.464111328125
],
[
0.5185546875,
0.448974609375,
0.9140625,
0.448974609375,
0.9140625,
0.462158203125,
0.5185546875,
0.462158203125
],
[
0.0849609375,
0.46533203125,
0.48193359375,
0.46533203125,
0.48193359375,
0.479248046875,
0.0849609375,
0.479248046875
],
[
0.5185546875,
0.464111328125,
0.91259765625,
0.464111328125,
0.91259765625,
0.474853515625,
0.5185546875,
0.474853515625
],
[
0.0849609375,
0.479736328125,
0.438720703125,
0.479736328125,
0.438720703125,
0.49365234375,
0.0849609375,
0.49365234375
],
[
0.51708984375,
0.47607421875,
0.9140625,
0.47607421875,
0.9140625,
0.4892578125,
0.51708984375,
0.4892578125
],
[
0.51708984375,
0.490478515625,
0.9140625,
0.490478515625,
0.9140625,
0.50390625,
0.51708984375,
0.50390625
],
[
0.1282958984375,
0.49755859375,
0.441162109375,
0.49755859375,
0.441162109375,
0.51220703125,
0.1282958984375,
0.51220703125
],
[
0.463134765625,
0.5,
0.4794921875,
0.5,
0.4794921875,
0.5087890625,
0.463134765625,
0.5087890625
],
[
0.5185546875,
0.505859375,
0.91259765625,
0.505859375,
0.91259765625,
0.5166015625,
0.5185546875,
0.5166015625
],
[
0.08660888671875,
0.5166015625,
0.48193359375,
0.5166015625,
0.48193359375,
0.52978515625,
0.08660888671875,
0.52978515625
],
[
0.51708984375,
0.517578125,
0.9140625,
0.517578125,
0.9140625,
0.53076171875,
0.51708984375,
0.53076171875
],
[
0.08660888671875,
0.52978515625,
0.48193359375,
0.52978515625,
0.48193359375,
0.54345703125,
0.08660888671875,
0.54345703125
],
[
0.51708984375,
0.5322265625,
0.91259765625,
0.5322265625,
0.91259765625,
0.54541015625,
0.51708984375,
0.54541015625
],
[
0.0849609375,
0.54541015625,
0.378173828125,
0.54541015625,
0.378173828125,
0.5595703125,
0.0849609375,
0.5595703125
],
[
0.5185546875,
0.54736328125,
0.91259765625,
0.54736328125,
0.91259765625,
0.55810546875,
0.5185546875,
0.55810546875
],
[
0.098876953125,
0.5595703125,
0.4794921875,
0.5595703125,
0.4794921875,
0.57275390625,
0.098876953125,
0.57275390625
],
[
0.51708984375,
0.56103515625,
0.5810546875,
0.56103515625,
0.5810546875,
0.5712890625,
0.51708984375,
0.5712890625
],
[
0.0849609375,
0.57373046875,
0.303955078125,
0.57373046875,
0.303955078125,
0.5869140625,
0.0849609375,
0.5869140625
],
[
0.53173828125,
0.57275390625,
0.9140625,
0.57275390625,
0.9140625,
0.5859375,
0.53173828125,
0.5859375
],
[
0.5185546875,
0.58837890625,
0.91259765625,
0.58837890625,
0.91259765625,
0.5986328125,
0.5185546875,
0.5986328125
],
[
0.0849609375,
0.599609375,
0.463134765625,
0.599609375,
0.463134765625,
0.61279296875,
0.0849609375,
0.61279296875
],
[
0.51953125,
0.60302734375,
0.9111328125,
0.60302734375,
0.9111328125,
0.61279296875,
0.51953125,
0.61279296875
],
[
0.0849609375,
0.6171875,
0.48046875,
0.6171875,
0.48046875,
0.630859375,
0.0849609375,
0.630859375
],
[
0.5185546875,
0.6162109375,
0.9111328125,
0.6162109375,
0.9111328125,
0.626953125,
0.5185546875,
0.626953125
],
[
0.0849609375,
0.6318359375,
0.483642578125,
0.6318359375,
0.483642578125,
0.64599609375,
0.0849609375,
0.64599609375
],
[
0.5185546875,
0.6298828125,
0.9111328125,
0.6298828125,
0.9111328125,
0.64013671875,
0.5185546875,
0.64013671875
],
[
0.08660888671875,
0.64794921875,
0.48046875,
0.64794921875,
0.48046875,
0.65771484375,
0.08660888671875,
0.65771484375
],
[
0.51708984375,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.51708984375,
0.65576171875
],
[
0.08416748046875,
0.65771484375,
0.48193359375,
0.6591796875,
0.48193359375,
0.673828125,
0.08416748046875,
0.67236328125
],
[
0.51708984375,
0.6572265625,
0.90625,
0.6572265625,
0.90625,
0.671875,
0.51708984375,
0.671875
],
[
0.08660888671875,
0.67236328125,
0.48046875,
0.67236328125,
0.48046875,
0.68603515625,
0.08660888671875,
0.68603515625
],
[
0.56884765625,
0.67822265625,
0.86279296875,
0.67822265625,
0.86279296875,
0.69873046875,
0.56884765625,
0.69873046875
],
[
0.08660888671875,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.69970703125,
0.08660888671875,
0.69970703125
],
[
0.896484375,
0.6845703125,
0.9111328125,
0.6845703125,
0.9111328125,
0.693359375,
0.896484375,
0.693359375
],
[
0.1004638671875,
0.69970703125,
0.48046875,
0.70068359375,
0.48046875,
0.71533203125,
0.1004638671875,
0.7138671875
],
[
0.51708984375,
0.70068359375,
0.8798828125,
0.70068359375,
0.8798828125,
0.7138671875,
0.51708984375,
0.7138671875
],
[
0.0849609375,
0.7138671875,
0.48193359375,
0.7138671875,
0.48193359375,
0.72802734375,
0.0849609375,
0.72802734375
],
[
0.53515625,
0.71630859375,
0.91259765625,
0.71630859375,
0.91259765625,
0.72998046875,
0.53515625,
0.72998046875
],
[
0.08660888671875,
0.72998046875,
0.4794921875,
0.72998046875,
0.4794921875,
0.74072265625,
0.08660888671875,
0.74072265625
],
[
0.51708984375,
0.72900390625,
0.9140625,
0.72998046875,
0.9140625,
0.744140625,
0.51708984375,
0.74365234375
],
[
0.08416748046875,
0.7421875,
0.48193359375,
0.7412109375,
0.48193359375,
0.7548828125,
0.08416748046875,
0.755859375
],
[
0.51708984375,
0.7421875,
0.9150390625,
0.74365234375,
0.9150390625,
0.75830078125,
0.51708984375,
0.7568359375
],
[
0.0849609375,
0.755859375,
0.48193359375,
0.755859375,
0.48193359375,
0.76953125,
0.0849609375,
0.76953125
],
[
0.69921875,
0.7646484375,
0.7197265625,
0.7646484375,
0.7197265625,
0.77978515625,
0.69921875,
0.77978515625
],
[
0.0849609375,
0.77099609375,
0.483642578125,
0.77099609375,
0.483642578125,
0.78515625,
0.0849609375,
0.78515625
],
[
0.70361328125,
0.77880859375,
0.75732421875,
0.77294921875,
0.76123046875,
0.79150390625,
0.70751953125,
0.79736328125
],
[
0.89453125,
0.77783203125,
0.9150390625,
0.77783203125,
0.9150390625,
0.79248046875,
0.89453125,
0.79248046875
],
[
0.70166015625,
0.78271484375,
0.71142578125,
0.78271484375,
0.71142578125,
0.787109375,
0.70166015625,
0.787109375
],
[
0.6708984375,
0.77978515625,
0.69140625,
0.77978515625,
0.69140625,
0.79052734375,
0.6708984375,
0.79052734375
],
[
0.0849609375,
0.787109375,
0.48193359375,
0.787109375,
0.48193359375,
0.80029296875,
0.0849609375,
0.80029296875
],
[
0.69775390625,
0.79248046875,
0.7138671875,
0.7890625,
0.71826171875,
0.80029296875,
0.70166015625,
0.8037109375
],
[
0.0849609375,
0.8017578125,
0.48193359375,
0.8017578125,
0.48193359375,
0.81494140625,
0.0849609375,
0.81494140625
],
[
0.5888671875,
0.81005859375,
0.62353515625,
0.81005859375,
0.62353515625,
0.82275390625,
0.5888671875,
0.82275390625
],
[
0.08416748046875,
0.81640625,
0.48193359375,
0.814453125,
0.48193359375,
0.82763671875,
0.08416748046875,
0.830078125
],
[
0.53515625,
0.82568359375,
0.55810546875,
0.82568359375,
0.55810546875,
0.833984375,
0.53515625,
0.833984375
],
[
0.701171875,
0.82080078125,
0.86376953125,
0.82080078125,
0.86376953125,
0.83544921875,
0.701171875,
0.83544921875
],
[
0.0849609375,
0.8291015625,
0.483642578125,
0.8291015625,
0.483642578125,
0.84228515625,
0.0849609375,
0.84228515625
],
[
0.5703125,
0.83203125,
0.650390625,
0.83203125,
0.650390625,
0.849609375,
0.5703125,
0.849609375
],
[
0.89794921875,
0.8271484375,
0.90673828125,
0.8271484375,
0.90673828125,
0.83203125,
0.89794921875,
0.83203125
],
[
0.101318359375,
0.8466796875,
0.483642578125,
0.8486328125,
0.483642578125,
0.8671875,
0.101318359375,
0.86572265625
],
[
0.5185546875,
0.85400390625,
0.91259765625,
0.85400390625,
0.91259765625,
0.8671875,
0.5185546875,
0.8671875
],
[
0.51708984375,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.136474609375,
0.875,
0.214111328125,
0.8720703125,
0.2156982421875,
0.88623046875,
0.1380615234375,
0.8896484375
],
[
0.461669921875,
0.87548828125,
0.483642578125,
0.87548828125,
0.483642578125,
0.88818359375,
0.461669921875,
0.88818359375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"GRU (BiGRU). To make the post encoder aware of the user's person-",
"where e'ifuse is the sum of the word embedding, segment embed-",
"alized information, we use the general user profile eG for initializa-",
"ding,and position embedding of the j-th word. These three embed-",
"tion. Consequently, the post X = (xi,...,xLx) is represented by",
"dingsareused ina similarwaylikeBERT[5].Therepresentation",
"hidden state sequence hP - (hP,.+-,hP.). These representations",
"of \"(CLS)\" token (eG) summarizes the information in the entire",
"will be dynamically aggregated as a personalized post representa-",
"historical responses, thus we call it the general user profile. ER con-",
"tion c by an attention mechanism at the decoding step t.",
"tains contextual representations ofwords inhistorical responses",
"For Ri,we denote its contextual representations as (e'f,.*,ei.",
"3.1.2 Decoder. In the decoder side, DHAP incorporates the per",
"sonalized information intwoperspectives",
"Transformern() is a N-layer bidirectional Transformer encoder",
"Part (3): User history memory and dynamic post-aware",
"identical totheoriginal implementationdescribed in[33].",
"user profile. In the general user profile, all historical responses",
"With the history encoder, we obtain: (1)a general user profile eG",
"are considered in a general view. However, for a specific input post",
"which summarizes all historical responses and contains personal-",
"X, the historical responses may play different roles. Intuitively",
"ized information of the user; and (2) the contextual representations",
"the historical responses relevant to X are more valuable than the",
"of words in the historical responses ER.These representations will",
"irrelevant ones. To highlight the historical responses relevant to",
"be used to build the dynamic post-aware user profile in Section 3.4",
"the input post, DHAP builds a key-value memory of historical post-",
"3.3PersonalizedPostEncoder",
"responsepairs ((Pi,R),...,(Pn,Rn)).Given ct as a personalized",
"representation of X,DHAP builds a dynamic post-aware user profile",
"In real-world applications, the posts are often very short, even",
"eD by selecting and aggregating the historical responses from the",
"ambiguous. Thus, building accurate encoding of the input post",
"user history memory. As eD is dynamic with different input posts X",
"is difficult, which further leads to poor quality of the generated",
"and summarizes related user history information, we call it dynamic",
"responses[9,28].Fortunately,withpersonalizedbackgroundknowl-",
"post-awareuserprofile",
"edge, the chatbot can get more input information and is promising",
"Part (4): Personalized decoder. Finally, the personalized post",
"tobetter capturethesemanticinformationofthepost.Letus usean",
"representation ct, general user profle e, and dynamic post-aware",
"exampleto explain this:given apost\"Thenew MACis sobeautiful\"",
"userprofileeparefusedtogethertodecodetheresponse sequen",
"different users may have different understandings.For a program-",
"tially. Inspired by CopyNet [7], DHAP can switch between gener-",
"mer, \"MAC\" may refer to the Apple's laptop; but for a fashion girl,",
"ating a word from a generic vocabulary (p(yr|mg) and copying a",
"she may associate\"MAC\"with the lipstick.Therefore, the user's",
"word from a user's personalized vocabulary (p(yt/mc) as:",
"history can help distinguish the word \"MAC\" and provide more",
"background knowledge for the input post, so that the post rep-",
"p(ytly<t,X,H) = p(mg)p(yt/mg) + p(mc)p(ytmc),",
"(2)",
"resentation can be significantly enhanced.We call this encoder",
"where p(mg) and p(m) are computed by our designed decoding",
"\"personalized post encoder\"- the encoded representations of the",
"switcher.p(yt|mg) and p(yt|mc) are calculated based on same in-",
"same post can be different for various users, because these users",
"puts(e.g., ct, eG and eD) with different functions.",
"mayhavedifferentprofiles and different understandingsof the",
" In the remaining part of this section, we will introduce these",
"samepost",
"four components of DHAP in detail.",
"Specifically, DHAP employs a BiGRU to encode the input post.",
"Herewe chooseRNN-basedarchitecturesbecausetheyarebetter",
"3.2 History Encoder and General UserProfile",
"atcapturinglocalcorrelationsandencodingpositionalinforma",
"Based on our observation, there is a large amount of personalized",
"tion than Transformers for short texts[19, 39].To facilitate the",
"information in the user's historical responses. For example, a fan of",
"encoderwithpersonalized information,DHAP uses thegeneral",
"cricket maytalk alot about crickettopicswith others.Furthermore",
"user profile eG to initialize hidden states of BiGRU. Given the post",
"different users can hold different speaking styles, such as enjoying",
"X = (xi,**,xLx), its representations (hf.**,h) are built as:",
"speaking slang. Therefore, our first idea is to devise a model for",
"hP - BiGRU(h-1x),h = ReLU(MLP(eG),",
"learningsuchpersonalizedlanguageinformationinageneralview",
"(5)",
"Inspired by the strong ability of Transformer [33] to aggregate",
"where x, is the embedding obtained by the embedding table.",
"context and model sequences, we use a Transformer encoder to",
"These hidden states draw personalized information from the",
"learn contextual representations of historicalresponses.Inpartic",
"general user profile, and in the decoding phase, they are aggregated",
"ular, we first add special tokens and concatenates all n historical",
"by the decoding hidden state hR through an attention mechanism:",
"responses as Rw = [(CLS); Ri; (SEP); - ; Rn; (SEP)], where each",
"Lx",
"response R, = (ri....,rLr.) contains LR, words, and [] is the con-",
"Zarihp",
"(6)",
"2",
"Ct:",
"catenation operation. A \"(SEP)\"token is added at the tail of each",
"i=1",
"response for segmentation, while a \"(CLS)\" token is added at the",
"exp(s",
"sequence head for summary.Then we map all words and special",
"αt.i",
"= vI,tanh (MLP (hR: hP)",
"tokens into embeddings, and learn their contextual representations:",
"Efxi exp(st)",
"",
"[eG;ER=TransformerN ([ecLs;Ri;esEP;..;Rn;esEPl), (3)",
"The detailed calculation and updating scheme of the decoding state",
"hR at time step t will be described in Section 3.5.4. Based on the",
"R, = [erifuse",
"(4)",
"attention mechanism, the personalized post representation ct can"
] |
{
"arxiv_id": "2108.09355",
"language": "en",
"page_idx": 3
}
|
||
2108.09355-en-4
|
2108.09355-en-4.jpg
|
dynamically focus on some important words of the post according to the current decoding state. In the next section, DHAP will use c$_t$ to build the dynamic post-aware user profile.
## 3.4 User History Memory and Dynamic Post-aware User Profile
With the general user profile, DHAP can capture personalized information of a user in a general view. However, when faced with a new input post, the historical data may play different roles. For example, when a crazy fan of cricket meets a post about cricket, they will be talkative and post a lot of things. But when they face other daily topics, they may behave much more gently. Hence, it is valuable to dynamically select the information that is most relevant to the input, and the chatbot can behave differently as the input varies. Following this idea, we propose to dynamically aggregate historical responses that are highly relevant to the current input post, and leverage them as a reference to drive the response generation.
To achieve this, DHAP uses a key-value memory network[18] to store the user's historical post-response pairs. Then the personalized post representation is used as the query to select highly related keys (historical posts) from the user history memory, and their corresponding values (historical responses) are aggregated as the dynamic post-aware user profile.
3.4.1 User History Memory. We firstly transform the historical
post-response pairs into key-value pairs and build the memory.
As we discussed earlier, the historical posts are usually issued by different users. Thus the language style and topic of them may be various. Under this circumstance, it is more reasonable to treat them independently, so DHAP applies a BiGRU to represent each historical post, respectively. In our implementation, this BiGRU shares parameters with the personalized post encoder (introduced in Section 3.3). Consider the $i$-th historical post $P_i$, its representation is computed by a summing pooling over the word dimension as p$_i^{c}=\sum_{j=1}^{L_{P_{i}}}\mathbf{h}_{j}^{i}$,where h$_j^{i}$ is the hidden state of the BiGRU for the $j$-th word in $P_i.$ For all historical posts $(P_1,\cdots,P_{n})$,their representations are denoted as $(\mathbf{p}_{1}^{c},\cdots,\mathbf{p}_{n}^{c}).$ Similarly, the representation of historical responses is also computed by the same pooling strat- $\text{egy as r}_{i}^{c}= \sum _{j= 1}^{L_{R_{i}}}\mathbf{e} _{j}^{r_{i}}, \mathrm{where~e}_{j}^{r_{i}}$ is the contextual representation of the $j$- th word in $R_{i}.$ Different from historical posts, the the $j$-th word in $R_i.$ Different from historical posts, the contextual word representations of historical responses are obtained by the history encoder in Equation (3). Thus, all historical responses are represented as $(\mathbf{r}_1^c,\cdots,\mathbf{r}_n^c).$
Finally, we build the user history memory by using the post representations as key and the corresponding response representations as value, $i.e.,\left\{\mathbf{k}_{1}^{M}:\mathbf{v}_{1}^{M},\cdots,\mathbf{k}_{n}^{M}:\mathbf{v}_{n}^{M}\right\}=\left\{\mathbf{p}_{1}^{c}:\mathbf{r}_{1}^{c},\cdots,\mathbf{p}_{n}^{c}:\mathbf{r}_{n}^{c}\right\}.$
3.4.2 Dynamic Post-aware User Profile. After building the user
history memory, DHAP can select and aggregate the most relevant historical responses and build the dynamic post-aware user profile based on the input post. Specifically, the personalized representation of the input post obtained in Equation (6) is used as the query and attends to the memory keys to find the most relevant historical posts. The relevance is measured by the attention weights. Then the corresponding historical responses are summed up based on
the normalized weights to construct the dynamic profile e$\mathbf{e}_t^D:$
$$\mathbf{e}_t^D=\sum_{i=1}^n\beta_{t,i}\mathbf{v}_i^M,$$
(8)
where $\beta_{t,i}$ is the attention weight of the $i$-th historical response based on the personalized post representation c$_t$ in a similar way like Equation (7). Note that the dynamic post-aware user profile is computed at each decoding time step $t.$ The most relevant information hidden in historical responses can thus be selected to help response generation.
# 3.5 Personalized Decoder
Finally, the generation probability of responses can be calculated by the personalized post representation c$_t$,the general user profile e$^G$,and the dynamic post-aware user profile $\mathbf{e}_t^D.$ Inspired by CopyNet [7], in addition to leveraging the personalized information captured by the implicit user profile, we construct a personalized vocabulary so that the model is allowed to directly select personalized words that the user frequently used in history.
Specifically, the probability of the word $y_t$ generated in the personalized decoder is computed as Equation (2), where $p\left(m_{g}\right)$ is the probability of general decoding mode and $p\left(m_{c}\right)$is the probability of copy decoding mode. They are computed by our proposed decoding switcher. $p\left(y_{t}|m_{g}\right)$ and $p\left(y_{t}|m_{c}\right)$ are the probabilities of generating $y_t$ under two modes, respectively.
It is worth noting that:(1) The switching probability $p\left(m_{g}\right)$ and $p\left(m_{c}\right)$are both in$\left[0,1\right].$Thus, it is a“soft”decoding switcher.$\left(2\right)$ The generic vocabulary also contains words in the personalized vocabulary. The generation probability of $y_t$ is obtained by the sum of probabilities under two decoding modes. Therefore, DHAP is just biased to the personalized words rather than lost in it.
3.5.1 Decoding switcher. The decoding switcher determines the probability of the two decoding modes, $i.e.$,predicting a word from the generic vocabulary to maintain sentence fluency, or copying a word directly from the personalized vocabulary to make the response more informative and personalized. Specifically, DHAP computes the switching probability based on the matching degree between the decoder state and the concatenation of the personalized post representation, general user profile, and dynamic post-aware user profile, and further calculates two decoding mode probabilities.
(9)
$$\begin{aligned}[p(m_{g}),p(m_{c})]&=\mathrm{Softmax}(\mathbf{d}_{t}),\\\mathbf{d}_{t}&=\mathrm{MLP}([\mathbf{h}_{t}^{R};\mathbf{c}_{t};\mathbf{e}^{G};\mathbf{e}_{t}^{D}]),\end{aligned}$$
(10)
where h$_t^R$ is the decoding hidden state at step $t$,and d$_t\in\mathbb{R}^2$ is the matching degree vector to estimate the two mode probabilities. The softmax function guarantees that $p(m_{g})+p(m_{c})=1.$
## 3.5.2 Personalized general decoding. While general decoding, the decoder should predict a word $y_t$ from the generic vocabulary:
$$p(y_{t}|m_{g})=\mathrm{Softmax}(\mathrm{MLP}([\mathbf{h}_{t}^{R};\mathbf{c}_{t};\mathbf{e}^{G};\mathbf{e}_{t}^{D}])).$$
(11)
3.5.3 Personalized copy decoding. The personalized vocabulary of user $u$ is composed of the words that appear in their historical responses. DHAP can directly select a word from this vocabulary
|
[
[
0.08660888671875,
0.106689453125,
0.48046875,
0.106689453125,
0.48046875,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.51708984375,
0.106689453125,
0.87744140625,
0.106689453125,
0.87744140625,
0.12060546875,
0.51708984375,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.373291015625,
0.1351318359375,
0.373291015625,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.70361328125,
0.136962890625,
0.71728515625,
0.136962890625,
0.71728515625,
0.146484375,
0.70361328125,
0.146484375
],
[
0.71728515625,
0.141357421875,
0.76123046875,
0.141357421875,
0.76123046875,
0.1558837890625,
0.71728515625,
0.1558837890625
],
[
0.8935546875,
0.141357421875,
0.9140625,
0.141357421875,
0.9140625,
0.1558837890625,
0.8935546875,
0.1558837890625
],
[
0.08660888671875,
0.17041015625,
0.416748046875,
0.17041015625,
0.416748046875,
0.184326171875,
0.08660888671875,
0.184326171875
],
[
0.67236328125,
0.143310546875,
0.6787109375,
0.143310546875,
0.6787109375,
0.1483154296875,
0.67236328125,
0.1483154296875
],
[
0.5185546875,
0.1749267578125,
0.9111328125,
0.1749267578125,
0.9111328125,
0.1849365234375,
0.5185546875,
0.1849365234375
],
[
0.12744140625,
0.1875,
0.318603515625,
0.1875,
0.318603515625,
0.2008056640625,
0.12744140625,
0.2008056640625
],
[
0.51708984375,
0.1881103515625,
0.91259765625,
0.1881103515625,
0.91259765625,
0.2020263671875,
0.51708984375,
0.2020263671875
],
[
0.08416748046875,
0.20263671875,
0.48193359375,
0.203857421875,
0.48193359375,
0.2183837890625,
0.08416748046875,
0.2171630859375
],
[
0.51708984375,
0.2020263671875,
0.91259765625,
0.2020263671875,
0.91259765625,
0.21533203125,
0.51708984375,
0.21533203125
],
[
0.0849609375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.51708984375,
0.216552734375,
0.91259765625,
0.216552734375,
0.91259765625,
0.2298583984375,
0.51708984375,
0.2298583984375
],
[
0.0849609375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.51708984375,
0.2298583984375,
0.9140625,
0.2298583984375,
0.9140625,
0.24365234375,
0.51708984375,
0.24365234375
],
[
0.0849609375,
0.246826171875,
0.48193359375,
0.246826171875,
0.48193359375,
0.260009765625,
0.0849609375,
0.260009765625
],
[
0.51708984375,
0.24560546875,
0.6435546875,
0.2442626953125,
0.6435546875,
0.255615234375,
0.51708984375,
0.2568359375
],
[
0.0849609375,
0.260009765625,
0.48046875,
0.260009765625,
0.48046875,
0.2734375,
0.0849609375,
0.2734375
],
[
0.08660888671875,
0.27587890625,
0.4794921875,
0.27587890625,
0.4794921875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.5185546875,
0.271484375,
0.734375,
0.271484375,
0.734375,
0.285888671875,
0.5185546875,
0.285888671875
],
[
0.08660888671875,
0.2890625,
0.48046875,
0.2890625,
0.48046875,
0.2998046875,
0.08660888671875,
0.2998046875
],
[
0.5185546875,
0.290283203125,
0.9140625,
0.290283203125,
0.9140625,
0.3037109375,
0.5185546875,
0.3037109375
],
[
0.08660888671875,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.08660888671875,
0.31494140625
],
[
0.5185546875,
0.30615234375,
0.9111328125,
0.30615234375,
0.9111328125,
0.31640625,
0.5185546875,
0.31640625
],
[
0.0849609375,
0.314453125,
0.48046875,
0.314453125,
0.48046875,
0.32763671875,
0.0849609375,
0.32763671875
],
[
0.5185546875,
0.317626953125,
0.91259765625,
0.317626953125,
0.91259765625,
0.330810546875,
0.5185546875,
0.330810546875
],
[
0.08660888671875,
0.330810546875,
0.48046875,
0.330810546875,
0.48046875,
0.341552734375,
0.08660888671875,
0.341552734375
],
[
0.51708984375,
0.33203125,
0.9140625,
0.33203125,
0.9140625,
0.34521484375,
0.51708984375,
0.34521484375
],
[
0.0849609375,
0.343505859375,
0.4599609375,
0.343505859375,
0.4599609375,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.5185546875,
0.347900390625,
0.91259765625,
0.347900390625,
0.91259765625,
0.35791015625,
0.5185546875,
0.35791015625
],
[
0.101318359375,
0.356689453125,
0.48046875,
0.356689453125,
0.48046875,
0.37060546875,
0.101318359375,
0.37060546875
],
[
0.51708984375,
0.359130859375,
0.9140625,
0.359130859375,
0.9140625,
0.37255859375,
0.51708984375,
0.37255859375
],
[
0.08660888671875,
0.37255859375,
0.48193359375,
0.37255859375,
0.48193359375,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
0.51708984375,
0.373779296875,
0.79736328125,
0.373779296875,
0.79736328125,
0.386962890625,
0.51708984375,
0.386962890625
],
[
0.0849609375,
0.3837890625,
0.48193359375,
0.3837890625,
0.48193359375,
0.397705078125,
0.0849609375,
0.397705078125
],
[
0.53173828125,
0.386474609375,
0.91259765625,
0.386474609375,
0.91259765625,
0.399658203125,
0.53173828125,
0.399658203125
],
[
0.0849609375,
0.3984375,
0.48046875,
0.3984375,
0.48046875,
0.412353515625,
0.0849609375,
0.412353515625
],
[
0.515625,
0.399658203125,
0.91259765625,
0.40087890625,
0.91259765625,
0.415283203125,
0.515625,
0.4140625
],
[
0.0849609375,
0.41357421875,
0.48046875,
0.41357421875,
0.48046875,
0.4267578125,
0.0849609375,
0.4267578125
],
[
0.51708984375,
0.415283203125,
0.9140625,
0.415283203125,
0.9140625,
0.4287109375,
0.51708984375,
0.4287109375
],
[
0.0849609375,
0.4267578125,
0.285888671875,
0.4267578125,
0.285888671875,
0.43994140625,
0.0849609375,
0.43994140625
],
[
0.51953125,
0.43115234375,
0.91259765625,
0.43115234375,
0.91259765625,
0.44140625,
0.51953125,
0.44140625
],
[
0.515625,
0.442626953125,
0.9140625,
0.44140625,
0.9140625,
0.455810546875,
0.515625,
0.45703125
],
[
0.0849609375,
0.455810546875,
0.48046875,
0.455810546875,
0.48046875,
0.4697265625,
0.0849609375,
0.4697265625
],
[
0.51708984375,
0.45703125,
0.79150390625,
0.45703125,
0.79150390625,
0.47021484375,
0.51708984375,
0.47021484375
],
[
0.08660888671875,
0.4716796875,
0.464111328125,
0.4716796875,
0.464111328125,
0.481689453125,
0.08660888671875,
0.481689453125
],
[
0.53173828125,
0.47021484375,
0.91259765625,
0.47021484375,
0.91259765625,
0.484130859375,
0.53173828125,
0.484130859375
],
[
0.10296630859375,
0.484130859375,
0.48193359375,
0.484130859375,
0.48193359375,
0.49755859375,
0.10296630859375,
0.49755859375
],
[
0.51708984375,
0.484130859375,
0.91259765625,
0.484130859375,
0.91259765625,
0.49755859375,
0.51708984375,
0.49755859375
],
[
0.0849609375,
0.49755859375,
0.48046875,
0.49755859375,
0.48046875,
0.51123046875,
0.0849609375,
0.51123046875
],
[
0.5185546875,
0.5,
0.9111328125,
0.5,
0.9111328125,
0.51025390625,
0.5185546875,
0.51025390625
],
[
0.08660888671875,
0.51220703125,
0.4794921875,
0.51220703125,
0.4794921875,
0.52294921875,
0.08660888671875,
0.52294921875
],
[
0.5185546875,
0.51318359375,
0.91259765625,
0.51318359375,
0.91259765625,
0.5234375,
0.5185546875,
0.5234375
],
[
0.0849609375,
0.5244140625,
0.48046875,
0.5244140625,
0.48046875,
0.53857421875,
0.0849609375,
0.53857421875
],
[
0.5185546875,
0.52587890625,
0.9140625,
0.52587890625,
0.9140625,
0.5390625,
0.5185546875,
0.5390625
],
[
0.0849609375,
0.53857421875,
0.48193359375,
0.53857421875,
0.48193359375,
0.55322265625,
0.0849609375,
0.55322265625
],
[
0.51708984375,
0.53857421875,
0.86767578125,
0.53857421875,
0.86767578125,
0.5517578125,
0.51708984375,
0.5517578125
],
[
0.08660888671875,
0.55322265625,
0.48046875,
0.55322265625,
0.48046875,
0.56640625,
0.08660888671875,
0.56640625
],
[
0.08660888671875,
0.568359375,
0.48046875,
0.568359375,
0.48046875,
0.5791015625,
0.08660888671875,
0.5791015625
],
[
0.5185546875,
0.5625,
0.91259765625,
0.5625,
0.91259765625,
0.57568359375,
0.5185546875,
0.57568359375
],
[
0.0849609375,
0.58056640625,
0.483642578125,
0.58056640625,
0.483642578125,
0.5947265625,
0.0849609375,
0.5947265625
],
[
0.51708984375,
0.5771484375,
0.9140625,
0.5771484375,
0.9140625,
0.59033203125,
0.51708984375,
0.59033203125
],
[
0.08416748046875,
0.59521484375,
0.15283203125,
0.59765625,
0.15283203125,
0.6123046875,
0.08416748046875,
0.60986328125
],
[
0.5185546875,
0.59130859375,
0.9111328125,
0.59130859375,
0.9111328125,
0.6015625,
0.5185546875,
0.6015625
],
[
0.1446533203125,
0.59765625,
0.48193359375,
0.59521484375,
0.48193359375,
0.60986328125,
0.1446533203125,
0.6123046875
],
[
0.5185546875,
0.60595703125,
0.9111328125,
0.60595703125,
0.9111328125,
0.6162109375,
0.5185546875,
0.6162109375
],
[
0.08660888671875,
0.6123046875,
0.484375,
0.6123046875,
0.484375,
0.62548828125,
0.08660888671875,
0.62548828125
],
[
0.5185546875,
0.61865234375,
0.91259765625,
0.61865234375,
0.91259765625,
0.6318359375,
0.5185546875,
0.6318359375
],
[
0.08660888671875,
0.626953125,
0.48193359375,
0.626953125,
0.48193359375,
0.64013671875,
0.08660888671875,
0.64013671875
],
[
0.515625,
0.6318359375,
0.91259765625,
0.6318359375,
0.91259765625,
0.64599609375,
0.515625,
0.64599609375
],
[
0.0849609375,
0.64013671875,
0.48193359375,
0.64013671875,
0.48193359375,
0.65380859375,
0.0849609375,
0.65380859375
],
[
0.1650390625,
0.65283203125,
0.484375,
0.65478515625,
0.484375,
0.67236328125,
0.1650390625,
0.67041015625
],
[
0.5185546875,
0.64794921875,
0.91259765625,
0.64794921875,
0.91259765625,
0.65771484375,
0.5185546875,
0.65771484375
],
[
0.0849609375,
0.6572265625,
0.17236328125,
0.6572265625,
0.17236328125,
0.671875,
0.0849609375,
0.671875
],
[
0.51708984375,
0.6591796875,
0.9140625,
0.6591796875,
0.9140625,
0.67236328125,
0.51708984375,
0.67236328125
],
[
0.0849609375,
0.67236328125,
0.48193359375,
0.67236328125,
0.48193359375,
0.68603515625,
0.0849609375,
0.68603515625
],
[
0.51708984375,
0.673828125,
0.9140625,
0.673828125,
0.9140625,
0.6875,
0.51708984375,
0.6875
],
[
0.08660888671875,
0.68798828125,
0.4794921875,
0.68798828125,
0.4794921875,
0.69873046875,
0.08660888671875,
0.69873046875
],
[
0.59228515625,
0.693359375,
0.78173828125,
0.693359375,
0.78173828125,
0.7109375,
0.59228515625,
0.7109375
],
[
0.0849609375,
0.70068359375,
0.48193359375,
0.70068359375,
0.48193359375,
0.7138671875,
0.0849609375,
0.7138671875
],
[
0.89208984375,
0.69580078125,
0.91650390625,
0.69580078125,
0.91650390625,
0.71337890625,
0.89208984375,
0.71337890625
],
[
0.0849609375,
0.71533203125,
0.250732421875,
0.71533203125,
0.250732421875,
0.72900390625,
0.0849609375,
0.72900390625
],
[
0.6689453125,
0.71533203125,
0.83642578125,
0.71533203125,
0.83642578125,
0.72998046875,
0.6689453125,
0.72998046875
],
[
0.8857421875,
0.71533203125,
0.9150390625,
0.71533203125,
0.9150390625,
0.73193359375,
0.8857421875,
0.73193359375
],
[
0.101318359375,
0.7265625,
0.4794921875,
0.7265625,
0.4794921875,
0.74072265625,
0.101318359375,
0.74072265625
],
[
0.08660888671875,
0.7412109375,
0.4794921875,
0.7412109375,
0.4794921875,
0.7548828125,
0.08660888671875,
0.7548828125
],
[
0.51708984375,
0.74072265625,
0.91259765625,
0.7412109375,
0.91259765625,
0.755859375,
0.51708984375,
0.7548828125
],
[
0.0849609375,
0.755859375,
0.45751953125,
0.755859375,
0.45751953125,
0.77099609375,
0.0849609375,
0.77099609375
],
[
0.51708984375,
0.755859375,
0.91259765625,
0.7548828125,
0.91259765625,
0.76953125,
0.51708984375,
0.77099609375
],
[
0.51708984375,
0.771484375,
0.84228515625,
0.771484375,
0.84228515625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.0849609375,
0.78515625,
0.48046875,
0.78515625,
0.48046875,
0.798828125,
0.0849609375,
0.798828125
],
[
0.5185546875,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.5185546875,
0.806640625
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.0849609375,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.51708984375,
0.80810546875,
0.89453125,
0.80810546875,
0.89453125,
0.8212890625,
0.51708984375,
0.8212890625
],
[
0.08660888671875,
0.8291015625,
0.4794921875,
0.8291015625,
0.4794921875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.57666015625,
0.830078125,
0.8505859375,
0.8291015625,
0.8505859375,
0.8466796875,
0.57666015625,
0.84765625
],
[
0.88623046875,
0.83203125,
0.91259765625,
0.83203125,
0.91259765625,
0.84521484375,
0.88623046875,
0.84521484375
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.483642578125,
0.85498046875,
0.483642578125,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.5185546875,
0.85498046875,
0.91015625,
0.85498046875,
0.91015625,
0.86865234375,
0.5185546875,
0.86865234375
],
[
0.08660888671875,
0.87060546875,
0.4794921875,
0.87060546875,
0.4794921875,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.51708984375,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.48193359375,
0.88330078125,
0.48193359375,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"dynamically focus on some important words of the post according",
"the normalized weights to construct the dynamic profile eD",
"to the current decoding state. In the next section, DHAP will use",
"ct to build the dynamic post-aware user profile.",
"1",
"Pt,ivM",
"(8)",
"3.4 User History Memory and Dynamic",
"L",
"where B,;is the attention weight of the i-th historical response",
"Post-aware UserProfile",
"based on the personalized post representation ct in a similar way",
"With the general user profile,DHAP can capture personalized infor-",
"like Equation (7). Note that the dynamic post-aware user profile is",
"mation of a user in a general view. However, when faced with a new",
"computed at each decoding time step t. The most relevant infor",
"input post, the historical data may play different roles. For example",
"mation hidden in historical responses can thus be selected to help",
"when a crazy fan of cricket meets a post about cricket, they will be",
"response generation.",
"talkative and post a lot of things. But when they face other daily",
"topics,theymaybehavemuch moregently.Hence,it is valuable",
"3.5 PersonalizedDecoder",
"to dynamically select the information that ismost relevant tothe",
"Finally, the generation probability of responses can be calculated",
"input, and the chatbot can behave differently as the input varies.",
"bythepersonalizedpostrepresentationct,thegeneraluserpro",
"Following this idea, we propose to dynamically aggregate historical",
"file eG, and the dynamic post-aware user profile eD.Inspired by",
"responses that are highlyrelevant to the current input post, and",
"CopyNet [7], in addition to leveraging the personalized information",
"leverage them as a reference to drive the response generation.",
"capturedbytheimplicituserprofile,weconstructapersonalized vo-",
"To achieve this, DHAP uses a key-value memory network [18]",
"cabulary so that the model is allowed to directly select personalized",
"to store the user's historical post-response pairs.Then thepersonal-",
"words that the user frequently used in history.",
"ized post representation is used as the query to select highly related",
"Specifically, the probability of the word yt generated in the per",
"keys (historical posts) from the user history memory, and their",
"sonalized decoder is computed as Equation (2), where p (mg) is the",
"corresponding values (historical responses) are aggregated as the",
"probability of general decoding mode and p (mc) is the probabil-",
"dynamic post-aware user profile.",
"ityofcopydecodingmode.Theyarecomputedbyourproposed",
"decoding switcher. p (yt mg) and p (yt|mc) are the probabilities of",
"3.4.1 User History Memory. We firstly transform the historical",
"generating yt under two modes, respectively.",
"post-responsepairsintokey-valuepairsandbuildthememory",
"It is worth noting that: (1) The switching probability p (mg) and",
"As we discussed earlier, the historical posts are usually issued",
"p (mc) are both in [o, 1]. Thus, it is a \"soft\" decoding switcher. (2)",
"by different users. Thus the language style and topic of them may",
"Thegeneric vocabulary also contains words in the personalized",
"be various. Under this circumstance,it is more reasonable to treat",
"vocabulary.Thegenerationprobability of yris obtained bythesum",
"them independently, so DHAP applies a BiGRU to represent each",
"of probabilities under two decoding modes. Therefore, DHAP is",
"historical post, respectively. In our implementation, this BiGRU",
"just biased to the personalized words rather than lost in it.",
"shares parameters with the personalized post encoder (introduced",
"inSection3.3).Considerthei-thhistoricalpostPi,itsrepresenta",
"3.5.1 Decoding switcher.The decoding switcher determines the",
"tion is computed by a summing pooling over the word dimension",
"probability of the two decoding modes, i.e., predicting a word from",
"asp=",
"thegenericvocabularytomaintainsentencefluency,orcopying",
"-h',whereh', is thehidden stateof the BiGRUfor the",
"a word directly from the personalized vocabulary to make the",
"j-th word in Pj.For all historical posts (Pi,..,Pn),their represen-",
"response more informative and personalized. Specifically,DHAP",
"tations are denoted as (p,**,P). Similarly,the representation",
"computes the switching probability based on the matching degree",
"of historical responses is also computed by the same pooling strat",
"LRel,wheree'j is the contextual representation of",
"betweenthedecoderstateandtheconcatenationofthepersonalized",
"egyasrg=2",
"post representation, general user profile, and dynamic post-aware",
"the j-th word in Ri. Different from historical posts, the contextual",
"user profle, and further calculates two decoding mode probabilities",
"word representations of historical responses are obtained bythe",
"[p(mg), p(me)] = Softmax(d),",
"history encoder in Equation (3). Thus, all historical responses are",
"(9)",
"represented as (rg....,r).",
"d, = MLP([hR:ct;eG;ePJ)",
"(10)",
"Finally, we build the user history memory by using the post rep",
"resentations as key and the corresponding response representations",
"where hR is the decoding hidden state at step t, and d, e R? is the",
"as value, i.e., [kM . vM..,kM : vM = (p : rg,-... ph : rh)",
"matching degree vector to estimate the two mode probabilities. The",
"softmax function guarantees that p(mg) + p(mc) = 1.",
"3.4.2 Dynamic Post-aware User Profile. After building the user",
"3.5.2 Personalized general decoding. While general decoding, the",
"history memory, DHAP can select and aggregate the most relevant",
"historical responses and build the dynamic post-aware user profle",
"decoder should predict a word yt from the generic vocabulary:",
"basedontheinputpost.Specifically,thepersonalizedrepresentation",
"p(yt|mg) = Softmax(MLP([hR:cr:;eG;eDD).",
"(11)",
"of the input post obtained in Equation (6) is used as the query and",
"attends to the memory keys to find the most relevant historical",
"3.5.3 Personalized copy decoding.The personalized vocabulary",
"posts.Therelevance ismeasured bytheattention weights.Then",
"of user u is composed of the words that appear in their historical",
"the corresponding historical responses are summed up based on",
"responses. DHAP can directly select a word from this vocabulary"
] |
{
"arxiv_id": "2108.09355",
"language": "en",
"page_idx": 4
}
|
||
1904.00634-en-0
|
1904.00634-en-0.jpg
|
CFSNet: Toward a Controllable Feature Space for Image Restoration
Wei Wang $^{*1}$, Ruiming Guo $^{*2}$, Yapeng Tian$^{3}$ and Wenming Yang ti
$^{1}$Tsinghua University, $^2$Chinese University of Hong Kong, 3University of Rochester
$స్లే$
$సి$
$等$
<FigureHere>
$উড়্তি$
Figure 1. The perception-distortion trade-off of image super-resolution (1st row) and the trade-off between noise reduction and detail preservation (2nd row). At test-time, when using CFSNet, users can easily adjust an input control variable $\alpha_{in}$ to attain the most satisfactory result according to personal preferences. In contrast, the fixed methods (e.g.,EDSR, DnCNN-B) can not always guarantee the optimal visual quality.
### $\mathbf{Abstract}$
age quality. Experiments on several typical image restoration tasks fully validate the effective benefits of the proposed method. Code is available at https://github.com/ $qibao77/CFSNet$
$ਛੁੱ$
### $1. \textbf{Introduction}$
意
Deep learning methods have witnessed the great progress in image restoration with specific metrics (e.g., $PSNR,SSIM).However,theperceptualqualityofthere-$ stored image is relatively subjective, and it is necessary for users to control the reconstruction result according to personal preferences or image characteristics, which cannot be done using existing deterministic networks. This motivates us to exquisitely design a unifted interactive frame- $workforgeneralimagerestorationtasks.Underthisframe-$ work, users can control continuous transition of different $objectives$, e.g., the perception-distortion trade-off of image super-resolution, the trade-off between noise reduction and detail preservation. We achieve this goal by controlling the latent features of the designed network. To be specific, our proposed framework, named Controllable Feature Space Network (CFSNet), is entangled by two branches based on different objectives. Our framework can adaptively learn the coupling coefficients of different layers and channels, which provides finer control of the restored im-
Image restoration is a classic ill-posed inverse problem that aims to recover high-quality images from damaged images affected by various kinds of degradations. According to the types of degradation, it can be categorized into different subtasks such as image super-resolution, image denoising, JPEG image deblocking, etc.
The rise of deep learning has greatly facilitated the development of these subtasks. But these methods are often goal-specific, and we need to retrain the network when we deal with images different from the training dataset. Furthermore, most methods usually aim to pursue high reconstruction accuracy in terms of PSNR or SSIM. However, image quality assessment from personal opinion is relatively subjective, and low reconstruction distortion is not always consistent with high visual quality [5]. In addition, in many
$^{*}$Denotes equal contribution.
$^{\dagger}$Denotes corresponding author.
|
[
[
0.1396484375,
0.1331787109375,
0.83251953125,
0.1331787109375,
0.83251953125,
0.15087890625,
0.1396484375,
0.15087890625
],
[
0.2197265625,
0.1944580078125,
0.7490234375,
0.1944580078125,
0.7490234375,
0.212158203125,
0.2197265625,
0.212158203125
],
[
0.1568603515625,
0.222900390625,
0.81396484375,
0.2247314453125,
0.81396484375,
0.242431640625,
0.1568603515625,
0.240478515625
],
[
0.021240234375,
0.255126953125,
0.063720703125,
0.255126953125,
0.063720703125,
0.4873046875,
0.021240234375,
0.4873046875
],
[
0.223876953125,
0.375732421875,
0.25,
0.375732421875,
0.25,
0.38330078125,
0.223876953125,
0.38330078125
],
[
0.32275390625,
0.375,
0.361083984375,
0.375732421875,
0.361083984375,
0.38330078125,
0.32275390625,
0.3818359375
],
[
0.53759765625,
0.376953125,
0.560546875,
0.376953125,
0.560546875,
0.3837890625,
0.53759765625,
0.3837890625
],
[
0.734375,
0.376953125,
0.7607421875,
0.376953125,
0.7607421875,
0.3837890625,
0.734375,
0.3837890625
],
[
0.84326171875,
0.376953125,
0.85888671875,
0.376953125,
0.85888671875,
0.3837890625,
0.84326171875,
0.3837890625
],
[
0.439453125,
0.37744140625,
0.459228515625,
0.37744140625,
0.459228515625,
0.3818359375,
0.439453125,
0.3818359375
],
[
0.318603515625,
0.37744140625,
0.325927734375,
0.37744140625,
0.325927734375,
0.380615234375,
0.318603515625,
0.380615234375
],
[
0.64453125,
0.37744140625,
0.65869140625,
0.37744140625,
0.65869140625,
0.3818359375,
0.64453125,
0.3818359375
],
[
0.71142578125,
0.37890625,
0.7255859375,
0.37890625,
0.7255859375,
0.38330078125,
0.71142578125,
0.38330078125
],
[
0.810546875,
0.380126953125,
0.8203125,
0.380126953125,
0.8203125,
0.38330078125,
0.810546875,
0.38330078125
],
[
0.125,
0.458984375,
0.143798828125,
0.458984375,
0.143798828125,
0.46728515625,
0.125,
0.46728515625
],
[
0.2059326171875,
0.460205078125,
0.2607421875,
0.460205078125,
0.2607421875,
0.46728515625,
0.2059326171875,
0.46728515625
],
[
0.31787109375,
0.460205078125,
0.35302734375,
0.460205078125,
0.35302734375,
0.46728515625,
0.31787109375,
0.46728515625
],
[
0.443603515625,
0.4609375,
0.455810546875,
0.4609375,
0.455810546875,
0.46533203125,
0.443603515625,
0.46533203125
],
[
0.54150390625,
0.4609375,
0.5537109375,
0.4609375,
0.5537109375,
0.46533203125,
0.54150390625,
0.46533203125
],
[
0.64453125,
0.4609375,
0.65673828125,
0.4609375,
0.65673828125,
0.46533203125,
0.64453125,
0.46533203125
],
[
0.73876953125,
0.460205078125,
0.75732421875,
0.460205078125,
0.75732421875,
0.46728515625,
0.73876953125,
0.46728515625
],
[
0.80810546875,
0.4609375,
0.830078125,
0.4609375,
0.830078125,
0.4697265625,
0.80810546875,
0.4697265625
],
[
0.84619140625,
0.4609375,
0.8544921875,
0.4609375,
0.8544921875,
0.46533203125,
0.84619140625,
0.46533203125
],
[
0.080078125,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.080078125,
0.4912109375
],
[
0.708984375,
0.462158203125,
0.73193359375,
0.462158203125,
0.73193359375,
0.4697265625,
0.708984375,
0.4697265625
],
[
0.0236968994140625,
0.4892578125,
0.058013916015625,
0.4892578125,
0.0555419921875,
0.6396484375,
0.021240234375,
0.6396484375
],
[
0.080078125,
0.492431640625,
0.8896484375,
0.492431640625,
0.8896484375,
0.505859375,
0.080078125,
0.505859375
],
[
0.080078125,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.080078125,
0.51953125
],
[
0.08087158203125,
0.521484375,
0.1650390625,
0.521484375,
0.1650390625,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.236083984375,
0.546875,
0.314453125,
0.546875,
0.314453125,
0.564453125,
0.236083984375,
0.564453125
],
[
0.5009765625,
0.5498046875,
0.8896484375,
0.548828125,
0.8896484375,
0.56298828125,
0.5009765625,
0.564453125
],
[
0.50244140625,
0.56494140625,
0.89208984375,
0.56494140625,
0.89208984375,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.1004638671875,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.1004638671875,
0.59326171875
],
[
0.5009765625,
0.5791015625,
0.89208984375,
0.580078125,
0.89208984375,
0.5947265625,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.080078125,
0.609375
],
[
0.50341796875,
0.59521484375,
0.646484375,
0.59521484375,
0.646484375,
0.609375,
0.50341796875,
0.609375
],
[
0.08087158203125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62353515625,
0.08087158203125,
0.62353515625
],
[
0.08087158203125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.08087158203125,
0.6396484375
],
[
0.0318603515625,
0.63720703125,
0.0555419921875,
0.63720703125,
0.0555419921875,
0.671875,
0.0318603515625,
0.671875
],
[
0.08087158203125,
0.642578125,
0.46728515625,
0.642578125,
0.46728515625,
0.65283203125,
0.08087158203125,
0.65283203125
],
[
0.50341796875,
0.64599609375,
0.63134765625,
0.64599609375,
0.63134765625,
0.66015625,
0.50341796875,
0.66015625
],
[
0.07843017578125,
0.65478515625,
0.4697265625,
0.65380859375,
0.4697265625,
0.66845703125,
0.07843017578125,
0.66943359375
],
[
0.080078125,
0.67041015625,
0.4697265625,
0.67041015625,
0.4697265625,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.51953125,
0.671875,
0.89208984375,
0.67236328125,
0.89208984375,
0.6875,
0.51953125,
0.68603515625
],
[
0.0318603515625,
0.6787109375,
0.051483154296875,
0.67822265625,
0.053924560546875,
0.70556640625,
0.035125732421875,
0.70703125
],
[
0.08087158203125,
0.6875,
0.46728515625,
0.6875,
0.46728515625,
0.69775390625,
0.08087158203125,
0.69775390625
],
[
0.50244140625,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.08087158203125,
0.69970703125,
0.46728515625,
0.69970703125,
0.46728515625,
0.71337890625,
0.08087158203125,
0.71337890625
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71630859375
],
[
0.08251953125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.08251953125,
0.72998046875
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.73095703125,
0.468017578125,
0.73095703125,
0.468017578125,
0.744140625,
0.08087158203125,
0.744140625
],
[
0.50341796875,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.07843017578125,
0.7470703125,
0.468017578125,
0.74560546875,
0.468017578125,
0.7587890625,
0.07843017578125,
0.76025390625
],
[
0.50244140625,
0.74853515625,
0.72119140625,
0.74853515625,
0.72119140625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.080078125,
0.76123046875,
0.46728515625,
0.76123046875,
0.46728515625,
0.7744140625,
0.080078125,
0.7744140625
],
[
0.521484375,
0.7646484375,
0.890625,
0.76513671875,
0.890625,
0.77978515625,
0.521484375,
0.77880859375
],
[
0.080078125,
0.7763671875,
0.46728515625,
0.7763671875,
0.46728515625,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.50341796875,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.08087158203125,
0.79150390625,
0.468017578125,
0.79150390625,
0.468017578125,
0.8046875,
0.08087158203125,
0.8046875
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.806640625,
0.4697265625,
0.806640625,
0.4697265625,
0.82080078125,
0.080078125,
0.82080078125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8212890625,
0.46728515625,
0.8212890625,
0.46728515625,
0.83544921875,
0.080078125,
0.83544921875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.83837890625,
0.46728515625,
0.83837890625,
0.46728515625,
0.8486328125,
0.08087158203125,
0.8486328125
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.8515625,
0.46728515625,
0.8515625,
0.46728515625,
0.86572265625,
0.08087158203125,
0.86572265625
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.1004638671875,
0.87548828125,
0.25,
0.87548828125,
0.25,
0.8896484375,
0.1004638671875,
0.8896484375
],
[
0.50244140625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88525390625
],
[
0.098876953125,
0.88720703125,
0.2646484375,
0.88818359375,
0.2646484375,
0.8994140625,
0.098876953125,
0.8984375
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"CFSNet: Toward a Controllable Feature Space for Image Restoration",
"Wei Wang *1, Ruiming Guo *2, Yapeng Tian’ and Wenming Yang +1",
"ITsinghua University, 2Chinese University of Hong Kong, 3University of Rochester",
"612 [s]",
"DSE",
"SRGAN",
"0.2",
"=0.6",
"1.0",
"",
"",
"0",
"",
"",
"HO",
"DnONN-E",
"RCNN",
"",
"0",
"0",
"0.",
"ain",
"1",
"Figure 1. The perception-distortion trade-off of image super-resolution (Ist row) and the trade-off between noise reduction and detail",
"ain",
"9000",
"preservation (2nd row). At test-time, when using CFSNet, users can easily adjust an input control variable Qin to attain the most satisfactory",
"result according to personal preferences. In contrast, the fixed methods (e.g, EDSR, DnCNN-B) can not always guarantee the optimal",
"visual quality.",
"Abstract",
"age quality. Experiments on several typical image restora-",
"tion tasks fully validate the effective benefits of the proposed",
"Deep learning methods have witnessed the great",
"method. Code is available at https://github.com/",
"progress in image restoration with specific metrics (e.g.)",
"qibao77/cFsNet.",
"PSNR, SSIM). However, the perceptual quality of the re-",
"stored image is relatively subjective, and it is necessary for",
":AI",
"userstocontrolthereconstructionresultaccordingtoper-",
"1. Introduction",
"sonal preferences or image characteristics, which cannot",
"be done using existing deterministic networks. This moti-",
"Image restoration is a classic ill-posed inverse problem",
"",
"vatesus to exquisitelydesign aunifiedinteractive frame-",
"that aims to recover high-quality images from damaged im-",
"work for general image restoration tasks. Under this frame-",
"ages affected by various kinds ofdegradations.According",
"work, users can control continuous transition of different",
"to the types of degradation, it can be categorized into differ-",
"objectives, e.g., the perception-distortion trade-off of im-",
"ent subtasks such as image super-resolution, image denois-",
"age super-resolution, the trade-off between noise reduction",
"ing,JPEGimagedeblocking,etc.",
"and detail preservation. We achieve this goal by control-",
"The rise of deep learning has greatly facilitated the de-",
"ling the latent features of the designed network. To be spe-",
"velopment of these subtasks. But these methods are often",
"cific, our proposed framework, named Controllable Fea-",
"goal-specific, and we need to retrain the network when we",
"ture Space Network (CFSNet), is entangled by two branches",
"deal with images different from the training dataset. Fur-",
"based on different objectives. Our framework can adap-",
"thermore, most methods usually aim to pursue high recon-",
"tivelylearnthecouplingcoefficientsofdifferentlayersand",
"struction accuracy in terms of PSNR or SSIM. However, im-",
"channels, which provides finer control of the restored im-",
"age quality assessment from personal opinion is relatively",
"Denotes equal contribution.",
"subjective, and low reconstruction distortion is not always",
"Denotes corresponding author.",
"consistent with high visual quality [5]. In addition, in many"
] |
{
"arxiv_id": "1904.00634",
"language": "en",
"page_idx": 0
}
|
||
1904.00634-en-1
|
1904.00634-en-1.jpg
|
practical applications $(e.g.$, mobile), it is often challenging to obtain user's preference and the real degradation level of the corrupted images. All of these appeal to an interactive image restoration framework which can be applied to a wide variety of subtasks. However, to the best of our knowledge, currently there are few available networks which can satisfy both interactivity and generality requirements.
Some designs have been proposed to improve the flexibility of deep methods. Take image denoising for example, data augmentation is widely used to improve the generalization of a model. Training with the dataset which contains a series of noise levels, a single model can be applied to blind denoising task [41]. However, this method still produces a fixed reconstruction result of the input, which does not necessarily guarantee satisfactory perceptual quality (as shown in Fig. 1). An alternative choice, Zhang et al.[43] concatenated a tunable noise level map with degraded images as input to handle blind image denoising task. Though this scheme is also user-friendly, it can not be generalized to other tasks. In image super-resolution, [28] added noise to the input to control the compromise between perceptual quality and distortion. However, this scheme is specific for image super-resolution and cannot guarantee smooth and continuous control.
In this paper, to rectify these weaknesses, we propose a novel framework equipped with controllability for human perception-oriented interactive image restoration. To be more specific, we realize the interactive control of the reconstruction result by tuning the features of each unit block, called coupling module. Each coupling module consists of a main block and a tuning block. The parameters of two blocks are obtained under two endpoint optimization objectives. Taking image super-resolution as an example, the main block is optimized for low distortion while the tuning block is optimized for high perceptual quality. Besides, as a key to achieving fine feature control, we assign the highdegree-of-freedom coupling coefficients adaptively learned from a control scalar to each coupling module.
Our main contributions can be summarized as follows: - We propose a novel controllable end-to-end framework
for interactive image restoration in a fine-grained way.
- We propose a coupling module and an adaptive learn-
ing strategy of coupling coefficients to improve recon-
struction performance.
$►$ Our CFSNet outperforms the state-of-the-art methods
on super-resolution, JPEG image deblocking and image denoising in terms of flexibility and visual quality.
## $2. \textbf{Related Work}$
Image Restoration. Deep learning methods have been widely used in image restoration.[19,22,33,34,39,44,45]
continuously deepen, widen or lighten the network structure, aiming at improving the super-resolution accuracy as much as possible. While [17,21,27,30] paid more attention to the design of loss function to improve visual quality. Besides,[5,24,37] explored the perception-distortion tradeoff. In [9], Dong et al. adopted ARCNN built with several stacked convolutional layers for JPEG image deblocking Zhang et al. [43] proposed FFDNet to make image denoising more flexible and effective. Guo et al.[14] designed CBDNet to handle blind denoising of real images. Different from these task-specific methods, [23,24,41,42] proposed some unified schemes that can be employed to different image restoration tasks. However, these fixed networks are not flexible enough to deal with volatile user needs and application requirements.
In high-level
Controllable Image Transformation.
vision task, many technologies have been explored to implement controllable image transformation.[25] and [40] incorporated facial attribute vector into network to control facial appearance $(e.g.$, gender, age, beard). In [35], deep feature interpolation was adopt to implement automatic high-resolution image transformation.[18] also proposed a scheme that controls adaptive instance normalization (AdaIN) in feature space to adjust high-level attributes. Shoshan et al. [32] inserted some tuning blocks in the main network to allow modification of the network. However, all of these methods are designed for high-level vision tasks and can not be directly applied to image restoration. To apply controllable image transformation to low-level vision tasks, Wang et al.[36] performed interpolation in the parameter space, but this method can not guarantee the optimality of the outputs, which inspires us to further explore fine-grain control of image restoration.
## $3. \textbf{Proposed Method}$
In this section, we first provide an overview of the proposed framework, called CFSNet and then, present the modeling process inspired by the image super-resolution problem. Instead of specializing for the specific superresolution task, we finally generalize our CFSNet to multiple image restoration tasks, including denoising and deblocking. Moreover, we give the explicit model interpretation based on the manifold learning to show the intrinsic rationality of the proposed network. At the end of this section, we show the superiority and improvements of the proposed CFSNet through the detailed comparison with the current typical related methods.
## 3.1. Basic Network Architecture
As shown in Fig. 2, our CFSNet consists of a main branch and a tuning branch. The main branch contains $M$ main blocks (residual blocks [22]) while the tuning branch contains $M$ tuning blocks with additional 2* M+ 3 fully
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.106689453125,
0.080078125,
0.106689453125
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50244140625,
0.1514892578125,
0.890625,
0.154052734375,
0.890625,
0.1673583984375,
0.50244140625,
0.1654052734375
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50244140625,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.384033203125,
0.18310546875,
0.384033203125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1817626953125,
0.8896484375,
0.1817626953125,
0.8896484375,
0.1956787109375,
0.50244140625,
0.1956787109375
],
[
0.096435546875,
0.1976318359375,
0.468017578125,
0.1962890625,
0.468017578125,
0.2108154296875,
0.096435546875,
0.212158203125
],
[
0.50244140625,
0.1976318359375,
0.8896484375,
0.1976318359375,
0.8896484375,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.08087158203125,
0.2279052734375,
0.4697265625,
0.2279052734375,
0.4697265625,
0.2412109375,
0.08087158203125,
0.2412109375
],
[
0.50244140625,
0.2279052734375,
0.89208984375,
0.2279052734375,
0.89208984375,
0.2412109375,
0.50244140625,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
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
],
[
0.50341796875,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.5009765625,
0.287353515625,
0.890625,
0.28857421875,
0.890625,
0.302978515625,
0.5009765625,
0.3017578125
],
[
0.08087158203125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.31640625,
0.08087158203125,
0.31640625
],
[
0.50341796875,
0.30615234375,
0.64208984375,
0.30615234375,
0.64208984375,
0.31640625,
0.50341796875,
0.31640625
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.52392578125,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.52392578125,
0.33203125
],
[
0.07843017578125,
0.333251953125,
0.46728515625,
0.333984375,
0.46728515625,
0.348388671875,
0.07843017578125,
0.347900390625
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.348388671875,
0.890625,
0.348388671875,
0.890625,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.5009765625,
0.37744140625,
0.89208984375,
0.37890625,
0.89208984375,
0.393310546875,
0.5009765625,
0.392578125
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.08087158203125,
0.4091796875,
0.46728515625,
0.4091796875,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.423583984375,
0.89208984375,
0.4375,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.43994140625,
0.2099609375,
0.43994140625,
0.2099609375,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.09722900390625,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.09722900390625,
0.4697265625
],
[
0.50341796875,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50341796875,
0.468505859375
],
[
0.080078125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50341796875,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.481689453125,
0.50341796875,
0.481689453125
],
[
0.080078125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.5009765625,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.5009765625,
0.5146484375
],
[
0.08087158203125,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.50244140625,
0.51513671875,
0.89208984375,
0.5146484375,
0.89208984375,
0.52783203125,
0.50244140625,
0.52880859375
],
[
0.08087158203125,
0.52978515625,
0.471435546875,
0.52978515625,
0.471435546875,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.5009765625,
0.52880859375,
0.890625,
0.52978515625,
0.890625,
0.54443359375,
0.5009765625,
0.54345703125
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.50244140625,
0.5595703125
],
[
0.08087158203125,
0.56005859375,
0.46728515625,
0.56005859375,
0.46728515625,
0.57373046875,
0.08087158203125,
0.57373046875
],
[
0.5009765625,
0.5595703125,
0.8935546875,
0.56005859375,
0.8935546875,
0.57470703125,
0.5009765625,
0.57373046875
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.57568359375,
0.7587890625,
0.57568359375,
0.7587890625,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.07843017578125,
0.59033203125,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.07843017578125,
0.60400390625
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50341796875,
0.60302734375,
0.67236328125,
0.60302734375,
0.67236328125,
0.6171875,
0.50341796875,
0.6171875
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.51953125,
0.626953125,
0.890625,
0.62744140625,
0.890625,
0.642578125,
0.51953125,
0.6416015625
],
[
0.08087158203125,
0.63623046875,
0.4697265625,
0.63623046875,
0.4697265625,
0.6494140625,
0.08087158203125,
0.6494140625
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.64306640625,
0.89208984375,
0.65771484375,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.6513671875,
0.384033203125,
0.6513671875,
0.384033203125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.09722900390625,
0.66552734375,
0.455810546875,
0.66552734375,
0.455810546875,
0.6787109375,
0.09722900390625,
0.6787109375
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.09478759765625,
0.68994140625,
0.46728515625,
0.68994140625,
0.46728515625,
0.7041015625,
0.09478759765625,
0.7041015625
],
[
0.50341796875,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50341796875,
0.70263671875
],
[
0.111083984375,
0.70556640625,
0.46728515625,
0.70703125,
0.46728515625,
0.7216796875,
0.111083984375,
0.72021484375
],
[
0.50341796875,
0.7041015625,
0.8935546875,
0.7041015625,
0.8935546875,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.50341796875,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.108642578125,
0.73095703125,
0.465576171875,
0.73095703125,
0.465576171875,
0.744140625,
0.108642578125,
0.744140625
],
[
0.50244140625,
0.734375,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74755859375
],
[
0.1094970703125,
0.7470703125,
0.468017578125,
0.74560546875,
0.468017578125,
0.76025390625,
0.1094970703125,
0.76123046875
],
[
0.1070556640625,
0.736328125,
0.11358642578125,
0.736328125,
0.11358642578125,
0.7412109375,
0.1070556640625,
0.7412109375
],
[
0.50244140625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.11273193359375,
0.76220703125,
0.2646484375,
0.76220703125,
0.2646484375,
0.77587890625,
0.11273193359375,
0.77587890625
],
[
0.50341796875,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.50341796875,
0.77978515625,
0.8935546875,
0.77978515625,
0.8935546875,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.0931396484375,
0.7861328125,
0.468017578125,
0.7861328125,
0.468017578125,
0.7998046875,
0.0931396484375,
0.7998046875
],
[
0.50244140625,
0.79443359375,
0.6611328125,
0.79345703125,
0.6611328125,
0.80810546875,
0.50244140625,
0.8095703125
],
[
0.11273193359375,
0.8017578125,
0.46728515625,
0.8017578125,
0.46728515625,
0.81494140625,
0.11273193359375,
0.81494140625
],
[
0.111083984375,
0.8173828125,
0.46728515625,
0.8173828125,
0.46728515625,
0.83056640625,
0.111083984375,
0.83056640625
],
[
0.50341796875,
0.8173828125,
0.7509765625,
0.8173828125,
0.7509765625,
0.83056640625,
0.50341796875,
0.83056640625
],
[
0.08087158203125,
0.84521484375,
0.2197265625,
0.84521484375,
0.2197265625,
0.85986328125,
0.08087158203125,
0.85986328125
],
[
0.51953125,
0.84033203125,
0.89208984375,
0.8388671875,
0.89208984375,
0.85400390625,
0.51953125,
0.85498046875
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.098876953125,
0.87060546875,
0.46728515625,
0.87060546875,
0.46728515625,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"practical applications (e.g., mobile), it is often challenging",
"continuously deepen, widen or lighten the network struc-",
"to obtain user's preference and the real degradation level of",
"ture, aiming at improving the super-resolution accuracy as",
"the corrupted images. All of these appeal to an interactive",
"much as possible. While [17,21,27, 30] paid more atten-",
"image restoration framework which can be applied to a wide",
"tion to the design of loss function to improve visual quality.",
"variety of subtasks.However,tothe best of ourknowledge.",
"Besides,[5,24,37]explored theperception-distortion trade-",
"currentlytherearefewavailablenetworks whichcan satisfy",
"off. In [9], Dong et al. adopted ARCNN built with several",
"both interactivity and generality requirements.",
"stacked convolutional layers for JPEG image deblocking",
"Some designs have been proposed to improve the flexi-",
"Zhang et al. [43] proposed FFDNet to make image denois-",
"bility of deep methods. Take image denoising for example,",
"ing more flexible and effective. Guo et al. [14] designed",
"data augmentation is widely used to improve the general-",
"CBDNet to handle blind denoising of real images.Different",
"ization of a model. Training with the dataset which con-",
"from these task-specific methods, [23, 24, 41,42] proposed",
"tains a series of noise levels, a single model can be applied",
"some unified schemes that can be employed to different im-",
"to blind denoising task [41]. However, this method still pro-",
"age restoration tasks. However, these fixed networks are not",
"duces a fixed reconstruction result of the input, which does",
"flexible enough to deal with volatile user needs and appli-",
"notnecessarilyguaranteesatisfactoryperceptualquality(as",
"cationrequirements.",
"showninFig.1).An alternativechoice,Zhang etal.[43]",
"Controllable Image Transformation.In high-level",
"concatenated a tunable noise level map with degraded im-",
"vision task, many technologies have been explored to im-",
"ages as inputtohandleblind image denoisingtask.Though",
"plement controllable image transformation. [25] and [40]",
"this scheme is also user-friendly, it can not be generalized",
"incorporated facial attribute vector into network to con-",
"toothertasks.Inimage super-resolution,[28]added noise",
"trol facial appearance (e.g., gender, age, beard). In [35],",
"to the input to control the compromise between perceptual",
"deep feature interpolation was adopt to implement auto-",
"quality and distortion. However, this scheme is specific for",
"matic high-resolution image transformation. [18] also pro-",
"image super-resolution and cannot guarantee smooth and",
"posed a scheme that controls adaptive instance normaliza-",
"continuous control.",
"tion (AdalN) in feature space to adjust high-level attributes.",
"In this paper, to rectify these weaknesses, we propose",
"Shoshan et al. [32] inserted some tuning blocks in the main",
"a novel frameworkequipped withcontrollabilityforhu-",
"networktoallowmodificationofthenetwork.However.all",
"man perception-oriented interactive image restoration. To",
"of these methods are designed for high-level vision tasks",
"be more specific, we realize the interactive control of the re-",
"and can not be directly applied to image restoration. To",
"constructionresultbytuningthefeaturesofeachunitblock.",
"applycontrollable imagetransformation tolow-levelvision",
"called coupling module. Each coupling module consists of",
"tasks, Wang et al. [36] performed interpolation in the pa-",
"a main block and a tuning block. The parameters of two",
"rameter space, but this method can not guarantee the opti-",
"blocks are obtained under two endpoint optimization ob-",
"mality of the outputs, which inspires us to further explore",
"jectives. Taking image super-resolution as an example, the",
"fine-grain control of image restoration.",
"main block is optimizedforlowdistortion while the tuning",
"block is optimized for high perceptual quality. Besides, as",
"3. Proposed Method",
"a key to achieving fine feature control, we assign the high-",
"In this section, we first provide an overview of the pro-",
"degree-of-freedom coupling coefficients adaptively learned",
"posed framework, called CFSNet and then, present the",
"from a control scalar to each coupling module",
"modeling process inspired by the image super-resolution",
"Our main contributions can be summarized as follows",
"problem.Instead of specializing for the specific super-",
"We propose a novel controllable end-to-end framework",
"resolution task, we finally generalize our CFSNet to mul-",
"for interactive image restoration in a fine-grained way.",
"tiple image restoration tasks, including denoising and de-",
"blocking. Moreover, we give the explicit model interpreta-",
" We propose a coupling module and an adaptive learn-",
"tion based on the manifold learning to show the intrinsicra-",
"ing strategy of coupling coefficients to improve recon-",
"",
"tionality of the proposed network. At the end of this section,",
"struction performance.",
"we show the superiority and improvements of the proposed",
"CFSNet through the detailed comparison with the current",
" Our CFSNet outperforms the state-of-the-art methods",
"typical related methods.",
"on super-resolution, JPEG image deblocking and im-",
" age denoising in terms of flexibility and visual quality.",
"3.1.Basic Network Architecture",
"2. Related Work",
"As shown in Fig.2, our CFSNet consists of a main",
"branch and a tuning branch. The main branch contains M",
"Image Restoration. Deep learning methods have been",
"main blocks (residual blocks [22]) while the tuning branch",
"widely used in image restoration. [19, 22, 33, 34, 39, 44, 45]",
"contains M tuning blocks with additional 2 * M + 3 fully"
] |
{
"arxiv_id": "1904.00634",
"language": "en",
"page_idx": 1
}
|
||
1904.00634-en-2
|
1904.00634-en-2.jpg
|
<FigureHere>
Figure 2. The framework of our proposed controllable feature space network (CFSNet).
in Fig. 2. Specifically, we utilize sub-pixel convolutional operation (convolution+pixel shuffle)[31] to upscale feature maps. Finally, we use a $3\times3$ convolutional layer to get the reconstructed image,
connected layers. A pair of main block and tuning block constitute a coupling module by using a coupling operation to combine the features of the two branches effectively. We take the original degraded image $I_{in}$ and the control scalar $\alpha_{in}$ as the input and output the restored image $I_rec$ as the final result.
In the beginning, we first use a $3\times3$ convolutional layer
to extract features from the degraded image $I_in$,
$$I_{rec}=F_{out}(B_M+B_0)\:or\:I_{rec}=F_{out}(B_{M+1}),$$
(4)
where $F_{out}(\cdot)$ denotes convolution operation. The overall
reconstruction process can be expressed as
(1)
$$B_0=F_{in}(I_{in}),$$
$$I_{rec}=F_{CFSN}(I_{in},\alpha_{in};\theta_{main},\theta_{tun},\theta_{\alpha}),$$
(5)
where $F_in(\cdot)$ represents the feature extraction function and $B_0$ serves as the input of next stage. Here, together with the input image $I_in$, we introduce a control scalar $\alpha_{in}$ to balance the different optimization goals. To be more specific, there are 3 shared fully connected layers to transform the input scalar $\alpha_{in}$ into multi-channels vectors and 2 independent fully connected layers to learn the optimal coupling $\dot{\text{coefficient for each coupling module:}}$
$\begin{array}{l}\mathrm{where~}F_{CFSN}(\cdot)\text{ represents the function of our proposed}\\\mathrm{CFSNet.~}\theta_{main},\:\theta_{tun}\mathrm{~and~}\theta_{\alpha}\text{ represent the parameters of}\end{array}$ main branch, all tuning blocks and all fully connected layers respectively.
Since the two branches of our framework are based on different optimization objectives, in further detail, our training process can be divided into two steps:
$$\alpha_m=F_m^{ind}(F_m^{sha}(\alpha_{in})),$$
(2)
Step 1 Set the control variable $\alpha_{in}$ as 0. Train the main
branch with the loss function $L_1(I_{rec},I_g;\theta_{main})$, where $I_g$ is the corresponding ground truth image.
Step2 Set the control variable $\alpha_{in}$ as 1. Map the con-
trol variable $\alpha_{in}$ to different coupling coefficients $\{\alpha_m\}$, fix parameters of the main branch and train the tuning branch with another loss function $L_{2}(I_{rec},I_{g};\theta_{tun},\theta_{\alpha}).$
where both $F_{m}^{sha}(\cdot)$ and $F_{m}^{ind}(\cdot)$ denote the function of the shared and independent fully connected layers, and $\alpha_m$ is the coupling coefficient vector of the $m$-th coupling module. Each coupling module couples the output of a main block and a tuning block as follows:
$$\begin{aligned}B_m&=F_m(R_m,T_m)\\&=F_m(F_m^{main}(B_{m-1}),F_m^{tun}(B_{m-1})),\end{aligned}$$
(3)
# 3.2. Coupling module
We now present the details of our coupling module. We mainly introduce our design from the perspective of image super-resolution. In order to balance the trade-off between perceptual quality and distortion, we usually realize it by modifying the penality parameter $\lambda$ of the loss terms[5],
where $F_m(\cdot)$ represents the $m$-th coupling operation, $R_m$ and $T_m$ denote the output features of $m$-th main block and $m$-th tuning block respectively, $F_m^{main}(\cdot)$ and $F_m^{tun}(\cdot)$ are the $m$-th main block function and $m$-th tuning block function respectively. To address the image super-resolution task, we add an extra coupling module consisting of the upscaling block before the last convolutional layer, as shown
(6)
$$L_{gen}=L_{distortion}+\lambda L_{adv},$$
|
[
[
0.314453125,
0.1142578125,
0.351318359375,
0.1142578125,
0.351318359375,
0.1256103515625,
0.314453125,
0.1256103515625
],
[
0.3798828125,
0.11553955078125,
0.420654296875,
0.11553955078125,
0.420654296875,
0.1256103515625,
0.3798828125,
0.1256103515625
],
[
0.455810546875,
0.11553955078125,
0.484375,
0.11553955078125,
0.484375,
0.12371826171875,
0.455810546875,
0.12371826171875
],
[
0.71142578125,
0.111083984375,
0.74853515625,
0.11297607421875,
0.74658203125,
0.1256103515625,
0.7099609375,
0.12371826171875
],
[
0.53125,
0.117431640625,
0.54833984375,
0.117431640625,
0.54833984375,
0.121826171875,
0.53125,
0.121826171875
],
[
0.318603515625,
0.125,
0.340576171875,
0.126953125,
0.340576171875,
0.1307373046875,
0.318603515625,
0.1287841796875
],
[
0.3896484375,
0.1275634765625,
0.41259765625,
0.1275634765625,
0.41259765625,
0.1307373046875,
0.3896484375,
0.1307373046875
],
[
0.7060546875,
0.1300048828125,
0.73779296875,
0.1319580078125,
0.736328125,
0.143310546875,
0.705078125,
0.141357421875
],
[
0.71337890625,
0.125,
0.744140625,
0.125,
0.744140625,
0.1319580078125,
0.71337890625,
0.1319580078125
],
[
0.318603515625,
0.1331787109375,
0.3349609375,
0.1331787109375,
0.3349609375,
0.141357421875,
0.318603515625,
0.141357421875
],
[
0.384033203125,
0.1331787109375,
0.39794921875,
0.1331787109375,
0.39794921875,
0.141357421875,
0.384033203125,
0.141357421875
],
[
0.453369140625,
0.1337890625,
0.464111328125,
0.1307373046875,
0.465576171875,
0.136962890625,
0.455078125,
0.1395263671875
],
[
0.525390625,
0.1331787109375,
0.54345703125,
0.1331787109375,
0.54345703125,
0.141357421875,
0.525390625,
0.141357421875
],
[
0.25,
0.14453125,
0.2548828125,
0.14453125,
0.2548828125,
0.15283203125,
0.25,
0.15283203125
],
[
0.318603515625,
0.146484375,
0.325927734375,
0.1514892578125,
0.318603515625,
0.1572265625,
0.31201171875,
0.1514892578125
],
[
0.381591796875,
0.1483154296875,
0.3857421875,
0.1483154296875,
0.3857421875,
0.1546630859375,
0.381591796875,
0.1546630859375
],
[
0.525390625,
0.149658203125,
0.53515625,
0.149658203125,
0.53515625,
0.1572265625,
0.525390625,
0.1572265625
],
[
0.69677734375,
0.1458740234375,
0.7158203125,
0.1514892578125,
0.7099609375,
0.1610107421875,
0.69140625,
0.1546630859375
],
[
0.2646484375,
0.1622314453125,
0.271240234375,
0.1622314453125,
0.271240234375,
0.1666259765625,
0.2646484375,
0.1666259765625
],
[
0.26318359375,
0.1514892578125,
0.27294921875,
0.1514892578125,
0.27294921875,
0.1634521484375,
0.26318359375,
0.1634521484375
],
[
0.373291015625,
0.1976318359375,
0.40283203125,
0.1932373046875,
0.406982421875,
0.212158203125,
0.37744140625,
0.216552734375
],
[
0.468017578125,
0.1976318359375,
0.4755859375,
0.199462890625,
0.4697265625,
0.210205078125,
0.463134765625,
0.2083740234375
],
[
0.5400390625,
0.1988525390625,
0.5458984375,
0.1988525390625,
0.5458984375,
0.20703125,
0.5400390625,
0.20703125
],
[
0.67724609375,
0.1988525390625,
0.7548828125,
0.1988525390625,
0.7548828125,
0.216552734375,
0.67724609375,
0.216552734375
],
[
0.330078125,
0.2008056640625,
0.334228515625,
0.2008056640625,
0.33251953125,
0.2083740234375,
0.328369140625,
0.2083740234375
],
[
0.281005859375,
0.1988525390625,
0.285888671875,
0.1988525390625,
0.285888671875,
0.2083740234375,
0.281005859375,
0.2083740234375
],
[
0.52294921875,
0.2008056640625,
0.53369140625,
0.2083740234375,
0.52685546875,
0.2139892578125,
0.51708984375,
0.20703125
],
[
0.313720703125,
0.20263671875,
0.32177734375,
0.20263671875,
0.32177734375,
0.212158203125,
0.313720703125,
0.212158203125
],
[
0.1895751953125,
0.20263671875,
0.1993408203125,
0.20263671875,
0.1993408203125,
0.210205078125,
0.1895751953125,
0.210205078125
],
[
0.25732421875,
0.2052001953125,
0.273681640625,
0.205810546875,
0.271240234375,
0.2291259765625,
0.2548828125,
0.2279052734375
],
[
0.450927734375,
0.2052001953125,
0.455810546875,
0.2052001953125,
0.455810546875,
0.2108154296875,
0.450927734375,
0.2108154296875
],
[
0.27783203125,
0.21337890625,
0.28515625,
0.21337890625,
0.28515625,
0.2197265625,
0.27783203125,
0.2197265625
],
[
0.281982421875,
0.2197265625,
0.289306640625,
0.2247314453125,
0.281982421875,
0.2291259765625,
0.275390625,
0.2235107421875
],
[
0.328369140625,
0.21533203125,
0.338134765625,
0.21533203125,
0.338134765625,
0.2342529296875,
0.328369140625,
0.2342529296875
],
[
0.393798828125,
0.2197265625,
0.40283203125,
0.2197265625,
0.40283203125,
0.2342529296875,
0.393798828125,
0.2342529296875
],
[
0.68310546875,
0.2197265625,
0.68701171875,
0.2215576171875,
0.68310546875,
0.2298583984375,
0.6787109375,
0.2279052734375
],
[
0.72119140625,
0.2197265625,
0.734375,
0.2197265625,
0.734375,
0.23486328125,
0.72119140625,
0.23486328125
],
[
0.5390625,
0.2235107421875,
0.54345703125,
0.2235107421875,
0.54345703125,
0.2291259765625,
0.5390625,
0.2291259765625
],
[
0.464111328125,
0.2203369140625,
0.473876953125,
0.2203369140625,
0.473876953125,
0.2342529296875,
0.464111328125,
0.2342529296875
],
[
0.281005859375,
0.2266845703125,
0.28759765625,
0.2298583984375,
0.281982421875,
0.23486328125,
0.275390625,
0.231689453125
],
[
0.2034912109375,
0.251953125,
0.325927734375,
0.25244140625,
0.325927734375,
0.266357421875,
0.2034912109375,
0.26513671875
],
[
0.49267578125,
0.25244140625,
0.55810546875,
0.25244140625,
0.55810546875,
0.26318359375,
0.49267578125,
0.26318359375
],
[
0.5947265625,
0.251953125,
0.74169921875,
0.25,
0.74169921875,
0.26318359375,
0.5947265625,
0.26513671875
],
[
0.2059326171875,
0.270751953125,
0.31787109375,
0.270751953125,
0.31787109375,
0.2841796875,
0.2059326171875,
0.2841796875
],
[
0.35693359375,
0.25390625,
0.44677734375,
0.25390625,
0.44677734375,
0.26513671875,
0.35693359375,
0.26513671875
],
[
0.35546875,
0.270751953125,
0.465576171875,
0.270751953125,
0.465576171875,
0.2841796875,
0.35546875,
0.2841796875
],
[
0.494384765625,
0.270751953125,
0.56787109375,
0.270751953125,
0.56787109375,
0.2822265625,
0.494384765625,
0.2822265625
],
[
0.59619140625,
0.270751953125,
0.6708984375,
0.270751953125,
0.6708984375,
0.2822265625,
0.59619140625,
0.2822265625
],
[
0.69921875,
0.270751953125,
0.78125,
0.270751953125,
0.78125,
0.2822265625,
0.69921875,
0.2822265625
],
[
0.2303466796875,
0.30810546875,
0.74169921875,
0.30810546875,
0.74169921875,
0.318115234375,
0.2303466796875,
0.318115234375
],
[
0.080078125,
0.34521484375,
0.4697265625,
0.34521484375,
0.4697265625,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.5009765625,
0.34521484375,
0.89208984375,
0.3447265625,
0.89208984375,
0.359130859375,
0.5009765625,
0.35986328125
],
[
0.07843017578125,
0.35986328125,
0.4697265625,
0.361083984375,
0.4697265625,
0.375732421875,
0.07843017578125,
0.375
],
[
0.5009765625,
0.361083984375,
0.890625,
0.35986328125,
0.890625,
0.375,
0.5009765625,
0.375732421875
],
[
0.080078125,
0.376953125,
0.468017578125,
0.376953125,
0.468017578125,
0.39013671875,
0.080078125,
0.39013671875
],
[
0.50244140625,
0.376953125,
0.89208984375,
0.376953125,
0.89208984375,
0.39013671875,
0.50244140625,
0.39013671875
],
[
0.080078125,
0.391357421875,
0.468017578125,
0.391357421875,
0.468017578125,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.50244140625,
0.392578125,
0.66650390625,
0.392578125,
0.66650390625,
0.406005859375,
0.50244140625,
0.406005859375
],
[
0.08087158203125,
0.4072265625,
0.4697265625,
0.4072265625,
0.4697265625,
0.42041015625,
0.08087158203125,
0.42041015625
],
[
0.08087158203125,
0.42236328125,
0.1544189453125,
0.42236328125,
0.1544189453125,
0.434326171875,
0.08087158203125,
0.434326171875
],
[
0.52783203125,
0.419189453125,
0.89208984375,
0.419189453125,
0.89208984375,
0.43310546875,
0.52783203125,
0.43310546875
],
[
0.09722900390625,
0.4375,
0.468017578125,
0.4375,
0.468017578125,
0.45068359375,
0.09722900390625,
0.45068359375
],
[
0.50341796875,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.460205078125,
0.50341796875,
0.460205078125
],
[
0.07843017578125,
0.45263671875,
0.396240234375,
0.45263671875,
0.396240234375,
0.466552734375,
0.07843017578125,
0.466552734375
],
[
0.50341796875,
0.46337890625,
0.78515625,
0.46337890625,
0.78515625,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.22216796875,
0.477783203125,
0.328369140625,
0.477783203125,
0.328369140625,
0.492431640625,
0.22216796875,
0.492431640625
],
[
0.44677734375,
0.477783203125,
0.4697265625,
0.477783203125,
0.4697265625,
0.492431640625,
0.44677734375,
0.492431640625
],
[
0.55810546875,
0.488037109375,
0.8388671875,
0.488037109375,
0.8388671875,
0.50244140625,
0.55810546875,
0.50244140625
],
[
0.869140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.501953125,
0.869140625,
0.501953125
],
[
0.08087158203125,
0.50390625,
0.4697265625,
0.50244140625,
0.4697265625,
0.517578125,
0.08087158203125,
0.51806640625
],
[
0.08251953125,
0.52099609375,
0.4697265625,
0.52099609375,
0.4697265625,
0.5341796875,
0.08251953125,
0.5341796875
],
[
0.5048828125,
0.5166015625,
0.89208984375,
0.5166015625,
0.89208984375,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.08087158203125,
0.53515625,
0.4697265625,
0.53515625,
0.4697265625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50244140625,
0.52978515625,
0.89208984375,
0.53076171875,
0.89208984375,
0.54541015625,
0.50244140625,
0.54443359375
],
[
0.07843017578125,
0.548828125,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.07843017578125,
0.5625
],
[
0.50244140625,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.08087158203125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.50341796875,
0.5625,
0.583984375,
0.5625,
0.583984375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.07843017578125,
0.580078125,
0.468017578125,
0.5791015625,
0.468017578125,
0.59326171875,
0.07843017578125,
0.5947265625
],
[
0.521484375,
0.57568359375,
0.89208984375,
0.57568359375,
0.89208984375,
0.5888671875,
0.521484375,
0.5888671875
],
[
0.080078125,
0.59521484375,
0.4697265625,
0.59521484375,
0.4697265625,
0.60986328125,
0.080078125,
0.60986328125
],
[
0.50341796875,
0.59033203125,
0.890625,
0.59033203125,
0.890625,
0.60400390625,
0.50341796875,
0.60400390625
],
[
0.080078125,
0.60986328125,
0.325927734375,
0.60986328125,
0.325927734375,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50341796875,
0.60595703125,
0.77685546875,
0.60595703125,
0.77685546875,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.188720703125,
0.63525390625,
0.359375,
0.63525390625,
0.359375,
0.65283203125,
0.188720703125,
0.65283203125
],
[
0.44677734375,
0.63623046875,
0.4697265625,
0.63623046875,
0.4697265625,
0.65087890625,
0.44677734375,
0.65087890625
],
[
0.5048828125,
0.6318359375,
0.89208984375,
0.63330078125,
0.89208984375,
0.64794921875,
0.5048828125,
0.646484375
],
[
0.560546875,
0.64794921875,
0.890625,
0.6484375,
0.890625,
0.66357421875,
0.560546875,
0.662109375
],
[
0.080078125,
0.66015625,
0.471435546875,
0.66015625,
0.471435546875,
0.67822265625,
0.080078125,
0.67822265625
],
[
0.560546875,
0.662109375,
0.890625,
0.6640625,
0.890625,
0.6787109375,
0.560546875,
0.6767578125
],
[
0.08087158203125,
0.6787109375,
0.468017578125,
0.6787109375,
0.468017578125,
0.6923828125,
0.08087158203125,
0.6923828125
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.5048828125,
0.6875,
0.8896484375,
0.6875,
0.8896484375,
0.7021484375,
0.5048828125,
0.7021484375
],
[
0.080078125,
0.70849609375,
0.4697265625,
0.70849609375,
0.4697265625,
0.7216796875,
0.080078125,
0.7216796875
],
[
0.56201171875,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71826171875,
0.56201171875,
0.71826171875
],
[
0.080078125,
0.7216796875,
0.31787109375,
0.7216796875,
0.31787109375,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.56201171875,
0.71826171875,
0.89208984375,
0.71728515625,
0.89208984375,
0.73193359375,
0.56201171875,
0.73291015625
],
[
0.56396484375,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.56396484375,
0.74755859375
],
[
0.134033203125,
0.74853515625,
0.271240234375,
0.74853515625,
0.271240234375,
0.76318359375,
0.134033203125,
0.76318359375
],
[
0.56201171875,
0.75,
0.708984375,
0.75,
0.708984375,
0.76318359375,
0.56201171875,
0.76318359375
],
[
0.44677734375,
0.76025390625,
0.4697265625,
0.76025390625,
0.4697265625,
0.77294921875,
0.44677734375,
0.77294921875
],
[
0.1641845703125,
0.76953125,
0.4150390625,
0.76953125,
0.4150390625,
0.7841796875,
0.1641845703125,
0.7841796875
],
[
0.50244140625,
0.7734375,
0.6689453125,
0.7744140625,
0.6689453125,
0.7890625,
0.50244140625,
0.78857421875
],
[
0.08087158203125,
0.79541015625,
0.46728515625,
0.79541015625,
0.46728515625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.52294921875,
0.798828125,
0.89208984375,
0.798828125,
0.89208984375,
0.8125,
0.52294921875,
0.8125
],
[
0.07843017578125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.814453125,
0.89208984375,
0.814453125,
0.89208984375,
0.82763671875,
0.50244140625,
0.82763671875
],
[
0.0767822265625,
0.82275390625,
0.471435546875,
0.82373046875,
0.471435546875,
0.84130859375,
0.0767822265625,
0.84033203125
],
[
0.50244140625,
0.830078125,
0.89208984375,
0.830078125,
0.89208984375,
0.84326171875,
0.50244140625,
0.84326171875
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.8447265625,
0.890625,
0.84326171875,
0.890625,
0.85791015625,
0.5009765625,
0.859375
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50244140625,
0.85986328125,
0.875,
0.85986328125,
0.875,
0.87353515625,
0.50244140625,
0.87353515625
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.6005859375,
0.88720703125,
0.79736328125,
0.88720703125,
0.79736328125,
0.9013671875,
0.6005859375,
0.9013671875
],
[
0.86767578125,
0.88525390625,
0.89453125,
0.88525390625,
0.89453125,
0.9013671875,
0.86767578125,
0.9013671875
]
] |
[
"64×64",
"64X64",
"64X64",
"64×64",
"6X",
"AXE",
"",
"am+1",
"64X64",
"ai",
"a2",
"",
"aM",
"T",
"",
"",
"M",
"TM+",
"2",
"",
"R置",
"D",
"o",
"#RM+1BM",
"",
"",
"R",
"R",
"lin",
"",
"",
"C",
"C",
"Inos",
"Os",
"",
"os",
"",
"105",
"",
"Fully conmected layer",
"Mainblock",
"Upscale coupling module",
"Convolutional layer",
"Activation layer",
"Coupling operation",
"Tuning block",
"Main branch",
"Tuning branch",
"Figure2.Theframeworkofourproposedcontrollablefeaturespacenetwork(CFSNet)",
"connected layers. A pair of main block and tuning block",
"in Fig.2. Specifically,we utilize sub-pixel convolutional",
"constitute a coupling module by using a coupling operation",
"operation (convolution + pixel shuffle) [31] to upscale fea-",
"to combine the features of the two branches effectively. We",
"ture maps. Finally, we use a 3 x 3 convolutional layer to get",
"take the original degraded image Iin and the control scalar",
"the reconstructed image,",
"Qin as the input and output the restored image Irec as the",
"final result.",
"Irec =Fout(BM +Bo) or Irec =Fout(BM+1),(4)",
"In the beginning, we first use a 3 × 3 convolutional layer",
"where Fout() denotes convolution operation. The overall",
"to extract features from the degraded image Iin,",
"reconstructionprocesscanbeexpressedas",
"Bo = Fin(Iin),",
"(1)",
"Irec =FcFSN(Iin,ain;Omain,Otun,Oα),",
"(5)",
"where Fin )represents the feature extraction function and",
"Bo serves as the input of next stage. Here, together with",
"where FcFsn() represents the function of our proposed",
"the input image Iin, we introduce a control scalar Qin to",
"CFSNet. Omain, Otun and α represent the parameters of",
"balance the different optimization goals. To be more spe-",
"main branch, all tuning blocks and all fully connected layers",
"cific, there are 3 shared fully connected layers to transform",
"respectively",
"the input scalar Qin into multi-channels vectors and 2 inde-",
"Since the two branches of our framework are based on",
"pendent fully connected layers to learn the optimal coupling",
"different optimization objectives, in further detail, our train-",
"coefficient for each coupling module:",
"ing process can be divided into two steps:",
"Qm = Find(Fsha(ain)),",
"(2)",
"Step 1 Set the control variable Qin as O. Train the main",
"branch with the loss function Li(Irec, Ig;Omain),",
"where both Fsha() and Find() denote the function of the",
"where Ig is the corresponding ground truth image.",
"shared and independent fully connected layers, and Qm is",
"the coupling coefficient vector of the m-th coupling mod-",
"Step 2 Set the control variable Qin as 1. Map the con-",
"ule. Each coupling module couples the output of a main",
"trol variable Qin to different coupling coefficients",
"block and a tuning block as follows:",
"[Qm],fix parameters of themain branch and",
"train the tuning branch with another loss function",
"Bm = Fm(Rm,Tm)",
"L2(Irec,Ig; Otun, 0α).",
"(3)",
"= Fm(Fmain(Bm-1), Ftun(Bm-1))",
"3.2. Coupling module",
"where Fm () represents the m-th coupling operation, Rm",
"We now present the details of our coupling module. We",
"and Tm denote the output features of m-th main block and",
"mainly introduce our design from the perspective of image",
"m-th tuning block respectively, Fmain() and Ftun() are",
"super-resolution. In order to balance the trade-off between",
"the m-th main block function and m-th tuning block func-",
"perceptual quality and distortion, we usually realize it by",
"tion respectively.To address theimage super-resolution",
"modifying the penality parameter ^ of the loss terms [5],",
"task, we add an extra coupling module consisting of the up-",
"scaling block before the last convolutional layer, as shown",
"Lgen =Ldistortion +ALadu,",
"(6)"
] |
{
"arxiv_id": "1904.00634",
"language": "en",
"page_idx": 2
}
|
||
1904.00634-en-3
|
1904.00634-en-3.jpg
|
where $L_distortion$ denotes distortion loss (e.g., MSE and MAE), $L_adv$ contains GAN loss [13,37] and perceptual loss [17], λ is a scalar. We usually pre-train the network with $L_distortion$ loss, then we fine-tune the network with combined loss $L_{gen}$ to reach a different working point for the trade-off determined by the value of $\lambda.$ That is to say, if we regard pre-trained results as a reference point, then we can start from the reference point and gradually convert it to the result of another optimization goal.
However, it is not efficient to train a network for each different value of $\lambda.$ In order to address this issue, we convert the control scalar $\lambda$ to the input and directly control the offset of the reference point in latent feature space. For this purpose, we implement a controllable coupling module to couple reference features learned with $L_{distortion}$ and new features learned with $L_gen$ together. We set the feature based on distortion optimization as the reference point which is denoted as $R_m.$ In the process of optimization based on perceptual quality, we keep the reference point unchanged and set $T_m-R_m$ as direction of change. In other words, in a coupling module, part of features are provided by reference information, and the other part are obtained from new exploration:
(7)
$$B_m=R_m+\alpha_m(T_m-R_m)=(1-\alpha_m)R_m+\alpha_mT_m,$$
where $T_m\in\mathbb{R}^{W\times H\times C},R_m\in\mathbb{R}^{W\times H\times C},\alpha_m\in\mathbb{R}^C$ denotes the $m$-th coefficient, and $C$ is the number of channels.
It is worth noting that different main blocks provide different reference information, so we should treat them differently. We expect to endow each coupling module a different coupling coefficient to make full use of reference information. Therefore, our control coefficients $\{\alpha_m\}$ are learned from optimization process. To be more specific, we use some fully connected layers to map a single input control scalar $\alpha_{in}$ into different coupling coefficients (Eq. 2). The proposed network will find the optimal coupling mode since our control coefficients $\{\alpha_m\}$ are adaptive not fixed.
Thanks to the coupling module and adaptive learning strategy, we can realize continuous and smooth transition by a single control variable $\alpha_{in}.$ Moreover, if this framework achieves an excellent trade-off between perceptual quality and distortion for super-resolution, then can we generalize this model to other restoration tasks? After the theoretical analysis and experimental tests, we find that this framework is applicable to a wide variety of image restoration tasks. In the next section, we will provide a more general theoretical explanation of our model.
## 3.3. Theoretical analysis
Suppose there is a high-dimensional space containing all natural images. The degradation process of a natural image can be regarded as continuous in the space. So approximately, these degraded images are adjacent in the space. It
<FigureHere>
Figure 3. Neural network mapping gradually disentangles data manifolds. We can represent unknown point with known point in latent space.
is possible to approximate the reconstruction result of unknown degradation level with the results of known degradation levels. Unfortunately, natural images lie on an approximate non-linear manifold [38]. As a result, simple image interpolation tends to introduce some ghosting artifacts or other unexpected details to final results.
Instead of operating in the pixel space, we naturally turn our attention to the feature space. Some literatures indicate that the data manifold can be flattened by neural network mapping and we can approximate the mapped manifold as a Euclidean space [2,6,32]. Based on this hypothesis, as shown in Fig. 3, we denote $X_i$ and $Y_i$ in the latent space as two endpoints respectively, and we can represent unknown point $Z_i$ as a affine combination of $X_i$ and $Y_i:$
$$Z_i\approx\alpha_iX_i+(1-\alpha_i)Y_i,$$
(8)
where $\alpha_i$ is the $i$-th combination coefficient. This is exactly the formula of the controllable coupling module Eq. 7. However, we should also note that this hypothesis is influenced by the depth and width of CNN. In other words, we do not know the degree of flattening that can be achieved by different channels and different layers. Thus the combination coefficients of different channels and different layers should be discrepant. Besides, we hope to find the optimal combination coefficients through the optimization process:
$$\alpha^{*}=\arg\min_{\alpha_{ij}}\:\sum_{i=1}^{M}\sum_{j=1}^{C}\left[Z_{ij}-\left(\alpha_{ij}X_{ij}+\left(1-\alpha_{ij}\right)Y_{ij}\right)\right],$$
where $\alpha^*=\{\alpha_{i,j}|i=1\cdots M,j=1\cdots C\}$ represents the
optimal solution. However, it is difficult to directly obtain $\hat{\text{the optimal }\alpha^*}$,because the unknown working point $Z_ij$ cannot in general be computed tractably. So we solve Eq. 9 in an implicit way. Specifically, we map the input control variable $\alpha_{in}$ to different combination coefficients with some stacked fully connected layers, and then, we can approximate the above process into optimizing the parameters of the linear mapping network:
(10)
$$\alpha^*\approx\hat{\alpha}=F_{alpha}(\alpha_{in};\theta_{\alpha}),$$
|
[
[
0.08087158203125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.08087158203125,
0.106689453125
],
[
0.66845703125,
0.0985107421875,
0.69921875,
0.0985107421875,
0.69921875,
0.121826171875,
0.66845703125,
0.121826171875
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.6044921875,
0.1142578125,
0.62353515625,
0.1142578125,
0.62353515625,
0.1817626953125,
0.6044921875,
0.1817626953125
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.66845703125,
0.12060546875,
0.72216796875,
0.12060546875,
0.72216796875,
0.1351318359375,
0.66845703125,
0.1351318359375
],
[
0.73876953125,
0.125,
0.75341796875,
0.125,
0.75244140625,
0.1767578125,
0.73779296875,
0.1767578125
],
[
0.765625,
0.1256103515625,
0.77783203125,
0.1351318359375,
0.76708984375,
0.143310546875,
0.7548828125,
0.1337890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.15283203125,
0.080078125,
0.15283203125
],
[
0.080078125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.7509765625,
0.15283203125,
0.765625,
0.146484375,
0.775390625,
0.160400390625,
0.7607421875,
0.1666259765625
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.75732421875,
0.17236328125,
0.7685546875,
0.17236328125,
0.7685546875,
0.1875,
0.75732421875,
0.1875
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.68310546875,
0.18310546875,
0.69140625,
0.18310546875,
0.69140625,
0.1875,
0.68310546875,
0.1875
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.7548828125,
0.1944580078125,
0.76708984375,
0.1944580078125,
0.76708984375,
0.2052001953125,
0.7548828125,
0.2052001953125
],
[
0.080078125,
0.21337890625,
0.35302734375,
0.21337890625,
0.35302734375,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.101318359375,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.240478515625,
0.101318359375,
0.240478515625
],
[
0.50244140625,
0.2279052734375,
0.89208984375,
0.2279052734375,
0.89208984375,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.242431640625,
0.89208984375,
0.242431640625,
0.89208984375,
0.255615234375,
0.50341796875,
0.255615234375
],
[
0.08087158203125,
0.25830078125,
0.4697265625,
0.25830078125,
0.4697265625,
0.271484375,
0.08087158203125,
0.271484375
],
[
0.50341796875,
0.25830078125,
0.59228515625,
0.25830078125,
0.59228515625,
0.26953125,
0.50341796875,
0.26953125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.080078125,
0.28857421875,
0.46728515625,
0.28857421875,
0.46728515625,
0.3017578125,
0.080078125,
0.3017578125
],
[
0.50341796875,
0.29345703125,
0.890625,
0.29345703125,
0.890625,
0.306884765625,
0.50341796875,
0.306884765625
],
[
0.07843017578125,
0.302978515625,
0.4697265625,
0.3037109375,
0.4697265625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.309326171875,
0.890625,
0.309326171875,
0.890625,
0.322509765625,
0.50244140625,
0.322509765625
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333984375,
0.080078125,
0.333984375
],
[
0.5009765625,
0.323974609375,
0.890625,
0.324462890625,
0.890625,
0.33837890625,
0.5009765625,
0.337158203125
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.50244140625,
0.354736328125,
0.89208984375,
0.354736328125,
0.89208984375,
0.36865234375,
0.50244140625,
0.36865234375
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.369384765625,
0.7646484375,
0.369384765625,
0.7646484375,
0.38330078125,
0.50244140625,
0.38330078125
],
[
0.07843017578125,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.07843017578125,
0.392578125
],
[
0.52294921875,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.52294921875,
0.3984375
],
[
0.08087158203125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.50244140625,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.50244140625,
0.4140625
],
[
0.08087158203125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.50341796875,
0.415283203125,
0.89208984375,
0.415283203125,
0.89208984375,
0.4287109375,
0.50341796875,
0.4287109375
],
[
0.080078125,
0.4248046875,
0.2303466796875,
0.4248046875,
0.2303466796875,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.5009765625,
0.429931640625,
0.890625,
0.429931640625,
0.890625,
0.443115234375,
0.5009765625,
0.443115234375
],
[
0.0882568359375,
0.449462890625,
0.468017578125,
0.45068359375,
0.468017578125,
0.46533203125,
0.0882568359375,
0.464111328125
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.443115234375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458984375
],
[
0.50341796875,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.07843017578125,
0.4716796875,
0.471435546875,
0.4716796875,
0.471435546875,
0.492431640625,
0.07843017578125,
0.492431640625
],
[
0.50244140625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.4892578125,
0.50244140625,
0.4892578125
],
[
0.080078125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.080078125,
0.5048828125
],
[
0.50244140625,
0.490478515625,
0.80615234375,
0.490478515625,
0.80615234375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.09722900390625,
0.5068359375,
0.468017578125,
0.5068359375,
0.468017578125,
0.52099609375,
0.09722900390625,
0.52099609375
],
[
0.61279296875,
0.51513671875,
0.78173828125,
0.51513671875,
0.78173828125,
0.52978515625,
0.61279296875,
0.52978515625
],
[
0.869140625,
0.51513671875,
0.8935546875,
0.51513671875,
0.8935546875,
0.52880859375,
0.869140625,
0.52880859375
],
[
0.080078125,
0.52294921875,
0.4697265625,
0.52294921875,
0.4697265625,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.080078125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.5517578125,
0.080078125,
0.5517578125
],
[
0.50732421875,
0.53857421875,
0.890625,
0.5390625,
0.890625,
0.5537109375,
0.50732421875,
0.55322265625
],
[
0.07843017578125,
0.55322265625,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.07843017578125,
0.56640625
],
[
0.5009765625,
0.5537109375,
0.890625,
0.5546875,
0.890625,
0.5693359375,
0.5009765625,
0.568359375
],
[
0.080078125,
0.568359375,
0.4697265625,
0.568359375,
0.4697265625,
0.58203125,
0.080078125,
0.58203125
],
[
0.50341796875,
0.57080078125,
0.89208984375,
0.57080078125,
0.89208984375,
0.583984375,
0.50341796875,
0.583984375
],
[
0.080078125,
0.58349609375,
0.4697265625,
0.58349609375,
0.4697265625,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.50341796875,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.08087158203125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.60986328125,
0.08087158203125,
0.60986328125
],
[
0.50244140625,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.080078125,
0.62548828125
],
[
0.50341796875,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.6298828125,
0.50341796875,
0.6298828125
],
[
0.07843017578125,
0.62744140625,
0.4697265625,
0.626953125,
0.4697265625,
0.6416015625,
0.07843017578125,
0.642578125
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.080078125,
0.64306640625,
0.42236328125,
0.64306640625,
0.42236328125,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.50244140625,
0.646484375,
0.89208984375,
0.646484375,
0.89208984375,
0.66015625,
0.50244140625,
0.66015625
],
[
0.09722900390625,
0.6572265625,
0.468017578125,
0.65771484375,
0.468017578125,
0.67236328125,
0.09722900390625,
0.671875
],
[
0.5009765625,
0.6611328125,
0.8896484375,
0.662109375,
0.8896484375,
0.67529296875,
0.5009765625,
0.6748046875
],
[
0.080078125,
0.673828125,
0.46728515625,
0.673828125,
0.46728515625,
0.6875,
0.080078125,
0.6875
],
[
0.61279296875,
0.681640625,
0.65673828125,
0.681640625,
0.65673828125,
0.6962890625,
0.61279296875,
0.6962890625
],
[
0.07843017578125,
0.68798828125,
0.4697265625,
0.68798828125,
0.4697265625,
0.7021484375,
0.07843017578125,
0.7021484375
],
[
0.50634765625,
0.69873046875,
0.60107421875,
0.69873046875,
0.60107421875,
0.71044921875,
0.50634765625,
0.71044921875
],
[
0.609375,
0.70263671875,
0.61767578125,
0.69970703125,
0.62158203125,
0.70703125,
0.61376953125,
0.71044921875
],
[
0.62744140625,
0.69873046875,
0.88623046875,
0.69873046875,
0.88623046875,
0.7119140625,
0.62744140625,
0.7119140625
],
[
0.56396484375,
0.71337890625,
0.58251953125,
0.71337890625,
0.58251953125,
0.7216796875,
0.56396484375,
0.7216796875
],
[
0.080078125,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.080078125,
0.71728515625
],
[
0.08087158203125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73095703125,
0.08087158203125,
0.73095703125
],
[
0.609375,
0.71533203125,
0.66015625,
0.71533203125,
0.66015625,
0.7265625,
0.609375,
0.7265625
],
[
0.87109375,
0.724609375,
0.89208984375,
0.724609375,
0.89208984375,
0.73779296875,
0.87109375,
0.73779296875
],
[
0.080078125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50341796875,
0.74072265625,
0.89208984375,
0.74072265625,
0.89208984375,
0.7548828125,
0.50341796875,
0.7548828125
],
[
0.08087158203125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.76123046875,
0.08087158203125,
0.76123046875
],
[
0.50244140625,
0.755859375,
0.89208984375,
0.755859375,
0.89208984375,
0.76953125,
0.50244140625,
0.76953125
],
[
0.080078125,
0.7646484375,
0.4697265625,
0.7646484375,
0.4697265625,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.5009765625,
0.76953125,
0.890625,
0.771484375,
0.890625,
0.7861328125,
0.5009765625,
0.7841796875
],
[
0.080078125,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.78515625,
0.890625,
0.7861328125,
0.890625,
0.80029296875,
0.50244140625,
0.7998046875
],
[
0.080078125,
0.79443359375,
0.25,
0.79443359375,
0.25,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.50244140625,
0.8017578125,
0.890625,
0.8017578125,
0.890625,
0.81494140625,
0.50244140625,
0.81494140625
],
[
0.080078125,
0.8173828125,
0.265625,
0.8173828125,
0.265625,
0.83203125,
0.080078125,
0.83203125
],
[
0.5009765625,
0.81640625,
0.89208984375,
0.81494140625,
0.89208984375,
0.830078125,
0.5009765625,
0.83056640625
],
[
0.50244140625,
0.83203125,
0.890625,
0.83203125,
0.890625,
0.84521484375,
0.50244140625,
0.84521484375
],
[
0.09722900390625,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.09722900390625,
0.85498046875
],
[
0.5009765625,
0.8466796875,
0.890625,
0.8466796875,
0.890625,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.8623046875,
0.708984375,
0.8623046875,
0.708984375,
0.87548828125,
0.50244140625,
0.87548828125
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.6044921875,
0.88525390625,
0.7890625,
0.88525390625,
0.7890625,
0.90283203125,
0.6044921875,
0.90283203125
],
[
0.859375,
0.88525390625,
0.89453125,
0.88525390625,
0.89453125,
0.90087890625,
0.859375,
0.90087890625
],
[
0.07843017578125,
0.88623046875,
0.4697265625,
0.88525390625,
0.4697265625,
0.8994140625,
0.07843017578125,
0.90087890625
]
] |
[
"where Ldistortion denotes distortion loss (e.g., MSE and",
"D",
"MAE), Ladu contains GAN loss [13, 37] and perceptual",
"er",
"loss [17], is a scalar. We usually pre-train the network",
"mapping",
"e",
"古",
"with Ldistortion loss, then we fine-tune the network with",
"combined loss Lgen to reach a different working point for",
"之",
"thetrade-off determined by thevalue of .That is to say,if",
"",
"we regard pre-trained results as a reference point, then we",
"O",
"can startfromthereferencepointand graduallyconvertit",
"Z.",
"to the result of another optimization goal.",
"However,itisnotefficienttotrain anetwork for each",
"Figure 3. Neural network mapping gradually disentangles data",
"different value of >. In order to address this issue, we con-",
"manifolds. We can represent unknown point with known point",
"vert the control scalar to the input and directly control",
"in latent space.",
"the offset of the reference point in latent feature space. For",
"this purpose, we implement a controllable coupling module",
"is possible to approximate the reconstruction result of un-",
"to couple reference features learned with Ldistortion and",
"known degradation level with the results of known degrada-",
"new features learned with Lgen together. We set the fea-",
"tion levels.Unfortunately,natural images lie on an approx-",
"ture based on distortion optimization as the reference point",
"imate non-linear manifold [38]. As a result, simple image",
"which is denoted as Rm-In the process of optimization",
"interpolation tends to introduce some ghosting artifacts or",
"based on perceptual quality, we keep the reference point un-",
"other unexpected details to final results.",
"changed and set Tm - Rm as direction of change. In other",
"Instead of operating in the pixel space, we naturally turn",
"words, in a coupling module, part of features are provided",
"our attention to the feature space. Some literatures indicate",
"by reference information, and the other part are obtained",
"that the data manifold can be flattened by neural network",
"from new exploration:",
"mapping and we can approximate the mapped manifold as",
"Bm=Rm+am(Tm-Rm)=(1-Qm)Rm+amTm, (7)",
"a Euclidean space [2, 6, 32]. Based on this hypothesis, as",
"shown in Fig. 3, we denote X, and Y, in the latent space as",
"where Tm E RWxHxC, Rm E RWxHxC, am E RC de-",
"two endpoints respectively, and we can represent unknown",
"notes the m-th coefficient, and C is the number of channels.",
"point Z as a affine combination of X; and Yi:",
"It is worth noting that different main blocks provide dif-",
"Z, ~ QiX, + (1 -Q)Yi,",
"(8)",
"ferent reference information, so we should treat them differ-",
"ently.We expectto endow each coupling module a different",
"where Q is the i-th combination coefficient. This is ex-",
"coupling coefficient to make full use of reference informa-",
"actly the formula of the controllable coupling module Eq",
"tion. Therefore, our control coefficients [am] are learned",
"7. However, we should also note that this hypothesis is in-",
"from optimization process. To be more specific, we use",
"fluenced by the depth and width of CNN. In other words, we",
"some fully connected layers to mapa single input control",
"do not know the degree of flattening that can be achieved by",
"scalar Qin into different coupling coefficients (Eg. 2). The",
"different channels and different layers. Thus the combina-",
"proposed network willfind the optimal coupling mode since",
"tion coeffcients of different channels and different layers",
"our control coefficients (αm] are adaptive not fixed.",
"should be discrepant. Besides, we hope to find the optimal",
"Thanks to the coupling module and adaptive learning",
"combination coefficientsthroughthe optimization process:",
"strategy, we can realize continuous and smooth transition by",
"MC",
"a single control variable Qin.Moreover, if this framework",
"α* = argmin",
"",
">[Zi- (aiXi + (1-Q)Y)],",
"aij",
"achieves an excellent trade-off between perceptual quality",
"and distortionfor super-resolution,thencanwegeneralize",
"i=1 j=1",
"(9)",
"this model to other restoration tasks? After the theoretical",
"where α* = {αi.li = 1... M, j = 1...C] represents the",
"analysisandexperimental tests,wefindthatthisframework",
"optimal solution. However, it is difficult to directly obtain",
"is applicable to a wide variety of image restoration tasks. In",
"the optimal α*, because the unknown working point Zij",
"the next section, we will provide a more general theoretical",
"cannot in general be computed tractably. So we solve Eq.",
"explanation of our model.",
"9 in an implicit way. Specifically, we map the input con-",
"3.3. Theoretical analysis",
"trol variable Qin to different combination coefficients with",
"some stacked fully connected layers, and then, we can ap-",
" Suppose there is a high-dimensional space containing all",
"proximate the above process into optimizing the parameters",
"natural images. The degradation process of a natural image",
"of the linear mapping network:",
"can be regarded as continuous in the space. So approxi-",
"a* ~ α= Falpha(αin; 0a),",
"(10)",
"mately, these degraded images are adjacent in the space. It"
] |
{
"arxiv_id": "1904.00634",
"language": "en",
"page_idx": 3
}
|
||
1904.00634-en-4
|
1904.00634-en-4.jpg
|
where $F_{alpha}$ denotes mapping function of $\alpha_in,\hat{\alpha}$ is the approximated solution of the optimal $\alpha^*.$ Fortunately, this network can be embedded into our framework. Therefore, we can optimize the parameters of the linear mapping network and the tuning blocks in one shot. The entire optimization process (corresponding to Step 2) can be expressed as
$$\theta_{tun},\theta_{\alpha}=\arg\min_{\theta_{tun},\theta_{\alpha}}\:L_{2}(F_{CFSN}(I_{in},\alpha_{in};\theta_{tun},\theta_{\alpha}),I_{g}),$$
(11)
## 3.4. Discussions
Difference to Dynamic-Net Recently, Dynamic-Net [32] realized interactive control of continuous image conversion. In contrast, there are two main differences between Dynamic-Net and our CFSNet. First of all, Dynamic-Net is mainly designed for image transformation tasks, like style transfer. It's difficult to achieve the desirable results when we use Dynamic-Net directly for image restoration tasks. While motivated by super-resolution, we design the proposed CFSNet for low-level image restoration. Secondly, in Dynamic-Net, as shown in Fig. 4, they directly tune multiple scalars $\{\alpha_m\}$ to get different outputs. While, in our framework coupling coefficient $\alpha_m$ is a vector and is learned adaptively from the input scalar $\alpha_{in}$ through the optimization process. This is more user-friendly and we explain the reasonability of this design in Sec. 3.3.
<FigureHere>
Figure 4. Basic module. Yellow and orange bar represent main
block and tuning block respectively.
Difference to Deep Network Interpolation Deep Network Interpolation (DNI) is another choice to control the compromise between perceptual quality and distortion [36, 37]. DNI also can be applied to many low-level vision tasks [36]. However, this method needs to train two networks with the same architecture but different losses and will generate a third network to control. In contrast, our framework can achieve better interactive control with a unified end-to-end network. Moreover, our framework makes better use of reference information using coupling module. DNI performs interpolation in the parameter space to generate continuous transition effects, and interpolation coefficients are kept the same in the whole parameter space. However, this simple strategy can not guarantee the optimality of the outputs. While in our CFSNet, we perform the interpolation in the feature space and the continuous transition of the reconstruction effect is consistent with the variation of
the control variable $\alpha_{in}.$ We can produce a better approximation of the unknown working point. See Sec. 4.3 for more experimental comparisons.
## $4. \textbf{Experiments}$
In this section, we first demonstrate implementation details of our framework. Then we validate the control mechanism of our CFSNet. Finally, we apply our CFSNet to three classic tasks: image super-resolution, image denoising, JPEG image deblocking. All experiments validate the effectiveness of our model. Due to space limitations, more examples and analyses are provided in the appendix.
### 4.1. Implement Details
For image super-resolution task, our framework contains 30 main blocks and 30 tuning blocks (i.e., $M=30).$ As for the other two tasks, the main branch parameters of our CFSNet are kept similar to that of the compared method [41](i.e., $M=10)$ for a fair comparison. Furthermore, we first generate a 512-dimensional vector with values of all 1. Then we multiply it by the control scalar $\alpha_{in}$ to produce a control input. All convolutional layers have 64 filters and the kernel size of each convolutional layer is $3\times3.$ We use the method in [15] to perform weight initialization. For both training stage of all tasks, we use the ADAM optimizer [20] by setting $\beta_1=0.9,\beta_2=0.999$, and $\varepsilon=10^-8$ with the initial learning rate le-4. We adopt 128 as the minibatch size in image denoising task and set it as 16 in the other three tasks. We use PyTorch to implement our network and perform all experiments on GTX 1080Ti GPU.
For image denoising and JPEG image deblocking, we follow the settings as in [41] and [9] respectively. The training loss function in Step 1 and Step 2 remains un- ${\mathrm{changed: }}$ $L_{1}( I_{rec}, I_{g})$ = $L_{2}( I_{rec}, I_{g}) .$ ${\mathrm{In~particular, ~for}}$ image denoising, we input the degraded images of noise level 25 when we train the main branch in Step 1 and we input the degraded images of noise level 50 when we train the tuning branch in Step 2. Training images are cut into $40\times40$ patches with a stride of 10. And the learning rate is reduced by 10 times every 50000 steps. For JPEG deblocking, we set quality factor as 10 in the first training stage and change it to 40 in the second training stage. Besides, we choose $48\times48$ as patch size and the learning rate is divided by 10 every 1000000 steps. For image super-resolution, we first train the main branch with objective MAE loss, then we train the tuning branch with ob- ${\mathrm{jective~}}L_{2}=L_{mae}+0.01L_{gan}+0.01L_{per}$,where $L_mae$ denotes mean absolute error $(\bar{\mathrm{MAE}}),L_{gan}$ represents wgangp loss [13] and $L_per$ is a variant of perceptual loss [37]. We set HR patch size as $128\times128$ and we multiply the learning rate by 0.6 every 400000 steps.
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.5009765625,
0.12371826171875,
0.7197265625,
0.12371826171875,
0.7197265625,
0.136962890625,
0.5009765625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.0767822265625,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.0767822265625,
0.1666259765625
],
[
0.50341796875,
0.160400390625,
0.63134765625,
0.160400390625,
0.63134765625,
0.1749267578125,
0.50341796875,
0.1749267578125
],
[
0.080078125,
0.1685791015625,
0.435546875,
0.1685791015625,
0.435546875,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.521484375,
0.1881103515625,
0.89208984375,
0.1881103515625,
0.89208984375,
0.2020263671875,
0.521484375,
0.2020263671875
],
[
0.08660888671875,
0.1976318359375,
0.4599609375,
0.1976318359375,
0.4599609375,
0.212158203125,
0.08660888671875,
0.212158203125
],
[
0.50244140625,
0.203857421875,
0.89208984375,
0.203857421875,
0.89208984375,
0.2171630859375,
0.50244140625,
0.2171630859375
],
[
0.1641845703125,
0.2108154296875,
0.214111328125,
0.2108154296875,
0.214111328125,
0.222900390625,
0.1641845703125,
0.222900390625
],
[
0.438720703125,
0.2215576171875,
0.468017578125,
0.2215576171875,
0.468017578125,
0.2342529296875,
0.438720703125,
0.2342529296875
],
[
0.50341796875,
0.2197265625,
0.89208984375,
0.2197265625,
0.89208984375,
0.23291015625,
0.50341796875,
0.23291015625
],
[
0.50244140625,
0.23486328125,
0.890625,
0.23486328125,
0.890625,
0.248779296875,
0.50244140625,
0.248779296875
],
[
0.07843017578125,
0.2474365234375,
0.2017822265625,
0.248779296875,
0.2017822265625,
0.26318359375,
0.07843017578125,
0.261962890625
],
[
0.50244140625,
0.25,
0.89208984375,
0.25,
0.89208984375,
0.264404296875,
0.50244140625,
0.264404296875
],
[
0.50244140625,
0.26513671875,
0.89208984375,
0.26513671875,
0.89208984375,
0.279052734375,
0.50244140625,
0.279052734375
],
[
0.09722900390625,
0.272705078125,
0.468017578125,
0.272705078125,
0.468017578125,
0.285888671875,
0.09722900390625,
0.285888671875
],
[
0.5009765625,
0.281005859375,
0.8486328125,
0.281005859375,
0.8486328125,
0.294189453125,
0.5009765625,
0.294189453125
],
[
0.08087158203125,
0.28857421875,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.08087158203125,
0.3017578125
],
[
0.08087158203125,
0.304931640625,
0.468017578125,
0.304931640625,
0.468017578125,
0.31494140625,
0.08087158203125,
0.31494140625
],
[
0.08087158203125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.50341796875,
0.313232421875,
0.6787109375,
0.313232421875,
0.6787109375,
0.32763671875,
0.50341796875,
0.32763671875
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.521484375,
0.339111328125,
0.890625,
0.339111328125,
0.890625,
0.354248046875,
0.521484375,
0.354248046875
],
[
0.08087158203125,
0.35107421875,
0.468017578125,
0.35107421875,
0.468017578125,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.354736328125,
0.8935546875,
0.354736328125,
0.8935546875,
0.369384765625,
0.50341796875,
0.369384765625
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50341796875,
0.37060546875,
0.89208984375,
0.37060546875,
0.89208984375,
0.3837890625,
0.50341796875,
0.3837890625
],
[
0.080078125,
0.376953125,
0.4697265625,
0.37744140625,
0.4697265625,
0.392578125,
0.080078125,
0.391357421875
],
[
0.50244140625,
0.385009765625,
0.890625,
0.385009765625,
0.890625,
0.3984375,
0.50244140625,
0.3984375
],
[
0.080078125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.50244140625,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.4091796875,
0.4697265625,
0.4091796875,
0.4697265625,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50244140625,
0.416748046875,
0.890625,
0.416748046875,
0.890625,
0.429931640625,
0.50244140625,
0.429931640625
],
[
0.07843017578125,
0.423583984375,
0.4697265625,
0.42236328125,
0.4697265625,
0.4375,
0.07843017578125,
0.438232421875
],
[
0.50244140625,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.080078125,
0.439453125,
0.4697265625,
0.439453125,
0.4697265625,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.50244140625,
0.460205078125
],
[
0.08087158203125,
0.455810546875,
0.46728515625,
0.455810546875,
0.46728515625,
0.466552734375,
0.08087158203125,
0.466552734375
],
[
0.50341796875,
0.4609375,
0.89208984375,
0.4609375,
0.89208984375,
0.474853515625,
0.50341796875,
0.474853515625
],
[
0.080078125,
0.4697265625,
0.2303466796875,
0.4697265625,
0.2303466796875,
0.48291015625,
0.080078125,
0.48291015625
],
[
0.226318359375,
0.474853515625,
0.233642578125,
0.474853515625,
0.233642578125,
0.477783203125,
0.226318359375,
0.477783203125
],
[
0.2296142578125,
0.4697265625,
0.4697265625,
0.4697265625,
0.4697265625,
0.48291015625,
0.2296142578125,
0.48291015625
],
[
0.50244140625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.4892578125,
0.50244140625,
0.4892578125
],
[
0.07843017578125,
0.484130859375,
0.41259765625,
0.48291015625,
0.41259765625,
0.49755859375,
0.07843017578125,
0.498779296875
],
[
0.50244140625,
0.490478515625,
0.890625,
0.4892578125,
0.890625,
0.50390625,
0.50244140625,
0.5048828125
],
[
0.50244140625,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.50244140625,
0.51953125
],
[
0.39453125,
0.51513671875,
0.449462890625,
0.51513671875,
0.449462890625,
0.5263671875,
0.39453125,
0.5263671875
],
[
0.369384765625,
0.521484375,
0.386474609375,
0.521484375,
0.386474609375,
0.52880859375,
0.369384765625,
0.52880859375
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.50244140625,
0.5361328125,
0.89208984375,
0.5361328125,
0.89208984375,
0.5498046875,
0.50244140625,
0.5498046875
],
[
0.1683349609375,
0.52294921875,
0.1805419921875,
0.52294921875,
0.1805419921875,
0.52880859375,
0.1683349609375,
0.52880859375
],
[
0.10296630859375,
0.54248046875,
0.11517333984375,
0.54443359375,
0.11273193359375,
0.5498046875,
0.101318359375,
0.54736328125
],
[
0.42236328125,
0.54345703125,
0.44287109375,
0.54345703125,
0.44287109375,
0.55029296875,
0.42236328125,
0.55029296875
],
[
0.50244140625,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.11517333984375,
0.5712890625,
0.25244140625,
0.5712890625,
0.25244140625,
0.58544921875,
0.11517333984375,
0.58544921875
],
[
0.50244140625,
0.56640625,
0.810546875,
0.56494140625,
0.810546875,
0.580078125,
0.50244140625,
0.58056640625
],
[
0.279296875,
0.57275390625,
0.44677734375,
0.57275390625,
0.44677734375,
0.58349609375,
0.279296875,
0.58349609375
],
[
0.521484375,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.080078125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.5009765625,
0.5986328125,
0.89208984375,
0.599609375,
0.89208984375,
0.6142578125,
0.5009765625,
0.61279296875
],
[
0.07843017578125,
0.6142578125,
0.294189453125,
0.6142578125,
0.294189453125,
0.62744140625,
0.07843017578125,
0.62744140625
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.098876953125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.098876953125,
0.65771484375
],
[
0.50244140625,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.5009765625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.705078125,
0.46728515625,
0.705078125,
0.46728515625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.5009765625,
0.7041015625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.5009765625,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.5048828125,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.4697265625,
0.767578125,
0.4697265625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.796875,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.5009765625,
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
],
[
0.498291015625,
0.82373046875,
0.89208984375,
0.82470703125,
0.89208984375,
0.84228515625,
0.498291015625,
0.84130859375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.89208984375,
0.84130859375,
0.89208984375,
0.85595703125,
0.5009765625,
0.85498046875
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.869140625
],
[
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
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.08087158203125,
0.88818359375,
0.468017578125,
0.88818359375,
0.468017578125,
0.8984375,
0.08087158203125,
0.8984375
],
[
0.50244140625,
0.88623046875,
0.7099609375,
0.88623046875,
0.7099609375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"where Falpha denotes mapping function of αin, Q is the ap-",
"the control variable Qin. We can produce a better approx-",
"proximated solution of the optimal α*. Fortunately, this net-",
"imation of the unknown working point. See Sec. 4.3 for",
"work can beembedded intoourframework.Therefore,we",
"more experimental comparisons.",
"can optimize the parameters of the linear mapping network",
"and the tuning blocks in one shot. The entire optimization",
"4.Experiments",
"process (corresponding to Step 2) can be expressed as",
"In this section, we first demonstrate implementation de-",
"Otun,Oα=argmin L2(FcFsN(Iin,Qin;Otun,Oα), Ig)",
"tails of our framework. Then we validate the control mech-",
"Otun.da",
"(11)",
"anism of our CFSNet. Finally, we apply our CFSNet to",
"three classic tasks: image super-resolution, image denois-",
"3.4.Discussions",
"ing, JPEG image deblocking. All experiments validate the",
"effectiveness of our model. Due to space limitations, more",
"Difference to Dynamic-Net Recently, Dynamic-Net",
"examples and analyses are provided in the appendix.",
"[32] realized interactive control of continuous image con-",
"version.Incontrast.therearetwomaindifferencesbetween",
"Dynamic-Net and our CFSNet. First of all, Dynamic-Net is",
"4.1. Implement Details",
"mainly designed for image transformation tasks, like style",
"For image super-resolution task, our framework contains",
"transfer.It's difficultto achieve thedesirableresults when",
"30 main blocks and 30 tuning blocks (i.e., M = 30). As",
"we use Dynamic-Net directly for image restoration tasks.",
"for the other two tasks, the main branch parameters of our",
"While motivated by super-resolution, we design the pro-",
"CFSNet are kept similar to that of the compared method",
"posed CFSNet for low-level image restoration.Secondly,",
"[41] (i.e., M = 10) for a fair comparison. Furthermore, we",
"in Dynamic-Net, as shown in Fig. 4, they directly tune",
"first generate a 512-dimensional vector with values of all 1.",
"multiple scalars am} to get different outputs. While, in",
"Then we multiply it by the control scalar Qin to produce a",
"our framework coupling coefficient am is a vector and is",
"control input. All convolutional layers have 64 filters and",
"learnedadaptivelyfromtheinputscalarainthroughthe",
"the kernel size of each convolutional layer is 3 x 3. We use",
"optimization process.",
"",
"This is more user-friendly and we",
"the method in [15] to perform weight initialization. For both",
"explain the reasonability of this design in Sec. 3.3.",
"training stage of all tasks, we use the ADAM optimizer [20]",
"by setting βi =0.9, β2 =0.999, and = 10-8 with the",
"ame[0,1]",
"Im",
"initial learning rate le-4.We adopt 128 as the minibatch",
"size in image denoising task and set it as 16 in the other",
"X",
"B",
"Bm",
"three tasks. We use PyTorch to implement our network and",
"(a)Basic module of CFSNet",
"perform all experiments onGTX 1080TiGPU.",
"(b)Basic moduleof Dynamic-Net",
"For image denoising and JPEG image deblocking, we",
"Figure 4. Basic module. Yellow and orange bar represent main",
"follow the settings as in [41] and [9] respectively.The",
"block and tuning block respectively.",
"training loss function in Step 1 and Step 2 remains un-",
"changed: Li(Irec,Ig) = L2(Irec,Ig). In particular, for",
"Difference to Deep Network Interpolation Deep Net-",
"image denoising, we input the degraded images of noise",
"work Interpolation (DNI) is another choice to control the",
"level 25 when we train the main branch in Step 1 and we",
"compromise between perceptual quality and distortion [36,",
"input the degraded images of noise level 50 when we train",
"37]. DNI also can be applied to many low-level vision",
"the tuning branch in Step 2. Training images are cut into",
"tasks [36]. However, this method needs to train two net-",
"40×40patcheswithastrideof10.Andthelearning",
"works with the same architecture but different losses and",
"rate is reduced by 10 times every 50000 steps. For JPEG",
"will generate a third network to control.In contrast, our",
"deblocking,we set quality factor as 10 in the first train-",
"framework can achieve better interactive control with a uni-",
"ing stage and change it to 40 in the second training stage.",
"fied end-to-end network.Moreover,our framework makes",
"Besides, we choose 48 × 48 as patch size and the learn-",
"better use of reference information using coupling module.",
"ing rate is divided by 10 every 100000 steps. For image",
"DNI performs interpolation in the parameter space to gen-",
"super-resolution,we firsttrain themain branch with objec-",
"erate continuous transition effects, and interpolation coeffi-",
"tive MAE loss, then we train the tuning branch with ob-",
"cients are kept the same in the whole parameter space. How-",
"jective L2= Lmae + 0.01Lgan + 0.01Lper, where Lmae",
"ever, this simple strategy can not guarantee the optimality of",
"denotes mean absolute error (MAE), Lgan represents wgan-",
"the outputs. While in our CFSNet, we perform the interpo-",
"gp loss [13] and Lper is a variant of perceptual loss [37]. We",
"lation in the feature space and the continuous transition of",
"set HR patch size as 128 × 128 and we multiply the learning",
"the reconstruction effectis consistent withthe variation of",
"rate by 0.6 every 400000 steps."
] |
{
"arxiv_id": "1904.00634",
"language": "en",
"page_idx": 4
}
|
||
2307.07093-en-0
|
2307.07093-en-0.jpg
|
MaxCorrMGNN: A Multi- Graph Neural Network Framework for Generalized Multimodal Fusion of
Medical Data for Outcome Prediction
Niharika S. D'Souza$^1$,Hongzhi Wang$^1$, Andrea Giovannini$^2$, Antonio Foncubierta-Rodriguez$^2$,Kristen L. Beck$^1$,Orest Boyko$^3$,and Tanveer
$\mathrm{Syeda-Mahmood}^{1}$
$^{1}$ IBM Research Almaden, San Jose, CA, USA
$^{2}{\mathrm{~IBM~Research,~Zurich,~Switzerland}}$
$^{3}$ Department of Radiology, VA Southern Nevada Healthcare System, NV, USA
Abstract. With the emergence of multimodal electronic health records, the evidence for an outcome may be captured across multiple modalities ranging from clinical to imaging and genomic data. Predicting outcomes effectively requires fusion frameworks capable of modeling fine-grained and multi-faceted complex interactions between modality features within and across patients. We develop an innovative fusion approach called MaxCorr MGNN that models non-linear modality correlations within and across patients through Hirschfeld-Gebelein-Reñyi maximal correlation (MaxCorr) embeddings, resulting in a multi-layered graph that preserves the identities of the modalities and patients. We then design, for the first time, a generalized multi-layered graph neural network (MGNN) for taskinformed reasoning in multi-layered graphs, that learns the parameters defining patient-modality graph connectivity and message passing in an end-to-end fashion. We evaluate our model an outcome prediction task on a Tuberculosis (TB) dataset consistently outperforming several state-of-the-art neural, graph-based and traditional fusion techniques. Keywords: Multimodal Fusion·Hirschfeld-Gebelein-Reñyi (HGR) max-
imal correlation $\cdot$ Multi-Layered Graphs $\cdot$ Multi-Graph Neural Networks
## 1 Introduction
In the age of modern medicine, it is now possible to capture information about a patient through multiple data-rich modalities to give a holistic view of a patient's condition. In complex diseases such as cancer [18], tuberculosis [16] or autism spectrum disorder [9,8,7], evidence for a diagnosis or treatment outcome may be present in multiple modalities such as clinical, genomic, molecular, pathological and radiological imaging. Reliable patient-tailored outcome prediction requires fusing information from modality data both within and across patients. This can be achieved by effectively modeling the fine-grained and multi-faceted complex interactions between modality features. In general, this is a challenging problem as it is largely unclear what information is best captured by each modality, how
|
[
[
0.2181396484375,
0.143310546875,
0.7861328125,
0.14453125,
0.7861328125,
0.1622314453125,
0.2181396484375,
0.1610107421875
],
[
0.223876953125,
0.1685791015625,
0.78369140625,
0.1685791015625,
0.78369140625,
0.18310546875,
0.223876953125,
0.18310546875
],
[
0.285888671875,
0.1881103515625,
0.71826171875,
0.189453125,
0.71826171875,
0.20703125,
0.285888671875,
0.205810546875
],
[
0.256591796875,
0.236083984375,
0.74853515625,
0.236083984375,
0.74853515625,
0.250732421875,
0.256591796875,
0.250732421875
],
[
0.250732421875,
0.25244140625,
0.75341796875,
0.25244140625,
0.75341796875,
0.266357421875,
0.250732421875,
0.266357421875
],
[
0.43701171875,
0.266357421875,
0.564453125,
0.266357421875,
0.564453125,
0.281005859375,
0.43701171875,
0.281005859375
],
[
0.0236968994140625,
0.2734375,
0.05963134765625,
0.2734375,
0.05963134765625,
0.339111328125,
0.0236968994140625,
0.339111328125
],
[
0.343017578125,
0.290283203125,
0.65869140625,
0.291748046875,
0.65869140625,
0.309326171875,
0.343017578125,
0.30810546875
],
[
0.38232421875,
0.30810546875,
0.6259765625,
0.30810546875,
0.6259765625,
0.3212890625,
0.38232421875,
0.3212890625
],
[
0.233642578125,
0.318115234375,
0.7685546875,
0.3193359375,
0.7685546875,
0.337158203125,
0.233642578125,
0.3359375
],
[
0.02532958984375,
0.330078125,
0.056365966796875,
0.330078125,
0.056365966796875,
0.371826171875,
0.02532958984375,
0.371826171875
],
[
0.02532958984375,
0.3623046875,
0.056365966796875,
0.3623046875,
0.056365966796875,
0.399658203125,
0.02532958984375,
0.399658203125
],
[
0.265625,
0.369384765625,
0.740234375,
0.37060546875,
0.740234375,
0.385009765625,
0.265625,
0.3837890625
],
[
0.2646484375,
0.385009765625,
0.740234375,
0.385009765625,
0.740234375,
0.3984375,
0.2646484375,
0.3984375
],
[
0.0228729248046875,
0.401611328125,
0.062103271484375,
0.401611328125,
0.062103271484375,
0.4873046875,
0.0228729248046875,
0.4873046875
],
[
0.265625,
0.399658203125,
0.73876953125,
0.399658203125,
0.73876953125,
0.41357421875,
0.265625,
0.41357421875
],
[
0.2646484375,
0.41357421875,
0.740234375,
0.41357421875,
0.740234375,
0.4267578125,
0.2646484375,
0.4267578125
],
[
0.265625,
0.4287109375,
0.73779296875,
0.4287109375,
0.73779296875,
0.439453125,
0.265625,
0.439453125
],
[
0.2646484375,
0.44140625,
0.740234375,
0.44140625,
0.740234375,
0.455078125,
0.2646484375,
0.455078125
],
[
0.26318359375,
0.455078125,
0.740234375,
0.455078125,
0.740234375,
0.468505859375,
0.26318359375,
0.468505859375
],
[
0.265625,
0.468505859375,
0.740234375,
0.468505859375,
0.740234375,
0.481689453125,
0.265625,
0.481689453125
],
[
0.2646484375,
0.481689453125,
0.740234375,
0.481689453125,
0.740234375,
0.49560546875,
0.2646484375,
0.49560546875
],
[
0.0228729248046875,
0.492431640625,
0.05963134765625,
0.492431640625,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.265625,
0.49755859375,
0.73876953125,
0.49755859375,
0.73876953125,
0.50830078125,
0.265625,
0.50830078125
],
[
0.2646484375,
0.51025390625,
0.740234375,
0.51025390625,
0.740234375,
0.5234375,
0.2646484375,
0.5234375
],
[
0.26318359375,
0.5234375,
0.74169921875,
0.5234375,
0.74169921875,
0.537109375,
0.26318359375,
0.537109375
],
[
0.2646484375,
0.537109375,
0.73876953125,
0.537109375,
0.73876953125,
0.55029296875,
0.2646484375,
0.55029296875
],
[
0.2646484375,
0.5517578125,
0.740234375,
0.5517578125,
0.740234375,
0.56494140625,
0.2646484375,
0.56494140625
],
[
0.2646484375,
0.56494140625,
0.740234375,
0.56494140625,
0.740234375,
0.5791015625,
0.2646484375,
0.5791015625
],
[
0.2646484375,
0.580078125,
0.7255859375,
0.580078125,
0.7255859375,
0.59326171875,
0.2646484375,
0.59326171875
],
[
0.265625,
0.60595703125,
0.740234375,
0.60595703125,
0.740234375,
0.619140625,
0.265625,
0.619140625
],
[
0.265625,
0.6220703125,
0.73876953125,
0.6220703125,
0.73876953125,
0.6318359375,
0.265625,
0.6318359375
],
[
0.2197265625,
0.6611328125,
0.2320556640625,
0.6611328125,
0.2320556640625,
0.671875,
0.2197265625,
0.671875
],
[
0.2484130859375,
0.66015625,
0.374267578125,
0.66015625,
0.374267578125,
0.673828125,
0.2484130859375,
0.673828125
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.21728515625,
0.70556640625,
0.78759765625,
0.70556640625,
0.78759765625,
0.72021484375,
0.21728515625,
0.72021484375
],
[
0.2181396484375,
0.7216796875,
0.7861328125,
0.7216796875,
0.7861328125,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.2181396484375,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.751953125,
0.7861328125,
0.751953125,
0.7861328125,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.78076171875,
0.2181396484375,
0.78076171875
],
[
0.2181396484375,
0.7822265625,
0.7861328125,
0.7822265625,
0.7861328125,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2156982421875,
0.8125,
0.78515625,
0.8125,
0.78515625,
0.82568359375,
0.2156982421875,
0.82568359375
],
[
0.21728515625,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"MaxCorrMGNN: A Multi-Graph Neural Network",
"Framework for Generalized Multimodal Fusion of",
"Medical Data for Outcome Prediction",
"Niharika S. D'Souzal, Hongzhi Wang', Andrea Giovannini?, Antonio",
"Foncubierta-Rodriguez?, Kristen L. Beck', Orest Boyko3, and Tanveer",
"Syeda-Mahmood1",
"",
"1 IBM Research Almaden, San Jose, CA, USA",
"IBMResearch,Zurich, Switzerland",
"3 Department of Radiology, VA Southern Nevada Healthcare System, NV, USA",
"[nr",
"",
"Abstract. With the emergence of multimodal electronic health records.",
"the evidence for an outcome may be captured across multiple modalities",
"[T'S]",
"ranging from clinical to imaging and genomic data. Predicting outcomes",
"effectively requires fusion frameworks capable of modeling fine-grained",
"and multi-faceted complex interactions between modality features within",
"and across patients. We develop an innovative fusion approach called",
"MaxCorr MGNN that models non-linear modality correlations within and",
"across patients through Hirschfeld-Gebelein-Renyi maximal correlation",
"(MaxCorr) embeddings, resulting in a multi-layered graph that preserves",
"1A600L07A0",
"the identities of the modalities and patients. We then design, for the first",
"time, a generalized multi-layered graph neural network (MGNN) for task-",
"informed reasoning in multi-layered graphs, that learns the parameters",
"defining patient-modality graph connectivity and message passing in",
"an end-to-end fashion. We evaluate our model an outcome prediction",
"task on a Tuberculosis (TB) dataset consistently outperforming several",
"state-of-the-art neural, graph-based and traditional fusion techniques.",
"Keywords: Multimodal Fusion + Hirschfeld-Gebelein-Renyi (HGR) max-",
"imal correlationMulti-Layered Graphs:Multi-Graph Neural Networks",
"1",
"Introduction",
"In the age of modern medicine, it is now possible to capture information about a",
"patient through multiple data-rich modalities to give a holistic view of a patient's",
"condition. In complex diseases such as cancer [18], tuberculosis [16] or autism",
"spectrum disorder [9,8,7], evidence for a diagnosis or treatment outcome may be",
"present in multiple modalities such as clinical, genomic, molecular, pathological",
"and radiological imaging. Reliable patient-tailored outcome prediction requires",
"fusing information from modality data both within and across patients. This can",
"be achieved by effectively modeling the fine-grained and multi-faceted complex",
"interactions between modality features. In general, this is a challenging problem",
"as it is largely unclear what information is best captured by each modality, how"
] |
{
"arxiv_id": "2307.07093",
"language": "en",
"page_idx": 0
}
|
||
2307.07093-en-1
|
2307.07093-en-1.jpg
|
2
N. S. $D^{\prime }$Souza et al.
best to combine modalities, and how to effectively extract predictive patterns
from data [14].
## 1.1 Related Works
Existing attempts to fuse modalities for outcome prediction can be divided into at least three approaches, namely, feature vector-based, statistical or graph-based approaches. The vector-based approaches perform early, intermediate, or late fusion [18,1]with the late fusion approach combining the results of prediction rather than fusing the modality features. Due to the restrictive nature of the underlying assumptions, these are often inadequate for characterizing the broader range of relationships among modality features and their relevance for prediction. In statistical approaches, methods such as canonical correlation analysis [19] and its deep learning variants [25] directly model feature correlations either in the native representation or in a latent space [18]. However, these are not guaranteed to learn discriminative patterns in the unsupervised setting and can suffer from scalability issues when integrated into larger predictive models [23]. Recently, graph-based approaches have been developed which form basic [2,26,5] or multiplexed graphs [6] from latent embeddings derived from modality features using concatenation [2] or weighted averaging [26]. Task-specific fusion is then achieved through inference via message passing walks between nodes in a graph neural network. In the basic collapsed graph construction, the inter-patient and intra-patient modality correlations are not fully distinguished. Conversely, in the multiplexed formulation [6], only a restricted form of multi-relational dependence is captured between nodes through vertical connections. Since the graph is defined using latent embedding directions, the modality semantics are not preserved. Additionally, the staged training of the graph construction and inference networks do not guarantee that the constructed graphs retain discriminable interaction patterns.
## 1.2 Our Contributions
We develop a novel end-to-end fusion framework that addresses the limitations mentioned above. The Maximal Correlation Multi-Layered Graph Neural $\underline{\mathrm{Network,~i.e.~MaxCorrMGNN,~is~a~general~yet~interpretable~framework~for~prob-}}$ lems of multimodal fusion with unstructured data. Specifically, our approach marries the design principles of statistical representation learning with deep learning models for reasoning from multi-graphs.
The main contributions of this work are three-fold:
-First, we propose to model intra and inter-patient modality relationships
explicitly through a novel patient-modality multi-layered graph as shown in Fig. 1. The edges in each layer (plane of the multi-graph) capture the intra-$modality\textit{ relations between patients, while the cross- edges between layers}$ capture $inter- modality\textit{relations across patients. }$
|
[
[
0.2181396484375,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.2181396484375,
0.1275634765625
],
[
0.27294921875,
0.11553955078125,
0.398681640625,
0.11553955078125,
0.398681640625,
0.1287841796875,
0.27294921875,
0.1287841796875
],
[
0.21728515625,
0.147705078125,
0.7861328125,
0.147705078125,
0.7861328125,
0.1622314453125,
0.21728515625,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.325927734375,
0.1634521484375,
0.325927734375,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.2181396484375,
0.20263671875,
0.2467041015625,
0.20263671875,
0.2467041015625,
0.2139892578125,
0.2181396484375,
0.2139892578125
],
[
0.2607421875,
0.2020263671875,
0.3857421875,
0.2020263671875,
0.3857421875,
0.21533203125,
0.2607421875,
0.21533203125
],
[
0.2181396484375,
0.2291259765625,
0.78515625,
0.2291259765625,
0.78515625,
0.242431640625,
0.2181396484375,
0.242431640625
],
[
0.2181396484375,
0.2442626953125,
0.7861328125,
0.2442626953125,
0.7861328125,
0.25830078125,
0.2181396484375,
0.25830078125
],
[
0.21728515625,
0.2587890625,
0.78515625,
0.25830078125,
0.78515625,
0.272705078125,
0.21728515625,
0.2734375
],
[
0.2181396484375,
0.274658203125,
0.78515625,
0.274658203125,
0.78515625,
0.2890625,
0.2181396484375,
0.2890625
],
[
0.21728515625,
0.2890625,
0.7861328125,
0.2890625,
0.7861328125,
0.3037109375,
0.21728515625,
0.3037109375
],
[
0.2181396484375,
0.3037109375,
0.7861328125,
0.3037109375,
0.7861328125,
0.318115234375,
0.2181396484375,
0.318115234375
],
[
0.2181396484375,
0.3193359375,
0.78515625,
0.3193359375,
0.78515625,
0.333251953125,
0.2181396484375,
0.333251953125
],
[
0.21728515625,
0.333251953125,
0.78515625,
0.333984375,
0.78515625,
0.348388671875,
0.21728515625,
0.347900390625
],
[
0.2181396484375,
0.349853515625,
0.7861328125,
0.349853515625,
0.7861328125,
0.363037109375,
0.2181396484375,
0.363037109375
],
[
0.2156982421875,
0.3642578125,
0.7861328125,
0.363037109375,
0.7861328125,
0.37744140625,
0.2156982421875,
0.37890625
],
[
0.2181396484375,
0.380126953125,
0.78515625,
0.380126953125,
0.78515625,
0.39453125,
0.2181396484375,
0.39453125
],
[
0.2181396484375,
0.395751953125,
0.78369140625,
0.395751953125,
0.78369140625,
0.4091796875,
0.2181396484375,
0.4091796875
],
[
0.2181396484375,
0.410400390625,
0.78515625,
0.410400390625,
0.78515625,
0.423583984375,
0.2181396484375,
0.423583984375
],
[
0.2181396484375,
0.4248046875,
0.7861328125,
0.4248046875,
0.7861328125,
0.439453125,
0.2181396484375,
0.439453125
],
[
0.2181396484375,
0.43994140625,
0.78515625,
0.43994140625,
0.78515625,
0.455078125,
0.2181396484375,
0.455078125
],
[
0.2156982421875,
0.45263671875,
0.7861328125,
0.453857421875,
0.7861328125,
0.4716796875,
0.2156982421875,
0.47021484375
],
[
0.2181396484375,
0.47021484375,
0.7861328125,
0.47021484375,
0.7861328125,
0.48486328125,
0.2181396484375,
0.48486328125
],
[
0.2197265625,
0.4873046875,
0.78515625,
0.4873046875,
0.78515625,
0.49755859375,
0.2197265625,
0.49755859375
],
[
0.2181396484375,
0.50048828125,
0.7861328125,
0.50048828125,
0.7861328125,
0.5146484375,
0.2181396484375,
0.5146484375
],
[
0.2181396484375,
0.51513671875,
0.7861328125,
0.51513671875,
0.7861328125,
0.52978515625,
0.2181396484375,
0.52978515625
],
[
0.2181396484375,
0.53076171875,
0.7890625,
0.53076171875,
0.7890625,
0.54443359375,
0.2181396484375,
0.54443359375
],
[
0.2181396484375,
0.54541015625,
0.7861328125,
0.54541015625,
0.7861328125,
0.56005859375,
0.2181396484375,
0.56005859375
],
[
0.21728515625,
0.56103515625,
0.7861328125,
0.56103515625,
0.7861328125,
0.57568359375,
0.21728515625,
0.57568359375
],
[
0.2181396484375,
0.57763671875,
0.283447265625,
0.57763671875,
0.283447265625,
0.5888671875,
0.2181396484375,
0.5888671875
],
[
0.2181396484375,
0.61572265625,
0.25244140625,
0.61572265625,
0.25244140625,
0.62744140625,
0.2181396484375,
0.62744140625
],
[
0.259033203125,
0.6142578125,
0.4150390625,
0.6142578125,
0.4150390625,
0.62890625,
0.259033203125,
0.62890625
],
[
0.2181396484375,
0.64013671875,
0.78759765625,
0.64013671875,
0.78759765625,
0.65478515625,
0.2181396484375,
0.65478515625
],
[
0.21728515625,
0.65576171875,
0.7861328125,
0.65478515625,
0.7861328125,
0.66943359375,
0.21728515625,
0.67041015625
],
[
0.2181396484375,
0.671875,
0.78759765625,
0.671875,
0.78759765625,
0.68505859375,
0.2181396484375,
0.68505859375
],
[
0.21728515625,
0.68505859375,
0.78515625,
0.68603515625,
0.78515625,
0.70068359375,
0.21728515625,
0.69970703125
],
[
0.2181396484375,
0.70263671875,
0.78515625,
0.70263671875,
0.78515625,
0.71630859375,
0.2181396484375,
0.71630859375
],
[
0.2181396484375,
0.71728515625,
0.56884765625,
0.71728515625,
0.56884765625,
0.73095703125,
0.2181396484375,
0.73095703125
],
[
0.21728515625,
0.7392578125,
0.58251953125,
0.73779296875,
0.58251953125,
0.75244140625,
0.21728515625,
0.75390625
],
[
0.2303466796875,
0.76953125,
0.2459716796875,
0.76953125,
0.2459716796875,
0.7763671875,
0.2303466796875,
0.7763671875
],
[
0.2459716796875,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.78076171875,
0.2459716796875,
0.78076171875
],
[
0.2467041015625,
0.7822265625,
0.7861328125,
0.7822265625,
0.7861328125,
0.79541015625,
0.2467041015625,
0.79541015625
],
[
0.2459716796875,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2459716796875,
0.81005859375
],
[
0.2442626953125,
0.8125,
0.78515625,
0.8125,
0.78515625,
0.82568359375,
0.2442626953125,
0.82568359375
],
[
0.2459716796875,
0.8291015625,
0.58642578125,
0.8291015625,
0.58642578125,
0.8388671875,
0.2459716796875,
0.8388671875
]
] |
[
"2",
"N.S. D'Souza et al.",
"best to combine modalities, and how to effectively extract predictive patterns",
"from data [14].",
"1.1",
"Related Works",
"Existing attempts to fuse modalities for outcome prediction can be divided into",
"at least three approaches, namely, feature vector-based, statistical or graph-based",
"approaches. The vector-based approaches perform early, intermediate, or late",
"fusion [18,1] with the late fusion approach combining the results of prediction",
"rather than fusing the modality features. Due to the restrictive nature of the",
"underlying assumptions, these are often inadequate for characterizing the broader",
"range of relationships among modality features and their relevance for prediction",
"In statistical approaches, methods such as canonical correlation analysis [19]",
"and its deep learning variants [25] directly model feature correlations either",
"in the native representation or in a latent space [18]. However, these are not",
"guaranteed to learn discriminative patterns in the unsupervised setting and can",
"suffer from scalability issues when integrated into larger predictive models [23]",
"Recently, graph-based approaches have been developed which form basic [2,26,5]",
"or multiplexed graphs [6] from latent embeddings derived from modality features",
"using concatenation [2] or weighted averaging [26]. Task-specific fusion is then",
"achieved through inference via message passing walks between nodes in a graph",
"neural network. In the basic collapsed graph construction, the inter-patient and",
"intra-patient modality correlations arenotfullydistinguished.Conversely,inthe",
"multiplexed formulation [6], only a restricted form of multi-relational dependence",
"is captured between nodes through vertical connections. Since the graph is defined",
"using latent embedding directions, the modality semantics are not preserved.",
"Additionally, the staged training of the graph construction and inference networks",
"do not guarantee that the constructed graphs retain discriminable interaction",
"patterns.",
"1.2",
"Our Contributions",
"We develop a novel end-to-end fusion framework that addresses the limita-",
"tions mentioned above. The Maximal Correlation Multi-Layered Graph Neural",
"Network, i.e. MaxCorrMGNN, is a general yet interpretable framework for prob-",
"lems of multimodal fusion with unstructured data. Specifically, our approach",
"marries the design principles of statistical representation learning with deep",
"learning models for reasoning from multi-graphs.",
"The main contributions of this work are three-fold:",
"一",
"First, we propose to model intra and inter-patient modality relationships",
"explicitly through a novel patient-modality multi-layered graph as shown",
"in Fig. 1. The edges in each layer (plane of the multi-graph) capture the",
"intra-modality relations between patients, while the cross-edges between layers",
"captureznter-modality relationsacrosspatients."
] |
{
"arxiv_id": "2307.07093",
"language": "en",
"page_idx": 1
}
|
||
2307.07093-en-2
|
2307.07093-en-2.jpg
|
Title Suppressed Due to Excessive Length
3
<FigureHere>
Fig.1. Our Generalized Framework for Multimodal Fusion Gray Box: The features from different modalities are input to the MaxCorr (HGR) formulation. The nodes are the patients of the Multi-Graph, and the planes are the modalities. Purple Box: The Multi-Graph Neural Network maps the multi-graph representation to the targets.
- Since these relationships are not known apriori for unstructured data, we
propose, for the first time, to use learnable Hirschfeld-Gebelein-Reñyi (HGR) maximal correlations. We introduce learnable soft-thresholding to uncover salient connectivity patterns automatically. Effectively, this procedure allows us to express any multimodal dataset as a patient-modality multilayered graph for fusion.
-Third, we develop a multilayered graph neural network (MGNN) from frst
principles for task-informed reasoning from multi-layered graphs
To demonstrate the generality of our approach, we evaluate our framework on a large Tuberculosis (TB) dataset for multi-outcome prediction. Through rigorous experimentation, we show our framework outperforms several state-of-the-art graph based and traditional fusion baselines.
2 MaxCorrMGNN formulation for Multimodal Fusion
We now describe the four main aspects of our formulation, namely, (a) multilayered graph representation, (b) formalism for maximal correlation (c) task-specific inference through graph neural networks, and (d) loss function for end-to-end learning of both graph connectivity and inference.
2.1 Patient-Modality Multi-layered Graph:
Given multimodal data about patients, we model the modality and patient information through a multi-layered graph [3] as shown in Fig.1.Herethe nodes are grouped into multiple planes, each plane representing edge-connectivity $according\textit{ to an individual modality while each patient is represented by a set off is presented by a set off is pres et off is pres et of is pres e of is pre s e of is pre s e of is p it s pre t is p i s p i s e p i s e of i s p i s o p i s o p i s o }$ $corresponding\textit{ nodes across the layers ( called a supra- node) . }$
|
[
[
0.450927734375,
0.11553955078125,
0.73193359375,
0.11553955078125,
0.73193359375,
0.1287841796875,
0.450927734375,
0.1287841796875
],
[
0.775390625,
0.11553955078125,
0.7890625,
0.11871337890625,
0.78369140625,
0.1287841796875,
0.771484375,
0.1256103515625
],
[
0.2646484375,
0.147705078125,
0.369384765625,
0.147705078125,
0.369384765625,
0.1514892578125,
0.2646484375,
0.1514892578125
],
[
0.54150390625,
0.1483154296875,
0.5703125,
0.1483154296875,
0.5703125,
0.1514892578125,
0.54150390625,
0.1514892578125
],
[
0.59619140625,
0.146484375,
0.67333984375,
0.146484375,
0.67333984375,
0.15087890625,
0.59619140625,
0.15087890625
],
[
0.29736328125,
0.1590576171875,
0.325927734375,
0.1590576171875,
0.325927734375,
0.1673583984375,
0.29736328125,
0.1673583984375
],
[
0.29736328125,
0.1749267578125,
0.318603515625,
0.1749267578125,
0.318603515625,
0.1787109375,
0.29736328125,
0.1787109375
],
[
0.48193359375,
0.169189453125,
0.515625,
0.169189453125,
0.515625,
0.17236328125,
0.48193359375,
0.17236328125
],
[
0.484375,
0.1817626953125,
0.5107421875,
0.1817626953125,
0.5107421875,
0.186279296875,
0.484375,
0.186279296875
],
[
0.40283203125,
0.2310791015625,
0.4306640625,
0.2310791015625,
0.4306640625,
0.2392578125,
0.40283203125,
0.2392578125
],
[
0.48779296875,
0.240478515625,
0.50634765625,
0.240478515625,
0.50634765625,
0.248779296875,
0.48779296875,
0.248779296875
],
[
0.295654296875,
0.255126953125,
0.330078125,
0.255126953125,
0.330078125,
0.266357421875,
0.295654296875,
0.266357421875
],
[
0.294189453125,
0.267578125,
0.33251953125,
0.267578125,
0.33251953125,
0.2841796875,
0.294189453125,
0.2841796875
],
[
0.221435546875,
0.304931640625,
0.78369140625,
0.304931640625,
0.78369140625,
0.31494140625,
0.221435546875,
0.31494140625
],
[
0.2181396484375,
0.317626953125,
0.78759765625,
0.317626953125,
0.78759765625,
0.330810546875,
0.2181396484375,
0.330810546875
],
[
0.2197265625,
0.333251953125,
0.78515625,
0.333251953125,
0.78515625,
0.343505859375,
0.2197265625,
0.343505859375
],
[
0.2181396484375,
0.34521484375,
0.7587890625,
0.34521484375,
0.7587890625,
0.359130859375,
0.2181396484375,
0.359130859375
],
[
0.233642578125,
0.412353515625,
0.78369140625,
0.412353515625,
0.78369140625,
0.425537109375,
0.233642578125,
0.425537109375
],
[
0.2442626953125,
0.427978515625,
0.7861328125,
0.4267578125,
0.7861328125,
0.44140625,
0.2442626953125,
0.442626953125
],
[
0.2442626953125,
0.44140625,
0.78515625,
0.442626953125,
0.78515625,
0.45703125,
0.2442626953125,
0.455810546875
],
[
0.2459716796875,
0.458251953125,
0.78515625,
0.458251953125,
0.78515625,
0.4716796875,
0.2459716796875,
0.4716796875
],
[
0.2459716796875,
0.472900390625,
0.7861328125,
0.472900390625,
0.7861328125,
0.486083984375,
0.2459716796875,
0.486083984375
],
[
0.2459716796875,
0.488037109375,
0.36767578125,
0.488037109375,
0.36767578125,
0.501953125,
0.2459716796875,
0.501953125
],
[
0.23779296875,
0.5068359375,
0.2442626953125,
0.5068359375,
0.2442626953125,
0.51220703125,
0.23779296875,
0.51220703125
],
[
0.240234375,
0.50244140625,
0.78369140625,
0.50244140625,
0.78369140625,
0.5166015625,
0.240234375,
0.5166015625
],
[
0.2467041015625,
0.51806640625,
0.7099609375,
0.51806640625,
0.7099609375,
0.5322265625,
0.2467041015625,
0.5322265625
],
[
0.2197265625,
0.54248046875,
0.78515625,
0.54248046875,
0.78515625,
0.556640625,
0.2197265625,
0.556640625
],
[
0.21728515625,
0.556640625,
0.78369140625,
0.556640625,
0.78369140625,
0.57080078125,
0.21728515625,
0.57080078125
],
[
0.2156982421875,
0.57373046875,
0.78515625,
0.57275390625,
0.78515625,
0.5859375,
0.2156982421875,
0.5869140625
],
[
0.2156982421875,
0.58837890625,
0.5400390625,
0.5869140625,
0.5400390625,
0.6015625,
0.2156982421875,
0.60302734375
],
[
0.21728515625,
0.62548828125,
0.7548828125,
0.62548828125,
0.7548828125,
0.64013671875,
0.21728515625,
0.64013671875
],
[
0.2181396484375,
0.65771484375,
0.78515625,
0.65771484375,
0.78515625,
0.671875,
0.2181396484375,
0.671875
],
[
0.2181396484375,
0.673828125,
0.78515625,
0.673828125,
0.78515625,
0.6875,
0.2181396484375,
0.6875
],
[
0.2181396484375,
0.68798828125,
0.7861328125,
0.68798828125,
0.7861328125,
0.70263671875,
0.2181396484375,
0.70263671875
],
[
0.2181396484375,
0.7041015625,
0.580078125,
0.7041015625,
0.580078125,
0.71728515625,
0.2181396484375,
0.71728515625
],
[
0.21728515625,
0.7392578125,
0.583984375,
0.74072265625,
0.583984375,
0.7548828125,
0.21728515625,
0.75390625
],
[
0.2181396484375,
0.76513671875,
0.7861328125,
0.76513671875,
0.7861328125,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78759765625,
0.78076171875,
0.78759765625,
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,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.646484375,
0.8271484375,
0.646484375,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"Title Suppressed Due to Excessive Length",
"3",
"aorr:LatentMuti-crapl",
"ant",
"alNathunrlforOi",
"X",
"",
"MGNNLAMGE",
"GINcOn",
"",
"C.A",
"f",
"①x\"",
"Fig.1.OurGeneralizedFrameworkforMultimodalFusionGrayBox:Thefeatures",
"from different modalities are input to the MaxCorr (HGR) formulation. The nodes are",
"the patients of the Multi-Graph, and the planes are the modalities. Purple Box: The",
"Multi-Graph Neural Network maps the multi-graph representation to the targets.",
"- Since these relationships are not known apriori for unstructured data, we",
"propose, for the first time, to use learnable Hirschfeld-Gebelein-Renyi (HGR)",
"maximal correlations. We introduce learnable soft-thresholding to uncover",
"salient connectivity patterns automatically. Effectively, this procedure allows",
"us to express any multimodal dataset as a patient-modality multilayered",
"graph for fusion.",
"",
"Third, we develop a multilayered graph neural network (MGNN) from first",
"principles for task-informed reasoning from multi-layered graphs.",
"To demonstrate the generality of our approach, we evaluate our framework on a",
"large Tuberculosis (TB) dataset for multi-outcome prediction.Through rigorous",
"experimentation, we show our framework outperforms several state-of-the-art",
"graph based and traditional fusion baselines.",
"2 MaxCorrMGNN formulation for Multimodal Fusion",
"We now describe the four main aspects of our formulation, namely, (a)multilayered",
"graph representation, (b) formalism for maximal correlation (c) task-specific",
"inference through graph neural networks, and (d) loss function for end-to-end",
"learning of both graph connectivity and inference.",
"2.1 Patient-Modality Multi-layered Graph:",
"Given multimodal data about patients, we model the modality and patient",
"information through a multi-layered graph [3] as shown in Fig. 1. Here the",
"nodes are grouped into multiple planes, each plane representing edge-connectivity",
"according to an individual modality while each patient is represented by a set of",
"corresponding nodes across the layers (called a supra-node)."
] |
{
"arxiv_id": "2307.07093",
"language": "en",
"page_idx": 2
}
|
||
2307.07093-en-3
|
2307.07093-en-3.jpg
|
4
N.S. D'Souza et al.
Mathematically, we represent the multi-layered graph as: $\mathcal{G} _{\mathrm{M} }= ( \mathcal{V} , \mathcal{E} _{\mathrm{M} }) $, where $| \mathcal{V} _{\mathrm{M} }| = | \mathcal{V} | \times K$ are the extended supra-nodes and $\mathcal{E}_{\mathrm{M}}=\{(i,j)\in\mathcal{V}_{\mathrm{M}}\times\mathcal{V}_{\mathrm{M}}\}$ are the edges between supra-nodes. There are $K$ modality planes, each with $\begin{array}{c}{\mathrm{adjacency~matrices~}\mathbf{A}_{(k)}\in\mathcal{R}^{P\times P}.\mathrm{~The~}K\times K\mathrm{~pairwise~cross~planar~connections}}\\{\mathrm{adjacency~matrices~}\mathbf{A}_{(k)}\in\mathcal{R}^{P\times P}.\mathrm{~The~}K\times K\mathrm{~pairwise~cross~planar~connections}}\end{array}$ are given by $\mathbf{C}_{(l,m)}\in\mathcal{R}^{P\times P}$, where $P=|\mathcal{V}|.$ All edge weights may take values $\operatorname{in}$ range$[0-1]^{'}$
2.2 HGR Maximal Correlations for Latent Multi-Graph Learning:
Recall that we would like to learn task informed patient-modality multi-graph representations automatically from unstructured modality data. To this end, we develop the framework illustrated in the Gray Box in Fig. 1.
$\operatorname*{Let}\{\mathbf{x}_{n}^{k}\in\mathcal{R}^{D_{k}\times1}\}_{k=1}^{K}$be the features from to modality $k$ for patient $n.$Since features from different modalities lie in different input subspaces, we develop parallel common space projections to explore the dependence between them. The Hirschfield, Gebelin, Rènyi (HGR) [23] framework in statistics is known to generalize the notion of dependence to abstract and non-linear functional spaces. Such non-linear projections can be parameterized by deep neural networks
Specifically, let the collection of modality-specific projection networks be given by $\{ f^{k}$( $\cdot$) : $\mathcal{R} ^{D_{k}\times 1}\to \mathcal{R} ^{D_{p}\times 1}\}$. The HGR maximal correlation is a sym- metric measure obtained by solving the following coupled pairwise constrained optimization problem:
(1)
$$\sup\limits_{\mathcal{C}_{E},\mathcal{C}_{\mathrm{Cov}}}\rho_{\mathrm{HGR}}(\mathbf{x}^{l},\mathbf{x}^{m})=\sup\limits_{\mathcal{C}_{E},\mathcal{C}_{\mathrm{Cov}}}\mathbb{E}\Big[[\mathbf{f}^{l}(\mathbf{x}^{l})]^{T}\mathbf{f}^{m}(\mathbf{x}^{m})\Big]$$
$\forall\{l.m\}$ s.t. $l\neq m$, where $\mathcal{I}_{D_{p}}$ is a $D_p\times D_p$ identity matrix. The constraint sets
are given by:
$$\begin{array}{c}\mathcal{C}_E:\{\mathbb{E}[\mathbf{f}^l(\cdot)]=\mathbb{E}[\mathbf{f}^m(\cdot)]=\mathbf{0}\}\\\mathcal{C}_{\mathrm{Cov}}:\{\mathbf{Cov}(\mathbf{f}^l(\mathbf{x}^l))=\mathbf{Cov}(\mathbf{f}^m(\mathbf{x}^m))=\mathcal{I}_{D_p}\}\end{array}$$
(2) (3)
Approaches such as deep CCA [25] can be thought of as a special case of this formulation which solve the whitening (empricial covariance) constraints (Eq. (3)) via explicit pairwise de-correlation.
However, for multiple modalities in large datasets, exact whitening is not scalable. To circumvent this issue, we can use the approach in [23]. This formulation proposes introduces a relaxation to the exact HGR, named soft-HGR, which consists of a trace regularizer in lieu of whitening. Eq. (1) can be relaxed as an empirical minimization problem min $\mathcal{L}_\mathrm{sHGR}$,where the sHGR loss is:
$$\mathcal{L}_{\mathrm{sHGR}}=-\frac{1}{N_{z}}\mathbb{E}\left[\mathbf{f}^{l}(\mathbf{x}^{l})^{T}\mathbf{f}^{m}(\mathbf{x}^{m})\right]+\frac{1}{2N_{z}}\operatorname{Tr}\left[\mathbf{Cov}[\mathbf{f}^{l}(\mathbf{x}^{l})]\mathbf{Cov}[\mathbf{f}^{m}(\mathbf{x}^{m})]\right]\quad(4)\\=-\frac{1}{N_{z}}\left(\sum_{l,m=1}^{K}\left[\sum_{n=1}^{N}\frac{\mathbf{f}^{l}(\mathbf{x}_{n}^{l})^{T}\mathbf{f}^{m}(\mathbf{x}_{n}^{m})}{(N-1)}-\frac{\mathrm{Tr}\left[\mathbf{Cov}[\mathbf{f}^{l}(\mathbf{x}^{l})]\mathbf{Cov}[\mathbf{f}^{m}(\mathbf{x}^{m})]\right]}{2}\right]\right)\:l\neq m$$
|
[
[
0.2181396484375,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.126953125,
0.2181396484375,
0.126953125
],
[
0.271240234375,
0.11553955078125,
0.398681640625,
0.11553955078125,
0.398681640625,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.24267578125,
0.146484375,
0.78173828125,
0.146484375,
0.78173828125,
0.1610107421875,
0.24267578125,
0.1610107421875
],
[
0.221435546875,
0.1634521484375,
0.78369140625,
0.1634521484375,
0.78369140625,
0.177978515625,
0.221435546875,
0.177978515625
],
[
0.2197265625,
0.177978515625,
0.78369140625,
0.177978515625,
0.78369140625,
0.1912841796875,
0.2197265625,
0.1912841796875
],
[
0.21728515625,
0.1900634765625,
0.78515625,
0.1900634765625,
0.78515625,
0.2083740234375,
0.21728515625,
0.2083740234375
],
[
0.214111328125,
0.2052001953125,
0.78759765625,
0.20263671875,
0.78759765625,
0.2235107421875,
0.214111328125,
0.2259521484375
],
[
0.21728515625,
0.2235107421875,
0.334228515625,
0.2235107421875,
0.334228515625,
0.2374267578125,
0.21728515625,
0.2374267578125
],
[
0.2156982421875,
0.2568359375,
0.77392578125,
0.2587890625,
0.77392578125,
0.276611328125,
0.2156982421875,
0.274658203125
],
[
0.21728515625,
0.285400390625,
0.78369140625,
0.285888671875,
0.78369140625,
0.300537109375,
0.21728515625,
0.2998046875
],
[
0.2197265625,
0.3017578125,
0.7861328125,
0.3017578125,
0.7861328125,
0.31494140625,
0.2197265625,
0.31494140625
],
[
0.2181396484375,
0.31640625,
0.65185546875,
0.31640625,
0.65185546875,
0.330078125,
0.2181396484375,
0.330078125
],
[
0.2418212890625,
0.328857421875,
0.7861328125,
0.328857421875,
0.7861328125,
0.3466796875,
0.2418212890625,
0.3466796875
],
[
0.2181396484375,
0.3466796875,
0.78515625,
0.3466796875,
0.78515625,
0.35986328125,
0.2181396484375,
0.35986328125
],
[
0.2197265625,
0.3623046875,
0.7861328125,
0.3623046875,
0.7861328125,
0.375732421875,
0.2197265625,
0.375732421875
],
[
0.2197265625,
0.376953125,
0.7861328125,
0.376953125,
0.7861328125,
0.39013671875,
0.2197265625,
0.39013671875
],
[
0.2197265625,
0.392578125,
0.78759765625,
0.392578125,
0.78759765625,
0.406005859375,
0.2197265625,
0.406005859375
],
[
0.2181396484375,
0.4072265625,
0.7548828125,
0.4072265625,
0.7548828125,
0.42041015625,
0.2181396484375,
0.42041015625
],
[
0.24267578125,
0.421630859375,
0.78515625,
0.421630859375,
0.78515625,
0.436279296875,
0.24267578125,
0.436279296875
],
[
0.214111328125,
0.43310546875,
0.7890625,
0.431884765625,
0.7890625,
0.45263671875,
0.214111328125,
0.453857421875
],
[
0.2181396484375,
0.45068359375,
0.78759765625,
0.451904296875,
0.78759765625,
0.466552734375,
0.2181396484375,
0.46533203125
],
[
0.2156982421875,
0.46728515625,
0.37744140625,
0.46728515625,
0.37744140625,
0.48095703125,
0.2156982421875,
0.48095703125
],
[
0.33251953125,
0.49560546875,
0.67724609375,
0.494384765625,
0.67724609375,
0.51318359375,
0.33251953125,
0.5146484375
],
[
0.76123046875,
0.496826171875,
0.78759765625,
0.496826171875,
0.78759765625,
0.5146484375,
0.76123046875,
0.5146484375
],
[
0.325927734375,
0.51025390625,
0.366943359375,
0.51025390625,
0.366943359375,
0.521484375,
0.325927734375,
0.521484375
],
[
0.491943359375,
0.51220703125,
0.515625,
0.51220703125,
0.515625,
0.51953125,
0.491943359375,
0.51953125
],
[
0.2181396484375,
0.52783203125,
0.78759765625,
0.52783203125,
0.78759765625,
0.54541015625,
0.2181396484375,
0.54541015625
],
[
0.2197265625,
0.546875,
0.31201171875,
0.546875,
0.31201171875,
0.556640625,
0.2197265625,
0.556640625
],
[
0.4248046875,
0.5693359375,
0.6337890625,
0.5693359375,
0.6337890625,
0.5869140625,
0.4248046875,
0.5869140625
],
[
0.76123046875,
0.57080078125,
0.78759765625,
0.57080078125,
0.78759765625,
0.58837890625,
0.76123046875,
0.58837890625
],
[
0.336669921875,
0.58837890625,
0.65625,
0.5888671875,
0.65625,
0.60693359375,
0.336669921875,
0.60595703125
],
[
0.76123046875,
0.5888671875,
0.7890625,
0.5888671875,
0.7890625,
0.60693359375,
0.76123046875,
0.60693359375
],
[
0.2197265625,
0.61865234375,
0.78515625,
0.61865234375,
0.78515625,
0.6318359375,
0.2197265625,
0.6318359375
],
[
0.21728515625,
0.63330078125,
0.78515625,
0.63330078125,
0.78515625,
0.646484375,
0.21728515625,
0.646484375
],
[
0.2181396484375,
0.6494140625,
0.465576171875,
0.6494140625,
0.465576171875,
0.66357421875,
0.2181396484375,
0.66357421875
],
[
0.24267578125,
0.66357421875,
0.7861328125,
0.66357421875,
0.7861328125,
0.6767578125,
0.24267578125,
0.6767578125
],
[
0.2181396484375,
0.6787109375,
0.78515625,
0.6787109375,
0.78515625,
0.6923828125,
0.2181396484375,
0.6923828125
],
[
0.2181396484375,
0.6943359375,
0.78515625,
0.6943359375,
0.78515625,
0.70849609375,
0.2181396484375,
0.70849609375
],
[
0.2181396484375,
0.708984375,
0.78369140625,
0.708984375,
0.78369140625,
0.72265625,
0.2181396484375,
0.72265625
],
[
0.2181396484375,
0.724609375,
0.7138671875,
0.724609375,
0.7138671875,
0.7392578125,
0.2181396484375,
0.7392578125
],
[
0.34716796875,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.7841796875,
0.34716796875,
0.7841796875
],
[
0.51318359375,
0.76513671875,
0.74609375,
0.7666015625,
0.74609375,
0.78515625,
0.51318359375,
0.7841796875
],
[
0.48046875,
0.7734375,
0.491943359375,
0.7734375,
0.491943359375,
0.78271484375,
0.48046875,
0.78271484375
],
[
0.316162109375,
0.77783203125,
0.34716796875,
0.77783203125,
0.34716796875,
0.7890625,
0.316162109375,
0.7890625
],
[
0.490234375,
0.77783203125,
0.525390625,
0.77783203125,
0.525390625,
0.7890625,
0.490234375,
0.7890625
],
[
0.35302734375,
0.79736328125,
0.7138671875,
0.79248046875,
0.7138671875,
0.81640625,
0.35302734375,
0.8212890625
],
[
0.75732421875,
0.81494140625,
0.79150390625,
0.81494140625,
0.79150390625,
0.82470703125,
0.75732421875,
0.82470703125
],
[
0.39453125,
0.8193359375,
0.453369140625,
0.8193359375,
0.453369140625,
0.83544921875,
0.39453125,
0.83544921875
],
[
0.238525390625,
0.76806640625,
0.31201171875,
0.77099609375,
0.31201171875,
0.78515625,
0.238525390625,
0.78271484375
],
[
0.60693359375,
0.82080078125,
0.619140625,
0.82470703125,
0.61279296875,
0.8359375,
0.6005859375,
0.83203125
]
] |
[
"4",
"N.S. D'Souza et al.",
"Mathematically,we represent the multi-layered graph as: GM = (V,M), where",
"[VM| = [VI × K are the extended supra-nodes and EM = [(i,j) E VM × VM)",
"are the edges between supra-nodes. There are K modality planes, each with",
"adjacency matrices A(k) E RPxP. The K × K pairwise cross planar connections",
"are given by C(l,m) e RPxP, where P= [Vl. All edge weights may take values",
"in range [0 - 1]",
"2.2 HGR Maximal Correlations for Latent Multi-Graph Learning:",
"Recall that we would like to learn task informed patient-modality multi-graph",
"representations automatically from unstructured modality data. To this end, we",
"develop the framework illustrated in the Gray Box in Fig. 1.",
"Let [xh e RDxx1)k- be the features from to modality k for patient n. Since",
"features from different modalities lie in different input subspaces, we develop",
"parallel common space projections to explore the dependence between them",
"The Hirschfield, Gebelin, Renyi (HGR) [23]l framework in statistics is known to",
"generalize the notion of dependence to abstract and non-linear functional spaces",
"Such non-linear projections can be parameterized by deep neural networks.",
"Specifically, let the collection of modality-specific projection networks be",
"given by (fh() : RDx×1 → RDpx1). The HGR maximal correlation is a sym-",
"metric measure obtained by solving the following coupled pairwise constrained",
"optimization problem:",
"sup PHGR(x*,x\")=,sup E['(x)]Tfm(x\")",
"(1)",
"CE,CCo",
"CF.",
"V[l.m) s.t. I + m, where ID, is a Dp × D, identity matrix. The constraint sets",
"are given by:",
"Ce : [E[f'()] = E[f\"m()] = 0]",
"(2)",
"Ccov : [Cov(f'(x\")) = Cov(f\"m(xm)) = LD,)",
"(3)",
"Approaches such as deep CCA [25] can be thought of as a special case of this",
"formulation which solve the whitening (empricial covariance) constraints (Eq. (3)",
"via explicit pairwise de-correlation",
"However, for multiple modalities in large datasets, exact whitening is not",
"scalable. To circumvent this issue, we can use the approach in [23]. This formula.",
"tion proposes introduces a relaxation to the exact HGR, named soft-HGR, which",
"consists of a trace regularizer in lieu of whitening. Eq. (1) can be relaxed as an",
"empirical minimization problem min CsHGR, where the sHGR loss is:",
"Ef'(x*)Trm(xm)| +",
"Tr [Cov[f'(x)]Cov[f\" (xm)]",
"十",
"N.",
"2N.",
" f(x) fm(x) Tr [Cov[f(x)]Cov[r\"(x\")]",
"丰m",
"(N- 1)",
"CsHGR=",
"2"
] |
{
"arxiv_id": "2307.07093",
"language": "en",
"page_idx": 3
}
|
||
2307.07093-en-4
|
2307.07093-en-4.jpg
|
Title Suppressed Due to Excessive Length 5
The expectation under the functional transformations $\mathbb{E}[\mathbf{f}^l(\mathbf{x}^l)]=\mathbb{E}[\mathbf{f}^m(\mathbf{x}^m)]=\mathbf{0}$ is enforced step-wise by mean subtraction during optimization. Here, Cov(·) is the empirical covariance matrix. We parameterize $\{\bar{\mathbf{f}}^k(\cdot)\}$ as a simple two layered fully connected neural network with a normalization factor as $N_z=M(M-1).$
By design, the MaxCorr formulation allows us to utilize the correlation $\rho_{\mathrm{sHGR}}(\mathbf{x}_{i}^{l},\mathbf{x}_{j}^{m})$ (computed after solving Eq. (4)) to model dependence between patients $i$ and $j$ according to the $l$ and $m$ modality features in a general setting The absolute value of this correlation measure define the edge weights between nodes in the patient-modality multi-graph. As opposed to existing
$Learnable\textit{ Adaptive Sparsity: Additionally, we would like to have our learning}$ framework automatically discover and retain salient edges that are relevant for prediction. To encourage sparsity in the edges, we utilize a learnable softthresholding formulation. We first define a symmetric block sparsity matrix S. Since edge weights in the multi-graph are in the range [0-1], we normalize it through the sigmoid function as $\tilde{\mathbf{S}}=\tilde{\mathbf{S}}^{T}=$Sigmoid$(\mathbf{S})\in\mathcal{R}^{\dot{K}\times K}.$ The entries of the soft-thresholding matrix $\tilde{\mathbf{S}}[l,m]$ define learnable thresholds for the cross modal connections when $l\neq m$ and in-plane connections when $l=m.$ Finally, the cross modal edges and in-plane edges of the multi-graph are given by
(5)
$$\mathbf{C}_{(l,m)}[i,j]=\mathrm{ReLU}(\tilde{\rho}_{\mathrm{sHGR}}(\mathbf{x}_{i}^{l},\mathbf{x}_{j}^{m})-\tilde{\mathbf{S}}[l,m])\\\mathbf{A}_{k}[i,j]=\mathrm{ReLU}(\tilde{\rho}_{\mathrm{sHGR}}(\mathbf{x}_{i}^{k},\mathbf{x}_{j}^{k})-\tilde{\mathbf{S}}[k,k])$$
(6)
with $\tilde{\rho}_{\mathrm{sHGR}}=|\rho_{\mathrm{sHGR}}|$ respectively. The adjacency matrices $\mathbf{A}_{(k)}$ model the dependence within the features of modality $k$,while the cross planar matrices $\{\mathbf{C}_{(l,m)}\}$ capture interactions across modalities. Overall, S acts as a regularizer that suppresses noisy weak dependencies. These regularization parameters are automatically inferred during training along with the MaxCorr projection parameters $\{\mathbf{f}^k(\cdot)\}.$ This effectively adds just $K(K+1)$ learnable parameters to the MaxCorrMGNN.
# 2.3 Multi-Graph Neural Networka
As a standalone optimization, the MaxCorr block is not guaranteed to learn discriminative projections of modality features. A natural next step is to couple the multi-graph representation learning with the classification task. Graph Neural Networks have recently become popular tools for reasoning from graphs and graph-signals. Given the patient-modality multi-graph, we design an extension of traditional graph neural networks to multi-graphs for inference tasks
Conventional GNNs flter information based on the graph topology (i.e. the adjacency matrix) to map the node features to the targets based on graph traversals. Conceptually, cascading $l$ GNN layers is analogous to filtered information pooling at each node from its $l$-hop neighbors [13] inferred from the powers of the graph adjacency matrix. These neighborhoods can be reached by seeding walks on the graph starting at the desired node. Inspired by this design, we craft a multi-graph neural network (Purple Box in Fig. 1) for outcome prediction. Our
|
[
[
0.451904296875,
0.11553955078125,
0.73388671875,
0.11553955078125,
0.73388671875,
0.1287841796875,
0.451904296875,
0.1287841796875
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.2181396484375,
0.147705078125,
0.7861328125,
0.147705078125,
0.7861328125,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.21728515625,
0.1767578125,
0.78515625,
0.177978515625,
0.78515625,
0.1925048828125,
0.21728515625,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.24267578125,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.24267578125,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.7861328125,
0.222900390625,
0.7861328125,
0.2374267578125,
0.2181396484375,
0.238037109375
],
[
0.2197265625,
0.238037109375,
0.78759765625,
0.238037109375,
0.78759765625,
0.25244140625,
0.2197265625,
0.25244140625
],
[
0.2197265625,
0.25390625,
0.7861328125,
0.25390625,
0.7861328125,
0.267578125,
0.2197265625,
0.267578125
],
[
0.2156982421875,
0.268310546875,
0.69287109375,
0.268310546875,
0.69287109375,
0.2822265625,
0.2156982421875,
0.2822265625
],
[
0.2181396484375,
0.296630859375,
0.7861328125,
0.29736328125,
0.7861328125,
0.3125,
0.2181396484375,
0.311279296875
],
[
0.2181396484375,
0.313232421875,
0.7861328125,
0.313232421875,
0.7861328125,
0.326904296875,
0.2181396484375,
0.326904296875
],
[
0.2181396484375,
0.328857421875,
0.7861328125,
0.328857421875,
0.7861328125,
0.34228515625,
0.2181396484375,
0.34228515625
],
[
0.2181396484375,
0.343505859375,
0.78759765625,
0.343505859375,
0.78759765625,
0.356689453125,
0.2181396484375,
0.356689453125
],
[
0.2181396484375,
0.359130859375,
0.7861328125,
0.359130859375,
0.7861328125,
0.37255859375,
0.2181396484375,
0.37255859375
],
[
0.21728515625,
0.37060546875,
0.7861328125,
0.37060546875,
0.7861328125,
0.38818359375,
0.21728515625,
0.38818359375
],
[
0.2181396484375,
0.38818359375,
0.7861328125,
0.38818359375,
0.7861328125,
0.40283203125,
0.2181396484375,
0.40283203125
],
[
0.21728515625,
0.40283203125,
0.78759765625,
0.404052734375,
0.78759765625,
0.41845703125,
0.21728515625,
0.417236328125
],
[
0.21728515625,
0.419189453125,
0.74169921875,
0.419189453125,
0.74169921875,
0.43310546875,
0.21728515625,
0.43310546875
],
[
0.338134765625,
0.443115234375,
0.65673828125,
0.443115234375,
0.65673828125,
0.4609375,
0.338134765625,
0.4609375
],
[
0.76123046875,
0.443115234375,
0.7890625,
0.443115234375,
0.7890625,
0.462158203125,
0.76123046875,
0.462158203125
],
[
0.363525390625,
0.46533203125,
0.65673828125,
0.46533203125,
0.65673828125,
0.48291015625,
0.363525390625,
0.48291015625
],
[
0.76123046875,
0.466552734375,
0.7890625,
0.466552734375,
0.7890625,
0.48291015625,
0.76123046875,
0.48291015625
],
[
0.2181396484375,
0.49365234375,
0.7861328125,
0.49365234375,
0.7861328125,
0.50830078125,
0.2181396484375,
0.50830078125
],
[
0.2181396484375,
0.5087890625,
0.7861328125,
0.5087890625,
0.7861328125,
0.52294921875,
0.2181396484375,
0.52294921875
],
[
0.2197265625,
0.5244140625,
0.78515625,
0.5244140625,
0.78515625,
0.5390625,
0.2197265625,
0.5390625
],
[
0.2181396484375,
0.5390625,
0.7861328125,
0.5390625,
0.7861328125,
0.55322265625,
0.2181396484375,
0.55322265625
],
[
0.21728515625,
0.5537109375,
0.78759765625,
0.5537109375,
0.78759765625,
0.5673828125,
0.21728515625,
0.5673828125
],
[
0.21728515625,
0.5673828125,
0.7861328125,
0.568359375,
0.7861328125,
0.58349609375,
0.21728515625,
0.58203125
],
[
0.21728515625,
0.583984375,
0.370849609375,
0.583984375,
0.370849609375,
0.59765625,
0.21728515625,
0.59765625
],
[
0.21728515625,
0.619140625,
0.5107421875,
0.61865234375,
0.5107421875,
0.63330078125,
0.21728515625,
0.6337890625
],
[
0.2181396484375,
0.64306640625,
0.78515625,
0.64306640625,
0.78515625,
0.65771484375,
0.2181396484375,
0.65771484375
],
[
0.2197265625,
0.662109375,
0.78369140625,
0.662109375,
0.78369140625,
0.67236328125,
0.2197265625,
0.67236328125
],
[
0.21728515625,
0.67529296875,
0.78759765625,
0.67529296875,
0.78759765625,
0.689453125,
0.21728515625,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.21728515625,
0.70556640625,
0.78759765625,
0.705078125,
0.78759765625,
0.7197265625,
0.21728515625,
0.72021484375
],
[
0.2181396484375,
0.72021484375,
0.7158203125,
0.72021484375,
0.7158203125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2442626953125,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2442626953125,
0.75
],
[
0.21728515625,
0.75,
0.78515625,
0.75048828125,
0.78515625,
0.76513671875,
0.21728515625,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78759765625,
0.78076171875,
0.78759765625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Title Suppressed Due to Excessive Length",
"5",
"The expectation under the functional transformations E[f'(x')] = E[f\"(x\")] = 0",
"is enforced step-wise by mean subtraction during optimization. Here, Cov() is",
"the empirical covariance matrix. We parameterize {fk()) as a simple two layered",
"fully connected neural network with a normalization factor as N, = M(M - 1),",
"By design, the MaxCorr formulation allows us to utilize the correlation",
"PsHGR(x',xm) (computed after solving Eq. (4)) to model dependence between",
"patients i and j according to the l and m modality features in a general setting.",
"The absolute value of this correlation measure define the edge weights between",
"nodes in the patient-modality multi-graph. As opposed to existing",
"Learnable Adaptive Sparsity:Additionally, we would like to have our learning",
"framework automatically discover and retain salient edges that are relevant",
"for prediction. To encourage sparsity in the edges, we utilize a learnable soft",
"thresholding formulation. We first define a symmetric block sparsity matrix S.",
"Since edge weights in the multi-graph are in the range [O-1], we normalize it",
"through the sigmoid function as S = ST = Sigmoid(S) e RKxK, The entries",
"of the soft-thresholding matrix S[l, m] define learnable thresholds for the cross",
"modal connections when I + m and in-plane connections when I = m.Finally,",
"the cross modal edges and in-plane edges of the multi-graph are given by",
"C(,m)[,j] = ReLU(psHGR(x*, xm) - S[1, ml)",
"(5)",
"A[i,j] = ReLU(psHGR(x,x,) - S[k, ])",
"(6)",
"with psHGR = IpsHGRl respectively. The adjacency matrices A(k) model the",
"dependence within the features of modality k, while the cross planar matrices",
"[C(l,m)J capture interactions across modalities. Overall, S acts as a regularizer",
"that suppresses noisy weak dependencies. These regularization parameters are",
"automatically inferred during training along with the MaxCorr projection pa-",
"rameters [fk(). This effectively adds just K(K +1) learnable parameters to",
"the MaxCorrMGNN.",
"2.3Multi-Graph Neural Network:",
"As a standalone optimization, the MaxCorr block is not guaranteed to learn",
"discriminative projections of modality features.A natural next step is to couple",
"the multi-graph representation learning with the classification task. Graph Neural",
"Networks have recently become popular tools for reasoning from graphs and",
"graph-signals. Given the patient-modality multi-graph, we design an extension of",
"traditional graph neural networks to multi-graphs for inference tasks.",
"Conventional GNNs filter information based on the graph topology (i.e. the",
"adjacency matrix) to map the node features to the targets based on graph traver",
"sals. Conceptually, cascading I GNN layers is analogous to filtered information",
"pooling at each node from its l-hop neighbors [13] inferred from the powers of",
"the graph adjacency matrix. These neighborhoods can be reached by seeding",
"walks on the graph starting at the desired node. Inspired by this design, we craft",
"a multi-graph neural network (Purple Box in Fig. 1) for outcome prediction. Our"
] |
{
"arxiv_id": "2307.07093",
"language": "en",
"page_idx": 4
}
|
||
2305.04268-en-1
|
2305.04268-en-1.jpg
|
a general enhancement to the NeRF-based backbone, equipping most NeRF-like methods with the ability to model complex reflection and refraction, as shown in Fig. 1.
Existing datasets have not paid enough attention to the 360-degree rendering of scenes containing mirror-like objects, such as RFFR[12] just has forward-facing scenes, and the Shiny dataset in [42] with small viewpoints changes and cannot exhibit view-dependent effects in large angle scale. Therefore we construct a novel dataset dedicated to evaluation for the 360-degree high-fidelity rendering of scenes containing complex reflections and refractions. In this dataset, we collect 25 synthesized scenes and 7 captured real-world scenes. Each synthesized scene consists of 120 images of 360-degree around reflective or refractive objects, with 100 randomly split for training, 10 for validation, and 10 for evaluation. Each real-world scene is captured randomly around scenes with reflective and refractive objects, consisting of 62 to 118 images, and organized under the convention of LLFF[24]. We then demonstrate the superiority and compatibility of our approach by comparisons, using three representative baseline models, $i.e.$, NeRF[25], Mip-NeRF[1], and Mip-NeRF 360[2], with and without our multi-space module. Experiments show that our approach improves performance by a large margin both quantitatively and qualitatively on scenes with reflection and refraction. Our main contributions are as follows:
· We propose a multi-space NeRF method that auto-
matically handles mirror-like objects in 360-degree high-fidelity scene rendering, achieving significant improvements over existing representative baselines both quantitatively and qualitatively.
· We design a lightweight module that can equip most NeRF-like methods with the ability to model reflection and refraction with small computational overheads.
· We construct a dataset dedicated to evaluation for the
360-degree high-fidelity rendering of scenes containing complex reflections and refractions, including 25 synthesized scenes and 7 real captured scenes.
## $2. \textbf{ Related works}$
Coordinate-based novel view synthesis. NeRF[25] has bridged the gap between computer vision and computer graphics, and reveals a promising way to render highquality photorealistic scenes with only posed images. The insights and the generalization ability of this scheme also facilitate various tasks both in CV and CG, i.e., 3D reconstruction [28,40], 3D-aware generation [4,15,27], 3D-aware edition [39,47], and avatar reconstruction and manipulation [9,18,52]. Besides, researchers have made great efforts to enhance this scheme. Mip-NeRF[1] enhances the
anti-aliasing ability of NeRF by featuring 3D conical frustum using integrated positional encoding.[14,23] adapt this scheme to HDR images.[2,50] extend NeRF and its variants to unbounded scenes. There are also many works trying to speed up the training and inference speed using explicit or hybrid representations [5,7,10,26,32,34,46].
Glossy materials with high specular have a great influence on NeRF-like methods,[38] is inspired by precomputation-based techniques [31] in computer graphics to represent and render view-dependent specular and reflection, but it fails to handle mirror-like reflective surfaces because the virtual images cannot be treated as textures. Guo et al.[12] propose to decompose reflective surfaces into a transmitted part and reflected part, which is the most relevant work to ours, but such decomposition cannot handle 360-degree views with mirror-like objects because the virtual images have no difference from real objects until the viewpoint moves beyond a certain angle.
Another line of work similar to ours is multiple neural radiance fields, but they do so for different purposes [11,27, 32,43,44].[27] uses object-level neural radiance fields for 3D-aware generation and composition.[32,44] uses multiple small MLPs for efficient rendering.[11,43] uses multiple object-level neural radiance fields for 3D scene decomposition and edition.
Commonly used datasets. Researchers have introduced or constructed many different datasets to facilitate the development of NeRF-based methods in various tasks. Mildenhall et al.[25] collect a dataset containing eight rendered sets of posed images about eight objects separately, and eight real captured forward-facing scenes with the camera poses and intrinsics estimated by COLMAP [33]. Nevertheless, these scenes lack reflection and refraction, which are very common. Wizadwongsa et $al.[42]$ propose a dataset, namely Shiny, that contains eight more challenging scenes to test NeRF-like methods on view-dependent effects, but they are captured in a roughly forward-facing manner. Verbin $et$ $al.[38]$ create a dataset of six glossy objects, namely Shiny Blender, which are rendered under similar conditions as done in NeRF to test methods in modeling more complex materials. For unbounded scenes, Barron $etal.[2]$ construct a dataset consisting of 5 outdoor scenes and 4 indoor scenes, while Zhang $et$ al.[50] adopt Tanks and Temples (T&T) dataset[19] and the Light Field dataset [48]. Bemanal et al.[3] capture a dataset consisting of refractive objects, which is composed of four scenes with cameras moving in a large range. Guo $etal.[12]$ collect six forward-facing scenes with reflective and semi-transparent materials, which is, to date, the most relevant dataset to ours, but ours is much more challenging. DTU dataset[17] and BlendedMVS dataset [45] are commonly used as benchmarks for the evaluation of 3D reconstruction.
2
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.07843017578125,
0.106689453125,
0.468017578125,
0.1060791015625,
0.468017578125,
0.12060546875,
0.07843017578125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.125,
0.428955078125,
0.125,
0.428955078125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.09722900390625,
0.136962890625,
0.4697265625,
0.1363525390625,
0.4697265625,
0.15087890625,
0.09722900390625,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1685791015625,
0.8037109375,
0.1685791015625,
0.8037109375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.184326171875,
0.4697265625,
0.184326171875,
0.4697265625,
0.1976318359375,
0.080078125,
0.1976318359375
],
[
0.521484375,
0.1875,
0.89208984375,
0.1875,
0.89208984375,
0.2008056640625,
0.521484375,
0.2008056640625
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.2020263671875,
0.890625,
0.2020263671875,
0.890625,
0.21533203125,
0.50244140625,
0.21533203125
],
[
0.08087158203125,
0.21533203125,
0.468017578125,
0.21533203125,
0.468017578125,
0.2259521484375,
0.08087158203125,
0.2259521484375
],
[
0.50244140625,
0.2171630859375,
0.890625,
0.2171630859375,
0.890625,
0.2310791015625,
0.50244140625,
0.2310791015625
],
[
0.08087158203125,
0.2298583984375,
0.4697265625,
0.2298583984375,
0.4697265625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50341796875,
0.2342529296875,
0.8896484375,
0.2342529296875,
0.8896484375,
0.2442626953125,
0.50341796875,
0.2442626953125
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.248779296875,
0.890625,
0.248779296875,
0.890625,
0.2587890625,
0.50341796875,
0.2587890625
],
[
0.08087158203125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.261962890625,
0.890625,
0.261962890625,
0.890625,
0.27587890625,
0.50244140625,
0.27587890625
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.27783203125,
0.89208984375,
0.27783203125,
0.89208984375,
0.291748046875,
0.50341796875,
0.291748046875
],
[
0.07843017578125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.07843017578125,
0.302978515625
],
[
0.50244140625,
0.29345703125,
0.89208984375,
0.29345703125,
0.89208984375,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.5048828125,
0.309326171875,
0.890625,
0.309326171875,
0.890625,
0.3193359375,
0.5048828125,
0.3193359375
],
[
0.080078125,
0.317626953125,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.330810546875
],
[
0.50244140625,
0.322509765625,
0.89208984375,
0.322509765625,
0.89208984375,
0.3359375,
0.50244140625,
0.3359375
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.5009765625,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.35107421875,
0.5009765625,
0.35107421875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50244140625,
0.354248046875,
0.77392578125,
0.354248046875,
0.77392578125,
0.367431640625,
0.50244140625,
0.367431640625
],
[
0.07843017578125,
0.363037109375,
0.468017578125,
0.3642578125,
0.468017578125,
0.37890625,
0.07843017578125,
0.37744140625
],
[
0.521484375,
0.371826171875,
0.89208984375,
0.371826171875,
0.89208984375,
0.385009765625,
0.521484375,
0.385009765625
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.39453125,
0.46728515625,
0.39453125,
0.46728515625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50341796875,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.50341796875,
0.416748046875
],
[
0.080078125,
0.4091796875,
0.4697265625,
0.4091796875,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50341796875,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.080078125,
0.4248046875,
0.46728515625,
0.4248046875,
0.46728515625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50341796875,
0.431884765625,
0.890625,
0.431884765625,
0.890625,
0.44580078125,
0.50341796875,
0.44580078125
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50341796875,
0.4609375
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.46337890625,
0.6396484375,
0.46337890625,
0.6396484375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.080078125,
0.47021484375,
0.39453125,
0.47021484375,
0.39453125,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.098876953125,
0.49560546875,
0.468017578125,
0.49560546875,
0.468017578125,
0.51025390625,
0.098876953125,
0.51025390625
],
[
0.50341796875,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50341796875,
0.5068359375
],
[
0.111083984375,
0.51123046875,
0.4697265625,
0.51123046875,
0.4697265625,
0.5244140625,
0.111083984375,
0.5244140625
],
[
0.50244140625,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.111083984375,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.111083984375,
0.54052734375
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.111083984375,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.111083984375,
0.5546875
],
[
0.50244140625,
0.54052734375,
0.8935546875,
0.54052734375,
0.8935546875,
0.5537109375,
0.50244140625,
0.5537109375
],
[
0.11273193359375,
0.55615234375,
0.318603515625,
0.55615234375,
0.318603515625,
0.5693359375,
0.11273193359375,
0.5693359375
],
[
0.50341796875,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50341796875,
0.568359375
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.098876953125,
0.58056640625,
0.4697265625,
0.58056640625,
0.4697265625,
0.59521484375,
0.098876953125,
0.59521484375
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.1094970703125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.1094970703125,
0.609375
],
[
0.50341796875,
0.6015625,
0.8896484375,
0.6015625,
0.8896484375,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.111083984375,
0.6123046875,
0.449462890625,
0.6123046875,
0.449462890625,
0.62548828125,
0.111083984375,
0.62548828125
],
[
0.50244140625,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50244140625,
0.62890625
],
[
0.50341796875,
0.630859375,
0.8935546875,
0.630859375,
0.8935546875,
0.64453125,
0.50341796875,
0.64453125
],
[
0.1004638671875,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.65087890625,
0.1004638671875,
0.65087890625
],
[
0.1094970703125,
0.6513671875,
0.465576171875,
0.6513671875,
0.465576171875,
0.66552734375,
0.1094970703125,
0.66552734375
],
[
0.50341796875,
0.646484375,
0.890625,
0.646484375,
0.890625,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.50244140625,
0.66015625,
0.8935546875,
0.66015625,
0.8935546875,
0.673828125,
0.50244140625,
0.673828125
],
[
0.111083984375,
0.66748046875,
0.46728515625,
0.66748046875,
0.46728515625,
0.68115234375,
0.111083984375,
0.68115234375
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.111083984375,
0.6845703125,
0.416748046875,
0.6845703125,
0.416748046875,
0.69775390625,
0.111083984375,
0.69775390625
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.08087158203125,
0.7119140625,
0.22216796875,
0.7119140625,
0.22216796875,
0.7265625,
0.08087158203125,
0.7265625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.5048828125,
0.7373046875,
0.88818359375,
0.7373046875,
0.88818359375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.50341796875,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.8896484375,
0.7666015625,
0.8896484375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5009765625,
0.78076171875,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
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,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.8935546875,
0.84033203125,
0.8935546875,
0.85400390625,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50244140625,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.74169921875,
0.88720703125,
0.74169921875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.4755859375,
0.923828125,
0.49267578125,
0.92041015625,
0.496826171875,
0.93603515625,
0.48046875,
0.939453125
]
] |
[
"a general enhancement to the NeRF-based backbone, equip-",
"anti-aliasing ability of NeRF by featuring 3D conical frus-",
"ping most NeRF-like methods with the ability to model",
"tum using integrated positional encoding. [14, 23] adapt this",
"complexreflection and refraction, as shown in Fig. 1.",
"scheme to HDR images. [2,50] extend NeRF and its variants",
"Existing datasets have not paid enough attention to the",
"to unbounded scenes. There are also many works trying to",
"360-degree rendering of scenes containing mirror-like ob-",
"speed up the training and inference speed using explicit or",
"jects, such as RFFR [12] just has forward-facing scenes,",
"hybrid representations [5,7, 10,26, 32,34, 46].",
"and the Shiny dataset in [42] with small viewpoints changes",
"Glossy materials with high specular have a great",
"and cannot exhibit view-dependent effects in large angle",
"influence on NeRF-like methods,[38] is inspired by",
"scale.Therefore we construct a novel dataset dedicated",
"precomputation-based techniques [31] in computer graph-",
"toevaluationforthe360-degreehigh-fidelityrendering of",
"icstorepresentandrenderview-dependentspecularandre-",
"scenes containing complex reflections and refractions.In",
"flection,butitfailstohandlemirror-likereflective surfaces",
"this dataset,wecollect25 synthesized scenesand7captured",
"because the virtual images cannot be treated as textures.",
"real-world scenes. Each synthesized scene consists of 120",
"Guo et al. [12] propose to decompose reflective surfaces",
"images of 360-degree around reflective or refractive objects,",
"into a transmitted part and reflected part, which is the most",
"with 100 randomly split for training, 10 for validation, and",
"relevantworktoours.butsuchdecompositioncannothan-",
"1o for evaluation.Each real-world scene is captured ran-",
"dle 360-degree views with mirror-like objects because the",
"domly around scenes with reflective and refractive objects,",
"virtual images have no difference from real objects until the",
"consisting of 62 to 118 images, and organized under the",
"viewpoint moves beyond a certain angle.",
"convention of LLFF [24]. We then demonstrate the supe-",
"Another line of work similar to ours is multiple neural",
"riority and compatibility of our approach by comparisons,",
"radiance fields, but they do so for different purposes [11,27,",
"using three representative baseline models, i.e., NeRF [25],",
"32, 43, 44]. [27] uses object-level neural radiance fields for",
"Mip-NeRF [1], and Mip-NeRF 360 [2], with and without",
"3D-aware generation and composition. [32, 44] uses multi-",
"our multi-space module. Experiments show that our ap-",
"ple small MLPs for efficient rendering. [11, 43] uses multi-",
"proach improves performance by a large margin both quan-",
"ple object-level neural radiance fields for 3D scene decom-",
"titatively and qualitatively on scenes with reflection and re-",
"position and edition.",
"fraction. Our main contributions are as follows:",
". We propose a multi-space NeRF method that auto-",
"Commonly used datasets. Researchers have introduced or",
"matically handles mirror-like objects in 360-degree",
"constructed many different datasets to facilitate the develop-",
"high-fidelity scene rendering, achieving significant im-",
"ment of NeRF-based methods in various tasks. Mildenhall",
"provements over existing representative baselines both",
"et al. [25] collect a dataset containing eight rendered sets of",
"quantitatively and qualitatively.",
"posed images about eight objects separately, and eight real",
"captured forward-facing scenes with the camera poses and",
". We design a lightweight module that can equip most",
"intrinsics estimated by COLMAP [33]. Nevertheless, these",
"NeRF-like methods with the ability to model reflection",
"scenes lack reflection and refraction, which are very com-",
"and refraction with small computational overheads.",
"mon. Wizadwongsa et al. [42] propose a dataset, namely",
"Shiny, that contains eight more challenging scenes to test",
".We construct a dataset dedicated to evaluation for the",
"360-degree high-fidelity rendering of scenes contain-",
"NeRF-like methods on view-dependent effects,but they are",
"captured in a roughly forward-facing manner. Verbin et",
"ing complex reflections and refractions, including 25",
"al.[38| create a dataset of six glossy objects, namely Shiny",
"synthesized scenes and 7 real captured scenes.",
"Blender, which are rendered under similar conditions as",
"done in NeRF to test methods in modeling more complex",
"2.Related works",
"materials. For unbounded scenes, Barron et al. [2] construct",
"adatasetconsistingof5outdoorscenesand4indoorscenes",
"Coordinate-based novel view synthesis. NeRF [25] has",
"while Zhang et al. [50] adopt Tanks and Temples (T&T)",
"bridged the gap between computer vision and computer",
"dataset [19] and the Light Field dataset [48]. Bemanal",
"graphics, and reveals a promising way to render high-",
"et al. [3] capture a dataset consisting of refractive objects,",
"quality photorealistic scenes with only posed images. The",
"which is composed of four scenes with cameras moving in",
"insights and the generalization ability of this scheme also",
"a large range. Guo et al. [12] collect six forward-facing",
"facilitate various tasks both in CV and CG, i.e., 3D recon-",
"scenes with reflective and semi-transparent materials, which",
"struction [28,40], 3D-aware generation [4,15,27], 3D-aware",
"is, to date, the most relevant dataset to ours, but ours is",
"edition [39, 47], and avatar reconstruction and manipula-",
"much more challenging. DTU dataset [17] and Blended-",
"tion [9, 18, 52]. Besides, researchers have made great ef-",
"MVS dataset [45] are commonly used as benchmarks for",
"forts to enhance this scheme. Mip-NeRF [1] enhances the",
"the evaluation of 3D reconstruction.",
"2"
] |
{
"arxiv_id": "2305.04268",
"language": "en",
"page_idx": 1
}
|
||
2305.04268-en-4
|
2305.04268-en-4.jpg
|
<table>
<tbody>
<tr>
<th>dataset</th>
<th>origin</th>
<th>applications</th>
<th>Type</th>
<th>viewpoints</th>
<th>properties</th>
<th>number</th>
</tr>
<tr>
<td>$\mathrm{Realistic~Synthetic~360}^{\circ}$</td>
<td>[25]</td>
<td>view synthesis</td>
<td>$S$</td>
<td>360-degree</td>
<td>non-Lambertian</td>
<td>8</td>
</tr>
<tr>
<td>Real Forward- Facing</td>
<td>[24,25]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>forward-facing</td>
<td>non-Lambertian</td>
<td>8</td>
</tr>
<tr>
<td>Shiny</td>
<td>[42]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>forward-facing</td>
<td>high-specular, refraction</td>
<td>8</td>
</tr>
<tr>
<td>Tanks and Temples(T&T)</td>
<td>[19]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>360-degree</td>
<td>unbounded scenes</td>
<td>4</td>
</tr>
<tr>
<td>Mip- NeRF360</td>
<td>[2]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>360-degree</td>
<td>unbounded 1 scenes</td>
<td>9</td>
</tr>
<tr>
<td>EikonalFields</td>
<td>[3]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>large view angle</td>
<td>refraction</td>
<td>4</td>
</tr>
<tr>
<td>RFFR</td>
<td>[12]</td>
<td>view synthesis</td>
<td>$R$</td>
<td>forward-facing</td>
<td>reflection, semi-transparent</td>
<td>6</td>
</tr>
<tr>
<td>$DTU$</td>
<td>[17]</td>
<td>reconstruction</td>
<td>$R$</td>
<td>360-degree</td>
<td>non-Lambertian objects</td>
<td>$15^{*}$</td>
</tr>
<tr>
<td>BlendedMVS</td>
<td>[45]</td>
<td>reconstruction</td>
<td>$S$</td>
<td>360-degree</td>
<td>non-Lambertian scenes</td>
<td>7*</td>
</tr>
<tr>
<td>Shiny Blender</td>
<td>[38]</td>
<td>view synthesis</td>
<td>S</td>
<td>360-degree</td>
<td>glossy materials</td>
<td>6</td>
</tr>
<tr>
<td>Ref-NeRF Real captured scenes</td>
<td>[13,. 38 </td>
<td>view synthesis</td>
<td>$R$</td>
<td>360-degree</td>
<td>glossy $\angle{\mathrm{materials}}$</td>
<td>3</td>
</tr>
</tbody>
</table>
Table 1. Properties of a commonly used dataset for NeRF-based methods. ‘S’and 'R’represent synthesized and real captured, respectively. We denote those unnamed datasets with the name of the methods. ’*’refers to the number of scenes commonly used by NeRF-based methods, as the original dataset contains more scenes than noted, and we do not take them into consideration
the conventional NeRF-like backbone networks, and we will show that our approach achieves significant enhancements in Sec. 5.2.
# 4. Dataset
parameters for the total sub-space number and the feature
dimension of the neural feature field, respectively.
We then integrate features along the ray in each subspace to collect $K$ feature maps that encode the color information and visibility of each sub-space from a certain viewpoint. As all pixels are calculated the same way, we denote each pixel as $\{\mathcal{F}^k\}$ for simplicity and describe the operation at the pixel level. Each pixel $\{\mathcal{F}^k\}$ of the feature maps is calculated using:
## 4.1. Existing datasets
We briefly revisit the commonly used or most relevant datasets to our task and list their properties in Tab. 1. As can be seen, there lacks a 360-degree benchmark for scenes with complex light paths, $e.g.$, a glass of water in front of a mirror.
$$\hat{\mathcal{F}^k}(\mathbf{r})=\sum_{i=1}^NT_i^k(1-\exp(-\sigma_i^k\delta_i))\mathbf{f}_i^k,$$
(4)
## 4.2. Our proposed dataset
where the superscript $k$ indicates the sub-space that the ray casts through. The $k$-th density $\sigma_i^k$ and feature $\mathbf{f}_i^k$ correspond to the $k$-th sub-space. $T_i^k=\exp(-\sum_{j=1}^{i-1}\sigma_j^k\delta_j)$ and $\delta_{i}=t_{i}-t_{i-1}$ are similarly computed as in Eq.(2).
Then, $\{\mathcal{F}^k\}$ is decoded by two small MLPs, each with just one hidden layer. The first is a Decoder MLP that takes $\{\mathcal{F}^k\}$ as input and outputs RGB vectors. The second is a Gate MLP that takes $\{\mathcal{F}^k\}$ as input and outputs weights that control the visibility of certain sub-space. Specifically, We use:
(5)
$$\{\mathcal{F}^{k}\}\stackrel{\Theta_{D}}{\longrightarrow}\{\mathbf{C}^{k}\},\{\mathcal{F}^{k}\}\stackrel{\Theta_{G}}{\longrightarrow}\{w^{k}\},$$
As summarized in Sec. 4.1, there lacks a 360-degree dataset consisting of complex reflection and refraction to facilitate the related research. Therefore we collect a 360- degree dataset comprising 25 synthetic scenes and 7 real captured scenes.
For our synthesized part shown in Fig. 6a, we use an open source software Blender[8] and design our scenes with 3D models from BlenderKit, a community for sharing 3D models. As our dataset consists of complete scenes instead of single objects, we fx the height of our camera position with the camera looking at the center of the scene and moving the camera around a circle to render the whole scene.
For all our scenes, we uniformly sample 120 points along the circle and randomly choose 100 images for the training set, 10 for the validation set, and 10 for the test set. The constructed dataset features a wide variety of scenes containing reflective and refractive objects. We include a variety of complexity of light paths, controlled by the number and the layout of the mirror(s) in the scene, where the number of mirrors ranges from 1 up to tens of small pieces. Note that even a scene in our dataset with only one mirror is more challenging than RFFR [25], as our camera moves from the
where $\Theta_D$ represents the Decoder MLP, and $\Theta_G$ represents the Gate MLP. In the end, the MS module applies the softmax function to $\{w^k\}$ as the color contribution of each subspace to form the final render results:
$$\hat{\mathbf C}(\mathbf r)=\frac{1}{\sum_{i=1}^K\exp(w^i)}\sum_{k=1}^K\exp(w^k)\mathbf C^k.$$
(6)
Eq. (6) needs no additional loss terms compared with the vanilla NeRF method. As a result, the above light-weighted MS module is able to serve as an enhancement module onto
5
|
[
[
0.1993408203125,
0.0947265625,
0.2484130859375,
0.0947265625,
0.2484130859375,
0.10919189453125,
0.1993408203125,
0.10919189453125
],
[
0.4794921875,
0.0953369140625,
0.51318359375,
0.0953369140625,
0.51318359375,
0.10797119140625,
0.4794921875,
0.10797119140625
],
[
0.7939453125,
0.0953369140625,
0.84326171875,
0.09344482421875,
0.8447265625,
0.106689453125,
0.7958984375,
0.10919189453125
],
[
0.39208984375,
0.09661865234375,
0.465576171875,
0.09661865234375,
0.465576171875,
0.10797119140625,
0.39208984375,
0.10797119140625
],
[
0.330810546875,
0.09661865234375,
0.371826171875,
0.09661865234375,
0.371826171875,
0.10797119140625,
0.330810546875,
0.10797119140625
],
[
0.5361328125,
0.09661865234375,
0.60107421875,
0.09661865234375,
0.60107421875,
0.10797119140625,
0.5361328125,
0.10797119140625
],
[
0.67626953125,
0.09661865234375,
0.73876953125,
0.09661865234375,
0.73876953125,
0.10797119140625,
0.67626953125,
0.10797119140625
],
[
0.148681640625,
0.1142578125,
0.294189453125,
0.1142578125,
0.294189453125,
0.1275634765625,
0.148681640625,
0.1275634765625
],
[
0.336669921875,
0.11297607421875,
0.369384765625,
0.11297607421875,
0.369384765625,
0.1287841796875,
0.336669921875,
0.1287841796875
],
[
0.38232421875,
0.11553955078125,
0.471435546875,
0.11553955078125,
0.471435546875,
0.126953125,
0.38232421875,
0.126953125
],
[
0.53515625,
0.11553955078125,
0.60302734375,
0.11553955078125,
0.60302734375,
0.126953125,
0.53515625,
0.126953125
],
[
0.65869140625,
0.11553955078125,
0.75634765625,
0.11553955078125,
0.75634765625,
0.126953125,
0.65869140625,
0.126953125
],
[
0.80615234375,
0.12060546875,
0.82177734375,
0.111083984375,
0.83251953125,
0.12060546875,
0.81787109375,
0.1300048828125
],
[
0.158447265625,
0.1275634765625,
0.285888671875,
0.1287841796875,
0.285888671875,
0.1427001953125,
0.158447265625,
0.141357421875
],
[
0.484375,
0.117431640625,
0.498291015625,
0.11236572265625,
0.5048828125,
0.12249755859375,
0.491943359375,
0.1275634765625
],
[
0.330078125,
0.1300048828125,
0.37744140625,
0.1300048828125,
0.37744140625,
0.141357421875,
0.330078125,
0.141357421875
],
[
0.38232421875,
0.1300048828125,
0.47216796875,
0.1300048828125,
0.47216796875,
0.141357421875,
0.38232421875,
0.141357421875
],
[
0.488525390625,
0.1300048828125,
0.5009765625,
0.1300048828125,
0.5009765625,
0.14013671875,
0.488525390625,
0.14013671875
],
[
0.52392578125,
0.1300048828125,
0.61376953125,
0.1300048828125,
0.61376953125,
0.141357421875,
0.52392578125,
0.141357421875
],
[
0.65869140625,
0.1300048828125,
0.7548828125,
0.1300048828125,
0.7548828125,
0.14013671875,
0.65869140625,
0.14013671875
],
[
0.81201171875,
0.1275634765625,
0.82861328125,
0.1275634765625,
0.82861328125,
0.143310546875,
0.81201171875,
0.143310546875
],
[
0.2034912109375,
0.14013671875,
0.24267578125,
0.1427001953125,
0.2418212890625,
0.1572265625,
0.2017822265625,
0.1546630859375
],
[
0.339111328125,
0.1427001953125,
0.36767578125,
0.1427001953125,
0.36767578125,
0.1572265625,
0.339111328125,
0.1572265625
],
[
0.63232421875,
0.1427001953125,
0.78125,
0.1427001953125,
0.78125,
0.1558837890625,
0.63232421875,
0.1558837890625
],
[
0.81201171875,
0.141357421875,
0.82861328125,
0.141357421875,
0.82861328125,
0.1558837890625,
0.81201171875,
0.1558837890625
],
[
0.488525390625,
0.143310546875,
0.5009765625,
0.143310546875,
0.5009765625,
0.154052734375,
0.488525390625,
0.154052734375
],
[
0.384033203125,
0.143310546875,
0.47216796875,
0.143310546875,
0.47216796875,
0.1546630859375,
0.384033203125,
0.1546630859375
],
[
0.52294921875,
0.143310546875,
0.61376953125,
0.143310546875,
0.61376953125,
0.1546630859375,
0.52294921875,
0.1546630859375
],
[
0.1446533203125,
0.1558837890625,
0.301513671875,
0.1558837890625,
0.301513671875,
0.169189453125,
0.1446533203125,
0.169189453125
],
[
0.53173828125,
0.154052734375,
0.6044921875,
0.1558837890625,
0.6044921875,
0.17041015625,
0.53173828125,
0.1685791015625
],
[
0.65185546875,
0.1558837890625,
0.76318359375,
0.1558837890625,
0.76318359375,
0.169189453125,
0.65185546875,
0.169189453125
],
[
0.384033203125,
0.1572265625,
0.473876953125,
0.1572265625,
0.473876953125,
0.1685791015625,
0.384033203125,
0.1685791015625
],
[
0.339111328125,
0.1572265625,
0.366943359375,
0.1572265625,
0.366943359375,
0.169189453125,
0.339111328125,
0.169189453125
],
[
0.488525390625,
0.1578369140625,
0.50244140625,
0.1578369140625,
0.50244140625,
0.1673583984375,
0.488525390625,
0.1673583984375
],
[
0.80615234375,
0.1622314453125,
0.82177734375,
0.15283203125,
0.83251953125,
0.1622314453125,
0.81787109375,
0.1717529296875
],
[
0.53173828125,
0.1685791015625,
0.6044921875,
0.17041015625,
0.6044921875,
0.1849365234375,
0.53173828125,
0.18310546875
],
[
0.81201171875,
0.169189453125,
0.82861328125,
0.169189453125,
0.82861328125,
0.184326171875,
0.81201171875,
0.184326171875
],
[
0.343017578125,
0.17041015625,
0.365234375,
0.17041015625,
0.365234375,
0.184326171875,
0.343017578125,
0.184326171875
],
[
0.17724609375,
0.1717529296875,
0.26953125,
0.1717529296875,
0.26953125,
0.1817626953125,
0.17724609375,
0.1817626953125
],
[
0.384033203125,
0.1717529296875,
0.47216796875,
0.1717529296875,
0.47216796875,
0.18310546875,
0.384033203125,
0.18310546875
],
[
0.488525390625,
0.1717529296875,
0.5009765625,
0.1717529296875,
0.5009765625,
0.1817626953125,
0.488525390625,
0.1817626953125
],
[
0.65185546875,
0.1717529296875,
0.76123046875,
0.1717529296875,
0.76123046875,
0.1817626953125,
0.65185546875,
0.1817626953125
],
[
0.343017578125,
0.184326171875,
0.365234375,
0.184326171875,
0.365234375,
0.1976318359375,
0.343017578125,
0.1976318359375
],
[
0.81201171875,
0.18310546875,
0.82861328125,
0.18310546875,
0.82861328125,
0.1976318359375,
0.81201171875,
0.1976318359375
],
[
0.384033203125,
0.1849365234375,
0.47216796875,
0.1849365234375,
0.47216796875,
0.1962890625,
0.384033203125,
0.1962890625
],
[
0.1805419921875,
0.1849365234375,
0.265625,
0.1849365234375,
0.265625,
0.1962890625,
0.1805419921875,
0.1962890625
],
[
0.490234375,
0.1849365234375,
0.5009765625,
0.1849365234375,
0.5009765625,
0.1956787109375,
0.490234375,
0.1956787109375
],
[
0.51953125,
0.1849365234375,
0.61669921875,
0.1849365234375,
0.61669921875,
0.1962890625,
0.51953125,
0.1962890625
],
[
0.67724609375,
0.1849365234375,
0.73779296875,
0.1849365234375,
0.73779296875,
0.1962890625,
0.67724609375,
0.1962890625
],
[
0.51953125,
0.1956787109375,
0.615234375,
0.1976318359375,
0.615234375,
0.212158203125,
0.51953125,
0.210205078125
],
[
0.625,
0.1976318359375,
0.7900390625,
0.1976318359375,
0.7900390625,
0.2108154296875,
0.625,
0.2108154296875
],
[
0.339111328125,
0.1988525390625,
0.366943359375,
0.1988525390625,
0.366943359375,
0.2108154296875,
0.339111328125,
0.2108154296875
],
[
0.2034912109375,
0.1988525390625,
0.24267578125,
0.1988525390625,
0.24267578125,
0.210205078125,
0.2034912109375,
0.210205078125
],
[
0.384033203125,
0.199462890625,
0.473876953125,
0.199462890625,
0.473876953125,
0.210205078125,
0.384033203125,
0.210205078125
],
[
0.488525390625,
0.199462890625,
0.5009765625,
0.199462890625,
0.5009765625,
0.208984375,
0.488525390625,
0.208984375
],
[
0.81396484375,
0.1988525390625,
0.826171875,
0.1988525390625,
0.826171875,
0.210205078125,
0.81396484375,
0.210205078125
],
[
0.20751953125,
0.212158203125,
0.240234375,
0.212158203125,
0.240234375,
0.2247314453125,
0.20751953125,
0.2247314453125
],
[
0.339111328125,
0.212158203125,
0.36767578125,
0.212158203125,
0.36767578125,
0.2266845703125,
0.339111328125,
0.2266845703125
],
[
0.63232421875,
0.212158203125,
0.77783203125,
0.212158203125,
0.77783203125,
0.2259521484375,
0.63232421875,
0.2259521484375
],
[
0.80810546875,
0.2108154296875,
0.83251953125,
0.2108154296875,
0.83251953125,
0.2235107421875,
0.80810546875,
0.2235107421875
],
[
0.488525390625,
0.21337890625,
0.5009765625,
0.21337890625,
0.5009765625,
0.2235107421875,
0.488525390625,
0.2235107421875
],
[
0.38232421875,
0.21337890625,
0.47216796875,
0.212158203125,
0.47216796875,
0.2235107421875,
0.38232421875,
0.2247314453125
],
[
0.53515625,
0.21337890625,
0.60302734375,
0.21337890625,
0.60302734375,
0.2247314453125,
0.53515625,
0.2247314453125
],
[
0.1805419921875,
0.2247314453125,
0.265625,
0.2247314453125,
0.265625,
0.2392578125,
0.1805419921875,
0.2392578125
],
[
0.339111328125,
0.2247314453125,
0.36767578125,
0.2266845703125,
0.365234375,
0.240478515625,
0.338134765625,
0.238037109375
],
[
0.495849609375,
0.2247314453125,
0.50341796875,
0.2310791015625,
0.495849609375,
0.2374267578125,
0.48779296875,
0.2310791015625
],
[
0.53369140625,
0.2259521484375,
0.6044921875,
0.2259521484375,
0.6044921875,
0.240478515625,
0.53369140625,
0.240478515625
],
[
0.6357421875,
0.2259521484375,
0.77783203125,
0.2259521484375,
0.77783203125,
0.2392578125,
0.6357421875,
0.2392578125
],
[
0.810546875,
0.2247314453125,
0.8310546875,
0.2247314453125,
0.8310546875,
0.238037109375,
0.810546875,
0.238037109375
],
[
0.53173828125,
0.238037109375,
0.6044921875,
0.240478515625,
0.6044921875,
0.25390625,
0.53173828125,
0.251953125
],
[
0.384033203125,
0.2266845703125,
0.47216796875,
0.2266845703125,
0.47216796875,
0.238037109375,
0.384033203125,
0.238037109375
],
[
0.1805419921875,
0.240478515625,
0.26708984375,
0.240478515625,
0.26708984375,
0.251953125,
0.1805419921875,
0.251953125
],
[
0.339111328125,
0.240478515625,
0.366943359375,
0.240478515625,
0.366943359375,
0.25244140625,
0.339111328125,
0.25244140625
],
[
0.384033203125,
0.240478515625,
0.47216796875,
0.240478515625,
0.47216796875,
0.251953125,
0.384033203125,
0.251953125
],
[
0.490234375,
0.2412109375,
0.5009765625,
0.2412109375,
0.5009765625,
0.250732421875,
0.490234375,
0.250732421875
],
[
0.65869140625,
0.2412109375,
0.75634765625,
0.2412109375,
0.75634765625,
0.25244140625,
0.65869140625,
0.25244140625
],
[
0.810546875,
0.2392578125,
0.82861328125,
0.2392578125,
0.82861328125,
0.25390625,
0.810546875,
0.25390625
],
[
0.53173828125,
0.251953125,
0.6044921875,
0.25390625,
0.6044921875,
0.268310546875,
0.53173828125,
0.266357421875
],
[
0.810546875,
0.25244140625,
0.82861328125,
0.25244140625,
0.82861328125,
0.266357421875,
0.810546875,
0.266357421875
],
[
0.325927734375,
0.255126953125,
0.378173828125,
0.255126953125,
0.378173828125,
0.26513671875,
0.325927734375,
0.26513671875
],
[
0.1282958984375,
0.255126953125,
0.328369140625,
0.255126953125,
0.328369140625,
0.26513671875,
0.1282958984375,
0.26513671875
],
[
0.38232421875,
0.255126953125,
0.47216796875,
0.255126953125,
0.47216796875,
0.266357421875,
0.38232421875,
0.266357421875
],
[
0.488525390625,
0.255126953125,
0.50244140625,
0.255126953125,
0.50244140625,
0.264404296875,
0.488525390625,
0.264404296875
],
[
0.65869140625,
0.255126953125,
0.75634765625,
0.255126953125,
0.75634765625,
0.266357421875,
0.65869140625,
0.266357421875
],
[
0.080078125,
0.2822265625,
0.8896484375,
0.2822265625,
0.8896484375,
0.29541015625,
0.080078125,
0.29541015625
],
[
0.375732421875,
0.25830078125,
0.3857421875,
0.25830078125,
0.3857421875,
0.261962890625,
0.375732421875,
0.261962890625
],
[
0.0767822265625,
0.292236328125,
0.8935546875,
0.29345703125,
0.8935546875,
0.311279296875,
0.0767822265625,
0.31005859375
],
[
0.07843017578125,
0.31005859375,
0.7236328125,
0.31005859375,
0.7236328125,
0.323974609375,
0.07843017578125,
0.323974609375
],
[
0.07843017578125,
0.34033203125,
0.471435546875,
0.339111328125,
0.471435546875,
0.352783203125,
0.07843017578125,
0.354248046875
],
[
0.50244140625,
0.339111328125,
0.89208984375,
0.339111328125,
0.89208984375,
0.352783203125,
0.50244140625,
0.352783203125
],
[
0.080078125,
0.354736328125,
0.410888671875,
0.354736328125,
0.410888671875,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.50244140625,
0.354736328125,
0.89208984375,
0.354736328125,
0.89208984375,
0.36865234375,
0.50244140625,
0.36865234375
],
[
0.1004638671875,
0.369384765625,
0.468017578125,
0.369384765625,
0.468017578125,
0.38330078125,
0.1004638671875,
0.38330078125
],
[
0.5009765625,
0.369384765625,
0.62353515625,
0.369384765625,
0.62353515625,
0.38330078125,
0.5009765625,
0.38330078125
],
[
0.080078125,
0.385009765625,
0.468017578125,
0.385009765625,
0.468017578125,
0.3984375,
0.080078125,
0.3984375
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.4140625,
0.080078125,
0.4140625
],
[
0.50244140625,
0.3984375,
0.5888671875,
0.3984375,
0.5888671875,
0.41357421875,
0.50244140625,
0.41357421875
],
[
0.07843017578125,
0.4140625,
0.4697265625,
0.415283203125,
0.4697265625,
0.4287109375,
0.07843017578125,
0.427978515625
],
[
0.50244140625,
0.42236328125,
0.66650390625,
0.42236328125,
0.66650390625,
0.4375,
0.50244140625,
0.4375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.080078125,
0.4443359375,
0.4697265625,
0.4443359375,
0.4697265625,
0.458984375,
0.080078125,
0.458984375
],
[
0.52294921875,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.460205078125,
0.52294921875,
0.460205078125
],
[
0.080078125,
0.460205078125,
0.2459716796875,
0.460205078125,
0.2459716796875,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.50341796875,
0.4609375,
0.89208984375,
0.4609375,
0.89208984375,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.50244140625,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.2115478515625,
0.490478515625,
0.2279052734375,
0.481689453125,
0.236083984375,
0.4912109375,
0.2197265625,
0.5
],
[
0.50244140625,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.50244140625,
0.505859375
],
[
0.4453125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.51318359375,
0.4453125,
0.51318359375
],
[
0.50341796875,
0.50830078125,
0.5498046875,
0.50830078125,
0.5498046875,
0.51953125,
0.50341796875,
0.51953125
],
[
0.146240234375,
0.49560546875,
0.40283203125,
0.49560546875,
0.40283203125,
0.5166015625,
0.146240234375,
0.5166015625
],
[
0.2115478515625,
0.5166015625,
0.236083984375,
0.5166015625,
0.236083984375,
0.52783203125,
0.2115478515625,
0.52783203125
],
[
0.50244140625,
0.52978515625,
0.705078125,
0.53076171875,
0.705078125,
0.54541015625,
0.50244140625,
0.54443359375
],
[
0.080078125,
0.53857421875,
0.468017578125,
0.53857421875,
0.468017578125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.07598876953125,
0.55322265625,
0.4697265625,
0.55029296875,
0.4697265625,
0.568359375,
0.07598876953125,
0.57080078125
],
[
0.521484375,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.568359375,
0.521484375,
0.568359375
],
[
0.07598876953125,
0.56640625,
0.473876953125,
0.56640625,
0.473876953125,
0.59033203125,
0.07598876953125,
0.59033203125
],
[
0.50244140625,
0.57080078125,
0.8935546875,
0.57080078125,
0.8935546875,
0.583984375,
0.50244140625,
0.583984375
],
[
0.07843017578125,
0.58544921875,
0.416748046875,
0.5859375,
0.416748046875,
0.599609375,
0.07843017578125,
0.5986328125
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.098876953125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.098876953125,
0.6142578125
],
[
0.50244140625,
0.60107421875,
0.89208984375,
0.60107421875,
0.89208984375,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.6162109375,
0.4697265625,
0.6162109375,
0.4697265625,
0.6298828125,
0.080078125,
0.6298828125
],
[
0.5009765625,
0.61572265625,
0.61279296875,
0.61572265625,
0.61279296875,
0.62890625,
0.5009765625,
0.62890625
],
[
0.08087158203125,
0.630859375,
0.4697265625,
0.630859375,
0.4697265625,
0.64453125,
0.08087158203125,
0.64453125
],
[
0.521484375,
0.6298828125,
0.8935546875,
0.6298828125,
0.8935546875,
0.64306640625,
0.521484375,
0.64306640625
],
[
0.080078125,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.66015625,
0.080078125,
0.66015625
],
[
0.50244140625,
0.64599609375,
0.8935546875,
0.64599609375,
0.8935546875,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.080078125,
0.662109375,
0.468017578125,
0.662109375,
0.468017578125,
0.67529296875,
0.080078125,
0.67529296875
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.6787109375,
0.13232421875,
0.6787109375,
0.13232421875,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.156005859375,
0.69873046875,
0.393798828125,
0.69970703125,
0.393798828125,
0.72021484375,
0.156005859375,
0.7197265625
],
[
0.44677734375,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.44677734375,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50244140625,
0.7216796875,
0.89208984375,
0.7216796875,
0.89208984375,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.080078125,
0.73193359375,
0.4697265625,
0.73193359375,
0.4697265625,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50341796875,
0.7373046875,
0.5439453125,
0.7373046875,
0.5439453125,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.080078125,
0.74755859375,
0.468017578125,
0.74755859375,
0.468017578125,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.521484375,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.521484375,
0.76513671875
],
[
0.080078125,
0.76220703125,
0.468017578125,
0.76220703125,
0.468017578125,
0.77587890625,
0.080078125,
0.77587890625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.77783203125,
0.325927734375,
0.77783203125,
0.325927734375,
0.79150390625,
0.080078125,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.240234375,
0.8095703125,
0.250732421875,
0.8095703125,
0.250732421875,
0.8212890625,
0.240234375,
0.8212890625
],
[
0.3056640625,
0.8037109375,
0.318603515625,
0.8037109375,
0.318603515625,
0.81298828125,
0.3056640625,
0.81298828125
],
[
0.1396484375,
0.81494140625,
0.1895751953125,
0.81494140625,
0.1895751953125,
0.83056640625,
0.1396484375,
0.83056640625
],
[
0.31787109375,
0.81640625,
0.406005859375,
0.8125,
0.406982421875,
0.830078125,
0.318603515625,
0.833984375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.4453125,
0.8173828125,
0.4697265625,
0.8173828125,
0.4697265625,
0.83056640625,
0.4453125,
0.83056640625
],
[
0.1953125,
0.82568359375,
0.324462890625,
0.82763671875,
0.324462890625,
0.8466796875,
0.1953125,
0.8447265625
],
[
0.205078125,
0.8271484375,
0.223876953125,
0.8271484375,
0.223876953125,
0.833984375,
0.205078125,
0.833984375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.1004638671875,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.1004638671875,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50244140625,
0.87353515625,
0.89208984375,
0.87353515625,
0.89208984375,
0.8837890625,
0.50244140625,
0.8837890625
],
[
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
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"dataset",
"Type",
"number",
"applications",
"origin",
"viewpoints",
"properties",
"Realistic Synthetic 3600",
"[25]",
"view synthesis",
"360-degree",
"non-Lambertian",
"8",
"Real Forward-Facing",
"S",
"[24,25]",
"view synthesis",
"R",
"forward-facing",
"non-Lambertian",
"8",
"Shiny",
"[42]",
"high-specular, refraction",
"8",
"R",
"view synthesis",
"forward-facing",
"Tanks and Temples(T&T)",
"360-degree",
"unbounded scenes",
"view synthesis",
"[19]",
"R",
"X",
"360-degree",
"9",
"[2] ",
"Mip-NeRF360",
"view synthesis",
"R",
"unbounded scenes",
"[3] ",
"4",
"view synthesis",
"EikonalFields",
"R",
"large angle view",
"refraction",
"forward-facing",
"reflection, semi-transparent",
"[12]",
"RFFR",
"view synthesis",
"R",
"6",
"DTU",
"[17]",
" non-Lambertian objects",
"15*",
"R",
"reconstruction",
"360-degree",
"BlendedMVS",
"[45]",
"S",
"360-degree",
"non-Lambertian scenes",
"7*",
"360-degree",
"reconstruction",
"Shiny Blender",
"[38]",
"view synthesis",
"S",
"glossy materials",
"6",
"360-degree",
"3",
"[13,38]",
"Ref-NeRFRealcapturedscenes",
"view synthesis",
"R",
"glossy materials",
"Table 1. Properties of a commonly used dataset for NeRF-based methods. 'S' and 'R' represent synthesized and real captured, respectively",
"",
"We denote those unnamed datasets with the name of the methods. *** refers to the number of scenes commonly used by NeRF-based",
"methods, as the original dataset contains more scenes than noted, and we do not take them into consideration.",
"parameters for the total sub-space number and the feature",
"the conventional NeRF-like backbone networks, and we",
"dimension of the neural feature field, respectively.",
"will show that our approach achieves significant enhance-",
"We then integrate features along the ray in each sub-",
"ments in Sec. 5.2.",
"space to collect K feature maps that encode the color in-",
"formation and visibility of each sub-space from a certain",
"4. Dataset",
"viewpoint.As all pixels are calculated the sameway,we",
"4.1. Existing datasets",
"denote each pixel as {Fk) for simplicity and describe the",
"operation at the pixel level. Each pixel [Fk} of the feature",
"We briefly revisit the commonly used or most relevant",
"maps is calculated using:",
"datasets to our task and list their properties in Tab. 1. As",
"can be seen, there lacks a 360-degree benchmark for scenes",
"",
"with complex light paths, e.g., a glass of water in front of a",
"(4)",
"mirror.",
"Fh(r) =ET(1 - exp(-o↓8,)fk,",
"i=1",
"4.2. Our proposed dataset",
"where the superscript k indicates the sub-space that the ray",
"casts through. The k-th density o and feature fk corre-",
"As summarized in Sec. 4.1, there lacks a 360-degree",
"spond to the k-th sub-space. T = exp(-j-lojo,) and",
"dataset consisting of complex reflection and refraction to",
";=t,-ti-1 are similarlycomputed as in Eq.(2)",
"facilitate the related research. Therefore we collect a 360-",
"Then, [Fk) is decoded by two small MLPs, each with",
"degree dataset comprising 25 synthetic scenes and 7 real",
"just one hidden layer. The first is a Decoder MLP that takes",
"captured scenes.",
"[Fk] as input and outputs RGB vectors. The second is a",
"For our synthesized part shown in Fig. 6a, we use an",
"Gate MLP that takes [Fk] as input and outputs weights",
"open source software Blender [8] and design our scenes",
"that control the visibility of certain sub-space. Specifically,",
"with 3D models from BlenderKit, a community for shar-",
"we use:",
"ing 3D models.: As our dataset consists of complete scenes",
"instead of single objects, we fix the height of our camera",
"[Fk] an (Ch], {Fh] [wh],",
"(5)",
"position with the camera looking at the center of the scene",
"and moving the camera around a circle to render the whole",
"where D represents the Decoder MLP, and G represents",
"scene.",
"the Gate MLP. In the end, the MS module applies the soft-",
"For all our scenes, we uniformly sample 120 points along",
"max function to fwk as the color contribution of each sub-",
"the circle and randomly choose 100 images for the training",
"space to form the final render results:",
"set,1forthevalidation set, and 10forthetest set.Thecon-",
"structed dataset features a wide variety of scenes containing",
"1",
"K",
"C(r) :",
"exp(wk)ck",
"reflective and refractive objects. We include a variety of",
"(6)",
"Ziexp(wi) 1",
"Jk",
"complexity of light paths, controlled by the number and the",
"layout of the mirror(s) in the scene, where the number of",
"Eq.(6)needsnoadditional losstermscomparedwiththe",
"mirrors ranges from I up to tens of small pieces. Note that",
"vanilla NeRF method. As a result, the above light-weighted",
"even a scene in ourdataset with only one mirror is more",
"MS module is able to serve as an enhancement module onto",
"challenging than RFFR [25], as our camera moves from the",
"5"
] |
{
"arxiv_id": "2305.04268",
"language": "en",
"page_idx": 4
}
|
||
2003.11288-en-0
|
2003.11288-en-0.jpg
|
SCATTER: Selective Context Attentional Scene Text Recognizer
Ron Litman$^*$,Oron Anschel$^*$,Shahar Tsiper, Roee Litman, Shai Mazor and R. Manmatha
Amazon Web Services
$\{$litmanr, oronans, tsiper, rlit, smazor, manmatha}@amazon.com
$ଟ୍ରି$
$പ്പു$
## Abstract
Scene Text Recognition (STR), the task of recognizing
$\sum_{\begin{array}{cc}\text{ text against complex image backgrounds, is an active area}\\&of\text{ research. Current state-of-the-art (SOTA) methods still}\\&struggle\text{ to recognize text written in arbitrary shapes. In this}\end{array}}$
paper, we introduce a novel Selective Context ATtentiona SCATTER utilizes a stacked b
mediate supervision duri
together with contextual fea
$\begin{array}{ccc}&\text{layer. The second attention step, simil}\end{array}$ intra- sequence relationships. Experiments sh
<FigureHere>
Figure 1: The proposed SCATT chitecture. We introduce int selective-decoding to stabil coder (The circle represents coding is done using a selec features from the CNN backbo BiLSTM encoder, while employ
## $1. \textbf{Introduction}$
We address the task of reading text in natural scenes, commonly referred to as Scene Text Recognition (STR). Although STR has been active since the late 90's, only recently accuracy reached a level that enables commercial applications, this is mostly due to advances in deep neural networks research for computer vision tasks. Applications for STR include, among others, recognizing street signs in autonomous driving, company logos, assistive technology for the blind and translation apps in mixed reality.
Text in natural scenes is characterised by a large variety of backgrounds, and arbitrary imaging conditions that can lead to low contrast, blur, distortion, low resolution, uneven illumination and other phenomena and artifacts. In addition, the sheer magnitude of possible font types and sizes add another layer of difficulty that STR algorithms must overcome. Generally, recognizing scene text can be divided into
$^{*}$Authors contribute equally.
two main tasks - text detection and text recognition. Text detection is the task of identifying the regions in a natural image, that contain arbitrary shapes of text. Text recognition deals with the task of decoding a cropped image that contains one or more words into a digital string of its contents.
In this paper, we propose a method for text recognition; we assume the input is a cropped image of text taken from a natural image, and the output is the recognized text string within the cropped image. As categorized by previous works [1, 16], text images can be divided into two categories: Irregular text for arbitrarily shaped text (e.g. curved text), as seen in Fig. 1, and regular text for text with nearly horizontally aligned characters (examples are provided in the supplementary material).
|
[
[
0.1641845703125,
0.1337890625,
0.80615234375,
0.1337890625,
0.80615234375,
0.1483154296875,
0.1641845703125,
0.1483154296875
],
[
0.1356201171875,
0.18310546875,
0.8349609375,
0.18310546875,
0.8349609375,
0.1962890625,
0.1356201171875,
0.1962890625
],
[
0.396240234375,
0.2008056640625,
0.576171875,
0.2008056640625,
0.576171875,
0.21533203125,
0.396240234375,
0.21533203125
],
[
0.2181396484375,
0.2197265625,
0.7548828125,
0.2197265625,
0.7548828125,
0.2342529296875,
0.2181396484375,
0.2342529296875
],
[
0.0228729248046875,
0.25830078125,
0.060455322265625,
0.25830078125,
0.060455322265625,
0.367431640625,
0.0228729248046875,
0.367431640625
],
[
0.23779296875,
0.26953125,
0.313720703125,
0.26953125,
0.313720703125,
0.2841796875,
0.23779296875,
0.2841796875
],
[
0.66015625,
0.271484375,
0.68896484375,
0.271484375,
0.68896484375,
0.279052734375,
0.66015625,
0.279052734375
],
[
0.80126953125,
0.28271484375,
0.81396484375,
0.28271484375,
0.81396484375,
0.29345703125,
0.80126953125,
0.29345703125
],
[
0.65185546875,
0.2890625,
0.69775390625,
0.2890625,
0.69775390625,
0.300537109375,
0.65185546875,
0.300537109375
],
[
0.75244140625,
0.2890625,
0.7958984375,
0.2890625,
0.7958984375,
0.300537109375,
0.75244140625,
0.300537109375
],
[
0.650390625,
0.298583984375,
0.69921875,
0.298583984375,
0.69921875,
0.31005859375,
0.650390625,
0.31005859375
],
[
0.7509765625,
0.298583984375,
0.80126953125,
0.298583984375,
0.80126953125,
0.31005859375,
0.7509765625,
0.31005859375
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.098876953125,
0.302978515625,
0.468017578125,
0.302978515625,
0.468017578125,
0.31640625,
0.098876953125,
0.31640625
],
[
0.64892578125,
0.323974609375,
0.701171875,
0.323974609375,
0.701171875,
0.33837890625,
0.64892578125,
0.33837890625
],
[
0.7509765625,
0.322509765625,
0.79833984375,
0.324462890625,
0.79736328125,
0.337158203125,
0.7490234375,
0.335205078125
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.55810546875,
0.335205078125,
0.5888671875,
0.335205078125,
0.5888671875,
0.3466796875,
0.55810546875,
0.3466796875
],
[
0.080078125,
0.348388671875,
0.4697265625,
0.348388671875,
0.4697265625,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.07843017578125,
0.363037109375,
0.4697265625,
0.3623046875,
0.4697265625,
0.376953125,
0.07843017578125,
0.37744140625
],
[
0.54833984375,
0.35986328125,
0.59716796875,
0.35986328125,
0.59716796875,
0.371826171875,
0.54833984375,
0.371826171875
],
[
0.81396484375,
0.369384765625,
0.85302734375,
0.369384765625,
0.85302734375,
0.37255859375,
0.81396484375,
0.37255859375
],
[
0.5498046875,
0.37060546875,
0.59619140625,
0.37060546875,
0.59619140625,
0.380615234375,
0.5498046875,
0.380615234375
],
[
0.650390625,
0.37060546875,
0.69921875,
0.37060546875,
0.69921875,
0.380615234375,
0.650390625,
0.380615234375
],
[
0.7509765625,
0.37060546875,
0.79833984375,
0.37060546875,
0.79833984375,
0.380615234375,
0.7509765625,
0.380615234375
],
[
0.02777099609375,
0.376953125,
0.0555419921875,
0.376953125,
0.0555419921875,
0.39453125,
0.02777099609375,
0.39453125
],
[
0.080078125,
0.37744140625,
0.46728515625,
0.37744140625,
0.46728515625,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.5498046875,
0.389404296875,
0.59619140625,
0.389404296875,
0.59619140625,
0.40087890625,
0.5498046875,
0.40087890625
],
[
0.080078125,
0.392578125,
0.465576171875,
0.392578125,
0.465576171875,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.02777099609375,
0.3642578125,
0.053924560546875,
0.3642578125,
0.053924560546875,
0.38330078125,
0.02777099609375,
0.38330078125
],
[
0.65185546875,
0.393310546875,
0.69775390625,
0.393310546875,
0.69775390625,
0.404052734375,
0.65185546875,
0.404052734375
],
[
0.75244140625,
0.393310546875,
0.80126953125,
0.393310546875,
0.80126953125,
0.40478515625,
0.75244140625,
0.40478515625
],
[
0.810546875,
0.3984375,
0.859375,
0.3984375,
0.859375,
0.4091796875,
0.810546875,
0.4091796875
],
[
0.82421875,
0.39453125,
0.85302734375,
0.39453125,
0.85302734375,
0.397705078125,
0.82421875,
0.397705078125
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.55615234375,
0.415283203125,
0.58837890625,
0.415283203125,
0.58837890625,
0.423583984375,
0.55615234375,
0.423583984375
],
[
0.65283203125,
0.4140625,
0.6953125,
0.4140625,
0.6953125,
0.4248046875,
0.65283203125,
0.4248046875
],
[
0.76318359375,
0.415283203125,
0.7890625,
0.415283203125,
0.7890625,
0.423583984375,
0.76318359375,
0.423583984375
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.07843017578125,
0.4375,
0.4697265625,
0.438232421875,
0.4697265625,
0.45263671875,
0.07843017578125,
0.451904296875
],
[
0.5537109375,
0.43505859375,
0.5908203125,
0.439453125,
0.58837890625,
0.448974609375,
0.55126953125,
0.44580078125
],
[
0.65283203125,
0.4375,
0.69287109375,
0.4375,
0.69287109375,
0.447509765625,
0.65283203125,
0.447509765625
],
[
0.75732421875,
0.4375,
0.7958984375,
0.4375,
0.7958984375,
0.447509765625,
0.75732421875,
0.447509765625
],
[
0.64453125,
0.447509765625,
0.67236328125,
0.447509765625,
0.67236328125,
0.455078125,
0.64453125,
0.455078125
],
[
0.740234375,
0.447509765625,
0.8095703125,
0.447509765625,
0.8095703125,
0.458251953125,
0.740234375,
0.458251953125
],
[
0.5458984375,
0.448974609375,
0.5986328125,
0.448974609375,
0.5986328125,
0.455810546875,
0.5458984375,
0.455810546875
],
[
0.0318603515625,
0.448974609375,
0.058013916015625,
0.448974609375,
0.058013916015625,
0.48095703125,
0.0318603515625,
0.48095703125
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.67724609375,
0.449462890625,
0.6953125,
0.449462890625,
0.6953125,
0.453857421875,
0.67724609375,
0.453857421875
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.56396484375,
0.46728515625,
0.5888671875,
0.4716796875,
0.58642578125,
0.481689453125,
0.560546875,
0.4765625
],
[
0.67333984375,
0.474853515625,
0.6845703125,
0.474853515625,
0.6845703125,
0.479248046875,
0.67333984375,
0.479248046875
],
[
0.77392578125,
0.472900390625,
0.7890625,
0.472900390625,
0.7890625,
0.479736328125,
0.77392578125,
0.479736328125
],
[
0.080078125,
0.484130859375,
0.4697265625,
0.484130859375,
0.4697265625,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.0236968994140625,
0.50048828125,
0.05963134765625,
0.50048828125,
0.05963134765625,
0.5673828125,
0.0236968994140625,
0.5673828125
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.50732421875,
0.50048828125,
0.708984375,
0.50048828125,
0.708984375,
0.51123046875,
0.50732421875,
0.51123046875
],
[
0.73876953125,
0.50048828125,
0.8095703125,
0.50048828125,
0.8095703125,
0.51123046875,
0.73876953125,
0.51123046875
],
[
0.53515625,
0.5087890625,
0.60546875,
0.5087890625,
0.60546875,
0.51953125,
0.53515625,
0.51953125
],
[
0.64892578125,
0.51220703125,
0.70166015625,
0.51220703125,
0.70166015625,
0.5234375,
0.64892578125,
0.5234375
],
[
0.75244140625,
0.51220703125,
0.7998046875,
0.51220703125,
0.7998046875,
0.5234375,
0.75244140625,
0.5234375
],
[
0.080078125,
0.52880859375,
0.46728515625,
0.52880859375,
0.46728515625,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.080078125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.080078125,
0.54443359375,
0.46728515625,
0.54443359375,
0.46728515625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50341796875,
0.53857421875,
0.8896484375,
0.53857421875,
0.8896484375,
0.548828125,
0.50341796875,
0.548828125
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.0269622802734375,
0.5595703125,
0.052276611328125,
0.5595703125,
0.052276611328125,
0.5888671875,
0.0269622802734375,
0.5888671875
],
[
0.043304443359375,
0.56103515625,
0.049835205078125,
0.564453125,
0.04412841796875,
0.5693359375,
0.037567138671875,
0.56640625
],
[
0.07843017578125,
0.55810546875,
0.43310546875,
0.5595703125,
0.43310546875,
0.57373046875,
0.07843017578125,
0.57275390625
],
[
0.50341796875,
0.56640625,
0.8896484375,
0.56640625,
0.8896484375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.041656494140625,
0.57568359375,
0.051483154296875,
0.57568359375,
0.051483154296875,
0.58349609375,
0.041656494140625,
0.58349609375
],
[
0.0196075439453125,
0.583984375,
0.058013916015625,
0.583984375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.041656494140625,
0.583984375,
0.0482177734375,
0.583984375,
0.0482177734375,
0.59130859375,
0.041656494140625,
0.59130859375
],
[
0.50244140625,
0.5791015625,
0.890625,
0.5791015625,
0.890625,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.08087158203125,
0.60693359375,
0.20751953125,
0.60693359375,
0.20751953125,
0.62060546875,
0.08087158203125,
0.62060546875
],
[
0.50244140625,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.619140625,
0.50244140625,
0.619140625
],
[
0.50244140625,
0.62060546875,
0.83837890625,
0.62060546875,
0.83837890625,
0.6337890625,
0.50244140625,
0.6337890625
],
[
0.1004638671875,
0.63330078125,
0.468017578125,
0.63330078125,
0.468017578125,
0.646484375,
0.1004638671875,
0.646484375
],
[
0.08087158203125,
0.6494140625,
0.465576171875,
0.6494140625,
0.465576171875,
0.66015625,
0.08087158203125,
0.66015625
],
[
0.08087158203125,
0.66357421875,
0.46728515625,
0.66357421875,
0.46728515625,
0.6767578125,
0.08087158203125,
0.6767578125
],
[
0.50244140625,
0.67041015625,
0.89208984375,
0.67041015625,
0.89208984375,
0.6845703125,
0.50244140625,
0.6845703125
],
[
0.080078125,
0.67822265625,
0.46728515625,
0.67822265625,
0.46728515625,
0.69140625,
0.080078125,
0.69140625
],
[
0.50244140625,
0.68603515625,
0.890625,
0.68603515625,
0.890625,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.08087158203125,
0.70849609375,
0.468017578125,
0.70849609375,
0.468017578125,
0.72265625,
0.08087158203125,
0.72265625
],
[
0.50244140625,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.080078125,
0.7236328125,
0.468017578125,
0.7236328125,
0.468017578125,
0.7373046875,
0.080078125,
0.7373046875
],
[
0.5009765625,
0.73095703125,
0.890625,
0.73095703125,
0.890625,
0.744140625,
0.5009765625,
0.744140625
],
[
0.080078125,
0.73779296875,
0.4697265625,
0.73779296875,
0.4697265625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50244140625,
0.74755859375,
0.5400390625,
0.74755859375,
0.5400390625,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.75390625,
0.386474609375,
0.75390625,
0.386474609375,
0.767578125,
0.080078125,
0.767578125
],
[
0.1004638671875,
0.76953125,
0.46728515625,
0.76953125,
0.46728515625,
0.78271484375,
0.1004638671875,
0.78271484375
],
[
0.521484375,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.521484375,
0.77978515625
],
[
0.080078125,
0.78271484375,
0.468017578125,
0.7841796875,
0.468017578125,
0.798828125,
0.080078125,
0.79736328125
],
[
0.5009765625,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.5009765625,
0.79541015625
],
[
0.08087158203125,
0.7998046875,
0.4697265625,
0.7998046875,
0.4697265625,
0.81298828125,
0.08087158203125,
0.81298828125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.07843017578125,
0.81298828125,
0.468017578125,
0.81494140625,
0.468017578125,
0.8291015625,
0.07843017578125,
0.8271484375
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.830078125,
0.4697265625,
0.830078125,
0.4697265625,
0.8447265625,
0.080078125,
0.8447265625
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84521484375,
0.4697265625,
0.84521484375,
0.4697265625,
0.859375,
0.080078125,
0.859375
],
[
0.5009765625,
0.84130859375,
0.89208984375,
0.84033203125,
0.89208984375,
0.85498046875,
0.5009765625,
0.85595703125
],
[
0.080078125,
0.861328125,
0.4697265625,
0.861328125,
0.4697265625,
0.875,
0.080078125,
0.875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.5009765625,
0.87060546875,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.09478759765625,
0.88623046875,
0.2467041015625,
0.88720703125,
0.2467041015625,
0.9013671875,
0.09478759765625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.69287109375,
0.88720703125,
0.69287109375,
0.9013671875,
0.50244140625,
0.9013671875
]
] |
[
"SCATTER: Selective Context Attentional Scene Text Recognizer",
"Ron Litman*, Oron Anschel*, Shahar Tsiper, Roee Litman. Shai Mazor and R. Manmatha",
"Amazon Web Services",
"(litmanr, oronans, tsiper,rlit, smazor,manmatha)@amazon.com",
"",
"Abstract",
"*BMW",
"O",
"Selective",
"Selective",
"Decoder",
"Decoder",
"text against complex image backgrounds, is an active area",
"Scene Text Recognition (STR), the task of recognizing",
"BiLSTM",
"BiLSTM",
"of research. Current state-of-the-art (SOTA) methods still",
"\"BOL\"",
"struggle to recognize text written in arbitrary shapes. In this",
"paper, we introduce a novel architecture for STR, named",
"Attention",
"Soieture",
"Decoder",
"BiLSTM",
"BiLSTM",
"2",
"Selective Context ATtentional Text Recognizer (SCATTER)",
"BiLSTM",
"SCATTER utilizes a stacked block architecture with inter",
"5",
"BiLSTM",
"BiLSTM",
"Decoder",
"ICLU",
"LAD",
"mediate supervision during training, that paves the way to",
"CNN",
"CNN",
"CNN",
"successfully train a deep BiLSTM encoder, thus improving",
"the encoding of contextual dependencies. Decoding is done",
"Image",
"Image",
"Image",
"Tran",
"Transform",
"Transform",
"SO",
"using a two-step ID attention mechanism. The first atten-",
"iorm",
"tion step re-weights visual features from a CNN backbone",
"M",
"",
"",
"together with contextual features computed by a BiLSTM",
"1887",
"layer The second attention step, similar to previous pa-",
"(a) General text recognition model (b) SCATTER (Ours)",
"(c) SCATTER (Ours)",
"(Beak et al., 2019)",
"Inference",
"Training",
"intra-sequence relationships. Experiments show that the",
"pers, treats the features as a sequence and attends to the",
"proposed approach surpasses SOTA performance on irreg-",
"Figure l:TheproposedSCATTERtrainingandinferencear-",
"chitecture. We introduce intermediate supervision combined with",
"",
"",
"ular text recognition benchmarks by 3.7% on average.",
"selective-decoding to stabilize the training of a deepBiLSTM en-",
"-",
"C00Z:AIX",
"",
"coder (The circle represents where a loss function is applied). De-",
"coding is done using a selective-decoder that operates on visual",
"1. Introduction",
"features from the CNN backbone and contextual features from the",
"BiLSTM encoder, while employing a two-step attention.",
"We address the task of reading text in natural scenes,",
"commonlyreferred to as Scene Text Recognition (STR)",
"Although STR has been active since the late 90's, only re-",
"two main tasks - text detection and text recognition. Text",
"cently accuracy reached a level that enables commercial ap-",
"detection is the task of identifying the regions in a natural",
"plications, this is mostly due to advances in deep neural net-",
"image, that contain arbitrary shapes of text. Text recogni-",
"works research for computer vision tasks. Applications for",
"tion deals with the task of decoding a cropped image that",
"STR include, among others, recognizing street signs in au-",
"contains one or more words into a digital string of its con-",
"tonomous driving, company logos, assistive technology for",
"tents.",
"the blind and translation apps in mixed reality.",
"Text in natural scenes is characterised by a large variety",
"In this paper, we propose a method for text recogni-",
"of backgrounds, and arbitrary imaging conditions that can",
"tion; we assume the input is a cropped image of text taken",
"lead to low contrast, blur, distortion, low resolution, uneven",
"from a natural image, and the output is the recognized text",
"illumination and other phenomena and artifacts.In addition,",
"string within the cropped image. As categorized by previ-",
"the sheer magnitude of possible font types and sizes add",
"ous works [1, 16], text images can be divided into two cate-",
"another layer of difficulty that STR algorithms must over-",
"gories: Irregular text for arbitrarily shaped text (e.g. curved",
"come. Generally, recognizing scene text can be divided into",
"text), as seen in Fig. 1, and regular text for text with nearly",
"horizontally aligned characters (examples are provided in",
"*Authors contribute equally.",
"the supplementary material)."
] |
{
"arxiv_id": "2003.11288",
"language": "en",
"page_idx": 0
}
|
||
2003.11288-en-1
|
2003.11288-en-1.jpg
|
<FigureHere>
Figure 2: Average test accuracy (IIT5K, SVT, IC03, IC13,IC15, SVTP, CUTE) at intermediate decoding steps, compared across different network depths used in training. Given a computation budget, improved results can be obtained by first training a deeper network, and then running only the first decoder(s) during inference.
Traditional text recognition methods [37,38,42] detect and recognize text character by character, however, these methods have an inherent limitation- they do not utilize sequential modeling and contextual dependencies between characters.
Modern methods treat STR as a sequence prediction problem. This technique alleviates the need for character-level annotations (per-character bounding box) while achieving superior accuracy. The majority of these sequence-based methods rely on Connectionist Temporal Classification (CTC)[31,7], or attention-based mechanisms[33,16]. Recently, Baek et al.[1]proposed a modular four-step STR framework, where the individual components are interchangeable allowing for different algorithms. This modular framework, along with its best performing component configuration, is depicted in Fig. l (a). In this work, we build upon this framework and extend it.
While accurately recognizing regular scene text remains an open problem, recent irregular STR benchmarks (e.g., ICD15, SVTP) have shifted research focus to the problem of recognizing text in arbitrary shapes. For instance, Sheng et al.[30] adopted the Transformer[35] model for STR, leveraging the transformers ability to capture longrange contextual dependencies. The authors in [16] passed the visual features from the CNN backbone through a 2D attention module down to their decoder. Mask TextSpotter [17] unified the detection and the recognition tasks with a shared backbone architecture. For the recognition stage, two types of prediction branches are used, and the final prediction is selected based on the output of the more confident branch. The first branch uses semantic segmentation of characters, and requires additional character-level annotations. The second branch employs a 2D spatial attentiondecoder.
Most of the aforementioned STR methods perform a sequential modeling step using a recursive neural network (RNN) or other sequential modeling layers (e.g., multihead attention [30]), usually in the encoder and/or the decoder. This step is performed to convert the visual feature map into a contextual feature map, which better captures long-term dependencies. In this work, we propose using a stacked block architecture for repeated feature processing, a concept similar to that used in other computer-vision tasks such as in [40] and later in [27,26]. The authors above showed that repeated processing used in conjunction with intermediate supervision could be used to increasingly refine predictions.
In this paper, we propose the Selective Context ATtentional TExt Recognizer (SCATTER) architecture. Our method, as depicted in Fig. 1, utilize a stacked block architecture for repetitive processing with intermediate supervision in training, and a novel selective-decoder. The selective-decoder receives features from two different lay-
ers of the network, namely, visual features from a CNN backbone and contextual features computed by a BiLSTM layer, while using a two-step 1D attention mechanism. Figure 2 shows the accuracy levels computed at the intermediate auxiliary decoders, for different stacking arrangements, thus demonstrating the increase in performance as additional blocks are added in succession. Interestingly, training with additional blocks in sequence leads to an improvement in the accuracy of the intermediate decoders as well (compared to training with a shallower stacking arrangement).
This paper presents two main contributions:
1. We propose a repetitive processing architecture for
text recognition, trained with intermediate selective decoders as supervision. Using this architecture we train a deep BiLSTM encoder, leading to SOTA results on irregular text.
2. A selective attention decoder, that simultaneously de-
codes both visual and contextual features by employing a two-step attention mechanism. The first attention step figures out which visual and contextual features to attend to. The second step treats the features as a sequence and attends the intra-sequence relations.
## $2. \textbf{Related Work}$
STR has attracted considerable attention over the past few years [34, 36, 3, 20]. Comprehensive surveys for scene text detection and recognition may be found in [43,1,21]. As mentioned above, STR may be divided into two categories: regular and irregular texts (further examples are provided in the supplementary material). Earlier papers[37, 38, 42], focused on regular text and used a bottom-up approach, which involved segmenting individual characters with a sliding window, and then recognizing the characters using hand-crafted features. A notable issue with the
|
[
[
0.1004638671875,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.1004638671875,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.49267578125,
0.106689453125,
0.5986328125,
0.0997314453125,
0.60302734375,
0.1395263671875,
0.498291015625,
0.146484375
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.07843017578125,
0.136962890625,
0.468017578125,
0.1363525390625,
0.468017578125,
0.15087890625,
0.07843017578125,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.156005859375,
0.1546630859375,
0.156005859375,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50634765625,
0.1546630859375,
0.67333984375,
0.1483154296875,
0.67626953125,
0.1875,
0.5087890625,
0.1932373046875
],
[
0.09722900390625,
0.169189453125,
0.46728515625,
0.1673583984375,
0.46728515625,
0.1817626953125,
0.09722900390625,
0.184326171875
],
[
0.5146484375,
0.17041015625,
0.521484375,
0.17041015625,
0.521484375,
0.175537109375,
0.5146484375,
0.175537109375
],
[
0.080078125,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1988525390625,
0.080078125,
0.1988525390625
],
[
0.080078125,
0.2008056640625,
0.468017578125,
0.2008056640625,
0.468017578125,
0.2139892578125,
0.080078125,
0.2139892578125
],
[
0.50732421875,
0.20703125,
0.7490234375,
0.20263671875,
0.7509765625,
0.2392578125,
0.5087890625,
0.24365234375
],
[
0.07843017578125,
0.21533203125,
0.46728515625,
0.2139892578125,
0.46728515625,
0.2291259765625,
0.07843017578125,
0.2298583984375
],
[
0.07843017578125,
0.2310791015625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.07843017578125,
0.2442626953125
],
[
0.080078125,
0.24560546875,
0.468017578125,
0.24560546875,
0.468017578125,
0.2587890625,
0.080078125,
0.2587890625
],
[
0.080078125,
0.261474609375,
0.4697265625,
0.261474609375,
0.4697265625,
0.274658203125,
0.080078125,
0.274658203125
],
[
0.5048828125,
0.25830078125,
0.8896484375,
0.25830078125,
0.8896484375,
0.268310546875,
0.5048828125,
0.268310546875
],
[
0.50244140625,
0.26953125,
0.890625,
0.270751953125,
0.890625,
0.2841796875,
0.50244140625,
0.28271484375
],
[
0.080078125,
0.27587890625,
0.4697265625,
0.27587890625,
0.4697265625,
0.2890625,
0.080078125,
0.2890625
],
[
0.50244140625,
0.285400390625,
0.89208984375,
0.285400390625,
0.89208984375,
0.298583984375,
0.50244140625,
0.298583984375
],
[
0.08087158203125,
0.292236328125,
0.468017578125,
0.292236328125,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.5,
0.29736328125,
0.89208984375,
0.298583984375,
0.89208984375,
0.3125,
0.5,
0.311279296875
],
[
0.080078125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.50244140625,
0.313232421875,
0.890625,
0.313232421875,
0.890625,
0.326904296875,
0.50244140625,
0.326904296875
],
[
0.07843017578125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.335205078125,
0.07843017578125,
0.335205078125
],
[
0.5009765625,
0.32763671875,
0.5361328125,
0.32763671875,
0.5361328125,
0.339111328125,
0.5009765625,
0.339111328125
],
[
0.080078125,
0.3359375,
0.370849609375,
0.3359375,
0.370849609375,
0.349853515625,
0.080078125,
0.349853515625
],
[
0.1004638671875,
0.354248046875,
0.4697265625,
0.354248046875,
0.4697265625,
0.367431640625,
0.1004638671875,
0.367431640625
],
[
0.0767822265625,
0.365478515625,
0.4697265625,
0.367431640625,
0.4697265625,
0.385009765625,
0.0767822265625,
0.38330078125
],
[
0.50244140625,
0.369384765625,
0.8896484375,
0.369384765625,
0.8896484375,
0.38330078125,
0.50244140625,
0.38330078125
],
[
0.080078125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.397705078125,
0.080078125,
0.397705078125
],
[
0.50341796875,
0.386962890625,
0.8896484375,
0.386962890625,
0.8896484375,
0.397705078125,
0.50341796875,
0.397705078125
],
[
0.080078125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.50244140625,
0.401611328125,
0.88818359375,
0.401611328125,
0.88818359375,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.080078125,
0.4140625,
0.4697265625,
0.4140625,
0.4697265625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.50341796875,
0.416748046875,
0.890625,
0.416748046875,
0.890625,
0.429931640625,
0.50341796875,
0.429931640625
],
[
0.080078125,
0.4287109375,
0.46728515625,
0.4287109375,
0.46728515625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.5009765625,
0.43115234375,
0.89208984375,
0.431884765625,
0.89208984375,
0.44580078125,
0.5009765625,
0.4443359375
],
[
0.08087158203125,
0.44580078125,
0.468017578125,
0.44580078125,
0.468017578125,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.50244140625,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.50244140625,
0.460205078125
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.5009765625,
0.460205078125,
0.89208984375,
0.4609375,
0.89208984375,
0.47607421875,
0.5009765625,
0.474853515625
],
[
0.08087158203125,
0.47607421875,
0.46728515625,
0.47607421875,
0.46728515625,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.50244140625,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.08087158203125,
0.4912109375,
0.46728515625,
0.4912109375,
0.46728515625,
0.501953125,
0.08087158203125,
0.501953125
],
[
0.50244140625,
0.4912109375,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.50244140625,
0.5048828125
],
[
0.07843017578125,
0.50244140625,
0.4697265625,
0.5048828125,
0.4697265625,
0.51953125,
0.07843017578125,
0.517578125
],
[
0.50244140625,
0.5068359375,
0.8759765625,
0.5068359375,
0.8759765625,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.521484375,
0.46728515625,
0.521484375,
0.46728515625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.51953125,
0.521484375,
0.81396484375,
0.52294921875,
0.81396484375,
0.537109375,
0.51953125,
0.5361328125
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.5146484375,
0.53857421875,
0.89208984375,
0.5390625,
0.89208984375,
0.5537109375,
0.5146484375,
0.55322265625
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.080078125,
0.564453125
],
[
0.53369140625,
0.5546875,
0.8935546875,
0.5537109375,
0.8935546875,
0.568359375,
0.53369140625,
0.5693359375
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.53759765625,
0.5712890625,
0.890625,
0.5712890625,
0.890625,
0.58203125,
0.53759765625,
0.58203125
],
[
0.080078125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.53515625,
0.58544921875,
0.8935546875,
0.58544921875,
0.8935546875,
0.5986328125,
0.53515625,
0.5986328125
],
[
0.08087158203125,
0.5966796875,
0.1380615234375,
0.5966796875,
0.1380615234375,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.5361328125,
0.6015625,
0.646484375,
0.6015625,
0.646484375,
0.6123046875,
0.5361328125,
0.6123046875
],
[
0.09722900390625,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.09722900390625,
0.62548828125
],
[
0.51318359375,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.51318359375,
0.62890625
],
[
0.080078125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.6416015625,
0.080078125,
0.6416015625
],
[
0.53515625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.53515625,
0.64453125
],
[
0.08087158203125,
0.64306640625,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.5361328125,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.5361328125,
0.65771484375
],
[
0.080078125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.080078125,
0.671875
],
[
0.53759765625,
0.662109375,
0.890625,
0.662109375,
0.890625,
0.67236328125,
0.53759765625,
0.67236328125
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68505859375,
0.08087158203125,
0.68505859375
],
[
0.53515625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.53515625,
0.689453125
],
[
0.080078125,
0.68798828125,
0.471435546875,
0.68798828125,
0.471435546875,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.53515625,
0.68994140625,
0.86279296875,
0.68994140625,
0.86279296875,
0.7041015625,
0.53515625,
0.7041015625
],
[
0.07843017578125,
0.7021484375,
0.4697265625,
0.70263671875,
0.4697265625,
0.71728515625,
0.07843017578125,
0.71630859375
],
[
0.080078125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50341796875,
0.7236328125,
0.640625,
0.7236328125,
0.640625,
0.73779296875,
0.50341796875,
0.73779296875
],
[
0.080078125,
0.73291015625,
0.4697265625,
0.73291015625,
0.4697265625,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.080078125,
0.74853515625,
0.471435546875,
0.74853515625,
0.471435546875,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.521484375,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.521484375,
0.76318359375
],
[
0.080078125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.77880859375,
0.468017578125,
0.77880859375,
0.468017578125,
0.79248046875,
0.080078125,
0.79248046875
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.79541015625,
0.188720703125,
0.79541015625,
0.188720703125,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.09722900390625,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.09722900390625,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
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
],
[
0.5009765625,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.5009765625,
0.87060546875
],
[
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,
0.468017578125,
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
]
] |
[
"Traditional text recognition methods [37, 38, 42] detect",
"and recognize text character bycharacter,however.these",
"日口",
"methods have an inherent limitation - they do not utilize",
"sequential modeling and contextual dependencies between",
"characters.",
"国0000",
"Modern methods treat STR as a sequence predic-",
"",
"tion problem. This technique alleviates the need for",
"character-level annotations (per-character bounding box)",
"0000000",
"while achieving superior accuracy. The majority of these",
"sequence-based methods rely on Connectionist Temporal",
"Classification (CTC) [31, 7], or attention-based mecha-",
"nisms [33, 16]. Recently, Baek et al. [1] proposed a modular",
"Figure2:Averagetestaccuracy(IIT5K.SVT,IC03.IC13.IC15.",
"SVTP, CUTE) at intermediate decoding steps, compared across",
"four-step STR framework, where the individual components",
"different network depths used in training. Given a computation",
"areinterchangeable allowingfor different algorithms.This",
"budget, improved results can be obtained by first training a deeper",
"modular framework, along with its best performing compo-",
"network, and then running only the first decoder(s) during infer-",
"nent configuration, is depicted in Fig. 1 (a). In this work,",
"ence.",
"we build upon this framework and extend it.",
"While accurately recognizing regular scene text remains",
"an open problem, recent irregular STR benchmarks (e.g.,",
"ers of the network, namely, visual features from a CNN",
"ICD15, SVTP) have shifted research focus to the prob-",
"backboneandcontextualfeaturescomputed byaBiLSTM",
"lem of recognizing text in arbitrary shapes.For instance,",
"layer, while using a two-step ID attention mechanism. Fig",
"Sheng et al. [30] adopted the Transformer [35] model for",
"ure 2 shows the accuracy levels computed at the intermedi-",
"STR, leveraging the transformers ability to capture long-",
"ate auxiliary decoders,for different stacking arrangements,",
"rangecontextualdependencies.Theauthorsin[16lpassed",
"thus demonstrating the increase in performance as addi-",
"the visual features from the CNN backbone through a 2D",
"tional blocks are added in succession. Interestingly, training",
"attentionmoduledowntotheirdecoder:MaskTextSpot-",
"with additional blocks in sequence leads to an improvement",
"ter[17lunifiedthedetectionandtherecognitiontaskswith",
"in the accuracy of the intermediate decoders as well (com-",
"a shared backbone architecture. For the recognition stage,",
"pared to training with a shallower stacking arrangement).",
"twotypes ofpredictionbranchesareused,andthefinal pre-",
"This paper presents two main contributions:",
"diction is selected based on the output of the more confi-",
"1. We propose a repetitive processing architecture for",
"dent branch. The first branch uses semantic segmentation",
"text recognition, trained with intermediate selective",
"of characters, and requires additional character-level anno-",
"decoders as supervision.Using this architecturewe",
"tations. The second branch employs a 2D spatial attention-",
"train a deep BiLSTM encoder, leading to SOTA results",
"decoder.",
"on irregulartext.",
"Most of the aforementioned STR methods perform a se-",
"2. A selective attention decoder, that simultaneously de-",
"quential modeling step using a recursive neural network",
"codes both visual and contextual features by employ-",
"(RNN) or other sequential modeling layers (e.g., multi-",
"ing a two-step attention mechanism. The first attention",
"head attention [30]), usually in the encoder and/or the de-",
"step figures out which visual and contextual features",
"coder.This step is performed to convert the visual feature",
"to attend to. The second step treats the features as a",
"map into a contextual feature map, which better captures",
"sequence and attends the intra-sequence relations.",
"long-term dependencies.In this work, we propose using a",
"stacked block architecture for repeated feature processing,a",
"2.Related Work",
"concept similar to that used in other computer-vision tasks",
"such as in [40] and later in [27, 26]. The authors above",
"STR has attracted considerable attention over the past",
"showed that repeated processing used in conjunction with",
"few years [34, 36, 3, 20]. Comprehensive surveys for scene",
"intermediate supervision could be used to increasingly re-",
"text detectionandrecognitionmaybefound in[43,l,21].",
"finepredictions.",
"As mentioned above, STR may be divided into two cate-",
"In this paper, we propose the Selective Context",
"gories: regular and irregular texts (further examples are pro-",
"ATtentional TExt Recognizer (SCATTER) architecture.",
"vided in the supplementary material). Earlier papers [37,",
"Our method, as depicted in Fig. 1, utilize a stacked block",
"38, 42], focused on regular text and used a bottom-up ap-",
"architecture forrepetitive processing with intermediate su-",
"proach, which involved segmenting individual characters",
"pervision in training, and a novel selective-decoder. The",
"with a sliding window, and then recognizing the charac-",
"selective-decoder receives features from two different lay-",
"ters using hand-crafted features. A notable issue with the"
] |
{
"arxiv_id": "2003.11288",
"language": "en",
"page_idx": 1
}
|
||
2003.11288-en-2
|
2003.11288-en-2.jpg
|
bottom-up approaches above, is that they struggle to use contextual information; instead they rely on accurate character classifiers. Shi et al. 2015[31] and He et al.[11] considered words as sequences of varying lengths, and employed RNNs to model the sequences without explicit character separation. Shi et al. 2016[32] presented a successful end-to-end trainable architecture using the sequence approach, without relying on character level annotations. Their solution employed a BiLSTM layer to extract the sequential feature vectors from the input feature maps, these vectors are then fed into an attention-Gated Recurrent Unit (GRU) module for decoding.
The methods mentioned above introduced significant improvements in STR accuracy on public benchmarks. Therefore, recent work has shifted focus to the more challenging problem of recognizing irregularly shaped text, hence promoting new lines of research. Topics such as input rectification, character-level segmentation, 2D attentional feature maps and self attention have emerged, pushing the envelope on irregular STR. Shi et al. 2018 [33] rectified oriented or curved text based on a Spatial Transformer Network (STN). Liu et al. 2018 [19] introduced a Character-Aware Neural Network (Char-Net) to detect and rectify individual characters. A combination of a CTC-Attention mechanism within an encoder-decoder framework, that was used for speech recognition tasks, was used for STR in [45], showing the benefits of joint CTC-Attention learning. The authors in[7] proposed two supervision branches to tackle explicit and implicit semantic information. In [4] a gate was inserted to the recurrent decoder, for controlling the transmissionweight of the previous embedded vector, demonstrating that context is not always needed for decoding. The authors of Mask TextSpotter[17] unified text detection and text recognition in an end-to-end fashion. For recognition they used two separate branches, a branch that uses visual (local) features and a branch which utilizes contextual information in the form of 2D attention.
More recent approaches have proposed leveraging various attention mechanisms for improved results. Li et al.[16] combined both visual and contextual features while utilizing a 2D attention within the encoder-decoder. Other researchers borrowed ideas from the Natural Language Processing (NLP) domain and adopted a transformer-based architecture[35]. One of them is Sheng et al.[30], that used a self-attention mechanism for both the encoder and the de- $coder.$
Our method differs from above approaches, by being the first to utilize a stacked block architecture for text recognition. Namely, we show that repetitive processing for text recognition, trained with intermediate selective decoders as
supervision (similar to [40,27,26]), increasingly refines
text predictions.
# 3. Methodology
As presented in Fig. 3, our proposed architecture consists
of four main components:
1. Transformation: the input text image is normalized
using a Spatial Transformer Network (STN)[13].
2. Feature Extraction: maps the input image to a feature
map representation while using a text attention mod-
$ule[7].$
3. Visual Feature Refinement: provides direct super-
vision for each column in the visual features. This part refines the representation in each of the feature columns, by classifying them into individual symbols
4. Selective-Contextual Refinement Block: Each block
consists of a two-layer BiLSTM encoder that outputs contextual features. The contextual features are concatenated to the visual features computed by the CNN backbone. This concatenated feature map is then fed into the selective-decoder, which employs a two-step 1D attention mechanism, as illustrated in Fig. 4
In this section we describe the training architecture of SCATTER , while addressing the differences between training and inference.
## $3. 1. \textbf{ Transformation}$
The transformation step operates on the cropped text image $X$, and transforms it into a normalized image $X^\prime.$ We use a Thin Plate Spline (TPS) transformation, a variant of the STN, as used in [1]. TPS employs a smooth spline interpolation between a set of fducial points. Specifically, it detects a pre-defined number of fiducial points at the top and bottom of the text region, and normalizes the predicted region to a constant predefined size.
## $3. 2. \textbf{ Feature Extraction}$
In this step a convolutional neural network (CNN) extracts features from the input image. We use a 29-layer ResNet as the CNN's backbone, as used in [5]. The output of the feature encoder is 512 channels by $N$ columns. Specifically, the feature encoder gets an input image $X^\prime$ and outputs a feature map $F=[f_1,f_2,...,f_N].$ Following the feature map extraction, we use a text attention module, similar to[7]. The attentional feature map can be regarded as a visual feature sequence of length $N$, denoted as $\tilde{V}=[v_{1},v_{2},...,v_{N}],\hat{\text{where each column represents a frame}}$ in the sequence.
## 3.3. Visual Feature Refinement
Here, the visual feature sequence $V$ is used for intermediate decoding. This intermediate supervision is aimed at refining the character embedding (representations) in each of the columns of $V$, and is done using CTC based decoding. We feed $V$ through a fully connected layer that outputs
|
[
[
0.080078125,
0.0921630859375,
0.47216796875,
0.0921630859375,
0.47216796875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.0889892578125,
0.63232421875,
0.091552734375,
0.63232421875,
0.1060791015625,
0.5009765625,
0.10418701171875
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.521484375,
0.117431640625,
0.89208984375,
0.117431640625,
0.89208984375,
0.1307373046875,
0.521484375,
0.1307373046875
],
[
0.50244140625,
0.1319580078125,
0.6748046875,
0.1319580078125,
0.6748046875,
0.1458740234375,
0.50244140625,
0.1458740234375
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.515625,
0.147705078125,
0.89208984375,
0.147705078125,
0.89208984375,
0.1610107421875,
0.515625,
0.1610107421875
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.53369140625,
0.1622314453125,
0.861328125,
0.1622314453125,
0.861328125,
0.1767578125,
0.53369140625,
0.1767578125
],
[
0.51318359375,
0.175537109375,
0.89208984375,
0.1767578125,
0.89208984375,
0.1912841796875,
0.51318359375,
0.1900634765625
],
[
0.080078125,
0.1817626953125,
0.46728515625,
0.1817626953125,
0.46728515625,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.080078125,
0.1988525390625,
0.46728515625,
0.1988525390625,
0.46728515625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.53369140625,
0.1932373046875,
0.890625,
0.1932373046875,
0.890625,
0.20703125,
0.53369140625,
0.20703125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5361328125,
0.208984375,
0.583984375,
0.208984375,
0.583984375,
0.2203369140625,
0.5361328125,
0.2203369140625
],
[
0.51123046875,
0.2215576171875,
0.890625,
0.222900390625,
0.890625,
0.2374267578125,
0.51123046875,
0.236083984375
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.53515625,
0.238037109375,
0.89208984375,
0.238037109375,
0.89208984375,
0.251953125,
0.53515625,
0.251953125
],
[
0.08087158203125,
0.2587890625,
0.271240234375,
0.2587890625,
0.271240234375,
0.272705078125,
0.08087158203125,
0.272705078125
],
[
0.53369140625,
0.25390625,
0.89208984375,
0.25244140625,
0.89208984375,
0.266357421875,
0.53369140625,
0.267578125
],
[
0.53515625,
0.267578125,
0.88818359375,
0.267578125,
0.88818359375,
0.281005859375,
0.53515625,
0.281005859375
],
[
0.098876953125,
0.2822265625,
0.468017578125,
0.2822265625,
0.468017578125,
0.29541015625,
0.098876953125,
0.29541015625
],
[
0.515625,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.515625,
0.296630859375
],
[
0.080078125,
0.29736328125,
0.468017578125,
0.29736328125,
0.468017578125,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.53369140625,
0.29736328125,
0.89208984375,
0.298583984375,
0.89208984375,
0.3125,
0.53369140625,
0.311279296875
],
[
0.07843017578125,
0.31005859375,
0.4697265625,
0.3125,
0.4697265625,
0.326904296875,
0.07843017578125,
0.324462890625
],
[
0.53515625,
0.313232421875,
0.89208984375,
0.313232421875,
0.89208984375,
0.326904296875,
0.53515625,
0.326904296875
],
[
0.080078125,
0.326904296875,
0.46728515625,
0.326904296875,
0.46728515625,
0.34033203125,
0.080078125,
0.34033203125
],
[
0.53515625,
0.32763671875,
0.890625,
0.326904296875,
0.890625,
0.341552734375,
0.53515625,
0.34228515625
],
[
0.08087158203125,
0.343505859375,
0.46728515625,
0.343505859375,
0.46728515625,
0.354248046875,
0.08087158203125,
0.354248046875
],
[
0.53515625,
0.343505859375,
0.89208984375,
0.343505859375,
0.89208984375,
0.356689453125,
0.53515625,
0.356689453125
],
[
0.080078125,
0.356689453125,
0.4697265625,
0.356689453125,
0.4697265625,
0.37060546875,
0.080078125,
0.37060546875
],
[
0.53515625,
0.359130859375,
0.89208984375,
0.359130859375,
0.89208984375,
0.37255859375,
0.53515625,
0.37255859375
],
[
0.080078125,
0.37255859375,
0.4697265625,
0.37255859375,
0.4697265625,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.53515625,
0.373779296875,
0.8505859375,
0.373779296875,
0.8505859375,
0.386962890625,
0.53515625,
0.386962890625
],
[
0.080078125,
0.386962890625,
0.4697265625,
0.386962890625,
0.4697265625,
0.40087890625,
0.080078125,
0.40087890625
],
[
0.50244140625,
0.389404296875,
0.89208984375,
0.389404296875,
0.89208984375,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.080078125,
0.401611328125,
0.46728515625,
0.401611328125,
0.46728515625,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.50244140625,
0.40478515625,
0.89208984375,
0.40478515625,
0.89208984375,
0.41845703125,
0.50244140625,
0.41845703125
],
[
0.080078125,
0.417236328125,
0.468017578125,
0.417236328125,
0.468017578125,
0.43115234375,
0.080078125,
0.43115234375
],
[
0.50244140625,
0.42041015625,
0.62353515625,
0.42041015625,
0.62353515625,
0.434326171875,
0.50244140625,
0.434326171875
],
[
0.08087158203125,
0.434326171875,
0.465576171875,
0.434326171875,
0.465576171875,
0.4443359375,
0.08087158203125,
0.4443359375
],
[
0.08087158203125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.460205078125,
0.08087158203125,
0.460205078125
],
[
0.50341796875,
0.44580078125,
0.65673828125,
0.44580078125,
0.65673828125,
0.460205078125,
0.50341796875,
0.460205078125
],
[
0.080078125,
0.46337890625,
0.4697265625,
0.46337890625,
0.4697265625,
0.4765625,
0.080078125,
0.4765625
],
[
0.52294921875,
0.4697265625,
0.890625,
0.4697265625,
0.890625,
0.48291015625,
0.52294921875,
0.48291015625
],
[
0.08087158203125,
0.479736328125,
0.46728515625,
0.479736328125,
0.46728515625,
0.490478515625,
0.08087158203125,
0.490478515625
],
[
0.50341796875,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.492431640625,
0.468017578125,
0.492431640625,
0.468017578125,
0.505859375,
0.080078125,
0.505859375
],
[
0.50244140625,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.50830078125,
0.468017578125,
0.50830078125,
0.468017578125,
0.521484375,
0.080078125,
0.521484375
],
[
0.50244140625,
0.51513671875,
0.890625,
0.51513671875,
0.890625,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.50244140625,
0.52978515625,
0.89208984375,
0.52978515625,
0.89208984375,
0.54345703125,
0.50244140625,
0.54345703125
],
[
0.07843017578125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.5517578125,
0.07843017578125,
0.5517578125
],
[
0.50244140625,
0.54541015625,
0.89208984375,
0.54541015625,
0.89208984375,
0.5595703125,
0.50244140625,
0.5595703125
],
[
0.080078125,
0.55322265625,
0.4697265625,
0.55322265625,
0.4697265625,
0.56640625,
0.080078125,
0.56640625
],
[
0.50244140625,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.080078125,
0.568359375,
0.4697265625,
0.568359375,
0.4697265625,
0.58203125,
0.080078125,
0.58203125
],
[
0.50244140625,
0.57568359375,
0.73876953125,
0.57568359375,
0.73876953125,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.080078125,
0.583984375,
0.46728515625,
0.583984375,
0.46728515625,
0.59765625,
0.080078125,
0.59765625
],
[
0.08087158203125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.50244140625,
0.599609375,
0.68115234375,
0.60107421875,
0.68115234375,
0.61572265625,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.08087158203125,
0.630859375,
0.46728515625,
0.630859375,
0.46728515625,
0.6416015625,
0.08087158203125,
0.6416015625
],
[
0.521484375,
0.625,
0.890625,
0.625,
0.890625,
0.6396484375,
0.521484375,
0.6396484375
],
[
0.080078125,
0.64453125,
0.2459716796875,
0.64453125,
0.2459716796875,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50244140625,
0.65380859375
],
[
0.50244140625,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.66943359375,
0.50244140625,
0.66943359375
],
[
0.09722900390625,
0.66552734375,
0.468017578125,
0.66650390625,
0.468017578125,
0.68115234375,
0.09722900390625,
0.6796875
],
[
0.50341796875,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.6845703125,
0.50341796875,
0.6845703125
],
[
0.07843017578125,
0.681640625,
0.4697265625,
0.68115234375,
0.4697265625,
0.69580078125,
0.07843017578125,
0.6962890625
],
[
0.50244140625,
0.68603515625,
0.890625,
0.68603515625,
0.890625,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.08087158203125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.708984375,
0.08087158203125,
0.708984375
],
[
0.5009765625,
0.69970703125,
0.8896484375,
0.69970703125,
0.8896484375,
0.7138671875,
0.5009765625,
0.7138671875
],
[
0.08087158203125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.080078125,
0.72802734375,
0.468017578125,
0.72802734375,
0.468017578125,
0.7412109375,
0.080078125,
0.7412109375
],
[
0.50341796875,
0.73291015625,
0.8896484375,
0.73291015625,
0.8896484375,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.080078125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.5048828125,
0.74755859375,
0.890625,
0.74755859375,
0.890625,
0.75830078125,
0.5048828125,
0.75830078125
],
[
0.08087158203125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.76953125,
0.08087158203125,
0.76953125
],
[
0.50341796875,
0.76123046875,
0.89208984375,
0.76123046875,
0.89208984375,
0.77587890625,
0.50341796875,
0.77587890625
],
[
0.07843017578125,
0.7734375,
0.46728515625,
0.7734375,
0.46728515625,
0.787109375,
0.07843017578125,
0.787109375
],
[
0.50244140625,
0.7763671875,
0.609375,
0.7763671875,
0.609375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.080078125,
0.79052734375,
0.1217041015625,
0.79052734375,
0.1217041015625,
0.80029296875,
0.080078125,
0.80029296875
],
[
0.50244140625,
0.80322265625,
0.73876953125,
0.80322265625,
0.73876953125,
0.81640625,
0.50244140625,
0.81640625
],
[
0.09722900390625,
0.8095703125,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.09722900390625,
0.82373046875
],
[
0.07843017578125,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.521484375,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.521484375,
0.84033203125
],
[
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,
0.4697265625,
0.85693359375,
0.4697265625,
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,
0.471435546875,
0.8720703125,
0.471435546875,
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,
0.1895751953125,
0.88720703125,
0.1895751953125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.8935546875,
0.88623046875,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"bottom-up approaches above, is that they struggle to use",
"3. Methodology",
"contextual information:instead they rely on accurate char-",
"acter classifiers. Shi et al. 2015 [31] and He et al. [11]",
"As presented in Fig.3, our proposed architecture consists",
"offourmain components:",
"considered words as sequences of varying lengths, and em-",
"ployed RNNs to model the sequences without explicit char-",
"1. Transformation: the input text image is normalized",
"acter separation. Shi et al. 2016 [32] presented a suc-",
"using a Spatial Transformer Network (STN) [13].",
"2. Feature Extraction: maps the input image to a feature",
"cessful end-to-end trainable architecture using the sequence",
"approach, without relying on character level annotations.",
"map representation while using a text attention mod-",
"Their solution employed a BiLSTM layer to extract the se-",
"ule [7].",
"3. Visual Feature Refinement: provides direct super-",
"quential feature vectors from the input feature maps, these",
"vectors are then fed into an attention-Gated Recurrent Unit",
"vision for each column in the visual features. This",
"(GRU) module for decoding.",
"part refines the representation in each of the feature",
"columns, by classifying them into individual symbols.",
"The methods mentioned above introduced significant im-",
"4. Selective-Contextual Refinement Block: Each block",
"provements in STR accuracy on public benchmarks. There-",
"consists of a two-layer BiLSTM encoder that outputs",
"fore, recent work has shifted focus to the more challenging",
"contextual features. The contextual features are con-",
"problem of recognizing irregularly shaped text, hence pro-",
"catenated to the visual features computed by the CNN",
"moting new lines of research. Topics such as input rectifi-",
"backbone.This concatenated feature map is then fed",
"cation, character-level segmentation, 2D attentional feature",
"into the selective-decoder, which employs a two-step",
"maps and self attention have emerged, pushing the envelope",
"1D attention mechanism, as illustrated in Fig. 4.",
"on irregular STR. Shi et al. 2018 [33] rectified oriented or",
"In this section we describe the training architecture of",
"curved text based on a Spatial Transformer Network (STN)",
"SCATTER, while addressing the differences between train-",
"Liu et al. 2018 [19] introduced a Character-Aware Neural",
"ing and inference.",
"Network(Char-Net)todetectandrectifyindividualcharac-",
"ters.A combination of a CTC-Attention mechanism within",
"3.1.Transformation",
"an encoder-decoder framework, that was used for speech",
"The transformation step operates on the cropped text im-",
"recognition tasks,was used for STRin[45l,showing the",
"age X, and transforms it into a normalized image X'. We",
"benefits of joint CTC-Attention learning. The authors in [7]",
"use a Thin Plate Spline (TPS) transformation, a variant of",
"proposed two supervision branches to tackle explicit and",
"the STN, as used in [1]. TPS employs a smooth spline in-",
"implicit semantic information. In [4] a gate was inserted",
"terpolation between a set of fiducial points. Specifically, it",
"to the recurrent decoder, for controlling the transmission-",
"detects a pre-defined number of fiducial points at the top",
"weight of the previous embedded vector, demonstrating that",
"and bottom of the text region, and normalizes the predicted",
"context is not always needed for decoding. The authors of",
"region to a constant predefined size.",
"Mask TextSpotter [17] unified text detection and text recog-",
"nitionin an end-to-end fashion.Forrecognition they used",
"3.2.Feature Extraction",
"two separate branches, a branch that uses visual (local) fea-",
"tures and a branch which utilizes contextual information in",
"In this step a convolutional neural network (CNN) ex-",
"the form of 2D attention.",
"tracts features from the input image.We use a 29-layer",
"ResNet as the CNN's backbone, as used in [5]. The out-",
"More recent approaches have proposed leveraging var-",
"put of the feature encoder is 512 channels by N columns.",
"ious attention mechanisms for improved results.Li et",
"Specifically, the feature encoder gets an input image X'",
"al.J16lcombinedbothvisualandcontextual featureswhile",
"and outputs a feature map F - [fi, f2,.., fl. Follow-",
"utilizinga2Dattentionwithintheencoder-decoder.Other",
"ing the feature map extraction, we use a text attention mod-",
"researchers borrowed ideas from the Natural Language Pro-",
"ule, similar to [7l. The attentional feature map can be re-",
"cessing (NLP) domain and adopted a transformer-based ar-",
"gardedas a visual feature sequence of length N,denoted as",
"chitecture[35].Oneof them is Sheng et al.[30l, that used",
"V = [ui, U2, .., Unl, where each column represents a frame",
"a self-attention mechanism for both the encoder and the de-",
"in the sequence.",
"coder.",
"3.3.Visual Feature Refinement",
"Our method differs from above approaches, by being the",
"first to utilize a stacked block architecture for text recogni-",
"Here, the visual feature sequence V is used for interme-",
"tion. Namely, we show that repetitive processing for text",
"diate decoding. This intermediate supervision is aimed at",
"recognition, trained with intermediate selective decoders as",
"refining the character embedding (representations) in each",
"supervision (similar to [40, 27, 26]), increasingly refines",
"of the columns of V, and is done using CTC based decod-",
"text predictions.",
"ing. We feed V through a fully connected layer that outputs"
] |
{
"arxiv_id": "2003.11288",
"language": "en",
"page_idx": 2
}
|
||
2003.11288-en-4
|
2003.11288-en-4.jpg
|
weights, $\alpha_t\in R^N:$
(3)
$$\begin{aligned}&e_{t,i}=w^{T}\tanh(Ws_{t-1}+Vd_{i}^{\prime}+b)\\&\alpha_{t,i}=\exp(e_{t,i})/\sum_{i^{*}=1}^{n}e_{t,i^{*}}\:,\end{aligned}$$
(4)
where $b,w,W,V$ are trainable parameters, $s_t-1$ is the hidden state of the recurrent cell within the decoder at time $t$, and $d\prime$ is a column of $D′.$ The decoder linearly combines the columns of $D^\prime$ into a vector $G$,which is called a glimpse:
(5)
$$g_t=\sum_{i=1}^n\alpha_{t,i}d_i'\:.$$
Next, a recurrent cell of the decoder is fed with
$$(x_t,s_t)=\text{RNN}\Big(s_{t-1},\big(g_t,f(y_{t-1})\big)\Big),$$
(6)
where $(g_t,f(y_{t-1}))$ denotes the concatenation between $g_t$
and the one-hot embedding of $y_{t-1}.$
The probability for a given character $p(y_t)$ can now be
recovered by:
$$p(y_t)=\mathrm{softmax}(W_ox_t+b_o)\:.$$
(7)
The loss for the $j$th block is the negative log-likelihood, de-
noted as $L_{\mathrm{Attn},j}$, as in [5,2].
## 3.5. Training Losses
The objective function is given by
(8)
$$L=\lambda_{\mathrm{CTC}}\cdot L_{\mathrm{CTC}}+\sum_{j=1}\lambda_{j}L_{\mathrm{Attn},j}\:,$$
where $L_\mathrm{CTC}$ is the loss function of the CTC decoder and $\sum_{j=1}\lambda_{j}L_{\mathrm{Attn},j}$is the sum of the losses from all of the Selective-Contextual Refinement blocks, as defined above. The $\lambda$ notation depicts a hyper-parameter used to balance the trade-off between the different supervisions, and specifically, $\lambda_CTC,\lambda_j$ are empirically set to 0.1,1.0 respectively for all $j.$
## 3.6. Inference
Once training is done, for test time we remove all of the intermediate decoders, as they are used only for additional supervision and refinement of intermediate feature. The visual features $V$ are processed by the BiLSTM layers in all blocks, and are also fed directly, via a skip connection, to the final selective-decoder. The final selective decoder is used to predict the output sequence of characters. A visualization of these changes can be seen in Fig. 3, where all the green colored operations are disabled during inference, and $\operatorname{in}$Fig. 1$(\mathbf{b}).$
## $4. \textbf{Experiments}$
In this section we empirically demonstrate the effectiveness of our proposed framework. We begin with a brief discussion regarding the datasets used for training and testing, and then describe our implementation and evaluation setup. Next, we compare our model against state-of-the-art methods on public benchmark datasets, including both regular and irregular text. Finally, we address the computational cost of our method.
### $4. 1. \textbf{Datasets}$
In this work, all SCATTER models are trained on three synthetic datasets. The model is evaluated on four regular scene-text datasets: ICDAR2003, ICDAR2013, IIIT5K, SVT, and three irregular text datasets: ICDAR2015, SVTP and CUTE
The training dataset is a union of three datasets:
MJSynth (MJ)[12] is a synthetic text in image dataset which contains 9 million word box images, generated from a lexicon of 90K English words.
$\mathbf{SynthText}\left ( \mathrm{ST}\right ) [ 10]$ is a synthetic text in image dataset, designed for scene-text detection, and recognition. We use a variant of the SynthText dataset composed of 5.5M samples, as used in [1]. This variant does not include any nonalphanumeric characters.
SynthAdd (SA)[16] is a synthetic text in image dataset, that contains 1.2 million word box images. This dataset was generated using the same synthetic engine as in ST, aiming to mitigate the lack of non-alphanumeric characters (e.g., punctuation marks) in the other datasets.
All experiments are evaluated on the seven real-word STR benchmark datasets described below. As in many STR manuscripts (e.g,[33,1,16]) the benchmark datasets are commonly divided into regular and irregular text, according to the text layout.
Regular text datasets include the following: IIIT5K[25] consists of 2000 training and 3000 testing images that are cropped from Google image searches. SVT[37] is a dataset collected from Google Street View images and contains 257 training and 647 testing word-box cropped images. ICDAR2003 [23] contains 867 word-box cropped images for testing. ICDAR2013 [15] contains 848 training and 1,015 testing word-box cropped images.
Irregular text datasets include the following: ICDAR2015[14] contains 4,468 training and 2,077 testing word-box cropped images, all captured by Google Glass, without careful positioning or focusing. SVTP [28] is a dataset collected from Google Street View images and consists of 645 cropped word-box images for testing. CUTE 80 [29] contains 288 cropped word-box images for testing, many of which are curved text images.
|
[
[
0.07843017578125,
0.0921630859375,
0.2059326171875,
0.0889892578125,
0.2059326171875,
0.10418701171875,
0.07843017578125,
0.106689453125
],
[
0.50244140625,
0.0889892578125,
0.63134765625,
0.09027099609375,
0.63134765625,
0.10479736328125,
0.50244140625,
0.10418701171875
],
[
0.15283203125,
0.11614990234375,
0.39453125,
0.11553955078125,
0.39453125,
0.1363525390625,
0.15283203125,
0.136962890625
],
[
0.44677734375,
0.121826171875,
0.4697265625,
0.121826171875,
0.4697265625,
0.1337890625,
0.44677734375,
0.1337890625
],
[
0.521484375,
0.11871337890625,
0.890625,
0.11871337890625,
0.890625,
0.1319580078125,
0.521484375,
0.1319580078125
],
[
0.50244140625,
0.1337890625,
0.890625,
0.1337890625,
0.890625,
0.147705078125,
0.50244140625,
0.147705078125
],
[
0.1519775390625,
0.146484375,
0.339111328125,
0.149658203125,
0.339111328125,
0.1685791015625,
0.1519775390625,
0.1654052734375
],
[
0.4453125,
0.1514892578125,
0.4697265625,
0.1514892578125,
0.4697265625,
0.1641845703125,
0.4453125,
0.1641845703125
],
[
0.5009765625,
0.1483154296875,
0.89208984375,
0.149658203125,
0.89208984375,
0.1634521484375,
0.5009765625,
0.1622314453125
],
[
0.273681640625,
0.169189453125,
0.298095703125,
0.169189453125,
0.298095703125,
0.1767578125,
0.273681640625,
0.1767578125
],
[
0.50244140625,
0.1634521484375,
0.8896484375,
0.1634521484375,
0.8896484375,
0.1767578125,
0.50244140625,
0.1767578125
],
[
0.5048828125,
0.18115234375,
0.8896484375,
0.18115234375,
0.8896484375,
0.1912841796875,
0.5048828125,
0.1912841796875
],
[
0.080078125,
0.1900634765625,
0.468017578125,
0.1900634765625,
0.468017578125,
0.203857421875,
0.080078125,
0.203857421875
],
[
0.50341796875,
0.1956787109375,
0.890625,
0.1956787109375,
0.890625,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.07843017578125,
0.203857421875,
0.4697265625,
0.2052001953125,
0.4697265625,
0.2197265625,
0.07843017578125,
0.2183837890625
],
[
0.50244140625,
0.208984375,
0.89208984375,
0.208984375,
0.89208984375,
0.222900390625,
0.50244140625,
0.222900390625
],
[
0.08087158203125,
0.2215576171875,
0.46728515625,
0.2215576171875,
0.46728515625,
0.231689453125,
0.08087158203125,
0.231689453125
],
[
0.50341796875,
0.2259521484375,
0.63134765625,
0.2259521484375,
0.63134765625,
0.236083984375,
0.50341796875,
0.236083984375
],
[
0.080078125,
0.23486328125,
0.45751953125,
0.23486328125,
0.45751953125,
0.248779296875,
0.080078125,
0.248779296875
],
[
0.5048828125,
0.2568359375,
0.6005859375,
0.2568359375,
0.6005859375,
0.268310546875,
0.5048828125,
0.268310546875
],
[
0.27685546875,
0.2734375,
0.324462890625,
0.2734375,
0.324462890625,
0.28857421875,
0.27685546875,
0.28857421875
],
[
0.44677734375,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.44677734375,
0.287353515625
],
[
0.521484375,
0.281005859375,
0.89208984375,
0.281005859375,
0.89208984375,
0.294189453125,
0.521484375,
0.294189453125
],
[
0.50244140625,
0.296630859375,
0.890625,
0.296630859375,
0.890625,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.221435546875,
0.27783203125,
0.238525390625,
0.27783203125,
0.238525390625,
0.285888671875,
0.221435546875,
0.285888671875
],
[
0.07843017578125,
0.311279296875,
0.3896484375,
0.31005859375,
0.3896484375,
0.324462890625,
0.07843017578125,
0.32568359375
],
[
0.50244140625,
0.311279296875,
0.890625,
0.311279296875,
0.890625,
0.324462890625,
0.50244140625,
0.324462890625
],
[
0.50341796875,
0.32568359375,
0.89208984375,
0.32568359375,
0.89208984375,
0.339111328125,
0.50341796875,
0.339111328125
],
[
0.143798828125,
0.34033203125,
0.259033203125,
0.33837890625,
0.259033203125,
0.356689453125,
0.143798828125,
0.359130859375
],
[
0.25732421875,
0.341552734375,
0.390625,
0.339111328125,
0.390625,
0.356689453125,
0.25732421875,
0.359130859375
],
[
0.44677734375,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.35595703125,
0.44677734375,
0.35595703125
],
[
0.5009765625,
0.34228515625,
0.57861328125,
0.34033203125,
0.57861328125,
0.354736328125,
0.5009765625,
0.356689453125
],
[
0.521484375,
0.35791015625,
0.83642578125,
0.359130859375,
0.83642578125,
0.373779296875,
0.521484375,
0.37255859375
],
[
0.08087158203125,
0.375,
0.468017578125,
0.375,
0.468017578125,
0.389404296875,
0.08087158203125,
0.389404296875
],
[
0.52294921875,
0.376953125,
0.89208984375,
0.376953125,
0.89208984375,
0.39013671875,
0.52294921875,
0.39013671875
],
[
0.080078125,
0.39013671875,
0.316162109375,
0.39013671875,
0.316162109375,
0.404052734375,
0.080078125,
0.404052734375
],
[
0.50244140625,
0.39013671875,
0.89208984375,
0.391357421875,
0.89208984375,
0.406005859375,
0.50244140625,
0.40478515625
],
[
0.098876953125,
0.40478515625,
0.4697265625,
0.40478515625,
0.4697265625,
0.419189453125,
0.098876953125,
0.419189453125
],
[
0.50244140625,
0.4072265625,
0.71728515625,
0.4072265625,
0.71728515625,
0.42041015625,
0.50244140625,
0.42041015625
],
[
0.08087158203125,
0.421630859375,
0.1707763671875,
0.421630859375,
0.1707763671875,
0.43310546875,
0.08087158203125,
0.43310546875
],
[
0.51953125,
0.42236328125,
0.89208984375,
0.423583984375,
0.89208984375,
0.438232421875,
0.51953125,
0.4375
],
[
0.5048828125,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.451904296875,
0.5048828125,
0.451904296875
],
[
0.1707763671875,
0.448974609375,
0.375732421875,
0.447509765625,
0.375732421875,
0.462158203125,
0.1707763671875,
0.46337890625
],
[
0.44677734375,
0.448974609375,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.44677734375,
0.462158203125
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.50244140625,
0.468505859375,
0.89208984375,
0.4697265625,
0.89208984375,
0.484130859375,
0.50244140625,
0.48291015625
],
[
0.08087158203125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.08087158203125,
0.4892578125
],
[
0.50341796875,
0.48486328125,
0.66845703125,
0.48486328125,
0.66845703125,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.492431640625,
0.2646484375,
0.492431640625,
0.2646484375,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.52294921875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.52294921875,
0.51513671875
],
[
0.080078125,
0.5166015625,
0.2344970703125,
0.5166015625,
0.2344970703125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.50341796875,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.546875,
0.50341796875,
0.546875
],
[
0.09722900390625,
0.5390625,
0.32275390625,
0.54052734375,
0.32275390625,
0.5546875,
0.09722900390625,
0.5537109375
],
[
0.5009765625,
0.546875,
0.890625,
0.54736328125,
0.890625,
0.5625,
0.5009765625,
0.56103515625
],
[
0.158447265625,
0.564453125,
0.38818359375,
0.5673828125,
0.38818359375,
0.5888671875,
0.158447265625,
0.5859375
],
[
0.50244140625,
0.56298828125,
0.76953125,
0.56298828125,
0.76953125,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.4453125,
0.5712890625,
0.4697265625,
0.5712890625,
0.4697265625,
0.58544921875,
0.4453125,
0.58544921875
],
[
0.52294921875,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.52294921875,
0.59326171875
],
[
0.295654296875,
0.5888671875,
0.316162109375,
0.5888671875,
0.316162109375,
0.59765625,
0.295654296875,
0.59765625
],
[
0.50341796875,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.50341796875,
0.60791015625
],
[
0.08087158203125,
0.611328125,
0.468017578125,
0.611328125,
0.468017578125,
0.625,
0.08087158203125,
0.625
],
[
0.50244140625,
0.60986328125,
0.89208984375,
0.60986328125,
0.89208984375,
0.62353515625,
0.50244140625,
0.62353515625
],
[
0.08087158203125,
0.626953125,
0.468017578125,
0.625,
0.468017578125,
0.6396484375,
0.08087158203125,
0.6416015625
],
[
0.50244140625,
0.62548828125,
0.89208984375,
0.62548828125,
0.89208984375,
0.6396484375,
0.50244140625,
0.6396484375
],
[
0.08087158203125,
0.6416015625,
0.46728515625,
0.6416015625,
0.46728515625,
0.65478515625,
0.08087158203125,
0.65478515625
],
[
0.50244140625,
0.6416015625,
0.64453125,
0.6416015625,
0.64453125,
0.65478515625,
0.50244140625,
0.65478515625
],
[
0.08087158203125,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.51953125,
0.6572265625,
0.89208984375,
0.65771484375,
0.89208984375,
0.67236328125,
0.51953125,
0.671875
],
[
0.07843017578125,
0.67041015625,
0.468017578125,
0.671875,
0.468017578125,
0.68603515625,
0.07843017578125,
0.68505859375
],
[
0.50341796875,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.07843017578125,
0.68603515625,
0.468017578125,
0.6875,
0.468017578125,
0.7021484375,
0.07843017578125,
0.70068359375
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.07843017578125,
0.7021484375,
0.1396484375,
0.7021484375,
0.1396484375,
0.71630859375,
0.07843017578125,
0.71630859375
],
[
0.50341796875,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71630859375,
0.50341796875,
0.71630859375
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.07843017578125,
0.7265625,
0.1871337890625,
0.7265625,
0.1871337890625,
0.7412109375,
0.07843017578125,
0.7412109375
],
[
0.50244140625,
0.73291015625,
0.890625,
0.73193359375,
0.890625,
0.7470703125,
0.50244140625,
0.74755859375
],
[
0.098876953125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.098876953125,
0.7646484375
],
[
0.5009765625,
0.74755859375,
0.890625,
0.74755859375,
0.890625,
0.76123046875,
0.5009765625,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7646484375,
0.8505859375,
0.7646484375,
0.8505859375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.521484375,
0.78076171875,
0.859375,
0.78076171875,
0.859375,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.8544921875,
0.7822265625,
0.8896484375,
0.7822265625,
0.8896484375,
0.79345703125,
0.8544921875,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
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
],
[
0.5009765625,
0.81103515625,
0.8896484375,
0.81103515625,
0.8896484375,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.08087158203125,
0.82763671875,
0.468017578125,
0.82763671875,
0.468017578125,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
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,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
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,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.07843017578125,
0.88525390625,
0.1666259765625,
0.88720703125,
0.1666259765625,
0.9013671875,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.7587890625,
0.88720703125,
0.7587890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"weights, Qt E RN.",
"4.Experiments",
"et,i = wT tanh(Wst-1 + Vd + b)",
"(3)",
"In this section we empirically demonstrate the effective-",
"ness of our proposed framework. We begin with a brief dis-",
"Qt,i = exp(et,i)/ et,i* ,",
"(4)",
"cussion regarding the datasets used for training and testing,",
"i*=1",
"and then describe our implementation and evaluation setup.",
"Next,wecompareourmodelagainst state-of-the-artmeth-",
"where b, w, W, V are trainable parameters, st-1 is the hid-",
"ods on publicbenchmarkdatasets,includingboth regular",
"den state of the recurrent cell within the decoder at time t,",
"and irregular text. Finally, we address the computational",
"anddiisacolumnofDiThedecoderlinearlycombinesthe",
"cost of our method.",
"columns of D' into a vector G, which is called a glimpse:",
"4.1.Datasets",
"at,id',",
"(5)",
"In this work, all SCATTER models are trained on three",
"synthetic datasets. The model is evaluated on four regu-",
"9t",
"Next, a recurrent cell of the decoder is fed with",
"lar scene-text datasets: ICDAR2003, ICDAR2013, IIIT5K,",
"SVT, and three irregular text datasets: ICDAR2015, SVTP",
"(rt, st) = RNN(",
"St-1, (gt, f(yt-1)",
"(6)",
"and CUTE.",
"The training dataset is a union of three datasets:",
"where (gt, f(yt-1)) denotes the concatenation between gt",
"MJSynth (MJ) [12] is a synthetic text in image dataset",
"and the one-hot embedding of yt-1.",
"which contains 9 million word box images, generated from",
"The probability for a given character p(yt) can now be",
"a lexicon of 90K English words.",
"recovered by:",
"SynthText (ST) [10] is a synthetic text in image dataset,",
"designedfor scene-textdetection,and recognition.Weuse",
"p(yt) = softmax(Wort +bo).",
"(7)",
"a variant of the SynthText dataset composed of 5.5M sam-",
"ples, as used in [I]. This variant does not include any non-",
"The loss for the jth block is the negative log-likelihood, de-",
"alphanumeric characters.",
"noted as LAtn,j, as in [5, 2].",
"SynthAdd (SA) [16] is a synthetic text in image dataset,",
"3.5. Training Losses",
"that contains 1.2 million word box images. This dataset was",
"generated using the same synthetic engine as in ST, aiming",
"The objectivefunction is given by",
"to mitigate the lack of non-alphanumeric characters (e.g.,",
"L = Acrc - Lcrc +入, LAm,j,",
"punctuation marks) in the other datasets.",
"(8)",
"All experiments are evaluated on the seven real-word",
"i=1",
"STR benchmark datasets described below. As in many STR",
"where LcTc is the loss function of the CTC decoder and",
"manuscripts (e.g, [33, 1, 16]) the benchmark datasets are",
"Zi=1^, LAttn,i is the sum of the losses from all of the",
"commonly divided into regular and irregular text, accord-",
"Seiective-Contextual Refinement blocks, as defined above.",
"ing to the text layout.",
"The notation depicts a hyper-parameter used to balance",
"Regulartext datasets includethefollowing:",
"the trade-off between the different supervisions, and specif-",
"I1IT5K [25] consists of 2000 training and 3000 test-",
"ically, AcTC, >, are empirically set to 0.1, 1.0 respectively",
"ing images that are cropped from Google image searches.",
"for all j.",
"SVT [37] is a dataset collected from Google Street View",
"images and contains 257 training and 647 testing word-box",
"3.6. Inference",
"cropped images.ICDAR2003[23] contains867word-box",
"Once training is done, for test time we remove all of the",
"cropped images for testing. ICDAR2013 [15] contains 848",
"intermediate decoders, as they are used only for additional",
"training and 1,015 testing word-box cropped images.",
"Irregular text datasets include the following:",
"IC-",
"sual features V are processed by the BiLSTM layers in all",
"supervision and refinement of intermediate feature.The vi-",
"DAR2015 [14] contains 4,468 training and 2,077 testing",
"blocks, and are also fed directly, via a skip connection, to",
"word-box cropped images, all captured by Google Glass,",
"the final selective-decoder.The final selective decoder is",
"without careful positioning or focusing. SVTP [28] is a",
"used to predict the output sequence of characters. A visual-",
"dataset collected from Google Street View images and con-",
"ization of these changes can be seen in Fig. 3, where all the",
"sists of 645 cropped word-box images for testing. CUTE",
"green colored operations are disabled during inference, and",
"80 [29] contains 288 cropped word-box images for testing,",
"in Fig. 1 (b).",
"many of which are curved text images."
] |
{
"arxiv_id": "2003.11288",
"language": "en",
"page_idx": 4
}
|
||
1903.11279-en-1
|
1903.11279-en-1.jpg
|
# 1.1 Challenges
IE from VRDs is a challenging task, and the difficulties mainly arise from how to effectively incorporate visual cues from the document and the scalability of the task.
First, text alone is not adequate to represent the semantic meaning in VRDs, and the contexts of the texts are usually expressed in visual cues. For example, there might be multiple dates in the purchase receipts. However, it is up to the visual part of the model to distinguish between the Invoice Date, Transaction Date, and Due Date. Another example is the tax amount in value-added tax invoice, as shown in Figure 1(b). There are multiple “money” entities in the document, and there is a lack of any textual context to determine which one is tax amount. To extract the tax amount correctly, we have to leverage the (relative) position of the text segment and visual features of the document in general.
Template matching based algorithms (Chiticariu et al., 2013; Dengel and Klein, 2002; Schuster et al., 2013) utilize visual features of the document to extract entities; however, we argue that they are mostly not scalable for the task in realworld business settings. There are easily thousands of vendors on the market, and the templates of purchase receipts from each vendor are not the same. Thousands of templates need to be created and maintained in this single scenario. It requires substantial efforts to update the template and make sure it's not conflicting with the rest every time a new template comes in, and the process is errorprone. Besides, user uploaded pictures introduce another dimension of variance from the template. An example is shown in Figure 1(b). Value-added tax invoice is a nation-wide tax document, and the layout is fxed. However, pictures taken by users are usually distorted, often blurred and sometimes contain interfering objects in the image. A simple template-based system performs poorly in such a scenario, while sophisticated rules require significant engineering efforts for each scenario, which we believe is not scalable.
## 1.2 Contributions
In this paper, we present a novel method for IE from VRDs. The method first computes graph embeddings for each text segment in the document using graph convolution. The graph embeddings represent the context of the current text segment
where the convolution operation combines both textual and visual features of the context. Then the graph embeddings are combined with text embeddings to feed into a standard BiLSTM for information extraction.
Extensive experiments have been conducted to show our method outperforms BiLSTM-CRF baselines by significant margins, on two realworld datasets. Additionally, ablation studies are also performed to evaluate the effectiveness of each component of our model. Furthermore, we also provide analysis and intuitions on why and how individual components work in our experiments.
## 2 Related Works
Our work is inspired by recent research in the area of graph convolution and information extraction.
## 2.1 Graph Convolution Network
Neural network architectures such as CNN and RNNs have demonstrated huge success on many artificial intelligence tasks where the underlying data has grid-like or sequential structure (Krizhevsky et al., 2012; Kim et al., 2016; Kim, 2014). Recently, there is a surge of interest in studying the neural network structure operating on graphs (Kipf and Welling, 2016; Hamilton et al., 2017), since much data in the real world is naturally represented as graphs. Many works attempt to generalize convolution on the graph structure. Some use a spectrum based approach where the learned model depends on the structure of the graph. As a result, the approach does not work well on dynamic graph structures. The others define convolution directly on the graph (Veličković et al., 2017; Hamilton et al., 2017; Xu et al., 2018; Johnson et al., 2018; Duvenaud et al., 2015). We follow the latter approach in our work to model the text segment graph of VRDs.
Different from existing works, this paper introduces explicit edge embeddings into the graph convolution network, which models the relationship between vertices directly. Similar to (Veličković et al., 2017), we apply self-attention (Vaswani et al., 2017) to define convolution on variable-sized neighbors, and the approach is computationally efficient since the operation is parallelizable across node pairs.
|
[
[
0.12005615234375,
0.07598876953125,
0.2493896484375,
0.07598876953125,
0.2493896484375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.5146484375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.5146484375,
0.0926513671875,
0.88330078125,
0.0926513671875,
0.88330078125,
0.105712890625,
0.5146484375,
0.105712890625
],
[
0.118408203125,
0.09857177734375,
0.489501953125,
0.09857177734375,
0.489501953125,
0.1134033203125,
0.118408203125,
0.1134033203125
],
[
0.51318359375,
0.10809326171875,
0.8818359375,
0.10687255859375,
0.8818359375,
0.1217041015625,
0.51318359375,
0.122314453125
],
[
0.12005615234375,
0.11517333984375,
0.48779296875,
0.11517333984375,
0.48779296875,
0.12890625,
0.12005615234375,
0.12890625
],
[
0.51318359375,
0.12469482421875,
0.8818359375,
0.12469482421875,
0.8818359375,
0.1383056640625,
0.51318359375,
0.1383056640625
],
[
0.1175537109375,
0.1318359375,
0.489501953125,
0.1300048828125,
0.489501953125,
0.1436767578125,
0.1175537109375,
0.1455078125
],
[
0.118408203125,
0.146728515625,
0.280517578125,
0.146728515625,
0.280517578125,
0.1602783203125,
0.118408203125,
0.1602783203125
],
[
0.51318359375,
0.141357421875,
0.67236328125,
0.141357421875,
0.67236328125,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.1351318359375,
0.163330078125,
0.489501953125,
0.1644287109375,
0.489501953125,
0.1787109375,
0.1351318359375,
0.1776123046875
],
[
0.52880859375,
0.157958984375,
0.88330078125,
0.1591796875,
0.88330078125,
0.17333984375,
0.52880859375,
0.1728515625
],
[
0.118408203125,
0.18115234375,
0.491943359375,
0.18115234375,
0.491943359375,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.51318359375,
0.17578125,
0.88330078125,
0.17578125,
0.88330078125,
0.189453125,
0.51318359375,
0.189453125
],
[
0.118408203125,
0.19775390625,
0.491943359375,
0.19775390625,
0.491943359375,
0.21142578125,
0.118408203125,
0.21142578125
],
[
0.51318359375,
0.1923828125,
0.8818359375,
0.1923828125,
0.8818359375,
0.207275390625,
0.51318359375,
0.207275390625
],
[
0.118408203125,
0.213134765625,
0.489501953125,
0.213134765625,
0.489501953125,
0.226806640625,
0.118408203125,
0.226806640625
],
[
0.51123046875,
0.2078857421875,
0.8818359375,
0.2078857421875,
0.8818359375,
0.221435546875,
0.51123046875,
0.221435546875
],
[
0.12005615234375,
0.2298583984375,
0.490234375,
0.2298583984375,
0.490234375,
0.243408203125,
0.12005615234375,
0.243408203125
],
[
0.51318359375,
0.2244873046875,
0.884765625,
0.2244873046875,
0.884765625,
0.2381591796875,
0.51318359375,
0.2381591796875
],
[
0.12005615234375,
0.2452392578125,
0.490234375,
0.2452392578125,
0.490234375,
0.2587890625,
0.12005615234375,
0.2587890625
],
[
0.51318359375,
0.240478515625,
0.88330078125,
0.240478515625,
0.88330078125,
0.254150390625,
0.51318359375,
0.254150390625
],
[
0.12005615234375,
0.26123046875,
0.490234375,
0.26123046875,
0.490234375,
0.27490234375,
0.12005615234375,
0.27490234375
],
[
0.51318359375,
0.257080078125,
0.88330078125,
0.257080078125,
0.88330078125,
0.270751953125,
0.51318359375,
0.270751953125
],
[
0.12005615234375,
0.275634765625,
0.48779296875,
0.275634765625,
0.48779296875,
0.289306640625,
0.12005615234375,
0.289306640625
],
[
0.51123046875,
0.271484375,
0.8818359375,
0.2724609375,
0.8818359375,
0.287353515625,
0.51123046875,
0.2861328125
],
[
0.12005615234375,
0.29345703125,
0.489501953125,
0.29345703125,
0.489501953125,
0.30712890625,
0.12005615234375,
0.30712890625
],
[
0.51318359375,
0.290283203125,
0.56494140625,
0.290283203125,
0.56494140625,
0.3017578125,
0.51318359375,
0.3017578125
],
[
0.118408203125,
0.30810546875,
0.489501953125,
0.308837890625,
0.489501953125,
0.32373046875,
0.118408203125,
0.322509765625
],
[
0.118408203125,
0.325927734375,
0.489501953125,
0.325927734375,
0.489501953125,
0.339111328125,
0.118408203125,
0.339111328125
],
[
0.5400390625,
0.32666015625,
0.67529296875,
0.32666015625,
0.67529296875,
0.341552734375,
0.5400390625,
0.341552734375
],
[
0.1175537109375,
0.34033203125,
0.48779296875,
0.341552734375,
0.48779296875,
0.355712890625,
0.1175537109375,
0.355224609375
],
[
0.118408203125,
0.35693359375,
0.48779296875,
0.35693359375,
0.48779296875,
0.37060546875,
0.118408203125,
0.37060546875
],
[
0.51123046875,
0.359375,
0.8818359375,
0.359375,
0.8818359375,
0.372314453125,
0.51123046875,
0.372314453125
],
[
0.118408203125,
0.37353515625,
0.48779296875,
0.37353515625,
0.48779296875,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.51318359375,
0.377197265625,
0.873046875,
0.377197265625,
0.873046875,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.118408203125,
0.39013671875,
0.197265625,
0.39013671875,
0.197265625,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.1368408203125,
0.40673828125,
0.489501953125,
0.40673828125,
0.489501953125,
0.421630859375,
0.1368408203125,
0.421630859375
],
[
0.5146484375,
0.413330078125,
0.78271484375,
0.413330078125,
0.78271484375,
0.427490234375,
0.5146484375,
0.427490234375
],
[
0.118408203125,
0.4228515625,
0.48779296875,
0.4228515625,
0.48779296875,
0.43701171875,
0.118408203125,
0.43701171875
],
[
0.118408203125,
0.439453125,
0.489501953125,
0.439453125,
0.489501953125,
0.453125,
0.118408203125,
0.453125
],
[
0.5146484375,
0.441162109375,
0.884765625,
0.441162109375,
0.884765625,
0.454833984375,
0.5146484375,
0.454833984375
],
[
0.1175537109375,
0.4560546875,
0.489501953125,
0.4560546875,
0.489501953125,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.51123046875,
0.4560546875,
0.88330078125,
0.457763671875,
0.88330078125,
0.471435546875,
0.51123046875,
0.4697265625
],
[
0.1175537109375,
0.471435546875,
0.48779296875,
0.471435546875,
0.48779296875,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.51318359375,
0.47265625,
0.8818359375,
0.47265625,
0.8818359375,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.118408203125,
0.48681640625,
0.490234375,
0.48681640625,
0.490234375,
0.501953125,
0.118408203125,
0.501953125
],
[
0.51318359375,
0.4892578125,
0.884765625,
0.4892578125,
0.884765625,
0.5029296875,
0.51318359375,
0.5029296875
],
[
0.12005615234375,
0.50439453125,
0.491943359375,
0.50439453125,
0.491943359375,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.51123046875,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.51123046875,
0.5185546875
],
[
0.118408203125,
0.51953125,
0.489501953125,
0.51953125,
0.489501953125,
0.533203125,
0.118408203125,
0.533203125
],
[
0.51318359375,
0.521484375,
0.884765625,
0.521484375,
0.884765625,
0.53515625,
0.51318359375,
0.53515625
],
[
0.12176513671875,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.54736328125,
0.12176513671875,
0.54736328125
],
[
0.51123046875,
0.5361328125,
0.884765625,
0.53759765625,
0.884765625,
0.5517578125,
0.51123046875,
0.55029296875
],
[
0.12176513671875,
0.55419921875,
0.48779296875,
0.55419921875,
0.48779296875,
0.56396484375,
0.12176513671875,
0.56396484375
],
[
0.51318359375,
0.55419921875,
0.88671875,
0.552734375,
0.88671875,
0.56787109375,
0.51318359375,
0.568359375
],
[
0.12005615234375,
0.568359375,
0.489501953125,
0.568359375,
0.489501953125,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.5146484375,
0.57080078125,
0.88330078125,
0.57080078125,
0.88330078125,
0.58447265625,
0.5146484375,
0.58447265625
],
[
0.118408203125,
0.58447265625,
0.489501953125,
0.58447265625,
0.489501953125,
0.59716796875,
0.118408203125,
0.59716796875
],
[
0.51123046875,
0.5849609375,
0.8818359375,
0.5849609375,
0.8818359375,
0.5986328125,
0.51123046875,
0.5986328125
],
[
0.12005615234375,
0.60107421875,
0.489501953125,
0.60107421875,
0.489501953125,
0.6142578125,
0.12005615234375,
0.6142578125
],
[
0.51318359375,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.12005615234375,
0.6162109375,
0.490234375,
0.6162109375,
0.490234375,
0.6298828125,
0.12005615234375,
0.6298828125
],
[
0.5146484375,
0.6181640625,
0.88330078125,
0.6181640625,
0.88330078125,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.12005615234375,
0.6328125,
0.489501953125,
0.6328125,
0.489501953125,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.51318359375,
0.6328125,
0.8818359375,
0.6328125,
0.8818359375,
0.646484375,
0.51318359375,
0.646484375
],
[
0.118408203125,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.5146484375,
0.65087890625,
0.88330078125,
0.65087890625,
0.88330078125,
0.66455078125,
0.5146484375,
0.66455078125
],
[
0.12005615234375,
0.66650390625,
0.48779296875,
0.66650390625,
0.48779296875,
0.6767578125,
0.12005615234375,
0.6767578125
],
[
0.5146484375,
0.66650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.6796875,
0.5146484375,
0.6796875
],
[
0.118408203125,
0.68115234375,
0.490234375,
0.68115234375,
0.490234375,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.51123046875,
0.681640625,
0.88330078125,
0.68115234375,
0.88330078125,
0.6953125,
0.51123046875,
0.69677734375
],
[
0.118408203125,
0.69677734375,
0.490234375,
0.69677734375,
0.490234375,
0.71044921875,
0.118408203125,
0.71044921875
],
[
0.51318359375,
0.69873046875,
0.88330078125,
0.69873046875,
0.88330078125,
0.7119140625,
0.51318359375,
0.7119140625
],
[
0.1175537109375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.509765625,
0.71435546875,
0.88330078125,
0.71337890625,
0.88330078125,
0.72802734375,
0.509765625,
0.728515625
],
[
0.118408203125,
0.72998046875,
0.491943359375,
0.72998046875,
0.491943359375,
0.74365234375,
0.118408203125,
0.74365234375
],
[
0.51318359375,
0.73095703125,
0.88330078125,
0.73095703125,
0.88330078125,
0.7451171875,
0.51318359375,
0.7451171875
],
[
0.12005615234375,
0.7451171875,
0.489501953125,
0.7451171875,
0.489501953125,
0.7587890625,
0.12005615234375,
0.7587890625
],
[
0.51318359375,
0.74755859375,
0.73046875,
0.74755859375,
0.73046875,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.118408203125,
0.76123046875,
0.489501953125,
0.76123046875,
0.489501953125,
0.775390625,
0.118408203125,
0.775390625
],
[
0.52880859375,
0.76416015625,
0.8818359375,
0.765625,
0.8818359375,
0.77978515625,
0.52880859375,
0.779296875
],
[
0.12176513671875,
0.779296875,
0.310546875,
0.779296875,
0.310546875,
0.7890625,
0.12176513671875,
0.7890625
],
[
0.51318359375,
0.7822265625,
0.88330078125,
0.7822265625,
0.88330078125,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.51318359375,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51318359375,
0.8125
],
[
0.12005615234375,
0.80810546875,
0.162841796875,
0.80810546875,
0.162841796875,
0.8193359375,
0.12005615234375,
0.8193359375
],
[
0.154541015625,
0.8056640625,
0.273681640625,
0.80712890625,
0.273681640625,
0.82177734375,
0.154541015625,
0.82080078125
],
[
0.51123046875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.845703125,
0.51318359375,
0.845703125
],
[
0.118408203125,
0.845703125,
0.491943359375,
0.8466796875,
0.491943359375,
0.86083984375,
0.118408203125,
0.85986328125
],
[
0.509765625,
0.845703125,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.509765625,
0.85986328125
],
[
0.118408203125,
0.86328125,
0.49365234375,
0.86328125,
0.49365234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.509765625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.509765625,
0.876953125
],
[
0.118408203125,
0.87890625,
0.491943359375,
0.87890625,
0.491943359375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.7138671875,
0.8955078125,
0.7138671875,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"1.1 Challenges",
"where the convolution operation combines both",
"textual and visual features of the context. Then",
"IE from VRDs is a challenging task, and the dif-",
"the graph embeddings are combined with text em-",
"ficulties mainly arise from how to effectively in-",
"beddings to feed into a standard BiLSTM for in-",
"corporate visual cues from the document and the",
"scalability of the task.",
"formation extraction.",
"First, text alone is not adequate to represent the",
"Extensive experiments have been conducted",
"semantic meaning in VRDs, and the contexts of",
"to show our method outperforms BiLSTM-CRF",
"the texts are usually expressed in visual cues. For",
"baselines by significant margins, on two real-",
"example, there might be multiple dates in the pur-",
"world datasets. Additionally, ablation studies are",
"chase receipts. However, it is up to the visual part",
"also performed to evaluate the effectiveness of",
"of the model to distinguish between the Invoice",
"each component of our model. Furthermore, we",
"Date, Transaction Date, and Due Date. Another",
"also provide analysis and intuitions on why and",
"example is the tax amount in value-added tax in-",
"how individual components work in our experi-",
"voice, as shown in Figure 1(b). There are multiple",
"ments.",
"\"money\" entities in the document, and there is a",
"lack of any textual context to determine which one",
"Related Works",
"is tax amount. To extract the tax amount correctly,",
"we have to leverage the (relative) position of the",
"Our work is inspired by recent research in the area",
"text segment and visual features of the document",
"of graph convolution and information extraction.",
"in general.",
"Template matching based algorithms (Chiti-",
"2.1 Graph Convolution Network",
"cariu et al., 2013; Dengel and Klein, 2002; Schus-",
"ter et al., 2013) utilize visual features of the doc-",
"Neural network architectures such as CNN and",
"ument to extract entities; however, we argue that",
"RNNs have demonstrated huge success on many",
"they are mostly not scalable for the task in real-",
"artificial intelligence tasks where the underly-",
"world business settings. There are easily thou-",
"ing data has grid-like or sequential structure",
"sands of vendors on the market, and the templates",
"(Krizhevsky et al., 2012; Kim et al., 2016; Kim,",
"of purchase receipts from each vendor are not the",
"2014). Recently, there is a surge of interest in",
"same.Thousands oftemplatesneed tobe created",
"studying the neural network structure operating on",
"and maintained in this single scenario.Itreguires",
"graphs (Kipf and Welling, 2016; Hamilton et al.,",
"substantial efforts to update the template and make",
"2017), since much data in the real world is natu-",
"sure it's not conflicting with the rest every time a",
"rally represented as graphs. Many works attempt",
"new template comes in, and the process is error-",
"to generalize convolution on the graph structure",
"prone. Besides, user uploaded pictures introduce",
"Some use a spectrum based approach where the",
"another dimension of variance from the template.",
"learned model depends on the structure of the",
"An example is shown in Figure l(b). Value-added",
"graph. As a result, the approach does not work",
"tax invoice is a nation-widetax document,and the",
"well on dynamic graph structures. The others de-",
"layout is fixed. However, pictures taken by users",
"fine convolution directly on the graph (Velickovic",
"are usually distorted, often blurred and sometimes",
"et al., 2017; Hamilton et al., 2017; Xu et al., 2018;",
"contain interfering objects in the image. A simple",
"Johnson et al., 2018; Duvenaud et al., 2015). We",
"template-based system performs poorly in such a",
"follow the latter approach in our work to model the",
"scenario, while sophisticated rules require signif-",
"text segment graph of VRDs.",
"icant engineering efforts for each scenario, which",
"Different from existing works,this paper in-",
"webelieve is not scalable.",
"troduces explicit edge embeddings into the graph",
"convolution network, which models the rela-",
"1.2",
" Contributions",
"tionship between vertices directly.Similar to",
"In this paper, we present a novel method for IE",
"(Velickovic et al., 2017), we apply self-attention",
"from VRDs. The method first computes graph em-",
"(Vaswani et al., 2017) to define convolution on",
"beddings for each text segment in the document",
" variable-sized neighbors, and the approach is com-",
"using graph convolution. The graph embeddings",
"putationally effcient since the operation is paral-",
"represent the context of the current text segment",
"lelizable across node pairs."
] |
{
"arxiv_id": "1903.11279",
"language": "en",
"page_idx": 1
}
|
||
1903.11279-en-2
|
1903.11279-en-2.jpg
|
## 2.2 Information Extraction
Recently, significant progress has been made in information extraction from unstructured or semistructured text. However, most works focus on plain text documents (Peng et al., 2017; Lample et al., 2016; Ma and Hovy, 2016; Chiu and Nichols, 2016). For information extraction from VRDs, (Palm et al., 2017) which uses a recurrent neural network (RNN) to extract entities of interest from VRDs (invoices) is the closest to our work, but does not take visual features into account. Besides, some of the studies (d'Andecy et al., 2018; Medvet et al., 2011; Rusinol et al., 2013) in the area of document understanding deal with a similar problem to our work, and explore using visual features to aid text extraction from VRDs; however, approaches they proposed are based on a large amount of heuristic knowledge and human-designed features, as well as limited in known templates, which are not scalable in real-world business settings. We also acknowledge a concurrent work of (Katti et al., 2018), which models 2-D document using convolution networks. However, there are several key differences. Our neural network architecture is graphbased, and our model operates on text segments instead of characters as in (Katti et al., 2018).
Besides, information extraction based on the graph structure has been developed most recently. (Peng et al., 2017; Song et al., 2018) present a graph LSTM to capture various dependencies among the input words and (Wang et al., 2018) designs a novel graph schema to extract entities and relations jointly. However, their models are not concerned with visual information directly.
### 3 Model Architecture
This section describes the document model and the architecture of our proposed model. Our model first encodes each text segment in the document into graph embedding, using multiple layers of graph convolution. The embedding represents the information in the text segment given its visual and textual context. By visual context, we refer to the layout of the document and relative positions of the individual segment to other segments. Textual context is the aggregate of text information in the document overall; our model learns to assign higher weights on texts from neighbor segments. Then we combine the graph embeddings with text embeddings and apply a standard BiLSTM-CRF
<FigureHere>
Figure 2: Document graph. Every node in the graph is
fully connected to each other.
model for entity extraction.
## 3.1 Document Modeling
We model each document as a graph of text segments (see Figure 2), where each text segment is comprised of the position of the segment and the text within it. The graph is comprised of nodes that represent text segments, and edges that represent visual dependencies, such as relative shapes and distance, between two nodes. Text segments are generated using an in-house Optical Character Recognition (OCR) system.
Mathematically, a document $\mathcal{D}$ is a tuple $( T, E) $, where $T$ = $\{ t_{1}, t_{2}, \cdots , t_{n}\} , t_{i}$ $\in$ $\mathcal{T}$ is a set of $n$ text boxes/nodes, $R=$ $\{r_{i1},r_{i2},\cdots,r_{ij}\},r_{ij}\in\mathcal{R}$ is a set of edges, and $E=T\times R\times T$ is a set of directed edges of the form $(t_i,r_{ij},t_j)$ where $t_i,t_j\in T$ and $r_ij\in R.$ In our experiments, every node is connected to each other.
## 3.2 Feature Extraction
For node $t_i$, we calculate node embedding t$_i$ using a single layer Bi-LSTM (Schuster and Paliwal, 1997) to extract features from the text content in the segment.
Edge embedding between node $t_i$ and node $t_j$ is
defined as follows,
(1)
$$\mathbf{r}_{ij}=[x_{ij},y_{ij},\frac{w_i}{h_i},\frac{h_j}{h_i},\frac{w_j}{h_i}],$$
where $x_ij$ and $y_{ij}$ are horizontal and vertical distance between the two text boxes respectively, and $w_i$ and $h_i$ are the width and height of the corresponding text box. The third, fourth and fifth value of the embedding are the aspect ratio of node $t_i$, relative height, and width of node $t_j$ respectively. Empirically, a visual distance between two segments is an important feature. For example, in general, the positions of relevant information are
|
[
[
0.5322265625,
0.0706787109375,
0.67919921875,
0.07183837890625,
0.67919921875,
0.08489990234375,
0.5322265625,
0.0843505859375
],
[
0.12005615234375,
0.07659912109375,
0.345947265625,
0.07659912109375,
0.345947265625,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.75732421875,
0.0736083984375,
0.84619140625,
0.0736083984375,
0.84619140625,
0.0843505859375,
0.75732421875,
0.0843505859375
],
[
0.1175537109375,
0.097412109375,
0.489501953125,
0.0968017578125,
0.489501953125,
0.11102294921875,
0.1175537109375,
0.11224365234375
],
[
0.118408203125,
0.114013671875,
0.489501953125,
0.114013671875,
0.489501953125,
0.127685546875,
0.118408203125,
0.127685546875
],
[
0.118408203125,
0.1300048828125,
0.490234375,
0.1300048828125,
0.490234375,
0.1436767578125,
0.118408203125,
0.1436767578125
],
[
0.1175537109375,
0.1455078125,
0.489501953125,
0.144287109375,
0.489501953125,
0.1591796875,
0.1175537109375,
0.1602783203125
],
[
0.12005615234375,
0.162109375,
0.489501953125,
0.162109375,
0.489501953125,
0.177001953125,
0.12005615234375,
0.177001953125
],
[
0.12005615234375,
0.1787109375,
0.489501953125,
0.1787109375,
0.489501953125,
0.1923828125,
0.12005615234375,
0.1923828125
],
[
0.51318359375,
0.1859130859375,
0.88330078125,
0.18701171875,
0.88330078125,
0.201904296875,
0.51318359375,
0.20068359375
],
[
0.118408203125,
0.193603515625,
0.48779296875,
0.193603515625,
0.48779296875,
0.207275390625,
0.118408203125,
0.207275390625
],
[
0.51318359375,
0.201904296875,
0.7138671875,
0.201904296875,
0.7138671875,
0.215576171875,
0.51318359375,
0.215576171875
],
[
0.118408203125,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.118408203125,
0.223876953125
],
[
0.118408203125,
0.226806640625,
0.489501953125,
0.226806640625,
0.489501953125,
0.240478515625,
0.118408203125,
0.240478515625
],
[
0.12005615234375,
0.2423095703125,
0.489501953125,
0.2423095703125,
0.489501953125,
0.257080078125,
0.12005615234375,
0.257080078125
],
[
0.51318359375,
0.243408203125,
0.71533203125,
0.243408203125,
0.71533203125,
0.257080078125,
0.51318359375,
0.257080078125
],
[
0.118408203125,
0.2587890625,
0.48779296875,
0.2587890625,
0.48779296875,
0.2724609375,
0.118408203125,
0.2724609375
],
[
0.118408203125,
0.273681640625,
0.486083984375,
0.273681640625,
0.486083984375,
0.287353515625,
0.118408203125,
0.287353515625
],
[
0.51123046875,
0.270751953125,
0.71875,
0.271484375,
0.71875,
0.2861328125,
0.51123046875,
0.28515625
],
[
0.12005615234375,
0.290283203125,
0.489501953125,
0.290283203125,
0.489501953125,
0.303955078125,
0.12005615234375,
0.303955078125
],
[
0.51123046875,
0.29150390625,
0.8818359375,
0.29345703125,
0.8818359375,
0.30712890625,
0.51123046875,
0.3046875
],
[
0.12005615234375,
0.30712890625,
0.490234375,
0.30712890625,
0.490234375,
0.320556640625,
0.12005615234375,
0.320556640625
],
[
0.51318359375,
0.308837890625,
0.88330078125,
0.308837890625,
0.88330078125,
0.322509765625,
0.51318359375,
0.322509765625
],
[
0.118408203125,
0.32373046875,
0.490234375,
0.322509765625,
0.490234375,
0.33740234375,
0.118408203125,
0.33837890625
],
[
0.51318359375,
0.32470703125,
0.88330078125,
0.32470703125,
0.88330078125,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.1175537109375,
0.33837890625,
0.490234375,
0.339111328125,
0.490234375,
0.35400390625,
0.1175537109375,
0.352783203125
],
[
0.51123046875,
0.339111328125,
0.884765625,
0.34033203125,
0.884765625,
0.355224609375,
0.51123046875,
0.35400390625
],
[
0.118408203125,
0.355712890625,
0.489501953125,
0.355712890625,
0.489501953125,
0.369384765625,
0.118408203125,
0.369384765625
],
[
0.51123046875,
0.355712890625,
0.880859375,
0.355712890625,
0.880859375,
0.369384765625,
0.51123046875,
0.369384765625
],
[
0.118408203125,
0.371826171875,
0.489501953125,
0.371826171875,
0.489501953125,
0.385498046875,
0.118408203125,
0.385498046875
],
[
0.51123046875,
0.371826171875,
0.88330078125,
0.372314453125,
0.88330078125,
0.38720703125,
0.51123046875,
0.385986328125
],
[
0.118408203125,
0.38720703125,
0.490234375,
0.38720703125,
0.490234375,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.51318359375,
0.3896484375,
0.884765625,
0.3896484375,
0.884765625,
0.402587890625,
0.51318359375,
0.402587890625
],
[
0.118408203125,
0.40380859375,
0.48779296875,
0.40380859375,
0.48779296875,
0.41748046875,
0.118408203125,
0.41748046875
],
[
0.51318359375,
0.405029296875,
0.884765625,
0.405029296875,
0.884765625,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.12005615234375,
0.419189453125,
0.489501953125,
0.419189453125,
0.489501953125,
0.432861328125,
0.12005615234375,
0.432861328125
],
[
0.51318359375,
0.421630859375,
0.724609375,
0.421630859375,
0.724609375,
0.435302734375,
0.51318359375,
0.435302734375
],
[
0.12005615234375,
0.435302734375,
0.489501953125,
0.435302734375,
0.489501953125,
0.449462890625,
0.12005615234375,
0.449462890625
],
[
0.52880859375,
0.435791015625,
0.88330078125,
0.43701171875,
0.88330078125,
0.451904296875,
0.52880859375,
0.45068359375
],
[
0.12005615234375,
0.451904296875,
0.48779296875,
0.451904296875,
0.48779296875,
0.465576171875,
0.12005615234375,
0.465576171875
],
[
0.5146484375,
0.453125,
0.85400390625,
0.454833984375,
0.85400390625,
0.4697265625,
0.5146484375,
0.46728515625
],
[
0.85986328125,
0.457763671875,
0.88330078125,
0.457763671875,
0.88330078125,
0.465576171875,
0.85986328125,
0.465576171875
],
[
0.12005615234375,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482177734375,
0.12005615234375,
0.482177734375
],
[
0.51318359375,
0.47021484375,
0.85205078125,
0.47021484375,
0.85205078125,
0.48388671875,
0.51318359375,
0.48388671875
],
[
0.118408203125,
0.482666015625,
0.489501953125,
0.48388671875,
0.489501953125,
0.498779296875,
0.118408203125,
0.49755859375
],
[
0.5146484375,
0.486328125,
0.88330078125,
0.485107421875,
0.88330078125,
0.49951171875,
0.5146484375,
0.50048828125
],
[
0.118408203125,
0.50048828125,
0.45166015625,
0.50048828125,
0.45166015625,
0.51416015625,
0.118408203125,
0.51416015625
],
[
0.5146484375,
0.501953125,
0.88330078125,
0.501953125,
0.88330078125,
0.515625,
0.5146484375,
0.515625
],
[
0.1376953125,
0.51708984375,
0.489501953125,
0.51708984375,
0.489501953125,
0.53076171875,
0.1376953125,
0.53076171875
],
[
0.51318359375,
0.5185546875,
0.88330078125,
0.5185546875,
0.88330078125,
0.533203125,
0.51318359375,
0.533203125
],
[
0.12005615234375,
0.53369140625,
0.484375,
0.53369140625,
0.484375,
0.54736328125,
0.12005615234375,
0.54736328125
],
[
0.51123046875,
0.53369140625,
0.88330078125,
0.533203125,
0.88330078125,
0.54736328125,
0.51123046875,
0.548828125
],
[
0.118408203125,
0.54736328125,
0.489501953125,
0.548828125,
0.489501953125,
0.5634765625,
0.118408203125,
0.5625
],
[
0.51123046875,
0.5498046875,
0.55908203125,
0.5498046875,
0.55908203125,
0.56396484375,
0.51123046875,
0.56396484375
],
[
0.118408203125,
0.5654296875,
0.489501953125,
0.56396484375,
0.489501953125,
0.5791015625,
0.118408203125,
0.580078125
],
[
0.118408203125,
0.58056640625,
0.489501953125,
0.58056640625,
0.489501953125,
0.59423828125,
0.118408203125,
0.59423828125
],
[
0.5146484375,
0.57763671875,
0.7060546875,
0.57763671875,
0.7060546875,
0.59130859375,
0.5146484375,
0.59130859375
],
[
0.12005615234375,
0.59716796875,
0.490234375,
0.59716796875,
0.490234375,
0.61083984375,
0.12005615234375,
0.61083984375
],
[
0.51123046875,
0.5986328125,
0.8818359375,
0.599609375,
0.8818359375,
0.6142578125,
0.51123046875,
0.61328125
],
[
0.1175537109375,
0.6123046875,
0.490234375,
0.61328125,
0.490234375,
0.62744140625,
0.1175537109375,
0.626953125
],
[
0.51318359375,
0.6162109375,
0.8818359375,
0.6162109375,
0.8818359375,
0.6298828125,
0.51318359375,
0.6298828125
],
[
0.1175537109375,
0.6298828125,
0.435791015625,
0.6298828125,
0.435791015625,
0.6435546875,
0.1175537109375,
0.6435546875
],
[
0.515625,
0.6318359375,
0.884765625,
0.6318359375,
0.884765625,
0.6455078125,
0.515625,
0.6455078125
],
[
0.5146484375,
0.6494140625,
0.60693359375,
0.6494140625,
0.60693359375,
0.6611328125,
0.5146484375,
0.6611328125
],
[
0.118408203125,
0.65673828125,
0.32080078125,
0.65771484375,
0.32080078125,
0.6728515625,
0.118408203125,
0.67138671875
],
[
0.5322265625,
0.6650390625,
0.88330078125,
0.6650390625,
0.88330078125,
0.6787109375,
0.5322265625,
0.6787109375
],
[
0.12005615234375,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.51123046875,
0.6796875,
0.65576171875,
0.68115234375,
0.65576171875,
0.6953125,
0.51123046875,
0.69482421875
],
[
0.118408203125,
0.70166015625,
0.48779296875,
0.70068359375,
0.48779296875,
0.71533203125,
0.118408203125,
0.71630859375
],
[
0.118408203125,
0.71728515625,
0.486083984375,
0.71728515625,
0.486083984375,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.59033203125,
0.7216796875,
0.80419921875,
0.736328125,
0.80029296875,
0.76123046875,
0.5859375,
0.74658203125
],
[
0.705078125,
0.71875,
0.80029296875,
0.7216796875,
0.79833984375,
0.7392578125,
0.70361328125,
0.736328125
],
[
0.118408203125,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.85791015625,
0.72998046875,
0.884765625,
0.72998046875,
0.884765625,
0.74365234375,
0.85791015625,
0.74365234375
],
[
0.118408203125,
0.75,
0.486083984375,
0.75,
0.486083984375,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.53076171875,
0.765625,
0.8818359375,
0.76416015625,
0.8818359375,
0.779296875,
0.53076171875,
0.77978515625
],
[
0.12005615234375,
0.7822265625,
0.490234375,
0.7822265625,
0.490234375,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51123046875,
0.7822265625,
0.88330078125,
0.78076171875,
0.88330078125,
0.7958984375,
0.51123046875,
0.79638671875
],
[
0.118408203125,
0.798828125,
0.490234375,
0.798828125,
0.490234375,
0.8125,
0.118408203125,
0.8125
],
[
0.51123046875,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.51123046875,
0.8125
],
[
0.1175537109375,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.1175537109375,
0.82763671875
],
[
0.51318359375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.845703125,
0.118408203125,
0.845703125
],
[
0.509765625,
0.830078125,
0.88330078125,
0.83056640625,
0.88330078125,
0.845703125,
0.509765625,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.491943359375,
0.84716796875,
0.491943359375,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.51123046875,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51123046875,
0.8623046875,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.51123046875,
0.876953125
],
[
0.118408203125,
0.87890625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.89453125,
0.489501953125,
0.8935546875,
0.489501953125,
0.90771484375,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"Text Segments of Document",
"2.2Information Extraction",
"DocumentGraph",
"Recently, significant progress has been made in in-",
"formation extraction from unstructured or semi-",
"structured text. However, most works focus on",
"plain text documents (Peng et al., 2017; Lam-",
"ple et al., 2016; Ma and Hovy, 2016; Chiu and",
"Nichols, 2016). For information extraction from",
"Figure 2: Document graph. Every node in the graph is",
"VRDs, (Palm et al., 2017) which uses a recurrent",
"fully connected to each other.",
"neural network (RNN) to extract entities of in-",
"terest from VRDs (invoices) is the closest to our",
"work, but does not take visual features into ac-",
"model for entity extraction.",
"count. Besides, some of the studies (d'Andecy",
"et al.. 2018; Medvet et al.. 2011; Rusinol et al..",
"3.1 Document Modeling",
"2013) in the area of document understanding deal",
"We model each document as a graph of text seg-",
"with a similar problem to our work, and explore",
"ments (see Figure 2), where each text segment is",
"using visual features to aid text extraction from",
"comprised of the position of the segment and the",
"VRDs;however,approaches they proposed are",
"text within it. The graph is comprised of nodes",
"based on a large amount of heuristic knowledge",
"that represent text segments, and edges that repre-",
"and human-designed features, as well as limited",
"sent visual dependencies, such as relative shapes",
"in known templates, which are not scalable in",
"and distance, between two nodes.Text segments",
"real-world business settings. We also acknowl-",
"are generated using an in-house Optical Character",
"edge a concurrent work of (Katti et al., 2018),",
"Recognition (OCR) system.",
"which models 2-D document using convolution",
"Mathematically, a document D is a tuple",
"networks. However, there are several key differ-",
"(T,E),where T={ti,t2,,tn),ti",
"E",
"ences. Our neural network architecture is graph-",
"T is a set of n text boxes/nodes, R",
"based, and our model operates on text segments",
"[rii, ri2, ...,riil, rii E R is a set of edges, and",
"instead of characters as in (Katti et al., 2018)",
"E = T × R × T is a set of directed edges of the",
"Besides, information extraction based on the",
"form (ti, rij,t,) where ti,t, E T and rij E R. In",
"graph structure has been developed most recently",
"our experiments, every node is connected to each",
"(Peng et al., 2017; Song et al., 2018) present",
"other.",
"a graph LSTM to capture various dependencies",
"among the input words and (Wang et al., 2018) de-",
"3.2Feature Extraction",
"signs a novel graph schema to extract entities and",
"For node ti, we calculate node embedding t, us-",
"relations jointly. However, their models are not",
"ing a single layer Bi-LSTM (Schuster and Paliwal,",
"concerned with visual information directly.",
"1997) to extract features from the text content in",
"the segment.",
"3Model Architecture",
"Edge embedding between node ti and node t, is",
"This section describes the document model and the",
"defined as follows,",
"architecture of our proposed model. Our model",
"first encodes each text segment in the document",
"=[,ih]",
"wi hj wij",
"into graph embedding, using multiple layers of",
"(1)",
"graph convolution. The embedding represents the",
"information in the text segment given its visual and",
"where ii and yii are horizontal and vertical dis-",
"textual context. By visual context, we refer to the",
"tance between the two text boxes respectively, and",
"layout of the document and relative positions of",
"Wi and hi are the width and height of the corre-",
"the individual segment to other segments. Tex-",
"sponding text box. The third, fourth and fifth value",
"tual context is the aggregate of text information in",
"of the embedding are the aspect ratio of node ti,",
"the document overall; our model learns to assign",
"relative height, and width of node t, respectively.",
"higher weights on texts from neighbor segments.",
"Empirically, a visual distance between two seg-",
"Then we combine the graph embeddings with text",
"ments is an important feature. For example, in",
"embeddings and apply a standard BiLSTM-CRF",
"general, the positions of relevant information are"
] |
{
"arxiv_id": "1903.11279",
"language": "en",
"page_idx": 2
}
|
||
1903.11279-en-3
|
1903.11279-en-3.jpg
|
<FigureHere>
Figure 3: Graph convolution of document graph. Convolution is defined on node-edge-node triplets $(t_i,r_{ij},t_j).$
Each layer produces new embeddings for both nodes and edges.
is to compute the output hidden representation of each node by attending to its neighbors. In its most general form, each node can attend to all the other nodes, assuming a fully connected graph.
Concretely the output embedding $\mathbf{t}_i^{\prime}$ of the layer
for node $t_i$ is computed by,
closer in one document, such as the key and value of an entity. Moreover, the shape of the text segment plays a critical role in representing semantic meanings. For example, the length of the text segment which has address information is usually longer than that of one which has a buyer name. Therefore, we use edge embedding to encode information regarding the visual distance between two segments, the shape of the source node, and the relative size of the destination node.
To summarize, node embedding encodes textual features, while edge embedding primarily represents visual features.
(3)
$$\mathbf{t}_i'=\sigma(\sum_{j\in\{1,\cdots,n\}}\alpha_{ij}\mathbf{h}_{ij}),$$
where $\alpha_{ij}$ are the attention coefficients, and $\sigma$ is an activation function. In our experiments, the attention mechanism is designed as the follows,
## 33 Graph Convolution
$$\alpha_{ij}=\frac{\exp(\mathrm{LeakyRelu}(\mathbf{w}_a^T\mathbf{h}_{ij}))}{\sum_{j\in\{1,\cdots,n\}}\exp(\mathrm{LeakyRelu}(\mathbf{w}_a^T\mathbf{h}_{ij}))},$$
(4)
Graph convolution is applied to compute visual text embeddings of text segments in the graph, as shown in Figure 3. Different from existing works, we define convolution on the node-edgenode triplets $(t_i,r_{ij},t_j)$ instead of on the node alone. We compare the performances of the models using nodes only and node-edge-node triplets in Section 5.3. For node $t_i$,we extract features h$_ij$ for each neighbour $t_j$ using a multi-layer perceptron (MLP) network,
where w$_a$ is a shared attention weight vector. We apply the LeakyRelu activation function to avoid the“dying Relu”problem and to increase the“contrast” of the attention coefficients potentially.
The edge embedding output of the graph convo-
lution layer is defined as,
(2)
(5)
$$\mathbf{h}_{ij}=g(\mathbf{t}_i,\mathbf{r}_{ij},\mathbf{t}_j)=\mathbf{MLP}([\mathbf{t}_i\|\mathbf{r}_{ij}\|\mathbf{t}_j]),$$
$$\mathbf{r}_{ij}'=\mathrm{MLP}(\mathbf{h}_{ij}).$$
Outputs $\mathbf t_{i}^{\prime},\mathbf r_{ij}^{\prime}$ are fed as inputs to the next layer of graph convolution (as computed in equation 2) or network modules for downstream tasks.
3.4 BiLSTM-CRF with Graph Embeddings
where $\|$ is the concatenate operation. There are several benefits of using this triplet feature set. First, it combines visual features directly into the neighbor representation. Furthermore, the information of the current node is copied across the neighbors. As a result, the neighbor features can potentially learn where to attend given the current node.
In our model, graph convolution is defined based on the self-attention mechanism. The idea
We combine graph embeddings with token embeddings and feed them into standard BiLSTM-CRF for entity extraction. As illustrated in Figure 4, visual text embedding generated from the graph convolution layers of the current text segment is
|
[
[
0.705078125,
0.0789794921875,
0.7119140625,
0.08197021484375,
0.7080078125,
0.0872802734375,
0.7001953125,
0.0843505859375
],
[
0.69775390625,
0.09857177734375,
0.71728515625,
0.09857177734375,
0.71728515625,
0.1134033203125,
0.69775390625,
0.1134033203125
],
[
0.7001953125,
0.1478271484375,
0.7138671875,
0.1478271484375,
0.7138671875,
0.15673828125,
0.7001953125,
0.15673828125
],
[
0.744140625,
0.201904296875,
0.76416015625,
0.201904296875,
0.76416015625,
0.208984375,
0.744140625,
0.208984375
],
[
0.1175537109375,
0.2392578125,
0.88330078125,
0.2410888671875,
0.88330078125,
0.2587890625,
0.1175537109375,
0.257080078125
],
[
0.12005615234375,
0.257080078125,
0.5498046875,
0.257080078125,
0.5498046875,
0.270751953125,
0.12005615234375,
0.270751953125
],
[
0.118408203125,
0.29638671875,
0.491943359375,
0.29638671875,
0.491943359375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51318359375,
0.297607421875,
0.884765625,
0.297607421875,
0.884765625,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.118408203125,
0.3134765625,
0.490234375,
0.3134765625,
0.490234375,
0.327880859375,
0.118408203125,
0.327880859375
],
[
0.51318359375,
0.314208984375,
0.884765625,
0.314208984375,
0.884765625,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.118408203125,
0.330078125,
0.490234375,
0.330078125,
0.490234375,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.51318359375,
0.330078125,
0.884765625,
0.330078125,
0.884765625,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.118408203125,
0.345703125,
0.491943359375,
0.345703125,
0.491943359375,
0.359375,
0.118408203125,
0.359375
],
[
0.51318359375,
0.345703125,
0.81689453125,
0.345703125,
0.81689453125,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.12005615234375,
0.3623046875,
0.489501953125,
0.3623046875,
0.489501953125,
0.3759765625,
0.12005615234375,
0.3759765625
],
[
0.5322265625,
0.3623046875,
0.88330078125,
0.3623046875,
0.88330078125,
0.3759765625,
0.5322265625,
0.3759765625
],
[
0.12005615234375,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.12005615234375,
0.39013671875
],
[
0.5146484375,
0.37890625,
0.71728515625,
0.37890625,
0.71728515625,
0.392578125,
0.5146484375,
0.392578125
],
[
0.12005615234375,
0.393798828125,
0.486083984375,
0.393798828125,
0.486083984375,
0.40673828125,
0.12005615234375,
0.40673828125
],
[
0.60546875,
0.40625,
0.66162109375,
0.407958984375,
0.66015625,
0.42041015625,
0.603515625,
0.418701171875
],
[
0.72314453125,
0.405029296875,
0.79248046875,
0.40673828125,
0.791015625,
0.4228515625,
0.72119140625,
0.42041015625
],
[
0.8564453125,
0.40625,
0.88671875,
0.40625,
0.88671875,
0.4228515625,
0.8564453125,
0.4228515625
],
[
0.12005615234375,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.12005615234375,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.12005615234375,
0.42333984375
],
[
0.66162109375,
0.42578125,
0.72705078125,
0.42578125,
0.72705078125,
0.43701171875,
0.66162109375,
0.43701171875
],
[
0.118408203125,
0.441162109375,
0.411376953125,
0.441162109375,
0.411376953125,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.53076171875,
0.448974609375,
0.88330078125,
0.448974609375,
0.88330078125,
0.4619140625,
0.53076171875,
0.4619140625
],
[
0.1368408203125,
0.457763671875,
0.489501953125,
0.457763671875,
0.489501953125,
0.471435546875,
0.1368408203125,
0.471435546875
],
[
0.51318359375,
0.465576171875,
0.88330078125,
0.465576171875,
0.88330078125,
0.478515625,
0.51318359375,
0.478515625
],
[
0.1175537109375,
0.473876953125,
0.48779296875,
0.474365234375,
0.48779296875,
0.4892578125,
0.1175537109375,
0.488037109375
],
[
0.51318359375,
0.48095703125,
0.86572265625,
0.48095703125,
0.86572265625,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.1175537109375,
0.4892578125,
0.273681640625,
0.490478515625,
0.273681640625,
0.5048828125,
0.1175537109375,
0.50439453125
],
[
0.1175537109375,
0.5166015625,
0.313232421875,
0.5166015625,
0.313232421875,
0.52978515625,
0.1175537109375,
0.52978515625
],
[
0.62646484375,
0.52099609375,
0.8134765625,
0.52099609375,
0.8134765625,
0.53515625,
0.62646484375,
0.53515625
],
[
0.118408203125,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.53076171875,
0.533203125,
0.57080078125,
0.533203125,
0.57080078125,
0.54443359375,
0.53076171875,
0.54443359375
],
[
0.57666015625,
0.5390625,
0.86572265625,
0.5361328125,
0.86572265625,
0.55419921875,
0.57666015625,
0.55712890625
],
[
0.118408203125,
0.552734375,
0.48779296875,
0.552734375,
0.48779296875,
0.56640625,
0.118408203125,
0.56640625
],
[
0.85791015625,
0.55810546875,
0.88330078125,
0.55810546875,
0.88330078125,
0.57080078125,
0.85791015625,
0.57080078125
],
[
0.118408203125,
0.568359375,
0.489501953125,
0.5693359375,
0.489501953125,
0.58447265625,
0.118408203125,
0.5830078125
],
[
0.12176513671875,
0.58740234375,
0.483642578125,
0.58740234375,
0.483642578125,
0.59716796875,
0.12176513671875,
0.59716796875
],
[
0.5322265625,
0.5830078125,
0.8818359375,
0.5830078125,
0.8818359375,
0.5966796875,
0.5322265625,
0.5966796875
],
[
0.12005615234375,
0.6015625,
0.489501953125,
0.6015625,
0.489501953125,
0.6162109375,
0.12005615234375,
0.6162109375
],
[
0.51318359375,
0.59716796875,
0.88330078125,
0.5986328125,
0.88330078125,
0.61328125,
0.51318359375,
0.6123046875
],
[
0.12005615234375,
0.61767578125,
0.48779296875,
0.61767578125,
0.48779296875,
0.6318359375,
0.12005615234375,
0.6318359375
],
[
0.51318359375,
0.6142578125,
0.880859375,
0.6142578125,
0.880859375,
0.62744140625,
0.51318359375,
0.62744140625
],
[
0.12005615234375,
0.6328125,
0.48779296875,
0.6328125,
0.48779296875,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.51318359375,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.1175537109375,
0.6484375,
0.48779296875,
0.6494140625,
0.48779296875,
0.66455078125,
0.1175537109375,
0.6630859375
],
[
0.51318359375,
0.64794921875,
0.5576171875,
0.64794921875,
0.5576171875,
0.662109375,
0.51318359375,
0.662109375
],
[
0.118408203125,
0.66650390625,
0.484375,
0.66650390625,
0.484375,
0.6796875,
0.118408203125,
0.6796875
],
[
0.5322265625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.5322265625,
0.677734375
],
[
0.118408203125,
0.681640625,
0.274658203125,
0.681640625,
0.274658203125,
0.6953125,
0.118408203125,
0.6953125
],
[
0.51318359375,
0.6787109375,
0.7021484375,
0.6796875,
0.7021484375,
0.69482421875,
0.51318359375,
0.693359375
],
[
0.6337890625,
0.7216796875,
0.76416015625,
0.7216796875,
0.76416015625,
0.7392578125,
0.6337890625,
0.7392578125
],
[
0.8564453125,
0.72265625,
0.88671875,
0.72265625,
0.88671875,
0.7392578125,
0.8564453125,
0.7392578125
],
[
0.1368408203125,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76416015625,
0.1368408203125,
0.76416015625
],
[
0.1427001953125,
0.7236328125,
0.489501953125,
0.7236328125,
0.489501953125,
0.73828125,
0.1427001953125,
0.73828125
],
[
0.5341796875,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.765625,
0.5341796875,
0.765625
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.118408203125,
0.78076171875,
0.48779296875,
0.78076171875,
0.48779296875,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.51318359375,
0.783203125,
0.8271484375,
0.7822265625,
0.8271484375,
0.79638671875,
0.51318359375,
0.79736328125
],
[
0.12005615234375,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.12005615234375,
0.8125
],
[
0.1175537109375,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.1175537109375,
0.82763671875
],
[
0.51318359375,
0.81005859375,
0.86962890625,
0.81005859375,
0.86962890625,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.118408203125,
0.83056640625,
0.490234375,
0.830078125,
0.490234375,
0.84423828125,
0.118408203125,
0.845703125
],
[
0.5146484375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.491943359375,
0.84716796875,
0.491943359375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.51123046875,
0.8466796875,
0.88330078125,
0.845703125,
0.88330078125,
0.85986328125,
0.51123046875,
0.86083984375
],
[
0.12176513671875,
0.865234375,
0.15869140625,
0.865234375,
0.15869140625,
0.87451171875,
0.12176513671875,
0.87451171875
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1368408203125,
0.87890625,
0.49365234375,
0.87890625,
0.49365234375,
0.8935546875,
0.1368408203125,
0.8935546875
],
[
0.51123046875,
0.87744140625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.51123046875,
0.892578125
],
[
0.12005615234375,
0.896484375,
0.490234375,
0.896484375,
0.490234375,
0.90673828125,
0.12005615234375,
0.90673828125
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"M",
"h13",
"015",
"ML",
"Figure 3: Graph convolution of document graph. Convolution is defined on node-edge-node triplets (ti, rij,t).",
"Each layer produces new embeddings for both nodes and edges.",
"closer in one document, such as the key and value",
"is to compute the output hidden representation of",
"of an entity. Moreover, the shape of the text seg-",
"each node by attending to its neighbors. In its most",
"ment plays a critical role in representing seman-",
"general form, each node can attend to all the other",
"tic meanings. For example, the length of the text",
"nodes, assuming a fully connected graph.",
"segment which has address information is usually",
"Concretely the output embedding t', of the layer",
"longer than that of one which has a buyer name.",
"for node t, is computed by,",
"Therefore, we use edge embedding to encode in-",
"t,=o(",
"Qijhi)",
"(3)",
"two segments, the shape of the source node, and",
"formation regarding the visual distance between",
"je(1,...,n)",
"the relative size of the destination node.",
"where Qii are the attention coefficients, and ",
"To summarize, node embedding encodes textual",
"is an activation function. In our experiments, the",
"features, while edge embedding primarily repre-",
"attention mechanism is designed as the follows,",
"sents visual features.",
"3.3Graph Convolution",
"exp(LeakyRelu(w hi)",
"Graph convolution is applied to compute visual",
"Qij=",
"Zje(1..n) exp(LeakyRelu(wT hig)",
"text embeddings of text segments in the graph,",
"(4)",
"as shown in Figure 3. Different from existing",
"works,wedefine convolutionon the node-edge",
"where wa is a shared attention weight vector",
"node triplets (ti, rij,tj) instead of on the node",
"We apply the LeakyRelu activation function to",
"alone. We compare the performances of the mod-",
"avoid the “dying Relu\" problem and to increase",
"els using nodes only and node-edge-node triplets",
"the “contrast\" of the attention coefficients poten-",
"in Section 5.3. For node ti, we extract features hij",
"tially.",
"for each neighbour t, using a multi-layer percep-",
"The edge embedding output of the graph convo-",
"tron (MLP) network,",
"lution layer is defined as,",
"re, = MLP(hij).",
"(5)",
"where ll is the concatenate operation. There",
"hi =g(ti,r,t,) =MLP([t,llrilt;l), (2)",
"Outputs t', ri, are fed as inputs to the next layer",
"are several benefits of using this triplet feature set.",
"of graph convolution (as computed in equation 2)",
"First, it combines visual features directly into the",
"or network modules for downstream tasks",
"neighbor representation. Furthermore, the infor-",
"mation of the current node is copied across the",
"3.4BiLSTM-CRF with Graph Embeddings",
"neighbors. As a result, the neighbor features can",
"We combine graph embeddings with token embed-",
"potentially learn where to attend given the current",
"dings and feed them into standard BiLSTM-CRF",
"node",
"for entity extraction. As illustrated in Figure 4,",
"In our model, graph convolution is defined",
"visual text embedding generated from the graph",
"based onthe self-attention mechanism.The idea",
"convolution layers of the current text segment is"
] |
{
"arxiv_id": "1903.11279",
"language": "en",
"page_idx": 3
}
|
||
1903.11279-en-4
|
1903.11279-en-4.jpg
|
<FigureHere>
Figure 4: BiLSTM-CRF with graph embeddings.
concatenated to each token embedding of the input sequence; Intuitively, graph embedding adds contextual information to the input sequence.
Formally, assume for node $t_i$, the input token sequence of the text segment is $x_1,x_2,\cdots,x_m$, and the graph embedding of the node is t$_i^{\prime}.$ The input embedding $\mathbf{u}_i$ is defined as,
(6)
$$\mathbf{u}_i=e(x_i)\|\mathbf{t}_i'$$
where $e$ is token embedding lookup function, and Word2Vec vectors are used as token embeddings in our experiments.
Then the input embeddings are fed into a BiLSTM network to be encoded, and the output is further passed to a fully connected network and then a CRF layer.
## 4 Model Supervision and Training
We build an annotation system to facilitate the labeling of the ground truth data. For each document, we label the values for each pre-defined entity, and their locations (bounding boxes). To generate training data, we frst identify the text segment each entity belongs to, and then we label the text in the segment according to IOB tagging format (Sang and Veenstra, 1999). We assign label O to all tokens in empty text segments.
Since human annotated bounding boxes cannot match OCR detected box exactly, we apply a simple heuristic to determine which text segment an entity belongs to based on overlap area. A text segment is considered to contain an entity if $A_overlap/\min(A_{annotator},A_{ocr})$ is bigger than a manually set threshold; Here $A_{annotator},A_{ocr},A_{overlap}$ are the area of the annotated box of the corresponding entity, the area of the ocr detected box and the area of the overlap between the two boxes respectively.
In our experiments, the graph convolution layers and BiLSTM-CRF extractors are trained jointly. Furthermore, to improve prediction accuracy, we add the segment classification task which classifies each text segment into a pre-defined tag as an auxiliary task and discuss the effect of multitask learning in Section 5.4.3. We feed the graph embedding of each text segment into a sigmoid classifer to predict the tag. Since the parameters of the graph convolution layers are shared across the extraction task and segment classification task, we employ a multi-task learning approach for model training. In multi-task training, the goal is to optimize the weighted sum of the two losses. In our experiments, the weight is determined using a principled approach as described in (Kendall et al., 2017). The idea is to adjust each task's relative weight in the loss function by considering task-dependant uncertainties.
## 5 Experiments
We apply our model for information extraction from two real-world datasets. They are ValueAdded Tax Invoices (VATI) and International Purchase Receipts (IPR).
### 5.1 Datasets Description
VATI consists of 3000 user-uploaded pictures and has 16 entities to exact. Example entities are the names of buyer/seller, date and tax amount. The invoices are in Chinese, and it has a fixed template since it is national standard invoice. However, there are many noises in the documents which include distracting objects in the image and skewed document orientation to name a few. IPR is a data set of 1500 scanned receipt documents in English which has 4 entities to exact (Invoice Number, Vendor Name, Payer Name and Total Amount). There exist 146 templates for the receipts. Variable templates introduce additional difficulties to the IPR dataset. For both datasets, we assign 70% of each dataset for training, 15% for validation and 15% for the test. The number of text segments varies per document from 100 to 300.
## 5.2 Baselines
We compare the performance of our system with two BiLSTM-CRF baselines. Baseline I applies BiLSTM-CRF to each text segment, where each text segment is an individual sentence. Baseline II applies the tagging model to the concatenated
|
[
[
0.252685546875,
0.0736083984375,
0.280517578125,
0.0736083984375,
0.280517578125,
0.08489990234375,
0.252685546875,
0.08489990234375
],
[
0.52880859375,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.08905029296875,
0.52880859375,
0.08905029296875
],
[
0.2418212890625,
0.0968017578125,
0.2939453125,
0.0968017578125,
0.2939453125,
0.10687255859375,
0.2418212890625,
0.10687255859375
],
[
0.5146484375,
0.0926513671875,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.51318359375,
0.10809326171875,
0.8818359375,
0.10809326171875,
0.8818359375,
0.122314453125,
0.51318359375,
0.122314453125
],
[
0.1292724609375,
0.1258544921875,
0.175537109375,
0.1258544921875,
0.175537109375,
0.13720703125,
0.1292724609375,
0.13720703125
],
[
0.51123046875,
0.12469482421875,
0.88330078125,
0.12353515625,
0.88330078125,
0.1383056640625,
0.51123046875,
0.1395263671875
],
[
0.5146484375,
0.141357421875,
0.88330078125,
0.141357421875,
0.88330078125,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.353515625,
0.15673828125,
0.46337890625,
0.157958984375,
0.46337890625,
0.169189453125,
0.353515625,
0.168701171875
],
[
0.51318359375,
0.15673828125,
0.8818359375,
0.15673828125,
0.8818359375,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.160400390625,
0.1728515625,
0.173828125,
0.1728515625,
0.173828125,
0.18115234375,
0.160400390625,
0.18115234375
],
[
0.359375,
0.1776123046875,
0.462646484375,
0.1776123046875,
0.462646484375,
0.189453125,
0.359375,
0.189453125
],
[
0.51318359375,
0.1728515625,
0.884765625,
0.1728515625,
0.884765625,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.51318359375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.51318359375,
0.203125
],
[
0.51123046875,
0.2037353515625,
0.88330078125,
0.204833984375,
0.88330078125,
0.2197265625,
0.51123046875,
0.218505859375
],
[
0.51318359375,
0.221435546875,
0.884765625,
0.221435546875,
0.884765625,
0.235107421875,
0.51318359375,
0.235107421875
],
[
0.1351318359375,
0.232177734375,
0.466796875,
0.2327880859375,
0.466796875,
0.2476806640625,
0.1351318359375,
0.2464599609375
],
[
0.51318359375,
0.2369384765625,
0.88330078125,
0.2369384765625,
0.88330078125,
0.25048828125,
0.51318359375,
0.25048828125
],
[
0.51123046875,
0.2529296875,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.51123046875,
0.267333984375
],
[
0.118408203125,
0.2724609375,
0.489501953125,
0.2724609375,
0.489501953125,
0.2861328125,
0.118408203125,
0.2861328125
],
[
0.51318359375,
0.268310546875,
0.88330078125,
0.26953125,
0.88330078125,
0.283935546875,
0.51318359375,
0.283203125
],
[
0.118408203125,
0.289306640625,
0.489501953125,
0.289306640625,
0.489501953125,
0.303955078125,
0.118408203125,
0.303955078125
],
[
0.5146484375,
0.2861328125,
0.884765625,
0.2861328125,
0.884765625,
0.2998046875,
0.5146484375,
0.2998046875
],
[
0.118408203125,
0.305908203125,
0.453369140625,
0.305908203125,
0.453369140625,
0.319580078125,
0.118408203125,
0.319580078125
],
[
0.5146484375,
0.3017578125,
0.880859375,
0.3017578125,
0.880859375,
0.3154296875,
0.5146484375,
0.3154296875
],
[
0.1368408203125,
0.32177734375,
0.490234375,
0.32177734375,
0.490234375,
0.3349609375,
0.1368408203125,
0.3349609375
],
[
0.5146484375,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.5146484375,
0.33203125
],
[
0.118408203125,
0.33740234375,
0.491943359375,
0.33740234375,
0.491943359375,
0.3515625,
0.118408203125,
0.3515625
],
[
0.51318359375,
0.33203125,
0.8818359375,
0.334228515625,
0.8818359375,
0.347412109375,
0.51318359375,
0.345703125
],
[
0.12005615234375,
0.352783203125,
0.490234375,
0.352783203125,
0.490234375,
0.36767578125,
0.12005615234375,
0.36767578125
],
[
0.51123046875,
0.3486328125,
0.880859375,
0.3486328125,
0.880859375,
0.3623046875,
0.51123046875,
0.3623046875
],
[
0.118408203125,
0.369384765625,
0.366943359375,
0.369384765625,
0.366943359375,
0.383056640625,
0.118408203125,
0.383056640625
],
[
0.5146484375,
0.366455078125,
0.79443359375,
0.366455078125,
0.79443359375,
0.380126953125,
0.5146484375,
0.380126953125
],
[
0.509765625,
0.392578125,
0.6572265625,
0.394287109375,
0.6572265625,
0.412109375,
0.509765625,
0.410400390625
],
[
0.2476806640625,
0.410888671875,
0.358642578125,
0.410888671875,
0.358642578125,
0.4287109375,
0.2476806640625,
0.4287109375
],
[
0.46337890625,
0.413330078125,
0.490234375,
0.413330078125,
0.490234375,
0.427001953125,
0.46337890625,
0.427001953125
],
[
0.51123046875,
0.419189453125,
0.88671875,
0.421630859375,
0.88671875,
0.439453125,
0.51123046875,
0.43701171875
],
[
0.1368408203125,
0.439453125,
0.489501953125,
0.439453125,
0.489501953125,
0.45361328125,
0.1368408203125,
0.45361328125
],
[
0.51123046875,
0.439453125,
0.8818359375,
0.439453125,
0.8818359375,
0.453125,
0.51123046875,
0.453125
],
[
0.118408203125,
0.454833984375,
0.489501953125,
0.4560546875,
0.489501953125,
0.47021484375,
0.118408203125,
0.4697265625
],
[
0.5146484375,
0.454833984375,
0.88330078125,
0.454833984375,
0.88330078125,
0.468505859375,
0.5146484375,
0.468505859375
],
[
0.118408203125,
0.47265625,
0.307373046875,
0.47265625,
0.307373046875,
0.486328125,
0.118408203125,
0.486328125
],
[
0.51318359375,
0.471435546875,
0.67529296875,
0.471435546875,
0.67529296875,
0.485107421875,
0.51318359375,
0.485107421875
],
[
0.1368408203125,
0.488037109375,
0.484375,
0.488037109375,
0.484375,
0.501953125,
0.1368408203125,
0.501953125
],
[
0.1175537109375,
0.5029296875,
0.490234375,
0.50439453125,
0.490234375,
0.5185546875,
0.1175537109375,
0.51708984375
],
[
0.5146484375,
0.49951171875,
0.71875,
0.49951171875,
0.71875,
0.51318359375,
0.5146484375,
0.51318359375
],
[
0.1175537109375,
0.51953125,
0.490234375,
0.52099609375,
0.490234375,
0.53515625,
0.1175537109375,
0.53369140625
],
[
0.5146484375,
0.521484375,
0.884765625,
0.521484375,
0.884765625,
0.53515625,
0.5146484375,
0.53515625
],
[
0.118408203125,
0.5361328125,
0.252685546875,
0.5361328125,
0.252685546875,
0.55029296875,
0.118408203125,
0.55029296875
],
[
0.51318359375,
0.5380859375,
0.884765625,
0.5380859375,
0.884765625,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.51318359375,
0.55419921875,
0.884765625,
0.55419921875,
0.884765625,
0.56787109375,
0.51318359375,
0.56787109375
],
[
0.1175537109375,
0.5625,
0.429931640625,
0.56396484375,
0.429931640625,
0.5791015625,
0.1175537109375,
0.57666015625
],
[
0.51123046875,
0.5693359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.5830078125,
0.51123046875,
0.5830078125
],
[
0.12005615234375,
0.5888671875,
0.490234375,
0.5888671875,
0.490234375,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.51123046875,
0.5859375,
0.884765625,
0.58740234375,
0.884765625,
0.60107421875,
0.51123046875,
0.599609375
],
[
0.118408203125,
0.60546875,
0.490234375,
0.60400390625,
0.490234375,
0.61767578125,
0.118408203125,
0.619140625
],
[
0.51318359375,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1175537109375,
0.62060546875,
0.490234375,
0.619140625,
0.490234375,
0.63427734375,
0.1175537109375,
0.63525390625
],
[
0.51318359375,
0.619140625,
0.880859375,
0.619140625,
0.880859375,
0.6298828125,
0.51318359375,
0.6298828125
],
[
0.1175537109375,
0.63720703125,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.1175537109375,
0.65185546875
],
[
0.51123046875,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64794921875,
0.51123046875,
0.64794921875
],
[
0.118408203125,
0.65283203125,
0.489501953125,
0.65380859375,
0.489501953125,
0.66748046875,
0.118408203125,
0.66650390625
],
[
0.51123046875,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.6630859375,
0.51123046875,
0.6630859375
],
[
0.118408203125,
0.66943359375,
0.489501953125,
0.66845703125,
0.489501953125,
0.68310546875,
0.118408203125,
0.68408203125
],
[
0.5146484375,
0.66748046875,
0.87890625,
0.66748046875,
0.87890625,
0.677734375,
0.5146484375,
0.677734375
],
[
0.12005615234375,
0.68603515625,
0.48779296875,
0.68603515625,
0.48779296875,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.5146484375,
0.68310546875,
0.8818359375,
0.68310546875,
0.8818359375,
0.69677734375,
0.5146484375,
0.69677734375
],
[
0.12005615234375,
0.70166015625,
0.489501953125,
0.70166015625,
0.489501953125,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.51318359375,
0.69775390625,
0.880859375,
0.69775390625,
0.880859375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71875,
0.3896484375,
0.71875,
0.3896484375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.5146484375,
0.71533203125,
0.88330078125,
0.71533203125,
0.88330078125,
0.728515625,
0.5146484375,
0.728515625
],
[
0.1368408203125,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74658203125,
0.1368408203125,
0.74658203125
],
[
0.51123046875,
0.72998046875,
0.88330078125,
0.73095703125,
0.88330078125,
0.7451171875,
0.51123046875,
0.74462890625
],
[
0.118408203125,
0.75048828125,
0.489501953125,
0.75048828125,
0.489501953125,
0.76416015625,
0.118408203125,
0.76416015625
],
[
0.51318359375,
0.74755859375,
0.88330078125,
0.74755859375,
0.88330078125,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.7626953125,
0.884765625,
0.76318359375,
0.884765625,
0.77783203125,
0.51318359375,
0.77685546875
],
[
0.118408203125,
0.783203125,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.118408203125,
0.79638671875
],
[
0.51318359375,
0.77978515625,
0.79248046875,
0.77978515625,
0.79248046875,
0.79345703125,
0.51318359375,
0.79345703125
],
[
0.118408203125,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.118408203125,
0.8125
],
[
0.1175537109375,
0.8134765625,
0.486083984375,
0.8134765625,
0.486083984375,
0.83056640625,
0.1175537109375,
0.83056640625
],
[
0.51318359375,
0.80810546875,
0.63232421875,
0.80810546875,
0.63232421875,
0.822265625,
0.51318359375,
0.822265625
],
[
0.1175537109375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.845703125,
0.1175537109375,
0.845703125
],
[
0.5146484375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.490234375,
0.84716796875,
0.490234375,
0.8623046875,
0.118408203125,
0.8623046875
],
[
0.51318359375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.491943359375,
0.86328125,
0.491943359375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.86083984375,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.51318359375,
0.8759765625
],
[
0.1175537109375,
0.87744140625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87744140625,
0.884765625,
0.892578125,
0.51318359375,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.3837890625,
0.8955078125,
0.3837890625,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"CRF",
"In our experiments, the graph convolution",
"FC Layer",
"layers and BiLSTM-CRF extractors are trained",
"jointly. Furthermore, to improve prediction accu-",
"BiLSTM",
"racy, we add the segment classification task which",
"classifies each text segment into a pre-defined tag",
">GraphEmbedding",
"as an auxiliary task and discuss the effect of multi-",
"ui",
"TokenEmbedding",
"task learning in Section 5.4.3. We feed the graph",
"embedding of each text segment into a sigmoid",
"classifier to predict the tag.Since the parame-",
"ters of the graph convolution layers are shared",
"Figure 4: BiLSTM-CRF with graph embeddings.",
"across the extraction task and segment classifi-",
"cation task, we employ a multi-task learning ap-",
"concatenated to each token embedding of the in-",
"proach for model training. In multi-task training,",
"put sequence; Intuitively, graph embedding adds",
"the goal is to optimize the weighted sum of the",
"contextual information to the input sequence.",
"two losses. In our experiments, the weight is de-",
"Formally, assume for node t, the input token",
"termined using a principled approach as described",
"sequence of the text segment is &1, &2,., m.",
"in (Kendall et al., 2017). The idea is to adjust each",
"and the graph embedding of the node is t'. The",
"task's relative weight in the loss function by con-",
"input embedding u, is defined as,",
"sidering task-dependant uncertainties.",
"5Experiments",
"ui = e(ci)llt)",
"(6)",
"We apply our model for information extraction",
"where e is token embedding lookup function,",
"from two real-world datasets. They are Value-",
"and Word2Vec vectors are used as token embed-",
"Added Tax Invoices (VATI) and International Pur-",
"dings in our experiments.",
"chase Receipts (IPR).",
"Then the input embeddings are fed into a Bi",
"LSTM network to be encoded, and the output is",
"5.1Datasets Description",
"further passed to a fully connected network and",
"VATI consists of 3000 user-uploaded pictures and",
"then a CRF layer.",
"has 16 entities to exact. Example entities are the",
"names of buyer/seller, date and tax amount. The",
"4Model Supervision and Training",
"invoices are in Chinese, and it has a fixed template",
"We build an annotation system to facilitate the la-",
"since it is national standard invoice. However,",
"beling of the ground truth data. For each docu-",
"there are many noises in the documents which in-",
"ment, we label the values for each pre-defined en-",
"cludedistracting obiects intheimageand skewed",
"tity, and their locations (bounding boxes). To gen-",
" document orientation to name a few. IPR is a data",
"erate training data, we first identify the text seg-",
"set of 1500 scanned receipt documents in English",
"ment each entity belongs to, and then we label the",
"which has 4 entities to exact (nvoice Number",
"text in the segment according to IOB tagging for-",
"Vendor Name, Payer Name and Total Amount).",
"mat (Sang and Veenstra, 1999). We assign label ",
"There exist 146 templates for the receipts. Vari-",
"to all tokens in empty text segments.",
"able templates introduce additional difficulties to",
"Since human annotated bounding boxes can-",
"the IPR dataset. For both datasets, we assign 70%",
"not match OCR detected box exactly, we ap-",
"of each dataset for training, 15% for validation and",
"ply a simple heuristic to determine which text",
"15% for the test. The number of text segments",
"segment an entity belongs to based on over-",
"varies per document from 100 to 300.",
"lap area. A text segment is considered to con-",
"tain an entity if Aoverlap/ min(Aannotator, Aocr)",
"5.2Baselines",
"is bigger than a manually set threshold; Here",
"We compare the performance of our system with",
"Aannotator, Aocr, Aoverlap are the area of the an-",
"two BiLSTM-CRF baselines. Baseline I applies",
"notated box of the corresponding entity, the area",
"BiLSTM-CRF to each text segment, where each",
"of the ocr detected box and the area of the overlap",
"text segment is an individual sentence. Baseline",
"between the two boxes respectively.",
"II applies the tagging model to the concatenated"
] |
{
"arxiv_id": "1903.11279",
"language": "en",
"page_idx": 4
}
|
||
2311.16298-en-0
|
2311.16298-en-0.jpg
|
Influence Scores at Scale for Efficient Language Data Sampling
### Nikhil Anand* and Joshua Tan* and Maria Minakova Amazon Alexa AI
### $\mathbf{Abstract}$
Modern ML systems ingest data aggregated from diverse sources, such as synthetic, humanannotated, and live customer traffic. Understanding which examples are important to the performance of a learning algorithm is crucial for efficient model training. Recently, a growing body of literature has given rise to various “influence scores,”which use training artifacts such as model confidence or checkpointed gradients to identify important subsets of data. However, these methods have primarily been developed in computer vision settings, and it remains unclear how well they generalize to language-based tasks using pretrained models. In this paper, we explore the applicability of influence scores in language classification tasks. We evaluate a diverse subset of these scores on the SNLI dataset by quantifying accuracy changes in response to pruning training data through random and influence-score-based sampling. We then stress-test one of the scores $-“$variance of gradients" (VoG) from Agarwal et al.(2022)-in an NLU model stack that was exposed to dynamic user speech patterns in a voice assistant type of setting. Our experiments demonstrate that in many cases, encoderbased language models can be finetuned on roughly 50% of the original data without degradation in performance metrics. Along the way, we summarize lessons learned from applying out-of-the-box implementations of influence scores, quantify the effects of noisy and classimbalanced data, and offer recommendations on score-based sampling for better accuracy and training efficiency.
## 1 Introduction
A salient challenge in training transformer-based models is selecting which examples are most important for learning. Understanding the relative importance of training examples towards model
$^*$These authors contributed equally, correspondence:
nkhlanan@amazon.com, jshtan@amazon.com.
performance can inform data selection strategies that minimize customer privacy risks associated with the collection of training data, estimate the impact of the removal of copyrighted or sensitive data, determine mixing strategies to augment monolingual and multilingual datasets to improve accuracy, and identify defective subsets of data. At the same time, in cases where it is desirable to train on as much data as possible- such as large language models - determining the influence of different data instances (both contextually and during pretraining) can help identify failure modes at the level of specific tokens (Grosse et al., 2023), determine the impact of removal of intellectual property, and significantly reduce costs through more efficient model training (Renduchintala et al., 2023).
A growing body of literature in the science of deep learning aims to capture this hierarchy of example importance and has led to a proliferation of a number of "difficulty" or "influence" scores (e.g., Paul et al. (2021); Agarwal et al. (2022); Toneva et al. (2019); Ethayarajh et al. (2022); Garima et al. (2020); Sorscher et al. (2022); Swayamdipta et al. (2020); see App. A for a more complete review). These scores use various training artifacts, such as the margin of confdence or the variance of loss gradients, to rank the relative contribution of each example to model performance. This ranking of examples can then be used in many downstream tasks that require intelligent data selection, such as pruning datasets while maintaining or even improving model accuracy (Paul et al., 2021; Sorscher et al., 2022; Marion et al., 2023); identifying outliers and misannotations in labeled data (Garima et al., 2020; Ethayarajh et al., 2022; Pleiss et al., 2020a; Carlini et al., 2019; Feldman and Zhang, 2020); or reweighting/reordering training examples to increase model robustness (Ren et al., 2018; Wu et al. , 2021) .
Apart from a few notable exceptions (Swayamdipta et al., 2020; Ethayarajh et al.,
|
[
[
0.172119140625,
0.0914306640625,
0.82373046875,
0.0926513671875,
0.82373046875,
0.10986328125,
0.172119140625,
0.1092529296875
],
[
0.270263671875,
0.1484375,
0.73046875,
0.1484375,
0.73046875,
0.162109375,
0.270263671875,
0.162109375
],
[
0.42822265625,
0.166259765625,
0.57666015625,
0.166259765625,
0.57666015625,
0.18115234375,
0.42822265625,
0.18115234375
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51123046875,
0.254150390625,
0.880859375,
0.2529296875,
0.880859375,
0.267333984375,
0.51123046875,
0.268310546875
],
[
0.51123046875,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.1435546875,
0.27978515625,
0.459228515625,
0.28076171875,
0.459228515625,
0.295654296875,
0.1435546875,
0.29443359375
],
[
0.51123046875,
0.2861328125,
0.88330078125,
0.2861328125,
0.88330078125,
0.2998046875,
0.51123046875,
0.2998046875
],
[
0.0201568603515625,
0.300537109375,
0.060455322265625,
0.300537109375,
0.060455322265625,
0.435302734375,
0.0201568603515625,
0.435302734375
],
[
0.14697265625,
0.29638671875,
0.45751953125,
0.29638671875,
0.45751953125,
0.30712890625,
0.14697265625,
0.30712890625
],
[
0.51318359375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51318359375,
0.31640625
],
[
0.145263671875,
0.308837890625,
0.4609375,
0.308837890625,
0.4609375,
0.322509765625,
0.145263671875,
0.322509765625
],
[
0.145263671875,
0.32373046875,
0.45751953125,
0.32373046875,
0.45751953125,
0.33740234375,
0.145263671875,
0.33740234375
],
[
0.51318359375,
0.320556640625,
0.8818359375,
0.320556640625,
0.8818359375,
0.330810546875,
0.51318359375,
0.330810546875
],
[
0.145263671875,
0.33837890625,
0.459228515625,
0.33837890625,
0.459228515625,
0.3515625,
0.145263671875,
0.3515625
],
[
0.51123046875,
0.334228515625,
0.8818359375,
0.3349609375,
0.8818359375,
0.349853515625,
0.51123046875,
0.3486328125
],
[
0.1435546875,
0.350830078125,
0.459228515625,
0.3515625,
0.459228515625,
0.366455078125,
0.1435546875,
0.365234375
],
[
0.51318359375,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.364013671875,
0.51318359375,
0.364013671875
],
[
0.145263671875,
0.366455078125,
0.4609375,
0.366455078125,
0.4609375,
0.380126953125,
0.145263671875,
0.380126953125
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.37890625,
0.51318359375,
0.37890625
],
[
0.145263671875,
0.38134765625,
0.459228515625,
0.38134765625,
0.459228515625,
0.394287109375,
0.145263671875,
0.394287109375
],
[
0.51318359375,
0.385498046875,
0.87890625,
0.385498046875,
0.87890625,
0.3955078125,
0.51318359375,
0.3955078125
],
[
0.1485595703125,
0.396728515625,
0.45751953125,
0.396728515625,
0.45751953125,
0.40673828125,
0.1485595703125,
0.40673828125
],
[
0.509765625,
0.399658203125,
0.8818359375,
0.39794921875,
0.8818359375,
0.410888671875,
0.509765625,
0.413330078125
],
[
0.14697265625,
0.4091796875,
0.459228515625,
0.4091796875,
0.459228515625,
0.4228515625,
0.14697265625,
0.4228515625
],
[
0.51123046875,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.145263671875,
0.42333984375,
0.459228515625,
0.42333984375,
0.459228515625,
0.43701171875,
0.145263671875,
0.43701171875
],
[
0.019317626953125,
0.43701171875,
0.0621337890625,
0.435791015625,
0.06463623046875,
0.51708984375,
0.0218353271484375,
0.5185546875
],
[
0.145263671875,
0.438232421875,
0.459228515625,
0.438232421875,
0.459228515625,
0.451904296875,
0.145263671875,
0.451904296875
],
[
0.51318359375,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.443603515625,
0.51318359375,
0.443603515625
],
[
0.145263671875,
0.451904296875,
0.459228515625,
0.451904296875,
0.459228515625,
0.465576171875,
0.145263671875,
0.465576171875
],
[
0.51123046875,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.145263671875,
0.466064453125,
0.45751953125,
0.466064453125,
0.45751953125,
0.479736328125,
0.145263671875,
0.479736328125
],
[
0.51123046875,
0.463134765625,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.145263671875,
0.485107421875,
0.45751953125,
0.485107421875,
0.45751953125,
0.498779296875,
0.145263671875,
0.498779296875
],
[
0.51123046875,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.493408203125,
0.51123046875,
0.493408203125
],
[
0.145263671875,
0.49951171875,
0.45751953125,
0.49951171875,
0.45751953125,
0.51318359375,
0.145263671875,
0.51318359375
],
[
0.509765625,
0.495361328125,
0.83447265625,
0.49462890625,
0.83447265625,
0.5087890625,
0.509765625,
0.51025390625
],
[
0.145263671875,
0.51318359375,
0.459228515625,
0.51318359375,
0.459228515625,
0.52685546875,
0.145263671875,
0.52685546875
],
[
0.52880859375,
0.51416015625,
0.88330078125,
0.51416015625,
0.88330078125,
0.529296875,
0.52880859375,
0.529296875
],
[
0.0218353271484375,
0.52099609375,
0.0621337890625,
0.52099609375,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.14697265625,
0.52783203125,
0.459228515625,
0.52783203125,
0.459228515625,
0.54150390625,
0.14697265625,
0.54150390625
],
[
0.51318359375,
0.5322265625,
0.8818359375,
0.5322265625,
0.8818359375,
0.5458984375,
0.51318359375,
0.5458984375
],
[
0.145263671875,
0.5419921875,
0.45751953125,
0.5419921875,
0.45751953125,
0.5556640625,
0.145263671875,
0.5556640625
],
[
0.51123046875,
0.54736328125,
0.88330078125,
0.54736328125,
0.88330078125,
0.56103515625,
0.51123046875,
0.56103515625
],
[
0.14697265625,
0.55810546875,
0.456787109375,
0.55810546875,
0.456787109375,
0.568359375,
0.14697265625,
0.568359375
],
[
0.509765625,
0.5625,
0.88330078125,
0.5634765625,
0.88330078125,
0.57763671875,
0.509765625,
0.57666015625
],
[
0.1435546875,
0.57080078125,
0.45751953125,
0.5693359375,
0.45751953125,
0.5830078125,
0.1435546875,
0.58447265625
],
[
0.51123046875,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.59326171875,
0.51123046875,
0.59326171875
],
[
0.1485595703125,
0.5849609375,
0.45751953125,
0.5849609375,
0.45751953125,
0.5986328125,
0.1485595703125,
0.5986328125
],
[
0.509765625,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.509765625,
0.60791015625
],
[
0.14697265625,
0.60107421875,
0.456787109375,
0.60107421875,
0.456787109375,
0.61083984375,
0.14697265625,
0.61083984375
],
[
0.145263671875,
0.61328125,
0.45751953125,
0.61328125,
0.45751953125,
0.626953125,
0.145263671875,
0.626953125
],
[
0.509765625,
0.60986328125,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.509765625,
0.62451171875
],
[
0.14697265625,
0.626953125,
0.45751953125,
0.626953125,
0.45751953125,
0.64013671875,
0.14697265625,
0.64013671875
],
[
0.51123046875,
0.62744140625,
0.88330078125,
0.62744140625,
0.88330078125,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.145263671875,
0.642578125,
0.459228515625,
0.642578125,
0.459228515625,
0.65625,
0.145263671875,
0.65625
],
[
0.509765625,
0.642578125,
0.8818359375,
0.6435546875,
0.8818359375,
0.65673828125,
0.509765625,
0.65625
],
[
0.145263671875,
0.65625,
0.459228515625,
0.65625,
0.459228515625,
0.66943359375,
0.145263671875,
0.66943359375
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.1485595703125,
0.67138671875,
0.45751953125,
0.67138671875,
0.45751953125,
0.681640625,
0.1485595703125,
0.681640625
],
[
0.51123046875,
0.67578125,
0.8818359375,
0.6748046875,
0.8818359375,
0.689453125,
0.51123046875,
0.6904296875
],
[
0.1435546875,
0.68310546875,
0.459228515625,
0.68408203125,
0.459228515625,
0.69873046875,
0.1435546875,
0.69775390625
],
[
0.51318359375,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.7060546875,
0.51318359375,
0.7060546875
],
[
0.145263671875,
0.69873046875,
0.459228515625,
0.69873046875,
0.459228515625,
0.7119140625,
0.145263671875,
0.7119140625
],
[
0.51318359375,
0.7080078125,
0.8818359375,
0.7080078125,
0.8818359375,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.14697265625,
0.71435546875,
0.45751953125,
0.71435546875,
0.45751953125,
0.724609375,
0.14697265625,
0.724609375
],
[
0.145263671875,
0.72705078125,
0.459228515625,
0.72705078125,
0.459228515625,
0.74072265625,
0.145263671875,
0.74072265625
],
[
0.51318359375,
0.724609375,
0.8818359375,
0.724609375,
0.8818359375,
0.73828125,
0.51318359375,
0.73828125
],
[
0.1435546875,
0.7412109375,
0.45751953125,
0.7412109375,
0.45751953125,
0.7548828125,
0.1435546875,
0.7548828125
],
[
0.509765625,
0.74072265625,
0.8818359375,
0.74072265625,
0.8818359375,
0.75439453125,
0.509765625,
0.75439453125
],
[
0.1435546875,
0.755859375,
0.45751953125,
0.755859375,
0.45751953125,
0.76953125,
0.1435546875,
0.76953125
],
[
0.51318359375,
0.75732421875,
0.8818359375,
0.75732421875,
0.8818359375,
0.77099609375,
0.51318359375,
0.77099609375
],
[
0.1427001953125,
0.7685546875,
0.301513671875,
0.76953125,
0.301513671875,
0.78369140625,
0.1427001953125,
0.783203125
],
[
0.51318359375,
0.7724609375,
0.8818359375,
0.7724609375,
0.8818359375,
0.7861328125,
0.51318359375,
0.7861328125
],
[
0.51123046875,
0.7890625,
0.8818359375,
0.7890625,
0.8818359375,
0.802734375,
0.51123046875,
0.802734375
],
[
0.118408203125,
0.798828125,
0.1318359375,
0.798828125,
0.1318359375,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.130126953125,
0.79638671875,
0.26025390625,
0.79638671875,
0.26025390625,
0.81103515625,
0.130126953125,
0.81103515625
],
[
0.51123046875,
0.8046875,
0.88671875,
0.8046875,
0.88671875,
0.818359375,
0.51123046875,
0.818359375
],
[
0.1175537109375,
0.822265625,
0.489501953125,
0.822265625,
0.489501953125,
0.8359375,
0.1175537109375,
0.8359375
],
[
0.51123046875,
0.8193359375,
0.88330078125,
0.82080078125,
0.88330078125,
0.8349609375,
0.51123046875,
0.83447265625
],
[
0.1158447265625,
0.8388671875,
0.49365234375,
0.8388671875,
0.49365234375,
0.8525390625,
0.1158447265625,
0.8525390625
],
[
0.51123046875,
0.8349609375,
0.88330078125,
0.83740234375,
0.88330078125,
0.8515625,
0.51123046875,
0.849609375
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51318359375,
0.8525390625,
0.88330078125,
0.8525390625,
0.88330078125,
0.8662109375,
0.51318359375,
0.8662109375
],
[
0.1175537109375,
0.87158203125,
0.48779296875,
0.87158203125,
0.48779296875,
0.88525390625,
0.1175537109375,
0.88525390625
],
[
0.509765625,
0.86767578125,
0.63623046875,
0.8681640625,
0.63623046875,
0.8828125,
0.509765625,
0.8818359375
],
[
0.1435546875,
0.8935546875,
0.490234375,
0.8935546875,
0.490234375,
0.90673828125,
0.1435546875,
0.90673828125
],
[
0.53076171875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.53076171875,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.40478515625,
0.90478515625,
0.40478515625,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.88330078125,
0.90478515625,
0.88330078125,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Influence Scores at Scale for Efficient Language Data Sampling",
"Nikhil Anand* and Joshua Tan* and Maria Minakova",
"Amazon Alexa AI",
"Abstract",
"performance can inform data selection strategies",
"that minimize customer privacy risks associated",
"Modern ML systems ingest data aggregated",
"with the collection of training data, estimate the im-",
" ",
"from diversesources,suchas synthetic,human",
"pact of the removal of copyrighted or sensitive data,",
"annotated, and live customer traffic. Under-",
"standing which examples are important to the",
"determinemixing strategiestoaugmentmonolin-",
"performance of a learning algorithm is crucial",
"gual and multilingual datasets to improve accuracy,",
"for efficient model training. Recently, a grow-",
"and identify defective subsets of data. At the same",
"ing body of literature has given rise to vari-",
"time, in cases where it is desirable to train on as",
"ous “influence scores,\" which use training arti-",
"much data as possible-such as largelanguage",
"factssuchasmodelconfidenceorcheckpointed",
"models -determining the influence of different",
"gradients to identify important subsets of data.",
"data instances (both contextually and during pre-",
"However, these methods have primarily been",
"[OT'S]",
"developed in computer vision settings, and it",
"training)canhelpidentifyfailuremodes atthelevel",
"remains unclear how well they generalize to",
"of specific tokens (Grosse et al., 2023), determine",
"language-based tasks using pretrained models.",
"the impact of removal of intellectual property, and",
"In this paper, we explore the applicability",
"significantly reduce costs through more efficient",
"of influence scores in language classification",
"model training (Renduchintala et al., 2023).",
"tasks. We evaluate a diverse subset of these",
"A growing body of literature in the science of",
"1186291167A",
"scores on the SNLI dataset by quantifying ac-",
"deep learning aims to capture this hierarchy of ex-",
"curacy changes in response to pruning training",
"ample importance and has led to a proliferation of",
"datathroughrandomand influence-score-based",
"a number of \"difficulty\" or\"influence\" scores (e.g.,",
"sampling. We then stress-test one of the scores",
"Paul et al. (2021); Agarwal et al. (2022); Toneva",
"*variance of gradients\" (VoG) from Agarwal",
"et al. (2019); Ethayarajh et al. (2022): Garima et al",
"etal. (2022)-in an NLU model stackthat",
"was exposed to dynamic user speech patterns",
"(2020); Sorscher et al. (2022); Swayamdipta et al.",
"in a voice assistant type of setting. Our experi-",
"(2020); see App. A for a more complete review)",
"ments demonstrate that in many cases, encoder-",
"These scores use various training artifacts, such as",
"based language models can be finetuned on",
"the margin of confidence or the variance of loss",
"roughly50%oftheoriginal datawithoutdegra-",
"gradients, to rank the relative contribution of each",
"dation in performance metrics. Along the way.",
"example to model performance. This ranking of",
"we summarize lessons learned from applying",
"examples can then be used in many downstream",
"out-of-the-box implementationsof influence",
"scores, quantify the effects of noisy and class-",
"tasks that require intelligent data selection, such",
"imbalanced data, and offer recommendations",
"as pruning datasets while maintaining or even im",
"on score-based sampling for better accuracy",
"proving model accuracy (Paul et al., 2021; Sorscher",
"and training efficiency.",
"et al., 2022; Marion et al., 2023); identifying out",
"liers and misannotations in labeled data (Garima",
"1",
"Introduction",
"et al., 2020; Ethayarajh et al., 2022; Pleiss et al.,",
"A salient challenge in training transformer-based",
"2020a; Carlini et al., 2019; Feldman and Zhang,",
"models is selecting which examples are most im-",
"2020); or reweighting/reordering training exam-",
"portant for learning: Understanding the relative",
"ples to increase model robustness (Ren et al., 2018;",
"importance of training examples towards model",
"Wu et al., 2021).",
"*These authors contributed equally, correspondence:",
"Apartfromafewnotableexceptions",
"nkhlanan@amazon.com, jshtan@amazon.com.",
"(Swayamdipta et al., 2020; Ethayarajh et al."
] |
{
"arxiv_id": "2311.16298",
"language": "en",
"page_idx": 0
}
|
||
2311.16298-en-1
|
2311.16298-en-1.jpg
|
2022; Marion et al., 2023), influence scores have primarily been developed and demonstrated in the context of image classification, and relatively little is known about their effcacy in downstream language-based tasks.$^{1}$ The application of these scores to data selection is further complicated by the fact that during fine-tuning, modern ML systems often ingest a vast amount of data that come from multiple sources, such as synthetic, weak signal,$^2$ live customer data, and humanannotated. Beyond quantifying the efficacy of influence scores in this highly mixed data setting, there is an operational question of the existence of a simple, scalable influence score that can be easily accommodated in a production workflow.
In this work, we take a first pass at answering these questions. First, we benchmark a subset of influence scores on the SNLI dataset (Bowman et al., 2015) in the downstream task of data reduction using a pretrained BERT model (Devlin et al., 2019). Given the task of pruning a language dataset for fine-tuning, are influence scores useful signals for determining optimal data selection strategies? If so, which scores work best? We evaluate these scores against a random sampling baseline, in both noisy and clean data settings.
User speech patterns are constantly evolving due to current events as well as user-system interactions that can be difficult to anticipate. Are influence scores still effective in surfacing data critical for model performance in this dynamic setting? To answer this question, we build upon on our initial findings on SNLI and implement one influence score ("variance of gradients" or "VoG", first presented in Agarwal et al. (2022)) in a generic, large-scale NLU model stack commonly found in commercial voice assistants. We present results for existing in-house test data as well as results for a live user study in which we leveraged VoG scores for the purpose of substantially reducing training data without incurring model-performance degradation.
Among the five influence scores we evaluated on SNLI, most out-of-the-box implementations do not beat a baseline of randomly pruning the dataset. The implementations can be improved to do better than the random-pruning baseline, but this typically
Large language models such as GPT-3 and T5 do imple-
ment some basic data mixing strategies (Brown et al., 2020; Raffel et al., 2020). Our focus here, however, is the setting of using a pretrained model in a downstream task.
$^{2}$For example, data not associated with customer interruptions in online traffic, which are then pseudo-annotated with labels according to the top model hypothesis.
requires careful experimentation to tune hyperparameters specific to each score. Out of the scores we tested, we find that VoG performs best relative to the random-pruning baseline, particularly at large pruning fractions. Test accuracy is mostly maintained after pruning $\sim45\%$ of the SNLI training data using VoG scores calculated in a “one-shot” fashion, i.e. from a single training run, without any score hyperparameter tuning.
In a large-scale user study performed using the NLU stack, we find that sampling by VoG scores is effective at surfacing training data that is particularly efficient for learning. We prune roughly 50% of training data without incurring statistically significant regressions in key metrics that track NLU errors, relative to a baseline model trained with all $data.$
2 Experiments on SNLI
2.1 Selection of influence scores
We considered five different influence scores (described in Table 1) to benchmark in data-reduction tasks on SNLI (Bowman et al., 2015), based on the following criteria: first, they should not require extensive computational resources to implement. For example, the score should not require extensive ensemble averaging by training many $(\gg1)$ copies of“replicate”models to refine the influence measurement of any particular example since many production models can only be trained once in operational workflows.$^3$ Second, the scores should have a natural definition in language models. This excluded some scores that were originally defined in the context of computer vision, such as inputpixel perturbation (Su et al., 2019). We report the implementation details of these scores in App. B.1. Our experiments on SNLI are run on BERTsMAL $( L$ = 4, $H$ = 512, 29.1M parameters), but we comment on the effects of model size in App. B.3.
# 2.2 Experimental Setup
We ran two sets of data-pruning experiments on SNLI to understand the effectiveness of pruning based on the influence scores in Table 1.
In Section 2.3, we describe data-pruning experiments on the original SNLI dataset. First, we generated the influence scores in Table 1 for the entire
${^{3}}$We report on the results of scores that require a moderate $\mathcal{O}(1)$ number of re-runs such as EL2N (Paul et al., 2021), but our main motivation is to determine if there are influence scores that can be used in a“one-shot” setting, using only training artifacts generated from a single run.
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.08489990234375,
0.8818359375,
0.08612060546875,
0.8818359375,
0.1009521484375,
0.5087890625,
0.09979248046875
],
[
0.1158447265625,
0.10272216796875,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1158447265625,
0.11639404296875
],
[
0.509765625,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.509765625,
0.114013671875
],
[
0.1158447265625,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.50732421875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.50732421875,
0.1318359375
],
[
0.1175537109375,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1455078125,
0.1175537109375,
0.1455078125
],
[
0.509765625,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.509765625,
0.1484375
],
[
0.1158447265625,
0.149658203125,
0.489501953125,
0.149658203125,
0.489501953125,
0.1644287109375,
0.1158447265625,
0.1644287109375
],
[
0.5087890625,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.163330078125,
0.5087890625,
0.163330078125
],
[
0.1158447265625,
0.1650390625,
0.486083984375,
0.1650390625,
0.486083984375,
0.1787109375,
0.1158447265625,
0.1787109375
],
[
0.509765625,
0.166259765625,
0.88330078125,
0.166259765625,
0.88330078125,
0.179931640625,
0.509765625,
0.179931640625
],
[
0.1158447265625,
0.1817626953125,
0.48779296875,
0.18115234375,
0.48779296875,
0.1953125,
0.1158447265625,
0.196533203125
],
[
0.5087890625,
0.1817626953125,
0.88330078125,
0.18115234375,
0.88330078125,
0.1953125,
0.5087890625,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.215576171875,
0.489501953125,
0.215576171875,
0.489501953125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.5087890625,
0.215576171875,
0.72900390625,
0.215576171875,
0.72900390625,
0.2286376953125,
0.5087890625,
0.2286376953125
],
[
0.1175537109375,
0.23095703125,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.5283203125,
0.232177734375,
0.880859375,
0.232177734375,
0.880859375,
0.2452392578125,
0.5283203125,
0.2452392578125
],
[
0.11419677734375,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.11419677734375,
0.26123046875
],
[
0.509765625,
0.2452392578125,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.509765625,
0.260009765625
],
[
0.11419677734375,
0.262451171875,
0.489501953125,
0.26318359375,
0.489501953125,
0.27783203125,
0.11419677734375,
0.276611328125
],
[
0.5087890625,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.275634765625,
0.5087890625,
0.275634765625
],
[
0.1175537109375,
0.27978515625,
0.489501953125,
0.27978515625,
0.489501953125,
0.29345703125,
0.1175537109375,
0.29345703125
],
[
0.509765625,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1158447265625,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1158447265625,
0.308837890625
],
[
0.5087890625,
0.29443359375,
0.880859375,
0.29443359375,
0.880859375,
0.30810546875,
0.5087890625,
0.30810546875
],
[
0.1158447265625,
0.311279296875,
0.424072265625,
0.311279296875,
0.424072265625,
0.32470703125,
0.1158447265625,
0.32470703125
],
[
0.509765625,
0.311279296875,
0.880859375,
0.31005859375,
0.880859375,
0.32470703125,
0.509765625,
0.325927734375
],
[
0.133544921875,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.133544921875,
0.341552734375
],
[
0.509765625,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.509765625,
0.341552734375
],
[
0.1175537109375,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.509765625,
0.342529296875,
0.5517578125,
0.344482421875,
0.5498046875,
0.35693359375,
0.5087890625,
0.355224609375
],
[
0.1175537109375,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.5205078125,
0.37353515625,
0.5263671875,
0.377685546875,
0.5205078125,
0.3818359375,
0.5146484375,
0.377685546875
],
[
0.54150390625,
0.371826171875,
0.7314453125,
0.371826171875,
0.7314453125,
0.385986328125,
0.54150390625,
0.385986328125
],
[
0.1158447265625,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.1158447265625,
0.40625
],
[
0.1175537109375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.385986328125,
0.1175537109375,
0.385986328125
],
[
0.509765625,
0.3955078125,
0.77490234375,
0.396728515625,
0.77490234375,
0.410888671875,
0.509765625,
0.410400390625
],
[
0.1175537109375,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.5087890625,
0.416259765625,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.5087890625,
0.43115234375
],
[
0.1158447265625,
0.424560546875,
0.486083984375,
0.424560546875,
0.486083984375,
0.438232421875,
0.1158447265625,
0.438232421875
],
[
0.509765625,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.44775390625,
0.509765625,
0.44775390625
],
[
0.1158447265625,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.454833984375,
0.1158447265625,
0.454833984375
],
[
0.1158447265625,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.4697265625,
0.1158447265625,
0.4697265625
],
[
0.51123046875,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.51123046875,
0.46435546875
],
[
0.1158447265625,
0.47265625,
0.48779296875,
0.47265625,
0.48779296875,
0.486328125,
0.1158447265625,
0.486328125
],
[
0.51123046875,
0.468505859375,
0.880859375,
0.468505859375,
0.880859375,
0.478515625,
0.51123046875,
0.478515625
],
[
0.11419677734375,
0.48681640625,
0.28955078125,
0.488037109375,
0.28955078125,
0.5029296875,
0.11419677734375,
0.501953125
],
[
0.51123046875,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.51123046875,
0.496337890625
],
[
0.1351318359375,
0.5048828125,
0.489501953125,
0.5048828125,
0.489501953125,
0.5185546875,
0.1351318359375,
0.5185546875
],
[
0.509765625,
0.49951171875,
0.88330078125,
0.49951171875,
0.88330078125,
0.51318359375,
0.509765625,
0.51318359375
],
[
0.1158447265625,
0.52099609375,
0.489501953125,
0.52099609375,
0.489501953125,
0.53369140625,
0.1158447265625,
0.53369140625
],
[
0.509765625,
0.515625,
0.88330078125,
0.515625,
0.88330078125,
0.529296875,
0.509765625,
0.529296875
],
[
0.1175537109375,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.1175537109375,
0.5498046875
],
[
0.5087890625,
0.53076171875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.5087890625,
0.5458984375
],
[
0.5087890625,
0.54638671875,
0.880859375,
0.54638671875,
0.880859375,
0.56005859375,
0.5087890625,
0.56005859375
],
[
0.1158447265625,
0.552734375,
0.48779296875,
0.552734375,
0.48779296875,
0.56640625,
0.1158447265625,
0.56640625
],
[
0.509765625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.509765625,
0.57666015625
],
[
0.118408203125,
0.57080078125,
0.48779296875,
0.57080078125,
0.48779296875,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.1158447265625,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.1158447265625,
0.59716796875
],
[
0.51123046875,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.5927734375,
0.51123046875,
0.5927734375
],
[
0.5087890625,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.5087890625,
0.60791015625
],
[
0.1175537109375,
0.6015625,
0.484375,
0.6015625,
0.484375,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.118408203125,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.63134765625,
0.118408203125,
0.63134765625
],
[
0.51123046875,
0.6123046875,
0.8818359375,
0.6123046875,
0.8818359375,
0.6259765625,
0.51123046875,
0.6259765625
],
[
0.1175537109375,
0.6328125,
0.48779296875,
0.6328125,
0.48779296875,
0.646484375,
0.1175537109375,
0.646484375
],
[
0.509765625,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.509765625,
0.64111328125
],
[
0.1158447265625,
0.64794921875,
0.48779296875,
0.6484375,
0.48779296875,
0.6630859375,
0.1158447265625,
0.662109375
],
[
0.51318359375,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.51318359375,
0.65771484375
],
[
0.1175537109375,
0.6650390625,
0.48779296875,
0.6650390625,
0.48779296875,
0.6787109375,
0.1175537109375,
0.6787109375
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.509765625,
0.6748046875,
0.880859375,
0.67578125,
0.880859375,
0.6904296875,
0.509765625,
0.689453125
],
[
0.1158447265625,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.1158447265625,
0.69677734375,
0.489501953125,
0.69677734375,
0.489501953125,
0.71044921875,
0.1158447265625,
0.71044921875
],
[
0.509765625,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.509765625,
0.705078125
],
[
0.1175537109375,
0.71337890625,
0.489501953125,
0.71337890625,
0.489501953125,
0.72802734375,
0.1175537109375,
0.72802734375
],
[
0.51123046875,
0.7080078125,
0.88330078125,
0.7080078125,
0.88330078125,
0.7216796875,
0.51123046875,
0.7216796875
],
[
0.11419677734375,
0.728515625,
0.4306640625,
0.72998046875,
0.4306640625,
0.74462890625,
0.11419677734375,
0.74365234375
],
[
0.51123046875,
0.73388671875,
0.7119140625,
0.7353515625,
0.7119140625,
0.75,
0.51123046875,
0.7490234375
],
[
0.1351318359375,
0.74658203125,
0.48779296875,
0.74658203125,
0.48779296875,
0.76025390625,
0.1351318359375,
0.76025390625
],
[
0.509765625,
0.7548828125,
0.8818359375,
0.755859375,
0.8818359375,
0.77099609375,
0.509765625,
0.76953125
],
[
0.1158447265625,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.1158447265625,
0.775390625
],
[
0.509765625,
0.77099609375,
0.8818359375,
0.7724609375,
0.8818359375,
0.78759765625,
0.509765625,
0.78515625
],
[
0.1175537109375,
0.779296875,
0.489501953125,
0.779296875,
0.489501953125,
0.7919921875,
0.1175537109375,
0.7919921875
],
[
0.51123046875,
0.7880859375,
0.8134765625,
0.7880859375,
0.8134765625,
0.8017578125,
0.51123046875,
0.8017578125
],
[
0.1158447265625,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.1158447265625,
0.80810546875
],
[
0.1158447265625,
0.80859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82470703125,
0.1158447265625,
0.82373046875
],
[
0.52880859375,
0.8046875,
0.88330078125,
0.8046875,
0.88330078125,
0.818359375,
0.52880859375,
0.818359375
],
[
0.509765625,
0.8193359375,
0.884765625,
0.82080078125,
0.884765625,
0.8349609375,
0.509765625,
0.83447265625
],
[
0.139404296875,
0.8330078125,
0.490234375,
0.8330078125,
0.490234375,
0.8466796875,
0.139404296875,
0.8466796875
],
[
0.5087890625,
0.83740234375,
0.880859375,
0.83740234375,
0.880859375,
0.85107421875,
0.5087890625,
0.85107421875
],
[
0.1158447265625,
0.845703125,
0.489501953125,
0.84423828125,
0.489501953125,
0.85791015625,
0.1158447265625,
0.859375
],
[
0.1158447265625,
0.859375,
0.486083984375,
0.859375,
0.486083984375,
0.869140625,
0.1158447265625,
0.869140625
],
[
0.5263671875,
0.85498046875,
0.8818359375,
0.85693359375,
0.8818359375,
0.87158203125,
0.5263671875,
0.869140625
],
[
0.1175537109375,
0.87060546875,
0.4013671875,
0.87060546875,
0.4013671875,
0.88134765625,
0.1175537109375,
0.88134765625
],
[
0.509765625,
0.8681640625,
0.88330078125,
0.869140625,
0.88330078125,
0.8828125,
0.509765625,
0.8818359375
],
[
0.1351318359375,
0.8798828125,
0.490234375,
0.8818359375,
0.490234375,
0.8955078125,
0.1351318359375,
0.8935546875
],
[
0.51123046875,
0.8828125,
0.880859375,
0.8828125,
0.880859375,
0.8935546875,
0.51123046875,
0.8935546875
],
[
0.1158447265625,
0.8935546875,
0.489501953125,
0.8935546875,
0.489501953125,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.51318359375,
0.89453125,
0.8818359375,
0.89453125,
0.8818359375,
0.90771484375,
0.51318359375,
0.90771484375
],
[
0.1158447265625,
0.90625,
0.39208984375,
0.90625,
0.39208984375,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.509765625,
0.90625,
0.78857421875,
0.90625,
0.78857421875,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"2022; Marion et al., 2023), influence scores have",
"requires careful experimentation to tune hyperpa",
"primarily been developed and demonstrated in",
"rametersspecifictoeach score.Out of the scores",
"the context of image classification, and relatively",
"we tested, we find that VoG performs best rela-",
"littleisknownabouttheirefficacyindownstream",
"tive to the random-pruning baseline, particularly",
"language-based tasks.1 The application of these",
" at large pruning fractions. Test accuracy is mostly",
"scores to data selection is further complicated",
"maintained after pruning ~45% of the SNLI train-",
"by the fact that during fine-tuning, modern ML",
"ing data using VoG scores calculated in a “one-shot\"",
"systems often ingest a vast amount of data that",
"fashion, i.e. from a single training run, without any",
"come from multiple sources, such as synthetic,",
" score hyperparameter tuning.",
"weak signal? live customer data, and human-",
"In a large-scale user study performed using the",
"annotated. Beyond quantifying the efficacy of",
"NLU stack, we find that sampling by VoG scores",
"influence scores in this highly mixed data setting,",
"is effective at surfacing training data that is particu-",
"there is an operational question of the existence of",
"larly efficient for learning. We prune roughly 50%",
"a simple, scalable influence score that can be easily",
"of training data without incurring statistically sig",
"accommodated in a production workflow.",
"nificant regressions in key metrics that track NLU",
"In this work, we take a first pass at answering",
"errors, relative to a baseline model trained with all",
"these questions. First, we benchmark a subset of in-",
"data.",
"fluence scores on the SNLI dataset (Bowman et al.,",
"5",
"Experiments on SNLI",
"ing a pretrained BERT model (Devlin et al., 2019)",
"2015)inthe downstreamtask of datareduction us-",
"2.1 Selection of influence scores",
"Given the task of pruning a language dataset for",
"We considered five different influence scores (de",
"fine-tuning, are influence scores useful signals for",
"scribed in Table 1) to benchmark in data-reduction",
"determining optimal data selection strategies? If so,",
"which scores work best? We evaluate these scores",
"tasks on SNLI (Bowman et al., 2015), based on",
"against a random sampling baseline, in both noisy",
"the following criteria:first.they should notrequire",
"and clean data settings.",
"extensive computational resources to implement.",
"User speech patterns are constantly evolving due",
"For example, the score should not require exten-",
"to current events as well as user-system interactions",
"sive ensemble averaging by training many (> 1)",
"that can be difficult to anticipate. Are influence",
"copies of “replicate\" models to refine the influence",
"measurement of any particular example since many",
"scores still effective in surfacing data critical for",
" production models can only be trained once in op-",
"modelperformanceinthisdynamic setting?Toan-",
"swer this question, we build upon on our initial find-",
"erational workflows.3 Second, the scores should",
" have a natural definition in language models. This",
"ings on SNLI and implement one influence score",
"(\"variance of gradients\" or \"VoG\", first presented",
"excluded some scores that were originally defined",
"in Agarwal et al. (2022) in a generic, large-scale",
"in the context of computer vision, such as input",
"NLU model stack commonly found in commercial",
"pixel perturbation (Su et al., 2019). We report the",
"voice assistants. We present results for existing",
"implementation details of these scores in App. B.1",
"Our experiments on SNLI are run on BERTsMALL",
"in-house test data as well as results for a live user",
"study in which we leveraged VoG scores for the pur-",
"(L = 4, H = 512, 29.1M parameters), but we",
"pose of substantially reducing training data without",
"comment on the effects of model size in App. B.3.",
"incurring model-performance degradation.",
"2.2 Experimental Setup",
"Among the five influence scores we evaluated",
"We ran two sets of data-pruning experiments on",
"on SNLI, most out-of-the-box implementations do",
"SNLI to understand the effectiveness of pruning",
"not beat a baseline of randomly pruning the dataset.",
"based on the influence scores in Table l.",
"The implementations can be improved to do better",
"than the random-pruning baseline, but this typically",
"In Section 2.3, we describe data-pruning experi-",
"ments on the original SNLI dataset. First, we gen-",
"'Large language models such as GPT-3 and T5 do imple-",
"erated the influence scores in Table 1 for the entire",
"ment some basic data mixing strategies (Brown et al., 2020;",
"Raffel et al., 2020).Our focus here, however, is the setting of",
"3we report on the results of scores that require a moderate",
"using a pretrained model in a downstream task.",
"O(1) number of re-runs such as EL2N (Paul et al., 2021).",
"2For example, data not associated with customer interrup-",
"but our main motivation is to determine if there are influence",
"tions in online traffic, which are then pseudo-annotated with",
"scores that can be used in a \"one-shot\" setting, using only",
"labels according to the top model hypothesis.",
" training artifacts generated from a single run."
] |
{
"arxiv_id": "2311.16298",
"language": "en",
"page_idx": 1
}
|
||
2311.16298-en-2
|
2311.16298-en-2.jpg
|
<table>
<tbody>
<tr>
<th>Score</th>
<th>$\mathbf{Description}$</th>
</tr>
<tr>
<td>VoG (Agarwal et al., 2 2022</td>
<td>Variance of gradients of model outputs with respect to the inputs.</td>
</tr>
<tr>
<td>EL2N (Paul et $et$ al. 2021)</td>
<td>Norm of the margin of confidence be tween the model prediction and the one-hot label.</td>
</tr>
<tr>
<td>${\overline{{\mathrm{Forgetting}}}}$ Score (Toneva et al. 2019)</td>
<td>${\mathrm{How~often~an~example~is~forgotten~i.e}}$ goes from being classified correctly at checkpoint $i$ to incorrectly at $i+1.$</td>
</tr>
<tr>
<td>${^{\bullet}}{^{\prime}}{^{\prime}}{^{\prime}}$ $PVI$ (Ethayarajh et al., 2022</td>
<td>Fine- grained $information-theoretic$ quantity whose expectation value is the amount of usable information (ir bits) by the model.</td>
</tr>
<tr>
<td>TracIn (Garima et al., 2020)</td>
<td>Influence of any example $z$ towards another example $z^\prime$ by tracking their gradient dot products. We generate the self-influence scores where $z=z^\prime.$</td>
</tr>
</tbody>
</table>
Table 1: Description of the influence scores we generated on the SNLI dataset. See App. B.1 for implementation details.
We use the terminology “easy”and“hard”for pedagogical reasons. Strictly speaking, we are running data-pruning experiments where examples are sampled from either the head or tail of the score distributions. Frequently, these examples do correspond to what a human would find easy and hard, respectively, but we clarify in Section 2.3 when they do not.
$^{5}$As originally prescribed in Agarwal et al.(2022). $^{6}$Details about the label noise are given in App. B.6. These scores were chosen in the noisy label setting due
to their reported efficacy in surfacing defective data.
SNLI training data. We then pruned the training data by using the scores to sample either “easy” or “hard”examples,$^4$ and measured test accuracy for a model trained on the reduced dataset. We compared these score-sampling pruning results to pruning by random sampling. We defer details of the implementation of influence scores and additional findings to App. B.1, but note here that we consider two normalization schemes for VoG scores: classnormalization$^{5}$, where the scores are normalized with respect to the mean and standard deviation of each class, and dataset-normalization, with respect to the full training dataset.
Results on a relatively clean public dataset like SNLI may not always translate to results on large, commercial datasets that are noisier and highly class-imbalanced. In Section 2.4, we address this concern by running similar pruning experiments on SNLI with increasing levels of randomly generated label noise.$^{6}$ We then computed VoG, TracIn, and PVI scores$^{7}$, pruned easy/hard examples based on those scores, and compared test accuracy to a random sampling baseline.
## 2.3 Results on SNLI
Fig. 1 shows test accuracy at the end of training as a function of percent data pruned for each of the five score-based pruning strategies.
General Findings: For most scores, we found that pruning the hardest examples resulted in models with poorer test accuracy compared to pruning the easiest examples. This supports the findings of Sorscher et al. (2022), which hypothesized that hard examples contain critical information about the decision boundaries of classes in larger, less noisy datasets. We also find that out-of-the-box implementations of influence scores- with the exception of VoG- do not result in test accuracy higher than the random sampling baseline without score hyperparameter tuning. For example, for EL2N scores, it is crucial that the scores are computed early during fine-tuning for best results. We explored different implementations and chose those that gave best results for data pruning, while adhering to the criteria listed in Sec. 2.1.
VoG: Remarkably, VoG required only a single model training run and no hyperparameter tuning. At 45% of training data removed, pruning classnormalized VoG-easy examples led to a test accuracy of 85.04±0.20%, compared to 85.52±0.14% with all of the data. At smaller pruning fractions $(\lesssim10\%)$, performance is roughly within the margin of error of sampling randomly. We find that sampling dataset-normalized scores generally performs worse than class-normalized (84.60±1.50% at 45% easy pruned), which is due to the overrepresentation of the“contradiction”class (Fig. 2) in the tail. We will revisit the merits of class versus dataset normalization in Sec 3.
EL2N: Best results were obtained by computing EL2N scores early in training; we found epoch $\sim2$ outperformed the random pruning baseline for small to moderate pruning fractions (between 0-25%), but worse beyond that.
EL2N is a margin-based metric, which means that examples drawn from the tail of the EL2N distribution should lie close to the decision boundary between classes (Paul et al., 2021; Sorscher et al., 2022). If that is true, then removing these examples should dissolve the decision boundary between different classes, and account for the drop in test accuracy. We provide some evidence for this in App. B.4 by clustering the t-SNE (Hinton and Roweis, 2002) encoder representations of the training and test data, before and after pruning EL2N-
|
[
[
0.124267578125,
0.0885009765625,
0.166259765625,
0.0885009765625,
0.166259765625,
0.09979248046875,
0.124267578125,
0.09979248046875
],
[
0.258544921875,
0.0885009765625,
0.339111328125,
0.0885009765625,
0.339111328125,
0.09979248046875,
0.258544921875,
0.09979248046875
],
[
0.51123046875,
0.08489990234375,
0.68408203125,
0.08489990234375,
0.68408203125,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.124267578125,
0.1051025390625,
0.160400390625,
0.1051025390625,
0.160400390625,
0.1175537109375,
0.124267578125,
0.1175537109375
],
[
0.1822509765625,
0.105712890625,
0.246826171875,
0.105712890625,
0.246826171875,
0.1175537109375,
0.1822509765625,
0.1175537109375
],
[
0.2568359375,
0.1051025390625,
0.496337890625,
0.1051025390625,
0.496337890625,
0.1181640625,
0.2568359375,
0.1181640625
],
[
0.5087890625,
0.10986328125,
0.880859375,
0.10986328125,
0.880859375,
0.12353515625,
0.5087890625,
0.12353515625
],
[
0.124267578125,
0.1175537109375,
0.202392578125,
0.1175537109375,
0.202392578125,
0.12890625,
0.124267578125,
0.12890625
],
[
0.258544921875,
0.1175537109375,
0.4189453125,
0.1175537109375,
0.4189453125,
0.1312255859375,
0.258544921875,
0.1312255859375
],
[
0.12176513671875,
0.1300048828125,
0.2476806640625,
0.1300048828125,
0.2476806640625,
0.1436767578125,
0.12176513671875,
0.1436767578125
],
[
0.258544921875,
0.1300048828125,
0.5029296875,
0.1300048828125,
0.5029296875,
0.1436767578125,
0.258544921875,
0.1436767578125
],
[
0.509765625,
0.1270751953125,
0.8818359375,
0.1270751953125,
0.8818359375,
0.14013671875,
0.509765625,
0.14013671875
],
[
0.1275634765625,
0.141357421875,
0.166259765625,
0.141357421875,
0.166259765625,
0.15380859375,
0.1275634765625,
0.15380859375
],
[
0.26123046875,
0.1436767578125,
0.495361328125,
0.1436767578125,
0.495361328125,
0.15380859375,
0.26123046875,
0.15380859375
],
[
0.509765625,
0.142578125,
0.771484375,
0.142578125,
0.771484375,
0.1561279296875,
0.509765625,
0.1561279296875
],
[
0.26025390625,
0.155029296875,
0.345947265625,
0.155029296875,
0.345947265625,
0.1650390625,
0.26025390625,
0.1650390625
],
[
0.52880859375,
0.1602783203125,
0.8818359375,
0.1602783203125,
0.8818359375,
0.17333984375,
0.52880859375,
0.17333984375
],
[
0.124267578125,
0.16748046875,
0.1956787109375,
0.16748046875,
0.1956787109375,
0.1776123046875,
0.124267578125,
0.1776123046875
],
[
0.204833984375,
0.16748046875,
0.2451171875,
0.16748046875,
0.2451171875,
0.1776123046875,
0.204833984375,
0.1776123046875
],
[
0.26025390625,
0.16748046875,
0.496337890625,
0.16748046875,
0.496337890625,
0.1776123046875,
0.26025390625,
0.1776123046875
],
[
0.1275634765625,
0.1776123046875,
0.181396484375,
0.1776123046875,
0.181396484375,
0.189453125,
0.1275634765625,
0.189453125
],
[
0.26123046875,
0.1787109375,
0.496337890625,
0.1787109375,
0.496337890625,
0.189453125,
0.26123046875,
0.189453125
],
[
0.51123046875,
0.177001953125,
0.8818359375,
0.177001953125,
0.8818359375,
0.190673828125,
0.51123046875,
0.190673828125
],
[
0.2275390625,
0.179931640625,
0.2435302734375,
0.179931640625,
0.2435302734375,
0.188232421875,
0.2275390625,
0.188232421875
],
[
0.19482421875,
0.179931640625,
0.208251953125,
0.179931640625,
0.208251953125,
0.188232421875,
0.19482421875,
0.188232421875
],
[
0.1259765625,
0.18701171875,
0.16796875,
0.189453125,
0.166259765625,
0.201904296875,
0.124267578125,
0.1995849609375
],
[
0.258544921875,
0.189453125,
0.47607421875,
0.189453125,
0.47607421875,
0.203125,
0.258544921875,
0.203125
],
[
0.509765625,
0.191162109375,
0.880859375,
0.191162109375,
0.880859375,
0.204833984375,
0.509765625,
0.204833984375
],
[
0.1234130859375,
0.201904296875,
0.156982421875,
0.201904296875,
0.156982421875,
0.213134765625,
0.1234130859375,
0.213134765625
],
[
0.166259765625,
0.201904296875,
0.2451171875,
0.203125,
0.2451171875,
0.21435546875,
0.166259765625,
0.213134765625
],
[
0.26025390625,
0.2037353515625,
0.34765625,
0.2037353515625,
0.34765625,
0.21435546875,
0.26025390625,
0.21435546875
],
[
0.3525390625,
0.2060546875,
0.35693359375,
0.2060546875,
0.35693359375,
0.21142578125,
0.3525390625,
0.21142578125
],
[
0.355224609375,
0.2037353515625,
0.491943359375,
0.2037353515625,
0.491943359375,
0.21435546875,
0.355224609375,
0.21435546875
],
[
0.258544921875,
0.213134765625,
0.49365234375,
0.213134765625,
0.49365234375,
0.226806640625,
0.258544921875,
0.226806640625
],
[
0.5087890625,
0.2078857421875,
0.880859375,
0.2078857421875,
0.880859375,
0.221435546875,
0.5087890625,
0.221435546875
],
[
0.1259765625,
0.21435546875,
0.202392578125,
0.21435546875,
0.202392578125,
0.2257080078125,
0.1259765625,
0.2257080078125
],
[
0.26025390625,
0.226806640625,
0.497802734375,
0.226806640625,
0.497802734375,
0.2369384765625,
0.26025390625,
0.2369384765625
],
[
0.509765625,
0.223876953125,
0.8818359375,
0.223876953125,
0.8818359375,
0.2369384765625,
0.509765625,
0.2369384765625
],
[
0.26123046875,
0.2381591796875,
0.376220703125,
0.2381591796875,
0.376220703125,
0.25048828125,
0.26123046875,
0.25048828125
],
[
0.509765625,
0.240478515625,
0.8818359375,
0.240478515625,
0.8818359375,
0.254150390625,
0.509765625,
0.254150390625
],
[
0.12176513671875,
0.2493896484375,
0.17041015625,
0.2493896484375,
0.17041015625,
0.26123046875,
0.12176513671875,
0.26123046875
],
[
0.19140625,
0.25048828125,
0.2476806640625,
0.25048828125,
0.2476806640625,
0.26123046875,
0.19140625,
0.26123046875
],
[
0.26025390625,
0.25048828125,
0.495361328125,
0.25048828125,
0.495361328125,
0.26123046875,
0.26025390625,
0.26123046875
],
[
0.5087890625,
0.254638671875,
0.8818359375,
0.255859375,
0.8818359375,
0.270751953125,
0.5087890625,
0.26953125
],
[
0.124267578125,
0.262451171875,
0.202392578125,
0.26123046875,
0.202392578125,
0.2724609375,
0.124267578125,
0.273681640625
],
[
0.26025390625,
0.26123046875,
0.495361328125,
0.26123046875,
0.495361328125,
0.27490234375,
0.26025390625,
0.27490234375
],
[
0.26025390625,
0.2724609375,
0.496337890625,
0.2724609375,
0.496337890625,
0.2861328125,
0.26025390625,
0.2861328125
],
[
0.5087890625,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.2861328125,
0.5087890625,
0.2861328125
],
[
0.258544921875,
0.283203125,
0.47021484375,
0.283203125,
0.47021484375,
0.29638671875,
0.258544921875,
0.29638671875
],
[
0.509765625,
0.2880859375,
0.88330078125,
0.2880859375,
0.88330078125,
0.3017578125,
0.509765625,
0.3017578125
],
[
0.51123046875,
0.3046875,
0.8818359375,
0.3046875,
0.8818359375,
0.318359375,
0.51123046875,
0.318359375
],
[
0.1175537109375,
0.3134765625,
0.490234375,
0.3134765625,
0.490234375,
0.32666015625,
0.1175537109375,
0.32666015625
],
[
0.5087890625,
0.319580078125,
0.88330078125,
0.320556640625,
0.88330078125,
0.3349609375,
0.5087890625,
0.334228515625
],
[
0.11419677734375,
0.325927734375,
0.486083984375,
0.325927734375,
0.486083984375,
0.339111328125,
0.11419677734375,
0.339111328125
],
[
0.1158447265625,
0.341552734375,
0.1956787109375,
0.341552734375,
0.1956787109375,
0.352783203125,
0.1158447265625,
0.352783203125
],
[
0.51123046875,
0.336181640625,
0.8818359375,
0.336181640625,
0.8818359375,
0.349853515625,
0.51123046875,
0.349853515625
],
[
0.51123046875,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.365234375,
0.51123046875,
0.365234375
],
[
0.5146484375,
0.37060546875,
0.880859375,
0.37060546875,
0.880859375,
0.38134765625,
0.5146484375,
0.38134765625
],
[
0.1175537109375,
0.385986328125,
0.48779296875,
0.385986328125,
0.48779296875,
0.399658203125,
0.1175537109375,
0.399658203125
],
[
0.51123046875,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3984375,
0.51123046875,
0.3984375
],
[
0.1158447265625,
0.402099609375,
0.48779296875,
0.40087890625,
0.48779296875,
0.41455078125,
0.1158447265625,
0.4150390625
],
[
0.509765625,
0.399658203125,
0.8818359375,
0.40087890625,
0.8818359375,
0.4150390625,
0.509765625,
0.41455078125
],
[
0.1158447265625,
0.418701171875,
0.48779296875,
0.418701171875,
0.48779296875,
0.431640625,
0.1158447265625,
0.431640625
],
[
0.509765625,
0.418701171875,
0.80615234375,
0.418701171875,
0.80615234375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.1158447265625,
0.435302734375,
0.489501953125,
0.435302734375,
0.489501953125,
0.448974609375,
0.1158447265625,
0.448974609375
],
[
0.52880859375,
0.432861328125,
0.880859375,
0.43408203125,
0.880859375,
0.448974609375,
0.52880859375,
0.44775390625
],
[
0.1175537109375,
0.451904296875,
0.489501953125,
0.451904296875,
0.489501953125,
0.465576171875,
0.1175537109375,
0.465576171875
],
[
0.509765625,
0.449462890625,
0.8818359375,
0.45068359375,
0.8818359375,
0.465576171875,
0.509765625,
0.46435546875
],
[
0.1175537109375,
0.468505859375,
0.484375,
0.468505859375,
0.484375,
0.478515625,
0.1175537109375,
0.478515625
],
[
0.51123046875,
0.465576171875,
0.88330078125,
0.466064453125,
0.88330078125,
0.48095703125,
0.51123046875,
0.479736328125
],
[
0.1158447265625,
0.482666015625,
0.48779296875,
0.48095703125,
0.48779296875,
0.49462890625,
0.1158447265625,
0.496337890625
],
[
0.509765625,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.509765625,
0.496337890625
],
[
0.1175537109375,
0.49951171875,
0.48779296875,
0.49951171875,
0.48779296875,
0.51318359375,
0.1175537109375,
0.51318359375
],
[
0.509765625,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.51318359375,
0.509765625,
0.51318359375
],
[
0.1158447265625,
0.515625,
0.489501953125,
0.515625,
0.489501953125,
0.529296875,
0.1158447265625,
0.529296875
],
[
0.51123046875,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.51123046875,
0.529296875
],
[
0.1175537109375,
0.53076171875,
0.48779296875,
0.53076171875,
0.48779296875,
0.54443359375,
0.1175537109375,
0.54443359375
],
[
0.51123046875,
0.5322265625,
0.8818359375,
0.5322265625,
0.8818359375,
0.5458984375,
0.51123046875,
0.5458984375
],
[
0.11419677734375,
0.54638671875,
0.486083984375,
0.54638671875,
0.486083984375,
0.56005859375,
0.11419677734375,
0.56005859375
],
[
0.509765625,
0.54736328125,
0.8818359375,
0.54638671875,
0.8818359375,
0.56103515625,
0.509765625,
0.5625
],
[
0.1158447265625,
0.5634765625,
0.48779296875,
0.5634765625,
0.48779296875,
0.57666015625,
0.1158447265625,
0.57666015625
],
[
0.51123046875,
0.56396484375,
0.88330078125,
0.56396484375,
0.88330078125,
0.57763671875,
0.51123046875,
0.57763671875
],
[
0.1158447265625,
0.580078125,
0.36865234375,
0.580078125,
0.36865234375,
0.59326171875,
0.1158447265625,
0.59326171875
],
[
0.51123046875,
0.580078125,
0.88330078125,
0.580078125,
0.88330078125,
0.59326171875,
0.51123046875,
0.59326171875
],
[
0.509765625,
0.595703125,
0.88330078125,
0.59423828125,
0.88330078125,
0.609375,
0.509765625,
0.60986328125
],
[
0.1351318359375,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.1351318359375,
0.615234375
],
[
0.1158447265625,
0.6162109375,
0.48779296875,
0.61767578125,
0.48779296875,
0.6318359375,
0.1158447265625,
0.63134765625
],
[
0.51123046875,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.51123046875,
0.6259765625
],
[
0.51123046875,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.118408203125,
0.63525390625,
0.483642578125,
0.63525390625,
0.483642578125,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.51123046875,
0.64453125,
0.736328125,
0.64453125,
0.736328125,
0.65478515625,
0.51123046875,
0.65478515625
],
[
0.118408203125,
0.65185546875,
0.486083984375,
0.65185546875,
0.486083984375,
0.662109375,
0.118408203125,
0.662109375
],
[
0.52880859375,
0.66015625,
0.8818359375,
0.662109375,
0.8818359375,
0.67578125,
0.52880859375,
0.67333984375
],
[
0.118408203125,
0.66748046875,
0.484375,
0.66748046875,
0.484375,
0.677734375,
0.118408203125,
0.677734375
],
[
0.1175537109375,
0.68310546875,
0.490234375,
0.68310546875,
0.490234375,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.51123046875,
0.677734375,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.51123046875,
0.69140625
],
[
0.11419677734375,
0.69677734375,
0.486083984375,
0.69677734375,
0.486083984375,
0.71142578125,
0.11419677734375,
0.71142578125
],
[
0.515625,
0.69677734375,
0.53662109375,
0.69677734375,
0.53662109375,
0.70361328125,
0.515625,
0.70361328125
],
[
0.5341796875,
0.6953125,
0.87890625,
0.6953125,
0.87890625,
0.7060546875,
0.5341796875,
0.7060546875
],
[
0.11419677734375,
0.71337890625,
0.48779296875,
0.71435546875,
0.48779296875,
0.728515625,
0.11419677734375,
0.72802734375
],
[
0.51318359375,
0.71044921875,
0.8818359375,
0.71044921875,
0.8818359375,
0.7236328125,
0.51318359375,
0.7236328125
],
[
0.1158447265625,
0.73095703125,
0.48779296875,
0.73095703125,
0.48779296875,
0.74462890625,
0.1158447265625,
0.74462890625
],
[
0.51318359375,
0.7255859375,
0.74658203125,
0.7255859375,
0.74658203125,
0.7392578125,
0.51318359375,
0.7392578125
],
[
0.11419677734375,
0.74658203125,
0.313232421875,
0.74658203125,
0.313232421875,
0.76025390625,
0.11419677734375,
0.76025390625
],
[
0.5283203125,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.5283203125,
0.755859375
],
[
0.509765625,
0.7587890625,
0.8818359375,
0.75830078125,
0.8818359375,
0.7724609375,
0.509765625,
0.77392578125
],
[
0.51123046875,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51123046875,
0.7880859375
],
[
0.1351318359375,
0.79345703125,
0.491943359375,
0.7958984375,
0.491943359375,
0.81005859375,
0.1351318359375,
0.80810546875
],
[
0.509765625,
0.79150390625,
0.8818359375,
0.79052734375,
0.8818359375,
0.8046875,
0.509765625,
0.8056640625
],
[
0.1175537109375,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.8056640625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.509765625,
0.82080078125
],
[
0.1175537109375,
0.82177734375,
0.48779296875,
0.82177734375,
0.48779296875,
0.83203125,
0.1175537109375,
0.83203125
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.8330078125,
0.48779296875,
0.8330078125,
0.48779296875,
0.84326171875,
0.118408203125,
0.84326171875
],
[
0.1175537109375,
0.84423828125,
0.486083984375,
0.84423828125,
0.486083984375,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.509765625,
0.8388671875,
0.8818359375,
0.84130859375,
0.8818359375,
0.85498046875,
0.509765625,
0.8525390625
],
[
0.1158447265625,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.8681640625,
0.1158447265625,
0.8681640625
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1376953125,
0.8662109375,
0.443359375,
0.86767578125,
0.443359375,
0.8818359375,
0.1376953125,
0.88134765625
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1376953125,
0.87890625,
0.453369140625,
0.8798828125,
0.453369140625,
0.89453125,
0.1376953125,
0.8935546875
],
[
0.1368408203125,
0.89111328125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.1368408203125,
0.90625
],
[
0.509765625,
0.88720703125,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.435791015625,
0.90478515625,
0.435791015625,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Score",
"Description",
"2.3 Results on SNLI",
"VoG",
"(Agarwal",
"Variance of gradients of model outputs",
"Fig. I shows test accuracy at the end of training as",
"et al., 2022)",
"with respect to the inputs.",
"EL2N (Paul et al.,",
"Norm of the margin of confidence be-",
"a function of percent data pruned for each of the",
"2021)",
"tween the model predictionand the",
"five score-based pruning strategies.",
"one-hot label.",
"General Findings: For most scores, we found",
"Forgetting",
"Score",
"How oftenan exampleis forgotteni.e.",
"(Toneva",
"goes from being classified correctly at",
"that pruning the hardest examples resulted in mod-",
"al..",
"et",
"2019)",
"checkpoint i to incorrectly at i + 1",
"els with poorer test accuracy compared to pruning",
"PVI",
"(Ethayarajh",
"Fine-grained",
"",
"information-theoretic",
"quantity whose expectation value is",
"the easiest examples. This supports the findings",
"et al., 2022)",
"the amount of usable information (in",
"of Sorscher et al. (2022), which hypothesized that",
"bits) by the model.",
"hard examples contain critical information about",
"TracIn",
"(Garima",
"Influence of any example ztowards",
"the decision boundaries of classes in larger, less",
"et al.,2020)",
"another example z' by tracking their",
"gradient dot products. We generate the",
"noisy datasets. We also find that out-of-the-box",
"self-influence scores where z = z",
"implementations of influence scores -- with the ex-",
"ception of VoG - do not result in test accuracy",
"Table l: Description of the influence scores we gener-",
"higher than the random sampling baseline with-",
" ated on the SNLI dataset. See App. B.1 for implementa-",
"tion details.",
"out score hyperparameter tuning.For example,",
"for EL2N scores, it is crucial that the scores are",
"computedearlyduringfine-tuningforbestresults",
"SNLI training data. We then pruned the training",
"We explored different implementations and chose",
"data by using the scores to sample either “easy\"or",
"those that gave best results for data pruning, while",
"\"hard\"examples, and measured test accuracy for",
"adhering to the criteria listed in Sec.2.1",
"a model trained on the reduced dataset. We com-",
"VoG: Remarkably, VoG required only a single",
"pared these score-sampling pruning results to prun-",
"model training run and no hyperparameter tuning",
"ingbyrandomsampling.Wedeferdetailsofthe",
"At 45% of training data removed, pruning class-",
"implementation of influence scores and additional",
" normalized VoG-easy examples led to a test accu-",
"findings to App. B.1, but note here that we consider",
"racy of 85.04±0.20%, compared to 85.52±0.14%",
"two normalization schemes for VoG scores: class-",
"with all of the data. At smaller pruning fractions",
"normalization, where the scores are normalized",
"(≤10%), performance is roughly within the mar-",
"with respect to the mean and standard deviation",
"gin of error of sampling randomly. We find that",
"of each class, and dataset-normalization, with",
"sampling dataset-normalized scores generally per-",
"respect to the full training dataset.",
"forms worse than class-normalized (84.60±1.50%",
"at 45% easy pruned), which is due to the over-",
"Results on a relatively clean public dataset like",
"SNLI may not always translate to results on large,",
"representation of the “contradiction\" class (Fig. 2)",
"in the tail. We will revisit the merits of class versus",
"commercial datasets that are noisier and highly",
"dataset normalization in Sec 3",
"class-imbalanced.In Section2.4.weaddressthis",
"EL2N:Best results wereobtained by computing",
"concern by running similar pruning experiments",
"on SNLI with increasing levels of randomly gener-",
"EL2N scores early in training; we found epoch",
" ated label noise.6 We then computed VoG, TracIn",
"",
"2outperformedtherandompruningbaseline",
"and PVI scores7,pruned easy/hard examples based",
"for small to moderate pruning fractions (between",
"on those scores, and compared test accuracy to a",
"0-25%), but worse beyond that.",
"random sampling baseline",
"EL2N is a margin-based metric, which means",
"that examples drawn from the tail of the EL2N",
"distribution should lie close to the decision bound-",
"we use the terminology“easy\"and“hard\"for pedagog-",
"ary between classes (Paul et al., 2021; Sorscher",
"ical reasons. Strictly speaking, we are running data-pruning",
"et al., 2022). If that is true, then removing these",
"experiments whereexamplesare sampledfrom either the head",
"examples should dissolve the decision boundary",
"ortail of the scoredistributions.Frequently,these examples",
"do correspond to what a human would find easy and hard.",
"between different classes, and account for the drop",
"respectively, but we clarify in Section 2.3 when they do not.",
"in test accuracy. We provide some evidence for this",
"'As originally prescribed in Agarwal et al. (2022).",
"in App. B.4 by clustering the t-SNE (Hinton and",
"6Details about the label noise are given in App. B.6.",
"7These scores were chosen in the noisy label setting due",
"Roweis, 2002) encoder representations of the train-",
"to their reported efficacy in surfacing defective data.",
"ing and test data, before and after pruning EL2N-"
] |
{
"arxiv_id": "2311.16298",
"language": "en",
"page_idx": 2
}
|
||
2311.16298-en-4
|
2311.16298-en-4.jpg
|
<FigureHere>
Figure 3: Shows the results of pruning VoG-easy (blue), TracIn-hard (orange), PVI-hard (purple), or randomly selected (gold) examples for noisy SNLI. The amount of added isotropic label noise is given in the orange inset. Points indicate the average of three runs and shaded regions indicate the 1-$\sigma$ envelope.
noisy settings, for large pruning fractions. In some cases, pruning based on VoG scores even clawed back a fraction of the initial accuracy loss due to noisy labels. However, somewhat surprisingly, this was likely not because VoG-based selection was pruning the misannotated examples themselves. The similarity index between the easiest VoG examples and all of the introduced misannotated examples in the 30% label-noise setting was only$\approx0.11.$ Compared to random sampling $(\approx0.11)$,we conclude that VoG does not do better than random chance at finding misannotated SNLI examples, but does reliably extract more influential examples that partially mitigate the effects of label noise. In all noisy settings, we found that pruning VoG-hard examples did worse than pruning randomly
Pruning by TracIn and PVI scores resulted in a small but persistent accuracy increase when $\sim5$- $10\%$ of the hard data was pruned. In the 30% noise setting, the similarity index between the TracInhard and PVI-hard examples and misannotated examples was$\approx0.11,0.09$, respectively, again indicating that the accuracy gains are not due to the removal of defects. The number of instances with a PVI score of < 0 (indicating a potential misannotation) comprises only 6% of the mislabeled data. Nevertheless, it appears beneficial to use these scores to prune 5-10% of overall hard data that adversely impacts training in a noisy setting.
typically found in large, general-purpose commercial voice assistants. This setting poses practical challenges often not reflected in public datasets, such as noisy and evolving speech patterns, diverse vocabularies, dialects, carrier phrases, and out-of-distribution named entities. As an added challenge, we focus on Japanese-data trained models and datasets to determine if VoG-based influence scoring could function in a lower-resource setting. The statistical-model component in this setting consisted of generic domain classifier (DC), intent classifier (IC), and named entity recognition (NER) models organized in a coupled, hierarchical manner: a single multi-class DC model is trained on all domains’ data to first predict the domain for a given utterance, which then invokes a domainspecific joint IC-NER model trained on in-domain data$^{11}.$ Both sets of models were based on distilled Japanese BERT models and VoG scores were computed using the procedure given in App. B.1.
Fig.4 shows the class-normalized scores for a subset of five domains that differ in the proportion of the overall training data they represent and intentlabel complexity$^{12}$ within that domain. We observe that smaller domains tend to have higher scores (e.g., HealthAndFitness vs. HomeAutomation) and more complex domains tend to have higher scores (Shopping vs. Video). In some cases, domains that are similar in size and complexity still exhibit different VoG score distributions (Music vs. Video) which reflects differing influence of domain data
## 3 $\textbf{VoG in the Context of NLU}$
Given its promising results for data reduction on SNLI, we set out to evaluate VoG in an environment
$\overline{\mathrm{ll}_{12}\text{See App. C.}{1}}$for additional context.
$^{12}$As measured by Shannon entropy; see App. C.6.
|
[
[
0.55517578125,
0.0885009765625,
0.6767578125,
0.0885009765625,
0.6767578125,
0.09857177734375,
0.55517578125,
0.09857177734375
],
[
0.72119140625,
0.0885009765625,
0.82861328125,
0.0885009765625,
0.82861328125,
0.09979248046875,
0.72119140625,
0.09979248046875
],
[
0.405517578125,
0.08905029296875,
0.50537109375,
0.08905029296875,
0.50537109375,
0.09979248046875,
0.405517578125,
0.09979248046875
],
[
0.2435302734375,
0.08905029296875,
0.35693359375,
0.08905029296875,
0.35693359375,
0.09979248046875,
0.2435302734375,
0.09979248046875
],
[
0.145263671875,
0.1009521484375,
0.16796875,
0.1009521484375,
0.16796875,
0.114013671875,
0.145263671875,
0.114013671875
],
[
0.14697265625,
0.1175537109375,
0.16796875,
0.1175537109375,
0.16796875,
0.1318359375,
0.14697265625,
0.1318359375
],
[
0.12005615234375,
0.141357421875,
0.14111328125,
0.141357421875,
0.14111328125,
0.2327880859375,
0.12005615234375,
0.2327880859375
],
[
0.14697265625,
0.135986328125,
0.16796875,
0.135986328125,
0.16796875,
0.15087890625,
0.14697265625,
0.15087890625
],
[
0.175537109375,
0.1478271484375,
0.274658203125,
0.1478271484375,
0.274658203125,
0.1591796875,
0.175537109375,
0.1591796875
],
[
0.5224609375,
0.1484375,
0.5517578125,
0.1484375,
0.5517578125,
0.15673828125,
0.5224609375,
0.15673828125
],
[
0.14697265625,
0.155029296875,
0.16796875,
0.155029296875,
0.16796875,
0.16748046875,
0.14697265625,
0.16748046875
],
[
0.1502685546875,
0.17333984375,
0.162841796875,
0.17333984375,
0.162841796875,
0.185302734375,
0.1502685546875,
0.185302734375
],
[
0.14697265625,
0.190673828125,
0.16796875,
0.190673828125,
0.16796875,
0.2037353515625,
0.14697265625,
0.2037353515625
],
[
0.177978515625,
0.201904296875,
0.204833984375,
0.201904296875,
0.204833984375,
0.213134765625,
0.177978515625,
0.213134765625
],
[
0.5224609375,
0.203125,
0.5498046875,
0.203125,
0.5498046875,
0.21142578125,
0.5224609375,
0.21142578125
],
[
0.14697265625,
0.208984375,
0.16796875,
0.208984375,
0.16796875,
0.22265625,
0.14697265625,
0.22265625
],
[
0.1485595703125,
0.22802734375,
0.16455078125,
0.22802734375,
0.16455078125,
0.2381591796875,
0.1485595703125,
0.2381591796875
],
[
0.14697265625,
0.24462890625,
0.1695556640625,
0.24462890625,
0.1695556640625,
0.25830078125,
0.14697265625,
0.25830078125
],
[
0.175537109375,
0.254638671875,
0.20654296875,
0.254638671875,
0.20654296875,
0.267333984375,
0.175537109375,
0.267333984375
],
[
0.5205078125,
0.255859375,
0.55322265625,
0.255859375,
0.55322265625,
0.267333984375,
0.5205078125,
0.267333984375
],
[
0.2392578125,
0.2724609375,
0.255126953125,
0.2724609375,
0.255126953125,
0.281982421875,
0.2392578125,
0.281982421875
],
[
0.1502685546875,
0.260009765625,
0.1796875,
0.26953125,
0.17041015625,
0.28076171875,
0.14111328125,
0.271484375
],
[
0.27783203125,
0.2724609375,
0.295654296875,
0.2724609375,
0.295654296875,
0.281982421875,
0.27783203125,
0.281982421875
],
[
0.313232421875,
0.2724609375,
0.3359375,
0.2724609375,
0.3359375,
0.283203125,
0.313232421875,
0.283203125
],
[
0.353515625,
0.2724609375,
0.372802734375,
0.2724609375,
0.372802734375,
0.281982421875,
0.353515625,
0.281982421875
],
[
0.391357421875,
0.2724609375,
0.41064453125,
0.2724609375,
0.41064453125,
0.281982421875,
0.391357421875,
0.281982421875
],
[
0.432373046875,
0.2724609375,
0.447509765625,
0.2724609375,
0.447509765625,
0.28076171875,
0.432373046875,
0.28076171875
],
[
0.47021484375,
0.2724609375,
0.48779296875,
0.2724609375,
0.48779296875,
0.28076171875,
0.47021484375,
0.28076171875
],
[
0.587890625,
0.2724609375,
0.60205078125,
0.2724609375,
0.60205078125,
0.28076171875,
0.587890625,
0.28076171875
],
[
0.62451171875,
0.2724609375,
0.64208984375,
0.2724609375,
0.64208984375,
0.281982421875,
0.62451171875,
0.281982421875
],
[
0.6630859375,
0.2724609375,
0.6826171875,
0.2724609375,
0.6826171875,
0.281982421875,
0.6630859375,
0.281982421875
],
[
0.7001953125,
0.2724609375,
0.7197265625,
0.2724609375,
0.7197265625,
0.281982421875,
0.7001953125,
0.281982421875
],
[
0.74072265625,
0.2724609375,
0.759765625,
0.2724609375,
0.759765625,
0.281982421875,
0.74072265625,
0.281982421875
],
[
0.77734375,
0.2724609375,
0.796875,
0.2724609375,
0.796875,
0.281982421875,
0.77734375,
0.281982421875
],
[
0.81787109375,
0.2724609375,
0.83642578125,
0.2724609375,
0.83642578125,
0.28076171875,
0.81787109375,
0.28076171875
],
[
0.85546875,
0.2724609375,
0.873046875,
0.2724609375,
0.873046875,
0.28076171875,
0.85546875,
0.28076171875
],
[
0.42822265625,
0.281982421875,
0.60546875,
0.281982421875,
0.60546875,
0.295654296875,
0.42822265625,
0.295654296875
],
[
0.2091064453125,
0.273681640625,
0.214111328125,
0.275634765625,
0.2091064453125,
0.28076171875,
0.204833984375,
0.279052734375
],
[
0.1175537109375,
0.308837890625,
0.8818359375,
0.308837890625,
0.8818359375,
0.322509765625,
0.1175537109375,
0.322509765625
],
[
0.1175537109375,
0.322509765625,
0.8818359375,
0.322509765625,
0.8818359375,
0.336181640625,
0.1175537109375,
0.336181640625
],
[
0.1158447265625,
0.33837890625,
0.69873046875,
0.33837890625,
0.69873046875,
0.3515625,
0.1158447265625,
0.3515625
],
[
0.1175537109375,
0.37890625,
0.489501953125,
0.37890625,
0.489501953125,
0.392578125,
0.1175537109375,
0.392578125
],
[
0.51123046875,
0.37890625,
0.88330078125,
0.37890625,
0.88330078125,
0.392578125,
0.51123046875,
0.392578125
],
[
0.1158447265625,
0.393798828125,
0.486083984375,
0.393798828125,
0.486083984375,
0.40673828125,
0.1158447265625,
0.40673828125
],
[
0.5087890625,
0.393798828125,
0.8818359375,
0.394287109375,
0.8818359375,
0.4091796875,
0.5087890625,
0.407958984375
],
[
0.1158447265625,
0.4091796875,
0.48779296875,
0.410400390625,
0.48779296875,
0.424560546875,
0.1158447265625,
0.42333984375
],
[
0.51123046875,
0.4091796875,
0.8818359375,
0.410400390625,
0.8818359375,
0.424560546875,
0.51123046875,
0.42333984375
],
[
0.1158447265625,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.1158447265625,
0.440673828125
],
[
0.509765625,
0.427001953125,
0.88330078125,
0.427001953125,
0.88330078125,
0.440673828125,
0.509765625,
0.440673828125
],
[
0.1158447265625,
0.4423828125,
0.48779296875,
0.4423828125,
0.48779296875,
0.4560546875,
0.1158447265625,
0.4560546875
],
[
0.509765625,
0.443603515625,
0.8818359375,
0.443603515625,
0.8818359375,
0.457275390625,
0.509765625,
0.457275390625
],
[
0.1175537109375,
0.458984375,
0.489501953125,
0.458984375,
0.489501953125,
0.47265625,
0.1175537109375,
0.47265625
],
[
0.51318359375,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.1175537109375,
0.474365234375,
0.49365234375,
0.474365234375,
0.49365234375,
0.488037109375,
0.1175537109375,
0.488037109375
],
[
0.51318359375,
0.474365234375,
0.88330078125,
0.474365234375,
0.88330078125,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.1175537109375,
0.4912109375,
0.49365234375,
0.4912109375,
0.49365234375,
0.5048828125,
0.1175537109375,
0.5048828125
],
[
0.51318359375,
0.4912109375,
0.884765625,
0.4912109375,
0.884765625,
0.5048828125,
0.51318359375,
0.5048828125
],
[
0.1158447265625,
0.50732421875,
0.490234375,
0.505859375,
0.490234375,
0.52099609375,
0.1158447265625,
0.521484375
],
[
0.5087890625,
0.50732421875,
0.8818359375,
0.505859375,
0.8818359375,
0.52099609375,
0.5087890625,
0.521484375
],
[
0.1175537109375,
0.52392578125,
0.490234375,
0.52392578125,
0.490234375,
0.53759765625,
0.1175537109375,
0.53759765625
],
[
0.5087890625,
0.52392578125,
0.880859375,
0.52392578125,
0.880859375,
0.53759765625,
0.5087890625,
0.53759765625
],
[
0.1158447265625,
0.5380859375,
0.48779296875,
0.5380859375,
0.48779296875,
0.5517578125,
0.1158447265625,
0.5517578125
],
[
0.5087890625,
0.5380859375,
0.88330078125,
0.5390625,
0.88330078125,
0.55419921875,
0.5087890625,
0.552734375
],
[
0.1158447265625,
0.55419921875,
0.489501953125,
0.55419921875,
0.489501953125,
0.568359375,
0.1158447265625,
0.568359375
],
[
0.509765625,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.509765625,
0.5693359375
],
[
0.1158447265625,
0.57177734375,
0.489501953125,
0.57177734375,
0.489501953125,
0.5849609375,
0.1158447265625,
0.5849609375
],
[
0.509765625,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.509765625,
0.5849609375
],
[
0.1158447265625,
0.58740234375,
0.48779296875,
0.5859375,
0.48779296875,
0.60107421875,
0.1158447265625,
0.6015625
],
[
0.5087890625,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.60107421875,
0.5087890625,
0.60107421875
],
[
0.1158447265625,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.6162109375,
0.1158447265625,
0.6162109375
],
[
0.51123046875,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.615234375,
0.51123046875,
0.615234375
],
[
0.1175537109375,
0.619140625,
0.441650390625,
0.619140625,
0.441650390625,
0.6328125,
0.1175537109375,
0.6328125
],
[
0.5087890625,
0.619140625,
0.88330078125,
0.6181640625,
0.88330078125,
0.6328125,
0.5087890625,
0.63427734375
],
[
0.133544921875,
0.63525390625,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.133544921875,
0.6494140625
],
[
0.509765625,
0.63525390625,
0.8818359375,
0.63427734375,
0.8818359375,
0.6484375,
0.509765625,
0.6494140625
],
[
0.1158447265625,
0.65185546875,
0.489501953125,
0.65185546875,
0.489501953125,
0.6650390625,
0.1158447265625,
0.6650390625
],
[
0.509765625,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.66455078125,
0.509765625,
0.66455078125
],
[
0.1158447265625,
0.66748046875,
0.486083984375,
0.66748046875,
0.486083984375,
0.68115234375,
0.1158447265625,
0.68115234375
],
[
0.509765625,
0.66650390625,
0.8818359375,
0.66748046875,
0.8818359375,
0.681640625,
0.509765625,
0.68115234375
],
[
0.1158447265625,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.1158447265625,
0.69873046875
],
[
0.51123046875,
0.68505859375,
0.87158203125,
0.68505859375,
0.87158203125,
0.69873046875,
0.51123046875,
0.69873046875
],
[
0.1158447265625,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.1158447265625,
0.71337890625
],
[
0.53076171875,
0.70166015625,
0.8818359375,
0.70166015625,
0.8818359375,
0.71533203125,
0.53076171875,
0.71533203125
],
[
0.1158447265625,
0.71630859375,
0.489501953125,
0.71630859375,
0.489501953125,
0.73095703125,
0.1158447265625,
0.73095703125
],
[
0.509765625,
0.71630859375,
0.8818359375,
0.71728515625,
0.8818359375,
0.732421875,
0.509765625,
0.73095703125
],
[
0.1175537109375,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.1175537109375,
0.74755859375
],
[
0.5087890625,
0.73291015625,
0.880859375,
0.73291015625,
0.880859375,
0.74658203125,
0.5087890625,
0.74658203125
],
[
0.1175537109375,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.76123046875,
0.1175537109375,
0.76123046875
],
[
0.51123046875,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.51123046875,
0.76318359375
],
[
0.11419677734375,
0.76416015625,
0.490234375,
0.765625,
0.490234375,
0.77978515625,
0.11419677734375,
0.779296875
],
[
0.5146484375,
0.76708984375,
0.880859375,
0.76708984375,
0.880859375,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.1158447265625,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.1158447265625,
0.7958984375
],
[
0.5087890625,
0.783203125,
0.8818359375,
0.78076171875,
0.8818359375,
0.79443359375,
0.5087890625,
0.79638671875
],
[
0.1175537109375,
0.79736328125,
0.489501953125,
0.79736328125,
0.489501953125,
0.8125,
0.1175537109375,
0.8125
],
[
0.51123046875,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51123046875,
0.8125
],
[
0.11419677734375,
0.8134765625,
0.486083984375,
0.8134765625,
0.486083984375,
0.82763671875,
0.11419677734375,
0.82763671875
],
[
0.51123046875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.8291015625,
0.51123046875,
0.8291015625
],
[
0.11419677734375,
0.8291015625,
0.476806640625,
0.830078125,
0.476806640625,
0.84423828125,
0.11419677734375,
0.84326171875
],
[
0.509765625,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.509765625,
0.84423828125
],
[
0.51318359375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.85986328125,
0.51318359375,
0.85986328125
],
[
0.1158447265625,
0.85986328125,
0.3837890625,
0.85986328125,
0.3837890625,
0.87353515625,
0.1158447265625,
0.87353515625
],
[
0.51123046875,
0.8623046875,
0.8818359375,
0.8623046875,
0.8818359375,
0.876953125,
0.51123046875,
0.876953125
],
[
0.1158447265625,
0.88720703125,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.52294921875,
0.88818359375,
0.759765625,
0.89013671875,
0.759765625,
0.90771484375,
0.52294921875,
0.90625
],
[
0.1158447265625,
0.90283203125,
0.489501953125,
0.90380859375,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91748046875
],
[
0.5283203125,
0.90478515625,
0.8330078125,
0.90478515625,
0.8330078125,
0.91943359375,
0.5283203125,
0.91943359375
]
] |
[
"pruneTracln-hard",
"prunePVI-hard",
"prunerandom",
"IpruneVoG-easy",
"86",
"85",
"[% e ",
"84",
"labelnoise=5%",
"10%",
"86",
"85",
"84",
"15%",
"20%",
"86",
"85",
"84",
"25%",
"30%",
"10",
"830",
"15",
"20",
"25",
"30",
"35",
"40",
"10",
"15",
"20",
"25",
"30",
"35",
"40",
"45",
"training data pruned [%]",
"",
"Figure 3: Shows the results of pruning VoG-easy (blue), TracIn-hard (orange), PVI-hard (purple), or randomly",
"selected (gold) examples for noisy SNLI. The amount of added isotropic label noise is given in the orange inset",
"Points indicate the average of three runs and shaded regions indicate the l-g envelope.",
"noisy settings, for large pruning fractions. In some",
"typically found in large, general-purpose commer-",
"cases, pruning based on VoG scores even clawed",
"cial voice assistants. This setting poses practical",
"back a fraction of the initial accuracy loss due to",
"challenges often not reflected in public datasets,",
"noisy labels. However, somewhat surprisingly, this",
"such as noisy and evolving speech patterns, di-",
"was likely not because VoG-based selection was",
"verse vocabularies, dialects, carrier phrases, and",
"pruning the misannotated examples themselves.",
"out-of-distribution named entities.As an added",
"The similarity index between the easiest VoG exam-",
"challenge, we focus on Japanese-data trained mod-",
"ples and all of the introduced misannotated exam-",
"els and datasets to determine if VoG-based influ-",
"ples in the 30% label-noise setting was only ~ 0.11.",
"ence scoring could function in a lower-resource",
"Compared to random sampling (~ 0.11), we con-",
"setting. The statistical-model component in this",
"clude that VoG does not do better than random",
"setting consisted of generic domain classifier (DC),",
"chance at finding misannotated SNLI examples,",
"intent classifier (IC), and named entity recognition",
"but does reliably extract more influential examples",
"(NER) models organized in a coupled, hierarchical",
"that partially mitigate the effects of label noise. In",
" manner: a single multi-class DC model is trained",
"all noisy settings, we found that pruning VoG-hard",
"onalldomainsdatatofirstpredictthedomainfor",
"examples did worse than pruning randomly.",
"a given utterance, which then invokes a domain-",
"Pruning by TracIn and PVI scores resulted in a",
"specific joint IC-NER model trained on in-domain",
"small but persistent accuracy increase when ~ 5-",
"datall. Both sets of models were based on dis",
"10% of the hard data was pruned. In the 30% noise",
"tilled Japanese BERT models and VoG scores were",
"setting, the similarity index between the Tracln-",
"computed using the procedure given in App. B.1",
"hard and PVI-hard examples and misannotated ex-",
"Fig. 4 shows the class-normalized scores for a",
"amples was ~ 0.11, 0.09, respectively, again indi-",
"subset of five domains that differ in the proportion",
"cating that the accuracy gains are not due to the",
" of the overall training data they represent and intent",
"removal of defects.The number of instances with",
"label complexity12 within that domain.We observe",
"a PVI score of < o (indicating a potential mis-",
"that smallerdomainstend tohavehigherscores",
"annotation) comprises only 6% of the mislabeled",
"(e.g.,HealthAndFitness vs.HomeAutomation)and",
"data. Nevertheless, it appears beneficial to use",
"more complex domains tend to have higher scores",
"these scores to prune 5-10% of overall hard data",
"(Shopping vs. Video). In some cases, domains",
"that adversely impacts training in a noisy setting.",
"that are similar in size and complexity still exhibit",
"different VoG score distributions (Music vs. Video)",
"3VoG in the Context of NLU",
"which reflects differing influence of domain data",
"Given its promising results for data reduction on",
"I\"See App. C.1 for additional context.",
"SNLI, we set out to evaluate VoG in an environment",
"12As measured by Shannon entropy; see App. C.6."
] |
{
"arxiv_id": "2311.16298",
"language": "en",
"page_idx": 4
}
|
||
2401.17979-en-0
|
2401.17979-en-0.jpg
|
$\textbf{Entity Linking in the Job Market Domain}$
Mike Zhang$^{O\text{由}}$ Rob van der Goot$^{O\text{由}}$ Barbara Plank$^{O\Delta\pi}$
$o_{\mathrm{Department~of~Computer~Science,~IT~University~of~Copenhagen,~Denmark}}$
$^{\textcircled{1}}$Pioneer Centre for Artificial Intelligence, Copenhagen, Denmark
$\Delta$MaiNLP, Center for Information and Language Processing, LMU Munich, Germany
$^{\text{ಸ್ಗ}}$Munich Center for Machine Learning (MCML), Munich, Germany
mikejj.zhang@gmail.com
### $\mathbf{Abstract}$
In Natural Language Processing, entity linking (EL) has centered around Wikipedia, but remains underexplored for the job market domain. Disambiguating skill mentions can help us to get insight into the labor market demands. In this work, we are the first to explore EL in this domain, specifically targeting the linkage of occupational skills to the ESCO taxonomy (le $\hat{\text{Vrang et al.,2014).Previous ef-}}$ forts linked coarse-grained (full) sentences to a corresponding ESCO skill. In this work, we link more fine-grained span-level mentions of skills. We tune two high-performing neural EL models, a bi-encoder (Wu et al., 2020) and an autoregressive model (Cao et al., 2021), on a synthetically generated mentionskill pair dataset and evaluate them on a humanannotated skill-linking benchmark. Our findings reveal that both models are capable of linking implicit mentions of skills to their correct taxonomy counterparts. Empirically, BLINK outperforms GENRE in strict evaluation, but GENRE performs better in loose evaluation (accuracy@$k).^1$
most commonly Wikipedia (He et al., 2013). Recent models address this problem by producing entity representations from a (sub)set of KB information, e.g., entity descriptions (Logeswaran et al., 2019; Wu et al., 2020), fine-grained entity types (Raiman and Raiman, 2018; Onoe and Durrett, 2020; Ayoola et al., 2022), or generation of the input text autoregressively (Cao et al., 2021, 2022). For skill linking specifically, we use the European Skills, Competences, Qualifications and Occupations (ESCO; le Vrang et al., 2014) taxonomy due to its comprehensiveness. Previous work classified spans to its taxonomy code via multi-class classification (Zhang et al., 2022b) without surrounding context and neither the full breadth of ESCO. Gnehm et al. (2022) approaches it as a sequence labeling task, but only uses more coarsegrained ESCO concepts, and not the full taxonomy. Last, others attempt to match the full sentence to their respective taxonomy title (Decorte et al., 2022, 2023; Clavié and Soulié, 2023).
The latter comes with a limitation: The taxonomy title does not indicate which subspan in the sentence it points to, without an exact match. We define this as an implicit skill, where mentions (spans) in the sentence do not have an exact string match with a skill in the ESCO taxonomy. The differences can range from single tokens to entire phrases. For example, we can link “being able to work together" to“plan teamwork”.2 If we know the exact span, this implicit skill can be added to the taxonomy as an alternative choice for the surface skill. As a result, this gives us a more nuanced view of the labor market skill demands. Therefore, we attempt to train models to the linking of both implicit and explicit skill mentions.
Contributions. Our findings can be summarized
## 1 Introduction
Labor market dynamics, influenced by technological changes, migration, and digitization, have led to the availability of job descriptions (JD) on platforms to attract qualified candidates (Brynjolfsson and McAfee, 2011, 2014; Balog et al., 2012). It is important to extract and link surface form skills to a unique taxonomy entry, allowing us to quantify the current labor market dynamics and determine the demands and needs. We attempt to tackle the problem of entity linking (EL) in the job market domain, specifically the linking of fine-grained span-level skill mentions to a specific taxonomy entry.
Generally, EL is the task of linking mentions of entities in unstructured text documents to their respective unique entities in a knowledge base (KB),
# as follows: ◯ We pose the task of skill linking as an entity linking problem, showing promising
IThe source code and data can be found at https://
github.com/mainlp/el_esco
$^{2}$See example here: https://t.1y/3VUJG.
|
[
[
0.28369140625,
0.08905029296875,
0.7138671875,
0.08905029296875,
0.7138671875,
0.10394287109375,
0.28369140625,
0.10394287109375
],
[
0.385498046875,
0.12054443359375,
0.5810546875,
0.1175537109375,
0.5810546875,
0.135986328125,
0.385498046875,
0.1395263671875
],
[
0.599609375,
0.12054443359375,
0.779296875,
0.11639404296875,
0.779296875,
0.1353759765625,
0.599609375,
0.1395263671875
],
[
0.193115234375,
0.13720703125,
0.81005859375,
0.1395263671875,
0.81005859375,
0.15673828125,
0.193115234375,
0.155029296875
],
[
0.2225341796875,
0.1217041015625,
0.364501953125,
0.1175537109375,
0.365234375,
0.1353759765625,
0.22412109375,
0.1395263671875
],
[
0.2275390625,
0.15380859375,
0.77490234375,
0.15673828125,
0.77490234375,
0.1776123046875,
0.2275390625,
0.174560546875
],
[
0.151123046875,
0.1728515625,
0.85400390625,
0.17578125,
0.85400390625,
0.196533203125,
0.151123046875,
0.193603515625
],
[
0.2225341796875,
0.1923828125,
0.77734375,
0.1953125,
0.77734375,
0.213134765625,
0.2225341796875,
0.210205078125
],
[
0.366943359375,
0.21435546875,
0.63818359375,
0.21435546875,
0.63818359375,
0.22802734375,
0.366943359375,
0.22802734375
],
[
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
],
[
0.1435546875,
0.2724609375,
0.459228515625,
0.273681640625,
0.459228515625,
0.2880859375,
0.1435546875,
0.287353515625
],
[
0.51123046875,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.145263671875,
0.2880859375,
0.459228515625,
0.2880859375,
0.459228515625,
0.302734375,
0.145263671875,
0.302734375
],
[
0.51318359375,
0.2861328125,
0.88330078125,
0.2861328125,
0.88330078125,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.14697265625,
0.303955078125,
0.45751953125,
0.303955078125,
0.45751953125,
0.314208984375,
0.14697265625,
0.314208984375
],
[
0.51123046875,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51123046875,
0.31640625
],
[
0.0218353271484375,
0.312255859375,
0.05877685546875,
0.312255859375,
0.0621337890625,
0.438232421875,
0.0251922607421875,
0.438232421875
],
[
0.145263671875,
0.317626953125,
0.45751953125,
0.317626953125,
0.45751953125,
0.330810546875,
0.145263671875,
0.330810546875
],
[
0.51123046875,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.33203125,
0.51123046875,
0.33203125
],
[
0.145263671875,
0.330810546875,
0.459228515625,
0.330810546875,
0.459228515625,
0.344482421875,
0.145263671875,
0.344482421875
],
[
0.51318359375,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.145263671875,
0.345703125,
0.45751953125,
0.345703125,
0.45751953125,
0.359375,
0.145263671875,
0.359375
],
[
0.51123046875,
0.349853515625,
0.880859375,
0.350830078125,
0.880859375,
0.365234375,
0.51123046875,
0.364013671875
],
[
0.14697265625,
0.35986328125,
0.459228515625,
0.35986328125,
0.459228515625,
0.37353515625,
0.14697265625,
0.37353515625
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.1435546875,
0.37353515625,
0.459228515625,
0.372314453125,
0.459228515625,
0.38720703125,
0.1435546875,
0.388427734375
],
[
0.5283203125,
0.3818359375,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.5283203125,
0.396728515625
],
[
0.145263671875,
0.388427734375,
0.4609375,
0.388427734375,
0.4609375,
0.402099609375,
0.145263671875,
0.402099609375
],
[
0.14697265625,
0.40380859375,
0.45751953125,
0.40380859375,
0.45751953125,
0.41455078125,
0.14697265625,
0.41455078125
],
[
0.51123046875,
0.399658203125,
0.8818359375,
0.3984375,
0.8818359375,
0.413330078125,
0.51123046875,
0.41455078125
],
[
0.14697265625,
0.418701171875,
0.45751953125,
0.418701171875,
0.45751953125,
0.4287109375,
0.14697265625,
0.4287109375
],
[
0.51123046875,
0.4150390625,
0.880859375,
0.416259765625,
0.880859375,
0.429931640625,
0.51123046875,
0.4287109375
],
[
0.042816162109375,
0.429931640625,
0.050384521484375,
0.432861328125,
0.047027587890625,
0.435791015625,
0.039459228515625,
0.432861328125
],
[
0.145263671875,
0.43115234375,
0.4609375,
0.43115234375,
0.4609375,
0.44482421875,
0.145263671875,
0.44482421875
],
[
0.509765625,
0.431640625,
0.880859375,
0.431640625,
0.880859375,
0.4453125,
0.509765625,
0.4453125
],
[
0.04534912109375,
0.435791015625,
0.050384521484375,
0.435791015625,
0.050384521484375,
0.441162109375,
0.04534912109375,
0.441162109375
],
[
0.0251922607421875,
0.435302734375,
0.0621337890625,
0.435302734375,
0.0621337890625,
0.51708984375,
0.0251922607421875,
0.51708984375
],
[
0.14697265625,
0.44482421875,
0.45751953125,
0.4453125,
0.45751953125,
0.458984375,
0.14697265625,
0.457763671875
],
[
0.51123046875,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.51123046875,
0.4619140625
],
[
0.1485595703125,
0.458984375,
0.459228515625,
0.458984375,
0.459228515625,
0.47265625,
0.1485595703125,
0.47265625
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.51123046875,
0.478515625
],
[
0.145263671875,
0.471435546875,
0.4609375,
0.47265625,
0.4609375,
0.48681640625,
0.145263671875,
0.486328125
],
[
0.145263671875,
0.486328125,
0.459228515625,
0.48681640625,
0.459228515625,
0.50048828125,
0.145263671875,
0.49951171875
],
[
0.51123046875,
0.48095703125,
0.88330078125,
0.48095703125,
0.88330078125,
0.49462890625,
0.51123046875,
0.49462890625
],
[
0.14697265625,
0.501953125,
0.4609375,
0.501953125,
0.4609375,
0.515625,
0.14697265625,
0.515625
],
[
0.51318359375,
0.496337890625,
0.88330078125,
0.496337890625,
0.88330078125,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.1485595703125,
0.51708984375,
0.45751953125,
0.51708984375,
0.45751953125,
0.52783203125,
0.1485595703125,
0.52783203125
],
[
0.51123046875,
0.5126953125,
0.88330078125,
0.5126953125,
0.88330078125,
0.52685546875,
0.51123046875,
0.52685546875
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.145263671875,
0.52978515625,
0.4609375,
0.52978515625,
0.4609375,
0.54345703125,
0.145263671875,
0.54345703125
],
[
0.51123046875,
0.529296875,
0.88330078125,
0.529296875,
0.88330078125,
0.5419921875,
0.51123046875,
0.5419921875
],
[
0.145263671875,
0.5458984375,
0.45751953125,
0.54443359375,
0.45751953125,
0.5556640625,
0.145263671875,
0.55712890625
],
[
0.51123046875,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.51123046875,
0.55810546875
],
[
0.145263671875,
0.55859375,
0.459228515625,
0.55859375,
0.459228515625,
0.572265625,
0.145263671875,
0.572265625
],
[
0.51123046875,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51123046875,
0.57373046875
],
[
0.145263671875,
0.572265625,
0.459228515625,
0.572265625,
0.459228515625,
0.5859375,
0.145263671875,
0.5859375
],
[
0.51123046875,
0.57666015625,
0.74658203125,
0.57666015625,
0.74658203125,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.145263671875,
0.58740234375,
0.459228515625,
0.58740234375,
0.459228515625,
0.60107421875,
0.145263671875,
0.60107421875
],
[
0.52880859375,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.52880859375,
0.60693359375
],
[
0.1435546875,
0.60107421875,
0.252685546875,
0.60107421875,
0.252685546875,
0.615234375,
0.1435546875,
0.615234375
],
[
0.51123046875,
0.60986328125,
0.8818359375,
0.60986328125,
0.8818359375,
0.62353515625,
0.51123046875,
0.62353515625
],
[
0.1435546875,
0.62158203125,
0.26025390625,
0.62158203125,
0.26025390625,
0.6357421875,
0.1435546875,
0.6357421875
],
[
0.509765625,
0.6259765625,
0.8818359375,
0.62451171875,
0.8818359375,
0.6396484375,
0.509765625,
0.64013671875
],
[
0.509765625,
0.64111328125,
0.8818359375,
0.64013671875,
0.8818359375,
0.65478515625,
0.509765625,
0.65625
],
[
0.12176513671875,
0.626953125,
0.130126953125,
0.626953125,
0.130126953125,
0.6318359375,
0.12176513671875,
0.6318359375
],
[
0.1175537109375,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.66015625,
0.1175537109375,
0.66015625
],
[
0.1175537109375,
0.6630859375,
0.48779296875,
0.6630859375,
0.48779296875,
0.6767578125,
0.1175537109375,
0.6767578125
],
[
0.509765625,
0.65771484375,
0.8818359375,
0.65771484375,
0.8818359375,
0.6728515625,
0.509765625,
0.6728515625
],
[
0.1175537109375,
0.6787109375,
0.490234375,
0.6787109375,
0.490234375,
0.6923828125,
0.1175537109375,
0.6923828125
],
[
0.509765625,
0.67333984375,
0.8818359375,
0.6728515625,
0.8818359375,
0.68701171875,
0.509765625,
0.6884765625
],
[
0.1158447265625,
0.693359375,
0.486083984375,
0.693359375,
0.486083984375,
0.70703125,
0.1158447265625,
0.70703125
],
[
0.51123046875,
0.6904296875,
0.8818359375,
0.6904296875,
0.8818359375,
0.70361328125,
0.51123046875,
0.70361328125
],
[
0.509765625,
0.705078125,
0.880859375,
0.705078125,
0.880859375,
0.7197265625,
0.509765625,
0.7197265625
],
[
0.1175537109375,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.1158447265625,
0.72705078125,
0.48779296875,
0.7255859375,
0.48779296875,
0.74072265625,
0.1158447265625,
0.7412109375
],
[
0.51123046875,
0.72265625,
0.8818359375,
0.72265625,
0.8818359375,
0.736328125,
0.51123046875,
0.736328125
],
[
0.509765625,
0.73828125,
0.880859375,
0.73828125,
0.880859375,
0.751953125,
0.509765625,
0.751953125
],
[
0.1158447265625,
0.74462890625,
0.484375,
0.74462890625,
0.484375,
0.7548828125,
0.1158447265625,
0.7548828125
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.7548828125,
0.8818359375,
0.7548828125,
0.8818359375,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.7724609375,
0.880859375,
0.7724609375,
0.880859375,
0.783203125,
0.51123046875,
0.783203125
],
[
0.5087890625,
0.78759765625,
0.88330078125,
0.7861328125,
0.88330078125,
0.7998046875,
0.5087890625,
0.80029296875
],
[
0.118408203125,
0.79345703125,
0.48779296875,
0.79345703125,
0.48779296875,
0.80419921875,
0.118408203125,
0.80419921875
],
[
0.1158447265625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.802734375,
0.8818359375,
0.802734375,
0.8818359375,
0.81787109375,
0.509765625,
0.81787109375
],
[
0.1158447265625,
0.822265625,
0.438232421875,
0.82373046875,
0.438232421875,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.51123046875,
0.818359375,
0.77490234375,
0.818359375,
0.77490234375,
0.83203125,
0.51123046875,
0.83203125
],
[
0.1351318359375,
0.8388671875,
0.489501953125,
0.8388671875,
0.489501953125,
0.8525390625,
0.1351318359375,
0.8525390625
],
[
0.51123046875,
0.84423828125,
0.8818359375,
0.845703125,
0.8818359375,
0.85986328125,
0.51123046875,
0.859375
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.5087890625,
0.86083984375,
0.8818359375,
0.8623046875,
0.8818359375,
0.876953125,
0.5087890625,
0.8759765625
],
[
0.1175537109375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.5087890625,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.8935546875,
0.5087890625,
0.8935546875
],
[
0.139404296875,
0.892578125,
0.48779296875,
0.892578125,
0.48779296875,
0.90673828125,
0.139404296875,
0.90673828125
],
[
0.1175537109375,
0.90673828125,
0.34765625,
0.90673828125,
0.34765625,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.5341796875,
0.90478515625,
0.82373046875,
0.90478515625,
0.82373046875,
0.91845703125,
0.5341796875,
0.91845703125
]
] |
[
"Entity Linking in the Job Market Domain",
"Rob van der Gooto",
"Barbara PlankoA",
"ODepartment of Computer Science, IT University of Copenhagen, Denmark",
"Mike Zhang由",
"pioneer Centre for Artificial Intelligence, Copenhagen, Denmark",
"AMaiNLP, Center for Information and Language Processing, LMU Munich, Germany",
"iMunich Center for Machine Learning (MCML), Munich, Germany",
"mikejj.zhang@gmail.com",
"Abstract",
"most commonly Wikipedia (He et al., 2013). Re-",
"In Natural Language Processing, entity link-",
"cent models address this problem by producing",
"ing (EL) has centered around Wikipedia, but",
"entity representations from a (sub)set of KB in-",
"remains underexploredforthejob marketdo-",
"formation, e.g., entity descriptions (Logeswaran",
" ",
"main. Disambiguating skill mentions can help",
"et al.,2019; Wu et al., 2020), fine-grained entity",
"us to get insight into the labor market demands",
"types (Raiman and Raiman, 2018; Onoe and Dur-",
"In this work, we are the first to explore EL",
"rett, 2020; Ayoola et al., 2022), or generation of the",
"in this domain, specifically targeting the link-",
"input text autoregressively (Cao et al., 2021, 2022)",
"age of occupational skills to the ESCO tax-",
"For skill linking specifically, we use the Euro",
"onomy (le Vrang et al., 2014). Previous ef-",
"forts linked coarse-grained (full) sentences to",
"pean Skills, Competences, Qualifications and Oc-",
"acorrespondingESCO skill.In this work.",
"cupations (ESCO; le Vrang et al., 2014) taxonomy",
"",
"we link more fine-grained span-level men-",
"due to its comprehensiveness. Previous work clas",
"",
"['S]",
"tions of skills. We tune two high-performing",
"sified spans to its taxonomy code via multi-class",
"neural EL models, a bi-encoder (Wu et al.,",
"classification (Zhang et al., 2022b) without sur-",
"2020) and an autoregressive model (Cao et al.,",
"2021),on a synthetically generated mention-",
"rounding context and neither the full breadth of",
"skill pair dataset and evaluate them on a human-",
"ESCO. Gnehm et al. (2022) approaches it as a se-",
"annotated skill-linking benchmark. Our find-",
"quence labeling task, but only uses more coarse-",
"1A6261102:A111",
"ings reveal that both models are capable of link-",
"grained ESCO concepts, and not the full taxonomy.",
"ing implicit mentions of skills to their correct",
"Last, others attempt to match the full sentence to",
"taxonomy counterparts. Empirically, BLINK",
"their respective taxonomy title (Decorte et al., 2022.",
"outperforms GENRE in strict evaluation, but",
"2023; Clavié and Soulié, 2023).",
"GENRE performs better in loose evaluation",
"The latter comes with a limitation: The taxon-",
"(accuracy@k).!",
"omy title does not indicate which subspan in the",
"Introduction",
"sentence it points to, without an exact match. We",
"define this as an implicit skill, where mentions",
"",
"Labor market dynamics, influenced by technologi-",
"cal changes, migration, and digitization, have led",
"(spans) in the sentence do not have an exact string",
"to the availability of job descriptions (JD) on plat-",
"match with a skill in the ESCO taxonomy. The",
"forms to attract qualified candidates (Brynjolfsson",
"differences can range from single tokens to entire",
"phrases. For example, we can link being able to",
"and McAfee, 2011, 2014; Balog et al., 2012). It is",
"important to extract and link surface form skills to a",
"work together\"to\"plan teamwork\"2 If we know",
"the exact span, this implicit skill can be added to",
"uniguetaxonomy entry.allowing us to guantifythe",
"current labor market dynamics and determine the",
"the taxonomy as an alternative choice for the sur-",
"demands and needs. We attempt to tackle the prob-",
"face skill.As a result, this gives us a more nuanced",
"view of the labor market skill demands. Therefore,",
"lemof entitylinking(EL)inthejobmarketdomain,",
"specifically the linking of fine-grained span-level",
"we attempt to train models to the linking of both",
"skill mentions to a specific taxonomy entry.",
"implicit and explicit skill mentions.",
"Generally, EL is the task of linking mentions of",
"Contributions.Our findings can be summarized",
"entities in unstructured text documents to their re-",
"as follows: We pose the task of skill linking",
"spective unique entities in a knowledge base (KB),",
" as an entity linking problem, showing promising",
"IThe source code and data can be found at https : / /",
"github.com/mainlp/el_esco",
"2See example here: https ://t.ly/3VUJG."
] |
{
"arxiv_id": "2401.17979",
"language": "en",
"page_idx": 0
}
|
||
2401.17979-en-1
|
2401.17979-en-1.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Instances</th>
<th>${\mathrm{Unique}}$ Titles</th>
<th>$UNK$</th>
</tr>
<tr>
<td>Train</td>
<td>123,619</td>
<td>12,984</td>
<td>14,641</td>
</tr>
<tr>
<td>$Dev.$</td>
<td>480</td>
<td>149</td>
<td>233</td>
</tr>
<tr>
<td>Test</td>
<td>1,824</td>
<td>455</td>
<td>813</td>
</tr>
</tbody>
</table>
Table 1: Data Statistics. Data distribution of train, dev, and test splits. UNK indicates skills mentions that are not linked to a corresponding taxonomy title.
results of linking with two entity linking systems. $\textcircled2 We present a qualitative analysis showing that$ the model successfully links implicit skills to their respective skill entry in ESCO.
# 2 Methodology
Definition. In EL,we process the input document ESCO in our case: $\mathcal{E}=\{e_1,\ldots,e_{13890}$,UNK$\}.$ The objective of an EL model is to generate a list of mention-entity pairs $\{(m_i,e_i)\}_{i=1}^n$, where each entity $e$ corresponds to an entry in a KB. We assume that both the titles and descriptions of the entities are available, which is a common scenario in EL research (Ganea and Hofmann, 2017; Logeswaran et al., 2019; Wu et al., 2020). We also assume that each mention in the document has a corresponding valid gold entity present in the knowledge base, including UNK. This scenario is typically referred to as “in-KB evaluation”. Similar to prior research efforts (Logeswaran et al., 2019; Wu et al., 2020), we also presuppose that the mentions within the document have already been tagged.
Data. We use ESCO titles as ground truth labels, containing 13,890 skills.$^{3}$ Table 1 presents the train, dev, and test data in our experiments. We leverage the train set introduced by Decorte et al. $(2023)^4 along with the dev and test sets provided$ in Decorte et al.(2022).$^{5}$ The train set is synthetically generated by Decorte et al. (2023) with the gpt-3.5-turbo-0301 model (OpenAI, 2023). Specifically, this involves taking each skill from ESCO and prompting the model to generate sentences resembling JD sentences that require that particular skill. The dev and test splits, conversely, are derived from actual job advertisements sourced from the study by Zhang et al.(2022a). These
$^{3}$Per version 1.1.1, accessed on 01 August 2023.
$\overset{4}{\operatorname*{\operatorname*{\operatorname*{https://t.ly/edqkp}}}}$ $^{5}$https://t.ly/Lcq$Q7$
JDs are annotated with spans corresponding to specific skills, and these spans have subsequently been manually linked to ESCO, as described in the work of Decorte et al.(2022). In cases where skills cannot be linked, two labels are used, namely UNDERSPECIFIED and LABEL NOT PRESENT. For the sake of uniformity, we map both of these labels to a generic UNK tag. We used several heuristics based on Levenshtein distance and sentence similarity to find the exact subspans if it exceeds certain thresholds, otherwise, it is UNK. This process is outlined in Appendix A. In addition, some data examples can be found in Appendix B. The number of UNKs in the data is also in Table 1. During inference, the UNK title is a prediction option for the models.
Models. We use two EL models, selected for
$\mathcal{D}$ = $\{ w_1, \ldots , w_r\} $, a collection of entity men- Models. We use two EL models, selected fons dired $\mathcal{D}_{\mathrm{ions~denoten~}}$ acollection of entity men- $\textbf{their robust performance in EL on Wikipedia. }^{6}$
BLINK (Wu et al., 2020). BLINK uses a biencoder architecture based on BERT (Devlin et al., 2019), for modeling pairs of mentions and entities. The model processes two inputs:
[CLS] ctxt$_\mathrm{l}[S]$ mention[E] ctxtr[SEP]
Where“mention”,“ctxt_l”, and“ctxt_r”corresponds to the wordpiece tokens of the mention, the left context, and the right context. The mention is denoted by special tokens [S] and [E]. The entity and its description are structured as follows:
[ CLS ] title [ ENT ] description [ SEP ]
Here,“title” and“description” represent the wordpiece tokens of the skills' title and description, respectively. [ENT] is a special token to separate the two representations. We train the model to maximize the dot product of the [ CLS] representation of the two inputs, for the correct skill in comparison to skills within the same batch. For each training $\operatorname{pair}\left(m_{i},e_{i}\right),\operatorname{the}\operatorname*{loss}_{B}$computed as$\mathcal{L}\left(m_{i},e_{i}\right)=$ -s$\left(m_{i},e_{i}\right)+\log\sum_{j=1}^{B}\exp\left(\mathrm{s}\left(m_{i},e_{j}\right)\right)$,where the objective is to minimize the distance between $m_i$ and $e_i$ while encouraging the model to assign a higher score to the correct pair and lower scores to randomly sampled incorrect pairs. Hard negatives are also used during training, these are obtained by finding the top 10 predicted skills for each training example. These extra hard negatives are added to the random in-batch negatives.
$^{6}$For the hyperparameter setups, we refer to Appendix C.
|
[
[
0.414794921875,
0.0872802734375,
0.453369140625,
0.0872802734375,
0.453369140625,
0.10272216796875,
0.414794921875,
0.10272216796875
],
[
0.509765625,
0.08489990234375,
0.87890625,
0.08489990234375,
0.87890625,
0.09857177734375,
0.509765625,
0.09857177734375
],
[
0.2822265625,
0.08905029296875,
0.385498046875,
0.08905029296875,
0.385498046875,
0.1009521484375,
0.2822265625,
0.1009521484375
],
[
0.21240234375,
0.0885009765625,
0.2880859375,
0.0885009765625,
0.2880859375,
0.09979248046875,
0.21240234375,
0.09979248046875
],
[
0.21826171875,
0.1051025390625,
0.27783203125,
0.10687255859375,
0.276123046875,
0.122314453125,
0.2166748046875,
0.12054443359375
],
[
0.334228515625,
0.10687255859375,
0.38623046875,
0.1051025390625,
0.387939453125,
0.12054443359375,
0.3359375,
0.122314453125
],
[
0.51123046875,
0.1015625,
0.880859375,
0.1009521484375,
0.880859375,
0.11517333984375,
0.51123046875,
0.11639404296875
],
[
0.154541015625,
0.10809326171875,
0.198974609375,
0.10809326171875,
0.198974609375,
0.122314453125,
0.154541015625,
0.122314453125
],
[
0.39794921875,
0.10809326171875,
0.45166015625,
0.10809326171875,
0.45166015625,
0.122314453125,
0.39794921875,
0.122314453125
],
[
0.4189453125,
0.1217041015625,
0.453369140625,
0.1217041015625,
0.453369140625,
0.13720703125,
0.4189453125,
0.13720703125
],
[
0.51123046875,
0.1175537109375,
0.8818359375,
0.1175537109375,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1561279296875,
0.12353515625,
0.193115234375,
0.12353515625,
0.193115234375,
0.135986328125,
0.1561279296875,
0.135986328125
],
[
0.2476806640625,
0.12353515625,
0.27783203125,
0.12353515625,
0.27783203125,
0.135986328125,
0.2476806640625,
0.135986328125
],
[
0.358642578125,
0.122314453125,
0.3896484375,
0.122314453125,
0.3896484375,
0.13720703125,
0.358642578125,
0.13720703125
],
[
0.355224609375,
0.135986328125,
0.3896484375,
0.135986328125,
0.3896484375,
0.1519775390625,
0.355224609375,
0.1519775390625
],
[
0.4189453125,
0.13720703125,
0.453369140625,
0.13720703125,
0.453369140625,
0.152587890625,
0.4189453125,
0.152587890625
],
[
0.51123046875,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.2376708984375,
0.1383056640625,
0.276123046875,
0.1383056640625,
0.276123046875,
0.15087890625,
0.2376708984375,
0.15087890625
],
[
0.1561279296875,
0.1383056640625,
0.189697265625,
0.1383056640625,
0.189697265625,
0.15087890625,
0.1561279296875,
0.15087890625
],
[
0.51123046875,
0.1484375,
0.880859375,
0.1484375,
0.880859375,
0.162109375,
0.51123046875,
0.162109375
],
[
0.1158447265625,
0.16748046875,
0.489501953125,
0.16748046875,
0.489501953125,
0.18115234375,
0.1158447265625,
0.18115234375
],
[
0.51123046875,
0.166259765625,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.193603515625,
0.1175537109375,
0.193603515625
],
[
0.51123046875,
0.1817626953125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.1953125
],
[
0.1158447265625,
0.196533203125,
0.420654296875,
0.196533203125,
0.420654296875,
0.210205078125,
0.1158447265625,
0.210205078125
],
[
0.509765625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.51123046875,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.1158447265625,
0.236328125,
0.489501953125,
0.2369384765625,
0.489501953125,
0.251708984375,
0.1158447265625,
0.25048828125
],
[
0.51318359375,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.51318359375,
0.24462890625
],
[
0.5087890625,
0.2452392578125,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.5087890625,
0.260009765625
],
[
0.1158447265625,
0.251708984375,
0.48779296875,
0.2529296875,
0.48779296875,
0.267333984375,
0.1158447265625,
0.2666015625
],
[
0.509765625,
0.262451171875,
0.8818359375,
0.262451171875,
0.8818359375,
0.275634765625,
0.509765625,
0.275634765625
],
[
0.1158447265625,
0.268310546875,
0.486083984375,
0.268310546875,
0.486083984375,
0.281982421875,
0.1158447265625,
0.281982421875
],
[
0.51318359375,
0.279052734375,
0.8818359375,
0.279052734375,
0.8818359375,
0.292236328125,
0.51318359375,
0.292236328125
],
[
0.11419677734375,
0.28515625,
0.344970703125,
0.283935546875,
0.344970703125,
0.298583984375,
0.11419677734375,
0.2998046875
],
[
0.51123046875,
0.295654296875,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.51123046875,
0.308837890625
],
[
0.11419677734375,
0.314208984375,
0.2626953125,
0.3154296875,
0.2626953125,
0.330078125,
0.11419677734375,
0.328857421875
],
[
0.51123046875,
0.31005859375,
0.880859375,
0.31005859375,
0.880859375,
0.32373046875,
0.51123046875,
0.32373046875
],
[
0.51123046875,
0.327880859375,
0.78515625,
0.327880859375,
0.78515625,
0.341552734375,
0.51123046875,
0.341552734375
],
[
0.1158447265625,
0.341552734375,
0.486083984375,
0.342529296875,
0.486083984375,
0.35693359375,
0.1158447265625,
0.355712890625
],
[
0.509765625,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.364013671875,
0.509765625,
0.364013671875
],
[
0.1175537109375,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.11419677734375,
0.37353515625,
0.490234375,
0.372314453125,
0.490234375,
0.39013671875,
0.11419677734375,
0.391357421875
],
[
0.5087890625,
0.3681640625,
0.861328125,
0.366455078125,
0.861328125,
0.38134765625,
0.5087890625,
0.383056640625
],
[
0.1175537109375,
0.391357421875,
0.489501953125,
0.391357421875,
0.489501953125,
0.40625,
0.1175537109375,
0.40625
],
[
0.509765625,
0.392578125,
0.8818359375,
0.392578125,
0.8818359375,
0.40673828125,
0.509765625,
0.40673828125
],
[
0.1158447265625,
0.40625,
0.48779296875,
0.40673828125,
0.48779296875,
0.421630859375,
0.1158447265625,
0.42041015625
],
[
0.51123046875,
0.410400390625,
0.88330078125,
0.410400390625,
0.88330078125,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.1158447265625,
0.42333984375,
0.48779296875,
0.42333984375,
0.48779296875,
0.438232421875,
0.1158447265625,
0.438232421875
],
[
0.509765625,
0.42578125,
0.884765625,
0.42578125,
0.884765625,
0.440673828125,
0.509765625,
0.440673828125
],
[
0.1175537109375,
0.439453125,
0.489501953125,
0.439453125,
0.489501953125,
0.453125,
0.1175537109375,
0.453125
],
[
0.51123046875,
0.440673828125,
0.75390625,
0.441162109375,
0.75390625,
0.4560546875,
0.51123046875,
0.454833984375
],
[
0.1158447265625,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.4697265625,
0.1158447265625,
0.4697265625
],
[
0.1158447265625,
0.471435546875,
0.489501953125,
0.471435546875,
0.489501953125,
0.485107421875,
0.1158447265625,
0.485107421875
],
[
0.5400390625,
0.471435546875,
0.85205078125,
0.471435546875,
0.85205078125,
0.485107421875,
0.5400390625,
0.485107421875
],
[
0.1158447265625,
0.488037109375,
0.48779296875,
0.488037109375,
0.48779296875,
0.501953125,
0.1158447265625,
0.501953125
],
[
0.1158447265625,
0.50439453125,
0.48779296875,
0.50439453125,
0.48779296875,
0.51708984375,
0.1158447265625,
0.51708984375
],
[
0.5087890625,
0.49951171875,
0.87890625,
0.49951171875,
0.87890625,
0.51416015625,
0.5087890625,
0.51416015625
],
[
0.1158447265625,
0.52099609375,
0.48779296875,
0.52099609375,
0.48779296875,
0.53369140625,
0.1158447265625,
0.53369140625
],
[
0.5087890625,
0.5166015625,
0.880859375,
0.5166015625,
0.880859375,
0.52978515625,
0.5087890625,
0.52978515625
],
[
0.1158447265625,
0.53515625,
0.486083984375,
0.53515625,
0.486083984375,
0.548828125,
0.1158447265625,
0.548828125
],
[
0.51318359375,
0.533203125,
0.8818359375,
0.533203125,
0.8818359375,
0.54638671875,
0.51318359375,
0.54638671875
],
[
0.1158447265625,
0.5517578125,
0.486083984375,
0.55029296875,
0.486083984375,
0.5654296875,
0.1158447265625,
0.56640625
],
[
0.5087890625,
0.54736328125,
0.880859375,
0.548828125,
0.880859375,
0.5634765625,
0.5087890625,
0.5625
],
[
0.1175537109375,
0.568359375,
0.484375,
0.568359375,
0.484375,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.51318359375,
0.56640625,
0.83642578125,
0.56640625,
0.83642578125,
0.57666015625,
0.51318359375,
0.57666015625
],
[
0.1175537109375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.1158447265625,
0.599609375,
0.48779296875,
0.5986328125,
0.48779296875,
0.61328125,
0.1158447265625,
0.6142578125
],
[
0.5556640625,
0.59423828125,
0.8388671875,
0.59423828125,
0.8388671875,
0.60791015625,
0.5556640625,
0.60791015625
],
[
0.1158447265625,
0.615234375,
0.38623046875,
0.6162109375,
0.38623046875,
0.63134765625,
0.1158447265625,
0.6298828125
],
[
0.51318359375,
0.62353515625,
0.8818359375,
0.62353515625,
0.8818359375,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.166259765625,
0.6435546875,
0.489501953125,
0.6435546875,
0.489501953125,
0.65673828125,
0.166259765625,
0.65673828125
],
[
0.51123046875,
0.64013671875,
0.8818359375,
0.64013671875,
0.8818359375,
0.65380859375,
0.51123046875,
0.65380859375
],
[
0.509765625,
0.65673828125,
0.8818359375,
0.65673828125,
0.8818359375,
0.67041015625,
0.509765625,
0.67041015625
],
[
0.118408203125,
0.64453125,
0.172119140625,
0.64453125,
0.172119140625,
0.65625,
0.118408203125,
0.65625
],
[
0.1158447265625,
0.65771484375,
0.48779296875,
0.65771484375,
0.48779296875,
0.6728515625,
0.1158447265625,
0.6728515625
],
[
0.11419677734375,
0.67578125,
0.48779296875,
0.6748046875,
0.48779296875,
0.689453125,
0.11419677734375,
0.6904296875
],
[
0.509765625,
0.6728515625,
0.88330078125,
0.6728515625,
0.88330078125,
0.68603515625,
0.509765625,
0.68603515625
],
[
0.1175537109375,
0.6923828125,
0.489501953125,
0.6923828125,
0.489501953125,
0.7060546875,
0.1175537109375,
0.7060546875
],
[
0.509765625,
0.6884765625,
0.88330078125,
0.6884765625,
0.88330078125,
0.70166015625,
0.509765625,
0.70166015625
],
[
0.1175537109375,
0.70703125,
0.486083984375,
0.70703125,
0.486083984375,
0.72021484375,
0.1175537109375,
0.72021484375
],
[
0.509765625,
0.705078125,
0.88330078125,
0.705078125,
0.88330078125,
0.71875,
0.509765625,
0.71875
],
[
0.1175537109375,
0.7236328125,
0.491943359375,
0.7236328125,
0.491943359375,
0.73828125,
0.1175537109375,
0.73828125
],
[
0.509765625,
0.72021484375,
0.8818359375,
0.72021484375,
0.8818359375,
0.73388671875,
0.509765625,
0.73388671875
],
[
0.1175537109375,
0.74072265625,
0.48779296875,
0.74072265625,
0.48779296875,
0.75439453125,
0.1175537109375,
0.75439453125
],
[
0.509765625,
0.736328125,
0.880859375,
0.7353515625,
0.880859375,
0.75,
0.509765625,
0.75048828125
],
[
0.51904296875,
0.74365234375,
0.88671875,
0.7490234375,
0.88671875,
0.77392578125,
0.51904296875,
0.7685546875
],
[
0.11419677734375,
0.755859375,
0.489501953125,
0.7548828125,
0.489501953125,
0.76953125,
0.11419677734375,
0.77099609375
],
[
0.5205078125,
0.75830078125,
0.52880859375,
0.75830078125,
0.52880859375,
0.7626953125,
0.5205078125,
0.7626953125
],
[
0.1175537109375,
0.7724609375,
0.489501953125,
0.7724609375,
0.489501953125,
0.7861328125,
0.1175537109375,
0.7861328125
],
[
0.51123046875,
0.7685546875,
0.880859375,
0.7685546875,
0.880859375,
0.7822265625,
0.51123046875,
0.7822265625
],
[
0.1158447265625,
0.78759765625,
0.489501953125,
0.7880859375,
0.489501953125,
0.802734375,
0.1158447265625,
0.8017578125
],
[
0.51123046875,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51123046875,
0.798828125
],
[
0.1158447265625,
0.8046875,
0.489501953125,
0.8046875,
0.489501953125,
0.818359375,
0.1158447265625,
0.818359375
],
[
0.51123046875,
0.8017578125,
0.880859375,
0.8017578125,
0.880859375,
0.8154296875,
0.51123046875,
0.8154296875
],
[
0.1158447265625,
0.82080078125,
0.489501953125,
0.82080078125,
0.489501953125,
0.8349609375,
0.1158447265625,
0.8349609375
],
[
0.509765625,
0.81787109375,
0.880859375,
0.81787109375,
0.880859375,
0.83056640625,
0.509765625,
0.83056640625
],
[
0.118408203125,
0.83837890625,
0.486083984375,
0.83837890625,
0.486083984375,
0.8486328125,
0.118408203125,
0.8486328125
],
[
0.509765625,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.8466796875,
0.509765625,
0.8466796875
],
[
0.1158447265625,
0.8525390625,
0.489501953125,
0.8515625,
0.489501953125,
0.8662109375,
0.1158447265625,
0.86767578125
],
[
0.51318359375,
0.849609375,
0.880859375,
0.849609375,
0.880859375,
0.86328125,
0.51318359375,
0.86328125
],
[
0.51318359375,
0.865234375,
0.8818359375,
0.865234375,
0.8818359375,
0.87890625,
0.51318359375,
0.87890625
],
[
0.139404296875,
0.87890625,
0.429931640625,
0.87890625,
0.429931640625,
0.8935546875,
0.139404296875,
0.8935546875
],
[
0.51123046875,
0.8818359375,
0.73876953125,
0.8818359375,
0.73876953125,
0.8955078125,
0.51123046875,
0.8955078125
],
[
0.1376953125,
0.892578125,
0.30908203125,
0.892578125,
0.30908203125,
0.90625,
0.1376953125,
0.90625
],
[
0.1368408203125,
0.90380859375,
0.30908203125,
0.90478515625,
0.30908203125,
0.91943359375,
0.1368408203125,
0.91845703125
],
[
0.5341796875,
0.90478515625,
0.8759765625,
0.90478515625,
0.8759765625,
0.91943359375,
0.5341796875,
0.91943359375
]
] |
[
"UNK",
"JDs are annotated with spans corresponding to",
"Unique Titles",
"Instances",
"123,619",
"12,984",
"specific skills, and these spans have subsequently",
"Train",
"14,641",
"233",
"been manually linked to ESCO, as described",
"Dev.",
"480",
"149 ",
"455",
"813",
"in the work of Decorte et al. (2022). In cases",
"1,824",
"Test",
"where skills cannot be linked. two labels are used.",
"Table l: Data Statistics. Data distribution of train, dev,",
"namely UNDERSPECIFIED and LABEL NOT",
"and test splits. UNK indicates skills mentions that are",
"PRESENT.For the sake of uniformity, we map",
"not linked to a corresponding taxonomy title.",
"both of these labels to a generic UNK tag. We used",
"several heuristics based on Levenshtein distance",
"results of linking with two entity linking systems.",
"and sentence similarity to find the exact subspans",
"if it exceeds certain thresholds, otherwise, it is",
"@ We present a qualitative analysis showing that",
"UNK. This process is outlined in Appendix A.",
"the model successfully links implicit skills to their",
"In addition, some data examples can be found",
"respective skill entry in ESCO.",
"in Appendix B. The number of UNKs in the data",
"2Methodology",
"is also in Table 1. During inference, the UNK title",
"is a prediction option for the models.",
"Definition.In EL, we process the input document",
"Models. We use two EL models, selected for",
"D = {wi,..., wr], a collection of entity men-",
"tions denoted as MD - {m1,..., mn], and a KB,",
"their robust performance in EL on Wikipedia. 6",
"ESCO in our case: & = e1,...,e13890, UNK].",
"BLINK (Wu et al., 2020). BLINK uses a bi-",
"The objective of an EL model is to generate a list",
"encoder architecture based on BERT (Devlin et al.,",
"of mention-entity pairs [(mi, e;)}\"-1, where each",
"2019), for modeling pairs of mentions and entities.",
"entity e corresponds to an entry in a KB. We assume",
"The model processes two inputs:",
"that both the titles and descriptions of the entities",
"are available, which is a common scenario in EL",
"[CLS] ctxti [S]mention [E] ctxtr [SEP]",
"research (Ganea and Hofmann, 2017; Logeswaran",
"et al., 2019; Wu et al., 2020). We also assume that",
"Where“mention\",\"ctxt\", and“\"ctxt.\"corresponds",
"each mention in the document has a corresponding",
"to the wordpiece tokens of the mention, the left",
"valid gold entity present in the knowledge base.",
"context, and the right context. The mention is de-",
"including UNK. This scenario is typically referred",
"noted by special tokens [S] and [E]. The entity",
"to as \"in-KB evaluation\". Similar to prior research",
"anditsdescriptionare.structuredasfollows:",
"efforts (Logeswaran et al., 2019; Wu et al., 2020),",
"we also presuppose that the mentions within the",
"[CLS] title [ENT] description[SEP]",
"document have already been tagged.",
"Here,\"title\"and“description\"represent the word-",
"We use ESCO titles as ground truth labels,",
"piece tokens of the skills' title and description, re-",
"spectively. [ENT] is a special token to separate",
"Data.",
"containing 13,890 skills.3 Table 1 presents the",
"train, dev, and test data in our experiments. We",
"the two representations. We train the model to max-",
"leverage the train set introduced by Decorte et al.",
"imize the dot product of the [cLs] representation",
"(2023)4 along with the dev and test sets provided",
" of the two inputs, for the correct skill in comparison",
"in Decorte et al. (2022).5 The train set is syn-",
" to skills within the same batch. For each training",
"thetically generated by Decorte et al. (2023) with",
"pair (mi, ei), the loss is computed as L (mi,ei) =",
"-s(mi,e) + log j-1 exp(s(mi,e;), where",
"the gpt-3.5-turbo-0301 model (OpenAI,",
"",
"2023). Specifically, this involves taking each skill",
"the objective is to minimize the distance between",
"from ESCO and prompting the model to generate",
"mi and e; while encouraging the model to assign a",
"sentences resembling JD sentences that require that",
"higher score to the correct pair and lower scores to",
"particular skill. The dev and test splits, conversely,",
"randomly sampled incorrect pairs. Hard negatives",
"arederivedfromactualjobadvertisementssourced",
"are also used during training, these are obtained by",
"from the study by Zhang et al. (2022a). These",
"finding the top 10 predicted skills for each training",
"example. These extra hard negatives are added to",
"3Per version 1.1.1, accessed on 01 August 2023.",
"the random in-batch negatives.",
"thttps://t.ly/edqkp",
"shttps://t.1y/LcqQ7",
"6For the hyperparameter setups, we refer to Appendix C."
] |
{
"arxiv_id": "2401.17979",
"language": "en",
"page_idx": 1
}
|
||
2401.17979-en-2
|
2401.17979-en-2.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>$\mathbf{Train}$ $\mathbf{Source}$ </th>
<th>$\mathbf{Acc@1}$</th>
<th>$\mathbf{Acc@4}$</th>
<th>$\mathbf{Acc@8}$</th>
<th>$\operatorname{Acc@16}$</th>
<th>$\mathbf{Acc@32}$</th>
</tr>
<tr>
<td>Random</td>
<td> </td>
<td>$0.22\pm0.00$</td>
<td>$0.88\pm0.00$</td>
<td>$1.76\pm0.00$</td>
<td>$3.52\pm0.00$</td>
<td>$7.04\pm0.00$</td>
</tr>
<tr>
<td>TF- IDF</td>
<td> </td>
<td>$2.25\pm0.00$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>BLINK( bert- base) </td>
<td>$ESCO$</td>
<td>$12.74\pm0.49$</td>
<td>$22.81\pm0.79$</td>
<td>$27.70\pm0.82$</td>
<td>$32.44\pm1.33$</td>
<td>$36.46\pm1.07$</td>
</tr>
<tr>
<td>$\mathbf BLINK$ (bert-large)</td>
<td>$ESCO$</td>
<td>$12.77\pm0.94$</td>
<td>$22.58\pm1.47$</td>
<td>$27.24\pm1.23$</td>
<td>$31.75\pm0.89$</td>
<td>$36.10\pm1.28$</td>
</tr>
<tr>
<td>$\mathbf BLINK$ (bert-large)</td>
<td>Wiki (0-shot)</td>
<td>$23.30\pm0.00$</td>
<td>$32.89\pm0.00$</td>
<td>$38.16\pm0.00$</td>
<td>$42.60\pm0.00$</td>
<td>$45.56\pm0.00$</td>
</tr>
<tr>
<td>BLINK ( bert- large) </td>
<td>Wiki+ESCO</td>
<td>$23.55\pm0.14$</td>
<td>$32.63\pm0.16$</td>
<td>$37.38\pm0.09$</td>
<td>$43.25\pm0.13$</td>
<td>$48.98\pm0.21$</td>
</tr>
<tr>
<td>GENRE (bart-base)</td>
<td>$ESCO$</td>
<td>$1.47\pm0.05$</td>
<td>$4.84\pm1.74$</td>
<td>$10.46\pm6.81$</td>
<td>$11.30\pm4.18$</td>
<td>$15.51\pm4.62$</td>
</tr>
<tr>
<td>GENRE (bart-large)</td>
<td>$ESCO$</td>
<td>$2.33\pm0.44$</td>
<td>$5.74\pm1.43$</td>
<td>$8.18\pm2.21$</td>
<td>$11.13\pm2.42$</td>
<td>$15.26\pm2.66$</td>
</tr>
<tr>
<td>GENRE (bart-large)</td>
<td>Wiki (0-shot)</td>
<td>$6.91\pm0.00$</td>
<td>$12.34\pm0.00$</td>
<td>$15.52\pm0.00$</td>
<td>$21.60\pm0.00$</td>
<td>$33.17\pm0.00$</td>
</tr>
<tr>
<td>GENRE $\operatorname{E}($bart-large)</td>
<td>Wiki+ESCO</td>
<td>$\mathbf{11.48\pm0.41}$</td>
<td>$21.26\pm0.43$</td>
<td>$27.40\pm0.78$</td>
<td>$37.21\pm0.69$</td>
<td>$49.78\pm1.05$</td>
</tr>
</tbody>
</table>
Table 2: Skill Linking Results. We show the results of the various models used. There are two base and four large models. Training sources are either ESCO or a combination of Wikipedia and ESCO. The results are the average and standard deviation over five seeds. For the 0-shot setup, we apply the fine-tuned models from the work of Wu et al. (2020) and Cao et al. (2021) to the ESCO test set once. We have a random and a TF-IDF-based baseline.
ground truth. Here, we use the evaluation metric Accuracy@k, following prior research (Logeswaran et al., 2019; Wu et al., 2020; Zaporojets et al., 2022). We calculate the correctness between mentions and entities in the KB as the sum of cor-
ing an autoregressive approach: $( e\mid x) = p_{\theta }( y\mid$ rect hits or true positives( TP) if the ground truth $x) = \prod _{i= 1}^{N}p_{\theta }( y_{i}\mid y_{< i}, x) $,where $y$ represents the for instance $i$ is in the top- $k$ predictions, formally: set of $N$ tokens in th
GENRE (Cao et al., 2021). GENRE formulates EL as a retrieval problem using a sequence-tosequence model based on BART (Lewis et al., 2020). This model generates textual entity identifers (i.e., skill titles) and ranks each entity $e\in\mathcal{E}$ usset of $N$ tokens in the identifier of entity $e$ (i.e., entity tile), and $\theta$ denotes the model parameters. During decoding, the model uses a constrained beam search to ensure the generation of valid identifiers (i.e., only producing valid titles that exist within the KB, including UNK).
${\mathrm{Accuracy@}}k=\frac{1}{n}\sum_{i=1}^{n}{\mathrm{TP~in~top-}}k{\mathrm{~for~instance~}}i.$
(1)
## 3 Results
Setup. We train a total of six models: for BLINK, these are BERT$_\mathrm{base}$ and BERT$_\mathrm{large}($uncased; Devlin et al., 2019) trained on ESCO, and another large version trained on Wikipedia and ESCO sequentially. GENRE has the same setup, but then with BART (Lewis et al., 2020). Additionally, we apply the released models from both BLINK and GENRE (large, trained on Wikipedia) in a zero-shot manner and evaluate their performance. The reason we use Wikipedia-based models is that we hypothe- models in this context are the BERT$_\mathrm{large}$ and
size this is due to many skills in ESCO also having corresponding Wikipedia pages (e.g., Python$^{7}$ or teamwork$^{8})$, thus could potentially help linking. Next, to address unknown entities (UNK), we include them as possible label outputs.
For evaluation, we assess the accuracy of generated mention-entity pairs in comparison to the
Table 2 presents the results. Each model is trained for five seeds, and we report the average and standard deviation. We make use of a random and TF-IDF-based baseline.
Firstly, we observe that the strict linking performance (i.e., Acc@ 1) is rather modest for both BLINK and GENRE. But most models outperform the baselines. Notably, the top-performing BART$_{\mathrm{large~models,~which~were~further~fine-tuned}}$ from Wikipedia EL with ESCO. As expected, scores improve considerably as we increase the value of $k.$ Secondly, for both BLINK and GENRE, model size seems not to have a substantial impact when trained only on ESCO. Specifically for BLINK, the performance remains consistent for Acc@1 and exhibits only a slight decline as we relax the number of candidates for performance evaluation. For GENRE, the observed trend remains largely unchanged, even with a larger $k.$
$^{7}$https://en.wikipedia.org/wiki/Python\_
(programming\_language)
$^{8}$https://en.wikipedia.org/wiki/
Teamwork
|
[
[
0.450927734375,
0.0872802734375,
0.50732421875,
0.0872802734375,
0.50732421875,
0.1015625,
0.450927734375,
0.1015625
],
[
0.54248046875,
0.0872802734375,
0.599609375,
0.0872802734375,
0.599609375,
0.1015625,
0.54248046875,
0.1015625
],
[
0.6337890625,
0.0872802734375,
0.69189453125,
0.0872802734375,
0.69189453125,
0.1015625,
0.6337890625,
0.1015625
],
[
0.71728515625,
0.0872802734375,
0.7841796875,
0.0872802734375,
0.7841796875,
0.1015625,
0.71728515625,
0.1015625
],
[
0.8076171875,
0.0872802734375,
0.875,
0.0872802734375,
0.875,
0.1015625,
0.8076171875,
0.1015625
],
[
0.124267578125,
0.10809326171875,
0.18896484375,
0.10809326171875,
0.18896484375,
0.122314453125,
0.124267578125,
0.122314453125
],
[
0.31982421875,
0.08905029296875,
0.4130859375,
0.08905029296875,
0.4130859375,
0.1009521484375,
0.31982421875,
0.1009521484375
],
[
0.5283203125,
0.10809326171875,
0.599609375,
0.10809326171875,
0.599609375,
0.122314453125,
0.5283203125,
0.122314453125
],
[
0.62109375,
0.10809326171875,
0.69189453125,
0.10809326171875,
0.69189453125,
0.122314453125,
0.62109375,
0.122314453125
],
[
0.7109375,
0.10809326171875,
0.7841796875,
0.10809326171875,
0.7841796875,
0.122314453125,
0.7109375,
0.122314453125
],
[
0.8037109375,
0.10809326171875,
0.875,
0.10809326171875,
0.875,
0.122314453125,
0.8037109375,
0.122314453125
],
[
0.124267578125,
0.122314453125,
0.183837890625,
0.122314453125,
0.183837890625,
0.13720703125,
0.124267578125,
0.13720703125
],
[
0.438232421875,
0.1092529296875,
0.50732421875,
0.1092529296875,
0.50732421875,
0.12054443359375,
0.438232421875,
0.12054443359375
],
[
0.4375,
0.12353515625,
0.509765625,
0.12353515625,
0.509765625,
0.1383056640625,
0.4375,
0.1383056640625
],
[
0.51904296875,
0.142578125,
0.599609375,
0.144287109375,
0.599609375,
0.1591796875,
0.51904296875,
0.15673828125
],
[
0.7021484375,
0.142578125,
0.7841796875,
0.144287109375,
0.7841796875,
0.1591796875,
0.7021484375,
0.15673828125
],
[
0.429931640625,
0.1436767578125,
0.5087890625,
0.1436767578125,
0.5087890625,
0.157958984375,
0.429931640625,
0.157958984375
],
[
0.366943359375,
0.1436767578125,
0.41650390625,
0.1436767578125,
0.41650390625,
0.157958984375,
0.366943359375,
0.157958984375
],
[
0.611328125,
0.1436767578125,
0.69189453125,
0.1436767578125,
0.69189453125,
0.157958984375,
0.611328125,
0.157958984375
],
[
0.7958984375,
0.1455078125,
0.873046875,
0.1455078125,
0.873046875,
0.15673828125,
0.7958984375,
0.15673828125
],
[
0.366943359375,
0.1561279296875,
0.414794921875,
0.157958984375,
0.4130859375,
0.17333984375,
0.365234375,
0.171630859375
],
[
0.124267578125,
0.144287109375,
0.287109375,
0.144287109375,
0.287109375,
0.157958984375,
0.124267578125,
0.157958984375
],
[
0.42822265625,
0.15673828125,
0.5087890625,
0.1591796875,
0.5087890625,
0.17333984375,
0.42822265625,
0.171630859375
],
[
0.51904296875,
0.15673828125,
0.599609375,
0.1591796875,
0.599609375,
0.17333984375,
0.51904296875,
0.171630859375
],
[
0.79443359375,
0.1591796875,
0.875,
0.1591796875,
0.875,
0.17333984375,
0.79443359375,
0.17333984375
],
[
0.61279296875,
0.1602783203125,
0.6884765625,
0.1602783203125,
0.6884765625,
0.171630859375,
0.61279296875,
0.171630859375
],
[
0.1259765625,
0.1602783203125,
0.297119140625,
0.1602783203125,
0.297119140625,
0.17333984375,
0.1259765625,
0.17333984375
],
[
0.705078125,
0.1602783203125,
0.78271484375,
0.1602783203125,
0.78271484375,
0.171630859375,
0.705078125,
0.171630859375
],
[
0.318115234375,
0.171630859375,
0.414794921875,
0.17333984375,
0.414794921875,
0.188232421875,
0.318115234375,
0.1859130859375
],
[
0.426513671875,
0.171630859375,
0.5087890625,
0.17333984375,
0.5087890625,
0.188232421875,
0.426513671875,
0.1859130859375
],
[
0.51904296875,
0.171630859375,
0.60107421875,
0.17333984375,
0.60107421875,
0.188232421875,
0.51904296875,
0.1859130859375
],
[
0.791015625,
0.1728515625,
0.875,
0.174560546875,
0.875,
0.189453125,
0.791015625,
0.18701171875
],
[
0.61279296875,
0.174560546875,
0.68994140625,
0.174560546875,
0.68994140625,
0.1859130859375,
0.61279296875,
0.1859130859375
],
[
0.124267578125,
0.1728515625,
0.297119140625,
0.17333984375,
0.297119140625,
0.188232421875,
0.124267578125,
0.18701171875
],
[
0.705078125,
0.174560546875,
0.7841796875,
0.174560546875,
0.7841796875,
0.1859130859375,
0.705078125,
0.1859130859375
],
[
0.1234130859375,
0.188232421875,
0.297119140625,
0.189453125,
0.297119140625,
0.2037353515625,
0.1234130859375,
0.203125
],
[
0.316650390625,
0.188232421875,
0.414794921875,
0.188232421875,
0.414794921875,
0.203125,
0.316650390625,
0.203125
],
[
0.42822265625,
0.188232421875,
0.5087890625,
0.188232421875,
0.5087890625,
0.203125,
0.42822265625,
0.203125
],
[
0.5205078125,
0.188232421875,
0.60107421875,
0.188232421875,
0.60107421875,
0.203125,
0.5205078125,
0.203125
],
[
0.611328125,
0.188232421875,
0.69189453125,
0.188232421875,
0.69189453125,
0.203125,
0.611328125,
0.203125
],
[
0.705078125,
0.190673828125,
0.7841796875,
0.190673828125,
0.7841796875,
0.201904296875,
0.705078125,
0.201904296875
],
[
0.79443359375,
0.189453125,
0.873046875,
0.189453125,
0.873046875,
0.2037353515625,
0.79443359375,
0.2037353515625
],
[
0.12176513671875,
0.2078857421875,
0.29296875,
0.208984375,
0.29296875,
0.223876953125,
0.12176513671875,
0.22265625
],
[
0.365234375,
0.2060546875,
0.41650390625,
0.2078857421875,
0.414794921875,
0.223876953125,
0.364501953125,
0.221435546875
],
[
0.4375,
0.208984375,
0.50732421875,
0.208984375,
0.50732421875,
0.223876953125,
0.4375,
0.223876953125
],
[
0.5263671875,
0.208984375,
0.60107421875,
0.208984375,
0.60107421875,
0.223876953125,
0.5263671875,
0.223876953125
],
[
0.611328125,
0.208984375,
0.68994140625,
0.208984375,
0.68994140625,
0.223876953125,
0.611328125,
0.223876953125
],
[
0.70361328125,
0.208984375,
0.7841796875,
0.208984375,
0.7841796875,
0.223876953125,
0.70361328125,
0.223876953125
],
[
0.79443359375,
0.208984375,
0.8759765625,
0.208984375,
0.8759765625,
0.223876953125,
0.79443359375,
0.223876953125
],
[
0.1234130859375,
0.223876953125,
0.301513671875,
0.2244873046875,
0.301513671875,
0.2392578125,
0.1234130859375,
0.2381591796875
],
[
0.365234375,
0.223876953125,
0.41650390625,
0.223876953125,
0.41650390625,
0.2381591796875,
0.365234375,
0.2381591796875
],
[
0.4375,
0.2257080078125,
0.50732421875,
0.2257080078125,
0.50732421875,
0.2369384765625,
0.4375,
0.2369384765625
],
[
0.52880859375,
0.2257080078125,
0.59765625,
0.2257080078125,
0.59765625,
0.2369384765625,
0.52880859375,
0.2369384765625
],
[
0.62060546875,
0.2257080078125,
0.68701171875,
0.2257080078125,
0.68701171875,
0.2369384765625,
0.62060546875,
0.2369384765625
],
[
0.705078125,
0.2257080078125,
0.78271484375,
0.2257080078125,
0.78271484375,
0.2369384765625,
0.705078125,
0.2369384765625
],
[
0.796875,
0.2257080078125,
0.873046875,
0.2257080078125,
0.873046875,
0.2369384765625,
0.796875,
0.2369384765625
],
[
0.1234130859375,
0.2381591796875,
0.301513671875,
0.2392578125,
0.301513671875,
0.254150390625,
0.1234130859375,
0.2529296875
],
[
0.61279296875,
0.2392578125,
0.69189453125,
0.2392578125,
0.69189453125,
0.254150390625,
0.61279296875,
0.254150390625
],
[
0.70361328125,
0.2392578125,
0.78515625,
0.2392578125,
0.78515625,
0.254150390625,
0.70361328125,
0.254150390625
],
[
0.79248046875,
0.2392578125,
0.875,
0.2392578125,
0.875,
0.254150390625,
0.79248046875,
0.254150390625
],
[
0.4375,
0.240478515625,
0.5087890625,
0.240478515625,
0.5087890625,
0.251708984375,
0.4375,
0.251708984375
],
[
0.32080078125,
0.240478515625,
0.4130859375,
0.240478515625,
0.4130859375,
0.251708984375,
0.32080078125,
0.251708984375
],
[
0.5224609375,
0.240478515625,
0.599609375,
0.240478515625,
0.599609375,
0.251708984375,
0.5224609375,
0.251708984375
],
[
0.124267578125,
0.254638671875,
0.30322265625,
0.254638671875,
0.30322265625,
0.268310546875,
0.124267578125,
0.268310546875
],
[
0.318115234375,
0.254638671875,
0.4130859375,
0.254638671875,
0.4130859375,
0.2666015625,
0.318115234375,
0.2666015625
],
[
0.42822265625,
0.254638671875,
0.50390625,
0.254638671875,
0.50390625,
0.2666015625,
0.42822265625,
0.2666015625
],
[
0.5224609375,
0.254638671875,
0.599609375,
0.254638671875,
0.599609375,
0.2666015625,
0.5224609375,
0.2666015625
],
[
0.61279296875,
0.254638671875,
0.68994140625,
0.254638671875,
0.68994140625,
0.2666015625,
0.61279296875,
0.2666015625
],
[
0.705078125,
0.254638671875,
0.7841796875,
0.254638671875,
0.7841796875,
0.2666015625,
0.705078125,
0.2666015625
],
[
0.7958984375,
0.254638671875,
0.873046875,
0.254638671875,
0.873046875,
0.2666015625,
0.7958984375,
0.2666015625
],
[
0.1158447265625,
0.283935546875,
0.88330078125,
0.283935546875,
0.88330078125,
0.297607421875,
0.1158447265625,
0.297607421875
],
[
0.1175537109375,
0.298583984375,
0.8818359375,
0.298583984375,
0.8818359375,
0.312255859375,
0.1175537109375,
0.312255859375
],
[
0.1158447265625,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.32666015625,
0.1158447265625,
0.32666015625
],
[
0.1158447265625,
0.32666015625,
0.880859375,
0.32666015625,
0.880859375,
0.34033203125,
0.1158447265625,
0.34033203125
],
[
0.1175537109375,
0.365234375,
0.48779296875,
0.365234375,
0.48779296875,
0.37890625,
0.1175537109375,
0.37890625
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.1175537109375,
0.383056640625,
0.495361328125,
0.383056640625,
0.495361328125,
0.396728515625,
0.1175537109375,
0.396728515625
],
[
0.5087890625,
0.3818359375,
0.88330078125,
0.383056640625,
0.88330078125,
0.39794921875,
0.5087890625,
0.396728515625
],
[
0.11419677734375,
0.3984375,
0.496337890625,
0.39794921875,
0.496337890625,
0.412109375,
0.11419677734375,
0.413330078125
],
[
0.509765625,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.413330078125,
0.509765625,
0.413330078125
],
[
0.1158447265625,
0.4150390625,
0.491943359375,
0.4150390625,
0.491943359375,
0.4287109375,
0.1158447265625,
0.4287109375
],
[
0.509765625,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.497802734375,
0.421630859375,
0.50537109375,
0.421630859375,
0.50537109375,
0.424560546875,
0.497802734375,
0.424560546875
],
[
0.1158447265625,
0.431640625,
0.491943359375,
0.431640625,
0.491943359375,
0.4453125,
0.1158447265625,
0.4453125
],
[
0.509765625,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.509765625,
0.4453125
],
[
0.1158447265625,
0.448974609375,
0.486083984375,
0.448974609375,
0.486083984375,
0.4619140625,
0.1158447265625,
0.4619140625
],
[
0.5087890625,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.5087890625,
0.4619140625
],
[
0.11248779296875,
0.460205078125,
0.491943359375,
0.46142578125,
0.491943359375,
0.482177734375,
0.11248779296875,
0.48095703125
],
[
0.509765625,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.509765625,
0.47802734375
],
[
0.1158447265625,
0.479736328125,
0.489501953125,
0.479736328125,
0.489501953125,
0.493408203125,
0.1158447265625,
0.493408203125
],
[
0.1158447265625,
0.496337890625,
0.489501953125,
0.496337890625,
0.489501953125,
0.51025390625,
0.1158447265625,
0.51025390625
],
[
0.640625,
0.5087890625,
0.6689453125,
0.5087890625,
0.6689453125,
0.52099609375,
0.640625,
0.52099609375
],
[
0.6630859375,
0.50439453125,
0.6767578125,
0.50439453125,
0.6767578125,
0.51318359375,
0.6630859375,
0.51318359375
],
[
0.1158447265625,
0.5126953125,
0.486083984375,
0.5126953125,
0.486083984375,
0.525390625,
0.1158447265625,
0.525390625
],
[
0.5146484375,
0.5166015625,
0.634765625,
0.515625,
0.634765625,
0.52978515625,
0.5146484375,
0.53076171875
],
[
0.67822265625,
0.51318359375,
0.87890625,
0.515625,
0.87890625,
0.533203125,
0.67822265625,
0.53076171875
],
[
0.1158447265625,
0.52783203125,
0.48779296875,
0.52685546875,
0.48779296875,
0.54150390625,
0.1158447265625,
0.5419921875
],
[
0.63623046875,
0.52978515625,
0.6533203125,
0.52392578125,
0.66015625,
0.533203125,
0.64404296875,
0.5390625
],
[
0.6572265625,
0.53369140625,
0.6826171875,
0.53369140625,
0.6826171875,
0.5458984375,
0.6572265625,
0.5458984375
],
[
0.1158447265625,
0.54345703125,
0.486083984375,
0.5419921875,
0.486083984375,
0.55712890625,
0.1158447265625,
0.55810546875
],
[
0.85791015625,
0.54345703125,
0.88330078125,
0.54345703125,
0.88330078125,
0.55712890625,
0.85791015625,
0.55712890625
],
[
0.1158447265625,
0.55859375,
0.298828125,
0.56005859375,
0.298828125,
0.57470703125,
0.1158447265625,
0.57373046875
],
[
0.51904296875,
0.57177734375,
0.52880859375,
0.5791015625,
0.5205078125,
0.58447265625,
0.51123046875,
0.57763671875
],
[
0.5380859375,
0.57177734375,
0.609375,
0.57177734375,
0.609375,
0.5859375,
0.5380859375,
0.5859375
],
[
0.1158447265625,
0.5849609375,
0.483642578125,
0.5849609375,
0.483642578125,
0.5986328125,
0.1158447265625,
0.5986328125
],
[
0.11419677734375,
0.5986328125,
0.489501953125,
0.60107421875,
0.489501953125,
0.6181640625,
0.11419677734375,
0.6162109375
],
[
0.51123046875,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.51123046875,
0.6123046875
],
[
0.1158447265625,
0.61767578125,
0.484375,
0.61767578125,
0.484375,
0.63134765625,
0.1158447265625,
0.63134765625
],
[
0.51123046875,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.51123046875,
0.62890625
],
[
0.1158447265625,
0.6328125,
0.489501953125,
0.63427734375,
0.489501953125,
0.6484375,
0.1158447265625,
0.64794921875
],
[
0.51123046875,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.51123046875,
0.64453125
],
[
0.5087890625,
0.64453125,
0.6884765625,
0.646484375,
0.6884765625,
0.6611328125,
0.5087890625,
0.6591796875
],
[
0.1175537109375,
0.65087890625,
0.48779296875,
0.65087890625,
0.48779296875,
0.6650390625,
0.1175537109375,
0.6650390625
],
[
0.1158447265625,
0.6650390625,
0.486083984375,
0.66650390625,
0.486083984375,
0.68115234375,
0.1158447265625,
0.6796875
],
[
0.5283203125,
0.662109375,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.5283203125,
0.6767578125
],
[
0.1175537109375,
0.68310546875,
0.489501953125,
0.68310546875,
0.489501953125,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.51318359375,
0.68115234375,
0.880859375,
0.68115234375,
0.880859375,
0.69140625,
0.51318359375,
0.69140625
],
[
0.1175537109375,
0.69970703125,
0.491943359375,
0.69970703125,
0.491943359375,
0.71337890625,
0.1175537109375,
0.71337890625
],
[
0.51123046875,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.1175537109375,
0.71533203125,
0.489501953125,
0.71533203125,
0.489501953125,
0.728515625,
0.1175537109375,
0.728515625
],
[
0.51123046875,
0.71044921875,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.51123046875,
0.724609375
],
[
0.50732421875,
0.7236328125,
0.88330078125,
0.7255859375,
0.88330078125,
0.74365234375,
0.50732421875,
0.7412109375
],
[
0.1175537109375,
0.732421875,
0.491943359375,
0.732421875,
0.491943359375,
0.7451171875,
0.1175537109375,
0.7451171875
],
[
0.1158447265625,
0.74658203125,
0.48779296875,
0.74658203125,
0.48779296875,
0.76123046875,
0.1158447265625,
0.76123046875
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75830078125,
0.51123046875,
0.75830078125
],
[
0.1158447265625,
0.76318359375,
0.490234375,
0.76318359375,
0.490234375,
0.77783203125,
0.1158447265625,
0.77783203125
],
[
0.509765625,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.509765625,
0.77392578125
],
[
0.1158447265625,
0.77783203125,
0.489501953125,
0.779296875,
0.489501953125,
0.79345703125,
0.1158447265625,
0.7919921875
],
[
0.509765625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.509765625,
0.7890625
],
[
0.1175537109375,
0.7958984375,
0.491943359375,
0.7958984375,
0.491943359375,
0.80859375,
0.1175537109375,
0.80859375
],
[
0.509765625,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8046875,
0.509765625,
0.8046875
],
[
0.1175537109375,
0.81103515625,
0.3896484375,
0.81103515625,
0.3896484375,
0.826171875,
0.1175537109375,
0.826171875
],
[
0.51123046875,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.1351318359375,
0.82666015625,
0.490234375,
0.82763671875,
0.490234375,
0.8427734375,
0.1351318359375,
0.84130859375
],
[
0.509765625,
0.82373046875,
0.88330078125,
0.822265625,
0.88330078125,
0.83740234375,
0.509765625,
0.83837890625
],
[
0.1175537109375,
0.84423828125,
0.489501953125,
0.84423828125,
0.489501953125,
0.85791015625,
0.1175537109375,
0.85791015625
],
[
0.51123046875,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.51123046875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1368408203125,
0.8662109375,
0.480224609375,
0.8681640625,
0.480224609375,
0.8828125,
0.1368408203125,
0.88134765625
],
[
0.509765625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.118408203125,
0.88134765625,
0.316650390625,
0.88134765625,
0.316650390625,
0.89453125,
0.118408203125,
0.89453125
],
[
0.14111328125,
0.8935546875,
0.414794921875,
0.8935546875,
0.414794921875,
0.90673828125,
0.14111328125,
0.90673828125
],
[
0.51123046875,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90625,
0.193115234375,
0.90625,
0.193115234375,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.8564453125,
0.90478515625,
0.8564453125,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Acc@1",
"Acc@4",
"Acc@8",
"Acc@16",
"Acc@32",
"Random",
"Train Source",
"0.88±0.00",
"1.76±0.00",
"3.52±0.00",
"7.04±0.00",
"TF-IDF",
"0.22±0.00",
"2.25±0.00",
"22.81±0.79",
"32.44±1.33",
"12.74±0.49",
"ESCO",
"27.70±0.82",
"36.46±1.07",
"ESCO",
"BLINK (bert-base)",
"12.77±0.94",
"22.58±1.47",
"36.10±1.28",
"27.24±1.23",
"BLINK (bert-large)",
"31.75±0.89",
"Wiki (O-shot)",
"23.30±0.00",
"32.89±0.00",
"45.56±0.00",
"38.16±0.00",
"BLINK (bert-large)",
"42.60±0.00",
"BLINK (bert-large)",
"Wiki + ESCO",
"23.55±0.14",
"32.63±0.16",
"37.38±0.09",
"43.25±0.13",
"48.98±0.21",
"GENRE (bart-base)",
"ESCO",
"1.47±0.05",
"4.84±1.74",
"10.46±6.81",
"11.30±4.18",
"15.51±4.62",
"GENRE (bart-large)",
"ESCO",
"2.33±0.44",
"5.74±1.43",
"8.18±2.21",
"11.13±2.42",
"15.26±2.66",
"GENRE (bart-large)",
"15.52±0.00",
"21.60±0.00",
"33.17±0.00",
"6.91±0.00",
"Wiki (0-shot)",
"12.34±0.00",
"GENRE (bart-large)",
"Wiki + ESCO",
"11.48±0.41",
"21.26±0.43",
"27.40±0.78",
"37.21±0.69",
"49.78±1.05",
"Table 2: Skill Linking Results. We show the results of the various models used. There are two ba se and four",
"large models. Training sources are either ESCO or a combination of Wikipedia and ESCO. The results are the",
"average and standard deviation over five seeds. For the O-shot setup, we apply the fine-tuned models from the work",
"of Wu et al. (2020) and Cao et al. (2021) to the ESCO test set once. We have a random and a TF-IDF-based baseline.",
"GENRE (Cao et al.. 2021)..GENRE formulates",
"ground truth. Here, we use the evaluation met-",
"EL as a retrieval problem using a sequence-to-",
"ric Accuracy@k, following prior research (Lo-",
"sequence model based on BART (Lewis et al.,",
"geswaran et al., 2019; Wu et al., 2020; Zaporojets",
"2020). This model generates textual entity identi-",
"et al., 2022). We calculate the correctness between",
"",
"fiers (i.e., skill titles) and ranks each entity e e & us-",
"mentions and entities in the KB as the sum of cor-",
"ing an autoregressive approach: s(e I ) = pe(y ",
"rect hits or true positives (TP) if the ground truth",
"a) - II;1 pe (yi | y<i,a), where y represents the",
"for instance i is in the top-k predictions, formally:",
"set of N tokens in the identifier of entity e (i.e., en-",
"tity tile), and denotes the model parameters. Dur-",
"17",
"n",
"ing decoding, the model uses a constrained beam",
"Accuracy@k =",
"TP in top-k for instance i.",
"search to ensure the generation of valid identifiers",
"n",
"i-1",
"(i.e., only producing valid titles that exist within",
"(1)",
"the KB, including UNK).",
"2",
"Results",
"Setup. We train a total of six models: for BLINK",
"these are BERTbase and BERTarge (uncased; Devlin",
"Table 2 presents the results. Each model is trained",
"et al., 2019) trained on ESCO, and another large",
"for five seeds, and we report the average and stan-",
"version trained on Wikipedia and ESCO sequen-",
"dard deviation. We make use of a random and",
"TF-IDF-based baseline.",
"tially. GENRE has the same setup, but then with",
"BART (Lewis et al., 2020). Additionally, we apply",
"Firstly, we observe that the strict linking per-",
"the released models from both BLINK and GENRE",
"formance(i.e.,Acc@1)israthermodestforboth",
"(large, trained on Wikipedia) in a zero-shot man-",
"BLINK and GENRE. But most models outper-",
"ner and evaluate their performance. The reason we",
"form the baselines. Notably, the top-performing",
"models in this context are the BERTiarge and",
"use Wikipedia-based models is that we hypothe-",
"size this is due to many skills in ESCO also having",
"BARTiarge models, which were further fine-tuned",
"corresponding Wikipedia pages (e.g., Python7 or",
"from Wikipedia EL with ESCO. As expected,",
"teamwork8), thus could potentially help linking.",
"scores improve considerably as we increase the",
"Next, to address unknown entities (UNK), we in-",
"value of k. Secondly, for both BLINK and GENRE,",
"clude them as possible label outputs.",
"model size seems not to have a substantial im-",
"For evaluation, we assess the accuracy of gen-",
"pact when trained only on ESCO. Specifically for",
"erated mention-entity pairs in comparison to the",
"BLINK, the performance remains consistent for",
"Acc@1 and exhibits only a slight decline as we",
"7https://en.wikipedia.org/wiki/python_",
"relax the number of candidates for performance",
"(programming_language)",
"\"https://en.wikipedia.org/wiki/",
"evaluation. For GENRE, the observed trend re-",
"Teamwork",
"mains largely unchanged, even with a larger k."
] |
{
"arxiv_id": "2401.17979",
"language": "en",
"page_idx": 2
}
|
||
2401.17979-en-3
|
2401.17979-en-3.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Mention}$</th>
<th>$\mathbf{BLINK}$</th>
<th>$\mathbf{GENRE}$</th>
</tr>
<tr>
<td>$\textcircled{1}$ Work in a way that is patient-centred and inclu- sive.</td>
<td>$(\mathbf{K}0913)$ person centred care</td>
<td>WOrk $\textbf{in an}$ organised manner (T)</td>
</tr>
<tr>
<td>You can ride a bike</td>
<td>sell bicycles (S1.6.1)</td>
<td rowspan="1">drive two-wheeled vehicles (S8.2.2)</td>
</tr>
<tr>
<td>$\textcircled{3}$ It is expected that you are a super user of the MS office tools.</td>
<td>use Microsoft Office (S5.6.1)</td>
<td>tools for software configuration management (0613)</td>
</tr>
<tr>
<td>4 Picking and packing</td>
<td>specialised packing carry r Out S for customers (S6.1.3)</td>
<td rowspan="1">perform loading and unloading (S6.2.1) operations</td>
</tr>
<tr>
<td>$\textcircled{5}$ You are expected to be able to further develop your team - both personally and professionally GOLD: manage a team (S4.8.1)</td>
<td>professional manage I personal T development (S1.14.1)</td>
<td>shape organisational teams based On competencies (S4.6.0)</td>
</tr>
<tr>
<td>$\textcircled{6}$ Our games are developed using Unity so we expect all our programmers to have solid knowledge of mobile game development in Unity3D and C#</td>
<td>$\mathbb{C}\#($K0613)</td>
<td>$\mathbb{C}\#($K0613)</td>
</tr>
</tbody>
</table>
Table 3: We show six qualitative examples. The mention is indicated with purple and we show the predictions (k= 1) of BLINK and GENRE. Green predictions mean correct, and red indicates wrong linking with respect to the ground truth. We also show the ESCO ID, indicating the differences in concepts. The results show successful linking of implicit mentions of skills. In example (5), we show how the linked results are still valid while being different concepts. However, evaluation does not count it as a correct hit.
taxonomy titles (e.g.,$\textcircled{1},\textcircled{2},\textcircled{4},\textcircled{6}).$
$\bullet$ In cases of hard skills (3,6), BLINK correctly matches “MS office tools”to“using Microsoft Office", which is not an exact match. Both models predict the explicit mention “C#" correctly to the C# taxonomy title.
· We found that the models predict paraphrased versions of skills that could also be considered correct $(\textcircled{4},\textcircled{5})$, even being entirely different concepts (i.e., different ESCO IDs).
Remarkably, the zero-shot setup performance of both BLINK and GENRE, when trained on Wikipedia, surpasses that of models trained solely on ESCO. For Wikipedia-based evaluation, GENRE usually outperforms BLINK. We notice the opposite in this case. For BLINK, this improvement is approximately 11 accuracy points for $k=1.$ Meanwhile, for GENRE, we observe an increase of roughly 9 accuracy points when trained on both Wikipedia and ESCO. This trend persists for a larger $k$, reaching up to a 12.5 accuracy point improvement for BLINK and a 34 accuracy point improvement for GENRE in the case of Acc@32. Furthermore, we show that further fine-tuning the Wikipedia-trained models on ESCO contributes to an improved EL performance at $k=\{1,16,32\}$ for both models. We confırm our hypothesis that Wikipedia has concepts that are also in ESCO, this gives the model strong prior knowledge of skills.
For UNK-specific results, we refer to Appendix D. Additionally, we show a direct comparison to previous work in Appendix E.
Evaluation Limitation. We qualitatively demonstrate the linking of skills that are implicit and/or valid. Empirically, we observe that the strict linking of skills leads to an underestimation of model performance. We believe this limitation is rooted in evaluation. In train, dev, and test, there is only one correct gold label. We reciprocate the findings by Li et al. (2020), where they found that a large number of predictions are“technically correct" but limitations in Wikipedia-based evaluation falsely penalized their model (i.e., a more or less precise version of the same entity). Especially $\textcircled{5}$ in Table 3 shows this challenge for ESCO, we can consider multiple links to be correct for a mention given a particular context. This highlights the need for appropriate EL evaluation sets, not only for ESCO, but for EL in general.
## 4 Discussion
Qualitative Analysis. We manually inspected a subset of the predictions. We present qualitative examples in Table 3. We found the following trends upon inspection:
$\bullet$ The EL models exhibit success in linking implicit and explicit mentions to their respective
|
[
[
0.1292724609375,
0.0885009765625,
0.18896484375,
0.0885009765625,
0.18896484375,
0.09979248046875,
0.1292724609375,
0.09979248046875
],
[
0.455078125,
0.08612060546875,
0.509765625,
0.08612060546875,
0.509765625,
0.1009521484375,
0.455078125,
0.1009521484375
],
[
0.6591796875,
0.0885009765625,
0.71728515625,
0.0885009765625,
0.71728515625,
0.09979248046875,
0.6591796875,
0.09979248046875
],
[
0.1259765625,
0.10394287109375,
0.441650390625,
0.105712890625,
0.441650390625,
0.12054443359375,
0.1259765625,
0.1181640625
],
[
0.45166015625,
0.105712890625,
0.63037109375,
0.1051025390625,
0.63037109375,
0.119384765625,
0.45166015625,
0.12054443359375
],
[
0.65576171875,
0.1051025390625,
0.84814453125,
0.105712890625,
0.84814453125,
0.12054443359375,
0.65576171875,
0.119384765625
],
[
0.1318359375,
0.12054443359375,
0.156982421875,
0.12054443359375,
0.156982421875,
0.12890625,
0.1318359375,
0.12890625
],
[
0.6591796875,
0.1181640625,
0.6826171875,
0.1181640625,
0.6826171875,
0.1318359375,
0.6591796875,
0.1318359375
],
[
0.1275634765625,
0.1353759765625,
0.276123046875,
0.13720703125,
0.276123046875,
0.1519775390625,
0.1275634765625,
0.149658203125
],
[
0.456787109375,
0.1383056640625,
0.5859375,
0.1383056640625,
0.5859375,
0.1519775390625,
0.456787109375,
0.1519775390625
],
[
0.6591796875,
0.13720703125,
0.84814453125,
0.13720703125,
0.84814453125,
0.15087890625,
0.6591796875,
0.15087890625
],
[
0.6591796875,
0.15087890625,
0.7109375,
0.15087890625,
0.7109375,
0.162109375,
0.6591796875,
0.162109375
],
[
0.1259765625,
0.168701171875,
0.441650390625,
0.168701171875,
0.441650390625,
0.1817626953125,
0.1259765625,
0.1817626953125
],
[
0.45166015625,
0.168701171875,
0.63623046875,
0.168701171875,
0.63623046875,
0.1817626953125,
0.45166015625,
0.1817626953125
],
[
0.6591796875,
0.168701171875,
0.84814453125,
0.168701171875,
0.84814453125,
0.1817626953125,
0.6591796875,
0.1817626953125
],
[
0.130126953125,
0.1817626953125,
0.2032470703125,
0.1817626953125,
0.2032470703125,
0.193603515625,
0.130126953125,
0.193603515625
],
[
0.6591796875,
0.18115234375,
0.7841796875,
0.18115234375,
0.7841796875,
0.19482421875,
0.6591796875,
0.19482421875
],
[
0.1275634765625,
0.198974609375,
0.287109375,
0.1995849609375,
0.287109375,
0.21435546875,
0.1275634765625,
0.213134765625
],
[
0.455078125,
0.20068359375,
0.64404296875,
0.20068359375,
0.64404296875,
0.21435546875,
0.455078125,
0.21435546875
],
[
0.65576171875,
0.20068359375,
0.84619140625,
0.20068359375,
0.84619140625,
0.21435546875,
0.65576171875,
0.21435546875
],
[
0.455078125,
0.21435546875,
0.591796875,
0.21435546875,
0.591796875,
0.2244873046875,
0.455078125,
0.2244873046875
],
[
0.65576171875,
0.2120361328125,
0.779296875,
0.2120361328125,
0.779296875,
0.2257080078125,
0.65576171875,
0.2257080078125
],
[
0.1259765625,
0.2298583984375,
0.43994140625,
0.23095703125,
0.43994140625,
0.2452392578125,
0.1259765625,
0.24462890625
],
[
0.6572265625,
0.2327880859375,
0.7021484375,
0.2327880859375,
0.7021484375,
0.24462890625,
0.6572265625,
0.24462890625
],
[
0.69775390625,
0.2327880859375,
0.8017578125,
0.2327880859375,
0.8017578125,
0.24462890625,
0.69775390625,
0.24462890625
],
[
0.79833984375,
0.2340087890625,
0.8447265625,
0.2340087890625,
0.8447265625,
0.24462890625,
0.79833984375,
0.24462890625
],
[
0.1292724609375,
0.243408203125,
0.443359375,
0.243408203125,
0.443359375,
0.257080078125,
0.1292724609375,
0.257080078125
],
[
0.455078125,
0.2340087890625,
0.64208984375,
0.2340087890625,
0.64208984375,
0.24462890625,
0.455078125,
0.24462890625
],
[
0.453369140625,
0.24462890625,
0.59619140625,
0.243408203125,
0.59619140625,
0.254638671875,
0.453369140625,
0.255859375
],
[
0.6572265625,
0.24462890625,
0.69873046875,
0.24462890625,
0.69873046875,
0.255859375,
0.6572265625,
0.255859375
],
[
0.7119140625,
0.2464599609375,
0.73876953125,
0.2464599609375,
0.73876953125,
0.254638671875,
0.7119140625,
0.254638671875
],
[
0.75732421875,
0.2452392578125,
0.8447265625,
0.2452392578125,
0.8447265625,
0.255859375,
0.75732421875,
0.255859375
],
[
0.1292724609375,
0.254150390625,
0.330078125,
0.254638671875,
0.330078125,
0.268310546875,
0.1292724609375,
0.267333984375
],
[
0.6572265625,
0.255859375,
0.7109375,
0.255859375,
0.7109375,
0.267333984375,
0.6572265625,
0.267333984375
],
[
0.1275634765625,
0.2724609375,
0.441650390625,
0.27490234375,
0.441650390625,
0.2880859375,
0.1275634765625,
0.2861328125
],
[
0.455078125,
0.275634765625,
0.53076171875,
0.275634765625,
0.53076171875,
0.287353515625,
0.455078125,
0.287353515625
],
[
0.6591796875,
0.275634765625,
0.73486328125,
0.275634765625,
0.73486328125,
0.287353515625,
0.6591796875,
0.287353515625
],
[
0.1292724609375,
0.2880859375,
0.43994140625,
0.2880859375,
0.43994140625,
0.298583984375,
0.1292724609375,
0.298583984375
],
[
0.1292724609375,
0.298583984375,
0.432373046875,
0.298583984375,
0.432373046875,
0.312255859375,
0.1292724609375,
0.312255859375
],
[
0.11419677734375,
0.32470703125,
0.8818359375,
0.325927734375,
0.8818359375,
0.34033203125,
0.11419677734375,
0.339111328125
],
[
0.1158447265625,
0.339111328125,
0.8818359375,
0.34033203125,
0.8818359375,
0.355224609375,
0.1158447265625,
0.35400390625
],
[
0.1158447265625,
0.35400390625,
0.880859375,
0.35400390625,
0.880859375,
0.36767578125,
0.1158447265625,
0.36767578125
],
[
0.1158447265625,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.1158447265625,
0.3818359375
],
[
0.1175537109375,
0.38427734375,
0.60791015625,
0.38427734375,
0.60791015625,
0.39794921875,
0.1175537109375,
0.39794921875
],
[
0.133544921875,
0.4228515625,
0.48779296875,
0.42333984375,
0.48779296875,
0.438232421875,
0.133544921875,
0.43701171875
],
[
0.5458984375,
0.421630859375,
0.82373046875,
0.42041015625,
0.82373046875,
0.438232421875,
0.5458984375,
0.439453125
],
[
0.1158447265625,
0.440673828125,
0.48779296875,
0.440673828125,
0.48779296875,
0.45361328125,
0.1158447265625,
0.45361328125
],
[
0.11419677734375,
0.454833984375,
0.48779296875,
0.4560546875,
0.48779296875,
0.47021484375,
0.11419677734375,
0.4697265625
],
[
0.5322265625,
0.454833984375,
0.8818359375,
0.454833984375,
0.8818359375,
0.468505859375,
0.5322265625,
0.468505859375
],
[
0.11419677734375,
0.47021484375,
0.489501953125,
0.471435546875,
0.489501953125,
0.486328125,
0.11419677734375,
0.485107421875
],
[
0.54736328125,
0.471435546875,
0.88330078125,
0.471435546875,
0.88330078125,
0.485107421875,
0.54736328125,
0.485107421875
],
[
0.1175537109375,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.1175537109375,
0.501953125
],
[
0.54833984375,
0.48681640625,
0.88330078125,
0.48681640625,
0.88330078125,
0.501953125,
0.54833984375,
0.501953125
],
[
0.11419677734375,
0.50439453125,
0.489501953125,
0.5029296875,
0.489501953125,
0.51708984375,
0.11419677734375,
0.5185546875
],
[
0.54833984375,
0.5029296875,
0.884765625,
0.501953125,
0.884765625,
0.5166015625,
0.54833984375,
0.51708984375
],
[
0.1158447265625,
0.52099609375,
0.489501953125,
0.52099609375,
0.489501953125,
0.53369140625,
0.1158447265625,
0.53369140625
],
[
0.54833984375,
0.51953125,
0.8037109375,
0.51953125,
0.8037109375,
0.53369140625,
0.54833984375,
0.53369140625
],
[
0.1175537109375,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.1175537109375,
0.5498046875
],
[
0.1158447265625,
0.552734375,
0.486083984375,
0.552734375,
0.486083984375,
0.56640625,
0.1158447265625,
0.56640625
],
[
0.5322265625,
0.55029296875,
0.87890625,
0.55029296875,
0.87890625,
0.56396484375,
0.5322265625,
0.56396484375
],
[
0.1175537109375,
0.568359375,
0.486083984375,
0.568359375,
0.486083984375,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.54833984375,
0.5693359375,
0.880859375,
0.5693359375,
0.880859375,
0.580078125,
0.54833984375,
0.580078125
],
[
0.1175537109375,
0.58447265625,
0.486083984375,
0.5849609375,
0.486083984375,
0.599609375,
0.1175537109375,
0.5986328125
],
[
0.54833984375,
0.58447265625,
0.8818359375,
0.58447265625,
0.8818359375,
0.59716796875,
0.54833984375,
0.59716796875
],
[
0.1175537109375,
0.60107421875,
0.486083984375,
0.60107421875,
0.486083984375,
0.615234375,
0.1175537109375,
0.615234375
],
[
0.54833984375,
0.599609375,
0.81201171875,
0.599609375,
0.81201171875,
0.6142578125,
0.54833984375,
0.6142578125
],
[
0.1175537109375,
0.61767578125,
0.486083984375,
0.61767578125,
0.486083984375,
0.63134765625,
0.1175537109375,
0.63134765625
],
[
0.118408203125,
0.63525390625,
0.484375,
0.63525390625,
0.484375,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.51318359375,
0.646484375,
0.7060546875,
0.646484375,
0.7060546875,
0.66015625,
0.51318359375,
0.66015625
],
[
0.6962890625,
0.6455078125,
0.88330078125,
0.646484375,
0.88330078125,
0.6611328125,
0.6962890625,
0.66015625
],
[
0.509765625,
0.6630859375,
0.88330078125,
0.6630859375,
0.88330078125,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.1158447265625,
0.64794921875,
0.486083984375,
0.64794921875,
0.486083984375,
0.6611328125,
0.1158447265625,
0.6611328125
],
[
0.11419677734375,
0.66455078125,
0.483642578125,
0.66455078125,
0.483642578125,
0.677734375,
0.11419677734375,
0.677734375
],
[
0.1158447265625,
0.6796875,
0.48779296875,
0.68115234375,
0.48779296875,
0.6953125,
0.1158447265625,
0.69482421875
],
[
0.509765625,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.509765625,
0.693359375
],
[
0.1158447265625,
0.69677734375,
0.489501953125,
0.69677734375,
0.489501953125,
0.71142578125,
0.1158447265625,
0.71142578125
],
[
0.5087890625,
0.6953125,
0.8818359375,
0.69482421875,
0.8818359375,
0.708984375,
0.5087890625,
0.71044921875
],
[
0.1158447265625,
0.71337890625,
0.478515625,
0.7119140625,
0.478515625,
0.72705078125,
0.1158447265625,
0.72802734375
],
[
0.509765625,
0.71142578125,
0.8818359375,
0.71044921875,
0.8818359375,
0.724609375,
0.509765625,
0.7255859375
],
[
0.133544921875,
0.72998046875,
0.491943359375,
0.72998046875,
0.491943359375,
0.74365234375,
0.133544921875,
0.74365234375
],
[
0.5087890625,
0.72705078125,
0.87890625,
0.72705078125,
0.87890625,
0.74072265625,
0.5087890625,
0.74072265625
],
[
0.1158447265625,
0.7451171875,
0.49365234375,
0.7451171875,
0.49365234375,
0.76025390625,
0.1158447265625,
0.76025390625
],
[
0.5087890625,
0.74365234375,
0.8818359375,
0.7421875,
0.8818359375,
0.75732421875,
0.5087890625,
0.75830078125
],
[
0.1158447265625,
0.7626953125,
0.387939453125,
0.7626953125,
0.387939453125,
0.775390625,
0.1158447265625,
0.775390625
],
[
0.51123046875,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.118408203125,
0.79150390625,
0.1318359375,
0.79150390625,
0.1318359375,
0.80029296875,
0.118408203125,
0.80029296875
],
[
0.1427001953125,
0.78759765625,
0.240966796875,
0.7890625,
0.240966796875,
0.80419921875,
0.1427001953125,
0.8017578125
],
[
0.1351318359375,
0.79345703125,
0.1427001953125,
0.79345703125,
0.1427001953125,
0.79736328125,
0.1351318359375,
0.79736328125
],
[
0.51123046875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51123046875,
0.8046875
],
[
0.51318359375,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1175537109375,
0.81396484375,
0.301513671875,
0.81396484375,
0.301513671875,
0.82763671875,
0.1175537109375,
0.82763671875
],
[
0.29296875,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.29296875,
0.82763671875
],
[
0.51123046875,
0.82373046875,
0.880859375,
0.822265625,
0.880859375,
0.83740234375,
0.51123046875,
0.83837890625
],
[
0.1175537109375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.1175537109375,
0.84423828125
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.11419677734375,
0.8466796875,
0.486083984375,
0.8466796875,
0.486083984375,
0.86083984375,
0.11419677734375,
0.86083984375
],
[
0.51123046875,
0.85693359375,
0.88330078125,
0.85693359375,
0.88330078125,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.86328125,
0.2418212890625,
0.86328125,
0.2418212890625,
0.876953125,
0.1158447265625,
0.876953125
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.509765625,
0.88720703125
],
[
0.1376953125,
0.88720703125,
0.49365234375,
0.88818359375,
0.49365234375,
0.90283203125,
0.1376953125,
0.90185546875
],
[
0.51123046875,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.15283203125,
0.90478515625,
0.49365234375,
0.90478515625,
0.49365234375,
0.91845703125,
0.15283203125,
0.91845703125
],
[
0.5087890625,
0.90283203125,
0.72900390625,
0.90380859375,
0.72900390625,
0.91845703125,
0.5087890625,
0.91748046875
]
] |
[
"Mention",
"BLINK",
"GENRE",
"@ Work in a way that is patient-centred and inclu-",
"person centred care (K0913)",
"work in an organised manner",
"sive",
"(T)",
"@ You can ride a bike.",
"sell bicycles (S1.6.1)",
"drive two-wheeled vehicles",
"(S8.2.2)",
" It is expected that you are a super user of the MS",
"use Microsoft Office (S5.6.1)",
"tools for software configuration",
"officetools",
"management (0613)",
"@ Picking and packing.",
"carry out specialised packing",
"perform loading and unloading",
"for customers (S6.1.3)",
"operations (S6.2.1)",
" You are expected to be able to further develop",
"shape",
"organisational",
"teams",
"your team - both personally and professionally.",
"manage personal professional",
"development (S1.14.1)",
" based",
"on",
"competencies",
"GOLD:manage a team (S4.8.1)",
"(S4.6.0)",
"@ Our games are developed using Unity so we",
"C#(K0613)",
"C#(K0613)",
"expectallourprogrammerstohavesolidknowledge",
"of mobile game development in Unity3D and C#.",
"Table 3: We show six qualitative examples. The mention is indicated with purple and we show the predictions",
"(k = 1) of BLINK and GENRE. Green predictions mean correct, and red indicates wrong linking with respect to",
"the ground truth. We also show the ESCO ID, indicating the differences in concepts. The results show successful",
"linking of implicit mentions of skills. In example (5), we show how the linked results are still valid while being",
"different concepts. However, evaluation does not count it as a correct hit.",
"Remarkably,the zero-shot setup performance",
"taxonomy titles (e.g., , @, , @).",
"of both BLINK and GENRE, when trained",
"on Wikipedia, surpasses that of models trained",
": In cases of hard skills (③, @), BLINK cor-",
" solely on ESCO. For Wikipedia-based evaluation,",
"rectly matches\"MS office tools\" to “using Mi-",
"GENRE usually outperforms BLINK. We notice",
"crosoft Office\", which is not an exact match.",
"the opposite in this case.For BLINK, this im-",
"Both models predict the explicit mention \"C#\"",
"provement is approximately 11 accuracy points for",
"correctly to the C# taxonomy title.",
"k = 1. Meanwhile, for GENRE, we observe an",
"increase of roughly 9 accuracy points when trained",
". We found that the models predict paraphrased",
"on both Wikipedia and ESCO. This trend persists",
"versionsofskillsthatcouldalsobeconsidered",
"for a larger k, reaching up to a 12.5 accuracy point",
"correct (@,), even being entirely different",
"improvement for BLINK and a 34 accuracy point",
"concepts (i.e., different ESCO IDs).",
"improvement for GENRE in the case of Acc@32.",
"Furthermore,weshowthatfurtherfine-tuningthe",
"Evaluation Limitation.",
"We qualitatively demon-",
" strate the linking of skills that are implicit and/or",
"Wikipedia-trained models on ESCO contributes to",
"an improved EL performance at k = 1, 16, 32!",
"for both models. We confirm our hypothesis that",
"valid. Empirically, we observe that the strict link-",
"Wikipedia has concepts that are also in ESCO, this",
"ing of skills leads to an underestimation of model",
"gives the model strong prior knowledge of skills.",
"performance. We believe this limitation is rooted",
"For UNK-specific results, we refer to Ap-",
"in evaluation. In train, dev, and test, there is only",
"pendix D. Additionally, we show a direct compari-",
" one correct gold label. We reciprocate the find-",
"son to previous work in Appendix E.",
"ings by Li et al. (2020), where they found that a",
"large number of predictions are“technically cor-",
"4",
"Discussion",
"",
"rect\" but limitations in Wikipedia-based evaluation",
"falsely penalized their model (i.e., a more or less",
"Qualitative Analysis.",
"We manually inspected a",
"precise version of the same entity). Especially ",
"subset of the predictions. We present qualitative",
"in Table 3 shows this challenge for ESCO, we can",
"examples in Table 3. We found the following trends",
"consider multiple links to be correct for a mention",
"upon inspection:",
"given a particular context. This highlights the need",
". The EL models exhibit success in linking im-",
"for appropriate EL evaluation sets, not only for",
"plicit and explicit mentions to their respective",
"ESCO, but for EL in general."
] |
{
"arxiv_id": "2401.17979",
"language": "en",
"page_idx": 3
}
|
||
2401.17979-en-4
|
2401.17979-en-4.jpg
|
## 5 Conclusion
We present entity linking in the job market domain, using two existing high-performing neural models. We demonstrate that the bi-encoder architecture of BLINK is more suited to the job market domain compared to the autoregressive GENRE model. While strict linking results favor BLINK over GENRE, if we relax the number of candidates, we observe that GENRE performs slightly better. From a qualitative perspective, the performance of strict linking results is modest due to limitations in the evaluation set, which considers only one skill correct per mention. However, upon examining the predictions, we identify valid links, suggesting the possibility of multiple correct links for a particular mention, highlighting the need for more comprehensive evaluation. We hope this work sparks interest in entity linking within the job market domain.
# Limitations
In the context of EL for ESCO, our approach has several limitations. Firstly, it only supports English, and might not generalize to other languages. However, several works are working on multilingual entity linking (e.g., Botha et al., 2020; De Cao et al., 2022) and ESCO itself consists of 28 European languages. This work could be extended by supporting it for more languages.
Secondly, our EL model is trained on synthetic training data, which may not fully capture the intricacies and variations present in real-world documents. The use of synthetic data could limit its performance on actual, real JD texts. Nevertheless, we have human-annotated evaluation data.
Moreover, in our evaluation process, we use only one gold-standard ESCO title as the correct answer. This approach may not adequately represent a realworld scenario, where multiple ESCO titles could be correct as shown in Table 3.
In Table 2, we show that providing in-domain data for continuous pre-training shows larger improvements for GENRE than for BLINK. We did not conduct a detailed analysis on the underlying reasons for these positive variations.
## Acknowledgements
We thank the MaiNLP and NLPnorth group for feedback on an earlier version of this paper, and the reviewers for their insightful comments. This research is supported by the Independent Research
Fund Denmark (DFF) grant 9131-00019B and in parts by ERC Consolidator Grant DIALECT 101043235.
# References
Tom Ayoola, Shubhi Tyagi, Joseph Fisher, Christos
Christodoulopoulos, and Andrea Pierleoni. 2022. ReFinED: An efficient zero-shot-capable approach to end-to-end entity linking. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, pages 209- 220,Hybrid: Seattle, Washington+Online. Association for Computational Linguistics.
Krisztian Balog, Yi Fang, Maarten De Rijke, Pavel
Serdyukov, and Luo Si. 2012. Expertise retrieval. Foundations and Trends in Information Retrieval, 6(2-3){:}127{-}256.
Jan A. Botha, Zifei Shan, and Daniel Gillick. 2020. Entity Linking in 100 Languages. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing $(EMNLP)$, pages 7833-7845, Online. Association for Computational Linguistics.
Erik Brynjolfsson and Andrew McAfee. 2011. Race
against the machine: How the digital revolution is accelerating innovation, driving productivity, and irreversibly transforming employment and the econ- $omy.$ Brynjolfsson and McAfee.
Erik Brynjolfsson and Andrew McAfee. 2014. The
second machine age: Work, progress, and prosperity in a time of brilliant technologies. WW Norton & Company.
de Nicola Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021. Autoregressive entity retrieval. In International Conference on Learning Representations.
Nicola de Cao, Ledell Wu, Kashyap Popat, Mikel
Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290.
Benjamin Clavié and Guillaume Soulié. 2023. Large
language models as batteries-included zero-shot esco skills matchers. arXiv preprint arXiv:2307.03539.
Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel
Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290.
Jens-Joris Decorte, Jeroen Van Hautte, Johannes
Deleu, Chris Develder, and Thomas Demeester 2022. Design of negative sampling strategies for
|
[
[
0.11419677734375,
0.0831298828125,
0.246826171875,
0.08489990234375,
0.246826171875,
0.1009521484375,
0.11419677734375,
0.09857177734375
],
[
0.51318359375,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.51123046875,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.1158447265625,
0.11224365234375,
0.489501953125,
0.11224365234375,
0.489501953125,
0.1258544921875,
0.1158447265625,
0.1258544921875
],
[
0.5087890625,
0.11517333984375,
0.60107421875,
0.1175537109375,
0.60107421875,
0.1318359375,
0.5087890625,
0.1300048828125
],
[
0.1175537109375,
0.12890625,
0.489501953125,
0.12890625,
0.489501953125,
0.142578125,
0.1175537109375,
0.142578125
],
[
0.1158447265625,
0.1436767578125,
0.48779296875,
0.1436767578125,
0.48779296875,
0.15673828125,
0.1158447265625,
0.15673828125
],
[
0.1175537109375,
0.160888671875,
0.48779296875,
0.160888671875,
0.48779296875,
0.174560546875,
0.1175537109375,
0.174560546875
],
[
0.51123046875,
0.1602783203125,
0.60791015625,
0.162109375,
0.60791015625,
0.177001953125,
0.51123046875,
0.174560546875
],
[
0.1158447265625,
0.177001953125,
0.48779296875,
0.17578125,
0.48779296875,
0.190673828125,
0.1158447265625,
0.191162109375
],
[
0.509765625,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.1158447265625,
0.193603515625,
0.48779296875,
0.193603515625,
0.48779296875,
0.207275390625,
0.1158447265625,
0.207275390625
],
[
0.52880859375,
0.1953125,
0.88330078125,
0.1953125,
0.88330078125,
0.208984375,
0.52880859375,
0.208984375
],
[
0.1158447265625,
0.208984375,
0.48779296875,
0.208984375,
0.48779296875,
0.22265625,
0.1158447265625,
0.22265625
],
[
0.52880859375,
0.208984375,
0.8818359375,
0.208984375,
0.8818359375,
0.22265625,
0.52880859375,
0.22265625
],
[
0.1175537109375,
0.226806640625,
0.48779296875,
0.226806640625,
0.48779296875,
0.2369384765625,
0.1175537109375,
0.2369384765625
],
[
0.52880859375,
0.22265625,
0.8818359375,
0.22265625,
0.8818359375,
0.236328125,
0.52880859375,
0.236328125
],
[
0.52880859375,
0.235107421875,
0.880859375,
0.235107421875,
0.880859375,
0.248779296875,
0.52880859375,
0.248779296875
],
[
0.1175537109375,
0.2410888671875,
0.48779296875,
0.2410888671875,
0.48779296875,
0.254638671875,
0.1175537109375,
0.254638671875
],
[
0.52880859375,
0.248779296875,
0.8818359375,
0.248779296875,
0.8818359375,
0.262451171875,
0.52880859375,
0.262451171875
],
[
0.1158447265625,
0.25830078125,
0.48779296875,
0.25830078125,
0.48779296875,
0.271484375,
0.1158447265625,
0.271484375
],
[
0.5283203125,
0.26123046875,
0.88330078125,
0.260009765625,
0.88330078125,
0.273681640625,
0.5283203125,
0.27490234375
],
[
0.1158447265625,
0.273681640625,
0.48779296875,
0.273681640625,
0.48779296875,
0.287353515625,
0.1158447265625,
0.287353515625
],
[
0.52880859375,
0.27490234375,
0.88330078125,
0.27490234375,
0.88330078125,
0.2880859375,
0.52880859375,
0.2880859375
],
[
0.1158447265625,
0.290283203125,
0.48779296875,
0.290283203125,
0.48779296875,
0.303955078125,
0.1158447265625,
0.303955078125
],
[
0.53076171875,
0.289306640625,
0.77099609375,
0.289306640625,
0.77099609375,
0.2998046875,
0.53076171875,
0.2998046875
],
[
0.1158447265625,
0.305908203125,
0.48779296875,
0.305908203125,
0.48779296875,
0.319580078125,
0.1158447265625,
0.319580078125
],
[
0.51318359375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.1175537109375,
0.32177734375,
0.48779296875,
0.32177734375,
0.48779296875,
0.336181640625,
0.1175537109375,
0.336181640625
],
[
0.52880859375,
0.32470703125,
0.88330078125,
0.32470703125,
0.88330078125,
0.33837890625,
0.52880859375,
0.33837890625
],
[
0.1158447265625,
0.33837890625,
0.490234375,
0.33837890625,
0.490234375,
0.3515625,
0.1158447265625,
0.3515625
],
[
0.53076171875,
0.339111328125,
0.88330078125,
0.339111328125,
0.88330078125,
0.349853515625,
0.53076171875,
0.349853515625
],
[
0.1158447265625,
0.35400390625,
0.489501953125,
0.35400390625,
0.489501953125,
0.36767578125,
0.1158447265625,
0.36767578125
],
[
0.52880859375,
0.350830078125,
0.64208984375,
0.3515625,
0.64208984375,
0.363525390625,
0.52880859375,
0.3623046875
],
[
0.1158447265625,
0.37060546875,
0.489501953125,
0.37060546875,
0.489501953125,
0.38427734375,
0.1158447265625,
0.38427734375
],
[
0.5087890625,
0.372314453125,
0.8818359375,
0.372314453125,
0.8818359375,
0.385986328125,
0.5087890625,
0.385986328125
],
[
0.52880859375,
0.38720703125,
0.8818359375,
0.38720703125,
0.8818359375,
0.40087890625,
0.52880859375,
0.40087890625
],
[
0.1158447265625,
0.39794921875,
0.2216796875,
0.399658203125,
0.2216796875,
0.41455078125,
0.1158447265625,
0.412109375
],
[
0.5283203125,
0.3984375,
0.8818359375,
0.399658203125,
0.8818359375,
0.41455078125,
0.5283203125,
0.413330078125
],
[
0.5283203125,
0.413330078125,
0.880859375,
0.413330078125,
0.880859375,
0.427001953125,
0.5283203125,
0.427001953125
],
[
0.1158447265625,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.441162109375,
0.1158447265625,
0.441162109375
],
[
0.5283203125,
0.42578125,
0.875,
0.42578125,
0.875,
0.439453125,
0.5283203125,
0.439453125
],
[
0.1175537109375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.1175537109375,
0.457275390625
],
[
0.51318359375,
0.449462890625,
0.8818359375,
0.449462890625,
0.8818359375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.1175537109375,
0.460205078125,
0.489501953125,
0.460205078125,
0.489501953125,
0.473876953125,
0.1175537109375,
0.473876953125
],
[
0.5283203125,
0.463134765625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.5283203125,
0.476806640625
],
[
0.1175537109375,
0.4755859375,
0.489501953125,
0.4755859375,
0.489501953125,
0.4892578125,
0.1175537109375,
0.4892578125
],
[
0.5283203125,
0.4755859375,
0.8818359375,
0.4755859375,
0.8818359375,
0.4892578125,
0.5283203125,
0.4892578125
],
[
0.1175537109375,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.1175537109375,
0.5048828125
],
[
0.5283203125,
0.48681640625,
0.88330078125,
0.488037109375,
0.88330078125,
0.5029296875,
0.5283203125,
0.501953125
],
[
0.1175537109375,
0.50732421875,
0.489501953125,
0.50732421875,
0.489501953125,
0.52099609375,
0.1175537109375,
0.52099609375
],
[
0.52880859375,
0.501953125,
0.75,
0.501953125,
0.75,
0.515625,
0.52880859375,
0.515625
],
[
0.1158447265625,
0.52392578125,
0.48779296875,
0.5224609375,
0.48779296875,
0.53759765625,
0.1158447265625,
0.5380859375
],
[
0.51123046875,
0.52490234375,
0.8818359375,
0.52490234375,
0.8818359375,
0.5380859375,
0.51123046875,
0.5380859375
],
[
0.1158447265625,
0.54052734375,
0.36279296875,
0.54052734375,
0.36279296875,
0.55419921875,
0.1158447265625,
0.55419921875
],
[
0.53076171875,
0.54052734375,
0.880859375,
0.54052734375,
0.880859375,
0.55029296875,
0.53076171875,
0.55029296875
],
[
0.133544921875,
0.55712890625,
0.48779296875,
0.55712890625,
0.48779296875,
0.57080078125,
0.133544921875,
0.57080078125
],
[
0.52880859375,
0.5517578125,
0.88330078125,
0.5517578125,
0.88330078125,
0.5654296875,
0.52880859375,
0.5654296875
],
[
0.1158447265625,
0.57177734375,
0.486083984375,
0.57177734375,
0.486083984375,
0.5849609375,
0.1158447265625,
0.5849609375
],
[
0.53076171875,
0.56640625,
0.599609375,
0.56640625,
0.599609375,
0.57763671875,
0.53076171875,
0.57763671875
],
[
0.1158447265625,
0.5888671875,
0.491943359375,
0.5888671875,
0.491943359375,
0.6025390625,
0.1158447265625,
0.6025390625
],
[
0.51318359375,
0.5888671875,
0.880859375,
0.5888671875,
0.880859375,
0.599609375,
0.51318359375,
0.599609375
],
[
0.1158447265625,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.1158447265625,
0.6181640625
],
[
0.52880859375,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.52880859375,
0.615234375
],
[
0.5283203125,
0.61328125,
0.88330078125,
0.6142578125,
0.88330078125,
0.62890625,
0.5283203125,
0.62744140625
],
[
0.1158447265625,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63525390625,
0.1158447265625,
0.63525390625
],
[
0.52880859375,
0.626953125,
0.5693359375,
0.62890625,
0.568359375,
0.64111328125,
0.5283203125,
0.6396484375
],
[
0.1158447265625,
0.63720703125,
0.4306640625,
0.63720703125,
0.4306640625,
0.65087890625,
0.1158447265625,
0.65087890625
],
[
0.1351318359375,
0.65380859375,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.1351318359375,
0.66748046875
],
[
0.5087890625,
0.6494140625,
0.8818359375,
0.65087890625,
0.8818359375,
0.6650390625,
0.5087890625,
0.66455078125
],
[
0.1175537109375,
0.66943359375,
0.489501953125,
0.66943359375,
0.489501953125,
0.68408203125,
0.1175537109375,
0.68408203125
],
[
0.52880859375,
0.66455078125,
0.8818359375,
0.66455078125,
0.8818359375,
0.677734375,
0.52880859375,
0.677734375
],
[
0.52880859375,
0.677734375,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.52880859375,
0.69140625
],
[
0.1175537109375,
0.68603515625,
0.489501953125,
0.68603515625,
0.489501953125,
0.69970703125,
0.1175537109375,
0.69970703125
],
[
0.52880859375,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.70361328125,
0.52880859375,
0.70361328125
],
[
0.1175537109375,
0.70068359375,
0.486083984375,
0.70068359375,
0.486083984375,
0.71435546875,
0.1175537109375,
0.71435546875
],
[
0.5283203125,
0.703125,
0.88330078125,
0.703125,
0.88330078125,
0.71630859375,
0.5283203125,
0.71630859375
],
[
0.1158447265625,
0.71875,
0.344970703125,
0.71875,
0.344970703125,
0.732421875,
0.1158447265625,
0.732421875
],
[
0.5263671875,
0.71630859375,
0.771484375,
0.71533203125,
0.771484375,
0.72998046875,
0.5263671875,
0.73095703125
],
[
0.1351318359375,
0.7353515625,
0.486083984375,
0.7353515625,
0.486083984375,
0.7490234375,
0.1351318359375,
0.7490234375
],
[
0.509765625,
0.7392578125,
0.8818359375,
0.74072265625,
0.8818359375,
0.75439453125,
0.509765625,
0.7529296875
],
[
0.1158447265625,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.1158447265625,
0.76318359375
],
[
0.52880859375,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76708984375,
0.52880859375,
0.76708984375
],
[
0.1158447265625,
0.7666015625,
0.48779296875,
0.765625,
0.48779296875,
0.77978515625,
0.1158447265625,
0.78076171875
],
[
0.52880859375,
0.7666015625,
0.869140625,
0.7666015625,
0.869140625,
0.77978515625,
0.52880859375,
0.77978515625
],
[
0.1158447265625,
0.78369140625,
0.48779296875,
0.78369140625,
0.48779296875,
0.79736328125,
0.1158447265625,
0.79736328125
],
[
0.51123046875,
0.79052734375,
0.8818359375,
0.79052734375,
0.8818359375,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.11248779296875,
0.7998046875,
0.385498046875,
0.7998046875,
0.385498046875,
0.8134765625,
0.11248779296875,
0.8134765625
],
[
0.53076171875,
0.802734375,
0.8818359375,
0.802734375,
0.8818359375,
0.81640625,
0.53076171875,
0.81640625
],
[
0.52880859375,
0.81640625,
0.8818359375,
0.81640625,
0.8818359375,
0.830078125,
0.52880859375,
0.830078125
],
[
0.1175537109375,
0.82763671875,
0.287109375,
0.8291015625,
0.287109375,
0.84326171875,
0.1175537109375,
0.8427734375
],
[
0.5283203125,
0.8291015625,
0.8818359375,
0.8291015625,
0.8818359375,
0.8427734375,
0.5283203125,
0.8427734375
],
[
0.52880859375,
0.8427734375,
0.88330078125,
0.8427734375,
0.88330078125,
0.85546875,
0.52880859375,
0.85546875
],
[
0.11419677734375,
0.85498046875,
0.48779296875,
0.85546875,
0.48779296875,
0.87060546875,
0.11419677734375,
0.869140625
],
[
0.5263671875,
0.85498046875,
0.771484375,
0.85400390625,
0.771484375,
0.8681640625,
0.5263671875,
0.869140625
],
[
0.1158447265625,
0.87158203125,
0.48779296875,
0.87255859375,
0.48779296875,
0.88720703125,
0.1158447265625,
0.8857421875
],
[
0.51123046875,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.51123046875,
0.892578125
],
[
0.1158447265625,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.52880859375,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90380859375,
0.52880859375,
0.90380859375
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.1158447265625,
0.91943359375
],
[
0.5283203125,
0.90380859375,
0.88330078125,
0.90478515625,
0.88330078125,
0.91943359375,
0.5283203125,
0.91845703125
]
] |
[
"5Conclusion",
"Fund Denmark (DFF) grant 9131-00019B and",
"in parts by ERC Consolidator Grant DIALECT",
"We present entity linking in the job market domain,",
"101043235.",
"using two existing high-performing neural mod-",
"els. We demonstrate that the bi-encoder architec-",
"ture of BLINK is more suited to the job market",
"References",
"domain compared to the autoregressive GENRE",
"Tom Ayoola, Shubhi Tyagi, Joseph Fisher, Christos",
"model. While strict linking results favor BLINK",
"Christodoulopoulos, and Andrea Pierleoni. 2022. Re-",
"over GENRE, if we relax the number of candidates,",
"FinED: An efficient zero-shot-capable approach to",
"weobservethatGENREperformsslightlybetter.",
"end-to-end entity linking. In Proceedings of the 2022",
"Conference of the North American Chapter of the",
"From a qualitative perspective, the performance of",
"Association for Computational Linguistics: Human",
"strict linking results is modest due to limitations in",
"Language Technologies: Industry Track, pages 209-",
"the evaluation set, which considers only one skill",
"220, Hybrid: Seattle, Washington + Online. Associa-",
"correct per mention. However, upon examining the",
"tion for Computational Linguistics.",
"predictions, we identify valid links, suggesting the",
"Krisztian Balog, Yi Fang, Maarten De Rijke, Pavel",
"possibility of multiple correct links for a particular",
"Serdyukov, and Luo Si. 2012. Expertise retrieval.",
"mention, highlighting the need for more compre-",
"Foundations and Trends in Information Retrieval,",
"hensive evaluation. We hope this work sparks inter-",
"6(2-3):127-256",
"est in entity linking within the job market domain.",
"Jan A. Botha, Zifei Shan, and Daniel Gillick. 2020. En-",
"tity Linking in 100 Languages. In Proceedings of the",
"Limitations",
"2020 Conference on Empirical Methods in Natural",
"Language Processing (EMNLP), pages 7833-7845",
"In the context of EL for ESCO, our approach has",
"Online. Association for Computational Linguistics.",
"several limitations. Firstly, it only supports En-",
"Erik Brynjolfsson and Andrew McAfee. 2011. Race",
"glish, and might not generalize to other languages.",
"against the machine: How the digital revolution is",
"However, several works are working on multilin-",
"accelerating innovation, driving productivity, and",
"gual entity linking (e.g., Botha et al., 2020; De Cao",
"irreversibly transforming employment and the econ-",
"et al., 2022) and ESCO itself consists of 28 Euro-",
"omy. Brynjolfsson and McAfee.",
"pean languages. This work could be extended by",
"Erik Brynjolfsson and Andrew McAfee. 2014.The",
" supporting it for more languages.",
"second machine age:Work,progress, and prosperity",
"Secondly, our EL model is trained on synthetic",
"in a time of brilliant technologies. WW Norton &",
"training data, which may not fully capture the in-",
"Company.",
"tricacies and variations present in real-world doc-",
"deNicola Cao, Gautier Izacard.Sebastian Riedel,and",
"uments. The use of synthetic data could limit its",
"Fabio Petroni. 2021. Autoregressive entity retrieval",
"In International Conference on Learning Representa-",
"performance on actual, real JD texts. Nevertheless,",
"tions.",
"we have human-annotated evaluation data.",
"Moreover, in our evaluation process, we use only",
"Nicola de Cao, Ledell Wu, Kashyap Popat, Mikel",
"one gold-standard ESCO title as the correct answer.",
"Artetxe, Naman Goyal, Mikhail Plekhanov, Luke",
"Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and",
"This approach may not adequately represent a real-",
"Fabio Petroni. 2022. Multilingual autoregressive en-",
"world scenario, where multiple ESCO titles could",
"tity linking.Transactions of the Association for Com-",
"be correct as shown in Table 3.",
"putational Linguistics, 10:274-290.",
"In Table 2, we show that providing in-domain",
"Benjamin Clavie and Guillaume Soulié. 2023. Large",
"data for continuous pre-training shows larger im-",
"language models as batteries-included zero-shot esco",
"provements for GENRE than for BLINK. We did",
"skills matchers. arXiv preprint arXiv:2307.03539.",
"not conduct a detailed analysis on the underlying",
"Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel",
" reasons for these positive variations.",
"Artetxe, Naman Goyal, Mikhail Plekhanov, Luke",
"Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and",
"Acknowledgements",
"Fabio Petroni. 2022. Multilingual autoregressive en-",
"tity linking. Transactions of the Association for Com-",
"We thank the MaiNLP and NLPnorth group for",
"putational Linguistics, 10:274-290.",
"feedback on an earlier version of this paper, and",
"Jens-Joris Decorte, Jeroen Van Hautte, Johannes",
"the reviewers for their insightful comments. This",
"Deleu, Chris Develder, and Thomas Demeester",
"research is supported by the Independent Research",
"2022. Design of negative sampling strategies for"
] |
{
"arxiv_id": "2401.17979",
"language": "en",
"page_idx": 4
}
|
||
2306.01153-en-0
|
2306.01153-en-0.jpg
|
## Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference
Yan Xu$^{1+\text{DeqianKong}^{2+\textbf{DehongXu}^2}}$ Ziwei Ji$^{1}$ Bo Pang$^{3}$ Pascale Fung$^{1+\text{YingNianWu}^{2+}}$
## $1. \textbf{Introduction}$
$\mathbf{Abstract}$
Open-domain dialogue systems aim at fulfilling humanmachine conversations by producing human-like responses to utterances from humans (Serban et al., 2016). The emergence of large-scale pre-trained language models (PLMs) has turbocharged the development of open-domain dialogue systems (Zhang et al., 2020; Roller et al., 2021). By maximizing the token-level likelihood of gold responses given dialogue history, dialogue systems can generate fluent and natural responses. However, challenges remain to ensure that responses are diverse and informative (Ghazvininejad et al., 2018), yet remain factual and accurate (Shuster et al., 2021). Prior approaches for improving the diversity of dialogue responses focus on preventing them from being dull and repetitive (Zhao et al., 2019; Xu et al., 2022), while optimizing for diversity alone tends to encourage the dialogue system to hallucinate non-factual responses (Ji et al., 2022a). $\dot{\text{ChatGPT(OpenAI,2023) tries to address this issue using}}$ a reward model trained with human preference. However, it is very resource-consuming. To address this limitation in generative dialogue systems, we need to ground system responses on external knowledge effectively.
The capability to generate responses with diversity and faithfulness using factual knowledge is paramount for creating a human-like, trustworthy dialogue system. Common strategies either adopt a two-step paradigm, which optimizes knowledge selection and response generation separately, and may overlook the inherent correlation between these two tasks, or leverage conditional variational method to jointly optimize knowledge selection and response generation by employing an inference network. In this paper, we present an end-to-end learning framework, termed Sequential Posterior Inference (SPI), capable of selecting knowledge and generating dialogues by approximately sampling from the posterior distribution. Unlike other methods, SPI does not require the inference network or assume a simple geometry of the posterior distribution. This straightforward and intuitive inference procedure of SPI directly queries the response generation model, allowing for accurate knowledge selection and generation of faithful responses. In addition to modeling contributions, our experimental results on two common dialogue datasets (Wizard of Wikipedia and Holl-E) demonstrate that SPI outperforms previous strong baselines according to both automatic and human evaluation metrics. The code and checkpoints are available at https : //github.com/deqiankong/SPI.
Knowledge-grounded dialogue (KGD) has been investigated in recent years (Dinan et al., 2018; Li et al., 2020; Xu et al., 2021; Yang et al., 2022). The objective is to enhance dialogue response generation to facilitate engaging and in-depth conversations, while avoiding the inclusion of non-factual information. The task can be achieved following a two-step paradigm:(1) knowledge selection;(2) response generation. Some previous works (Lian et al.; Kim et al., 2019; Chen et al., 2020) optimize these two steps individually. They first utilize variational inference (Kingma & Welling, 2014) for knowledge selection, where the prior distribution is conditioned on dialogue history, and the posterior distribution depends on both response and dialogue history. Then they optimize the response generation task based on the selected knowledge. Since knowledge selection in KGD tasks is a complex one-to-many problem, it is not trivial to generate a factual response with dialogue history and selected knowledge solely, not to mention that inaccurate knowledge may be chosen even with a complex knowledge selection module. Other works (Liu et al., 2021) bypass the
$^{\dagger}$Equal contribution$^{\dagger}$Equal advising$^{\text{l}}$Center for Artificial Intelligence Research (CAiRE), The Hong Kong University of Science and Technology, Hong Kong $^{2}$Department of Statistics, UCLA, CA, USA $^3$Salesforce Research. Correspondence to: Yan Xu<yxucb@connect.ust.hk>, Deqian Kong <deqiankong@ucla.edu>. Proceedings of the 40$^{th}$ International Conference on Machine $Learning$,Honolulu, Hawaii, USA. PMLR 202,2023. Copyright 2023 by the author(s).
1
|
[
[
0.1478271484375,
0.1142578125,
0.82666015625,
0.1142578125,
0.82666015625,
0.1287841796875,
0.1478271484375,
0.1287841796875
],
[
0.3349609375,
0.1363525390625,
0.640625,
0.1363525390625,
0.640625,
0.154052734375,
0.3349609375,
0.154052734375
],
[
0.1405029296875,
0.1988525390625,
0.8310546875,
0.1976318359375,
0.8310546875,
0.21533203125,
0.1405029296875,
0.216552734375
],
[
0.24267578125,
0.242431640625,
0.32177734375,
0.242431640625,
0.32177734375,
0.260009765625,
0.24267578125,
0.260009765625
],
[
0.5009765625,
0.2442626953125,
0.62841796875,
0.2442626953125,
0.62841796875,
0.2587890625,
0.5009765625,
0.2587890625
],
[
0.0187835693359375,
0.26318359375,
0.066162109375,
0.26318359375,
0.068603515625,
0.3837890625,
0.021240234375,
0.3837890625
],
[
0.5009765625,
0.26953125,
0.88623046875,
0.26953125,
0.88623046875,
0.28271484375,
0.5009765625,
0.28271484375
],
[
0.5009765625,
0.285400390625,
0.8857421875,
0.285400390625,
0.8857421875,
0.298583984375,
0.5009765625,
0.298583984375
],
[
0.12091064453125,
0.3017578125,
0.441162109375,
0.3017578125,
0.441162109375,
0.31494140625,
0.12091064453125,
0.31494140625
],
[
0.5009765625,
0.3017578125,
0.8857421875,
0.3017578125,
0.8857421875,
0.3125,
0.5009765625,
0.3125
],
[
0.1217041015625,
0.317626953125,
0.441162109375,
0.317626953125,
0.441162109375,
0.330810546875,
0.1217041015625,
0.330810546875
],
[
0.5009765625,
0.31494140625,
0.88623046875,
0.31494140625,
0.88623046875,
0.328857421875,
0.5009765625,
0.328857421875
],
[
0.12091064453125,
0.333251953125,
0.441162109375,
0.33203125,
0.441162109375,
0.34521484375,
0.12091064453125,
0.3466796875
],
[
0.5009765625,
0.330810546875,
0.8857421875,
0.330810546875,
0.8857421875,
0.3447265625,
0.5009765625,
0.3447265625
],
[
0.1192626953125,
0.3466796875,
0.439453125,
0.3466796875,
0.439453125,
0.35986328125,
0.1192626953125,
0.35986328125
],
[
0.498291015625,
0.34521484375,
0.88623046875,
0.3447265625,
0.88623046875,
0.359130859375,
0.498291015625,
0.35986328125
],
[
0.12091064453125,
0.363037109375,
0.441162109375,
0.363037109375,
0.441162109375,
0.376953125,
0.12091064453125,
0.376953125
],
[
0.5009765625,
0.361083984375,
0.8857421875,
0.361083984375,
0.8857421875,
0.375,
0.5009765625,
0.375
],
[
0.0269622802734375,
0.37255859375,
0.0555419921875,
0.37255859375,
0.0555419921875,
0.391357421875,
0.0269622802734375,
0.391357421875
],
[
0.1192626953125,
0.376953125,
0.441162109375,
0.37744140625,
0.441162109375,
0.392578125,
0.1192626953125,
0.391357421875
],
[
0.5,
0.376953125,
0.88623046875,
0.376953125,
0.88623046875,
0.39013671875,
0.5,
0.39013671875
],
[
0.12091064453125,
0.393310546875,
0.441162109375,
0.393310546875,
0.441162109375,
0.4072265625,
0.12091064453125,
0.4072265625
],
[
0.5,
0.39013671875,
0.8857421875,
0.389404296875,
0.8857421875,
0.404052734375,
0.5,
0.40478515625
],
[
0.0196075439453125,
0.399658203125,
0.05963134765625,
0.3984375,
0.062103271484375,
0.48095703125,
0.0228729248046875,
0.481689453125
],
[
0.12091064453125,
0.40771484375,
0.441162109375,
0.40771484375,
0.441162109375,
0.421630859375,
0.12091064453125,
0.421630859375
],
[
0.5009765625,
0.406005859375,
0.88623046875,
0.406005859375,
0.88623046875,
0.419189453125,
0.5009765625,
0.419189453125
],
[
0.12091064453125,
0.423583984375,
0.441162109375,
0.423583984375,
0.441162109375,
0.4375,
0.12091064453125,
0.4375
],
[
0.5009765625,
0.421630859375,
0.88623046875,
0.421630859375,
0.88623046875,
0.43505859375,
0.5009765625,
0.43505859375
],
[
0.12091064453125,
0.438232421875,
0.441162109375,
0.438232421875,
0.441162109375,
0.451904296875,
0.12091064453125,
0.451904296875
],
[
0.5009765625,
0.436279296875,
0.88623046875,
0.436279296875,
0.88623046875,
0.449462890625,
0.5009765625,
0.449462890625
],
[
0.1217041015625,
0.453857421875,
0.441162109375,
0.453857421875,
0.441162109375,
0.46728515625,
0.1217041015625,
0.46728515625
],
[
0.5009765625,
0.451904296875,
0.8857421875,
0.451904296875,
0.8857421875,
0.46533203125,
0.5009765625,
0.46533203125
],
[
0.1217041015625,
0.468505859375,
0.44287109375,
0.468505859375,
0.44287109375,
0.481689453125,
0.1217041015625,
0.481689453125
],
[
0.5009765625,
0.466552734375,
0.88623046875,
0.466552734375,
0.88623046875,
0.479736328125,
0.5009765625,
0.479736328125
],
[
0.0228729248046875,
0.48291015625,
0.060455322265625,
0.48291015625,
0.058013916015625,
0.70556640625,
0.0196075439453125,
0.70556640625
],
[
0.12091064453125,
0.484130859375,
0.44287109375,
0.484130859375,
0.44287109375,
0.49755859375,
0.12091064453125,
0.49755859375
],
[
0.5,
0.48095703125,
0.88623046875,
0.481689453125,
0.88623046875,
0.49560546875,
0.5,
0.494384765625
],
[
0.12091064453125,
0.498779296875,
0.441162109375,
0.498779296875,
0.441162109375,
0.51220703125,
0.12091064453125,
0.51220703125
],
[
0.5009765625,
0.496826171875,
0.88623046875,
0.496826171875,
0.88623046875,
0.51025390625,
0.5009765625,
0.51025390625
],
[
0.12091064453125,
0.51318359375,
0.44287109375,
0.51318359375,
0.44287109375,
0.52783203125,
0.12091064453125,
0.52783203125
],
[
0.5009765625,
0.51220703125,
0.88623046875,
0.51220703125,
0.88623046875,
0.52587890625,
0.5009765625,
0.52587890625
],
[
0.1192626953125,
0.52783203125,
0.44287109375,
0.52880859375,
0.44287109375,
0.54248046875,
0.1192626953125,
0.54150390625
],
[
0.5009765625,
0.5263671875,
0.88623046875,
0.5263671875,
0.88623046875,
0.54052734375,
0.5009765625,
0.54052734375
],
[
0.12091064453125,
0.54443359375,
0.438720703125,
0.54443359375,
0.438720703125,
0.5546875,
0.12091064453125,
0.5546875
],
[
0.5,
0.54248046875,
0.88623046875,
0.54248046875,
0.88623046875,
0.55615234375,
0.5,
0.55615234375
],
[
0.1217041015625,
0.5595703125,
0.441162109375,
0.5595703125,
0.441162109375,
0.57275390625,
0.1217041015625,
0.57275390625
],
[
0.498291015625,
0.55615234375,
0.8857421875,
0.55615234375,
0.8857421875,
0.57080078125,
0.498291015625,
0.57080078125
],
[
0.12091064453125,
0.57470703125,
0.441162109375,
0.57470703125,
0.441162109375,
0.58837890625,
0.12091064453125,
0.58837890625
],
[
0.5,
0.57275390625,
0.7958984375,
0.57275390625,
0.7958984375,
0.5859375,
0.5,
0.5859375
],
[
0.1192626953125,
0.58837890625,
0.441162109375,
0.5888671875,
0.441162109375,
0.60400390625,
0.1192626953125,
0.60302734375
],
[
0.5009765625,
0.59326171875,
0.8857421875,
0.59326171875,
0.8857421875,
0.60791015625,
0.5009765625,
0.60791015625
],
[
0.1217041015625,
0.60595703125,
0.441162109375,
0.60595703125,
0.441162109375,
0.6162109375,
0.1217041015625,
0.6162109375
],
[
0.498291015625,
0.609375,
0.8857421875,
0.60791015625,
0.8857421875,
0.62255859375,
0.498291015625,
0.62353515625
],
[
0.12335205078125,
0.62060546875,
0.439453125,
0.62060546875,
0.439453125,
0.630859375,
0.12335205078125,
0.630859375
],
[
0.5009765625,
0.625,
0.88623046875,
0.625,
0.88623046875,
0.63818359375,
0.5009765625,
0.63818359375
],
[
0.1217041015625,
0.63623046875,
0.439453125,
0.63623046875,
0.439453125,
0.646484375,
0.1217041015625,
0.646484375
],
[
0.5,
0.64013671875,
0.8857421875,
0.64013671875,
0.8857421875,
0.65380859375,
0.5,
0.65380859375
],
[
0.12091064453125,
0.6494140625,
0.441162109375,
0.6494140625,
0.441162109375,
0.66357421875,
0.12091064453125,
0.66357421875
],
[
0.5,
0.65478515625,
0.88623046875,
0.65478515625,
0.88623046875,
0.66845703125,
0.5,
0.66845703125
],
[
0.12091064453125,
0.6640625,
0.441162109375,
0.6640625,
0.441162109375,
0.67822265625,
0.12091064453125,
0.67822265625
],
[
0.50244140625,
0.671875,
0.8857421875,
0.671875,
0.8857421875,
0.681640625,
0.50244140625,
0.681640625
],
[
0.12091064453125,
0.6796875,
0.441162109375,
0.6796875,
0.441162109375,
0.693359375,
0.12091064453125,
0.693359375
],
[
0.50244140625,
0.6875,
0.8837890625,
0.6875,
0.8837890625,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.12091064453125,
0.69580078125,
0.439453125,
0.69580078125,
0.439453125,
0.70556640625,
0.12091064453125,
0.70556640625
],
[
0.5,
0.70068359375,
0.8857421875,
0.70068359375,
0.8857421875,
0.7138671875,
0.5,
0.7138671875
],
[
0.12091064453125,
0.71044921875,
0.443603515625,
0.71044921875,
0.443603515625,
0.7236328125,
0.12091064453125,
0.7236328125
],
[
0.5009765625,
0.71533203125,
0.88623046875,
0.71533203125,
0.88623046875,
0.72900390625,
0.5009765625,
0.72900390625
],
[
0.12091064453125,
0.72607421875,
0.390625,
0.72607421875,
0.390625,
0.7392578125,
0.12091064453125,
0.7392578125
],
[
0.5009765625,
0.73095703125,
0.88623046875,
0.73095703125,
0.88623046875,
0.744140625,
0.5009765625,
0.744140625
],
[
0.5,
0.74365234375,
0.88623046875,
0.74560546875,
0.88623046875,
0.76025390625,
0.5,
0.75830078125
],
[
0.5009765625,
0.76025390625,
0.88623046875,
0.76025390625,
0.88623046875,
0.7734375,
0.5009765625,
0.7734375
],
[
0.5,
0.77587890625,
0.88623046875,
0.77587890625,
0.88623046875,
0.7890625,
0.5,
0.7890625
],
[
0.10540771484375,
0.79052734375,
0.4755859375,
0.7890625,
0.4755859375,
0.8037109375,
0.10540771484375,
0.8046875
],
[
0.5,
0.79150390625,
0.88623046875,
0.79150390625,
0.88623046875,
0.8046875,
0.5,
0.8046875
],
[
0.0882568359375,
0.8046875,
0.4755859375,
0.8046875,
0.4755859375,
0.818359375,
0.0882568359375,
0.818359375
],
[
0.498291015625,
0.80615234375,
0.88623046875,
0.80615234375,
0.88623046875,
0.8193359375,
0.498291015625,
0.8193359375
],
[
0.08905029296875,
0.81640625,
0.4755859375,
0.81640625,
0.4755859375,
0.83056640625,
0.08905029296875,
0.83056640625
],
[
0.5009765625,
0.8212890625,
0.88623046875,
0.8212890625,
0.88623046875,
0.83544921875,
0.5009765625,
0.83544921875
],
[
0.0882568359375,
0.82763671875,
0.473876953125,
0.830078125,
0.473876953125,
0.84326171875,
0.0882568359375,
0.84130859375
],
[
0.5,
0.8359375,
0.88623046875,
0.8359375,
0.88623046875,
0.849609375,
0.5,
0.849609375
],
[
0.0882568359375,
0.84228515625,
0.471435546875,
0.84228515625,
0.471435546875,
0.85595703125,
0.0882568359375,
0.85595703125
],
[
0.5,
0.8515625,
0.8857421875,
0.8515625,
0.8857421875,
0.86572265625,
0.5,
0.86572265625
],
[
0.08660888671875,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.5009765625,
0.86865234375,
0.8837890625,
0.86865234375,
0.8837890625,
0.87890625,
0.5009765625,
0.87890625
],
[
0.0882568359375,
0.8798828125,
0.4755859375,
0.8798828125,
0.4755859375,
0.89306640625,
0.0882568359375,
0.89306640625
],
[
0.5,
0.8818359375,
0.8857421875,
0.8818359375,
0.8857421875,
0.89599609375,
0.5,
0.89599609375
],
[
0.0882568359375,
0.892578125,
0.221435546875,
0.892578125,
0.221435546875,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.48193359375,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.9345703125,
0.48193359375,
0.9345703125
]
] |
[
"Diverse and Faithful Knowledge-Grounded Dialogue Generation via",
"Sequential Posterior Inference",
"Yan XuI t Deqian Kong 2 t Dehong Xu? Ziwei Jil Bo Pang3 Pascale Fung I+ Ying Nian Wu?",
"Abstract",
"1. Introduction",
"2",
"Open-domain dialogue systems aim at fulfilling human-",
"machine conversations by producing human-like responses",
"The capability to generate responses with diver-",
"to utterances from humans (Serban et al.,2016).Theemer-",
"sity and faithfulness using factual knowledge is",
"gence of large-scale pre-trained language models (PLMs)",
"paramount for creating a human-like, trustworthy",
"has turbocharged the development of open-domain dialogue",
"dialogue system. Common strategies either adopt",
"systems (Zhang et al., 2020; Roller et al., 2021).By maxi-",
"a two-step paradigm, which optimizes knowledge",
"mizing the token-level likelihood of gold responses given",
"5",
"selection and response generation separately, and",
"dialogue history, dialogue systems can generate fluent and",
"may overlook the inherent correlation between",
"natural responses. However, challenges remain to ensure",
"['s]",
"these two tasks, or leverage conditional varia-",
"that responses are diverse and informative (Ghazvininejad",
"tional method to jointly optimize knowledge se-",
"et al., 2018), yet remain factual and accurate (Shuster et al.,",
"lection and response generation by employing",
"2021).Prior approaches for improving the diversity of dia-",
"an inference network. In this paper, we present",
"logue responses focus on preventing them from being dull",
"an end-to-end learning framework, termed Se-",
"and repetitive (Zhao et al., 2019; Xu et al., 2022), while op-",
"A05007A0",
"quential Posterior Inference (SPI), capable of se-",
"timizingfor diversity alone tends to encourage the dialogue",
"lecting knowledge and generating dialogues by",
"system to hallucinate non-factual responses (Ji et al., 2022a).",
"approximately sampling from the posterior dis-",
"ChatGPT (OpenAI, 2023) tries to address this issue using",
"tribution. Unlike other methods, SPI does not",
"a reward model trained with human preference. However,",
"reguire the inference network or assume a sim",
"it is very resource-consuming. To address this limitation",
"ple geometry of the posterior distribution. This",
"in generative dialogue systems, we need to ground system",
"straightforward and intuitive inference procedure",
"responses on external knowledge effectively.",
"of SPI directly queries the response generation",
"Knowledge-grounded dialogue (KGD) has been investigated",
"model,allowingforaccurateknowledgeselection",
"in recent years (Dinan et al., 2018; Li et al., 2020; Xu",
"andgenerationoffaithfulresponses.Inaddition",
"et al., 2021; Yang et al., 2022). The objective is to en-",
"tomodeling contributions,ourexperimentalre",
"hance dialogue response generation to facilitate engaging",
"sults on two common dialogue datasets (Wizard",
"and in-depth conversations, while avoiding the inclusion",
"of Wikipedia and Holl-E) demonstrate that SPI",
"ofnon-factualinformation.Thetaskcanbeachievedfol-",
"outperforms previous strong baselines according",
"lowingatwo-stepparadigm:(1)knowledgeselection;(2)",
"to both automatic and human evaluationmetrics",
"response generation. Some previous works (Lian et al.; Kim",
"The code and checkpoints are available at ht t ps :",
"et al., 2019; Chen et al., 2020) optimize these two steps in-",
"//github.com/deqiankong/spI.",
"dividually. They first utilize variational inference (Kingma",
"& Welling,2014) for knowledge selection, where the prior",
"distribution is conditioned on dialogue history, and the poste-",
"rior distribution depends on both response and dialogue his-",
"tEqual contribution *Equal advising 'Center for Artificial Intel-",
"tory. Then they optimize the response generation task based",
"ligence Research (CAiRE), The Hong Kong University of Science",
"on the selected knowledge. Since knowledge selection in",
"and Technology, Hong Kong 'Department of Statistics, UCLA,",
"KGD tasks is a complex one-to-many problem, it is not triv-",
"CA,USA3SalesforceResearch.Correspondence to:Yan Xu<yx-",
"ial to generate a factual response with dialogue history and",
"ucb@connect.ust.hk>, Deqian Kong <deqiankong@ucla.edu>.",
"selected knowledge solely, not to mention that inaccurate",
"Proceedings of the 4oth International Conference on Machine",
"knowledgemaybechosenevenwithacomplexknowledge",
"Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright",
"selection module. Other works (Liu et al., 2021) bypass the",
"2023 by the author(s).",
"1"
] |
{
"arxiv_id": "2306.01153",
"language": "en",
"page_idx": 0
}
|
||
2306.01153-en-4
|
2306.01153-en-4.jpg
|
Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference
we propose the following sampling procedure,
The learnable knowledge prior with parameter $\alpha_1$ also functions akin to a multi-label classifier. We use both gold annotations and posterior knowledge selection as ground-truth labels, mirroring the updating of the initializer.
$$\begin{aligned}z^0&\sim p_{\alpha_2}(z|C_s),\\z^{t+1}&=z^t+\delta\nabla_z\log p_\theta(z^t|R,C_s)+\sqrt{2\delta}\epsilon_t,\end{aligned}$$
(21)
Generation We can get a response by greedy search as summarized in Algorithm 2. Given dialogue context $C=$ $(H,\mathbf{K})$,we can first select the knowledge candidate with the highest logit from the initializer $f_\gamma(C)$ or learnable prior $f_{\alpha_1}(C)$ based on different prior choices,
where $t=1,\ldots,T$, and the initial state for the Markov chain is sampled from the context-conditioned prior distribution. The total length of the Markov chain is rather small (e.g. $T=5).$ Further theoretical underpinnings of this approximate sampling and learning method can be found in Appendix A.
(25)
### 2.4. Algorithms
$$s=\arg\max_{i}f_{\gamma}(C_{i}),\:i\in\{1,\ldots,S\}.$$
$$s=\arg\max_{i}f_{\alpha_{1}}(C_{i}),\:i\in\{1,\ldots,M\}.$$
(26)
The choice of prior for the discrete variable $s$ leads to minor
variations in the learning and generation algorithms.
Then we use sample mean of the response latent variable
$p_{\alpha_2}(z|C_s)\sim\mathcal{N}(\hat{f}_{\alpha_2}(C_s),\mathbf{I})$ to generate dialogue,
$$\hat{R}\sim p_\beta(R|z=f_{\alpha_2}(C_s),C_s).$$
(27)
Learning with Uniform Knowledge Prior Given learning iterations $\tau=1,\ldots,T_L$, the generative model with parameters $\theta=\{\alpha=\alpha_2,\beta\}$ can be updated through
$$\begin{aligned}&\theta_{\tau+1}=\theta_{\tau}+\eta_{1}\Delta\theta,\\&\Delta\theta=\frac{1}{N}\sum_{n=1}^{N}\mathbb{E}_{p_{\theta_{\tau}}(s^{n},z^{n}|R^{n},C^{n})}[\nabla_{\theta}\log p_{\alpha_{2}}(z^{n}|C_{s}^{n})\\&+\nabla_{\theta}\log p_{\beta}(R|z^{n},C_{s}^{n})].&\text{(2}\end{aligned}$$
The additional top-$S$ initializer $f_\gamma(C)$ with parameters $\gamma$ can be viewed as a multi-label classifer or multiple binary classifiers and be updated by cross-entropy loss,
<table>
<tbody>
<tr>
<td>$\overline{\mathbf{lg}}$</td>
<td>thm 1 Learning with Sequential Posterior Inference</td>
</tr>
<tr>
<td>$T$ 0</td>
<td>ep size $\delta$, initial weights $\theta_0,\gamma_0.$ $\textbf{ut: Updated weights }\theta _{T}, \gamma _{T}.$</td>
</tr>
</tbody>
</table>
$$\mathcal{L}_{\mathrm{CE}}(y,C)=-\sum_{i=1}^{M}y_{i}\log f_{\gamma}(C_{i})+(1-y_{i})\log(1-f_{\gamma}(C_{i})),\\\gamma_{\tau+1}=\gamma_{\tau}-\eta_{2}\frac{1}{N}\sum_{n=1}^{N}\nabla_{\gamma}\mathcal{L}_{\mathrm{CE}}(y^{n},C^{n}),\text{(23)}$$
where $\{y^n\}_{n=1}^N$ denotes labels which can be obtained by posterior knowledge selection and/or annotations (if we use both posterior knowledge selection and annotations, it is possible that two of $y_i$ 's equal to 1). Therefore, the learned initializer can output top-$S$ candidates that likely include the gold knowledge or the one from posterior knowledge selection
Learning with Learnable Knowledge Prior The generative model with parameters $\theta=\{\alpha=(\alpha_1,\alpha_2),\beta\}$ can be## updated through
<table>
<tbody>
<tr>
<td>Algorithm 2 Knowledge-Grounded Response</td>
<td>$\overline{A}$</td>
<td>rithm 2 Knowledge-Grounded Response Generation</td>
</tr>
</tbody>
</table>
$$\begin{aligned}&\theta_{\tau+1}=\theta_{\tau}+\eta_{1}\Delta\theta,\\&\Delta\theta=\frac{1}{N}\sum_{n=1}^{N}\mathbb{E}_{p_{\theta_{\tau}}(s^{n},z^{n}|R^{n},C^{n})}[\nabla_{\theta}\log p_{\alpha_{1}}(s^{n}|C^{n})\\&+\nabla_{\theta}\log p_{\alpha_{2}}(z^{n}|C_{s}^{n})+\nabla_{\theta}\log p_{\beta}(R|z^{n},C_{s}^{n})].\end{aligned}$$
(24)
5
|
[
[
0.1748046875,
0.056182861328125,
0.7998046875,
0.05682373046875,
0.7998046875,
0.0706787109375,
0.1748046875,
0.0694580078125
],
[
0.0882568359375,
0.08587646484375,
0.39208984375,
0.08587646484375,
0.39208984375,
0.10101318359375,
0.0882568359375,
0.10101318359375
],
[
0.5009765625,
0.08587646484375,
0.88623046875,
0.08587646484375,
0.88623046875,
0.0997314453125,
0.5009765625,
0.0997314453125
],
[
0.1192626953125,
0.106689453125,
0.233642578125,
0.10919189453125,
0.233642578125,
0.126953125,
0.1192626953125,
0.125
],
[
0.5009765625,
0.10101318359375,
0.8857421875,
0.10101318359375,
0.8857421875,
0.1142578125,
0.5009765625,
0.1142578125
],
[
0.5,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.5,
0.1300048828125
],
[
0.1192626953125,
0.1275634765625,
0.42236328125,
0.1300048828125,
0.42236328125,
0.1483154296875,
0.1192626953125,
0.146484375
],
[
0.44287109375,
0.1331787109375,
0.4755859375,
0.1331787109375,
0.4755859375,
0.14453125,
0.44287109375,
0.14453125
],
[
0.498291015625,
0.1319580078125,
0.8056640625,
0.1319580078125,
0.8056640625,
0.1427001953125,
0.498291015625,
0.1427001953125
],
[
0.0882568359375,
0.1572265625,
0.1846923828125,
0.1572265625,
0.1846923828125,
0.1673583984375,
0.0882568359375,
0.1673583984375
],
[
0.200927734375,
0.1558837890625,
0.473876953125,
0.1558837890625,
0.473876953125,
0.17041015625,
0.200927734375,
0.17041015625
],
[
0.5,
0.160400390625,
0.88623046875,
0.1610107421875,
0.88623046875,
0.175537109375,
0.5,
0.1749267578125
],
[
0.0882568359375,
0.1717529296875,
0.473876953125,
0.1717529296875,
0.473876953125,
0.186279296875,
0.0882568359375,
0.186279296875
],
[
0.5009765625,
0.1767578125,
0.8837890625,
0.1767578125,
0.8837890625,
0.1900634765625,
0.5009765625,
0.1900634765625
],
[
0.08905029296875,
0.1875,
0.473876953125,
0.1875,
0.473876953125,
0.2008056640625,
0.08905029296875,
0.2008056640625
],
[
0.5,
0.1912841796875,
0.8857421875,
0.1912841796875,
0.8857421875,
0.2052001953125,
0.5,
0.2052001953125
],
[
0.08660888671875,
0.2020263671875,
0.473876953125,
0.2008056640625,
0.473876953125,
0.21533203125,
0.08660888671875,
0.216552734375
],
[
0.5009765625,
0.20703125,
0.88623046875,
0.20703125,
0.88623046875,
0.2203369140625,
0.5009765625,
0.2203369140625
],
[
0.0882568359375,
0.2171630859375,
0.473876953125,
0.216552734375,
0.473876953125,
0.2310791015625,
0.0882568359375,
0.231689453125
],
[
0.5009765625,
0.2215576171875,
0.765625,
0.2215576171875,
0.765625,
0.236083984375,
0.5009765625,
0.236083984375
],
[
0.0882568359375,
0.2342529296875,
0.1732177734375,
0.2342529296875,
0.1732177734375,
0.24560546875,
0.0882568359375,
0.24560546875
],
[
0.56640625,
0.24365234375,
0.81884765625,
0.242431640625,
0.81884765625,
0.260009765625,
0.56640625,
0.261474609375
],
[
0.85546875,
0.2442626953125,
0.88818359375,
0.2442626953125,
0.88818359375,
0.260009765625,
0.85546875,
0.260009765625
],
[
0.08660888671875,
0.261962890625,
0.1993408203125,
0.26318359375,
0.1993408203125,
0.27783203125,
0.08660888671875,
0.276611328125
],
[
0.56201171875,
0.2841796875,
0.82275390625,
0.2841796875,
0.82275390625,
0.298583984375,
0.56201171875,
0.298583984375
],
[
0.85546875,
0.285400390625,
0.88623046875,
0.285400390625,
0.88623046875,
0.29736328125,
0.85546875,
0.29736328125
],
[
0.08660888671875,
0.3017578125,
0.4306640625,
0.3017578125,
0.4306640625,
0.31494140625,
0.08660888671875,
0.31494140625
],
[
0.0882568359375,
0.287353515625,
0.473876953125,
0.287353515625,
0.473876953125,
0.300537109375,
0.0882568359375,
0.300537109375
],
[
0.498291015625,
0.3125,
0.88623046875,
0.313232421875,
0.88623046875,
0.32763671875,
0.498291015625,
0.326904296875
],
[
0.08905029296875,
0.333251953125,
0.473876953125,
0.333251953125,
0.473876953125,
0.3466796875,
0.08905029296875,
0.3466796875
],
[
0.498291015625,
0.330078125,
0.8310546875,
0.328857421875,
0.8310546875,
0.343505859375,
0.498291015625,
0.3447265625
],
[
0.0882568359375,
0.348388671875,
0.473876953125,
0.347900390625,
0.473876953125,
0.3623046875,
0.0882568359375,
0.363037109375
],
[
0.59326171875,
0.348388671875,
0.79150390625,
0.35107421875,
0.79150390625,
0.36865234375,
0.59326171875,
0.3662109375
],
[
0.85546875,
0.3515625,
0.88818359375,
0.3515625,
0.88818359375,
0.367431640625,
0.85546875,
0.367431640625
],
[
0.0849609375,
0.361083984375,
0.438720703125,
0.3623046875,
0.438720703125,
0.380126953125,
0.0849609375,
0.37890625
],
[
0.111083984375,
0.38818359375,
0.2418212890625,
0.38818359375,
0.2418212890625,
0.401611328125,
0.111083984375,
0.401611328125
],
[
0.5,
0.39013671875,
0.8837890625,
0.391357421875,
0.8837890625,
0.406005859375,
0.5,
0.40478515625
],
[
0.181396484375,
0.4091796875,
0.1953125,
0.4091796875,
0.1953125,
0.417236328125,
0.181396484375,
0.417236328125
],
[
0.51123046875,
0.406005859375,
0.88818359375,
0.4072265625,
0.88818359375,
0.427978515625,
0.51123046875,
0.4267578125
],
[
0.158447265625,
0.4140625,
0.17236328125,
0.4140625,
0.17236328125,
0.423583984375,
0.158447265625,
0.423583984375
],
[
0.1829833984375,
0.41845703125,
0.1993408203125,
0.4248046875,
0.193603515625,
0.431884765625,
0.17724609375,
0.425537109375
],
[
0.188720703125,
0.419189453125,
0.451904296875,
0.417236328125,
0.451904296875,
0.436279296875,
0.188720703125,
0.438232421875
],
[
0.51708984375,
0.4267578125,
0.88623046875,
0.4267578125,
0.88623046875,
0.43994140625,
0.51708984375,
0.43994140625
],
[
0.111083984375,
0.42041015625,
0.146240234375,
0.42041015625,
0.146240234375,
0.43310546875,
0.111083984375,
0.43310546875
],
[
0.1544189453125,
0.4287109375,
0.181396484375,
0.4287109375,
0.181396484375,
0.43994140625,
0.1544189453125,
0.43994140625
],
[
0.1829833984375,
0.43994140625,
0.1953125,
0.43994140625,
0.1953125,
0.4462890625,
0.1829833984375,
0.4462890625
],
[
0.51708984375,
0.44140625,
0.8857421875,
0.44140625,
0.8857421875,
0.455078125,
0.51708984375,
0.455078125
],
[
0.146240234375,
0.451904296875,
0.310546875,
0.451904296875,
0.310546875,
0.466552734375,
0.146240234375,
0.466552734375
],
[
0.443603515625,
0.45068359375,
0.476318359375,
0.45068359375,
0.476318359375,
0.466552734375,
0.443603515625,
0.466552734375
],
[
0.5146484375,
0.455078125,
0.75341796875,
0.45703125,
0.75341796875,
0.4716796875,
0.5146484375,
0.4697265625
],
[
0.51318359375,
0.4697265625,
0.740234375,
0.47021484375,
0.740234375,
0.488037109375,
0.51318359375,
0.4873046875
],
[
0.08660888671875,
0.48291015625,
0.473876953125,
0.484130859375,
0.473876953125,
0.498779296875,
0.08660888671875,
0.49755859375
],
[
0.515625,
0.488037109375,
0.63623046875,
0.488037109375,
0.63623046875,
0.498779296875,
0.515625,
0.498779296875
],
[
0.0882568359375,
0.498779296875,
0.473876953125,
0.5,
0.473876953125,
0.5146484375,
0.0882568359375,
0.51318359375
],
[
0.53515625,
0.501953125,
0.78759765625,
0.501953125,
0.78759765625,
0.51513671875,
0.53515625,
0.51513671875
],
[
0.0882568359375,
0.51513671875,
0.404296875,
0.51513671875,
0.404296875,
0.52880859375,
0.0882568359375,
0.52880859375
],
[
0.53369140625,
0.517578125,
0.7646484375,
0.517578125,
0.7646484375,
0.53076171875,
0.53369140625,
0.53076171875
],
[
0.2034912109375,
0.537109375,
0.2197265625,
0.537109375,
0.2197265625,
0.546875,
0.2034912109375,
0.546875
],
[
0.5498046875,
0.53271484375,
0.77783203125,
0.53271484375,
0.77783203125,
0.546875,
0.5498046875,
0.546875
],
[
0.0906982421875,
0.55029296875,
0.1805419921875,
0.55029296875,
0.1805419921875,
0.56494140625,
0.0906982421875,
0.56494140625
],
[
0.1993408203125,
0.55029296875,
0.498291015625,
0.55029296875,
0.498291015625,
0.56494140625,
0.1993408203125,
0.56494140625
],
[
0.5498046875,
0.54736328125,
0.7099609375,
0.54736328125,
0.7099609375,
0.56103515625,
0.5498046875,
0.56103515625
],
[
0.5498046875,
0.5625,
0.88623046875,
0.5625,
0.88623046875,
0.57568359375,
0.5498046875,
0.57568359375
],
[
0.2034912109375,
0.57080078125,
0.2181396484375,
0.57080078125,
0.2181396484375,
0.57470703125,
0.2034912109375,
0.57470703125
],
[
0.55126953125,
0.5791015625,
0.86376953125,
0.5791015625,
0.86376953125,
0.5888671875,
0.55126953125,
0.5888671875
],
[
0.2442626953125,
0.5869140625,
0.250732421875,
0.59228515625,
0.2459716796875,
0.5966796875,
0.238525390625,
0.59130859375
],
[
0.2548828125,
0.5869140625,
0.268798828125,
0.5771484375,
0.281005859375,
0.5859375,
0.26708984375,
0.5966796875
],
[
0.1217041015625,
0.59326171875,
0.40283203125,
0.5888671875,
0.40283203125,
0.61572265625,
0.1217041015625,
0.619140625
],
[
0.443603515625,
0.59326171875,
0.476318359375,
0.59326171875,
0.476318359375,
0.609375,
0.443603515625,
0.609375
],
[
0.5498046875,
0.59326171875,
0.72802734375,
0.59326171875,
0.72802734375,
0.60693359375,
0.5498046875,
0.60693359375
],
[
0.2548828125,
0.611328125,
0.281982421875,
0.611328125,
0.281982421875,
0.62255859375,
0.2548828125,
0.62255859375
],
[
0.54833984375,
0.60693359375,
0.82666015625,
0.60693359375,
0.82666015625,
0.6220703125,
0.54833984375,
0.6220703125
],
[
0.5498046875,
0.6220703125,
0.8857421875,
0.6220703125,
0.8857421875,
0.63525390625,
0.5498046875,
0.63525390625
],
[
0.08660888671875,
0.630859375,
0.473876953125,
0.630859375,
0.473876953125,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.54736328125,
0.63623046875,
0.8857421875,
0.63720703125,
0.8857421875,
0.6513671875,
0.54736328125,
0.65087890625
],
[
0.0882568359375,
0.64794921875,
0.473876953125,
0.64794921875,
0.473876953125,
0.6611328125,
0.0882568359375,
0.6611328125
],
[
0.53369140625,
0.65380859375,
0.576171875,
0.65380859375,
0.576171875,
0.66552734375,
0.53369140625,
0.66552734375
],
[
0.0906982421875,
0.6640625,
0.47216796875,
0.6640625,
0.47216796875,
0.6748046875,
0.0906982421875,
0.6748046875
],
[
0.53173828125,
0.66650390625,
0.736328125,
0.66748046875,
0.736328125,
0.681640625,
0.53173828125,
0.68115234375
],
[
0.08905029296875,
0.67822265625,
0.473876953125,
0.67822265625,
0.473876953125,
0.69140625,
0.08905029296875,
0.69140625
],
[
0.53173828125,
0.68310546875,
0.83251953125,
0.68310546875,
0.83251953125,
0.69775390625,
0.53173828125,
0.69775390625
],
[
0.08905029296875,
0.693359375,
0.473876953125,
0.693359375,
0.473876953125,
0.70703125,
0.08905029296875,
0.70703125
],
[
0.515625,
0.69775390625,
0.5703125,
0.69775390625,
0.5703125,
0.7119140625,
0.515625,
0.7119140625
],
[
0.0882568359375,
0.70849609375,
0.47216796875,
0.70849609375,
0.47216796875,
0.7216796875,
0.0882568359375,
0.7216796875
],
[
0.08905029296875,
0.7236328125,
0.1519775390625,
0.7236328125,
0.1519775390625,
0.73486328125,
0.08905029296875,
0.73486328125
],
[
0.5,
0.7421875,
0.88134765625,
0.74365234375,
0.88134765625,
0.75830078125,
0.5,
0.7568359375
],
[
0.0882568359375,
0.75244140625,
0.47216796875,
0.75244140625,
0.47216796875,
0.7666015625,
0.0882568359375,
0.7666015625
],
[
0.51318359375,
0.76123046875,
0.765625,
0.76123046875,
0.765625,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.0882568359375,
0.767578125,
0.47216796875,
0.767578125,
0.47216796875,
0.7822265625,
0.0882568359375,
0.7822265625
],
[
0.51318359375,
0.7744140625,
0.708984375,
0.77783203125,
0.708984375,
0.798828125,
0.51318359375,
0.79541015625
],
[
0.0882568359375,
0.7841796875,
0.1993408203125,
0.7841796875,
0.1993408203125,
0.79736328125,
0.0882568359375,
0.79736328125
],
[
0.515625,
0.79541015625,
0.65673828125,
0.79541015625,
0.65673828125,
0.81005859375,
0.515625,
0.81005859375
],
[
0.10455322265625,
0.806640625,
0.238525390625,
0.80615234375,
0.238525390625,
0.82080078125,
0.10455322265625,
0.8212890625
],
[
0.53369140625,
0.81005859375,
0.69775390625,
0.8095703125,
0.69775390625,
0.82373046875,
0.53369140625,
0.82470703125
],
[
0.1748046875,
0.82568359375,
0.19287109375,
0.82568359375,
0.19287109375,
0.8388671875,
0.1748046875,
0.8388671875
],
[
0.53173828125,
0.82470703125,
0.8857421875,
0.82373046875,
0.8857421875,
0.83837890625,
0.53173828125,
0.8388671875
],
[
0.1405029296875,
0.83056640625,
0.455078125,
0.82373046875,
0.455810546875,
0.85986328125,
0.1422119140625,
0.8671875
],
[
0.1544189453125,
0.83349609375,
0.1666259765625,
0.83349609375,
0.1666259765625,
0.84326171875,
0.1544189453125,
0.84326171875
],
[
0.108642578125,
0.84130859375,
0.1405029296875,
0.84130859375,
0.1405029296875,
0.85302734375,
0.108642578125,
0.85302734375
],
[
0.53173828125,
0.84033203125,
0.79736328125,
0.8388671875,
0.79736328125,
0.85400390625,
0.53173828125,
0.85498046875
],
[
0.176513671875,
0.85791015625,
0.1968994140625,
0.85791015625,
0.1968994140625,
0.86572265625,
0.176513671875,
0.86572265625
],
[
0.53369140625,
0.85595703125,
0.87158203125,
0.85595703125,
0.87158203125,
0.87060546875,
0.53369140625,
0.87060546875
],
[
0.1405029296875,
0.87060546875,
0.455810546875,
0.87060546875,
0.455810546875,
0.88818359375,
0.1405029296875,
0.88818359375
],
[
0.53125,
0.869140625,
0.78759765625,
0.87060546875,
0.78759765625,
0.88525390625,
0.53125,
0.8837890625
],
[
0.443603515625,
0.88623046875,
0.4755859375,
0.88623046875,
0.4755859375,
0.9013671875,
0.443603515625,
0.9013671875
],
[
0.51708984375,
0.88720703125,
0.56884765625,
0.88720703125,
0.56884765625,
0.8984375,
0.51708984375,
0.8984375
],
[
0.48193359375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93505859375,
0.48193359375,
0.93505859375
]
] |
[
"Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference",
"we propose the following sampling procedure,",
"The learnableknowledgeprior with parameter Qi also func-",
"20 ~ Pa2(z|Cs),",
"tions akin to a multi-label classifier. We use both gold anno-",
"tations and posterior knowledge selection as ground-truth",
"zt+1=zt+$V,log pe(z|R,Cs)+V28et,",
"(21)",
"labels.mirroring the updating of the initializer.",
"where t = 1.",
", T, and the initial state for the Markov",
"Generation We can get a response by greedy search as",
"chain is sampled from the context-conditioned prior distri-",
"summarized in Algorithm 2. Given dialogue context C ",
"bution. The total length of the Markov chain is rather small",
"(H,K), we can first select the knowledge candidate with the",
"(e.g. T = 5). Further theoretical underpinnings of this",
"highest logit from the initializerf(C) or learnable prior",
"approximate sampling and learning method can be found in",
"fa (C) based on different prior choices,",
"Appendix A.",
"s = arg max f(Ci), ie [1,.., S).",
"(25)",
"2.4. Algorithms",
"s = argmax fa, (C), ie (1,..., M).",
"(26)",
"variations in the learning and generation algorithms.",
"The choice of prior for the discrete variable s leads to minor",
"Then we use sample mean of the response latent variable",
"Learning with Uniform Knowledge PriorGiven learn-",
"Pa2 (z|Cs) ~ N(fa2 (Cs), I) to generate dialogue,",
"ing iterations T = 1,..., TL, the generative model with",
"R ~ pg(R|z = fa2(C), Cs).",
"(27)",
"parameters = [α = α2, β] can be updated through",
"0+1=0+m140,",
"Algorithm 1 Learning with Sequential Posterior Inference",
"N",
"input: Observed examples (Cn,R\", toal training",
"1.",
"",
"CEpe- (s\",zn|Rn,Cn)[Ve log Pa (z\"]Cm)",
"epochs TL, learning rate n, number of candidates S in",
"A0=",
"N",
"=",
"knowledge selection initializer, number of Langevin steps",
"+Vlogpe(Rz\",Cn)",
"(22)",
"T, step size , initial weights o, %o.",
"output: Updated weights OT, T.",
"The additional top-S initializer f(C) with parameters ",
"for T = 1 to T do",
"can be viewed as a multi-label classifier or multiple binary",
"1. Draw observed examples fCn, Rn]",
"classifiers and be updated by cross-entropy loss,",
"2. Sequential Posterior Inference",
"M",
"2.1Posterior Knowledge Selection",
"LcE(y, C) =",
">yi log f(C) + (1 - yi) log(1 - f~(C))",
"If uniform prior",
"(a) Select knowledge candidates by top-S initializer.",
"7",
"(b)Infersfromthetop-Scandidatesusing(14)",
"",
"Z",
"++1 =%- m2Cce(g\", C\"),",
"(23)",
"If learnable prior",
"n=1",
"Infer s from the learnable prior using (17)",
"2.2Posterior Inference of Response Latent Variable",
"where (ynN- denotes labels which can be obtained by",
"Infer z by T-step short-run MCMC (21) with step",
"posteriorknowledge selection and/orannotations (if weuse",
"size d.",
"bothposteriorknowledgeselectionandannotations,itis",
"3.Update Model Parameters",
"possible that two of yi's equal to 1). Therefore, the learned",
"Update and according to (22),(23) or (24)",
"initializer can output top-S candidates that likely include",
"end for",
"the gold knowledge or the one from posterior knowledge",
"selection.",
"Algorithm2Knowledge-Grounded ResponseGeneration",
"Learning with Learnable Knowledge PriorThe genera",
"input: Observed examples [Cn]n-]",
"tive model with parameters = [α = (αi, Q2), β) can be",
"output: Response [R\")N].",
"updated through",
"for n = 1 to Ntest do",
"0++1=0++10",
"1. Draw test example Cn",
"N",
"2. Select s using the initializer f(Cn) or learnable",
" ()",
"1",
"△0:",
"prior fa(C) according to (25) or (26).",
"7-1",
"3. Set z as the sample mean of the prior pa2 (zCn).",
"+Ve log pa2(z\"|Cn) + Vg log pe(R/z\",Cn).",
"4. Generate Rn by decoder using (27).",
"(24)",
"end for",
"5"
] |
{
"arxiv_id": "2306.01153",
"language": "en",
"page_idx": 4
}
|
||
1811.03532-en-1
|
1811.03532-en-1.jpg
|
straints, we apply robust optimization to kidney exchange—a graph-based market clearing or resource allocation problem.
Our Contributions. To our knowledge, weight uncertainty has not been addressed in the kidney exchange literature. Our approach is similar to that of Bertsimas and Sim (2004) and Poss (2014), and uses some of their results. Several approaches have been proposed for existence uncertainty, primarily based on stochastic optimization (Dickerson et al. 2016; Anderson et al. 2015; Dickerson et al. 2018) or hierarchical optimization (Manlove and O'Malley 2015). The primary disadvantage of these approaches—in addition to tractability—is their reliance on, and sensitivity to, the explicit estimation of the probability of each particular potential transplant. This probability is extremely difficult to determine (Dickerson et al. 2018; Glorie 2012), and prevents the translation of those methods into practice. Our approach uses a simpler notion of edge existence uncertainty—an upperbound on the number of non-existent edges—which is easier to interpret and estimate. Glorie (2014) proposed a related robust formulation that is exponentially larger than ours, and is intractable for realistically-sized exchanges.
In addition, we introduce a new scalable formulation for kidney exchange that combines concepts from two state-ofthe-art formulations (Anderson et al.2015; Dickerson $et$ $al.2016)$, handles long or uncapped NDD-initiated chains without requiring expensive constraint generation, and ties into a developed literature on fairness in kidney exchange— thus addressing use cases that are becoming more common in fielded exchanges (Anderson et al. 2015).
# 2 Preliminaries
Model for kidney exchange. A kidney exchange can be represented formally by a directed compatibility graph $G=$ $(V,E).$ Here, vertices $v\in V$ represent participants in the exchange, and are partitioned as $\hat{V}=P\hat{\cup}N$ into $P$,the set of all patient-donor pairs, and $N$, the set of all NDDs (Roth $et$ $al. 2004; 2005;$ Abraham $et$ $al. 2007) .$ Each potential transplant from a donor at vertex $u$ to a patient at vertex $v$ is represented by a directed edge $e=(\hat{u},v)\in E$,which has an associated weight $w_e\in \textbf{w; weights are set by policy- }$ makers, and reflect both the medical utility of the transplant, as well as ethical considerations (e.g., prioritizing patients by waiting time, age, and so on). Cycles in $G$ correspond to cyclic trades between multiple patient-donor pairs in $P;$ chains, correspond to donations that begin with an NDD in $N$ and continue through multiple patient-donor pairs in $P.$ The kidney exchange clearingproblem is to select a feasible set of transplants (edges in $E$) that maximize overall weight. Let $\mathcal{M}$ be the set of all feasible matchings(i.e., solutions) to a kidney exchange problem; the general formulation of this problem is max$_\mathbf{x}\in\mathcal{M}\mathbf{x}\cdot\mathbf{w}$,where binary decision variables x represent edges, or cycles and chains. This problem is NP- and APX-hard (Abraham et al. 2007; $\text{Bir\' {o}etal. 2009) . }$
Robust optimization. Robust optimization is a common approach to optimization under uncertainty, which is often more tractable and requires less accurate uncertainty information
than other approaches (Bertsimas et al. 2011a). This approach begins by defining an uncertainty set U for the uncertain optimization parameter; $\mathcal{U}$ contains different realizations of this parameter. Consider the example of edge weight uncertainty: we might design an edge weight uncertainty set $\mathcal{U}_w$ that contains the realized (i.e.“true”) edge weights $\hat{\mathbf{w}}$ with $\operatorname{high}$ probability, $P(\hat{\mathbf{w}}\in\mathcal{U}_{w})\geq1-\epsilon,\operatorname{for}0<\epsilon\ll1.$ The parameter $\epsilon$ is referred to as the protection level, and is often used to control the number of realizations in $\mathcal{U}.$
After designing $\mathcal{U}$, the robust approach fnds the best solution, assuming the worst-case realization within $\mathcal{U}.$ For kidney exchange (a maximization problem), this corresponds to a minimization over $\mathcal{U};$ for example, Problem (1) is the robust formulation with uncertain edge weights.
(1)
$$\max_{\mathbf{x}\in\mathcal{M}}\min_{\hat{\mathbf{w}}\in\mathcal{U}}\quad\mathbf{x}\cdot\hat{\mathbf{w}}$$
The robustness of this approach depends on the proportion of possible realizations contained in $\mathcal{U}.$ If $\hat{U}$ contains all possible realizations, the approach may be too conservative; if $\mathcal{U}$ only contains one possible realization of $\hat{\mathbf{w}}$, the solution may be too myopic. The number of realizations in $U$ is often controlled by a parameter: either an uncertainty budget $\Gamma$, or the protection level $\epsilon.$ Next we introduce the first type of uncertainty considered in this paper: edge weight uncertainty.
## 3 $\textbf{Optimization in the Presence of Edge}$ Weight Uncertainty
Edge weights in kidney exchange represent the medical and social utility gained by a single kidney transplant. Weights are determined by policymakers, and are subject to several types of uncertainty.$^{\mathrm{l}}$ Part of this uncertainty is due to insufficient knowledge of the future: a patient or donor's health may change, raising or lowering the“true”weight of their transplant edges. Another type of uncertainty stems from disagreement between policymakers regarding the social utility of a transplant. For example, some policymakers might prioritize young patients over older patients; other policymakers might prioritize the sickest patients above all ${\mathrm{healthier~patients.~Policymakers~aggregate~these~value~judg-}}$ ments to assign a single weight to each transplant edge, but this aggregation is a contentious and imperfect process (although recent work from the AI community has begun to address this using techniques from computational social choice and machine learning (Freedman et al.2018; Noothigattu $et\textit{al. 2018}) ) .$ Still, there is no way to measure the“true” social utility of a transplant, and therefore this uncertainty is not easily measured.
Interval weight uncertainty. It is beyond the scope of this work to characterize these sources of uncertainty. We simply assume that the nominal edge weights w, provided by policymakers, are an uncertain estimate of the realized edge weights $\mathbf{\hat{w}}$,i.e., the“true”value of each transplant. Next, we formalize edge weight uncertainty and our robust approach. This section focuses on edge weights, so we write our for-
The process used to set weights by the UNOS US-wide kidney
exchange is published publicly (UNOS 2015).
|
[
[
0.08660888671875,
0.0694580078125,
0.464111328125,
0.0694580078125,
0.464111328125,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.521484375,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.521484375,
0.08331298828125
],
[
0.08660888671875,
0.08331298828125,
0.4794921875,
0.08331298828125,
0.4794921875,
0.09661865234375,
0.08660888671875,
0.09661865234375
],
[
0.51953125,
0.083984375,
0.91259765625,
0.083984375,
0.91259765625,
0.09783935546875,
0.51953125,
0.09783935546875
],
[
0.51953125,
0.09783935546875,
0.9140625,
0.09783935546875,
0.9140625,
0.111083984375,
0.51953125,
0.111083984375
],
[
0.08660888671875,
0.10479736328125,
0.476318359375,
0.10479736328125,
0.476318359375,
0.11871337890625,
0.08660888671875,
0.11871337890625
],
[
0.5185546875,
0.111083984375,
0.91259765625,
0.11236572265625,
0.91259765625,
0.1256103515625,
0.5185546875,
0.125
],
[
0.0882568359375,
0.12060546875,
0.47802734375,
0.12060546875,
0.47802734375,
0.1307373046875,
0.0882568359375,
0.1307373046875
],
[
0.51953125,
0.1256103515625,
0.9111328125,
0.1256103515625,
0.9111328125,
0.1395263671875,
0.51953125,
0.1395263671875
],
[
0.08660888671875,
0.1331787109375,
0.4794921875,
0.1331787109375,
0.4794921875,
0.146484375,
0.08660888671875,
0.146484375
],
[
0.51953125,
0.1395263671875,
0.91259765625,
0.1395263671875,
0.91259765625,
0.15283203125,
0.51953125,
0.15283203125
],
[
0.08660888671875,
0.146484375,
0.47802734375,
0.146484375,
0.47802734375,
0.160400390625,
0.08660888671875,
0.160400390625
],
[
0.51953125,
0.15283203125,
0.91259765625,
0.15283203125,
0.91259765625,
0.1666259765625,
0.51953125,
0.1666259765625
],
[
0.08660888671875,
0.160400390625,
0.4794921875,
0.160400390625,
0.4794921875,
0.173583984375,
0.08660888671875,
0.173583984375
],
[
0.5185546875,
0.1673583984375,
0.9140625,
0.1666259765625,
0.9140625,
0.179931640625,
0.5185546875,
0.18115234375
],
[
0.0882568359375,
0.175537109375,
0.47802734375,
0.175537109375,
0.47802734375,
0.186279296875,
0.0882568359375,
0.186279296875
],
[
0.51953125,
0.18115234375,
0.830078125,
0.18115234375,
0.830078125,
0.1944580078125,
0.51953125,
0.1944580078125
],
[
0.08660888671875,
0.1875,
0.4794921875,
0.1875,
0.4794921875,
0.2008056640625,
0.08660888671875,
0.2008056640625
],
[
0.5390625,
0.1956787109375,
0.9140625,
0.1956787109375,
0.9140625,
0.208984375,
0.5390625,
0.208984375
],
[
0.0849609375,
0.2020263671875,
0.47802734375,
0.2020263671875,
0.47802734375,
0.21533203125,
0.0849609375,
0.21533203125
],
[
0.51953125,
0.208984375,
0.9140625,
0.208984375,
0.9140625,
0.222900390625,
0.51953125,
0.222900390625
],
[
0.08660888671875,
0.216552734375,
0.4794921875,
0.216552734375,
0.4794921875,
0.2298583984375,
0.08660888671875,
0.2298583984375
],
[
0.51953125,
0.2215576171875,
0.91259765625,
0.222900390625,
0.91259765625,
0.236083984375,
0.51953125,
0.23486328125
],
[
0.08660888671875,
0.2310791015625,
0.47802734375,
0.2310791015625,
0.47802734375,
0.2412109375,
0.08660888671875,
0.2412109375
],
[
0.51953125,
0.2374267578125,
0.91259765625,
0.2374267578125,
0.91259765625,
0.250732421875,
0.51953125,
0.250732421875
],
[
0.08660888671875,
0.24365234375,
0.47802734375,
0.24365234375,
0.47802734375,
0.2568359375,
0.08660888671875,
0.2568359375
],
[
0.51953125,
0.250732421875,
0.8349609375,
0.250732421875,
0.8349609375,
0.264404296875,
0.51953125,
0.264404296875
],
[
0.08660888671875,
0.2587890625,
0.47802734375,
0.2587890625,
0.47802734375,
0.26953125,
0.08660888671875,
0.26953125
],
[
0.0849609375,
0.271484375,
0.4794921875,
0.271484375,
0.4794921875,
0.285400390625,
0.0849609375,
0.285400390625
],
[
0.73388671875,
0.2734375,
0.775390625,
0.2734375,
0.775390625,
0.28857421875,
0.73388671875,
0.28857421875
],
[
0.8896484375,
0.271484375,
0.9150390625,
0.271484375,
0.9150390625,
0.2890625,
0.8896484375,
0.2890625
],
[
0.0849609375,
0.285400390625,
0.4794921875,
0.285400390625,
0.4794921875,
0.298583984375,
0.0849609375,
0.298583984375
],
[
0.65869140625,
0.27587890625,
0.7236328125,
0.27587890625,
0.7236328125,
0.287353515625,
0.65869140625,
0.287353515625
],
[
0.65673828125,
0.285888671875,
0.7236328125,
0.285888671875,
0.7236328125,
0.296630859375,
0.65673828125,
0.296630859375
],
[
0.08660888671875,
0.2998046875,
0.404296875,
0.2998046875,
0.404296875,
0.313232421875,
0.08660888671875,
0.313232421875
],
[
0.40283203125,
0.300537109375,
0.47802734375,
0.300537109375,
0.47802734375,
0.3125,
0.40283203125,
0.3125
],
[
0.5361328125,
0.302978515625,
0.91259765625,
0.3037109375,
0.91259765625,
0.318115234375,
0.5361328125,
0.317626953125
],
[
0.08660888671875,
0.314453125,
0.47802734375,
0.314453125,
0.47802734375,
0.324462890625,
0.08660888671875,
0.324462890625
],
[
0.521484375,
0.318115234375,
0.91259765625,
0.318115234375,
0.91259765625,
0.33203125,
0.521484375,
0.33203125
],
[
0.0849609375,
0.326904296875,
0.47802734375,
0.326904296875,
0.47802734375,
0.34033203125,
0.0849609375,
0.34033203125
],
[
0.521484375,
0.333251953125,
0.91259765625,
0.333251953125,
0.91259765625,
0.3466796875,
0.521484375,
0.3466796875
],
[
0.0849609375,
0.34033203125,
0.4794921875,
0.34033203125,
0.4794921875,
0.354248046875,
0.0849609375,
0.354248046875
],
[
0.521484375,
0.347900390625,
0.9111328125,
0.347900390625,
0.9111328125,
0.35791015625,
0.521484375,
0.35791015625
],
[
0.08660888671875,
0.35595703125,
0.390625,
0.35595703125,
0.390625,
0.3662109375,
0.08660888671875,
0.3662109375
],
[
0.51953125,
0.35986328125,
0.91259765625,
0.35986328125,
0.91259765625,
0.373779296875,
0.51953125,
0.373779296875
],
[
0.10296630859375,
0.36865234375,
0.4794921875,
0.36865234375,
0.4794921875,
0.3818359375,
0.10296630859375,
0.3818359375
],
[
0.51953125,
0.373779296875,
0.9140625,
0.373779296875,
0.9140625,
0.386962890625,
0.51953125,
0.386962890625
],
[
0.0882568359375,
0.3837890625,
0.47802734375,
0.3837890625,
0.47802734375,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.521484375,
0.389404296875,
0.9111328125,
0.389404296875,
0.9111328125,
0.399658203125,
0.521484375,
0.399658203125
],
[
0.0882568359375,
0.397705078125,
0.47802734375,
0.397705078125,
0.47802734375,
0.40771484375,
0.0882568359375,
0.40771484375
],
[
0.51953125,
0.401611328125,
0.90869140625,
0.401611328125,
0.90869140625,
0.415283203125,
0.51953125,
0.415283203125
],
[
0.0849609375,
0.410400390625,
0.47802734375,
0.410400390625,
0.47802734375,
0.423583984375,
0.0849609375,
0.423583984375
],
[
0.0849609375,
0.4248046875,
0.47802734375,
0.4248046875,
0.47802734375,
0.438232421875,
0.0849609375,
0.438232421875
],
[
0.5458984375,
0.431884765625,
0.57275390625,
0.431884765625,
0.57275390625,
0.443115234375,
0.5458984375,
0.443115234375
],
[
0.56640625,
0.43115234375,
0.8857421875,
0.43115234375,
0.8857421875,
0.44580078125,
0.56640625,
0.44580078125
],
[
0.0849609375,
0.438232421875,
0.471435546875,
0.438232421875,
0.471435546875,
0.451904296875,
0.0849609375,
0.451904296875
],
[
0.0849609375,
0.451904296875,
0.4794921875,
0.451904296875,
0.4794921875,
0.46533203125,
0.0849609375,
0.46533203125
],
[
0.646484375,
0.447509765625,
0.81396484375,
0.449462890625,
0.81396484375,
0.464111328125,
0.646484375,
0.462158203125
],
[
0.08416748046875,
0.46533203125,
0.375732421875,
0.46533203125,
0.375732421875,
0.479248046875,
0.08416748046875,
0.479248046875
],
[
0.521484375,
0.468505859375,
0.91259765625,
0.468505859375,
0.91259765625,
0.481689453125,
0.521484375,
0.481689453125
],
[
0.51953125,
0.48291015625,
0.91259765625,
0.48291015625,
0.91259765625,
0.496826171875,
0.51953125,
0.496826171875
],
[
0.2091064453125,
0.494384765625,
0.35546875,
0.494384765625,
0.35546875,
0.5087890625,
0.2091064453125,
0.5087890625
],
[
0.51953125,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.51123046875,
0.51953125,
0.51123046875
],
[
0.51953125,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.51953125,
0.5234375
],
[
0.08660888671875,
0.521484375,
0.47802734375,
0.521484375,
0.47802734375,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.51953125,
0.5244140625,
0.91259765625,
0.5244140625,
0.91259765625,
0.53857421875,
0.51953125,
0.53857421875
],
[
0.0849609375,
0.5361328125,
0.476318359375,
0.5361328125,
0.476318359375,
0.5498046875,
0.0849609375,
0.5498046875
],
[
0.51953125,
0.53857421875,
0.91259765625,
0.53857421875,
0.91259765625,
0.55322265625,
0.51953125,
0.55322265625
],
[
0.0849609375,
0.5498046875,
0.47802734375,
0.5498046875,
0.47802734375,
0.56298828125,
0.0849609375,
0.56298828125
],
[
0.521484375,
0.5537109375,
0.9111328125,
0.5537109375,
0.9111328125,
0.564453125,
0.521484375,
0.564453125
],
[
0.08416748046875,
0.56298828125,
0.4794921875,
0.5625,
0.4794921875,
0.5771484375,
0.08416748046875,
0.57763671875
],
[
0.521484375,
0.56494140625,
0.9140625,
0.56494140625,
0.9140625,
0.5791015625,
0.521484375,
0.5791015625
],
[
0.08660888671875,
0.5791015625,
0.476318359375,
0.5791015625,
0.476318359375,
0.5888671875,
0.08660888671875,
0.5888671875
],
[
0.51953125,
0.580078125,
0.91259765625,
0.580078125,
0.91259765625,
0.59326171875,
0.51953125,
0.59326171875
],
[
0.0849609375,
0.59130859375,
0.4794921875,
0.59130859375,
0.4794921875,
0.60498046875,
0.0849609375,
0.60498046875
],
[
0.52294921875,
0.59521484375,
0.9111328125,
0.59521484375,
0.9111328125,
0.60595703125,
0.52294921875,
0.60595703125
],
[
0.0882568359375,
0.60595703125,
0.4794921875,
0.60595703125,
0.4794921875,
0.619140625,
0.0882568359375,
0.619140625
],
[
0.521484375,
0.60791015625,
0.91259765625,
0.60791015625,
0.91259765625,
0.6220703125,
0.521484375,
0.6220703125
],
[
0.08660888671875,
0.61865234375,
0.47802734375,
0.61865234375,
0.47802734375,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.51953125,
0.6220703125,
0.9140625,
0.6220703125,
0.9140625,
0.63623046875,
0.51953125,
0.63623046875
],
[
0.0849609375,
0.63330078125,
0.47802734375,
0.63330078125,
0.47802734375,
0.646484375,
0.0849609375,
0.646484375
],
[
0.521484375,
0.63525390625,
0.9140625,
0.63525390625,
0.9140625,
0.6484375,
0.521484375,
0.6484375
],
[
0.0849609375,
0.646484375,
0.4794921875,
0.646484375,
0.4794921875,
0.66015625,
0.0849609375,
0.66015625
],
[
0.51953125,
0.6484375,
0.9140625,
0.6484375,
0.9140625,
0.662109375,
0.51953125,
0.662109375
],
[
0.0849609375,
0.6591796875,
0.47802734375,
0.6611328125,
0.47802734375,
0.6748046875,
0.0849609375,
0.67236328125
],
[
0.52294921875,
0.6640625,
0.9111328125,
0.6640625,
0.9111328125,
0.6748046875,
0.52294921875,
0.6748046875
],
[
0.08660888671875,
0.6748046875,
0.47802734375,
0.6748046875,
0.47802734375,
0.68798828125,
0.08660888671875,
0.68798828125
],
[
0.521484375,
0.6767578125,
0.91259765625,
0.6767578125,
0.91259765625,
0.68994140625,
0.521484375,
0.68994140625
],
[
0.0849609375,
0.68798828125,
0.4794921875,
0.68798828125,
0.4794921875,
0.7021484375,
0.0849609375,
0.7021484375
],
[
0.51953125,
0.68994140625,
0.9140625,
0.68994140625,
0.9140625,
0.7041015625,
0.51953125,
0.7041015625
],
[
0.08660888671875,
0.7021484375,
0.4794921875,
0.7021484375,
0.4794921875,
0.71533203125,
0.08660888671875,
0.71533203125
],
[
0.521484375,
0.70556640625,
0.9111328125,
0.70556640625,
0.9111328125,
0.71630859375,
0.521484375,
0.71630859375
],
[
0.08660888671875,
0.71630859375,
0.4794921875,
0.71630859375,
0.4794921875,
0.72998046875,
0.08660888671875,
0.72998046875
],
[
0.51953125,
0.71826171875,
0.91259765625,
0.71826171875,
0.91259765625,
0.73193359375,
0.51953125,
0.73193359375
],
[
0.08660888671875,
0.72998046875,
0.4794921875,
0.72998046875,
0.4794921875,
0.74365234375,
0.08660888671875,
0.74365234375
],
[
0.51953125,
0.73291015625,
0.74853515625,
0.73291015625,
0.74853515625,
0.7470703125,
0.51953125,
0.7470703125
],
[
0.0849609375,
0.74365234375,
0.47802734375,
0.74365234375,
0.47802734375,
0.7568359375,
0.0849609375,
0.7568359375
],
[
0.521484375,
0.75244140625,
0.91259765625,
0.75244140625,
0.91259765625,
0.7666015625,
0.521484375,
0.7666015625
],
[
0.08660888671875,
0.7587890625,
0.47802734375,
0.7587890625,
0.47802734375,
0.76953125,
0.08660888671875,
0.76953125
],
[
0.0849609375,
0.77099609375,
0.47802734375,
0.77099609375,
0.47802734375,
0.7841796875,
0.0849609375,
0.7841796875
],
[
0.521484375,
0.767578125,
0.9140625,
0.767578125,
0.9140625,
0.78076171875,
0.521484375,
0.78076171875
],
[
0.0849609375,
0.7841796875,
0.47802734375,
0.7841796875,
0.47802734375,
0.79736328125,
0.0849609375,
0.79736328125
],
[
0.521484375,
0.78076171875,
0.9111328125,
0.78076171875,
0.9111328125,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.0849609375,
0.798828125,
0.4794921875,
0.798828125,
0.4794921875,
0.8125,
0.0849609375,
0.8125
],
[
0.51953125,
0.79541015625,
0.9111328125,
0.79541015625,
0.9111328125,
0.8095703125,
0.51953125,
0.8095703125
],
[
0.08660888671875,
0.8125,
0.4794921875,
0.8125,
0.4794921875,
0.82568359375,
0.08660888671875,
0.82568359375
],
[
0.51953125,
0.8095703125,
0.9140625,
0.8095703125,
0.9140625,
0.82275390625,
0.51953125,
0.82275390625
],
[
0.08660888671875,
0.8271484375,
0.1993408203125,
0.8271484375,
0.1993408203125,
0.84033203125,
0.08660888671875,
0.84033203125
],
[
0.51953125,
0.82373046875,
0.9140625,
0.82373046875,
0.9140625,
0.8369140625,
0.51953125,
0.8369140625
],
[
0.521484375,
0.8369140625,
0.9140625,
0.8369140625,
0.9140625,
0.85107421875,
0.521484375,
0.85107421875
],
[
0.0849609375,
0.8466796875,
0.47802734375,
0.84765625,
0.47802734375,
0.8623046875,
0.0849609375,
0.861328125
],
[
0.0849609375,
0.8623046875,
0.4794921875,
0.861328125,
0.4794921875,
0.875,
0.0849609375,
0.87548828125
],
[
0.5400390625,
0.861328125,
0.91259765625,
0.86279296875,
0.91259765625,
0.87744140625,
0.5400390625,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.4794921875,
0.87548828125,
0.4794921875,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51708984375,
0.87548828125,
0.79345703125,
0.875,
0.79345703125,
0.8896484375,
0.51708984375,
0.89013671875
]
] |
[
"straints, we apply robust optimization to kidney exchange",
"than other approaches (Bertsimas et al. 201la). This approach",
"graph-based market clearing or resource allocation problem",
"begins by defining an uncertainty set U for the uncertain op-",
"timization parameter; U contains different realizations of",
"Our Contributions. To our knowledge, weight uncertainty",
"this parameter.Consider the example of edge weight uncer-",
"has not been addressed in the kidney exchange literature",
"tainty: we might design an edge weight uncertainty set Uw",
"Our approach is similar to that of Bertsimas and Sim (2004)",
"that contains the realized (i.e. \"true\") edge weights w with",
"and Poss (2014), and uses some of their results. Several",
"high probability, P(w E Uw) ≥ 1 - e, for 0 < e < 1. The",
"approaches have been proposed for existence uncertainty,",
"parameter eis referred to astheprotection level,and is often",
"primarily based on stochastic optimization (Dickerson et al.",
"used to control the number of realizations in U.",
"2016; Anderson et al. 2015; Dickerson et al. 2018) or hier-",
"After designing U, the robust approach finds the best so-",
"archical optimization (Manlove and O'Malley 2015). The",
"lution, assuming the worst-case realization within U. For",
"primary disadvantage of these approaches-in addition to",
"kidney exchange (a maximization problem),this corresponds",
"tractability-is their reliance on.and sensitivity to.the ex-",
"to a minimization over U; for example, Problem (1) is the",
"plicit estimation of the probability of each particular potential",
"robust formulation with uncertain edge weights.",
"transplant.This probabilityis extremely difficult to deter-",
"mine (Dickerson et al. 2018; Glorie 2012), and prevents the",
"X.W",
"(1)",
"translation of those methods into practice. Our approach uses",
"max min",
"xEMwEu",
"a simpler notion of edge existence uncertainty-",
"-an upper",
"The robustness of this approach depends on the proportion",
"boundonthenumberofnon-existentedges-whichiseasier",
"of possible realizations contained in u. If u contains all",
"to interpret and estimate. Glorie (2014) proposed a related",
"possible realizations, the approach may be too conservative;",
"robust formulation that is exponentially larger than ours, and",
"if U onlycontains onepossiblerealization ofw,thesolution",
"isintractableforrealistically-sizedexchanges.",
"may be too myopic. The number of realizations in U is often",
"In addition, we introduce a new scalable formulation for",
"controlled by a parameter: either an uncertainty budget T,",
"kidneyexchangethat combines conceptsfromtwostate-of",
"ortheprotectionlevel e.Nextweintroducethefirsttypeof",
"the-art formulations (Anderson et al. 2015:Dickerson et",
"uncertainty considered in this paper: edge weight uncertainty.",
"al. 2016), handles long or uncapped NDD-initiated chains",
"without requiring expensive constraint generation, and ties",
"3",
"Optimization in the Presence of Edge",
"into a developed literature on fairness in kidney exchange-",
"thus addressing use cases that are becoming more common",
"Weight Uncertainty",
"in fielded exchanges (Anderson et al. 2015).",
"Edge weights in kidney exchange represent the medical and",
"social utility gained by a single kidney transplant. Weights",
"2Preliminaries",
"are determined by policymakers, and are subject to sev-",
"eral types of uncertainty.1 Part of this uncertainty is due",
"Model for kidney exchange. A kidney exchange can be rep-",
"to insufficient knowledge of the future: a patient or donor's",
"resented formally by a directed compatibility graph G -",
"health may change, raising or lowering the \"true\" weight",
"(V, E). Here, vertices V represent participants in the",
"of theirtransplantedges.Anothertype ofuncertaintystems",
"exchange, and are partitioned as V = PU N into P, the set",
"from disagreement between policymakers regarding the so-",
"ofall patient-donorpairs,and N,thesetofallNDDs(Roth",
"cial utility of a transplant. For example, some policymakers",
"et al. 2004; 2005; Abraham et al. 2007).Each potential trans-",
"might prioritize young patients over older patients: other",
"plant from a donor at vertex u to a patient at vertex u is",
"policymakers might prioritize the sickest patients above all",
"represented by a directed edge e = (u, v) E E, which has",
"healthier patients. Policymakers aggregate these value judg-",
"an associated weight we E w; weights are set by policy",
"ments to assign a single weight to each transplant edge,",
"makers, and reflect both the medical utility of the transplant,",
"but this aggregation is a contentious and imperfect process",
"as well as ethical considerations (e.g., prioritizing patients",
"(althoughrecentworkfromtheAIcommunityhasbegun",
"by waiting time, age, and so on). Cycles in G correspond",
"to address this using techniques from computational so-",
"to cyclic trades between multiple patient-donor pairs in P;",
"cial choice and machine learning (Freedman et al. 2018;",
"chains, correspond to donations that begin with an NDD in",
"Noothigattu et al.2018)).Still, there is no way tomeasure",
"N and continue through multiple patient-donor pairs in P",
"the \"true\" social utility of a transplant, and therefore this",
"The kidney exchange clearing problem is to select a feasi-",
"uncertainty is not easily measured.",
"ble set of transplants (edges in E) that maximize overall",
"Interval weight uncertainty. It is beyond the scope of this",
"weight. Let M be the set of all feasible matchings (i.e..",
" solutions) to a kidney exchange problem; the general for",
"work to characterize these sources of uncertainty. We sim-",
"mulation of this problem is maxxeM x -w, where binary",
"ply assume that the nominal edge weights w, provided by",
"decision variables x represent edges, or cycles and chains.",
"policymakers, are an uncertain estimate of the realized edge",
"This problem is NP- and APX-hard (Abraham et al. 2007;",
"weights w, i.e., the \"true\" value of each transplant. Next, we",
"Biro et al. 2009).",
"formalize edge weight uncertainty and our robust approach.",
"This section focuses on edge weights, so we write our for-",
"Robust optimization. Robust optimization is a common ap-",
"proachto optimization underuncertainty,whichis often more",
"'The process used to set weights by the UNOS US-wide kidney",
"tractable and requires less accurate uncertainty information",
"exchange is published publicly (UNOS 2015)."
] |
{
"arxiv_id": "1811.03532",
"language": "en",
"page_idx": 1
}
|
||
1811.03532-en-4
|
1811.03532-en-4.jpg
|
This uncertainty set is less intuitive than $\mathcal{U}_\Gamma^{ex}$, but more suited to the robust approach. In Appendix B we show that $\mathcal{U}_\Gamma^w$ and $\mathcal{U}_{\Gamma}^{ex}$ are equivalent for integer $\Gamma$,and thus can be used for our robust approach.
# 4.1 Robust Optimization Approach
In this section we briefly describe our robust approach; for a full discussion and derivation, please see Appendix B. Our robust formulation for uncertainty set $\mathcal{U}_\Gamma^w$ follows a similar approach to Section 3. First, we directly minimize the kidney exchange objective over $\mathcal{U}_\Gamma^w$, for some feasible solution $\mathbf{x}\in\mathcal{M}.$ We express this minimization as a function $Z(\mathbf{x}):$ in effect, $Z(\mathbf{x})$ discounts the $\Gamma$ largest-weight cycles and chains. We then linearize $Z(\mathbf{x})$ using several variables and constraints—this requires a formulation with variables tracking individual total chain weights—which is not possible in any existing compact kidney exchange formulations. For this purpose, we introduce a new kidney exchange formulation.
The PI-TSP formulation. We propose the position-indexed TSP formulation (PI-TSP); for details, please see Appendix B. Our formulation combines innovations from the two leading kidney exchange clearing approaches: PICEF (Dickerson et al. 2016) and PC-TSP (Anderson et al.2015). PICEF introduced an indexing schema that enables a more compact formulation in the context of long chains; our formulation builds on this to allow tracking of individual chain weights, a necessity that PICEF could not do. PC-TSP builds on techniques from the prize-collecting travelling salesperson problem (Balas 1989) to provide a tight linear programming relaxation; in general, the PC-TSP formulation has exponentially many constraints and thus requires constraint generation to solve. Our formulation uses an efficient version of position indexing that also requires only $O(|E|)+O(|V|\cdot|N|)$ constraints. Unlike PICEF, our formulation does not grow with the chain cap $L{:}$ PICEF uses $O(|V|^3)$ variables (when $L\to|V|);$for large graphs, the PICEF model becomes too large to fit into memory (Dickerson et al.2016). Our formulation uses a fxed number of variables—$O(|V|^{2})$—for any chain cap, alleviating this memory problem. This is especially relevant to existing exchanges, as long chains can significantly increase efficiency in kidney exchange (Ashlagi et al. 2012). PI-TSP uses the following parameters:
$\bullet$ $G{: }$kidney exchange graph,
$\bullet C$:a set of cycles on exchange graph $G$,
$\bullet$ $L$: chain cap (maximum number of edges used in a chain),
$\bullet$ $w_e{: }$ edge weights for each edge $e\in E$, $\bullet w_{c}^{C}:$cycle weights for each cycle $c\in C$, and the following decision variables:
$\bullet~p_{e}\geq1{:}$the position of edge $e$ in any chain,
$\bullet$ $p_v\geq 1{: }$the position of patient-donor vertex $v$ in any chain, $\bullet$ $\hat{p} _{e}\geq 0$: equal to $p_e$ if $e$ is used in a chain, and 0 otherwise, $\bullet$ $z_{c}\in \{ 0, 1\} \nobreak {: } 1$ if cycle $c$ is used in the matching, and 0
otherwise,
$\bullet$ $y_{e}\in \{ 0, 1\} {: }1$ if edge $e$ is used in a chain, and 0 otherwise,
$\bullet$ $y_e^n\in \{ 0, 1\} : 1$ if edge $e$ is used in a chain starting with
NDD $n$, and 0 otherwise,
$\bullet w_n^N:$total weight of the chain starting with NDD $n$, $\bullet$ $f_{v}^{i}$ and $f_v^o{:}$ chain flow into and out of vertex $v\in P$, $f_v^{i,n}$ and $f_v^i,n:$ chain flow into and out of vertex $v\in P$,
from the chain starting with NDD $n\in N.$
The PI-TSP formulation with chain cap $L$ is given in Problem 4. We use the notation $\delta^-(v)$ for the set of edges into vertex $v$ and $\delta^+(v)$ for the set of edges out of $v.$
<table>
<tbody>
<tr>
<th>$\sum$ max $n\in N$</th>
<th> </th>
<th>(4a)</th>
</tr>
<tr>
<td>s.t.</td>
<td>$n\in N$</td>
<td>$(4b)$</td>
</tr>
<tr>
<td> </td>
<td>$e\in E$</td>
<td>(4c)</td>
</tr>
<tr>
<td> </td>
<td>$v\in V$</td>
<td>(4d)</td>
</tr>
<tr>
<td> </td>
<td>$v\in V$</td>
<td>(4e)</td>
</tr>
<tr>
<td> </td>
<td>$v\in V,n\in N$</td>
<td>(4f)</td>
</tr>
<tr>
<td>$\hat{e}\in$</td>
<td>$v\in V,n\in N$</td>
<td>(4g)</td>
</tr>
<tr>
<td>$e\in i$ $\sum$ + $z_c\leq$ 全</td>
<td>$v\in P$</td>
<td>(4h)</td>
</tr>
<tr>
<td>$\bullet U$ 一</td>
<td>$v\in N$</td>
<td>(4i)</td>
</tr>
<tr>
<td> </td>
<td>$e\in\delta^{+}(N)$</td>
<td>(4j)</td>
</tr>
<tr>
<td> </td>
<td>$e\in E$</td>
<td>$(4k)$</td>
</tr>
<tr>
<td>$F$</td>
<td>$v\in P$</td>
<td>(41)</td>
</tr>
<tr>
<td> </td>
<td>$v\in P.e\in\delta^{\top}$</td>
<td>$(v)$ 1 1 1</td>
</tr>
<tr>
<td> </td>
<td>$n\in N$</td>
<td>$(4m)$ (4n)</td>
</tr>
<tr>
<td> </td>
<td>$v\in V,n\in N$</td>
<td>(40)</td>
</tr>
<tr>
<td> </td>
<td>$e\in E$</td>
<td>(4p)</td>
</tr>
<tr>
<td> </td>
<td>$c\in C$</td>
<td>$(4q)$</td>
</tr>
<tr>
<td> </td>
<td>$\Lambda$ TA</td>
<td>$(4r^{\circ}$</td>
</tr>
</tbody>
</table>
The ability to express individual chain weights as decision variables has applications beyond robustness. For particularly valuable NDDs (such as those with so-called“universal donor" blood-type O), exchanges may enforce a $min-$ imum chain length or chain weight, to ensure that these rare NDDs are not “used up”on short chains; such a policy was formerly used by the Ûnited Network for Organ Sharing (Dickerson et al. 2012), using a much less scalable form of optimization—that also does not consider uncertainty— than our approach (Abraham et al. 2007). Such a policy can be implemented efficiently with PI-TSP, inefficiently with PC-TSP, and not with PICEF, where decision variables do not indicate from which NDD a chain originated. In Appendix B we show-using real kidney exchange data-that PI-TSP can enforce a minimum chain length, and that this restriction has almost no impact on overall matching score.
## 5 Experimental Results
In this section, we compare each robust formulation against the leading non-robust formulation, PICEF (Dickerson $et$
|
[
[
0.08660888671875,
0.0694580078125,
0.4794921875,
0.0694580078125,
0.4794921875,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.52685546875,
0.07452392578125,
0.5390625,
0.07452392578125,
0.5390625,
0.080810546875,
0.52685546875,
0.080810546875
],
[
0.53369140625,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.53369140625,
0.08331298828125
],
[
0.0849609375,
0.083984375,
0.4794921875,
0.083984375,
0.4794921875,
0.09783935546875,
0.0849609375,
0.09783935546875
],
[
0.53369140625,
0.08209228515625,
0.705078125,
0.08331298828125,
0.705078125,
0.09783935546875,
0.53369140625,
0.09661865234375
],
[
0.0849609375,
0.09783935546875,
0.4794921875,
0.09783935546875,
0.4794921875,
0.111083984375,
0.0849609375,
0.111083984375
],
[
0.51953125,
0.0997314453125,
0.87353515625,
0.10101318359375,
0.87353515625,
0.11614990234375,
0.51953125,
0.11553955078125
],
[
0.0849609375,
0.11236572265625,
0.1968994140625,
0.11236572265625,
0.1968994140625,
0.1256103515625,
0.0849609375,
0.1256103515625
],
[
0.521484375,
0.121826171875,
0.87353515625,
0.121826171875,
0.87353515625,
0.1363525390625,
0.521484375,
0.1363525390625
],
[
0.0849609375,
0.1395263671875,
0.36767578125,
0.1395263671875,
0.36767578125,
0.154052734375,
0.0849609375,
0.154052734375
],
[
0.5185546875,
0.1395263671875,
0.9140625,
0.1383056640625,
0.9140625,
0.15283203125,
0.5185546875,
0.154052734375
],
[
0.08416748046875,
0.1578369140625,
0.4794921875,
0.1590576171875,
0.4794921875,
0.173583984375,
0.08416748046875,
0.17236328125
],
[
0.5361328125,
0.154052734375,
0.81396484375,
0.154052734375,
0.81396484375,
0.1673583984375,
0.5361328125,
0.1673583984375
],
[
0.0849609375,
0.173583984375,
0.4794921875,
0.173583984375,
0.4794921875,
0.1875,
0.0849609375,
0.1875
],
[
0.5185546875,
0.169189453125,
0.9140625,
0.17041015625,
0.9140625,
0.1849365234375,
0.5185546875,
0.184326171875
],
[
0.08660888671875,
0.189453125,
0.47802734375,
0.189453125,
0.47802734375,
0.199462890625,
0.08660888671875,
0.199462890625
],
[
0.51953125,
0.1849365234375,
0.9140625,
0.1849365234375,
0.9140625,
0.1988525390625,
0.51953125,
0.1988525390625
],
[
0.0849609375,
0.2020263671875,
0.4794921875,
0.2020263671875,
0.4794921875,
0.21533203125,
0.0849609375,
0.21533203125
],
[
0.5185546875,
0.2008056640625,
0.7890625,
0.2008056640625,
0.7890625,
0.2139892578125,
0.5185546875,
0.2139892578125
],
[
0.0849609375,
0.21533203125,
0.4794921875,
0.21533203125,
0.4794921875,
0.2291259765625,
0.0849609375,
0.2291259765625
],
[
0.67333984375,
0.2203369140625,
0.765625,
0.2203369140625,
0.765625,
0.238037109375,
0.67333984375,
0.238037109375
],
[
0.89208984375,
0.222900390625,
0.9150390625,
0.222900390625,
0.9150390625,
0.236083984375,
0.89208984375,
0.236083984375
],
[
0.65625,
0.2259521484375,
0.6748046875,
0.2259521484375,
0.6748046875,
0.2342529296875,
0.65625,
0.2342529296875
],
[
0.62353515625,
0.2259521484375,
0.64208984375,
0.2259521484375,
0.64208984375,
0.2342529296875,
0.62353515625,
0.2342529296875
],
[
0.67333984375,
0.2266845703125,
0.6787109375,
0.2266845703125,
0.6787109375,
0.231689453125,
0.67333984375,
0.231689453125
],
[
0.65185546875,
0.23486328125,
0.6748046875,
0.23486328125,
0.6748046875,
0.242431640625,
0.65185546875,
0.242431640625
],
[
0.08416748046875,
0.2279052734375,
0.48046875,
0.2298583984375,
0.48046875,
0.24365234375,
0.08416748046875,
0.2412109375
],
[
0.7099609375,
0.23486328125,
0.73193359375,
0.23486328125,
0.73193359375,
0.242431640625,
0.7099609375,
0.242431640625
],
[
0.0849609375,
0.24365234375,
0.4794921875,
0.24365234375,
0.4794921875,
0.2568359375,
0.0849609375,
0.2568359375
],
[
0.67626953125,
0.24560546875,
0.765625,
0.24560546875,
0.765625,
0.26318359375,
0.67626953125,
0.26318359375
],
[
0.810546875,
0.248779296875,
0.8486328125,
0.248779296875,
0.8486328125,
0.260009765625,
0.810546875,
0.260009765625
],
[
0.890625,
0.2474365234375,
0.9150390625,
0.2474365234375,
0.9150390625,
0.261474609375,
0.890625,
0.261474609375
],
[
0.0849609375,
0.2568359375,
0.48046875,
0.2568359375,
0.48046875,
0.270751953125,
0.0849609375,
0.270751953125
],
[
0.654296875,
0.25,
0.66845703125,
0.25,
0.66845703125,
0.2587890625,
0.654296875,
0.2587890625
],
[
0.67724609375,
0.2587890625,
0.69775390625,
0.2587890625,
0.69775390625,
0.26953125,
0.67724609375,
0.26953125
],
[
0.685546875,
0.26318359375,
0.77392578125,
0.26318359375,
0.77392578125,
0.302978515625,
0.685546875,
0.302978515625
],
[
0.08660888671875,
0.272705078125,
0.1650390625,
0.272705078125,
0.1650390625,
0.28271484375,
0.08660888671875,
0.28271484375
],
[
0.1683349609375,
0.272705078125,
0.47802734375,
0.272705078125,
0.47802734375,
0.28271484375,
0.1683349609375,
0.28271484375
],
[
0.8095703125,
0.272705078125,
0.84619140625,
0.272705078125,
0.84619140625,
0.2841796875,
0.8095703125,
0.2841796875
],
[
0.89208984375,
0.271484375,
0.9150390625,
0.271484375,
0.9150390625,
0.285400390625,
0.89208984375,
0.285400390625
],
[
0.0849609375,
0.2841796875,
0.4794921875,
0.2841796875,
0.4794921875,
0.29736328125,
0.0849609375,
0.29736328125
],
[
0.69140625,
0.294189453125,
0.7099609375,
0.294189453125,
0.7099609375,
0.311279296875,
0.69140625,
0.311279296875
],
[
0.72119140625,
0.29541015625,
0.76708984375,
0.29541015625,
0.76708984375,
0.31005859375,
0.72119140625,
0.31005859375
],
[
0.8095703125,
0.296630859375,
0.84716796875,
0.296630859375,
0.84716796875,
0.31005859375,
0.8095703125,
0.31005859375
],
[
0.89208984375,
0.296630859375,
0.9150390625,
0.296630859375,
0.9150390625,
0.309326171875,
0.89208984375,
0.309326171875
],
[
0.6787109375,
0.309326171875,
0.7236328125,
0.309326171875,
0.7236328125,
0.32080078125,
0.6787109375,
0.32080078125
],
[
0.0849609375,
0.298583984375,
0.4794921875,
0.298583984375,
0.4794921875,
0.3125,
0.0849609375,
0.3125
],
[
0.08660888671875,
0.314453125,
0.47216796875,
0.314453125,
0.47216796875,
0.324462890625,
0.08660888671875,
0.324462890625
],
[
0.69140625,
0.3212890625,
0.7099609375,
0.3212890625,
0.7099609375,
0.3359375,
0.69140625,
0.3359375
],
[
0.71728515625,
0.322509765625,
0.76708984375,
0.322509765625,
0.76708984375,
0.337158203125,
0.71728515625,
0.337158203125
],
[
0.810546875,
0.323974609375,
0.84619140625,
0.323974609375,
0.84619140625,
0.3359375,
0.810546875,
0.3359375
],
[
0.89208984375,
0.323974609375,
0.9150390625,
0.323974609375,
0.9150390625,
0.337158203125,
0.89208984375,
0.337158203125
],
[
0.08416748046875,
0.33203125,
0.4794921875,
0.333251953125,
0.4794921875,
0.347900390625,
0.08416748046875,
0.3466796875
],
[
0.67724609375,
0.3359375,
0.7255859375,
0.3359375,
0.7255859375,
0.347900390625,
0.67724609375,
0.347900390625
],
[
0.08660888671875,
0.3466796875,
0.47802734375,
0.3466796875,
0.47802734375,
0.35986328125,
0.08660888671875,
0.35986328125
],
[
0.6787109375,
0.348388671875,
0.69921875,
0.348388671875,
0.69921875,
0.363037109375,
0.6787109375,
0.363037109375
],
[
0.705078125,
0.349853515625,
0.7646484375,
0.347900390625,
0.765625,
0.3623046875,
0.7060546875,
0.3642578125
],
[
0.810546875,
0.3515625,
0.8837890625,
0.3515625,
0.8837890625,
0.3623046875,
0.810546875,
0.3623046875
],
[
0.890625,
0.3515625,
0.9140625,
0.3515625,
0.9140625,
0.363037109375,
0.890625,
0.363037109375
],
[
0.0849609375,
0.361083984375,
0.4794921875,
0.361083984375,
0.4794921875,
0.375,
0.0849609375,
0.375
],
[
0.66650390625,
0.363037109375,
0.71142578125,
0.363037109375,
0.71142578125,
0.375,
0.66650390625,
0.375
],
[
0.08416748046875,
0.375,
0.4794921875,
0.373779296875,
0.4794921875,
0.38818359375,
0.08416748046875,
0.389404296875
],
[
0.67724609375,
0.37744140625,
0.69775390625,
0.37744140625,
0.69775390625,
0.39013671875,
0.67724609375,
0.39013671875
],
[
0.70361328125,
0.376953125,
0.7646484375,
0.375,
0.765625,
0.39013671875,
0.705078125,
0.392578125
],
[
0.810546875,
0.37890625,
0.88232421875,
0.37890625,
0.88232421875,
0.389404296875,
0.810546875,
0.389404296875
],
[
0.890625,
0.37890625,
0.9140625,
0.37890625,
0.9140625,
0.39013671875,
0.890625,
0.39013671875
],
[
0.0849609375,
0.38818359375,
0.4794921875,
0.38818359375,
0.4794921875,
0.401611328125,
0.0849609375,
0.401611328125
],
[
0.6640625,
0.39013671875,
0.7099609375,
0.39013671875,
0.7099609375,
0.40087890625,
0.6640625,
0.40087890625
],
[
0.08660888671875,
0.404052734375,
0.47802734375,
0.404052734375,
0.47802734375,
0.4140625,
0.08660888671875,
0.4140625
],
[
0.609375,
0.404052734375,
0.68701171875,
0.404052734375,
0.68701171875,
0.41845703125,
0.609375,
0.41845703125
],
[
0.69775390625,
0.40283203125,
0.71728515625,
0.40283203125,
0.71728515625,
0.416748046875,
0.69775390625,
0.416748046875
],
[
0.810546875,
0.40478515625,
0.84619140625,
0.40478515625,
0.84619140625,
0.417236328125,
0.810546875,
0.417236328125
],
[
0.89208984375,
0.40478515625,
0.9150390625,
0.40478515625,
0.9150390625,
0.41845703125,
0.89208984375,
0.41845703125
],
[
0.5888671875,
0.406005859375,
0.6044921875,
0.406005859375,
0.6044921875,
0.4140625,
0.5888671875,
0.4140625
],
[
0.5458984375,
0.406005859375,
0.57666015625,
0.406005859375,
0.57666015625,
0.417236328125,
0.5458984375,
0.417236328125
],
[
0.7294921875,
0.406005859375,
0.765625,
0.406005859375,
0.765625,
0.417236328125,
0.7294921875,
0.417236328125
],
[
0.08660888671875,
0.417236328125,
0.48046875,
0.417236328125,
0.48046875,
0.43115234375,
0.08660888671875,
0.43115234375
],
[
0.57275390625,
0.417236328125,
0.61669921875,
0.417236328125,
0.61669921875,
0.4248046875,
0.57275390625,
0.4248046875
],
[
0.68701171875,
0.417236328125,
0.7294921875,
0.417236328125,
0.7294921875,
0.4248046875,
0.68701171875,
0.4248046875
],
[
0.72607421875,
0.4267578125,
0.7685546875,
0.4267578125,
0.7685546875,
0.442626953125,
0.72607421875,
0.442626953125
],
[
0.80810546875,
0.4267578125,
0.8486328125,
0.4267578125,
0.8486328125,
0.44140625,
0.80810546875,
0.44140625
],
[
0.8935546875,
0.427978515625,
0.9150390625,
0.427978515625,
0.9150390625,
0.44140625,
0.8935546875,
0.44140625
],
[
0.80615234375,
0.43994140625,
0.87158203125,
0.442626953125,
0.87158203125,
0.45703125,
0.80615234375,
0.455078125
],
[
0.08660888671875,
0.43115234375,
0.4794921875,
0.43115234375,
0.4794921875,
0.4443359375,
0.08660888671875,
0.4443359375
],
[
0.08416748046875,
0.4443359375,
0.47802734375,
0.443115234375,
0.47802734375,
0.45703125,
0.08416748046875,
0.458251953125
],
[
0.7294921875,
0.4443359375,
0.76708984375,
0.4443359375,
0.76708984375,
0.455810546875,
0.7294921875,
0.455810546875
],
[
0.8935546875,
0.443115234375,
0.9150390625,
0.443115234375,
0.9150390625,
0.45703125,
0.8935546875,
0.45703125
],
[
0.08416748046875,
0.45703125,
0.4794921875,
0.458251953125,
0.4794921875,
0.472900390625,
0.08416748046875,
0.4716796875
],
[
0.8095703125,
0.45703125,
0.84716796875,
0.45703125,
0.84716796875,
0.47021484375,
0.8095703125,
0.47021484375
],
[
0.890625,
0.45703125,
0.9150390625,
0.45703125,
0.9150390625,
0.4697265625,
0.890625,
0.4697265625
],
[
0.71826171875,
0.47021484375,
0.740234375,
0.47021484375,
0.740234375,
0.484130859375,
0.71826171875,
0.484130859375
],
[
0.708984375,
0.458251953125,
0.765625,
0.458251953125,
0.765625,
0.4697265625,
0.708984375,
0.4697265625
],
[
0.8095703125,
0.47021484375,
0.84716796875,
0.47021484375,
0.84716796875,
0.484130859375,
0.8095703125,
0.484130859375
],
[
0.8935546875,
0.4716796875,
0.9150390625,
0.4716796875,
0.9150390625,
0.48486328125,
0.8935546875,
0.48486328125
],
[
0.67626953125,
0.472900390625,
0.7060546875,
0.472900390625,
0.7060546875,
0.484130859375,
0.67626953125,
0.484130859375
],
[
0.08660888671875,
0.472900390625,
0.4794921875,
0.472900390625,
0.4794921875,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.75244140625,
0.473388671875,
0.7646484375,
0.473388671875,
0.7646484375,
0.48291015625,
0.75244140625,
0.48291015625
],
[
0.0849609375,
0.48486328125,
0.47802734375,
0.486083984375,
0.47802734375,
0.5,
0.0849609375,
0.498779296875
],
[
0.70751953125,
0.484130859375,
0.75244140625,
0.484130859375,
0.75244140625,
0.494384765625,
0.70751953125,
0.494384765625
],
[
0.701171875,
0.5,
0.7646484375,
0.496826171875,
0.765625,
0.51123046875,
0.70166015625,
0.5146484375
],
[
0.810546875,
0.498779296875,
0.90283203125,
0.498779296875,
0.90283203125,
0.51025390625,
0.810546875,
0.51025390625
],
[
0.890625,
0.51025390625,
0.91259765625,
0.51025390625,
0.91259765625,
0.51806640625,
0.890625,
0.51806640625
],
[
0.08660888671875,
0.501953125,
0.4794921875,
0.501953125,
0.4794921875,
0.51220703125,
0.08660888671875,
0.51220703125
],
[
0.0849609375,
0.5146484375,
0.4794921875,
0.5146484375,
0.4794921875,
0.52783203125,
0.0849609375,
0.52783203125
],
[
0.70361328125,
0.52099609375,
0.76708984375,
0.52294921875,
0.765625,
0.53857421875,
0.70166015625,
0.5361328125
],
[
0.810546875,
0.5234375,
0.84716796875,
0.5234375,
0.84716796875,
0.53515625,
0.810546875,
0.53515625
],
[
0.89208984375,
0.52294921875,
0.9150390625,
0.52294921875,
0.9150390625,
0.5361328125,
0.89208984375,
0.5361328125
],
[
0.08660888671875,
0.52880859375,
0.476318359375,
0.52880859375,
0.476318359375,
0.5390625,
0.08660888671875,
0.5390625
],
[
0.70166015625,
0.5322265625,
0.7255859375,
0.5322265625,
0.7255859375,
0.54443359375,
0.70166015625,
0.54443359375
],
[
0.0849609375,
0.54052734375,
0.4794921875,
0.54052734375,
0.4794921875,
0.5546875,
0.0849609375,
0.5546875
],
[
0.67724609375,
0.546875,
0.76708984375,
0.54443359375,
0.76708984375,
0.5595703125,
0.67724609375,
0.56103515625
],
[
0.810546875,
0.548828125,
0.8837890625,
0.548828125,
0.8837890625,
0.5595703125,
0.810546875,
0.5595703125
],
[
0.88623046875,
0.548828125,
0.9140625,
0.548828125,
0.9140625,
0.56005859375,
0.88623046875,
0.56005859375
],
[
0.7060546875,
0.56103515625,
0.76708984375,
0.56103515625,
0.76708984375,
0.57275390625,
0.7060546875,
0.57275390625
],
[
0.08660888671875,
0.5537109375,
0.47802734375,
0.5537109375,
0.47802734375,
0.5673828125,
0.08660888671875,
0.5673828125
],
[
0.8095703125,
0.56103515625,
0.84619140625,
0.56103515625,
0.84619140625,
0.57275390625,
0.8095703125,
0.57275390625
],
[
0.890625,
0.56103515625,
0.9150390625,
0.56103515625,
0.9150390625,
0.57373046875,
0.890625,
0.57373046875
],
[
0.0882568359375,
0.568359375,
0.4794921875,
0.568359375,
0.4794921875,
0.58203125,
0.0882568359375,
0.58203125
],
[
0.7060546875,
0.57275390625,
0.7685546875,
0.57275390625,
0.7685546875,
0.5869140625,
0.7060546875,
0.5869140625
],
[
0.80615234375,
0.5712890625,
0.84619140625,
0.5712890625,
0.84619140625,
0.5869140625,
0.80615234375,
0.5869140625
],
[
0.890625,
0.57373046875,
0.9150390625,
0.57373046875,
0.9150390625,
0.5859375,
0.890625,
0.5859375
],
[
0.0849609375,
0.58349609375,
0.48046875,
0.58349609375,
0.48046875,
0.5966796875,
0.0849609375,
0.5966796875
],
[
0.705078125,
0.58837890625,
0.76708984375,
0.58837890625,
0.76708984375,
0.599609375,
0.705078125,
0.599609375
],
[
0.810546875,
0.5869140625,
0.8857421875,
0.5869140625,
0.8857421875,
0.5986328125,
0.810546875,
0.5986328125
],
[
0.8935546875,
0.58837890625,
0.9140625,
0.58837890625,
0.9140625,
0.5986328125,
0.8935546875,
0.5986328125
],
[
0.0849609375,
0.5966796875,
0.4794921875,
0.5966796875,
0.4794921875,
0.60986328125,
0.0849609375,
0.60986328125
],
[
0.0849609375,
0.60986328125,
0.4794921875,
0.60986328125,
0.4794921875,
0.62353515625,
0.0849609375,
0.62353515625
],
[
0.5361328125,
0.60693359375,
0.9140625,
0.60693359375,
0.9140625,
0.62060546875,
0.5361328125,
0.62060546875
],
[
0.0849609375,
0.625,
0.4794921875,
0.625,
0.4794921875,
0.63818359375,
0.0849609375,
0.63818359375
],
[
0.51953125,
0.6220703125,
0.9150390625,
0.6220703125,
0.9150390625,
0.63525390625,
0.51953125,
0.63525390625
],
[
0.08416748046875,
0.63818359375,
0.4794921875,
0.63818359375,
0.4794921875,
0.6513671875,
0.08416748046875,
0.6513671875
],
[
0.521484375,
0.63720703125,
0.91259765625,
0.63720703125,
0.91259765625,
0.64794921875,
0.521484375,
0.64794921875
],
[
0.08416748046875,
0.65087890625,
0.423095703125,
0.6513671875,
0.423095703125,
0.66650390625,
0.08416748046875,
0.66552734375
],
[
0.51953125,
0.6484375,
0.9140625,
0.6484375,
0.9140625,
0.662109375,
0.51953125,
0.662109375
],
[
0.51953125,
0.66357421875,
0.9140625,
0.66357421875,
0.9140625,
0.6767578125,
0.51953125,
0.6767578125
],
[
0.0849609375,
0.67041015625,
0.281005859375,
0.671875,
0.281005859375,
0.68603515625,
0.0849609375,
0.68505859375
],
[
0.51953125,
0.6767578125,
0.91259765625,
0.6767578125,
0.91259765625,
0.68994140625,
0.51953125,
0.68994140625
],
[
0.0849609375,
0.689453125,
0.365234375,
0.689453125,
0.365234375,
0.7041015625,
0.0849609375,
0.7041015625
],
[
0.51953125,
0.68994140625,
0.9140625,
0.68994140625,
0.9140625,
0.7041015625,
0.51953125,
0.7041015625
],
[
0.08660888671875,
0.70556640625,
0.47802734375,
0.70556640625,
0.47802734375,
0.7197265625,
0.08660888671875,
0.7197265625
],
[
0.51953125,
0.705078125,
0.9140625,
0.705078125,
0.9140625,
0.71826171875,
0.51953125,
0.71826171875
],
[
0.08416748046875,
0.72265625,
0.35693359375,
0.72265625,
0.35693359375,
0.736328125,
0.08416748046875,
0.736328125
],
[
0.51953125,
0.71826171875,
0.89892578125,
0.71826171875,
0.89892578125,
0.73193359375,
0.51953125,
0.73193359375
],
[
0.5185546875,
0.73193359375,
0.9140625,
0.73095703125,
0.9140625,
0.74560546875,
0.5185546875,
0.7470703125
],
[
0.0849609375,
0.74072265625,
0.369384765625,
0.74072265625,
0.369384765625,
0.75390625,
0.0849609375,
0.75390625
],
[
0.521484375,
0.74755859375,
0.91259765625,
0.74755859375,
0.91259765625,
0.75830078125,
0.521484375,
0.75830078125
],
[
0.08416748046875,
0.7587890625,
0.330078125,
0.76025390625,
0.330078125,
0.7744140625,
0.08416748046875,
0.7734375
],
[
0.51953125,
0.76025390625,
0.91259765625,
0.76025390625,
0.91259765625,
0.7734375,
0.51953125,
0.7734375
],
[
0.51953125,
0.7734375,
0.91259765625,
0.7734375,
0.91259765625,
0.787109375,
0.51953125,
0.787109375
],
[
0.08416748046875,
0.78076171875,
0.3857421875,
0.77978515625,
0.3857421875,
0.79443359375,
0.08416748046875,
0.79541015625
],
[
0.5185546875,
0.78857421875,
0.9140625,
0.78857421875,
0.9140625,
0.8017578125,
0.5185546875,
0.8017578125
],
[
0.08660888671875,
0.80029296875,
0.4794921875,
0.80029296875,
0.4794921875,
0.81494140625,
0.08660888671875,
0.81494140625
],
[
0.5185546875,
0.8017578125,
0.91259765625,
0.8017578125,
0.91259765625,
0.81494140625,
0.5185546875,
0.81494140625
],
[
0.5185546875,
0.81494140625,
0.810546875,
0.81494140625,
0.810546875,
0.8291015625,
0.5185546875,
0.8291015625
],
[
0.08660888671875,
0.8212890625,
0.4794921875,
0.8212890625,
0.4794921875,
0.83544921875,
0.08660888671875,
0.83544921875
],
[
0.08660888671875,
0.84130859375,
0.47802734375,
0.84130859375,
0.47802734375,
0.85498046875,
0.08660888671875,
0.85498046875
],
[
0.609375,
0.84228515625,
0.82177734375,
0.84228515625,
0.82177734375,
0.85693359375,
0.609375,
0.85693359375
],
[
0.10296630859375,
0.85693359375,
0.1732177734375,
0.85693359375,
0.1732177734375,
0.86865234375,
0.10296630859375,
0.86865234375
],
[
0.51953125,
0.8623046875,
0.9140625,
0.8623046875,
0.9140625,
0.87548828125,
0.51953125,
0.87548828125
],
[
0.08660888671875,
0.87548828125,
0.48046875,
0.87548828125,
0.48046875,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.5185546875,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.5185546875,
0.8896484375
]
] |
[
"This uncertainty set is less intuitive than Ue, but more suited",
"",
"yn E f0, 1]: 1 if edge e is used in a chain starting with",
"to the robust approach. In Appendix B we show that U and",
"NDD n, and 0 otherwise,",
"Ue are equivalent for integer I, and thus can be used for our",
". wN: total weight of the chain starting with NDD n,",
"robust approach.",
": fu and fo: chain flow into and out of vertex u E P,",
"4.1 Robust Optimization Approach",
": fi,n and fin: chain flow into and out of vertex w P,",
"In this section we briefly describe our robust approach; for a",
"from the chain starting with NDD n. E N.",
"full discussion and derivation, please see Appendix B. Our",
"The PI-TSP formulation with chain cap L is given in Problem",
"robustformulationforuncertaintysetUwfollowsasimilar",
"4. We use the notation - (v) for the set of edges into vertex",
"approach to Section 3. First, we directly minimize the kid-",
" and &+ (u) for the set of edges out of v.",
"ney exchange objective over U, for some feasible solution",
"wN+wOz",
"(4a) ",
">",
"max",
"",
"nEN",
"x E M.Weexpress this minimization asafunction Z(x):",
"CEC",
"in effect, Z(x) discounts the I largest-weight cycles and",
"Mweve=wn",
"nEN",
"(4b)",
"chains. We then linearize Z(x) using several variables and",
"s.t.",
"eEE",
"",
"constraints-",
"-thisrequiresaformulationwithvariablestrack-",
"eEE",
"(4e)",
"ing individual total chain weightswhich is not possible in",
"Z",
"ye=ft",
"UEV",
"(4d)",
"eE8-(u)",
"any existing compact kidney exchange formulations. For this",
"purpose,weintroduceanewkidneyexchangeformulation.",
"M",
"ye=fo",
"uEv",
"(4e)",
"The PI-TSP formulation. We propose the position-indexed",
"ees+(v)",
"TSP formulation (PI-TSP): for details, please see Appendix B",
"",
"ye=f.n",
"UEVnEN",
"(4)",
"Our formulation combines innovations from the two leading",
"eEs-(u)",
"kidney exchange clearing approaches: PICEF (Dickerson",
"M",
"we=fo.n",
"uEV,nEN",
"(4g)",
"et al. 2016) and PC-TSP (Anderson et al. 2015). PICEF",
"eEs+(0)",
"introduced an indexing schema that enables a more com-",
"zfu+",
"M",
"UEP",
"(4h)",
"T",
"fo+",
"2e≤1",
"pact formulation in the context of long chains; our formu-",
"CEC:VE",
"CEC:VEC",
"fo≤1",
"UEN",
"(4i)",
"cE&+(N)",
"lation builds on this to allow tracking of individual chain",
"weights, a necessity that PICEF could not do. PC-TSP builds",
"Pe.= 1",
"(4i)",
"on techniques from the prize-collecting travelling salesperson",
"eEE",
"(4k)",
"n",
"Pe=peye",
"UEP",
"(41)",
"Pu=",
"problem (Balas 1989) to provide a tight linear programming",
"pe",
"relaxation; in general, the PC-TSPformulation has exponen",
"eES-(U)",
"Pe=pu+1",
"vEP,eE&+(u)",
"(4m)",
"tially many constraints andthus requires constraintgenera-",
"tion to solve. Our formulation uses an efficient version of",
"Zu<L",
"nEN",
"(4n)",
"positionindexingthatalsorequiresonlyO(ED+O(VN",
"eEE",
"constraints. Unlike PICEF, our formulation does not grow",
"fonyi.w<1",
"UEVnEN",
"(40)",
"Ye E [0,1]",
"with the chain cap L: PICEF uses O(/VJ3) variables (when",
"eEE",
"(4p)",
"L → [VD; for large graphs, the PICEF model becomes too",
"ze E (0, 1]",
"CEC",
"(4g)",
"large to fit into memory (Dickerson et al. 2016). Our for-",
"ee [0,1]",
"eEE,nEN",
"(4r)",
"mulation uses a fixed number of variables-O(/V/2)-for",
" any chain cap, alleviating this memory problem. This is es-",
"The ability to express individual chain weights as decision",
"pecially relevant to existing exchanges, as long chains can",
"variables has applications beyond robustness. For particu-",
"significantly increase efficiency in kidney exchange (Ashlagi",
"larly valuableNDDs(suchasthosewith so-called\"univer-",
"et al.2012).PI-TSPuses the following parameters:",
"sal donor\" blood-type O), exchanges may enforce a min-",
"imum chain length or chain weight, to ensure that these",
". G: kidney exchange graph,",
"rare NDDs are not “used up on short chains; such a policy",
". C: a set of cycles on exchange graph G,",
"was formerly used by the United Network for Organ Shar-",
". L: chain cap (maximum number of edges used in a chain)",
"ing (Dickerson et al. 2012), using a much less scalable form",
" We: edge weights for each edge e E E,",
"of optimization-that also does not consider uncertainty",
"than our approach (Abraham et al. 2007). Such a policy can",
". w: cycle weights for each cycle c E C,",
"beimplementedefficientlywithPI-TSP,inefficientlywith",
"and the following decision variables:",
"PC-TSP, and not with PICEF, where decision variables do not",
"indicate from which NDD a chain originated. In Appendix B",
". pe ≥ l: the position of edge e in any chain,",
"we show-using real kidney exchange data-that PI-TSP can",
". pu ≥ 1: the position of patient-donor vertex u in any chain,",
"enforce a minimum chain length, and that this restriction has",
"almost no impact on overall matching score.",
". pe ≥ O: equal to pe if e is used in a chain, and 0 otherwise,",
". ze E {0, 1]: 1 if cycle c is used in the matching, and 0",
"5Experimental Results",
"otherwise,",
"In this section, we compare each robust formulation against",
". Ye E {0, 1]: l if edge e is used in a chain, and 0 otherwise,",
"the leading non-robust formulation, PICEF (Dickerson et"
] |
{
"arxiv_id": "1811.03532",
"language": "en",
"page_idx": 4
}
|
||
1602.07776-en-0
|
1602.07776-en-0.jpg
|
Recurrent Neural Network Grammars
Chris Dyer[INDEX] Adhiguna Kuncoro[INDEX] Miguel Ballesteros[INDEX] Noah A.Smith$^{\circ}$ $^{\bullet}$School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA
$^{\textcircled{1}}$NLP Group, Pompeu Fabra University, Barcelona, Spain
$^{\textcircled{\circ}}$Computer Science & Engineering, University of Washington, Seattle, WA, USA
$\{$cdyer,akuncoro$\}$@cs.cmu.edu, miguel.ballesteros@upf.edu, nasmith@cs.washington.edu
This is modifed version of a paper originally published at NAACL 2016 that contains a corrigendum at the end, with improved results after fixing an implementation bug in the RNNG composition function.
### $\mathbf{Abstract}$
We introduce recurrent neural network grammars, probabilistic models of sentences with explicit phrase structure. We explain efficient inference procedures that allow application to both parsing and language modeling. Experiments show that they provide better parsing in English than any single previously published supervised generative model and better language modeling than state-of-the-art sequential RNNs in English and Chinese$^{\mathrm{l}}.$
probabilistic model of sentences that explicitly models nested, hierarchical relationships among words and phrases. RNNGs operate via a recursive syntactic process reminiscent of probabilistic context-free grammar generation, but decisions are parameterized using RNNs that condition on the entire syntactic derivation history, greatly relaxing context-free independence assumptions.
The foundation of this work is a top-down variant of transition-based parsing (§3). We give two variants of the algorithm, one for parsing (given an observed sentence, transform it into a tree), and one for generation. While several transition-based neural models of syntactic generation exist (Henderson, 2003, 2004; Emami and Jelinek, 2005; Titov and Henderson, 2007; Buys and Blunsom, 2015b), these have relied on structure building operations based on parsing actions in shift-reduce and leftcorner parsers which operate in a largely bottomup fashion. While this construction is appealing because inference is relatively straightforward, it limits the use of top-down grammar information, which is helpful for generation (Roark, 2001).2 RNNGs maintain the algorithmic convenience of transitionbased parsing but incorporate top-down (i.e., rootto-terminal) syntactic information (§4).
The top-down transition set that RNNGs are based on lends itself to discriminative modeling as well, where sequences of transitions are modeled conditional on the full input sentence along with the incrementally constructed syntactic structures. Similar to previously published discriminative bottomup transition-based parsers (Henderson, 2004; Sagae and Lavie, 2005; Zhang and Clark, 2011, inter alia), greedy prediction with our model yields a linear-
$^2$The left-corner parsers used by Henderson (2003,2004) incorporate limited top-down information, but a complete path
### 1 Introduction
Sequential recurrent neural networks (RNNs) are remarkably effective models of natural language. In the last few years, language model results that substantially improve over long-established state-ofthe-art baselines have been obtained using RNNs (Zaremba et al., 2015; Mikolov et al., 2010) as well as in various conditional language modeling tasks such as machine translation (Bahdanau et al., 2015), image caption generation (Xu et al., 2015), and dialogue generation (Wen et al., 2015). Despite these impressive results, sequential models are $a\textit{priori}$ inappropriate models of natural language, since relationships among words are largely organized in terms of latent nested structures rather than sequential surface order (Chomsky, 1957).
In this paper, we introduce recurrent neural network grammars (RNNGs; §2), a new generative
from the root of the tree to a terminal is not generally present when a terminal is generated. Refer to Henderson (2003, Fig. 1) for an example.
IThe code to reproduce our results after the bug fix is pub-
licly available at https://github.com/clab/rnng.
|
[
[
0.306396484375,
0.09344482421875,
0.69384765625,
0.09344482421875,
0.69384765625,
0.10797119140625,
0.306396484375,
0.10797119140625
],
[
0.188720703125,
0.12249755859375,
0.81640625,
0.12249755859375,
0.81640625,
0.136962890625,
0.188720703125,
0.136962890625
],
[
0.1829833984375,
0.1395263671875,
0.81884765625,
0.14013671875,
0.81884765625,
0.1578369140625,
0.1829833984375,
0.1572265625
],
[
0.26953125,
0.1572265625,
0.73193359375,
0.1590576171875,
0.73193359375,
0.1767578125,
0.26953125,
0.1749267578125
],
[
0.178955078125,
0.17236328125,
0.82421875,
0.1749267578125,
0.82421875,
0.1956787109375,
0.178955078125,
0.1932373046875
],
[
0.10455322265625,
0.1962890625,
0.90283203125,
0.1962890625,
0.90283203125,
0.2108154296875,
0.10455322265625,
0.2108154296875
],
[
0.5087890625,
0.236083984375,
0.88134765625,
0.23486328125,
0.88134765625,
0.25,
0.5087890625,
0.250732421875
],
[
0.1282958984375,
0.246826171875,
0.47216796875,
0.2474365234375,
0.47216796875,
0.26513671875,
0.1282958984375,
0.264404296875
],
[
0.5087890625,
0.25390625,
0.88232421875,
0.25390625,
0.88232421875,
0.268310546875,
0.5087890625,
0.268310546875
],
[
0.0228729248046875,
0.26513671875,
0.05963134765625,
0.26513671875,
0.05963134765625,
0.417236328125,
0.0228729248046875,
0.417236328125
],
[
0.1298828125,
0.266357421875,
0.471435546875,
0.26513671875,
0.471435546875,
0.27978515625,
0.1298828125,
0.281005859375
],
[
0.50732421875,
0.26953125,
0.88232421875,
0.270751953125,
0.88232421875,
0.285400390625,
0.50732421875,
0.2841796875
],
[
0.1298828125,
0.2841796875,
0.47216796875,
0.28271484375,
0.47216796875,
0.29736328125,
0.1298828125,
0.298583984375
],
[
0.50732421875,
0.28857421875,
0.88232421875,
0.28857421875,
0.88232421875,
0.3017578125,
0.50732421875,
0.3017578125
],
[
0.1298828125,
0.300537109375,
0.473876953125,
0.2998046875,
0.473876953125,
0.314453125,
0.1298828125,
0.31494140625
],
[
0.5087890625,
0.304931640625,
0.88134765625,
0.3037109375,
0.88134765625,
0.318115234375,
0.5087890625,
0.3193359375
],
[
0.1298828125,
0.3193359375,
0.291748046875,
0.317626953125,
0.291748046875,
0.330810546875,
0.1298828125,
0.333251953125
],
[
0.5087890625,
0.3212890625,
0.88134765625,
0.3212890625,
0.88134765625,
0.3359375,
0.5087890625,
0.3359375
],
[
0.5087890625,
0.339111328125,
0.88232421875,
0.339111328125,
0.88232421875,
0.354248046875,
0.5087890625,
0.354248046875
],
[
0.2646484375,
0.3515625,
0.34228515625,
0.3515625,
0.34228515625,
0.369384765625,
0.2646484375,
0.369384765625
],
[
0.50732421875,
0.354736328125,
0.7099609375,
0.35595703125,
0.7099609375,
0.37060546875,
0.50732421875,
0.369384765625
],
[
0.52685546875,
0.37255859375,
0.8798828125,
0.37255859375,
0.8798828125,
0.386962890625,
0.52685546875,
0.386962890625
],
[
0.1502685546875,
0.38330078125,
0.451904296875,
0.3837890625,
0.451904296875,
0.3984375,
0.1502685546875,
0.397705078125
],
[
0.5107421875,
0.39013671875,
0.88232421875,
0.39013671875,
0.88232421875,
0.40478515625,
0.5107421875,
0.40478515625
],
[
0.02532958984375,
0.40478515625,
0.05963134765625,
0.40478515625,
0.05963134765625,
0.49560546875,
0.02532958984375,
0.49560546875
],
[
0.1519775390625,
0.399658203125,
0.451904296875,
0.3984375,
0.451904296875,
0.41357421875,
0.1519775390625,
0.4140625
],
[
0.5107421875,
0.40771484375,
0.88134765625,
0.40771484375,
0.88134765625,
0.421630859375,
0.5107421875,
0.421630859375
],
[
0.15283203125,
0.415283203125,
0.453369140625,
0.415283203125,
0.453369140625,
0.4287109375,
0.15283203125,
0.4287109375
],
[
0.1502685546875,
0.4287109375,
0.453369140625,
0.429931640625,
0.453369140625,
0.4443359375,
0.1502685546875,
0.443115234375
],
[
0.5107421875,
0.4248046875,
0.88232421875,
0.4248046875,
0.88232421875,
0.438232421875,
0.5107421875,
0.438232421875
],
[
0.1502685546875,
0.44580078125,
0.451904296875,
0.44580078125,
0.451904296875,
0.458984375,
0.1502685546875,
0.458984375
],
[
0.5107421875,
0.44140625,
0.88134765625,
0.44140625,
0.88134765625,
0.455078125,
0.5107421875,
0.455078125
],
[
0.1519775390625,
0.460205078125,
0.453369140625,
0.460205078125,
0.453369140625,
0.473388671875,
0.1519775390625,
0.473388671875
],
[
0.5087890625,
0.458984375,
0.88134765625,
0.458984375,
0.88134765625,
0.472900390625,
0.5087890625,
0.472900390625
],
[
0.02532958984375,
0.47607421875,
0.058013916015625,
0.47607421875,
0.058013916015625,
0.7109375,
0.02532958984375,
0.7109375
],
[
0.1519775390625,
0.474853515625,
0.453369140625,
0.474853515625,
0.453369140625,
0.4892578125,
0.1519775390625,
0.4892578125
],
[
0.5087890625,
0.47607421875,
0.88134765625,
0.47607421875,
0.88134765625,
0.4892578125,
0.5087890625,
0.4892578125
],
[
0.1502685546875,
0.490478515625,
0.453369140625,
0.490478515625,
0.453369140625,
0.50390625,
0.1502685546875,
0.50390625
],
[
0.5087890625,
0.492431640625,
0.88134765625,
0.492431640625,
0.88134765625,
0.505859375,
0.5087890625,
0.505859375
],
[
0.1519775390625,
0.505859375,
0.453369140625,
0.505859375,
0.453369140625,
0.51953125,
0.1519775390625,
0.51953125
],
[
0.5087890625,
0.51025390625,
0.88232421875,
0.51025390625,
0.88232421875,
0.5234375,
0.5087890625,
0.5234375
],
[
0.1502685546875,
0.52099609375,
0.38232421875,
0.52099609375,
0.38232421875,
0.5341796875,
0.1502685546875,
0.5341796875
],
[
0.5087890625,
0.5263671875,
0.88134765625,
0.5263671875,
0.88134765625,
0.54150390625,
0.5087890625,
0.54150390625
],
[
0.5087890625,
0.54443359375,
0.88134765625,
0.54443359375,
0.88134765625,
0.55810546875,
0.5087890625,
0.55810546875
],
[
0.1396484375,
0.5595703125,
0.253173828125,
0.5595703125,
0.253173828125,
0.57373046875,
0.1396484375,
0.57373046875
],
[
0.5087890625,
0.56005859375,
0.88134765625,
0.56005859375,
0.88134765625,
0.57470703125,
0.5087890625,
0.57470703125
],
[
0.5107421875,
0.5791015625,
0.88232421875,
0.5791015625,
0.88232421875,
0.59228515625,
0.5107421875,
0.59228515625
],
[
0.1168212890625,
0.5859375,
0.490234375,
0.5859375,
0.490234375,
0.599609375,
0.1168212890625,
0.599609375
],
[
0.5087890625,
0.59521484375,
0.88232421875,
0.59521484375,
0.88232421875,
0.60986328125,
0.5087890625,
0.60986328125
],
[
0.1168212890625,
0.60400390625,
0.488525390625,
0.60400390625,
0.488525390625,
0.6171875,
0.1168212890625,
0.6171875
],
[
0.50634765625,
0.6123046875,
0.88232421875,
0.6123046875,
0.88232421875,
0.626953125,
0.50634765625,
0.626953125
],
[
0.11517333984375,
0.62060546875,
0.490234375,
0.62060546875,
0.490234375,
0.6337890625,
0.11517333984375,
0.6337890625
],
[
0.5087890625,
0.6298828125,
0.88134765625,
0.6298828125,
0.88134765625,
0.64306640625,
0.5087890625,
0.64306640625
],
[
0.1168212890625,
0.63720703125,
0.48779296875,
0.63720703125,
0.48779296875,
0.6513671875,
0.1168212890625,
0.6513671875
],
[
0.50732421875,
0.64599609375,
0.8798828125,
0.64599609375,
0.8798828125,
0.66015625,
0.50732421875,
0.66015625
],
[
0.1168212890625,
0.65380859375,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.1168212890625,
0.66748046875
],
[
0.5087890625,
0.6640625,
0.79150390625,
0.6640625,
0.79150390625,
0.67822265625,
0.5087890625,
0.67822265625
],
[
0.1168212890625,
0.671875,
0.490234375,
0.671875,
0.490234375,
0.68505859375,
0.1168212890625,
0.68505859375
],
[
0.525390625,
0.6796875,
0.88134765625,
0.6796875,
0.88134765625,
0.693359375,
0.525390625,
0.693359375
],
[
0.11358642578125,
0.6875,
0.490234375,
0.68798828125,
0.490234375,
0.70263671875,
0.11358642578125,
0.7021484375
],
[
0.5087890625,
0.69775390625,
0.88232421875,
0.69775390625,
0.88232421875,
0.7109375,
0.5087890625,
0.7109375
],
[
0.11517333984375,
0.70556640625,
0.490234375,
0.70556640625,
0.490234375,
0.7197265625,
0.11517333984375,
0.7197265625
],
[
0.5087890625,
0.71533203125,
0.88232421875,
0.71533203125,
0.88232421875,
0.72900390625,
0.5087890625,
0.72900390625
],
[
0.11517333984375,
0.7236328125,
0.488525390625,
0.7236328125,
0.488525390625,
0.7373046875,
0.11517333984375,
0.7373046875
],
[
0.5087890625,
0.73291015625,
0.88134765625,
0.73291015625,
0.88134765625,
0.7470703125,
0.5087890625,
0.7470703125
],
[
0.11358642578125,
0.74072265625,
0.491943359375,
0.7392578125,
0.491943359375,
0.75390625,
0.11358642578125,
0.7548828125
],
[
0.1168212890625,
0.755859375,
0.488525390625,
0.755859375,
0.488525390625,
0.77099609375,
0.1168212890625,
0.77099609375
],
[
0.5087890625,
0.75,
0.88134765625,
0.75,
0.88134765625,
0.76318359375,
0.5087890625,
0.76318359375
],
[
0.5087890625,
0.767578125,
0.88134765625,
0.767578125,
0.88134765625,
0.78076171875,
0.5087890625,
0.78076171875
],
[
0.11517333984375,
0.7744140625,
0.48779296875,
0.7744140625,
0.48779296875,
0.78857421875,
0.11517333984375,
0.78857421875
],
[
0.5107421875,
0.7841796875,
0.88134765625,
0.7841796875,
0.88134765625,
0.79736328125,
0.5107421875,
0.79736328125
],
[
0.11517333984375,
0.79150390625,
0.491943359375,
0.79150390625,
0.491943359375,
0.80615234375,
0.11517333984375,
0.80615234375
],
[
0.5087890625,
0.7998046875,
0.88134765625,
0.80029296875,
0.88134765625,
0.81494140625,
0.5087890625,
0.814453125
],
[
0.11517333984375,
0.8095703125,
0.488525390625,
0.8095703125,
0.488525390625,
0.82275390625,
0.11517333984375,
0.82275390625
],
[
0.51123046875,
0.818359375,
0.88134765625,
0.818359375,
0.88134765625,
0.83203125,
0.51123046875,
0.83203125
],
[
0.11517333984375,
0.82568359375,
0.371826171875,
0.82568359375,
0.371826171875,
0.8388671875,
0.11517333984375,
0.8388671875
],
[
0.131591796875,
0.84228515625,
0.488525390625,
0.84130859375,
0.488525390625,
0.85595703125,
0.131591796875,
0.85693359375
],
[
0.52783203125,
0.84033203125,
0.88232421875,
0.84130859375,
0.88232421875,
0.85595703125,
0.52783203125,
0.85498046875
],
[
0.11358642578125,
0.85791015625,
0.49267578125,
0.859375,
0.49267578125,
0.87353515625,
0.11358642578125,
0.87255859375
],
[
0.50634765625,
0.85693359375,
0.88134765625,
0.85791015625,
0.88134765625,
0.869140625,
0.50634765625,
0.86865234375
],
[
0.5087890625,
0.87060546875,
0.88232421875,
0.8720703125,
0.88232421875,
0.88330078125,
0.5087890625,
0.8818359375
],
[
0.1396484375,
0.8837890625,
0.490234375,
0.8837890625,
0.490234375,
0.89794921875,
0.1396484375,
0.89794921875
],
[
0.5087890625,
0.88525390625,
0.88134765625,
0.88525390625,
0.88134765625,
0.8984375,
0.5087890625,
0.8984375
],
[
0.11358642578125,
0.8984375,
0.46728515625,
0.8984375,
0.46728515625,
0.912109375,
0.11358642578125,
0.912109375
],
[
0.5107421875,
0.8984375,
0.62109375,
0.8984375,
0.62109375,
0.912109375,
0.5107421875,
0.912109375
]
] |
[
"Recurrent Neural Network Grammars",
"Chris Dver*Adhiguna Kuncoro* Miguel BallesterosNoah A.Smith",
"*School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA",
"NLP Group, Pompeu Fabra University, Barcelona, Spain",
"Computer Science & Engineering, University of Washington, Seattle, WA, USA",
"[cdyer,akuncoro]@cs.cmu.edu, miguel.ballesteros@upf.edu, nasmith@cs.washington.edu",
"probabilistic model of sentences that explicitly mod-",
"This is modified version of a paper originally",
"els nested, hierarchical relationships among words",
"911[",
"published at NAACL 2016 that contains a cor-",
"and phrases. RNNGs operate via a recursive syntac-",
"rigendum at the end, with improved results af-",
"ticprocess reminiscent of probabilistic context-free",
"ter fixing an implementation bug in the RNNG",
"grammar generation, but decisions are parameter-",
"composition function.",
"ized using RNNs that condition on the entire syntac-",
"tic derivation history, greatly relaxing context-free",
"Abstract",
"independence assumptions.",
"The foundation of this work is a top-down vari-",
"We introduce recurrent neural network gram-",
"ant of transition-based parsing (g3). We give two",
"[T'S]",
"mars,probabilistic models of sentences with",
"variants of the algorithm, one for parsing (given an",
"explicit phrase structure. We explain efficient",
"inference procedures that allow application to",
"observed sentence, transform it into a tree), and one",
"both parsing and language modeling.Experi-",
"for generation. While several transition-based neu-",
"ments show that they provide better parsing in",
"ral models of syntactic generation exist (Hender-",
"19110209110",
"English than any single previously published",
"son, 2003, 2004; Emami and Jelinek, 2005; Titov",
"supervised generative model and better lan-",
"and Henderson, 2007; Buys and Blunsom, 2015b),",
"guage modeling than state-of-the-art sequen-",
"these have relied on structure building operations",
"tial RNNs in English and Chinesel",
"based on parsing actions in shift-reduce and left-",
"corner parsers which operate in a largely bottom-",
"Introduction",
"up fashion. While this construction is appealing be-",
"cause inference is relatively straightforward, it lim-",
"Sequential recurrent neural networks (RNNs) are",
"its the use of top-down grammar information, which",
"remarkably effective models of natural language.",
"is helpful for generation (Roark, 2001).2 RNNGs",
"In the last few years, language model results that",
"maintain the algorithmic convenience of transition-",
"substantially improve over long-established state-of-",
"based parsing but incorporate top-down (i.e., root-",
"the-art baselines have been obtained using RNNs",
"to-terminal) syntactic information ($4).",
"(Zaremba et al., 2015; Mikolov et al., 2010) as well",
"The top-down transition set that RNNGs are",
"as in various conditional language modeling tasks",
"based on lends itself to discriminative modeling as",
"such as machine translation (Bahdanau et al., 2015),",
"well, where sequences of transitions are modeled",
"image caption generation (Xu et al., 2015), and dia-",
"conditional on the full input sentence along with the",
"logue generation (Wen et al., 2015). Despite these",
"impressive results, sequential models are a priori",
"incrementally constructed syntactic structures. Sim-",
"ilar to previously published discriminative bottom-",
"inappropriate models of natural language, since re-",
"up transition-based parsers (Henderson,2004; Sagae",
"lationships among words are largely organized in",
"and Lavie, 2005; Zhang and Clark, 2011, inter alia),",
"terms of latent nested structures rather than sequen-",
"greedy prediction with our model yields a linear-",
"tial surface order (Chomsky, 1957).",
"In this paper, we introduce recurrent neural net-",
"2The left-corner parsers used by Henderson (2003, 2004)",
"work grammars (RNNGs; s2), a new generative",
"incorporatelimitedtop-downinformation,butacompletepath",
"fromtherootof thetreetoaterminal is notgenerallypresent",
"The code to reproduce our results after the bug fix is pub-",
"when a terminal is generated. Refer to Henderson (2003, Fig",
"licly available at https://github.com/clab/rnng.",
"1) for an example."
] |
{
"arxiv_id": "1602.07776",
"language": "en",
"page_idx": 0
}
|
||
1602.07776-en-1
|
1602.07776-en-1.jpg
|
on the number of actions taken between processever, our algorithm generates arbitrary tree structures directly, without the binarization required by shift-reduce parsers. The discriminative model also lets us use ancestor sampling to obtain samples of parse trees for sentences, and this is used to solve a second practical challenge with RNNGs: approximating the marginal likelihood and MAP tree of a sentence under the generative model. We present a simple importance sampling algorithm which uses samples from the discriminative parser to solve inference problems in the generative model (§5).
Experiments show that RNNGs are effective for both language modeling and parsing (§6). Our generative model obtains (i) the best-known parsing results using a single supervised generative model and (ii) better perplexities in language modeling than state-of-the-art sequential LSTM language models. Surprisingly—although in line with previous parsing results showing the effectiveness of generative models (Henderson, 2004; Johnson, 2001)— parsing with the generative model obtains signifrcantly better results than parsing with the discriminative model.
## 2 $\mathbf{RNNGrammars}$
Formally, an RNNG is a triple $(N,\Sigma,\Theta)$ consisting of a finite set of nonterminal symbols (N), a finite set of terminal symbols (Σ) such that $N\cap\Sigma=\emptyset$, and a collection of neural network parameters $\Theta.$ It does not explicitly define rules since these are implicitly characterized by $\Theta.$ The algorithm that the grammar uses to generate trees and strings in the language is characterized in terms of a transition-based algorithm, which is outlined in the next section. In the section after that, the semantics of the parameters that are used to turn this into a stochastic algorithm that generates pairs of trees and strings are discussed.
## 3 Top-down Parsing and Generation
RNNGs are based on a top-down generation algorithm that relies on a stack data structure of partially completed syntactic constituents. To emphasize the similarity of our algorithm to more familiar
time deterministic parser (provided an upper bound bottom-up shift-reduce recognition algorithms, we
first present the parsing (rather than generation) ver-
ing subsequent terminal symbols is imposed); how- sion of our algorithm (§3.1) and then present modi-
fications to turn it into a generator $(\S3.2).$
## 3.1 Parser Transitions
The parsing algorithm transforms a sequence of words $x$ into a parse tree $y$ using two data structures (a stack and an input buffer). As with the bottomup algorithm of Sagae and Lavie (2005), our algorithm begins with the stack $(S)$ empty and the complete sequence of words in the input buffer (B). The buffer contains unprocessed terminal symbols, and the stack contains terminal symbols,“open”nonterminal symbols, and completed constituents. At each timestep, one of the following three classes of operations (Fig. 1) is selected by a classifier, based on the current contents on the stack and buffer:
$\bullet$ NT(X) introduces an“open nonterminal”X onto the top of the stack. Open nonterminals are written as a nonterminal symbol preceded by an open parenthesis, e.g.,“(VP", and they represent a nonterminal whose child nodes have not yet been fully constructed. Open nonterminals are “closed” to form complete constituents by subsequent REDUCE operations.
$\bullet$ SHIFT removes the terminal symbol $x$ from the front of the input buffer, and pushes it onto the top of the stack
$\bullet$ REDUCE repeatedly pops completed subtrees or terminal symbols from the stack until an open nonterminal is encountered, and then this open NT is popped and used as the label of a new constituent that has the popped subtrees as its children. This new completed constituent is pushed onto the stack as a single composite item. A single REDUCE operation can thus create constituents with an unbounded number of children.
The parsing algorithm terminates when there is a single completed constituent on the stack and the buffer is empty. Fig.2 shows an example parse using our transition set. Note that in this paper we do not model preterminal symbols (i.e., part-ofspeech tags) and our examples therefore do not include them.$^{3}$
$^3$Preterminal symbols are, from the parsing algorithm's
point of view, just another kind of nonterminal symbol that re-
|
[
[
0.11517333984375,
0.0921630859375,
0.491943359375,
0.091552734375,
0.491943359375,
0.1060791015625,
0.11517333984375,
0.106689453125
],
[
0.50634765625,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.106689453125,
0.50634765625,
0.106689453125
],
[
0.11517333984375,
0.10986328125,
0.490234375,
0.111083984375,
0.490234375,
0.125,
0.11517333984375,
0.12371826171875
],
[
0.5087890625,
0.111083984375,
0.8798828125,
0.111083984375,
0.8798828125,
0.125,
0.5087890625,
0.125
],
[
0.1168212890625,
0.1275634765625,
0.491943359375,
0.1275634765625,
0.491943359375,
0.141357421875,
0.1168212890625,
0.141357421875
],
[
0.5087890625,
0.1275634765625,
0.88134765625,
0.1275634765625,
0.88134765625,
0.141357421875,
0.5087890625,
0.141357421875
],
[
0.1168212890625,
0.143310546875,
0.488525390625,
0.143310546875,
0.488525390625,
0.1578369140625,
0.1168212890625,
0.1578369140625
],
[
0.5087890625,
0.14453125,
0.8037109375,
0.14453125,
0.8037109375,
0.1590576171875,
0.5087890625,
0.1590576171875
],
[
0.1168212890625,
0.1622314453125,
0.488525390625,
0.1622314453125,
0.488525390625,
0.175537109375,
0.1168212890625,
0.175537109375
],
[
0.50732421875,
0.1717529296875,
0.69287109375,
0.1717529296875,
0.69287109375,
0.1849365234375,
0.50732421875,
0.1849365234375
],
[
0.1168212890625,
0.177978515625,
0.488525390625,
0.177978515625,
0.488525390625,
0.1912841796875,
0.1168212890625,
0.1912841796875
],
[
0.1168212890625,
0.1962890625,
0.49267578125,
0.1962890625,
0.49267578125,
0.210205078125,
0.1168212890625,
0.210205078125
],
[
0.5048828125,
0.1925048828125,
0.88134765625,
0.1925048828125,
0.88134765625,
0.20703125,
0.5048828125,
0.20703125
],
[
0.11358642578125,
0.21337890625,
0.494384765625,
0.212158203125,
0.494384765625,
0.2266845703125,
0.11358642578125,
0.2279052734375
],
[
0.5087890625,
0.2108154296875,
0.88232421875,
0.2108154296875,
0.88232421875,
0.2247314453125,
0.5087890625,
0.2247314453125
],
[
0.11517333984375,
0.2298583984375,
0.49267578125,
0.2298583984375,
0.49267578125,
0.2442626953125,
0.11517333984375,
0.2442626953125
],
[
0.5087890625,
0.2279052734375,
0.88232421875,
0.2279052734375,
0.88232421875,
0.2412109375,
0.5087890625,
0.2412109375
],
[
0.11517333984375,
0.2474365234375,
0.495849609375,
0.2474365234375,
0.495849609375,
0.261474609375,
0.11517333984375,
0.261474609375
],
[
0.5087890625,
0.2442626953125,
0.88232421875,
0.2442626953125,
0.88232421875,
0.2587890625,
0.5087890625,
0.2587890625
],
[
0.11358642578125,
0.264404296875,
0.494384765625,
0.26318359375,
0.494384765625,
0.27783203125,
0.11358642578125,
0.279052734375
],
[
0.5087890625,
0.261962890625,
0.88134765625,
0.261962890625,
0.88134765625,
0.27587890625,
0.5087890625,
0.27587890625
],
[
0.11358642578125,
0.281005859375,
0.49267578125,
0.27978515625,
0.49267578125,
0.294189453125,
0.11358642578125,
0.29541015625
],
[
0.50732421875,
0.279052734375,
0.88232421875,
0.27783203125,
0.88232421875,
0.292236328125,
0.50732421875,
0.29345703125
],
[
0.11517333984375,
0.298583984375,
0.490234375,
0.298583984375,
0.490234375,
0.3125,
0.11517333984375,
0.3125
],
[
0.5087890625,
0.296630859375,
0.88232421875,
0.296630859375,
0.88232421875,
0.311279296875,
0.5087890625,
0.311279296875
],
[
0.11517333984375,
0.31640625,
0.449462890625,
0.31640625,
0.449462890625,
0.330078125,
0.11517333984375,
0.330078125
],
[
0.50732421875,
0.3125,
0.88134765625,
0.313232421875,
0.88134765625,
0.32763671875,
0.50732421875,
0.326904296875
],
[
0.134033203125,
0.33203125,
0.488525390625,
0.33203125,
0.488525390625,
0.34521484375,
0.134033203125,
0.34521484375
],
[
0.5087890625,
0.330810546875,
0.88232421875,
0.330810546875,
0.88232421875,
0.3447265625,
0.5087890625,
0.3447265625
],
[
0.11358642578125,
0.349853515625,
0.488525390625,
0.35107421875,
0.488525390625,
0.365478515625,
0.11358642578125,
0.3642578125
],
[
0.5087890625,
0.347900390625,
0.8798828125,
0.347900390625,
0.8798828125,
0.3623046875,
0.5087890625,
0.3623046875
],
[
0.11517333984375,
0.36865234375,
0.486083984375,
0.36865234375,
0.486083984375,
0.3818359375,
0.11517333984375,
0.3818359375
],
[
0.5087890625,
0.3642578125,
0.88232421875,
0.3642578125,
0.88232421875,
0.37890625,
0.5087890625,
0.37890625
],
[
0.1168212890625,
0.3837890625,
0.48779296875,
0.3837890625,
0.48779296875,
0.397705078125,
0.1168212890625,
0.397705078125
],
[
0.5087890625,
0.3818359375,
0.826171875,
0.3818359375,
0.826171875,
0.395751953125,
0.5087890625,
0.395751953125
],
[
0.11517333984375,
0.40283203125,
0.488525390625,
0.40283203125,
0.488525390625,
0.416748046875,
0.11517333984375,
0.416748046875
],
[
0.51123046875,
0.4072265625,
0.88134765625,
0.4072265625,
0.88134765625,
0.42041015625,
0.51123046875,
0.42041015625
],
[
0.1168212890625,
0.41845703125,
0.486083984375,
0.41845703125,
0.486083984375,
0.43310546875,
0.1168212890625,
0.43310546875
],
[
0.52685546875,
0.4248046875,
0.88134765625,
0.4248046875,
0.88134765625,
0.438232421875,
0.52685546875,
0.438232421875
],
[
0.11517333984375,
0.43505859375,
0.488525390625,
0.436279296875,
0.488525390625,
0.45068359375,
0.11517333984375,
0.449462890625
],
[
0.52685546875,
0.44140625,
0.88134765625,
0.44140625,
0.88134765625,
0.455078125,
0.52685546875,
0.455078125
],
[
0.11517333984375,
0.453857421875,
0.488525390625,
0.453857421875,
0.488525390625,
0.46728515625,
0.11517333984375,
0.46728515625
],
[
0.52685546875,
0.458984375,
0.88134765625,
0.458984375,
0.88134765625,
0.473388671875,
0.52685546875,
0.473388671875
],
[
0.1168212890625,
0.4697265625,
0.48779296875,
0.4697265625,
0.48779296875,
0.484130859375,
0.1168212890625,
0.484130859375
],
[
0.52685546875,
0.47607421875,
0.88134765625,
0.47607421875,
0.88134765625,
0.490478515625,
0.52685546875,
0.490478515625
],
[
0.11358642578125,
0.4873046875,
0.488525390625,
0.486083984375,
0.488525390625,
0.50048828125,
0.11358642578125,
0.501953125
],
[
0.52685546875,
0.49365234375,
0.88134765625,
0.49365234375,
0.88134765625,
0.5068359375,
0.52685546875,
0.5068359375
],
[
0.11517333984375,
0.50390625,
0.488525390625,
0.50390625,
0.488525390625,
0.51806640625,
0.11517333984375,
0.51806640625
],
[
0.52685546875,
0.51025390625,
0.8798828125,
0.51025390625,
0.8798828125,
0.5234375,
0.52685546875,
0.5234375
],
[
0.11517333984375,
0.521484375,
0.221435546875,
0.521484375,
0.221435546875,
0.5361328125,
0.11517333984375,
0.5361328125
],
[
0.525390625,
0.52783203125,
0.7197265625,
0.5263671875,
0.7197265625,
0.54150390625,
0.525390625,
0.54248046875
],
[
0.51123046875,
0.54541015625,
0.88134765625,
0.54541015625,
0.88134765625,
0.5595703125,
0.51123046875,
0.5595703125
],
[
0.11358642578125,
0.5517578125,
0.28759765625,
0.5517578125,
0.28759765625,
0.56640625,
0.11358642578125,
0.56640625
],
[
0.52685546875,
0.56298828125,
0.88134765625,
0.56298828125,
0.88134765625,
0.57763671875,
0.52685546875,
0.57763671875
],
[
0.11517333984375,
0.58056640625,
0.488525390625,
0.58056640625,
0.488525390625,
0.59521484375,
0.11517333984375,
0.59521484375
],
[
0.52685546875,
0.580078125,
0.64208984375,
0.580078125,
0.64208984375,
0.5947265625,
0.52685546875,
0.5947265625
],
[
0.11517333984375,
0.5986328125,
0.488525390625,
0.5986328125,
0.488525390625,
0.6123046875,
0.11517333984375,
0.6123046875
],
[
0.51123046875,
0.5986328125,
0.88232421875,
0.5986328125,
0.88232421875,
0.61279296875,
0.51123046875,
0.61279296875
],
[
0.11517333984375,
0.61572265625,
0.488525390625,
0.61572265625,
0.488525390625,
0.62890625,
0.11517333984375,
0.62890625
],
[
0.52685546875,
0.6162109375,
0.88134765625,
0.6162109375,
0.88134765625,
0.6298828125,
0.52685546875,
0.6298828125
],
[
0.11517333984375,
0.6318359375,
0.488525390625,
0.6318359375,
0.488525390625,
0.64599609375,
0.11517333984375,
0.64599609375
],
[
0.525390625,
0.6318359375,
0.88134765625,
0.63330078125,
0.88134765625,
0.64794921875,
0.525390625,
0.646484375
],
[
0.11517333984375,
0.6494140625,
0.486083984375,
0.6494140625,
0.486083984375,
0.66357421875,
0.11517333984375,
0.66357421875
],
[
0.52685546875,
0.6494140625,
0.88134765625,
0.6494140625,
0.88134765625,
0.6640625,
0.52685546875,
0.6640625
],
[
0.1168212890625,
0.66748046875,
0.488525390625,
0.66748046875,
0.488525390625,
0.68115234375,
0.1168212890625,
0.68115234375
],
[
0.52685546875,
0.66748046875,
0.8798828125,
0.66748046875,
0.8798828125,
0.681640625,
0.52685546875,
0.681640625
],
[
0.1168212890625,
0.68505859375,
0.486083984375,
0.68505859375,
0.486083984375,
0.69873046875,
0.1168212890625,
0.69873046875
],
[
0.52685546875,
0.68505859375,
0.88134765625,
0.68505859375,
0.88134765625,
0.69873046875,
0.52685546875,
0.69873046875
],
[
0.1168212890625,
0.7021484375,
0.488525390625,
0.7021484375,
0.488525390625,
0.71533203125,
0.1168212890625,
0.71533203125
],
[
0.52685546875,
0.7021484375,
0.88134765625,
0.7021484375,
0.88134765625,
0.71533203125,
0.52685546875,
0.71533203125
],
[
0.1168212890625,
0.71826171875,
0.488525390625,
0.71826171875,
0.488525390625,
0.73193359375,
0.1168212890625,
0.73193359375
],
[
0.52685546875,
0.71826171875,
0.88232421875,
0.71826171875,
0.88232421875,
0.73193359375,
0.52685546875,
0.73193359375
],
[
0.1168212890625,
0.7373046875,
0.484375,
0.7373046875,
0.484375,
0.74755859375,
0.1168212890625,
0.74755859375
],
[
0.529296875,
0.7373046875,
0.810546875,
0.7373046875,
0.810546875,
0.74755859375,
0.529296875,
0.74755859375
],
[
0.11358642578125,
0.75244140625,
0.486083984375,
0.751953125,
0.486083984375,
0.76513671875,
0.11358642578125,
0.7666015625
],
[
0.51123046875,
0.76025390625,
0.88232421875,
0.76025390625,
0.88232421875,
0.7744140625,
0.51123046875,
0.7744140625
],
[
0.1168212890625,
0.76953125,
0.488525390625,
0.76953125,
0.488525390625,
0.7841796875,
0.1168212890625,
0.7841796875
],
[
0.51123046875,
0.77880859375,
0.88134765625,
0.77880859375,
0.88134765625,
0.79248046875,
0.51123046875,
0.79248046875
],
[
0.1168212890625,
0.787109375,
0.188720703125,
0.787109375,
0.188720703125,
0.798828125,
0.1168212890625,
0.798828125
],
[
0.5107421875,
0.79541015625,
0.88232421875,
0.79541015625,
0.88232421875,
0.81005859375,
0.5107421875,
0.81005859375
],
[
0.11273193359375,
0.8173828125,
0.4345703125,
0.8173828125,
0.4345703125,
0.83203125,
0.11273193359375,
0.83203125
],
[
0.5107421875,
0.8125,
0.88232421875,
0.8125,
0.88232421875,
0.8271484375,
0.5107421875,
0.8271484375
],
[
0.5087890625,
0.830078125,
0.8798828125,
0.830078125,
0.8798828125,
0.84326171875,
0.5087890625,
0.84326171875
],
[
0.11358642578125,
0.84326171875,
0.48779296875,
0.8447265625,
0.48779296875,
0.8623046875,
0.11358642578125,
0.861328125
],
[
0.50732421875,
0.8466796875,
0.88134765625,
0.84521484375,
0.88134765625,
0.85986328125,
0.50732421875,
0.861328125
],
[
0.11517333984375,
0.8642578125,
0.490234375,
0.8642578125,
0.490234375,
0.87744140625,
0.11517333984375,
0.87744140625
],
[
0.5087890625,
0.86279296875,
0.60107421875,
0.86279296875,
0.60107421875,
0.87744140625,
0.5087890625,
0.87744140625
],
[
0.11517333984375,
0.880859375,
0.488525390625,
0.880859375,
0.488525390625,
0.89453125,
0.11517333984375,
0.89453125
],
[
0.52783203125,
0.880859375,
0.8837890625,
0.8818359375,
0.8837890625,
0.8994140625,
0.52783203125,
0.8984375
],
[
0.11358642578125,
0.89794921875,
0.491943359375,
0.896484375,
0.491943359375,
0.9111328125,
0.11358642578125,
0.912109375
],
[
0.50732421875,
0.8984375,
0.88134765625,
0.8984375,
0.88134765625,
0.912109375,
0.50732421875,
0.912109375
]
] |
[
"time deterministic parser (provided an upper bound",
"bottom-up shift-reduce recognition algorithms, we",
"on the number of actions taken between process-",
"first present the parsing (rather than generation) ver-",
"ing subsequent terminal symbols is imposed); how-",
"sion of our algorithm (3.1) and then present modi-",
"ever, our algorithm generates arbitrary tree struc-",
"fications to turn it into a generator (g3.2).",
"tures directly, without the binarization required by",
"3.1 Parser Transitions",
"shift-reduce parsers. The discriminative model also",
"lets us use ancestor sampling to obtain samples of",
"The parsing algorithm transforms a sequence of",
"parse trees for sentences, and this is used to solve",
"words a into a parse tree y using two data structures",
"a second practical challenge with RNNGs: approx-",
"(a stack and an input buffer). As with the bottom-",
"imating the marginal likelihood and MAP tree of a",
"up algorithm of Sagae and Lavie (2005), our algo-",
"sentence under the generative model. We present a",
"rithm begins with the stack (S) empty and the com-",
"simple importance sampling algorithm which uses",
"plete sequence of words in the input buffer (B). The",
"samples from the discriminative parser to solve in-",
"buffer contains unprocessed terminal symbols, and",
"ference problems in the generative model (s5).",
"the stack contains terminal symbols,“\"open\"nonter-",
"Experiments show that RNNGs are effective for",
"minal symbols, and completed constituents. At each",
"both language modeling and parsing (s6). Our gen-",
"timestep, one of the following three classes of op-",
"erative model obtains (i) the best-known parsing re-",
"erations (Fig. 1) is selected by a classifier, based on",
"sults using a single supervised generative model and",
"the current contents on the stack and buffer:",
"(ii) better perplexities in language modeling than",
". NT(X) introduces an “open nonterminal\" X onto",
"state-of-the-art sequential LSTM language models",
"the top of the stack.Open nonterminals are",
"Surprisingly-although in line with previous pars-",
"written as a nonterminal symbol preceded by an",
"ing results showing the effectiveness of genera-",
"open parenthesis, e.g., “\"(VP\", and they represent",
"tive models (Henderson, 2004; Johnson, 2001)",
"a nonterminal whose child nodes have not yet",
"parsing with the generative model obtains signifi-",
"been fully constructed. Open nonterminals are",
"cantly better results than parsing with the discrim-",
"\"closed\"to form complete constituents by subse-",
"inative model.",
"quent REDUCE operations.",
" SHIFT removes the terminal symbol & from the",
"2 RNN Grammars",
"front of the input buffer, and pushes it onto the",
"Formally, an RNNG is a triple (N, 2, ) consisting",
"top of the stack.",
"of a finite set of nonterminal symbols (N), a finite",
"· REDUCE repeatedly pops completed subtrees or",
"set of terminal symbols () such that N n = 0,",
"terminal symbols from the stack until an open",
"and a collection of neural network parameters . It",
"nonterminal is encountered, and then this open",
"does not explicitly define rules since these are im-",
"NT is popped and used as the label of a new con-",
"plicitly characterized by . The algorithm that the",
"stituent that has the popped subtrees as its chil-",
"grammar uses to generate trees and strings in the lan-",
"dren. This new completed constituent is pushed",
"guage is characterized in terms of a transition-based",
"onto the stack as a single composite item. A single",
"algorithm, which is outlined in the next section. In",
"REDUCE operation can thus create constituents",
"the section after that.the semantics of the param",
"withanunboundednumberof children.",
"eters that are used to turn this into a stochastic al-",
"The parsing algorithm terminates when there is a",
"gorithm that generates pairs of trees and strings are",
"single completed constituent on the stack and the",
"discussed",
"buffer is empty.Fig.2 shows an example parse",
"3Top-down Parsing and Generation",
"using our transition set. Note that in this paper",
"we do not model preterminal symbols (i.e., part-of-",
"RNNGs are based on a top-down generation algo-",
"speech tags) and our examples therefore do not in-",
"rithm that relies on a stack data structure of par-",
"clude them.3",
"tially completed syntactic constituents. To empha-",
"'Preterminal symbols are, from the parsing algorithm's",
"size the similarity of our algorithm to more familiar",
"point of view, just another kind of nonterminal symbol that re-"
] |
{
"arxiv_id": "1602.07776",
"language": "en",
"page_idx": 1
}
|
||
1602.07776-en-2
|
1602.07776-en-2.jpg
|
Our transition set is closely related to the operations used in Earley's algorithm which likewise introduces nonterminals symbols with its PREDICT operation and later COMPLETEs them after consuming terminal symbols one at a time using SCAN (Earley, 1970). It is likewise closely related to the “linearized” parse trees proposed by Vinyals et al. (2015) and to the top-down, left-to-right decompositions of trees used in previous generative parsing and language modeling work (Roark, 2001, 2004; Charniak, 2010).
A further connection is to $LL(^*)$ parsing which uses an unbounded lookahead (compactly represented by a DFA) to distinguish between parse alternatives in a top-down parser (Parr and Fisher, 2011); however, our parser uses an RNN encoding of the lookahead rather than a DFA.
Constraints on parser transitions. To guarantee that only well-formed phrase-structure trees are produced by the parser, we impose the following constraints on the transitions that can be applied at each step which are a function of the parser state $(B,S,n)$ where $n$ is the number of open nonterminals on the $\bullet$ The REDUCE operation can only be applied if the
stack:
$\bullet$ The NT(X) operation can only be applied if $B$ is
not empty and $n<100.^4$
$\bullet$ The SHIFT operation can only be applied if $B$ is
not empty and $n\geq1.$
$\bullet$ The REDUCE operation can only be applied if the top of the stack is not an open nonterminal symbol.
$\bullet$ The REDUCE operation can only be applied if $n\geq$
2 or if the buffer is empty.
To designate the set of valid parser transitions, we
# write $\mathcal{A}_D(B,S,n).$
$\overline{\text{quires no special}}$ handling. However, leaving them out reduces the number of transitions by $O(n)$ and also reduces the number of action types, both of which reduce the runtime. Furthermore, standard parsing evaluation scores do not depend on preterminal prediction accuracy.
$^{4}$Since our parser allows unary nonterminal productions, there are an infinite number of valid trees for finite-length sentences. The $n<100$ constraint prevents the classifier from misbehaving and generating excessively large numbers of nonterminals. Similar constraints have been proposed to deal with the analogous problem in bottom-up shift-reduce parsers (Sagae and Lavie, 2005).
## 3.2 Generator Transitions
The parsing algorithm that maps from sequences of words to parse trees can be adapted with minor changes to produce an algorithm that stochastically generates trees and terminal symbols. Two changes are required: (i) there is no input buffer of unprocessed words, rather there is an output buffer (T), and (ii) instead of a SHIFT operation there are GEN$(x)$ operations which generate terminal symbol $x\in\Sigma$ and add it to the top of the stack and the output buffer. At each timestep an action is stochastically selected according to a conditional distribution that depends on the current contents of $S$ and $T.$ The algorithm terminates when a single completed constituent remains on the stack. Fig.4 shows an example generation sequence.
## Constraints on generator transitions. The generation algorithm also requires slightly modified constraints. These are:
$\bullet$ The GEN$(x)$ operation can only be applied if $n\geq$
1.
top of the stack is not an open nonterminal symbol
and $n\geq1.$
To designate the set of valid generator transitions,
we write $\mathcal{A}_G(T,S,n).$
This transition set generates trees using nearly the same structure building actions and stack configurations as the“top-down PDA”construction proposed by Abney et al. (1999), albeit without the restriction that the trees be in Chomsky normal form.
## 3.3 Transition Sequences from Trees
Any parse tree can be converted to a sequence of transitions via a depth-first, left-to-right traversal of a parse tree. Since there is a unique depth-first, leftro-right traversal of a tree, there is exactly one transition sequence of each tree. For a tree $y$ and a sequence of symbols $x$, we write $a(x,y)$ to indicate the corresponding sequence of generation transitions, and $b(x,y)$ to indicate the parser transitions.
## 3.4 Runtime Analysis
A detailed analysis of the algorithmic properties of our top-down parser is beyond the scope of this paper; however, we briefly state several facts. As-
|
[
[
0.1356201171875,
0.0921630859375,
0.5,
0.0921630859375,
0.5,
0.1060791015625,
0.1356201171875,
0.1060791015625
],
[
0.50634765625,
0.0921630859375,
0.72216796875,
0.0921630859375,
0.72216796875,
0.106689453125,
0.50634765625,
0.106689453125
],
[
0.11517333984375,
0.10919189453125,
0.49267578125,
0.10986328125,
0.49267578125,
0.125,
0.11517333984375,
0.12371826171875
],
[
0.5087890625,
0.11553955078125,
0.88232421875,
0.11553955078125,
0.88232421875,
0.1300048828125,
0.5087890625,
0.1300048828125
],
[
0.1168212890625,
0.1275634765625,
0.490234375,
0.1275634765625,
0.490234375,
0.141357421875,
0.1168212890625,
0.141357421875
],
[
0.50634765625,
0.1331787109375,
0.88134765625,
0.1331787109375,
0.88134765625,
0.146484375,
0.50634765625,
0.146484375
],
[
0.11517333984375,
0.14453125,
0.488525390625,
0.143310546875,
0.488525390625,
0.1578369140625,
0.11517333984375,
0.1590576171875
],
[
0.5087890625,
0.149658203125,
0.88134765625,
0.149658203125,
0.88134765625,
0.1634521484375,
0.5087890625,
0.1634521484375
],
[
0.1168212890625,
0.1622314453125,
0.491943359375,
0.1622314453125,
0.491943359375,
0.175537109375,
0.1168212890625,
0.175537109375
],
[
0.5087890625,
0.1673583984375,
0.88134765625,
0.1673583984375,
0.88134765625,
0.18115234375,
0.5087890625,
0.18115234375
],
[
0.1168212890625,
0.1787109375,
0.490234375,
0.1787109375,
0.490234375,
0.1925048828125,
0.1168212890625,
0.1925048828125
],
[
0.5087890625,
0.184326171875,
0.88232421875,
0.184326171875,
0.88232421875,
0.1976318359375,
0.5087890625,
0.1976318359375
],
[
0.1168212890625,
0.1962890625,
0.490234375,
0.1962890625,
0.490234375,
0.2108154296875,
0.1168212890625,
0.2108154296875
],
[
0.50732421875,
0.2008056640625,
0.88134765625,
0.199462890625,
0.88134765625,
0.2139892578125,
0.50732421875,
0.21533203125
],
[
0.11517333984375,
0.212158203125,
0.490234375,
0.21337890625,
0.490234375,
0.2279052734375,
0.11517333984375,
0.2266845703125
],
[
0.5087890625,
0.2183837890625,
0.88134765625,
0.2183837890625,
0.88134765625,
0.231689453125,
0.5087890625,
0.231689453125
],
[
0.11517333984375,
0.2298583984375,
0.488525390625,
0.2298583984375,
0.488525390625,
0.2442626953125,
0.11517333984375,
0.2442626953125
],
[
0.5087890625,
0.23486328125,
0.88134765625,
0.23486328125,
0.88134765625,
0.248779296875,
0.5087890625,
0.248779296875
],
[
0.11517333984375,
0.2474365234375,
0.488525390625,
0.2474365234375,
0.488525390625,
0.261474609375,
0.11517333984375,
0.261474609375
],
[
0.50634765625,
0.251953125,
0.8798828125,
0.251953125,
0.8798828125,
0.26513671875,
0.50634765625,
0.26513671875
],
[
0.11517333984375,
0.264404296875,
0.238525390625,
0.264404296875,
0.238525390625,
0.27783203125,
0.11517333984375,
0.27783203125
],
[
0.5107421875,
0.26953125,
0.8798828125,
0.26953125,
0.8798828125,
0.28271484375,
0.5107421875,
0.28271484375
],
[
0.13232421875,
0.2822265625,
0.490234375,
0.2822265625,
0.490234375,
0.296630859375,
0.13232421875,
0.296630859375
],
[
0.5087890625,
0.287353515625,
0.88232421875,
0.287353515625,
0.88232421875,
0.300537109375,
0.5087890625,
0.300537109375
],
[
0.11358642578125,
0.2998046875,
0.494384765625,
0.2998046875,
0.494384765625,
0.313232421875,
0.11358642578125,
0.313232421875
],
[
0.5087890625,
0.3037109375,
0.88232421875,
0.3037109375,
0.88232421875,
0.317626953125,
0.5087890625,
0.317626953125
],
[
0.11517333984375,
0.31640625,
0.491943359375,
0.31640625,
0.491943359375,
0.330078125,
0.11517333984375,
0.330078125
],
[
0.50732421875,
0.32080078125,
0.88232421875,
0.32080078125,
0.88232421875,
0.333984375,
0.50732421875,
0.333984375
],
[
0.11358642578125,
0.333251953125,
0.490234375,
0.33203125,
0.490234375,
0.3466796875,
0.11358642578125,
0.347900390625
],
[
0.5048828125,
0.3359375,
0.88134765625,
0.337158203125,
0.88134765625,
0.3515625,
0.5048828125,
0.35107421875
],
[
0.11517333984375,
0.35107421875,
0.490234375,
0.35107421875,
0.490234375,
0.3642578125,
0.11517333984375,
0.3642578125
],
[
0.5087890625,
0.354736328125,
0.68896484375,
0.354736328125,
0.68896484375,
0.369384765625,
0.5087890625,
0.369384765625
],
[
0.11358642578125,
0.3662109375,
0.325927734375,
0.3662109375,
0.325927734375,
0.380126953125,
0.11358642578125,
0.380126953125
],
[
0.5107421875,
0.380126953125,
0.88134765625,
0.380126953125,
0.88134765625,
0.39453125,
0.5107421875,
0.39453125
],
[
0.1168212890625,
0.39453125,
0.494384765625,
0.39453125,
0.494384765625,
0.4091796875,
0.1168212890625,
0.4091796875
],
[
0.50732421875,
0.397705078125,
0.88232421875,
0.397705078125,
0.88232421875,
0.412353515625,
0.50732421875,
0.412353515625
],
[
0.11517333984375,
0.410888671875,
0.491943359375,
0.412353515625,
0.491943359375,
0.4267578125,
0.11517333984375,
0.425537109375
],
[
0.50634765625,
0.415283203125,
0.6748046875,
0.415283203125,
0.6748046875,
0.4287109375,
0.50634765625,
0.4287109375
],
[
0.11517333984375,
0.429931640625,
0.488525390625,
0.429931640625,
0.488525390625,
0.443115234375,
0.11517333984375,
0.443115234375
],
[
0.1168212890625,
0.4462890625,
0.491943359375,
0.4462890625,
0.491943359375,
0.4609375,
0.1168212890625,
0.4609375
],
[
0.5048828125,
0.442626953125,
0.88134765625,
0.442626953125,
0.88134765625,
0.45703125,
0.5048828125,
0.45703125
],
[
0.11517333984375,
0.46337890625,
0.488525390625,
0.46337890625,
0.488525390625,
0.477783203125,
0.11517333984375,
0.477783203125
],
[
0.52783203125,
0.462158203125,
0.54345703125,
0.462158203125,
0.54345703125,
0.472900390625,
0.52783203125,
0.472900390625
],
[
0.11517333984375,
0.48095703125,
0.490234375,
0.48095703125,
0.490234375,
0.494384765625,
0.11517333984375,
0.494384765625
],
[
0.5087890625,
0.479736328125,
0.88134765625,
0.479736328125,
0.88134765625,
0.494384765625,
0.5087890625,
0.494384765625
],
[
0.11358642578125,
0.496826171875,
0.1641845703125,
0.496826171875,
0.1641845703125,
0.51220703125,
0.11358642578125,
0.51220703125
],
[
0.52685546875,
0.49755859375,
0.88232421875,
0.49755859375,
0.88232421875,
0.51123046875,
0.52685546875,
0.51123046875
],
[
0.52783203125,
0.51513671875,
0.60693359375,
0.51513671875,
0.60693359375,
0.5263671875,
0.52783203125,
0.5263671875
],
[
0.1168212890625,
0.52783203125,
0.488525390625,
0.52783203125,
0.488525390625,
0.54248046875,
0.1168212890625,
0.54248046875
],
[
0.1282958984375,
0.54541015625,
0.313720703125,
0.54345703125,
0.313720703125,
0.55810546875,
0.1282958984375,
0.56005859375
],
[
0.5107421875,
0.54052734375,
0.8798828125,
0.54052734375,
0.8798828125,
0.5546875,
0.5107421875,
0.5546875
],
[
0.1168212890625,
0.56494140625,
0.491943359375,
0.56494140625,
0.491943359375,
0.5791015625,
0.1168212890625,
0.5791015625
],
[
0.5107421875,
0.5595703125,
0.67236328125,
0.5595703125,
0.67236328125,
0.57373046875,
0.5107421875,
0.57373046875
],
[
0.525390625,
0.57470703125,
0.88134765625,
0.57568359375,
0.88134765625,
0.59033203125,
0.525390625,
0.5888671875
],
[
0.1282958984375,
0.58203125,
0.2900390625,
0.58056640625,
0.2900390625,
0.59521484375,
0.1282958984375,
0.5966796875
],
[
0.5087890625,
0.59228515625,
0.8798828125,
0.59326171875,
0.8798828125,
0.60791015625,
0.5087890625,
0.60693359375
],
[
0.1168212890625,
0.6015625,
0.490234375,
0.6015625,
0.490234375,
0.6162109375,
0.1168212890625,
0.6162109375
],
[
0.5087890625,
0.609375,
0.88232421875,
0.60986328125,
0.88232421875,
0.625,
0.5087890625,
0.62353515625
],
[
0.131591796875,
0.619140625,
0.484375,
0.619140625,
0.484375,
0.63330078125,
0.131591796875,
0.63330078125
],
[
0.5087890625,
0.62744140625,
0.88232421875,
0.62744140625,
0.88232421875,
0.6416015625,
0.5087890625,
0.6416015625
],
[
0.131591796875,
0.63720703125,
0.16259765625,
0.63720703125,
0.16259765625,
0.6494140625,
0.131591796875,
0.6494140625
],
[
0.5107421875,
0.64453125,
0.814453125,
0.64453125,
0.814453125,
0.6591796875,
0.5107421875,
0.6591796875
],
[
0.1168212890625,
0.65576171875,
0.486083984375,
0.65576171875,
0.486083984375,
0.67041015625,
0.1168212890625,
0.67041015625
],
[
0.1298828125,
0.671875,
0.3203125,
0.673828125,
0.3203125,
0.689453125,
0.1298828125,
0.6875
],
[
0.5087890625,
0.671875,
0.80126953125,
0.671875,
0.80126953125,
0.68603515625,
0.5087890625,
0.68603515625
],
[
0.51123046875,
0.6943359375,
0.8837890625,
0.6943359375,
0.8837890625,
0.708984375,
0.51123046875,
0.708984375
],
[
0.11358642578125,
0.7021484375,
0.491943359375,
0.70263671875,
0.491943359375,
0.71728515625,
0.11358642578125,
0.71630859375
],
[
0.5107421875,
0.7119140625,
0.88232421875,
0.7119140625,
0.88232421875,
0.72607421875,
0.5107421875,
0.72607421875
],
[
0.1168212890625,
0.72021484375,
0.256591796875,
0.72021484375,
0.256591796875,
0.73486328125,
0.1168212890625,
0.73486328125
],
[
0.50732421875,
0.72900390625,
0.88232421875,
0.72802734375,
0.88232421875,
0.7421875,
0.50732421875,
0.74365234375
],
[
0.11517333984375,
0.74560546875,
0.494384765625,
0.74560546875,
0.494384765625,
0.7587890625,
0.11517333984375,
0.7587890625
],
[
0.50634765625,
0.74560546875,
0.88232421875,
0.744140625,
0.88232421875,
0.7587890625,
0.50634765625,
0.76025390625
],
[
0.11517333984375,
0.7587890625,
0.490234375,
0.7587890625,
0.490234375,
0.77294921875,
0.11517333984375,
0.77294921875
],
[
0.50634765625,
0.76318359375,
0.88232421875,
0.76318359375,
0.88232421875,
0.7763671875,
0.50634765625,
0.7763671875
],
[
0.11358642578125,
0.7734375,
0.488525390625,
0.7734375,
0.488525390625,
0.787109375,
0.11358642578125,
0.787109375
],
[
0.5087890625,
0.77978515625,
0.88134765625,
0.77978515625,
0.88134765625,
0.79443359375,
0.5087890625,
0.79443359375
],
[
0.1168212890625,
0.7890625,
0.488525390625,
0.7890625,
0.488525390625,
0.7998046875,
0.1168212890625,
0.7998046875
],
[
0.11358642578125,
0.8017578125,
0.2344970703125,
0.8017578125,
0.2344970703125,
0.81494140625,
0.11358642578125,
0.81494140625
],
[
0.5087890625,
0.79736328125,
0.88134765625,
0.79736328125,
0.88134765625,
0.81103515625,
0.5087890625,
0.81103515625
],
[
0.1356201171875,
0.81494140625,
0.491943359375,
0.81494140625,
0.491943359375,
0.8291015625,
0.1356201171875,
0.8291015625
],
[
0.5087890625,
0.814453125,
0.88134765625,
0.814453125,
0.88134765625,
0.8291015625,
0.5087890625,
0.8291015625
],
[
0.1168212890625,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84130859375,
0.1168212890625,
0.84130859375
],
[
0.11358642578125,
0.84326171875,
0.494384765625,
0.84326171875,
0.494384765625,
0.85693359375,
0.11358642578125,
0.85693359375
],
[
0.5087890625,
0.84130859375,
0.689453125,
0.84130859375,
0.689453125,
0.85498046875,
0.5087890625,
0.85498046875
],
[
0.11358642578125,
0.85693359375,
0.490234375,
0.85693359375,
0.490234375,
0.87060546875,
0.11358642578125,
0.87060546875
],
[
0.5107421875,
0.86279296875,
0.8837890625,
0.86279296875,
0.8837890625,
0.87744140625,
0.5107421875,
0.87744140625
],
[
0.1168212890625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.88330078125,
0.1168212890625,
0.88330078125
],
[
0.11517333984375,
0.88525390625,
0.491943359375,
0.88525390625,
0.491943359375,
0.8984375,
0.11517333984375,
0.8984375
],
[
0.5087890625,
0.880859375,
0.88134765625,
0.880859375,
0.88134765625,
0.89599609375,
0.5087890625,
0.89599609375
],
[
0.11358642578125,
0.8984375,
0.22216796875,
0.8984375,
0.22216796875,
0.912109375,
0.11358642578125,
0.912109375
],
[
0.5107421875,
0.89794921875,
0.88134765625,
0.89794921875,
0.88134765625,
0.912109375,
0.5107421875,
0.912109375
]
] |
[
"Our transition set is closely related to the op-",
"3.2 Generator Transitions",
"erations used in Earley's algorithm which likewise",
"The parsing algorithm that maps from sequences",
"introduces nonterminals symbols with its PREDICT",
"of words to parse trees can be adapted with mi-",
"operation and later COMPLETEs them after consum-",
"nor changes to produce an algorithm that stochas-",
"ing terminal symbols one at a time using scAN",
"tically generates trees and terminal symbols. Two",
"(Earley, 1970). It is likewise closely related to the",
"changes are required: (i) there is no input buffer of",
"\"linearized\" parse trees proposed by Vinyals et al.",
"unprocessed words, rather there is an output buffer",
"(2015) and to the top-down, left-to-right decompo-",
"(T), and (ii) instead of a SHIFT operation there are",
"sitions of trees used in previous generative parsing",
"GEN() operations which generate terminal symbol",
"and language modeling work (Roark, 2001, 2004;",
" α E and add it to the top of the stack and the out",
"Charniak, 2010).",
"put buffer. At each timestep an action is stochasti-",
"A further connection is to LL(*) parsing which",
"cally selected according to a conditional distribution",
"uses an unbounded lookahead (compactly repre-",
"that depends on the current contents of S and T. The",
"sented by a DFA) to distinguish between parse alter-",
"algorithm terminates when a single completed con-",
"natives in a top-down parser (Parr and Fisher, 2011);",
"stituent remains on the stack. Fig. 4 shows an exam-",
"however, our parser uses an RNN encoding of the",
"ple generation sequence.",
"lookahead rather than a DFA.",
"Constraints on generator transitions.The gen-",
"Constraints on parser transitions. To guarantee",
"eration algorithm also requires slightly modified",
"that only well-formed phrase-structure trees are pro-",
"constraints. These are:",
"duced by the parser, we impose the following con-",
"straints on the transitions that can be applied at each",
". The GEN() operation can only be applied if n ≥",
"step which are a function of the parser state (B, S, n)",
"1.",
"where n is the number of open nonterminals on the",
"· The REDUCE operation can only be applied if the",
"stack:",
"top of the stack is not an open nonterminal symbol",
"and n > 1.",
". The NT(X) operation can only be applied if B is",
"not empty and n < 100.4",
"To designate the set of valid generator transitions",
". The SHIFT operation can only be applied if B is",
"we write AG(T, S, n).",
"This transition set generates trees using nearly the",
"not empty and n ≥ 1.",
"same structure building actions and stack configura-",
"The REDUCE operation can only be applied if the",
"tions as the“top-down PDA\" construction proposed",
"top of the stack is not an open nonterminal sym",
"by Abney et al. (1999), albeit without the restriction",
"bol.",
"that the trees be in Chomsky normal form.",
" The REDUCE operation can only be applied if n >",
"2 or if the buffer is empty.",
"3.3 Transition Sequences from Trees",
"Any parse tree can be converted to a sequence of",
"To designate the set of valid parser transitions, we",
"transitions via a depth-first, left-to-right traversal of",
"write AD(B, S, n).",
"a parse tree. Since there is a unique depth-first, left-",
"quires no special handling. However, leaving them out reduces",
"ro-right traversal of a tree, there is exactly one tran-",
"the number of transitions by O(n) and also reduces the number",
"sition sequence of each tree. For a tree y and a",
"of action types, both of which reduce the runtime. Furthermore,",
"sequence of symbols a, we write a(aα, y) to indi-",
"standard parsing evaluation scores do notdependon preterminal",
"prediction accuracy",
"cate the corresponding sequence of generation tran-",
"4since our parser allows unary nonterminal productions,",
"sitions, and b(α, y) to indicate the parser transitions.",
"there are an infinite number of valid trees for finite-length sen-",
"tences.The n < 100 constraint prevents the classifier from",
"3.4Runtime Analysis",
"misbehaving and generating excessively large numbers of non-",
"A detailed analysis of the algorithmic properties of",
"terminals. Similar constraints have been proposed todeal with",
"the analogous problem in bottom-up shift-reduce parsers (Sagae",
"our top-down parser is beyond the scope of this pa-",
"and Lavie, 2005).",
"per; however, we briefly state several facts. As-"
] |
{
"arxiv_id": "1602.07776",
"language": "en",
"page_idx": 2
}
|
||
1602.07776-en-3
|
1602.07776-en-3.jpg
|
$\overline{S}$
<table>
<tbody>
<tr>
<th>$\mathbf{ck}_t$</th>
<th>$Buffer_i$</th>
<th>Open $NTs_t$</th>
<th>$\mathbf{Action}$</th>
<th>$t+1$</th>
<th>$\mathbf{Buffer}$ $\dot{\cdot}t+1$</th>
<th>Open $\mathbf{NTs}_t$</th>
</tr>
<tr>
<td> </td>
<td>$B$</td>
<td>$m$</td>
<td>(X $NT$</td>
<td>A</td>
<td>$\overline{B}$</td>
<td>$n+1$</td>
</tr>
<tr>
<td> </td>
<td>$B$ $x$</td>
<td>$m$</td>
<td>SHIFT</td>
<td> </td>
<td>$B$</td>
<td>$m$</td>
</tr>
<tr>
<td>$(\mathbf{X}$ $\tau_{1}$ $\tau_{\ell}$</td>
<td>$B$</td>
<td>$m$</td>
<td>REDUCE</td>
<td>$\tau_{1}$ $\tau_{f}$</td>
<td>$B$</td>
<td>一 1 $m$ </td>
</tr>
</tbody>
</table>
$S$
Figure 1: Parser transitions showing the stack, buffer, and open nonterminal count before and after each action type. $S$ represents the stack, which contains open nonterminals and completed subtrees; $B$ represents the buffer of unprocessed terminal symbols; $x$ is a terminal symbol, X is a nonterminal symbol, and each $\tau$ is a completed subtree. The top of the stack is to the right, and the buffer is consumed from left to right. Elements on the stack and buffer are delimited by a vertical bar(|).
<table>
<tbody>
<tr>
<th> </th>
<th>Stack</th>
<th> </th>
<th> </th>
<th>$\mathbf{Action}$</th>
</tr>
<tr>
<td>0</td>
<td> </td>
<td>cat meows</td>
<td> ·</td>
<td>$NT(S)$</td>
</tr>
<tr>
<td></td>
<td>(S</td>
<td>cat meows</td>
<td>1 ·</td>
<td>NT( NP) </td>
</tr>
<tr>
<td>2</td>
<td>(S$\mid($NP</td>
<td>cat meows</td>
<td>1 ·</td>
<td>SHIFT</td>
</tr>
<tr>
<td>3</td>
<td>(S$\mid($NP$\mid The$</td>
<td>meows</td>
<td> </td>
<td>SHIFT</td>
</tr>
<tr>
<td>4</td>
<td>(S$\mid($NP$\mid The\mid hungry$</td>
<td> </td>
<td> </td>
<td>SHIFT</td>
</tr>
<tr>
<td></td>
<td>(S$\mid$NP$\mid$The$\mid$hungry$\mid \textit{c}at$</td>
<td> </td>
<td> </td>
<td>REDUCE</td>
</tr>
<tr>
<td>6</td>
<td>$(\mathcal{S}\mid($NP The hungry cat)</td>
<td> </td>
<td> </td>
<td>$\operatorname{NT}(\operatorname{VP})$</td>
</tr>
<tr>
<td>7</td>
<td>(S$\mid($NP The hungry cat$)\mid($VP</td>
<td> </td>
<td> </td>
<td>SHIFT</td>
</tr>
<tr>
<td></td>
<td>(S$\mid($NP The hungry cat$)\mid($VP pmeows</td>
<td> </td>
<td> </td>
<td>REDUCE</td>
</tr>
<tr>
<td>9</td>
<td>(S$\mid($NP The hungry cat$)\mid($VP meows)</td>
<td> </td>
<td> </td>
<td>SHIFT</td>
</tr>
<tr>
<td>10</td>
<td>(S$\mid($NP The hungry cat$)\mid($VP meows$)\mid$</td>
<td> </td>
<td> </td>
<td>REDUCE</td>
</tr>
<tr>
<td>1 1</td>
<td>$$($S (NP The hungry cat) (VP meows) .)$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>
Figure 2: Top-down parsing example.
$\underline{\mathbf{s}_{t+1}}$
$S$
$\textbf{Stack}_{t}$
<table>
<tbody>
<tr>
<th>$it$</th>
<th>$\mathbf{Terms}_{t}$</th>
<th>Open $\mathbf{NTs}_t$</th>
<th>$\mathbf{Action}$</th>
<th>$\mathbf{Stack}_{t+1}$</th>
<th>1</th>
<th>$\mathbf{Terms}_{t+1}$</th>
<th>Open $N$</th>
</tr>
<tr>
<td> </td>
<td>$T$</td>
<td>$m$</td>
<td>(X) $NT$</td>
<td>$\overline{S\mid(\mathrm{X}}$</td>
<td> </td>
<td>$T$</td>
<td>$n$ +1</td>
</tr>
<tr>
<td> </td>
<td>$T$</td>
<td>$m$</td>
<td>$GEN$ $x$</td>
<td>$S$ $x$</td>
<td> </td>
<td>$T$ $x$</td>
<td>$m$</td>
</tr>
<tr>
<td>$\tau_{1}$ $\tau_{\ell}$</td>
<td>$T$</td>
<td>$m$</td>
<td>REDUCE</td>
<td>$S$ $\mid($X $\tau_{1}$</td>
<td>$\tau_{\ell}$</td>
<td>$T$</td>
<td>1 $m$ 一</td>
</tr>
</tbody>
</table>
$S$
Figure 3: Generator transitions. Symbols defined as in Fig. 1 with the addition of $T$ representing the history of generated terminals
<table>
<tbody>
<tr>
<th> </th>
<th>Stack</th>
<th>$\mathbf{Terminals}$</th>
<th> </th>
<th> </th>
<th>$\mathbf{Action}$</th>
</tr>
<tr>
<td>0</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>$NT(S)$</td>
</tr>
<tr>
<td></td>
<td>(S</td>
<td> </td>
<td> </td>
<td> </td>
<td>$\operatorname{NT}(\operatorname{NP})$</td>
</tr>
<tr>
<td>2</td>
<td>(S$\mid($NP</td>
<td> </td>
<td> </td>
<td> </td>
<td>The $GEN$</td>
</tr>
<tr>
<td>3</td>
<td>(S$\mid($NP $P\mid The$</td>
<td>The</td>
<td> </td>
<td> </td>
<td>$(hungry$ $GEN$</td>
</tr>
<tr>
<td>4</td>
<td>(S$\mid($NP$\mid The\mid hungry$</td>
<td>The hungry</td>
<td> </td>
<td> </td>
<td>$GEN$ cat</td>
</tr>
<tr>
<td></td>
<td>(S$\mid$NP$\mid$The$\mid$hungry$\mid \textit{c}at$</td>
<td>The hungry</td>
<td>${:}at$</td>
<td> </td>
<td>REDUCE</td>
</tr>
<tr>
<td>6</td>
<td>$(\mathcal{S}\mid($NP The hungry cat)</td>
<td>The hungry</td>
<td>${:}at$</td>
<td> </td>
<td>$\operatorname{NT}(\operatorname{VP})$</td>
</tr>
<tr>
<td>7</td>
<td>(S$\mid($NP The hungry cat$)\mid($VP</td>
<td>The hungry</td>
<td>${:}at$</td>
<td> </td>
<td>$GEN$ meows</td>
</tr>
<tr>
<td></td>
<td>(S$\mid($NP The hungry cat$)\mid($VP meows</td>
<td>The hungry</td>
<td>$:at$ $m\epsilon$</td>
<td>neows</td>
<td>REDUCE</td>
</tr>
<tr>
<td>9</td>
<td>(S$\mid($NP The hungry cat$)\mid($VP meows)</td>
<td>The hungry</td>
<td>$:at$ $m\epsilon$</td>
<td>neows</td>
<td>$GEN$</td>
</tr>
<tr>
<td></td>
<td>(S$\mid($NP The hungry cat$)\mid($VP meows$)\mid.$</td>
<td>The hungry</td>
<td>$:at$ $m\epsilon$ 、</td>
<td>neows</td>
<td>REDUCE</td>
</tr>
<tr>
<td> </td>
<td>$$($S (NP The hungrv cat) (VP meows) .)$</td>
<td>The hungry</td>
<td>${:}at$ $m\epsilon$</td>
<td>neows</td>
<td> </td>
</tr>
</tbody>
</table>
Figure 4: Joint generation of a parse tree and sentence.
|
[
[
0.54345703125,
0.0921630859375,
0.61669921875,
0.0953369140625,
0.615234375,
0.11297607421875,
0.54150390625,
0.10986328125
],
[
0.69140625,
0.0921630859375,
0.765625,
0.0953369140625,
0.7646484375,
0.11297607421875,
0.689453125,
0.10986328125
],
[
0.779296875,
0.0921630859375,
0.89208984375,
0.0947265625,
0.89208984375,
0.11297607421875,
0.779296875,
0.111083984375
],
[
0.461669921875,
0.09344482421875,
0.51953125,
0.0953369140625,
0.5185546875,
0.111083984375,
0.4599609375,
0.10919189453125
],
[
0.285888671875,
0.09344482421875,
0.346435546875,
0.0953369140625,
0.3447265625,
0.111083984375,
0.28515625,
0.10919189453125
],
[
0.361083984375,
0.0947265625,
0.44775390625,
0.0921630859375,
0.44775390625,
0.10986328125,
0.361083984375,
0.11236572265625
],
[
0.1192626953125,
0.11236572265625,
0.143798828125,
0.11236572265625,
0.143798828125,
0.1300048828125,
0.1192626953125,
0.1300048828125
],
[
0.12335205078125,
0.0953369140625,
0.176513671875,
0.0953369140625,
0.176513671875,
0.111083984375,
0.12335205078125,
0.111083984375
],
[
0.285888671875,
0.11297607421875,
0.3056640625,
0.11297607421875,
0.3056640625,
0.1287841796875,
0.285888671875,
0.1287841796875
],
[
0.54345703125,
0.11236572265625,
0.6005859375,
0.11236572265625,
0.6005859375,
0.1307373046875,
0.54345703125,
0.1307373046875
],
[
0.689453125,
0.11297607421875,
0.7099609375,
0.11297607421875,
0.7099609375,
0.1287841796875,
0.689453125,
0.1287841796875
],
[
0.463134765625,
0.1142578125,
0.515625,
0.1142578125,
0.515625,
0.1300048828125,
0.463134765625,
0.1300048828125
],
[
0.359375,
0.1142578125,
0.37744140625,
0.1142578125,
0.37744140625,
0.1307373046875,
0.359375,
0.1307373046875
],
[
0.78125,
0.1142578125,
0.8310546875,
0.1142578125,
0.8310546875,
0.1300048828125,
0.78125,
0.1300048828125
],
[
0.5458984375,
0.1287841796875,
0.5703125,
0.1287841796875,
0.5703125,
0.146484375,
0.5458984375,
0.146484375
],
[
0.68896484375,
0.1287841796875,
0.71142578125,
0.1287841796875,
0.71142578125,
0.147705078125,
0.68896484375,
0.147705078125
],
[
0.281982421875,
0.1300048828125,
0.330810546875,
0.1275634765625,
0.33251953125,
0.146484375,
0.283447265625,
0.1483154296875
],
[
0.125,
0.1300048828125,
0.1422119140625,
0.1300048828125,
0.1422119140625,
0.146484375,
0.125,
0.146484375
],
[
0.359375,
0.1319580078125,
0.37744140625,
0.1319580078125,
0.37744140625,
0.147705078125,
0.359375,
0.147705078125
],
[
0.464111328125,
0.1307373046875,
0.5146484375,
0.1307373046875,
0.5146484375,
0.146484375,
0.464111328125,
0.146484375
],
[
0.57177734375,
0.1319580078125,
0.59326171875,
0.1319580078125,
0.59326171875,
0.146484375,
0.57177734375,
0.146484375
],
[
0.78125,
0.1319580078125,
0.79833984375,
0.1319580078125,
0.79833984375,
0.147705078125,
0.78125,
0.147705078125
],
[
0.1217041015625,
0.1458740234375,
0.275390625,
0.146484375,
0.275390625,
0.1641845703125,
0.1217041015625,
0.1634521484375
],
[
0.265625,
0.146484375,
0.306396484375,
0.146484375,
0.306396484375,
0.1622314453125,
0.265625,
0.1622314453125
],
[
0.68896484375,
0.147705078125,
0.708984375,
0.147705078125,
0.708984375,
0.1622314453125,
0.68896484375,
0.1622314453125
],
[
0.78125,
0.147705078125,
0.830078125,
0.147705078125,
0.830078125,
0.1634521484375,
0.78125,
0.1634521484375
],
[
0.463134765625,
0.1483154296875,
0.53173828125,
0.1483154296875,
0.53173828125,
0.1634521484375,
0.463134765625,
0.1634521484375
],
[
0.359375,
0.1483154296875,
0.37744140625,
0.1483154296875,
0.37744140625,
0.1634521484375,
0.359375,
0.1634521484375
],
[
0.5458984375,
0.1483154296875,
0.6748046875,
0.1483154296875,
0.6748046875,
0.1634521484375,
0.5458984375,
0.1634521484375
],
[
0.11517333984375,
0.1767578125,
0.8837890625,
0.1767578125,
0.8837890625,
0.1900634765625,
0.11517333984375,
0.1900634765625
],
[
0.11517333984375,
0.1944580078125,
0.8837890625,
0.1944580078125,
0.8837890625,
0.2083740234375,
0.11517333984375,
0.2083740234375
],
[
0.11517333984375,
0.2108154296875,
0.88232421875,
0.2108154296875,
0.88232421875,
0.2247314453125,
0.11517333984375,
0.2247314453125
],
[
0.11358642578125,
0.2266845703125,
0.734375,
0.2279052734375,
0.734375,
0.242431640625,
0.11358642578125,
0.2412109375
],
[
0.2034912109375,
0.2587890625,
0.4306640625,
0.260009765625,
0.4306640625,
0.274658203125,
0.2034912109375,
0.2734375
],
[
0.2181396484375,
0.27978515625,
0.26708984375,
0.27978515625,
0.26708984375,
0.29541015625,
0.2181396484375,
0.29541015625
],
[
0.52392578125,
0.27978515625,
0.5810546875,
0.27978515625,
0.5810546875,
0.29541015625,
0.52392578125,
0.29541015625
],
[
0.740234375,
0.276611328125,
0.7998046875,
0.279052734375,
0.79833984375,
0.29736328125,
0.73876953125,
0.29541015625
],
[
0.52685546875,
0.298583984375,
0.7236328125,
0.298583984375,
0.7236328125,
0.313232421875,
0.52685546875,
0.313232421875
],
[
0.740234375,
0.298583984375,
0.7900390625,
0.298583984375,
0.7900390625,
0.314453125,
0.740234375,
0.314453125
],
[
0.525390625,
0.314453125,
0.7197265625,
0.314453125,
0.7197265625,
0.33203125,
0.525390625,
0.33203125
],
[
0.1953125,
0.2998046875,
0.20751953125,
0.3017578125,
0.2034912109375,
0.31494140625,
0.191162109375,
0.313232421875
],
[
0.2197265625,
0.31494140625,
0.240234375,
0.31494140625,
0.240234375,
0.330810546875,
0.2197265625,
0.330810546875
],
[
0.74169921875,
0.31494140625,
0.80224609375,
0.31494140625,
0.80224609375,
0.330810546875,
0.74169921875,
0.330810546875
],
[
0.193603515625,
0.3212890625,
0.200927734375,
0.32080078125,
0.2017822265625,
0.32763671875,
0.1953125,
0.328857421875
],
[
0.2181396484375,
0.330078125,
0.281005859375,
0.330078125,
0.281005859375,
0.348388671875,
0.2181396484375,
0.348388671875
],
[
0.52783203125,
0.333251953125,
0.7236328125,
0.333251953125,
0.7236328125,
0.347900390625,
0.52783203125,
0.347900390625
],
[
0.74169921875,
0.33203125,
0.79150390625,
0.33203125,
0.79150390625,
0.347900390625,
0.74169921875,
0.347900390625
],
[
0.2197265625,
0.347900390625,
0.31787109375,
0.347900390625,
0.31787109375,
0.365478515625,
0.2197265625,
0.365478515625
],
[
0.1993408203125,
0.337158203125,
0.2059326171875,
0.34228515625,
0.1993408203125,
0.347900390625,
0.19287109375,
0.34228515625
],
[
0.525390625,
0.348388671875,
0.68896484375,
0.348388671875,
0.68896484375,
0.3662109375,
0.525390625,
0.3662109375
],
[
0.74169921875,
0.349853515625,
0.79150390625,
0.349853515625,
0.79150390625,
0.3642578125,
0.74169921875,
0.3642578125
],
[
0.1953125,
0.35595703125,
0.2017822265625,
0.35595703125,
0.2017822265625,
0.3623046875,
0.1953125,
0.3623046875
],
[
0.221435546875,
0.367431640625,
0.375732421875,
0.367431640625,
0.375732421875,
0.3818359375,
0.221435546875,
0.3818359375
],
[
0.525390625,
0.3662109375,
0.62353515625,
0.3642578125,
0.62353515625,
0.3818359375,
0.525390625,
0.3837890625
],
[
0.1968994140625,
0.37255859375,
0.2017822265625,
0.37255859375,
0.2017822265625,
0.380126953125,
0.1968994140625,
0.380126953125
],
[
0.74169921875,
0.367431640625,
0.79150390625,
0.367431640625,
0.79150390625,
0.3818359375,
0.74169921875,
0.3818359375
],
[
0.2181396484375,
0.3818359375,
0.410888671875,
0.38330078125,
0.410888671875,
0.40087890625,
0.2181396484375,
0.399658203125
],
[
0.52685546875,
0.3837890625,
0.5888671875,
0.3837890625,
0.5888671875,
0.399658203125,
0.52685546875,
0.399658203125
],
[
0.74169921875,
0.3837890625,
0.8095703125,
0.3837890625,
0.8095703125,
0.3984375,
0.74169921875,
0.3984375
],
[
0.2181396484375,
0.3984375,
0.39794921875,
0.399658203125,
0.39794921875,
0.417236328125,
0.2181396484375,
0.416748046875
],
[
0.193603515625,
0.386962890625,
0.205078125,
0.386962890625,
0.205078125,
0.397705078125,
0.193603515625,
0.397705078125
],
[
0.525390625,
0.401611328125,
0.59228515625,
0.401611328125,
0.59228515625,
0.416748046875,
0.525390625,
0.416748046875
],
[
0.74169921875,
0.40087890625,
0.80224609375,
0.40087890625,
0.80224609375,
0.416748046875,
0.74169921875,
0.416748046875
],
[
0.221435546875,
0.41845703125,
0.435546875,
0.41845703125,
0.435546875,
0.43310546875,
0.221435546875,
0.43310546875
],
[
0.193603515625,
0.40478515625,
0.205078125,
0.40478515625,
0.205078125,
0.4140625,
0.193603515625,
0.4140625
],
[
0.525390625,
0.41845703125,
0.59228515625,
0.41845703125,
0.59228515625,
0.43310546875,
0.525390625,
0.43310546875
],
[
0.74169921875,
0.41845703125,
0.7900390625,
0.41845703125,
0.7900390625,
0.43310546875,
0.74169921875,
0.43310546875
],
[
0.2181396484375,
0.43310546875,
0.491943359375,
0.434326171875,
0.491943359375,
0.451904296875,
0.2181396484375,
0.45068359375
],
[
0.1953125,
0.423583984375,
0.2017822265625,
0.423583984375,
0.2017822265625,
0.429931640625,
0.1953125,
0.429931640625
],
[
0.1968994140625,
0.438232421875,
0.205078125,
0.44140625,
0.200927734375,
0.448974609375,
0.19287109375,
0.44580078125
],
[
0.74267578125,
0.4375,
0.80810546875,
0.4375,
0.80810546875,
0.448974609375,
0.74267578125,
0.448974609375
],
[
0.2181396484375,
0.449462890625,
0.495849609375,
0.45068359375,
0.495849609375,
0.468505859375,
0.2181396484375,
0.46728515625
],
[
0.1953125,
0.45703125,
0.2017822265625,
0.45703125,
0.2017822265625,
0.464111328125,
0.1953125,
0.464111328125
],
[
0.74169921875,
0.45263671875,
0.7900390625,
0.45263671875,
0.7900390625,
0.46728515625,
0.74169921875,
0.46728515625
],
[
0.221435546875,
0.468505859375,
0.50732421875,
0.468505859375,
0.50732421875,
0.486083984375,
0.221435546875,
0.486083984375
],
[
0.74267578125,
0.4697265625,
0.80810546875,
0.4697265625,
0.80810546875,
0.484130859375,
0.74267578125,
0.484130859375
],
[
0.2181396484375,
0.484130859375,
0.498291015625,
0.48486328125,
0.498291015625,
0.50244140625,
0.2181396484375,
0.501953125
],
[
0.1871337890625,
0.473388671875,
0.200927734375,
0.47021484375,
0.205078125,
0.481689453125,
0.191162109375,
0.48486328125
],
[
0.191162109375,
0.490478515625,
0.2017822265625,
0.490478515625,
0.2017822265625,
0.5,
0.191162109375,
0.5
],
[
0.386474609375,
0.51025390625,
0.611328125,
0.51025390625,
0.611328125,
0.5244140625,
0.386474609375,
0.5244140625
],
[
0.54345703125,
0.54248046875,
0.61376953125,
0.54541015625,
0.61279296875,
0.56103515625,
0.54150390625,
0.55810546875
],
[
0.689453125,
0.54248046875,
0.7646484375,
0.54541015625,
0.76318359375,
0.56103515625,
0.68896484375,
0.55810546875
],
[
0.77685546875,
0.54150390625,
0.88623046875,
0.54345703125,
0.88623046875,
0.5625,
0.77685546875,
0.56005859375
],
[
0.285888671875,
0.54443359375,
0.343017578125,
0.54443359375,
0.343017578125,
0.56005859375,
0.285888671875,
0.56005859375
],
[
0.12335205078125,
0.54443359375,
0.176513671875,
0.54443359375,
0.176513671875,
0.56005859375,
0.12335205078125,
0.56005859375
],
[
0.361083984375,
0.54541015625,
0.44677734375,
0.54541015625,
0.44677734375,
0.56005859375,
0.361083984375,
0.56005859375
],
[
0.463134765625,
0.54443359375,
0.5185546875,
0.54443359375,
0.5185546875,
0.56005859375,
0.463134765625,
0.56005859375
],
[
0.1192626953125,
0.56103515625,
0.143798828125,
0.56103515625,
0.143798828125,
0.580078125,
0.1192626953125,
0.580078125
],
[
0.283447265625,
0.56103515625,
0.306396484375,
0.56103515625,
0.306396484375,
0.580078125,
0.283447265625,
0.580078125
],
[
0.35693359375,
0.56298828125,
0.374267578125,
0.56298828125,
0.374267578125,
0.580078125,
0.35693359375,
0.580078125
],
[
0.461669921875,
0.56298828125,
0.5146484375,
0.56298828125,
0.5146484375,
0.5791015625,
0.461669921875,
0.5791015625
],
[
0.54345703125,
0.56298828125,
0.59619140625,
0.56298828125,
0.59619140625,
0.5791015625,
0.54345703125,
0.5791015625
],
[
0.68310546875,
0.56298828125,
0.70361328125,
0.5595703125,
0.708984375,
0.57568359375,
0.68896484375,
0.580078125
],
[
0.775390625,
0.5625,
0.826171875,
0.56005859375,
0.82666015625,
0.5791015625,
0.77685546875,
0.58056640625
],
[
0.283447265625,
0.5771484375,
0.306396484375,
0.5771484375,
0.306396484375,
0.5966796875,
0.283447265625,
0.5966796875
],
[
0.685546875,
0.5771484375,
0.73193359375,
0.5771484375,
0.73193359375,
0.5966796875,
0.685546875,
0.5966796875
],
[
0.358642578125,
0.580078125,
0.375732421875,
0.580078125,
0.375732421875,
0.59765625,
0.358642578125,
0.59765625
],
[
0.125,
0.5791015625,
0.1422119140625,
0.5791015625,
0.1422119140625,
0.59521484375,
0.125,
0.59521484375
],
[
0.461669921875,
0.580078125,
0.52294921875,
0.580078125,
0.52294921875,
0.59521484375,
0.461669921875,
0.59521484375
],
[
0.5458984375,
0.580078125,
0.56787109375,
0.580078125,
0.56787109375,
0.5947265625,
0.5458984375,
0.5947265625
],
[
0.56640625,
0.58203125,
0.5908203125,
0.58203125,
0.5908203125,
0.59521484375,
0.56640625,
0.59521484375
],
[
0.77685546875,
0.58056640625,
0.7958984375,
0.58056640625,
0.7958984375,
0.5966796875,
0.77685546875,
0.5966796875
],
[
0.1258544921875,
0.5966796875,
0.27294921875,
0.5966796875,
0.27294921875,
0.60986328125,
0.1258544921875,
0.60986328125
],
[
0.358642578125,
0.5966796875,
0.375732421875,
0.5966796875,
0.375732421875,
0.61279296875,
0.358642578125,
0.61279296875
],
[
0.461669921875,
0.5966796875,
0.53125,
0.5966796875,
0.53125,
0.611328125,
0.461669921875,
0.611328125
],
[
0.54345703125,
0.59765625,
0.67236328125,
0.59765625,
0.67236328125,
0.6123046875,
0.54345703125,
0.6123046875
],
[
0.6845703125,
0.5966796875,
0.705078125,
0.5966796875,
0.705078125,
0.60986328125,
0.6845703125,
0.60986328125
],
[
0.77294921875,
0.59765625,
0.82666015625,
0.59326171875,
0.830078125,
0.611328125,
0.775390625,
0.61572265625
],
[
0.1168212890625,
0.625,
0.88134765625,
0.625,
0.88134765625,
0.63818359375,
0.1168212890625,
0.63818359375
],
[
0.275390625,
0.599609375,
0.301513671875,
0.59326171875,
0.306396484375,
0.60693359375,
0.281005859375,
0.61279296875
],
[
0.200927734375,
0.65771484375,
0.25,
0.65771484375,
0.25,
0.673828125,
0.200927734375,
0.673828125
],
[
0.50634765625,
0.6572265625,
0.59228515625,
0.6572265625,
0.59228515625,
0.6748046875,
0.50634765625,
0.6748046875
],
[
0.7236328125,
0.6591796875,
0.779296875,
0.6591796875,
0.779296875,
0.6748046875,
0.7236328125,
0.6748046875
],
[
0.17724609375,
0.68115234375,
0.1895751953125,
0.68115234375,
0.1895751953125,
0.6923828125,
0.17724609375,
0.6923828125
],
[
0.72216796875,
0.6767578125,
0.77294921875,
0.6767578125,
0.77294921875,
0.6923828125,
0.72216796875,
0.6923828125
],
[
0.2034912109375,
0.6943359375,
0.22216796875,
0.6943359375,
0.22216796875,
0.708984375,
0.2034912109375,
0.708984375
],
[
0.7236328125,
0.6943359375,
0.78369140625,
0.6943359375,
0.78369140625,
0.708984375,
0.7236328125,
0.708984375
],
[
0.2034912109375,
0.71044921875,
0.26318359375,
0.71044921875,
0.26318359375,
0.72607421875,
0.2034912109375,
0.72607421875
],
[
0.1805419921875,
0.69970703125,
0.1854248046875,
0.69970703125,
0.1854248046875,
0.70703125,
0.1805419921875,
0.70703125
],
[
0.178955078125,
0.71533203125,
0.1895751953125,
0.71533203125,
0.1895751953125,
0.724609375,
0.178955078125,
0.724609375
],
[
0.7255859375,
0.7109375,
0.7998046875,
0.7109375,
0.7998046875,
0.72607421875,
0.7255859375,
0.72607421875
],
[
0.5087890625,
0.7265625,
0.54150390625,
0.7265625,
0.54150390625,
0.74365234375,
0.5087890625,
0.74365234375
],
[
0.7236328125,
0.724609375,
0.826171875,
0.72802734375,
0.82421875,
0.74560546875,
0.72216796875,
0.7421875
],
[
0.1829833984375,
0.73193359375,
0.1895751953125,
0.7373046875,
0.1829833984375,
0.7421875,
0.176513671875,
0.7373046875
],
[
0.2034912109375,
0.72802734375,
0.298095703125,
0.72802734375,
0.298095703125,
0.7421875,
0.2034912109375,
0.7421875
],
[
0.50634765625,
0.7421875,
0.60302734375,
0.744140625,
0.60302734375,
0.76220703125,
0.50634765625,
0.76025390625
],
[
0.7236328125,
0.7421875,
0.79736328125,
0.744140625,
0.7958984375,
0.76220703125,
0.72216796875,
0.76025390625
],
[
0.181396484375,
0.74853515625,
0.19287109375,
0.75244140625,
0.1871337890625,
0.76123046875,
0.176513671875,
0.7568359375
],
[
0.200927734375,
0.74365234375,
0.361083984375,
0.744140625,
0.361083984375,
0.76220703125,
0.200927734375,
0.76123046875
],
[
0.200927734375,
0.76025390625,
0.39453125,
0.76123046875,
0.39453125,
0.77880859375,
0.200927734375,
0.77783203125
],
[
0.5087890625,
0.76123046875,
0.6357421875,
0.76123046875,
0.6357421875,
0.77880859375,
0.5087890625,
0.77880859375
],
[
0.7255859375,
0.7646484375,
0.7900390625,
0.7646484375,
0.7900390625,
0.77587890625,
0.7255859375,
0.77587890625
],
[
0.1732177734375,
0.77099609375,
0.1829833984375,
0.76513671875,
0.1895751953125,
0.771484375,
0.1805419921875,
0.7763671875
],
[
0.2034912109375,
0.77978515625,
0.3798828125,
0.77978515625,
0.3798828125,
0.79443359375,
0.2034912109375,
0.79443359375
],
[
0.5107421875,
0.77978515625,
0.6337890625,
0.77978515625,
0.6337890625,
0.79443359375,
0.5107421875,
0.79443359375
],
[
0.7255859375,
0.77880859375,
0.78515625,
0.77880859375,
0.78515625,
0.79443359375,
0.7255859375,
0.79443359375
],
[
0.176513671875,
0.7890625,
0.1829833984375,
0.7841796875,
0.188720703125,
0.78857421875,
0.181396484375,
0.79345703125
],
[
0.200927734375,
0.79541015625,
0.420654296875,
0.79541015625,
0.420654296875,
0.81298828125,
0.200927734375,
0.81298828125
],
[
0.50732421875,
0.79541015625,
0.6357421875,
0.79541015625,
0.6357421875,
0.81298828125,
0.50732421875,
0.81298828125
],
[
0.7255859375,
0.796875,
0.82177734375,
0.796875,
0.82177734375,
0.81103515625,
0.7255859375,
0.81103515625
],
[
0.50634765625,
0.81103515625,
0.69677734375,
0.8125,
0.69677734375,
0.830078125,
0.50634765625,
0.8291015625
],
[
0.181396484375,
0.7998046875,
0.188720703125,
0.8037109375,
0.1829833984375,
0.8095703125,
0.176513671875,
0.80615234375
],
[
0.2017822265625,
0.8125,
0.4755859375,
0.8125,
0.4755859375,
0.830078125,
0.2017822265625,
0.830078125
],
[
0.72216796875,
0.81103515625,
0.79150390625,
0.81298828125,
0.79150390625,
0.8291015625,
0.72216796875,
0.8271484375
],
[
0.178955078125,
0.8173828125,
0.1895751953125,
0.8173828125,
0.1895751953125,
0.8271484375,
0.178955078125,
0.8271484375
],
[
0.2017822265625,
0.83056640625,
0.47802734375,
0.83056640625,
0.47802734375,
0.84521484375,
0.2017822265625,
0.84521484375
],
[
0.50634765625,
0.82763671875,
0.6953125,
0.83056640625,
0.6953125,
0.849609375,
0.50634765625,
0.8466796875
],
[
0.7255859375,
0.83056640625,
0.779296875,
0.83056640625,
0.779296875,
0.8466796875,
0.7255859375,
0.8466796875
],
[
0.5087890625,
0.8466796875,
0.70751953125,
0.8466796875,
0.70751953125,
0.8642578125,
0.5087890625,
0.8642578125
],
[
0.1805419921875,
0.83203125,
0.191162109375,
0.8359375,
0.1846923828125,
0.84521484375,
0.1732177734375,
0.84130859375
],
[
0.2034912109375,
0.84765625,
0.494384765625,
0.84765625,
0.494384765625,
0.8623046875,
0.2034912109375,
0.8623046875
],
[
0.7255859375,
0.84765625,
0.79150390625,
0.84765625,
0.79150390625,
0.8623046875,
0.7255859375,
0.8623046875
],
[
0.200927734375,
0.86572265625,
0.48046875,
0.86572265625,
0.48046875,
0.8798828125,
0.200927734375,
0.8798828125
],
[
0.1732177734375,
0.849609375,
0.191162109375,
0.849609375,
0.191162109375,
0.86279296875,
0.1732177734375,
0.86279296875
],
[
0.1690673828125,
0.8671875,
0.19287109375,
0.8671875,
0.19287109375,
0.8798828125,
0.1690673828125,
0.8798828125
],
[
0.5107421875,
0.86572265625,
0.70361328125,
0.86572265625,
0.70361328125,
0.8798828125,
0.5107421875,
0.8798828125
],
[
0.336669921875,
0.8916015625,
0.66259765625,
0.8916015625,
0.66259765625,
0.90478515625,
0.336669921875,
0.90478515625
]
] |
[
"Stackt+1",
"Buffert+1",
"Open NTst+1",
"Action",
"Buffert",
"Open NTst",
"s",
"Stackt",
"B",
"sI(x",
"B",
"NT(X)",
"n",
"n+1",
"si",
"B",
"|B",
"s",
"n",
"SHIFT",
"a",
"n",
"SI(XI TiI.. I Te",
"B",
"B",
"n-1",
"REDUCE",
"n",
"S I(X Ti .. Te)",
"Figure 1: Parser transitions showing the stack, buffer, and open nonterminal count before and after each action type. S represents",
"the stack, which contains open nonterminals and completed subtrees; B represents the buffer of unprocessed terminal symbols; ",
"is a terminal symbol, X is a nonterminal symbol, and each r is a completed subtree. The top of the stack is to the right, and the",
"buffer is consumed from left to right. Elements on the stack and buffer are delimited by a vertical bar (D)",
"Input: The hungry cat meows.",
"Stack",
"Buffer",
"Action",
"Thehungry cat meows.",
"NT(S)",
"The | hungry | cat [ meows ]",
"0",
"(S",
"NT(NP)",
"",
"(S /(NP",
"The | hungry |cat meows ].",
"SHIFT",
"(S I (NP| The",
"2",
"hungry I cat I meows I.",
"SHIFT",
"3",
"(S |(NPThe | hungry",
"cat | meows ",
"4",
"SHIFT",
"(S I (NP| The | hungry | cat",
"meows |",
"REDUCE",
"(S I (NP The hungry cat)",
"5",
"meows .",
"NT(VP)",
"(S / (NP The hungry cat) / (VP",
"6",
"meows |.",
"SHIFT",
"(S I (NP The hungry cat) I (VP meows",
"7",
"8",
"REDUCE",
"(S i (NP The hungry cat) I (VP meows)",
"9",
"SHIFT",
"(S / (NP The hungry cat) / (VP meows) I,",
"REDUCE",
"(S (NP The hungry cat) (VP meows) .)",
"10",
"11",
"Figure 2: Top-down parsing example.",
"Stackt+1",
"TermSt+1",
"Open NTst+1",
"Termst",
"Stackt",
"Open NTst",
"Action",
"s",
"T",
"n",
"NT(X)",
"SI(X",
"T",
"n+1",
"T",
"TIc",
"n",
"s",
"GEN(α)",
"S",
"",
"n",
"SI(XITI I..I Te",
"n",
"REDUCE",
"SI (X TI ... Te)",
"T",
"n-1",
"Figure 3: Generator transitions. Symbols defined as in Fig. 1 with the addition of T representing the history of generated terminals",
"T",
"Stack",
" Terminals",
"Action",
"0",
"NT(S)",
"(S",
"NT(NP)",
"(S /(NP",
"I",
"2",
"GEN(The)",
"The",
"GEN(hungry)",
"A",
"(S |(NP|The",
"The | hungry",
"GEN(cat)",
"A",
"(S | (NP| The | hungry",
"(S ] (NP| The | hungry | cat",
"The | hungry | cat",
"REDUCE",
"S",
"(S | (NP The hungry cat)",
"The[hungry |cat",
"NT(VP)",
"6",
"(S I(NP The hungry cat) I(VP",
"The | hungry I cat",
"GEN(meows)",
"The | hungry | cat meows",
"",
"(S [ (NP The hungry cat) I (VP meows",
"REDUCE",
"8",
"(S |(NP The hungry cat) [(VP meows)",
"The | hungry | cat | meows",
"GEN()",
"The | hungry | cat | meows ].",
"9",
"(S /(NP The hungry cat) / (VP meows) /:",
"REDUCE",
"(S (NP The hungry cat) (VP meows) .)",
"10",
"11",
"The[hungry [cat [meows]",
"Figure 4: Jointgeneration of a parsetree and sentence."
] |
{
"arxiv_id": "1602.07776",
"language": "en",
"page_idx": 3
}
|
||
1602.07776-en-4
|
1602.07776-en-4.jpg
|
suming the availability of constant time push and pop operations, the runtime is linear in the number of the nodes in the parse tree that is generated by the parser/generator (intuitively, this is true since although an individual REDUCE operation may require applying a number of pops that is linear in the number of input symbols, the total number of pop operations across an entire parse/generation run will also be linear). Since there is no way to bound the number of output nodes in a parse tree as a function of the number of input words, stating the runtime complexity of the parsing algorithm as a function of the input size requires further assumptions. Assuming our fixed constraint on maximum depth, it is linear.
## 3.5 Comparison to Other Models
Our generation algorithm algorithm differs from previous stack-based parsing/generation algorithms in two ways. First, it constructs rooted tree structures top down (rather than bottom up), and second, the transition operators are capable of directly generating arbitrary tree structures rather than, e.g., assuming binarized trees, as is the case in much prior work that has used transition-based algorithms to produce phrase-structure trees (Sagae and Lavie, 2005;Zhang and Clark, 2011; Zhu et al., 2013).
### 4 Generative Model
RNNGs use the generator transition set just presented to define a joint distribution on syntax trees (y) and words ($x$). This distribution is defined as a sequence model over generator transitions that is parameterized using a continuous space embedding of the algorithm state at each time step (u$_t);$i.e.,
$$\begin{aligned}p(\boldsymbol{x},\boldsymbol{y})&=\prod_{t=1}^{|\boldsymbol{a}(\boldsymbol{x},\boldsymbol{y})|}p(a_t\mid\boldsymbol{a}_{<t})\\&=\prod_{t=1}^{|\boldsymbol{a}(\boldsymbol{x},\boldsymbol{y})|}\frac{\exp\mathbf{r}_{a_t}^\top\mathbf{u}_t+b_{a_t}}{\sum_{a^{\prime}\in\mathcal{A}_G(T_t,S_t,n_t)}\exp\mathbf{r}_{a^{\prime}}^\top\mathbf{u}_t+b_{a^{\prime}}},\end{aligned}$$
and where action-specific embeddings r$_a$ and bias
vector b are parameters in $\Theta.$
The representation of the algorithm state at time $t,\textbf{u}_{t}$, is computed by combining the representation of the generator's three data structures: the output buffer $(T_t)$, represented by an embedding $\mathbf{o}_t$, the stack $(S_t)$, represented by an embedding s$_t$,and the
history of actions $(a_<t)$ taken by the generator, rep-
resented by an embedding h$_t$,
$$\mathbf{u}_t=\tanh\left(\mathbf{W}[\mathbf{o}_t;\mathbf{s}_t;\mathbf{h}_t]+\mathbf{c}\right),$$
where W and c are parameters. Refer to Figure 5
for an illustration of the architecture.
The output buffer, stack, and history are sequences that grow unboundedly, and to obtain representations of them we use recurrent neural networks to“encode” their contents (Cho et al., 2014). Since the output buffer and history of actions are only appended to and only contain symbols from a finite alphabet, it is straightforward to apply a standard RNN encoding architecture. The stack $(S)$ is more complicated for two reasons. First, the elements of the stack are more complicated objects than symbols from a discrete alphabet: open nonterminals, terminals, and full trees, are all present on the stack. Second, it is manipulated using both push and pop operations. To efficiently obtain representations of $S$ under push and pop operations, we use stack LSTMs (Dyer et al., 2015). To represent complex parse trees, we define a new syntactic composition function that recursively defines representations of trees.
## 4.1 Syntactic Composition Function
When a REDUCE operation is executed, the parser pops a sequence of completed subtrees and/or tokens (together with their vector embeddings) from the stack and makes them children of the most recent open nonterminal on the stack,“completing”the constituent. To compute an embedding of this new subtree, we use a composition function based on bidirectional LSTMs, which is illustrated in Fig. 6.
<FigureHere>
Figure 6: Syntactic composition function based on bidirectional LSTMs that is executed during a REDUCE operation; the network on the right models the structure on the left.
|
[
[
0.1168212890625,
0.0921630859375,
0.491943359375,
0.0921630859375,
0.491943359375,
0.106689453125,
0.1168212890625,
0.106689453125
],
[
0.5048828125,
0.09027099609375,
0.8798828125,
0.091552734375,
0.8798828125,
0.10919189453125,
0.5048828125,
0.10797119140625
],
[
0.11517333984375,
0.10986328125,
0.491943359375,
0.10919189453125,
0.491943359375,
0.12371826171875,
0.11517333984375,
0.125
],
[
0.50634765625,
0.111083984375,
0.7255859375,
0.111083984375,
0.7255859375,
0.125,
0.50634765625,
0.125
],
[
0.1168212890625,
0.1275634765625,
0.486083984375,
0.1275634765625,
0.486083984375,
0.141357421875,
0.1168212890625,
0.141357421875
],
[
0.1168212890625,
0.14453125,
0.486083984375,
0.14453125,
0.486083984375,
0.1590576171875,
0.1168212890625,
0.1590576171875
],
[
0.58251953125,
0.1458740234375,
0.8095703125,
0.1458740234375,
0.8095703125,
0.160400390625,
0.58251953125,
0.160400390625
],
[
0.1168212890625,
0.1610107421875,
0.490234375,
0.1610107421875,
0.490234375,
0.175537109375,
0.1168212890625,
0.175537109375
],
[
0.11517333984375,
0.1787109375,
0.491943359375,
0.177978515625,
0.491943359375,
0.1925048828125,
0.11517333984375,
0.1932373046875
],
[
0.5087890625,
0.179931640625,
0.88232421875,
0.179931640625,
0.88232421875,
0.1944580078125,
0.5087890625,
0.1944580078125
],
[
0.11517333984375,
0.1962890625,
0.491943359375,
0.1962890625,
0.491943359375,
0.210205078125,
0.11517333984375,
0.210205078125
],
[
0.5087890625,
0.1976318359375,
0.77392578125,
0.1976318359375,
0.77392578125,
0.2108154296875,
0.5087890625,
0.2108154296875
],
[
0.11517333984375,
0.21337890625,
0.490234375,
0.21337890625,
0.490234375,
0.2279052734375,
0.11517333984375,
0.2279052734375
],
[
0.52685546875,
0.21533203125,
0.88134765625,
0.21533203125,
0.88134765625,
0.2291259765625,
0.52685546875,
0.2291259765625
],
[
0.11517333984375,
0.2298583984375,
0.488525390625,
0.2298583984375,
0.488525390625,
0.2442626953125,
0.11517333984375,
0.2442626953125
],
[
0.5087890625,
0.23291015625,
0.8798828125,
0.23291015625,
0.8798828125,
0.246826171875,
0.5087890625,
0.246826171875
],
[
0.11517333984375,
0.2474365234375,
0.490234375,
0.2474365234375,
0.490234375,
0.261474609375,
0.11517333984375,
0.261474609375
],
[
0.50634765625,
0.25,
0.88134765625,
0.25,
0.88134765625,
0.26318359375,
0.50634765625,
0.26318359375
],
[
0.11517333984375,
0.264404296875,
0.490234375,
0.264404296875,
0.490234375,
0.27783203125,
0.11517333984375,
0.27783203125
],
[
0.5087890625,
0.266357421875,
0.88134765625,
0.266357421875,
0.88134765625,
0.27978515625,
0.5087890625,
0.27978515625
],
[
0.11517333984375,
0.281005859375,
0.491943359375,
0.27978515625,
0.491943359375,
0.294189453125,
0.11517333984375,
0.29541015625
],
[
0.5087890625,
0.2841796875,
0.88134765625,
0.2841796875,
0.88134765625,
0.29736328125,
0.5087890625,
0.29736328125
],
[
0.11358642578125,
0.29736328125,
0.491943359375,
0.298583984375,
0.491943359375,
0.313232421875,
0.11358642578125,
0.3125
],
[
0.50634765625,
0.3017578125,
0.88232421875,
0.3017578125,
0.88232421875,
0.31494140625,
0.50634765625,
0.31494140625
],
[
0.1168212890625,
0.317626953125,
0.483642578125,
0.317626953125,
0.483642578125,
0.32763671875,
0.1168212890625,
0.32763671875
],
[
0.5087890625,
0.317626953125,
0.8798828125,
0.318115234375,
0.8798828125,
0.333251953125,
0.5087890625,
0.33203125
],
[
0.5107421875,
0.333984375,
0.88232421875,
0.335205078125,
0.88232421875,
0.349853515625,
0.5107421875,
0.348388671875
],
[
0.11517333984375,
0.34228515625,
0.38232421875,
0.34228515625,
0.38232421875,
0.356689453125,
0.11517333984375,
0.356689453125
],
[
0.51123046875,
0.352783203125,
0.88134765625,
0.352783203125,
0.88134765625,
0.3662109375,
0.51123046875,
0.3662109375
],
[
0.1168212890625,
0.3642578125,
0.490234375,
0.3642578125,
0.490234375,
0.37890625,
0.1168212890625,
0.37890625
],
[
0.5087890625,
0.37060546875,
0.88134765625,
0.37060546875,
0.88134765625,
0.3837890625,
0.5087890625,
0.3837890625
],
[
0.11517333984375,
0.3818359375,
0.490234375,
0.380615234375,
0.490234375,
0.395751953125,
0.11517333984375,
0.396484375
],
[
0.5087890625,
0.386962890625,
0.88134765625,
0.386962890625,
0.88134765625,
0.40087890625,
0.5087890625,
0.40087890625
],
[
0.11517333984375,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.41357421875,
0.11517333984375,
0.41357421875
],
[
0.5087890625,
0.404052734375,
0.88232421875,
0.404052734375,
0.88232421875,
0.417236328125,
0.5087890625,
0.417236328125
],
[
0.11517333984375,
0.416748046875,
0.490234375,
0.416748046875,
0.490234375,
0.429931640625,
0.11517333984375,
0.429931640625
],
[
0.50732421875,
0.419189453125,
0.8837890625,
0.42041015625,
0.8837890625,
0.43505859375,
0.50732421875,
0.434326171875
],
[
0.11358642578125,
0.431884765625,
0.488525390625,
0.43310546875,
0.488525390625,
0.447509765625,
0.11358642578125,
0.4462890625
],
[
0.50732421875,
0.438232421875,
0.88232421875,
0.4375,
0.88232421875,
0.451904296875,
0.50732421875,
0.45263671875
],
[
0.1168212890625,
0.45068359375,
0.488525390625,
0.45068359375,
0.488525390625,
0.46533203125,
0.1168212890625,
0.46533203125
],
[
0.50634765625,
0.453857421875,
0.88232421875,
0.455078125,
0.88232421875,
0.4697265625,
0.50634765625,
0.468505859375
],
[
0.11517333984375,
0.46728515625,
0.490234375,
0.466552734375,
0.490234375,
0.48095703125,
0.11517333984375,
0.481689453125
],
[
0.50732421875,
0.47021484375,
0.88232421875,
0.4716796875,
0.88232421875,
0.486083984375,
0.50732421875,
0.48486328125
],
[
0.1168212890625,
0.48486328125,
0.488525390625,
0.48486328125,
0.488525390625,
0.498779296875,
0.1168212890625,
0.498779296875
],
[
0.5107421875,
0.4892578125,
0.88134765625,
0.4892578125,
0.88134765625,
0.50390625,
0.5107421875,
0.50390625
],
[
0.11517333984375,
0.501953125,
0.486083984375,
0.501953125,
0.486083984375,
0.51513671875,
0.11517333984375,
0.51513671875
],
[
0.5107421875,
0.5068359375,
0.8837890625,
0.5068359375,
0.8837890625,
0.521484375,
0.5107421875,
0.521484375
],
[
0.11358642578125,
0.51806640625,
0.455810546875,
0.51806640625,
0.455810546875,
0.5322265625,
0.11358642578125,
0.5322265625
],
[
0.5107421875,
0.52587890625,
0.5498046875,
0.52587890625,
0.5498046875,
0.537109375,
0.5107421875,
0.537109375
],
[
0.1168212890625,
0.548828125,
0.148681640625,
0.548828125,
0.148681640625,
0.56005859375,
0.1168212890625,
0.56005859375
],
[
0.1422119140625,
0.546875,
0.298095703125,
0.546875,
0.298095703125,
0.56103515625,
0.1422119140625,
0.56103515625
],
[
0.5107421875,
0.5546875,
0.79736328125,
0.5546875,
0.79736328125,
0.5693359375,
0.5107421875,
0.5693359375
],
[
0.11517333984375,
0.57275390625,
0.490234375,
0.57373046875,
0.490234375,
0.58837890625,
0.11517333984375,
0.5869140625
],
[
0.5087890625,
0.5791015625,
0.88232421875,
0.58056640625,
0.88232421875,
0.59521484375,
0.5087890625,
0.59326171875
],
[
0.11517333984375,
0.59130859375,
0.490234375,
0.59130859375,
0.490234375,
0.60498046875,
0.11517333984375,
0.60498046875
],
[
0.5087890625,
0.59765625,
0.8798828125,
0.59765625,
0.8798828125,
0.6123046875,
0.5087890625,
0.6123046875
],
[
0.11358642578125,
0.60791015625,
0.491943359375,
0.60693359375,
0.491943359375,
0.6220703125,
0.11358642578125,
0.62255859375
],
[
0.50634765625,
0.61279296875,
0.88232421875,
0.6142578125,
0.88232421875,
0.62890625,
0.50634765625,
0.62744140625
],
[
0.11517333984375,
0.625,
0.486083984375,
0.62548828125,
0.486083984375,
0.64013671875,
0.11517333984375,
0.6396484375
],
[
0.5087890625,
0.6318359375,
0.88134765625,
0.6318359375,
0.88134765625,
0.64599609375,
0.5087890625,
0.64599609375
],
[
0.1168212890625,
0.642578125,
0.491943359375,
0.642578125,
0.491943359375,
0.6572265625,
0.1168212890625,
0.6572265625
],
[
0.5107421875,
0.6494140625,
0.88134765625,
0.6494140625,
0.88134765625,
0.66357421875,
0.5107421875,
0.66357421875
],
[
0.11517333984375,
0.6591796875,
0.443603515625,
0.6591796875,
0.443603515625,
0.67236328125,
0.11517333984375,
0.67236328125
],
[
0.5107421875,
0.66650390625,
0.8798828125,
0.66650390625,
0.8798828125,
0.6796875,
0.5107421875,
0.6796875
],
[
0.193603515625,
0.68603515625,
0.2467041015625,
0.68603515625,
0.2467041015625,
0.70068359375,
0.193603515625,
0.70068359375
],
[
0.5087890625,
0.68310546875,
0.88134765625,
0.68310546875,
0.88134765625,
0.6962890625,
0.5087890625,
0.6962890625
],
[
0.11273193359375,
0.70263671875,
0.2091064453125,
0.70263671875,
0.2091064453125,
0.72021484375,
0.11273193359375,
0.72021484375
],
[
0.1993408203125,
0.70263671875,
0.328369140625,
0.70263671875,
0.328369140625,
0.72021484375,
0.1993408203125,
0.72021484375
],
[
0.5087890625,
0.69970703125,
0.87353515625,
0.69970703125,
0.87353515625,
0.71337890625,
0.5087890625,
0.71337890625
],
[
0.20751953125,
0.7216796875,
0.2344970703125,
0.7216796875,
0.2344970703125,
0.734375,
0.20751953125,
0.734375
],
[
0.193603515625,
0.73486328125,
0.2467041015625,
0.73486328125,
0.2467041015625,
0.75048828125,
0.193603515625,
0.75048828125
],
[
0.30810546875,
0.74365234375,
0.419189453125,
0.74365234375,
0.419189453125,
0.75830078125,
0.30810546875,
0.75830078125
],
[
0.2132568359375,
0.755859375,
0.2279052734375,
0.75830078125,
0.2254638671875,
0.76513671875,
0.2099609375,
0.76220703125
],
[
0.240234375,
0.7646484375,
0.48779296875,
0.76123046875,
0.48779296875,
0.7822265625,
0.240234375,
0.78515625
],
[
0.2091064453125,
0.7734375,
0.2344970703125,
0.7734375,
0.2344970703125,
0.78076171875,
0.2091064453125,
0.78076171875
],
[
0.1168212890625,
0.79443359375,
0.488525390625,
0.79443359375,
0.488525390625,
0.8095703125,
0.1168212890625,
0.8095703125
],
[
0.1168212890625,
0.8125,
0.326904296875,
0.8125,
0.326904296875,
0.8271484375,
0.1168212890625,
0.8271484375
],
[
0.134033203125,
0.830078125,
0.488525390625,
0.830078125,
0.488525390625,
0.84326171875,
0.134033203125,
0.84326171875
],
[
0.11517333984375,
0.8466796875,
0.490234375,
0.8466796875,
0.490234375,
0.861328125,
0.11517333984375,
0.861328125
],
[
0.8408203125,
0.8486328125,
0.8486328125,
0.8486328125,
0.8486328125,
0.85302734375,
0.8408203125,
0.85302734375
],
[
0.11517333984375,
0.86279296875,
0.494384765625,
0.86279296875,
0.494384765625,
0.87744140625,
0.11517333984375,
0.87744140625
],
[
0.50732421875,
0.861328125,
0.88232421875,
0.85986328125,
0.88232421875,
0.875,
0.50732421875,
0.87548828125
],
[
0.11358642578125,
0.87890625,
0.494384765625,
0.8798828125,
0.494384765625,
0.89453125,
0.11358642578125,
0.89306640625
],
[
0.50634765625,
0.87890625,
0.88232421875,
0.87890625,
0.88232421875,
0.892578125,
0.50634765625,
0.892578125
],
[
0.11517333984375,
0.89794921875,
0.491943359375,
0.89794921875,
0.491943359375,
0.912109375,
0.11517333984375,
0.912109375
],
[
0.50634765625,
0.89599609375,
0.81884765625,
0.89599609375,
0.81884765625,
0.9091796875,
0.50634765625,
0.9091796875
]
] |
[
"suming the availability of constant time push and",
"history of actions (a<t) taken by the generator, rep-",
"pop operations, the runtime is linear in the number",
"resented by an embedding ht,",
"of the nodes in the parse tree that is generated by",
"the parser/generator (intuitively, this is true since al-",
"ut = tanh (W[ot; st; ht] + c),",
"though an individual REDUCE operation may require",
"applying a number of pops that is linear in the num-",
"where W and c are parameters. Refer to Figure 5",
"ber of input symbols, the total number of pop opera-",
"for an illustration of the architecture.",
"tions across an entire parse/generation run will also",
"The output buffer, stack, and history are se-",
"be linear). Since there is no way to bound the num-",
"quences that grow unboundedly, and to obtain rep-",
"ber of output nodes in a parse tree as a function of",
"resentations of them we use recurrent neural net-",
"the number of input words, stating the runtime com-",
"works to \"encode\" their contents (Cho et al., 2014).",
"plexity of the parsing algorithm as a function of the",
"Since the output buffer and history of actions are",
"input size requires further assumptions. Assuming",
" only appended to and only contain symbols from a",
"ourfixed constrainton maximum depth.itis linear.",
"finite alphabet,it is straightforward to apply a stan-",
"dard RNN encoding architecture. The stack (S) is",
"3.5Comparison to Other Models",
"more complicated for two reasons. First, the ele-",
"Our generation algorithm algorithm differs from",
"ments of the stack are more complicated objects than",
"previous stack-based parsing/generation algorithms",
"symbols from a discrete alphabet: open nontermi-",
"in two ways. First, it constructs rooted tree struc-",
"nals, terminals, and full trees, are all present on the",
"tures top down (rather than bottom up), and sec-",
"stack. Second, it is manipulated using both push and",
"ond, the transition operators are capable of directly",
"pop operations. To efficiently obtain representations",
"generating arbitrary tree structures rather than, e.g.,",
"of S under push and pop operations, we use stack",
"assuming binarized trees, as is the case in much",
"LSTMs (Dyer et al., 2015). To represent complex",
"prior work that has used transition-based algorithms",
"parse trees, we define a new syntactic composition",
"to produce phrase-structure trees (Sagae and Lavie.",
"function that recursively defines representations of",
"2005; Zhang and Clark, 2011; Zhu et al., 2013)",
"trees.",
"4",
"Generative Model",
"4.1 Syntactic Composition Function",
"RNNGs use the generator transition set just pre-",
"When a REDUCE operation is executed, the parser",
"sented to define a joint distribution on syntax trees",
"pops a sequence of completed subtrees and/or to-",
"(y) and words (α). This distribution is defined as a",
"kens (together with their vector embeddings) from",
"sequence model over generator transitions that is pa-",
"the stack and makes them children of the most recent",
"rameterized using a continuous space embedding of",
"open nonterminal on the stack, \"completing\"the",
"the algorithm state at each time step (ut); i.e.,",
"constituent. To compute an embedding of this new",
"[a(a,y)]",
"subtree, we use a composition function based on",
"p(c,y) = ",
"II p(at | a<t)",
"bidirectional LSTMs, which is illustrated in Fig. 6.",
"t=1",
"[a(ac,y)l",
"expr. ut + bat",
"",
"a'Ac(T,St,nt) exprdut + ba.",
"t-1.",
"and where action-specific embeddings ra and bias",
"vector b are parameters in .",
"The representation of the algorithm state at time",
"t, ut, is computed by combining the representation",
"",
"of the generator's three data structures: the output",
"Figure 6: Syntactic composition function based on bidirec-",
"buffer (T),represented by an embedding ot,the",
"tional LSTMs that is executed during a REDUCE operation; the",
"stack (St), represented by an embedding St, and the",
"network on the right models the structure on the left."
] |
{
"arxiv_id": "1602.07776",
"language": "en",
"page_idx": 4
}
|
||
1708.09268-en-0
|
1708.09268-en-0.jpg
|
## Two-stream Flow-guided Convolutional Attention Networks for Action $\textbf{Recognition}$
An Tran
Loong-Fah Cheong
Department of Electrical & Computer Engineering, National University of Singapore
an.tran@u.nus.edu eleclf@nus.edu.sg
## $\mathbf{Abstract}$
This paper proposes a two-stream flow-guided convolutional attention networks for action recognition in videos. The central idea is that optical flows, when properly compensatedforthe camera motion, can be used to guide attention to the human foreground. We thus develop crosslink layers from the temporal network (trained on flows) to the spatial network (trained on RGB frames). These crosslink layers guide the spatial-stream to pay more attention to the human foreground areas and be less affected by background clutter. We obtain promising performances with our approach on the UCF101, HMDB51 and Hollywood2 datasets.
<FigureHere>
Figure 1: Our proposed Two-stream Flow-guided Convolutional Attention Networks (Two-stream FCANs) for action recognition. A video clip is represented by RGB frames and optical flows. Two streams of data are fed into two separate CNNs: spatial stream that models scene and object contexts (blue), while (compensated) temporal-stream likely provides some motion-based attentions on foreground actions (orange). We leverage attentions provided from temporal-stream to assist recognition processes in spatialstream by cross-link layers (pink). The attention weighted feature maps are fused by element-wise multiplication to preserve spatial-temporal structure in videos. The Twostream FCAN refers to the entire architecture of two streams with the late fusion stage, whereas we call the area inside dashed lines the FCAN model. Best viewed in color.
## $1. \textbf{Introduction}$
Human action recognition in video is an important and challenging problem in computer vision. Like many other
$lems.$
Over the last decade, there is a great evolution of features for action recognition from short video clips [29, 10,22,26]. The research works can be roughly divided into two mainstreams. The first type of representation is hand-crafted local features in combination with the Bag-ofFeatures (BoFs) paradigm [14, 11, 29]. Probably the most successful approach of local features representation is to extract improved dense trajectory features [29] and deploy Fisher vector representation [20]. The second approach is to utilize deep learning algorithms to learn features automatically from data $(e.g.$, RGB frames or optical flows) [22,10,26,28,30]. Probably the most successful approach of local features representation is to extract improved dense trajectory features [29] and deploy Fisher vector representation [20]. High performances of neural network architectures have been recently reported on video action recognition, specially those of two-stream convolutional networks
### [22,30].
Due to different network architectures and types of data in these two-stream networks, the learned features should have characteristics that help to deal with the different types of nuances in these specific data streams. RGB frames in video usually provide scene and object contexts in the background together with the human forms in the foreground. However, the spatial area occupied by the human foreground is usually much smaller than the area of the background such that it might not be effectively represented. On the contrary, optical flows in videos, when properly compensated for camera motion, immediately isolate the moving human silhouettes (see Figure 2), and provide motion
1
|
[
[
0.134033203125,
0.1337890625,
0.83837890625,
0.1337890625,
0.83837890625,
0.1483154296875,
0.134033203125,
0.1483154296875
],
[
0.423095703125,
0.154052734375,
0.54736328125,
0.1558837890625,
0.54736328125,
0.173583984375,
0.423095703125,
0.1717529296875
],
[
0.35546875,
0.2052001953125,
0.4248046875,
0.2052001953125,
0.4248046875,
0.2197265625,
0.35546875,
0.2197265625
],
[
0.461669921875,
0.2052001953125,
0.61767578125,
0.2052001953125,
0.61767578125,
0.2197265625,
0.461669921875,
0.2197265625
],
[
0.1519775390625,
0.2215576171875,
0.82177734375,
0.2215576171875,
0.82177734375,
0.2392578125,
0.1519775390625,
0.2392578125
],
[
0.32275390625,
0.242431640625,
0.48193359375,
0.242431640625,
0.48193359375,
0.255615234375,
0.32275390625,
0.255615234375
],
[
0.48046875,
0.2412109375,
0.64892578125,
0.242431640625,
0.64892578125,
0.2568359375,
0.48046875,
0.255615234375
],
[
0.0187835693359375,
0.255615234375,
0.062103271484375,
0.255126953125,
0.06781005859375,
0.373779296875,
0.0236968994140625,
0.375
],
[
0.23779296875,
0.292236328125,
0.313720703125,
0.292236328125,
0.313720703125,
0.306884765625,
0.23779296875,
0.306884765625
],
[
0.5361328125,
0.302978515625,
0.5537109375,
0.302978515625,
0.5537109375,
0.306884765625,
0.5361328125,
0.306884765625
],
[
0.098876953125,
0.32568359375,
0.46728515625,
0.32568359375,
0.46728515625,
0.34033203125,
0.098876953125,
0.34033203125
],
[
0.08087158203125,
0.341552734375,
0.468017578125,
0.341552734375,
0.468017578125,
0.354736328125,
0.08087158203125,
0.354736328125
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.35595703125,
0.468017578125,
0.37060546875,
0.080078125,
0.369384765625
],
[
0.0269622802734375,
0.363037109375,
0.0555419921875,
0.363037109375,
0.0555419921875,
0.395751953125,
0.0269622802734375,
0.395751953125
],
[
0.080078125,
0.371826171875,
0.468017578125,
0.371826171875,
0.468017578125,
0.385009765625,
0.080078125,
0.385009765625
],
[
0.51708984375,
0.37890625,
0.56640625,
0.37890625,
0.56640625,
0.386474609375,
0.51708984375,
0.386474609375
],
[
0.08087158203125,
0.386474609375,
0.46728515625,
0.386474609375,
0.46728515625,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.08087158203125,
0.401611328125,
0.468017578125,
0.401611328125,
0.468017578125,
0.415283203125,
0.08087158203125,
0.415283203125
],
[
0.080078125,
0.416748046875,
0.464111328125,
0.416748046875,
0.464111328125,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.080078125,
0.429931640625,
0.468017578125,
0.43115234375,
0.468017578125,
0.44580078125,
0.080078125,
0.4443359375
],
[
0.0318603515625,
0.44580078125,
0.058013916015625,
0.4462890625,
0.0555419921875,
0.481689453125,
0.0294189453125,
0.48095703125
],
[
0.080078125,
0.4462890625,
0.468017578125,
0.4462890625,
0.468017578125,
0.4609375,
0.080078125,
0.4609375
],
[
0.50341796875,
0.4462890625,
0.890625,
0.4462890625,
0.890625,
0.460205078125,
0.50341796875,
0.460205078125
],
[
0.080078125,
0.462158203125,
0.4697265625,
0.462158203125,
0.4697265625,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.5048828125,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.473388671875,
0.5048828125,
0.473388671875
],
[
0.07843017578125,
0.4765625,
0.468017578125,
0.47607421875,
0.468017578125,
0.490478515625,
0.07843017578125,
0.4912109375
],
[
0.50244140625,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.080078125,
0.49365234375,
0.1380615234375,
0.49365234375,
0.1380615234375,
0.5048828125,
0.080078125,
0.5048828125
],
[
0.50341796875,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.50341796875,
0.505859375
],
[
0.50244140625,
0.50830078125,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.50244140625,
0.521484375
],
[
0.0228729248046875,
0.49560546875,
0.05963134765625,
0.49560546875,
0.05963134765625,
0.64013671875,
0.0228729248046875,
0.64013671875
],
[
0.50244140625,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.08087158203125,
0.54150390625,
0.2091064453125,
0.54150390625,
0.2091064453125,
0.55615234375,
0.08087158203125,
0.55615234375
],
[
0.50244140625,
0.537109375,
0.89208984375,
0.5361328125,
0.89208984375,
0.55029296875,
0.50244140625,
0.5517578125
],
[
0.50341796875,
0.55322265625,
0.89208984375,
0.55322265625,
0.89208984375,
0.56640625,
0.50341796875,
0.56640625
],
[
0.098876953125,
0.56640625,
0.468017578125,
0.5673828125,
0.468017578125,
0.58203125,
0.098876953125,
0.58056640625
],
[
0.50244140625,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50244140625,
0.58203125
],
[
0.08087158203125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.59765625,
0.08087158203125,
0.59765625
],
[
0.50341796875,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.080078125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.50244140625,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.080078125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.080078125,
0.626953125
],
[
0.5009765625,
0.61279296875,
0.890625,
0.6123046875,
0.890625,
0.626953125,
0.5009765625,
0.62744140625
],
[
0.080078125,
0.6298828125,
0.1168212890625,
0.6298828125,
0.1168212890625,
0.6416015625,
0.080078125,
0.6416015625
],
[
0.50244140625,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.0310516357421875,
0.63623046875,
0.056365966796875,
0.63623046875,
0.056365966796875,
0.66650390625,
0.0310516357421875,
0.66650390625
],
[
0.098876953125,
0.64453125,
0.46728515625,
0.64453125,
0.46728515625,
0.65771484375,
0.098876953125,
0.65771484375
],
[
0.5009765625,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.080078125,
0.673828125
],
[
0.5009765625,
0.6572265625,
0.84716796875,
0.6572265625,
0.84716796875,
0.67041015625,
0.5009765625,
0.67041015625
],
[
0.0318603515625,
0.6787109375,
0.052276611328125,
0.67822265625,
0.0555419921875,
0.70556640625,
0.035125732421875,
0.70703125
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.689453125,
0.08087158203125,
0.689453125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70263671875,
0.560546875,
0.70263671875,
0.560546875,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.08087158203125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.08087158203125,
0.734375
],
[
0.521484375,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.521484375,
0.734375
],
[
0.080078125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.5009765625,
0.75,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.5009765625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
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
]
] |
[
"Two-stream Flow-guided Convolutional Attention Networks for Action",
"Recognition",
"An Tran",
"Loong-Fah Cheong",
"Department of Electrical & Computer Engineering, National University of Singapore",
"an.tran@u.nus.edu",
"eleclf@nus.edu.sg",
"LIOz SnV",
"Abstract",
"ECN",
"This paper proposes a two-stream flow-guided convolu-",
"tional attention networks for action recognition in videos.",
"The central idea is that optical flows, when properly com-",
"3",
"pensated for the camera motion, can be used to guide at-",
"temporal-CNN",
"tention to the human foreground. We thus develop cross-",
"LAD",
"link layers from the temporal network (trained on flows) to",
"the spatial network (trained on RGB frames). These cross",
"link layers guide the spatial-stream to paymore attention to",
"SO",
"the human foreground areas and be less afected by back-",
"Figure 1: Our proposed Two-stream Flow-guided Convolu-",
"ground clutter.We obtain promising performances with",
"tionalAttentionNetworks(Two-streamFCANs)foraction",
"our approach on the UCF101, HMDB51 and Hollywood2",
"recognition. A video clip is represented by RGB frames",
"datasets.",
"and optical flows. Two streams of data are fed into two",
"separate CNNs: spatial stream that models scene and ob-",
"189260802",
"ject contexts (blue), while (compensated) temporal-stream",
"1. Introduction",
"likely provides some motion-based attentions on foreground",
"actions (orange).We leverage attentions provided from",
"Human action recognition in video is an important and",
"temporal-stream to assist recognition processes in spatial-",
"challenging problem in computer vision. Like many other",
"stream by cross-link layers (pink). The attention weighted",
"computer vision problems, an effective visual representa-",
"feature maps are fused by element-wise multiplication to",
"tion of actions in video data is vital to deal with these prob-",
"preserve spatial-temporal structure in videos.The Two-",
"lems.",
"stream FCAN refers to the entire architecture of two streams",
"[:A",
"Over the last decade, there is a great evolution of fea-",
"with the late fusion stage, whereas we call the area inside",
"tures for action recognition from short video clips [29,",
"dashed lines the FCAN model. Best viewed in color.",
"e",
"10, 22, 26]. The research works can be roughly divided",
"into two mainstreams. The first type of representation is",
"hand-crafted local features in combination with the Bag-of-",
"[22, 30].",
"Features (BoFs) paradigm [14, 11, 29]. Probably the most",
"Due to different network architectures and types of data",
"successful approach of local features representation is to",
"inthesetwo-streamnetworks.thelearnedfeatures should",
"extract improved dense trajectory features [29] and deploy",
"have characteristics that help to deal with the different types",
"Fisher vector representation [20]. The second approach is",
"of nuances in these specific data streams. RGB frames in",
"to utilize deep learning algorithms to learn features auto-",
"videousuallyprovidesceneand object contextsintheback-",
"matically from data (e.g., RGB frames or optical flows)",
"ground together with the human forms in the foreground.",
"[22, 10, 26, 28, 30]. Probably the most successful approach",
"However, the spatial area occupied by the human fore-",
"of local features representation is to extract improved dense",
"ground is usually much smaller than the area of the back-",
"trajectory features [29] and deploy Fisher vector represen-",
"ground such that it might not be effectively represented. On",
"tation [20]. High performances of neural network architec-",
"the contrary, optical flows in videos, when properly com-",
"tures have been recently reported on video action recogni-",
"pensated for camera motion, immediately isolate the mov-",
"tion, specially those of two-stream convolutional networks",
"ing human silhouettes (see Figure 2), and provide motion"
] |
{
"arxiv_id": "1708.09268",
"language": "en",
"page_idx": 0
}
|
||
1708.09268-en-1
|
1708.09268-en-1.jpg
|
<FigureHere>
Figure 2: Columns represent examples of three modalities of inputs: RGB frames, optical flows (x,y-directions), and compensated optical flows. First two rows consist of two frames of a cartwheel video in HMDB51 dataset, and last two ones consist of two frames of a handstand walking video in UCF101 dataset.
cues. In view of the preceding discussion, feature responses of a CNN model $(i.e.$, called spatial-CNN) on RGB data are likely to be activations more on background contexts rather than on human foreground actions. In contrast, a CNN model $(i.e.$, called $temporal-CNN)$ trained on flow data would often fire more on the human forms and movements.
In this paper, we propose a novel flow-guided convolutional attention networks (FCANs) for action recognition based on the aforementioned two-stream network architecture (see Figure 1). This attention guiding is partly motivated by the primate visual system, whereby it is known that there is connectivity between the motion pathway and the form pathway [27]. To model the attention guidance, we propose cross-link layers from the temporal stream to the spatial stream. There can be multiple such cross-link layers, but as we shall show later, the optimal number is in the range of one to two layers. Each cross-link layer has three components: (1) a convolution layer to reduce the dimension of the flow feature tensor; (2) a mean-variance normalization layer; (3) a sigmoid function. The final cross-link output is an attention map, which is used to control the level of activation in the corresponding layer in the spatial-stream via an element-wise multiplication.
Our contributions are two-fold. First, we propose a flowguided convolution based attention mechanism for action recognition task. Second, we perform comprehensive evaluations two-stream FCANs based on 3d-convolution operations; we also explore the effects of different number of cross-link layers to understand where they are most effective. We visualize attentions provided by cross-link layers in our FCAN model (3D version) to show the attentive capacity of the (compensated) flows. Lastly, we achieve promising results on the HMDB51 and the UCF101 datasets. All codes and models $^{1}$ are implemented in Caffe framework [9].
## $2. \textbf{Related Work}$
Visual features. Many hand-crafted features have been proposed in the history of action recognition community, such as HOG/HOF [14], HOG3D [11], MBH [29], etc.
Inspired by recent successes in image classification [12], there have been extensions of the neural networks to the video action recognition problem [22,10,26]. CNN architectures play significant roles in these works, either as an individual module or as an encoder module for a type of recurrent neural networks (RNNs). Karpathy et $al.$ [10] propose a large-scale video dataset, namely Sport1M, and investigate different ways to embed temporal information into the current CNN architecture. Two-stream CNN model [22] has demonstrated good performance on
$^{\mathrm{l}}$https://github.com/antran89/two-stream-fcan
the UCF101 dataset [24] by combining predictions from two CNNs: spatial-CNNs trained from RGB frames and $temporal-CNNs$ trained from optical flows. Recently, Tran et $\dot{al}.[26]$ extend 2D CNNs to 3D CNNs by developing 3d-convolution and 3d-pooling layers. In [31,4], the output of a CNN's last layer is fed into a recurrent sequence model usually formed by LSTM cells. Interesting works [28,30] have attempted to model longer temporal information of videos.
Attention for action recognition. Attention is a mechanism used to confer more weights on a subset of features. The attention mechanism has also been applied to action recognition [16,21,1,2]. Bazzani et al.use additional human fxation data to train mixture density network for saliency prediction and apply it to action recognition with the so called C3D [26] features obtained from the 3D CNNs. On the contrary, our approach does not use any additional data except flows to predict attentions. Furthermore, Sharma et al. [21] extract image features in each frame with the VGG16 CNN model [23] and predict visual attention in each frame using a recurrent model with LSTM cell. The work most similar to ours is VideoLSTM [16]. VideoLSTM[16] uses convolutional LSTM trained on optical flow to predict attention for a second convolutional LSTM layer. Our FCAN is a convolution-based network that embeds attention in the process of action classification.
Before the deep learning era, there have been works incorporating saliency into action recognition from videos. Several saliency measures have been proposed for actions in [25,17] and they show improvements in the recognition accuracy when focusing attention on the foreground.
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.07843017578125,
0.1060791015625,
0.4697265625,
0.106689453125,
0.4697265625,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.080078125,
0.184326171875,
0.125,
0.184326171875,
0.125,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.098876953125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.098876953125,
0.212158203125
],
[
0.07843017578125,
0.212158203125,
0.4697265625,
0.21337890625,
0.4697265625,
0.2279052734375,
0.07843017578125,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.46728515625,
0.2291259765625,
0.46728515625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.08087158203125,
0.27587890625,
0.46728515625,
0.27587890625,
0.46728515625,
0.285888671875,
0.08087158203125,
0.285888671875
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.296630859375,
0.890625,
0.296630859375,
0.890625,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50244140625,
0.3125,
0.890625,
0.3125,
0.890625,
0.32568359375,
0.50244140625,
0.32568359375
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50341796875,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.080078125,
0.335205078125,
0.4697265625,
0.335205078125,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.50244140625,
0.35595703125
],
[
0.08087158203125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.50244140625,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.50244140625,
0.371826171875
],
[
0.08087158203125,
0.3662109375,
0.465576171875,
0.3662109375,
0.465576171875,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.50244140625,
0.37255859375,
0.67236328125,
0.37255859375,
0.67236328125,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.080078125,
0.425537109375,
0.4697265625,
0.425537109375,
0.4697265625,
0.439453125,
0.080078125,
0.439453125
],
[
0.50341796875,
0.423583984375,
0.89208984375,
0.423583984375,
0.89208984375,
0.4375,
0.50341796875,
0.4375
],
[
0.080078125,
0.43994140625,
0.31201171875,
0.43994140625,
0.31201171875,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.09722900390625,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.09722900390625,
0.4697265625
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.080078125,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.50244140625,
0.4697265625,
0.89208984375,
0.4697265625,
0.89208984375,
0.48291015625,
0.50244140625,
0.48291015625
],
[
0.080078125,
0.486083984375,
0.468017578125,
0.486083984375,
0.468017578125,
0.5,
0.080078125,
0.5
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50341796875,
0.5,
0.89208984375,
0.5,
0.89208984375,
0.51318359375,
0.50341796875,
0.51318359375
],
[
0.07843017578125,
0.5166015625,
0.4697265625,
0.51513671875,
0.4697265625,
0.52978515625,
0.07843017578125,
0.53076171875
],
[
0.50244140625,
0.5146484375,
0.89208984375,
0.5146484375,
0.89208984375,
0.52783203125,
0.50244140625,
0.52783203125
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.50244140625,
0.52880859375,
0.89208984375,
0.52978515625,
0.89208984375,
0.54443359375,
0.50244140625,
0.54345703125
],
[
0.07843017578125,
0.54541015625,
0.471435546875,
0.546875,
0.471435546875,
0.56103515625,
0.07843017578125,
0.56005859375
],
[
0.50341796875,
0.546875,
0.6005859375,
0.546875,
0.6005859375,
0.55810546875,
0.50341796875,
0.55810546875
],
[
0.080078125,
0.56103515625,
0.4697265625,
0.56103515625,
0.4697265625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.521484375,
0.564453125,
0.89208984375,
0.564453125,
0.89208984375,
0.57763671875,
0.521484375,
0.57763671875
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.080078125,
0.59130859375,
0.46728515625,
0.59130859375,
0.46728515625,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.5009765625,
0.59326171875,
0.89208984375,
0.5947265625,
0.89208984375,
0.609375,
0.5009765625,
0.60791015625
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50341796875,
0.60986328125,
0.890625,
0.60986328125,
0.890625,
0.62353515625,
0.50341796875,
0.62353515625
],
[
0.080078125,
0.62353515625,
0.181396484375,
0.62353515625,
0.181396484375,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.5048828125,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.63720703125,
0.5048828125,
0.63720703125
],
[
0.50244140625,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50244140625,
0.65380859375
],
[
0.080078125,
0.65283203125,
0.2197265625,
0.65283203125,
0.2197265625,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.50341796875,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.66943359375,
0.50341796875,
0.66943359375
],
[
0.50341796875,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.098876953125,
0.6767578125,
0.468017578125,
0.67822265625,
0.468017578125,
0.6923828125,
0.098876953125,
0.69140625
],
[
0.5048828125,
0.6875,
0.8896484375,
0.6875,
0.8896484375,
0.69775390625,
0.5048828125,
0.69775390625
],
[
0.080078125,
0.6943359375,
0.468017578125,
0.6943359375,
0.468017578125,
0.70849609375,
0.080078125,
0.70849609375
],
[
0.50341796875,
0.70068359375,
0.89208984375,
0.70068359375,
0.89208984375,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.080078125,
0.708984375,
0.439453125,
0.708984375,
0.439453125,
0.72265625,
0.080078125,
0.72265625
],
[
0.50341796875,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50341796875,
0.72998046875
],
[
0.09722900390625,
0.724609375,
0.46728515625,
0.724609375,
0.46728515625,
0.73779296875,
0.09722900390625,
0.73779296875
],
[
0.5009765625,
0.73095703125,
0.89208984375,
0.72900390625,
0.89208984375,
0.7421875,
0.5009765625,
0.744140625
],
[
0.080078125,
0.74072265625,
0.468017578125,
0.74072265625,
0.468017578125,
0.75390625,
0.080078125,
0.75390625
],
[
0.50341796875,
0.74560546875,
0.89208984375,
0.74560546875,
0.89208984375,
0.7587890625,
0.50341796875,
0.7587890625
],
[
0.07843017578125,
0.7548828125,
0.468017578125,
0.75390625,
0.468017578125,
0.76806640625,
0.07843017578125,
0.76953125
],
[
0.50341796875,
0.76123046875,
0.890625,
0.76123046875,
0.890625,
0.7744140625,
0.50341796875,
0.7744140625
],
[
0.080078125,
0.77099609375,
0.4697265625,
0.77099609375,
0.4697265625,
0.7841796875,
0.080078125,
0.7841796875
],
[
0.5009765625,
0.77587890625,
0.890625,
0.7744140625,
0.890625,
0.7890625,
0.5009765625,
0.79052734375
],
[
0.07843017578125,
0.7841796875,
0.468017578125,
0.78515625,
0.468017578125,
0.7998046875,
0.07843017578125,
0.798828125
],
[
0.50244140625,
0.79150390625,
0.890625,
0.79150390625,
0.890625,
0.8046875,
0.50244140625,
0.8046875
],
[
0.08087158203125,
0.80029296875,
0.468017578125,
0.80029296875,
0.468017578125,
0.814453125,
0.08087158203125,
0.814453125
],
[
0.50244140625,
0.806640625,
0.79736328125,
0.806640625,
0.79736328125,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.08087158203125,
0.81640625,
0.46728515625,
0.81640625,
0.46728515625,
0.830078125,
0.08087158203125,
0.830078125
],
[
0.080078125,
0.83056640625,
0.468017578125,
0.83056640625,
0.468017578125,
0.8447265625,
0.080078125,
0.8447265625
],
[
0.521484375,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.080078125,
0.84521484375,
0.468017578125,
0.84521484375,
0.468017578125,
0.859375,
0.080078125,
0.859375
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.859375,
0.4697265625,
0.85986328125,
0.4697265625,
0.875,
0.080078125,
0.87353515625
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88623046875,
0.44287109375,
0.88720703125,
0.44287109375,
0.90087890625,
0.09722900390625,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.8486328125,
0.88720703125,
0.8486328125,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"cues. In view of the preceding discussion, feature responses",
"of a CNN model (i.e., called spatial-CNN) on RGB data",
"arelikelyto be activations more on background contexts",
"rather than on human foreground actions. In contrast, a",
"CNN model (i.e., called temporal-CNN) trained on flow",
"data would often fire more on the human forms and move-",
"ments.",
"In this paper, we propose a novel flow-guided convolu-",
"tional attention networks (FCANs) for action recognition",
"based on the aforementioned two-stream network architec-",
"ture (see Figure 1). This attention guiding is partly moti-",
"vated by the primate visual system, whereby it is known",
"thatthereisconnectivitybetweenthemotionpathwayand",
"the form pathway [27]. To model the attention guidance, we",
"Figure 2: Columns represent examples of three modali-",
"propose cross-link layers from the temporal stream to the",
"ties of inputs: RGB frames, optical flows (x,y-directions),",
"spatial stream. There can be multiple such cross-link lay-",
"and compensated optical flows. First two rows consist of",
"ers, but as we shall show later, the optimal number is in the",
"two frames of a cartwheel video in HMDB51 dataset, and",
"range of oneto two layers.Each cross-link layerhas three",
"last two ones consist of two frames of a handstand walking",
"components:(1)a convolutionlayertoreducethe dimen-",
"video in UCF101 dataset.",
"sion of the flow feature tensor; (2) a mean-variance normal-",
"ization layer; (3) a sigmoid function. The final cross-link",
"output is an attention map, which is used to control the level",
"of activation in the corresponding layer in the spatial-stream",
"the UCF101 dataset [24] by combining predictions from",
"via an element-wise multiplication.",
"two CNNs: spatial-CNNs trained from RGB frames and",
"Our contributions are two-fold. First, we propose a flow-",
"temporal-CNNs trained from optical flows. Recently, Tran",
"guided convolution based attention mechanism for action",
"et al. [26] extend 2D CNNs to 3D CNNs by developing",
"recognition task. Second, we perform comprehensive eval-",
"3d-convolution and 3d-pooling layers. In [31, 4], the out-",
"uations two-stream FCANs based on 3d-convolution op-",
"put of a CNN's last layer is fed into a recurrent sequence",
"erations; we also explore the effects of different number",
"model usually formed by LSTM cells. Interesting works",
"of cross-link layers to understand where they are most ef-",
"[28, 30] have attempted to model longer temporal informa-",
"fective.We visualize attentions provided by cross-link",
"tion of videos.",
"layers in our FCAN model (3D version) to show the at-",
"Attention for action recognition. Attention is a mech-",
"tentive capacity of the (compensated)flows.Lastly,we",
"anism used to confer more weights on a subset of fea-",
"achieve promising results on the HMDB51 and the UCF101",
"tures. The attention mechanism has also been applied to",
"datasets. All codes and models I are implemented in Caffe",
"action recognition [16, 21, 1, 2]. Bazzani et al. use addi-",
"framework[9].",
"tionalhumanfixationdatatotrainmixturedensitynetwork",
"for saliency prediction and apply it to action recognition",
"2. Related Work",
"with the so called C3D [26] features obtained from the 3D",
"CNNs. On the contrary, our approach does not use any addi-",
"Visual features. Many hand-crafted features have been",
"tionaldata exceptflowstopredictattentions.Furthermore.",
"proposed in the history of action recognition community,",
"Sharma et al. [21] extract image features in each frame with",
"such as HOG/HOF [14], HOG3D [11], MBH [29], etc.",
"the VGG16 CNN model [23] and predict visual attention in",
"Inspired by recent successes in image classification [12]",
"each frame using a recurrent model with LSTM cell. The",
"there have been extensions of the neural networks to the",
"work most similar to ours is VideoLSTM [16]. VideoL-",
"video action recognition problem [22, 10, 26]. CNN ar-",
"STM [16] uses convolutional LSTM trained on optical flow",
"chitectures play significant roles in these works, either as",
"to predict attention for a second convolutional LSTM layer",
"an individual module or as an encoder module for a type",
"Our FCAN is a convolution-based network that embeds at-",
"of recurrent neural networks (RNNs). Karpathy et al.",
"tention in the process of action classification.",
"[10] propose a large-scale video dataset, namely Sport1M,",
"and investigate different ways to embed temporal infor-",
"Before the deep learning era, there have been works in-",
"mation into the current CNN architecture. Two-stream",
"corporating saliency into action recognition from videos.",
"CNN model [22] has demonstrated good performance on",
"Several saliency measures have been proposed for actions",
"in [25, 17] and they show improvements in the recognition",
"lhttps://github.com/antran89/two-stream-fcan",
"accuracy when focusing attention on the foreground."
] |
{
"arxiv_id": "1708.09268",
"language": "en",
"page_idx": 1
}
|
||
1708.09268-en-2
|
1708.09268-en-2.jpg
|
### 3.Flow-guided Convolutional Attention Net- $\mathbf{works}\left(\mathbf{FCANs}\right)$
We propose the cross-link layers to model the interactions of the two networks in the two-stream convolutional network[22]. The whole network is differential, so it can be trained end-to-end with the stochastic gradient descent (SGD) and back-propagation algorithm [15]. The overall architecture of the FCAN is shown in Figure 1. In the ensuing discussion, we describe the 3D-FCAN model, which is the 3D version based on 3D convolution (e.g., C3D [26]) building blocks (please refer to the Supplementary Materials for the 2D-FCAN model based on 2D convolution (e.g., $\operatorname{Alexnet}\left[12\right])).$
### 3.1.3D version of flow-guided convolutional attention networks
With the assumption that the magnitudes of optical flows, when appropriately compensated for camera motions, usually correlate with the foreground regions, we develop the framework of flow-guided convolutional attention networks (FCANs) as shown in Figure 1. The C3D network[26] provides explicit representation of the time dimension in the architecture. Both the 3D-FCAN and 2DFCAN have similar structures, except for the operations in the convolution and mean-variance normalization layer. Let $x\_ rgb^{l}$ $\in$ $\mathbb{R} ^{C_{l}\times T_{l}\times H_{l}\times W_{l}}$, $x\_ flow^{l}$ $\in$ $\mathbb{R} ^{C_{l}\times T_{l}\times H_{l}\times W_{l}}$ be the feature map of layer $l\in\{0,1,...,L\}$ in the spatialand temporal-C3D respectively, with $C_l,T_l,H_l,W_l$ being the number of channels, temporal length, height and width of the feature map. Specifically, in the proposed FCANs, $x\_rgb$ and $x\_flow$ are the feature maps from a pooling layer in the C3D network. We develop attentive cross-link layers between the early pooling layers from the $temporal-C3D$ to the spatial-C3D. As we shall show later, the optimal number of cross-link layers is between one and two, because the activations from the temporal-C3D at these early stages are still largely retinotopic. They directly point to the foreground regions and help the spatial-C3D learn distributed feature representation focused around these regions for the label prediction task. In the following, we report results for the case of only one attentive cross-link layer. Such cross-link layer includes the following three steps: reducing dimensions of a flow feature tensor $x_-flow^{\bar{l}}$ (Equ. 1), mean-variance normalization (Equ. 2), and attention prediction (Equ. 3). We use a 3d-convolutional layer to reduce a flow feature tensor $x\_flow^l\in\mathbb{R}^{C_l\times T_l\times H_l\times W_l}$ to $x\_link^l\in\mathbb{R}^{1\times T_l\times H_l\times W_l}:$
$$x_link^l=W_{3D_link}\circledast x_flow^l.$$
(1)
where $\otimes$ is a 3d-convolution operation along the channel dimension $C_l.$ We initialize the filter weights $W_{3D\_link}$ to $\frac1{C_l}$
in the training phase. Then, we normalize the feature tensor $x\_link^l$ by the mean $\mu$ and variance $\sigma$ of all the spatialtemporal feature activations in $x\_link^l:$
$$\hat{x}_{t,h,w}^l=\frac{x_link_{t,h,w}^l-\mu}{\sigma}.$$
(2)
The mean-variance normalization layer transforms the raw attention scores $x\textit{link}^l$ into a normalized range $\hat{x}^l\in$ [-1,1]. Finally, the normalized attention score $\hat{x}^l$ is converted to an attention probability score $a^l\in[0,1]$ by a sigmoid function:
(3)
$$a_{t,h,w}^l=\mathrm{sigmoid}(\hat{x}_{t,h,w}^l).$$
where $a_{h,w}^l\in\mathbb{R}^{1\times T_{l}\times H_{l}\times W_{l}}.$
We apply the flow-guided attention map on the feature map $x\_rgb^l$ of the spatial-C3D by multiplicative interaction:
$$x_rgb_{att}^l=\mathbf{r}(a^l,C_l)\odot x_rgb^l.$$
(4)
where r$(a^l,C_l)$ is the $C_l$-times replication of the predictive attention map $a^l$ along the channel dimension, and $\odot$ denotes element-wise multiplication operation.
The attended feature map $x\_rgb_{att}^l$ is forwarded into the
next layer $l+1$ to learn more abstract attended features:
$$x_rgb^{l+1}=f_{spatial_C3D}^{l+1}(x_rgb_{att}^{l}).$$
(5)
where $f_spatial\_C3D$ is the operation in the next layer $(e.g.$, convolution layer). Recall that we choose to have only one attentive cross-link layers because the activations in the higher layers of the temporal-C3D would be more abstract and not necessarily correspond to the notions of foreground objects.
## $4. \textbf{Experiments}$
### 4.1.Data sets
We evaluate the two-stream FCANs on three datasets for
action recognition.
UCF101[24]. This dataset is among the largest available action recognition benchmarks. UCF101 has 101 action classes and about 13320 videos (180 frames/video on average). There are three splits of training/testing data, and the performance is measured by mean classification accuracy across the splits
HMDB51[13]. HMDB51 has 51 action categories and 6,766 videos. The dataset has two versions (original and motion-stabilized), and we use the original version which is more challenging for action recognition. The dataset has diverse background contexts and variations in motion pattern. It has three train/test splits with 3,570 training and 1,530 test videos.
|
[
[
0.08087158203125,
0.09027099609375,
0.4697265625,
0.09027099609375,
0.4697265625,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.1094970703125,
0.10919189453125,
0.24267578125,
0.10919189453125,
0.24267578125,
0.12371826171875,
0.1094970703125,
0.12371826171875
],
[
0.50341796875,
0.106689453125,
0.890625,
0.106689453125,
0.890625,
0.12060546875,
0.50341796875,
0.12060546875
],
[
0.5009765625,
0.12249755859375,
0.7646484375,
0.121826171875,
0.7646484375,
0.1363525390625,
0.5009765625,
0.136962890625
],
[
0.1004638671875,
0.1337890625,
0.46728515625,
0.1337890625,
0.46728515625,
0.147705078125,
0.1004638671875,
0.147705078125
],
[
0.080078125,
0.149658203125,
0.468017578125,
0.149658203125,
0.468017578125,
0.1634521484375,
0.080078125,
0.1634521484375
],
[
0.6689453125,
0.147705078125,
0.78125,
0.1514892578125,
0.779296875,
0.169189453125,
0.66845703125,
0.1654052734375
],
[
0.611328125,
0.1558837890625,
0.646484375,
0.1654052734375,
0.6396484375,
0.179931640625,
0.6044921875,
0.17041015625
],
[
0.080078125,
0.1641845703125,
0.4697265625,
0.1641845703125,
0.4697265625,
0.177978515625,
0.080078125,
0.177978515625
],
[
0.86767578125,
0.1590576171875,
0.89453125,
0.1590576171875,
0.89453125,
0.175537109375,
0.86767578125,
0.175537109375
],
[
0.7236328125,
0.173583984375,
0.7294921875,
0.177978515625,
0.7255859375,
0.18115234375,
0.7197265625,
0.1767578125
],
[
0.08087158203125,
0.179931640625,
0.46728515625,
0.179931640625,
0.46728515625,
0.1900634765625,
0.08087158203125,
0.1900634765625
],
[
0.08087158203125,
0.1944580078125,
0.4697265625,
0.1944580078125,
0.4697265625,
0.2083740234375,
0.08087158203125,
0.2083740234375
],
[
0.52294921875,
0.1925048828125,
0.89208984375,
0.1925048828125,
0.89208984375,
0.205810546875,
0.52294921875,
0.205810546875
],
[
0.080078125,
0.210205078125,
0.468017578125,
0.210205078125,
0.468017578125,
0.2235107421875,
0.080078125,
0.2235107421875
],
[
0.50341796875,
0.2083740234375,
0.890625,
0.2083740234375,
0.890625,
0.2215576171875,
0.50341796875,
0.2215576171875
],
[
0.080078125,
0.2247314453125,
0.468017578125,
0.2247314453125,
0.468017578125,
0.238037109375,
0.080078125,
0.238037109375
],
[
0.5048828125,
0.2235107421875,
0.89208984375,
0.2235107421875,
0.89208984375,
0.2374267578125,
0.5048828125,
0.2374267578125
],
[
0.080078125,
0.240478515625,
0.468017578125,
0.240478515625,
0.468017578125,
0.25390625,
0.080078125,
0.25390625
],
[
0.5009765625,
0.2374267578125,
0.890625,
0.2392578125,
0.890625,
0.25390625,
0.5009765625,
0.251953125
],
[
0.07843017578125,
0.25390625,
0.468017578125,
0.255126953125,
0.468017578125,
0.26953125,
0.07843017578125,
0.268310546875
],
[
0.50341796875,
0.255126953125,
0.6044921875,
0.255126953125,
0.6044921875,
0.266357421875,
0.50341796875,
0.266357421875
],
[
0.07843017578125,
0.268310546875,
0.468017578125,
0.271484375,
0.468017578125,
0.285888671875,
0.07843017578125,
0.28271484375
],
[
0.60693359375,
0.281005859375,
0.78369140625,
0.27978515625,
0.78369140625,
0.298583984375,
0.60693359375,
0.2998046875
],
[
0.86767578125,
0.281005859375,
0.89453125,
0.281005859375,
0.89453125,
0.29736328125,
0.86767578125,
0.29736328125
],
[
0.0767822265625,
0.309326171875,
0.4697265625,
0.31005859375,
0.4697265625,
0.32763671875,
0.0767822265625,
0.326904296875
],
[
0.07843017578125,
0.28271484375,
0.1805419921875,
0.285400390625,
0.1805419921875,
0.2998046875,
0.07843017578125,
0.29736328125
],
[
0.5009765625,
0.31005859375,
0.6953125,
0.30615234375,
0.6953125,
0.324462890625,
0.5009765625,
0.328857421875
],
[
0.12091064453125,
0.32763671875,
0.2296142578125,
0.32763671875,
0.2296142578125,
0.341552734375,
0.12091064453125,
0.341552734375
],
[
0.52294921875,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.34228515625,
0.52294921875,
0.34228515625
],
[
0.50244140625,
0.343505859375,
0.890625,
0.343505859375,
0.890625,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.098876953125,
0.35107421875,
0.468017578125,
0.35107421875,
0.468017578125,
0.3642578125,
0.098876953125,
0.3642578125
],
[
0.50244140625,
0.35986328125,
0.5361328125,
0.35986328125,
0.5361328125,
0.371826171875,
0.50244140625,
0.371826171875
],
[
0.080078125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.5908203125,
0.37255859375,
0.79833984375,
0.371826171875,
0.79833984375,
0.389404296875,
0.5908203125,
0.39013671875
],
[
0.080078125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.080078125,
0.39453125
],
[
0.87109375,
0.375,
0.8935546875,
0.375,
0.8935546875,
0.38818359375,
0.87109375,
0.38818359375
],
[
0.080078125,
0.396484375,
0.471435546875,
0.396484375,
0.471435546875,
0.410400390625,
0.080078125,
0.410400390625
],
[
0.50244140625,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.412353515625,
0.50244140625,
0.412353515625
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.4267578125,
0.468017578125,
0.4267578125,
0.468017578125,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.5009765625,
0.427978515625,
0.79736328125,
0.4287109375,
0.79736328125,
0.443115234375,
0.5009765625,
0.442626953125
],
[
0.08087158203125,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.455078125,
0.08087158203125,
0.455078125
],
[
0.52294921875,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458984375,
0.52294921875,
0.458984375
],
[
0.07843017578125,
0.455078125,
0.4697265625,
0.455810546875,
0.4697265625,
0.47021484375,
0.07843017578125,
0.4697265625
],
[
0.50244140625,
0.460205078125,
0.87353515625,
0.460205078125,
0.87353515625,
0.473388671875,
0.50244140625,
0.473388671875
],
[
0.080078125,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.07843017578125,
0.48291015625,
0.4697265625,
0.48291015625,
0.4697265625,
0.50048828125,
0.07843017578125,
0.50048828125
],
[
0.57177734375,
0.481689453125,
0.81884765625,
0.48486328125,
0.81884765625,
0.51025390625,
0.57177734375,
0.5068359375
],
[
0.67236328125,
0.4892578125,
0.68310546875,
0.4892578125,
0.68310546875,
0.492431640625,
0.67236328125,
0.492431640625
],
[
0.869140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.869140625,
0.501953125
],
[
0.080078125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.51953125,
0.57421875,
0.5166015625,
0.576171875,
0.53076171875,
0.50341796875,
0.5341796875
],
[
0.56396484375,
0.51953125,
0.89208984375,
0.517578125,
0.89208984375,
0.53515625,
0.56396484375,
0.537109375
],
[
0.080078125,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50244140625,
0.53515625,
0.8896484375,
0.53515625,
0.8896484375,
0.548828125,
0.50244140625,
0.548828125
],
[
0.07843017578125,
0.546875,
0.468017578125,
0.54541015625,
0.468017578125,
0.56005859375,
0.07843017578125,
0.56103515625
],
[
0.5048828125,
0.5517578125,
0.8896484375,
0.5517578125,
0.8896484375,
0.5625,
0.5048828125,
0.5625
],
[
0.08087158203125,
0.5625,
0.4697265625,
0.5625,
0.4697265625,
0.57568359375,
0.08087158203125,
0.57568359375
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.080078125,
0.59228515625,
0.4697265625,
0.59228515625,
0.4697265625,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.50244140625,
0.59521484375,
0.55810546875,
0.59521484375,
0.55810546875,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.08087158203125,
0.609375,
0.46728515625,
0.609375,
0.46728515625,
0.619140625,
0.08087158203125,
0.619140625
],
[
0.080078125,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63623046875,
0.080078125,
0.63623046875
],
[
0.50341796875,
0.62548828125,
0.63134765625,
0.62548828125,
0.63134765625,
0.64013671875,
0.50341796875,
0.64013671875
],
[
0.080078125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.07843017578125,
0.65283203125,
0.468017578125,
0.65283203125,
0.468017578125,
0.66650390625,
0.07843017578125,
0.66650390625
],
[
0.50341796875,
0.6494140625,
0.60546875,
0.6494140625,
0.60546875,
0.6640625,
0.50341796875,
0.6640625
],
[
0.08087158203125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.08087158203125,
0.681640625
],
[
0.521484375,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.521484375,
0.6875
],
[
0.08251953125,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69580078125,
0.08251953125,
0.69580078125
],
[
0.50244140625,
0.68994140625,
0.62841796875,
0.68994140625,
0.62841796875,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.08087158203125,
0.7119140625
],
[
0.521484375,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71826171875,
0.521484375,
0.71826171875
],
[
0.080078125,
0.71337890625,
0.4697265625,
0.71337890625,
0.4697265625,
0.7265625,
0.080078125,
0.7265625
],
[
0.50341796875,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.08087158203125,
0.72900390625,
0.46728515625,
0.72900390625,
0.46728515625,
0.7421875,
0.08087158203125,
0.7421875
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.50341796875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08087158203125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.77294921875,
0.08087158203125,
0.77294921875
],
[
0.080078125,
0.7734375,
0.468017578125,
0.7734375,
0.468017578125,
0.787109375,
0.080078125,
0.787109375
],
[
0.5048828125,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.0767822265625,
0.7861328125,
0.47216796875,
0.7841796875,
0.47216796875,
0.80322265625,
0.0767822265625,
0.8046875
],
[
0.5009765625,
0.78076171875,
0.64453125,
0.78076171875,
0.64453125,
0.79443359375,
0.5009765625,
0.79443359375
],
[
0.521484375,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.521484375,
0.8095703125
],
[
0.07843017578125,
0.8037109375,
0.253173828125,
0.798828125,
0.253173828125,
0.81298828125,
0.07843017578125,
0.818359375
],
[
0.5009765625,
0.8095703125,
0.89208984375,
0.81005859375,
0.89208984375,
0.82470703125,
0.5009765625,
0.82373046875
],
[
0.16259765625,
0.830078125,
0.381591796875,
0.8291015625,
0.381591796875,
0.84765625,
0.16259765625,
0.8486328125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.44677734375,
0.83349609375,
0.4697265625,
0.83349609375,
0.4697265625,
0.8466796875,
0.44677734375,
0.8466796875
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.861328125,
0.468017578125,
0.85986328125,
0.468017578125,
0.875,
0.080078125,
0.87548828125
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.0767822265625,
0.87890625,
0.449462890625,
0.880859375,
0.449462890625,
0.8984375,
0.0767822265625,
0.896484375
],
[
0.443603515625,
0.8916015625,
0.46728515625,
0.8916015625,
0.46728515625,
0.90380859375,
0.443603515625,
0.90380859375
],
[
0.50341796875,
0.88818359375,
0.580078125,
0.88818359375,
0.580078125,
0.8994140625,
0.50341796875,
0.8994140625
]
] |
[
"3. Flow-guided Convolutional Attention Net-",
"in the training phase. Then, we normalize the feature tensor",
"works (FCANs)",
"r_linkl by the mean μ and variance of all the spatial-",
"temporal feature activations in r_linkl:",
"We propose the cross-link layers to model the interac-",
"tions of the two networks in the two-stream convolutional",
"a_linkt,h,w-μ",
"tth,",
"network [22]. The whole network is differential, so it can",
"(2)",
"0",
"be trained end-to-end with the stochastic gradient descent",
"(SGD) and back-propagation algorithm [15]. The overall",
"The mean-variance normalization layer transforms the",
"architecture of the FCAN is shown in Figure 1. In the en-",
"raw attention scores r_linkinto a normalized range E",
"suing discussion, we describe the 3D-FCAN model, which",
"[-1,1].Finally, the normalized attention score is con-",
"is the 3D version based on 3D convolution (e.g., C3D [26])",
"verted to an attention probability score a e [o, 1] by a sig-",
"building blocks (please refer to the Supplementary Materi-",
"moid function:",
"als for the 2D-FCAN model based on 2D convolution (e.g.",
"at,h,w = sigmoid(t,h,w)",
"(3)",
"3.1. 3D version of flow-guided convolutional atten-",
"Alexnet [12]]).",
"where dh.w. E IR1xT+xH,xW,",
"tion networks",
"We apply the flow-guided attention map on the feature",
"map &-rgbl of the spatial-C3D by multiplicative interac-",
"With the assumption that the magnitudes of optical",
"tion:",
"flows, when appropriately compensated for camera mo-",
"r-rgbatt = r(al,Ci) oa-rgbl.",
"tions, usually correlate with the foreground regions, we de-",
"(4)",
"velop the framework of flow-guided convolutional attention",
"where r(a', Cr) is the Cr-times replication of the predictive",
"networks (FCANs) as shown in Figure 1. The C3D net-",
"attention map a' along the channel dimension, and de-",
"work [26] provides explicit representation of the time di-",
"notes element-wise multiplication operation.",
"mension in the architecture. Both the 3D-FCAN and 2D-",
"The attended feature map -rgbatt is forwarded into the",
"FCAN have similar structures, except for the operations in",
"next layer I + 1 to learn more abstract attended features:",
"the convolution and mean-variance normalization layer. Let",
"-rgbl e IRCixTixHixWt, a_flowl e RCixTixHixWi be",
"a-rgb+1 = futtial.C3D(a-rgbatt).",
"+",
"(5)",
"the feature map of layer I e fo, 1, ..., L) in the spatial-",
"and temporal-C3D respectively, with Ci, Ti, Hi, Wi being",
"where fl+",
"patial.C3D is the operation in the next layer (e.g.,",
"the number of channels, temporal length, height and width",
"convolution layer). Recall that we choose to have only",
"of the feature map.Specifically,in the proposed FCANs,",
"oneattentivecross-linklayersbecausetheactivationsinthe",
"r_rgb and _flow are the feature maps from a pooling layer",
"higher layers of the temporal-C3D would be more abstract",
"in the C3D network. We develop attentive cross-link layers",
"and not necessarily correspond to the notions of foreground",
"between the early pooling layers from the temporal-C3D to",
"objects.",
"the spatial-C3D.As we shall showlater.the optimal num-",
"ber of cross-link layers is between one and two, because",
"4.Experiments",
"the activations from the temporal-C3D at these early stages",
"are still largely retinotopic. They directly point to the fore-",
"4.1. Data sets",
"ground regions and help the spatial-C3D learn distributed",
"We evaluate the two-stream FCANs on three datasets for",
"featurerepresentationfocusedaroundtheseregionsforthe",
"action recognition.",
"label prediction task. In the following, we report results",
"UCF101 [24]. This dataset is among the largest avail-",
"for the case of only one attentive cross-link layer.Such",
"able action recognition benchmarks.UCF101 has 101 ac-",
"cross-link layer includes the following three steps: reduc-",
"tion classes and about 13320 videos (180 frames/video on",
"ing dimensions of a flow feature tensor _flow (Equ. 1),",
"average). There are three splits of training/testing data, and",
"mean-variance normalization (Equ. 2), and attention pre-",
"diction (Equ. 3).We use a 3d-convolutional layer to re-",
"theperformanceismeasuredbymeanclassificationaccu-",
"duce a flow feature tensor r-flowt e RCixTrxHixwi to",
"racy across the splits.",
"HMDB51 [13]. HMDB51 has 51 action categories and",
"r_linkleRixTixHixWi.",
"6,766 videos. The dataset has two versions (original and",
"rlink=WaDlink ?a-flowl.",
"motion-stabilized), and we use the original version which is",
"(1)",
"more challenging for action recognition. The dataset has di-",
"where is a 3d-convolution operation along the channel di-",
"verse background contexts and variations in motion pattern.",
"It has three train/test splits with 3,570 training and 1,530",
"mension Ci. We initialize the filter weights W3D.link to",
"Ci",
"test videos."
] |
{
"arxiv_id": "1708.09268",
"language": "en",
"page_idx": 2
}
|
||
2110.01529-en-0
|
2110.01529-en-0.jpg
|
# A Proposed Conceptual Framework for a $\textbf{Representational Approach to Information Retrieval}$
Jimmy Lin
David R. Cheriton School of Computer Science
University of Waterloo
## Abstract
This paper outlines a conceptual framework for understanding recent developments in information retrieval and natural language processing that attempts to integrate dense and sparse retrieval methods. I propose a representational approach that breaks the core text retrieval problem into a logical scoring model and a physical retrieval model. The scoring model is defined in terms of encoders, which map queries and documents into a representational space, and a comparison function $\hat{\text{that computes query-document scores. The physical retrieval model defines how}}$ a system produces the top-$k$ scoring documents from an arbitrarily large corpus with respect to a query. The scoring model can be further analyzed along two dimensions: dense vs. sparse representations and supervised (learned) vs. unsupervised approaches. I show that many recently proposed retrieval methods, including multi-stage ranking designs, can be seen as different parameterizations in this framework, and that a unified view suggests a number of open research questions, providing a roadmap for future work. As a bonus, this conceptual framework establishes connections to sentence similarity tasks in natural language processing and information access “technologies”prior to the dawn of computing.
### 1 Introduction
For the past half a century, information retrieval has been dominated by bag-of-words exact-match scoring models such as BM25 executed at scale using inverted indexes and efficient query-at-atime retrieval algorithms. Even in the context of feature-based learning to rank and, more recently, neural models, these bag-of-words models remain of fundamental importance because they provide potentially relevant texts for downstream reranking in the context of multi-stage pipelines. This role is usually referred to as fırst-stage retrieval or candidate generation. Multi-stage ranking architectures have been studied extensively by academic researchers [Matveeva et al., 2006, Cambazoglu et al., 2010, Wang et al., 2011, Tonellotto et al., 2013, Asadi and Lin, 2013, Capannini et al., 2016, Clarke et al., 2016, Chen et al., 2017, Mackenzie et al., 2018] and there is substantial documentation that many commercial applications are designed in this manner [Pedersen, 2010, Liu et al., 2017, Huang et al., 2020, Zou et al., 2021].
There has, of late, been much interest and excitement surrounding so-called“dense retrieval”techniques, or ranking with learned dense representations. This general approach, often called a bi-encoder design [Humeau et al., 2020], is perhaps best exemplified by DPR [Karpukhin et al., 2020] and ANCE[Xiong et al., 2021], but other examples abound [Gao et al., 2021b, Hofstätter et al., 2020, Qu et al., 2021, Hofstätter et al., 2021, Qu et al., 2021, Zhan et al., 2021, Lin et al., 2021c]. Dense retrieval is formulated as a representational learning problem where the task is to learn (nowadays, transformer-based) encoders that map queries and documents into dense fixed-width vectors (768 dimensions is typical). The goal is to maximize inner products between queries and relevant documents and to minimize inner products between queries and non-relevant documents. This is framed as a supervised machine learning problem, with relevance signals coming from a large dataset such
|
[
[
0.2484130859375,
0.125,
0.7490234375,
0.126953125,
0.7490234375,
0.147705078125,
0.2484130859375,
0.1458740234375
],
[
0.1829833984375,
0.15087890625,
0.81787109375,
0.149658203125,
0.81787109375,
0.17041015625,
0.1829833984375,
0.1717529296875
],
[
0.461669921875,
0.2266845703125,
0.54150390625,
0.2266845703125,
0.54150390625,
0.2412109375,
0.461669921875,
0.2412109375
],
[
0.343017578125,
0.246826171875,
0.65673828125,
0.246826171875,
0.65673828125,
0.261474609375,
0.343017578125,
0.261474609375
],
[
0.423095703125,
0.261474609375,
0.576171875,
0.261474609375,
0.576171875,
0.274658203125,
0.423095703125,
0.274658203125
],
[
0.0228729248046875,
0.272705078125,
0.060455322265625,
0.272705078125,
0.060455322265625,
0.380615234375,
0.0228729248046875,
0.380615234375
],
[
0.459228515625,
0.309326171875,
0.5400390625,
0.311279296875,
0.5400390625,
0.328857421875,
0.459228515625,
0.326904296875
],
[
0.233642578125,
0.34033203125,
0.765625,
0.341552734375,
0.765625,
0.35595703125,
0.233642578125,
0.354736328125
],
[
0.233642578125,
0.35595703125,
0.76708984375,
0.35595703125,
0.76708984375,
0.37060546875,
0.233642578125,
0.37060546875
],
[
0.02532958984375,
0.369384765625,
0.056365966796875,
0.369384765625,
0.056365966796875,
0.40771484375,
0.02532958984375,
0.40771484375
],
[
0.233642578125,
0.37060546875,
0.76708984375,
0.37060546875,
0.76708984375,
0.3837890625,
0.233642578125,
0.3837890625
],
[
0.233642578125,
0.3837890625,
0.76708984375,
0.3837890625,
0.76708984375,
0.397705078125,
0.233642578125,
0.397705078125
],
[
0.233642578125,
0.396484375,
0.765625,
0.396484375,
0.765625,
0.410888671875,
0.233642578125,
0.410888671875
],
[
0.02532958984375,
0.41357421875,
0.060455322265625,
0.41357421875,
0.060455322265625,
0.488037109375,
0.02532958984375,
0.488037109375
],
[
0.233642578125,
0.410888671875,
0.765625,
0.410888671875,
0.765625,
0.4248046875,
0.233642578125,
0.4248046875
],
[
0.233642578125,
0.425537109375,
0.765625,
0.425537109375,
0.765625,
0.439453125,
0.233642578125,
0.439453125
],
[
0.233642578125,
0.438232421875,
0.765625,
0.438232421875,
0.765625,
0.45263671875,
0.233642578125,
0.45263671875
],
[
0.233642578125,
0.45263671875,
0.39453125,
0.45263671875,
0.39453125,
0.466552734375,
0.233642578125,
0.466552734375
],
[
0.3896484375,
0.45263671875,
0.7646484375,
0.45263671875,
0.7646484375,
0.466552734375,
0.3896484375,
0.466552734375
],
[
0.233642578125,
0.466552734375,
0.76708984375,
0.466552734375,
0.76708984375,
0.479736328125,
0.233642578125,
0.479736328125
],
[
0.047393798828125,
0.481689453125,
0.052276611328125,
0.481689453125,
0.052276611328125,
0.4873046875,
0.047393798828125,
0.4873046875
],
[
0.233642578125,
0.479736328125,
0.765625,
0.479736328125,
0.765625,
0.49365234375,
0.233642578125,
0.49365234375
],
[
0.021240234375,
0.4892578125,
0.05963134765625,
0.4892578125,
0.05963134765625,
0.7119140625,
0.021240234375,
0.7119140625
],
[
0.2344970703125,
0.49560546875,
0.7646484375,
0.49560546875,
0.7646484375,
0.505859375,
0.2344970703125,
0.505859375
],
[
0.233642578125,
0.50830078125,
0.76708984375,
0.50830078125,
0.76708984375,
0.521484375,
0.233642578125,
0.521484375
],
[
0.233642578125,
0.521484375,
0.76708984375,
0.521484375,
0.76708984375,
0.53515625,
0.233642578125,
0.53515625
],
[
0.233642578125,
0.53515625,
0.76708984375,
0.53515625,
0.76708984375,
0.548828125,
0.233642578125,
0.548828125
],
[
0.233642578125,
0.5498046875,
0.69677734375,
0.5498046875,
0.69677734375,
0.56298828125,
0.233642578125,
0.56298828125
],
[
0.176513671875,
0.5888671875,
0.19287109375,
0.5888671875,
0.19287109375,
0.5986328125,
0.176513671875,
0.5986328125
],
[
0.1993408203125,
0.5859375,
0.31201171875,
0.58837890625,
0.31201171875,
0.60302734375,
0.1993408203125,
0.60107421875
],
[
0.1732177734375,
0.6171875,
0.82421875,
0.6171875,
0.82421875,
0.630859375,
0.1732177734375,
0.630859375
],
[
0.1732177734375,
0.630859375,
0.826171875,
0.6318359375,
0.826171875,
0.646484375,
0.1732177734375,
0.64599609375
],
[
0.1748046875,
0.646484375,
0.826171875,
0.646484375,
0.826171875,
0.66015625,
0.1748046875,
0.66015625
],
[
0.1748046875,
0.66015625,
0.826171875,
0.66015625,
0.826171875,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1748046875,
0.673828125,
0.826171875,
0.673828125,
0.826171875,
0.6875,
0.1748046875,
0.6875
],
[
0.1748046875,
0.6875,
0.82421875,
0.6875,
0.82421875,
0.70068359375,
0.1748046875,
0.70068359375
],
[
0.1732177734375,
0.69970703125,
0.82666015625,
0.70068359375,
0.82666015625,
0.71533203125,
0.1732177734375,
0.7138671875
],
[
0.1748046875,
0.7138671875,
0.82421875,
0.7138671875,
0.82421875,
0.72900390625,
0.1748046875,
0.72900390625
],
[
0.1748046875,
0.72802734375,
0.82421875,
0.72802734375,
0.82421875,
0.7412109375,
0.1748046875,
0.7412109375
],
[
0.1748046875,
0.7421875,
0.82421875,
0.7421875,
0.82421875,
0.755859375,
0.1748046875,
0.755859375
],
[
0.1748046875,
0.7548828125,
0.370849609375,
0.7548828125,
0.370849609375,
0.76953125,
0.1748046875,
0.76953125
],
[
0.1748046875,
0.77587890625,
0.82666015625,
0.77587890625,
0.82666015625,
0.79052734375,
0.1748046875,
0.79052734375
],
[
0.1748046875,
0.79052734375,
0.82421875,
0.79052734375,
0.82421875,
0.8046875,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.8037109375,
0.82421875,
0.8037109375,
0.82421875,
0.8173828125,
0.1748046875,
0.8173828125
],
[
0.1732177734375,
0.8173828125,
0.82421875,
0.8173828125,
0.82421875,
0.83203125,
0.1732177734375,
0.83203125
],
[
0.1748046875,
0.83056640625,
0.82421875,
0.83056640625,
0.82421875,
0.8447265625,
0.1748046875,
0.8447265625
],
[
0.1748046875,
0.84521484375,
0.826171875,
0.84521484375,
0.826171875,
0.859375,
0.1748046875,
0.859375
],
[
0.17236328125,
0.85791015625,
0.82421875,
0.859375,
0.82421875,
0.87353515625,
0.17236328125,
0.87255859375
],
[
0.1732177734375,
0.87255859375,
0.82275390625,
0.87255859375,
0.82275390625,
0.88720703125,
0.1732177734375,
0.88720703125
],
[
0.1748046875,
0.88720703125,
0.826171875,
0.88720703125,
0.826171875,
0.90087890625,
0.1748046875,
0.90087890625
],
[
0.1732177734375,
0.90087890625,
0.82421875,
0.90087890625,
0.82421875,
0.9140625,
0.1732177734375,
0.9140625
]
] |
[
"A Proposed Conceptual Framework for a",
"Representational Approach to Information Retrieval",
"JimmyLin",
"David R. Cheriton School of Computer Science",
"University of Waterloo",
"0",
"Abstract",
"This paper outlines a conceptual framework for understanding recent developments",
"in information retrieval and natural language processing that attempts to integrate",
"8",
"dense and sparse retrieval methods. I propose a representational approach that",
"breaks the core text retrieval problem into a logical scoring model and a physical",
"retrieval model. The scoring model is defined in terms of encoders, which map",
"['s]",
"queries and documents into a representational space, and a comparison function",
"that computes query-document scores. The physical retrieval model defines how",
"a system produces the top-k scoring documents from an arbitrarily large corpus",
"with respect to a query.",
"The scoring model can be further analyzed along two",
"dimensions: dense vs. sparse representations and supervised (learned) vs. unsuper-",
"",
"vised approaches. I show that many recently proposed retrieval methods, including",
"A62S0011:A10",
"multi-stageranking designs,can be seen as differentparameterizations inthis",
"framework, and that a unified view suggests a number of open research questions,",
"providing a roadmap for future work. As a bonus, this conceptual framework",
"establishes connections to sentence similarity tasks in natural language processing",
"and information access\"technologies\"prior to the dawn of computing.",
"1.",
"Introduction",
"For the past half a century, information retrieval has been dominated by bag-of-words exact-match",
"scoring models such as BM25 executed at scale using inverted indexes and efficient query-at-a-",
"time retrieval algorithms. Even in the context of feature-based learning to rank and, more recently,",
"neural models, these bag-of-words models remain of fundamental importance because they provide",
"potentially relevant texts for downstream reranking in the context of multi-stage pipelines. This role",
"is usually referred to as first-stage retrieval or candidate generation. Multi-stage ranking architectures",
"have been studied extensively by academic researchers [Matveeva et al., 2006, Cambazoglu et al.,",
"2010, Wang et al., 2011, Tonellotto et al., 2013, Asadi and Lin, 2013, Capannini et al., 2016, Clarke",
"et al., 2016, Chen et al., 2017, Mackenzie et al., 2018] and there is substantial documentation that",
"many commercial applications are designed in this manner [Pedersen, 2010, Liu et al., 2017, Huang",
"et al., 2020, Zou et al., 2021].",
"There has, of late, been much interest and excitement surrounding so-called \"dense retrieval\" tech-",
"niques, or ranking with learned dense representations. This general approach, often called a bi-encoder",
"design [Humeau et al., 2020], is perhaps best exemplified by DPR [Karpukhin et al., 2020] and",
"ANCE [Xiong et al., 2021], but other examples abound [Gao et al., 2021b, Hofstatter et al., 2020,",
"Qu et al., 2021, Hofstatter et al., 2021, Qu et al., 2021, Zhan et al., 2021, Lin et al., 2021cl. Dense",
"retrieval is formulated as a representational learning problem where the task is to learn (nowadays,",
"transformer-based) encoders that map queries and documents into dense fixed-width vectors (768",
"dimensions is typical). The goal is to maximize inner products between queries and relevant docu.",
"ments and to minimize inner products between queries and non-relevant documents. This is framed",
"as a supervised machine learning problem, with relevance signals coming from a large dataset such"
] |
{
"arxiv_id": "2110.01529",
"language": "en",
"page_idx": 0
}
|
||
2110.01529-en-1
|
2110.01529-en-1.jpg
|
as the MS MARCO passage ranking test collection [Bajaj et al., 2018]. Lin et al.[2021b] provide a recent survey of this general approach within the broader context of text ranking using BERT and other pretrained transformer-based language models.
Experiments have shown that dense retrieval methods outperform “sparse retrieval” methods, usually referring to bag-of-words exact-match methods such as BM25.$^{1}$ This appears to be a robust and widely replicated finding, and dense retrieval models are known to have been deployed in real-world search applications, for example, by Bing [Xiong et al., 2021] and Facebook [Huang et al., 2020]. Scaling such methods requires infrastructure that is very different from sparse retrieval: instead of relying on inverted indexes for query evaluation, as BM25 does, dense retrieval typically relies on approximate nearest neighbor (ANN) search; one standard technique exploits hierarchical navigable small world graphs (HNSW) [Malkov and Yashunin, 2020].
Thus, recent literature appears to have established a contrast between dense retrieval and sparse retrieval. The standard portrayal is that they represent fundamentally different approaches, requiring different problem formulations, different models, and different software infrastructures for efficient execution at scale. I argue, however, that this is not the case. Aspects of the ideas and observations presented here were originally captured in two previous papers [Lin and Ma, 2021,Lin et al., 2021a]. I build on both, with additional analysis and synthesis.
The goal of this paper is to provide a conceptual framework that unites dense and sparse retrieval by demonstrating that they, in fact, have the same functional form, just with different parameterizations. This framework adopts a representational approach and breaks the core text retrieval problem into a logical scoring model and a physical retrieval model, allowing a researcher to separate how document relevance scores are computed from how retrieval is performed at scale. In terms of scoring models, dense and sparse retrieval can be characterized along two dimensions: the contrast between dense vs. sparse vector representations, and the contrast between supervised (learned) vs. unsupervised $\tilde{\mathrm{approaches.}}$
The main contribution of this conceptual framework is that it provides abstractions to help researchers make sense of the panoply of recently proposed retrieval models that, at first glance, defy orderly categorization. The proposed framework suggests a number of open research questions, providing a roadmap for future research, potentially tying together multiple sub-fields within information retrieval As a bonus, this conceptual framework establishes interesting connections to sentence similarity tasks in natural language processing and information access “technologies”prior to the dawn of computing.
# 2 A Conceptual Framework
The formulation of text retrieval (alternatively, text ranking)—what information retrieval researchers more precisely call $ad\textit{hoc retrieval— is typically defined as follows: Given an information need}$ expressed as a query $q$,the text retrieval task is to return a ranked list of $k$ documents$^2\left\{d_1,d_2\ldots d_k\right\}$ from an arbitrarily large but finite collection of documents $\mathcal{D}=\{d_i\}$ that maximizes a metric of interest, for example, nDCG, AP, etc. These metrics vary, but they all aim to quantify the “goodness” of the results with respect to the information need; in some cases, metrics can be understood more formally in terms of the utility that a user would derive from consuming the results. The retrieval task is also called top-$k$ retrieval (or ranking), where $k$ is the length of the ranked list (also known as the retrieval or ranking depth).
We can break the text retrieval problem down into two distinct components, as follows:
Logical Scoring Model Let us define $\eta_q(q)$ and $\eta_d(d)$ as two arbitrary functions that take a query and a document (both sequences of terms), respectively, and map each into a fixed-width vector representation. As will become clear below, I will call these two functions “encoders”.
Referring to bag-of-words exact-match methods as“sparse retrieval”is a relatively new invention, primarily to establish contrast with dense retrieval methods. Nevertheless, I will use this terminology throughout the paper. $^{2}$Consistent with parlance in information retrieval, I use“document”throughout this paper in a generic sense to refer to the unit of retrieved text, even though in truth it may be a passage, a web page, a PDF, or some arbitrary span of text.
2
|
[
[
0.1748046875,
0.091552734375,
0.82275390625,
0.091552734375,
0.82275390625,
0.1060791015625,
0.1748046875,
0.1060791015625
],
[
0.1748046875,
0.106689453125,
0.82421875,
0.106689453125,
0.82421875,
0.12060546875,
0.1748046875,
0.12060546875
],
[
0.1748046875,
0.12060546875,
0.52392578125,
0.12060546875,
0.52392578125,
0.1337890625,
0.1748046875,
0.1337890625
],
[
0.1748046875,
0.14013671875,
0.82177734375,
0.14013671875,
0.82177734375,
0.154052734375,
0.1748046875,
0.154052734375
],
[
0.17236328125,
0.154052734375,
0.82275390625,
0.15283203125,
0.82275390625,
0.1673583984375,
0.17236328125,
0.1685791015625
],
[
0.1748046875,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.1748046875,
0.1817626953125
],
[
0.1748046875,
0.18310546875,
0.82275390625,
0.18310546875,
0.82275390625,
0.1962890625,
0.1748046875,
0.1962890625
],
[
0.1748046875,
0.1962890625,
0.826171875,
0.1962890625,
0.826171875,
0.210205078125,
0.1748046875,
0.210205078125
],
[
0.1748046875,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1748046875,
0.2235107421875,
0.82275390625,
0.2235107421875,
0.82275390625,
0.2374267578125,
0.1748046875,
0.2374267578125
],
[
0.17236328125,
0.2374267578125,
0.56884765625,
0.236083984375,
0.56884765625,
0.250732421875,
0.17236328125,
0.251953125
],
[
0.1732177734375,
0.25830078125,
0.82177734375,
0.25830078125,
0.82177734375,
0.271484375,
0.1732177734375,
0.271484375
],
[
0.1748046875,
0.271484375,
0.82275390625,
0.271484375,
0.82275390625,
0.285888671875,
0.1748046875,
0.285888671875
],
[
0.1748046875,
0.285400390625,
0.82421875,
0.285400390625,
0.82421875,
0.298583984375,
0.1748046875,
0.298583984375
],
[
0.1748046875,
0.2998046875,
0.82275390625,
0.2998046875,
0.82275390625,
0.313232421875,
0.1748046875,
0.313232421875
],
[
0.1748046875,
0.313232421875,
0.826171875,
0.313232421875,
0.826171875,
0.326904296875,
0.1748046875,
0.326904296875
],
[
0.1732177734375,
0.32568359375,
0.529296875,
0.32568359375,
0.529296875,
0.339111328125,
0.1732177734375,
0.339111328125
],
[
0.1748046875,
0.347900390625,
0.82275390625,
0.347900390625,
0.82275390625,
0.361083984375,
0.1748046875,
0.361083984375
],
[
0.176513671875,
0.363037109375,
0.82275390625,
0.363037109375,
0.82275390625,
0.373779296875,
0.176513671875,
0.373779296875
],
[
0.1748046875,
0.375732421875,
0.82421875,
0.375732421875,
0.82421875,
0.389404296875,
0.1748046875,
0.389404296875
],
[
0.1748046875,
0.389404296875,
0.82275390625,
0.389404296875,
0.82275390625,
0.40283203125,
0.1748046875,
0.40283203125
],
[
0.1748046875,
0.404052734375,
0.82275390625,
0.404052734375,
0.82275390625,
0.4140625,
0.1748046875,
0.4140625
],
[
0.1732177734375,
0.417236328125,
0.82275390625,
0.417236328125,
0.82275390625,
0.43115234375,
0.1732177734375,
0.43115234375
],
[
0.1732177734375,
0.43115234375,
0.82275390625,
0.43115234375,
0.82275390625,
0.4443359375,
0.1732177734375,
0.4443359375
],
[
0.17236328125,
0.4443359375,
0.253173828125,
0.442626953125,
0.253173828125,
0.45703125,
0.17236328125,
0.458984375
],
[
0.1732177734375,
0.464111328125,
0.82275390625,
0.464111328125,
0.82275390625,
0.477783203125,
0.1732177734375,
0.477783203125
],
[
0.1748046875,
0.479248046875,
0.82275390625,
0.479248046875,
0.82275390625,
0.492431640625,
0.1748046875,
0.492431640625
],
[
0.1748046875,
0.492431640625,
0.82275390625,
0.492431640625,
0.82275390625,
0.505859375,
0.1748046875,
0.505859375
],
[
0.1732177734375,
0.505859375,
0.826171875,
0.5048828125,
0.826171875,
0.51953125,
0.1732177734375,
0.52099609375
],
[
0.176513671875,
0.521484375,
0.82177734375,
0.521484375,
0.82177734375,
0.5322265625,
0.176513671875,
0.5322265625
],
[
0.1732177734375,
0.5341796875,
0.826171875,
0.5341796875,
0.826171875,
0.54736328125,
0.1732177734375,
0.54736328125
],
[
0.1748046875,
0.57568359375,
0.423095703125,
0.57568359375,
0.423095703125,
0.59033203125,
0.1748046875,
0.59033203125
],
[
0.176513671875,
0.61279296875,
0.8203125,
0.61279296875,
0.8203125,
0.62353515625,
0.176513671875,
0.62353515625
],
[
0.1748046875,
0.62548828125,
0.82275390625,
0.62548828125,
0.82275390625,
0.6396484375,
0.1748046875,
0.6396484375
],
[
0.1732177734375,
0.63818359375,
0.82177734375,
0.6396484375,
0.82177734375,
0.65380859375,
0.1732177734375,
0.65283203125
],
[
0.176513671875,
0.65478515625,
0.82177734375,
0.65478515625,
0.82177734375,
0.66552734375,
0.176513671875,
0.66552734375
],
[
0.1748046875,
0.66748046875,
0.82421875,
0.66748046875,
0.82421875,
0.68115234375,
0.1748046875,
0.68115234375
],
[
0.176513671875,
0.68115234375,
0.82177734375,
0.68115234375,
0.82177734375,
0.6943359375,
0.176513671875,
0.6943359375
],
[
0.1748046875,
0.6943359375,
0.82275390625,
0.6943359375,
0.82275390625,
0.70849609375,
0.1748046875,
0.70849609375
],
[
0.1732177734375,
0.708984375,
0.82421875,
0.708984375,
0.82421875,
0.72265625,
0.1732177734375,
0.72265625
],
[
0.1748046875,
0.72265625,
0.351318359375,
0.72265625,
0.351318359375,
0.736328125,
0.1748046875,
0.736328125
],
[
0.1732177734375,
0.7412109375,
0.74267578125,
0.7421875,
0.74267578125,
0.7568359375,
0.1732177734375,
0.755859375
],
[
0.1748046875,
0.77978515625,
0.82275390625,
0.77978515625,
0.82275390625,
0.79443359375,
0.1748046875,
0.79443359375
],
[
0.176513671875,
0.79541015625,
0.82275390625,
0.79541015625,
0.82275390625,
0.80615234375,
0.176513671875,
0.80615234375
],
[
0.1748046875,
0.80810546875,
0.74169921875,
0.80810546875,
0.74169921875,
0.8212890625,
0.1748046875,
0.8212890625
],
[
0.1732177734375,
0.84326171875,
0.82666015625,
0.84521484375,
0.82666015625,
0.86279296875,
0.1732177734375,
0.861328125
],
[
0.178955078125,
0.85791015625,
0.82275390625,
0.85986328125,
0.82275390625,
0.87353515625,
0.178955078125,
0.8720703125
],
[
0.1732177734375,
0.8720703125,
0.82421875,
0.87353515625,
0.82421875,
0.88818359375,
0.1732177734375,
0.88623046875
],
[
0.1805419921875,
0.88818359375,
0.82275390625,
0.88818359375,
0.82275390625,
0.8984375,
0.1805419921875,
0.8984375
],
[
0.1805419921875,
0.90087890625,
0.2548828125,
0.90087890625,
0.2548828125,
0.9111328125,
0.1805419921875,
0.9111328125
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"as the MS MARCO passage ranking test collection [Bajaj et al., 2018]. Lin et al. [2021b] provide a",
"recent survey of this general approach within the broader context of text ranking using BERT and",
"other pretrained transformer-based language models.",
"Experiments have shown that dense retrieval methods outperform \"sparse retrieval\" methods, usually",
"referring to bag-of-words exact-match methods such as BM25.1 This appears to be a robust and",
"widely replicated finding, and dense retrieval models are known to have been deployed in real-world",
"search applications, for example, by Bing [Xiong et al., 2021] and Facebook [Huang et al., 2020]",
"Scaling such methods requires infrastructure that is very different from sparse retrieval: instead of",
"relying on inverted indexes for query evaluation, as BM25 does, dense retrieval typically relies on",
"approximate nearest neighbor (ANN) search; one standard technique exploits hierarchical navigable",
"small world graphs (HNSW) [Malkov and Yashunin, 2020].",
"Thus, recent literature appears to have established a contrast between dense retrieval and sparse",
"retrieval. The standard portrayal is that they represent fundamentally different approaches, requiring",
"different problem formulations, different models, and different software infrastructures for efficient",
"execution at scale. I argue, however, that this is not the case. Aspects of the ideas and observations",
"presented here were originally captured in two previous papers [Lin and Ma, 2021, Lin et al., 2021a].",
"I build on both, with additional analysis and synthesis.",
"The goal of this paper is to provide a conceptual framework that unites dense and sparse retrieval by",
"demonstrating thatthey,infact,have the samefunctional form,just with different parameterizations",
"This framework adopts a representational approach and breaks the core text retrieval problem into a",
"logical scoring model and a physical retrieval model, allowing a researcher to separate how document",
"relevancescoresarecomputedfromhowretrievalisperformed atscale.Interms of scoringmodels",
"dense and sparse retrieval can be characterized along two dimensions: the contrast between dense",
"vs. sparse vector representations, and the contrast between supervised (learned) vs. unsupervised",
"approaches.",
"The main contribution of this conceptual framework is that it provides abstractions to help researchers",
"make sense of the panoply of recently proposed retrieval models that, at first glance, defy orderly",
"categorization. The proposed framework suggests a number of open research questions, providing a",
"roadmap for future research, potentially tying together multiple sub-fields within information retrieval.",
"Asabonus,this conceptualframework establishesinterestingconnectionsto sentencesimilaritytasks",
"in natural language processing and information access technologies\"prior to the dawn of computing.",
"2AConceptualFramework",
"Theformulationoftextretrieval(alternatively.textranking)whatinformationretrieval researchers",
"more precisely call ad hoc retrieval-is typically defined as follows: Given an information need",
"expressed as a query q, the text retrieval task is to return a ranked list of k documents? (di, d2 . . . d]",
"froman arbitrarily largebutfinite collection ofdocuments D=fd,thatmaximizesametric of",
"interest, for example, nDCG, AP, etc. These metrics vary, but they all aim to quantify the “goodness\"",
"of the results with respect to the information need; in some cases, metrics can be understood more",
"formally in terms of the utility that a user would derive from consuming the results. The retrieval task",
"is also called top-k retrieval (or ranking), where k is the length of the ranked list (also known as the",
"retrieval or ranking depth).",
"We can break the text retrieval problem down into two distinct components, as follows:",
"Logical Scoring Model Let us define ng(q) and na(d) as two arbitrary functions that take a query",
"and adocument(bothsequencesofterms),respectively,andmap eachinto afixed-widthvector",
"representation. As will become clear below, I will call these two functions \"encoders\".",
"'Referring to bag-of-words exact-match methods as \"sparse retrieval\" is a relatively new invention, primarily to",
"establish contrast with dense retrieval methods. Nevertheless, I will use this terminology throughout the paper.",
"2Consistent with parlance in information retrieval, I use “document\" throughout this paper in a generic sense to",
"refer to the unit of retrieved text, even though in truth it may be a passage, a webpage, a PDF, or some arbitrary",
"spanoftext.",
"2"
] |
{
"arxiv_id": "2110.01529",
"language": "en",
"page_idx": 1
}
|
||
2110.01529-en-2
|
2110.01529-en-2.jpg
|
Let us further define a comparison function $\phi$ that takes these fixed-width vector representations and
computes a score. We have:
$$s(q,d)\stackrel{\Delta}{=}\phi(\eta_q(q),\eta_d(d))$$
(1)
We can interpret the score $s$ as quantifying the degree to which $d$ is relevant to query $q$, i.e., the basis for ranking a set of documents with respect to a query. For example, we desire to maximize scores for queries and their relevant documents and minimize scores for queries and non-relevant documents (note how this statement can be straightforwardly operationalized into a loss function). For dense retrieval methods, this design is commonly called a bi-encoder [Humeau et al., 2020].
More intuitively, we can understand the score $s$ as capturing the probability of relevance:
$$P(\text{Relevant}=1|d,q)\stackrel{\Delta}{=}s(q,d).$$
(2)
Note that the domain of $\eta_q$ comprises arbitrary sequences of terms, including sequences that have never been encountered before. In contrast, the domain of $\eta_d$ is typically $\mathcal{D}$, since we aretrieving from a given collection of documents (i.e., the corpus).
The logical scoring model, as defined in Eq.(1), nicely captures why I characterize this proposed conceptual framework as a“representational approach”, since it focuses on matching representations derived from queries (information needs) and documents (texts to be searched). In the context of bag-of-words representations, this formulation puts the vocabulary mismatch problem [Furnas et al., 1987]—overcoming the fact that information seekers and authors use different words to express the same concepts—front and center in the design of retrieval models. As I will discuss in detail later, neural models are simply the source of (better) representations—the structure of the $ad\textit{hoc retrieval}$ problem remains the same. In fact, across many diverse formulations of retrieval models, $\phi$ is defined as the inner product.
Physical Retrieval Model Given the setup above, top-$k$ retrieval can be defined as:
(3)
$$\underset{d\in\mathcal{D}}{\text{arg top-k}}\phi(\eta_q(q),\eta_d(d))$$
That is, given $q$, we wish to identify from $\mathcal{D}$ the $k$ documents $d_1\ldots d_k$ that have the highest scores $s_{1}\ldots s_{k}.$ These $\{(d_{i},s_{i})\}_{i=0}^{k}\tilde{\text{pairs are usually referred to as the ranked list of results (sometimes)}}$ called the “hits”).
If $s$ is interpreted as a probability of relevance, as per Eq.(2), then the physical retrieval model represents a direct realization of the $Probability\textit{Ranking Principle[ Robertson, 1977] , which states}$ that documents should be ranked in decreasing order of the estimated probability of relevance with respect to the query.
We might think of the logical scoring model and the physical retrieval model as providing what I argue to be the “right”abstractions for the text retrieval problem. So far, however, nothing in the presentation above captures information that isn't already common knowledge. I have simply adopted notation that may seem slightly peculiar, compared to how the text retrieval problem is usually presented (for example, in standard textbooks). Nevertheless, I will attempt to convince the reader that this isn't a pointless symbol manipulation exercise, but rather this framing of the problem provides a conceptual framework that bridges dense and sparse retrieval methods.
2.1 Applications to Dense and Sparse Retrieval
Let us consider DPR [Karpukhin et al., 2020], a popular representative dense retrieval model, and see how it can be understood within this conceptual framework. DPR uses separate transformerbased encoders for queries and documents, $\eta_q$ and $\eta_d$, respectively. Both encoders take the[CLS] representation from BERT[Devlin et al., 2019] as its output representation. In other words, the DPR encoders project queries and documents into fixed-width vector representations in some latent semantic space (by default, 768 dimensions).
Relevance between query representations and document representations—the comparison function
$\phi$— is defined in terms of inner products
$$\phi(\eta_q(q),\eta_d(d))=\eta_q(q)^\intercal\eta_d(d)$$
(4)
3
|
[
[
0.1748046875,
0.0921630859375,
0.82421875,
0.0921630859375,
0.82421875,
0.1060791015625,
0.1748046875,
0.1060791015625
],
[
0.1748046875,
0.106689453125,
0.359375,
0.106689453125,
0.359375,
0.12060546875,
0.1748046875,
0.12060546875
],
[
0.41259765625,
0.12249755859375,
0.583984375,
0.12249755859375,
0.583984375,
0.14013671875,
0.41259765625,
0.14013671875
],
[
0.80224609375,
0.1256103515625,
0.826171875,
0.1256103515625,
0.826171875,
0.1395263671875,
0.80224609375,
0.1395263671875
],
[
0.17236328125,
0.147705078125,
0.82421875,
0.1483154296875,
0.82421875,
0.1634521484375,
0.17236328125,
0.1622314453125
],
[
0.1748046875,
0.1634521484375,
0.82421875,
0.1634521484375,
0.82421875,
0.1767578125,
0.1748046875,
0.1767578125
],
[
0.1748046875,
0.1767578125,
0.82421875,
0.1767578125,
0.82421875,
0.1900634765625,
0.1748046875,
0.1900634765625
],
[
0.1732177734375,
0.1912841796875,
0.82421875,
0.1912841796875,
0.82421875,
0.2052001953125,
0.1732177734375,
0.2052001953125
],
[
0.1732177734375,
0.2052001953125,
0.73779296875,
0.2052001953125,
0.73779296875,
0.2183837890625,
0.1732177734375,
0.2183837890625
],
[
0.1748046875,
0.2247314453125,
0.75341796875,
0.2247314453125,
0.75341796875,
0.238037109375,
0.1748046875,
0.238037109375
],
[
0.390625,
0.24560546875,
0.60546875,
0.2474365234375,
0.60546875,
0.26513671875,
0.390625,
0.26318359375
],
[
0.8037109375,
0.250732421875,
0.82421875,
0.250732421875,
0.82421875,
0.264404296875,
0.8037109375,
0.264404296875
],
[
0.1732177734375,
0.274658203125,
0.82275390625,
0.27587890625,
0.82275390625,
0.290283203125,
0.1732177734375,
0.2890625
],
[
0.1732177734375,
0.2890625,
0.82275390625,
0.290283203125,
0.82275390625,
0.304931640625,
0.1732177734375,
0.3037109375
],
[
0.1732177734375,
0.304931640625,
0.5361328125,
0.304931640625,
0.5361328125,
0.318115234375,
0.1732177734375,
0.318115234375
],
[
0.1732177734375,
0.323974609375,
0.82421875,
0.324462890625,
0.82421875,
0.339111328125,
0.1732177734375,
0.33837890625
],
[
0.176513671875,
0.339111328125,
0.82275390625,
0.339111328125,
0.82275390625,
0.352783203125,
0.176513671875,
0.352783203125
],
[
0.1748046875,
0.352783203125,
0.82421875,
0.352783203125,
0.82421875,
0.3662109375,
0.1748046875,
0.3662109375
],
[
0.1748046875,
0.3662109375,
0.826171875,
0.3662109375,
0.826171875,
0.380126953125,
0.1748046875,
0.380126953125
],
[
0.1748046875,
0.380126953125,
0.82421875,
0.380126953125,
0.82421875,
0.393310546875,
0.1748046875,
0.393310546875
],
[
0.1732177734375,
0.393310546875,
0.826171875,
0.39453125,
0.826171875,
0.40771484375,
0.1732177734375,
0.4072265625
],
[
0.1748046875,
0.40771484375,
0.82275390625,
0.40771484375,
0.82275390625,
0.421630859375,
0.1748046875,
0.421630859375
],
[
0.1732177734375,
0.421630859375,
0.82421875,
0.421630859375,
0.82421875,
0.43505859375,
0.1732177734375,
0.43505859375
],
[
0.1732177734375,
0.43505859375,
0.3095703125,
0.43505859375,
0.3095703125,
0.448974609375,
0.1732177734375,
0.448974609375
],
[
0.1732177734375,
0.46337890625,
0.358642578125,
0.46337890625,
0.358642578125,
0.4765625,
0.1732177734375,
0.4765625
],
[
0.35302734375,
0.464111328125,
0.73193359375,
0.464111328125,
0.73193359375,
0.477783203125,
0.35302734375,
0.477783203125
],
[
0.4150390625,
0.48291015625,
0.5810546875,
0.484130859375,
0.5810546875,
0.498779296875,
0.4150390625,
0.49755859375
],
[
0.80224609375,
0.484130859375,
0.826171875,
0.484130859375,
0.826171875,
0.498779296875,
0.80224609375,
0.498779296875
],
[
0.435546875,
0.496826171875,
0.464111328125,
0.496826171875,
0.464111328125,
0.50830078125,
0.435546875,
0.50830078125
],
[
0.1748046875,
0.51220703125,
0.82275390625,
0.51220703125,
0.82275390625,
0.5263671875,
0.1748046875,
0.5263671875
],
[
0.1748046875,
0.52783203125,
0.82421875,
0.52783203125,
0.82421875,
0.54248046875,
0.1748046875,
0.54248046875
],
[
0.1732177734375,
0.54052734375,
0.291748046875,
0.54150390625,
0.291748046875,
0.55615234375,
0.1732177734375,
0.5546875
],
[
0.1732177734375,
0.56103515625,
0.82275390625,
0.56103515625,
0.82275390625,
0.57470703125,
0.1732177734375,
0.57470703125
],
[
0.1732177734375,
0.57470703125,
0.82177734375,
0.57470703125,
0.82177734375,
0.58837890625,
0.1732177734375,
0.58837890625
],
[
0.1732177734375,
0.5888671875,
0.82421875,
0.5888671875,
0.82421875,
0.60302734375,
0.1732177734375,
0.60302734375
],
[
0.1732177734375,
0.60400390625,
0.30810546875,
0.60400390625,
0.30810546875,
0.6171875,
0.1732177734375,
0.6171875
],
[
0.1748046875,
0.63623046875,
0.82177734375,
0.63623046875,
0.82177734375,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.1748046875,
0.65087890625,
0.82275390625,
0.65087890625,
0.82275390625,
0.6640625,
0.1748046875,
0.6640625
],
[
0.1748046875,
0.6640625,
0.82275390625,
0.6640625,
0.82275390625,
0.67822265625,
0.1748046875,
0.67822265625
],
[
0.176513671875,
0.6796875,
0.82177734375,
0.6796875,
0.82177734375,
0.68994140625,
0.176513671875,
0.68994140625
],
[
0.1748046875,
0.6923828125,
0.82275390625,
0.6923828125,
0.82275390625,
0.70556640625,
0.1748046875,
0.70556640625
],
[
0.1732177734375,
0.70556640625,
0.82275390625,
0.70556640625,
0.82275390625,
0.7197265625,
0.1732177734375,
0.7197265625
],
[
0.1748046875,
0.7197265625,
0.708984375,
0.7197265625,
0.708984375,
0.73291015625,
0.1748046875,
0.73291015625
],
[
0.1732177734375,
0.74853515625,
0.515625,
0.74853515625,
0.515625,
0.76318359375,
0.1732177734375,
0.76318359375
],
[
0.1748046875,
0.7734375,
0.82275390625,
0.7734375,
0.82275390625,
0.78857421875,
0.1748046875,
0.78857421875
],
[
0.1748046875,
0.787109375,
0.82275390625,
0.787109375,
0.82275390625,
0.80029296875,
0.1748046875,
0.80029296875
],
[
0.1732177734375,
0.80322265625,
0.8203125,
0.80322265625,
0.8203125,
0.81640625,
0.1732177734375,
0.81640625
],
[
0.1748046875,
0.81640625,
0.82275390625,
0.81640625,
0.82275390625,
0.830078125,
0.1748046875,
0.830078125
],
[
0.1732177734375,
0.8291015625,
0.82421875,
0.830078125,
0.82421875,
0.8447265625,
0.1732177734375,
0.84326171875
],
[
0.1748046875,
0.8447265625,
0.468017578125,
0.8447265625,
0.468017578125,
0.85791015625,
0.1748046875,
0.85791015625
],
[
0.1748046875,
0.86279296875,
0.82421875,
0.86279296875,
0.82421875,
0.87744140625,
0.1748046875,
0.87744140625
],
[
0.1748046875,
0.87890625,
0.443603515625,
0.87890625,
0.443603515625,
0.892578125,
0.1748046875,
0.892578125
],
[
0.39794921875,
0.8984375,
0.60107421875,
0.8984375,
0.60107421875,
0.91357421875,
0.39794921875,
0.91357421875
],
[
0.80126953125,
0.896484375,
0.82666015625,
0.896484375,
0.82666015625,
0.9140625,
0.80126953125,
0.9140625
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"Let us further define a comparison function β that takes these fixed-width vector representations and",
"computes a score. We have:",
"s(q, d) 二(ng(q), na(d))",
"(1)",
"We can interpret the score s as quantifying the degree to which d is relevant to query q,i.e., the basis",
"for ranking a set of documents with respect to a query. For example, we desire to maximize scores for",
"queries and their relevant documents and minimize scores for queries and non-relevant documents",
"(note how this statement can be straightforwardly operationalized into a loss function). For dense",
"retrieval methods, this design is commonly called a bi-encoder [Humeau et al., 2020].",
"More intuitively, we can understand the score s as capturing the probability of relevance",
"P(Relevant = 1|d, q) ≤ s(q, d),",
"(2)",
"Note that the domain of ng comprises arbitrary sequences of terms, including sequences that have",
"never been encountered before. In contrast, the domain of nd is typically D, since we are retrieving",
"from a given collection of documents (i.e., the corpus).",
"The logical scoring model, as defined in Eq. (1), nicely captures why I characterize this proposed",
"conceptual framework as a “representational approach, since it focuses on matching representations",
"derived from queries (information needs) and documents (texts to be searched). In the context of",
"bag-of-words representations, this formulation puts the vocabulary mismatch problem [Furnas et al.,",
"1987]overcoming the fact that information seekers and authors use different words to express the",
"same concepts-front and center in the design of retrieval models. As I will discuss in detail later,",
"neural models are simply the source of (better) representations-the structure of the ad hoc retrieval",
"problem remains the same. In fact, across many diverse formulations of retrieval models, is defined",
"as the inner product.",
"Physical Retrieval Model",
"Given the setup above, top-k retrieval can be defined as:",
"arg top-k p(ng(q), na(d))",
"(3)",
"deD",
"That is, given g, we wish to identify from D the k documents di .. . d, that have the highest scores",
"S1 - .. Sk. These ((ds, si))-o pairs are usually referred to as the ranked list of results (sometimes",
"called the “hits\").",
"If s is interpreted as a probability of relevance, as per Eq. (2), then the physical retrieval model",
"represents a direct realization of the Probability Ranking Principle [Robertson, 1977], which states",
"that documents should be ranked in decreasing order of the estimated probability of relevance with",
"respect to the query.",
"We might think of the logical scoring model and the physical retrieval model as providing wha",
"I argue to be the “right\" abstractions for the text retrieval problem. So far, however, nothing in",
"the presentation above captures information that isn't already common knowledge. I have simply",
"adoptednotationthatmayseemslightlypeculiar,comparedtohowthetextretrievalproblemis",
"usually presented (for example, in standard textbooks). Nevertheless, I will attempt to convince the",
"reader that this isn't a pointless symbol manipulation exercise, but rather this framing of the problem",
"provides a conceptual framework that bridges dense and sparse retrieval methods.",
"2.1 Applications to Dense and Sparse Retrieval",
"Let us consider DPR [Karpukhin et al., 2020], a popular representative dense retrieval model, and",
"see how it can be understood within this conceptual framework. DPR uses separate transformer",
"based encoders for queries and documents, ng and nd, respectively. Both encoders take the [cLs]",
"representation from BERT [Devlin et al., 20i9] as its output representation. In other words, the",
"DPR encoders project queries and documents into fixed-width vector representations in some latent",
"semantic space (by default, 768 dimensions)",
"Relevance between query representations and document representations-the comparison function",
"βis defined in terms of inner products:",
"p(ng(q), nd(d)) = ng(q)Tnd(d)",
"(4)",
"3"
] |
{
"arxiv_id": "2110.01529",
"language": "en",
"page_idx": 2
}
|
||
2110.01529-en-3
|
2110.01529-en-3.jpg
|
The model is trained as follows: let $\mathcal{R}=\{\langle q_i,d_i^+,d_{i,1}^-,d_{i,2}^-,\ldots d_{i,n}^-\rangle\}_{i=1}^m$ be the training set comprising $m$ instances. Each instance contains a query $q$, a relevant passage $d^+$, and $n$ non-relevant passages $d_1^- , d_2^- , . . . d_n^- . \textbf{DPR is trained with the following loss function: }$
(5)
$$\mathcal{L}(q,d^+,d_1^-,d_2^-,...d_n^-)=-\log\frac{\exp\left[\phi(\eta_q(q),\eta_d(d^+))\right]}{\exp\left[\phi(\eta_q(q),\eta_d(d^+))\right]+\sum_{j=1}^n\exp\left[\phi(\eta_q(q),\eta_d(d^-))\right]}.$$
Non-relevant passages for a query are selected via in-batch negative sampling [Henderson et al., 2017], from examples associated with other queries in the same training batch. However, this is a technical detail and other models select negative examples in different ways. For example, ANCE [Xiong et al, 2021] searches for“hard negatives”based on an earlier version of the document encoder itself.
I have just described DPR in terms of the proposed conceptual framework outlined above. Now let's try to recast BM25 [Robertson et al., 1995] in the same framework. In fact, the mapping is pretty straightforward: The query encoder $\eta_q$ and the document encoder $\eta_d$ both generate sparse bag-of-words vector representations of dimension $|V|$, where $V$ is the vocabulary of the corpus. For the output of the document encoder $\eta_d$, as with any bag-of-words representation, each dimension corresponds to a term in the vocabulary, and each term is assigned a weight according to the BM25 scoring function. The query encoder $\eta_q$ uses a multi-hot representation, with a weight of one if the term is present in the query, and zero otherwise.$^3$ The comparison function $\phi$ is, like DPR, defined in terms of the inner product.
Viewed in this manner, we can clearly see that BM25 and DPR have the same functional form, parameterized by $\eta_q,\eta_d$, and $\phi$, and in fact, $\phi$ is the inner product in both cases. Explained in terms of abstractions such as interfaces in programming languages, by analogy the logical scoring model defines the abstract methods $(\eta_q,\eta_d$, and $\phi)$ that specific retrieval models override with custom implementations, and here I have demonstrated that the abstraction covers both BM25 and DPR. This framework can be applied to the recent panoply of proposed dense retrieval methods in the literature, as well as nearly all families of bag-of-words exact-match models beyond BM25's probabilistic formulation, e.g, tf-idf, query likelihood, divergence from randomness, etc. This conceptual framework allows us to draw a direct connection between dense retrieval and sparse retrieval as parametric variations of the same underlying logical scoring model.
Finally, what about cross-encoders? Typical of this design is the monoBERT model [Nogueira and Cho, $\dot{2}019$, Lin et al., 2021b], where a query and a document are fed into a pretrained transformer as part of an input template, and the contextual representation of the[cLS] token is used for relevance classification. Here, we can say that the comparison function $\phi$ is defined in terms of the transformer, and thus cross-encoders are still captured by the logical scoring model defined in Eq. (1).
“Hiding” transformer inference in the comparison function $\phi$ might seem like a sleight of hand, but the PreTTR reranking model proposed by MacAvaney et al.[2020] connects a“full”crossencoder like monoBERT on the one hand to $\phi$-as-inner-product methods like DPR on the other hand. MacAvaney et al. began with the simple observation that query-document attention prevents document representations from being computed offline; recall that in $\tilde{\mathrm{DPR}},\eta_d(\cdot)$ does not depend on the query. Yet, it is precisely query-document attention that allows cross-encoders to obtain high levels of effectiveness. PreTTR was designed with this insight: What if we limited query-document attention to only the upper layers of the transformer? In such a design, document representations in the lower layers could be precomputed (and hence cached to accelerate inference). At one extreme end of the PreTTR design space, if all query-document attention is eliminated, then we have essentially“cleaved”monoBERT into two disconnected networks, and the result looks quite similar to DPR, where each of the disconnected networks serves as an encoder (and all document representations can be precomputed and indexed for low-latency retrieval). At the other extreme, if no query-document attention is eliminated, we have monoBERT. Thus, PreTTR provides the conceptual linkage that allows us to understand bi-encoders and cross-encoders as the two extreme cases of a single underlying design: it's all in the definition of the comparison function $\phi.$
$^3$This is a slight simplification; the original formulation of BM25[Robertson et al, 1995] included a query weighting component, but this term is usually omitted in modern implementations [Kamphuis et al., 2020].
4
|
[
[
0.1732177734375,
0.0889892578125,
0.82666015625,
0.091552734375,
0.82666015625,
0.10919189453125,
0.1732177734375,
0.106689453125
],
[
0.1748046875,
0.10919189453125,
0.82177734375,
0.10919189453125,
0.82177734375,
0.12249755859375,
0.1748046875,
0.12249755859375
],
[
0.1732177734375,
0.12371826171875,
0.64794921875,
0.12249755859375,
0.64794921875,
0.136962890625,
0.1732177734375,
0.1383056640625
],
[
0.515625,
0.1578369140625,
0.6689453125,
0.1578369140625,
0.6689453125,
0.17236328125,
0.515625,
0.17236328125
],
[
0.1953125,
0.1641845703125,
0.4150390625,
0.1654052734375,
0.4150390625,
0.18115234375,
0.1953125,
0.179931640625
],
[
0.404296875,
0.17236328125,
0.78125,
0.17236328125,
0.78125,
0.1932373046875,
0.404296875,
0.1932373046875
],
[
0.80224609375,
0.1673583984375,
0.82275390625,
0.1673583984375,
0.82275390625,
0.179931640625,
0.80224609375,
0.179931640625
],
[
0.1748046875,
0.2171630859375,
0.82275390625,
0.2171630859375,
0.82275390625,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.1748046875,
0.231689453125,
0.82275390625,
0.231689453125,
0.82275390625,
0.24560546875,
0.1748046875,
0.24560546875
],
[
0.1748046875,
0.24560546875,
0.826171875,
0.24560546875,
0.826171875,
0.2587890625,
0.1748046875,
0.2587890625
],
[
0.1748046875,
0.2587890625,
0.79345703125,
0.2587890625,
0.79345703125,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1732177734375,
0.27783203125,
0.82177734375,
0.27783203125,
0.82177734375,
0.291748046875,
0.1732177734375,
0.291748046875
],
[
0.1748046875,
0.29345703125,
0.82275390625,
0.29345703125,
0.82275390625,
0.306884765625,
0.1748046875,
0.306884765625
],
[
0.1748046875,
0.30810546875,
0.82275390625,
0.30810546875,
0.82275390625,
0.3212890625,
0.1748046875,
0.3212890625
],
[
0.1748046875,
0.3212890625,
0.826171875,
0.3212890625,
0.826171875,
0.335205078125,
0.1748046875,
0.335205078125
],
[
0.1748046875,
0.335205078125,
0.82275390625,
0.335205078125,
0.82275390625,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1748046875,
0.348388671875,
0.82275390625,
0.348388671875,
0.82275390625,
0.3623046875,
0.1748046875,
0.3623046875
],
[
0.1732177734375,
0.3623046875,
0.82275390625,
0.361083984375,
0.82275390625,
0.375732421875,
0.1732177734375,
0.376953125
],
[
0.1748046875,
0.37744140625,
0.82275390625,
0.37744140625,
0.82275390625,
0.392578125,
0.1748046875,
0.392578125
],
[
0.1748046875,
0.392578125,
0.348876953125,
0.392578125,
0.348876953125,
0.40283203125,
0.1748046875,
0.40283203125
],
[
0.1732177734375,
0.410400390625,
0.82421875,
0.410888671875,
0.82421875,
0.425537109375,
0.1732177734375,
0.4248046875
],
[
0.1732177734375,
0.425537109375,
0.82275390625,
0.4248046875,
0.82275390625,
0.439453125,
0.1732177734375,
0.43994140625
],
[
0.1748046875,
0.43994140625,
0.82275390625,
0.43994140625,
0.82275390625,
0.453857421875,
0.1748046875,
0.453857421875
],
[
0.1748046875,
0.453857421875,
0.82275390625,
0.453857421875,
0.82275390625,
0.46728515625,
0.1748046875,
0.46728515625
],
[
0.1748046875,
0.46728515625,
0.82421875,
0.46728515625,
0.82421875,
0.48095703125,
0.1748046875,
0.48095703125
],
[
0.176513671875,
0.48095703125,
0.82275390625,
0.48095703125,
0.82275390625,
0.494384765625,
0.176513671875,
0.494384765625
],
[
0.1732177734375,
0.494384765625,
0.82275390625,
0.49365234375,
0.82275390625,
0.5068359375,
0.1732177734375,
0.50830078125
],
[
0.1732177734375,
0.5087890625,
0.82275390625,
0.5087890625,
0.82275390625,
0.52294921875,
0.1732177734375,
0.52294921875
],
[
0.1748046875,
0.52294921875,
0.82421875,
0.52294921875,
0.82421875,
0.5361328125,
0.1748046875,
0.5361328125
],
[
0.1732177734375,
0.537109375,
0.69140625,
0.537109375,
0.69140625,
0.55029296875,
0.1732177734375,
0.55029296875
],
[
0.1748046875,
0.55615234375,
0.82421875,
0.55615234375,
0.82421875,
0.5693359375,
0.1748046875,
0.5693359375
],
[
0.1732177734375,
0.5693359375,
0.82421875,
0.57080078125,
0.82421875,
0.583984375,
0.1732177734375,
0.58349609375
],
[
0.1748046875,
0.583984375,
0.82177734375,
0.583984375,
0.82177734375,
0.59765625,
0.1748046875,
0.59765625
],
[
0.1748046875,
0.5986328125,
0.82666015625,
0.5986328125,
0.82666015625,
0.6123046875,
0.1748046875,
0.6123046875
],
[
0.1748046875,
0.61279296875,
0.7548828125,
0.61279296875,
0.7548828125,
0.626953125,
0.1748046875,
0.626953125
],
[
0.1732177734375,
0.6318359375,
0.826171875,
0.6318359375,
0.826171875,
0.64599609375,
0.1732177734375,
0.64599609375
],
[
0.1748046875,
0.64599609375,
0.826171875,
0.64599609375,
0.826171875,
0.6591796875,
0.1748046875,
0.6591796875
],
[
0.1748046875,
0.66015625,
0.82421875,
0.66015625,
0.82421875,
0.673828125,
0.1748046875,
0.673828125
],
[
0.1732177734375,
0.673828125,
0.82275390625,
0.673828125,
0.82275390625,
0.6875,
0.1732177734375,
0.6875
],
[
0.1748046875,
0.68798828125,
0.82421875,
0.68798828125,
0.82421875,
0.70263671875,
0.1748046875,
0.70263671875
],
[
0.1748046875,
0.7021484375,
0.82275390625,
0.7021484375,
0.82275390625,
0.71533203125,
0.1748046875,
0.71533203125
],
[
0.1732177734375,
0.71533203125,
0.82421875,
0.71533203125,
0.82421875,
0.72900390625,
0.1732177734375,
0.72900390625
],
[
0.176513671875,
0.73095703125,
0.82275390625,
0.73095703125,
0.82275390625,
0.7412109375,
0.176513671875,
0.7412109375
],
[
0.1748046875,
0.74365234375,
0.82421875,
0.74365234375,
0.82421875,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1748046875,
0.7568359375,
0.82275390625,
0.7568359375,
0.82275390625,
0.77099609375,
0.1748046875,
0.77099609375
],
[
0.176513671875,
0.77294921875,
0.82177734375,
0.77294921875,
0.82177734375,
0.78271484375,
0.176513671875,
0.78271484375
],
[
0.17236328125,
0.78271484375,
0.82177734375,
0.78271484375,
0.82177734375,
0.796875,
0.17236328125,
0.796875
],
[
0.1732177734375,
0.798828125,
0.82275390625,
0.79736328125,
0.82275390625,
0.81103515625,
0.1732177734375,
0.8125
],
[
0.1748046875,
0.8125,
0.82275390625,
0.81103515625,
0.82275390625,
0.82470703125,
0.1748046875,
0.82568359375
],
[
0.1748046875,
0.82568359375,
0.826171875,
0.82568359375,
0.826171875,
0.8388671875,
0.1748046875,
0.8388671875
],
[
0.1748046875,
0.84033203125,
0.69140625,
0.84033203125,
0.69140625,
0.85400390625,
0.1748046875,
0.85400390625
],
[
0.1748046875,
0.88330078125,
0.82275390625,
0.88525390625,
0.82275390625,
0.8994140625,
0.1748046875,
0.89794921875
],
[
0.178955078125,
0.8984375,
0.8095703125,
0.8984375,
0.8095703125,
0.912109375,
0.178955078125,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"The model is trained as follows: let R = {(qi, dt, di.1, dj.2, ... di.n))m, be the training set com-",
"prising m instances. Each instance contains a query q, a relevant passage d+, and n non-relevan",
"passages di,d2,..d.DPR is trained with the followingloss function:",
"exp [(ng(q), na(d+))]",
"C(q,d+,di,dz,..d)=-log",
"exp [0(ng(a), na(d+)] + ,-1exp [(ng(g), na(d, )]",
"(5)",
"Non-relevant passages for a query are selected via in-batch negative sampling [Henderson et al., 2017]",
"from examples associated with other queries in the same training batch. However, this is a technical",
"detail and other models select negative examples in different ways. For example, ANCE [Xiong et al.,",
"2021] searches for \"hard negatives\" based on an earlier version of the document encoder itseif.",
"I have just described DPR in terms of the proposed conceptual framework outlined above. Now",
"let's try to recast BM25 [Robertson et al., 1995] in the same framework. In fact, the mapping is",
"pretty straightforward: The query encoder ng and the document encoder nd both generate sparse",
"bag-of-words vector representations of dimension /Vl, where V is the vocabulary of the corpus. For",
"the output of the document encoder nd, as with any bag-of-words representation, each dimension",
"corresponds to a term in the vocabulary, and each term is assigned a weight according to the BM25",
"scoring function. The query encoder ng uses a multi-hot representation, with a weight of one if the",
"term is present in the query, and zero otherwise.3 The comparison function β is, like DPR, defined in",
"terms of the inner product.",
"Viewed in this manner, we can clearly see that BM25 and DPR have the same functional form.",
"parameterized by ng, nd, and p, and in fact, is the inner product in both cases. Explained in",
"terms of abstractions such as interfaces in programming languages, by analogy the logical scoring",
"model defines the abstract methods (ng, Nd, and ) that specific retrieval models override with",
"custom implementations, and here I have demonstrated that the abstraction covers both BM25 and",
"DPR. This framework can be applied to the recent panoply of proposed dense retrieval methods in",
"the literature, as well as nearly all families of bag-of-words exact-match models beyond BM25's",
"probabilistic formulation, e.g., tf-idf, query likelihood, divergence from randomness, etc. This",
"conceptual framework allows us to draw a direct connection between dense retrieval and sparse",
"retrieval as parametric variations of the same underlying logical scoring model.",
"Finally,what about cross-encoders?Typical of this design is the monoBERT model Nogueira and",
"Cho,2019,Lin et al.,2021b], where a query and a document are fed into a pretrained transformer as",
"part of an input template, and the contextual representation of the [cLs] token is used for relevance",
"classification. Here, we can say that the comparison function is defined in terms of the transformer,",
"and thus cross-encoders are still captured by the logical scoring model defined in Eq. (1).",
"“Hiding\"transformer inference in the comparison function might seem like a sleight of hand,",
"but thePreTTR reranking model proposed byMacAvaney et al.[2020] connects a“full\"cross-",
"encoder like monoBERT on the one hand to @-as-inner-product methods like DPR on the other",
"hand. MacAvaney et al. began with the simple observation that query-document attention prevents",
"document representations from being computed offline; recall that in DPR, na() does not depend on",
"the query. Yet, it is precisely query-document attention that allows cross-encoders to obtain high",
"levels of effectiveness. PreTTR was designed with this insight: What if we limited query-document",
"attentiontoonlytheupperlayers ofthetransformer?In suchadesign,documentrepresentations",
"in the lower layers could be precomputed (and hence cached to accelerate inference).At one",
"extreme end of the PreTTR design space, if all query-document attention is eliminated, then we",
"haveessentiallycleaved\"monoBERTintotwodisconnectednetworks,andtheresultlooksquite",
"similar to DPR, where each of the disconnected networks serves as an encoder (and all document",
"representations can beprecomputed and indexed for low-latency retrieval).At the other extreme,if no",
"query-document attention is eliminated, we have monoBERT. Thus, PreTTR provides the conceptual",
"linkage that allows us to understand bi-encoders and cross-encoders as the two extreme cases of a",
"single underlying design: it's all in the definition of the comparison function .",
"3This is a slight simplification; the original formulation of BM25 [Robertson et al.,1995] included a query",
"weighting component, but this term is usually omitted in modern implementations [Kamphuis et al., 2020].",
"4"
] |
{
"arxiv_id": "2110.01529",
"language": "en",
"page_idx": 3
}
|
||
2110.01529-en-4
|
2110.01529-en-4.jpg
|
$\boxed{\:\text{Dense}}$
$\underline{Sparse}$
$\overline{\mathrm{DPR,~ANCE}\:\text{DeepImpact, uniCOIL}}$
${\mathrm{Supervised}}$
BM25, tf-idf
$\dot{\mathbf{Unsupervised}}$ LSI, LDA
Table 1: A taxonomy of logical scoring models.
2.2 Generalization of Logical Scoring Models
Dense retrieval models such as DPR are often compared against sparse retrieval models such as BM25 in experimental evaluations, as Karpukhin et al.[2020] did in their paper. Not surprisingly, results show that dense retrieval models obtain higher effectiveness.
This, however, is not a fair comparison. Dense retrieval methods represent an instance of representational learning—the key here is $learning.$ The output of the encoders are learned representations that benefit from (large amounts of) training data under a standard supervised machine learning paradigm. In contrast, BM25 is unsupervised.$^4$ Comparing a supervised method to an unsupervised method is fundamentally an apples-to-oranges juxtaposition; it should not be surprising that a supervised technique is more effective.
As previously argued in Lin and Ma [2021], the encoders $\eta$. should be organized along two distinct dimensions or properties: The first dimension contrasts dense vs. sparse vector representations for queries and documents. The second dimension distinguishes between supervised (learned) and unsupervised representations. Table 1 illustrates this taxonomy. DPR (along with nearly all dense retrieval methods today) are instances of learned dense representations. BM25 is an instance of an unsupervised sparse representation.
This taxonomy immediately points to the existence of two other classes of logical scoring models. In fact, they correspond to models described in the literature that we can now categorize and unify in a single conceptual framework:
Learned sparse representations The existence of learned dense representations such as DPR and unsupervised sparse representations such as BM25 suggests that there should exist a class of learned sparse representations.
Learning sparse representations is by no means a new idea. If we fix the dimensions of the output representation to be the vocabulary (i.e., retaining a bag-of-words assumption), models for learned sparse representations become term weighting models—that is, a supervised machine learning approach to learning term weights. The earliest example I am aware of is Gordon [1988], who applied (what we might today call) representational learning on boolean vectors of descriptors using genetic algorithms, based on a small set of relevance judgments. These experiments might today be characterized as“toy”, but all the key elements of learned sparse retrieval models (quite amazingly!) are present. Another example along these lines is the work of Wilbur [2001], who attempted to learn global term weights using TREC data. A bit later, Trotman [2005] used genetic programming to discover better BM25-like scoring functions. Quite simply, there is plenty of evidence that learned sparse representations aren't new.
The fırst example of learned sparse representations in the “BERT era”is DeepCT [Dai and Callan, 2019], which uses a transformer to learn term weights based on a regression model, with the supervision signal coming from the MS MARCO passage ranking test collection. DeepCT has an interesting“quirk”: in truth, it only learns the term frequency (tf) component of term weights, but still relies on the remaining parts of the BM25 scoring function via the generation of pseudodocuments. The method also has a weakness: it only assigns weights to terms that are already present in the document, which limits retrieval to exact match. More generally, if we retain a bag-of-words assumption, term weighting models cannot address the vocabulary mismatch problem (more below). Note that dense representations do not have this issue since the dimensions of the vector representation capture some latent semantic space, not specific terms in the corpus vocabulary, and thus are able to capture what researchers call “semantic matching”.
The exact-match weakness of DeepCT discussed above was resolved by the DeepImpact model [Mallia et al., 2021], which brought together two key ideas: the use of document expansion to
$^{4}$Leaving aside simple tuning of parameters such as $k_1$ and $b.$
5
|
[
[
0.431396484375,
0.0889892578125,
0.473876953125,
0.0889892578125,
0.473876953125,
0.10101318359375,
0.431396484375,
0.10101318359375
],
[
0.52294921875,
0.08837890625,
0.57421875,
0.08837890625,
0.57421875,
0.1029052734375,
0.52294921875,
0.1029052734375
],
[
0.325927734375,
0.1029052734375,
0.398681640625,
0.10418701171875,
0.398681640625,
0.11553955078125,
0.325927734375,
0.1142578125
],
[
0.431396484375,
0.10418701171875,
0.5107421875,
0.10418701171875,
0.5107421875,
0.11553955078125,
0.431396484375,
0.11553955078125
],
[
0.52392578125,
0.1029052734375,
0.66650390625,
0.1029052734375,
0.66650390625,
0.11614990234375,
0.52392578125,
0.11614990234375
],
[
0.328369140625,
0.11614990234375,
0.4150390625,
0.11614990234375,
0.4150390625,
0.1275634765625,
0.328369140625,
0.1275634765625
],
[
0.428955078125,
0.11614990234375,
0.494384765625,
0.11614990234375,
0.494384765625,
0.1275634765625,
0.428955078125,
0.1275634765625
],
[
0.525390625,
0.1142578125,
0.609375,
0.1142578125,
0.609375,
0.1287841796875,
0.525390625,
0.1287841796875
],
[
0.340576171875,
0.1331787109375,
0.654296875,
0.1337890625,
0.654296875,
0.1483154296875,
0.340576171875,
0.147705078125
],
[
0.1748046875,
0.184326171875,
0.5009765625,
0.184326171875,
0.5009765625,
0.1976318359375,
0.1748046875,
0.1976318359375
],
[
0.1732177734375,
0.2083740234375,
0.82275390625,
0.208984375,
0.82275390625,
0.2235107421875,
0.1732177734375,
0.222900390625
],
[
0.1732177734375,
0.2235107421875,
0.82421875,
0.2235107421875,
0.82421875,
0.238037109375,
0.1732177734375,
0.238037109375
],
[
0.1732177734375,
0.238037109375,
0.61767578125,
0.238037109375,
0.61767578125,
0.251953125,
0.1732177734375,
0.251953125
],
[
0.1748046875,
0.25830078125,
0.826171875,
0.25830078125,
0.826171875,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1732177734375,
0.271484375,
0.82275390625,
0.271484375,
0.82275390625,
0.285400390625,
0.1732177734375,
0.285400390625
],
[
0.1748046875,
0.285888671875,
0.826171875,
0.285888671875,
0.826171875,
0.2998046875,
0.1748046875,
0.2998046875
],
[
0.1748046875,
0.2998046875,
0.82421875,
0.2998046875,
0.82421875,
0.313232421875,
0.1748046875,
0.313232421875
],
[
0.1732177734375,
0.313232421875,
0.82275390625,
0.313232421875,
0.82275390625,
0.326904296875,
0.1732177734375,
0.326904296875
],
[
0.1732177734375,
0.326904296875,
0.35546875,
0.32763671875,
0.35546875,
0.341552734375,
0.1732177734375,
0.34033203125
],
[
0.1748046875,
0.347900390625,
0.82421875,
0.347900390625,
0.82421875,
0.361083984375,
0.1748046875,
0.361083984375
],
[
0.1748046875,
0.3623046875,
0.82275390625,
0.3623046875,
0.82275390625,
0.375732421875,
0.1748046875,
0.375732421875
],
[
0.1732177734375,
0.375,
0.82275390625,
0.375,
0.82275390625,
0.38818359375,
0.1732177734375,
0.38818359375
],
[
0.1748046875,
0.389404296875,
0.82421875,
0.389404296875,
0.82421875,
0.40283203125,
0.1748046875,
0.40283203125
],
[
0.17236328125,
0.404052734375,
0.82275390625,
0.404052734375,
0.82275390625,
0.417236328125,
0.17236328125,
0.417236328125
],
[
0.1732177734375,
0.417236328125,
0.410888671875,
0.417236328125,
0.410888671875,
0.43115234375,
0.1732177734375,
0.43115234375
],
[
0.1748046875,
0.4375,
0.82275390625,
0.4375,
0.82275390625,
0.45068359375,
0.1748046875,
0.45068359375
],
[
0.1748046875,
0.451904296875,
0.82421875,
0.451904296875,
0.82421875,
0.46533203125,
0.1748046875,
0.46533203125
],
[
0.1732177734375,
0.46533203125,
0.371826171875,
0.46533203125,
0.371826171875,
0.479248046875,
0.1732177734375,
0.479248046875
],
[
0.1748046875,
0.494384765625,
0.82275390625,
0.494384765625,
0.82275390625,
0.50830078125,
0.1748046875,
0.50830078125
],
[
0.1748046875,
0.5087890625,
0.82275390625,
0.5087890625,
0.82275390625,
0.52294921875,
0.1748046875,
0.52294921875
],
[
0.1732177734375,
0.52294921875,
0.324462890625,
0.52294921875,
0.324462890625,
0.5361328125,
0.1732177734375,
0.5361328125
],
[
0.1732177734375,
0.54248046875,
0.82275390625,
0.54248046875,
0.82275390625,
0.556640625,
0.1732177734375,
0.556640625
],
[
0.1732177734375,
0.556640625,
0.82421875,
0.556640625,
0.82421875,
0.57080078125,
0.1732177734375,
0.57080078125
],
[
0.1748046875,
0.57080078125,
0.82275390625,
0.57080078125,
0.82275390625,
0.583984375,
0.1748046875,
0.583984375
],
[
0.17236328125,
0.583984375,
0.82421875,
0.58349609375,
0.82421875,
0.59765625,
0.17236328125,
0.5986328125
],
[
0.1732177734375,
0.59765625,
0.82275390625,
0.59765625,
0.82275390625,
0.6123046875,
0.1732177734375,
0.6123046875
],
[
0.1748046875,
0.6123046875,
0.82275390625,
0.6123046875,
0.82275390625,
0.62548828125,
0.1748046875,
0.62548828125
],
[
0.1732177734375,
0.62548828125,
0.82275390625,
0.62548828125,
0.82275390625,
0.6396484375,
0.1732177734375,
0.6396484375
],
[
0.1748046875,
0.6396484375,
0.82421875,
0.6396484375,
0.82421875,
0.65283203125,
0.1748046875,
0.65283203125
],
[
0.176513671875,
0.65478515625,
0.82177734375,
0.65478515625,
0.82177734375,
0.66552734375,
0.176513671875,
0.66552734375
],
[
0.176513671875,
0.66748046875,
0.82666015625,
0.66748046875,
0.82666015625,
0.68115234375,
0.176513671875,
0.68115234375
],
[
0.1748046875,
0.68115234375,
0.5361328125,
0.68115234375,
0.5361328125,
0.6943359375,
0.1748046875,
0.6943359375
],
[
0.1748046875,
0.70068359375,
0.826171875,
0.70068359375,
0.826171875,
0.71533203125,
0.1748046875,
0.71533203125
],
[
0.1732177734375,
0.7138671875,
0.82275390625,
0.7138671875,
0.82275390625,
0.72802734375,
0.1732177734375,
0.72802734375
],
[
0.176513671875,
0.73095703125,
0.82275390625,
0.73095703125,
0.82275390625,
0.7412109375,
0.176513671875,
0.7412109375
],
[
0.1748046875,
0.7421875,
0.826171875,
0.7421875,
0.826171875,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1732177734375,
0.7548828125,
0.826171875,
0.755859375,
0.826171875,
0.77099609375,
0.1732177734375,
0.76953125
],
[
0.1748046875,
0.77099609375,
0.82421875,
0.77099609375,
0.82421875,
0.7841796875,
0.1748046875,
0.7841796875
],
[
0.1732177734375,
0.7841796875,
0.82275390625,
0.7841796875,
0.82275390625,
0.79736328125,
0.1732177734375,
0.79736328125
],
[
0.176513671875,
0.7998046875,
0.82421875,
0.7998046875,
0.82421875,
0.81005859375,
0.176513671875,
0.81005859375
],
[
0.176513671875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.82373046875,
0.176513671875,
0.82373046875
],
[
0.1748046875,
0.82568359375,
0.826171875,
0.82568359375,
0.826171875,
0.8388671875,
0.1748046875,
0.8388671875
],
[
0.1732177734375,
0.83837890625,
0.50732421875,
0.8388671875,
0.50732421875,
0.85400390625,
0.1732177734375,
0.85302734375
],
[
0.1748046875,
0.859375,
0.82275390625,
0.859375,
0.82275390625,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.1748046875,
0.87353515625,
0.82421875,
0.87353515625,
0.82421875,
0.88720703125,
0.1748046875,
0.88720703125
],
[
0.1748046875,
0.896484375,
0.53515625,
0.89794921875,
0.53515625,
0.912109375,
0.1748046875,
0.9111328125
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Dense",
"Sparse",
"Supervised",
"DPR,ANCE",
"Deeplmpact, uniCOIL",
"Unsupervised",
"LSI, LDA",
"BM25, tf-idf",
"Table I: A taxonomy of logical scoring models.",
"2.2Generalization of Logical Scoring Models",
"Dense retrieval models such as DPR are often compared against sparse retrieval models such as",
"BM25 in experimental evaluations, as Karpukhin et al. [2020] did in their paper. Not surprisingly.",
"results show that dense retrieval models obtain higher effectiveness.",
"This, however, is not a fair comparison. Dense retrieval methods represent an instance of representa-",
"tional learning-the key here is learning. The output of the encoders are learned representations that",
"benefit from (large amounts of) training data under a standard supervised machine learning paradigm.",
"In contrast, BM25 is unsupervised.4 Comparing a supervised method to an unsupervised method",
"is fundamentally an apples-to-oranges juxtaposition; it should not be surprising that a supervised",
"technique is more effective.",
"As previously argued in Lin and Ma [2021]l, the encoders n. should be organized along two distinct",
"dimensions or properties: The first dimension contrasts dense vs. sparse vector representations for",
"queries and documents. The second dimension distinguishes between supervised (learned) and",
"unsupervised representations. Table 1 illustrates this taxonomy. DPR (along with nearly all dense",
"retrieval methods today) are instances of learned dense representations. BM25 is an instance of an",
"unsupervised sparse representation.",
"This taxonomy immediately points to the existence of two other classes of logical scoring models. In",
"fact, they correspond to models described in the literature that we can now categorize and unify in a",
"single conceptual framework:",
"Learned sparse representationsThe existence of learned dense representations such as DPR and",
"unsupervised sparse representations such as BM25 suggests that there should exist a class of learned",
"sparse representations.",
"Learning sparse representations is by no means a new idea. If we fix the dimensions of the output",
"representation to be the vocabulary (i.e., retaining a bag-of-words assumption), models for learned",
"sparse representations become term weighting modelsthat is, a supervised machine learning",
"approach to learning term weights. The earliest example I am aware of is Gordon [1988], who",
"applied (what we might today call) representational learning on boolean vectors of descriptors",
"using genetic algorithms, based on a small set of relevance judgments. These experiments might",
"today be characterized as “toy\", but all the key elements of learned sparse retrieval models (quite",
"amazingly!) are present. Another example along these lines is the work of Wilbur [2001], who",
"attemptedtolearnglobaltermweightsusingTRECdata.Abitlater,Trotman[2005lusedgenetic",
"programming to discover better BM25-like scoring functions. Quite simply, there is plenty of",
"evidence that learned sparse representations aren't new",
"The first example of learned sparse representations in the “BERT era\" is DeepCT [Dai and Callan,",
"2019], which uses a transformer to learn term weights based on a regression model, with the",
"supervisionsignalcomingfromtheMSMARCOpassagerankingtestcollection.DeepCThas",
"an interesting quirk\": in truth, it only learns the term frequency (tf) component of term weights,",
"but still relies on the remaining parts of the BM25 scoring function via the generation of pseudo-",
"documents. The method also has a weakness: it only assigns weights to terms that are already present",
"in the document, which limits retrieval to exact match. More generally, if we retain a bag-of-words",
"assumption,termweightingmodelscannotaddressthevocabularymismatchproblem(morebelow)",
"Notethatdenserepresentationsdonothavethisissuesincethedimensions ofthevectorrepresentation",
"capture some latent semantic space, not specific terms in the corpus vocabulary, and thus are able to",
"capture what researchers call“semantic matching\"",
"The exact-match weakness of DeepCT discussed above was resolved by the Deeplmpact model",
"[Mallia et al., 2021], which brought together two key ideas: the use of document expansion to",
"+Leaving aside simple tuning of parameters such as ki and b.",
"5"
] |
{
"arxiv_id": "2110.01529",
"language": "en",
"page_idx": 4
}
|
||
2309.05192-en-0
|
2309.05192-en-0.jpg
|
Towards Viewpoint Robustness in Bird's Eye View Segmentation
$ଟ୍ରି$
Zan Gojcic$^{2}$ Jungseock Joo$^{2,5}$ Ramesh Raskar$^{1}$ Sanja Fidler$^{2,3,4}$ Or Litany$^{2}$Zan Gojcic$^{2}$ Jungseock Joo$^{2,5}$ Ramesh Raskar$^{1}$ Sanja Fidler$^{2,3,4}$ Jose M. Alvarez$^{2,4}$IMIT $^{2\text{NVIDIA }}^{3}$University of Toronto$^{4}$
$Projectpage:$ https://nvlabs.github.io/viewpoint-robustness
## Abstract
<FigureHere>
Figure 1. Impact of Changed Camera Viewpoint: We find that
the performance of state-of-the-art methods for bird's eye view (BEV) segmentation quickly drop with small changes to viewpoint at inference. Above we see predictions from Cross View Transformers [29] trained on data from a source rig (top). The target rig pitch is reduced by 10° (bottom), leading a 17% drop in IoU.
$\begin{array}{ccc}\text{G}&\text{Autonomous vehicles (AV) require that neural networks}\\\text{G}&\text{used for perception be robust to different viewpoints if they}\\\text{G}&\text{are to be deployed across many types of vehicles without the}\end{array}$
repeated cost of data collection and labeling for each. AV companies typically focus on collecting data from diverse scenarios and locations, but not camera rig configurations, due to cost. As a result, only a small number of rig variations exist across most fleets. In this paper, we study how $AV$ perception models are affected by changes in camera viewpoint and propose a way to scale them across vehicle types without repeated data collection and labeling. Using bird's eye view (BEV) segmentation as a motivating task, we find through extensive experiments that existing perception models are surprisingly sensitive to changes in camera viewpoint. When trained with data from one camera rig, small changes to pitch, yaw, depth, or height of the camera at inference time lead to large drops in performance. We introduce a technique for novel view synthesis and use it to transform collected data to the viewpoint of target rigs, allowing us to train BEV segmentation models for diverse target rigs without any additional data collection or labeling cost. To analyze the impact of viewpoint changes, we leverage synthetic data to mitigate other gaps (content, ISP, etc). Our approach is then trained on real data and evaluated on synthetic data, enabling evaluation on diverse target rigs. We release all data for use in future work. Our method is able to recover an average of 14.7% of the IoU that is otherwise lost when deploying to new rigs.
to OOD data, such as adversarial corruptions [10], unseen weather conditions [15], and new geographic regions [6]. While each of these pose a significant challenge for safety critical applications, we focus on another distribution shift, which, thus far, has been understudied in the research literature- changes in camera viewpoint between train data and test data. Because camera viewpoint changes are realistic in AVs, we study their impact on AV perception tasks.
AVs use cameras around the ego-vehicle to perceive their surroundings. Using images from each camera, NNs detect and segment objects in the scene, such as vehicles, pedestrians, roads, and more. This information is used by trajectory planners to decide how the ego-vehicle navigates. Camera viewpoint for AVs may differ between train and test in several real-world scenarios. First, the camera viewpoint may change over time due to wear and tear or damage. Second, camera viewpoint may change due to installation variation. Third, and most relevant for our work, if a single NN is to be deployed across different types of vehicles, it must be able to generalize to the camera viewpoints of each car. Collecting and labeling train data for each target rig is not scalable and quickly becomes intractable for AV companies wishing to scale across many types of vehicles due to cost, thus motivating our work to transform collected data into the viewpoint of diverse target rigs to use for training.
The goal of this paper is to bring understanding and a
### $1. \textbf{Introduction}$
Neural networks (NNs) are becoming ubiquitous across domains. Safety critical applications, such as autonomous vehicles (AVs), rely on these NN to be robust to out of distribution (OOD) data. Yet, recent work has drawn attention to the susceptibility of NNs to failure when exposed
$^{*}$Work done during an internship at NVIDIA.
$^{\dagger}$Corresponding author: Jose M. Alvarez (josea@nvidia.com).
1
|
[
[
0.1641845703125,
0.1337890625,
0.80810546875,
0.1337890625,
0.80810546875,
0.1483154296875,
0.1641845703125,
0.1483154296875
],
[
0.19287109375,
0.1787109375,
0.68896484375,
0.1787109375,
0.68896484375,
0.199462890625,
0.19287109375,
0.199462890625
],
[
0.6748046875,
0.1817626953125,
0.775390625,
0.179931640625,
0.775390625,
0.1976318359375,
0.6748046875,
0.199462890625
],
[
0.134033203125,
0.1976318359375,
0.833984375,
0.1944580078125,
0.833984375,
0.216552734375,
0.134033203125,
0.2197265625
],
[
0.2459716796875,
0.216552734375,
0.72607421875,
0.216552734375,
0.72607421875,
0.2342529296875,
0.2459716796875,
0.2342529296875
],
[
0.291748046875,
0.2392578125,
0.6787109375,
0.240478515625,
0.6787109375,
0.255126953125,
0.291748046875,
0.25390625
],
[
0.0236968994140625,
0.26318359375,
0.05963134765625,
0.26318359375,
0.05963134765625,
0.326904296875,
0.0236968994140625,
0.326904296875
],
[
0.236083984375,
0.2890625,
0.314453125,
0.2890625,
0.314453125,
0.306884765625,
0.236083984375,
0.306884765625
],
[
0.6435546875,
0.292236328125,
0.7900390625,
0.292236328125,
0.7900390625,
0.302978515625,
0.6435546875,
0.302978515625
],
[
0.02777099609375,
0.3125,
0.063720703125,
0.31494140625,
0.056365966796875,
0.367431640625,
0.021240234375,
0.3642578125
],
[
0.09722900390625,
0.322509765625,
0.471435546875,
0.322509765625,
0.471435546875,
0.3359375,
0.09722900390625,
0.3359375
],
[
0.68115234375,
0.330810546875,
0.74267578125,
0.328857421875,
0.744140625,
0.348388671875,
0.68310546875,
0.35107421875
],
[
0.080078125,
0.33837890625,
0.4697265625,
0.33837890625,
0.4697265625,
0.3515625,
0.080078125,
0.3515625
],
[
0.5810546875,
0.343505859375,
0.61669921875,
0.343505859375,
0.61669921875,
0.347900390625,
0.5810546875,
0.347900390625
],
[
0.080078125,
0.352783203125,
0.468017578125,
0.3515625,
0.468017578125,
0.3662109375,
0.080078125,
0.367431640625
],
[
0.685546875,
0.35791015625,
0.73388671875,
0.35791015625,
0.73388671875,
0.37744140625,
0.685546875,
0.37744140625
],
[
0.0318603515625,
0.371826171875,
0.037567138671875,
0.371826171875,
0.037567138671875,
0.386962890625,
0.0318603515625,
0.386962890625
],
[
0.08087158203125,
0.36865234375,
0.468017578125,
0.36865234375,
0.468017578125,
0.3818359375,
0.08087158203125,
0.3818359375
],
[
0.08087158203125,
0.3837890625,
0.4697265625,
0.3837890625,
0.4697265625,
0.397705078125,
0.08087158203125,
0.397705078125
],
[
0.08087158203125,
0.40087890625,
0.46728515625,
0.40087890625,
0.46728515625,
0.410888671875,
0.08087158203125,
0.410888671875
],
[
0.5048828125,
0.40478515625,
0.890625,
0.40478515625,
0.890625,
0.415283203125,
0.5048828125,
0.415283203125
],
[
0.080078125,
0.4140625,
0.46728515625,
0.4140625,
0.46728515625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.50244140625,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.08087158203125,
0.4443359375,
0.4697265625,
0.4443359375,
0.4697265625,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.03350830078125,
0.45068359375,
0.0555419921875,
0.45068359375,
0.0555419921875,
0.479736328125,
0.03350830078125,
0.479736328125
],
[
0.50341796875,
0.4462890625,
0.8896484375,
0.4462890625,
0.8896484375,
0.45703125,
0.50341796875,
0.45703125
],
[
0.07843017578125,
0.458984375,
0.468017578125,
0.458251953125,
0.468017578125,
0.472900390625,
0.07843017578125,
0.473388671875
],
[
0.50244140625,
0.458984375,
0.89208984375,
0.458984375,
0.89208984375,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.080078125,
0.473388671875,
0.468017578125,
0.473388671875,
0.468017578125,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.50244140625,
0.472900390625,
0.86376953125,
0.472900390625,
0.86376953125,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.080078125,
0.4892578125,
0.468017578125,
0.4892578125,
0.468017578125,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.02777099609375,
0.50048828125,
0.058013916015625,
0.50244140625,
0.052276611328125,
0.55615234375,
0.0228729248046875,
0.5537109375
],
[
0.080078125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.50244140625,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.51953125,
0.4697265625,
0.51953125,
0.4697265625,
0.53271484375,
0.080078125,
0.53271484375
],
[
0.50341796875,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.50244140625,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.080078125,
0.5498046875,
0.4697265625,
0.5498046875,
0.4697265625,
0.56298828125,
0.080078125,
0.56298828125
],
[
0.50244140625,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.50244140625,
0.56640625
],
[
0.07843017578125,
0.564453125,
0.4697265625,
0.56298828125,
0.4697265625,
0.57763671875,
0.07843017578125,
0.5791015625
],
[
0.0269622802734375,
0.5546875,
0.0555419921875,
0.5546875,
0.0555419921875,
0.5869140625,
0.0269622802734375,
0.5869140625
],
[
0.5048828125,
0.5693359375,
0.8896484375,
0.5693359375,
0.8896484375,
0.580078125,
0.5048828125,
0.580078125
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.02777099609375,
0.59033203125,
0.053924560546875,
0.59033203125,
0.053924560546875,
0.62060546875,
0.02777099609375,
0.62060546875
],
[
0.037567138671875,
0.5859375,
0.049835205078125,
0.5859375,
0.049835205078125,
0.59228515625,
0.037567138671875,
0.59228515625
],
[
0.080078125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.080078125,
0.609375
],
[
0.50341796875,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.02777099609375,
0.61279296875,
0.053924560546875,
0.61279296875,
0.053924560546875,
0.6318359375,
0.02777099609375,
0.6318359375
],
[
0.08087158203125,
0.60986328125,
0.4697265625,
0.60986328125,
0.4697265625,
0.62353515625,
0.08087158203125,
0.62353515625
],
[
0.50244140625,
0.61279296875,
0.8408203125,
0.61279296875,
0.8408203125,
0.626953125,
0.50244140625,
0.626953125
],
[
0.080078125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.52294921875,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.52294921875,
0.642578125
],
[
0.03350830078125,
0.64013671875,
0.053924560546875,
0.64013671875,
0.053924560546875,
0.66552734375,
0.03350830078125,
0.66552734375
],
[
0.0294189453125,
0.6298828125,
0.052276611328125,
0.6298828125,
0.052276611328125,
0.64306640625,
0.0294189453125,
0.64306640625
],
[
0.080078125,
0.64013671875,
0.4697265625,
0.64013671875,
0.4697265625,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.5009765625,
0.6591796875,
0.8896484375,
0.6591796875,
0.8896484375,
0.67236328125,
0.5009765625,
0.67236328125
],
[
0.0400390625,
0.66845703125,
0.051483154296875,
0.66845703125,
0.051483154296875,
0.6796875,
0.0400390625,
0.6796875
],
[
0.041656494140625,
0.66552734375,
0.049835205078125,
0.66552734375,
0.049835205078125,
0.66943359375,
0.041656494140625,
0.66943359375
],
[
0.080078125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.0269622802734375,
0.67822265625,
0.052276611328125,
0.67529296875,
0.056365966796875,
0.70703125,
0.0310516357421875,
0.708984375
],
[
0.50244140625,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.69970703125,
0.46728515625,
0.69970703125,
0.46728515625,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72900390625,
0.080078125,
0.72900390625
],
[
0.50244140625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.07843017578125,
0.72998046875,
0.404296875,
0.73095703125,
0.404296875,
0.74560546875,
0.07843017578125,
0.744140625
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.08087158203125,
0.76513671875,
0.20751953125,
0.76513671875,
0.20751953125,
0.77880859375,
0.08087158203125,
0.77880859375
],
[
0.50244140625,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.50244140625,
0.77978515625,
0.8935546875,
0.77978515625,
0.8935546875,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.098876953125,
0.79052734375,
0.4697265625,
0.79052734375,
0.4697265625,
0.8037109375,
0.098876953125,
0.8037109375
],
[
0.50244140625,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.07843017578125,
0.8037109375,
0.4697265625,
0.8046875,
0.4697265625,
0.8193359375,
0.07843017578125,
0.818359375
],
[
0.50244140625,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.080078125,
0.82080078125,
0.4697265625,
0.82080078125,
0.4697265625,
0.833984375,
0.080078125,
0.833984375
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.080078125,
0.83544921875,
0.46728515625,
0.83544921875,
0.46728515625,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85107421875,
0.468017578125,
0.85107421875,
0.468017578125,
0.8642578125,
0.080078125,
0.8642578125
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50244140625,
0.869140625
],
[
0.096435546875,
0.875,
0.3349609375,
0.87353515625,
0.3349609375,
0.88818359375,
0.096435546875,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.85693359375,
0.8720703125,
0.85693359375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88720703125,
0.423095703125,
0.88720703125,
0.423095703125,
0.90087890625,
0.09722900390625,
0.90087890625
],
[
0.521484375,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.521484375,
0.8994140625
],
[
0.48193359375,
0.92236328125,
0.49267578125,
0.92822265625,
0.48779296875,
0.93603515625,
0.476318359375,
0.931640625
]
] |
[
"Towards Viewpoint Robustness in Bird's Eve View Segmentation",
"Tzofi Klinghofferl,2* Jonah Philion2.3.4Wenzheng Chen2.3.4",
"Or Litany?",
"Zan Gojcic2 Jungseock Joo2.5 Ramesh Raskarl Sanja Fidler2.3,4 Jose M. Alvarez2t",
"'MIT 2NVIDIA 3University of Toronto 4Vector Institute 5UCLA",
"Project page: https://nvlabs.github.io/viewpoint-robustness",
"",
"Abstract",
"BEVSegmentationInference",
"das",
"Autonomous vehicles (AV) require that neural networks",
"001",
"used for perception be robust to different viewpoints if they",
"burceRi",
"are to be deployed across many types of vehicles without the",
"一一",
"",
"repeated cost of data collection and labeling for each. Av",
"companies typically focus on collecting data from diverse",
"scenariosand locations,butnotcamerarig configurations,",
"Figure I.Impact of Changed Camera Viewpoint:Wefind that",
"due to cost. As a result, only a small number of rig varia-",
"LAD",
"the performance of state-of-the-art methods for bird's eye view",
"tions exist across most fleets.In this paper, we study how",
"(BEV) segmentation quickly drop with small changes to viewpoint",
"AV perception models are affected by changes in camera",
"S",
"atinference.AboveweseepredictionsfromCrossViewTrans-",
"viewpoint and propose a way to scale them across vehicle",
"formers [29] trained on data from a source rig (top). The target rig",
"types without repeated data collection and labeling. Using",
"pitch is reduced by 10° (bottom), leading a 17% drop in IoU.",
"bird's eye view (BEV) segmentation as a motivating task,",
"16",
"we find through extensive experiments that existing percep-",
"to OOD data, such as adversarial corruptions [10], unseen",
"tion models are surprisingly sensitive to changes in camera",
"weather conditions [15], and new geographic regions [6].",
"viewpoint. When trained with data from one camera rig,",
"While each of these pose a significant challenge for safety",
"small changes to pitch, yaw, depth, or height of the camera",
"critical applications, we focus on another distribution shift,",
"at inference time lead to large drops in performance. We",
"5",
"which, thus far, has been understudied in the research liter-",
"introduce a technique for novel view synthesis and use it",
"ature - changes in camera viewpoint between train data and",
"60",
"",
"to transform collected data to the viewpoint of target rigs,",
"testdata.Becausecameraviewpointchangesarerealisticin",
"3",
"allowing us to train BEV segmentation models for diverse",
"AVs, we study their impact on AV perception tasks.",
"target rigs without any additional data collection or label-",
"AVs use cameras around the ego-vehicle to perceive their",
".A",
"2",
"ing cost. To analyze the impact of viewpoint changes, we",
"surroundings. Using images from each camera, NNs detect",
"leverage synthetic data to mitigate other gaps (content, IsP,",
"and segment objects in the scene, such as vehicles, pedestri-",
"",
"",
"etc). Our approach is then trained on real data and evalu-",
"Ke",
"ans, roads, and more. This information is used by trajectory",
"ated on synthetic data, enabling evaluation on diverse tar-",
"planners to decide how the ego-vehicle navigates. Camera",
"get rigs. We release all data for use in future work. Our",
"viewpoint for AVs may differ between train and test in sev-",
"method is able to recover an average of 14.7% of the loU",
"eral real-world scenarios. First, the camera viewpoint may",
"that is otherwise lost when deploying to new rigs.",
"change over time due to wear and tear or damage. Second,",
"camera viewpoint may change due to installation variation.",
"1. Introduction",
"Third, and most relevant for our work, if a single NN is",
"to be deployed across different types of vehicles, it must",
"Neural networks (NNs) are becoming ubiquitous across",
"be able to generalize to the camera viewpoints of each car.",
"domains. Safety critical applications, such as autonomous",
"Collecting and labeling train data for each target rig is not",
"vehicles (AVs), rely on these NN to be robust to out of dis-",
"scalable and quickly becomes intractable for AV companies",
"tribution (OOD) data. Yet, recent work has drawn atten-",
"wishing to scale across many types of vehicles due to cost,",
"tion to the susceptibility of NNs to failure when exposed",
"thus motivating our work to transform collected data into",
"*Work done during an internship at NVIDIA",
"the viewpoint of diverse target rigs to use for training.",
"'Corresponding author: Jose M. Alvarez (josea@nvidia.com).",
"The goal of this paper is to bring understanding and a",
"1"
] |
{
"arxiv_id": "2309.05192",
"language": "en",
"page_idx": 0
}
|
||
2309.05192-en-1
|
2309.05192-en-1.jpg
|
first approach to a real-world problem in the AV space that has yet to receive attention in the research literature-generalization from a source to target camera rig. We focus on bird's eye view (BEV) segmentation from RGB data to motivate how changing camera viewpoint can affect AV perception models. We study this problem by conducting an in-depth analysis on the impact changing the camera viewpoint at inference time has on recent BEV segmentation models. Our findings indicate that even small changes in camera placement at inference time degrade BEV segmentation accuracy, as illustrated in Fig. $\tilde{1}.$ We then propose a method to improve generalization to a target rig by simulating views in the target perspective. We show that incorporating data generated from novel view synthesis into $\hat{\text{training can significantly reduce the viewpoint domain gap,}}$ bringing the BEV segmentation model to the same level of accuracy as when there is no change in camera viewpoint, without having to collect or label any additional data. We compare our approach with other strategies, such as augmenting the camera extrinsics and labels during training, and find that our approach leads to better accuracy. Little work has focused on the impact of viewpoint changes for AV perception, and, to the best of our knowledge, we are the first to study the impact of diverse camera viewpoint changes on 3D AV perception tasks, such as BEV segmentation. We hope that this paper will encourage more research on the important problem of viewpoint robustness in AV.
Our paper makes the following contributions:
$\bullet$ We highlight the understudied problem of viewpoint
robustness in bird's eye view segmentation for autonomous vehicles (AV) through an in-depth analysis revealing that recent models fail to generalize to different camera viewpoints at inference time.
$\cdot$ We propose a viewpoint augmentation framework for AV; we develop a novel view synthesis method that can be used to transform training data to target viewpoints and show that it improves the robustness of bird's eye view segmentation models to viewpoint changes.
$\cdot$ We provide datasets that can be used to benchmark fu-
ture work on viewpoint robustness in AV.
Because real-world AV datasets from a diverse set of camera rigs are not publicly available, we use simulated data both for (1) training and evaluation in our analysis and (2) evaluation of our proposed technique. Our synthetic datasets can be used for future efforts to benchmark the generalization abilities of different AV perception methods to viewpoint changes. Datasets are publicly available on our project page. The first dataset, rendered from CARLA, consists of both training and testing data, allowing for isolated analysis of the impact of viewpoint changes on BEV segmentation models (example images in Fig. 2). The second
dataset, rendered with NVIDIA DRIVE Sim [20], is significantly more photorealistic and consists of test sets from a diverse set of camera viewpoints. Thus, it can be used to evaluate models trained on real data, as we show in Section 5. Both datasets include 3D bounding box labels.
## $2. \textbf{ Related Work}$
## $2. 1. \textbf{ Viewpoint Robustness}$
Recent work has drawn attention to the susceptibility of NNs to misclassify when presented with distributions not seen during training. Madan et al.[14] show that both convolutional- and transformer-based classiffers are fooled by small viewpoint changes, and they introduce a search strategy for finding adversarial viewpoints, which leads to misclassifications over 71% of the time. Similarly, [13] shows that small viewpoint changes degrade classification performance, especially when paired with out of distribution (OOD) categories, and demonstrates that increasing the diversity of training data is an effective strategy to mitigate this issue. Do et al.[4] use homography to move images closer to the distribution of training data at inference time. Coors et al. [3] study the impact of viewpoint changes for 2D semantic segmentation for AV, but do not explore 3D tasks. In contrast, we focus on providing a thorough analysis and a solution to the problem of viewpoint robustness for 3D AV perception tasks, focusing on BEV segmentation.
## $2. 2. \textbf{Novel View Synthesis}$
Novel view synthesis (NVS) provides a way to render images from unseen viewpoints of a scene, and thus could be used to improve the robustness of perception models to viewpoint changes. Many methods have been proposed for NVS in recent years [18, 16, 19], many of which are based on Neural Radiance Fields (NeRF)[17]. However, NeRF still faces two challenges that limit its applicability for our use case: (1) getting NeRF to generalize to dynamic scenes, which are common in AV, is an open research problem, and while there is promising work in this direction [28,22], the setup is often too constrained and simplified to fit the AV problem setting, and (2) NeRF is challenging to scale due to lack of generalizability, so multiple NeRFs must be trained to perform NVS across scenes. $\hat{\text{While there is work aimed}}$ at generalizing NeRF[12], it remains an open problem and current methods are often constrained. Other methods for NVS rely on monocular depth estimation and can generalize across scenes when the depth estimation network is trained on diverse data. We leverage Worldsheet in our work[11], which is described in more detail in Sec. 4.1.
## $2. 3. \textbf{ Bird' s Eye View Segmentation}$
Bird's eye view (BEV) segmentation -the task of segmenting a scene in the top-down view (BEV) from 2D im-
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,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
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,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.5009765625,
0.1514892578125,
0.82666015625,
0.15283203125,
0.82666015625,
0.1673583984375,
0.5009765625,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.080078125,
0.1817626953125,
0.46728515625,
0.1817626953125,
0.46728515625,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.50341796875,
0.1817626953125,
0.640625,
0.1817626953125,
0.640625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.07843017578125,
0.1988525390625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.50341796875,
0.20703125,
0.7060546875,
0.20703125,
0.7060546875,
0.2215576171875,
0.50341796875,
0.2215576171875
],
[
0.080078125,
0.21337890625,
0.471435546875,
0.21337890625,
0.471435546875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.4697265625,
0.2279052734375,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.521484375,
0.2298583984375,
0.8935546875,
0.2298583984375,
0.8935546875,
0.24365234375,
0.521484375,
0.24365234375
],
[
0.080078125,
0.24365234375,
0.471435546875,
0.24365234375,
0.471435546875,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24560546875,
0.89208984375,
0.24560546875,
0.89208984375,
0.2587890625,
0.50244140625,
0.2587890625
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.291748046875,
0.89208984375,
0.291748046875,
0.89208984375,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.07843017578125,
0.302978515625,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50341796875,
0.30810546875,
0.890625,
0.30810546875,
0.890625,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.08087158203125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50341796875,
0.322509765625,
0.890625,
0.322509765625,
0.890625,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.3359375,
0.89208984375,
0.3359375,
0.89208984375,
0.349853515625,
0.50341796875,
0.349853515625
],
[
0.08087158203125,
0.35107421875,
0.468017578125,
0.35107421875,
0.468017578125,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50341796875,
0.365478515625
],
[
0.07843017578125,
0.363037109375,
0.468017578125,
0.3642578125,
0.468017578125,
0.37890625,
0.07843017578125,
0.37744140625
],
[
0.50244140625,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.50244140625,
0.380126953125
],
[
0.07843017578125,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.07843017578125,
0.392578125
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.080078125,
0.393310546875,
0.46728515625,
0.393310546875,
0.46728515625,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.5009765625,
0.395751953125,
0.89208984375,
0.396484375,
0.89208984375,
0.410888671875,
0.5009765625,
0.410400390625
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.412353515625,
0.8896484375,
0.412353515625,
0.8896484375,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4267578125,
0.89208984375,
0.4267578125,
0.89208984375,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50341796875,
0.458251953125,
0.890625,
0.458251953125,
0.890625,
0.4716796875,
0.50341796875,
0.4716796875
],
[
0.080078125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.080078125,
0.48486328125,
0.44775390625,
0.48486328125,
0.44775390625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.5009765625,
0.486083984375,
0.8759765625,
0.4873046875,
0.8759765625,
0.501953125,
0.5009765625,
0.50048828125
],
[
0.08087158203125,
0.5068359375,
0.378173828125,
0.5068359375,
0.378173828125,
0.521484375,
0.08087158203125,
0.521484375
],
[
0.50244140625,
0.51025390625,
0.701171875,
0.51123046875,
0.701171875,
0.52587890625,
0.50244140625,
0.5244140625
],
[
0.09722900390625,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.09722900390625,
0.54345703125
],
[
0.521484375,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.5498046875,
0.521484375,
0.5498046875
],
[
0.1094970703125,
0.54443359375,
0.46728515625,
0.54443359375,
0.46728515625,
0.55810546875,
0.1094970703125,
0.55810546875
],
[
0.50244140625,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.111083984375,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.111083984375,
0.57373046875
],
[
0.50244140625,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.50244140625,
0.580078125
],
[
0.111083984375,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.111083984375,
0.5888671875
],
[
0.50244140625,
0.58056640625,
0.8896484375,
0.58056640625,
0.8896484375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.111083984375,
0.59130859375,
0.378173828125,
0.59130859375,
0.378173828125,
0.60498046875,
0.111083984375,
0.60498046875
],
[
0.5009765625,
0.5947265625,
0.89208984375,
0.59521484375,
0.89208984375,
0.60986328125,
0.5009765625,
0.609375
],
[
0.09722900390625,
0.6123046875,
0.4697265625,
0.61279296875,
0.4697265625,
0.62744140625,
0.09722900390625,
0.626953125
],
[
0.50244140625,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.50244140625,
0.625
],
[
0.11358642578125,
0.6298828125,
0.46728515625,
0.6298828125,
0.46728515625,
0.64013671875,
0.11358642578125,
0.64013671875
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.111083984375,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.111083984375,
0.65771484375
],
[
0.5009765625,
0.6416015625,
0.890625,
0.64013671875,
0.890625,
0.65478515625,
0.5009765625,
0.65576171875
],
[
0.111083984375,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.111083984375,
0.673828125
],
[
0.50341796875,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50341796875,
0.67041015625
],
[
0.11273193359375,
0.6748046875,
0.43310546875,
0.6748046875,
0.43310546875,
0.68798828125,
0.11273193359375,
0.68798828125
],
[
0.50341796875,
0.671875,
0.890625,
0.671875,
0.890625,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.50244140625,
0.68603515625,
0.8896484375,
0.68603515625,
0.8896484375,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.098876953125,
0.6962890625,
0.468017578125,
0.6962890625,
0.468017578125,
0.7109375,
0.098876953125,
0.7109375
],
[
0.5048828125,
0.7041015625,
0.8896484375,
0.7041015625,
0.8896484375,
0.7138671875,
0.5048828125,
0.7138671875
],
[
0.111083984375,
0.71337890625,
0.384033203125,
0.71337890625,
0.384033203125,
0.7265625,
0.111083984375,
0.7265625
],
[
0.50341796875,
0.71728515625,
0.890625,
0.71728515625,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.50244140625,
0.7470703125
],
[
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,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.7646484375,
0.890625,
0.7646484375,
0.890625,
0.7744140625,
0.50341796875,
0.7744140625
],
[
0.08251953125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.50244140625,
0.77783203125,
0.89208984375,
0.77783203125,
0.89208984375,
0.79150390625,
0.50244140625,
0.79150390625
],
[
0.07843017578125,
0.79541015625,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.07843017578125,
0.8095703125
],
[
0.50244140625,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50244140625,
0.806640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.89208984375,
0.80810546875,
0.89208984375,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.8212890625,
0.7998046875,
0.82275390625,
0.7998046875,
0.8369140625,
0.5009765625,
0.8359375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84765625,
0.7685546875,
0.84765625,
0.7685546875,
0.8623046875,
0.50341796875,
0.8623046875
],
[
0.07843017578125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.0767822265625,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.0767822265625,
0.88525390625
],
[
0.51953125,
0.869140625,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.51953125,
0.8837890625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"first approach to a real-world problem in the AV space that",
"dataset, rendered with NVIDIA DRIVE Sim [20], is signif-",
"has yet to receive attention in the research literature - gen-",
"icantly more photorealistic and consists of test sets from a",
"eralization from a source to target camera rig. We focus on",
"diverse set of camera viewpoints. Thus, it can be used to",
"bird's eye view (BEV) segmentation from RGB data to mo-",
"evaluate models trained on real data, as we show in Section",
"tivatehow changing camera viewpoint canaffect AV per-",
"5. Both datasets include 3D bounding box labels.",
"ception models. We study this problem by conducting an",
"in-depth analysis on the impact changing the camera view-",
"2.Related Work",
"point at inference time has on recent BEV segmentation",
"2.1. Viewpoint Robustness",
"models. Our findings indicate that even small changes in",
"camera placement at inference time degrade BEV segmen-",
"Recent work has drawn attention to the susceptibility of",
"tation accuracy,as illustrated in Fig.1.Wethen propose",
"NNs to misclassify when presented with distributions not",
"a method to improve generalization to a target rig by sim-",
"seen during training. Madan et al. [14] show that both",
"ulating views in the target perspective.We show that in-",
"convolutional- and transformer-based classifiers are fooled",
"corporating data generated from novel view synthesis into",
"by small viewpoint changes, and they introduce a search",
"training can significantly reduce the viewpoint domain gap,",
"strategy for finding adversarial viewpoints, whichleads to",
"bringingtheBEVsegmentationmodeltothesamelevelof",
"misclassifications over71%of the time.Similarly[13]",
"accuracy as when there is no change in camera viewpoint,",
"shows that small viewpoint changes degrade classification",
"withouthavingto collect orlabel any additional data.We",
"performance, especially when paired with out of distribu-",
"compare our approach with other strategies, such as aug-",
"tion (OOD) categories, and demonstrates that increasing the",
"menting the camera extrinsics and labels during training,",
"diversity of training data is an effective strategy to mitigate",
"and find that our approach leads to better accuracy. Lit-",
"this issue. Do et al. [4] use homography to move images",
"tle work has focused on the impact of viewpoint changes",
"closer to the distribution of training data at inference time",
"for AV perception, and, to the best of our knowledge, we",
"Coors et al. [3] study the impact of viewpoint changes for",
"are the first to study the impact of diverse camera viewpoint",
"2D semantic segmentation for AV, but do not explore 3D",
"changes on 3D AV perception tasks, such as BEV segmenta-",
"tasks. In contrast, we focus on providing a thorough analy-",
"tion. We hope that this paper will encourage more research",
"sis and a solution to the problem of viewpoint robustness for",
"on the important problem of viewpoint robustness in AV",
"3D AV perception tasks, focusing on BEV segmentation.",
"Our paper makes the following contributions:",
"2.2.Novel View Synthesis",
". We highlight the understudied problem of viewpoint",
"Novel view synthesis (NVS) provides a way to render",
"robustness in bird's eye view segmentation for au-",
"images from unseen viewpoints of a scene, and thus could",
"tonomous vehicles (AV) through an in-depth analysis",
"be used to improve the robustness of perception models to",
"revealing that recent models fail to generalize to differ-",
"viewpoint changes.Many methods have been proposed for",
"ent camera viewpoints at inference time.",
"NvS in recent years [18, 16, 19], many of which are based",
". We propose a viewpoint augmentation framework for",
"on Neural Radiance Fields (NeRF) [17]. However, NeRF",
"AV:wedevelopanovel viewsynthesismethodthatcan",
"still faces two challenges that limit its applicability for our",
"be used to transform training data to target viewpoints",
"use case: (1) getting NeRF to generalize to dynamic scenes,",
"and show that it improves the robustness of bird's eye",
"which are common in Av, is an open research problem, and",
"view segmentation models to viewpoint changes.",
"while there is promising work in this direction [28, 22], the",
"setup is often too constrained and simplified to fit the AV",
". We provide datasets that can be used to benchmark fu-",
"problemsetting,and(2)NeRFischallengingtoscaledueto",
"ture work on viewpoint robustness in AV.",
"lack of generalizability, so multiple NeRFs must be trained",
"Because real-world AV datasets from a diverse set of cam-",
"to perform NvS across scenes. While there is work aimed",
"era rigs are not publicly available, we use simulated data",
"at generalizing NeRF [12], it remains an open problem and",
"both for (1) training and evaluation in our analysis and",
"currentmethods are often constrained.Othermethodsfor",
"(2)evaluation of our proposedtechnique.Our synthetic",
"NVS rely on monocular depth estimation and can generalize",
"datasets can be used for future efforts to benchmark the gen-",
"across scenes when the depth estimation network is trained",
"eralization abilities of different AV perception methods to",
"on diverse data. We leverage Worldsheet in our work [11],",
"viewpoint changes. Datasets are publicly available on our",
"which is described in more detail in Sec. 4.1.",
"project page. The first dataset, rendered from CARLA, con-",
"2.3. Bird's Eye View Segmentation",
" sists of both training and testing data, allowing for isolated",
"analysis of the impact of viewpoint changes on BEV seg-",
"Bird's eye view (BEV) segmentation -the task of seg-",
"mentation models (example images in Fig. 2). The second",
"menting a scene in the top-down view (BEV) from 2D im-",
"2"
] |
{
"arxiv_id": "2309.05192",
"language": "en",
"page_idx": 1
}
|
||
2309.05192-en-2
|
2309.05192-en-2.jpg
|
<FigureHere>
Figure 2. Datasets rendered in CARLA across viewpoints: For the analysis part of our work, we use CARLA to simulate different
viewpoints. We rendered datasets from a total of 36 viewpoints, a few of which are highlighted above, including the source rig (extrinsics
from nuScenes [2] dataset), +12° yaw, +12° pitch, +21 inch height, and -12° pitch and +18 inch height together
<FigureHere>
Figure 3. Analysis of impact of viewpoint changes in CARLA: We train a source BEV model using Lift Splat Shoot (LSS)[21] and
Cross View Transformers (CVT) [29], denoted at point 0 on the $x$ axis of each graph. We then test the model across different target rigs where the camera pitch, yaw, height, or pitch and height are changed, as denoted by the different points along the $x$ axes. We also trained each model on the target rig directly and refer to this model as the "oracle", as it reflects the expected upper bound IoU for each viewpoint.
ages-is a useful task for benchmarking AV perception [21,24,1]. BEV segmentation requires a 2D to 3D unprojection to predict the position of objects surrounding the ego-vehicle from the BEV perspective. BEV segmentation models usually consist of an image encoder, which extracts the features from images from the camera rig, and a decoder, which uses the image features to predict the objects of interest in the BEV coordinate frame. Existing methods condition on the extrinsics and intrinsics of each camera in different ways. Lift-Splat-Shoot (LSS)[21] and Orthographic Feature Transform (OFT)[25] unproject features into a point cloud according to each camera's intrinsic and extrinsic parameters. LSS performs sum pooling along each pillar in the map-view, while OFT performs average pooling. Other methods, such as Cross $\hat{\text{View Transform-}}$ ers (CVT) [29], treat camera intrinsics and extrinsics as a feature, rather than explicitly unprojecting. We use LSS and CVT to conduct benchmarks, since these two methods encompass both convolutional and transformer-based architectures and explicit and implicit geometric representations.
viewpoint at inference time for BEV segmentation models trained on a single, source rig. We use simulated data from CARLA[5] for this analysis for two reasons: (1) using simulated data allows us to isolate the domain gaps between training and testing such that only camera viewpoint changes, and (2) real AV datasets with large differences in camera position are not publicly available. Examples of different camera viewpoints rendered in CARLA are shown in Fig. 2. For simplicity and ease of interpretation of our results, we conduct all experiments on a single camera rig, containing a front facing camera, which we refer to as the source rig. We first train a BEV segmentation model on data rendered from the source rig. For this rig, we use the camera parameters of sessions from the nuScenes dataset[2]. Then, we render train and test datasets from different target rigs, which contain variations to the yaw, pitch, height, or pitch and height of the camera. The train datasets are used to train an oracle for each target rig, while the test datasets are used to evaluate the model trained on the source rig in comparison to the oracle. For completeness, we sweep over a large range of each extrinsic and render a train and test dataset on regular intervals. For pitch and yaw, we sweep from -20° to 20°, rendering a dataset every $4^\circ.$ For height, we sweep from 0 in to 30 in, rendering a dataset every 3 in. For height and pitch together, we sweep from $0^{\circ}$ and 0 in to
## $3. \textbf{Measuring the Impact of Camera Viewpoint}$ Variations on BEV Segmentation
Method: In this section, we introduce our approach and results for measuring the impact of changing the camera
3
|
[
[
0.08087158203125,
0.1666259765625,
0.89208984375,
0.1666259765625,
0.89208984375,
0.1767578125,
0.08087158203125,
0.1767578125
],
[
0.080078125,
0.179931640625,
0.8935546875,
0.179931640625,
0.8935546875,
0.1932373046875,
0.080078125,
0.1932373046875
],
[
0.080078125,
0.1925048828125,
0.74169921875,
0.1925048828125,
0.74169921875,
0.205810546875,
0.080078125,
0.205810546875
],
[
0.1192626953125,
0.2442626953125,
0.143798828125,
0.2442626953125,
0.143798828125,
0.2998046875,
0.1192626953125,
0.2998046875
],
[
0.1422119140625,
0.261474609375,
0.156005859375,
0.261474609375,
0.156005859375,
0.276611328125,
0.1422119140625,
0.276611328125
],
[
0.1446533203125,
0.272705078125,
0.1544189453125,
0.272705078125,
0.1544189453125,
0.285400390625,
0.1446533203125,
0.285400390625
],
[
0.08087158203125,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.45263671875,
0.08087158203125,
0.45263671875
],
[
0.07843017578125,
0.451904296875,
0.89208984375,
0.45263671875,
0.89208984375,
0.466552734375,
0.07843017578125,
0.46533203125
],
[
0.080078125,
0.466552734375,
0.89208984375,
0.466552734375,
0.89208984375,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.07843017578125,
0.48095703125,
0.89208984375,
0.48095703125,
0.89208984375,
0.494384765625,
0.07843017578125,
0.494384765625
],
[
0.1298828125,
0.51123046875,
0.47216796875,
0.51220703125,
0.47216796875,
0.5263671875,
0.1298828125,
0.52587890625
],
[
0.50244140625,
0.51220703125,
0.89208984375,
0.51220703125,
0.89208984375,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.080078125,
0.52783203125,
0.4697265625,
0.52783203125,
0.4697265625,
0.54150390625,
0.080078125,
0.54150390625
],
[
0.08087158203125,
0.5146484375,
0.12091064453125,
0.5146484375,
0.12091064453125,
0.52587890625,
0.08087158203125,
0.52587890625
],
[
0.50244140625,
0.52783203125,
0.89208984375,
0.52783203125,
0.89208984375,
0.54150390625,
0.50244140625,
0.54150390625
],
[
0.080078125,
0.54345703125,
0.4697265625,
0.54345703125,
0.4697265625,
0.556640625,
0.080078125,
0.556640625
],
[
0.50244140625,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.5712890625,
0.080078125,
0.5712890625
],
[
0.50244140625,
0.55810546875,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.080078125,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.5869140625,
0.080078125,
0.5869140625
],
[
0.50244140625,
0.57373046875,
0.89208984375,
0.57373046875,
0.89208984375,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.080078125,
0.5888671875,
0.4697265625,
0.5888671875,
0.4697265625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.5888671875,
0.8935546875,
0.5888671875,
0.8935546875,
0.60302734375,
0.50244140625,
0.60302734375
],
[
0.07843017578125,
0.60400390625,
0.468017578125,
0.60400390625,
0.468017578125,
0.6171875,
0.07843017578125,
0.6171875
],
[
0.5009765625,
0.60302734375,
0.8896484375,
0.60302734375,
0.8896484375,
0.6162109375,
0.5009765625,
0.6162109375
],
[
0.07843017578125,
0.61865234375,
0.468017578125,
0.6171875,
0.468017578125,
0.6318359375,
0.07843017578125,
0.63330078125
],
[
0.50341796875,
0.62060546875,
0.890625,
0.62060546875,
0.890625,
0.630859375,
0.50341796875,
0.630859375
],
[
0.080078125,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.080078125,
0.64794921875
],
[
0.50341796875,
0.63623046875,
0.8896484375,
0.63623046875,
0.8896484375,
0.646484375,
0.50341796875,
0.646484375
],
[
0.08087158203125,
0.65087890625,
0.46728515625,
0.65087890625,
0.46728515625,
0.6611328125,
0.08087158203125,
0.6611328125
],
[
0.5009765625,
0.6484375,
0.890625,
0.6494140625,
0.890625,
0.66357421875,
0.5009765625,
0.662109375
],
[
0.080078125,
0.66357421875,
0.46728515625,
0.66357421875,
0.46728515625,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.5009765625,
0.6640625,
0.89208984375,
0.6640625,
0.89208984375,
0.67822265625,
0.5009765625,
0.67822265625
],
[
0.080078125,
0.6787109375,
0.4697265625,
0.6787109375,
0.4697265625,
0.6923828125,
0.080078125,
0.6923828125
],
[
0.50244140625,
0.6787109375,
0.89208984375,
0.6787109375,
0.89208984375,
0.6923828125,
0.50244140625,
0.6923828125
],
[
0.080078125,
0.6943359375,
0.468017578125,
0.6943359375,
0.468017578125,
0.70849609375,
0.080078125,
0.70849609375
],
[
0.50244140625,
0.6943359375,
0.890625,
0.6943359375,
0.890625,
0.70849609375,
0.50244140625,
0.70849609375
],
[
0.07843017578125,
0.70849609375,
0.4697265625,
0.708984375,
0.4697265625,
0.7236328125,
0.07843017578125,
0.72265625
],
[
0.50244140625,
0.71044921875,
0.890625,
0.71044921875,
0.890625,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.07843017578125,
0.724609375,
0.468017578125,
0.7236328125,
0.468017578125,
0.7373046875,
0.07843017578125,
0.73779296875
],
[
0.50244140625,
0.7236328125,
0.89208984375,
0.724609375,
0.89208984375,
0.73779296875,
0.50244140625,
0.7373046875
],
[
0.08087158203125,
0.7412109375,
0.468017578125,
0.7412109375,
0.468017578125,
0.751953125,
0.08087158203125,
0.751953125
],
[
0.50244140625,
0.74072265625,
0.89208984375,
0.74072265625,
0.89208984375,
0.75390625,
0.50244140625,
0.75390625
],
[
0.080078125,
0.7548828125,
0.468017578125,
0.7548828125,
0.468017578125,
0.76806640625,
0.080078125,
0.76806640625
],
[
0.50244140625,
0.7548828125,
0.89208984375,
0.7548828125,
0.89208984375,
0.76806640625,
0.50244140625,
0.76806640625
],
[
0.08087158203125,
0.771484375,
0.468017578125,
0.771484375,
0.468017578125,
0.7822265625,
0.08087158203125,
0.7822265625
],
[
0.50341796875,
0.771484375,
0.890625,
0.771484375,
0.890625,
0.7822265625,
0.50341796875,
0.7822265625
],
[
0.080078125,
0.78515625,
0.468017578125,
0.78515625,
0.468017578125,
0.798828125,
0.080078125,
0.798828125
],
[
0.50341796875,
0.78515625,
0.89208984375,
0.78515625,
0.89208984375,
0.798828125,
0.50341796875,
0.798828125
],
[
0.080078125,
0.8017578125,
0.46728515625,
0.8017578125,
0.46728515625,
0.8125,
0.080078125,
0.8125
],
[
0.50244140625,
0.80029296875,
0.89208984375,
0.80029296875,
0.89208984375,
0.814453125,
0.50244140625,
0.814453125
],
[
0.5048828125,
0.8173828125,
0.890625,
0.8173828125,
0.890625,
0.82763671875,
0.5048828125,
0.82763671875
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.82763671875,
0.468017578125,
0.84326171875,
0.080078125,
0.84228515625
],
[
0.50244140625,
0.830078125,
0.89208984375,
0.830078125,
0.89208984375,
0.84326171875,
0.50244140625,
0.84326171875
],
[
0.111083984375,
0.8466796875,
0.3896484375,
0.8466796875,
0.3896484375,
0.861328125,
0.111083984375,
0.861328125
],
[
0.5009765625,
0.84326171875,
0.890625,
0.8447265625,
0.890625,
0.859375,
0.5009765625,
0.85791015625
],
[
0.50341796875,
0.85986328125,
0.890625,
0.85986328125,
0.890625,
0.87353515625,
0.50341796875,
0.87353515625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.87548828125,
0.89208984375,
0.87548828125,
0.89208984375,
0.8896484375,
0.50244140625,
0.8896484375
],
[
0.07843017578125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"Figure 2.Datasets rendered in CARLA across viewpoints: For the analysis part of our work, we use CARLA to simulate different",
"viewpoints. We rendered datasets from a total of 36 viewpoints, a few of which are highlighted above, including the source rig (extrinsics",
"from nuScenes [2] dataset), +12° yaw, +12° pitch, +21 inch height, and -12° pitch and +18 inch height together.",
"goous eids un",
"",
"",
"Figure 3. Analysis of impact of viewpoint changes in CARLA: We train a source BEV model using Lift Splat Shoot (LSS) [21] and",
"Cross View Transformers (CVT) [29], denoted at point O on the axis of each graph. We then test the model across different target rigs",
"where the camera pitch, yaw, height, or pitch and height are changed, as denoted by the different points along the r axes. We also trained",
"each model on the target rig directly and refer to this model as the \"oracle\", as it reflects the expected upper bound IoU for each viewpoint.",
"is a useful task for benchmarking AV perception",
"viewpoint at inference time for BEV segmentation mod-",
"[21, 24, 1]. BEV segmentation requires a 2D to 3D un-",
"ages",
"els trained on a single, source rig. We use simulated data",
"projection to predict the position of objects surrounding",
"from CARLA [5] for this analysis for two reasons: (1) us-",
"the ego-vehicle from the BEV perspective. BEV segmen-",
"ing simulated data allows us to isolate the domain gaps be-",
"tation models usually consist of an image encoder, which",
"tween training and testing such that only camera viewpoint",
"extracts the features from images from the camera rig, and",
"changes, and (2) real AV datasets with large differences in",
"a decoder, which uses the image features to predict the ob-",
"camera position are not publicly available. Examples of dif-",
"jects of interest in the BEV coordinate frame. Existing",
"ferentcamera viewpointsrenderedinCARLAareshownin",
"methods condition on the extrinsics and intrinsics of each",
"Fig.2.For simplicity and ease of interpretation of our re-",
"camera in different ways.Lift-Splat-Shoot(LSS)[21land",
"sults, we conduct all experiments on a single camera rig,",
"Orthographic Feature Transform (OFT) [25] unproject fea-",
"containing a front facing camera, which we refer to as the",
"tures into a point cloud according to each camera's intrinsic",
"source rig. We first train a BEV segmentation model on data",
"and extrinsic parameters. LSS performs sum pooling along",
"rendered from the source rig. For this rig, we use the cam-",
"each pillar in the map-view, while OFT performs average",
"era parameters of sessions from the nuScenes dataset [2].",
"pooling.Othermethods,suchasCrossViewTransform-",
"Then, we render train and test datasets from different target",
"ers (CVT)[29l,treat camera intrinsics and extrinsics as a",
"rigs, which contain variations to the yaw, pitch, height, or",
"feature, rather than explicitly unprojecting. We use LSs",
"pitch and height of the camera. The train datasets are used",
"and CVT to conductbenchmarks,sincethesetwo methods",
"to train an oracle for each target rig, while the test datasets",
"encompass both convolutional and transformer-based archi-",
"are used to evaluate the model trained on the source rig in",
"tectures andexplicitand implicitgeometricrepresentations.",
"comparison to the oracle. For completeness, we sweep over",
"a large range of each extrinsic and render a train and test",
"3. Measuring the Impact of Camera Viewpoint",
"dataset on regular intervals. For pitch and yaw, we sweep",
"Variations on BEV Segmentation",
"from -20 to 20°, rendering a dataset every 4.For height",
"we sweep from 0 in to 30 in, rendering a dataset every 3 in.",
"Method: In this section, we introduce our approach and",
"For height and pitch together, we sweep from O° and O in to",
"results for measuring the impact of changing the camera",
"3"
] |
{
"arxiv_id": "2309.05192",
"language": "en",
"page_idx": 2
}
|
||
2309.05192-en-3
|
2309.05192-en-3.jpg
|
$-20^{\circ} and 30 in, rendering a dataset at every -4^{\circ} and 6 in.$
To understand the domain gap introduced by changes in camera position, we test the“source model”, which is the model trained on data from the source rig, across each test dataset, where each test dataset contains changes to either yaw, pitch, height, or pitch and height together. We then compare the test accuracy of the source model to the test accuracy of the oracle model, which was only trained on data from the target rig. The oracle model serves as an upper bound on model performance since there is no domain gap between the train and test datasets
Model Details: We conduct our analysis using two BEV segmentation models, Lift Splat Shoot (LSS) and Cross View Transformers (CVT). LSS uses an explicit geometric operation to map objects in the camera coordinate system to the bird's eye coordinate system. It does this by using each camera's intrinsic and extrinsic parameters to construct a frustum shaped point cloud per camera where predicted objects are placed inside. A convolutional encoder maps images to features and depths, which are unprojected into the frustum, and a cumulative summing operation is done over the features in the vertical pillars of the frustum before the decoder then predicts the final segmentation. In contrast, CVT uses a transformer to learn features over images, extrinsics, and intrinsics. The extrinsic and intrinsic parameters are used to condition the segmentation network, such that it implicitly learns correlations between the parameters and positions of objects relative to the ego-vehicle. We use these two architectures because they cover both explicit and implicit geometric representations and convolutional and transformer backbones, allowing us to test the impact of each on generalization to viewpoint changes.
Results: Results of our analysis are shown in Fig. 3. We see that the performance of both LSS and CVT suffers drastically with even small changes to camera viewpoint, whether it be pitch, yaw, height, or pitch and height together. Because of the architecture of LSS, which includes cumulative summing in the vertical pillars within each frustum, changes to camera height have a relatively small impact on downstream BEV segmentation performance in comparison to other viewpoint changes. CVT lacks this generalization to changes in camera height because it does not sum features in the height dimension, but rather conditions on the camera extrinsics. We also note that because the training dataset is acquired in simulation, the extrinsics of the source rig have no noise or calibration error, and, thus, are always the same during training. As a result, we found that CVT learns to ignore the extrinsic embedding during training, indicating that the degradations we see to test performance in Fig. 3 are the result of the images being out of distribution. In contrast, our experiments in Sec. 5 involve training CVT on real world data, which has calibration error, and, as a result, CVT learns to use the extrinsic embedding to inform
predictions, but still lacks generalization to target rigs.
During our analysis, we also found that while changes to yaw have a negative impact on performance on LSS, the resulting segmentation predictions are transformed based on the difference in yaw between training and testing. To mitigate this, a post-processing step can be applied where the predictions are rotated to the viewpoint of the target rig. Post-processing can be used to mitigate the effect of changes in yaw, but does not generalize to other extrinsic parameters, such as pitch and height.
Lastly, we note two biases in the oracle models. First, we observe that the LSS oracle model trained on negative pitches performs poorly. Second, both LSS and CVT achieve higher test IoU when trained and tested with rigs with a larger camera height. While higher IoU could be explained by fewer occlusions due to a higher viewpoint, and thus more ground truth pixels, the number of ground truth objects is consistent across each of the test datasets (7 objects per frame on average), and so this bias is not explained by differences in the number of ground truth pixels. We note these biases, but they are not the main focus of our work.
Training Details: We train each BEV segmentation model three times and show the mean and standard deviation in test IoU in Fig. 3. Each model is trained on 25,000 images rendered from the front center camera (same camera parameters as in nuScenes) with the CARLA Simulator [5]. Train datasets are created for all camera viewpoints tested so that an oracle model can be constructed. For evaluation, we use 5,000 test images from each target rig, where the target rigs include changes to camera pitch, yaw, height, or pitch and height together, and are rendered from a different CARLA map than the training sets. Each model is trained for 30 epochs. We will release all 36 train and test datasets with this paper. The 36 datasets include train and test data for the source rig, 10 pitch rigs, 10 yaw rigs, 10 height rigs, and 5 height and pitch rigs.
## 4. Viewpoint Robustness via NVS
We present a new method that improves generalization of BEV segmentation models to different camera positions using novel view synthesis (NVS). As described in Sec. 3, BEV segmentation models fail to generalize to even small changes in camera viewpoint. However, collecting new data from each target rig, especially when AV companies may wish to deploy models across many types of cars, is impractical due to the cost of collection and annotation. Thus, we focus on NVS as it provides an opportunity to reuse labeled data from the source rig by transforming it into the viewpoint of each target rig. We can then train a new model on the transformed data for each target rig. We first define our NVS method. The key difference between our NVS method and past work is how we generalize to complex, dy-
4
|
[
[
0.080078125,
0.09027099609375,
0.44775390625,
0.091552734375,
0.44775390625,
0.1060791015625,
0.080078125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.85888671875,
0.0921630859375,
0.85888671875,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.09722900390625,
0.10986328125,
0.468017578125,
0.10986328125,
0.468017578125,
0.125,
0.09722900390625,
0.125
],
[
0.521484375,
0.10919189453125,
0.89208984375,
0.10919189453125,
0.89208984375,
0.12249755859375,
0.521484375,
0.12249755859375
],
[
0.07843017578125,
0.1256103515625,
0.468017578125,
0.125,
0.468017578125,
0.1395263671875,
0.07843017578125,
0.14013671875
],
[
0.5009765625,
0.125,
0.89208984375,
0.125,
0.89208984375,
0.1383056640625,
0.5009765625,
0.1383056640625
],
[
0.080078125,
0.141357421875,
0.4697265625,
0.141357421875,
0.4697265625,
0.1546630859375,
0.080078125,
0.1546630859375
],
[
0.50244140625,
0.1395263671875,
0.89208984375,
0.1395263671875,
0.89208984375,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.080078125,
0.1546630859375,
0.468017578125,
0.1546630859375,
0.468017578125,
0.1685791015625,
0.080078125,
0.1685791015625
],
[
0.5009765625,
0.1546630859375,
0.89208984375,
0.1546630859375,
0.89208984375,
0.1685791015625,
0.5009765625,
0.1685791015625
],
[
0.0767822265625,
0.17041015625,
0.4697265625,
0.169189453125,
0.4697265625,
0.184326171875,
0.0767822265625,
0.1849365234375
],
[
0.50244140625,
0.169189453125,
0.89208984375,
0.169189453125,
0.89208984375,
0.18310546875,
0.50244140625,
0.18310546875
],
[
0.07843017578125,
0.186279296875,
0.4697265625,
0.1849365234375,
0.4697265625,
0.1988525390625,
0.07843017578125,
0.199462890625
],
[
0.50244140625,
0.1849365234375,
0.89208984375,
0.1849365234375,
0.89208984375,
0.1988525390625,
0.50244140625,
0.1988525390625
],
[
0.080078125,
0.20263671875,
0.468017578125,
0.20263671875,
0.468017578125,
0.21337890625,
0.080078125,
0.21337890625
],
[
0.5009765625,
0.199462890625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.21337890625,
0.5009765625,
0.2139892578125
],
[
0.080078125,
0.216552734375,
0.468017578125,
0.216552734375,
0.468017578125,
0.2298583984375,
0.080078125,
0.2298583984375
],
[
0.50244140625,
0.21533203125,
0.89208984375,
0.21533203125,
0.89208984375,
0.2291259765625,
0.50244140625,
0.2291259765625
],
[
0.080078125,
0.231689453125,
0.468017578125,
0.231689453125,
0.468017578125,
0.24560546875,
0.080078125,
0.24560546875
],
[
0.50244140625,
0.2310791015625,
0.74609375,
0.2310791015625,
0.74609375,
0.2442626953125,
0.50244140625,
0.2442626953125
],
[
0.07843017578125,
0.246826171875,
0.3095703125,
0.246826171875,
0.3095703125,
0.260009765625,
0.07843017578125,
0.260009765625
],
[
0.521484375,
0.246826171875,
0.8935546875,
0.246826171875,
0.8935546875,
0.260009765625,
0.521484375,
0.260009765625
],
[
0.080078125,
0.264404296875,
0.468017578125,
0.264404296875,
0.468017578125,
0.27783203125,
0.080078125,
0.27783203125
],
[
0.50244140625,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.080078125,
0.27978515625,
0.47216796875,
0.27978515625,
0.47216796875,
0.29345703125,
0.080078125,
0.29345703125
],
[
0.5009765625,
0.27587890625,
0.8896484375,
0.27587890625,
0.8896484375,
0.2890625,
0.5009765625,
0.2890625
],
[
0.080078125,
0.294189453125,
0.46728515625,
0.294189453125,
0.46728515625,
0.30810546875,
0.080078125,
0.30810546875
],
[
0.50244140625,
0.292236328125,
0.89208984375,
0.292236328125,
0.89208984375,
0.30615234375,
0.50244140625,
0.30615234375
],
[
0.080078125,
0.31005859375,
0.4697265625,
0.31005859375,
0.4697265625,
0.323974609375,
0.080078125,
0.323974609375
],
[
0.50244140625,
0.306884765625,
0.890625,
0.306884765625,
0.890625,
0.32080078125,
0.50244140625,
0.32080078125
],
[
0.0767822265625,
0.323974609375,
0.4697265625,
0.324462890625,
0.4697265625,
0.339111328125,
0.0767822265625,
0.33837890625
],
[
0.50244140625,
0.322509765625,
0.89208984375,
0.322509765625,
0.89208984375,
0.3359375,
0.50244140625,
0.3359375
],
[
0.080078125,
0.34033203125,
0.4697265625,
0.34033203125,
0.4697265625,
0.354248046875,
0.080078125,
0.354248046875
],
[
0.5009765625,
0.337158203125,
0.89208984375,
0.3359375,
0.89208984375,
0.35107421875,
0.5009765625,
0.3515625
],
[
0.07843017578125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.07843017578125,
0.369384765625
],
[
0.50341796875,
0.354248046875,
0.8896484375,
0.354248046875,
0.8896484375,
0.3642578125,
0.50341796875,
0.3642578125
],
[
0.080078125,
0.37060546875,
0.471435546875,
0.37060546875,
0.471435546875,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.50244140625,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50244140625,
0.3818359375
],
[
0.08087158203125,
0.386962890625,
0.468017578125,
0.386962890625,
0.468017578125,
0.397705078125,
0.08087158203125,
0.397705078125
],
[
0.5009765625,
0.3818359375,
0.890625,
0.3818359375,
0.890625,
0.395751953125,
0.5009765625,
0.395751953125
],
[
0.080078125,
0.40087890625,
0.471435546875,
0.40087890625,
0.471435546875,
0.4140625,
0.080078125,
0.4140625
],
[
0.50341796875,
0.399658203125,
0.875,
0.399658203125,
0.875,
0.410400390625,
0.50341796875,
0.410400390625
],
[
0.08087158203125,
0.417236328125,
0.46728515625,
0.417236328125,
0.46728515625,
0.427978515625,
0.08087158203125,
0.427978515625
],
[
0.50244140625,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.43115234375,
0.4697265625,
0.43115234375,
0.4697265625,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.50341796875,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.44580078125,
0.50341796875,
0.44580078125
],
[
0.080078125,
0.44580078125,
0.4697265625,
0.44580078125,
0.4697265625,
0.458984375,
0.080078125,
0.458984375
],
[
0.50244140625,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.07843017578125,
0.4609375,
0.4697265625,
0.4609375,
0.4697265625,
0.474853515625,
0.07843017578125,
0.474853515625
],
[
0.50244140625,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.080078125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.50244140625,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.0767822265625,
0.490478515625,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.0767822265625,
0.50390625
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.08087158203125,
0.50830078125,
0.46728515625,
0.50830078125,
0.46728515625,
0.51806640625,
0.08087158203125,
0.51806640625
],
[
0.5009765625,
0.50830078125,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.5009765625,
0.521484375
],
[
0.080078125,
0.521484375,
0.4697265625,
0.521484375,
0.4697265625,
0.53515625,
0.080078125,
0.53515625
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.08087158203125,
0.53857421875,
0.468017578125,
0.53857421875,
0.468017578125,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50244140625,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.080078125,
0.5517578125,
0.4697265625,
0.5517578125,
0.4697265625,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.080078125,
0.56640625,
0.438720703125,
0.56640625,
0.438720703125,
0.580078125,
0.080078125,
0.580078125
],
[
0.50341796875,
0.57080078125,
0.890625,
0.57080078125,
0.890625,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.08087158203125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.5966796875,
0.08087158203125,
0.5966796875
],
[
0.50244140625,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.080078125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.080078125,
0.630859375,
0.468017578125,
0.630859375,
0.468017578125,
0.64453125,
0.080078125,
0.64453125
],
[
0.5009765625,
0.62890625,
0.68115234375,
0.6298828125,
0.68115234375,
0.64453125,
0.5009765625,
0.64306640625
],
[
0.08087158203125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.66357421875,
0.78173828125,
0.66357421875,
0.78173828125,
0.67822265625,
0.50341796875,
0.67822265625
],
[
0.07843017578125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.689453125,
0.07843017578125,
0.689453125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.51953125,
0.68798828125,
0.89208984375,
0.689453125,
0.89208984375,
0.7041015625,
0.51953125,
0.70263671875
],
[
0.07843017578125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.07843017578125,
0.71826171875
],
[
0.5009765625,
0.7041015625,
0.89208984375,
0.705078125,
0.89208984375,
0.7197265625,
0.5009765625,
0.71826171875
],
[
0.07843017578125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.07843017578125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.07843017578125,
0.75,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.07843017578125,
0.7646484375
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
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,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
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,
0.468017578125,
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,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5009765625,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.0767822265625,
0.88525390625,
0.471435546875,
0.88623046875,
0.471435546875,
0.90087890625,
0.0767822265625,
0.8994140625
],
[
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,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"-20° and 30 in, rendering a dataset at every -4 and 6 in.",
"predictions, but still lacks generalization to target rigs.",
"To understand the domain gap introduced by changes in",
"During our analysis, we also found that while changes to",
"camera position, we test the \"source model\", which is the",
"yaw have a negative impact on performance on LsS, the",
"model trained on data from the source rig, across each test",
"resulting segmentation predictions are transformed based",
"dataset, where each test dataset contains changes to either",
"on the difference in yaw between training and testing. To",
"yaw, pitch, height, or pitch and height together. We then",
"mitigate this, a post-processing step can be applied where",
"compare the test accuracy of the source model to the test",
"the predictions are rotated to the viewpoint of the target",
"accuracy of the oracle model, which was onlytrained on",
"rig. Post-processing can be used to mitigate the effect of",
"data from the target rig. The oracle model serves as an upper",
"changes in yaw, but does not generalize to other extrinsic",
"bound on model performance since there is no domain gap",
"parameters, such as pitch and height.",
"between the train and test datasets.",
"Lastly, we note two biases in the oracle models. First,",
"Model Details: We conduct our analysis using two BEV",
"we observe that the LsS oracle model trained on nega-",
"segmentation models, Lift Splat Shoot (LSS) and Cross",
"tive pitches performs poorly. Second, both LSS and CVT",
"View Transformers (CVT). LSS uses an explicit geometric",
"achieve higher test IoU when trained and tested with rigs",
"operation to map objects in the camera coordinate system",
"with a larger camera height. While higher loU could be ex-",
"to the bird's eye coordinate system. It does this by using",
"plained by fewer occlusions due to a higher viewpoint, and",
"each camera's intrinsic and extrinsic parameters to construct",
"thus more ground truth pixels, the number of ground truth",
"a frustum shaped point cloud per camera where predicted",
"objects is consistent across each of the test datasets (7 ob-",
"objects are placed inside.A convolutional encoder maps",
"jects per frame on average), and so this bias is not explained",
"images to features and depths,which are unprojectedinto",
"by differences in the number of ground truth pixels. We note",
"the frustum, and a cumulative summing operation is done",
"these biases, but they are not the main focus of our work.",
"overthe features in the vertical pillars of the frustum be-",
"Training Details: We train each BEV segmentation model",
"fore the decoder then predicts the final segmentation. In",
"three times and show the mean and standard deviation in",
"contrast, CvT uses a transformer to learn features over im-",
"test IoU in Fig. 3. Each model is trained on 25,000 im-",
"ages, extrinsics, and intrinsics. The extrinsic and intrinsic",
"ages rendered from the front center camera (same camera",
"parameters are used to condition the segmentation network,",
"parameters as in nuScenes) with the CARLA Simulator [5].",
"such that it implicitly learns correlations between the pa-",
"Train datasets are created for all camera viewpoints tested",
"rametersandpositionsofobjectsrelativetotheego-vehicle.",
"so that an oracle model can be constructed. For evaluation,",
"We use these two architectures because they cover both ex-",
"we use 5,000 test images from each target rig, where the",
"plicit and implicit geometric representations and convolu-",
"target rigs include changes to camera pitch, yaw, height, or",
"tional and transformer backbones, allowing us to test the",
"pitch and height together, and are rendered from a different",
"impact of each on generalization to viewpoint changes.",
"CARLAmap than the training sets.Eachmodel istrained",
"Results:Results of our analysis are shown in Fig.3.We see",
"for 30 epochs. We will release all 36 train and test datasets",
"that the performance of both LSS and CVT suffers drasti-",
"with this paper. The 36 datasets include train and test data",
"cally with even small changes to camera viewpoint, whether",
"for the source rig, 10 pitch rigs, 10 yaw rigs, 10 height rigs,",
"it be pitch, yaw, height, or pitch and height together. Be-",
"and 5 height and pitch rigs.",
"cause of the architecture of LSS, which includes cumula-",
"tive summing in the vertical pillars within each frustum,",
"4. Viewpoint Robustness via NVS",
"changes to camera height have a relatively small impact on",
"downstream BEV segmentation performance in comparison",
"We present a new method that improves generalization",
"to other viewpoint changes. CVT lacks this generalization",
"of BEV segmentation models to different camera positions",
"to changes in camera height because it does not sum fea-",
"using novel view synthesis (NVS). As described in Sec. 3,",
"tures in the height dimension, but rather conditions on the",
"BEV segmentation models fail to generalize to even small",
"camera extrinsics. We also note that because the training",
"changesin cameraviewpoint.However,collectingnew data",
"dataset is acquired in simulation, the extrinsics of the source",
"from each target rig, especially when AV companies may",
"rig have no noise or calibration error, and, thus, are always",
"wish to deploy models across many types of cars, is im-",
"the same during training. As a result, we found that CVT",
"practical due to the cost of collection and annotation. Thus,",
"learns to ignore the extrinsic embedding during training, in-",
"we focus on NVS as it provides an opportunity to reuse la-",
"dicating that the degradations we see to test performance in",
"beled data from the source rig by transforming it into the",
"Fig. 3 are the result of the images being out of distribution.",
"viewpoint of each target rig. We can then train a new model",
"In contrast, our experiments in Sec. 5 involve training CVT",
"on the transformed data for each target rig. We first define",
" on real world data, which has calibration error, and, as a re-",
"our NVS method. The key difference between our NVS",
" sult, CVT learns to use the extrinsic embedding to inform",
"method and past work is how we generalize to complex, dy-",
"4"
] |
{
"arxiv_id": "2309.05192",
"language": "en",
"page_idx": 3
}
|
||
2309.05192-en-4
|
2309.05192-en-4.jpg
|
<FigureHere>
Figure 4. Proposed Pipeline. Current methods for bird's eye view (BEV) segmentation are trained on data captured from one set of camera rigs (the source rig). At inference time, these models perform well on that camera rig, but, according to our analysis, even small changes in camera viewpoint lead to large drops in BEV segmentation accuracy. Our solution is to use novel view synthesis to augment the training dataset. We find this simple solution drastically improves the robustness of BEV segmentation models to data from a target camera rig, even when no real data from the target rig is available during training
<FigureHere>
Figure 5. NVS Qualitative Comparison: We compare the unrec-
namic AV scenes. Then, we show how the transformed data can be used to train BEV segmentation models for diverse target rigs without access to real data from the target rig. We use real data to train our NVS and BEV segmentation models. To evaluate over diverse target rigs, we use synthetic data rendered with NVIDIA DRIVE Sim since real data only provides one rig setting. We compare test performance achieved with models trained with data transformed to the target viewpoint vs. only data from a source rig. Our approach is summarized in Fig. 4.
tified NVS results (top) and depth results (bottom) from Worldsheet [11] (right) to our method (middle and left). SSIM is SSIM loss, ML is min loss, AM is automasking, LS is lidar supervision.
## 4.1. Preliminaries
camera pose, $\theta_t.$ The target view is rendered and compared to the GT with L1 and perceptual losses. A pix2pixHD generator inpaints parts of the scene in the generated target view that were not visible in the input. In contrast, we omit the pix2pixHD generator and use lidar depth supervision (LS), SSIM loss [27], automasking (AM) & minimum loss (ML) over neighboring frames [7] to build an NVS model that generalizes to complex, dynamic, AV scenes.
We build off of Worldsheet [11], a recent method for single image NVS of static scenes, extending it to work on complex AV scenes that have dynamic objects and occlusions. While NeRF-type approaches generate impressive NVS results, generalizing to dynamic scenes and across many scenes is still an active area of research. Worldsheet, on the other hand, is able to generalize across scenes, which is why we choose to use it in our work. The goal of Worldsheet is to build a 3D scene mesh, $M$, by warping a $W\times H$ lattice grid onto the scene based on predicted depths and vertex offsets. Given an input image, $I$, a ResNet-50[9] is trained to predict depth, $z$, and grid offset of each vertex, $V_{(x,y)}$ at each $(x,y)$ in $I.z$ and $V_(x,y)$ are used to build $M=(\{V_{(x,y)}\},\{F\})$,where $F$ are the mesh faces. A differentiable texture sampler is then used to splat the RGB pixel intensities from the original image onto the mesh's UV texture map. The pipeline is trained end-to-end on a multiview consistency loss. Given two views of the scene, an input and a target, the mesh is predicted from the input view and then projected to the target view based on the target
## $4. 2. \textbf{ Novel View Synthesis for AV Data}$
Overview: Because AV sessions, $S$, are composed of temporally sequential images, $\{I_0,I_1,...,I_n\}$ $\in S$, temporal consistency, rather than multi-view consistency, can be enforced between neighboring images to train our NVS model, assuming a sufficiently high frame rate so parts of the scene are visible in the input and target images. For every input image, $I_n$, we enforce consistency between $I_n-1$ and $I_n+ 1\textbf{by transforming }I_{n- 1}$ and $I_n+1$ into the viewpoint of $I_n$ and comparing each predicted novel view $\hat{I}_n$ to GT $I_n:$
5
|
[
[
0.09478759765625,
0.0889892578125,
0.265625,
0.09027099609375,
0.265625,
0.10479736328125,
0.09478759765625,
0.10418701171875
],
[
0.294189453125,
0.10986328125,
0.46728515625,
0.11297607421875,
0.465576171875,
0.1634521484375,
0.293212890625,
0.160400390625
],
[
0.61279296875,
0.1142578125,
0.6806640625,
0.12060546875,
0.67626953125,
0.1458740234375,
0.60888671875,
0.1395263671875
],
[
0.615234375,
0.1634521484375,
0.6748046875,
0.1634521484375,
0.6748046875,
0.1817626953125,
0.615234375,
0.1817626953125
],
[
0.221435546875,
0.1900634765625,
0.363525390625,
0.1900634765625,
0.363525390625,
0.203857421875,
0.221435546875,
0.203857421875
],
[
0.439453125,
0.1912841796875,
0.51123046875,
0.1912841796875,
0.51123046875,
0.20263671875,
0.439453125,
0.20263671875
],
[
0.54736328125,
0.1912841796875,
0.63623046875,
0.1912841796875,
0.63623046875,
0.2020263671875,
0.54736328125,
0.2020263671875
],
[
0.64892578125,
0.189453125,
0.71337890625,
0.1912841796875,
0.71337890625,
0.20263671875,
0.64892578125,
0.2008056640625
],
[
0.338134765625,
0.25,
0.398681640625,
0.25,
0.398681640625,
0.261474609375,
0.338134765625,
0.261474609375
],
[
0.328369140625,
0.26318359375,
0.408447265625,
0.26318359375,
0.408447265625,
0.2734375,
0.328369140625,
0.2734375
],
[
0.56396484375,
0.27587890625,
0.6357421875,
0.27587890625,
0.6357421875,
0.287353515625,
0.56396484375,
0.287353515625
],
[
0.2303466796875,
0.285400390625,
0.28515625,
0.285400390625,
0.28515625,
0.29541015625,
0.2303466796875,
0.29541015625
],
[
0.47216796875,
0.28271484375,
0.52294921875,
0.28271484375,
0.52294921875,
0.29345703125,
0.47216796875,
0.29345703125
],
[
0.62744140625,
0.3017578125,
0.77783203125,
0.3017578125,
0.77783203125,
0.3125,
0.62744140625,
0.3125
],
[
0.080078125,
0.31494140625,
0.89208984375,
0.31640625,
0.89208984375,
0.330810546875,
0.080078125,
0.330078125
],
[
0.080078125,
0.330810546875,
0.89208984375,
0.330810546875,
0.89208984375,
0.3447265625,
0.080078125,
0.3447265625
],
[
0.08251953125,
0.3466796875,
0.890625,
0.3466796875,
0.890625,
0.356689453125,
0.08251953125,
0.356689453125
],
[
0.080078125,
0.359130859375,
0.89208984375,
0.359130859375,
0.89208984375,
0.37255859375,
0.080078125,
0.37255859375
],
[
0.080078125,
0.37255859375,
0.490234375,
0.37255859375,
0.490234375,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.55615234375,
0.40771484375,
0.60302734375,
0.40771484375,
0.60302734375,
0.41845703125,
0.55615234375,
0.41845703125
],
[
0.6640625,
0.4072265625,
0.74169921875,
0.4091796875,
0.74169921875,
0.419189453125,
0.6640625,
0.417236328125
],
[
0.08087158203125,
0.41357421875,
0.46728515625,
0.41357421875,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.77392578125,
0.40771484375,
0.88623046875,
0.40771484375,
0.88623046875,
0.41845703125,
0.77392578125,
0.41845703125
],
[
0.080078125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.6748046875,
0.431884765625,
0.7158203125,
0.431884765625,
0.7158203125,
0.447509765625,
0.6748046875,
0.447509765625
],
[
0.080078125,
0.442626953125,
0.46728515625,
0.442626953125,
0.46728515625,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.080078125,
0.45703125,
0.468017578125,
0.45703125,
0.468017578125,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.080078125,
0.472900390625,
0.46728515625,
0.472900390625,
0.46728515625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.59228515625,
0.48291015625,
0.6435546875,
0.48486328125,
0.64208984375,
0.5048828125,
0.5908203125,
0.50244140625
],
[
0.6396484375,
0.479248046875,
0.77783203125,
0.47607421875,
0.779296875,
0.51025390625,
0.640625,
0.51318359375
],
[
0.63818359375,
0.486083984375,
0.65185546875,
0.4892578125,
0.64794921875,
0.50244140625,
0.6337890625,
0.5
],
[
0.77783203125,
0.48291015625,
0.86669921875,
0.48291015625,
0.86669921875,
0.5087890625,
0.77783203125,
0.5087890625
],
[
0.57275390625,
0.496826171875,
0.5810546875,
0.496826171875,
0.5810546875,
0.5,
0.57275390625,
0.5
],
[
0.080078125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.080078125,
0.50244140625,
0.468017578125,
0.50244140625,
0.468017578125,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.08087158203125,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.52978515625,
0.08087158203125,
0.52978515625
],
[
0.08087158203125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.54541015625,
0.08087158203125,
0.54541015625
],
[
0.50244140625,
0.5322265625,
0.8896484375,
0.5322265625,
0.8896484375,
0.54541015625,
0.50244140625,
0.54541015625
],
[
0.07843017578125,
0.548828125,
0.303955078125,
0.548828125,
0.303955078125,
0.5625,
0.07843017578125,
0.5625
],
[
0.5009765625,
0.54443359375,
0.890625,
0.54541015625,
0.890625,
0.56005859375,
0.5009765625,
0.5595703125
],
[
0.50244140625,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50244140625,
0.57373046875
],
[
0.080078125,
0.57373046875,
0.2181396484375,
0.57470703125,
0.2181396484375,
0.5888671875,
0.080078125,
0.58837890625
],
[
0.50341796875,
0.57568359375,
0.88818359375,
0.57568359375,
0.88818359375,
0.5859375,
0.50341796875,
0.5859375
],
[
0.1004638671875,
0.5986328125,
0.46728515625,
0.5986328125,
0.46728515625,
0.6123046875,
0.1004638671875,
0.6123046875
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.080078125,
0.62890625,
0.46728515625,
0.62890625,
0.46728515625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.07843017578125,
0.64306640625,
0.468017578125,
0.64453125,
0.468017578125,
0.6591796875,
0.07843017578125,
0.65771484375
],
[
0.50244140625,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50244140625,
0.671875
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.67236328125,
0.890625,
0.671875,
0.890625,
0.68603515625,
0.50244140625,
0.6875
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.07843017578125,
0.705078125,
0.46728515625,
0.7041015625,
0.46728515625,
0.71826171875,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.7021484375,
0.89208984375,
0.71630859375,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.50341796875,
0.7197265625,
0.7998046875,
0.7197265625,
0.7998046875,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.5048828125,
0.7421875,
0.7939453125,
0.7421875,
0.7939453125,
0.755859375,
0.5048828125,
0.755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.08087158203125,
0.77978515625,
0.471435546875,
0.77978515625,
0.471435546875,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.08251953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81103515625,
0.08251953125,
0.81103515625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82568359375,
0.468017578125,
0.82470703125,
0.468017578125,
0.8388671875,
0.07843017578125,
0.84033203125
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.465576171875,
0.85791015625,
0.465576171875,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.8896484375,
0.85693359375,
0.8896484375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"NovelView Synthesis Pipeline",
"",
"口",
"",
"Monocular Depth Estimation",
"MeshCreation",
"ChangeViewpoint",
"RenderView",
"Viewpoint",
"Augmentation",
"BEV Training",
"Source Rig",
"TargetRig",
"BEVSegmentationPipeline",
"Figure 4. Proposed Pipeline. Current methods for bird's eye view (BEV) segmentation are trained on data captured from one set of camera",
"rigs (the source rig). At inference time, these models perform well on that camera rig, but, according to our analysis, even small changes",
"incameraviewpointleadtolargedropsinBEV segmentationaccuracy.Our solution isto usenovel viewsynthesistoaugmentthetraining",
"dataset. We find this simple solution drastically improves the robustness of BEV segmentation models to data from a target camera rig,",
"even when no real data from the target rig is available during training.",
"Original ws",
"WS+ SSIM, ML, AM",
"namic AVscenes.Then,we showhowthetransformed data",
"WS+SSIM,ML, AM,LS(Ours",
"can be used to train BEV segmentation models for diverse",
"",
"target rigs without access to real data from the target rig.",
"We use real data to train our NVS and BEV segmentation",
"models. To evaluate over diverse target rigs, we use syn-",
"",
"",
"P",
"",
"",
"thetic data rendered with NVIDIA DRIVE Sim since real",
"data only provides one rig setting. We compare test perfor-",
"manceachieved withmodelstrained withdatatransformed",
"to thetarget viewpoint vs. only data from a source rig.Our",
"Figure 5. NVS Qualitative Comparison: We compare the unrec-",
"approach is summarized in Fig. 4.",
"tified NvS results (top) and depth results (bottom) from World-",
"sheet [11] (right) to our method (middle and left). SSIM is SSIM",
"4.1. Preliminaries",
"loss,MLis minloss,AMis automasking,LS islidar supervision.",
"We build off of Worldsheet [11], a recent method for",
"single image NVS of static scenes, extending it to work",
"camera pose, Ot. The target view is rendered and compared",
"on complex AV scenes that have dynamic objects and oc-",
"to the GT with L1 and perceptual losses. A pix2pixHD gen-",
"clusions..While NeRF-type approaches generate impres-",
"erator inpaints parts of the scene in the generated target view",
"sive NVS results, generalizing to dynamic scenes and across",
"that were not visible in the input. In contrast, we omit the",
"many scenes is still an active area of research. Worldsheet,",
"pix2pixHD generator and use lidar depth supervision (LS),",
"on the other hand, is able to generalize across scenes, which",
"SSIM loss [27], automasking (AM) & minimum loss (ML)",
"is why we choose to use it in our work. The goal of World-",
"over neighboring frames [7] to build an NvS model that",
"sheet is to build a 3D scene mesh, M, by warping a W x H",
"generalizes to complex, dynamic, AV scenes.",
"lattice grid ontothe scene based on predicted depths and",
"4.2. Novel View Synthesis for AV Data",
"vertex offsets. Given an input image, I, a ResNet-50 [9] is",
"trained to predict depth, z, and grid offset of each vertex,",
"Overview: Because AV sessions, S, are composed of tem-",
"V(r,y) at each (r, y) in I. z and V(r,y) are used to build",
"porally sequential images, [Io, Ii,.., In] E S, tempo-",
"M = ({V(r,)], {F)), where F are the mesh faces. A dif-",
"ral consistency, rather than multi-view consistency, can be",
"ferentiable texture sampler is then used to splat the RGB",
"enforced between neighboring images to train our NVS",
"pixel intensities from the original image onto the mesh's UV",
"model, assuming a sufficiently high frame rate so parts of",
"texture map. The pipeline is trained end-to-end on a multi-",
"the scene are visible in the input and target images. For ev-",
"viewconsistencyloss.Giventwoviewsofthescene,an in-",
"ery input image, In, we enforce consistency between In-1",
"put and a target, the mesh is predicted from the input view",
"and In+1 by transforming In-1 and In+1 into the viewpoint",
"and then projected to the target view based on the target",
"of In and comparing each predicted novel view In to GT In:",
"5"
] |
{
"arxiv_id": "2309.05192",
"language": "en",
"page_idx": 4
}
|
||
2006.02598-en-0
|
2006.02598-en-0.jpg
|
## Info3D: Representation Learning on 3D Objects using Mutual Information Maximization and Contrastive Learning
Aditya Sanghi
${\mathrm{Autodesk~AI~Lab,~Toronto,~Canada}}$ aditya. sanghi@autodesk. com
Abstract. A major endeavor of computer vision is to represent, understand and extract structure from 3D data. Towards this goal, unsupervised learning is a powerful and necessary tool. Most current unsupervised methods for 3D shape analysis use datasets that are aligned, require objects to be reconstructed and suffer from deteriorated performance on downstream tasks. To solve these issues, we propose to extend the InfoMax and contrastive learning principles on 3D shapes. We show that we can maximize the mutual information between 3D objects and their“chunks” to improve the representations in aligned datasets. Furthermore, we can achieve rotation invariance in SO(3) group by maximizing the mutual information between the 3D objects and their geometric transformed versions. Finally, we conduct several experiments such as clustering, transfer learning, shape retrieval, and achieve state of art results.
Keywords: 3D Shape Analysis, Unsupervised Learning, Rotation In-
$\dot{\text{variance, InfoMax, Contrastive Learning}}$
### 1 Introduction
Recently, several unsupervised methods have managed to extract powerful features for 3D objects such as in [16],[49],[1],[30] and [26]. However, these methods assume all 3D objects are aligned and have the same pose in the given category. In real world scenarios, this is not the case. For example, when a robot is identifying and picking up an object, the object is in an unknown pose. Even in online repositories of 3D shapes, most of the data is randomly oriented as users create objects in different poses. To use these methods effectively we require to align all objects for a given category which is a very expensive and time consuming process.
Furthermore, these unsupervised methods require reconstruction of 3D shapes which is not ideal for many reasons. Firstly, it is not always feasible to reconstruct the 3D representation of a shape. For example, due to the discrete nature of meshes, reconstructing their representation may not be attainable. Moreover, in cases where we need invariant representation, it's hard to reconstruct back the
|
[
[
0.223876953125,
0.1458740234375,
0.78125,
0.1458740234375,
0.78125,
0.1634521484375,
0.223876953125,
0.1634521484375
],
[
0.24267578125,
0.1673583984375,
0.7607421875,
0.1666259765625,
0.7607421875,
0.184326171875,
0.24267578125,
0.1849365234375
],
[
0.3798828125,
0.1881103515625,
0.625,
0.1900634765625,
0.625,
0.208984375,
0.3798828125,
0.20703125
],
[
0.451904296875,
0.2374267578125,
0.5556640625,
0.2374267578125,
0.5556640625,
0.251953125,
0.451904296875,
0.251953125
],
[
0.0196075439453125,
0.255126953125,
0.06781005859375,
0.255126953125,
0.06781005859375,
0.4091796875,
0.0196075439453125,
0.4091796875
],
[
0.384033203125,
0.264404296875,
0.62109375,
0.264404296875,
0.62109375,
0.279052734375,
0.384033203125,
0.279052734375
],
[
0.40185546875,
0.281005859375,
0.60302734375,
0.281005859375,
0.60302734375,
0.291748046875,
0.40185546875,
0.291748046875
],
[
0.2646484375,
0.330810546875,
0.73779296875,
0.33203125,
0.73779296875,
0.3466796875,
0.2646484375,
0.34521484375
],
[
0.2646484375,
0.3466796875,
0.740234375,
0.3466796875,
0.740234375,
0.35986328125,
0.2646484375,
0.35986328125
],
[
0.2646484375,
0.361083984375,
0.740234375,
0.361083984375,
0.740234375,
0.375,
0.2646484375,
0.375
],
[
0.2646484375,
0.375,
0.73876953125,
0.375,
0.73876953125,
0.38818359375,
0.2646484375,
0.38818359375
],
[
0.2646484375,
0.38818359375,
0.73876953125,
0.38818359375,
0.73876953125,
0.401611328125,
0.2646484375,
0.401611328125
],
[
0.0228729248046875,
0.397705078125,
0.062103271484375,
0.397705078125,
0.062103271484375,
0.498779296875,
0.0228729248046875,
0.498779296875
],
[
0.265625,
0.404052734375,
0.73876953125,
0.404052734375,
0.73876953125,
0.4140625,
0.265625,
0.4140625
],
[
0.2646484375,
0.416748046875,
0.73876953125,
0.416748046875,
0.73876953125,
0.429931640625,
0.2646484375,
0.429931640625
],
[
0.2646484375,
0.429931640625,
0.740234375,
0.429931640625,
0.740234375,
0.443115234375,
0.2646484375,
0.443115234375
],
[
0.2646484375,
0.443115234375,
0.73876953125,
0.443115234375,
0.73876953125,
0.45703125,
0.2646484375,
0.45703125
],
[
0.2646484375,
0.458251953125,
0.73876953125,
0.458251953125,
0.73876953125,
0.4716796875,
0.2646484375,
0.4716796875
],
[
0.2646484375,
0.4716796875,
0.740234375,
0.4716796875,
0.740234375,
0.48486328125,
0.2646484375,
0.48486328125
],
[
0.0228729248046875,
0.48486328125,
0.05963134765625,
0.48486328125,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.265625,
0.4873046875,
0.465576171875,
0.4873046875,
0.465576171875,
0.49755859375,
0.265625,
0.49755859375
],
[
0.463134765625,
0.4873046875,
0.73876953125,
0.4873046875,
0.73876953125,
0.49755859375,
0.463134765625,
0.49755859375
],
[
0.26318359375,
0.5,
0.73876953125,
0.5,
0.73876953125,
0.51318359375,
0.26318359375,
0.51318359375
],
[
0.2646484375,
0.5146484375,
0.30224609375,
0.5146484375,
0.30224609375,
0.52783203125,
0.2646484375,
0.52783203125
],
[
0.2646484375,
0.54052734375,
0.73779296875,
0.54052734375,
0.73779296875,
0.5537109375,
0.2646484375,
0.5537109375
],
[
0.26318359375,
0.55322265625,
0.53515625,
0.5546875,
0.53515625,
0.5693359375,
0.26318359375,
0.5673828125
],
[
0.221435546875,
0.59765625,
0.2344970703125,
0.59765625,
0.2344970703125,
0.609375,
0.221435546875,
0.609375
],
[
0.2484130859375,
0.5966796875,
0.373291015625,
0.5966796875,
0.373291015625,
0.611328125,
0.2484130859375,
0.611328125
],
[
0.2181396484375,
0.6298828125,
0.78369140625,
0.6298828125,
0.78369140625,
0.64306640625,
0.2181396484375,
0.64306640625
],
[
0.2181396484375,
0.64599609375,
0.7861328125,
0.64599609375,
0.7861328125,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.21728515625,
0.6591796875,
0.78515625,
0.66015625,
0.78515625,
0.6748046875,
0.21728515625,
0.673828125
],
[
0.2181396484375,
0.67529296875,
0.78515625,
0.67529296875,
0.78515625,
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,
0.78759765625,
0.70556640625,
0.78759765625,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.2181396484375,
0.736328125,
0.78759765625,
0.736328125,
0.78759765625,
0.75048828125,
0.2181396484375,
0.75048828125
],
[
0.2181396484375,
0.75244140625,
0.275390625,
0.75244140625,
0.275390625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2418212890625,
0.76513671875,
0.78759765625,
0.7666015625,
0.78759765625,
0.78076171875,
0.2418212890625,
0.77978515625
],
[
0.2181396484375,
0.77978515625,
0.78515625,
0.78076171875,
0.78515625,
0.79541015625,
0.2181396484375,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78515625,
0.8125,
0.78515625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.21728515625,
0.82763671875,
0.78759765625,
0.82763671875,
0.78759765625,
0.84130859375,
0.21728515625,
0.84130859375
]
] |
[
"Info3D: Representation Learning on 3D Objects",
"using Mutual Information Maximization and",
"Contrastive Learning",
"Aditya Sanghi",
"",
"Autodesk AI Lab, Toronto, Canada",
"aditya.sanghi@autodesk.com",
"Abstract. A major endeavor of computer vision is to represent, un-",
"derstand and extract structure from 3D data. Towards this goal, unsu-",
"pervised learning is a powerful and necessary tool. Most current unsu-",
"pervised methods for 3D shape analysis use datasets that are aligned,",
"require objects to be reconstructed and suffer from deteriorated perfor-",
"['S]",
"mance on downstream tasks.To solvethese issues, we proposeto extend",
"the InfoMax and contrastive learning principles on 3D shapes. We show",
"that we can maximize the mutual information between 3D objects and",
"their “chunks\" to improve the representations in aligned datasets. Fur-",
"thermore, we can achieve rotation invariance in SO(3) group by maximiz-",
"ing the mutual information between the 3D objects and their geometric",
"865200021",
"transformed versions.Finally,",
"weconduct several experiments suchas",
"clustering, transfer learning, shape retrieval, and achieve state of art re-",
"sults.",
"Keywords: 3D Shape Analysis, Unsupervised Learning, Rotation In-",
"variance, InfoMax, Contrastive Learning",
"1.",
"Introduction",
"Recently, several unsupervised methods have managed to extract powerful fea-",
"tures for 3D objects such as in [16], [49], [1], [30] and [26].However, these methods",
"assume all 3D objects are aligned and have the same pose in the given category.",
"In real world scenarios, this is not the case. For example, when a robot is identi-",
"fying and picking up an object, the object is in an unknown pose. Even in online",
"repositories of 3D shapes, most of the data is randomly oriented as users create",
"objects in different poses. To use these methods effectively we require to align",
"all objects for a given category which is a very expensive and time consuming",
"process.",
"Furthermore, these unsupervised methods require reconstruction of 3D shapes",
"which is not ideal for many reasons. Firstly, it is not always feasible to recon-",
"struct the 3D representation of a shape. For example, due to the discrete nature",
"of meshes, reconstructing their representation may not be attainable. Moreover,",
"in cases where we need invariant representation, it's hard to reconstruct back the"
] |
{
"arxiv_id": "2006.02598",
"language": "en",
"page_idx": 0
}
|
||
2006.02598-en-1
|
2006.02598-en-1.jpg
|
2
A. Sanghi
<FigureHere>
Fig.1: General idea of the method. We try to bring the 3D shape and the different view of the 3D shape closer in latent space while pushing away the representation of other objects in the dataset further away.
shape from this invariant representation. For example, if you wanted to create rotation invariant embeddings, you need to lose pose information from the embeddings of 3D objects. If you lose the pose information, it will not be possible to reconstruct the shape back as you need to reconstruct it with that given pose.
To overcome the challenges mentioned above we propose a decoder-free, unsupervised representation learning mechanism which is rotation insensitive. The method we introduce takes inspiration from the Contrastive Predictive Coding [29] and Deep InfoMax [43] approach. These methods usually require a different “view” of the object, which is used to maximize the mutual information with the object. This other view can be different modalities, data augmentation of the object, local substructure of the object, etc.
We consider two different views of a given object in this work. First, we consider maximizing the mutual information between a local chunk of a 3D object and the whole 3D object. The intuition for using this view is that the 3D shape is forced to learn about its local region as it has to distinguish it from other parts of different objects. This greatly enhances the representation learnt in aligned objects. Second, we consider maximizing the mutual information between 3D shape and a geometric transformed version of the 3D shape. The advantage of maximizing the mutual information in this scenario is that it can create global geometry invariant representations. This is very useful in the case of achieving rotation insensitive representation in SO(3). Figure 1 illustrates the rough intuition behind the method. Note, despite using objects from different category in the figure, we push away every other shape in the dataset. This might even include object from the same category. This method can be thought as instance discrimination [48].
The key contributions of our work are as follows:
- We introduce a decoder-free representation learning method which can easily
be extended to any 3D descriptor without needing to construct complex
decoder architectures.
|
[
[
0.2181396484375,
0.11236572265625,
0.2344970703125,
0.11614990234375,
0.2279052734375,
0.1307373046875,
0.2115478515625,
0.126953125
],
[
0.27294921875,
0.117431640625,
0.339111328125,
0.117431640625,
0.339111328125,
0.1287841796875,
0.27294921875,
0.1287841796875
],
[
0.2181396484375,
0.317626953125,
0.7861328125,
0.318115234375,
0.7861328125,
0.333251953125,
0.2181396484375,
0.33203125
],
[
0.21728515625,
0.333251953125,
0.7861328125,
0.333984375,
0.7861328125,
0.348388671875,
0.21728515625,
0.347900390625
],
[
0.2181396484375,
0.349853515625,
0.640625,
0.349853515625,
0.640625,
0.363037109375,
0.2181396484375,
0.363037109375
],
[
0.2181396484375,
0.39453125,
0.78515625,
0.39453125,
0.78515625,
0.40771484375,
0.2181396484375,
0.40771484375
],
[
0.2181396484375,
0.410400390625,
0.78369140625,
0.410400390625,
0.78369140625,
0.423583984375,
0.2181396484375,
0.423583984375
],
[
0.2181396484375,
0.4248046875,
0.7861328125,
0.4248046875,
0.7861328125,
0.438232421875,
0.2181396484375,
0.438232421875
],
[
0.21728515625,
0.43994140625,
0.78515625,
0.43994140625,
0.78515625,
0.453857421875,
0.21728515625,
0.453857421875
],
[
0.24267578125,
0.455078125,
0.78369140625,
0.455078125,
0.78369140625,
0.468505859375,
0.24267578125,
0.468505859375
],
[
0.2181396484375,
0.47021484375,
0.7861328125,
0.47021484375,
0.7861328125,
0.484130859375,
0.2181396484375,
0.484130859375
],
[
0.21728515625,
0.484130859375,
0.7861328125,
0.48486328125,
0.7861328125,
0.5,
0.21728515625,
0.498779296875
],
[
0.2197265625,
0.50048828125,
0.78515625,
0.50048828125,
0.78515625,
0.5146484375,
0.2197265625,
0.5146484375
],
[
0.2197265625,
0.51513671875,
0.78515625,
0.51513671875,
0.78515625,
0.52880859375,
0.2197265625,
0.52880859375
],
[
0.21728515625,
0.52880859375,
0.78759765625,
0.52978515625,
0.78759765625,
0.54443359375,
0.21728515625,
0.54345703125
],
[
0.2181396484375,
0.54443359375,
0.560546875,
0.54541015625,
0.560546875,
0.56005859375,
0.2181396484375,
0.5595703125
],
[
0.2442626953125,
0.56103515625,
0.7861328125,
0.56103515625,
0.7861328125,
0.57470703125,
0.2442626953125,
0.57470703125
],
[
0.21728515625,
0.57568359375,
0.7861328125,
0.57470703125,
0.7861328125,
0.5888671875,
0.21728515625,
0.59033203125
],
[
0.2181396484375,
0.59130859375,
0.7861328125,
0.59130859375,
0.7861328125,
0.60498046875,
0.2181396484375,
0.60498046875
],
[
0.2181396484375,
0.60595703125,
0.78759765625,
0.60595703125,
0.78759765625,
0.619140625,
0.2181396484375,
0.619140625
],
[
0.2181396484375,
0.62060546875,
0.7861328125,
0.62060546875,
0.7861328125,
0.63525390625,
0.2181396484375,
0.63525390625
],
[
0.2181396484375,
0.63623046875,
0.7861328125,
0.63623046875,
0.7861328125,
0.6494140625,
0.2181396484375,
0.6494140625
],
[
0.21728515625,
0.6513671875,
0.7861328125,
0.6513671875,
0.7861328125,
0.66552734375,
0.21728515625,
0.66552734375
],
[
0.2181396484375,
0.66650390625,
0.7861328125,
0.66650390625,
0.7861328125,
0.6796875,
0.2181396484375,
0.6796875
],
[
0.2181396484375,
0.681640625,
0.78515625,
0.681640625,
0.78515625,
0.69580078125,
0.2181396484375,
0.69580078125
],
[
0.21728515625,
0.6962890625,
0.7861328125,
0.6962890625,
0.7861328125,
0.71044921875,
0.21728515625,
0.71044921875
],
[
0.2181396484375,
0.7119140625,
0.78515625,
0.7119140625,
0.78515625,
0.72607421875,
0.2181396484375,
0.72607421875
],
[
0.2181396484375,
0.72802734375,
0.7861328125,
0.72802734375,
0.7861328125,
0.7412109375,
0.2181396484375,
0.7412109375
],
[
0.2181396484375,
0.7421875,
0.78515625,
0.7421875,
0.78515625,
0.755859375,
0.2181396484375,
0.755859375
],
[
0.2181396484375,
0.7568359375,
0.441162109375,
0.7568359375,
0.441162109375,
0.77099609375,
0.2181396484375,
0.77099609375
],
[
0.24267578125,
0.771484375,
0.59326171875,
0.771484375,
0.59326171875,
0.7861328125,
0.24267578125,
0.7861328125
],
[
0.2459716796875,
0.79541015625,
0.78515625,
0.79541015625,
0.78515625,
0.81005859375,
0.2459716796875,
0.81005859375
],
[
0.2459716796875,
0.8125,
0.7861328125,
0.8125,
0.7861328125,
0.82568359375,
0.2459716796875,
0.82568359375
],
[
0.2459716796875,
0.8271484375,
0.404296875,
0.8271484375,
0.404296875,
0.84033203125,
0.2459716796875,
0.84033203125
]
] |
[
"2",
"A.Sanghi",
"Fig.l: General idea of the method. We try to bring the 3D shape and the",
"different view of the 3D shape closer in latent space while pushing away the",
"representation of other objects in the dataset further away.",
"shape from this invariant representation. For example, if you wanted to create",
"rotation invariant embeddings, you need to lose pose information from the em-",
"beddings of 3D objects. If you lose the pose information, it will not be possible",
"to reconstruct the shape back as you need to reconstruct it with that given pose.",
"To overcome the challenges mentioned above we propose a decoder-free, un-",
"supervised representation learning mechanism which is rotation insensitive. The",
"method we introduce takes inspiration from the Contrastive Predictive Coding",
"[29] and Deep InfoMax [43] approach. These methods usually require a different",
"\"view\" of the object, which is used to maximize the mutual information with",
"the object. This other view can be different modalities, data augmentation of",
"the object, local substructure of the object, etc.",
"We consider two different views of a given object in this work, First, we",
"consider maximizing the mutual information between a local chunk of a 3D",
"object and the whole 3D object. The intuition for using this view is that the",
"3D shape is forced to learn about its local region as it has to distinguish it",
"from other parts of different objects. This greatly enhances the representation",
"learnt in aligned objects. Second, we consider maximizing the mutual information",
"between 3D shape and a geometric transformed version of the 3D shape. The",
"advantage of maximizing the mutual information in this scenario is that it can",
"create global geometry invariant representations. This is very useful in the case",
"of achieving rotation insensitive representation in SO(3). Figure 1 illustrates the",
"rough intuition behind the method. Note, despite using objects from different",
"category in the figure, we push away every other shape in the dataset.This",
"might even include object from the same category. This method can be thought",
"as instance discrimination [48].",
"The key contributions of our work are as follows:",
"We introduce a decoder-free representation learning method which can easily",
"be extended to any 3D descriptor without needing to construct complex",
"decoder architectures."
] |
{
"arxiv_id": "2006.02598",
"language": "en",
"page_idx": 1
}
|
||
2006.02598-en-2
|
2006.02598-en-2.jpg
|
Info3D 3
- We show how local chunks of 3D objects can be used to get very effective
representations for downstream tasks.
We demonstrate the effectiveness of the method on rotated inputs and show
how it is insensitive to such rotations in SO(3) group.
We conduct several experiments to show the effcacy of our method and achieve state of art results in transfer learning, clustering and semi-supervised $\operatorname{learning}.$
## 2 Background
Representation Learning on 3D objects. Much progress has been made to learn good representations of 3D objects in an unsupervised manner which can then be used in several downstream tasks. In point clouds, works such as [1],[49],[16],[18],[36] and [52] have been proposed. For voxels and implicit representations, work such as [46],[26],[10],[27],[30] create powerful representation features which are used for several downstream tasks such as shape completion, shape generation and classification. Recently, there has been a lot of progress in auto-encoding meshes such as in [40],[11],[40]. One disadvantage of the above approaches are that they require you to reconstruct or generate the 3D shapes. As stated earlier, it might be expensive or not possible to reconstruct the shape. A recent approach [51] does not require to reconstruct the shape and instead does two stages of training. It frst uses parts of a shape to learn features using contrastive learning and then uses pseudoclusters to cluster all the data. Our method does not require two stages of training, and furthermore allows us to create rotation invariant embeddings.
Maximizing Mutual Information and Contrastive Learning.A lot of methods have used mutual information to do unsupervised learning. Historically, works such as [24],[4],[6],[5] have explored the InfoMax principle and mutual information maximization. More recently [29] proposed the Contrastive Predicting Coding framework which uses the embeddings to capture maximal information about future samples. The Deep InfoMax (DIM) [43] approach is similar but has the advantage of doing orderless autoregression. In concurrent to these works, works such as [48] and [53], have extended these ideas from the metric learning point of view. These methods were extended in [3],[19],[42] by considering multiple views of the data and achieved state of the art representation learning on images. The InfoMax principle has been extended to graphs [43],[39] and for state representation in reinforcement learning [2].Our method is inspired by these methods and we extend them to 3D representations. For the multiple views, we use geometric transformations and chunks of a 3D object. Finally, in concurrent to our work several news works such as [9],[28] have been proposed.
$\textbf{Rotation Invariance on 3D objects. Traditionally, several methods have fo- }$ cused on hand engineered features to get local rotation invariant descriptors, such as [34],[37],[22]. More recently, methods such as [13] and [14],first, get
|
[
[
0.68310546875,
0.11553955078125,
0.734375,
0.11553955078125,
0.734375,
0.1300048828125,
0.68310546875,
0.1300048828125
],
[
0.77294921875,
0.11553955078125,
0.7890625,
0.11553955078125,
0.7890625,
0.1300048828125,
0.77294921875,
0.1300048828125
],
[
0.2296142578125,
0.146484375,
0.78515625,
0.147705078125,
0.78515625,
0.1622314453125,
0.2296142578125,
0.1610107421875
],
[
0.2467041015625,
0.1634521484375,
0.515625,
0.1634521484375,
0.515625,
0.177978515625,
0.2467041015625,
0.177978515625
],
[
0.2296142578125,
0.1787109375,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.2296142578125,
0.1925048828125
],
[
0.2442626953125,
0.1932373046875,
0.6298828125,
0.1944580078125,
0.6298828125,
0.208984375,
0.2442626953125,
0.2083740234375
],
[
0.2303466796875,
0.208984375,
0.7861328125,
0.208984375,
0.7861328125,
0.222900390625,
0.2303466796875,
0.222900390625
],
[
0.2442626953125,
0.2247314453125,
0.7861328125,
0.2247314453125,
0.7861328125,
0.238037109375,
0.2442626953125,
0.238037109375
],
[
0.24267578125,
0.238037109375,
0.310546875,
0.240478515625,
0.310546875,
0.255126953125,
0.24267578125,
0.25244140625
],
[
0.21728515625,
0.28271484375,
0.2303466796875,
0.279052734375,
0.236083984375,
0.2890625,
0.22216796875,
0.29345703125
],
[
0.2467041015625,
0.27783203125,
0.369384765625,
0.27783203125,
0.369384765625,
0.29541015625,
0.2467041015625,
0.29541015625
],
[
0.2197265625,
0.3125,
0.78515625,
0.3125,
0.78515625,
0.326904296875,
0.2197265625,
0.326904296875
],
[
0.2181396484375,
0.32763671875,
0.7861328125,
0.32763671875,
0.7861328125,
0.341552734375,
0.2181396484375,
0.341552734375
],
[
0.2197265625,
0.343505859375,
0.7861328125,
0.343505859375,
0.7861328125,
0.356689453125,
0.2197265625,
0.356689453125
],
[
0.2197265625,
0.35791015625,
0.7861328125,
0.35791015625,
0.7861328125,
0.371826171875,
0.2197265625,
0.371826171875
],
[
0.2181396484375,
0.373779296875,
0.78369140625,
0.373779296875,
0.78369140625,
0.386962890625,
0.2181396484375,
0.386962890625
],
[
0.21728515625,
0.386962890625,
0.7861328125,
0.38818359375,
0.7861328125,
0.40283203125,
0.21728515625,
0.401611328125
],
[
0.2181396484375,
0.40283203125,
0.7861328125,
0.40283203125,
0.7861328125,
0.416748046875,
0.2181396484375,
0.416748046875
],
[
0.2181396484375,
0.41845703125,
0.78369140625,
0.41845703125,
0.78369140625,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.2181396484375,
0.434326171875,
0.78515625,
0.434326171875,
0.78515625,
0.447509765625,
0.2181396484375,
0.447509765625
],
[
0.2181396484375,
0.448974609375,
0.7861328125,
0.448974609375,
0.7861328125,
0.462158203125,
0.2181396484375,
0.462158203125
],
[
0.21728515625,
0.46337890625,
0.78515625,
0.46337890625,
0.78515625,
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,
0.78369140625,
0.494384765625,
0.78369140625,
0.50830078125,
0.21728515625,
0.50830078125
],
[
0.2181396484375,
0.5087890625,
0.78759765625,
0.5087890625,
0.78759765625,
0.52294921875,
0.2181396484375,
0.52294921875
],
[
0.21728515625,
0.5244140625,
0.78369140625,
0.5244140625,
0.78369140625,
0.53857421875,
0.21728515625,
0.53857421875
],
[
0.2181396484375,
0.5537109375,
0.78759765625,
0.5537109375,
0.78759765625,
0.568359375,
0.2181396484375,
0.568359375
],
[
0.2181396484375,
0.5693359375,
0.78369140625,
0.5693359375,
0.78369140625,
0.58349609375,
0.2181396484375,
0.58349609375
],
[
0.2197265625,
0.58544921875,
0.78369140625,
0.58544921875,
0.78369140625,
0.5986328125,
0.2197265625,
0.5986328125
],
[
0.21728515625,
0.599609375,
0.78515625,
0.599609375,
0.78515625,
0.61279296875,
0.21728515625,
0.61279296875
],
[
0.2181396484375,
0.61572265625,
0.7861328125,
0.61572265625,
0.7861328125,
0.62890625,
0.2181396484375,
0.62890625
],
[
0.2181396484375,
0.6298828125,
0.7861328125,
0.6298828125,
0.7861328125,
0.64453125,
0.2181396484375,
0.64453125
],
[
0.21728515625,
0.64453125,
0.78369140625,
0.64453125,
0.78369140625,
0.65771484375,
0.21728515625,
0.65771484375
],
[
0.2181396484375,
0.66015625,
0.78369140625,
0.66015625,
0.78369140625,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.21728515625,
0.6748046875,
0.7861328125,
0.6748046875,
0.7861328125,
0.689453125,
0.21728515625,
0.689453125
],
[
0.2181396484375,
0.68994140625,
0.7861328125,
0.68994140625,
0.7861328125,
0.7041015625,
0.2181396484375,
0.7041015625
],
[
0.2181396484375,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.21728515625,
0.73486328125,
0.78369140625,
0.73486328125,
0.78369140625,
0.74853515625,
0.21728515625,
0.74853515625
],
[
0.2181396484375,
0.75048828125,
0.78369140625,
0.75048828125,
0.78369140625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.21728515625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.21728515625,
0.77978515625
],
[
0.2181396484375,
0.79541015625,
0.78369140625,
0.79541015625,
0.78369140625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78369140625,
0.8125,
0.78369140625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.21728515625,
0.82568359375,
0.78515625,
0.8271484375,
0.78515625,
0.84130859375,
0.21728515625,
0.84033203125
]
] |
[
"Info3D",
"3",
"- We show how local chunks of 3D objects can be used to get very effective",
"representations for downstream tasks.",
"- We demonstrate the effectiveness of the method on rotated inputs and show",
"how it is insensitive to such rotations in SO(3)group.",
"We conduct several experiments to show the efficacy of our method and",
"achieve state of art results in transfer learning, clustering and semi-supervised",
"learning.",
"2",
" Background",
"Representation Learning on 3D objects. Much progress has been made",
"to learn good representations of 3D objects in an unsupervised manner which",
"can then be used in several downstream tasks. In point clouds, works such as",
"[1], [49], [16], [18], [36] and [52] have been proposed. For voxels and implicit",
"representations, work such as [46], [26], [10], [27], [30] create powerful repre-",
"sentation features which are used for several downstream tasks such as shape",
"completion, shape generation and classification. Recently, there has been a lot",
"of progress in auto-encoding meshes such as in [40], [11], [40]. One disadvan-",
"tage of the above approaches are that they require you to reconstruct or gen-",
"erate the 3D shapes. As stated earlier, it might be expensive or not possible",
"to reconstruct the shape. A recent approach [51] does not require to recon-",
"struct the shape and instead does two stages of training. It first uses parts",
"of a shape to learn features using contrastive learning and then uses pseudo-",
"clusters to cluster all the data. Our method does not require two stages of",
"training, and furthermore allows us to create rotation invariant embeddings.",
"Maximizing Mutual Information and Contrastive Learning. A lot of",
"methods have used mutual information to do unsupervised learning. Historically",
"works such as [24], [4], [6], [5] have explored the InfoMax principle and mutual in-",
"formation maximization. More recently [29l proposed the Contrastive Predicting",
"Coding framework which uses the embeddings to capture maximal information",
"about future samples. The Deep InfoMax (DIM) [43] approach is similar but has",
"the advantage of doing orderless autoregression. In concurrent to these works.",
"works such as [48] and [53], have extended these ideas from the metric learn-",
"ing point of view. These methods were extended in [3], [19], [42] by considering",
"multiple views of the data and achieved state of the art representation learning",
"on images. The InfoMax principle has been extended to graphs [43], [39] and",
"for state representation in reinforcement learning [2]. Our method is inspired",
"by these methods and we extend them to 3D representations. For the multi-",
"ple views, we use geometric transformations and chunks of a 3D object. Finally.",
"in concurrent to our work several news works such as [9], [28] have been proposed.",
"Rotation Invariance on 3D objects. Traditionally, several methods have fo-",
"cused on hand engineered features to get local rotation invariant descriptors.",
"such as [34], [37], [22]. More recently, methods such as [13] and [14], first, get"
] |
{
"arxiv_id": "2006.02598",
"language": "en",
"page_idx": 2
}
|
||
2006.02598-en-3
|
2006.02598-en-3.jpg
|
4
A. Sanghi
local invariant features by encoding local geometry into patches and then use autoencoder to reconstruct the local features. These approaches require creating hand-crafted local features and need normal information. Methods such as MVCNN [38], Rot-SO-Net [23] and [35] explicitly force invariance by taking multiple poses of the object and aggregating them over the poses. However, such methods only work on discrete rotations or can only reconstruct objects rotated along one axis. A lot of deep learning methods have also attempted to use equivariance based architectures to achieve local and global rotation equivariance in 2D and 3D data. Methods such as [12],[45],[15],[41] either use constrained filters that achieve rotation equivariance or use filter orbit which are themselves equivariant. It is usually difficult to create such architectures for different 3D representations. Furthermore, to generate invariant representation to rotation from equivariant representation usually requires a post-processing step. Our method uses the InfoMax loss with rotation transformation to enforce rotation invariance. This method can be easily extended to voxels, implicit representation and meshes without needing to create complex architectures.
3
<FigureHere>
Fig.2: Overview of the method. A 3D object and different view of that object is encoded using the same encoder. The features across these views are made similar while a memory bank is used for negative examples. We also store the features obtained in the memory bank.
Our goal is to extract good features from 3D shapes in an unsupervised manner. To achieve this, we maximize the mutual information on the features extracted from a 3D shape and a different view of the 3D shape. We consider two such views in this work for achieving different purposes. First, we consider using a“chunk” of a 3D object to improve the representation of aligned shapes. The goal of using chunks of a 3D object is to incorporate some form of locality and structure in input into the objective. We investigate different ways to extract a
|
[
[
0.21728515625,
0.11614990234375,
0.2303466796875,
0.11614990234375,
0.2303466796875,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.27294921875,
0.117431640625,
0.339111328125,
0.117431640625,
0.339111328125,
0.1287841796875,
0.27294921875,
0.1287841796875
],
[
0.21728515625,
0.1483154296875,
0.78759765625,
0.1483154296875,
0.78759765625,
0.1622314453125,
0.21728515625,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.21728515625,
0.1767578125,
0.7861328125,
0.177978515625,
0.7861328125,
0.1925048828125,
0.21728515625,
0.1912841796875
],
[
0.21728515625,
0.1925048828125,
0.78369140625,
0.1932373046875,
0.78369140625,
0.2083740234375,
0.21728515625,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.7861328125,
0.208984375,
0.7861328125,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2197265625,
0.2259521484375,
0.7861328125,
0.2259521484375,
0.7861328125,
0.236083984375,
0.2197265625,
0.236083984375
],
[
0.21728515625,
0.2392578125,
0.78515625,
0.2392578125,
0.78515625,
0.25244140625,
0.21728515625,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.7861328125,
0.25390625,
0.7861328125,
0.268310546875,
0.2181396484375,
0.268310546875
],
[
0.2181396484375,
0.26953125,
0.78759765625,
0.26953125,
0.78759765625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2181396484375,
0.2841796875,
0.78759765625,
0.2841796875,
0.78759765625,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.21728515625,
0.298583984375,
0.78369140625,
0.2998046875,
0.78369140625,
0.314453125,
0.21728515625,
0.313232421875
],
[
0.2181396484375,
0.31494140625,
0.78515625,
0.31494140625,
0.78515625,
0.328857421875,
0.2181396484375,
0.328857421875
],
[
0.2181396484375,
0.330078125,
0.7861328125,
0.330078125,
0.7861328125,
0.3447265625,
0.2181396484375,
0.3447265625
],
[
0.2181396484375,
0.3447265625,
0.78369140625,
0.3447265625,
0.78369140625,
0.35791015625,
0.2181396484375,
0.35791015625
],
[
0.2181396484375,
0.35986328125,
0.7861328125,
0.35986328125,
0.7861328125,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.21728515625,
0.373779296875,
0.62353515625,
0.375732421875,
0.62353515625,
0.389404296875,
0.21728515625,
0.386962890625
],
[
0.2181396484375,
0.41357421875,
0.23779296875,
0.41357421875,
0.23779296875,
0.425537109375,
0.2181396484375,
0.425537109375
],
[
0.23779296875,
0.40771484375,
0.378173828125,
0.410888671875,
0.378173828125,
0.429931640625,
0.23779296875,
0.4267578125
],
[
0.2344970703125,
0.416748046875,
0.2442626953125,
0.416748046875,
0.2442626953125,
0.421630859375,
0.2344970703125,
0.421630859375
],
[
0.72216796875,
0.46337890625,
0.76318359375,
0.466552734375,
0.7607421875,
0.492431640625,
0.71826171875,
0.4892578125
],
[
0.279296875,
0.47607421875,
0.60302734375,
0.47607421875,
0.60302734375,
0.52587890625,
0.279296875,
0.52587890625
],
[
0.58251953125,
0.490478515625,
0.6005859375,
0.490478515625,
0.6005859375,
0.51025390625,
0.58251953125,
0.51025390625
],
[
0.325927734375,
0.556640625,
0.6708984375,
0.556640625,
0.6708984375,
0.60693359375,
0.325927734375,
0.60693359375
],
[
0.2181396484375,
0.64453125,
0.7861328125,
0.64599609375,
0.7861328125,
0.66015625,
0.2181396484375,
0.6591796875
],
[
0.21728515625,
0.6611328125,
0.78515625,
0.6611328125,
0.78515625,
0.6748046875,
0.21728515625,
0.6748046875
],
[
0.2181396484375,
0.6767578125,
0.7861328125,
0.6767578125,
0.7861328125,
0.68994140625,
0.2181396484375,
0.68994140625
],
[
0.2181396484375,
0.69140625,
0.496826171875,
0.69140625,
0.496826171875,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2442626953125,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2442626953125,
0.75
],
[
0.21728515625,
0.75,
0.7861328125,
0.75048828125,
0.7861328125,
0.76513671875,
0.21728515625,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.7822265625,
0.7861328125,
0.7822265625,
0.7861328125,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78759765625,
0.8125,
0.78759765625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"A. Sanghi",
"local invariant features by encoding local geometry into patches and then use",
"autoencoder to reconstruct the local features. These approaches require creat-",
"ing hand-crafted local features and need normal information. Methods such as",
"MVCNN [38], Rot-SO-Net [23] and [35] explicitly force invariance by taking mul-",
"tiple poses of the object and aggregating them over the poses. However, such",
"methodsonlyworkon discreterotationsorcan onlyreconstructobjectsrotated",
"along one axis. A lot of deep learning methods have also attempted to use equiv-",
"ariance based architectures to achieve local and global rotation equivariance in",
"2D and 3D data.Methods such as[12],[45],[15],[41] either use constrained",
"filters that achieve rotation equivariance or use filter orbit which are themselves",
"equivariant. It is usually difficult to create such architectures for different 3D rep-",
"resentations. Furthermore, to generate invariant representation to rotation from",
"equivariant representation usually requires a post-processing step. Our method",
"uses the InfoMax loss with rotation transformation to enforce rotation invari-",
"ance. This method can be easily extended to voxels, implicit representation and",
"meshes withoutneedingto createcomplexarchitectures.",
"3 ",
"Methodology",
"",
"Manay",
"",
"",
"",
"Fig. 2: Overview of the method. A 3D object and different view of that object",
"is encoded using the same encoder. The features across these views are made",
"similar while a memory bank is used for negative examples. We also store the",
"features obtained in the memory bank.",
"Our goal is to extract good features from 3D shapes in an unsupervised",
"manner. To achieve this, we maximize the mutual information on the features",
"extracted from a 3D shape and a different view of the 3D shape. We consider two",
"such views in this work for achieving different purposes. First, we consider using",
"a “chunk\" of a 3D object to improve the representation of aligned shapes. The",
"goal of using chunks of a 3D object is to incorporate some form of locality and",
"structure in input into the objective. We investigate different ways to extract a"
] |
{
"arxiv_id": "2006.02598",
"language": "en",
"page_idx": 3
}
|
||
2006.02598-en-4
|
2006.02598-en-4.jpg
|
5
$Info3D$
chunk of a 3D object and show how this can improve the representations learned. Next, we consider geometric transformation as the second view. Intuitively, the goal is to make the shape and its geometric transformed version closer in the latent space while distancing itself from other objects. We show how this can create transformation invariant embeddings to a large extent. We discuss this method in the context of point clouds but it should be straightforward to extend it to other 3D representations. The method in a pictorial form is shown in Figure 2.
Let $x_i$ and $T(x_i)$ represent a 3D shape and the object obtained after applying the transformation as mentioned above. We use an encoder, $f(.)$, to transform $x_i$ and $T(x_i)$ into the latent representation $z_i^a$ and $z_i^b.$ Note that this encoder also includes a final normalizing layer which makes the embeddings unit vectors. In the next sections, we detail the motivation and mechanism for using chunks of a 3D object and geometric transformations as second view.
3.1 Local Chunks
<FigureHere>
Fig.3: Types of chunks. The first object represents a sample 3D object. The rest of the objects are chunks obtained from that object. Note, for the last three chunks we use the same random point. Despite using the same random point, very different chunks can be obtained from different distance metrics.
As mentioned above, the chunks provide a way to define a locality structure to be incorporated in the objective. We do this by first defning a local subset of the 3D object. For point clouds, if you randomly select a subset of points, you will just get a coarser representation of the 3D object. So, we investigate some potential ways to define local sub-structures in point clouds. Next, we force the network to distinguish between its own local subset and other objects' local subsets using the InfoMax principle. This objective forces the network to learn about its own local sub-structures and creates more informative embeddings.
We define chunks by considering two mechanisms. In the frst approach, we randomly select a point from the point cloud. Then, we use a distance measure in Euclidean space to select a subset of points. The distance measure we consider are Euclidean distance, cosine distance and Chebyshev distance. Once we select a subset of points, we normalize them using a bounding sphere. In the second approach, we take a chunk of a 3D object based on chopping the object randomly
|
[
[
0.6845703125,
0.11553955078125,
0.734375,
0.11553955078125,
0.734375,
0.1300048828125,
0.6845703125,
0.1300048828125
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.77392578125,
0.1287841796875
],
[
0.2197265625,
0.147705078125,
0.78515625,
0.147705078125,
0.78515625,
0.1622314453125,
0.2197265625,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.2181396484375,
0.1787109375,
0.7861328125,
0.1787109375,
0.7861328125,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.2181396484375,
0.1944580078125,
0.7861328125,
0.1944580078125,
0.7861328125,
0.2083740234375,
0.2181396484375,
0.2083740234375
],
[
0.2181396484375,
0.208984375,
0.7861328125,
0.208984375,
0.7861328125,
0.2235107421875,
0.2181396484375,
0.2235107421875
],
[
0.2181396484375,
0.2235107421875,
0.78759765625,
0.2235107421875,
0.78759765625,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.2156982421875,
0.238037109375,
0.7861328125,
0.2392578125,
0.7861328125,
0.25390625,
0.2156982421875,
0.25244140625
],
[
0.2197265625,
0.255615234375,
0.2320556640625,
0.255615234375,
0.2320556640625,
0.266357421875,
0.2197265625,
0.266357421875
],
[
0.24267578125,
0.26953125,
0.7861328125,
0.26953125,
0.7861328125,
0.2841796875,
0.24267578125,
0.2841796875
],
[
0.2181396484375,
0.285400390625,
0.7861328125,
0.285400390625,
0.7861328125,
0.298583984375,
0.2181396484375,
0.298583984375
],
[
0.2181396484375,
0.300537109375,
0.7861328125,
0.300537109375,
0.7861328125,
0.314453125,
0.2181396484375,
0.314453125
],
[
0.21728515625,
0.314453125,
0.78515625,
0.31494140625,
0.78515625,
0.330078125,
0.21728515625,
0.328857421875
],
[
0.2181396484375,
0.330810546875,
0.78759765625,
0.330810546875,
0.78759765625,
0.3447265625,
0.2181396484375,
0.3447265625
],
[
0.2181396484375,
0.3466796875,
0.66015625,
0.3466796875,
0.66015625,
0.35986328125,
0.2181396484375,
0.35986328125
],
[
0.2197265625,
0.38818359375,
0.25244140625,
0.38818359375,
0.25244140625,
0.399658203125,
0.2197265625,
0.399658203125
],
[
0.250732421875,
0.391357421875,
0.2607421875,
0.391357421875,
0.2607421875,
0.396484375,
0.250732421875,
0.396484375
],
[
0.256591796875,
0.386962890625,
0.374267578125,
0.386962890625,
0.374267578125,
0.40087890625,
0.256591796875,
0.40087890625
],
[
0.38232421875,
0.43994140625,
0.42724609375,
0.442626953125,
0.426513671875,
0.453857421875,
0.381591796875,
0.451904296875
],
[
0.4755859375,
0.43994140625,
0.53125,
0.442626953125,
0.53125,
0.453857421875,
0.4755859375,
0.451904296875
],
[
0.580078125,
0.43994140625,
0.64453125,
0.442626953125,
0.64453125,
0.453857421875,
0.580078125,
0.451904296875
],
[
0.67333984375,
0.442626953125,
0.74169921875,
0.442626953125,
0.74169921875,
0.45263671875,
0.67333984375,
0.45263671875
],
[
0.583984375,
0.468505859375,
0.63134765625,
0.47021484375,
0.6298828125,
0.490478515625,
0.58251953125,
0.488037109375
],
[
0.685546875,
0.47021484375,
0.72802734375,
0.47021484375,
0.72802734375,
0.486083984375,
0.685546875,
0.486083984375
],
[
0.47802734375,
0.472900390625,
0.51708984375,
0.472900390625,
0.51708984375,
0.48291015625,
0.47802734375,
0.48291015625
],
[
0.301513671875,
0.472900390625,
0.324462890625,
0.474853515625,
0.32275390625,
0.48291015625,
0.2998046875,
0.48095703125
],
[
0.39453125,
0.4765625,
0.406982421875,
0.4765625,
0.406982421875,
0.48095703125,
0.39453125,
0.48095703125
],
[
0.2181396484375,
0.5341796875,
0.78515625,
0.5341796875,
0.78515625,
0.54736328125,
0.2181396484375,
0.54736328125
],
[
0.2181396484375,
0.5498046875,
0.78515625,
0.5498046875,
0.78515625,
0.56298828125,
0.2181396484375,
0.56298828125
],
[
0.21728515625,
0.56298828125,
0.78515625,
0.564453125,
0.78515625,
0.5791015625,
0.21728515625,
0.57763671875
],
[
0.2181396484375,
0.580078125,
0.7158203125,
0.580078125,
0.7158203125,
0.59326171875,
0.2181396484375,
0.59326171875
],
[
0.2442626953125,
0.62890625,
0.7861328125,
0.62890625,
0.7861328125,
0.64306640625,
0.2442626953125,
0.64306640625
],
[
0.2181396484375,
0.64599609375,
0.78515625,
0.64599609375,
0.78515625,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.21728515625,
0.65771484375,
0.78515625,
0.6591796875,
0.78515625,
0.673828125,
0.21728515625,
0.67236328125
],
[
0.21728515625,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.689453125,
0.21728515625,
0.689453125
],
[
0.2181396484375,
0.68994140625,
0.78515625,
0.68994140625,
0.78515625,
0.7041015625,
0.2181396484375,
0.7041015625
],
[
0.2181396484375,
0.705078125,
0.78759765625,
0.705078125,
0.78759765625,
0.71826171875,
0.2181396484375,
0.71826171875
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2197265625,
0.7373046875,
0.765625,
0.7373046875,
0.765625,
0.74755859375,
0.2197265625,
0.74755859375
],
[
0.2442626953125,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2442626953125,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.78076171875,
0.7861328125,
0.78076171875,
0.7861328125,
0.79541015625,
0.21728515625,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78759765625,
0.8125,
0.78759765625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.21728515625,
0.82763671875,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.21728515625,
0.84130859375
]
] |
[
"Info3D",
"5",
"chunk of a 3D object and show how this can improve the representations learned.",
"Next, we consider geometric transformation as the second view. Intuitively, the",
"goal is to make the shape and its geometric transformed version closer in the",
"latent space while distancing itself from other objects. We show how this can",
"create transformation invariant embeddings to a large extent. We discuss this",
"method in the context of point clouds but it should be straightforward to extend",
"it to other 3D representations. The method in a pictorial form is shown in Figure",
"2.",
"Let r, and T(ci) represent a 3D shape and the object obtained after applying",
"the transformation as mentioned above. We use an encoder, f(), to transform",
"ri and T(c) into the latent representation z and z. Note that this encoder",
"also includes a final normalizing layer which makes the embeddings unit vectors.",
"In the next sections, we detail the motivation and mechanism for using chunks",
"of a 3D object and geometric transformations as second view.",
"3.1",
"",
"Local Chunks",
"Axis choppe",
"Cosine distance",
"Euclidean distance",
"Chebyshev distance",
"",
"",
"",
"",
"",
"Fig.3: Types of chunks. The first object represents a sample 3D object. The",
"rest of the objects are chunks obtained from that object. Note, for the last three",
"chunks we use the same random point. Despite using the same random point,",
"very different chunks can be obtained from different distance metrics.",
"As mentioned above, the chunks provide a way to define a locality structure",
"to be incorporated in the objective. We do this by first defining a local subset",
"of the 3D object. For point clouds, if you randomly select a subset of points,",
"you will just get a coarser representation of the 3D object. So, we investigate",
"some potential ways to define local sub-structures in point clouds. Next, we force",
"the network to distinguish between its own local subset and other objects' local",
"subsets using the InfoMax principle. This objective forces the network to learn",
"aboutits own local sub-structures and createsmoreinformative embeddings",
"We define chunks by considering two mechanisms. In the first approach, we",
"randomly select a point from the point cloud. Then, we use a distance measure",
"in Euclidean space to select a subset of points. The distance measure we consider",
"are Euclidean distance, cosine distance and Chebyshev distance. Once we select",
"a subset of points, we normalize them using a bounding sphere. In the second",
"approach,we takea chunk ofa3Dobjectbased onchopping theobject randomly"
] |
{
"arxiv_id": "2006.02598",
"language": "en",
"page_idx": 4
}
|
||
2311.00273-en-2
|
2311.00273-en-2.jpg
|
2 Human-centered Mental Health LLM
2.1 SoulChatCorpus Collection
We consider an one-on-one psychological counseling conversational setting where a user and a psychological consultant engage in multiple rounds of dialogue. However, such conversation data is
To this end, we first constructed 215,813 different psychological counseling questions about 12 topics and 619,725 answers through data outsourcing services. Rule-based cleaning, manual rewriting and human proofreading are applied to ensure that there is no sensitive or privacy-related content in the dataset. Then, we use ChatGPT to convert these single-turn long text psychological counseling conversations to multi-turn empathy conversations. We also conducted manual proofreading and data cleansing for multi-turn dialogues rewritten by ChatGPT to further strengthen the expression of empathy, such as questioning, comfort, recognition, listening, trust, emotional support, etc. In the end, we obtained a multi-turn empathy conversation dataset, named SoulChatCorpus, with 2,300,248 samples. To our knowledge, it is the first millionscale multi-turn empathy conversation dataset in the field of mental health or emotional support. We conduct experiments by using ChatGLM-6B as the base model for fine-tuning on SoulChatCorpus. Results demonstrate that LLMs' empathy, listening, and comfort abilities can be improved significantly through fine-tuning with million-scale multi-turn empathy conversation dataset.
<FigureHere>
Figure 3: Distribution of counseling topics.
not publicly available due to the privacy protection and ethical standards of psychological counseling. To construct high-quality multi-turn empathy conversation dataset, We selected 12 topics of psychological counseling to construct 215,813 long-text questions and 619,725 long-text answer through crowdsourcing. The distribution of topics is shown in Figure 3. Then, we used ChatGPT (99% called gpt-3.5-turbo api and 1% called gpt-4 api) as a text rewriting tool following the prompt as shown in Figure 2 to convert single-turn psychological counseling conversations to multi-turn empathy conversations, in which one turn is in the form of "用户：<user\_utt>\n心理咨询师：<psy\_utt>". The response of "心理咨询师" was asked to be rewritten to reflect human-centered expressions such as empathy, listening, comfort, etc. Finally, after manual proofreading, we removed 105,134 low-quality samples and ultimately obtained 2,300,248 samples. As shown in Figure 4, the word cloud map of the utterances expressed by psychological consultants indicated that the rewritten multi-turn empathy conversation has high level of empathy.
<FigureHere>
Figure 4: Word cloud map of psychological consultants'
utterances (English version: Appendix D).
### 2.2 SoulChat Model
We utilized the ChatGLM-6B (Du et al., 2022; Zeng et al., 2023) as the base LLM architecture to develop the SoulChat. ChatGLM-6B is an opensource, bilingual LLM based on the General Language Model (GLM) (Du et al., 2022) framework with 6.2 billion parameters. The input of model is defined as:
$$input=u_1^u+'\:\backslash n'+u_1^p+...+u_N^u+'\:\backslash n'+u_N^p$$
where the utterance of User on $i$ turn $u_i^u=`$用户： (User:)' + utterance$_i^u$, utterance of Psychologist on $i$ turn $u_i^p=`$心理咨询师：(Psychologist:)'+ utterance$_i^p\left(i<N\right),u_N^p=`$心理咨询师：(Psychol-
|
[
[
0.1368408203125,
0.08612060546875,
0.497802734375,
0.08612060546875,
0.497802734375,
0.09979248046875,
0.1368408203125,
0.09979248046875
],
[
0.509765625,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.1158447265625,
0.1015625,
0.489501953125,
0.1009521484375,
0.489501953125,
0.11517333984375,
0.1158447265625,
0.11639404296875
],
[
0.509765625,
0.10272216796875,
0.8818359375,
0.10272216796875,
0.8818359375,
0.11639404296875,
0.509765625,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.491943359375,
0.1181640625,
0.491943359375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.509765625,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1158447265625,
0.1341552734375,
0.489501953125,
0.133056640625,
0.489501953125,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.509765625,
0.1353759765625,
0.87890625,
0.1353759765625,
0.87890625,
0.1455078125,
0.509765625,
0.1455078125
],
[
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
],
[
0.1175537109375,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51318359375,
0.166259765625,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1158447265625,
0.18115234375,
0.486083984375,
0.18115234375,
0.486083984375,
0.19482421875,
0.1158447265625,
0.19482421875
],
[
0.51123046875,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.1158447265625,
0.198974609375,
0.496337890625,
0.198974609375,
0.496337890625,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.5087890625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.5087890625,
0.2120361328125
],
[
0.1158447265625,
0.215576171875,
0.49365234375,
0.215576171875,
0.49365234375,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.51318359375,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.2286376953125,
0.51318359375,
0.2286376953125
],
[
0.11419677734375,
0.23095703125,
0.486083984375,
0.23095703125,
0.486083984375,
0.24462890625,
0.11419677734375,
0.24462890625
],
[
0.509765625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.509765625,
0.24462890625
],
[
0.1175537109375,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.51123046875,
0.2476806640625,
0.88330078125,
0.2476806640625,
0.88330078125,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.11419677734375,
0.262451171875,
0.48779296875,
0.262451171875,
0.48779296875,
0.275634765625,
0.11419677734375,
0.275634765625
],
[
0.50732421875,
0.26318359375,
0.8818359375,
0.262451171875,
0.8818359375,
0.276611328125,
0.50732421875,
0.27783203125
],
[
0.1158447265625,
0.279052734375,
0.489501953125,
0.279052734375,
0.489501953125,
0.292236328125,
0.1158447265625,
0.292236328125
],
[
0.509765625,
0.27783203125,
0.87890625,
0.27783203125,
0.87890625,
0.29150390625,
0.509765625,
0.29150390625
],
[
0.1175537109375,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51123046875,
0.29345703125,
0.8818359375,
0.29443359375,
0.8818359375,
0.308837890625,
0.51123046875,
0.30810546875
],
[
0.1158447265625,
0.31005859375,
0.48779296875,
0.311279296875,
0.48779296875,
0.325927734375,
0.1158447265625,
0.32470703125
],
[
0.50732421875,
0.308837890625,
0.88330078125,
0.30810546875,
0.88330078125,
0.325927734375,
0.50732421875,
0.32666015625
],
[
0.1158447265625,
0.325927734375,
0.486083984375,
0.325927734375,
0.486083984375,
0.339111328125,
0.1158447265625,
0.339111328125
],
[
0.51123046875,
0.32666015625,
0.8818359375,
0.32666015625,
0.8818359375,
0.34033203125,
0.51123046875,
0.34033203125
],
[
0.1175537109375,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.51318359375,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.1175537109375,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.51123046875,
0.359375,
0.8818359375,
0.359375,
0.8818359375,
0.37353515625,
0.51123046875,
0.37353515625
],
[
0.1175537109375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.509765625,
0.374755859375,
0.8818359375,
0.37353515625,
0.8818359375,
0.388427734375,
0.509765625,
0.3896484375
],
[
0.1158447265625,
0.391357421875,
0.48779296875,
0.39013671875,
0.48779296875,
0.405029296875,
0.1158447265625,
0.40625
],
[
0.51123046875,
0.392578125,
0.8818359375,
0.39013671875,
0.8818359375,
0.40380859375,
0.51123046875,
0.40625
],
[
0.1158447265625,
0.407958984375,
0.489501953125,
0.407958984375,
0.489501953125,
0.421630859375,
0.1158447265625,
0.421630859375
],
[
0.51318359375,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.51318359375,
0.421630859375
],
[
0.1175537109375,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.509765625,
0.42333984375,
0.88330078125,
0.424560546875,
0.88330078125,
0.438232421875,
0.509765625,
0.43701171875
],
[
0.11419677734375,
0.440673828125,
0.484375,
0.440673828125,
0.484375,
0.45361328125,
0.11419677734375,
0.45361328125
],
[
0.51123046875,
0.440673828125,
0.82958984375,
0.440673828125,
0.82958984375,
0.45361328125,
0.51123046875,
0.45361328125
],
[
0.1175537109375,
0.457275390625,
0.486083984375,
0.457275390625,
0.486083984375,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.11419677734375,
0.47265625,
0.337646484375,
0.47265625,
0.337646484375,
0.486328125,
0.11419677734375,
0.486328125
],
[
0.62744140625,
0.47265625,
0.70947265625,
0.47265625,
0.70947265625,
0.482666015625,
0.62744140625,
0.482666015625
],
[
0.70361328125,
0.47265625,
0.755859375,
0.474365234375,
0.755859375,
0.482177734375,
0.70361328125,
0.479736328125
],
[
0.791015625,
0.48388671875,
0.8037109375,
0.48388671875,
0.8037109375,
0.48681640625,
0.791015625,
0.48681640625
],
[
0.8017578125,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.4921875,
0.8017578125,
0.4921875
],
[
0.75830078125,
0.490478515625,
0.86572265625,
0.4921875,
0.86572265625,
0.51025390625,
0.75830078125,
0.50830078125
],
[
0.5263671875,
0.495361328125,
0.765625,
0.474365234375,
0.77490234375,
0.52099609375,
0.53564453125,
0.54150390625
],
[
0.312255859375,
0.5029296875,
0.45166015625,
0.5029296875,
0.45166015625,
0.5166015625,
0.312255859375,
0.5166015625
],
[
0.32421875,
0.5185546875,
0.399658203125,
0.5185546875,
0.399658203125,
0.52978515625,
0.32421875,
0.52978515625
],
[
0.81005859375,
0.525390625,
0.84228515625,
0.5224609375,
0.84619140625,
0.54052734375,
0.8154296875,
0.54443359375
],
[
0.2216796875,
0.52685546875,
0.26025390625,
0.52685546875,
0.26025390625,
0.5380859375,
0.2216796875,
0.5380859375
],
[
0.5283203125,
0.485107421875,
0.5615234375,
0.482666015625,
0.5634765625,
0.495361328125,
0.52880859375,
0.49755859375
],
[
0.361083984375,
0.5322265625,
0.4189453125,
0.5322265625,
0.4189453125,
0.5419921875,
0.361083984375,
0.5419921875
],
[
0.74658203125,
0.52783203125,
0.81689453125,
0.52783203125,
0.81689453125,
0.55029296875,
0.74658203125,
0.55029296875
],
[
0.8095703125,
0.5322265625,
0.8154296875,
0.5322265625,
0.8154296875,
0.54052734375,
0.8095703125,
0.54052734375
],
[
0.307373046875,
0.54345703125,
0.328369140625,
0.54345703125,
0.328369140625,
0.5517578125,
0.307373046875,
0.5517578125
],
[
0.60791015625,
0.5380859375,
0.6396484375,
0.5380859375,
0.6396484375,
0.5498046875,
0.60791015625,
0.5498046875
],
[
0.25439453125,
0.5498046875,
0.287109375,
0.5498046875,
0.287109375,
0.56103515625,
0.25439453125,
0.56103515625
],
[
0.333251953125,
0.552734375,
0.364501953125,
0.552734375,
0.364501953125,
0.56396484375,
0.333251953125,
0.56396484375
],
[
0.7001953125,
0.552734375,
0.7109375,
0.552734375,
0.7109375,
0.55810546875,
0.7001953125,
0.55810546875
],
[
0.69189453125,
0.56005859375,
0.74072265625,
0.56005859375,
0.74072265625,
0.57177734375,
0.69189453125,
0.57177734375
],
[
0.74462890625,
0.55810546875,
0.80029296875,
0.55810546875,
0.80029296875,
0.5693359375,
0.74462890625,
0.5693359375
],
[
0.60791015625,
0.56396484375,
0.6396484375,
0.56396484375,
0.6396484375,
0.572265625,
0.60791015625,
0.572265625
],
[
0.424072265625,
0.572265625,
0.474365234375,
0.572265625,
0.474365234375,
0.5830078125,
0.424072265625,
0.5830078125
],
[
0.66650390625,
0.57373046875,
0.765625,
0.57763671875,
0.76416015625,
0.5927734375,
0.6650390625,
0.58837890625
],
[
0.771484375,
0.57763671875,
0.7958984375,
0.57763671875,
0.7958984375,
0.5859375,
0.771484375,
0.5859375
],
[
0.599609375,
0.580078125,
0.7021484375,
0.580078125,
0.7021484375,
0.6015625,
0.599609375,
0.6015625
],
[
0.1234130859375,
0.576171875,
0.17626953125,
0.57763671875,
0.175537109375,
0.5927734375,
0.12176513671875,
0.59033203125
],
[
0.3779296875,
0.58447265625,
0.414794921875,
0.58447265625,
0.414794921875,
0.595703125,
0.3779296875,
0.595703125
],
[
0.5322265625,
0.58447265625,
0.5576171875,
0.58447265625,
0.5576171875,
0.5927734375,
0.5322265625,
0.5927734375
],
[
0.60546875,
0.59326171875,
0.623046875,
0.59326171875,
0.623046875,
0.5966796875,
0.60546875,
0.5966796875
],
[
0.765625,
0.59326171875,
0.85205078125,
0.59326171875,
0.85205078125,
0.60107421875,
0.765625,
0.60107421875
],
[
0.51318359375,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.187255859375,
0.58837890625,
0.22412109375,
0.58837890625,
0.22412109375,
0.60107421875,
0.187255859375,
0.60107421875
],
[
0.177978515625,
0.6357421875,
0.214111328125,
0.6357421875,
0.214111328125,
0.64794921875,
0.177978515625,
0.64794921875
],
[
0.3896484375,
0.6357421875,
0.435791015625,
0.6357421875,
0.435791015625,
0.65087890625,
0.3896484375,
0.65087890625
],
[
0.51318359375,
0.6328125,
0.796875,
0.6328125,
0.796875,
0.646484375,
0.51318359375,
0.646484375
],
[
0.139404296875,
0.64013671875,
0.1695556640625,
0.64013671875,
0.1695556640625,
0.65185546875,
0.139404296875,
0.65185546875
],
[
0.4375,
0.6435546875,
0.49365234375,
0.6435546875,
0.49365234375,
0.65478515625,
0.4375,
0.65478515625
],
[
0.193115234375,
0.6591796875,
0.2216796875,
0.6591796875,
0.2216796875,
0.66748046875,
0.193115234375,
0.66748046875
],
[
0.1318359375,
0.66748046875,
0.181396484375,
0.66748046875,
0.181396484375,
0.681640625,
0.1318359375,
0.681640625
],
[
0.35107421875,
0.681640625,
0.3896484375,
0.681640625,
0.3896484375,
0.693359375,
0.35107421875,
0.693359375
],
[
0.51318359375,
0.6767578125,
0.68603515625,
0.6767578125,
0.68603515625,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.225830078125,
0.68505859375,
0.26123046875,
0.68505859375,
0.26123046875,
0.69775390625,
0.225830078125,
0.69775390625
],
[
0.2880859375,
0.69873046875,
0.32080078125,
0.69873046875,
0.32080078125,
0.71044921875,
0.2880859375,
0.71044921875
],
[
0.3896484375,
0.70166015625,
0.47265625,
0.70166015625,
0.47265625,
0.71337890625,
0.3896484375,
0.71337890625
],
[
0.51318359375,
0.69775390625,
0.8818359375,
0.69775390625,
0.8818359375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.151123046875,
0.7060546875,
0.2225341796875,
0.7060546875,
0.2225341796875,
0.71728515625,
0.151123046875,
0.71728515625
],
[
0.51123046875,
0.71435546875,
0.8818359375,
0.71435546875,
0.8818359375,
0.72802734375,
0.51123046875,
0.72802734375
],
[
0.306396484375,
0.7255859375,
0.343505859375,
0.7255859375,
0.343505859375,
0.7353515625,
0.306396484375,
0.7353515625
],
[
0.509765625,
0.72998046875,
0.8818359375,
0.73095703125,
0.8818359375,
0.7451171875,
0.509765625,
0.74462890625
],
[
0.51123046875,
0.74658203125,
0.88330078125,
0.74658203125,
0.88330078125,
0.76123046875,
0.51123046875,
0.76123046875
],
[
0.1561279296875,
0.7666015625,
0.447509765625,
0.7666015625,
0.447509765625,
0.77978515625,
0.1561279296875,
0.77978515625
],
[
0.51123046875,
0.76318359375,
0.8818359375,
0.7626953125,
0.8818359375,
0.77685546875,
0.51123046875,
0.77783203125
],
[
0.51123046875,
0.779296875,
0.8818359375,
0.779296875,
0.8818359375,
0.79345703125,
0.51123046875,
0.79345703125
],
[
0.5146484375,
0.7958984375,
0.59375,
0.7958984375,
0.59375,
0.80712890625,
0.5146484375,
0.80712890625
],
[
0.118408203125,
0.81103515625,
0.1318359375,
0.81103515625,
0.1318359375,
0.8193359375,
0.118408203125,
0.8193359375
],
[
0.1351318359375,
0.8134765625,
0.1485595703125,
0.8134765625,
0.1485595703125,
0.81787109375,
0.1351318359375,
0.81787109375
],
[
0.1435546875,
0.80859375,
0.47265625,
0.80859375,
0.47265625,
0.822265625,
0.1435546875,
0.822265625
],
[
0.51318359375,
0.82470703125,
0.8759765625,
0.82470703125,
0.8759765625,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.1175537109375,
0.8349609375,
0.376220703125,
0.8349609375,
0.376220703125,
0.8486328125,
0.1175537109375,
0.8486328125
],
[
0.11419677734375,
0.85498046875,
0.490234375,
0.85546875,
0.490234375,
0.87060546875,
0.11419677734375,
0.869140625
],
[
0.509765625,
0.85546875,
0.86962890625,
0.85400390625,
0.86962890625,
0.8681640625,
0.509765625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1158447265625,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.509765625,
0.88720703125,
0.88330078125,
0.88720703125,
0.88330078125,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90380859375,
0.8818359375,
0.90380859375,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"To this end, we first constructed 215,813 differ-",
"not publicly available due to the privacy protection",
"ent psychological counseling questions about 12",
" and ethical standards of psychological counseling",
"topics and 619,725 answers through data outsourc-",
"To construct high-quality multi-turn empathy con",
"ing services. Rule-based cleaning, manual rewrit-",
"versationdatasetWeselected12topicsofpsvcho",
"ing and human proofreading are applied to ensure",
"logicalcounselingtoconstruct215.813long-text",
"that there is no sensitive or privacy-related content",
"questions and 619,725 long-text answer through",
"in the dataset. Then, we use ChatGPT to convert",
"crowdsourcing. The distribution of topics is shown",
"these single-turn long text psychological counsel-",
"in Figure 3. Then, we used ChatGPT (99% called",
"ing conversations to multi-turn empathy conversa-",
"gpt-3.5-turbo api and 1% called gpt-4 api) as a text",
"tions. We also conducted manual proofreading and",
"rewriting tool following the prompt as shown in",
"data cleansing for multi-turn dialogues rewritten",
"Figure 2 to convert single-turn psychological coun-",
"by ChatGPT to further strengthen the expression of",
"seling conversations to multi-turn empathy conver-",
"empathy, such as questioning, comfort, recognition,",
"sations, in which one turn is in the form of \"用",
"listening, trust, emotional support, etc. In the end,",
"户：<user_utt>n心理咨询师：<psy_utt>\".The",
"we obtained a multi-turn empathy conversation",
"response of\"心理咨询师\"was asked tobe rewrit-",
"dataset. named SoulChatCorpus, with 2.300.248",
"ten to reflect human-centered expressions such as",
"samples. To our knowledge, it is the first million-",
"empathy, listening, comfort, etc. Finally, after man",
"scale multi-turn empathy conversation dataset in",
"ual proofreading, we removed 105,134 low-quality",
"the field of mental health or emotional support. We",
" samples and ultimately obtained 2,300,248 sam-",
"conduct experiments by using ChatGLM-6B as the",
"ples. As shown in Figure 4, the word cloud map of",
"base model for fine-tuning on SoulChatCorpus. Re-",
"the utterances expressed by psychological consul-",
"sults demonstrate that LLMs' empathy, listening,",
"tants indicated that the rewritten multi-turn empa-",
" and comfort abilities can be improved significantly",
"thy conversation has high level of empathy.",
"through fine-tuning with million-scale multi-turn",
"empathy conversation dataset.",
"支持相信不",
"不是同时",
"",
"能出",
"而安",
"很会口以",
"Psychological knowledge",
"Social events",
"中",
"Family",
"都",
"Treatment",
"计不",
"",
"0.70",
"更好",
"9.7%",
"7.9%",
"5T",
"我到个",
"要",
"事情",
"Emotion",
"客这种明友此",
"人",
"帮助",
"Marriage",
"11.7%",
"奇",
"",
"好寻求想法口",
"Figure 4: Word cloud map of psychological consultants",
"16.3%",
"3.7%",
"11.5%",
"utterances (English version: Appendix D).",
"Sex",
"Behavior",
"5.8%",
"Growth",
"11.1%",
"2.2 SoulChat Model",
"10.0%",
"9.5%",
"Self-awareness",
"We utilized the ChatGLM-6B (Du et al., 2022;",
"Relationship",
"Zeng et al., 2023) as the base LLM architecture",
"Career",
"to develop the SoulChat. ChatGLM-6B is an open+",
"source, bilingual LLM based on the General Lan-",
"Figure 3: Distribution of counseling topics.",
"guage Model (GLM) (Du et al., 2022) framework",
"with 6.2 billion parameters. The input of model is",
"defined as:",
"2",
"",
"Human-centered Mental Health LLM",
"input=uu+'In'+up+...+uw +'n'+u",
"2.1 SoulChatCorpus Collection",
"We consider an one-on-one psychological counsel-",
"where the utteranceofUser on iturn w=用户：",
"ing conversational setting where a user and a psy-",
"(User:)' + utterance', utterance of Psychologist",
"chological consultant engage in multiple rounds",
"oniturn up-\"心理咨询师：(Psychologist:)’+",
"of dialogue. However, such conversation data is",
"utterancep(i<N),u=心理咨询师：(Psychol"
] |
{
"arxiv_id": "2311.00273",
"language": "en",
"page_idx": 2
}
|
||
2208.00183-en-0
|
2208.00183-en-0.jpg
|
# Few-shot Single-view 3D Reconstruction with Memory Prior Contrastive Network
Zhen Xing$^1$, Yijiang Chen$^1$,Zhixin Ling$^1$,
Xiangdong Zhou$^1$, and Yu Xiang$^2$
$^{1}$School of Computer Science, Fudan University, Shanghai, China, 200433
$^{2}$The University of Texas at Dallas
$\textbf{Abstract. 3D reconstruction of novel categories based on few- shot learn- }$ ing is appealing in real-world applications and attracts increasing research interests. Previous approaches mainly focus on how to design shape prior models for different categories. Their performance on unseen categories is not very competitive. In this paper, we present a Memory Prior Contrastive Network (MPCN) that can store shape prior knowledge in a few-shot learning based 3D reconstruction framework. With the shape memory, a multi-head attention module is proposed to capture different parts of a candidate shape prior and fuse these parts together to guide 3D reconstruction of novel categories. Besides, we introduce a 3D-aware contrastive learning method, which can not only complement the retrieval accuracy of memory network, but also better organize image features for downstream tasks. Compared with previous few-shot 3D reconstruction methods, MPCN can handle the inter-class variability without category annotations. Experimental results on a benchmark synthetic dataset and the Pascal3D+ real-world dataset show that our model outperforms the current state-of-the-art methods significantly Keywords: Few-shot learning, 3D Reconstruction, Memory Network
## 1 Introduction
Reconstructing 3D shapes from RGB images is valuable in many real-world applications such as autonomous driving, virtual reality, CAD and robotics. Traditional methods for 3D reconstruction such as Structure From Motion (SFM) [29] and Simultaneous Localization and Mapping (SLAM) [2] often require significant efforts in data acquisition. For example, a large number of images need to be captured and the camera parameters need to be calibrated, which limit the applications of these traditional methods.
In recent years, 3D reconstruction from single image based on deep neural networks attracts great research interests. However, previous methods of singleview 3D reconstruction are mostly category-specific [42,8]. Therefore, they only perform well in the specific training categories. These methods also require a large number of labeled training images, which is time consuming and costly to obtain.
|
[
[
0.238525390625,
0.1458740234375,
0.76318359375,
0.1458740234375,
0.76318359375,
0.160400390625,
0.238525390625,
0.160400390625
],
[
0.294189453125,
0.1685791015625,
0.7099609375,
0.1685791015625,
0.7099609375,
0.18310546875,
0.294189453125,
0.18310546875
],
[
0.351318359375,
0.212158203125,
0.65185546875,
0.212158203125,
0.65185546875,
0.2298583984375,
0.351318359375,
0.2298583984375
],
[
0.3798828125,
0.2298583984375,
0.62353515625,
0.2298583984375,
0.62353515625,
0.2442626953125,
0.3798828125,
0.2442626953125
],
[
0.250732421875,
0.25244140625,
0.75244140625,
0.255126953125,
0.75244140625,
0.272705078125,
0.250732421875,
0.270751953125
],
[
0.0228729248046875,
0.270751953125,
0.060455322265625,
0.270751953125,
0.060455322265625,
0.4072265625,
0.0228729248046875,
0.4072265625
],
[
0.386474609375,
0.270751953125,
0.615234375,
0.270751953125,
0.615234375,
0.2841796875,
0.386474609375,
0.2841796875
],
[
0.265625,
0.322509765625,
0.74169921875,
0.322509765625,
0.74169921875,
0.3359375,
0.265625,
0.3359375
],
[
0.2646484375,
0.337158203125,
0.740234375,
0.337158203125,
0.740234375,
0.35107421875,
0.2646484375,
0.35107421875
],
[
0.2646484375,
0.35107421875,
0.73876953125,
0.35107421875,
0.73876953125,
0.365478515625,
0.2646484375,
0.365478515625
],
[
0.26318359375,
0.3642578125,
0.740234375,
0.363037109375,
0.740234375,
0.37744140625,
0.26318359375,
0.37890625
],
[
0.2646484375,
0.37890625,
0.73876953125,
0.37890625,
0.73876953125,
0.392578125,
0.2646484375,
0.392578125
],
[
0.0228729248046875,
0.397705078125,
0.062103271484375,
0.397705078125,
0.062103271484375,
0.494384765625,
0.0228729248046875,
0.494384765625
],
[
0.2646484375,
0.392578125,
0.73876953125,
0.392578125,
0.73876953125,
0.406005859375,
0.2646484375,
0.406005859375
],
[
0.2646484375,
0.406005859375,
0.740234375,
0.406005859375,
0.740234375,
0.419189453125,
0.2646484375,
0.419189453125
],
[
0.2646484375,
0.42041015625,
0.73876953125,
0.42041015625,
0.73876953125,
0.434326171875,
0.2646484375,
0.434326171875
],
[
0.2646484375,
0.434326171875,
0.740234375,
0.434326171875,
0.740234375,
0.447509765625,
0.2646484375,
0.447509765625
],
[
0.2646484375,
0.447509765625,
0.740234375,
0.447509765625,
0.740234375,
0.4609375,
0.2646484375,
0.4609375
],
[
0.2646484375,
0.4609375,
0.740234375,
0.4609375,
0.740234375,
0.474853515625,
0.2646484375,
0.474853515625
],
[
0.2646484375,
0.47607421875,
0.740234375,
0.47607421875,
0.740234375,
0.4892578125,
0.2646484375,
0.4892578125
],
[
0.0228729248046875,
0.4912109375,
0.05963134765625,
0.4912109375,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.2646484375,
0.4892578125,
0.740234375,
0.4892578125,
0.740234375,
0.50244140625,
0.2646484375,
0.50244140625
],
[
0.2646484375,
0.50244140625,
0.740234375,
0.50244140625,
0.740234375,
0.5166015625,
0.2646484375,
0.5166015625
],
[
0.2646484375,
0.517578125,
0.740234375,
0.517578125,
0.740234375,
0.53076171875,
0.2646484375,
0.53076171875
],
[
0.2646484375,
0.53076171875,
0.740234375,
0.53076171875,
0.740234375,
0.54443359375,
0.2646484375,
0.54443359375
],
[
0.265625,
0.54541015625,
0.71728515625,
0.54541015625,
0.71728515625,
0.55615234375,
0.265625,
0.55615234375
],
[
0.2646484375,
0.5712890625,
0.72802734375,
0.5712890625,
0.72802734375,
0.5859375,
0.2646484375,
0.5859375
],
[
0.22216796875,
0.611328125,
0.236083984375,
0.6162109375,
0.2296142578125,
0.626953125,
0.2156982421875,
0.6220703125
],
[
0.25,
0.61279296875,
0.374267578125,
0.61279296875,
0.374267578125,
0.62744140625,
0.25,
0.62744140625
],
[
0.2181396484375,
0.64599609375,
0.78369140625,
0.64599609375,
0.78369140625,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.2181396484375,
0.6611328125,
0.78515625,
0.6611328125,
0.78515625,
0.6748046875,
0.2181396484375,
0.6748046875
],
[
0.2181396484375,
0.67529296875,
0.7861328125,
0.67529296875,
0.7861328125,
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,
0.78759765625,
0.70556640625,
0.78759765625,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.78759765625,
0.72021484375,
0.78759765625,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.21728515625,
0.736328125,
0.515625,
0.73486328125,
0.515625,
0.75,
0.21728515625,
0.75048828125
],
[
0.24267578125,
0.75048828125,
0.78759765625,
0.75048828125,
0.78759765625,
0.76513671875,
0.24267578125,
0.76513671875
],
[
0.2181396484375,
0.767578125,
0.78515625,
0.767578125,
0.78515625,
0.78076171875,
0.2181396484375,
0.78076171875
],
[
0.2181396484375,
0.7822265625,
0.78515625,
0.7822265625,
0.78515625,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.21728515625,
0.8125,
0.7861328125,
0.8125,
0.7861328125,
0.82568359375,
0.21728515625,
0.82568359375
],
[
0.21728515625,
0.82568359375,
0.26953125,
0.82763671875,
0.268798828125,
0.84228515625,
0.2156982421875,
0.84033203125
]
] |
[
"Few-shot Single-view 3D Reconstruction with",
"Memory Prior Contrastive Network",
"Zhen Xingl, Yijiang Chen', Zhixin Ling',",
"Xiangdong Zhou', and Yu Xiang?",
"1 School of Computer Science, Fudan University, Shanghai, China, 200433",
"",
"The University of Texas at Dallas",
"Abstract. 3D reconstruction of novel categories based on few-shot learn-",
"ing is appealing in real-world applications and attracts increasing re-",
"search interests. Previous approaches mainly focus on how to design",
"shape prior models for different categories. Their performance on unseen",
"categories is not very competitive. In this paper, we present a Memory",
"['S]",
"Prior Contrastive Network (MPCN) that can store shape prior knowl-",
"edge in a few-shot learning based 3D reconstruction framework. With",
"the shape memory, a multi-head attention module is proposed to capture",
"different parts of a candidate shape prior and fuse these parts together",
"to guide 3D reconstruction of novel categories. Besides, we introduce a",
"3D-aware contrastive learning method, which can not only complement",
"the retrieval accuracy of memory network, but also better organize im-",
"128100802:A1",
"age features for downstream tasks. Compared with previous few-shot",
"3D reconstruction methods, MPCN can handle the inter-class variabil-",
"ity without category annotations. Experimental results on a benchmark",
"synthetic dataset and the Pascal3D+ real-world dataset show that our",
"model outperforms the current state-of-the-art methods significantly",
"Keywords: Few-shot learning, 3D Reconstruction, Memory Network",
"1",
"Introduction",
"Reconstructing 3D shapes from RGB images is valuable in many real-world ap-",
"plications such as autonomous driving, virtual reality, CAD and robotics. Tradi-",
"tional methods for 3D reconstruction such as Structure From Motion (SFM) [29]",
"and Simultaneous Localization and Mapping (SLAM) [2] often require signifi-",
"cant efforts in data acquisition, For example, a large number of images need to",
"be captured and the camera parameters need to be calibrated, which limit the",
"applications of thesetraditional methods.",
"In recent years, 3D reconstruction from single image based on deep neural",
"networks attracts great research interests. However, previous methods of single-",
"view 3D reconstruction are mostly category-specific [42,8].Therefore, they only",
"perform well in the specific training categories. These methods also require a",
"large number of labeled training images, which is time consuming and costly to",
"obtain."
] |
{
"arxiv_id": "2208.00183",
"language": "en",
"page_idx": 0
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.