Add files using upload-large-folder tool

parse/train/B1eY_pVYvB/B1eY_pVYvB.md

@@ -0,0 +1,284 @@
+# EFFICIENT AND INFORMATION-PRESERVING FUTURE FRAME PREDICTION AND BEYOND
+
+Wei ${ { \bf { Y } } { \bf { u } } ^ { 1 } }$ , Yichao $\mathbf { L u } ^ { 1 }$ , Steve Easterbrook1, Sanja Fidler1,2,3
+
+1Department of Computer Science, University of Toronto 2Vector Institute, Canada   
+3NVIDIA   
+{gnosis,yichao,sme,fidler}@cs.toronto.edu
+
+# ABSTRACT
+
+Applying resolution-preserving blocks is a common practice to maximize information preservation in video prediction, yet their high memory consumption greatly limits their application scenarios. We propose CrevNet, a Conditionally Reversible Network that uses reversible architectures to build a bijective two-way autoencoder and its complementary recurrent predictor. Our model enjoys the theoretically guaranteed property of no information loss during the feature extraction, much lower memory consumption and computational efficiency. The lightweight nature of our model enables us to incorporate 3D convolutions without concern of memory bottleneck, enhancing the model’s ability to capture both short-term and long-term temporal dependencies. Our proposed approach achieves state-of-the-art results on Moving MNIST, Traffic4cast and KITTI datasets. We further demonstrate the transferability of our self-supervised learning method by exploiting its learnt features for object detection on KITTI. Our competitive results indicate the potential of using CrevNet as a generative pre-training strategy to guide downstream tasks.
+
+# 1 INTRODUCTION
+
+Deep learning has enjoyed tremendous success in recent years due to its ability to capture complex dependencies and non-linearities in large datasets (Krizhevsky et al. (2012); He et al. (2016); Gomez et al. (2017)). Excellent performance has been achieved on a wide range of supervised machine learning tasks, ranging from image classification (He et al. (2016)) and object detection (Ren et al. (2015)) to speech recognition (Amodei et al. (2016)). Despite the significant breakthrough in supervised learning, the potential of applying deep architectures to unsupervised learning problems remains largely unexplored. Lately there has been a surge of interest in the task of video prediction, i.e., to predict future frames of a video sequence (Wang et al. (2017; 2018); Denton et al. (2017); Denton & Fergus (2018); Villegas et al. (2017); Lee et al. (2018)). The significance of video prediction primarily lies in its potential of discovering dynamics in the physical world. The self-supervised nature of video prediction aligns well with how humans learn, without requiring large amounts of labeled data. In addition, videos can provide an abundant and virtually unlimited source of visual information. This allows video prediction models to serve as a generative pre-training strategy of feature representation learning for a variety of downstream supervised tasks.
+
+To date, most of the existing models for video prediction employ a hybrid of convolutional and recurrent layers as the underlying architecture (Wang et al. (2017); Shi et al. (2015); Lotter et al. (2016)). Such architectural design enables the model to simultaneously exploit the ability of convolutional units to model spatial relationships and the potential of recurrent units to capture temporal dependencies. Despite their prevalence in the literature, classical video prediction architectures suffer from two major limitations. Firstly, in dense prediction tasks such as video prediction, models are required to make pixel-wise predictions, which emphasizes the demand for the preservation of information through layers. Prior works attempt to address such demand through the extensive use of resolution-preserving blocks (Wang et al. (2017; 2018); Kalchbrenner et al. (2016)). Nevertheless, these resolution-preserving blocks are not guaranteed to preserve all the relevant information, and they greatly increase the memory consumption and computational cost of the models. The second drawback of existing video prediction models is that they cannot efficiently take advantage of 3D convolutions, as that would make these already cumbersome architectures even larger. 3D convolutions have been shown to be a very effective alternative to RNNs to capture temporal relations in a variety of video tasks (Liu et al. (2018); Carreira & Zisserman (2017)), and thus desirable to exploit.
+
+![](images/2814f3e7c41cd24b84d6f7d7d8910bd06c6aa74d5183b36c2eadb0881c09f737.jpg)  
+Figure 1: The pipeline of our proposed CrevNet where a single two-way autoencoder serves as both encoder and decoder. CrevNet first observes a warm-up video sequence and then starts a multi-frame video prediction without refeeding its own predictions.
+
+Recently, reversible architectures (Dinh et al. (2014); Gomez et al. (2017); Jacobsen et al. (2018)) have attracted attention due to their light memory demand and their information preserving property by design. However, the effectiveness of reversible models remains greatly unexplored in the video literature. In this paper, we introduce a novel, conditionally reversible video prediction model, CrevNet, in the sense that when conditioned on previous hidden states, it can exactly reconstruct the input from its predictions. The contribution of this work can be summarized as follows:
+
+• We introduce a two-way autoencoder that uses the forward and backward passes of an invertible network as encoder and decoder (Fig 1). The volume-preserving two-way autoencoder not only greatly reduces the memory demand and computational cost, but also enjoys the theoretically guaranteed property of no information loss. The lightweight nature of our model enables us to incorporate 3D convolutions without concern of memory bottleneck. We propose the reversible predictive module (RPM), as illustrated in Fig 2b, which extends the reversibility from spatial to temporal domain. RPM, together with the two-way autoencoder, provides a conditionally reversible architecture (CrevNet) for spatiotemporal learning. CrevNet achieves the state-of-the-art results on Moving MNIST, Traffic4cast and KITTI. We evaluate the effectiveness of features learnt from self-supervision by adapting our CrevNet for object detection on KITTI. Our competitive results indicate the potential of using CrevNet as a generative pre-training strategy to guide downstream CV tasks.
+
+# 2 APPROACH
+
+We first outline the general pipeline of our method. Our CrevNet consists of two subnetworks, an autonencoder network with an encoder $\mathcal { E }$ , decoder $\mathcal { D }$ and a recurrent predictor $\mathcal { P }$ bridging encoder and decoder. Let $\boldsymbol { x } _ { t } \in \mathbb { R } ^ { w \times h \times c }$ represent the $t _ { \mathrm { t h } }$ frame in video $x$ , where $w , h .$ , and $c$ denote its width, height, and the number of channels. Given $x _ { 0 : t - 1 }$ , the model predicts the next frame $\hat { x } _ { t }$ as follows:
+
+$$
+\hat { x } _ { t } = \mathcal { D } ( \mathcal { P } ( \mathcal { E } ( x _ { t - 1 } ) | x _ { 0 : t - 2 } ) )
+$$
+
+In the case of 3D convolution, $\boldsymbol { x } _ { t } \in \mathbb { R } ^ { k \times w \times h \times c }$ denotes the short video clip from $t$ to $t + k - 1$ instead of a single frame at timestep $t$ , where $k$ is the temporal dimension of input or output. During the multi-frame generation process without access to the ground truth frames, the model uses its previous predictions instead.
+
+# 2.1 THE INVERTIBLE TWO-WAY AUTOENCODER
+
+We propose a bijective two-way autoencoder based on the additive coupling layer introduced in NICE (Dinh et al. (2014)). We begin with describing the building block of the two-way autoencoder
+
+![](images/82c62b6a2698f2565cb8ca6a3f84285551bbb5913c9507f6e1009f46cb4a594c.jpg)  
+Figure 2: The network architecture of CrevNet (Better viewed in color). The input video frames are first reshaped and split channelwise into two groups. These two groups are passed to the two-way autoencoder (a) for feature extraction, and then to the predictor made up of multiple reversible predictive modules (b). The transformed high-level features produced by predictor are then passed back through the decoding pass of (a), shown here as a representative block (c) to yield its prediction.
+
+(Fig 2a). Formally, the input $x$ is first reshaped and split channelwise into two groups, denoted as $x ^ { 1 }$ and $x ^ { 2 }$ . During the forward pass of each building block, one group, e.g. $x ^ { 1 }$ , passes through several convolutions and activations and is then added to another group, $x ^ { \bar { 2 } }$ , like a residual block:
+
+$$
+\hat { x } ^ { 2 } = x ^ { 2 } + \mathcal { F } _ { 1 } ( x ^ { 1 } ) \qquad \hat { x } ^ { 1 } = x ^ { 1 } + \mathcal { F } _ { 2 } ( \hat { x } ^ { 2 } )
+$$
+
+where $\mathcal { F }$ is a composite non-linear operator consisting of convolutions and activations, and ${ \hat { x } } ^ { 1 }$ and ${ \hat { x } } ^ { 2 }$ are the updated $x ^ { 1 }$ and $x ^ { 2 }$ . Note that $x ^ { 1 }$ and $x ^ { 2 }$ can be simply recovered from ${ \hat { x } } ^ { 2 }$ and ${ \hat { x } } ^ { 1 }$ by the inverse computation (Fig 2c) as follows:
+
+$$
+x ^ { 1 } = \hat { x } ^ { 1 } - \mathcal { F } _ { 2 } ( \hat { x } ^ { 2 } ) \qquad x ^ { 2 } = \hat { x } ^ { 2 } - \mathcal { F } _ { 1 } ( x ^ { 1 } )
+$$
+
+Multiple building blocks are stacked in an alternating fashion between $x ^ { 1 }$ and $x ^ { 2 }$ to construct a two-way autoencoder, as shown in Fig 2a. A series of the forward and inverse computations builds a one-to-one and onto, i.e. bijective , mapping between the input and features. Such invertibility ensures that there is no information loss during the feature extraction, which is presumably more favorable for video prediction since the model is expected to restore the future frames with finegrained details. To enable the invertibility of the entire autoencoder, our two-way autoencoder uses a bijective downsampling, pixel shuffle layer (Shi et al. (2016)), that changes the shape of feature from $( w , h , c )$ to $( w / n , \bar { h } / n , \bar { c } \times n ^ { 2 } )$ . The resulting volume-preserving architecture can greatly reduce its memory consumption compared with the existing resolution-preserving methods.
+
+We further argue that for generative tasks, e.g. video prediction, we can effectively utilize a single two-way autoencoder, and to use its forward and backward pass as the encoder and the decoder, respectively. The predicted frame $\hat { x } _ { t }$ is thus given by
+
+$$
+\hat { x } _ { t } = \mathcal { E } ^ { - 1 } ( \mathcal { P } ( \mathcal { E } ( x _ { t - 1 } ) | x _ { 0 : t - 2 } ) )
+$$
+
+where ${ { \mathcal { E } } ^ { - 1 } }$ is the backward pass of $\mathcal { E }$ . Our rationale is that, such setting would not only reduce the number of parameters in the model, but also encourage the model to explore the shared feature space between the inputs and the targets. As a result, our method does not require any form of information sharing, e.g. skip connection, between the encoder and decoder. In addition, our two-way autoencoder can enjoy a lower computational cost at the multi-frame prediction phase where the encoding pass is no longer needed and the predictor directly takes the output from previous timestep as input, as shown in Fig 1, since $\mathcal { E } ( \mathcal { E } ^ { - 1 } )$ is an identity mapping .
+
+# 2.2 REVERSIBLE PREDICTIVE MODULE
+
+In this section, we describe the second part of our video prediction model, the predictor $\mathcal { P }$ , which computes dependencies along both the space and time dimensions. Although the traditional stackedConvRNN layers architecture is the most straightforward choice of predictor, we find that it fails to establish a consistent temporal dependency when equipped with our two-way autoencoder through experiments. Therefore, we propose a novel reversible predictive module (RPM), which can be regarded as a recurrent extension of the two-way autoencoder. In the RPM, we substitute all standard convolutions with layers from the ConvRNN family (e.g. ConvLSTM or spatiotemporal LSTM) and introduce a soft attention (weighting gates) mechanism to form a weighted sum of the two groups instead of the direct addition. The main operations of RPM used in this paper are given as follows:
+
+ConvRNN
+
+$$
+\begin{array} { r l } & { h _ { t } ^ { 1 } = \mathrm { C o n v R N N } ( x _ { t } ^ { 1 } , h _ { t - 1 } ^ { 1 } ) } \\ & { g _ { t } = \phi ( W _ { 2 } * \mathrm { R e L U } ( W _ { 1 } * h _ { t } ^ { 1 } + b _ { 1 } ) + b _ { 2 } ) } \\ & { \hat { x _ { t } ^ { 2 } } = ( 1 - g _ { t } ) \odot x _ { t } ^ { 2 } + g _ { t } \odot h _ { t } ^ { 1 } } \end{array}
+$$
+
+Attention module
+
+Weighted sum
+
+where $\boldsymbol { x } _ { t } ^ { 1 }$ and $\boldsymbol { x } _ { t } ^ { 2 }$ denote two groups of features at timestep $t$ , $h _ { t } ^ { 1 }$ denote the hidden states of ConvRNN layer, $\phi$ is sigmoid activation, $^ *$ is the standard convolution operator and $\odot$ is the Hadamard product. The architecture of reversible predictive module is also shown in Fig 2b. RPM adopts a similar architectural design as the two-way autoencoder to ensure a pixel-wise alignment between the input and the output, i.e. each position of features can be traced back to certain pixel, and thus make it compatible with our two-way autoencoder. It also mitigates the vanishing gradient issues across stacked layers since the coupling layer provides a nice property w.r.t. the Jacobian (Dinh et al. (2014)). In addition, the attention mechanism in the RPM enables the model to focus on objects in motion instead of background, which further improves the video prediction quality. Similarly, multiple RPMs alternate between the two groups to form a predictor. We call this predictor conditionally reversible since, given $h _ { t - 1 }$ , we are able to reconstruct $x _ { t - 1 }$ from $\hat { x } _ { t }$ if there are no numerical errors:
+
+$$
+x _ { t - 1 } = \mathcal { E } ^ { - 1 } ( \mathcal { P } ^ { - 1 } ( \mathcal { E } ( \hat { x } _ { t } ) | h _ { t - 1 } ) )
+$$
+
+where $\mathcal { P } ^ { - 1 }$ is the inverse computation of the predictor $\mathcal { P }$ . We name the video prediction model using two-way autoencoder as its backbone and RPMs as its predictor CrevNet. Another key factor of RPM is the choice of ConvRNN. In this paper, we mainly employ ConvLSTM (Shi et al. (2015)) and spatiotemporal LSTM (ST-LSTM, Wang et al. (2017)) to enable a fair comparison with baselines.
+
+# 2.3 3D CONVOLUTIONS
+
+3D convolutions are proposed to address the shortcomings of standard 2D convolutions. The major difference between 2D-CNNs and 3D-CNNs is that at each time step 2D-CNNs take as input one video frame, while 3D-CNNs read in and output a short video clip containing $k$ continuous video frames. By applying convolutions on the temporal dimension along with the spatial dimension, models equipped with 3D convolution filters can not only extract representative spatiotemporal features, but also learn to produce consistent video clip at each generation, which further improve the quality of long-term prediction. In some cases, e.g. sequences are too short, we will use 2 consecutive frames stacked in the channel dimension instead as input at each timestep to assemble a valid warm-up sequence for ConvRNN.
+
+# 3 EXPERIMENTS
+
+# 3.1 LONG-TERM PREDICTION—MOVING MNIST
+
+Moving MNIST (Srivastava et al. (2015)) is a synthetically generated dataset that contains an infinite number of sequences of length 20. Each sequence shows how 2 digits move at a constant speed and bounce inside a $6 4 \times 6 4$ frame, where each handwritten digit is randomly sampled from the MNIST dataset. By assigning different initial locations and velocities to each digit, it is possible to generate an unlimited number of sequences, thus enabling us to accurately evaluate the performance of each model without the concern of data insufficiency issues. In the default setting, models are trained to predict 10 future frames after observing 10 prior frames in the sequence. Although the dynamics of Moving MNIST seems to be simple at first glance. It is quite hard to generate consistent future frames in the task of long-term prediction as digits can bounce or occlude each other frequently.
+
+Datasets and Setup: The general architecture of CrevNet used on Moving MNIST is composed of a 36-layer two-way autoencoder and 8 RPMs. All variants of CrevNet are trained by using the Adam optimizer with a starting learning rate of $5 \times 1 0 ^ { - 4 }$ to minimize MSE. The training process is stopped
+
+after 300, 000 iterations with the batch size of 16 and evaluated with a fixed test set containing 5, 000 sequences. To ensure that all samples in the test set are unseen by the model, digits in the training set and the testing set are separately sampled from two mutually exclusive subsets of MNIST.   
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">MNIST-2</td><td colspan="2">MNIST-3 MSE</td><td rowspan="2">Memory (1 sample)</td><td rowspan="2">FLOPS (1 frame)</td></tr><tr><td>SSIM</td><td>MSE</td><td>Human</td><td>SSIM</td><td></td></tr><tr><td>ConvLSTM (Shi et al. (2015))</td><td>0.707</td><td>103.3</td><td>0.923</td><td>0.695</td><td>127.3</td><td>1043 MB</td><td>107.4 G</td></tr><tr><td>FRNN (Oliu et al. (2018))</td><td>0.819</td><td>68.4</td><td>0.848</td><td>0.791</td><td>90.4</td><td>717MB</td><td>80.1G</td></tr><tr><td>VPN (Kalchbrenner et al. (2016))</td><td>0.870</td><td>70.0</td><td>0.831</td><td>0.820</td><td>85.6</td><td>5206MB</td><td>309.6 G</td></tr><tr><td>PredRNN (Wang et al. (2017))</td><td>0.869</td><td>56.8</td><td>0.837</td><td>0.822</td><td>83.1</td><td>1666 MB</td><td>192.9 G</td></tr><tr><td>PredRNN++ (Wang et al. (2018))</td><td>0.898</td><td>46.5</td><td>0.781</td><td>0.864</td><td>68.4</td><td>2017MB</td><td>106.8 G</td></tr><tr><td>E3D-LSTM (Wang et al. (2019))</td><td>0.910</td><td>41.3</td><td>0.706</td><td>0.870</td><td>62.4</td><td>2695 MB</td><td>381.3 G</td></tr><tr><td>CrevNet+ConvLSTM</td><td>0.928</td><td>38.5</td><td>0.602</td><td>0.886</td><td>57.2</td><td>130 MB</td><td>0.919 G</td></tr><tr><td>CrevNet + ST-LSTM</td><td>0.949</td><td>22.3</td><td>0.558</td><td>0.916</td><td>40.6</td><td>195MB</td><td>1.618 G</td></tr></table>
+
+Table 1: Quantitative evaluation of different methods on Moving MNIST. All metrics are averaged over the 10 predictions. Lower MSE and higher SSIM indicates better prediction accuracy.
+
+We compare CrevNet to six popular benchmark models from the literature: (i) ConvLSTM (Shi et al. (2015)), (ii) FRNN (Oliu et al. (2018)), (iii) VPN (Kalchbrenner et al. (2016)), (iv) PredRNN (Wang et al. (2017)) , (v) PredRNN $^ { + + }$ (Wang et al. (2018)), and (vi) E3D-LSTM (Wang et al. (2019)),. All baselines are implemented and optimized by following their corresponding protocols. To test our model in a more challenging setting, we also extend Moving MNIST to a 3-digit version where digits are more likely to occlude each other.
+
+Results: The performance of each model in terms of per-frame MSE and the Structural Similarity Index Measure (SSIM) (Wang et al. (2004)) is presented in Table 1. CrevNet outperforms all previous methods by a wide margin on both metrics while memory consumption of all CrevNet variants is significantly lower than that of other baselines. In particular, CrevNet with ConvLSTM only uses 130 MB memory per sample and is still capable of yielding results better than any baselines.
+
+To analyze the contribution of each module, we conduct an ablation study on both ConvLSTM and ST-LSTM with respect to 3D convolution, two-way autoencoder and RPM and summarize the results in Table 2. Note that we do not include the quantitative results of the combination of two-way autoencoder and stack-ConvRNN predictor because it fails to produce consistent long-term generations and we choose UNet (Ronneberger et al. (2015)) as an alternative to our two-way autoencoder. We can observe a significant improvement over ConvLSTM after we embed it into our CrevNet framework, indicating the effectiveness of reversible architectures. Also, integrating 3D convolution can consistently enhance the performance of all architectures. To further show the superior performance of CrevNet, we evaluate it on a harder 3-digit setting. Results are shown in the right column of Table 1. Compared with the 2-digit setting, all models suffer a deterioration in quantitative performance due to the more frequent occurrence of overlapping digits. Nevertheless, our CrevNet still achieves the best result.
+
+![](images/7ed8d7ba2d38f9bb78e2fb72587a3164387b1cd5d302ac1569447f621063ef89.jpg)  
+Figure 3: An extremely hard sequence of Moving MNIST where two digits are continuously overlapped during the warm-up phase.
+
+<table><tr><td>Model</td><td>ConvLSTM MSE SSIM|</td><td>ST-LSTM MSE SSIM</td></tr><tr><td>Stacked-RNN</td><td>103.3 0.707</td><td>56.8 0.869</td></tr><tr><td>3D Stacked-RNN</td><td>85.8 0.785</td><td>46.2 0.878</td></tr><tr><td>UNet+StackedRNN 83.5</td><td>0.793</td><td>58.8 0.865</td></tr><tr><td>UNet+RPM</td><td>63.4 0.855</td><td>50.2 0.896</td></tr><tr><td>CrevNet w/o 3D</td><td>50.2 0.888</td><td>40.4 0.916</td></tr><tr><td>CrevNet + 3D</td><td>38.5 0.928</td><td>22.3 0.949</td></tr></table>
+
+Table 2: An ablation study w.r.t. 3D convolution, RPM and two-way autoencoder. All metrics are averaged over the 10 predictions.
+
+In Fig 3 , our qualitative analysis shows how each model performs on an extremely hard case of Moving MNIST where two digits are continuously overlapped during the warm-up phase.
+
+![](images/5d5e85adb45b4e2b127effb0619d1ce9ff93eb2c78c8c9d4df842a4141a5c165.jpg)  
+Figure 4: The visual comparison of Traffic4cast. The red boxes track some dynamics successfully captured by our CrevNet. Better viewed large.
+
+As we can see, our model is the only model that can differentiate the overlapping digits. The information-preserving property of the two-way autoencoder enables our method to reconstruct every fine detail of moving digits after occlusion while baselines typically only restore the basic shape of these numbers. In fact, our CrevNet works almost perfectly on Moving MNIST, with most of its generations being visually indistinguishable from groundtruth.
+
+We perform a human study to assess the fidelity of the video clips generated by different models. We presented pairs of video clips to human judges, where each pair consists of a video clip from the test set together with the prediction generated by the model. The judges were asked to decide which of the two video clips is more likely to be the groundtruth. To make each trail blind, the judges were not informed which model is used for generation and two sequences were randomly displayed on either side of the screen. We totally collected 2439 responses made by 58 human subjects and then calculated the probability that human judges answered correctly. The results are reported in Table 1. The accuracy of $5 5 . 8 \%$ suggests that subjects could hardly detect the difference and their decisions were very close to random guesses.
+
+# 3.2 SHORT-TERM PREDICTION—TRAFFIC FLOW FORECASTING
+
+Next, we evaluate our model on a more complicated real-world dataset, Traffic4cast (IARAI (2019)), which collects the traffic statuses of 3 big cities over a year at a 5-minute interval. Traffic forecasting can be straightforwardly defined as video prediction task by its spatiotemporal nature. However, this dataset is quite challenging for the following reasons. (1). High resolution: The frame resolution of Traffic4cast is $4 9 5 \times 4 3 6$ , which is the highest among all datasets. Existing resolution-preserving methods can hardly be adapted to this dataset since they all require extremely large memory and computation. Even if these models can be fitted in GPUs, they still do not have large enough receptive fields to capture the meaningful dynamics as vehicles can move up to 100 pixels between consecutive frames. (2). Complicated nonlinear dynamics: Valid data points only reside on the hidden roadmap of each city, which is not explicitly provided in this dataset. Moving vehicles on these curved roads along with tangled road conditions will produce very complex nonlinear behaviours. It also involves many unobservable conditions or random events like weather and car accidents.
+
+Datasets and Setup: Each frame in Traffic4cast dataset is a $4 9 5 \times 4 3 6 \times 3$ heatmap, where the last dimension records 3 traffic statuses representing volume, mean speed and major direction at given location. The architecture of CrevNet is the same as the one we used on Moving MNIST. As we mentioned before, the existing resolution-preserving methods cannot handle such high resolution input. Thus, to make the comparison possible, we add U-Net encoder-decoder to the baseline models including ConvLSTM and ST-LSTM. We train each model to predict next 3 frames (the next 15 minutes) from 9 observations and evaluate prediction with MSE criterion.
+
+<table><tr><td>Model</td><td>CrevNet</td><td>UNet+ST-LSTM</td><td>UNet+ConvLSTM丨Best Team</td><td></td><td>2nd Best Team</td></tr><tr><td>MSE</td><td>9.340×10-3</td><td>9.725× 10-3</td><td>9.846 ×10-3</td><td>9.559×10-3</td><td>9.717 × 10-3</td></tr></table>
+
+Table 3: Quantitative evaluation on Traffic4cast. Lower MSE indicates better prediction accuracy.
+
+![](images/d59d0905d69ce4a69c2e0d6fb0d4d47d0f31ad6570a2d67a3b99ee83bd6a76ef.jpg)  
+Figure 5: The visual comparison of next-frame predictions on Caltech Pedestrian.
+
+Results: The quantitative comparison including the best two results on the leaderboard before the submission of this paper is provided in Table 3. Unlike all previous state-of-the-art methods, CrevNet does not suffer from high memory consumption so that we were able to train our model in a single V100 GPU. The invertibility of two-way autoencoder preserves all necessary information for spatiotemporal learning and allows our model to generate sharp and reasonable predictions. As illustrated in Fig 4, our model can identify and remember the hidden roadmap of each city through the learning of complicated nonlinear dynamics and accurately predict how traffic system will evolve.
+
+# 3.3 NEXT-FRAME PREDICTION AND BEYOND—CAR-MOUNTED CAMERA VIDEO
+
+The real-world videos are usually long-term unpredictable because of the intrinsic randomness and the lack of necessary information. Thus, the common practice for datasets like KITTI (Geiger et al. (2012)), a car-mounted camera video dataset, is to perform next-frame prediction. In this section, we further demonstrate the superior performance of our CrevNet by conducting experiments on KITTI and Caltech Pedestrian (Dollár et al. (2009)). Compared with the previous two settings, carmounted camera videos dataset presents another level of difficulty for video prediction as it describes various nonlinear three-dimensional dynamics of multiple moving objects including backgrounds. Furthermore, as our well-trained model is capable of generating authentic future frames, it should spontaneously learn at least the shape and location of all moving objects, which indicates that the learnt features are very informative for downstream tasks. For example, in the case of object detection, these features can be incorporated to estimate more accurate locations and sizes of bounding boxes. Therefore, we also explore the effectiveness of our self-supervised learning method on the 2D object detection on KITTI.
+
+# 3.3.1 VIDEO PREDICTION
+
+Datasets and Setup: We follow the same protocol used in PredNet (Lotter et al. (2016)) for preprocessing and evaluation. We first center-crop all video frames and resize them into $1 2 8 \times 1 6 0$ . We compare our proposed method with 4 state-of-the-art benchmark models. Models are trained on KITTI dataset to predict the next frame after 10-frame warm-up and are evaluated on Caltech Pedestrian. The architecture of CrevNet used on KITTI is composed of a 48-layer two-way autoencoder and 40 RPMs. Note that it is the memory efficiency of our method that allows us to deploy such deep model. Since our model also possess good capability of long-term prediction. We add a 12-frame prediction comparison with CycleGAN (Kwon & Park (2019)) and PredNet (Lotter et al. (2016)).
+
+![](images/0deecc352bfbe70c65085939887ec28bba178661485724005c1bcff8bc8ec03e.jpg)  
+Figure 6: The visual comparison of 12-frame prediction on Caltech Pedestrian. Notice how well CrevNet captures the detail and geometry of the buildings in the background, and the overall shading.
+
+<table><tr><td>Model</td><td colspan="2">Next-Frame</td><td>3rd</td><td>6th</td><td>9th</td><td>12th</td><td>Average</td></tr><tr><td></td><td>PSNR</td><td>SSIM</td><td colspan="5">SSIM</td></tr><tr><td>Copy-Last-Frame Dual Motion GAN (Liang et al. (2017))</td><td>23.3</td><td>0.779 0.899</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ContextVP (Byeon et al. (2018)) PredNet (Lotter et al. (2016))</td><td>28.7 27.6</td><td>0.921 0.905</td><td>0.72</td><td>0.66</td><td>0.61</td><td>0.58</td><td>0.701</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CycleGAN (Kwon &amp; Park (2019))</td><td>29.2</td><td>0.919</td><td>0.83</td><td>0.73</td><td>0.67</td><td>0.63</td><td>0.752</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CrevNet</td><td>29.3</td><td>0.925</td><td>0.84</td><td>0.76</td><td>0.70</td><td>0.65</td><td>0.776</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></table>
+
+Table 4: Quantitative evaluation of different methods on the Caltech Pedestrian dataset. Higher PSNR or SSIM means better prediction accuracy.
+
+Results: Performance of different models in terms of PSNR and SSIM is displayed in Table 4. CrevNet outperforms all baselines in both next-frame and multi-frame prediction regimes. Visual comparisons are provided in Fig 5 and Fig 6. Especially, in the case of 12-frame generation, we can observe that compared with our method, PredNet suffers severely from the famous error propagation of RNN issue while CycleGAN produces realistic yet physically inconsistent predictions.
+
+# 3.3.2 2D OBJECT DETECTION
+
+Datasets and Setup: KITTI provides three prior frames of unlabeled data for each labeled image. This allows us to run our CrevNet to extract useful spatiotemporal features for object detection. All video sequences were recorded at $1 0 \ : \mathrm { H z }$ with resolution of $1 2 4 2 \times 3 7 5$ . We first resize each frame to $4 1 6 \times 1 2 8$ and finetune our best model on the video prediction task solely. The combinations of features extracted by our two-way autoencoder and attention masks of the target frame are then fed into the detection head for the further training. Note that we do not update the weights of CrevNet at this stage to purely demonstrate the power of self-supervised learning. Two image-based detection models, SqueezeDet and RRC, are compared as baselines. We also add an experiment on transfer learning of features learnt by PredNet on KITTI as comparison. To be consistent with related work, we use SSD (Liu et al. (2016)) as detection head.
+
+Table 5: Quantitative evaluation of different methods on 2D KITTI detection in term of Average Precision (AP). The numbers in the "Best Results on the Leaderboard " are achieved by different best models on each class according to the default ranking mechanism on the Leaderboard.   
+
+<table><tr><td rowspan="2">Backbone</td><td rowspan="2">Add-ons</td><td colspan="3">Car L</td><td colspan="3">Pedestrain M</td><td colspan="3">Cyclist</td><td rowspan="2">mAP</td></tr><tr><td>E</td><td>M</td><td>H</td><td>E</td><td></td><td>H</td><td>E</td><td>M</td><td>H</td></tr><tr><td>SqueezeDet RRC</td><td>(Wu et al. (2017)) (Ren et al. (2017))</td><td>90.4 90.61</td><td>87.1 90.23</td><td>78.9 87.44</td><td>81.4 84.16</td><td>71.3 75.33</td><td>68.5 70.39</td><td>87.6 84.96</td><td>80.3 76.49</td><td>78.1 65.46</td><td>80.4 80.56</td></tr><tr><td colspan="2">Best Results on the Leaderboard</td><td>91.96</td><td>91.97</td><td>84.57</td><td>88.27</td><td>81.73</td><td>75.29</td><td>84.28</td><td>79.24</td><td>71.22</td><td>83.17</td></tr><tr><td>PredNet</td><td>(Lotter et al. (2016))</td><td>59.05</td><td>41.61</td><td>37.53</td><td>50.88</td><td>47.51</td><td>43.44</td><td>46.25</td><td>43.79</td><td>38.66</td><td>45.41</td></tr><tr><td>CrevNet</td><td>attention mask + extracted features</td><td>91.53 91.94</td><td>90.95 91.84</td><td>85.71 85.97</td><td>89.31 89.66</td><td>82.55 83.17</td><td>75.21 75.80</td><td>85.51 87.33</td><td>80.41 80.91</td><td>71.52 72.21</td><td>83.63 84.31</td></tr></table>
+
+Results: The results of all experiments and baselines can be found in Table 5. Surprisingly, our CrevNet even outperforms the combination of the best model on each class in term of mAP. Since our model is capable of capturing the motion information, it is sensitive to the small (hard) moving objects. However, the motion information alone is not sufficient for object detection due to the appearance of relatively static objects. Therefore, we can observe a performance boost after we incorporate the features extracted by our two-way autoencoder. Another advantage of our method is that it can provide a better localization of bounding box since the learnt features of CrevNet remain the pixel-wise alignment with the input and output frame. Finally, thanks to the lightweight nature of our CrevNet, our best detection model can run at 6.8 FPS at the testing time.
+
+# 4 RELATED WORK
+
+Deep Learning in Video Prediction: Mainstream video prediction models can mostly be categorized into two frameworks, stacked ConvRNNs and encoder-predictor-decoder models. The former framework attempts to design a new spatiotemporal module and then stacks multiple such modules to form the final model, while the latter usually utilizes an autoencoder to project video frames into their latent representations and then employs a recurrent neural network to model the temporal transformations. PredNet (Lotter et al. (2016)) is a good representative of stacked ConvRNNs framework. In PredNet, each ConvLSTM layer produces a layer-specific prediction at every time step to transmit an error term to the next layer. This model works well for predicting the next frame, but fails to maintain its performance in a long-term setting. To tackle long-term predictions, PredRNN (Wang et al. (2017)) proposed a new spatiotemporal LSTM, which allows memory to flow both vertically and horizontally. PredRNN $^ { + + }$ (Wang et al. (2018)) further improved the results by rearranging spatial and temporal memory in a cascaded mechanism, and by using a gradient highway architecture to ease the optimization. E3D-LSTM (Wang et al. (2019)) effectively recalled the previous memory states and also proposed to include 3D convolutions to enhance its performance. ContextVP (Byeon et al. (2018)) introduced a fully context-aware architecture consisting of parallel multi-dimensional LSTM units and blending units. Methods from stacked ConvRNNs family usually yield more accurate deterministic predictions but they consume considerable GPU memory and computational power as they abandon downsampling to prevent information loss.
+
+The encoder-predictor-decoder framework, on the other hand, provides more flexibility than its counterpart. MCNET (Villegas et al. (2017)) and DrNet (Denton et al. (2017)) decompose the content and motion in videos by building their corresponding encoders and then integrate this disentangled information to yield the next frame. Retrospective CycleGAN (Kwon & Park (2019)) combines sequential adversarial loss with frame adversarial loss, which encourages the model to generate frames that are visually similar to authentic images. In terms of modeling stochasticity, SVG (Denton & Fergus (2018)) and SAVG (Lee et al. (2018)) utilize a prior inference network to mimic the uncertainty in the environment, and then embed it into a deterministic generative model to produce stochastic video frames. VPN (Kalchbrenner et al. (2016)) estimates the discrete joint distribution of the raw pixel values in a video using the well-established PixelCNNs. It is worth noticing that VPN employs a resolution-preserving encoder to circumvent the information loss, showing the need for an efficient information-preserving encoder in the community.
+
+Comparison with Related Works: To the best of our knowledge, our CrevNet is the first conditionally reversible model in the video prediction literature. There are three prior arts, E3D-LSTM (Wang et al. (2019)), FRNN (Oliu et al. (2018)) and VideoFlow (Kumar et al. (2019)), having some similarities with our CrevNet. While E3D-LSTM also employs 3D convolutions, their implementation is essentially equivalent to applying two 2D convolutional operations, as there is no shared filter on the temporal dimension. Similar to CrevNet, FRNN reduces its computational cost by eliminating the need to re-encode the output of decoder. However, FRNN has a substantially different architecture compared to CrevNet. While the encoder and decoder in our model do not need information sharing at all, FRNN relies heavily on the sharing of the hidden states between them. Although VideoFlow also utilizes invertible transformation. This approach is very different from ours because: (1). VideoFlow is built upon Glow, a very memory-consuming architecture. Such memory limits preclude the use of 3D convolutions, or even from training the model with Adam. (2). They use ANN to model temporal relationship. As such, VideoFlow cannot capture complex dynamics. (3). So far, VideoFlow has only been applied to stochastic video generation instead of deterministic video prediction.
+
+The Reversible and Invertible Architectures: The idea of the coupling layer was initially introduced in NICE (Dinh et al. (2014)) so as to make the computation of the determinant of the Jacobian and inverse Jacobian trival. Inspired by additive coupling layer, RevNet (Gomez et al. (2017)) introduced a reversible block that allowed the reconstruction of activations of each layer from that of the next layer, thus eliminating the need to store activations between downsampling and significantly reducing its memory consumption. The follow-up work by (Jacobsen et al. (2018)) further proposed an invertible extension, i-RevNet, which enabled the model to preserve all information of input through layers while still being capable of extracting a useful representation for classification.
+
+# 5 CONCLUSION
+
+We described a novel conditionally reversible network, CrevNet, for pixel-level prediction of future frames in videos. The originality of our model lies in our use of the reversible two-way autoencoder and the accompanying reversible predictive module. Such architectural design enables the model to preserve fine-grained information without significant memory and computation overhead. CrevNet achieves state-of-the-art results on both synthetic and real-world datasets. The subsequent detection experiments demonstrate the potential of CrevNet to be a continuous self-supervised learning system to enhance downstream CV tasks, as shown in the case of BERT (Devlin et al. (2018)) for NLP tasks.
+
+# REFERENCES
+
+Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: Endto-end speech recognition in english and mandarin. In International Conference on Machine Learning, pp. 173–182, 2016.
+
+Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextvp: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 753–769, 2018.
+
+Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6299–6308, 2017.
+
+Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 1174–1183, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http://proceedings.mlr.press/ v80/denton18a.html.
+
+Emily L Denton et al. Unsupervised learning of disentangled representations from video. In Advances in Neural Information Processing Systems, pp. 4417–4426, 2017.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
+
+Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014.
+
+Piotr Dollár, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009.
+
+Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 3354–3361. IEEE, 2012.
+
+Aidan N Gomez, Mengye Ren, Raquel Urtasun, and Roger B Grosse. The reversible residual network: Backpropagation without storing activations. In Advances in Neural Information Processing Systems, pp. 2211–2221, 2017.
+
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.
+
+IARAI. Traffic4cast: Traffic map movie forecasting. https://www.iarai.ac.at/ traffic4cast, 2019. Accessed: 2019-09-25.
+
+Jörn-Henrik Jacobsen, Arnold Smeulders, and Edouard Oyallon. i-revnet: Deep invertible networks. arXiv preprint arXiv:1802.07088, 2018.
+
+Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. arXiv preprint arXiv:1610.00527, 2016.
+
+Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.
+
+Manoj Kumar, Mohammad Babaeizadeh, Dumitru Erhan, Chelsea Finn, Sergey Levine, Laurent Dinh, and Durk Kingma. Videoflow: A flow-based generative model for video. arXiv preprint arXiv:1903.01434, 2019.
+
+Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1811–1820, 2019.
+
+Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018.
+
+Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1744–1752, 2017.
+
+Kun Liu, Wu Liu, Chuang Gan, Mingkui Tan, and Huadong Ma. T-c3d: Temporal convolutional 3d network for real-time action recognition. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018.
+
+Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pp. 21–37. Springer, 2016.
+
+William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. arXiv preprint arXiv:1605.08104, 2016.
+
+Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 716–731, 2018.
+
+Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan, Yu-Wing Tai, and Li Xu. Accurate single stage detector using recurrent rolling convolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5420–5428, 2017.
+
+Shaoqing Ren, Kaiming He, Ross Girprerhick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015.
+
+Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pp. 234–241. Springer, 2015.
+
+Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1874–1883, 2016.
+
+Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional lstm network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems, pp. 802–810, 2015.
+
+Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pp. 843–852, 2015.
+
+Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017.
+
+Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and S Yu Philip. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. In Advances in Neural Information Processing Systems, pp. 879–888, 2017.
+
+Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and Philip S Yu. PredRNN $^ { + + }$ : Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 5123–5132, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http://proceedings.mlr.press/ v80/wang18b.html.
+
+Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ B1lKS2AqtX.
+
+Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004.
+
+Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 129–137, 2017.
+
+# A CONVLSTM AND ST-LSTM
+
+The key equations of ConvLSTM are shown as belows.
+
+$$
+\begin{array} { r l } & { i _ { t } = \sigma \big ( W _ { x i } \ast \mathcal { X } _ { t } + W _ { h i } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { i } \big ) } \\ & { f _ { t } = \sigma \big ( W _ { x f } \ast \mathcal { X } _ { t } + W _ { h f } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { f } \big ) } \\ & { \mathcal { C } _ { t } ^ { l } = f _ { t } \circ \mathcal { C } _ { t - 1 } ^ { l } + i _ { t } \circ \operatorname { t a n h } \bigl ( W _ { x c } \ast \mathcal { X } _ { t } + W _ { h c } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { c } \bigr ) } \\ & { o _ { t } = \sigma \big ( W _ { x o } \ast \mathcal { X } _ { t } + W _ { h o } \ast \mathcal { H } _ { t - 1 } ^ { l } + W _ { c o } \circ \mathcal { C } _ { t } ^ { l } + b _ { o } \big ) } \\ & { \mathcal { H } _ { t } ^ { l } = o _ { t } \circ \operatorname { t a n h } ( \mathcal { C } _ { t } ^ { l } ) } \end{array}
+$$
+
+where $^ *$ denotes the convolution operator and $\circ$ denotes the Hadamard product. Based on ConvLSTM, spatiotemporal LSTM (ST-LSTM) in PredRNN adds another vertical memory flow to enhance the long-term temporal dependency as follow.
+
+$$
+\begin{array} { r l } & { i _ { t } = \sigma ( W _ { x i } \ast \mathcal { X } _ { t } + W _ { h i } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { i } ) } \\ & { f _ { t } = \sigma ( W _ { x f } \ast \mathcal { X } _ { t } + W _ { h f } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { f } ) } \\ & { \mathcal { C } _ { t } ^ { l } = f _ { t } \circ \mathcal { C } _ { t - 1 } ^ { l } + i _ { t } \circ \operatorname { t a n h } ( W _ { x c } \ast \mathcal { X } _ { t } + W _ { h c } \ast \mathcal { H } _ { t - 1 } ^ { l } + b _ { c } ) } \\ & { \dot { u } _ { t } ^ { l } = \sigma ( W _ { x i } ^ { r } \ast \mathcal { X } _ { t } + W _ { m i } \ast \mathcal { M } _ { t } ^ { l - 1 } + b _ { i } ^ { i } ) } \\ & { f _ { t } ^ { \prime } = \sigma ( W _ { x f } ^ { r } \ast \mathcal { X } _ { t } + W _ { m f } \ast \mathcal { M } _ { t } ^ { l - 1 } + b _ { f } ^ { \prime } ) } \\ & { \mathcal { M } _ { t } ^ { l } = f _ { t } ^ { \prime } \circ \mathcal { M } _ { t } ^ { l - 1 } + i _ { t } ^ { \prime } \circ \operatorname { t a n h } ( W _ { x m } \ast \mathcal { X } _ { t } + W _ { m m } \ast \mathcal { M } _ { t } ^ { l - 1 } + b _ { m } ) } \\ & { o _ { t } = \sigma ( W _ { x o } \ast \mathcal { X } _ { t } + W _ { h o } \ast \mathcal { H } _ { t - 1 } ^ { l } + W _ { c o } \circ \mathcal { C } _ { t } ^ { l } + W _ { m o } \circ \mathcal { M } _ { t } ^ { l } + b _ { o } ) } \\ & { \mathcal { H } _ { t } = o _ { t } \mathrm { o t } \operatorname { t a n h } ( W _ { 1 x } | C _ { t } ^ { l } , \mathcal { M } _ { t } ^ { l } | ) } \end{array}
+$$
+
+where blue part overlaps ConvLSTM. Note that $\mathcal { M } _ { t } ^ { l }$ usually receives information from the previous layer instead of the previous state and the special case is that $\boldsymbol { \mathcal { M } } _ { t } ^ { 1 }$ receives $\mathcal { M } _ { t - 1 } ^ { L }$ to constitute a zigzag information flow. As we can see, ST-LSTM basically doubles the size of feature map and the number of parameters compared with ConvLSTM.
+
+# B CONDITIONAL REVERSIBILITY
+
+As we mentioned in Section 2.2, conditional reversibility is an interesting property of our CrevNet. In this section, we will provide more details about it. Given $\hat { x } _ { t } ^ { 2 } , x _ { t } ^ { 1 }$ and $h _ { t - 1 } ^ { 1 }$ , the reversible predictive module can recover $\boldsymbol { x } _ { t } ^ { 2 }$ as follow
+
+$$
+\begin{array} { r l } & { h _ { t } ^ { 1 } = \mathrm { C o n v R N N } ( x _ { t } ^ { 1 } , h _ { t - 1 } ^ { 1 } ) } \\ & { g _ { t } = \phi ( W * h _ { t } ^ { 1 } + b ) } \\ & { x _ { t } ^ { 2 } = ( \hat { x } _ { t } ^ { 2 } - g _ { t } \odot h _ { t } ^ { 1 } ) \odot 1 / 1 - g _ { t } } \end{array}
+$$
+
+Ideally, if there is no numerical error during the calculation, we can get the perfect reconstruction of input by applying this inverse operation repeatedly. In practice, while the most of reverse generations are successful, the inevitable numerical error will still result in some failing cases, especially in the case of a very deep architecture because errors will be amplified layer by layer.
+
+C OBJECT DETECTION
+
+![](images/7b2150b3b04ec4bfef448c6064e3b09a1763f16086583fdf4984344b7d0c05cb.jpg)  
+Figure 7: Visualization of 2D Object Detection on KITTI.

parse/train/B1eY_pVYvB/B1eY_pVYvB_content_list.json

@@ -0,0 +1,1528 @@
+[
+    {
+        "type": "text",
+        "text": "EFFICIENT AND INFORMATION-PRESERVING FUTURE FRAME PREDICTION AND BEYOND ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            99,
+            821,
+            146
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wei ${ { \\bf { Y } } { \\bf { u } } ^ { 1 } }$ , Yichao $\\mathbf { L u } ^ { 1 }$ , Steve Easterbrook1, Sanja Fidler1,2,3 ",
+        "bbox": [
+            183,
+            169,
+            588,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1Department of Computer Science, University of Toronto 2Vector Institute, Canada   \n3NVIDIA   \n{gnosis,yichao,sme,fidler}@cs.toronto.edu ",
+        "bbox": [
+            184,
+            185,
+            555,
+            242
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            279,
+            544,
+            294
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Applying resolution-preserving blocks is a common practice to maximize information preservation in video prediction, yet their high memory consumption greatly limits their application scenarios. We propose CrevNet, a Conditionally Reversible Network that uses reversible architectures to build a bijective two-way autoencoder and its complementary recurrent predictor. Our model enjoys the theoretically guaranteed property of no information loss during the feature extraction, much lower memory consumption and computational efficiency. The lightweight nature of our model enables us to incorporate 3D convolutions without concern of memory bottleneck, enhancing the model’s ability to capture both short-term and long-term temporal dependencies. Our proposed approach achieves state-of-the-art results on Moving MNIST, Traffic4cast and KITTI datasets. We further demonstrate the transferability of our self-supervised learning method by exploiting its learnt features for object detection on KITTI. Our competitive results indicate the potential of using CrevNet as a generative pre-training strategy to guide downstream tasks. ",
+        "bbox": [
+            233,
+            313,
+            766,
+            506
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            536,
+            334,
+            553
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep learning has enjoyed tremendous success in recent years due to its ability to capture complex dependencies and non-linearities in large datasets (Krizhevsky et al. (2012); He et al. (2016); Gomez et al. (2017)). Excellent performance has been achieved on a wide range of supervised machine learning tasks, ranging from image classification (He et al. (2016)) and object detection (Ren et al. (2015)) to speech recognition (Amodei et al. (2016)). Despite the significant breakthrough in supervised learning, the potential of applying deep architectures to unsupervised learning problems remains largely unexplored. Lately there has been a surge of interest in the task of video prediction, i.e., to predict future frames of a video sequence (Wang et al. (2017; 2018); Denton et al. (2017); Denton & Fergus (2018); Villegas et al. (2017); Lee et al. (2018)). The significance of video prediction primarily lies in its potential of discovering dynamics in the physical world. The self-supervised nature of video prediction aligns well with how humans learn, without requiring large amounts of labeled data. In addition, videos can provide an abundant and virtually unlimited source of visual information. This allows video prediction models to serve as a generative pre-training strategy of feature representation learning for a variety of downstream supervised tasks. ",
+        "bbox": [
+            174,
+            569,
+            825,
+            763
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To date, most of the existing models for video prediction employ a hybrid of convolutional and recurrent layers as the underlying architecture (Wang et al. (2017); Shi et al. (2015); Lotter et al. (2016)). Such architectural design enables the model to simultaneously exploit the ability of convolutional units to model spatial relationships and the potential of recurrent units to capture temporal dependencies. Despite their prevalence in the literature, classical video prediction architectures suffer from two major limitations. Firstly, in dense prediction tasks such as video prediction, models are required to make pixel-wise predictions, which emphasizes the demand for the preservation of information through layers. Prior works attempt to address such demand through the extensive use of resolution-preserving blocks (Wang et al. (2017; 2018); Kalchbrenner et al. (2016)). Nevertheless, these resolution-preserving blocks are not guaranteed to preserve all the relevant information, and they greatly increase the memory consumption and computational cost of the models. The second drawback of existing video prediction models is that they cannot efficiently take advantage of 3D convolutions, as that would make these already cumbersome architectures even larger. 3D convolutions have been shown to be a very effective alternative to RNNs to capture temporal relations in a variety of video tasks (Liu et al. (2018); Carreira & Zisserman (2017)), and thus desirable to exploit. ",
+        "bbox": [
+            174,
+            771,
+            825,
+            924
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2814f3e7c41cd24b84d6f7d7d8910bd06c6aa74d5183b36c2eadb0881c09f737.jpg",
+        "image_caption": [
+            "Figure 1: The pipeline of our proposed CrevNet where a single two-way autoencoder serves as both encoder and decoder. CrevNet first observes a warm-up video sequence and then starts a multi-frame video prediction without refeeding its own predictions. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            204,
+            101,
+            790,
+            262
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            338,
+            825,
+            393
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, reversible architectures (Dinh et al. (2014); Gomez et al. (2017); Jacobsen et al. (2018)) have attracted attention due to their light memory demand and their information preserving property by design. However, the effectiveness of reversible models remains greatly unexplored in the video literature. In this paper, we introduce a novel, conditionally reversible video prediction model, CrevNet, in the sense that when conditioned on previous hidden states, it can exactly reconstruct the input from its predictions. The contribution of this work can be summarized as follows: ",
+        "bbox": [
+            174,
+            400,
+            825,
+            484
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "• We introduce a two-way autoencoder that uses the forward and backward passes of an invertible network as encoder and decoder (Fig 1). The volume-preserving two-way autoencoder not only greatly reduces the memory demand and computational cost, but also enjoys the theoretically guaranteed property of no information loss. The lightweight nature of our model enables us to incorporate 3D convolutions without concern of memory bottleneck. We propose the reversible predictive module (RPM), as illustrated in Fig 2b, which extends the reversibility from spatial to temporal domain. RPM, together with the two-way autoencoder, provides a conditionally reversible architecture (CrevNet) for spatiotemporal learning. CrevNet achieves the state-of-the-art results on Moving MNIST, Traffic4cast and KITTI. We evaluate the effectiveness of features learnt from self-supervision by adapting our CrevNet for object detection on KITTI. Our competitive results indicate the potential of using CrevNet as a generative pre-training strategy to guide downstream CV tasks. ",
+        "bbox": [
+            215,
+            494,
+            826,
+            670
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 APPROACH",
+        "text_level": 1,
+        "bbox": [
+            174,
+            689,
+            299,
+            705
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We first outline the general pipeline of our method. Our CrevNet consists of two subnetworks, an autonencoder network with an encoder $\\mathcal { E }$ , decoder $\\mathcal { D }$ and a recurrent predictor $\\mathcal { P }$ bridging encoder and decoder. Let $\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { w \\times h \\times c }$ represent the $t _ { \\mathrm { t h } }$ frame in video $x$ , where $w , h .$ , and $c$ denote its width, height, and the number of channels. Given $x _ { 0 : t - 1 }$ , the model predicts the next frame $\\hat { x } _ { t }$ as follows: ",
+        "bbox": [
+            173,
+            719,
+            825,
+            776
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "img_path": "images/61b3e3a7b453b84623249f224e30da8594b7e2c1c2b38c1201522f8bf16742ea.jpg",
+        "text": "$$\n\\hat { x } _ { t } = \\mathcal { D } ( \\mathcal { P } ( \\mathcal { E } ( x _ { t - 1 } ) | x _ { 0 : t - 2 } ) )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            401,
+            779,
+            594,
+            795
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In the case of 3D convolution, $\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { k \\times w \\times h \\times c }$ denotes the short video clip from $t$ to $t + k - 1$ instead of a single frame at timestep $t$ , where $k$ is the temporal dimension of input or output. During the multi-frame generation process without access to the ground truth frames, the model uses its previous predictions instead. ",
+        "bbox": [
+            174,
+            796,
+            825,
+            853
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 THE INVERTIBLE TWO-WAY AUTOENCODER ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            869,
+            521,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose a bijective two-way autoencoder based on the additive coupling layer introduced in NICE (Dinh et al. (2014)). We begin with describing the building block of the two-way autoencoder ",
+        "bbox": [
+            176,
+            895,
+            823,
+            924
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/82c62b6a2698f2565cb8ca6a3f84285551bbb5913c9507f6e1009f46cb4a594c.jpg",
+        "image_caption": [
+            "Figure 2: The network architecture of CrevNet (Better viewed in color). The input video frames are first reshaped and split channelwise into two groups. These two groups are passed to the two-way autoencoder (a) for feature extraction, and then to the predictor made up of multiple reversible predictive modules (b). The transformed high-level features produced by predictor are then passed back through the decoding pass of (a), shown here as a representative block (c) to yield its prediction. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            207,
+            101,
+            790,
+            285
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Fig 2a). Formally, the input $x$ is first reshaped and split channelwise into two groups, denoted as $x ^ { 1 }$ and $x ^ { 2 }$ . During the forward pass of each building block, one group, e.g. $x ^ { 1 }$ , passes through several convolutions and activations and is then added to another group, $x ^ { \\bar { 2 } }$ , like a residual block: ",
+        "bbox": [
+            173,
+            390,
+            823,
+            433
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/d95aecb8c2c14af02e752cc513afd78afeaacd1d65061f104df5d9e2f2402584.jpg",
+        "text": "$$\n\\hat { x } ^ { 2 } = x ^ { 2 } + \\mathcal { F } _ { 1 } ( x ^ { 1 } ) \\qquad \\hat { x } ^ { 1 } = x ^ { 1 } + \\mathcal { F } _ { 2 } ( \\hat { x } ^ { 2 } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            339,
+            436,
+            656,
+            455
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where $\\mathcal { F }$ is a composite non-linear operator consisting of convolutions and activations, and ${ \\hat { x } } ^ { 1 }$ and ${ \\hat { x } } ^ { 2 }$ are the updated $x ^ { 1 }$ and $x ^ { 2 }$ . Note that $x ^ { 1 }$ and $x ^ { 2 }$ can be simply recovered from ${ \\hat { x } } ^ { 2 }$ and ${ \\hat { x } } ^ { 1 }$ by the inverse computation (Fig 2c) as follows: ",
+        "bbox": [
+            178,
+            463,
+            826,
+            506
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/9380049f9f3aed3c215b5c5411bf98e9d94ec930e6518078f41fcbb1bbbd2d77.jpg",
+        "text": "$$\nx ^ { 1 } = \\hat { x } ^ { 1 } - \\mathcal { F } _ { 2 } ( \\hat { x } ^ { 2 } ) \\qquad x ^ { 2 } = \\hat { x } ^ { 2 } - \\mathcal { F } _ { 1 } ( x ^ { 1 } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            341,
+            511,
+            658,
+            530
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Multiple building blocks are stacked in an alternating fashion between $x ^ { 1 }$ and $x ^ { 2 }$ to construct a two-way autoencoder, as shown in Fig 2a. A series of the forward and inverse computations builds a one-to-one and onto, i.e. bijective , mapping between the input and features. Such invertibility ensures that there is no information loss during the feature extraction, which is presumably more favorable for video prediction since the model is expected to restore the future frames with finegrained details. To enable the invertibility of the entire autoencoder, our two-way autoencoder uses a bijective downsampling, pixel shuffle layer (Shi et al. (2016)), that changes the shape of feature from $( w , h , c )$ to $( w / n , \\bar { h } / n , \\bar { c } \\times n ^ { 2 } )$ . The resulting volume-preserving architecture can greatly reduce its memory consumption compared with the existing resolution-preserving methods. ",
+        "bbox": [
+            173,
+            536,
+            825,
+            664
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We further argue that for generative tasks, e.g. video prediction, we can effectively utilize a single two-way autoencoder, and to use its forward and backward pass as the encoder and the decoder, respectively. The predicted frame $\\hat { x } _ { t }$ is thus given by ",
+        "bbox": [
+            174,
+            670,
+            825,
+            712
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/a3839b7c74b4c0b80a36d8f0e702cd15824e7b5565ad9f60df012db0d06e7793.jpg",
+        "text": "$$\n\\hat { x } _ { t } = \\mathcal { E } ^ { - 1 } ( \\mathcal { P } ( \\mathcal { E } ( x _ { t - 1 } ) | x _ { 0 : t - 2 } ) )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            395,
+            717,
+            602,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where ${ { \\mathcal { E } } ^ { - 1 } }$ is the backward pass of $\\mathcal { E }$ . Our rationale is that, such setting would not only reduce the number of parameters in the model, but also encourage the model to explore the shared feature space between the inputs and the targets. As a result, our method does not require any form of information sharing, e.g. skip connection, between the encoder and decoder. In addition, our two-way autoencoder can enjoy a lower computational cost at the multi-frame prediction phase where the encoding pass is no longer needed and the predictor directly takes the output from previous timestep as input, as shown in Fig 1, since $\\mathcal { E } ( \\mathcal { E } ^ { - 1 } )$ is an identity mapping . ",
+        "bbox": [
+            174,
+            741,
+            825,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 REVERSIBLE PREDICTIVE MODULE ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            856,
+            462,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we describe the second part of our video prediction model, the predictor $\\mathcal { P }$ , which computes dependencies along both the space and time dimensions. Although the traditional stackedConvRNN layers architecture is the most straightforward choice of predictor, we find that it fails to establish a consistent temporal dependency when equipped with our two-way autoencoder through experiments. Therefore, we propose a novel reversible predictive module (RPM), which can be regarded as a recurrent extension of the two-way autoencoder. In the RPM, we substitute all standard convolutions with layers from the ConvRNN family (e.g. ConvLSTM or spatiotemporal LSTM) and introduce a soft attention (weighting gates) mechanism to form a weighted sum of the two groups instead of the direct addition. The main operations of RPM used in this paper are given as follows: ",
+        "bbox": [
+            174,
+            881,
+            825,
+            924
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            103,
+            823,
+            188
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ConvRNN ",
+        "bbox": [
+            622,
+            195,
+            691,
+            209
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/530f97d8faf577eafc39dd240b00c9a6595d8f264fd0aafc776adb39772c3705.jpg",
+        "text": "$$\n\\begin{array} { r l } & { h _ { t } ^ { 1 } = \\mathrm { C o n v R N N } ( x _ { t } ^ { 1 } , h _ { t - 1 } ^ { 1 } ) } \\\\ & { g _ { t } = \\phi ( W _ { 2 } * \\mathrm { R e L U } ( W _ { 1 } * h _ { t } ^ { 1 } + b _ { 1 } ) + b _ { 2 } ) } \\\\ & { \\hat { x _ { t } ^ { 2 } } = ( 1 - g _ { t } ) \\odot x _ { t } ^ { 2 } + g _ { t } \\odot h _ { t } ^ { 1 } } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            259,
+            191,
+            539,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Attention module ",
+        "bbox": [
+            622,
+            217,
+            736,
+            231
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Weighted sum ",
+        "bbox": [
+            622,
+            238,
+            715,
+            252
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $\\boldsymbol { x } _ { t } ^ { 1 }$ and $\\boldsymbol { x } _ { t } ^ { 2 }$ denote two groups of features at timestep $t$ , $h _ { t } ^ { 1 }$ denote the hidden states of ConvRNN layer, $\\phi$ is sigmoid activation, $^ *$ is the standard convolution operator and $\\odot$ is the Hadamard product. The architecture of reversible predictive module is also shown in Fig 2b. RPM adopts a similar architectural design as the two-way autoencoder to ensure a pixel-wise alignment between the input and the output, i.e. each position of features can be traced back to certain pixel, and thus make it compatible with our two-way autoencoder. It also mitigates the vanishing gradient issues across stacked layers since the coupling layer provides a nice property w.r.t. the Jacobian (Dinh et al. (2014)). In addition, the attention mechanism in the RPM enables the model to focus on objects in motion instead of background, which further improves the video prediction quality. Similarly, multiple RPMs alternate between the two groups to form a predictor. We call this predictor conditionally reversible since, given $h _ { t - 1 }$ , we are able to reconstruct $x _ { t - 1 }$ from $\\hat { x } _ { t }$ if there are no numerical errors: ",
+        "bbox": [
+            173,
+            262,
+            825,
+            416
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/3de1193dd535592a443543cd614eef93c5c7e331cdfd62cac2995db79a841eba.jpg",
+        "text": "$$\nx _ { t - 1 } = \\mathcal { E } ^ { - 1 } ( \\mathcal { P } ^ { - 1 } ( \\mathcal { E } ( \\hat { x } _ { t } ) | h _ { t - 1 } ) )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            392,
+            421,
+            606,
+            440
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $\\mathcal { P } ^ { - 1 }$ is the inverse computation of the predictor $\\mathcal { P }$ . We name the video prediction model using two-way autoencoder as its backbone and RPMs as its predictor CrevNet. Another key factor of RPM is the choice of ConvRNN. In this paper, we mainly employ ConvLSTM (Shi et al. (2015)) and spatiotemporal LSTM (ST-LSTM, Wang et al. (2017)) to enable a fair comparison with baselines. ",
+        "bbox": [
+            174,
+            446,
+            825,
+            503
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.3 3D CONVOLUTIONS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            520,
+            354,
+            534
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3D convolutions are proposed to address the shortcomings of standard 2D convolutions. The major difference between 2D-CNNs and 3D-CNNs is that at each time step 2D-CNNs take as input one video frame, while 3D-CNNs read in and output a short video clip containing $k$ continuous video frames. By applying convolutions on the temporal dimension along with the spatial dimension, models equipped with 3D convolution filters can not only extract representative spatiotemporal features, but also learn to produce consistent video clip at each generation, which further improve the quality of long-term prediction. In some cases, e.g. sequences are too short, we will use 2 consecutive frames stacked in the channel dimension instead as input at each timestep to assemble a valid warm-up sequence for ConvRNN. ",
+        "bbox": [
+            174,
+            545,
+            825,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 EXPERIMENTS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            691,
+            326,
+            708
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 LONG-TERM PREDICTION—MOVING MNIST ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            723,
+            529,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Moving MNIST (Srivastava et al. (2015)) is a synthetically generated dataset that contains an infinite number of sequences of length 20. Each sequence shows how 2 digits move at a constant speed and bounce inside a $6 4 \\times 6 4$ frame, where each handwritten digit is randomly sampled from the MNIST dataset. By assigning different initial locations and velocities to each digit, it is possible to generate an unlimited number of sequences, thus enabling us to accurately evaluate the performance of each model without the concern of data insufficiency issues. In the default setting, models are trained to predict 10 future frames after observing 10 prior frames in the sequence. Although the dynamics of Moving MNIST seems to be simple at first glance. It is quite hard to generate consistent future frames in the task of long-term prediction as digits can bounce or occlude each other frequently. ",
+        "bbox": [
+            174,
+            748,
+            825,
+            875
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Setup: The general architecture of CrevNet used on Moving MNIST is composed of a 36-layer two-way autoencoder and 8 RPMs. All variants of CrevNet are trained by using the Adam optimizer with a starting learning rate of $5 \\times 1 0 ^ { - 4 }$ to minimize MSE. The training process is stopped ",
+        "bbox": [
+            176,
+            882,
+            823,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/1aa6e3d4cafbe149996ed511c28554bb07fd62b3aaffd7270ba4ba6cd6c446ea.jpg",
+        "table_caption": [
+            "after 300, 000 iterations with the batch size of 16 and evaluated with a fixed test set containing 5, 000 sequences. To ensure that all samples in the test set are unseen by the model, digits in the training set and the testing set are separately sampled from two mutually exclusive subsets of MNIST. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">MNIST-2</td><td colspan=\"2\">MNIST-3 MSE</td><td rowspan=\"2\">Memory (1 sample)</td><td rowspan=\"2\">FLOPS (1 frame)</td></tr><tr><td>SSIM</td><td>MSE</td><td>Human</td><td>SSIM</td><td></td></tr><tr><td>ConvLSTM (Shi et al. (2015))</td><td>0.707</td><td>103.3</td><td>0.923</td><td>0.695</td><td>127.3</td><td>1043 MB</td><td>107.4 G</td></tr><tr><td>FRNN (Oliu et al. (2018))</td><td>0.819</td><td>68.4</td><td>0.848</td><td>0.791</td><td>90.4</td><td>717MB</td><td>80.1G</td></tr><tr><td>VPN (Kalchbrenner et al. (2016))</td><td>0.870</td><td>70.0</td><td>0.831</td><td>0.820</td><td>85.6</td><td>5206MB</td><td>309.6 G</td></tr><tr><td>PredRNN (Wang et al. (2017))</td><td>0.869</td><td>56.8</td><td>0.837</td><td>0.822</td><td>83.1</td><td>1666 MB</td><td>192.9 G</td></tr><tr><td>PredRNN++ (Wang et al. (2018))</td><td>0.898</td><td>46.5</td><td>0.781</td><td>0.864</td><td>68.4</td><td>2017MB</td><td>106.8 G</td></tr><tr><td>E3D-LSTM (Wang et al. (2019))</td><td>0.910</td><td>41.3</td><td>0.706</td><td>0.870</td><td>62.4</td><td>2695 MB</td><td>381.3 G</td></tr><tr><td>CrevNet+ConvLSTM</td><td>0.928</td><td>38.5</td><td>0.602</td><td>0.886</td><td>57.2</td><td>130 MB</td><td>0.919 G</td></tr><tr><td>CrevNet + ST-LSTM</td><td>0.949</td><td>22.3</td><td>0.558</td><td>0.916</td><td>40.6</td><td>195MB</td><td>1.618 G</td></tr></table>",
+        "bbox": [
+            171,
+            157,
+            825,
+            301
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Quantitative evaluation of different methods on Moving MNIST. All metrics are averaged over the 10 predictions. Lower MSE and higher SSIM indicates better prediction accuracy. ",
+        "bbox": [
+            174,
+            310,
+            823,
+            340
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compare CrevNet to six popular benchmark models from the literature: (i) ConvLSTM (Shi et al. (2015)), (ii) FRNN (Oliu et al. (2018)), (iii) VPN (Kalchbrenner et al. (2016)), (iv) PredRNN (Wang et al. (2017)) , (v) PredRNN $^ { + + }$ (Wang et al. (2018)), and (vi) E3D-LSTM (Wang et al. (2019)),. All baselines are implemented and optimized by following their corresponding protocols. To test our model in a more challenging setting, we also extend Moving MNIST to a 3-digit version where digits are more likely to occlude each other. ",
+        "bbox": [
+            173,
+            354,
+            825,
+            438
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Results: The performance of each model in terms of per-frame MSE and the Structural Similarity Index Measure (SSIM) (Wang et al. (2004)) is presented in Table 1. CrevNet outperforms all previous methods by a wide margin on both metrics while memory consumption of all CrevNet variants is significantly lower than that of other baselines. In particular, CrevNet with ConvLSTM only uses 130 MB memory per sample and is still capable of yielding results better than any baselines. ",
+        "bbox": [
+            174,
+            444,
+            825,
+            515
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To analyze the contribution of each module, we conduct an ablation study on both ConvLSTM and ST-LSTM with respect to 3D convolution, two-way autoencoder and RPM and summarize the results in Table 2. Note that we do not include the quantitative results of the combination of two-way autoencoder and stack-ConvRNN predictor because it fails to produce consistent long-term generations and we choose UNet (Ronneberger et al. (2015)) as an alternative to our two-way autoencoder. We can observe a significant improvement over ConvLSTM after we embed it into our CrevNet framework, indicating the effectiveness of reversible architectures. Also, integrating 3D convolution can consistently enhance the performance of all architectures. To further show the superior performance of CrevNet, we evaluate it on a harder 3-digit setting. Results are shown in the right column of Table 1. Compared with the 2-digit setting, all models suffer a deterioration in quantitative performance due to the more frequent occurrence of overlapping digits. Nevertheless, our CrevNet still achieves the best result. ",
+        "bbox": [
+            173,
+            522,
+            825,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7ed8d7ba2d38f9bb78e2fb72587a3164387b1cd5d302ac1569447f621063ef89.jpg",
+        "image_caption": [
+            "Figure 3: An extremely hard sequence of Moving MNIST where two digits are continuously overlapped during the warm-up phase. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            705,
+            482,
+            837
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/4208b05051866c5c53c107bf9ad21ea65eed4d4e2e2e3847cf81ae1e953d4187.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>ConvLSTM MSE SSIM|</td><td>ST-LSTM MSE SSIM</td></tr><tr><td>Stacked-RNN</td><td>103.3 0.707</td><td>56.8 0.869</td></tr><tr><td>3D Stacked-RNN</td><td>85.8 0.785</td><td>46.2 0.878</td></tr><tr><td>UNet+StackedRNN 83.5</td><td>0.793</td><td>58.8 0.865</td></tr><tr><td>UNet+RPM</td><td>63.4 0.855</td><td>50.2 0.896</td></tr><tr><td>CrevNet w/o 3D</td><td>50.2 0.888</td><td>40.4 0.916</td></tr><tr><td>CrevNet + 3D</td><td>38.5 0.928</td><td>22.3 0.949</td></tr></table>",
+        "bbox": [
+            509,
+            705,
+            825,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2: An ablation study w.r.t. 3D convolution, RPM and two-way autoencoder. All metrics are averaged over the 10 predictions. ",
+        "bbox": [
+            509,
+            847,
+            825,
+            888
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In Fig 3 , our qualitative analysis shows how each model performs on an extremely hard case of Moving MNIST where two digits are continuously overlapped during the warm-up phase. ",
+        "bbox": [
+            171,
+            895,
+            826,
+            924
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d5e85adb45b4e2b127effb0619d1ce9ff93eb2c78c8c9d4df842a4141a5c165.jpg",
+        "image_caption": [
+            "Figure 4: The visual comparison of Traffic4cast. The red boxes track some dynamics successfully captured by our CrevNet. Better viewed large. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            102,
+            820,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As we can see, our model is the only model that can differentiate the overlapping digits. The information-preserving property of the two-way autoencoder enables our method to reconstruct every fine detail of moving digits after occlusion while baselines typically only restore the basic shape of these numbers. In fact, our CrevNet works almost perfectly on Moving MNIST, with most of its generations being visually indistinguishable from groundtruth. ",
+        "bbox": [
+            174,
+            304,
+            825,
+            375
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We perform a human study to assess the fidelity of the video clips generated by different models. We presented pairs of video clips to human judges, where each pair consists of a video clip from the test set together with the prediction generated by the model. The judges were asked to decide which of the two video clips is more likely to be the groundtruth. To make each trail blind, the judges were not informed which model is used for generation and two sequences were randomly displayed on either side of the screen. We totally collected 2439 responses made by 58 human subjects and then calculated the probability that human judges answered correctly. The results are reported in Table 1. The accuracy of $5 5 . 8 \\%$ suggests that subjects could hardly detect the difference and their decisions were very close to random guesses. ",
+        "bbox": [
+            173,
+            395,
+            825,
+            520
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.2 SHORT-TERM PREDICTION—TRAFFIC FLOW FORECASTING ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            540,
+            619,
+            554
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Next, we evaluate our model on a more complicated real-world dataset, Traffic4cast (IARAI (2019)), which collects the traffic statuses of 3 big cities over a year at a 5-minute interval. Traffic forecasting can be straightforwardly defined as video prediction task by its spatiotemporal nature. However, this dataset is quite challenging for the following reasons. (1). High resolution: The frame resolution of Traffic4cast is $4 9 5 \\times 4 3 6$ , which is the highest among all datasets. Existing resolution-preserving methods can hardly be adapted to this dataset since they all require extremely large memory and computation. Even if these models can be fitted in GPUs, they still do not have large enough receptive fields to capture the meaningful dynamics as vehicles can move up to 100 pixels between consecutive frames. (2). Complicated nonlinear dynamics: Valid data points only reside on the hidden roadmap of each city, which is not explicitly provided in this dataset. Moving vehicles on these curved roads along with tangled road conditions will produce very complex nonlinear behaviours. It also involves many unobservable conditions or random events like weather and car accidents. ",
+        "bbox": [
+            174,
+            565,
+            825,
+            733
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Setup: Each frame in Traffic4cast dataset is a $4 9 5 \\times 4 3 6 \\times 3$ heatmap, where the last dimension records 3 traffic statuses representing volume, mean speed and major direction at given location. The architecture of CrevNet is the same as the one we used on Moving MNIST. As we mentioned before, the existing resolution-preserving methods cannot handle such high resolution input. Thus, to make the comparison possible, we add U-Net encoder-decoder to the baseline models including ConvLSTM and ST-LSTM. We train each model to predict next 3 frames (the next 15 minutes) from 9 observations and evaluate prediction with MSE criterion. ",
+        "bbox": [
+            174,
+            739,
+            825,
+            838
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/95fbe680383c4a027540e26210dcfbb0d13f584aa464962f6bed9037f37a43bd.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Quantitative evaluation on Traffic4cast. Lower MSE indicates better prediction accuracy. "
+        ],
+        "table_body": "<table><tr><td>Model</td><td>CrevNet</td><td>UNet+ST-LSTM</td><td>UNet+ConvLSTM丨Best Team</td><td></td><td>2nd Best Team</td></tr><tr><td>MSE</td><td>9.340×10-3</td><td>9.725× 10-3</td><td>9.846 ×10-3</td><td>9.559×10-3</td><td>9.717 × 10-3</td></tr></table>",
+        "bbox": [
+            171,
+            853,
+            826,
+            895
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d59d0905d69ce4a69c2e0d6fb0d4d47d0f31ad6570a2d67a3b99ee83bd6a76ef.jpg",
+        "image_caption": [
+            "Figure 5: The visual comparison of next-frame predictions on Caltech Pedestrian. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            176,
+            103,
+            818,
+            449
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results: The quantitative comparison including the best two results on the leaderboard before the submission of this paper is provided in Table 3. Unlike all previous state-of-the-art methods, CrevNet does not suffer from high memory consumption so that we were able to train our model in a single V100 GPU. The invertibility of two-way autoencoder preserves all necessary information for spatiotemporal learning and allows our model to generate sharp and reasonable predictions. As illustrated in Fig 4, our model can identify and remember the hidden roadmap of each city through the learning of complicated nonlinear dynamics and accurately predict how traffic system will evolve. ",
+        "bbox": [
+            174,
+            494,
+            825,
+            592
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.3 NEXT-FRAME PREDICTION AND BEYOND—CAR-MOUNTED CAMERA VIDEO ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            608,
+            735,
+            622
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The real-world videos are usually long-term unpredictable because of the intrinsic randomness and the lack of necessary information. Thus, the common practice for datasets like KITTI (Geiger et al. (2012)), a car-mounted camera video dataset, is to perform next-frame prediction. In this section, we further demonstrate the superior performance of our CrevNet by conducting experiments on KITTI and Caltech Pedestrian (Dollár et al. (2009)). Compared with the previous two settings, carmounted camera videos dataset presents another level of difficulty for video prediction as it describes various nonlinear three-dimensional dynamics of multiple moving objects including backgrounds. Furthermore, as our well-trained model is capable of generating authentic future frames, it should spontaneously learn at least the shape and location of all moving objects, which indicates that the learnt features are very informative for downstream tasks. For example, in the case of object detection, these features can be incorporated to estimate more accurate locations and sizes of bounding boxes. Therefore, we also explore the effectiveness of our self-supervised learning method on the 2D object detection on KITTI. ",
+        "bbox": [
+            174,
+            633,
+            825,
+            814
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 VIDEO PREDICTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            830,
+            362,
+            843
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Setup: We follow the same protocol used in PredNet (Lotter et al. (2016)) for preprocessing and evaluation. We first center-crop all video frames and resize them into $1 2 8 \\times 1 6 0$ . We compare our proposed method with 4 state-of-the-art benchmark models. Models are trained on KITTI dataset to predict the next frame after 10-frame warm-up and are evaluated on Caltech Pedestrian. The architecture of CrevNet used on KITTI is composed of a 48-layer two-way autoencoder and 40 RPMs. Note that it is the memory efficiency of our method that allows us to deploy such deep model. Since our model also possess good capability of long-term prediction. We add a 12-frame prediction comparison with CycleGAN (Kwon & Park (2019)) and PredNet (Lotter et al. (2016)). ",
+        "bbox": [
+            174,
+            854,
+            825,
+            922
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0deecc352bfbe70c65085939887ec28bba178661485724005c1bcff8bc8ec03e.jpg",
+        "image_caption": [
+            "Figure 6: The visual comparison of 12-frame prediction on Caltech Pedestrian. Notice how well CrevNet captures the detail and geometry of the buildings in the background, and the overall shading. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            179,
+            102,
+            820,
+            407
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            468,
+            825,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f84c2ec775ee9046b89ae75fff242d4d36554a7ec0dca297d612d9e21ac999d3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td colspan=\"2\">Next-Frame</td><td>3rd</td><td>6th</td><td>9th</td><td>12th</td><td>Average</td></tr><tr><td></td><td>PSNR</td><td>SSIM</td><td colspan=\"5\">SSIM</td></tr><tr><td>Copy-Last-Frame Dual Motion GAN (Liang et al. (2017))</td><td>23.3</td><td>0.779 0.899</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ContextVP (Byeon et al. (2018)) PredNet (Lotter et al. (2016))</td><td>28.7 27.6</td><td>0.921 0.905</td><td>0.72</td><td>0.66</td><td>0.61</td><td>0.58</td><td>0.701</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CycleGAN (Kwon &amp; Park (2019))</td><td>29.2</td><td>0.919</td><td>0.83</td><td>0.73</td><td>0.67</td><td>0.63</td><td>0.752</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CrevNet</td><td>29.3</td><td>0.925</td><td>0.84</td><td>0.76</td><td>0.70</td><td>0.65</td><td>0.776</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></table>",
+        "bbox": [
+            176,
+            525,
+            820,
+            661
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Quantitative evaluation of different methods on the Caltech Pedestrian dataset. Higher PSNR or SSIM means better prediction accuracy. ",
+        "bbox": [
+            173,
+            671,
+            823,
+            699
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Results: Performance of different models in terms of PSNR and SSIM is displayed in Table 4. CrevNet outperforms all baselines in both next-frame and multi-frame prediction regimes. Visual comparisons are provided in Fig 5 and Fig 6. Especially, in the case of 12-frame generation, we can observe that compared with our method, PredNet suffers severely from the famous error propagation of RNN issue while CycleGAN produces realistic yet physically inconsistent predictions. ",
+        "bbox": [
+            174,
+            715,
+            825,
+            786
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 2D OBJECT DETECTION ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            801,
+            397,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Setup: KITTI provides three prior frames of unlabeled data for each labeled image. This allows us to run our CrevNet to extract useful spatiotemporal features for object detection. All video sequences were recorded at $1 0 \\ : \\mathrm { H z }$ with resolution of $1 2 4 2 \\times 3 7 5$ . We first resize each frame to $4 1 6 \\times 1 2 8$ and finetune our best model on the video prediction task solely. The combinations of features extracted by our two-way autoencoder and attention masks of the target frame are then fed into the detection head for the further training. Note that we do not update the weights of CrevNet at this stage to purely demonstrate the power of self-supervised learning. Two image-based detection models, SqueezeDet and RRC, are compared as baselines. We also add an experiment on transfer learning of features learnt by PredNet on KITTI as comparison. To be consistent with related work, we use SSD (Liu et al. (2016)) as detection head. ",
+        "bbox": [
+            174,
+            825,
+            825,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            103,
+            825,
+            146
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/4b6cfaf8ca9ffb0affdd5164b8168c8fc38add6fd288de6a9eb46b2929055867.jpg",
+        "table_caption": [
+            "Table 5: Quantitative evaluation of different methods on 2D KITTI detection in term of Average Precision (AP). The numbers in the \"Best Results on the Leaderboard \" are achieved by different best models on each class according to the default ranking mechanism on the Leaderboard. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">Add-ons</td><td colspan=\"3\">Car L</td><td colspan=\"3\">Pedestrain M</td><td colspan=\"3\">Cyclist</td><td rowspan=\"2\">mAP</td></tr><tr><td>E</td><td>M</td><td>H</td><td>E</td><td></td><td>H</td><td>E</td><td>M</td><td>H</td></tr><tr><td>SqueezeDet RRC</td><td>(Wu et al. (2017)) (Ren et al. (2017))</td><td>90.4 90.61</td><td>87.1 90.23</td><td>78.9 87.44</td><td>81.4 84.16</td><td>71.3 75.33</td><td>68.5 70.39</td><td>87.6 84.96</td><td>80.3 76.49</td><td>78.1 65.46</td><td>80.4 80.56</td></tr><tr><td colspan=\"2\">Best Results on the Leaderboard</td><td>91.96</td><td>91.97</td><td>84.57</td><td>88.27</td><td>81.73</td><td>75.29</td><td>84.28</td><td>79.24</td><td>71.22</td><td>83.17</td></tr><tr><td>PredNet</td><td>(Lotter et al. (2016))</td><td>59.05</td><td>41.61</td><td>37.53</td><td>50.88</td><td>47.51</td><td>43.44</td><td>46.25</td><td>43.79</td><td>38.66</td><td>45.41</td></tr><tr><td>CrevNet</td><td>attention mask + extracted features</td><td>91.53 91.94</td><td>90.95 91.84</td><td>85.71 85.97</td><td>89.31 89.66</td><td>82.55 83.17</td><td>75.21 75.80</td><td>85.51 87.33</td><td>80.41 80.91</td><td>71.52 72.21</td><td>83.63 84.31</td></tr></table>",
+        "bbox": [
+            176,
+            160,
+            821,
+            276
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Results: The results of all experiments and baselines can be found in Table 5. Surprisingly, our CrevNet even outperforms the combination of the best model on each class in term of mAP. Since our model is capable of capturing the motion information, it is sensitive to the small (hard) moving objects. However, the motion information alone is not sufficient for object detection due to the appearance of relatively static objects. Therefore, we can observe a performance boost after we incorporate the features extracted by our two-way autoencoder. Another advantage of our method is that it can provide a better localization of bounding box since the learnt features of CrevNet remain the pixel-wise alignment with the input and output frame. Finally, thanks to the lightweight nature of our CrevNet, our best detection model can run at 6.8 FPS at the testing time. ",
+        "bbox": [
+            173,
+            347,
+            825,
+            472
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4 RELATED WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            496,
+            341,
+            511
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Deep Learning in Video Prediction: Mainstream video prediction models can mostly be categorized into two frameworks, stacked ConvRNNs and encoder-predictor-decoder models. The former framework attempts to design a new spatiotemporal module and then stacks multiple such modules to form the final model, while the latter usually utilizes an autoencoder to project video frames into their latent representations and then employs a recurrent neural network to model the temporal transformations. PredNet (Lotter et al. (2016)) is a good representative of stacked ConvRNNs framework. In PredNet, each ConvLSTM layer produces a layer-specific prediction at every time step to transmit an error term to the next layer. This model works well for predicting the next frame, but fails to maintain its performance in a long-term setting. To tackle long-term predictions, PredRNN (Wang et al. (2017)) proposed a new spatiotemporal LSTM, which allows memory to flow both vertically and horizontally. PredRNN $^ { + + }$ (Wang et al. (2018)) further improved the results by rearranging spatial and temporal memory in a cascaded mechanism, and by using a gradient highway architecture to ease the optimization. E3D-LSTM (Wang et al. (2019)) effectively recalled the previous memory states and also proposed to include 3D convolutions to enhance its performance. ContextVP (Byeon et al. (2018)) introduced a fully context-aware architecture consisting of parallel multi-dimensional LSTM units and blending units. Methods from stacked ConvRNNs family usually yield more accurate deterministic predictions but they consume considerable GPU memory and computational power as they abandon downsampling to prevent information loss. ",
+        "bbox": [
+            174,
+            529,
+            825,
+            777
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The encoder-predictor-decoder framework, on the other hand, provides more flexibility than its counterpart. MCNET (Villegas et al. (2017)) and DrNet (Denton et al. (2017)) decompose the content and motion in videos by building their corresponding encoders and then integrate this disentangled information to yield the next frame. Retrospective CycleGAN (Kwon & Park (2019)) combines sequential adversarial loss with frame adversarial loss, which encourages the model to generate frames that are visually similar to authentic images. In terms of modeling stochasticity, SVG (Denton & Fergus (2018)) and SAVG (Lee et al. (2018)) utilize a prior inference network to mimic the uncertainty in the environment, and then embed it into a deterministic generative model to produce stochastic video frames. VPN (Kalchbrenner et al. (2016)) estimates the discrete joint distribution of the raw pixel values in a video using the well-established PixelCNNs. It is worth noticing that VPN employs a resolution-preserving encoder to circumvent the information loss, showing the need for an efficient information-preserving encoder in the community. ",
+        "bbox": [
+            174,
+            785,
+            825,
+            924
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            103,
+            823,
+            132
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Related Works: To the best of our knowledge, our CrevNet is the first conditionally reversible model in the video prediction literature. There are three prior arts, E3D-LSTM (Wang et al. (2019)), FRNN (Oliu et al. (2018)) and VideoFlow (Kumar et al. (2019)), having some similarities with our CrevNet. While E3D-LSTM also employs 3D convolutions, their implementation is essentially equivalent to applying two 2D convolutional operations, as there is no shared filter on the temporal dimension. Similar to CrevNet, FRNN reduces its computational cost by eliminating the need to re-encode the output of decoder. However, FRNN has a substantially different architecture compared to CrevNet. While the encoder and decoder in our model do not need information sharing at all, FRNN relies heavily on the sharing of the hidden states between them. Although VideoFlow also utilizes invertible transformation. This approach is very different from ours because: (1). VideoFlow is built upon Glow, a very memory-consuming architecture. Such memory limits preclude the use of 3D convolutions, or even from training the model with Adam. (2). They use ANN to model temporal relationship. As such, VideoFlow cannot capture complex dynamics. (3). So far, VideoFlow has only been applied to stochastic video generation instead of deterministic video prediction. ",
+        "bbox": [
+            174,
+            138,
+            825,
+            333
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The Reversible and Invertible Architectures: The idea of the coupling layer was initially introduced in NICE (Dinh et al. (2014)) so as to make the computation of the determinant of the Jacobian and inverse Jacobian trival. Inspired by additive coupling layer, RevNet (Gomez et al. (2017)) introduced a reversible block that allowed the reconstruction of activations of each layer from that of the next layer, thus eliminating the need to store activations between downsampling and significantly reducing its memory consumption. The follow-up work by (Jacobsen et al. (2018)) further proposed an invertible extension, i-RevNet, which enabled the model to preserve all information of input through layers while still being capable of extracting a useful representation for classification. ",
+        "bbox": [
+            174,
+            340,
+            825,
+            452
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "5 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            472,
+            318,
+            488
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We described a novel conditionally reversible network, CrevNet, for pixel-level prediction of future frames in videos. The originality of our model lies in our use of the reversible two-way autoencoder and the accompanying reversible predictive module. Such architectural design enables the model to preserve fine-grained information without significant memory and computation overhead. CrevNet achieves state-of-the-art results on both synthetic and real-world datasets. The subsequent detection experiments demonstrate the potential of CrevNet to be a continuous self-supervised learning system to enhance downstream CV tasks, as shown in the case of BERT (Devlin et al. (2018)) for NLP tasks. ",
+        "bbox": [
+            174,
+            503,
+            825,
+            601
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            622,
+            285,
+            637
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: Endto-end speech recognition in english and mandarin. In International Conference on Machine Learning, pp. 173–182, 2016. ",
+        "bbox": [
+            173,
+            646,
+            825,
+            702
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextvp: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 753–769, 2018. ",
+        "bbox": [
+            173,
+            712,
+            823,
+            753
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6299–6308, 2017. ",
+        "bbox": [
+            171,
+            763,
+            825,
+            805
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 1174–1183, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http://proceedings.mlr.press/ v80/denton18a.html. ",
+        "bbox": [
+            174,
+            815,
+            825,
+            885
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Emily L Denton et al. Unsupervised learning of disentangled representations from video. In Advances in Neural Information Processing Systems, pp. 4417–4426, 2017. ",
+        "bbox": [
+            173,
+            895,
+            821,
+            924
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. ",
+        "bbox": [
+            171,
+            103,
+            823,
+            133
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. ",
+        "bbox": [
+            171,
+            141,
+            823,
+            171
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Piotr Dollár, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009. ",
+        "bbox": [
+            173,
+            179,
+            823,
+            209
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 3354–3361. IEEE, 2012. ",
+        "bbox": [
+            173,
+            218,
+            825,
+            261
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Aidan N Gomez, Mengye Ren, Raquel Urtasun, and Roger B Grosse. The reversible residual network: Backpropagation without storing activations. In Advances in Neural Information Processing Systems, pp. 2211–2221, 2017. ",
+        "bbox": [
+            173,
+            270,
+            826,
+            314
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. ",
+        "bbox": [
+            174,
+            323,
+            826,
+            366
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "IARAI. Traffic4cast: Traffic map movie forecasting. https://www.iarai.ac.at/ traffic4cast, 2019. Accessed: 2019-09-25. ",
+        "bbox": [
+            173,
+            376,
+            825,
+            405
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jörn-Henrik Jacobsen, Arnold Smeulders, and Edouard Oyallon. i-revnet: Deep invertible networks. arXiv preprint arXiv:1802.07088, 2018. ",
+        "bbox": [
+            171,
+            412,
+            826,
+            443
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. arXiv preprint arXiv:1610.00527, 2016. ",
+        "bbox": [
+            169,
+            452,
+            825,
+            482
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ",
+        "bbox": [
+            174,
+            489,
+            825,
+            534
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Manoj Kumar, Mohammad Babaeizadeh, Dumitru Erhan, Chelsea Finn, Sergey Levine, Laurent Dinh, and Durk Kingma. Videoflow: A flow-based generative model for video. arXiv preprint arXiv:1903.01434, 2019. ",
+        "bbox": [
+            173,
+            542,
+            825,
+            585
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1811–1820, 2019. ",
+        "bbox": [
+            174,
+            594,
+            826,
+            637
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. ",
+        "bbox": [
+            171,
+            647,
+            823,
+            678
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1744–1752, 2017. ",
+        "bbox": [
+            173,
+            685,
+            826,
+            729
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kun Liu, Wu Liu, Chuang Gan, Mingkui Tan, and Huadong Ma. T-c3d: Temporal convolutional 3d network for real-time action recognition. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. ",
+        "bbox": [
+            173,
+            738,
+            825,
+            781
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pp. 21–37. Springer, 2016. ",
+        "bbox": [
+            174,
+            790,
+            825,
+            833
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. arXiv preprint arXiv:1605.08104, 2016. ",
+        "bbox": [
+            171,
+            842,
+            825,
+            872
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 716–731, 2018. ",
+        "bbox": [
+            174,
+            881,
+            825,
+            922
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan, Yu-Wing Tai, and Li Xu. Accurate single stage detector using recurrent rolling convolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5420–5428, 2017. ",
+        "bbox": [
+            176,
+            103,
+            823,
+            146
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Shaoqing Ren, Kaiming He, Ross Girprerhick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015. ",
+        "bbox": [
+            174,
+            155,
+            823,
+            196
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pp. 234–241. Springer, 2015. ",
+        "bbox": [
+            174,
+            207,
+            826,
+            250
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1874–1883, 2016. ",
+        "bbox": [
+            174,
+            257,
+            826,
+            314
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional lstm network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems, pp. 802–810, 2015. ",
+        "bbox": [
+            174,
+            323,
+            826,
+            366
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pp. 843–852, 2015. ",
+        "bbox": [
+            173,
+            375,
+            825,
+            405
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. ",
+        "bbox": [
+            171,
+            412,
+            825,
+            441
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and S Yu Philip. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. In Advances in Neural Information Processing Systems, pp. 879–888, 2017. ",
+        "bbox": [
+            174,
+            450,
+            825,
+            493
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and Philip S Yu. PredRNN $^ { + + }$ : Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 5123–5132, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http://proceedings.mlr.press/ v80/wang18b.html. ",
+        "bbox": [
+            176,
+            501,
+            826,
+            587
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ B1lKS2AqtX. ",
+        "bbox": [
+            176,
+            594,
+            825,
+            638
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. ",
+        "bbox": [
+            174,
+            646,
+            826,
+            689
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 129–137, 2017. ",
+        "bbox": [
+            174,
+            698,
+            826,
+            755
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A CONVLSTM AND ST-LSTM ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            101,
+            450,
+            118
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The key equations of ConvLSTM are shown as belows. ",
+        "bbox": [
+            173,
+            132,
+            537,
+            148
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "equation",
+        "img_path": "images/bc8d3cae7fbdc020a524e33ba4634a1e04c5b570b9b5f0930852ec16a4abb7ba.jpg",
+        "text": "$$\n\\begin{array} { r l } & { i _ { t } = \\sigma \\big ( W _ { x i } \\ast \\mathcal { X } _ { t } + W _ { h i } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { i } \\big ) } \\\\ & { f _ { t } = \\sigma \\big ( W _ { x f } \\ast \\mathcal { X } _ { t } + W _ { h f } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { f } \\big ) } \\\\ & { \\mathcal { C } _ { t } ^ { l } = f _ { t } \\circ \\mathcal { C } _ { t - 1 } ^ { l } + i _ { t } \\circ \\operatorname { t a n h } \\bigl ( W _ { x c } \\ast \\mathcal { X } _ { t } + W _ { h c } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { c } \\bigr ) } \\\\ & { o _ { t } = \\sigma \\big ( W _ { x o } \\ast \\mathcal { X } _ { t } + W _ { h o } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + W _ { c o } \\circ \\mathcal { C } _ { t } ^ { l } + b _ { o } \\big ) } \\\\ & { \\mathcal { H } _ { t } ^ { l } = o _ { t } \\circ \\operatorname { t a n h } ( \\mathcal { C } _ { t } ^ { l } ) } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            305,
+            154,
+            692,
+            258
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "where $^ *$ denotes the convolution operator and $\\circ$ denotes the Hadamard product. Based on ConvLSTM, spatiotemporal LSTM (ST-LSTM) in PredRNN adds another vertical memory flow to enhance the long-term temporal dependency as follow. ",
+        "bbox": [
+            176,
+            262,
+            823,
+            304
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "equation",
+        "img_path": "images/51d1cb21befee84cd58881dd1d342717c376c9b2ef0223032870cbb6324fa49d.jpg",
+        "text": "$$\n\\begin{array} { r l } & { i _ { t } = \\sigma ( W _ { x i } \\ast \\mathcal { X } _ { t } + W _ { h i } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { i } ) } \\\\ & { f _ { t } = \\sigma ( W _ { x f } \\ast \\mathcal { X } _ { t } + W _ { h f } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { f } ) } \\\\ & { \\mathcal { C } _ { t } ^ { l } = f _ { t } \\circ \\mathcal { C } _ { t - 1 } ^ { l } + i _ { t } \\circ \\operatorname { t a n h } ( W _ { x c } \\ast \\mathcal { X } _ { t } + W _ { h c } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + b _ { c } ) } \\\\ & { \\dot { u } _ { t } ^ { l } = \\sigma ( W _ { x i } ^ { r } \\ast \\mathcal { X } _ { t } + W _ { m i } \\ast \\mathcal { M } _ { t } ^ { l - 1 } + b _ { i } ^ { i } ) } \\\\ & { f _ { t } ^ { \\prime } = \\sigma ( W _ { x f } ^ { r } \\ast \\mathcal { X } _ { t } + W _ { m f } \\ast \\mathcal { M } _ { t } ^ { l - 1 } + b _ { f } ^ { \\prime } ) } \\\\ & { \\mathcal { M } _ { t } ^ { l } = f _ { t } ^ { \\prime } \\circ \\mathcal { M } _ { t } ^ { l - 1 } + i _ { t } ^ { \\prime } \\circ \\operatorname { t a n h } ( W _ { x m } \\ast \\mathcal { X } _ { t } + W _ { m m } \\ast \\mathcal { M } _ { t } ^ { l - 1 } + b _ { m } ) } \\\\ & { o _ { t } = \\sigma ( W _ { x o } \\ast \\mathcal { X } _ { t } + W _ { h o } \\ast \\mathcal { H } _ { t - 1 } ^ { l } + W _ { c o } \\circ \\mathcal { C } _ { t } ^ { l } + W _ { m o } \\circ \\mathcal { M } _ { t } ^ { l } + b _ { o } ) } \\\\ & { \\mathcal { H } _ { t } = o _ { t } \\mathrm { o t } \\operatorname { t a n h } ( W _ { 1 x } | C _ { t } ^ { l } , \\mathcal { M } _ { t } ^ { l } | ) } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            281,
+            310,
+            717,
+            477
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "where blue part overlaps ConvLSTM. Note that $\\mathcal { M } _ { t } ^ { l }$ usually receives information from the previous layer instead of the previous state and the special case is that $\\boldsymbol { \\mathcal { M } } _ { t } ^ { 1 }$ receives $\\mathcal { M } _ { t - 1 } ^ { L }$ to constitute a zigzag information flow. As we can see, ST-LSTM basically doubles the size of feature map and the number of parameters compared with ConvLSTM. ",
+        "bbox": [
+            173,
+            489,
+            826,
+            547
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B CONDITIONAL REVERSIBILITY ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            568,
+            462,
+            583
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "As we mentioned in Section 2.2, conditional reversibility is an interesting property of our CrevNet. In this section, we will provide more details about it. Given $\\hat { x } _ { t } ^ { 2 } , x _ { t } ^ { 1 }$ and $h _ { t - 1 } ^ { 1 }$ , the reversible predictive module can recover $\\boldsymbol { x } _ { t } ^ { 2 }$ as follow ",
+        "bbox": [
+            174,
+            598,
+            826,
+            643
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "equation",
+        "img_path": "images/e2f29cb29b20d2464cc7b54fce55720e7ca814c376913f261367084175ff0b1c.jpg",
+        "text": "$$\n\\begin{array} { r l } & { h _ { t } ^ { 1 } = \\mathrm { C o n v R N N } ( x _ { t } ^ { 1 } , h _ { t - 1 } ^ { 1 } ) } \\\\ & { g _ { t } = \\phi ( W * h _ { t } ^ { 1 } + b ) } \\\\ & { x _ { t } ^ { 2 } = ( \\hat { x } _ { t } ^ { 2 } - g _ { t } \\odot h _ { t } ^ { 1 } ) \\odot 1 / 1 - g _ { t } } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            388,
+            648,
+            607,
+            712
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Ideally, if there is no numerical error during the calculation, we can get the perfect reconstruction of input by applying this inverse operation repeatedly. In practice, while the most of reverse generations are successful, the inevitable numerical error will still result in some failing cases, especially in the case of a very deep architecture because errors will be amplified layer by layer. ",
+        "bbox": [
+            173,
+            715,
+            826,
+            772
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C OBJECT DETECTION ",
+        "bbox": [
+            174,
+            791,
+            380,
+            808
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b2150b3b04ec4bfef448c6064e3b09a1763f16086583fdf4984344b7d0c05cb.jpg",
+        "image_caption": [
+            "Figure 7: Visualization of 2D Object Detection on KITTI. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            118,
+            797,
+            864
+        ],
+        "page_idx": 13
+    }
+]

parse/train/HyGTuv9eg/HyGTuv9eg.md

@@ -0,0 +1,224 @@
+# INCORPORATING LONG-RANGE CONSISTENCY I NCNN-BASED TEXTURE GENERATION
+
+Guillaume Berger & Roland Memisevic
+
+Department of Computer Science and Operations Research   
+University of Montreal   
+guillaume.berger@umontreal.ca, memisevr@iro.umontreal.ca
+
+# ABSTRACT
+
+Gatys et al. (2015a) showed that pair-wise products of features in a convolutional network are a very effective representation of image textures. We propose a simple modification to that representation which makes it possible to incorporate longrange structure into image generation, and to render images that satisfy various symmetry constraints. We show how this can greatly improve rendering of regular textures and of images that contain other kinds of symmetric structure. We also present applications to inpainting and season transfer.
+
+# 1 INTRODUCTION
+
+There are currently two dominant approaches to texture synthesis: non-parametric techniques, which synthesize a texture by extracting pixels (or patches) from a reference image that are resampled for rendering (Efros & Leung, 1999; Kwatra et al., 2003), and parametric statistical models, which optimize reconstructions to match certain statistics computed on filter responses (Heeger & Bergen, 1995; Portilla & Simoncelli, 2000). Recently, the second approach has seen a significant advancement, after Gatys et al. (2015a) showed that a CNN pre-trained on an object classification task, such as ImageNet (Russakovsky et al., 2015), can be very effective at generating textures. Gatys et al. (2015a) propose to minimize with respect to the input image a loss function, that measures how well certain high-level features of a reference image are preserved. The reference image constitutes an example of the texture to be generated. The high-level features to be preserved are pair-wise products of feature responses, averaged over the whole image, referred to as the “Gramian” in that work. In Gatys et al. (2015b), the same authors show that by adding a second term to the cost, which matches the content of another image, one can render that other image in the “style” (texture) of the first. Numerous follow-up works have since then analysed and extended this approach (Ulyanov et al., 2016; Johnson et al., 2016; Ustyuzhaninov et al., 2016).
+
+As shown in Figure 1, this method produces impressive results. However, it fails to take into account non-local structure, and consequently cannot generate results that exhibit long-range correlations in images. An example of the importance of long-range structure is the regular brick wall texture in the middle of the figure. Another example is the task of inpainting, where the goal is to fill in a missing part of an image, such that it is faithful to the non-missing pixels. Our main contribution is to introduce a way to deal with long-range structure using a simple modification to the product-based texture features. Our approach is based on imposing a “Markov-structure” on high-level features, allowing us to establish feature constraints that range across sites instead of being local. Unlike classical approaches to preserving spatial structure in image generation, such as Markov Random Fields and learning-based extensions (Roth & Black, 2005), our approach does not impose any explicit local constraints on pixels themselves. Rather, inspired by Gatys et al. (2015a), it encourages consistency to be satisfied on high-level features and on average across the whole image. We present applications to texture generation, inpainting and season transfer.
+
+# 2 THE ARTISTIC STYLE ALGORITHM
+
+# 2.1 SYNTHESIS PROCEDURE
+
+Given a reference texture, $x$ , the algorithm described in Gatys et al. (2015a) permits to synthesize by optimization a new texture $\hat { x }$ similar to $x$ . To achieve this, the algorithm exploits an ImageNet
+
+![](images/fdfb6b9e81db4b326e7af8723287b7dd54410b08bb7ba3391927271ae0a5b270.jpg)  
+Figure 1: Reference image (left) and generated texture (right) using the procedure described in Gatys et al. (2015a).
+
+![](images/d62b0a03199a11debffeb0150e39840133ebb843e2d97dbc40ddfdcc856e8665.jpg)  
+Figure 2: Summary of the texture synthesis procedure described in Gatys et al. (2015a). We use a VGG-19 network Simonyan & Zisserman (2014) as the pre-trained CNN.
+
+pre-trained model to define metrics suitable for describing textures: “Gram” matrices of feature maps, computed on top of $L$ selected layers. Formally, let $\hat { N } ^ { l }$ be the number of maps in layer $l$ of a pre-trained CNN. The corresponding Gram matrix $G ^ { \bar { l } }$ is a $N ^ { l } \times N ^ { l }$ matrix defined as:
+
+$$
+G _ { i j } ^ { l } = \frac { 1 } { M ^ { l } } \sum _ { k = 1 } ^ { M ^ { l } } F _ { i k } ^ { l } F _ { j k } ^ { l } = \frac { 1 } { M ^ { l } } \langle F _ { i : } ^ { l } , F _ { j : } ^ { l } \rangle
+$$
+
+where $F _ { i : } ^ { l }$ is the $i ^ { t h }$ vectorized feature map of layer $l$ , $M ^ { l }$ is the number of elements in each map of this layer, and where $\langle \cdot , \cdot \rangle$ denotes the inner product. Equation 1 makes it clear that $G ^ { l }$ captures how feature maps from layer $l$ are correlated to each other. Diagonal terms, $G _ { i i } ^ { l }$ are the squared Frobenius norm of the $i ^ { t h }$ map $\left\| F _ { i : } ^ { l } \right\| _ { F } ^ { 2 }$ , so they represent its spatially averaged energy. We will discuss the Gramians in more detail in the next paragraph. Once the Gram matrices $\{ G ^ { l } \} _ { l \in [ 1 , L ] }$ of the reference texture are computed, the synthesis procedure by Gatys et al. (2015a) amounts to constructing an image that produces Gram matrices $\{ \hat { G } ^ { l } \} _ { l \in [ 1 , L ] }$ that match the ones of the reference texture. More precisely, the following loss function is minimized with respect to the image being constructed:
+
+$$
+\mathcal { L } _ { s t y l e } = \sum _ { l = 1 } ^ { L } w _ { l } \left. \hat { \boldsymbol { G } } ^ { l } - \boldsymbol { G } ^ { l } \right. _ { F } ^ { 2 } = \sum _ { l = 1 } ^ { L } w _ { l } \mathcal { L } _ { s t y l e } ^ { l }
+$$
+
+where $w _ { l }$ is a normalizing constant similar to Gatys et al. (2015a).
+
+The overall process is summarized in Figure 2. While the procedure can be computationally expensive, there have been successful attempts reported recently which reduce the generation time (Ulyanov et al., 2016; Johnson et al., 2016).
+
+# 2.2 WHY GRAM MATRICES WORK
+
+Feature Gram matrices are effective at representing texture, because they capture global statistics across the image due to spatial averaging. Since textures are static, averaging over positions is required and makes Gram matrices fully blind to the global arrangement of objects inside the reference image. This property permits to generate very diverse textures by just changing the starting point of the optimization. Despite averaging over positions, coherence across multiple features needs to be preserved (locally) to model visually sensible textures. This requirement is taken care of by the off-diagonal terms in the Gram matrix, which capture the co-occurence of different features at a single spatial location. Indeed, Figure 3 shows that restricting the texture representation to the squared Frobenius norm of feature maps (i.e. diagonal terms) makes distinct object-parts from the reference texture encroach on each other in the reconstruction, as local coherence is not captured by the model. Exploiting off-diagonal terms improves the quality of the reconstruction as consistency across feature maps is enforced (on average across the image).
+
+![](images/81370cb9bdd3f71e6f5fa79f9a4d5298de9b1732fa438e0865bb44a9d80ca202.jpg)  
+Figure 3: Exploiting Gram matrices of feature maps as in Gatys et al. (2015a) $( 1 ^ { s t } r o w )$ or only the squared Frobenius norm of feature maps $( 2 ^ { n d } r o w )$ for increasingly deep layers (from left to right).
+
+The importance of local coherence can be intuitively understood in the case of linear features (or in the lowest layer of a convolutional network): when decomposing an image using Gabor-like features, local structure can be expressed as the relative offsets in the Fourier phase angles between multiple different filter responses. A sharp step-edge, for example, requires the phases of local Fourier components at different frequencies to align in a different way than a blurry edge or a ridge (Morrone & Burr, 1988; Kovesi, 1999). Also, natural images exhibit very specific phaserelationships across frequency components in general, and destroying these makes the image look unnatural (Wang & Simoncelli, 2003). The same is not true of Fourier amplitudes (represented on the diagonals of the Gramian), which play a much less important role in the visual appearance (Oppenheim & Lim, 1981). In the case of deeper representations, the situation is more complex, but it is still local co-occurrence averaged over the whole image that captures texture.
+
+Unfortunately, average local coherence falls short of capturing long-range structure in images. Spatial consistency is hard to capture within a single filter bank, because of combinatorial effects. Indeed, since Gram matrices capture coherence at a single spatial location, every feature would have to be matched to multiple transformed versions of itself. A corollary is that every feature would have to appear in the form of multiple transformed copies of itself in order to capture spatial consistency. However, this requirement clashes with the limited number of features available in each CNN layer. One way to address this is to use higher-layer features, whose receptive fields are larger. Unfortunately, as illustrated in Figure 3, even if using layers up to pool5 whose input receptive field covers the whole image1 (first row, last column), the reconstruction remains mainly unstructured and the method fails to produce spatial regularities.
+
+# 3 MODELING SPATIAL CO-OCCURENCES
+
+To account for spatial structure in images, we propose encoding this structure in the feature selfsimilarity matrices themselves. To this end, we suggest that, instead of computing co-occurences between multiple features within a map, we compute co-occurences between feature maps $F ^ { l }$ and spatially transformed feature maps $T ( F ^ { l } )$ , where $T$ denotes a spatial transformation. In the simplest case, $T$ represents local translation, which amounts to measuring similarities between local features and other neighbouring features. We denote by $T _ { x , + \delta }$ the operation consisting in horizontally translating feature maps by $\delta$ pixels and define the transformed Gramian:
+
+$$
+G _ { x , \delta , i j } ^ { l } = \frac { 1 } { M ^ { l } } \langle T _ { x , + \delta } \left( F _ { i : } ^ { l } \right) , T _ { x , - \delta } \left( F _ { j : } ^ { l } \right) \rangle
+$$
+
+![](images/a94944d3a051ed2f9aae98e632a49046bd878a60099da97d9cc1efc68fe4cea9.jpg)  
+Figure 4: Computing the shifted Gram matrix for a given layer with feature maps of width $X$ .
+
+where $T _ { x , - \delta }$ performs a translation in the opposite direction. As illustrated in Figure 4, the transformation in practice simply amounts to removing the $\delta$ first or last columns from the raw feature maps. Therefore, the inner product now captures how features at position $( i , j )$ are correlated with features located at position $( i , j + \delta )$ in average. While Figure 4 illustrates the case where feature maps are horizontally shifted, one would typically use translations along both the $x$ -axis and the $y$ -axis. Our transformed Gramians are related to Gray-Level Co-occurrence Matrices (GLCM) (Haralick et al., 1973) which compute the unnormalized frequencies of pixel values for a given offset in an image. While GLCMs have been mainly used for analysis, some work tried to use these features for texture synthesis (Lohmann, 1995). Usually, GLCMs are defined along 4 directions: $0 ^ { \circ }$ and $9 0 °$ (i.e. horizontal and vertical offsets), as well as $4 5 ^ { \circ }$ and $1 3 5 ^ { \circ }$ (i.e. diagonal offsets). In comparison, our method does not consider diagonal offsets and captures spatial coherence on high-level features, making use of a pre-trained CNN, instead of working directly in the pixel domain.
+
+With this definition for transformed Gram matrices, we propose defining the loss as: $\begin{array} { r l } { \mathcal { L } } & { { } = } \end{array}$ $\mathcal { L } _ { s t y l e } + \mathcal { L } _ { c c }$ , where $c c$ stands for cross-correlation. Like $\mathcal { L } _ { s t y l e }$ , $\mathcal { L } _ { c c }$ is a weighted sum of multiple losses $\mathcal { L } _ { c c , \delta } ^ { l }$ defined for several selected layers as the mean squared error between transformed Gram matrices of the reference texture and the one being constructed:
+
+$$
+\mathcal { L } _ { c c , \delta } ^ { l } = \frac { 1 } { 2 } \left( \mathcal { L } _ { c c , x , \delta } ^ { l } + \mathcal { L } _ { c c , y , \delta } ^ { l } \right) = \frac { 1 } { 2 } \left( \left. \hat { G } _ { x , \delta } ^ { l } - G _ { x , \delta } ^ { l } \right. _ { F } ^ { 2 } + \left. \hat { G } _ { y , \delta } ^ { l } - G _ { y , \delta } ^ { l } \right. _ { F } ^ { 2 } \right)
+$$
+
+Although this amounts to adding more terms to a representation that was already high-dimensional and overparametrized, we found that these additional terms do not hurt the diversity of generated textures. Indeed, the new loss remains blind to the global arrangement of objects. In fact, there exists a specific situation where our approach is strictly equivalent to computing Gram matrices of deeper layers: with linear activations and “one-hot” convolution kernels2, deeper layers would simply contain translated versions of lower feature maps. In that case, computing Gram matrices of deeper layers would permit to directly capture cross-correlation statistics in lower ones. Nevertheless, this situation is very unlikely when using a pretrained CNN, as the network probably learned operations more useful than simple translations during its supervised training.
+
+While we focus on translation for most of our results, we shall discuss other types of transformation in the experiments Section.
+
+# 4 EXPERIMENTS
+
+In our experiments, we exploit the same normalized version of the VGG-19 network3 (Simonyan & Zisserman, 2014) as in Gatys et al. (2015a;b) and layers conv1 1, pool1, pool2, pool3, and pool4 are always used to define the standard Gram matrices. In our method, we did not use conv1 1 to define cross-correlation terms, as the large number of neurons at this stage makes the computation of Gram matrices costly. Corresponding $\delta$ values for each layer are discussed in the next paragraph.
+
+![](images/a4fe332fd30d43110d6b3cb6270896f60861c155adfa9dd5fdca304e451c2575.jpg)  
+Figure 5: P ool2 reconstruction using different values for $\delta$ in $\mathcal { L } _ { c c , \delta } ^ { p o o l 2 }$ . $\boldsymbol { \cdot } \boldsymbol { 1 } ^ { s t }$ column): Without crosscorrelation terms. (Other columns): Using a single cross-correlation term with a fixed value $( 1 ^ { s t }$ row) or using multiple cross-correlation terms with distinct $\delta$ values $( 2 ^ { n d } r o w )$ .
+
+Finally, our implementation4 uses Lasagne (Dieleman et al., 2015). Each image is of size $3 8 4 \times 3 8 4$ Most textures used as references in this paper were taken from textures.com and pixabay.com.
+
+# 4.1 EXPERIMENTS WITH TRANSLATION-GRAMIANS
+
+The $\delta$ parameter is of central importance as it dictates the range of the spatial constraints. We observed that the optimal value depends on both the considered layer in the pre-trained CNN and the reference texture, making it difficult to choose a value automatically.
+
+For instance, Figure 5 shows generated images from the brick wall texture using only the pool2 layer with different $\delta$ configurations. The first row depicts the results when considering single values of $\delta$ only. While $\delta = 4$ or $\delta = 8$ are good choices, considering extreme long-range correlations does not help for this particular texture: a brick depends mostly on its neighbouring bricks and not the far-away ones. More precisely, a translation of more than 16 pixels in the pool2 layer makes the input receptive field move more than 64 pixels. Therefore $\delta = 1 6$ or $\delta = 3 2$ do not capture any information about neighbouring bricks. Unfortunately, this is not true for all textures, and $\delta = 1 6$ or $\delta = 3 2$ might be good choices for another image that exhibits longer structures.
+
+Searching systematically for a $\delta$ configuration that works well with the reference texture being considered would be a tedious task: even for a very regular texture with a periodic horizontal (or vertical) pattern, it is hard to guess the optimal $\delta$ values for each layer (for deeper ones in particular). Instead, we propose to use a fixed but wide set of $\delta$ values per layer, by defining the cost to be: $\mathcal { L } _ { c c } ^ { l } = \sum _ { k } \mathcal { L } _ { c c , \delta _ { k } } ^ { l }$ . A potential concern is that combining many loss terms can hurt the reconstruction or the diversity of generated textures. Figure 5 (second row) shows contrarily that there is no visual effect from using $\delta$ values that are not specifically useful for the reference texture being considered: the rendering benefits from using $\bar { \delta } \in \{ 2 , \dot { 4 } , 8 \}$ while considering bigger values $( \delta \in \{ 2 , 4 , 8 , 1 6 , 3 2 \}$ , e.g.) does not help, but does not hurt the reconstruction either. We found the same to be true for other textures as well, and our results (shown in the next Section) show that combining the loss terms is able to generate very diverse textures. A drawback from using multiple cross-correlation terms per layers, however, is computational. We found that in our experimental setups, adding cross-correlation terms increases the generation time by roughly $8 0 \%$ .
+
+As a guide-line, for image sizes of roughly $3 8 4 \times 3 8 4$ pixels, we recommend the following $\delta$ values per layer (which we used in all our following experiments): $\{ 2 , 4 , 8 , 1 6 , 3 2 , 6 4 \}$ for pool1, $\{ 2 , 4 , 8 , 1 6 , 3 2 \}$ for pool2, $\{ 2 , 4 , 8 , 1 6 \}$ for pool3, and $\{ 2 , 4 , 8 \}$ for pool4. The number and the range of $\delta$ values decrease with depth because feature maps are getting smaller due to $2 \times 2$ pooling layers. This configuration should be sufficient to account for spatial structure in any $3 8 4 \times 3 8 4$ image.
+
+![](images/0733b00648dcc0afd39301437db05c2b2d3c89d6c966c8fe19133bf1c9e0c450.jpg)  
+Figure 6: Some results of our approach compared with Gatys et al. (2015a). Only the initialization differs in the last two columns. Further results are shown in the supplementary material.
+
+# 4.2 SYNTHESIS OF STRUCTURED TEXTURE EXAMPLES
+
+Figure 6 shows the result of our approach applied to various structured and unstructured textures. It demonstrates that the method is effective at capturing long-range correlations without simply copying the content of the original texture. For instance, note how our model captures the depth aspect of the reference image in the third and fourth rows.
+
+![](images/b3c7674352750d3be3a9314c029171b2f99b3d3205948b2e93a221b3ce6f6dba.jpg)  
+Figure 7: Texture generation applied to in-painting. More in-painted images can be found in the supplementary material.
+
+![](images/dd51b4bc2b046d0074a607813aa7e8c25954f2a16b93452368583276119ba3e2.jpg)  
+Figure 8: Season transfer examples.
+
+The problem of synthesizing near-regular structures is challenging because stochasticity and regularity are adversarial properties (Lin et al., 2006). Non-parametric patch-based techniques, such as Efros & Freeman (2001), are better suited for this task because they can $t i l e ^ { 5 }$ the reference image. On the other hand, regular structures are usually more problematic for parametric statistical models. Nevertheless, the two first rows of Figure 6 demonstrate that our approach can produce good visual results and can reduce the gap to patch-based methods on these kinds of texture.
+
+Even if the reference image is not a texture, the generated images in the last row (Leonardo Dicaprio’s face) provide a good visual illustration of the effect of translation terms. In contrast to Gatys et al., our approach preserves longer-range structure, such as the alignment and similar appearance of the eyes, hair on top of the forehead, the chin below the mouth, etc. Finally, when the reference texture is unstructured (fifth row), our solution does not necessarily provide a benefit, but it also does not hurt the visual quality or the diversity of the generated textures.
+
+# 4.3 INPAINTING APPLICATION
+
+Modelling long-range correlations can make it possible to apply texture generation to inpainting, because it allows us to impose consistency constraints between the newly rendered region and the unmodified parts of the original image.
+
+To apply our approach to texture inpainting, we extracted two patches from the original image: one that covers the whole area to inpaint, and another one that serves as the reference texture. Then, approximately the same process as for texture generation is used, with the following two modifications: First, instead of random noise, the optimization starts from the masked content patch (the one to inpaint) showing a grey area and its non-missing surrounding. Second, we encourage the borders of the output to not change much with respect to the original image using an $L _ { 2 }$ penalty. We apply the penalty both in the Gatys et al. rendering and in ours. Some inpainted images are shown in Figure 7. As seen in the figure, our solution significantly outperforms Gatys et al. in terms of visual quality. Further results are shown in the supplementary material.
+
+![](images/1fbcc91968bc06348295a657a024bdfd88b21ab99f4a0f4a49d20e11ff2acdfd.jpg)  
+Figure 9: Generation of abstract symmetric textures.
+
+# 4.4 SEASON TRANSFER
+
+Figure 8 shows the result of applying our approach to a style transfer task, as in Gatys et al. (2015b): transferring the “season” of a landscape image to another one. On this task, the results from our approach are similar to those from Gatys et al. (2015b). Nevertheless, in contrast to the Gatys et al. results, our approach seems to better capture global information, such as sky color and leaves (bottom row), or the appearance of branches in the winter image (top row).
+
+# 4.5 INCORPORATING OTHER TYPES OF STRUCTURE
+
+While we focused on feature map translations in most of our experiments, other transformations can be applied as well. To illustrate this point, we explored a way to generate symmetric textures using another simple transformation. To this end, we propose flipping one of the two feature maps before computing the Gram matrices: $G _ { l r , i j } ^ { l } = \langle F _ { i : } ^ { l } , T _ { l r } \ ' ( F _ { j : } ^ { l } ) \rangle$ . Here $T _ { l r }$ corresponds to the left-right flipping operation, but we also considered up-down flipping of feature maps: $\mathcal { L } ^ { l } = \mathcal { L } _ { s t y l e } ^ { l } + \mathcal { L } _ { l r } ^ { l } +$ L  lud .
+
+As can be seen in Figure 9, in contrast to Gatys et al., the additional loss terms capture which objects are symmetric in the reference texture, and enforce these same objects to be symmetric in the reconstruction as well. Other kinds of transformation could be used, depending on the type of property in the source texture one desires to preserve.
+
+# 5 CONCLUSION
+
+We presented an approach to satisfying long-range consistency constraints in the generation of images. It is based on a variation of the method by Gatys et al., and considers spatial co-occurences of local features (instead of only co-occurences across features). We showed that the approach permits to generate textures with various global symmetry properties and that it makes it possible to apply texture generation to in-painting. Since it preserves correlations across sites, the approach is reminiscent of an MRF, but in contrast to an MRF or other graphical models, it defines correlationconstraints on high-level features of a (pre-trained) CNN rather than on pixels.
+
+# REFERENCES
+
+Sander Dieleman, Jan Schlter, Colin Raffel, et al. Lasagne: First release., August 2015. URL http://dx. doi.org/10.5281/zenodo.27878.   
+A. A. Efros and T. K. Leung. Texture synthesis by non-parametric sampling. In Proceedings of the International Conference on Computer Vision-Volume 2, ICCV ’99, pp. 1033–, 1999.   
+Alexei A. Efros and William T. Freeman. Image Quilting for Texture Synthesis and Transfer. Proceedings of SIGGRAPH 2001, pp. 341–346, August 2001.   
+L. A. Gatys, A. S. Ecker, and M. Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems 28, 2015a. URL http://arxiv.org/abs/1505.07376.   
+L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm of artistic style. CoRR, abs/1508.06576, 2015b. URL http://arxiv.org/abs/1508.06576.   
+R. Haralick, K. Shanmugam, and I. Dinstein. Texture features for image classification. IEEE Transactions on Systems, Man, and Cybernetics 3 (6), 1973.   
+David J. Heeger and James R. Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1995, pp. 229–238, 1995.   
+J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. CoRR, abs/1603.08155, 2016. URL http://arxiv.org/abs/1603.08155.   
+P. Kovesi. Image features from phase congruency. Videre, 1(3):2–27, 1999.   
+V. Kwatra, A. Schdl, I. Essa, G. Turk, and A. Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Transactions on Graphics, SIGGRAPH 2003, 22(3):277–286, 2003.   
+Wen-Chieh Lin, James Hays, Chenyu Wu, V. Kwatra, and Yanxi Liu. Quantitative evaluation on near regular texture synthesis. In CVPR ’06, volume 1, pp. 427 – 434, June 2006.   
+G. Lohmann. Analysis and synthesis of textures: A co-occurrence-based approach. Computers and Graphics, 19(1), pp. 29–36, 1995.   
+M. C. Morrone and D. C. Burr. Feature detection in human vision: A phase-dependent energy model. Proceedings of the Royal Society of London B: Biological Sciences, 235(1280):221–245, 1988.   
+A. V. Oppenheim and J. S. Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529–541, 1981. ISSN 0018-9219.   
+J. Portilla and E. P. Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. Int. J. Comput. Vision, 40(1):49–70, 2000.   
+S. Roth and M. J. Black. Fields of experts: A framework for learning image priors. In IEEE Conf. on Computer Vision and Pattern Recognition, volume 2, pp. 860–867, June 2005.   
+O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015.   
+K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014.   
+D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In International Conference on Machine Learning (ICML), 2016.   
+I. Ustyuzhaninov, W. Brendel, L. A. Gatys, and M. Bethge. Texture synthesis using shallow convolutional networks with random filters. CoRR, abs/1606.00021, 2016. URL http://arxiv.org/abs/1606. 00021.   
+Z. Wang and E. P. Simoncelli. Local phase coherence and the perception of blur. In Advances in Neural Information Processing Systems 16 [NIPS 2003], pp. 1435–1442, 2003.   
+Ciyou Zhu, Richard H. Byrd, Peihuang Lu, and Jorge Nocedal. L-bfgs-b - fortran subroutines for large-scale bound constrained optimization. Technical report, ACM Trans. Math. Software, 1994.
+
+# Supplementary material
+
+A TEXTURE GENERATION
+
+![](images/7996b087aac1120f4135dc33c1be02fad01fda7f4f161d9c89a75c42da2f811e.jpg)
+
+![](images/10a8cc26ef285ca401b2e76fa624cfbae157f2de6a8075520d9bece66f04319f.jpg)
+
+# B INPAINTING
+
+![](images/6201af3b350a06711412fc7b240a36cc657b5ef4cdbcb06c4bde514444bda001.jpg)
+
+![](images/455b2d1f33375b4133b4b298627eb4879109e41a4f9f13fe6e66715d6daae2d2.jpg)
+
+![](images/267bf72c6aca6e0e1e6e883796f77dbea9556357f14c4d5ea7c1d3cb0399c383.jpg)
+
+![](images/d2dfcfb15fed6909b957b975612920289da35cb18b6401e73af4ecbb17c18f92.jpg)  
+Gatys et al.
+
+![](images/f5cbb699d907e2eac9878e99f3830544c0028945aa7c44e4d26985e0004266f9.jpg)  
+Inpainted image
+
+![](images/675e22dc39763b4f655d38d605aba2ad35e17199acb9830163b57035fd91d414.jpg)  
+Gatys et al.
+
+![](images/d9f6f36c89d7a8c86ebae09f8f61b55f4a623c6064c1fded1a16a1696709d7e0.jpg)
+
+![](images/3bae23c5542923774b38a947f266d94da913576141432c0f47b112ea97f15e13.jpg)
+
+Table 1: Final Gatys et al. loss for all inpainted textures (by order of appearance).   
+
+<table><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=2>Gatys et al. loss</td></tr><tr><td rowspan=1 colspan=1>Texture</td><td rowspan=1 colspan=1>Gatys et al.</td><td rowspan=1 colspan=1>Ours</td></tr><tr><td rowspan=1 colspan=1>&quot;red_leaves”</td><td rowspan=1 colspan=1>3.38e-4</td><td rowspan=1 colspan=1>4.72e-4</td></tr><tr><td rowspan=1 colspan=1>“floor”</td><td rowspan=1 colspan=1>3.68e-4</td><td rowspan=1 colspan=1>4.73e-4</td></tr><tr><td rowspan=1 colspan=1>Supplementa</td><td rowspan=1 colspan=1>rymaterialtex</td><td rowspan=1 colspan=1>tures</td></tr><tr><td rowspan=1 colspan=1>&quot;leaves&quot;</td><td rowspan=1 colspan=1>4.08e-4</td><td rowspan=1 colspan=1>5.14e-4</td></tr><tr><td rowspan=1 colspan=1>“cargo”</td><td rowspan=1 colspan=1>3.39e-4</td><td rowspan=1 colspan=1>4.73e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;building”</td><td rowspan=1 colspan=1>8.90e-4</td><td rowspan=1 colspan=1>6.67e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;building2&quot;</td><td rowspan=1 colspan=1>7.19e-4</td><td rowspan=1 colspan=1>7.78e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;cargo2&quot;</td><td rowspan=1 colspan=1>1.05e-3</td><td rowspan=1 colspan=1>1.18e-3</td></tr><tr><td rowspan=1 colspan=1>&quot;small_bricks&quot;</td><td rowspan=1 colspan=1>1.21e-3</td><td rowspan=1 colspan=1>6.82e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;football_team&quot;</td><td rowspan=1 colspan=1>4.56e-3</td><td rowspan=1 colspan=1>5.68e-3</td></tr></table>
+
+Table 1 reports the final Gatys et al. losses obtained by both approaches for all inpainted textures showed previously. In most cases, the Gatys et al. (2015a) approach converges to a smaller value, which is not surprising since our method optimizes a modified version of the loss. Table 1, combined with the visual aspect of Gatys et al. inpainted renderings, illustrates that only imposing local coherence is not enough to obtain good inpainted textures. Our texture representation seems to be better suited for this task.
+
+![](images/2231dc59db139e901ed523d9c558bec18b16d8935cd502936456659b797a3cde.jpg)
+
+![](images/e0fa62ff4230f9c419d3d39bd52bfa9c05b0730b5ff3652b369555d4cfae2a42.jpg)
+
+![](images/7cd2b08cfe5550333150be8478841c746a249da2ce20540ee9b25ad7a09d0aed.jpg)  
+Style
+
+![](images/b11d2a55d9752e5cc1349d9e99b1d8ff309f0a3b2e36271d669c7c05e237643f.jpg)  
+Content (&init)
+
+![](images/32cae7131934dfbed3305598754f0968fe23b5e3a724f9b5c832ee8385ab0c1a.jpg)  
+Ours
+
+![](images/6ee592082e45d87c3ff944ef1fa30d5b08a60546fcc2adc7e12c8e37868b8f0b.jpg)  
+Gatys
+
+![](images/4979210ace33ad8aadca5edafe85398ff90db694cd7f16502bfe8ea8c4682526.jpg)
+
+![](images/66541a0ba406f23dbfab5e6aa52db140a57333d83d1f90070ddc9efc7e3eea6b.jpg)
+
+![](images/af9da92cc614db1ca298004e54e4372697b826ce6efbb764deff7b922b87e852.jpg)
+
+# C SYMMETRIC TEXTURES
+
+![](images/aee878193ca11f6507957dcf5903f3dd8195b3beabcbeae2cdf46ef0386165a5.jpg)

parse/train/HyGTuv9eg/HyGTuv9eg_content_list.json

@@ -0,0 +1,1126 @@
+[
+    {
+        "type": "text",
+        "text": "INCORPORATING LONG-RANGE CONSISTENCY I NCNN-BASED TEXTURE GENERATION",
+        "text_level": 1,
+        "bbox": [
+            176,
+            101,
+            821,
+            146
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Guillaume Berger & Roland Memisevic ",
+        "bbox": [
+            186,
+            170,
+            460,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science and Operations Research   \nUniversity of Montreal   \nguillaume.berger@umontreal.ca, memisevr@iro.umontreal.ca ",
+        "bbox": [
+            184,
+            185,
+            717,
+            226
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            452,
+            247,
+            544,
+            263
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Gatys et al. (2015a) showed that pair-wise products of features in a convolutional network are a very effective representation of image textures. We propose a simple modification to that representation which makes it possible to incorporate longrange structure into image generation, and to render images that satisfy various symmetry constraints. We show how this can greatly improve rendering of regular textures and of images that contain other kinds of symmetric structure. We also present applications to inpainting and season transfer. ",
+        "bbox": [
+            232,
+            276,
+            764,
+            375
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            392,
+            336,
+            409
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "There are currently two dominant approaches to texture synthesis: non-parametric techniques, which synthesize a texture by extracting pixels (or patches) from a reference image that are resampled for rendering (Efros & Leung, 1999; Kwatra et al., 2003), and parametric statistical models, which optimize reconstructions to match certain statistics computed on filter responses (Heeger & Bergen, 1995; Portilla & Simoncelli, 2000). Recently, the second approach has seen a significant advancement, after Gatys et al. (2015a) showed that a CNN pre-trained on an object classification task, such as ImageNet (Russakovsky et al., 2015), can be very effective at generating textures. Gatys et al. (2015a) propose to minimize with respect to the input image a loss function, that measures how well certain high-level features of a reference image are preserved. The reference image constitutes an example of the texture to be generated. The high-level features to be preserved are pair-wise products of feature responses, averaged over the whole image, referred to as the “Gramian” in that work. In Gatys et al. (2015b), the same authors show that by adding a second term to the cost, which matches the content of another image, one can render that other image in the “style” (texture) of the first. Numerous follow-up works have since then analysed and extended this approach (Ulyanov et al., 2016; Johnson et al., 2016; Ustyuzhaninov et al., 2016). ",
+        "bbox": [
+            174,
+            424,
+            825,
+            632
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 1, this method produces impressive results. However, it fails to take into account non-local structure, and consequently cannot generate results that exhibit long-range correlations in images. An example of the importance of long-range structure is the regular brick wall texture in the middle of the figure. Another example is the task of inpainting, where the goal is to fill in a missing part of an image, such that it is faithful to the non-missing pixels. Our main contribution is to introduce a way to deal with long-range structure using a simple modification to the product-based texture features. Our approach is based on imposing a “Markov-structure” on high-level features, allowing us to establish feature constraints that range across sites instead of being local. Unlike classical approaches to preserving spatial structure in image generation, such as Markov Random Fields and learning-based extensions (Roth & Black, 2005), our approach does not impose any explicit local constraints on pixels themselves. Rather, inspired by Gatys et al. (2015a), it encourages consistency to be satisfied on high-level features and on average across the whole image. We present applications to texture generation, inpainting and season transfer. ",
+        "bbox": [
+            173,
+            638,
+            825,
+            819
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 THE ARTISTIC STYLE ALGORITHM ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            839,
+            490,
+            854
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2.1 SYNTHESIS PROCEDURE ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            869,
+            383,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Given a reference texture, $x$ , the algorithm described in Gatys et al. (2015a) permits to synthesize by optimization a new texture $\\hat { x }$ similar to $x$ . To achieve this, the algorithm exploits an ImageNet ",
+        "bbox": [
+            174,
+            895,
+            823,
+            924
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/fdfb6b9e81db4b326e7af8723287b7dd54410b08bb7ba3391927271ae0a5b270.jpg",
+        "image_caption": [
+            "Figure 1: Reference image (left) and generated texture (right) using the procedure described in Gatys et al. (2015a). "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            178,
+            102,
+            821,
+            183
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/d62b0a03199a11debffeb0150e39840133ebb843e2d97dbc40ddfdcc856e8665.jpg",
+        "image_caption": [
+            "Figure 2: Summary of the texture synthesis procedure described in Gatys et al. (2015a). We use a VGG-19 network Simonyan & Zisserman (2014) as the pre-trained CNN. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            233,
+            803,
+            367
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "pre-trained model to define metrics suitable for describing textures: “Gram” matrices of feature maps, computed on top of $L$ selected layers. Formally, let $\\hat { N } ^ { l }$ be the number of maps in layer $l$ of a pre-trained CNN. The corresponding Gram matrix $G ^ { \\bar { l } }$ is a $N ^ { l } \\times N ^ { l }$ matrix defined as: ",
+        "bbox": [
+            174,
+            435,
+            825,
+            478
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "img_path": "images/52794122ec98d0dcfd2541a4c6babb5236a6429591371cb6356d46b63d7c5389.jpg",
+        "text": "$$\nG _ { i j } ^ { l } = \\frac { 1 } { M ^ { l } } \\sum _ { k = 1 } ^ { M ^ { l } } F _ { i k } ^ { l } F _ { j k } ^ { l } = \\frac { 1 } { M ^ { l } } \\langle F _ { i : } ^ { l } , F _ { j : } ^ { l } \\rangle\n$$",
+        "text_format": "latex",
+        "bbox": [
+            367,
+            486,
+            632,
+            531
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where $F _ { i : } ^ { l }$ is the $i ^ { t h }$ vectorized feature map of layer $l$ , $M ^ { l }$ is the number of elements in each map of this layer, and where $\\langle \\cdot , \\cdot \\rangle$ denotes the inner product. Equation 1 makes it clear that $G ^ { l }$ captures how feature maps from layer $l$ are correlated to each other. Diagonal terms, $G _ { i i } ^ { l }$ are the squared Frobenius norm of the $i ^ { t h }$ map $\\left\\| F _ { i : } ^ { l } \\right\\| _ { F } ^ { 2 }$ , so they represent its spatially averaged energy. We will discuss the Gramians in more detail in the next paragraph. Once the Gram matrices $\\{ G ^ { l } \\} _ { l \\in [ 1 , L ] }$ of the reference texture are computed, the synthesis procedure by Gatys et al. (2015a) amounts to constructing an image that produces Gram matrices $\\{ \\hat { G } ^ { l } \\} _ { l \\in [ 1 , L ] }$ that match the ones of the reference texture. More precisely, the following loss function is minimized with respect to the image being constructed: ",
+        "bbox": [
+            173,
+            539,
+            825,
+            675
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "img_path": "images/63a4422daa98d2165b447d1017c54d6cb11dad89bdff68b005809c83129c56bc.jpg",
+        "text": "$$\n\\mathcal { L } _ { s t y l e } = \\sum _ { l = 1 } ^ { L } w _ { l } \\left. \\hat { \\boldsymbol { G } } ^ { l } - \\boldsymbol { G } ^ { l } \\right. _ { F } ^ { 2 } = \\sum _ { l = 1 } ^ { L } w _ { l } \\mathcal { L } _ { s t y l e } ^ { l }\n$$",
+        "text_format": "latex",
+        "bbox": [
+            348,
+            671,
+            650,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where $w _ { l }$ is a normalizing constant similar to Gatys et al. (2015a). ",
+        "bbox": [
+            174,
+            718,
+            607,
+            733
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The overall process is summarized in Figure 2. While the procedure can be computationally expensive, there have been successful attempts reported recently which reduce the generation time (Ulyanov et al., 2016; Johnson et al., 2016). ",
+        "bbox": [
+            174,
+            739,
+            821,
+            782
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 WHY GRAM MATRICES WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            799,
+            421,
+            814
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Feature Gram matrices are effective at representing texture, because they capture global statistics across the image due to spatial averaging. Since textures are static, averaging over positions is required and makes Gram matrices fully blind to the global arrangement of objects inside the reference image. This property permits to generate very diverse textures by just changing the starting point of the optimization. Despite averaging over positions, coherence across multiple features needs to be preserved (locally) to model visually sensible textures. This requirement is taken care of by the off-diagonal terms in the Gram matrix, which capture the co-occurence of different features at a single spatial location. Indeed, Figure 3 shows that restricting the texture representation to the squared Frobenius norm of feature maps (i.e. diagonal terms) makes distinct object-parts from the reference texture encroach on each other in the reconstruction, as local coherence is not captured by the model. Exploiting off-diagonal terms improves the quality of the reconstruction as consistency across feature maps is enforced (on average across the image). ",
+        "bbox": [
+            173,
+            825,
+            825,
+            924
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/81370cb9bdd3f71e6f5fa79f9a4d5298de9b1732fa438e0865bb44a9d80ca202.jpg",
+        "image_caption": [
+            "Figure 3: Exploiting Gram matrices of feature maps as in Gatys et al. (2015a) $( 1 ^ { s t } r o w )$ or only the squared Frobenius norm of feature maps $( 2 ^ { n d } r o w )$ for increasingly deep layers (from left to right). "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            178,
+            101,
+            820,
+            277
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            344,
+            825,
+            415
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The importance of local coherence can be intuitively understood in the case of linear features (or in the lowest layer of a convolutional network): when decomposing an image using Gabor-like features, local structure can be expressed as the relative offsets in the Fourier phase angles between multiple different filter responses. A sharp step-edge, for example, requires the phases of local Fourier components at different frequencies to align in a different way than a blurry edge or a ridge (Morrone & Burr, 1988; Kovesi, 1999). Also, natural images exhibit very specific phaserelationships across frequency components in general, and destroying these makes the image look unnatural (Wang & Simoncelli, 2003). The same is not true of Fourier amplitudes (represented on the diagonals of the Gramian), which play a much less important role in the visual appearance (Oppenheim & Lim, 1981). In the case of deeper representations, the situation is more complex, but it is still local co-occurrence averaged over the whole image that captures texture. ",
+        "bbox": [
+            174,
+            421,
+            825,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Unfortunately, average local coherence falls short of capturing long-range structure in images. Spatial consistency is hard to capture within a single filter bank, because of combinatorial effects. Indeed, since Gram matrices capture coherence at a single spatial location, every feature would have to be matched to multiple transformed versions of itself. A corollary is that every feature would have to appear in the form of multiple transformed copies of itself in order to capture spatial consistency. However, this requirement clashes with the limited number of features available in each CNN layer. One way to address this is to use higher-layer features, whose receptive fields are larger. Unfortunately, as illustrated in Figure 3, even if using layers up to pool5 whose input receptive field covers the whole image1 (first row, last column), the reconstruction remains mainly unstructured and the method fails to produce spatial regularities. ",
+        "bbox": [
+            173,
+            580,
+            825,
+            719
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 MODELING SPATIAL CO-OCCURENCES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            741,
+            522,
+            756
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To account for spatial structure in images, we propose encoding this structure in the feature selfsimilarity matrices themselves. To this end, we suggest that, instead of computing co-occurences between multiple features within a map, we compute co-occurences between feature maps $F ^ { l }$ and spatially transformed feature maps $T ( F ^ { l } )$ , where $T$ denotes a spatial transformation. In the simplest case, $T$ represents local translation, which amounts to measuring similarities between local features and other neighbouring features. We denote by $T _ { x , + \\delta }$ the operation consisting in horizontally translating feature maps by $\\delta$ pixels and define the transformed Gramian: ",
+        "bbox": [
+            173,
+            770,
+            825,
+            868
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/f989a0266377fab63ae12c9434f6fda93bd38e9c405a3ac84da59025e6b7910d.jpg",
+        "text": "$$\nG _ { x , \\delta , i j } ^ { l } = \\frac { 1 } { M ^ { l } } \\langle T _ { x , + \\delta } \\left( F _ { i : } ^ { l } \\right) , T _ { x , - \\delta } \\left( F _ { j : } ^ { l } \\right) \\rangle\n$$",
+        "text_format": "latex",
+        "bbox": [
+            361,
+            872,
+            635,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a94944d3a051ed2f9aae98e632a49046bd878a60099da97d9cc1efc68fe4cea9.jpg",
+        "image_caption": [
+            "Figure 4: Computing the shifted Gram matrix for a given layer with feature maps of width $X$ . "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            178,
+            98,
+            823,
+            229
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $T _ { x , - \\delta }$ performs a translation in the opposite direction. As illustrated in Figure 4, the transformation in practice simply amounts to removing the $\\delta$ first or last columns from the raw feature maps. Therefore, the inner product now captures how features at position $( i , j )$ are correlated with features located at position $( i , j + \\delta )$ in average. While Figure 4 illustrates the case where feature maps are horizontally shifted, one would typically use translations along both the $x$ -axis and the $y$ -axis. Our transformed Gramians are related to Gray-Level Co-occurrence Matrices (GLCM) (Haralick et al., 1973) which compute the unnormalized frequencies of pixel values for a given offset in an image. While GLCMs have been mainly used for analysis, some work tried to use these features for texture synthesis (Lohmann, 1995). Usually, GLCMs are defined along 4 directions: $0 ^ { \\circ }$ and $9 0 °$ (i.e. horizontal and vertical offsets), as well as $4 5 ^ { \\circ }$ and $1 3 5 ^ { \\circ }$ (i.e. diagonal offsets). In comparison, our method does not consider diagonal offsets and captures spatial coherence on high-level features, making use of a pre-trained CNN, instead of working directly in the pixel domain. ",
+        "bbox": [
+            174,
+            285,
+            825,
+            453
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With this definition for transformed Gram matrices, we propose defining the loss as: $\\begin{array} { r l } { \\mathcal { L } } & { { } = } \\end{array}$ $\\mathcal { L } _ { s t y l e } + \\mathcal { L } _ { c c }$ , where $c c$ stands for cross-correlation. Like $\\mathcal { L } _ { s t y l e }$ , $\\mathcal { L } _ { c c }$ is a weighted sum of multiple losses $\\mathcal { L } _ { c c , \\delta } ^ { l }$ defined for several selected layers as the mean squared error between transformed Gram matrices of the reference texture and the one being constructed: ",
+        "bbox": [
+            174,
+            459,
+            825,
+            518
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/64689127ebd1aa3c0ded3747badfff4cb5be8504f48c528eed1f148a5298e960.jpg",
+        "text": "$$\n\\mathcal { L } _ { c c , \\delta } ^ { l } = \\frac { 1 } { 2 } \\left( \\mathcal { L } _ { c c , x , \\delta } ^ { l } + \\mathcal { L } _ { c c , y , \\delta } ^ { l } \\right) = \\frac { 1 } { 2 } \\left( \\left. \\hat { G } _ { x , \\delta } ^ { l } - G _ { x , \\delta } ^ { l } \\right. _ { F } ^ { 2 } + \\left. \\hat { G } _ { y , \\delta } ^ { l } - G _ { y , \\delta } ^ { l } \\right. _ { F } ^ { 2 } \\right)\n$$",
+        "text_format": "latex",
+        "bbox": [
+            248,
+            537,
+            750,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Although this amounts to adding more terms to a representation that was already high-dimensional and overparametrized, we found that these additional terms do not hurt the diversity of generated textures. Indeed, the new loss remains blind to the global arrangement of objects. In fact, there exists a specific situation where our approach is strictly equivalent to computing Gram matrices of deeper layers: with linear activations and “one-hot” convolution kernels2, deeper layers would simply contain translated versions of lower feature maps. In that case, computing Gram matrices of deeper layers would permit to directly capture cross-correlation statistics in lower ones. Nevertheless, this situation is very unlikely when using a pretrained CNN, as the network probably learned operations more useful than simple translations during its supervised training. ",
+        "bbox": [
+            173,
+            583,
+            825,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While we focus on translation for most of our results, we shall discuss other types of transformation in the experiments Section. ",
+        "bbox": [
+            174,
+            729,
+            823,
+            757
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 EXPERIMENTS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            780,
+            326,
+            796
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, we exploit the same normalized version of the VGG-19 network3 (Simonyan & Zisserman, 2014) as in Gatys et al. (2015a;b) and layers conv1 1, pool1, pool2, pool3, and pool4 are always used to define the standard Gram matrices. In our method, we did not use conv1 1 to define cross-correlation terms, as the large number of neurons at this stage makes the computation of Gram matrices costly. Corresponding $\\delta$ values for each layer are discussed in the next paragraph. ",
+        "bbox": [
+            176,
+            813,
+            825,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a4fe332fd30d43110d6b3cb6270896f60861c155adfa9dd5fdca304e451c2575.jpg",
+        "image_caption": [
+            "Figure 5: P ool2 reconstruction using different values for $\\delta$ in $\\mathcal { L } _ { c c , \\delta } ^ { p o o l 2 }$ . $\\boldsymbol { \\cdot } \\boldsymbol { 1 } ^ { s t }$ column): Without crosscorrelation terms. (Other columns): Using a single cross-correlation term with a fixed value $( 1 ^ { s t }$ row) or using multiple cross-correlation terms with distinct $\\delta$ values $( 2 ^ { n d } r o w )$ . "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            98,
+            795,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, our implementation4 uses Lasagne (Dieleman et al., 2015). Each image is of size $3 8 4 \\times 3 8 4$ Most textures used as references in this paper were taken from textures.com and pixabay.com. ",
+        "bbox": [
+            174,
+            373,
+            821,
+            402
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 EXPERIMENTS WITH TRANSLATION-GRAMIANS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            421,
+            544,
+            435
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The $\\delta$ parameter is of central importance as it dictates the range of the spatial constraints. We observed that the optimal value depends on both the considered layer in the pre-trained CNN and the reference texture, making it difficult to choose a value automatically. ",
+        "bbox": [
+            174,
+            449,
+            825,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For instance, Figure 5 shows generated images from the brick wall texture using only the pool2 layer with different $\\delta$ configurations. The first row depicts the results when considering single values of $\\delta$ only. While $\\delta = 4$ or $\\delta = 8$ are good choices, considering extreme long-range correlations does not help for this particular texture: a brick depends mostly on its neighbouring bricks and not the far-away ones. More precisely, a translation of more than 16 pixels in the pool2 layer makes the input receptive field move more than 64 pixels. Therefore $\\delta = 1 6$ or $\\delta = 3 2$ do not capture any information about neighbouring bricks. Unfortunately, this is not true for all textures, and $\\delta = 1 6$ or $\\delta = 3 2$ might be good choices for another image that exhibits longer structures. ",
+        "bbox": [
+            174,
+            497,
+            825,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Searching systematically for a $\\delta$ configuration that works well with the reference texture being considered would be a tedious task: even for a very regular texture with a periodic horizontal (or vertical) pattern, it is hard to guess the optimal $\\delta$ values for each layer (for deeper ones in particular). Instead, we propose to use a fixed but wide set of $\\delta$ values per layer, by defining the cost to be: $\\mathcal { L } _ { c c } ^ { l } = \\sum _ { k } \\mathcal { L } _ { c c , \\delta _ { k } } ^ { l }$ . A potential concern is that combining many loss terms can hurt the reconstruction or the diversity of generated textures. Figure 5 (second row) shows contrarily that there is no visual effect from using $\\delta$ values that are not specifically useful for the reference texture being considered: the rendering benefits from using $\\bar { \\delta } \\in \\{ 2 , \\dot { 4 } , 8 \\}$ while considering bigger values $( \\delta \\in \\{ 2 , 4 , 8 , 1 6 , 3 2 \\}$ , e.g.) does not help, but does not hurt the reconstruction either. We found the same to be true for other textures as well, and our results (shown in the next Section) show that combining the loss terms is able to generate very diverse textures. A drawback from using multiple cross-correlation terms per layers, however, is computational. We found that in our experimental setups, adding cross-correlation terms increases the generation time by roughly $8 0 \\%$ . ",
+        "bbox": [
+            173,
+            616,
+            825,
+            804
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As a guide-line, for image sizes of roughly $3 8 4 \\times 3 8 4$ pixels, we recommend the following $\\delta$ values per layer (which we used in all our following experiments): $\\{ 2 , 4 , 8 , 1 6 , 3 2 , 6 4 \\}$ for pool1, $\\{ 2 , 4 , 8 , 1 6 , 3 2 \\}$ for pool2, $\\{ 2 , 4 , 8 , 1 6 \\}$ for pool3, and $\\{ 2 , 4 , 8 \\}$ for pool4. The number and the range of $\\delta$ values decrease with depth because feature maps are getting smaller due to $2 \\times 2$ pooling layers. This configuration should be sufficient to account for spatial structure in any $3 8 4 \\times 3 8 4$ image. ",
+        "bbox": [
+            174,
+            811,
+            825,
+            896
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0733b00648dcc0afd39301437db05c2b2d3c89d6c966c8fe19133bf1c9e0c450.jpg",
+        "image_caption": [
+            "Figure 6: Some results of our approach compared with Gatys et al. (2015a). Only the initialization differs in the last two columns. Further results are shown in the supplementary material. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            220,
+            83,
+            776,
+            775
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 SYNTHESIS OF STRUCTURED TEXTURE EXAMPLES ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            842,
+            558,
+            856
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 6 shows the result of our approach applied to various structured and unstructured textures. It demonstrates that the method is effective at capturing long-range correlations without simply copying the content of the original texture. For instance, note how our model captures the depth aspect of the reference image in the third and fourth rows. ",
+        "bbox": [
+            174,
+            867,
+            825,
+            924
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3c7674352750d3be3a9314c029171b2f99b3d3205948b2e93a221b3ce6f6dba.jpg",
+        "image_caption": [
+            "Figure 7: Texture generation applied to in-painting. More in-painted images can be found in the supplementary material. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            209,
+            102,
+            790,
+            299
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd51b4bc2b046d0074a607813aa7e8c25954f2a16b93452368583276119ba3e2.jpg",
+        "image_caption": [
+            "Figure 8: Season transfer examples. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            222,
+            359,
+            777,
+            526
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The problem of synthesizing near-regular structures is challenging because stochasticity and regularity are adversarial properties (Lin et al., 2006). Non-parametric patch-based techniques, such as Efros & Freeman (2001), are better suited for this task because they can $t i l e ^ { 5 }$ the reference image. On the other hand, regular structures are usually more problematic for parametric statistical models. Nevertheless, the two first rows of Figure 6 demonstrate that our approach can produce good visual results and can reduce the gap to patch-based methods on these kinds of texture. ",
+        "bbox": [
+            174,
+            585,
+            825,
+            669
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Even if the reference image is not a texture, the generated images in the last row (Leonardo Dicaprio’s face) provide a good visual illustration of the effect of translation terms. In contrast to Gatys et al., our approach preserves longer-range structure, such as the alignment and similar appearance of the eyes, hair on top of the forehead, the chin below the mouth, etc. Finally, when the reference texture is unstructured (fifth row), our solution does not necessarily provide a benefit, but it also does not hurt the visual quality or the diversity of the generated textures. ",
+        "bbox": [
+            174,
+            676,
+            825,
+            761
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 INPAINTING APPLICATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            779,
+            395,
+            792
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Modelling long-range correlations can make it possible to apply texture generation to inpainting, because it allows us to impose consistency constraints between the newly rendered region and the unmodified parts of the original image. ",
+        "bbox": [
+            174,
+            805,
+            825,
+            848
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To apply our approach to texture inpainting, we extracted two patches from the original image: one that covers the whole area to inpaint, and another one that serves as the reference texture. Then, approximately the same process as for texture generation is used, with the following two modifications: First, instead of random noise, the optimization starts from the masked content patch (the one to inpaint) showing a grey area and its non-missing surrounding. Second, we encourage the borders of the output to not change much with respect to the original image using an $L _ { 2 }$ penalty. We apply the penalty both in the Gatys et al. rendering and in ours. Some inpainted images are shown in Figure 7. As seen in the figure, our solution significantly outperforms Gatys et al. in terms of visual quality. Further results are shown in the supplementary material. ",
+        "bbox": [
+            176,
+            854,
+            823,
+            897
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1fbcc91968bc06348295a657a024bdfd88b21ab99f4a0f4a49d20e11ff2acdfd.jpg",
+        "image_caption": [
+            "Figure 9: Generation of abstract symmetric textures. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            243,
+            101,
+            754,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            488,
+            825,
+            573
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4 SEASON TRANSFER ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            604,
+            348,
+            618
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 8 shows the result of applying our approach to a style transfer task, as in Gatys et al. (2015b): transferring the “season” of a landscape image to another one. On this task, the results from our approach are similar to those from Gatys et al. (2015b). Nevertheless, in contrast to the Gatys et al. results, our approach seems to better capture global information, such as sky color and leaves (bottom row), or the appearance of branches in the winter image (top row). ",
+        "bbox": [
+            173,
+            636,
+            825,
+            707
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 INCORPORATING OTHER TYPES OF STRUCTURE ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            739,
+            539,
+            752
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "While we focused on feature map translations in most of our experiments, other transformations can be applied as well. To illustrate this point, we explored a way to generate symmetric textures using another simple transformation. To this end, we propose flipping one of the two feature maps before computing the Gram matrices: $G _ { l r , i j } ^ { l } = \\langle F _ { i : } ^ { l } , T _ { l r } \\ ' ( F _ { j : } ^ { l } ) \\rangle$ . Here $T _ { l r }$ corresponds to the left-right flipping operation, but we also considered up-down flipping of feature maps: $\\mathcal { L } ^ { l } = \\mathcal { L } _ { s t y l e } ^ { l } + \\mathcal { L } _ { l r } ^ { l } +$ L  lud . ",
+        "bbox": [
+            173,
+            770,
+            825,
+            862
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As can be seen in Figure 9, in contrast to Gatys et al., the additional loss terms capture which objects are symmetric in the reference texture, and enforce these same objects to be symmetric in the reconstruction as well. Other kinds of transformation could be used, depending on the type of property in the source texture one desires to preserve. ",
+        "bbox": [
+            174,
+            867,
+            825,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            318,
+            117
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We presented an approach to satisfying long-range consistency constraints in the generation of images. It is based on a variation of the method by Gatys et al., and considers spatial co-occurences of local features (instead of only co-occurences across features). We showed that the approach permits to generate textures with various global symmetry properties and that it makes it possible to apply texture generation to in-painting. Since it preserves correlations across sites, the approach is reminiscent of an MRF, but in contrast to an MRF or other graphical models, it defines correlationconstraints on high-level features of a (pre-trained) CNN rather than on pixels. ",
+        "bbox": [
+            174,
+            133,
+            825,
+            231
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            251,
+            285,
+            265
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Sander Dieleman, Jan Schlter, Colin Raffel, et al. Lasagne: First release., August 2015. URL http://dx. doi.org/10.5281/zenodo.27878.   \nA. A. Efros and T. K. Leung. Texture synthesis by non-parametric sampling. In Proceedings of the International Conference on Computer Vision-Volume 2, ICCV ’99, pp. 1033–, 1999.   \nAlexei A. Efros and William T. Freeman. Image Quilting for Texture Synthesis and Transfer. Proceedings of SIGGRAPH 2001, pp. 341–346, August 2001.   \nL. A. Gatys, A. S. Ecker, and M. Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems 28, 2015a. URL http://arxiv.org/abs/1505.07376.   \nL. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm of artistic style. CoRR, abs/1508.06576, 2015b. URL http://arxiv.org/abs/1508.06576.   \nR. Haralick, K. Shanmugam, and I. Dinstein. Texture features for image classification. IEEE Transactions on Systems, Man, and Cybernetics 3 (6), 1973.   \nDavid J. Heeger and James R. Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1995, pp. 229–238, 1995.   \nJ. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. CoRR, abs/1603.08155, 2016. URL http://arxiv.org/abs/1603.08155.   \nP. Kovesi. Image features from phase congruency. Videre, 1(3):2–27, 1999.   \nV. Kwatra, A. Schdl, I. Essa, G. Turk, and A. Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Transactions on Graphics, SIGGRAPH 2003, 22(3):277–286, 2003.   \nWen-Chieh Lin, James Hays, Chenyu Wu, V. Kwatra, and Yanxi Liu. Quantitative evaluation on near regular texture synthesis. In CVPR ’06, volume 1, pp. 427 – 434, June 2006.   \nG. Lohmann. Analysis and synthesis of textures: A co-occurrence-based approach. Computers and Graphics, 19(1), pp. 29–36, 1995.   \nM. C. Morrone and D. C. Burr. Feature detection in human vision: A phase-dependent energy model. Proceedings of the Royal Society of London B: Biological Sciences, 235(1280):221–245, 1988.   \nA. V. Oppenheim and J. S. Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529–541, 1981. ISSN 0018-9219.   \nJ. Portilla and E. P. Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. Int. J. Comput. Vision, 40(1):49–70, 2000.   \nS. Roth and M. J. Black. Fields of experts: A framework for learning image priors. In IEEE Conf. on Computer Vision and Pattern Recognition, volume 2, pp. 860–867, June 2005.   \nO. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015.   \nK. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014.   \nD. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In International Conference on Machine Learning (ICML), 2016.   \nI. Ustyuzhaninov, W. Brendel, L. A. Gatys, and M. Bethge. Texture synthesis using shallow convolutional networks with random filters. CoRR, abs/1606.00021, 2016. URL http://arxiv.org/abs/1606. 00021.   \nZ. Wang and E. P. Simoncelli. Local phase coherence and the perception of blur. In Advances in Neural Information Processing Systems 16 [NIPS 2003], pp. 1435–1442, 2003.   \nCiyou Zhu, Richard H. Byrd, Peihuang Lu, and Jorge Nocedal. L-bfgs-b - fortran subroutines for large-scale bound constrained optimization. Technical report, ACM Trans. Math. Software, 1994. ",
+        "bbox": [
+            171,
+            266,
+            826,
+            926
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            104,
+            826,
+            213
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Supplementary material ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            99,
+            529,
+            127
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A TEXTURE GENERATION ",
+        "bbox": [
+            176,
+            150,
+            405,
+            165
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/7996b087aac1120f4135dc33c1be02fad01fda7f4f161d9c89a75c42da2f811e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            181,
+            174,
+            820,
+            828
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/10a8cc26ef285ca401b2e76fa624cfbae157f2de6a8075520d9bece66f04319f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            178,
+            176,
+            821,
+            830
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B INPAINTING ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            102,
+            310,
+            117
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/6201af3b350a06711412fc7b240a36cc657b5ef4cdbcb06c4bde514444bda001.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            140,
+            825,
+            308
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/455b2d1f33375b4133b4b298627eb4879109e41a4f9f13fe6e66715d6daae2d2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            245,
+            373,
+            753,
+            540
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/267bf72c6aca6e0e1e6e883796f77dbea9556357f14c4d5ea7c1d3cb0399c383.jpg",
+        "image_caption": [
+            ""
+        ],
+        "image_footnote": [],
+        "bbox": [
+            243,
+            569,
+            450,
+            727
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2dfcfb15fed6909b957b975612920289da35cb18b6401e73af4ecbb17c18f92.jpg",
+        "image_caption": [
+            "Gatys et al. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            545,
+            569,
+            751,
+            728
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/f5cbb699d907e2eac9878e99f3830544c0028945aa7c44e4d26985e0004266f9.jpg",
+        "image_caption": [
+            "Inpainted image "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            245,
+            760,
+            751,
+            904
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/675e22dc39763b4f655d38d605aba2ad35e17199acb9830163b57035fd91d414.jpg",
+        "image_caption": [
+            "Gatys et al. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            102,
+            797,
+            305
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9f6f36c89d7a8c86ebae09f8f61b55f4a623c6064c1fded1a16a1696709d7e0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            383,
+            802,
+            838
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/3bae23c5542923774b38a947f266d94da913576141432c0f47b112ea97f15e13.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            104,
+            799,
+            311
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/4bed86a0169d629cb1ae84d288ad1049dcdacfd700a429b72c1293c00fe392e8.jpg",
+        "table_caption": [
+            "Table 1: Final Gatys et al. loss for all inpainted textures (by order of appearance). "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=2>Gatys et al. loss</td></tr><tr><td rowspan=1 colspan=1>Texture</td><td rowspan=1 colspan=1>Gatys et al.</td><td rowspan=1 colspan=1>Ours</td></tr><tr><td rowspan=1 colspan=1>&quot;red_leaves”</td><td rowspan=1 colspan=1>3.38e-4</td><td rowspan=1 colspan=1>4.72e-4</td></tr><tr><td rowspan=1 colspan=1>“floor”</td><td rowspan=1 colspan=1>3.68e-4</td><td rowspan=1 colspan=1>4.73e-4</td></tr><tr><td rowspan=1 colspan=1>Supplementa</td><td rowspan=1 colspan=1>rymaterialtex</td><td rowspan=1 colspan=1>tures</td></tr><tr><td rowspan=1 colspan=1>&quot;leaves&quot;</td><td rowspan=1 colspan=1>4.08e-4</td><td rowspan=1 colspan=1>5.14e-4</td></tr><tr><td rowspan=1 colspan=1>“cargo”</td><td rowspan=1 colspan=1>3.39e-4</td><td rowspan=1 colspan=1>4.73e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;building”</td><td rowspan=1 colspan=1>8.90e-4</td><td rowspan=1 colspan=1>6.67e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;building2&quot;</td><td rowspan=1 colspan=1>7.19e-4</td><td rowspan=1 colspan=1>7.78e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;cargo2&quot;</td><td rowspan=1 colspan=1>1.05e-3</td><td rowspan=1 colspan=1>1.18e-3</td></tr><tr><td rowspan=1 colspan=1>&quot;small_bricks&quot;</td><td rowspan=1 colspan=1>1.21e-3</td><td rowspan=1 colspan=1>6.82e-4</td></tr><tr><td rowspan=1 colspan=1>&quot;football_team&quot;</td><td rowspan=1 colspan=1>4.56e-3</td><td rowspan=1 colspan=1>5.68e-3</td></tr></table>",
+        "bbox": [
+            354,
+            335,
+            643,
+            510
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 1 reports the final Gatys et al. losses obtained by both approaches for all inpainted textures showed previously. In most cases, the Gatys et al. (2015a) approach converges to a smaller value, which is not surprising since our method optimizes a modified version of the loss. Table 1, combined with the visual aspect of Gatys et al. inpainted renderings, illustrates that only imposing local coherence is not enough to obtain good inpainted textures. Our texture representation seems to be better suited for this task. ",
+        "bbox": [
+            173,
+            559,
+            825,
+            642
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/2231dc59db139e901ed523d9c558bec18b16d8935cd502936456659b797a3cde.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            75,
+            803,
+            917
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/e0fa62ff4230f9c419d3d39bd52bfa9c05b0730b5ff3652b369555d4cfae2a42.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            78,
+            799,
+            698
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/7cd2b08cfe5550333150be8478841c746a249da2ce20540ee9b25ad7a09d0aed.jpg",
+        "image_caption": [
+            "Style "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            154,
+            338,
+            318
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/b11d2a55d9752e5cc1349d9e99b1d8ff309f0a3b2e36271d669c7c05e237643f.jpg",
+        "image_caption": [
+            "Content (&init) "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            155,
+            491,
+            233
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/32cae7131934dfbed3305598754f0968fe23b5e3a724f9b5c832ee8385ab0c1a.jpg",
+        "image_caption": [
+            "Ours "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            154,
+            643,
+            233
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/6ee592082e45d87c3ff944ef1fa30d5b08a60546fcc2adc7e12c8e37868b8f0b.jpg",
+        "image_caption": [
+            "Gatys "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            239,
+            643,
+            319
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/4979210ace33ad8aadca5edafe85398ff90db694cd7f16502bfe8ea8c4682526.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            239,
+            490,
+            319
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/66541a0ba406f23dbfab5e6aa52db140a57333d83d1f90070ddc9efc7e3eea6b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            414,
+            800,
+            598
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/af9da92cc614db1ca298004e54e4372697b826ce6efbb764deff7b922b87e852.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            690,
+            802,
+            875
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "text",
+        "text": "C SYMMETRIC TEXTURES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            102,
+            406,
+            117
+        ],
+        "page_idx": 18
+    },
+    {
+        "type": "image",
+        "img_path": "images/aee878193ca11f6507957dcf5903f3dd8195b3beabcbeae2cdf46ef0386165a5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            132,
+            799,
+            868
+        ],
+        "page_idx": 18
+    }
+]

parse/train/NJS8kp15zzH/NJS8kp15zzH_content_list.json

@@ -0,0 +1,850 @@
+[
+    {
+        "type": "text",
+        "text": "Rethinking Calibration of Deep Neural Networks: Do Not Be Afraid of Overconfidence ",
+        "text_level": 1,
+        "bbox": [
+            199,
+            122,
+            803,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deng-Bao Wang,1,2 Lei Feng,3 Min-Ling Zhang1,2∗ ",
+        "bbox": [
+            320,
+            219,
+            681,
+            236
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1School of Computer Science and Engineering, Southeast University, Nanjing 210096, China   \n2Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China 3College of Computer Science, Chongqing University, Chongqing, 400044, China wangdb@seu.edu.cn, lfeng@cqu.edu.cn, zhangml@seu.edu.cn ",
+        "bbox": [
+            196,
+            237,
+            805,
+            308
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract ",
+        "text_level": 1,
+        "bbox": [
+            462,
+            319,
+            535,
+            334
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Capturing accurate uncertainty quantification of the predictions from deep neural networks is important in many real-world decision-making applications. A reliable predictor is expected to be accurate when it is confident about its predictions and indicate high uncertainty when it is likely to be inaccurate. However, modern neural networks have been found to be poorly calibrated, primarily in the direction of overconfidence. In recent years, there is a surge of research on model calibration by leveraging implicit or explicit regularization techniques during training, which achieve well calibration performance by avoiding overconfident outputs. In our study, we empirically found that despite the predictions obtained from these regularized models are better calibrated, they suffer from not being as calibratable, namely, it is harder to further calibrate these predictions with post-hoc calibration methods like temperature scaling and histogram binning. We conduct a series of empirical studies showing that overconfidence may not hurt final calibration performance if post-hoc calibration is allowed, rather, the penalty of confident outputs will compress the room of potential improvement in post-hoc calibration phase. Our experimental findings point out a new direction to improve calibration of DNNs by considering main training and post-hoc calibration as a unified framework. ",
+        "bbox": [
+            233,
+            337,
+            766,
+            570
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            594,
+            310,
+            611
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Modern over-parameterized deep neural networks (DNNs) have been shown to be very powerful modeling tools for many prediction tasks involving complex input patterns [37]. In addition to obtaining accurate predictions, it is also important to capture accurate quantification of prediction uncertainty from deep neural networks in many real-world decision-making applications. A reliable predictive model should be accurate when it is confident about its predictions and indicate high uncertainty when it is likely to be inaccurate. However, modern DNNs trained with cross-entropy (CE) loss, despite being highly accurate, have been recently found to predict poorly calibrated probabilities, unlike traditional models trained with the same objective [4]. The overconfident predictions of DNNs could cause undesired consequences in safety-critical applications such as medical diagnosis and autonomous driving. Bayesian DNNs, which indirectly infer prediction uncertainty through weight uncertainties, have innate abilities to represent the model uncertainty [2, 16]. But training and inferring those bayesian models are computationally more expensive and conceptually more complicated than non-bayesian models, and their performance depends on the form of approximation made due to computational constraints. Therefore, the study on uncertainty calibration of deterministic DNNs is important for both development practice and the perspective of understanding DNNs. ",
+        "bbox": [
+            174,
+            626,
+            825,
+            845
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Post-hoc calibration addresses the miscalibration problem by equipping a given neural network with an additional parameterized calibration component, which can be tuned with a hold-out validation dataset. Guo et al. [4] experimented with several classical calibration fixes and found that simple post-hoc methods like Temperature Scaling (TS) [25] and Histogram Binning (HB) [33] are significantly effective for DNNs. The authors of [10] and [27] proposed to learn linear and non-linear transformation functions to rescale the original output logits respectively. Gupta et al. [5] proposed to obtain a calibration function by approximating the empirical cumulative distribution of output probabilities via splines. Kumar et al. [11] proposed to integrate TS with HB to achieve more stable calibration performance. Patel et al. [24] proposed a mutual information maximization-based binning strategy to solve the severe sample-inefficiency issue in HB. ",
+        "bbox": [
+            178,
+            853,
+            823,
+            881
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            90,
+            825,
+            202
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, there is another line of research which presents a possibility of improving the calibration quality of deterministic DNNs via regularization during training. Guo et al. [4] found that training DNNs with strong weight decay, which used to be the predominant regularization mechanism for training neural networks, has a positive impact on calibration. Müller et al. [19] showed that training models using the standard CE loss with label smoothing [28], instead of one-hot labels, has a very favourable effect on model calibration. Mukhoti et al. [18] proposed to improve uncertainty calibration by replacing the conventionally used CE loss with the focal loss proposed in [14] when training DNNs. It is important to note that CE loss with label smoothing and focal loss can be considered as standard CE with an additional maximum-entropy regularizer, which means minimizing these losses is equivalent to minimizing CE loss and maximizing the entropy of the predicted distribution simultaneously [18, 17]. Following these studies, a recent work [7] explored several explicit regularization techniques for improving the predictive uncertainty calibration directly. ",
+        "bbox": [
+            174,
+            208,
+            825,
+            375
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we conduct an empirical study showing that despite the predictions obtained from the regularized models are well calibrated, they suffer from worse calibratable, namely, it is harder to further improve the calibrate performance with post-hoc calibration methods like temperature scaling and histogram binning. We found that the regularization works by simply aligning the average confidence of the whole dataset to the accuracy with some specific regularization strengths, and cannot achieve fine-grained calibration. The comparison results show that when post-hoc calibration methods are allowed, the standard CE loss yields better calibration performance than those regularization methods. The extended experiments demonstrate that regularization will make DNNs lose the important information about the hardness of samples, which results in compressing the room of potential improvement by post-hoc calibration. Based on the experimental findings, we raise a natural question: can we design new loss functions in the opposite direction of these regularization methods to further improve the calibration performance? To this end, we propose inverse focal loss, and empirically found that it can learn more calibratable models in some cases compared with the CE loss, though it causes severer overconfidence problem without post-hoc calibration. Most importantly, our findings show that overconfidence of DNNs is not the nightmare in uncertainty qualification and point out a new direction to improve the calibration of DNNs by considering main training and post-hoc calibration as a unified framework. ",
+        "bbox": [
+            174,
+            381,
+            825,
+            616
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preliminaries ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            637,
+            318,
+            655
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Let $\\mathcal { V } = \\{ { 1 , . . . , K } \\}$ denote the label space and $\\mathcal { X } = \\mathbb { R } ^ { d }$ denote the feature space. Given a sample $( { \\pmb x } , y ) \\in \\mathcal { X } \\times \\mathcal { Y }$ sampled from an unknown distribution, a learned neural network classifier $f ^ { \\theta } :$ $\\mathcal { X } \\to \\Delta ^ { K }$ can produce a probability distribution for $_ { \\textbf { \\em x } }$ on $K$ classes, where $\\Delta ^ { K }$ denotes the $K - 1$ dimensional unit simplex. Here we assume $f ^ { \\theta }$ as a composition of a non-probabilistic $K$ -way classifier $g ^ { \\theta }$ and a softmax function $\\sigma$ , i.e. $\\dot { \\pmb { f } } ^ { \\theta } = \\pmb { g } ^ { \\theta } \\circ \\pmb { \\sigma }$ . For a query instance $_ { \\textbf { \\em x } }$ , $f ^ { \\theta }$ gives its probability of assigning it to label i as exp(g i P (x))Kk=1 exp(gθk(x)) , where $g _ { i } ^ { \\theta } ( { \\pmb x } )$ denotes the $i$ -th element of the logit vector produced by $g ^ { \\theta }$ . Then, ${ \\hat { y } } : = \\arg \\operatorname* { m a x } _ { i } f _ { i } ^ { \\theta } ( { \\pmb x } )$ can be returned as the predicted label and ${ \\hat { p } } : = \\operatorname* { m a x } _ { i } f _ { i } ^ { \\theta } ( { \\pmb x } )$ can be treated as the associated confidence score. ",
+        "bbox": [
+            173,
+            670,
+            825,
+            796
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Expected Calibration Error (ECE) For a well-calibrated model, $\\hat { p }$ is expected to represent the true probability of correctness. Formally, a perfectly calibrated model satisfies $\\mathbb { P } ( \\boldsymbol { \\hat { y } } = \\boldsymbol { y } | \\boldsymbol { \\hat { p } } = \\boldsymbol { p } ) = \\boldsymbol { p }$ for any $p \\in [ 0 , 1 ]$ . In practice, ECE [20] is a commonly used calibration metric from finite samples. It works by firstly grouping all samples (let $n$ denote the number of samples) into $M$ equally interval bins betw $\\{ \\bar { B _ { m } } \\} _ { m = 1 } ^ { M }$ with respect to their confideuracy and average confidence: c e. $\\begin{array} { r } { \\mathrm { E C E } = \\sum _ { m = 1 } ^ { M } \\frac { | B _ { m } | } { n } | \\mathrm { a c c } ( B _ { m } ) - \\mathrm { a v g } \\mathrm { C o n f } ( B _ { m } ) | . } \\end{array}$ ",
+        "bbox": [
+            174,
+            801,
+            825,
+            891
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Temperature Scaling By scaling the logits produced by $g ^ { \\theta }$ with a temperature $T$ , the sharpness of output probabilities can be changed. Formally, after adding TS, the new prediction confidence can be expressed as: $\\begin{array} { r } { \\hat { p } = \\operatorname* { m a x } _ { i } \\frac { \\exp ( g _ { i } ^ { \\theta } ( \\pmb { x } ) / T ) } { \\sum _ { k = 1 } ^ { K } \\exp ( g _ { k } ^ { \\theta } ( \\pmb { x } ) / T ) } } \\end{array}$ . The temperature softens the output probability with and sharpens the probability with $T < 1$ . As , the output probability collapses to one-hot vector. As $T \\to \\infty$ , the output probability approaches to a uniform distribution. After training of the model, $T$ can be tuned on a hold-out validation set by optimization methods. ",
+        "bbox": [
+            173,
+            896,
+            821,
+            912
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            90,
+            825,
+            170
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Histogram Binning is a non-parametric calibration approach. Given an uncalibrated model, all the prediction confidences of validation samples can be divided into mutually exclusive $N$ bins $\\{ B _ { n } \\} _ { n = 1 } ^ { N }$ according to a set of intervals $\\{ I _ { n } \\} _ { n = 1 } ^ { N + 1 }$ which partitions [0, 1]. Each bin is assigned a confidence score , which can be simply set to the corresponding accuracy of samples in each bin. If the uncalibrated confidence $\\hat { p }$ of a query instance falls into bin $B _ { n }$ , then the calibrated confidence is $\\eta _ { n }$ . The bins can be chosen by two simple schemes: equal size binning (uniformly partitioning the probability interval in $[ 0 , 1 ] \\cdot$ ) and equal mass binning (uniformly distributing samples over bins). Note that although the HB scheme is simple to implement and was demonstrated to achieve good calibration results in some datasets, it makes the predictor only produce very sparse confidence distribution, and compromises the many legitimately confident predictions. ",
+        "bbox": [
+            173,
+            175,
+            826,
+            316
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Regularization in Neural Networks for Calibration ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            334,
+            629,
+            352
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In recent years, there is a surge of research on model calibration by leveraging implicit or explicit regularization techniques during training of DNNs, which makes better calibrated predictions by avoiding the overconfident outputs. In this section, we firstly review three representative regularization methods and then empirically show their improvements on ECE compared with the baseline. ",
+        "bbox": [
+            174,
+            366,
+            825,
+            421
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Label Smoothing is widely used as a means to reduce overfitting of DNNs. The mechanism of LS is simple: when training with CE loss, the one-hot label vector $\\textbf {  { y } }$ is replaced with soft label vector $\\widetilde { \\pmb { y } }$ , whose elements can be formally denoted as $\\widetilde { y _ { i } } = ( 1 - \\epsilon ) y _ { i } + \\epsilon / K , \\forall i \\in \\{ 1 , . . . , K \\}$ , where $\\epsilon > 0$ e eis a strength coefficient. Müller et al. [19] demonstrated that label smoothing implicitly calibrates DNNs by preventing the networks from becoming overconfident. Let $\\mathcal { L } _ { c e }$ denote the CE loss, then the following equation holds: ",
+        "bbox": [
+            173,
+            426,
+            825,
+            511
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/e71628b96ddde4746861a10ae3be295dc323b28f5b37aa41962e64381ebf6a96.jpg",
+        "text": "$$\n\\mathcal { L } _ { c e } ( \\widetilde { \\pmb { y } } , \\pmb { f } ^ { \\theta } ) = ( 1 - \\epsilon ) \\mathcal { L } _ { c e } ( \\pmb { y } , \\pmb { f } ^ { \\theta } ) + \\epsilon \\mathcal { L } _ { c e } ( \\pmb { u } , \\pmb { f } ^ { \\theta } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            339,
+            516,
+            658,
+            535
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This can be simply proved. Therefore, minimizing CE loss between smoothed labels and the model outputs is equivalent to adding a confidence penalty term, i.e., a weighted CE loss between the uniform distribution $\\textbf { \\em u }$ and the model outputs, to the original CE loss. ",
+        "bbox": [
+            174,
+            540,
+            825,
+            583
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "$L _ { p }$ Norm in the Function Space is one of the explicit regularization methods for calibration investigated by the recent work [7]. For a real number $p \\geq 1$ , the $L _ { p }$ Norm of a vector $_ { z }$ with dimension $n$ can be expressed as: $\\begin{array} { r } { \\| z \\| _ { p } = ( \\sum _ { i = 1 } ^ { n } | z _ { i } | ^ { p } ) ^ { 1 / p } } \\end{array}$ . By adding $L _ { p }$ Norm of logits $g ^ { \\theta }$ with a weighting coefficient $\\alpha$ into final objective function, i.e. $\\mathcal { L } _ { L _ { p } } ( y , f ^ { \\theta } ) = \\mathcal { L } _ { c e } ( y , f ^ { \\theta } ) + \\alpha \\left\\| g ^ { \\theta } \\right\\| _ { p }$ , the function complexity of neural networks can be directly penalized during training. ",
+        "bbox": [
+            174,
+            588,
+            825,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Focal Loss is originally proposed to address the class imbalance problem in object detection. By reshaping the standard CE loss through weighting loss components of all samples according to how well the model fits them, focal loss focuses on fitting hard samples and prevents the easy samples from overwhelming the training procedure. Formally, for classification tasks where the target distribution is one-hot encoding, it is defined as: $\\mathcal { L } _ { f } = - \\dot { ( } 1 - f _ { y } ^ { \\theta } ) ^ { \\gamma } \\log f _ { y } ^ { \\theta }$ , where $\\gamma$ is a predefined coefficient. Mukhoti et al. [18] found that the models learned by focal loss produce output probabilities which are already very well calibrated. Interestingly, they also showed that focal loss is an upper bound of the regularized KL-divergence, which can be expressed formally as follows: ",
+        "bbox": [
+            173,
+            671,
+            825,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/f30c0b546825f3f7824b27ff24d053a857730eaf5ca82188b85567b4e85a0cd7.jpg",
+        "text": "$$\n\\mathcal { L } _ { f } \\geq \\mathrm { K L } ( \\pmb { y } | | \\pmb { f } ^ { \\theta } ) - \\gamma \\mathrm { H } ( \\pmb { f } ^ { \\theta } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            401,
+            789,
+            594,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where $\\operatorname { H } ( p )$ denotes the entropy of distribution $\\pmb { p }$ . This upper bound property shows that replacing the CE loss with focal loss has the effect of adding a maximum-entropy regularizer. ",
+        "bbox": [
+            173,
+            813,
+            825,
+            842
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Empirical Comparison ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            857,
+            372,
+            872
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We conduct a comparison study of the above regularization methods on four commonly used datasets. We train ResNet-32 [6] models on SVHN [21], CIFAR-10/100 [9] and train a 8-layer 1D-CNN model on 20Newsgroups [13], using the standard CE loss and the above regularized losses respectively, with state-of-the-art learning policy settings (see implementation details in Appendix). For Norm regularization, we use $L _ { 1 }$ Norm, which has been shown effective for calibration despite its simple form [7]. Table 1 shows the comparison of these methods. Note that for each of the above three regularization methods, there is a coefficient, i.e., \u000f, $\\alpha$ and $\\gamma$ , controlling the strength of regularization. We conduct experiments using these methods with the following coefficient settings: $\\{ 0 . 0 1 , 0 . 0 3 , 0 . 0 5 , 0 . 0 7 , 0 . 0 \\bar { 9 } \\}$ for label smoothing, $\\{ 0 . 0 0 1 , 0 . 0 0 5 , 0 . 0 1 , 0 . 0 5 , 0 . 1 \\}$ for $L _ { 1 }$ Norm and $\\{ 1 , 3 , 5 , 7 , 9 \\}$ for focal loss. And we choose the best coefficient for each method and dataset, according to their ECE directly on test data. ",
+        "bbox": [
+            173,
+            882,
+            823,
+            911
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/aa924487088cb9c81fc7600c2d9242b8190b39772084005cb0c773f18f2b4cbc.jpg",
+        "table_caption": [
+            "Table 1: Comparison results (mean±std) of ECE $( \\% )$ with $M = 1 5$ and predictive accuracy $( \\% )$ over 5 random runs. The values with underline in first row represent the chosen coefficients of each regularization method on four datasets according to the ECE on test data. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td></td><td>Cross-Entropy</td><td>Label Smoothing 0.01/0.05/0.09/0.09</td><td>L1Norm 0.01/0.05/0.01/0.01</td><td>Focal Loss 1/3/5/5</td></tr><tr><td rowspan=\"2\">SVHN</td><td>ECE</td><td>3.03±0.16</td><td>1.84±0.19</td><td>1.85±0.04</td><td>1.01±0.21</td></tr><tr><td>Accuracy</td><td>95.00±0.27</td><td>95.21±0.23</td><td>95.29±0.13</td><td>94.77±0.19</td></tr><tr><td rowspan=\"2\">CIFAR-10</td><td>ECE</td><td>6.43±0.22</td><td>2.72±0.32</td><td>2.93±0.39</td><td>3.00±0.26</td></tr><tr><td>Accuracy</td><td>90.46±0.23</td><td>90.09±0.41</td><td>90.06±0.59</td><td>87.84±0.17</td></tr><tr><td rowspan=\"2\">CIFAR-100</td><td>ECE</td><td>19.53±0.36</td><td>2.27±0.48</td><td>8.07±0.44</td><td>2.34±0.35</td></tr><tr><td>Accuracy</td><td>64.64±0.43</td><td>63.73±0.67</td><td>63.07±0.29</td><td>60.36±0.44</td></tr><tr><td rowspan=\"2\">20 Newsgroups</td><td>ECE</td><td>20.82±0.93</td><td>5.85±0.64</td><td>13.31±0.56</td><td>3.82±0.51</td></tr><tr><td>Accuracy</td><td>72.85±0.89</td><td>72.81±0.26</td><td>73.61±0.80</td><td>59.17±1.81</td></tr></table>",
+        "bbox": [
+            181,
+            137,
+            812,
+            299
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            316,
+            826,
+            441
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "From the results of Table 1, it is obvious that the regularization methods significantly decrease the ECE on all datasets, compared with the standard CE loss. The prediction accuracy results are also reported. When the strength coefficients of $L _ { 1 }$ Norm and focal loss are large, their predictive performances are harmed. Especially, $L _ { 1 }$ Norm fails on CIFAR-100 and 20Newsgroups when $\\alpha \\ge 0 . 0 5$ , thus $\\alpha$ is chosen from $\\{ 0 . 0 0 1 , 0 . 0 0 5 , 0 . 0 1 \\}$ on these two datasets. ",
+        "bbox": [
+            174,
+            448,
+            825,
+            517
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Does Regularization Really Help Calibration? ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            536,
+            588,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown by the above empirical results, the regularization methods do help the calibration of DNNs during training, especially alleviate the overconfidence issue caused by the standard CE loss. In this section, we empirically investigate their calibration performance when integrating them with post-hoc calibration. After training, we use the post-hoc methods TS and HB to further calibrate the output probabilities. For TS, we simply search the best temperature in the temperature pool {0.01,0.02...,10} on the validation set (see data splits in Appendix). For HB, we use equal size binning scheme on the top-1 prediction of all classes with bin number set as 15. The experimental details used in this section are the same with those in Section 3. ",
+        "bbox": [
+            173,
+            566,
+            825,
+            678
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Comparison Results Table 2 shows the comparison results of ECE with the help of TS and HB. We can see that: (1) The standard CE loss achieves the best calibration performance on most cases. (2) The searched temperatures of models trained with the CE loss are significantly higher than those of other losses, which indicates that CE loss causes higher predictive confidences. These results demonstrate that despite the regularized models can produce better calibrated predictions, it is harder to further improve them with post-hoc calibration methods after main training. In other words, the penalty of confident predictions will compress the room of potential improvement by post-hoc methods. We also conduct experiments on CIFAR-10 and CIFAR-100 using a deeper model ResNet-110 and similar comparison results are obtained (see Appendix Table A). ",
+        "bbox": [
+            173,
+            684,
+            825,
+            809
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Coefficient Sensitivity Figure 1(a), 1(b) and 1(c) illustrate the ECEs of these three regularization methods with varied coefficient strengths. As we can see, since the complexity of the used datasets are different, the best coefficients of these methods markedly vary across the datasets, and a small change on these coefficients may cause large ECE increase. This means that we need to carefully choose the coefficient of each method when employing them on new datasets, to achieve good calibration. We can also observe that for SVHN, on which the accuracy is highest among four datasets, the regularization methods obtain lowest ECE with small coefficients. For CIFAR-100 and 20Newsgroups, on which the accuracy is relatively lower, the regularization methods need larger coefficients for better calibration. Based on this observation, we conduct another experiment for investigating the correlation between the regularization coefficients and accuracy. We learn networks on CIFAR-10 by controlling training data size, which leads to varied predictive accuracies, and choose the best coefficient for each case. Here, $\\epsilon$ , $\\alpha$ and $\\gamma$ are chosen from $\\{ 0 . 0 1 , 0 . 0 2 , . . . , 0 . 2 5 \\}$ , $\\{ 0 . 0 1 , 0 . 0 2 , . . . , 0 . 1 \\}$ and $\\{ 1 , 3 , 5 , 7 , 9 \\}$ respectively. Figure 2(d) shows that with the increase of training data size, which results in increase of predictive accuracy, the best coefficients of the regularization methods keep decreasing. ",
+        "bbox": [
+            174,
+            814,
+            825,
+            911
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/e27983661b963d061d1c8cf9f994232e27f3e9cf665f331d112e3b22208663af.jpg",
+        "table_caption": [
+            "Table 2: Comparison results (mean±std) of ECE $( \\% )$ with $M = 1 5$ over 5 random runs. The coefficients of the regularization methods on each dataset are same with those in Table 1. N/N and $\\diagup$ indicate that the average ECE of regularization methods are higher and lower than standard CE, where $\\blacktriangle$ and $\\boldsymbol { \\vee }$ are based on two-sample t-test at 0.05 significance level. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\"></td><td>Cross-Entropy</td><td>Label Smoothing 0.01/0.05/0.09/0.09</td><td>L1Norm 0.01/0.05/0.01/0.01</td><td>Focal Loss 1/3/5/5</td></tr><tr><td rowspan=\"3\">SVHN</td><td>with TS Temperature</td><td>0.72±0.26 1.82</td><td>1.35±0.11 1.11</td><td>1.22±0.08 1.12</td><td>0.80±0.22 1.10</td></tr><tr><td>with HB</td><td>0.68±0.22</td><td>0.70±0.21</td><td>0.73±0.20</td><td>0.96±0.14</td></tr><tr><td>with TS Temperature</td><td>0.95±0.19 2.51</td><td>2.54±0.11 0.96</td><td>2.71±0.36</td><td>1.39±0.28</td></tr><tr><td rowspan=\"2\">CIFAR-10</td><td>with HB</td><td>0.74±0.15</td><td>0.94±0.21</td><td>0.95 1.16±0.54</td><td>0.76 1.65±0.31</td></tr><tr><td>with TS Temperature</td><td>1.35±0.19 2.19</td><td>1.37±0.27 1.04</td><td>3.92±0.21</td><td>2.14±0.42</td></tr><tr><td rowspan=\"2\">CIFAR-100</td><td>with HB</td><td>1.27±0.27</td><td>2.01±0.22</td><td>1.24 1.56±0.44</td><td>0.97 1.83±0.30</td></tr><tr><td>with TS</td><td>3.11±0.33</td><td>5.22±0.60</td><td>2.71±0.25</td><td>3.77±0.41</td></tr><tr><td rowspan=\"2\">20 Newsgroups</td><td>Temperature</td><td>4.18</td><td>1.06</td><td>1.48</td><td>0.89</td></tr><tr><td>with HB</td><td>2.52±0.47</td><td>2.67±0.82</td><td>2.61±0.95</td><td>3.16±0.97</td></tr></table>",
+        "bbox": [
+            173,
+            150,
+            825,
+            353
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/dc28ce39124194e523cb322a5691e5dbda67d0d11c34ed73d324ccc32182fcab.jpg",
+        "image_caption": [
+            "Figure 1: (a-c): ECE $( \\% )$ with $M = 1 5$ of regularization methods with controlled regularization strength. (d): Best coefficients of regularization methods with respect to ECE with controlled training data size. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            372,
+            815,
+            468
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            537,
+            825,
+            636
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Reliability Diagram We use reliability diagram to visually represent the gap between predictive confidence and accuracy of each method. Due to the space limitation, here we only present the diagrams of CIFAR-10, and the rest figures are presented in Appendix. We can see that these visual results are similar with the comparison results of ECE reported in Table 2. Although the gap between confidence and accuracy is large when using the standard CE loss, it can be significantly diminished after using TS. However, the improvements of TS for the regularization methods are not obvious. Most importantly, no matter whether TS is used or not, the regularization methods suffer from overconfidence on samples which have high predictive uncertainty, especially on label smoothing and $L _ { 1 }$ Norm, which contradicts the traditional view. Combining with the observation in Figure 2(d), it is indicated that the regularization methods work by simply aligning the average predictive confidence of the whole dataset to the accuracy with some specific regularization strengths, and does not produce fine-grained calibration with respect to the difference of samples. ",
+        "bbox": [
+            173,
+            641,
+            825,
+            808
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Impact of Validation Size We also wonder how does the validation data size impact the post-hoc calibration. Figure 3 shows the ECE results with controlled validation data size. We can see that quite low ECE can be obtained with only a small size of validation data when using TS, which offers high efficiency for practice development. Relatively, HB needs more validation samples to obtain better calibration performance. Nevertheless, the standard CE loss stably achieves better calibration across varied validation data size with both TS and HB. ",
+        "bbox": [
+            174,
+            813,
+            825,
+            896
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/27966219feac50faf7c7caca796c0660612cfdd81581b6a556ebfa0cf0f27b7d.jpg",
+        "image_caption": [
+            "Figure 2: Reliability diagrams of each methods (after TS calibration) on CIFAR-10. The results are chosen from one of the 5 random runs of Table 1. Darker color of bars indicates that more samples are assigned with the corresponding confidence intervals. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            161,
+            97,
+            823,
+            212
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d6d05b4aeef156c24bb978f3ec868ed4e45ea0130abecc85e4cbd7d333e3ab38.jpg",
+        "image_caption": [
+            "Figure 3: ECE $( \\% )$ (after post-hoc calibration) with of regularization methods with controlled validation data size. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            158,
+            281,
+            825,
+            512
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 From Calibrated to Calibratable: A Closer Look ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            574,
+            616,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results reported in above section show the degradation of regularization methods when integrating them with post-hoc calibration methods, which indicates that though regularization helps DNNs obtain well-calibrated predictions, it makes these predictions worse calibratable. In this Section, we further investigate this phenomenon by a series of illustrative experiments. We firstly attempt to empirically understand the reason of the calibration degradation from the view of information loss. Then, we propose an inverse form of focal loss to give a closer look at the correlation between the loss functions used in training and the calibration performance. The implementation details used in this section are also the same with those in Section 3. ",
+        "bbox": [
+            173,
+            606,
+            825,
+            717
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Information Loss of Regularized Models ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            733,
+            495,
+            748
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ECE among Epochs We start by investigating the ECEs of temperature-scaled outputs over epochs during model training. To avoid the impact of the bias of validation data, we directly search the best temperature on test data in each training epoch. We denote the corresponding ECE with this searched temperature as optimal ECE, which is the lower bound of ECE with temperatures searched on validation data. Figure 4 shows the curves of optimal ECE during epochs using label smoothing with different smoothing coefficients. We can observe that the optimal ECE rises after some learning epochs: On SVHN and CIFAR-10, it starts to significantly rise around the 10th epoch, and on CIFAR-100 and 20Newsgroups, it tends to rise after 100 and 50 epochs, where the learning rate drops by a factor of 10. Another observation is that larger smoothing strength $\\epsilon$ results in worse calibration performance and more remarkable (also earlier) ECE rising. According to the memorization effect [35], DNNs usually learn easy samples at the early stage of training and tend to fit the hard ones later. ",
+        "bbox": [
+            173,
+            758,
+            825,
+            911
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/33c00a14094f2b53a1e6919bddc2dd98da1353492d763f7256f12cd39dba3625.jpg",
+        "image_caption": [
+            "Figure 4: Curves of optimal ECE $( \\% )$ during learning epochs using label smoothing with different coefficients. Dark colors show the mean results of 5 random runs and light colors show the ranges between minimal and maximum results of 5 runs. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            156,
+            104,
+            825,
+            212
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0ae6e6cf4cbd74214819296e296f6755197c03aceea1c5842458826ed114e888.jpg",
+        "image_caption": [
+            "Figure 5: Histograms of maximum logits produced from models trained with different methods. Different colors of bars represent distributions of samples with different learned epochs. The first row and second row are results of CIFAR-10 and CIFAR-100 respectively. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            178,
+            277,
+            807,
+            483
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Therefore, a simple conjecture for the ECE rising is that after some learning epochs, DNNs start to fit hard samples, at the same time the regularizer would penalize the confidences of easy samples, which makes the predictive confidences of those easy and hard samples difficult to be distinguished. Similar phenomenon is also observed when using $L _ { 1 }$ Norm (see Appendix Figure A(a-d)) except 20Newsgroups dataset, on which large norm coefficient will hurt the calibration. For focal loss (see Appendix Figure A(e-h)), the optimal ECEs trained with large regularization strengths keep high without the remarkable rising. ",
+        "bbox": [
+            173,
+            554,
+            825,
+            652
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Histogram of Logits We use histograms to visualize what the patterns of model outputs learned with different methods look like. Before that, we define learned epoch 2 of an individual training sample as the epoch, since which the sample can be correctly classified till the final learning epoch. As we mentioned above, DNNs usually learn hard samples after easy ones, hence the learned epoch of a sample can be used to indicate its corresponding hardness degree to be learned. Based on this, we want to investigate if the samples with different hardness degrees can be distinguished by model itself after training. To this end, we record the learned epochs of all training samples of CIFAR-10 and CIFAR-100 during training and statistic the distributions of their maximum logit outputs (i.e. $\\operatorname* { m a x } _ { i } g _ { i } ^ { \\theta } ( { \\pmb x } ) )$ ). As shown in Figure 5, the logits of models trained with the standard CE loss cover much larger ranges, and the regularization methods compress the distributions too tight without distinction between samples with different learned epochs, especially in label smoothing and $L _ { 1 }$ Norm. This visual observation further confirms that the regularization of DNNs works by only penalizing the confidence of the whole dataset to a low level with a specific regularization strength. This will result in loss of the important information about the hardness of samples as an undesirable side effects, and compress the room of potential improvement by post-hoc calibration. On the contrary, models trained with the standard CE loss manage to preserve this information to a certain extent during training, hence achieve better results after post-hoc calibration. ",
+        "bbox": [
+            173,
+            659,
+            826,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2faab9556d8c6d6e4dedd9e2959d0614bc01c6bb24ba522a2a3c58f25b426b59.jpg",
+        "image_caption": [
+            "Figure 6: (a): Visual representation of focal loss, CE loss and inverse focal loss. (b): Predictive accuracies $( \\% )$ of different methods. (c): ECEs $( \\% )$ with $M = 1 5$ of different methods without post-hoc calibration. (d): Searched temperatures on validation data. (e-f): ECEs $( \\% )$ with $M = 1 5$ of different methods with the help of post-hoc calibration. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            168,
+            97,
+            813,
+            313
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            393,
+            825,
+            421
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 Is Cross-Entropy the Best for Calibration? ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            438,
+            509,
+            453
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Based on our experimental findings, one natural question is that can we design some loss functions in the opposite direction of these regularization methods to further improve the calibration? For label smoothing and $L _ { p }$ Norm, we can simply set the regularization coefficients of these methods as negative values. However, we empirically found this will cause extremely low predictive accuracies even failures of training using only very small weighting coefficients. Fortunately, we can design an inverse version of focal loss without prediction degradation by mimicking the original focal loss3. Recall the form of focal loss, we see that it works by assigning larger weights to the samples with smaller confidences. This makes the optimizer pay more attention to those hard samples when updating model parameters. Actually, this weighting scheme also implicitly exists in the standard CE loss, and this can be expressed by the gradients of CE loss function w.r.t. model parameters $\\theta$ : ",
+        "bbox": [
+            173,
+            462,
+            826,
+            602
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "equation",
+        "img_path": "images/02af05181eac4fd87c3a8945b6f7f105efa5091df5f950dccc7c1756a7f495e9.jpg",
+        "text": "$$\n\\frac { \\partial \\mathcal { L } _ { c e } ( y , f ^ { \\theta } ( { \\pmb x } ) ) } { \\partial \\theta } = - \\frac { 1 } { f _ { y } ^ { \\theta } ( { \\pmb x } ) } \\nabla _ { \\theta } f _ { y } ^ { \\theta } ( { \\pmb x } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            372,
+            607,
+            625,
+            645
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "where the factor term $\\frac { 1 } { f _ { y } ^ { \\theta } ( { \\pmb x } ) }$ indicates that samples with smaller confidences are weighted larger in gradient calculation. Opposite to the principle of focal loss, we propose inverse focal loss as follows: ",
+        "bbox": [
+            173,
+            651,
+            825,
+            685
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "equation",
+        "img_path": "images/0200a73870b96ea968c9568c8e59073233c1487db65118647ab9a816c940b0fc.jpg",
+        "text": "$$\n\\mathcal { L } _ { \\bar { f } } = - ( 1 + f _ { y } ^ { \\theta } ) ^ { \\bar { \\gamma } } \\log f _ { y } ^ { \\theta }\n$$",
+        "text_format": "latex",
+        "bbox": [
+            415,
+            690,
+            583,
+            712
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "By a simple modification on the weighting term of original focal loss, the inverse focal loss assigns larger weights to the samples with larger output confidences. Similar with original focal loss, the choice of coefficient $\\bar { \\gamma }$ has a huge impact on the property of this loss. In Figure 6(a), we plot the curves of inverse focal loss with varied $\\bar { \\gamma }$ and also plot the standard CE loss and focal loss for comparison. We can see that different from the original focal loss, the curves of inverse focal loss are steeper when confidence is large, and larger $\\bar { \\gamma }$ gives steeper curves. ",
+        "bbox": [
+            173,
+            715,
+            825,
+            799
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conduct another experiment to evaluate the inverse focal loss, and also investigate what will happen when we increase its coefficient $\\bar { \\gamma }$ . Figure 6(c) shows the ECE results without post-hoc calibration. The ECEs of inverse focal loss are larger than CE and focal loss in most cases. This is consistent to our expectation since inverse focal loss aggravates the overconfidence issue of DNNs by weighting larger on the easy samples. Figure 6(e) and 6(f) show the ECE results with the help of post-hoc calibration. When using HB, the ECEs of inverse focal loss are worse than CE on SVHN and CIFAR-10, while better than CE on CIFAR-100. Generally speaking, there is no clear trend when we increase $\\bar { \\gamma }$ . More interesting results appear when using TS: (1) On CIFAR-10 and CIFAR-100, the ECE results of inverse focal loss are better than that of the standard CE loss; and (2) there is a descend-then-ascend trend from focal loss with $\\gamma = 3$ to inverse focal loss with $\\bar { \\gamma } = 3$ . From these observations, we may say that the best loss function for calibration is varied across different tasks according to the characteristics of datasets. On SVHN, which is a relatively easy dataset, standard CE loss yields pretty good results; on CIFAR-10 and CIFAR-100, which is more complex and difficult, the best results are obtained using inverse focal loss; on 20Newsgroups, which has fewest training samples among four datasets, the best result is obtained when using focal loss with $\\gamma = 1$ . The searched temperatures when using TS are presented in Figure 6(d). The increasing of best temperatures indicates that the overconfidence problem is severer when using inverse focal loss with larger $\\bar { \\gamma }$ . The predictive accuracies are presented in Figure 6(a). As is shown that inverse focal loss yields highly competitive results compared with the standard CE loss on SVHN, CIFAR-10 and CIFAR-100. On 20Newsgroups, when using large $\\bar { \\gamma }$ , the predictive performance of inverse focal loss is worse than the CE loss. ",
+        "bbox": [
+            173,
+            805,
+            825,
+            876
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            90,
+            825,
+            313
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            333,
+            321,
+            351
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In machine learning, calibration has long been studied [25, 33, 34, 22], and many classical methods, like Platt Scaling [25] and Histogram Binning [33], have been proposed in the literature. In recent years, deep neural networks trained with commonly used CE loss, have been empirically found to predict poorly calibrated probabilities. The early researches for this problem focus on bayesian models [2, 16, 3, 1], which indirectly infer prediction uncertainty through weight uncertainties. But training and inferring the bayesian DNNs are computationally more expensive and conceptually more complicated than deterministic DNNs. Therefore, the uncertainty qualification of the nonbayesian models has always been an important topic, which also attracts a lot of researchers from the perspective of understanding DNNs. ",
+        "bbox": [
+            174,
+            366,
+            825,
+            491
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Guo et al. [4] systematically investigated the miscalibration problem of the deterministic DNNs and empirically compared several conventional post-hoc calibration fixes. Two key findings are suggested in their paper: (1) Increasing model capacity and regularization strength negatively affect the calibration. (2) Simple post-hoc methods like TS [25] and HB [33] can reduce the calibration error to a quite low level. Following their work, there is a surge of research that proposed new post-hoc calibration methods [10, 27, 5, 11, 24, 36, 26]. Different from post-hoc calibration methods, another line of research aims to learn calibrated networks during training by modifying the training process [29, 12, 8]. Thulasidasan et al. [29] found that DNNs trained with mixup are significantly better calibrated than DNNs trained in the regular fashion. Kumar et al. [12] proposed a RKHS kernel based measure of calibration that is efficiently trainable alongside the standard CE loss, which can minimize an explicit calibration error during training. Krishnan and Tichoo [8] introduced a differentiable accuracy versus uncertainty calibration loss function that allows a model to learn to provide well-calibrated uncertainties, in addition to improved accuracy. Recently, inspired by the findings in [4], several studies were proposed to leverage the regularization of DNNs to improve calibration performance during training [19, 18, 7]. As we described in Section 3, these implicit or explicit regularization techniques can improve calibration by penalizing the predictive confidences of DNNs. It is worth nothing that besides the studies on improving calibration performance, there are also several studies that focus on the measure of calibration performance [23, 31, 32, 5]. ",
+        "bbox": [
+            174,
+            497,
+            825,
+            746
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            767,
+            297,
+            785
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we investigate the uncertainty calibration problem of DNNs by a series of experiments. The empirical study shows that despite the predictions obtained from the regularized models are better calibrated, worse results would be obtained if we employ post-hoc calibration methods on these regularized models. Extended experiments demonstrate that the regularized DNNs will lose the important information about the hardness of samples, which results in the harm of post-hoc calibration. Based on the experimental observations, we design a new loss function in the opposite direction of previous regularization methods, and empirically show the superiority of this loss in calibration with the help of post-hoc methods, even though it causes severer overconfidence issue in the main training phase. Our findings suggest that overconfidence of DNNs is not the nightmare in model calibration and point out a new direction to improve the calibration performance of DNNs by considering main training and post-hoc calibration as a unified framework. Moreover, the study of the phenomena of deep learning uncertainty under distribution shift is very interesting as one of the future work, since the behaviour with distribution shifts might be most important in practice. ",
+        "bbox": [
+            174,
+            800,
+            825,
+            911
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            90,
+            825,
+            161
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "8 Acknowledgments ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            181,
+            357,
+            199
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The authors wish to thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Science Foundation of China (62176055), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX21_0151) and the China University S&T Innovation Plan Guided by the Ministry of Education. We thank the Big Data Center of Southeast University for providing the facility support on the numerical calculations in this paper. ",
+        "bbox": [
+            174,
+            213,
+            826,
+            284
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            303,
+            266,
+            319
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "[1] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International Conference on Machine Learning, pages 1613–1622, 2015.   \n[2] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, pages 1050–1059, 2016.   \n[3] Alex Graves. Practical variational inference for neural networks. In Advances in Neural Information Processing Systems, pages 2348–2356, 2011.   \n[4] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330, 2017.   \n[5] Kartik Gupta, Amir Rahimi, Thalaiyasingam Ajanthan, Thomas Mensink, Cristian Sminchisescu, and Richard Hartley. Calibration of neural networks using splines. In International Conference on Representation Learning, 2021.   \n[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.   \n[7] Taejong Joo and Uijung Chung. Revisiting explicit regularization in neural networks for well-calibrated predictive uncertainty. arXiv:2006.06399, 2021.   \n[8] Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. In Advances in Neural Information Processing Systems, pages 18237– 18248, 2020.   \n[9] Alex Krizhevsky. Learning multiple layers of features from tiny images. In Tech Report, 2009.   \n[10] Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. In Advances in Neural Information Processing Systems, pages 12295–12305, 2019.   \n[11] Ananya Kumar, Percy S Liang, and Tengyu Ma. Verified uncertainty calibration. In Advances in Neural Information Processing Systems, pages 3787–3798, 2019.   \n[12] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805–2814, 2018.   \n[13] Ken Lang. Newsweeder: Learning to filter netnews. In International Conference on Machine Learning, pages 331–339, 1995.   \n[14] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In IEEE International Conference on Computer Vision, pages 2980–2988, 2017.   \n[15] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In International Conference on Neural Information Processing Systems, pages 1647–1657, 2018.   \n[16] Wesley J Maddox, Pavel Izmailov, Timur Garipov, Dmitry P Vetrov, and Andrew Gordon Wilson. A simple baseline for bayesian uncertainty in deep learning. In Advances in Neural Information Processing Systems, pages 13153–13164, 2019.   \n[17] Clara Meister, Elizabeth Salesky, and Ryan Cotterell. Generalized entropy regularization or: There’s nothing special about label smoothing. In Annual Meeting of the Association for Computational Linguistics, pages 6870–6886, 2020.   \n[18] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. In Advances in Neural Information Processing Systems, pages 15288–15299, 2020.   \n[19] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In Advances in Neural Information Processing Systems, pages 4696–4705, 2019.   \n[20] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI Conference on Artificial Intelligence, pages 2901–2907, 2015.   \n[21] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In Advances in Neural Information Processing Systems Workshops, 2011.   \n[22] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In International Conference on Machine learning, pages 625–632, 2005.   \n[23] Jeremy Nixon, Michael W. Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 38–41, 2019.   \n[24] Kanil Patel, William Beluch, Bin Yang, Michael Pfeiffer, and Dan Zhang. Multi-class uncertainty calibration via mutual information maximization-based binning. In International Conference on Representation Learning, 2021.   \n[25] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers, 10(3):61–74, 1999.   \n[26] Amir Rahimi, Kartik Gupta, Thalaiyasingam Ajanthan, Thomas Mensink, Cristian Sminchisescu, and Richard Hartley. Post-hoc calibration of neural networks. arXiv:2006.12807, 2020.   \n[27] Amir Rahimi, Amirreza Shaban, Ching-An Cheng, Richard Hartley, and Byron Boots. Intra order-preserving functions for calibration of multi-class neural networks. In Advances in Neural Information Processing Systems, 2020.   \n[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2818–2826, 2016.   \n[29] Sunil Thulasidasan, Gopinath Chennupati, Jeff A. Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In Advances in Neural Information Processing Systems, pages 13888–13899, 2019.   \n[30] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. An empirical study of example forgetting during deep neural network learning. In International Conference on Representation Learning, 2019.   \n[31] Juozas Vaicenavicius, David Widmann, Carl R. Andersson, Fredrik Lindsten, Jacob Roll, and Thomas B. Schön. Evaluating model calibration in classification. In International Conference on Artificial Intelligence and Statistics, pages 3459–3467, 2019.   \n[32] David Widmann, Fredrik Lindsten, and Dave Zachariah. Calibration tests in multi-class classification: A unifying framework. In Advances in Neural Information Processing Systems, pages 12236–12246, 2019.   \n[33] Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In International Conference on Machine Learning, pages 609–616, 2001.   \n[34] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 694–699, 2002.   \n[35] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In International Conference on Representation Learning, 2017.   \n[36] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International Conference on Machine Learning, pages 11117–11128, 2020.   \n[37] Zhi-Hua Zhou. Why over-parameterization of deep neural networks does not overfit? Science China Information Sciences, 64(1):1–3, 2021. ",
+        "bbox": [
+            174,
+            319,
+            826,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            59,
+            826,
+            916
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            89,
+            826,
+            430
+        ],
+        "page_idx": 11
+    }
+]

parse/train/SJzwb2RcK7/SJzwb2RcK7.md

@@ -0,0 +1,323 @@
+# ADVERSARIAL DECOMPOSITION OF TEXT REPRESENTATION
+
+Anonymous authors Paper under double-blind review
+
+# ABSTRACT
+
+In this paper, we present a method for adversarial decomposition of text representation. This method can be used to decompose a representation of an input sentence into several independent vectors, where each vector is responsible for a specific aspect of the input sentence. We evaluate the proposed method on two case studies: the conversion between different social registers and diachronic language change. We show that the proposed method is capable of fine-grained controlled change of these aspects of the input sentence. For example, our model is capable of learning a continuous (rather than categorical) representation of the style of the sentence, in line with the reality of language use. The model uses adversarial-motivational training and includes a special motivational loss, which acts opposite to the discriminator and encourages a better decomposition. Finally, we evaluate the obtained meaning embeddings on a downstream task of paraphrase detection and show that they are significantly better than embeddings of a regular autoencoder.
+
+# 1 INTRODUCTION
+
+Despite the recent successes in using neural models for representation learning for natural language text, learning a meaningful representation of input sentences remains an open research problem. A variety of approaches, from sequence-to-sequence models that followed the work of Sutskever et al. (2014) to the more recent proposals (Arora et al., 2017; Nangia et al., 2017; Conneau et al., 2017; Logeswaran & Lee, 2018; Subramanian et al., 2018; Cer et al., 2018) share one common drawback. Namely, all of them encode the input sentence into just one single vector of a fixed size. One way to bypass the limitations of a single vector representation is to use an attention mechanism (Bahdanau et al., 2014; Vaswani et al., 2017). We propose to approach this problem differently and design a method for adversarial decomposition of the learned input representation into multiple components. Our method encodes the input sentence into several vectors, where each vector is responsible for a specific aspect of the sentence.
+
+In terms of learning different separable components of input representation, our work most closely relates to the style transfer work, which has been applied to a variety of different aspects of language, from diachronic language differences (Xu et al., 2012) to authors’ personalities (Lipton et al., 2015) and even sentiment (Hu et al., 2017; Fu et al., 2017). The style transfer work effectively relies on the more classical distinction between meaning and form (de Saussure, 1959), which accounts for the fact that multiple surface realizations are possible for the same meaning. For simplicity, we will use this terminology throughout the rest of the paper.
+
+Consider the case when we encode an input sentence into a meaning vector and a form vector. We are then able to perform a controllable change of meaning or form by a simple change applied to these vectors. For example, we can encode two sentences written in two different styles, then swap the form vectors while leaving the meaning vectors intact. We can then generate new unique sentences with the original meaning, but written in a different style.
+
+In the present work, we propose a novel model for this type of decomposition based on adversarialmotivational training and design an architecture inspired by the GANs (Goodfellow et al., 2014) and adversarial autoencoders (Makhzani et al., 2015). In addition to the adversarial loss, we use a special motivator (Albanie et al., 2017), which, in contrast to the discriminator, is used to provide a motivational loss to encourage the model to better decomposition of the meaning and the form, as well as specific aspects of meaning. We make all the code publicly available on GitHub 1.
+
+We evaluate the proposed methods for learning separate aspects of input representation on the following case studies:
+
+1. Learning to separate out a representation of the specific diachronic slice of language. One may express the same meaning using the Early Modern English (e.g. What would she have?) and the contemporary English ( What does she want?)   
+2. Learning a representation for a social register (Halliday et al., 1968) – that is, subsets of language appropriate in a given context or characteristic of a certain group of speakers. These include formal and informal language, the language used in different genres (e.g., fiction vs. newspapers vs. academic texts), different dialects, and even literary idiostyles. We experiment with the registers corresponding to the titles of scientific papers vs. newspaper articles.
+
+# 2 RELATED WORK
+
+As mentioned above, the most relevant previous work comes from the style transfer research, and it can be divided into two groups:
+
+1. Approaches that aim to generate text in a given form. For example, the task may be to produce just any verse as long as it is in the “style” of the target poet. 2. Approaches that aim to induce a change in either the “form” or the “meaning” of an existing utterance. For example, “Good bye, Mr. Anderson.” can be transformed to “Fare you well, good Master Anderson” (Xu et al., 2012)).
+
+An example of the first group is the work by Potash et al. (2015), who trained several separate networks on verses by different hip-hip artists. An LSTM network successfully generated verses that were stylistically similar to the verses of the target artist (as measured by cosine distance on TfIdf vectors). More complicated approaches use language models that are conditioned in some way. For example, Lipton et al. (2015) produced product reviews with a target rating by passing the rating as an additional input at each timestep of an LSTM model. Tang et al. (2016) generated reviews not only with a given rating but also for a specific product. At each timestep a special context vector was provided as input, gated so as to enable the model to decide how much attention to pay to that vector and the current hidden state. Li et al. (2016) used “speaker” vectors as an additional input to a conversational model, improving consistency of dialog responses. Finally, Ficler & Goldberg (2017) performed an extensive evaluation of conditioned language models based on “content” (theme and sentiment) and “style” (professional, personal, length, descriptiveness). Importantly, they showed that it is possible to control both “content” and “style” simultaneously.
+
+Work from the second group can further be divided into two clusters by the nature of the training data: parallel aligned corpora, or non-aligned datasets. The aligned corpora enable approaching the problem of form shift as a paraphrasing or machine translation problem. Xu et al. (2012) used statistical and dictionary-based systems on a dataset of original plays by Shakespeare and their contemporary translations. Carlson et al. (2017) trained an LSTM network on 33 versions of the Bible. Jhamtani et al. (2017) used a Pointer Network (Vinyals et al., 2015), an architecture that was successfully applied to a wide variety of tasks (Merity et al., 2016; Gulcehre et al., 2016; Potash et al., 2017), to enable direct copying of the input tokens to the output. Note that these works use BLEU (Papineni et al., 2002) as the main, or even the only evaluation measure. This is only possible in cases where a parallel corpus is available.
+
+Recently, new approaches that do not require a parallel corpora were developed in both CV (Zhu et al., 2017) and NLP. Hu et al. (2017) succeeded in changing tense and sentiment of sentences with a two steps procedure based on a variational auto-encoder (VAE) (Kingma & Welling, 2013). After training a VAE, a discriminator and a generator are trained in an alternate manner, where the discriminator tries to correctly classify the target sentence attributes. A special loss component forces the hidden representation of the encoded sentence to not have any information about the target sentence attributes. Mueller et al. (2017) used a VAE to produce a hidden representation of a sentence, and then modify it to match the desired form. Unlike Hu et al. (2017), they do not separate the form and meaning embeddings. Shen et al. (2017) applied a GAN to align the hidden representation of sentences from two corpora and force them to do not have any information about the form via adversarial loss. During the decoding, similarly the work by Lipton et al. (2015), special “style” vectors are passed to the decoder at every timestep to produce a sentence with the desired properties. The model is trained using the Professor-Forcing algorithm (Lamb et al., 2016). Kim et al. (2017) worked directly on hidden space vectors that are constrained with the same adversarial loss instead of outputs of the generator, and use two different generators for two different “styles”. Finally, Fu et al. (2017) proposed two models for generating sentences with the target properties using an adversarial loss, similarly to Shen et al. (2017) and Kim et al. (2017).
+
+Comparison with previous work In contrast to the proposals of $\mathrm { X u }$ et al. (2012), Carlson et al. (2017), Jhamtani et al. (2017), our solution does not require a parallel corpus. Furthermore, unlike the model by Shen et al. (2017), our model works directly on representation of sentences in the hidden space.
+
+Most importantly, in contrast to the proposals by Mueller et al. (2017), Hu et al. (2017), Kim et al. (2017), Fu et al. (2017), our model produces a representation for both meaning and form and does not treat the form as a categorical (in the vast majority of works, binary) variable. Although the form was represented as dense vectors in previous work, it is still just a binary feature, as they use a single pre-defined vector for each form, with all sentences of the same form assigned the same form vector. In contrast, our work treats form as a truly continuous variable, where each sentence has its own, unique, form vector.
+
+Treating meaning and form not as binary/categorical, but as continuous is more consistent with the reality of language use, since there are different degrees of overlap between the language used by different registers or in different diachronic slices. Indeed, language change is gradual, and the acceptability of expressions in a given register also forms a continuum, so one expects a substantial overlap between the grammar and vocabulary used, for example, on Twitter and by New York Times. To the best of our knowledge, this is the first model that considers linguistic form in the task of text generation as a continuous variable.
+
+One significant consequence of learning a continuous representation for form is that it allows the model to work with a large, and potentially infinite, number of forms. Note that in this case the locations of areas of specific forms in the vector style space would reflect the similarity between these forms. For example, the proposed model could be directly applied to the authorship attribution problem. In this case, each author would have their own area in the form space, and the more similar the authors are in terms of writing style, the closer these areas would be to each other. We performed preliminary experiments on this and report the results in Appendix A.
+
+# 3 FORMULATION
+
+Let us formulate the problem of decomposition of text representation on an example of controlled change of linguistic form and conversion of Shakespeare plays in the original Early Modern to contemporary English. Let $X ^ { a }$ be a corpus of texts $\pmb { x } _ { i } ^ { a } \in \mathcal { X } ^ { a }$ in Early Middle English $\mathbf { f } ^ { a } \in \mathcal { F }$ , and $X ^ { b }$ be a corpus of texts $\pmb { x } _ { i } ^ { b } \in \mathcal { X } ^ { b }$ in modern English $\mathbf { f } ^ { b } \in \mathcal { F }$ . We assume that the texts in both $X ^ { a }$ and $X ^ { b }$ has the same distribution of meaning $\mathbf { m } \in \mathcal { M }$ . The form f , however, is different and generated from a mixture of two distributions:
+
+$$
+\mathbf { f } _ { i } = \alpha _ { i } ^ { a } p ( \mathbf { f } ^ { a } ) + \alpha _ { i } ^ { b } p ( \mathbf { f } ^ { b } )
+$$
+
+where $\mathbf { f } ^ { a }$ and $\mathbf { f } ^ { b }$ are two different languages (Early Modern and contemporary English). Intuitively, we say that a sample $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { i } }$ has the form $\mathbf { f } ^ { a }$ if $\alpha _ { i } ^ { a } > \alpha _ { i } ^ { b }$ , and it has the form $\mathbf { f } ^ { b }$ if $\alpha _ { i } ^ { b } > \alpha _ { i } ^ { a }$ .
+
+The goal of dissociation meaning and form is to learn two encoders $E _ { \mathbf { m } } : \mathcal { X }  \mathcal { M }$ and $E _ { \mathbf { f } } : \mathcal { X }  \mathcal { F }$ for the meaning and form correspondingly, and the generator $G : { \mathcal { M } } , { \mathcal { F } } \to { \mathcal { X } }$ such that
+
+$$
+\forall j \in \{ a , b \} , \forall k \in \{ a , b \} : G ( E _ { \mathbf { m } } ( x ^ { k } ) , E _ { \mathbf { f } } ( x ^ { j } ) )  \mathcal { X } ^ { j }
+$$
+
+That is, the form of a generated sample depends exclusively on the provided $\mathbf { f } _ { j }$ and can be the in the same domain for two different ${ \bf m } _ { u }$ and $\mathbf { m } _ { v }$ from two samples from different domains $\mathcal { X } ^ { a }$ and $\mathcal { X } ^ { b }$ .
+
+Note that, in contrast to the previously proposals, the form f is not a categorical variable but a continuous vector. This enables fine-grained controllable change of form: the original form $\mathbf { f } _ { i }$ is changed to reflect the form of the specific target sentence $\mathbf { f } _ { j }$ with its own unique $\alpha ^ { a }$ and $\alpha ^ { b }$ while preserving the original meaning $\mathbf { m } _ { i }$ .
+
+An important caveat concerns the core assumption of the similar meaning distribution in the two corpora, which is also made in all other works reviewed in Section 2. It limits the possible use of this approach to cases where the distributions are in fact similar (i.e. parallel or at least comparable corpora are available). It does not apply to many cases that could be analyzed in terms of meaning and form. For example, books for children and scholarly papers are both registers, they have their own form (i.e. specific subsets of linguistic means and structure conventions) – but there is little overlap in the content. This would make it hard even for a professional writer to turn a research paper into a fairy tale.
+
+# 4 METHOD DESCRIPTION
+
+Inspired by Makhzani et al. (2015), Kim et al. (2017), and Albanie et al. (2017), we propose ADNet, a new model for adversarial decomposition of text representation (Figure 1).
+
+![](images/3abb02e83857746da9d726ec7398bd1454ecd5ed2a8931c9876ed697853ada2e.jpg)  
+Figure 1: Overview of ADNet. Encoder encodes the inputs sentences into two latent vectors m and f. The Generator takes them as the input and produces the output sentence. During the training, the Discriminator is used for an adversarial loss that forces m to do not carry any information about the form, and the M otivator is used for a motivational loss that encourages f to carry the needed information about the form.
+
+Our solution is based on a widely used sequence-to-sequence framework (Sutskever et al., 2014) and consists of four main parts. The encoder $E$ encodes the inputs sequence $_ { \textbf { \em x } }$ into two latent vectors $\mathbf { m }$ and f which capture the meaning and the form of the sentence correspondingly. The generator $G$ then takes these two vectors as the input and produces a reconstruction of the original input sequence $\hat { \pmb x }$ .
+
+The encoder and generator by themselves will likely not achieve the dissociation of the meaning and form. We encourage this behavior in a way similar to Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), which had an overwhelming success the past few years and have been proven to be a good way of enforcing a specific distribution and characteristics on the output of a model.
+
+Inspired by the work of Albanie et al. (2017) and the principle of ”carrot and stick” (Safire, 1995), in contrast to the majority of work that promotes pure adversarial approach (Goodfel
+
+low et al., 2014; Shen et al., 2017; Fu et al., 2017; Zhu et al., 2017), we propose two additional components, the discriminator $D$ and the motivator $M$ to force and motivate the model to learn the dissociation of the meaning and the form. Similarly to a regular GAN model, the adversarial discriminator $D$ tries to classify the form f based on the latent meaning vector $\mathbf { m }$ , and the encoder $E$ is penalized to make this task as hard as possible.
+
+Opposed to such vicious behaviour, the motivator $M$ tries to classify the form based on the latent form vector f, as it should be done, and encourages the encoder $E$ to make this task as simple as possible. We could apply the adversarial approach here as well and force the distribution of the form vectors to fit a mixture of Gaussians (in this particular case, a mixture of two Guassians) with another discriminator, as it is done by Makhzani et al. (2015), but we opted for the “dualistic” path of two complimentary forces.
+
+# 4.1 ENCODER-DECODER
+
+Both the encoder $E$ and the generator $G$ are modeled with a neural network. Gated Recurrent Unit (GRU) (Chung et al., 2014) is used for $E$ to encode the input sentence $_ { \textbf { \em x } }$ into a hidden vector $\pmb { h } = \mathbf { G } \mathbf { R } \mathbf { U } ( \pmb { x } )$ .
+
+The vector $^ { h }$ is then passed through two different fully connected layers to produce the latent vectors of the form and the meaning of the input sentence:
+
+$$
+{ \bf m } = \operatorname { t a n h } ( W _ { m } h + b _ { m } ) { \bf f } = \operatorname { t a n h } ( W _ { f } h + b _ { f } )
+$$
+
+We use $\theta _ { E }$ to denote the parameters of the encoder $E \colon W _ { m } , b _ { m } , W _ { f } , b _ { f }$ , and the parameters of the GRU unit.
+
+The generator $G$ is also modelled with a GRU unit. The generator takes as input the meaning vector $\mathbf { m }$ and the form vector f, concatenates them, and passes trough a fully-connected layer to obtain a hidden vector $_ z$ that represents both meaning and form of the original input sentence:
+
+$$
+z = \operatorname { t a n h } ( W _ { z } [ \mathbf { m } ; \mathbf { f } ] + b _ { m } )
+$$
+
+After that, we use a GRU unit to generate the output sentence as a probability distribution over the vocabulary tokens:
+
+$$
+p ( \hat { \pmb x } ) = \prod _ { t = 1 } ^ { T } p ( \hat { \pmb x } _ { t } | \boldsymbol z , \hat { \pmb x } _ { 1 } , \dots , \hat { \pmb x } _ { t - 1 } )
+$$
+
+We use $\theta _ { G }$ to denote the parameters of the generator $G$ : $W _ { z }$ , $b _ { m }$ , and the parameters of the used GRU. The encoder and generator are trained using the standard reconstruction loss:
+
+$$
+\begin{array} { r } { \mathcal { L } _ { \mathrm { r e c } } ( \pmb { \theta } _ { E } , \pmb { \theta } _ { G } ) = \mathbb { E } _ { \pmb { x } \sim \pmb { X } ^ { a } } [ - \log p ( \hat { \pmb x } | \pmb { x } ) ] + \mathbb { E } _ { \pmb { x } \sim \pmb { X } ^ { b } } [ - \log p ( \hat { \pmb x } | \pmb { x } ) ] } \end{array}
+$$
+
+# 4.2 DISCRIMINATOR
+
+The representation of the meaning m produced by the encoder $E$ should not contain any information about the form f . We achieve this by using an adversarial approach. First, we train a discriminator $D$ , consisting of several fully connected layers with ELU activation function (Clevert et al., 2015) between them, to predict the form f of a sentence by its meaning vector: $\hat { f } _ { D } = D ( { \bf m } )$ , where $\hat { f }$ is the score (logit) reflecting the probability of the sentence $_ { \textbf { \em x } }$ to belong to one of the form domains.
+
+Motivated by the Wasserstein GAN (Arjovsky et al., 2017), we use the following loss function instead of the standard cross-entropy:
+
+$$
+\mathcal { L } _ { D } ( \pmb { \theta } _ { D } ) = \mathbb { E } _ { \pmb { x } \sim \pmb { X } ^ { a } } [ D ( E _ { \mathbf { m } } ( \pmb { x } ) ) ] - \mathbb { E } _ { \pmb { x } \sim \pmb { X } ^ { b } } [ D ( E _ { \mathbf { m } } ( \pmb { x } ) ) ]
+$$
+
+Thus, a successful discriminator will produce negative scores $\hat { f }$ for sentences from $X ^ { a }$ and positive scores for sentences from $X ^ { b }$ . This discriminator is then used in an adversarial manner to provide a learning signal for the encoder and force dissociation of the meaning and form by maximizing $\mathcal { L } _ { D }$ $: \mathcal { L } _ { \mathrm { a d v } } ( \pmb { \theta } _ { E } ) = - \lambda _ { \mathrm { a d v } } \mathcal { L } _ { D }$ , where $\lambda _ { \mathrm { a d v } }$ is a hyperparameter reflecting the strength of the adversarial loss. Note that this loss applies to the parameters of the encoder.
+
+# 4.3 MOTIVATOR
+
+Our experiments showed that it is enough to have just the discriminator $D$ and the adversarial loss $\mathcal { L } _ { \mathrm { a d v } }$ to force the model to dissociate the form and the meaning. However, in order to achieve a better dissociation, we propose to use a motivator $M$ (Albanie et al., 2017) and the corresponding motivational loss. Conceptually, this is the opposite of the adversarial loss, hence the name. As the discriminator $D$ , the motivator $M$ learns to classify the form f of the input sentence. However, its input is not not the meaning vector but the form vector: ${ \hat { f } } _ { M } = M ( \mathbf { f } )$ .
+
+The motivator has the same architecture as the discriminator, and the same loss function. While the adversarial loss forces the encoder $E$ to produce a meaning vector m with no information about the form f , the motivational loss encourages $E$ to encode this information in the form vector by minimizing ${ \mathcal { L } } _ { M }$ ${ \bf \chi } _ { \ / t } \colon \mathcal { L } _ { \mathrm { m o t i v } } ( \pmb { \theta } _ { E } ) = \lambda _ { \mathrm { m o t i v } } \mathcal { L } _ { \ / M }$ .
+
+# 4.4 TRAINING PROCEDURE
+
+The overall training procedure follows the methods for training GANs (Goodfellow et al., 2014; Arjovsky et al., 2017) and consists of two stages: training the discriminator $D$ and the motivator $M$ , and training the encoder $E$ and the generator $G$ .
+
+In contrast to Arjovsky et al. (2017), we do not train the $D$ and $M$ more than the $E$ and the $G$ . In our experiments we found that simple training in two stages is enough to achieve dissociation of the meaning and the form. Encoder and generator are trained with the following loss function that combines reconstruction loss with the losses from the discriminator and the motivator:
+
+$$
+\mathcal { L } _ { \mathrm { t o t a l } } ( \theta _ { E } , \theta _ { G } ) = \mathcal { L } _ { \mathrm { r e c } } + \mathcal { L } _ { \mathrm { a d v } } + \mathcal { L } _ { \mathrm { m o t i v } }
+$$
+
+# 5 EXPERIMENTAL SETUP
+
+# 5.1 EVALUATION
+
+Similarly to the evaluation of style transfer in CV (Isola et al., 2017), evaluation of this task is difficult. We follow the approach of Isola et al. (2017); Shen et al. (2017) and recently proposed by Fu et al. (2017) methods of evaluation of “transfer strength” and “content preservation”. The authors showed the proposed automatic metrics to a large degree correlate with human judgment and can serve as a proxy. Below we give an overview of these metrics.
+
+Transfer Strength. The goal of this metric is to capture whether the form has been changed successfully. To do that, a classifier $C$ is trained on the two corpora, $X ^ { a }$ and $X ^ { b }$ to recognize the linguistic “form” typical of each of them. After that, a sentence the form/meaning of which was changed is passed to the classifier. The overall accuracy reflects the degree of success of changing the form/meaning. This approach is widely used in CV (Isola et al., 2017), and was applied in NLP as well (Shen et al., 2017).
+
+In our experiments we used a GRU unit followed by four fully-connected layers with ELU activation functions between them as the classifier.
+
+Content preservation Note that transfer strength by itself does not capture the overall quality of a changed sentence. A extremely overfitted model that produces the same, the most characteristic sentence of one corpus all the time would have a high score according to this metric. Thus, we need to measure how much of the meaning was preserved while changing the form. To do that, Fu et al. (2017) proposed to use a cosine similarity based metric using pretrained word embeddings. First, a sentence embedding is computed by concatenation of max, mean, and average pooling over the timesteps:
+
+$$
+\pmb { v } = [ \operatorname* { m a x } ( \pmb { v } _ { 1 } , \dots , \pmb { v } _ { T } ) ; \operatorname* { m i n } ( \pmb { v } _ { 1 } , \dots , \pmb { v } _ { T } ) ; \operatorname* { m e a n } ( \pmb { v } _ { 1 } , \dots , \pmb { v } _ { T } ) ]
+$$
+
+Next, the cosine similarity score $s _ { i }$ between the embedding $\pmb { v } _ { i } ^ { s }$ of the original source sentence and the target sentence with the changed form $\mathbf { \Delta } \mathbf { \vec { v } } _ { i } ^ { t }$ is computed, and the scores across the dataset are averaged to obtain the total score:
+
+$$
+\mathbf { s } = \frac { 1 } { 2 } \left[ \frac { 1 } { | X ^ { a } | } \sum _ { i = 1 } ^ { | X ^ { a } | } \mathrm { s } _ { i } + \frac { 1 } { | X ^ { b } | } \sum _ { i = 1 } ^ { | X ^ { b } | } \mathrm { s } _ { i } \right]
+$$
+
+# 5.1.1 CONTINUOUS FORM
+
+The metrics described above treat the form as a categorical (in most cases, even binary) variable. This was not a problem in previous work since the change of form could be done by just inverting the form vector. Our work, in contrast, treats the form as a continuous variable, and, therefore, we cannot just use the proposed metrics directly. To enable a fair comparison, we propose the following procedure.
+
+For each sentence $s _ { s } ^ { a }$ in the test set from the corpus $X ^ { a }$ we sample $k = 1 0$ random sentence from the corpus $X ^ { b }$ of the opposite form. After that, we encode them into the meaning $m _ { i }$ and form $f _ { i }$ vectors, and average the form vectors to obtain a single form vector $\begin{array} { r } { { f } _ { \mathrm { a v g } } = \frac { 1 } { k } \sum _ { i = 1 } ^ { k } { f } _ { i } } \end{array}$ . We then generate a new sentence with its original meaning vector $m _ { s }$ and the resulting form vector $\pmb { f } _ { \mathrm { a v g } }$ , and use it for evalation. This process enables a fair comparison with the previous works that treat form as a binary variable.
+
+# 5.2 DATASETS
+
+We performed an extensive evaluation of the proposed method on several dataset that reflect different changes of meaning, form, or specific aspects of meaning, such as sentiment polarity.
+
+Changing form: register This experiment is conducted with a dataset of titles of scientific papers and news articles published by Fu et al. (2017). This dataset (referred to as “Headlines”) contains titles of scientific articles crawled from online digital libraries, such as “ACM Digital Library” and “arXiv”. The titles of the news articles are taken from the “News Aggregator Data Set” from UCI Machine Learning Repository (Dheeru & Karra Taniskidou, 2017)
+
+Changing form: language diachrony Diachronic language change is explored with the dataset composed by Xu et al. (2012). It includes the texts of 17 plays by William Shakespeare in the original Early Modern English, and their translations into contemporary English. We randomly permuted all sentences from all plays and sampled the training, validation, and test sets. Note that this is the smallest dataset in our experiments.
+
+Previous work on style transfer for text also included the experiments with changing sentiment polarity (Shen et al., 2017; Fu et al., 2017). We do not report the experiments with sentiment data, since the change in sentiment polarity corresponds to a change in a specific aspect of meaning, rather than form. We therefore believe the comparison with these data would not be instructive.
+
+# 6 RESULTS AND DISCUSSION
+
+![](images/345122e7aea7fa589082eadec03c850ada45c1d27c374c52dc7d77b4925102de.jpg)  
+Figure 2: Transfer strength vs Content preservation (see subsection 5.1) for different combination of the size of the meaning and form vectors. Each point is labeled with ”¡meaning vector size¿, ¡form vector size¿”.
+
+Probably, the most recent and similar to our work is the model proposed by Fu et al. (2017), in particular the “style-embedding” model. We implemented this model to provide a baseline for comparison.
+
+The classifier used in the transfer strength metric achieves very high accuracy (0.832 and 0.99 for the Shakespeare and Headlines datasets correspondingly). These results concur with the results of Shen et al. (2017) and Fu et al. (2017), and show that the two forms in the corpora are significantly different.
+
+Following Fu et al. (2017), we show the result of different configuration of the size of the form and meaning vectors on Figure 2. Namely, we report combinations of 64 and 256-dimensional vectors. Note that the sizes of the form vector are important. The larger is the form vector, the higher is the transfer strength, but smaller is content preservation. This is consistent with Fu et al. (2017), where they observed a similar behaviour.
+
+It is clear that the proposed method achieves significantly better transfer strength then the previously proposed model. It also has a lower content preservation score, which means that it repeats fewer exact words from the source sentence. Note that a low transfer strength and very high (0.9) content ˜ preservation score means that the model was not able to successfully learn to transfer the form and the target sentence is almost identical to the source sentence. The Shakespeare dataset is the hardest for the model in terms of transfer strength, probably because it is the smallest dataset, but the proposed method performs consistently well in transfer of both form and meaning and, in contrast to the baseline.
+
+Fluency of generated sentences Note that there is no guarantee that the generated sentences would be coherent after switching the form vector. In order to estimate how this switch affects the fluency of generated sentences, we trained a language model on the Shakespeare dataset and calculated the perplexity of the generated sentences using the original form vector and the average of form vectors of $k$ random sentences from the opposite style (see subsubsection 5.1.1). While the perplexity of such sentences does go up, this change is not big (6.89 vs 9.74).
+
+# 6.1 IMPACT OF THE MOTIVATIONAL TRAINING
+
+To investigate the impact of the motivator, we visualized form and meaning embeddings of 1000 random samples from the Headlines dataset using t-SNE algorithm (Van Der Maaten, 2014) with the Multicore-TSNE library (Ulyanov, 2016). The result is presented in Figure 3.
+
+There are three important observations. First, there is no clear separation in the meaning embeddings, which means that any accurate form transfer is due to the form embeddings, and the dissociation of form and meaning was successful.
+
+Second, even without the motivator the model is able to produce the form embeddings that are clustered into two group. Recall from section 4 that without the motivational loss there are no forces that influence the form embeddings, but nevertheless the model learns to separate them.
+
+However, the separation effect is much more pronounced in the presence of motivator. This explains why the motivator consistently improved transfer strength of ADNet, as shown in Figure 2.
+
+![](images/177a04829132ea7eb43a11912ab21cf3b18fabd8252eb4e7e976664a6e7b58b0.jpg)  
+Figure 3: t-SNE visualization of the form and meaning embeddings of 1000 random sentences. Green point represent sentences form news headlines, and red points represent titles of scientific articles.
+
+# 6.2 QUALITATIVE EVALUATION
+
+Table 1 and Table 2 show several examples of the successful form/meaning transfer achieved by ADNet. Table 1 presentes the results of an experiment that to some extent replicates the approach taken by the authors who treat linguistic form as a binary variable (Shen et al., 2017; Fu et al., 2017). The sentences the original Shakespeare plays were averaged to get the “typical” Early Modern English form vector. This averaged vector was used to decode a sentence from the modern English translation back into the original. The same was done in the opposite direction.
+
+Table 1: Decoding of the source sentence from Early Modern English (EME) into contemporary English (CE), and vice versa.   
+
+<table><tr><td>Aye,sir. (EME)</td><td>→ Yes,sir. (CE)</td></tr><tr><td>Fare thee well, my lord (EME)</td><td>Fare you well, my lord (CE)</td></tr><tr><td>This guy will tell us everything. (CE)</td><td>This man will tell us everything. (EME)</td></tr><tr><td>Ive done no more to caesar than you will do to me.(CE)</td><td>Ihave done no more to caesar than,you shall do to me.(EME)</td></tr></table>
+
+Table 2 illustrates the possibilities of ADNet on fine-grained transfer applied to the change of register. We encoded two sentences in different registers from the Headlines dataset to produce form and meaning embeddings, and then we decoded the first sentence with the meaning embedding of the second, and vice versa. As can be seen from Table 2, the model correctly captures the meaning of sentences and decodes them using the form of the source sentences. Note how the model preserves specific words and the structure of the source sentence. In particular, note how in the first example, the model decided to put the colon after the “crisis management”, as the source form sentence has this syntactic structure (“A review:”). This is not possible in the previously proposed models, as they treat form as just a binary variable.
+
+<table><tr><td>A review: detection techniques for LTE system Situation management knowledge from social media</td><td>→Crisis management: media practices in telecommunication management →Areview study against intelligence internet</td></tr><tr><td>Security flaw could notaffect digital devices,experts say Semantic approach to event processing</td><td>→Semantic approach approach: current multimedia networks as modeling processes →Security flaw to verifyleaks</td></tr></table>
+
+Table 2: Flipping the meaning and the form embeddings of two sentence from different registers. Note the use of the colon in the first example, and the use of the “to”-constructions in the second example, consistent with the form of the source sentences.
+
+# 6.3 PERFORMANCE OF MEANING EMBEDDINGS ON DOWNSTREAM TASKS
+
+We conducted some experiments to test the assumption that the derived meaning embeddings should improve performance on downstream tasks that require understanding of the meaning of the sentences regardless of their form. We evaluated embeddings produced by the ADNet, trained in the Headlines dataset, on a task of paraphrase detection. We used the SentEval toolkit (Conneau et al., 2017) and the Microsoft Research Paraphrase Corpus (Dolan et al., 2004). The F1 scores on this task for different models are presented in Table 3. Note that all models, except InferSent, are unsupervised. The InferSent model was trained on a big SNLI dataset, consisting of more than 500,000 manually annotated pairs. ADNet achieves the the highest score among the unsupervised systems and outperforms the regular sequence-to-sequence autoencoder with a large gap.
+
+Table 3: F1 scores on the task of paraphrase detection using the SentEval toolkit (Conneau et al., 2017)   
+
+<table><tr><td>BoW</td><td>Seq2Seq</td><td>InferSent</td><td>Fu et al. (2017)</td><td>ADNet</td></tr><tr><td>80.82</td><td>74.68</td><td>83.17</td><td>78.88</td><td>81.38</td></tr></table>
+
+# 7 CONCLUSION
+
+In this paper, we presented ADNet, a new model that performs adversarial decomposition of text representation. In contrast to previous work, it does not require a parallel training corpus and works directly on hidden representations of sentences. Most importantly, is does not treat the form as a binary variable (as done in most previously proposed models), enabling a fine-grained change of the form of sentences or specific aspects of meaning. We evaluate ADNet on two tasks: the shift of language register and diachronic language change. Our solution achieves superior results, and t-SNE visualizations of the learned meaning and style embeddings illustrate that the proposed motivational loss leads to significantly better separation of the form embeddings.
+
+# REFERENCES
+
+Samuel Albanie, Sebastien Ehrhardt, and Jo ´ ao F Henriques. Stopping gan violence: Generative ˜ unadversarial networks. arXiv preprint arXiv:1703.02528, 2017.
+
+Martin Arjovsky, Soumith Chintala, and Leon Bottou. Wasserstein gan. ´ arXiv preprint arXiv:1701.07875, 2017.
+
+Sanjeev Arora, Yingyu Liang, and Tengyu Ma. A simple but tough-to-beat baseline for sentence embeddings. In International Conference on Learning Representation, 2017.
+
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.
+
+Keith Carlson, Allen Riddell, and Daniel Rockmore. Zero-shot style transfer in text using recurrent neural networks. arXiv preprint arXiv:1711.04731, 2017.
+
+Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. Universal sentence encoder. arXiv preprint arXiv:1803.11175, 2018.
+
+Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
+
+Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´ learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.
+
+Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo¨ıc Barrault, and Antoine Bordes. Supervised learning of universal sentence representations from natural language inference data. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 670–680, 2017.
+
+Ferdinand de Saussure. Course in General Linguistics. New York : Philosophical Library, 1959. URL http://archive.org/details/courseingenerall00saus.
+
+Dua Dheeru and Efi Karra Taniskidou. UCI machine learning repository, 2017. URL http: //archive.ics.uci.edu/ml.
+
+Bill Dolan, Chris Quirk, and Chris Brockett. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In Proceedings of the 20th international conference on Computational Linguistics, pp. 350. Association for Computational Linguistics, 2004.
+
+Jessica Ficler and Yoav Goldberg. Controlling linguistic style aspects in neural language generation. In Proceedings of the Workshop on Stylistic Variation, pp. 94–104, 2017.
+
+Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. Style transfer in text: Exploration and evaluation. arXiv preprint arXiv:1711.06861, 2017.
+
+Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014.
+
+Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pp. 140–149, 2016.
+
+M. A. K. Halliday, A. McIntosh, and P. Stevens. The Linguistic Sciences and Language Teaching. Longmans, Green and Co., London, 1968.
+
+Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. Toward controlled generation of text. In International Conference on Machine Learning, pp. 1587–1596, 2017.
+
+Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1125–1134, 2017.
+
+Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. Shakespearizing modern language using copy-enriched sequence-to-sequence models. arXiv preprint arXiv:1707.01161, 2017.
+
+Yoon Kim, Kelly Zhang, Alexander M Rush, Yann LeCun, et al. Adversarially regularized autoencoders for generating discrete structures. arXiv preprint arXiv:1706.04223, 2017.
+
+Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
+
+Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances In Neural Information Processing Systems, pp. 4601–4609, 2016.
+
+Jiwei Li, Michel Galley, Chris Brockett, Georgios Spithourakis, Jianfeng Gao, and Bill Dolan. A persona-based neural conversation model. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pp. 994–1003, 2016.
+
+Zachary C Lipton, Sharad Vikram, and Julian McAuley. Generative concatenative nets jointly learn to write and classify reviews. arXiv preprint arXiv:1511.03683, 2015.
+
+Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018.
+
+Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015.
+
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016.
+
+Jonas Mueller, David Gifford, and Tommi Jaakkola. Sequence to better sequence: continuous revision of combinatorial structures. In International Conference on Machine Learning, pp. 2536– 2544, 2017.
+
+Nikita Nangia, Adina Williams, Angeliki Lazaridou, and Samuel R Bowman. The repeval 2017 shared task: Multi-genre natural language inference with sentence representations. arXiv preprint arXiv:1707.08172, 2017.
+
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pp. 311–318. Association for Computational Linguistics, 2002.
+
+Peter Potash, Alexey Romanov, and Anna Rumshisky. Ghostwriter: using an lstm for automatic rap lyric generation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1919–1924, 2015.
+
+Peter Potash, Alexey Romanov, and Anna Rumshisky. Here’s my point: Joint pointer architecture for argument mining. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 1364–1373, 2017.
+
+William Safire. On Language Gotcha! Gang Strikes Again, 1995. URL https://www.nytimes.com/1995/12/31/magazine/ on-language-gotcha-gang-strikes-again.html.
+
+Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. Style transfer from non-parallel text by cross-alignment. In Advances in Neural Information Processing Systems, pp. 6833–6844, 2017.
+
+Efstathios Stamatatos. Authorship attribution using text distortion. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, volume 1, pp. 1138–1149, 2017.
+
+Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018.
+
+Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pp. 3104–3112, 2014.   
+Jian Tang, Yifan Yang, Sam Carton, Ming Zhang, and Qiaozhu Mei. Context-aware natural language generation with recurrent neural networks. arXiv preprint arXiv:1611.09900, 2016.   
+Dmitry Ulyanov. Multicore-tsne. https://github.com/DmitryUlyanov/ Multicore-TSNE, 2016.   
+Laurens Van Der Maaten. Accelerating t-sne using tree-based algorithms. Journal of machine learning research, 15(1):3221–3245, 2014.   
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp. 6000–6010, 2017.   
+Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in Neural Information Processing Systems, pp. 2692–2700, 2015.   
+Wei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and Colin Cherry. Paraphrasing for style. Proceedings of COLING 2012, pp. 2899–2914, 2012.   
+Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2223–2232, 2017.
+
+# A MULTIPLE FORMS AND STYLISTIC SIMILARITIES
+
+![](images/0e733eb345569ac00898384b916d68adac3771ae1b15c0710524723af5f8dfcc.jpg)  
+Figure 4: t-SNE visualization of the form and meaning embeddings. Each color corresponds to a different author.
+
+In order to go beyond just two different forms, we experimented with training the model on a set of literature novels from six different authors from Project Gutenberg2 written in two different time periods. A t-SNE visualization of the resulting meaning and form embeddings is presented in Figure 4. Note how form embeddings create a six-pointed star. After further examination, we observed that common phrases (for example, “Good morning” or “Hello!”) were embedded into the center of the star, whereas the most specific sentences from a given author were placed into the rays of the star. In particular, some sentences included character names, thus further research is required to mitigate this problem. Stamatatos (2017) provides a promising direction for solving this.

parse/train/SJzwb2RcK7/SJzwb2RcK7_content_list.json

@@ -0,0 +1,1700 @@
+[
+    {
+        "type": "text",
+        "text": "ADVERSARIAL DECOMPOSITION OF TEXT REPRESENTATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            98,
+            821,
+            145
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Anonymous authors Paper under double-blind review ",
+        "bbox": [
+            183,
+            171,
+            398,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            236,
+            544,
+            251
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present a method for adversarial decomposition of text representation. This method can be used to decompose a representation of an input sentence into several independent vectors, where each vector is responsible for a specific aspect of the input sentence. We evaluate the proposed method on two case studies: the conversion between different social registers and diachronic language change. We show that the proposed method is capable of fine-grained controlled change of these aspects of the input sentence. For example, our model is capable of learning a continuous (rather than categorical) representation of the style of the sentence, in line with the reality of language use. The model uses adversarial-motivational training and includes a special motivational loss, which acts opposite to the discriminator and encourages a better decomposition. Finally, we evaluate the obtained meaning embeddings on a downstream task of paraphrase detection and show that they are significantly better than embeddings of a regular autoencoder. ",
+        "bbox": [
+            233,
+            268,
+            764,
+            462
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            493,
+            336,
+            510
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the recent successes in using neural models for representation learning for natural language text, learning a meaningful representation of input sentences remains an open research problem. A variety of approaches, from sequence-to-sequence models that followed the work of Sutskever et al. (2014) to the more recent proposals (Arora et al., 2017; Nangia et al., 2017; Conneau et al., 2017; Logeswaran & Lee, 2018; Subramanian et al., 2018; Cer et al., 2018) share one common drawback. Namely, all of them encode the input sentence into just one single vector of a fixed size. One way to bypass the limitations of a single vector representation is to use an attention mechanism (Bahdanau et al., 2014; Vaswani et al., 2017). We propose to approach this problem differently and design a method for adversarial decomposition of the learned input representation into multiple components. Our method encodes the input sentence into several vectors, where each vector is responsible for a specific aspect of the sentence. ",
+        "bbox": [
+            174,
+            526,
+            825,
+            679
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In terms of learning different separable components of input representation, our work most closely relates to the style transfer work, which has been applied to a variety of different aspects of language, from diachronic language differences (Xu et al., 2012) to authors’ personalities (Lipton et al., 2015) and even sentiment (Hu et al., 2017; Fu et al., 2017). The style transfer work effectively relies on the more classical distinction between meaning and form (de Saussure, 1959), which accounts for the fact that multiple surface realizations are possible for the same meaning. For simplicity, we will use this terminology throughout the rest of the paper. ",
+        "bbox": [
+            174,
+            688,
+            825,
+            784
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Consider the case when we encode an input sentence into a meaning vector and a form vector. We are then able to perform a controllable change of meaning or form by a simple change applied to these vectors. For example, we can encode two sentences written in two different styles, then swap the form vectors while leaving the meaning vectors intact. We can then generate new unique sentences with the original meaning, but written in a different style. ",
+        "bbox": [
+            174,
+            791,
+            823,
+            861
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the present work, we propose a novel model for this type of decomposition based on adversarialmotivational training and design an architecture inspired by the GANs (Goodfellow et al., 2014) and adversarial autoencoders (Makhzani et al., 2015). In addition to the adversarial loss, we use a special motivator (Albanie et al., 2017), which, in contrast to the discriminator, is used to provide a motivational loss to encourage the model to better decomposition of the meaning and the form, as well as specific aspects of meaning. We make all the code publicly available on GitHub 1. ",
+        "bbox": [
+            176,
+            868,
+            823,
+            922
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            103,
+            821,
+            132
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the proposed methods for learning separate aspects of input representation on the following case studies: ",
+        "bbox": [
+            173,
+            138,
+            820,
+            166
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. Learning to separate out a representation of the specific diachronic slice of language. One may express the same meaning using the Early Modern English (e.g. What would she have?) and the contemporary English ( What does she want?)   \n2. Learning a representation for a social register (Halliday et al., 1968) – that is, subsets of language appropriate in a given context or characteristic of a certain group of speakers. These include formal and informal language, the language used in different genres (e.g., fiction vs. newspapers vs. academic texts), different dialects, and even literary idiostyles. We experiment with the registers corresponding to the titles of scientific papers vs. newspaper articles. ",
+        "bbox": [
+            210,
+            179,
+            825,
+            310
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 RELATED WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            333,
+            341,
+            348
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As mentioned above, the most relevant previous work comes from the style transfer research, and it can be divided into two groups: ",
+        "bbox": [
+            174,
+            364,
+            823,
+            393
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. Approaches that aim to generate text in a given form. For example, the task may be to produce just any verse as long as it is in the “style” of the target poet. 2. Approaches that aim to induce a change in either the “form” or the “meaning” of an existing utterance. For example, “Good bye, Mr. Anderson.” can be transformed to “Fare you well, good Master Anderson” (Xu et al., 2012)). ",
+        "bbox": [
+            212,
+            406,
+            825,
+            482
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "An example of the first group is the work by Potash et al. (2015), who trained several separate networks on verses by different hip-hip artists. An LSTM network successfully generated verses that were stylistically similar to the verses of the target artist (as measured by cosine distance on TfIdf vectors). More complicated approaches use language models that are conditioned in some way. For example, Lipton et al. (2015) produced product reviews with a target rating by passing the rating as an additional input at each timestep of an LSTM model. Tang et al. (2016) generated reviews not only with a given rating but also for a specific product. At each timestep a special context vector was provided as input, gated so as to enable the model to decide how much attention to pay to that vector and the current hidden state. Li et al. (2016) used “speaker” vectors as an additional input to a conversational model, improving consistency of dialog responses. Finally, Ficler & Goldberg (2017) performed an extensive evaluation of conditioned language models based on “content” (theme and sentiment) and “style” (professional, personal, length, descriptiveness). Importantly, they showed that it is possible to control both “content” and “style” simultaneously. ",
+        "bbox": [
+            174,
+            494,
+            825,
+            674
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Work from the second group can further be divided into two clusters by the nature of the training data: parallel aligned corpora, or non-aligned datasets. The aligned corpora enable approaching the problem of form shift as a paraphrasing or machine translation problem. Xu et al. (2012) used statistical and dictionary-based systems on a dataset of original plays by Shakespeare and their contemporary translations. Carlson et al. (2017) trained an LSTM network on 33 versions of the Bible. Jhamtani et al. (2017) used a Pointer Network (Vinyals et al., 2015), an architecture that was successfully applied to a wide variety of tasks (Merity et al., 2016; Gulcehre et al., 2016; Potash et al., 2017), to enable direct copying of the input tokens to the output. Note that these works use BLEU (Papineni et al., 2002) as the main, or even the only evaluation measure. This is only possible in cases where a parallel corpus is available. ",
+        "bbox": [
+            174,
+            681,
+            825,
+            820
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, new approaches that do not require a parallel corpora were developed in both CV (Zhu et al., 2017) and NLP. Hu et al. (2017) succeeded in changing tense and sentiment of sentences with a two steps procedure based on a variational auto-encoder (VAE) (Kingma & Welling, 2013). After training a VAE, a discriminator and a generator are trained in an alternate manner, where the discriminator tries to correctly classify the target sentence attributes. A special loss component forces the hidden representation of the encoded sentence to not have any information about the target sentence attributes. Mueller et al. (2017) used a VAE to produce a hidden representation of a sentence, and then modify it to match the desired form. Unlike Hu et al. (2017), they do not separate the form and meaning embeddings. Shen et al. (2017) applied a GAN to align the hidden representation of sentences from two corpora and force them to do not have any information about the form via adversarial loss. During the decoding, similarly the work by Lipton et al. (2015), special “style” vectors are passed to the decoder at every timestep to produce a sentence with the desired properties. The model is trained using the Professor-Forcing algorithm (Lamb et al., 2016). Kim et al. (2017) worked directly on hidden space vectors that are constrained with the same adversarial loss instead of outputs of the generator, and use two different generators for two different “styles”. Finally, Fu et al. (2017) proposed two models for generating sentences with the target properties using an adversarial loss, similarly to Shen et al. (2017) and Kim et al. (2017). ",
+        "bbox": [
+            174,
+            828,
+            823,
+            897
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            103,
+            825,
+            270
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Comparison with previous work In contrast to the proposals of $\\mathrm { X u }$ et al. (2012), Carlson et al. (2017), Jhamtani et al. (2017), our solution does not require a parallel corpus. Furthermore, unlike the model by Shen et al. (2017), our model works directly on representation of sentences in the hidden space. ",
+        "bbox": [
+            174,
+            285,
+            825,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Most importantly, in contrast to the proposals by Mueller et al. (2017), Hu et al. (2017), Kim et al. (2017), Fu et al. (2017), our model produces a representation for both meaning and form and does not treat the form as a categorical (in the vast majority of works, binary) variable. Although the form was represented as dense vectors in previous work, it is still just a binary feature, as they use a single pre-defined vector for each form, with all sentences of the same form assigned the same form vector. In contrast, our work treats form as a truly continuous variable, where each sentence has its own, unique, form vector. ",
+        "bbox": [
+            174,
+            348,
+            825,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Treating meaning and form not as binary/categorical, but as continuous is more consistent with the reality of language use, since there are different degrees of overlap between the language used by different registers or in different diachronic slices. Indeed, language change is gradual, and the acceptability of expressions in a given register also forms a continuum, so one expects a substantial overlap between the grammar and vocabulary used, for example, on Twitter and by New York Times. To the best of our knowledge, this is the first model that considers linguistic form in the task of text generation as a continuous variable. ",
+        "bbox": [
+            174,
+            452,
+            825,
+            549
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "One significant consequence of learning a continuous representation for form is that it allows the model to work with a large, and potentially infinite, number of forms. Note that in this case the locations of areas of specific forms in the vector style space would reflect the similarity between these forms. For example, the proposed model could be directly applied to the authorship attribution problem. In this case, each author would have their own area in the form space, and the more similar the authors are in terms of writing style, the closer these areas would be to each other. We performed preliminary experiments on this and report the results in Appendix A. ",
+        "bbox": [
+            174,
+            556,
+            825,
+            655
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 FORMULATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            674,
+            331,
+            690
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Let us formulate the problem of decomposition of text representation on an example of controlled change of linguistic form and conversion of Shakespeare plays in the original Early Modern to contemporary English. Let $X ^ { a }$ be a corpus of texts $\\pmb { x } _ { i } ^ { a } \\in \\mathcal { X } ^ { a }$ in Early Middle English $\\mathbf { f } ^ { a } \\in \\mathcal { F }$ , and $X ^ { b }$ be a corpus of texts $\\pmb { x } _ { i } ^ { b } \\in \\mathcal { X } ^ { b }$ in modern English $\\mathbf { f } ^ { b } \\in \\mathcal { F }$ . We assume that the texts in both $X ^ { a }$ and $X ^ { b }$ has the same distribution of meaning $\\mathbf { m } \\in \\mathcal { M }$ . The form f , however, is different and generated from a mixture of two distributions: ",
+        "bbox": [
+            174,
+            704,
+            825,
+            792
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/4a11e7c60e2c8b0d0355c151e4bfe43c3f0471022bc86dfcef0ddc385ae6c31f.jpg",
+        "text": "$$\n\\mathbf { f } _ { i } = \\alpha _ { i } ^ { a } p ( \\mathbf { f } ^ { a } ) + \\alpha _ { i } ^ { b } p ( \\mathbf { f } ^ { b } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            415,
+            797,
+            581,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where $\\mathbf { f } ^ { a }$ and $\\mathbf { f } ^ { b }$ are two different languages (Early Modern and contemporary English). Intuitively, we say that a sample $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }$ has the form $\\mathbf { f } ^ { a }$ if $\\alpha _ { i } ^ { a } > \\alpha _ { i } ^ { b }$ , and it has the form $\\mathbf { f } ^ { b }$ if $\\alpha _ { i } ^ { b } > \\alpha _ { i } ^ { a }$ . ",
+        "bbox": [
+            174,
+            823,
+            823,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The goal of dissociation meaning and form is to learn two encoders $E _ { \\mathbf { m } } : \\mathcal { X }  \\mathcal { M }$ and $E _ { \\mathbf { f } } : \\mathcal { X }  \\mathcal { F }$ for the meaning and form correspondingly, and the generator $G : { \\mathcal { M } } , { \\mathcal { F } } \\to { \\mathcal { X } }$ such that ",
+        "bbox": [
+            173,
+            859,
+            828,
+            890
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/a441574f5e8c112d32d76863ae817cd1ac446e2f802d171019672f92738679d7.jpg",
+        "text": "$$\n\\forall j \\in \\{ a , b \\} , \\forall k \\in \\{ a , b \\} : G ( E _ { \\mathbf { m } } ( x ^ { k } ) , E _ { \\mathbf { f } } ( x ^ { j } ) )  \\mathcal { X } ^ { j }\n$$",
+        "text_format": "latex",
+        "bbox": [
+            316,
+            906,
+            679,
+            925
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "That is, the form of a generated sample depends exclusively on the provided $\\mathbf { f } _ { j }$ and can be the in the same domain for two different ${ \\bf m } _ { u }$ and $\\mathbf { m } _ { v }$ from two samples from different domains $\\mathcal { X } ^ { a }$ and $\\mathcal { X } ^ { b }$ . ",
+        "bbox": [
+            173,
+            103,
+            823,
+            133
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Note that, in contrast to the previously proposals, the form f is not a categorical variable but a continuous vector. This enables fine-grained controllable change of form: the original form $\\mathbf { f } _ { i }$ is changed to reflect the form of the specific target sentence $\\mathbf { f } _ { j }$ with its own unique $\\alpha ^ { a }$ and $\\alpha ^ { b }$ while preserving the original meaning $\\mathbf { m } _ { i }$ . ",
+        "bbox": [
+            174,
+            140,
+            825,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "An important caveat concerns the core assumption of the similar meaning distribution in the two corpora, which is also made in all other works reviewed in Section 2. It limits the possible use of this approach to cases where the distributions are in fact similar (i.e. parallel or at least comparable corpora are available). It does not apply to many cases that could be analyzed in terms of meaning and form. For example, books for children and scholarly papers are both registers, they have their own form (i.e. specific subsets of linguistic means and structure conventions) – but there is little overlap in the content. This would make it hard even for a professional writer to turn a research paper into a fairy tale. ",
+        "bbox": [
+            174,
+            204,
+            825,
+            315
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 METHOD DESCRIPTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            362,
+            398,
+            377
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by Makhzani et al. (2015), Kim et al. (2017), and Albanie et al. (2017), we propose ADNet, a new model for adversarial decomposition of text representation (Figure 1). ",
+        "bbox": [
+            173,
+            409,
+            823,
+            436
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3abb02e83857746da9d726ec7398bd1454ecd5ed2a8931c9876ed697853ada2e.jpg",
+        "image_caption": [
+            "Figure 1: Overview of ADNet. Encoder encodes the inputs sentences into two latent vectors m and f. The Generator takes them as the input and produces the output sentence. During the training, the Discriminator is used for an adversarial loss that forces m to do not carry any information about the form, and the M otivator is used for a motivational loss that encourages f to carry the needed information about the form. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            455,
+            473,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our solution is based on a widely used sequence-to-sequence framework (Sutskever et al., 2014) and consists of four main parts. The encoder $E$ encodes the inputs sequence $_ { \\textbf { \\em x } }$ into two latent vectors $\\mathbf { m }$ and f which capture the meaning and the form of the sentence correspondingly. The generator $G$ then takes these two vectors as the input and produces a reconstruction of the original input sequence $\\hat { \\pmb x }$ . ",
+        "bbox": [
+            514,
+            444,
+            823,
+            569
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The encoder and generator by themselves will likely not achieve the dissociation of the meaning and form. We encourage this behavior in a way similar to Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), which had an overwhelming success the past few years and have been proven to be a good way of enforcing a specific distribution and characteristics on the output of a model. ",
+        "bbox": [
+            514,
+            577,
+            823,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the work of Albanie et al. (2017) and the principle of ”carrot and stick” (Safire, 1995), in contrast to the majority of work that promotes pure adversarial approach (Goodfel",
+        "bbox": [
+            514,
+            708,
+            823,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "low et al., 2014; Shen et al., 2017; Fu et al., 2017; Zhu et al., 2017), we propose two additional components, the discriminator $D$ and the motivator $M$ to force and motivate the model to learn the dissociation of the meaning and the form. Similarly to a regular GAN model, the adversarial discriminator $D$ tries to classify the form f based on the latent meaning vector $\\mathbf { m }$ , and the encoder $E$ is penalized to make this task as hard as possible. ",
+        "bbox": [
+            174,
+            763,
+            823,
+            833
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Opposed to such vicious behaviour, the motivator $M$ tries to classify the form based on the latent form vector f, as it should be done, and encourages the encoder $E$ to make this task as simple as possible. We could apply the adversarial approach here as well and force the distribution of the form vectors to fit a mixture of Gaussians (in this particular case, a mixture of two Guassians) with another discriminator, as it is done by Makhzani et al. (2015), but we opted for the “dualistic” path of two complimentary forces. ",
+        "bbox": [
+            174,
+            840,
+            825,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 ENCODER-DECODER ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            103,
+            359,
+            118
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Both the encoder $E$ and the generator $G$ are modeled with a neural network. Gated Recurrent Unit (GRU) (Chung et al., 2014) is used for $E$ to encode the input sentence $_ { \\textbf { \\em x } }$ into a hidden vector $\\pmb { h } = \\mathbf { G } \\mathbf { R } \\mathbf { U } ( \\pmb { x } )$ . ",
+        "bbox": [
+            174,
+            128,
+            825,
+            172
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The vector $^ { h }$ is then passed through two different fully connected layers to produce the latent vectors of the form and the meaning of the input sentence: ",
+        "bbox": [
+            171,
+            178,
+            823,
+            207
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/0deb193caf16da6d6a73f2f3fc17dc117c64c3f456de8a4cea96a726fc6ed20e.jpg",
+        "text": "$$\n{ \\bf m } = \\operatorname { t a n h } ( W _ { m } h + b _ { m } ) { \\bf f } = \\operatorname { t a n h } ( W _ { f } h + b _ { f } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            326,
+            210,
+            669,
+            228
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use $\\theta _ { E }$ to denote the parameters of the encoder $E \\colon W _ { m } , b _ { m } , W _ { f } , b _ { f }$ , and the parameters of the GRU unit. ",
+        "bbox": [
+            173,
+            237,
+            825,
+            266
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The generator $G$ is also modelled with a GRU unit. The generator takes as input the meaning vector $\\mathbf { m }$ and the form vector f, concatenates them, and passes trough a fully-connected layer to obtain a hidden vector $_ z$ that represents both meaning and form of the original input sentence: ",
+        "bbox": [
+            173,
+            272,
+            825,
+            315
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/ae7a560cbf35d009c903ccd101880e0426bc40be0e6a42687f1a980136116b70.jpg",
+        "text": "$$\nz = \\operatorname { t a n h } ( W _ { z } [ \\mathbf { m } ; \\mathbf { f } ] + b _ { m } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            405,
+            319,
+            591,
+            337
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After that, we use a GRU unit to generate the output sentence as a probability distribution over the vocabulary tokens: ",
+        "bbox": [
+            176,
+            339,
+            823,
+            366
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/60b5cf97e32b9c2307e0e93e84c169b7d5520a7ec2f571dbfbb0930d11558f63.jpg",
+        "text": "$$\np ( \\hat { \\pmb x } ) = \\prod _ { t = 1 } ^ { T } p ( \\hat { \\pmb x } _ { t } | \\boldsymbol z , \\hat { \\pmb x } _ { 1 } , \\dots , \\hat { \\pmb x } _ { t - 1 } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            383,
+            364,
+            612,
+            409
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use $\\theta _ { G }$ to denote the parameters of the generator $G$ : $W _ { z }$ , $b _ { m }$ , and the parameters of the used GRU. The encoder and generator are trained using the standard reconstruction loss: ",
+        "bbox": [
+            171,
+            415,
+            825,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/8810459513e31102b13681a71fb72547ce03cc5c0aa9cff380ff21ef787e92ec.jpg",
+        "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { r e c } } ( \\pmb { \\theta } _ { E } , \\pmb { \\theta } _ { G } ) = \\mathbb { E } _ { \\pmb { x } \\sim \\pmb { X } ^ { a } } [ - \\log p ( \\hat { \\pmb x } | \\pmb { x } ) ] + \\mathbb { E } _ { \\pmb { x } \\sim \\pmb { X } ^ { b } } [ - \\log p ( \\hat { \\pmb x } | \\pmb { x } ) ] } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            287,
+            463,
+            710,
+            481
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 DISCRIMINATOR ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            496,
+            330,
+            511
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The representation of the meaning m produced by the encoder $E$ should not contain any information about the form f . We achieve this by using an adversarial approach. First, we train a discriminator $D$ , consisting of several fully connected layers with ELU activation function (Clevert et al., 2015) between them, to predict the form f of a sentence by its meaning vector: $\\hat { f } _ { D } = D ( { \\bf m } )$ , where $\\hat { f }$ is the score (logit) reflecting the probability of the sentence $_ { \\textbf { \\em x } }$ to belong to one of the form domains. ",
+        "bbox": [
+            173,
+            521,
+            825,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Motivated by the Wasserstein GAN (Arjovsky et al., 2017), we use the following loss function instead of the standard cross-entropy: ",
+        "bbox": [
+            173,
+            599,
+            823,
+            628
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/74103d494cb6da9fc3f8a3f091a472533016f4cb31134796ded0580940289afd.jpg",
+        "text": "$$\n\\mathcal { L } _ { D } ( \\pmb { \\theta } _ { D } ) = \\mathbb { E } _ { \\pmb { x } \\sim \\pmb { X } ^ { a } } [ D ( E _ { \\mathbf { m } } ( \\pmb { x } ) ) ] - \\mathbb { E } _ { \\pmb { x } \\sim \\pmb { X } ^ { b } } [ D ( E _ { \\mathbf { m } } ( \\pmb { x } ) ) ]\n$$",
+        "text_format": "latex",
+        "bbox": [
+            313,
+            632,
+            684,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Thus, a successful discriminator will produce negative scores $\\hat { f }$ for sentences from $X ^ { a }$ and positive scores for sentences from $X ^ { b }$ . This discriminator is then used in an adversarial manner to provide a learning signal for the encoder and force dissociation of the meaning and form by maximizing $\\mathcal { L } _ { D }$ $: \\mathcal { L } _ { \\mathrm { a d v } } ( \\pmb { \\theta } _ { E } ) = - \\lambda _ { \\mathrm { a d v } } \\mathcal { L } _ { D }$ , where $\\lambda _ { \\mathrm { a d v } }$ is a hyperparameter reflecting the strength of the adversarial loss. Note that this loss applies to the parameters of the encoder. ",
+        "bbox": [
+            173,
+            661,
+            825,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 MOTIVATOR ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            748,
+            299,
+            763
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our experiments showed that it is enough to have just the discriminator $D$ and the adversarial loss $\\mathcal { L } _ { \\mathrm { a d v } }$ to force the model to dissociate the form and the meaning. However, in order to achieve a better dissociation, we propose to use a motivator $M$ (Albanie et al., 2017) and the corresponding motivational loss. Conceptually, this is the opposite of the adversarial loss, hence the name. As the discriminator $D$ , the motivator $M$ learns to classify the form f of the input sentence. However, its input is not not the meaning vector but the form vector: ${ \\hat { f } } _ { M } = M ( \\mathbf { f } )$ . ",
+        "bbox": [
+            173,
+            773,
+            825,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The motivator has the same architecture as the discriminator, and the same loss function. While the adversarial loss forces the encoder $E$ to produce a meaning vector m with no information about the form f , the motivational loss encourages $E$ to encode this information in the form vector by minimizing ${ \\mathcal { L } } _ { M }$ ${ \\bf \\chi } _ { \\ / t } \\colon \\mathcal { L } _ { \\mathrm { m o t i v } } ( \\pmb { \\theta } _ { E } ) = \\lambda _ { \\mathrm { m o t i v } } \\mathcal { L } _ { \\ / M }$ . ",
+        "bbox": [
+            174,
+            867,
+            823,
+            925
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4 TRAINING PROCEDURE ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            103,
+            375,
+            117
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall training procedure follows the methods for training GANs (Goodfellow et al., 2014; Arjovsky et al., 2017) and consists of two stages: training the discriminator $D$ and the motivator $M$ , and training the encoder $E$ and the generator $G$ . ",
+        "bbox": [
+            174,
+            128,
+            823,
+            171
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In contrast to Arjovsky et al. (2017), we do not train the $D$ and $M$ more than the $E$ and the $G$ . In our experiments we found that simple training in two stages is enough to achieve dissociation of the meaning and the form. Encoder and generator are trained with the following loss function that combines reconstruction loss with the losses from the discriminator and the motivator: ",
+        "bbox": [
+            174,
+            178,
+            825,
+            234
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "img_path": "images/0f9ae7dacfb13a6ce4f138bbdd554f89aea0ca4c9ec776ed1fbdc79d0c83fb7c.jpg",
+        "text": "$$\n\\mathcal { L } _ { \\mathrm { t o t a l } } ( \\theta _ { E } , \\theta _ { G } ) = \\mathcal { L } _ { \\mathrm { r e c } } + \\mathcal { L } _ { \\mathrm { a d v } } + \\mathcal { L } _ { \\mathrm { m o t i v } }\n$$",
+        "text_format": "latex",
+        "bbox": [
+            374,
+            242,
+            622,
+            260
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 EXPERIMENTAL SETUP ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            279,
+            397,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 EVALUATION ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            310,
+            307,
+            324
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Similarly to the evaluation of style transfer in CV (Isola et al., 2017), evaluation of this task is difficult. We follow the approach of Isola et al. (2017); Shen et al. (2017) and recently proposed by Fu et al. (2017) methods of evaluation of “transfer strength” and “content preservation”. The authors showed the proposed automatic metrics to a large degree correlate with human judgment and can serve as a proxy. Below we give an overview of these metrics. ",
+        "bbox": [
+            174,
+            335,
+            825,
+            406
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Transfer Strength. The goal of this metric is to capture whether the form has been changed successfully. To do that, a classifier $C$ is trained on the two corpora, $X ^ { a }$ and $X ^ { b }$ to recognize the linguistic “form” typical of each of them. After that, a sentence the form/meaning of which was changed is passed to the classifier. The overall accuracy reflects the degree of success of changing the form/meaning. This approach is widely used in CV (Isola et al., 2017), and was applied in NLP as well (Shen et al., 2017). ",
+        "bbox": [
+            174,
+            421,
+            825,
+            506
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In our experiments we used a GRU unit followed by four fully-connected layers with ELU activation functions between them as the classifier. ",
+        "bbox": [
+            173,
+            512,
+            823,
+            541
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Content preservation Note that transfer strength by itself does not capture the overall quality of a changed sentence. A extremely overfitted model that produces the same, the most characteristic sentence of one corpus all the time would have a high score according to this metric. Thus, we need to measure how much of the meaning was preserved while changing the form. To do that, Fu et al. (2017) proposed to use a cosine similarity based metric using pretrained word embeddings. First, a sentence embedding is computed by concatenation of max, mean, and average pooling over the timesteps: ",
+        "bbox": [
+            173,
+            556,
+            825,
+            654
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "img_path": "images/aa6b3cce295505e9e1362bbb90560bbaca4d0e390facec904ab9a870af5729a0.jpg",
+        "text": "$$\n\\pmb { v } = [ \\operatorname* { m a x } ( \\pmb { v } _ { 1 } , \\dots , \\pmb { v } _ { T } ) ; \\operatorname* { m i n } ( \\pmb { v } _ { 1 } , \\dots , \\pmb { v } _ { T } ) ; \\operatorname* { m e a n } ( \\pmb { v } _ { 1 } , \\dots , \\pmb { v } _ { T } ) ]\n$$",
+        "text_format": "latex",
+        "bbox": [
+            294,
+            652,
+            704,
+            670
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Next, the cosine similarity score $s _ { i }$ between the embedding $\\pmb { v } _ { i } ^ { s }$ of the original source sentence and the target sentence with the changed form $\\mathbf { \\Delta } \\mathbf { \\vec { v } } _ { i } ^ { t }$ is computed, and the scores across the dataset are averaged to obtain the total score: ",
+        "bbox": [
+            174,
+            680,
+            825,
+            723
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "img_path": "images/26568db59e4c4e1fd47436a124b10e742b95274a340c236bb4c228029957fc3c.jpg",
+        "text": "$$\n\\mathbf { s } = \\frac { 1 } { 2 } \\left[ \\frac { 1 } { | X ^ { a } | } \\sum _ { i = 1 } ^ { | X ^ { a } | } \\mathrm { s } _ { i } + \\frac { 1 } { | X ^ { b } | } \\sum _ { i = 1 } ^ { | X ^ { b } | } \\mathrm { s } _ { i } \\right]\n$$",
+        "text_format": "latex",
+        "bbox": [
+            372,
+            729,
+            624,
+            781
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1.1 CONTINUOUS FORM ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            794,
+            369,
+            809
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The metrics described above treat the form as a categorical (in most cases, even binary) variable. This was not a problem in previous work since the change of form could be done by just inverting the form vector. Our work, in contrast, treats the form as a continuous variable, and, therefore, we cannot just use the proposed metrics directly. To enable a fair comparison, we propose the following procedure. ",
+        "bbox": [
+            174,
+            818,
+            825,
+            888
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For each sentence $s _ { s } ^ { a }$ in the test set from the corpus $X ^ { a }$ we sample $k = 1 0$ random sentence from the corpus $X ^ { b }$ of the opposite form. After that, we encode them into the meaning $m _ { i }$ and form $f _ { i }$ vectors, and average the form vectors to obtain a single form vector $\\begin{array} { r } { { f } _ { \\mathrm { a v g } } = \\frac { 1 } { k } \\sum _ { i = 1 } ^ { k } { f } _ { i } } \\end{array}$ . We then generate a new sentence with its original meaning vector $m _ { s }$ and the resulting form vector $\\pmb { f } _ { \\mathrm { a v g } }$ , and use it for evalation. This process enables a fair comparison with the previous works that treat form as a binary variable. ",
+        "bbox": [
+            173,
+            895,
+            823,
+            924
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            101,
+            825,
+            160
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 DATASETS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            176,
+            287,
+            190
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We performed an extensive evaluation of the proposed method on several dataset that reflect different changes of meaning, form, or specific aspects of meaning, such as sentiment polarity. ",
+        "bbox": [
+            173,
+            202,
+            823,
+            231
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Changing form: register This experiment is conducted with a dataset of titles of scientific papers and news articles published by Fu et al. (2017). This dataset (referred to as “Headlines”) contains titles of scientific articles crawled from online digital libraries, such as “ACM Digital Library” and “arXiv”. The titles of the news articles are taken from the “News Aggregator Data Set” from UCI Machine Learning Repository (Dheeru & Karra Taniskidou, 2017) ",
+        "bbox": [
+            174,
+            246,
+            825,
+            315
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Changing form: language diachrony Diachronic language change is explored with the dataset composed by Xu et al. (2012). It includes the texts of 17 plays by William Shakespeare in the original Early Modern English, and their translations into contemporary English. We randomly permuted all sentences from all plays and sampled the training, validation, and test sets. Note that this is the smallest dataset in our experiments. ",
+        "bbox": [
+            174,
+            330,
+            825,
+            400
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Previous work on style transfer for text also included the experiments with changing sentiment polarity (Shen et al., 2017; Fu et al., 2017). We do not report the experiments with sentiment data, since the change in sentiment polarity corresponds to a change in a specific aspect of meaning, rather than form. We therefore believe the comparison with these data would not be instructive. ",
+        "bbox": [
+            174,
+            406,
+            825,
+            463
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 RESULTS AND DISCUSSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            483,
+            429,
+            498
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/345122e7aea7fa589082eadec03c850ada45c1d27c374c52dc7d77b4925102de.jpg",
+        "image_caption": [
+            "Figure 2: Transfer strength vs Content preservation (see subsection 5.1) for different combination of the size of the meaning and form vectors. Each point is labeled with ”¡meaning vector size¿, ¡form vector size¿”. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            209,
+            525,
+            787,
+            699
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Probably, the most recent and similar to our work is the model proposed by Fu et al. (2017), in particular the “style-embedding” model. We implemented this model to provide a baseline for comparison. ",
+        "bbox": [
+            173,
+            770,
+            825,
+            811
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The classifier used in the transfer strength metric achieves very high accuracy (0.832 and 0.99 for the Shakespeare and Headlines datasets correspondingly). These results concur with the results of Shen et al. (2017) and Fu et al. (2017), and show that the two forms in the corpora are significantly different. ",
+        "bbox": [
+            174,
+            819,
+            825,
+            875
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Following Fu et al. (2017), we show the result of different configuration of the size of the form and meaning vectors on Figure 2. Namely, we report combinations of 64 and 256-dimensional vectors. Note that the sizes of the form vector are important. The larger is the form vector, the higher is the transfer strength, but smaller is content preservation. This is consistent with Fu et al. (2017), where they observed a similar behaviour. ",
+        "bbox": [
+            176,
+            881,
+            823,
+            924
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            103,
+            823,
+            132
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It is clear that the proposed method achieves significantly better transfer strength then the previously proposed model. It also has a lower content preservation score, which means that it repeats fewer exact words from the source sentence. Note that a low transfer strength and very high (0.9) content ˜ preservation score means that the model was not able to successfully learn to transfer the form and the target sentence is almost identical to the source sentence. The Shakespeare dataset is the hardest for the model in terms of transfer strength, probably because it is the smallest dataset, but the proposed method performs consistently well in transfer of both form and meaning and, in contrast to the baseline. ",
+        "bbox": [
+            174,
+            138,
+            825,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Fluency of generated sentences Note that there is no guarantee that the generated sentences would be coherent after switching the form vector. In order to estimate how this switch affects the fluency of generated sentences, we trained a language model on the Shakespeare dataset and calculated the perplexity of the generated sentences using the original form vector and the average of form vectors of $k$ random sentences from the opposite style (see subsubsection 5.1.1). While the perplexity of such sentences does go up, this change is not big (6.89 vs 9.74). ",
+        "bbox": [
+            174,
+            267,
+            825,
+            351
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.1 IMPACT OF THE MOTIVATIONAL TRAINING ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            371,
+            506,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate the impact of the motivator, we visualized form and meaning embeddings of 1000 random samples from the Headlines dataset using t-SNE algorithm (Van Der Maaten, 2014) with the Multicore-TSNE library (Ulyanov, 2016). The result is presented in Figure 3. ",
+        "bbox": [
+            174,
+            396,
+            825,
+            438
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "There are three important observations. First, there is no clear separation in the meaning embeddings, which means that any accurate form transfer is due to the form embeddings, and the dissociation of form and meaning was successful. ",
+        "bbox": [
+            174,
+            445,
+            823,
+            487
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Second, even without the motivator the model is able to produce the form embeddings that are clustered into two group. Recall from section 4 that without the motivational loss there are no forces that influence the form embeddings, but nevertheless the model learns to separate them. ",
+        "bbox": [
+            174,
+            494,
+            825,
+            536
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "However, the separation effect is much more pronounced in the presence of motivator. This explains why the motivator consistently improved transfer strength of ADNet, as shown in Figure 2. ",
+        "bbox": [
+            174,
+            544,
+            825,
+            573
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/177a04829132ea7eb43a11912ab21cf3b18fabd8252eb4e7e976664a6e7b58b0.jpg",
+        "image_caption": [
+            "Figure 3: t-SNE visualization of the form and meaning embeddings of 1000 random sentences. Green point represent sentences form news headlines, and red points represent titles of scientific articles. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            186,
+            589,
+            813,
+            731
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2 QUALITATIVE EVALUATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            814,
+            403,
+            828
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 1 and Table 2 show several examples of the successful form/meaning transfer achieved by ADNet. Table 1 presentes the results of an experiment that to some extent replicates the approach taken by the authors who treat linguistic form as a binary variable (Shen et al., 2017; Fu et al., 2017). The sentences the original Shakespeare plays were averaged to get the “typical” Early Modern English form vector. This averaged vector was used to decode a sentence from the modern English translation back into the original. The same was done in the opposite direction. ",
+        "bbox": [
+            174,
+            840,
+            825,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/48fab46f9e741a871d5a244a76c9f4e66d510443109411ad64023b2ea1d9ac5a.jpg",
+        "table_caption": [
+            "Table 1: Decoding of the source sentence from Early Modern English (EME) into contemporary English (CE), and vice versa. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Aye,sir. (EME)</td><td>→ Yes,sir. (CE)</td></tr><tr><td>Fare thee well, my lord (EME)</td><td>Fare you well, my lord (CE)</td></tr><tr><td>This guy will tell us everything. (CE)</td><td>This man will tell us everything. (EME)</td></tr><tr><td>Ive done no more to caesar than you will do to me.(CE)</td><td>Ihave done no more to caesar than,you shall do to me.(EME)</td></tr></table>",
+        "bbox": [
+            181,
+            101,
+            816,
+            150
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 2 illustrates the possibilities of ADNet on fine-grained transfer applied to the change of register. We encoded two sentences in different registers from the Headlines dataset to produce form and meaning embeddings, and then we decoded the first sentence with the meaning embedding of the second, and vice versa. As can be seen from Table 2, the model correctly captures the meaning of sentences and decodes them using the form of the source sentences. Note how the model preserves specific words and the structure of the source sentence. In particular, note how in the first example, the model decided to put the colon after the “crisis management”, as the source form sentence has this syntactic structure (“A review:”). This is not possible in the previously proposed models, as they treat form as just a binary variable. ",
+        "bbox": [
+            173,
+            218,
+            825,
+            344
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/189a9802191980e6625dd56d88134683cf1e22672a45fcbea58d6eadb8c9badc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>A review: detection techniques for LTE system Situation management knowledge from social media</td><td>→Crisis management: media practices in telecommunication management →Areview study against intelligence internet</td></tr><tr><td>Security flaw could notaffect digital devices,experts say Semantic approach to event processing</td><td>→Semantic approach approach: current multimedia networks as modeling processes →Security flaw to verifyleaks</td></tr></table>",
+        "bbox": [
+            181,
+            361,
+            816,
+            406
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Flipping the meaning and the form embeddings of two sentence from different registers. Note the use of the colon in the first example, and the use of the “to”-constructions in the second example, consistent with the form of the source sentences. ",
+        "bbox": [
+            174,
+            419,
+            823,
+            460
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6.3 PERFORMANCE OF MEANING EMBEDDINGS ON DOWNSTREAM TASKS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            494,
+            691,
+            507
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We conducted some experiments to test the assumption that the derived meaning embeddings should improve performance on downstream tasks that require understanding of the meaning of the sentences regardless of their form. We evaluated embeddings produced by the ADNet, trained in the Headlines dataset, on a task of paraphrase detection. We used the SentEval toolkit (Conneau et al., 2017) and the Microsoft Research Paraphrase Corpus (Dolan et al., 2004). The F1 scores on this task for different models are presented in Table 3. Note that all models, except InferSent, are unsupervised. The InferSent model was trained on a big SNLI dataset, consisting of more than 500,000 manually annotated pairs. ADNet achieves the the highest score among the unsupervised systems and outperforms the regular sequence-to-sequence autoencoder with a large gap. ",
+        "bbox": [
+            173,
+            520,
+            825,
+            645
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/ab266bb82c389d88561ac65cd20daea25d30d98240a874e61f8c0cf0bba5b08d.jpg",
+        "table_caption": [
+            "Table 3: F1 scores on the task of paraphrase detection using the SentEval toolkit (Conneau et al., 2017) "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>BoW</td><td>Seq2Seq</td><td>InferSent</td><td>Fu et al. (2017)</td><td>ADNet</td></tr><tr><td>80.82</td><td>74.68</td><td>83.17</td><td>78.88</td><td>81.38</td></tr></table>",
+        "bbox": [
+            305,
+            661,
+            692,
+            705
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            779,
+            318,
+            795
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we presented ADNet, a new model that performs adversarial decomposition of text representation. In contrast to previous work, it does not require a parallel training corpus and works directly on hidden representations of sentences. Most importantly, is does not treat the form as a binary variable (as done in most previously proposed models), enabling a fine-grained change of the form of sentences or specific aspects of meaning. We evaluate ADNet on two tasks: the shift of language register and diachronic language change. Our solution achieves superior results, and t-SNE visualizations of the learned meaning and style embeddings illustrate that the proposed motivational loss leads to significantly better separation of the form embeddings. ",
+        "bbox": [
+            174,
+            811,
+            825,
+            924
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            103,
+            287,
+            117
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Samuel Albanie, Sebastien Ehrhardt, and Jo ´ ao F Henriques. Stopping gan violence: Generative ˜ unadversarial networks. arXiv preprint arXiv:1703.02528, 2017. ",
+        "bbox": [
+            174,
+            126,
+            823,
+            155
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Martin Arjovsky, Soumith Chintala, and Leon Bottou. Wasserstein gan. ´ arXiv preprint arXiv:1701.07875, 2017. ",
+        "bbox": [
+            173,
+            162,
+            820,
+            191
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Sanjeev Arora, Yingyu Liang, and Tengyu Ma. A simple but tough-to-beat baseline for sentence embeddings. In International Conference on Learning Representation, 2017. ",
+        "bbox": [
+            174,
+            199,
+            820,
+            229
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. ",
+        "bbox": [
+            174,
+            237,
+            823,
+            267
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Keith Carlson, Allen Riddell, and Daniel Rockmore. Zero-shot style transfer in text using recurrent neural networks. arXiv preprint arXiv:1711.04731, 2017. ",
+        "bbox": [
+            173,
+            275,
+            823,
+            304
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. Universal sentence encoder. arXiv preprint arXiv:1803.11175, 2018. ",
+        "bbox": [
+            174,
+            311,
+            823,
+            354
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. ",
+        "bbox": [
+            171,
+            363,
+            823,
+            393
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´ learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015. ",
+        "bbox": [
+            171,
+            400,
+            823,
+            430
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo¨ıc Barrault, and Antoine Bordes. Supervised learning of universal sentence representations from natural language inference data. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 670–680, 2017. ",
+        "bbox": [
+            174,
+            436,
+            823,
+            494
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ferdinand de Saussure. Course in General Linguistics. New York : Philosophical Library, 1959. URL http://archive.org/details/courseingenerall00saus. ",
+        "bbox": [
+            171,
+            502,
+            823,
+            534
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Dua Dheeru and Efi Karra Taniskidou. UCI machine learning repository, 2017. URL http: //archive.ics.uci.edu/ml. ",
+        "bbox": [
+            173,
+            540,
+            821,
+            570
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Bill Dolan, Chris Quirk, and Chris Brockett. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In Proceedings of the 20th international conference on Computational Linguistics, pp. 350. Association for Computational Linguistics, 2004. ",
+        "bbox": [
+            176,
+            578,
+            823,
+            622
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jessica Ficler and Yoav Goldberg. Controlling linguistic style aspects in neural language generation. In Proceedings of the Workshop on Stylistic Variation, pp. 94–104, 2017. ",
+        "bbox": [
+            169,
+            628,
+            820,
+            659
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. Style transfer in text: Exploration and evaluation. arXiv preprint arXiv:1711.06861, 2017. ",
+        "bbox": [
+            171,
+            666,
+            823,
+            695
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014. ",
+        "bbox": [
+            174,
+            704,
+            823,
+            747
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pp. 140–149, 2016. ",
+        "bbox": [
+            174,
+            755,
+            823,
+            799
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "M. A. K. Halliday, A. McIntosh, and P. Stevens. The Linguistic Sciences and Language Teaching. Longmans, Green and Co., London, 1968. ",
+        "bbox": [
+            176,
+            806,
+            821,
+            835
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. Toward controlled generation of text. In International Conference on Machine Learning, pp. 1587–1596, 2017. ",
+        "bbox": [
+            174,
+            843,
+            821,
+            873
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1125–1134, 2017. ",
+        "bbox": [
+            174,
+            882,
+            825,
+            924
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. Shakespearizing modern language using copy-enriched sequence-to-sequence models. arXiv preprint arXiv:1707.01161, 2017. ",
+        "bbox": [
+            173,
+            103,
+            823,
+            133
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Yoon Kim, Kelly Zhang, Alexander M Rush, Yann LeCun, et al. Adversarially regularized autoencoders for generating discrete structures. arXiv preprint arXiv:1706.04223, 2017. ",
+        "bbox": [
+            173,
+            140,
+            820,
+            170
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. ",
+        "bbox": [
+            173,
+            176,
+            823,
+            205
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances In Neural Information Processing Systems, pp. 4601–4609, 2016. ",
+        "bbox": [
+            176,
+            213,
+            823,
+            257
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jiwei Li, Michel Galley, Chris Brockett, Georgios Spithourakis, Jianfeng Gao, and Bill Dolan. A persona-based neural conversation model. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pp. 994–1003, 2016. ",
+        "bbox": [
+            173,
+            263,
+            825,
+            320
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Zachary C Lipton, Sharad Vikram, and Julian McAuley. Generative concatenative nets jointly learn to write and classify reviews. arXiv preprint arXiv:1511.03683, 2015. ",
+        "bbox": [
+            169,
+            328,
+            823,
+            358
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018. ",
+        "bbox": [
+            171,
+            366,
+            823,
+            395
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015. ",
+        "bbox": [
+            173,
+            401,
+            823,
+            431
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. ",
+        "bbox": [
+            173,
+            439,
+            823,
+            468
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jonas Mueller, David Gifford, and Tommi Jaakkola. Sequence to better sequence: continuous revision of combinatorial structures. In International Conference on Machine Learning, pp. 2536– 2544, 2017. ",
+        "bbox": [
+            173,
+            476,
+            821,
+            518
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Nikita Nangia, Adina Williams, Angeliki Lazaridou, and Samuel R Bowman. The repeval 2017 shared task: Multi-genre natural language inference with sentence representations. arXiv preprint arXiv:1707.08172, 2017. ",
+        "bbox": [
+            171,
+            526,
+            823,
+            569
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pp. 311–318. Association for Computational Linguistics, 2002. ",
+        "bbox": [
+            174,
+            577,
+            823,
+            621
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Peter Potash, Alexey Romanov, and Anna Rumshisky. Ghostwriter: using an lstm for automatic rap lyric generation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1919–1924, 2015. ",
+        "bbox": [
+            174,
+            627,
+            823,
+            671
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Peter Potash, Alexey Romanov, and Anna Rumshisky. Here’s my point: Joint pointer architecture for argument mining. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 1364–1373, 2017. ",
+        "bbox": [
+            174,
+            679,
+            821,
+            722
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "William Safire. On Language Gotcha! Gang Strikes Again, 1995. URL https://www.nytimes.com/1995/12/31/magazine/ on-language-gotcha-gang-strikes-again.html. ",
+        "bbox": [
+            176,
+            728,
+            821,
+            772
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. Style transfer from non-parallel text by cross-alignment. In Advances in Neural Information Processing Systems, pp. 6833–6844, 2017. ",
+        "bbox": [
+            174,
+            780,
+            823,
+            821
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Efstathios Stamatatos. Authorship attribution using text distortion. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, volume 1, pp. 1138–1149, 2017. ",
+        "bbox": [
+            178,
+            830,
+            823,
+            875
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. ",
+        "bbox": [
+            174,
+            881,
+            825,
+            924
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pp. 3104–3112, 2014.   \nJian Tang, Yifan Yang, Sam Carton, Ming Zhang, and Qiaozhu Mei. Context-aware natural language generation with recurrent neural networks. arXiv preprint arXiv:1611.09900, 2016.   \nDmitry Ulyanov. Multicore-tsne. https://github.com/DmitryUlyanov/ Multicore-TSNE, 2016.   \nLaurens Van Der Maaten. Accelerating t-sne using tree-based algorithms. Journal of machine learning research, 15(1):3221–3245, 2014.   \nAshish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp. 6000–6010, 2017.   \nOriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in Neural Information Processing Systems, pp. 2692–2700, 2015.   \nWei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and Colin Cherry. Paraphrasing for style. Proceedings of COLING 2012, pp. 2899–2914, 2012.   \nJun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2223–2232, 2017. ",
+        "bbox": [
+            169,
+            102,
+            826,
+            425
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A MULTIPLE FORMS AND STYLISTIC SIMILARITIES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            452,
+            612,
+            467
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/0e733eb345569ac00898384b916d68adac3771ae1b15c0710524723af5f8dfcc.jpg",
+        "image_caption": [
+            "Figure 4: t-SNE visualization of the form and meaning embeddings. Each color corresponds to a different author. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            258,
+            483,
+            743,
+            628
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In order to go beyond just two different forms, we experimented with training the model on a set of literature novels from six different authors from Project Gutenberg2 written in two different time periods. A t-SNE visualization of the resulting meaning and form embeddings is presented in Figure 4. Note how form embeddings create a six-pointed star. After further examination, we observed that common phrases (for example, “Good morning” or “Hello!”) were embedded into the center of the star, whereas the most specific sentences from a given author were placed into the rays of the star. In particular, some sentences included character names, thus further research is required to mitigate this problem. Stamatatos (2017) provides a promising direction for solving this. ",
+        "bbox": [
+            173,
+            686,
+            825,
+            799
+        ],
+        "page_idx": 11
+    }
+]

parse/train/SygXPaEYvH/SygXPaEYvH.md

@@ -0,0 +1,325 @@
+# VL-BERT: PRE-TRAINING OF GENERIC VISUALLINGUISTIC REPRESENTATIONS
+
+Weijie $\mathbf { S u } ^ { 1 , 2 * }$ , Xizhou $\mathbf { Z } \mathbf { h } \mathbf { u } ^ { 1 , 2 * }$ , Yue $\mathbf { C a o ^ { 2 } }$ , Bin $\mathbf { L i } ^ { 1 }$ , Lewei $\mathbf { L } \mathbf { u } ^ { 2 }$ , Furu Wei2, Jifeng Dai2†
+
+1University of Science and Technology of China   
+2Microsoft Research Asia   
+{jackroos,ezra0408}@mail.ustc.edu.cn, binli@ustc.edu.cn   
+{yuecao,lewlu,fuwei,jifdai}@microsoft.com
+
+# ABSTRACT
+
+We introduce a new pre-trainable generic representation for visual-linguistic tasks, called Visual-Linguistic BERT (VL-BERT for short). VL-BERT adopts the simple yet powerful Transformer model as the backbone, and extends it to take both visual and linguistic embedded features as input. In it, each element of the input is either of a word from the input sentence, or a region-of-interest (RoI) from the input image. It is designed to fit for most of the visual-linguistic downstream tasks. To better exploit the generic representation, we pre-train VL-BERT on the massive-scale Conceptual Captions dataset, together with text-only corpus. Extensive empirical analysis demonstrates that the pre-training procedure can better align the visual-linguistic clues and benefit the downstream tasks, such as visual commonsense reasoning, visual question answering and referring expression comprehension. It is worth noting that VL-BERT achieved the first place of single model on the leaderboard of the VCR benchmark. Code is released at https://github.com/jackroos/VL-BERT.
+
+# 1 INTRODUCTION
+
+Pre-training of generic feature representations applicable to a variety of tasks in a domain is a hallmark of the success of deep networks. Firstly in computer vision, backbone networks designed for and pre-trained on ImageNet (Deng et al., 2009) classification are found to be effective for improving numerous image recognition tasks. Recently in natural language processing (NLP), Transformer networks (Vaswani et al., 2017) pre-trained with “masked language model” (MLM) objective (Devlin et al., 2018) on large language corpus excel at a variety of NLP tasks.
+
+Meanwhile, for tasks at the intersection of vision and language, such as image captioning (Young et al., 2014; Chen et al., 2015; Sharma et al., 2018), visual question answering (VQA) (Antol et al., 2015; Johnson et al., 2017; Goyal et al., 2017; Hudson & Manning, 2019), visual commonsense reasoning (VCR) (Zellers et al., 2019; Gao et al., 2019), there lacks such pre-trained generic feature representations. The previous practice is to combine base networks pre-trained for image recognition and NLP respectively in a task-specific way. The task-specific model is directly finetuned for the specific target task, without any generic visual-linguistic pre-training. The task-specific model may well suffer from overfitting when the data for the target task is scarce. Also, due to the task-specific model design, it is difficult to benefit from pre-training, where the pre-training task may well be different from the target. There lacks a common ground for studying the feature design and pretraining of visual-linguistic tasks in general.
+
+In the various network architectures designed for different visual-linguistic tasks, a key goal is to effectively aggregate the multi-modal information in both the visual and linguistic domains. For example, to pick the right answer in the VQA task, the network should empower integrating linguistic information from the question and the answers, and aggregating visual information from the input image, together with aligning the linguistic meanings with the visual clues. Thus, we seek to derive generic representations that can effectively aggregate and align visual and linguistic information.
+
+In the meantime, we see the successful application of Transformer attention (Vaswani et al., 2017) in NLP, together with its MLM-based pre-training technique in BERT (Devlin et al., 2018). The attention module is powerful and flexible in aggregating and aligning word embedded features in sentences, while the pre-training in BERT further enhances the capability.
+
+Inspired by that, we developed VL-BERT, a pre-trainable generic representation for visual-linguistic tasks, as shown in Figure 1. The backbone of VL-BERT is of (multi-modal) Transformer attention module taking both visual and linguistic embedded features as input. In it, each element is either of a word from the input sentence, or a region-of-interest (RoI) from the input image, together with certain special elements to disambiguate different input formats. Each element can adaptively aggregate information from all the other elements according to the compatibility defined on their contents, positions, categories, and etc. The content features of a word / an RoI are domain specific (WordPiece embeddings (Wu et al., 2016) as word features, Fast R-CNN (Girshick, 2015) features for RoIs). By stacking multiple layers of multi-modal Transformer attention modules, the derived representation is of rich capability in aggregating and aligning visual-linguistic clues. And taskspecific branches can be added above for specific visual-linguistic tasks.
+
+To better exploit the generic representation, we pre-train VL-BERT at both large visual-linguistic corpus and text-only datasets1. The pre-training loss on the visual-linguistic corpus is incurred via predicting randomly masked words or RoIs. Such pre-training sharpens the capability of VL-BERT in aggregating and aligning visual-linguistic clues. While the loss on the text-only corpus is of the standard MLM loss in BERT, improving the generalization on long and complex sentences.
+
+Comprehensive empirical evidence demonstrates that the proposed VL-BERT achieves state-of-theart performance on various downstream visual-linguistic tasks, such as visual commonsense reasoning, visual question answering and referring expression comprehension. In particular, we achieved the first place of single model on the leaderboard of visual commonsense reasoning.
+
+# 2 RELATED WORK
+
+Pre-training for Computer Vision Prior to the era of deep networks, it is far from mature to share features among different tasks and to improve the features via pre-training. The models for various computer vision tasks are of too diverse design choices to derive a generic representation. With the success of AlexNet (Krizhevsky et al., 2012) in ImageNet (Deng et al., 2009) classification, we see the renaissance of convolutional neural networks (CNNs) in the vision community. Soon after that, researchers found that ImageNet pre-trained CNNs can serve well as generic feature representation for various downstream tasks (Donahue et al., 2014), such as object detection (Girshick et al., 2014), semantic segmentation (Long et al., 2015), instance segmentation (Hariharan et al., 2014). The improvement in backbone networks for ImageNet classification further improves the downstream tasks. Recently there are research works on directly training CNNs from scratch on massive-scale target datasets, without ImageNet pre-training (He et al., 2018). They achieved performance on par with those with ImageNet pre-training. While they also note that pre-training on a proper massive dataset is vital for improving performance on target tasks with scarce data.
+
+Pre-training for Natural Language Processing (NLP) It is interesting to note that the development of pre-training techniques in NLP lags quite behind computer vision. There are previous research works on improving word embedding (Mikolov et al., 2013; Pennington et al., 2014; Kiros et al., 2015), which is a low-level linguistic feature representation. On top of that, numerous diverse architectures are designed for various NLP tasks. In the milestone work of Transformers (Vaswani et al., 2017), the Transformer attention module is proposed as a generic building block for various NLP tasks. After that, a serious of approaches are proposed for pre-training the generic representation, mainly based on Transformers, such as GPT (Radford et al., 2018), BERT (Devlin et al., 2018), GPT-2 (Radford et al., 2019), XLNet (Yang et al., 2019), XLM (Lample & Conneau, 2019), and RoBERTa (Liu et al., 2019). Among them, BERT is perhaps the most popular one due to its simplicity and superior performance.
+
+Pre-training for Visual-Linguistic Tasks. The development course of models for visual-linguistic tasks is also quite similar to those in the computer vision and NLP communities. Previously, taskspecific models are designed, wherein the features derived from off-the-shelf computer vision and NLP models are combined in an ad-hoc way for specific tasks. Model training is performed on the dataset for the specific task only.
+
+VideoBERT (Sun et al., 2019b) is the first work seeking to conduct pre-training for visual-linguistic tasks. In it, video clips are processed by off-the-shelf networks for action recognition, and are assigned to different clusters (visual words) based on the derived features. The pre-training loss is incurred via predicting the cluster ids of masked video clips. Due to the abrupt clustering of the video clips, it losses considerable visual content information and hinders updating visual network parameters. In the following work of CBT (Sun et al., 2019a), such clustering mechanism is removed. Both works are applied on videos, which are of linear structure in the time dimension, same as sentences. It is highly desired to study at the well-established image-based visual-linguistic tasks.
+
+Concurrent to our work, multiple works released on Arxiv very recently also seek to derive a pretrainable generic representation for visual-linguistic tasks. Table 5 in Appendix compares among them. We briefly discuss some of these works here.
+
+In ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019), which are under review or just got accepted, the network architectures are of two single-modal networks applied on input sentences and images respectively, followed by a cross-modal Transformer combining information from the two sources. The attention pattern in the cross-modal Transformer is restricted, where the authors believe to improve the performance. The authors of ViLBERT claim that such two-stream design is superior than a single-stream unified model. Meanwhile, in the proposed VL-BERT, it is of a unified architecture based on Transformers without any restriction on the attention patterns. The visual and linguistic contents are fed as input to VL-BERT, wherein they interact early and freely. We found that our unified model of VL-BERT outperforms such two-stream designs.
+
+VisualBert (Li et al., 2019b), B2T2 (Alberti et al., 2019), and Unicoder-VL (Li et al., 2019a), which are of work in progress or under review, are also of unified single-stream architecture. The differences of these works are compared in Table 5. The concurrent emergency of these research works indicates the importance of deriving a generic pre-trainable representation for visual-linguistic tasks.
+
+In addition, there are three noticeable differences between VL-BERT and other concurrent works in pre-training. Their effects are validated in Section 4.3. (1) We found the task of Sentence-Image Relationship Prediction used in all of the other concurrent works (e.g., ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019)) is of no help in pre-training visual-linguistic representations. Thus such a task is not incorporated in VL-BERT. (2) We pre-train VL-BERT on both visual-linguistic and text-only datasets. We found such joint pre-training improves the generalization over long and complex sentences. (3) Improved tuning of the visual representation. In VL-BERT, the parameters of Fast R-CNN, deriving the visual features, are also updated. To avoid visual clue leakage in the pre-training task of Masked RoI Classification with Linguistic Clues, the masking operation is conducted on the input raw pixels, other than the feature maps produced by layers of convolution.
+
+# 3 VL-BERT
+
+# 3.1 REVISIT BERT MODEL
+
+Let $\boldsymbol { x } = \{ x _ { 1 } , . . . , x _ { N } \}$ be the input elements in BERT (Devlin et al., 2018), which are of embedded features encoding sentence words. They are processed by a multi-layer bidirectional Transformer (Vaswani et al., 2017), where the embedding features of each element are transformed layer-by-layer in the fashion of aggregating features from the other elements with adaptive attention weights. Let $x ^ { l } = \{ x _ { 1 } ^ { l } , . . . , x _ { N } ^ { l } \}$ be the features of the $l$ -th layer $\cdot x ^ { 0 }$ is set as the input $x$ ). The features of the $( l + 1 )$ -th layer, $x ^ { l + 1 }$ , is computed by
+
+$$
+\begin{array} { r l } & { \tilde { h } _ { i } ^ { l + 1 } = \displaystyle \sum _ { m = 1 } ^ { M } W _ { m } ^ { l + 1 } \Big \{ \sum _ { j = 1 } ^ { N } A _ { i , j } ^ { m } \cdot V _ { m } ^ { l + 1 } x _ { j } ^ { l } \Big \} } \\ & { { \displaystyle h _ { i } ^ { l + 1 } = \mathrm { L a y e r N o r m } ( x _ { i } ^ { l } + \tilde { h } _ { i } ^ { l + 1 } ) } } \\ & { \tilde { x } _ { i } ^ { l + 1 } = W _ { 2 } ^ { l + 1 } \cdot \mathrm { G E L U } ( W _ { 1 } ^ { l + 1 } h _ { i } ^ { l + 1 } + b _ { 1 } ^ { l + 1 } ) + b _ { 2 } ^ { l + 1 } } \\ & { x _ { i } ^ { l + 1 } = \mathrm { L a y e r N o r m } ( h _ { i } ^ { l + 1 } + \tilde { x } _ { i } ^ { l + 1 } ) } \end{array}
+$$
+
+where notes t $m$ in Eq. 1 indexes over the attention attention weights between elements eadsand and in t $A _ { i , j } ^ { m } \ \propto \ \exp [ ( Q _ { m } ^ { l + 1 } x _ { i } ^ { l } ) ^ { T } ( K _ { m } ^ { l + 1 } x _ { j } ^ { l } ) ]$ de- by $i$ $j$ $m$   
+$\textstyle \sum _ { j = 1 } ^ { N } A _ { i , j } ^ { m } \ = \ 1$ . $W _ { m } ^ { l + 1 }$ , $Q _ { m } ^ { l + 1 }$ , $K _ { m } ^ { l + 1 }$ and $V _ { m } ^ { l + 1 }$ are learnable weights for $m ^ { \mathrm { t h } }$ attention head, $W _ { 1 } ^ { l + 1 }$ , $W _ { 2 } ^ { l + 1 }$ and $b _ { 1 } ^ { l + 1 } , b _ { 2 } ^ { l + 1 }$ in Eq. 3 are learnable weights and biases, respectively. Note that, the operations in Eq. $1 \sim 4$ is irrelevant to the order of input sequence, i.e. the final BERT representation of permuted input is same as the final BERT representation of the original input after the same permutation. The position of an element in BERT is encoded in its own embedding features by sequence positional embedding. Thanks to such decoupled representation, the BERT model is flexible enough to be pre-trained and finetuned for a variety of NLP tasks.
+
+In BERT pre-training, the masked language modeling (MLM) task is introduced. The embedded features of a certain input word would be randomly masked out (the token embedding channels capturing the word content is replaced by a special [MASK] token). The BERT model is trained to predict the masked word from linguistic clues of all the other unmasked elements. As explained in Wang & Cho (2019), the overall MLM-based training of BERT is equivalent to optimizing the following joint probability distribution
+
+$$
+\log P ( x | \theta ) = \frac { 1 } { Z ( \theta ) } \sum _ { i = 1 } ^ { N } \log \phi _ { i } ( x | \theta ) ,
+$$
+
+where $\phi _ { i } ( x | \theta )$ is the potential function for the $i$ -th input element, with parameters $\theta$ , and $Z ( \theta )$ is the partition function. Each log-potential term $\log \phi _ { i } ( x )$ is defined as
+
+$$
+\log \phi _ { i } ( x | \theta ) = x _ { i } ^ { T } f _ { i } ( x _ { \backslash i } | \theta ) _ { i } ,
+$$
+
+where $f _ { i } ( x _ { \backslash i } | \theta )$ denotes the final output feature of BERT corresponding to the $i$ -th element for input $x _ { \backslash i }$ , where $x _ { \backslash i }$ is defined as $x _ { \backslash i } = \{ x _ { 1 } , . . . , x _ { i - 1 } , [ \mathbf { M A S K } ] , x _ { i + 1 } , . . . , x _ { N } \} .$ . The incurred MLM-based loss is as
+
+$$
+L _ { \mathrm { M L M } } ( \theta ) = - E _ { x \sim D , i \sim \{ 1 , \ldots , N \} } \log \phi _ { i } ( x ) ,
+$$
+
+where $x$ is a randomly sampled sentence from the training set $D$ , and $i$ is a randomly sampled location for masking words.
+
+The second pre-training task, Next Sentence Prediction, focuses on modeling the relationship between two sentences. Two sentences are sampled from the input document, and the model should predict whether the second sentence is the direct successor of the first. In BERT, the sampled two sentences are concatenated into one input sequence, with special elements [CLS] and [SEP] inserted prior to the first and the second sentences, respectively. A Sigmoid classifier is appended on the final output feature corresponding to the [CLS] element to make the prediction. Let $x$ be the input sequence, $t \in \{ 0 , 1 \}$ indicates the relationship between the two sentences. The loss function is defined as
+
+$$
+L _ { \mathrm { N S P } } ( \theta ) = - E _ { ( x , t ) \sim D } \big [ t \log ( g ( x _ { 0 } ^ { L } ) ) + ( 1 - t ) \log ( 1 - g ( x _ { 0 } ^ { L } ) ) \big ] ,
+$$
+
+where $x _ { 0 } ^ { L }$ is the final output feature of the [CLS] element (at the $L$ -th layer), and $g ( x _ { 0 } ^ { L } )$ is the classifier output.
+
+# 3.2 MODEL ARCHITECTURE
+
+Figure 1 illustrates the architecture of VL-BERT. Basically, it modifies the original BERT (Devlin et al., 2018) model by adding new elements to accommodate the visual contents, and a new type of visual feature embedding to the input feature embeddings. Similar to BERT, the backbone is of multi-layer bidirectional Transformer encoder (Vaswani et al., 2017), enabling dependency modeling among all the input elements. Different to BERT processing sentence words only, VL-BERT takes both visual and linguistic elements as input, which are of features defined on regions-of-interest (RoIs) in images and sub-words from input sentences, respectively. The RoIs can either be bounding boxes produced by object detectors, or be annotated ones in certain tasks.
+
+It is worth noting that the input formats vary for different visual-linguistic tasks (e.g., <Caption, Image $>$ for image captioning, and $<$ <Question, Answer, Image $>$ for VQA (Antol et al., 2015; Johnson et al., 2017; Goyal et al., 2017; Hudson & Manning, 2019) and VCR (Zellers et al., 2019; Gao et al., 2019)). But thanks to the unordered representation nature of Transformer attention (e.g., the position of a word in sentence is encoded by the positional embedding only, other than the order in the input sequence), a generic representation can be derived as long as the input elements and embedding features are properly designed. Three types of input elements are involved, namely, visual, linguistic, and special elements for disambiguating different input formats. The input sequence always starts with a special classification element ([CLS]), then goes on with linguistic elements, then follows up with visual elements, and ends with a special ending element ([END]). A special separation element ([SEP]) is inserted in between different sentences in the linguistic elements, and between the linguistic and visual elements. For each input element, its embedding feature is the summation of four types of embedding, namely, token embedding, visual feature embedding, segment embedding, and sequence position embedding. Among them, the visual feature embedding is newly introduced for capturing visual clues, while the other three embeddings follow the design in the original BERT paper.
+
+![](images/3540328e50418d41b6f05a0aa4b47e6155f02e42de19e4c5d81baa32c0b03377.jpg)  
+Figure 1: Architecture for pre-training VL-BERT. All the parameters in this architecture including VL-BERT and Fast R-CNN are jointly trained in both pre-training and fine-tuning phases.
+
+Token Embedding Following the practice in BERT, the linguistic words are embedded with WordPiece embeddings (Wu et al., 2016) with a 30,000 vocabulary. A special token is assigned to each special element. For the visual elements, a special [IMG] token is assigned for each one of them.
+
+Visual Feature Embedding We firstly describe visual appearance feature and visual geometry embedding separately, and then how to combine them to form the visual feature embedding.
+
+For the visual element corresponding to an RoI, the visual appearance feature is extracted by applying a Fast R-CNN (Girshick, 2015) detector (i.e., the detection branch in Faster R-CNN (Ren et al., 2015)), where the feature vector prior to the output layer of each RoI is utilized as the visual feature embedding (of 2048-d in paper). For the non-visual elements, the corresponding visual appearance features are of features extracted on the whole input image. They are obtained by applying Faster R-CNN on an RoI covering the whole input image.
+
+The visual geometry embedding is designed to inform VL-BERT the geometry location of each inelemand ge. Each RoI is characterized by a 4-d vector, as denote the coordinate of the top-left and bottom $\textstyle \left( { \frac { x _ { \mathrm { L T } } } { W } } , { \frac { y _ { \mathrm { L T } } } { H } } , { \frac { x _ { \mathrm { R B } } } { W } } , { \frac { h _ { \mathrm { R B } } } { H } } \right)$ , wherectively, $( x _ { \mathrm { L T } } , y _ { \mathrm { L T } } )$ $\left( x _ { \mathrm { R B } } , y _ { \mathrm { R B } } \right)$ and $W , H$ are of the width and height of the input image. Following the practice in Relation Networks (Hu et al., 2018), the 4-d vector is embedded into a high-dimensional representation (of 2048-d in paper) by computing sine and cosine functions of different wavelengths.
+
+The visual feature embedding is attached to each of the input elements, which is the output of a fully connected layer taking the concatenation of visual appearance feature and visual geometry embedding as input.
+
+Segment Embedding Three types of segment, $A , B , C$ , are defined to separate input elements from different sources, namely, $A$ and $B$ for the words from the first and second input sentence respectively, and $C$ for the RoIs from the input image. For example, for input format of $<$ Question, Answer, Image $>$ , $A$ denotes Question, $B$ denotes Answer, and $C$ denotes Image. For input format of <Caption, Image>, $A$ denotes Caption, and $C$ denotes Image. A learned segment embedding is added to every input element for indicating which segment it belongs to.
+
+Sequence Position Embedding A learnable sequence position embedding is added to every input element indicating its order in the input sequence, same as BERT. Because there is no natural order among input visual elements, any permutation of them in the input sequence should achieve the same result. Thus the sequence position embedding for all visual elements are the same.
+
+# 3.3 PRE-TRAINING VL-BERT
+
+The generic feature representation of VL-BERT enables us to pre-train it on massive-scale datasets, with properly designed pre-training tasks. We pre-train VL-BERT on both visual-linguistic and textonly datasets. Here we utilize the Conceptual Captions dataset (Sharma et al., 2018) as the visuallinguistic corpus. It contains around 3.3 million images annotated with captions, which are harvested from web data and processed through an automatic pipeline. The issue with the Conceptual Captions dataset is that the captions are mainly simple clauses, which are too short and simple for many downstream tasks. To avoid overfitting on such short and simple text scenario, we also pre-train VL-BERT on text-only corpus with long and complex sentences. We utilize the BooksCorpus (Zhu et al., 2015) and the English Wikipedia datasets, which are also utilized in pre-training BERT.
+
+In SGD training, in each mini-batch, samples are randomly drawn from both Conceptual Captions and BooksCorpus & English Wikipedia (at a ratio of 1:1). For a sample drawn from Conceptual Captions, the input format to VL-BERT is of $<$ Caption, Image $>$ , where the RoIs in the image are localized and categorized by a pre-trained Faster R-CNN object detector. Two pre-training tasks are exploited to incur loss, which are as follows.
+
+Task #1: Masked Language Modeling with Visual Clues This task is very similar to the Masked Language Modeling (MLM) task utilized in BERT. The key difference is that visual clues are incorporated in VL-BERT for capturing the dependencies among visual and linguistic contents. During pre-training, each word in the input sentence(s) is randomly masked (at a probability of $1 5 \%$ ). For the masked word, its token is replaced with a special token of [MASK]. The model is trained to predict the masked words, based on the unmasked words and the visual features. The task drives the network to not only model the dependencies in sentence words, but also to align the visual and linguistic contents. For example, in Figure 1 “kitten drinking from [MASK]”, without the input image, the masked word could be any containers, such as “bowl”, “spoon” and “bottle”. The representation should capture the correspondence of the word “bottle” and the corresponding RoIs in the image to make the right guess. During pre-training, the final output feature corresponding to the masked word is fed into a classifier over the whole vocabulary, driven by Softmax cross-entropy loss.
+
+Task #2: Masked RoI Classification with Linguistic Clues This is a dual task of Task #1. Each RoI in image is randomly masked out (with $15 \%$ probability), and the pre-training task is to predict the category label of the masked RoI from the other clues. To avoid any visual clue leakage from the visual feature embedding of other elements, the pixels laid in the masked RoI are set as zeros before applying Fast R-CNN. During pre-training, the final output feature corresponding to the masked RoI is fed into a classifier with Softmax cross-entropy loss for object category classification. The category label predicted by pre-trained Faster R-CNN is set as the ground-truth. An example is shown in Figure 1. The RoI corresponding to cat in image is masked out, and the corresponding category cannot be predicted from any visual clues. But with the input caption of “kitten drinking from bottle”, the model can infer the category by exploiting the linguistic clues.
+
+For a sample drawn from the BooksCorpus & English Wikipedia datasets, the input format to VLBERT degenerates to be $<$ Text, $\mathcal { O } >$ , where no visual information is involved. The “visual feature embedding” term in Figure 1 is a learnable embedding shared for all words. The training loss is from the standard task of Masked Language Modeling (MLM) as in BERT.
+
+In summary, the pre-training on visual-linguistic corpus improves the detailed alignment between visual and linguistic contents. Such detailed alignment is vital for many downstream tasks (for example, in Visual Grounding (Kazemzadeh et al., 2014), the model locates the most relevant object or region in an image based on a natural language query). While the pre-training on text-only corpus facilitates downstream tasks involving understanding of long and complex sentences.
+
+# 3.4 FINE-TUNING VL-BERT
+
+VL-BERT is designed to be a generic feature representation for various visual-linguistic tasks. It is relatively simple to finetune VL-BERT for various downstream tasks. We simply need to feed VLBERT with properly formatted input and output, and finetune all the network parameters end-to-end. For the input, the typical formats of $<$ <Caption, Image> and $<$ Question, Answer, Image> cover the majority visual-linguistic tasks. VL-BERT also supports more sentences and more images as long as appropriate segment embeddings are introduced to identify different input sources. At the output, typically, the final output feature of the [CLS] element is used for sentence-image-relation level prediction. The final output features of words or RoIs are for word-level or RoI-level prediction. In addition to the input and output format, task-specific loss functions and training strategies also need to be tuned. See Section 4.2 for the detailed design choices and settings.
+
+# 4 EXPERIMENT
+
+# 4.1 PRE-TRAINING
+
+As described in Section 3.3, we pre-train VL-BERT jointly on Conceptual Captions (Sharma et al., 2018) as visual-linguistic corpus, and BooksCorpus (Zhu et al., 2015) & English Wikipedia as textonly corpus. As VL-BERT is developed via adding new inputs capturing visual information to the original BERT model, we initialize the parameters to be the same as the original BERT described in (Devlin et al., 2018). VL-BERTBASE and VL-BERTLARGE denote models developed from the original BERTBASE and BERTLARGE models, respectively. The newly added parameters in VL-BERT are randomly initialized from a Gaussian distribution with mean of 0 and standard deviation of 0.02. Visual content embedding is produced by Faster $\mathrm { R { \mathrm { - } C N N + \mathrm { R e s N e t { - } 1 0 1 } } }$ , initialized from parameters pre-trained on Visual Genome (Krishna et al., 2017) for object detection (see BUTD (Anderson et al., 2018)).
+
+Prior to pre-training on Conceptual Captions, the pre-trained Faster R-CNN is applied to extract RoIs. Specifically, at most 100 RoIs with detection scores higher than 0.5 are selected for each image. At minimum, 10 RoIs are selected from one image, regardless of the detection score threshold. The detailed parameter settings are in Appendix.
+
+# 4.2 FINE-TUNING ON DOWNSTREAM TASKS
+
+The pre-trained VL-BERT model can be fine-tuned for various downstream visual-linguistic tasks, with simple modifications on the input format, output prediction, loss function and training strategy.
+
+4.2.1 VISUAL COMMONSENSE REASONING (VCR)   
+Table 1: Comparison to the state-of-the-art methods with single model on the VCR dataset. $\dagger$ indicates concurrent works.   
+
+<table><tr><td>Model</td><td>Q→A val</td><td>test</td><td>QA→R val</td><td>test</td><td>Q→AR val</td><td>test</td></tr><tr><td>R2C (Zellers et al., 2019)</td><td>63.8</td><td>65.1</td><td>67.2</td><td>67.3</td><td>43.1</td><td>44.0</td></tr><tr><td>ViLBERT (Lu et al.,2019)+</td><td>72.4</td><td>73.3</td><td>74.5</td><td>74.6</td><td>54.0</td><td>54.8</td></tr><tr><td>VisualBERT (Li et al.,2019b)†</td><td>70.8</td><td>71.6</td><td>73.2</td><td>73.2</td><td>52.2</td><td>52.4</td></tr><tr><td>B2T2 (Alberti et al.,2019)+</td><td>71.9</td><td>72.6</td><td>76.0</td><td>75.7</td><td>54.9</td><td>55.0</td></tr><tr><td>VL-BERTBAsE W/O pre-training</td><td>73.1</td><td>-</td><td>73.8</td><td>-</td><td>54.2</td><td>1</td></tr><tr><td>VL-BERTBASE</td><td>73.8</td><td>=</td><td>74.4</td><td>=</td><td>55.2</td><td>1</td></tr><tr><td>VL-BERTLARGE</td><td>75.5</td><td>75.8</td><td>77.9</td><td>78.4</td><td>58.9</td><td>59.7</td></tr></table>
+
+Visual Commonsense Reasoning (VCR) focuses on higher-order cognitive and commonsense understanding of the given image. In the dataset of Zellers et al. (2019), given an image and a list of categorized RoIs, a question at cognition level is raised. The model should pick the right answer to the question and provide the rationale explanation. For each question, there are 4 candidate answers and 4 candidate rationales. This holistic task $( \mathbf { Q } \to \mathbf { A R } )$ ) is decomposed into two sub-tasks wherein researchers can train specific individual models: question answering $( \mathbf { Q }  \mathbf { A } )$ ) and answer justification $( \mathrm { Q A } \to \mathrm { R }$ ). The released VCR dataset consists of $2 6 5 \mathrm { k }$ pairs of questions, answers, and rationales, over $1 0 0 \mathrm { k }$ unique movie scenes ( $1 0 0 \mathrm { k }$ images). They are split into training, validation, and test sets consisting of 213k questions and 80k images, 27k questions and 10k images, and $2 5 \mathrm { k }$ questions and $1 0 \mathrm { k }$ images, respectively.
+
+![](images/8d93389419b06a119b0e9a5c8dba2e7e7897e9be07d33e6073828e5814699bba.jpg)  
+Figure 2: Input and output formats for fine-tuning different visual-linguistic downstream tasks.
+
+Our experimental protocol for VCR follows that in R2C (Zellers et al., 2019). The model is trained on the train split, and is evaluated at the val and test sets. In the original work R2C, task-specific “Grounding”, “Contextualization” and “Reasoning” modules are designed. Here we simply adopt the generic representation of VL-BERT for the task. Figure 2 (a) illustrates the input format, <Question, Answer, Image>. For the sub-task of $Q  \mathbf { A }$ , ‘Q’ and ‘A’ are filled to the Question section and Answer section respectively. For the sub-task of $\mathrm { Q A } \to \mathrm { R }$ , the concatenation of $\mathbf { \bar { Q } } ^ { \prime }$ and ‘A’ is filled to the Question section, and ‘R’ is filled to the Answer section. The input RoIs to VL-BERT are the ground-truth annotations in the dataset. The final output feature of [CLS] element is fed to a Softmax classifier for predicting whether the given Answer is the correct choice. During fine-tuning, we adopt two losses, the classification over the correctness of the answers and the RoI classification with linguistic clues. The detailed parameter settings are in Appendix.
+
+Table 1 presents the experiment results. Pre-training VL-BERT improves the performance by $1 . 0 \%$ in the final $\mathrm { Q }  \mathrm { A R }$ task, which validates the effectiveness of pre-training. Compared with R2C, we do not use ad-hoc task-specific modules. Instead, we simply adopt the generic representation of VL-BERT and jointly train the whole model end-to-end. Despite the same input, output and experimental protocol as R2C, VL-BERT outperforms R2C by large margins, indicating the power of our simple cross-modal architecture. Compared with other concurrent works, i.e., ViLBERT, VisualBERT and B2T2, our VL-BERT achieves the state-of-the-art performance.
+
+# 4.2.2 VISUAL QUESTION ANSWERING (VQA)
+
+In the VQA task, given a natural image, a question at the perceptual level is asked, and the algorithm should generate / choose the correct answer. Here we conduct experiments on the widely-used VQA v2.0 dataset (Goyal et al., 2017), which is built based on the COCO (Lin et al., 2014) images. The VQA $\mathrm { { v } } 2 . 0$ dataset is split into train ( $8 3 \mathrm { k }$ images and 444k questions), validation (41k images and
+
+Table 2: Comparison to the state-of-the-art methods with single model on the VQA dataset. $\dagger$ indicates concurrent works.   
+
+<table><tr><td>Model</td><td>test-dev</td><td>test-std</td></tr><tr><td>BUTD (Anderson et al.,2018)</td><td>65.32</td><td>65.67</td></tr><tr><td>ViLBERT (Lu et al.,2019)†</td><td>70.55</td><td>70.92</td></tr><tr><td>VisualBERT (Li et al.,2019b)t LXMERT(Tan &amp; Bansal, 2019)†</td><td>70.80</td><td>71.00</td></tr><tr><td>VL-BERTBAsE W/o pre-training</td><td>72.42</td><td>72.54</td></tr><tr><td>VL-BERTBASE</td><td>69.58 71.16</td><td>1</td></tr><tr><td>VL-BERTLARGE</td><td></td><td>-</td></tr><tr><td></td><td>71.79</td><td>72.22</td></tr></table>
+
+$2 1 4 \mathrm { k }$ questions), and test (81k images and $4 4 8 \mathrm { k }$ questions) sets. Following the experimental protocol in BUTD (Anderson et al., 2018), for each question, the algorithm should pick the corresponding answer from a shared set consisting of 3,129 answers.
+
+Figure 2 (b) illustrates the input format for the VQA task, which is of $<$ <Question, Answer, Image>. As the possible answers are from a shared pool independent to the question, we only fill a [MASK] element to the Answer section. As in BUTD (Anderson et al., 2018), the input RoIs in VL-BERT are generated by a Faster R-CNN detector pre-trained on Visual Genome (Krishna et al., 2017). The answer prediction is made from a multi-class classifier based upon the output feature of the [MASK] element. During fine-tuning, the network training is driven by the multi-class cross-entropy loss over the possible answers. The detailed parameter settings are in Appendix.
+
+Table 2 presents our experimental results. Pre-training VL-BERT improves the performance by $1 . 6 \%$ , which validates the importance of pre-training. VL-BERT shares the same input (i.e., question, image, and RoIs), output and experimental protocol with BUTD, a prevalent model specifically designed for the task. Still, VL-BERT surpasses BUTD by over $5 \%$ in accuracy. Except for LXMERT, our VL-BERT achieves better performance than the other concurrent works. This is because LXMERT is pre-trained on massive visual question answering data (aggregating almost all the VQA datasets based on COCO and Visual Genome). While our model is only pre-trained on captioning and text-only dataset, where there is still gap with the VQA task.
+
+# 4.2.3 REFERRING EXPRESSION COMPREHENSION
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Ground-truth Regions</td><td colspan="3">Detected Regions</td></tr><tr><td>val</td><td>testA</td><td>testB</td><td>val</td><td>testA</td><td>testB</td></tr><tr><td>MAttNet (Yu et al.,2018)</td><td>71.01</td><td>75.13</td><td>66.17</td><td>65.33</td><td>71.62</td><td>56.02</td></tr><tr><td>ViLBERT (Lu et al.,2019)†</td><td>-</td><td>-</td><td>1</td><td>72.34</td><td>78.52</td><td>62.61</td></tr><tr><td>VL-BERTBAsE W/o pre-training</td><td>74.41</td><td>77.28</td><td>67.52</td><td>66.03</td><td>71.87</td><td>56.13</td></tr><tr><td>VL-BERTBASE</td><td>79.88</td><td>82.40</td><td>75.01</td><td>71.60</td><td>77.72</td><td>60.99</td></tr><tr><td>VL-BERTLARGE</td><td>80.31</td><td>83.62</td><td>75.45</td><td>72.59</td><td>78.57</td><td>62.30</td></tr></table>
+
+Table 3: Comparison to the state-of-the-art methods with single model on the RefCOCO $^ +$ dataset. $\dagger$ indicates concurrent work.
+
+A referring expression is a natural language phrase that refers to an object in an image. The referring expression comprehension task is to localize the object in an image with the given referring expression. We adopt the RefCOCO $^ +$ (Kazemzadeh et al., 2014) dataset for evaluation, consisting of 141k expressions for 50k referred objects in $2 0 \mathrm { k }$ images in the COCO dataset (Lin et al., 2014). The referring expressions in $\operatorname { R e f C O C O + }$ are forbidden from using absolute location words, e.g. left dog. Therefore the referring expressions focus on purely appearance-based descriptions. $\operatorname { R e f C O C O + }$ are split into four sets, training set (train), validation set (val), and two testing sets (testA and testB). Images containing multiple people are in testA set, while images containing multiple objects of other categories are in testB set. There is no overlap between the training, validation and testing images.
+
+Figure 2 (c) illustrates the input format for referring expression comprehension , where the input format is of $<$ <Query, Image>. Model training and evaluation are conducted either on the groundtruth RoIs or on the detected boxes in MAttNet (Yu et al., 2018). And the results are reported either in the track of ground-truth regions or that of detected regions, respectively. During training, we compute the classification scores for all the input RoIs. For each RoI, a binary classification loss is applied. During inference, we directly choose the RoI with the highest classification score as the referred object of the input referring expression. The detailed parameter settings are in Appendix.
+
+Table 3 presents our experimental results. Pre-trained VL-BERT significantly improves the performance. Compared with MAttNet, VL-BERT is much simpler without task-specific architecture designs, yet much better. VL-BERT achieves comparable performance with the concurrent work of ViLBERT.
+
+# 4.3 ABLATION STUDY
+
+Table 4: Ablation study for VL-BERTBASE with $0 . 5 \times$ fine-tuning epochs.   
+
+<table><tr><td>Settings</td><td>MaskedLanguage Modeling with Visual Clues</td><td>Masked RoI Classification with Linguistic Clues</td><td>Sentence-Image Relationship Prediction</td><td>with Text-only Corpus</td><td>Tuning Fast R-CNN</td><td>VCR Q→A val val</td><td>QA→R test-dev</td><td>VQA</td><td>RefCOCO+ Detected Regions val</td></tr><tr><td>w/o pre-training</td><td></td><td></td><td></td><td></td><td></td><td>72.9</td><td>73.0</td><td>69.5</td><td>62.7</td></tr><tr><td>(a)</td><td></td><td></td><td></td><td></td><td></td><td>72.9</td><td>73.1</td><td>71.0</td><td>69.1</td></tr><tr><td>(b)</td><td>√ √</td><td>√</td><td></td><td></td><td></td><td>73.0</td><td>73.1</td><td>71.1</td><td>70.7</td></tr><tr><td>(c)</td><td>√</td><td>√</td><td>√</td><td></td><td></td><td>72.2</td><td>72.4</td><td>70.3</td><td>69.5</td></tr><tr><td>(d)</td><td>√</td><td>√</td><td></td><td>√</td><td></td><td>73.4</td><td>73.8</td><td>71.1</td><td>70.7</td></tr><tr><td>VL-BERTBASE</td><td>√</td><td>√</td><td></td><td>√</td><td>√</td><td>73.8</td><td>73.9</td><td>71.2</td><td>71.1</td></tr></table>
+
+Table 4 ablates key design choices in pre-training VL-BERT. For experimental efficiency, the finetuning epoches of VL-BERT are of $0 . 5 \times$ of those in Section 4.2, with only VL-BERTBASE model.
+
+Overall, the pre-training of VL-BERT improves the performance over all the three down-stream tasks (by comparing setting “w/o pre-training” and VL-BERTBASE). The improvement amplitude varies for different tasks. By comparing setting (a) to that of “w/o pre-training”, we see the benefits of Task #1, Masked Language Modeling with Visual Clues. By further incorporating Task #2, Masked RoI Classification with Linguistic Clues, the accuracy further improves on $\operatorname { R e f C O C O + }$ , but gets stuck at VCR and VQA. This might be because only $\operatorname { R e f C O C O + }$ utilizes the final output feature corresponding to [IMG] tokens for prediction. Thus the pre-training of such features is beneficial. Setting (c) incorporates the task of Sentence-Image Relationship Prediction as in ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019). It would hurt accuracy on all the three downstream tasks. We guess the reason is because the task of Sentence-Image Relationship Prediction would introduce unmatched image and caption pairs as negative examples. Such unmatched samples would hamper the training of other tasks. Setting (d) adds text-only corpus during pre-training. Compared with setting (b), it improves the performance over all three down-stream tasks, and is most significant on VCR. This is because the task of VCR involves more complex and longer sentences than those in VQA and $_ \mathrm { R e f C O C O + } 2$ . By further finetuning the network parameters of Fast R-CNN, which generates the visual features, we get the final setting of VL-BERTBASE. Such end-to-end training of the entire network is helpful for all the downstream tasks.
+
+# 5 CONCLUSION
+
+In this paper, we developed VL-BERT, a new pre-trainable generic representation for visuallinguistic tasks. Instead of using ad-hoc task-specific modules, VL-BERT adopts the simple yet powerful Transformer model as the backbone. It is pre-trained on the massive-scale Conceptual Captions dataset, together with text-only corpus. Extensive empirical analysis demonstrates that the pre-training procedure can better align the visual-linguistic clues, and thus benefit the downstream tasks. In the future, we would like to seek better pre-training tasks, which could beneficial more downstream tasks (e.g., Image Caption Generation).
+
+# ACKNOWLEDGMENTS
+
+The work is partially supported by the National Natural Science Foundation of China under grand No.U19B2044 and No.61836011.
+
+# REFERENCES
+
+Chris Alberti, Jeffrey Ling, Michael Collins, and David Reitter. Fusion of detected objects in text for visual question answering. arXiv preprint arXiv:1908.05054, 2019.
+
+Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6077–6086, 2018.
+
+Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pp. 2425–2433, 2015.
+
+Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and ´ C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015.
+
+Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee, 2009.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
+
+Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In International conference on machine learning, pp. 647–655, 2014.
+
+Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. From two graphs to n questions: A vqa dataset for compositional reasoning on vision and commonsense. arXiv preprint arXiv:1908.02962, 2019.
+
+Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pp. 1440–1448, 2015.
+
+Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 580–587, 2014.
+
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6904–6913, 2017.
+
+Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Simultaneous detection and ´ segmentation. In European Conference on Computer Vision, pp. 297–312. Springer, 2014.
+
+Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imagenet pre-training. ´ arXiv preprint arXiv:1811.08883, 2018.
+
+Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3588–3597, 2018.
+
+Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6700–6709, 2019.
+
+Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2901–2910, 2017.
+
+Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 787–798, 2014.
+
+Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
+
+Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In Advances in neural information processing systems, pp. 3294–3302, 2015.
+
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32–73, 2017.
+
+Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.
+
+Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291, 2019.
+
+Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training, 2019a.
+
+Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019b.
+
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ European conference on computer vision, pp. 740–755. Springer, 2014.
+
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.
+
+Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, 2015.
+
+Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019.
+
+Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013.
+
+Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014.
+
+Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018.
+
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8), 2019.
+
+Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015.
+
+Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2556–2565, 2018.
+
+Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive bidirectional transformer for temporal representation learning. arXiv preprint arXiv:1906.05743, 2019a.
+
+Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. arXiv preprint arXiv:1904.01766, 2019b.
+
+Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019.
+
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017.
+
+Jesse Vig. A multiscale visualization of attention in the transformer model. arXiv preprint arXiv:1906.05714, 2019. URL https://arxiv.org/abs/1906.05714.
+
+Alex Wang and Kyunghyun Cho. Bert has a mouth, and it must speak: Bert as a markov random field language model. arXiv preprint arXiv:1902.04094, 2019.
+
+Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016.
+
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237, 2019.
+
+Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67–78, 2014.
+
+Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1307–1315, 2018.
+
+Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6720–6731, 2019.
+
+Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4995–5004, 2016.
+
+Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pp. 19–27, 2015.
+
+# A APPENDIX
+
+# A.1 COMPARISON AMONG VL-BERT AND OTHER WORKS
+
+Table 5 compares among VL-BERT and other concurrent works for pre-training generic visuallinguistic representations.
+
+<table><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>Method</td><td rowspan=1 colspan=1>Architecture</td><td rowspan=1 colspan=1>Visual Token</td><td rowspan=1 colspan=1>Pre-train Datasets</td><td rowspan=1 colspan=1>Pre-train Tasks</td><td rowspan=1 colspan=1>Downstream Tasks</td></tr><tr><td rowspan=1 colspan=1>PublishedWorks</td><td rowspan=1 colspan=1>VideoBERT(Sun et al., 2019b)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>video frame</td><td rowspan=1 colspan=1>Cooking312K(Sun et al., 2019b)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-words prediction</td><td rowspan=1 colspan=1>1) zero-shot action classification2) video captioning</td></tr><tr><td rowspan=7 colspan=1>WorksUnderReview /Just GotAccepted</td><td rowspan=1 colspan=1>CBT(Sun et al., 2019a)</td><td rowspan=1 colspan=1>two single-modal Transformer(vision &amp; language respectively)+ one cross-modal Transformer</td><td rowspan=1 colspan=1>video frame</td><td rowspan=1 colspan=1>Cooking312K(Sun et al., 2019b)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature regression</td><td rowspan=1 colspan=1>1) action anticipation2) video captioning</td></tr><tr><td rowspan=1 colspan=1>ViLBERT(Lu et al., 2019)</td><td rowspan=1 colspan=1>one single-modal Transformer(language)+ one cross-modal Transformer(with restricted attention pattern)</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al., 2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature classification</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) grounding referring expressions4) image retrieval5) zero-shot image retrieval</td></tr><tr><td rowspan=1 colspan=1>B2T2(Alberti et al., 2019)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al.,2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling</td><td rowspan=1 colspan=1>1) visual commonsense reasoning</td></tr><tr><td rowspan=1 colspan=1>LXMERT(Tan &amp; Bansal, 2019)</td><td rowspan=1 colspan=1>two single-modal Transformer(vision &amp; language respectively)+ one cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>COCO Caption+ VG Caption+ VG QA+VQA+GQA</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature classification4) masked visual-feature regression5) visual question answering</td><td rowspan=1 colspan=1>1) visual question answering2) natural language visual reasoning</td></tr><tr><td rowspan=1 colspan=1>VisualBERT(Li et al., 2019b)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>COCO Caption(Chen et al., 2015)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) natural language visual reasoning4) grounding phrases</td></tr><tr><td rowspan=1 colspan=1>Unicoder-VL(Li et al., 2019a)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al.,2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3)maskedvisual-featureclassification</td><td rowspan=1 colspan=1>1) image-text retrieval2)zero-shotimage-textretrieval</td></tr><tr><td rowspan=1 colspan=1>Our VL-BERT</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al., 2018)+ BooksCorpus(Zhu et al.,2015)+ English Wikipedia</td><td rowspan=1 colspan=1>1) masked language modeling2) masked visual-feature classification</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) grounding referring expressions</td></tr></table>
+
+‡ LXMERT is pre-trained on COCO Caption (Chen et al., 2015), VG Caption (Krishna et al., 2017), VG QA (Zhu et al., 2016), VQA (Antol et al., 2015) and GQA (Hudson & Manning, 2019).
+
+Table 5: Comparison among our VL-BERT and other works seeking to derive pre-trainable generic representations for visual-linguistic tasks.
+
+# A.2 DETAILED EXPERIMENT SETTINGS
+
+Pre-training is conducted on 16 Tesla V100 GPUs for $2 5 0 \mathrm { k }$ iterations by SGD. In each mini-batch, 256 samples are drawn. Among them, 128 samples are of ${ < } \mathbf { C }$ aption, Image $>$ pairs from Conceptual Captions, and the rest 128 samples are sequential tokens (at most 64 tokens for each sequence) from BooksCorpus & English Wikipedia. In SGD, Adam optimizer (Kingma & Ba, 2014) is applied, with base learning rate of $2 \times 1 0 ^ { - 5 }$ , $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 0 ^ { - 4 }$ , learning rate warmed up over the first 8,000 steps, and linear decay of the learning rate. All the parameters in VL-BERT and Fast R-CNN are jointly trained in both pre-training and fine-tuning phase. The visual feature input for textual corpus is a learnable embedding shared for all words. In the task of Masked RoI Classification with Linguistic Clues, the pixels lying in all the masked RoIs are set as zeros in the image. A box covering the whole image is added as a RoI and would not be masked.
+
+For VCR, the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each mini-batch, 256 triplets of <Question, Answer, Image $>$ are sampled. In SGD, the basic mini-batch gradient descent is conducted, with base learning rate of $5 \times \mathrm { 1 \bar { 0 } ^ { - 3 } }$ , momentum of 0.9, and weight decay of $1 0 ^ { - 4 }$ . The learning rate is linearly warmed up in the first 1,000 steps from an initial learning rate of 0, and is decayed by 0.1 at the 14-th and the 18-th epochs.
+
+For VQA, the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each mini-batch, 256 triplets of $<$ <Question, Answer, Image $>$ are sampled. In SGD, Adam optimizer is applied, with base learning rate of $1 \times 1 0 ^ { - 4 }$ , $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 0 ^ { - 4 }$ , learning rate warmed up over the first 2,000 steps, and linear decay of the learning rate.
+
+For $\operatorname { R e f C O C O + }$ , the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each minibatch, 256 pairs of $<$ <Query, Image $>$ are sampled. In SGD, Adam optimizer is applied, with base learning rate of $1 \times 1 0 ^ { - 4 }$ , $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 \dot { 0 } ^ { - 4 }$ , learning rate warmed up over the first 500 steps, and linear decay of the learning rate.
+
+![](images/27d26e5998a446d208137b24004ce833aead6fd5de0e5fa774c2524c34106ac3.jpg)  
+Figure 3: Visualization of attention maps in pre-trained VL-BERTBASE. Line intensity indicates the magnitude of attention probability with the text token as query and the image RoI as key. The intensity is affinely rescaled to set the maximum value as 1 and the minimum as 0, across different heads in each layer. The index of network layer and attention head is counted from 0.
+
+# A.3 VISUALIZATION OF ATTENTION MAPS IN VL-BERT
+
+To better understand what VL-BERT learns from pre-training, we visualized the attention maps of pre-trained VL-BERT (without fine-tuning on downstream tasks) using BertViz3(Vig, 2019).
+
+Some visualization results on COCO (Lin et al., 2014; Chen et al., 2015) val2017 set are shown in Figure 3. We can see different attention patterns across attention heads. For some attention heads, text tokens attend more on the associated image RoIs. While in some other heads, text tokens attend uniformly to all RoIs. It demonstrates the ability of VL-BERT in aggregating and aligning visuallinguistic contents.

parse/train/SygXPaEYvH/SygXPaEYvH_content_list.json

@@ -0,0 +1,1766 @@
+[
+    {
+        "type": "text",
+        "text": "VL-BERT: PRE-TRAINING OF GENERIC VISUALLINGUISTIC REPRESENTATIONS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            98,
+            821,
+            146
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Weijie $\\mathbf { S u } ^ { 1 , 2 * }$ , Xizhou $\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 1 , 2 * }$ , Yue $\\mathbf { C a o ^ { 2 } }$ , Bin $\\mathbf { L i } ^ { 1 }$ , Lewei $\\mathbf { L } \\mathbf { u } ^ { 2 }$ , Furu Wei2, Jifeng Dai2† ",
+        "bbox": [
+            184,
+            169,
+            779,
+            185
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1University of Science and Technology of China   \n2Microsoft Research Asia   \n{jackroos,ezra0408}@mail.ustc.edu.cn, binli@ustc.edu.cn   \n{yuecao,lewlu,fuwei,jifdai}@microsoft.com ",
+        "bbox": [
+            184,
+            186,
+            707,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            277,
+            544,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce a new pre-trainable generic representation for visual-linguistic tasks, called Visual-Linguistic BERT (VL-BERT for short). VL-BERT adopts the simple yet powerful Transformer model as the backbone, and extends it to take both visual and linguistic embedded features as input. In it, each element of the input is either of a word from the input sentence, or a region-of-interest (RoI) from the input image. It is designed to fit for most of the visual-linguistic downstream tasks. To better exploit the generic representation, we pre-train VL-BERT on the massive-scale Conceptual Captions dataset, together with text-only corpus. Extensive empirical analysis demonstrates that the pre-training procedure can better align the visual-linguistic clues and benefit the downstream tasks, such as visual commonsense reasoning, visual question answering and referring expression comprehension. It is worth noting that VL-BERT achieved the first place of single model on the leaderboard of the VCR benchmark. Code is released at https://github.com/jackroos/VL-BERT. ",
+        "bbox": [
+            233,
+            308,
+            764,
+            501
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            525,
+            336,
+            541
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Pre-training of generic feature representations applicable to a variety of tasks in a domain is a hallmark of the success of deep networks. Firstly in computer vision, backbone networks designed for and pre-trained on ImageNet (Deng et al., 2009) classification are found to be effective for improving numerous image recognition tasks. Recently in natural language processing (NLP), Transformer networks (Vaswani et al., 2017) pre-trained with “masked language model” (MLM) objective (Devlin et al., 2018) on large language corpus excel at a variety of NLP tasks. ",
+        "bbox": [
+            174,
+            556,
+            823,
+            640
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Meanwhile, for tasks at the intersection of vision and language, such as image captioning (Young et al., 2014; Chen et al., 2015; Sharma et al., 2018), visual question answering (VQA) (Antol et al., 2015; Johnson et al., 2017; Goyal et al., 2017; Hudson & Manning, 2019), visual commonsense reasoning (VCR) (Zellers et al., 2019; Gao et al., 2019), there lacks such pre-trained generic feature representations. The previous practice is to combine base networks pre-trained for image recognition and NLP respectively in a task-specific way. The task-specific model is directly finetuned for the specific target task, without any generic visual-linguistic pre-training. The task-specific model may well suffer from overfitting when the data for the target task is scarce. Also, due to the task-specific model design, it is difficult to benefit from pre-training, where the pre-training task may well be different from the target. There lacks a common ground for studying the feature design and pretraining of visual-linguistic tasks in general. ",
+        "bbox": [
+            174,
+            647,
+            825,
+            800
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the various network architectures designed for different visual-linguistic tasks, a key goal is to effectively aggregate the multi-modal information in both the visual and linguistic domains. For example, to pick the right answer in the VQA task, the network should empower integrating linguistic information from the question and the answers, and aggregating visual information from the input image, together with aligning the linguistic meanings with the visual clues. Thus, we seek to derive generic representations that can effectively aggregate and align visual and linguistic information. ",
+        "bbox": [
+            174,
+            806,
+            825,
+            890
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the meantime, we see the successful application of Transformer attention (Vaswani et al., 2017) in NLP, together with its MLM-based pre-training technique in BERT (Devlin et al., 2018). The attention module is powerful and flexible in aggregating and aligning word embedded features in sentences, while the pre-training in BERT further enhances the capability. ",
+        "bbox": [
+            174,
+            103,
+            823,
+            159
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Inspired by that, we developed VL-BERT, a pre-trainable generic representation for visual-linguistic tasks, as shown in Figure 1. The backbone of VL-BERT is of (multi-modal) Transformer attention module taking both visual and linguistic embedded features as input. In it, each element is either of a word from the input sentence, or a region-of-interest (RoI) from the input image, together with certain special elements to disambiguate different input formats. Each element can adaptively aggregate information from all the other elements according to the compatibility defined on their contents, positions, categories, and etc. The content features of a word / an RoI are domain specific (WordPiece embeddings (Wu et al., 2016) as word features, Fast R-CNN (Girshick, 2015) features for RoIs). By stacking multiple layers of multi-modal Transformer attention modules, the derived representation is of rich capability in aggregating and aligning visual-linguistic clues. And taskspecific branches can be added above for specific visual-linguistic tasks. ",
+        "bbox": [
+            174,
+            166,
+            825,
+            319
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To better exploit the generic representation, we pre-train VL-BERT at both large visual-linguistic corpus and text-only datasets1. The pre-training loss on the visual-linguistic corpus is incurred via predicting randomly masked words or RoIs. Such pre-training sharpens the capability of VL-BERT in aggregating and aligning visual-linguistic clues. While the loss on the text-only corpus is of the standard MLM loss in BERT, improving the generalization on long and complex sentences. ",
+        "bbox": [
+            174,
+            327,
+            823,
+            396
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Comprehensive empirical evidence demonstrates that the proposed VL-BERT achieves state-of-theart performance on various downstream visual-linguistic tasks, such as visual commonsense reasoning, visual question answering and referring expression comprehension. In particular, we achieved the first place of single model on the leaderboard of visual commonsense reasoning. ",
+        "bbox": [
+            176,
+            404,
+            823,
+            458
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 RELATED WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            479,
+            344,
+            496
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pre-training for Computer Vision Prior to the era of deep networks, it is far from mature to share features among different tasks and to improve the features via pre-training. The models for various computer vision tasks are of too diverse design choices to derive a generic representation. With the success of AlexNet (Krizhevsky et al., 2012) in ImageNet (Deng et al., 2009) classification, we see the renaissance of convolutional neural networks (CNNs) in the vision community. Soon after that, researchers found that ImageNet pre-trained CNNs can serve well as generic feature representation for various downstream tasks (Donahue et al., 2014), such as object detection (Girshick et al., 2014), semantic segmentation (Long et al., 2015), instance segmentation (Hariharan et al., 2014). The improvement in backbone networks for ImageNet classification further improves the downstream tasks. Recently there are research works on directly training CNNs from scratch on massive-scale target datasets, without ImageNet pre-training (He et al., 2018). They achieved performance on par with those with ImageNet pre-training. While they also note that pre-training on a proper massive dataset is vital for improving performance on target tasks with scarce data. ",
+        "bbox": [
+            173,
+            511,
+            825,
+            691
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pre-training for Natural Language Processing (NLP) It is interesting to note that the development of pre-training techniques in NLP lags quite behind computer vision. There are previous research works on improving word embedding (Mikolov et al., 2013; Pennington et al., 2014; Kiros et al., 2015), which is a low-level linguistic feature representation. On top of that, numerous diverse architectures are designed for various NLP tasks. In the milestone work of Transformers (Vaswani et al., 2017), the Transformer attention module is proposed as a generic building block for various NLP tasks. After that, a serious of approaches are proposed for pre-training the generic representation, mainly based on Transformers, such as GPT (Radford et al., 2018), BERT (Devlin et al., 2018), GPT-2 (Radford et al., 2019), XLNet (Yang et al., 2019), XLM (Lample & Conneau, 2019), and RoBERTa (Liu et al., 2019). Among them, BERT is perhaps the most popular one due to its simplicity and superior performance. ",
+        "bbox": [
+            173,
+            699,
+            825,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pre-training for Visual-Linguistic Tasks. The development course of models for visual-linguistic tasks is also quite similar to those in the computer vision and NLP communities. Previously, taskspecific models are designed, wherein the features derived from off-the-shelf computer vision and NLP models are combined in an ad-hoc way for specific tasks. Model training is performed on the dataset for the specific task only. ",
+        "bbox": [
+            174,
+            858,
+            821,
+            886
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            176,
+            103,
+            823,
+            146
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "VideoBERT (Sun et al., 2019b) is the first work seeking to conduct pre-training for visual-linguistic tasks. In it, video clips are processed by off-the-shelf networks for action recognition, and are assigned to different clusters (visual words) based on the derived features. The pre-training loss is incurred via predicting the cluster ids of masked video clips. Due to the abrupt clustering of the video clips, it losses considerable visual content information and hinders updating visual network parameters. In the following work of CBT (Sun et al., 2019a), such clustering mechanism is removed. Both works are applied on videos, which are of linear structure in the time dimension, same as sentences. It is highly desired to study at the well-established image-based visual-linguistic tasks. ",
+        "bbox": [
+            174,
+            152,
+            825,
+            263
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Concurrent to our work, multiple works released on Arxiv very recently also seek to derive a pretrainable generic representation for visual-linguistic tasks. Table 5 in Appendix compares among them. We briefly discuss some of these works here. ",
+        "bbox": [
+            174,
+            271,
+            825,
+            313
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019), which are under review or just got accepted, the network architectures are of two single-modal networks applied on input sentences and images respectively, followed by a cross-modal Transformer combining information from the two sources. The attention pattern in the cross-modal Transformer is restricted, where the authors believe to improve the performance. The authors of ViLBERT claim that such two-stream design is superior than a single-stream unified model. Meanwhile, in the proposed VL-BERT, it is of a unified architecture based on Transformers without any restriction on the attention patterns. The visual and linguistic contents are fed as input to VL-BERT, wherein they interact early and freely. We found that our unified model of VL-BERT outperforms such two-stream designs. ",
+        "bbox": [
+            174,
+            319,
+            825,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "VisualBert (Li et al., 2019b), B2T2 (Alberti et al., 2019), and Unicoder-VL (Li et al., 2019a), which are of work in progress or under review, are also of unified single-stream architecture. The differences of these works are compared in Table 5. The concurrent emergency of these research works indicates the importance of deriving a generic pre-trainable representation for visual-linguistic tasks. ",
+        "bbox": [
+            174,
+            452,
+            825,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In addition, there are three noticeable differences between VL-BERT and other concurrent works in pre-training. Their effects are validated in Section 4.3. (1) We found the task of Sentence-Image Relationship Prediction used in all of the other concurrent works (e.g., ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019)) is of no help in pre-training visual-linguistic representations. Thus such a task is not incorporated in VL-BERT. (2) We pre-train VL-BERT on both visual-linguistic and text-only datasets. We found such joint pre-training improves the generalization over long and complex sentences. (3) Improved tuning of the visual representation. In VL-BERT, the parameters of Fast R-CNN, deriving the visual features, are also updated. To avoid visual clue leakage in the pre-training task of Masked RoI Classification with Linguistic Clues, the masking operation is conducted on the input raw pixels, other than the feature maps produced by layers of convolution. ",
+        "bbox": [
+            174,
+            515,
+            825,
+            654
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 VL-BERT ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            672,
+            294,
+            689
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 REVISIT BERT MODEL ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            704,
+            377,
+            719
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Let $\\boldsymbol { x } = \\{ x _ { 1 } , . . . , x _ { N } \\}$ be the input elements in BERT (Devlin et al., 2018), which are of embedded features encoding sentence words. They are processed by a multi-layer bidirectional Transformer (Vaswani et al., 2017), where the embedding features of each element are transformed layer-by-layer in the fashion of aggregating features from the other elements with adaptive attention weights. Let $x ^ { l } = \\{ x _ { 1 } ^ { l } , . . . , x _ { N } ^ { l } \\}$ be the features of the $l$ -th layer $\\cdot x ^ { 0 }$ is set as the input $x$ ). The features of the $( l + 1 )$ -th layer, $x ^ { l + 1 }$ , is computed by ",
+        "bbox": [
+            174,
+            729,
+            823,
+            815
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/2e4a808109378150e7d0b6717120ffaa59711097c4679cd6591ea4d843a81b5f.jpg",
+        "text": "$$\n\\begin{array} { r l } & { \\tilde { h } _ { i } ^ { l + 1 } = \\displaystyle \\sum _ { m = 1 } ^ { M } W _ { m } ^ { l + 1 } \\Big \\{ \\sum _ { j = 1 } ^ { N } A _ { i , j } ^ { m } \\cdot V _ { m } ^ { l + 1 } x _ { j } ^ { l } \\Big \\} } \\\\ & { { \\displaystyle h _ { i } ^ { l + 1 } = \\mathrm { L a y e r N o r m } ( x _ { i } ^ { l } + \\tilde { h } _ { i } ^ { l + 1 } ) } } \\\\ & { \\tilde { x } _ { i } ^ { l + 1 } = W _ { 2 } ^ { l + 1 } \\cdot \\mathrm { G E L U } ( W _ { 1 } ^ { l + 1 } h _ { i } ^ { l + 1 } + b _ { 1 } ^ { l + 1 } ) + b _ { 2 } ^ { l + 1 } } \\\\ & { x _ { i } ^ { l + 1 } = \\mathrm { L a y e r N o r m } ( h _ { i } ^ { l + 1 } + \\tilde { x } _ { i } ^ { l + 1 } ) } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            232,
+            819,
+            563,
+            928
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where notes t $m$ in Eq. 1 indexes over the attention attention weights between elements eadsand and in t $A _ { i , j } ^ { m } \\ \\propto \\ \\exp [ ( Q _ { m } ^ { l + 1 } x _ { i } ^ { l } ) ^ { T } ( K _ { m } ^ { l + 1 } x _ { j } ^ { l } ) ]$ de- by $i$ $j$ $m$   \n$\\textstyle \\sum _ { j = 1 } ^ { N } A _ { i , j } ^ { m } \\ = \\ 1$ . $W _ { m } ^ { l + 1 }$ , $Q _ { m } ^ { l + 1 }$ , $K _ { m } ^ { l + 1 }$ and $V _ { m } ^ { l + 1 }$ are learnable weights for $m ^ { \\mathrm { t h } }$ attention head, $W _ { 1 } ^ { l + 1 }$ , $W _ { 2 } ^ { l + 1 }$ and $b _ { 1 } ^ { l + 1 } , b _ { 2 } ^ { l + 1 }$ in Eq. 3 are learnable weights and biases, respectively. Note that, the operations in Eq. $1 \\sim 4$ is irrelevant to the order of input sequence, i.e. the final BERT representation of permuted input is same as the final BERT representation of the original input after the same permutation. The position of an element in BERT is encoded in its own embedding features by sequence positional embedding. Thanks to such decoupled representation, the BERT model is flexible enough to be pre-trained and finetuned for a variety of NLP tasks. ",
+        "bbox": [
+            173,
+            102,
+            825,
+            236
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In BERT pre-training, the masked language modeling (MLM) task is introduced. The embedded features of a certain input word would be randomly masked out (the token embedding channels capturing the word content is replaced by a special [MASK] token). The BERT model is trained to predict the masked word from linguistic clues of all the other unmasked elements. As explained in Wang & Cho (2019), the overall MLM-based training of BERT is equivalent to optimizing the following joint probability distribution ",
+        "bbox": [
+            173,
+            242,
+            825,
+            327
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/30682f9d76d69ce90a4eeabbef027427b634882329b5562fdd326495360cdec5.jpg",
+        "text": "$$\n\\log P ( x | \\theta ) = \\frac { 1 } { Z ( \\theta ) } \\sum _ { i = 1 } ^ { N } \\log \\phi _ { i } ( x | \\theta ) ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            379,
+            333,
+            620,
+            377
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $\\phi _ { i } ( x | \\theta )$ is the potential function for the $i$ -th input element, with parameters $\\theta$ , and $Z ( \\theta )$ is the partition function. Each log-potential term $\\log \\phi _ { i } ( x )$ is defined as ",
+        "bbox": [
+            173,
+            383,
+            821,
+            412
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/28a662382c252e2d7d89ff5b9107134d9751453249fafeb60a412060f603c2c2.jpg",
+        "text": "$$\n\\log \\phi _ { i } ( x | \\theta ) = x _ { i } ^ { T } f _ { i } ( x _ { \\backslash i } | \\theta ) _ { i } ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            405,
+            420,
+            593,
+            439
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $f _ { i } ( x _ { \\backslash i } | \\theta )$ denotes the final output feature of BERT corresponding to the $i$ -th element for input $x _ { \\backslash i }$ , where $x _ { \\backslash i }$ is defined as $x _ { \\backslash i } = \\{ x _ { 1 } , . . . , x _ { i - 1 } , [ \\mathbf { M A S K } ] , x _ { i + 1 } , . . . , x _ { N } \\} .$ . The incurred MLM-based loss is as ",
+        "bbox": [
+            174,
+            445,
+            825,
+            489
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/863816be059da7c7082b3c75b103caed3b5488d837679796e60ec550f696d518.jpg",
+        "text": "$$\nL _ { \\mathrm { M L M } } ( \\theta ) = - E _ { x \\sim D , i \\sim \\{ 1 , \\ldots , N \\} } \\log \\phi _ { i } ( x ) ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            361,
+            488,
+            637,
+            507
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $x$ is a randomly sampled sentence from the training set $D$ , and $i$ is a randomly sampled location for masking words. ",
+        "bbox": [
+            173,
+            510,
+            825,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The second pre-training task, Next Sentence Prediction, focuses on modeling the relationship between two sentences. Two sentences are sampled from the input document, and the model should predict whether the second sentence is the direct successor of the first. In BERT, the sampled two sentences are concatenated into one input sequence, with special elements [CLS] and [SEP] inserted prior to the first and the second sentences, respectively. A Sigmoid classifier is appended on the final output feature corresponding to the [CLS] element to make the prediction. Let $x$ be the input sequence, $t \\in \\{ 0 , 1 \\}$ indicates the relationship between the two sentences. The loss function is defined as ",
+        "bbox": [
+            173,
+            544,
+            825,
+            655
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/d32f28293707e48357d3f6f38933a97f173f17b9f3def06d3b34c42da3fcabf4.jpg",
+        "text": "$$\nL _ { \\mathrm { N S P } } ( \\theta ) = - E _ { ( x , t ) \\sim D } \\big [ t \\log ( g ( x _ { 0 } ^ { L } ) ) + ( 1 - t ) \\log ( 1 - g ( x _ { 0 } ^ { L } ) ) \\big ] ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            285,
+            652,
+            710,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $x _ { 0 } ^ { L }$ is the final output feature of the [CLS] element (at the $L$ -th layer), and $g ( x _ { 0 } ^ { L } )$ is the classifier output. ",
+        "bbox": [
+            176,
+            678,
+            823,
+            707
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 MODEL ARCHITECTURE ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            723,
+            383,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 illustrates the architecture of VL-BERT. Basically, it modifies the original BERT (Devlin et al., 2018) model by adding new elements to accommodate the visual contents, and a new type of visual feature embedding to the input feature embeddings. Similar to BERT, the backbone is of multi-layer bidirectional Transformer encoder (Vaswani et al., 2017), enabling dependency modeling among all the input elements. Different to BERT processing sentence words only, VL-BERT takes both visual and linguistic elements as input, which are of features defined on regions-of-interest (RoIs) in images and sub-words from input sentences, respectively. The RoIs can either be bounding boxes produced by object detectors, or be annotated ones in certain tasks. ",
+        "bbox": [
+            173,
+            748,
+            825,
+            861
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is worth noting that the input formats vary for different visual-linguistic tasks (e.g., <Caption, Image $>$ for image captioning, and $<$ <Question, Answer, Image $>$ for VQA (Antol et al., 2015; Johnson et al., 2017; Goyal et al., 2017; Hudson & Manning, 2019) and VCR (Zellers et al., 2019; Gao et al., 2019)). But thanks to the unordered representation nature of Transformer attention (e.g., the position of a word in sentence is encoded by the positional embedding only, other than the order in the input sequence), a generic representation can be derived as long as the input elements and embedding features are properly designed. Three types of input elements are involved, namely, visual, linguistic, and special elements for disambiguating different input formats. The input sequence always starts with a special classification element ([CLS]), then goes on with linguistic elements, then follows up with visual elements, and ends with a special ending element ([END]). A special separation element ([SEP]) is inserted in between different sentences in the linguistic elements, and between the linguistic and visual elements. For each input element, its embedding feature is the summation of four types of embedding, namely, token embedding, visual feature embedding, segment embedding, and sequence position embedding. Among them, the visual feature embedding is newly introduced for capturing visual clues, while the other three embeddings follow the design in the original BERT paper. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3540328e50418d41b6f05a0aa4b47e6155f02e42de19e4c5d81baa32c0b03377.jpg",
+        "image_caption": [
+            "Figure 1: Architecture for pre-training VL-BERT. All the parameters in this architecture including VL-BERT and Fast R-CNN are jointly trained in both pre-training and fine-tuning phases. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            99,
+            823,
+            314
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            378,
+            825,
+            545
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Token Embedding Following the practice in BERT, the linguistic words are embedded with WordPiece embeddings (Wu et al., 2016) with a 30,000 vocabulary. A special token is assigned to each special element. For the visual elements, a special [IMG] token is assigned for each one of them. ",
+        "bbox": [
+            176,
+            551,
+            821,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Visual Feature Embedding We firstly describe visual appearance feature and visual geometry embedding separately, and then how to combine them to form the visual feature embedding. ",
+        "bbox": [
+            173,
+            601,
+            820,
+            628
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the visual element corresponding to an RoI, the visual appearance feature is extracted by applying a Fast R-CNN (Girshick, 2015) detector (i.e., the detection branch in Faster R-CNN (Ren et al., 2015)), where the feature vector prior to the output layer of each RoI is utilized as the visual feature embedding (of 2048-d in paper). For the non-visual elements, the corresponding visual appearance features are of features extracted on the whole input image. They are obtained by applying Faster R-CNN on an RoI covering the whole input image. ",
+        "bbox": [
+            174,
+            636,
+            825,
+            719
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The visual geometry embedding is designed to inform VL-BERT the geometry location of each inelemand ge. Each RoI is characterized by a 4-d vector, as denote the coordinate of the top-left and bottom $\\textstyle \\left( { \\frac { x _ { \\mathrm { L T } } } { W } } , { \\frac { y _ { \\mathrm { L T } } } { H } } , { \\frac { x _ { \\mathrm { R B } } } { W } } , { \\frac { h _ { \\mathrm { R B } } } { H } } \\right)$ , wherectively, $( x _ { \\mathrm { L T } } , y _ { \\mathrm { L T } } )$ $\\left( x _ { \\mathrm { R B } } , y _ { \\mathrm { R B } } \\right)$ and $W , H$ are of the width and height of the input image. Following the practice in Relation Networks (Hu et al., 2018), the 4-d vector is embedded into a high-dimensional representation (of 2048-d in paper) by computing sine and cosine functions of different wavelengths. ",
+        "bbox": [
+            174,
+            727,
+            823,
+            813
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The visual feature embedding is attached to each of the input elements, which is the output of a fully connected layer taking the concatenation of visual appearance feature and visual geometry embedding as input. ",
+        "bbox": [
+            174,
+            819,
+            823,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Segment Embedding Three types of segment, $A , B , C$ , are defined to separate input elements from different sources, namely, $A$ and $B$ for the words from the first and second input sentence respectively, and $C$ for the RoIs from the input image. For example, for input format of $<$ Question, Answer, Image $>$ , $A$ denotes Question, $B$ denotes Answer, and $C$ denotes Image. For input format of <Caption, Image>, $A$ denotes Caption, and $C$ denotes Image. A learned segment embedding is added to every input element for indicating which segment it belongs to. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            103,
+            823,
+            132
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Sequence Position Embedding A learnable sequence position embedding is added to every input element indicating its order in the input sequence, same as BERT. Because there is no natural order among input visual elements, any permutation of them in the input sequence should achieve the same result. Thus the sequence position embedding for all visual elements are the same. ",
+        "bbox": [
+            174,
+            138,
+            825,
+            194
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3 PRE-TRAINING VL-BERT ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            229,
+            395,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The generic feature representation of VL-BERT enables us to pre-train it on massive-scale datasets, with properly designed pre-training tasks. We pre-train VL-BERT on both visual-linguistic and textonly datasets. Here we utilize the Conceptual Captions dataset (Sharma et al., 2018) as the visuallinguistic corpus. It contains around 3.3 million images annotated with captions, which are harvested from web data and processed through an automatic pipeline. The issue with the Conceptual Captions dataset is that the captions are mainly simple clauses, which are too short and simple for many downstream tasks. To avoid overfitting on such short and simple text scenario, we also pre-train VL-BERT on text-only corpus with long and complex sentences. We utilize the BooksCorpus (Zhu et al., 2015) and the English Wikipedia datasets, which are also utilized in pre-training BERT. ",
+        "bbox": [
+            174,
+            262,
+            825,
+            387
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In SGD training, in each mini-batch, samples are randomly drawn from both Conceptual Captions and BooksCorpus & English Wikipedia (at a ratio of 1:1). For a sample drawn from Conceptual Captions, the input format to VL-BERT is of $<$ Caption, Image $>$ , where the RoIs in the image are localized and categorized by a pre-trained Faster R-CNN object detector. Two pre-training tasks are exploited to incur loss, which are as follows. ",
+        "bbox": [
+            174,
+            395,
+            825,
+            464
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Task #1: Masked Language Modeling with Visual Clues This task is very similar to the Masked Language Modeling (MLM) task utilized in BERT. The key difference is that visual clues are incorporated in VL-BERT for capturing the dependencies among visual and linguistic contents. During pre-training, each word in the input sentence(s) is randomly masked (at a probability of $1 5 \\%$ ). For the masked word, its token is replaced with a special token of [MASK]. The model is trained to predict the masked words, based on the unmasked words and the visual features. The task drives the network to not only model the dependencies in sentence words, but also to align the visual and linguistic contents. For example, in Figure 1 “kitten drinking from [MASK]”, without the input image, the masked word could be any containers, such as “bowl”, “spoon” and “bottle”. The representation should capture the correspondence of the word “bottle” and the corresponding RoIs in the image to make the right guess. During pre-training, the final output feature corresponding to the masked word is fed into a classifier over the whole vocabulary, driven by Softmax cross-entropy loss. ",
+        "bbox": [
+            174,
+            470,
+            825,
+            637
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Task #2: Masked RoI Classification with Linguistic Clues This is a dual task of Task #1. Each RoI in image is randomly masked out (with $15 \\%$ probability), and the pre-training task is to predict the category label of the masked RoI from the other clues. To avoid any visual clue leakage from the visual feature embedding of other elements, the pixels laid in the masked RoI are set as zeros before applying Fast R-CNN. During pre-training, the final output feature corresponding to the masked RoI is fed into a classifier with Softmax cross-entropy loss for object category classification. The category label predicted by pre-trained Faster R-CNN is set as the ground-truth. An example is shown in Figure 1. The RoI corresponding to cat in image is masked out, and the corresponding category cannot be predicted from any visual clues. But with the input caption of “kitten drinking from bottle”, the model can infer the category by exploiting the linguistic clues. ",
+        "bbox": [
+            174,
+            645,
+            825,
+            784
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For a sample drawn from the BooksCorpus & English Wikipedia datasets, the input format to VLBERT degenerates to be $<$ Text, $\\mathcal { O } >$ , where no visual information is involved. The “visual feature embedding” term in Figure 1 is a learnable embedding shared for all words. The training loss is from the standard task of Masked Language Modeling (MLM) as in BERT. ",
+        "bbox": [
+            174,
+            790,
+            823,
+            847
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In summary, the pre-training on visual-linguistic corpus improves the detailed alignment between visual and linguistic contents. Such detailed alignment is vital for many downstream tasks (for example, in Visual Grounding (Kazemzadeh et al., 2014), the model locates the most relevant object or region in an image based on a natural language query). While the pre-training on text-only corpus facilitates downstream tasks involving understanding of long and complex sentences. ",
+        "bbox": [
+            174,
+            854,
+            823,
+            924
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4 FINE-TUNING VL-BERT ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            103,
+            387,
+            117
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "VL-BERT is designed to be a generic feature representation for various visual-linguistic tasks. It is relatively simple to finetune VL-BERT for various downstream tasks. We simply need to feed VLBERT with properly formatted input and output, and finetune all the network parameters end-to-end. For the input, the typical formats of $<$ <Caption, Image> and $<$ Question, Answer, Image> cover the majority visual-linguistic tasks. VL-BERT also supports more sentences and more images as long as appropriate segment embeddings are introduced to identify different input sources. At the output, typically, the final output feature of the [CLS] element is used for sentence-image-relation level prediction. The final output features of words or RoIs are for word-level or RoI-level prediction. In addition to the input and output format, task-specific loss functions and training strategies also need to be tuned. See Section 4.2 for the detailed design choices and settings. ",
+        "bbox": [
+            173,
+            130,
+            825,
+            268
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4 EXPERIMENT ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            289,
+            316,
+            304
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.1 PRE-TRAINING ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            320,
+            318,
+            333
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As described in Section 3.3, we pre-train VL-BERT jointly on Conceptual Captions (Sharma et al., 2018) as visual-linguistic corpus, and BooksCorpus (Zhu et al., 2015) & English Wikipedia as textonly corpus. As VL-BERT is developed via adding new inputs capturing visual information to the original BERT model, we initialize the parameters to be the same as the original BERT described in (Devlin et al., 2018). VL-BERTBASE and VL-BERTLARGE denote models developed from the original BERTBASE and BERTLARGE models, respectively. The newly added parameters in VL-BERT are randomly initialized from a Gaussian distribution with mean of 0 and standard deviation of 0.02. Visual content embedding is produced by Faster $\\mathrm { R { \\mathrm { - } C N N + \\mathrm { R e s N e t { - } 1 0 1 } } }$ , initialized from parameters pre-trained on Visual Genome (Krishna et al., 2017) for object detection (see BUTD (Anderson et al., 2018)). ",
+        "bbox": [
+            173,
+            345,
+            825,
+            484
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Prior to pre-training on Conceptual Captions, the pre-trained Faster R-CNN is applied to extract RoIs. Specifically, at most 100 RoIs with detection scores higher than 0.5 are selected for each image. At minimum, 10 RoIs are selected from one image, regardless of the detection score threshold. The detailed parameter settings are in Appendix. ",
+        "bbox": [
+            174,
+            491,
+            825,
+            547
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 FINE-TUNING ON DOWNSTREAM TASKS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            564,
+            491,
+            579
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The pre-trained VL-BERT model can be fine-tuned for various downstream visual-linguistic tasks, with simple modifications on the input format, output prediction, loss function and training strategy. ",
+        "bbox": [
+            169,
+            590,
+            823,
+            619
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/727292a804b0fe1aa47e2567ee471ff29b56dcad4bbf30fbba66a15e829d02cc.jpg",
+        "table_caption": [
+            "4.2.1 VISUAL COMMONSENSE REASONING (VCR) ",
+            "Table 1: Comparison to the state-of-the-art methods with single model on the VCR dataset. $\\dagger$ indicates concurrent works. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Q→A val</td><td>test</td><td>QA→R val</td><td>test</td><td>Q→AR val</td><td>test</td></tr><tr><td>R2C (Zellers et al., 2019)</td><td>63.8</td><td>65.1</td><td>67.2</td><td>67.3</td><td>43.1</td><td>44.0</td></tr><tr><td>ViLBERT (Lu et al.,2019)+</td><td>72.4</td><td>73.3</td><td>74.5</td><td>74.6</td><td>54.0</td><td>54.8</td></tr><tr><td>VisualBERT (Li et al.,2019b)†</td><td>70.8</td><td>71.6</td><td>73.2</td><td>73.2</td><td>52.2</td><td>52.4</td></tr><tr><td>B2T2 (Alberti et al.,2019)+</td><td>71.9</td><td>72.6</td><td>76.0</td><td>75.7</td><td>54.9</td><td>55.0</td></tr><tr><td>VL-BERTBAsE W/O pre-training</td><td>73.1</td><td>-</td><td>73.8</td><td>-</td><td>54.2</td><td>1</td></tr><tr><td>VL-BERTBASE</td><td>73.8</td><td>=</td><td>74.4</td><td>=</td><td>55.2</td><td>1</td></tr><tr><td>VL-BERTLARGE</td><td>75.5</td><td>75.8</td><td>77.9</td><td>78.4</td><td>58.9</td><td>59.7</td></tr></table>",
+        "bbox": [
+            258,
+            664,
+            740,
+            787
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Visual Commonsense Reasoning (VCR) focuses on higher-order cognitive and commonsense understanding of the given image. In the dataset of Zellers et al. (2019), given an image and a list of categorized RoIs, a question at cognition level is raised. The model should pick the right answer to the question and provide the rationale explanation. For each question, there are 4 candidate answers and 4 candidate rationales. This holistic task $( \\mathbf { Q } \\to \\mathbf { A R } )$ ) is decomposed into two sub-tasks wherein researchers can train specific individual models: question answering $( \\mathbf { Q }  \\mathbf { A } )$ ) and answer justification $( \\mathrm { Q A } \\to \\mathrm { R }$ ). The released VCR dataset consists of $2 6 5 \\mathrm { k }$ pairs of questions, answers, and rationales, over $1 0 0 \\mathrm { k }$ unique movie scenes ( $1 0 0 \\mathrm { k }$ images). They are split into training, validation, and test sets consisting of 213k questions and 80k images, 27k questions and 10k images, and $2 5 \\mathrm { k }$ questions and $1 0 \\mathrm { k }$ images, respectively. ",
+        "bbox": [
+            174,
+            839,
+            825,
+            924
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8d93389419b06a119b0e9a5c8dba2e7e7897e9be07d33e6073828e5814699bba.jpg",
+        "image_caption": [
+            "Figure 2: Input and output formats for fine-tuning different visual-linguistic downstream tasks. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            253,
+            102,
+            759,
+            450
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            502,
+            825,
+            559
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our experimental protocol for VCR follows that in R2C (Zellers et al., 2019). The model is trained on the train split, and is evaluated at the val and test sets. In the original work R2C, task-specific “Grounding”, “Contextualization” and “Reasoning” modules are designed. Here we simply adopt the generic representation of VL-BERT for the task. Figure 2 (a) illustrates the input format, <Question, Answer, Image>. For the sub-task of $Q  \\mathbf { A }$ , ‘Q’ and ‘A’ are filled to the Question section and Answer section respectively. For the sub-task of $\\mathrm { Q A } \\to \\mathrm { R }$ , the concatenation of $\\mathbf { \\bar { Q } } ^ { \\prime }$ and ‘A’ is filled to the Question section, and ‘R’ is filled to the Answer section. The input RoIs to VL-BERT are the ground-truth annotations in the dataset. The final output feature of [CLS] element is fed to a Softmax classifier for predicting whether the given Answer is the correct choice. During fine-tuning, we adopt two losses, the classification over the correctness of the answers and the RoI classification with linguistic clues. The detailed parameter settings are in Appendix. ",
+        "bbox": [
+            174,
+            565,
+            825,
+            718
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 1 presents the experiment results. Pre-training VL-BERT improves the performance by $1 . 0 \\%$ in the final $\\mathrm { Q }  \\mathrm { A R }$ task, which validates the effectiveness of pre-training. Compared with R2C, we do not use ad-hoc task-specific modules. Instead, we simply adopt the generic representation of VL-BERT and jointly train the whole model end-to-end. Despite the same input, output and experimental protocol as R2C, VL-BERT outperforms R2C by large margins, indicating the power of our simple cross-modal architecture. Compared with other concurrent works, i.e., ViLBERT, VisualBERT and B2T2, our VL-BERT achieves the state-of-the-art performance. ",
+        "bbox": [
+            174,
+            724,
+            825,
+            823
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2.2 VISUAL QUESTION ANSWERING (VQA) ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            842,
+            508,
+            856
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the VQA task, given a natural image, a question at the perceptual level is asked, and the algorithm should generate / choose the correct answer. Here we conduct experiments on the widely-used VQA v2.0 dataset (Goyal et al., 2017), which is built based on the COCO (Lin et al., 2014) images. The VQA $\\mathrm { { v } } 2 . 0$ dataset is split into train ( $8 3 \\mathrm { k }$ images and 444k questions), validation (41k images and ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a733ca5a4f2a572b929e196e4727d7672b7b14f6b8acafac61e1d7c3d2af6b2c.jpg",
+        "table_caption": [
+            "Table 2: Comparison to the state-of-the-art methods with single model on the VQA dataset. $\\dagger$ indicates concurrent works. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>test-dev</td><td>test-std</td></tr><tr><td>BUTD (Anderson et al.,2018)</td><td>65.32</td><td>65.67</td></tr><tr><td>ViLBERT (Lu et al.,2019)†</td><td>70.55</td><td>70.92</td></tr><tr><td>VisualBERT (Li et al.,2019b)t LXMERT(Tan &amp; Bansal, 2019)†</td><td>70.80</td><td>71.00</td></tr><tr><td>VL-BERTBAsE W/o pre-training</td><td>72.42</td><td>72.54</td></tr><tr><td>VL-BERTBASE</td><td>69.58 71.16</td><td>1</td></tr><tr><td>VL-BERTLARGE</td><td></td><td>-</td></tr><tr><td></td><td>71.79</td><td>72.22</td></tr></table>",
+        "bbox": [
+            325,
+            101,
+            673,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "$2 1 4 \\mathrm { k }$ questions), and test (81k images and $4 4 8 \\mathrm { k }$ questions) sets. Following the experimental protocol in BUTD (Anderson et al., 2018), for each question, the algorithm should pick the corresponding answer from a shared set consisting of 3,129 answers. ",
+        "bbox": [
+            173,
+            275,
+            823,
+            318
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 (b) illustrates the input format for the VQA task, which is of $<$ <Question, Answer, Image>. As the possible answers are from a shared pool independent to the question, we only fill a [MASK] element to the Answer section. As in BUTD (Anderson et al., 2018), the input RoIs in VL-BERT are generated by a Faster R-CNN detector pre-trained on Visual Genome (Krishna et al., 2017). The answer prediction is made from a multi-class classifier based upon the output feature of the [MASK] element. During fine-tuning, the network training is driven by the multi-class cross-entropy loss over the possible answers. The detailed parameter settings are in Appendix. ",
+        "bbox": [
+            173,
+            324,
+            825,
+            422
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents our experimental results. Pre-training VL-BERT improves the performance by $1 . 6 \\%$ , which validates the importance of pre-training. VL-BERT shares the same input (i.e., question, image, and RoIs), output and experimental protocol with BUTD, a prevalent model specifically designed for the task. Still, VL-BERT surpasses BUTD by over $5 \\%$ in accuracy. Except for LXMERT, our VL-BERT achieves better performance than the other concurrent works. This is because LXMERT is pre-trained on massive visual question answering data (aggregating almost all the VQA datasets based on COCO and Visual Genome). While our model is only pre-trained on captioning and text-only dataset, where there is still gap with the VQA task. ",
+        "bbox": [
+            173,
+            429,
+            825,
+            541
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.2.3 REFERRING EXPRESSION COMPREHENSION ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            556,
+            531,
+            570
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/6588f65e136bfe960738b933792a9621aec18761277a989c2b13c7c502852dd4.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Comparison to the state-of-the-art methods with single model on the RefCOCO $^ +$ dataset. $\\dagger$ indicates concurrent work. "
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">Ground-truth Regions</td><td colspan=\"3\">Detected Regions</td></tr><tr><td>val</td><td>testA</td><td>testB</td><td>val</td><td>testA</td><td>testB</td></tr><tr><td>MAttNet (Yu et al.,2018)</td><td>71.01</td><td>75.13</td><td>66.17</td><td>65.33</td><td>71.62</td><td>56.02</td></tr><tr><td>ViLBERT (Lu et al.,2019)†</td><td>-</td><td>-</td><td>1</td><td>72.34</td><td>78.52</td><td>62.61</td></tr><tr><td>VL-BERTBAsE W/o pre-training</td><td>74.41</td><td>77.28</td><td>67.52</td><td>66.03</td><td>71.87</td><td>56.13</td></tr><tr><td>VL-BERTBASE</td><td>79.88</td><td>82.40</td><td>75.01</td><td>71.60</td><td>77.72</td><td>60.99</td></tr><tr><td>VL-BERTLARGE</td><td>80.31</td><td>83.62</td><td>75.45</td><td>72.59</td><td>78.57</td><td>62.30</td></tr></table>",
+        "bbox": [
+            235,
+            584,
+            761,
+            680
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A referring expression is a natural language phrase that refers to an object in an image. The referring expression comprehension task is to localize the object in an image with the given referring expression. We adopt the RefCOCO $^ +$ (Kazemzadeh et al., 2014) dataset for evaluation, consisting of 141k expressions for 50k referred objects in $2 0 \\mathrm { k }$ images in the COCO dataset (Lin et al., 2014). The referring expressions in $\\operatorname { R e f C O C O + }$ are forbidden from using absolute location words, e.g. left dog. Therefore the referring expressions focus on purely appearance-based descriptions. $\\operatorname { R e f C O C O + }$ are split into four sets, training set (train), validation set (val), and two testing sets (testA and testB). Images containing multiple people are in testA set, while images containing multiple objects of other categories are in testB set. There is no overlap between the training, validation and testing images. ",
+        "bbox": [
+            173,
+            734,
+            825,
+            861
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 (c) illustrates the input format for referring expression comprehension , where the input format is of $<$ <Query, Image>. Model training and evaluation are conducted either on the groundtruth RoIs or on the detected boxes in MAttNet (Yu et al., 2018). And the results are reported either in the track of ground-truth regions or that of detected regions, respectively. During training, we compute the classification scores for all the input RoIs. For each RoI, a binary classification loss is applied. During inference, we directly choose the RoI with the highest classification score as the referred object of the input referring expression. The detailed parameter settings are in Appendix. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            103,
+            823,
+            146
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 3 presents our experimental results. Pre-trained VL-BERT significantly improves the performance. Compared with MAttNet, VL-BERT is much simpler without task-specific architecture designs, yet much better. VL-BERT achieves comparable performance with the concurrent work of ViLBERT. ",
+        "bbox": [
+            174,
+            152,
+            825,
+            208
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "4.3 ABLATION STUDY ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            227,
+            341,
+            241
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5f0ef1e3e5d02f52868f717e3368a22827af7daa1f6956e7952aeb9e88d5210.jpg",
+        "table_caption": [
+            "Table 4: Ablation study for VL-BERTBASE with $0 . 5 \\times$ fine-tuning epochs. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Settings</td><td>MaskedLanguage Modeling with Visual Clues</td><td>Masked RoI Classification with Linguistic Clues</td><td>Sentence-Image Relationship Prediction</td><td>with Text-only Corpus</td><td>Tuning Fast R-CNN</td><td>VCR Q→A val val</td><td>QA→R test-dev</td><td>VQA</td><td>RefCOCO+ Detected Regions val</td></tr><tr><td>w/o pre-training</td><td></td><td></td><td></td><td></td><td></td><td>72.9</td><td>73.0</td><td>69.5</td><td>62.7</td></tr><tr><td>(a)</td><td></td><td></td><td></td><td></td><td></td><td>72.9</td><td>73.1</td><td>71.0</td><td>69.1</td></tr><tr><td>(b)</td><td>√ √</td><td>√</td><td></td><td></td><td></td><td>73.0</td><td>73.1</td><td>71.1</td><td>70.7</td></tr><tr><td>(c)</td><td>√</td><td>√</td><td>√</td><td></td><td></td><td>72.2</td><td>72.4</td><td>70.3</td><td>69.5</td></tr><tr><td>(d)</td><td>√</td><td>√</td><td></td><td>√</td><td></td><td>73.4</td><td>73.8</td><td>71.1</td><td>70.7</td></tr><tr><td>VL-BERTBASE</td><td>√</td><td>√</td><td></td><td>√</td><td>√</td><td>73.8</td><td>73.9</td><td>71.2</td><td>71.1</td></tr></table>",
+        "bbox": [
+            176,
+            257,
+            825,
+            354
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 4 ablates key design choices in pre-training VL-BERT. For experimental efficiency, the finetuning epoches of VL-BERT are of $0 . 5 \\times$ of those in Section 4.2, with only VL-BERTBASE model. ",
+        "bbox": [
+            174,
+            397,
+            825,
+            425
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Overall, the pre-training of VL-BERT improves the performance over all the three down-stream tasks (by comparing setting “w/o pre-training” and VL-BERTBASE). The improvement amplitude varies for different tasks. By comparing setting (a) to that of “w/o pre-training”, we see the benefits of Task #1, Masked Language Modeling with Visual Clues. By further incorporating Task #2, Masked RoI Classification with Linguistic Clues, the accuracy further improves on $\\operatorname { R e f C O C O + }$ , but gets stuck at VCR and VQA. This might be because only $\\operatorname { R e f C O C O + }$ utilizes the final output feature corresponding to [IMG] tokens for prediction. Thus the pre-training of such features is beneficial. Setting (c) incorporates the task of Sentence-Image Relationship Prediction as in ViLBERT (Lu et al., 2019) and LXMERT (Tan & Bansal, 2019). It would hurt accuracy on all the three downstream tasks. We guess the reason is because the task of Sentence-Image Relationship Prediction would introduce unmatched image and caption pairs as negative examples. Such unmatched samples would hamper the training of other tasks. Setting (d) adds text-only corpus during pre-training. Compared with setting (b), it improves the performance over all three down-stream tasks, and is most significant on VCR. This is because the task of VCR involves more complex and longer sentences than those in VQA and $_ \\mathrm { R e f C O C O + } 2$ . By further finetuning the network parameters of Fast R-CNN, which generates the visual features, we get the final setting of VL-BERTBASE. Such end-to-end training of the entire network is helpful for all the downstream tasks. ",
+        "bbox": [
+            173,
+            433,
+            825,
+            667
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "5 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            689,
+            318,
+            705
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we developed VL-BERT, a new pre-trainable generic representation for visuallinguistic tasks. Instead of using ad-hoc task-specific modules, VL-BERT adopts the simple yet powerful Transformer model as the backbone. It is pre-trained on the massive-scale Conceptual Captions dataset, together with text-only corpus. Extensive empirical analysis demonstrates that the pre-training procedure can better align the visual-linguistic clues, and thus benefit the downstream tasks. In the future, we would like to seek better pre-training tasks, which could beneficial more downstream tasks (e.g., Image Caption Generation). ",
+        "bbox": [
+            174,
+            722,
+            825,
+            819
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "ACKNOWLEDGMENTS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            837,
+            326,
+            849
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The work is partially supported by the National Natural Science Foundation of China under grand No.U19B2044 and No.61836011. ",
+        "bbox": [
+            176,
+            859,
+            821,
+            887
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            287,
+            118
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Chris Alberti, Jeffrey Ling, Michael Collins, and David Reitter. Fusion of detected objects in text for visual question answering. arXiv preprint arXiv:1908.05054, 2019. ",
+        "bbox": [
+            173,
+            126,
+            823,
+            155
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6077–6086, 2018. ",
+        "bbox": [
+            174,
+            165,
+            826,
+            222
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pp. 2425–2433, 2015. ",
+        "bbox": [
+            174,
+            232,
+            825,
+            275
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and ´ C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. ",
+        "bbox": [
+            173,
+            285,
+            826,
+            329
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee, 2009. ",
+        "bbox": [
+            173,
+            338,
+            823,
+            381
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. ",
+        "bbox": [
+            169,
+            391,
+            823,
+            421
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In International conference on machine learning, pp. 647–655, 2014. ",
+        "bbox": [
+            173,
+            430,
+            825,
+            474
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. From two graphs to n questions: A vqa dataset for compositional reasoning on vision and commonsense. arXiv preprint arXiv:1908.02962, 2019. ",
+        "bbox": [
+            173,
+            484,
+            825,
+            526
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pp. 1440–1448, 2015. ",
+        "bbox": [
+            173,
+            536,
+            823,
+            566
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 580–587, 2014. ",
+        "bbox": [
+            174,
+            575,
+            823,
+            619
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6904–6913, 2017. ",
+        "bbox": [
+            176,
+            628,
+            823,
+            672
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Simultaneous detection and ´ segmentation. In European Conference on Computer Vision, pp. 297–312. Springer, 2014. ",
+        "bbox": [
+            174,
+            683,
+            823,
+            712
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imagenet pre-training. ´ arXiv preprint arXiv:1811.08883, 2018. ",
+        "bbox": [
+            173,
+            722,
+            825,
+            751
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3588–3597, 2018. ",
+        "bbox": [
+            173,
+            761,
+            825,
+            804
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6700–6709, 2019. ",
+        "bbox": [
+            173,
+            814,
+            823,
+            857
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2901–2910, 2017. ",
+        "bbox": [
+            174,
+            867,
+            825,
+            924
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 787–798, 2014. ",
+        "bbox": [
+            176,
+            103,
+            821,
+            147
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ",
+        "bbox": [
+            171,
+            154,
+            823,
+            183
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In Advances in neural information processing systems, pp. 3294–3302, 2015. ",
+        "bbox": [
+            176,
+            190,
+            823,
+            233
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32–73, 2017. ",
+        "bbox": [
+            173,
+            241,
+            825,
+            297
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ",
+        "bbox": [
+            173,
+            306,
+            823,
+            348
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291, 2019. ",
+        "bbox": [
+            174,
+            356,
+            823,
+            386
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training, 2019a. ",
+        "bbox": [
+            171,
+            392,
+            823,
+            422
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019b. ",
+        "bbox": [
+            171,
+            429,
+            823,
+            459
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ European conference on computer vision, pp. 740–755. Springer, 2014. ",
+        "bbox": [
+            176,
+            467,
+            823,
+            511
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. ",
+        "bbox": [
+            176,
+            517,
+            823,
+            560
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, 2015. ",
+        "bbox": [
+            171,
+            568,
+            823,
+            611
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019. ",
+        "bbox": [
+            173,
+            618,
+            820,
+            648
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. ",
+        "bbox": [
+            171,
+            655,
+            821,
+            685
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014. ",
+        "bbox": [
+            173,
+            691,
+            823,
+            736
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. ",
+        "bbox": [
+            169,
+            742,
+            823,
+            772
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8), 2019. ",
+        "bbox": [
+            171,
+            780,
+            823,
+            809
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015. ",
+        "bbox": [
+            174,
+            816,
+            823,
+            859
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2556–2565, 2018. ",
+        "bbox": [
+            174,
+            867,
+            825,
+            924
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive bidirectional transformer for temporal representation learning. arXiv preprint arXiv:1906.05743, 2019a. ",
+        "bbox": [
+            173,
+            103,
+            823,
+            133
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. arXiv preprint arXiv:1904.01766, 2019b. ",
+        "bbox": [
+            173,
+            140,
+            823,
+            171
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. ",
+        "bbox": [
+            176,
+            179,
+            821,
+            222
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. ",
+        "bbox": [
+            174,
+            231,
+            823,
+            273
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Jesse Vig. A multiscale visualization of attention in the transformer model. arXiv preprint arXiv:1906.05714, 2019. URL https://arxiv.org/abs/1906.05714. ",
+        "bbox": [
+            174,
+            281,
+            820,
+            311
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Alex Wang and Kyunghyun Cho. Bert has a mouth, and it must speak: Bert as a markov random field language model. arXiv preprint arXiv:1902.04094, 2019. ",
+        "bbox": [
+            173,
+            319,
+            820,
+            349
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. ",
+        "bbox": [
+            173,
+            357,
+            825,
+            414
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237, 2019. ",
+        "bbox": [
+            173,
+            422,
+            823,
+            465
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67–78, 2014. ",
+        "bbox": [
+            173,
+            473,
+            823,
+            517
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1307–1315, 2018. ",
+        "bbox": [
+            173,
+            525,
+            823,
+            569
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6720–6731, 2019. ",
+        "bbox": [
+            174,
+            577,
+            823,
+            621
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4995–5004, 2016. ",
+        "bbox": [
+            173,
+            628,
+            823,
+            672
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pp. 19–27, 2015. ",
+        "bbox": [
+            174,
+            680,
+            826,
+            738
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A APPENDIX ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            297,
+            117
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A.1 COMPARISON AMONG VL-BERT AND OTHER WORKS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            138,
+            588,
+            152
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 5 compares among VL-BERT and other concurrent works for pre-training generic visuallinguistic representations. ",
+        "bbox": [
+            174,
+            167,
+            821,
+            196
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/3268a63a92f11643105fc42a016f6c677a3a1bbb39847efa4d29a6d2f018d993.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "‡ LXMERT is pre-trained on COCO Caption (Chen et al., 2015), VG Caption (Krishna et al., 2017), VG QA (Zhu et al., 2016), VQA (Antol et al., 2015) and GQA (Hudson & Manning, 2019). "
+        ],
+        "table_body": "<table><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>Method</td><td rowspan=1 colspan=1>Architecture</td><td rowspan=1 colspan=1>Visual Token</td><td rowspan=1 colspan=1>Pre-train Datasets</td><td rowspan=1 colspan=1>Pre-train Tasks</td><td rowspan=1 colspan=1>Downstream Tasks</td></tr><tr><td rowspan=1 colspan=1>PublishedWorks</td><td rowspan=1 colspan=1>VideoBERT(Sun et al., 2019b)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>video frame</td><td rowspan=1 colspan=1>Cooking312K(Sun et al., 2019b)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-words prediction</td><td rowspan=1 colspan=1>1) zero-shot action classification2) video captioning</td></tr><tr><td rowspan=7 colspan=1>WorksUnderReview /Just GotAccepted</td><td rowspan=1 colspan=1>CBT(Sun et al., 2019a)</td><td rowspan=1 colspan=1>two single-modal Transformer(vision &amp; language respectively)+ one cross-modal Transformer</td><td rowspan=1 colspan=1>video frame</td><td rowspan=1 colspan=1>Cooking312K(Sun et al., 2019b)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature regression</td><td rowspan=1 colspan=1>1) action anticipation2) video captioning</td></tr><tr><td rowspan=1 colspan=1>ViLBERT(Lu et al., 2019)</td><td rowspan=1 colspan=1>one single-modal Transformer(language)+ one cross-modal Transformer(with restricted attention pattern)</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al., 2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature classification</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) grounding referring expressions4) image retrieval5) zero-shot image retrieval</td></tr><tr><td rowspan=1 colspan=1>B2T2(Alberti et al., 2019)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al.,2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling</td><td rowspan=1 colspan=1>1) visual commonsense reasoning</td></tr><tr><td rowspan=1 colspan=1>LXMERT(Tan &amp; Bansal, 2019)</td><td rowspan=1 colspan=1>two single-modal Transformer(vision &amp; language respectively)+ one cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>COCO Caption+ VG Caption+ VG QA+VQA+GQA</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3) masked visual-feature classification4) masked visual-feature regression5) visual question answering</td><td rowspan=1 colspan=1>1) visual question answering2) natural language visual reasoning</td></tr><tr><td rowspan=1 colspan=1>VisualBERT(Li et al., 2019b)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>COCO Caption(Chen et al., 2015)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) natural language visual reasoning4) grounding phrases</td></tr><tr><td rowspan=1 colspan=1>Unicoder-VL(Li et al., 2019a)</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al.,2018)</td><td rowspan=1 colspan=1>1) sentence-image alignment2) masked language modeling3)maskedvisual-featureclassification</td><td rowspan=1 colspan=1>1) image-text retrieval2)zero-shotimage-textretrieval</td></tr><tr><td rowspan=1 colspan=1>Our VL-BERT</td><td rowspan=1 colspan=1>single cross-modal Transformer</td><td rowspan=1 colspan=1>image RoI</td><td rowspan=1 colspan=1>Conceptual Captions(Sharma et al., 2018)+ BooksCorpus(Zhu et al.,2015)+ English Wikipedia</td><td rowspan=1 colspan=1>1) masked language modeling2) masked visual-feature classification</td><td rowspan=1 colspan=1>1) visual question answering2) visual commonsense reasoning3) grounding referring expressions</td></tr></table>",
+        "bbox": [
+            176,
+            217,
+            823,
+            463
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Comparison among our VL-BERT and other works seeking to derive pre-trainable generic representations for visual-linguistic tasks. ",
+        "bbox": [
+            174,
+            483,
+            821,
+            511
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A.2 DETAILED EXPERIMENT SETTINGS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            553,
+            457,
+            566
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Pre-training is conducted on 16 Tesla V100 GPUs for $2 5 0 \\mathrm { k }$ iterations by SGD. In each mini-batch, 256 samples are drawn. Among them, 128 samples are of ${ < } \\mathbf { C }$ aption, Image $>$ pairs from Conceptual Captions, and the rest 128 samples are sequential tokens (at most 64 tokens for each sequence) from BooksCorpus & English Wikipedia. In SGD, Adam optimizer (Kingma & Ba, 2014) is applied, with base learning rate of $2 \\times 1 0 ^ { - 5 }$ , $\\beta _ { 1 } = 0 . 9$ , $\\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 0 ^ { - 4 }$ , learning rate warmed up over the first 8,000 steps, and linear decay of the learning rate. All the parameters in VL-BERT and Fast R-CNN are jointly trained in both pre-training and fine-tuning phase. The visual feature input for textual corpus is a learnable embedding shared for all words. In the task of Masked RoI Classification with Linguistic Clues, the pixels lying in all the masked RoIs are set as zeros in the image. A box covering the whole image is added as a RoI and would not be masked. ",
+        "bbox": [
+            173,
+            582,
+            825,
+            722
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "For VCR, the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each mini-batch, 256 triplets of <Question, Answer, Image $>$ are sampled. In SGD, the basic mini-batch gradient descent is conducted, with base learning rate of $5 \\times \\mathrm { 1 \\bar { 0 } ^ { - 3 } }$ , momentum of 0.9, and weight decay of $1 0 ^ { - 4 }$ . The learning rate is linearly warmed up in the first 1,000 steps from an initial learning rate of 0, and is decayed by 0.1 at the 14-th and the 18-th epochs. ",
+        "bbox": [
+            174,
+            728,
+            825,
+            797
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "For VQA, the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each mini-batch, 256 triplets of $<$ <Question, Answer, Image $>$ are sampled. In SGD, Adam optimizer is applied, with base learning rate of $1 \\times 1 0 ^ { - 4 }$ , $\\beta _ { 1 } = 0 . 9$ , $\\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 0 ^ { - 4 }$ , learning rate warmed up over the first 2,000 steps, and linear decay of the learning rate. ",
+        "bbox": [
+            174,
+            804,
+            823,
+            861
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "For $\\operatorname { R e f C O C O + }$ , the fine-tuning is conducted on 16 Tesla V100 GPUs for 20 epochs. In each minibatch, 256 pairs of $<$ <Query, Image $>$ are sampled. In SGD, Adam optimizer is applied, with base learning rate of $1 \\times 1 0 ^ { - 4 }$ , $\\beta _ { 1 } = 0 . 9$ , $\\beta _ { 2 } = 0 . 9 9 9$ , weight decay of $1 \\dot { 0 } ^ { - 4 }$ , learning rate warmed up over the first 500 steps, and linear decay of the learning rate. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/27d26e5998a446d208137b24004ce833aead6fd5de0e5fa774c2524c34106ac3.jpg",
+        "image_caption": [
+            "Figure 3: Visualization of attention maps in pre-trained VL-BERTBASE. Line intensity indicates the magnitude of attention probability with the text token as query and the image RoI as key. The intensity is affinely rescaled to set the maximum value as 1 and the minimum as 0, across different heads in each layer. The index of network layer and attention head is counted from 0. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            225,
+            89,
+            799,
+            852
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "A.3 VISUALIZATION OF ATTENTION MAPS IN VL-BERT ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            103,
+            576,
+            117
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "To better understand what VL-BERT learns from pre-training, we visualized the attention maps of pre-trained VL-BERT (without fine-tuning on downstream tasks) using BertViz3(Vig, 2019). ",
+        "bbox": [
+            171,
+            128,
+            823,
+            159
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Some visualization results on COCO (Lin et al., 2014; Chen et al., 2015) val2017 set are shown in Figure 3. We can see different attention patterns across attention heads. For some attention heads, text tokens attend more on the associated image RoIs. While in some other heads, text tokens attend uniformly to all RoIs. It demonstrates the ability of VL-BERT in aggregating and aligning visuallinguistic contents. ",
+        "bbox": [
+            174,
+            165,
+            825,
+            234
+        ],
+        "page_idx": 15
+    }
+]